1 /* 2 * Copyright (c) 1997, 2024, 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 #include "precompiled.hpp" 27 #include "asm/assembler.hpp" 28 #include "asm/assembler.inline.hpp" 29 #include "ci/ciEnv.hpp" 30 #include "code/compiledIC.hpp" 31 #include "compiler/compileTask.hpp" 32 #include "compiler/disassembler.hpp" 33 #include "compiler/oopMap.hpp" 34 #include "gc/shared/barrierSet.hpp" 35 #include "gc/shared/barrierSetAssembler.hpp" 36 #include "gc/shared/cardTableBarrierSet.hpp" 37 #include "gc/shared/cardTable.hpp" 38 #include "gc/shared/collectedHeap.hpp" 39 #include "gc/shared/tlab_globals.hpp" 40 #include "interpreter/bytecodeHistogram.hpp" 41 #include "interpreter/interpreter.hpp" 42 #include "jvm.h" 43 #include "memory/resourceArea.hpp" 44 #include "memory/universe.hpp" 45 #include "nativeInst_aarch64.hpp" 46 #include "oops/accessDecorators.hpp" 47 #include "oops/compressedKlass.inline.hpp" 48 #include "oops/compressedOops.inline.hpp" 49 #include "oops/klass.inline.hpp" 50 #include "runtime/continuation.hpp" 51 #include "runtime/icache.hpp" 52 #include "runtime/interfaceSupport.inline.hpp" 53 #include "runtime/javaThread.hpp" 54 #include "runtime/jniHandles.inline.hpp" 55 #include "runtime/sharedRuntime.hpp" 56 #include "runtime/stubRoutines.hpp" 57 #include "utilities/globalDefinitions.hpp" 58 #include "utilities/powerOfTwo.hpp" 59 #ifdef COMPILER1 60 #include "c1/c1_LIRAssembler.hpp" 61 #endif 62 #ifdef COMPILER2 63 #include "oops/oop.hpp" 64 #include "opto/compile.hpp" 65 #include "opto/node.hpp" 66 #include "opto/output.hpp" 67 #endif 68 69 #include <sys/types.h> 70 71 #ifdef PRODUCT 72 #define BLOCK_COMMENT(str) /* nothing */ 73 #else 74 #define BLOCK_COMMENT(str) block_comment(str) 75 #endif 76 #define STOP(str) stop(str); 77 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 78 79 #ifdef ASSERT 80 extern "C" void disnm(intptr_t p); 81 #endif 82 // Target-dependent relocation processing 83 // 84 // Instruction sequences whose target may need to be retrieved or 85 // patched are distinguished by their leading instruction, sorting 86 // them into three main instruction groups and related subgroups. 87 // 88 // 1) Branch, Exception and System (insn count = 1) 89 // 1a) Unconditional branch (immediate): 90 // b/bl imm19 91 // 1b) Compare & branch (immediate): 92 // cbz/cbnz Rt imm19 93 // 1c) Test & branch (immediate): 94 // tbz/tbnz Rt imm14 95 // 1d) Conditional branch (immediate): 96 // b.cond imm19 97 // 98 // 2) Loads and Stores (insn count = 1) 99 // 2a) Load register literal: 100 // ldr Rt imm19 101 // 102 // 3) Data Processing Immediate (insn count = 2 or 3) 103 // 3a) PC-rel. addressing 104 // adr/adrp Rx imm21; ldr/str Ry Rx #imm12 105 // adr/adrp Rx imm21; add Ry Rx #imm12 106 // adr/adrp Rx imm21; movk Rx #imm16<<32; ldr/str Ry, [Rx, #offset_in_page] 107 // adr/adrp Rx imm21 108 // adr/adrp Rx imm21; movk Rx #imm16<<32 109 // adr/adrp Rx imm21; movk Rx #imm16<<32; add Ry, Rx, #offset_in_page 110 // The latter form can only happen when the target is an 111 // ExternalAddress, and (by definition) ExternalAddresses don't 112 // move. Because of that property, there is never any need to 113 // patch the last of the three instructions. However, 114 // MacroAssembler::target_addr_for_insn takes all three 115 // instructions into account and returns the correct address. 116 // 3b) Move wide (immediate) 117 // movz Rx #imm16; movk Rx #imm16 << 16; movk Rx #imm16 << 32; 118 // 119 // A switch on a subset of the instruction's bits provides an 120 // efficient dispatch to these subcases. 121 // 122 // insn[28:26] -> main group ('x' == don't care) 123 // 00x -> UNALLOCATED 124 // 100 -> Data Processing Immediate 125 // 101 -> Branch, Exception and System 126 // x1x -> Loads and Stores 127 // 128 // insn[30:25] -> subgroup ('_' == group, 'x' == don't care). 129 // n.b. in some cases extra bits need to be checked to verify the 130 // instruction is as expected 131 // 132 // 1) ... xx101x Branch, Exception and System 133 // 1a) 00___x Unconditional branch (immediate) 134 // 1b) 01___0 Compare & branch (immediate) 135 // 1c) 01___1 Test & branch (immediate) 136 // 1d) 10___0 Conditional branch (immediate) 137 // other Should not happen 138 // 139 // 2) ... xxx1x0 Loads and Stores 140 // 2a) xx1__00 Load/Store register (insn[28] == 1 && insn[24] == 0) 141 // 2aa) x01__00 Load register literal (i.e. requires insn[29] == 0) 142 // strictly should be 64 bit non-FP/SIMD i.e. 143 // 0101_000 (i.e. requires insn[31:24] == 01011000) 144 // 145 // 3) ... xx100x Data Processing Immediate 146 // 3a) xx___00 PC-rel. addressing (n.b. requires insn[24] == 0) 147 // 3b) xx___101 Move wide (immediate) (n.b. requires insn[24:23] == 01) 148 // strictly should be 64 bit movz #imm16<<0 149 // 110___10100 (i.e. requires insn[31:21] == 11010010100) 150 // 151 class RelocActions { 152 protected: 153 typedef int (*reloc_insn)(address insn_addr, address &target); 154 155 virtual reloc_insn adrpMem() = 0; 156 virtual reloc_insn adrpAdd() = 0; 157 virtual reloc_insn adrpMovk() = 0; 158 159 const address _insn_addr; 160 const uint32_t _insn; 161 162 static uint32_t insn_at(address insn_addr, int n) { 163 return ((uint32_t*)insn_addr)[n]; 164 } 165 uint32_t insn_at(int n) const { 166 return insn_at(_insn_addr, n); 167 } 168 169 public: 170 171 RelocActions(address insn_addr) : _insn_addr(insn_addr), _insn(insn_at(insn_addr, 0)) {} 172 RelocActions(address insn_addr, uint32_t insn) 173 : _insn_addr(insn_addr), _insn(insn) {} 174 175 virtual int unconditionalBranch(address insn_addr, address &target) = 0; 176 virtual int conditionalBranch(address insn_addr, address &target) = 0; 177 virtual int testAndBranch(address insn_addr, address &target) = 0; 178 virtual int loadStore(address insn_addr, address &target) = 0; 179 virtual int adr(address insn_addr, address &target) = 0; 180 virtual int adrp(address insn_addr, address &target, reloc_insn inner) = 0; 181 virtual int immediate(address insn_addr, address &target) = 0; 182 virtual void verify(address insn_addr, address &target) = 0; 183 184 int ALWAYSINLINE run(address insn_addr, address &target) { 185 int instructions = 1; 186 187 uint32_t dispatch = Instruction_aarch64::extract(_insn, 30, 25); 188 switch(dispatch) { 189 case 0b001010: 190 case 0b001011: { 191 instructions = unconditionalBranch(insn_addr, target); 192 break; 193 } 194 case 0b101010: // Conditional branch (immediate) 195 case 0b011010: { // Compare & branch (immediate) 196 instructions = conditionalBranch(insn_addr, target); 197 break; 198 } 199 case 0b011011: { 200 instructions = testAndBranch(insn_addr, target); 201 break; 202 } 203 case 0b001100: 204 case 0b001110: 205 case 0b011100: 206 case 0b011110: 207 case 0b101100: 208 case 0b101110: 209 case 0b111100: 210 case 0b111110: { 211 // load/store 212 if ((Instruction_aarch64::extract(_insn, 29, 24) & 0b111011) == 0b011000) { 213 // Load register (literal) 214 instructions = loadStore(insn_addr, target); 215 break; 216 } else { 217 // nothing to do 218 assert(target == 0, "did not expect to relocate target for polling page load"); 219 } 220 break; 221 } 222 case 0b001000: 223 case 0b011000: 224 case 0b101000: 225 case 0b111000: { 226 // adr/adrp 227 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be"); 228 int shift = Instruction_aarch64::extract(_insn, 31, 31); 229 if (shift) { 230 uint32_t insn2 = insn_at(1); 231 if (Instruction_aarch64::extract(insn2, 29, 24) == 0b111001 && 232 Instruction_aarch64::extract(_insn, 4, 0) == 233 Instruction_aarch64::extract(insn2, 9, 5)) { 234 instructions = adrp(insn_addr, target, adrpMem()); 235 } else if (Instruction_aarch64::extract(insn2, 31, 22) == 0b1001000100 && 236 Instruction_aarch64::extract(_insn, 4, 0) == 237 Instruction_aarch64::extract(insn2, 4, 0)) { 238 instructions = adrp(insn_addr, target, adrpAdd()); 239 } else if (Instruction_aarch64::extract(insn2, 31, 21) == 0b11110010110 && 240 Instruction_aarch64::extract(_insn, 4, 0) == 241 Instruction_aarch64::extract(insn2, 4, 0)) { 242 instructions = adrp(insn_addr, target, adrpMovk()); 243 } else { 244 ShouldNotReachHere(); 245 } 246 } else { 247 instructions = adr(insn_addr, target); 248 } 249 break; 250 } 251 case 0b001001: 252 case 0b011001: 253 case 0b101001: 254 case 0b111001: { 255 instructions = immediate(insn_addr, target); 256 break; 257 } 258 default: { 259 ShouldNotReachHere(); 260 } 261 } 262 263 verify(insn_addr, target); 264 return instructions * NativeInstruction::instruction_size; 265 } 266 }; 267 268 class Patcher : public RelocActions { 269 virtual reloc_insn adrpMem() { return &Patcher::adrpMem_impl; } 270 virtual reloc_insn adrpAdd() { return &Patcher::adrpAdd_impl; } 271 virtual reloc_insn adrpMovk() { return &Patcher::adrpMovk_impl; } 272 273 public: 274 Patcher(address insn_addr) : RelocActions(insn_addr) {} 275 276 virtual int unconditionalBranch(address insn_addr, address &target) { 277 intptr_t offset = (target - insn_addr) >> 2; 278 Instruction_aarch64::spatch(insn_addr, 25, 0, offset); 279 return 1; 280 } 281 virtual int conditionalBranch(address insn_addr, address &target) { 282 intptr_t offset = (target - insn_addr) >> 2; 283 Instruction_aarch64::spatch(insn_addr, 23, 5, offset); 284 return 1; 285 } 286 virtual int testAndBranch(address insn_addr, address &target) { 287 intptr_t offset = (target - insn_addr) >> 2; 288 Instruction_aarch64::spatch(insn_addr, 18, 5, offset); 289 return 1; 290 } 291 virtual int loadStore(address insn_addr, address &target) { 292 intptr_t offset = (target - insn_addr) >> 2; 293 Instruction_aarch64::spatch(insn_addr, 23, 5, offset); 294 return 1; 295 } 296 virtual int adr(address insn_addr, address &target) { 297 #ifdef ASSERT 298 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be"); 299 #endif 300 // PC-rel. addressing 301 ptrdiff_t offset = target - insn_addr; 302 int offset_lo = offset & 3; 303 offset >>= 2; 304 Instruction_aarch64::spatch(insn_addr, 23, 5, offset); 305 Instruction_aarch64::patch(insn_addr, 30, 29, offset_lo); 306 return 1; 307 } 308 virtual int adrp(address insn_addr, address &target, reloc_insn inner) { 309 int instructions = 1; 310 #ifdef ASSERT 311 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be"); 312 #endif 313 ptrdiff_t offset = target - insn_addr; 314 instructions = 2; 315 precond(inner != nullptr); 316 // Give the inner reloc a chance to modify the target. 317 address adjusted_target = target; 318 instructions = (*inner)(insn_addr, adjusted_target); 319 uintptr_t pc_page = (uintptr_t)insn_addr >> 12; 320 uintptr_t adr_page = (uintptr_t)adjusted_target >> 12; 321 offset = adr_page - pc_page; 322 int offset_lo = offset & 3; 323 offset >>= 2; 324 Instruction_aarch64::spatch(insn_addr, 23, 5, offset); 325 Instruction_aarch64::patch(insn_addr, 30, 29, offset_lo); 326 return instructions; 327 } 328 static int adrpMem_impl(address insn_addr, address &target) { 329 uintptr_t dest = (uintptr_t)target; 330 int offset_lo = dest & 0xfff; 331 uint32_t insn2 = insn_at(insn_addr, 1); 332 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30); 333 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 21, 10, offset_lo >> size); 334 guarantee(((dest >> size) << size) == dest, "misaligned target"); 335 return 2; 336 } 337 static int adrpAdd_impl(address insn_addr, address &target) { 338 uintptr_t dest = (uintptr_t)target; 339 int offset_lo = dest & 0xfff; 340 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 21, 10, offset_lo); 341 return 2; 342 } 343 static int adrpMovk_impl(address insn_addr, address &target) { 344 uintptr_t dest = uintptr_t(target); 345 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 20, 5, (uintptr_t)target >> 32); 346 dest = (dest & 0xffffffffULL) | (uintptr_t(insn_addr) & 0xffff00000000ULL); 347 target = address(dest); 348 return 2; 349 } 350 virtual int immediate(address insn_addr, address &target) { 351 assert(Instruction_aarch64::extract(_insn, 31, 21) == 0b11010010100, "must be"); 352 uint64_t dest = (uint64_t)target; 353 // Move wide constant 354 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch"); 355 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch"); 356 Instruction_aarch64::patch(insn_addr, 20, 5, dest & 0xffff); 357 Instruction_aarch64::patch(insn_addr+4, 20, 5, (dest >>= 16) & 0xffff); 358 Instruction_aarch64::patch(insn_addr+8, 20, 5, (dest >>= 16) & 0xffff); 359 return 3; 360 } 361 virtual void verify(address insn_addr, address &target) { 362 #ifdef ASSERT 363 address address_is = MacroAssembler::target_addr_for_insn(insn_addr); 364 if (!(address_is == target)) { 365 tty->print_cr("%p at %p should be %p", address_is, insn_addr, target); 366 disnm((intptr_t)insn_addr); 367 assert(address_is == target, "should be"); 368 } 369 #endif 370 } 371 }; 372 373 // If insn1 and insn2 use the same register to form an address, either 374 // by an offsetted LDR or a simple ADD, return the offset. If the 375 // second instruction is an LDR, the offset may be scaled. 376 static bool offset_for(uint32_t insn1, uint32_t insn2, ptrdiff_t &byte_offset) { 377 if (Instruction_aarch64::extract(insn2, 29, 24) == 0b111001 && 378 Instruction_aarch64::extract(insn1, 4, 0) == 379 Instruction_aarch64::extract(insn2, 9, 5)) { 380 // Load/store register (unsigned immediate) 381 byte_offset = Instruction_aarch64::extract(insn2, 21, 10); 382 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30); 383 byte_offset <<= size; 384 return true; 385 } else if (Instruction_aarch64::extract(insn2, 31, 22) == 0b1001000100 && 386 Instruction_aarch64::extract(insn1, 4, 0) == 387 Instruction_aarch64::extract(insn2, 4, 0)) { 388 // add (immediate) 389 byte_offset = Instruction_aarch64::extract(insn2, 21, 10); 390 return true; 391 } 392 return false; 393 } 394 395 class AArch64Decoder : public RelocActions { 396 virtual reloc_insn adrpMem() { return &AArch64Decoder::adrpMem_impl; } 397 virtual reloc_insn adrpAdd() { return &AArch64Decoder::adrpAdd_impl; } 398 virtual reloc_insn adrpMovk() { return &AArch64Decoder::adrpMovk_impl; } 399 400 public: 401 AArch64Decoder(address insn_addr, uint32_t insn) : RelocActions(insn_addr, insn) {} 402 403 virtual int loadStore(address insn_addr, address &target) { 404 intptr_t offset = Instruction_aarch64::sextract(_insn, 23, 5); 405 target = insn_addr + (offset << 2); 406 return 1; 407 } 408 virtual int unconditionalBranch(address insn_addr, address &target) { 409 intptr_t offset = Instruction_aarch64::sextract(_insn, 25, 0); 410 target = insn_addr + (offset << 2); 411 return 1; 412 } 413 virtual int conditionalBranch(address insn_addr, address &target) { 414 intptr_t offset = Instruction_aarch64::sextract(_insn, 23, 5); 415 target = address(((uint64_t)insn_addr + (offset << 2))); 416 return 1; 417 } 418 virtual int testAndBranch(address insn_addr, address &target) { 419 intptr_t offset = Instruction_aarch64::sextract(_insn, 18, 5); 420 target = address(((uint64_t)insn_addr + (offset << 2))); 421 return 1; 422 } 423 virtual int adr(address insn_addr, address &target) { 424 // PC-rel. addressing 425 intptr_t offset = Instruction_aarch64::extract(_insn, 30, 29); 426 offset |= Instruction_aarch64::sextract(_insn, 23, 5) << 2; 427 target = address((uint64_t)insn_addr + offset); 428 return 1; 429 } 430 virtual int adrp(address insn_addr, address &target, reloc_insn inner) { 431 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be"); 432 intptr_t offset = Instruction_aarch64::extract(_insn, 30, 29); 433 offset |= Instruction_aarch64::sextract(_insn, 23, 5) << 2; 434 int shift = 12; 435 offset <<= shift; 436 uint64_t target_page = ((uint64_t)insn_addr) + offset; 437 target_page &= ((uint64_t)-1) << shift; 438 uint32_t insn2 = insn_at(1); 439 target = address(target_page); 440 precond(inner != nullptr); 441 (*inner)(insn_addr, target); 442 return 2; 443 } 444 static int adrpMem_impl(address insn_addr, address &target) { 445 uint32_t insn2 = insn_at(insn_addr, 1); 446 // Load/store register (unsigned immediate) 447 ptrdiff_t byte_offset = Instruction_aarch64::extract(insn2, 21, 10); 448 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30); 449 byte_offset <<= size; 450 target += byte_offset; 451 return 2; 452 } 453 static int adrpAdd_impl(address insn_addr, address &target) { 454 uint32_t insn2 = insn_at(insn_addr, 1); 455 // add (immediate) 456 ptrdiff_t byte_offset = Instruction_aarch64::extract(insn2, 21, 10); 457 target += byte_offset; 458 return 2; 459 } 460 static int adrpMovk_impl(address insn_addr, address &target) { 461 uint32_t insn2 = insn_at(insn_addr, 1); 462 uint64_t dest = uint64_t(target); 463 dest = (dest & 0xffff0000ffffffff) | 464 ((uint64_t)Instruction_aarch64::extract(insn2, 20, 5) << 32); 465 target = address(dest); 466 467 // We know the destination 4k page. Maybe we have a third 468 // instruction. 469 uint32_t insn = insn_at(insn_addr, 0); 470 uint32_t insn3 = insn_at(insn_addr, 2); 471 ptrdiff_t byte_offset; 472 if (offset_for(insn, insn3, byte_offset)) { 473 target += byte_offset; 474 return 3; 475 } else { 476 return 2; 477 } 478 } 479 virtual int immediate(address insn_addr, address &target) { 480 uint32_t *insns = (uint32_t *)insn_addr; 481 assert(Instruction_aarch64::extract(_insn, 31, 21) == 0b11010010100, "must be"); 482 // Move wide constant: movz, movk, movk. See movptr(). 483 assert(nativeInstruction_at(insns+1)->is_movk(), "wrong insns in patch"); 484 assert(nativeInstruction_at(insns+2)->is_movk(), "wrong insns in patch"); 485 target = address(uint64_t(Instruction_aarch64::extract(_insn, 20, 5)) 486 + (uint64_t(Instruction_aarch64::extract(insns[1], 20, 5)) << 16) 487 + (uint64_t(Instruction_aarch64::extract(insns[2], 20, 5)) << 32)); 488 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch"); 489 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch"); 490 return 3; 491 } 492 virtual void verify(address insn_addr, address &target) { 493 } 494 }; 495 496 address MacroAssembler::target_addr_for_insn(address insn_addr, uint32_t insn) { 497 AArch64Decoder decoder(insn_addr, insn); 498 address target; 499 decoder.run(insn_addr, target); 500 return target; 501 } 502 503 // Patch any kind of instruction; there may be several instructions. 504 // Return the total length (in bytes) of the instructions. 505 int MacroAssembler::pd_patch_instruction_size(address insn_addr, address target) { 506 Patcher patcher(insn_addr); 507 return patcher.run(insn_addr, target); 508 } 509 510 int MacroAssembler::patch_oop(address insn_addr, address o) { 511 int instructions; 512 unsigned insn = *(unsigned*)insn_addr; 513 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch"); 514 515 // OOPs are either narrow (32 bits) or wide (48 bits). We encode 516 // narrow OOPs by setting the upper 16 bits in the first 517 // instruction. 518 if (Instruction_aarch64::extract(insn, 31, 21) == 0b11010010101) { 519 // Move narrow OOP 520 uint32_t n = CompressedOops::narrow_oop_value(cast_to_oop(o)); 521 Instruction_aarch64::patch(insn_addr, 20, 5, n >> 16); 522 Instruction_aarch64::patch(insn_addr+4, 20, 5, n & 0xffff); 523 instructions = 2; 524 } else { 525 // Move wide OOP 526 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch"); 527 uintptr_t dest = (uintptr_t)o; 528 Instruction_aarch64::patch(insn_addr, 20, 5, dest & 0xffff); 529 Instruction_aarch64::patch(insn_addr+4, 20, 5, (dest >>= 16) & 0xffff); 530 Instruction_aarch64::patch(insn_addr+8, 20, 5, (dest >>= 16) & 0xffff); 531 instructions = 3; 532 } 533 return instructions * NativeInstruction::instruction_size; 534 } 535 536 int MacroAssembler::patch_narrow_klass(address insn_addr, narrowKlass n) { 537 // Metadata pointers are either narrow (32 bits) or wide (48 bits). 538 // We encode narrow ones by setting the upper 16 bits in the first 539 // instruction. 540 NativeInstruction *insn = nativeInstruction_at(insn_addr); 541 assert(Instruction_aarch64::extract(insn->encoding(), 31, 21) == 0b11010010101 && 542 nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch"); 543 544 Instruction_aarch64::patch(insn_addr, 20, 5, n >> 16); 545 Instruction_aarch64::patch(insn_addr+4, 20, 5, n & 0xffff); 546 return 2 * NativeInstruction::instruction_size; 547 } 548 549 address MacroAssembler::target_addr_for_insn_or_null(address insn_addr, unsigned insn) { 550 if (NativeInstruction::is_ldrw_to_zr(address(&insn))) { 551 return nullptr; 552 } 553 return MacroAssembler::target_addr_for_insn(insn_addr, insn); 554 } 555 556 void MacroAssembler::safepoint_poll(Label& slow_path, bool at_return, bool acquire, bool in_nmethod, Register tmp) { 557 if (acquire) { 558 lea(tmp, Address(rthread, JavaThread::polling_word_offset())); 559 ldar(tmp, tmp); 560 } else { 561 ldr(tmp, Address(rthread, JavaThread::polling_word_offset())); 562 } 563 if (at_return) { 564 // Note that when in_nmethod is set, the stack pointer is incremented before the poll. Therefore, 565 // we may safely use the sp instead to perform the stack watermark check. 566 cmp(in_nmethod ? sp : rfp, tmp); 567 br(Assembler::HI, slow_path); 568 } else { 569 tbnz(tmp, log2i_exact(SafepointMechanism::poll_bit()), slow_path); 570 } 571 } 572 573 void MacroAssembler::rt_call(address dest, Register tmp) { 574 CodeBlob *cb = CodeCache::find_blob(dest); 575 if (cb) { 576 far_call(RuntimeAddress(dest)); 577 } else { 578 lea(tmp, RuntimeAddress(dest)); 579 blr(tmp); 580 } 581 } 582 583 void MacroAssembler::push_cont_fastpath(Register java_thread) { 584 if (!Continuations::enabled()) return; 585 Label done; 586 ldr(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset())); 587 cmp(sp, rscratch1); 588 br(Assembler::LS, done); 589 mov(rscratch1, sp); // we can't use sp as the source in str 590 str(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset())); 591 bind(done); 592 } 593 594 void MacroAssembler::pop_cont_fastpath(Register java_thread) { 595 if (!Continuations::enabled()) return; 596 Label done; 597 ldr(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset())); 598 cmp(sp, rscratch1); 599 br(Assembler::LO, done); 600 str(zr, Address(java_thread, JavaThread::cont_fastpath_offset())); 601 bind(done); 602 } 603 604 void MacroAssembler::reset_last_Java_frame(bool clear_fp) { 605 // we must set sp to zero to clear frame 606 str(zr, Address(rthread, JavaThread::last_Java_sp_offset())); 607 608 // must clear fp, so that compiled frames are not confused; it is 609 // possible that we need it only for debugging 610 if (clear_fp) { 611 str(zr, Address(rthread, JavaThread::last_Java_fp_offset())); 612 } 613 614 // Always clear the pc because it could have been set by make_walkable() 615 str(zr, Address(rthread, JavaThread::last_Java_pc_offset())); 616 } 617 618 // Calls to C land 619 // 620 // When entering C land, the rfp, & resp of the last Java frame have to be recorded 621 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp 622 // has to be reset to 0. This is required to allow proper stack traversal. 623 void MacroAssembler::set_last_Java_frame(Register last_java_sp, 624 Register last_java_fp, 625 Register last_java_pc, 626 Register scratch) { 627 628 if (last_java_pc->is_valid()) { 629 str(last_java_pc, Address(rthread, 630 JavaThread::frame_anchor_offset() 631 + JavaFrameAnchor::last_Java_pc_offset())); 632 } 633 634 // determine last_java_sp register 635 if (last_java_sp == sp) { 636 mov(scratch, sp); 637 last_java_sp = scratch; 638 } else if (!last_java_sp->is_valid()) { 639 last_java_sp = esp; 640 } 641 642 str(last_java_sp, Address(rthread, JavaThread::last_Java_sp_offset())); 643 644 // last_java_fp is optional 645 if (last_java_fp->is_valid()) { 646 str(last_java_fp, Address(rthread, JavaThread::last_Java_fp_offset())); 647 } 648 } 649 650 void MacroAssembler::set_last_Java_frame(Register last_java_sp, 651 Register last_java_fp, 652 address last_java_pc, 653 Register scratch) { 654 assert(last_java_pc != nullptr, "must provide a valid PC"); 655 656 adr(scratch, last_java_pc); 657 str(scratch, Address(rthread, 658 JavaThread::frame_anchor_offset() 659 + JavaFrameAnchor::last_Java_pc_offset())); 660 661 set_last_Java_frame(last_java_sp, last_java_fp, noreg, scratch); 662 } 663 664 void MacroAssembler::set_last_Java_frame(Register last_java_sp, 665 Register last_java_fp, 666 Label &L, 667 Register scratch) { 668 if (L.is_bound()) { 669 set_last_Java_frame(last_java_sp, last_java_fp, target(L), scratch); 670 } else { 671 InstructionMark im(this); 672 L.add_patch_at(code(), locator()); 673 set_last_Java_frame(last_java_sp, last_java_fp, pc() /* Patched later */, scratch); 674 } 675 } 676 677 static inline bool target_needs_far_branch(address addr) { 678 // codecache size <= 128M 679 if (!MacroAssembler::far_branches()) { 680 return false; 681 } 682 // codecache size > 240M 683 if (MacroAssembler::codestub_branch_needs_far_jump()) { 684 return true; 685 } 686 // codecache size: 128M..240M 687 return !CodeCache::is_non_nmethod(addr); 688 } 689 690 void MacroAssembler::far_call(Address entry, Register tmp) { 691 assert(ReservedCodeCacheSize < 4*G, "branch out of range"); 692 assert(CodeCache::find_blob(entry.target()) != nullptr, 693 "destination of far call not found in code cache"); 694 assert(entry.rspec().type() == relocInfo::external_word_type 695 || entry.rspec().type() == relocInfo::runtime_call_type 696 || entry.rspec().type() == relocInfo::none, "wrong entry relocInfo type"); 697 if (target_needs_far_branch(entry.target())) { 698 uint64_t offset; 699 // We can use ADRP here because we know that the total size of 700 // the code cache cannot exceed 2Gb (ADRP limit is 4GB). 701 adrp(tmp, entry, offset); 702 add(tmp, tmp, offset); 703 blr(tmp); 704 } else { 705 bl(entry); 706 } 707 } 708 709 int MacroAssembler::far_jump(Address entry, Register tmp) { 710 assert(ReservedCodeCacheSize < 4*G, "branch out of range"); 711 assert(CodeCache::find_blob(entry.target()) != nullptr, 712 "destination of far call not found in code cache"); 713 assert(entry.rspec().type() == relocInfo::external_word_type 714 || entry.rspec().type() == relocInfo::runtime_call_type 715 || entry.rspec().type() == relocInfo::none, "wrong entry relocInfo type"); 716 address start = pc(); 717 if (target_needs_far_branch(entry.target())) { 718 uint64_t offset; 719 // We can use ADRP here because we know that the total size of 720 // the code cache cannot exceed 2Gb (ADRP limit is 4GB). 721 adrp(tmp, entry, offset); 722 add(tmp, tmp, offset); 723 br(tmp); 724 } else { 725 b(entry); 726 } 727 return pc() - start; 728 } 729 730 void MacroAssembler::reserved_stack_check() { 731 // testing if reserved zone needs to be enabled 732 Label no_reserved_zone_enabling; 733 734 ldr(rscratch1, Address(rthread, JavaThread::reserved_stack_activation_offset())); 735 cmp(sp, rscratch1); 736 br(Assembler::LO, no_reserved_zone_enabling); 737 738 enter(); // LR and FP are live. 739 lea(rscratch1, CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone)); 740 mov(c_rarg0, rthread); 741 blr(rscratch1); 742 leave(); 743 744 // We have already removed our own frame. 745 // throw_delayed_StackOverflowError will think that it's been 746 // called by our caller. 747 lea(rscratch1, RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry())); 748 br(rscratch1); 749 should_not_reach_here(); 750 751 bind(no_reserved_zone_enabling); 752 } 753 754 static void pass_arg0(MacroAssembler* masm, Register arg) { 755 if (c_rarg0 != arg ) { 756 masm->mov(c_rarg0, arg); 757 } 758 } 759 760 static void pass_arg1(MacroAssembler* masm, Register arg) { 761 if (c_rarg1 != arg ) { 762 masm->mov(c_rarg1, arg); 763 } 764 } 765 766 static void pass_arg2(MacroAssembler* masm, Register arg) { 767 if (c_rarg2 != arg ) { 768 masm->mov(c_rarg2, arg); 769 } 770 } 771 772 static void pass_arg3(MacroAssembler* masm, Register arg) { 773 if (c_rarg3 != arg ) { 774 masm->mov(c_rarg3, arg); 775 } 776 } 777 778 void MacroAssembler::call_VM_base(Register oop_result, 779 Register java_thread, 780 Register last_java_sp, 781 address entry_point, 782 int number_of_arguments, 783 bool check_exceptions) { 784 // determine java_thread register 785 if (!java_thread->is_valid()) { 786 java_thread = rthread; 787 } 788 789 // determine last_java_sp register 790 if (!last_java_sp->is_valid()) { 791 last_java_sp = esp; 792 } 793 794 // debugging support 795 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); 796 assert(java_thread == rthread, "unexpected register"); 797 #ifdef ASSERT 798 // TraceBytecodes does not use r12 but saves it over the call, so don't verify 799 // if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?"); 800 #endif // ASSERT 801 802 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result"); 803 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp"); 804 805 // push java thread (becomes first argument of C function) 806 807 mov(c_rarg0, java_thread); 808 809 // set last Java frame before call 810 assert(last_java_sp != rfp, "can't use rfp"); 811 812 Label l; 813 set_last_Java_frame(last_java_sp, rfp, l, rscratch1); 814 815 // do the call, remove parameters 816 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments, &l); 817 818 // lr could be poisoned with PAC signature during throw_pending_exception 819 // if it was tail-call optimized by compiler, since lr is not callee-saved 820 // reload it with proper value 821 adr(lr, l); 822 823 // reset last Java frame 824 // Only interpreter should have to clear fp 825 reset_last_Java_frame(true); 826 827 // C++ interp handles this in the interpreter 828 check_and_handle_popframe(java_thread); 829 check_and_handle_earlyret(java_thread); 830 831 if (check_exceptions) { 832 // check for pending exceptions (java_thread is set upon return) 833 ldr(rscratch1, Address(java_thread, in_bytes(Thread::pending_exception_offset()))); 834 Label ok; 835 cbz(rscratch1, ok); 836 lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry())); 837 br(rscratch1); 838 bind(ok); 839 } 840 841 // get oop result if there is one and reset the value in the thread 842 if (oop_result->is_valid()) { 843 get_vm_result(oop_result, java_thread); 844 } 845 } 846 847 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { 848 call_VM_base(oop_result, noreg, noreg, entry_point, number_of_arguments, check_exceptions); 849 } 850 851 // Check the entry target is always reachable from any branch. 