sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 1 | |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 2 | /* Tests in detail the core arithmetic for pcmp{e,i}str{i,m} using |
| 3 | pcmpistri to drive it. Does not check the e-vs-i or i-vs-m |
| 4 | aspect. */ |
| 5 | |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 6 | #include <string.h> |
| 7 | #include <stdio.h> |
| 8 | #include <assert.h> |
| 9 | |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 10 | typedef unsigned int UInt; |
| 11 | typedef signed int Int; |
| 12 | typedef unsigned char UChar; |
sewardj | 15df336 | 2012-09-17 13:43:57 +0000 | [diff] [blame] | 13 | typedef signed char Char; |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 14 | typedef unsigned long long int ULong; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 15 | typedef UChar Bool; |
| 16 | #define False ((Bool)0) |
| 17 | #define True ((Bool)1) |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 18 | |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 19 | //typedef unsigned char V128[16]; |
| 20 | typedef |
| 21 | union { |
| 22 | UChar uChar[16]; |
| 23 | UInt uInt[4]; |
| 24 | } |
| 25 | V128; |
| 26 | |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 27 | #define SHIFT_O 11 |
| 28 | #define SHIFT_S 7 |
| 29 | #define SHIFT_Z 6 |
| 30 | #define SHIFT_A 4 |
| 31 | #define SHIFT_C 0 |
| 32 | #define SHIFT_P 2 |
| 33 | |
| 34 | #define MASK_O (1ULL << SHIFT_O) |
| 35 | #define MASK_S (1ULL << SHIFT_S) |
| 36 | #define MASK_Z (1ULL << SHIFT_Z) |
| 37 | #define MASK_A (1ULL << SHIFT_A) |
| 38 | #define MASK_C (1ULL << SHIFT_C) |
| 39 | #define MASK_P (1ULL << SHIFT_P) |
| 40 | |
| 41 | |
| 42 | UInt clz32 ( UInt x ) |
| 43 | { |
| 44 | Int y, m, n; |
| 45 | y = -(x >> 16); |
| 46 | m = (y >> 16) & 16; |
| 47 | n = 16 - m; |
| 48 | x = x >> m; |
| 49 | y = x - 0x100; |
| 50 | m = (y >> 16) & 8; |
| 51 | n = n + m; |
| 52 | x = x << m; |
| 53 | y = x - 0x1000; |
| 54 | m = (y >> 16) & 4; |
| 55 | n = n + m; |
| 56 | x = x << m; |
| 57 | y = x - 0x4000; |
| 58 | m = (y >> 16) & 2; |
| 59 | n = n + m; |
| 60 | x = x << m; |
| 61 | y = x >> 14; |
| 62 | m = y & ~(y >> 1); |
| 63 | return n + 2 - m; |
| 64 | } |
| 65 | |
| 66 | UInt ctz32 ( UInt x ) |
| 67 | { |
| 68 | return 32 - clz32((~x) & (x-1)); |
| 69 | } |
| 70 | |
| 71 | void expand ( V128* dst, char* summary ) |
| 72 | { |
| 73 | Int i; |
| 74 | assert( strlen(summary) == 16 ); |
| 75 | for (i = 0; i < 16; i++) { |
| 76 | UChar xx = 0; |
| 77 | UChar x = summary[15-i]; |
| 78 | if (x >= '0' && x <= '9') { xx = x - '0'; } |
| 79 | else if (x >= 'A' && x <= 'F') { xx = x - 'A' + 10; } |
| 80 | else if (x >= 'a' && x <= 'f') { xx = x - 'a' + 10; } |
| 81 | else assert(0); |
| 82 | |
| 83 | assert(xx < 16); |
| 84 | xx = (xx << 4) | xx; |
| 85 | assert(xx < 256); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 86 | dst->uChar[i] = xx; |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 87 | } |
| 88 | } |
| 89 | |
| 90 | void try_istri ( char* which, |
| 91 | UInt(*h_fn)(V128*,V128*), |
| 92 | UInt(*s_fn)(V128*,V128*), |
| 93 | char* summL, char* summR ) |
| 94 | { |
| 95 | assert(strlen(which) == 2); |
| 96 | V128 argL, argR; |
| 97 | expand(&argL, summL); |
| 98 | expand(&argR, summR); |
| 99 | UInt h_res = h_fn(&argL, &argR); |
| 100 | UInt s_res = s_fn(&argL, &argR); |
| 101 | printf("istri %s %s %s -> %08x %08x %s\n", |
| 102 | which, summL, summR, h_res, s_res, h_res == s_res ? "" : "!!!!"); |
| 103 | } |
| 104 | |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 105 | UInt zmask_from_V128 ( V128* arg ) |
| 106 | { |
| 107 | UInt i, res = 0; |
| 108 | for (i = 0; i < 16; i++) { |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 109 | res |= ((arg->uChar[i] == 0) ? 1 : 0) << i; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 110 | } |
| 111 | return res; |
| 112 | } |
| 113 | |
| 114 | ////////////////////////////////////////////////////////// |
| 115 | // // |
| 116 | // GENERAL // |
| 117 | // // |
| 118 | ////////////////////////////////////////////////////////// |
| 119 | |
| 120 | |
| 121 | /* Given partial results from a pcmpXstrX operation (intRes1, |
| 122 | basically), generate an I format (index value for ECX) output, and |
| 123 | also the new OSZACP flags. |
| 124 | */ |
| 125 | static |
| 126 | void pcmpXstrX_WRK_gen_output_fmt_I(/*OUT*/V128* resV, |
| 127 | /*OUT*/UInt* resOSZACP, |
| 128 | UInt intRes1, |
| 129 | UInt zmaskL, UInt zmaskR, |
| 130 | UInt validL, |
| 131 | UInt pol, UInt idx ) |
| 132 | { |
| 133 | assert((pol >> 2) == 0); |
| 134 | assert((idx >> 1) == 0); |
| 135 | |
| 136 | UInt intRes2 = 0; |
| 137 | switch (pol) { |
| 138 | case 0: intRes2 = intRes1; break; // pol + |
| 139 | case 1: intRes2 = ~intRes1; break; // pol - |
| 140 | case 2: intRes2 = intRes1; break; // pol m+ |
| 141 | case 3: intRes2 = intRes1 ^ validL; break; // pol m- |
| 142 | } |
| 143 | intRes2 &= 0xFFFF; |
| 144 | |
| 145 | // generate ecx value |
| 146 | UInt newECX = 0; |
| 147 | if (idx) { |
| 148 | // index of ms-1-bit |
| 149 | newECX = intRes2 == 0 ? 16 : (31 - clz32(intRes2)); |
| 150 | } else { |
| 151 | // index of ls-1-bit |
| 152 | newECX = intRes2 == 0 ? 16 : ctz32(intRes2); |
| 153 | } |
| 154 | |
| 155 | *(UInt*)(&resV[0]) = newECX; |
| 156 | |
| 157 | // generate new flags, common to all ISTRI and ISTRM cases |
| 158 | *resOSZACP // A, P are zero |
| 159 | = ((intRes2 == 0) ? 0 : MASK_C) // C == 0 iff intRes2 == 0 |
| 160 | | ((zmaskL == 0) ? 0 : MASK_Z) // Z == 1 iff any in argL is 0 |
| 161 | | ((zmaskR == 0) ? 0 : MASK_S) // S == 1 iff any in argR is 0 |
| 162 | | ((intRes2 & 1) << SHIFT_O); // O == IntRes2[0] |
| 163 | } |
| 164 | |
| 165 | |
| 166 | /* Compute result and new OSZACP flags for all PCMP{E,I}STR{I,M} |
| 167 | variants. |
| 168 | |
| 169 | For xSTRI variants, the new ECX value is placed in the 32 bits |
| 170 | pointed to by *resV. For xSTRM variants, the result is a 128 bit |
| 171 | value and is placed at *resV in the obvious way. |
| 172 | |
| 173 | For all variants, the new OSZACP value is placed at *resOSZACP. |
| 174 | |
| 175 | argLV and argRV are the vector args. The caller must prepare a |
| 176 | 16-bit mask for each, zmaskL and zmaskR. For ISTRx variants this |
| 177 | must be 1 for each zero byte of of the respective arg. For ESTRx |
| 178 | variants this is derived from the explicit length indication, and |
| 179 | must be 0 in all places except at the bit index corresponding to |
| 180 | the valid length (0 .. 16). If the valid length is 16 then the |
| 181 | mask must be all zeroes. In all cases, bits 31:16 must be zero. |
| 182 | |
| 183 | imm8 is the original immediate from the instruction. isSTRM |
| 184 | indicates whether this is a xSTRM or xSTRI variant, which controls |
| 185 | how much of *res is written. |
| 186 | |
| 187 | If the given imm8 case can be handled, the return value is True. |
| 188 | If not, False is returned, and neither *res not *resOSZACP are |
| 189 | altered. |
| 190 | */ |
| 191 | |
| 192 | Bool pcmpXstrX_WRK ( /*OUT*/V128* resV, |
| 193 | /*OUT*/UInt* resOSZACP, |
| 194 | V128* argLV, V128* argRV, |
| 195 | UInt zmaskL, UInt zmaskR, |
| 196 | UInt imm8, Bool isSTRM ) |
| 197 | { |
| 198 | assert(imm8 < 0x80); |
| 199 | assert((zmaskL >> 16) == 0); |
| 200 | assert((zmaskR >> 16) == 0); |
| 201 | |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 202 | /* Explicitly reject any imm8 values that haven't been validated, |
| 203 | even if they would probably work. Life is too short to have |
| 204 | unvalidated cases in the code base. */ |
| 205 | switch (imm8) { |
sewardj | a2bf389 | 2014-05-21 14:43:11 +0000 | [diff] [blame] | 206 | case 0x00: case 0x02: case 0x08: case 0x0A: case 0x0C: case 0x0E: |
sewardj | 053f436 | 2015-08-13 14:19:48 +0000 | [diff] [blame] | 207 | case 0x12: case 0x14: case 0x18: case 0x1A: |
sewardj | a2bf389 | 2014-05-21 14:43:11 +0000 | [diff] [blame] | 208 | case 0x30: case 0x34: case 0x38: case 0x3A: |
sewardj | 7f0dfdd | 2015-08-13 13:47:02 +0000 | [diff] [blame] | 209 | case 0x40: case 0x42: case 0x44: case 0x46: case 0x4A: |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 210 | break; |
| 211 | default: |
| 212 | return False; |
| 213 | } |
| 214 | |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 215 | UInt fmt = (imm8 >> 0) & 3; // imm8[1:0] data format |
| 216 | UInt agg = (imm8 >> 2) & 3; // imm8[3:2] aggregation fn |
| 217 | UInt pol = (imm8 >> 4) & 3; // imm8[5:4] polarity |
| 218 | UInt idx = (imm8 >> 6) & 1; // imm8[6] 1==msb/bytemask |
| 219 | |
| 220 | /*----------------------------------------*/ |
| 221 | /*-- strcmp on byte data --*/ |
| 222 | /*----------------------------------------*/ |
| 223 | |
| 224 | if (agg == 2/*equal each, aka strcmp*/ |
| 225 | && (fmt == 0/*ub*/ || fmt == 2/*sb*/) |
| 226 | && !isSTRM) { |
| 227 | Int i; |
| 228 | UChar* argL = (UChar*)argLV; |
| 229 | UChar* argR = (UChar*)argRV; |
| 230 | UInt boolResII = 0; |
| 231 | for (i = 15; i >= 0; i--) { |
| 232 | UChar cL = argL[i]; |
| 233 | UChar cR = argR[i]; |
| 234 | boolResII = (boolResII << 1) | (cL == cR ? 1 : 0); |
| 235 | } |
| 236 | UInt validL = ~(zmaskL | -zmaskL); // not(left(zmaskL)) |
| 237 | UInt validR = ~(zmaskR | -zmaskR); // not(left(zmaskR)) |
| 238 | |
| 239 | // do invalidation, common to all equal-each cases |
| 240 | UInt intRes1 |
| 241 | = (boolResII & validL & validR) // if both valid, use cmpres |
| 242 | | (~ (validL | validR)); // if both invalid, force 1 |
| 243 | // else force 0 |
| 244 | intRes1 &= 0xFFFF; |
| 245 | |
| 246 | // generate I-format output |
| 247 | pcmpXstrX_WRK_gen_output_fmt_I( |
| 248 | resV, resOSZACP, |
| 249 | intRes1, zmaskL, zmaskR, validL, pol, idx |
| 250 | ); |
| 251 | |
| 252 | return True; |
| 253 | } |
| 254 | |
| 255 | /*----------------------------------------*/ |
| 256 | /*-- set membership on byte data --*/ |
| 257 | /*----------------------------------------*/ |
| 258 | |
| 259 | if (agg == 0/*equal any, aka find chars in a set*/ |
| 260 | && (fmt == 0/*ub*/ || fmt == 2/*sb*/) |
| 261 | && !isSTRM) { |
| 262 | /* argL: the string, argR: charset */ |
| 263 | UInt si, ci; |
| 264 | UChar* argL = (UChar*)argLV; |
| 265 | UChar* argR = (UChar*)argRV; |
| 266 | UInt boolRes = 0; |
| 267 | UInt validL = ~(zmaskL | -zmaskL); // not(left(zmaskL)) |
| 268 | UInt validR = ~(zmaskR | -zmaskR); // not(left(zmaskR)) |
| 269 | |
| 270 | for (si = 0; si < 16; si++) { |
| 271 | if ((validL & (1 << si)) == 0) |
