1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * Copyright 2012, 2015 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 "frame_ppc.hpp" 32 #include "interpreter/interpreter.hpp" 33 #include "interpreter/interp_masm.hpp" 34 #include "oops/compiledICHolder.hpp" 35 #include "prims/jvmtiRedefineClassesTrace.hpp" 36 #include "runtime/sharedRuntime.hpp" 37 #include "runtime/vframeArray.hpp" 38 #include "vmreg_ppc.inline.hpp" 39 #ifdef COMPILER1 40 #include "c1/c1_Runtime1.hpp" 41 #endif 42 #ifdef COMPILER2 43 #include "adfiles/ad_ppc_64.hpp" 44 #include "opto/runtime.hpp" 45 #endif 46 47 #define __ masm-> 48 49 #ifdef PRODUCT 50 #define BLOCK_COMMENT(str) // nothing 51 #else 52 #define BLOCK_COMMENT(str) __ block_comment(str) 53 #endif 54 55 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") 56 57 58 class RegisterSaver { 59 // Used for saving volatile registers. 60 public: 61 62 // Support different return pc locations. 63 enum ReturnPCLocation { 64 return_pc_is_lr, 65 return_pc_is_r4, 66 return_pc_is_thread_saved_exception_pc 67 }; 68 69 static OopMap* push_frame_reg_args_and_save_live_registers(MacroAssembler* masm, 70 int* out_frame_size_in_bytes, 71 bool generate_oop_map, 72 int return_pc_adjustment, 73 ReturnPCLocation return_pc_location); 74 static void restore_live_registers_and_pop_frame(MacroAssembler* masm, 75 int frame_size_in_bytes, 76 bool restore_ctr); 77 78 static void push_frame_and_save_argument_registers(MacroAssembler* masm, 79 Register r_temp, 80 int frame_size, 81 int total_args, 82 const VMRegPair *regs, const VMRegPair *regs2 = NULL); 83 static void restore_argument_registers_and_pop_frame(MacroAssembler*masm, 84 int frame_size, 85 int total_args, 86 const VMRegPair *regs, const VMRegPair *regs2 = NULL); 87 88 // During deoptimization only the result registers need to be restored 89 // all the other values have already been extracted. 90 static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes); 91 92 // Constants and data structures: 93 94 typedef enum { 95 int_reg = 0, 96 float_reg = 1, 97 special_reg = 2 98 } RegisterType; 99 100 typedef enum { 101 reg_size = 8, 102 half_reg_size = reg_size / 2, 103 } RegisterConstants; 104 105 typedef struct { 106 RegisterType reg_type; 107 int reg_num; 108 VMReg vmreg; 109 } LiveRegType; 110 }; 111 112 113 #define RegisterSaver_LiveSpecialReg(regname) \ 114 { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() } 115 116 #define RegisterSaver_LiveIntReg(regname) \ 117 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() } 118 119 #define RegisterSaver_LiveFloatReg(regname) \ 120 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() } 121 122 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = { 123 // Live registers which get spilled to the stack. Register 124 // positions in this array correspond directly to the stack layout. 125 126 // 127 // live special registers: 128 // 129 RegisterSaver_LiveSpecialReg(SR_CTR), 130 // 131 // live float registers: 132 // 133 RegisterSaver_LiveFloatReg( F0 ), 134 RegisterSaver_LiveFloatReg( F1 ), 135 RegisterSaver_LiveFloatReg( F2 ), 136 RegisterSaver_LiveFloatReg( F3 ), 137 RegisterSaver_LiveFloatReg( F4 ), 138 RegisterSaver_LiveFloatReg( F5 ), 139 RegisterSaver_LiveFloatReg( F6 ), 140 RegisterSaver_LiveFloatReg( F7 ), 141 RegisterSaver_LiveFloatReg( F8 ), 142 RegisterSaver_LiveFloatReg( F9 ), 143 RegisterSaver_LiveFloatReg( F10 ), 144 RegisterSaver_LiveFloatReg( F11 ), 145 RegisterSaver_LiveFloatReg( F12 ), 146 RegisterSaver_LiveFloatReg( F13 ), 147 RegisterSaver_LiveFloatReg( F14 ), 148 RegisterSaver_LiveFloatReg( F15 ), 149 RegisterSaver_LiveFloatReg( F16 ), 150 RegisterSaver_LiveFloatReg( F17 ), 151 RegisterSaver_LiveFloatReg( F18 ), 152 RegisterSaver_LiveFloatReg( F19 ), 153 RegisterSaver_LiveFloatReg( F20 ), 154 RegisterSaver_LiveFloatReg( F21 ), 155 RegisterSaver_LiveFloatReg( F22 ), 156 RegisterSaver_LiveFloatReg( F23 ), 157 RegisterSaver_LiveFloatReg( F24 ), 158 RegisterSaver_LiveFloatReg( F25 ), 159 RegisterSaver_LiveFloatReg( F26 ), 160 RegisterSaver_LiveFloatReg( F27 ), 161 RegisterSaver_LiveFloatReg( F28 ), 162 RegisterSaver_LiveFloatReg( F29 ), 163 RegisterSaver_LiveFloatReg( F30 ), 164 RegisterSaver_LiveFloatReg( F31 ), 165 // 166 // live integer registers: 167 // 168 RegisterSaver_LiveIntReg( R0 ), 169 //RegisterSaver_LiveIntReg( R1 ), // stack pointer 170 RegisterSaver_LiveIntReg( R2 ), 171 RegisterSaver_LiveIntReg( R3 ), 172 RegisterSaver_LiveIntReg( R4 ), 173 RegisterSaver_LiveIntReg( R5 ), 174 RegisterSaver_LiveIntReg( R6 ), 175 RegisterSaver_LiveIntReg( R7 ), 176 RegisterSaver_LiveIntReg( R8 ), 177 RegisterSaver_LiveIntReg( R9 ), 178 RegisterSaver_LiveIntReg( R10 ), 179 RegisterSaver_LiveIntReg( R11 ), 180 RegisterSaver_LiveIntReg( R12 ), 181 //RegisterSaver_LiveIntReg( R13 ), // system thread id 182 RegisterSaver_LiveIntReg( R14 ), 183 RegisterSaver_LiveIntReg( R15 ), 184 RegisterSaver_LiveIntReg( R16 ), 185 RegisterSaver_LiveIntReg( R17 ), 186 RegisterSaver_LiveIntReg( R18 ), 187 RegisterSaver_LiveIntReg( R19 ), 188 RegisterSaver_LiveIntReg( R20 ), 189 RegisterSaver_LiveIntReg( R21 ), 190 RegisterSaver_LiveIntReg( R22 ), 191 RegisterSaver_LiveIntReg( R23 ), 192 RegisterSaver_LiveIntReg( R24 ), 193 RegisterSaver_LiveIntReg( R25 ), 194 RegisterSaver_LiveIntReg( R26 ), 195 RegisterSaver_LiveIntReg( R27 ), 196 RegisterSaver_LiveIntReg( R28 ), 197 RegisterSaver_LiveIntReg( R29 ), 198 RegisterSaver_LiveIntReg( R30 ), 199 RegisterSaver_LiveIntReg( R31 ), // must be the last register (see save/restore functions below) 200 }; 201 202 OopMap* RegisterSaver::push_frame_reg_args_and_save_live_registers(MacroAssembler* masm, 203 int* out_frame_size_in_bytes, 204 bool generate_oop_map, 205 int return_pc_adjustment, 206 ReturnPCLocation return_pc_location) { 207 // Push an abi_reg_args-frame and store all registers which may be live. 208 // If requested, create an OopMap: Record volatile registers as 209 // callee-save values in an OopMap so their save locations will be 210 // propagated to the RegisterMap of the caller frame during 211 // StackFrameStream construction (needed for deoptimization; see 212 // compiledVFrame::create_stack_value). 213 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment. 214 215 int i; 216 int offset; 217 218 // calcualte frame size 219 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 220 sizeof(RegisterSaver::LiveRegType); 221 const int register_save_size = regstosave_num * reg_size; 222 const int frame_size_in_bytes = round_to(register_save_size, frame::alignment_in_bytes) 223 + frame::abi_reg_args_size; 224 *out_frame_size_in_bytes = frame_size_in_bytes; 225 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint); 226 const int register_save_offset = frame_size_in_bytes - register_save_size; 227 228 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words. 229 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : NULL; 230 231 BLOCK_COMMENT("push_frame_reg_args_and_save_live_registers {"); 232 233 // Save r31 in the last slot of the not yet pushed frame so that we 234 // can use it as scratch reg. 235 __ std(R31, -reg_size, R1_SP); 236 assert(-reg_size == register_save_offset - frame_size_in_bytes + ((regstosave_num-1)*reg_size), 237 "consistency check"); 238 239 // save the flags 240 // Do the save_LR_CR by hand and adjust the return pc if requested. 241 __ mfcr(R31); 242 __ std(R31, _abi(cr), R1_SP); 243 switch (return_pc_location) { 244 case return_pc_is_lr: __ mflr(R31); break; 245 case return_pc_is_r4: __ mr(R31, R4); break; 246 case return_pc_is_thread_saved_exception_pc: 247 __ ld(R31, thread_(saved_exception_pc)); break; 248 default: ShouldNotReachHere(); 249 } 250 if (return_pc_adjustment != 0) { 251 __ addi(R31, R31, return_pc_adjustment); 252 } 253 __ std(R31, _abi(lr), R1_SP); 254 255 // push a new frame 256 __ push_frame(frame_size_in_bytes, R31); 257 258 // save all registers (ints and floats) 259 offset = register_save_offset; 260 for (int i = 0; i < regstosave_num; i++) { 261 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 262 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 263 264 switch (reg_type) { 265 case RegisterSaver::int_reg: { 266 if (reg_num != 31) { // We spilled R31 right at the beginning. 267 __ std(as_Register(reg_num), offset, R1_SP); 268 } 269 break; 270 } 271 case RegisterSaver::float_reg: { 272 __ stfd(as_FloatRegister(reg_num), offset, R1_SP); 273 break; 274 } 275 case RegisterSaver::special_reg: { 276 if (reg_num == SR_CTR_SpecialRegisterEnumValue) { 277 __ mfctr(R31); 278 __ std(R31, offset, R1_SP); 279 } else { 280 Unimplemented(); 281 } 282 break; 283 } 284 default: 285 ShouldNotReachHere(); 286 } 287 288 if (generate_oop_map) { 289 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), 290 RegisterSaver_LiveRegs[i].vmreg); 291 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size)>>2), 292 RegisterSaver_LiveRegs[i].vmreg->next()); 293 } 294 offset += reg_size; 295 } 296 297 BLOCK_COMMENT("} push_frame_reg_args_and_save_live_registers"); 298 299 // And we're done. 300 return map; 301 } 302 303 304 // Pop the current frame and restore all the registers that we 305 // saved. 306 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm, 307 int frame_size_in_bytes, 308 bool restore_ctr) { 309 int i; 310 int offset; 311 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 312 sizeof(RegisterSaver::LiveRegType); 313 const int register_save_size = regstosave_num * reg_size; 314 const int register_save_offset = frame_size_in_bytes - register_save_size; 315 316 BLOCK_COMMENT("restore_live_registers_and_pop_frame {"); 317 318 // restore all registers (ints and floats) 319 offset = register_save_offset; 320 for (int i = 0; i < regstosave_num; i++) { 321 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 322 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 323 324 switch (reg_type) { 325 case RegisterSaver::int_reg: { 326 if (reg_num != 31) // R31 restored at the end, it's the tmp reg! 327 __ ld(as_Register(reg_num), offset, R1_SP); 328 break; 329 } 330 case RegisterSaver::float_reg: { 331 __ lfd(as_FloatRegister(reg_num), offset, R1_SP); 332 break; 333 } 334 case RegisterSaver::special_reg: { 335 if (reg_num == SR_CTR_SpecialRegisterEnumValue) { 336 if (restore_ctr) { // Nothing to do here if ctr already contains the next address. 337 __ ld(R31, offset, R1_SP); 338 __ mtctr(R31); 339 } 340 } else { 341 Unimplemented(); 342 } 343 break; 344 } 345 default: 346 ShouldNotReachHere(); 347 } 348 offset += reg_size; 349 } 350 351 // pop the frame 352 __ pop_frame(); 353 354 // restore the flags 355 __ restore_LR_CR(R31); 356 357 // restore scratch register's value 358 __ ld(R31, -reg_size, R1_SP); 359 360 BLOCK_COMMENT("} restore_live_registers_and_pop_frame"); 361 } 362 363 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp, 364 int frame_size,int total_args, const VMRegPair *regs, 365 const VMRegPair *regs2) { 366 __ push_frame(frame_size, r_temp); 367 int st_off = frame_size - wordSize; 368 for (int i = 0; i < total_args; i++) { 369 VMReg r_1 = regs[i].first(); 370 VMReg r_2 = regs[i].second(); 371 if (!r_1->is_valid()) { 372 assert(!r_2->is_valid(), ""); 373 continue; 374 } 375 if (r_1->is_Register()) { 376 Register r = r_1->as_Register(); 377 __ std(r, st_off, R1_SP); 378 st_off -= wordSize; 379 } else if (r_1->is_FloatRegister()) { 380 FloatRegister f = r_1->as_FloatRegister(); 381 __ stfd(f, st_off, R1_SP); 382 st_off -= wordSize; 383 } 384 } 385 if (regs2 != NULL) { 386 for (int i = 0; i < total_args; i++) { 387 VMReg r_1 = regs2[i].first(); 388 VMReg r_2 = regs2[i].second(); 389 if (!r_1->is_valid()) { 390 assert(!r_2->is_valid(), ""); 391 continue; 392 } 393 if (r_1->is_Register()) { 394 Register r = r_1->as_Register(); 395 __ std(r, st_off, R1_SP); 396 st_off -= wordSize; 397 } else if (r_1->is_FloatRegister()) { 398 FloatRegister f = r_1->as_FloatRegister(); 399 __ stfd(f, st_off, R1_SP); 400 st_off -= wordSize; 401 } 402 } 403 } 404 } 405 406 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size, 407 int total_args, const VMRegPair *regs, 408 const VMRegPair *regs2) { 409 int st_off = frame_size - wordSize; 410 for (int i = 0; i < total_args; i++) { 411 VMReg r_1 = regs[i].first(); 412 VMReg r_2 = regs[i].second(); 413 if (r_1->is_Register()) { 414 Register r = r_1->as_Register(); 415 __ ld(r, st_off, R1_SP); 416 st_off -= wordSize; 417 } else if (r_1->is_FloatRegister()) { 418 FloatRegister f = r_1->as_FloatRegister(); 419 __ lfd(f, st_off, R1_SP); 420 st_off -= wordSize; 421 } 422 } 423 if (regs2 != NULL) 424 for (int i = 0; i < total_args; i++) { 425 VMReg r_1 = regs2[i].first(); 426 VMReg r_2 = regs2[i].second(); 427 if (r_1->is_Register()) { 428 Register r = r_1->as_Register(); 429 __ ld(r, st_off, R1_SP); 430 st_off -= wordSize; 431 } else if (r_1->is_FloatRegister()) { 432 FloatRegister f = r_1->as_FloatRegister(); 433 __ lfd(f, st_off, R1_SP); 434 st_off -= wordSize; 435 } 436 } 437 __ pop_frame(); 438 } 439 440 // Restore the registers that might be holding a result. 441 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) { 442 int i; 443 int offset; 444 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) / 445 sizeof(RegisterSaver::LiveRegType); 446 const int register_save_size = regstosave_num * reg_size; 447 const int register_save_offset = frame_size_in_bytes - register_save_size; 448 449 // restore all result registers (ints and floats) 450 offset = register_save_offset; 451 for (int i = 0; i < regstosave_num; i++) { 452 int reg_num = RegisterSaver_LiveRegs[i].reg_num; 453 int reg_type = RegisterSaver_LiveRegs[i].reg_type; 454 switch (reg_type) { 455 case RegisterSaver::int_reg: { 456 if (as_Register(reg_num)==R3_RET) // int result_reg 457 __ ld(as_Register(reg_num), offset, R1_SP); 458 break; 459 } 460 case RegisterSaver::float_reg: { 461 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg 462 __ lfd(as_FloatRegister(reg_num), offset, R1_SP); 463 break; 464 } 465 case RegisterSaver::special_reg: { 466 // Special registers don't hold a result. 