weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 1 | |
| 2 | /*--------------------------------------------------------------------*/ |
| 3 | /*--- Cache simulation. ---*/ |
| 4 | /*--- sim.c ---*/ |
| 5 | /*--------------------------------------------------------------------*/ |
| 6 | |
| 7 | /* |
njn | 9a0cba4 | 2007-04-15 22:15:57 +0000 | [diff] [blame] | 8 | This file is part of Callgrind, a Valgrind tool for call graph |
| 9 | profiling programs. |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 10 | |
njn | 9a0cba4 | 2007-04-15 22:15:57 +0000 | [diff] [blame] | 11 | Copyright (C) 2003-2005, Josef Weidendorfer (Josef.Weidendorfer@gmx.de) |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 12 | |
njn | 9a0cba4 | 2007-04-15 22:15:57 +0000 | [diff] [blame] | 13 | This tool is derived from and contains code from Cachegrind |
sewardj | 4d474d0 | 2008-02-11 11:34:59 +0000 | [diff] [blame^] | 14 | Copyright (C) 2002-2008 Nicholas Nethercote (njn@valgrind.org) |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 15 | |
| 16 | This program is free software; you can redistribute it and/or |
| 17 | modify it under the terms of the GNU General Public License as |
| 18 | published by the Free Software Foundation; either version 2 of the |
| 19 | License, or (at your option) any later version. |
| 20 | |
| 21 | This program is distributed in the hope that it will be useful, but |
| 22 | WITHOUT ANY WARRANTY; without even the implied warranty of |
| 23 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 24 | General Public License for more details. |
| 25 | |
| 26 | You should have received a copy of the GNU General Public License |
| 27 | along with this program; if not, write to the Free Software |
| 28 | Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA |
| 29 | 02111-1307, USA. |
| 30 | |
| 31 | The GNU General Public License is contained in the file COPYING. |
| 32 | */ |
| 33 | |
| 34 | #include "global.h" |
| 35 | |
| 36 | |
| 37 | /* Notes: |
| 38 | - simulates a write-allocate cache |
| 39 | - (block --> set) hash function uses simple bit selection |
| 40 | - handling of references straddling two cache blocks: |
| 41 | - counts as only one cache access (not two) |
| 42 | - both blocks hit --> one hit |
| 43 | - one block hits, the other misses --> one miss |
| 44 | - both blocks miss --> one miss (not two) |
| 45 | */ |
| 46 | |
| 47 | /* Cache configuration */ |
| 48 | #include "cg_arch.h" |
| 49 | |
| 50 | /* additional structures for cache use info, separated |
| 51 | * according usage frequency: |
| 52 | * - line_loaded : pointer to cost center of instruction |
| 53 | * which loaded the line into cache. |
| 54 | * Needed to increment counters when line is evicted. |
| 55 | * - line_use : updated on every access |
| 56 | */ |
| 57 | typedef struct { |
| 58 | UInt count; |
| 59 | UInt mask; /* e.g. for 64Byte line size 1bit/2Byte */ |
| 60 | } line_use; |
| 61 | |
| 62 | typedef struct { |
| 63 | Addr memline, iaddr; |
| 64 | line_use* dep_use; /* point to higher-level cacheblock for this memline */ |
| 65 | ULong* use_base; |
| 66 | } line_loaded; |
| 67 | |
| 68 | /* Cache state */ |
| 69 | typedef struct { |
| 70 | char* name; |
| 71 | int size; /* bytes */ |
| 72 | int assoc; |
| 73 | int line_size; /* bytes */ |
| 74 | Bool sectored; /* prefetch nearside cacheline on read */ |
| 75 | int sets; |
| 76 | int sets_min_1; |
| 77 | int assoc_bits; |
| 78 | int line_size_bits; |
| 79 | int tag_shift; |
| 80 | UWord tag_mask; |
| 81 | char desc_line[128]; |
| 82 | UWord* tags; |
| 83 | |
| 84 | /* for cache use */ |
| 85 | int line_size_mask; |
| 86 | int* line_start_mask; |
| 87 | int* line_end_mask; |
| 88 | line_loaded* loaded; |
| 89 | line_use* use; |
| 90 | } cache_t2; |
| 91 | |
| 92 | /* |
| 93 | * States of flat caches in our model. |
| 94 | * We use a 2-level hierarchy, |
| 95 | */ |
| 96 | static cache_t2 I1, D1, L2; |
| 97 | |
| 98 | /* Lower bits of cache tags are used as flags for a cache line */ |
| 99 | #define CACHELINE_FLAGMASK (MIN_LINE_SIZE-1) |
| 100 | #define CACHELINE_DIRTY 1 |
| 101 | |
| 102 | |
| 103 | /* Cache simulator Options */ |
| 104 | static Bool clo_simulate_writeback = False; |
| 105 | static Bool clo_simulate_hwpref = False; |
| 106 | static Bool clo_simulate_sectors = False; |
| 107 | static Bool clo_collect_cacheuse = False; |
| 108 | |
| 109 | /* Following global vars are setup before by |
| 110 | * setup_bbcc()/cachesim_after_bbsetup(): |
| 111 | * |
| 112 | * - Addr bb_base (instruction start address of original BB) |
| 113 | * - ULong* cost_base (start of cost array for BB) |
| 114 | * - BBCC* nonskipped (only != 0 when in a function not skipped) |
| 115 | */ |
| 116 | |
| 117 | /* Offset to events in event set, used in log_* functions */ |
| 118 | static Int off_D0_Ir; |
| 119 | static Int off_D1r_Ir; |
| 120 | static Int off_D1r_Dr; |
| 121 | static Int off_D1w_Ir; |
| 122 | static Int off_D1w_Dw; |
| 123 | static Int off_D2_Ir; |
| 124 | static Int off_D2_Dr; |
| 125 | static Int off_D2_Dw; |
| 126 | |
| 127 | static Addr bb_base; |
| 128 | static ULong* cost_base; |
| 129 | static InstrInfo* current_ii; |
| 130 | |
| 131 | /* Cache use offsets */ |
| 132 | /* FIXME: The offsets are only correct because all eventsets get |
| 133 | * the "Use" set added first ! |
| 134 | */ |
| 135 | static Int off_I1_AcCost = 0; |
| 136 | static Int off_I1_SpLoss = 1; |
| 137 | static Int off_D1_AcCost = 0; |
| 138 | static Int off_D1_SpLoss = 1; |
| 139 | static Int off_L2_AcCost = 2; |
| 140 | static Int off_L2_SpLoss = 3; |
| 141 | |
| 142 | /* Cache access types */ |
| 143 | typedef enum { Read = 0, Write = CACHELINE_DIRTY } RefType; |
| 144 | |
| 145 | /* Result of a reference into a flat cache */ |
| 146 | typedef enum { Hit = 0, Miss, MissDirty } CacheResult; |
| 147 | |
| 148 | /* Result of a reference into a hierarchical cache model */ |
| 149 | typedef enum { |
| 150 | L1_Hit, |
| 151 | L2_Hit, |
| 152 | MemAccess, |
| 153 | WriteBackMemAccess } CacheModelResult; |
| 154 | |
| 155 | typedef CacheModelResult (*simcall_type)(Addr, UChar); |
| 156 | |
| 157 | static struct { |
| 158 | simcall_type I1_Read; |
| 159 | simcall_type D1_Read; |
| 160 | simcall_type D1_Write; |
| 161 | } simulator; |
| 162 | |
| 163 | /*------------------------------------------------------------*/ |
| 164 | /*--- Cache Simulator Initialization ---*/ |
| 165 | /*------------------------------------------------------------*/ |
| 166 | |
| 167 | static void cachesim_clearcache(cache_t2* c) |
| 168 | { |
| 169 | Int i; |
| 170 | |
| 171 | for (i = 0; i < c->sets * c->assoc; i++) |
| 172 | c->tags[i] = 0; |
| 173 | if (c->use) { |
| 174 | for (i = 0; i < c->sets * c->assoc; i++) { |
| 175 | c->loaded[i].memline = 0; |
| 176 | c->loaded[i].use_base = 0; |
| 177 | c->loaded[i].dep_use = 0; |
| 178 | c->loaded[i].iaddr = 0; |
| 179 | c->use[i].mask = 0; |
| 180 | c->use[i].count = 0; |
| 181 | c->tags[i] = i % c->assoc; /* init lower bits as pointer */ |
| 182 | } |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | static void cacheuse_initcache(cache_t2* c); |
| 187 | |
| 188 | /* By this point, the size/assoc/line_size has been checked. */ |
| 189 | static void cachesim_initcache(cache_t config, cache_t2* c) |
| 190 | { |
| 191 | c->size = config.size; |
| 192 | c->assoc = config.assoc; |
| 193 | c->line_size = config.line_size; |
| 194 | c->sectored = False; // FIXME |
| 195 | |
| 196 | c->sets = (c->size / c->line_size) / c->assoc; |
| 197 | c->sets_min_1 = c->sets - 1; |
| 198 | c->assoc_bits = VG_(log2)(c->assoc); |
| 199 | c->line_size_bits = VG_(log2)(c->line_size); |
| 200 | c->tag_shift = c->line_size_bits + VG_(log2)(c->sets); |
| 201 | c->tag_mask = ~((1<<c->tag_shift)-1); |
| 202 | |
| 203 | /* Can bits in tag entries be used for flags? |
| 204 | * Should be always true as MIN_LINE_SIZE >= 16 */ |
| 205 | CLG_ASSERT( (c->tag_mask & CACHELINE_FLAGMASK) == 0); |
| 206 | |
| 207 | if (c->assoc == 1) { |
| 208 | VG_(sprintf)(c->desc_line, "%d B, %d B, direct-mapped%s", |
| 209 | c->size, c->line_size, |
| 210 | c->sectored ? ", sectored":""); |
| 211 | } else { |
| 212 | VG_(sprintf)(c->desc_line, "%d B, %d B, %d-way associative%s", |
| 213 | c->size, c->line_size, c->assoc, |
| 214 | c->sectored ? ", sectored":""); |
| 215 | } |
| 216 | |
| 217 | c->tags = (UWord*) CLG_MALLOC(sizeof(UWord) * c->sets * c->assoc); |
| 218 | if (clo_collect_cacheuse) |
| 219 | cacheuse_initcache(c); |
| 220 | else |
| 221 | c->use = 0; |
| 222 | cachesim_clearcache(c); |
| 223 | } |
| 224 | |
| 225 | |
| 226 | #if 0 |
| 227 | static void print_cache(cache_t2* c) |
| 228 | { |
| 229 | UInt set, way, i; |
| 230 | |
| 231 | /* Note initialisation and update of 'i'. */ |
| 232 | for (i = 0, set = 0; set < c->sets; set++) { |
| 233 | for (way = 0; way < c->assoc; way++, i++) { |
| 234 | VG_(printf)("%8x ", c->tags[i]); |
| 235 | } |
| 236 | VG_(printf)("\n"); |
| 237 | } |
| 238 | } |
| 239 | #endif |
| 240 | |
| 241 | |
| 242 | /*------------------------------------------------------------*/ |
| 243 | /*--- Write Through Cache Simulation ---*/ |
| 244 | /*------------------------------------------------------------*/ |
| 245 | |
| 246 | /* |
| 247 | * Simple model: L1 & L2 Write Through |
| 248 | * Does not distinguish among read and write references |
| 249 | * |
| 250 | * Simulator functions: |
| 251 | * CacheModelResult cachesim_I1_ref(Addr a, UChar size) |
| 252 | * CacheModelResult cachesim_D1_ref(Addr a, UChar size) |
| 253 | */ |
| 254 | |
| 255 | static __inline__ |
| 256 | CacheResult cachesim_setref(cache_t2* c, UInt set_no, UWord tag) |
| 257 | { |
| 258 | int i, j; |
| 259 | UWord *set; |
| 260 | |
| 261 | /* Shifting is a bit faster than multiplying */ |
| 262 | set = &(c->tags[set_no << c->assoc_bits]); |
| 263 | |
| 264 | /* This loop is unrolled for just the first case, which is the most */ |
| 265 | /* common. We can't unroll any further because it would screw up */ |
| 266 | /* if we have a direct-mapped (1-way) cache. */ |
| 267 | if (tag == set[0]) |
| 268 | return Hit; |
| 269 | |
| 270 | /* If the tag is one other than the MRU, move it into the MRU spot */ |
| 271 | /* and shuffle the rest down. */ |
| 272 | for (i = 1; i < c->assoc; i++) { |
| 273 | if (tag == set[i]) { |
| 274 | for (j = i; j > 0; j--) { |
| 275 | set[j] = set[j - 1]; |
| 276 | } |
| 277 | set[0] = tag; |
| 278 | return Hit; |
