Aubrey Li | 155fd76 | 2007-04-05 18:31:18 +0800 | [diff] [blame] | 1 | /* Copyright (C) 2003-2007 Analog Devices Inc. |
Aubrey Li | 26bf7de | 2007-03-19 01:24:52 +0800 | [diff] [blame] | 2 | * |
| 3 | * This file is subject to the terms and conditions of the GNU General Public |
| 4 | * License. |
| 5 | */ |
Aubrey Li | 155fd76 | 2007-04-05 18:31:18 +0800 | [diff] [blame] | 6 | |
Aubrey Li | 26bf7de | 2007-03-19 01:24:52 +0800 | [diff] [blame] | 7 | #define ASSEMBLY |
| 8 | |
| 9 | #include <asm/linkage.h> |
| 10 | #include <asm/cplb.h> |
| 11 | #include <config.h> |
| 12 | #include <asm/blackfin.h> |
| 13 | |
| 14 | .text |
| 15 | |
| 16 | /* This is an external function being called by the user |
| 17 | * application through __flush_cache_all. Currently this function |
| 18 | * serves the purpose of flushing all the pending writes in |
| 19 | * in the instruction cache. |
| 20 | */ |
| 21 | |
| 22 | ENTRY(_flush_instruction_cache) |
| 23 | [--SP] = ( R7:6, P5:4 ); |
| 24 | LINK 12; |
| 25 | SP += -12; |
| 26 | P5.H = (ICPLB_ADDR0 >> 16); |
| 27 | P5.L = (ICPLB_ADDR0 & 0xFFFF); |
| 28 | P4.H = (ICPLB_DATA0 >> 16); |
| 29 | P4.L = (ICPLB_DATA0 & 0xFFFF); |
| 30 | R7 = CPLB_VALID | CPLB_L1_CHBL; |
| 31 | R6 = 16; |
| 32 | inext: R0 = [P5++]; |
| 33 | R1 = [P4++]; |
| 34 | [--SP] = RETS; |
| 35 | CALL _icplb_flush; /* R0 = page, R1 = data*/ |
| 36 | RETS = [SP++]; |
| 37 | iskip: R6 += -1; |
| 38 | CC = R6; |
| 39 | IF CC JUMP inext; |
| 40 | SSYNC; |
| 41 | SP += 12; |
| 42 | UNLINK; |
| 43 | ( R7:6, P5:4 ) = [SP++]; |
| 44 | RTS; |
| 45 | |
| 46 | /* This is an internal function to flush all pending |
| 47 | * writes in the cache associated with a particular ICPLB. |
| 48 | * |
| 49 | * R0 - page's start address |
| 50 | * R1 - CPLB's data field. |
| 51 | */ |
| 52 | |
| 53 | .align 2 |
| 54 | ENTRY(_icplb_flush) |
| 55 | [--SP] = ( R7:0, P5:0 ); |
| 56 | [--SP] = LC0; |
| 57 | [--SP] = LT0; |
| 58 | [--SP] = LB0; |
| 59 | [--SP] = LC1; |
| 60 | [--SP] = LT1; |
| 61 | [--SP] = LB1; |
| 62 | |
| 63 | /* If it's a 1K or 4K page, then it's quickest to |
| 64 | * just systematically flush all the addresses in |
| 65 | * the page, regardless of whether they're in the |
| 66 | * cache, or dirty. If it's a 1M or 4M page, there |
| 67 | * are too many addresses, and we have to search the |
| 68 | * cache for lines corresponding to the page. |
| 69 | */ |
| 70 | |
| 71 | CC = BITTST(R1, 17); /* 1MB or 4MB */ |
| 72 | IF !CC JUMP iflush_whole_page; |
| 73 | |
| 74 | /* We're only interested in the page's size, so extract |
| 75 | * this from the CPLB (bits 17:16), and scale to give an |
| 76 | * offset into the page_size and page_prefix tables. |
| 77 | */ |
| 78 | |
| 79 | R1 <<= 14; |
| 80 | R1 >>= 30; |
| 81 | R1 <<= 2; |
| 82 | |
