1 /* 2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved. 3 * Copyright 2012, 2013 SAP AG. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #include "precompiled.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "code/icBuffer.hpp" 30 #include "code/vtableStubs.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "oops/compiledICHolder.hpp" 33 #include "prims/jvmtiRedefineClassesTrace.hpp" 34 #include "runtime/sharedRuntime.hpp" 35 #include "runtime/vframeArray.hpp" 36 #include "vmreg_ppc.inline.hpp" 37 #ifdef COMPILER1 38 #include "c1/c1_Runtime1.hpp" 39 #endif 40 #ifdef COMPILER2 41 #include "opto/runtime.hpp" 42 #endif 43 44 #define __ masm-> 45 46 #ifdef PRODUCT 47 #define BLOCK_COMMENT(str) // nothing 48 #else 49 #define BLOCK_COMMENT(str) __ block_comment(str) 50 #endif 51 52 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 53 54 55 // Used by generate_deopt_blob. Defined in .ad file. 56 extern uint size_deopt_handler(); 57 58 59 class RegisterSaver { 60 // Used for saving volatile registers. 61 public: 62 63 // Support different return pc locations. 64 enum ReturnPCLocation { 65 return_pc_is_lr, 66 return_pc_is_r4, 67 return_pc_is_thread_saved_exception_pc 68 }; 69 70 static OopMap* push_frame_abi112_and_save_live_registers(MacroAssembler* masm, 71 int* out_frame_size_in_bytes, 72 bool generate_oop_map, 73 int return_pc_adjustment, 74 ReturnPCLocation return_pc_location); 75 static void restore_live_registers_and_pop_frame(MacroAssembler* masm, 76 int frame_size_in_bytes, 77 bool restore_ctr); 78 79 static void push_frame_and_save_argument_registers(MacroAssembler* masm, 80 Register r_temp, 81 int frame_size, 82 int total_args, 83 const VMRegPair *regs, const VMRegPair *regs2 = NULL); 84 static void restore_argument_registers_and_pop_frame(MacroAssembler*masm, 85 int frame_size, 86 int total_args, 87 const VMRegPair *regs, const VMRegPair *regs2 = NULL); 88 89 // During deoptimization only the result registers need to be restored 90 // all the other values have already been extracted. 91 static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes); 92 93 // Constants and data structures: 94 95 typedef enum { 96 int_reg = 0, 97 float_reg = 1, 98 special_reg = 2 99 } RegisterType; 100 101 typedef enum { 102 reg_size = 8, 103 half_reg_size = reg_size / 2, 104 } RegisterConstants; 105 106 typedef struct { 107 RegisterType reg_type; 108 int reg_num; 109 VMReg vmreg; 110 } LiveRegType; 111 }; 112 113 114 #define RegisterSaver_LiveSpecialReg(regname) \ 115 { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() } 116 117 #define RegisterSaver_LiveIntReg(regname) \ 118 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() } 119 120 #define RegisterSaver_LiveFloatReg(regname) \ 121 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() } 122 123 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = { 124 // Live registers which get spilled to the stack. Register 125 // positions in this array correspond directly to the stack layout. 126 127 // 128 // live special registers: 129 // 130 RegisterSaver_LiveSpecialReg(SR_CTR), 131 // 132 // live float registers: 133 // 134 RegisterSaver_LiveFloatReg( F0 ), 135 RegisterSaver_LiveFloatReg( F1 ), 136 RegisterSaver_LiveFloatReg( F2 ), 137 RegisterSaver_LiveFloatReg( F3 ), 138 RegisterSaver_LiveFloatReg( F4 ), 139 RegisterSaver_LiveFloatReg( F5 ), 140 RegisterSaver_LiveFloatReg( F6 ), 141 RegisterSaver_LiveFloatReg( F7 ), 142 RegisterSaver_LiveFloatReg( F8 ), 143 RegisterSaver_LiveFloatReg( F9 ), 144 RegisterSaver_LiveFloatReg( F10 ), 145 RegisterSaver_LiveFloatReg( F11 ), 146 RegisterSaver_LiveFloatReg( F12 ), 147 RegisterSaver_LiveFloatReg( F13 ), 148 RegisterSaver_LiveFloatReg( F14 ), 149 RegisterSaver_LiveFloatReg( F15 ), 150 RegisterSaver_LiveFloatReg( F16 ), 151 RegisterSaver_LiveFloatReg( F17 ), 152 RegisterSaver_LiveFloatReg( F18 ), 153 RegisterSaver_LiveFloatReg( F19 ), 154 RegisterSaver_LiveFloatReg( F20 ), 155 RegisterSaver_LiveFloatReg( F21 ), 156 RegisterSaver_LiveFloatReg( F22 ), 157 RegisterSaver_LiveFloatReg( F23 ), 158 RegisterSaver_LiveFloatReg( F24 ), 159 RegisterSaver_LiveFloatReg( F25 ), 160 RegisterSaver_LiveFloatReg( F26 ), 161 RegisterSaver_LiveFloatReg( F27 ), 162 RegisterSaver_LiveFloatReg( F28 ), 163 RegisterSaver_LiveFloatReg( F29 ), 164 RegisterSaver_LiveFloatReg( F30 ), 165 RegisterSaver_LiveFloatReg( F31 ), 166 // 167 // live integer registers: 168 // 169 RegisterSaver_LiveIntReg( R0 ), 170 //RegisterSaver_LiveIntReg( R1 ), // stack pointer 171 RegisterSaver_LiveIntReg( R2 ), 172 RegisterSaver_LiveIntReg( R3 ), 173 RegisterSaver_LiveIntReg( R4 ), 174 RegisterSaver_LiveIntReg( R5 ), 175 RegisterSaver_LiveIntReg( R6 ), 176 RegisterSaver_LiveIntReg( R7 ), 177 RegisterSaver_LiveIntReg( R8 ), 178 RegisterSaver_LiveIntReg( R9 ), 179 RegisterSaver_LiveIntReg( R10 ), 180 RegisterSaver_LiveIntReg( R11 ), 181 RegisterSaver_LiveIntReg( R12 ), 182 //RegisterSaver_LiveIntReg( R13 ), // system thread id 183 RegisterSaver_LiveIntReg( R14 ), 184 RegisterSaver_LiveIntReg( R15 ), 185 RegisterSaver_LiveIntReg( R16 ), 186 RegisterSaver_LiveIntReg( R17 ), 187 RegisterSaver_LiveIntReg( R18 ), 188 RegisterSaver_LiveIntReg( R19 ), 189 RegisterSaver_LiveIntReg( R20 ), 190 RegisterSaver_LiveIntReg( R21 ), 191 RegisterSaver_LiveIntReg( R22 ), 192 RegisterSaver_LiveIntReg( R23 ), 193 RegisterSaver_LiveIntReg( R24 ), 194 RegisterSaver_LiveIntReg( R25 ), 195 RegisterSaver_LiveIntReg( R26 ), 196 RegisterSaver_LiveIntReg( R27 ), 197 RegisterSaver_LiveIntReg( R28 ), 198 RegisterSaver_LiveIntReg( R29 ), 199 RegisterSaver_LiveIntReg( R31 ), 200 RegisterSaver_LiveIntReg( R30 ), // r30 must be the last register 201 }; 202 203 OopMap* RegisterSaver::push_frame_abi112_and_save_live_registers(MacroAssembler* masm, 204 int* out_frame_size_in_bytes, 205 bool generate_oop_map, 206 int return_pc_adjustment, 207 ReturnPCLocation return_pc_location) { 208 // Push an abi112-frame and store all registers which may be live. 209 // If requested, create an OopMap: Record volatile registers as 210 // callee-save values in an OopMap so their save locations will be 211 // propagated to the RegisterMap of the caller frame during 212 // StackFrameStream construction (needed for deoptimization; see 213 // compiledVFrame::create_stack_value). 214 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment. 215 216 int i; 217 int offset; 218 219 // calcualte frame size 220 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 221 sizeof(RegisterSaver::LiveRegType); 222 const int register_save_size = regstosave_num * reg_size; 223 const int frame_size_in_bytes = round_to(register_save_size, frame::alignment_in_bytes) 224 + frame::abi_112_size; 225 *out_frame_size_in_bytes = frame_size_in_bytes; 226 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint); 227 const int register_save_offset = frame_size_in_bytes - register_save_size; 228 229 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words. 230 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : NULL; 231 232 BLOCK_COMMENT("push_frame_abi112_and_save_live_registers {"); 233 234 // Save r30 in the last slot of the not yet pushed frame so that we 235 // can use it as scratch reg. 236 __ std(R30, -reg_size, R1_SP); 237 assert(-reg_size == register_save_offset - frame_size_in_bytes + ((regstosave_num-1)*reg_size), 238 "consistency check"); 239 240 // save the flags 241 // Do the save_LR_CR by hand and adjust the return pc if requested. 242 __ mfcr(R30); 243 __ std(R30, _abi(cr), R1_SP); 244 switch (return_pc_location) { 245 case return_pc_is_lr: __ mflr(R30); break; 246 case return_pc_is_r4: __ mr(R30, R4); break; 247 case return_pc_is_thread_saved_exception_pc: 248 __ ld(R30, thread_(saved_exception_pc)); break; 249 default: ShouldNotReachHere(); 250 } 251 if (return_pc_adjustment != 0) 252 __ addi(R30, R30, return_pc_adjustment); 253 __ std(R30, _abi(lr), R1_SP); 254 255 // push a new frame 256 __ push_frame(frame_size_in_bytes, R30); 257 258 // save all registers (ints and floats) 259 offset = register_save_offset; 260 for (int i = 0; i < regstosave_num; i++) { 261 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 262 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 263 264 switch (reg_type) { 265 case RegisterSaver::int_reg: { 266 if (reg_num != 30) { // We spilled R30 right at the beginning. 267 __ std(as_Register(reg_num), offset, R1_SP); 268 } 269 break; 270 } 271 case RegisterSaver::float_reg: { 272 __ stfd(as_FloatRegister(reg_num), offset, R1_SP); 273 break; 274 } 275 case RegisterSaver::special_reg: { 276 if (reg_num == SR_CTR_SpecialRegisterEnumValue) { 277 __ mfctr(R30); 278 __ std(R30, offset, R1_SP); 279 } else { 280 Unimplemented(); 281 } 282 break; 283 } 284 default: 285 ShouldNotReachHere(); 286 } 287 288 if (generate_oop_map) { 289 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), 290 RegisterSaver_LiveRegs[i].vmreg); 291 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size)>>2), 292 RegisterSaver_LiveRegs[i].vmreg->next()); 293 } 294 offset += reg_size; 295 } 296 297 BLOCK_COMMENT("} push_frame_abi112_and_save_live_registers"); 298 299 // And we're done. 300 return map; 301 } 302 303 304 // Pop the current frame and restore all the registers that we 305 // saved. 306 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm, 307 int frame_size_in_bytes, 308 bool restore_ctr) { 309 int i; 310 int offset; 311 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 312 sizeof(RegisterSaver::LiveRegType); 313 const int register_save_size = regstosave_num * reg_size; 314 const int register_save_offset = frame_size_in_bytes - register_save_size; 315 316 BLOCK_COMMENT("restore_live_registers_and_pop_frame {"); 317 318 // restore all registers (ints and floats) 319 offset = register_save_offset; 320 for (int i = 0; i < regstosave_num; i++) { 321 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 322 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 323 324 switch (reg_type) { 325 case RegisterSaver::int_reg: { 326 if (reg_num != 30) // R30 restored at the end, it's the tmp reg! 327 __ ld(as_Register(reg_num), offset, R1_SP); 328 break; 329 } 330 case RegisterSaver::float_reg: { 331 __ lfd(as_FloatRegister(reg_num), offset, R1_SP); 332 break; 333 } 334 case RegisterSaver::special_reg: { 335 if (reg_num == SR_CTR_SpecialRegisterEnumValue) { 336 if (restore_ctr) { // Nothing to do here if ctr already contains the next address. 337 __ ld(R30, offset, R1_SP); 338 __ mtctr(R30); 339 } 340 } else { 341 Unimplemented(); 342 } 343 break; 344 } 345 default: 346 ShouldNotReachHere(); 347 } 348 offset += reg_size; 349 } 350 351 // pop the frame 352 __ pop_frame(); 353 354 // restore the flags 355 __ restore_LR_CR(R30); 356 357 // restore scratch register's value 358 __ ld(R30, -reg_size, R1_SP); 359 360 BLOCK_COMMENT("} restore_live_registers_and_pop_frame"); 361 } 362 363 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp, 364 int frame_size,int total_args, const VMRegPair *regs, 365 const VMRegPair *regs2) { 366 __ push_frame(frame_size, r_temp); 367 int st_off = frame_size - wordSize; 368 for (int i = 0; i < total_args; i++) { 369 VMReg r_1 = regs[i].first(); 370 VMReg r_2 = regs[i].second(); 371 if (!r_1->is_valid()) { 372 assert(!r_2->is_valid(), ""); 373 continue; 374 } 375 if (r_1->is_Register()) { 376 Register r = r_1->as_Register(); 377 __ std(r, st_off, R1_SP); 378 st_off -= wordSize; 379 } else if (r_1->is_FloatRegister()) { 380 FloatRegister f = r_1->as_FloatRegister(); 381 __ stfd(f, st_off, R1_SP); 382 st_off -= wordSize; 383 } 384 } 385 if (regs2 != NULL) { 386 for (int i = 0; i < total_args; i++) { 387 VMReg r_1 = regs2[i].first(); 388 VMReg r_2 = regs2[i].second(); 389 if (!r_1->is_valid()) { 390 assert(!r_2->is_valid(), ""); 391 continue; 392 } 393 if (r_1->is_Register()) { 394 Register r = r_1->as_Register(); 395 __ std(r, st_off, R1_SP); 396 st_off -= wordSize; 397 } else if (r_1->is_FloatRegister()) { 398 FloatRegister f = r_1->as_FloatRegister(); 399 __ stfd(f, st_off, R1_SP); 400 st_off -= wordSize; 401 } 402 } 403 } 404 } 405 406 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size, 407 int total_args, const VMRegPair *regs, 408 const VMRegPair *regs2) { 409 int st_off = frame_size - wordSize; 410 for (int i = 0; i < total_args; i++) { 411 VMReg r_1 = regs[i].first(); 412 VMReg r_2 = regs[i].second(); 413 if (r_1->is_Register()) { 414 Register r = r_1->as_Register(); 415 __ ld(r, st_off, R1_SP); 416 st_off -= wordSize; 417 } else if (r_1->is_FloatRegister()) { 418 FloatRegister f = r_1->as_FloatRegister(); 419 __ lfd(f, st_off, R1_SP); 420 st_off -= wordSize; 421 } 422 } 423 if (regs2 != NULL) 424 for (int i = 0; i < total_args; i++) { 425 VMReg r_1 = regs2[i].first(); 426 VMReg r_2 = regs2[i].second(); 427 if (r_1->is_Register()) { 428 Register r = r_1->as_Register(); 429 __ ld(r, st_off, R1_SP); 430 st_off -= wordSize; 431 } else if (r_1->is_FloatRegister()) { 432 FloatRegister f = r_1->as_FloatRegister(); 433 __ lfd(f, st_off, R1_SP); 434 st_off -= wordSize; 435 } 436 } 437 __ pop_frame(); 438 } 439 440 // Restore the registers that might be holding a result. 441 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) { 442 int i; 443 int offset; 444 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 445 sizeof(RegisterSaver::LiveRegType); 446 const int register_save_size = regstosave_num * reg_size; 447 const int register_save_offset = frame_size_in_bytes - register_save_size; 448 449 // restore all result registers (ints and floats) 450 offset = register_save_offset; 451 for (int i = 0; i < regstosave_num; i++) { 452 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 453 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 454 switch (reg_type) { 455 case RegisterSaver::int_reg: { 456 if (as_Register(reg_num)==R3_RET) // int result_reg 457 __ ld(as_Register(reg_num), offset, R1_SP); 458 break; 459 } 460 case RegisterSaver::float_reg: { 461 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg 462 __ lfd(as_FloatRegister(reg_num), offset, R1_SP); 463 break; 464 } 465 case RegisterSaver::special_reg: { 466 // Special registers don't hold a result. 