1 /* 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 3 * Copyright 2012, 2014 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 "interpreter/interp_masm.hpp" 33 #include "oops/compiledICHolder.hpp" 34 #include "prims/jvmtiRedefineClassesTrace.hpp" 35 #include "runtime/sharedRuntime.hpp" 36 #include "runtime/vframeArray.hpp" 37 #include "vmreg_ppc.inline.hpp" 38 #ifdef COMPILER1 39 #include "c1/c1_Runtime1.hpp" 40 #endif 41 #ifdef COMPILER2 42 #include "adfiles/ad_ppc_64.hpp" 43 #include "opto/runtime.hpp" 44 #endif 45 46 #define __ masm-> 47 48 #ifdef PRODUCT 49 #define BLOCK_COMMENT(str) // nothing 50 #else 51 #define BLOCK_COMMENT(str) __ block_comment(str) 52 #endif 53 54 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 55 56 57 class RegisterSaver { 58 // Used for saving volatile registers. 59 public: 60 61 // Support different return pc locations. 62 enum ReturnPCLocation { 63 return_pc_is_lr, 64 return_pc_is_r4, 65 return_pc_is_thread_saved_exception_pc 66 }; 67 68 static OopMap* push_frame_reg_args_and_save_live_registers(MacroAssembler* masm, 69 int* out_frame_size_in_bytes, 70 bool generate_oop_map, 71 int return_pc_adjustment, 72 ReturnPCLocation return_pc_location); 73 static void restore_live_registers_and_pop_frame(MacroAssembler* masm, 74 int frame_size_in_bytes, 75 bool restore_ctr); 76 77 static void push_frame_and_save_argument_registers(MacroAssembler* masm, 78 Register r_temp, 79 int frame_size, 80 int total_args, 81 const VMRegPair *regs, const VMRegPair *regs2 = NULL); 82 static void restore_argument_registers_and_pop_frame(MacroAssembler*masm, 83 int frame_size, 84 int total_args, 85 const VMRegPair *regs, const VMRegPair *regs2 = NULL); 86 87 // During deoptimization only the result registers need to be restored 88 // all the other values have already been extracted. 89 static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes); 90 91 // Constants and data structures: 92 93 typedef enum { 94 int_reg = 0, 95 float_reg = 1, 96 special_reg = 2 97 } RegisterType; 98 99 typedef enum { 100 reg_size = 8, 101 half_reg_size = reg_size / 2, 102 } RegisterConstants; 103 104 typedef struct { 105 RegisterType reg_type; 106 int reg_num; 107 VMReg vmreg; 108 } LiveRegType; 109 }; 110 111 112 #define RegisterSaver_LiveSpecialReg(regname) \ 113 { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() } 114 115 #define RegisterSaver_LiveIntReg(regname) \ 116 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() } 117 118 #define RegisterSaver_LiveFloatReg(regname) \ 119 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() } 120 121 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = { 122 // Live registers which get spilled to the stack. Register 123 // positions in this array correspond directly to the stack layout. 124 125 // 126 // live special registers: 127 // 128 RegisterSaver_LiveSpecialReg(SR_CTR), 129 // 130 // live float registers: 131 // 132 RegisterSaver_LiveFloatReg( F0 ), 133 RegisterSaver_LiveFloatReg( F1 ), 134 RegisterSaver_LiveFloatReg( F2 ), 135 RegisterSaver_LiveFloatReg( F3 ), 136 RegisterSaver_LiveFloatReg( F4 ), 137 RegisterSaver_LiveFloatReg( F5 ), 138 RegisterSaver_LiveFloatReg( F6 ), 139 RegisterSaver_LiveFloatReg( F7 ), 140 RegisterSaver_LiveFloatReg( F8 ), 141 RegisterSaver_LiveFloatReg( F9 ), 142 RegisterSaver_LiveFloatReg( F10 ), 143 RegisterSaver_LiveFloatReg( F11 ), 144 RegisterSaver_LiveFloatReg( F12 ), 145 RegisterSaver_LiveFloatReg( F13 ), 146 RegisterSaver_LiveFloatReg( F14 ), 147 RegisterSaver_LiveFloatReg( F15 ), 148 RegisterSaver_LiveFloatReg( F16 ), 149 RegisterSaver_LiveFloatReg( F17 ), 150 RegisterSaver_LiveFloatReg( F18 ), 151 RegisterSaver_LiveFloatReg( F19 ), 152 RegisterSaver_LiveFloatReg( F20 ), 153 RegisterSaver_LiveFloatReg( F21 ), 154 RegisterSaver_LiveFloatReg( F22 ), 155 RegisterSaver_LiveFloatReg( F23 ), 156 RegisterSaver_LiveFloatReg( F24 ), 157 RegisterSaver_LiveFloatReg( F25 ), 158 RegisterSaver_LiveFloatReg( F26 ), 159 RegisterSaver_LiveFloatReg( F27 ), 160 RegisterSaver_LiveFloatReg( F28 ), 161 RegisterSaver_LiveFloatReg( F29 ), 162 RegisterSaver_LiveFloatReg( F30 ), 163 RegisterSaver_LiveFloatReg( F31 ), 164 // 165 // live integer registers: 166 // 167 RegisterSaver_LiveIntReg( R0 ), 168 //RegisterSaver_LiveIntReg( R1 ), // stack pointer 169 RegisterSaver_LiveIntReg( R2 ), 170 RegisterSaver_LiveIntReg( R3 ), 171 RegisterSaver_LiveIntReg( R4 ), 172 RegisterSaver_LiveIntReg( R5 ), 173 RegisterSaver_LiveIntReg( R6 ), 174 RegisterSaver_LiveIntReg( R7 ), 175 RegisterSaver_LiveIntReg( R8 ), 176 RegisterSaver_LiveIntReg( R9 ), 177 RegisterSaver_LiveIntReg( R10 ), 178 RegisterSaver_LiveIntReg( R11 ), 179 RegisterSaver_LiveIntReg( R12 ), 180 //RegisterSaver_LiveIntReg( R13 ), // system thread id 181 RegisterSaver_LiveIntReg( R14 ), 182 RegisterSaver_LiveIntReg( R15 ), 183 RegisterSaver_LiveIntReg( R16 ), 184 RegisterSaver_LiveIntReg( R17 ), 185 RegisterSaver_LiveIntReg( R18 ), 186 RegisterSaver_LiveIntReg( R19 ), 187 RegisterSaver_LiveIntReg( R20 ), 188 RegisterSaver_LiveIntReg( R21 ), 189 RegisterSaver_LiveIntReg( R22 ), 190 RegisterSaver_LiveIntReg( R23 ), 191 RegisterSaver_LiveIntReg( R24 ), 192 RegisterSaver_LiveIntReg( R25 ), 193 RegisterSaver_LiveIntReg( R26 ), 194 RegisterSaver_LiveIntReg( R27 ), 195 RegisterSaver_LiveIntReg( R28 ), 196 RegisterSaver_LiveIntReg( R29 ), 197 RegisterSaver_LiveIntReg( R31 ), 198 RegisterSaver_LiveIntReg( R30 ), // r30 must be the last register 199 }; 200 201 OopMap* RegisterSaver::push_frame_reg_args_and_save_live_registers(MacroAssembler* masm, 202 int* out_frame_size_in_bytes, 203 bool generate_oop_map, 204 int return_pc_adjustment, 205 ReturnPCLocation return_pc_location) { 206 // Push an abi_reg_args-frame and store all registers which may be live. 207 // If requested, create an OopMap: Record volatile registers as 208 // callee-save values in an OopMap so their save locations will be 209 // propagated to the RegisterMap of the caller frame during 210 // StackFrameStream construction (needed for deoptimization; see 211 // compiledVFrame::create_stack_value). 212 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment. 213 214 int i; 215 int offset; 216 217 // calcualte frame size 218 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 219 sizeof(RegisterSaver::LiveRegType); 220 const int register_save_size = regstosave_num * reg_size; 221 const int frame_size_in_bytes = round_to(register_save_size, frame::alignment_in_bytes) 222 + frame::abi_reg_args_size; 223 *out_frame_size_in_bytes = frame_size_in_bytes; 224 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint); 225 const int register_save_offset = frame_size_in_bytes - register_save_size; 226 227 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words. 228 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : NULL; 229 230 BLOCK_COMMENT("push_frame_reg_args_and_save_live_registers {"); 231 232 // Save r30 in the last slot of the not yet pushed frame so that we 233 // can use it as scratch reg. 234 __ std(R30, -reg_size, R1_SP); 235 assert(-reg_size == register_save_offset - frame_size_in_bytes + ((regstosave_num-1)*reg_size), 236 "consistency check"); 237 238 // save the flags 239 // Do the save_LR_CR by hand and adjust the return pc if requested. 240 __ mfcr(R30); 241 __ std(R30, _abi(cr), R1_SP); 242 switch (return_pc_location) { 243 case return_pc_is_lr: __ mflr(R30); break; 244 case return_pc_is_r4: __ mr(R30, R4); break; 245 case return_pc_is_thread_saved_exception_pc: 246 __ ld(R30, thread_(saved_exception_pc)); break; 247 default: ShouldNotReachHere(); 248 } 249 if (return_pc_adjustment != 0) 250 __ addi(R30, R30, return_pc_adjustment); 251 __ std(R30, _abi(lr), R1_SP); 252 253 // push a new frame 254 __ push_frame(frame_size_in_bytes, R30); 255 256 // save all registers (ints and floats) 257 offset = register_save_offset; 258 for (int i = 0; i < regstosave_num; i++) { 259 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 260 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 261 262 switch (reg_type) { 263 case RegisterSaver::int_reg: { 264 if (reg_num != 30) { // We spilled R30 right at the beginning. 265 __ std(as_Register(reg_num), offset, R1_SP); 266 } 267 break; 268 } 269 case RegisterSaver::float_reg: { 270 __ stfd(as_FloatRegister(reg_num), offset, R1_SP); 271 break; 272 } 273 case RegisterSaver::special_reg: { 274 if (reg_num == SR_CTR_SpecialRegisterEnumValue) { 275 __ mfctr(R30); 276 __ std(R30, offset, R1_SP); 277 } else { 278 Unimplemented(); 279 } 280 break; 281 } 282 default: 283 ShouldNotReachHere(); 284 } 285 286 if (generate_oop_map) { 287 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), 288 RegisterSaver_LiveRegs[i].vmreg); 289 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size)>>2), 290 RegisterSaver_LiveRegs[i].vmreg->next()); 291 } 292 offset += reg_size; 293 } 294 295 BLOCK_COMMENT("} push_frame_reg_args_and_save_live_registers"); 296 297 // And we're done. 298 return map; 299 } 300 301 302 // Pop the current frame and restore all the registers that we 303 // saved. 304 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm, 305 int frame_size_in_bytes, 306 bool restore_ctr) { 307 int i; 308 int offset; 309 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 310 sizeof(RegisterSaver::LiveRegType); 311 const int register_save_size = regstosave_num * reg_size; 312 const int register_save_offset = frame_size_in_bytes - register_save_size; 313 314 BLOCK_COMMENT("restore_live_registers_and_pop_frame {"); 315 316 // restore all registers (ints and floats) 317 offset = register_save_offset; 318 for (int i = 0; i < regstosave_num; i++) { 319 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 320 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 321 322 switch (reg_type) { 323 case RegisterSaver::int_reg: { 324 if (reg_num != 30) // R30 restored at the end, it's the tmp reg! 325 __ ld(as_Register(reg_num), offset, R1_SP); 326 break; 327 } 328 case RegisterSaver::float_reg: { 329 __ lfd(as_FloatRegister(reg_num), offset, R1_SP); 330 break; 331 } 332 case RegisterSaver::special_reg: { 333 if (reg_num == SR_CTR_SpecialRegisterEnumValue) { 334 if (restore_ctr) { // Nothing to do here if ctr already contains the next address. 335 __ ld(R30, offset, R1_SP); 336 __ mtctr(R30); 337 } 338 } else { 339 Unimplemented(); 340 } 341 break; 342 } 343 default: 344 ShouldNotReachHere(); 345 } 346 offset += reg_size; 347 } 348 349 // pop the frame 350 __ pop_frame(); 351 352 // restore the flags 353 __ restore_LR_CR(R30); 354 355 // restore scratch register's value 356 __ ld(R30, -reg_size, R1_SP); 357 358 BLOCK_COMMENT("} restore_live_registers_and_pop_frame"); 359 } 360 361 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp, 362 int frame_size,int total_args, const VMRegPair *regs, 363 const VMRegPair *regs2) { 364 __ push_frame(frame_size, r_temp); 365 int st_off = frame_size - wordSize; 366 for (int i = 0; i < total_args; i++) { 367 VMReg r_1 = regs[i].first(); 368 VMReg r_2 = regs[i].second(); 369 if (!r_1->is_valid()) { 370 assert(!r_2->is_valid(), ""); 371 continue; 372 } 373 if (r_1->is_Register()) { 374 Register r = r_1->as_Register(); 375 __ std(r, st_off, R1_SP); 376 st_off -= wordSize; 377 } else if (r_1->is_FloatRegister()) { 378 FloatRegister f = r_1->as_FloatRegister(); 379 __ stfd(f, st_off, R1_SP); 380 st_off -= wordSize; 381 } 382 } 383 if (regs2 != NULL) { 384 for (int i = 0; i < total_args; i++) { 385 VMReg r_1 = regs2[i].first(); 386 VMReg r_2 = regs2[i].second(); 387 if (!r_1->is_valid()) { 388 assert(!r_2->is_valid(), ""); 389 continue; 390 } 391 if (r_1->is_Register()) { 392 Register r = r_1->as_Register(); 393 __ std(r, st_off, R1_SP); 394 st_off -= wordSize; 395 } else if (r_1->is_FloatRegister()) { 396 FloatRegister f = r_1->as_FloatRegister(); 397 __ stfd(f, st_off, R1_SP); 398 st_off -= wordSize; 399 } 400 } 401 } 402 } 403 404 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size, 405 int total_args, const VMRegPair *regs, 406 const VMRegPair *regs2) { 407 int st_off = frame_size - wordSize; 408 for (int i = 0; i < total_args; i++) { 409 VMReg r_1 = regs[i].first(); 410 VMReg r_2 = regs[i].second(); 411 if (r_1->is_Register()) { 412 Register r = r_1->as_Register(); 413 __ ld(r, st_off, R1_SP); 414 st_off -= wordSize; 415 } else if (r_1->is_FloatRegister()) { 416 FloatRegister f = r_1->as_FloatRegister(); 417 __ lfd(f, st_off, R1_SP); 418 st_off -= wordSize; 419 } 420 } 421 if (regs2 != NULL) 422 for (int i = 0; i < total_args; i++) { 423 VMReg r_1 = regs2[i].first(); 424 VMReg r_2 = regs2[i].second(); 425 if (r_1->is_Register()) { 426 Register r = r_1->as_Register(); 427 __ ld(r, st_off, R1_SP); 428 st_off -= wordSize; 429 } else if (r_1->is_FloatRegister()) { 430 FloatRegister f = r_1->as_FloatRegister(); 431 __ lfd(f, st_off, R1_SP); 432 st_off -= wordSize; 433 } 434 } 435 __ pop_frame(); 436 } 437 438 // Restore the registers that might be holding a result. 439 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) { 440 int i; 441 int offset; 442 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 443 sizeof(RegisterSaver::LiveRegType); 444 const int register_save_size = regstosave_num * reg_size; 445 const int register_save_offset = frame_size_in_bytes - register_save_size; 446 447 // restore all result registers (ints and floats) 448 offset = register_save_offset; 449 for (int i = 0; i < regstosave_num; i++) { 450 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 451 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 452 switch (reg_type) { 453 case RegisterSaver::int_reg: { 454 if (as_Register(reg_num)==R3_RET) // int result_reg 455 __ ld(as_Register(reg_num), offset, R1_SP); 456 break; 457 } 458 case RegisterSaver::float_reg: { 459 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg 460 __ lfd(as_FloatRegister(reg_num), offset, R1_SP); 461 break; 462 } 463 case RegisterSaver::special_reg: { 464 // Special registers don't hold a result. 