852 static bool is_always_within_branch_range(Address entry) { 853 const address target = entry.target(); 854 855 if (!CodeCache::contains(target)) { 856 // We always use trampolines for callees outside CodeCache. 857 assert(entry.rspec().type() == relocInfo::runtime_call_type, "non-runtime call of an external target"); 858 return false; 859 } 860 861 if (!MacroAssembler::far_branches()) { 862 return true; 863 } 864 865 if (entry.rspec().type() == relocInfo::runtime_call_type) { 866 // Runtime calls are calls of a non-compiled method (stubs, adapters). 867 // Non-compiled methods stay forever in CodeCache. 868 // We check whether the longest possible branch is within the branch range. 869 assert(CodeCache::find_blob(target) != nullptr && 870 !CodeCache::find_blob(target)->is_nmethod(), 871 "runtime call of compiled method"); 872 const address right_longest_branch_start = CodeCache::high_bound() - NativeInstruction::instruction_size; 873 const address left_longest_branch_start = CodeCache::low_bound(); 874 const bool is_reachable = Assembler::reachable_from_branch_at(left_longest_branch_start, target) && 875 Assembler::reachable_from_branch_at(right_longest_branch_start, target); 876 return is_reachable; 877 } 878 879 return false; 880 } 881 882 // Maybe emit a call via a trampoline. If the code cache is small 883 // trampolines won't be emitted. 884 address MacroAssembler::trampoline_call(Address entry) { 885 assert(entry.rspec().type() == relocInfo::runtime_call_type 886 || entry.rspec().type() == relocInfo::opt_virtual_call_type 887 || entry.rspec().type() == relocInfo::static_call_type 888 || entry.rspec().type() == relocInfo::virtual_call_type, "wrong reloc type"); 889 890 address target = entry.target(); 891 892 if (!is_always_within_branch_range(entry)) { 893 if (!in_scratch_emit_size()) { 894 // We don't want to emit a trampoline if C2 is generating dummy 895 // code during its branch shortening phase. 896 if (entry.rspec().type() == relocInfo::runtime_call_type) { 897 assert(CodeBuffer::supports_shared_stubs(), "must support shared stubs"); 898 code()->share_trampoline_for(entry.target(), offset()); 899 } else { 900 address stub = emit_trampoline_stub(offset(), target); 901 if (stub == nullptr) { 902 postcond(pc() == badAddress); 903 return nullptr; // CodeCache is full 904 } 905 } 906 } 907 target = pc(); 908 } 909 910 address call_pc = pc(); 911 relocate(entry.rspec()); 912 bl(target); 913 914 postcond(pc() != badAddress); 915 return call_pc; 916 } 917 918 // Emit a trampoline stub for a call to a target which is too far away. 919 // 920 // code sequences: 921 // 922 // call-site: 923 // branch-and-link to <destination> or <trampoline stub> 924 // 925 // Related trampoline stub for this call site in the stub section: 926 // load the call target from the constant pool 927 // branch (LR still points to the call site above) 928 929 address MacroAssembler::emit_trampoline_stub(int insts_call_instruction_offset, 930 address dest) { 931 // Max stub size: alignment nop, TrampolineStub. 932 address stub = start_a_stub(max_trampoline_stub_size()); 933 if (stub == nullptr) { 934 return nullptr; // CodeBuffer::expand failed 935 } 936 937 // Create a trampoline stub relocation which relates this trampoline stub 938 // with the call instruction at insts_call_instruction_offset in the 939 // instructions code-section. 940 align(wordSize); 941 relocate(trampoline_stub_Relocation::spec(code()->insts()->start() 942 + insts_call_instruction_offset)); 943 const int stub_start_offset = offset(); 944 945 // Now, create the trampoline stub's code: 946 // - load the call 947 // - call 948 Label target; 949 ldr(rscratch1, target); 950 br(rscratch1); 951 bind(target); 952 assert(offset() - stub_start_offset == NativeCallTrampolineStub::data_offset, 953 "should be"); 954 emit_int64((int64_t)dest); 955 956 const address stub_start_addr = addr_at(stub_start_offset); 957 958 assert(is_NativeCallTrampolineStub_at(stub_start_addr), "doesn't look like a trampoline"); 959 960 end_a_stub(); 961 return stub_start_addr; 962 } 963 964 int MacroAssembler::max_trampoline_stub_size() { 965 // Max stub size: alignment nop, TrampolineStub. 966 return NativeInstruction::instruction_size + NativeCallTrampolineStub::instruction_size; 967 } 968 969 void MacroAssembler::emit_static_call_stub() { 970 // CompiledDirectCall::set_to_interpreted knows the 971 // exact layout of this stub. 972 973 isb(); 974 mov_metadata(rmethod, nullptr); 975 976 // Jump to the entry point of the c2i stub. 977 movptr(rscratch1, 0); 978 br(rscratch1); 979 } 980 981 int MacroAssembler::static_call_stub_size() { 982 // isb; movk; movz; movz; movk; movz; movz; br 983 return 8 * NativeInstruction::instruction_size; 984 } 985 986 void MacroAssembler::c2bool(Register x) { 987 // implements x == 0 ? 0 : 1 988 // note: must only look at least-significant byte of x 989 // since C-style booleans are stored in one byte 990 // only! (was bug) 991 tst(x, 0xff); 992 cset(x, Assembler::NE); 993 } 994 995 address MacroAssembler::ic_call(address entry, jint method_index) { 996 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index); 997 // address const_ptr = long_constant((jlong)Universe::non_oop_word()); 998 // uintptr_t offset; 999 // ldr_constant(rscratch2, const_ptr); 1000 movptr(rscratch2, (intptr_t)Universe::non_oop_word()); 1001 return trampoline_call(Address(entry, rh)); 1002 } 1003 1004 int MacroAssembler::ic_check_size() { 1005 if (target_needs_far_branch(CAST_FROM_FN_PTR(address, SharedRuntime::get_ic_miss_stub()))) { 1006 return NativeInstruction::instruction_size * 7; 1007 } else { 1008 return NativeInstruction::instruction_size * 5; 1009 } 1010 } 1011 1012 int MacroAssembler::ic_check(int end_alignment) { 1013 Register receiver = j_rarg0; 1014 Register data = rscratch2; 1015 Register tmp1 = rscratch1; 1016 Register tmp2 = r10; 1017 1018 // The UEP of a code blob ensures that the VEP is padded. However, the padding of the UEP is placed 1019 // before the inline cache check, so we don't have to execute any nop instructions when dispatching 1020 // through the UEP, yet we can ensure that the VEP is aligned appropriately. That's why we align 1021 // before the inline cache check here, and not after 1022 align(end_alignment, offset() + ic_check_size()); 1023 1024 int uep_offset = offset(); 1025 1026 if (UseCompressedClassPointers) { 1027 ldrw(tmp1, Address(receiver, oopDesc::klass_offset_in_bytes())); 1028 ldrw(tmp2, Address(data, CompiledICData::speculated_klass_offset())); 1029 cmpw(tmp1, tmp2); 1030 } else { 1031 ldr(tmp1, Address(receiver, oopDesc::klass_offset_in_bytes())); 1032 ldr(tmp2, Address(data, CompiledICData::speculated_klass_offset())); 1033 cmp(tmp1, tmp2); 1034 } 1035 1036 Label dont; 1037 br(Assembler::EQ, dont); 1038 far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1039 bind(dont); 1040 assert((offset() % end_alignment) == 0, "Misaligned verified entry point"); 1041 1042 return uep_offset; 1043 } 1044 1045 // Implementation of call_VM versions 1046 1047 void MacroAssembler::call_VM(Register oop_result, 1048 address entry_point, 1049 bool check_exceptions) { 1050 call_VM_helper(oop_result, entry_point, 0, check_exceptions); 1051 } 1052 1053 void MacroAssembler::call_VM(Register oop_result, 1054 address entry_point, 1055 Register arg_1, 1056 bool check_exceptions) { 1057 pass_arg1(this, arg_1); 1058 call_VM_helper(oop_result, entry_point, 1, check_exceptions); 1059 } 1060 1061 void MacroAssembler::call_VM(Register oop_result, 1062 address entry_point, 1063 Register arg_1, 1064 Register arg_2, 1065 bool check_exceptions) { 1066 assert_different_registers(arg_1, c_rarg2); 1067 pass_arg2(this, arg_2); 1068 pass_arg1(this, arg_1); 1069 call_VM_helper(oop_result, entry_point, 2, check_exceptions); 1070 } 1071 1072 void MacroAssembler::call_VM(Register oop_result, 1073 address entry_point, 1074 Register arg_1, 1075 Register arg_2, 1076 Register arg_3, 1077 bool check_exceptions) { 1078 assert_different_registers(arg_1, c_rarg2, c_rarg3); 1079 assert_different_registers(arg_2, c_rarg3); 1080 pass_arg3(this, arg_3); 1081 1082 pass_arg2(this, arg_2); 1083 1084 pass_arg1(this, arg_1); 1085 call_VM_helper(oop_result, entry_point, 3, check_exceptions); 1086 } 1087 1088 void MacroAssembler::call_VM(Register oop_result, 1089 Register last_java_sp, 1090 address entry_point, 1091 int number_of_arguments, 1092 bool check_exceptions) { 1093 call_VM_base(oop_result, rthread, last_java_sp, entry_point, number_of_arguments, check_exceptions); 1094 } 1095 1096 void MacroAssembler::call_VM(Register oop_result, 1097 Register last_java_sp, 1098 address entry_point, 1099 Register arg_1, 1100 bool check_exceptions) { 1101 pass_arg1(this, arg_1); 1102 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); 1103 } 1104 1105 void MacroAssembler::call_VM(Register oop_result, 1106 Register last_java_sp, 1107 address entry_point, 1108 Register arg_1, 1109 Register arg_2, 1110 bool check_exceptions) { 1111 1112 assert_different_registers(arg_1, c_rarg2); 1113 pass_arg2(this, arg_2); 1114 pass_arg1(this, arg_1); 1115 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); 1116 } 1117 1118 void MacroAssembler::call_VM(Register oop_result, 1119 Register last_java_sp, 1120 address entry_point, 1121 Register arg_1, 1122 Register arg_2, 1123 Register arg_3, 1124 bool check_exceptions) { 1125 assert_different_registers(arg_1, c_rarg2, c_rarg3); 1126 assert_different_registers(arg_2, c_rarg3); 1127 pass_arg3(this, arg_3); 1128 pass_arg2(this, arg_2); 1129 pass_arg1(this, arg_1); 1130 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); 1131 } 1132 1133 1134 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) { 1135 ldr(oop_result, Address(java_thread, JavaThread::vm_result_offset())); 1136 str(zr, Address(java_thread, JavaThread::vm_result_offset())); 1137 verify_oop_msg(oop_result, "broken oop in call_VM_base"); 1138 } 1139 1140 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) { 1141 ldr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset())); 1142 str(zr, Address(java_thread, JavaThread::vm_result_2_offset())); 1143 } 1144 1145 void MacroAssembler::align(int modulus) { 1146 align(modulus, offset()); 1147 } 1148 1149 // Ensure that the code at target bytes offset from the current offset() is aligned 1150 // according to modulus. 1151 void MacroAssembler::align(int modulus, int target) { 1152 int delta = target - offset(); 1153 while ((offset() + delta) % modulus != 0) nop(); 1154 } 1155 1156 void MacroAssembler::post_call_nop() { 1157 if (!Continuations::enabled()) { 1158 return; 1159 } 1160 InstructionMark im(this); 1161 relocate(post_call_nop_Relocation::spec()); 1162 InlineSkippedInstructionsCounter skipCounter(this); 1163 nop(); 1164 movk(zr, 0); 1165 movk(zr, 0); 1166 } 1167 1168 // these are no-ops overridden by InterpreterMacroAssembler 1169 1170 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { } 1171 1172 void MacroAssembler::check_and_handle_popframe(Register java_thread) { } 1173 1174 // Look up the method for a megamorphic invokeinterface call. 1175 // The target method is determined by <intf_klass, itable_index>. 1176 // The receiver klass is in recv_klass. 1177 // On success, the result will be in method_result, and execution falls through. 1178 // On failure, execution transfers to the given label. 1179 void MacroAssembler::lookup_interface_method(Register recv_klass, 1180 Register intf_klass, 1181 RegisterOrConstant itable_index, 1182 Register method_result, 1183 Register scan_temp, 1184 Label& L_no_such_interface, 1185 bool return_method) { 1186 assert_different_registers(recv_klass, intf_klass, scan_temp); 1187 assert_different_registers(method_result, intf_klass, scan_temp); 1188 assert(recv_klass != method_result || !return_method, 1189 "recv_klass can be destroyed when method isn't needed"); 1190 assert(itable_index.is_constant() || itable_index.as_register() == method_result, 1191 "caller must use same register for non-constant itable index as for method"); 1192 1193 // Compute start of first itableOffsetEntry (which is at the end of the vtable) 1194 int vtable_base = in_bytes(Klass::vtable_start_offset()); 1195 int itentry_off = in_bytes(itableMethodEntry::method_offset()); 1196 int scan_step = itableOffsetEntry::size() * wordSize; 1197 int vte_size = vtableEntry::size_in_bytes(); 1198 assert(vte_size == wordSize, "else adjust times_vte_scale"); 1199 1200 ldrw(scan_temp, Address(recv_klass, Klass::vtable_length_offset())); 1201 1202 // Could store the aligned, prescaled offset in the klass. 1203 // lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base)); 1204 lea(scan_temp, Address(recv_klass, scan_temp, Address::lsl(3))); 1205 add(scan_temp, scan_temp, vtable_base); 1206 1207 if (return_method) { 1208 // Adjust recv_klass by scaled itable_index, so we can free itable_index. 1209 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); 1210 // lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off)); 1211 lea(recv_klass, Address(recv_klass, itable_index, Address::lsl(3))); 1212 if (itentry_off) 1213 add(recv_klass, recv_klass, itentry_off); 1214 } 1215 1216 // for (scan = klass->itable(); scan->interface() != nullptr; scan += scan_step) { 1217 // if (scan->interface() == intf) { 1218 // result = (klass + scan->offset() + itable_index); 1219 // } 1220 // } 1221 Label search, found_method; 1222 1223 ldr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset())); 1224 cmp(intf_klass, method_result); 1225 br(Assembler::EQ, found_method); 1226 bind(search); 1227 // Check that the previous entry is non-null. A null entry means that 1228 // the receiver class doesn't implement the interface, and wasn't the 1229 // same as when the caller was compiled. 1230 cbz(method_result, L_no_such_interface); 1231 if (itableOffsetEntry::interface_offset() != 0) { 1232 add(scan_temp, scan_temp, scan_step); 1233 ldr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset())); 1234 } else { 1235 ldr(method_result, Address(pre(scan_temp, scan_step))); 1236 } 1237 cmp(intf_klass, method_result); 1238 br(Assembler::NE, search); 1239 1240 bind(found_method); 1241 1242 // Got a hit. 1243 if (return_method) { 1244 ldrw(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset())); 1245 ldr(method_result, Address(recv_klass, scan_temp, Address::uxtw(0))); 1246 } 1247 } 1248 1249 // Look up the method for a megamorphic invokeinterface call in a single pass over itable: 1250 // - check recv_klass (actual object class) is a subtype of resolved_klass from CompiledICData 1251 // - find a holder_klass (class that implements the method) vtable offset and get the method from vtable by index 1252 // The target method is determined by <holder_klass, itable_index>. 1253 // The receiver klass is in recv_klass. 1254 // On success, the result will be in method_result, and execution falls through. 1255 // On failure, execution transfers to the given label. 1256 void MacroAssembler::lookup_interface_method_stub(Register recv_klass, 1257 Register holder_klass, 1258 Register resolved_klass, 1259 Register method_result, 1260 Register temp_itbl_klass, 1261 Register scan_temp, 1262 int itable_index, 1263 Label& L_no_such_interface) { 1264 // 'method_result' is only used as output register at the very end of this method. 1265 // Until then we can reuse it as 'holder_offset'. 1266 Register holder_offset = method_result; 1267 assert_different_registers(resolved_klass, recv_klass, holder_klass, temp_itbl_klass, scan_temp, holder_offset); 1268 1269 int vtable_start_offset = in_bytes(Klass::vtable_start_offset()); 1270 int itable_offset_entry_size = itableOffsetEntry::size() * wordSize; 1271 int ioffset = in_bytes(itableOffsetEntry::interface_offset()); 1272 int ooffset = in_bytes(itableOffsetEntry::offset_offset()); 1273 1274 Label L_loop_search_resolved_entry, L_resolved_found, L_holder_found; 1275 1276 ldrw(scan_temp, Address(recv_klass, Klass::vtable_length_offset())); 1277 add(recv_klass, recv_klass, vtable_start_offset + ioffset); 1278 // itableOffsetEntry[] itable = recv_klass + Klass::vtable_start_offset() + sizeof(vtableEntry) * recv_klass->_vtable_len; 1279 // temp_itbl_klass = itable[0]._interface; 1280 int vtblEntrySize = vtableEntry::size_in_bytes(); 1281 assert(vtblEntrySize == wordSize, "ldr lsl shift amount must be 3"); 1282 ldr(temp_itbl_klass, Address(recv_klass, scan_temp, Address::lsl(exact_log2(vtblEntrySize)))); 1283 mov(holder_offset, zr); 1284 // scan_temp = &(itable[0]._interface) 1285 lea(scan_temp, Address(recv_klass, scan_temp, Address::lsl(exact_log2(vtblEntrySize)))); 1286 1287 // Initial checks: 1288 // - if (holder_klass != resolved_klass), go to "scan for resolved" 1289 // - if (itable[0] == holder_klass), shortcut to "holder found" 1290 // - if (itable[0] == 0), no such interface 1291 cmp(resolved_klass, holder_klass); 1292 br(Assembler::NE, L_loop_search_resolved_entry); 1293 cmp(holder_klass, temp_itbl_klass); 1294 br(Assembler::EQ, L_holder_found); 1295 cbz(temp_itbl_klass, L_no_such_interface); 1296 1297 // Loop: Look for holder_klass record in itable 1298 // do { 1299 // temp_itbl_klass = *(scan_temp += itable_offset_entry_size); 1300 // if (temp_itbl_klass == holder_klass) { 1301 // goto L_holder_found; // Found! 1302 // } 1303 // } while (temp_itbl_klass != 0); 1304 // goto L_no_such_interface // Not found. 1305 Label L_search_holder; 1306 bind(L_search_holder); 1307 ldr(temp_itbl_klass, Address(pre(scan_temp, itable_offset_entry_size))); 1308 cmp(holder_klass, temp_itbl_klass); 1309 br(Assembler::EQ, L_holder_found); 1310 cbnz(temp_itbl_klass, L_search_holder); 1311 1312 b(L_no_such_interface); 1313 1314 // Loop: Look for resolved_class record in itable 1315 // while (true) { 1316 // temp_itbl_klass = *(scan_temp += itable_offset_entry_size); 1317 // if (temp_itbl_klass == 0) { 1318 // goto L_no_such_interface; 1319 // } 1320 // if (temp_itbl_klass == resolved_klass) { 1321 // goto L_resolved_found; // Found! 1322 // } 1323 // if (temp_itbl_klass == holder_klass) { 1324 // holder_offset = scan_temp; 1325 // } 1326 // } 1327 // 1328 Label L_loop_search_resolved; 1329 bind(L_loop_search_resolved); 1330 ldr(temp_itbl_klass, Address(pre(scan_temp, itable_offset_entry_size))); 1331 bind(L_loop_search_resolved_entry); 1332 cbz(temp_itbl_klass, L_no_such_interface); 1333 cmp(resolved_klass, temp_itbl_klass); 1334 br(Assembler::EQ, L_resolved_found); 1335 cmp(holder_klass, temp_itbl_klass); 1336 br(Assembler::NE, L_loop_search_resolved); 1337 mov(holder_offset, scan_temp); 1338 b(L_loop_search_resolved); 1339 1340 // See if we already have a holder klass. If not, go and scan for it. 1341 bind(L_resolved_found); 1342 cbz(holder_offset, L_search_holder); 1343 mov(scan_temp, holder_offset); 1344 1345 // Finally, scan_temp contains holder_klass vtable offset 1346 bind(L_holder_found); 1347 ldrw(method_result, Address(scan_temp, ooffset - ioffset)); 1348 add(recv_klass, recv_klass, itable_index * wordSize + in_bytes(itableMethodEntry::method_offset()) 1349 - vtable_start_offset - ioffset); // substract offsets to restore the original value of recv_klass 1350 ldr(method_result, Address(recv_klass, method_result, Address::uxtw(0))); 1351 } 1352 1353 // virtual method calling 1354 void MacroAssembler::lookup_virtual_method(Register recv_klass, 1355 RegisterOrConstant vtable_index, 1356 Register method_result) { 1357 assert(vtableEntry::size() * wordSize == 8, 1358 "adjust the scaling in the code below"); 1359 int64_t vtable_offset_in_bytes = in_bytes(Klass::vtable_start_offset() + vtableEntry::method_offset()); 1360 1361 if (vtable_index.is_register()) { 1362 lea(method_result, Address(recv_klass, 1363 vtable_index.as_register(), 1364 Address::lsl(LogBytesPerWord))); 1365 ldr(method_result, Address(method_result, vtable_offset_in_bytes)); 1366 } else { 1367 vtable_offset_in_bytes += vtable_index.as_constant() * wordSize; 1368 ldr(method_result, 1369 form_address(rscratch1, recv_klass, vtable_offset_in_bytes, 0)); 1370 } 1371 } 1372 1373 void MacroAssembler::check_klass_subtype(Register sub_klass, 1374 Register super_klass, 1375 Register temp_reg, 1376 Label& L_success) { 1377 Label L_failure; 1378 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, nullptr); 1379 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, nullptr); 1380 bind(L_failure); 1381 } 1382 1383 1384 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, 1385 Register super_klass, 1386 Register temp_reg, 1387 Label* L_success, 1388 Label* L_failure, 1389 Label* L_slow_path, 1390 RegisterOrConstant super_check_offset) { 1391 assert_different_registers(sub_klass, super_klass, temp_reg); 1392 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); 1393 if (super_check_offset.is_register()) { 1394 assert_different_registers(sub_klass, super_klass, 1395 super_check_offset.as_register()); 1396 } else if (must_load_sco) { 1397 assert(temp_reg != noreg, "supply either a temp or a register offset"); 1398 } 1399 1400 Label L_fallthrough; 1401 int label_nulls = 0; 1402 if (L_success == nullptr) { L_success = &L_fallthrough; label_nulls++; } 1403 if (L_failure == nullptr) { L_failure = &L_fallthrough; label_nulls++; } 1404 if (L_slow_path == nullptr) { L_slow_path = &L_fallthrough; label_nulls++; } 1405 assert(label_nulls <= 1, "at most one null in the batch"); 1406 1407 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 1408 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 1409 Address super_check_offset_addr(super_klass, sco_offset); 1410 1411 // Hacked jmp, which may only be used just before L_fallthrough. 1412 #define final_jmp(label) \ 1413 if (&(label) == &L_fallthrough) { /*do nothing*/ } \ 1414 else b(label) /*omit semi*/ 1415 1416 // If the pointers are equal, we are done (e.g., String[] elements). 1417 // This self-check enables sharing of secondary supertype arrays among 1418 // non-primary types such as array-of-interface. Otherwise, each such 1419 // type would need its own customized SSA. 1420 // We move this check to the front of the fast path because many 1421 // type checks are in fact trivially successful in this manner, 1422 // so we get a nicely predicted branch right at the start of the check. 1423 cmp(sub_klass, super_klass); 1424 br(Assembler::EQ, *L_success); 1425 1426 // Check the supertype display: 1427 if (must_load_sco) { 1428 ldrw(temp_reg, super_check_offset_addr); 1429 super_check_offset = RegisterOrConstant(temp_reg); 1430 } 1431 Address super_check_addr(sub_klass, super_check_offset); 1432 ldr(rscratch1, super_check_addr); 1433 cmp(super_klass, rscratch1); // load displayed supertype 1434 1435 // This check has worked decisively for primary supers. 1436 // Secondary supers are sought in the super_cache ('super_cache_addr'). 1437 // (Secondary supers are interfaces and very deeply nested subtypes.) 1438 // This works in the same check above because of a tricky aliasing 1439 // between the super_cache and the primary super display elements. 1440 // (The 'super_check_addr' can address either, as the case requires.) 1441 // Note that the cache is updated below if it does not help us find 1442 // what we need immediately. 1443 // So if it was a primary super, we can just fail immediately. 1444 // Otherwise, it's the slow path for us (no success at this point). 1445 1446 if (super_check_offset.is_register()) { 1447 br(Assembler::EQ, *L_success); 1448 subs(zr, super_check_offset.as_register(), sc_offset); 1449 if (L_failure == &L_fallthrough) { 1450 br(Assembler::EQ, *L_slow_path); 1451 } else { 1452 br(Assembler::NE, *L_failure); 1453 final_jmp(*L_slow_path); 1454 } 1455 } else if (super_check_offset.as_constant() == sc_offset) { 1456 // Need a slow path; fast failure is impossible. 1457 if (L_slow_path == &L_fallthrough) { 1458 br(Assembler::EQ, *L_success); 1459 } else { 1460 br(Assembler::NE, *L_slow_path); 1461 final_jmp(*L_success); 1462 } 1463 } else { 1464 // No slow path; it's a fast decision. 1465 if (L_failure == &L_fallthrough) { 1466 br(Assembler::EQ, *L_success); 1467 } else { 1468 br(Assembler::NE, *L_failure); 1469 final_jmp(*L_success); 1470 } 1471 } 1472 1473 bind(L_fallthrough); 1474 1475 #undef final_jmp 1476 } 1477 1478 // These two are taken from x86, but they look generally useful 1479 1480 // scans count pointer sized words at [addr] for occurrence of value, 1481 // generic 1482 void MacroAssembler::repne_scan(Register addr, Register value, Register count, 1483 Register scratch) { 1484 Label Lloop, Lexit; 1485 cbz(count, Lexit); 1486 bind(Lloop); 1487 ldr(scratch, post(addr, wordSize)); 1488 cmp(value, scratch); 1489 br(EQ, Lexit); 1490 sub(count, count, 1); 1491 cbnz(count, Lloop); 1492 bind(Lexit); 1493 } 1494 1495 // scans count 4 byte words at [addr] for occurrence of value, 1496 // generic 1497 void MacroAssembler::repne_scanw(Register addr, Register value, Register count, 1498 Register scratch) { 1499 Label Lloop, Lexit; 1500 cbz(count, Lexit); 1501 bind(Lloop); 1502 ldrw(scratch, post(addr, wordSize)); 1503 cmpw(value, scratch); 1504 br(EQ, Lexit); 1505 sub(count, count, 1); 1506 cbnz(count, Lloop); 1507 bind(Lexit); 1508 } 1509 1510 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, 1511 Register super_klass, 1512 Register temp_reg, 1513 Register temp2_reg, 1514 Label* L_success, 1515 Label* L_failure, 1516 bool set_cond_codes) { 1517 // NB! Callers may assume that, when temp2_reg is a valid register, 1518 // this code sets it to a nonzero value. 1519 1520 assert_different_registers(sub_klass, super_klass, temp_reg); 1521 if (temp2_reg != noreg) 1522 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg, rscratch1); 1523 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg) 1524 1525 Label L_fallthrough; 1526 int label_nulls = 0; 1527 if (L_success == nullptr) { L_success = &L_fallthrough; label_nulls++; } 1528 if (L_failure == nullptr) { L_failure = &L_fallthrough; label_nulls++; } 1529 assert(label_nulls <= 1, "at most one null in the batch"); 1530 1531 // a couple of useful fields in sub_klass: 1532 int ss_offset = in_bytes(Klass::secondary_supers_offset()); 1533 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); 1534 Address secondary_supers_addr(sub_klass, ss_offset); 1535 Address super_cache_addr( sub_klass, sc_offset); 1536 1537 BLOCK_COMMENT("check_klass_subtype_slow_path"); 1538 1539 // Do a linear scan of the secondary super-klass chain. 1540 // This code is rarely used, so simplicity is a virtue here. 1541 // The repne_scan instruction uses fixed registers, which we must spill. 1542 // Don't worry too much about pre-existing connections with the input regs. 1543 1544 assert(sub_klass != r0, "killed reg"); // killed by mov(r0, super) 1545 assert(sub_klass != r2, "killed reg"); // killed by lea(r2, &pst_counter) 1546 1547 RegSet pushed_registers; 1548 if (!IS_A_TEMP(r2)) pushed_registers += r2; 1549 if (!IS_A_TEMP(r5)) pushed_registers += r5; 1550 1551 if (super_klass != r0) { 1552 if (!IS_A_TEMP(r0)) pushed_registers += r0; 1553 } 1554 1555 push(pushed_registers, sp); 1556 1557 // Get super_klass value into r0 (even if it was in r5 or r2). 1558 if (super_klass != r0) { 1559 mov(r0, super_klass); 1560 } 1561 1562 #ifndef PRODUCT 1563 incrementw(ExternalAddress((address)&SharedRuntime::_partial_subtype_ctr)); 1564 #endif //PRODUCT 1565 1566 // We will consult the secondary-super array. 1567 ldr(r5, secondary_supers_addr); 1568 // Load the array length. 1569 ldrw(r2, Address(r5, Array<Klass*>::length_offset_in_bytes())); 1570 // Skip to start of data. 1571 add(r5, r5, Array<Klass*>::base_offset_in_bytes()); 1572 1573 cmp(sp, zr); // Clear Z flag; SP is never zero 1574 // Scan R2 words at [R5] for an occurrence of R0. 