| 272 | // run off the end of the string. |
| 273 | break; |
| 274 | UInt m = 0; |
| 275 | for (ci = 0; ci < 16; ci++) { |
| 276 | if ((validR & (1 << ci)) == 0) break; |
| 277 | if (argR[ci] == argL[si]) { m = 1; break; } |
| 278 | } |
| 279 | boolRes |= (m << si); |
| 280 | } |
| 281 | |
| 282 | // boolRes is "pre-invalidated" |
| 283 | UInt intRes1 = boolRes & 0xFFFF; |
| 284 | |
| 285 | // generate I-format output |
| 286 | pcmpXstrX_WRK_gen_output_fmt_I( |
| 287 | resV, resOSZACP, |
| 288 | intRes1, zmaskL, zmaskR, validL, pol, idx |
| 289 | ); |
| 290 | |
| 291 | return True; |
| 292 | } |
| 293 | |
| 294 | /*----------------------------------------*/ |
| 295 | /*-- substring search on byte data --*/ |
| 296 | /*----------------------------------------*/ |
| 297 | |
| 298 | if (agg == 3/*equal ordered, aka substring search*/ |
| 299 | && (fmt == 0/*ub*/ || fmt == 2/*sb*/) |
| 300 | && !isSTRM) { |
| 301 | |
| 302 | /* argL: haystack, argR: needle */ |
| 303 | UInt ni, hi; |
| 304 | UChar* argL = (UChar*)argLV; |
| 305 | UChar* argR = (UChar*)argRV; |
| 306 | UInt boolRes = 0; |
| 307 | UInt validL = ~(zmaskL | -zmaskL); // not(left(zmaskL)) |
| 308 | UInt validR = ~(zmaskR | -zmaskR); // not(left(zmaskR)) |
| 309 | for (hi = 0; hi < 16; hi++) { |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 310 | UInt m = 1; |
| 311 | for (ni = 0; ni < 16; ni++) { |
| 312 | if ((validR & (1 << ni)) == 0) break; |
| 313 | UInt i = ni + hi; |
| 314 | if (i >= 16) break; |
| 315 | if (argL[i] != argR[ni]) { m = 0; break; } |
| 316 | } |
| 317 | boolRes |= (m << hi); |
weidendo | c5274ae | 2012-07-25 09:49:26 +0000 | [diff] [blame] | 318 | if ((validL & (1 << hi)) == 0) |
| 319 | // run off the end of the haystack |
| 320 | break; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 321 | } |
| 322 | |
| 323 | // boolRes is "pre-invalidated" |
| 324 | UInt intRes1 = boolRes & 0xFFFF; |
| 325 | |
| 326 | // generate I-format output |
| 327 | pcmpXstrX_WRK_gen_output_fmt_I( |
| 328 | resV, resOSZACP, |
| 329 | intRes1, zmaskL, zmaskR, validL, pol, idx |
| 330 | ); |
| 331 | |
| 332 | return True; |
| 333 | } |
| 334 | |
| 335 | /*----------------------------------------*/ |
| 336 | /*-- ranges, unsigned byte data --*/ |
| 337 | /*----------------------------------------*/ |
| 338 | |
| 339 | if (agg == 1/*ranges*/ |
| 340 | && fmt == 0/*ub*/ |
| 341 | && !isSTRM) { |
| 342 | |
| 343 | /* argL: string, argR: range-pairs */ |
| 344 | UInt ri, si; |
| 345 | UChar* argL = (UChar*)argLV; |
| 346 | UChar* argR = (UChar*)argRV; |
| 347 | UInt boolRes = 0; |
| 348 | UInt validL = ~(zmaskL | -zmaskL); // not(left(zmaskL)) |
| 349 | UInt validR = ~(zmaskR | -zmaskR); // not(left(zmaskR)) |
| 350 | for (si = 0; si < 16; si++) { |
| 351 | if ((validL & (1 << si)) == 0) |
| 352 | // run off the end of the string |
| 353 | break; |
| 354 | UInt m = 0; |
| 355 | for (ri = 0; ri < 16; ri += 2) { |
| 356 | if ((validR & (3 << ri)) != (3 << ri)) break; |
| 357 | if (argR[ri] <= argL[si] && argL[si] <= argR[ri+1]) { |
| 358 | m = 1; break; |
| 359 | } |
| 360 | } |
| 361 | boolRes |= (m << si); |
| 362 | } |
| 363 | |
| 364 | // boolRes is "pre-invalidated" |
| 365 | UInt intRes1 = boolRes & 0xFFFF; |
| 366 | |
| 367 | // generate I-format output |
| 368 | pcmpXstrX_WRK_gen_output_fmt_I( |
| 369 | resV, resOSZACP, |
| 370 | intRes1, zmaskL, zmaskR, validL, pol, idx |
| 371 | ); |
| 372 | |
| 373 | return True; |
| 374 | } |
| 375 | |
sewardj | 15df336 | 2012-09-17 13:43:57 +0000 | [diff] [blame] | 376 | /*----------------------------------------*/ |
| 377 | /*-- ranges, signed byte data --*/ |
| 378 | /*----------------------------------------*/ |
| 379 | |
| 380 | if (agg == 1/*ranges*/ |
| 381 | && fmt == 2/*sb*/ |
| 382 | && !isSTRM) { |
| 383 | |
| 384 | /* argL: string, argR: range-pairs */ |
| 385 | UInt ri, si; |
| 386 | Char* argL = (Char*)argLV; |
| 387 | Char* argR = (Char*)argRV; |
| 388 | UInt boolRes = 0; |
| 389 | UInt validL = ~(zmaskL | -zmaskL); // not(left(zmaskL)) |
| 390 | UInt validR = ~(zmaskR | -zmaskR); // not(left(zmaskR)) |
| 391 | for (si = 0; si < 16; si++) { |
| 392 | if ((validL & (1 << si)) == 0) |
| 393 | // run off the end of the string |
| 394 | break; |
| 395 | UInt m = 0; |
| 396 | for (ri = 0; ri < 16; ri += 2) { |
| 397 | if ((validR & (3 << ri)) != (3 << ri)) break; |
| 398 | if (argR[ri] <= argL[si] && argL[si] <= argR[ri+1]) { |
| 399 | m = 1; break; |
| 400 | } |
| 401 | } |
| 402 | boolRes |= (m << si); |
| 403 | } |
| 404 | |
| 405 | // boolRes is "pre-invalidated" |
| 406 | UInt intRes1 = boolRes & 0xFFFF; |
| 407 | |
| 408 | // generate I-format output |
| 409 | pcmpXstrX_WRK_gen_output_fmt_I( |
| 410 | resV, resOSZACP, |
| 411 | intRes1, zmaskL, zmaskR, validL, pol, idx |
| 412 | ); |
| 413 | |
| 414 | return True; |
| 415 | } |
| 416 | |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 417 | return False; |
| 418 | } |
| 419 | |
| 420 | |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 421 | ////////////////////////////////////////////////////////// |
| 422 | // // |
| 423 | // ISTRI_4A // |
| 424 | // // |
| 425 | ////////////////////////////////////////////////////////// |
| 426 | |
| 427 | UInt h_pcmpistri_4A ( V128* argL, V128* argR ) |
| 428 | { |
| 429 | V128 block[2]; |
| 430 | memcpy(&block[0], argL, sizeof(V128)); |
| 431 | memcpy(&block[1], argR, sizeof(V128)); |
| 432 | ULong res, flags; |
| 433 | __asm__ __volatile__( |
| 434 | "subq $1024, %%rsp" "\n\t" |
| 435 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 436 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 437 | "pcmpistri $0x4A, %%xmm2, %%xmm11" "\n\t" |
| 438 | "pushfq" "\n\t" |
| 439 | "popq %%rdx" "\n\t" |
| 440 | "movq %%rcx, %0" "\n\t" |
| 441 | "movq %%rdx, %1" "\n\t" |
| 442 | "addq $1024, %%rsp" "\n\t" |
| 443 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 444 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 445 | ); |
| 446 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 447 | } |
| 448 | |
| 449 | UInt s_pcmpistri_4A ( V128* argLU, V128* argRU ) |
| 450 | { |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 451 | V128 resV; |
| 452 | UInt resOSZACP, resECX; |
| 453 | Bool ok |
| 454 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 455 | zmask_from_V128(argLU), |
| 456 | zmask_from_V128(argRU), |
| 457 | 0x4A, False/*!isSTRM*/ |
| 458 | ); |
| 459 | assert(ok); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 460 | resECX = resV.uInt[0]; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 461 | return (resOSZACP << 16) | resECX; |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 462 | } |
| 463 | |
| 464 | void istri_4A ( void ) |
| 465 | { |
| 466 | char* wot = "4A"; |
| 467 | UInt(*h)(V128*,V128*) = h_pcmpistri_4A; |
| 468 | UInt(*s)(V128*,V128*) = s_pcmpistri_4A; |
| 469 | |
| 470 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 471 | |
| 472 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 473 | try_istri(wot,h,s, "aaaa2aaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 474 | try_istri(wot,h,s, "aaaaaaaaa2aaaaaa", "aaaaaaaaaaaaaaaa"); |
| 475 | try_istri(wot,h,s, "aaaaaaaaaaaaa2aa", "aaaaaaaaaaaaaaaa"); |
| 476 | |
| 477 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaa2aaaaaaaaaaa"); |
| 478 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaa2aaaaaa"); |
| 479 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaa2a"); |
| 480 | |
| 481 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 482 | try_istri(wot,h,s, "baaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 483 | try_istri(wot,h,s, "b9aaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 484 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 485 | |
| 486 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 487 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaa7aaa"); |
| 488 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaa2aaa4aaa"); |
| 489 | |
| 490 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 491 | |
| 492 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 493 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 494 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaa0aaa"); |
| 495 | |
| 496 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 497 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 498 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 499 | |
| 500 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 501 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaa0aaaaaaa"); |
| 502 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaa0aaaaaaa"); |
| 503 | |
| 504 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaa0aaaaaaa"); |
| 505 | try_istri(wot,h,s, "8000000000000000", "aaaaaaaa0aaaaaaa"); |
| 506 | try_istri(wot,h,s, "0000000000000001", "aaaaaaaa0aaaaaaa"); |
| 507 | |
| 508 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaaaaaaaaaa"); |
| 509 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "0000000000000000"); |
| 510 | } |
| 511 | |
| 512 | ////////////////////////////////////////////////////////// |
| 513 | // // |
| 514 | // ISTRI_3A // |
| 515 | // // |
| 516 | ////////////////////////////////////////////////////////// |
| 517 | |
| 518 | UInt h_pcmpistri_3A ( V128* argL, V128* argR ) |
| 519 | { |
| 520 | V128 block[2]; |
| 521 | memcpy(&block[0], argL, sizeof(V128)); |
| 522 | memcpy(&block[1], argR, sizeof(V128)); |
| 523 | ULong res, flags; |
| 524 | __asm__ __volatile__( |
| 525 | "subq $1024, %%rsp" "\n\t" |
| 526 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 527 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 528 | "pcmpistri $0x3A, %%xmm2, %%xmm11" "\n\t" |
| 529 | "pushfq" "\n\t" |
| 530 | "popq %%rdx" "\n\t" |
| 531 | "movq %%rcx, %0" "\n\t" |
| 532 | "movq %%rdx, %1" "\n\t" |
| 533 | "addq $1024, %%rsp" "\n\t" |
| 534 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 535 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 536 | ); |
| 537 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 538 | } |
| 539 | |
| 540 | UInt s_pcmpistri_3A ( V128* argLU, V128* argRU ) |
| 541 | { |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 542 | V128 resV; |
| 543 | UInt resOSZACP, resECX; |
| 544 | Bool ok |
| 545 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 546 | zmask_from_V128(argLU), |
| 547 | zmask_from_V128(argRU), |
| 548 | 0x3A, False/*!isSTRM*/ |
| 549 | ); |
