1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * Copyright (c) 2012, 2015 SAP SE. 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 // Note, MaxVectorSize == 8 on PPC64. 479 assert(size <= 8, "%d bytes vectors are not supported", size); 480 return size > 8; 481 } 482 #ifdef COMPILER2 483 static int reg2slot(VMReg r) { 484 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 485 } 486 487 static int reg2offset(VMReg r) { 488 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 489 } 490 #endif 491 492 // --------------------------------------------------------------------------- 493 // Read the array of BasicTypes from a signature, and compute where the 494 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 495 // quantities. Values less than VMRegImpl::stack0 are registers, those above 496 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 497 // as framesizes are fixed. 498 // VMRegImpl::stack0 refers to the first slot 0(sp). 499 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register 500 // up to RegisterImpl::number_of_registers) are the 64-bit 501 // integer registers. 502 503 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 504 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 505 // units regardless of build. Of course for i486 there is no 64 bit build 506 507 // The Java calling convention is a "shifted" version of the C ABI. 508 // By skipping the first C ABI register we can call non-static jni methods 509 // with small numbers of arguments without having to shuffle the arguments 510 // at all. Since we control the java ABI we ought to at least get some 511 // advantage out of it. 512 513 const VMReg java_iarg_reg[8] = { 514 R3->as_VMReg(), 515 R4->as_VMReg(), 516 R5->as_VMReg(), 517 R6->as_VMReg(), 518 R7->as_VMReg(), 519 R8->as_VMReg(), 520 R9->as_VMReg(), 521 R10->as_VMReg() 522 }; 523 524 const VMReg java_farg_reg[13] = { 525 F1->as_VMReg(), 526 F2->as_VMReg(), 527 F3->as_VMReg(), 528 F4->as_VMReg(), 529 F5->as_VMReg(), 530 F6->as_VMReg(), 531 F7->as_VMReg(), 532 F8->as_VMReg(), 533 F9->as_VMReg(), 534 F10->as_VMReg(), 535 F11->as_VMReg(), 536 F12->as_VMReg(), 537 F13->as_VMReg() 538 }; 539 540 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]); 541 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]); 542 543 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 544 VMRegPair *regs, 545 int total_args_passed, 546 int is_outgoing) { 547 // C2c calling conventions for compiled-compiled calls. 548 // Put 8 ints/longs into registers _AND_ 13 float/doubles into 549 // registers _AND_ put the rest on the stack. 550 551 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats 552 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles 553 554 int i; 555 VMReg reg; 556 int stk = 0; 557 int ireg = 0; 558 int freg = 0; 559 560 // We put the first 8 arguments into registers and the rest on the 561 // stack, float arguments are already in their argument registers 562 // due to c2c calling conventions (see calling_convention). 563 for (int i = 0; i < total_args_passed; ++i) { 564 switch(sig_bt[i]) { 565 case T_BOOLEAN: 566 case T_CHAR: 567 case T_BYTE: 568 case T_SHORT: 569 case T_INT: 570 if (ireg < num_java_iarg_registers) { 571 // Put int/ptr in register 572 reg = java_iarg_reg[ireg]; 573 ++ireg; 574 } else { 575 // Put int/ptr on stack. 576 reg = VMRegImpl::stack2reg(stk); 577 stk += inc_stk_for_intfloat; 578 } 579 regs[i].set1(reg); 580 break; 581 case T_LONG: 582 assert(sig_bt[i+1] == T_VOID, "expecting half"); 583 if (ireg < num_java_iarg_registers) { 584 // Put long in register. 585 reg = java_iarg_reg[ireg]; 586 ++ireg; 587 } else { 588 // Put long on stack. They must be aligned to 2 slots. 589 if (stk & 0x1) ++stk; 590 reg = VMRegImpl::stack2reg(stk); 591 stk += inc_stk_for_longdouble; 592 } 593 regs[i].set2(reg); 594 break; 595 case T_OBJECT: 596 case T_ARRAY: 597 case T_ADDRESS: 598 if (ireg < num_java_iarg_registers) { 599 // Put ptr in register. 600 reg = java_iarg_reg[ireg]; 601 ++ireg; 602 } else { 603 // Put ptr on stack. Objects must be aligned to 2 slots too, 604 // because "64-bit pointers record oop-ishness on 2 aligned 605 // adjacent registers." (see OopFlow::build_oop_map). 606 if (stk & 0x1) ++stk; 607 reg = VMRegImpl::stack2reg(stk); 608 stk += inc_stk_for_longdouble; 609 } 610 regs[i].set2(reg); 611 break; 612 case T_FLOAT: 613 if (freg < num_java_farg_registers) { 614 // Put float in register. 615 reg = java_farg_reg[freg]; 616 ++freg; 617 } else { 618 // Put float on stack. 619 reg = VMRegImpl::stack2reg(stk); 620 stk += inc_stk_for_intfloat; 621 } 622 regs[i].set1(reg); 623 break; 624 case T_DOUBLE: 625 assert(sig_bt[i+1] == T_VOID, "expecting half"); 626 if (freg < num_java_farg_registers) { 627 // Put double in register. 628 reg = java_farg_reg[freg]; 629 ++freg; 630 } else { 631 // Put double on stack. They must be aligned to 2 slots. 632 if (stk & 0x1) ++stk; 633 reg = VMRegImpl::stack2reg(stk); 634 stk += inc_stk_for_longdouble; 635 } 636 regs[i].set2(reg); 637 break; 638 case T_VOID: 639 // Do not count halves. 640 regs[i].set_bad(); 641 break; 642 default: 643 ShouldNotReachHere(); 644 } 645 } 646 return round_to(stk, 2); 647 } 648 649 #ifdef COMPILER2 650 // Calling convention for calling C code. 651 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 652 VMRegPair *regs, 653 VMRegPair *regs2, 654 int total_args_passed) { 655 // Calling conventions for C runtime calls and calls to JNI native methods. 656 // 657 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8 658 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist 659 // the first 13 flt/dbl's in the first 13 fp regs but additionally 660 // copy flt/dbl to the stack if they are beyond the 8th argument. 661 662 const VMReg iarg_reg[8] = { 663 R3->as_VMReg(), 664 R4->as_VMReg(), 665 R5->as_VMReg(), 666 R6->as_VMReg(), 667 R7->as_VMReg(), 668 R8->as_VMReg(), 669 R9->as_VMReg(), 670 R10->as_VMReg() 671 }; 672 673 const VMReg farg_reg[13] = { 674 F1->as_VMReg(), 675 F2->as_VMReg(), 676 F3->as_VMReg(), 677 F4->as_VMReg(), 678 F5->as_VMReg(), 679 F6->as_VMReg(), 680 F7->as_VMReg(), 681 F8->as_VMReg(), 682 F9->as_VMReg(), 683 F10->as_VMReg(), 684 F11->as_VMReg(), 685 F12->as_VMReg(), 686 F13->as_VMReg() 687 }; 688 689 // Check calling conventions consistency. 