467 break; 468 } 469 default: 470 ShouldNotReachHere(); 471 } 472 offset += reg_size; 473 } 474 } 475 476 // Is vector's size (in bytes) bigger than a size saved by default? 477 bool SharedRuntime::is_wide_vector(int size) { 478 ResourceMark rm; 479 // Note, MaxVectorSize == 8 on PPC64. 480 assert(size <= 8, "%d bytes vectors are not supported", size); 481 return size > 8; 482 } 483 #ifdef COMPILER2 484 static int reg2slot(VMReg r) { 485 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 486 } 487 488 static int reg2offset(VMReg r) { 489 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 490 } 491 #endif 492 493 // --------------------------------------------------------------------------- 494 // Read the array of BasicTypes from a signature, and compute where the 495 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 496 // quantities. Values less than VMRegImpl::stack0 are registers, those above 497 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 498 // as framesizes are fixed. 499 // VMRegImpl::stack0 refers to the first slot 0(sp). 500 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register 501 // up to RegisterImpl::number_of_registers) are the 64-bit 502 // integer registers. 503 504 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 505 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 506 // units regardless of build. Of course for i486 there is no 64 bit build 507 508 // The Java calling convention is a "shifted" version of the C ABI. 509 // By skipping the first C ABI register we can call non-static jni methods 510 // with small numbers of arguments without having to shuffle the arguments 511 // at all. Since we control the java ABI we ought to at least get some 512 // advantage out of it. 513 514 const VMReg java_iarg_reg[8] = { 515 R3->as_VMReg(), 516 R4->as_VMReg(), 517 R5->as_VMReg(), 518 R6->as_VMReg(), 519 R7->as_VMReg(), 520 R8->as_VMReg(), 521 R9->as_VMReg(), 522 R10->as_VMReg() 523 }; 524 525 const VMReg java_farg_reg[13] = { 526 F1->as_VMReg(), 527 F2->as_VMReg(), 528 F3->as_VMReg(), 529 F4->as_VMReg(), 530 F5->as_VMReg(), 531 F6->as_VMReg(), 532 F7->as_VMReg(), 533 F8->as_VMReg(), 534 F9->as_VMReg(), 535 F10->as_VMReg(), 536 F11->as_VMReg(), 537 F12->as_VMReg(), 538 F13->as_VMReg() 539 }; 540 541 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]); 542 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]); 543 544 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 545 VMRegPair *regs, 546 int total_args_passed, 547 int is_outgoing) { 548 // C2c calling conventions for compiled-compiled calls. 549 // Put 8 ints/longs into registers _AND_ 13 float/doubles into 550 // registers _AND_ put the rest on the stack. 551 552 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats 553 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles 554 555 int i; 556 VMReg reg; 557 int stk = 0; 558 int ireg = 0; 559 int freg = 0; 560 561 // We put the first 8 arguments into registers and the rest on the 562 // stack, float arguments are already in their argument registers 563 // due to c2c calling conventions (see calling_convention). 564 for (int i = 0; i < total_args_passed; ++i) { 565 switch(sig_bt[i]) { 566 case T_BOOLEAN: 567 case T_CHAR: 568 case T_BYTE: 569 case T_SHORT: 570 case T_INT: 571 if (ireg < num_java_iarg_registers) { 572 // Put int/ptr in register 573 reg = java_iarg_reg[ireg]; 574 ++ireg; 575 } else { 576 // Put int/ptr on stack. 577 reg = VMRegImpl::stack2reg(stk); 578 stk += inc_stk_for_intfloat; 579 } 580 regs[i].set1(reg); 581 break; 582 case T_LONG: 583 assert(sig_bt[i+1] == T_VOID, "expecting half"); 584 if (ireg < num_java_iarg_registers) { 585 // Put long in register. 586 reg = java_iarg_reg[ireg]; 587 ++ireg; 588 } else { 589 // Put long on stack. They must be aligned to 2 slots. 590 if (stk & 0x1) ++stk; 591 reg = VMRegImpl::stack2reg(stk); 592 stk += inc_stk_for_longdouble; 593 } 594 regs[i].set2(reg); 595 break; 596 case T_OBJECT: 597 case T_ARRAY: 598 case T_ADDRESS: 599 if (ireg < num_java_iarg_registers) { 600 // Put ptr in register. 601 reg = java_iarg_reg[ireg]; 602 ++ireg; 603 } else { 604 // Put ptr on stack. Objects must be aligned to 2 slots too, 605 // because "64-bit pointers record oop-ishness on 2 aligned 606 // adjacent registers." (see OopFlow::build_oop_map). 607 if (stk & 0x1) ++stk; 608 reg = VMRegImpl::stack2reg(stk); 609 stk += inc_stk_for_longdouble; 610 } 611 regs[i].set2(reg); 612 break; 613 case T_FLOAT: 614 if (freg < num_java_farg_registers) { 615 // Put float in register. 616 reg = java_farg_reg[freg]; 617 ++freg; 618 } else { 619 // Put float on stack. 620 reg = VMRegImpl::stack2reg(stk); 621 stk += inc_stk_for_intfloat; 622 } 623 regs[i].set1(reg); 624 break; 625 case T_DOUBLE: 626 assert(sig_bt[i+1] == T_VOID, "expecting half"); 627 if (freg < num_java_farg_registers) { 628 // Put double in register. 629 reg = java_farg_reg[freg]; 630 ++freg; 631 } else { 632 // Put double on stack. They must be aligned to 2 slots. 633 if (stk & 0x1) ++stk; 634 reg = VMRegImpl::stack2reg(stk); 635 stk += inc_stk_for_longdouble; 636 } 637 regs[i].set2(reg); 638 break; 639 case T_VOID: 640 // Do not count halves. 641 regs[i].set_bad(); 642 break; 643 default: 644 ShouldNotReachHere(); 645 } 646 } 647 return round_to(stk, 2); 648 } 649 650 #ifdef COMPILER2 651 // Calling convention for calling C code. 652 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 653 VMRegPair *regs, 654 VMRegPair *regs2, 655 int total_args_passed) { 656 // Calling conventions for C runtime calls and calls to JNI native methods. 657 // 658 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8 659 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist 660 // the first 13 flt/dbl's in the first 13 fp regs but additionally 661 // copy flt/dbl to the stack if they are beyond the 8th argument. 662 663 const VMReg iarg_reg[8] = { 664 R3->as_VMReg(), 665 R4->as_VMReg(), 666 R5->as_VMReg(), 667 R6->as_VMReg(), 668 R7->as_VMReg(), 669 R8->as_VMReg(), 670 R9->as_VMReg(), 671 R10->as_VMReg() 672 }; 673 674 const VMReg farg_reg[13] = { 675 F1->as_VMReg(), 676 F2->as_VMReg(), 677 F3->as_VMReg(), 678 F4->as_VMReg(), 679 F5->as_VMReg(), 680 F6->as_VMReg(), 681 F7->as_VMReg(), 682 F8->as_VMReg(), 683 F9->as_VMReg(), 684 F10->as_VMReg(), 685 F11->as_VMReg(), 686 F12->as_VMReg(), 687 F13->as_VMReg() 688 }; 689 690 // Check calling conventions consistency. 691 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c && 692 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c, 693 "consistency"); 694 695 // `Stk' counts stack slots. Due to alignment, 32 bit values occupy 696 // 2 such slots, like 64 bit values do. 697 const int inc_stk_for_intfloat = 2; // 2 slots for ints and floats 698 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles 699 700 int i; 701 VMReg reg; 702 // Leave room for C-compatible ABI_REG_ARGS. 703 int stk = (frame::abi_reg_args_size - frame::jit_out_preserve_size) / VMRegImpl::stack_slot_size; 704 int arg = 0; 705 int freg = 0; 706 707 // Avoid passing C arguments in the wrong stack slots. 708 #if defined(ABI_ELFv2) 709 assert((SharedRuntime::out_preserve_stack_slots() + stk) * VMRegImpl::stack_slot_size == 96, 710 "passing C arguments in wrong stack slots"); 711 #else 712 assert((SharedRuntime::out_preserve_stack_slots() + stk) * VMRegImpl::stack_slot_size == 112, 713 "passing C arguments in wrong stack slots"); 714 #endif 715 // We fill-out regs AND regs2 if an argument must be passed in a 716 // register AND in a stack slot. If regs2 is NULL in such a 717 // situation, we bail-out with a fatal error. 718 for (int i = 0; i < total_args_passed; ++i, ++arg) { 719 // Initialize regs2 to BAD. 720 if (regs2 != NULL) regs2[i].set_bad(); 721 722 switch(sig_bt[i]) { 723 724 // 725 // If arguments 0-7 are integers, they are passed in integer registers. 726 // Argument i is placed in iarg_reg[i]. 727 // 728 case T_BOOLEAN: 729 case T_CHAR: 730 case T_BYTE: 731 case T_SHORT: 732 case T_INT: 733 // We must cast ints to longs and use full 64 bit stack slots 734 // here. Thus fall through, handle as long. 735 case T_LONG: 736 case T_OBJECT: 737 case T_ARRAY: 738 case T_ADDRESS: 739 case T_METADATA: 740 // Oops are already boxed if required (JNI). 741 if (arg < Argument::n_int_register_parameters_c) { 742 reg = iarg_reg[arg]; 743 } else { 744 reg = VMRegImpl::stack2reg(stk); 745 stk += inc_stk_for_longdouble; 746 } 747 regs[i].set2(reg); 748 break; 749 750 // 751 // Floats are treated differently from int regs: The first 13 float arguments 752 // are passed in registers (not the float args among the first 13 args). 753 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed 754 // in farg_reg[j] if argument i is the j-th float argument of this call. 755 // 756 case T_FLOAT: 757 if (freg < Argument::n_float_register_parameters_c) { 758 // Put float in register ... 759 reg = farg_reg[freg]; 760 ++freg; 761 762 // Argument i for i > 8 is placed on the stack even if it's 763 // placed in a register (if it's a float arg). Aix disassembly 764 // shows that xlC places these float args on the stack AND in 765 // a register. This is not documented, but we follow this 766 // convention, too. 767 if (arg >= Argument::n_regs_not_on_stack_c) { 768 // ... and on the stack. 769 guarantee(regs2 != NULL, "must pass float in register and stack slot"); 770 VMReg reg2 = VMRegImpl::stack2reg(stk LINUX_ONLY(+1)); 771 regs2[i].set1(reg2); 772 stk += inc_stk_for_intfloat; 773 } 774 775 } else { 776 // Put float on stack. 777 reg = VMRegImpl::stack2reg(stk LINUX_ONLY(+1)); 778 stk += inc_stk_for_intfloat; 779 } 780 regs[i].set1(reg); 781 break; 782 case T_DOUBLE: 783 assert(sig_bt[i+1] == T_VOID, "expecting half"); 784 if (freg < Argument::n_float_register_parameters_c) { 785 // Put double in register ... 786 reg = farg_reg[freg]; 787 ++freg; 788 789 // Argument i for i > 8 is placed on the stack even if it's 790 // placed in a register (if it's a double arg). Aix disassembly 791 // shows that xlC places these float args on the stack AND in 792 // a register. This is not documented, but we follow this 793 // convention, too. 794 if (arg >= Argument::n_regs_not_on_stack_c) { 795 // ... and on the stack. 796 guarantee(regs2 != NULL, "must pass float in register and stack slot"); 797 VMReg reg2 = VMRegImpl::stack2reg(stk); 798 regs2[i].set2(reg2); 799 stk += inc_stk_for_longdouble; 800 } 801 } else { 802 // Put double on stack. 803 reg = VMRegImpl::stack2reg(stk); 804 stk += inc_stk_for_longdouble; 805 } 806 regs[i].set2(reg); 807 break; 808 809 case T_VOID: 810 // Do not count halves. 811 regs[i].set_bad(); 812 --arg; 813 break; 814 default: 815 ShouldNotReachHere(); 816 } 817 } 818 819 return round_to(stk, 2); 820 } 821 #endif // COMPILER2 822 823 static address gen_c2i_adapter(MacroAssembler *masm, 824 int total_args_passed, 825 int comp_args_on_stack, 826 const BasicType *sig_bt, 827 const VMRegPair *regs, 828 Label& call_interpreter, 829 const Register& ientry) { 830 831 address c2i_entrypoint; 832 833 const Register sender_SP = R21_sender_SP; // == R21_tmp1 834 const Register code = R22_tmp2; 835 //const Register ientry = R23_tmp3; 836 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 }; 837 const int num_value_regs = sizeof(value_regs) / sizeof(Register); 838 int value_regs_index = 0; 839 840 const Register return_pc = R27_tmp7; 841 const Register tmp = R28_tmp8; 842 843 assert_different_registers(sender_SP, code, ientry, return_pc, tmp); 844 845 // Adapter needs TOP_IJAVA_FRAME_ABI. 846 const int adapter_size = frame::top_ijava_frame_abi_size + 847 round_to(total_args_passed * wordSize, frame::alignment_in_bytes); 848 849 // regular (verified) c2i entry point 850 c2i_entrypoint = __ pc(); 851 852 // Does compiled code exists? If yes, patch the caller's callsite. 853 __ ld(code, method_(code)); 854 __ cmpdi(CCR0, code, 0); 855 __ ld(ientry, method_(interpreter_entry)); // preloaded 856 __ beq(CCR0, call_interpreter); 857 858 859 // Patch caller's callsite, method_(code) was not NULL which means that 860 // compiled code exists. 861 __ mflr(return_pc); 862 __ std(return_pc, _abi(lr), R1_SP); 863 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs); 864 865 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc); 866 867 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs); 868 __ ld(return_pc, _abi(lr), R1_SP); 869 __ ld(ientry, method_(interpreter_entry)); // preloaded 870 __ mtlr(return_pc); 871 872 873 // Call the interpreter. 874 __ BIND(call_interpreter); 875 __ mtctr(ientry); 876 877 // Get a copy of the current SP for loading caller's arguments. 878 __ mr(sender_SP, R1_SP); 879 880 // Add space for the adapter. 881 __ resize_frame(-adapter_size, R12_scratch2); 882 883 int st_off = adapter_size - wordSize; 884 885 // Write the args into the outgoing interpreter space. 