| 550 | assert(ok); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 551 | resECX = resV.uInt[0]; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 552 | return (resOSZACP << 16) | resECX; |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 553 | } |
| 554 | |
| 555 | void istri_3A ( void ) |
| 556 | { |
| 557 | char* wot = "3A"; |
| 558 | UInt(*h)(V128*,V128*) = h_pcmpistri_3A; |
| 559 | UInt(*s)(V128*,V128*) = s_pcmpistri_3A; |
| 560 | |
| 561 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 562 | |
| 563 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 564 | try_istri(wot,h,s, "aaaa2aaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 565 | try_istri(wot,h,s, "aaaaaaaaa2aaaaaa", "aaaaaaaaaaaaaaaa"); |
| 566 | try_istri(wot,h,s, "aaaaaaaaaaaaa2aa", "aaaaaaaaaaaaaaaa"); |
| 567 | |
| 568 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaa2aaaaaaaaaaa"); |
| 569 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaa2aaaaaa"); |
| 570 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaa2a"); |
| 571 | |
| 572 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 573 | try_istri(wot,h,s, "baaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 574 | try_istri(wot,h,s, "b9aaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 575 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 576 | |
| 577 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 578 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaa7aaa"); |
| 579 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaa2aaa4aaa"); |
| 580 | |
| 581 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 582 | |
| 583 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 584 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 585 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaa0aaa"); |
| 586 | |
| 587 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 588 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 589 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 590 | |
| 591 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 592 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaa0aaaaaaa"); |
| 593 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaa0aaaaaaa"); |
| 594 | |
| 595 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaa0aaaaaaa"); |
| 596 | try_istri(wot,h,s, "8000000000000000", "aaaaaaaa0aaaaaaa"); |
| 597 | try_istri(wot,h,s, "0000000000000001", "aaaaaaaa0aaaaaaa"); |
| 598 | |
| 599 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaaaaaaaaaa"); |
| 600 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "0000000000000000"); |
| 601 | } |
| 602 | |
| 603 | |
| 604 | |
| 605 | ////////////////////////////////////////////////////////// |
| 606 | // // |
| 607 | // ISTRI_0C // |
| 608 | // // |
| 609 | ////////////////////////////////////////////////////////// |
| 610 | |
| 611 | __attribute__((noinline)) |
| 612 | UInt h_pcmpistri_0C ( V128* argL, V128* argR ) |
| 613 | { |
| 614 | V128 block[2]; |
| 615 | memcpy(&block[0], argL, sizeof(V128)); |
| 616 | memcpy(&block[1], argR, sizeof(V128)); |
| 617 | ULong res = 0, flags = 0; |
| 618 | __asm__ __volatile__( |
weidendo | c5274ae | 2012-07-25 09:49:26 +0000 | [diff] [blame] | 619 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 620 | "movdqu 16(%2), %%xmm11" "\n\t" |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 621 | "pcmpistri $0x0C, %%xmm2, %%xmm11" "\n\t" |
| 622 | //"pcmpistrm $0x0C, %%xmm2, %%xmm11" "\n\t" |
| 623 | //"movd %%xmm0, %%ecx" "\n\t" |
| 624 | "pushfq" "\n\t" |
| 625 | "popq %%rdx" "\n\t" |
| 626 | "movq %%rcx, %0" "\n\t" |
| 627 | "movq %%rdx, %1" "\n\t" |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 628 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 629 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 630 | ); |
| 631 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 632 | } |
| 633 | |
| 634 | UInt s_pcmpistri_0C ( V128* argLU, V128* argRU ) |
| 635 | { |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 636 | V128 resV; |
| 637 | UInt resOSZACP, resECX; |
| 638 | Bool ok |
| 639 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 640 | zmask_from_V128(argLU), |
| 641 | zmask_from_V128(argRU), |
| 642 | 0x0C, False/*!isSTRM*/ |
| 643 | ); |
| 644 | assert(ok); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 645 | resECX = resV.uInt[0]; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 646 | return (resOSZACP << 16) | resECX; |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 647 | } |
| 648 | |
| 649 | void istri_0C ( void ) |
| 650 | { |
| 651 | char* wot = "0C"; |
| 652 | UInt(*h)(V128*,V128*) = h_pcmpistri_0C; |
| 653 | UInt(*s)(V128*,V128*) = s_pcmpistri_0C; |
| 654 | |
| 655 | try_istri(wot,h,s, "111111111abcde11", "00000000000abcde"); |
| 656 | |
| 657 | try_istri(wot,h,s, "111111111abcde11", "0000abcde00abcde"); |
| 658 | |
| 659 | try_istri(wot,h,s, "1111111111abcde1", "00000000000abcde"); |
| 660 | try_istri(wot,h,s, "11111111111abcde", "00000000000abcde"); |
| 661 | try_istri(wot,h,s, "111111111111abcd", "00000000000abcde"); |
| 662 | |
| 663 | try_istri(wot,h,s, "111abcde1abcde11", "00000000000abcde"); |
| 664 | |
| 665 | try_istri(wot,h,s, "11abcde11abcde11", "00000000000abcde"); |
| 666 | try_istri(wot,h,s, "1abcde111abcde11", "00000000000abcde"); |
| 667 | try_istri(wot,h,s, "abcde1111abcde11", "00000000000abcde"); |
| 668 | try_istri(wot,h,s, "bcde11111abcde11", "00000000000abcde"); |
| 669 | try_istri(wot,h,s, "cde111111abcde11", "00000000000abcde"); |
| 670 | |
| 671 | try_istri(wot,h,s, "01abcde11abcde11", "00000000000abcde"); |
| 672 | try_istri(wot,h,s, "00abcde11abcde11", "00000000000abcde"); |
| 673 | try_istri(wot,h,s, "000bcde11abcde11", "00000000000abcde"); |
| 674 | |
| 675 | try_istri(wot,h,s, "00abcde10abcde11", "00000000000abcde"); |
| 676 | try_istri(wot,h,s, "00abcde100bcde11", "00000000000abcde"); |
| 677 | |
| 678 | try_istri(wot,h,s, "1111111111111234", "0000000000000000"); |
| 679 | try_istri(wot,h,s, "1111111111111234", "0000000000000001"); |
| 680 | try_istri(wot,h,s, "1111111111111234", "0000000000000011"); |
| 681 | |
| 682 | try_istri(wot,h,s, "1111111111111234", "1111111111111234"); |
| 683 | try_istri(wot,h,s, "a111111111111111", "000000000000000a"); |
| 684 | try_istri(wot,h,s, "b111111111111111", "000000000000000a"); |
weidendo | c5274ae | 2012-07-25 09:49:26 +0000 | [diff] [blame] | 685 | |
| 686 | try_istri(wot,h,s, "b111111111111111", "0000000000000000"); |
| 687 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 688 | try_istri(wot,h,s, "123456789abcdef1", "0000000000000000"); |
| 689 | try_istri(wot,h,s, "0000000000000000", "123456789abcdef1"); |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 690 | } |
| 691 | |
| 692 | |
| 693 | ////////////////////////////////////////////////////////// |
| 694 | // // |
| 695 | // ISTRI_08 // |
| 696 | // // |
| 697 | ////////////////////////////////////////////////////////// |
| 698 | |
| 699 | UInt h_pcmpistri_08 ( V128* argL, V128* argR ) |
| 700 | { |
| 701 | V128 block[2]; |
| 702 | memcpy(&block[0], argL, sizeof(V128)); |
| 703 | memcpy(&block[1], argR, sizeof(V128)); |
| 704 | ULong res, flags; |
| 705 | __asm__ __volatile__( |
| 706 | "subq $1024, %%rsp" "\n\t" |
| 707 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 708 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 709 | "pcmpistri $0x08, %%xmm2, %%xmm11" "\n\t" |
| 710 | "pushfq" "\n\t" |
| 711 | "popq %%rdx" "\n\t" |
| 712 | "movq %%rcx, %0" "\n\t" |
| 713 | "movq %%rdx, %1" "\n\t" |
| 714 | "addq $1024, %%rsp" "\n\t" |
| 715 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 716 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 717 | ); |
| 718 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 719 | } |
| 720 | |
| 721 | UInt s_pcmpistri_08 ( V128* argLU, V128* argRU ) |
| 722 | { |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 723 | V128 resV; |
| 724 | UInt resOSZACP, resECX; |
| 725 | Bool ok |
| 726 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 727 | zmask_from_V128(argLU), |
| 728 | zmask_from_V128(argRU), |
| 729 | 0x08, False/*!isSTRM*/ |
| 730 | ); |
| 731 | assert(ok); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 732 | resECX = resV.uInt[0]; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 733 | return (resOSZACP << 16) | resECX; |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 734 | } |
| 735 | |
| 736 | void istri_08 ( void ) |
| 737 | { |
| 738 | char* wot = "08"; |
| 739 | UInt(*h)(V128*,V128*) = h_pcmpistri_08; |
| 740 | UInt(*s)(V128*,V128*) = s_pcmpistri_08; |
| 741 | |
| 742 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 743 | |
| 744 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 745 | try_istri(wot,h,s, "aaaa2aaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 746 | try_istri(wot,h,s, "aaaaaaaaa2aaaaaa", "aaaaaaaaaaaaaaaa"); |
| 747 | try_istri(wot,h,s, "aaaaaaaaaaaaa2aa", "aaaaaaaaaaaaaaaa"); |
| 748 | |
| 749 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaa2aaaaaaaaaaa"); |
| 750 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaa2aaaaaa"); |
| 751 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaa2a"); |
| 752 | |
| 753 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 754 | try_istri(wot,h,s, "baaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 755 | try_istri(wot,h,s, "b9aaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 756 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 757 | |
| 758 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 759 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaa7aaa"); |
| 760 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaa2aaa4aaa"); |
| 761 | |
| 762 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 763 | |
| 764 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 765 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 766 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaa0aaa"); |
| 767 | |
| 768 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 769 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 770 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 771 | |
| 772 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 773 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaa0aaaaaaa"); |
| 774 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaa0aaaaaaa"); |
| 775 | |
| 776 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaa0aaaaaaa"); |
| 777 | try_istri(wot,h,s, "8000000000000000", "aaaaaaaa0aaaaaaa"); |
| 778 | try_istri(wot,h,s, "0000000000000001", "aaaaaaaa0aaaaaaa"); |
| 779 | |