690 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c && 691 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c, 692 "consistency"); 693 694 // `Stk' counts stack slots. Due to alignment, 32 bit values occupy 695 // 2 such slots, like 64 bit values do. 696 const int inc_stk_for_intfloat = 2; // 2 slots for ints and floats 697 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles 698 699 int i; 700 VMReg reg; 701 // Leave room for C-compatible ABI_REG_ARGS. 702 int stk = (frame::abi_reg_args_size - frame::jit_out_preserve_size) / VMRegImpl::stack_slot_size; 703 int arg = 0; 704 int freg = 0; 705 706 // Avoid passing C arguments in the wrong stack slots. 707 #if defined(ABI_ELFv2) 708 assert((SharedRuntime::out_preserve_stack_slots() + stk) * VMRegImpl::stack_slot_size == 96, 709 "passing C arguments in wrong stack slots"); 710 #else 711 assert((SharedRuntime::out_preserve_stack_slots() + stk) * VMRegImpl::stack_slot_size == 112, 712 "passing C arguments in wrong stack slots"); 713 #endif 714 // We fill-out regs AND regs2 if an argument must be passed in a 715 // register AND in a stack slot. If regs2 is NULL in such a 716 // situation, we bail-out with a fatal error. 717 for (int i = 0; i < total_args_passed; ++i, ++arg) { 718 // Initialize regs2 to BAD. 719 if (regs2 != NULL) regs2[i].set_bad(); 720 721 switch(sig_bt[i]) { 722 723 // 724 // If arguments 0-7 are integers, they are passed in integer registers. 725 // Argument i is placed in iarg_reg[i]. 726 // 727 case T_BOOLEAN: 728 case T_CHAR: 729 case T_BYTE: 730 case T_SHORT: 731 case T_INT: 732 // We must cast ints to longs and use full 64 bit stack slots 733 // here. Thus fall through, handle as long. 734 case T_LONG: 735 case T_OBJECT: 736 case T_ARRAY: 737 case T_ADDRESS: 738 case T_METADATA: 739 // Oops are already boxed if required (JNI). 740 if (arg < Argument::n_int_register_parameters_c) { 741 reg = iarg_reg[arg]; 742 } else { 743 reg = VMRegImpl::stack2reg(stk); 744 stk += inc_stk_for_longdouble; 745 } 746 regs[i].set2(reg); 747 break; 748 749 // 750 // Floats are treated differently from int regs: The first 13 float arguments 751 // are passed in registers (not the float args among the first 13 args). 752 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed 753 // in farg_reg[j] if argument i is the j-th float argument of this call. 754 // 755 case T_FLOAT: 756 #if defined(LINUX) 757 // Linux uses ELF ABI. Both original ELF and ELFv2 ABIs have float 758 // in the least significant word of an argument slot. 759 #if defined(VM_LITTLE_ENDIAN) 760 #define FLOAT_WORD_OFFSET_IN_SLOT 0 761 #else 762 #define FLOAT_WORD_OFFSET_IN_SLOT 1 763 #endif 764 #elif defined(AIX) 765 // Although AIX runs on big endian CPU, float is in the most 766 // significant word of an argument slot. 767 #define FLOAT_WORD_OFFSET_IN_SLOT 0 768 #else 769 #error "unknown OS" 770 #endif 771 if (freg < Argument::n_float_register_parameters_c) { 772 // Put float in register ... 773 reg = farg_reg[freg]; 774 ++freg; 775 776 // Argument i for i > 8 is placed on the stack even if it's 777 // placed in a register (if it's a float arg). Aix disassembly 778 // shows that xlC places these float args on the stack AND in 779 // a register. This is not documented, but we follow this 780 // convention, too. 781 if (arg >= Argument::n_regs_not_on_stack_c) { 782 // ... and on the stack. 783 guarantee(regs2 != NULL, "must pass float in register and stack slot"); 784 VMReg reg2 = VMRegImpl::stack2reg(stk + FLOAT_WORD_OFFSET_IN_SLOT); 785 regs2[i].set1(reg2); 786 stk += inc_stk_for_intfloat; 787 } 788 789 } else { 790 // Put float on stack. 791 reg = VMRegImpl::stack2reg(stk + FLOAT_WORD_OFFSET_IN_SLOT); 792 stk += inc_stk_for_intfloat; 793 } 794 regs[i].set1(reg); 795 break; 796 case T_DOUBLE: 797 assert(sig_bt[i+1] == T_VOID, "expecting half"); 798 if (freg < Argument::n_float_register_parameters_c) { 799 // Put double in register ... 800 reg = farg_reg[freg]; 801 ++freg; 802 803 // Argument i for i > 8 is placed on the stack even if it's 804 // placed in a register (if it's a double arg). Aix disassembly 805 // shows that xlC places these float args on the stack AND in 806 // a register. This is not documented, but we follow this 807 // convention, too. 808 if (arg >= Argument::n_regs_not_on_stack_c) { 809 // ... and on the stack. 810 guarantee(regs2 != NULL, "must pass float in register and stack slot"); 811 VMReg reg2 = VMRegImpl::stack2reg(stk); 812 regs2[i].set2(reg2); 813 stk += inc_stk_for_longdouble; 814 } 815 } else { 816 // Put double on stack. 817 reg = VMRegImpl::stack2reg(stk); 818 stk += inc_stk_for_longdouble; 819 } 820 regs[i].set2(reg); 821 break; 822 823 case T_VOID: 824 // Do not count halves. 825 regs[i].set_bad(); 826 --arg; 827 break; 828 default: 829 ShouldNotReachHere(); 830 } 831 } 832 833 return round_to(stk, 2); 834 } 835 #endif // COMPILER2 836 837 static address gen_c2i_adapter(MacroAssembler *masm, 838 int total_args_passed, 839 int comp_args_on_stack, 840 const BasicType *sig_bt, 841 const VMRegPair *regs, 842 Label& call_interpreter, 843 const Register& ientry) { 844 845 address c2i_entrypoint; 846 847 const Register sender_SP = R21_sender_SP; // == R21_tmp1 848 const Register code = R22_tmp2; 849 //const Register ientry = R23_tmp3; 850 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 }; 851 const int num_value_regs = sizeof(value_regs) / sizeof(Register); 852 int value_regs_index = 0; 853 854 const Register return_pc = R27_tmp7; 855 const Register tmp = R28_tmp8; 856 857 assert_different_registers(sender_SP, code, ientry, return_pc, tmp); 858 859 // Adapter needs TOP_IJAVA_FRAME_ABI. 860 const int adapter_size = frame::top_ijava_frame_abi_size + 861 round_to(total_args_passed * wordSize, frame::alignment_in_bytes); 862 863 // regular (verified) c2i entry point 864 c2i_entrypoint = __ pc(); 865 866 // Does compiled code exists? If yes, patch the caller's callsite. 867 __ ld(code, method_(code)); 868 __ cmpdi(CCR0, code, 0); 869 __ ld(ientry, method_(interpreter_entry)); // preloaded 870 __ beq(CCR0, call_interpreter); 871 872 873 // Patch caller's callsite, method_(code) was not NULL which means that 874 // compiled code exists. 