| 279 | } |
| 280 | } |
| 281 | |
| 282 | /* A miss; install this tag as MRU, shuffle rest down. */ |
| 283 | for (j = c->assoc - 1; j > 0; j--) { |
| 284 | set[j] = set[j - 1]; |
| 285 | } |
| 286 | set[0] = tag; |
| 287 | |
| 288 | return Miss; |
| 289 | } |
| 290 | |
| 291 | static CacheResult cachesim_ref(cache_t2* c, Addr a, UChar size) |
| 292 | { |
| 293 | UInt set1 = ( a >> c->line_size_bits) & (c->sets_min_1); |
| 294 | UInt set2 = ((a+size-1) >> c->line_size_bits) & (c->sets_min_1); |
| 295 | UWord tag = a >> c->tag_shift; |
| 296 | |
| 297 | /* Access entirely within line. */ |
| 298 | if (set1 == set2) |
| 299 | return cachesim_setref(c, set1, tag); |
| 300 | |
| 301 | /* Access straddles two lines. */ |
| 302 | /* Nb: this is a fast way of doing ((set1+1) % c->sets) */ |
| 303 | else if (((set1 + 1) & (c->sets-1)) == set2) { |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 304 | UWord tag2 = (a+size-1) >> c->tag_shift; |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 305 | |
| 306 | /* the call updates cache structures as side effect */ |
| 307 | CacheResult res1 = cachesim_setref(c, set1, tag); |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 308 | CacheResult res2 = cachesim_setref(c, set2, tag2); |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 309 | return ((res1 == Miss) || (res2 == Miss)) ? Miss : Hit; |
| 310 | |
| 311 | } else { |
njn | 8a7b41b | 2007-09-23 00:51:24 +0000 | [diff] [blame] | 312 | VG_(printf)("addr: %lx size: %u sets: %d %d", a, size, set1, set2); |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 313 | VG_(tool_panic)("item straddles more than two cache sets"); |
| 314 | } |
| 315 | return Hit; |
| 316 | } |
| 317 | |
| 318 | static |
| 319 | CacheModelResult cachesim_I1_ref(Addr a, UChar size) |
| 320 | { |
| 321 | if ( cachesim_ref( &I1, a, size) == Hit ) return L1_Hit; |
| 322 | if ( cachesim_ref( &L2, a, size) == Hit ) return L2_Hit; |
| 323 | return MemAccess; |
| 324 | } |
| 325 | |
| 326 | static |
| 327 | CacheModelResult cachesim_D1_ref(Addr a, UChar size) |
| 328 | { |
| 329 | if ( cachesim_ref( &D1, a, size) == Hit ) return L1_Hit; |
| 330 | if ( cachesim_ref( &L2, a, size) == Hit ) return L2_Hit; |
| 331 | return MemAccess; |
| 332 | } |
| 333 | |
| 334 | |
| 335 | /*------------------------------------------------------------*/ |
| 336 | /*--- Write Back Cache Simulation ---*/ |
| 337 | /*------------------------------------------------------------*/ |
| 338 | |
| 339 | /* |
| 340 | * More complex model: L1 Write-through, L2 Write-back |
| 341 | * This needs to distinguish among read and write references. |
| 342 | * |
| 343 | * Simulator functions: |
| 344 | * CacheModelResult cachesim_I1_Read(Addr a, UChar size) |
| 345 | * CacheModelResult cachesim_D1_Read(Addr a, UChar size) |
| 346 | * CacheModelResult cachesim_D1_Write(Addr a, UChar size) |
| 347 | */ |
| 348 | |
| 349 | /* |
| 350 | * With write-back, result can be a miss evicting a dirty line |
| 351 | * The dirty state of a cache line is stored in Bit0 of the tag for |
| 352 | * this cache line (CACHELINE_DIRTY = 1). By OR'ing the reference |
| 353 | * type (Read/Write), the line gets dirty on a write. |
| 354 | */ |
| 355 | static __inline__ |
| 356 | CacheResult cachesim_setref_wb(cache_t2* c, RefType ref, UInt set_no, UWord tag) |
| 357 | { |
| 358 | int i, j; |
| 359 | UWord *set, tmp_tag; |
| 360 | |
| 361 | /* Shifting is a bit faster than multiplying */ |
| 362 | set = &(c->tags[set_no << c->assoc_bits]); |
| 363 | |
| 364 | /* This loop is unrolled for just the first case, which is the most */ |
| 365 | /* common. We can't unroll any further because it would screw up */ |
| 366 | /* if we have a direct-mapped (1-way) cache. */ |
| 367 | if (tag == (set[0] & ~CACHELINE_DIRTY)) { |
| 368 | set[0] |= ref; |
| 369 | return Hit; |
| 370 | } |
| 371 | /* If the tag is one other than the MRU, move it into the MRU spot */ |
| 372 | /* and shuffle the rest down. */ |
| 373 | for (i = 1; i < c->assoc; i++) { |
| 374 | if (tag == (set[i] & ~CACHELINE_DIRTY)) { |
| 375 | tmp_tag = set[i] | ref; // update dirty flag |
| 376 | for (j = i; j > 0; j--) { |
| 377 | set[j] = set[j - 1]; |
| 378 | } |
| 379 | set[0] = tmp_tag; |
| 380 | return Hit; |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | /* A miss; install this tag as MRU, shuffle rest down. */ |
| 385 | tmp_tag = set[c->assoc - 1]; |
| 386 | for (j = c->assoc - 1; j > 0; j--) { |
| 387 | set[j] = set[j - 1]; |
| 388 | } |
| 389 | set[0] = tag | ref; |
| 390 | |
| 391 | return (tmp_tag & CACHELINE_DIRTY) ? MissDirty : Miss; |
| 392 | } |
| 393 | |
| 394 | |
| 395 | static __inline__ |
| 396 | CacheResult cachesim_ref_wb(cache_t2* c, RefType ref, Addr a, UChar size) |
| 397 | { |
| 398 | UInt set1 = ( a >> c->line_size_bits) & (c->sets_min_1); |
| 399 | UInt set2 = ((a+size-1) >> c->line_size_bits) & (c->sets_min_1); |
| 400 | UWord tag = a & c->tag_mask; |
| 401 | |
| 402 | /* Access entirely within line. */ |
| 403 | if (set1 == set2) |
| 404 | return cachesim_setref_wb(c, ref, set1, tag); |
| 405 | |
| 406 | /* Access straddles two lines. */ |
| 407 | /* Nb: this is a fast way of doing ((set1+1) % c->sets) */ |
| 408 | else if (((set1 + 1) & (c->sets-1)) == set2) { |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 409 | UWord tag2 = (a+size-1) >> c->tag_shift; |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 410 | |
| 411 | /* the call updates cache structures as side effect */ |
| 412 | CacheResult res1 = cachesim_setref_wb(c, ref, set1, tag); |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 413 | CacheResult res2 = cachesim_setref_wb(c, ref, set2, tag2); |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 414 | |
| 415 | if ((res1 == MissDirty) || (res2 == MissDirty)) return MissDirty; |
| 416 | return ((res1 == Miss) || (res2 == Miss)) ? Miss : Hit; |
| 417 | |
| 418 | } else { |
njn | 8a7b41b | 2007-09-23 00:51:24 +0000 | [diff] [blame] | 419 | VG_(printf)("addr: %lx size: %u sets: %d %d", a, size, set1, set2); |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 420 | VG_(tool_panic)("item straddles more than two cache sets"); |
| 421 | } |
| 422 | return Hit; |
| 423 | } |
| 424 | |
| 425 | |
| 426 | static |
| 427 | CacheModelResult cachesim_I1_Read(Addr a, UChar size) |
| 428 | { |
| 429 | if ( cachesim_ref( &I1, a, size) == Hit ) return L1_Hit; |
| 430 | switch( cachesim_ref_wb( &L2, Read, a, size) ) { |
| 431 | case Hit: return L2_Hit; |
| 432 | case Miss: return MemAccess; |
| 433 | default: break; |
| 434 | } |
| 435 | return WriteBackMemAccess; |
| 436 | } |
| 437 | |
| 438 | static |
| 439 | CacheModelResult cachesim_D1_Read(Addr a, UChar size) |
| 440 | { |
| 441 | if ( cachesim_ref( &D1, a, size) == Hit ) return L1_Hit; |
| 442 | switch( cachesim_ref_wb( &L2, Read, a, size) ) { |
| 443 | case Hit: return L2_Hit; |
| 444 | case Miss: return MemAccess; |
| 445 | default: break; |
| 446 | } |
| 447 | return WriteBackMemAccess; |
| 448 | } |
| 449 | |
| 450 | static |
| 451 | CacheModelResult cachesim_D1_Write(Addr a, UChar size) |
| 452 | { |
| 453 | if ( cachesim_ref( &D1, a, size) == Hit ) { |
| 454 | /* Even for a L1 hit, the write-trough L1 passes |
| 455 | * the write to the L2 to make the L2 line dirty. |
| 456 | * But this causes no latency, so return the hit. |
| 457 | */ |
| 458 | cachesim_ref_wb( &L2, Write, a, size); |
| 459 | return L1_Hit; |
| 460 | } |
| 461 | switch( cachesim_ref_wb( &L2, Write, a, size) ) { |
| 462 | case Hit: return L2_Hit; |
| 463 | case Miss: return MemAccess; |
| 464 | default: break; |
| 465 | } |
| 466 | return WriteBackMemAccess; |
| 467 | } |
| 468 | |
| 469 | |
| 470 | /*------------------------------------------------------------*/ |
| 471 | /*--- Hardware Prefetch Simulation ---*/ |
| 472 | /*------------------------------------------------------------*/ |
| 473 | |
| 474 | static ULong prefetch_up = 0; |
| 475 | static ULong prefetch_down = 0; |
| 476 | |
| 477 | #define PF_STREAMS 8 |
| 478 | #define PF_PAGEBITS 12 |
| 479 | |
| 480 | static UInt pf_lastblock[PF_STREAMS]; |
| 481 | static Int pf_seqblocks[PF_STREAMS]; |
| 482 | |
| 483 | static |
| 484 | void prefetch_clear(void) |
| 485 | { |
| 486 | int i; |
| 487 | for(i=0;i<PF_STREAMS;i++) |
| 488 | pf_lastblock[i] = pf_seqblocks[i] = 0; |
| 489 | } |
| 490 | |
| 491 | /* |
| 492 | * HW Prefetch emulation |
| 493 | * Start prefetching when detecting sequential access to 3 memory blocks. |
| 494 | * One stream can be detected per 4k page. |
| 495 | */ |
| 496 | static __inline__ |
| 497 | void prefetch_L2_doref(Addr a, UChar size) |
| 498 | { |
| 499 | UInt stream = (a >> PF_PAGEBITS) % PF_STREAMS; |
| 500 | UInt block = ( a >> L2.line_size_bits); |
| 501 | |
| 502 | if (block != pf_lastblock[stream]) { |
| 503 | if (pf_seqblocks[stream] == 0) { |
| 504 | if (pf_lastblock[stream] +1 == block) pf_seqblocks[stream]++; |
| 505 | else if (pf_lastblock[stream] -1 == block) pf_seqblocks[stream]--; |
| 506 | } |
| 507 | else if (pf_seqblocks[stream] >0) { |
| 508 | if (pf_lastblock[stream] +1 == block) { |
| 509 | pf_seqblocks[stream]++; |
| 510 | if (pf_seqblocks[stream] >= 2) { |
| 511 | prefetch_up++; |
| 512 | cachesim_ref(&L2, a + 5 * L2.line_size,1); |
| 513 | } |
| 514 | } |
| 515 | else pf_seqblocks[stream] = 0; |
| 516 | } |
| 517 | else if (pf_seqblocks[stream] <0) { |
| 518 | if (pf_lastblock[stream] -1 == block) { |
| 519 | pf_seqblocks[stream]--; |
| 520 | if (pf_seqblocks[stream] <= -2) { |
| 521 | prefetch_down++; |
| 522 | cachesim_ref(&L2, a - 5 * L2.line_size,1); |
| 523 | } |
| 524 | } |
| 525 | else pf_seqblocks[stream] = 0; |
| 526 | } |
| 527 | pf_lastblock[stream] = block; |
| 528 | } |
| 529 | } |
| 530 | |
| 531 | /* simple model with hardware prefetch */ |
| 532 | |
| 533 | static |
| 534 | CacheModelResult prefetch_I1_ref(Addr a, UChar size) |
| 535 | { |
| 536 | if ( cachesim_ref( &I1, a, size) == Hit ) return L1_Hit; |
| 537 | prefetch_L2_doref(a,size); |
| 538 | if ( cachesim_ref( &L2, a, size) == Hit ) return L2_Hit; |
| 539 | return MemAccess; |
| 540 | } |
| 541 | |
| 542 | static |
| 543 | CacheModelResult prefetch_D1_ref(Addr a, UChar size) |
| 544 | { |
| 545 | if ( cachesim_ref( &D1, a, size) == Hit ) return L1_Hit; |
| 546 | prefetch_L2_doref(a,size); |
| 547 | if ( cachesim_ref( &L2, a, size) == Hit ) return L2_Hit; |
| 548 | return MemAccess; |
| 549 | } |
| 550 | |