1575 // Set NZ/Z based on last compare. 1576 repne_scan(r5, r0, r2, rscratch1); 1577 1578 // Unspill the temp. registers: 1579 pop(pushed_registers, sp); 1580 1581 br(Assembler::NE, *L_failure); 1582 1583 // Success. Cache the super we found and proceed in triumph. 1584 str(super_klass, super_cache_addr); 1585 1586 if (L_success != &L_fallthrough) { 1587 b(*L_success); 1588 } 1589 1590 #undef IS_A_TEMP 1591 1592 bind(L_fallthrough); 1593 } 1594 1595 // Ensure that the inline code and the stub are using the same registers. 1596 #define LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS \ 1597 do { \ 1598 assert(r_super_klass == r0 && \ 1599 r_array_base == r1 && \ 1600 r_array_length == r2 && \ 1601 (r_array_index == r3 || r_array_index == noreg) && \ 1602 (r_sub_klass == r4 || r_sub_klass == noreg) && \ 1603 (r_bitmap == rscratch2 || r_bitmap == noreg) && \ 1604 (result == r5 || result == noreg), "registers must match aarch64.ad"); \ 1605 } while(0) 1606 1607 // Return true: we succeeded in generating this code 1608 bool MacroAssembler::lookup_secondary_supers_table(Register r_sub_klass, 1609 Register r_super_klass, 1610 Register temp1, 1611 Register temp2, 1612 Register temp3, 1613 FloatRegister vtemp, 1614 Register result, 1615 u1 super_klass_slot, 1616 bool stub_is_near) { 1617 assert_different_registers(r_sub_klass, temp1, temp2, temp3, result, rscratch1, rscratch2); 1618 1619 Label L_fallthrough; 1620 1621 BLOCK_COMMENT("lookup_secondary_supers_table {"); 1622 1623 const Register 1624 r_array_base = temp1, // r1 1625 r_array_length = temp2, // r2 1626 r_array_index = temp3, // r3 1627 r_bitmap = rscratch2; 1628 1629 LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS; 1630 1631 u1 bit = super_klass_slot; 1632 1633 // Make sure that result is nonzero if the TBZ below misses. 1634 mov(result, 1); 1635 1636 // We're going to need the bitmap in a vector reg and in a core reg, 1637 // so load both now. 1638 ldr(r_bitmap, Address(r_sub_klass, Klass::bitmap_offset())); 1639 if (bit != 0) { 1640 ldrd(vtemp, Address(r_sub_klass, Klass::bitmap_offset())); 1641 } 1642 // First check the bitmap to see if super_klass might be present. If 1643 // the bit is zero, we are certain that super_klass is not one of 1644 // the secondary supers. 1645 tbz(r_bitmap, bit, L_fallthrough); 1646 1647 // Get the first array index that can contain super_klass into r_array_index. 1648 if (bit != 0) { 1649 shld(vtemp, vtemp, Klass::SECONDARY_SUPERS_TABLE_MASK - bit); 1650 cnt(vtemp, T8B, vtemp); 1651 addv(vtemp, T8B, vtemp); 1652 fmovd(r_array_index, vtemp); 1653 } else { 1654 mov(r_array_index, (u1)1); 1655 } 1656 // NB! r_array_index is off by 1. It is compensated by keeping r_array_base off by 1 word. 1657 1658 // We will consult the secondary-super array. 1659 ldr(r_array_base, Address(r_sub_klass, in_bytes(Klass::secondary_supers_offset()))); 1660 1661 // The value i in r_array_index is >= 1, so even though r_array_base 1662 // points to the length, we don't need to adjust it to point to the 1663 // data. 1664 assert(Array<Klass*>::base_offset_in_bytes() == wordSize, "Adjust this code"); 1665 assert(Array<Klass*>::length_offset_in_bytes() == 0, "Adjust this code"); 1666 1667 ldr(result, Address(r_array_base, r_array_index, Address::lsl(LogBytesPerWord))); 1668 eor(result, result, r_super_klass); 1669 cbz(result, L_fallthrough); // Found a match 1670 1671 // Is there another entry to check? Consult the bitmap. 1672 tbz(r_bitmap, (bit + 1) & Klass::SECONDARY_SUPERS_TABLE_MASK, L_fallthrough); 1673 1674 // Linear probe. 1675 if (bit != 0) { 1676 ror(r_bitmap, r_bitmap, bit); 1677 } 1678 1679 // The slot we just inspected is at secondary_supers[r_array_index - 1]. 1680 // The next slot to be inspected, by the stub we're about to call, 1681 // is secondary_supers[r_array_index]. Bits 0 and 1 in the bitmap 1682 // have been checked. 1683 Address stub = RuntimeAddress(StubRoutines::lookup_secondary_supers_table_slow_path_stub()); 1684 if (stub_is_near) { 1685 bl(stub); 1686 } else { 1687 address call = trampoline_call(stub); 1688 if (call == nullptr) { 1689 return false; // trampoline allocation failed 1690 } 1691 } 1692 1693 BLOCK_COMMENT("} lookup_secondary_supers_table"); 1694 1695 bind(L_fallthrough); 1696 1697 if (VerifySecondarySupers) { 1698 verify_secondary_supers_table(r_sub_klass, r_super_klass, // r4, r0 1699 temp1, temp2, result); // r1, r2, r5 1700 } 1701 return true; 1702 } 1703 1704 // Called by code generated by check_klass_subtype_slow_path 1705 // above. This is called when there is a collision in the hashed 1706 // lookup in the secondary supers array. 1707 void MacroAssembler::lookup_secondary_supers_table_slow_path(Register r_super_klass, 1708 Register r_array_base, 1709 Register r_array_index, 1710 Register r_bitmap, 1711 Register temp1, 1712 Register result) { 1713 assert_different_registers(r_super_klass, r_array_base, r_array_index, r_bitmap, temp1, result, rscratch1); 1714 1715 const Register 1716 r_array_length = temp1, 1717 r_sub_klass = noreg; // unused 1718 1719 LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS; 1720 1721 Label L_fallthrough, L_huge; 1722 1723 // Load the array length. 1724 ldrw(r_array_length, Address(r_array_base, Array<Klass*>::length_offset_in_bytes())); 1725 // And adjust the array base to point to the data. 1726 // NB! Effectively increments current slot index by 1. 1727 assert(Array<Klass*>::base_offset_in_bytes() == wordSize, ""); 1728 add(r_array_base, r_array_base, Array<Klass*>::base_offset_in_bytes()); 1729 1730 // The bitmap is full to bursting. 1731 // Implicit invariant: BITMAP_FULL implies (length > 0) 1732 assert(Klass::SECONDARY_SUPERS_BITMAP_FULL == ~uintx(0), ""); 1733 cmn(r_bitmap, (u1)1); 1734 br(EQ, L_huge); 1735 1736 // NB! Our caller has checked bits 0 and 1 in the bitmap. The 1737 // current slot (at secondary_supers[r_array_index]) has not yet 1738 // been inspected, and r_array_index may be out of bounds if we 1739 // wrapped around the end of the array. 1740 1741 { // This is conventional linear probing, but instead of terminating 1742 // when a null entry is found in the table, we maintain a bitmap 1743 // in which a 0 indicates missing entries. 1744 // The check above guarantees there are 0s in the bitmap, so the loop 1745 // eventually terminates. 1746 Label L_loop; 1747 bind(L_loop); 1748 1749 // Check for wraparound. 1750 cmp(r_array_index, r_array_length); 1751 csel(r_array_index, zr, r_array_index, GE); 1752 1753 ldr(rscratch1, Address(r_array_base, r_array_index, Address::lsl(LogBytesPerWord))); 1754 eor(result, rscratch1, r_super_klass); 1755 cbz(result, L_fallthrough); 1756 1757 tbz(r_bitmap, 2, L_fallthrough); // look-ahead check (Bit 2); result is non-zero 1758 1759 ror(r_bitmap, r_bitmap, 1); 1760 add(r_array_index, r_array_index, 1); 1761 b(L_loop); 1762 } 1763 1764 { // Degenerate case: more than 64 secondary supers. 1765 // FIXME: We could do something smarter here, maybe a vectorized 1766 // comparison or a binary search, but is that worth any added 1767 // complexity? 1768 bind(L_huge); 1769 cmp(sp, zr); // Clear Z flag; SP is never zero 1770 repne_scan(r_array_base, r_super_klass, r_array_length, rscratch1); 1771 cset(result, NE); // result == 0 iff we got a match. 1772 } 1773 1774 bind(L_fallthrough); 1775 } 1776 1777 // Make sure that the hashed lookup and a linear scan agree. 1778 void MacroAssembler::verify_secondary_supers_table(Register r_sub_klass, 1779 Register r_super_klass, 1780 Register temp1, 1781 Register temp2, 1782 Register result) { 1783 assert_different_registers(r_sub_klass, r_super_klass, temp1, temp2, result, rscratch1); 1784 1785 const Register 1786 r_array_base = temp1, 1787 r_array_length = temp2, 1788 r_array_index = noreg, // unused 1789 r_bitmap = noreg; // unused 1790 1791 LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS; 1792 1793 BLOCK_COMMENT("verify_secondary_supers_table {"); 1794 1795 // We will consult the secondary-super array. 1796 ldr(r_array_base, Address(r_sub_klass, in_bytes(Klass::secondary_supers_offset()))); 1797 1798 // Load the array length. 1799 ldrw(r_array_length, Address(r_array_base, Array<Klass*>::length_offset_in_bytes())); 1800 // And adjust the array base to point to the data. 1801 add(r_array_base, r_array_base, Array<Klass*>::base_offset_in_bytes()); 1802 1803 cmp(sp, zr); // Clear Z flag; SP is never zero 1804 // Scan R2 words at [R5] for an occurrence of R0. 1805 // Set NZ/Z based on last compare. 1806 repne_scan(/*addr*/r_array_base, /*value*/r_super_klass, /*count*/r_array_length, rscratch2); 1807 // rscratch1 == 0 iff we got a match. 1808 cset(rscratch1, NE); 1809 1810 Label passed; 1811 cmp(result, zr); 1812 cset(result, NE); // normalize result to 0/1 for comparison 1813 1814 cmp(rscratch1, result); 1815 br(EQ, passed); 1816 { 1817 mov(r0, r_super_klass); // r0 <- r0 1818 mov(r1, r_sub_klass); // r1 <- r4 1819 mov(r2, /*expected*/rscratch1); // r2 <- r8 1820 mov(r3, result); // r3 <- r5 1821 mov(r4, (address)("mismatch")); // r4 <- const 1822 rt_call(CAST_FROM_FN_PTR(address, Klass::on_secondary_supers_verification_failure), rscratch2); 1823 should_not_reach_here(); 1824 } 1825 bind(passed); 1826 1827 BLOCK_COMMENT("} verify_secondary_supers_table"); 1828 } 1829 1830 void MacroAssembler::clinit_barrier(Register klass, Register scratch, Label* L_fast_path, Label* L_slow_path) { 1831 assert(L_fast_path != nullptr || L_slow_path != nullptr, "at least one is required"); 1832 assert_different_registers(klass, rthread, scratch); 1833 1834 Label L_fallthrough, L_tmp; 1835 if (L_fast_path == nullptr) { 1836 L_fast_path = &L_fallthrough; 1837 } else if (L_slow_path == nullptr) { 1838 L_slow_path = &L_fallthrough; 1839 } 1840 // Fast path check: class is fully initialized 1841 ldrb(scratch, Address(klass, InstanceKlass::init_state_offset())); 1842 subs(zr, scratch, InstanceKlass::fully_initialized); 1843 br(Assembler::EQ, *L_fast_path); 1844 1845 // Fast path check: current thread is initializer thread 1846 ldr(scratch, Address(klass, InstanceKlass::init_thread_offset())); 1847 cmp(rthread, scratch); 1848 1849 if (L_slow_path == &L_fallthrough) { 1850 br(Assembler::EQ, *L_fast_path); 1851 bind(*L_slow_path); 1852 } else if (L_fast_path == &L_fallthrough) { 1853 br(Assembler::NE, *L_slow_path); 1854 bind(*L_fast_path); 1855 } else { 1856 Unimplemented(); 1857 } 1858 } 1859 1860 void MacroAssembler::_verify_oop(Register reg, const char* s, const char* file, int line) { 1861 if (!VerifyOops) return; 1862 1863 // Pass register number to verify_oop_subroutine 1864 const char* b = nullptr; 1865 { 1866 ResourceMark rm; 1867 stringStream ss; 1868 ss.print("verify_oop: %s: %s (%s:%d)", reg->name(), s, file, line); 1869 b = code_string(ss.as_string()); 1870 } 1871 BLOCK_COMMENT("verify_oop {"); 1872 1873 strip_return_address(); // This might happen within a stack frame. 1874 protect_return_address(); 1875 stp(r0, rscratch1, Address(pre(sp, -2 * wordSize))); 1876 stp(rscratch2, lr, Address(pre(sp, -2 * wordSize))); 1877 1878 mov(r0, reg); 1879 movptr(rscratch1, (uintptr_t)(address)b); 1880 1881 // call indirectly to solve generation ordering problem 1882 lea(rscratch2, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 1883 ldr(rscratch2, Address(rscratch2)); 1884 blr(rscratch2); 1885 1886 ldp(rscratch2, lr, Address(post(sp, 2 * wordSize))); 1887 ldp(r0, rscratch1, Address(post(sp, 2 * wordSize))); 1888 authenticate_return_address(); 1889 1890 BLOCK_COMMENT("} verify_oop"); 1891 } 1892 1893 void MacroAssembler::_verify_oop_addr(Address addr, const char* s, const char* file, int line) { 1894 if (!VerifyOops) return; 1895 1896 const char* b = nullptr; 1897 { 1898 ResourceMark rm; 1899 stringStream ss; 1900 ss.print("verify_oop_addr: %s (%s:%d)", s, file, line); 1901 b = code_string(ss.as_string()); 1902 } 1903 BLOCK_COMMENT("verify_oop_addr {"); 1904 1905 strip_return_address(); // This might happen within a stack frame. 1906 protect_return_address(); 1907 stp(r0, rscratch1, Address(pre(sp, -2 * wordSize))); 1908 stp(rscratch2, lr, Address(pre(sp, -2 * wordSize))); 1909 1910 // addr may contain sp so we will have to adjust it based on the 1911 // pushes that we just did. 1912 if (addr.uses(sp)) { 1913 lea(r0, addr); 1914 ldr(r0, Address(r0, 4 * wordSize)); 1915 } else { 1916 ldr(r0, addr); 1917 } 1918 movptr(rscratch1, (uintptr_t)(address)b); 1919 1920 // call indirectly to solve generation ordering problem 1921 lea(rscratch2, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); 1922 ldr(rscratch2, Address(rscratch2)); 1923 blr(rscratch2); 1924 1925 ldp(rscratch2, lr, Address(post(sp, 2 * wordSize))); 1926 ldp(r0, rscratch1, Address(post(sp, 2 * wordSize))); 1927 authenticate_return_address(); 1928 1929 BLOCK_COMMENT("} verify_oop_addr"); 1930 } 1931 1932 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, 1933 int extra_slot_offset) { 1934 // cf. TemplateTable::prepare_invoke(), if (load_receiver). 1935 int stackElementSize = Interpreter::stackElementSize; 1936 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0); 1937 #ifdef ASSERT 1938 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1); 1939 assert(offset1 - offset == stackElementSize, "correct arithmetic"); 1940 #endif 1941 if (arg_slot.is_constant()) { 1942 return Address(esp, arg_slot.as_constant() * stackElementSize 1943 + offset); 1944 } else { 1945 add(rscratch1, esp, arg_slot.as_register(), 1946 ext::uxtx, exact_log2(stackElementSize)); 1947 return Address(rscratch1, offset); 1948 } 1949 } 1950 1951 void MacroAssembler::call_VM_leaf_base(address entry_point, 1952 int number_of_arguments, 1953 Label *retaddr) { 1954 Label E, L; 1955 1956 stp(rscratch1, rmethod, Address(pre(sp, -2 * wordSize))); 1957 1958 mov(rscratch1, entry_point); 1959 blr(rscratch1); 1960 if (retaddr) 1961 bind(*retaddr); 1962 1963 ldp(rscratch1, rmethod, Address(post(sp, 2 * wordSize))); 1964 } 1965 1966 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) { 1967 call_VM_leaf_base(entry_point, number_of_arguments); 1968 } 1969 1970 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) { 1971 pass_arg0(this, arg_0); 1972 call_VM_leaf_base(entry_point, 1); 1973 } 1974 1975 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 1976 assert_different_registers(arg_1, c_rarg0); 1977 pass_arg0(this, arg_0); 1978 pass_arg1(this, arg_1); 1979 call_VM_leaf_base(entry_point, 2); 1980 } 1981 1982 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, 1983 Register arg_1, Register arg_2) { 1984 assert_different_registers(arg_1, c_rarg0); 1985 assert_different_registers(arg_2, c_rarg0, c_rarg1); 1986 pass_arg0(this, arg_0); 1987 pass_arg1(this, arg_1); 1988 pass_arg2(this, arg_2); 1989 call_VM_leaf_base(entry_point, 3); 1990 } 1991 1992 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) { 1993 pass_arg0(this, arg_0); 1994 MacroAssembler::call_VM_leaf_base(entry_point, 1); 1995 } 1996 1997 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { 1998 1999 assert_different_registers(arg_0, c_rarg1); 2000 pass_arg1(this, arg_1); 2001 pass_arg0(this, arg_0); 2002 MacroAssembler::call_VM_leaf_base(entry_point, 2); 2003 } 2004 2005 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { 2006 assert_different_registers(arg_0, c_rarg1, c_rarg2); 2007 assert_different_registers(arg_1, c_rarg2); 2008 pass_arg2(this, arg_2); 2009 pass_arg1(this, arg_1); 2010 pass_arg0(this, arg_0); 2011 MacroAssembler::call_VM_leaf_base(entry_point, 3); 2012 } 2013 2014 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) { 2015 assert_different_registers(arg_0, c_rarg1, c_rarg2, c_rarg3); 2016 assert_different_registers(arg_1, c_rarg2, c_rarg3); 2017 assert_different_registers(arg_2, c_rarg3); 2018 pass_arg3(this, arg_3); 2019 pass_arg2(this, arg_2); 2020 pass_arg1(this, arg_1); 2021 pass_arg0(this, arg_0); 2022 MacroAssembler::call_VM_leaf_base(entry_point, 4); 2023 } 2024 2025 void MacroAssembler::null_check(Register reg, int offset) { 2026 if (needs_explicit_null_check(offset)) { 2027 // provoke OS null exception if reg is null by 2028 // accessing M[reg] w/o changing any registers 2029 // NOTE: this is plenty to provoke a segv 2030 ldr(zr, Address(reg)); 2031 } else { 2032 // nothing to do, (later) access of M[reg + offset] 2033 // will provoke OS null exception if reg is null 2034 } 2035 } 2036 2037 // MacroAssembler protected routines needed to implement 2038 // public methods 2039 2040 void MacroAssembler::mov(Register r, Address dest) { 2041 code_section()->relocate(pc(), dest.rspec()); 2042 uint64_t imm64 = (uint64_t)dest.target(); 2043 movptr(r, imm64); 2044 } 2045 2046 // Move a constant pointer into r. In AArch64 mode the virtual 2047 // address space is 48 bits in size, so we only need three 2048 // instructions to create a patchable instruction sequence that can 2049 // reach anywhere. 2050 void MacroAssembler::movptr(Register r, uintptr_t imm64) { 2051 #ifndef PRODUCT 2052 { 2053 char buffer[64]; 2054 snprintf(buffer, sizeof(buffer), "0x%" PRIX64, (uint64_t)imm64); 2055 block_comment(buffer); 2056 } 2057 #endif 2058 assert(imm64 < (1ull << 48), "48-bit overflow in address constant"); 2059 movz(r, imm64 & 0xffff); 2060 imm64 >>= 16; 2061 movk(r, imm64 & 0xffff, 16); 2062 imm64 >>= 16; 2063 movk(r, imm64 & 0xffff, 32); 2064 } 2065 2066 // Macro to mov replicated immediate to vector register. 2067 // imm64: only the lower 8/16/32 bits are considered for B/H/S type. That is, 2068 // the upper 56/48/32 bits must be zeros for B/H/S type. 2069 // Vd will get the following values for different arrangements in T 2070 // imm64 == hex 000000gh T8B: Vd = ghghghghghghghgh 2071 // imm64 == hex 000000gh T16B: Vd = ghghghghghghghghghghghghghghghgh 2072 // imm64 == hex 0000efgh T4H: Vd = efghefghefghefgh 2073 // imm64 == hex 0000efgh T8H: Vd = efghefghefghefghefghefghefghefgh 2074 // imm64 == hex abcdefgh T2S: Vd = abcdefghabcdefgh 2075 // imm64 == hex abcdefgh T4S: Vd = abcdefghabcdefghabcdefghabcdefgh 2076 // imm64 == hex abcdefgh T1D: Vd = 00000000abcdefgh 2077 // imm64 == hex abcdefgh T2D: Vd = 00000000abcdefgh00000000abcdefgh 2078 // Clobbers rscratch1 2079 void MacroAssembler::mov(FloatRegister Vd, SIMD_Arrangement T, uint64_t imm64) { 2080 assert(T != T1Q, "unsupported"); 2081 if (T == T1D || T == T2D) { 2082 int imm = operand_valid_for_movi_immediate(imm64, T); 2083 if (-1 != imm) { 2084 movi(Vd, T, imm); 2085 } else { 2086 mov(rscratch1, imm64); 2087 dup(Vd, T, rscratch1); 2088 } 2089 return; 2090 } 2091 2092 #ifdef ASSERT 2093 if (T == T8B || T == T16B) assert((imm64 & ~0xff) == 0, "extraneous bits (T8B/T16B)"); 2094 if (T == T4H || T == T8H) assert((imm64 & ~0xffff) == 0, "extraneous bits (T4H/T8H)"); 2095 if (T == T2S || T == T4S) assert((imm64 & ~0xffffffff) == 0, "extraneous bits (T2S/T4S)"); 2096 #endif 2097 int shift = operand_valid_for_movi_immediate(imm64, T); 2098 uint32_t imm32 = imm64 & 0xffffffffULL; 2099 if (shift >= 0) { 2100 movi(Vd, T, (imm32 >> shift) & 0xff, shift); 2101 } else { 2102 movw(rscratch1, imm32); 2103 dup(Vd, T, rscratch1); 2104 } 2105 } 2106 2107 void MacroAssembler::mov_immediate64(Register dst, uint64_t imm64) 2108 { 2109 #ifndef PRODUCT 2110 { 2111 char buffer[64]; 2112 snprintf(buffer, sizeof(buffer), "0x%" PRIX64, imm64); 2113 block_comment(buffer); 2114 } 2115 #endif 2116 if (operand_valid_for_logical_immediate(false, imm64)) { 2117 orr(dst, zr, imm64); 2118 } else { 2119 // we can use a combination of MOVZ or MOVN with 2120 // MOVK to build up the constant 2121 uint64_t imm_h[4]; 2122 int zero_count = 0; 2123 int neg_count = 0; 2124 int i; 2125 for (i = 0; i < 4; i++) { 2126 imm_h[i] = ((imm64 >> (i * 16)) & 0xffffL); 2127 if (imm_h[i] == 0) { 2128 zero_count++; 2129 } else if (imm_h[i] == 0xffffL) { 2130 neg_count++; 2131 } 2132 } 2133 if (zero_count == 4) { 2134 // one MOVZ will do 2135 movz(dst, 0); 2136 } else if (neg_count == 4) { 2137 // one MOVN will do 2138 movn(dst, 0); 2139 } else if (zero_count == 3) { 2140 for (i = 0; i < 4; i++) { 2141 if (imm_h[i] != 0L) { 2142 movz(dst, (uint32_t)imm_h[i], (i << 4)); 2143 break; 2144 } 2145 } 2146 } else if (neg_count == 3) { 2147 // one MOVN will do 2148 for (int i = 0; i < 4; i++) { 2149 if (imm_h[i] != 0xffffL) { 2150 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4)); 2151 break; 2152 } 2153 } 2154 } else if (zero_count == 2) { 2155 // one MOVZ and one MOVK will do 2156 for (i = 0; i < 3; i++) { 2157 if (imm_h[i] != 0L) { 2158 movz(dst, (uint32_t)imm_h[i], (i << 4)); 2159 i++; 2160 break; 2161 } 2162 } 2163 for (;i < 4; i++) { 2164 if (imm_h[i] != 0L) { 2165 movk(dst, (uint32_t)imm_h[i], (i << 4)); 2166 } 2167 } 2168 } else if (neg_count == 2) { 2169 // one MOVN and one MOVK will do 2170 for (i = 0; i < 4; i++) { 2171 if (imm_h[i] != 0xffffL) { 2172 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4)); 2173 i++; 2174 break; 2175 } 2176 } 2177 for (;i < 4; i++) { 2178 if (imm_h[i] != 0xffffL) { 2179 movk(dst, (uint32_t)imm_h[i], (i << 4)); 2180 } 2181 } 2182 } else if (zero_count == 1) { 2183 // one MOVZ and two MOVKs will do 2184 for (i = 0; i < 4; i++) { 2185 if (imm_h[i] != 0L) { 2186 movz(dst, (uint32_t)imm_h[i], (i << 4)); 2187 i++; 2188 break; 2189 } 2190 } 2191 for (;i < 4; i++) { 2192 if (imm_h[i] != 0x0L) { 2193 movk(dst, (uint32_t)imm_h[i], (i << 4)); 2194 } 2195 } 2196 } else if (neg_count == 1) { 2197 // one MOVN and two MOVKs will do 2198 for (i = 0; i < 4; i++) { 2199 if (imm_h[i] != 0xffffL) { 2200 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4)); 2201 i++; 2202 break; 2203 } 2204 } 2205 for (;i < 4; i++) { 2206 if (imm_h[i] != 0xffffL) { 2207 movk(dst, (uint32_t)imm_h[i], (i << 4)); 2208 } 2209 } 2210 } else { 2211 // use a MOVZ and 3 MOVKs (makes it easier to debug) 2212 movz(dst, (uint32_t)imm_h[0], 0); 2213 for (i = 1; i < 4; i++) { 2214 movk(dst, (uint32_t)imm_h[i], (i << 4)); 2215 } 2216 } 2217 } 2218 } 2219 2220 void MacroAssembler::mov_immediate32(Register dst, uint32_t imm32) 2221 { 2222 #ifndef PRODUCT 2223 { 2224 char buffer[64]; 2225 snprintf(buffer, sizeof(buffer), "0x%" PRIX32, imm32); 2226 block_comment(buffer); 2227 } 2228 #endif 2229 if (operand_valid_for_logical_immediate(true, imm32)) { 2230 orrw(dst, zr, imm32); 2231 } else { 2232 // we can use MOVZ, MOVN or two calls to MOVK to build up the 2233 // constant 2234 uint32_t imm_h[2]; 2235 imm_h[0] = imm32 & 0xffff; 2236 imm_h[1] = ((imm32 >> 16) & 0xffff); 2237 if (imm_h[0] == 0) { 2238 movzw(dst, imm_h[1], 16); 2239 } else if (imm_h[0] == 0xffff) { 2240 movnw(dst, imm_h[1] ^ 0xffff, 16); 2241 } else if (imm_h[1] == 0) { 2242 movzw(dst, imm_h[0], 0); 2243 } else if (imm_h[1] == 0xffff) { 2244 movnw(dst, imm_h[0] ^ 0xffff, 0); 2245 } else { 2246 // use a MOVZ and MOVK (makes it easier to debug) 2247 movzw(dst, imm_h[0], 0); 2248 movkw(dst, imm_h[1], 16); 2249 } 2250 } 2251 } 2252 2253 // Form an address from base + offset in Rd. Rd may or may 2254 // not actually be used: you must use the Address that is returned. 2255 // It is up to you to ensure that the shift provided matches the size 2256 // of your data. 2257 Address MacroAssembler::form_address(Register Rd, Register base, int64_t byte_offset, int shift) { 2258 if (Address::offset_ok_for_immed(byte_offset, shift)) 2259 // It fits; no need for any heroics 2260 return Address(base, byte_offset); 2261 2262 // Don't do anything clever with negative or misaligned offsets 2263 unsigned mask = (1 << shift) - 1; 2264 if (byte_offset < 0 || byte_offset & mask) { 2265 mov(Rd, byte_offset); 2266 add(Rd, base, Rd); 2267 return Address(Rd); 2268 } 2269 2270 // See if we can do this with two 12-bit offsets 2271 { 2272 uint64_t word_offset = byte_offset >> shift; 2273 uint64_t masked_offset = word_offset & 0xfff000; 2274 if (Address::offset_ok_for_immed(word_offset - masked_offset, 0) 2275 && Assembler::operand_valid_for_add_sub_immediate(masked_offset << shift)) { 2276 add(Rd, base, masked_offset << shift); 2277 word_offset -= masked_offset; 2278 return Address(Rd, word_offset << shift); 2279 } 2280 } 2281 2282 // Do it the hard way 2283 mov(Rd, byte_offset); 2284 add(Rd, base, Rd); 2285 return Address(Rd); 2286 } 2287 2288 int MacroAssembler::corrected_idivl(Register result, Register ra, Register rb, 2289 bool want_remainder, Register scratch) 2290 { 2291 // Full implementation of Java idiv and irem. The function 2292 // returns the (pc) offset of the div instruction - may be needed 2293 // for implicit exceptions. 2294 // 2295 // constraint : ra/rb =/= scratch 2296 // normal case 2297 // 2298 // input : ra: dividend 2299 // rb: divisor 2300 // 2301 // result: either 2302 // quotient (= ra idiv rb) 2303 // remainder (= ra irem rb) 2304 2305 assert(ra != scratch && rb != scratch, "reg cannot be scratch"); 2306 2307 int idivl_offset = offset(); 2308 if (! want_remainder) { 2309 sdivw(result, ra, rb); 2310 } else { 2311 sdivw(scratch, ra, rb); 2312 Assembler::msubw(result, scratch, rb, ra); 2313 } 2314 2315 return idivl_offset; 2316 } 2317 2318 int MacroAssembler::corrected_idivq(Register result, Register ra, Register rb, 2319 bool want_remainder, Register scratch) 2320 { 2321 // Full implementation of Java ldiv and lrem. The function 2322 // returns the (pc) offset of the div instruction - may be needed 2323 // for implicit exceptions. 2324 // 2325 // constraint : ra/rb =/= scratch 2326 // normal case 2327 // 2328 // input : ra: dividend 2329 // rb: divisor 2330 // 2331 // result: either 2332 // quotient (= ra idiv rb) 2333 // remainder (= ra irem rb) 2334 2335 assert(ra != scratch && rb != scratch, "reg cannot be scratch"); 2336 2337 int idivq_offset = offset(); 2338 if (! want_remainder) { 2339 sdiv(result, ra, rb); 2340 } else { 2341 sdiv(scratch, ra, rb); 2342 Assembler::msub(result, scratch, rb, ra); 2343 } 2344 2345 return idivq_offset; 2346 } 2347 2348 void MacroAssembler::membar(Membar_mask_bits order_constraint) { 2349 address prev = pc() - NativeMembar::instruction_size; 2350 address last = code()->last_insn(); 2351 if (last != nullptr && nativeInstruction_at(last)->is_Membar() && prev == last) { 2352 NativeMembar *bar = NativeMembar_at(prev); 2353 // We are merging two memory barrier instructions. On AArch64 we 2354 // can do this simply by ORing them together. 2355 bar->set_kind(bar->get_kind() | order_constraint); 2356 BLOCK_COMMENT("merged membar"); 2357 } else { 2358 code()->set_last_insn(pc()); 2359 dmb(Assembler::barrier(order_constraint)); 2360 } 2361 } 2362 2363 bool MacroAssembler::try_merge_ldst(Register rt, const Address &adr, size_t size_in_bytes, bool is_store) { 2364 if (ldst_can_merge(rt, adr, size_in_bytes, is_store)) { 2365 merge_ldst(rt, adr, size_in_bytes, is_store); 2366 code()->clear_last_insn(); 2367 return true; 2368 } else { 2369 assert(size_in_bytes == 8 || size_in_bytes == 4, "only 8 bytes or 4 bytes load/store is supported."); 2370 const uint64_t mask = size_in_bytes - 1; 2371 if (adr.getMode() == Address::base_plus_offset && 2372 (adr.offset() & mask) == 0) { // only supports base_plus_offset. 