875 __ mflr(return_pc); 876 __ std(return_pc, _abi(lr), R1_SP); 877 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs); 878 879 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc); 880 881 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs); 882 __ ld(return_pc, _abi(lr), R1_SP); 883 __ ld(ientry, method_(interpreter_entry)); // preloaded 884 __ mtlr(return_pc); 885 886 887 // Call the interpreter. 888 __ BIND(call_interpreter); 889 __ mtctr(ientry); 890 891 // Get a copy of the current SP for loading caller's arguments. 892 __ mr(sender_SP, R1_SP); 893 894 // Add space for the adapter. 895 __ resize_frame(-adapter_size, R12_scratch2); 896 897 int st_off = adapter_size - wordSize; 898 899 // Write the args into the outgoing interpreter space. 900 for (int i = 0; i < total_args_passed; i++) { 901 VMReg r_1 = regs[i].first(); 902 VMReg r_2 = regs[i].second(); 903 if (!r_1->is_valid()) { 904 assert(!r_2->is_valid(), ""); 905 continue; 906 } 907 if (r_1->is_stack()) { 908 Register tmp_reg = value_regs[value_regs_index]; 909 value_regs_index = (value_regs_index + 1) % num_value_regs; 910 // The calling convention produces OptoRegs that ignore the out 911 // preserve area (JIT's ABI). We must account for it here. 912 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 913 if (!r_2->is_valid()) { 914 __ lwz(tmp_reg, ld_off, sender_SP); 915 } else { 916 __ ld(tmp_reg, ld_off, sender_SP); 917 } 918 // Pretend stack targets were loaded into tmp_reg. 919 r_1 = tmp_reg->as_VMReg(); 920 } 921 922 if (r_1->is_Register()) { 923 Register r = r_1->as_Register(); 924 if (!r_2->is_valid()) { 925 __ stw(r, st_off, R1_SP); 926 st_off-=wordSize; 927 } else { 928 // Longs are given 2 64-bit slots in the interpreter, but the 929 // data is passed in only 1 slot. 930 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 931 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); ) 932 st_off-=wordSize; 933 } 934 __ std(r, st_off, R1_SP); 935 st_off-=wordSize; 936 } 937 } else { 938 assert(r_1->is_FloatRegister(), ""); 939 FloatRegister f = r_1->as_FloatRegister(); 940 if (!r_2->is_valid()) { 941 __ stfs(f, st_off, R1_SP); 942 st_off-=wordSize; 943 } else { 944 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the 945 // data is passed in only 1 slot. 946 // One of these should get known junk... 947 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); ) 948 st_off-=wordSize; 949 __ stfd(f, st_off, R1_SP); 950 st_off-=wordSize; 951 } 952 } 953 } 954 955 // Jump to the interpreter just as if interpreter was doing it. 956 957 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 958 959 // load TOS 960 __ addi(R15_esp, R1_SP, st_off); 961 962 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1. 963 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register"); 964 __ bctr(); 965 966 return c2i_entrypoint; 967 } 968 969 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 970 int total_args_passed, 971 int comp_args_on_stack, 972 const BasicType *sig_bt, 973 const VMRegPair *regs) { 974 975 // Load method's entry-point from method. 976 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method); 977 __ mtctr(R12_scratch2); 978 979 // We will only enter here from an interpreted frame and never from after 980 // passing thru a c2i. Azul allowed this but we do not. If we lose the 981 // race and use a c2i we will remain interpreted for the race loser(s). 982 // This removes all sorts of headaches on the x86 side and also eliminates 983 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions. 984 985 // Note: r13 contains the senderSP on entry. We must preserve it since 986 // we may do a i2c -> c2i transition if we lose a race where compiled 987 // code goes non-entrant while we get args ready. 988 // In addition we use r13 to locate all the interpreter args as 989 // we must align the stack to 16 bytes on an i2c entry else we 990 // lose alignment we expect in all compiled code and register 991 // save code can segv when fxsave instructions find improperly 992 // aligned stack pointer. 993 994 const Register ld_ptr = R15_esp; 995 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 }; 996 const int num_value_regs = sizeof(value_regs) / sizeof(Register); 997 int value_regs_index = 0; 998 999 int ld_offset = total_args_passed*wordSize; 1000 1001 // Cut-out for having no stack args. Since up to 2 int/oop args are passed 1002 // in registers, we will occasionally have no stack args. 1003 int comp_words_on_stack = 0; 1004 if (comp_args_on_stack) { 1005 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in 1006 // registers are below. By subtracting stack0, we either get a negative 1007 // number (all values in registers) or the maximum stack slot accessed. 1008 1009 // Convert 4-byte c2 stack slots to words. 1010 comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 1011 // Round up to miminum stack alignment, in wordSize. 1012 comp_words_on_stack = round_to(comp_words_on_stack, 2); 1013 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1); 1014 } 1015 1016 // Now generate the shuffle code. Pick up all register args and move the 1017 // rest through register value=Z_R12. 1018 BLOCK_COMMENT("Shuffle arguments"); 1019 for (int i = 0; i < total_args_passed; i++) { 1020 if (sig_bt[i] == T_VOID) { 1021 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 1022 continue; 1023 } 1024 1025 // Pick up 0, 1 or 2 words from ld_ptr. 1026 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 1027 "scrambled load targets?"); 1028 VMReg r_1 = regs[i].first(); 1029 VMReg r_2 = regs[i].second(); 1030 if (!r_1->is_valid()) { 1031 assert(!r_2->is_valid(), ""); 1032 continue; 1033 } 1034 if (r_1->is_FloatRegister()) { 1035 if (!r_2->is_valid()) { 1036 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr); 1037 ld_offset-=wordSize; 1038 } else { 1039 // Skip the unused interpreter slot. 1040 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr); 1041 ld_offset-=2*wordSize; 1042 } 1043 } else { 1044 Register r; 1045 if (r_1->is_stack()) { 1046 // Must do a memory to memory move thru "value". 