465 break; 466 } 467 default: 468 ShouldNotReachHere(); 469 } 470 offset += reg_size; 471 } 472 } 473 474 // Is vector's size (in bytes) bigger than a size saved by default? 475 bool SharedRuntime::is_wide_vector(int size) { 476 ResourceMark rm; 477 // Note, MaxVectorSize == 8 on PPC64. 478 assert(size <= 8, err_msg_res("%d bytes vectors are not supported", size)); 479 return size > 8; 480 } 481 #ifdef COMPILER2 482 static int reg2slot(VMReg r) { 483 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 484 } 485 486 static int reg2offset(VMReg r) { 487 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 488 } 489 #endif 490 491 // --------------------------------------------------------------------------- 492 // Read the array of BasicTypes from a signature, and compute where the 493 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 494 // quantities. Values less than VMRegImpl::stack0 are registers, those above 495 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 496 // as framesizes are fixed. 497 // VMRegImpl::stack0 refers to the first slot 0(sp). 498 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register 499 // up to RegisterImpl::number_of_registers) are the 64-bit 500 // integer registers. 501 502 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 503 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 504 // units regardless of build. Of course for i486 there is no 64 bit build 505 506 // The Java calling convention is a "shifted" version of the C ABI. 507 // By skipping the first C ABI register we can call non-static jni methods 508 // with small numbers of arguments without having to shuffle the arguments 509 // at all. Since we control the java ABI we ought to at least get some 510 // advantage out of it. 511 512 const VMReg java_iarg_reg[8] = { 513 R3->as_VMReg(), 514 R4->as_VMReg(), 515 R5->as_VMReg(), 516 R6->as_VMReg(), 517 R7->as_VMReg(), 518 R8->as_VMReg(), 519 R9->as_VMReg(), 520 R10->as_VMReg() 521 }; 522 523 const VMReg java_farg_reg[13] = { 524 F1->as_VMReg(), 525 F2->as_VMReg(), 526 F3->as_VMReg(), 527 F4->as_VMReg(), 528 F5->as_VMReg(), 529 F6->as_VMReg(), 530 F7->as_VMReg(), 531 F8->as_VMReg(), 532 F9->as_VMReg(), 533 F10->as_VMReg(), 534 F11->as_VMReg(), 535 F12->as_VMReg(), 536 F13->as_VMReg() 537 }; 538 539 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]); 540 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]); 541 542 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 543 VMRegPair *regs, 544 int total_args_passed, 545 int is_outgoing) { 546 // C2c calling conventions for compiled-compiled calls. 547 // Put 8 ints/longs into registers _AND_ 13 float/doubles into 548 // registers _AND_ put the rest on the stack. 549 550 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats 551 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles 552 553 int i; 554 VMReg reg; 555 int stk = 0; 556 int ireg = 0; 557 int freg = 0; 558 559 // We put the first 8 arguments into registers and the rest on the 560 // stack, float arguments are already in their argument registers 561 // due to c2c calling conventions (see calling_convention). 562 for (int i = 0; i < total_args_passed; ++i) { 563 switch(sig_bt[i]) { 564 case T_BOOLEAN: 565 case T_CHAR: 566 case T_BYTE: 567 case T_SHORT: 568 case T_INT: 569 if (ireg < num_java_iarg_registers) { 570 // Put int/ptr in register 571 reg = java_iarg_reg[ireg]; 572 ++ireg; 573 } else { 574 // Put int/ptr on stack. 575 reg = VMRegImpl::stack2reg(stk); 576 stk += inc_stk_for_intfloat; 577 } 578 regs[i].set1(reg); 579 break; 580 case T_LONG: 581 assert(sig_bt[i+1] == T_VOID, "expecting half"); 582 if (ireg < num_java_iarg_registers) { 583 // Put long in register. 584 reg = java_iarg_reg[ireg]; 585 ++ireg; 586 } else { 587 // Put long on stack. They must be aligned to 2 slots. 588 if (stk & 0x1) ++stk; 589 reg = VMRegImpl::stack2reg(stk); 590 stk += inc_stk_for_longdouble; 591 } 592 regs[i].set2(reg); 593 break; 594 case T_OBJECT: 595 case T_ARRAY: 596 case T_ADDRESS: 597 if (ireg < num_java_iarg_registers) { 598 // Put ptr in register. 599 reg = java_iarg_reg[ireg]; 600 ++ireg; 601 } else { 602 // Put ptr on stack. Objects must be aligned to 2 slots too, 603 // because "64-bit pointers record oop-ishness on 2 aligned 604 // adjacent registers." (see OopFlow::build_oop_map). 605 if (stk & 0x1) ++stk; 606 reg = VMRegImpl::stack2reg(stk); 607 stk += inc_stk_for_longdouble; 608 } 609 regs[i].set2(reg); 610 break; 611 case T_FLOAT: 612 if (freg < num_java_farg_registers) { 613 // Put float in register. 614 reg = java_farg_reg[freg]; 615 ++freg; 616 } else { 617 // Put float on stack. 618 reg = VMRegImpl::stack2reg(stk); 619 stk += inc_stk_for_intfloat; 620 } 621 regs[i].set1(reg); 622 break; 623 case T_DOUBLE: 624 assert(sig_bt[i+1] == T_VOID, "expecting half"); 625 if (freg < num_java_farg_registers) { 626 // Put double in register. 627 reg = java_farg_reg[freg]; 628 ++freg; 629 } else { 630 // Put double on stack. They must be aligned to 2 slots. 631 if (stk & 0x1) ++stk; 632 reg = VMRegImpl::stack2reg(stk); 633 stk += inc_stk_for_longdouble; 634 } 635 regs[i].set2(reg); 636 break; 637 case T_VOID: 638 // Do not count halves. 639 regs[i].set_bad(); 640 break; 641 default: 642 ShouldNotReachHere(); 643 } 644 } 645 return round_to(stk, 2); 646 } 647 648 #ifdef COMPILER2 649 // Calling convention for calling C code. 650 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 651 VMRegPair *regs, 652 VMRegPair *regs2, 653 int total_args_passed) { 654 // Calling conventions for C runtime calls and calls to JNI native methods. 655 // 656 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8 657 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist 658 // the first 13 flt/dbl's in the first 13 fp regs but additionally 659 // copy flt/dbl to the stack if they are beyond the 8th argument. 660 661 const VMReg iarg_reg[8] = { 662 R3->as_VMReg(), 663 R4->as_VMReg(), 664 R5->as_VMReg(), 665 R6->as_VMReg(), 666 R7->as_VMReg(), 667 R8->as_VMReg(), 668 R9->as_VMReg(), 669 R10->as_VMReg() 670 }; 671 672 const VMReg farg_reg[13] = { 673 F1->as_VMReg(), 674 F2->as_VMReg(), 675 F3->as_VMReg(), 676 F4->as_VMReg(), 677 F5->as_VMReg(), 678 F6->as_VMReg(), 679 F7->as_VMReg(), 680 F8->as_VMReg(), 681 F9->as_VMReg(), 682 F10->as_VMReg(), 683 F11->as_VMReg(), 684 F12->as_VMReg(), 685 F13->as_VMReg() 686 }; 687 688 // Check calling conventions consistency. 689 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c && 690 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c, 691 "consistency"); 692 693 // `Stk' counts stack slots. Due to alignment, 32 bit values occupy 694 // 2 such slots, like 64 bit values do. 695 const int inc_stk_for_intfloat = 2; // 2 slots for ints and floats 696 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles 697 698 int i; 699 VMReg reg; 700 // Leave room for C-compatible ABI_REG_ARGS. 701 int stk = (frame::abi_reg_args_size - frame::jit_out_preserve_size) / VMRegImpl::stack_slot_size; 702 int arg = 0; 703 int freg = 0; 704 705 // Avoid passing C arguments in the wrong stack slots. 706 #if defined(ABI_ELFv2) 707 assert((SharedRuntime::out_preserve_stack_slots() + stk) * VMRegImpl::stack_slot_size == 96, 708 "passing C arguments in wrong stack slots"); 709 #else 710 assert((SharedRuntime::out_preserve_stack_slots() + stk) * VMRegImpl::stack_slot_size == 112, 711 "passing C arguments in wrong stack slots"); 712 #endif 713 // We fill-out regs AND regs2 if an argument must be passed in a 714 // register AND in a stack slot. If regs2 is NULL in such a 715 // situation, we bail-out with a fatal error. 716 for (int i = 0; i < total_args_passed; ++i, ++arg) { 717 // Initialize regs2 to BAD. 718 if (regs2 != NULL) regs2[i].set_bad(); 719 720 switch(sig_bt[i]) { 721 722 // 723 // If arguments 0-7 are integers, they are passed in integer registers. 724 // Argument i is placed in iarg_reg[i]. 725 // 726 case T_BOOLEAN: 727 case T_CHAR: 728 case T_BYTE: 729 case T_SHORT: 730 case T_INT: 731 // We must cast ints to longs and use full 64 bit stack slots 732 // here. We do the cast in GraphKit::gen_stub() and just guard 733 // here against loosing that change. 734 assert(CCallingConventionRequiresIntsAsLongs, 735 "argument of type int should be promoted to type long"); 736 guarantee(i > 0 && sig_bt[i-1] == T_LONG, 737 "argument of type (bt) should have been promoted to type (T_LONG,bt) for bt in " 738 "{T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}"); 739 // Do not count halves. 740 regs[i].set_bad(); 741 --arg; 742 break; 743 case T_LONG: 744 guarantee(sig_bt[i+1] == T_VOID || 745 sig_bt[i+1] == T_BOOLEAN || sig_bt[i+1] == T_CHAR || 746 sig_bt[i+1] == T_BYTE || sig_bt[i+1] == T_SHORT || 747 sig_bt[i+1] == T_INT, 748 "expecting type (T_LONG,half) or type (T_LONG,bt) with bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}"); 749 case T_OBJECT: 750 case T_ARRAY: 751 case T_ADDRESS: 752 case T_METADATA: 753 // Oops are already boxed if required (JNI). 754 if (arg < Argument::n_int_register_parameters_c) { 755 reg = iarg_reg[arg]; 756 } else { 757 reg = VMRegImpl::stack2reg(stk); 758 stk += inc_stk_for_longdouble; 759 } 760 regs[i].set2(reg); 761 break; 762 763 // 764 // Floats are treated differently from int regs: The first 13 float arguments 765 // are passed in registers (not the float args among the first 13 args). 766 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed 767 // in farg_reg[j] if argument i is the j-th float argument of this call. 768 // 769 case T_FLOAT: 770 if (freg < Argument::n_float_register_parameters_c) { 771 // Put float in register ... 772 reg = farg_reg[freg]; 773 ++freg; 774 775 // Argument i for i > 8 is placed on the stack even if it's 776 // placed in a register (if it's a float arg). Aix disassembly 777 // shows that xlC places these float args on the stack AND in 778 // a register. This is not documented, but we follow this 779 // convention, too. 780 if (arg >= Argument::n_regs_not_on_stack_c) { 781 // ... and on the stack. 782 guarantee(regs2 != NULL, "must pass float in register and stack slot"); 783 VMReg reg2 = VMRegImpl::stack2reg(stk LINUX_ONLY(+1)); 784 regs2[i].set1(reg2); 785 stk += inc_stk_for_intfloat; 786 } 787 788 } else { 789 // Put float on stack. 790 reg = VMRegImpl::stack2reg(stk LINUX_ONLY(+1)); 791 stk += inc_stk_for_intfloat; 792 } 793 regs[i].set1(reg); 794 break; 795 case T_DOUBLE: 796 assert(sig_bt[i+1] == T_VOID, "expecting half"); 797 if (freg < Argument::n_float_register_parameters_c) { 798 // Put double in register ... 799 reg = farg_reg[freg]; 800 ++freg; 801 802 // Argument i for i > 8 is placed on the stack even if it's 803 // placed in a register (if it's a double arg). Aix disassembly 804 // shows that xlC places these float args on the stack AND in 805 // a register. This is not documented, but we follow this 806 // convention, too. 807 if (arg >= Argument::n_regs_not_on_stack_c) { 808 // ... and on the stack. 809 guarantee(regs2 != NULL, "must pass float in register and stack slot"); 810 VMReg reg2 = VMRegImpl::stack2reg(stk); 811 regs2[i].set2(reg2); 812 stk += inc_stk_for_longdouble; 813 } 814 } else { 815 // Put double on stack. 816 reg = VMRegImpl::stack2reg(stk); 817 stk += inc_stk_for_longdouble; 818 } 819 regs[i].set2(reg); 820 break; 821 822 case T_VOID: 823 // Do not count halves. 824 regs[i].set_bad(); 825 --arg; 826 break; 827 default: 828 ShouldNotReachHere(); 829 } 830 } 831 832 return round_to(stk, 2); 833 } 834 #endif // COMPILER2 835 836 static address gen_c2i_adapter(MacroAssembler *masm, 837 int total_args_passed, 838 int comp_args_on_stack, 839 const BasicType *sig_bt, 840 const VMRegPair *regs, 841 Label& call_interpreter, 842 const Register& ientry) { 843 844 address c2i_entrypoint; 845 846 const Register sender_SP = R21_sender_SP; // == R21_tmp1 847 const Register code = R22_tmp2; 848 //const Register ientry = R23_tmp3; 849 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 }; 850 const int num_value_regs = sizeof(value_regs) / sizeof(Register); 851 int value_regs_index = 0; 852 853 const Register return_pc = R27_tmp7; 854 const Register tmp = R28_tmp8; 855 856 assert_different_registers(sender_SP, code, ientry, return_pc, tmp); 857 858 // Adapter needs TOP_IJAVA_FRAME_ABI. 859 const int adapter_size = frame::top_ijava_frame_abi_size + 860 round_to(total_args_passed * wordSize, frame::alignment_in_bytes); 861 862 // regular (verified) c2i entry point 863 c2i_entrypoint = __ pc(); 864 865 // Does compiled code exists? If yes, patch the caller's callsite. 866 __ ld(code, method_(code)); 867 __ cmpdi(CCR0, code, 0); 868 __ ld(ientry, method_(interpreter_entry)); // preloaded 869 __ beq(CCR0, call_interpreter); 870 871 872 // Patch caller's callsite, method_(code) was not NULL which means that 873 // compiled code exists. 874 __ mflr(return_pc); 875 __ std(return_pc, _abi(lr), R1_SP); 876 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs); 877 878 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc); 879 880 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs); 881 __ ld(return_pc, _abi(lr), R1_SP); 882 __ ld(ientry, method_(interpreter_entry)); // preloaded 883 __ mtlr(return_pc); 884 885 886 // Call the interpreter. 887 __ BIND(call_interpreter); 888 __ mtctr(ientry); 889 890 // Get a copy of the current SP for loading caller's arguments. 891 __ mr(sender_SP, R1_SP); 892 893 // Add space for the adapter. 894 __ resize_frame(-adapter_size, R12_scratch2); 895 896 int st_off = adapter_size - wordSize; 897 898 // Write the args into the outgoing interpreter space. 