886 for (int i = 0; i < total_args_passed; i++) { 887 VMReg r_1 = regs[i].first(); 888 VMReg r_2 = regs[i].second(); 889 if (!r_1->is_valid()) { 890 assert(!r_2->is_valid(), ""); 891 continue; 892 } 893 if (r_1->is_stack()) { 894 Register tmp_reg = value_regs[value_regs_index]; 895 value_regs_index = (value_regs_index + 1) % num_value_regs; 896 // The calling convention produces OptoRegs that ignore the out 897 // preserve area (JIT's ABI). We must account for it here. 898 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 899 if (!r_2->is_valid()) { 900 __ lwz(tmp_reg, ld_off, sender_SP); 901 } else { 902 __ ld(tmp_reg, ld_off, sender_SP); 903 } 904 // Pretend stack targets were loaded into tmp_reg. 905 r_1 = tmp_reg->as_VMReg(); 906 } 907 908 if (r_1->is_Register()) { 909 Register r = r_1->as_Register(); 910 if (!r_2->is_valid()) { 911 __ stw(r, st_off, R1_SP); 912 st_off-=wordSize; 913 } else { 914 // Longs are given 2 64-bit slots in the interpreter, but the 915 // data is passed in only 1 slot. 916 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 917 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); ) 918 st_off-=wordSize; 919 } 920 __ std(r, st_off, R1_SP); 921 st_off-=wordSize; 922 } 923 } else { 924 assert(r_1->is_FloatRegister(), ""); 925 FloatRegister f = r_1->as_FloatRegister(); 926 if (!r_2->is_valid()) { 927 __ stfs(f, st_off, R1_SP); 928 st_off-=wordSize; 929 } else { 930 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the 931 // data is passed in only 1 slot. 932 // One of these should get known junk... 933 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); ) 934 st_off-=wordSize; 935 __ stfd(f, st_off, R1_SP); 936 st_off-=wordSize; 937 } 938 } 939 } 940 941 // Jump to the interpreter just as if interpreter was doing it. 942 943 #ifdef CC_INTERP 944 const Register tos = R17_tos; 945 #else 946 const Register tos = R15_esp; 947 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 948 #endif 949 950 // load TOS 951 __ addi(tos, R1_SP, st_off); 952 953 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1. 954 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register"); 955 __ bctr(); 956 957 return c2i_entrypoint; 958 } 959 960 static void gen_i2c_adapter(MacroAssembler *masm, 961 int total_args_passed, 962 int comp_args_on_stack, 963 const BasicType *sig_bt, 964 const VMRegPair *regs) { 965 966 // Load method's entry-point from method. 967 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method); 968 __ mtctr(R12_scratch2); 969 970 // We will only enter here from an interpreted frame and never from after 971 // passing thru a c2i. Azul allowed this but we do not. If we lose the 972 // race and use a c2i we will remain interpreted for the race loser(s). 973 // This removes all sorts of headaches on the x86 side and also eliminates 974 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions. 975 976 // Note: r13 contains the senderSP on entry. We must preserve it since 977 // we may do a i2c -> c2i transition if we lose a race where compiled 978 // code goes non-entrant while we get args ready. 979 // In addition we use r13 to locate all the interpreter args as 980 // we must align the stack to 16 bytes on an i2c entry else we 981 // lose alignment we expect in all compiled code and register 982 // save code can segv when fxsave instructions find improperly 983 // aligned stack pointer. 984 985 #ifdef CC_INTERP 986 const Register ld_ptr = R17_tos; 987 #else 988 const Register ld_ptr = R15_esp; 989 #endif 990 991 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 }; 992 const int num_value_regs = sizeof(value_regs) / sizeof(Register); 993 int value_regs_index = 0; 994 995 int ld_offset = total_args_passed*wordSize; 996 997 // Cut-out for having no stack args. Since up to 2 int/oop args are passed 998 // in registers, we will occasionally have no stack args. 999 int comp_words_on_stack = 0; 1000 if (comp_args_on_stack) { 1001 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in 1002 // registers are below. By subtracting stack0, we either get a negative 1003 // number (all values in registers) or the maximum stack slot accessed. 1004 1005 // Convert 4-byte c2 stack slots to words. 1006 comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 1007 // Round up to miminum stack alignment, in wordSize. 1008 comp_words_on_stack = round_to(comp_words_on_stack, 2); 1009 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1); 1010 } 1011 1012 // Now generate the shuffle code. Pick up all register args and move the 1013 // rest through register value=Z_R12. 1014 BLOCK_COMMENT("Shuffle arguments"); 1015 for (int i = 0; i < total_args_passed; i++) { 1016 if (sig_bt[i] == T_VOID) { 1017 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 1018 continue; 1019 } 1020 1021 // Pick up 0, 1 or 2 words from ld_ptr. 1022 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 1023 "scrambled load targets?"); 1024 VMReg r_1 = regs[i].first(); 1025 VMReg r_2 = regs[i].second(); 1026 if (!r_1->is_valid()) { 1027 assert(!r_2->is_valid(), ""); 1028 continue; 1029 } 1030 if (r_1->is_FloatRegister()) { 1031 if (!r_2->is_valid()) { 1032 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr); 1033 ld_offset-=wordSize; 1034 } else { 1035 // Skip the unused interpreter slot. 1036 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr); 1037 ld_offset-=2*wordSize; 1038 } 1039 } else { 1040 Register r; 1041 if (r_1->is_stack()) { 1042 // Must do a memory to memory move thru "value". 1043 r = value_regs[value_regs_index]; 1044 value_regs_index = (value_regs_index + 1) % num_value_regs; 1045 } else { 1046 r = r_1->as_Register(); 1047 } 1048 if (!r_2->is_valid()) { 1049 // Not sure we need to do this but it shouldn't hurt. 1050 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ADDRESS || sig_bt[i] == T_ARRAY) { 1051 __ ld(r, ld_offset, ld_ptr); 1052 ld_offset-=wordSize; 1053 } else { 1054 __ lwz(r, ld_offset, ld_ptr); 1055 ld_offset-=wordSize; 1056 } 1057 } else { 1058 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the 1059 // data is passed in only 1 slot. 1060 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 1061 ld_offset-=wordSize; 1062 } 1063 __ ld(r, ld_offset, ld_ptr); 1064 ld_offset-=wordSize; 1065 } 1066 1067 if (r_1->is_stack()) { 1068 // Now store value where the compiler expects it 1069 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size; 1070 1071 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN || 1072 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) { 1073 __ stw(r, st_off, R1_SP); 1074 } else { 1075 __ std(r, st_off, R1_SP); 1076 } 1077 } 1078 } 1079 } 1080 1081 BLOCK_COMMENT("Store method"); 1082 // Store method into thread->callee_target. 1083 // We might end up in handle_wrong_method if the callee is 1084 // deoptimized as we race thru here. If that happens we don't want 1085 // to take a safepoint because the caller frame will look 1086 // interpreted and arguments are now "compiled" so it is much better 1087 // to make this transition invisible to the stack walking 1088 // code. Unfortunately if we try and find the callee by normal means 1089 // a safepoint is possible. So we stash the desired callee in the 1090 // thread and the vm will find there should this case occur. 1091 __ std(R19_method, thread_(callee_target)); 1092 1093 // Jump to the compiled code just as if compiled code was doing it. 1094 __ bctr(); 1095 } 1096 1097 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 1098 int total_args_passed, 1099 int comp_args_on_stack, 1100 const BasicType *sig_bt, 1101 const VMRegPair *regs, 1102 AdapterFingerPrint* fingerprint) { 1103 address i2c_entry; 1104 address c2i_unverified_entry; 1105 address c2i_entry; 1106 1107 1108 // entry: i2c 1109 1110 __ align(CodeEntryAlignment); 1111 i2c_entry = __ pc(); 1112 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 1113 1114 1115 // entry: c2i unverified 1116 1117 __ align(CodeEntryAlignment); 1118 BLOCK_COMMENT("c2i unverified entry"); 1119 c2i_unverified_entry = __ pc(); 1120 1121 // inline_cache contains a compiledICHolder 1122 const Register ic = R19_method; 1123 const Register ic_klass = R11_scratch1; 1124 const Register receiver_klass = R12_scratch2; 1125 const Register code = R21_tmp1; 1126 const Register ientry = R23_tmp3; 1127 1128 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry); 1129 assert(R11_scratch1 == R11, "need prologue scratch register"); 1130 1131 Label call_interpreter; 1132 1133 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), 1134 "klass offset should reach into any page"); 1135 // Check for NULL argument if we don't have implicit null checks. 1136 if (!ImplicitNullChecks || !os::zero_page_read_protected()) { 1137 if (TrapBasedNullChecks) { 1138 __ trap_null_check(R3_ARG1); 1139 } else { 1140 Label valid; 1141 __ cmpdi(CCR0, R3_ARG1, 0); 1142 __ bne_predict_taken(CCR0, valid); 1143 // We have a null argument, branch to ic_miss_stub. 1144 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1145 relocInfo::runtime_call_type); 1146 __ BIND(valid); 1147 } 1148 } 1149 // Assume argument is not NULL, load klass from receiver. 1150 __ load_klass(receiver_klass, R3_ARG1); 1151 1152 __ ld(ic_klass, CompiledICHolder::holder_klass_offset(), ic); 1153 1154 if (TrapBasedICMissChecks) { 1155 __ trap_ic_miss_check(receiver_klass, ic_klass); 1156 } else { 1157 Label valid; 1158 __ cmpd(CCR0, receiver_klass, ic_klass); 1159 __ beq_predict_taken(CCR0, valid); 1160 // We have an unexpected klass, branch to ic_miss_stub. 1161 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1162 relocInfo::runtime_call_type); 1163 __ BIND(valid); 1164 } 1165 1166 // Argument is valid and klass is as expected, continue. 1167 1168 // Extract method from inline cache, verified entry point needs it. 1169 __ ld(R19_method, CompiledICHolder::holder_method_offset(), ic); 1170 assert(R19_method == ic, "the inline cache register is dead here"); 1171 1172 __ ld(code, method_(code)); 1173 __ cmpdi(CCR0, code, 0); 1174 __ ld(ientry, method_(interpreter_entry)); // preloaded 1175 __ beq_predict_taken(CCR0, call_interpreter); 1176 1177 // Branch to ic_miss_stub. 1178 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type); 1179 1180 // entry: c2i 1181 1182 c2i_entry = gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry); 1183 1184 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 1185 } 1186 1187 #ifdef COMPILER2 1188 // An oop arg. Must pass a handle not the oop itself. 1189 static void object_move(MacroAssembler* masm, 1190 int frame_size_in_slots, 1191 OopMap* oop_map, int oop_handle_offset, 1192 bool is_receiver, int* receiver_offset, 1193 VMRegPair src, VMRegPair dst, 1194 Register r_caller_sp, Register r_temp_1, Register r_temp_2) { 1195 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)), 1196 "receiver has already been moved"); 1197 1198 // We must pass a handle. First figure out the location we use as a handle. 1199 1200 if (src.first()->is_stack()) { 1201 // stack to stack or reg 1202 1203 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register(); 1204 Label skip; 1205 const int oop_slot_in_callers_frame = reg2slot(src.first()); 1206 1207 guarantee(!is_receiver, "expecting receiver in register"); 1208 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots)); 1209 1210 __ addi(r_handle, r_caller_sp, reg2offset(src.first())); 1211 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp); 1212 __ cmpdi(CCR0, r_temp_2, 0); 1213 __ bne(CCR0, skip); 1214 // Use a NULL handle if oop is NULL. 1215 __ li(r_handle, 0); 1216 __ bind(skip); 1217 1218 if (dst.first()->is_stack()) { 1219 // stack to stack 1220 __ std(r_handle, reg2offset(dst.first()), R1_SP); 1221 } else { 1222 // stack to reg 1223 // Nothing to do, r_handle is already the dst register. 1224 } 1225 } else { 1226 // reg to stack or reg 1227 const Register r_oop = src.first()->as_Register(); 1228 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register(); 1229 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word 1230 + oop_handle_offset; // in slots 1231 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size; 1232 Label skip; 1233 1234 if (is_receiver) { 1235 *receiver_offset = oop_offset; 1236 } 1237 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1238 1239 __ std( r_oop, oop_offset, R1_SP); 1240 __ addi(r_handle, R1_SP, oop_offset); 1241 1242 __ cmpdi(CCR0, r_oop, 0); 1243 __ bne(CCR0, skip); 1244 // Use a NULL handle if oop is NULL. 1245 __ li(r_handle, 0); 1246 __ bind(skip); 1247 1248 if (dst.first()->is_stack()) { 1249 // reg to stack 1250 __ std(r_handle, reg2offset(dst.first()), R1_SP); 1251 } else { 1252 // reg to reg 1253 // Nothing to do, r_handle is already the dst register. 