1047 r = value_regs[value_regs_index]; 1048 value_regs_index = (value_regs_index + 1) % num_value_regs; 1049 } else { 1050 r = r_1->as_Register(); 1051 } 1052 if (!r_2->is_valid()) { 1053 // Not sure we need to do this but it shouldn't hurt. 1054 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ADDRESS || sig_bt[i] == T_ARRAY) { 1055 __ ld(r, ld_offset, ld_ptr); 1056 ld_offset-=wordSize; 1057 } else { 1058 __ lwz(r, ld_offset, ld_ptr); 1059 ld_offset-=wordSize; 1060 } 1061 } else { 1062 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the 1063 // data is passed in only 1 slot. 1064 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 1065 ld_offset-=wordSize; 1066 } 1067 __ ld(r, ld_offset, ld_ptr); 1068 ld_offset-=wordSize; 1069 } 1070 1071 if (r_1->is_stack()) { 1072 // Now store value where the compiler expects it 1073 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size; 1074 1075 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN || 1076 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) { 1077 __ stw(r, st_off, R1_SP); 1078 } else { 1079 __ std(r, st_off, R1_SP); 1080 } 1081 } 1082 } 1083 } 1084 1085 BLOCK_COMMENT("Store method"); 1086 // Store method into thread->callee_target. 1087 // We might end up in handle_wrong_method if the callee is 1088 // deoptimized as we race thru here. If that happens we don't want 1089 // to take a safepoint because the caller frame will look 1090 // interpreted and arguments are now "compiled" so it is much better 1091 // to make this transition invisible to the stack walking 1092 // code. Unfortunately if we try and find the callee by normal means 1093 // a safepoint is possible. So we stash the desired callee in the 1094 // thread and the vm will find there should this case occur. 1095 __ std(R19_method, thread_(callee_target)); 1096 1097 // Jump to the compiled code just as if compiled code was doing it. 1098 __ bctr(); 1099 } 1100 1101 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 1102 int total_args_passed, 1103 int comp_args_on_stack, 1104 const BasicType *sig_bt, 1105 const VMRegPair *regs, 1106 AdapterFingerPrint* fingerprint) { 1107 address i2c_entry; 1108 address c2i_unverified_entry; 1109 address c2i_entry; 1110 1111 1112 // entry: i2c 1113 1114 __ align(CodeEntryAlignment); 1115 i2c_entry = __ pc(); 1116 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 1117 1118 1119 // entry: c2i unverified 1120 1121 __ align(CodeEntryAlignment); 1122 BLOCK_COMMENT("c2i unverified entry"); 1123 c2i_unverified_entry = __ pc(); 1124 1125 // inline_cache contains a compiledICHolder 1126 const Register ic = R19_method; 1127 const Register ic_klass = R11_scratch1; 1128 const Register receiver_klass = R12_scratch2; 1129 const Register code = R21_tmp1; 1130 const Register ientry = R23_tmp3; 1131 1132 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry); 1133 assert(R11_scratch1 == R11, "need prologue scratch register"); 1134 1135 Label call_interpreter; 1136 1137 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), 1138 "klass offset should reach into any page"); 1139 // Check for NULL argument if we don't have implicit null checks. 1140 if (!ImplicitNullChecks || !os::zero_page_read_protected()) { 1141 if (TrapBasedNullChecks) { 1142 __ trap_null_check(R3_ARG1); 1143 } else { 1144 Label valid; 1145 __ cmpdi(CCR0, R3_ARG1, 0); 1146 __ bne_predict_taken(CCR0, valid); 1147 // We have a null argument, branch to ic_miss_stub. 1148 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1149 relocInfo::runtime_call_type); 1150 __ BIND(valid); 1151 } 1152 } 1153 // Assume argument is not NULL, load klass from receiver. 1154 __ load_klass(receiver_klass, R3_ARG1); 1155 1156 __ ld(ic_klass, CompiledICHolder::holder_klass_offset(), ic); 1157 1158 if (TrapBasedICMissChecks) { 1159 __ trap_ic_miss_check(receiver_klass, ic_klass); 1160 } else { 1161 Label valid; 1162 __ cmpd(CCR0, receiver_klass, ic_klass); 1163 __ beq_predict_taken(CCR0, valid); 1164 // We have an unexpected klass, branch to ic_miss_stub. 1165 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 1166 relocInfo::runtime_call_type); 1167 __ BIND(valid); 1168 } 1169 1170 // Argument is valid and klass is as expected, continue. 1171 1172 // Extract method from inline cache, verified entry point needs it. 1173 __ ld(R19_method, CompiledICHolder::holder_method_offset(), ic); 1174 assert(R19_method == ic, "the inline cache register is dead here"); 1175 1176 __ ld(code, method_(code)); 1177 __ cmpdi(CCR0, code, 0); 1178 __ ld(ientry, method_(interpreter_entry)); // preloaded 1179 __ beq_predict_taken(CCR0, call_interpreter); 1180 1181 // Branch to ic_miss_stub. 1182 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type); 1183 1184 // entry: c2i 1185 1186 c2i_entry = gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry); 1187 1188 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 1189 } 1190 1191 #ifdef COMPILER2 1192 // An oop arg. Must pass a handle not the oop itself. 1193 static void object_move(MacroAssembler* masm, 1194 int frame_size_in_slots, 1195 OopMap* oop_map, int oop_handle_offset, 1196 bool is_receiver, int* receiver_offset, 1197 VMRegPair src, VMRegPair dst, 1198 Register r_caller_sp, Register r_temp_1, Register r_temp_2) { 1199 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)), 1200 "receiver has already been moved"); 1201 1202 // We must pass a handle. First figure out the location we use as a handle. 1203 1204 if (src.first()->is_stack()) { 1205 // stack to stack or reg 1206 1207 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register(); 1208 Label skip; 1209 const int oop_slot_in_callers_frame = reg2slot(src.first()); 1210 1211 guarantee(!is_receiver, "expecting receiver in register"); 1212 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots)); 1213 1214 __ addi(r_handle, r_caller_sp, reg2offset(src.first())); 1215 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp); 1216 __ cmpdi(CCR0, r_temp_2, 0); 1217 __ bne(CCR0, skip); 1218 // Use a NULL handle if oop is NULL. 1219 __ li(r_handle, 0); 1220 __ bind(skip); 1221 1222 if (dst.first()->is_stack()) { 1223 // stack to stack 1224 __ std(r_handle, reg2offset(dst.first()), R1_SP); 1225 } else { 1226 // stack to reg 1227 // Nothing to do, r_handle is already the dst register. 