467 break; 468 } 469 default: 470 ShouldNotReachHere(); 471 } 472 offset += reg_size; 473 } 474 } 475 476 // Is vector's size (in bytes) bigger than a size saved by default? 477 bool SharedRuntime::is_wide_vector(int size) { 478 ResourceMark rm; 479 // Note, MaxVectorSize == 8 on PPC64. 480 assert(size <= 8, err_msg_res("%d bytes vectors are not supported", size)); 481 return size > 8; 482 } 483 #ifdef COMPILER2 484 static int reg2slot(VMReg r) { 485 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 486 } 487 488 static int reg2offset(VMReg r) { 489 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 490 } 491 #endif 492 493 // --------------------------------------------------------------------------- 494 // Read the array of BasicTypes from a signature, and compute where the 495 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 496 // quantities. Values less than VMRegImpl::stack0 are registers, those above 497 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 498 // as framesizes are fixed. 499 // VMRegImpl::stack0 refers to the first slot 0(sp). 500 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register 501 // up to RegisterImpl::number_of_registers) are the 64-bit 502 // integer registers. 503 504 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 505 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 506 // units regardless of build. Of course for i486 there is no 64 bit build 507 508 // The Java calling convention is a "shifted" version of the C ABI. 509 // By skipping the first C ABI register we can call non-static jni methods 510 // with small numbers of arguments without having to shuffle the arguments 511 // at all. Since we control the java ABI we ought to at least get some 512 // advantage out of it. 513 514 const VMReg java_iarg_reg[8] = { 515 R3->as_VMReg(), 516 R4->as_VMReg(), 517 R5->as_VMReg(), 518 R6->as_VMReg(), 519 R7->as_VMReg(), 520 R8->as_VMReg(), 521 R9->as_VMReg(), 522 R10->as_VMReg() 523 }; 524 525 const VMReg java_farg_reg[13] = { 526 F1->as_VMReg(), 527 F2->as_VMReg(), 528 F3->as_VMReg(), 529 F4->as_VMReg(), 530 F5->as_VMReg(), 531 F6->as_VMReg(), 532 F7->as_VMReg(), 533 F8->as_VMReg(), 534 F9->as_VMReg(), 535 F10->as_VMReg(), 536 F11->as_VMReg(), 537 F12->as_VMReg(), 538 F13->as_VMReg() 539 }; 540 541 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]); 542 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]); 543 544 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 545 VMRegPair *regs, 546 int total_args_passed, 547 int is_outgoing) { 548 // C2c calling conventions for compiled-compiled calls. 549 // Put 8 ints/longs into registers _AND_ 13 float/doubles into 550 // registers _AND_ put the rest on the stack. 551 552 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats 553 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles 554 555 int i; 556 VMReg reg; 557 int stk = 0; 558 int ireg = 0; 559 int freg = 0; 560 561 // We put the first 8 arguments into registers and the rest on the 562 // stack, float arguments are already in their argument registers 563 // due to c2c calling conventions (see calling_convention). 564 for (int i = 0; i < total_args_passed; ++i) { 565 switch(sig_bt[i]) { 566 case T_BOOLEAN: 567 case T_CHAR: 568 case T_BYTE: 569 case T_SHORT: 570 case T_INT: 571 if (ireg < num_java_iarg_registers) { 572 // Put int/ptr in register 573 reg = java_iarg_reg[ireg]; 574 ++ireg; 575 } else { 576 // Put int/ptr on stack. 577 reg = VMRegImpl::stack2reg(stk); 578 stk += inc_stk_for_intfloat; 579 } 580 regs[i].set1(reg); 581 break; 582 case T_LONG: 583 assert(sig_bt[i+1] == T_VOID, "expecting half"); 584 if (ireg < num_java_iarg_registers) { 585 // Put long in register. 586 reg = java_iarg_reg[ireg]; 587 ++ireg; 588 } else { 589 // Put long on stack. They must be aligned to 2 slots. 590 if (stk & 0x1) ++stk; 591 reg = VMRegImpl::stack2reg(stk); 592 stk += inc_stk_for_longdouble; 593 } 594 regs[i].set2(reg); 595 break; 596 case T_OBJECT: 597 case T_ARRAY: 598 case T_ADDRESS: 599 if (ireg < num_java_iarg_registers) { 600 // Put ptr in register. 601 reg = java_iarg_reg[ireg]; 602 ++ireg; 603 } else { 604 // Put ptr on stack. Objects must be aligned to 2 slots too, 605 // because "64-bit pointers record oop-ishness on 2 aligned 606 // adjacent registers." (see OopFlow::build_oop_map). 607 if (stk & 0x1) ++stk; 608 reg = VMRegImpl::stack2reg(stk); 609 stk += inc_stk_for_longdouble; 610 } 611 regs[i].set2(reg); 612 break; 613 case T_FLOAT: 614 if (freg < num_java_farg_registers) { 615 // Put float in register. 616 reg = java_farg_reg[freg]; 617 ++freg; 618 } else { 619 // Put float on stack. 620 reg = VMRegImpl::stack2reg(stk); 621 stk += inc_stk_for_intfloat; 622 } 623 regs[i].set1(reg); 624 break; 625 case T_DOUBLE: 626 assert(sig_bt[i+1] == T_VOID, "expecting half"); 627 if (freg < num_java_farg_registers) { 628 // Put double in register. 629 reg = java_farg_reg[freg]; 630 ++freg; 631 } else { 632 // Put double on stack. They must be aligned to 2 slots. 633 if (stk & 0x1) ++stk; 634 reg = VMRegImpl::stack2reg(stk); 635 stk += inc_stk_for_longdouble; 636 } 637 regs[i].set2(reg); 638 break; 639 case T_VOID: 640 // Do not count halves. 641 regs[i].set_bad(); 642 break; 643 default: 644 ShouldNotReachHere(); 645 } 646 } 647 return round_to(stk, 2); 648 } 649 650 #ifdef COMPILER2 651 // Calling convention for calling C code. 652 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 653 VMRegPair *regs, 654 VMRegPair *regs2, 655 int total_args_passed) { 656 // Calling conventions for C runtime calls and calls to JNI native methods. 657 // 658 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8 659 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist 660 // the first 13 flt/dbl's in the first 13 fp regs but additionally 661 // copy flt/dbl to the stack if they are beyond the 8th argument. 662 663 const VMReg iarg_reg[8] = { 664 R3->as_VMReg(), 665 R4->as_VMReg(), 666 R5->as_VMReg(), 667 R6->as_VMReg(), 668 R7->as_VMReg(), 669 R8->as_VMReg(), 670 R9->as_VMReg(), 671 R10->as_VMReg() 672 }; 673 674 const VMReg farg_reg[13] = { 675 F1->as_VMReg(), 676 F2->as_VMReg(), 677 F3->as_VMReg(), 678 F4->as_VMReg(), 679 F5->as_VMReg(), 680 F6->as_VMReg(), 681 F7->as_VMReg(), 682 F8->as_VMReg(), 683 F9->as_VMReg(), 684 F10->as_VMReg(), 685 F11->as_VMReg(), 686 F12->as_VMReg(), 687 F13->as_VMReg() 688 }; 689 690 const int num_iarg_registers = sizeof(iarg_reg) / sizeof(iarg_reg[0]); 691 const int num_farg_registers = sizeof(farg_reg) / sizeof(farg_reg[0]); 692 693 // The first 8 arguments are not passed on the stack. 694 const int num_args_in_regs = 8; 695 #define put_arg_in_reg(arg) ((arg) < num_args_in_regs) 696 697 // Check calling conventions consistency. 698 assert(num_iarg_registers == num_args_in_regs 699 && num_iarg_registers == 8 700 && num_farg_registers == 13, 701 "consistency"); 702 703 // `Stk' counts stack slots. Due to alignment, 32 bit values occupy 704 // 2 such slots, like 64 bit values do. 705 const int inc_stk_for_intfloat = 2; // 2 slots for ints and floats 706 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles 707 708 int ill_i = 0; 709 int ill_t = 0; 710 int i; 711 VMReg reg; 712 // Leave room for C-compatible ABI_112. 713 int stk = (frame::abi_112_size - frame::jit_out_preserve_size) / VMRegImpl::stack_slot_size; 714 int arg = 0; 715 int freg = 0; 716 717 // Avoid passing C arguments in the wrong stack slots. 718 assert((SharedRuntime::out_preserve_stack_slots() + stk) * VMRegImpl::stack_slot_size == 112, 719 "passing C arguments in wrong stack slots"); 720 721 // We fill-out regs AND regs2 if an argument must be passed in a 722 // register AND in a stack slot. If regs2 is NULL in such a 723 // situation, we bail-out with a fatal error. 724 for (int i = 0; i < total_args_passed; ++i, ++arg) { 725 // Initialize regs2 to BAD. 726 if (regs2 != NULL) regs2[i].set_bad(); 727 728 switch(sig_bt[i]) { 729 case T_BOOLEAN: 730 case T_CHAR: 731 case T_BYTE: 732 case T_SHORT: 733 case T_INT: 734 // We must cast ints to longs and use full 64 bit stack slots 735 // here. We do the cast in GraphKit::gen_stub() and just guard 736 // here against loosing that change. 737 Unimplemented(); // TODO: PPC port 738 /* 739 assert(SharedRuntime::c_calling_convention_requires_ints_as_longs(), 740 "argument of type int should be promoted to type long"); 741 */ 742 guarantee(i > 0 && sig_bt[i-1] == T_LONG, 743 "argument of type (bt) should have been promoted to type (T_LONG,bt) for bt in " 744 "{T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}"); 745 // Do not count halves. 746 regs[i].set_bad(); 747 --arg; 748 break; 749 case T_LONG: 750 guarantee(sig_bt[i+1] == T_VOID || 751 sig_bt[i+1] == T_BOOLEAN || sig_bt[i+1] == T_CHAR || 752 sig_bt[i+1] == T_BYTE || sig_bt[i+1] == T_SHORT || 753 sig_bt[i+1] == T_INT, 754 "expecting type (T_LONG,half) or type (T_LONG,bt) with bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}"); 755 case T_OBJECT: 756 case T_ARRAY: 757 case T_ADDRESS: 758 case T_METADATA: 759 // Oops are already boxed if required (JNI). 760 if (put_arg_in_reg(arg)) { 761 reg = iarg_reg[arg]; 762 } else { 763 reg = VMRegImpl::stack2reg(stk); 764 stk += inc_stk_for_longdouble; 765 } 766 regs[i].set2(reg); 767 break; 768 case T_FLOAT: 769 if (put_arg_in_reg(arg)) { 770 reg = farg_reg[freg]; 771 } else { 772 // Put float on stack 773 # if defined(LINUX) 774 reg = VMRegImpl::stack2reg(stk+1); 775 # elif defined(AIX) 776 reg = VMRegImpl::stack2reg(stk); 777 # else 778 # error "unknown OS" 779 # endif 780 stk += inc_stk_for_intfloat; 781 } 782 783 if (freg < num_farg_registers) { 784 // There are still some float argument registers left. Put the 785 // float in a register if not already done. 786 if (reg != farg_reg[freg]) { 787 guarantee(regs2 != NULL, "must pass float in register and stack slot"); 788 VMReg reg2 = farg_reg[freg]; 789 regs2[i].set1(reg2); 790 } 791 ++freg; 792 } 793 794 regs[i].set1(reg); 795 break; 796 case T_DOUBLE: 797 assert(sig_bt[i+1] == T_VOID, "expecting half"); 798 if (put_arg_in_reg(arg)) { 799 reg = farg_reg[freg]; 800 } else { 801 // Put double on stack. 802 reg = VMRegImpl::stack2reg(stk); 803 stk += inc_stk_for_longdouble; 804 } 805 806 if (freg < num_farg_registers) { 807 // There are still some float argument registers left. Put the 808 // float in a register if not already done. 809 if (reg != farg_reg[freg]) { 810 guarantee(regs2 != NULL, "must pass float in register and stack slot"); 811 VMReg reg2 = farg_reg[freg]; 812 regs2[i].set2(reg2); 813 } 814 ++freg; 815 } 816 817 regs[i].set2(reg); 818 break; 819 case T_VOID: 820 // Do not count halves. 821 regs[i].set_bad(); 822 --arg; 823 break; 824 default: 825 ShouldNotReachHere(); 826 } 827 } 828 829 return round_to(stk, 2); 830 } 831 #endif // COMPILER2 832 833 static address gen_c2i_adapter(MacroAssembler *masm, 834 int total_args_passed, 835 int comp_args_on_stack, 836 const BasicType *sig_bt, 837 const VMRegPair *regs, 838 Label& call_interpreter, 839 const Register& ientry) { 840 841 address c2i_entrypoint; 842 843 const Register sender_SP = R21_sender_SP; // == R21_tmp1 844 const Register code = R22_tmp2; 845 //const Register ientry = R23_tmp3; 846 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 }; 847 const int num_value_regs = sizeof(value_regs) / sizeof(Register); 848 int value_regs_index = 0; 849 850 const Register return_pc = R27_tmp7; 851 const Register tmp = R28_tmp8; 852 853 assert_different_registers(sender_SP, code, ientry, return_pc, tmp); 854 855 // Adapter needs TOP_IJAVA_FRAME_ABI. 856 const int adapter_size = frame::top_ijava_frame_abi_size + 857 round_to(total_args_passed * wordSize, frame::alignment_in_bytes); 858 859 860 // regular (verified) c2i entry point 861 c2i_entrypoint = __ pc(); 862 863 // Does compiled code exists? If yes, patch the caller's callsite. 864 __ ld(code, method_(code)); 865 __ cmpdi(CCR0, code, 0); 866 __ ld(ientry, method_(interpreter_entry)); // preloaded 867 __ beq(CCR0, call_interpreter); 868 869 870 // Patch caller's callsite, method_(code) was not NULL which means that 871 // compiled code exists. 872 __ mflr(return_pc); 873 __ std(return_pc, _abi(lr), R1_SP); 874 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs); 875 876 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc); 877 878 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs); 879 __ ld(return_pc, _abi(lr), R1_SP); 880 __ ld(ientry, method_(interpreter_entry)); // preloaded 881 __ mtlr(return_pc); 882 883 884 // call the interpreter 885 __ BIND(call_interpreter); 886 __ mtctr(ientry); 887 888 // Get a copy of the current SP for loading caller's arguments. 889 __ mr(sender_SP, R1_SP); 890 891 // Add space for the adapter. 892 __ resize_frame(-adapter_size, R12_scratch2); 893 894 int st_off = adapter_size - wordSize; 895 896 // Write the args into the outgoing interpreter space. 