Bryan Wu | 1394f03 | 2007-05-06 14:50:22 -0700 | [diff] [blame] | 1 | /* |
Robin Getz | 96f1050 | 2009-09-24 14:11:24 +0000 | [diff] [blame^] | 2 | * Copyright 2004-2009 Analog Devices Inc. |
Bryan Wu | 1394f03 | 2007-05-06 14:50:22 -0700 | [diff] [blame] | 3 | * |
Robin Getz | 96f1050 | 2009-09-24 14:11:24 +0000 | [diff] [blame^] | 4 | * Licensed under the ADI BSD license or the GPL-2 (or later) |
Bryan Wu | 1394f03 | 2007-05-06 14:50:22 -0700 | [diff] [blame] | 5 | */ |
| 6 | |
| 7 | #include <linux/linkage.h> |
| 8 | |
| 9 | #define CARRY AC0 |
| 10 | |
| 11 | #ifdef CONFIG_ARITHMETIC_OPS_L1 |
| 12 | .section .l1.text |
| 13 | #else |
| 14 | .text |
| 15 | #endif |
| 16 | |
| 17 | |
| 18 | ENTRY(___udivsi3) |
| 19 | |
| 20 | CC = R0 < R1 (IU); /* If X < Y, always return 0 */ |
| 21 | IF CC JUMP .Lreturn_ident; |
| 22 | |
| 23 | R2 = R1 << 16; |
| 24 | CC = R2 <= R0 (IU); |
| 25 | IF CC JUMP .Lidents; |
| 26 | |
| 27 | R2 = R0 >> 31; /* if X is a 31-bit number */ |
| 28 | R3 = R1 >> 15; /* and Y is a 15-bit number */ |
| 29 | R2 = R2 | R3; /* then it's okay to use the DIVQ builtins (fallthrough to fast)*/ |
| 30 | CC = R2; |
| 31 | IF CC JUMP .Ly_16bit; |
| 32 | |
| 33 | /* METHOD 1: FAST DIVQ |
| 34 | We know we have a 31-bit dividend, and 15-bit divisor so we can use the |
| 35 | simple divq approach (first setting AQ to 0 - implying unsigned division, |
| 36 | then 16 DIVQ's). |
| 37 | */ |
| 38 | |
| 39 | AQ = CC; /* Clear AQ (CC==0) */ |
| 40 | |
| 41 | /* ISR States: When dividing two integers (32.0/16.0) using divide primitives, |
| 42 | we need to shift the dividend one bit to the left. |
| 43 | We have already checked that we have a 31-bit number so we are safe to do |
| 44 | that. |
| 45 | */ |
| 46 | R0 <<= 1; |
| 47 | DIVQ(R0, R1); // 1 |
| 48 | DIVQ(R0, R1); // 2 |
| 49 | DIVQ(R0, R1); // 3 |
| 50 | DIVQ(R0, R1); // 4 |
| 51 | DIVQ(R0, R1); // 5 |
| 52 | DIVQ(R0, R1); // 6 |
| 53 | DIVQ(R0, R1); // 7 |
| 54 | DIVQ(R0, R1); // 8 |
| 55 | DIVQ(R0, R1); // 9 |
| 56 | DIVQ(R0, R1); // 10 |
| 57 | DIVQ(R0, R1); // 11 |
| 58 | DIVQ(R0, R1); // 12 |
| 59 | DIVQ(R0, R1); // 13 |
| 60 | DIVQ(R0, R1); // 14 |
| 61 | DIVQ(R0, R1); // 15 |
| 62 | DIVQ(R0, R1); // 16 |
| 63 | R0 = R0.L (Z); |
| 64 | RTS; |
| 65 | |
| 66 | .Ly_16bit: |
| 67 | /* We know that the upper 17 bits of Y might have bits set, |
| 68 | ** or that the sign bit of X might have a bit. If Y is a |
| 69 | ** 16-bit number, but not bigger, then we can use the builtins |
| 70 | ** with a post-divide correction. |
| 71 | ** R3 currently holds Y>>15, which means R3's LSB is the |
| 72 | ** bit we're interested in. |
| 73 | */ |
| 74 | |
| 75 | /* According to the ISR, to use the Divide primitives for |
| 76 | ** unsigned integer divide, the useable range is 31 bits |
| 77 | */ |
| 78 | CC = ! BITTST(R0, 31); |
| 79 | |
| 80 | /* IF condition is true we can scale our inputs and use the divide primitives, |
| 81 | ** with some post-adjustment |
| 82 | */ |
| 83 | R3 += -1; /* if so, Y is 0x00008nnn */ |
| 84 | CC &= AZ; |
| 85 | |
| 86 | /* If condition is true we can scale our inputs and use the divide primitives, |
| 87 | ** with some post-adjustment |
| 88 | */ |
| 89 | R3 = R1 >> 1; /* Pre-scaled divisor for primitive case */ |
| 90 | R2 = R0 >> 16; |
| 91 | |
| 92 | R2 = R3 - R2; /* shifted divisor < upper 16 bits of dividend */ |
| 93 | CC &= CARRY; |