1254 } 1255 } 1256 } 1257 1258 static void int_move(MacroAssembler*masm, 1259 VMRegPair src, VMRegPair dst, 1260 Register r_caller_sp, Register r_temp) { 1261 assert(src.first()->is_valid(), "incoming must be int"); 1262 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long"); 1263 1264 if (src.first()->is_stack()) { 1265 if (dst.first()->is_stack()) { 1266 // stack to stack 1267 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp); 1268 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1269 } else { 1270 // stack to reg 1271 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1272 } 1273 } else if (dst.first()->is_stack()) { 1274 // reg to stack 1275 __ extsw(r_temp, src.first()->as_Register()); 1276 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1277 } else { 1278 // reg to reg 1279 __ extsw(dst.first()->as_Register(), src.first()->as_Register()); 1280 } 1281 } 1282 1283 static void long_move(MacroAssembler*masm, 1284 VMRegPair src, VMRegPair dst, 1285 Register r_caller_sp, Register r_temp) { 1286 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long"); 1287 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long"); 1288 1289 if (src.first()->is_stack()) { 1290 if (dst.first()->is_stack()) { 1291 // stack to stack 1292 __ ld( r_temp, reg2offset(src.first()), r_caller_sp); 1293 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1294 } else { 1295 // stack to reg 1296 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1297 } 1298 } else if (dst.first()->is_stack()) { 1299 // reg to stack 1300 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP); 1301 } else { 1302 // reg to reg 1303 if (dst.first()->as_Register() != src.first()->as_Register()) 1304 __ mr(dst.first()->as_Register(), src.first()->as_Register()); 1305 } 1306 } 1307 1308 static void float_move(MacroAssembler*masm, 1309 VMRegPair src, VMRegPair dst, 1310 Register r_caller_sp, Register r_temp) { 1311 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float"); 1312 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float"); 1313 1314 if (src.first()->is_stack()) { 1315 if (dst.first()->is_stack()) { 1316 // stack to stack 1317 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp); 1318 __ stw(r_temp, reg2offset(dst.first()), R1_SP); 1319 } else { 1320 // stack to reg 1321 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp); 1322 } 1323 } else if (dst.first()->is_stack()) { 1324 // reg to stack 1325 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP); 1326 } else { 1327 // reg to reg 1328 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister()) 1329 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1330 } 1331 } 1332 1333 static void double_move(MacroAssembler*masm, 1334 VMRegPair src, VMRegPair dst, 1335 Register r_caller_sp, Register r_temp) { 1336 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double"); 1337 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double"); 1338 1339 if (src.first()->is_stack()) { 1340 if (dst.first()->is_stack()) { 1341 // stack to stack 1342 __ ld( r_temp, reg2offset(src.first()), r_caller_sp); 1343 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1344 } else { 1345 // stack to reg 1346 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp); 1347 } 1348 } else if (dst.first()->is_stack()) { 1349 // reg to stack 1350 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP); 1351 } else { 1352 // reg to reg 1353 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister()) 1354 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1355 } 1356 } 1357 1358 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1359 switch (ret_type) { 1360 case T_BOOLEAN: 1361 case T_CHAR: 1362 case T_BYTE: 1363 case T_SHORT: 1364 case T_INT: 1365 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1366 break; 1367 case T_ARRAY: 1368 case T_OBJECT: 1369 case T_LONG: 1370 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1371 break; 1372 case T_FLOAT: 1373 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1374 break; 1375 case T_DOUBLE: 1376 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1377 break; 1378 case T_VOID: 1379 break; 1380 default: 1381 ShouldNotReachHere(); 1382 break; 1383 } 1384 } 1385 1386 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1387 switch (ret_type) { 1388 case T_BOOLEAN: 1389 case T_CHAR: 1390 case T_BYTE: 1391 case T_SHORT: 1392 case T_INT: 1393 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1394 break; 1395 case T_ARRAY: 1396 case T_OBJECT: 1397 case T_LONG: 1398 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1399 break; 1400 case T_FLOAT: 1401 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1402 break; 1403 case T_DOUBLE: 1404 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1405 break; 1406 case T_VOID: 1407 break; 1408 default: 1409 ShouldNotReachHere(); 1410 break; 1411 } 1412 } 1413 1414 static void save_or_restore_arguments(MacroAssembler* masm, 1415 const int stack_slots, 1416 const int total_in_args, 1417 const int arg_save_area, 1418 OopMap* map, 1419 VMRegPair* in_regs, 1420 BasicType* in_sig_bt) { 1421 // If map is non-NULL then the code should store the values, 1422 // otherwise it should load them. 1423 int slot = arg_save_area; 1424 // Save down double word first. 1425 for (int i = 0; i < total_in_args; i++) { 1426 if (in_regs[i].first()->is_FloatRegister() && in_sig_bt[i] == T_DOUBLE) { 1427 int offset = slot * VMRegImpl::stack_slot_size; 1428 slot += VMRegImpl::slots_per_word; 1429 assert(slot <= stack_slots, "overflow (after DOUBLE stack slot)"); 1430 if (map != NULL) { 1431 __ stfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1432 } else { 1433 __ lfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1434 } 1435 } else if (in_regs[i].first()->is_Register() && 1436 (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_ARRAY)) { 1437 int offset = slot * VMRegImpl::stack_slot_size; 1438 if (map != NULL) { 1439 __ std(in_regs[i].first()->as_Register(), offset, R1_SP); 1440 if (in_sig_bt[i] == T_ARRAY) { 1441 map->set_oop(VMRegImpl::stack2reg(slot)); 1442 } 1443 } else { 1444 __ ld(in_regs[i].first()->as_Register(), offset, R1_SP); 1445 } 1446 slot += VMRegImpl::slots_per_word; 1447 assert(slot <= stack_slots, "overflow (after LONG/ARRAY stack slot)"); 1448 } 1449 } 1450 // Save or restore single word registers. 1451 for (int i = 0; i < total_in_args; i++) { 1452 // PPC64: pass ints as longs: must only deal with floats here. 1453 if (in_regs[i].first()->is_FloatRegister()) { 1454 if (in_sig_bt[i] == T_FLOAT) { 1455 int offset = slot * VMRegImpl::stack_slot_size; 1456 slot++; 1457 assert(slot <= stack_slots, "overflow (after FLOAT stack slot)"); 1458 if (map != NULL) { 1459 __ stfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1460 } else { 1461 __ lfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1462 } 1463 } 1464 } else if (in_regs[i].first()->is_stack()) { 1465 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1466 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1467 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots)); 1468 } 1469 } 1470 } 1471 } 1472 1473 // Check GC_locker::needs_gc and enter the runtime if it's true. This 1474 // keeps a new JNI critical region from starting until a GC has been 1475 // forced. Save down any oops in registers and describe them in an 1476 // OopMap. 1477 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1478 const int stack_slots, 1479 const int total_in_args, 1480 const int arg_save_area, 1481 OopMapSet* oop_maps, 1482 VMRegPair* in_regs, 1483 BasicType* in_sig_bt, 1484 Register tmp_reg ) { 1485 __ block_comment("check GC_locker::needs_gc"); 1486 Label cont; 1487 __ lbz(tmp_reg, (RegisterOrConstant)(intptr_t)GC_locker::needs_gc_address()); 1488 __ cmplwi(CCR0, tmp_reg, 0); 1489 __ beq(CCR0, cont); 1490 1491 // Save down any values that are live in registers and call into the 1492 // runtime to halt for a GC. 1493 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1494 save_or_restore_arguments(masm, stack_slots, total_in_args, 1495 arg_save_area, map, in_regs, in_sig_bt); 1496 1497 __ mr(R3_ARG1, R16_thread); 1498 __ set_last_Java_frame(R1_SP, noreg); 1499 1500 __ block_comment("block_for_jni_critical"); 1501 address entry_point = CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical); 1502 #if defined(ABI_ELFv2) 1503 __ call_c(entry_point, relocInfo::runtime_call_type); 1504 #else 1505 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, entry_point), relocInfo::runtime_call_type); 1506 #endif 1507 address start = __ pc() - __ offset(), 1508 calls_return_pc = __ last_calls_return_pc(); 1509 oop_maps->add_gc_map(calls_return_pc - start, map); 1510 1511 __ reset_last_Java_frame(); 1512 1513 // Reload all the register arguments. 1514 save_or_restore_arguments(masm, stack_slots, total_in_args, 1515 arg_save_area, NULL, in_regs, in_sig_bt); 1516 1517 __ BIND(cont); 1518 1519 #ifdef ASSERT 1520 if (StressCriticalJNINatives) { 1521 // Stress register saving. 1522 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1523 save_or_restore_arguments(masm, stack_slots, total_in_args, 1524 arg_save_area, map, in_regs, in_sig_bt); 1525 // Destroy argument registers. 1526 for (int i = 0; i < total_in_args; i++) { 1527 if (in_regs[i].first()->is_Register()) { 1528 const Register reg = in_regs[i].first()->as_Register(); 1529 __ neg(reg, reg); 1530 } else if (in_regs[i].first()->is_FloatRegister()) { 1531 __ fneg(in_regs[i].first()->as_FloatRegister(), in_regs[i].first()->as_FloatRegister()); 1532 } 1533 } 1534 1535 save_or_restore_arguments(masm, stack_slots, total_in_args, 1536 arg_save_area, NULL, in_regs, in_sig_bt); 1537 } 1538 #endif 1539 } 1540 1541 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst, Register r_caller_sp, Register r_temp) { 1542 if (src.first()->is_stack()) { 1543 if (dst.first()->is_stack()) { 1544 // stack to stack 1545 __ ld(r_temp, reg2offset(src.first()), r_caller_sp); 1546 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1547 } else { 1548 // stack to reg 1549 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1550 } 1551 } else if (dst.first()->is_stack()) { 1552 // reg to stack 1553 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP); 1554 } else { 1555 if (dst.first() != src.first()) { 1556 __ mr(dst.first()->as_Register(), src.first()->as_Register()); 1557 } 1558 } 1559 } 1560 1561 // Unpack an array argument into a pointer to the body and the length 1562 // if the array is non-null, otherwise pass 0 for both. 1563 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, 1564 VMRegPair body_arg, VMRegPair length_arg, Register r_caller_sp, 1565 Register tmp_reg, Register tmp2_reg) { 1566 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg, 1567 "possible collision"); 1568 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg, 1569 "possible collision"); 1570 1571 // Pass the length, ptr pair. 1572 Label set_out_args; 1573 VMRegPair tmp, tmp2; 1574 tmp.set_ptr(tmp_reg->as_VMReg()); 1575 tmp2.set_ptr(tmp2_reg->as_VMReg()); 1576 if (reg.first()->is_stack()) { 1577 // Load the arg up from the stack. 1578 move_ptr(masm, reg, tmp, r_caller_sp, /*unused*/ R0); 1579 reg = tmp; 1580 } 1581 __ li(tmp2_reg, 0); // Pass zeros if Array=null. 1582 if (tmp_reg != reg.first()->as_Register()) __ li(tmp_reg, 0); 1583 __ cmpdi(CCR0, reg.first()->as_Register(), 0); 1584 __ beq(CCR0, set_out_args); 1585 __ lwa(tmp2_reg, arrayOopDesc::length_offset_in_bytes(), reg.first()->as_Register()); 1586 __ addi(tmp_reg, reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type)); 1587 __ bind(set_out_args); 1588 move_ptr(masm, tmp, body_arg, r_caller_sp, /*unused*/ R0); 1589 move_ptr(masm, tmp2, length_arg, r_caller_sp, /*unused*/ R0); // Same as move32_64 on PPC64. 1590 } 1591 1592 static void verify_oop_args(MacroAssembler* masm, 1593 methodHandle method, 1594 const BasicType* sig_bt, 1595 const VMRegPair* regs) { 1596 Register temp_reg = R19_method; // not part of any compiled calling seq 1597 if (VerifyOops) { 1598 for (int i = 0; i < method->size_of_parameters(); i++) { 1599 if (sig_bt[i] == T_OBJECT || 1600 sig_bt[i] == T_ARRAY) { 1601 VMReg r = regs[i].first(); 1602 assert(r->is_valid(), "bad oop arg"); 1603 if (r->is_stack()) { 1604 __ ld(temp_reg, reg2offset(r), R1_SP); 1605 __ verify_oop(temp_reg); 1606 } else { 1607 __ verify_oop(r->as_Register()); 1608 } 1609 } 1610 } 1611 } 1612 } 1613 1614 static void gen_special_dispatch(MacroAssembler* masm, 1615 methodHandle method, 1616 const BasicType* sig_bt, 1617 const VMRegPair* regs) { 1618 verify_oop_args(masm, method, sig_bt, regs); 1619 vmIntrinsics::ID iid = method->intrinsic_id(); 1620 1621 // Now write the args into the outgoing interpreter space 1622 bool has_receiver = false; 1623 Register receiver_reg = noreg; 1624 int member_arg_pos = -1; 1625 Register member_reg = noreg; 1626 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1627 if (ref_kind != 0) { 1628 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1629 member_reg = R19_method; // known to be free at this point 1630 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1631 } else if (iid == vmIntrinsics::_invokeBasic) { 1632 has_receiver = true; 1633 } else { 1634 fatal("unexpected intrinsic id %d", iid); 1635 } 1636 1637 if (member_reg != noreg) { 1638 // Load the member_arg into register, if necessary. 1639 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1640 VMReg r = regs[member_arg_pos].first(); 1641 if (r->is_stack()) { 1642 __ ld(member_reg, reg2offset(r), R1_SP); 1643 } else { 1644 // no data motion is needed 1645 member_reg = r->as_Register(); 1646 } 1647 } 1648 1649 if (has_receiver) { 1650 // Make sure the receiver is loaded into a register. 1651 assert(method->size_of_parameters() > 0, "oob"); 1652 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1653 VMReg r = regs[0].first(); 1654 assert(r->is_valid(), "bad receiver arg"); 1655 if (r->is_stack()) { 1656 // Porting note: This assumes that compiled calling conventions always 1657 // pass the receiver oop in a register. If this is not true on some 1658 // platform, pick a temp and load the receiver from stack. 1659 fatal("receiver always in a register"); 1660 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point? 1661 __ ld(receiver_reg, reg2offset(r), R1_SP); 1662 } else { 1663 // no data motion is needed 1664 receiver_reg = r->as_Register(); 1665 } 1666 } 1667 1668 // Figure out which address we are really jumping to: 1669 MethodHandles::generate_method_handle_dispatch(masm, iid, 1670 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1671 } 1672 1673 #endif // COMPILER2 1674 1675 // --------------------------------------------------------------------------- 1676 // Generate a native wrapper for a given method. The method takes arguments 1677 // in the Java compiled code convention, marshals them to the native 1678 // convention (handlizes oops, etc), transitions to native, makes the call, 1679 // returns to java state (possibly blocking), unhandlizes any result and 1680 // returns. 