1 /* 2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2014, 2024, Red Hat Inc. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #ifndef CPU_AARCH64_ASSEMBLER_AARCH64_HPP 27 #define CPU_AARCH64_ASSEMBLER_AARCH64_HPP 28 29 #include "asm/register.hpp" 30 #include "metaprogramming/enableIf.hpp" 31 #include "utilities/checkedCast.hpp" 32 #include "utilities/debug.hpp" 33 #include "utilities/globalDefinitions.hpp" 34 #include "utilities/macros.hpp" 35 #include <type_traits> 36 37 #ifdef __GNUC__ 38 39 // __nop needs volatile so that compiler doesn't optimize it away 40 #define NOP() asm volatile ("nop"); 41 42 #elif defined(_MSC_VER) 43 44 // Use MSVC intrinsic: https://docs.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=vs-2019#I 45 #define NOP() __nop(); 46 47 #endif 48 49 50 // definitions of various symbolic names for machine registers 51 52 // First intercalls between C and Java which use 8 general registers 53 // and 8 floating registers 54 55 // we also have to copy between x86 and ARM registers but that's a 56 // secondary complication -- not all code employing C call convention 57 // executes as x86 code though -- we generate some of it 58 59 class Argument { 60 public: 61 enum { 62 n_int_register_parameters_c = 8, // r0, r1, ... r7 (c_rarg0, c_rarg1, ...) 63 n_float_register_parameters_c = 8, // v0, v1, ... v7 (c_farg0, c_farg1, ... ) 64 65 n_int_register_parameters_j = 8, // r1, ... r7, r0 (rj_rarg0, j_rarg1, ... 66 n_float_register_parameters_j = 8 // v0, v1, ... v7 (j_farg0, j_farg1, ... 67 }; 68 }; 69 70 constexpr Register c_rarg0 = r0; 71 constexpr Register c_rarg1 = r1; 72 constexpr Register c_rarg2 = r2; 73 constexpr Register c_rarg3 = r3; 74 constexpr Register c_rarg4 = r4; 75 constexpr Register c_rarg5 = r5; 76 constexpr Register c_rarg6 = r6; 77 constexpr Register c_rarg7 = r7; 78 79 constexpr FloatRegister c_farg0 = v0; 80 constexpr FloatRegister c_farg1 = v1; 81 constexpr FloatRegister c_farg2 = v2; 82 constexpr FloatRegister c_farg3 = v3; 83 constexpr FloatRegister c_farg4 = v4; 84 constexpr FloatRegister c_farg5 = v5; 85 constexpr FloatRegister c_farg6 = v6; 86 constexpr FloatRegister c_farg7 = v7; 87 88 // Symbolically name the register arguments used by the Java calling convention. 89 // We have control over the convention for java so we can do what we please. 90 // What pleases us is to offset the java calling convention so that when 91 // we call a suitable jni method the arguments are lined up and we don't 92 // have to do much shuffling. A suitable jni method is non-static and a 93 // small number of arguments 94 // 95 // |--------------------------------------------------------------------| 96 // | c_rarg0 c_rarg1 c_rarg2 c_rarg3 c_rarg4 c_rarg5 c_rarg6 c_rarg7 | 97 // |--------------------------------------------------------------------| 98 // | r0 r1 r2 r3 r4 r5 r6 r7 | 99 // |--------------------------------------------------------------------| 100 // | j_rarg7 j_rarg0 j_rarg1 j_rarg2 j_rarg3 j_rarg4 j_rarg5 j_rarg6 | 101 // |--------------------------------------------------------------------| 102 103 104 constexpr Register j_rarg0 = c_rarg1; 105 constexpr Register j_rarg1 = c_rarg2; 106 constexpr Register j_rarg2 = c_rarg3; 107 constexpr Register j_rarg3 = c_rarg4; 108 constexpr Register j_rarg4 = c_rarg5; 109 constexpr Register j_rarg5 = c_rarg6; 110 constexpr Register j_rarg6 = c_rarg7; 111 constexpr Register j_rarg7 = c_rarg0; 112 113 // Java floating args are passed as per C 114 115 constexpr FloatRegister j_farg0 = v0; 116 constexpr FloatRegister j_farg1 = v1; 117 constexpr FloatRegister j_farg2 = v2; 118 constexpr FloatRegister j_farg3 = v3; 119 constexpr FloatRegister j_farg4 = v4; 120 constexpr FloatRegister j_farg5 = v5; 121 constexpr FloatRegister j_farg6 = v6; 122 constexpr FloatRegister j_farg7 = v7; 123 124 // registers used to hold VM data either temporarily within a method 125 // or across method calls 126 127 // volatile (caller-save) registers 128 129 // r8 is used for indirect result location return 130 // we use it and r9 as scratch registers 131 constexpr Register rscratch1 = r8; 132 constexpr Register rscratch2 = r9; 133 134 // current method -- must be in a call-clobbered register 135 constexpr Register rmethod = r12; 136 137 // non-volatile (callee-save) registers are r16-29 138 // of which the following are dedicated global state 139 140 constexpr Register lr = r30; // link register 141 constexpr Register rfp = r29; // frame pointer 142 constexpr Register rthread = r28; // current thread 143 constexpr Register rheapbase = r27; // base of heap 144 constexpr Register rcpool = r26; // constant pool cache 145 constexpr Register rlocals = r24; // locals on stack 146 constexpr Register rbcp = r22; // bytecode pointer 147 constexpr Register rdispatch = r21; // dispatch table base 148 constexpr Register esp = r20; // Java expression stack pointer 149 constexpr Register r19_sender_sp = r19; // sender's SP while in interpreter 150 151 // Preserved predicate register with all elements set TRUE. 152 constexpr PRegister ptrue = p7; 153 154 #define assert_cond(ARG1) assert(ARG1, #ARG1) 155 156 namespace asm_util { 157 uint32_t encode_logical_immediate(bool is32, uint64_t imm); 158 uint32_t encode_sve_logical_immediate(unsigned elembits, uint64_t imm); 159 bool operand_valid_for_immediate_bits(int64_t imm, unsigned nbits); 160 }; 161 162 using namespace asm_util; 163 164 165 class Assembler; 166 167 class Instruction_aarch64 { 168 unsigned insn; 169 #ifdef ASSERT 170 unsigned bits; 171 #endif 172 Assembler *assem; 173 174 public: 175 176 Instruction_aarch64(class Assembler *as) { 177 #ifdef ASSERT 178 bits = 0; 179 #endif 180 insn = 0; 181 assem = as; 182 } 183 184 inline ~Instruction_aarch64(); 185 186 unsigned &get_insn() { return insn; } 187 #ifdef ASSERT 188 unsigned &get_bits() { return bits; } 189 #endif 190 191 static inline int32_t extend(unsigned val, int hi = 31, int lo = 0) { 192 union { 193 unsigned u; 194 int n; 195 }; 196 197 u = val << (31 - hi); 198 n = n >> (31 - hi + lo); 199 return n; 200 } 201 202 static inline uint32_t extract(uint32_t val, int msb, int lsb) { 203 int nbits = msb - lsb + 1; 204 assert_cond(msb >= lsb); 205 uint32_t mask = checked_cast<uint32_t>(right_n_bits(nbits)); 206 uint32_t result = val >> lsb; 207 result &= mask; 208 return result; 209 } 210 211 static inline int32_t sextract(uint32_t val, int msb, int lsb) { 212 uint32_t uval = extract(val, msb, lsb); 213 return extend(uval, msb - lsb); 214 } 215 216 static ALWAYSINLINE void patch(address a, int msb, int lsb, uint64_t val) { 217 int nbits = msb - lsb + 1; 218 guarantee(val < (1ULL << nbits), "Field too big for insn"); 219 assert_cond(msb >= lsb); 220 unsigned mask = checked_cast<unsigned>(right_n_bits(nbits)); 221 val <<= lsb; 222 mask <<= lsb; 223 unsigned target = *(unsigned *)a; 224 target &= ~mask; 225 target |= (unsigned)val; 226 *(unsigned *)a = target; 227 } 228 229 static void spatch(address a, int msb, int lsb, int64_t val) { 230 int nbits = msb - lsb + 1; 231 int64_t chk = val >> (nbits - 1); 232 guarantee (chk == -1 || chk == 0, "Field too big for insn at " INTPTR_FORMAT, p2i(a)); 233 uint64_t uval = val; 234 unsigned mask = checked_cast<unsigned>(right_n_bits(nbits)); 235 uval &= mask; 236 uval <<= lsb; 237 mask <<= lsb; 238 unsigned target = *(unsigned *)a; 239 target &= ~mask; 240 target |= (unsigned)uval; 241 *(unsigned *)a = target; 242 } 243 244 void f(unsigned val, int msb, int lsb) { 245 int nbits = msb - lsb + 1; 246 guarantee(val < (1ULL << nbits), "Field too big for insn"); 247 assert_cond(msb >= lsb); 248 val <<= lsb; 249 insn |= val; 250 #ifdef ASSERT 251 unsigned mask = checked_cast<unsigned>(right_n_bits(nbits)); 252 mask <<= lsb; 253 assert_cond((bits & mask) == 0); 254 bits |= mask; 255 #endif 256 } 257 258 void f(unsigned val, int bit) { 259 f(val, bit, bit); 260 } 261 262 void sf(int64_t val, int msb, int lsb) { 263 int nbits = msb - lsb + 1; 264 int64_t chk = val >> (nbits - 1); 265 guarantee (chk == -1 || chk == 0, "Field too big for insn"); 266 uint64_t uval = val; 267 unsigned mask = checked_cast<unsigned>(right_n_bits(nbits)); 268 uval &= mask; 269 f((unsigned)uval, lsb + nbits - 1, lsb); 270 } 271 272 void rf(Register r, int lsb) { 273 f(r->raw_encoding(), lsb + 4, lsb); 274 } 275 276 // reg|ZR 277 void zrf(Register r, int lsb) { 278 f(r->raw_encoding() - (r == zr), lsb + 4, lsb); 279 } 280 281 // reg|SP 282 void srf(Register r, int lsb) { 283 f(r == sp ? 31 : r->raw_encoding(), lsb + 4, lsb); 284 } 285 286 void rf(FloatRegister r, int lsb) { 287 f(r->raw_encoding(), lsb + 4, lsb); 288 } 289 290 //<0-15>reg: As `rf(FloatRegister)`, but only the lower 16 FloatRegisters are allowed. 291 void lrf(FloatRegister r, int lsb) { 292 f(r->raw_encoding(), lsb + 3, lsb); 293 } 294 295 void prf(PRegister r, int lsb) { 296 f(r->raw_encoding(), lsb + 3, lsb); 297 } 298 299 void pgrf(PRegister r, int lsb) { 300 f(r->raw_encoding(), lsb + 2, lsb); 301 } 302 303 unsigned get(int msb = 31, int lsb = 0) { 304 int nbits = msb - lsb + 1; 305 unsigned mask = checked_cast<unsigned>(right_n_bits(nbits)) << lsb; 306 assert_cond((bits & mask) == mask); 307 return (insn & mask) >> lsb; 308 } 309 }; 310 311 #define starti Instruction_aarch64 current_insn(this); 312 313 class PrePost { 314 int _offset; 315 Register _r; 316 protected: 317 PrePost(Register reg, int o) : _offset(o), _r(reg) { } 318 ~PrePost() = default; 319 PrePost(const PrePost&) = default; 320 PrePost& operator=(const PrePost&) = default; 321 public: 322 int offset() const { return _offset; } 323 Register reg() const { return _r; } 324 }; 325 326 class Pre : public PrePost { 327 public: 328 Pre(Register reg, int o) : PrePost(reg, o) { } 329 }; 330 331 class Post : public PrePost { 332 Register _idx; 333 bool _is_postreg; 334 public: 335 Post(Register reg, int o) : PrePost(reg, o), _idx(noreg), _is_postreg(false) {} 336 Post(Register reg, Register idx) : PrePost(reg, 0), _idx(idx), _is_postreg(true) {} 337 Register idx_reg() const { return _idx; } 338 bool is_postreg() const { return _is_postreg; } 339 }; 340 341 namespace ext 342 { 343 enum operation { uxtb, uxth, uxtw, uxtx, sxtb, sxth, sxtw, sxtx }; 344 }; 345 346 // Addressing modes 347 class Address { 348 public: 349 350 enum mode { no_mode, base_plus_offset, pre, post, post_reg, 351 base_plus_offset_reg, literal }; 352 353 // Shift and extend for base reg + reg offset addressing 354 class extend { 355 int _option, _shift; 356 ext::operation _op; 357 public: 358 extend() { } 359 extend(int s, int o, ext::operation op) : _option(o), _shift(s), _op(op) { } 360 int option() const{ return _option; } 361 int shift() const { return _shift; } 362 ext::operation op() const { return _op; } 363 }; 364 365 static extend uxtw(int shift = -1) { return extend(shift, 0b010, ext::uxtw); } 366 static extend lsl(int shift = -1) { return extend(shift, 0b011, ext::uxtx); } 367 static extend sxtw(int shift = -1) { return extend(shift, 0b110, ext::sxtw); } 368 static extend sxtx(int shift = -1) { return extend(shift, 0b111, ext::sxtx); } 369 370 private: 371 struct Nonliteral { 372 Nonliteral(Register base, Register index, int64_t offset, extend ext = extend()) 373 : _base(base), _index(index), _offset(offset), _ext(ext) {} 374 Register _base; 375 Register _index; 376 int64_t _offset; 377 extend _ext; 378 }; 379 380 struct Literal { 381 Literal(address target, const RelocationHolder& rspec) 382 : _target(target), _rspec(rspec) {} 383 384 // If the target is far we'll need to load the ea of this to a 385 // register to reach it. Otherwise if near we can do PC-relative 386 // addressing. 387 address _target; 388 389 RelocationHolder _rspec; 390 }; 391 392 void assert_is_nonliteral() const NOT_DEBUG_RETURN; 393 void assert_is_literal() const NOT_DEBUG_RETURN; 394 395 // Discriminated union, based on _mode. 396 // - no_mode: uses dummy _nonliteral, for ease of copying. 397 // - literal: only _literal is used. 398 // - others: only _nonliteral is used. 399 enum mode _mode; 400 union { 401 Nonliteral _nonliteral; 402 Literal _literal; 403 }; 404 405 // Helper for copy constructor and assignment operator. 406 // Copy mode-relevant part of a into this. 407 void copy_data(const Address& a) { 408 assert(_mode == a._mode, "precondition"); 409 if (_mode == literal) { 410 new (&_literal) Literal(a._literal); 411 } else { 412 // non-literal mode or no_mode. 413 new (&_nonliteral) Nonliteral(a._nonliteral); 414 } 415 } 416 417 public: 418 // no_mode initializes _nonliteral for ease of copying. 419 Address() : 420 _mode(no_mode), 421 _nonliteral(noreg, noreg, 0) 422 {} 423 424 Address(Register r) : 425 _mode(base_plus_offset), 426 _nonliteral(r, noreg, 0) 427 {} 428 429 template<typename T, ENABLE_IF(std::is_integral<T>::value)> 430 Address(Register r, T o) : 431 _mode(base_plus_offset), 432 _nonliteral(r, noreg, o) 433 {} 434 435 Address(Register r, ByteSize disp) : Address(r, in_bytes(disp)) {} 436 437 Address(Register r, Register r1, extend ext = lsl()) : 438 _mode(base_plus_offset_reg), 439 _nonliteral(r, r1, 0, ext) 440 {} 441 442 Address(Pre p) : 443 _mode(pre), 444 _nonliteral(p.reg(), noreg, p.offset()) 445 {} 446 447 Address(Post p) : 448 _mode(p.is_postreg() ? post_reg : post), 449 _nonliteral(p.reg(), p.idx_reg(), p.offset()) 450 {} 451 452 Address(address target, const RelocationHolder& rspec) : 453 _mode(literal), 454 _literal(target, rspec) 455 {} 456 457 Address(address target, relocInfo::relocType rtype = relocInfo::external_word_type); 458 459 Address(Register base, RegisterOrConstant index, extend ext = lsl()) { 460 if (index.is_register()) { 461 _mode = base_plus_offset_reg; 462 new (&_nonliteral) Nonliteral(base, index.as_register(), 0, ext); 463 } else { 464 guarantee(ext.option() == ext::uxtx, "should be"); 465 assert(index.is_constant(), "should be"); 466 _mode = base_plus_offset; 467 new (&_nonliteral) Nonliteral(base, 468 noreg, 469 index.as_constant() << ext.shift()); 470 } 471 } 472 473 Address(const Address& a) : _mode(a._mode) { copy_data(a); } 474 475 // Verify the value is trivially destructible regardless of mode, so our 476 // destructor can also be trivial, and so our assignment operator doesn't 477 // need to destruct the old value before copying over it. 478 static_assert(std::is_trivially_destructible<Literal>::value, "must be"); 479 static_assert(std::is_trivially_destructible<Nonliteral>::value, "must be"); 480 481 Address& operator=(const Address& a) { 482 _mode = a._mode; 483 copy_data(a); 484 return *this; 485 } 486 487 ~Address() = default; 488 489 Register base() const { 490 assert_is_nonliteral(); 491 return _nonliteral._base; 492 } 493 494 int64_t offset() const { 495 assert_is_nonliteral(); 496 return _nonliteral._offset; 497 } 498 499 Register index() const { 500 assert_is_nonliteral(); 501 return _nonliteral._index; 502 } 503 504 extend ext() const { 505 assert_is_nonliteral(); 506 return _nonliteral._ext; 507 } 508 509 mode getMode() const { 510 return _mode; 511 } 512 513 bool uses(Register reg) const { 514 switch (_mode) { 515 case literal: 516 case no_mode: 517 return false; 518 case base_plus_offset: 519 case base_plus_offset_reg: 520 case pre: 521 case post: 522 case post_reg: 523 return base() == reg || index() == reg; 524 default: 525 ShouldNotReachHere(); 526 return false; 527 } 528 } 529 530 address target() const { 531 assert_is_literal(); 532 return _literal._target; 533 } 534 535 const RelocationHolder& rspec() const { 536 assert_is_literal(); 537 return _literal._rspec; 538 } 539 540 void encode(Instruction_aarch64 *i) const { 541 i->f(0b111, 29, 27); 542 i->srf(base(), 5); 543 544 switch(_mode) { 545 case base_plus_offset: 546 { 547 unsigned size = i->get(31, 30); 548 if (i->get(26, 26) && i->get(23, 23)) { 549 // SIMD Q Type - Size = 128 bits 550 assert(size == 0, "bad size"); 551 size = 0b100; 552 } 553 assert(offset_ok_for_immed(offset(), size), 554 "must be, was: " INT64_FORMAT ", %d", offset(), size); 555 unsigned mask = (1 << size) - 1; 556 if (offset() < 0 || offset() & mask) { 557 i->f(0b00, 25, 24); 558 i->f(0, 21), i->f(0b00, 11, 10); 559 i->sf(offset(), 20, 12); 560 } else { 561 i->f(0b01, 25, 24); 562 i->f(checked_cast<unsigned>(offset() >> size), 21, 10); 563 } 564 } 565 break; 566 567 case base_plus_offset_reg: 568 { 569 i->f(0b00, 25, 24); 570 i->f(1, 21); 571 i->rf(index(), 16); 572 i->f(ext().option(), 15, 13); 573 unsigned size = i->get(31, 30); 574 if (i->get(26, 26) && i->get(23, 23)) { 575 // SIMD Q Type - Size = 128 bits 576 assert(size == 0, "bad size"); 577 size = 0b100; 578 } 579 if (size == 0) // It's a byte 580 i->f(ext().shift() >= 0, 12); 581 else { 582 guarantee(ext().shift() <= 0 || ext().shift() == (int)size, "bad shift"); 583 i->f(ext().shift() > 0, 12); 584 } 585 i->f(0b10, 11, 10); 586 } 587 break; 588 589 case pre: 590 i->f(0b00, 25, 24); 591 i->f(0, 21), i->f(0b11, 11, 10); 592 i->sf(offset(), 20, 12); 593 break; 594 595 case post: 596 i->f(0b00, 25, 24); 597 i->f(0, 21), i->f(0b01, 11, 10); 598 i->sf(offset(), 20, 12); 599 break; 600 601 default: 602 ShouldNotReachHere(); 603 } 604 } 605 606 void encode_pair(Instruction_aarch64 *i) const { 607 switch(_mode) { 608 case base_plus_offset: 609 i->f(0b010, 25, 23); 610 break; 611 case pre: 612 i->f(0b011, 25, 23); 613 break; 614 case post: 615 i->f(0b001, 25, 23); 616 break; 617 default: 618 ShouldNotReachHere(); 619 } 620 621 unsigned size; // Operand shift in 32-bit words 622 623 if (i->get(26, 26)) { // float 624 switch(i->get(31, 30)) { 625 case 0b10: 626 size = 2; break; 627 case 0b01: 628 size = 1; break; 629 case 0b00: 630 size = 0; break; 631 default: 632 ShouldNotReachHere(); 633 size = 0; // unreachable 634 } 635 } else { 636 size = i->get(31, 31); 637 } 638 639 size = 4 << size; 640 guarantee(offset() % size == 0, "bad offset"); 641 i->sf(offset() / size, 21, 15); 642 i->srf(base(), 5); 643 } 644 645 void encode_nontemporal_pair(Instruction_aarch64 *i) const { 646 guarantee(_mode == base_plus_offset, "Bad addressing mode for nontemporal op"); 647 i->f(0b000, 25, 23); 648 unsigned size = i->get(31, 31); 649 size = 4 << size; 650 guarantee(offset() % size == 0, "bad offset"); 651 i->sf(offset() / size, 21, 15); 652 i->srf(base(), 5); 653 } 654 655 void lea(MacroAssembler *, Register) const; 656 657 static bool offset_ok_for_immed(int64_t offset, uint shift); 658 659 static bool offset_ok_for_sve_immed(int64_t offset, int shift, int vl /* sve vector length */) { 660 if (offset % vl == 0) { 661 // Convert address offset into sve imm offset (MUL VL). 