899 for (int i = 0; i < total_args_passed; i++) { 900 VMReg r_1 = regs[i].first(); 901 VMReg r_2 = regs[i].second(); 902 if (!r_1->is_valid()) { 903 assert(!r_2->is_valid(), ""); 904 continue; 905 } 906 if (r_1->is_stack()) { 907 Register tmp_reg = value_regs[value_regs_index]; 908 value_regs_index = (value_regs_index + 1) % num_value_regs; 909 // The calling convention produces OptoRegs that ignore the out 910 // preserve area (JIT's ABI). We must account for it here. 911 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 912 if (!r_2->is_valid()) { 913 __ lwz(tmp_reg, ld_off, sender_SP); 914 } else { 915 __ ld(tmp_reg, ld_off, sender_SP); 916 } 917 // Pretend stack targets were loaded into tmp_reg. 918 r_1 = tmp_reg->as_VMReg(); 919 } 920 921 if (r_1->is_Register()) { 922 Register r = r_1->as_Register(); 923 if (!r_2->is_valid()) { 924 __ stw(r, st_off, R1_SP); 925 st_off-=wordSize; 926 } else { 927 // Longs are given 2 64-bit slots in the interpreter, but the 928 // data is passed in only 1 slot. 929 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 930 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); ) 931 st_off-=wordSize; 932 } 933 __ std(r, st_off, R1_SP); 934 st_off-=wordSize; 935 } 936 } else { 937 assert(r_1->is_FloatRegister(), ""); 938 FloatRegister f = r_1->as_FloatRegister(); 939 if (!r_2->is_valid()) { 940 __ stfs(f, st_off, R1_SP); 941 st_off-=wordSize; 942 } else { 943 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the 944 // data is passed in only 1 slot. 945 // One of these should get known junk... 946 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); ) 947 st_off-=wordSize; 948 __ stfd(f, st_off, R1_SP); 949 st_off-=wordSize; 950 } 951 } 952 } 953 954 // Jump to the interpreter just as if interpreter was doing it. 955 956 #ifdef CC_INTERP 957 const Register tos = R17_tos; 958 #else 959 const Register tos = R15_esp; 960 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 961 #endif 962 963 // load TOS 964 __ addi(tos, R1_SP, st_off); 965 966 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1. 967 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register"); 968 __ bctr(); 969 970 return c2i_entrypoint; 971 } 972 973 static void gen_i2c_adapter(MacroAssembler *masm, 974 int total_args_passed, 975 int comp_args_on_stack, 976 const BasicType *sig_bt, 977 const VMRegPair *regs) { 978 979 // Load method's entry-point from method. 980 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method); 981 __ mtctr(R12_scratch2); 982 983 // We will only enter here from an interpreted frame and never from after 984 // passing thru a c2i. Azul allowed this but we do not. If we lose the 985 // race and use a c2i we will remain interpreted for the race loser(s). 986 // This removes all sorts of headaches on the x86 side and also eliminates 987 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions. 988 989 // Note: r13 contains the senderSP on entry. We must preserve it since 990 // we may do a i2c -> c2i transition if we lose a race where compiled 991 // code goes non-entrant while we get args ready. 992 // In addition we use r13 to locate all the interpreter args as 993 // we must align the stack to 16 bytes on an i2c entry else we 994 // lose alignment we expect in all compiled code and register 995 // save code can segv when fxsave instructions find improperly 996 // aligned stack pointer. 997 998 #ifdef CC_INTERP 999 const Register ld_ptr = R17_tos; 1000 #else 1001 const Register ld_ptr = R15_esp; 1002 #endif 1003 1004 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 }; 1005 const int num_value_regs = sizeof(value_regs) / sizeof(Register); 1006 int value_regs_index = 0; 1007 1008 int ld_offset = total_args_passed*wordSize; 1009 1010 // Cut-out for having no stack args. Since up to 2 int/oop args are passed 1011 // in registers, we will occasionally have no stack args. 1012 int comp_words_on_stack = 0; 1013 if (comp_args_on_stack) { 1014 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in 1015 // registers are below. By subtracting stack0, we either get a negative 1016 // number (all values in registers) or the maximum stack slot accessed. 1017 1018 // Convert 4-byte c2 stack slots to words. 1019 comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 1020 // Round up to miminum stack alignment, in wordSize. 1021 comp_words_on_stack = round_to(comp_words_on_stack, 2); 1022 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1); 1023 } 1024 1025 // Now generate the shuffle code. Pick up all register args and move the 1026 // rest through register value=Z_R12. 1027 BLOCK_COMMENT("Shuffle arguments"); 1028 for (int i = 0; i < total_args_passed; i++) { 1029 if (sig_bt[i] == T_VOID) { 1030 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 1031 continue; 1032 } 1033 1034 // Pick up 0, 1 or 2 words from ld_ptr. 1035 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 1036 "scrambled load targets?"); 1037 VMReg r_1 = regs[i].first(); 1038 VMReg r_2 = regs[i].second(); 1039 if (!r_1->is_valid()) { 1040 assert(!r_2->is_valid(), ""); 1041 continue; 1042 } 1043 if (r_1->is_FloatRegister()) { 1044 if (!r_2->is_valid()) { 1045 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr); 1046 ld_offset-=wordSize; 1047 } else { 1048 // Skip the unused interpreter slot. 1049 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr); 1050 ld_offset-=2*wordSize; 1051 } 1052 } else { 1053 Register r; 1054 if (r_1->is_stack()) { 1055 // Must do a memory to memory move thru "value". 1056 r = value_regs[value_regs_index]; 1057 value_regs_index = (value_regs_index + 1) % num_value_regs; 1058 } else { 1059 r = r_1->as_Register(); 1060 } 1061 if (!r_2->is_valid()) { 1062 // Not sure we need to do this but it shouldn't hurt. 1063 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ADDRESS || sig_bt[i] == T_ARRAY) { 1064 __ ld(r, ld_offset, ld_ptr); 1065 ld_offset-=wordSize; 1066 } else { 1067 __ lwz(r, ld_offset, ld_ptr); 1068 ld_offset-=wordSize; 1069 } 1070 } else { 1071 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the 1072 // data is passed in only 1 slot. 1073 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 1074 ld_offset-=wordSize; 1075 } 1076 __ ld(r, ld_offset, ld_ptr); 1077 ld_offset-=wordSize; 1078 } 1079 1080 if (r_1->is_stack()) { 1081 // Now store value where the compiler expects it 1082 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size; 1083 1084 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN || 1085 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) { 1086 __ stw(r, st_off, R1_SP); 1087 } else { 1088 __ std(r, st_off, R1_SP); 1089 } 1090 } 1091 } 1092 } 1093 1094 BLOCK_COMMENT("Store method"); 1095 // Store method into thread->callee_target. 1096 // We might end up in handle_wrong_method if the callee is 1097 // deoptimized as we race thru here. If that happens we don't want 1098 // to take a safepoint because the caller frame will look 1099 // interpreted and arguments are now "compiled" so it is much better 1100 // to make this transition invisible to the stack walking 1101 // code. Unfortunately if we try and find the callee by normal means 1102 // a safepoint is possible. So we stash the desired callee in the 1103 // thread and the vm will find there should this case occur. 1104 __ std(R19_method, thread_(callee_target)); 1105 1106 // Jump to the compiled code just as if compiled code was doing it. 1107 __ bctr(); 1108 } 1109 1110 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 1111 int total_args_passed, 1112 int comp_args_on_stack, 1113 const BasicType *sig_bt, 1114 const VMRegPair *regs, 1115 AdapterFingerPrint* fingerprint) { 1116 address i2c_entry; 1117 address c2i_unverified_entry; 1118 address c2i_entry; 1119 1120 1121 // entry: i2c 1122 1123 __ align(CodeEntryAlignment); 1124 i2c_entry = __ pc(); 1125 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 1126 1127 1128 // entry: c2i unverified 1129 1130 __ align(CodeEntryAlignment); 1131 BLOCK_COMMENT("c2i unverified entry"); 1132 c2i_unverified_entry = __ pc(); 1133 1134 // inline_cache contains a compiledICHolder 1135 const Register ic = R19_method; 1136 const Register ic_klass = R11_scratch1; 1137 const Register receiver_klass = R12_scratch2; 1138 const Register code = R21_tmp1; 1139 const Register ientry = R23_tmp3; 1140 1141 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry); 1142 assert(R11_scratch1 == R11, "need prologue scratch register"); 1143 1144 Label call_interpreter; 1145 1146 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), 1147 "klass offset should reach into any page"); 1148 // Check for NULL argument if we don't have implicit null checks. 1149 if (!ImplicitNullChecks || !os::zero_page_read_protected()) { 1150 if (TrapBasedNullChecks) { 1151 __ trap_null_check(R3_ARG1); 1152 } else { 1153 Label valid; 1154 __ cmpdi(CCR0, R3_ARG1, 0); 1155 __ bne_predict_taken(CCR0, valid); 1156 // We have a null argument, branch to ic_miss_stub. 1157 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1158 relocInfo::runtime_call_type); 1159 __ BIND(valid); 1160 } 1161 } 1162 // Assume argument is not NULL, load klass from receiver. 1163 __ load_klass(receiver_klass, R3_ARG1); 1164 1165 __ ld(ic_klass, CompiledICHolder::holder_klass_offset(), ic); 1166 1167 if (TrapBasedICMissChecks) { 1168 __ trap_ic_miss_check(receiver_klass, ic_klass); 1169 } else { 1170 Label valid; 1171 __ cmpd(CCR0, receiver_klass, ic_klass); 1172 __ beq_predict_taken(CCR0, valid); 1173 // We have an unexpected klass, branch to ic_miss_stub. 1174 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1175 relocInfo::runtime_call_type); 1176 __ BIND(valid); 1177 } 1178 1179 // Argument is valid and klass is as expected, continue. 1180 1181 // Extract method from inline cache, verified entry point needs it. 1182 __ ld(R19_method, CompiledICHolder::holder_method_offset(), ic); 1183 assert(R19_method == ic, "the inline cache register is dead here"); 1184 1185 __ ld(code, method_(code)); 1186 __ cmpdi(CCR0, code, 0); 1187 __ ld(ientry, method_(interpreter_entry)); // preloaded 1188 __ beq_predict_taken(CCR0, call_interpreter); 1189 1190 // Branch to ic_miss_stub. 1191 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type); 1192 1193 // entry: c2i 1194 1195 c2i_entry = gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry); 1196 1197 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 1198 } 1199 1200 #ifdef COMPILER2 1201 // An oop arg. Must pass a handle not the oop itself. 1202 static void object_move(MacroAssembler* masm, 1203 int frame_size_in_slots, 1204 OopMap* oop_map, int oop_handle_offset, 1205 bool is_receiver, int* receiver_offset, 1206 VMRegPair src, VMRegPair dst, 1207 Register r_caller_sp, Register r_temp_1, Register r_temp_2) { 1208 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)), 1209 "receiver has already been moved"); 1210 1211 // We must pass a handle. First figure out the location we use as a handle. 1212 1213 if (src.first()->is_stack()) { 1214 // stack to stack or reg 1215 1216 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register(); 1217 Label skip; 1218 const int oop_slot_in_callers_frame = reg2slot(src.first()); 1219 1220 guarantee(!is_receiver, "expecting receiver in register"); 1221 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots)); 1222 1223 __ addi(r_handle, r_caller_sp, reg2offset(src.first())); 1224 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp); 1225 __ cmpdi(CCR0, r_temp_2, 0); 1226 __ bne(CCR0, skip); 1227 // Use a NULL handle if oop is NULL. 1228 __ li(r_handle, 0); 1229 __ bind(skip); 1230 1231 if (dst.first()->is_stack()) { 1232 // stack to stack 1233 __ std(r_handle, reg2offset(dst.first()), R1_SP); 1234 } else { 1235 // stack to reg 1236 // Nothing to do, r_handle is already the dst register. 1237 } 1238 } else { 1239 // reg to stack or reg 1240 const Register r_oop = src.first()->as_Register(); 1241 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register(); 1242 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word 1243 + oop_handle_offset; // in slots 1244 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size; 1245 Label skip; 1246 1247 if (is_receiver) { 1248 *receiver_offset = oop_offset; 1249 } 1250 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1251 1252 __ std( r_oop, oop_offset, R1_SP); 1253 __ addi(r_handle, R1_SP, oop_offset); 1254 1255 __ cmpdi(CCR0, r_oop, 0); 1256 __ bne(CCR0, skip); 1257 // Use a NULL handle if oop is NULL. 1258 __ li(r_handle, 0); 1259 __ bind(skip); 1260 1261 if (dst.first()->is_stack()) { 1262 // reg to stack 1263 __ std(r_handle, reg2offset(dst.first()), R1_SP); 1264 } else { 1265 // reg to reg 1266 // Nothing to do, r_handle is already the dst register. 