1681 // 1682 // Critical native functions are a shorthand for the use of 1683 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1684 // functions. The wrapper is expected to unpack the arguments before 1685 // passing them to the callee and perform checks before and after the 1686 // native call to ensure that they GC_locker 1687 // lock_critical/unlock_critical semantics are followed. Some other 1688 // parts of JNI setup are skipped like the tear down of the JNI handle 1689 // block and the check for pending exceptions it's impossible for them 1690 // to be thrown. 1691 // 1692 // They are roughly structured like this: 1693 // if (GC_locker::needs_gc()) 1694 // SharedRuntime::block_for_jni_critical(); 1695 // tranistion to thread_in_native 1696 // unpack arrray arguments and call native entry point 1697 // check for safepoint in progress 1698 // check if any thread suspend flags are set 1699 // call into JVM and possible unlock the JNI critical 1700 // if a GC was suppressed while in the critical native. 1701 // transition back to thread_in_Java 1702 // return to caller 1703 // 1704 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm, 1705 methodHandle method, 1706 int compile_id, 1707 BasicType *in_sig_bt, 1708 VMRegPair *in_regs, 1709 BasicType ret_type) { 1710 #ifdef COMPILER2 1711 if (method->is_method_handle_intrinsic()) { 1712 vmIntrinsics::ID iid = method->intrinsic_id(); 1713 intptr_t start = (intptr_t)__ pc(); 1714 int vep_offset = ((intptr_t)__ pc()) - start; 1715 gen_special_dispatch(masm, 1716 method, 1717 in_sig_bt, 1718 in_regs); 1719 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1720 __ flush(); 1721 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1722 return nmethod::new_native_nmethod(method, 1723 compile_id, 1724 masm->code(), 1725 vep_offset, 1726 frame_complete, 1727 stack_slots / VMRegImpl::slots_per_word, 1728 in_ByteSize(-1), 1729 in_ByteSize(-1), 1730 (OopMapSet*)NULL); 1731 } 1732 1733 bool is_critical_native = true; 1734 address native_func = method->critical_native_function(); 1735 if (native_func == NULL) { 1736 native_func = method->native_function(); 1737 is_critical_native = false; 1738 } 1739 assert(native_func != NULL, "must have function"); 1740 1741 // First, create signature for outgoing C call 1742 // -------------------------------------------------------------------------- 1743 1744 int total_in_args = method->size_of_parameters(); 1745 // We have received a description of where all the java args are located 1746 // on entry to the wrapper. We need to convert these args to where 1747 // the jni function will expect them. To figure out where they go 1748 // we convert the java signature to a C signature by inserting 1749 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1750 1751 // Calculate the total number of C arguments and create arrays for the 1752 // signature and the outgoing registers. 1753 // On ppc64, we have two arrays for the outgoing registers, because 1754 // some floating-point arguments must be passed in registers _and_ 1755 // in stack locations. 1756 bool method_is_static = method->is_static(); 1757 int total_c_args = total_in_args; 1758 1759 if (!is_critical_native) { 1760 int n_hidden_args = method_is_static ? 2 : 1; 1761 total_c_args += n_hidden_args; 1762 } else { 1763 // No JNIEnv*, no this*, but unpacked arrays (base+length). 1764 for (int i = 0; i < total_in_args; i++) { 1765 if (in_sig_bt[i] == T_ARRAY) { 1766 total_c_args++; 1767 } 1768 } 1769 } 1770 1771 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1772 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1773 VMRegPair *out_regs2 = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1774 BasicType* in_elem_bt = NULL; 1775 1776 // Create the signature for the C call: 1777 // 1) add the JNIEnv* 1778 // 2) add the class if the method is static 1779 // 3) copy the rest of the incoming signature (shifted by the number of 1780 // hidden arguments). 1781 1782 int argc = 0; 1783 if (!is_critical_native) { 1784 out_sig_bt[argc++] = T_ADDRESS; 1785 if (method->is_static()) { 1786 out_sig_bt[argc++] = T_OBJECT; 1787 } 1788 1789 for (int i = 0; i < total_in_args ; i++ ) { 1790 out_sig_bt[argc++] = in_sig_bt[i]; 1791 } 1792 } else { 1793 Thread* THREAD = Thread::current(); 1794 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1795 SignatureStream ss(method->signature()); 1796 int o = 0; 1797 for (int i = 0; i < total_in_args ; i++, o++) { 1798 if (in_sig_bt[i] == T_ARRAY) { 1799 // Arrays are passed as int, elem* pair 1800 Symbol* atype = ss.as_symbol(CHECK_NULL); 1801 const char* at = atype->as_C_string(); 1802 if (strlen(at) == 2) { 1803 assert(at[0] == '[', "must be"); 1804 switch (at[1]) { 1805 case 'B': in_elem_bt[o] = T_BYTE; break; 1806 case 'C': in_elem_bt[o] = T_CHAR; break; 1807 case 'D': in_elem_bt[o] = T_DOUBLE; break; 1808 case 'F': in_elem_bt[o] = T_FLOAT; break; 1809 case 'I': in_elem_bt[o] = T_INT; break; 1810 case 'J': in_elem_bt[o] = T_LONG; break; 1811 case 'S': in_elem_bt[o] = T_SHORT; break; 1812 case 'Z': in_elem_bt[o] = T_BOOLEAN; break; 1813 default: ShouldNotReachHere(); 1814 } 1815 } 1816 } else { 1817 in_elem_bt[o] = T_VOID; 1818 } 1819 if (in_sig_bt[i] != T_VOID) { 1820 assert(in_sig_bt[i] == ss.type(), "must match"); 1821 ss.next(); 1822 } 1823 } 1824 1825 for (int i = 0; i < total_in_args ; i++ ) { 1826 if (in_sig_bt[i] == T_ARRAY) { 1827 // Arrays are passed as int, elem* pair. 1828 out_sig_bt[argc++] = T_INT; 1829 out_sig_bt[argc++] = T_ADDRESS; 1830 } else { 1831 out_sig_bt[argc++] = in_sig_bt[i]; 1832 } 1833 } 1834 } 1835 1836 1837 // Compute the wrapper's frame size. 1838 // -------------------------------------------------------------------------- 1839 1840 // Now figure out where the args must be stored and how much stack space 1841 // they require. 1842 // 1843 // Compute framesize for the wrapper. We need to handlize all oops in 1844 // incoming registers. 1845 // 1846 // Calculate the total number of stack slots we will need: 1847 // 1) abi requirements 1848 // 2) outgoing arguments 1849 // 3) space for inbound oop handle area 1850 // 4) space for handlizing a klass if static method 1851 // 5) space for a lock if synchronized method 1852 // 6) workspace for saving return values, int <-> float reg moves, etc. 1853 // 7) alignment 1854 // 1855 // Layout of the native wrapper frame: 1856 // (stack grows upwards, memory grows downwards) 1857 // 1858 // NW [ABI_REG_ARGS] <-- 1) R1_SP 1859 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset 1860 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset (save area for critical natives) 1861 // klass <-- 4) R1_SP + klass_offset 1862 // lock <-- 5) R1_SP + lock_offset 1863 // [workspace] <-- 6) R1_SP + workspace_offset 1864 // [alignment] (optional) <-- 7) 1865 // caller [JIT_TOP_ABI_48] <-- r_callers_sp 1866 // 1867 // - *_slot_offset Indicates offset from SP in number of stack slots. 1868 // - *_offset Indicates offset from SP in bytes. 1869 1870 int stack_slots = c_calling_convention(out_sig_bt, out_regs, out_regs2, total_c_args) // 1+2) 1871 + SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention. 1872 1873 // Now the space for the inbound oop handle area. 1874 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word; 1875 if (is_critical_native) { 1876 // Critical natives may have to call out so they need a save area 1877 // for register arguments. 1878 int double_slots = 0; 1879 int single_slots = 0; 1880 for (int i = 0; i < total_in_args; i++) { 1881 if (in_regs[i].first()->is_Register()) { 1882 const Register reg = in_regs[i].first()->as_Register(); 1883 switch (in_sig_bt[i]) { 1884 case T_BOOLEAN: 1885 case T_BYTE: 1886 case T_SHORT: 1887 case T_CHAR: 1888 case T_INT: 1889 // Fall through. 1890 case T_ARRAY: 1891 case T_LONG: double_slots++; break; 1892 default: ShouldNotReachHere(); 1893 } 1894 } else if (in_regs[i].first()->is_FloatRegister()) { 1895 switch (in_sig_bt[i]) { 1896 case T_FLOAT: single_slots++; break; 1897 case T_DOUBLE: double_slots++; break; 1898 default: ShouldNotReachHere(); 1899 } 1900 } 1901 } 1902 total_save_slots = double_slots * 2 + round_to(single_slots, 2); // round to even 1903 } 1904 1905 int oop_handle_slot_offset = stack_slots; 1906 stack_slots += total_save_slots; // 3) 1907 1908 int klass_slot_offset = 0; 1909 int klass_offset = -1; 1910 if (method_is_static && !is_critical_native) { // 4) 1911 klass_slot_offset = stack_slots; 1912 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1913 stack_slots += VMRegImpl::slots_per_word; 1914 } 1915 1916 int lock_slot_offset = 0; 1917 int lock_offset = -1; 1918 if (method->is_synchronized()) { // 5) 1919 lock_slot_offset = stack_slots; 1920 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size; 1921 stack_slots += VMRegImpl::slots_per_word; 1922 } 1923 1924 int workspace_slot_offset = stack_slots; // 6) 1925 stack_slots += 2; 1926 1927 // Now compute actual number of stack words we need. 1928 // Rounding to make stack properly aligned. 1929 stack_slots = round_to(stack_slots, // 7) 1930 frame::alignment_in_bytes / VMRegImpl::stack_slot_size); 1931 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size; 1932 1933 1934 // Now we can start generating code. 1935 // -------------------------------------------------------------------------- 1936 1937 intptr_t start_pc = (intptr_t)__ pc(); 1938 intptr_t vep_start_pc; 1939 intptr_t frame_done_pc; 1940 intptr_t oopmap_pc; 1941 1942 Label ic_miss; 1943 Label handle_pending_exception; 1944 1945 Register r_callers_sp = R21; 1946 Register r_temp_1 = R22; 1947 Register r_temp_2 = R23; 1948 Register r_temp_3 = R24; 1949 Register r_temp_4 = R25; 1950 Register r_temp_5 = R26; 1951 Register r_temp_6 = R27; 1952 Register r_return_pc = R28; 1953 1954 Register r_carg1_jnienv = noreg; 1955 Register r_carg2_classorobject = noreg; 1956 if (!is_critical_native) { 1957 r_carg1_jnienv = out_regs[0].first()->as_Register(); 1958 r_carg2_classorobject = out_regs[1].first()->as_Register(); 1959 } 1960 1961 1962 // Generate the Unverified Entry Point (UEP). 1963 // -------------------------------------------------------------------------- 1964 assert(start_pc == (intptr_t)__ pc(), "uep must be at start"); 1965 1966 // Check ic: object class == cached class? 1967 if (!method_is_static) { 1968 Register ic = as_Register(Matcher::inline_cache_reg_encode()); 1969 Register receiver_klass = r_temp_1; 1970 1971 __ cmpdi(CCR0, R3_ARG1, 0); 1972 __ beq(CCR0, ic_miss); 1973 __ verify_oop(R3_ARG1); 1974 __ load_klass(receiver_klass, R3_ARG1); 1975 1976 __ cmpd(CCR0, receiver_klass, ic); 1977 __ bne(CCR0, ic_miss); 1978 } 1979 1980 1981 // Generate the Verified Entry Point (VEP). 1982 // -------------------------------------------------------------------------- 1983 vep_start_pc = (intptr_t)__ pc(); 1984 1985 __ save_LR_CR(r_temp_1); 1986 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame. 1987 __ mr(r_callers_sp, R1_SP); // Remember frame pointer. 1988 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame. 1989 frame_done_pc = (intptr_t)__ pc(); 1990 1991 __ verify_thread(); 1992 1993 // Native nmethod wrappers never take possesion of the oop arguments. 1994 // So the caller will gc the arguments. 1995 // The only thing we need an oopMap for is if the call is static. 1996 // 1997 // An OopMap for lock (and class if static), and one for the VM call itself. 1998 OopMapSet *oop_maps = new OopMapSet(); 1999 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 2000 2001 if (is_critical_native) { 2002 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); 2003 } 2004 2005 // Move arguments from register/stack to register/stack. 2006 // -------------------------------------------------------------------------- 2007 // 2008 // We immediately shuffle the arguments so that for any vm call we have 2009 // to make from here on out (sync slow path, jvmti, etc.) we will have 2010 // captured the oops from our caller and have a valid oopMap for them. 2011 // 2012 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv* 2013 // (derived from JavaThread* which is in R16_thread) and, if static, 2014 // the class mirror instead of a receiver. This pretty much guarantees that 2015 // register layout will not match. We ignore these extra arguments during 2016 // the shuffle. The shuffle is described by the two calling convention 2017 // vectors we have in our possession. We simply walk the java vector to 2018 // get the source locations and the c vector to get the destinations. 2019 2020 // Record sp-based slot for receiver on stack for non-static methods. 2021 int receiver_offset = -1; 2022 2023 // We move the arguments backward because the floating point registers 2024 // destination will always be to a register with a greater or equal 2025 // register number or the stack. 2026 // in is the index of the incoming Java arguments 2027 // out is the index of the outgoing C arguments 2028 2029 #ifdef ASSERT 2030 bool reg_destroyed[RegisterImpl::number_of_registers]; 2031 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 2032 for (int r = 0 ; r < RegisterImpl::number_of_registers ; r++) { 2033 reg_destroyed[r] = false; 2034 } 2035 for (int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++) { 2036 freg_destroyed[f] = false; 2037 } 2038 #endif // ASSERT 2039 2040 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) { 2041 2042 #ifdef ASSERT 2043 if (in_regs[in].first()->is_Register()) { 2044 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!"); 2045 } else if (in_regs[in].first()->is_FloatRegister()) { 2046 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!"); 2047 } 2048 if (out_regs[out].first()->is_Register()) { 2049 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true; 2050 } else if (out_regs[out].first()->is_FloatRegister()) { 2051 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true; 2052 } 2053 if (out_regs2[out].first()->is_Register()) { 2054 reg_destroyed[out_regs2[out].first()->as_Register()->encoding()] = true; 2055 } else if (out_regs2[out].first()->is_FloatRegister()) { 2056 freg_destroyed[out_regs2[out].first()->as_FloatRegister()->encoding()] = true; 2057 } 2058 #endif // ASSERT 2059 2060 switch (in_sig_bt[in]) { 2061 case T_BOOLEAN: 2062 case T_CHAR: 2063 case T_BYTE: 2064 case T_SHORT: 2065 case T_INT: 2066 // Move int and do sign extension. 