| 83 | /* We can also determine the sub-bank used, because this is |
| 84 | * taken from bits 13:12 of the address. |
| 85 | */ |
| 86 | |
| 87 | R3 = ((12<<8)|2); /* Extraction pattern */ |
| 88 | nop; /* Anamoly 05000209 */ |
| 89 | R4 = EXTRACT(R0, R3.L) (Z); /* Extract bits */ |
| 90 | |
| 91 | /* Save in extraction pattern for later deposit. */ |
| 92 | R3.H = R4.L << 0; |
| 93 | |
| 94 | /* So: |
| 95 | * R0 = Page start |
| 96 | * R1 = Page length (actually, offset into size/prefix tables) |
| 97 | * R3 = sub-bank deposit values |
| 98 | * |
| 99 | * The cache has 2 Ways, and 64 sets, so we iterate through |
| 100 | * the sets, accessing the tag for each Way, for our Bank and |
| 101 | * sub-bank, looking for dirty, valid tags that match our |
| 102 | * address prefix. |
| 103 | */ |
| 104 | |
| 105 | P5.L = (ITEST_COMMAND & 0xFFFF); |
| 106 | P5.H = (ITEST_COMMAND >> 16); |
| 107 | P4.L = (ITEST_DATA0 & 0xFFFF); |
| 108 | P4.H = (ITEST_DATA0 >> 16); |
| 109 | |
| 110 | P0.L = page_prefix_table; |
| 111 | P0.H = page_prefix_table; |
| 112 | P1 = R1; |
| 113 | R5 = 0; /* Set counter*/ |
| 114 | P0 = P1 + P0; |
| 115 | R4 = [P0]; /* This is the address prefix*/ |
| 116 | |
| 117 | /* We're reading (bit 1==0) the tag (bit 2==0), and we |
| 118 | * don't care about which double-word, since we're only |
| 119 | * fetching tags, so we only have to set Set, Bank, |
| 120 | * Sub-bank and Way. |
| 121 | */ |
| 122 | |
| 123 | P2 = 4; |
| 124 | LSETUP (ifs1, ife1) LC1 = P2; |
| 125 | ifs1: P0 = 32; /* iterate over all sets*/ |
| 126 | LSETUP (ifs0, ife0) LC0 = P0; |
| 127 | ifs0: R6 = R5 << 5; /* Combine set*/ |
| 128 | R6.H = R3.H << 0 ; /* and sub-bank*/ |
| 129 | [P5] = R6; /* Issue Command*/ |
| 130 | SSYNC; /* CSYNC will not work here :(*/ |
| 131 | R7 = [P4]; /* and read Tag.*/ |
| 132 | CC = BITTST(R7, 0); /* Check if valid*/ |
| 133 | IF !CC JUMP ifskip; /* and skip if not.*/ |
| 134 | |
| 135 | /* Compare against the page address. First, plant bits 13:12 |
| 136 | * into the tag, since those aren't part of the returned data. |
| 137 | */ |
| 138 | |
| 139 | R7 = DEPOSIT(R7, R3); /* set 13:12*/ |
| 140 | R1 = R7 & R4; /* Mask off lower bits*/ |
| 141 | CC = R1 == R0; /* Compare against page start.*/ |
| 142 | IF !CC JUMP ifskip; /* Skip it if it doesn't match.*/ |
| 143 | |
| 144 | /* Tag address matches against page, so this is an entry |
| 145 | * we must flush. |
| 146 | */ |
| 147 | |
| 148 | R7 >>= 10; /* Mask off the non-address bits*/ |
| 149 | R7 <<= 10; |
| 150 | P3 = R7; |
| 151 | IFLUSH [P3]; /* And flush the entry*/ |
| 152 | ifskip: |
| 153 | ife0: R5 += 1; /* Advance to next Set*/ |
| 154 | ife1: NOP; |
| 155 | |
| 156 | ifinished: |
| 157 | SSYNC; /* Ensure the data gets out to mem.*/ |
| 158 | |
| 159 | /*Finished. Restore context.*/ |