662 int64_t sve_offset = offset / vl; 663 int32_t range = 1 << (shift - 1); 664 if ((-range <= sve_offset) && (sve_offset < range)) { 665 // sve_offset can be encoded 666 return true; 667 } 668 } 669 return false; 670 } 671 }; 672 673 // Convenience classes 674 class RuntimeAddress: public Address { 675 676 public: 677 678 RuntimeAddress(address target) : Address(target, relocInfo::runtime_call_type) {} 679 680 }; 681 682 class OopAddress: public Address { 683 684 public: 685 686 OopAddress(address target) : Address(target, relocInfo::oop_type){} 687 688 }; 689 690 class ExternalAddress: public Address { 691 private: 692 static relocInfo::relocType reloc_for_target(address target) { 693 // Sometimes ExternalAddress is used for values which aren't 694 // exactly addresses, like the card table base. 695 // external_word_type can't be used for values in the first page 696 // so just skip the reloc in that case. 697 return external_word_Relocation::can_be_relocated(target) ? relocInfo::external_word_type : relocInfo::none; 698 } 699 700 public: 701 702 ExternalAddress(address target) : Address(target, reloc_for_target(target)) {} 703 704 }; 705 706 class InternalAddress: public Address { 707 708 public: 709 710 InternalAddress(address target) : Address(target, relocInfo::internal_word_type) {} 711 }; 712 713 const int FPUStateSizeInWords = FloatRegister::number_of_registers * FloatRegister::save_slots_per_register; 714 715 typedef enum { 716 PLDL1KEEP = 0b00000, PLDL1STRM, PLDL2KEEP, PLDL2STRM, PLDL3KEEP, PLDL3STRM, 717 PSTL1KEEP = 0b10000, PSTL1STRM, PSTL2KEEP, PSTL2STRM, PSTL3KEEP, PSTL3STRM, 718 PLIL1KEEP = 0b01000, PLIL1STRM, PLIL2KEEP, PLIL2STRM, PLIL3KEEP, PLIL3STRM 719 } prfop; 720 721 class Assembler : public AbstractAssembler { 722 723 public: 724 725 #ifndef PRODUCT 726 static const uintptr_t asm_bp; 727 728 void emit_int32(jint x) { 729 if ((uintptr_t)pc() == asm_bp) 730 NOP(); 731 AbstractAssembler::emit_int32(x); 732 } 733 #else 734 void emit_int32(jint x) { 735 AbstractAssembler::emit_int32(x); 736 } 737 #endif 738 739 enum { instruction_size = 4 }; 740 741 //---< calculate length of instruction >--- 742 // We just use the values set above. 743 // instruction must start at passed address 744 static unsigned int instr_len(unsigned char *instr) { return instruction_size; } 745 746 //---< longest instructions >--- 747 static unsigned int instr_maxlen() { return instruction_size; } 748 749 Address adjust(Register base, int offset, bool preIncrement) { 750 if (preIncrement) 751 return Address(Pre(base, offset)); 752 else 753 return Address(Post(base, offset)); 754 } 755 756 Address pre(Register base, int offset) { 757 return adjust(base, offset, true); 758 } 759 760 Address post(Register base, int offset) { 761 return adjust(base, offset, false); 762 } 763 764 Address post(Register base, Register idx) { 765 return Address(Post(base, idx)); 766 } 767 768 static address locate_next_instruction(address inst); 769 770 #define f current_insn.f 771 #define sf current_insn.sf 772 #define rf current_insn.rf 773 #define lrf current_insn.lrf 774 #define srf current_insn.srf 775 #define zrf current_insn.zrf 776 #define prf current_insn.prf 777 #define pgrf current_insn.pgrf 778 779 typedef void (Assembler::* uncond_branch_insn)(address dest); 780 typedef void (Assembler::* compare_and_branch_insn)(Register Rt, address dest); 781 typedef void (Assembler::* test_and_branch_insn)(Register Rt, int bitpos, address dest); 782 typedef void (Assembler::* prefetch_insn)(address target, prfop); 783 784 void wrap_label(Label &L, uncond_branch_insn insn); 785 void wrap_label(Register r, Label &L, compare_and_branch_insn insn); 786 void wrap_label(Register r, int bitpos, Label &L, test_and_branch_insn insn); 787 void wrap_label(Label &L, prfop, prefetch_insn insn); 788 789 // PC-rel. addressing 790 791 void adr(Register Rd, address dest); 792 void _adrp(Register Rd, address dest); 793 794 void adr(Register Rd, const Address &dest); 795 void _adrp(Register Rd, const Address &dest); 796 797 void adr(Register Rd, Label &L) { 798 wrap_label(Rd, L, &Assembler::Assembler::adr); 799 } 800 void _adrp(Register Rd, Label &L) { 801 wrap_label(Rd, L, &Assembler::_adrp); 802 } 803 804 void adrp(Register Rd, const Address &dest, uint64_t &offset) = delete; 805 806 void prfm(const Address &adr, prfop pfop = PLDL1KEEP); 807 808 #undef INSN 809 810 void add_sub_immediate(Instruction_aarch64 ¤t_insn, Register Rd, Register Rn, 811 unsigned uimm, int op, int negated_op); 812 813 // Add/subtract (immediate) 814 #define INSN(NAME, decode, negated) \ 815 void NAME(Register Rd, Register Rn, unsigned imm, unsigned shift) { \ 816 starti; \ 817 f(decode, 31, 29), f(0b10001, 28, 24), f(shift, 23, 22), f(imm, 21, 10); \ 818 zrf(Rd, 0), srf(Rn, 5); \ 819 } \ 820 \ 821 void NAME(Register Rd, Register Rn, unsigned imm) { \ 822 starti; \ 823 add_sub_immediate(current_insn, Rd, Rn, imm, decode, negated); \ 824 } 825 826 INSN(addsw, 0b001, 0b011); 827 INSN(subsw, 0b011, 0b001); 828 INSN(adds, 0b101, 0b111); 829 INSN(subs, 0b111, 0b101); 830 831 #undef INSN 832 833 #define INSN(NAME, decode, negated) \ 834 void NAME(Register Rd, Register Rn, unsigned imm) { \ 835 starti; \ 836 add_sub_immediate(current_insn, Rd, Rn, imm, decode, negated); \ 837 } 838 839 INSN(addw, 0b000, 0b010); 840 INSN(subw, 0b010, 0b000); 841 INSN(add, 0b100, 0b110); 842 INSN(sub, 0b110, 0b100); 843 844 #undef INSN 845 846 // Logical (immediate) 847 #define INSN(NAME, decode, is32) \ 848 void NAME(Register Rd, Register Rn, uint64_t imm) { \ 849 starti; \ 850 uint32_t val = encode_logical_immediate(is32, imm); \ 851 f(decode, 31, 29), f(0b100100, 28, 23), f(val, 22, 10); \ 852 srf(Rd, 0), zrf(Rn, 5); \ 853 } 854 855 INSN(andw, 0b000, true); 856 INSN(orrw, 0b001, true); 857 INSN(eorw, 0b010, true); 858 INSN(andr, 0b100, false); 859 INSN(orr, 0b101, false); 860 INSN(eor, 0b110, false); 861 862 #undef INSN 863 864 #define INSN(NAME, decode, is32) \ 865 void NAME(Register Rd, Register Rn, uint64_t imm) { \ 866 starti; \ 867 uint32_t val = encode_logical_immediate(is32, imm); \ 868 f(decode, 31, 29), f(0b100100, 28, 23), f(val, 22, 10); \ 869 zrf(Rd, 0), zrf(Rn, 5); \ 870 } 871 872 INSN(ands, 0b111, false); 873 INSN(andsw, 0b011, true); 874 875 #undef INSN 876 877 // Move wide (immediate) 878 #define INSN(NAME, opcode) \ 879 void NAME(Register Rd, unsigned imm, unsigned shift = 0) { \ 880 assert_cond((shift/16)*16 == shift); \ 881 starti; \ 882 f(opcode, 31, 29), f(0b100101, 28, 23), f(shift/16, 22, 21), \ 883 f(imm, 20, 5); \ 884 zrf(Rd, 0); \ 885 } 886 887 INSN(movnw, 0b000); 888 INSN(movzw, 0b010); 889 INSN(movkw, 0b011); 890 INSN(movn, 0b100); 891 INSN(movz, 0b110); 892 INSN(movk, 0b111); 893 894 #undef INSN 895 896 // Bitfield 897 #define INSN(NAME, opcode, size) \ 898 void NAME(Register Rd, Register Rn, unsigned immr, unsigned imms) { \ 899 starti; \ 900 guarantee(size == 1 || (immr < 32 && imms < 32), "incorrect immr/imms");\ 901 f(opcode, 31, 22), f(immr, 21, 16), f(imms, 15, 10); \ 902 zrf(Rn, 5), rf(Rd, 0); \ 903 } 904 905 INSN(sbfmw, 0b0001001100, 0); 906 INSN(bfmw, 0b0011001100, 0); 907 INSN(ubfmw, 0b0101001100, 0); 908 INSN(sbfm, 0b1001001101, 1); 909 INSN(bfm, 0b1011001101, 1); 910 INSN(ubfm, 0b1101001101, 1); 911 912 #undef INSN 913 914 // Extract 915 #define INSN(NAME, opcode, size) \ 916 void NAME(Register Rd, Register Rn, Register Rm, unsigned imms) { \ 917 starti; \ 918 guarantee(size == 1 || imms < 32, "incorrect imms"); \ 919 f(opcode, 31, 21), f(imms, 15, 10); \ 920 zrf(Rm, 16), zrf(Rn, 5), zrf(Rd, 0); \ 921 } 922 923 INSN(extrw, 0b00010011100, 0); 924 INSN(extr, 0b10010011110, 1); 925 926 #undef INSN 927 928 // The maximum range of a branch is fixed for the AArch64 929 // architecture. In debug mode we shrink it in order to test 930 // trampolines, but not so small that branches in the interpreter 931 // are out of range. 932 static const uint64_t branch_range = NOT_DEBUG(128 * M) DEBUG_ONLY(2 * M); 933 934 static bool reachable_from_branch_at(address branch, address target) { 935 return uabs(target - branch) < branch_range; 936 } 937 938 // Unconditional branch (immediate) 939 #define INSN(NAME, opcode) \ 940 void NAME(address dest) { \ 941 starti; \ 942 int64_t offset = (dest - pc()) >> 2; \ 943 DEBUG_ONLY(assert(reachable_from_branch_at(pc(), dest), "debug only")); \ 944 f(opcode, 31), f(0b00101, 30, 26), sf(offset, 25, 0); \ 945 } \ 946 void NAME(Label &L) { \ 947 wrap_label(L, &Assembler::NAME); \ 948 } \ 949 void NAME(const Address &dest); 950 951 INSN(b, 0); 952 INSN(bl, 1); 953 954 #undef INSN 955 956 // Compare & branch (immediate) 957 #define INSN(NAME, opcode) \ 958 void NAME(Register Rt, address dest) { \ 959 int64_t offset = (dest - pc()) >> 2; \ 960 starti; \ 961 f(opcode, 31, 24), sf(offset, 23, 5), rf(Rt, 0); \ 962 } \ 963 void NAME(Register Rt, Label &L) { \ 964 wrap_label(Rt, L, &Assembler::NAME); \ 965 } 966 967 INSN(cbzw, 0b00110100); 968 INSN(cbnzw, 0b00110101); 969 INSN(cbz, 0b10110100); 970 INSN(cbnz, 0b10110101); 971 972 #undef INSN 973 974 // Test & branch (immediate) 975 #define INSN(NAME, opcode) \ 976 void NAME(Register Rt, int bitpos, address dest) { \ 977 int64_t offset = (dest - pc()) >> 2; \ 978 int b5 = bitpos >> 5; \ 979 bitpos &= 0x1f; \ 980 starti; \ 981 f(b5, 31), f(opcode, 30, 24), f(bitpos, 23, 19), sf(offset, 18, 5); \ 982 rf(Rt, 0); \ 983 } \ 984 void NAME(Register Rt, int bitpos, Label &L) { \ 985 wrap_label(Rt, bitpos, L, &Assembler::NAME); \ 986 } 987 988 INSN(tbz, 0b0110110); 989 INSN(tbnz, 0b0110111); 990 991 #undef INSN 992 993 // Conditional branch (immediate) 994 enum Condition 995 {EQ, NE, HS, CS=HS, LO, CC=LO, MI, PL, VS, VC, HI, LS, GE, LT, GT, LE, AL, NV}; 996 997 void br(Condition cond, address dest) { 998 int64_t offset = (dest - pc()) >> 2; 999 starti; 1000 f(0b0101010, 31, 25), f(0, 24), sf(offset, 23, 5), f(0, 4), f(cond, 3, 0); 1001 } 1002 1003 #define INSN(NAME, cond) \ 1004 void NAME(address dest) { \ 1005 br(cond, dest); \ 1006 } 1007 1008 INSN(beq, EQ); 1009 INSN(bne, NE); 1010 INSN(bhs, HS); 1011 INSN(bcs, CS); 1012 INSN(blo, LO); 1013 INSN(bcc, CC); 1014 INSN(bmi, MI); 1015 INSN(bpl, PL); 1016 INSN(bvs, VS); 1017 INSN(bvc, VC); 1018 INSN(bhi, HI); 1019 INSN(bls, LS); 1020 INSN(bge, GE); 1021 INSN(blt, LT); 1022 INSN(bgt, GT); 1023 INSN(ble, LE); 1024 INSN(bal, AL); 1025 INSN(bnv, NV); 1026 1027 void br(Condition cc, Label &L); 1028 1029 #undef INSN 1030 1031 // Exception generation 1032 void generate_exception(int opc, int op2, int LL, unsigned imm) { 1033 starti; 1034 f(0b11010100, 31, 24); 1035 f(opc, 23, 21), f(imm, 20, 5), f(op2, 4, 2), f(LL, 1, 0); 1036 } 1037 1038 #define INSN(NAME, opc, op2, LL) \ 1039 void NAME(unsigned imm) { \ 1040 generate_exception(opc, op2, LL, imm); \ 1041 } 1042 1043 INSN(svc, 0b000, 0, 0b01); 1044 INSN(hvc, 0b000, 0, 0b10); 1045 INSN(smc, 0b000, 0, 0b11); 1046 INSN(brk, 0b001, 0, 0b00); 1047 INSN(hlt, 0b010, 0, 0b00); 1048 INSN(dcps1, 0b101, 0, 0b01); 1049 INSN(dcps2, 0b101, 0, 0b10); 1050 INSN(dcps3, 0b101, 0, 0b11); 1051 1052 #undef INSN 1053 1054 // System 1055 void system(int op0, int op1, int CRn, int CRm, int op2, 1056 Register rt = dummy_reg) 1057 { 1058 starti; 1059 f(0b11010101000, 31, 21); 1060 f(op0, 20, 19); 1061 f(op1, 18, 16); 1062 f(CRn, 15, 12); 1063 f(CRm, 11, 8); 1064 f(op2, 7, 5); 1065 rf(rt, 0); 1066 } 1067 1068 // Hint instructions 1069 1070 #define INSN(NAME, crm, op2) \ 1071 void NAME() { \ 1072 system(0b00, 0b011, 0b0010, crm, op2); \ 1073 } 1074 1075 INSN(nop, 0b000, 0b0000); 1076 INSN(yield, 0b000, 0b0001); 1077 INSN(wfe, 0b000, 0b0010); 1078 INSN(wfi, 0b000, 0b0011); 1079 INSN(sev, 0b000, 0b0100); 1080 INSN(sevl, 0b000, 0b0101); 1081 1082 INSN(autia1716, 0b0001, 0b100); 1083 INSN(autiasp, 0b0011, 0b101); 1084 INSN(autiaz, 0b0011, 0b100); 1085 INSN(autib1716, 0b0001, 0b110); 1086 INSN(autibsp, 0b0011, 0b111); 1087 INSN(autibz, 0b0011, 0b110); 1088 INSN(pacia1716, 0b0001, 0b000); 1089 INSN(paciasp, 0b0011, 0b001); 1090 INSN(paciaz, 0b0011, 0b000); 1091 INSN(pacib1716, 0b0001, 0b010); 1092 INSN(pacibsp, 0b0011, 0b011); 1093 INSN(pacibz, 0b0011, 0b010); 1094 INSN(xpaclri, 0b0000, 0b111); 1095 1096 #undef INSN 1097 1098 // we only provide mrs and msr for the special purpose system 1099 // registers where op1 (instr[20:19]) == 11 1100 // n.b msr has L (instr[21]) == 0 mrs has L == 1 1101 1102 void msr(int op1, int CRn, int CRm, int op2, Register rt) { 1103 starti; 1104 f(0b1101010100011, 31, 19); 1105 f(op1, 18, 16); 1106 f(CRn, 15, 12); 1107 f(CRm, 11, 8); 1108 f(op2, 7, 5); 1109 // writing zr is ok 1110 zrf(rt, 0); 1111 } 1112 1113 void mrs(int op1, int CRn, int CRm, int op2, Register rt) { 1114 starti; 1115 f(0b1101010100111, 31, 19); 1116 f(op1, 18, 16); 1117 f(CRn, 15, 12); 1118 f(CRm, 11, 8); 1119 f(op2, 7, 5); 1120 // reading to zr is a mistake 1121 rf(rt, 0); 1122 } 1123 1124 enum barrier {OSHLD = 0b0001, OSHST, OSH, NSHLD=0b0101, NSHST, NSH, 1125 ISHLD = 0b1001, ISHST, ISH, LD=0b1101, ST, SY}; 1126 1127 void dsb(barrier imm) { 1128 system(0b00, 0b011, 0b00011, imm, 0b100); 1129 } 1130 1131 void dmb(barrier imm) { 1132 system(0b00, 0b011, 0b00011, imm, 0b101); 1133 } 1134 1135 void isb() { 1136 system(0b00, 0b011, 0b00011, SY, 0b110); 1137 } 1138 1139 void sys(int op1, int CRn, int CRm, int op2, 1140 Register rt = as_Register(0b11111)) { 1141 system(0b01, op1, CRn, CRm, op2, rt); 1142 } 1143 1144 // Only implement operations accessible from EL0 or higher, i.e., 1145 // op1 CRn CRm op2 1146 // IC IVAU 3 7 5 1 1147 // DC CVAC 3 7 10 1 1148 // DC CVAP 3 7 12 1 1149 // DC CVAU 3 7 11 1 1150 // DC CIVAC 3 7 14 1 1151 // DC ZVA 3 7 4 1 1152 // So only deal with the CRm field. 1153 enum icache_maintenance {IVAU = 0b0101}; 1154 enum dcache_maintenance {CVAC = 0b1010, CVAP = 0b1100, CVAU = 0b1011, CIVAC = 0b1110, ZVA = 0b100}; 1155 1156 void dc(dcache_maintenance cm, Register Rt) { 1157 sys(0b011, 0b0111, cm, 0b001, Rt); 1158 } 1159 1160 void ic(icache_maintenance cm, Register Rt) { 1161 sys(0b011, 0b0111, cm, 0b001, Rt); 1162 } 1163 1164 // A more convenient access to dmb for our purposes 1165 enum Membar_mask_bits { 1166 // We can use ISH for a barrier because the Arm ARM says "This 1167 // architecture assumes that all Processing Elements that use the 1168 // same operating system or hypervisor are in the same Inner 1169 // Shareable shareability domain." 1170 StoreStore = ISHST, 1171 LoadStore = ISHLD, 1172 LoadLoad = ISHLD, 1173 StoreLoad = ISH, 1174 AnyAny = ISH 1175 }; 1176 1177 void membar(Membar_mask_bits order_constraint) { 1178 dmb(Assembler::barrier(order_constraint)); 1179 } 1180 1181 // Unconditional branch (register) 1182 1183 void branch_reg(int OP, int A, int M, Register RN, Register RM) { 1184 starti; 1185 f(0b1101011, 31, 25); 1186 f(OP, 24, 21); 1187 f(0b111110000, 20, 12); 1188 f(A, 11, 11); 1189 f(M, 10, 10); 1190 rf(RN, 5); 1191 rf(RM, 0); 1192 } 1193 1194 #define INSN(NAME, opc) \ 1195 void NAME(Register RN) { \ 1196 branch_reg(opc, 0, 0, RN, r0); \ 1197 } 1198 1199 INSN(br, 0b0000); 1200 INSN(blr, 0b0001); 1201 INSN(ret, 0b0010); 1202 1203 void ret(void *p); // This forces a compile-time error for ret(0) 1204 1205 #undef INSN 1206 1207 #define INSN(NAME, opc) \ 1208 void NAME() { \ 1209 branch_reg(opc, 0, 0, dummy_reg, r0); \ 1210 } 1211 1212 INSN(eret, 0b0100); 1213 INSN(drps, 0b0101); 1214 1215 #undef INSN 1216 1217 #define INSN(NAME, M) \ 1218 void NAME() { \ 1219 branch_reg(0b0010, 1, M, dummy_reg, dummy_reg); \ 1220 } 1221 1222 INSN(retaa, 0); 1223 INSN(retab, 1); 1224 1225 #undef INSN 1226 1227 #define INSN(NAME, OP, M) \ 1228 void NAME(Register rn) { \ 1229 branch_reg(OP, 1, M, rn, dummy_reg); \ 1230 } 1231 1232 INSN(braaz, 0b0000, 0); 1233 INSN(brabz, 0b0000, 1); 1234 INSN(blraaz, 0b0001, 0); 1235 INSN(blrabz, 0b0001, 1); 1236 1237 #undef INSN 1238 1239 #define INSN(NAME, OP, M) \ 1240 void NAME(Register rn, Register rm) { \ 1241 branch_reg(OP, 1, M, rn, rm); \ 1242 } 1243 1244 INSN(braa, 0b1000, 0); 1245 INSN(brab, 0b1000, 1); 1246 INSN(blraa, 0b1001, 0); 1247 INSN(blrab, 0b1001, 1); 1248 1249 #undef INSN 1250 1251 // Load/store exclusive 1252 enum operand_size { byte, halfword, word, xword }; 1253 1254 void load_store_exclusive(Register Rs, Register Rt1, Register Rt2, 1255 Register Rn, enum operand_size sz, int op, bool ordered) { 1256 starti; 1257 f(sz, 31, 30), f(0b001000, 29, 24), f(op, 23, 21); 1258 rf(Rs, 16), f(ordered, 15), zrf(Rt2, 10), srf(Rn, 5), zrf(Rt1, 0); 1259 } 1260 1261 void load_exclusive(Register dst, Register addr, 1262 enum operand_size sz, bool ordered) { 1263 load_store_exclusive(dummy_reg, dst, dummy_reg, addr, 1264 sz, 0b010, ordered); 1265 } 1266 1267 void store_exclusive(Register status, Register new_val, Register addr, 1268 enum operand_size sz, bool ordered) { 1269 load_store_exclusive(status, new_val, dummy_reg, addr, 1270 sz, 0b000, ordered); 1271 } 1272 1273 #define INSN4(NAME, sz, op, o0) /* Four registers */ \ 1274 void NAME(Register Rs, Register Rt1, Register Rt2, Register Rn) { \ 1275 guarantee(Rs != Rn && Rs != Rt1 && Rs != Rt2, "unpredictable instruction"); \ 1276 load_store_exclusive(Rs, Rt1, Rt2, Rn, sz, op, o0); \ 1277 } 1278 1279 #define INSN3(NAME, sz, op, o0) /* Three registers */ \ 1280 void NAME(Register Rs, Register Rt, Register Rn) { \ 1281 guarantee(Rs != Rn && Rs != Rt, "unpredictable instruction"); \ 1282 load_store_exclusive(Rs, Rt, dummy_reg, Rn, sz, op, o0); \ 1283 } 1284 1285 #define INSN2(NAME, sz, op, o0) /* Two registers */ \ 1286 void NAME(Register Rt, Register Rn) { \ 1287 load_store_exclusive(dummy_reg, Rt, dummy_reg, \ 1288 Rn, sz, op, o0); \ 1289 } 1290 1291 #define INSN_FOO(NAME, sz, op, o0) /* Three registers, encoded differently */ \ 1292 void NAME(Register Rt1, Register Rt2, Register Rn) { \ 1293 guarantee(Rt1 != Rt2, "unpredictable instruction"); \ 1294 load_store_exclusive(dummy_reg, Rt1, Rt2, Rn, sz, op, o0); \ 1295 } 1296 1297 // bytes 1298 INSN3(stxrb, byte, 0b000, 0); 1299 INSN3(stlxrb, byte, 0b000, 1); 1300 INSN2(ldxrb, byte, 0b010, 0); 1301 INSN2(ldaxrb, byte, 0b010, 1); 1302 INSN2(stlrb, byte, 0b100, 1); 1303 INSN2(ldarb, byte, 0b110, 1); 1304 1305 // halfwords 1306 INSN3(stxrh, halfword, 0b000, 0); 1307 INSN3(stlxrh, halfword, 0b000, 1); 1308 INSN2(ldxrh, halfword, 0b010, 0); 1309 INSN2(ldaxrh, halfword, 0b010, 1); 1310 INSN2(stlrh, halfword, 0b100, 1); 1311 INSN2(ldarh, halfword, 0b110, 1); 1312 1313 // words 1314 INSN3(stxrw, word, 0b000, 0); 1315 INSN3(stlxrw, word, 0b000, 1); 1316 INSN4(stxpw, word, 0b001, 0); 1317 INSN4(stlxpw, word, 0b001, 1); 1318 INSN2(ldxrw, word, 0b010, 0); 1319 INSN2(ldaxrw, word, 0b010, 1); 1320 INSN2(stlrw, word, 0b100, 1); 1321 INSN2(ldarw, word, 0b110, 1); 1322 // pairs of words 1323 INSN_FOO(ldxpw, word, 0b011, 0); 1324 INSN_FOO(ldaxpw, word, 0b011, 1); 1325 1326 // xwords 1327 INSN3(stxr, xword, 0b000, 0); 1328 INSN3(stlxr, xword, 0b000, 1); 1329 INSN4(stxp, xword, 0b001, 0); 1330 INSN4(stlxp, xword, 0b001, 1); 1331 INSN2(ldxr, xword, 0b010, 0); 1332 INSN2(ldaxr, xword, 0b010, 1); 1333 INSN2(stlr, xword, 0b100, 1); 1334 INSN2(ldar, xword, 0b110, 1); 1335 // pairs of xwords 1336 INSN_FOO(ldxp, xword, 0b011, 0); 1337 INSN_FOO(ldaxp, xword, 0b011, 1); 1338 1339 #undef INSN2 1340 #undef INSN3 1341 #undef INSN4 1342 #undef INSN_FOO 1343 1344 // 8.1 Compare and swap extensions 1345 void lse_cas(Register Rs, Register Rt, Register Rn, 1346 enum operand_size sz, bool a, bool r, bool not_pair) { 1347 starti; 1348 if (! not_pair) { // Pair 1349 assert(sz == word || sz == xword, "invalid size"); 1350 /* The size bit is in bit 30, not 31 */ 1351 sz = (operand_size)(sz == word ? 