2067 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2068 break; 2069 case T_LONG: 2070 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2071 break; 2072 case T_ARRAY: 2073 if (is_critical_native) { 2074 int body_arg = out; 2075 out -= 1; // Point to length arg. 2076 unpack_array_argument(masm, in_regs[in], in_elem_bt[in], out_regs[body_arg], out_regs[out], 2077 r_callers_sp, r_temp_1, r_temp_2); 2078 break; 2079 } 2080 case T_OBJECT: 2081 assert(!is_critical_native, "no oop arguments"); 2082 object_move(masm, stack_slots, 2083 oop_map, oop_handle_slot_offset, 2084 ((in == 0) && (!method_is_static)), &receiver_offset, 2085 in_regs[in], out_regs[out], 2086 r_callers_sp, r_temp_1, r_temp_2); 2087 break; 2088 case T_VOID: 2089 break; 2090 case T_FLOAT: 2091 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2092 if (out_regs2[out].first()->is_valid()) { 2093 float_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1); 2094 } 2095 break; 2096 case T_DOUBLE: 2097 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2098 if (out_regs2[out].first()->is_valid()) { 2099 double_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1); 2100 } 2101 break; 2102 case T_ADDRESS: 2103 fatal("found type (T_ADDRESS) in java args"); 2104 break; 2105 default: 2106 ShouldNotReachHere(); 2107 break; 2108 } 2109 } 2110 2111 // Pre-load a static method's oop into ARG2. 2112 // Used both by locking code and the normal JNI call code. 2113 if (method_is_static && !is_critical_native) { 2114 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()), 2115 r_carg2_classorobject); 2116 2117 // Now handlize the static class mirror in carg2. It's known not-null. 2118 __ std(r_carg2_classorobject, klass_offset, R1_SP); 2119 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 2120 __ addi(r_carg2_classorobject, R1_SP, klass_offset); 2121 } 2122 2123 // Get JNIEnv* which is first argument to native. 2124 if (!is_critical_native) { 2125 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset())); 2126 } 2127 2128 // NOTE: 2129 // 2130 // We have all of the arguments setup at this point. 2131 // We MUST NOT touch any outgoing regs from this point on. 2132 // So if we must call out we must push a new frame. 2133 2134 // Get current pc for oopmap, and load it patchable relative to global toc. 2135 oopmap_pc = (intptr_t) __ pc(); 2136 __ calculate_address_from_global_toc(r_return_pc, (address)oopmap_pc, true, true, true, true); 2137 2138 // We use the same pc/oopMap repeatedly when we call out. 2139 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map); 2140 2141 // r_return_pc now has the pc loaded that we will use when we finally call 2142 // to native. 2143 2144 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below. 2145 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register"); 2146 2147 # if 0 2148 // DTrace method entry 2149 # endif 2150 2151 // Lock a synchronized method. 2152 // -------------------------------------------------------------------------- 2153 2154 if (method->is_synchronized()) { 2155 assert(!is_critical_native, "unhandled"); 2156 ConditionRegister r_flag = CCR1; 2157 Register r_oop = r_temp_4; 2158 const Register r_box = r_temp_5; 2159 Label done, locked; 2160 2161 // Load the oop for the object or class. r_carg2_classorobject contains 2162 // either the handlized oop from the incoming arguments or the handlized 2163 // class mirror (if the method is static). 2164 __ ld(r_oop, 0, r_carg2_classorobject); 2165 2166 // Get the lock box slot's address. 2167 __ addi(r_box, R1_SP, lock_offset); 2168 2169 # ifdef ASSERT 2170 if (UseBiasedLocking) { 2171 // Making the box point to itself will make it clear it went unused 2172 // but also be obviously invalid. 2173 __ std(r_box, 0, r_box); 2174 } 2175 # endif // ASSERT 2176 2177 // Try fastpath for locking. 2178 // fast_lock kills r_temp_1, r_temp_2, r_temp_3. 2179 __ compiler_fast_lock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3); 2180 __ beq(r_flag, locked); 2181 2182 // None of the above fast optimizations worked so we have to get into the 2183 // slow case of monitor enter. Inline a special case of call_VM that 2184 // disallows any pending_exception. 2185 2186 // Save argument registers and leave room for C-compatible ABI_REG_ARGS. 2187 int frame_size = frame::abi_reg_args_size + 2188 round_to(total_c_args * wordSize, frame::alignment_in_bytes); 2189 __ mr(R11_scratch1, R1_SP); 2190 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs, out_regs2); 2191 2192 // Do the call. 2193 __ set_last_Java_frame(R11_scratch1, r_return_pc); 2194 assert(r_return_pc->is_nonvolatile(), "expecting return pc to be in non-volatile register"); 2195 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread); 2196 __ reset_last_Java_frame(); 2197 2198 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs, out_regs2); 2199 2200 __ asm_assert_mem8_is_zero(thread_(pending_exception), 2201 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C", 0); 2202 2203 __ bind(locked); 2204 } 2205 2206 2207 // Publish thread state 2208 // -------------------------------------------------------------------------- 2209 2210 // Use that pc we placed in r_return_pc a while back as the current frame anchor. 2211 __ set_last_Java_frame(R1_SP, r_return_pc); 2212 2213 // Transition from _thread_in_Java to _thread_in_native. 2214 __ li(R0, _thread_in_native); 2215 __ release(); 2216 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2217 __ stw(R0, thread_(thread_state)); 2218 if (UseMembar) { 2219 __ fence(); 2220 } 2221 2222 2223 // The JNI call 2224 // -------------------------------------------------------------------------- 2225 #if defined(ABI_ELFv2) 2226 __ call_c(native_func, relocInfo::runtime_call_type); 2227 #else 2228 FunctionDescriptor* fd_native_method = (FunctionDescriptor*) native_func; 2229 __ call_c(fd_native_method, relocInfo::runtime_call_type); 2230 #endif 2231 2232 2233 // Now, we are back from the native code. 2234 2235 2236 // Unpack the native result. 2237 // -------------------------------------------------------------------------- 2238 2239 // For int-types, we do any needed sign-extension required. 2240 // Care must be taken that the return values (R3_RET and F1_RET) 2241 // will survive any VM calls for blocking or unlocking. 2242 // An OOP result (handle) is done specially in the slow-path code. 2243 2244 switch (ret_type) { 2245 case T_VOID: break; // Nothing to do! 2246 case T_FLOAT: break; // Got it where we want it (unless slow-path). 2247 case T_DOUBLE: break; // Got it where we want it (unless slow-path). 2248 case T_LONG: break; // Got it where we want it (unless slow-path). 2249 case T_OBJECT: break; // Really a handle. 2250 // Cannot de-handlize until after reclaiming jvm_lock. 2251 case T_ARRAY: break; 2252 2253 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1) 2254 Label skip_modify; 2255 __ cmpwi(CCR0, R3_RET, 0); 2256 __ beq(CCR0, skip_modify); 2257 __ li(R3_RET, 1); 2258 __ bind(skip_modify); 2259 break; 2260 } 2261 case T_BYTE: { // sign extension 2262 __ extsb(R3_RET, R3_RET); 2263 break; 2264 } 2265 case T_CHAR: { // unsigned result 2266 __ andi(R3_RET, R3_RET, 0xffff); 2267 break; 2268 } 2269 case T_SHORT: { // sign extension 2270 __ extsh(R3_RET, R3_RET); 2271 break; 2272 } 2273 case T_INT: // nothing to do 2274 break; 2275 default: 2276 ShouldNotReachHere(); 2277 break; 2278 } 2279 2280 2281 // Publish thread state 2282 // -------------------------------------------------------------------------- 2283 2284 // Switch thread to "native transition" state before reading the 2285 // synchronization state. This additional state is necessary because reading 2286 // and testing the synchronization state is not atomic w.r.t. GC, as this 2287 // scenario demonstrates: 2288 // - Java thread A, in _thread_in_native state, loads _not_synchronized 2289 // and is preempted. 2290 // - VM thread changes sync state to synchronizing and suspends threads 2291 // for GC. 2292 // - Thread A is resumed to finish this native method, but doesn't block 2293 // here since it didn't see any synchronization in progress, and escapes. 2294 2295 // Transition from _thread_in_native to _thread_in_native_trans. 2296 __ li(R0, _thread_in_native_trans); 2297 __ release(); 2298 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2299 __ stw(R0, thread_(thread_state)); 2300 2301 2302 // Must we block? 2303 // -------------------------------------------------------------------------- 2304 2305 // Block, if necessary, before resuming in _thread_in_Java state. 2306 // In order for GC to work, don't clear the last_Java_sp until after blocking. 2307 Label after_transition; 2308 { 2309 Label no_block, sync; 2310 2311 if (os::is_MP()) { 2312 if (UseMembar) { 2313 // Force this write out before the read below. 2314 __ fence(); 2315 } else { 2316 // Write serialization page so VM thread can do a pseudo remote membar. 2317 // We use the current thread pointer to calculate a thread specific 2318 // offset to write to within the page. This minimizes bus traffic 2319 // due to cache line collision. 2320 __ serialize_memory(R16_thread, r_temp_4, r_temp_5); 2321 } 2322 } 2323 2324 Register sync_state_addr = r_temp_4; 2325 Register sync_state = r_temp_5; 2326 Register suspend_flags = r_temp_6; 2327 2328 __ load_const(sync_state_addr, SafepointSynchronize::address_of_state(), /*temp*/ sync_state); 2329 2330 // TODO: PPC port assert(4 == SafepointSynchronize::sz_state(), "unexpected field size"); 2331 __ lwz(sync_state, 0, sync_state_addr); 2332 2333 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size"); 2334 __ lwz(suspend_flags, thread_(suspend_flags)); 2335 2336 __ acquire(); 2337 2338 Label do_safepoint; 2339 // No synchronization in progress nor yet synchronized. 2340 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized); 2341 // Not suspended. 2342 __ cmpwi(CCR1, suspend_flags, 0); 2343 2344 __ bne(CCR0, sync); 2345 __ beq(CCR1, no_block); 2346 2347 // Block. Save any potential method result value before the operation and 2348 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this 2349 // lets us share the oopMap we used when we went native rather than create 2350 // a distinct one for this pc. 2351 __ bind(sync); 2352 2353 address entry_point = is_critical_native 2354 ? CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition) 2355 : CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans); 2356 save_native_result(masm, ret_type, workspace_slot_offset); 2357 __ call_VM_leaf(entry_point, R16_thread); 2358 restore_native_result(masm, ret_type, workspace_slot_offset); 2359 2360 if (is_critical_native) { 2361 __ b(after_transition); // No thread state transition here. 2362 } 2363 __ bind(no_block); 2364 } 2365 2366 // Publish thread state. 2367 // -------------------------------------------------------------------------- 2368 2369 // Thread state is thread_in_native_trans. Any safepoint blocking has 2370 // already happened so we can now change state to _thread_in_Java. 2371 2372 // Transition from _thread_in_native_trans to _thread_in_Java. 2373 __ li(R0, _thread_in_Java); 2374 __ release(); 2375 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2376 __ stw(R0, thread_(thread_state)); 2377 if (UseMembar) { 2378 __ fence(); 2379 } 2380 __ bind(after_transition); 2381 2382 // Reguard any pages if necessary. 2383 // -------------------------------------------------------------------------- 2384 2385 Label no_reguard; 2386 __ lwz(r_temp_1, thread_(stack_guard_state)); 2387 __ cmpwi(CCR0, r_temp_1, JavaThread::stack_guard_yellow_disabled); 2388 __ bne(CCR0, no_reguard); 2389 2390 save_native_result(masm, ret_type, workspace_slot_offset); 2391 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); 2392 restore_native_result(masm, ret_type, workspace_slot_offset); 2393 2394 __ bind(no_reguard); 2395 2396 2397 // Unlock 2398 // -------------------------------------------------------------------------- 2399 2400 if (method->is_synchronized()) { 2401 2402 ConditionRegister r_flag = CCR1; 2403 const Register r_oop = r_temp_4; 2404 const Register r_box = r_temp_5; 2405 const Register r_exception = r_temp_6; 2406 Label done; 2407 2408 // Get oop and address of lock object box. 2409 if (method_is_static) { 2410 assert(klass_offset != -1, ""); 2411 __ ld(r_oop, klass_offset, R1_SP); 2412 } else { 2413 assert(receiver_offset != -1, ""); 2414 __ ld(r_oop, receiver_offset, R1_SP); 2415 } 2416 __ addi(r_box, R1_SP, lock_offset); 2417 2418 // Try fastpath for unlocking. 2419 __ compiler_fast_unlock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3); 2420 __ beq(r_flag, done); 2421 2422 // Save and restore any potential method result value around the unlocking operation. 2423 save_native_result(masm, ret_type, workspace_slot_offset); 2424 2425 // Must save pending exception around the slow-path VM call. Since it's a 2426 // leaf call, the pending exception (if any) can be kept in a register. 2427 __ ld(r_exception, thread_(pending_exception)); 2428 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile"); 2429 __ li(R0, 0); 2430 __ std(R0, thread_(pending_exception)); 2431 2432 // Slow case of monitor enter. 2433 // Inline a special case of call_VM that disallows any pending_exception. 