897 for (int i = 0; i < total_args_passed; i++) { 898 VMReg r_1 = regs[i].first(); 899 VMReg r_2 = regs[i].second(); 900 if (!r_1->is_valid()) { 901 assert(!r_2->is_valid(), ""); 902 continue; 903 } 904 if (r_1->is_stack()) { 905 Register tmp_reg = value_regs[value_regs_index]; 906 value_regs_index = (value_regs_index + 1) % num_value_regs; 907 // The calling convention produces OptoRegs that ignore the out 908 // preserve area (JIT's ABI). We must account for it here. 909 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 910 if (!r_2->is_valid()) { 911 __ lwz(tmp_reg, ld_off, sender_SP); 912 } else { 913 __ ld(tmp_reg, ld_off, sender_SP); 914 } 915 // Pretend stack targets were loaded into tmp_reg. 916 r_1 = tmp_reg->as_VMReg(); 917 } 918 919 if (r_1->is_Register()) { 920 Register r = r_1->as_Register(); 921 if (!r_2->is_valid()) { 922 __ stw(r, st_off, R1_SP); 923 st_off-=wordSize; 924 } else { 925 // Longs are given 2 64-bit slots in the interpreter, but the 926 // data is passed in only 1 slot. 927 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 928 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); ) 929 st_off-=wordSize; 930 } 931 __ std(r, st_off, R1_SP); 932 st_off-=wordSize; 933 } 934 } else { 935 assert(r_1->is_FloatRegister(), ""); 936 FloatRegister f = r_1->as_FloatRegister(); 937 if (!r_2->is_valid()) { 938 __ stfs(f, st_off, R1_SP); 939 st_off-=wordSize; 940 } else { 941 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the 942 // data is passed in only 1 slot. 943 // One of these should get known junk... 944 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); ) 945 st_off-=wordSize; 946 __ stfd(f, st_off, R1_SP); 947 st_off-=wordSize; 948 } 949 } 950 } 951 952 // Jump to the interpreter just as if interpreter was doing it. 953 954 // load TOS 955 __ addi(R17_tos, R1_SP, st_off); 956 957 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1. 958 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register"); 959 __ bctr(); 960 961 return c2i_entrypoint; 962 } 963 964 static void gen_i2c_adapter(MacroAssembler *masm, 965 int total_args_passed, 966 int comp_args_on_stack, 967 const BasicType *sig_bt, 968 const VMRegPair *regs) { 969 970 // Load method's entry-point from methodOop. 971 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method); 972 __ mtctr(R12_scratch2); 973 974 // We will only enter here from an interpreted frame and never from after 975 // passing thru a c2i. Azul allowed this but we do not. If we lose the 976 // race and use a c2i we will remain interpreted for the race loser(s). 977 // This removes all sorts of headaches on the x86 side and also eliminates 978 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions. 979 980 // Note: r13 contains the senderSP on entry. We must preserve it since 981 // we may do a i2c -> c2i transition if we lose a race where compiled 982 // code goes non-entrant while we get args ready. 983 // In addition we use r13 to locate all the interpreter args as 984 // we must align the stack to 16 bytes on an i2c entry else we 985 // lose alignment we expect in all compiled code and register 986 // save code can segv when fxsave instructions find improperly 987 // aligned stack pointer. 988 989 const Register ld_ptr = R17_tos; 990 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 }; 991 const int num_value_regs = sizeof(value_regs) / sizeof(Register); 992 int value_regs_index = 0; 993 994 int ld_offset = total_args_passed*wordSize; 995 996 // Cut-out for having no stack args. Since up to 2 int/oop args are passed 997 // in registers, we will occasionally have no stack args. 998 int comp_words_on_stack = 0; 999 if (comp_args_on_stack) { 1000 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in 1001 // registers are below. By subtracting stack0, we either get a negative 1002 // number (all values in registers) or the maximum stack slot accessed. 1003 1004 // Convert 4-byte c2 stack slots to words. 1005 comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 1006 // Round up to miminum stack alignment, in wordSize. 1007 comp_words_on_stack = round_to(comp_words_on_stack, 2); 1008 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1); 1009 } 1010 1011 // Now generate the shuffle code. Pick up all register args and move the 1012 // rest through register value=Z_R12. 1013 BLOCK_COMMENT("Shuffle arguments"); 1014 for (int i = 0; i < total_args_passed; i++) { 1015 if (sig_bt[i] == T_VOID) { 1016 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 1017 continue; 1018 } 1019 1020 // Pick up 0, 1 or 2 words from ld_ptr. 1021 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 1022 "scrambled load targets?"); 1023 VMReg r_1 = regs[i].first(); 1024 VMReg r_2 = regs[i].second(); 1025 if (!r_1->is_valid()) { 1026 assert(!r_2->is_valid(), ""); 1027 continue; 1028 } 1029 if (r_1->is_FloatRegister()) { 1030 if (!r_2->is_valid()) { 1031 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr); 1032 ld_offset-=wordSize; 1033 } else { 1034 // Skip the unused interpreter slot. 1035 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr); 1036 ld_offset-=2*wordSize; 1037 } 1038 } else { 1039 Register r; 1040 if (r_1->is_stack()) { 1041 // Must do a memory to memory move thru "value". 1042 r = value_regs[value_regs_index]; 1043 value_regs_index = (value_regs_index + 1) % num_value_regs; 1044 } else { 1045 r = r_1->as_Register(); 1046 } 1047 if (!r_2->is_valid()) { 1048 // Not sure we need to do this but it shouldn't hurt. 1049 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ADDRESS || sig_bt[i] == T_ARRAY) { 1050 __ ld(r, ld_offset, ld_ptr); 1051 ld_offset-=wordSize; 1052 } else { 1053 __ lwz(r, ld_offset, ld_ptr); 1054 ld_offset-=wordSize; 1055 } 1056 } else { 1057 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the 1058 // data is passed in only 1 slot. 1059 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 1060 ld_offset-=wordSize; 1061 } 1062 __ ld(r, ld_offset, ld_ptr); 1063 ld_offset-=wordSize; 1064 } 1065 1066 if (r_1->is_stack()) { 1067 // Now store value where the compiler expects it 1068 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size; 1069 1070 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN || 1071 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) { 1072 __ stw(r, st_off, R1_SP); 1073 } else { 1074 __ std(r, st_off, R1_SP); 1075 } 1076 } 1077 } 1078 } 1079 1080 BLOCK_COMMENT("Store method oop"); 1081 // Store method oop into thread->callee_target. 1082 // We might end up in handle_wrong_method if the callee is 1083 // deoptimized as we race thru here. If that happens we don't want 1084 // to take a safepoint because the caller frame will look 1085 // interpreted and arguments are now "compiled" so it is much better 1086 // to make this transition invisible to the stack walking 1087 // code. Unfortunately if we try and find the callee by normal means 1088 // a safepoint is possible. So we stash the desired callee in the 1089 // thread and the vm will find there should this case occur. 1090 __ std(R19_method, thread_(callee_target)); 1091 1092 // Jump to the compiled code just as if compiled code was doing it. 1093 __ bctr(); 1094 } 1095 1096 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 1097 int total_args_passed, 1098 int comp_args_on_stack, 1099 const BasicType *sig_bt, 1100 const VMRegPair *regs, 1101 AdapterFingerPrint* fingerprint) { 1102 address i2c_entry; 1103 address c2i_unverified_entry; 1104 address c2i_entry; 1105 1106 1107 // entry: i2c 1108 1109 __ align(CodeEntryAlignment); 1110 i2c_entry = __ pc(); 1111 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 1112 1113 1114 // entry: c2i unverified 1115 1116 __ align(CodeEntryAlignment); 1117 BLOCK_COMMENT("c2i unverified entry"); 1118 c2i_unverified_entry = __ pc(); 1119 1120 // inline_cache contains a compiledICHolder 1121 const Register ic = R19_method; 1122 const Register ic_klass = R11_scratch1; 1123 const Register receiver_klass = R12_scratch2; 1124 const Register code = R21_tmp1; 1125 const Register ientry = R23_tmp3; 1126 1127 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry); 1128 assert(R11_scratch1 == R11, "need prologue scratch register"); 1129 1130 Label call_interpreter; 1131 1132 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), 1133 "klass offset should reach into any page"); 1134 // Check for NULL argument if we don't have implicit null checks. 1135 if (!ImplicitNullChecks NOT_LINUX(|| true) /*!os::zero_page_read_protected()*/) { 1136 if (TrapBasedNullChecks) { 1137 __ trap_null_check(R3_ARG1); 1138 } else { 1139 Label valid; 1140 __ cmpdi(CCR0, R3_ARG1, 0); 1141 __ bne_predict_taken(CCR0, valid); 1142 // We have a null argument, branch to ic_miss_stub. 1143 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1144 relocInfo::runtime_call_type); 1145 __ BIND(valid); 1146 } 1147 } 1148 // Assume argument is not NULL, load klass from receiver. 1149 __ load_klass(receiver_klass, R3_ARG1); 1150 1151 __ ld(ic_klass, CompiledICHolder::holder_klass_offset(), ic); 1152 1153 if (TrapBasedICMissChecks) { 1154 __ trap_ic_miss_check(receiver_klass, ic_klass); 1155 } else { 1156 Label valid; 1157 __ cmpd(CCR0, receiver_klass, ic_klass); 1158 __ beq_predict_taken(CCR0, valid); 1159 // We have an unexpected klass, branch to ic_miss_stub. 1160 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1161 relocInfo::runtime_call_type); 1162 __ BIND(valid); 1163 } 1164 1165 // Argument is valid and klass is as expected, continue. 1166 1167 // Extract method from inline cache, verified entry point needs it. 1168 __ ld(R19_method, CompiledICHolder::holder_method_offset(), ic); 1169 assert(R19_method == ic, "the inline cache register is dead here"); 1170 1171 __ ld(code, method_(code)); 1172 __ cmpdi(CCR0, code, 0); 1173 __ ld(ientry, method_(interpreter_entry)); // preloaded 1174 __ beq_predict_taken(CCR0, call_interpreter); 1175 1176 // Branch to ic_miss_stub. 1177 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1178 relocInfo::runtime_call_type); 1179 1180 // entry: c2i 1181 1182 c2i_entry = gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry); 1183 1184 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 1185 } 1186 1187 #ifdef COMPILER2 1188 // An oop arg. Must pass a handle not the oop itself. 1189 static void object_move(MacroAssembler* masm, 1190 int frame_size_in_slots, 1191 OopMap* oop_map, int oop_handle_offset, 1192 bool is_receiver, int* receiver_offset, 1193 VMRegPair src, VMRegPair dst, 1194 Register r_caller_sp, Register r_temp_1, Register r_temp_2) { 1195 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)), 1196 "receiver has already been moved"); 1197 1198 // We must pass a handle. First figure out the location we use as a handle. 1199 1200 if (src.first()->is_stack()) { 1201 // stack to stack or reg 1202 1203 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register(); 1204 Label skip; 1205 const int oop_slot_in_callers_frame = reg2slot(src.first()); 1206 1207 guarantee(!is_receiver, "expecting receiver in register"); 1208 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots)); 1209 1210 __ addi(r_handle, r_caller_sp, reg2offset(src.first())); 1211 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp); 1212 __ cmpdi(CCR0, r_temp_2, 0); 1213 __ bne(CCR0, skip); 1214 // Use a NULL handle if oop is NULL. 1215 __ li(r_handle, 0); 1216 __ bind(skip); 1217 1218 if (dst.first()->is_stack()) { 1219 // stack to stack 1220 __ std(r_handle, reg2offset(dst.first()), R1_SP); 1221 } else { 1222 // stack to reg 1223 // Nothing to do, r_handle is already the dst register. 1224 } 1225 } else { 1226 // reg to stack or reg 1227 const Register r_oop = src.first()->as_Register(); 1228 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register(); 1229 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word 1230 + oop_handle_offset; // in slots 1231 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size; 1232 Label skip; 1233 1234 if (is_receiver) { 1235 *receiver_offset = oop_offset; 1236 } 1237 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1238 1239 __ std( r_oop, oop_offset, R1_SP); 1240 __ addi(r_handle, R1_SP, oop_offset); 1241 1242 __ cmpdi(CCR0, r_oop, 0); 1243 __ bne(CCR0, skip); 1244 // Use a NULL handle if oop is NULL. 1245 __ li(r_handle, 0); 1246 __ bind(skip); 1247 1248 if (dst.first()->is_stack()) { 1249 // reg to stack 1250 __ std(r_handle, reg2offset(dst.first()), R1_SP); 1251 } else { 1252 // reg to reg 1253 // Nothing to do, r_handle is already the dst register. 