1 /* 2 * Copyright (c) 2007, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 */ 23 24 #include "precompiled.hpp" 25 #include "memory/allocation.inline.hpp" 26 #include "opto/connode.hpp" 27 #include "opto/vectornode.hpp" 28 29 //------------------------------VectorNode-------------------------------------- 30 31 // Return the vector operator for the specified scalar operation 32 // and vector length. 33 int VectorNode::opcode(int sopc, BasicType bt) { 34 switch (sopc) { 35 case Op_AddI: 36 switch (bt) { 37 case T_BOOLEAN: 38 case T_BYTE: return Op_AddVB; 39 case T_CHAR: 40 case T_SHORT: return Op_AddVS; 41 case T_INT: return Op_AddVI; 42 } 43 ShouldNotReachHere(); 44 case Op_AddL: 45 assert(bt == T_LONG, "must be"); 46 return Op_AddVL; 47 case Op_AddF: 48 assert(bt == T_FLOAT, "must be"); 49 return Op_AddVF; 50 case Op_AddD: 51 assert(bt == T_DOUBLE, "must be"); 52 return Op_AddVD; 53 case Op_SubI: 54 switch (bt) { 55 case T_BOOLEAN: 56 case T_BYTE: return Op_SubVB; 57 case T_CHAR: 58 case T_SHORT: return Op_SubVS; 59 case T_INT: return Op_SubVI; 60 } 61 ShouldNotReachHere(); 62 case Op_SubL: 63 assert(bt == T_LONG, "must be"); 64 return Op_SubVL; 65 case Op_SubF: 66 assert(bt == T_FLOAT, "must be"); 67 return Op_SubVF; 68 case Op_SubD: 69 assert(bt == T_DOUBLE, "must be"); 70 return Op_SubVD; 71 case Op_MulI: 72 switch (bt) { 73 case T_BOOLEAN: 74 case T_BYTE: return 0; // Unimplemented 75 case T_CHAR: 76 case T_SHORT: return Op_MulVS; 77 case T_INT: return Op_MulVI; 78 } 79 ShouldNotReachHere(); 80 case Op_MulL: 81 assert(bt == T_LONG, "must be"); 82 return Op_MulVL; 83 case Op_MulF: 84 assert(bt == T_FLOAT, "must be"); 85 return Op_MulVF; 86 case Op_MulD: 87 assert(bt == T_DOUBLE, "must be"); 88 return Op_MulVD; 89 case Op_FmaD: 90 assert(bt == T_DOUBLE, "must be"); 91 return Op_FmaVD; 92 case Op_FmaF: 93 assert(bt == T_FLOAT, "must be"); 94 return Op_FmaVF; 95 case Op_CMoveD: 96 assert(bt == T_DOUBLE, "must be"); 97 return Op_CMoveVD; 98 case Op_DivF: 99 assert(bt == T_FLOAT, "must be"); 100 return Op_DivVF; 101 case Op_DivD: 102 assert(bt == T_DOUBLE, "must be"); 103 return Op_DivVD; 104 case Op_AbsF: 105 assert(bt == T_FLOAT, "must be"); 106 return Op_AbsVF; 107 case Op_AbsD: 108 assert(bt == T_DOUBLE, "must be"); 109 return Op_AbsVD; 110 case Op_NegF: 111 assert(bt == T_FLOAT, "must be"); 112 return Op_NegVF; 113 case Op_NegD: 114 assert(bt == T_DOUBLE, "must be"); 115 return Op_NegVD; 116 case Op_SqrtD: 117 assert(bt == T_DOUBLE, "must be"); 118 return Op_SqrtVD; 119 case Op_LShiftI: 120 switch (bt) { 121 case T_BOOLEAN: 122 case T_BYTE: return Op_LShiftVB; 123 case T_CHAR: 124 case T_SHORT: return Op_LShiftVS; 125 case T_INT: return Op_LShiftVI; 126 } 127 ShouldNotReachHere(); 128 case Op_LShiftL: 129 assert(bt == T_LONG, "must be"); 130 return Op_LShiftVL; 131 case Op_RShiftI: 132 switch (bt) { 133 case T_BOOLEAN:return Op_URShiftVB; // boolean is unsigned value 134 case T_CHAR: return Op_URShiftVS; // char is unsigned value 135 case T_BYTE: return Op_RShiftVB; 136 case T_SHORT: return Op_RShiftVS; 137 case T_INT: return Op_RShiftVI; 138 } 139 ShouldNotReachHere(); 140 case Op_RShiftL: 141 assert(bt == T_LONG, "must be"); 142 return Op_RShiftVL; 143 case Op_URShiftI: 144 switch (bt) { 145 case T_BOOLEAN:return Op_URShiftVB; 146 case T_CHAR: return Op_URShiftVS; 147 case T_BYTE: 148 case T_SHORT: return 0; // Vector logical right shift for signed short 149 // values produces incorrect Java result for 150 // negative data because java code should convert 151 // a short value into int value with sign 152 // extension before a shift. 