| 160 | LB1 = [SP++]; |
| 161 | LT1 = [SP++]; |
| 162 | LC1 = [SP++]; |
| 163 | LB0 = [SP++]; |
| 164 | LT0 = [SP++]; |
| 165 | LC0 = [SP++]; |
| 166 | ( R7:0, P5:0 ) = [SP++]; |
| 167 | RTS; |
| 168 | |
| 169 | iflush_whole_page: |
| 170 | /* It's a 1K or 4K page, so quicker to just flush the |
| 171 | * entire page. |
| 172 | */ |
| 173 | |
| 174 | P1 = 32; /* For 1K pages*/ |
| 175 | P2 = P1 << 2; /* For 4K pages*/ |
| 176 | P0 = R0; /* Start of page*/ |
| 177 | CC = BITTST(R1, 16); /* Whether 1K or 4K*/ |
| 178 | IF CC P1 = P2; |
| 179 | P1 += -1; /* Unroll one iteration*/ |
| 180 | SSYNC; |
| 181 | IFLUSH [P0++]; /* because CSYNC can't end loops.*/ |
| 182 | LSETUP (isall, ieall) LC0 = P1; |
| 183 | isall:IFLUSH [P0++]; |
| 184 | ieall: NOP; |
| 185 | SSYNC; |
| 186 | JUMP ifinished; |
| 187 | |
| 188 | /* This is an external function being called by the user |
| 189 | * application through __flush_cache_all. Currently this function |
| 190 | * serves the purpose of flushing all the pending writes in |
| 191 | * in the data cache. |
| 192 | */ |
| 193 | |
| 194 | ENTRY(_flush_data_cache) |
| 195 | [--SP] = ( R7:6, P5:4 ); |
| 196 | LINK 12; |
| 197 | SP += -12; |
| 198 | P5.H = (DCPLB_ADDR0 >> 16); |
| 199 | P5.L = (DCPLB_ADDR0 & 0xFFFF); |
| 200 | P4.H = (DCPLB_DATA0 >> 16); |
| 201 | P4.L = (DCPLB_DATA0 & 0xFFFF); |
| 202 | R7 = CPLB_VALID | CPLB_L1_CHBL | CPLB_DIRTY (Z); |
| 203 | R6 = 16; |
| 204 | next: R0 = [P5++]; |
| 205 | R1 = [P4++]; |
| 206 | CC = BITTST(R1, 14); /* Is it write-through?*/ |
| 207 | IF CC JUMP skip; /* If so, ignore it.*/ |
| 208 | R2 = R1 & R7; /* Is it a dirty, cached page?*/ |
| 209 | CC = R2; |
| 210 | IF !CC JUMP skip; /* If not, ignore it.*/ |
| 211 | [--SP] = RETS; |
| 212 | CALL _dcplb_flush; /* R0 = page, R1 = data*/ |
| 213 | RETS = [SP++]; |
| 214 | skip: R6 += -1; |
| 215 | CC = R6; |
| 216 | IF CC JUMP next; |
| 217 | SSYNC; |
| 218 | SP += 12; |
| 219 | UNLINK; |
| 220 | ( R7:6, P5:4 ) = [SP++]; |
| 221 | RTS; |
| 222 | |
| 223 | /* This is an internal function to flush all pending |
| 224 | * writes in the cache associated with a particular DCPLB. |
| 225 | * |
| 226 | * R0 - page's start address |
| 227 | * R1 - CPLB's data field. |
| 228 | */ |
| 229 | |
| 230 | .align 2 |
| 231 | ENTRY(_dcplb_flush) |
| 232 | [--SP] = ( R7:0, P5:0 ); |
| 233 | [--SP] = LC0; |
| 234 | [--SP] = LT0; |
| 235 | [--SP] = LB0; |
| 236 | [--SP] = LC1; |
| 237 | [--SP] = LT1; |
| 238 | [--SP] = LB1; |
| 239 | |
| 240 | /* If it's a 1K or 4K page, then it's quickest to |
| 241 | * just systematically flush all the addresses in |
| 242 | * the page, regardless of whether they're in the |
| 243 | * cache, or dirty. If it's a 1M or 4M page, there |
| 244 | * are too many addresses, and we have to search the |
| 245 | * cache for lines corresponding to the page. |