1254 } 1255 } 1256 } 1257 1258 static void int_move(MacroAssembler*masm, 1259 VMRegPair src, VMRegPair dst, 1260 Register r_caller_sp, Register r_temp) { 1261 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long-int"); 1262 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long"); 1263 1264 if (src.first()->is_stack()) { 1265 if (dst.first()->is_stack()) { 1266 // stack to stack 1267 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp); 1268 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1269 } else { 1270 // stack to reg 1271 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1272 } 1273 } else if (dst.first()->is_stack()) { 1274 // reg to stack 1275 __ extsw(r_temp, src.first()->as_Register()); 1276 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1277 } else { 1278 // reg to reg 1279 __ extsw(dst.first()->as_Register(), src.first()->as_Register()); 1280 } 1281 } 1282 1283 static void long_move(MacroAssembler*masm, 1284 VMRegPair src, VMRegPair dst, 1285 Register r_caller_sp, Register r_temp) { 1286 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long"); 1287 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long"); 1288 1289 if (src.first()->is_stack()) { 1290 if (dst.first()->is_stack()) { 1291 // stack to stack 1292 __ ld( r_temp, reg2offset(src.first()), r_caller_sp); 1293 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1294 } else { 1295 // stack to reg 1296 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1297 } 1298 } else if (dst.first()->is_stack()) { 1299 // reg to stack 1300 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP); 1301 } else { 1302 // reg to reg 1303 if (dst.first()->as_Register() != src.first()->as_Register()) 1304 __ mr(dst.first()->as_Register(), src.first()->as_Register()); 1305 } 1306 } 1307 1308 static void float_move(MacroAssembler*masm, 1309 VMRegPair src, VMRegPair dst, 1310 Register r_caller_sp, Register r_temp) { 1311 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float"); 1312 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float"); 1313 1314 if (src.first()->is_stack()) { 1315 if (dst.first()->is_stack()) { 1316 // stack to stack 1317 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp); 1318 __ stw(r_temp, reg2offset(dst.first()), R1_SP); 1319 } else { 1320 // stack to reg 1321 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp); 1322 } 1323 } else if (dst.first()->is_stack()) { 1324 // reg to stack 1325 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP); 1326 } else { 1327 // reg to reg 1328 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister()) 1329 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1330 } 1331 } 1332 1333 static void double_move(MacroAssembler*masm, 1334 VMRegPair src, VMRegPair dst, 1335 Register r_caller_sp, Register r_temp) { 1336 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double"); 1337 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double"); 1338 1339 if (src.first()->is_stack()) { 1340 if (dst.first()->is_stack()) { 1341 // stack to stack 1342 __ ld( r_temp, reg2offset(src.first()), r_caller_sp); 1343 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1344 } else { 1345 // stack to reg 1346 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp); 1347 } 1348 } else if (dst.first()->is_stack()) { 1349 // reg to stack 1350 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP); 1351 } else { 1352 // reg to reg 1353 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister()) 1354 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1355 } 1356 } 1357 1358 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1359 switch (ret_type) { 1360 case T_BOOLEAN: 1361 case T_CHAR: 1362 case T_BYTE: 1363 case T_SHORT: 1364 case T_INT: 1365 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1366 break; 1367 case T_ARRAY: 1368 case T_OBJECT: 1369 case T_LONG: 1370 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1371 break; 1372 case T_FLOAT: 1373 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1374 break; 1375 case T_DOUBLE: 1376 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1377 break; 1378 case T_VOID: 1379 break; 1380 default: 1381 ShouldNotReachHere(); 1382 break; 1383 } 1384 } 1385 1386 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1387 switch (ret_type) { 1388 case T_BOOLEAN: 1389 case T_CHAR: 1390 case T_BYTE: 1391 case T_SHORT: 1392 case T_INT: 1393 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1394 break; 1395 case T_ARRAY: 1396 case T_OBJECT: 1397 case T_LONG: 1398 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1399 break; 1400 case T_FLOAT: 1401 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1402 break; 1403 case T_DOUBLE: 1404 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1405 break; 1406 case T_VOID: 1407 break; 1408 default: 1409 ShouldNotReachHere(); 1410 break; 1411 } 1412 } 1413 1414 static void save_or_restore_arguments(MacroAssembler* masm, 1415 const int stack_slots, 1416 const int total_in_args, 1417 const int arg_save_area, 1418 OopMap* map, 1419 VMRegPair* in_regs, 1420 BasicType* in_sig_bt) { 1421 // If map is non-NULL then the code should store the values, 1422 // otherwise it should load them. 1423 int slot = arg_save_area; 1424 // Save down double word first. 1425 for (int i = 0; i < total_in_args; i++) { 1426 if (in_regs[i].first()->is_FloatRegister() && in_sig_bt[i] == T_DOUBLE) { 1427 int offset = slot * VMRegImpl::stack_slot_size; 1428 slot += VMRegImpl::slots_per_word; 1429 assert(slot <= stack_slots, "overflow (after DOUBLE stack slot)"); 1430 if (map != NULL) { 1431 __ stfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1432 } else { 1433 __ lfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1434 } 1435 } else if (in_regs[i].first()->is_Register() && 1436 (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_ARRAY)) { 1437 int offset = slot * VMRegImpl::stack_slot_size; 1438 if (map != NULL) { 1439 __ std(in_regs[i].first()->as_Register(), offset, R1_SP); 1440 if (in_sig_bt[i] == T_ARRAY) { 1441 map->set_oop(VMRegImpl::stack2reg(slot)); 1442 } 1443 } else { 1444 __ ld(in_regs[i].first()->as_Register(), offset, R1_SP); 1445 } 1446 slot += VMRegImpl::slots_per_word; 1447 assert(slot <= stack_slots, "overflow (after LONG/ARRAY stack slot)"); 1448 } 1449 } 1450 // Save or restore single word registers. 1451 for (int i = 0; i < total_in_args; i++) { 1452 // PPC64: pass ints as longs: must only deal with floats here. 1453 if (in_regs[i].first()->is_FloatRegister()) { 1454 if (in_sig_bt[i] == T_FLOAT) { 1455 int offset = slot * VMRegImpl::stack_slot_size; 1456 slot++; 1457 assert(slot <= stack_slots, "overflow (after FLOAT stack slot)"); 1458 if (map != NULL) { 1459 __ stfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1460 } else { 1461 __ lfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1462 } 1463 } 1464 } else if (in_regs[i].first()->is_stack()) { 1465 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1466 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1467 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots)); 1468 } 1469 } 1470 } 1471 } 1472 1473 // Check GC_locker::needs_gc and enter the runtime if it's true. This 1474 // keeps a new JNI critical region from starting until a GC has been 1475 // forced. Save down any oops in registers and describe them in an 1476 // OopMap. 1477 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1478 const int stack_slots, 1479 const int total_in_args, 1480 const int arg_save_area, 1481 OopMapSet* oop_maps, 1482 VMRegPair* in_regs, 1483 BasicType* in_sig_bt, 1484 Register tmp_reg ) { 1485 __ block_comment("check GC_locker::needs_gc"); 1486 Label cont; 1487 __ lbz(tmp_reg, (RegisterOrConstant)(intptr_t)GC_locker::needs_gc_address()); 1488 __ cmplwi(CCR0, tmp_reg, 0); 1489 __ beq(CCR0, cont); 1490 1491 // Save down any values that are live in registers and call into the 1492 // runtime to halt for a GC. 1493 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1494 save_or_restore_arguments(masm, stack_slots, total_in_args, 1495 arg_save_area, map, in_regs, in_sig_bt); 1496 1497 __ mr(R3_ARG1, R16_thread); 1498 __ set_last_Java_frame(R1_SP, noreg); 1499 1500 __ block_comment("block_for_jni_critical"); 1501 address entry_point = CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical); 1502 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, entry_point), relocInfo::runtime_call_type); 1503 address start = __ pc() - __ offset(), 1504 calls_return_pc = __ last_calls_return_pc(); 1505 oop_maps->add_gc_map(calls_return_pc - start, map); 1506 1507 __ reset_last_Java_frame(); 1508 1509 // Reload all the register arguments. 1510 save_or_restore_arguments(masm, stack_slots, total_in_args, 1511 arg_save_area, NULL, in_regs, in_sig_bt); 1512 1513 __ BIND(cont); 1514 1515 #ifdef ASSERT 1516 if (StressCriticalJNINatives) { 1517 // Stress register saving. 1518 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1519 save_or_restore_arguments(masm, stack_slots, total_in_args, 1520 arg_save_area, map, in_regs, in_sig_bt); 1521 // Destroy argument registers. 1522 for (int i = 0; i < total_in_args; i++) { 1523 if (in_regs[i].first()->is_Register()) { 1524 const Register reg = in_regs[i].first()->as_Register(); 1525 __ neg(reg, reg); 1526 } else if (in_regs[i].first()->is_FloatRegister()) { 1527 __ fneg(in_regs[i].first()->as_FloatRegister(), in_regs[i].first()->as_FloatRegister()); 1528 } 1529 } 1530 1531 save_or_restore_arguments(masm, stack_slots, total_in_args, 1532 arg_save_area, NULL, in_regs, in_sig_bt); 1533 } 1534 #endif 1535 } 1536 1537 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst, Register r_caller_sp, Register r_temp) { 1538 if (src.first()->is_stack()) { 1539 if (dst.first()->is_stack()) { 1540 // stack to stack 1541 __ ld(r_temp, reg2offset(src.first()), r_caller_sp); 1542 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1543 } else { 1544 // stack to reg 1545 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1546 } 1547 } else if (dst.first()->is_stack()) { 1548 // reg to stack 1549 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP); 1550 } else { 1551 if (dst.first() != src.first()) { 1552 __ mr(dst.first()->as_Register(), src.first()->as_Register()); 1553 } 1554 } 1555 } 1556 1557 // Unpack an array argument into a pointer to the body and the length 1558 // if the array is non-null, otherwise pass 0 for both. 1559 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, 1560 VMRegPair body_arg, VMRegPair length_arg, Register r_caller_sp, 1561 Register tmp_reg, Register tmp2_reg) { 1562 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg, 1563 "possible collision"); 1564 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg, 1565 "possible collision"); 1566 1567 // Pass the length, ptr pair. 1568 Label set_out_args; 1569 VMRegPair tmp, tmp2; 1570 tmp.set_ptr(tmp_reg->as_VMReg()); 1571 tmp2.set_ptr(tmp2_reg->as_VMReg()); 1572 if (reg.first()->is_stack()) { 1573 // Load the arg up from the stack. 1574 move_ptr(masm, reg, tmp, r_caller_sp, /*unused*/ R0); 1575 reg = tmp; 1576 } 1577 __ li(tmp2_reg, 0); // Pass zeros if Array=null. 1578 if (tmp_reg != reg.first()->as_Register()) __ li(tmp_reg, 0); 1579 __ cmpdi(CCR0, reg.first()->as_Register(), 0); 1580 __ beq(CCR0, set_out_args); 1581 __ lwa(tmp2_reg, arrayOopDesc::length_offset_in_bytes(), reg.first()->as_Register()); 1582 __ addi(tmp_reg, reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type)); 1583 __ bind(set_out_args); 1584 move_ptr(masm, tmp, body_arg, r_caller_sp, /*unused*/ R0); 1585 move_ptr(masm, tmp2, length_arg, r_caller_sp, /*unused*/ R0); // Same as move32_64 on PPC64. 1586 } 1587 1588 static void verify_oop_args(MacroAssembler* masm, 1589 methodHandle method, 1590 const BasicType* sig_bt, 1591 const VMRegPair* regs) { 1592 Register temp_reg = R19_method; // not part of any compiled calling seq 1593 if (VerifyOops) { 1594 for (int i = 0; i < method->size_of_parameters(); i++) { 1595 if (sig_bt[i] == T_OBJECT || 1596 sig_bt[i] == T_ARRAY) { 1597 VMReg r = regs[i].first(); 1598 assert(r->is_valid(), "bad oop arg"); 1599 if (r->is_stack()) { 1600 __ ld(temp_reg, reg2offset(r), R1_SP); 1601 __ verify_oop(temp_reg); 1602 } else { 1603 __ verify_oop(r->as_Register()); 1604 } 1605 } 1606 } 1607 } 1608 } 1609 1610 static void gen_special_dispatch(MacroAssembler* masm, 1611 methodHandle method, 1612 const BasicType* sig_bt, 1613 const VMRegPair* regs) { 1614 verify_oop_args(masm, method, sig_bt, regs); 1615 vmIntrinsics::ID iid = method->intrinsic_id(); 1616 1617 // Now write the args into the outgoing interpreter space 1618 bool has_receiver = false; 1619 Register receiver_reg = noreg; 1620 int member_arg_pos = -1; 1621 Register member_reg = noreg; 1622 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1623 if (ref_kind != 0) { 1624 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1625 member_reg = R19_method; // known to be free at this point 1626 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1627 } else if (iid == vmIntrinsics::_invokeBasic) { 1628 has_receiver = true; 1629 } else { 1630 fatal(err_msg_res("unexpected intrinsic id %d", iid)); 1631 } 1632 1633 if (member_reg != noreg) { 1634 // Load the member_arg into register, if necessary. 1635 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1636 VMReg r = regs[member_arg_pos].first(); 1637 if (r->is_stack()) { 1638 __ ld(member_reg, reg2offset(r), R1_SP); 1639 } else { 1640 // no data motion is needed 1641 member_reg = r->as_Register(); 1642 } 1643 } 1644 1645 if (has_receiver) { 1646 // Make sure the receiver is loaded into a register. 1647 assert(method->size_of_parameters() > 0, "oob"); 1648 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1649 VMReg r = regs[0].first(); 1650 assert(r->is_valid(), "bad receiver arg"); 1651 if (r->is_stack()) { 1652 // Porting note: This assumes that compiled calling conventions always 1653 // pass the receiver oop in a register. If this is not true on some 1654 // platform, pick a temp and load the receiver from stack. 1655 fatal("receiver always in a register"); 1656 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point? 1657 __ ld(receiver_reg, reg2offset(r), R1_SP); 1658 } else { 1659 // no data motion is needed 1660 receiver_reg = r->as_Register(); 1661 } 1662 } 1663 1664 // Figure out which address we are really jumping to: 1665 MethodHandles::generate_method_handle_dispatch(masm, iid, 1666 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1667 } 1668 1669 #endif // COMPILER2 1670 1671 // --------------------------------------------------------------------------- 1672 // Generate a native wrapper for a given method. The method takes arguments 1673 // in the Java compiled code convention, marshals them to the native 1674 // convention (handlizes oops, etc), transitions to native, makes the call, 1675 // returns to java state (possibly blocking), unhandlizes any result and 1676 // returns. 1677 // 1678 // Critical native functions are a shorthand for the use of 1679 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1680 // functions. The wrapper is expected to unpack the arguments before 1681 // passing them to the callee and perform checks before and after the 1682 // native call to ensure that they GC_locker 1683 // lock_critical/unlock_critical semantics are followed. Some other 1684 // parts of JNI setup are skipped like the tear down of the JNI handle 1685 // block and the check for pending exceptions it's impossible for them 1686 // to be thrown. 