0b00:0b01); 1352 } 1353 f(sz, 31, 30), f(0b001000, 29, 24), f(not_pair ? 1 : 0, 23), f(a, 22), f(1, 21); 1354 zrf(Rs, 16), f(r, 15), f(0b11111, 14, 10), srf(Rn, 5), zrf(Rt, 0); 1355 } 1356 1357 // CAS 1358 #define INSN(NAME, a, r) \ 1359 void NAME(operand_size sz, Register Rs, Register Rt, Register Rn) { \ 1360 assert(Rs != Rn && Rs != Rt, "unpredictable instruction"); \ 1361 lse_cas(Rs, Rt, Rn, sz, a, r, true); \ 1362 } 1363 INSN(cas, false, false) 1364 INSN(casa, true, false) 1365 INSN(casl, false, true) 1366 INSN(casal, true, true) 1367 #undef INSN 1368 1369 // CASP 1370 #define INSN(NAME, a, r) \ 1371 void NAME(operand_size sz, Register Rs, Register Rs1, \ 1372 Register Rt, Register Rt1, Register Rn) { \ 1373 assert((Rs->encoding() & 1) == 0 && (Rt->encoding() & 1) == 0 && \ 1374 Rs->successor() == Rs1 && Rt->successor() == Rt1 && \ 1375 Rs != Rn && Rs1 != Rn && Rs != Rt, "invalid registers"); \ 1376 lse_cas(Rs, Rt, Rn, sz, a, r, false); \ 1377 } 1378 INSN(casp, false, false) 1379 INSN(caspa, true, false) 1380 INSN(caspl, false, true) 1381 INSN(caspal, true, true) 1382 #undef INSN 1383 1384 // 8.1 Atomic operations 1385 void lse_atomic(Register Rs, Register Rt, Register Rn, 1386 enum operand_size sz, int op1, int op2, bool a, bool r) { 1387 starti; 1388 f(sz, 31, 30), f(0b111000, 29, 24), f(a, 23), f(r, 22), f(1, 21); 1389 zrf(Rs, 16), f(op1, 15), f(op2, 14, 12), f(0, 11, 10), srf(Rn, 5), zrf(Rt, 0); 1390 } 1391 1392 #define INSN(NAME, NAME_A, NAME_L, NAME_AL, op1, op2) \ 1393 void NAME(operand_size sz, Register Rs, Register Rt, Register Rn) { \ 1394 lse_atomic(Rs, Rt, Rn, sz, op1, op2, false, false); \ 1395 } \ 1396 void NAME_A(operand_size sz, Register Rs, Register Rt, Register Rn) { \ 1397 lse_atomic(Rs, Rt, Rn, sz, op1, op2, true, false); \ 1398 } \ 1399 void NAME_L(operand_size sz, Register Rs, Register Rt, Register Rn) { \ 1400 lse_atomic(Rs, Rt, Rn, sz, op1, op2, false, true); \ 1401 } \ 1402 void NAME_AL(operand_size sz, Register Rs, Register Rt, Register Rn) {\ 1403 lse_atomic(Rs, Rt, Rn, sz, op1, op2, true, true); \ 1404 } 1405 INSN(ldadd, ldadda, ldaddl, ldaddal, 0, 0b000); 1406 INSN(ldbic, ldbica, ldbicl, ldbical, 0, 0b001); 1407 INSN(ldeor, ldeora, ldeorl, ldeoral, 0, 0b010); 1408 INSN(ldorr, ldorra, ldorrl, ldorral, 0, 0b011); 1409 INSN(ldsmax, ldsmaxa, ldsmaxl, ldsmaxal, 0, 0b100); 1410 INSN(ldsmin, ldsmina, ldsminl, ldsminal, 0, 0b101); 1411 INSN(ldumax, ldumaxa, ldumaxl, ldumaxal, 0, 0b110); 1412 INSN(ldumin, ldumina, lduminl, lduminal, 0, 0b111); 1413 INSN(swp, swpa, swpl, swpal, 1, 0b000); 1414 #undef INSN 1415 1416 // Load register (literal) 1417 #define INSN(NAME, opc, V) \ 1418 void NAME(Register Rt, address dest) { \ 1419 int64_t offset = (dest - pc()) >> 2; \ 1420 starti; \ 1421 f(opc, 31, 30), f(0b011, 29, 27), f(V, 26), f(0b00, 25, 24), \ 1422 sf(offset, 23, 5); \ 1423 rf(Rt, 0); \ 1424 } \ 1425 void NAME(Register Rt, address dest, relocInfo::relocType rtype) { \ 1426 InstructionMark im(this); \ 1427 guarantee(rtype == relocInfo::internal_word_type, \ 1428 "only internal_word_type relocs make sense here"); \ 1429 code_section()->relocate(inst_mark(), InternalAddress(dest).rspec()); \ 1430 NAME(Rt, dest); \ 1431 } \ 1432 void NAME(Register Rt, Label &L) { \ 1433 wrap_label(Rt, L, &Assembler::NAME); \ 1434 } 1435 1436 INSN(ldrw, 0b00, 0); 1437 INSN(ldr, 0b01, 0); 1438 INSN(ldrsw, 0b10, 0); 1439 1440 #undef INSN 1441 1442 #define INSN(NAME, opc, V) \ 1443 void NAME(FloatRegister Rt, address dest) { \ 1444 int64_t offset = (dest - pc()) >> 2; \ 1445 starti; \ 1446 f(opc, 31, 30), f(0b011, 29, 27), f(V, 26), f(0b00, 25, 24), \ 1447 sf(offset, 23, 5); \ 1448 rf(as_Register(Rt), 0); \ 1449 } 1450 1451 INSN(ldrs, 0b00, 1); 1452 INSN(ldrd, 0b01, 1); 1453 INSN(ldrq, 0b10, 1); 1454 1455 #undef INSN 1456 1457 #define INSN(NAME, size, opc) \ 1458 void NAME(FloatRegister Rt, Register Rn) { \ 1459 starti; \ 1460 f(size, 31, 30), f(0b111100, 29, 24), f(opc, 23, 22), f(0, 21); \ 1461 f(0, 20, 12), f(0b01, 11, 10); \ 1462 rf(Rn, 5), rf(as_Register(Rt), 0); \ 1463 } 1464 1465 INSN(ldrs, 0b10, 0b01); 1466 INSN(ldrd, 0b11, 0b01); 1467 INSN(ldrq, 0b00, 0b11); 1468 1469 #undef INSN 1470 1471 1472 #define INSN(NAME, opc, V) \ 1473 void NAME(address dest, prfop op = PLDL1KEEP) { \ 1474 int64_t offset = (dest - pc()) >> 2; \ 1475 starti; \ 1476 f(opc, 31, 30), f(0b011, 29, 27), f(V, 26), f(0b00, 25, 24), \ 1477 sf(offset, 23, 5); \ 1478 f(op, 4, 0); \ 1479 } \ 1480 void NAME(Label &L, prfop op = PLDL1KEEP) { \ 1481 wrap_label(L, op, &Assembler::NAME); \ 1482 } 1483 1484 INSN(prfm, 0b11, 0); 1485 1486 #undef INSN 1487 1488 // Load/store 1489 void ld_st1(int opc, int p1, int V, int L, 1490 Register Rt1, Register Rt2, Address adr, bool no_allocate) { 1491 starti; 1492 f(opc, 31, 30), f(p1, 29, 27), f(V, 26), f(L, 22); 1493 zrf(Rt2, 10), zrf(Rt1, 0); 1494 if (no_allocate) { 1495 adr.encode_nontemporal_pair(¤t_insn); 1496 } else { 1497 adr.encode_pair(¤t_insn); 1498 } 1499 } 1500 1501 // Load/store register pair (offset) 1502 #define INSN(NAME, size, p1, V, L, no_allocate) \ 1503 void NAME(Register Rt1, Register Rt2, Address adr) { \ 1504 ld_st1(size, p1, V, L, Rt1, Rt2, adr, no_allocate); \ 1505 } 1506 1507 INSN(stpw, 0b00, 0b101, 0, 0, false); 1508 INSN(ldpw, 0b00, 0b101, 0, 1, false); 1509 INSN(ldpsw, 0b01, 0b101, 0, 1, false); 1510 INSN(stp, 0b10, 0b101, 0, 0, false); 1511 INSN(ldp, 0b10, 0b101, 0, 1, false); 1512 1513 // Load/store no-allocate pair (offset) 1514 INSN(stnpw, 0b00, 0b101, 0, 0, true); 1515 INSN(ldnpw, 0b00, 0b101, 0, 1, true); 1516 INSN(stnp, 0b10, 0b101, 0, 0, true); 1517 INSN(ldnp, 0b10, 0b101, 0, 1, true); 1518 1519 #undef INSN 1520 1521 #define INSN(NAME, size, p1, V, L, no_allocate) \ 1522 void NAME(FloatRegister Rt1, FloatRegister Rt2, Address adr) { \ 1523 ld_st1(size, p1, V, L, \ 1524 as_Register(Rt1), as_Register(Rt2), adr, no_allocate); \ 1525 } 1526 1527 INSN(stps, 0b00, 0b101, 1, 0, false); 1528 INSN(ldps, 0b00, 0b101, 1, 1, false); 1529 INSN(stpd, 0b01, 0b101, 1, 0, false); 1530 INSN(ldpd, 0b01, 0b101, 1, 1, false); 1531 INSN(stpq, 0b10, 0b101, 1, 0, false); 1532 INSN(ldpq, 0b10, 0b101, 1, 1, false); 1533 1534 #undef INSN 1535 1536 // Load/store register (all modes) 1537 void ld_st2(Register Rt, const Address &adr, int size, int op, int V = 0) { 1538 starti; 1539 1540 f(V, 26); // general reg? 1541 zrf(Rt, 0); 1542 1543 // Encoding for literal loads is done here (rather than pushed 1544 // down into Address::encode) because the encoding of this 1545 // instruction is too different from all of the other forms to 1546 // make it worth sharing. 1547 if (adr.getMode() == Address::literal) { 1548 assert(size == 0b10 || size == 0b11, "bad operand size in ldr"); 1549 assert(op == 0b01, "literal form can only be used with loads"); 1550 f(size & 0b01, 31, 30), f(0b011, 29, 27), f(0b00, 25, 24); 1551 int64_t offset = (adr.target() - pc()) >> 2; 1552 sf(offset, 23, 5); 1553 code_section()->relocate(pc(), adr.rspec()); 1554 return; 1555 } 1556 1557 f(size, 31, 30); 1558 f(op, 23, 22); // str 1559 adr.encode(¤t_insn); 1560 } 1561 1562 #define INSN(NAME, size, op) \ 1563 void NAME(Register Rt, const Address &adr) { \ 1564 ld_st2(Rt, adr, size, op); \ 1565 } \ 1566 1567 INSN(str, 0b11, 0b00); 1568 INSN(strw, 0b10, 0b00); 1569 INSN(strb, 0b00, 0b00); 1570 INSN(strh, 0b01, 0b00); 1571 1572 INSN(ldr, 0b11, 0b01); 1573 INSN(ldrw, 0b10, 0b01); 1574 INSN(ldrb, 0b00, 0b01); 1575 INSN(ldrh, 0b01, 0b01); 1576 1577 INSN(ldrsb, 0b00, 0b10); 1578 INSN(ldrsbw, 0b00, 0b11); 1579 INSN(ldrsh, 0b01, 0b10); 1580 INSN(ldrshw, 0b01, 0b11); 1581 INSN(ldrsw, 0b10, 0b10); 1582 1583 #undef INSN 1584 1585 #define INSN(NAME, size, op) \ 1586 void NAME(FloatRegister Rt, const Address &adr) { \ 1587 ld_st2(as_Register(Rt), adr, size, op, 1); \ 1588 } 1589 1590 INSN(strd, 0b11, 0b00); 1591 INSN(strs, 0b10, 0b00); 1592 INSN(ldrd, 0b11, 0b01); 1593 INSN(ldrs, 0b10, 0b01); 1594 INSN(strq, 0b00, 0b10); 1595 INSN(ldrq, 0x00, 0b11); 1596 1597 #undef INSN 1598 1599 // Load/store a register, but with a BasicType parameter. Loaded signed integer values are 1600 // extended to 64 bits. 1601 void load(Register Rt, const Address &adr, BasicType bt) { 1602 int op = (is_signed_subword_type(bt) || bt == T_INT) ? 0b10 : 0b01; 1603 ld_st2(Rt, adr, exact_log2(type2aelembytes(bt)), op); 1604 } 1605 void store(Register Rt, const Address &adr, BasicType bt) { 1606 ld_st2(Rt, adr, exact_log2(type2aelembytes(bt)), 0b00); 1607 } 1608 1609 /* SIMD extensions 1610 * 1611 * We just use FloatRegister in the following. They are exactly the same 1612 * as SIMD registers. 1613 */ 1614 public: 1615 1616 enum SIMD_Arrangement { 1617 T8B, T16B, T4H, T8H, T2S, T4S, T1D, T2D, T1Q, INVALID_ARRANGEMENT 1618 }; 1619 1620 enum SIMD_RegVariant { 1621 B, H, S, D, Q, INVALID 1622 }; 1623 1624 private: 1625 1626 static SIMD_Arrangement _esize2arrangement_table[9][2]; 1627 static SIMD_RegVariant _esize2regvariant[9]; 1628 1629 public: 1630 1631 static SIMD_Arrangement esize2arrangement(unsigned esize, bool isQ); 1632 static SIMD_RegVariant elemType_to_regVariant(BasicType bt); 1633 static SIMD_RegVariant elemBytes_to_regVariant(unsigned esize); 1634 // Return the corresponding bits for different SIMD_RegVariant value. 1635 static unsigned regVariant_to_elemBits(SIMD_RegVariant T); 1636 1637 enum shift_kind { LSL, LSR, ASR, ROR }; 1638 1639 void op_shifted_reg(Instruction_aarch64 ¤t_insn, unsigned decode, 1640 enum shift_kind kind, unsigned shift, 1641 unsigned size, unsigned op) { 1642 f(size, 31); 1643 f(op, 30, 29); 1644 f(decode, 28, 24); 1645 f(shift, 15, 10); 1646 f(kind, 23, 22); 1647 } 1648 1649 // Logical (shifted register) 1650 #define INSN(NAME, size, op, N) \ 1651 void NAME(Register Rd, Register Rn, Register Rm, \ 1652 enum shift_kind kind = LSL, unsigned shift = 0) { \ 1653 starti; \ 1654 guarantee(size == 1 || shift < 32, "incorrect shift"); \ 1655 f(N, 21); \ 1656 zrf(Rm, 16), zrf(Rn, 5), zrf(Rd, 0); \ 1657 op_shifted_reg(current_insn, 0b01010, kind, shift, size, op); \ 1658 } 1659 1660 INSN(andr, 1, 0b00, 0); 1661 INSN(orr, 1, 0b01, 0); 1662 INSN(eor, 1, 0b10, 0); 1663 INSN(ands, 1, 0b11, 0); 1664 INSN(andw, 0, 0b00, 0); 1665 INSN(orrw, 0, 0b01, 0); 1666 INSN(eorw, 0, 0b10, 0); 1667 INSN(andsw, 0, 0b11, 0); 1668 1669 #undef INSN 1670 1671 #define INSN(NAME, size, op, N) \ 1672 void NAME(Register Rd, Register Rn, Register Rm, \ 1673 enum shift_kind kind = LSL, unsigned shift = 0) { \ 1674 starti; \ 1675 f(N, 21); \ 1676 zrf(Rm, 16), zrf(Rn, 5), zrf(Rd, 0); \ 1677 op_shifted_reg(current_insn, 0b01010, kind, shift, size, op); \ 1678 } \ 1679 \ 1680 /* These instructions have no immediate form. Provide an overload so \ 1681 that if anyone does try to use an immediate operand -- this has \ 1682 happened! -- we'll get a compile-time error. */ \ 1683 void NAME(Register Rd, Register Rn, unsigned imm, \ 1684 enum shift_kind kind = LSL, unsigned shift = 0) { \ 1685 assert(false, " can't be used with immediate operand"); \ 1686 } 1687 1688 INSN(bic, 1, 0b00, 1); 1689 INSN(orn, 1, 0b01, 1); 1690 INSN(eon, 1, 0b10, 1); 1691 INSN(bics, 1, 0b11, 1); 1692 INSN(bicw, 0, 0b00, 1); 1693 INSN(ornw, 0, 0b01, 1); 1694 INSN(eonw, 0, 0b10, 1); 1695 INSN(bicsw, 0, 0b11, 1); 1696 1697 #undef INSN 1698 1699 #ifdef _WIN64 1700 // In MSVC, `mvn` is defined as a macro and it affects compilation 1701 #undef mvn 1702 #endif 1703 1704 // Aliases for short forms of orn 1705 void mvn(Register Rd, Register Rm, 1706 enum shift_kind kind = LSL, unsigned shift = 0) { 1707 orn(Rd, zr, Rm, kind, shift); 1708 } 1709 1710 void mvnw(Register Rd, Register Rm, 1711 enum shift_kind kind = LSL, unsigned shift = 0) { 1712 ornw(Rd, zr, Rm, kind, shift); 1713 } 1714 1715 // Add/subtract (shifted register) 1716 #define INSN(NAME, size, op) \ 1717 void NAME(Register Rd, Register Rn, Register Rm, \ 1718 enum shift_kind kind, unsigned shift = 0) { \ 1719 starti; \ 1720 f(0, 21); \ 1721 assert_cond(kind != ROR); \ 1722 guarantee(size == 1 || shift < 32, "incorrect shift");\ 1723 zrf(Rd, 0), zrf(Rn, 5), zrf(Rm, 16); \ 1724 op_shifted_reg(current_insn, 0b01011, kind, shift, size, op); \ 1725 } 1726 1727 INSN(add, 1, 0b000); 1728 INSN(sub, 1, 0b10); 1729 INSN(addw, 0, 0b000); 1730 INSN(subw, 0, 0b10); 1731 1732 INSN(adds, 1, 0b001); 1733 INSN(subs, 1, 0b11); 1734 INSN(addsw, 0, 0b001); 1735 INSN(subsw, 0, 0b11); 1736 1737 #undef INSN 1738 1739 // Add/subtract (extended register) 1740 #define INSN(NAME, op) \ 1741 void NAME(Register Rd, Register Rn, Register Rm, \ 1742 ext::operation option, int amount = 0) { \ 1743 starti; \ 1744 zrf(Rm, 16), srf(Rn, 5), srf(Rd, 0); \ 1745 add_sub_extended_reg(current_insn, op, 0b01011, Rd, Rn, Rm, 0b00, option, amount); \ 1746 } 1747 1748 void add_sub_extended_reg(Instruction_aarch64 ¤t_insn, unsigned op, unsigned decode, 1749 Register Rd, Register Rn, Register Rm, 1750 unsigned opt, ext::operation option, unsigned imm) { 1751 guarantee(imm <= 4, "shift amount must be <= 4"); 1752 f(op, 31, 29), f(decode, 28, 24), f(opt, 23, 22), f(1, 21); 1753 f(option, 15, 13), f(imm, 12, 10); 1754 } 1755 1756 INSN(addw, 0b000); 1757 INSN(subw, 0b010); 1758 INSN(add, 0b100); 1759 INSN(sub, 0b110); 1760 1761 #undef INSN 1762 1763 #define INSN(NAME, op) \ 1764 void NAME(Register Rd, Register Rn, Register Rm, \ 1765 ext::operation option, int amount = 0) { \ 1766 starti; \ 1767 zrf(Rm, 16), srf(Rn, 5), zrf(Rd, 0); \ 1768 add_sub_extended_reg(current_insn, op, 0b01011, Rd, Rn, Rm, 0b00, option, amount); \ 1769 } 1770 1771 INSN(addsw, 0b001); 1772 INSN(subsw, 0b011); 1773 INSN(adds, 0b101); 1774 INSN(subs, 0b111); 1775 1776 #undef INSN 1777 1778 // Aliases for short forms of add and sub 1779 #define INSN(NAME) \ 1780 void NAME(Register Rd, Register Rn, Register Rm) { \ 1781 if (Rd == sp || Rn == sp) \ 1782 NAME(Rd, Rn, Rm, ext::uxtx); \ 1783 else \ 1784 NAME(Rd, Rn, Rm, LSL); \ 1785 } 1786 1787 INSN(addw); 1788 INSN(subw); 1789 INSN(add); 1790 INSN(sub); 1791 1792 INSN(addsw); 1793 INSN(subsw); 1794 INSN(adds); 1795 INSN(subs); 1796 1797 #undef INSN 1798 1799 // Add/subtract (with carry) 1800 void add_sub_carry(unsigned op, Register Rd, Register Rn, Register Rm) { 1801 starti; 1802 f(op, 31, 29); 1803 f(0b11010000, 28, 21); 1804 f(0b000000, 15, 10); 1805 zrf(Rm, 16), zrf(Rn, 5), zrf(Rd, 0); 1806 } 1807 1808 #define INSN(NAME, op) \ 1809 void NAME(Register Rd, Register Rn, Register Rm) { \ 1810 add_sub_carry(op, Rd, Rn, Rm); \ 1811 } 1812 1813 INSN(adcw, 0b000); 1814 INSN(adcsw, 0b001); 1815 INSN(sbcw, 0b010); 1816 INSN(sbcsw, 0b011); 1817 INSN(adc, 0b100); 1818 INSN(adcs, 0b101); 1819 INSN(sbc, 0b110); 1820 INSN(sbcs, 0b111); 1821 1822 #undef INSN 1823 1824 // Conditional compare (both kinds) 1825 void conditional_compare(unsigned op, int o1, int o2, int o3, 1826 Register Rn, unsigned imm5, unsigned nzcv, 1827 unsigned cond) { 1828 starti; 1829 f(op, 31, 29); 1830 f(0b11010010, 28, 21); 1831 f(cond, 15, 12); 1832 f(o1, 11); 1833 f(o2, 10); 1834 f(o3, 4); 1835 f(nzcv, 3, 0); 1836 f(imm5, 20, 16), zrf(Rn, 5); 1837 } 1838 1839 #define INSN(NAME, op) \ 1840 void NAME(Register Rn, Register Rm, int imm, Condition cond) { \ 1841 int regNumber = (Rm == zr ? 