2373 code()->set_last_insn(pc()); 2374 } 2375 return false; 2376 } 2377 } 2378 2379 void MacroAssembler::ldr(Register Rx, const Address &adr) { 2380 // We always try to merge two adjacent loads into one ldp. 2381 if (!try_merge_ldst(Rx, adr, 8, false)) { 2382 Assembler::ldr(Rx, adr); 2383 } 2384 } 2385 2386 void MacroAssembler::ldrw(Register Rw, const Address &adr) { 2387 // We always try to merge two adjacent loads into one ldp. 2388 if (!try_merge_ldst(Rw, adr, 4, false)) { 2389 Assembler::ldrw(Rw, adr); 2390 } 2391 } 2392 2393 void MacroAssembler::str(Register Rx, const Address &adr) { 2394 // We always try to merge two adjacent stores into one stp. 2395 if (!try_merge_ldst(Rx, adr, 8, true)) { 2396 Assembler::str(Rx, adr); 2397 } 2398 } 2399 2400 void MacroAssembler::strw(Register Rw, const Address &adr) { 2401 // We always try to merge two adjacent stores into one stp. 2402 if (!try_merge_ldst(Rw, adr, 4, true)) { 2403 Assembler::strw(Rw, adr); 2404 } 2405 } 2406 2407 // MacroAssembler routines found actually to be needed 2408 2409 void MacroAssembler::push(Register src) 2410 { 2411 str(src, Address(pre(esp, -1 * wordSize))); 2412 } 2413 2414 void MacroAssembler::pop(Register dst) 2415 { 2416 ldr(dst, Address(post(esp, 1 * wordSize))); 2417 } 2418 2419 // Note: load_unsigned_short used to be called load_unsigned_word. 2420 int MacroAssembler::load_unsigned_short(Register dst, Address src) { 2421 int off = offset(); 2422 ldrh(dst, src); 2423 return off; 2424 } 2425 2426 int MacroAssembler::load_unsigned_byte(Register dst, Address src) { 2427 int off = offset(); 2428 ldrb(dst, src); 2429 return off; 2430 } 2431 2432 int MacroAssembler::load_signed_short(Register dst, Address src) { 2433 int off = offset(); 2434 ldrsh(dst, src); 2435 return off; 2436 } 2437 2438 int MacroAssembler::load_signed_byte(Register dst, Address src) { 2439 int off = offset(); 2440 ldrsb(dst, src); 2441 return off; 2442 } 2443 2444 int MacroAssembler::load_signed_short32(Register dst, Address src) { 2445 int off = offset(); 2446 ldrshw(dst, src); 2447 return off; 2448 } 2449 2450 int MacroAssembler::load_signed_byte32(Register dst, Address src) { 2451 int off = offset(); 2452 ldrsbw(dst, src); 2453 return off; 2454 } 2455 2456 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed) { 2457 switch (size_in_bytes) { 2458 case 8: ldr(dst, src); break; 2459 case 4: ldrw(dst, src); break; 2460 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break; 2461 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break; 2462 default: ShouldNotReachHere(); 2463 } 2464 } 2465 2466 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes) { 2467 switch (size_in_bytes) { 2468 case 8: str(src, dst); break; 2469 case 4: strw(src, dst); break; 2470 case 2: strh(src, dst); break; 2471 case 1: strb(src, dst); break; 2472 default: ShouldNotReachHere(); 2473 } 2474 } 2475 2476 void MacroAssembler::decrementw(Register reg, int value) 2477 { 2478 if (value < 0) { incrementw(reg, -value); return; } 2479 if (value == 0) { return; } 2480 if (value < (1 << 12)) { subw(reg, reg, value); return; } 2481 /* else */ { 2482 guarantee(reg != rscratch2, "invalid dst for register decrement"); 2483 movw(rscratch2, (unsigned)value); 2484 subw(reg, reg, rscratch2); 2485 } 2486 } 2487 2488 void MacroAssembler::decrement(Register reg, int value) 2489 { 2490 if (value < 0) { increment(reg, -value); return; } 2491 if (value == 0) { return; } 2492 if (value < (1 << 12)) { sub(reg, reg, value); return; } 2493 /* else */ { 2494 assert(reg != rscratch2, "invalid dst for register decrement"); 2495 mov(rscratch2, (uint64_t)value); 2496 sub(reg, reg, rscratch2); 2497 } 2498 } 2499 2500 void MacroAssembler::decrementw(Address dst, int value) 2501 { 2502 assert(!dst.uses(rscratch1), "invalid dst for address decrement"); 2503 if (dst.getMode() == Address::literal) { 2504 assert(abs(value) < (1 << 12), "invalid value and address mode combination"); 2505 lea(rscratch2, dst); 2506 dst = Address(rscratch2); 2507 } 2508 ldrw(rscratch1, dst); 2509 decrementw(rscratch1, value); 2510 strw(rscratch1, dst); 2511 } 2512 2513 void MacroAssembler::decrement(Address dst, int value) 2514 { 2515 assert(!dst.uses(rscratch1), "invalid address for decrement"); 2516 if (dst.getMode() == Address::literal) { 2517 assert(abs(value) < (1 << 12), "invalid value and address mode combination"); 2518 lea(rscratch2, dst); 2519 dst = Address(rscratch2); 2520 } 2521 ldr(rscratch1, dst); 2522 decrement(rscratch1, value); 2523 str(rscratch1, dst); 2524 } 2525 2526 void MacroAssembler::incrementw(Register reg, int value) 2527 { 2528 if (value < 0) { decrementw(reg, -value); return; } 2529 if (value == 0) { return; } 2530 if (value < (1 << 12)) { addw(reg, reg, value); return; } 2531 /* else */ { 2532 assert(reg != rscratch2, "invalid dst for register increment"); 2533 movw(rscratch2, (unsigned)value); 2534 addw(reg, reg, rscratch2); 2535 } 2536 } 2537 2538 void MacroAssembler::increment(Register reg, int value) 2539 { 2540 if (value < 0) { decrement(reg, -value); return; } 2541 if (value == 0) { return; } 2542 if (value < (1 << 12)) { add(reg, reg, value); return; } 2543 /* else */ { 2544 assert(reg != rscratch2, "invalid dst for register increment"); 2545 movw(rscratch2, (unsigned)value); 2546 add(reg, reg, rscratch2); 2547 } 2548 } 2549 2550 void MacroAssembler::incrementw(Address dst, int value) 2551 { 2552 assert(!dst.uses(rscratch1), "invalid dst for address increment"); 2553 if (dst.getMode() == Address::literal) { 2554 assert(abs(value) < (1 << 12), "invalid value and address mode combination"); 2555 lea(rscratch2, dst); 2556 dst = Address(rscratch2); 2557 } 2558 ldrw(rscratch1, dst); 2559 incrementw(rscratch1, value); 2560 strw(rscratch1, dst); 2561 } 2562 2563 void MacroAssembler::increment(Address dst, int value) 2564 { 2565 assert(!dst.uses(rscratch1), "invalid dst for address increment"); 2566 if (dst.getMode() == Address::literal) { 2567 assert(abs(value) < (1 << 12), "invalid value and address mode combination"); 2568 lea(rscratch2, dst); 2569 dst = Address(rscratch2); 2570 } 2571 ldr(rscratch1, dst); 2572 increment(rscratch1, value); 2573 str(rscratch1, dst); 2574 } 2575 2576 // Push lots of registers in the bit set supplied. Don't push sp. 2577 // Return the number of words pushed 2578 int MacroAssembler::push(unsigned int bitset, Register stack) { 2579 int words_pushed = 0; 2580 2581 // Scan bitset to accumulate register pairs 2582 unsigned char regs[32]; 2583 int count = 0; 2584 for (int reg = 0; reg <= 30; reg++) { 2585 if (1 & bitset) 2586 regs[count++] = reg; 2587 bitset >>= 1; 2588 } 2589 regs[count++] = zr->raw_encoding(); 2590 count &= ~1; // Only push an even number of regs 2591 2592 if (count) { 2593 stp(as_Register(regs[0]), as_Register(regs[1]), 2594 Address(pre(stack, -count * wordSize))); 2595 words_pushed += 2; 2596 } 2597 for (int i = 2; i < count; i += 2) { 2598 stp(as_Register(regs[i]), as_Register(regs[i+1]), 2599 Address(stack, i * wordSize)); 2600 words_pushed += 2; 2601 } 2602 2603 assert(words_pushed == count, "oops, pushed != count"); 2604 2605 return count; 2606 } 2607 2608 int MacroAssembler::pop(unsigned int bitset, Register stack) { 2609 int words_pushed = 0; 2610 2611 // Scan bitset to accumulate register pairs 2612 unsigned char regs[32]; 2613 int count = 0; 2614 for (int reg = 0; reg <= 30; reg++) { 2615 if (1 & bitset) 2616 regs[count++] = reg; 2617 bitset >>= 1; 2618 } 2619 regs[count++] = zr->raw_encoding(); 2620 count &= ~1; 2621 2622 for (int i = 2; i < count; i += 2) { 2623 ldp(as_Register(regs[i]), as_Register(regs[i+1]), 2624 Address(stack, i * wordSize)); 2625 words_pushed += 2; 2626 } 2627 if (count) { 2628 ldp(as_Register(regs[0]), as_Register(regs[1]), 2629 Address(post(stack, count * wordSize))); 2630 words_pushed += 2; 2631 } 2632 2633 assert(words_pushed == count, "oops, pushed != count"); 2634 2635 return count; 2636 } 2637 2638 // Push lots of registers in the bit set supplied. Don't push sp. 2639 // Return the number of dwords pushed 2640 int MacroAssembler::push_fp(unsigned int bitset, Register stack, FpPushPopMode mode) { 2641 int words_pushed = 0; 2642 bool use_sve = false; 2643 int sve_vector_size_in_bytes = 0; 2644 2645 #ifdef COMPILER2 2646 use_sve = Matcher::supports_scalable_vector(); 2647 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE); 2648 #endif 2649 2650 // Scan bitset to accumulate register pairs 2651 unsigned char regs[32]; 2652 int count = 0; 2653 for (int reg = 0; reg <= 31; reg++) { 2654 if (1 & bitset) 2655 regs[count++] = reg; 2656 bitset >>= 1; 2657 } 2658 2659 if (count == 0) { 2660 return 0; 2661 } 2662 2663 if (mode == PushPopFull) { 2664 if (use_sve && sve_vector_size_in_bytes > 16) { 2665 mode = PushPopSVE; 2666 } else { 2667 mode = PushPopNeon; 2668 } 2669 } 2670 2671 #ifndef PRODUCT 2672 { 2673 char buffer[48]; 2674 if (mode == PushPopSVE) { 2675 snprintf(buffer, sizeof(buffer), "push_fp: %d SVE registers", count); 2676 } else if (mode == PushPopNeon) { 2677 snprintf(buffer, sizeof(buffer), "push_fp: %d Neon registers", count); 2678 } else { 2679 snprintf(buffer, sizeof(buffer), "push_fp: %d fp registers", count); 2680 } 2681 block_comment(buffer); 2682 } 2683 #endif 2684 2685 if (mode == PushPopSVE) { 2686 sub(stack, stack, sve_vector_size_in_bytes * count); 2687 for (int i = 0; i < count; i++) { 2688 sve_str(as_FloatRegister(regs[i]), Address(stack, i)); 2689 } 2690 return count * sve_vector_size_in_bytes / 8; 2691 } 2692 2693 if (mode == PushPopNeon) { 2694 if (count == 1) { 2695 strq(as_FloatRegister(regs[0]), Address(pre(stack, -wordSize * 2))); 2696 return 2; 2697 } 2698 2699 bool odd = (count & 1) == 1; 2700 int push_slots = count + (odd ? 1 : 0); 2701 2702 // Always pushing full 128 bit registers. 2703 stpq(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(pre(stack, -push_slots * wordSize * 2))); 2704 words_pushed += 2; 2705 2706 for (int i = 2; i + 1 < count; i += 2) { 2707 stpq(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize * 2)); 2708 words_pushed += 2; 2709 } 2710 2711 if (odd) { 2712 strq(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize * 2)); 2713 words_pushed++; 2714 } 2715 2716 assert(words_pushed == count, "oops, pushed(%d) != count(%d)", words_pushed, count); 2717 return count * 2; 2718 } 2719 2720 if (mode == PushPopFp) { 2721 bool odd = (count & 1) == 1; 2722 int push_slots = count + (odd ? 1 : 0); 2723 2724 if (count == 1) { 2725 // Stack pointer must be 16 bytes aligned 2726 strd(as_FloatRegister(regs[0]), Address(pre(stack, -push_slots * wordSize))); 2727 return 1; 2728 } 2729 2730 stpd(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(pre(stack, -push_slots * wordSize))); 2731 words_pushed += 2; 2732 2733 for (int i = 2; i + 1 < count; i += 2) { 2734 stpd(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize)); 2735 words_pushed += 2; 2736 } 2737 2738 if (odd) { 2739 // Stack pointer must be 16 bytes aligned 2740 strd(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize)); 2741 words_pushed++; 2742 } 2743 2744 assert(words_pushed == count, "oops, pushed != count"); 2745 2746 return count; 2747 } 2748 2749 return 0; 2750 } 2751 2752 // Return the number of dwords popped 2753 int MacroAssembler::pop_fp(unsigned int bitset, Register stack, FpPushPopMode mode) { 2754 int words_pushed = 0; 2755 bool use_sve = false; 2756 int sve_vector_size_in_bytes = 0; 2757 2758 #ifdef COMPILER2 2759 use_sve = Matcher::supports_scalable_vector(); 2760 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE); 2761 #endif 2762 // Scan bitset to accumulate register pairs 2763 unsigned char regs[32]; 2764 int count = 0; 2765 for (int reg = 0; reg <= 31; reg++) { 2766 if (1 & bitset) 2767 regs[count++] = reg; 2768 bitset >>= 1; 2769 } 2770 2771 if (count == 0) { 2772 return 0; 2773 } 2774 2775 if (mode == PushPopFull) { 2776 if (use_sve && sve_vector_size_in_bytes > 16) { 2777 mode = PushPopSVE; 2778 } else { 2779 mode = PushPopNeon; 2780 } 2781 } 2782 2783 #ifndef PRODUCT 2784 { 2785 char buffer[48]; 2786 if (mode == PushPopSVE) { 2787 snprintf(buffer, sizeof(buffer), "pop_fp: %d SVE registers", count); 2788 } else if (mode == PushPopNeon) { 2789 snprintf(buffer, sizeof(buffer), "pop_fp: %d Neon registers", count); 2790 } else { 2791 snprintf(buffer, sizeof(buffer), "pop_fp: %d fp registers", count); 2792 } 2793 block_comment(buffer); 2794 } 2795 #endif 2796 2797 if (mode == PushPopSVE) { 2798 for (int i = count - 1; i >= 0; i--) { 2799 sve_ldr(as_FloatRegister(regs[i]), Address(stack, i)); 2800 } 2801 add(stack, stack, sve_vector_size_in_bytes * count); 2802 return count * sve_vector_size_in_bytes / 8; 2803 } 2804 2805 if (mode == PushPopNeon) { 2806 if (count == 1) { 2807 ldrq(as_FloatRegister(regs[0]), Address(post(stack, wordSize * 2))); 2808 return 2; 2809 } 2810 2811 bool odd = (count & 1) == 1; 2812 int push_slots = count + (odd ? 1 : 0); 2813 2814 if (odd) { 2815 ldrq(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize * 2)); 2816 words_pushed++; 2817 } 2818 2819 for (int i = 2; i + 1 < count; i += 2) { 2820 ldpq(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize * 2)); 2821 words_pushed += 2; 2822 } 2823 2824 ldpq(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(post(stack, push_slots * wordSize * 2))); 2825 words_pushed += 2; 2826 2827 assert(words_pushed == count, "oops, pushed(%d) != count(%d)", words_pushed, count); 2828 2829 return count * 2; 2830 } 2831 2832 if (mode == PushPopFp) { 2833 bool odd = (count & 1) == 1; 2834 int push_slots = count + (odd ? 1 : 0); 2835 2836 if (count == 1) { 2837 ldrd(as_FloatRegister(regs[0]), Address(post(stack, push_slots * wordSize))); 2838 return 1; 2839 } 2840 2841 if (odd) { 2842 ldrd(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize)); 2843 words_pushed++; 2844 } 2845 2846 for (int i = 2; i + 1 < count; i += 2) { 2847 ldpd(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize)); 2848 words_pushed += 2; 2849 } 2850 2851 ldpd(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(post(stack, push_slots * wordSize))); 2852 words_pushed += 2; 2853 2854 assert(words_pushed == count, "oops, pushed != count"); 2855 2856 return count; 2857 } 2858 2859 return 0; 2860 } 2861 2862 // Return the number of dwords pushed 2863 int MacroAssembler::push_p(unsigned int bitset, Register stack) { 2864 bool use_sve = false; 2865 int sve_predicate_size_in_slots = 0; 2866 2867 #ifdef COMPILER2 2868 use_sve = Matcher::supports_scalable_vector(); 2869 if (use_sve) { 2870 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots(); 2871 } 2872 #endif 2873 2874 if (!use_sve) { 2875 return 0; 2876 } 2877 2878 unsigned char regs[PRegister::number_of_registers]; 2879 int count = 0; 2880 for (int reg = 0; reg < PRegister::number_of_registers; reg++) { 2881 if (1 & bitset) 2882 regs[count++] = reg; 2883 bitset >>= 1; 2884 } 2885 2886 if (count == 0) { 2887 return 0; 2888 } 2889 2890 int total_push_bytes = align_up(sve_predicate_size_in_slots * 2891 VMRegImpl::stack_slot_size * count, 16); 2892 sub(stack, stack, total_push_bytes); 2893 for (int i = 0; i < count; i++) { 2894 sve_str(as_PRegister(regs[i]), Address(stack, i)); 2895 } 2896 return total_push_bytes / 8; 2897 } 2898 2899 // Return the number of dwords popped 2900 int MacroAssembler::pop_p(unsigned int bitset, Register stack) { 2901 bool use_sve = false; 2902 int sve_predicate_size_in_slots = 0; 2903 2904 #ifdef COMPILER2 2905 use_sve = Matcher::supports_scalable_vector(); 2906 if (use_sve) { 2907 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots(); 2908 } 2909 #endif 2910 2911 if (!use_sve) { 2912 return 0; 2913 } 2914 2915 unsigned char regs[PRegister::number_of_registers]; 2916 int count = 0; 2917 for (int reg = 0; reg < PRegister::number_of_registers; reg++) { 2918 if (1 & bitset) 2919 regs[count++] = reg; 2920 bitset >>= 1; 2921 } 2922 2923 if (count == 0) { 2924 return 0; 2925 } 2926 2927 int total_pop_bytes = align_up(sve_predicate_size_in_slots * 2928 VMRegImpl::stack_slot_size * count, 16); 2929 for (int i = count - 1; i >= 0; i--) { 2930 sve_ldr(as_PRegister(regs[i]), Address(stack, i)); 2931 } 2932 add(stack, stack, total_pop_bytes); 2933 return total_pop_bytes / 8; 2934 } 2935 2936 #ifdef ASSERT 2937 void MacroAssembler::verify_heapbase(const char* msg) { 2938 #if 0 2939 assert (UseCompressedOops || UseCompressedClassPointers, "should be compressed"); 2940 assert (Universe::heap() != nullptr, "java heap should be initialized"); 2941 if (!UseCompressedOops || Universe::ptr_base() == nullptr) { 2942 // rheapbase is allocated as general register 2943 return; 2944 } 2945 if (CheckCompressedOops) { 2946 Label ok; 2947 push(1 << rscratch1->encoding(), sp); // cmpptr trashes rscratch1 2948 cmpptr(rheapbase, ExternalAddress(CompressedOops::ptrs_base_addr())); 2949 br(Assembler::EQ, ok); 2950 stop(msg); 2951 bind(ok); 2952 pop(1 << rscratch1->encoding(), sp); 2953 } 2954 #endif 2955 } 2956 #endif 2957 2958 void MacroAssembler::resolve_jobject(Register value, Register tmp1, Register tmp2) { 2959 assert_different_registers(value, tmp1, tmp2); 2960 Label done, tagged, weak_tagged; 2961 2962 cbz(value, done); // Use null as-is. 2963 tst(value, JNIHandles::tag_mask); // Test for tag. 2964 br(Assembler::NE, tagged); 2965 2966 // Resolve local handle 2967 access_load_at(T_OBJECT, IN_NATIVE | AS_RAW, value, Address(value, 0), tmp1, tmp2); 2968 verify_oop(value); 2969 b(done); 2970 2971 bind(tagged); 2972 STATIC_ASSERT(JNIHandles::TypeTag::weak_global == 0b1); 2973 tbnz(value, 0, weak_tagged); // Test for weak tag. 2974 2975 // Resolve global handle 2976 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, -JNIHandles::TypeTag::global), tmp1, tmp2); 2977 verify_oop(value); 2978 b(done); 2979 2980 bind(weak_tagged); 2981 // Resolve jweak. 2982 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF, 2983 value, Address(value, -JNIHandles::TypeTag::weak_global), tmp1, tmp2); 2984 verify_oop(value); 2985 2986 bind(done); 2987 } 2988 2989 void MacroAssembler::resolve_global_jobject(Register value, Register tmp1, Register tmp2) { 2990 assert_different_registers(value, tmp1, tmp2); 2991 Label done; 2992 2993 cbz(value, done); // Use null as-is. 2994 2995 #ifdef ASSERT 2996 { 2997 STATIC_ASSERT(JNIHandles::TypeTag::global == 0b10); 2998 Label valid_global_tag; 2999 tbnz(value, 1, valid_global_tag); // Test for global tag 3000 stop("non global jobject using resolve_global_jobject"); 3001 bind(valid_global_tag); 3002 } 3003 #endif 3004 3005 // Resolve global handle 3006 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, -JNIHandles::TypeTag::global), tmp1, tmp2); 3007 verify_oop(value); 3008 3009 bind(done); 3010 } 3011 3012 void MacroAssembler::stop(const char* msg) { 3013 BLOCK_COMMENT(msg); 3014 dcps1(0xdeae); 3015 emit_int64((uintptr_t)msg); 3016 } 3017 3018 void MacroAssembler::unimplemented(const char* what) { 3019 const char* buf = nullptr; 3020 { 3021 ResourceMark rm; 3022 stringStream ss; 3023 ss.print("unimplemented: %s", what); 3024 buf = code_string(ss.as_string()); 3025 } 3026 stop(buf); 3027 } 3028 3029 void MacroAssembler::_assert_asm(Assembler::Condition cc, const char* msg) { 3030 #ifdef ASSERT 3031 Label OK; 3032 br(cc, OK); 3033 stop(msg); 3034 bind(OK); 3035 #endif 3036 } 3037 3038 // If a constant does not fit in an immediate field, generate some 3039 // number of MOV instructions and then perform the operation. 3040 void MacroAssembler::wrap_add_sub_imm_insn(Register Rd, Register Rn, uint64_t imm, 3041 add_sub_imm_insn insn1, 3042 add_sub_reg_insn insn2, 3043 bool is32) { 3044 assert(Rd != zr, "Rd = zr and not setting flags?"); 3045 bool fits = operand_valid_for_add_sub_immediate(is32 ? (int32_t)imm : imm); 3046 if (fits) { 3047 (this->*insn1)(Rd, Rn, imm); 3048 } else { 3049 if (uabs(imm) < (1 << 24)) { 3050 (this->*insn1)(Rd, Rn, imm & -(1 << 12)); 3051 (this->*insn1)(Rd, Rd, imm & ((1 << 12)-1)); 3052 } else { 3053 assert_different_registers(Rd, Rn); 3054 mov(Rd, imm); 3055 (this->*insn2)(Rd, Rn, Rd, LSL, 0); 3056 } 3057 } 3058 } 3059 3060 // Separate vsn which sets the flags. Optimisations are more restricted 3061 // because we must set the flags correctly. 3062 void MacroAssembler::wrap_adds_subs_imm_insn(Register Rd, Register Rn, uint64_t imm, 3063 add_sub_imm_insn insn1, 3064 add_sub_reg_insn insn2, 3065 bool is32) { 3066 bool fits = operand_valid_for_add_sub_immediate(is32 ? (int32_t)imm : imm); 3067 if (fits) { 3068 (this->*insn1)(Rd, Rn, imm); 3069 } else { 3070 assert_different_registers(Rd, Rn); 3071 assert(Rd != zr, "overflow in immediate operand"); 3072 mov(Rd, imm); 3073 (this->*insn2)(Rd, Rn, Rd, LSL, 0); 3074 } 3075 } 3076 3077 3078 void MacroAssembler::add(Register Rd, Register Rn, RegisterOrConstant increment) { 3079 if (increment.is_register()) { 3080 add(Rd, Rn, increment.as_register()); 3081 } else { 3082 add(Rd, Rn, increment.as_constant()); 3083 } 3084 } 3085 3086 void MacroAssembler::addw(Register Rd, Register Rn, RegisterOrConstant increment) { 3087 if (increment.is_register()) { 3088 addw(Rd, Rn, increment.as_register()); 3089 } else { 3090 addw(Rd, Rn, increment.as_constant()); 3091 } 3092 } 3093 3094 void MacroAssembler::sub(Register Rd, Register Rn, RegisterOrConstant decrement) { 3095 if (decrement.is_register()) { 3096 sub(Rd, Rn, decrement.as_register()); 3097 } else { 3098 sub(Rd, Rn, decrement.as_constant()); 3099 } 3100 } 3101 3102 void MacroAssembler::subw(Register Rd, Register Rn, RegisterOrConstant decrement) { 3103 if (decrement.is_register()) { 3104 subw(Rd, Rn, decrement.as_register()); 3105 } else { 3106 subw(Rd, Rn, decrement.as_constant()); 3107 } 3108 } 3109 3110 void MacroAssembler::reinit_heapbase() 3111 { 3112 if (UseCompressedOops) { 3113 if (Universe::is_fully_initialized()) { 3114 mov(rheapbase, CompressedOops::ptrs_base()); 3115 } else { 3116 lea(rheapbase, ExternalAddress(CompressedOops::ptrs_base_addr())); 3117 ldr(rheapbase, Address(rheapbase)); 3118 } 3119 } 3120 } 3121 3122 // this simulates the behaviour of the x86 cmpxchg instruction using a 3123 // load linked/store conditional pair. we use the acquire/release 3124 // versions of these instructions so that we flush pending writes as 3125 // per Java semantics. 3126 3127 // n.b the x86 version assumes the old value to be compared against is 3128 // in rax and updates rax with the value located in memory if the 3129 // cmpxchg fails. we supply a register for the old value explicitly 3130 3131 // the aarch64 load linked/store conditional instructions do not 3132 // accept an offset. so, unlike x86, we must provide a plain register 3133 // to identify the memory word to be compared/exchanged rather than a 3134 // register+offset Address. 3135 3136 void MacroAssembler::cmpxchgptr(Register oldv, Register newv, Register addr, Register tmp, 3137 Label &succeed, Label *fail) { 3138 // oldv holds comparison value 3139 // newv holds value to write in exchange 3140 // addr identifies memory word to compare against/update 3141 if (UseLSE) { 3142 mov(tmp, oldv); 3143 casal(Assembler::xword, oldv, newv, addr); 3144 cmp(tmp, oldv); 3145 br(Assembler::EQ, succeed); 3146 membar(AnyAny); 3147 } else { 3148 Label retry_load, nope; 3149 prfm(Address(addr), PSTL1STRM); 3150 bind(retry_load); 3151 // flush and load exclusive from the memory location 3152 // and fail if it is not what we expect 3153 ldaxr(tmp, addr); 3154 cmp(tmp, oldv); 3155 br(Assembler::NE, nope); 3156 // if we store+flush with no intervening write tmp will be zero 3157 stlxr(tmp, newv, addr); 3158 cbzw(tmp, succeed); 3159 // retry so we only ever return after a load fails to compare 3160 // ensures we don't return a stale value after a failed write. 3161 b(retry_load); 3162 // if the memory word differs we return it in oldv and signal a fail 3163 bind(nope); 3164 membar(AnyAny); 3165 mov(oldv, tmp); 3166 } 3167 if (fail) 3168 b(*fail); 3169 } 3170 3171 void MacroAssembler::cmpxchg_obj_header(Register oldv, Register newv, Register obj, Register tmp, 3172 Label &succeed, Label *fail) { 3173 assert(oopDesc::mark_offset_in_bytes() == 0, "assumption"); 3174 cmpxchgptr(oldv, newv, obj, tmp, succeed, fail); 3175 } 3176 3177 void MacroAssembler::cmpxchgw(Register oldv, Register newv, Register addr, Register tmp, 3178 Label &succeed, Label *fail) { 3179 // oldv holds comparison value 3180 // newv holds value to write in exchange 3181 // addr identifies memory word to compare against/update 3182 // tmp returns 0/1 for success/failure 3183 if (UseLSE) { 3184 mov(tmp, oldv); 3185 casal(Assembler::word, oldv, newv, addr); 3186 cmp(tmp, oldv); 3187 br(Assembler::EQ, succeed); 3188 membar(AnyAny); 3189 } else { 3190 Label retry_load, nope; 3191 prfm(Address(addr), PSTL1STRM); 3192 bind(retry_load); 3193 // flush and load exclusive from the memory location 3194 // and fail if it is not what we expect 3195 ldaxrw(tmp, addr); 3196 cmp(tmp, oldv); 3197 br(Assembler::NE, nope); 3198 // if we store+flush with no intervening write tmp will be zero 3199 stlxrw(tmp, newv, addr); 3200 cbzw(tmp, succeed); 3201 // retry so we only ever return after a load fails to compare 3202 // ensures we don't return a stale value after a failed write. 3203 b(retry_load); 3204 // if the memory word differs we return it in oldv and signal a fail 3205 bind(nope); 3206 membar(AnyAny); 3207 mov(oldv, tmp); 3208 } 3209 if (fail) 3210 b(*fail); 3211 } 3212 3213 // A generic CAS; success or failure is in the EQ flag. A weak CAS 3214 // doesn't retry and may fail spuriously. If the oldval is wanted, 3215 // Pass a register for the result, otherwise pass noreg. 3216 3217 // Clobbers rscratch1 3218 void MacroAssembler::cmpxchg(Register addr, Register expected, 3219 Register new_val, 3220 enum operand_size size, 3221 bool acquire, bool release, 3222 bool weak, 3223 Register result) { 3224 if (result == noreg) result = rscratch1; 3225 BLOCK_COMMENT("cmpxchg {"); 3226 if (UseLSE) { 3227 mov(result, expected); 3228 lse_cas(result, new_val, addr, size, acquire, release, /*not_pair*/ true); 3229 compare_eq(result, expected, size); 3230 #ifdef ASSERT 3231 // Poison rscratch1 which is written on !UseLSE branch 3232 mov(rscratch1, 0x1f1f1f1f1f1f1f1f); 3233 #endif 3234 } else { 3235 Label retry_load, done; 3236 prfm(Address(addr), PSTL1STRM); 3237 bind(retry_load); 3238 load_exclusive(result, addr, size, acquire); 3239 compare_eq(result, expected, size); 3240 br(Assembler::NE, done); 3241 store_exclusive(rscratch1, new_val, addr, size, release); 3242 if (weak) { 3243 cmpw(rscratch1, 0u); // If the store fails, return NE to our caller. 3244 } else { 3245 cbnzw(rscratch1, retry_load); 3246 } 3247 bind(done); 3248 } 3249 BLOCK_COMMENT("} cmpxchg"); 3250 } 3251 3252 // A generic comparison. Only compares for equality, clobbers rscratch1. 3253 void MacroAssembler::compare_eq(Register rm, Register rn, enum operand_size size) { 3254 if (size == xword) { 3255 cmp(rm, rn); 3256 } else if (size == word) { 3257 cmpw(rm, rn); 3258 } else if (size == halfword) { 3259 eorw(rscratch1, rm, rn); 3260 ands(zr, rscratch1, 0xffff); 3261 } else if (size == byte) { 3262 eorw(rscratch1, rm, rn); 3263 ands(zr, rscratch1, 0xff); 3264 } else { 3265 ShouldNotReachHere(); 3266 } 3267 } 3268 3269 3270 static bool different(Register a, RegisterOrConstant b, Register c) { 3271 if (b.is_constant()) 3272 return a != c; 3273 else 3274 return a != b.as_register() && a != c && b.as_register() != c; 3275 } 3276 3277 #define ATOMIC_OP(NAME, LDXR, OP, IOP, AOP, STXR, sz) \ 3278 void MacroAssembler::atomic_##NAME(Register prev, RegisterOrConstant incr, Register addr) { \ 3279 if (UseLSE) { \ 3280 prev = prev->is_valid() ? prev : zr; \ 3281 if (incr.is_register()) { \ 3282 AOP(sz, incr.as_register(), prev, addr); \ 3283 } else { \ 3284 mov(rscratch2, incr.as_constant()); \ 3285 AOP(sz, rscratch2, prev, addr); \ 3286 } \ 3287 return; \ 3288 } \ 3289 Register result = rscratch2; \ 3290 if (prev->is_valid()) \ 3291 result = different(prev, incr, addr) ? prev : rscratch2; \ 3292 \ 3293 Label retry_load; \ 3294 prfm(Address(addr), PSTL1STRM); \ 3295 bind(retry_load); \ 3296 LDXR(result, addr); \ 3297 OP(rscratch1, result, incr); \ 3298 STXR(rscratch2, rscratch1, addr); \ 3299 cbnzw(rscratch2, retry_load); \ 3300 if (prev->is_valid() && prev != result) { \ 3301 IOP(prev, rscratch1, incr); \ 3302 } \ 3303 } 3304 3305 ATOMIC_OP(add, ldxr, add, sub, ldadd, stxr, Assembler::xword) 3306 ATOMIC_OP(addw, ldxrw, addw, subw, ldadd, stxrw, Assembler::word) 3307 ATOMIC_OP(addal, ldaxr, add, sub, ldaddal, stlxr, Assembler::xword) 3308 ATOMIC_OP(addalw, ldaxrw, addw, subw, ldaddal, stlxrw, Assembler::word) 3309 3310 #undef ATOMIC_OP 3311 3312 #define ATOMIC_XCHG(OP, AOP, LDXR, STXR, sz) \ 3313 void MacroAssembler::atomic_##OP(Register prev, Register newv, Register addr) { \ 3314 if (UseLSE) { \ 3315 prev = prev->is_valid() ? prev : zr; \ 3316 AOP(sz, newv, prev, addr); \ 3317 return; \ 3318 } \ 3319 Register result = rscratch2; \ 3320 if (prev->is_valid()) \ 3321 result = different(prev, newv, addr) ? prev : rscratch2; \ 3322 \ 3323 Label retry_load; \ 3324 prfm(Address(addr), PSTL1STRM); \ 3325 bind(retry_load); \ 3326 LDXR(result, addr); \ 3327 STXR(rscratch1, newv, addr); \ 3328 cbnzw(rscratch1, retry_load); \ 3329 if (prev->is_valid() && prev != result) \ 3330 mov(prev, result); \ 3331 } 3332 3333 ATOMIC_XCHG(xchg, swp, ldxr, stxr, Assembler::xword) 3334 ATOMIC_XCHG(xchgw, swp, ldxrw, stxrw, Assembler::word) 3335 ATOMIC_XCHG(xchgl, swpl, ldxr, stlxr, Assembler::xword) 3336 ATOMIC_XCHG(xchglw, swpl, ldxrw, stlxrw, Assembler::word) 3337 ATOMIC_XCHG(xchgal, swpal, ldaxr, stlxr, Assembler::xword) 3338 ATOMIC_XCHG(xchgalw, swpal, ldaxrw, stlxrw, Assembler::word) 3339 3340 #undef ATOMIC_XCHG 3341 3342 #ifndef PRODUCT 3343 extern "C" void findpc(intptr_t x); 3344 #endif 3345 3346 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) 3347 { 3348 // In order to get locks to work, we need to fake a in_VM state 3349 if (ShowMessageBoxOnError ) { 3350 JavaThread* thread = JavaThread::current(); 3351 JavaThreadState saved_state = thread->thread_state(); 3352 thread->set_thread_state(_thread_in_vm); 3353 #ifndef PRODUCT 3354 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { 3355 ttyLocker ttyl; 3356 BytecodeCounter::print(); 3357 } 3358 #endif 3359 if (os::message_box(msg, "Execution stopped, print registers?")) { 3360 ttyLocker ttyl; 3361 tty->print_cr(" pc = 0x%016" PRIx64, pc); 3362 #ifndef PRODUCT 3363 tty->cr(); 3364 findpc(pc); 3365 tty->cr(); 3366 #endif 3367 tty->print_cr(" r0 = 0x%016" PRIx64, regs[0]); 3368 tty->print_cr(" r1 = 0x%016" PRIx64, regs[1]); 3369 tty->print_cr(" r2 = 0x%016" PRIx64, regs[2]); 3370 tty->print_cr(" r3 = 0x%016" PRIx64, regs[3]); 3371 tty->print_cr(" r4 = 0x%016" PRIx64, regs[4]); 3372 tty->print_cr(" r5 = 0x%016" PRIx64, regs[5]); 3373 tty->print_cr(" r6 = 0x%016" PRIx64, regs[6]); 3374 tty->print_cr(" r7 = 0x%016" PRIx64, regs[7]); 3375 tty->print_cr(" r8 = 0x%016" PRIx64, regs[8]); 3376 tty->print_cr(" r9 = 0x%016" PRIx64, regs[9]); 3377 tty->print_cr("r10 = 0x%016" PRIx64, regs[10]); 3378 tty->print_cr("r11 = 0x%016" PRIx64, regs[11]); 3379 tty->print_cr("r12 = 0x%016" PRIx64, regs[12]); 3380 tty->print_cr("r13 = 0x%016" PRIx64, regs[13]); 3381 tty->print_cr("r14 = 0x%016" PRIx64, regs[14]); 3382 tty->print_cr("r15 = 0x%016" PRIx64, regs[15]); 3383 tty->print_cr("r16 = 0x%016" PRIx64, regs[16]); 3384 tty->print_cr("r17 = 0x%016" PRIx64, regs[17]); 3385 tty->print_cr("r18 = 0x%016" PRIx64, regs[18]); 3386 tty->print_cr("r19 = 0x%016" PRIx64, regs[19]); 3387 tty->print_cr("r20 = 0x%016" PRIx64, regs[20]); 3388 tty->print_cr("r21 = 0x%016" PRIx64, regs[21]); 3389 tty->print_cr("r22 = 0x%016" PRIx64, regs[22]); 3390 tty->print_cr("r23 = 0x%016" PRIx64, regs[23]); 3391 tty->print_cr("r24 = 0x%016" PRIx64, regs[24]); 3392 tty->print_cr("r25 = 0x%016" PRIx64, regs[25]); 3393 tty->print_cr("r26 = 0x%016" PRIx64, regs[26]); 3394 tty->print_cr("r27 = 0x%016" PRIx64, regs[27]); 3395 tty->print_cr("r28 = 0x%016" PRIx64, regs[28]); 3396 tty->print_cr("r30 = 0x%016" PRIx64, regs[30]); 3397 tty->print_cr("r31 = 0x%016" PRIx64, regs[31]); 3398 BREAKPOINT; 3399 } 3400 } 3401 fatal("DEBUG MESSAGE: %s", msg); 3402 } 3403 3404 RegSet MacroAssembler::call_clobbered_gp_registers() { 3405 RegSet regs = RegSet::range(r0, r17) - RegSet::of(rscratch1, rscratch2); 3406 #ifndef R18_RESERVED 3407 regs += r18_tls; 3408 #endif 3409 return regs; 3410 } 3411 3412 void MacroAssembler::push_call_clobbered_registers_except(RegSet exclude) { 3413 int step = 4 * wordSize; 3414 push(call_clobbered_gp_registers() - exclude, sp); 3415 sub(sp, sp, step); 3416 mov(rscratch1, -step); 3417 // Push v0-v7, v16-v31. 