1687 // 1688 // They are roughly structured like this: 1689 // if (GC_locker::needs_gc()) 1690 // SharedRuntime::block_for_jni_critical(); 1691 // tranistion to thread_in_native 1692 // unpack arrray arguments and call native entry point 1693 // check for safepoint in progress 1694 // check if any thread suspend flags are set 1695 // call into JVM and possible unlock the JNI critical 1696 // if a GC was suppressed while in the critical native. 1697 // transition back to thread_in_Java 1698 // return to caller 1699 // 1700 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm, 1701 methodHandle method, 1702 int compile_id, 1703 BasicType *in_sig_bt, 1704 VMRegPair *in_regs, 1705 BasicType ret_type) { 1706 #ifdef COMPILER2 1707 if (method->is_method_handle_intrinsic()) { 1708 vmIntrinsics::ID iid = method->intrinsic_id(); 1709 intptr_t start = (intptr_t)__ pc(); 1710 int vep_offset = ((intptr_t)__ pc()) - start; 1711 gen_special_dispatch(masm, 1712 method, 1713 in_sig_bt, 1714 in_regs); 1715 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1716 __ flush(); 1717 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1718 return nmethod::new_native_nmethod(method, 1719 compile_id, 1720 masm->code(), 1721 vep_offset, 1722 frame_complete, 1723 stack_slots / VMRegImpl::slots_per_word, 1724 in_ByteSize(-1), 1725 in_ByteSize(-1), 1726 (OopMapSet*)NULL); 1727 } 1728 1729 bool is_critical_native = true; 1730 address native_func = method->critical_native_function(); 1731 if (native_func == NULL) { 1732 native_func = method->native_function(); 1733 is_critical_native = false; 1734 } 1735 assert(native_func != NULL, "must have function"); 1736 1737 // First, create signature for outgoing C call 1738 // -------------------------------------------------------------------------- 1739 1740 int total_in_args = method->size_of_parameters(); 1741 // We have received a description of where all the java args are located 1742 // on entry to the wrapper. We need to convert these args to where 1743 // the jni function will expect them. To figure out where they go 1744 // we convert the java signature to a C signature by inserting 1745 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1746 // 1747 // Additionally, on ppc64 we must convert integers to longs in the C 1748 // signature. We do this in advance in order to have no trouble with 1749 // indexes into the bt-arrays. 1750 // So convert the signature and registers now, and adjust the total number 1751 // of in-arguments accordingly. 1752 int i2l_argcnt = convert_ints_to_longints_argcnt(total_in_args, in_sig_bt); // PPC64: pass ints as longs. 1753 1754 // Calculate the total number of C arguments and create arrays for the 1755 // signature and the outgoing registers. 1756 // On ppc64, we have two arrays for the outgoing registers, because 1757 // some floating-point arguments must be passed in registers _and_ 1758 // in stack locations. 1759 bool method_is_static = method->is_static(); 1760 int total_c_args = i2l_argcnt; 1761 1762 if (!is_critical_native) { 1763 int n_hidden_args = method_is_static ? 2 : 1; 1764 total_c_args += n_hidden_args; 1765 } else { 1766 // No JNIEnv*, no this*, but unpacked arrays (base+length). 1767 for (int i = 0; i < total_in_args; i++) { 1768 if (in_sig_bt[i] == T_ARRAY) { 1769 total_c_args += 2; // PPC64: T_LONG, T_INT, T_ADDRESS (see convert_ints_to_longints and c_calling_convention) 1770 } 1771 } 1772 } 1773 1774 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1775 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1776 VMRegPair *out_regs2 = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1777 BasicType* in_elem_bt = NULL; 1778 1779 // Create the signature for the C call: 1780 // 1) add the JNIEnv* 1781 // 2) add the class if the method is static 1782 // 3) copy the rest of the incoming signature (shifted by the number of 1783 // hidden arguments). 1784 1785 int argc = 0; 1786 if (!is_critical_native) { 1787 convert_ints_to_longints(i2l_argcnt, total_in_args, in_sig_bt, in_regs); // PPC64: pass ints as longs. 1788 1789 out_sig_bt[argc++] = T_ADDRESS; 1790 if (method->is_static()) { 1791 out_sig_bt[argc++] = T_OBJECT; 1792 } 1793 1794 for (int i = 0; i < total_in_args ; i++ ) { 1795 out_sig_bt[argc++] = in_sig_bt[i]; 1796 } 1797 } else { 1798 Thread* THREAD = Thread::current(); 1799 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, i2l_argcnt); 1800 SignatureStream ss(method->signature()); 1801 int o = 0; 1802 for (int i = 0; i < total_in_args ; i++, o++) { 1803 if (in_sig_bt[i] == T_ARRAY) { 1804 // Arrays are passed as int, elem* pair 1805 Symbol* atype = ss.as_symbol(CHECK_NULL); 1806 const char* at = atype->as_C_string(); 1807 if (strlen(at) == 2) { 1808 assert(at[0] == '[', "must be"); 1809 switch (at[1]) { 1810 case 'B': in_elem_bt[o] = T_BYTE; break; 1811 case 'C': in_elem_bt[o] = T_CHAR; break; 1812 case 'D': in_elem_bt[o] = T_DOUBLE; break; 1813 case 'F': in_elem_bt[o] = T_FLOAT; break; 1814 case 'I': in_elem_bt[o] = T_INT; break; 1815 case 'J': in_elem_bt[o] = T_LONG; break; 1816 case 'S': in_elem_bt[o] = T_SHORT; break; 1817 case 'Z': in_elem_bt[o] = T_BOOLEAN; break; 1818 default: ShouldNotReachHere(); 1819 } 1820 } 1821 } else { 1822 in_elem_bt[o] = T_VOID; 1823 switch(in_sig_bt[i]) { // PPC64: pass ints as longs. 1824 case T_BOOLEAN: 1825 case T_CHAR: 1826 case T_BYTE: 1827 case T_SHORT: 1828 case T_INT: in_elem_bt[++o] = T_VOID; break; 1829 default: break; 1830 } 1831 } 1832 if (in_sig_bt[i] != T_VOID) { 1833 assert(in_sig_bt[i] == ss.type(), "must match"); 1834 ss.next(); 1835 } 1836 } 1837 assert(i2l_argcnt==o, "must match"); 1838 1839 convert_ints_to_longints(i2l_argcnt, total_in_args, in_sig_bt, in_regs); // PPC64: pass ints as longs. 1840 1841 for (int i = 0; i < total_in_args ; i++ ) { 1842 if (in_sig_bt[i] == T_ARRAY) { 1843 // Arrays are passed as int, elem* pair. 1844 out_sig_bt[argc++] = T_LONG; // PPC64: pass ints as longs. 1845 out_sig_bt[argc++] = T_INT; 1846 out_sig_bt[argc++] = T_ADDRESS; 1847 } else { 1848 out_sig_bt[argc++] = in_sig_bt[i]; 1849 } 1850 } 1851 } 1852 1853 1854 // Compute the wrapper's frame size. 1855 // -------------------------------------------------------------------------- 1856 1857 // Now figure out where the args must be stored and how much stack space 1858 // they require. 1859 // 1860 // Compute framesize for the wrapper. We need to handlize all oops in 1861 // incoming registers. 1862 // 1863 // Calculate the total number of stack slots we will need: 1864 // 1) abi requirements 1865 // 2) outgoing arguments 1866 // 3) space for inbound oop handle area 1867 // 4) space for handlizing a klass if static method 1868 // 5) space for a lock if synchronized method 1869 // 6) workspace for saving return values, int <-> float reg moves, etc. 1870 // 7) alignment 1871 // 1872 // Layout of the native wrapper frame: 1873 // (stack grows upwards, memory grows downwards) 1874 // 1875 // NW [ABI_112] <-- 1) R1_SP 1876 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset 1877 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset (save area for critical natives) 1878 // klass <-- 4) R1_SP + klass_offset 1879 // lock <-- 5) R1_SP + lock_offset 1880 // [workspace] <-- 6) R1_SP + workspace_offset 1881 // [alignment] (optional) <-- 7) 1882 // caller [JIT_TOP_ABI_48] <-- r_callers_sp 1883 // 1884 // - *_slot_offset Indicates offset from SP in number of stack slots. 1885 // - *_offset Indicates offset from SP in bytes. 1886 1887 int stack_slots = c_calling_convention(out_sig_bt, out_regs, out_regs2, total_c_args) // 1+2) 1888 + SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention. 1889 1890 // Now the space for the inbound oop handle area. 1891 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word; 1892 if (is_critical_native) { 1893 // Critical natives may have to call out so they need a save area 1894 // for register arguments. 1895 int double_slots = 0; 1896 int single_slots = 0; 1897 for (int i = 0; i < total_in_args; i++) { 1898 if (in_regs[i].first()->is_Register()) { 1899 const Register reg = in_regs[i].first()->as_Register(); 1900 switch (in_sig_bt[i]) { 1901 case T_BOOLEAN: 1902 case T_BYTE: 1903 case T_SHORT: 1904 case T_CHAR: 1905 case T_INT: /*single_slots++;*/ break; // PPC64: pass ints as longs. 1906 case T_ARRAY: 1907 case T_LONG: double_slots++; break; 1908 default: ShouldNotReachHere(); 1909 } 1910 } else if (in_regs[i].first()->is_FloatRegister()) { 1911 switch (in_sig_bt[i]) { 1912 case T_FLOAT: single_slots++; break; 1913 case T_DOUBLE: double_slots++; break; 1914 default: ShouldNotReachHere(); 1915 } 1916 } 1917 } 1918 total_save_slots = double_slots * 2 + round_to(single_slots, 2); // round to even 1919 } 1920 1921 int oop_handle_slot_offset = stack_slots; 1922 stack_slots += total_save_slots; // 3) 1923 1924 int klass_slot_offset = 0; 1925 int klass_offset = -1; 1926 if (method_is_static && !is_critical_native) { // 4) 1927 klass_slot_offset = stack_slots; 1928 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1929 stack_slots += VMRegImpl::slots_per_word; 1930 } 1931 1932 int lock_slot_offset = 0; 1933 int lock_offset = -1; 1934 if (method->is_synchronized()) { // 5) 1935 lock_slot_offset = stack_slots; 1936 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size; 1937 stack_slots += VMRegImpl::slots_per_word; 1938 } 1939 1940 int workspace_slot_offset = stack_slots; // 6) 1941 stack_slots += 2; 1942 1943 // Now compute actual number of stack words we need. 1944 // Rounding to make stack properly aligned. 1945 stack_slots = round_to(stack_slots, // 7) 1946 frame::alignment_in_bytes / VMRegImpl::stack_slot_size); 1947 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size; 1948 1949 1950 // Now we can start generating code. 1951 // -------------------------------------------------------------------------- 1952 1953 intptr_t start_pc = (intptr_t)__ pc(); 1954 intptr_t vep_start_pc; 1955 intptr_t frame_done_pc; 1956 intptr_t oopmap_pc; 1957 1958 Label ic_miss; 1959 Label handle_pending_exception; 1960 1961 Register r_callers_sp = R21; 1962 Register r_temp_1 = R22; 1963 Register r_temp_2 = R23; 1964 Register r_temp_3 = R24; 1965 Register r_temp_4 = R25; 1966 Register r_temp_5 = R26; 1967 Register r_temp_6 = R27; 1968 Register r_return_pc = R28; 1969 1970 Register r_carg1_jnienv = noreg; 1971 Register r_carg2_classorobject = noreg; 1972 if (!is_critical_native) { 1973 r_carg1_jnienv = out_regs[0].first()->as_Register(); 1974 r_carg2_classorobject = out_regs[1].first()->as_Register(); 1975 } 1976 1977 1978 // Generate the Unverified Entry Point (UEP). 1979 // -------------------------------------------------------------------------- 1980 assert(start_pc == (intptr_t)__ pc(), "uep must be at start"); 1981 1982 // Check ic: object class == cached class? 1983 if (!method_is_static) { 1984 Register ic = as_Register(Matcher::inline_cache_reg_encode()); 1985 Register receiver_klass = r_temp_1; 1986 1987 __ cmpdi(CCR0, R3_ARG1, 0); 1988 __ beq(CCR0, ic_miss); 1989 __ verify_oop(R3_ARG1); 1990 __ load_klass(receiver_klass, R3_ARG1); 1991 1992 __ cmpd(CCR0, receiver_klass, ic); 1993 __ bne(CCR0, ic_miss); 1994 } 1995 1996 1997 // Generate the Verified Entry Point (VEP). 1998 // -------------------------------------------------------------------------- 1999 vep_start_pc = (intptr_t)__ pc(); 2000 2001 __ save_LR_CR(r_temp_1); 2002 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame. 2003 __ mr(r_callers_sp, R1_SP); // Remember frame pointer. 2004 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame. 2005 frame_done_pc = (intptr_t)__ pc(); 2006 2007 // Native nmethod wrappers never take possesion of the oop arguments. 2008 // So the caller will gc the arguments. 2009 // The only thing we need an oopMap for is if the call is static. 2010 // 2011 // An OopMap for lock (and class if static), and one for the VM call itself. 2012 OopMapSet *oop_maps = new OopMapSet(); 2013 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 2014 2015 if (is_critical_native) { 2016 check_needs_gc_for_critical_native(masm, stack_slots, total_in_args, oop_handle_slot_offset, oop_maps, in_regs, in_sig_bt, r_temp_1); 2017 } 2018 2019 // Move arguments from register/stack to register/stack. 2020 // -------------------------------------------------------------------------- 2021 // 2022 // We immediately shuffle the arguments so that for any vm call we have 2023 // to make from here on out (sync slow path, jvmti, etc.) we will have 2024 // captured the oops from our caller and have a valid oopMap for them. 2025 // 2026 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv* 2027 // (derived from JavaThread* which is in R16_thread) and, if static, 2028 // the class mirror instead of a receiver. This pretty much guarantees that 2029 // register layout will not match. We ignore these extra arguments during 2030 // the shuffle. The shuffle is described by the two calling convention 2031 // vectors we have in our possession. We simply walk the java vector to 2032 // get the source locations and the c vector to get the destinations. 2033 2034 // Record sp-based slot for receiver on stack for non-static methods. 2035 int receiver_offset = -1; 2036 2037 // We move the arguments backward because the floating point registers 2038 // destination will always be to a register with a greater or equal 2039 // register number or the stack. 