1267 } 1268 } 1269 } 1270 1271 static void int_move(MacroAssembler*masm, 1272 VMRegPair src, VMRegPair dst, 1273 Register r_caller_sp, Register r_temp) { 1274 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long-int"); 1275 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long"); 1276 1277 if (src.first()->is_stack()) { 1278 if (dst.first()->is_stack()) { 1279 // stack to stack 1280 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp); 1281 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1282 } else { 1283 // stack to reg 1284 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1285 } 1286 } else if (dst.first()->is_stack()) { 1287 // reg to stack 1288 __ extsw(r_temp, src.first()->as_Register()); 1289 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1290 } else { 1291 // reg to reg 1292 __ extsw(dst.first()->as_Register(), src.first()->as_Register()); 1293 } 1294 } 1295 1296 static void long_move(MacroAssembler*masm, 1297 VMRegPair src, VMRegPair dst, 1298 Register r_caller_sp, Register r_temp) { 1299 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long"); 1300 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long"); 1301 1302 if (src.first()->is_stack()) { 1303 if (dst.first()->is_stack()) { 1304 // stack to stack 1305 __ ld( r_temp, reg2offset(src.first()), r_caller_sp); 1306 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1307 } else { 1308 // stack to reg 1309 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1310 } 1311 } else if (dst.first()->is_stack()) { 1312 // reg to stack 1313 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP); 1314 } else { 1315 // reg to reg 1316 if (dst.first()->as_Register() != src.first()->as_Register()) 1317 __ mr(dst.first()->as_Register(), src.first()->as_Register()); 1318 } 1319 } 1320 1321 static void float_move(MacroAssembler*masm, 1322 VMRegPair src, VMRegPair dst, 1323 Register r_caller_sp, Register r_temp) { 1324 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float"); 1325 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float"); 1326 1327 if (src.first()->is_stack()) { 1328 if (dst.first()->is_stack()) { 1329 // stack to stack 1330 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp); 1331 __ stw(r_temp, reg2offset(dst.first()), R1_SP); 1332 } else { 1333 // stack to reg 1334 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp); 1335 } 1336 } else if (dst.first()->is_stack()) { 1337 // reg to stack 1338 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP); 1339 } else { 1340 // reg to reg 1341 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister()) 1342 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1343 } 1344 } 1345 1346 static void double_move(MacroAssembler*masm, 1347 VMRegPair src, VMRegPair dst, 1348 Register r_caller_sp, Register r_temp) { 1349 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double"); 1350 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double"); 1351 1352 if (src.first()->is_stack()) { 1353 if (dst.first()->is_stack()) { 1354 // stack to stack 1355 __ ld( r_temp, reg2offset(src.first()), r_caller_sp); 1356 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1357 } else { 1358 // stack to reg 1359 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp); 1360 } 1361 } else if (dst.first()->is_stack()) { 1362 // reg to stack 1363 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP); 1364 } else { 1365 // reg to reg 1366 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister()) 1367 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1368 } 1369 } 1370 1371 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1372 switch (ret_type) { 1373 case T_BOOLEAN: 1374 case T_CHAR: 1375 case T_BYTE: 1376 case T_SHORT: 1377 case T_INT: 1378 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1379 break; 1380 case T_ARRAY: 1381 case T_OBJECT: 1382 case T_LONG: 1383 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1384 break; 1385 case T_FLOAT: 1386 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1387 break; 1388 case T_DOUBLE: 1389 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1390 break; 1391 case T_VOID: 1392 break; 1393 default: 1394 ShouldNotReachHere(); 1395 break; 1396 } 1397 } 1398 1399 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1400 switch (ret_type) { 1401 case T_BOOLEAN: 1402 case T_CHAR: 1403 case T_BYTE: 1404 case T_SHORT: 1405 case T_INT: 1406 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1407 break; 1408 case T_ARRAY: 1409 case T_OBJECT: 1410 case T_LONG: 1411 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1412 break; 1413 case T_FLOAT: 1414 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1415 break; 1416 case T_DOUBLE: 1417 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1418 break; 1419 case T_VOID: 1420 break; 1421 default: 1422 ShouldNotReachHere(); 1423 break; 1424 } 1425 } 1426 1427 static void save_or_restore_arguments(MacroAssembler* masm, 1428 const int stack_slots, 1429 const int total_in_args, 1430 const int arg_save_area, 1431 OopMap* map, 1432 VMRegPair* in_regs, 1433 BasicType* in_sig_bt) { 1434 // If map is non-NULL then the code should store the values, 1435 // otherwise it should load them. 1436 int slot = arg_save_area; 1437 // Save down double word first. 1438 for (int i = 0; i < total_in_args; i++) { 1439 if (in_regs[i].first()->is_FloatRegister() && in_sig_bt[i] == T_DOUBLE) { 1440 int offset = slot * VMRegImpl::stack_slot_size; 1441 slot += VMRegImpl::slots_per_word; 1442 assert(slot <= stack_slots, "overflow (after DOUBLE stack slot)"); 1443 if (map != NULL) { 1444 __ stfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1445 } else { 1446 __ lfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1447 } 1448 } else if (in_regs[i].first()->is_Register() && 1449 (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_ARRAY)) { 1450 int offset = slot * VMRegImpl::stack_slot_size; 1451 if (map != NULL) { 1452 __ std(in_regs[i].first()->as_Register(), offset, R1_SP); 1453 if (in_sig_bt[i] == T_ARRAY) { 1454 map->set_oop(VMRegImpl::stack2reg(slot)); 1455 } 1456 } else { 1457 __ ld(in_regs[i].first()->as_Register(), offset, R1_SP); 1458 } 1459 slot += VMRegImpl::slots_per_word; 1460 assert(slot <= stack_slots, "overflow (after LONG/ARRAY stack slot)"); 1461 } 1462 } 1463 // Save or restore single word registers. 1464 for (int i = 0; i < total_in_args; i++) { 1465 // PPC64: pass ints as longs: must only deal with floats here. 1466 if (in_regs[i].first()->is_FloatRegister()) { 1467 if (in_sig_bt[i] == T_FLOAT) { 1468 int offset = slot * VMRegImpl::stack_slot_size; 1469 slot++; 1470 assert(slot <= stack_slots, "overflow (after FLOAT stack slot)"); 1471 if (map != NULL) { 1472 __ stfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1473 } else { 1474 __ lfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1475 } 1476 } 1477 } else if (in_regs[i].first()->is_stack()) { 1478 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1479 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1480 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots)); 1481 } 1482 } 1483 } 1484 } 1485 1486 // Check GC_locker::needs_gc and enter the runtime if it's true. This 1487 // keeps a new JNI critical region from starting until a GC has been 1488 // forced. Save down any oops in registers and describe them in an 1489 // OopMap. 1490 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1491 const int stack_slots, 1492 const int total_in_args, 1493 const int arg_save_area, 1494 OopMapSet* oop_maps, 1495 VMRegPair* in_regs, 1496 BasicType* in_sig_bt, 1497 Register tmp_reg ) { 1498 __ block_comment("check GC_locker::needs_gc"); 1499 Label cont; 1500 __ lbz(tmp_reg, (RegisterOrConstant)(intptr_t)GC_locker::needs_gc_address()); 1501 __ cmplwi(CCR0, tmp_reg, 0); 1502 __ beq(CCR0, cont); 1503 1504 // Save down any values that are live in registers and call into the 1505 // runtime to halt for a GC. 1506 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1507 save_or_restore_arguments(masm, stack_slots, total_in_args, 1508 arg_save_area, map, in_regs, in_sig_bt); 1509 1510 __ mr(R3_ARG1, R16_thread); 1511 __ set_last_Java_frame(R1_SP, noreg); 1512 1513 __ block_comment("block_for_jni_critical"); 1514 address entry_point = CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical); 1515 #if defined(ABI_ELFv2) 1516 __ call_c(entry_point, relocInfo::runtime_call_type); 1517 #else 1518 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, entry_point), relocInfo::runtime_call_type); 1519 #endif 1520 address start = __ pc() - __ offset(), 1521 calls_return_pc = __ last_calls_return_pc(); 1522 oop_maps->add_gc_map(calls_return_pc - start, map); 1523 1524 __ reset_last_Java_frame(); 1525 1526 // Reload all the register arguments. 1527 save_or_restore_arguments(masm, stack_slots, total_in_args, 1528 arg_save_area, NULL, in_regs, in_sig_bt); 1529 1530 __ BIND(cont); 1531 1532 #ifdef ASSERT 1533 if (StressCriticalJNINatives) { 1534 // Stress register saving. 1535 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1536 save_or_restore_arguments(masm, stack_slots, total_in_args, 1537 arg_save_area, map, in_regs, in_sig_bt); 1538 // Destroy argument registers. 1539 for (int i = 0; i < total_in_args; i++) { 1540 if (in_regs[i].first()->is_Register()) { 1541 const Register reg = in_regs[i].first()->as_Register(); 1542 __ neg(reg, reg); 1543 } else if (in_regs[i].first()->is_FloatRegister()) { 1544 __ fneg(in_regs[i].first()->as_FloatRegister(), in_regs[i].first()->as_FloatRegister()); 1545 } 1546 } 1547 1548 save_or_restore_arguments(masm, stack_slots, total_in_args, 1549 arg_save_area, NULL, in_regs, in_sig_bt); 1550 } 1551 #endif 1552 } 1553 1554 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst, Register r_caller_sp, Register r_temp) { 1555 if (src.first()->is_stack()) { 1556 if (dst.first()->is_stack()) { 1557 // stack to stack 1558 __ ld(r_temp, reg2offset(src.first()), r_caller_sp); 1559 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1560 } else { 1561 // stack to reg 1562 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1563 } 1564 } else if (dst.first()->is_stack()) { 1565 // reg to stack 1566 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP); 1567 } else { 1568 if (dst.first() != src.first()) { 1569 __ mr(dst.first()->as_Register(), src.first()->as_Register()); 1570 } 1571 } 1572 } 1573 1574 // Unpack an array argument into a pointer to the body and the length 1575 // if the array is non-null, otherwise pass 0 for both. 1576 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, 1577 VMRegPair body_arg, VMRegPair length_arg, Register r_caller_sp, 1578 Register tmp_reg, Register tmp2_reg) { 1579 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg, 1580 "possible collision"); 1581 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg, 1582 "possible collision"); 1583 1584 // Pass the length, ptr pair. 1585 Label set_out_args; 1586 VMRegPair tmp, tmp2; 1587 tmp.set_ptr(tmp_reg->as_VMReg()); 1588 tmp2.set_ptr(tmp2_reg->as_VMReg()); 1589 if (reg.first()->is_stack()) { 1590 // Load the arg up from the stack. 1591 move_ptr(masm, reg, tmp, r_caller_sp, /*unused*/ R0); 1592 reg = tmp; 1593 } 1594 __ li(tmp2_reg, 0); // Pass zeros if Array=null. 1595 if (tmp_reg != reg.first()->as_Register()) __ li(tmp_reg, 0); 1596 __ cmpdi(CCR0, reg.first()->as_Register(), 0); 1597 __ beq(CCR0, set_out_args); 1598 __ lwa(tmp2_reg, arrayOopDesc::length_offset_in_bytes(), reg.first()->as_Register()); 1599 __ addi(tmp_reg, reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type)); 1600 __ bind(set_out_args); 1601 move_ptr(masm, tmp, body_arg, r_caller_sp, /*unused*/ R0); 1602 move_ptr(masm, tmp2, length_arg, r_caller_sp, /*unused*/ R0); // Same as move32_64 on PPC64. 1603 } 1604 1605 static void verify_oop_args(MacroAssembler* masm, 1606 methodHandle method, 1607 const BasicType* sig_bt, 1608 const VMRegPair* regs) { 1609 Register temp_reg = R19_method; // not part of any compiled calling seq 1610 if (VerifyOops) { 1611 for (int i = 0; i < method->size_of_parameters(); i++) { 1612 if (sig_bt[i] == T_OBJECT || 1613 sig_bt[i] == T_ARRAY) { 1614 VMReg r = regs[i].first(); 1615 assert(r->is_valid(), "bad oop arg"); 1616 if (r->is_stack()) { 1617 __ ld(temp_reg, reg2offset(r), R1_SP); 1618 __ verify_oop(temp_reg); 1619 } else { 1620 __ verify_oop(r->as_Register()); 1621 } 1622 } 1623 } 1624 } 1625 } 1626 1627 static void gen_special_dispatch(MacroAssembler* masm, 1628 methodHandle method, 1629 const BasicType* sig_bt, 1630 const VMRegPair* regs) { 1631 verify_oop_args(masm, method, sig_bt, regs); 1632 vmIntrinsics::ID iid = method->intrinsic_id(); 1633 1634 // Now write the args into the outgoing interpreter space 1635 bool has_receiver = false; 1636 Register receiver_reg = noreg; 1637 int member_arg_pos = -1; 1638 Register member_reg = noreg; 1639 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1640 if (ref_kind != 0) { 1641 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1642 member_reg = R19_method; // known to be free at this point 1643 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1644 } else if (iid == vmIntrinsics::_invokeBasic) { 1645 has_receiver = true; 1646 } else { 1647 fatal(err_msg_res("unexpected intrinsic id %d", iid)); 1648 } 1649 1650 if (member_reg != noreg) { 1651 // Load the member_arg into register, if necessary. 1652 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1653 VMReg r = regs[member_arg_pos].first(); 1654 if (r->is_stack()) { 1655 __ ld(member_reg, reg2offset(r), R1_SP); 1656 } else { 1657 // no data motion is needed 1658 member_reg = r->as_Register(); 1659 } 1660 } 1661 1662 if (has_receiver) { 1663 // Make sure the receiver is loaded into a register. 1664 assert(method->size_of_parameters() > 0, "oob"); 1665 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1666 VMReg r = regs[0].first(); 1667 assert(r->is_valid(), "bad receiver arg"); 1668 if (r->is_stack()) { 1669 // Porting note: This assumes that compiled calling conventions always 1670 // pass the receiver oop in a register. If this is not true on some 1671 // platform, pick a temp and load the receiver from stack. 1672 fatal("receiver always in a register"); 1673 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point? 1674 __ ld(receiver_reg, reg2offset(r), R1_SP); 1675 } else { 1676 // no data motion is needed 1677 receiver_reg = r->as_Register(); 1678 } 1679 } 1680 1681 // Figure out which address we are really jumping to: 1682 MethodHandles::generate_method_handle_dispatch(masm, iid, 1683 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1684 } 1685 1686 #endif // COMPILER2 1687 1688 // --------------------------------------------------------------------------- 1689 // Generate a native wrapper for a given method. The method takes arguments 1690 // in the Java compiled code convention, marshals them to the native 1691 // convention (handlizes oops, etc), transitions to native, makes the call, 1692 // returns to java state (possibly blocking), unhandlizes any result and 1693 // returns. 1694 // 1695 // Critical native functions are a shorthand for the use of 1696 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1697 // functions. The wrapper is expected to unpack the arguments before 1698 // passing them to the callee and perform checks before and after the 1699 // native call to ensure that they GC_locker 1700 // lock_critical/unlock_critical semantics are followed. Some other 1701 // parts of JNI setup are skipped like the tear down of the JNI handle 1702 // block and the check for pending exceptions it's impossible for them 1703 // to be thrown. 