2434 // Arguments are (oop obj, BasicLock* lock, JavaThread* thread). 2435 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box, R16_thread); 2436 2437 __ asm_assert_mem8_is_zero(thread_(pending_exception), 2438 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C", 0); 2439 2440 restore_native_result(masm, ret_type, workspace_slot_offset); 2441 2442 // Check_forward_pending_exception jump to forward_exception if any pending 2443 // exception is set. The forward_exception routine expects to see the 2444 // exception in pending_exception and not in a register. Kind of clumsy, 2445 // since all folks who branch to forward_exception must have tested 2446 // pending_exception first and hence have it in a register already. 2447 __ std(r_exception, thread_(pending_exception)); 2448 2449 __ bind(done); 2450 } 2451 2452 # if 0 2453 // DTrace method exit 2454 # endif 2455 2456 // Clear "last Java frame" SP and PC. 2457 // -------------------------------------------------------------------------- 2458 2459 __ reset_last_Java_frame(); 2460 2461 // Unpack oop result. 2462 // -------------------------------------------------------------------------- 2463 2464 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2465 Label skip_unboxing; 2466 __ cmpdi(CCR0, R3_RET, 0); 2467 __ beq(CCR0, skip_unboxing); 2468 __ ld(R3_RET, 0, R3_RET); 2469 __ bind(skip_unboxing); 2470 __ verify_oop(R3_RET); 2471 } 2472 2473 2474 // Reset handle block. 2475 // -------------------------------------------------------------------------- 2476 if (!is_critical_native) { 2477 __ ld(r_temp_1, thread_(active_handles)); 2478 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size"); 2479 __ li(r_temp_2, 0); 2480 __ stw(r_temp_2, JNIHandleBlock::top_offset_in_bytes(), r_temp_1); 2481 2482 2483 // Check for pending exceptions. 2484 // -------------------------------------------------------------------------- 2485 __ ld(r_temp_2, thread_(pending_exception)); 2486 __ cmpdi(CCR0, r_temp_2, 0); 2487 __ bne(CCR0, handle_pending_exception); 2488 } 2489 2490 // Return 2491 // -------------------------------------------------------------------------- 2492 2493 __ pop_frame(); 2494 __ restore_LR_CR(R11); 2495 __ blr(); 2496 2497 2498 // Handler for pending exceptions (out-of-line). 2499 // -------------------------------------------------------------------------- 2500 2501 // Since this is a native call, we know the proper exception handler 2502 // is the empty function. We just pop this frame and then jump to 2503 // forward_exception_entry. 2504 if (!is_critical_native) { 2505 __ align(InteriorEntryAlignment); 2506 __ bind(handle_pending_exception); 2507 2508 __ pop_frame(); 2509 __ restore_LR_CR(R11); 2510 __ b64_patchable((address)StubRoutines::forward_exception_entry(), 2511 relocInfo::runtime_call_type); 2512 } 2513 2514 // Handler for a cache miss (out-of-line). 2515 // -------------------------------------------------------------------------- 2516 2517 if (!method_is_static) { 2518 __ align(InteriorEntryAlignment); 2519 __ bind(ic_miss); 2520 2521 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 2522 relocInfo::runtime_call_type); 2523 } 2524 2525 // Done. 2526 // -------------------------------------------------------------------------- 2527 2528 __ flush(); 2529 2530 nmethod *nm = nmethod::new_native_nmethod(method, 2531 compile_id, 2532 masm->code(), 2533 vep_start_pc-start_pc, 2534 frame_done_pc-start_pc, 2535 stack_slots / VMRegImpl::slots_per_word, 2536 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2537 in_ByteSize(lock_offset), 2538 oop_maps); 2539 2540 if (is_critical_native) { 2541 nm->set_lazy_critical_native(true); 2542 } 2543 2544 return nm; 2545 #else 2546 ShouldNotReachHere(); 2547 return NULL; 2548 #endif // COMPILER2 2549 } 2550 2551 // This function returns the adjust size (in number of words) to a c2i adapter 2552 // activation for use during deoptimization. 2553 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2554 return round_to((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::alignment_in_bytes); 2555 } 2556 2557 uint SharedRuntime::out_preserve_stack_slots() { 2558 #if defined(COMPILER1) || defined(COMPILER2) 2559 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size; 2560 #else 2561 return 0; 2562 #endif 2563 } 2564 2565 #ifdef COMPILER2 2566 // Frame generation for deopt and uncommon trap blobs. 2567 static void push_skeleton_frame(MacroAssembler* masm, bool deopt, 2568 /* Read */ 2569 Register unroll_block_reg, 2570 /* Update */ 2571 Register frame_sizes_reg, 2572 Register number_of_frames_reg, 2573 Register pcs_reg, 2574 /* Invalidate */ 2575 Register frame_size_reg, 2576 Register pc_reg) { 2577 2578 __ ld(pc_reg, 0, pcs_reg); 2579 __ ld(frame_size_reg, 0, frame_sizes_reg); 2580 __ std(pc_reg, _abi(lr), R1_SP); 2581 __ push_frame(frame_size_reg, R0/*tmp*/); 2582 #ifdef CC_INTERP 2583 __ std(R1_SP, _parent_ijava_frame_abi(initial_caller_sp), R1_SP); 2584 #else 2585 #ifdef ASSERT 2586 __ load_const_optimized(pc_reg, 0x5afe); 2587 __ std(pc_reg, _ijava_state_neg(ijava_reserved), R1_SP); 2588 #endif 2589 __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP); 2590 #endif // CC_INTERP 2591 __ addi(number_of_frames_reg, number_of_frames_reg, -1); 2592 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize); 2593 __ addi(pcs_reg, pcs_reg, wordSize); 2594 } 2595 2596 // Loop through the UnrollBlock info and create new frames. 2597 static void push_skeleton_frames(MacroAssembler* masm, bool deopt, 2598 /* read */ 2599 Register unroll_block_reg, 2600 /* invalidate */ 2601 Register frame_sizes_reg, 2602 Register number_of_frames_reg, 2603 Register pcs_reg, 2604 Register frame_size_reg, 2605 Register pc_reg) { 2606 Label loop; 2607 2608 // _number_of_frames is of type int (deoptimization.hpp) 2609 __ lwa(number_of_frames_reg, 2610 Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(), 2611 unroll_block_reg); 2612 __ ld(pcs_reg, 2613 Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(), 2614 unroll_block_reg); 2615 __ ld(frame_sizes_reg, 2616 Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(), 2617 unroll_block_reg); 2618 2619 // stack: (caller_of_deoptee, ...). 2620 2621 // At this point we either have an interpreter frame or a compiled 2622 // frame on top of stack. If it is a compiled frame we push a new c2i 2623 // adapter here 2624 2625 // Memorize top-frame stack-pointer. 2626 __ mr(frame_size_reg/*old_sp*/, R1_SP); 2627 2628 // Resize interpreter top frame OR C2I adapter. 2629 2630 // At this moment, the top frame (which is the caller of the deoptee) is 2631 // an interpreter frame or a newly pushed C2I adapter or an entry frame. 2632 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the 2633 // outgoing arguments. 2634 // 2635 // In order to push the interpreter frame for the deoptee, we need to 2636 // resize the top frame such that we are able to place the deoptee's 2637 // locals in the frame. 2638 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI 2639 // into a valid PARENT_IJAVA_FRAME_ABI. 2640 2641 __ lwa(R11_scratch1, 2642 Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(), 2643 unroll_block_reg); 2644 __ neg(R11_scratch1, R11_scratch1); 2645 2646 // R11_scratch1 contains size of locals for frame resizing. 2647 // R12_scratch2 contains top frame's lr. 2648 2649 // Resize frame by complete frame size prevents TOC from being 2650 // overwritten by locals. A more stack space saving way would be 2651 // to copy the TOC to its location in the new abi. 2652 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size); 2653 2654 // now, resize the frame 2655 __ resize_frame(R11_scratch1, pc_reg/*tmp*/); 2656 2657 // In the case where we have resized a c2i frame above, the optional 2658 // alignment below the locals has size 32 (why?). 2659 __ std(R12_scratch2, _abi(lr), R1_SP); 2660 2661 // Initialize initial_caller_sp. 2662 #ifdef CC_INTERP 2663 __ std(frame_size_reg/*old_sp*/, _parent_ijava_frame_abi(initial_caller_sp), R1_SP); 2664 #else 2665 #ifdef ASSERT 2666 __ load_const_optimized(pc_reg, 0x5afe); 2667 __ std(pc_reg, _ijava_state_neg(ijava_reserved), R1_SP); 2668 #endif 2669 __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP); 2670 #endif // CC_INTERP 2671 2672 #ifdef ASSERT 2673 // Make sure that there is at least one entry in the array. 2674 __ cmpdi(CCR0, number_of_frames_reg, 0); 2675 __ asm_assert_ne("array_size must be > 0", 0x205); 2676 #endif 2677 2678 // Now push the new interpreter frames. 2679 // 2680 __ bind(loop); 2681 // Allocate a new frame, fill in the pc. 2682 push_skeleton_frame(masm, deopt, 2683 unroll_block_reg, 2684 frame_sizes_reg, 2685 number_of_frames_reg, 2686 pcs_reg, 2687 frame_size_reg, 2688 pc_reg); 2689 __ cmpdi(CCR0, number_of_frames_reg, 0); 2690 __ bne(CCR0, loop); 2691 2692 // Get the return address pointing into the frame manager. 2693 __ ld(R0, 0, pcs_reg); 2694 // Store it in the top interpreter frame. 2695 __ std(R0, _abi(lr), R1_SP); 2696 // Initialize frame_manager_lr of interpreter top frame. 2697 #ifdef CC_INTERP 2698 __ std(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP); 2699 #endif 2700 } 2701 #endif 2702 2703 void SharedRuntime::generate_deopt_blob() { 2704 // Allocate space for the code 2705 ResourceMark rm; 2706 // Setup code generation tools 2707 CodeBuffer buffer("deopt_blob", 2048, 1024); 2708 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer); 2709 Label exec_mode_initialized; 2710 int frame_size_in_words; 2711 OopMap* map = NULL; 2712 OopMapSet *oop_maps = new OopMapSet(); 2713 2714 // size of ABI112 plus spill slots for R3_RET and F1_RET. 2715 const int frame_size_in_bytes = frame::abi_reg_args_spill_size; 2716 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint); 2717 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info. 2718 2719 const Register exec_mode_reg = R21_tmp1; 2720 2721 const address start = __ pc(); 2722 2723 #ifdef COMPILER2 2724 // -------------------------------------------------------------------------- 2725 // Prolog for non exception case! 2726 2727 // We have been called from the deopt handler of the deoptee. 2728 // 2729 // deoptee: 2730 // ... 2731 // call X 2732 // ... 2733 // deopt_handler: call_deopt_stub 2734 // cur. return pc --> ... 2735 // 2736 // So currently SR_LR points behind the call in the deopt handler. 2737 // We adjust it such that it points to the start of the deopt handler. 2738 // The return_pc has been stored in the frame of the deoptee and 2739 // will replace the address of the deopt_handler in the call 2740 // to Deoptimization::fetch_unroll_info below. 2741 // We can't grab a free register here, because all registers may 2742 // contain live values, so let the RegisterSaver do the adjustment 2743 // of the return pc. 2744 const int return_pc_adjustment_no_exception = -HandlerImpl::size_deopt_handler(); 2745 2746 // Push the "unpack frame" 2747 // Save everything in sight. 2748 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 2749 &first_frame_size_in_bytes, 2750 /*generate_oop_map=*/ true, 2751 return_pc_adjustment_no_exception, 2752 RegisterSaver::return_pc_is_lr); 2753 assert(map != NULL, "OopMap must have been created"); 2754 2755 __ li(exec_mode_reg, Deoptimization::Unpack_deopt); 2756 // Save exec mode for unpack_frames. 2757 __ b(exec_mode_initialized); 2758 2759 // -------------------------------------------------------------------------- 2760 // Prolog for exception case 2761 2762 // An exception is pending. 2763 // We have been called with a return (interpreter) or a jump (exception blob). 2764 // 2765 // - R3_ARG1: exception oop 2766 // - R4_ARG2: exception pc 2767 2768 int exception_offset = __ pc() - start; 2769 2770 BLOCK_COMMENT("Prolog for exception case"); 2771 2772 // The RegisterSaves doesn't need to adjust the return pc for this situation. 2773 const int return_pc_adjustment_exception = 0; 2774 2775 // Push the "unpack frame". 2776 // Save everything in sight. 2777 assert(R4 == R4_ARG2, "exception pc must be in r4"); 2778 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 2779 &first_frame_size_in_bytes, 2780 /*generate_oop_map=*/ false, 2781 return_pc_adjustment_exception, 2782 RegisterSaver::return_pc_is_r4); 2783 2784 // Deopt during an exception. Save exec mode for unpack_frames. 2785 __ li(exec_mode_reg, Deoptimization::Unpack_exception); 2786 2787 // Store exception oop and pc in thread (location known to GC). 2788 // This is needed since the call to "fetch_unroll_info()" may safepoint. 2789 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2790 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2791 2792 // fall through 2793 2794 // -------------------------------------------------------------------------- 2795 __ BIND(exec_mode_initialized); 2796 2797 { 2798 const Register unroll_block_reg = R22_tmp2; 2799 2800 // We need to set `last_Java_frame' because `fetch_unroll_info' will 2801 // call `last_Java_frame()'. The value of the pc in the frame is not 2802 // particularly important. It just needs to identify this blob. 2803 __ set_last_Java_frame(R1_SP, noreg); 2804 2805 // With EscapeAnalysis turned on, this call may safepoint! 2806 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread); 2807 address calls_return_pc = __ last_calls_return_pc(); 2808 // Set an oopmap for the call site that describes all our saved registers. 2809 oop_maps->add_gc_map(calls_return_pc - start, map); 2810 2811 __ reset_last_Java_frame(); 2812 // Save the return value. 2813 __ mr(unroll_block_reg, R3_RET); 2814 2815 // Restore only the result registers that have been saved 2816 // by save_volatile_registers(...). 