Mercurial > hg > graal-compiler
annotate src/cpu/x86/vm/stubGenerator_x86_64.cpp @ 1972:f95d63e2154a
6989984: Use standard include model for Hospot
Summary: Replaced MakeDeps and the includeDB files with more standardized solutions.
Reviewed-by: coleenp, kvn, kamg
| author | stefank |
|---|---|
| date | Tue, 23 Nov 2010 13:22:55 -0800 |
| parents | d6f45b55c972 |
| children | bbefa3ca1543 |
| rev | line source |
|---|---|
| 0 | 1 /* |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1506
diff
changeset
|
2 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1506
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1506
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1506
diff
changeset
|
21 * questions. |
| 0 | 22 * |
| 23 */ | |
| 24 | |
| 1972 | 25 #include "precompiled.hpp" |
| 26 #include "asm/assembler.hpp" | |
| 27 #include "assembler_x86.inline.hpp" | |
| 28 #include "interpreter/interpreter.hpp" | |
| 29 #include "nativeInst_x86.hpp" | |
| 30 #include "oops/instanceOop.hpp" | |
| 31 #include "oops/methodOop.hpp" | |
| 32 #include "oops/objArrayKlass.hpp" | |
| 33 #include "oops/oop.inline.hpp" | |
| 34 #include "prims/methodHandles.hpp" | |
| 35 #include "runtime/frame.inline.hpp" | |
| 36 #include "runtime/handles.inline.hpp" | |
| 37 #include "runtime/sharedRuntime.hpp" | |
| 38 #include "runtime/stubCodeGenerator.hpp" | |
| 39 #include "runtime/stubRoutines.hpp" | |
| 40 #include "utilities/top.hpp" | |
| 41 #ifdef TARGET_OS_FAMILY_linux | |
| 42 # include "thread_linux.inline.hpp" | |
| 43 #endif | |
| 44 #ifdef TARGET_OS_FAMILY_solaris | |
| 45 # include "thread_solaris.inline.hpp" | |
| 46 #endif | |
| 47 #ifdef TARGET_OS_FAMILY_windows | |
| 48 # include "thread_windows.inline.hpp" | |
| 49 #endif | |
| 50 #ifdef COMPILER2 | |
| 51 #include "opto/runtime.hpp" | |
| 52 #endif | |
| 0 | 53 |
| 54 // Declaration and definition of StubGenerator (no .hpp file). | |
| 55 // For a more detailed description of the stub routine structure | |
| 56 // see the comment in stubRoutines.hpp | |
| 57 | |
| 58 #define __ _masm-> | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
59 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8) |
| 304 | 60 #define a__ ((Assembler*)_masm)-> |
| 0 | 61 |
| 62 #ifdef PRODUCT | |
| 63 #define BLOCK_COMMENT(str) /* nothing */ | |
| 64 #else | |
| 65 #define BLOCK_COMMENT(str) __ block_comment(str) | |
| 66 #endif | |
| 67 | |
| 68 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") | |
| 69 const int MXCSR_MASK = 0xFFC0; // Mask out any pending exceptions | |
| 70 | |
| 71 // Stub Code definitions | |
| 72 | |
| 73 static address handle_unsafe_access() { | |
| 74 JavaThread* thread = JavaThread::current(); | |
| 75 address pc = thread->saved_exception_pc(); | |
| 76 // pc is the instruction which we must emulate | |
| 77 // doing a no-op is fine: return garbage from the load | |
| 78 // therefore, compute npc | |
| 79 address npc = Assembler::locate_next_instruction(pc); | |
| 80 | |
| 81 // request an async exception | |
| 82 thread->set_pending_unsafe_access_error(); | |
| 83 | |
| 84 // return address of next instruction to execute | |
| 85 return npc; | |
| 86 } | |
| 87 | |
| 88 class StubGenerator: public StubCodeGenerator { | |
| 89 private: | |
| 90 | |
| 91 #ifdef PRODUCT | |
| 92 #define inc_counter_np(counter) (0) | |
| 93 #else | |
| 94 void inc_counter_np_(int& counter) { | |
| 95 __ incrementl(ExternalAddress((address)&counter)); | |
| 96 } | |
| 97 #define inc_counter_np(counter) \ | |
| 98 BLOCK_COMMENT("inc_counter " #counter); \ | |
| 99 inc_counter_np_(counter); | |
| 100 #endif | |
| 101 | |
| 102 // Call stubs are used to call Java from C | |
| 103 // | |
| 104 // Linux Arguments: | |
| 105 // c_rarg0: call wrapper address address | |
| 106 // c_rarg1: result address | |
| 107 // c_rarg2: result type BasicType | |
| 108 // c_rarg3: method methodOop | |
| 109 // c_rarg4: (interpreter) entry point address | |
| 110 // c_rarg5: parameters intptr_t* | |
| 111 // 16(rbp): parameter size (in words) int | |
| 112 // 24(rbp): thread Thread* | |
| 113 // | |
| 114 // [ return_from_Java ] <--- rsp | |
| 115 // [ argument word n ] | |
| 116 // ... | |
| 117 // -12 [ argument word 1 ] | |
| 118 // -11 [ saved r15 ] <--- rsp_after_call | |
| 119 // -10 [ saved r14 ] | |
| 120 // -9 [ saved r13 ] | |
| 121 // -8 [ saved r12 ] | |
| 122 // -7 [ saved rbx ] | |
| 123 // -6 [ call wrapper ] | |
| 124 // -5 [ result ] | |
| 125 // -4 [ result type ] | |
| 126 // -3 [ method ] | |
| 127 // -2 [ entry point ] | |
| 128 // -1 [ parameters ] | |
| 129 // 0 [ saved rbp ] <--- rbp | |
| 130 // 1 [ return address ] | |
| 131 // 2 [ parameter size ] | |
| 132 // 3 [ thread ] | |
| 133 // | |
| 134 // Windows Arguments: | |
| 135 // c_rarg0: call wrapper address address | |
| 136 // c_rarg1: result address | |
| 137 // c_rarg2: result type BasicType | |
| 138 // c_rarg3: method methodOop | |
| 139 // 48(rbp): (interpreter) entry point address | |
| 140 // 56(rbp): parameters intptr_t* | |
| 141 // 64(rbp): parameter size (in words) int | |
| 142 // 72(rbp): thread Thread* | |
| 143 // | |
| 144 // [ return_from_Java ] <--- rsp | |
| 145 // [ argument word n ] | |
| 146 // ... | |
| 147 // -8 [ argument word 1 ] | |
| 148 // -7 [ saved r15 ] <--- rsp_after_call | |
| 149 // -6 [ saved r14 ] | |
| 150 // -5 [ saved r13 ] | |
| 151 // -4 [ saved r12 ] | |
| 152 // -3 [ saved rdi ] | |
| 153 // -2 [ saved rsi ] | |
| 154 // -1 [ saved rbx ] | |
| 155 // 0 [ saved rbp ] <--- rbp | |
| 156 // 1 [ return address ] | |
| 157 // 2 [ call wrapper ] | |
| 158 // 3 [ result ] | |
| 159 // 4 [ result type ] | |
| 160 // 5 [ method ] | |
| 161 // 6 [ entry point ] | |
| 162 // 7 [ parameters ] | |
| 163 // 8 [ parameter size ] | |
| 164 // 9 [ thread ] | |
| 165 // | |
| 166 // Windows reserves the callers stack space for arguments 1-4. | |
| 167 // We spill c_rarg0-c_rarg3 to this space. | |
| 168 | |
| 169 // Call stub stack layout word offsets from rbp | |
| 170 enum call_stub_layout { | |
| 171 #ifdef _WIN64 | |
| 172 rsp_after_call_off = -7, | |
| 173 r15_off = rsp_after_call_off, | |
| 174 r14_off = -6, | |
| 175 r13_off = -5, | |
| 176 r12_off = -4, | |
| 177 rdi_off = -3, | |
| 178 rsi_off = -2, | |
| 179 rbx_off = -1, | |
| 180 rbp_off = 0, | |
| 181 retaddr_off = 1, | |
| 182 call_wrapper_off = 2, | |
| 183 result_off = 3, | |
| 184 result_type_off = 4, | |
| 185 method_off = 5, | |
| 186 entry_point_off = 6, | |
| 187 parameters_off = 7, | |
| 188 parameter_size_off = 8, | |
| 189 thread_off = 9 | |
| 190 #else | |
| 191 rsp_after_call_off = -12, | |
| 192 mxcsr_off = rsp_after_call_off, | |
| 193 r15_off = -11, | |
| 194 r14_off = -10, | |
| 195 r13_off = -9, | |
| 196 r12_off = -8, | |
| 197 rbx_off = -7, | |
| 198 call_wrapper_off = -6, | |
| 199 result_off = -5, | |
| 200 result_type_off = -4, | |
| 201 method_off = -3, | |
| 202 entry_point_off = -2, | |
| 203 parameters_off = -1, | |
| 204 rbp_off = 0, | |
| 205 retaddr_off = 1, | |
| 206 parameter_size_off = 2, | |
| 207 thread_off = 3 | |
| 208 #endif | |
| 209 }; | |
| 210 | |
| 211 address generate_call_stub(address& return_address) { | |
| 212 assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 && | |
| 213 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off, | |
| 214 "adjust this code"); | |
| 215 StubCodeMark mark(this, "StubRoutines", "call_stub"); | |
| 216 address start = __ pc(); | |
| 217 | |
| 218 // same as in generate_catch_exception()! | |
| 219 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize); | |
| 220 | |
| 221 const Address call_wrapper (rbp, call_wrapper_off * wordSize); | |
| 222 const Address result (rbp, result_off * wordSize); | |
| 223 const Address result_type (rbp, result_type_off * wordSize); | |
| 224 const Address method (rbp, method_off * wordSize); | |
| 225 const Address entry_point (rbp, entry_point_off * wordSize); | |
| 226 const Address parameters (rbp, parameters_off * wordSize); | |
| 227 const Address parameter_size(rbp, parameter_size_off * wordSize); | |
| 228 | |
| 229 // same as in generate_catch_exception()! | |
| 230 const Address thread (rbp, thread_off * wordSize); | |
| 231 | |
| 232 const Address r15_save(rbp, r15_off * wordSize); | |
| 233 const Address r14_save(rbp, r14_off * wordSize); | |
| 234 const Address r13_save(rbp, r13_off * wordSize); | |
| 235 const Address r12_save(rbp, r12_off * wordSize); | |
| 236 const Address rbx_save(rbp, rbx_off * wordSize); | |
| 237 | |
| 238 // stub code | |
| 239 __ enter(); | |
| 304 | 240 __ subptr(rsp, -rsp_after_call_off * wordSize); |
| 0 | 241 |
| 242 // save register parameters | |
| 243 #ifndef _WIN64 | |
| 304 | 244 __ movptr(parameters, c_rarg5); // parameters |
| 245 __ movptr(entry_point, c_rarg4); // entry_point | |
| 0 | 246 #endif |
| 247 | |
| 304 | 248 __ movptr(method, c_rarg3); // method |
| 249 __ movl(result_type, c_rarg2); // result type | |
| 250 __ movptr(result, c_rarg1); // result | |
| 251 __ movptr(call_wrapper, c_rarg0); // call wrapper | |
| 0 | 252 |
| 253 // save regs belonging to calling function | |
| 304 | 254 __ movptr(rbx_save, rbx); |
| 255 __ movptr(r12_save, r12); | |
| 256 __ movptr(r13_save, r13); | |
| 257 __ movptr(r14_save, r14); | |
| 258 __ movptr(r15_save, r15); | |
| 0 | 259 |
| 260 #ifdef _WIN64 | |
| 261 const Address rdi_save(rbp, rdi_off * wordSize); | |
| 262 const Address rsi_save(rbp, rsi_off * wordSize); | |
| 263 | |
| 304 | 264 __ movptr(rsi_save, rsi); |
| 265 __ movptr(rdi_save, rdi); | |
| 0 | 266 #else |
| 267 const Address mxcsr_save(rbp, mxcsr_off * wordSize); | |
| 268 { | |
| 269 Label skip_ldmx; | |
| 270 __ stmxcsr(mxcsr_save); | |
| 271 __ movl(rax, mxcsr_save); | |
| 272 __ andl(rax, MXCSR_MASK); // Only check control and mask bits | |
| 304 | 273 ExternalAddress mxcsr_std(StubRoutines::x86::mxcsr_std()); |
| 0 | 274 __ cmp32(rax, mxcsr_std); |
| 275 __ jcc(Assembler::equal, skip_ldmx); | |
| 276 __ ldmxcsr(mxcsr_std); | |
| 277 __ bind(skip_ldmx); | |
| 278 } | |
| 279 #endif | |
| 280 | |
| 281 // Load up thread register | |
| 304 | 282 __ movptr(r15_thread, thread); |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
283 __ reinit_heapbase(); |
| 0 | 284 |
| 285 #ifdef ASSERT | |
| 286 // make sure we have no pending exceptions | |
| 287 { | |
| 288 Label L; | |
| 304 | 289 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); |
| 0 | 290 __ jcc(Assembler::equal, L); |
| 291 __ stop("StubRoutines::call_stub: entered with pending exception"); | |
| 292 __ bind(L); | |
| 293 } | |
| 294 #endif | |
| 295 | |
| 296 // pass parameters if any | |
| 297 BLOCK_COMMENT("pass parameters if any"); | |
| 298 Label parameters_done; | |
| 299 __ movl(c_rarg3, parameter_size); | |
| 300 __ testl(c_rarg3, c_rarg3); | |
| 301 __ jcc(Assembler::zero, parameters_done); | |
| 302 | |
| 303 Label loop; | |
| 304 | 304 __ movptr(c_rarg2, parameters); // parameter pointer |
| 305 __ movl(c_rarg1, c_rarg3); // parameter counter is in c_rarg1 | |
| 0 | 306 __ BIND(loop); |
| 304 | 307 __ movptr(rax, Address(c_rarg2, 0));// get parameter |
| 308 __ addptr(c_rarg2, wordSize); // advance to next parameter | |
| 309 __ decrementl(c_rarg1); // decrement counter | |
| 310 __ push(rax); // pass parameter | |
| 0 | 311 __ jcc(Assembler::notZero, loop); |
| 312 | |
| 313 // call Java function | |
| 314 __ BIND(parameters_done); | |
| 304 | 315 __ movptr(rbx, method); // get methodOop |
| 316 __ movptr(c_rarg1, entry_point); // get entry_point | |
| 317 __ mov(r13, rsp); // set sender sp | |
| 0 | 318 BLOCK_COMMENT("call Java function"); |
| 319 __ call(c_rarg1); | |
| 320 | |
| 321 BLOCK_COMMENT("call_stub_return_address:"); | |
| 322 return_address = __ pc(); | |
| 323 | |
| 324 // store result depending on type (everything that is not | |
| 325 // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT) | |
| 304 | 326 __ movptr(c_rarg0, result); |
| 0 | 327 Label is_long, is_float, is_double, exit; |
| 328 __ movl(c_rarg1, result_type); | |
| 329 __ cmpl(c_rarg1, T_OBJECT); | |
| 330 __ jcc(Assembler::equal, is_long); | |
| 331 __ cmpl(c_rarg1, T_LONG); | |
| 332 __ jcc(Assembler::equal, is_long); | |
| 333 __ cmpl(c_rarg1, T_FLOAT); | |
| 334 __ jcc(Assembler::equal, is_float); | |
| 335 __ cmpl(c_rarg1, T_DOUBLE); | |
| 336 __ jcc(Assembler::equal, is_double); | |
| 337 | |
| 338 // handle T_INT case | |
| 339 __ movl(Address(c_rarg0, 0), rax); | |
| 340 | |
| 341 __ BIND(exit); | |
| 342 | |
| 343 // pop parameters | |
| 304 | 344 __ lea(rsp, rsp_after_call); |
| 0 | 345 |
| 346 #ifdef ASSERT | |
| 347 // verify that threads correspond | |
| 348 { | |
| 349 Label L, S; | |
| 304 | 350 __ cmpptr(r15_thread, thread); |
| 0 | 351 __ jcc(Assembler::notEqual, S); |
| 352 __ get_thread(rbx); | |
| 304 | 353 __ cmpptr(r15_thread, rbx); |
| 0 | 354 __ jcc(Assembler::equal, L); |
| 355 __ bind(S); | |
| 356 __ jcc(Assembler::equal, L); | |
| 357 __ stop("StubRoutines::call_stub: threads must correspond"); | |
| 358 __ bind(L); | |
| 359 } | |
| 360 #endif | |
| 361 | |
| 362 // restore regs belonging to calling function | |
| 304 | 363 __ movptr(r15, r15_save); |
| 364 __ movptr(r14, r14_save); | |
| 365 __ movptr(r13, r13_save); | |
| 366 __ movptr(r12, r12_save); | |
| 367 __ movptr(rbx, rbx_save); | |
| 0 | 368 |
| 369 #ifdef _WIN64 | |
| 304 | 370 __ movptr(rdi, rdi_save); |
| 371 __ movptr(rsi, rsi_save); | |
| 0 | 372 #else |
| 373 __ ldmxcsr(mxcsr_save); | |
| 374 #endif | |
| 375 | |
| 376 // restore rsp | |
| 304 | 377 __ addptr(rsp, -rsp_after_call_off * wordSize); |
| 0 | 378 |
| 379 // return | |
| 304 | 380 __ pop(rbp); |
| 0 | 381 __ ret(0); |
| 382 | |
| 383 // handle return types different from T_INT | |
| 384 __ BIND(is_long); | |
| 385 __ movq(Address(c_rarg0, 0), rax); | |
| 386 __ jmp(exit); | |
| 387 | |
| 388 __ BIND(is_float); | |
| 389 __ movflt(Address(c_rarg0, 0), xmm0); | |
| 390 __ jmp(exit); | |
| 391 | |
| 392 __ BIND(is_double); | |
| 393 __ movdbl(Address(c_rarg0, 0), xmm0); | |
| 394 __ jmp(exit); | |
| 395 | |
| 396 return start; | |
| 397 } | |
| 398 | |
| 399 // Return point for a Java call if there's an exception thrown in | |
| 400 // Java code. The exception is caught and transformed into a | |
| 401 // pending exception stored in JavaThread that can be tested from | |
| 402 // within the VM. | |
| 403 // | |
| 404 // Note: Usually the parameters are removed by the callee. In case | |
| 405 // of an exception crossing an activation frame boundary, that is | |
| 406 // not the case if the callee is compiled code => need to setup the | |
| 407 // rsp. | |
| 408 // | |
| 409 // rax: exception oop | |
| 410 | |
| 411 address generate_catch_exception() { | |
| 412 StubCodeMark mark(this, "StubRoutines", "catch_exception"); | |
| 413 address start = __ pc(); | |
| 414 | |
| 415 // same as in generate_call_stub(): | |
| 416 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize); | |
| 417 const Address thread (rbp, thread_off * wordSize); | |
| 418 | |
| 419 #ifdef ASSERT | |
| 420 // verify that threads correspond | |
| 421 { | |
| 422 Label L, S; | |
| 304 | 423 __ cmpptr(r15_thread, thread); |
| 0 | 424 __ jcc(Assembler::notEqual, S); |
| 425 __ get_thread(rbx); | |
| 304 | 426 __ cmpptr(r15_thread, rbx); |
| 0 | 427 __ jcc(Assembler::equal, L); |
| 428 __ bind(S); | |
| 429 __ stop("StubRoutines::catch_exception: threads must correspond"); | |
| 430 __ bind(L); | |
| 431 } | |
| 432 #endif | |
| 433 | |
| 434 // set pending exception | |
| 435 __ verify_oop(rax); | |
| 436 | |
| 304 | 437 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax); |
| 0 | 438 __ lea(rscratch1, ExternalAddress((address)__FILE__)); |
| 304 | 439 __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1); |
| 0 | 440 __ movl(Address(r15_thread, Thread::exception_line_offset()), (int) __LINE__); |
| 441 | |
| 442 // complete return to VM | |
| 443 assert(StubRoutines::_call_stub_return_address != NULL, | |
| 444 "_call_stub_return_address must have been generated before"); | |
| 445 __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address)); | |
| 446 | |
| 447 return start; | |
| 448 } | |
| 449 | |
| 450 // Continuation point for runtime calls returning with a pending | |
| 451 // exception. The pending exception check happened in the runtime | |
| 452 // or native call stub. The pending exception in Thread is | |
| 453 // converted into a Java-level exception. | |
| 454 // | |
| 455 // Contract with Java-level exception handlers: | |
| 456 // rax: exception | |
| 457 // rdx: throwing pc | |
| 458 // | |
| 459 // NOTE: At entry of this stub, exception-pc must be on stack !! | |
| 460 | |
| 461 address generate_forward_exception() { | |
| 462 StubCodeMark mark(this, "StubRoutines", "forward exception"); | |
| 463 address start = __ pc(); | |
| 464 | |
| 465 // Upon entry, the sp points to the return address returning into | |
| 466 // Java (interpreted or compiled) code; i.e., the return address | |
| 467 // becomes the throwing pc. | |
| 468 // | |
| 469 // Arguments pushed before the runtime call are still on the stack | |
| 470 // but the exception handler will reset the stack pointer -> | |
| 471 // ignore them. A potential result in registers can be ignored as | |