153 case T_INT: return Op_URShiftVI; 154 } 155 ShouldNotReachHere(); 156 case Op_URShiftL: 157 assert(bt == T_LONG, "must be"); 158 return Op_URShiftVL; 159 case Op_AndI: 160 case Op_AndL: 161 return Op_AndV; 162 case Op_OrI: 163 case Op_OrL: 164 return Op_OrV; 165 case Op_XorI: 166 case Op_XorL: 167 return Op_XorV; 168 169 case Op_LoadB: 170 case Op_LoadUB: 171 case Op_LoadUS: 172 case Op_LoadS: 173 case Op_LoadI: 174 case Op_LoadL: 175 case Op_LoadF: 176 case Op_LoadD: 177 return Op_LoadVector; 178 179 case Op_StoreB: 180 case Op_StoreC: 181 case Op_StoreI: 182 case Op_StoreL: 183 case Op_StoreF: 184 case Op_StoreD: 185 return Op_StoreVector; 186 } 187 return 0; // Unimplemented 188 } 189 190 // Also used to check if the code generator 191 // supports the vector operation. 192 bool VectorNode::implemented(int opc, uint vlen, BasicType bt) { 193 if (is_java_primitive(bt) && 194 (vlen > 1) && is_power_of_2(vlen) && 195 Matcher::vector_size_supported(bt, vlen)) { 196 int vopc = VectorNode::opcode(opc, bt); 197 return vopc > 0 && Matcher::match_rule_supported_vector(vopc, vlen); 198 } 199 return false; 200 } 201 202 bool VectorNode::is_shift(Node* n) { 203 switch (n->Opcode()) { 204 case Op_LShiftI: 205 case Op_LShiftL: 206 case Op_RShiftI: 207 case Op_RShiftL: 208 case Op_URShiftI: 209 case Op_URShiftL: 210 return true; 211 } 212 return false; 213 } 214 215 // Check if input is loop invariant vector. 216 bool VectorNode::is_invariant_vector(Node* n) { 217 // Only Replicate vector nodes are loop invariant for now. 218 switch (n->Opcode()) { 219 case Op_ReplicateB: 220 case Op_ReplicateS: 221 case Op_ReplicateI: 222 case Op_ReplicateL: 223 case Op_ReplicateF: 224 case Op_ReplicateD: 225 return true; 226 } 227 return false; 228 } 229 230 // [Start, end) half-open range defining which operands are vectors 231 void VectorNode::vector_operands(Node* n, uint* start, uint* end) { 232 switch (n->Opcode()) { 233 case Op_LoadB: case Op_LoadUB: 234 case Op_LoadS: case Op_LoadUS: 235 case Op_LoadI: case Op_LoadL: 236 case Op_LoadF: case Op_LoadD: 237 case Op_LoadP: case Op_LoadN: 238 *start = 0; 239 *end = 0; // no vector operands 240 break; 241 case Op_StoreB: case Op_StoreC: 242 case Op_StoreI: case Op_StoreL: 243 case Op_StoreF: case Op_StoreD: 244 case Op_StoreP: case Op_StoreN: 245 *start = MemNode::ValueIn; 246 *end = MemNode::ValueIn + 1; // 1 vector operand 247 break; 248 case Op_LShiftI: case Op_LShiftL: 249 case Op_RShiftI: case Op_RShiftL: 250 case Op_URShiftI: case Op_URShiftL: 251 *start = 1; 252 *end = 2; // 1 vector operand 253 break; 254 case Op_AddI: case Op_AddL: case Op_AddF: case Op_AddD: 255 case Op_SubI: case Op_SubL: case Op_SubF: case Op_SubD: 256 case Op_MulI: case Op_MulL: case Op_MulF: case Op_MulD: 257 case Op_DivF: case Op_DivD: 258 case Op_AndI: case Op_AndL: 259 case Op_OrI: case Op_OrL: 260 case Op_XorI: case Op_XorL: 261 *start = 1; 262 *end = 3; // 2 vector operands 263 break; 264 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD: 265 *start = 2; 266 *end = n->req(); 267 break; 268 case Op_FmaD: 269 case Op_FmaF: 270 *start = 1; 271 *end = 4; // 3 vector operands 272 break; 273 default: 274 *start = 1; 275 *end = n->req(); // default is all operands 276 } 277 } 278 279 // Return the vector version of a scalar operation node. 280 VectorNode* VectorNode::make(int opc, Node* n1, Node* n2, uint vlen, BasicType bt) { 281 const TypeVect* vt = TypeVect::make(bt, vlen); 282 int vopc = VectorNode::opcode(opc, bt); 283 // This method should not be called for unimplemented vectors. 