| 551 | |
| 552 | /* complex model with hardware prefetch */ |
| 553 | |
| 554 | static |
| 555 | CacheModelResult prefetch_I1_Read(Addr a, UChar size) |
| 556 | { |
| 557 | if ( cachesim_ref( &I1, a, size) == Hit ) return L1_Hit; |
| 558 | prefetch_L2_doref(a,size); |
| 559 | switch( cachesim_ref_wb( &L2, Read, a, size) ) { |
| 560 | case Hit: return L2_Hit; |
| 561 | case Miss: return MemAccess; |
| 562 | default: break; |
| 563 | } |
| 564 | return WriteBackMemAccess; |
| 565 | } |
| 566 | |
| 567 | static |
| 568 | CacheModelResult prefetch_D1_Read(Addr a, UChar size) |
| 569 | { |
| 570 | if ( cachesim_ref( &D1, a, size) == Hit ) return L1_Hit; |
| 571 | prefetch_L2_doref(a,size); |
| 572 | switch( cachesim_ref_wb( &L2, Read, a, size) ) { |
| 573 | case Hit: return L2_Hit; |
| 574 | case Miss: return MemAccess; |
| 575 | default: break; |
| 576 | } |
| 577 | return WriteBackMemAccess; |
| 578 | } |
| 579 | |
| 580 | static |
| 581 | CacheModelResult prefetch_D1_Write(Addr a, UChar size) |
| 582 | { |
| 583 | prefetch_L2_doref(a,size); |
| 584 | if ( cachesim_ref( &D1, a, size) == Hit ) { |
| 585 | /* Even for a L1 hit, the write-trough L1 passes |
| 586 | * the write to the L2 to make the L2 line dirty. |
| 587 | * But this causes no latency, so return the hit. |
| 588 | */ |
| 589 | cachesim_ref_wb( &L2, Write, a, size); |
| 590 | return L1_Hit; |
| 591 | } |
| 592 | switch( cachesim_ref_wb( &L2, Write, a, size) ) { |
| 593 | case Hit: return L2_Hit; |
| 594 | case Miss: return MemAccess; |
| 595 | default: break; |
| 596 | } |
| 597 | return WriteBackMemAccess; |
| 598 | } |
| 599 | |
| 600 | |
| 601 | /*------------------------------------------------------------*/ |
| 602 | /*--- Cache Simulation with use metric collection ---*/ |
| 603 | /*------------------------------------------------------------*/ |
| 604 | |
| 605 | /* can not be combined with write-back or prefetch */ |
| 606 | |
| 607 | static |
| 608 | void cacheuse_initcache(cache_t2* c) |
| 609 | { |
| 610 | int i; |
| 611 | unsigned int start_mask, start_val; |
| 612 | unsigned int end_mask, end_val; |
| 613 | |
| 614 | c->use = CLG_MALLOC(sizeof(line_use) * c->sets * c->assoc); |
| 615 | c->loaded = CLG_MALLOC(sizeof(line_loaded) * c->sets * c->assoc); |
| 616 | c->line_start_mask = CLG_MALLOC(sizeof(int) * c->line_size); |
| 617 | c->line_end_mask = CLG_MALLOC(sizeof(int) * c->line_size); |
| 618 | |
| 619 | |
| 620 | c->line_size_mask = c->line_size-1; |
| 621 | |
| 622 | /* Meaning of line_start_mask/line_end_mask |
| 623 | * Example: for a given cache line, you get an access starting at |
| 624 | * byte offset 5, length 4, byte 5 - 8 was touched. For a cache |
| 625 | * line size of 32, you have 1 bit per byte in the mask: |
| 626 | * |
| 627 | * bit31 bit8 bit5 bit 0 |
| 628 | * | | | | |
| 629 | * 11..111111100000 line_start_mask[5] |
| 630 | * 00..000111111111 line_end_mask[(5+4)-1] |
| 631 | * |
| 632 | * use_mask |= line_start_mask[5] && line_end_mask[8] |
| 633 | * |
| 634 | */ |
| 635 | start_val = end_val = ~0; |
| 636 | if (c->line_size < 32) { |
| 637 | int bits_per_byte = 32/c->line_size; |
| 638 | start_mask = (1<<bits_per_byte)-1; |
| 639 | end_mask = start_mask << (32-bits_per_byte); |
| 640 | for(i=0;i<c->line_size;i++) { |
| 641 | c->line_start_mask[i] = start_val; |
| 642 | start_val = start_val & ~start_mask; |
| 643 | start_mask = start_mask << bits_per_byte; |
| 644 | |
| 645 | c->line_end_mask[c->line_size-i-1] = end_val; |
| 646 | end_val = end_val & ~end_mask; |
| 647 | end_mask = end_mask >> bits_per_byte; |
| 648 | } |
| 649 | } |
| 650 | else { |
| 651 | int bytes_per_bit = c->line_size/32; |
| 652 | start_mask = 1; |
| 653 | end_mask = 1 << 31; |
| 654 | for(i=0;i<c->line_size;i++) { |
| 655 | c->line_start_mask[i] = start_val; |
| 656 | c->line_end_mask[c->line_size-i-1] = end_val; |
| 657 | if ( ((i+1)%bytes_per_bit) == 0) { |
| 658 | start_val &= ~start_mask; |
| 659 | end_val &= ~end_mask; |
| 660 | start_mask <<= 1; |
| 661 | end_mask >>= 1; |
| 662 | } |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | CLG_DEBUG(6, "Config %s:\n", c->desc_line); |
| 667 | for(i=0;i<c->line_size;i++) { |
| 668 | CLG_DEBUG(6, " [%2d]: start mask %8x, end mask %8x\n", |
| 669 | i, c->line_start_mask[i], c->line_end_mask[i]); |
| 670 | } |
| 671 | |
| 672 | /* We use lower tag bits as offset pointers to cache use info. |
| 673 | * I.e. some cache parameters don't work. |
| 674 | */ |
| 675 | if (c->tag_shift < c->assoc_bits) { |
| 676 | VG_(message)(Vg_DebugMsg, |
| 677 | "error: Use associativity < %d for cache use statistics!", |
| 678 | (1<<c->tag_shift) ); |
| 679 | VG_(tool_panic)("Unsupported cache configuration"); |
| 680 | } |
| 681 | } |
| 682 | |
| 683 | /* FIXME: A little tricky */ |
| 684 | #if 0 |
| 685 | |
| 686 | static __inline__ |
| 687 | void cacheuse_update_hit(cache_t2* c, UInt high_idx, UInt low_idx, UInt use_mask) |
| 688 | { |
| 689 | int idx = (high_idx << c->assoc_bits) | low_idx; |
| 690 | |
| 691 | c->use[idx].count ++; |
| 692 | c->use[idx].mask |= use_mask; |
| 693 | |
| 694 | CLG_DEBUG(6," Hit [idx %d] (line %p from %p): %x => %08x, count %d\n", |
| 695 | idx, c->loaded[idx].memline, c->loaded[idx].iaddr, |
| 696 | use_mask, c->use[idx].mask, c->use[idx].count); |
| 697 | } |
| 698 | |
| 699 | /* only used for I1, D1 */ |
| 700 | |
| 701 | static __inline__ |
| 702 | CacheResult cacheuse_setref(cache_t2* c, UInt set_no, UWord tag) |
| 703 | { |
| 704 | int i, j, idx; |
| 705 | UWord *set, tmp_tag; |
| 706 | UInt use_mask; |
| 707 | |
| 708 | /* Shifting is a bit faster than multiplying */ |
| 709 | set = &(c->tags[set_no << c->assoc_bits]); |
| 710 | use_mask = |
| 711 | c->line_start_mask[a & c->line_size_mask] & |
| 712 | c->line_end_mask[(a+size-1) & c->line_size_mask]; |
| 713 | |
| 714 | /* This loop is unrolled for just the first case, which is the most */ |
| 715 | /* common. We can't unroll any further because it would screw up */ |
| 716 | /* if we have a direct-mapped (1-way) cache. */ |
| 717 | if (tag == (set[0] & c->tag_mask)) { |
| 718 | cacheuse_update(c, set_no, set[0] & ~c->tag_mask, use_mask); |
| 719 | return L1_Hit; |
| 720 | } |
| 721 | |
| 722 | /* If the tag is one other than the MRU, move it into the MRU spot */ |
| 723 | /* and shuffle the rest down. */ |
| 724 | for (i = 1; i < c->assoc; i++) { |
| 725 | if (tag == (set[i] & c->tag_mask)) { |
| 726 | tmp_tag = set[i]; |
| 727 | for (j = i; j > 0; j--) { |
| 728 | set[j] = set[j - 1]; |
| 729 | } |
| 730 | set[0] = tmp_tag; |
| 731 | |
| 732 | cacheuse_update(c, set_no, tmp_tag & ~c->tag_mask, use_mask); |
| 733 | return L1_Hit; |
| 734 | } |
| 735 | } |
| 736 | |
| 737 | /* A miss; install this tag as MRU, shuffle rest down. */ |
| 738 | tmp_tag = set[L.assoc - 1] & ~c->tag_mask; |
| 739 | for (j = c->assoc - 1; j > 0; j--) { |
| 740 | set[j] = set[j - 1]; |
| 741 | } |
| 742 | set[0] = tag | tmp_tag; |
| 743 | |
| 744 | cacheuse_L2_miss(c, (set_no << c->assoc_bits) | tmp_tag, |
| 745 | use_mask, a & ~c->line_size_mask); |
| 746 | |
| 747 | return Miss; |
| 748 | } |
| 749 | |
| 750 | |
| 751 | static CacheResult cacheuse_ref(cache_t2* c, Addr a, UChar size) |
| 752 | { |
| 753 | UInt set1 = ( a >> c->line_size_bits) & (c->sets_min_1); |
| 754 | UInt set2 = ((a+size-1) >> c->line_size_bits) & (c->sets_min_1); |
| 755 | UWord tag = a >> c->tag_shift; |
| 756 | |
| 757 | /* Access entirely within line. */ |
| 758 | if (set1 == set2) |
| 759 | return cacheuse_setref(c, set1, tag); |
| 760 | |
| 761 | /* Access straddles two lines. */ |
| 762 | /* Nb: this is a fast way of doing ((set1+1) % c->sets) */ |
| 763 | else if (((set1 + 1) & (c->sets-1)) == set2) { |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 764 | UWord tag2 = a >> c->tag_shift; |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 765 | |
| 766 | /* the call updates cache structures as side effect */ |
| 767 | CacheResult res1 = cacheuse_isMiss(c, set1, tag); |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 768 | CacheResult res2 = cacheuse_isMiss(c, set2, tag2); |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 769 | return ((res1 == Miss) || (res2 == Miss)) ? Miss : Hit; |
| 770 | |
| 771 | } else { |
| 772 | VG_(printf)("addr: %x size: %u sets: %d %d", a, size, set1, set2); |
| 773 | VG_(tool_panic)("item straddles more than two cache sets"); |
| 774 | } |
| 775 | return Hit; |
| 776 | } |
| 777 | #endif |
| 778 | |
| 779 | |
| 780 | /* for I1/D1 caches */ |
| 781 | #define CACHEUSE(L) \ |
| 782 | \ |
| 783 | static CacheModelResult cacheuse##_##L##_doRead(Addr a, UChar size) \ |
| 784 | { \ |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 785 | UInt set1 = ( a >> L.line_size_bits) & (L.sets_min_1); \ |
| 786 | UInt set2 = ((a+size-1) >> L.line_size_bits) & (L.sets_min_1); \ |
| 787 | UWord tag = a & L.tag_mask; \ |
| 788 | UWord tag2; \ |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 789 | int i, j, idx; \ |
| 790 | UWord *set, tmp_tag; \ |
| 791 | UInt use_mask; \ |
| 792 | \ |
| 793 | CLG_DEBUG(6,"%s.Acc(Addr %p, size %d): Sets [%d/%d]\n", \ |
| 794 | L.name, a, size, set1, set2); \ |
| 795 | \ |
| 796 | /* First case: word entirely within line. */ \ |
| 797 | if (set1 == set2) { \ |
| 798 | \ |
| 799 | /* Shifting is a bit faster than multiplying */ \ |
| 800 | set = &(L.tags[set1 << L.assoc_bits]); \ |
| 801 | use_mask = L.line_start_mask[a & L.line_size_mask] & \ |
| 802 | L.line_end_mask[(a+size-1) & L.line_size_mask]; \ |
| 803 | \ |
| 804 | /* This loop is unrolled for just the first case, which is the most */\ |
| 805 | /* common. We can't unroll any further because it would screw up */\ |
| 806 | /* if we have a direct-mapped (1-way) cache. */\ |
| 807 | if (tag == (set[0] & L.tag_mask)) { \ |
| 808 | idx = (set1 << L.assoc_bits) | (set[0] & ~L.tag_mask); \ |
| 809 | L.use[idx].count ++; \ |
| 810 | L.use[idx].mask |= use_mask; \ |
| 811 | CLG_DEBUG(6," Hit0 [idx %d] (line %p from %p): %x => %08x, count %d\n",\ |
| 812 | idx, L.loaded[idx].memline, L.loaded[idx].iaddr, \ |
| 813 | use_mask, L.use[idx].mask, L.use[idx].count); \ |
| 814 | return L1_Hit; \ |
| 815 | } \ |
| 816 | /* If the tag is one other than the MRU, move it into the MRU spot */\ |
| 817 | /* and shuffle the rest down. */\ |
| 818 | for (i = 1; i < L.assoc; i++) { \ |
| 819 | if (tag == (set[i] & L.tag_mask)) { \ |
| 820 | tmp_tag = set[i]; \ |