| 780 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaaaaaaaaaa"); |
| 781 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "0000000000000000"); |
| 782 | } |
| 783 | |
| 784 | |
| 785 | |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 786 | ////////////////////////////////////////////////////////// |
| 787 | // // |
sewardj | 053f436 | 2015-08-13 14:19:48 +0000 | [diff] [blame] | 788 | // ISTRI_18 // |
| 789 | // // |
| 790 | ////////////////////////////////////////////////////////// |
| 791 | |
| 792 | UInt h_pcmpistri_18 ( V128* argL, V128* argR ) |
| 793 | { |
| 794 | V128 block[2]; |
| 795 | memcpy(&block[0], argL, sizeof(V128)); |
| 796 | memcpy(&block[1], argR, sizeof(V128)); |
| 797 | ULong res, flags; |
| 798 | __asm__ __volatile__( |
| 799 | "subq $1024, %%rsp" "\n\t" |
| 800 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 801 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 802 | "pcmpistri $0x18, %%xmm2, %%xmm11" "\n\t" |
| 803 | "pushfq" "\n\t" |
| 804 | "popq %%rdx" "\n\t" |
| 805 | "movq %%rcx, %0" "\n\t" |
| 806 | "movq %%rdx, %1" "\n\t" |
| 807 | "addq $1024, %%rsp" "\n\t" |
| 808 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 809 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 810 | ); |
| 811 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 812 | } |
| 813 | |
| 814 | UInt s_pcmpistri_18 ( V128* argLU, V128* argRU ) |
| 815 | { |
| 816 | V128 resV; |
| 817 | UInt resOSZACP, resECX; |
| 818 | Bool ok |
| 819 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 820 | zmask_from_V128(argLU), |
| 821 | zmask_from_V128(argRU), |
| 822 | 0x18, False/*!isSTRM*/ |
| 823 | ); |
| 824 | assert(ok); |
| 825 | resECX = resV.uInt[0]; |
| 826 | return (resOSZACP << 16) | resECX; |
| 827 | } |
| 828 | |
| 829 | void istri_18 ( void ) |
| 830 | { |
| 831 | char* wot = "18"; |
| 832 | UInt(*h)(V128*,V128*) = h_pcmpistri_18; |
| 833 | UInt(*s)(V128*,V128*) = s_pcmpistri_18; |
| 834 | |
| 835 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 836 | |
| 837 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 838 | try_istri(wot,h,s, "aaaa2aaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 839 | try_istri(wot,h,s, "aaaaaaaaa2aaaaaa", "aaaaaaaaaaaaaaaa"); |
| 840 | try_istri(wot,h,s, "aaaaaaaaaaaaa2aa", "aaaaaaaaaaaaaaaa"); |
| 841 | |
| 842 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaa2aaaaaaaaaaa"); |
| 843 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaa2aaaaaa"); |
| 844 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaa2a"); |
| 845 | |
| 846 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 847 | try_istri(wot,h,s, "baaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 848 | try_istri(wot,h,s, "b9aaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 849 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 850 | |
| 851 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 852 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaa7aaa"); |
| 853 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaa2aaa4aaa"); |
| 854 | |
| 855 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 856 | |
| 857 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 858 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 859 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaa0aaa"); |
| 860 | |
| 861 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 862 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 863 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 864 | |
| 865 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 866 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaa0aaaaaaa"); |
| 867 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaa0aaaaaaa"); |
| 868 | |
| 869 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaa0aaaaaaa"); |
| 870 | try_istri(wot,h,s, "8000000000000000", "aaaaaaaa0aaaaaaa"); |
| 871 | try_istri(wot,h,s, "0000000000000001", "aaaaaaaa0aaaaaaa"); |
| 872 | |
| 873 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaaaaaaaaaa"); |
| 874 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "0000000000000000"); |
| 875 | } |
| 876 | |
| 877 | |
| 878 | |
| 879 | ////////////////////////////////////////////////////////// |
| 880 | // // |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 881 | // ISTRI_1A // |
| 882 | // // |
| 883 | ////////////////////////////////////////////////////////// |
| 884 | |
| 885 | UInt h_pcmpistri_1A ( V128* argL, V128* argR ) |
| 886 | { |
| 887 | V128 block[2]; |
| 888 | memcpy(&block[0], argL, sizeof(V128)); |
| 889 | memcpy(&block[1], argR, sizeof(V128)); |
| 890 | ULong res, flags; |
| 891 | __asm__ __volatile__( |
| 892 | "subq $1024, %%rsp" "\n\t" |
| 893 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 894 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 895 | "pcmpistri $0x1A, %%xmm2, %%xmm11" "\n\t" |
| 896 | "pushfq" "\n\t" |
| 897 | "popq %%rdx" "\n\t" |
| 898 | "movq %%rcx, %0" "\n\t" |
| 899 | "movq %%rdx, %1" "\n\t" |
| 900 | "addq $1024, %%rsp" "\n\t" |
| 901 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 902 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 903 | ); |
| 904 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 905 | } |
| 906 | |
| 907 | UInt s_pcmpistri_1A ( V128* argLU, V128* argRU ) |
| 908 | { |
| 909 | V128 resV; |
| 910 | UInt resOSZACP, resECX; |
| 911 | Bool ok |
| 912 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 913 | zmask_from_V128(argLU), |
| 914 | zmask_from_V128(argRU), |
| 915 | 0x1A, False/*!isSTRM*/ |
| 916 | ); |
| 917 | assert(ok); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 918 | resECX = resV.uInt[0]; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 919 | return (resOSZACP << 16) | resECX; |
| 920 | } |
| 921 | |
| 922 | void istri_1A ( void ) |
| 923 | { |
| 924 | char* wot = "1A"; |
| 925 | UInt(*h)(V128*,V128*) = h_pcmpistri_1A; |
| 926 | UInt(*s)(V128*,V128*) = s_pcmpistri_1A; |
| 927 | |
| 928 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 929 | |
| 930 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 931 | try_istri(wot,h,s, "aaaa2aaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 932 | try_istri(wot,h,s, "aaaaaaaaa2aaaaaa", "aaaaaaaaaaaaaaaa"); |
| 933 | try_istri(wot,h,s, "aaaaaaaaaaaaa2aa", "aaaaaaaaaaaaaaaa"); |
| 934 | |
| 935 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaa2aaaaaaaaaaa"); |
| 936 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaa2aaaaaa"); |
| 937 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaa2a"); |
| 938 | |
| 939 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 940 | try_istri(wot,h,s, "baaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 941 | try_istri(wot,h,s, "b9aaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 942 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 943 | |
| 944 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 945 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaa7aaa"); |
| 946 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaa2aaa4aaa"); |
| 947 | |
| 948 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 949 | |
| 950 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 951 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 952 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaa0aaa"); |
| 953 | |
| 954 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 955 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 956 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 957 | |
| 958 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 959 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaa0aaaaaaa"); |
| 960 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaa0aaaaaaa"); |
| 961 | |
| 962 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaa0aaaaaaa"); |
| 963 | try_istri(wot,h,s, "8000000000000000", "aaaaaaaa0aaaaaaa"); |
| 964 | try_istri(wot,h,s, "0000000000000001", "aaaaaaaa0aaaaaaa"); |
| 965 | |
| 966 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaaaaaaaaaa"); |
| 967 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "0000000000000000"); |
| 968 | } |
| 969 | |
| 970 | |
| 971 | |
| 972 | ////////////////////////////////////////////////////////// |
| 973 | // // |
| 974 | // ISTRI_02 // |
| 975 | // // |
| 976 | ////////////////////////////////////////////////////////// |
| 977 | |
| 978 | UInt h_pcmpistri_02 ( V128* argL, V128* argR ) |
| 979 | { |
| 980 | V128 block[2]; |
| 981 | memcpy(&block[0], argL, sizeof(V128)); |
| 982 | memcpy(&block[1], argR, sizeof(V128)); |
| 983 | ULong res, flags; |
| 984 | __asm__ __volatile__( |
| 985 | "subq $1024, %%rsp" "\n\t" |
| 986 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 987 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 988 | "pcmpistri $0x02, %%xmm2, %%xmm11" "\n\t" |
| 989 | //"pcmpistrm $0x02, %%xmm2, %%xmm11" "\n\t" |
| 990 | //"movd %%xmm0, %%ecx" "\n\t" |
| 991 | "pushfq" "\n\t" |
| 992 | "popq %%rdx" "\n\t" |
| 993 | "movq %%rcx, %0" "\n\t" |
| 994 | "movq %%rdx, %1" "\n\t" |
| 995 | "addq $1024, %%rsp" "\n\t" |
| 996 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 997 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 998 | ); |
| 999 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1000 | } |
| 1001 | |
| 1002 | UInt s_pcmpistri_02 ( V128* argLU, V128* argRU ) |
| 1003 | { |
| 1004 | V128 resV; |
| 1005 | UInt resOSZACP, resECX; |
| 1006 | Bool ok |
| 1007 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1008 | zmask_from_V128(argLU), |
| 1009 | zmask_from_V128(argRU), |
| 1010 | 0x02, False/*!isSTRM*/ |
| 1011 | ); |
| 1012 | assert(ok); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 1013 | resECX = resV.uInt[0]; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 1014 | return (resOSZACP << 16) | resECX; |
| 1015 | } |
| 1016 | |
| 1017 | void istri_02 ( void ) |
| 1018 | { |
| 1019 | char* wot = "02"; |
| 1020 | UInt(*h)(V128*,V128*) = h_pcmpistri_02; |
| 1021 | UInt(*s)(V128*,V128*) = s_pcmpistri_02; |
| 1022 | |
| 1023 | try_istri(wot,h,s, "abcdacbdabcdabcd", "000000000000000a"); |
| 1024 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000000b"); |
| 1025 | try_istri(wot,h,s, "abcdabcdabcdabcd", "00000000000000ab"); |
| 1026 | try_istri(wot,h,s, "abcdabc0abcdabcd", "000000000000abcd"); |
| 1027 | |
| 1028 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1029 | try_istri(wot,h,s, "0bcdabcdabcdabcd", "000000000000abcd"); |
| 1030 | try_istri(wot,h,s, "abcdabcdabcda0cd", "000000000000abcd"); |
| 1031 | try_istri(wot,h,s, "abcdabcdabcdab0d", "000000000000abcd"); |