| 472 // well. | |
| 473 | |
| 474 #ifdef ASSERT | |
| 475 // make sure this code is only executed if there is a pending exception | |
| 476 { | |
| 477 Label L; | |
| 304 | 478 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t) NULL); |
| 0 | 479 __ jcc(Assembler::notEqual, L); |
| 480 __ stop("StubRoutines::forward exception: no pending exception (1)"); | |
| 481 __ bind(L); | |
| 482 } | |
| 483 #endif | |
| 484 | |
| 485 // compute exception handler into rbx | |
| 304 | 486 __ movptr(c_rarg0, Address(rsp, 0)); |
| 0 | 487 BLOCK_COMMENT("call exception_handler_for_return_address"); |
| 488 __ call_VM_leaf(CAST_FROM_FN_PTR(address, | |
| 489 SharedRuntime::exception_handler_for_return_address), | |
| 1295 | 490 r15_thread, c_rarg0); |
| 304 | 491 __ mov(rbx, rax); |
| 0 | 492 |
| 493 // setup rax & rdx, remove return address & clear pending exception | |
| 304 | 494 __ pop(rdx); |
| 495 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset())); | |
|
512
db4caa99ef11
6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents:
405
diff
changeset
|
496 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); |
| 0 | 497 |
| 498 #ifdef ASSERT | |
| 499 // make sure exception is set | |
| 500 { | |
| 501 Label L; | |
| 304 | 502 __ testptr(rax, rax); |
| 0 | 503 __ jcc(Assembler::notEqual, L); |
| 504 __ stop("StubRoutines::forward exception: no pending exception (2)"); | |
| 505 __ bind(L); | |
| 506 } | |
| 507 #endif | |
| 508 | |
| 509 // continue at exception handler (return address removed) | |
| 510 // rax: exception | |
| 511 // rbx: exception handler | |
| 512 // rdx: throwing pc | |
| 513 __ verify_oop(rax); | |
| 514 __ jmp(rbx); | |
| 515 | |
| 516 return start; | |
| 517 } | |
| 518 | |
| 519 // Support for jint atomic::xchg(jint exchange_value, volatile jint* dest) | |
| 520 // | |
| 521 // Arguments : | |
| 522 // c_rarg0: exchange_value | |
| 523 // c_rarg0: dest | |
| 524 // | |
| 525 // Result: | |
| 526 // *dest <- ex, return (orig *dest) | |
| 527 address generate_atomic_xchg() { | |
| 528 StubCodeMark mark(this, "StubRoutines", "atomic_xchg"); | |
| 529 address start = __ pc(); | |
| 530 | |
| 531 __ movl(rax, c_rarg0); // Copy to eax we need a return value anyhow | |
| 532 __ xchgl(rax, Address(c_rarg1, 0)); // automatic LOCK | |
| 533 __ ret(0); | |
| 534 | |
| 535 return start; | |
| 536 } | |
| 537 | |
| 538 // Support for intptr_t atomic::xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest) | |
| 539 // | |
| 540 // Arguments : | |
| 541 // c_rarg0: exchange_value | |
| 542 // c_rarg1: dest | |
| 543 // | |
| 544 // Result: | |
| 545 // *dest <- ex, return (orig *dest) | |
| 546 address generate_atomic_xchg_ptr() { | |
| 547 StubCodeMark mark(this, "StubRoutines", "atomic_xchg_ptr"); | |
| 548 address start = __ pc(); | |
| 549 | |
| 304 | 550 __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow |
| 551 __ xchgptr(rax, Address(c_rarg1, 0)); // automatic LOCK | |
| 0 | 552 __ ret(0); |
| 553 | |
| 554 return start; | |
| 555 } | |
| 556 | |
| 557 // Support for jint atomic::atomic_cmpxchg(jint exchange_value, volatile jint* dest, | |
| 558 // jint compare_value) | |
| 559 // | |
| 560 // Arguments : | |
| 561 // c_rarg0: exchange_value | |
| 562 // c_rarg1: dest | |
| 563 // c_rarg2: compare_value | |
| 564 // | |
| 565 // Result: | |
| 566 // if ( compare_value == *dest ) { | |
| 567 // *dest = exchange_value | |
| 568 // return compare_value; | |
| 569 // else | |
| 570 // return *dest; | |
| 571 address generate_atomic_cmpxchg() { | |
| 572 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg"); | |
| 573 address start = __ pc(); | |
| 574 | |
| 575 __ movl(rax, c_rarg2); | |
| 576 if ( os::is_MP() ) __ lock(); | |
| 577 __ cmpxchgl(c_rarg0, Address(c_rarg1, 0)); | |
| 578 __ ret(0); | |
| 579 | |
| 580 return start; | |
| 581 } | |
| 582 | |
| 583 // Support for jint atomic::atomic_cmpxchg_long(jlong exchange_value, | |
| 584 // volatile jlong* dest, | |
| 585 // jlong compare_value) | |
| 586 // Arguments : | |
| 587 // c_rarg0: exchange_value | |
| 588 // c_rarg1: dest | |
| 589 // c_rarg2: compare_value | |
| 590 // | |
| 591 // Result: | |
| 592 // if ( compare_value == *dest ) { | |
| 593 // *dest = exchange_value | |
| 594 // return compare_value; | |
| 595 // else | |
| 596 // return *dest; | |
| 597 address generate_atomic_cmpxchg_long() { | |
| 598 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_long"); | |
| 599 address start = __ pc(); | |
| 600 | |
| 601 __ movq(rax, c_rarg2); | |
| 602 if ( os::is_MP() ) __ lock(); | |
| 603 __ cmpxchgq(c_rarg0, Address(c_rarg1, 0)); | |
| 604 __ ret(0); | |
| 605 | |
| 606 return start; | |
| 607 } | |
| 608 | |
| 609 // Support for jint atomic::add(jint add_value, volatile jint* dest) | |
| 610 // | |
| 611 // Arguments : | |
| 612 // c_rarg0: add_value | |
| 613 // c_rarg1: dest | |
| 614 // | |
| 615 // Result: | |
| 616 // *dest += add_value | |
| 617 // return *dest; | |
| 618 address generate_atomic_add() { | |
| 619 StubCodeMark mark(this, "StubRoutines", "atomic_add"); | |
| 620 address start = __ pc(); | |
| 621 | |
| 622 __ movl(rax, c_rarg0); | |
| 623 if ( os::is_MP() ) __ lock(); | |
| 624 __ xaddl(Address(c_rarg1, 0), c_rarg0); | |
| 625 __ addl(rax, c_rarg0); | |
| 626 __ ret(0); | |
| 627 | |
| 628 return start; | |
| 629 } | |
| 630 | |
| 631 // Support for intptr_t atomic::add_ptr(intptr_t add_value, volatile intptr_t* dest) | |
| 632 // | |
| 633 // Arguments : | |
| 634 // c_rarg0: add_value | |
| 635 // c_rarg1: dest | |
| 636 // | |
| 637 // Result: | |
| 638 // *dest += add_value | |
| 639 // return *dest; | |
| 640 address generate_atomic_add_ptr() { | |
| 641 StubCodeMark mark(this, "StubRoutines", "atomic_add_ptr"); | |
| 642 address start = __ pc(); | |
| 643 | |
| 304 | 644 __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow |
| 0 | 645 if ( os::is_MP() ) __ lock(); |
| 304 | 646 __ xaddptr(Address(c_rarg1, 0), c_rarg0); |
| 647 __ addptr(rax, c_rarg0); | |
| 0 | 648 __ ret(0); |
| 649 | |
| 650 return start; | |
| 651 } | |
| 652 | |
| 653 // Support for intptr_t OrderAccess::fence() | |
| 654 // | |
| 655 // Arguments : | |
| 656 // | |
| 657 // Result: | |
| 658 address generate_orderaccess_fence() { | |
| 659 StubCodeMark mark(this, "StubRoutines", "orderaccess_fence"); | |
| 660 address start = __ pc(); | |
|
671
d0994e5bebce
6822204: volatile fences should prefer lock:addl to actual mfence instructions
never
parents:
647
diff
changeset
|
661 __ membar(Assembler::StoreLoad); |
| 0 | 662 __ ret(0); |
| 663 | |
| 664 return start; | |
| 665 } | |
| 666 | |
| 667 // Support for intptr_t get_previous_fp() | |
| 668 // | |
| 669 // This routine is used to find the previous frame pointer for the | |
| 670 // caller (current_frame_guess). This is used as part of debugging | |
| 671 // ps() is seemingly lost trying to find frames. | |
| 672 // This code assumes that caller current_frame_guess) has a frame. | |
| 673 address generate_get_previous_fp() { | |
| 674 StubCodeMark mark(this, "StubRoutines", "get_previous_fp"); | |
| 675 const Address old_fp(rbp, 0); | |
| 676 const Address older_fp(rax, 0); | |
| 677 address start = __ pc(); | |
| 678 | |
| 679 __ enter(); | |
| 304 | 680 __ movptr(rax, old_fp); // callers fp |
| 681 __ movptr(rax, older_fp); // the frame for ps() | |
| 682 __ pop(rbp); | |
| 0 | 683 __ ret(0); |
| 684 | |
| 685 return start; | |
| 686 } | |
| 687 | |
| 688 //---------------------------------------------------------------------------------------------------- | |
| 689 // Support for void verify_mxcsr() | |
| 690 // | |
| 691 // This routine is used with -Xcheck:jni to verify that native | |
| 692 // JNI code does not return to Java code without restoring the | |
| 693 // MXCSR register to our expected state. | |
| 694 | |
| 695 address generate_verify_mxcsr() { | |
| 696 StubCodeMark mark(this, "StubRoutines", "verify_mxcsr"); | |
| 697 address start = __ pc(); | |
| 698 | |
| 699 const Address mxcsr_save(rsp, 0); | |
| 700 | |
| 701 if (CheckJNICalls) { | |
| 702 Label ok_ret; | |
| 304 | 703 __ push(rax); |
| 704 __ subptr(rsp, wordSize); // allocate a temp location | |
| 0 | 705 __ stmxcsr(mxcsr_save); |
| 706 __ movl(rax, mxcsr_save); | |
| 707 __ andl(rax, MXCSR_MASK); // Only check control and mask bits | |
| 304 | 708 __ cmpl(rax, *(int *)(StubRoutines::x86::mxcsr_std())); |
| 0 | 709 __ jcc(Assembler::equal, ok_ret); |
| 710 | |
| 711 __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall"); | |
| 712 | |
| 304 | 713 __ ldmxcsr(ExternalAddress(StubRoutines::x86::mxcsr_std())); |
| 0 | 714 |
| 715 __ bind(ok_ret); | |
| 304 | 716 __ addptr(rsp, wordSize); |
| 717 __ pop(rax); | |
| 0 | 718 } |
| 719 | |
| 720 __ ret(0); | |
| 721 | |
| 722 return start; | |
| 723 } | |
| 724 | |
| 725 address generate_f2i_fixup() { | |
| 726 StubCodeMark mark(this, "StubRoutines", "f2i_fixup"); | |
| 727 Address inout(rsp, 5 * wordSize); // return address + 4 saves | |
| 728 | |
| 729 address start = __ pc(); | |
| 730 | |
| 731 Label L; | |
| 732 | |
| 304 | 733 __ push(rax); |
| 734 __ push(c_rarg3); | |
| 735 __ push(c_rarg2); | |
| 736 __ push(c_rarg1); | |
| 0 | 737 |
| 738 __ movl(rax, 0x7f800000); | |
| 739 __ xorl(c_rarg3, c_rarg3); | |
| 740 __ movl(c_rarg2, inout); | |
| 741 __ movl(c_rarg1, c_rarg2); | |
| 742 __ andl(c_rarg1, 0x7fffffff); | |
| 743 __ cmpl(rax, c_rarg1); // NaN? -> 0 | |
| 744 __ jcc(Assembler::negative, L); | |
| 745 __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint | |
| 746 __ movl(c_rarg3, 0x80000000); | |
| 747 __ movl(rax, 0x7fffffff); | |
| 748 __ cmovl(Assembler::positive, c_rarg3, rax); | |
| 749 | |
| 750 __ bind(L); | |
| 304 | 751 __ movptr(inout, c_rarg3); |
| 752 | |
| 753 __ pop(c_rarg1); | |
| 754 __ pop(c_rarg2); | |
| 755 __ pop(c_rarg3); | |
| 756 __ pop(rax); | |
| 0 | 757 |
| 758 __ ret(0); | |
| 759 | |
| 760 return start; | |
| 761 } | |
| 762 | |
| 763 address generate_f2l_fixup() { | |
| 764 StubCodeMark mark(this, "StubRoutines", "f2l_fixup"); | |
| 765 Address inout(rsp, 5 * wordSize); // return address + 4 saves | |
| 766 address start = __ pc(); | |
| 767 | |
| 768 Label L; | |
| 769 | |
| 304 | 770 __ push(rax); |
| 771 __ push(c_rarg3); | |
| 772 __ push(c_rarg2); | |
| 773 __ push(c_rarg1); | |
| 0 | 774 |
| 775 __ movl(rax, 0x7f800000); | |
| 776 __ xorl(c_rarg3, c_rarg3); | |
| 777 __ movl(c_rarg2, inout); | |
| 778 __ movl(c_rarg1, c_rarg2); | |
| 779 __ andl(c_rarg1, 0x7fffffff); | |
| 780 __ cmpl(rax, c_rarg1); // NaN? -> 0 | |
| 781 __ jcc(Assembler::negative, L); | |
| 782 __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong | |
| 783 __ mov64(c_rarg3, 0x8000000000000000); | |
| 784 __ mov64(rax, 0x7fffffffffffffff); | |
| 304 | 785 __ cmov(Assembler::positive, c_rarg3, rax); |
| 0 | 786 |
| 787 __ bind(L); | |
| 304 | 788 __ movptr(inout, c_rarg3); |
| 789 | |
| 790 __ pop(c_rarg1); | |
| 791 __ pop(c_rarg2); | |
| 792 __ pop(c_rarg3); | |
| 793 __ pop(rax); | |
| 0 | 794 |
| 795 __ ret(0); | |
| 796 | |
| 797 return start; | |
| 798 } | |
| 799 | |
| 800 address generate_d2i_fixup() { | |
| 801 StubCodeMark mark(this, "StubRoutines", "d2i_fixup"); | |
| 802 Address inout(rsp, 6 * wordSize); // return address + 5 saves | |
| 803 | |
| 804 address start = __ pc(); | |
| 805 | |
| 806 Label L; | |
| 807 | |
| 304 | 808 __ push(rax); |
| 809 __ push(c_rarg3); | |
| 810 __ push(c_rarg2); | |
| 811 __ push(c_rarg1); | |
| 812 __ push(c_rarg0); | |
| 0 | 813 |
| 814 __ movl(rax, 0x7ff00000); | |
| 815 __ movq(c_rarg2, inout); | |
| 816 __ movl(c_rarg3, c_rarg2); | |
| 304 | 817 __ mov(c_rarg1, c_rarg2); |
| 818 __ mov(c_rarg0, c_rarg2); | |
| 0 | 819 __ negl(c_rarg3); |
| 304 | 820 __ shrptr(c_rarg1, 0x20); |
| 0 | 821 __ orl(c_rarg3, c_rarg2); |
| 822 __ andl(c_rarg1, 0x7fffffff); | |
| 823 __ xorl(c_rarg2, c_rarg2); | |
| 824 __ shrl(c_rarg3, 0x1f); | |
| 825 __ orl(c_rarg1, c_rarg3); | |
| 826 __ cmpl(rax, c_rarg1); | |
| 827 __ jcc(Assembler::negative, L); // NaN -> 0 | |
| 304 | 828 __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint |
| 0 | 829 __ movl(c_rarg2, 0x80000000); |
| 830 __ movl(rax, 0x7fffffff); | |
| 304 | 831 __ cmov(Assembler::positive, c_rarg2, rax); |
| 0 | 832 |
| 833 __ bind(L); | |
| 304 | 834 __ movptr(inout, c_rarg2); |
| 835 | |
| 836 __ pop(c_rarg0); | |
| 837 __ pop(c_rarg1); | |
| 838 __ pop(c_rarg2); | |
| 839 __ pop(c_rarg3); | |
| 840 __ pop(rax); | |
| 0 | 841 |
| 842 __ ret(0); | |
| 843 | |
| 844 return start; | |
| 845 } | |
| 846 | |
| 847 address generate_d2l_fixup() { | |
| 848 StubCodeMark mark(this, "StubRoutines", "d2l_fixup"); | |
| 849 Address inout(rsp, 6 * wordSize); // return address + 5 saves | |
| 850 | |
| 851 address start = __ pc(); | |
| 852 | |
| 853 Label L; | |
| 854 | |
| 304 | 855 __ push(rax); |
| 856 __ push(c_rarg3); | |
| 857 __ push(c_rarg2); | |
| 858 __ push(c_rarg1); | |
| 859 __ push(c_rarg0); | |
| 0 | 860 |
| 861 __ movl(rax, 0x7ff00000); | |
| 862 __ movq(c_rarg2, inout); | |
| 863 __ movl(c_rarg3, c_rarg2); | |
| 304 | 864 __ mov(c_rarg1, c_rarg2); |
| 865 __ mov(c_rarg0, c_rarg2); | |
| 0 | 866 __ negl(c_rarg3); |
| 304 | 867 __ shrptr(c_rarg1, 0x20); |
| 0 | 868 __ orl(c_rarg3, c_rarg2); |
| 869 __ andl(c_rarg1, 0x7fffffff); | |
| 870 __ xorl(c_rarg2, c_rarg2); | |
| 871 __ shrl(c_rarg3, 0x1f); | |
| 872 __ orl(c_rarg1, c_rarg3); | |
| 873 __ cmpl(rax, c_rarg1); | |
| 874 __ jcc(Assembler::negative, L); // NaN -> 0 | |
| 875 __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong | |
| 876 __ mov64(c_rarg2, 0x8000000000000000); | |
| 877 __ mov64(rax, 0x7fffffffffffffff); | |
| 878 __ cmovq(Assembler::positive, c_rarg2, rax); | |
| 879 | |
| 880 __ bind(L); | |
| 881 __ movq(inout, c_rarg2); | |
| 882 | |
| 304 | 883 __ pop(c_rarg0); |
| 884 __ pop(c_rarg1); | |
| 885 __ pop(c_rarg2); | |
| 886 __ pop(c_rarg3); | |
| 887 __ pop(rax); | |
| 0 | 888 |
| 889 __ ret(0); | |
| 890 | |
| 891 return start; | |
| 892 } | |
| 893 | |
| 894 address generate_fp_mask(const char *stub_name, int64_t mask) { | |
| 1365 | 895 __ align(CodeEntryAlignment); |
| 0 | 896 StubCodeMark mark(this, "StubRoutines", stub_name); |
| 897 address start = __ pc(); | |
| 898 | |
| 899 __ emit_data64( mask, relocInfo::none ); | |
| 900 __ emit_data64( mask, relocInfo::none ); | |
| 901 | |
| 902 return start; | |
| 903 } | |
| 904 | |
| 905 // The following routine generates a subroutine to throw an | |
| 906 // asynchronous UnknownError when an unsafe access gets a fault that | |
| 907 // could not be reasonably prevented by the programmer. (Example: | |
| 908 // SIGBUS/OBJERR.) | |
| 909 address generate_handler_for_unsafe_access() { | |
| 910 StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access"); | |
| 911 address start = __ pc(); | |
| 912 | |
| 304 | 913 __ push(0); // hole for return address-to-be |
| 914 __ pusha(); // push registers | |
| 0 | 915 Address next_pc(rsp, RegisterImpl::number_of_registers * BytesPerWord); |
| 916 | |
| 304 | 917 __ subptr(rsp, frame::arg_reg_save_area_bytes); |
| 0 | 918 BLOCK_COMMENT("call handle_unsafe_access"); |
| 919 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, handle_unsafe_access))); | |
| 304 | 920 __ addptr(rsp, frame::arg_reg_save_area_bytes); |
| 921 | |
| 922 __ movptr(next_pc, rax); // stuff next address | |
| 923 __ popa(); | |
| 0 | 924 __ ret(0); // jump to next address |
| 925 | |
| 926 return start; | |
| 927 } | |
| 928 | |
| 929 // Non-destructive plausibility checks for oops | |
| 930 // | |
| 931 // Arguments: | |
| 932 // all args on stack! | |
| 933 // | |
| 934 // Stack after saving c_rarg3: | |
| 935 // [tos + 0]: saved c_rarg3 | |
| 936 // [tos + 1]: saved c_rarg2 | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
937 // [tos + 2]: saved r12 (several TemplateTable methods use it) |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
938 // [tos + 3]: saved flags |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
939 // [tos + 4]: return address |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
940 // * [tos + 5]: error message (char*) |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
941 // * [tos + 6]: object to verify (oop) |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
942 // * [tos + 7]: saved rax - saved by caller and bashed |
| 1583 | 943 // * [tos + 8]: saved r10 (rscratch1) - saved by caller |
| 0 | 944 // * = popped on exit |
| 945 address generate_verify_oop() { | |
| 946 StubCodeMark mark(this, "StubRoutines", "verify_oop"); | |
| 947 address start = __ pc(); | |
| 948 | |
| 949 Label exit, error; | |
| 950 | |
| 304 | 951 __ pushf(); |
| 0 | 952 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr())); |
| 953 | |
| 304 | 954 __ push(r12); |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
955 |
| 0 | 956 // save c_rarg2 and c_rarg3 |
| 304 | 957 __ push(c_rarg2); |
| 958 __ push(c_rarg3); | |
| 0 | 959 |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
960 enum { |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
961 // After previous pushes. |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
962 oop_to_verify = 6 * wordSize, |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
963 saved_rax = 7 * wordSize, |