| 821 | for (j = i; j > 0; j--) { \ |
| 822 | set[j] = set[j - 1]; \ |
| 823 | } \ |
| 824 | set[0] = tmp_tag; \ |
| 825 | idx = (set1 << L.assoc_bits) | (tmp_tag & ~L.tag_mask); \ |
| 826 | L.use[idx].count ++; \ |
| 827 | L.use[idx].mask |= use_mask; \ |
| 828 | CLG_DEBUG(6," Hit%d [idx %d] (line %p from %p): %x => %08x, count %d\n",\ |
| 829 | i, idx, L.loaded[idx].memline, L.loaded[idx].iaddr, \ |
| 830 | use_mask, L.use[idx].mask, L.use[idx].count); \ |
| 831 | return L1_Hit; \ |
| 832 | } \ |
| 833 | } \ |
| 834 | \ |
| 835 | /* A miss; install this tag as MRU, shuffle rest down. */ \ |
| 836 | tmp_tag = set[L.assoc - 1] & ~L.tag_mask; \ |
| 837 | for (j = L.assoc - 1; j > 0; j--) { \ |
| 838 | set[j] = set[j - 1]; \ |
| 839 | } \ |
| 840 | set[0] = tag | tmp_tag; \ |
| 841 | idx = (set1 << L.assoc_bits) | tmp_tag; \ |
| 842 | return update_##L##_use(&L, idx, \ |
| 843 | use_mask, a &~ L.line_size_mask); \ |
| 844 | \ |
| 845 | /* Second case: word straddles two lines. */ \ |
| 846 | /* Nb: this is a fast way of doing ((set1+1) % L.sets) */ \ |
| 847 | } else if (((set1 + 1) & (L.sets-1)) == set2) { \ |
| 848 | Int miss1=0, miss2=0; /* 0: L1 hit, 1:L1 miss, 2:L2 miss */ \ |
| 849 | set = &(L.tags[set1 << L.assoc_bits]); \ |
| 850 | use_mask = L.line_start_mask[a & L.line_size_mask]; \ |
| 851 | if (tag == (set[0] & L.tag_mask)) { \ |
| 852 | idx = (set1 << L.assoc_bits) | (set[0] & ~L.tag_mask); \ |
| 853 | L.use[idx].count ++; \ |
| 854 | L.use[idx].mask |= use_mask; \ |
| 855 | CLG_DEBUG(6," Hit0 [idx %d] (line %p from %p): %x => %08x, count %d\n",\ |
| 856 | idx, L.loaded[idx].memline, L.loaded[idx].iaddr, \ |
| 857 | use_mask, L.use[idx].mask, L.use[idx].count); \ |
| 858 | goto block2; \ |
| 859 | } \ |
| 860 | for (i = 1; i < L.assoc; i++) { \ |
| 861 | if (tag == (set[i] & L.tag_mask)) { \ |
| 862 | tmp_tag = set[i]; \ |
| 863 | for (j = i; j > 0; j--) { \ |
| 864 | set[j] = set[j - 1]; \ |
| 865 | } \ |
| 866 | set[0] = tmp_tag; \ |
| 867 | idx = (set1 << L.assoc_bits) | (tmp_tag & ~L.tag_mask); \ |
| 868 | L.use[idx].count ++; \ |
| 869 | L.use[idx].mask |= use_mask; \ |
| 870 | CLG_DEBUG(6," Hit%d [idx %d] (line %p from %p): %x => %08x, count %d\n",\ |
| 871 | i, idx, L.loaded[idx].memline, L.loaded[idx].iaddr, \ |
| 872 | use_mask, L.use[idx].mask, L.use[idx].count); \ |
| 873 | goto block2; \ |
| 874 | } \ |
| 875 | } \ |
| 876 | tmp_tag = set[L.assoc - 1] & ~L.tag_mask; \ |
| 877 | for (j = L.assoc - 1; j > 0; j--) { \ |
| 878 | set[j] = set[j - 1]; \ |
| 879 | } \ |
| 880 | set[0] = tag | tmp_tag; \ |
| 881 | idx = (set1 << L.assoc_bits) | tmp_tag; \ |
| 882 | miss1 = update_##L##_use(&L, idx, \ |
| 883 | use_mask, a &~ L.line_size_mask); \ |
| 884 | block2: \ |
| 885 | set = &(L.tags[set2 << L.assoc_bits]); \ |
| 886 | use_mask = L.line_end_mask[(a+size-1) & L.line_size_mask]; \ |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 887 | tag2 = (a+size-1) & L.tag_mask; \ |
| 888 | if (tag2 == (set[0] & L.tag_mask)) { \ |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 889 | idx = (set2 << L.assoc_bits) | (set[0] & ~L.tag_mask); \ |
| 890 | L.use[idx].count ++; \ |
| 891 | L.use[idx].mask |= use_mask; \ |
| 892 | CLG_DEBUG(6," Hit0 [idx %d] (line %p from %p): %x => %08x, count %d\n",\ |
| 893 | idx, L.loaded[idx].memline, L.loaded[idx].iaddr, \ |
| 894 | use_mask, L.use[idx].mask, L.use[idx].count); \ |
| 895 | return miss1; \ |
| 896 | } \ |
| 897 | for (i = 1; i < L.assoc; i++) { \ |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 898 | if (tag2 == (set[i] & L.tag_mask)) { \ |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 899 | tmp_tag = set[i]; \ |
| 900 | for (j = i; j > 0; j--) { \ |
| 901 | set[j] = set[j - 1]; \ |
| 902 | } \ |
| 903 | set[0] = tmp_tag; \ |
| 904 | idx = (set2 << L.assoc_bits) | (tmp_tag & ~L.tag_mask); \ |
| 905 | L.use[idx].count ++; \ |
| 906 | L.use[idx].mask |= use_mask; \ |
| 907 | CLG_DEBUG(6," Hit%d [idx %d] (line %p from %p): %x => %08x, count %d\n",\ |
| 908 | i, idx, L.loaded[idx].memline, L.loaded[idx].iaddr, \ |
| 909 | use_mask, L.use[idx].mask, L.use[idx].count); \ |
| 910 | return miss1; \ |
| 911 | } \ |
| 912 | } \ |
| 913 | tmp_tag = set[L.assoc - 1] & ~L.tag_mask; \ |
| 914 | for (j = L.assoc - 1; j > 0; j--) { \ |
| 915 | set[j] = set[j - 1]; \ |
| 916 | } \ |
weidendo | 28e2a14 | 2006-11-22 21:00:53 +0000 | [diff] [blame] | 917 | set[0] = tag2 | tmp_tag; \ |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 918 | idx = (set2 << L.assoc_bits) | tmp_tag; \ |
| 919 | miss2 = update_##L##_use(&L, idx, \ |
| 920 | use_mask, (a+size-1) &~ L.line_size_mask); \ |
| 921 | return (miss1==MemAccess || miss2==MemAccess) ? MemAccess:L2_Hit; \ |
| 922 | \ |
| 923 | } else { \ |
| 924 | VG_(printf)("addr: %p size: %u sets: %d %d", a, size, set1, set2); \ |
| 925 | VG_(tool_panic)("item straddles more than two cache sets"); \ |
| 926 | } \ |
| 927 | return 0; \ |
| 928 | } |
| 929 | |
| 930 | |
| 931 | /* logarithmic bitcounting algorithm, see |
| 932 | * http://graphics.stanford.edu/~seander/bithacks.html |
| 933 | */ |
| 934 | static __inline__ unsigned int countBits(unsigned int bits) |
| 935 | { |
| 936 | unsigned int c; // store the total here |
| 937 | const int S[] = {1, 2, 4, 8, 16}; // Magic Binary Numbers |
| 938 | const int B[] = {0x55555555, 0x33333333, 0x0F0F0F0F, 0x00FF00FF, 0x0000FFFF}; |
| 939 | |
| 940 | c = bits; |
| 941 | c = ((c >> S[0]) & B[0]) + (c & B[0]); |
| 942 | c = ((c >> S[1]) & B[1]) + (c & B[1]); |
| 943 | c = ((c >> S[2]) & B[2]) + (c & B[2]); |
| 944 | c = ((c >> S[3]) & B[3]) + (c & B[3]); |
| 945 | c = ((c >> S[4]) & B[4]) + (c & B[4]); |
| 946 | return c; |
| 947 | } |
| 948 | |
| 949 | static void update_L2_use(int idx, Addr memline) |
| 950 | { |
| 951 | line_loaded* loaded = &(L2.loaded[idx]); |
| 952 | line_use* use = &(L2.use[idx]); |
| 953 | int i = ((32 - countBits(use->mask)) * L2.line_size)>>5; |
| 954 | |
| 955 | CLG_DEBUG(2, " L2.miss [%d]: at %p accessing memline %p\n", |
| 956 | idx, bb_base + current_ii->instr_offset, memline); |
| 957 | if (use->count>0) { |
| 958 | CLG_DEBUG(2, " old: used %d, loss bits %d (%08x) [line %p from %p]\n", |
| 959 | use->count, i, use->mask, loaded->memline, loaded->iaddr); |
| 960 | CLG_DEBUG(2, " collect: %d, use_base %p\n", |
| 961 | CLG_(current_state).collect, loaded->use_base); |
| 962 | |
| 963 | if (CLG_(current_state).collect && loaded->use_base) { |
| 964 | (loaded->use_base)[off_L2_AcCost] += 1000 / use->count; |
| 965 | (loaded->use_base)[off_L2_SpLoss] += i; |
| 966 | } |
| 967 | } |
| 968 | |
| 969 | use->count = 0; |
| 970 | use->mask = 0; |
| 971 | |
| 972 | loaded->memline = memline; |
| 973 | loaded->iaddr = bb_base + current_ii->instr_offset; |
| 974 | loaded->use_base = (CLG_(current_state).nonskipped) ? |
| 975 | CLG_(current_state).nonskipped->skipped : |
| 976 | cost_base + current_ii->cost_offset; |
| 977 | } |
| 978 | |
| 979 | static |
| 980 | CacheModelResult cacheuse_L2_access(Addr memline, line_loaded* l1_loaded) |
| 981 | { |
| 982 | UInt setNo = (memline >> L2.line_size_bits) & (L2.sets_min_1); |
| 983 | UWord* set = &(L2.tags[setNo << L2.assoc_bits]); |
| 984 | UWord tag = memline & L2.tag_mask; |
| 985 | |
| 986 | int i, j, idx; |
| 987 | UWord tmp_tag; |
| 988 | |
| 989 | CLG_DEBUG(6,"L2.Acc(Memline %p): Set %d\n", memline, setNo); |
| 990 | |
| 991 | if (tag == (set[0] & L2.tag_mask)) { |
| 992 | idx = (setNo << L2.assoc_bits) | (set[0] & ~L2.tag_mask); |
| 993 | l1_loaded->dep_use = &(L2.use[idx]); |
| 994 | |
| 995 | CLG_DEBUG(6," Hit0 [idx %d] (line %p from %p): => %08x, count %d\n", |
| 996 | idx, L2.loaded[idx].memline, L2.loaded[idx].iaddr, |
| 997 | L2.use[idx].mask, L2.use[idx].count); |
| 998 | return L2_Hit; |
| 999 | } |
| 1000 | for (i = 1; i < L2.assoc; i++) { |
| 1001 | if (tag == (set[i] & L2.tag_mask)) { |
| 1002 | tmp_tag = set[i]; |
| 1003 | for (j = i; j > 0; j--) { |
| 1004 | set[j] = set[j - 1]; |
| 1005 | } |
| 1006 | set[0] = tmp_tag; |
| 1007 | idx = (setNo << L2.assoc_bits) | (tmp_tag & ~L2.tag_mask); |
| 1008 | l1_loaded->dep_use = &(L2.use[idx]); |
| 1009 | |
| 1010 | CLG_DEBUG(6," Hit%d [idx %d] (line %p from %p): => %08x, count %d\n", |
| 1011 | i, idx, L2.loaded[idx].memline, L2.loaded[idx].iaddr, |
| 1012 | L2.use[idx].mask, L2.use[idx].count); |
| 1013 | return L2_Hit; |
| 1014 | } |
| 1015 | } |
| 1016 | |
| 1017 | /* A miss; install this tag as MRU, shuffle rest down. */ |
| 1018 | tmp_tag = set[L2.assoc - 1] & ~L2.tag_mask; |
| 1019 | for (j = L2.assoc - 1; j > 0; j--) { |
| 1020 | set[j] = set[j - 1]; |
| 1021 | } |
| 1022 | set[0] = tag | tmp_tag; |
| 1023 | idx = (setNo << L2.assoc_bits) | tmp_tag; |
| 1024 | l1_loaded->dep_use = &(L2.use[idx]); |
| 1025 | |
| 1026 | update_L2_use(idx, memline); |
| 1027 | |
| 1028 | return MemAccess; |
| 1029 | } |
| 1030 | |
| 1031 | |
| 1032 | |
| 1033 | |
| 1034 | #define UPDATE_USE(L) \ |
| 1035 | \ |
| 1036 | static CacheModelResult update##_##L##_use(cache_t2* cache, int idx, \ |
| 1037 | UInt mask, Addr memline) \ |
| 1038 | { \ |
| 1039 | line_loaded* loaded = &(cache->loaded[idx]); \ |
| 1040 | line_use* use = &(cache->use[idx]); \ |
| 1041 | int c = ((32 - countBits(use->mask)) * cache->line_size)>>5; \ |
| 1042 | \ |
| 1043 | CLG_DEBUG(2, " %s.miss [%d]: at %p accessing memline %p (mask %08x)\n", \ |
| 1044 | cache->name, idx, bb_base + current_ii->instr_offset, memline, mask); \ |
| 1045 | if (use->count>0) { \ |
| 1046 | CLG_DEBUG(2, " old: used %d, loss bits %d (%08x) [line %p from %p]\n",\ |
| 1047 | use->count, c, use->mask, loaded->memline, loaded->iaddr); \ |
| 1048 | CLG_DEBUG(2, " collect: %d, use_base %p\n", \ |
| 1049 | CLG_(current_state).collect, loaded->use_base); \ |
| 1050 | \ |
| 1051 | if (CLG_(current_state).collect && loaded->use_base) { \ |
| 1052 | (loaded->use_base)[off_##L##_AcCost] += 1000 / use->count; \ |
| 1053 | (loaded->use_base)[off_##L##_SpLoss] += c; \ |
| 1054 | \ |
| 1055 | /* FIXME (?): L1/L2 line sizes must be equal ! */ \ |
| 1056 | loaded->dep_use->mask |= use->mask; \ |
| 1057 | loaded->dep_use->count += use->count; \ |
| 1058 | } \ |
| 1059 | } \ |
| 1060 | \ |
| 1061 | use->count = 1; \ |
| 1062 | use->mask = mask; \ |
| 1063 | loaded->memline = memline; \ |
| 1064 | loaded->iaddr = bb_base + current_ii->instr_offset; \ |
| 1065 | loaded->use_base = (CLG_(current_state).nonskipped) ? \ |
| 1066 | CLG_(current_state).nonskipped->skipped : \ |