31 : Rm->encoding()); \ 1842 conditional_compare(op, 0, 0, 0, Rn, regNumber, imm, cond); \ 1843 } \ 1844 \ 1845 void NAME(Register Rn, int imm5, int imm, Condition cond) { \ 1846 conditional_compare(op, 1, 0, 0, Rn, imm5, imm, cond); \ 1847 } 1848 1849 INSN(ccmnw, 0b001); 1850 INSN(ccmpw, 0b011); 1851 INSN(ccmn, 0b101); 1852 INSN(ccmp, 0b111); 1853 1854 #undef INSN 1855 1856 // Conditional select 1857 void conditional_select(unsigned op, unsigned op2, 1858 Register Rd, Register Rn, Register Rm, 1859 unsigned cond) { 1860 starti; 1861 f(op, 31, 29); 1862 f(0b11010100, 28, 21); 1863 f(cond, 15, 12); 1864 f(op2, 11, 10); 1865 zrf(Rm, 16), zrf(Rn, 5), rf(Rd, 0); 1866 } 1867 1868 #define INSN(NAME, op, op2) \ 1869 void NAME(Register Rd, Register Rn, Register Rm, Condition cond) { \ 1870 conditional_select(op, op2, Rd, Rn, Rm, cond); \ 1871 } 1872 1873 INSN(cselw, 0b000, 0b00); 1874 INSN(csincw, 0b000, 0b01); 1875 INSN(csinvw, 0b010, 0b00); 1876 INSN(csnegw, 0b010, 0b01); 1877 INSN(csel, 0b100, 0b00); 1878 INSN(csinc, 0b100, 0b01); 1879 INSN(csinv, 0b110, 0b00); 1880 INSN(csneg, 0b110, 0b01); 1881 1882 #undef INSN 1883 1884 // Data processing 1885 void data_processing(Instruction_aarch64 ¤t_insn, unsigned op29, unsigned opcode, 1886 Register Rd, Register Rn) { 1887 f(op29, 31, 29), f(0b11010110, 28, 21); 1888 f(opcode, 15, 10); 1889 rf(Rn, 5), rf(Rd, 0); 1890 } 1891 1892 // (1 source) 1893 #define INSN(NAME, op29, opcode2, opcode) \ 1894 void NAME(Register Rd, Register Rn) { \ 1895 starti; \ 1896 f(opcode2, 20, 16); \ 1897 data_processing(current_insn, op29, opcode, Rd, Rn); \ 1898 } 1899 1900 INSN(rbitw, 0b010, 0b00000, 0b00000); 1901 INSN(rev16w, 0b010, 0b00000, 0b00001); 1902 INSN(revw, 0b010, 0b00000, 0b00010); 1903 INSN(clzw, 0b010, 0b00000, 0b00100); 1904 INSN(clsw, 0b010, 0b00000, 0b00101); 1905 1906 INSN(rbit, 0b110, 0b00000, 0b00000); 1907 INSN(rev16, 0b110, 0b00000, 0b00001); 1908 INSN(rev32, 0b110, 0b00000, 0b00010); 1909 INSN(rev, 0b110, 0b00000, 0b00011); 1910 INSN(clz, 0b110, 0b00000, 0b00100); 1911 INSN(cls, 0b110, 0b00000, 0b00101); 1912 1913 // PAC instructions 1914 INSN(pacia, 0b110, 0b00001, 0b00000); 1915 INSN(pacib, 0b110, 0b00001, 0b00001); 1916 INSN(pacda, 0b110, 0b00001, 0b00010); 1917 INSN(pacdb, 0b110, 0b00001, 0b00011); 1918 INSN(autia, 0b110, 0b00001, 0b00100); 1919 INSN(autib, 0b110, 0b00001, 0b00101); 1920 INSN(autda, 0b110, 0b00001, 0b00110); 1921 INSN(autdb, 0b110, 0b00001, 0b00111); 1922 1923 #undef INSN 1924 1925 #define INSN(NAME, op29, opcode2, opcode) \ 1926 void NAME(Register Rd) { \ 1927 starti; \ 1928 f(opcode2, 20, 16); \ 1929 data_processing(current_insn, op29, opcode, Rd, dummy_reg); \ 1930 } 1931 1932 // PAC instructions (with zero modifier) 1933 INSN(paciza, 0b110, 0b00001, 0b01000); 1934 INSN(pacizb, 0b110, 0b00001, 0b01001); 1935 INSN(pacdza, 0b110, 0b00001, 0b01010); 1936 INSN(pacdzb, 0b110, 0b00001, 0b01011); 1937 INSN(autiza, 0b110, 0b00001, 0b01100); 1938 INSN(autizb, 0b110, 0b00001, 0b01101); 1939 INSN(autdza, 0b110, 0b00001, 0b01110); 1940 INSN(autdzb, 0b110, 0b00001, 0b01111); 1941 INSN(xpaci, 0b110, 0b00001, 0b10000); 1942 INSN(xpacd, 0b110, 0b00001, 0b10001); 1943 1944 #undef INSN 1945 1946 // Data-processing (2 source) 1947 #define INSN(NAME, op29, opcode) \ 1948 void NAME(Register Rd, Register Rn, Register Rm) { \ 1949 starti; \ 1950 rf(Rm, 16); \ 1951 data_processing(current_insn, op29, opcode, Rd, Rn); \ 1952 } 1953 1954 INSN(udivw, 0b000, 0b000010); 1955 INSN(sdivw, 0b000, 0b000011); 1956 INSN(lslvw, 0b000, 0b001000); 1957 INSN(lsrvw, 0b000, 0b001001); 1958 INSN(asrvw, 0b000, 0b001010); 1959 INSN(rorvw, 0b000, 0b001011); 1960 1961 INSN(udiv, 0b100, 0b000010); 1962 INSN(sdiv, 0b100, 0b000011); 1963 INSN(lslv, 0b100, 0b001000); 1964 INSN(lsrv, 0b100, 0b001001); 1965 INSN(asrv, 0b100, 0b001010); 1966 INSN(rorv, 0b100, 0b001011); 1967 1968 #undef INSN 1969 1970 // Data-processing (3 source) 1971 void data_processing(unsigned op54, unsigned op31, unsigned o0, 1972 Register Rd, Register Rn, Register Rm, 1973 Register Ra) { 1974 starti; 1975 f(op54, 31, 29), f(0b11011, 28, 24); 1976 f(op31, 23, 21), f(o0, 15); 1977 zrf(Rm, 16), zrf(Ra, 10), zrf(Rn, 5), zrf(Rd, 0); 1978 } 1979 1980 #define INSN(NAME, op54, op31, o0) \ 1981 void NAME(Register Rd, Register Rn, Register Rm, Register Ra) { \ 1982 data_processing(op54, op31, o0, Rd, Rn, Rm, Ra); \ 1983 } 1984 1985 INSN(maddw, 0b000, 0b000, 0); 1986 INSN(msubw, 0b000, 0b000, 1); 1987 INSN(madd, 0b100, 0b000, 0); 1988 INSN(msub, 0b100, 0b000, 1); 1989 INSN(smaddl, 0b100, 0b001, 0); 1990 INSN(smsubl, 0b100, 0b001, 1); 1991 INSN(umaddl, 0b100, 0b101, 0); 1992 INSN(umsubl, 0b100, 0b101, 1); 1993 1994 #undef INSN 1995 1996 #define INSN(NAME, op54, op31, o0) \ 1997 void NAME(Register Rd, Register Rn, Register Rm) { \ 1998 data_processing(op54, op31, o0, Rd, Rn, Rm, as_Register(31)); \ 1999 } 2000 2001 INSN(smulh, 0b100, 0b010, 0); 2002 INSN(umulh, 0b100, 0b110, 0); 2003 2004 #undef INSN 2005 2006 // Floating-point data-processing (1 source) 2007 void data_processing(unsigned type, unsigned opcode, 2008 FloatRegister Vd, FloatRegister Vn) { 2009 starti; 2010 f(0b000, 31, 29); 2011 f(0b11110, 28, 24); 2012 f(type, 23, 22), f(1, 21), f(opcode, 20, 15), f(0b10000, 14, 10); 2013 rf(Vn, 5), rf(Vd, 0); 2014 } 2015 2016 #define INSN(NAME, type, opcode) \ 2017 void NAME(FloatRegister Vd, FloatRegister Vn) { \ 2018 data_processing(type, opcode, Vd, Vn); \ 2019 } 2020 2021 INSN(fmovs, 0b00, 0b000000); 2022 INSN(fabss, 0b00, 0b000001); 2023 INSN(fnegs, 0b00, 0b000010); 2024 INSN(fsqrts, 0b00, 0b000011); 2025 INSN(fcvts, 0b00, 0b000101); // Single-precision to double-precision 2026 INSN(fcvths, 0b11, 0b000100); // Half-precision to single-precision 2027 INSN(fcvtsh, 0b00, 0b000111); // Single-precision to half-precision 2028 2029 INSN(fmovd, 0b01, 0b000000); 2030 INSN(fabsd, 0b01, 0b000001); 2031 INSN(fnegd, 0b01, 0b000010); 2032 INSN(fsqrtd, 0b01, 0b000011); 2033 INSN(fcvtd, 0b01, 0b000100); // Double-precision to single-precision 2034 2035 private: 2036 void _fcvt_narrow_extend(FloatRegister Vd, SIMD_Arrangement Ta, 2037 FloatRegister Vn, SIMD_Arrangement Tb, bool do_extend) { 2038 assert((do_extend && (Tb >> 1) + 1 == (Ta >> 1)) 2039 || (!do_extend && (Ta >> 1) + 1 == (Tb >> 1)), "Incompatible arrangement"); 2040 starti; 2041 int op30 = (do_extend ? Tb : Ta) & 1; 2042 int op22 = ((do_extend ? Ta : Tb) >> 1) & 1; 2043 f(0, 31), f(op30, 30), f(0b0011100, 29, 23), f(op22, 22); 2044 f(0b100001011, 21, 13), f(do_extend ? 1 : 0, 12), f(0b10, 11, 10); 2045 rf(Vn, 5), rf(Vd, 0); 2046 } 2047 2048 public: 2049 void fcvtl(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb) { 2050 assert(Tb == T4H || Tb == T8H|| Tb == T2S || Tb == T4S, "invalid arrangement"); 2051 _fcvt_narrow_extend(Vd, Ta, Vn, Tb, true); 2052 } 2053 2054 void fcvtn(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb) { 2055 assert(Ta == T4H || Ta == T8H|| Ta == T2S || Ta == T4S, "invalid arrangement"); 2056 _fcvt_narrow_extend(Vd, Ta, Vn, Tb, false); 2057 } 2058 2059 #undef INSN 2060 2061 // Floating-point data-processing (2 source) 2062 void data_processing(unsigned op31, unsigned type, unsigned opcode, 2063 FloatRegister Vd, FloatRegister Vn, FloatRegister Vm) { 2064 starti; 2065 f(op31, 31, 29); 2066 f(0b11110, 28, 24); 2067 f(type, 23, 22), f(1, 21), f(opcode, 15, 10); 2068 rf(Vm, 16), rf(Vn, 5), rf(Vd, 0); 2069 } 2070 2071 #define INSN(NAME, op31, type, opcode) \ 2072 void NAME(FloatRegister Vd, FloatRegister Vn, FloatRegister Vm) { \ 2073 data_processing(op31, type, opcode, Vd, Vn, Vm); \ 2074 } 2075 2076 INSN(fabds, 0b011, 0b10, 0b110101); 2077 INSN(fmuls, 0b000, 0b00, 0b000010); 2078 INSN(fdivs, 0b000, 0b00, 0b000110); 2079 INSN(fadds, 0b000, 0b00, 0b001010); 2080 INSN(fsubs, 0b000, 0b00, 0b001110); 2081 INSN(fmaxs, 0b000, 0b00, 0b010010); 2082 INSN(fmins, 0b000, 0b00, 0b010110); 2083 INSN(fnmuls, 0b000, 0b00, 0b100010); 2084 2085 INSN(fabdd, 0b011, 0b11, 0b110101); 2086 INSN(fmuld, 0b000, 0b01, 0b000010); 2087 INSN(fdivd, 0b000, 0b01, 0b000110); 2088 INSN(faddd, 0b000, 0b01, 0b001010); 2089 INSN(fsubd, 0b000, 0b01, 0b001110); 2090 INSN(fmaxd, 0b000, 0b01, 0b010010); 2091 INSN(fmind, 0b000, 0b01, 0b010110); 2092 INSN(fnmuld, 0b000, 0b01, 0b100010); 2093 2094 #undef INSN 2095 2096 // Floating-point data-processing (3 source) 2097 void data_processing(unsigned op31, unsigned type, unsigned o1, unsigned o0, 2098 FloatRegister Vd, FloatRegister Vn, FloatRegister Vm, 2099 FloatRegister Va) { 2100 starti; 2101 f(op31, 31, 29); 2102 f(0b11111, 28, 24); 2103 f(type, 23, 22), f(o1, 21), f(o0, 15); 2104 rf(Vm, 16), rf(Va, 10), rf(Vn, 5), rf(Vd, 0); 2105 } 2106 2107 #define INSN(NAME, op31, type, o1, o0) \ 2108 void NAME(FloatRegister Vd, FloatRegister Vn, FloatRegister Vm, \ 2109 FloatRegister Va) { \ 2110 data_processing(op31, type, o1, o0, Vd, Vn, Vm, Va); \ 2111 } 2112 2113 INSN(fmadds, 0b000, 0b00, 0, 0); 2114 INSN(fmsubs, 0b000, 0b00, 0, 1); 2115 INSN(fnmadds, 0b000, 0b00, 1, 0); 2116 INSN(fnmsubs, 0b000, 0b00, 1, 1); 2117 2118 INSN(fmaddd, 0b000, 0b01, 0, 0); 2119 INSN(fmsubd, 0b000, 0b01, 0, 1); 2120 INSN(fnmaddd, 0b000, 0b01, 1, 0); 2121 INSN(fnmsub, 0b000, 0b01, 1, 1); 2122 2123 #undef INSN 2124 2125 // Floating-point conditional select 2126 void fp_conditional_select(unsigned op31, unsigned type, 2127 unsigned op1, unsigned op2, 2128 Condition cond, FloatRegister Vd, 2129 FloatRegister Vn, FloatRegister Vm) { 2130 starti; 2131 f(op31, 31, 29); 2132 f(0b11110, 28, 24); 2133 f(type, 23, 22); 2134 f(op1, 21, 21); 2135 f(op2, 11, 10); 2136 f(cond, 15, 12); 2137 rf(Vm, 16), rf(Vn, 5), rf(Vd, 0); 2138 } 2139 2140 #define INSN(NAME, op31, type, op1, op2) \ 2141 void NAME(FloatRegister Vd, FloatRegister Vn, \ 2142 FloatRegister Vm, Condition cond) { \ 2143 fp_conditional_select(op31, type, op1, op2, cond, Vd, Vn, Vm); \ 2144 } 2145 2146 INSN(fcsels, 0b000, 0b00, 0b1, 0b11); 2147 INSN(fcseld, 0b000, 0b01, 0b1, 0b11); 2148 2149 #undef INSN 2150 2151 // Conversion between floating-point and integer 2152 void float_int_convert(unsigned sflag, unsigned ftype, 2153 unsigned rmode, unsigned opcode, 2154 Register Rd, Register Rn) { 2155 starti; 2156 f(sflag, 31); 2157 f(0b00, 30, 29); 2158 f(0b11110, 28, 24); 2159 f(ftype, 23, 22), f(1, 21), f(rmode, 20, 19); 2160 f(opcode, 18, 16), f(0b000000, 15, 10); 2161 zrf(Rn, 5), zrf(Rd, 0); 2162 } 2163 2164 #define INSN(NAME, sflag, ftype, rmode, opcode) \ 2165 void NAME(Register Rd, FloatRegister Vn) { \ 2166 float_int_convert(sflag, ftype, rmode, opcode, Rd, as_Register(Vn)); \ 2167 } 2168 2169 INSN(fcvtzsw, 0b0, 0b00, 0b11, 0b000); 2170 INSN(fcvtzs, 0b1, 0b00, 0b11, 0b000); 2171 INSN(fcvtzdw, 0b0, 0b01, 0b11, 0b000); 2172 INSN(fcvtzd, 0b1, 0b01, 0b11, 0b000); 2173 2174 // RoundToNearestTiesAway 2175 INSN(fcvtassw, 0b0, 0b00, 0b00, 0b100); // float -> signed word 2176 INSN(fcvtasd, 0b1, 0b01, 0b00, 0b100); // double -> signed xword 2177 2178 // RoundTowardsNegative 2179 INSN(fcvtmssw, 0b0, 0b00, 0b10, 0b000); // float -> signed word 2180 INSN(fcvtmsd, 0b1, 0b01, 0b10, 0b000); // double -> signed xword 2181 2182 INSN(fmovs, 0b0, 0b00, 0b00, 0b110); 2183 INSN(fmovd, 0b1, 0b01, 0b00, 0b110); 2184 2185 INSN(fmovhid, 0b1, 0b10, 0b01, 0b110); 2186 2187 #undef INSN 2188 2189 #define INSN(NAME, sflag, type, rmode, opcode) \ 2190 void NAME(FloatRegister Vd, Register Rn) { \ 2191 float_int_convert(sflag, type, rmode, opcode, as_Register(Vd), Rn); \ 2192 } 2193 2194 INSN(fmovs, 0b0, 0b00, 0b00, 0b111); 2195 INSN(fmovd, 0b1, 0b01, 0b00, 0b111); 2196 2197 INSN(scvtfws, 0b0, 0b00, 0b00, 0b010); 2198 INSN(scvtfs, 0b1, 0b00, 0b00, 0b010); 2199 INSN(scvtfwd, 0b0, 0b01, 0b00, 0b010); 2200 INSN(scvtfd, 0b1, 0b01, 0b00, 0b010); 2201 2202 // INSN(fmovhid, 0b100, 0b10, 0b01, 0b111); 2203 2204 #undef INSN 2205 2206 private: 2207 void _xcvtf_vector_integer(bool is_unsigned, SIMD_Arrangement T, 2208 FloatRegister Rd, FloatRegister Rn) { 2209 assert(T == T2S || T == T4S || T == T2D, "invalid arrangement"); 2210 starti; 2211 f(0, 31), f(T & 1, 30), f(is_unsigned ? 1 : 0, 29); 2212 f(0b011100, 28, 23), f((T >> 1) & 1, 22), f(0b100001110110, 21, 10); 2213 rf(Rn, 5), rf(Rd, 0); 2214 } 2215 2216 public: 2217 2218 void scvtfv(SIMD_Arrangement T, FloatRegister Rd, FloatRegister Rn) { 2219 _xcvtf_vector_integer(/* is_unsigned */ false, T, Rd, Rn); 2220 } 2221 2222 // Floating-point compare 2223 void float_compare(unsigned op31, unsigned type, 2224 unsigned op, unsigned op2, 2225 FloatRegister Vn, FloatRegister Vm = as_FloatRegister(0)) { 2226 starti; 2227 f(op31, 31, 29); 2228 f(0b11110, 28, 24); 2229 f(type, 23, 22), f(1, 21); 2230 f(op, 15, 14), f(0b1000, 13, 10), f(op2, 4, 0); 2231 rf(Vn, 5), rf(Vm, 16); 2232 } 2233 2234 2235 #define INSN(NAME, op31, type, op, op2) \ 2236 void NAME(FloatRegister Vn, FloatRegister Vm) { \ 2237 float_compare(op31, type, op, op2, Vn, Vm); \ 2238 } 2239 2240 #define INSN1(NAME, op31, type, op, op2) \ 2241 void NAME(FloatRegister Vn, double d) { \ 2242 assert_cond(d == 0.0); \ 2243 float_compare(op31, type, op, op2, Vn); \ 2244 } 2245 2246 INSN(fcmps, 0b000, 0b00, 0b00, 0b00000); 2247 INSN1(fcmps, 0b000, 0b00, 0b00, 0b01000); 2248 // INSN(fcmpes, 0b000, 0b00, 0b00, 0b10000); 2249 // INSN1(fcmpes, 0b000, 0b00, 0b00, 0b11000); 2250 2251 INSN(fcmpd, 0b000, 0b01, 0b00, 0b00000); 2252 INSN1(fcmpd, 0b000, 0b01, 0b00, 0b01000); 2253 // INSN(fcmped, 0b000, 0b01, 0b00, 0b10000); 2254 // INSN1(fcmped, 0b000, 0b01, 0b00, 0b11000); 2255 2256 #undef INSN 2257 #undef INSN1 2258 2259 // Floating-point compare. 3-registers versions (scalar). 2260 #define INSN(NAME, sz, e) \ 2261 void NAME(FloatRegister Vd, FloatRegister Vn, FloatRegister Vm) { \ 2262 starti; \ 2263 f(0b01111110, 31, 24), f(e, 23), f(sz, 22), f(1, 21), rf(Vm, 16); \ 2264 f(0b111011, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2265 } \ 2266 2267 INSN(facged, 1, 0); // facge-double 2268 INSN(facges, 0, 0); // facge-single 2269 INSN(facgtd, 1, 1); // facgt-double 2270 INSN(facgts, 0, 1); // facgt-single 2271 2272 #undef INSN 2273 2274 // Floating-point Move (immediate) 2275 private: 2276 unsigned pack(double value); 2277 2278 void fmov_imm(FloatRegister Vn, double value, unsigned size) { 2279 starti; 2280 f(0b00011110, 31, 24), f(size, 23, 22), f(1, 21); 2281 f(pack(value), 20, 13), f(0b10000000, 12, 5); 2282 rf(Vn, 0); 2283 } 2284 2285 public: 2286 2287 void fmovs(FloatRegister Vn, double value) { 2288 if (value) 2289 fmov_imm(Vn, value, 0b00); 2290 else 2291 movi(Vn, T2S, 0); 2292 } 2293 void fmovd(FloatRegister Vn, double value) { 2294 if (value) 2295 fmov_imm(Vn, value, 0b01); 2296 else 2297 movi(Vn, T1D, 0); 2298 } 2299 2300 // Floating-point data-processing (1 source) 2301 2302 // Floating-point rounding 2303 // type: half-precision = 11 2304 // single = 00 2305 // double = 01 2306 // rmode: A = Away = 100 2307 // I = current = 111 2308 // M = MinusInf = 010 2309 // N = eveN = 000 2310 // P = PlusInf = 001 2311 // X = eXact = 110 2312 // Z = Zero = 011 2313 void float_round(unsigned type, unsigned rmode, FloatRegister Rd, FloatRegister Rn) { 2314 starti; 2315 f(0b00011110, 31, 24); 2316 f(type, 23, 22); 2317 f(0b1001, 21, 18); 2318 f(rmode, 17, 15); 2319 f(0b10000, 14, 10); 2320 rf(Rn, 5), rf(Rd, 0); 2321 } 2322 #define INSN(NAME, type, rmode) \ 2323 void NAME(FloatRegister Vd, FloatRegister Vn) { \ 2324 float_round(type, rmode, Vd, Vn); \ 2325 } 2326 2327 public: 2328 INSN(frintah, 0b11, 0b100); 2329 INSN(frintih, 0b11, 0b111); 2330 INSN(frintmh, 0b11, 0b010); 2331 INSN(frintnh, 0b11, 0b000); 2332 INSN(frintph, 0b11, 0b001); 2333 INSN(frintxh, 0b11, 0b110); 2334 INSN(frintzh, 0b11, 0b011); 2335 2336 INSN(frintas, 0b00, 0b100); 2337 INSN(frintis, 0b00, 0b111); 2338 INSN(frintms, 0b00, 0b010); 2339 INSN(frintns, 0b00, 0b000); 2340 INSN(frintps, 0b00, 0b001); 2341 INSN(frintxs, 0b00, 0b110); 2342 INSN(frintzs, 0b00, 0b011); 2343 2344 INSN(frintad, 0b01, 0b100); 2345 INSN(frintid, 0b01, 0b111); 2346 INSN(frintmd, 0b01, 0b010); 2347 INSN(frintnd, 0b01, 0b000); 2348 INSN(frintpd, 0b01, 0b001); 2349 INSN(frintxd, 0b01, 0b110); 2350 INSN(frintzd, 0b01, 0b011); 2351 #undef INSN 2352 2353 private: 2354 static short SIMD_Size_in_bytes[]; 2355 2356 public: 2357 #define INSN(NAME, op) \ 2358 void NAME(FloatRegister Rt, SIMD_RegVariant T, const Address &adr) { \ 2359 ld_st2(as_Register(Rt), adr, (int)T & 3, op + ((T==Q) ? 0b10:0b00), 1); \ 2360 } 2361 2362 INSN(ldr, 1); 2363 INSN(str, 0); 2364 2365 #undef INSN 2366 2367 private: 2368 2369 void ld_st(FloatRegister Vt, SIMD_Arrangement T, Register Xn, int op1, int op2) { 2370 starti; 2371 f(0,31), f((int)T & 1, 30); 2372 f(op1, 29, 21), f(0, 20, 16), f(op2, 15, 12); 2373 f((int)T >> 1, 11, 10), srf(Xn, 5), rf(Vt, 0); 2374 } 2375 void ld_st(FloatRegister Vt, SIMD_Arrangement T, Register Xn, 2376 int imm, int op1, int op2, int regs) { 2377 2378 bool replicate = op2 >> 2 == 3; 2379 // post-index value (imm) is formed differently for replicate/non-replicate ld* instructions 2380 int expectedImmediate = replicate ? regs * (1 << (T >> 1)) : SIMD_Size_in_bytes[T] * regs; 2381 guarantee(T < T1Q , "incorrect arrangement"); 2382 guarantee(imm == expectedImmediate, "bad offset"); 2383 starti; 2384 f(0,31), f((int)T & 1, 30); 2385 f(op1 | 0b100, 29, 21), f(0b11111, 20, 16), f(op2, 15, 12); 2386 f((int)T >> 1, 11, 10), srf(Xn, 5), rf(Vt, 0); 2387 } 2388 void ld_st(FloatRegister Vt, SIMD_Arrangement T, Register Xn, 2389 Register Xm, int op1, int op2) { 2390 starti; 2391 f(0,31), f((int)T & 1, 30); 2392 f(op1 | 0b100, 29, 21), rf(Xm, 16), f(op2, 15, 12); 2393 f((int)T >> 1, 11, 10), srf(Xn, 5), rf(Vt, 0); 2394 } 2395 2396 void ld_st(FloatRegister Vt, SIMD_Arrangement T, Address a, int op1, int op2, int regs) { 2397 switch (a.getMode()) { 2398 case Address::base_plus_offset: 2399 guarantee(a.offset() == 0, "no offset allowed here"); 2400 ld_st(Vt, T, a.base(), op1, op2); 2401 break; 2402 case Address::post: 2403 ld_st(Vt, T, a.base(), checked_cast<int>(a.offset()), op1, op2, regs); 2404 break; 2405 case Address::post_reg: 2406 ld_st(Vt, T, a.base(), a.index(), op1, op2); 2407 break; 2408 default: 2409 ShouldNotReachHere(); 2410 } 2411 } 2412 2413 // Single-structure load/store method (all addressing variants) 2414 void ld_st(FloatRegister Vt, SIMD_RegVariant T, int index, Address a, 2415 int op1, int op2, int regs) { 2416 int expectedImmediate = (regVariant_to_elemBits(T) >> 3) * regs; 2417 int sVal = (T < D) ? (index >> (2 - T)) & 0x01 : 0; 2418 int opcode = (T < D) ? (T << 2) : ((T & 0x02) << 2); 2419 int size = (T < D) ? (index & (0x3 << T)) : 1; // only care about low 2b 2420 Register Xn = a.base(); 2421 int Rm; 2422 2423 switch (a.getMode()) { 2424 case Address::base_plus_offset: 2425 guarantee(a.offset() == 0, "no offset allowed here"); 2426 Rm = 0; 2427 break; 2428 case Address::post: 2429 guarantee(a.offset() == expectedImmediate, "bad offset"); 2430 op1 |= 0b100; 2431 Rm = 0b11111; 2432 break; 2433 case Address::post_reg: 2434 op1 |= 0b100; 2435 Rm = a.index()->encoding(); 2436 break; 2437 default: 2438 ShouldNotReachHere(); 2439 Rm = 0; // unreachable 2440 } 2441 2442 starti; 2443 f(0,31), f((index >> (3 - T)), 30); 2444 f(op1, 29, 21), f(Rm, 20, 16), f(op2 | opcode | sVal, 15, 12); 2445 f(size, 11, 10), srf(Xn, 5), rf(Vt, 0); 2446 } 2447 2448 public: 2449 2450 #define INSN1(NAME, op1, op2) \ 2451 void NAME(FloatRegister Vt, SIMD_Arrangement T, const Address &a) { \ 2452 ld_st(Vt, T, a, op1, op2, 1); \ 2453 } 2454 2455 #define INSN2(NAME, op1, op2) \ 2456 void NAME(FloatRegister Vt, FloatRegister Vt2, SIMD_Arrangement T, const Address &a) { \ 2457 assert(Vt->successor() == Vt2, "Registers must be ordered"); \ 2458 ld_st(Vt, T, a, op1, op2, 2); \ 2459 } 2460 2461 #define INSN3(NAME, op1, op2) \ 2462 void NAME(FloatRegister Vt, FloatRegister Vt2, FloatRegister Vt3, \ 2463 SIMD_Arrangement T, const Address &a) { \ 2464 assert(Vt->successor() == Vt2 && Vt2->successor() == Vt3, \ 2465 "Registers must be ordered"); \ 2466 ld_st(Vt, T, a, op1, op2, 3); \ 2467 } 2468 2469 #define INSN4(NAME, op1, op2) \ 2470 void NAME(FloatRegister Vt, FloatRegister Vt2, FloatRegister Vt3, \ 2471 FloatRegister Vt4, SIMD_Arrangement T, const Address &a) { \ 2472 assert(Vt->successor() == Vt2 && Vt2->successor() == Vt3 && \ 2473 Vt3->successor() == Vt4, "Registers must be ordered"); \ 2474 ld_st(Vt, T, a, op1, op2, 4); \ 2475 } 2476 2477 INSN1(ld1, 0b001100010, 0b0111); 2478 INSN2(ld1, 0b001100010, 0b1010); 2479 INSN3(ld1, 0b001100010, 0b0110); 2480 INSN4(ld1, 0b001100010, 0b0010); 2481 2482 INSN2(ld2, 0b001100010, 0b1000); 2483 INSN3(ld3, 0b001100010, 0b0100); 2484 INSN4(ld4, 0b001100010, 0b0000); 2485 2486 INSN1(st1, 0b001100000, 0b0111); 2487 INSN2(st1, 0b001100000, 0b1010); 2488 INSN3(st1, 0b001100000, 0b0110); 2489 INSN4(st1, 0b001100000, 0b0010); 2490 2491 INSN2(st2, 0b001100000, 0b1000); 2492 INSN3(st3, 0b001100000, 0b0100); 2493 INSN4(st4, 0b001100000, 0b0000); 2494 2495 INSN1(ld1r, 0b001101010, 0b1100); 2496 INSN2(ld2r, 0b001101011, 0b1100); 2497 INSN3(ld3r, 0b001101010, 0b1110); 2498 INSN4(ld4r, 0b001101011, 0b1110); 2499 2500 #undef INSN1 2501 #undef INSN2 2502 #undef INSN3 2503 #undef INSN4 2504 2505 // Handle common single-structure ld/st parameter sanity checks 2506 // for all variations (1 to 4) of SIMD reigster inputs. This 2507 // method will call the routine that generates the opcode. 