| 1583 | 964 saved_r10 = 8 * wordSize, |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
965 |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
966 // Before the call to MacroAssembler::debug(), see below. |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
967 return_addr = 16 * wordSize, |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
968 error_msg = 17 * wordSize |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
969 }; |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
970 |
| 0 | 971 // get object |
| 304 | 972 __ movptr(rax, Address(rsp, oop_to_verify)); |
| 0 | 973 |
| 974 // make sure object is 'reasonable' | |
| 304 | 975 __ testptr(rax, rax); |
| 0 | 976 __ jcc(Assembler::zero, exit); // if obj is NULL it is OK |
| 977 // Check if the oop is in the right area of memory | |
| 304 | 978 __ movptr(c_rarg2, rax); |
|
512
db4caa99ef11
6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents:
405
diff
changeset
|
979 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_mask()); |
| 304 | 980 __ andptr(c_rarg2, c_rarg3); |
|
512
db4caa99ef11
6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents:
405
diff
changeset
|
981 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_bits()); |
| 304 | 982 __ cmpptr(c_rarg2, c_rarg3); |
| 0 | 983 __ jcc(Assembler::notZero, error); |
| 984 | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
985 // set r12 to heapbase for load_klass() |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
986 __ reinit_heapbase(); |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
987 |
| 0 | 988 // make sure klass is 'reasonable' |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
989 __ load_klass(rax, rax); // get klass |
| 304 | 990 __ testptr(rax, rax); |
| 0 | 991 __ jcc(Assembler::zero, error); // if klass is NULL it is broken |
| 992 // Check if the klass is in the right area of memory | |
| 304 | 993 __ mov(c_rarg2, rax); |
|
512
db4caa99ef11
6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents:
405
diff
changeset
|
994 __ movptr(c_rarg3, (intptr_t) Universe::verify_klass_mask()); |
| 304 | 995 __ andptr(c_rarg2, c_rarg3); |
|
512
db4caa99ef11
6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents:
405
diff
changeset
|
996 __ movptr(c_rarg3, (intptr_t) Universe::verify_klass_bits()); |
| 304 | 997 __ cmpptr(c_rarg2, c_rarg3); |
| 0 | 998 __ jcc(Assembler::notZero, error); |
| 999 | |
| 1000 // make sure klass' klass is 'reasonable' | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1001 __ load_klass(rax, rax); |
| 304 | 1002 __ testptr(rax, rax); |
| 0 | 1003 __ jcc(Assembler::zero, error); // if klass' klass is NULL it is broken |
| 1004 // Check if the klass' klass is in the right area of memory | |
|
512
db4caa99ef11
6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents:
405
diff
changeset
|
1005 __ movptr(c_rarg3, (intptr_t) Universe::verify_klass_mask()); |
| 304 | 1006 __ andptr(rax, c_rarg3); |
|
512
db4caa99ef11
6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents:
405
diff
changeset
|
1007 __ movptr(c_rarg3, (intptr_t) Universe::verify_klass_bits()); |
| 304 | 1008 __ cmpptr(rax, c_rarg3); |
| 0 | 1009 __ jcc(Assembler::notZero, error); |
| 1010 | |
| 1011 // return if everything seems ok | |
| 1012 __ bind(exit); | |
| 304 | 1013 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back |
| 1583 | 1014 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back |
| 304 | 1015 __ pop(c_rarg3); // restore c_rarg3 |
| 1016 __ pop(c_rarg2); // restore c_rarg2 | |
| 1017 __ pop(r12); // restore r12 | |
| 1018 __ popf(); // restore flags | |
| 1583 | 1019 __ ret(4 * wordSize); // pop caller saved stuff |
| 0 | 1020 |
| 1021 // handle errors | |
| 1022 __ bind(error); | |
| 304 | 1023 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back |
| 1583 | 1024 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back |
| 304 | 1025 __ pop(c_rarg3); // get saved c_rarg3 back |
| 1026 __ pop(c_rarg2); // get saved c_rarg2 back | |
| 1027 __ pop(r12); // get saved r12 back | |
| 1028 __ popf(); // get saved flags off stack -- | |
| 0 | 1029 // will be ignored |
| 1030 | |
| 304 | 1031 __ pusha(); // push registers |
| 0 | 1032 // (rip is already |
| 1033 // already pushed) | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1034 // debug(char* msg, int64_t pc, int64_t regs[]) |
| 0 | 1035 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and |
| 1036 // pushed all the registers, so now the stack looks like: | |
| 1037 // [tos + 0] 16 saved registers | |
| 1038 // [tos + 16] return address | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1039 // * [tos + 17] error message (char*) |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1040 // * [tos + 18] object to verify (oop) |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1041 // * [tos + 19] saved rax - saved by caller and bashed |
| 1583 | 1042 // * [tos + 20] saved r10 (rscratch1) - saved by caller |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1043 // * = popped on exit |
|
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1044 |
| 304 | 1045 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message |
| 1046 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address | |
| 1047 __ movq(c_rarg2, rsp); // pass address of regs on stack | |
| 1048 __ mov(r12, rsp); // remember rsp | |
| 1049 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows | |
| 1050 __ andptr(rsp, -16); // align stack as required by ABI | |
| 0 | 1051 BLOCK_COMMENT("call MacroAssembler::debug"); |
| 304 | 1052 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64))); |
| 1053 __ mov(rsp, r12); // restore rsp | |
| 1054 __ popa(); // pop registers (includes r12) | |
| 1583 | 1055 __ ret(4 * wordSize); // pop caller saved stuff |
| 0 | 1056 |
| 1057 return start; | |
| 1058 } | |
| 1059 | |
| 1060 static address disjoint_byte_copy_entry; | |
| 1061 static address disjoint_short_copy_entry; | |
| 1062 static address disjoint_int_copy_entry; | |
| 1063 static address disjoint_long_copy_entry; | |
| 1064 static address disjoint_oop_copy_entry; | |
| 1065 | |
| 1066 static address byte_copy_entry; | |
| 1067 static address short_copy_entry; | |
| 1068 static address int_copy_entry; | |
| 1069 static address long_copy_entry; | |
| 1070 static address oop_copy_entry; | |
| 1071 | |
| 1072 static address checkcast_copy_entry; | |
| 1073 | |
| 1074 // | |
| 1075 // Verify that a register contains clean 32-bits positive value | |
| 1076 // (high 32-bits are 0) so it could be used in 64-bits shifts. | |
| 1077 // | |
| 1078 // Input: | |
| 1079 // Rint - 32-bits value | |
| 1080 // Rtmp - scratch | |
| 1081 // | |
| 1082 void assert_clean_int(Register Rint, Register Rtmp) { | |
| 1083 #ifdef ASSERT | |
| 1084 Label L; | |
| 1085 assert_different_registers(Rtmp, Rint); | |
| 1086 __ movslq(Rtmp, Rint); | |
| 1087 __ cmpq(Rtmp, Rint); | |
|
124
b130b98db9cf
6689060: Escape Analysis does not work with Compressed Oops
kvn
parents:
113
diff
changeset
|
1088 __ jcc(Assembler::equal, L); |
| 0 | 1089 __ stop("high 32-bits of int value are not 0"); |
| 1090 __ bind(L); | |
| 1091 #endif | |
| 1092 } | |
| 1093 | |
| 1094 // Generate overlap test for array copy stubs | |
| 1095 // | |
| 1096 // Input: | |
| 1097 // c_rarg0 - from | |
| 1098 // c_rarg1 - to | |
| 1099 // c_rarg2 - element count | |
| 1100 // | |
| 1101 // Output: | |
| 1102 // rax - &from[element count - 1] | |
| 1103 // | |
| 1104 void array_overlap_test(address no_overlap_target, Address::ScaleFactor sf) { | |
| 1105 assert(no_overlap_target != NULL, "must be generated"); | |
| 1106 array_overlap_test(no_overlap_target, NULL, sf); | |
| 1107 } | |
| 1108 void array_overlap_test(Label& L_no_overlap, Address::ScaleFactor sf) { | |
| 1109 array_overlap_test(NULL, &L_no_overlap, sf); | |
| 1110 } | |
| 1111 void array_overlap_test(address no_overlap_target, Label* NOLp, Address::ScaleFactor sf) { | |
| 1112 const Register from = c_rarg0; | |
| 1113 const Register to = c_rarg1; | |
| 1114 const Register count = c_rarg2; | |
| 1115 const Register end_from = rax; | |
| 1116 | |
| 304 | 1117 __ cmpptr(to, from); |
| 1118 __ lea(end_from, Address(from, count, sf, 0)); | |
| 0 | 1119 if (NOLp == NULL) { |
| 1120 ExternalAddress no_overlap(no_overlap_target); | |
| 1121 __ jump_cc(Assembler::belowEqual, no_overlap); | |
| 304 | 1122 __ cmpptr(to, end_from); |
| 0 | 1123 __ jump_cc(Assembler::aboveEqual, no_overlap); |
| 1124 } else { | |
| 1125 __ jcc(Assembler::belowEqual, (*NOLp)); | |
| 304 | 1126 __ cmpptr(to, end_from); |
| 0 | 1127 __ jcc(Assembler::aboveEqual, (*NOLp)); |
| 1128 } | |
| 1129 } | |
| 1130 | |
| 1131 // Shuffle first three arg regs on Windows into Linux/Solaris locations. | |
| 1132 // | |
| 1133 // Outputs: | |
| 1134 // rdi - rcx | |
| 1135 // rsi - rdx | |
| 1136 // rdx - r8 | |
| 1137 // rcx - r9 | |
| 1138 // | |
| 1139 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter | |
| 1140 // are non-volatile. r9 and r10 should not be used by the caller. | |
| 1141 // | |
| 1142 void setup_arg_regs(int nargs = 3) { | |
| 1143 const Register saved_rdi = r9; | |
| 1144 const Register saved_rsi = r10; | |
| 1145 assert(nargs == 3 || nargs == 4, "else fix"); | |
| 1146 #ifdef _WIN64 | |
| 1147 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9, | |
| 1148 "unexpected argument registers"); | |
| 1149 if (nargs >= 4) | |
| 304 | 1150 __ mov(rax, r9); // r9 is also saved_rdi |
| 1151 __ movptr(saved_rdi, rdi); | |
| 1152 __ movptr(saved_rsi, rsi); | |
| 1153 __ mov(rdi, rcx); // c_rarg0 | |
| 1154 __ mov(rsi, rdx); // c_rarg1 | |
| 1155 __ mov(rdx, r8); // c_rarg2 | |
| 0 | 1156 if (nargs >= 4) |
| 304 | 1157 __ mov(rcx, rax); // c_rarg3 (via rax) |
| 0 | 1158 #else |
| 1159 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx, | |
| 1160 "unexpected argument registers"); | |
| 1161 #endif | |
| 1162 } | |
| 1163 | |
| 1164 void restore_arg_regs() { | |
| 1165 const Register saved_rdi = r9; | |
| 1166 const Register saved_rsi = r10; | |
| 1167 #ifdef _WIN64 | |
| 304 | 1168 __ movptr(rdi, saved_rdi); |
| 1169 __ movptr(rsi, saved_rsi); | |
| 0 | 1170 #endif |
| 1171 } | |
| 1172 | |
| 1173 // Generate code for an array write pre barrier | |
| 1174 // | |
| 1175 // addr - starting address | |
| 1176 // count - element count | |
| 1177 // | |
| 1178 // Destroy no registers! | |
| 1179 // | |
| 1180 void gen_write_ref_array_pre_barrier(Register addr, Register count) { | |
| 1181 BarrierSet* bs = Universe::heap()->barrier_set(); | |
| 1182 switch (bs->kind()) { | |
| 1183 case BarrierSet::G1SATBCT: | |
| 1184 case BarrierSet::G1SATBCTLogging: | |
| 1185 { | |
| 304 | 1186 __ pusha(); // push registers |
|
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1187 if (count == c_rarg0) { |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1188 if (addr == c_rarg1) { |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1189 // exactly backwards!! |
| 362 | 1190 __ xchgptr(c_rarg1, c_rarg0); |
|
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1191 } else { |
| 362 | 1192 __ movptr(c_rarg1, count); |
| 1193 __ movptr(c_rarg0, addr); | |
|
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1194 } |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1195 |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1196 } else { |
| 362 | 1197 __ movptr(c_rarg0, addr); |
| 1198 __ movptr(c_rarg1, count); | |
|
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1199 } |
|
1192
776fb94f33cc
6918006: G1: spill space must be reserved on the stack for barrier calls on Windows x64
apetrusenko
parents:
1174
diff
changeset
|
1200 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), 2); |
| 304 | 1201 __ popa(); |
| 0 | 1202 } |
| 1203 break; | |
| 1204 case BarrierSet::CardTableModRef: | |
| 1205 case BarrierSet::CardTableExtension: | |
| 1206 case BarrierSet::ModRef: | |
| 1207 break; | |
|
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1208 default: |
| 0 | 1209 ShouldNotReachHere(); |
| 1210 | |
| 1211 } | |
| 1212 } | |
| 1213 | |
| 1214 // | |
| 1215 // Generate code for an array write post barrier | |
| 1216 // | |
| 1217 // Input: | |
| 1218 // start - register containing starting address of destination array | |
| 1219 // end - register containing ending address of destination array | |
| 1220 // scratch - scratch register | |
| 1221 // | |
| 1222 // The input registers are overwritten. | |
| 1223 // The ending address is inclusive. | |
| 1224 void gen_write_ref_array_post_barrier(Register start, Register end, Register scratch) { | |
| 1225 assert_different_registers(start, end, scratch); | |
| 1226 BarrierSet* bs = Universe::heap()->barrier_set(); | |
| 1227 switch (bs->kind()) { | |
| 1228 case BarrierSet::G1SATBCT: | |
| 1229 case BarrierSet::G1SATBCTLogging: | |
| 1230 | |
| 1231 { | |
| 304 | 1232 __ pusha(); // push registers (overkill) |
| 0 | 1233 // must compute element count unless barrier set interface is changed (other platforms supply count) |
| 1234 assert_different_registers(start, end, scratch); | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
671
diff
changeset
|
1235 __ lea(scratch, Address(end, BytesPerHeapOop)); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
671
diff
changeset
|
1236 __ subptr(scratch, start); // subtract start to get #bytes |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
671
diff
changeset
|
1237 __ shrptr(scratch, LogBytesPerHeapOop); // convert to element count |
| 304 | 1238 __ mov(c_rarg0, start); |
| 1239 __ mov(c_rarg1, scratch); | |
|
1192
776fb94f33cc
6918006: G1: spill space must be reserved on the stack for barrier calls on Windows x64
apetrusenko
parents:
1174
diff
changeset
|
1240 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), 2); |
| 304 | 1241 __ popa(); |
| 0 | 1242 } |
| 1243 break; | |
| 1244 case BarrierSet::CardTableModRef: | |
| 1245 case BarrierSet::CardTableExtension: | |
| 1246 { | |
| 1247 CardTableModRefBS* ct = (CardTableModRefBS*)bs; | |
| 1248 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); | |
| 1249 | |
| 1250 Label L_loop; | |
| 1251 | |
| 304 | 1252 __ shrptr(start, CardTableModRefBS::card_shift); |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
671
diff
changeset
|
1253 __ addptr(end, BytesPerHeapOop); |
| 304 | 1254 __ shrptr(end, CardTableModRefBS::card_shift); |
| 1255 __ subptr(end, start); // number of bytes to copy | |
| 0 | 1256 |
|
249
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1257 intptr_t disp = (intptr_t) ct->byte_map_base; |
|
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1258 if (__ is_simm32(disp)) { |
|
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1259 Address cardtable(noreg, noreg, Address::no_scale, disp); |
|
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1260 __ lea(scratch, cardtable); |
|
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1261 } else { |
|
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1262 ExternalAddress cardtable((address)disp); |
|
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1263 __ lea(scratch, cardtable); |
|
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1264 } |
|
910a4cb98e9e
6717457: Internal Error (src/share/vm/code/relocInfo.hpp:1089)
never
parents:
196
diff
changeset
|
1265 |
| 0 | 1266 const Register count = end; // 'end' register contains bytes count now |
| 304 | 1267 __ addptr(start, scratch); |
| 0 | 1268 __ BIND(L_loop); |
| 1269 __ movb(Address(start, count, Address::times_1), 0); | |
| 304 | 1270 __ decrement(count); |
| 0 | 1271 __ jcc(Assembler::greaterEqual, L_loop); |
| 1272 } | |
|
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1273 break; |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1274 default: |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1275 ShouldNotReachHere(); |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1276 |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1277 } |
|
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
124
diff
changeset
|
1278 } |
| 0 | 1279 |
| 405 | 1280 |
| 0 | 1281 // Copy big chunks forward |
| 1282 // | |
| 1283 // Inputs: | |
| 1284 // end_from - source arrays end address | |
| 1285 // end_to - destination array end address | |
| 1286 // qword_count - 64-bits element count, negative | |
| 1287 // to - scratch | |
| 1288 // L_copy_32_bytes - entry label | |
| 1289 // L_copy_8_bytes - exit label | |
| 1290 // | |
| 1291 void copy_32_bytes_forward(Register end_from, Register end_to, | |
| 1292 Register qword_count, Register to, | |
| 1293 Label& L_copy_32_bytes, Label& L_copy_8_bytes) { | |
| 1294 DEBUG_ONLY(__ stop("enter at entry label, not here")); | |
| 1295 Label L_loop; | |
| 1365 | 1296 __ align(OptoLoopAlignment); |
| 0 | 1297 __ BIND(L_loop); |
| 405 | 1298 if(UseUnalignedLoadStores) { |
| 1299 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24)); | |
| 1300 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0); | |
| 1301 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, - 8)); | |
| 1302 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm1); | |
| 1303 | |
| 1304 } else { | |
| 1305 __ movq(to, Address(end_from, qword_count, Address::times_8, -24)); | |
| 1306 __ movq(Address(end_to, qword_count, Address::times_8, -24), to); | |