2040 // in is the index of the incoming Java arguments 2041 // out is the index of the outgoing C arguments 2042 2043 #ifdef ASSERT 2044 bool reg_destroyed[RegisterImpl::number_of_registers]; 2045 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 2046 for (int r = 0 ; r < RegisterImpl::number_of_registers ; r++) { 2047 reg_destroyed[r] = false; 2048 } 2049 for (int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++) { 2050 freg_destroyed[f] = false; 2051 } 2052 #endif // ASSERT 2053 2054 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) { 2055 2056 #ifdef ASSERT 2057 if (in_regs[in].first()->is_Register()) { 2058 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!"); 2059 } else if (in_regs[in].first()->is_FloatRegister()) { 2060 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!"); 2061 } 2062 if (out_regs[out].first()->is_Register()) { 2063 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true; 2064 } else if (out_regs[out].first()->is_FloatRegister()) { 2065 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true; 2066 } 2067 if (out_regs2[out].first()->is_Register()) { 2068 reg_destroyed[out_regs2[out].first()->as_Register()->encoding()] = true; 2069 } else if (out_regs2[out].first()->is_FloatRegister()) { 2070 freg_destroyed[out_regs2[out].first()->as_FloatRegister()->encoding()] = true; 2071 } 2072 #endif // ASSERT 2073 2074 switch (in_sig_bt[in]) { 2075 case T_BOOLEAN: 2076 case T_CHAR: 2077 case T_BYTE: 2078 case T_SHORT: 2079 case T_INT: 2080 guarantee(in > 0 && in_sig_bt[in-1] == T_LONG, 2081 "expecting type (T_LONG,bt) for bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}"); 2082 break; 2083 case T_LONG: 2084 if (in_sig_bt[in+1] == T_VOID) { 2085 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2086 } else { 2087 guarantee(in_sig_bt[in+1] == T_BOOLEAN || in_sig_bt[in+1] == T_CHAR || 2088 in_sig_bt[in+1] == T_BYTE || in_sig_bt[in+1] == T_SHORT || 2089 in_sig_bt[in+1] == T_INT, 2090 "expecting type (T_LONG,bt) for bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}"); 2091 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2092 } 2093 break; 2094 case T_ARRAY: 2095 if (is_critical_native) { 2096 int body_arg = out; 2097 out -= 2; // Point to length arg. PPC64: pass ints as longs. 2098 unpack_array_argument(masm, in_regs[in], in_elem_bt[in], out_regs[body_arg], out_regs[out], 2099 r_callers_sp, r_temp_1, r_temp_2); 2100 break; 2101 } 2102 case T_OBJECT: 2103 assert(!is_critical_native, "no oop arguments"); 2104 object_move(masm, stack_slots, 2105 oop_map, oop_handle_slot_offset, 2106 ((in == 0) && (!method_is_static)), &receiver_offset, 2107 in_regs[in], out_regs[out], 2108 r_callers_sp, r_temp_1, r_temp_2); 2109 break; 2110 case T_VOID: 2111 break; 2112 case T_FLOAT: 2113 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2114 if (out_regs2[out].first()->is_valid()) { 2115 float_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1); 2116 } 2117 break; 2118 case T_DOUBLE: 2119 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2120 if (out_regs2[out].first()->is_valid()) { 2121 double_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1); 2122 } 2123 break; 2124 case T_ADDRESS: 2125 fatal("found type (T_ADDRESS) in java args"); 2126 break; 2127 default: 2128 ShouldNotReachHere(); 2129 break; 2130 } 2131 } 2132 2133 // Pre-load a static method's oop into ARG2. 2134 // Used both by locking code and the normal JNI call code. 2135 if (method_is_static && !is_critical_native) { 2136 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()), 2137 r_carg2_classorobject); 2138 2139 // Now handlize the static class mirror in carg2. It's known not-null. 2140 __ std(r_carg2_classorobject, klass_offset, R1_SP); 2141 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 2142 __ addi(r_carg2_classorobject, R1_SP, klass_offset); 2143 } 2144 2145 // Get JNIEnv* which is first argument to native. 2146 if (!is_critical_native) { 2147 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset())); 2148 } 2149 2150 // NOTE: 2151 // 2152 // We have all of the arguments setup at this point. 2153 // We MUST NOT touch any outgoing regs from this point on. 2154 // So if we must call out we must push a new frame. 2155 2156 // Get current pc for oopmap, and load it patchable relative to global toc. 2157 oopmap_pc = (intptr_t) __ pc(); 2158 __ calculate_address_from_global_toc(r_return_pc, (address)oopmap_pc, true, true, true, true); 2159 2160 // We use the same pc/oopMap repeatedly when we call out. 2161 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map); 2162 2163 // r_return_pc now has the pc loaded that we will use when we finally call 2164 // to native. 2165 2166 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below. 2167 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register"); 2168 2169 2170 # if 0 2171 // DTrace method entry 2172 # endif 2173 2174 // Lock a synchronized method. 2175 // -------------------------------------------------------------------------- 2176 2177 if (method->is_synchronized()) { 2178 assert(!is_critical_native, "unhandled"); 2179 ConditionRegister r_flag = CCR1; 2180 Register r_oop = r_temp_4; 2181 const Register r_box = r_temp_5; 2182 Label done, locked; 2183 2184 // Load the oop for the object or class. r_carg2_classorobject contains 2185 // either the handlized oop from the incoming arguments or the handlized 2186 // class mirror (if the method is static). 2187 __ ld(r_oop, 0, r_carg2_classorobject); 2188 2189 // Get the lock box slot's address. 2190 __ addi(r_box, R1_SP, lock_offset); 2191 2192 # ifdef ASSERT 2193 if (UseBiasedLocking) { 2194 // Making the box point to itself will make it clear it went unused 2195 // but also be obviously invalid. 2196 __ std(r_box, 0, r_box); 2197 } 2198 # endif // ASSERT 2199 2200 // Try fastpath for locking. 2201 // fast_lock kills r_temp_1, r_temp_2, r_temp_3. 2202 __ compiler_fast_lock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3); 2203 __ beq(r_flag, locked); 2204 2205 // None of the above fast optimizations worked so we have to get into the 2206 // slow case of monitor enter. Inline a special case of call_VM that 2207 // disallows any pending_exception. 2208 2209 // Save argument registers and leave room for C-compatible ABI_112. 2210 int frame_size = frame::abi_112_size + 2211 round_to(total_c_args * wordSize, frame::alignment_in_bytes); 2212 __ mr(R11_scratch1, R1_SP); 2213 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs, out_regs2); 2214 2215 // Do the call. 2216 __ set_last_Java_frame(R11_scratch1, r_return_pc); 2217 assert(r_return_pc->is_nonvolatile(), "expecting return pc to be in non-volatile register"); 2218 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread); 2219 __ reset_last_Java_frame(); 2220 2221 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs, out_regs2); 2222 2223 __ asm_assert_mem8_is_zero(thread_(pending_exception), 2224 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C", 0); 2225 2226 __ bind(locked); 2227 } 2228 2229 2230 // Publish thread state 2231 // -------------------------------------------------------------------------- 2232 2233 // Use that pc we placed in r_return_pc a while back as the current frame anchor. 2234 __ set_last_Java_frame(R1_SP, r_return_pc); 2235 2236 // Transition from _thread_in_Java to _thread_in_native. 2237 __ li(R0, _thread_in_native); 2238 __ release(); 2239 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2240 __ stw(R0, thread_(thread_state)); 2241 if (UseMembar) { 2242 __ fence(); 2243 } 2244 2245 2246 // The JNI call 2247 // -------------------------------------------------------------------------- 2248 2249 FunctionDescriptor* fd_native_method = (FunctionDescriptor*) native_func; 2250 __ call_c(fd_native_method, relocInfo::runtime_call_type); 2251 2252 2253 // Now, we are back from the native code. 2254 2255 2256 // Unpack the native result. 2257 // -------------------------------------------------------------------------- 2258 2259 // For int-types, we do any needed sign-extension required. 2260 // Care must be taken that the return values (R3_RET and F1_RET) 2261 // will survive any VM calls for blocking or unlocking. 2262 // An OOP result (handle) is done specially in the slow-path code. 2263 2264 switch (ret_type) { 2265 case T_VOID: break; // Nothing to do! 2266 case T_FLOAT: break; // Got it where we want it (unless slow-path). 2267 case T_DOUBLE: break; // Got it where we want it (unless slow-path). 2268 case T_LONG: break; // Got it where we want it (unless slow-path). 2269 case T_OBJECT: break; // Really a handle. 2270 // Cannot de-handlize until after reclaiming jvm_lock. 2271 case T_ARRAY: break; 2272 2273 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1) 2274 Label skip_modify; 2275 __ cmpwi(CCR0, R3_RET, 0); 2276 __ beq(CCR0, skip_modify); 2277 __ li(R3_RET, 1); 2278 __ bind(skip_modify); 2279 break; 2280 } 2281 case T_BYTE: { // sign extension 2282 __ extsb(R3_RET, R3_RET); 2283 break; 2284 } 2285 case T_CHAR: { // unsigned result 2286 __ andi(R3_RET, R3_RET, 0xffff); 2287 break; 2288 } 2289 case T_SHORT: { // sign extension 2290 __ extsh(R3_RET, R3_RET); 2291 break; 2292 } 2293 case T_INT: // nothing to do 2294 break; 2295 default: 2296 ShouldNotReachHere(); 2297 break; 2298 } 2299 2300 2301 // Publish thread state 2302 // -------------------------------------------------------------------------- 2303 2304 // Switch thread to "native transition" state before reading the 2305 // synchronization state. This additional state is necessary because reading 2306 // and testing the synchronization state is not atomic w.r.t. GC, as this 2307 // scenario demonstrates: 2308 // - Java thread A, in _thread_in_native state, loads _not_synchronized 2309 // and is preempted. 2310 // - VM thread changes sync state to synchronizing and suspends threads 2311 // for GC. 2312 // - Thread A is resumed to finish this native method, but doesn't block 2313 // here since it didn't see any synchronization in progress, and escapes. 2314 2315 // Transition from _thread_in_native to _thread_in_native_trans. 2316 __ li(R0, _thread_in_native_trans); 2317 __ release(); 2318 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2319 __ stw(R0, thread_(thread_state)); 2320 2321 2322 // Must we block? 2323 // -------------------------------------------------------------------------- 2324 2325 // Block, if necessary, before resuming in _thread_in_Java state. 2326 // In order for GC to work, don't clear the last_Java_sp until after blocking. 2327 Label after_transition; 2328 { 2329 Label no_block, sync; 2330 2331 if (os::is_MP()) { 2332 if (UseMembar) { 2333 // Force this write out before the read below. 2334 __ fence(); 2335 } else { 2336 // Write serialization page so VM thread can do a pseudo remote membar. 2337 // We use the current thread pointer to calculate a thread specific 2338 // offset to write to within the page. This minimizes bus traffic 2339 // due to cache line collision. 2340 __ serialize_memory(R16_thread, r_temp_4, r_temp_5); 2341 } 2342 } 2343 2344 Register sync_state_addr = r_temp_4; 2345 Register sync_state = r_temp_5; 2346 Register suspend_flags = r_temp_6; 2347 2348 __ load_const(sync_state_addr, SafepointSynchronize::address_of_state(), /*temp*/ sync_state); 2349 2350 // TODO: PPC port assert(4 == SafepointSynchronize::sz_state(), "unexpected field size"); 2351 __ lwz(sync_state, 0, sync_state_addr); 2352 2353 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size"); 2354 __ lwz(suspend_flags, thread_(suspend_flags)); 2355 2356 __ acquire(); 2357 2358 Label do_safepoint; 2359 // No synchronization in progress nor yet synchronized. 2360 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized); 2361 // Not suspended. 2362 __ cmpwi(CCR1, suspend_flags, 0); 2363 2364 __ bne(CCR0, sync); 2365 __ beq(CCR1, no_block); 2366 2367 // Block. Save any potential method result value before the operation and 2368 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this 2369 // lets us share the oopMap we used when we went native rather than create 2370 // a distinct one for this pc. 2371 __ bind(sync); 2372 2373 address entry_point = is_critical_native 2374 ? CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition) 2375 : CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans); 2376 save_native_result(masm, ret_type, workspace_slot_offset); 2377 __ call_VM_leaf(entry_point, R16_thread); 2378 restore_native_result(masm, ret_type, workspace_slot_offset); 2379 2380 if (is_critical_native) { 2381 __ b(after_transition); // No thread state transition here. 2382 } 2383 __ bind(no_block); 2384 } 2385 2386 // Publish thread state. 2387 // -------------------------------------------------------------------------- 2388 2389 // Thread state is thread_in_native_trans. Any safepoint blocking has 2390 // already happened so we can now change state to _thread_in_Java. 2391 2392 // Transition from _thread_in_native_trans to _thread_in_Java. 2393 __ li(R0, _thread_in_Java); 2394 __ release(); 2395 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2396 __ stw(R0, thread_(thread_state)); 2397 if (UseMembar) { 2398 __ fence(); 2399 } 2400 __ bind(after_transition); 2401 2402 // Reguard any pages if necessary. 2403 // -------------------------------------------------------------------------- 2404 2405 Label no_reguard; 2406 __ lwz(r_temp_1, thread_(stack_guard_state)); 2407 __ cmpwi(CCR0, r_temp_1, JavaThread::stack_guard_yellow_disabled); 2408 __ bne(CCR0, no_reguard); 2409 2410 save_native_result(masm, ret_type, workspace_slot_offset); 2411 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); 2412 restore_native_result(masm, ret_type, workspace_slot_offset); 2413 2414 __ bind(no_reguard); 2415 2416 2417 // Unlock 2418 // -------------------------------------------------------------------------- 2419 2420 if (method->is_synchronized()) { 2421 2422 ConditionRegister r_flag = CCR1; 2423 const Register r_oop = r_temp_4; 2424 const Register r_box = r_temp_5; 2425 const Register r_exception = r_temp_6; 2426 Label done; 2427 2428 // Get oop and address of lock object box. 2429 if (method_is_static) { 2430 assert(klass_offset != -1, ""); 2431 __ ld(r_oop, klass_offset, R1_SP); 2432 } else { 2433 assert(receiver_offset != -1, ""); 2434 __ ld(r_oop, receiver_offset, R1_SP); 2435 } 2436 __ addi(r_box, R1_SP, lock_offset); 2437 2438 // Try fastpath for unlocking. 