2508 template<typename R, typename... Rx> 2509 void ldst_sstr(SIMD_RegVariant T, int index, const Address &a, 2510 int op1, int op2, R firstReg, Rx... otherRegs) { 2511 const FloatRegister vtSet[] = { firstReg, otherRegs... }; 2512 const int regCount = sizeof...(otherRegs) + 1; 2513 assert(index >= 0 && (T <= D) && ((T == B && index <= 15) || 2514 (T == H && index <= 7) || (T == S && index <= 3) || 2515 (T == D && index <= 1)), "invalid index"); 2516 assert(regCount >= 1 && regCount <= 4, "illegal register count"); 2517 2518 // Check to make sure when multiple SIMD registers are used 2519 // that they are in successive order. 2520 for (int i = 0; i < regCount - 1; i++) { 2521 assert(vtSet[i]->successor() == vtSet[i + 1], 2522 "Registers must be ordered"); 2523 } 2524 2525 ld_st(firstReg, T, index, a, op1, op2, regCount); 2526 } 2527 2528 // Define a set of INSN1/2/3/4 macros to handle single-structure 2529 // load/store instructions. 2530 #define INSN1(NAME, op1, op2) \ 2531 void NAME(FloatRegister Vt, SIMD_RegVariant T, int index, \ 2532 const Address &a) { \ 2533 ldst_sstr(T, index, a, op1, op2, Vt); \ 2534 } 2535 2536 #define INSN2(NAME, op1, op2) \ 2537 void NAME(FloatRegister Vt, FloatRegister Vt2, SIMD_RegVariant T, \ 2538 int index, const Address &a) { \ 2539 ldst_sstr(T, index, a, op1, op2, Vt, Vt2); \ 2540 } 2541 2542 #define INSN3(NAME, op1, op2) \ 2543 void NAME(FloatRegister Vt, FloatRegister Vt2, FloatRegister Vt3, \ 2544 SIMD_RegVariant T, int index, const Address &a) { \ 2545 ldst_sstr(T, index, a, op1, op2, Vt, Vt2, Vt3); \ 2546 } 2547 2548 #define INSN4(NAME, op1, op2) \ 2549 void NAME(FloatRegister Vt, FloatRegister Vt2, FloatRegister Vt3, \ 2550 FloatRegister Vt4, SIMD_RegVariant T, int index, \ 2551 const Address &a) { \ 2552 ldst_sstr(T, index, a, op1, op2, Vt, Vt2, Vt3, Vt4); \ 2553 } 2554 2555 INSN1(ld1, 0b001101010, 0b0000); 2556 INSN2(ld2, 0b001101011, 0b0000); 2557 INSN3(ld3, 0b001101010, 0b0010); 2558 INSN4(ld4, 0b001101011, 0b0010); 2559 2560 INSN1(st1, 0b001101000, 0b0000); 2561 INSN2(st2, 0b001101001, 0b0000); 2562 INSN3(st3, 0b001101000, 0b0010); 2563 INSN4(st4, 0b001101001, 0b0010); 2564 2565 #undef INSN1 2566 #undef INSN2 2567 #undef INSN3 2568 #undef INSN4 2569 2570 #define INSN(NAME, opc) \ 2571 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \ 2572 starti; \ 2573 assert(T == T8B || T == T16B, "must be T8B or T16B"); \ 2574 f(0, 31), f((int)T & 1, 30), f(opc, 29, 21); \ 2575 rf(Vm, 16), f(0b000111, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2576 } 2577 2578 INSN(eor, 0b101110001); 2579 INSN(orr, 0b001110101); 2580 INSN(andr, 0b001110001); 2581 INSN(bic, 0b001110011); 2582 INSN(bif, 0b101110111); 2583 INSN(bit, 0b101110101); 2584 INSN(bsl, 0b101110011); 2585 INSN(orn, 0b001110111); 2586 2587 #undef INSN 2588 2589 // Advanced SIMD three different 2590 #define INSN(NAME, opc, opc2, acceptT2D) \ 2591 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \ 2592 guarantee(T != T1Q && T != T1D, "incorrect arrangement"); \ 2593 if (!acceptT2D) guarantee(T != T2D, "incorrect arrangement"); \ 2594 if (opc2 == 0b101101) guarantee(T != T8B && T != T16B, "incorrect arrangement"); \ 2595 starti; \ 2596 f(0, 31), f((int)T & 1, 30), f(opc, 29), f(0b01110, 28, 24); \ 2597 f((int)T >> 1, 23, 22), f(1, 21), rf(Vm, 16), f(opc2, 15, 10); \ 2598 rf(Vn, 5), rf(Vd, 0); \ 2599 } 2600 2601 INSN(addv, 0, 0b100001, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2602 INSN(subv, 1, 0b100001, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2603 INSN(sqaddv, 0, 0b000011, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2604 INSN(sqsubv, 0, 0b001011, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2605 INSN(uqaddv, 1, 0b000011, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2606 INSN(uqsubv, 1, 0b001011, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2607 INSN(mulv, 0, 0b100111, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2608 INSN(mlav, 0, 0b100101, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2609 INSN(mlsv, 1, 0b100101, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2610 INSN(sshl, 0, 0b010001, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2611 INSN(ushl, 1, 0b010001, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2612 INSN(addpv, 0, 0b101111, true); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2613 INSN(smullv, 0, 0b110000, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2614 INSN(umullv, 1, 0b110000, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2615 INSN(smlalv, 0, 0b100000, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2616 INSN(umlalv, 1, 0b100000, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2617 INSN(maxv, 0, 0b011001, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2618 INSN(minv, 0, 0b011011, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2619 INSN(umaxv, 1, 0b011001, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2620 INSN(uminv, 1, 0b011011, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2621 INSN(smaxp, 0, 0b101001, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2622 INSN(sminp, 0, 0b101011, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2623 INSN(sqdmulh,0, 0b101101, false); // accepted arrangements: T4H, T8H, T2S, T4S 2624 INSN(shsubv, 0, 0b001001, false); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2625 2626 #undef INSN 2627 2628 // Advanced SIMD across lanes 2629 #define INSN(NAME, opc, opc2, accepted) \ 2630 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn) { \ 2631 guarantee(T != T1Q && T != T1D, "incorrect arrangement"); \ 2632 if (accepted < 3) guarantee(T != T2D, "incorrect arrangement"); \ 2633 if (accepted < 2) guarantee(T != T2S, "incorrect arrangement"); \ 2634 if (accepted < 1) guarantee(T == T8B || T == T16B, "incorrect arrangement"); \ 2635 starti; \ 2636 f(0, 31), f((int)T & 1, 30), f(opc, 29), f(0b01110, 28, 24); \ 2637 f((int)T >> 1, 23, 22), f(opc2, 21, 10); \ 2638 rf(Vn, 5), rf(Vd, 0); \ 2639 } 2640 2641 INSN(absr, 0, 0b100000101110, 3); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2642 INSN(negr, 1, 0b100000101110, 3); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S, T2D 2643 INSN(notr, 1, 0b100000010110, 0); // accepted arrangements: T8B, T16B 2644 INSN(addv, 0, 0b110001101110, 1); // accepted arrangements: T8B, T16B, T4H, T8H, T4S 2645 INSN(smaxv, 0, 0b110000101010, 1); // accepted arrangements: T8B, T16B, T4H, T8H, T4S 2646 INSN(umaxv, 1, 0b110000101010, 1); // accepted arrangements: T8B, T16B, T4H, T8H, T4S 2647 INSN(sminv, 0, 0b110001101010, 1); // accepted arrangements: T8B, T16B, T4H, T8H, T4S 2648 INSN(uminv, 1, 0b110001101010, 1); // accepted arrangements: T8B, T16B, T4H, T8H, T4S 2649 INSN(cls, 0, 0b100000010010, 2); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2650 INSN(clz, 1, 0b100000010010, 2); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2651 INSN(cnt, 0, 0b100000010110, 0); // accepted arrangements: T8B, T16B 2652 INSN(uaddlp, 1, 0b100000001010, 2); // accepted arrangements: T8B, T16B, T4H, T8H, T2S, T4S 2653 INSN(uaddlv, 1, 0b110000001110, 1); // accepted arrangements: T8B, T16B, T4H, T8H, T4S 2654 2655 #undef INSN 2656 2657 #define INSN(NAME, opc) \ 2658 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn) { \ 2659 starti; \ 2660 assert(T == T4S, "arrangement must be T4S"); \ 2661 f(0, 31), f((int)T & 1, 30), f(0b101110, 29, 24), f(opc, 23), \ 2662 f(T == T4S ? 0 : 1, 22), f(0b110000111110, 21, 10); rf(Vn, 5), rf(Vd, 0); \ 2663 } 2664 2665 INSN(fmaxv, 0); 2666 INSN(fminv, 1); 2667 2668 #undef INSN 2669 2670 // Advanced SIMD modified immediate 2671 #define INSN(NAME, op0, cmode0) \ 2672 void NAME(FloatRegister Vd, SIMD_Arrangement T, unsigned imm8, unsigned lsl = 0) { \ 2673 unsigned cmode = cmode0; \ 2674 unsigned op = op0; \ 2675 starti; \ 2676 assert(lsl == 0 || \ 2677 ((T == T4H || T == T8H) && lsl == 8) || \ 2678 ((T == T2S || T == T4S) && ((lsl >> 3) < 4) && ((lsl & 7) == 0)), "invalid shift");\ 2679 cmode |= lsl >> 2; \ 2680 if (T == T4H || T == T8H) cmode |= 0b1000; \ 2681 if (!(T == T4H || T == T8H || T == T2S || T == T4S)) { \ 2682 assert(op == 0 && cmode0 == 0, "must be MOVI"); \ 2683 cmode = 0b1110; \ 2684 if (T == T1D || T == T2D) op = 1; \ 2685 } \ 2686 f(0, 31), f((int)T & 1, 30), f(op, 29), f(0b0111100000, 28, 19); \ 2687 f(imm8 >> 5, 18, 16), f(cmode, 15, 12), f(0x01, 11, 10), f(imm8 & 0b11111, 9, 5); \ 2688 rf(Vd, 0); \ 2689 } 2690 2691 INSN(movi, 0, 0); 2692 INSN(orri, 0, 1); 2693 INSN(mvni, 1, 0); 2694 INSN(bici, 1, 1); 2695 2696 #undef INSN 2697 2698 #define INSN(NAME, op, cmode) \ 2699 void NAME(FloatRegister Vd, SIMD_Arrangement T, double imm) { \ 2700 unsigned imm8 = pack(imm); \ 2701 starti; \ 2702 f(0, 31), f((int)T & 1, 30), f(op, 29), f(0b0111100000, 28, 19); \ 2703 f(imm8 >> 5, 18, 16), f(cmode, 15, 12), f(0x01, 11, 10), f(imm8 & 0b11111, 9, 5); \ 2704 rf(Vd, 0); \ 2705 } 2706 2707 INSN(fmovs, 0, 0b1111); 2708 INSN(fmovd, 1, 0b1111); 2709 2710 #undef INSN 2711 2712 // Advanced SIMD three same 2713 #define INSN(NAME, op1, op2, op3) \ 2714 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \ 2715 starti; \ 2716 assert(T == T2S || T == T4S || T == T2D, "invalid arrangement"); \ 2717 f(0, 31), f((int)T & 1, 30), f(op1, 29), f(0b01110, 28, 24), f(op2, 23); \ 2718 f(T==T2D ? 1:0, 22); f(1, 21), rf(Vm, 16), f(op3, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2719 } 2720 2721 INSN(fabd, 1, 1, 0b110101); 2722 INSN(fadd, 0, 0, 0b110101); 2723 INSN(fdiv, 1, 0, 0b111111); 2724 INSN(faddp, 1, 0, 0b110101); 2725 INSN(fmul, 1, 0, 0b110111); 2726 INSN(fsub, 0, 1, 0b110101); 2727 INSN(fmla, 0, 0, 0b110011); 2728 INSN(fmls, 0, 1, 0b110011); 2729 INSN(fmax, 0, 0, 0b111101); 2730 INSN(fmin, 0, 1, 0b111101); 2731 INSN(facgt, 1, 1, 0b111011); 2732 2733 #undef INSN 2734 2735 // AdvSIMD vector compare 2736 void cm(Condition cond, FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { 2737 starti; 2738 assert(T != T1Q && T != T1D, "incorrect arrangement"); 2739 int cond_op; 2740 switch (cond) { 2741 case EQ: cond_op = 0b110001; break; 2742 case GT: cond_op = 0b000110; break; 2743 case GE: cond_op = 0b000111; break; 2744 case HI: cond_op = 0b100110; break; 2745 case HS: cond_op = 0b100111; break; 2746 default: 2747 ShouldNotReachHere(); 2748 break; 2749 } 2750 2751 f(0, 31), f((int)T & 1, 30), f((cond_op >> 5) & 1, 29); 2752 f(0b01110, 28, 24), f((int)T >> 1, 23, 22), f(1, 21), rf(Vm, 16); 2753 f(cond_op & 0b11111, 15, 11), f(1, 10), rf(Vn, 5), rf(Vd, 0); 2754 } 2755 2756 // AdvSIMD Floating-point vector compare 2757 void fcm(Condition cond, FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { 2758 starti; 2759 assert(T == T2S || T == T4S || T == T2D, "invalid arrangement"); 2760 int cond_op; 2761 switch (cond) { 2762 case EQ: cond_op = 0b00; break; 2763 case GT: cond_op = 0b11; break; 2764 case GE: cond_op = 0b10; break; 2765 default: 2766 ShouldNotReachHere(); 2767 break; 2768 } 2769 2770 f(0, 31), f((int)T & 1, 30), f((cond_op >> 1) & 1, 29); 2771 f(0b01110, 28, 24), f(cond_op & 1, 23), f(T == T2D ? 1 : 0, 22); 2772 f(1, 21), rf(Vm, 16), f(0b111001, 15, 10), rf(Vn, 5), rf(Vd, 0); 2773 } 2774 2775 #define INSN(NAME, opc) \ 2776 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \ 2777 starti; \ 2778 assert(T == T4S, "arrangement must be T4S"); \ 2779 f(0b01011110000, 31, 21), rf(Vm, 16), f(opc, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2780 } 2781 2782 INSN(sha1c, 0b000000); 2783 INSN(sha1m, 0b001000); 2784 INSN(sha1p, 0b000100); 2785 INSN(sha1su0, 0b001100); 2786 INSN(sha256h2, 0b010100); 2787 INSN(sha256h, 0b010000); 2788 INSN(sha256su1, 0b011000); 2789 2790 #undef INSN 2791 2792 #define INSN(NAME, opc) \ 2793 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn) { \ 2794 starti; \ 2795 assert(T == T4S, "arrangement must be T4S"); \ 2796 f(0b0101111000101000, 31, 16), f(opc, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2797 } 2798 2799 INSN(sha1h, 0b000010); 2800 INSN(sha1su1, 0b000110); 2801 INSN(sha256su0, 0b001010); 2802 2803 #undef INSN 2804 2805 #define INSN(NAME, opc) \ 2806 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \ 2807 starti; \ 2808 assert(T == T2D, "arrangement must be T2D"); \ 2809 f(0b11001110011, 31, 21), rf(Vm, 16), f(opc, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2810 } 2811 2812 INSN(sha512h, 0b100000); 2813 INSN(sha512h2, 0b100001); 2814 INSN(sha512su1, 0b100010); 2815 2816 #undef INSN 2817 2818 #define INSN(NAME, opc) \ 2819 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn) { \ 2820 starti; \ 2821 assert(T == T2D, "arrangement must be T2D"); \ 2822 f(opc, 31, 10), rf(Vn, 5), rf(Vd, 0); \ 2823 } 2824 2825 INSN(sha512su0, 0b1100111011000000100000); 2826 2827 #undef INSN 2828 2829 #define INSN(NAME, opc) \ 2830 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm, FloatRegister Va) { \ 2831 starti; \ 2832 assert(T == T16B, "arrangement must be T16B"); \ 2833 f(0b11001110, 31, 24), f(opc, 23, 21), rf(Vm, 16), f(0b0, 15, 15), rf(Va, 10), rf(Vn, 5), rf(Vd, 0); \ 2834 } 2835 2836 INSN(eor3, 0b000); 2837 INSN(bcax, 0b001); 2838 2839 #undef INSN 2840 2841 #define INSN(NAME, opc) \ 2842 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm, unsigned imm) { \ 2843 starti; \ 2844 assert(T == T2D, "arrangement must be T2D"); \ 2845 f(0b11001110, 31, 24), f(opc, 23, 21), rf(Vm, 16), f(imm, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2846 } 2847 2848 INSN(xar, 0b100); 2849 2850 #undef INSN 2851 2852 #define INSN(NAME, opc) \ 2853 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \ 2854 starti; \ 2855 assert(T == T2D, "arrangement must be T2D"); \ 2856 f(0b11001110, 31, 24), f(opc, 23, 21), rf(Vm, 16), f(0b100011, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2857 } 2858 2859 INSN(rax1, 0b011); 2860 2861 #undef INSN 2862 2863 #define INSN(NAME, opc) \ 2864 void NAME(FloatRegister Vd, FloatRegister Vn) { \ 2865 starti; \ 2866 f(opc, 31, 10), rf(Vn, 5), rf(Vd, 0); \ 2867 } 2868 2869 INSN(aese, 0b0100111000101000010010); 2870 INSN(aesd, 0b0100111000101000010110); 2871 INSN(aesmc, 0b0100111000101000011010); 2872 INSN(aesimc, 0b0100111000101000011110); 2873 2874 #undef INSN 2875 2876 #define INSN(NAME, op1, op2) \ 2877 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm, int index = 0) { \ 2878 starti; \ 2879 assert(T == T2S || T == T4S || T == T2D, "invalid arrangement"); \ 2880 assert(index >= 0 && ((T == T2D && index <= 1) || (T != T2D && index <= 3)), "invalid index"); \ 2881 f(0, 31), f((int)T & 1, 30), f(op1, 29); f(0b011111, 28, 23); \ 2882 f(T == T2D ? 1 : 0, 22), f(T == T2D ? 0 : index & 1, 21), rf(Vm, 16); \ 2883 f(op2, 15, 12), f(T == T2D ? index : (index >> 1), 11), f(0, 10); \ 2884 rf(Vn, 5), rf(Vd, 0); \ 2885 } 2886 2887 // FMLA/FMLS - Vector - Scalar 2888 INSN(fmlavs, 0, 0b0001); 2889 INSN(fmlsvs, 0, 0b0101); 2890 // FMULX - Vector - Scalar 2891 INSN(fmulxvs, 1, 0b1001); 2892 2893 #undef INSN 2894 2895 #define INSN(NAME, op1, op2) \ 2896 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm, int index) { \ 2897 starti; \ 2898 assert(T == T4H || T == T8H || T == T2S || T == T4S, "invalid arrangement"); \ 2899 assert(index >= 0 && \ 2900 ((T == T2S && index <= 1) || (T != T2S && index <= 3) || (T == T8H && index <= 7)), \ 2901 "invalid index"); \ 2902 assert((T != T4H && T != T8H) || Vm->encoding() < 16, "invalid source SIMD&FP register"); \ 2903 f(0, 31), f((int)T & 1, 30), f(op1, 29), f(0b01111, 28, 24); \ 2904 if (T == T4H || T == T8H) { \ 2905 f(0b01, 23, 22), f(index & 0b11, 21, 20), lrf(Vm, 16), f(index >> 2 & 1, 11); \ 2906 } else { \ 2907 f(0b10, 23, 22), f(index & 1, 21), rf(Vm, 16), f(index >> 1, 11); \ 2908 } \ 2909 f(op2, 15, 12), f(0, 10), rf(Vn, 5), rf(Vd, 0); \ 2910 } 2911 2912 // MUL - Vector - Scalar 2913 INSN(mulvs, 0, 0b1000); 2914 2915 #undef INSN 2916 2917 // Floating-point Reciprocal Estimate 2918 void frecpe(FloatRegister Vd, FloatRegister Vn, SIMD_RegVariant type) { 2919 assert(type == D || type == S, "Wrong type for frecpe"); 2920 starti; 2921 f(0b010111101, 31, 23); 2922 f(type == D ? 