| 1307 __ movq(to, Address(end_from, qword_count, Address::times_8, -16)); | |
| 1308 __ movq(Address(end_to, qword_count, Address::times_8, -16), to); | |
| 1309 __ movq(to, Address(end_from, qword_count, Address::times_8, - 8)); | |
| 1310 __ movq(Address(end_to, qword_count, Address::times_8, - 8), to); | |
| 1311 __ movq(to, Address(end_from, qword_count, Address::times_8, - 0)); | |
| 1312 __ movq(Address(end_to, qword_count, Address::times_8, - 0), to); | |
| 1313 } | |
| 0 | 1314 __ BIND(L_copy_32_bytes); |
| 304 | 1315 __ addptr(qword_count, 4); |
| 0 | 1316 __ jcc(Assembler::lessEqual, L_loop); |
| 304 | 1317 __ subptr(qword_count, 4); |
| 0 | 1318 __ jcc(Assembler::less, L_copy_8_bytes); // Copy trailing qwords |
| 1319 } | |
| 1320 | |
| 1321 | |
| 1322 // Copy big chunks backward | |
| 1323 // | |
| 1324 // Inputs: | |
| 1325 // from - source arrays address | |
| 1326 // dest - destination array address | |
| 1327 // qword_count - 64-bits element count | |
| 1328 // to - scratch | |
| 1329 // L_copy_32_bytes - entry label | |
| 1330 // L_copy_8_bytes - exit label | |
| 1331 // | |
| 1332 void copy_32_bytes_backward(Register from, Register dest, | |
| 1333 Register qword_count, Register to, | |
| 1334 Label& L_copy_32_bytes, Label& L_copy_8_bytes) { | |
| 1335 DEBUG_ONLY(__ stop("enter at entry label, not here")); | |
| 1336 Label L_loop; | |
| 1365 | 1337 __ align(OptoLoopAlignment); |
| 0 | 1338 __ BIND(L_loop); |
| 405 | 1339 if(UseUnalignedLoadStores) { |
| 1340 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 16)); | |
| 1341 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm0); | |
| 1342 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 0)); | |
| 1343 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm1); | |
| 1344 | |
| 1345 } else { | |
| 1346 __ movq(to, Address(from, qword_count, Address::times_8, 24)); | |
| 1347 __ movq(Address(dest, qword_count, Address::times_8, 24), to); | |
| 1348 __ movq(to, Address(from, qword_count, Address::times_8, 16)); | |
| 1349 __ movq(Address(dest, qword_count, Address::times_8, 16), to); | |
| 1350 __ movq(to, Address(from, qword_count, Address::times_8, 8)); | |
| 1351 __ movq(Address(dest, qword_count, Address::times_8, 8), to); | |
| 1352 __ movq(to, Address(from, qword_count, Address::times_8, 0)); | |
| 1353 __ movq(Address(dest, qword_count, Address::times_8, 0), to); | |
| 1354 } | |
| 0 | 1355 __ BIND(L_copy_32_bytes); |
| 304 | 1356 __ subptr(qword_count, 4); |
| 0 | 1357 __ jcc(Assembler::greaterEqual, L_loop); |
| 304 | 1358 __ addptr(qword_count, 4); |
| 0 | 1359 __ jcc(Assembler::greater, L_copy_8_bytes); // Copy trailing qwords |
| 1360 } | |
| 1361 | |
| 1362 | |
| 1363 // Arguments: | |
| 1364 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary | |
| 1365 // ignored | |
| 1366 // name - stub name string | |
| 1367 // | |
| 1368 // Inputs: | |
| 1369 // c_rarg0 - source array address | |
| 1370 // c_rarg1 - destination array address | |
| 1371 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 1372 // | |
| 1373 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries, | |
| 1374 // we let the hardware handle it. The one to eight bytes within words, | |
| 1375 // dwords or qwords that span cache line boundaries will still be loaded | |
| 1376 // and stored atomically. | |
| 1377 // | |
| 1378 // Side Effects: | |
| 1379 // disjoint_byte_copy_entry is set to the no-overlap entry point | |
| 1380 // used by generate_conjoint_byte_copy(). | |
| 1381 // | |
| 1382 address generate_disjoint_byte_copy(bool aligned, const char *name) { | |
| 1383 __ align(CodeEntryAlignment); | |
| 1384 StubCodeMark mark(this, "StubRoutines", name); | |
| 1385 address start = __ pc(); | |
| 1386 | |
| 1387 Label L_copy_32_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes; | |
| 1388 Label L_copy_byte, L_exit; | |
| 1389 const Register from = rdi; // source array address | |
| 1390 const Register to = rsi; // destination array address | |
| 1391 const Register count = rdx; // elements count | |
| 1392 const Register byte_count = rcx; | |
| 1393 const Register qword_count = count; | |
| 1394 const Register end_from = from; // source array end address | |
| 1395 const Register end_to = to; // destination array end address | |
| 1396 // End pointers are inclusive, and if count is not zero they point | |
| 1397 // to the last unit copied: end_to[0] := end_from[0] | |
| 1398 | |
| 1399 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 1400 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int. | |
| 1401 | |
| 1402 disjoint_byte_copy_entry = __ pc(); | |
| 1403 BLOCK_COMMENT("Entry:"); | |
| 1404 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory) | |
| 1405 | |
| 1406 setup_arg_regs(); // from => rdi, to => rsi, count => rdx | |
| 1407 // r9 and r10 may be used to save non-volatile registers | |
| 1408 | |
| 1409 // 'from', 'to' and 'count' are now valid | |
| 304 | 1410 __ movptr(byte_count, count); |
| 1411 __ shrptr(count, 3); // count => qword_count | |
| 0 | 1412 |
| 1413 // Copy from low to high addresses. Use 'to' as scratch. | |
| 304 | 1414 __ lea(end_from, Address(from, qword_count, Address::times_8, -8)); |
| 1415 __ lea(end_to, Address(to, qword_count, Address::times_8, -8)); | |
| 1416 __ negptr(qword_count); // make the count negative | |
| 0 | 1417 __ jmp(L_copy_32_bytes); |
| 1418 | |
| 1419 // Copy trailing qwords | |
| 1420 __ BIND(L_copy_8_bytes); | |
| 1421 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8)); | |
| 1422 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax); | |
| 304 | 1423 __ increment(qword_count); |
| 0 | 1424 __ jcc(Assembler::notZero, L_copy_8_bytes); |
| 1425 | |
| 1426 // Check for and copy trailing dword | |
| 1427 __ BIND(L_copy_4_bytes); | |
| 304 | 1428 __ testl(byte_count, 4); |
| 0 | 1429 __ jccb(Assembler::zero, L_copy_2_bytes); |
| 1430 __ movl(rax, Address(end_from, 8)); | |
| 1431 __ movl(Address(end_to, 8), rax); | |
| 1432 | |
| 304 | 1433 __ addptr(end_from, 4); |
| 1434 __ addptr(end_to, 4); | |
| 0 | 1435 |
| 1436 // Check for and copy trailing word | |
| 1437 __ BIND(L_copy_2_bytes); | |
| 304 | 1438 __ testl(byte_count, 2); |
| 0 | 1439 __ jccb(Assembler::zero, L_copy_byte); |
| 1440 __ movw(rax, Address(end_from, 8)); | |
| 1441 __ movw(Address(end_to, 8), rax); | |
| 1442 | |
| 304 | 1443 __ addptr(end_from, 2); |
| 1444 __ addptr(end_to, 2); | |
| 0 | 1445 |
| 1446 // Check for and copy trailing byte | |
| 1447 __ BIND(L_copy_byte); | |
| 304 | 1448 __ testl(byte_count, 1); |
| 0 | 1449 __ jccb(Assembler::zero, L_exit); |
| 1450 __ movb(rax, Address(end_from, 8)); | |
| 1451 __ movb(Address(end_to, 8), rax); | |
| 1452 | |
| 1453 __ BIND(L_exit); | |
| 1454 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); | |
| 1455 restore_arg_regs(); | |
| 304 | 1456 __ xorptr(rax, rax); // return 0 |
| 0 | 1457 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1458 __ ret(0); | |
| 1459 | |
| 1460 // Copy in 32-bytes chunks | |
| 1461 copy_32_bytes_forward(end_from, end_to, qword_count, rax, L_copy_32_bytes, L_copy_8_bytes); | |
| 1462 __ jmp(L_copy_4_bytes); | |
| 1463 | |
| 1464 return start; | |
| 1465 } | |
| 1466 | |
| 1467 // Arguments: | |
| 1468 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary | |
| 1469 // ignored | |
| 1470 // name - stub name string | |
| 1471 // | |
| 1472 // Inputs: | |
| 1473 // c_rarg0 - source array address | |
| 1474 // c_rarg1 - destination array address | |
| 1475 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 1476 // | |
| 1477 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries, | |
| 1478 // we let the hardware handle it. The one to eight bytes within words, | |
| 1479 // dwords or qwords that span cache line boundaries will still be loaded | |
| 1480 // and stored atomically. | |
| 1481 // | |
| 1482 address generate_conjoint_byte_copy(bool aligned, const char *name) { | |
| 1483 __ align(CodeEntryAlignment); | |
| 1484 StubCodeMark mark(this, "StubRoutines", name); | |
| 1485 address start = __ pc(); | |
| 1486 | |
| 1487 Label L_copy_32_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes; | |
| 1488 const Register from = rdi; // source array address | |
| 1489 const Register to = rsi; // destination array address | |
| 1490 const Register count = rdx; // elements count | |
| 1491 const Register byte_count = rcx; | |
| 1492 const Register qword_count = count; | |
| 1493 | |
| 1494 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 1495 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int. | |
| 1496 | |
| 1497 byte_copy_entry = __ pc(); | |
| 1498 BLOCK_COMMENT("Entry:"); | |
| 1499 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory) | |
| 1500 | |
| 1501 array_overlap_test(disjoint_byte_copy_entry, Address::times_1); | |
| 1502 setup_arg_regs(); // from => rdi, to => rsi, count => rdx | |
| 1503 // r9 and r10 may be used to save non-volatile registers | |
| 1504 | |
| 1505 // 'from', 'to' and 'count' are now valid | |
| 304 | 1506 __ movptr(byte_count, count); |
| 1507 __ shrptr(count, 3); // count => qword_count | |
| 0 | 1508 |
| 1509 // Copy from high to low addresses. | |
| 1510 | |
| 1511 // Check for and copy trailing byte | |
| 304 | 1512 __ testl(byte_count, 1); |
| 0 | 1513 __ jcc(Assembler::zero, L_copy_2_bytes); |
| 1514 __ movb(rax, Address(from, byte_count, Address::times_1, -1)); | |
| 1515 __ movb(Address(to, byte_count, Address::times_1, -1), rax); | |
| 304 | 1516 __ decrement(byte_count); // Adjust for possible trailing word |
| 0 | 1517 |
| 1518 // Check for and copy trailing word | |
| 1519 __ BIND(L_copy_2_bytes); | |
| 304 | 1520 __ testl(byte_count, 2); |
| 0 | 1521 __ jcc(Assembler::zero, L_copy_4_bytes); |
| 1522 __ movw(rax, Address(from, byte_count, Address::times_1, -2)); | |
| 1523 __ movw(Address(to, byte_count, Address::times_1, -2), rax); | |
| 1524 | |
| 1525 // Check for and copy trailing dword | |
| 1526 __ BIND(L_copy_4_bytes); | |
| 304 | 1527 __ testl(byte_count, 4); |
| 0 | 1528 __ jcc(Assembler::zero, L_copy_32_bytes); |
| 1529 __ movl(rax, Address(from, qword_count, Address::times_8)); | |
| 1530 __ movl(Address(to, qword_count, Address::times_8), rax); | |
| 1531 __ jmp(L_copy_32_bytes); | |
| 1532 | |
| 1533 // Copy trailing qwords | |
| 1534 __ BIND(L_copy_8_bytes); | |
| 1535 __ movq(rax, Address(from, qword_count, Address::times_8, -8)); | |
| 1536 __ movq(Address(to, qword_count, Address::times_8, -8), rax); | |
| 304 | 1537 __ decrement(qword_count); |
| 0 | 1538 __ jcc(Assembler::notZero, L_copy_8_bytes); |
| 1539 | |
| 1540 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); | |
| 1541 restore_arg_regs(); | |
| 304 | 1542 __ xorptr(rax, rax); // return 0 |
| 0 | 1543 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1544 __ ret(0); | |
| 1545 | |
| 1546 // Copy in 32-bytes chunks | |
| 1547 copy_32_bytes_backward(from, to, qword_count, rax, L_copy_32_bytes, L_copy_8_bytes); | |
| 1548 | |
| 1549 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); | |
| 1550 restore_arg_regs(); | |
| 304 | 1551 __ xorptr(rax, rax); // return 0 |
| 0 | 1552 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1553 __ ret(0); | |
| 1554 | |
| 1555 return start; | |
| 1556 } | |
| 1557 | |
| 1558 // Arguments: | |
| 1559 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary | |
| 1560 // ignored | |
| 1561 // name - stub name string | |
| 1562 // | |
| 1563 // Inputs: | |
| 1564 // c_rarg0 - source array address | |
| 1565 // c_rarg1 - destination array address | |
| 1566 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 1567 // | |
| 1568 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we | |
| 1569 // let the hardware handle it. The two or four words within dwords | |
| 1570 // or qwords that span cache line boundaries will still be loaded | |
| 1571 // and stored atomically. | |
| 1572 // | |
| 1573 // Side Effects: | |
| 1574 // disjoint_short_copy_entry is set to the no-overlap entry point | |
| 1575 // used by generate_conjoint_short_copy(). | |
| 1576 // | |
| 1577 address generate_disjoint_short_copy(bool aligned, const char *name) { | |
| 1578 __ align(CodeEntryAlignment); | |
| 1579 StubCodeMark mark(this, "StubRoutines", name); | |
| 1580 address start = __ pc(); | |
| 1581 | |
| 1582 Label L_copy_32_bytes, L_copy_8_bytes, L_copy_4_bytes,L_copy_2_bytes,L_exit; | |
| 1583 const Register from = rdi; // source array address | |
| 1584 const Register to = rsi; // destination array address | |
| 1585 const Register count = rdx; // elements count | |
| 1586 const Register word_count = rcx; | |
| 1587 const Register qword_count = count; | |
| 1588 const Register end_from = from; // source array end address | |
| 1589 const Register end_to = to; // destination array end address | |
| 1590 // End pointers are inclusive, and if count is not zero they point | |
| 1591 // to the last unit copied: end_to[0] := end_from[0] | |
| 1592 | |
| 1593 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 1594 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int. | |
| 1595 | |
| 1596 disjoint_short_copy_entry = __ pc(); | |
| 1597 BLOCK_COMMENT("Entry:"); | |
| 1598 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory) | |
| 1599 | |
| 1600 setup_arg_regs(); // from => rdi, to => rsi, count => rdx | |
| 1601 // r9 and r10 may be used to save non-volatile registers | |
| 1602 | |
| 1603 // 'from', 'to' and 'count' are now valid | |
| 304 | 1604 __ movptr(word_count, count); |
| 1605 __ shrptr(count, 2); // count => qword_count | |
| 0 | 1606 |
| 1607 // Copy from low to high addresses. Use 'to' as scratch. | |
| 304 | 1608 __ lea(end_from, Address(from, qword_count, Address::times_8, -8)); |
| 1609 __ lea(end_to, Address(to, qword_count, Address::times_8, -8)); | |
| 1610 __ negptr(qword_count); | |
| 0 | 1611 __ jmp(L_copy_32_bytes); |
| 1612 | |
| 1613 // Copy trailing qwords | |
| 1614 __ BIND(L_copy_8_bytes); | |
| 1615 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8)); | |
| 1616 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax); | |
| 304 | 1617 __ increment(qword_count); |
| 0 | 1618 __ jcc(Assembler::notZero, L_copy_8_bytes); |
| 1619 | |
| 1620 // Original 'dest' is trashed, so we can't use it as a | |
| 1621 // base register for a possible trailing word copy | |
| 1622 | |
| 1623 // Check for and copy trailing dword | |
| 1624 __ BIND(L_copy_4_bytes); | |
| 304 | 1625 __ testl(word_count, 2); |
| 0 | 1626 __ jccb(Assembler::zero, L_copy_2_bytes); |
| 1627 __ movl(rax, Address(end_from, 8)); | |
| 1628 __ movl(Address(end_to, 8), rax); | |
| 1629 | |
| 304 | 1630 __ addptr(end_from, 4); |
| 1631 __ addptr(end_to, 4); | |
| 0 | 1632 |
| 1633 // Check for and copy trailing word | |
| 1634 __ BIND(L_copy_2_bytes); | |
| 304 | 1635 __ testl(word_count, 1); |
| 0 | 1636 __ jccb(Assembler::zero, L_exit); |
| 1637 __ movw(rax, Address(end_from, 8)); | |
| 1638 __ movw(Address(end_to, 8), rax); | |
| 1639 | |
| 1640 __ BIND(L_exit); | |
| 1641 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); | |
| 1642 restore_arg_regs(); | |
| 304 | 1643 __ xorptr(rax, rax); // return 0 |
| 0 | 1644 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1645 __ ret(0); | |
| 1646 | |
| 1647 // Copy in 32-bytes chunks | |
| 1648 copy_32_bytes_forward(end_from, end_to, qword_count, rax, L_copy_32_bytes, L_copy_8_bytes); | |
| 1649 __ jmp(L_copy_4_bytes); | |
| 1650 | |
| 1651 return start; | |
| 1652 } | |
| 1653 | |
| 1763 | 1654 address generate_fill(BasicType t, bool aligned, const char *name) { |
| 1655 __ align(CodeEntryAlignment); | |
| 1656 StubCodeMark mark(this, "StubRoutines", name); | |
| 1657 address start = __ pc(); | |
| 1658 | |
| 1659 BLOCK_COMMENT("Entry:"); | |
| 1660 | |
| 1661 const Register to = c_rarg0; // source array address | |
| 1662 const Register value = c_rarg1; // value | |
| 1663 const Register count = c_rarg2; // elements count | |
| 1664 | |
| 1665 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 1666 | |
| 1667 __ generate_fill(t, aligned, to, value, count, rax, xmm0); | |
| 1668 | |
| 1669 __ leave(); // required for proper stackwalking of RuntimeStub frame | |
| 1670 __ ret(0); | |
| 1671 return start; | |
| 1672 } | |
| 1673 | |
| 0 | 1674 // Arguments: |
| 1675 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary | |
| 1676 // ignored | |
| 1677 // name - stub name string | |
| 1678 // | |
| 1679 // Inputs: | |
| 1680 // c_rarg0 - source array address | |
| 1681 // c_rarg1 - destination array address | |
| 1682 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 1683 // | |
| 1684 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we | |
| 1685 // let the hardware handle it. The two or four words within dwords | |
| 1686 // or qwords that span cache line boundaries will still be loaded | |
| 1687 // and stored atomically. | |
| 1688 // | |
| 1689 address generate_conjoint_short_copy(bool aligned, const char *name) { | |
| 1690 __ align(CodeEntryAlignment); | |
| 1691 StubCodeMark mark(this, "StubRoutines", name); | |
| 1692 address start = __ pc(); | |
| 1693 | |
| 1694 Label L_copy_32_bytes, L_copy_8_bytes, L_copy_4_bytes; | |
| 1695 const Register from = rdi; // source array address | |