2817 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes); 2818 2819 // In excp_deopt_mode, restore and clear exception oop which we 2820 // stored in the thread during exception entry above. The exception 2821 // oop will be the return value of this stub. 2822 Label skip_restore_excp; 2823 __ cmpdi(CCR0, exec_mode_reg, Deoptimization::Unpack_exception); 2824 __ bne(CCR0, skip_restore_excp); 2825 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2826 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2827 __ li(R0, 0); 2828 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2829 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2830 __ BIND(skip_restore_excp); 2831 2832 __ pop_frame(); 2833 2834 // stack: (deoptee, optional i2c, caller of deoptee, ...). 2835 2836 // pop the deoptee's frame 2837 __ pop_frame(); 2838 2839 // stack: (caller_of_deoptee, ...). 2840 2841 // Loop through the `UnrollBlock' info and create interpreter frames. 2842 push_skeleton_frames(masm, true/*deopt*/, 2843 unroll_block_reg, 2844 R23_tmp3, 2845 R24_tmp4, 2846 R25_tmp5, 2847 R26_tmp6, 2848 R27_tmp7); 2849 2850 // stack: (skeletal interpreter frame, ..., optional skeletal 2851 // interpreter frame, optional c2i, caller of deoptee, ...). 2852 } 2853 2854 // push an `unpack_frame' taking care of float / int return values. 2855 __ push_frame(frame_size_in_bytes, R0/*tmp*/); 2856 2857 // stack: (unpack frame, skeletal interpreter frame, ..., optional 2858 // skeletal interpreter frame, optional c2i, caller of deoptee, 2859 // ...). 2860 2861 // Spill live volatile registers since we'll do a call. 2862 __ std( R3_RET, _abi_reg_args_spill(spill_ret), R1_SP); 2863 __ stfd(F1_RET, _abi_reg_args_spill(spill_fret), R1_SP); 2864 2865 // Let the unpacker layout information in the skeletal frames just 2866 // allocated. 2867 __ get_PC_trash_LR(R3_RET); 2868 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET); 2869 // This is a call to a LEAF method, so no oop map is required. 2870 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), 2871 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/); 2872 __ reset_last_Java_frame(); 2873 2874 // Restore the volatiles saved above. 2875 __ ld( R3_RET, _abi_reg_args_spill(spill_ret), R1_SP); 2876 __ lfd(F1_RET, _abi_reg_args_spill(spill_fret), R1_SP); 2877 2878 // Pop the unpack frame. 2879 __ pop_frame(); 2880 __ restore_LR_CR(R0); 2881 2882 // stack: (top interpreter frame, ..., optional interpreter frame, 2883 // optional c2i, caller of deoptee, ...). 2884 2885 // Initialize R14_state. 2886 #ifdef CC_INTERP 2887 __ ld(R14_state, 0, R1_SP); 2888 __ addi(R14_state, R14_state, -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 2889 // Also inititialize R15_prev_state. 2890 __ restore_prev_state(); 2891 #else 2892 __ restore_interpreter_state(R11_scratch1); 2893 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 2894 #endif // CC_INTERP 2895 2896 2897 // Return to the interpreter entry point. 2898 __ blr(); 2899 __ flush(); 2900 #else // COMPILER2 2901 __ unimplemented("deopt blob needed only with compiler"); 2902 int exception_offset = __ pc() - start; 2903 #endif // COMPILER2 2904 2905 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, 0, first_frame_size_in_bytes / wordSize); 2906 } 2907 2908 #ifdef COMPILER2 2909 void SharedRuntime::generate_uncommon_trap_blob() { 2910 // Allocate space for the code. 2911 ResourceMark rm; 2912 // Setup code generation tools. 2913 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2914 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer); 2915 address start = __ pc(); 2916 2917 Register unroll_block_reg = R21_tmp1; 2918 Register klass_index_reg = R22_tmp2; 2919 Register unc_trap_reg = R23_tmp3; 2920 2921 OopMapSet* oop_maps = new OopMapSet(); 2922 int frame_size_in_bytes = frame::abi_reg_args_size; 2923 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0); 2924 2925 // stack: (deoptee, optional i2c, caller_of_deoptee, ...). 2926 2927 // Push a dummy `unpack_frame' and call 2928 // `Deoptimization::uncommon_trap' to pack the compiled frame into a 2929 // vframe array and return the `UnrollBlock' information. 2930 2931 // Save LR to compiled frame. 2932 __ save_LR_CR(R11_scratch1); 2933 2934 // Push an "uncommon_trap" frame. 2935 __ push_frame_reg_args(0, R11_scratch1); 2936 2937 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...). 2938 2939 // Set the `unpack_frame' as last_Java_frame. 2940 // `Deoptimization::uncommon_trap' expects it and considers its 2941 // sender frame as the deoptee frame. 2942 // Remember the offset of the instruction whose address will be 2943 // moved to R11_scratch1. 2944 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1); 2945 2946 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 2947 2948 __ mr(klass_index_reg, R3); 2949 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), 2950 R16_thread, klass_index_reg); 2951 2952 // Set an oopmap for the call site. 2953 oop_maps->add_gc_map(gc_map_pc - start, map); 2954 2955 __ reset_last_Java_frame(); 2956 2957 // Pop the `unpack frame'. 2958 __ pop_frame(); 2959 2960 // stack: (deoptee, optional i2c, caller_of_deoptee, ...). 2961 2962 // Save the return value. 2963 __ mr(unroll_block_reg, R3_RET); 2964 2965 // Pop the uncommon_trap frame. 2966 __ pop_frame(); 2967 2968 // stack: (caller_of_deoptee, ...). 2969 2970 // Allocate new interpreter frame(s) and possibly a c2i adapter 2971 // frame. 2972 push_skeleton_frames(masm, false/*deopt*/, 2973 unroll_block_reg, 2974 R22_tmp2, 2975 R23_tmp3, 2976 R24_tmp4, 2977 R25_tmp5, 2978 R26_tmp6); 2979 2980 // stack: (skeletal interpreter frame, ..., optional skeletal 2981 // interpreter frame, optional c2i, caller of deoptee, ...). 2982 2983 // Push a dummy `unpack_frame' taking care of float return values. 2984 // Call `Deoptimization::unpack_frames' to layout information in the 2985 // interpreter frames just created. 2986 2987 // Push a simple "unpack frame" here. 2988 __ push_frame_reg_args(0, R11_scratch1); 2989 2990 // stack: (unpack frame, skeletal interpreter frame, ..., optional 2991 // skeletal interpreter frame, optional c2i, caller of deoptee, 2992 // ...). 2993 2994 // Set the "unpack_frame" as last_Java_frame. 2995 __ get_PC_trash_LR(R11_scratch1); 2996 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 2997 2998 // Indicate it is the uncommon trap case. 2999 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap); 3000 // Let the unpacker layout information in the skeletal frames just 3001 // allocated. 3002 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), 3003 R16_thread, unc_trap_reg); 3004 3005 __ reset_last_Java_frame(); 3006 // Pop the `unpack frame'. 3007 __ pop_frame(); 3008 // Restore LR from top interpreter frame. 3009 __ restore_LR_CR(R11_scratch1); 3010 3011 // stack: (top interpreter frame, ..., optional interpreter frame, 3012 // optional c2i, caller of deoptee, ...). 3013 3014 #ifdef CC_INTERP 3015 // Initialize R14_state, ... 3016 __ ld(R11_scratch1, 0, R1_SP); 3017 __ addi(R14_state, R11_scratch1, -frame::interpreter_frame_cinterpreterstate_size_in_bytes()); 3018 // also initialize R15_prev_state. 3019 __ restore_prev_state(); 3020 #else 3021 __ restore_interpreter_state(R11_scratch1); 3022 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 3023 #endif // CC_INTERP 3024 3025 // Return to the interpreter entry point. 3026 __ blr(); 3027 3028 masm->flush(); 3029 3030 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize); 3031 } 3032 #endif // COMPILER2 3033 3034 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap. 3035 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 3036 assert(StubRoutines::forward_exception_entry() != NULL, 3037 "must be generated before"); 3038 3039 ResourceMark rm; 3040 OopMapSet *oop_maps = new OopMapSet(); 3041 OopMap* map; 3042 3043 // Allocate space for the code. Setup code generation tools. 3044 CodeBuffer buffer("handler_blob", 2048, 1024); 3045 MacroAssembler* masm = new MacroAssembler(&buffer); 3046 3047 address start = __ pc(); 3048 int frame_size_in_bytes = 0; 3049 3050 RegisterSaver::ReturnPCLocation return_pc_location; 3051 bool cause_return = (poll_type == POLL_AT_RETURN); 3052 if (cause_return) { 3053 // Nothing to do here. The frame has already been popped in MachEpilogNode. 3054 // Register LR already contains the return pc. 3055 return_pc_location = RegisterSaver::return_pc_is_lr; 3056 } else { 3057 // Use thread()->saved_exception_pc() as return pc. 3058 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc; 3059 } 3060 3061 // Save registers, fpu state, and flags. 3062 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 3063 &frame_size_in_bytes, 3064 /*generate_oop_map=*/ true, 3065 /*return_pc_adjustment=*/0, 3066 return_pc_location); 3067 3068 // The following is basically a call_VM. However, we need the precise 3069 // address of the call in order to generate an oopmap. Hence, we do all the 3070 // work outselves. 3071 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg); 3072 3073 // The return address must always be correct so that the frame constructor 3074 // never sees an invalid pc. 3075 3076 // Do the call 3077 __ call_VM_leaf(call_ptr, R16_thread); 3078 address calls_return_pc = __ last_calls_return_pc(); 3079 3080 // Set an oopmap for the call site. This oopmap will map all 3081 // oop-registers and debug-info registers as callee-saved. This 3082 // will allow deoptimization at this safepoint to find all possible 3083 // debug-info recordings, as well as let GC find all oops. 3084 oop_maps->add_gc_map(calls_return_pc - start, map); 3085 3086 Label noException; 3087 3088 // Clear the last Java frame. 3089 __ reset_last_Java_frame(); 3090 3091 BLOCK_COMMENT(" Check pending exception."); 3092 const Register pending_exception = R0; 3093 __ ld(pending_exception, thread_(pending_exception)); 3094 __ cmpdi(CCR0, pending_exception, 0); 3095 __ beq(CCR0, noException); 3096 3097 // Exception pending 3098 RegisterSaver::restore_live_registers_and_pop_frame(masm, 3099 frame_size_in_bytes, 3100 /*restore_ctr=*/true); 3101 3102 BLOCK_COMMENT(" Jump to forward_exception_entry."); 3103 // Jump to forward_exception_entry, with the issuing PC in LR 3104 // so it looks like the original nmethod called forward_exception_entry. 3105 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 3106 3107 // No exception case. 3108 __ BIND(noException); 3109 3110 3111 // Normal exit, restore registers and exit. 3112 RegisterSaver::restore_live_registers_and_pop_frame(masm, 3113 frame_size_in_bytes, 3114 /*restore_ctr=*/true); 3115 3116 __ blr(); 3117 3118 // Make sure all code is generated 3119 masm->flush(); 3120 3121 // Fill-out other meta info 3122 // CodeBlob frame size is in words. 3123 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize); 3124 } 3125 3126 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss) 3127 // 3128 // Generate a stub that calls into the vm to find out the proper destination 3129 // of a java call. All the argument registers are live at this point 3130 // but since this is generic code we don't know what they are and the caller 3131 // must do any gc of the args. 3132 // 3133 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3134 3135 // allocate space for the code 3136 ResourceMark rm; 3137 3138 CodeBuffer buffer(name, 1000, 512); 3139 MacroAssembler* masm = new MacroAssembler(&buffer); 3140 3141 int frame_size_in_bytes; 3142 3143 OopMapSet *oop_maps = new OopMapSet(); 3144 OopMap* map = NULL; 3145 3146 address start = __ pc(); 3147 3148 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 3149 &frame_size_in_bytes, 3150 /*generate_oop_map*/ true, 3151 /*return_pc_adjustment*/ 0, 3152 RegisterSaver::return_pc_is_lr); 3153 3154 // Use noreg as last_Java_pc, the return pc will be reconstructed 3155 // from the physical frame. 3156 __ set_last_Java_frame(/*sp*/R1_SP, noreg); 3157 3158 int frame_complete = __ offset(); 3159 3160 // Pass R19_method as 2nd (optional) argument, used by 3161 // counter_overflow_stub. 3162 __ call_VM_leaf(destination, R16_thread, R19_method); 3163 address calls_return_pc = __ last_calls_return_pc(); 3164 // Set an oopmap for the call site. 3165 // We need this not only for callee-saved registers, but also for volatile 3166 // registers that the compiler might be keeping live across a safepoint. 3167 // Create the oopmap for the call's return pc. 3168 oop_maps->add_gc_map(calls_return_pc - start, map); 3169 3170 // R3_RET contains the address we are going to jump to assuming no exception got installed. 3171 3172 // clear last_Java_sp 3173 __ reset_last_Java_frame(); 3174 3175 // Check for pending exceptions. 3176 BLOCK_COMMENT("Check for pending exceptions."); 3177 Label pending; 3178 __ ld(R11_scratch1, thread_(pending_exception)); 3179 __ cmpdi(CCR0, R11_scratch1, 0); 3180 __ bne(CCR0, pending); 3181 3182 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame. 3183 3184 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false); 3185 3186 // Get the returned method. 3187 __ get_vm_result_2(R19_method); 3188 3189 __ bctr(); 3190 3191 3192 // Pending exception after the safepoint. 3193 __ BIND(pending); 3194 3195 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true); 3196 3197 // exception pending => remove activation and forward to exception handler 3198 3199 __ li(R11_scratch1, 0); 3200 __ ld(R3_ARG1, thread_(pending_exception)); 3201 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_offset()), R16_thread); 3202 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 3203 3204 // ------------- 3205 // Make sure all code is generated. 3206 masm->flush(); 3207 3208 // return the blob 3209 // frame_size_words or bytes?? 3210 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize, 3211 oop_maps, true); 3212 }