| 1032 | try_istri(wot,h,s, "abcdabcdabcdabc0", "000000000000abcd"); |
| 1033 | |
| 1034 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1035 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000a0cd"); |
| 1036 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000ab0d"); |
| 1037 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abc0"); |
| 1038 | |
| 1039 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1040 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1041 | |
| 1042 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000abcd"); |
| 1043 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000dcba"); |
| 1044 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000bbbb"); |
| 1045 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000baba"); |
| 1046 | |
| 1047 | try_istri(wot,h,s, "0000abcdabcdabcd", "00000000000baba0"); |
| 1048 | |
| 1049 | try_istri(wot,h,s, "0ddc0ffeebadf00d", "00000000cafebabe"); |
| 1050 | try_istri(wot,h,s, "0ddc0ffeebadfeed", "00000000cafebabe"); |
| 1051 | } |
| 1052 | |
| 1053 | |
| 1054 | ////////////////////////////////////////////////////////// |
| 1055 | // // |
| 1056 | // ISTRI_12 // |
| 1057 | // // |
| 1058 | ////////////////////////////////////////////////////////// |
| 1059 | |
| 1060 | UInt h_pcmpistri_12 ( V128* argL, V128* argR ) |
| 1061 | { |
| 1062 | V128 block[2]; |
| 1063 | memcpy(&block[0], argL, sizeof(V128)); |
| 1064 | memcpy(&block[1], argR, sizeof(V128)); |
| 1065 | ULong res, flags; |
| 1066 | __asm__ __volatile__( |
| 1067 | "subq $1024, %%rsp" "\n\t" |
| 1068 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1069 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1070 | "pcmpistri $0x12, %%xmm2, %%xmm11" "\n\t" |
| 1071 | //"pcmpistrm $0x12, %%xmm2, %%xmm11" "\n\t" |
| 1072 | //"movd %%xmm0, %%ecx" "\n\t" |
| 1073 | "pushfq" "\n\t" |
| 1074 | "popq %%rdx" "\n\t" |
| 1075 | "movq %%rcx, %0" "\n\t" |
| 1076 | "movq %%rdx, %1" "\n\t" |
| 1077 | "addq $1024, %%rsp" "\n\t" |
| 1078 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1079 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1080 | ); |
| 1081 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1082 | } |
| 1083 | |
| 1084 | UInt s_pcmpistri_12 ( V128* argLU, V128* argRU ) |
| 1085 | { |
| 1086 | V128 resV; |
| 1087 | UInt resOSZACP, resECX; |
| 1088 | Bool ok |
| 1089 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1090 | zmask_from_V128(argLU), |
| 1091 | zmask_from_V128(argRU), |
| 1092 | 0x12, False/*!isSTRM*/ |
| 1093 | ); |
| 1094 | assert(ok); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 1095 | resECX = resV.uInt[0]; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 1096 | return (resOSZACP << 16) | resECX; |
| 1097 | } |
| 1098 | |
| 1099 | void istri_12 ( void ) |
| 1100 | { |
| 1101 | char* wot = "12"; |
| 1102 | UInt(*h)(V128*,V128*) = h_pcmpistri_12; |
| 1103 | UInt(*s)(V128*,V128*) = s_pcmpistri_12; |
| 1104 | |
| 1105 | try_istri(wot,h,s, "abcdacbdabcdabcd", "000000000000000a"); |
| 1106 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000000b"); |
| 1107 | try_istri(wot,h,s, "abcdabcdabcdabcd", "00000000000000ab"); |
| 1108 | try_istri(wot,h,s, "abcdabc0abcdabcd", "000000000000abcd"); |
| 1109 | |
| 1110 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1111 | try_istri(wot,h,s, "0bcdabcdabcdabcd", "000000000000abcd"); |
| 1112 | try_istri(wot,h,s, "abcdabcdabcda0cd", "000000000000abcd"); |
| 1113 | try_istri(wot,h,s, "abcdabcdabcdab0d", "000000000000abcd"); |
| 1114 | try_istri(wot,h,s, "abcdabcdabcdabc0", "000000000000abcd"); |
| 1115 | |
| 1116 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1117 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000a0cd"); |
| 1118 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000ab0d"); |
| 1119 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abc0"); |
| 1120 | |
| 1121 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1122 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1123 | |
| 1124 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000abcd"); |
| 1125 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000dcba"); |
| 1126 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000bbbb"); |
| 1127 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000baba"); |
| 1128 | |
| 1129 | try_istri(wot,h,s, "0000abcdabcdabcd", "00000000000baba0"); |
| 1130 | |
| 1131 | try_istri(wot,h,s, "0ddc0ffeebadf00d", "00000000cafebabe"); |
| 1132 | try_istri(wot,h,s, "0ddc0ffeebadfeed", "00000000cafebabe"); |
| 1133 | } |
| 1134 | |
| 1135 | |
| 1136 | |
| 1137 | ////////////////////////////////////////////////////////// |
| 1138 | // // |
| 1139 | // ISTRI_44 // |
| 1140 | // // |
| 1141 | ////////////////////////////////////////////////////////// |
| 1142 | |
| 1143 | UInt h_pcmpistri_44 ( V128* argL, V128* argR ) |
| 1144 | { |
| 1145 | V128 block[2]; |
| 1146 | memcpy(&block[0], argL, sizeof(V128)); |
| 1147 | memcpy(&block[1], argR, sizeof(V128)); |
| 1148 | ULong res, flags; |
| 1149 | __asm__ __volatile__( |
| 1150 | "subq $1024, %%rsp" "\n\t" |
| 1151 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1152 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1153 | "pcmpistri $0x44, %%xmm2, %%xmm11" "\n\t" |
| 1154 | //"pcmpistrm $0x04, %%xmm2, %%xmm11" "\n\t" |
| 1155 | //"movd %%xmm0, %%ecx" "\n\t" |
| 1156 | "pushfq" "\n\t" |
| 1157 | "popq %%rdx" "\n\t" |
| 1158 | "movq %%rcx, %0" "\n\t" |
| 1159 | "movq %%rdx, %1" "\n\t" |
| 1160 | "addq $1024, %%rsp" "\n\t" |
| 1161 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1162 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1163 | ); |
| 1164 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1165 | } |
| 1166 | |
| 1167 | UInt s_pcmpistri_44 ( V128* argLU, V128* argRU ) |
| 1168 | { |
| 1169 | V128 resV; |
| 1170 | UInt resOSZACP, resECX; |
| 1171 | Bool ok |
| 1172 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1173 | zmask_from_V128(argLU), |
| 1174 | zmask_from_V128(argRU), |
| 1175 | 0x44, False/*!isSTRM*/ |
| 1176 | ); |
| 1177 | assert(ok); |
sewardj | 7f3019b | 2010-08-17 22:39:46 +0000 | [diff] [blame] | 1178 | resECX = resV.uInt[0]; |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 1179 | return (resOSZACP << 16) | resECX; |
| 1180 | } |
| 1181 | |
| 1182 | void istri_44 ( void ) |
| 1183 | { |
| 1184 | char* wot = "44"; |
| 1185 | UInt(*h)(V128*,V128*) = h_pcmpistri_44; |
| 1186 | UInt(*s)(V128*,V128*) = s_pcmpistri_44; |
| 1187 | |
| 1188 | try_istri(wot,h,s, "aaaabbbbccccdddd", "00000000000000bc"); |
| 1189 | try_istri(wot,h,s, "aaaabbbbccccdddd", "00000000000000cb"); |
| 1190 | try_istri(wot,h,s, "baaabbbbccccdddd", "00000000000000cb"); |
| 1191 | try_istri(wot,h,s, "baaabbbbccccdddc", "00000000000000cb"); |
| 1192 | |
| 1193 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "00000000000000cb"); |
| 1194 | try_istri(wot,h,s, "bbbbbbbb0bbbbbbb", "00000000000000cb"); |
| 1195 | try_istri(wot,h,s, "bbbbbbbbbbbbbb0b", "00000000000000cb"); |
| 1196 | try_istri(wot,h,s, "bbbbbbbbbbbbbbb0", "00000000000000cb"); |
| 1197 | try_istri(wot,h,s, "0000000000000000", "00000000000000cb"); |
| 1198 | |
| 1199 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1200 | |
| 1201 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "00000000000000cb"); |
| 1202 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "000000000000000b"); |
| 1203 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000062cb"); |
| 1204 | |
| 1205 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000002cb"); |
| 1206 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000000cb"); |
| 1207 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "000000000000000b"); |
| 1208 | |
| 1209 | try_istri(wot,h,s, "0123456789abcdef", "000000fecb975421"); |
| 1210 | try_istri(wot,h,s, "123456789abcdef1", "000000fecb975421"); |
| 1211 | |
| 1212 | try_istri(wot,h,s, "0123456789abcdef", "00000000dca86532"); |
| 1213 | try_istri(wot,h,s, "123456789abcdef1", "00000000dca86532"); |
sewardj | 15df336 | 2012-09-17 13:43:57 +0000 | [diff] [blame] | 1214 | |
| 1215 | try_istri(wot,h,s, "163887ec041a9b72", "fcd75adb9b3e895a"); |
| 1216 | try_istri(wot,h,s, "fc937cbfbf53f8e2", "0d136bcb024d3fb7"); |
| 1217 | try_istri(wot,h,s, "2ca34182c29a82ab", "302ebd646775ab54"); |
| 1218 | try_istri(wot,h,s, "3f2987608c11be6f", "a9ecb661f8e0a8cb"); |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 1219 | } |
| 1220 | |
| 1221 | |
sewardj | 3b20b17 | 2011-01-17 23:10:39 +0000 | [diff] [blame] | 1222 | ////////////////////////////////////////////////////////// |
| 1223 | // // |
| 1224 | // ISTRI_00 // |
| 1225 | // // |
| 1226 | ////////////////////////////////////////////////////////// |
| 1227 | |
| 1228 | UInt h_pcmpistri_00 ( V128* argL, V128* argR ) |
| 1229 | { |
| 1230 | V128 block[2]; |
| 1231 | memcpy(&block[0], argL, sizeof(V128)); |
| 1232 | memcpy(&block[1], argR, sizeof(V128)); |
| 1233 | ULong res, flags; |
| 1234 | __asm__ __volatile__( |
| 1235 | "subq $1024, %%rsp" "\n\t" |
| 1236 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1237 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1238 | "pcmpistri $0x00, %%xmm2, %%xmm11" "\n\t" |
| 1239 | //"pcmpistrm $0x00, %%xmm2, %%xmm11" "\n\t" |
| 1240 | //"movd %%xmm0, %%ecx" "\n\t" |
| 1241 | "pushfq" "\n\t" |
| 1242 | "popq %%rdx" "\n\t" |
| 1243 | "movq %%rcx, %0" "\n\t" |
| 1244 | "movq %%rdx, %1" "\n\t" |
| 1245 | "addq $1024, %%rsp" "\n\t" |
| 1246 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1247 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1248 | ); |
| 1249 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1250 | } |
| 1251 | |
| 1252 | UInt s_pcmpistri_00 ( V128* argLU, V128* argRU ) |
| 1253 | { |
| 1254 | V128 resV; |
| 1255 | UInt resOSZACP, resECX; |
| 1256 | Bool ok |
| 1257 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1258 | zmask_from_V128(argLU), |
| 1259 | zmask_from_V128(argRU), |
| 1260 | 0x00, False/*!isSTRM*/ |
| 1261 | ); |
| 1262 | assert(ok); |
| 1263 | resECX = resV.uInt[0]; |
| 1264 | return (resOSZACP << 16) | resECX; |
| 1265 | } |
| 1266 | |
| 1267 | void istri_00 ( void ) |
| 1268 | { |
| 1269 | char* wot = "00"; |
| 1270 | UInt(*h)(V128*,V128*) = h_pcmpistri_00; |
| 1271 | UInt(*s)(V128*,V128*) = s_pcmpistri_00; |
| 1272 | |
| 1273 | try_istri(wot,h,s, "abcdacbdabcdabcd", "000000000000000a"); |
| 1274 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000000b"); |
| 1275 | try_istri(wot,h,s, "abcdabcdabcdabcd", "00000000000000ab"); |