| 246 | */ |
| 247 | |
| 248 | CC = BITTST(R1, 17); /* 1MB or 4MB */ |
| 249 | IF !CC JUMP dflush_whole_page; |
| 250 | |
| 251 | /* We're only interested in the page's size, so extract |
| 252 | * this from the CPLB (bits 17:16), and scale to give an |
| 253 | * offset into the page_size and page_prefix tables. |
| 254 | */ |
| 255 | |
| 256 | R1 <<= 14; |
| 257 | R1 >>= 30; |
| 258 | R1 <<= 2; |
| 259 | |
| 260 | /* The page could be mapped into Bank A or Bank B, depending |
| 261 | * on (a) whether both banks are configured as cache, and |
| 262 | * (b) on whether address bit A[x] is set. x is determined |
| 263 | * by DCBS in DMEM_CONTROL |
| 264 | */ |
| 265 | |
| 266 | R2 = 0; /* Default to Bank A (Bank B would be 1)*/ |
| 267 | |
| 268 | P0.L = (DMEM_CONTROL & 0xFFFF); |
| 269 | P0.H = (DMEM_CONTROL >> 16); |
| 270 | |
| 271 | R3 = [P0]; /* If Bank B is not enabled as cache*/ |
| 272 | CC = BITTST(R3, 2); /* then Bank A is our only option.*/ |
| 273 | IF CC JUMP bank_chosen; |
| 274 | |
| 275 | R4 = 1<<14; /* If DCBS==0, use A[14].*/ |
| 276 | R5 = R4 << 7; /* If DCBS==1, use A[23];*/ |
| 277 | CC = BITTST(R3, 4); |
| 278 | IF CC R4 = R5; /* R4 now has either bit 14 or bit 23 set.*/ |
| 279 | R5 = R0 & R4; /* Use it to test the Page address*/ |
| 280 | CC = R5; /* and if that bit is set, we use Bank B,*/ |
| 281 | R2 = CC; /* else we use Bank A.*/ |
| 282 | R2 <<= 23; /* The Bank selection's at posn 23.*/ |
| 283 | |
| 284 | bank_chosen: |
| 285 | |
| 286 | /* We can also determine the sub-bank used, because this is |
| 287 | * taken from bits 13:12 of the address. |
| 288 | */ |
| 289 | |
| 290 | R3 = ((12<<8)|2); /* Extraction pattern */ |
| 291 | nop; /*Anamoly 05000209*/ |
| 292 | R4 = EXTRACT(R0, R3.L) (Z); /* Extract bits*/ |
| 293 | /* Save in extraction pattern for later deposit.*/ |
| 294 | R3.H = R4.L << 0; |
| 295 | |
| 296 | /* So: |
| 297 | * R0 = Page start |
| 298 | * R1 = Page length (actually, offset into size/prefix tables) |
| 299 | * R2 = Bank select mask |
| 300 | * R3 = sub-bank deposit values |
| 301 | * |
| 302 | * The cache has 2 Ways, and 64 sets, so we iterate through |
| 303 | * the sets, accessing the tag for each Way, for our Bank and |
| 304 | * sub-bank, looking for dirty, valid tags that match our |
| 305 | * address prefix. |
| 306 | */ |
| 307 | |
| 308 | P5.L = (DTEST_COMMAND & 0xFFFF); |
| 309 | P5.H = (DTEST_COMMAND >> 16); |
| 310 | P4.L = (DTEST_DATA0 & 0xFFFF); |
| 311 | P4.H = (DTEST_DATA0 >> 16); |
| 312 | |
| 313 | P0.L = page_prefix_table; |
| 314 | P0.H = page_prefix_table; |
| 315 | P1 = R1; |
| 316 | R5 = 0; /* Set counter*/ |
| 317 | P0 = P1 + P0; |
| 318 | R4 = [P0]; /* This is the address prefix*/ |
| 319 | |
| 320 | |
| 321 | /* We're reading (bit 1==0) the tag (bit 2==0), and we |