1228 } 1229 } else { 1230 // reg to stack or reg 1231 const Register r_oop = src.first()->as_Register(); 1232 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register(); 1233 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word 1234 + oop_handle_offset; // in slots 1235 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size; 1236 Label skip; 1237 1238 if (is_receiver) { 1239 *receiver_offset = oop_offset; 1240 } 1241 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1242 1243 __ std( r_oop, oop_offset, R1_SP); 1244 __ addi(r_handle, R1_SP, oop_offset); 1245 1246 __ cmpdi(CCR0, r_oop, 0); 1247 __ bne(CCR0, skip); 1248 // Use a NULL handle if oop is NULL. 1249 __ li(r_handle, 0); 1250 __ bind(skip); 1251 1252 if (dst.first()->is_stack()) { 1253 // reg to stack 1254 __ std(r_handle, reg2offset(dst.first()), R1_SP); 1255 } else { 1256 // reg to reg 1257 // Nothing to do, r_handle is already the dst register. 1258 } 1259 } 1260 } 1261 1262 static void int_move(MacroAssembler*masm, 1263 VMRegPair src, VMRegPair dst, 1264 Register r_caller_sp, Register r_temp) { 1265 assert(src.first()->is_valid(), "incoming must be int"); 1266 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long"); 1267 1268 if (src.first()->is_stack()) { 1269 if (dst.first()->is_stack()) { 1270 // stack to stack 1271 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp); 1272 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1273 } else { 1274 // stack to reg 1275 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1276 } 1277 } else if (dst.first()->is_stack()) { 1278 // reg to stack 1279 __ extsw(r_temp, src.first()->as_Register()); 1280 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1281 } else { 1282 // reg to reg 1283 __ extsw(dst.first()->as_Register(), src.first()->as_Register()); 1284 } 1285 } 1286 1287 static void long_move(MacroAssembler*masm, 1288 VMRegPair src, VMRegPair dst, 1289 Register r_caller_sp, Register r_temp) { 1290 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long"); 1291 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long"); 1292 1293 if (src.first()->is_stack()) { 1294 if (dst.first()->is_stack()) { 1295 // stack to stack 1296 __ ld( r_temp, reg2offset(src.first()), r_caller_sp); 1297 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1298 } else { 1299 // stack to reg 1300 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1301 } 1302 } else if (dst.first()->is_stack()) { 1303 // reg to stack 1304 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP); 1305 } else { 1306 // reg to reg 1307 if (dst.first()->as_Register() != src.first()->as_Register()) 1308 __ mr(dst.first()->as_Register(), src.first()->as_Register()); 1309 } 1310 } 1311 1312 static void float_move(MacroAssembler*masm, 1313 VMRegPair src, VMRegPair dst, 1314 Register r_caller_sp, Register r_temp) { 1315 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float"); 1316 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float"); 1317 1318 if (src.first()->is_stack()) { 1319 if (dst.first()->is_stack()) { 1320 // stack to stack 1321 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp); 1322 __ stw(r_temp, reg2offset(dst.first()), R1_SP); 1323 } else { 1324 // stack to reg 1325 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp); 1326 } 1327 } else if (dst.first()->is_stack()) { 1328 // reg to stack 1329 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP); 1330 } else { 1331 // reg to reg 1332 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister()) 1333 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1334 } 1335 } 1336 1337 static void double_move(MacroAssembler*masm, 1338 VMRegPair src, VMRegPair dst, 1339 Register r_caller_sp, Register r_temp) { 1340 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double"); 1341 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double"); 1342 1343 if (src.first()->is_stack()) { 1344 if (dst.first()->is_stack()) { 1345 // stack to stack 1346 __ ld( r_temp, reg2offset(src.first()), r_caller_sp); 1347 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1348 } else { 1349 // stack to reg 1350 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp); 1351 } 1352 } else if (dst.first()->is_stack()) { 1353 // reg to stack 1354 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP); 1355 } else { 1356 // reg to reg 1357 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister()) 1358 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister()); 1359 } 1360 } 1361 1362 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1363 switch (ret_type) { 1364 case T_BOOLEAN: 1365 case T_CHAR: 1366 case T_BYTE: 1367 case T_SHORT: 1368 case T_INT: 1369 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1370 break; 1371 case T_ARRAY: 1372 case T_OBJECT: 1373 case T_LONG: 1374 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1375 break; 1376 case T_FLOAT: 1377 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1378 break; 1379 case T_DOUBLE: 1380 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1381 break; 1382 case T_VOID: 1383 break; 1384 default: 1385 ShouldNotReachHere(); 1386 break; 1387 } 1388 } 1389 1390 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1391 switch (ret_type) { 1392 case T_BOOLEAN: 1393 case T_CHAR: 1394 case T_BYTE: 1395 case T_SHORT: 1396 case T_INT: 1397 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1398 break; 1399 case T_ARRAY: 1400 case T_OBJECT: 1401 case T_LONG: 1402 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1403 break; 1404 case T_FLOAT: 1405 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1406 break; 1407 case T_DOUBLE: 1408 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP); 1409 break; 1410 case T_VOID: 1411 break; 1412 default: 1413 ShouldNotReachHere(); 1414 break; 1415 } 1416 } 1417 1418 static void save_or_restore_arguments(MacroAssembler* masm, 1419 const int stack_slots, 1420 const int total_in_args, 1421 const int arg_save_area, 1422 OopMap* map, 1423 VMRegPair* in_regs, 1424 BasicType* in_sig_bt) { 1425 // If map is non-NULL then the code should store the values, 1426 // otherwise it should load them. 1427 int slot = arg_save_area; 1428 // Save down double word first. 