| 1276 | try_istri(wot,h,s, "abcdabc0abcdabcd", "000000000000abcd"); |
| 1277 | |
| 1278 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1279 | try_istri(wot,h,s, "0bcdabcdabcdabcd", "000000000000abcd"); |
| 1280 | try_istri(wot,h,s, "abcdabcdabcda0cd", "000000000000abcd"); |
| 1281 | try_istri(wot,h,s, "abcdabcdabcdab0d", "000000000000abcd"); |
| 1282 | try_istri(wot,h,s, "abcdabcdabcdabc0", "000000000000abcd"); |
| 1283 | |
| 1284 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1285 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000a0cd"); |
| 1286 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000ab0d"); |
| 1287 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abc0"); |
| 1288 | |
| 1289 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1290 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1291 | |
| 1292 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000abcd"); |
| 1293 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000dcba"); |
| 1294 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000bbbb"); |
| 1295 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000baba"); |
| 1296 | |
| 1297 | try_istri(wot,h,s, "0000abcdabcdabcd", "00000000000baba0"); |
| 1298 | |
| 1299 | try_istri(wot,h,s, "0ddc0ffeebadf00d", "00000000cafebabe"); |
| 1300 | try_istri(wot,h,s, "0ddc0ffeebadfeed", "00000000cafebabe"); |
| 1301 | } |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 1302 | |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 1303 | |
sewardj | e801ed2 | 2011-10-19 20:17:50 +0000 | [diff] [blame] | 1304 | ////////////////////////////////////////////////////////// |
| 1305 | // // |
| 1306 | // ISTRI_38 // |
| 1307 | // // |
| 1308 | ////////////////////////////////////////////////////////// |
| 1309 | |
| 1310 | UInt h_pcmpistri_38 ( V128* argL, V128* argR ) |
| 1311 | { |
| 1312 | V128 block[2]; |
| 1313 | memcpy(&block[0], argL, sizeof(V128)); |
| 1314 | memcpy(&block[1], argR, sizeof(V128)); |
| 1315 | ULong res, flags; |
| 1316 | __asm__ __volatile__( |
| 1317 | "subq $1024, %%rsp" "\n\t" |
| 1318 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1319 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1320 | "pcmpistri $0x38, %%xmm2, %%xmm11" "\n\t" |
| 1321 | "pushfq" "\n\t" |
| 1322 | "popq %%rdx" "\n\t" |
| 1323 | "movq %%rcx, %0" "\n\t" |
| 1324 | "movq %%rdx, %1" "\n\t" |
| 1325 | "addq $1024, %%rsp" "\n\t" |
| 1326 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1327 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1328 | ); |
| 1329 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1330 | } |
| 1331 | |
| 1332 | UInt s_pcmpistri_38 ( V128* argLU, V128* argRU ) |
| 1333 | { |
| 1334 | V128 resV; |
| 1335 | UInt resOSZACP, resECX; |
| 1336 | Bool ok |
| 1337 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1338 | zmask_from_V128(argLU), |
| 1339 | zmask_from_V128(argRU), |
| 1340 | 0x38, False/*!isSTRM*/ |
| 1341 | ); |
| 1342 | assert(ok); |
| 1343 | resECX = resV.uInt[0]; |
| 1344 | return (resOSZACP << 16) | resECX; |
| 1345 | } |
| 1346 | |
| 1347 | void istri_38 ( void ) |
| 1348 | { |
| 1349 | char* wot = "38"; |
| 1350 | UInt(*h)(V128*,V128*) = h_pcmpistri_38; |
| 1351 | UInt(*s)(V128*,V128*) = s_pcmpistri_38; |
| 1352 | |
| 1353 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1354 | |
| 1355 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1356 | try_istri(wot,h,s, "aaaa2aaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1357 | try_istri(wot,h,s, "aaaaaaaaa2aaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1358 | try_istri(wot,h,s, "aaaaaaaaaaaaa2aa", "aaaaaaaaaaaaaaaa"); |
| 1359 | |
| 1360 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaa2aaaaaaaaaaa"); |
| 1361 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaa2aaaaaa"); |
| 1362 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaa2a"); |
| 1363 | |
| 1364 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1365 | try_istri(wot,h,s, "baaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1366 | try_istri(wot,h,s, "b9aaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1367 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1368 | |
| 1369 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1370 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaaaaaa7aaa"); |
| 1371 | try_istri(wot,h,s, "b9baaaaaaaaaaaaa", "aaaaaaaa2aaa4aaa"); |
| 1372 | |
| 1373 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1374 | |
| 1375 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 1376 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 1377 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaa0aaa"); |
| 1378 | |
| 1379 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1380 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 1381 | try_istri(wot,h,s, "aaaaaaaa0aaaaaaa", "aaaaaaaaaaaa0aaa"); |
| 1382 | |
| 1383 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaaaaaaaaaa"); |
| 1384 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaa0aaaaaaa"); |
| 1385 | try_istri(wot,h,s, "aaaaaaaaaaaa0aaa", "aaaaaaaa0aaaaaaa"); |
| 1386 | |
| 1387 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaa0aaaaaaa"); |
| 1388 | try_istri(wot,h,s, "8000000000000000", "aaaaaaaa0aaaaaaa"); |
| 1389 | try_istri(wot,h,s, "0000000000000001", "aaaaaaaa0aaaaaaa"); |
| 1390 | |
| 1391 | try_istri(wot,h,s, "0000000000000000", "aaaaaaaaaaaaaaaa"); |
| 1392 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "0000000000000000"); |
| 1393 | } |
| 1394 | |
| 1395 | |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 1396 | |
| 1397 | ////////////////////////////////////////////////////////// |
| 1398 | // // |
sewardj | 15df336 | 2012-09-17 13:43:57 +0000 | [diff] [blame] | 1399 | // ISTRI_46 // |
| 1400 | // // |
| 1401 | ////////////////////////////////////////////////////////// |
| 1402 | |
| 1403 | UInt h_pcmpistri_46 ( V128* argL, V128* argR ) |
| 1404 | { |
| 1405 | V128 block[2]; |
| 1406 | memcpy(&block[0], argL, sizeof(V128)); |
| 1407 | memcpy(&block[1], argR, sizeof(V128)); |
| 1408 | ULong res, flags; |
| 1409 | __asm__ __volatile__( |
| 1410 | "subq $1024, %%rsp" "\n\t" |
| 1411 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1412 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1413 | "pcmpistri $0x46, %%xmm2, %%xmm11" "\n\t" |
| 1414 | "pushfq" "\n\t" |
| 1415 | "popq %%rdx" "\n\t" |
| 1416 | "movq %%rcx, %0" "\n\t" |
| 1417 | "movq %%rdx, %1" "\n\t" |
| 1418 | "addq $1024, %%rsp" "\n\t" |
| 1419 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1420 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1421 | ); |
| 1422 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1423 | } |
| 1424 | |
| 1425 | UInt s_pcmpistri_46 ( V128* argLU, V128* argRU ) |
| 1426 | { |
| 1427 | V128 resV; |
| 1428 | UInt resOSZACP, resECX; |
| 1429 | Bool ok |
| 1430 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1431 | zmask_from_V128(argLU), |
| 1432 | zmask_from_V128(argRU), |
| 1433 | 0x46, False/*!isSTRM*/ |
| 1434 | ); |
| 1435 | assert(ok); |
| 1436 | resECX = resV.uInt[0]; |
| 1437 | return (resOSZACP << 16) | resECX; |
| 1438 | } |
| 1439 | |
| 1440 | void istri_46 ( void ) |
| 1441 | { |
| 1442 | char* wot = "46"; |
| 1443 | UInt(*h)(V128*,V128*) = h_pcmpistri_46; |
| 1444 | UInt(*s)(V128*,V128*) = s_pcmpistri_46; |
| 1445 | |
| 1446 | try_istri(wot,h,s, "aaaabbbbccccdddd", "00000000000000bc"); |
| 1447 | try_istri(wot,h,s, "aaaabbbbccccdddd", "00000000000000cb"); |
| 1448 | try_istri(wot,h,s, "baaabbbbccccdddd", "00000000000000cb"); |
| 1449 | try_istri(wot,h,s, "baaabbbbccccdddc", "00000000000000cb"); |
| 1450 | |
| 1451 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "00000000000000cb"); |
| 1452 | try_istri(wot,h,s, "bbbbbbbb0bbbbbbb", "00000000000000cb"); |
| 1453 | try_istri(wot,h,s, "bbbbbbbbbbbbbb0b", "00000000000000cb"); |
| 1454 | try_istri(wot,h,s, "bbbbbbbbbbbbbbb0", "00000000000000cb"); |
| 1455 | try_istri(wot,h,s, "0000000000000000", "00000000000000cb"); |
| 1456 | |
| 1457 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1458 | |
| 1459 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "00000000000000cb"); |
| 1460 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "000000000000000b"); |
| 1461 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000062cb"); |
| 1462 | |
| 1463 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000002cb"); |
| 1464 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000000cb"); |
| 1465 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "000000000000000b"); |
| 1466 | |
| 1467 | try_istri(wot,h,s, "0123456789abcdef", "000000fecb975421"); |
| 1468 | try_istri(wot,h,s, "123456789abcdef1", "000000fecb975421"); |
| 1469 | |
| 1470 | try_istri(wot,h,s, "0123456789abcdef", "00000000dca86532"); |
| 1471 | try_istri(wot,h,s, "123456789abcdef1", "00000000dca86532"); |
| 1472 | |
| 1473 | try_istri(wot,h,s, "163887ec041a9b72", "fcd75adb9b3e895a"); |
| 1474 | try_istri(wot,h,s, "fc937cbfbf53f8e2", "0d136bcb024d3fb7"); |
| 1475 | try_istri(wot,h,s, "2ca34182c29a82ab", "302ebd646775ab54"); |
| 1476 | try_istri(wot,h,s, "3f2987608c11be6f", "a9ecb661f8e0a8cb"); |
| 1477 | } |
| 1478 | |
| 1479 | |
| 1480 | ////////////////////////////////////////////////////////// |
| 1481 | // // |
sewardj | 473d0ba | 2013-10-18 10:46:52 +0000 | [diff] [blame] | 1482 | // ISTRI_30 // |
| 1483 | // // |
| 1484 | ////////////////////////////////////////////////////////// |
| 1485 | |
| 1486 | UInt h_pcmpistri_30 ( V128* argL, V128* argR ) |
| 1487 | { |
| 1488 | V128 block[2]; |
| 1489 | memcpy(&block[0], argL, sizeof(V128)); |
| 1490 | memcpy(&block[1], argR, sizeof(V128)); |
| 1491 | ULong res, flags; |
| 1492 | __asm__ __volatile__( |
| 1493 | "subq $1024, %%rsp" "\n\t" |
| 1494 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1495 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1496 | "pcmpistri $0x30, %%xmm2, %%xmm11" "\n\t" |
| 1497 | "pushfq" "\n\t" |
| 1498 | "popq %%rdx" "\n\t" |
| 1499 | "movq %%rcx, %0" "\n\t" |
| 1500 | "movq %%rdx, %1" "\n\t" |
| 1501 | "addq $1024, %%rsp" "\n\t" |
| 1502 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1503 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1504 | ); |
| 1505 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1506 | } |
| 1507 | |
| 1508 | UInt s_pcmpistri_30 ( V128* argLU, V128* argRU ) |
| 1509 | { |
| 1510 | V128 resV; |
| 1511 | UInt resOSZACP, resECX; |
| 1512 | Bool ok |