| 1067 | cost_base + current_ii->cost_offset; \ |
| 1068 | \ |
| 1069 | if (memline == 0) return L2_Hit; \ |
| 1070 | return cacheuse_L2_access(memline, loaded); \ |
| 1071 | } |
| 1072 | |
| 1073 | UPDATE_USE(I1); |
| 1074 | UPDATE_USE(D1); |
| 1075 | |
| 1076 | CACHEUSE(I1); |
| 1077 | CACHEUSE(D1); |
| 1078 | |
| 1079 | |
| 1080 | static |
| 1081 | void cacheuse_finish(void) |
| 1082 | { |
| 1083 | int i; |
| 1084 | InstrInfo ii = { 0,0,0,0,0 }; |
| 1085 | |
| 1086 | if (!CLG_(current_state).collect) return; |
| 1087 | |
| 1088 | bb_base = 0; |
| 1089 | current_ii = ⅈ |
| 1090 | cost_base = 0; |
| 1091 | |
| 1092 | /* update usage counters */ |
| 1093 | if (I1.use) |
| 1094 | for (i = 0; i < I1.sets * I1.assoc; i++) |
| 1095 | if (I1.loaded[i].use_base) |
| 1096 | update_I1_use( &I1, i, 0,0); |
| 1097 | |
| 1098 | if (D1.use) |
| 1099 | for (i = 0; i < D1.sets * D1.assoc; i++) |
| 1100 | if (D1.loaded[i].use_base) |
| 1101 | update_D1_use( &D1, i, 0,0); |
| 1102 | |
| 1103 | if (L2.use) |
| 1104 | for (i = 0; i < L2.sets * L2.assoc; i++) |
| 1105 | if (L2.loaded[i].use_base) |
| 1106 | update_L2_use(i, 0); |
| 1107 | } |
| 1108 | |
| 1109 | |
| 1110 | |
| 1111 | /*------------------------------------------------------------*/ |
| 1112 | /*--- Helper functions called by instrumented code ---*/ |
| 1113 | /*------------------------------------------------------------*/ |
| 1114 | |
| 1115 | |
| 1116 | static __inline__ |
| 1117 | void inc_costs(CacheModelResult r, ULong* c1, ULong* c2) |
| 1118 | { |
| 1119 | switch(r) { |
| 1120 | case WriteBackMemAccess: |
| 1121 | if (clo_simulate_writeback) { |
| 1122 | c1[3]++; |
| 1123 | c2[3]++; |
| 1124 | } |
| 1125 | // fall through |
| 1126 | |
| 1127 | case MemAccess: |
| 1128 | c1[2]++; |
| 1129 | c2[2]++; |
| 1130 | // fall through |
| 1131 | |
| 1132 | case L2_Hit: |
| 1133 | c1[1]++; |
| 1134 | c2[1]++; |
| 1135 | // fall through |
| 1136 | |
| 1137 | default: |
| 1138 | c1[0]++; |
| 1139 | c2[0]++; |
| 1140 | } |
| 1141 | } |
| 1142 | |
| 1143 | |
| 1144 | VG_REGPARM(1) |
| 1145 | static void log_1I0D(InstrInfo* ii) |
| 1146 | { |
| 1147 | CacheModelResult IrRes; |
| 1148 | |
| 1149 | current_ii = ii; |
| 1150 | IrRes = (*simulator.I1_Read)(bb_base + ii->instr_offset, ii->instr_size); |
| 1151 | |
| 1152 | CLG_DEBUG(6, "log_1I0D: Ir=%p/%u => Ir %d\n", |
| 1153 | bb_base + ii->instr_offset, ii->instr_size, IrRes); |
| 1154 | |
| 1155 | if (CLG_(current_state).collect) { |
| 1156 | ULong* cost_Ir; |
| 1157 | |
| 1158 | if (CLG_(current_state).nonskipped) |
| 1159 | cost_Ir = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_full_Ir; |
| 1160 | else |
| 1161 | cost_Ir = cost_base + ii->cost_offset + off_D0_Ir; |
| 1162 | |
| 1163 | inc_costs(IrRes, cost_Ir, |
| 1164 | CLG_(current_state).cost + CLG_(sets).off_full_Ir ); |
| 1165 | } |
| 1166 | } |
| 1167 | |
| 1168 | |
| 1169 | /* Instruction doing a read access */ |
| 1170 | |
| 1171 | VG_REGPARM(2) |
| 1172 | static void log_1I1Dr(InstrInfo* ii, Addr data) |
| 1173 | { |
| 1174 | CacheModelResult IrRes, DrRes; |
| 1175 | |
| 1176 | current_ii = ii; |
| 1177 | IrRes = (*simulator.I1_Read)(bb_base + ii->instr_offset, ii->instr_size); |
| 1178 | DrRes = (*simulator.D1_Read)(data, ii->data_size); |
| 1179 | |
| 1180 | CLG_DEBUG(6, "log_1I1Dr: Ir=%p/%u, Dr=%p/%u => Ir %d, Dr %d\n", |
| 1181 | bb_base + ii->instr_offset, ii->instr_size, |
| 1182 | data, ii->data_size, IrRes, DrRes); |
| 1183 | |
| 1184 | if (CLG_(current_state).collect) { |
| 1185 | ULong *cost_Ir, *cost_Dr; |
| 1186 | |
| 1187 | if (CLG_(current_state).nonskipped) { |
| 1188 | cost_Ir = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_full_Ir; |
| 1189 | cost_Dr = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_full_Dr; |
| 1190 | } |
| 1191 | else { |
| 1192 | cost_Ir = cost_base + ii->cost_offset + off_D1r_Ir; |
| 1193 | cost_Dr = cost_base + ii->cost_offset + off_D1r_Dr; |
| 1194 | } |
| 1195 | |
| 1196 | inc_costs(IrRes, cost_Ir, |
| 1197 | CLG_(current_state).cost + CLG_(sets).off_full_Ir ); |
| 1198 | inc_costs(DrRes, cost_Dr, |
| 1199 | CLG_(current_state).cost + CLG_(sets).off_full_Dr ); |
| 1200 | } |
| 1201 | } |
| 1202 | |
| 1203 | |
| 1204 | VG_REGPARM(2) |
| 1205 | static void log_0I1Dr(InstrInfo* ii, Addr data) |
| 1206 | { |
| 1207 | CacheModelResult DrRes; |
| 1208 | |
| 1209 | current_ii = ii; |
| 1210 | DrRes = (*simulator.D1_Read)(data, ii->data_size); |
| 1211 | |
| 1212 | CLG_DEBUG(6, "log_0I1Dr: Dr=%p/%u => Dr %d\n", |
| 1213 | data, ii->data_size, DrRes); |
| 1214 | |
| 1215 | if (CLG_(current_state).collect) { |
| 1216 | ULong *cost_Dr; |
| 1217 | |
| 1218 | if (CLG_(current_state).nonskipped) { |
| 1219 | cost_Dr = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_full_Dr; |
| 1220 | } |
| 1221 | else { |
| 1222 | cost_Dr = cost_base + ii->cost_offset + off_D1r_Dr; |
| 1223 | } |
| 1224 | |
| 1225 | inc_costs(DrRes, cost_Dr, |
| 1226 | CLG_(current_state).cost + CLG_(sets).off_full_Dr ); |
| 1227 | } |
| 1228 | } |
| 1229 | |
| 1230 | |
| 1231 | /* Instruction doing a write access */ |
| 1232 | |
| 1233 | VG_REGPARM(2) |
| 1234 | static void log_1I1Dw(InstrInfo* ii, Addr data) |
| 1235 | { |
| 1236 | CacheModelResult IrRes, DwRes; |
| 1237 | |
| 1238 | current_ii = ii; |
| 1239 | IrRes = (*simulator.I1_Read)(bb_base + ii->instr_offset, ii->instr_size); |
| 1240 | DwRes = (*simulator.D1_Write)(data, ii->data_size); |
| 1241 | |
| 1242 | CLG_DEBUG(6, "log_1I1Dw: Ir=%p/%u, Dw=%p/%u => Ir %d, Dw %d\n", |
| 1243 | bb_base + ii->instr_offset, ii->instr_size, |
| 1244 | data, ii->data_size, IrRes, DwRes); |
| 1245 | |
| 1246 | if (CLG_(current_state).collect) { |
| 1247 | ULong *cost_Ir, *cost_Dw; |
| 1248 | |
| 1249 | if (CLG_(current_state).nonskipped) { |
| 1250 | cost_Ir = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_sim_Ir; |
| 1251 | cost_Dw = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_sim_Dw; |
| 1252 | } |
| 1253 | else { |
| 1254 | cost_Ir = cost_base + ii->cost_offset + off_D1w_Ir; |
| 1255 | cost_Dw = cost_base + ii->cost_offset + off_D1w_Dw; |
| 1256 | } |
| 1257 | |
| 1258 | inc_costs(IrRes, cost_Ir, |
| 1259 | CLG_(current_state).cost + CLG_(sets).off_full_Ir ); |
| 1260 | inc_costs(DwRes, cost_Dw, |
| 1261 | CLG_(current_state).cost + CLG_(sets).off_full_Dw ); |
| 1262 | } |
| 1263 | } |
| 1264 | |
| 1265 | VG_REGPARM(2) |
| 1266 | static void log_0I1Dw(InstrInfo* ii, Addr data) |
| 1267 | { |
| 1268 | CacheModelResult DwRes; |
| 1269 | |
| 1270 | current_ii = ii; |
| 1271 | DwRes = (*simulator.D1_Write)(data, ii->data_size); |
| 1272 | |
| 1273 | CLG_DEBUG(6, "log_0I1Dw: Dw=%p/%u => Dw %d\n", |
| 1274 | data, ii->data_size, DwRes); |
| 1275 | |
| 1276 | if (CLG_(current_state).collect) { |
| 1277 | ULong *cost_Dw; |
| 1278 | |
| 1279 | if (CLG_(current_state).nonskipped) { |
| 1280 | cost_Dw = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_full_Dw; |
| 1281 | } |
| 1282 | else { |
| 1283 | cost_Dw = cost_base + ii->cost_offset + off_D1w_Dw; |
| 1284 | } |
| 1285 | |
| 1286 | inc_costs(DwRes, cost_Dw, |
| 1287 | CLG_(current_state).cost + CLG_(sets).off_full_Dw ); |
| 1288 | } |
| 1289 | } |
| 1290 | |
| 1291 | /* Instruction doing a read and a write access */ |
| 1292 | |
| 1293 | VG_REGPARM(3) |
| 1294 | static void log_1I2D(InstrInfo* ii, Addr data1, Addr data2) |
| 1295 | { |
| 1296 | CacheModelResult IrRes, DrRes, DwRes; |
| 1297 | |
| 1298 | current_ii = ii; |
| 1299 | IrRes = (*simulator.I1_Read)(bb_base + ii->instr_offset, ii->instr_size); |
| 1300 | DrRes = (*simulator.D1_Read)(data1, ii->data_size); |
| 1301 | DwRes = (*simulator.D1_Write)(data2, ii->data_size); |
| 1302 | |
| 1303 | CLG_DEBUG(6, |
| 1304 | "log_1I2D: Ir=%p/%u, Dr=%p/%u, Dw=%p/%u => Ir %d, Dr %d, Dw %d\n", |
| 1305 | bb_base + ii->instr_offset, ii->instr_size, |
| 1306 | data1, ii->data_size, data2, ii->data_size, IrRes, DrRes, DwRes); |
| 1307 | |
| 1308 | if (CLG_(current_state).collect) { |
| 1309 | ULong *cost_Ir, *cost_Dr, *cost_Dw; |
| 1310 | |
| 1311 | if (CLG_(current_state).nonskipped) { |
| 1312 | cost_Ir = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_sim_Ir; |
| 1313 | cost_Dr = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_sim_Dr; |
| 1314 | cost_Dw = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_sim_Dw; |
| 1315 | } |
| 1316 | else { |
| 1317 | cost_Ir = cost_base + ii->cost_offset + off_D2_Ir; |
| 1318 | cost_Dr = cost_base + ii->cost_offset + off_D2_Dr; |
| 1319 | cost_Dw = cost_base + ii->cost_offset + off_D2_Dw; |
| 1320 | } |
| 1321 | |
| 1322 | inc_costs(IrRes, cost_Ir, |
| 1323 | CLG_(current_state).cost + CLG_(sets).off_full_Ir ); |
| 1324 | inc_costs(DrRes, cost_Dr, |
| 1325 | CLG_(current_state).cost + CLG_(sets).off_full_Dr ); |
| 1326 | inc_costs(DwRes, cost_Dw, |
| 1327 | CLG_(current_state).cost + CLG_(sets).off_full_Dw ); |
| 1328 | } |
| 1329 | } |
| 1330 | |
| 1331 | VG_REGPARM(3) |
| 1332 | static void log_0I2D(InstrInfo* ii, Addr data1, Addr data2) |
| 1333 | { |
| 1334 | CacheModelResult DrRes, DwRes; |
| 1335 | |
| 1336 | current_ii = ii; |
| 1337 | DrRes = (*simulator.D1_Read)(data1, ii->data_size); |
| 1338 | DwRes = (*simulator.D1_Write)(data2, ii->data_size); |
| 1339 | |
| 1340 | CLG_DEBUG(6, |
| 1341 | "log_0D2D: Dr=%p/%u, Dw=%p/%u => Dr %d, Dw %d\n", |
| 1342 | data1, ii->data_size, data2, ii->data_size, DrRes, DwRes); |
| 1343 | |
| 1344 | if (CLG_(current_state).collect) { |
| 1345 | ULong *cost_Dr, *cost_Dw; |
| 1346 | |
| 1347 | if (CLG_(current_state).nonskipped) { |
| 1348 | cost_Dr = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_sim_Dr; |
| 1349 | cost_Dw = CLG_(current_state).nonskipped->skipped + CLG_(sets).off_sim_Dw; |
| 1350 | } |
| 1351 | else { |
| 1352 | cost_Dr = cost_base + ii->cost_offset + off_D2_Dr; |
| 1353 | cost_Dw = cost_base + ii->cost_offset + off_D2_Dw; |
| 1354 | } |
| 1355 | |
| 1356 | inc_costs(DrRes, cost_Dr, |
| 1357 | CLG_(current_state).cost + CLG_(sets).off_full_Dr ); |
| 1358 | inc_costs(DwRes, cost_Dw, |
| 1359 | CLG_(current_state).cost + CLG_(sets).off_full_Dw ); |
| 1360 | } |
| 1361 | } |
| 1362 | |
| 1363 | |
| 1364 | /*------------------------------------------------------------*/ |
| 1365 | /*--- Cache configuration ---*/ |
| 1366 | /*------------------------------------------------------------*/ |
| 1367 | |
| 1368 | #define UNDEFINED_CACHE ((cache_t) { -1, -1, -1 }) |
| 1369 | |