3418 for (int i = 31; i>= 4; i -= 4) { 3419 if (i <= v7->encoding() || i >= v16->encoding()) 3420 st1(as_FloatRegister(i-3), as_FloatRegister(i-2), as_FloatRegister(i-1), 3421 as_FloatRegister(i), T1D, Address(post(sp, rscratch1))); 3422 } 3423 st1(as_FloatRegister(0), as_FloatRegister(1), as_FloatRegister(2), 3424 as_FloatRegister(3), T1D, Address(sp)); 3425 } 3426 3427 void MacroAssembler::pop_call_clobbered_registers_except(RegSet exclude) { 3428 for (int i = 0; i < 32; i += 4) { 3429 if (i <= v7->encoding() || i >= v16->encoding()) 3430 ld1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2), 3431 as_FloatRegister(i+3), T1D, Address(post(sp, 4 * wordSize))); 3432 } 3433 3434 reinitialize_ptrue(); 3435 3436 pop(call_clobbered_gp_registers() - exclude, sp); 3437 } 3438 3439 void MacroAssembler::push_CPU_state(bool save_vectors, bool use_sve, 3440 int sve_vector_size_in_bytes, int total_predicate_in_bytes) { 3441 push(RegSet::range(r0, r29), sp); // integer registers except lr & sp 3442 if (save_vectors && use_sve && sve_vector_size_in_bytes > 16) { 3443 sub(sp, sp, sve_vector_size_in_bytes * FloatRegister::number_of_registers); 3444 for (int i = 0; i < FloatRegister::number_of_registers; i++) { 3445 sve_str(as_FloatRegister(i), Address(sp, i)); 3446 } 3447 } else { 3448 int step = (save_vectors ? 8 : 4) * wordSize; 3449 mov(rscratch1, -step); 3450 sub(sp, sp, step); 3451 for (int i = 28; i >= 4; i -= 4) { 3452 st1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2), 3453 as_FloatRegister(i+3), save_vectors ? T2D : T1D, Address(post(sp, rscratch1))); 3454 } 3455 st1(v0, v1, v2, v3, save_vectors ? T2D : T1D, sp); 3456 } 3457 if (save_vectors && use_sve && total_predicate_in_bytes > 0) { 3458 sub(sp, sp, total_predicate_in_bytes); 3459 for (int i = 0; i < PRegister::number_of_registers; i++) { 3460 sve_str(as_PRegister(i), Address(sp, i)); 3461 } 3462 } 3463 } 3464 3465 void MacroAssembler::pop_CPU_state(bool restore_vectors, bool use_sve, 3466 int sve_vector_size_in_bytes, int total_predicate_in_bytes) { 3467 if (restore_vectors && use_sve && total_predicate_in_bytes > 0) { 3468 for (int i = PRegister::number_of_registers - 1; i >= 0; i--) { 3469 sve_ldr(as_PRegister(i), Address(sp, i)); 3470 } 3471 add(sp, sp, total_predicate_in_bytes); 3472 } 3473 if (restore_vectors && use_sve && sve_vector_size_in_bytes > 16) { 3474 for (int i = FloatRegister::number_of_registers - 1; i >= 0; i--) { 3475 sve_ldr(as_FloatRegister(i), Address(sp, i)); 3476 } 3477 add(sp, sp, sve_vector_size_in_bytes * FloatRegister::number_of_registers); 3478 } else { 3479 int step = (restore_vectors ? 8 : 4) * wordSize; 3480 for (int i = 0; i <= 28; i += 4) 3481 ld1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2), 3482 as_FloatRegister(i+3), restore_vectors ? T2D : T1D, Address(post(sp, step))); 3483 } 3484 3485 // We may use predicate registers and rely on ptrue with SVE, 3486 // regardless of wide vector (> 8 bytes) used or not. 3487 if (use_sve) { 3488 reinitialize_ptrue(); 3489 } 3490 3491 // integer registers except lr & sp 3492 pop(RegSet::range(r0, r17), sp); 3493 #ifdef R18_RESERVED 3494 ldp(zr, r19, Address(post(sp, 2 * wordSize))); 3495 pop(RegSet::range(r20, r29), sp); 3496 #else 3497 pop(RegSet::range(r18_tls, r29), sp); 3498 #endif 3499 } 3500 3501 /** 3502 * Helpers for multiply_to_len(). 3503 */ 3504 void MacroAssembler::add2_with_carry(Register final_dest_hi, Register dest_hi, Register dest_lo, 3505 Register src1, Register src2) { 3506 adds(dest_lo, dest_lo, src1); 3507 adc(dest_hi, dest_hi, zr); 3508 adds(dest_lo, dest_lo, src2); 3509 adc(final_dest_hi, dest_hi, zr); 3510 } 3511 3512 // Generate an address from (r + r1 extend offset). "size" is the 3513 // size of the operand. The result may be in rscratch2. 3514 Address MacroAssembler::offsetted_address(Register r, Register r1, 3515 Address::extend ext, int offset, int size) { 3516 if (offset || (ext.shift() % size != 0)) { 3517 lea(rscratch2, Address(r, r1, ext)); 3518 return Address(rscratch2, offset); 3519 } else { 3520 return Address(r, r1, ext); 3521 } 3522 } 3523 3524 Address MacroAssembler::spill_address(int size, int offset, Register tmp) 3525 { 3526 assert(offset >= 0, "spill to negative address?"); 3527 // Offset reachable ? 3528 // Not aligned - 9 bits signed offset 3529 // Aligned - 12 bits unsigned offset shifted 3530 Register base = sp; 3531 if ((offset & (size-1)) && offset >= (1<<8)) { 3532 add(tmp, base, offset & ((1<<12)-1)); 3533 base = tmp; 3534 offset &= -1u<<12; 3535 } 3536 3537 if (offset >= (1<<12) * size) { 3538 add(tmp, base, offset & (((1<<12)-1)<<12)); 3539 base = tmp; 3540 offset &= ~(((1<<12)-1)<<12); 3541 } 3542 3543 return Address(base, offset); 3544 } 3545 3546 Address MacroAssembler::sve_spill_address(int sve_reg_size_in_bytes, int offset, Register tmp) { 3547 assert(offset >= 0, "spill to negative address?"); 3548 3549 Register base = sp; 3550 3551 // An immediate offset in the range 0 to 255 which is multiplied 3552 // by the current vector or predicate register size in bytes. 3553 if (offset % sve_reg_size_in_bytes == 0 && offset < ((1<<8)*sve_reg_size_in_bytes)) { 3554 return Address(base, offset / sve_reg_size_in_bytes); 3555 } 3556 3557 add(tmp, base, offset); 3558 return Address(tmp); 3559 } 3560 3561 // Checks whether offset is aligned. 3562 // Returns true if it is, else false. 3563 bool MacroAssembler::merge_alignment_check(Register base, 3564 size_t size, 3565 int64_t cur_offset, 3566 int64_t prev_offset) const { 3567 if (AvoidUnalignedAccesses) { 3568 if (base == sp) { 3569 // Checks whether low offset if aligned to pair of registers. 3570 int64_t pair_mask = size * 2 - 1; 3571 int64_t offset = prev_offset > cur_offset ? cur_offset : prev_offset; 3572 return (offset & pair_mask) == 0; 3573 } else { // If base is not sp, we can't guarantee the access is aligned. 3574 return false; 3575 } 3576 } else { 3577 int64_t mask = size - 1; 3578 // Load/store pair instruction only supports element size aligned offset. 3579 return (cur_offset & mask) == 0 && (prev_offset & mask) == 0; 3580 } 3581 } 3582 3583 // Checks whether current and previous loads/stores can be merged. 3584 // Returns true if it can be merged, else false. 3585 bool MacroAssembler::ldst_can_merge(Register rt, 3586 const Address &adr, 3587 size_t cur_size_in_bytes, 3588 bool is_store) const { 3589 address prev = pc() - NativeInstruction::instruction_size; 3590 address last = code()->last_insn(); 3591 3592 if (last == nullptr || !nativeInstruction_at(last)->is_Imm_LdSt()) { 3593 return false; 3594 } 3595 3596 if (adr.getMode() != Address::base_plus_offset || prev != last) { 3597 return false; 3598 } 3599 3600 NativeLdSt* prev_ldst = NativeLdSt_at(prev); 3601 size_t prev_size_in_bytes = prev_ldst->size_in_bytes(); 3602 3603 assert(prev_size_in_bytes == 4 || prev_size_in_bytes == 8, "only supports 64/32bit merging."); 3604 assert(cur_size_in_bytes == 4 || cur_size_in_bytes == 8, "only supports 64/32bit merging."); 3605 3606 if (cur_size_in_bytes != prev_size_in_bytes || is_store != prev_ldst->is_store()) { 3607 return false; 3608 } 3609 3610 int64_t max_offset = 63 * prev_size_in_bytes; 3611 int64_t min_offset = -64 * prev_size_in_bytes; 3612 3613 assert(prev_ldst->is_not_pre_post_index(), "pre-index or post-index is not supported to be merged."); 3614 3615 // Only same base can be merged. 3616 if (adr.base() != prev_ldst->base()) { 3617 return false; 3618 } 3619 3620 int64_t cur_offset = adr.offset(); 3621 int64_t prev_offset = prev_ldst->offset(); 3622 size_t diff = abs(cur_offset - prev_offset); 3623 if (diff != prev_size_in_bytes) { 3624 return false; 3625 } 3626 3627 // Following cases can not be merged: 3628 // ldr x2, [x2, #8] 3629 // ldr x3, [x2, #16] 3630 // or: 3631 // ldr x2, [x3, #8] 3632 // ldr x2, [x3, #16] 3633 // If t1 and t2 is the same in "ldp t1, t2, [xn, #imm]", we'll get SIGILL. 3634 if (!is_store && (adr.base() == prev_ldst->target() || rt == prev_ldst->target())) { 3635 return false; 3636 } 3637 3638 int64_t low_offset = prev_offset > cur_offset ? cur_offset : prev_offset; 3639 // Offset range must be in ldp/stp instruction's range. 3640 if (low_offset > max_offset || low_offset < min_offset) { 3641 return false; 3642 } 3643 3644 if (merge_alignment_check(adr.base(), prev_size_in_bytes, cur_offset, prev_offset)) { 3645 return true; 3646 } 3647 3648 return false; 3649 } 3650 3651 // Merge current load/store with previous load/store into ldp/stp. 3652 void MacroAssembler::merge_ldst(Register rt, 3653 const Address &adr, 3654 size_t cur_size_in_bytes, 3655 bool is_store) { 3656 3657 assert(ldst_can_merge(rt, adr, cur_size_in_bytes, is_store) == true, "cur and prev must be able to be merged."); 3658 3659 Register rt_low, rt_high; 3660 address prev = pc() - NativeInstruction::instruction_size; 3661 NativeLdSt* prev_ldst = NativeLdSt_at(prev); 3662 3663 int64_t offset; 3664 3665 if (adr.offset() < prev_ldst->offset()) { 3666 offset = adr.offset(); 3667 rt_low = rt; 3668 rt_high = prev_ldst->target(); 3669 } else { 3670 offset = prev_ldst->offset(); 3671 rt_low = prev_ldst->target(); 3672 rt_high = rt; 3673 } 3674 3675 Address adr_p = Address(prev_ldst->base(), offset); 3676 // Overwrite previous generated binary. 3677 code_section()->set_end(prev); 3678 3679 const size_t sz = prev_ldst->size_in_bytes(); 3680 assert(sz == 8 || sz == 4, "only supports 64/32bit merging."); 3681 if (!is_store) { 3682 BLOCK_COMMENT("merged ldr pair"); 3683 if (sz == 8) { 3684 ldp(rt_low, rt_high, adr_p); 3685 } else { 3686 ldpw(rt_low, rt_high, adr_p); 3687 } 3688 } else { 3689 BLOCK_COMMENT("merged str pair"); 3690 if (sz == 8) { 3691 stp(rt_low, rt_high, adr_p); 3692 } else { 3693 stpw(rt_low, rt_high, adr_p); 3694 } 3695 } 3696 } 3697 3698 /** 3699 * Multiply 64 bit by 64 bit first loop. 3700 */ 3701 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart, 3702 Register y, Register y_idx, Register z, 3703 Register carry, Register product, 3704 Register idx, Register kdx) { 3705 // 3706 // jlong carry, x[], y[], z[]; 3707 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 3708 // huge_128 product = y[idx] * x[xstart] + carry; 3709 // z[kdx] = (jlong)product; 3710 // carry = (jlong)(product >>> 64); 3711 // } 3712 // z[xstart] = carry; 3713 // 3714 3715 Label L_first_loop, L_first_loop_exit; 3716 Label L_one_x, L_one_y, L_multiply; 3717 3718 subsw(xstart, xstart, 1); 3719 br(Assembler::MI, L_one_x); 3720 3721 lea(rscratch1, Address(x, xstart, Address::lsl(LogBytesPerInt))); 3722 ldr(x_xstart, Address(rscratch1)); 3723 ror(x_xstart, x_xstart, 32); // convert big-endian to little-endian 3724 3725 bind(L_first_loop); 3726 subsw(idx, idx, 1); 3727 br(Assembler::MI, L_first_loop_exit); 3728 subsw(idx, idx, 1); 3729 br(Assembler::MI, L_one_y); 3730 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt))); 3731 ldr(y_idx, Address(rscratch1)); 3732 ror(y_idx, y_idx, 32); // convert big-endian to little-endian 3733 bind(L_multiply); 3734 3735 // AArch64 has a multiply-accumulate instruction that we can't use 3736 // here because it has no way to process carries, so we have to use 3737 // separate add and adc instructions. Bah. 3738 umulh(rscratch1, x_xstart, y_idx); // x_xstart * y_idx -> rscratch1:product 3739 mul(product, x_xstart, y_idx); 3740 adds(product, product, carry); 3741 adc(carry, rscratch1, zr); // x_xstart * y_idx + carry -> carry:product 3742 3743 subw(kdx, kdx, 2); 3744 ror(product, product, 32); // back to big-endian 3745 str(product, offsetted_address(z, kdx, Address::uxtw(LogBytesPerInt), 0, BytesPerLong)); 3746 3747 b(L_first_loop); 3748 3749 bind(L_one_y); 3750 ldrw(y_idx, Address(y, 0)); 3751 b(L_multiply); 3752 3753 bind(L_one_x); 3754 ldrw(x_xstart, Address(x, 0)); 3755 b(L_first_loop); 3756 3757 bind(L_first_loop_exit); 3758 } 3759 3760 /** 3761 * Multiply 128 bit by 128. Unrolled inner loop. 3762 * 3763 */ 3764 void MacroAssembler::multiply_128_x_128_loop(Register y, Register z, 3765 Register carry, Register carry2, 3766 Register idx, Register jdx, 3767 Register yz_idx1, Register yz_idx2, 3768 Register tmp, Register tmp3, Register tmp4, 3769 Register tmp6, Register product_hi) { 3770 3771 // jlong carry, x[], y[], z[]; 3772 // int kdx = ystart+1; 3773 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop 3774 // huge_128 tmp3 = (y[idx+1] * product_hi) + z[kdx+idx+1] + carry; 3775 // jlong carry2 = (jlong)(tmp3 >>> 64); 3776 // huge_128 tmp4 = (y[idx] * product_hi) + z[kdx+idx] + carry2; 3777 // carry = (jlong)(tmp4 >>> 64); 3778 // z[kdx+idx+1] = (jlong)tmp3; 3779 // z[kdx+idx] = (jlong)tmp4; 3780 // } 3781 // idx += 2; 3782 // if (idx > 0) { 3783 // yz_idx1 = (y[idx] * product_hi) + z[kdx+idx] + carry; 3784 // z[kdx+idx] = (jlong)yz_idx1; 3785 // carry = (jlong)(yz_idx1 >>> 64); 3786 // } 3787 // 3788 3789 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done; 3790 3791 lsrw(jdx, idx, 2); 3792 3793 bind(L_third_loop); 3794 3795 subsw(jdx, jdx, 1); 3796 br(Assembler::MI, L_third_loop_exit); 3797 subw(idx, idx, 4); 3798 3799 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt))); 3800 3801 ldp(yz_idx2, yz_idx1, Address(rscratch1, 0)); 3802 3803 lea(tmp6, Address(z, idx, Address::uxtw(LogBytesPerInt))); 3804 3805 ror(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian 3806 ror(yz_idx2, yz_idx2, 32); 3807 3808 ldp(rscratch2, rscratch1, Address(tmp6, 0)); 3809 3810 mul(tmp3, product_hi, yz_idx1); // yz_idx1 * product_hi -> tmp4:tmp3 3811 umulh(tmp4, product_hi, yz_idx1); 3812 3813 ror(rscratch1, rscratch1, 32); // convert big-endian to little-endian 3814 ror(rscratch2, rscratch2, 32); 3815 3816 mul(tmp, product_hi, yz_idx2); // yz_idx2 * product_hi -> carry2:tmp 3817 umulh(carry2, product_hi, yz_idx2); 3818 3819 // propagate sum of both multiplications into carry:tmp4:tmp3 3820 adds(tmp3, tmp3, carry); 3821 adc(tmp4, tmp4, zr); 3822 adds(tmp3, tmp3, rscratch1); 3823 adcs(tmp4, tmp4, tmp); 3824 adc(carry, carry2, zr); 3825 adds(tmp4, tmp4, rscratch2); 3826 adc(carry, carry, zr); 3827 3828 ror(tmp3, tmp3, 32); // convert little-endian to big-endian 3829 ror(tmp4, tmp4, 32); 3830 stp(tmp4, tmp3, Address(tmp6, 0)); 3831 3832 b(L_third_loop); 3833 bind (L_third_loop_exit); 3834 3835 andw (idx, idx, 0x3); 3836 cbz(idx, L_post_third_loop_done); 3837 3838 Label L_check_1; 3839 subsw(idx, idx, 2); 3840 br(Assembler::MI, L_check_1); 3841 3842 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt))); 3843 ldr(yz_idx1, Address(rscratch1, 0)); 3844 ror(yz_idx1, yz_idx1, 32); 3845 mul(tmp3, product_hi, yz_idx1); // yz_idx1 * product_hi -> tmp4:tmp3 3846 umulh(tmp4, product_hi, yz_idx1); 3847 lea(rscratch1, Address(z, idx, Address::uxtw(LogBytesPerInt))); 3848 ldr(yz_idx2, Address(rscratch1, 0)); 3849 ror(yz_idx2, yz_idx2, 32); 3850 3851 add2_with_carry(carry, tmp4, tmp3, carry, yz_idx2); 3852 3853 ror(tmp3, tmp3, 32); 3854 str(tmp3, Address(rscratch1, 0)); 3855 3856 bind (L_check_1); 3857 3858 andw (idx, idx, 0x1); 3859 subsw(idx, idx, 1); 3860 br(Assembler::MI, L_post_third_loop_done); 3861 ldrw(tmp4, Address(y, idx, Address::uxtw(LogBytesPerInt))); 3862 mul(tmp3, tmp4, product_hi); // tmp4 * product_hi -> carry2:tmp3 3863 umulh(carry2, tmp4, product_hi); 3864 ldrw(tmp4, Address(z, idx, Address::uxtw(LogBytesPerInt))); 3865 3866 add2_with_carry(carry2, tmp3, tmp4, carry); 3867 3868 strw(tmp3, Address(z, idx, Address::uxtw(LogBytesPerInt))); 3869 extr(carry, carry2, tmp3, 32); 3870 3871 bind(L_post_third_loop_done); 3872 } 3873 3874 /** 3875 * Code for BigInteger::multiplyToLen() intrinsic. 3876 * 3877 * r0: x 3878 * r1: xlen 3879 * r2: y 3880 * r3: ylen 3881 * r4: z 3882 * r5: zlen 3883 * r10: tmp1 3884 * r11: tmp2 3885 * r12: tmp3 3886 * r13: tmp4 3887 * r14: tmp5 3888 * r15: tmp6 3889 * r16: tmp7 3890 * 3891 */ 3892 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, 3893 Register z, Register zlen, 3894 Register tmp1, Register tmp2, Register tmp3, Register tmp4, 3895 Register tmp5, Register tmp6, Register product_hi) { 3896 3897 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6); 3898 3899 const Register idx = tmp1; 3900 const Register kdx = tmp2; 3901 const Register xstart = tmp3; 3902 3903 const Register y_idx = tmp4; 3904 const Register carry = tmp5; 3905 const Register product = xlen; 3906 const Register x_xstart = zlen; // reuse register 3907 3908 // First Loop. 3909 // 3910 // final static long LONG_MASK = 0xffffffffL; 3911 // int xstart = xlen - 1; 3912 // int ystart = ylen - 1; 3913 // long carry = 0; 3914 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) { 3915 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry; 3916 // z[kdx] = (int)product; 3917 // carry = product >>> 32; 3918 // } 3919 // z[xstart] = (int)carry; 3920 // 3921 3922 movw(idx, ylen); // idx = ylen; 3923 movw(kdx, zlen); // kdx = xlen+ylen; 3924 mov(carry, zr); // carry = 0; 3925 3926 Label L_done; 3927 3928 movw(xstart, xlen); 3929 subsw(xstart, xstart, 1); 3930 br(Assembler::MI, L_done); 3931 3932 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx); 3933 3934 Label L_second_loop; 3935 cbzw(kdx, L_second_loop); 3936 3937 Label L_carry; 3938 subw(kdx, kdx, 1); 3939 cbzw(kdx, L_carry); 3940 3941 strw(carry, Address(z, kdx, Address::uxtw(LogBytesPerInt))); 3942 lsr(carry, carry, 32); 3943 subw(kdx, kdx, 1); 3944 3945 bind(L_carry); 3946 strw(carry, Address(z, kdx, Address::uxtw(LogBytesPerInt))); 3947 3948 // Second and third (nested) loops. 3949 // 3950 // for (int i = xstart-1; i >= 0; i--) { // Second loop 3951 // carry = 0; 3952 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop 3953 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) + 3954 // (z[k] & LONG_MASK) + carry; 3955 // z[k] = (int)product; 3956 // carry = product >>> 32; 3957 // } 3958 // z[i] = (int)carry; 3959 // } 3960 // 3961 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = product_hi 3962 3963 const Register jdx = tmp1; 3964 3965 bind(L_second_loop); 3966 mov(carry, zr); // carry = 0; 3967 movw(jdx, ylen); // j = ystart+1 3968 3969 subsw(xstart, xstart, 1); // i = xstart-1; 3970 br(Assembler::MI, L_done); 3971 3972 str(z, Address(pre(sp, -4 * wordSize))); 3973 3974 Label L_last_x; 3975 lea(z, offsetted_address(z, xstart, Address::uxtw(LogBytesPerInt), 4, BytesPerInt)); // z = z + k - j 3976 subsw(xstart, xstart, 1); // i = xstart-1; 3977 br(Assembler::MI, L_last_x); 3978 3979 lea(rscratch1, Address(x, xstart, Address::uxtw(LogBytesPerInt))); 3980 ldr(product_hi, Address(rscratch1)); 3981 ror(product_hi, product_hi, 32); // convert big-endian to little-endian 3982 3983 Label L_third_loop_prologue; 3984 bind(L_third_loop_prologue); 3985 3986 str(ylen, Address(sp, wordSize)); 3987 stp(x, xstart, Address(sp, 2 * wordSize)); 3988 multiply_128_x_128_loop(y, z, carry, x, jdx, ylen, product, 3989 tmp2, x_xstart, tmp3, tmp4, tmp6, product_hi); 3990 ldp(z, ylen, Address(post(sp, 2 * wordSize))); 3991 ldp(x, xlen, Address(post(sp, 2 * wordSize))); // copy old xstart -> xlen 3992 3993 addw(tmp3, xlen, 1); 3994 strw(carry, Address(z, tmp3, Address::uxtw(LogBytesPerInt))); 3995 subsw(tmp3, tmp3, 1); 3996 br(Assembler::MI, L_done); 3997 3998 lsr(carry, carry, 32); 3999 strw(carry, Address(z, tmp3, Address::uxtw(LogBytesPerInt))); 4000 b(L_second_loop); 4001 4002 // Next infrequent code is moved outside loops. 4003 bind(L_last_x); 4004 ldrw(product_hi, Address(x, 0)); 4005 b(L_third_loop_prologue); 4006 4007 bind(L_done); 4008 } 4009 4010 // Code for BigInteger::mulAdd intrinsic 4011 // out = r0 4012 // in = r1 4013 // offset = r2 (already out.length-offset) 4014 // len = r3 4015 // k = r4 4016 // 4017 // pseudo code from java implementation: 4018 // carry = 0; 4019 // offset = out.length-offset - 1; 4020 // for (int j=len-1; j >= 0; j--) { 4021 // product = (in[j] & LONG_MASK) * kLong + (out[offset] & LONG_MASK) + carry; 4022 // out[offset--] = (int)product; 4023 // carry = product >>> 32; 4024 // } 4025 // return (int)carry; 4026 void MacroAssembler::mul_add(Register out, Register in, Register offset, 4027 Register len, Register k) { 4028 Label LOOP, END; 4029 // pre-loop 4030 cmp(len, zr); // cmp, not cbz/cbnz: to use condition twice => less branches 4031 csel(out, zr, out, Assembler::EQ); 4032 br(Assembler::EQ, END); 4033 add(in, in, len, LSL, 2); // in[j+1] address 4034 add(offset, out, offset, LSL, 2); // out[offset + 1] address 4035 mov(out, zr); // used to keep carry now 4036 BIND(LOOP); 4037 ldrw(rscratch1, Address(pre(in, -4))); 4038 madd(rscratch1, rscratch1, k, out); 4039 ldrw(rscratch2, Address(pre(offset, -4))); 4040 add(rscratch1, rscratch1, rscratch2); 4041 strw(rscratch1, Address(offset)); 4042 lsr(out, rscratch1, 32); 4043 subs(len, len, 1); 4044 br(Assembler::NE, LOOP); 4045 BIND(END); 4046 } 4047 4048 /** 4049 * Emits code to update CRC-32 with a byte value according to constants in table 4050 * 4051 * @param [in,out]crc Register containing the crc. 4052 * @param [in]val Register containing the byte to fold into the CRC. 4053 * @param [in]table Register containing the table of crc constants. 4054 * 4055 * uint32_t crc; 4056 * val = crc_table[(val ^ crc) & 0xFF]; 4057 * crc = val ^ (crc >> 8); 4058 * 4059 */ 4060 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { 4061 eor(val, val, crc); 4062 andr(val, val, 0xff); 4063 ldrw(val, Address(table, val, Address::lsl(2))); 4064 eor(crc, val, crc, Assembler::LSR, 8); 4065 } 4066 4067 /** 4068 * Emits code to update CRC-32 with a 32-bit value according to tables 0 to 3 4069 * 4070 * @param [in,out]crc Register containing the crc. 4071 * @param [in]v Register containing the 32-bit to fold into the CRC. 4072 * @param [in]table0 Register containing table 0 of crc constants. 4073 * @param [in]table1 Register containing table 1 of crc constants. 4074 * @param [in]table2 Register containing table 2 of crc constants. 4075 * @param [in]table3 Register containing table 3 of crc constants. 4076 * 4077 * uint32_t crc; 4078 * v = crc ^ v 4079 * crc = table3[v&0xff]^table2[(v>>8)&0xff]^table1[(v>>16)&0xff]^table0[v>>24] 4080 * 4081 */ 4082 void MacroAssembler::update_word_crc32(Register crc, Register v, Register tmp, 4083 Register table0, Register table1, Register table2, Register table3, 4084 bool upper) { 4085 eor(v, crc, v, upper ? LSR:LSL, upper ? 32:0); 4086 uxtb(tmp, v); 4087 ldrw(crc, Address(table3, tmp, Address::lsl(2))); 4088 ubfx(tmp, v, 8, 8); 4089 ldrw(tmp, Address(table2, tmp, Address::lsl(2))); 4090 eor(crc, crc, tmp); 4091 ubfx(tmp, v, 16, 8); 4092 ldrw(tmp, Address(table1, tmp, Address::lsl(2))); 4093 eor(crc, crc, tmp); 4094 ubfx(tmp, v, 24, 8); 4095 ldrw(tmp, Address(table0, tmp, Address::lsl(2))); 4096 eor(crc, crc, tmp); 4097 } 4098 4099 void MacroAssembler::kernel_crc32_using_crypto_pmull(Register crc, Register buf, 4100 Register len, Register tmp0, Register tmp1, Register tmp2, Register tmp3) { 4101 Label CRC_by4_loop, CRC_by1_loop, CRC_less128, CRC_by128_pre, CRC_by32_loop, CRC_less32, L_exit; 4102 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2); 4103 4104 subs(tmp0, len, 384); 4105 mvnw(crc, crc); 4106 br(Assembler::GE, CRC_by128_pre); 4107 BIND(CRC_less128); 4108 subs(len, len, 32); 4109 br(Assembler::GE, CRC_by32_loop); 4110 BIND(CRC_less32); 4111 adds(len, len, 32 - 4); 4112 br(Assembler::GE, CRC_by4_loop); 4113 adds(len, len, 4); 4114 br(Assembler::GT, CRC_by1_loop); 4115 b(L_exit); 4116 4117 BIND(CRC_by32_loop); 4118 ldp(tmp0, tmp1, Address(buf)); 4119 crc32x(crc, crc, tmp0); 4120 ldp(tmp2, tmp3, Address(buf, 16)); 4121 crc32x(crc, crc, tmp1); 4122 add(buf, buf, 32); 4123 crc32x(crc, crc, tmp2); 4124 subs(len, len, 32); 4125 crc32x(crc, crc, tmp3); 4126 br(Assembler::GE, CRC_by32_loop); 4127 cmn(len, (u1)32); 4128 br(Assembler::NE, CRC_less32); 4129 b(L_exit); 4130 4131 BIND(CRC_by4_loop); 4132 ldrw(tmp0, Address(post(buf, 4))); 4133 subs(len, len, 4); 4134 crc32w(crc, crc, tmp0); 4135 br(Assembler::GE, CRC_by4_loop); 4136 adds(len, len, 4); 4137 br(Assembler::LE, L_exit); 4138 BIND(CRC_by1_loop); 4139 ldrb(tmp0, Address(post(buf, 1))); 4140 subs(len, len, 1); 4141 crc32b(crc, crc, tmp0); 4142 br(Assembler::GT, CRC_by1_loop); 4143 b(L_exit); 4144 4145 BIND(CRC_by128_pre); 4146 kernel_crc32_common_fold_using_crypto_pmull(crc, buf, len, tmp0, tmp1, tmp2, 4147 4*256*sizeof(juint) + 8*sizeof(juint)); 4148 mov(crc, 0); 4149 crc32x(crc, crc, tmp0); 4150 crc32x(crc, crc, tmp1); 4151 4152 cbnz(len, CRC_less128); 4153 4154 BIND(L_exit); 4155 mvnw(crc, crc); 4156 } 4157 4158 void MacroAssembler::kernel_crc32_using_crc32(Register crc, Register buf, 4159 Register len, Register tmp0, Register tmp1, Register tmp2, 4160 Register tmp3) { 4161 Label CRC_by64_loop, CRC_by4_loop, CRC_by1_loop, CRC_less64, CRC_by64_pre, CRC_by32_loop, CRC_less32, L_exit; 4162 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2, tmp3); 4163 4164 mvnw(crc, crc); 4165 4166 subs(len, len, 128); 4167 br(Assembler::GE, CRC_by64_pre); 4168 BIND(CRC_less64); 4169 adds(len, len, 128-32); 4170 br(Assembler::GE, CRC_by32_loop); 4171 BIND(CRC_less32); 4172 adds(len, len, 32-4); 4173 br(Assembler::GE, CRC_by4_loop); 4174 adds(len, len, 4); 4175 br(Assembler::GT, CRC_by1_loop); 4176 b(L_exit); 4177 4178 BIND(CRC_by32_loop); 4179 ldp(tmp0, tmp1, Address(post(buf, 16))); 4180 subs(len, len, 32); 4181 crc32x(crc, crc, tmp0); 4182 ldr(tmp2, Address(post(buf, 8))); 4183 crc32x(crc, crc, tmp1); 4184 ldr(tmp3, Address(post(buf, 8))); 4185 crc32x(crc, crc, tmp2); 4186 crc32x(crc, crc, tmp3); 4187 br(Assembler::GE, CRC_by32_loop); 4188 cmn(len, (u1)32); 4189 br(Assembler::NE, CRC_less32); 4190 b(L_exit); 4191 4192 BIND(CRC_by4_loop); 4193 ldrw(tmp0, Address(post(buf, 4))); 4194 subs(len, len, 4); 4195 crc32w(crc, crc, tmp0); 4196 br(Assembler::GE, CRC_by4_loop); 4197 adds(len, len, 4); 4198 br(Assembler::LE, L_exit); 4199 BIND(CRC_by1_loop); 4200 ldrb(tmp0, Address(post(buf, 1))); 4201 subs(len, len, 1); 4202 crc32b(crc, crc, tmp0); 4203 br(Assembler::GT, CRC_by1_loop); 4204 b(L_exit); 4205 4206 BIND(CRC_by64_pre); 4207 sub(buf, buf, 8); 4208 ldp(tmp0, tmp1, Address(buf, 8)); 4209 crc32x(crc, crc, tmp0); 4210 ldr(tmp2, Address(buf, 24)); 4211 crc32x(crc, crc, tmp1); 4212 ldr(tmp3, Address(buf, 32)); 4213 crc32x(crc, crc, tmp2); 4214 ldr(tmp0, Address(buf, 40)); 4215 crc32x(crc, crc, tmp3); 4216 ldr(tmp1, Address(buf, 48)); 4217 crc32x(crc, crc, tmp0); 4218 ldr(tmp2, Address(buf, 56)); 4219 crc32x(crc, crc, tmp1); 4220 ldr(tmp3, Address(pre(buf, 64))); 4221 4222 b(CRC_by64_loop); 4223 4224 align(CodeEntryAlignment); 4225 BIND(CRC_by64_loop); 4226 subs(len, len, 64); 4227 crc32x(crc, crc, tmp2); 4228 ldr(tmp0, Address(buf, 8)); 4229 crc32x(crc, crc, tmp3); 4230 ldr(tmp1, Address(buf, 16)); 4231 crc32x(crc, crc, tmp0); 4232 ldr(tmp2, Address(buf, 24)); 4233 crc32x(crc, crc, tmp1); 4234 ldr(tmp3, Address(buf, 32)); 4235 crc32x(crc, crc, tmp2); 4236 ldr(tmp0, Address(buf, 40)); 4237 crc32x(crc, crc, tmp3); 4238 ldr(tmp1, Address(buf, 48)); 4239 crc32x(crc, crc, tmp0); 4240 ldr(tmp2, Address(buf, 56)); 4241 crc32x(crc, crc, tmp1); 4242 ldr(tmp3, Address(pre(buf, 64))); 4243 br(Assembler::GE, CRC_by64_loop); 