| 1696 const Register to = rsi; // destination array address | |
| 1697 const Register count = rdx; // elements count | |
| 1698 const Register word_count = rcx; | |
| 1699 const Register qword_count = count; | |
| 1700 | |
| 1701 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 1702 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int. | |
| 1703 | |
| 1704 short_copy_entry = __ pc(); | |
| 1705 BLOCK_COMMENT("Entry:"); | |
| 1706 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory) | |
| 1707 | |
| 1708 array_overlap_test(disjoint_short_copy_entry, Address::times_2); | |
| 1709 setup_arg_regs(); // from => rdi, to => rsi, count => rdx | |
| 1710 // r9 and r10 may be used to save non-volatile registers | |
| 1711 | |
| 1712 // 'from', 'to' and 'count' are now valid | |
| 304 | 1713 __ movptr(word_count, count); |
| 1714 __ shrptr(count, 2); // count => qword_count | |
| 0 | 1715 |
| 1716 // Copy from high to low addresses. Use 'to' as scratch. | |
| 1717 | |
| 1718 // Check for and copy trailing word | |
| 304 | 1719 __ testl(word_count, 1); |
| 0 | 1720 __ jccb(Assembler::zero, L_copy_4_bytes); |
| 1721 __ movw(rax, Address(from, word_count, Address::times_2, -2)); | |
| 1722 __ movw(Address(to, word_count, Address::times_2, -2), rax); | |
| 1723 | |
| 1724 // Check for and copy trailing dword | |
| 1725 __ BIND(L_copy_4_bytes); | |
| 304 | 1726 __ testl(word_count, 2); |
| 0 | 1727 __ jcc(Assembler::zero, L_copy_32_bytes); |
| 1728 __ movl(rax, Address(from, qword_count, Address::times_8)); | |
| 1729 __ movl(Address(to, qword_count, Address::times_8), rax); | |
| 1730 __ jmp(L_copy_32_bytes); | |
| 1731 | |
| 1732 // Copy trailing qwords | |
| 1733 __ BIND(L_copy_8_bytes); | |
| 1734 __ movq(rax, Address(from, qword_count, Address::times_8, -8)); | |
| 1735 __ movq(Address(to, qword_count, Address::times_8, -8), rax); | |
| 304 | 1736 __ decrement(qword_count); |
| 0 | 1737 __ jcc(Assembler::notZero, L_copy_8_bytes); |
| 1738 | |
| 1739 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); | |
| 1740 restore_arg_regs(); | |
| 304 | 1741 __ xorptr(rax, rax); // return 0 |
| 0 | 1742 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1743 __ ret(0); | |
| 1744 | |
| 1745 // Copy in 32-bytes chunks | |
| 1746 copy_32_bytes_backward(from, to, qword_count, rax, L_copy_32_bytes, L_copy_8_bytes); | |
| 1747 | |
| 1748 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); | |
| 1749 restore_arg_regs(); | |
| 304 | 1750 __ xorptr(rax, rax); // return 0 |
| 0 | 1751 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1752 __ ret(0); | |
| 1753 | |
| 1754 return start; | |
| 1755 } | |
| 1756 | |
| 1757 // Arguments: | |
| 1758 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary | |
| 1759 // ignored | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1760 // is_oop - true => oop array, so generate store check code |
| 0 | 1761 // name - stub name string |
| 1762 // | |
| 1763 // Inputs: | |
| 1764 // c_rarg0 - source array address | |
| 1765 // c_rarg1 - destination array address | |
| 1766 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 1767 // | |
| 1768 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let | |
| 1769 // the hardware handle it. The two dwords within qwords that span | |
| 1770 // cache line boundaries will still be loaded and stored atomicly. | |
| 1771 // | |
| 1772 // Side Effects: | |
| 1773 // disjoint_int_copy_entry is set to the no-overlap entry point | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1774 // used by generate_conjoint_int_oop_copy(). |
| 0 | 1775 // |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1776 address generate_disjoint_int_oop_copy(bool aligned, bool is_oop, const char *name) { |
| 0 | 1777 __ align(CodeEntryAlignment); |
| 1778 StubCodeMark mark(this, "StubRoutines", name); | |
| 1779 address start = __ pc(); | |
| 1780 | |
| 1781 Label L_copy_32_bytes, L_copy_8_bytes, L_copy_4_bytes, L_exit; | |
| 1782 const Register from = rdi; // source array address | |
| 1783 const Register to = rsi; // destination array address | |
| 1784 const Register count = rdx; // elements count | |
| 1785 const Register dword_count = rcx; | |
| 1786 const Register qword_count = count; | |
| 1787 const Register end_from = from; // source array end address | |
| 1788 const Register end_to = to; // destination array end address | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1789 const Register saved_to = r11; // saved destination array address |
| 0 | 1790 // End pointers are inclusive, and if count is not zero they point |
| 1791 // to the last unit copied: end_to[0] := end_from[0] | |
| 1792 | |
| 1793 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 1794 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int. | |
| 1795 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1796 (is_oop ? disjoint_oop_copy_entry : disjoint_int_copy_entry) = __ pc(); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1797 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1798 if (is_oop) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1799 // no registers are destroyed by this call |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1800 gen_write_ref_array_pre_barrier(/* dest */ c_rarg1, /* count */ c_rarg2); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1801 } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1802 |
| 0 | 1803 BLOCK_COMMENT("Entry:"); |
| 1804 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory) | |
| 1805 | |
| 1806 setup_arg_regs(); // from => rdi, to => rsi, count => rdx | |
| 1807 // r9 and r10 may be used to save non-volatile registers | |
| 1808 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1809 if (is_oop) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1810 __ movq(saved_to, to); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1811 } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1812 |
| 0 | 1813 // 'from', 'to' and 'count' are now valid |
| 304 | 1814 __ movptr(dword_count, count); |
| 1815 __ shrptr(count, 1); // count => qword_count | |
| 0 | 1816 |
| 1817 // Copy from low to high addresses. Use 'to' as scratch. | |
| 304 | 1818 __ lea(end_from, Address(from, qword_count, Address::times_8, -8)); |
| 1819 __ lea(end_to, Address(to, qword_count, Address::times_8, -8)); | |
| 1820 __ negptr(qword_count); | |
| 0 | 1821 __ jmp(L_copy_32_bytes); |
| 1822 | |
| 1823 // Copy trailing qwords | |
| 1824 __ BIND(L_copy_8_bytes); | |
| 1825 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8)); | |
| 1826 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax); | |
| 304 | 1827 __ increment(qword_count); |
| 0 | 1828 __ jcc(Assembler::notZero, L_copy_8_bytes); |
| 1829 | |
| 1830 // Check for and copy trailing dword | |
| 1831 __ BIND(L_copy_4_bytes); | |
| 304 | 1832 __ testl(dword_count, 1); // Only byte test since the value is 0 or 1 |
| 0 | 1833 __ jccb(Assembler::zero, L_exit); |
| 1834 __ movl(rax, Address(end_from, 8)); | |
| 1835 __ movl(Address(end_to, 8), rax); | |
| 1836 | |
| 1837 __ BIND(L_exit); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1838 if (is_oop) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1839 __ leaq(end_to, Address(saved_to, dword_count, Address::times_4, -4)); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1840 gen_write_ref_array_post_barrier(saved_to, end_to, rax); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1841 } |
| 0 | 1842 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); |
| 1843 restore_arg_regs(); | |
| 304 | 1844 __ xorptr(rax, rax); // return 0 |
| 0 | 1845 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1846 __ ret(0); | |
| 1847 | |
| 1848 // Copy 32-bytes chunks | |
| 1849 copy_32_bytes_forward(end_from, end_to, qword_count, rax, L_copy_32_bytes, L_copy_8_bytes); | |
| 1850 __ jmp(L_copy_4_bytes); | |
| 1851 | |
| 1852 return start; | |
| 1853 } | |
| 1854 | |
| 1855 // Arguments: | |
| 1856 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary | |
| 1857 // ignored | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1858 // is_oop - true => oop array, so generate store check code |
| 0 | 1859 // name - stub name string |
| 1860 // | |
| 1861 // Inputs: | |
| 1862 // c_rarg0 - source array address | |
| 1863 // c_rarg1 - destination array address | |
| 1864 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 1865 // | |
| 1866 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let | |
| 1867 // the hardware handle it. The two dwords within qwords that span | |
| 1868 // cache line boundaries will still be loaded and stored atomicly. | |
| 1869 // | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1870 address generate_conjoint_int_oop_copy(bool aligned, bool is_oop, const char *name) { |
| 0 | 1871 __ align(CodeEntryAlignment); |
| 1872 StubCodeMark mark(this, "StubRoutines", name); | |
| 1873 address start = __ pc(); | |
| 1874 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1875 Label L_copy_32_bytes, L_copy_8_bytes, L_copy_2_bytes, L_exit; |
| 0 | 1876 const Register from = rdi; // source array address |
| 1877 const Register to = rsi; // destination array address | |
| 1878 const Register count = rdx; // elements count | |
| 1879 const Register dword_count = rcx; | |
| 1880 const Register qword_count = count; | |
| 1881 | |
| 1882 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 1883 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int. | |
| 1884 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1885 if (is_oop) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1886 // no registers are destroyed by this call |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1887 gen_write_ref_array_pre_barrier(/* dest */ c_rarg1, /* count */ c_rarg2); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1888 } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1889 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1890 (is_oop ? oop_copy_entry : int_copy_entry) = __ pc(); |
| 0 | 1891 BLOCK_COMMENT("Entry:"); |
| 1892 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory) | |
| 1893 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1894 array_overlap_test(is_oop ? disjoint_oop_copy_entry : disjoint_int_copy_entry, |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1895 Address::times_4); |
| 0 | 1896 setup_arg_regs(); // from => rdi, to => rsi, count => rdx |
| 1897 // r9 and r10 may be used to save non-volatile registers | |
| 1898 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1899 assert_clean_int(count, rax); // Make sure 'count' is clean int. |
| 0 | 1900 // 'from', 'to' and 'count' are now valid |
| 304 | 1901 __ movptr(dword_count, count); |
| 1902 __ shrptr(count, 1); // count => qword_count | |
| 0 | 1903 |
| 1904 // Copy from high to low addresses. Use 'to' as scratch. | |
| 1905 | |
| 1906 // Check for and copy trailing dword | |
| 304 | 1907 __ testl(dword_count, 1); |
| 0 | 1908 __ jcc(Assembler::zero, L_copy_32_bytes); |
| 1909 __ movl(rax, Address(from, dword_count, Address::times_4, -4)); | |
| 1910 __ movl(Address(to, dword_count, Address::times_4, -4), rax); | |
| 1911 __ jmp(L_copy_32_bytes); | |
| 1912 | |
| 1913 // Copy trailing qwords | |
| 1914 __ BIND(L_copy_8_bytes); | |
| 1915 __ movq(rax, Address(from, qword_count, Address::times_8, -8)); | |
| 1916 __ movq(Address(to, qword_count, Address::times_8, -8), rax); | |
| 304 | 1917 __ decrement(qword_count); |
| 0 | 1918 __ jcc(Assembler::notZero, L_copy_8_bytes); |
| 1919 | |
| 1920 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1921 if (is_oop) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1922 __ jmp(L_exit); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1923 } |
| 0 | 1924 restore_arg_regs(); |
| 304 | 1925 __ xorptr(rax, rax); // return 0 |
| 0 | 1926 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1927 __ ret(0); | |
| 1928 | |
| 1929 // Copy in 32-bytes chunks | |
| 1930 copy_32_bytes_backward(from, to, qword_count, rax, L_copy_32_bytes, L_copy_8_bytes); | |
| 1931 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1932 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1933 __ bind(L_exit); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1934 if (is_oop) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1935 Register end_to = rdx; |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1936 __ leaq(end_to, Address(to, dword_count, Address::times_4, -4)); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1937 gen_write_ref_array_post_barrier(to, end_to, rax); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1938 } |
| 0 | 1939 restore_arg_regs(); |
| 304 | 1940 __ xorptr(rax, rax); // return 0 |
| 0 | 1941 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 1942 __ ret(0); | |
| 1943 | |
| 1944 return start; | |
| 1945 } | |
| 1946 | |
| 1947 // Arguments: | |
| 1948 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes | |
| 1949 // ignored | |
| 1950 // is_oop - true => oop array, so generate store check code | |
| 1951 // name - stub name string | |
| 1952 // | |
| 1953 // Inputs: | |
| 1954 // c_rarg0 - source array address | |
| 1955 // c_rarg1 - destination array address | |
| 1956 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 1957 // | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
1958 // Side Effects: |
| 0 | 1959 // disjoint_oop_copy_entry or disjoint_long_copy_entry is set to the |
| 1960 // no-overlap entry point used by generate_conjoint_long_oop_copy(). | |
| 1961 // | |
| 1962 address generate_disjoint_long_oop_copy(bool aligned, bool is_oop, const char *name) { | |
| 1963 __ align(CodeEntryAlignment); | |
| 1964 StubCodeMark mark(this, "StubRoutines", name); | |
| 1965 address start = __ pc(); | |
| 1966 | |
| 1967 Label L_copy_32_bytes, L_copy_8_bytes, L_exit; | |
| 1968 const Register from = rdi; // source array address | |
| 1969 const Register to = rsi; // destination array address | |
| 1970 const Register qword_count = rdx; // elements count | |
| 1971 const Register end_from = from; // source array end address | |
| 1972 const Register end_to = rcx; // destination array end address | |
| 1973 const Register saved_to = to; | |
| 1974 // End pointers are inclusive, and if count is not zero they point | |
| 1975 // to the last unit copied: end_to[0] := end_from[0] | |
| 1976 | |
| 1977 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 1978 // Save no-overlap entry point for generate_conjoint_long_oop_copy() | |
| 1979 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int. | |
| 1980 | |
| 1981 if (is_oop) { | |
| 1982 disjoint_oop_copy_entry = __ pc(); | |
| 1983 // no registers are destroyed by this call | |
| 1984 gen_write_ref_array_pre_barrier(/* dest */ c_rarg1, /* count */ c_rarg2); | |
| 1985 } else { | |
| 1986 disjoint_long_copy_entry = __ pc(); | |
| 1987 } | |
| 1988 BLOCK_COMMENT("Entry:"); | |
| 1989 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory) | |
| 1990 | |
| 1991 setup_arg_regs(); // from => rdi, to => rsi, count => rdx | |
| 1992 // r9 and r10 may be used to save non-volatile registers | |
| 1993 | |
| 1994 // 'from', 'to' and 'qword_count' are now valid | |
| 1995 | |
| 1996 // Copy from low to high addresses. Use 'to' as scratch. | |
| 304 | 1997 __ lea(end_from, Address(from, qword_count, Address::times_8, -8)); |
| 1998 __ lea(end_to, Address(to, qword_count, Address::times_8, -8)); | |
| 1999 __ negptr(qword_count); | |
| 0 | 2000 __ jmp(L_copy_32_bytes); |
| 2001 | |
| 2002 // Copy trailing qwords | |
| 2003 __ BIND(L_copy_8_bytes); | |
| 2004 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8)); | |
| 2005 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax); | |
| 304 | 2006 __ increment(qword_count); |
| 0 | 2007 __ jcc(Assembler::notZero, L_copy_8_bytes); |
| 2008 | |
| 2009 if (is_oop) { | |
| 2010 __ jmp(L_exit); | |
| 2011 } else { | |
| 2012 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); | |
| 2013 restore_arg_regs(); | |
| 304 | 2014 __ xorptr(rax, rax); // return 0 |
| 0 | 2015 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 2016 __ ret(0); | |
| 2017 } | |
| 2018 | |
| 2019 // Copy 64-byte chunks | |
| 2020 copy_32_bytes_forward(end_from, end_to, qword_count, rax, L_copy_32_bytes, L_copy_8_bytes); | |
| 2021 | |
| 2022 if (is_oop) { | |
| 2023 __ BIND(L_exit); | |
| 2024 gen_write_ref_array_post_barrier(saved_to, end_to, rax); | |
| 2025 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); | |
| 2026 } else { | |
| 2027 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); | |
| 2028 } | |
| 2029 restore_arg_regs(); | |
| 304 | 2030 __ xorptr(rax, rax); // return 0 |
| 0 | 2031 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 2032 __ ret(0); | |
| 2033 | |
| 2034 return start; | |
| 2035 } | |
| 2036 | |
| 2037 // Arguments: | |
| 2038 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes | |
| 2039 // ignored | |
| 2040 // is_oop - true => oop array, so generate store check code | |
| 2041 // name - stub name string | |
| 2042 // | |
| 2043 // Inputs: | |
| 2044 // c_rarg0 - source array address | |
| 2045 // c_rarg1 - destination array address | |
| 2046 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 2047 // | |