| 1513 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1514 | zmask_from_V128(argLU), |
| 1515 | zmask_from_V128(argRU), |
| 1516 | 0x30, False/*!isSTRM*/ |
| 1517 | ); |
| 1518 | assert(ok); |
| 1519 | resECX = resV.uInt[0]; |
| 1520 | return (resOSZACP << 16) | resECX; |
| 1521 | } |
| 1522 | |
| 1523 | void istri_30 ( void ) |
| 1524 | { |
| 1525 | char* wot = "30"; |
| 1526 | UInt(*h)(V128*,V128*) = h_pcmpistri_30; |
| 1527 | UInt(*s)(V128*,V128*) = s_pcmpistri_30; |
| 1528 | |
| 1529 | try_istri(wot,h,s, "abcdacbdabcdabcd", "000000000000000a"); |
| 1530 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000000b"); |
| 1531 | try_istri(wot,h,s, "abcdabcdabcdabcd", "00000000000000ab"); |
| 1532 | try_istri(wot,h,s, "abcdabc0abcdabcd", "000000000000abcd"); |
| 1533 | |
| 1534 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1535 | try_istri(wot,h,s, "0bcdabcdabcdabcd", "000000000000abcd"); |
| 1536 | try_istri(wot,h,s, "abcdabcdabcda0cd", "000000000000abcd"); |
| 1537 | try_istri(wot,h,s, "abcdabcdabcdab0d", "000000000000abcd"); |
| 1538 | try_istri(wot,h,s, "abcdabcdabcdabc0", "000000000000abcd"); |
| 1539 | |
| 1540 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1541 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000a0cd"); |
| 1542 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000ab0d"); |
| 1543 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abc0"); |
| 1544 | |
| 1545 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1546 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1547 | |
| 1548 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000abcd"); |
| 1549 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000dcba"); |
| 1550 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000bbbb"); |
| 1551 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000baba"); |
| 1552 | |
| 1553 | try_istri(wot,h,s, "0000abcdabcdabcd", "00000000000baba0"); |
| 1554 | |
| 1555 | try_istri(wot,h,s, "0ddc0ffeebadf00d", "00000000cafebabe"); |
| 1556 | try_istri(wot,h,s, "0ddc0ffeebadfeed", "00000000cafebabe"); |
| 1557 | } |
| 1558 | |
| 1559 | |
| 1560 | ////////////////////////////////////////////////////////// |
| 1561 | // // |
| 1562 | // ISTRI_40 // |
| 1563 | // // |
| 1564 | ////////////////////////////////////////////////////////// |
| 1565 | |
| 1566 | UInt h_pcmpistri_40 ( V128* argL, V128* argR ) |
| 1567 | { |
| 1568 | V128 block[2]; |
| 1569 | memcpy(&block[0], argL, sizeof(V128)); |
| 1570 | memcpy(&block[1], argR, sizeof(V128)); |
| 1571 | ULong res, flags; |
| 1572 | __asm__ __volatile__( |
| 1573 | "subq $1024, %%rsp" "\n\t" |
| 1574 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1575 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1576 | "pcmpistri $0x40, %%xmm2, %%xmm11" "\n\t" |
| 1577 | "pushfq" "\n\t" |
| 1578 | "popq %%rdx" "\n\t" |
| 1579 | "movq %%rcx, %0" "\n\t" |
| 1580 | "movq %%rdx, %1" "\n\t" |
| 1581 | "addq $1024, %%rsp" "\n\t" |
| 1582 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1583 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1584 | ); |
| 1585 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1586 | } |
| 1587 | |
| 1588 | UInt s_pcmpistri_40 ( V128* argLU, V128* argRU ) |
| 1589 | { |
| 1590 | V128 resV; |
| 1591 | UInt resOSZACP, resECX; |
| 1592 | Bool ok |
| 1593 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1594 | zmask_from_V128(argLU), |
| 1595 | zmask_from_V128(argRU), |
| 1596 | 0x40, False/*!isSTRM*/ |
| 1597 | ); |
| 1598 | assert(ok); |
| 1599 | resECX = resV.uInt[0]; |
| 1600 | return (resOSZACP << 16) | resECX; |
| 1601 | } |
| 1602 | |
| 1603 | void istri_40 ( void ) |
| 1604 | { |
| 1605 | char* wot = "40"; |
| 1606 | UInt(*h)(V128*,V128*) = h_pcmpistri_40; |
| 1607 | UInt(*s)(V128*,V128*) = s_pcmpistri_40; |
| 1608 | |
| 1609 | try_istri(wot,h,s, "abcdacbdabcdabcd", "000000000000000a"); |
| 1610 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000000b"); |
| 1611 | try_istri(wot,h,s, "abcdabcdabcdabcd", "00000000000000ab"); |
| 1612 | try_istri(wot,h,s, "abcdabc0abcdabcd", "000000000000abcd"); |
| 1613 | |
| 1614 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1615 | try_istri(wot,h,s, "0bcdabcdabcdabcd", "000000000000abcd"); |
| 1616 | try_istri(wot,h,s, "abcdabcdabcda0cd", "000000000000abcd"); |
| 1617 | try_istri(wot,h,s, "abcdabcdabcdab0d", "000000000000abcd"); |
| 1618 | try_istri(wot,h,s, "abcdabcdabcdabc0", "000000000000abcd"); |
| 1619 | |
| 1620 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1621 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000a0cd"); |
| 1622 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000ab0d"); |
| 1623 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abc0"); |
| 1624 | |
| 1625 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1626 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1627 | |
| 1628 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000abcd"); |
| 1629 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000dcba"); |
| 1630 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000bbbb"); |
| 1631 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000baba"); |
| 1632 | |
| 1633 | try_istri(wot,h,s, "0000abcdabcdabcd", "00000000000baba0"); |
| 1634 | |
| 1635 | try_istri(wot,h,s, "0ddc0ffeebadf00d", "00000000cafebabe"); |
| 1636 | try_istri(wot,h,s, "0ddc0ffeebadfeed", "00000000cafebabe"); |
| 1637 | } |
| 1638 | |
| 1639 | |
| 1640 | ////////////////////////////////////////////////////////// |
| 1641 | // // |
sewardj | 7f0dfdd | 2015-08-13 13:47:02 +0000 | [diff] [blame] | 1642 | // ISTRI_42 // |
| 1643 | // // |
| 1644 | ////////////////////////////////////////////////////////// |
| 1645 | |
| 1646 | UInt h_pcmpistri_42 ( V128* argL, V128* argR ) |
| 1647 | { |
| 1648 | V128 block[2]; |
| 1649 | memcpy(&block[0], argL, sizeof(V128)); |
| 1650 | memcpy(&block[1], argR, sizeof(V128)); |
| 1651 | ULong res, flags; |
| 1652 | __asm__ __volatile__( |
| 1653 | "subq $1024, %%rsp" "\n\t" |
| 1654 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1655 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1656 | "pcmpistri $0x42, %%xmm2, %%xmm11" "\n\t" |
| 1657 | "pushfq" "\n\t" |
| 1658 | "popq %%rdx" "\n\t" |
| 1659 | "movq %%rcx, %0" "\n\t" |
| 1660 | "movq %%rdx, %1" "\n\t" |
| 1661 | "addq $1024, %%rsp" "\n\t" |
| 1662 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1663 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1664 | ); |
| 1665 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1666 | } |
| 1667 | |
| 1668 | UInt s_pcmpistri_42 ( V128* argLU, V128* argRU ) |
| 1669 | { |
| 1670 | V128 resV; |
| 1671 | UInt resOSZACP, resECX; |
| 1672 | Bool ok |
| 1673 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1674 | zmask_from_V128(argLU), |
| 1675 | zmask_from_V128(argRU), |
| 1676 | 0x42, False/*!isSTRM*/ |
| 1677 | ); |
| 1678 | assert(ok); |
| 1679 | resECX = resV.uInt[0]; |
| 1680 | return (resOSZACP << 16) | resECX; |
| 1681 | } |
| 1682 | |
| 1683 | void istri_42 ( void ) |
| 1684 | { |
| 1685 | char* wot = "42"; |
| 1686 | UInt(*h)(V128*,V128*) = h_pcmpistri_42; |
| 1687 | UInt(*s)(V128*,V128*) = s_pcmpistri_42; |
| 1688 | |
| 1689 | try_istri(wot,h,s, "abcdacbdabcdabcd", "000000000000000a"); |
| 1690 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000000b"); |
| 1691 | try_istri(wot,h,s, "abcdabcdabcdabcd", "00000000000000ab"); |
| 1692 | try_istri(wot,h,s, "abcdabc0abcdabcd", "000000000000abcd"); |
| 1693 | |
| 1694 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1695 | try_istri(wot,h,s, "0bcdabcdabcdabcd", "000000000000abcd"); |
| 1696 | try_istri(wot,h,s, "abcdabcdabcda0cd", "000000000000abcd"); |
| 1697 | try_istri(wot,h,s, "abcdabcdabcdab0d", "000000000000abcd"); |
| 1698 | try_istri(wot,h,s, "abcdabcdabcdabc0", "000000000000abcd"); |
| 1699 | |
| 1700 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); |
| 1701 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000a0cd"); |
| 1702 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000ab0d"); |
| 1703 | try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abc0"); |
| 1704 | |
| 1705 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1706 | try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); |
| 1707 | |
| 1708 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000abcd"); |
| 1709 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000dcba"); |
| 1710 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000bbbb"); |
| 1711 | try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000baba"); |
| 1712 | |
| 1713 | try_istri(wot,h,s, "0000abcdabcdabcd", "00000000000baba0"); |
| 1714 | |
| 1715 | try_istri(wot,h,s, "0ddc0ffeebadf00d", "00000000cafebabe"); |
| 1716 | try_istri(wot,h,s, "0ddc0ffeebadfeed", "00000000cafebabe"); |
| 1717 | } |
| 1718 | |
| 1719 | |
| 1720 | ////////////////////////////////////////////////////////// |
| 1721 | // // |
sewardj | a2bf389 | 2014-05-21 14:43:11 +0000 | [diff] [blame] | 1722 | // ISTRI_0E // |
| 1723 | // // |
| 1724 | ////////////////////////////////////////////////////////// |
| 1725 | |
| 1726 | __attribute__((noinline)) |
| 1727 | UInt h_pcmpistri_0E ( V128* argL, V128* argR ) |
| 1728 | { |
| 1729 | V128 block[2]; |
| 1730 | memcpy(&block[0], argL, sizeof(V128)); |
| 1731 | memcpy(&block[1], argR, sizeof(V128)); |
| 1732 | ULong res = 0, flags = 0; |
| 1733 | __asm__ __volatile__( |
| 1734 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1735 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1736 | "pcmpistri $0x0E, %%xmm2, %%xmm11" "\n\t" |
| 1737 | "pushfq" "\n\t" |
| 1738 | "popq %%rdx" "\n\t" |
| 1739 | "movq %%rcx, %0" "\n\t" |
| 1740 | "movq %%rdx, %1" "\n\t" |
| 1741 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1742 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1743 | ); |
| 1744 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1745 | } |
| 1746 | |
| 1747 | UInt s_pcmpistri_0E ( V128* argLU, V128* argRU ) |
| 1748 | { |
| 1749 | V128 resV; |
| 1750 | UInt resOSZACP, resECX; |
| 1751 | Bool ok |
| 1752 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1753 | zmask_from_V128(argLU), |
| 1754 | zmask_from_V128(argRU), |
| 1755 | 0x0E, False/*!isSTRM*/ |
| 1756 | ); |
| 1757 | assert(ok); |
| 1758 | resECX = resV.uInt[0]; |
| 1759 | return (resOSZACP << 16) | resECX; |
| 1760 | } |
| 1761 | |
| 1762 | void istri_0E ( void ) |
| 1763 | { |
| 1764 | char* wot = "0E"; |
| 1765 | UInt(*h)(V128*,V128*) = h_pcmpistri_0E; |
| 1766 | UInt(*s)(V128*,V128*) = s_pcmpistri_0E; |