| 322 | * don't care about which double-word, since we're only |
| 323 | * fetching tags, so we only have to set Set, Bank, |
| 324 | * Sub-bank and Way. |
| 325 | */ |
| 326 | |
| 327 | P2 = 2; |
| 328 | LSETUP (fs1, fe1) LC1 = P2; |
| 329 | fs1: P0 = 64; /* iterate over all sets*/ |
| 330 | LSETUP (fs0, fe0) LC0 = P0; |
| 331 | fs0: R6 = R5 << 5; /* Combine set*/ |
| 332 | R6.H = R3.H << 0 ; /* and sub-bank*/ |
| 333 | R6 = R6 | R2; /* and Bank. Leave Way==0 at first.*/ |
| 334 | BITSET(R6,14); |
| 335 | [P5] = R6; /* Issue Command*/ |
| 336 | SSYNC; |
| 337 | R7 = [P4]; /* and read Tag.*/ |
| 338 | CC = BITTST(R7, 0); /* Check if valid*/ |
| 339 | IF !CC JUMP fskip; /* and skip if not.*/ |
| 340 | CC = BITTST(R7, 1); /* Check if dirty*/ |
| 341 | IF !CC JUMP fskip; /* and skip if not.*/ |
| 342 | |
| 343 | /* Compare against the page address. First, plant bits 13:12 |
| 344 | * into the tag, since those aren't part of the returned data. |
| 345 | */ |
| 346 | |
| 347 | R7 = DEPOSIT(R7, R3); /* set 13:12*/ |
| 348 | R1 = R7 & R4; /* Mask off lower bits*/ |
| 349 | CC = R1 == R0; /* Compare against page start.*/ |
| 350 | IF !CC JUMP fskip; /* Skip it if it doesn't match.*/ |
| 351 | |
| 352 | /* Tag address matches against page, so this is an entry |
| 353 | * we must flush. |
| 354 | */ |
| 355 | |
| 356 | R7 >>= 10; /* Mask off the non-address bits*/ |
| 357 | R7 <<= 10; |
| 358 | P3 = R7; |
| 359 | SSYNC; |
| 360 | FLUSHINV [P3]; /* And flush the entry*/ |
| 361 | fskip: |
| 362 | fe0: R5 += 1; /* Advance to next Set*/ |
| 363 | fe1: BITSET(R2, 26); /* Go to next Way.*/ |
| 364 | |
| 365 | dfinished: |
| 366 | SSYNC; /* Ensure the data gets out to mem.*/ |
| 367 | |
| 368 | /*Finished. Restore context.*/ |
| 369 | LB1 = [SP++]; |
| 370 | LT1 = [SP++]; |
| 371 | LC1 = [SP++]; |
| 372 | LB0 = [SP++]; |
| 373 | LT0 = [SP++]; |
| 374 | LC0 = [SP++]; |
| 375 | ( R7:0, P5:0 ) = [SP++]; |
| 376 | RTS; |
| 377 | |
| 378 | dflush_whole_page: |
| 379 | |
| 380 | /* It's a 1K or 4K page, so quicker to just flush the |
| 381 | * entire page. |
| 382 | */ |
| 383 | |
| 384 | P1 = 32; /* For 1K pages*/ |
| 385 | P2 = P1 << 2; /* For 4K pages*/ |
| 386 | P0 = R0; /* Start of page*/ |
| 387 | CC = BITTST(R1, 16); /* Whether 1K or 4K*/ |
| 388 | IF CC P1 = P2; |
| 389 | P1 += -1; /* Unroll one iteration*/ |
| 390 | SSYNC; |
| 391 | FLUSHINV [P0++]; /* because CSYNC can't end loops.*/ |
| 392 | LSETUP (eall, eall) LC0 = P1; |
| 393 | eall: FLUSHINV [P0++]; |
| 394 | SSYNC; |
| 395 | JUMP dfinished; |
| 396 | |
| 397 | .align 4; |
| 398 | page_prefix_table: |
| 399 | .byte4 0xFFFFFC00; /* 1K */ |
| 400 | .byte4 0xFFFFF000; /* 4K */ |
| 401 | .byte4 0xFFF00000; /* 1M */ |
| 402 | .byte4 0xFFC00000; /* 4M */ |
| 403 | .page_prefix_table.end: |