284 guarantee(vopc > 0, "Vector for '%s' is not implemented", NodeClassNames[opc]); 285 switch (vopc) { 286 case Op_AddVB: return new AddVBNode(n1, n2, vt); 287 case Op_AddVS: return new AddVSNode(n1, n2, vt); 288 case Op_AddVI: return new AddVINode(n1, n2, vt); 289 case Op_AddVL: return new AddVLNode(n1, n2, vt); 290 case Op_AddVF: return new AddVFNode(n1, n2, vt); 291 case Op_AddVD: return new AddVDNode(n1, n2, vt); 292 293 case Op_SubVB: return new SubVBNode(n1, n2, vt); 294 case Op_SubVS: return new SubVSNode(n1, n2, vt); 295 case Op_SubVI: return new SubVINode(n1, n2, vt); 296 case Op_SubVL: return new SubVLNode(n1, n2, vt); 297 case Op_SubVF: return new SubVFNode(n1, n2, vt); 298 case Op_SubVD: return new SubVDNode(n1, n2, vt); 299 300 case Op_MulVS: return new MulVSNode(n1, n2, vt); 301 case Op_MulVI: return new MulVINode(n1, n2, vt); 302 case Op_MulVL: return new MulVLNode(n1, n2, vt); 303 case Op_MulVF: return new MulVFNode(n1, n2, vt); 304 case Op_MulVD: return new MulVDNode(n1, n2, vt); 305 306 case Op_DivVF: return new DivVFNode(n1, n2, vt); 307 case Op_DivVD: return new DivVDNode(n1, n2, vt); 308 309 case Op_AbsVF: return new AbsVFNode(n1, vt); 310 case Op_AbsVD: return new AbsVDNode(n1, vt); 311 312 case Op_NegVF: return new NegVFNode(n1, vt); 313 case Op_NegVD: return new NegVDNode(n1, vt); 314 315 // Currently only supports double precision sqrt 316 case Op_SqrtVD: return new SqrtVDNode(n1, vt); 317 318 case Op_LShiftVB: return new LShiftVBNode(n1, n2, vt); 319 case Op_LShiftVS: return new LShiftVSNode(n1, n2, vt); 320 case Op_LShiftVI: return new LShiftVINode(n1, n2, vt); 321 case Op_LShiftVL: return new LShiftVLNode(n1, n2, vt); 322 323 case Op_RShiftVB: return new RShiftVBNode(n1, n2, vt); 324 case Op_RShiftVS: return new RShiftVSNode(n1, n2, vt); 325 case Op_RShiftVI: return new RShiftVINode(n1, n2, vt); 326 case Op_RShiftVL: return new RShiftVLNode(n1, n2, vt); 327 328 case Op_URShiftVB: return new URShiftVBNode(n1, n2, vt); 329 case Op_URShiftVS: return new URShiftVSNode(n1, n2, vt); 330 case Op_URShiftVI: return new URShiftVINode(n1, n2, vt); 331 case Op_URShiftVL: return new URShiftVLNode(n1, n2, vt); 332 333 case Op_AndV: return new AndVNode(n1, n2, vt); 334 case Op_OrV: return new OrVNode (n1, n2, vt); 335 case Op_XorV: return new XorVNode(n1, n2, vt); 336 } 337 fatal("Missed vector creation for '%s'", NodeClassNames[vopc]); 338 return NULL; 339 340 } 341 342 VectorNode* VectorNode::make(int opc, Node* n1, Node* n2, Node* n3, uint vlen, BasicType bt) { 343 const TypeVect* vt = TypeVect::make(bt, vlen); 344 int vopc = VectorNode::opcode(opc, bt); 345 // This method should not be called for unimplemented vectors. 