| 1767 | |
| 1768 | try_istri(wot,h,s, "111111111abcde11", "00000000000abcde"); |
| 1769 | |
| 1770 | try_istri(wot,h,s, "111111111abcde11", "0000abcde00abcde"); |
| 1771 | |
| 1772 | try_istri(wot,h,s, "1111111111abcde1", "00000000000abcde"); |
| 1773 | try_istri(wot,h,s, "11111111111abcde", "00000000000abcde"); |
| 1774 | try_istri(wot,h,s, "111111111111abcd", "00000000000abcde"); |
| 1775 | |
| 1776 | try_istri(wot,h,s, "111abcde1abcde11", "00000000000abcde"); |
| 1777 | |
| 1778 | try_istri(wot,h,s, "11abcde11abcde11", "00000000000abcde"); |
| 1779 | try_istri(wot,h,s, "1abcde111abcde11", "00000000000abcde"); |
| 1780 | try_istri(wot,h,s, "abcde1111abcde11", "00000000000abcde"); |
| 1781 | try_istri(wot,h,s, "bcde11111abcde11", "00000000000abcde"); |
| 1782 | try_istri(wot,h,s, "cde111111abcde11", "00000000000abcde"); |
| 1783 | |
| 1784 | try_istri(wot,h,s, "01abcde11abcde11", "00000000000abcde"); |
| 1785 | try_istri(wot,h,s, "00abcde11abcde11", "00000000000abcde"); |
| 1786 | try_istri(wot,h,s, "000bcde11abcde11", "00000000000abcde"); |
| 1787 | |
| 1788 | try_istri(wot,h,s, "00abcde10abcde11", "00000000000abcde"); |
| 1789 | try_istri(wot,h,s, "00abcde100bcde11", "00000000000abcde"); |
| 1790 | |
| 1791 | try_istri(wot,h,s, "1111111111111234", "0000000000000000"); |
| 1792 | try_istri(wot,h,s, "1111111111111234", "0000000000000001"); |
| 1793 | try_istri(wot,h,s, "1111111111111234", "0000000000000011"); |
| 1794 | |
| 1795 | try_istri(wot,h,s, "1111111111111234", "1111111111111234"); |
| 1796 | try_istri(wot,h,s, "a111111111111111", "000000000000000a"); |
| 1797 | try_istri(wot,h,s, "b111111111111111", "000000000000000a"); |
| 1798 | |
| 1799 | try_istri(wot,h,s, "b111111111111111", "0000000000000000"); |
| 1800 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1801 | try_istri(wot,h,s, "123456789abcdef1", "0000000000000000"); |
| 1802 | try_istri(wot,h,s, "0000000000000000", "123456789abcdef1"); |
| 1803 | } |
| 1804 | |
| 1805 | |
| 1806 | ////////////////////////////////////////////////////////// |
| 1807 | // // |
| 1808 | // ISTRI_34 // |
| 1809 | // // |
| 1810 | ////////////////////////////////////////////////////////// |
| 1811 | |
| 1812 | UInt h_pcmpistri_34 ( V128* argL, V128* argR ) |
| 1813 | { |
| 1814 | V128 block[2]; |
| 1815 | memcpy(&block[0], argL, sizeof(V128)); |
| 1816 | memcpy(&block[1], argR, sizeof(V128)); |
| 1817 | ULong res, flags; |
| 1818 | __asm__ __volatile__( |
| 1819 | "subq $1024, %%rsp" "\n\t" |
| 1820 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1821 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1822 | "pcmpistri $0x34, %%xmm2, %%xmm11" "\n\t" |
| 1823 | "pushfq" "\n\t" |
| 1824 | "popq %%rdx" "\n\t" |
| 1825 | "movq %%rcx, %0" "\n\t" |
| 1826 | "movq %%rdx, %1" "\n\t" |
| 1827 | "addq $1024, %%rsp" "\n\t" |
| 1828 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1829 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1830 | ); |
| 1831 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1832 | } |
| 1833 | |
| 1834 | UInt s_pcmpistri_34 ( V128* argLU, V128* argRU ) |
| 1835 | { |
| 1836 | V128 resV; |
| 1837 | UInt resOSZACP, resECX; |
| 1838 | Bool ok |
| 1839 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1840 | zmask_from_V128(argLU), |
| 1841 | zmask_from_V128(argRU), |
| 1842 | 0x34, False/*!isSTRM*/ |
| 1843 | ); |
| 1844 | assert(ok); |
| 1845 | resECX = resV.uInt[0]; |
| 1846 | return (resOSZACP << 16) | resECX; |
| 1847 | } |
| 1848 | |
| 1849 | void istri_34 ( void ) |
| 1850 | { |
| 1851 | char* wot = "34"; |
| 1852 | UInt(*h)(V128*,V128*) = h_pcmpistri_34; |
| 1853 | UInt(*s)(V128*,V128*) = s_pcmpistri_34; |
| 1854 | |
| 1855 | try_istri(wot,h,s, "aaaabbbbccccdddd", "00000000000000bc"); |
| 1856 | try_istri(wot,h,s, "aaaabbbbccccdddd", "00000000000000cb"); |
| 1857 | try_istri(wot,h,s, "baaabbbbccccdddd", "00000000000000cb"); |
| 1858 | try_istri(wot,h,s, "baaabbbbccccdddc", "00000000000000cb"); |
| 1859 | |
| 1860 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "00000000000000cb"); |
| 1861 | try_istri(wot,h,s, "bbbbbbbb0bbbbbbb", "00000000000000cb"); |
| 1862 | try_istri(wot,h,s, "bbbbbbbbbbbbbb0b", "00000000000000cb"); |
| 1863 | try_istri(wot,h,s, "bbbbbbbbbbbbbbb0", "00000000000000cb"); |
| 1864 | try_istri(wot,h,s, "0000000000000000", "00000000000000cb"); |
| 1865 | |
| 1866 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1867 | |
| 1868 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "00000000000000cb"); |
| 1869 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "000000000000000b"); |
| 1870 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000062cb"); |
| 1871 | |
| 1872 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000002cb"); |
| 1873 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000000cb"); |
| 1874 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "000000000000000b"); |
| 1875 | |
| 1876 | try_istri(wot,h,s, "0123456789abcdef", "000000fecb975421"); |
| 1877 | try_istri(wot,h,s, "123456789abcdef1", "000000fecb975421"); |
| 1878 | |
| 1879 | try_istri(wot,h,s, "0123456789abcdef", "00000000dca86532"); |
| 1880 | try_istri(wot,h,s, "123456789abcdef1", "00000000dca86532"); |
| 1881 | |
| 1882 | try_istri(wot,h,s, "163887ec041a9b72", "fcd75adb9b3e895a"); |
| 1883 | try_istri(wot,h,s, "fc937cbfbf53f8e2", "0d136bcb024d3fb7"); |
| 1884 | try_istri(wot,h,s, "2ca34182c29a82ab", "302ebd646775ab54"); |
| 1885 | try_istri(wot,h,s, "3f2987608c11be6f", "a9ecb661f8e0a8cb"); |
| 1886 | } |
| 1887 | |
| 1888 | |
| 1889 | ////////////////////////////////////////////////////////// |
| 1890 | // // |
| 1891 | // ISTRI_14 // |
| 1892 | // // |
| 1893 | ////////////////////////////////////////////////////////// |
| 1894 | |
| 1895 | UInt h_pcmpistri_14 ( V128* argL, V128* argR ) |
| 1896 | { |
| 1897 | V128 block[2]; |
| 1898 | memcpy(&block[0], argL, sizeof(V128)); |
| 1899 | memcpy(&block[1], argR, sizeof(V128)); |
| 1900 | ULong res, flags; |
| 1901 | __asm__ __volatile__( |
| 1902 | "subq $1024, %%rsp" "\n\t" |
| 1903 | "movdqu 0(%2), %%xmm2" "\n\t" |
| 1904 | "movdqu 16(%2), %%xmm11" "\n\t" |
| 1905 | "pcmpistri $0x14, %%xmm2, %%xmm11" "\n\t" |
| 1906 | "pushfq" "\n\t" |
| 1907 | "popq %%rdx" "\n\t" |
| 1908 | "movq %%rcx, %0" "\n\t" |
| 1909 | "movq %%rdx, %1" "\n\t" |
| 1910 | "addq $1024, %%rsp" "\n\t" |
| 1911 | : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0]) |
| 1912 | : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory" |
| 1913 | ); |
| 1914 | return ((flags & 0x8D5) << 16) | (res & 0xFFFF); |
| 1915 | } |
| 1916 | |
| 1917 | UInt s_pcmpistri_14 ( V128* argLU, V128* argRU ) |
| 1918 | { |
| 1919 | V128 resV; |
| 1920 | UInt resOSZACP, resECX; |
| 1921 | Bool ok |
| 1922 | = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU, |
| 1923 | zmask_from_V128(argLU), |
| 1924 | zmask_from_V128(argRU), |
| 1925 | 0x14, False/*!isSTRM*/ |
| 1926 | ); |
| 1927 | assert(ok); |
| 1928 | resECX = resV.uInt[0]; |
| 1929 | return (resOSZACP << 16) | resECX; |
| 1930 | } |
| 1931 | |
| 1932 | void istri_14 ( void ) |
| 1933 | { |
| 1934 | char* wot = "14"; |
| 1935 | UInt(*h)(V128*,V128*) = h_pcmpistri_14; |
| 1936 | UInt(*s)(V128*,V128*) = s_pcmpistri_14; |
| 1937 | |
| 1938 | try_istri(wot,h,s, "aaaabbbbccccdddd", "00000000000000bc"); |
| 1939 | try_istri(wot,h,s, "aaaabbbbccccdddd", "00000000000000cb"); |
| 1940 | try_istri(wot,h,s, "baaabbbbccccdddd", "00000000000000cb"); |
| 1941 | try_istri(wot,h,s, "baaabbbbccccdddc", "00000000000000cb"); |
| 1942 | |
| 1943 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "00000000000000cb"); |
| 1944 | try_istri(wot,h,s, "bbbbbbbb0bbbbbbb", "00000000000000cb"); |
| 1945 | try_istri(wot,h,s, "bbbbbbbbbbbbbb0b", "00000000000000cb"); |
| 1946 | try_istri(wot,h,s, "bbbbbbbbbbbbbbb0", "00000000000000cb"); |
| 1947 | try_istri(wot,h,s, "0000000000000000", "00000000000000cb"); |
| 1948 | |
| 1949 | try_istri(wot,h,s, "0000000000000000", "0000000000000000"); |
| 1950 | |
| 1951 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "00000000000000cb"); |
| 1952 | try_istri(wot,h,s, "bbbbbbbbbbbbbbbb", "000000000000000b"); |
| 1953 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000062cb"); |
| 1954 | |
| 1955 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000002cb"); |
| 1956 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "00000000000000cb"); |
| 1957 | try_istri(wot,h,s, "b4b4b4b4b4b4b4b4", "000000000000000b"); |
| 1958 | |
| 1959 | try_istri(wot,h,s, "0123456789abcdef", "000000fecb975421"); |
| 1960 | try_istri(wot,h,s, "123456789abcdef1", "000000fecb975421"); |
| 1961 | |
| 1962 | try_istri(wot,h,s, "0123456789abcdef", "00000000dca86532"); |
| 1963 | try_istri(wot,h,s, "123456789abcdef1", "00000000dca86532"); |
| 1964 | |
| 1965 | try_istri(wot,h,s, "163887ec041a9b72", "fcd75adb9b3e895a"); |
| 1966 | try_istri(wot,h,s, "fc937cbfbf53f8e2", "0d136bcb024d3fb7"); |
| 1967 | try_istri(wot,h,s, "2ca34182c29a82ab", "302ebd646775ab54"); |
| 1968 | try_istri(wot,h,s, "3f2987608c11be6f", "a9ecb661f8e0a8cb"); |
| 1969 | } |
| 1970 | |
| 1971 | |
| 1972 | ////////////////////////////////////////////////////////// |
| 1973 | // // |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 1974 | // main // |
| 1975 | // // |
| 1976 | ////////////////////////////////////////////////////////// |
| 1977 | |
| 1978 | int main ( void ) |
| 1979 | { |
| 1980 | istri_4A(); |
| 1981 | istri_3A(); |
| 1982 | istri_08(); |
sewardj | 053f436 | 2015-08-13 14:19:48 +0000 | [diff] [blame] | 1983 | istri_18(); |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 1984 | istri_1A(); |
| 1985 | istri_02(); |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 1986 | istri_0C(); |
sewardj | 5ac9906 | 2010-08-17 00:19:25 +0000 | [diff] [blame] | 1987 | istri_12(); |
| 1988 | istri_44(); |
sewardj | 3b20b17 | 2011-01-17 23:10:39 +0000 | [diff] [blame] | 1989 | istri_00(); |
sewardj | e801ed2 | 2011-10-19 20:17:50 +0000 | [diff] [blame] | 1990 | istri_38(); |
sewardj | 15df336 | 2012-09-17 13:43:57 +0000 | [diff] [blame] | 1991 | istri_46(); |
sewardj | 473d0ba | 2013-10-18 10:46:52 +0000 | [diff] [blame] | 1992 | istri_30(); |
| 1993 | istri_40(); |
sewardj | 7f0dfdd | 2015-08-13 13:47:02 +0000 | [diff] [blame] | 1994 | istri_42(); |
sewardj | a2bf389 | 2014-05-21 14:43:11 +0000 | [diff] [blame] | 1995 | istri_0E(); |
| 1996 | istri_14(); |
| 1997 | istri_34(); |
sewardj | 0a87e8d | 2010-08-06 08:10:45 +0000 | [diff] [blame] | 1998 | return 0; |
| 1999 | } |