1704 // 1705 // They are roughly structured like this: 1706 // if (GC_locker::needs_gc()) 1707 // SharedRuntime::block_for_jni_critical(); 1708 // tranistion to thread_in_native 1709 // unpack arrray arguments and call native entry point 1710 // check for safepoint in progress 1711 // check if any thread suspend flags are set 1712 // call into JVM and possible unlock the JNI critical 1713 // if a GC was suppressed while in the critical native. 1714 // transition back to thread_in_Java 1715 // return to caller 1716 // 1717 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm, 1718 methodHandle method, 1719 int compile_id, 1720 BasicType *in_sig_bt, 1721 VMRegPair *in_regs, 1722 BasicType ret_type) { 1723 #ifdef COMPILER2 1724 if (method->is_method_handle_intrinsic()) { 1725 vmIntrinsics::ID iid = method->intrinsic_id(); 1726 intptr_t start = (intptr_t)__ pc(); 1727 int vep_offset = ((intptr_t)__ pc()) - start; 1728 gen_special_dispatch(masm, 1729 method, 1730 in_sig_bt, 1731 in_regs); 1732 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1733 __ flush(); 1734 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1735 return nmethod::new_native_nmethod(method, 1736 compile_id, 1737 masm->code(), 1738 vep_offset, 1739 frame_complete, 1740 stack_slots / VMRegImpl::slots_per_word, 1741 in_ByteSize(-1), 1742 in_ByteSize(-1), 1743 (OopMapSet*)NULL); 1744 } 1745 1746 bool is_critical_native = true; 1747 address native_func = method->critical_native_function(); 1748 if (native_func == NULL) { 1749 native_func = method->native_function(); 1750 is_critical_native = false; 1751 } 1752 assert(native_func != NULL, "must have function"); 1753 1754 // First, create signature for outgoing C call 1755 // -------------------------------------------------------------------------- 1756 1757 int total_in_args = method->size_of_parameters(); 1758 // We have received a description of where all the java args are located 1759 // on entry to the wrapper. We need to convert these args to where 1760 // the jni function will expect them. To figure out where they go 1761 // we convert the java signature to a C signature by inserting 1762 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1763 // 1764 // Additionally, on ppc64 we must convert integers to longs in the C 1765 // signature. We do this in advance in order to have no trouble with 1766 // indexes into the bt-arrays. 1767 // So convert the signature and registers now, and adjust the total number 1768 // of in-arguments accordingly. 1769 int i2l_argcnt = convert_ints_to_longints_argcnt(total_in_args, in_sig_bt); // PPC64: pass ints as longs. 1770 1771 // Calculate the total number of C arguments and create arrays for the 1772 // signature and the outgoing registers. 1773 // On ppc64, we have two arrays for the outgoing registers, because 1774 // some floating-point arguments must be passed in registers _and_ 1775 // in stack locations. 1776 bool method_is_static = method->is_static(); 1777 int total_c_args = i2l_argcnt; 1778 1779 if (!is_critical_native) { 1780 int n_hidden_args = method_is_static ? 2 : 1; 1781 total_c_args += n_hidden_args; 1782 } else { 1783 // No JNIEnv*, no this*, but unpacked arrays (base+length). 1784 for (int i = 0; i < total_in_args; i++) { 1785 if (in_sig_bt[i] == T_ARRAY) { 1786 total_c_args += 2; // PPC64: T_LONG, T_INT, T_ADDRESS (see convert_ints_to_longints and c_calling_convention) 1787 } 1788 } 1789 } 1790 1791 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1792 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1793 VMRegPair *out_regs2 = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1794 BasicType* in_elem_bt = NULL; 1795 1796 // Create the signature for the C call: 1797 // 1) add the JNIEnv* 1798 // 2) add the class if the method is static 1799 // 3) copy the rest of the incoming signature (shifted by the number of 1800 // hidden arguments). 1801 1802 int argc = 0; 1803 if (!is_critical_native) { 1804 convert_ints_to_longints(i2l_argcnt, total_in_args, in_sig_bt, in_regs); // PPC64: pass ints as longs. 1805 1806 out_sig_bt[argc++] = T_ADDRESS; 1807 if (method->is_static()) { 1808 out_sig_bt[argc++] = T_OBJECT; 1809 } 1810 1811 for (int i = 0; i < total_in_args ; i++ ) { 1812 out_sig_bt[argc++] = in_sig_bt[i]; 1813 } 1814 } else { 1815 Thread* THREAD = Thread::current(); 1816 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, i2l_argcnt); 1817 SignatureStream ss(method->signature()); 1818 int o = 0; 1819 for (int i = 0; i < total_in_args ; i++, o++) { 1820 if (in_sig_bt[i] == T_ARRAY) { 1821 // Arrays are passed as int, elem* pair 1822 Symbol* atype = ss.as_symbol(CHECK_NULL); 1823 const char* at = atype->as_C_string(); 1824 if (strlen(at) == 2) { 1825 assert(at[0] == '[', "must be"); 1826 switch (at[1]) { 1827 case 'B': in_elem_bt[o] = T_BYTE; break; 1828 case 'C': in_elem_bt[o] = T_CHAR; break; 1829 case 'D': in_elem_bt[o] = T_DOUBLE; break; 1830 case 'F': in_elem_bt[o] = T_FLOAT; break; 1831 case 'I': in_elem_bt[o] = T_INT; break; 1832 case 'J': in_elem_bt[o] = T_LONG; break; 1833 case 'S': in_elem_bt[o] = T_SHORT; break; 1834 case 'Z': in_elem_bt[o] = T_BOOLEAN; break; 1835 default: ShouldNotReachHere(); 1836 } 1837 } 1838 } else { 1839 in_elem_bt[o] = T_VOID; 1840 switch(in_sig_bt[i]) { // PPC64: pass ints as longs. 1841 case T_BOOLEAN: 1842 case T_CHAR: 1843 case T_BYTE: 1844 case T_SHORT: 1845 case T_INT: in_elem_bt[++o] = T_VOID; break; 1846 default: break; 1847 } 1848 } 1849 if (in_sig_bt[i] != T_VOID) { 1850 assert(in_sig_bt[i] == ss.type(), "must match"); 1851 ss.next(); 1852 } 1853 } 1854 assert(i2l_argcnt==o, "must match"); 1855 1856 convert_ints_to_longints(i2l_argcnt, total_in_args, in_sig_bt, in_regs); // PPC64: pass ints as longs. 1857 1858 for (int i = 0; i < total_in_args ; i++ ) { 1859 if (in_sig_bt[i] == T_ARRAY) { 1860 // Arrays are passed as int, elem* pair. 1861 out_sig_bt[argc++] = T_LONG; // PPC64: pass ints as longs. 1862 out_sig_bt[argc++] = T_INT; 1863 out_sig_bt[argc++] = T_ADDRESS; 1864 } else { 1865 out_sig_bt[argc++] = in_sig_bt[i]; 1866 } 1867 } 1868 } 1869 1870 1871 // Compute the wrapper's frame size. 1872 // -------------------------------------------------------------------------- 1873 1874 // Now figure out where the args must be stored and how much stack space 1875 // they require. 1876 // 1877 // Compute framesize for the wrapper. We need to handlize all oops in 1878 // incoming registers. 1879 // 1880 // Calculate the total number of stack slots we will need: 1881 // 1) abi requirements 1882 // 2) outgoing arguments 1883 // 3) space for inbound oop handle area 1884 // 4) space for handlizing a klass if static method 1885 // 5) space for a lock if synchronized method 1886 // 6) workspace for saving return values, int <-> float reg moves, etc. 1887 // 7) alignment 1888 // 1889 // Layout of the native wrapper frame: 1890 // (stack grows upwards, memory grows downwards) 1891 // 1892 // NW [ABI_REG_ARGS] <-- 1) R1_SP 1893 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset 1894 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset (save area for critical natives) 1895 // klass <-- 4) R1_SP + klass_offset 1896 // lock <-- 5) R1_SP + lock_offset 1897 // [workspace] <-- 6) R1_SP + workspace_offset 1898 // [alignment] (optional) <-- 7) 1899 // caller [JIT_TOP_ABI_48] <-- r_callers_sp 1900 // 1901 // - *_slot_offset Indicates offset from SP in number of stack slots. 1902 // - *_offset Indicates offset from SP in bytes. 1903 1904 int stack_slots = c_calling_convention(out_sig_bt, out_regs, out_regs2, total_c_args) // 1+2) 1905 + SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention. 1906 1907 // Now the space for the inbound oop handle area. 1908 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word; 1909 if (is_critical_native) { 1910 // Critical natives may have to call out so they need a save area 1911 // for register arguments. 1912 int double_slots = 0; 1913 int single_slots = 0; 1914 for (int i = 0; i < total_in_args; i++) { 1915 if (in_regs[i].first()->is_Register()) { 1916 const Register reg = in_regs[i].first()->as_Register(); 1917 switch (in_sig_bt[i]) { 1918 case T_BOOLEAN: 1919 case T_BYTE: 1920 case T_SHORT: 1921 case T_CHAR: 1922 case T_INT: /*single_slots++;*/ break; // PPC64: pass ints as longs. 1923 case T_ARRAY: 1924 case T_LONG: double_slots++; break; 1925 default: ShouldNotReachHere(); 1926 } 1927 } else if (in_regs[i].first()->is_FloatRegister()) { 1928 switch (in_sig_bt[i]) { 1929 case T_FLOAT: single_slots++; break; 1930 case T_DOUBLE: double_slots++; break; 1931 default: ShouldNotReachHere(); 1932 } 1933 } 1934 } 1935 total_save_slots = double_slots * 2 + round_to(single_slots, 2); // round to even 1936 } 1937 1938 int oop_handle_slot_offset = stack_slots; 1939 stack_slots += total_save_slots; // 3) 1940 1941 int klass_slot_offset = 0; 1942 int klass_offset = -1; 1943 if (method_is_static && !is_critical_native) { // 4) 1944 klass_slot_offset = stack_slots; 1945 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1946 stack_slots += VMRegImpl::slots_per_word; 1947 } 1948 1949 int lock_slot_offset = 0; 1950 int lock_offset = -1; 1951 if (method->is_synchronized()) { // 5) 1952 lock_slot_offset = stack_slots; 1953 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size; 1954 stack_slots += VMRegImpl::slots_per_word; 1955 } 1956 1957 int workspace_slot_offset = stack_slots; // 6) 1958 stack_slots += 2; 1959 1960 // Now compute actual number of stack words we need. 1961 // Rounding to make stack properly aligned. 1962 stack_slots = round_to(stack_slots, // 7) 1963 frame::alignment_in_bytes / VMRegImpl::stack_slot_size); 1964 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size; 1965 1966 1967 // Now we can start generating code. 1968 // -------------------------------------------------------------------------- 1969 1970 intptr_t start_pc = (intptr_t)__ pc(); 1971 intptr_t vep_start_pc; 1972 intptr_t frame_done_pc; 1973 intptr_t oopmap_pc; 1974 1975 Label ic_miss; 1976 Label handle_pending_exception; 1977 1978 Register r_callers_sp = R21; 1979 Register r_temp_1 = R22; 1980 Register r_temp_2 = R23; 1981 Register r_temp_3 = R24; 1982 Register r_temp_4 = R25; 1983 Register r_temp_5 = R26; 1984 Register r_temp_6 = R27; 1985 Register r_return_pc = R28; 1986 1987 Register r_carg1_jnienv = noreg; 1988 Register r_carg2_classorobject = noreg; 1989 if (!is_critical_native) { 1990 r_carg1_jnienv = out_regs[0].first()->as_Register(); 1991 r_carg2_classorobject = out_regs[1].first()->as_Register(); 1992 } 1993 1994 1995 // Generate the Unverified Entry Point (UEP). 1996 // -------------------------------------------------------------------------- 1997 assert(start_pc == (intptr_t)__ pc(), "uep must be at start"); 1998 1999 // Check ic: object class == cached class? 2000 if (!method_is_static) { 2001 Register ic = as_Register(Matcher::inline_cache_reg_encode()); 2002 Register receiver_klass = r_temp_1; 2003 2004 __ cmpdi(CCR0, R3_ARG1, 0); 2005 __ beq(CCR0, ic_miss); 2006 __ verify_oop(R3_ARG1); 2007 __ load_klass(receiver_klass, R3_ARG1); 2008 2009 __ cmpd(CCR0, receiver_klass, ic); 2010 __ bne(CCR0, ic_miss); 2011 } 2012 2013 2014 // Generate the Verified Entry Point (VEP). 2015 // -------------------------------------------------------------------------- 2016 vep_start_pc = (intptr_t)__ pc(); 2017 2018 __ save_LR_CR(r_temp_1); 2019 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame. 2020 __ mr(r_callers_sp, R1_SP); // Remember frame pointer. 2021 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame. 2022 frame_done_pc = (intptr_t)__ pc(); 2023 2024 // Native nmethod wrappers never take possesion of the oop arguments. 2025 // So the caller will gc the arguments. 2026 // The only thing we need an oopMap for is if the call is static. 2027 // 2028 // An OopMap for lock (and class if static), and one for the VM call itself. 2029 OopMapSet *oop_maps = new OopMapSet(); 2030 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 2031 2032 if (is_critical_native) { 2033 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); 2034 } 2035 2036 // Move arguments from register/stack to register/stack. 2037 // -------------------------------------------------------------------------- 2038 // 2039 // We immediately shuffle the arguments so that for any vm call we have 2040 // to make from here on out (sync slow path, jvmti, etc.) we will have 2041 // captured the oops from our caller and have a valid oopMap for them. 2042 // 2043 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv* 2044 // (derived from JavaThread* which is in R16_thread) and, if static, 2045 // the class mirror instead of a receiver. This pretty much guarantees that 2046 // register layout will not match. We ignore these extra arguments during 2047 // the shuffle. The shuffle is described by the two calling convention 2048 // vectors we have in our possession. We simply walk the java vector to 2049 // get the source locations and the c vector to get the destinations. 2050 2051 // Record sp-based slot for receiver on stack for non-static methods. 2052 int receiver_offset = -1; 2053 2054 // We move the arguments backward because the floating point registers 2055 // destination will always be to a register with a greater or equal 2056 // register number or the stack. 