4244 4245 // post-loop 4246 crc32x(crc, crc, tmp2); 4247 crc32x(crc, crc, tmp3); 4248 4249 sub(len, len, 64); 4250 add(buf, buf, 8); 4251 cmn(len, (u1)128); 4252 br(Assembler::NE, CRC_less64); 4253 BIND(L_exit); 4254 mvnw(crc, crc); 4255 } 4256 4257 /** 4258 * @param crc register containing existing CRC (32-bit) 4259 * @param buf register pointing to input byte buffer (byte*) 4260 * @param len register containing number of bytes 4261 * @param table register that will contain address of CRC table 4262 * @param tmp scratch register 4263 */ 4264 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, 4265 Register table0, Register table1, Register table2, Register table3, 4266 Register tmp, Register tmp2, Register tmp3) { 4267 Label L_by16, L_by16_loop, L_by4, L_by4_loop, L_by1, L_by1_loop, L_exit; 4268 4269 if (UseCryptoPmullForCRC32) { 4270 kernel_crc32_using_crypto_pmull(crc, buf, len, table0, table1, table2, table3); 4271 return; 4272 } 4273 4274 if (UseCRC32) { 4275 kernel_crc32_using_crc32(crc, buf, len, table0, table1, table2, table3); 4276 return; 4277 } 4278 4279 mvnw(crc, crc); 4280 4281 { 4282 uint64_t offset; 4283 adrp(table0, ExternalAddress(StubRoutines::crc_table_addr()), offset); 4284 add(table0, table0, offset); 4285 } 4286 add(table1, table0, 1*256*sizeof(juint)); 4287 add(table2, table0, 2*256*sizeof(juint)); 4288 add(table3, table0, 3*256*sizeof(juint)); 4289 4290 { // Neon code start 4291 cmp(len, (u1)64); 4292 br(Assembler::LT, L_by16); 4293 eor(v16, T16B, v16, v16); 4294 4295 Label L_fold; 4296 4297 add(tmp, table0, 4*256*sizeof(juint)); // Point at the Neon constants 4298 4299 ld1(v0, v1, T2D, post(buf, 32)); 4300 ld1r(v4, T2D, post(tmp, 8)); 4301 ld1r(v5, T2D, post(tmp, 8)); 4302 ld1r(v6, T2D, post(tmp, 8)); 4303 ld1r(v7, T2D, post(tmp, 8)); 4304 mov(v16, S, 0, crc); 4305 4306 eor(v0, T16B, v0, v16); 4307 sub(len, len, 64); 4308 4309 BIND(L_fold); 4310 pmull(v22, T8H, v0, v5, T8B); 4311 pmull(v20, T8H, v0, v7, T8B); 4312 pmull(v23, T8H, v0, v4, T8B); 4313 pmull(v21, T8H, v0, v6, T8B); 4314 4315 pmull2(v18, T8H, v0, v5, T16B); 4316 pmull2(v16, T8H, v0, v7, T16B); 4317 pmull2(v19, T8H, v0, v4, T16B); 4318 pmull2(v17, T8H, v0, v6, T16B); 4319 4320 uzp1(v24, T8H, v20, v22); 4321 uzp2(v25, T8H, v20, v22); 4322 eor(v20, T16B, v24, v25); 4323 4324 uzp1(v26, T8H, v16, v18); 4325 uzp2(v27, T8H, v16, v18); 4326 eor(v16, T16B, v26, v27); 4327 4328 ushll2(v22, T4S, v20, T8H, 8); 4329 ushll(v20, T4S, v20, T4H, 8); 4330 4331 ushll2(v18, T4S, v16, T8H, 8); 4332 ushll(v16, T4S, v16, T4H, 8); 4333 4334 eor(v22, T16B, v23, v22); 4335 eor(v18, T16B, v19, v18); 4336 eor(v20, T16B, v21, v20); 4337 eor(v16, T16B, v17, v16); 4338 4339 uzp1(v17, T2D, v16, v20); 4340 uzp2(v21, T2D, v16, v20); 4341 eor(v17, T16B, v17, v21); 4342 4343 ushll2(v20, T2D, v17, T4S, 16); 4344 ushll(v16, T2D, v17, T2S, 16); 4345 4346 eor(v20, T16B, v20, v22); 4347 eor(v16, T16B, v16, v18); 4348 4349 uzp1(v17, T2D, v20, v16); 4350 uzp2(v21, T2D, v20, v16); 4351 eor(v28, T16B, v17, v21); 4352 4353 pmull(v22, T8H, v1, v5, T8B); 4354 pmull(v20, T8H, v1, v7, T8B); 4355 pmull(v23, T8H, v1, v4, T8B); 4356 pmull(v21, T8H, v1, v6, T8B); 4357 4358 pmull2(v18, T8H, v1, v5, T16B); 4359 pmull2(v16, T8H, v1, v7, T16B); 4360 pmull2(v19, T8H, v1, v4, T16B); 4361 pmull2(v17, T8H, v1, v6, T16B); 4362 4363 ld1(v0, v1, T2D, post(buf, 32)); 4364 4365 uzp1(v24, T8H, v20, v22); 4366 uzp2(v25, T8H, v20, v22); 4367 eor(v20, T16B, v24, v25); 4368 4369 uzp1(v26, T8H, v16, v18); 4370 uzp2(v27, T8H, v16, v18); 4371 eor(v16, T16B, v26, v27); 4372 4373 ushll2(v22, T4S, v20, T8H, 8); 4374 ushll(v20, T4S, v20, T4H, 8); 4375 4376 ushll2(v18, T4S, v16, T8H, 8); 4377 ushll(v16, T4S, v16, T4H, 8); 4378 4379 eor(v22, T16B, v23, v22); 4380 eor(v18, T16B, v19, v18); 4381 eor(v20, T16B, v21, v20); 4382 eor(v16, T16B, v17, v16); 4383 4384 uzp1(v17, T2D, v16, v20); 4385 uzp2(v21, T2D, v16, v20); 4386 eor(v16, T16B, v17, v21); 4387 4388 ushll2(v20, T2D, v16, T4S, 16); 4389 ushll(v16, T2D, v16, T2S, 16); 4390 4391 eor(v20, T16B, v22, v20); 4392 eor(v16, T16B, v16, v18); 4393 4394 uzp1(v17, T2D, v20, v16); 4395 uzp2(v21, T2D, v20, v16); 4396 eor(v20, T16B, v17, v21); 4397 4398 shl(v16, T2D, v28, 1); 4399 shl(v17, T2D, v20, 1); 4400 4401 eor(v0, T16B, v0, v16); 4402 eor(v1, T16B, v1, v17); 4403 4404 subs(len, len, 32); 4405 br(Assembler::GE, L_fold); 4406 4407 mov(crc, 0); 4408 mov(tmp, v0, D, 0); 4409 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false); 4410 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true); 4411 mov(tmp, v0, D, 1); 4412 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false); 4413 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true); 4414 mov(tmp, v1, D, 0); 4415 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false); 4416 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true); 4417 mov(tmp, v1, D, 1); 4418 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false); 4419 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true); 4420 4421 add(len, len, 32); 4422 } // Neon code end 4423 4424 BIND(L_by16); 4425 subs(len, len, 16); 4426 br(Assembler::GE, L_by16_loop); 4427 adds(len, len, 16-4); 4428 br(Assembler::GE, L_by4_loop); 4429 adds(len, len, 4); 4430 br(Assembler::GT, L_by1_loop); 4431 b(L_exit); 4432 4433 BIND(L_by4_loop); 4434 ldrw(tmp, Address(post(buf, 4))); 4435 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3); 4436 subs(len, len, 4); 4437 br(Assembler::GE, L_by4_loop); 4438 adds(len, len, 4); 4439 br(Assembler::LE, L_exit); 4440 BIND(L_by1_loop); 4441 subs(len, len, 1); 4442 ldrb(tmp, Address(post(buf, 1))); 4443 update_byte_crc32(crc, tmp, table0); 4444 br(Assembler::GT, L_by1_loop); 4445 b(L_exit); 4446 4447 align(CodeEntryAlignment); 4448 BIND(L_by16_loop); 4449 subs(len, len, 16); 4450 ldp(tmp, tmp3, Address(post(buf, 16))); 4451 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false); 4452 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true); 4453 update_word_crc32(crc, tmp3, tmp2, table0, table1, table2, table3, false); 4454 update_word_crc32(crc, tmp3, tmp2, table0, table1, table2, table3, true); 4455 br(Assembler::GE, L_by16_loop); 4456 adds(len, len, 16-4); 4457 br(Assembler::GE, L_by4_loop); 4458 adds(len, len, 4); 4459 br(Assembler::GT, L_by1_loop); 4460 BIND(L_exit); 4461 mvnw(crc, crc); 4462 } 4463 4464 void MacroAssembler::kernel_crc32c_using_crypto_pmull(Register crc, Register buf, 4465 Register len, Register tmp0, Register tmp1, Register tmp2, Register tmp3) { 4466 Label CRC_by4_loop, CRC_by1_loop, CRC_less128, CRC_by128_pre, CRC_by32_loop, CRC_less32, L_exit; 4467 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2); 4468 4469 subs(tmp0, len, 384); 4470 br(Assembler::GE, CRC_by128_pre); 4471 BIND(CRC_less128); 4472 subs(len, len, 32); 4473 br(Assembler::GE, CRC_by32_loop); 4474 BIND(CRC_less32); 4475 adds(len, len, 32 - 4); 4476 br(Assembler::GE, CRC_by4_loop); 4477 adds(len, len, 4); 4478 br(Assembler::GT, CRC_by1_loop); 4479 b(L_exit); 4480 4481 BIND(CRC_by32_loop); 4482 ldp(tmp0, tmp1, Address(buf)); 4483 crc32cx(crc, crc, tmp0); 4484 ldr(tmp2, Address(buf, 16)); 4485 crc32cx(crc, crc, tmp1); 4486 ldr(tmp3, Address(buf, 24)); 4487 crc32cx(crc, crc, tmp2); 4488 add(buf, buf, 32); 4489 subs(len, len, 32); 4490 crc32cx(crc, crc, tmp3); 4491 br(Assembler::GE, CRC_by32_loop); 4492 cmn(len, (u1)32); 4493 br(Assembler::NE, CRC_less32); 4494 b(L_exit); 4495 4496 BIND(CRC_by4_loop); 4497 ldrw(tmp0, Address(post(buf, 4))); 4498 subs(len, len, 4); 4499 crc32cw(crc, crc, tmp0); 4500 br(Assembler::GE, CRC_by4_loop); 4501 adds(len, len, 4); 4502 br(Assembler::LE, L_exit); 4503 BIND(CRC_by1_loop); 4504 ldrb(tmp0, Address(post(buf, 1))); 4505 subs(len, len, 1); 4506 crc32cb(crc, crc, tmp0); 4507 br(Assembler::GT, CRC_by1_loop); 4508 b(L_exit); 4509 4510 BIND(CRC_by128_pre); 4511 kernel_crc32_common_fold_using_crypto_pmull(crc, buf, len, tmp0, tmp1, tmp2, 4512 4*256*sizeof(juint) + 8*sizeof(juint) + 0x50); 4513 mov(crc, 0); 4514 crc32cx(crc, crc, tmp0); 4515 crc32cx(crc, crc, tmp1); 4516 4517 cbnz(len, CRC_less128); 4518 4519 BIND(L_exit); 4520 } 4521 4522 void MacroAssembler::kernel_crc32c_using_crc32c(Register crc, Register buf, 4523 Register len, Register tmp0, Register tmp1, Register tmp2, 4524 Register tmp3) { 4525 Label CRC_by64_loop, CRC_by4_loop, CRC_by1_loop, CRC_less64, CRC_by64_pre, CRC_by32_loop, CRC_less32, L_exit; 4526 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2, tmp3); 4527 4528 subs(len, len, 128); 4529 br(Assembler::GE, CRC_by64_pre); 4530 BIND(CRC_less64); 4531 adds(len, len, 128-32); 4532 br(Assembler::GE, CRC_by32_loop); 4533 BIND(CRC_less32); 4534 adds(len, len, 32-4); 4535 br(Assembler::GE, CRC_by4_loop); 4536 adds(len, len, 4); 4537 br(Assembler::GT, CRC_by1_loop); 4538 b(L_exit); 4539 4540 BIND(CRC_by32_loop); 4541 ldp(tmp0, tmp1, Address(post(buf, 16))); 4542 subs(len, len, 32); 4543 crc32cx(crc, crc, tmp0); 4544 ldr(tmp2, Address(post(buf, 8))); 4545 crc32cx(crc, crc, tmp1); 4546 ldr(tmp3, Address(post(buf, 8))); 4547 crc32cx(crc, crc, tmp2); 4548 crc32cx(crc, crc, tmp3); 4549 br(Assembler::GE, CRC_by32_loop); 4550 cmn(len, (u1)32); 4551 br(Assembler::NE, CRC_less32); 4552 b(L_exit); 4553 4554 BIND(CRC_by4_loop); 4555 ldrw(tmp0, Address(post(buf, 4))); 4556 subs(len, len, 4); 4557 crc32cw(crc, crc, tmp0); 4558 br(Assembler::GE, CRC_by4_loop); 4559 adds(len, len, 4); 4560 br(Assembler::LE, L_exit); 4561 BIND(CRC_by1_loop); 4562 ldrb(tmp0, Address(post(buf, 1))); 4563 subs(len, len, 1); 4564 crc32cb(crc, crc, tmp0); 4565 br(Assembler::GT, CRC_by1_loop); 4566 b(L_exit); 4567 4568 BIND(CRC_by64_pre); 4569 sub(buf, buf, 8); 4570 ldp(tmp0, tmp1, Address(buf, 8)); 4571 crc32cx(crc, crc, tmp0); 4572 ldr(tmp2, Address(buf, 24)); 4573 crc32cx(crc, crc, tmp1); 4574 ldr(tmp3, Address(buf, 32)); 4575 crc32cx(crc, crc, tmp2); 4576 ldr(tmp0, Address(buf, 40)); 4577 crc32cx(crc, crc, tmp3); 4578 ldr(tmp1, Address(buf, 48)); 4579 crc32cx(crc, crc, tmp0); 4580 ldr(tmp2, Address(buf, 56)); 4581 crc32cx(crc, crc, tmp1); 4582 ldr(tmp3, Address(pre(buf, 64))); 4583 4584 b(CRC_by64_loop); 4585 4586 align(CodeEntryAlignment); 4587 BIND(CRC_by64_loop); 4588 subs(len, len, 64); 4589 crc32cx(crc, crc, tmp2); 4590 ldr(tmp0, Address(buf, 8)); 4591 crc32cx(crc, crc, tmp3); 4592 ldr(tmp1, Address(buf, 16)); 4593 crc32cx(crc, crc, tmp0); 4594 ldr(tmp2, Address(buf, 24)); 4595 crc32cx(crc, crc, tmp1); 4596 ldr(tmp3, Address(buf, 32)); 4597 crc32cx(crc, crc, tmp2); 4598 ldr(tmp0, Address(buf, 40)); 4599 crc32cx(crc, crc, tmp3); 4600 ldr(tmp1, Address(buf, 48)); 4601 crc32cx(crc, crc, tmp0); 4602 ldr(tmp2, Address(buf, 56)); 4603 crc32cx(crc, crc, tmp1); 4604 ldr(tmp3, Address(pre(buf, 64))); 4605 br(Assembler::GE, CRC_by64_loop); 4606 4607 // post-loop 4608 crc32cx(crc, crc, tmp2); 4609 crc32cx(crc, crc, tmp3); 4610 4611 sub(len, len, 64); 4612 add(buf, buf, 8); 4613 cmn(len, (u1)128); 4614 br(Assembler::NE, CRC_less64); 4615 BIND(L_exit); 4616 } 4617 4618 /** 4619 * @param crc register containing existing CRC (32-bit) 4620 * @param buf register pointing to input byte buffer (byte*) 4621 * @param len register containing number of bytes 4622 * @param table register that will contain address of CRC table 4623 * @param tmp scratch register 4624 */ 4625 void MacroAssembler::kernel_crc32c(Register crc, Register buf, Register len, 4626 Register table0, Register table1, Register table2, Register table3, 4627 Register tmp, Register tmp2, Register tmp3) { 4628 if (UseCryptoPmullForCRC32) { 4629 kernel_crc32c_using_crypto_pmull(crc, buf, len, table0, table1, table2, table3); 4630 } else { 4631 kernel_crc32c_using_crc32c(crc, buf, len, table0, table1, table2, table3); 4632 } 4633 } 4634 4635 void MacroAssembler::kernel_crc32_common_fold_using_crypto_pmull(Register crc, Register buf, 4636 Register len, Register tmp0, Register tmp1, Register tmp2, size_t table_offset) { 4637 Label CRC_by128_loop; 4638 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2); 4639 4640 sub(len, len, 256); 4641 Register table = tmp0; 4642 { 4643 uint64_t offset; 4644 adrp(table, ExternalAddress(StubRoutines::crc_table_addr()), offset); 4645 add(table, table, offset); 4646 } 4647 add(table, table, table_offset); 4648 4649 // Registers v0..v7 are used as data registers. 4650 // Registers v16..v31 are used as tmp registers. 4651 sub(buf, buf, 0x10); 4652 ldrq(v0, Address(buf, 0x10)); 4653 ldrq(v1, Address(buf, 0x20)); 4654 ldrq(v2, Address(buf, 0x30)); 4655 ldrq(v3, Address(buf, 0x40)); 4656 ldrq(v4, Address(buf, 0x50)); 4657 ldrq(v5, Address(buf, 0x60)); 4658 ldrq(v6, Address(buf, 0x70)); 4659 ldrq(v7, Address(pre(buf, 0x80))); 4660 4661 movi(v31, T4S, 0); 4662 mov(v31, S, 0, crc); 4663 eor(v0, T16B, v0, v31); 4664 4665 // Register v16 contains constants from the crc table. 4666 ldrq(v16, Address(table)); 4667 b(CRC_by128_loop); 4668 4669 align(OptoLoopAlignment); 4670 BIND(CRC_by128_loop); 4671 pmull (v17, T1Q, v0, v16, T1D); 4672 pmull2(v18, T1Q, v0, v16, T2D); 4673 ldrq(v0, Address(buf, 0x10)); 4674 eor3(v0, T16B, v17, v18, v0); 4675 4676 pmull (v19, T1Q, v1, v16, T1D); 4677 pmull2(v20, T1Q, v1, v16, T2D); 4678 ldrq(v1, Address(buf, 0x20)); 4679 eor3(v1, T16B, v19, v20, v1); 4680 4681 pmull (v21, T1Q, v2, v16, T1D); 4682 pmull2(v22, T1Q, v2, v16, T2D); 4683 ldrq(v2, Address(buf, 0x30)); 4684 eor3(v2, T16B, v21, v22, v2); 4685 4686 pmull (v23, T1Q, v3, v16, T1D); 4687 pmull2(v24, T1Q, v3, v16, T2D); 4688 ldrq(v3, Address(buf, 0x40)); 4689 eor3(v3, T16B, v23, v24, v3); 4690 4691 pmull (v25, T1Q, v4, v16, T1D); 4692 pmull2(v26, T1Q, v4, v16, T2D); 4693 ldrq(v4, Address(buf, 0x50)); 4694 eor3(v4, T16B, v25, v26, v4); 4695 4696 pmull (v27, T1Q, v5, v16, T1D); 4697 pmull2(v28, T1Q, v5, v16, T2D); 4698 ldrq(v5, Address(buf, 0x60)); 4699 eor3(v5, T16B, v27, v28, v5); 4700 4701 pmull (v29, T1Q, v6, v16, T1D); 4702 pmull2(v30, T1Q, v6, v16, T2D); 4703 ldrq(v6, Address(buf, 0x70)); 4704 eor3(v6, T16B, v29, v30, v6); 4705 4706 // Reuse registers v23, v24. 4707 // Using them won't block the first instruction of the next iteration. 4708 pmull (v23, T1Q, v7, v16, T1D); 4709 pmull2(v24, T1Q, v7, v16, T2D); 4710 ldrq(v7, Address(pre(buf, 0x80))); 4711 eor3(v7, T16B, v23, v24, v7); 4712 4713 subs(len, len, 0x80); 4714 br(Assembler::GE, CRC_by128_loop); 4715 4716 // fold into 512 bits 4717 // Use v31 for constants because v16 can be still in use. 4718 ldrq(v31, Address(table, 0x10)); 4719 4720 pmull (v17, T1Q, v0, v31, T1D); 4721 pmull2(v18, T1Q, v0, v31, T2D); 4722 eor3(v0, T16B, v17, v18, v4); 4723 4724 pmull (v19, T1Q, v1, v31, T1D); 4725 pmull2(v20, T1Q, v1, v31, T2D); 4726 eor3(v1, T16B, v19, v20, v5); 4727 4728 pmull (v21, T1Q, v2, v31, T1D); 4729 pmull2(v22, T1Q, v2, v31, T2D); 4730 eor3(v2, T16B, v21, v22, v6); 4731 4732 pmull (v23, T1Q, v3, v31, T1D); 4733 pmull2(v24, T1Q, v3, v31, T2D); 4734 eor3(v3, T16B, v23, v24, v7); 4735 4736 // fold into 128 bits 4737 // Use v17 for constants because v31 can be still in use. 4738 ldrq(v17, Address(table, 0x20)); 4739 pmull (v25, T1Q, v0, v17, T1D); 4740 pmull2(v26, T1Q, v0, v17, T2D); 4741 eor3(v3, T16B, v3, v25, v26); 4742 4743 // Use v18 for constants because v17 can be still in use. 4744 ldrq(v18, Address(table, 0x30)); 4745 pmull (v27, T1Q, v1, v18, T1D); 4746 pmull2(v28, T1Q, v1, v18, T2D); 4747 eor3(v3, T16B, v3, v27, v28); 4748 4749 // Use v19 for constants because v18 can be still in use. 4750 ldrq(v19, Address(table, 0x40)); 4751 pmull (v29, T1Q, v2, v19, T1D); 4752 pmull2(v30, T1Q, v2, v19, T2D); 4753 eor3(v0, T16B, v3, v29, v30); 4754 4755 add(len, len, 0x80); 4756 add(buf, buf, 0x10); 4757 4758 mov(tmp0, v0, D, 0); 4759 mov(tmp1, v0, D, 1); 4760 } 4761 4762 SkipIfEqual::SkipIfEqual( 4763 MacroAssembler* masm, const bool* flag_addr, bool value) { 4764 _masm = masm; 4765 uint64_t offset; 4766 _masm->adrp(rscratch1, ExternalAddress((address)flag_addr), offset); 4767 _masm->ldrb(rscratch1, Address(rscratch1, offset)); 4768 if (value) { 4769 _masm->cbnzw(rscratch1, _label); 4770 } else { 4771 _masm->cbzw(rscratch1, _label); 4772 } 4773 } 4774 4775 SkipIfEqual::~SkipIfEqual() { 4776 _masm->bind(_label); 4777 } 4778 4779 void MacroAssembler::addptr(const Address &dst, int32_t src) { 4780 Address adr; 4781 switch(dst.getMode()) { 4782 case Address::base_plus_offset: 4783 // This is the expected mode, although we allow all the other 4784 // forms below. 4785 adr = form_address(rscratch2, dst.base(), dst.offset(), LogBytesPerWord); 4786 break; 4787 default: 4788 lea(rscratch2, dst); 4789 adr = Address(rscratch2); 4790 break; 4791 } 4792 ldr(rscratch1, adr); 4793 add(rscratch1, rscratch1, src); 4794 str(rscratch1, adr); 4795 } 4796 4797 void MacroAssembler::cmpptr(Register src1, Address src2) { 4798 uint64_t offset; 4799 adrp(rscratch1, src2, offset); 4800 ldr(rscratch1, Address(rscratch1, offset)); 4801 cmp(src1, rscratch1); 4802 } 4803 4804 void MacroAssembler::cmpoop(Register obj1, Register obj2) { 4805 cmp(obj1, obj2); 4806 } 4807 4808 void MacroAssembler::load_method_holder_cld(Register rresult, Register rmethod) { 4809 load_method_holder(rresult, rmethod); 4810 ldr(rresult, Address(rresult, InstanceKlass::class_loader_data_offset())); 4811 } 4812 4813 void MacroAssembler::load_method_holder(Register holder, Register method) { 4814 ldr(holder, Address(method, Method::const_offset())); // ConstMethod* 4815 ldr(holder, Address(holder, ConstMethod::constants_offset())); // ConstantPool* 4816 ldr(holder, Address(holder, ConstantPool::pool_holder_offset())); // InstanceKlass* 4817 } 4818 4819 void MacroAssembler::load_klass(Register dst, Register src) { 4820 if (UseCompressedClassPointers) { 4821 ldrw(dst, Address(src, oopDesc::klass_offset_in_bytes())); 4822 decode_klass_not_null(dst); 4823 } else { 4824 ldr(dst, Address(src, oopDesc::klass_offset_in_bytes())); 4825 } 4826 } 4827 4828 void MacroAssembler::restore_cpu_control_state_after_jni(Register tmp1, Register tmp2) { 4829 if (RestoreMXCSROnJNICalls) { 4830 Label OK; 4831 get_fpcr(tmp1); 4832 mov(tmp2, tmp1); 4833 // Set FPCR to the state we need. We do want Round to Nearest. We 4834 // don't want non-IEEE rounding modes or floating-point traps. 4835 bfi(tmp1, zr, 22, 4); // Clear DN, FZ, and Rmode 4836 bfi(tmp1, zr, 8, 5); // Clear exception-control bits (8-12) 4837 bfi(tmp1, zr, 0, 2); // Clear AH:FIZ 4838 eor(tmp2, tmp1, tmp2); 4839 cbz(tmp2, OK); // Only reset FPCR if it's wrong 4840 set_fpcr(tmp1); 4841 bind(OK); 4842 } 4843 } 4844 4845 // ((OopHandle)result).resolve(); 4846 void MacroAssembler::resolve_oop_handle(Register result, Register tmp1, Register tmp2) { 4847 // OopHandle::resolve is an indirection. 4848 access_load_at(T_OBJECT, IN_NATIVE, result, Address(result, 0), tmp1, tmp2); 4849 } 4850 4851 // ((WeakHandle)result).resolve(); 4852 void MacroAssembler::resolve_weak_handle(Register result, Register tmp1, Register tmp2) { 4853 assert_different_registers(result, tmp1, tmp2); 4854 Label resolved; 4855 4856 // A null weak handle resolves to null. 4857 cbz(result, resolved); 4858 4859 // Only 64 bit platforms support GCs that require a tmp register 4860 // WeakHandle::resolve is an indirection like jweak. 4861 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF, 4862 result, Address(result), tmp1, tmp2); 4863 bind(resolved); 4864 } 4865 4866 void MacroAssembler::load_mirror(Register dst, Register method, Register tmp1, Register tmp2) { 4867 const int mirror_offset = in_bytes(Klass::java_mirror_offset()); 4868 ldr(dst, Address(rmethod, Method::const_offset())); 4869 ldr(dst, Address(dst, ConstMethod::constants_offset())); 4870 ldr(dst, Address(dst, ConstantPool::pool_holder_offset())); 4871 ldr(dst, Address(dst, mirror_offset)); 4872 resolve_oop_handle(dst, tmp1, tmp2); 4873 } 4874 4875 void MacroAssembler::cmp_klass(Register oop, Register trial_klass, Register tmp) { 4876 if (UseCompressedClassPointers) { 4877 ldrw(tmp, Address(oop, oopDesc::klass_offset_in_bytes())); 4878 if (CompressedKlassPointers::base() == nullptr) { 4879 cmp(trial_klass, tmp, LSL, CompressedKlassPointers::shift()); 4880 return; 4881 } else if (((uint64_t)CompressedKlassPointers::base() & 0xffffffff) == 0 4882 && CompressedKlassPointers::shift() == 0) { 4883 // Only the bottom 32 bits matter 4884 cmpw(trial_klass, tmp); 4885 return; 4886 } 4887 decode_klass_not_null(tmp); 4888 } else { 4889 ldr(tmp, Address(oop, oopDesc::klass_offset_in_bytes())); 4890 } 4891 cmp(trial_klass, tmp); 4892 } 4893 4894 void MacroAssembler::store_klass(Register dst, Register src) { 4895 // FIXME: Should this be a store release? concurrent gcs assumes 4896 // klass length is valid if klass field is not null. 4897 if (UseCompressedClassPointers) { 4898 encode_klass_not_null(src); 4899 strw(src, Address(dst, oopDesc::klass_offset_in_bytes())); 4900 } else { 4901 str(src, Address(dst, oopDesc::klass_offset_in_bytes())); 4902 } 4903 } 4904 4905 void MacroAssembler::store_klass_gap(Register dst, Register src) { 4906 if (UseCompressedClassPointers) { 4907 // Store to klass gap in destination 4908 strw(src, Address(dst, oopDesc::klass_gap_offset_in_bytes())); 4909 } 4910 } 4911 4912 // Algorithm must match CompressedOops::encode. 4913 void MacroAssembler::encode_heap_oop(Register d, Register s) { 4914 #ifdef ASSERT 4915 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?"); 4916 #endif 4917 verify_oop_msg(s, "broken oop in encode_heap_oop"); 4918 if (CompressedOops::base() == nullptr) { 4919 if (CompressedOops::shift() != 0) { 4920 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 4921 lsr(d, s, LogMinObjAlignmentInBytes); 4922 } else { 4923 mov(d, s); 4924 } 4925 } else { 4926 subs(d, s, rheapbase); 4927 csel(d, d, zr, Assembler::HS); 4928 lsr(d, d, LogMinObjAlignmentInBytes); 4929 4930 /* Old algorithm: is this any worse? 4931 Label nonnull; 4932 cbnz(r, nonnull); 4933 sub(r, r, rheapbase); 4934 bind(nonnull); 4935 lsr(r, r, LogMinObjAlignmentInBytes); 4936 */ 4937 } 4938 } 4939 4940 void MacroAssembler::encode_heap_oop_not_null(Register r) { 4941 #ifdef ASSERT 4942 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?"); 4943 if (CheckCompressedOops) { 4944 Label ok; 4945 cbnz(r, ok); 4946 stop("null oop passed to encode_heap_oop_not_null"); 4947 bind(ok); 4948 } 4949 #endif 4950 verify_oop_msg(r, "broken oop in encode_heap_oop_not_null"); 4951 if (CompressedOops::base() != nullptr) { 4952 sub(r, r, rheapbase); 4953 } 4954 if (CompressedOops::shift() != 0) { 4955 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 4956 lsr(r, r, LogMinObjAlignmentInBytes); 4957 } 4958 } 4959 4960 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) { 4961 #ifdef ASSERT 4962 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?"); 4963 if (CheckCompressedOops) { 4964 Label ok; 4965 cbnz(src, ok); 4966 stop("null oop passed to encode_heap_oop_not_null2"); 4967 bind(ok); 4968 } 4969 #endif 4970 verify_oop_msg(src, "broken oop in encode_heap_oop_not_null2"); 4971 4972 Register data = src; 4973 if (CompressedOops::base() != nullptr) { 4974 sub(dst, src, rheapbase); 4975 data = dst; 4976 } 4977 if (CompressedOops::shift() != 0) { 4978 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 4979 lsr(dst, data, LogMinObjAlignmentInBytes); 4980 data = dst; 4981 } 4982 if (data == src) 4983 mov(dst, src); 4984 } 4985 4986 void MacroAssembler::decode_heap_oop(Register d, Register s) { 4987 #ifdef ASSERT 4988 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?"); 4989 #endif 4990 if (CompressedOops::base() == nullptr) { 4991 if (CompressedOops::shift() != 0 || d != s) { 4992 lsl(d, s, CompressedOops::shift()); 4993 } 4994 } else { 4995 Label done; 4996 if (d != s) 4997 mov(d, s); 4998 cbz(s, done); 4999 add(d, rheapbase, s, Assembler::LSL, LogMinObjAlignmentInBytes); 5000 bind(done); 5001 } 5002 verify_oop_msg(d, "broken oop in decode_heap_oop"); 5003 } 5004 5005 void MacroAssembler::decode_heap_oop_not_null(Register r) { 5006 assert (UseCompressedOops, "should only be used for compressed headers"); 5007 assert (Universe::heap() != nullptr, "java heap should be initialized"); 5008 // Cannot assert, unverified entry point counts instructions (see .ad file) 5009 // vtableStubs also counts instructions in pd_code_size_limit. 5010 // Also do not verify_oop as this is called by verify_oop. 5011 if (CompressedOops::shift() != 0) { 5012 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 5013 if (CompressedOops::base() != nullptr) { 5014 add(r, rheapbase, r, Assembler::LSL, LogMinObjAlignmentInBytes); 5015 } else { 5016 add(r, zr, r, Assembler::LSL, LogMinObjAlignmentInBytes); 5017 } 5018 } else { 5019 assert (CompressedOops::base() == nullptr, "sanity"); 5020 } 5021 } 5022 5023 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) { 5024 assert (UseCompressedOops, "should only be used for compressed headers"); 5025 assert (Universe::heap() != nullptr, "java heap should be initialized"); 5026 // Cannot assert, unverified entry point counts instructions (see .ad file) 5027 // vtableStubs also counts instructions in pd_code_size_limit. 5028 // Also do not verify_oop as this is called by verify_oop. 5029 if (CompressedOops::shift() != 0) { 5030 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong"); 5031 if (CompressedOops::base() != nullptr) { 5032 add(dst, rheapbase, src, Assembler::LSL, LogMinObjAlignmentInBytes); 5033 } else { 5034 add(dst, zr, src, Assembler::LSL, LogMinObjAlignmentInBytes); 5035 } 5036 } else { 5037 assert (CompressedOops::base() == nullptr, "sanity"); 5038 if (dst != src) { 5039 mov(dst, src); 5040 } 5041 } 5042 } 5043 5044 MacroAssembler::KlassDecodeMode MacroAssembler::_klass_decode_mode(KlassDecodeNone); 5045 5046 MacroAssembler::KlassDecodeMode MacroAssembler::klass_decode_mode() { 5047 assert(UseCompressedClassPointers, "not using compressed class pointers"); 5048 assert(Metaspace::initialized(), "metaspace not initialized yet"); 5049 5050 if (_klass_decode_mode != KlassDecodeNone) { 5051 return _klass_decode_mode; 5052 } 5053 5054 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift() 5055 || 0 == CompressedKlassPointers::shift(), "decode alg wrong"); 5056 5057 if (CompressedKlassPointers::base() == nullptr) { 5058 return (_klass_decode_mode = KlassDecodeZero); 5059 } 5060 5061 if (operand_valid_for_logical_immediate( 5062 /*is32*/false, (uint64_t)CompressedKlassPointers::base())) { 5063 const uint64_t range_mask = 5064 (1ULL << log2i(CompressedKlassPointers::range())) - 1; 5065 if (((uint64_t)CompressedKlassPointers::base() & range_mask) == 0) { 5066 return (_klass_decode_mode = KlassDecodeXor); 5067 } 5068 } 5069 5070 const uint64_t shifted_base = 5071 (uint64_t)CompressedKlassPointers::base() >> CompressedKlassPointers::shift(); 5072 guarantee((shifted_base & 0xffff0000ffffffff) == 0, 5073 "compressed class base bad alignment"); 5074 5075 return (_klass_decode_mode = KlassDecodeMovk); 5076 } 5077 5078 void MacroAssembler::encode_klass_not_null(Register dst, Register src) { 5079 switch (klass_decode_mode()) { 5080 case KlassDecodeZero: 5081 if (CompressedKlassPointers::shift() != 0) { 5082 lsr(dst, src, LogKlassAlignmentInBytes); 5083 } else { 5084 if (dst != src) mov(dst, src); 5085 } 5086 break; 5087 5088 case KlassDecodeXor: 5089 if (CompressedKlassPointers::shift() != 0) { 5090 eor(dst, src, (uint64_t)CompressedKlassPointers::base()); 5091 lsr(dst, dst, LogKlassAlignmentInBytes); 5092 } else { 5093 eor(dst, src, (uint64_t)CompressedKlassPointers::base()); 5094 } 5095 break; 5096 5097 case KlassDecodeMovk: 5098 if (CompressedKlassPointers::shift() != 0) { 5099 ubfx(dst, src, LogKlassAlignmentInBytes, 32); 5100 } else { 5101 movw(dst, src); 5102 } 5103 break; 5104 5105 case KlassDecodeNone: 5106 ShouldNotReachHere(); 5107 break; 5108 } 5109 } 5110 5111 void MacroAssembler::encode_klass_not_null(Register r) { 5112 encode_klass_not_null(r, r); 5113 } 5114 5115 void MacroAssembler::decode_klass_not_null(Register dst, Register src) { 5116 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5117 5118 switch (klass_decode_mode()) { 5119 case KlassDecodeZero: 5120 if (CompressedKlassPointers::shift() != 0) { 5121 lsl(dst, src, LogKlassAlignmentInBytes); 5122 } else { 5123 if (dst != src) mov(dst, src); 5124 } 5125 break; 5126 5127 case KlassDecodeXor: 5128 if (CompressedKlassPointers::shift() != 0) { 5129 lsl(dst, src, LogKlassAlignmentInBytes); 5130 eor(dst, dst, (uint64_t)CompressedKlassPointers::base()); 5131 } else { 5132 eor(dst, src, (uint64_t)CompressedKlassPointers::base()); 5133 } 5134 break; 5135 5136 case KlassDecodeMovk: { 5137 const uint64_t shifted_base = 5138 (uint64_t)CompressedKlassPointers::base() >> CompressedKlassPointers::shift(); 5139 5140 if (dst != src) movw(dst, src); 5141 movk(dst, shifted_base >> 32, 32); 5142 5143 if (CompressedKlassPointers::shift() != 0) { 5144 lsl(dst, dst, LogKlassAlignmentInBytes); 5145 } 5146 5147 break; 5148 } 5149 5150 case KlassDecodeNone: 5151 ShouldNotReachHere(); 5152 break; 5153 } 5154 } 5155 5156 void MacroAssembler::decode_klass_not_null(Register r) { 5157 decode_klass_not_null(r, r); 5158 } 5159 5160 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) { 5161 #ifdef ASSERT 5162 { 5163 ThreadInVMfromUnknown tiv; 5164 assert (UseCompressedOops, "should only be used for compressed oops"); 5165 assert (Universe::heap() != nullptr, "java heap should be initialized"); 5166 assert (oop_recorder() != nullptr, "this assembler needs an OopRecorder"); 5167 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "should be real oop"); 5168 } 5169 #endif 5170 int oop_index = oop_recorder()->find_index(obj); 5171 InstructionMark im(this); 5172 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5173 code_section()->relocate(inst_mark(), rspec); 5174 movz(dst, 0xDEAD, 16); 5175 movk(dst, 0xBEEF); 5176 } 5177 5178 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) { 5179 assert (UseCompressedClassPointers, "should only be used for compressed headers"); 5180 assert (oop_recorder() != nullptr, "this assembler needs an OopRecorder"); 5181 int index = oop_recorder()->find_index(k); 5182 assert(! Universe::heap()->is_in(k), "should not be an oop"); 5183 5184 InstructionMark im(this); 5185 RelocationHolder rspec = metadata_Relocation::spec(index); 5186 code_section()->relocate(inst_mark(), rspec); 5187 narrowKlass nk = CompressedKlassPointers::encode(k); 5188 movz(dst, (nk >> 16), 16); 5189 movk(dst, nk & 0xffff); 5190 } 5191 5192 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators, 5193 Register dst, Address src, 5194 Register tmp1, Register tmp2) { 5195 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5196 decorators = AccessInternal::decorator_fixup(decorators, type); 5197 bool as_raw = (decorators & AS_RAW) != 0; 5198 if (as_raw) { 5199 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, tmp2); 5200 } else { 5201 bs->load_at(this, decorators, type, dst, src, tmp1, tmp2); 5202 } 5203 } 5204 5205 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators, 5206 Address dst, Register val, 5207 Register tmp1, Register tmp2, Register tmp3) { 5208 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5209 decorators = AccessInternal::decorator_fixup(decorators, type); 5210 bool as_raw = (decorators & AS_RAW) != 0; 5211 if (as_raw) { 5212 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, val, tmp1, tmp2, tmp3); 5213 } else { 5214 bs->store_at(this, decorators, type, dst, val, tmp1, tmp2, tmp3); 5215 } 5216 } 5217 5218 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1, 5219 Register tmp2, DecoratorSet decorators) { 5220 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2); 5221 } 5222 5223 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1, 5224 Register tmp2, DecoratorSet decorators) { 5225 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, tmp2); 5226 } 5227 5228 void MacroAssembler::store_heap_oop(Address dst, Register val, Register tmp1, 5229 Register tmp2, Register tmp3, DecoratorSet decorators) { 5230 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, val, tmp1, tmp2, tmp3); 5231 } 5232 5233 // Used for storing nulls. 