1429 for (int i = 0; i < total_in_args; i++) { 1430 if (in_regs[i].first()->is_FloatRegister() && in_sig_bt[i] == T_DOUBLE) { 1431 int offset = slot * VMRegImpl::stack_slot_size; 1432 slot += VMRegImpl::slots_per_word; 1433 assert(slot <= stack_slots, "overflow (after DOUBLE stack slot)"); 1434 if (map != NULL) { 1435 __ stfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1436 } else { 1437 __ lfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1438 } 1439 } else if (in_regs[i].first()->is_Register() && 1440 (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_ARRAY)) { 1441 int offset = slot * VMRegImpl::stack_slot_size; 1442 if (map != NULL) { 1443 __ std(in_regs[i].first()->as_Register(), offset, R1_SP); 1444 if (in_sig_bt[i] == T_ARRAY) { 1445 map->set_oop(VMRegImpl::stack2reg(slot)); 1446 } 1447 } else { 1448 __ ld(in_regs[i].first()->as_Register(), offset, R1_SP); 1449 } 1450 slot += VMRegImpl::slots_per_word; 1451 assert(slot <= stack_slots, "overflow (after LONG/ARRAY stack slot)"); 1452 } 1453 } 1454 // Save or restore single word registers. 1455 for (int i = 0; i < total_in_args; i++) { 1456 // PPC64: pass ints as longs: must only deal with floats here. 1457 if (in_regs[i].first()->is_FloatRegister()) { 1458 if (in_sig_bt[i] == T_FLOAT) { 1459 int offset = slot * VMRegImpl::stack_slot_size; 1460 slot++; 1461 assert(slot <= stack_slots, "overflow (after FLOAT stack slot)"); 1462 if (map != NULL) { 1463 __ stfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1464 } else { 1465 __ lfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP); 1466 } 1467 } 1468 } else if (in_regs[i].first()->is_stack()) { 1469 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1470 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1471 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots)); 1472 } 1473 } 1474 } 1475 } 1476 1477 // Check GC_locker::needs_gc and enter the runtime if it's true. This 1478 // keeps a new JNI critical region from starting until a GC has been 1479 // forced. Save down any oops in registers and describe them in an 1480 // OopMap. 1481 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1482 const int stack_slots, 1483 const int total_in_args, 1484 const int arg_save_area, 1485 OopMapSet* oop_maps, 1486 VMRegPair* in_regs, 1487 BasicType* in_sig_bt, 1488 Register tmp_reg ) { 1489 __ block_comment("check GC_locker::needs_gc"); 1490 Label cont; 1491 __ lbz(tmp_reg, (RegisterOrConstant)(intptr_t)GC_locker::needs_gc_address()); 1492 __ cmplwi(CCR0, tmp_reg, 0); 1493 __ beq(CCR0, cont); 1494 1495 // Save down any values that are live in registers and call into the 1496 // runtime to halt for a GC. 1497 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1498 save_or_restore_arguments(masm, stack_slots, total_in_args, 1499 arg_save_area, map, in_regs, in_sig_bt); 1500 1501 __ mr(R3_ARG1, R16_thread); 1502 __ set_last_Java_frame(R1_SP, noreg); 1503 1504 __ block_comment("block_for_jni_critical"); 1505 address entry_point = CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical); 1506 #if defined(ABI_ELFv2) 1507 __ call_c(entry_point, relocInfo::runtime_call_type); 1508 #else 1509 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, entry_point), relocInfo::runtime_call_type); 1510 #endif 1511 address start = __ pc() - __ offset(), 1512 calls_return_pc = __ last_calls_return_pc(); 1513 oop_maps->add_gc_map(calls_return_pc - start, map); 1514 1515 __ reset_last_Java_frame(); 1516 1517 // Reload all the register arguments. 1518 save_or_restore_arguments(masm, stack_slots, total_in_args, 1519 arg_save_area, NULL, in_regs, in_sig_bt); 1520 1521 __ BIND(cont); 1522 1523 #ifdef ASSERT 1524 if (StressCriticalJNINatives) { 1525 // Stress register saving. 1526 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1527 save_or_restore_arguments(masm, stack_slots, total_in_args, 1528 arg_save_area, map, in_regs, in_sig_bt); 1529 // Destroy argument registers. 1530 for (int i = 0; i < total_in_args; i++) { 1531 if (in_regs[i].first()->is_Register()) { 1532 const Register reg = in_regs[i].first()->as_Register(); 1533 __ neg(reg, reg); 1534 } else if (in_regs[i].first()->is_FloatRegister()) { 1535 __ fneg(in_regs[i].first()->as_FloatRegister(), in_regs[i].first()->as_FloatRegister()); 1536 } 1537 } 1538 1539 save_or_restore_arguments(masm, stack_slots, total_in_args, 1540 arg_save_area, NULL, in_regs, in_sig_bt); 1541 } 1542 #endif 1543 } 1544 1545 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst, Register r_caller_sp, Register r_temp) { 1546 if (src.first()->is_stack()) { 1547 if (dst.first()->is_stack()) { 1548 // stack to stack 1549 __ ld(r_temp, reg2offset(src.first()), r_caller_sp); 1550 __ std(r_temp, reg2offset(dst.first()), R1_SP); 1551 } else { 1552 // stack to reg 1553 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp); 1554 } 1555 } else if (dst.first()->is_stack()) { 1556 // reg to stack 1557 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP); 1558 } else { 1559 if (dst.first() != src.first()) { 1560 __ mr(dst.first()->as_Register(), src.first()->as_Register()); 1561 } 1562 } 1563 } 1564 1565 // Unpack an array argument into a pointer to the body and the length 1566 // if the array is non-null, otherwise pass 0 for both. 1567 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, 1568 VMRegPair body_arg, VMRegPair length_arg, Register r_caller_sp, 1569 Register tmp_reg, Register tmp2_reg) { 1570 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg, 1571 "possible collision"); 1572 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg, 1573 "possible collision"); 1574 1575 // Pass the length, ptr pair. 1576 Label set_out_args; 1577 VMRegPair tmp, tmp2; 1578 tmp.set_ptr(tmp_reg->as_VMReg()); 1579 tmp2.set_ptr(tmp2_reg->as_VMReg()); 1580 if (reg.first()->is_stack()) { 1581 // Load the arg up from the stack. 1582 move_ptr(masm, reg, tmp, r_caller_sp, /*unused*/ R0); 1583 reg = tmp; 1584 } 1585 __ li(tmp2_reg, 0); // Pass zeros if Array=null. 1586 if (tmp_reg != reg.first()->as_Register()) __ li(tmp_reg, 0); 1587 __ cmpdi(CCR0, reg.first()->as_Register(), 0); 1588 __ beq(CCR0, set_out_args); 1589 __ lwa(tmp2_reg, arrayOopDesc::length_offset_in_bytes(), reg.first()->as_Register()); 1590 __ addi(tmp_reg, reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type)); 1591 __ bind(set_out_args); 1592 move_ptr(masm, tmp, body_arg, r_caller_sp, /*unused*/ R0); 1593 move_ptr(masm, tmp2, length_arg, r_caller_sp, /*unused*/ R0); // Same as move32_64 on PPC64. 