| 2048 address generate_conjoint_long_oop_copy(bool aligned, bool is_oop, const char *name) { | |
| 2049 __ align(CodeEntryAlignment); | |
| 2050 StubCodeMark mark(this, "StubRoutines", name); | |
| 2051 address start = __ pc(); | |
| 2052 | |
| 2053 Label L_copy_32_bytes, L_copy_8_bytes, L_exit; | |
| 2054 const Register from = rdi; // source array address | |
| 2055 const Register to = rsi; // destination array address | |
| 2056 const Register qword_count = rdx; // elements count | |
| 2057 const Register saved_count = rcx; | |
| 2058 | |
| 2059 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 2060 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int. | |
| 2061 | |
| 2062 address disjoint_copy_entry = NULL; | |
| 2063 if (is_oop) { | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2064 assert(!UseCompressedOops, "shouldn't be called for compressed oops"); |
| 0 | 2065 disjoint_copy_entry = disjoint_oop_copy_entry; |
| 2066 oop_copy_entry = __ pc(); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2067 array_overlap_test(disjoint_oop_copy_entry, Address::times_8); |
| 0 | 2068 } else { |
| 2069 disjoint_copy_entry = disjoint_long_copy_entry; | |
| 2070 long_copy_entry = __ pc(); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2071 array_overlap_test(disjoint_long_copy_entry, Address::times_8); |
| 0 | 2072 } |
| 2073 BLOCK_COMMENT("Entry:"); | |
| 2074 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory) | |
| 2075 | |
| 2076 array_overlap_test(disjoint_copy_entry, Address::times_8); | |
| 2077 setup_arg_regs(); // from => rdi, to => rsi, count => rdx | |
| 2078 // r9 and r10 may be used to save non-volatile registers | |
| 2079 | |
| 2080 // 'from', 'to' and 'qword_count' are now valid | |
| 2081 | |
| 2082 if (is_oop) { | |
| 2083 // Save to and count for store barrier | |
| 304 | 2084 __ movptr(saved_count, qword_count); |
| 0 | 2085 // No registers are destroyed by this call |
| 2086 gen_write_ref_array_pre_barrier(to, saved_count); | |
| 2087 } | |
| 2088 | |
| 2089 __ jmp(L_copy_32_bytes); | |
| 2090 | |
| 2091 // Copy trailing qwords | |
| 2092 __ BIND(L_copy_8_bytes); | |
| 2093 __ movq(rax, Address(from, qword_count, Address::times_8, -8)); | |
| 2094 __ movq(Address(to, qword_count, Address::times_8, -8), rax); | |
| 304 | 2095 __ decrement(qword_count); |
| 0 | 2096 __ jcc(Assembler::notZero, L_copy_8_bytes); |
| 2097 | |
| 2098 if (is_oop) { | |
| 2099 __ jmp(L_exit); | |
| 2100 } else { | |
| 2101 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); | |
| 2102 restore_arg_regs(); | |
| 304 | 2103 __ xorptr(rax, rax); // return 0 |
| 0 | 2104 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 2105 __ ret(0); | |
| 2106 } | |
| 2107 | |
| 2108 // Copy in 32-bytes chunks | |
| 2109 copy_32_bytes_backward(from, to, qword_count, rax, L_copy_32_bytes, L_copy_8_bytes); | |
| 2110 | |
| 2111 if (is_oop) { | |
| 2112 __ BIND(L_exit); | |
| 304 | 2113 __ lea(rcx, Address(to, saved_count, Address::times_8, -8)); |
| 0 | 2114 gen_write_ref_array_post_barrier(to, rcx, rax); |
| 2115 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); | |
| 2116 } else { | |
| 2117 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); | |
| 2118 } | |
| 2119 restore_arg_regs(); | |
| 304 | 2120 __ xorptr(rax, rax); // return 0 |
| 0 | 2121 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 2122 __ ret(0); | |
| 2123 | |
| 2124 return start; | |
| 2125 } | |
| 2126 | |
| 2127 | |
| 2128 // Helper for generating a dynamic type check. | |
| 2129 // Smashes no registers. | |
| 2130 void generate_type_check(Register sub_klass, | |
| 2131 Register super_check_offset, | |
| 2132 Register super_klass, | |
| 2133 Label& L_success) { | |
| 2134 assert_different_registers(sub_klass, super_check_offset, super_klass); | |
| 2135 | |
| 2136 BLOCK_COMMENT("type_check:"); | |
| 2137 | |
| 2138 Label L_miss; | |
| 2139 | |
|
644
c517646eef23
6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents:
512
diff
changeset
|
2140 __ check_klass_subtype_fast_path(sub_klass, super_klass, noreg, &L_success, &L_miss, NULL, |
|
c517646eef23
6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents:
512
diff
changeset
|
2141 super_check_offset); |
|
c517646eef23
6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents:
512
diff
changeset
|
2142 __ check_klass_subtype_slow_path(sub_klass, super_klass, noreg, noreg, &L_success, NULL); |
| 0 | 2143 |
| 2144 // Fall through on failure! | |
| 2145 __ BIND(L_miss); | |
| 2146 } | |
| 2147 | |
| 2148 // | |
| 2149 // Generate checkcasting array copy stub | |
| 2150 // | |
| 2151 // Input: | |
| 2152 // c_rarg0 - source array address | |
| 2153 // c_rarg1 - destination array address | |
| 2154 // c_rarg2 - element count, treated as ssize_t, can be zero | |
| 2155 // c_rarg3 - size_t ckoff (super_check_offset) | |
| 2156 // not Win64 | |
| 2157 // c_rarg4 - oop ckval (super_klass) | |
| 2158 // Win64 | |
| 2159 // rsp+40 - oop ckval (super_klass) | |
| 2160 // | |
| 2161 // Output: | |
| 2162 // rax == 0 - success | |
| 2163 // rax == -1^K - failure, where K is partial transfer count | |
| 2164 // | |
| 2165 address generate_checkcast_copy(const char *name) { | |
| 2166 | |
| 2167 Label L_load_element, L_store_element, L_do_card_marks, L_done; | |
| 2168 | |
| 2169 // Input registers (after setup_arg_regs) | |
| 2170 const Register from = rdi; // source array address | |
| 2171 const Register to = rsi; // destination array address | |
| 2172 const Register length = rdx; // elements count | |
| 2173 const Register ckoff = rcx; // super_check_offset | |
| 2174 const Register ckval = r8; // super_klass | |
| 2175 | |
| 2176 // Registers used as temps (r13, r14 are save-on-entry) | |
| 2177 const Register end_from = from; // source array end address | |
| 2178 const Register end_to = r13; // destination array end address | |
| 2179 const Register count = rdx; // -(count_remaining) | |
| 2180 const Register r14_length = r14; // saved copy of length | |
| 2181 // End pointers are inclusive, and if length is not zero they point | |
| 2182 // to the last unit copied: end_to[0] := end_from[0] | |
| 2183 | |
| 2184 const Register rax_oop = rax; // actual oop copied | |
| 2185 const Register r11_klass = r11; // oop._klass | |
| 2186 | |
| 2187 //--------------------------------------------------------------- | |
| 2188 // Assembler stub will be used for this call to arraycopy | |
| 2189 // if the two arrays are subtypes of Object[] but the | |
| 2190 // destination array type is not equal to or a supertype | |
| 2191 // of the source type. Each element must be separately | |
| 2192 // checked. | |
| 2193 | |
| 2194 __ align(CodeEntryAlignment); | |
| 2195 StubCodeMark mark(this, "StubRoutines", name); | |
| 2196 address start = __ pc(); | |
| 2197 | |
| 2198 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 2199 | |
| 2200 checkcast_copy_entry = __ pc(); | |
| 2201 BLOCK_COMMENT("Entry:"); | |
| 2202 | |
| 2203 #ifdef ASSERT | |
| 2204 // caller guarantees that the arrays really are different | |
| 2205 // otherwise, we would have to make conjoint checks | |
| 2206 { Label L; | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2207 array_overlap_test(L, TIMES_OOP); |
| 0 | 2208 __ stop("checkcast_copy within a single array"); |
| 2209 __ bind(L); | |
| 2210 } | |
| 2211 #endif //ASSERT | |
| 2212 | |
| 2213 // allocate spill slots for r13, r14 | |
| 2214 enum { | |
| 2215 saved_r13_offset, | |
| 2216 saved_r14_offset, | |
| 2217 saved_rbp_offset, | |
| 2218 saved_rip_offset, | |
| 2219 saved_rarg0_offset | |
| 2220 }; | |
| 304 | 2221 __ subptr(rsp, saved_rbp_offset * wordSize); |
| 2222 __ movptr(Address(rsp, saved_r13_offset * wordSize), r13); | |
| 2223 __ movptr(Address(rsp, saved_r14_offset * wordSize), r14); | |
| 0 | 2224 setup_arg_regs(4); // from => rdi, to => rsi, length => rdx |
| 2225 // ckoff => rcx, ckval => r8 | |
| 2226 // r9 and r10 may be used to save non-volatile registers | |
| 2227 #ifdef _WIN64 | |
| 2228 // last argument (#4) is on stack on Win64 | |
| 2229 const int ckval_offset = saved_rarg0_offset + 4; | |
| 304 | 2230 __ movptr(ckval, Address(rsp, ckval_offset * wordSize)); |
| 0 | 2231 #endif |
| 2232 | |
| 2233 // check that int operands are properly extended to size_t | |
| 2234 assert_clean_int(length, rax); | |
| 2235 assert_clean_int(ckoff, rax); | |
| 2236 | |
| 2237 #ifdef ASSERT | |
| 2238 BLOCK_COMMENT("assert consistent ckoff/ckval"); | |
| 2239 // The ckoff and ckval must be mutually consistent, | |
| 2240 // even though caller generates both. | |
| 2241 { Label L; | |
| 2242 int sco_offset = (klassOopDesc::header_size() * HeapWordSize + | |
| 2243 Klass::super_check_offset_offset_in_bytes()); | |
| 2244 __ cmpl(ckoff, Address(ckval, sco_offset)); | |
| 2245 __ jcc(Assembler::equal, L); | |
| 2246 __ stop("super_check_offset inconsistent"); | |
| 2247 __ bind(L); | |
| 2248 } | |
| 2249 #endif //ASSERT | |
| 2250 | |
| 2251 // Loop-invariant addresses. They are exclusive end pointers. | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2252 Address end_from_addr(from, length, TIMES_OOP, 0); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2253 Address end_to_addr(to, length, TIMES_OOP, 0); |
| 0 | 2254 // Loop-variant addresses. They assume post-incremented count < 0. |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2255 Address from_element_addr(end_from, count, TIMES_OOP, 0); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2256 Address to_element_addr(end_to, count, TIMES_OOP, 0); |
| 0 | 2257 |
| 2258 gen_write_ref_array_pre_barrier(to, count); | |
| 2259 | |
| 2260 // Copy from low to high addresses, indexed from the end of each array. | |
| 304 | 2261 __ lea(end_from, end_from_addr); |
| 2262 __ lea(end_to, end_to_addr); | |
| 2263 __ movptr(r14_length, length); // save a copy of the length | |
| 2264 assert(length == count, ""); // else fix next line: | |
| 2265 __ negptr(count); // negate and test the length | |
| 0 | 2266 __ jcc(Assembler::notZero, L_load_element); |
| 2267 | |
| 2268 // Empty array: Nothing to do. | |
| 304 | 2269 __ xorptr(rax, rax); // return 0 on (trivial) success |
| 0 | 2270 __ jmp(L_done); |
| 2271 | |
| 2272 // ======== begin loop ======== | |
| 2273 // (Loop is rotated; its entry is L_load_element.) | |
| 2274 // Loop control: | |
| 2275 // for (count = -count; count != 0; count++) | |
| 2276 // Base pointers src, dst are biased by 8*(count-1),to last element. | |
| 1365 | 2277 __ align(OptoLoopAlignment); |
| 0 | 2278 |
| 2279 __ BIND(L_store_element); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2280 __ store_heap_oop(to_element_addr, rax_oop); // store the oop |
| 304 | 2281 __ increment(count); // increment the count toward zero |
| 0 | 2282 __ jcc(Assembler::zero, L_do_card_marks); |
| 2283 | |
| 2284 // ======== loop entry is here ======== | |
| 2285 __ BIND(L_load_element); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2286 __ load_heap_oop(rax_oop, from_element_addr); // load the oop |
| 304 | 2287 __ testptr(rax_oop, rax_oop); |
| 0 | 2288 __ jcc(Assembler::zero, L_store_element); |
| 2289 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2290 __ load_klass(r11_klass, rax_oop);// query the object klass |
| 0 | 2291 generate_type_check(r11_klass, ckoff, ckval, L_store_element); |
| 2292 // ======== end loop ======== | |
| 2293 | |
| 2294 // It was a real error; we must depend on the caller to finish the job. | |
| 2295 // Register rdx = -1 * number of *remaining* oops, r14 = *total* oops. | |
| 2296 // Emit GC store barriers for the oops we have copied (r14 + rdx), | |
| 2297 // and report their number to the caller. | |
| 2298 assert_different_registers(rax, r14_length, count, to, end_to, rcx); | |
| 304 | 2299 __ lea(end_to, to_element_addr); |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
671
diff
changeset
|
2300 __ addptr(end_to, -heapOopSize); // make an inclusive end pointer |
| 362 | 2301 gen_write_ref_array_post_barrier(to, end_to, rscratch1); |
| 304 | 2302 __ movptr(rax, r14_length); // original oops |
| 2303 __ addptr(rax, count); // K = (original - remaining) oops | |
| 2304 __ notptr(rax); // report (-1^K) to caller | |
| 0 | 2305 __ jmp(L_done); |
| 2306 | |
| 2307 // Come here on success only. | |
| 2308 __ BIND(L_do_card_marks); | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
671
diff
changeset
|
2309 __ addptr(end_to, -heapOopSize); // make an inclusive end pointer |
| 362 | 2310 gen_write_ref_array_post_barrier(to, end_to, rscratch1); |
| 304 | 2311 __ xorptr(rax, rax); // return 0 on success |
| 0 | 2312 |
| 2313 // Common exit point (success or failure). | |
| 2314 __ BIND(L_done); | |
| 304 | 2315 __ movptr(r13, Address(rsp, saved_r13_offset * wordSize)); |
| 2316 __ movptr(r14, Address(rsp, saved_r14_offset * wordSize)); | |
| 0 | 2317 inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr); |
| 2318 restore_arg_regs(); | |
| 2319 __ leave(); // required for proper stackwalking of RuntimeStub frame | |
| 2320 __ ret(0); | |
| 2321 | |
| 2322 return start; | |
| 2323 } | |
| 2324 | |
| 2325 // | |
| 2326 // Generate 'unsafe' array copy stub | |
| 2327 // Though just as safe as the other stubs, it takes an unscaled | |
| 2328 // size_t argument instead of an element count. | |
| 2329 // | |
| 2330 // Input: | |
| 2331 // c_rarg0 - source array address | |
| 2332 // c_rarg1 - destination array address | |
| 2333 // c_rarg2 - byte count, treated as ssize_t, can be zero | |
| 2334 // | |
| 2335 // Examines the alignment of the operands and dispatches | |
| 2336 // to a long, int, short, or byte copy loop. | |
| 2337 // | |
| 2338 address generate_unsafe_copy(const char *name) { | |
| 2339 | |
| 2340 Label L_long_aligned, L_int_aligned, L_short_aligned; | |
| 2341 | |
| 2342 // Input registers (before setup_arg_regs) | |
| 2343 const Register from = c_rarg0; // source array address | |
| 2344 const Register to = c_rarg1; // destination array address | |
| 2345 const Register size = c_rarg2; // byte count (size_t) | |
| 2346 | |
| 2347 // Register used as a temp | |
| 2348 const Register bits = rax; // test copy of low bits | |
| 2349 | |
| 2350 __ align(CodeEntryAlignment); | |
| 2351 StubCodeMark mark(this, "StubRoutines", name); | |
| 2352 address start = __ pc(); | |
| 2353 | |
| 2354 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 2355 | |
| 2356 // bump this on entry, not on exit: | |
| 2357 inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr); | |
| 2358 | |
| 304 | 2359 __ mov(bits, from); |
| 2360 __ orptr(bits, to); | |
| 2361 __ orptr(bits, size); | |
| 0 | 2362 |
| 2363 __ testb(bits, BytesPerLong-1); | |
| 2364 __ jccb(Assembler::zero, L_long_aligned); | |
| 2365 | |
| 2366 __ testb(bits, BytesPerInt-1); | |
| 2367 __ jccb(Assembler::zero, L_int_aligned); | |
| 2368 | |
| 2369 __ testb(bits, BytesPerShort-1); | |
| 2370 __ jump_cc(Assembler::notZero, RuntimeAddress(byte_copy_entry)); | |
| 2371 | |
| 2372 __ BIND(L_short_aligned); | |
| 304 | 2373 __ shrptr(size, LogBytesPerShort); // size => short_count |
| 0 | 2374 __ jump(RuntimeAddress(short_copy_entry)); |
| 2375 | |
| 2376 __ BIND(L_int_aligned); | |
| 304 | 2377 __ shrptr(size, LogBytesPerInt); // size => int_count |
| 0 | 2378 __ jump(RuntimeAddress(int_copy_entry)); |
| 2379 | |
| 2380 __ BIND(L_long_aligned); | |
| 304 | 2381 __ shrptr(size, LogBytesPerLong); // size => qword_count |
| 0 | 2382 __ jump(RuntimeAddress(long_copy_entry)); |
| 2383 | |
| 2384 return start; | |
| 2385 } | |
| 2386 | |
| 2387 // Perform range checks on the proposed arraycopy. | |
| 2388 // Kills temp, but nothing else. | |
| 2389 // Also, clean the sign bits of src_pos and dst_pos. | |
| 2390 void arraycopy_range_checks(Register src, // source array oop (c_rarg0) | |
| 2391 Register src_pos, // source position (c_rarg1) | |
| 2392 Register dst, // destination array oo (c_rarg2) | |
| 2393 Register dst_pos, // destination position (c_rarg3) | |
| 2394 Register length, | |
| 2395 Register temp, | |
| 2396 Label& L_failed) { | |
| 2397 BLOCK_COMMENT("arraycopy_range_checks:"); | |
| 2398 | |
| 2399 // if (src_pos + length > arrayOop(src)->length()) FAIL; | |
| 2400 __ movl(temp, length); | |
| 2401 __ addl(temp, src_pos); // src_pos + length | |
| 2402 __ cmpl(temp, Address(src, arrayOopDesc::length_offset_in_bytes())); | |
| 2403 __ jcc(Assembler::above, L_failed); | |
| 2404 | |
| 2405 // if (dst_pos + length > arrayOop(dst)->length()) FAIL; | |
| 2406 __ movl(temp, length); | |
| 2407 __ addl(temp, dst_pos); // dst_pos + length | |
| 2408 __ cmpl(temp, Address(dst, arrayOopDesc::length_offset_in_bytes())); | |
| 2409 __ jcc(Assembler::above, L_failed); | |
| 2410 | |
| 2411 // Have to clean up high 32-bits of 'src_pos' and 'dst_pos'. | |
| 2412 // Move with sign extension can be used since they are positive. | |
| 2413 __ movslq(src_pos, src_pos); | |
| 2414 __ movslq(dst_pos, dst_pos); | |
| 2415 | |
| 2416 BLOCK_COMMENT("arraycopy_range_checks done"); | |
| 2417 } | |
| 2418 | |
| 2419 // | |
| 2420 // Generate generic array copy stubs | |
| 2421 // | |
| 2422 // Input: | |
| 2423 // c_rarg0 - src oop | |
| 2424 // c_rarg1 - src_pos (32-bits) | |
| 2425 // c_rarg2 - dst oop | |
| 2426 // c_rarg3 - dst_pos (32-bits) | |