346 guarantee(vopc > 0, "Vector for '%s' is not implemented", NodeClassNames[opc]); 347 switch (vopc) { 348 case Op_FmaVD: return new FmaVDNode(n1, n2, n3, vt); 349 case Op_FmaVF: return new FmaVFNode(n1, n2, n3, vt); 350 } 351 fatal("Missed vector creation for '%s'", NodeClassNames[vopc]); 352 return NULL; 353 } 354 355 // Scalar promotion 356 VectorNode* VectorNode::scalar2vector(Node* s, uint vlen, const Type* opd_t) { 357 BasicType bt = opd_t->array_element_basic_type(); 358 const TypeVect* vt = opd_t->singleton() ? TypeVect::make(opd_t, vlen) 359 : TypeVect::make(bt, vlen); 360 switch (bt) { 361 case T_BOOLEAN: 362 case T_BYTE: 363 return new ReplicateBNode(s, vt); 364 case T_CHAR: 365 case T_SHORT: 366 return new ReplicateSNode(s, vt); 367 case T_INT: 368 return new ReplicateINode(s, vt); 369 case T_LONG: 370 return new ReplicateLNode(s, vt); 371 case T_FLOAT: 372 return new ReplicateFNode(s, vt); 373 case T_DOUBLE: 374 return new ReplicateDNode(s, vt); 375 } 376 fatal("Type '%s' is not supported for vectors", type2name(bt)); 377 return NULL; 378 } 379 380 VectorNode* VectorNode::shift_count(Node* shift, Node* cnt, uint vlen, BasicType bt) { 381 assert(VectorNode::is_shift(shift) && !cnt->is_Con(), "only variable shift count"); 382 // Match shift count type with shift vector type. 383 const TypeVect* vt = TypeVect::make(bt, vlen); 384 switch (shift->Opcode()) { 385 case Op_LShiftI: 386 case Op_LShiftL: 387 return new LShiftCntVNode(cnt, vt); 388 case Op_RShiftI: 389 case Op_RShiftL: 390 case Op_URShiftI: 391 case Op_URShiftL: 392 return new RShiftCntVNode(cnt, vt); 393 } 394 fatal("Missed vector creation for '%s'", NodeClassNames[shift->Opcode()]); 395 return NULL; 396 } 397 398 // Return initial Pack node. Additional operands added with add_opd() calls. 399 PackNode* PackNode::make(Node* s, uint vlen, BasicType bt) { 400 const TypeVect* vt = TypeVect::make(bt, vlen); 401 switch (bt) { 402 case T_BOOLEAN: 403 case T_BYTE: 404 return new PackBNode(s, vt); 405 case T_CHAR: 406 case T_SHORT: 407 return new PackSNode(s, vt); 408 case T_INT: 409 return new PackINode(s, vt); 410 case T_LONG: 411 return new PackLNode(s, vt); 412 case T_FLOAT: 413 return new PackFNode(s, vt); 414 case T_DOUBLE: 415 return new PackDNode(s, vt); 416 } 417 fatal("Type '%s' is not supported for vectors", type2name(bt)); 418 return NULL; 419 } 420 421 // Create a binary tree form for Packs. [lo, hi) (half-open) range 422 PackNode* PackNode::binary_tree_pack(int lo, int hi) { 423 int ct = hi - lo; 424 assert(is_power_of_2(ct), "power of 2"); 425 if (ct == 2) { 426 PackNode* pk = PackNode::make(in(lo), 2, vect_type()->element_basic_type()); 427 pk->add_opd(in(lo+1)); 428 return pk; 429 430 } else { 431 int mid = lo + ct/2; 432 PackNode* n1 = binary_tree_pack(lo, mid); 433 PackNode* n2 = binary_tree_pack(mid, hi ); 434 435 BasicType bt = n1->vect_type()->element_basic_type(); 436 assert(bt == n2->vect_type()->element_basic_type(), "should be the same"); 437 switch (bt) { 438 case T_BOOLEAN: 439 case T_BYTE: 440 return new PackSNode(n1, n2, TypeVect::make(T_SHORT, 2)); 441 case T_CHAR: 442 case T_SHORT: 443 return new PackINode(n1, n2, TypeVect::make(T_INT, 2)); 444 case T_INT: 445 return new PackLNode(n1, n2, TypeVect::make(T_LONG, 2)); 446 case T_LONG: 447 return new Pack2LNode(n1, n2, TypeVect::make(T_LONG, 2)); 448 case T_FLOAT: 449 return new PackDNode(n1, n2, TypeVect::make(T_DOUBLE, 2)); 450 case T_DOUBLE: 451 return new Pack2DNode(n1, n2, TypeVect::make(T_DOUBLE, 2)); 452 } 453 fatal("Type '%s' is not supported for vectors", type2name(bt)); 454 } 455 return NULL; 456 } 457 458 // Return the vector version of a scalar load node. 459 LoadVectorNode* LoadVectorNode::make(int opc, Node* ctl, Node* mem, 460 Node* adr, const TypePtr* atyp, 461 uint vlen, BasicType bt, 462 ControlDependency control_dependency) { 463 const TypeVect* vt = TypeVect::make(bt, vlen); 464 return new LoadVectorNode(ctl, mem, adr, atyp, vt, control_dependency); 465 } 466 467 // Return the vector version of a scalar store node. 468 StoreVectorNode* StoreVectorNode::make(int opc, Node* ctl, Node* mem, 469 Node* adr, const TypePtr* atyp, Node* val, 470 uint vlen) { 471 return new StoreVectorNode(ctl, mem, adr, atyp, val); 472 } 473 474 // Extract a scalar element of vector. 