2439 __ compiler_fast_unlock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3); 2440 __ beq(r_flag, done); 2441 2442 // Save and restore any potential method result value around the unlocking operation. 2443 save_native_result(masm, ret_type, workspace_slot_offset); 2444 2445 // Must save pending exception around the slow-path VM call. Since it's a 2446 // leaf call, the pending exception (if any) can be kept in a register. 2447 __ ld(r_exception, thread_(pending_exception)); 2448 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile"); 2449 __ li(R0, 0); 2450 __ std(R0, thread_(pending_exception)); 2451 2452 // Slow case of monitor enter. 2453 // Inline a special case of call_VM that disallows any pending_exception. 2454 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box); 2455 2456 __ asm_assert_mem8_is_zero(thread_(pending_exception), 2457 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C", 0); 2458 2459 restore_native_result(masm, ret_type, workspace_slot_offset); 2460 2461 // Check_forward_pending_exception jump to forward_exception if any pending 2462 // exception is set. The forward_exception routine expects to see the 2463 // exception in pending_exception and not in a register. Kind of clumsy, 2464 // since all folks who branch to forward_exception must have tested 2465 // pending_exception first and hence have it in a register already. 2466 __ std(r_exception, thread_(pending_exception)); 2467 2468 __ bind(done); 2469 } 2470 2471 # if 0 2472 // DTrace method exit 2473 # endif 2474 2475 // Clear "last Java frame" SP and PC. 2476 // -------------------------------------------------------------------------- 2477 2478 __ reset_last_Java_frame(); 2479 2480 // Unpack oop result. 2481 // -------------------------------------------------------------------------- 2482 2483 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2484 Label skip_unboxing; 2485 __ cmpdi(CCR0, R3_RET, 0); 2486 __ beq(CCR0, skip_unboxing); 2487 __ ld(R3_RET, 0, R3_RET); 2488 __ bind(skip_unboxing); 2489 __ verify_oop(R3_RET); 2490 } 2491 2492 2493 // Reset handle block. 2494 // -------------------------------------------------------------------------- 2495 if (!is_critical_native) { 2496 __ ld(r_temp_1, thread_(active_handles)); 2497 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size"); 2498 __ li(r_temp_2, 0); 2499 __ stw(r_temp_2, JNIHandleBlock::top_offset_in_bytes(), r_temp_1); 2500 2501 2502 // Check for pending exceptions. 2503 // -------------------------------------------------------------------------- 2504 __ ld(r_temp_2, thread_(pending_exception)); 2505 __ cmpdi(CCR0, r_temp_2, 0); 2506 __ bne(CCR0, handle_pending_exception); 2507 } 2508 2509 // Return 2510 // -------------------------------------------------------------------------- 2511 2512 __ pop_frame(); 2513 __ restore_LR_CR(R11); 2514 __ blr(); 2515 2516 2517 // Handler for pending exceptions (out-of-line). 2518 // -------------------------------------------------------------------------- 2519 2520 // Since this is a native call, we know the proper exception handler 2521 // is the empty function. We just pop this frame and then jump to 2522 // forward_exception_entry. 2523 if (!is_critical_native) { 2524 __ align(InteriorEntryAlignment); 2525 __ bind(handle_pending_exception); 2526 2527 __ pop_frame(); 2528 __ restore_LR_CR(R11); 2529 __ b64_patchable((address)StubRoutines::forward_exception_entry(), 2530 relocInfo::runtime_call_type); 2531 } 2532 2533 // Handler for a cache miss (out-of-line). 2534 // -------------------------------------------------------------------------- 2535 2536 if (!method_is_static) { 2537 __ align(InteriorEntryAlignment); 2538 __ bind(ic_miss); 2539 2540 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 2541 relocInfo::runtime_call_type); 2542 } 2543 2544 // Done. 2545 // -------------------------------------------------------------------------- 2546 2547 __ flush(); 2548 2549 nmethod *nm = nmethod::new_native_nmethod(method, 2550 compile_id, 2551 masm->code(), 2552 vep_start_pc-start_pc, 2553 frame_done_pc-start_pc, 2554 stack_slots / VMRegImpl::slots_per_word, 2555 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2556 in_ByteSize(lock_offset), 2557 oop_maps); 2558 2559 if (is_critical_native) { 2560 nm->set_lazy_critical_native(true); 2561 } 2562 2563 return nm; 2564 #else 2565 ShouldNotReachHere(); 2566 return NULL; 2567 #endif // COMPILER2 2568 } 2569 2570 // This function returns the adjust size (in number of words) to a c2i adapter 2571 // activation for use during deoptimization. 2572 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2573 return round_to((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::alignment_in_bytes); 2574 } 2575 2576 uint SharedRuntime::out_preserve_stack_slots() { 2577 #ifdef COMPILER2 2578 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size; 2579 #else 2580 return 0; 2581 #endif 2582 } 2583 2584 #ifdef COMPILER2 2585 // Frame generation for deopt and uncommon trap blobs. 2586 static void push_skeleton_frame(MacroAssembler* masm, bool deopt, 2587 /* Read */ 2588 Register unroll_block_reg, 2589 /* Update */ 2590 Register frame_sizes_reg, 2591 Register number_of_frames_reg, 2592 Register pcs_reg, 2593 /* Invalidate */ 2594 Register frame_size_reg, 2595 Register pc_reg) { 2596 2597 __ ld(pc_reg, 0, pcs_reg); 2598 __ ld(frame_size_reg, 0, frame_sizes_reg); 2599 __ std(pc_reg, _abi(lr), R1_SP); 2600 __ push_frame(frame_size_reg, R0/*tmp*/); 2601 __ std(R1_SP, _parent_ijava_frame_abi(initial_caller_sp), R1_SP); 2602 __ addi(number_of_frames_reg, number_of_frames_reg, -1); 2603 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize); 2604 __ addi(pcs_reg, pcs_reg, wordSize); 2605 } 2606 2607 // Loop through the UnrollBlock info and create new frames. 2608 static void push_skeleton_frames(MacroAssembler* masm, bool deopt, 2609 /* read */ 2610 Register unroll_block_reg, 2611 /* invalidate */ 2612 Register frame_sizes_reg, 2613 Register number_of_frames_reg, 2614 Register pcs_reg, 2615 Register frame_size_reg, 2616 Register pc_reg) { 2617 Label loop; 2618 2619 // _number_of_frames is of type int (deoptimization.hpp) 2620 __ lwa(number_of_frames_reg, 2621 Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(), 2622 unroll_block_reg); 2623 __ ld(pcs_reg, 2624 Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(), 2625 unroll_block_reg); 2626 __ ld(frame_sizes_reg, 2627 Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(), 2628 unroll_block_reg); 2629 2630 // stack: (caller_of_deoptee, ...). 2631 2632 // At this point we either have an interpreter frame or a compiled 2633 // frame on top of stack. If it is a compiled frame we push a new c2i 2634 // adapter here 2635 2636 // Memorize top-frame stack-pointer. 2637 __ mr(frame_size_reg/*old_sp*/, R1_SP); 2638 2639 // Resize interpreter top frame OR C2I adapter. 2640 2641 // At this moment, the top frame (which is the caller of the deoptee) is 2642 // an interpreter frame or a newly pushed C2I adapter or an entry frame. 2643 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the 2644 // outgoing arguments. 2645 // 2646 // In order to push the interpreter frame for the deoptee, we need to 2647 // resize the top frame such that we are able to place the deoptee's 2648 // locals in the frame. 2649 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI 2650 // into a valid PARENT_IJAVA_FRAME_ABI. 2651 2652 __ lwa(R11_scratch1, 2653 Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(), 2654 unroll_block_reg); 2655 __ neg(R11_scratch1, R11_scratch1); 2656 2657 // R11_scratch1 contains size of locals for frame resizing. 2658 // R12_scratch2 contains top frame's lr. 2659 2660 // Resize frame by complete frame size prevents TOC from being 2661 // overwritten by locals. A more stack space saving way would be 2662 // to copy the TOC to its location in the new abi. 2663 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size); 2664 2665 // now, resize the frame 2666 __ resize_frame(R11_scratch1, pc_reg/*tmp*/); 2667 2668 // In the case where we have resized a c2i frame above, the optional 2669 // alignment below the locals has size 32 (why?). 2670 __ std(R12_scratch2, _abi(lr), R1_SP); 2671 2672 // Initialize initial_caller_sp. 2673 __ std(frame_size_reg/*old_sp*/, _parent_ijava_frame_abi(initial_caller_sp), R1_SP); 2674 2675 #ifdef ASSERT 2676 // Make sure that there is at least one entry in the array. 2677 __ cmpdi(CCR0, number_of_frames_reg, 0); 2678 __ asm_assert_ne("array_size must be > 0", 0x205); 2679 #endif 2680 2681 // Now push the new interpreter frames. 2682 // 2683 __ bind(loop); 2684 // Allocate a new frame, fill in the pc. 2685 push_skeleton_frame(masm, deopt, 2686 unroll_block_reg, 2687 frame_sizes_reg, 2688 number_of_frames_reg, 2689 pcs_reg, 2690 frame_size_reg, 2691 pc_reg); 2692 __ cmpdi(CCR0, number_of_frames_reg, 0); 2693 __ bne(CCR0, loop); 2694 2695 // Get the return address pointing into the frame manager. 2696 __ ld(R0, 0, pcs_reg); 2697 // Store it in the top interpreter frame. 2698 __ std(R0, _abi(lr), R1_SP); 2699 // Initialize frame_manager_lr of interpreter top frame. 2700 __ std(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 2701 } 2702 #endif 2703 2704 void SharedRuntime::generate_deopt_blob() { 2705 // Allocate space for the code 2706 ResourceMark rm; 2707 // Setup code generation tools 2708 CodeBuffer buffer("deopt_blob", 2048, 1024); 2709 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer); 2710 Label exec_mode_initialized; 2711 int frame_size_in_words; 2712 OopMap* map = NULL; 2713 OopMapSet *oop_maps = new OopMapSet(); 2714 2715 // size of ABI112 plus spill slots for R3_RET and F1_RET. 2716 const int frame_size_in_bytes = frame::abi_112_spill_size; 2717 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint); 2718 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info. 2719 2720 const Register exec_mode_reg = R21_tmp1; 2721 2722 const address start = __ pc(); 2723 2724 #ifdef COMPILER2 2725 // -------------------------------------------------------------------------- 2726 // Prolog for non exception case! 2727 2728 // We have been called from the deopt handler of the deoptee. 2729 // 2730 // deoptee: 2731 // ... 2732 // call X 2733 // ... 2734 // deopt_handler: call_deopt_stub 2735 // cur. return pc --> ... 2736 // 2737 // So currently SR_LR points behind the call in the deopt handler. 2738 // We adjust it such that it points to the start of the deopt handler. 2739 // The return_pc has been stored in the frame of the deoptee and 2740 // will replace the address of the deopt_handler in the call 2741 // to Deoptimization::fetch_unroll_info below. 2742 // We can't grab a free register here, because all registers may 2743 // contain live values, so let the RegisterSaver do the adjustment 2744 // of the return pc. 2745 const int return_pc_adjustment_no_exception = -size_deopt_handler(); 2746 2747 // Push the "unpack frame" 2748 // Save everything in sight. 2749 map = RegisterSaver::push_frame_abi112_and_save_live_registers(masm, 2750 &first_frame_size_in_bytes, 2751 /*generate_oop_map=*/ true, 2752 return_pc_adjustment_no_exception, 2753 RegisterSaver::return_pc_is_lr); 2754 assert(map != NULL, "OopMap must have been created"); 2755 2756 __ li(exec_mode_reg, Deoptimization::Unpack_deopt); 2757 // Save exec mode for unpack_frames. 2758 __ b(exec_mode_initialized); 2759 2760 // -------------------------------------------------------------------------- 2761 // Prolog for exception case 2762 2763 // An exception is pending. 2764 // We have been called with a return (interpreter) or a jump (exception blob). 2765 // 2766 // - R3_ARG1: exception oop 2767 // - R4_ARG2: exception pc 2768 2769 int exception_offset = __ pc() - start; 2770 2771 BLOCK_COMMENT("Prolog for exception case"); 2772 2773 // The RegisterSaves doesn't need to adjust the return pc for this situation. 2774 const int return_pc_adjustment_exception = 0; 2775 2776 // Push the "unpack frame". 2777 // Save everything in sight. 2778 assert(R4 == R4_ARG2, "exception pc must be in r4"); 2779 RegisterSaver::push_frame_abi112_and_save_live_registers(masm, 2780 &first_frame_size_in_bytes, 2781 /*generate_oop_map=*/ false, 2782 return_pc_adjustment_exception, 2783 RegisterSaver::return_pc_is_r4); 2784 2785 // Deopt during an exception. Save exec mode for unpack_frames. 2786 __ li(exec_mode_reg, Deoptimization::Unpack_exception); 2787 2788 // Store exception oop and pc in thread (location known to GC). 2789 // This is needed since the call to "fetch_unroll_info()" may safepoint. 2790 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2791 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2792 2793 // fall through 2794 2795 // -------------------------------------------------------------------------- 2796 __ BIND(exec_mode_initialized); 2797 2798 { 2799 const Register unroll_block_reg = R22_tmp2; 2800 2801 // We need to set `last_Java_frame' because `fetch_unroll_info' will 2802 // call `last_Java_frame()'. The value of the pc in the frame is not 2803 // particularly important. It just needs to identify this blob. 2804 __ set_last_Java_frame(R1_SP, noreg); 2805 2806 // With EscapeAnalysis turned on, this call may safepoint! 2807 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread); 2808 address calls_return_pc = __ last_calls_return_pc(); 2809 // Set an oopmap for the call site that describes all our saved registers. 2810 oop_maps->add_gc_map(calls_return_pc - start, map); 2811 2812 __ reset_last_Java_frame(); 2813 // Save the return value. 