| 2427 // not Win64 | |
| 2428 // c_rarg4 - element count (32-bits) | |
| 2429 // Win64 | |
| 2430 // rsp+40 - element count (32-bits) | |
| 2431 // | |
| 2432 // Output: | |
| 2433 // rax == 0 - success | |
| 2434 // rax == -1^K - failure, where K is partial transfer count | |
| 2435 // | |
| 2436 address generate_generic_copy(const char *name) { | |
| 2437 | |
| 2438 Label L_failed, L_failed_0, L_objArray; | |
| 2439 Label L_copy_bytes, L_copy_shorts, L_copy_ints, L_copy_longs; | |
| 2440 | |
| 2441 // Input registers | |
| 2442 const Register src = c_rarg0; // source array oop | |
| 2443 const Register src_pos = c_rarg1; // source position | |
| 2444 const Register dst = c_rarg2; // destination array oop | |
| 2445 const Register dst_pos = c_rarg3; // destination position | |
| 2446 // elements count is on stack on Win64 | |
| 2447 #ifdef _WIN64 | |
| 2448 #define C_RARG4 Address(rsp, 6 * wordSize) | |
| 2449 #else | |
| 2450 #define C_RARG4 c_rarg4 | |
| 2451 #endif | |
| 2452 | |
| 2453 { int modulus = CodeEntryAlignment; | |
| 2454 int target = modulus - 5; // 5 = sizeof jmp(L_failed) | |
| 2455 int advance = target - (__ offset() % modulus); | |
| 2456 if (advance < 0) advance += modulus; | |
| 2457 if (advance > 0) __ nop(advance); | |
| 2458 } | |
| 2459 StubCodeMark mark(this, "StubRoutines", name); | |
| 2460 | |
| 2461 // Short-hop target to L_failed. Makes for denser prologue code. | |
| 2462 __ BIND(L_failed_0); | |
| 2463 __ jmp(L_failed); | |
| 2464 assert(__ offset() % CodeEntryAlignment == 0, "no further alignment needed"); | |
| 2465 | |
| 2466 __ align(CodeEntryAlignment); | |
| 2467 address start = __ pc(); | |
| 2468 | |
| 2469 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 2470 | |
| 2471 // bump this on entry, not on exit: | |
| 2472 inc_counter_np(SharedRuntime::_generic_array_copy_ctr); | |
| 2473 | |
| 2474 //----------------------------------------------------------------------- | |
| 2475 // Assembler stub will be used for this call to arraycopy | |
| 2476 // if the following conditions are met: | |
| 2477 // | |
| 2478 // (1) src and dst must not be null. | |
| 2479 // (2) src_pos must not be negative. | |
| 2480 // (3) dst_pos must not be negative. | |
| 2481 // (4) length must not be negative. | |
| 2482 // (5) src klass and dst klass should be the same and not NULL. | |
| 2483 // (6) src and dst should be arrays. | |
| 2484 // (7) src_pos + length must not exceed length of src. | |
| 2485 // (8) dst_pos + length must not exceed length of dst. | |
| 2486 // | |
| 2487 | |
| 2488 // if (src == NULL) return -1; | |
| 304 | 2489 __ testptr(src, src); // src oop |
| 0 | 2490 size_t j1off = __ offset(); |
| 2491 __ jccb(Assembler::zero, L_failed_0); | |
| 2492 | |
| 2493 // if (src_pos < 0) return -1; | |
| 2494 __ testl(src_pos, src_pos); // src_pos (32-bits) | |
| 2495 __ jccb(Assembler::negative, L_failed_0); | |
| 2496 | |
| 2497 // if (dst == NULL) return -1; | |
| 304 | 2498 __ testptr(dst, dst); // dst oop |
| 0 | 2499 __ jccb(Assembler::zero, L_failed_0); |
| 2500 | |
| 2501 // if (dst_pos < 0) return -1; | |
| 2502 __ testl(dst_pos, dst_pos); // dst_pos (32-bits) | |
| 2503 size_t j4off = __ offset(); | |
| 2504 __ jccb(Assembler::negative, L_failed_0); | |
| 2505 | |
| 2506 // The first four tests are very dense code, | |
| 2507 // but not quite dense enough to put four | |
| 2508 // jumps in a 16-byte instruction fetch buffer. | |
| 2509 // That's good, because some branch predicters | |
| 2510 // do not like jumps so close together. | |
| 2511 // Make sure of this. | |
| 2512 guarantee(((j1off ^ j4off) & ~15) != 0, "I$ line of 1st & 4th jumps"); | |
| 2513 | |
| 2514 // registers used as temp | |
| 2515 const Register r11_length = r11; // elements count to copy | |
| 2516 const Register r10_src_klass = r10; // array klass | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2517 const Register r9_dst_klass = r9; // dest array klass |
| 0 | 2518 |
| 2519 // if (length < 0) return -1; | |
| 2520 __ movl(r11_length, C_RARG4); // length (elements count, 32-bits value) | |
| 2521 __ testl(r11_length, r11_length); | |
| 2522 __ jccb(Assembler::negative, L_failed_0); | |
| 2523 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2524 __ load_klass(r10_src_klass, src); |
| 0 | 2525 #ifdef ASSERT |
| 2526 // assert(src->klass() != NULL); | |
| 2527 BLOCK_COMMENT("assert klasses not null"); | |
| 2528 { Label L1, L2; | |
| 304 | 2529 __ testptr(r10_src_klass, r10_src_klass); |
| 0 | 2530 __ jcc(Assembler::notZero, L2); // it is broken if klass is NULL |
| 2531 __ bind(L1); | |
| 2532 __ stop("broken null klass"); | |
| 2533 __ bind(L2); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2534 __ load_klass(r9_dst_klass, dst); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2535 __ cmpq(r9_dst_klass, 0); |
| 0 | 2536 __ jcc(Assembler::equal, L1); // this would be broken also |
| 2537 BLOCK_COMMENT("assert done"); | |
| 2538 } | |
| 2539 #endif | |
| 2540 | |
| 2541 // Load layout helper (32-bits) | |
| 2542 // | |
| 2543 // |array_tag| | header_size | element_type | |log2_element_size| | |
| 2544 // 32 30 24 16 8 2 0 | |
| 2545 // | |
| 2546 // array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0 | |
| 2547 // | |
| 2548 | |
| 2549 int lh_offset = klassOopDesc::header_size() * HeapWordSize + | |
| 2550 Klass::layout_helper_offset_in_bytes(); | |
| 2551 | |
| 2552 const Register rax_lh = rax; // layout helper | |
| 2553 | |
| 2554 __ movl(rax_lh, Address(r10_src_klass, lh_offset)); | |
| 2555 | |
| 2556 // Handle objArrays completely differently... | |
| 2557 jint objArray_lh = Klass::array_layout_helper(T_OBJECT); | |
| 2558 __ cmpl(rax_lh, objArray_lh); | |
| 2559 __ jcc(Assembler::equal, L_objArray); | |
| 2560 | |
| 2561 // if (src->klass() != dst->klass()) return -1; | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2562 __ load_klass(r9_dst_klass, dst); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2563 __ cmpq(r10_src_klass, r9_dst_klass); |
| 0 | 2564 __ jcc(Assembler::notEqual, L_failed); |
| 2565 | |
| 2566 // if (!src->is_Array()) return -1; | |
| 2567 __ cmpl(rax_lh, Klass::_lh_neutral_value); | |
| 2568 __ jcc(Assembler::greaterEqual, L_failed); | |
| 2569 | |
| 2570 // At this point, it is known to be a typeArray (array_tag 0x3). | |
| 2571 #ifdef ASSERT | |
| 2572 { Label L; | |
| 2573 __ cmpl(rax_lh, (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift)); | |
| 2574 __ jcc(Assembler::greaterEqual, L); | |
| 2575 __ stop("must be a primitive array"); | |
| 2576 __ bind(L); | |
| 2577 } | |
| 2578 #endif | |
| 2579 | |
| 2580 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length, | |
| 2581 r10, L_failed); | |
| 2582 | |
| 2583 // typeArrayKlass | |
| 2584 // | |
| 2585 // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize); | |
| 2586 // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize); | |
| 2587 // | |
| 2588 | |
| 2589 const Register r10_offset = r10; // array offset | |
| 2590 const Register rax_elsize = rax_lh; // element size | |
| 2591 | |
| 2592 __ movl(r10_offset, rax_lh); | |
| 2593 __ shrl(r10_offset, Klass::_lh_header_size_shift); | |
| 304 | 2594 __ andptr(r10_offset, Klass::_lh_header_size_mask); // array_offset |
| 2595 __ addptr(src, r10_offset); // src array offset | |
| 2596 __ addptr(dst, r10_offset); // dst array offset | |
| 0 | 2597 BLOCK_COMMENT("choose copy loop based on element size"); |
| 2598 __ andl(rax_lh, Klass::_lh_log2_element_size_mask); // rax_lh -> rax_elsize | |
| 2599 | |
| 2600 // next registers should be set before the jump to corresponding stub | |
| 2601 const Register from = c_rarg0; // source array address | |
| 2602 const Register to = c_rarg1; // destination array address | |
| 2603 const Register count = c_rarg2; // elements count | |
| 2604 | |
| 2605 // 'from', 'to', 'count' registers should be set in such order | |
| 2606 // since they are the same as 'src', 'src_pos', 'dst'. | |
| 2607 | |
| 2608 __ BIND(L_copy_bytes); | |
| 2609 __ cmpl(rax_elsize, 0); | |
| 2610 __ jccb(Assembler::notEqual, L_copy_shorts); | |
| 304 | 2611 __ lea(from, Address(src, src_pos, Address::times_1, 0));// src_addr |
| 2612 __ lea(to, Address(dst, dst_pos, Address::times_1, 0));// dst_addr | |
| 2613 __ movl2ptr(count, r11_length); // length | |
| 0 | 2614 __ jump(RuntimeAddress(byte_copy_entry)); |
| 2615 | |
| 2616 __ BIND(L_copy_shorts); | |
| 2617 __ cmpl(rax_elsize, LogBytesPerShort); | |
| 2618 __ jccb(Assembler::notEqual, L_copy_ints); | |
| 304 | 2619 __ lea(from, Address(src, src_pos, Address::times_2, 0));// src_addr |
| 2620 __ lea(to, Address(dst, dst_pos, Address::times_2, 0));// dst_addr | |
| 2621 __ movl2ptr(count, r11_length); // length | |
| 0 | 2622 __ jump(RuntimeAddress(short_copy_entry)); |
| 2623 | |
| 2624 __ BIND(L_copy_ints); | |
| 2625 __ cmpl(rax_elsize, LogBytesPerInt); | |
| 2626 __ jccb(Assembler::notEqual, L_copy_longs); | |
| 304 | 2627 __ lea(from, Address(src, src_pos, Address::times_4, 0));// src_addr |
| 2628 __ lea(to, Address(dst, dst_pos, Address::times_4, 0));// dst_addr | |
| 2629 __ movl2ptr(count, r11_length); // length | |
| 0 | 2630 __ jump(RuntimeAddress(int_copy_entry)); |
| 2631 | |
| 2632 __ BIND(L_copy_longs); | |
| 2633 #ifdef ASSERT | |
| 2634 { Label L; | |
| 2635 __ cmpl(rax_elsize, LogBytesPerLong); | |
| 2636 __ jcc(Assembler::equal, L); | |
| 2637 __ stop("must be long copy, but elsize is wrong"); | |
| 2638 __ bind(L); | |
| 2639 } | |
| 2640 #endif | |
| 304 | 2641 __ lea(from, Address(src, src_pos, Address::times_8, 0));// src_addr |
| 2642 __ lea(to, Address(dst, dst_pos, Address::times_8, 0));// dst_addr | |
| 2643 __ movl2ptr(count, r11_length); // length | |
| 0 | 2644 __ jump(RuntimeAddress(long_copy_entry)); |
| 2645 | |
| 2646 // objArrayKlass | |
| 2647 __ BIND(L_objArray); | |
| 2648 // live at this point: r10_src_klass, src[_pos], dst[_pos] | |
| 2649 | |
| 2650 Label L_plain_copy, L_checkcast_copy; | |
| 2651 // test array classes for subtyping | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2652 __ load_klass(r9_dst_klass, dst); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2653 __ cmpq(r10_src_klass, r9_dst_klass); // usual case is exact equality |
| 0 | 2654 __ jcc(Assembler::notEqual, L_checkcast_copy); |
| 2655 | |
| 2656 // Identically typed arrays can be copied without element-wise checks. | |
| 2657 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length, | |
| 2658 r10, L_failed); | |
| 2659 | |
| 304 | 2660 __ lea(from, Address(src, src_pos, TIMES_OOP, |
| 0 | 2661 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // src_addr |
| 304 | 2662 __ lea(to, Address(dst, dst_pos, TIMES_OOP, |
| 2663 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // dst_addr | |
| 2664 __ movl2ptr(count, r11_length); // length | |
| 0 | 2665 __ BIND(L_plain_copy); |
| 2666 __ jump(RuntimeAddress(oop_copy_entry)); | |
| 2667 | |
| 2668 __ BIND(L_checkcast_copy); | |
| 2669 // live at this point: r10_src_klass, !r11_length | |
| 2670 { | |
| 2671 // assert(r11_length == C_RARG4); // will reload from here | |
| 2672 Register r11_dst_klass = r11; | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2673 __ load_klass(r11_dst_klass, dst); |
| 0 | 2674 |
| 2675 // Before looking at dst.length, make sure dst is also an objArray. | |
| 2676 __ cmpl(Address(r11_dst_klass, lh_offset), objArray_lh); | |
| 2677 __ jcc(Assembler::notEqual, L_failed); | |
| 2678 | |
| 2679 // It is safe to examine both src.length and dst.length. | |
| 2680 #ifndef _WIN64 | |
| 2681 arraycopy_range_checks(src, src_pos, dst, dst_pos, C_RARG4, | |
| 2682 rax, L_failed); | |
| 2683 #else | |
| 2684 __ movl(r11_length, C_RARG4); // reload | |
| 2685 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length, | |
| 2686 rax, L_failed); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2687 __ load_klass(r11_dst_klass, dst); // reload |
| 0 | 2688 #endif |
| 2689 | |
| 2690 // Marshal the base address arguments now, freeing registers. | |
| 304 | 2691 __ lea(from, Address(src, src_pos, TIMES_OOP, |
| 0 | 2692 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); |
| 304 | 2693 __ lea(to, Address(dst, dst_pos, TIMES_OOP, |
| 0 | 2694 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); |
| 2695 __ movl(count, C_RARG4); // length (reloaded) | |
| 2696 Register sco_temp = c_rarg3; // this register is free now | |
| 2697 assert_different_registers(from, to, count, sco_temp, | |
| 2698 r11_dst_klass, r10_src_klass); | |
| 2699 assert_clean_int(count, sco_temp); | |
| 2700 | |
| 2701 // Generate the type check. | |
| 2702 int sco_offset = (klassOopDesc::header_size() * HeapWordSize + | |
| 2703 Klass::super_check_offset_offset_in_bytes()); | |
| 2704 __ movl(sco_temp, Address(r11_dst_klass, sco_offset)); | |
| 2705 assert_clean_int(sco_temp, rax); | |
| 2706 generate_type_check(r10_src_klass, sco_temp, r11_dst_klass, L_plain_copy); | |
| 2707 | |
| 2708 // Fetch destination element klass from the objArrayKlass header. | |
| 2709 int ek_offset = (klassOopDesc::header_size() * HeapWordSize + | |
| 2710 objArrayKlass::element_klass_offset_in_bytes()); | |
| 304 | 2711 __ movptr(r11_dst_klass, Address(r11_dst_klass, ek_offset)); |
| 0 | 2712 __ movl(sco_temp, Address(r11_dst_klass, sco_offset)); |
| 2713 assert_clean_int(sco_temp, rax); | |
| 2714 | |
| 2715 // the checkcast_copy loop needs two extra arguments: | |
| 2716 assert(c_rarg3 == sco_temp, "#3 already in place"); | |
| 304 | 2717 __ movptr(C_RARG4, r11_dst_klass); // dst.klass.element_klass |
| 0 | 2718 __ jump(RuntimeAddress(checkcast_copy_entry)); |
| 2719 } | |
| 2720 | |
| 2721 __ BIND(L_failed); | |
| 304 | 2722 __ xorptr(rax, rax); |
| 2723 __ notptr(rax); // return -1 | |
| 0 | 2724 __ leave(); // required for proper stackwalking of RuntimeStub frame |
| 2725 __ ret(0); | |
| 2726 | |
| 2727 return start; | |
| 2728 } | |
| 2729 | |
| 2730 #undef length_arg | |
| 2731 | |
| 2732 void generate_arraycopy_stubs() { | |
| 2733 // Call the conjoint generation methods immediately after | |
| 2734 // the disjoint ones so that short branches from the former | |
| 2735 // to the latter can be generated. | |
| 2736 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, "jbyte_disjoint_arraycopy"); | |
| 2737 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, "jbyte_arraycopy"); | |
| 2738 | |
| 2739 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, "jshort_disjoint_arraycopy"); | |
| 2740 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, "jshort_arraycopy"); | |
| 2741 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2742 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, false, "jint_disjoint_arraycopy"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2743 StubRoutines::_jint_arraycopy = generate_conjoint_int_oop_copy(false, false, "jint_arraycopy"); |
| 0 | 2744 |
| 2745 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, false, "jlong_disjoint_arraycopy"); | |
| 2746 StubRoutines::_jlong_arraycopy = generate_conjoint_long_oop_copy(false, false, "jlong_arraycopy"); | |
| 2747 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2748 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2749 if (UseCompressedOops) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2750 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, true, "oop_disjoint_arraycopy"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2751 StubRoutines::_oop_arraycopy = generate_conjoint_int_oop_copy(false, true, "oop_arraycopy"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2752 } else { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2753 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, true, "oop_disjoint_arraycopy"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2754 StubRoutines::_oop_arraycopy = generate_conjoint_long_oop_copy(false, true, "oop_arraycopy"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
16
diff
changeset
|
2755 } |
| 0 | 2756 |
| 2757 StubRoutines::_checkcast_arraycopy = generate_checkcast_copy("checkcast_arraycopy"); | |
| 2758 StubRoutines::_unsafe_arraycopy = generate_unsafe_copy("unsafe_arraycopy"); | |
| 2759 StubRoutines::_generic_arraycopy = generate_generic_copy("generic_arraycopy"); | |
| 2760 | |
| 1763 | 2761 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill"); |
| 2762 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill"); | |
| 2763 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill"); | |
| 2764 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill"); | |
| 2765 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill"); | |
| 2766 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill"); | |
| 2767 | |
| 0 | 2768 // We don't generate specialized code for HeapWord-aligned source |
| 2769 // arrays, so just use the code we've already generated | |
| 2770 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = StubRoutines::_jbyte_disjoint_arraycopy; | |
| 2771 StubRoutines::_arrayof_jbyte_arraycopy = StubRoutines::_jbyte_arraycopy; | |
| 2772 | |
| 2773 StubRoutines::_arrayof_jshort_disjoint_arraycopy = StubRoutines::_jshort_disjoint_arraycopy; | |
| 2774 StubRoutines::_arrayof_jshort_arraycopy = StubRoutines::_jshort_arraycopy; | |