| 1370 | static cache_t clo_I1_cache = UNDEFINED_CACHE; |
| 1371 | static cache_t clo_D1_cache = UNDEFINED_CACHE; |
| 1372 | static cache_t clo_L2_cache = UNDEFINED_CACHE; |
| 1373 | |
| 1374 | |
| 1375 | /* Checks cache config is ok; makes it so if not. */ |
| 1376 | static |
| 1377 | void check_cache(cache_t* cache, Char *name) |
| 1378 | { |
| 1379 | /* First check they're all powers of two */ |
| 1380 | if (-1 == VG_(log2)(cache->size)) { |
| 1381 | VG_(message)(Vg_UserMsg, |
| 1382 | "error: %s size of %dB not a power of two; aborting.", |
| 1383 | name, cache->size); |
| 1384 | VG_(exit)(1); |
| 1385 | } |
| 1386 | |
| 1387 | if (-1 == VG_(log2)(cache->assoc)) { |
| 1388 | VG_(message)(Vg_UserMsg, |
| 1389 | "error: %s associativity of %d not a power of two; aborting.", |
| 1390 | name, cache->assoc); |
| 1391 | VG_(exit)(1); |
| 1392 | } |
| 1393 | |
| 1394 | if (-1 == VG_(log2)(cache->line_size)) { |
| 1395 | VG_(message)(Vg_UserMsg, |
| 1396 | "error: %s line size of %dB not a power of two; aborting.", |
| 1397 | name, cache->line_size); |
| 1398 | VG_(exit)(1); |
| 1399 | } |
| 1400 | |
| 1401 | // Then check line size >= 16 -- any smaller and a single instruction could |
| 1402 | // straddle three cache lines, which breaks a simulation assertion and is |
| 1403 | // stupid anyway. |
| 1404 | if (cache->line_size < MIN_LINE_SIZE) { |
| 1405 | VG_(message)(Vg_UserMsg, |
| 1406 | "error: %s line size of %dB too small; aborting.", |
| 1407 | name, cache->line_size); |
| 1408 | VG_(exit)(1); |
| 1409 | } |
| 1410 | |
| 1411 | /* Then check cache size > line size (causes seg faults if not). */ |
| 1412 | if (cache->size <= cache->line_size) { |
| 1413 | VG_(message)(Vg_UserMsg, |
| 1414 | "error: %s cache size of %dB <= line size of %dB; aborting.", |
| 1415 | name, cache->size, cache->line_size); |
| 1416 | VG_(exit)(1); |
| 1417 | } |
| 1418 | |
| 1419 | /* Then check assoc <= (size / line size) (seg faults otherwise). */ |
| 1420 | if (cache->assoc > (cache->size / cache->line_size)) { |
| 1421 | VG_(message)(Vg_UserMsg, |
| 1422 | "warning: %s associativity > (size / line size); aborting.", name); |
| 1423 | VG_(exit)(1); |
| 1424 | } |
| 1425 | } |
| 1426 | |
| 1427 | static |
| 1428 | void configure_caches(cache_t* I1c, cache_t* D1c, cache_t* L2c) |
| 1429 | { |
| 1430 | #define DEFINED(L) (-1 != L.size || -1 != L.assoc || -1 != L.line_size) |
| 1431 | |
| 1432 | Int n_clos = 0; |
| 1433 | |
| 1434 | // Count how many were defined on the command line. |
| 1435 | if (DEFINED(clo_I1_cache)) { n_clos++; } |
| 1436 | if (DEFINED(clo_D1_cache)) { n_clos++; } |
| 1437 | if (DEFINED(clo_L2_cache)) { n_clos++; } |
| 1438 | |
| 1439 | // Set the cache config (using auto-detection, if supported by the |
| 1440 | // architecture) |
| 1441 | VG_(configure_caches)( I1c, D1c, L2c, (3 == n_clos) ); |
| 1442 | |
| 1443 | // Then replace with any defined on the command line. |
| 1444 | if (DEFINED(clo_I1_cache)) { *I1c = clo_I1_cache; } |
| 1445 | if (DEFINED(clo_D1_cache)) { *D1c = clo_D1_cache; } |
| 1446 | if (DEFINED(clo_L2_cache)) { *L2c = clo_L2_cache; } |
| 1447 | |
| 1448 | // Then check values and fix if not acceptable. |
| 1449 | check_cache(I1c, "I1"); |
| 1450 | check_cache(D1c, "D1"); |
| 1451 | check_cache(L2c, "L2"); |
| 1452 | |
| 1453 | if (VG_(clo_verbosity) > 1) { |
| 1454 | VG_(message)(Vg_UserMsg, "Cache configuration used:"); |
| 1455 | VG_(message)(Vg_UserMsg, " I1: %dB, %d-way, %dB lines", |
| 1456 | I1c->size, I1c->assoc, I1c->line_size); |
| 1457 | VG_(message)(Vg_UserMsg, " D1: %dB, %d-way, %dB lines", |
| 1458 | D1c->size, D1c->assoc, D1c->line_size); |
| 1459 | VG_(message)(Vg_UserMsg, " L2: %dB, %d-way, %dB lines", |
| 1460 | L2c->size, L2c->assoc, L2c->line_size); |
| 1461 | } |
| 1462 | #undef CMD_LINE_DEFINED |
| 1463 | } |
| 1464 | |
| 1465 | |
| 1466 | /* Initialize and clear simulator state */ |
| 1467 | static void cachesim_post_clo_init(void) |
| 1468 | { |
| 1469 | /* Cache configurations. */ |
| 1470 | cache_t I1c, D1c, L2c; |
| 1471 | |
| 1472 | /* Initialize access handlers */ |
| 1473 | if (!CLG_(clo).simulate_cache) { |
| 1474 | CLG_(cachesim).log_1I0D = 0; |
| 1475 | CLG_(cachesim).log_1I0D_name = "(no function)"; |
| 1476 | |
| 1477 | CLG_(cachesim).log_1I1Dr = 0; |
| 1478 | CLG_(cachesim).log_1I1Dw = 0; |
| 1479 | CLG_(cachesim).log_1I2D = 0; |
| 1480 | CLG_(cachesim).log_1I1Dr_name = "(no function)"; |
| 1481 | CLG_(cachesim).log_1I1Dw_name = "(no function)"; |
| 1482 | CLG_(cachesim).log_1I2D_name = "(no function)"; |
| 1483 | |
| 1484 | CLG_(cachesim).log_0I1Dr = 0; |
| 1485 | CLG_(cachesim).log_0I1Dw = 0; |
| 1486 | CLG_(cachesim).log_0I2D = 0; |
| 1487 | CLG_(cachesim).log_0I1Dr_name = "(no function)"; |
| 1488 | CLG_(cachesim).log_0I1Dw_name = "(no function)"; |
| 1489 | CLG_(cachesim).log_0I2D_name = "(no function)"; |
| 1490 | return; |
| 1491 | } |
| 1492 | |
| 1493 | /* Configuration of caches only needed with real cache simulation */ |
| 1494 | configure_caches(&I1c, &D1c, &L2c); |
| 1495 | |
| 1496 | I1.name = "I1"; |
| 1497 | D1.name = "D1"; |
| 1498 | L2.name = "L2"; |
| 1499 | |
| 1500 | cachesim_initcache(I1c, &I1); |
| 1501 | cachesim_initcache(D1c, &D1); |
| 1502 | cachesim_initcache(L2c, &L2); |
| 1503 | |
| 1504 | /* the other cache simulators use the standard helpers |
| 1505 | * with dispatching via simulator struct */ |
| 1506 | |
| 1507 | CLG_(cachesim).log_1I0D = log_1I0D; |
| 1508 | CLG_(cachesim).log_1I0D_name = "log_1I0D"; |
| 1509 | |
| 1510 | CLG_(cachesim).log_1I1Dr = log_1I1Dr; |
| 1511 | CLG_(cachesim).log_1I1Dw = log_1I1Dw; |
| 1512 | CLG_(cachesim).log_1I2D = log_1I2D; |
| 1513 | CLG_(cachesim).log_1I1Dr_name = "log_1I1Dr"; |
| 1514 | CLG_(cachesim).log_1I1Dw_name = "log_1I1Dw"; |
| 1515 | CLG_(cachesim).log_1I2D_name = "log_1I2D"; |
| 1516 | |
| 1517 | CLG_(cachesim).log_0I1Dr = log_0I1Dr; |
| 1518 | CLG_(cachesim).log_0I1Dw = log_0I1Dw; |
| 1519 | CLG_(cachesim).log_0I2D = log_0I2D; |
| 1520 | CLG_(cachesim).log_0I1Dr_name = "log_0I1Dr"; |
| 1521 | CLG_(cachesim).log_0I1Dw_name = "log_0I1Dw"; |
| 1522 | CLG_(cachesim).log_0I2D_name = "log_0I2D"; |
| 1523 | |
| 1524 | if (clo_collect_cacheuse) { |
| 1525 | |
| 1526 | /* Output warning for not supported option combinations */ |
| 1527 | if (clo_simulate_hwpref) { |
| 1528 | VG_(message)(Vg_DebugMsg, |
| 1529 | "warning: prefetch simulation can not be used with cache usage"); |
| 1530 | clo_simulate_hwpref = False; |
| 1531 | } |
| 1532 | |
| 1533 | if (clo_simulate_writeback) { |
| 1534 | VG_(message)(Vg_DebugMsg, |
| 1535 | "warning: write-back simulation can not be used with cache usage"); |
| 1536 | clo_simulate_writeback = False; |
| 1537 | } |
| 1538 | |
| 1539 | simulator.I1_Read = cacheuse_I1_doRead; |
| 1540 | simulator.D1_Read = cacheuse_D1_doRead; |
| 1541 | simulator.D1_Write = cacheuse_D1_doRead; |
| 1542 | return; |
| 1543 | } |
| 1544 | |
| 1545 | if (clo_simulate_hwpref) { |
| 1546 | prefetch_clear(); |
| 1547 | |
| 1548 | if (clo_simulate_writeback) { |
| 1549 | simulator.I1_Read = prefetch_I1_Read; |
| 1550 | simulator.D1_Read = prefetch_D1_Read; |
| 1551 | simulator.D1_Write = prefetch_D1_Write; |
| 1552 | } |
| 1553 | else { |
| 1554 | simulator.I1_Read = prefetch_I1_ref; |
| 1555 | simulator.D1_Read = prefetch_D1_ref; |
| 1556 | simulator.D1_Write = prefetch_D1_ref; |
| 1557 | } |
| 1558 | |
| 1559 | return; |
| 1560 | } |
| 1561 | |
| 1562 | if (clo_simulate_writeback) { |
| 1563 | simulator.I1_Read = cachesim_I1_Read; |
| 1564 | simulator.D1_Read = cachesim_D1_Read; |
| 1565 | simulator.D1_Write = cachesim_D1_Write; |
| 1566 | } |
| 1567 | else { |
| 1568 | simulator.I1_Read = cachesim_I1_ref; |
| 1569 | simulator.D1_Read = cachesim_D1_ref; |
| 1570 | simulator.D1_Write = cachesim_D1_ref; |
| 1571 | } |
| 1572 | } |
| 1573 | |
| 1574 | |
| 1575 | /* Clear simulator state. Has to be initialized before */ |
| 1576 | static |
| 1577 | void cachesim_clear(void) |
| 1578 | { |
| 1579 | cachesim_clearcache(&I1); |
| 1580 | cachesim_clearcache(&D1); |
| 1581 | cachesim_clearcache(&L2); |
| 1582 | |
| 1583 | prefetch_clear(); |
| 1584 | } |
| 1585 | |
| 1586 | |
| 1587 | static void cachesim_getdesc(Char* buf) |
| 1588 | { |
| 1589 | Int p; |
| 1590 | p = VG_(sprintf)(buf, "\ndesc: I1 cache: %s\n", I1.desc_line); |
| 1591 | p += VG_(sprintf)(buf+p, "desc: D1 cache: %s\n", D1.desc_line); |
| 1592 | VG_(sprintf)(buf+p, "desc: L2 cache: %s\n", L2.desc_line); |
| 1593 | } |
| 1594 | |
| 1595 | static |
| 1596 | void cachesim_print_opts(void) |
| 1597 | { |
| 1598 | VG_(printf)( |
| 1599 | "\n cache simulator options:\n" |
| 1600 | " --simulate-cache=no|yes Do cache simulation [no]\n" |
| 1601 | " --simulate-wb=no|yes Count write-back events [no]\n" |
| 1602 | " --simulate-hwpref=no|yes Simulate hardware prefetch [no]\n" |
| 1603 | #if CLG_EXPERIMENTAL |
| 1604 | " --simulate-sectors=no|yes Simulate sectored behaviour [no]\n" |
| 1605 | #endif |
| 1606 | " --cacheuse=no|yes Collect cache block use [no]\n" |
| 1607 | " --I1=<size>,<assoc>,<line_size> set I1 cache manually\n" |
| 1608 | " --D1=<size>,<assoc>,<line_size> set D1 cache manually\n" |
| 1609 | " --L2=<size>,<assoc>,<line_size> set L2 cache manually\n" |
| 1610 | ); |
| 1611 | } |
| 1612 | |
| 1613 | static void parse_opt ( cache_t* cache, char* orig_opt, int opt_len ) |
| 1614 | { |
| 1615 | int i1, i2, i3; |
| 1616 | int i; |
| 1617 | char *opt = VG_(strdup)(orig_opt); |
| 1618 | |
| 1619 | i = i1 = opt_len; |
| 1620 | |
| 1621 | /* Option looks like "--I1=65536,2,64". |
| 1622 | * Find commas, replace with NULs to make three independent |
| 1623 | * strings, then extract numbers. Yuck. */ |
| 1624 | while (VG_(isdigit)(opt[i])) i++; |
| 1625 | if (',' == opt[i]) { |
| 1626 | opt[i++] = '\0'; |
| 1627 | i2 = i; |
| 1628 | } else goto bad; |
| 1629 | while (VG_(isdigit)(opt[i])) i++; |
| 1630 | if (',' == opt[i]) { |
| 1631 | opt[i++] = '\0'; |
| 1632 | i3 = i; |
| 1633 | } else goto bad; |
| 1634 | while (VG_(isdigit)(opt[i])) i++; |
| 1635 | if ('\0' != opt[i]) goto bad; |
| 1636 | |
| 1637 | cache->size = (Int)VG_(atoll)(opt + i1); |
| 1638 | cache->assoc = (Int)VG_(atoll)(opt + i2); |
| 1639 | cache->line_size = (Int)VG_(atoll)(opt + i3); |
| 1640 | |
| 1641 | VG_(free)(opt); |
| 1642 | |
| 1643 | return; |
| 1644 | |
| 1645 | bad: |
sewardj | 6893d65 | 2006-10-15 01:25:13 +0000 | [diff] [blame] | 1646 | VG_(err_bad_option)(orig_opt); |