2057 // in is the index of the incoming Java arguments 2058 // out is the index of the outgoing C arguments 2059 2060 #ifdef ASSERT 2061 bool reg_destroyed[RegisterImpl::number_of_registers]; 2062 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 2063 for (int r = 0 ; r < RegisterImpl::number_of_registers ; r++) { 2064 reg_destroyed[r] = false; 2065 } 2066 for (int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++) { 2067 freg_destroyed[f] = false; 2068 } 2069 #endif // ASSERT 2070 2071 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) { 2072 2073 #ifdef ASSERT 2074 if (in_regs[in].first()->is_Register()) { 2075 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!"); 2076 } else if (in_regs[in].first()->is_FloatRegister()) { 2077 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!"); 2078 } 2079 if (out_regs[out].first()->is_Register()) { 2080 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true; 2081 } else if (out_regs[out].first()->is_FloatRegister()) { 2082 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true; 2083 } 2084 if (out_regs2[out].first()->is_Register()) { 2085 reg_destroyed[out_regs2[out].first()->as_Register()->encoding()] = true; 2086 } else if (out_regs2[out].first()->is_FloatRegister()) { 2087 freg_destroyed[out_regs2[out].first()->as_FloatRegister()->encoding()] = true; 2088 } 2089 #endif // ASSERT 2090 2091 switch (in_sig_bt[in]) { 2092 case T_BOOLEAN: 2093 case T_CHAR: 2094 case T_BYTE: 2095 case T_SHORT: 2096 case T_INT: 2097 guarantee(in > 0 && in_sig_bt[in-1] == T_LONG, 2098 "expecting type (T_LONG,bt) for bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}"); 2099 break; 2100 case T_LONG: 2101 if (in_sig_bt[in+1] == T_VOID) { 2102 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2103 } else { 2104 guarantee(in_sig_bt[in+1] == T_BOOLEAN || in_sig_bt[in+1] == T_CHAR || 2105 in_sig_bt[in+1] == T_BYTE || in_sig_bt[in+1] == T_SHORT || 2106 in_sig_bt[in+1] == T_INT, 2107 "expecting type (T_LONG,bt) for bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}"); 2108 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2109 } 2110 break; 2111 case T_ARRAY: 2112 if (is_critical_native) { 2113 int body_arg = out; 2114 out -= 2; // Point to length arg. PPC64: pass ints as longs. 2115 unpack_array_argument(masm, in_regs[in], in_elem_bt[in], out_regs[body_arg], out_regs[out], 2116 r_callers_sp, r_temp_1, r_temp_2); 2117 break; 2118 } 2119 case T_OBJECT: 2120 assert(!is_critical_native, "no oop arguments"); 2121 object_move(masm, stack_slots, 2122 oop_map, oop_handle_slot_offset, 2123 ((in == 0) && (!method_is_static)), &receiver_offset, 2124 in_regs[in], out_regs[out], 2125 r_callers_sp, r_temp_1, r_temp_2); 2126 break; 2127 case T_VOID: 2128 break; 2129 case T_FLOAT: 2130 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2131 if (out_regs2[out].first()->is_valid()) { 2132 float_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1); 2133 } 2134 break; 2135 case T_DOUBLE: 2136 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2137 if (out_regs2[out].first()->is_valid()) { 2138 double_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1); 2139 } 2140 break; 2141 case T_ADDRESS: 2142 fatal("found type (T_ADDRESS) in java args"); 2143 break; 2144 default: 2145 ShouldNotReachHere(); 2146 break; 2147 } 2148 } 2149 2150 // Pre-load a static method's oop into ARG2. 2151 // Used both by locking code and the normal JNI call code. 2152 if (method_is_static && !is_critical_native) { 2153 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()), 2154 r_carg2_classorobject); 2155 2156 // Now handlize the static class mirror in carg2. It's known not-null. 2157 __ std(r_carg2_classorobject, klass_offset, R1_SP); 2158 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 2159 __ addi(r_carg2_classorobject, R1_SP, klass_offset); 2160 } 2161 2162 // Get JNIEnv* which is first argument to native. 2163 if (!is_critical_native) { 2164 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset())); 2165 } 2166 2167 // NOTE: 2168 // 2169 // We have all of the arguments setup at this point. 2170 // We MUST NOT touch any outgoing regs from this point on. 2171 // So if we must call out we must push a new frame. 2172 2173 // Get current pc for oopmap, and load it patchable relative to global toc. 2174 oopmap_pc = (intptr_t) __ pc(); 2175 __ calculate_address_from_global_toc(r_return_pc, (address)oopmap_pc, true, true, true, true); 2176 2177 // We use the same pc/oopMap repeatedly when we call out. 2178 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map); 2179 2180 // r_return_pc now has the pc loaded that we will use when we finally call 2181 // to native. 2182 2183 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below. 2184 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register"); 2185 2186 2187 # if 0 2188 // DTrace method entry 2189 # endif 2190 2191 // Lock a synchronized method. 2192 // -------------------------------------------------------------------------- 2193 2194 if (method->is_synchronized()) { 2195 assert(!is_critical_native, "unhandled"); 2196 ConditionRegister r_flag = CCR1; 2197 Register r_oop = r_temp_4; 2198 const Register r_box = r_temp_5; 2199 Label done, locked; 2200 2201 // Load the oop for the object or class. r_carg2_classorobject contains 2202 // either the handlized oop from the incoming arguments or the handlized 2203 // class mirror (if the method is static). 2204 __ ld(r_oop, 0, r_carg2_classorobject); 2205 2206 // Get the lock box slot's address. 2207 __ addi(r_box, R1_SP, lock_offset); 2208 2209 # ifdef ASSERT 2210 if (UseBiasedLocking) { 2211 // Making the box point to itself will make it clear it went unused 2212 // but also be obviously invalid. 2213 __ std(r_box, 0, r_box); 2214 } 2215 # endif // ASSERT 2216 2217 // Try fastpath for locking. 2218 // fast_lock kills r_temp_1, r_temp_2, r_temp_3. 2219 __ compiler_fast_lock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3); 2220 __ beq(r_flag, locked); 2221 2222 // None of the above fast optimizations worked so we have to get into the 2223 // slow case of monitor enter. Inline a special case of call_VM that 2224 // disallows any pending_exception. 2225 2226 // Save argument registers and leave room for C-compatible ABI_REG_ARGS. 2227 int frame_size = frame::abi_reg_args_size + 2228 round_to(total_c_args * wordSize, frame::alignment_in_bytes); 2229 __ mr(R11_scratch1, R1_SP); 2230 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs, out_regs2); 2231 2232 // Do the call. 2233 __ set_last_Java_frame(R11_scratch1, r_return_pc); 2234 assert(r_return_pc->is_nonvolatile(), "expecting return pc to be in non-volatile register"); 2235 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread); 2236 __ reset_last_Java_frame(); 2237 2238 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs, out_regs2); 2239 2240 __ asm_assert_mem8_is_zero(thread_(pending_exception), 2241 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C", 0); 2242 2243 __ bind(locked); 2244 } 2245 2246 2247 // Publish thread state 2248 // -------------------------------------------------------------------------- 2249 2250 // Use that pc we placed in r_return_pc a while back as the current frame anchor. 2251 __ set_last_Java_frame(R1_SP, r_return_pc); 2252 2253 // Transition from _thread_in_Java to _thread_in_native. 2254 __ li(R0, _thread_in_native); 2255 __ release(); 2256 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2257 __ stw(R0, thread_(thread_state)); 2258 if (UseMembar) { 2259 __ fence(); 2260 } 2261 2262 2263 // The JNI call 2264 // -------------------------------------------------------------------------- 2265 #if defined(ABI_ELFv2) 2266 __ call_c(native_func, relocInfo::runtime_call_type); 2267 #else 2268 FunctionDescriptor* fd_native_method = (FunctionDescriptor*) native_func; 2269 __ call_c(fd_native_method, relocInfo::runtime_call_type); 2270 #endif 2271 2272 2273 // Now, we are back from the native code. 2274 2275 2276 // Unpack the native result. 2277 // -------------------------------------------------------------------------- 2278 2279 // For int-types, we do any needed sign-extension required. 2280 // Care must be taken that the return values (R3_RET and F1_RET) 2281 // will survive any VM calls for blocking or unlocking. 2282 // An OOP result (handle) is done specially in the slow-path code. 2283 2284 switch (ret_type) { 2285 case T_VOID: break; // Nothing to do! 2286 case T_FLOAT: break; // Got it where we want it (unless slow-path). 2287 case T_DOUBLE: break; // Got it where we want it (unless slow-path). 2288 case T_LONG: break; // Got it where we want it (unless slow-path). 2289 case T_OBJECT: break; // Really a handle. 2290 // Cannot de-handlize until after reclaiming jvm_lock. 2291 case T_ARRAY: break; 2292 2293 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1) 2294 Label skip_modify; 2295 __ cmpwi(CCR0, R3_RET, 0); 2296 __ beq(CCR0, skip_modify); 2297 __ li(R3_RET, 1); 2298 __ bind(skip_modify); 2299 break; 2300 } 2301 case T_BYTE: { // sign extension 2302 __ extsb(R3_RET, R3_RET); 2303 break; 2304 } 2305 case T_CHAR: { // unsigned result 2306 __ andi(R3_RET, R3_RET, 0xffff); 2307 break; 2308 } 2309 case T_SHORT: { // sign extension 2310 __ extsh(R3_RET, R3_RET); 2311 break; 2312 } 2313 case T_INT: // nothing to do 2314 break; 2315 default: 2316 ShouldNotReachHere(); 2317 break; 2318 } 2319 2320 2321 // Publish thread state 2322 // -------------------------------------------------------------------------- 2323 2324 // Switch thread to "native transition" state before reading the 2325 // synchronization state. This additional state is necessary because reading 2326 // and testing the synchronization state is not atomic w.r.t. GC, as this 2327 // scenario demonstrates: 2328 // - Java thread A, in _thread_in_native state, loads _not_synchronized 2329 // and is preempted. 2330 // - VM thread changes sync state to synchronizing and suspends threads 2331 // for GC. 2332 // - Thread A is resumed to finish this native method, but doesn't block 2333 // here since it didn't see any synchronization in progress, and escapes. 2334 2335 // Transition from _thread_in_native to _thread_in_native_trans. 2336 __ li(R0, _thread_in_native_trans); 2337 __ release(); 2338 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2339 __ stw(R0, thread_(thread_state)); 2340 2341 2342 // Must we block? 2343 // -------------------------------------------------------------------------- 2344 2345 // Block, if necessary, before resuming in _thread_in_Java state. 2346 // In order for GC to work, don't clear the last_Java_sp until after blocking. 2347 Label after_transition; 2348 { 2349 Label no_block, sync; 2350 2351 if (os::is_MP()) { 2352 if (UseMembar) { 2353 // Force this write out before the read below. 2354 __ fence(); 2355 } else { 2356 // Write serialization page so VM thread can do a pseudo remote membar. 2357 // We use the current thread pointer to calculate a thread specific 2358 // offset to write to within the page. This minimizes bus traffic 2359 // due to cache line collision. 2360 __ serialize_memory(R16_thread, r_temp_4, r_temp_5); 2361 } 2362 } 2363 2364 Register sync_state_addr = r_temp_4; 2365 Register sync_state = r_temp_5; 2366 Register suspend_flags = r_temp_6; 2367 2368 __ load_const(sync_state_addr, SafepointSynchronize::address_of_state(), /*temp*/ sync_state); 2369 2370 // TODO: PPC port assert(4 == SafepointSynchronize::sz_state(), "unexpected field size"); 2371 __ lwz(sync_state, 0, sync_state_addr); 2372 2373 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size"); 2374 __ lwz(suspend_flags, thread_(suspend_flags)); 2375 2376 __ acquire(); 2377 2378 Label do_safepoint; 2379 // No synchronization in progress nor yet synchronized. 2380 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized); 2381 // Not suspended. 2382 __ cmpwi(CCR1, suspend_flags, 0); 2383 2384 __ bne(CCR0, sync); 2385 __ beq(CCR1, no_block); 2386 2387 // Block. Save any potential method result value before the operation and 2388 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this 2389 // lets us share the oopMap we used when we went native rather than create 2390 // a distinct one for this pc. 2391 __ bind(sync); 2392 2393 address entry_point = is_critical_native 2394 ? CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition) 2395 : CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans); 2396 save_native_result(masm, ret_type, workspace_slot_offset); 2397 __ call_VM_leaf(entry_point, R16_thread); 2398 restore_native_result(masm, ret_type, workspace_slot_offset); 2399 2400 if (is_critical_native) { 2401 __ b(after_transition); // No thread state transition here. 2402 } 2403 __ bind(no_block); 2404 } 2405 2406 // Publish thread state. 2407 // -------------------------------------------------------------------------- 2408 2409 // Thread state is thread_in_native_trans. Any safepoint blocking has 2410 // already happened so we can now change state to _thread_in_Java. 2411 2412 // Transition from _thread_in_native_trans to _thread_in_Java. 2413 __ li(R0, _thread_in_Java); 2414 __ release(); 2415 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2416 __ stw(R0, thread_(thread_state)); 2417 if (UseMembar) { 2418 __ fence(); 2419 } 2420 __ bind(after_transition); 2421 2422 // Reguard any pages if necessary. 2423 // -------------------------------------------------------------------------- 2424 2425 Label no_reguard; 2426 __ lwz(r_temp_1, thread_(stack_guard_state)); 2427 __ cmpwi(CCR0, r_temp_1, JavaThread::stack_guard_yellow_disabled); 2428 __ bne(CCR0, no_reguard); 2429 2430 save_native_result(masm, ret_type, workspace_slot_offset); 2431 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); 2432 restore_native_result(masm, ret_type, workspace_slot_offset); 2433 2434 __ bind(no_reguard); 2435 2436 2437 // Unlock 2438 // -------------------------------------------------------------------------- 2439 2440 if (method->is_synchronized()) { 2441 2442 ConditionRegister r_flag = CCR1; 2443 const Register r_oop = r_temp_4; 2444 const Register r_box = r_temp_5; 2445 const Register r_exception = r_temp_6; 2446 Label done; 2447 2448 // Get oop and address of lock object box. 2449 if (method_is_static) { 2450 assert(klass_offset != -1, ""); 2451 __ ld(r_oop, klass_offset, R1_SP); 2452 } else { 2453 assert(receiver_offset != -1, ""); 2454 __ ld(r_oop, receiver_offset, R1_SP); 2455 } 2456 __ addi(r_box, R1_SP, lock_offset); 2457 2458 // Try fastpath for unlocking. 2459 __ compiler_fast_unlock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3); 2460 __ beq(r_flag, done); 2461 2462 // Save and restore any potential method result value around the unlocking operation. 