475 Node* ExtractNode::make(Node* v, uint position, BasicType bt) { 476 assert((int)position < Matcher::max_vector_size(bt), "pos in range"); 477 ConINode* pos = ConINode::make((int)position); 478 switch (bt) { 479 case T_BOOLEAN: 480 return new ExtractUBNode(v, pos); 481 case T_BYTE: 482 return new ExtractBNode(v, pos); 483 case T_CHAR: 484 return new ExtractCNode(v, pos); 485 case T_SHORT: 486 return new ExtractSNode(v, pos); 487 case T_INT: 488 return new ExtractINode(v, pos); 489 case T_LONG: 490 return new ExtractLNode(v, pos); 491 case T_FLOAT: 492 return new ExtractFNode(v, pos); 493 case T_DOUBLE: 494 return new ExtractDNode(v, pos); 495 } 496 fatal("Type '%s' is not supported for vectors", type2name(bt)); 497 return NULL; 498 } 499 500 int ReductionNode::opcode(int opc, BasicType bt) { 501 int vopc = opc; 502 switch (opc) { 503 case Op_AddI: 504 assert(bt == T_INT, "must be"); 505 vopc = Op_AddReductionVI; 506 break; 507 case Op_AddL: 508 assert(bt == T_LONG, "must be"); 509 vopc = Op_AddReductionVL; 510 break; 511 case Op_AddF: 512 assert(bt == T_FLOAT, "must be"); 513 vopc = Op_AddReductionVF; 514 break; 515 case Op_AddD: 516 assert(bt == T_DOUBLE, "must be"); 517 vopc = Op_AddReductionVD; 518 break; 519 case Op_MulI: 520 assert(bt == T_INT, "must be"); 521 vopc = Op_MulReductionVI; 522 break; 523 case Op_MulL: 524 assert(bt == T_LONG, "must be"); 525 vopc = Op_MulReductionVL; 526 break; 527 case Op_MulF: 528 assert(bt == T_FLOAT, "must be"); 529 vopc = Op_MulReductionVF; 530 break; 531 case Op_MulD: 532 assert(bt == T_DOUBLE, "must be"); 533 vopc = Op_MulReductionVD; 534 break; 535 // TODO: add MulL for targets that support it 536 default: 537 break; 538 } 539 return vopc; 540 } 541 542 // Return the appropriate reduction node. 543 ReductionNode* ReductionNode::make(int opc, Node *ctrl, Node* n1, Node* n2, BasicType bt) { 544 545 int vopc = opcode(opc, bt); 546 547 // This method should not be called for unimplemented vectors. 548 guarantee(vopc != opc, "Vector for '%s' is not implemented", NodeClassNames[opc]); 549 550 switch (vopc) { 551 case Op_AddReductionVI: return new AddReductionVINode(ctrl, n1, n2); 552 case Op_AddReductionVL: return new AddReductionVLNode(ctrl, n1, n2); 553 case Op_AddReductionVF: return new AddReductionVFNode(ctrl, n1, n2); 554 case Op_AddReductionVD: return new AddReductionVDNode(ctrl, n1, n2); 555 case Op_MulReductionVI: return new MulReductionVINode(ctrl, n1, n2); 556 case Op_MulReductionVL: return new MulReductionVLNode(ctrl, n1, n2); 557 case Op_MulReductionVF: return new MulReductionVFNode(ctrl, n1, n2); 558 case Op_MulReductionVD: return new MulReductionVDNode(ctrl, n1, n2); 559 } 560 fatal("Missed vector creation for '%s'", NodeClassNames[vopc]); 561 return NULL; 562 } 563 564 bool ReductionNode::implemented(int opc, uint vlen, BasicType bt) { 565 if (is_java_primitive(bt) && 566 (vlen > 1) && is_power_of_2(vlen) && 567 Matcher::vector_size_supported(bt, vlen)) { 568 int vopc = ReductionNode::opcode(opc, bt); 569 return vopc != opc && Matcher::match_rule_supported(vopc); 570 } 571 return false; 572 } 573