5234 void MacroAssembler::store_heap_oop_null(Address dst) { 5235 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg, noreg); 5236 } 5237 5238 Address MacroAssembler::allocate_metadata_address(Metadata* obj) { 5239 assert(oop_recorder() != nullptr, "this assembler needs a Recorder"); 5240 int index = oop_recorder()->allocate_metadata_index(obj); 5241 RelocationHolder rspec = metadata_Relocation::spec(index); 5242 return Address((address)obj, rspec); 5243 } 5244 5245 // Move an oop into a register. 5246 void MacroAssembler::movoop(Register dst, jobject obj) { 5247 int oop_index; 5248 if (obj == nullptr) { 5249 oop_index = oop_recorder()->allocate_oop_index(obj); 5250 } else { 5251 #ifdef ASSERT 5252 { 5253 ThreadInVMfromUnknown tiv; 5254 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "should be real oop"); 5255 } 5256 #endif 5257 oop_index = oop_recorder()->find_index(obj); 5258 } 5259 RelocationHolder rspec = oop_Relocation::spec(oop_index); 5260 5261 if (BarrierSet::barrier_set()->barrier_set_assembler()->supports_instruction_patching()) { 5262 mov(dst, Address((address)obj, rspec)); 5263 } else { 5264 address dummy = address(uintptr_t(pc()) & -wordSize); // A nearby aligned address 5265 ldr_constant(dst, Address(dummy, rspec)); 5266 } 5267 5268 } 5269 5270 // Move a metadata address into a register. 5271 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { 5272 int oop_index; 5273 if (obj == nullptr) { 5274 oop_index = oop_recorder()->allocate_metadata_index(obj); 5275 } else { 5276 oop_index = oop_recorder()->find_index(obj); 5277 } 5278 RelocationHolder rspec = metadata_Relocation::spec(oop_index); 5279 mov(dst, Address((address)obj, rspec)); 5280 } 5281 5282 Address MacroAssembler::constant_oop_address(jobject obj) { 5283 #ifdef ASSERT 5284 { 5285 ThreadInVMfromUnknown tiv; 5286 assert(oop_recorder() != nullptr, "this assembler needs an OopRecorder"); 5287 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "not an oop"); 5288 } 5289 #endif 5290 int oop_index = oop_recorder()->find_index(obj); 5291 return Address((address)obj, oop_Relocation::spec(oop_index)); 5292 } 5293 5294 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes. 5295 void MacroAssembler::tlab_allocate(Register obj, 5296 Register var_size_in_bytes, 5297 int con_size_in_bytes, 5298 Register t1, 5299 Register t2, 5300 Label& slow_case) { 5301 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler(); 5302 bs->tlab_allocate(this, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case); 5303 } 5304 5305 void MacroAssembler::verify_tlab() { 5306 #ifdef ASSERT 5307 if (UseTLAB && VerifyOops) { 5308 Label next, ok; 5309 5310 stp(rscratch2, rscratch1, Address(pre(sp, -16))); 5311 5312 ldr(rscratch2, Address(rthread, in_bytes(JavaThread::tlab_top_offset()))); 5313 ldr(rscratch1, Address(rthread, in_bytes(JavaThread::tlab_start_offset()))); 5314 cmp(rscratch2, rscratch1); 5315 br(Assembler::HS, next); 5316 STOP("assert(top >= start)"); 5317 should_not_reach_here(); 5318 5319 bind(next); 5320 ldr(rscratch2, Address(rthread, in_bytes(JavaThread::tlab_end_offset()))); 5321 ldr(rscratch1, Address(rthread, in_bytes(JavaThread::tlab_top_offset()))); 5322 cmp(rscratch2, rscratch1); 5323 br(Assembler::HS, ok); 5324 STOP("assert(top <= end)"); 5325 should_not_reach_here(); 5326 5327 bind(ok); 5328 ldp(rscratch2, rscratch1, Address(post(sp, 16))); 5329 } 5330 #endif 5331 } 5332 5333 // Writes to stack successive pages until offset reached to check for 5334 // stack overflow + shadow pages. This clobbers tmp. 5335 void MacroAssembler::bang_stack_size(Register size, Register tmp) { 5336 assert_different_registers(tmp, size, rscratch1); 5337 mov(tmp, sp); 5338 // Bang stack for total size given plus shadow page size. 5339 // Bang one page at a time because large size can bang beyond yellow and 5340 // red zones. 5341 Label loop; 5342 mov(rscratch1, (int)os::vm_page_size()); 5343 bind(loop); 5344 lea(tmp, Address(tmp, -(int)os::vm_page_size())); 5345 subsw(size, size, rscratch1); 5346 str(size, Address(tmp)); 5347 br(Assembler::GT, loop); 5348 5349 // Bang down shadow pages too. 5350 // At this point, (tmp-0) is the last address touched, so don't 5351 // touch it again. (It was touched as (tmp-pagesize) but then tmp 5352 // was post-decremented.) Skip this address by starting at i=1, and 5353 // touch a few more pages below. N.B. It is important to touch all 5354 // the way down to and including i=StackShadowPages. 5355 for (int i = 0; i < (int)(StackOverflow::stack_shadow_zone_size() / (int)os::vm_page_size()) - 1; i++) { 5356 // this could be any sized move but this is can be a debugging crumb 5357 // so the bigger the better. 5358 lea(tmp, Address(tmp, -(int)os::vm_page_size())); 5359 str(size, Address(tmp)); 5360 } 5361 } 5362 5363 // Move the address of the polling page into dest. 5364 void MacroAssembler::get_polling_page(Register dest, relocInfo::relocType rtype) { 5365 ldr(dest, Address(rthread, JavaThread::polling_page_offset())); 5366 } 5367 5368 // Read the polling page. The address of the polling page must 5369 // already be in r. 5370 address MacroAssembler::read_polling_page(Register r, relocInfo::relocType rtype) { 5371 address mark; 5372 { 5373 InstructionMark im(this); 5374 code_section()->relocate(inst_mark(), rtype); 5375 ldrw(zr, Address(r, 0)); 5376 mark = inst_mark(); 5377 } 5378 verify_cross_modify_fence_not_required(); 5379 return mark; 5380 } 5381 5382 void MacroAssembler::adrp(Register reg1, const Address &dest, uint64_t &byte_offset) { 5383 relocInfo::relocType rtype = dest.rspec().reloc()->type(); 5384 uint64_t low_page = (uint64_t)CodeCache::low_bound() >> 12; 5385 uint64_t high_page = (uint64_t)(CodeCache::high_bound()-1) >> 12; 5386 uint64_t dest_page = (uint64_t)dest.target() >> 12; 5387 int64_t offset_low = dest_page - low_page; 5388 int64_t offset_high = dest_page - high_page; 5389 5390 assert(is_valid_AArch64_address(dest.target()), "bad address"); 5391 assert(dest.getMode() == Address::literal, "ADRP must be applied to a literal address"); 5392 5393 InstructionMark im(this); 5394 code_section()->relocate(inst_mark(), dest.rspec()); 5395 // 8143067: Ensure that the adrp can reach the dest from anywhere within 5396 // the code cache so that if it is relocated we know it will still reach 5397 if (offset_high >= -(1<<20) && offset_low < (1<<20)) { 5398 _adrp(reg1, dest.target()); 5399 } else { 5400 uint64_t target = (uint64_t)dest.target(); 5401 uint64_t adrp_target 5402 = (target & 0xffffffffULL) | ((uint64_t)pc() & 0xffff00000000ULL); 5403 5404 _adrp(reg1, (address)adrp_target); 5405 movk(reg1, target >> 32, 32); 5406 } 5407 byte_offset = (uint64_t)dest.target() & 0xfff; 5408 } 5409 5410 void MacroAssembler::load_byte_map_base(Register reg) { 5411 CardTable::CardValue* byte_map_base = 5412 ((CardTableBarrierSet*)(BarrierSet::barrier_set()))->card_table()->byte_map_base(); 5413 5414 // Strictly speaking the byte_map_base isn't an address at all, and it might 5415 // even be negative. It is thus materialised as a constant. 5416 mov(reg, (uint64_t)byte_map_base); 5417 } 5418 5419 void MacroAssembler::build_frame(int framesize) { 5420 assert(framesize >= 2 * wordSize, "framesize must include space for FP/LR"); 5421 assert(framesize % (2*wordSize) == 0, "must preserve 2*wordSize alignment"); 5422 protect_return_address(); 5423 if (framesize < ((1 << 9) + 2 * wordSize)) { 5424 sub(sp, sp, framesize); 5425 stp(rfp, lr, Address(sp, framesize - 2 * wordSize)); 5426 if (PreserveFramePointer) add(rfp, sp, framesize - 2 * wordSize); 5427 } else { 5428 stp(rfp, lr, Address(pre(sp, -2 * wordSize))); 5429 if (PreserveFramePointer) mov(rfp, sp); 5430 if (framesize < ((1 << 12) + 2 * wordSize)) 5431 sub(sp, sp, framesize - 2 * wordSize); 5432 else { 5433 mov(rscratch1, framesize - 2 * wordSize); 5434 sub(sp, sp, rscratch1); 5435 } 5436 } 5437 verify_cross_modify_fence_not_required(); 5438 } 5439 5440 void MacroAssembler::remove_frame(int framesize) { 5441 assert(framesize >= 2 * wordSize, "framesize must include space for FP/LR"); 5442 assert(framesize % (2*wordSize) == 0, "must preserve 2*wordSize alignment"); 5443 if (framesize < ((1 << 9) + 2 * wordSize)) { 5444 ldp(rfp, lr, Address(sp, framesize - 2 * wordSize)); 5445 add(sp, sp, framesize); 5446 } else { 5447 if (framesize < ((1 << 12) + 2 * wordSize)) 5448 add(sp, sp, framesize - 2 * wordSize); 5449 else { 5450 mov(rscratch1, framesize - 2 * wordSize); 5451 add(sp, sp, rscratch1); 5452 } 5453 ldp(rfp, lr, Address(post(sp, 2 * wordSize))); 5454 } 5455 authenticate_return_address(); 5456 } 5457 5458 5459 // This method counts leading positive bytes (highest bit not set) in provided byte array 5460 address MacroAssembler::count_positives(Register ary1, Register len, Register result) { 5461 // Simple and most common case of aligned small array which is not at the 5462 // end of memory page is placed here. All other cases are in stub. 5463 Label LOOP, END, STUB, STUB_LONG, SET_RESULT, DONE; 5464 const uint64_t UPPER_BIT_MASK=0x8080808080808080; 5465 assert_different_registers(ary1, len, result); 5466 5467 mov(result, len); 5468 cmpw(len, 0); 5469 br(LE, DONE); 5470 cmpw(len, 4 * wordSize); 5471 br(GE, STUB_LONG); // size > 32 then go to stub 5472 5473 int shift = 64 - exact_log2(os::vm_page_size()); 5474 lsl(rscratch1, ary1, shift); 5475 mov(rscratch2, (size_t)(4 * wordSize) << shift); 5476 adds(rscratch2, rscratch1, rscratch2); // At end of page? 5477 br(CS, STUB); // at the end of page then go to stub 5478 subs(len, len, wordSize); 5479 br(LT, END); 5480 5481 BIND(LOOP); 5482 ldr(rscratch1, Address(post(ary1, wordSize))); 5483 tst(rscratch1, UPPER_BIT_MASK); 5484 br(NE, SET_RESULT); 5485 subs(len, len, wordSize); 5486 br(GE, LOOP); 5487 cmpw(len, -wordSize); 5488 br(EQ, DONE); 5489 5490 BIND(END); 5491 ldr(rscratch1, Address(ary1)); 5492 sub(rscratch2, zr, len, LSL, 3); // LSL 3 is to get bits from bytes 5493 lslv(rscratch1, rscratch1, rscratch2); 5494 tst(rscratch1, UPPER_BIT_MASK); 5495 br(NE, SET_RESULT); 5496 b(DONE); 5497 5498 BIND(STUB); 5499 RuntimeAddress count_pos = RuntimeAddress(StubRoutines::aarch64::count_positives()); 5500 assert(count_pos.target() != nullptr, "count_positives stub has not been generated"); 5501 address tpc1 = trampoline_call(count_pos); 5502 if (tpc1 == nullptr) { 5503 DEBUG_ONLY(reset_labels(STUB_LONG, SET_RESULT, DONE)); 5504 postcond(pc() == badAddress); 5505 return nullptr; 5506 } 5507 b(DONE); 5508 5509 BIND(STUB_LONG); 5510 RuntimeAddress count_pos_long = RuntimeAddress(StubRoutines::aarch64::count_positives_long()); 5511 assert(count_pos_long.target() != nullptr, "count_positives_long stub has not been generated"); 5512 address tpc2 = trampoline_call(count_pos_long); 5513 if (tpc2 == nullptr) { 5514 DEBUG_ONLY(reset_labels(SET_RESULT, DONE)); 5515 postcond(pc() == badAddress); 5516 return nullptr; 5517 } 5518 b(DONE); 5519 5520 BIND(SET_RESULT); 5521 5522 add(len, len, wordSize); 5523 sub(result, result, len); 5524 5525 BIND(DONE); 5526 postcond(pc() != badAddress); 5527 return pc(); 5528 } 5529 5530 // Clobbers: rscratch1, rscratch2, rflags 5531 // May also clobber v0-v7 when (!UseSimpleArrayEquals && UseSIMDForArrayEquals) 5532 address MacroAssembler::arrays_equals(Register a1, Register a2, Register tmp3, 5533 Register tmp4, Register tmp5, Register result, 5534 Register cnt1, int elem_size) { 5535 Label DONE, SAME; 5536 Register tmp1 = rscratch1; 5537 Register tmp2 = rscratch2; 5538 int elem_per_word = wordSize/elem_size; 5539 int log_elem_size = exact_log2(elem_size); 5540 int klass_offset = arrayOopDesc::klass_offset_in_bytes(); 5541 int length_offset = arrayOopDesc::length_offset_in_bytes(); 5542 int base_offset 5543 = arrayOopDesc::base_offset_in_bytes(elem_size == 2 ? T_CHAR : T_BYTE); 5544 // When the length offset is not aligned to 8 bytes, 5545 // then we align it down. This is valid because the new 5546 // offset will always be the klass which is the same 5547 // for type arrays. 5548 int start_offset = align_down(length_offset, BytesPerWord); 5549 int extra_length = base_offset - start_offset; 5550 assert(start_offset == length_offset || start_offset == klass_offset, 5551 "start offset must be 8-byte-aligned or be the klass offset"); 5552 assert(base_offset != start_offset, "must include the length field"); 5553 extra_length = extra_length / elem_size; // We count in elements, not bytes. 5554 int stubBytesThreshold = 3 * 64 + (UseSIMDForArrayEquals ? 0 : 16); 5555 5556 assert(elem_size == 1 || elem_size == 2, "must be char or byte"); 5557 assert_different_registers(a1, a2, result, cnt1, rscratch1, rscratch2); 5558 5559 #ifndef PRODUCT 5560 { 5561 const char kind = (elem_size == 2) ? 'U' : 'L'; 5562 char comment[64]; 5563 snprintf(comment, sizeof comment, "array_equals%c{", kind); 5564 BLOCK_COMMENT(comment); 5565 } 5566 #endif 5567 5568 // if (a1 == a2) 5569 // return true; 5570 cmpoop(a1, a2); // May have read barriers for a1 and a2. 5571 br(EQ, SAME); 5572 5573 if (UseSimpleArrayEquals) { 5574 Label NEXT_WORD, SHORT, TAIL03, TAIL01, A_MIGHT_BE_NULL, A_IS_NOT_NULL; 5575 // if (a1 == nullptr || a2 == nullptr) 5576 // return false; 5577 // a1 & a2 == 0 means (some-pointer is null) or 5578 // (very-rare-or-even-probably-impossible-pointer-values) 5579 // so, we can save one branch in most cases 5580 tst(a1, a2); 5581 mov(result, false); 5582 br(EQ, A_MIGHT_BE_NULL); 5583 // if (a1.length != a2.length) 5584 // return false; 5585 bind(A_IS_NOT_NULL); 5586 ldrw(cnt1, Address(a1, length_offset)); 5587 // Increase loop counter by diff between base- and actual start-offset. 5588 addw(cnt1, cnt1, extra_length); 5589 lea(a1, Address(a1, start_offset)); 5590 lea(a2, Address(a2, start_offset)); 5591 // Check for short strings, i.e. smaller than wordSize. 5592 subs(cnt1, cnt1, elem_per_word); 5593 br(Assembler::LT, SHORT); 5594 // Main 8 byte comparison loop. 5595 bind(NEXT_WORD); { 5596 ldr(tmp1, Address(post(a1, wordSize))); 5597 ldr(tmp2, Address(post(a2, wordSize))); 5598 subs(cnt1, cnt1, elem_per_word); 5599 eor(tmp5, tmp1, tmp2); 5600 cbnz(tmp5, DONE); 5601 } br(GT, NEXT_WORD); 5602 // Last longword. In the case where length == 4 we compare the 5603 // same longword twice, but that's still faster than another 5604 // conditional branch. 5605 // cnt1 could be 0, -1, -2, -3, -4 for chars; -4 only happens when 5606 // length == 4. 5607 if (log_elem_size > 0) 5608 lsl(cnt1, cnt1, log_elem_size); 5609 ldr(tmp3, Address(a1, cnt1)); 5610 ldr(tmp4, Address(a2, cnt1)); 5611 eor(tmp5, tmp3, tmp4); 5612 cbnz(tmp5, DONE); 5613 b(SAME); 5614 bind(A_MIGHT_BE_NULL); 5615 // in case both a1 and a2 are not-null, proceed with loads 5616 cbz(a1, DONE); 5617 cbz(a2, DONE); 5618 b(A_IS_NOT_NULL); 5619 bind(SHORT); 5620 5621 tbz(cnt1, 2 - log_elem_size, TAIL03); // 0-7 bytes left. 5622 { 5623 ldrw(tmp1, Address(post(a1, 4))); 5624 ldrw(tmp2, Address(post(a2, 4))); 5625 eorw(tmp5, tmp1, tmp2); 5626 cbnzw(tmp5, DONE); 5627 } 5628 bind(TAIL03); 5629 tbz(cnt1, 1 - log_elem_size, TAIL01); // 0-3 bytes left. 5630 { 5631 ldrh(tmp3, Address(post(a1, 2))); 5632 ldrh(tmp4, Address(post(a2, 2))); 5633 eorw(tmp5, tmp3, tmp4); 5634 cbnzw(tmp5, DONE); 5635 } 5636 bind(TAIL01); 5637 if (elem_size == 1) { // Only needed when comparing byte arrays. 5638 tbz(cnt1, 0, SAME); // 0-1 bytes left. 5639 { 5640 ldrb(tmp1, a1); 5641 ldrb(tmp2, a2); 5642 eorw(tmp5, tmp1, tmp2); 5643 cbnzw(tmp5, DONE); 5644 } 5645 } 5646 } else { 5647 Label NEXT_DWORD, SHORT, TAIL, TAIL2, STUB, 5648 CSET_EQ, LAST_CHECK; 5649 mov(result, false); 5650 cbz(a1, DONE); 5651 ldrw(cnt1, Address(a1, length_offset)); 5652 cbz(a2, DONE); 5653 // Increase loop counter by diff between base- and actual start-offset. 5654 addw(cnt1, cnt1, extra_length); 5655 5656 // on most CPUs a2 is still "locked"(surprisingly) in ldrw and it's 5657 // faster to perform another branch before comparing a1 and a2 5658 cmp(cnt1, (u1)elem_per_word); 5659 br(LE, SHORT); // short or same 5660 ldr(tmp3, Address(pre(a1, start_offset))); 5661 subs(zr, cnt1, stubBytesThreshold); 5662 br(GE, STUB); 5663 ldr(tmp4, Address(pre(a2, start_offset))); 5664 sub(tmp5, zr, cnt1, LSL, 3 + log_elem_size); 5665 5666 // Main 16 byte comparison loop with 2 exits 5667 bind(NEXT_DWORD); { 5668 ldr(tmp1, Address(pre(a1, wordSize))); 5669 ldr(tmp2, Address(pre(a2, wordSize))); 5670 subs(cnt1, cnt1, 2 * elem_per_word); 5671 br(LE, TAIL); 5672 eor(tmp4, tmp3, tmp4); 5673 cbnz(tmp4, DONE); 5674 ldr(tmp3, Address(pre(a1, wordSize))); 5675 ldr(tmp4, Address(pre(a2, wordSize))); 5676 cmp(cnt1, (u1)elem_per_word); 5677 br(LE, TAIL2); 5678 cmp(tmp1, tmp2); 5679 } br(EQ, NEXT_DWORD); 5680 b(DONE); 5681 5682 bind(TAIL); 5683 eor(tmp4, tmp3, tmp4); 5684 eor(tmp2, tmp1, tmp2); 5685 lslv(tmp2, tmp2, tmp5); 5686 orr(tmp5, tmp4, tmp2); 5687 cmp(tmp5, zr); 5688 b(CSET_EQ); 5689 5690 bind(TAIL2); 5691 eor(tmp2, tmp1, tmp2); 5692 cbnz(tmp2, DONE); 5693 b(LAST_CHECK); 5694 5695 bind(STUB); 5696 ldr(tmp4, Address(pre(a2, start_offset))); 5697 if (elem_size == 2) { // convert to byte counter 5698 lsl(cnt1, cnt1, 1); 5699 } 5700 eor(tmp5, tmp3, tmp4); 5701 cbnz(tmp5, DONE); 5702 RuntimeAddress stub = RuntimeAddress(StubRoutines::aarch64::large_array_equals()); 5703 assert(stub.target() != nullptr, "array_equals_long stub has not been generated"); 5704 address tpc = trampoline_call(stub); 5705 if (tpc == nullptr) { 5706 DEBUG_ONLY(reset_labels(SHORT, LAST_CHECK, CSET_EQ, SAME, DONE)); 5707 postcond(pc() == badAddress); 5708 return nullptr; 5709 } 5710 b(DONE); 5711 5712 // (a1 != null && a2 == null) || (a1 != null && a2 != null && a1 == a2) 5713 // so, if a2 == null => return false(0), else return true, so we can return a2 5714 mov(result, a2); 5715 b(DONE); 5716 bind(SHORT); 5717 sub(tmp5, zr, cnt1, LSL, 3 + log_elem_size); 5718 ldr(tmp3, Address(a1, start_offset)); 5719 ldr(tmp4, Address(a2, start_offset)); 5720 bind(LAST_CHECK); 5721 eor(tmp4, tmp3, tmp4); 5722 lslv(tmp5, tmp4, tmp5); 5723 cmp(tmp5, zr); 5724 bind(CSET_EQ); 5725 cset(result, EQ); 5726 b(DONE); 5727 } 5728 5729 bind(SAME); 5730 mov(result, true); 5731 // That's it. 5732 bind(DONE); 5733 5734 BLOCK_COMMENT("} array_equals"); 5735 postcond(pc() != badAddress); 5736 return pc(); 5737 } 5738 5739 // Compare Strings 5740 5741 // For Strings we're passed the address of the first characters in a1 5742 // and a2 and the length in cnt1. 5743 // There are two implementations. For arrays >= 8 bytes, all 5744 // comparisons (including the final one, which may overlap) are 5745 // performed 8 bytes at a time. For strings < 8 bytes, we compare a 5746 // halfword, then a short, and then a byte. 5747 5748 void MacroAssembler::string_equals(Register a1, Register a2, 5749 Register result, Register cnt1) 5750 { 5751 Label SAME, DONE, SHORT, NEXT_WORD; 5752 Register tmp1 = rscratch1; 5753 Register tmp2 = rscratch2; 5754 Register cnt2 = tmp2; // cnt2 only used in array length compare 5755 5756 assert_different_registers(a1, a2, result, cnt1, rscratch1, rscratch2); 5757 5758 #ifndef PRODUCT 5759 { 5760 char comment[64]; 5761 snprintf(comment, sizeof comment, "{string_equalsL"); 5762 BLOCK_COMMENT(comment); 5763 } 5764 #endif 5765 5766 mov(result, false); 5767 5768 // Check for short strings, i.e. smaller than wordSize. 5769 subs(cnt1, cnt1, wordSize); 5770 br(Assembler::LT, SHORT); 5771 // Main 8 byte comparison loop. 5772 bind(NEXT_WORD); { 5773 ldr(tmp1, Address(post(a1, wordSize))); 5774 ldr(tmp2, Address(post(a2, wordSize))); 5775 subs(cnt1, cnt1, wordSize); 5776 eor(tmp1, tmp1, tmp2); 5777 cbnz(tmp1, DONE); 5778 } br(GT, NEXT_WORD); 5779 // Last longword. In the case where length == 4 we compare the 5780 // same longword twice, but that's still faster than another 5781 // conditional branch. 5782 // cnt1 could be 0, -1, -2, -3, -4 for chars; -4 only happens when 5783 // length == 4. 5784 ldr(tmp1, Address(a1, cnt1)); 5785 ldr(tmp2, Address(a2, cnt1)); 5786 eor(tmp2, tmp1, tmp2); 5787 cbnz(tmp2, DONE); 5788 b(SAME); 5789 5790 bind(SHORT); 5791 Label TAIL03, TAIL01; 5792 5793 tbz(cnt1, 2, TAIL03); // 0-7 bytes left. 5794 { 5795 ldrw(tmp1, Address(post(a1, 4))); 5796 ldrw(tmp2, Address(post(a2, 4))); 5797 eorw(tmp1, tmp1, tmp2); 5798 cbnzw(tmp1, DONE); 5799 } 5800 bind(TAIL03); 5801 tbz(cnt1, 1, TAIL01); // 0-3 bytes left. 5802 { 5803 ldrh(tmp1, Address(post(a1, 2))); 5804 ldrh(tmp2, Address(post(a2, 2))); 5805 eorw(tmp1, tmp1, tmp2); 5806 cbnzw(tmp1, DONE); 5807 } 5808 bind(TAIL01); 5809 tbz(cnt1, 0, SAME); // 0-1 bytes left. 5810 { 5811 ldrb(tmp1, a1); 5812 ldrb(tmp2, a2); 5813 eorw(tmp1, tmp1, tmp2); 5814 cbnzw(tmp1, DONE); 5815 } 5816 // Arrays are equal. 5817 bind(SAME); 5818 mov(result, true); 5819 5820 // That's it. 5821 bind(DONE); 5822 BLOCK_COMMENT("} string_equals"); 5823 } 5824 5825 5826 // The size of the blocks erased by the zero_blocks stub. We must 5827 // handle anything smaller than this ourselves in zero_words(). 5828 const int MacroAssembler::zero_words_block_size = 8; 5829 5830 // zero_words() is used by C2 ClearArray patterns and by 5831 // C1_MacroAssembler. It is as small as possible, handling small word 5832 // counts locally and delegating anything larger to the zero_blocks 5833 // stub. It is expanded many times in compiled code, so it is 5834 // important to keep it short. 5835 5836 // ptr: Address of a buffer to be zeroed. 5837 // cnt: Count in HeapWords. 5838 // 5839 // ptr, cnt, rscratch1, and rscratch2 are clobbered. 5840 address MacroAssembler::zero_words(Register ptr, Register cnt) 5841 { 5842 assert(is_power_of_2(zero_words_block_size), "adjust this"); 5843 5844 BLOCK_COMMENT("zero_words {"); 5845 assert(ptr == r10 && cnt == r11, "mismatch in register usage"); 5846 RuntimeAddress zero_blocks = RuntimeAddress(StubRoutines::aarch64::zero_blocks()); 5847 assert(zero_blocks.target() != nullptr, "zero_blocks stub has not been generated"); 5848 5849 subs(rscratch1, cnt, zero_words_block_size); 5850 Label around; 5851 br(LO, around); 5852 { 5853 RuntimeAddress zero_blocks = RuntimeAddress(StubRoutines::aarch64::zero_blocks()); 5854 assert(zero_blocks.target() != nullptr, "zero_blocks stub has not been generated"); 5855 // Make sure this is a C2 compilation. C1 allocates space only for 5856 // trampoline stubs generated by Call LIR ops, and in any case it 5857 // makes sense for a C1 compilation task to proceed as quickly as 5858 // possible. 5859 CompileTask* task; 5860 if (StubRoutines::aarch64::complete() 5861 && Thread::current()->is_Compiler_thread() 5862 && (task = ciEnv::current()->task()) 5863 && is_c2_compile(task->comp_level())) { 5864 address tpc = trampoline_call(zero_blocks); 5865 if (tpc == nullptr) { 5866 DEBUG_ONLY(reset_labels(around)); 5867 return nullptr; 5868 } 5869 } else { 5870 far_call(zero_blocks); 5871 } 5872 } 5873 bind(around); 5874 5875 // We have a few words left to do. zero_blocks has adjusted r10 and r11 5876 // for us. 5877 for (int i = zero_words_block_size >> 1; i > 1; i >>= 1) { 5878 Label l; 5879 tbz(cnt, exact_log2(i), l); 5880 for (int j = 0; j < i; j += 2) { 5881 stp(zr, zr, post(ptr, 2 * BytesPerWord)); 5882 } 5883 bind(l); 5884 } 5885 { 5886 Label l; 5887 tbz(cnt, 0, l); 5888 str(zr, Address(ptr)); 5889 bind(l); 5890 } 5891 5892 BLOCK_COMMENT("} zero_words"); 5893 return pc(); 5894 } 5895 5896 // base: Address of a buffer to be zeroed, 8 bytes aligned. 5897 // cnt: Immediate count in HeapWords. 5898 // 5899 // r10, r11, rscratch1, and rscratch2 are clobbered. 5900 address MacroAssembler::zero_words(Register base, uint64_t cnt) 5901 { 5902 assert(wordSize <= BlockZeroingLowLimit, 5903 "increase BlockZeroingLowLimit"); 5904 address result = nullptr; 5905 if (cnt <= (uint64_t)BlockZeroingLowLimit / BytesPerWord) { 5906 #ifndef PRODUCT 5907 { 5908 char buf[64]; 5909 snprintf(buf, sizeof buf, "zero_words (count = %" PRIu64 ") {", cnt); 5910 BLOCK_COMMENT(buf); 5911 } 5912 #endif 5913 if (cnt >= 16) { 5914 uint64_t loops = cnt/16; 5915 if (loops > 1) { 5916 mov(rscratch2, loops - 1); 5917 } 5918 { 5919 Label loop; 5920 bind(loop); 5921 for (int i = 0; i < 16; i += 2) { 5922 stp(zr, zr, Address(base, i * BytesPerWord)); 5923 } 5924 add(base, base, 16 * BytesPerWord); 5925 if (loops > 1) { 5926 subs(rscratch2, rscratch2, 1); 5927 br(GE, loop); 5928 } 5929 } 5930 } 5931 cnt %= 16; 5932 int i = cnt & 1; // store any odd word to start 5933 if (i) str(zr, Address(base)); 5934 for (; i < (int)cnt; i += 2) { 5935 stp(zr, zr, Address(base, i * wordSize)); 5936 } 5937 BLOCK_COMMENT("} zero_words"); 5938 result = pc(); 5939 } else { 5940 mov(r10, base); mov(r11, cnt); 5941 result = zero_words(r10, r11); 5942 } 5943 return result; 5944 } 5945 5946 // Zero blocks of memory by using DC ZVA. 5947 // 5948 // Aligns the base address first sufficiently for DC ZVA, then uses 5949 // DC ZVA repeatedly for every full block. cnt is the size to be 5950 // zeroed in HeapWords. Returns the count of words left to be zeroed 5951 // in cnt. 5952 // 5953 // NOTE: This is intended to be used in the zero_blocks() stub. If 5954 // you want to use it elsewhere, note that cnt must be >= 2*zva_length. 5955 void MacroAssembler::zero_dcache_blocks(Register base, Register cnt) { 5956 Register tmp = rscratch1; 5957 Register tmp2 = rscratch2; 5958 int zva_length = VM_Version::zva_length(); 5959 Label initial_table_end, loop_zva; 5960 Label fini; 5961 5962 // Base must be 16 byte aligned. If not just return and let caller handle it 5963 tst(base, 0x0f); 5964 br(Assembler::NE, fini); 5965 // Align base with ZVA length. 5966 neg(tmp, base); 5967 andr(tmp, tmp, zva_length - 1); 5968 5969 // tmp: the number of bytes to be filled to align the base with ZVA length. 