2463 save_native_result(masm, ret_type, workspace_slot_offset); 2464 2465 // Must save pending exception around the slow-path VM call. Since it's a 2466 // leaf call, the pending exception (if any) can be kept in a register. 2467 __ ld(r_exception, thread_(pending_exception)); 2468 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile"); 2469 __ li(R0, 0); 2470 __ std(R0, thread_(pending_exception)); 2471 2472 // Slow case of monitor enter. 2473 // Inline a special case of call_VM that disallows any pending_exception. 2474 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box); 2475 2476 __ asm_assert_mem8_is_zero(thread_(pending_exception), 2477 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C", 0); 2478 2479 restore_native_result(masm, ret_type, workspace_slot_offset); 2480 2481 // Check_forward_pending_exception jump to forward_exception if any pending 2482 // exception is set. The forward_exception routine expects to see the 2483 // exception in pending_exception and not in a register. Kind of clumsy, 2484 // since all folks who branch to forward_exception must have tested 2485 // pending_exception first and hence have it in a register already. 2486 __ std(r_exception, thread_(pending_exception)); 2487 2488 __ bind(done); 2489 } 2490 2491 # if 0 2492 // DTrace method exit 2493 # endif 2494 2495 // Clear "last Java frame" SP and PC. 2496 // -------------------------------------------------------------------------- 2497 2498 __ reset_last_Java_frame(); 2499 2500 // Unpack oop result. 2501 // -------------------------------------------------------------------------- 2502 2503 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2504 Label skip_unboxing; 2505 __ cmpdi(CCR0, R3_RET, 0); 2506 __ beq(CCR0, skip_unboxing); 2507 __ ld(R3_RET, 0, R3_RET); 2508 __ bind(skip_unboxing); 2509 __ verify_oop(R3_RET); 2510 } 2511 2512 2513 // Reset handle block. 2514 // -------------------------------------------------------------------------- 2515 if (!is_critical_native) { 2516 __ ld(r_temp_1, thread_(active_handles)); 2517 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size"); 2518 __ li(r_temp_2, 0); 2519 __ stw(r_temp_2, JNIHandleBlock::top_offset_in_bytes(), r_temp_1); 2520 2521 2522 // Check for pending exceptions. 2523 // -------------------------------------------------------------------------- 2524 __ ld(r_temp_2, thread_(pending_exception)); 2525 __ cmpdi(CCR0, r_temp_2, 0); 2526 __ bne(CCR0, handle_pending_exception); 2527 } 2528 2529 // Return 2530 // -------------------------------------------------------------------------- 2531 2532 __ pop_frame(); 2533 __ restore_LR_CR(R11); 2534 __ blr(); 2535 2536 2537 // Handler for pending exceptions (out-of-line). 2538 // -------------------------------------------------------------------------- 2539 2540 // Since this is a native call, we know the proper exception handler 2541 // is the empty function. We just pop this frame and then jump to 2542 // forward_exception_entry. 2543 if (!is_critical_native) { 2544 __ align(InteriorEntryAlignment); 2545 __ bind(handle_pending_exception); 2546 2547 __ pop_frame(); 2548 __ restore_LR_CR(R11); 2549 __ b64_patchable((address)StubRoutines::forward_exception_entry(), 2550 relocInfo::runtime_call_type); 2551 } 2552 2553 // Handler for a cache miss (out-of-line). 2554 // -------------------------------------------------------------------------- 2555 2556 if (!method_is_static) { 2557 __ align(InteriorEntryAlignment); 2558 __ bind(ic_miss); 2559 2560 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 2561 relocInfo::runtime_call_type); 2562 } 2563 2564 // Done. 2565 // -------------------------------------------------------------------------- 2566 2567 __ flush(); 2568 2569 nmethod *nm = nmethod::new_native_nmethod(method, 2570 compile_id, 2571 masm->code(), 2572 vep_start_pc-start_pc, 2573 frame_done_pc-start_pc, 2574 stack_slots / VMRegImpl::slots_per_word, 2575 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2576 in_ByteSize(lock_offset), 2577 oop_maps); 2578 2579 if (is_critical_native) { 2580 nm->set_lazy_critical_native(true); 2581 } 2582 2583 return nm; 2584 #else 2585 ShouldNotReachHere(); 2586 return NULL; 2587 #endif // COMPILER2 2588 } 2589 2590 // This function returns the adjust size (in number of words) to a c2i adapter 2591 // activation for use during deoptimization. 2592 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2593 return round_to((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::alignment_in_bytes); 2594 } 2595 2596 uint SharedRuntime::out_preserve_stack_slots() { 2597 #ifdef COMPILER2 2598 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size; 2599 #else 2600 return 0; 2601 #endif 2602 } 2603 2604 #ifdef COMPILER2 2605 // Frame generation for deopt and uncommon trap blobs. 2606 static void push_skeleton_frame(MacroAssembler* masm, bool deopt, 2607 /* Read */ 2608 Register unroll_block_reg, 2609 /* Update */ 2610 Register frame_sizes_reg, 2611 Register number_of_frames_reg, 2612 Register pcs_reg, 2613 /* Invalidate */ 2614 Register frame_size_reg, 2615 Register pc_reg) { 2616 2617 __ ld(pc_reg, 0, pcs_reg); 2618 __ ld(frame_size_reg, 0, frame_sizes_reg); 2619 __ std(pc_reg, _abi(lr), R1_SP); 2620 __ push_frame(frame_size_reg, R0/*tmp*/); 2621 #ifdef CC_INTERP 2622 __ std(R1_SP, _parent_ijava_frame_abi(initial_caller_sp), R1_SP); 2623 #else 2624 #ifdef ASSERT 2625 __ load_const_optimized(pc_reg, 0x5afe); 2626 __ std(pc_reg, _ijava_state_neg(ijava_reserved), R1_SP); 2627 #endif 2628 __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP); 2629 #endif // CC_INTERP 2630 __ addi(number_of_frames_reg, number_of_frames_reg, -1); 2631 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize); 2632 __ addi(pcs_reg, pcs_reg, wordSize); 2633 } 2634 2635 // Loop through the UnrollBlock info and create new frames. 2636 static void push_skeleton_frames(MacroAssembler* masm, bool deopt, 2637 /* read */ 2638 Register unroll_block_reg, 2639 /* invalidate */ 2640 Register frame_sizes_reg, 2641 Register number_of_frames_reg, 2642 Register pcs_reg, 2643 Register frame_size_reg, 2644 Register pc_reg) { 2645 Label loop; 2646 2647 // _number_of_frames is of type int (deoptimization.hpp) 2648 __ lwa(number_of_frames_reg, 2649 Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(), 2650 unroll_block_reg); 2651 __ ld(pcs_reg, 2652 Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(), 2653 unroll_block_reg); 2654 __ ld(frame_sizes_reg, 2655 Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(), 2656 unroll_block_reg); 2657 2658 // stack: (caller_of_deoptee, ...). 2659 2660 // At this point we either have an interpreter frame or a compiled 2661 // frame on top of stack. If it is a compiled frame we push a new c2i 2662 // adapter here 2663 2664 // Memorize top-frame stack-pointer. 2665 __ mr(frame_size_reg/*old_sp*/, R1_SP); 2666 2667 // Resize interpreter top frame OR C2I adapter. 2668 2669 // At this moment, the top frame (which is the caller of the deoptee) is 2670 // an interpreter frame or a newly pushed C2I adapter or an entry frame. 2671 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the 2672 // outgoing arguments. 2673 // 2674 // In order to push the interpreter frame for the deoptee, we need to 2675 // resize the top frame such that we are able to place the deoptee's 2676 // locals in the frame. 2677 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI 2678 // into a valid PARENT_IJAVA_FRAME_ABI. 2679 2680 __ lwa(R11_scratch1, 2681 Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(), 2682 unroll_block_reg); 2683 __ neg(R11_scratch1, R11_scratch1); 2684 2685 // R11_scratch1 contains size of locals for frame resizing. 2686 // R12_scratch2 contains top frame's lr. 2687 2688 // Resize frame by complete frame size prevents TOC from being 2689 // overwritten by locals. A more stack space saving way would be 2690 // to copy the TOC to its location in the new abi. 2691 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size); 2692 2693 // now, resize the frame 2694 __ resize_frame(R11_scratch1, pc_reg/*tmp*/); 2695 2696 // In the case where we have resized a c2i frame above, the optional 2697 // alignment below the locals has size 32 (why?). 2698 __ std(R12_scratch2, _abi(lr), R1_SP); 2699 2700 // Initialize initial_caller_sp. 2701 #ifdef CC_INTERP 2702 __ std(frame_size_reg/*old_sp*/, _parent_ijava_frame_abi(initial_caller_sp), R1_SP); 2703 #else 2704 #ifdef ASSERT 2705 __ load_const_optimized(pc_reg, 0x5afe); 2706 __ std(pc_reg, _ijava_state_neg(ijava_reserved), R1_SP); 2707 #endif 2708 __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP); 2709 #endif // CC_INTERP 2710 2711 #ifdef ASSERT 2712 // Make sure that there is at least one entry in the array. 2713 __ cmpdi(CCR0, number_of_frames_reg, 0); 2714 __ asm_assert_ne("array_size must be > 0", 0x205); 2715 #endif 2716 2717 // Now push the new interpreter frames. 2718 // 2719 __ bind(loop); 2720 // Allocate a new frame, fill in the pc. 2721 push_skeleton_frame(masm, deopt, 2722 unroll_block_reg, 2723 frame_sizes_reg, 2724 number_of_frames_reg, 2725 pcs_reg, 2726 frame_size_reg, 2727 pc_reg); 2728 __ cmpdi(CCR0, number_of_frames_reg, 0); 2729 __ bne(CCR0, loop); 2730 2731 // Get the return address pointing into the frame manager. 2732 __ ld(R0, 0, pcs_reg); 2733 // Store it in the top interpreter frame. 2734 __ std(R0, _abi(lr), R1_SP); 2735 // Initialize frame_manager_lr of interpreter top frame. 2736 #ifdef CC_INTERP 2737 __ std(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 2738 #endif 2739 } 2740 #endif 2741 2742 void SharedRuntime::generate_deopt_blob() { 2743 // Allocate space for the code 2744 ResourceMark rm; 2745 // Setup code generation tools 2746 CodeBuffer buffer("deopt_blob", 2048, 1024); 2747 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer); 2748 Label exec_mode_initialized; 2749 int frame_size_in_words; 2750 OopMap* map = NULL; 2751 OopMapSet *oop_maps = new OopMapSet(); 2752 2753 // size of ABI112 plus spill slots for R3_RET and F1_RET. 2754 const int frame_size_in_bytes = frame::abi_reg_args_spill_size; 2755 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint); 2756 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info. 2757 2758 const Register exec_mode_reg = R21_tmp1; 2759 2760 const address start = __ pc(); 2761 2762 #ifdef COMPILER2 2763 // -------------------------------------------------------------------------- 2764 // Prolog for non exception case! 2765 2766 // We have been called from the deopt handler of the deoptee. 2767 // 2768 // deoptee: 2769 // ... 2770 // call X 2771 // ... 2772 // deopt_handler: call_deopt_stub 2773 // cur. return pc --> ... 2774 // 2775 // So currently SR_LR points behind the call in the deopt handler. 2776 // We adjust it such that it points to the start of the deopt handler. 2777 // The return_pc has been stored in the frame of the deoptee and 2778 // will replace the address of the deopt_handler in the call 2779 // to Deoptimization::fetch_unroll_info below. 2780 // We can't grab a free register here, because all registers may 2781 // contain live values, so let the RegisterSaver do the adjustment 2782 // of the return pc. 2783 const int return_pc_adjustment_no_exception = -HandlerImpl::size_deopt_handler(); 2784 2785 // Push the "unpack frame" 2786 // Save everything in sight. 2787 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 2788 &first_frame_size_in_bytes, 2789 /*generate_oop_map=*/ true, 2790 return_pc_adjustment_no_exception, 2791 RegisterSaver::return_pc_is_lr); 2792 assert(map != NULL, "OopMap must have been created"); 2793 2794 __ li(exec_mode_reg, Deoptimization::Unpack_deopt); 2795 // Save exec mode for unpack_frames. 2796 __ b(exec_mode_initialized); 2797 2798 // -------------------------------------------------------------------------- 2799 // Prolog for exception case 2800 2801 // An exception is pending. 2802 // We have been called with a return (interpreter) or a jump (exception blob). 2803 // 2804 // - R3_ARG1: exception oop 2805 // - R4_ARG2: exception pc 2806 2807 int exception_offset = __ pc() - start; 2808 2809 BLOCK_COMMENT("Prolog for exception case"); 2810 2811 // The RegisterSaves doesn't need to adjust the return pc for this situation. 2812 const int return_pc_adjustment_exception = 0; 2813 2814 // Push the "unpack frame". 2815 // Save everything in sight. 2816 assert(R4 == R4_ARG2, "exception pc must be in r4"); 2817 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 2818 &first_frame_size_in_bytes, 2819 /*generate_oop_map=*/ false, 2820 return_pc_adjustment_exception, 2821 RegisterSaver::return_pc_is_r4); 2822 2823 // Deopt during an exception. Save exec mode for unpack_frames. 2824 __ li(exec_mode_reg, Deoptimization::Unpack_exception); 2825 2826 // Store exception oop and pc in thread (location known to GC). 2827 // This is needed since the call to "fetch_unroll_info()" may safepoint. 2828 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2829 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2830 2831 // fall through 2832 2833 // -------------------------------------------------------------------------- 2834 __ BIND(exec_mode_initialized); 2835 2836 { 2837 const Register unroll_block_reg = R22_tmp2; 2838 2839 // We need to set `last_Java_frame' because `fetch_unroll_info' will 2840 // call `last_Java_frame()'. The value of the pc in the frame is not 2841 // particularly important. It just needs to identify this blob. 2842 __ set_last_Java_frame(R1_SP, noreg); 2843 2844 // With EscapeAnalysis turned on, this call may safepoint! 2845 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread); 2846 address calls_return_pc = __ last_calls_return_pc(); 2847 // Set an oopmap for the call site that describes all our saved registers. 2848 oop_maps->add_gc_map(calls_return_pc - start, map); 2849 2850 __ reset_last_Java_frame(); 2851 // Save the return value. 2852 __ mr(unroll_block_reg, R3_RET); 2853 2854 // Restore only the result registers that have been saved 2855 // by save_volatile_registers(...). 