2814 __ mr(unroll_block_reg, R3_RET); 2815 2816 // Restore only the result registers that have been saved 2817 // by save_volatile_registers(...). 2818 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes); 2819 2820 // In excp_deopt_mode, restore and clear exception oop which we 2821 // stored in the thread during exception entry above. The exception 2822 // oop will be the return value of this stub. 2823 Label skip_restore_excp; 2824 __ cmpdi(CCR0, exec_mode_reg, Deoptimization::Unpack_exception); 2825 __ bne(CCR0, skip_restore_excp); 2826 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2827 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2828 __ li(R0, 0); 2829 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2830 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2831 __ BIND(skip_restore_excp); 2832 2833 // reload narrro_oop_base 2834 if (UseCompressedOops && Universe::narrow_oop_base() != 0) { 2835 __ load_const_optimized(R30, Universe::narrow_oop_base()); 2836 } 2837 2838 __ pop_frame(); 2839 2840 // stack: (deoptee, optional i2c, caller of deoptee, ...). 2841 2842 // pop the deoptee's frame 2843 __ pop_frame(); 2844 2845 // stack: (caller_of_deoptee, ...). 2846 2847 // Loop through the `UnrollBlock' info and create interpreter frames. 2848 push_skeleton_frames(masm, true/*deopt*/, 2849 unroll_block_reg, 2850 R23_tmp3, 2851 R24_tmp4, 2852 R25_tmp5, 2853 R26_tmp6, 2854 R27_tmp7); 2855 2856 // stack: (skeletal interpreter frame, ..., optional skeletal 2857 // interpreter frame, optional c2i, caller of deoptee, ...). 2858 } 2859 2860 // push an `unpack_frame' taking care of float / int return values. 2861 __ push_frame(frame_size_in_bytes, R0/*tmp*/); 2862 2863 // stack: (unpack frame, skeletal interpreter frame, ..., optional 2864 // skeletal interpreter frame, optional c2i, caller of deoptee, 2865 // ...). 2866 2867 // Spill live volatile registers since we'll do a call. 2868 __ std( R3_RET, _abi_112_spill(spill_ret), R1_SP); 2869 __ stfd(F1_RET, _abi_112_spill(spill_fret), R1_SP); 2870 2871 // Let the unpacker layout information in the skeletal frames just 2872 // allocated. 2873 __ get_PC_trash_LR(R3_RET); 2874 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET); 2875 // This is a call to a LEAF method, so no oop map is required. 2876 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), 2877 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/); 2878 __ reset_last_Java_frame(); 2879 2880 // Restore the volatiles saved above. 2881 __ ld( R3_RET, _abi_112_spill(spill_ret), R1_SP); 2882 __ lfd(F1_RET, _abi_112_spill(spill_fret), R1_SP); 2883 2884 // Pop the unpack frame. 2885 __ pop_frame(); 2886 __ restore_LR_CR(R0); 2887 2888 // stack: (top interpreter frame, ..., optional interpreter frame, 2889 // optional c2i, caller of deoptee, ...). 2890 2891 // Initialize R14_state. 2892 __ ld(R14_state, 0, R1_SP); 2893 __ addi(R14_state, R14_state, 2894 -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 2895 // Also inititialize R15_prev_state. 2896 __ restore_prev_state(); 2897 2898 // Return to the interpreter entry point. 2899 __ blr(); 2900 __ flush(); 2901 #else // COMPILER2 2902 __ unimplemented("deopt blob needed only with compiler"); 2903 int exception_offset = __ pc() - start; 2904 #endif // COMPILER2 2905 2906 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, 0, first_frame_size_in_bytes / wordSize); 2907 } 2908 2909 #ifdef COMPILER2 2910 void SharedRuntime::generate_uncommon_trap_blob() { 2911 // Allocate space for the code. 2912 ResourceMark rm; 2913 // Setup code generation tools. 2914 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2915 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer); 2916 address start = __ pc(); 2917 2918 Register unroll_block_reg = R21_tmp1; 2919 Register klass_index_reg = R22_tmp2; 2920 Register unc_trap_reg = R23_tmp3; 2921 2922 OopMapSet* oop_maps = new OopMapSet(); 2923 int frame_size_in_bytes = frame::abi_112_size; 2924 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0); 2925 2926 // stack: (deoptee, optional i2c, caller_of_deoptee, ...). 2927 2928 // Push a dummy `unpack_frame' and call 2929 // `Deoptimization::uncommon_trap' to pack the compiled frame into a 2930 // vframe array and return the `UnrollBlock' information. 2931 2932 // Save LR to compiled frame. 2933 __ save_LR_CR(R11_scratch1); 2934 2935 // Push an "uncommon_trap" frame. 2936 __ push_frame_abi112(0, R11_scratch1); 2937 2938 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...). 2939 2940 // Set the `unpack_frame' as last_Java_frame. 2941 // `Deoptimization::uncommon_trap' expects it and considers its 2942 // sender frame as the deoptee frame. 2943 // Remember the offset of the instruction whose address will be 2944 // moved to R11_scratch1. 2945 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1); 2946 2947 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 2948 2949 __ mr(klass_index_reg, R3); 2950 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), 2951 R16_thread, klass_index_reg); 2952 2953 // Set an oopmap for the call site. 2954 oop_maps->add_gc_map(gc_map_pc - start, map); 2955 2956 __ reset_last_Java_frame(); 2957 2958 // Pop the `unpack frame'. 2959 __ pop_frame(); 2960 2961 // stack: (deoptee, optional i2c, caller_of_deoptee, ...). 2962 2963 // Save the return value. 2964 __ mr(unroll_block_reg, R3_RET); 2965 2966 // Pop the uncommon_trap frame. 2967 __ pop_frame(); 2968 2969 // stack: (caller_of_deoptee, ...). 2970 2971 // Allocate new interpreter frame(s) and possibly a c2i adapter 2972 // frame. 2973 push_skeleton_frames(masm, false/*deopt*/, 2974 unroll_block_reg, 2975 R22_tmp2, 2976 R23_tmp3, 2977 R24_tmp4, 2978 R25_tmp5, 2979 R26_tmp6); 2980 2981 // stack: (skeletal interpreter frame, ..., optional skeletal 2982 // interpreter frame, optional c2i, caller of deoptee, ...). 2983 2984 // Push a dummy `unpack_frame' taking care of float return values. 2985 // Call `Deoptimization::unpack_frames' to layout information in the 2986 // interpreter frames just created. 2987 2988 // Push a simple "unpack frame" here. 2989 __ push_frame_abi112(0, R11_scratch1); 2990 2991 // stack: (unpack frame, skeletal interpreter frame, ..., optional 2992 // skeletal interpreter frame, optional c2i, caller of deoptee, 2993 // ...). 2994 2995 // Set the "unpack_frame" as last_Java_frame. 2996 __ get_PC_trash_LR(R11_scratch1); 2997 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 2998 2999 // Indicate it is the uncommon trap case. 3000 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap); 3001 // Let the unpacker layout information in the skeletal frames just 3002 // allocated. 3003 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), 3004 R16_thread, unc_trap_reg); 3005 3006 __ reset_last_Java_frame(); 3007 // Pop the `unpack frame'. 3008 __ pop_frame(); 3009 // Restore LR from top interpreter frame. 3010 __ restore_LR_CR(R11_scratch1); 3011 3012 // stack: (top interpreter frame, ..., optional interpreter frame, 3013 // optional c2i, caller of deoptee, ...). 3014 3015 // Initialize R14_state, ... 3016 __ ld(R11_scratch1, 0, R1_SP); 3017 __ addi(R14_state, R11_scratch1, 3018 -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 3019 // also initialize R15_prev_state. 3020 __ restore_prev_state(); 3021 // Return to the interpreter entry point. 3022 __ blr(); 3023 3024 masm->flush(); 3025 3026 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize); 3027 } 3028 #endif // COMPILER2 3029 3030 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap. 3031 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 3032 assert(StubRoutines::forward_exception_entry() != NULL, 3033 "must be generated before"); 3034 3035 ResourceMark rm; 3036 OopMapSet *oop_maps = new OopMapSet(); 3037 OopMap* map; 3038 3039 // Allocate space for the code. Setup code generation tools. 3040 CodeBuffer buffer("handler_blob", 2048, 1024); 3041 MacroAssembler* masm = new MacroAssembler(&buffer); 3042 3043 address start = __ pc(); 3044 int frame_size_in_bytes = 0; 3045 3046 RegisterSaver::ReturnPCLocation return_pc_location; 3047 bool cause_return = (poll_type == POLL_AT_RETURN); 3048 if (cause_return) { 3049 // Nothing to do here. The frame has already been popped in MachEpilogNode. 3050 // Register LR already contains the return pc. 3051 return_pc_location = RegisterSaver::return_pc_is_lr; 3052 } else { 3053 // Use thread()->saved_exception_pc() as return pc. 3054 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc; 3055 } 3056 3057 // Save registers, fpu state, and flags. 3058 map = RegisterSaver::push_frame_abi112_and_save_live_registers(masm, 3059 &frame_size_in_bytes, 3060 /*generate_oop_map=*/ true, 3061 /*return_pc_adjustment=*/0, 3062 return_pc_location); 3063 3064 // The following is basically a call_VM. However, we need the precise 3065 // address of the call in order to generate an oopmap. Hence, we do all the 3066 // work outselves. 3067 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg); 3068 3069 // The return address must always be correct so that the frame constructor 3070 // never sees an invalid pc. 3071 3072 // Do the call 3073 __ call_VM_leaf(call_ptr, R16_thread); 3074 address calls_return_pc = __ last_calls_return_pc(); 3075 3076 // Set an oopmap for the call site. This oopmap will map all 3077 // oop-registers and debug-info registers as callee-saved. This 3078 // will allow deoptimization at this safepoint to find all possible 3079 // debug-info recordings, as well as let GC find all oops. 3080 oop_maps->add_gc_map(calls_return_pc - start, map); 3081 3082 Label noException; 3083 3084 // Clear the last Java frame. 3085 __ reset_last_Java_frame(); 3086 3087 BLOCK_COMMENT(" Check pending exception."); 3088 const Register pending_exception = R0; 3089 __ ld(pending_exception, thread_(pending_exception)); 3090 __ cmpdi(CCR0, pending_exception, 0); 3091 __ beq(CCR0, noException); 3092 3093 // Exception pending 3094 RegisterSaver::restore_live_registers_and_pop_frame(masm, 3095 frame_size_in_bytes, 3096 /*restore_ctr=*/true); 3097 3098 3099 BLOCK_COMMENT(" Jump to forward_exception_entry."); 3100 // Jump to forward_exception_entry, with the issuing PC in LR 3101 // so it looks like the original nmethod called forward_exception_entry. 3102 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 3103 3104 // No exception case. 3105 __ BIND(noException); 3106 3107 3108 // Normal exit, restore registers and exit. 3109 RegisterSaver::restore_live_registers_and_pop_frame(masm, 3110 frame_size_in_bytes, 3111 /*restore_ctr=*/true); 3112 3113 __ blr(); 3114 3115 // Make sure all code is generated 3116 masm->flush(); 3117 3118 // Fill-out other meta info 3119 // CodeBlob frame size is in words. 3120 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize); 3121 } 3122 3123 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss) 3124 // 3125 // Generate a stub that calls into the vm to find out the proper destination 3126 // of a java call. All the argument registers are live at this point 3127 // but since this is generic code we don't know what they are and the caller 3128 // must do any gc of the args. 3129 // 3130 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3131 3132 // allocate space for the code 3133 ResourceMark rm; 3134 3135 CodeBuffer buffer(name, 1000, 512); 3136 MacroAssembler* masm = new MacroAssembler(&buffer); 3137 3138 int frame_size_in_bytes; 3139 3140 OopMapSet *oop_maps = new OopMapSet(); 3141 OopMap* map = NULL; 3142 3143 address start = __ pc(); 3144 3145 map = RegisterSaver::push_frame_abi112_and_save_live_registers(masm, 3146 &frame_size_in_bytes, 3147 /*generate_oop_map*/ true, 3148 /*return_pc_adjustment*/ 0, 3149 RegisterSaver::return_pc_is_lr); 3150 3151 // Use noreg as last_Java_pc, the return pc will be reconstructed 3152 // from the physical frame. 3153 __ set_last_Java_frame(/*sp*/R1_SP, noreg); 3154 3155 int frame_complete = __ offset(); 3156 3157 // Pass R19_method as 2nd (optional) argument, used by 3158 // counter_overflow_stub. 3159 __ call_VM_leaf(destination, R16_thread, R19_method); 3160 address calls_return_pc = __ last_calls_return_pc(); 3161 // Set an oopmap for the call site. 3162 // We need this not only for callee-saved registers, but also for volatile 3163 // registers that the compiler might be keeping live across a safepoint. 3164 // Create the oopmap for the call's return pc. 3165 oop_maps->add_gc_map(calls_return_pc - start, map); 3166 3167 // R3_RET contains the address we are going to jump to assuming no exception got installed. 3168 3169 // clear last_Java_sp 3170 __ reset_last_Java_frame(); 3171 3172 // Check for pending exceptions. 3173 BLOCK_COMMENT("Check for pending exceptions."); 3174 Label pending; 3175 __ ld(R11_scratch1, thread_(pending_exception)); 3176 __ cmpdi(CCR0, R11_scratch1, 0); 3177 __ bne(CCR0, pending); 3178 3179 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame. 3180 3181 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false); 3182 3183 // Get the returned methodOop. 3184 __ get_vm_result_2(R19_method); 3185 3186 __ bctr(); 3187 3188 3189 // Pending exception after the safepoint. 3190 __ BIND(pending); 3191 3192 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true); 3193 3194 // exception pending => remove activation and forward to exception handler 3195 3196 __ li(R11_scratch1, 0); 3197 __ ld(R3_ARG1, thread_(pending_exception)); 3198 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_offset()), R16_thread); 3199 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 3200 3201 // ------------- 3202 // Make sure all code is generated. 3203 masm->flush(); 3204 3205 // return the blob 3206 // frame_size_words or bytes?? 3207 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize, 3208 oop_maps, true); 3209 }