weidendo | a17f2a3 | 2006-03-20 10:27:30 +0000 | [diff] [blame] | 1647 | } |
| 1648 | |
| 1649 | /* Check for command line option for cache configuration. |
| 1650 | * Return False if unknown and not handled. |
| 1651 | * |
| 1652 | * Called from CLG_(process_cmd_line_option)() in clo.c |
| 1653 | */ |
| 1654 | static Bool cachesim_parse_opt(Char* arg) |
| 1655 | { |
| 1656 | if (0 == VG_(strcmp)(arg, "--simulate-wb=yes")) |
| 1657 | clo_simulate_writeback = True; |
| 1658 | else if (0 == VG_(strcmp)(arg, "--simulate-wb=no")) |
| 1659 | clo_simulate_writeback = False; |
| 1660 | |
| 1661 | else if (0 == VG_(strcmp)(arg, "--simulate-hwpref=yes")) |
| 1662 | clo_simulate_hwpref = True; |
| 1663 | else if (0 == VG_(strcmp)(arg, "--simulate-hwpref=no")) |
| 1664 | clo_simulate_hwpref = False; |
| 1665 | |
| 1666 | else if (0 == VG_(strcmp)(arg, "--simulate-sectors=yes")) |
| 1667 | clo_simulate_sectors = True; |
| 1668 | else if (0 == VG_(strcmp)(arg, "--simulate-sectors=no")) |
| 1669 | clo_simulate_sectors = False; |
| 1670 | |
| 1671 | else if (0 == VG_(strcmp)(arg, "--cacheuse=yes")) { |
| 1672 | clo_collect_cacheuse = True; |
| 1673 | /* Use counters only make sense with fine dumping */ |
| 1674 | CLG_(clo).dump_instr = True; |
| 1675 | } |
| 1676 | else if (0 == VG_(strcmp)(arg, "--cacheuse=no")) |
| 1677 | clo_collect_cacheuse = False; |
| 1678 | |
| 1679 | /* 5 is length of "--I1=" */ |
| 1680 | else if (0 == VG_(strncmp)(arg, "--I1=", 5)) |
| 1681 | parse_opt(&clo_I1_cache, arg, 5); |
| 1682 | else if (0 == VG_(strncmp)(arg, "--D1=", 5)) |
| 1683 | parse_opt(&clo_D1_cache, arg, 5); |
| 1684 | else if (0 == VG_(strncmp)(arg, "--L2=", 5)) |
| 1685 | parse_opt(&clo_L2_cache, arg, 5); |
| 1686 | else |
| 1687 | return False; |
| 1688 | |
| 1689 | return True; |
| 1690 | } |
| 1691 | |
| 1692 | /* Adds commas to ULong, right justifying in a field field_width wide, returns |
| 1693 | * the string in buf. */ |
| 1694 | static |
| 1695 | Int commify(ULong n, int field_width, char* buf) |
| 1696 | { |
| 1697 | int len, n_commas, i, j, new_len, space; |
| 1698 | |
| 1699 | VG_(sprintf)(buf, "%llu", n); |
| 1700 | len = VG_(strlen)(buf); |
| 1701 | n_commas = (len - 1) / 3; |
| 1702 | new_len = len + n_commas; |
| 1703 | space = field_width - new_len; |
| 1704 | |
| 1705 | /* Allow for printing a number in a field_width smaller than it's size */ |
| 1706 | if (space < 0) space = 0; |
| 1707 | |
| 1708 | /* Make j = -1 because we copy the '\0' before doing the numbers in groups |
| 1709 | * of three. */ |
| 1710 | for (j = -1, i = len ; i >= 0; i--) { |
| 1711 | buf[i + n_commas + space] = buf[i]; |
| 1712 | |
| 1713 | if ((i>0) && (3 == ++j)) { |
| 1714 | j = 0; |
| 1715 | n_commas--; |
| 1716 | buf[i + n_commas + space] = ','; |
| 1717 | } |
| 1718 | } |
| 1719 | /* Right justify in field. */ |
| 1720 | for (i = 0; i < space; i++) buf[i] = ' '; |
| 1721 | return new_len; |
| 1722 | } |
| 1723 | |
| 1724 | static |
| 1725 | void percentify(Int n, Int ex, Int field_width, char buf[]) |
| 1726 | { |
| 1727 | int i, len, space; |
| 1728 | |
| 1729 | VG_(sprintf)(buf, "%d.%d%%", n / ex, n % ex); |
| 1730 | len = VG_(strlen)(buf); |
| 1731 | space = field_width - len; |
| 1732 | if (space < 0) space = 0; /* Allow for v. small field_width */ |
| 1733 | i = len; |
| 1734 | |
| 1735 | /* Right justify in field */ |
| 1736 | for ( ; i >= 0; i--) buf[i + space] = buf[i]; |
| 1737 | for (i = 0; i < space; i++) buf[i] = ' '; |
| 1738 | } |
| 1739 | |
| 1740 | static |
| 1741 | void cachesim_printstat(void) |
| 1742 | { |
| 1743 | FullCost total = CLG_(total_cost), D_total = 0; |
| 1744 | ULong L2_total_m, L2_total_mr, L2_total_mw, |
| 1745 | L2_total, L2_total_r, L2_total_w; |
| 1746 | char buf1[RESULTS_BUF_LEN], |
| 1747 | buf2[RESULTS_BUF_LEN], |
| 1748 | buf3[RESULTS_BUF_LEN]; |
| 1749 | Int l1, l2, l3; |
| 1750 | Int p; |
| 1751 | |
| 1752 | if ((VG_(clo_verbosity) >1) && clo_simulate_hwpref) { |
| 1753 | VG_(message)(Vg_DebugMsg, "Prefetch Up: %llu", |
| 1754 | prefetch_up); |
| 1755 | VG_(message)(Vg_DebugMsg, "Prefetch Down: %llu", |
| 1756 | prefetch_down); |
| 1757 | VG_(message)(Vg_DebugMsg, ""); |
| 1758 | } |
| 1759 | |
| 1760 | /* I cache results. Use the I_refs value to determine the first column |
| 1761 | * width. */ |
| 1762 | l1 = commify(total[CLG_(sets).off_full_Ir], 0, buf1); |
| 1763 | VG_(message)(Vg_UserMsg, "I refs: %s", buf1); |
| 1764 | |
| 1765 | if (!CLG_(clo).simulate_cache) return; |
| 1766 | |
| 1767 | commify(total[CLG_(sets).off_full_Ir +1], l1, buf1); |
| 1768 | VG_(message)(Vg_UserMsg, "I1 misses: %s", buf1); |
| 1769 | |
| 1770 | commify(total[CLG_(sets).off_full_Ir +2], l1, buf1); |
| 1771 | VG_(message)(Vg_UserMsg, "L2i misses: %s", buf1); |
| 1772 | |
| 1773 | p = 100; |
| 1774 | |
| 1775 | if (0 == total[CLG_(sets).off_full_Ir]) |
| 1776 | total[CLG_(sets).off_full_Ir] = 1; |
| 1777 | |
| 1778 | percentify(total[CLG_(sets).off_full_Ir+1] * 100 * p / |
| 1779 | total[CLG_(sets).off_full_Ir], p, l1+1, buf1); |
| 1780 | VG_(message)(Vg_UserMsg, "I1 miss rate: %s", buf1); |
| 1781 | |
| 1782 | percentify(total[CLG_(sets).off_full_Ir+2] * 100 * p / |
| 1783 | total[CLG_(sets).off_full_Ir], p, l1+1, buf1); |
| 1784 | VG_(message)(Vg_UserMsg, "L2i miss rate: %s", buf1); |
| 1785 | VG_(message)(Vg_UserMsg, ""); |
| 1786 | |
| 1787 | /* D cache results. |
| 1788 | Use the D_refs.rd and D_refs.wr values to determine the |
| 1789 | * width of columns 2 & 3. */ |
| 1790 | |
| 1791 | D_total = CLG_(get_eventset_cost)( CLG_(sets).full ); |
| 1792 | CLG_(init_cost)( CLG_(sets).full, D_total); |
| 1793 | CLG_(copy_cost)( CLG_(sets).Dr, D_total, total + CLG_(sets).off_full_Dr ); |
| 1794 | CLG_(add_cost) ( CLG_(sets).Dw, D_total, total + CLG_(sets).off_full_Dw ); |
| 1795 | |
| 1796 | commify( D_total[0], l1, buf1); |
| 1797 | l2 = commify(total[CLG_(sets).off_full_Dr], 0, buf2); |
| 1798 | l3 = commify(total[CLG_(sets).off_full_Dw], 0, buf3); |
| 1799 | VG_(message)(Vg_UserMsg, "D refs: %s (%s rd + %s wr)", |
| 1800 | buf1, buf2, buf3); |
| 1801 | |
| 1802 | commify( D_total[1], l1, buf1); |
| 1803 | commify(total[CLG_(sets).off_full_Dr+1], l2, buf2); |
| 1804 | commify(total[CLG_(sets).off_full_Dw+1], l3, buf3); |
| 1805 | VG_(message)(Vg_UserMsg, "D1 misses: %s (%s rd + %s wr)", |
| 1806 | buf1, buf2, buf3); |
| 1807 | |
| 1808 | commify( D_total[2], l1, buf1); |
| 1809 | commify(total[CLG_(sets).off_full_Dr+2], l2, buf2); |
| 1810 | commify(total[CLG_(sets).off_full_Dw+2], l3, buf3); |
| 1811 | VG_(message)(Vg_UserMsg, "L2d misses: %s (%s rd + %s wr)", |
| 1812 | buf1, buf2, buf3); |
| 1813 | |
| 1814 | p = 10; |
| 1815 | |
| 1816 | if (0 == D_total[0]) D_total[0] = 1; |
| 1817 | if (0 == total[CLG_(sets).off_full_Dr]) total[CLG_(sets).off_full_Dr] = 1; |
| 1818 | if (0 == total[CLG_(sets).off_full_Dw]) total[CLG_(sets).off_full_Dw] = 1; |
| 1819 | |
| 1820 | percentify( D_total[1] * 100 * p / D_total[0], p, l1+1, buf1); |
| 1821 | percentify(total[CLG_(sets).off_full_Dr+1] * 100 * p / |
| 1822 | total[CLG_(sets).off_full_Dr], p, l2+1, buf2); |
| 1823 | percentify(total[CLG_(sets).off_full_Dw+1] * 100 * p / |
| 1824 | total[CLG_(sets).off_full_Dw], p, l3+1, buf3); |
| 1825 | VG_(message)(Vg_UserMsg, "D1 miss rate: %s (%s + %s )", buf1, buf2,buf3); |
| 1826 | |
| 1827 | percentify( D_total[2] * 100 * p / D_total[0], p, l1+1, buf1); |
| 1828 | percentify(total[CLG_(sets).off_full_Dr+2] * 100 * p / |
| 1829 | total[CLG_(sets).off_full_Dr], p, l2+1, buf2); |
| 1830 | percentify(total[CLG_(sets).off_full_Dw+2] * 100 * p / |
| 1831 | total[CLG_(sets).off_full_Dw], p, l3+1, buf3); |
| 1832 | VG_(message)(Vg_UserMsg, "L2d miss rate: %s (%s + %s )", buf1, buf2,buf3); |
| 1833 | VG_(message)(Vg_UserMsg, ""); |
| 1834 | |
| 1835 | |
| 1836 | |
| 1837 | /* L2 overall results */ |
| 1838 | |
| 1839 | L2_total = |
| 1840 | total[CLG_(sets).off_full_Dr +1] + |
| 1841 | total[CLG_(sets).off_full_Dw +1] + |
| 1842 | total[CLG_(sets).off_full_Ir +1]; |
| 1843 | L2_total_r = |
| 1844 | total[CLG_(sets).off_full_Dr +1] + |
| 1845 | total[CLG_(sets).off_full_Ir +1]; |
| 1846 | L2_total_w = total[CLG_(sets).off_full_Dw +1]; |
| 1847 | commify(L2_total, l1, buf1); |
| 1848 | commify(L2_total_r, l2, buf2); |
| 1849 | commify(L2_total_w, l3, buf3); |
| 1850 | VG_(message)(Vg_UserMsg, "L2 refs: %s (%s rd + %s wr)", |
| 1851 | buf1, buf2, buf3); |
| 1852 | |
| 1853 | L2_total_m = |
| 1854 | total[CLG_(sets).off_full_Dr +2] + |
| 1855 | total[CLG_(sets).off_full_Dw +2] + |
| 1856 | total[CLG_(sets).off_full_Ir +2]; |
| 1857 | L2_total_mr = |
| 1858 | total[CLG_(sets).off_full_Dr +2] + |
| 1859 | total[CLG_(sets).off_full_Ir +2]; |
| 1860 | L2_total_mw = total[CLG_(sets).off_full_Dw +2]; |
| 1861 | commify(L2_total_m, l1, buf1); |
| 1862 | commify(L2_total_mr, l2, buf2); |
| 1863 | commify(L2_total_mw, l3, buf3); |
| 1864 | VG_(message)(Vg_UserMsg, "L2 misses: %s (%s rd + %s wr)", |
| 1865 | buf1, buf2, buf3); |
| 1866 | |
| 1867 | percentify(L2_total_m * 100 * p / |
| 1868 | (total[CLG_(sets).off_full_Ir] + D_total[0]), p, l1+1, buf1); |
| 1869 | percentify(L2_total_mr * 100 * p / |
| 1870 | (total[CLG_(sets).off_full_Ir] + total[CLG_(sets).off_full_Dr]), |
| 1871 | p, l2+1, buf2); |
| 1872 | percentify(L2_total_mw * 100 * p / |
| 1873 | total[CLG_(sets).off_full_Dw], p, l3+1, buf3); |
| 1874 | VG_(message)(Vg_UserMsg, "L2 miss rate: %s (%s + %s )", |
| 1875 | buf1, buf2,buf3); |
| 1876 | } |
| 1877 | |
| 1878 | |
| 1879 | /*------------------------------------------------------------*/ |
| 1880 | /*--- Setup for Event set. ---*/ |
| 1881 | /*------------------------------------------------------------*/ |
| 1882 | |
| 1883 | struct event_sets CLG_(sets); |
| 1884 | |
| 1885 | void CLG_(init_eventsets)(Int max_user) |
| 1886 | { |
| 1887 | EventType * e1, *e2, *e3, *e4; |
| 1888 | EventSet *Ir, *Dr, *Dw; |
| 1889 | EventSet *D0, *D1r, *D1w, *D2; |
| 1890 | EventSet *sim, *full; |
| 1891 | EventSet *use; |
| 1892 | int sizeOfUseIr; |
| 1893 | |
| 1894 | use = CLG_(get_eventset)("Use", 4); |
| 1895 | if (clo_collect_cacheuse) { |
| 1896 | /* if TUse is 0, there was never a load, and no loss, too */ |
| 1897 | e1 = CLG_(register_eventtype)("AcCost1"); |
| 1898 | CLG_(add_eventtype)(use, e1); |
| 1899 | e1 = CLG_(register_eventtype)("SpLoss1"); |
| 1900 | CLG_(add_eventtype)(use, e1); |
| 1901 | e1 = CLG_(register_eventtype)("AcCost2"); |