2856 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes); 2857 2858 // In excp_deopt_mode, restore and clear exception oop which we 2859 // stored in the thread during exception entry above. The exception 2860 // oop will be the return value of this stub. 2861 Label skip_restore_excp; 2862 __ cmpdi(CCR0, exec_mode_reg, Deoptimization::Unpack_exception); 2863 __ bne(CCR0, skip_restore_excp); 2864 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2865 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2866 __ li(R0, 0); 2867 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2868 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2869 __ BIND(skip_restore_excp); 2870 2871 // reload narrro_oop_base 2872 if (UseCompressedOops && Universe::narrow_oop_base() != 0) { 2873 __ load_const_optimized(R30, Universe::narrow_oop_base()); 2874 } 2875 2876 __ pop_frame(); 2877 2878 // stack: (deoptee, optional i2c, caller of deoptee, ...). 2879 2880 // pop the deoptee's frame 2881 __ pop_frame(); 2882 2883 // stack: (caller_of_deoptee, ...). 2884 2885 // Loop through the `UnrollBlock' info and create interpreter frames. 2886 push_skeleton_frames(masm, true/*deopt*/, 2887 unroll_block_reg, 2888 R23_tmp3, 2889 R24_tmp4, 2890 R25_tmp5, 2891 R26_tmp6, 2892 R27_tmp7); 2893 2894 // stack: (skeletal interpreter frame, ..., optional skeletal 2895 // interpreter frame, optional c2i, caller of deoptee, ...). 2896 } 2897 2898 // push an `unpack_frame' taking care of float / int return values. 2899 __ push_frame(frame_size_in_bytes, R0/*tmp*/); 2900 2901 // stack: (unpack frame, skeletal interpreter frame, ..., optional 2902 // skeletal interpreter frame, optional c2i, caller of deoptee, 2903 // ...). 2904 2905 // Spill live volatile registers since we'll do a call. 2906 __ std( R3_RET, _abi_reg_args_spill(spill_ret), R1_SP); 2907 __ stfd(F1_RET, _abi_reg_args_spill(spill_fret), R1_SP); 2908 2909 // Let the unpacker layout information in the skeletal frames just 2910 // allocated. 2911 __ get_PC_trash_LR(R3_RET); 2912 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET); 2913 // This is a call to a LEAF method, so no oop map is required. 2914 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), 2915 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/); 2916 __ reset_last_Java_frame(); 2917 2918 // Restore the volatiles saved above. 2919 __ ld( R3_RET, _abi_reg_args_spill(spill_ret), R1_SP); 2920 __ lfd(F1_RET, _abi_reg_args_spill(spill_fret), R1_SP); 2921 2922 // Pop the unpack frame. 2923 __ pop_frame(); 2924 __ restore_LR_CR(R0); 2925 2926 // stack: (top interpreter frame, ..., optional interpreter frame, 2927 // optional c2i, caller of deoptee, ...). 2928 2929 // Initialize R14_state. 2930 #ifdef CC_INTERP 2931 __ ld(R14_state, 0, R1_SP); 2932 __ addi(R14_state, R14_state, -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 2933 // Also inititialize R15_prev_state. 2934 __ restore_prev_state(); 2935 #else 2936 __ restore_interpreter_state(R11_scratch1); 2937 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 2938 #endif // CC_INTERP 2939 2940 2941 // Return to the interpreter entry point. 2942 __ blr(); 2943 __ flush(); 2944 #else // COMPILER2 2945 __ unimplemented("deopt blob needed only with compiler"); 2946 int exception_offset = __ pc() - start; 2947 #endif // COMPILER2 2948 2949 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, 0, first_frame_size_in_bytes / wordSize); 2950 } 2951 2952 #ifdef COMPILER2 2953 void SharedRuntime::generate_uncommon_trap_blob() { 2954 // Allocate space for the code. 2955 ResourceMark rm; 2956 // Setup code generation tools. 2957 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2958 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer); 2959 address start = __ pc(); 2960 2961 Register unroll_block_reg = R21_tmp1; 2962 Register klass_index_reg = R22_tmp2; 2963 Register unc_trap_reg = R23_tmp3; 2964 2965 OopMapSet* oop_maps = new OopMapSet(); 2966 int frame_size_in_bytes = frame::abi_reg_args_size; 2967 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0); 2968 2969 // stack: (deoptee, optional i2c, caller_of_deoptee, ...). 2970 2971 // Push a dummy `unpack_frame' and call 2972 // `Deoptimization::uncommon_trap' to pack the compiled frame into a 2973 // vframe array and return the `UnrollBlock' information. 2974 2975 // Save LR to compiled frame. 2976 __ save_LR_CR(R11_scratch1); 2977 2978 // Push an "uncommon_trap" frame. 2979 __ push_frame_reg_args(0, R11_scratch1); 2980 2981 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...). 2982 2983 // Set the `unpack_frame' as last_Java_frame. 2984 // `Deoptimization::uncommon_trap' expects it and considers its 2985 // sender frame as the deoptee frame. 2986 // Remember the offset of the instruction whose address will be 2987 // moved to R11_scratch1. 2988 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1); 2989 2990 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 2991 2992 __ mr(klass_index_reg, R3); 2993 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), 2994 R16_thread, klass_index_reg); 2995 2996 // Set an oopmap for the call site. 2997 oop_maps->add_gc_map(gc_map_pc - start, map); 2998 2999 __ reset_last_Java_frame(); 3000 3001 // Pop the `unpack frame'. 3002 __ pop_frame(); 3003 3004 // stack: (deoptee, optional i2c, caller_of_deoptee, ...). 3005 3006 // Save the return value. 3007 __ mr(unroll_block_reg, R3_RET); 3008 3009 // Pop the uncommon_trap frame. 3010 __ pop_frame(); 3011 3012 // stack: (caller_of_deoptee, ...). 3013 3014 // Allocate new interpreter frame(s) and possibly a c2i adapter 3015 // frame. 3016 push_skeleton_frames(masm, false/*deopt*/, 3017 unroll_block_reg, 3018 R22_tmp2, 3019 R23_tmp3, 3020 R24_tmp4, 3021 R25_tmp5, 3022 R26_tmp6); 3023 3024 // stack: (skeletal interpreter frame, ..., optional skeletal 3025 // interpreter frame, optional c2i, caller of deoptee, ...). 3026 3027 // Push a dummy `unpack_frame' taking care of float return values. 3028 // Call `Deoptimization::unpack_frames' to layout information in the 3029 // interpreter frames just created. 3030 3031 // Push a simple "unpack frame" here. 3032 __ push_frame_reg_args(0, R11_scratch1); 3033 3034 // stack: (unpack frame, skeletal interpreter frame, ..., optional 3035 // skeletal interpreter frame, optional c2i, caller of deoptee, 3036 // ...). 3037 3038 // Set the "unpack_frame" as last_Java_frame. 3039 __ get_PC_trash_LR(R11_scratch1); 3040 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 3041 3042 // Indicate it is the uncommon trap case. 3043 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap); 3044 // Let the unpacker layout information in the skeletal frames just 3045 // allocated. 3046 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), 3047 R16_thread, unc_trap_reg); 3048 3049 __ reset_last_Java_frame(); 3050 // Pop the `unpack frame'. 3051 __ pop_frame(); 3052 // Restore LR from top interpreter frame. 3053 __ restore_LR_CR(R11_scratch1); 3054 3055 // stack: (top interpreter frame, ..., optional interpreter frame, 3056 // optional c2i, caller of deoptee, ...). 3057 3058 #ifdef CC_INTERP 3059 // Initialize R14_state, ... 3060 __ ld(R11_scratch1, 0, R1_SP); 3061 __ addi(R14_state, R11_scratch1, -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 3062 // also initialize R15_prev_state. 3063 __ restore_prev_state(); 3064 #else 3065 __ restore_interpreter_state(R11_scratch1); 3066 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 3067 #endif // CC_INTERP 3068 3069 // Return to the interpreter entry point. 3070 __ blr(); 3071 3072 masm->flush(); 3073 3074 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize); 3075 } 3076 #endif // COMPILER2 3077 3078 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap. 3079 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 3080 assert(StubRoutines::forward_exception_entry() != NULL, 3081 "must be generated before"); 3082 3083 ResourceMark rm; 3084 OopMapSet *oop_maps = new OopMapSet(); 3085 OopMap* map; 3086 3087 // Allocate space for the code. Setup code generation tools. 3088 CodeBuffer buffer("handler_blob", 2048, 1024); 3089 MacroAssembler* masm = new MacroAssembler(&buffer); 3090 3091 address start = __ pc(); 3092 int frame_size_in_bytes = 0; 3093 3094 RegisterSaver::ReturnPCLocation return_pc_location; 3095 bool cause_return = (poll_type == POLL_AT_RETURN); 3096 if (cause_return) { 3097 // Nothing to do here. The frame has already been popped in MachEpilogNode. 3098 // Register LR already contains the return pc. 3099 return_pc_location = RegisterSaver::return_pc_is_lr; 3100 } else { 3101 // Use thread()->saved_exception_pc() as return pc. 3102 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc; 3103 } 3104 3105 // Save registers, fpu state, and flags. 3106 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 3107 &frame_size_in_bytes, 3108 /*generate_oop_map=*/ true, 3109 /*return_pc_adjustment=*/0, 3110 return_pc_location); 3111 3112 // The following is basically a call_VM. However, we need the precise 3113 // address of the call in order to generate an oopmap. Hence, we do all the 3114 // work outselves. 3115 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg); 3116 3117 // The return address must always be correct so that the frame constructor 3118 // never sees an invalid pc. 3119 3120 // Do the call 3121 __ call_VM_leaf(call_ptr, R16_thread); 3122 address calls_return_pc = __ last_calls_return_pc(); 3123 3124 // Set an oopmap for the call site. This oopmap will map all 3125 // oop-registers and debug-info registers as callee-saved. This 3126 // will allow deoptimization at this safepoint to find all possible 3127 // debug-info recordings, as well as let GC find all oops. 3128 oop_maps->add_gc_map(calls_return_pc - start, map); 3129 3130 Label noException; 3131 3132 // Clear the last Java frame. 3133 __ reset_last_Java_frame(); 3134 3135 BLOCK_COMMENT(" Check pending exception."); 3136 const Register pending_exception = R0; 3137 __ ld(pending_exception, thread_(pending_exception)); 3138 __ cmpdi(CCR0, pending_exception, 0); 3139 __ beq(CCR0, noException); 3140 3141 // Exception pending 3142 RegisterSaver::restore_live_registers_and_pop_frame(masm, 3143 frame_size_in_bytes, 3144 /*restore_ctr=*/true); 3145 3146 BLOCK_COMMENT(" Jump to forward_exception_entry."); 3147 // Jump to forward_exception_entry, with the issuing PC in LR 3148 // so it looks like the original nmethod called forward_exception_entry. 3149 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 3150 3151 // No exception case. 3152 __ BIND(noException); 3153 3154 3155 // Normal exit, restore registers and exit. 3156 RegisterSaver::restore_live_registers_and_pop_frame(masm, 3157 frame_size_in_bytes, 3158 /*restore_ctr=*/true); 3159 3160 __ blr(); 3161 3162 // Make sure all code is generated 3163 masm->flush(); 3164 3165 // Fill-out other meta info 3166 // CodeBlob frame size is in words. 3167 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize); 3168 } 3169 3170 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss) 3171 // 3172 // Generate a stub that calls into the vm to find out the proper destination 3173 // of a java call. All the argument registers are live at this point 3174 // but since this is generic code we don't know what they are and the caller 3175 // must do any gc of the args. 3176 // 3177 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3178 3179 // allocate space for the code 3180 ResourceMark rm; 3181 3182 CodeBuffer buffer(name, 1000, 512); 3183 MacroAssembler* masm = new MacroAssembler(&buffer); 3184 3185 int frame_size_in_bytes; 3186 3187 OopMapSet *oop_maps = new OopMapSet(); 3188 OopMap* map = NULL; 3189 3190 address start = __ pc(); 3191 3192 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 3193 &frame_size_in_bytes, 3194 /*generate_oop_map*/ true, 3195 /*return_pc_adjustment*/ 0, 3196 RegisterSaver::return_pc_is_lr); 3197 3198 // Use noreg as last_Java_pc, the return pc will be reconstructed 3199 // from the physical frame. 3200 __ set_last_Java_frame(/*sp*/R1_SP, noreg); 3201 3202 int frame_complete = __ offset(); 3203 3204 // Pass R19_method as 2nd (optional) argument, used by 3205 // counter_overflow_stub. 3206 __ call_VM_leaf(destination, R16_thread, R19_method); 3207 address calls_return_pc = __ last_calls_return_pc(); 3208 // Set an oopmap for the call site. 3209 // We need this not only for callee-saved registers, but also for volatile 3210 // registers that the compiler might be keeping live across a safepoint. 3211 // Create the oopmap for the call's return pc. 3212 oop_maps->add_gc_map(calls_return_pc - start, map); 3213 3214 // R3_RET contains the address we are going to jump to assuming no exception got installed. 3215 3216 // clear last_Java_sp 3217 __ reset_last_Java_frame(); 3218 3219 // Check for pending exceptions. 3220 BLOCK_COMMENT("Check for pending exceptions."); 3221 Label pending; 3222 __ ld(R11_scratch1, thread_(pending_exception)); 3223 __ cmpdi(CCR0, R11_scratch1, 0); 3224 __ bne(CCR0, pending); 3225 3226 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame. 3227 3228 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false); 3229 3230 // Get the returned method. 3231 __ get_vm_result_2(R19_method); 3232 3233 __ bctr(); 3234 3235 3236 // Pending exception after the safepoint. 3237 __ BIND(pending); 3238 3239 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true); 3240 3241 // exception pending => remove activation and forward to exception handler 3242 3243 __ li(R11_scratch1, 0); 3244 __ ld(R3_ARG1, thread_(pending_exception)); 3245 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_offset()), R16_thread); 3246 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 3247 3248 // ------------- 3249 // Make sure all code is generated. 3250 masm->flush(); 3251 3252 // return the blob 3253 // frame_size_words or bytes?? 3254 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize, 3255 oop_maps, true); 3256 }