1 : 0, 22); 2923 f(0b100001110110, 21, 10); 2924 rf(Vn, 5), rf(Vd, 0); 2925 } 2926 2927 // (long) {a, b} -> (a + b) 2928 void addpd(FloatRegister Vd, FloatRegister Vn) { 2929 starti; 2930 f(0b0101111011110001101110, 31, 10); 2931 rf(Vn, 5), rf(Vd, 0); 2932 } 2933 2934 // Floating-point AdvSIMD scalar pairwise 2935 #define INSN(NAME, op1, op2) \ 2936 void NAME(FloatRegister Vd, FloatRegister Vn, SIMD_RegVariant type) { \ 2937 starti; \ 2938 assert(type == D || type == S, "Wrong type for faddp/fmaxp/fminp"); \ 2939 f(0b0111111, 31, 25), f(op1, 24, 23), \ 2940 f(type == S ? 0 : 1, 22), f(0b11000, 21, 17), f(op2, 16, 10), rf(Vn, 5), rf(Vd, 0); \ 2941 } 2942 2943 INSN(faddp, 0b00, 0b0110110); 2944 INSN(fmaxp, 0b00, 0b0111110); 2945 INSN(fminp, 0b01, 0b0111110); 2946 2947 #undef INSN 2948 2949 void ins(FloatRegister Vd, SIMD_RegVariant T, FloatRegister Vn, int didx, int sidx) { 2950 starti; 2951 assert(T != Q, "invalid register variant"); 2952 f(0b01101110000, 31, 21), f(((didx<<1)|1)<<(int)T, 20, 16), f(0, 15); 2953 f(sidx<<(int)T, 14, 11), f(1, 10), rf(Vn, 5), rf(Vd, 0); 2954 } 2955 2956 #define INSN(NAME, cond, op1, op2) \ 2957 void NAME(Register Rd, FloatRegister Vn, SIMD_RegVariant T, int idx) { \ 2958 starti; \ 2959 assert(cond, "invalid register variant"); \ 2960 f(0, 31), f(op1, 30), f(0b001110000, 29, 21); \ 2961 f(((idx << 1) | 1) << (int)T, 20, 16), f(op2, 15, 10); \ 2962 rf(Vn, 5), rf(Rd, 0); \ 2963 } 2964 2965 INSN(umov, (T != Q), (T == D ? 1 : 0), 0b001111); 2966 INSN(smov, (T < D), 1, 0b001011); 2967 2968 #undef INSN 2969 2970 #define INSN(NAME, opc, opc2, isSHR) \ 2971 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, int shift){ \ 2972 starti; \ 2973 /* The encodings for the immh:immb fields (bits 22:16) in *SHR are \ 2974 * 0001 xxx 8B/16B, shift = 16 - UInt(immh:immb) \ 2975 * 001x xxx 4H/8H, shift = 32 - UInt(immh:immb) \ 2976 * 01xx xxx 2S/4S, shift = 64 - UInt(immh:immb) \ 2977 * 1xxx xxx 1D/2D, shift = 128 - UInt(immh:immb) \ 2978 * (1D is RESERVED) \ 2979 * for SHL shift is calculated as: \ 2980 * 0001 xxx 8B/16B, shift = UInt(immh:immb) - 8 \ 2981 * 001x xxx 4H/8H, shift = UInt(immh:immb) - 16 \ 2982 * 01xx xxx 2S/4S, shift = UInt(immh:immb) - 32 \ 2983 * 1xxx xxx 1D/2D, shift = UInt(immh:immb) - 64 \ 2984 * (1D is RESERVED) \ 2985 */ \ 2986 guarantee(!isSHR || (isSHR && (shift != 0)), "impossible encoding");\ 2987 assert((1 << ((T>>1)+3)) > shift, "Invalid Shift value"); \ 2988 int cVal = (1 << (((T >> 1) + 3) + (isSHR ? 1 : 0))); \ 2989 int encodedShift = isSHR ? cVal - shift : cVal + shift; \ 2990 f(0, 31), f(T & 1, 30), f(opc, 29), f(0b011110, 28, 23), \ 2991 f(encodedShift, 22, 16); f(opc2, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 2992 } 2993 2994 INSN(shl, 0, 0b010101, /* isSHR = */ false); 2995 INSN(sshr, 0, 0b000001, /* isSHR = */ true); 2996 INSN(ushr, 1, 0b000001, /* isSHR = */ true); 2997 INSN(usra, 1, 0b000101, /* isSHR = */ true); 2998 INSN(ssra, 0, 0b000101, /* isSHR = */ true); 2999 INSN(sli, 1, 0b010101, /* isSHR = */ false); 3000 3001 #undef INSN 3002 3003 #define INSN(NAME, opc, opc2, isSHR) \ 3004 void NAME(FloatRegister Vd, FloatRegister Vn, int shift){ \ 3005 starti; \ 3006 int encodedShift = isSHR ? 128 - shift : 64 + shift; \ 3007 f(0b01, 31, 30), f(opc, 29), f(0b111110, 28, 23), \ 3008 f(encodedShift, 22, 16); f(opc2, 15, 10), rf(Vn, 5), rf(Vd, 0); \ 3009 } 3010 3011 INSN(shld, 0, 0b010101, /* isSHR = */ false); 3012 INSN(sshrd, 0, 0b000001, /* isSHR = */ true); 3013 INSN(ushrd, 1, 0b000001, /* isSHR = */ true); 3014 3015 #undef INSN 3016 3017 protected: 3018 void _xshll(bool is_unsigned, FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb, int shift) { 3019 starti; 3020 /* The encodings for the immh:immb fields (bits 22:16) are 3021 * 0001 xxx 8H, 8B/16B shift = xxx 3022 * 001x xxx 4S, 4H/8H shift = xxxx 3023 * 01xx xxx 2D, 2S/4S shift = xxxxx 3024 * 1xxx xxx RESERVED 3025 */ 3026 assert((Tb >> 1) + 1 == (Ta >> 1), "Incompatible arrangement"); 3027 assert((1 << ((Tb>>1)+3)) > shift, "Invalid shift value"); 3028 f(0, 31), f(Tb & 1, 30), f(is_unsigned ? 1 : 0, 29), f(0b011110, 28, 23); 3029 f((1 << ((Tb>>1)+3))|shift, 22, 16); 3030 f(0b101001, 15, 10), rf(Vn, 5), rf(Vd, 0); 3031 } 3032 3033 public: 3034 void ushll(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb, int shift) { 3035 assert(Tb == T8B || Tb == T4H || Tb == T2S, "invalid arrangement"); 3036 _xshll(/* is_unsigned */ true, Vd, Ta, Vn, Tb, shift); 3037 } 3038 3039 void ushll2(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb, int shift) { 3040 assert(Tb == T16B || Tb == T8H || Tb == T4S, "invalid arrangement"); 3041 _xshll(/* is_unsigned */ true, Vd, Ta, Vn, Tb, shift); 3042 } 3043 3044 void uxtl(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb) { 3045 ushll(Vd, Ta, Vn, Tb, 0); 3046 } 3047 3048 void sshll(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb, int shift) { 3049 assert(Tb == T8B || Tb == T4H || Tb == T2S, "invalid arrangement"); 3050 _xshll(/* is_unsigned */ false, Vd, Ta, Vn, Tb, shift); 3051 } 3052 3053 void sshll2(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb, int shift) { 3054 assert(Tb == T16B || Tb == T8H || Tb == T4S, "invalid arrangement"); 3055 _xshll(/* is_unsigned */ false, Vd, Ta, Vn, Tb, shift); 3056 } 3057 3058 void sxtl(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb) { 3059 sshll(Vd, Ta, Vn, Tb, 0); 3060 } 3061 3062 // Move from general purpose register 3063 // mov Vd.T[index], Rn 3064 void mov(FloatRegister Vd, SIMD_RegVariant T, int index, Register Xn) { 3065 guarantee(T != Q, "invalid register variant"); 3066 starti; 3067 f(0b01001110000, 31, 21), f(((1 << T) | (index << (T + 1))), 20, 16); 3068 f(0b000111, 15, 10), zrf(Xn, 5), rf(Vd, 0); 3069 } 3070 3071 // Move to general purpose register 3072 // mov Rd, Vn.T[index] 3073 void mov(Register Xd, FloatRegister Vn, SIMD_RegVariant T, int index) { 3074 guarantee(T == S || T == D, "invalid register variant"); 3075 umov(Xd, Vn, T, index); 3076 } 3077 3078 protected: 3079 void _xaddwv(bool is_unsigned, FloatRegister Vd, FloatRegister Vn, SIMD_Arrangement Ta, 3080 FloatRegister Vm, SIMD_Arrangement Tb) { 3081 starti; 3082 assert((Tb >> 1) + 1 == (Ta >> 1), "Incompatible arrangement"); 3083 f(0, 31), f((int)Tb & 1, 30), f(is_unsigned ? 1 : 0, 29), f(0b01110, 28, 24); 3084 f((int)(Ta >> 1) - 1, 23, 22), f(1, 21), rf(Vm, 16), f(0b000100, 15, 10), rf(Vn, 5), rf(Vd, 0); 3085 } 3086 3087 public: 3088 #define INSN(NAME, assertion, is_unsigned) \ 3089 void NAME(FloatRegister Vd, FloatRegister Vn, SIMD_Arrangement Ta, FloatRegister Vm, \ 3090 SIMD_Arrangement Tb) { \ 3091 assert((assertion), "invalid arrangement"); \ 3092 _xaddwv(is_unsigned, Vd, Vn, Ta, Vm, Tb); \ 3093 } 3094 3095 public: 3096 3097 INSN(uaddwv, Tb == T8B || Tb == T4H || Tb == T2S, /*is_unsigned*/true) 3098 INSN(uaddwv2, Tb == T16B || Tb == T8H || Tb == T4S, /*is_unsigned*/true) 3099 INSN(saddwv, Tb == T8B || Tb == T4H || Tb == T2S, /*is_unsigned*/false) 3100 INSN(saddwv2, Tb == T16B || Tb == T8H || Tb == T4S, /*is_unsigned*/false) 3101 3102 #undef INSN 3103 3104 3105 private: 3106 void _pmull(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, FloatRegister Vm, SIMD_Arrangement Tb) { 3107 starti; 3108 assert((Ta == T1Q && (Tb == T1D || Tb == T2D)) || 3109 (Ta == T8H && (Tb == T8B || Tb == T16B)), "Invalid Size specifier"); 3110 int size = (Ta == T1Q) ? 0b11 : 0b00; 3111 f(0, 31), f(Tb & 1, 30), f(0b001110, 29, 24), f(size, 23, 22); 3112 f(1, 21), rf(Vm, 16), f(0b111000, 15, 10), rf(Vn, 5), rf(Vd, 0); 3113 } 3114 3115 public: 3116 void pmull(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, FloatRegister Vm, SIMD_Arrangement Tb) { 3117 assert(Tb == T1D || Tb == T8B, "pmull assumes T1D or T8B as the second size specifier"); 3118 _pmull(Vd, Ta, Vn, Vm, Tb); 3119 } 3120 3121 void pmull2(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, FloatRegister Vm, SIMD_Arrangement Tb) { 3122 assert(Tb == T2D || Tb == T16B, "pmull2 assumes T2D or T16B as the second size specifier"); 3123 _pmull(Vd, Ta, Vn, Vm, Tb); 3124 } 3125 3126 void uqxtn(FloatRegister Vd, SIMD_Arrangement Tb, FloatRegister Vn, SIMD_Arrangement Ta) { 3127 starti; 3128 int size_b = (int)Tb >> 1; 3129 int size_a = (int)Ta >> 1; 3130 assert(size_b < 3 && size_b == size_a - 1, "Invalid size specifier"); 3131 f(0, 31), f(Tb & 1, 30), f(0b101110, 29, 24), f(size_b, 23, 22); 3132 f(0b100001010010, 21, 10), rf(Vn, 5), rf(Vd, 0); 3133 } 3134 3135 void xtn(FloatRegister Vd, SIMD_Arrangement Tb, FloatRegister Vn, SIMD_Arrangement Ta) { 3136 starti; 3137 int size_b = (int)Tb >> 1; 3138 int size_a = (int)Ta >> 1; 3139 assert(size_b < 3 && size_b == size_a - 1, "Invalid size specifier"); 3140 f(0, 31), f(Tb & 1, 30), f(0b001110, 29, 24), f(size_b, 23, 22); 3141 f(0b100001001010, 21, 10), rf(Vn, 5), rf(Vd, 0); 3142 } 3143 3144 void dup(FloatRegister Vd, SIMD_Arrangement T, Register Xs) 3145 { 3146 starti; 3147 assert(T != T1D, "reserved encoding"); 3148 f(0,31), f((int)T & 1, 30), f(0b001110000, 29, 21); 3149 f((1 << (T >> 1)), 20, 16), f(0b000011, 15, 10), zrf(Xs, 5), rf(Vd, 0); 3150 } 3151 3152 void dup(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, int index = 0) 3153 { 3154 starti; 3155 assert(T != T1D, "reserved encoding"); 3156 f(0, 31), f((int)T & 1, 30), f(0b001110000, 29, 21); 3157 f(((1 << (T >> 1)) | (index << ((T >> 1) + 1))), 20, 16); 3158 f(0b000001, 15, 10), rf(Vn, 5), rf(Vd, 0); 3159 } 3160 3161 // Advanced SIMD scalar copy 3162 void dup(FloatRegister Vd, SIMD_RegVariant T, FloatRegister Vn, int index = 0) 3163 { 3164 starti; 3165 assert(T != Q, "invalid size"); 3166 f(0b01011110000, 31, 21); 3167 f((1 << T) | (index << (T + 1)), 20, 16); 3168 f(0b000001, 15, 10), rf(Vn, 5), rf(Vd, 0); 3169 } 3170 3171 // AdvSIMD ZIP/UZP/TRN 3172 #define INSN(NAME, opcode) \ 3173 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \ 3174 guarantee(T != T1D && T != T1Q, "invalid arrangement"); \ 3175 starti; \ 3176 f(0, 31), f(0b001110, 29, 24), f(0, 21), f(0, 15); \ 3177 f(opcode, 14, 12), f(0b10, 11, 10); \ 3178 rf(Vm, 16), rf(Vn, 5), rf(Vd, 0); \ 3179 f(T & 1, 30), f(T >> 1, 23, 22); \ 3180 } 3181 3182 INSN(uzp1, 0b001); 3183 INSN(trn1, 0b010); 3184 INSN(zip1, 0b011); 3185 INSN(uzp2, 0b101); 3186 INSN(trn2, 0b110); 3187 INSN(zip2, 0b111); 3188 3189 #undef INSN 3190 3191 // CRC32 instructions 3192 #define INSN(NAME, c, sf, sz) \ 3193 void NAME(Register Rd, Register Rn, Register Rm) { \ 3194 starti; \ 3195 f(sf, 31), f(0b0011010110, 30, 21), f(0b010, 15, 13), f(c, 12); \ 3196 f(sz, 11, 10), rf(Rm, 16), rf(Rn, 5), rf(Rd, 0); \ 3197 } 3198 3199 INSN(crc32b, 0, 0, 0b00); 3200 INSN(crc32h, 0, 0, 0b01); 3201 INSN(crc32w, 0, 0, 0b10); 3202 INSN(crc32x, 0, 1, 0b11); 3203 INSN(crc32cb, 1, 0, 0b00); 3204 INSN(crc32ch, 1, 0, 0b01); 3205 INSN(crc32cw, 1, 0, 0b10); 3206 INSN(crc32cx, 1, 1, 0b11); 3207 3208 #undef INSN 3209 3210 // Table vector lookup 3211 #define INSN(NAME, op) \ 3212 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, unsigned registers, FloatRegister Vm) { \ 3213 starti; \ 3214 assert(T == T8B || T == T16B, "invalid arrangement"); \ 3215 assert(0 < registers && registers <= 4, "invalid number of registers"); \ 3216 f(0, 31), f((int)T & 1, 30), f(0b001110000, 29, 21), rf(Vm, 16), f(0, 15); \ 3217 f(registers - 1, 14, 13), f(op, 12),f(0b00, 11, 10), rf(Vn, 5), rf(Vd, 0); \ 3218 } 3219 3220 INSN(tbl, 0); 3221 INSN(tbx, 1); 3222 3223 #undef INSN 3224 3225 // AdvSIMD two-reg misc 3226 // In this instruction group, the 2 bits in the size field ([23:22]) may be 3227 // fixed or determined by the "SIMD_Arrangement T", or both. The additional 3228 // parameter "tmask" is a 2-bit mask used to indicate which bits in the size 3229 // field are determined by the SIMD_Arrangement. The bit of "tmask" should be 3230 // set to 1 if corresponding bit marked as "x" in the ArmARM. 3231 #define INSN(NAME, U, size, tmask, opcode) \ 3232 void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn) { \ 3233 starti; \ 3234 assert((ASSERTION), MSG); \ 3235 f(0, 31), f((int)T & 1, 30), f(U, 29), f(0b01110, 28, 24); \ 3236 f(size | ((int)(T >> 1) & tmask), 23, 22), f(0b10000, 21, 17); \ 3237 f(opcode, 16, 12), f(0b10, 11, 10), rf(Vn, 5), rf(Vd, 0); \ 3238 } 3239 3240 #define MSG "invalid arrangement" 3241 3242 #define ASSERTION (T == T2S || T == T4S || T == T2D) 3243 INSN(fsqrt, 1, 0b10, 0b01, 0b11111); 3244 INSN(fabs, 0, 0b10, 0b01, 0b01111); 3245 INSN(fneg, 1, 0b10, 0b01, 0b01111); 3246 INSN(frintn, 0, 0b00, 0b01, 0b11000); 3247 INSN(frintm, 0, 0b00, 0b01, 0b11001); 3248 INSN(frintp, 0, 0b10, 0b01, 0b11000); 3249 INSN(fcvtas, 0, 0b00, 0b01, 0b11100); 3250 INSN(fcvtzs, 0, 0b10, 0b01, 0b11011); 3251 INSN(fcvtms, 0, 0b00, 0b01, 0b11011); 3252 #undef ASSERTION 3253 3254 #define ASSERTION (T == T8B || T == T16B || T == T4H || T == T8H || T == T2S || T == T4S) 3255 INSN(rev64, 0, 0b00, 0b11, 0b00000); 3256 #undef ASSERTION 3257 3258 #define ASSERTION (T == T8B || T == T16B || T == T4H || T == T8H) 3259 INSN(rev32, 1, 0b00, 0b11, 0b00000); 3260 #undef ASSERTION 3261 3262 #define ASSERTION (T == T8B || T == T16B) 3263 INSN(rev16, 0, 0b00, 0b11, 0b00001); 3264 INSN(rbit, 1, 0b01, 0b00, 0b00101); 3265 #undef ASSERTION 3266 3267 #undef MSG 3268 3269 #undef INSN 3270 3271 // AdvSIMD compare with zero (vector) 3272 void cm(Condition cond, FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn) { 3273 starti; 3274 assert(T != T1Q && T != T1D, "invalid arrangement"); 3275 int cond_op; 3276 switch (cond) { 3277 case EQ: cond_op = 0b001; break; 3278 case GE: cond_op = 0b100; break; 3279 case GT: cond_op = 0b000; break; 3280 case LE: cond_op = 0b101; break; 3281 case LT: cond_op = 0b010; break; 3282 default: 3283 ShouldNotReachHere(); 3284 break; 3285 } 3286 3287 f(0, 31), f((int)T & 1, 30), f((cond_op >> 2) & 1, 29); 3288 f(0b01110, 28, 24), f((int)T >> 1, 23, 22), f(0b10000010, 21, 14); 3289 f(cond_op & 0b11, 13, 12), f(0b10, 11, 10), rf(Vn, 5), rf(Vd, 0); 3290 } 3291 3292 // AdvSIMD Floating-point compare with zero (vector) 3293 void fcm(Condition cond, FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn) { 3294 starti; 3295 assert(T == T2S || T == T4S || T == T2D, "invalid arrangement"); 3296 int cond_op; 3297 switch (cond) { 3298 case EQ: cond_op = 0b010; break; 3299 case GT: cond_op = 0b000; break; 3300 case GE: cond_op = 0b001; break; 3301 case LE: cond_op = 0b011; break; 3302 case LT: cond_op = 0b100; break; 3303 default: 3304 ShouldNotReachHere(); 3305 break; 3306 } 3307 3308 f(0, 31), f((int)T & 1, 30), f(cond_op & 1, 29), f(0b011101, 28, 23); 3309 f(((int)(T >> 1) & 1), 22), f(0b10000011, 21, 14); 3310 f((cond_op >> 1) & 0b11, 13, 12), f(0b10, 11, 10), rf(Vn, 5), rf(Vd, 0); 3311 } 3312 3313 void ext(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm, int index) 3314 { 3315 starti; 3316 assert(T == T8B || T == T16B, "invalid arrangement"); 3317 assert((T == T8B && index <= 0b0111) || (T == T16B && index <= 0b1111), "Invalid index value"); 3318 f(0, 31), f((int)T & 1, 30), f(0b101110000, 29, 21); 3319 rf(Vm, 16), f(0, 15), f(index, 14, 11); 3320 f(0, 10), rf(Vn, 5), rf(Vd, 0); 3321 } 3322 3323 // SVE arithmetic - unpredicated 3324 #define INSN(NAME, opcode) \ 3325 void NAME(FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn, FloatRegister Zm) { \ 3326 starti; \ 3327 assert(T != Q, "invalid register variant"); \ 3328 f(0b00000100, 31, 24), f(T, 23, 22), f(1, 21), \ 3329 rf(Zm, 16), f(0, 15, 13), f(opcode, 12, 10), rf(Zn, 5), rf(Zd, 0); \ 3330 } 3331 INSN(sve_add, 0b000); 3332 INSN(sve_sub, 0b001); 3333 INSN(sve_sqadd, 0b100); 3334 INSN(sve_sqsub, 0b110); 3335 INSN(sve_uqadd, 0b101); 3336 INSN(sve_uqsub, 0b111); 3337 #undef INSN 3338 3339 // SVE integer add/subtract immediate (unpredicated) 3340 #define INSN(NAME, op) \ 3341 void NAME(FloatRegister Zd, SIMD_RegVariant T, unsigned imm8) { \ 3342 starti; \ 3343 /* The immediate is an unsigned value in the range 0 to 255, and \ 3344 * for element width of 16 bits or higher it may also be a \ 3345 * positive multiple of 256 in the range 256 to 65280. \ 3346 */ \ 3347 assert(T != Q, "invalid size"); \ 3348 int sh = 0; \ 3349 if (imm8 <= 0xff) { \ 3350 sh = 0; \ 3351 } else if (T != B && imm8 <= 0xff00 && (imm8 & 0xff) == 0) { \ 3352 sh = 1; \ 3353 imm8 = (imm8 >> 8); \ 3354 } else { \ 3355 guarantee(false, "invalid immediate"); \ 3356 } \ 3357 f(0b00100101, 31, 24), f(T, 23, 22), f(0b10000, 21, 17); \ 3358 f(op, 16, 14), f(sh, 13), f(imm8, 12, 5), rf(Zd, 0); \ 3359 } 3360 3361 INSN(sve_add, 0b011); 3362 INSN(sve_sub, 0b111); 3363 #undef INSN 3364 3365 // SVE floating-point arithmetic - unpredicated 3366 #define INSN(NAME, opcode) \ 3367 void NAME(FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn, FloatRegister Zm) { \ 3368 starti; \ 3369 assert(T == S || T == D, "invalid register variant"); \ 3370 f(0b01100101, 31, 24), f(T, 23, 22), f(0, 21), \ 3371 rf(Zm, 16), f(0, 15, 13), f(opcode, 12, 10), rf(Zn, 5), rf(Zd, 0); \ 3372 } 3373 3374 INSN(sve_fadd, 0b000); 3375 INSN(sve_fmul, 0b010); 3376 INSN(sve_fsub, 0b001); 3377 #undef INSN 3378 3379 private: 3380 void sve_predicate_reg_insn(unsigned op24, unsigned op13, 3381 FloatRegister Zd_or_Vd, SIMD_RegVariant T, 3382 PRegister Pg, FloatRegister Zn_or_Vn) { 3383 starti; 3384 f(op24, 31, 24), f(T, 23, 22), f(op13, 21, 13); 3385 pgrf(Pg, 10), rf(Zn_or_Vn, 5), rf(Zd_or_Vd, 0); 3386 } 3387 3388 void sve_shift_imm_encoding(SIMD_RegVariant T, int shift, bool isSHR, 3389 int& tszh, int& tszl_imm) { 3390 /* The encodings for the tszh:tszl:imm3 fields 3391 * for shift right is calculated as: 3392 * 0001 xxx B, shift = 16 - UInt(tszh:tszl:imm3) 3393 * 001x xxx H, shift = 32 - UInt(tszh:tszl:imm3) 3394 * 01xx xxx S, shift = 64 - UInt(tszh:tszl:imm3) 3395 * 1xxx xxx D, shift = 128 - UInt(tszh:tszl:imm3) 3396 * for shift left is calculated as: 3397 * 0001 xxx B, shift = UInt(tszh:tszl:imm3) - 8 3398 * 001x xxx H, shift = UInt(tszh:tszl:imm3) - 16 3399 * 01xx xxx S, shift = UInt(tszh:tszl:imm3) - 32 3400 * 1xxx xxx D, shift = UInt(tszh:tszl:imm3) - 64 3401 */ 3402 assert(T != Q, "Invalid register variant"); 3403 if (isSHR) { 3404 assert(((1 << (T + 3)) >= shift) && (shift > 0) , "Invalid shift value"); 3405 } else { 3406 assert(((1 << (T + 3)) > shift) && (shift >= 0) , "Invalid shift value"); 3407 } 3408 int cVal = (1 << ((T + 3) + (isSHR ? 