| 1902 | CLG_(add_eventtype)(use, e1); |
| 1903 | e1 = CLG_(register_eventtype)("SpLoss2"); |
| 1904 | CLG_(add_eventtype)(use, e1); |
| 1905 | } |
| 1906 | |
| 1907 | Ir = CLG_(get_eventset)("Ir", 4); |
| 1908 | Dr = CLG_(get_eventset)("Dr", 4); |
| 1909 | Dw = CLG_(get_eventset)("Dw", 4); |
| 1910 | if (CLG_(clo).simulate_cache) { |
| 1911 | e1 = CLG_(register_eventtype)("Ir"); |
| 1912 | e2 = CLG_(register_eventtype)("I1mr"); |
| 1913 | e3 = CLG_(register_eventtype)("I2mr"); |
| 1914 | if (clo_simulate_writeback) { |
| 1915 | e4 = CLG_(register_eventtype)("I2dmr"); |
| 1916 | CLG_(add_dep_event4)(Ir, e1,e2,e3,e4); |
| 1917 | } |
| 1918 | else |
| 1919 | CLG_(add_dep_event3)(Ir, e1,e2,e3); |
| 1920 | |
| 1921 | e1 = CLG_(register_eventtype)("Dr"); |
| 1922 | e2 = CLG_(register_eventtype)("D1mr"); |
| 1923 | e3 = CLG_(register_eventtype)("D2mr"); |
| 1924 | if (clo_simulate_writeback) { |
| 1925 | e4 = CLG_(register_eventtype)("D2dmr"); |
| 1926 | CLG_(add_dep_event4)(Dr, e1,e2,e3,e4); |
| 1927 | } |
| 1928 | else |
| 1929 | CLG_(add_dep_event3)(Dr, e1,e2,e3); |
| 1930 | |
| 1931 | e1 = CLG_(register_eventtype)("Dw"); |
| 1932 | e2 = CLG_(register_eventtype)("D1mw"); |
| 1933 | e3 = CLG_(register_eventtype)("D2mw"); |
| 1934 | if (clo_simulate_writeback) { |
| 1935 | e4 = CLG_(register_eventtype)("D2dmw"); |
| 1936 | CLG_(add_dep_event4)(Dw, e1,e2,e3,e4); |
| 1937 | } |
| 1938 | else |
| 1939 | CLG_(add_dep_event3)(Dw, e1,e2,e3); |
| 1940 | |
| 1941 | } |
| 1942 | else { |
| 1943 | e1 = CLG_(register_eventtype)("Ir"); |
| 1944 | CLG_(add_eventtype)(Ir, e1); |
| 1945 | } |
| 1946 | |
| 1947 | sizeOfUseIr = use->size + Ir->size; |
| 1948 | D0 = CLG_(get_eventset)("D0", sizeOfUseIr); |
| 1949 | CLG_(add_eventset)(D0, use); |
| 1950 | off_D0_Ir = CLG_(add_eventset)(D0, Ir); |
| 1951 | |
| 1952 | D1r = CLG_(get_eventset)("D1r", sizeOfUseIr + Dr->size); |
| 1953 | CLG_(add_eventset)(D1r, use); |
| 1954 | off_D1r_Ir = CLG_(add_eventset)(D1r, Ir); |
| 1955 | off_D1r_Dr = CLG_(add_eventset)(D1r, Dr); |
| 1956 | |
| 1957 | D1w = CLG_(get_eventset)("D1w", sizeOfUseIr + Dw->size); |
| 1958 | CLG_(add_eventset)(D1w, use); |
| 1959 | off_D1w_Ir = CLG_(add_eventset)(D1w, Ir); |
| 1960 | off_D1w_Dw = CLG_(add_eventset)(D1w, Dw); |
| 1961 | |
| 1962 | D2 = CLG_(get_eventset)("D2", sizeOfUseIr + Dr->size + Dw->size); |
| 1963 | CLG_(add_eventset)(D2, use); |
| 1964 | off_D2_Ir = CLG_(add_eventset)(D2, Ir); |
| 1965 | off_D2_Dr = CLG_(add_eventset)(D2, Dr); |
| 1966 | off_D2_Dw = CLG_(add_eventset)(D2, Dw); |
| 1967 | |
| 1968 | sim = CLG_(get_eventset)("sim", sizeOfUseIr + Dr->size + Dw->size); |
| 1969 | CLG_(add_eventset)(sim, use); |
| 1970 | CLG_(sets).off_sim_Ir = CLG_(add_eventset)(sim, Ir); |
| 1971 | CLG_(sets).off_sim_Dr = CLG_(add_eventset)(sim, Dr); |
| 1972 | CLG_(sets).off_sim_Dw = CLG_(add_eventset)(sim, Dw); |
| 1973 | |
| 1974 | if (CLG_(clo).collect_alloc) max_user += 2; |
| 1975 | if (CLG_(clo).collect_systime) max_user += 2; |
| 1976 | |
| 1977 | full = CLG_(get_eventset)("full", sim->size + max_user); |
| 1978 | CLG_(add_eventset)(full, sim); |
| 1979 | CLG_(sets).off_full_Ir = CLG_(sets).off_sim_Ir; |
| 1980 | CLG_(sets).off_full_Dr = CLG_(sets).off_sim_Dr; |
| 1981 | CLG_(sets).off_full_Dw = CLG_(sets).off_sim_Dw; |
| 1982 | |
| 1983 | CLG_(sets).use = use; |
| 1984 | CLG_(sets).Ir = Ir; |
| 1985 | CLG_(sets).Dr = Dr; |
| 1986 | CLG_(sets).Dw = Dw; |
| 1987 | |
| 1988 | CLG_(sets).D0 = D0; |
| 1989 | CLG_(sets).D1r = D1r; |
| 1990 | CLG_(sets).D1w = D1w; |
| 1991 | CLG_(sets).D2 = D2; |
| 1992 | |
| 1993 | CLG_(sets).sim = sim; |
| 1994 | CLG_(sets).full = full; |
| 1995 | |
| 1996 | if (CLG_(clo).collect_alloc) { |
| 1997 | e1 = CLG_(register_eventtype)("allocCount"); |
| 1998 | e2 = CLG_(register_eventtype)("allocSize"); |
| 1999 | CLG_(sets).off_full_user = CLG_(add_dep_event2)(full, e1,e2); |
| 2000 | } |
| 2001 | |
| 2002 | if (CLG_(clo).collect_systime) { |
| 2003 | e1 = CLG_(register_eventtype)("sysCount"); |
| 2004 | e2 = CLG_(register_eventtype)("sysTime"); |
| 2005 | CLG_(sets).off_full_systime = CLG_(add_dep_event2)(full, e1,e2); |
| 2006 | } |
| 2007 | |
| 2008 | CLG_DEBUGIF(1) { |
| 2009 | CLG_DEBUG(1, "EventSets:\n"); |
| 2010 | CLG_(print_eventset)(-2, use); |
| 2011 | CLG_(print_eventset)(-2, Ir); |
| 2012 | CLG_(print_eventset)(-2, Dr); |
| 2013 | CLG_(print_eventset)(-2, Dw); |
| 2014 | CLG_(print_eventset)(-2, sim); |
| 2015 | CLG_(print_eventset)(-2, full); |
| 2016 | } |
| 2017 | |
| 2018 | /* Not-existing events are silently ignored */ |
| 2019 | CLG_(dumpmap) = CLG_(get_eventmapping)(full); |
| 2020 | CLG_(append_event)(CLG_(dumpmap), "Ir"); |
| 2021 | CLG_(append_event)(CLG_(dumpmap), "Dr"); |
| 2022 | CLG_(append_event)(CLG_(dumpmap), "Dw"); |
| 2023 | CLG_(append_event)(CLG_(dumpmap), "I1mr"); |
| 2024 | CLG_(append_event)(CLG_(dumpmap), "D1mr"); |
| 2025 | CLG_(append_event)(CLG_(dumpmap), "D1mw"); |
| 2026 | CLG_(append_event)(CLG_(dumpmap), "I2mr"); |
| 2027 | CLG_(append_event)(CLG_(dumpmap), "D2mr"); |
| 2028 | CLG_(append_event)(CLG_(dumpmap), "D2mw"); |
| 2029 | CLG_(append_event)(CLG_(dumpmap), "I2dmr"); |
| 2030 | CLG_(append_event)(CLG_(dumpmap), "D2dmr"); |
| 2031 | CLG_(append_event)(CLG_(dumpmap), "D2dmw"); |
| 2032 | CLG_(append_event)(CLG_(dumpmap), "AcCost1"); |
| 2033 | CLG_(append_event)(CLG_(dumpmap), "SpLoss1"); |
| 2034 | CLG_(append_event)(CLG_(dumpmap), "AcCost2"); |
| 2035 | CLG_(append_event)(CLG_(dumpmap), "SpLoss2"); |
| 2036 | CLG_(append_event)(CLG_(dumpmap), "allocCount"); |
| 2037 | CLG_(append_event)(CLG_(dumpmap), "allocSize"); |
| 2038 | CLG_(append_event)(CLG_(dumpmap), "sysCount"); |
| 2039 | CLG_(append_event)(CLG_(dumpmap), "sysTime"); |
| 2040 | |
| 2041 | } |
| 2042 | |
| 2043 | |
| 2044 | |
| 2045 | static |
| 2046 | void add_and_zero_Dx(EventSet* es, SimCost dst, ULong* cost) |
| 2047 | { |
| 2048 | /* if eventset use is defined, it is always first (hardcoded!) */ |
| 2049 | CLG_(add_and_zero_cost)( CLG_(sets).use, dst, cost); |
| 2050 | |
| 2051 | /* FIXME: This is hardcoded... */ |
| 2052 | if (es == CLG_(sets).D0) { |
| 2053 | CLG_(add_and_zero_cost)( CLG_(sets).Ir, dst + CLG_(sets).off_sim_Ir, |
| 2054 | cost + off_D0_Ir); |
| 2055 | } |
| 2056 | else if (es == CLG_(sets).D1r) { |
| 2057 | CLG_(add_and_zero_cost)( CLG_(sets).Ir, dst + CLG_(sets).off_sim_Ir, |
| 2058 | cost + off_D1r_Ir); |
| 2059 | CLG_(add_and_zero_cost)( CLG_(sets).Dr, dst + CLG_(sets).off_sim_Dr, |
| 2060 | cost + off_D1r_Dr); |
| 2061 | } |
| 2062 | else if (es == CLG_(sets).D1w) { |
| 2063 | CLG_(add_and_zero_cost)( CLG_(sets).Ir, dst + CLG_(sets).off_sim_Ir, |
| 2064 | cost + off_D1w_Ir); |
| 2065 | CLG_(add_and_zero_cost)( CLG_(sets).Dw, dst + CLG_(sets).off_sim_Dw, |
| 2066 | cost + off_D1w_Dw); |
| 2067 | } |
| 2068 | else { |
| 2069 | CLG_ASSERT(es == CLG_(sets).D2); |
| 2070 | CLG_(add_and_zero_cost)( CLG_(sets).Ir, dst + CLG_(sets).off_sim_Ir, |
| 2071 | cost + off_D2_Ir); |
| 2072 | CLG_(add_and_zero_cost)( CLG_(sets).Dr, dst + CLG_(sets).off_sim_Dr, |
| 2073 | cost + off_D2_Dr); |
| 2074 | CLG_(add_and_zero_cost)( CLG_(sets).Dw, dst + CLG_(sets).off_sim_Dw, |
| 2075 | cost + off_D2_Dw); |
| 2076 | } |
| 2077 | } |
| 2078 | |
| 2079 | /* this is called at dump time for every instruction executed */ |
| 2080 | static void cachesim_add_icost(SimCost cost, BBCC* bbcc, |
| 2081 | InstrInfo* ii, ULong exe_count) |
| 2082 | { |
| 2083 | if (!CLG_(clo).simulate_cache) |
| 2084 | cost[CLG_(sets).off_sim_Ir] += exe_count; |
| 2085 | else { |
| 2086 | |
| 2087 | #if 0 |
| 2088 | /* There is always a trivial case where exe_count and Ir can be |
| 2089 | * slightly different because ecounter is updated when executing |
| 2090 | * the next BB. E.g. for last BB executed, or when toggling collection |
| 2091 | */ |
| 2092 | /* FIXME: Hardcoded that each eventset has Ir as first */ |
| 2093 | if ((bbcc->cost + ii->cost_offset)[0] != exe_count) { |
| 2094 | VG_(printf)("==> Ir %llu, exe %llu\n", |
| 2095 | (bbcc->cost + ii->cost_offset)[0], exe_count); |
| 2096 | CLG_(print_bbcc_cost)(-2, bbcc); |
| 2097 | //CLG_ASSERT((bbcc->cost + ii->cost_offset)[0] == exe_count); |
| 2098 | } |
| 2099 | #endif |
| 2100 | |
| 2101 | add_and_zero_Dx(ii->eventset, cost, |
| 2102 | bbcc->cost + ii->cost_offset); |
| 2103 | } |
| 2104 | } |
| 2105 | |
| 2106 | static |
| 2107 | void cachesim_after_bbsetup(void) |
| 2108 | { |
| 2109 | BBCC* bbcc = CLG_(current_state).bbcc; |
| 2110 | |
| 2111 | if (CLG_(clo).simulate_cache) { |
| 2112 | BB* bb = bbcc->bb; |
| 2113 | |
| 2114 | /* only needed if log_* functions are called */ |
| 2115 | bb_base = bb->obj->offset + bb->offset; |
| 2116 | cost_base = bbcc->cost; |
| 2117 | } |
| 2118 | } |
| 2119 | |
| 2120 | static |
| 2121 | void cachesim_finish(void) |
| 2122 | { |
| 2123 | if (clo_collect_cacheuse) |
| 2124 | cacheuse_finish(); |
| 2125 | } |
| 2126 | |
| 2127 | /*------------------------------------------------------------*/ |
| 2128 | /*--- The simulator defined in this file ---*/ |
| 2129 | /*------------------------------------------------------------*/ |
| 2130 | |
| 2131 | struct cachesim_if CLG_(cachesim) = { |
| 2132 | .print_opts = cachesim_print_opts, |
| 2133 | .parse_opt = cachesim_parse_opt, |
| 2134 | .post_clo_init = cachesim_post_clo_init, |
| 2135 | .clear = cachesim_clear, |
| 2136 | .getdesc = cachesim_getdesc, |
| 2137 | .printstat = cachesim_printstat, |
| 2138 | .add_icost = cachesim_add_icost, |
| 2139 | .after_bbsetup = cachesim_after_bbsetup, |
| 2140 | .finish = cachesim_finish, |
| 2141 | |
| 2142 | /* these will be set by cachesim_post_clo_init */ |
| 2143 | .log_1I0D = 0, |
| 2144 | |
| 2145 | .log_1I1Dr = 0, |
| 2146 | .log_1I1Dw = 0, |
| 2147 | .log_1I2D = 0, |
| 2148 | |
| 2149 | .log_0I1Dr = 0, |
| 2150 | .log_0I1Dw = 0, |
| 2151 | .log_0I2D = 0, |
| 2152 | |
| 2153 | .log_1I0D_name = "(no function)", |
| 2154 | |
| 2155 | .log_1I1Dr_name = "(no function)", |
| 2156 | .log_1I1Dw_name = "(no function)", |
| 2157 | .log_1I2D_name = "(no function)", |
| 2158 | |
| 2159 | .log_0I1Dr_name = "(no function)", |
| 2160 | .log_0I1Dw_name = "(no function)", |
| 2161 | .log_0I2D_name = "(no function)" |
| 2162 | }; |
| 2163 | |
| 2164 | |
| 2165 | /*--------------------------------------------------------------------*/ |
| 2166 | /*--- end ct_sim.c ---*/ |
| 2167 | /*--------------------------------------------------------------------*/ |
| 2168 | |