5970 add(base, base, tmp); 5971 sub(cnt, cnt, tmp, Assembler::ASR, 3); 5972 adr(tmp2, initial_table_end); 5973 sub(tmp2, tmp2, tmp, Assembler::LSR, 2); 5974 br(tmp2); 5975 5976 for (int i = -zva_length + 16; i < 0; i += 16) 5977 stp(zr, zr, Address(base, i)); 5978 bind(initial_table_end); 5979 5980 sub(cnt, cnt, zva_length >> 3); 5981 bind(loop_zva); 5982 dc(Assembler::ZVA, base); 5983 subs(cnt, cnt, zva_length >> 3); 5984 add(base, base, zva_length); 5985 br(Assembler::GE, loop_zva); 5986 add(cnt, cnt, zva_length >> 3); // count not zeroed by DC ZVA 5987 bind(fini); 5988 } 5989 5990 // base: Address of a buffer to be filled, 8 bytes aligned. 5991 // cnt: Count in 8-byte unit. 5992 // value: Value to be filled with. 5993 // base will point to the end of the buffer after filling. 5994 void MacroAssembler::fill_words(Register base, Register cnt, Register value) 5995 { 5996 // Algorithm: 5997 // 5998 // if (cnt == 0) { 5999 // return; 6000 // } 6001 // if ((p & 8) != 0) { 6002 // *p++ = v; 6003 // } 6004 // 6005 // scratch1 = cnt & 14; 6006 // cnt -= scratch1; 6007 // p += scratch1; 6008 // switch (scratch1 / 2) { 6009 // do { 6010 // cnt -= 16; 6011 // p[-16] = v; 6012 // p[-15] = v; 6013 // case 7: 6014 // p[-14] = v; 6015 // p[-13] = v; 6016 // case 6: 6017 // p[-12] = v; 6018 // p[-11] = v; 6019 // // ... 6020 // case 1: 6021 // p[-2] = v; 6022 // p[-1] = v; 6023 // case 0: 6024 // p += 16; 6025 // } while (cnt); 6026 // } 6027 // if ((cnt & 1) == 1) { 6028 // *p++ = v; 6029 // } 6030 6031 assert_different_registers(base, cnt, value, rscratch1, rscratch2); 6032 6033 Label fini, skip, entry, loop; 6034 const int unroll = 8; // Number of stp instructions we'll unroll 6035 6036 cbz(cnt, fini); 6037 tbz(base, 3, skip); 6038 str(value, Address(post(base, 8))); 6039 sub(cnt, cnt, 1); 6040 bind(skip); 6041 6042 andr(rscratch1, cnt, (unroll-1) * 2); 6043 sub(cnt, cnt, rscratch1); 6044 add(base, base, rscratch1, Assembler::LSL, 3); 6045 adr(rscratch2, entry); 6046 sub(rscratch2, rscratch2, rscratch1, Assembler::LSL, 1); 6047 br(rscratch2); 6048 6049 bind(loop); 6050 add(base, base, unroll * 16); 6051 for (int i = -unroll; i < 0; i++) 6052 stp(value, value, Address(base, i * 16)); 6053 bind(entry); 6054 subs(cnt, cnt, unroll * 2); 6055 br(Assembler::GE, loop); 6056 6057 tbz(cnt, 0, fini); 6058 str(value, Address(post(base, 8))); 6059 bind(fini); 6060 } 6061 6062 // Intrinsic for 6063 // 6064 // - sun/nio/cs/ISO_8859_1$Encoder.implEncodeISOArray 6065 // return the number of characters copied. 6066 // - java/lang/StringUTF16.compress 6067 // return index of non-latin1 character if copy fails, otherwise 'len'. 6068 // 6069 // This version always returns the number of characters copied, and does not 6070 // clobber the 'len' register. A successful copy will complete with the post- 6071 // condition: 'res' == 'len', while an unsuccessful copy will exit with the 6072 // post-condition: 0 <= 'res' < 'len'. 6073 // 6074 // NOTE: Attempts to use 'ld2' (and 'umaxv' in the ISO part) has proven to 6075 // degrade performance (on Ampere Altra - Neoverse N1), to an extent 6076 // beyond the acceptable, even though the footprint would be smaller. 6077 // Using 'umaxv' in the ASCII-case comes with a small penalty but does 6078 // avoid additional bloat. 6079 // 6080 // Clobbers: src, dst, res, rscratch1, rscratch2, rflags 6081 void MacroAssembler::encode_iso_array(Register src, Register dst, 6082 Register len, Register res, bool ascii, 6083 FloatRegister vtmp0, FloatRegister vtmp1, 6084 FloatRegister vtmp2, FloatRegister vtmp3, 6085 FloatRegister vtmp4, FloatRegister vtmp5) 6086 { 6087 Register cnt = res; 6088 Register max = rscratch1; 6089 Register chk = rscratch2; 6090 6091 prfm(Address(src), PLDL1STRM); 6092 movw(cnt, len); 6093 6094 #define ASCII(insn) do { if (ascii) { insn; } } while (0) 6095 6096 Label LOOP_32, DONE_32, FAIL_32; 6097 6098 BIND(LOOP_32); 6099 { 6100 cmpw(cnt, 32); 6101 br(LT, DONE_32); 6102 ld1(vtmp0, vtmp1, vtmp2, vtmp3, T8H, Address(post(src, 64))); 6103 // Extract lower bytes. 6104 FloatRegister vlo0 = vtmp4; 6105 FloatRegister vlo1 = vtmp5; 6106 uzp1(vlo0, T16B, vtmp0, vtmp1); 6107 uzp1(vlo1, T16B, vtmp2, vtmp3); 6108 // Merge bits... 6109 orr(vtmp0, T16B, vtmp0, vtmp1); 6110 orr(vtmp2, T16B, vtmp2, vtmp3); 6111 // Extract merged upper bytes. 6112 FloatRegister vhix = vtmp0; 6113 uzp2(vhix, T16B, vtmp0, vtmp2); 6114 // ISO-check on hi-parts (all zero). 6115 // ASCII-check on lo-parts (no sign). 6116 FloatRegister vlox = vtmp1; // Merge lower bytes. 6117 ASCII(orr(vlox, T16B, vlo0, vlo1)); 6118 umov(chk, vhix, D, 1); ASCII(cm(LT, vlox, T16B, vlox)); 6119 fmovd(max, vhix); ASCII(umaxv(vlox, T16B, vlox)); 6120 orr(chk, chk, max); ASCII(umov(max, vlox, B, 0)); 6121 ASCII(orr(chk, chk, max)); 6122 cbnz(chk, FAIL_32); 6123 subw(cnt, cnt, 32); 6124 st1(vlo0, vlo1, T16B, Address(post(dst, 32))); 6125 b(LOOP_32); 6126 } 6127 BIND(FAIL_32); 6128 sub(src, src, 64); 6129 BIND(DONE_32); 6130 6131 Label LOOP_8, SKIP_8; 6132 6133 BIND(LOOP_8); 6134 { 6135 cmpw(cnt, 8); 6136 br(LT, SKIP_8); 6137 FloatRegister vhi = vtmp0; 6138 FloatRegister vlo = vtmp1; 6139 ld1(vtmp3, T8H, src); 6140 uzp1(vlo, T16B, vtmp3, vtmp3); 6141 uzp2(vhi, T16B, vtmp3, vtmp3); 6142 // ISO-check on hi-parts (all zero). 6143 // ASCII-check on lo-parts (no sign). 6144 ASCII(cm(LT, vtmp2, T16B, vlo)); 6145 fmovd(chk, vhi); ASCII(umaxv(vtmp2, T16B, vtmp2)); 6146 ASCII(umov(max, vtmp2, B, 0)); 6147 ASCII(orr(chk, chk, max)); 6148 cbnz(chk, SKIP_8); 6149 6150 strd(vlo, Address(post(dst, 8))); 6151 subw(cnt, cnt, 8); 6152 add(src, src, 16); 6153 b(LOOP_8); 6154 } 6155 BIND(SKIP_8); 6156 6157 #undef ASCII 6158 6159 Label LOOP, DONE; 6160 6161 cbz(cnt, DONE); 6162 BIND(LOOP); 6163 { 6164 Register chr = rscratch1; 6165 ldrh(chr, Address(post(src, 2))); 6166 tst(chr, ascii ? 0xff80 : 0xff00); 6167 br(NE, DONE); 6168 strb(chr, Address(post(dst, 1))); 6169 subs(cnt, cnt, 1); 6170 br(GT, LOOP); 6171 } 6172 BIND(DONE); 6173 // Return index where we stopped. 6174 subw(res, len, cnt); 6175 } 6176 6177 // Inflate byte[] array to char[]. 6178 // Clobbers: src, dst, len, rflags, rscratch1, v0-v6 6179 address MacroAssembler::byte_array_inflate(Register src, Register dst, Register len, 6180 FloatRegister vtmp1, FloatRegister vtmp2, 6181 FloatRegister vtmp3, Register tmp4) { 6182 Label big, done, after_init, to_stub; 6183 6184 assert_different_registers(src, dst, len, tmp4, rscratch1); 6185 6186 fmovd(vtmp1, 0.0); 6187 lsrw(tmp4, len, 3); 6188 bind(after_init); 6189 cbnzw(tmp4, big); 6190 // Short string: less than 8 bytes. 6191 { 6192 Label loop, tiny; 6193 6194 cmpw(len, 4); 6195 br(LT, tiny); 6196 // Use SIMD to do 4 bytes. 6197 ldrs(vtmp2, post(src, 4)); 6198 zip1(vtmp3, T8B, vtmp2, vtmp1); 6199 subw(len, len, 4); 6200 strd(vtmp3, post(dst, 8)); 6201 6202 cbzw(len, done); 6203 6204 // Do the remaining bytes by steam. 6205 bind(loop); 6206 ldrb(tmp4, post(src, 1)); 6207 strh(tmp4, post(dst, 2)); 6208 subw(len, len, 1); 6209 6210 bind(tiny); 6211 cbnz(len, loop); 6212 6213 b(done); 6214 } 6215 6216 if (SoftwarePrefetchHintDistance >= 0) { 6217 bind(to_stub); 6218 RuntimeAddress stub = RuntimeAddress(StubRoutines::aarch64::large_byte_array_inflate()); 6219 assert(stub.target() != nullptr, "large_byte_array_inflate stub has not been generated"); 6220 address tpc = trampoline_call(stub); 6221 if (tpc == nullptr) { 6222 DEBUG_ONLY(reset_labels(big, done)); 6223 postcond(pc() == badAddress); 6224 return nullptr; 6225 } 6226 b(after_init); 6227 } 6228 6229 // Unpack the bytes 8 at a time. 6230 bind(big); 6231 { 6232 Label loop, around, loop_last, loop_start; 6233 6234 if (SoftwarePrefetchHintDistance >= 0) { 6235 const int large_loop_threshold = (64 + 16)/8; 6236 ldrd(vtmp2, post(src, 8)); 6237 andw(len, len, 7); 6238 cmp(tmp4, (u1)large_loop_threshold); 6239 br(GE, to_stub); 6240 b(loop_start); 6241 6242 bind(loop); 6243 ldrd(vtmp2, post(src, 8)); 6244 bind(loop_start); 6245 subs(tmp4, tmp4, 1); 6246 br(EQ, loop_last); 6247 zip1(vtmp2, T16B, vtmp2, vtmp1); 6248 ldrd(vtmp3, post(src, 8)); 6249 st1(vtmp2, T8H, post(dst, 16)); 6250 subs(tmp4, tmp4, 1); 6251 zip1(vtmp3, T16B, vtmp3, vtmp1); 6252 st1(vtmp3, T8H, post(dst, 16)); 6253 br(NE, loop); 6254 b(around); 6255 bind(loop_last); 6256 zip1(vtmp2, T16B, vtmp2, vtmp1); 6257 st1(vtmp2, T8H, post(dst, 16)); 6258 bind(around); 6259 cbz(len, done); 6260 } else { 6261 andw(len, len, 7); 6262 bind(loop); 6263 ldrd(vtmp2, post(src, 8)); 6264 sub(tmp4, tmp4, 1); 6265 zip1(vtmp3, T16B, vtmp2, vtmp1); 6266 st1(vtmp3, T8H, post(dst, 16)); 6267 cbnz(tmp4, loop); 6268 } 6269 } 6270 6271 // Do the tail of up to 8 bytes. 6272 add(src, src, len); 6273 ldrd(vtmp3, Address(src, -8)); 6274 add(dst, dst, len, ext::uxtw, 1); 6275 zip1(vtmp3, T16B, vtmp3, vtmp1); 6276 strq(vtmp3, Address(dst, -16)); 6277 6278 bind(done); 6279 postcond(pc() != badAddress); 6280 return pc(); 6281 } 6282 6283 // Compress char[] array to byte[]. 6284 // Intrinsic for java.lang.StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 6285 // Return the array length if every element in array can be encoded, 6286 // otherwise, the index of first non-latin1 (> 0xff) character. 6287 void MacroAssembler::char_array_compress(Register src, Register dst, Register len, 6288 Register res, 6289 FloatRegister tmp0, FloatRegister tmp1, 6290 FloatRegister tmp2, FloatRegister tmp3, 6291 FloatRegister tmp4, FloatRegister tmp5) { 6292 encode_iso_array(src, dst, len, res, false, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5); 6293 } 6294 6295 // java.math.round(double a) 6296 // Returns the closest long to the argument, with ties rounding to 6297 // positive infinity. This requires some fiddling for corner 6298 // cases. We take care to avoid double rounding in e.g. (jlong)(a + 0.5). 6299 void MacroAssembler::java_round_double(Register dst, FloatRegister src, 6300 FloatRegister ftmp) { 6301 Label DONE; 6302 BLOCK_COMMENT("java_round_double: { "); 6303 fmovd(rscratch1, src); 6304 // Use RoundToNearestTiesAway unless src small and -ve. 6305 fcvtasd(dst, src); 6306 // Test if src >= 0 || abs(src) >= 0x1.0p52 6307 eor(rscratch1, rscratch1, UCONST64(1) << 63); // flip sign bit 6308 mov(rscratch2, julong_cast(0x1.0p52)); 6309 cmp(rscratch1, rscratch2); 6310 br(HS, DONE); { 6311 // src < 0 && abs(src) < 0x1.0p52 6312 // src may have a fractional part, so add 0.5 6313 fmovd(ftmp, 0.5); 6314 faddd(ftmp, src, ftmp); 6315 // Convert double to jlong, use RoundTowardsNegative 6316 fcvtmsd(dst, ftmp); 6317 } 6318 bind(DONE); 6319 BLOCK_COMMENT("} java_round_double"); 6320 } 6321 6322 void MacroAssembler::java_round_float(Register dst, FloatRegister src, 6323 FloatRegister ftmp) { 6324 Label DONE; 6325 BLOCK_COMMENT("java_round_float: { "); 6326 fmovs(rscratch1, src); 6327 // Use RoundToNearestTiesAway unless src small and -ve. 6328 fcvtassw(dst, src); 6329 // Test if src >= 0 || abs(src) >= 0x1.0p23 6330 eor(rscratch1, rscratch1, 0x80000000); // flip sign bit 6331 mov(rscratch2, jint_cast(0x1.0p23f)); 6332 cmp(rscratch1, rscratch2); 6333 br(HS, DONE); { 6334 // src < 0 && |src| < 0x1.0p23 6335 // src may have a fractional part, so add 0.5 6336 fmovs(ftmp, 0.5f); 6337 fadds(ftmp, src, ftmp); 6338 // Convert float to jint, use RoundTowardsNegative 6339 fcvtmssw(dst, ftmp); 6340 } 6341 bind(DONE); 6342 BLOCK_COMMENT("} java_round_float"); 6343 } 6344 6345 // get_thread() can be called anywhere inside generated code so we 6346 // need to save whatever non-callee save context might get clobbered 6347 // by the call to JavaThread::aarch64_get_thread_helper() or, indeed, 6348 // the call setup code. 6349 // 6350 // On Linux, aarch64_get_thread_helper() clobbers only r0, r1, and flags. 6351 // On other systems, the helper is a usual C function. 6352 // 6353 void MacroAssembler::get_thread(Register dst) { 6354 RegSet saved_regs = 6355 LINUX_ONLY(RegSet::range(r0, r1) + lr - dst) 6356 NOT_LINUX (RegSet::range(r0, r17) + lr - dst); 6357 6358 protect_return_address(); 6359 push(saved_regs, sp); 6360 6361 mov(lr, CAST_FROM_FN_PTR(address, JavaThread::aarch64_get_thread_helper)); 6362 blr(lr); 6363 if (dst != c_rarg0) { 6364 mov(dst, c_rarg0); 6365 } 6366 6367 pop(saved_regs, sp); 6368 authenticate_return_address(); 6369 } 6370 6371 void MacroAssembler::cache_wb(Address line) { 6372 assert(line.getMode() == Address::base_plus_offset, "mode should be base_plus_offset"); 6373 assert(line.index() == noreg, "index should be noreg"); 6374 assert(line.offset() == 0, "offset should be 0"); 6375 // would like to assert this 6376 // assert(line._ext.shift == 0, "shift should be zero"); 6377 if (VM_Version::supports_dcpop()) { 6378 // writeback using clear virtual address to point of persistence 6379 dc(Assembler::CVAP, line.base()); 6380 } else { 6381 // no need to generate anything as Unsafe.writebackMemory should 6382 // never invoke this stub 6383 } 6384 } 6385 6386 void MacroAssembler::cache_wbsync(bool is_pre) { 6387 // we only need a barrier post sync 6388 if (!is_pre) { 6389 membar(Assembler::AnyAny); 6390 } 6391 } 6392 6393 void MacroAssembler::verify_sve_vector_length(Register tmp) { 6394 // Make sure that native code does not change SVE vector length. 6395 if (!UseSVE) return; 6396 Label verify_ok; 6397 movw(tmp, zr); 6398 sve_inc(tmp, B); 6399 subsw(zr, tmp, VM_Version::get_initial_sve_vector_length()); 6400 br(EQ, verify_ok); 6401 stop("Error: SVE vector length has changed since jvm startup"); 6402 bind(verify_ok); 6403 } 6404 6405 void MacroAssembler::verify_ptrue() { 6406 Label verify_ok; 6407 if (!UseSVE) { 6408 return; 6409 } 6410 sve_cntp(rscratch1, B, ptrue, ptrue); // get true elements count. 6411 sve_dec(rscratch1, B); 6412 cbz(rscratch1, verify_ok); 6413 stop("Error: the preserved predicate register (p7) elements are not all true"); 6414 bind(verify_ok); 6415 } 6416 6417 void MacroAssembler::safepoint_isb() { 6418 isb(); 6419 #ifndef PRODUCT 6420 if (VerifyCrossModifyFence) { 6421 // Clear the thread state. 6422 strb(zr, Address(rthread, in_bytes(JavaThread::requires_cross_modify_fence_offset()))); 6423 } 6424 #endif 6425 } 6426 6427 #ifndef PRODUCT 6428 void MacroAssembler::verify_cross_modify_fence_not_required() { 6429 if (VerifyCrossModifyFence) { 6430 // Check if thread needs a cross modify fence. 6431 ldrb(rscratch1, Address(rthread, in_bytes(JavaThread::requires_cross_modify_fence_offset()))); 6432 Label fence_not_required; 6433 cbz(rscratch1, fence_not_required); 6434 // If it does then fail. 6435 lea(rscratch1, CAST_FROM_FN_PTR(address, JavaThread::verify_cross_modify_fence_failure)); 6436 mov(c_rarg0, rthread); 6437 blr(rscratch1); 6438 bind(fence_not_required); 6439 } 6440 } 6441 #endif 6442 6443 void MacroAssembler::spin_wait() { 6444 for (int i = 0; i < VM_Version::spin_wait_desc().inst_count(); ++i) { 6445 switch (VM_Version::spin_wait_desc().inst()) { 6446 case SpinWait::NOP: 6447 nop(); 6448 break; 6449 case SpinWait::ISB: 6450 isb(); 6451 break; 6452 case SpinWait::YIELD: 6453 yield(); 6454 break; 6455 default: 6456 ShouldNotReachHere(); 6457 } 6458 } 6459 } 6460 6461 // Stack frame creation/removal 6462 6463 void MacroAssembler::enter(bool strip_ret_addr) { 6464 if (strip_ret_addr) { 6465 // Addresses can only be signed once. If there are multiple nested frames being created 6466 // in the same function, then the return address needs stripping first. 6467 strip_return_address(); 6468 } 6469 protect_return_address(); 6470 stp(rfp, lr, Address(pre(sp, -2 * wordSize))); 6471 mov(rfp, sp); 6472 } 6473 6474 void MacroAssembler::leave() { 6475 mov(sp, rfp); 6476 ldp(rfp, lr, Address(post(sp, 2 * wordSize))); 6477 authenticate_return_address(); 6478 } 6479 6480 // ROP Protection 6481 // Use the AArch64 PAC feature to add ROP protection for generated code. Use whenever creating/ 6482 // destroying stack frames or whenever directly loading/storing the LR to memory. 6483 // If ROP protection is not set then these functions are no-ops. 6484 // For more details on PAC see pauth_aarch64.hpp. 6485 6486 // Sign the LR. Use during construction of a stack frame, before storing the LR to memory. 6487 // Uses value zero as the modifier. 6488 // 6489 void MacroAssembler::protect_return_address() { 6490 if (VM_Version::use_rop_protection()) { 6491 check_return_address(); 6492 paciaz(); 6493 } 6494 } 6495 6496 // Sign the return value in the given register. Use before updating the LR in the existing stack 6497 // frame for the current function. 6498 // Uses value zero as the modifier. 6499 // 6500 void MacroAssembler::protect_return_address(Register return_reg) { 6501 if (VM_Version::use_rop_protection()) { 6502 check_return_address(return_reg); 6503 paciza(return_reg); 6504 } 6505 } 6506 6507 // Authenticate the LR. Use before function return, after restoring FP and loading LR from memory. 6508 // Uses value zero as the modifier. 6509 // 6510 void MacroAssembler::authenticate_return_address() { 6511 if (VM_Version::use_rop_protection()) { 6512 autiaz(); 6513 check_return_address(); 6514 } 6515 } 6516 6517 // Authenticate the return value in the given register. Use before updating the LR in the existing 6518 // stack frame for the current function. 6519 // Uses value zero as the modifier. 6520 // 6521 void MacroAssembler::authenticate_return_address(Register return_reg) { 6522 if (VM_Version::use_rop_protection()) { 6523 autiza(return_reg); 6524 check_return_address(return_reg); 6525 } 6526 } 6527 6528 // Strip any PAC data from LR without performing any authentication. Use with caution - only if 6529 // there is no guaranteed way of authenticating the LR. 6530 // 6531 void MacroAssembler::strip_return_address() { 6532 if (VM_Version::use_rop_protection()) { 6533 xpaclri(); 6534 } 6535 } 6536 6537 #ifndef PRODUCT 6538 // PAC failures can be difficult to debug. After an authentication failure, a segfault will only 6539 // occur when the pointer is used - ie when the program returns to the invalid LR. At this point 6540 // it is difficult to debug back to the callee function. 6541 // This function simply loads from the address in the given register. 6542 // Use directly after authentication to catch authentication failures. 6543 // Also use before signing to check that the pointer is valid and hasn't already been signed. 6544 // 6545 void MacroAssembler::check_return_address(Register return_reg) { 6546 if (VM_Version::use_rop_protection()) { 6547 ldr(zr, Address(return_reg)); 6548 } 6549 } 6550 #endif 6551 6552 // The java_calling_convention describes stack locations as ideal slots on 6553 // a frame with no abi restrictions. Since we must observe abi restrictions 6554 // (like the placement of the register window) the slots must be biased by 6555 // the following value. 6556 static int reg2offset_in(VMReg r) { 6557 // Account for saved rfp and lr 6558 // This should really be in_preserve_stack_slots 6559 return (r->reg2stack() + 4) * VMRegImpl::stack_slot_size; 6560 } 6561 6562 static int reg2offset_out(VMReg r) { 6563 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 6564 } 6565 6566 // On 64bit we will store integer like items to the stack as 6567 // 64bits items (AArch64 ABI) even though java would only store 6568 // 32bits for a parameter. On 32bit it will simply be 32bits 6569 // So this routine will do 32->32 on 32bit and 32->64 on 64bit 6570 void MacroAssembler::move32_64(VMRegPair src, VMRegPair dst, Register tmp) { 6571 if (src.first()->is_stack()) { 6572 if (dst.first()->is_stack()) { 6573 // stack to stack 6574 ldr(tmp, Address(rfp, reg2offset_in(src.first()))); 6575 str(tmp, Address(sp, reg2offset_out(dst.first()))); 6576 } else { 6577 // stack to reg 6578 ldrsw(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first()))); 6579 } 6580 } else if (dst.first()->is_stack()) { 6581 // reg to stack 6582 str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first()))); 6583 } else { 6584 if (dst.first() != src.first()) { 6585 sxtw(dst.first()->as_Register(), src.first()->as_Register()); 6586 } 6587 } 6588 } 6589 6590 // An oop arg. Must pass a handle not the oop itself 6591 void MacroAssembler::object_move( 6592 OopMap* map, 6593 int oop_handle_offset, 6594 int framesize_in_slots, 6595 VMRegPair src, 6596 VMRegPair dst, 6597 bool is_receiver, 6598 int* receiver_offset) { 6599 6600 // must pass a handle. First figure out the location we use as a handle 6601 6602 Register rHandle = dst.first()->is_stack() ? rscratch2 : dst.first()->as_Register(); 6603 6604 // See if oop is null if it is we need no handle 6605 6606 if (src.first()->is_stack()) { 6607 6608 // Oop is already on the stack as an argument 6609 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 6610 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 6611 if (is_receiver) { 6612 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 6613 } 6614 6615 ldr(rscratch1, Address(rfp, reg2offset_in(src.first()))); 6616 lea(rHandle, Address(rfp, reg2offset_in(src.first()))); 6617 // conditionally move a null 6618 cmp(rscratch1, zr); 6619 csel(rHandle, zr, rHandle, Assembler::EQ); 6620 } else { 6621 6622 // Oop is in an a register we must store it to the space we reserve 6623 // on the stack for oop_handles and pass a handle if oop is non-null 6624 6625 const Register rOop = src.first()->as_Register(); 6626 int oop_slot; 6627 if (rOop == j_rarg0) 6628 oop_slot = 0; 6629 else if (rOop == j_rarg1) 6630 oop_slot = 1; 6631 else if (rOop == j_rarg2) 6632 oop_slot = 2; 6633 else if (rOop == j_rarg3) 6634 oop_slot = 3; 6635 else if (rOop == j_rarg4) 6636 oop_slot = 4; 6637 else if (rOop == j_rarg5) 6638 oop_slot = 5; 6639 else if (rOop == j_rarg6) 6640 oop_slot = 6; 6641 else { 6642 assert(rOop == j_rarg7, "wrong register"); 6643 oop_slot = 7; 6644 } 6645 6646 oop_slot = oop_slot * VMRegImpl::slots_per_word + oop_handle_offset; 6647 int offset = oop_slot*VMRegImpl::stack_slot_size; 6648 6649 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 6650 // Store oop in handle area, may be null 6651 str(rOop, Address(sp, offset)); 6652 if (is_receiver) { 6653 *receiver_offset = offset; 6654 } 6655 6656 cmp(rOop, zr); 6657 lea(rHandle, Address(sp, offset)); 6658 // conditionally move a null 6659 csel(rHandle, zr, rHandle, Assembler::EQ); 6660 } 6661 6662 // If arg is on the stack then place it otherwise it is already in correct reg. 6663 if (dst.first()->is_stack()) { 6664 str(rHandle, Address(sp, reg2offset_out(dst.first()))); 6665 } 6666 } 6667 6668 // A float arg may have to do float reg int reg conversion 6669 void MacroAssembler::float_move(VMRegPair src, VMRegPair dst, Register tmp) { 6670 if (src.first()->is_stack()) { 6671 if (dst.first()->is_stack()) { 6672 ldrw(tmp, Address(rfp, reg2offset_in(src.first()))); 6673 strw(tmp, Address(sp, reg2offset_out(dst.first()))); 6674 } else { 6675 ldrs(dst.first()->as_FloatRegister(), Address(rfp, reg2offset_in(src.first()))); 6676 } 6677 } else if (src.first() != dst.first()) { 6678 if (src.is_single_phys_reg() && dst.is_single_phys_reg()) 6679 fmovs(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 6680 else 6681 strs(src.first()->as_FloatRegister(), Address(sp, reg2offset_out(dst.first()))); 6682 } 6683 } 6684 6685 // A long move 6686 void MacroAssembler::long_move(VMRegPair src, VMRegPair dst, Register tmp) { 6687 if (src.first()->is_stack()) { 6688 if (dst.first()->is_stack()) { 6689 // stack to stack 6690 ldr(tmp, Address(rfp, reg2offset_in(src.first()))); 6691 str(tmp, Address(sp, reg2offset_out(dst.first()))); 6692 } else { 6693 // stack to reg 6694 ldr(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first()))); 6695 } 6696 } else if (dst.first()->is_stack()) { 6697 // reg to stack 6698 // Do we really have to sign extend??? 6699 // __ movslq(src.first()->as_Register(), src.first()->as_Register()); 6700 str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first()))); 6701 } else { 6702 if (dst.first() != src.first()) { 6703 mov(dst.first()->as_Register(), src.first()->as_Register()); 6704 } 6705 } 6706 } 6707 6708 6709 // A double move 6710 void MacroAssembler::double_move(VMRegPair src, VMRegPair dst, Register tmp) { 6711 if (src.first()->is_stack()) { 6712 if (dst.first()->is_stack()) { 6713 ldr(tmp, Address(rfp, reg2offset_in(src.first()))); 6714 str(tmp, Address(sp, reg2offset_out(dst.first()))); 6715 } else { 6716 ldrd(dst.first()->as_FloatRegister(), Address(rfp, reg2offset_in(src.first()))); 6717 } 6718 } else if (src.first() != dst.first()) { 6719 if (src.is_single_phys_reg() && dst.is_single_phys_reg()) 6720 fmovd(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 6721 else 6722 strd(src.first()->as_FloatRegister(), Address(sp, reg2offset_out(dst.first()))); 6723 } 6724 } 6725 6726 // Implements lightweight-locking. 6727 // 6728 // - obj: the object to be locked 6729 // - t1, t2, t3: temporary registers, will be destroyed 6730 // - slow: branched to if locking fails, absolute offset may larger than 32KB (imm14 encoding). 6731 void MacroAssembler::lightweight_lock(Register obj, Register t1, Register t2, Register t3, Label& slow) { 6732 assert(LockingMode == LM_LIGHTWEIGHT, "only used with new lightweight locking"); 6733 assert_different_registers(obj, t1, t2, t3, rscratch1); 6734 6735 Label push; 6736 const Register top = t1; 6737 const Register mark = t2; 6738 const Register t = t3; 6739 6740 // Preload the markWord. It is important that this is the first 6741 // instruction emitted as it is part of C1's null check semantics. 6742 ldr(mark, Address(obj, oopDesc::mark_offset_in_bytes())); 6743 6744 // Check if the lock-stack is full. 6745 ldrw(top, Address(rthread, JavaThread::lock_stack_top_offset())); 6746 cmpw(top, (unsigned)LockStack::end_offset()); 6747 br(Assembler::GE, slow); 6748 6749 // Check for recursion. 6750 subw(t, top, oopSize); 6751 ldr(t, Address(rthread, t)); 6752 cmp(obj, t); 6753 br(Assembler::EQ, push); 6754 6755 // Check header for monitor (0b10). 6756 tst(mark, markWord::monitor_value); 6757 br(Assembler::NE, slow); 6758 6759 // Try to lock. Transition lock bits 0b01 => 0b00 6760 assert(oopDesc::mark_offset_in_bytes() == 0, "required to avoid lea"); 6761 orr(mark, mark, markWord::unlocked_value); 6762 eor(t, mark, markWord::unlocked_value); 6763 cmpxchg(/*addr*/ obj, /*expected*/ mark, /*new*/ t, Assembler::xword, 6764 /*acquire*/ true, /*release*/ false, /*weak*/ false, noreg); 6765 br(Assembler::NE, slow); 6766 6767 bind(push); 6768 // After successful lock, push object on lock-stack. 6769 str(obj, Address(rthread, top)); 6770 addw(top, top, oopSize); 6771 strw(top, Address(rthread, JavaThread::lock_stack_top_offset())); 6772 } 6773 6774 // Implements lightweight-unlocking. 6775 // 6776 // - obj: the object to be unlocked 6777 // - t1, t2, t3: temporary registers 6778 // - slow: branched to if unlocking fails, absolute offset may larger than 32KB (imm14 encoding). 6779 void MacroAssembler::lightweight_unlock(Register obj, Register t1, Register t2, Register t3, Label& slow) { 6780 assert(LockingMode == LM_LIGHTWEIGHT, "only used with new lightweight locking"); 6781 // cmpxchg clobbers rscratch1. 6782 assert_different_registers(obj, t1, t2, t3, rscratch1); 6783 6784 #ifdef ASSERT 6785 { 6786 // Check for lock-stack underflow. 6787 Label stack_ok; 6788 ldrw(t1, Address(rthread, JavaThread::lock_stack_top_offset())); 6789 cmpw(t1, (unsigned)LockStack::start_offset()); 6790 br(Assembler::GE, stack_ok); 6791 STOP("Lock-stack underflow"); 6792 bind(stack_ok); 6793 } 6794 #endif 6795 6796 Label unlocked, push_and_slow; 6797 const Register top = t1; 6798 const Register mark = t2; 6799 const Register t = t3; 6800 6801 // Check if obj is top of lock-stack. 6802 ldrw(top, Address(rthread, JavaThread::lock_stack_top_offset())); 6803 subw(top, top, oopSize); 6804 ldr(t, Address(rthread, top)); 6805 cmp(obj, t); 6806 br(Assembler::NE, slow); 6807 6808 // Pop lock-stack. 6809 DEBUG_ONLY(str(zr, Address(rthread, top));) 6810 strw(top, Address(rthread, JavaThread::lock_stack_top_offset())); 6811 6812 // Check if recursive. 6813 subw(t, top, oopSize); 6814 ldr(t, Address(rthread, t)); 6815 cmp(obj, t); 6816 br(Assembler::EQ, unlocked); 6817 6818 // Not recursive. Check header for monitor (0b10). 6819 ldr(mark, Address(obj, oopDesc::mark_offset_in_bytes())); 6820 tbnz(mark, log2i_exact(markWord::monitor_value), push_and_slow); 6821 6822 #ifdef ASSERT 6823 // Check header not unlocked (0b01). 6824 Label not_unlocked; 6825 tbz(mark, log2i_exact(markWord::unlocked_value), not_unlocked); 6826 stop("lightweight_unlock already unlocked"); 6827 bind(not_unlocked); 6828 #endif 6829 6830 // Try to unlock. Transition lock bits 0b00 => 0b01 6831 assert(oopDesc::mark_offset_in_bytes() == 0, "required to avoid lea"); 6832 orr(t, mark, markWord::unlocked_value); 6833 cmpxchg(obj, mark, t, Assembler::xword, 6834 /*acquire*/ false, /*release*/ true, /*weak*/ false, noreg); 6835 br(Assembler::EQ, unlocked); 6836 6837 bind(push_and_slow); 6838 // Restore lock-stack and handle the unlock in runtime. 6839 DEBUG_ONLY(str(obj, Address(rthread, top));) 6840 addw(top, top, oopSize); 6841 strw(top, Address(rthread, JavaThread::lock_stack_top_offset())); 6842 b(slow); 6843 6844 bind(unlocked); 6845 }