1594 } 1595 1596 static void verify_oop_args(MacroAssembler* masm, 1597 methodHandle method, 1598 const BasicType* sig_bt, 1599 const VMRegPair* regs) { 1600 Register temp_reg = R19_method; // not part of any compiled calling seq 1601 if (VerifyOops) { 1602 for (int i = 0; i < method->size_of_parameters(); i++) { 1603 if (sig_bt[i] == T_OBJECT || 1604 sig_bt[i] == T_ARRAY) { 1605 VMReg r = regs[i].first(); 1606 assert(r->is_valid(), "bad oop arg"); 1607 if (r->is_stack()) { 1608 __ ld(temp_reg, reg2offset(r), R1_SP); 1609 __ verify_oop(temp_reg); 1610 } else { 1611 __ verify_oop(r->as_Register()); 1612 } 1613 } 1614 } 1615 } 1616 } 1617 1618 static void gen_special_dispatch(MacroAssembler* masm, 1619 methodHandle method, 1620 const BasicType* sig_bt, 1621 const VMRegPair* regs) { 1622 verify_oop_args(masm, method, sig_bt, regs); 1623 vmIntrinsics::ID iid = method->intrinsic_id(); 1624 1625 // Now write the args into the outgoing interpreter space 1626 bool has_receiver = false; 1627 Register receiver_reg = noreg; 1628 int member_arg_pos = -1; 1629 Register member_reg = noreg; 1630 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1631 if (ref_kind != 0) { 1632 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1633 member_reg = R19_method; // known to be free at this point 1634 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1635 } else if (iid == vmIntrinsics::_invokeBasic) { 1636 has_receiver = true; 1637 } else { 1638 fatal("unexpected intrinsic id %d", iid); 1639 } 1640 1641 if (member_reg != noreg) { 1642 // Load the member_arg into register, if necessary. 1643 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1644 VMReg r = regs[member_arg_pos].first(); 1645 if (r->is_stack()) { 1646 __ ld(member_reg, reg2offset(r), R1_SP); 1647 } else { 1648 // no data motion is needed 1649 member_reg = r->as_Register(); 1650 } 1651 } 1652 1653 if (has_receiver) { 1654 // Make sure the receiver is loaded into a register. 1655 assert(method->size_of_parameters() > 0, "oob"); 1656 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1657 VMReg r = regs[0].first(); 1658 assert(r->is_valid(), "bad receiver arg"); 1659 if (r->is_stack()) { 1660 // Porting note: This assumes that compiled calling conventions always 1661 // pass the receiver oop in a register. If this is not true on some 1662 // platform, pick a temp and load the receiver from stack. 1663 fatal("receiver always in a register"); 1664 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point? 1665 __ ld(receiver_reg, reg2offset(r), R1_SP); 1666 } else { 1667 // no data motion is needed 1668 receiver_reg = r->as_Register(); 1669 } 1670 } 1671 1672 // Figure out which address we are really jumping to: 1673 MethodHandles::generate_method_handle_dispatch(masm, iid, 1674 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1675 } 1676 1677 #endif // COMPILER2 1678 1679 // --------------------------------------------------------------------------- 1680 // Generate a native wrapper for a given method. The method takes arguments 1681 // in the Java compiled code convention, marshals them to the native 1682 // convention (handlizes oops, etc), transitions to native, makes the call, 1683 // returns to java state (possibly blocking), unhandlizes any result and 1684 // returns. 1685 // 1686 // Critical native functions are a shorthand for the use of 1687 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1688 // functions. The wrapper is expected to unpack the arguments before 1689 // passing them to the callee and perform checks before and after the 1690 // native call to ensure that they GC_locker 1691 // lock_critical/unlock_critical semantics are followed. Some other 1692 // parts of JNI setup are skipped like the tear down of the JNI handle 1693 // block and the check for pending exceptions it's impossible for them 1694 // to be thrown. 1695 // 1696 // They are roughly structured like this: 1697 // if (GC_locker::needs_gc()) 1698 // SharedRuntime::block_for_jni_critical(); 1699 // tranistion to thread_in_native 1700 // unpack arrray arguments and call native entry point 1701 // check for safepoint in progress 1702 // check if any thread suspend flags are set 1703 // call into JVM and possible unlock the JNI critical 1704 // if a GC was suppressed while in the critical native. 1705 // transition back to thread_in_Java 1706 // return to caller 1707 // 1708 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm, 1709 const methodHandle& method, 1710 int compile_id, 1711 BasicType *in_sig_bt, 1712 VMRegPair *in_regs, 1713 BasicType ret_type) { 1714 #ifdef COMPILER2 1715 if (method->is_method_handle_intrinsic()) { 1716 vmIntrinsics::ID iid = method->intrinsic_id(); 1717 intptr_t start = (intptr_t)__ pc(); 1718 int vep_offset = ((intptr_t)__ pc()) - start; 1719 gen_special_dispatch(masm, 1720 method, 1721 in_sig_bt, 1722 in_regs); 1723 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1724 __ flush(); 1725 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1726 return nmethod::new_native_nmethod(method, 1727 compile_id, 1728 masm->code(), 1729 vep_offset, 1730 frame_complete, 1731 stack_slots / VMRegImpl::slots_per_word, 1732 in_ByteSize(-1), 1733 in_ByteSize(-1), 1734 (OopMapSet*)NULL); 1735 } 1736 1737 bool is_critical_native = true; 1738 address native_func = method->critical_native_function(); 1739 if (native_func == NULL) { 1740 native_func = method->native_function(); 1741 is_critical_native = false; 1742 } 1743 assert(native_func != NULL, "must have function"); 1744 1745 // First, create signature for outgoing C call 1746 // -------------------------------------------------------------------------- 1747 1748 int total_in_args = method->size_of_parameters(); 1749 // We have received a description of where all the java args are located 1750 // on entry to the wrapper. We need to convert these args to where 1751 // the jni function will expect them. To figure out where they go 1752 // we convert the java signature to a C signature by inserting 1753 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1754 1755 // Calculate the total number of C arguments and create arrays for the 1756 // signature and the outgoing registers. 1757 // On ppc64, we have two arrays for the outgoing registers, because 1758 // some floating-point arguments must be passed in registers _and_ 1759 // in stack locations. 