| 94 | IF CC JUMP .Lshift_and_correct; |
| 95 | |
| 96 | /* Fall through to the identities */ |
| 97 | |
| 98 | /* METHOD 2: identities and manual calculation |
| 99 | We are not able to use the divide primites, but may still catch some special |
| 100 | cases. |
| 101 | */ |
| 102 | .Lidents: |
| 103 | /* Test for common identities. Value to be returned is placed in R2. */ |
| 104 | CC = R0 == 0; /* 0/Y => 0 */ |
| 105 | IF CC JUMP .Lreturn_r0; |
| 106 | CC = R0 == R1; /* X==Y => 1 */ |
| 107 | IF CC JUMP .Lreturn_ident; |
| 108 | CC = R1 == 1; /* X/1 => X */ |
| 109 | IF CC JUMP .Lreturn_ident; |
| 110 | |
| 111 | R2.L = ONES R1; |
| 112 | R2 = R2.L (Z); |
| 113 | CC = R2 == 1; |
| 114 | IF CC JUMP .Lpower_of_two; |
| 115 | |
| 116 | [--SP] = (R7:5); /* Push registers R5-R7 */ |
| 117 | |
| 118 | /* Idents don't match. Go for the full operation. */ |
| 119 | |
| 120 | |
| 121 | R6 = 2; /* assume we'll shift two */ |
| 122 | R3 = 1; |
| 123 | |
| 124 | P2 = R1; |
| 125 | /* If either R0 or R1 have sign set, */ |
| 126 | /* divide them by two, and note it's */ |
| 127 | /* been done. */ |
| 128 | CC = R1 < 0; |
| 129 | R2 = R1 >> 1; |
| 130 | IF CC R1 = R2; /* Possibly-shifted R1 */ |
| 131 | IF !CC R6 = R3; /* R1 doesn't, so at most 1 shifted */ |
| 132 | |
| 133 | P0 = 0; |
| 134 | R3 = -R1; |
| 135 | [--SP] = R3; |
| 136 | R2 = R0 >> 1; |
| 137 | R2 = R0 >> 1; |
| 138 | CC = R0 < 0; |
| 139 | IF CC P0 = R6; /* Number of values divided */ |
| 140 | IF !CC R2 = R0; /* Shifted R0 */ |
| 141 | |
| 142 | /* P0 is 0, 1 (NR/=2) or 2 (NR/=2, DR/=2) */ |
| 143 | |
| 144 | /* r2 holds Copy dividend */ |
| 145 | R3 = 0; /* Clear partial remainder */ |
| 146 | R7 = 0; /* Initialise quotient bit */ |
| 147 | |
| 148 | P1 = 32; /* Set loop counter */ |
| 149 | LSETUP(.Lulst, .Lulend) LC0 = P1; /* Set loop counter */ |
| 150 | .Lulst: R6 = R2 >> 31; /* R6 = sign bit of R2, for carry */ |
| 151 | R2 = R2 << 1; /* Shift 64 bit dividend up by 1 bit */ |
| 152 | R3 = R3 << 1 || R5 = [SP]; |
| 153 | R3 = R3 | R6; /* Include any carry */ |
| 154 | CC = R7 < 0; /* Check quotient(AQ) */ |
| 155 | /* If AQ==0, we'll sub divisor */ |
| 156 | IF CC R5 = R1; /* and if AQ==1, we'll add it. */ |
| 157 | R3 = R3 + R5; /* Add/sub divsor to partial remainder */ |
| 158 | R7 = R3 ^ R1; /* Generate next quotient bit */ |
| 159 | |
| 160 | R5 = R7 >> 31; /* Get AQ */ |
| 161 | BITTGL(R5, 0); /* Invert it, to get what we'll shift */ |
| 162 | .Lulend: R2 = R2 + R5; /* and "shift" it in. */ |
| 163 | |
| 164 | CC = P0 == 0; /* Check how many inputs we shifted */ |
| 165 | IF CC JUMP .Lno_mult; /* if none... */ |
| 166 | R6 = R2 << 1; |
| 167 | CC = P0 == 1; |
| 168 | IF CC R2 = R6; /* if 1, Q = Q*2 */ |
| 169 | IF !CC R1 = P2; /* if 2, restore stored divisor */ |
| 170 | |
| 171 | R3 = R2; /* Copy of R2 */ |
| 172 | R3 *= R1; /* Q * divisor */ |
| 173 | R5 = R0 - R3; /* Z = (dividend - Q * divisor) */ |
| 174 | CC = R1 <= R5 (IU); /* Check if divisor <= Z? */ |
| 175 | R6 = CC; /* if yes, R6 = 1 */ |
| 176 | R2 = R2 + R6; /* if yes, add one to quotient(Q) */ |
| 177 | .Lno_mult: |
| 178 | SP += 4; |
| 179 | (R7:5) = [SP++]; /* Pop registers R5-R7 */ |
| 180 | R0 = R2; /* Store quotient */ |
| 181 | RTS; |
| 182 | |
| 183 | .Lreturn_ident: |
| 184 | CC = R0 < R1 (IU); /* If X < Y, always return 0 */ |
| 185 | R2 = 0; |
| 186 | IF CC JUMP .Ltrue_return_ident; |
| 187 | R2 = -1 (X); /* X/0 => 0xFFFFFFFF */ |