1 : 0))); 3409 int encodedShift = isSHR ? cVal - shift : cVal + shift; 3410 tszh = encodedShift >> 5; 3411 tszl_imm = encodedShift & 0x1f; 3412 } 3413 3414 public: 3415 3416 // SVE integer arithmetic - predicate 3417 #define INSN(NAME, op1, op2) \ 3418 void NAME(FloatRegister Zdn_or_Zd_or_Vd, SIMD_RegVariant T, PRegister Pg, FloatRegister Znm_or_Vn) { \ 3419 assert(T != Q, "invalid register variant"); \ 3420 sve_predicate_reg_insn(op1, op2, Zdn_or_Zd_or_Vd, T, Pg, Znm_or_Vn); \ 3421 } 3422 3423 INSN(sve_abs, 0b00000100, 0b010110101); // vector abs, unary 3424 INSN(sve_add, 0b00000100, 0b000000000); // vector add 3425 INSN(sve_and, 0b00000100, 0b011010000); // vector and 3426 INSN(sve_andv, 0b00000100, 0b011010001); // bitwise and reduction to scalar 3427 INSN(sve_asr, 0b00000100, 0b010000100); // vector arithmetic shift right 3428 INSN(sve_bic, 0b00000100, 0b011011000); // vector bitwise clear 3429 INSN(sve_clz, 0b00000100, 0b011001101); // vector count leading zero bits 3430 INSN(sve_cnt, 0b00000100, 0b011010101); // count non-zero bits 3431 INSN(sve_cpy, 0b00000101, 0b100000100); // copy scalar to each active vector element 3432 INSN(sve_eor, 0b00000100, 0b011001000); // vector eor 3433 INSN(sve_eorv, 0b00000100, 0b011001001); // bitwise xor reduction to scalar 3434 INSN(sve_lsl, 0b00000100, 0b010011100); // vector logical shift left 3435 INSN(sve_lsr, 0b00000100, 0b010001100); // vector logical shift right 3436 INSN(sve_mul, 0b00000100, 0b010000000); // vector mul 3437 INSN(sve_neg, 0b00000100, 0b010111101); // vector neg, unary 3438 INSN(sve_not, 0b00000100, 0b011110101); // bitwise invert vector, unary 3439 INSN(sve_orr, 0b00000100, 0b011000000); // vector or 3440 INSN(sve_orv, 0b00000100, 0b011000001); // bitwise or reduction to scalar 3441 INSN(sve_smax, 0b00000100, 0b001000000); // signed maximum vectors 3442 INSN(sve_smaxv, 0b00000100, 0b001000001); // signed maximum reduction to scalar 3443 INSN(sve_smin, 0b00000100, 0b001010000); // signed minimum vectors 3444 INSN(sve_sminv, 0b00000100, 0b001010001); // signed minimum reduction to scalar 3445 INSN(sve_sub, 0b00000100, 0b000001000); // vector sub 3446 INSN(sve_uaddv, 0b00000100, 0b000001001); // unsigned add reduction to scalar 3447 INSN(sve_umax, 0b00000100, 0b001001000); // unsigned maximum vectors 3448 INSN(sve_umin, 0b00000100, 0b001011000); // unsigned minimum vectors 3449 #undef INSN 3450 3451 // SVE floating-point arithmetic - predicate 3452 #define INSN(NAME, op1, op2) \ 3453 void NAME(FloatRegister Zd_or_Zdn_or_Vd, SIMD_RegVariant T, PRegister Pg, FloatRegister Zn_or_Zm) { \ 3454 assert(T == S || T == D, "invalid register variant"); \ 3455 sve_predicate_reg_insn(op1, op2, Zd_or_Zdn_or_Vd, T, Pg, Zn_or_Zm); \ 3456 } 3457 3458 INSN(sve_fabd, 0b01100101, 0b001000100); // floating-point absolute difference 3459 INSN(sve_fabs, 0b00000100, 0b011100101); 3460 INSN(sve_fadd, 0b01100101, 0b000000100); 3461 INSN(sve_fadda, 0b01100101, 0b011000001); // add strictly-ordered reduction to scalar Vd 3462 INSN(sve_fdiv, 0b01100101, 0b001101100); 3463 INSN(sve_fmax, 0b01100101, 0b000110100); // floating-point maximum 3464 INSN(sve_fmaxv, 0b01100101, 0b000110001); // floating-point maximum recursive reduction to scalar 3465 INSN(sve_fmin, 0b01100101, 0b000111100); // floating-point minimum 3466 INSN(sve_fminv, 0b01100101, 0b000111001); // floating-point minimum recursive reduction to scalar 3467 INSN(sve_fmul, 0b01100101, 0b000010100); 3468 INSN(sve_fneg, 0b00000100, 0b011101101); 3469 INSN(sve_frintm, 0b01100101, 0b000010101); // floating-point round to integral value, toward minus infinity 3470 INSN(sve_frintn, 0b01100101, 0b000000101); // floating-point round to integral value, nearest with ties to even 3471 INSN(sve_frinta, 0b01100101, 0b000100101); // floating-point round to integral value, nearest with ties to away 3472 INSN(sve_frintp, 0b01100101, 0b000001101); // floating-point round to integral value, toward plus infinity 3473 INSN(sve_fsqrt, 0b01100101, 0b001101101); 3474 INSN(sve_fsub, 0b01100101, 0b000001100); 3475 #undef INSN 3476 3477 // SVE multiple-add/sub - predicated 3478 #define INSN(NAME, op0, op1, op2) \ 3479 void NAME(FloatRegister Zda, SIMD_RegVariant T, PRegister Pg, FloatRegister Zn, FloatRegister Zm) { \ 3480 starti; \ 3481 assert(T != Q, "invalid size"); \ 3482 f(op0, 31, 24), f(T, 23, 22), f(op1, 21), rf(Zm, 16); \ 3483 f(op2, 15, 13), pgrf(Pg, 10), rf(Zn, 5), rf(Zda, 0); \ 3484 } 3485 3486 INSN(sve_fmla, 0b01100101, 1, 0b000); // floating-point fused multiply-add, writing addend: Zda = Zda + Zn * Zm 3487 INSN(sve_fmls, 0b01100101, 1, 0b001); // floating-point fused multiply-subtract: Zda = Zda + -Zn * Zm 3488 INSN(sve_fnmla, 0b01100101, 1, 0b010); // floating-point negated fused multiply-add: Zda = -Zda + -Zn * Zm 3489 INSN(sve_fnmls, 0b01100101, 1, 0b011); // floating-point negated fused multiply-subtract: Zda = -Zda + Zn * Zm 3490 INSN(sve_fmad, 0b01100101, 1, 0b100); // floating-point fused multiply-add, writing multiplicand: Zda = Zm + Zda * Zn 3491 INSN(sve_fmsb, 0b01100101, 1, 0b101); // floating-point fused multiply-subtract, writing multiplicand: Zda = Zm + -Zda * Zn 3492 INSN(sve_fnmad, 0b01100101, 1, 0b110); // floating-point negated fused multiply-add, writing multiplicand: Zda = -Zm + -Zda * Zn 3493 INSN(sve_fnmsb, 0b01100101, 1, 0b111); // floating-point negated fused multiply-subtract, writing multiplicand: Zda = -Zm + Zda * Zn 3494 INSN(sve_mla, 0b00000100, 0, 0b010); // multiply-add, writing addend: Zda = Zda + Zn*Zm 3495 INSN(sve_mls, 0b00000100, 0, 0b011); // multiply-subtract, writing addend: Zda = Zda + -Zn*Zm 3496 #undef INSN 3497 3498 // SVE bitwise logical - unpredicated 3499 #define INSN(NAME, opc) \ 3500 void NAME(FloatRegister Zd, FloatRegister Zn, FloatRegister Zm) { \ 3501 starti; \ 3502 f(0b00000100, 31, 24), f(opc, 23, 22), f(1, 21), \ 3503 rf(Zm, 16), f(0b001100, 15, 10), rf(Zn, 5), rf(Zd, 0); \ 3504 } 3505 INSN(sve_and, 0b00); 3506 INSN(sve_eor, 0b10); 3507 INSN(sve_orr, 0b01); 3508 INSN(sve_bic, 0b11); 3509 #undef INSN 3510 3511 // SVE bitwise logical with immediate (unpredicated) 3512 #define INSN(NAME, opc) \ 3513 void NAME(FloatRegister Zd, SIMD_RegVariant T, uint64_t imm) { \ 3514 starti; \ 3515 unsigned elembits = regVariant_to_elemBits(T); \ 3516 uint32_t val = encode_sve_logical_immediate(elembits, imm); \ 3517 f(0b00000101, 31, 24), f(opc, 23, 22), f(0b0000, 21, 18); \ 3518 f(val, 17, 5), rf(Zd, 0); \ 3519 } 3520 INSN(sve_and, 0b10); 3521 INSN(sve_eor, 0b01); 3522 INSN(sve_orr, 0b00); 3523 #undef INSN 3524 3525 // SVE shift immediate - unpredicated 3526 #define INSN(NAME, opc, isSHR) \ 3527 void NAME(FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn, int shift) { \ 3528 starti; \ 3529 int tszh, tszl_imm; \ 3530 sve_shift_imm_encoding(T, shift, isSHR, tszh, tszl_imm); \ 3531 f(0b00000100, 31, 24); \ 3532 f(tszh, 23, 22), f(1,21), f(tszl_imm, 20, 16); \ 3533 f(0b100, 15, 13), f(opc, 12, 10), rf(Zn, 5), rf(Zd, 0); \ 3534 } 3535 3536 INSN(sve_asr, 0b100, /* isSHR = */ true); 3537 INSN(sve_lsl, 0b111, /* isSHR = */ false); 3538 INSN(sve_lsr, 0b101, /* isSHR = */ true); 3539 #undef INSN 3540 3541 // SVE bitwise shift by immediate (predicated) 3542 #define INSN(NAME, opc, isSHR) \ 3543 void NAME(FloatRegister Zdn, SIMD_RegVariant T, PRegister Pg, int shift) { \ 3544 starti; \ 3545 int tszh, tszl_imm; \ 3546 sve_shift_imm_encoding(T, shift, isSHR, tszh, tszl_imm); \ 3547 f(0b00000100, 31, 24), f(tszh, 23, 22), f(0b00, 21, 20), f(opc, 19, 16); \ 3548 f(0b100, 15, 13), pgrf(Pg, 10), f(tszl_imm, 9, 5), rf(Zdn, 0); \ 3549 } 3550 3551 INSN(sve_asr, 0b0000, /* isSHR = */ true); 3552 INSN(sve_lsl, 0b0011, /* isSHR = */ false); 3553 INSN(sve_lsr, 0b0001, /* isSHR = */ true); 3554 #undef INSN 3555 3556 private: 3557 3558 // Scalar base + immediate index 3559 void sve_ld_st1(FloatRegister Zt, Register Xn, int imm, PRegister Pg, 3560 SIMD_RegVariant T, int op1, int type, int op2) { 3561 starti; 3562 assert_cond(T >= type); 3563 f(op1, 31, 25), f(type, 24, 23), f(T, 22, 21); 3564 f(0, 20), sf(imm, 19, 16), f(op2, 15, 13); 3565 pgrf(Pg, 10), srf(Xn, 5), rf(Zt, 0); 3566 } 3567 3568 // Scalar base + scalar index 3569 void sve_ld_st1(FloatRegister Zt, Register Xn, Register Xm, PRegister Pg, 3570 SIMD_RegVariant T, int op1, int type, int op2) { 3571 starti; 3572 assert_cond(T >= type); 3573 f(op1, 31, 25), f(type, 24, 23), f(T, 22, 21); 3574 rf(Xm, 16), f(op2, 15, 13); 3575 pgrf(Pg, 10), srf(Xn, 5), rf(Zt, 0); 3576 } 3577 3578 void sve_ld_st1(FloatRegister Zt, PRegister Pg, 3579 SIMD_RegVariant T, const Address &a, 3580 int op1, int type, int imm_op2, int scalar_op2) { 3581 switch (a.getMode()) { 3582 case Address::base_plus_offset: 3583 sve_ld_st1(Zt, a.base(), checked_cast<int>(a.offset()), Pg, T, op1, type, imm_op2); 3584 break; 3585 case Address::base_plus_offset_reg: 3586 sve_ld_st1(Zt, a.base(), a.index(), Pg, T, op1, type, scalar_op2); 3587 break; 3588 default: 3589 ShouldNotReachHere(); 3590 } 3591 } 3592 3593 public: 3594 3595 // SVE contiguous load/store 3596 #define INSN(NAME, op1, type, imm_op2, scalar_op2) \ 3597 void NAME(FloatRegister Zt, SIMD_RegVariant T, PRegister Pg, const Address &a) { \ 3598 assert(T != Q, "invalid register variant"); \ 3599 sve_ld_st1(Zt, Pg, T, a, op1, type, imm_op2, scalar_op2); \ 3600 } 3601 3602 INSN(sve_ld1b, 0b1010010, 0b00, 0b101, 0b010); 3603 INSN(sve_st1b, 0b1110010, 0b00, 0b111, 0b010); 3604 INSN(sve_ld1h, 0b1010010, 0b01, 0b101, 0b010); 3605 INSN(sve_st1h, 0b1110010, 0b01, 0b111, 0b010); 3606 INSN(sve_ld1w, 0b1010010, 0b10, 0b101, 0b010); 3607 INSN(sve_st1w, 0b1110010, 0b10, 0b111, 0b010); 3608 INSN(sve_ld1d, 0b1010010, 0b11, 0b101, 0b010); 3609 INSN(sve_st1d, 0b1110010, 0b11, 0b111, 0b010); 3610 #undef INSN 3611 3612 // Gather/scatter load/store (SVE) - scalar plus vector 3613 #define INSN(NAME, op1, type, op2, op3) \ 3614 void NAME(FloatRegister Zt, PRegister Pg, Register Xn, FloatRegister Zm) { \ 3615 starti; \ 3616 f(op1, 31, 25), f(type, 24, 23), f(op2, 22, 21), rf(Zm, 16); \ 3617 f(op3, 15, 13), pgrf(Pg, 10), srf(Xn, 5), rf(Zt, 0); \ 3618 } 3619 // SVE 32-bit gather load words (scalar plus 32-bit scaled offsets) 3620 INSN(sve_ld1w_gather, 0b1000010, 0b10, 0b01, 0b010); 3621 // SVE 64-bit gather load (scalar plus 32-bit unpacked scaled offsets) 3622 INSN(sve_ld1d_gather, 0b1100010, 0b11, 0b01, 0b010); 3623 // SVE 32-bit scatter store (scalar plus 32-bit scaled offsets) 3624 INSN(sve_st1w_scatter, 0b1110010, 0b10, 0b11, 0b100); 3625 // SVE 64-bit scatter store (scalar plus unpacked 32-bit scaled offsets) 3626 INSN(sve_st1d_scatter, 0b1110010, 0b11, 0b01, 0b100); 3627 #undef INSN 3628 3629 // SVE load/store - unpredicated 3630 #define INSN(NAME, op1) \ 3631 void NAME(FloatRegister Zt, const Address &a) { \ 3632 starti; \ 3633 assert(a.index() == noreg, "invalid address variant"); \ 3634 f(op1, 31, 29), f(0b0010110, 28, 22), sf(a.offset() >> 3, 21, 16), \ 3635 f(0b010, 15, 13), f(a.offset() & 0x7, 12, 10), srf(a.base(), 5), rf(Zt, 0); \ 3636 } 3637 3638 INSN(sve_ldr, 0b100); // LDR (vector) 3639 INSN(sve_str, 0b111); // STR (vector) 3640 #undef INSN 3641 3642 // SVE stack frame adjustment 3643 #define INSN(NAME, op) \ 3644 void NAME(Register Xd, Register Xn, int imm6) { \ 3645 starti; \ 3646 f(0b000001000, 31, 23), f(op, 22, 21); \ 3647 srf(Xn, 16), f(0b01010, 15, 11), sf(imm6, 10, 5), srf(Xd, 0); \ 3648 } 3649 3650 INSN(sve_addvl, 0b01); // Add multiple of vector register size to scalar register 3651 INSN(sve_addpl, 0b11); // Add multiple of predicate register size to scalar register 3652 #undef INSN 3653 3654 // SVE inc/dec register by element count 3655 #define INSN(NAME, op) \ 3656 void NAME(Register Xdn, SIMD_RegVariant T, unsigned imm4 = 1, int pattern = 0b11111) { \ 3657 starti; \ 3658 assert(T != Q, "invalid size"); \ 3659 f(0b00000100,31, 24), f(T, 23, 22), f(0b11, 21, 20); \ 3660 f(imm4 - 1, 19, 16), f(0b11100, 15, 11), f(op, 10), f(pattern, 9, 5), rf(Xdn, 0); \ 3661 } 3662 3663 INSN(sve_inc, 0); 3664 INSN(sve_dec, 1); 3665 #undef INSN 3666 3667 // SVE predicate logical operations 3668 #define INSN(NAME, op1, op2, op3) \ 3669 void NAME(PRegister Pd, PRegister Pg, PRegister Pn, PRegister Pm) { \ 3670 starti; \ 3671 f(0b00100101, 31, 24), f(op1, 23, 22), f(0b00, 21, 20); \ 3672 prf(Pm, 16), f(0b01, 15, 14), prf(Pg, 10), f(op2, 9); \ 3673 prf(Pn, 5), f(op3, 4), prf(Pd, 0); \ 3674 } 3675 3676 INSN(sve_and, 0b00, 0b0, 0b0); 3677 INSN(sve_ands, 0b01, 0b0, 0b0); 3678 INSN(sve_eor, 0b00, 0b1, 0b0); 3679 INSN(sve_eors, 0b01, 0b1, 0b0); 3680 INSN(sve_orr, 0b10, 0b0, 0b0); 3681 INSN(sve_orrs, 0b11, 0b0, 0b0); 3682 INSN(sve_bic, 0b00, 0b0, 0b1); 3683 #undef INSN 3684 3685 // SVE increment register by predicate count 3686 void sve_incp(const Register rd, SIMD_RegVariant T, PRegister pg) { 3687 starti; 3688 assert(T != Q, "invalid size"); 3689 f(0b00100101, 31, 24), f(T, 23, 22), f(0b1011001000100, 21, 9), 3690 prf(pg, 5), rf(rd, 0); 3691 } 3692 3693 // SVE broadcast general-purpose register to vector elements (unpredicated) 3694 void sve_dup(FloatRegister Zd, SIMD_RegVariant T, Register Rn) { 3695 starti; 3696 assert(T != Q, "invalid size"); 3697 f(0b00000101, 31, 24), f(T, 23, 22), f(0b100000001110, 21, 10); 3698 srf(Rn, 5), rf(Zd, 0); 3699 } 3700 3701 // SVE broadcast signed immediate to vector elements (unpredicated) 3702 void sve_dup(FloatRegister Zd, SIMD_RegVariant T, int imm8) { 3703 starti; 3704 assert(T != Q, "invalid size"); 3705 int sh = 0; 3706 if (imm8 <= 127 && imm8 >= -128) { 3707 sh = 0; 3708 } else if (T != B && imm8 <= 32512 && imm8 >= -32768 && (imm8 & 0xff) == 0) { 3709 sh = 1; 3710 imm8 = (imm8 >> 8); 3711 } else { 3712 guarantee(false, "invalid immediate"); 3713 } 3714 f(0b00100101, 31, 24), f(T, 23, 22), f(0b11100011, 21, 14); 3715 f(sh, 13), sf(imm8, 12, 5), rf(Zd, 0); 3716 } 3717 3718 // SVE predicate test 3719 void sve_ptest(PRegister Pg, PRegister Pn) { 3720 starti; 3721 f(0b001001010101000011, 31, 14), prf(Pg, 10), f(0, 9), prf(Pn, 5), f(0, 4, 0); 3722 } 3723 3724 // SVE predicate initialize 3725 void sve_ptrue(PRegister pd, SIMD_RegVariant esize, int pattern = 0b11111) { 3726 starti; 3727 f(0b00100101, 31, 24), f(esize, 23, 22), f(0b011000111000, 21, 10); 3728 f(pattern, 9, 5), f(0b0, 4), prf(pd, 0); 3729 } 3730 3731 // SVE predicate zero 3732 void sve_pfalse(PRegister pd) { 3733 starti; 3734 f(0b00100101, 31, 24), f(0b00, 23, 22), f(0b011000111001, 21, 10); 3735 f(0b000000, 9, 4), prf(pd, 0); 3736 } 3737 3738 // SVE load/store predicate register 3739 #define INSN(NAME, op1) \ 3740 void NAME(PRegister Pt, const Address &a) { \ 3741 starti; \ 3742 assert(a.index() == noreg, "invalid address variant"); \ 3743 f(op1, 31, 29), f(0b0010110, 28, 22), sf(a.offset() >> 3, 21, 16), \ 3744 f(0b000, 15, 13), f(a.offset() & 0x7, 12, 10), srf(a.base(), 5), \ 3745 f(0, 4), prf(Pt, 0); \ 3746 } 3747 3748 INSN(sve_ldr, 0b100); // LDR (predicate) 3749 INSN(sve_str, 0b111); // STR (predicate) 3750 #undef INSN 3751 3752 // SVE move predicate register 3753 void sve_mov(PRegister Pd, PRegister Pn) { 3754 starti; 3755 f(0b001001011000, 31, 20), prf(Pn, 16), f(0b01, 15, 14), prf(Pn, 10); 3756 f(0, 9), prf(Pn, 5), f(0, 4), prf(Pd, 0); 3757 } 3758 3759 // SVE copy general-purpose register to vector elements (predicated) 3760 void sve_cpy(FloatRegister Zd, SIMD_RegVariant T, PRegister Pg, Register Rn) { 3761 starti; 3762 assert(T != Q, "invalid size"); 3763 f(0b00000101, 31, 24), f(T, 23, 22), f(0b101000101, 21, 13); 3764 pgrf(Pg, 10), srf(Rn, 5), rf(Zd, 0); 3765 } 3766 3767 private: 3768 void sve_cpy(FloatRegister Zd, SIMD_RegVariant T, PRegister Pg, int imm8, 3769 bool isMerge, bool isFloat) { 3770 starti; 3771 assert(T != Q, "invalid size"); 3772 int sh = 0; 3773 if (imm8 <= 127 && imm8 >= -128) { 3774 sh = 0; 3775 } else if (T != B && imm8 <= 32512 && imm8 >= -32768 && (imm8 & 0xff) == 0) { 3776 sh = 1; 3777 imm8 = (imm8 >> 8); 3778 } else { 3779 guarantee(false, "invalid immediate"); 3780 } 3781 int m = isMerge ? 1 : 0; 3782 f(0b00000101, 31, 24), f(T, 23, 22), f(0b01, 21, 20); 3783 prf(Pg, 16), f(isFloat ? 1 : 0, 15), f(m, 14), f(sh, 13), sf(imm8, 12, 5), rf(Zd, 0); 3784 } 3785 3786 public: 3787 // SVE copy signed integer immediate to vector elements (predicated) 3788 void sve_cpy(FloatRegister Zd, SIMD_RegVariant T, PRegister Pg, int imm8, bool isMerge) { 3789 sve_cpy(Zd, T, Pg, imm8, isMerge, /*isFloat*/false); 3790 } 3791 // SVE copy floating-point immediate to vector elements (predicated) 3792 void sve_cpy(FloatRegister Zd, SIMD_RegVariant T, PRegister Pg, double d) { 3793 sve_cpy(Zd, T, Pg, checked_cast<int8_t>(pack(d)), /*isMerge*/true, /*isFloat*/true); 3794 } 3795 3796 // SVE conditionally select elements from two vectors 3797 void sve_sel(FloatRegister Zd, SIMD_RegVariant T, PRegister Pg, 3798 FloatRegister Zn, FloatRegister Zm) { 3799 starti; 3800 assert(T != Q, "invalid size"); 3801 f(0b00000101, 31, 24), f(T, 23, 22), f(0b1, 21), rf(Zm, 16); 3802 f(0b11, 15, 14), prf(Pg, 10), rf(Zn, 5), rf(Zd, 0); 3803 } 3804 3805 // SVE Permute Vector - Extract 3806 void sve_ext(FloatRegister Zdn, FloatRegister Zm, int imm8) { 3807 starti; 3808 f(0b00000101001, 31, 21), f(imm8 >> 3, 20, 16), f(0b000, 15, 13); 3809 f(imm8 & 0b111, 12, 10), rf(Zm, 5), rf(Zdn, 0); 3810 } 3811 3812 // SVE Integer/Floating-Point Compare - Vectors 3813 #define INSN(NAME, op1, op2, fp) \ 3814 void NAME(Condition cond, PRegister Pd, SIMD_RegVariant T, PRegister Pg, \ 3815 FloatRegister Zn, FloatRegister Zm) { \ 3816 starti; \ 3817 assert(T != Q, "invalid size"); \ 3818 bool is_absolute = op2 == 0b11; \ 3819 if (fp == 1) { \ 3820 assert(T != B, "invalid size"); \ 3821 if (is_absolute) { \ 3822 assert(cond == GT || cond == GE, "invalid condition for fac"); \ 3823 } else { \ 3824 assert(cond != HI && cond != HS, "invalid condition for fcm"); \ 3825 } \ 3826 } \ 3827 int cond_op; \ 3828 switch(cond) { \ 3829 case EQ: cond_op = (op2 << 2) | 0b10; break; \ 3830 case NE: cond_op = (op2 << 2) | 0b11; break; \ 3831 case GE: cond_op = (op2 << 2) | (is_absolute ? 