| 2775 | |
| 2776 StubRoutines::_arrayof_jint_disjoint_arraycopy = StubRoutines::_jint_disjoint_arraycopy; | |
| 2777 StubRoutines::_arrayof_jint_arraycopy = StubRoutines::_jint_arraycopy; | |
| 2778 | |
| 2779 StubRoutines::_arrayof_jlong_disjoint_arraycopy = StubRoutines::_jlong_disjoint_arraycopy; | |
| 2780 StubRoutines::_arrayof_jlong_arraycopy = StubRoutines::_jlong_arraycopy; | |
| 2781 | |
| 2782 StubRoutines::_arrayof_oop_disjoint_arraycopy = StubRoutines::_oop_disjoint_arraycopy; | |
| 2783 StubRoutines::_arrayof_oop_arraycopy = StubRoutines::_oop_arraycopy; | |
| 2784 } | |
| 2785 | |
|
1174
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2786 void generate_math_stubs() { |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2787 { |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2788 StubCodeMark mark(this, "StubRoutines", "log"); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2789 StubRoutines::_intrinsic_log = (double (*)(double)) __ pc(); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2790 |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2791 __ subq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2792 __ movdbl(Address(rsp, 0), xmm0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2793 __ fld_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2794 __ flog(); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2795 __ fstp_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2796 __ movdbl(xmm0, Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2797 __ addq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2798 __ ret(0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2799 } |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2800 { |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2801 StubCodeMark mark(this, "StubRoutines", "log10"); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2802 StubRoutines::_intrinsic_log10 = (double (*)(double)) __ pc(); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2803 |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2804 __ subq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2805 __ movdbl(Address(rsp, 0), xmm0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2806 __ fld_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2807 __ flog10(); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2808 __ fstp_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2809 __ movdbl(xmm0, Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2810 __ addq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2811 __ ret(0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2812 } |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2813 { |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2814 StubCodeMark mark(this, "StubRoutines", "sin"); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2815 StubRoutines::_intrinsic_sin = (double (*)(double)) __ pc(); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2816 |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2817 __ subq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2818 __ movdbl(Address(rsp, 0), xmm0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2819 __ fld_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2820 __ trigfunc('s'); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2821 __ fstp_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2822 __ movdbl(xmm0, Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2823 __ addq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2824 __ ret(0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2825 } |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2826 { |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2827 StubCodeMark mark(this, "StubRoutines", "cos"); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2828 StubRoutines::_intrinsic_cos = (double (*)(double)) __ pc(); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2829 |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2830 __ subq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2831 __ movdbl(Address(rsp, 0), xmm0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2832 __ fld_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2833 __ trigfunc('c'); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2834 __ fstp_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2835 __ movdbl(xmm0, Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2836 __ addq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2837 __ ret(0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2838 } |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2839 { |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2840 StubCodeMark mark(this, "StubRoutines", "tan"); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2841 StubRoutines::_intrinsic_tan = (double (*)(double)) __ pc(); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2842 |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2843 __ subq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2844 __ movdbl(Address(rsp, 0), xmm0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2845 __ fld_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2846 __ trigfunc('t'); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2847 __ fstp_d(Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2848 __ movdbl(xmm0, Address(rsp, 0)); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2849 __ addq(rsp, 8); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2850 __ ret(0); |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2851 } |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2852 |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2853 // The intrinsic version of these seem to return the same value as |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2854 // the strict version. |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2855 StubRoutines::_intrinsic_exp = SharedRuntime::dexp; |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2856 StubRoutines::_intrinsic_pow = SharedRuntime::dpow; |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2857 } |
|
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
2858 |
| 0 | 2859 #undef __ |
| 2860 #define __ masm-> | |
| 2861 | |
| 2862 // Continuation point for throwing of implicit exceptions that are | |
| 2863 // not handled in the current activation. Fabricates an exception | |
| 2864 // oop and initiates normal exception dispatching in this | |
| 2865 // frame. Since we need to preserve callee-saved values (currently | |
| 2866 // only for C2, but done for C1 as well) we need a callee-saved oop | |
| 2867 // map and therefore have to make these stubs into RuntimeStubs | |
| 2868 // rather than BufferBlobs. If the compiler needs all registers to | |
| 2869 // be preserved between the fault point and the exception handler | |
| 2870 // then it must assume responsibility for that in | |
| 2871 // AbstractCompiler::continuation_for_implicit_null_exception or | |
| 2872 // continuation_for_implicit_division_by_zero_exception. All other | |
| 2873 // implicit exceptions (e.g., NullPointerException or | |
| 2874 // AbstractMethodError on entry) are either at call sites or | |
| 2875 // otherwise assume that stack unwinding will be initiated, so | |
| 2876 // caller saved registers were assumed volatile in the compiler. | |
| 2877 address generate_throw_exception(const char* name, | |
| 2878 address runtime_entry, | |
| 2879 bool restore_saved_exception_pc) { | |
| 2880 // Information about frame layout at time of blocking runtime call. | |
| 2881 // Note that we only have to preserve callee-saved registers since | |
| 2882 // the compilers are responsible for supplying a continuation point | |
| 2883 // if they expect all registers to be preserved. | |
| 2884 enum layout { | |
| 2885 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt, | |
| 2886 rbp_off2, | |
| 2887 return_off, | |
| 2888 return_off2, | |
| 2889 framesize // inclusive of return address | |
| 2890 }; | |
| 2891 | |
| 2892 int insts_size = 512; | |
| 2893 int locs_size = 64; | |
| 2894 | |
| 2895 CodeBuffer code(name, insts_size, locs_size); | |
| 2896 OopMapSet* oop_maps = new OopMapSet(); | |
| 2897 MacroAssembler* masm = new MacroAssembler(&code); | |
| 2898 | |
| 2899 address start = __ pc(); | |
| 2900 | |
| 2901 // This is an inlined and slightly modified version of call_VM | |
| 2902 // which has the ability to fetch the return PC out of | |
| 2903 // thread-local storage and also sets up last_Java_sp slightly | |
| 2904 // differently than the real call_VM | |
| 2905 if (restore_saved_exception_pc) { | |
| 304 | 2906 __ movptr(rax, |
| 2907 Address(r15_thread, | |
| 2908 in_bytes(JavaThread::saved_exception_pc_offset()))); | |
| 2909 __ push(rax); | |
| 0 | 2910 } |
| 2911 | |
| 2912 __ enter(); // required for proper stackwalking of RuntimeStub frame | |
| 2913 | |
| 2914 assert(is_even(framesize/2), "sp not 16-byte aligned"); | |
| 2915 | |
| 2916 // return address and rbp are already in place | |
| 304 | 2917 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog |
| 0 | 2918 |
| 2919 int frame_complete = __ pc() - start; | |
| 2920 | |
| 2921 // Set up last_Java_sp and last_Java_fp | |
| 2922 __ set_last_Java_frame(rsp, rbp, NULL); | |
| 2923 | |
| 2924 // Call runtime | |
| 304 | 2925 __ movptr(c_rarg0, r15_thread); |
| 0 | 2926 BLOCK_COMMENT("call runtime_entry"); |
| 2927 __ call(RuntimeAddress(runtime_entry)); | |
| 2928 | |
| 2929 // Generate oop map | |
| 2930 OopMap* map = new OopMap(framesize, 0); | |
| 2931 | |
| 2932 oop_maps->add_gc_map(__ pc() - start, map); | |
| 2933 | |
| 2934 __ reset_last_Java_frame(true, false); | |
| 2935 | |
| 2936 __ leave(); // required for proper stackwalking of RuntimeStub frame | |
| 2937 | |
| 2938 // check for pending exceptions | |
| 2939 #ifdef ASSERT | |
| 2940 Label L; | |
| 304 | 2941 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), |
| 2942 (int32_t) NULL_WORD); | |
| 0 | 2943 __ jcc(Assembler::notEqual, L); |
| 2944 __ should_not_reach_here(); | |
| 2945 __ bind(L); | |
| 2946 #endif // ASSERT | |
| 2947 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); | |
| 2948 | |
| 2949 | |
| 2950 // codeBlob framesize is in words (not VMRegImpl::slot_size) | |
| 2951 RuntimeStub* stub = | |
| 2952 RuntimeStub::new_runtime_stub(name, | |
| 2953 &code, | |
| 2954 frame_complete, | |
| 2955 (framesize >> (LogBytesPerWord - LogBytesPerInt)), | |
| 2956 oop_maps, false); | |
| 2957 return stub->entry_point(); | |
| 2958 } | |
| 2959 | |
| 2960 // Initialization | |
| 2961 void generate_initial() { | |
| 2962 // Generates all stubs and initializes the entry points | |
| 2963 | |
| 2964 // This platform-specific stub is needed by generate_call_stub() | |
| 304 | 2965 StubRoutines::x86::_mxcsr_std = generate_fp_mask("mxcsr_std", 0x0000000000001F80); |
| 0 | 2966 |
| 2967 // entry points that exist in all platforms Note: This is code | |
| 2968 // that could be shared among different platforms - however the | |
| 2969 // benefit seems to be smaller than the disadvantage of having a | |
| 2970 // much more complicated generator structure. See also comment in | |
| 2971 // stubRoutines.hpp. | |
| 2972 | |
| 2973 StubRoutines::_forward_exception_entry = generate_forward_exception(); | |
| 2974 | |
| 2975 StubRoutines::_call_stub_entry = | |
| 2976 generate_call_stub(StubRoutines::_call_stub_return_address); | |
| 2977 | |
| 2978 // is referenced by megamorphic call | |
| 2979 StubRoutines::_catch_exception_entry = generate_catch_exception(); | |
| 2980 | |
| 2981 // atomic calls | |
| 2982 StubRoutines::_atomic_xchg_entry = generate_atomic_xchg(); | |
| 2983 StubRoutines::_atomic_xchg_ptr_entry = generate_atomic_xchg_ptr(); | |
| 2984 StubRoutines::_atomic_cmpxchg_entry = generate_atomic_cmpxchg(); | |
| 2985 StubRoutines::_atomic_cmpxchg_long_entry = generate_atomic_cmpxchg_long(); | |
| 2986 StubRoutines::_atomic_add_entry = generate_atomic_add(); | |
| 2987 StubRoutines::_atomic_add_ptr_entry = generate_atomic_add_ptr(); | |
| 2988 StubRoutines::_fence_entry = generate_orderaccess_fence(); | |
| 2989 | |
| 2990 StubRoutines::_handler_for_unsafe_access_entry = | |
| 2991 generate_handler_for_unsafe_access(); | |
| 2992 | |
| 2993 // platform dependent | |
| 304 | 2994 StubRoutines::x86::_get_previous_fp_entry = generate_get_previous_fp(); |
| 2995 | |
| 2996 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr(); | |
| 0 | 2997 } |
| 2998 | |
| 2999 void generate_all() { | |
| 3000 // Generates all stubs and initializes the entry points | |
| 3001 | |
| 3002 // These entry points require SharedInfo::stack0 to be set up in | |
| 3003 // non-core builds and need to be relocatable, so they each | |
| 3004 // fabricate a RuntimeStub internally. | |
| 3005 StubRoutines::_throw_AbstractMethodError_entry = | |
| 3006 generate_throw_exception("AbstractMethodError throw_exception", | |
| 3007 CAST_FROM_FN_PTR(address, | |
| 3008 SharedRuntime:: | |
| 3009 throw_AbstractMethodError), | |
| 3010 false); | |
| 3011 | |
|
16
f8236e79048a
6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents:
0
diff
changeset
|
3012 StubRoutines::_throw_IncompatibleClassChangeError_entry = |
|
f8236e79048a
6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents:
0
diff
changeset
|
3013 generate_throw_exception("IncompatibleClassChangeError throw_exception", |
|
f8236e79048a
6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents:
0
diff
changeset
|
3014 CAST_FROM_FN_PTR(address, |
|
f8236e79048a
6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents:
0
diff
changeset
|
3015 SharedRuntime:: |
|
f8236e79048a
6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents:
0
diff
changeset
|
3016 throw_IncompatibleClassChangeError), |
|
f8236e79048a
6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents:
0
diff
changeset
|
3017 false); |
|
f8236e79048a
6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents:
0
diff
changeset
|
3018 |
| 0 | 3019 StubRoutines::_throw_ArithmeticException_entry = |
| 3020 generate_throw_exception("ArithmeticException throw_exception", | |
| 3021 CAST_FROM_FN_PTR(address, | |
| 3022 SharedRuntime:: | |
| 3023 throw_ArithmeticException), | |
| 3024 true); | |
| 3025 | |
| 3026 StubRoutines::_throw_NullPointerException_entry = | |
| 3027 generate_throw_exception("NullPointerException throw_exception", | |
| 3028 CAST_FROM_FN_PTR(address, | |
| 3029 SharedRuntime:: | |
| 3030 throw_NullPointerException), | |
| 3031 true); | |
| 3032 | |
| 3033 StubRoutines::_throw_NullPointerException_at_call_entry = | |
| 3034 generate_throw_exception("NullPointerException at call throw_exception", | |
| 3035 CAST_FROM_FN_PTR(address, | |
| 3036 SharedRuntime:: | |
| 3037 throw_NullPointerException_at_call), | |
| 3038 false); | |
| 3039 | |
| 3040 StubRoutines::_throw_StackOverflowError_entry = | |
| 3041 generate_throw_exception("StackOverflowError throw_exception", | |
| 3042 CAST_FROM_FN_PTR(address, | |
| 3043 SharedRuntime:: | |
| 3044 throw_StackOverflowError), | |
| 3045 false); | |
| 3046 | |
| 3047 // entry points that are platform specific | |
| 304 | 3048 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup(); |
| 3049 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup(); | |
| 3050 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup(); | |
| 3051 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup(); | |
| 3052 | |
| 3053 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF); | |
| 3054 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000); | |
| 3055 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF); | |
| 3056 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000); | |
| 0 | 3057 |
| 3058 // support for verify_oop (must happen after universe_init) | |
| 3059 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop(); | |
| 3060 | |
| 3061 // arraycopy stubs used by compilers | |
| 3062 generate_arraycopy_stubs(); | |
| 1108 | 3063 |
|
1174
ddb7834449d0
6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents:
1108
diff
changeset
|
3064 generate_math_stubs(); |
| 0 | 3065 } |
| 3066 | |
| 3067 public: | |
| 3068 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) { | |
| 3069 if (all) { | |
| 3070 generate_all(); | |
| 3071 } else { | |
| 3072 generate_initial(); | |
| 3073 } | |
| 3074 } | |
| 3075 }; // end class declaration | |
| 3076 | |
| 3077 address StubGenerator::disjoint_byte_copy_entry = NULL; | |
| 3078 address StubGenerator::disjoint_short_copy_entry = NULL; | |
| 3079 address StubGenerator::disjoint_int_copy_entry = NULL; | |
| 3080 address StubGenerator::disjoint_long_copy_entry = NULL; | |
| 3081 address StubGenerator::disjoint_oop_copy_entry = NULL; | |
| 3082 | |
| 3083 address StubGenerator::byte_copy_entry = NULL; | |
| 3084 address StubGenerator::short_copy_entry = NULL; | |
| 3085 address StubGenerator::int_copy_entry = NULL; | |
| 3086 address StubGenerator::long_copy_entry = NULL; | |
| 3087 address StubGenerator::oop_copy_entry = NULL; | |
| 3088 | |
| 3089 address StubGenerator::checkcast_copy_entry = NULL; | |
| 3090 | |
| 3091 void StubGenerator_generate(CodeBuffer* code, bool all) { | |
| 3092 StubGenerator g(code, all); | |
| 3093 } |