| 188 | CC = R1 == 0; |
| 189 | IF CC JUMP .Ltrue_return_ident; |
| 190 | R2 = -R2; /* R2 now 1 */ |
| 191 | CC = R0 == R1; /* X==Y => 1 */ |
| 192 | IF CC JUMP .Ltrue_return_ident; |
| 193 | R2 = R0; /* X/1 => X */ |
| 194 | /*FALLTHRU*/ |
| 195 | |
| 196 | .Ltrue_return_ident: |
| 197 | R0 = R2; |
| 198 | .Lreturn_r0: |
| 199 | RTS; |
| 200 | |
| 201 | .Lpower_of_two: |
| 202 | /* Y has a single bit set, which means it's a power of two. |
| 203 | ** That means we can perform the division just by shifting |
| 204 | ** X to the right the appropriate number of bits |
| 205 | */ |
| 206 | |
| 207 | /* signbits returns the number of sign bits, minus one. |
| 208 | ** 1=>30, 2=>29, ..., 0x40000000=>0. Which means we need |
| 209 | ** to shift right n-signbits spaces. It also means 0x80000000 |
| 210 | ** is a special case, because that *also* gives a signbits of 0 |
| 211 | */ |
| 212 | |
| 213 | R2 = R0 >> 31; |
| 214 | CC = R1 < 0; |
| 215 | IF CC JUMP .Ltrue_return_ident; |
| 216 | |
| 217 | R1.l = SIGNBITS R1; |
| 218 | R1 = R1.L (Z); |
| 219 | R1 += -30; |
| 220 | R0 = LSHIFT R0 by R1.L; |
| 221 | RTS; |
| 222 | |
| 223 | /* METHOD 3: PRESCALE AND USE THE DIVIDE PRIMITIVES WITH SOME POST-CORRECTION |
| 224 | Two scaling operations are required to use the divide primitives with a |
| 225 | divisor > 0x7FFFF. |
| 226 | Firstly (as in method 1) we need to shift the dividend 1 to the left for |
| 227 | integer division. |
| 228 | Secondly we need to shift both the divisor and dividend 1 to the right so |
| 229 | both are in range for the primitives. |
| 230 | The left/right shift of the dividend does nothing so we can skip it. |
| 231 | */ |
| 232 | .Lshift_and_correct: |
| 233 | R2 = R0; |
| 234 | // R3 is already R1 >> 1 |
| 235 | CC=!CC; |
| 236 | AQ = CC; /* Clear AQ, got here with CC = 0 */ |
| 237 | DIVQ(R2, R3); // 1 |
| 238 | DIVQ(R2, R3); // 2 |
| 239 | DIVQ(R2, R3); // 3 |
| 240 | DIVQ(R2, R3); // 4 |
| 241 | DIVQ(R2, R3); // 5 |
| 242 | DIVQ(R2, R3); // 6 |
| 243 | DIVQ(R2, R3); // 7 |
| 244 | DIVQ(R2, R3); // 8 |
| 245 | DIVQ(R2, R3); // 9 |
| 246 | DIVQ(R2, R3); // 10 |
| 247 | DIVQ(R2, R3); // 11 |
| 248 | DIVQ(R2, R3); // 12 |
| 249 | DIVQ(R2, R3); // 13 |
| 250 | DIVQ(R2, R3); // 14 |
| 251 | DIVQ(R2, R3); // 15 |
| 252 | DIVQ(R2, R3); // 16 |
| 253 | |
| 254 | /* According to the Instruction Set Reference: |
| 255 | To divide by a divisor > 0x7FFF, |
| 256 | 1. prescale and perform divide to obtain quotient (Q) (done above), |
| 257 | 2. multiply quotient by unscaled divisor (result M) |
| 258 | 3. subtract the product from the divident to get an error (E = X - M) |
| 259 | 4. if E < divisor (Y) subtract 1, if E > divisor (Y) add 1, else return quotient (Q) |
| 260 | */ |
| 261 | R3 = R2.L (Z); /* Q = X' / Y' */ |
| 262 | R2 = R3; /* Preserve Q */ |
| 263 | R2 *= R1; /* M = Q * Y */ |
| 264 | R2 = R0 - R2; /* E = X - M */ |
| 265 | R0 = R3; /* Copy Q into result reg */ |
| 266 | |
| 267 | /* Correction: If result of the multiply is negative, we overflowed |
| 268 | and need to correct the result by subtracting 1 from the result.*/ |
| 269 | R3 = 0xFFFF (Z); |
| 270 | R2 = R2 >> 16; /* E >> 16 */ |
| 271 | CC = R2 == R3; |
| 272 | R3 = 1 ; |
| 273 | R1 = R0 - R3; |
| 274 | IF CC R0 = R1; |
| 275 | RTS; |
Mike Frysinger | 51be24c | 2007-06-11 15:31:30 +0800 | [diff] [blame] | 276 | |
| 277 | ENDPROC(___udivsi3) |