0b01 : 0b00); break; \ 3832 case GT: cond_op = (op2 << 2) | (is_absolute ? 0b11 : 0b01); break; \ 3833 case HI: cond_op = 0b0001; break; \ 3834 case HS: cond_op = 0b0000; break; \ 3835 default: \ 3836 ShouldNotReachHere(); \ 3837 } \ 3838 f(op1, 31, 24), f(T, 23, 22), f(0, 21), rf(Zm, 16), f((cond_op >> 1) & 7, 15, 13); \ 3839 pgrf(Pg, 10), rf(Zn, 5), f(cond_op & 1, 4), prf(Pd, 0); \ 3840 } 3841 3842 INSN(sve_cmp, 0b00100100, 0b10, 0); // Integer compare vectors 3843 INSN(sve_fcm, 0b01100101, 0b01, 1); // Floating-point compare vectors 3844 INSN(sve_fac, 0b01100101, 0b11, 1); // Floating-point absolute compare vectors 3845 #undef INSN 3846 3847 private: 3848 // Convert Assembler::Condition to op encoding - used by sve integer compare encoding 3849 static int assembler_cond_to_sve_op(Condition cond, bool &is_unsigned) { 3850 if (cond == HI || cond == HS || cond == LO || cond == LS) { 3851 is_unsigned = true; 3852 } else { 3853 is_unsigned = false; 3854 } 3855 3856 switch (cond) { 3857 case HI: 3858 case GT: 3859 return 0b0001; 3860 case HS: 3861 case GE: 3862 return 0b0000; 3863 case LO: 3864 case LT: 3865 return 0b0010; 3866 case LS: 3867 case LE: 3868 return 0b0011; 3869 case EQ: 3870 return 0b1000; 3871 case NE: 3872 return 0b1001; 3873 default: 3874 ShouldNotReachHere(); 3875 return -1; 3876 } 3877 } 3878 3879 public: 3880 // SVE Integer Compare - 5 bits signed imm and 7 bits unsigned imm 3881 void sve_cmp(Condition cond, PRegister Pd, SIMD_RegVariant T, 3882 PRegister Pg, FloatRegister Zn, int imm) { 3883 starti; 3884 assert(T != Q, "invalid size"); 3885 bool is_unsigned = false; 3886 int cond_op = assembler_cond_to_sve_op(cond, is_unsigned); 3887 f(is_unsigned ? 0b00100100 : 0b00100101, 31, 24), f(T, 23, 22); 3888 f(is_unsigned ? 0b1 : 0b0, 21); 3889 if (is_unsigned) { 3890 f(imm, 20, 14), f((cond_op >> 1) & 0x1, 13); 3891 } else { 3892 sf(imm, 20, 16), f((cond_op >> 1) & 0x7, 15, 13); 3893 } 3894 pgrf(Pg, 10), rf(Zn, 5), f(cond_op & 0x1, 4), prf(Pd, 0); 3895 } 3896 3897 // SVE Floating-point compare vector with zero 3898 void sve_fcm(Condition cond, PRegister Pd, SIMD_RegVariant T, 3899 PRegister Pg, FloatRegister Zn, double d) { 3900 starti; 3901 assert(T != Q, "invalid size"); 3902 guarantee(d == 0.0, "invalid immediate"); 3903 int cond_op; 3904 switch(cond) { 3905 case EQ: cond_op = 0b100; break; 3906 case GT: cond_op = 0b001; break; 3907 case GE: cond_op = 0b000; break; 3908 case LT: cond_op = 0b010; break; 3909 case LE: cond_op = 0b011; break; 3910 case NE: cond_op = 0b110; break; 3911 default: 3912 ShouldNotReachHere(); 3913 } 3914 f(0b01100101, 31, 24), f(T, 23, 22), f(0b0100, 21, 18), 3915 f((cond_op >> 1) & 0x3, 17, 16), f(0b001, 15, 13), 3916 pgrf(Pg, 10), rf(Zn, 5); 3917 f(cond_op & 0x1, 4), prf(Pd, 0); 3918 } 3919 3920 // SVE unpack vector elements 3921 protected: 3922 void _sve_xunpk(bool is_unsigned, bool is_high, FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn) { 3923 starti; 3924 assert(T != B && T != Q, "invalid size"); 3925 f(0b00000101, 31, 24), f(T, 23, 22), f(0b1100, 21, 18); 3926 f(is_unsigned ? 1 : 0, 17), f(is_high ? 1 : 0, 16), 3927 f(0b001110, 15, 10), rf(Zn, 5), rf(Zd, 0); 3928 } 3929 3930 public: 3931 #define INSN(NAME, is_unsigned, is_high) \ 3932 void NAME(FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn) { \ 3933 _sve_xunpk(is_unsigned, is_high, Zd, T, Zn); \ 3934 } 3935 3936 INSN(sve_uunpkhi, true, true ); // Unsigned unpack and extend half of vector - high half 3937 INSN(sve_uunpklo, true, false); // Unsigned unpack and extend half of vector - low half 3938 INSN(sve_sunpkhi, false, true ); // Signed unpack and extend half of vector - high half 3939 INSN(sve_sunpklo, false, false); // Signed unpack and extend half of vector - low half 3940 #undef INSN 3941 3942 // SVE unpack predicate elements 3943 #define INSN(NAME, op) \ 3944 void NAME(PRegister Pd, PRegister Pn) { \ 3945 starti; \ 3946 f(0b000001010011000, 31, 17), f(op, 16), f(0b0100000, 15, 9); \ 3947 prf(Pn, 5), f(0b0, 4), prf(Pd, 0); \ 3948 } 3949 3950 INSN(sve_punpkhi, 0b1); // Unpack and widen high half of predicate 3951 INSN(sve_punpklo, 0b0); // Unpack and widen low half of predicate 3952 #undef INSN 3953 3954 // SVE permute vector elements 3955 #define INSN(NAME, op) \ 3956 void NAME(FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn, FloatRegister Zm) { \ 3957 starti; \ 3958 assert(T != Q, "invalid size"); \ 3959 f(0b00000101, 31, 24), f(T, 23, 22), f(0b1, 21), rf(Zm, 16); \ 3960 f(0b01101, 15, 11), f(op, 10), rf(Zn, 5), rf(Zd, 0); \ 3961 } 3962 3963 INSN(sve_uzp1, 0b0); // Concatenate even elements from two vectors 3964 INSN(sve_uzp2, 0b1); // Concatenate odd elements from two vectors 3965 #undef INSN 3966 3967 // SVE permute predicate elements 3968 #define INSN(NAME, op) \ 3969 void NAME(PRegister Pd, SIMD_RegVariant T, PRegister Pn, PRegister Pm) { \ 3970 starti; \ 3971 assert(T != Q, "invalid size"); \ 3972 f(0b00000101, 31, 24), f(T, 23, 22), f(0b10, 21, 20), prf(Pm, 16); \ 3973 f(0b01001, 15, 11), f(op, 10), f(0b0, 9), prf(Pn, 5), f(0b0, 4), prf(Pd, 0); \ 3974 } 3975 3976 INSN(sve_uzp1, 0b0); // Concatenate even elements from two predicates 3977 INSN(sve_uzp2, 0b1); // Concatenate odd elements from two predicates 3978 #undef INSN 3979 3980 // SVE integer compare scalar count and limit 3981 #define INSN(NAME, sf, op) \ 3982 void NAME(PRegister Pd, SIMD_RegVariant T, Register Rn, Register Rm) { \ 3983 starti; \ 3984 assert(T != Q, "invalid register variant"); \ 3985 f(0b00100101, 31, 24), f(T, 23, 22), f(1, 21), \ 3986 zrf(Rm, 16), f(0, 15, 13), f(sf, 12), f(op >> 1, 11, 10), \ 3987 zrf(Rn, 5), f(op & 1, 4), prf(Pd, 0); \ 3988 } 3989 // While incrementing signed scalar less than scalar 3990 INSN(sve_whileltw, 0b0, 0b010); 3991 INSN(sve_whilelt, 0b1, 0b010); 3992 // While incrementing signed scalar less than or equal to scalar 3993 INSN(sve_whilelew, 0b0, 0b011); 3994 INSN(sve_whilele, 0b1, 0b011); 3995 // While incrementing unsigned scalar lower than scalar 3996 INSN(sve_whilelow, 0b0, 0b110); 3997 INSN(sve_whilelo, 0b1, 0b110); 3998 // While incrementing unsigned scalar lower than or the same as scalar 3999 INSN(sve_whilelsw, 0b0, 0b111); 4000 INSN(sve_whilels, 0b1, 0b111); 4001 #undef INSN 4002 4003 // SVE predicate reverse 4004 void sve_rev(PRegister Pd, SIMD_RegVariant T, PRegister Pn) { 4005 starti; 4006 assert(T != Q, "invalid size"); 4007 f(0b00000101, 31, 24), f(T, 23, 22), f(0b1101000100000, 21, 9); 4008 prf(Pn, 5), f(0, 4), prf(Pd, 0); 4009 } 4010 4011 // SVE partition break condition 4012 #define INSN(NAME, op) \ 4013 void NAME(PRegister Pd, PRegister Pg, PRegister Pn, bool isMerge) { \ 4014 starti; \ 4015 f(0b00100101, 31, 24), f(op, 23, 22), f(0b01000001, 21, 14); \ 4016 prf(Pg, 10), f(0b0, 9), prf(Pn, 5), f(isMerge ? 1 : 0, 4), prf(Pd, 0); \ 4017 } 4018 4019 INSN(sve_brka, 0b00); // Break after first true condition 4020 INSN(sve_brkb, 0b10); // Break before first true condition 4021 #undef INSN 4022 4023 // Element count and increment scalar (SVE) 4024 #define INSN(NAME, TYPE) \ 4025 void NAME(Register Xdn, unsigned imm4 = 1, int pattern = 0b11111) { \ 4026 starti; \ 4027 f(0b00000100, 31, 24), f(TYPE, 23, 22), f(0b10, 21, 20); \ 4028 f(imm4 - 1, 19, 16), f(0b11100, 15, 11), f(0, 10), f(pattern, 9, 5), rf(Xdn, 0); \ 4029 } 4030 4031 INSN(sve_cntb, B); // Set scalar to multiple of 8-bit predicate constraint element count 4032 INSN(sve_cnth, H); // Set scalar to multiple of 16-bit predicate constraint element count 4033 INSN(sve_cntw, S); // Set scalar to multiple of 32-bit predicate constraint element count 4034 INSN(sve_cntd, D); // Set scalar to multiple of 64-bit predicate constraint element count 4035 #undef INSN 4036 4037 // Set scalar to active predicate element count 4038 void sve_cntp(Register Xd, SIMD_RegVariant T, PRegister Pg, PRegister Pn) { 4039 starti; 4040 assert(T != Q, "invalid size"); 4041 f(0b00100101, 31, 24), f(T, 23, 22), f(0b10000010, 21, 14); 4042 prf(Pg, 10), f(0, 9), prf(Pn, 5), rf(Xd, 0); 4043 } 4044 4045 // SVE convert signed integer to floating-point (predicated) 4046 void sve_scvtf(FloatRegister Zd, SIMD_RegVariant T_dst, PRegister Pg, 4047 FloatRegister Zn, SIMD_RegVariant T_src) { 4048 starti; 4049 assert(T_src != B && T_dst != B && T_src != Q && T_dst != Q && 4050 (T_src != H || T_dst == T_src), "invalid register variant"); 4051 int opc = T_dst; 4052 int opc2 = T_src; 4053 // In most cases we can treat T_dst, T_src as opc, opc2, 4054 // except for the following two combinations. 4055 // +-----+------+---+------------------------------------+ 4056 // | opc | opc2 | U | Instruction Details | 4057 // +-----+------+---+------------------------------------+ 4058 // | 11 | 00 | 0 | SCVTF - 32-bit to double-precision | 4059 // | 11 | 10 | 0 | SCVTF - 64-bit to single-precision | 4060 // +-----+------+---+------------------------------------+ 4061 if (T_src == S && T_dst == D) { 4062 opc = 0b11; 4063 opc2 = 0b00; 4064 } else if (T_src == D && T_dst == S) { 4065 opc = 0b11; 4066 opc2 = 0b10; 4067 } 4068 f(0b01100101, 31, 24), f(opc, 23, 22), f(0b010, 21, 19); 4069 f(opc2, 18, 17), f(0b0101, 16, 13); 4070 pgrf(Pg, 10), rf(Zn, 5), rf(Zd, 0); 4071 } 4072 4073 // SVE floating-point convert to signed integer, rounding toward zero (predicated) 4074 void sve_fcvtzs(FloatRegister Zd, SIMD_RegVariant T_dst, PRegister Pg, 4075 FloatRegister Zn, SIMD_RegVariant T_src) { 4076 starti; 4077 assert(T_src != B && T_dst != B && T_src != Q && T_dst != Q && 4078 (T_dst != H || T_src == H), "invalid register variant"); 4079 int opc = T_src; 4080 int opc2 = T_dst; 4081 // In most cases we can treat T_src, T_dst as opc, opc2, 4082 // except for the following two combinations. 4083 // +-----+------+---+-------------------------------------+ 4084 // | opc | opc2 | U | Instruction Details | 4085 // +-----+------+---+-------------------------------------+ 4086 // | 11 | 10 | 0 | FCVTZS - single-precision to 64-bit | 4087 // | 11 | 00 | 0 | FCVTZS - double-precision to 32-bit | 4088 // +-----+------+---+-------------------------------------+ 4089 if (T_src == S && T_dst == D) { 4090 opc = 0b11; 4091 opc2 = 0b10; 4092 } else if (T_src == D && T_dst == S) { 4093 opc = 0b11; 4094 opc2 = 0b00; 4095 } 4096 f(0b01100101, 31, 24), f(opc, 23, 22), f(0b011, 21, 19); 4097 f(opc2, 18, 17), f(0b0101, 16, 13); 4098 pgrf(Pg, 10), rf(Zn, 5), rf(Zd, 0); 4099 } 4100 4101 // SVE floating-point convert precision (predicated) 4102 void sve_fcvt(FloatRegister Zd, SIMD_RegVariant T_dst, PRegister Pg, 4103 FloatRegister Zn, SIMD_RegVariant T_src) { 4104 starti; 4105 assert(T_src != B && T_dst != B && T_src != Q && T_dst != Q && 4106 T_src != T_dst, "invalid register variant"); 4107 // The encodings of fields op1 (bits 17-16) and op2 (bits 23-22) 4108 // depend on T_src and T_dst as given below - 4109 // +-----+------+---------------------------------------------+ 4110 // | op2 | op1 | Instruction Details | 4111 // +-----+------+---------------------------------------------+ 4112 // | 10 | 01 | FCVT - half-precision to single-precision | 4113 // | 11 | 01 | FCVT - half-precision to double-precision | 4114 // | 10 | 00 | FCVT - single-precision to half-precision | 4115 // | 11 | 11 | FCVT - single-precision to double-precision | 4116 // | 11 | 00 | FCVT - double-preciison to half-precision | 4117 // | 11 | 10 | FCVT - double-precision to single-precision | 4118 // +-----+------+---+-----------------------------------------+ 4119 int op1 = 0b00; 4120 int op2 = (T_src == D || T_dst == D) ? 0b11 : 0b10; 4121 if (T_src == H) { 4122 op1 = 0b01; 4123 } else if (T_dst == S) { 4124 op1 = 0b10; 4125 } else if (T_dst == D) { 4126 op1 = 0b11; 4127 } 4128 f(0b01100101, 31, 24), f(op2, 23, 22), f(0b0010, 21, 18); 4129 f(op1, 17, 16), f(0b101, 15, 13); 4130 pgrf(Pg, 10), rf(Zn, 5), rf(Zd, 0); 4131 } 4132 4133 // SVE extract element to general-purpose register 4134 #define INSN(NAME, before) \ 4135 void NAME(Register Rd, SIMD_RegVariant T, PRegister Pg, FloatRegister Zn) { \ 4136 starti; \ 4137 f(0b00000101, 31, 24), f(T, 23, 22), f(0b10000, 21, 17); \ 4138 f(before, 16), f(0b101, 15, 13); \ 4139 pgrf(Pg, 10), rf(Zn, 5), rf(Rd, 0); \ 4140 } 4141 4142 INSN(sve_lasta, 0b0); 4143 INSN(sve_lastb, 0b1); 4144 #undef INSN 4145 4146 // SVE extract element to SIMD&FP scalar register 4147 #define INSN(NAME, before) \ 4148 void NAME(FloatRegister Vd, SIMD_RegVariant T, PRegister Pg, FloatRegister Zn) { \ 4149 starti; \ 4150 f(0b00000101, 31, 24), f(T, 23, 22), f(0b10001, 21, 17); \ 4151 f(before, 16), f(0b100, 15, 13); \ 4152 pgrf(Pg, 10), rf(Zn, 5), rf(Vd, 0); \ 4153 } 4154 4155 INSN(sve_lasta, 0b0); 4156 INSN(sve_lastb, 0b1); 4157 #undef INSN 4158 4159 // SVE reverse within elements 4160 #define INSN(NAME, opc, cond) \ 4161 void NAME(FloatRegister Zd, SIMD_RegVariant T, PRegister Pg, FloatRegister Zn) { \ 4162 starti; \ 4163 assert(cond, "invalid size"); \ 4164 f(0b00000101, 31, 24), f(T, 23, 22), f(0b1001, 21, 18), f(opc, 17, 16); \ 4165 f(0b100, 15, 13), pgrf(Pg, 10), rf(Zn, 5), rf(Zd, 0); \ 4166 } 4167 4168 INSN(sve_revb, 0b00, T == H || T == S || T == D); 4169 INSN(sve_rbit, 0b11, T != Q); 4170 #undef INSN 4171 4172 // SVE Create index starting from general-purpose register and incremented by immediate 4173 void sve_index(FloatRegister Zd, SIMD_RegVariant T, Register Rn, int imm) { 4174 starti; 4175 assert(T != Q, "invalid size"); 4176 f(0b00000100, 31, 24), f(T, 23, 22), f(0b1, 21); 4177 sf(imm, 20, 16), f(0b010001, 15, 10); 4178 rf(Rn, 5), rf(Zd, 0); 4179 } 4180 4181 // SVE create index starting from and incremented by immediate 4182 void sve_index(FloatRegister Zd, SIMD_RegVariant T, int imm1, int imm2) { 4183 starti; 4184 assert(T != Q, "invalid size"); 4185 f(0b00000100, 31, 24), f(T, 23, 22), f(0b1, 21); 4186 sf(imm2, 20, 16), f(0b010000, 15, 10); 4187 sf(imm1, 9, 5), rf(Zd, 0); 4188 } 4189 4190 // SVE programmable table lookup/permute using vector of element indices 4191 void sve_tbl(FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn, FloatRegister Zm) { 4192 starti; 4193 assert(T != Q, "invalid size"); 4194 f(0b00000101, 31, 24), f(T, 23, 22), f(0b1, 21), rf(Zm, 16); 4195 f(0b001100, 15, 10), rf(Zn, 5), rf(Zd, 0); 4196 } 4197 4198 // Shuffle active elements of vector to the right and fill with zero 4199 void sve_compact(FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn, PRegister Pg) { 4200 starti; 4201 assert(T == S || T == D, "invalid size"); 4202 f(0b00000101, 31, 24), f(T, 23, 22), f(0b100001100, 21, 13); 4203 pgrf(Pg, 10), rf(Zn, 5), rf(Zd, 0); 4204 } 4205 4206 // SVE2 Count matching elements in vector 4207 void sve_histcnt(FloatRegister Zd, SIMD_RegVariant T, PRegister Pg, 4208 FloatRegister Zn, FloatRegister Zm) { 4209 starti; 4210 assert(T == S || T == D, "invalid size"); 4211 f(0b01000101, 31, 24), f(T, 23, 22), f(0b1, 21), rf(Zm, 16); 4212 f(0b110, 15, 13), pgrf(Pg, 10), rf(Zn, 5), rf(Zd, 0); 4213 } 4214 4215 // SVE2 bitwise permute 4216 #define INSN(NAME, opc) \ 4217 void NAME(FloatRegister Zd, SIMD_RegVariant T, FloatRegister Zn, FloatRegister Zm) { \ 4218 starti; \ 4219 assert(T != Q, "invalid size"); \ 4220 f(0b01000101, 31, 24), f(T, 23, 22), f(0b0, 21); \ 4221 rf(Zm, 16), f(0b1011, 15, 12), f(opc, 11, 10); \ 4222 rf(Zn, 5), rf(Zd, 0); \ 4223 } 4224 4225 INSN(sve_bext, 0b00); 4226 INSN(sve_bdep, 0b01); 4227 #undef INSN 4228 4229 // SVE2 bitwise ternary operations 4230 #define INSN(NAME, opc) \ 4231 void NAME(FloatRegister Zdn, FloatRegister Zm, FloatRegister Zk) { \ 4232 starti; \ 4233 f(0b00000100, 31, 24), f(opc, 23, 21), rf(Zm, 16); \ 4234 f(0b001110, 15, 10), rf(Zk, 5), rf(Zdn, 0); \ 4235 } 4236 4237 INSN(sve_eor3, 0b001); // Bitwise exclusive OR of three vectors 4238 #undef INSN 4239 4240 // SVE2 saturating operations - predicate 4241 #define INSN(NAME, op1, op2) \ 4242 void NAME(FloatRegister Zdn, SIMD_RegVariant T, PRegister Pg, FloatRegister Znm) { \ 4243 assert(T != Q, "invalid register variant"); \ 4244 sve_predicate_reg_insn(op1, op2, Zdn, T, Pg, Znm); \ 4245 } 4246 4247 INSN(sve_sqadd, 0b01000100, 0b011000100); // signed saturating add 4248 INSN(sve_sqsub, 0b01000100, 0b011010100); // signed saturating sub 4249 INSN(sve_uqadd, 0b01000100, 0b011001100); // unsigned saturating add 4250 INSN(sve_uqsub, 0b01000100, 0b011011100); // unsigned saturating sub 4251 4252 #undef INSN 4253 4254 Assembler(CodeBuffer* code) : AbstractAssembler(code) { 4255 } 4256 4257 // Stack overflow checking 4258 virtual void bang_stack_with_offset(int offset); 4259 4260 static bool operand_valid_for_logical_immediate(bool is32, uint64_t imm); 4261 static bool operand_valid_for_sve_logical_immediate(unsigned elembits, uint64_t imm); 4262 static bool operand_valid_for_add_sub_immediate(int64_t imm); 4263 static bool operand_valid_for_sve_add_sub_immediate(int64_t imm); 4264 static bool operand_valid_for_float_immediate(double imm); 4265 static int operand_valid_for_movi_immediate(uint64_t imm64, SIMD_Arrangement T); 4266 4267 void emit_data64(jlong data, relocInfo::relocType rtype, int format = 0); 4268 void emit_data64(jlong data, RelocationHolder const& rspec, int format = 0); 4269 }; 4270 4271 inline Assembler::Membar_mask_bits operator|(Assembler::Membar_mask_bits a, 4272 Assembler::Membar_mask_bits b) { 4273 return Assembler::Membar_mask_bits(unsigned(a)|unsigned(b)); 4274 } 4275 4276 Instruction_aarch64::~Instruction_aarch64() { 4277 assem->emit_int32(insn); 4278 assert_cond(get_bits() == 0xffffffff); 4279 } 4280 4281 #undef f 4282 #undef sf 4283 #undef rf 4284 #undef srf 4285 #undef zrf 4286 #undef prf 4287 #undef pgrf 4288 #undef fixed 4289 4290 #undef starti 4291 4292 // Invert a condition 4293 inline Assembler::Condition operator~(const Assembler::Condition cond) { 4294 return Assembler::Condition(int(cond) ^ 1); 4295 } 4296 4297 extern "C" void das(uint64_t start, int len); 4298 4299 #endif // CPU_AARCH64_ASSEMBLER_AARCH64_HPP