1760 bool method_is_static = method->is_static(); 1761 int total_c_args = total_in_args; 1762 1763 if (!is_critical_native) { 1764 int n_hidden_args = method_is_static ? 2 : 1; 1765 total_c_args += n_hidden_args; 1766 } else { 1767 // No JNIEnv*, no this*, but unpacked arrays (base+length). 1768 for (int i = 0; i < total_in_args; i++) { 1769 if (in_sig_bt[i] == T_ARRAY) { 1770 total_c_args++; 1771 } 1772 } 1773 } 1774 1775 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1776 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1777 VMRegPair *out_regs2 = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1778 BasicType* in_elem_bt = NULL; 1779 1780 // Create the signature for the C call: 1781 // 1) add the JNIEnv* 1782 // 2) add the class if the method is static 1783 // 3) copy the rest of the incoming signature (shifted by the number of 1784 // hidden arguments). 1785 1786 int argc = 0; 1787 if (!is_critical_native) { 1788 out_sig_bt[argc++] = T_ADDRESS; 1789 if (method->is_static()) { 1790 out_sig_bt[argc++] = T_OBJECT; 1791 } 1792 1793 for (int i = 0; i < total_in_args ; i++ ) { 1794 out_sig_bt[argc++] = in_sig_bt[i]; 1795 } 1796 } else { 1797 Thread* THREAD = Thread::current(); 1798 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1799 SignatureStream ss(method->signature()); 1800 int o = 0; 1801 for (int i = 0; i < total_in_args ; i++, o++) { 1802 if (in_sig_bt[i] == T_ARRAY) { 1803 // Arrays are passed as int, elem* pair 1804 Symbol* atype = ss.as_symbol(CHECK_NULL); 1805 const char* at = atype->as_C_string(); 1806 if (strlen(at) == 2) { 1807 assert(at[0] == '[', "must be"); 1808 switch (at[1]) { 1809 case 'B': in_elem_bt[o] = T_BYTE; break; 1810 case 'C': in_elem_bt[o] = T_CHAR; break; 1811 case 'D': in_elem_bt[o] = T_DOUBLE; break; 1812 case 'F': in_elem_bt[o] = T_FLOAT; break; 1813 case 'I': in_elem_bt[o] = T_INT; break; 1814 case 'J': in_elem_bt[o] = T_LONG; break; 1815 case 'S': in_elem_bt[o] = T_SHORT; break; 1816 case 'Z': in_elem_bt[o] = T_BOOLEAN; break; 1817 default: ShouldNotReachHere(); 1818 } 1819 } 1820 } else { 1821 in_elem_bt[o] = T_VOID; 1822 } 1823 if (in_sig_bt[i] != T_VOID) { 1824 assert(in_sig_bt[i] == ss.type(), "must match"); 1825 ss.next(); 1826 } 1827 } 1828 1829 for (int i = 0; i < total_in_args ; i++ ) { 1830 if (in_sig_bt[i] == T_ARRAY) { 1831 // Arrays are passed as int, elem* pair. 1832 out_sig_bt[argc++] = T_INT; 1833 out_sig_bt[argc++] = T_ADDRESS; 1834 } else { 1835 out_sig_bt[argc++] = in_sig_bt[i]; 1836 } 1837 } 1838 } 1839 1840 1841 // Compute the wrapper's frame size. 1842 // -------------------------------------------------------------------------- 1843 1844 // Now figure out where the args must be stored and how much stack space 1845 // they require. 1846 // 1847 // Compute framesize for the wrapper. We need to handlize all oops in 1848 // incoming registers. 1849 // 1850 // Calculate the total number of stack slots we will need: 1851 // 1) abi requirements 1852 // 2) outgoing arguments 1853 // 3) space for inbound oop handle area 1854 // 4) space for handlizing a klass if static method 1855 // 5) space for a lock if synchronized method 1856 // 6) workspace for saving return values, int <-> float reg moves, etc. 1857 // 7) alignment 1858 // 1859 // Layout of the native wrapper frame: 1860 // (stack grows upwards, memory grows downwards) 1861 // 1862 // NW [ABI_REG_ARGS] <-- 1) R1_SP 1863 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset 1864 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset (save area for critical natives) 1865 // klass <-- 4) R1_SP + klass_offset 1866 // lock <-- 5) R1_SP + lock_offset 1867 // [workspace] <-- 6) R1_SP + workspace_offset 1868 // [alignment] (optional) <-- 7) 1869 // caller [JIT_TOP_ABI_48] <-- r_callers_sp 1870 // 1871 // - *_slot_offset Indicates offset from SP in number of stack slots. 1872 // - *_offset Indicates offset from SP in bytes. 1873 1874 int stack_slots = c_calling_convention(out_sig_bt, out_regs, out_regs2, total_c_args) // 1+2) 1875 + SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention. 1876 1877 // Now the space for the inbound oop handle area. 1878 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word; 1879 if (is_critical_native) { 1880 // Critical natives may have to call out so they need a save area 1881 // for register arguments. 1882 int double_slots = 0; 1883 int single_slots = 0; 1884 for (int i = 0; i < total_in_args; i++) { 1885 if (in_regs[i].first()->is_Register()) { 1886 const Register reg = in_regs[i].first()->as_Register(); 1887 switch (in_sig_bt[i]) { 1888 case T_BOOLEAN: 1889 case T_BYTE: 1890 case T_SHORT: 1891 case T_CHAR: 1892 case T_INT: 1893 // Fall through. 1894 case T_ARRAY: 1895 case T_LONG: double_slots++; break; 1896 default: ShouldNotReachHere(); 1897 } 1898 } else if (in_regs[i].first()->is_FloatRegister()) { 1899 switch (in_sig_bt[i]) { 1900 case T_FLOAT: single_slots++; break; 1901 case T_DOUBLE: double_slots++; break; 1902 default: ShouldNotReachHere(); 1903 } 1904 } 1905 } 1906 total_save_slots = double_slots * 2 + round_to(single_slots, 2); // round to even 1907 } 1908 1909 int oop_handle_slot_offset = stack_slots; 1910 stack_slots += total_save_slots; // 3) 1911 1912 int klass_slot_offset = 0; 1913 int klass_offset = -1; 1914 if (method_is_static && !is_critical_native) { // 4) 1915 klass_slot_offset = stack_slots; 1916 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1917 stack_slots += VMRegImpl::slots_per_word; 1918 } 1919 1920 int lock_slot_offset = 0; 1921 int lock_offset = -1; 1922 if (method->is_synchronized()) { // 5) 1923 lock_slot_offset = stack_slots; 1924 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size; 1925 stack_slots += VMRegImpl::slots_per_word; 1926 } 1927 1928 int workspace_slot_offset = stack_slots; // 6) 1929 stack_slots += 2; 1930 1931 // Now compute actual number of stack words we need. 1932 // Rounding to make stack properly aligned. 1933 stack_slots = round_to(stack_slots, // 7) 1934 frame::alignment_in_bytes / VMRegImpl::stack_slot_size); 1935 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size; 1936 1937 1938 // Now we can start generating code. 1939 // -------------------------------------------------------------------------- 1940 1941 intptr_t start_pc = (intptr_t)__ pc(); 1942 intptr_t vep_start_pc; 1943 intptr_t frame_done_pc; 1944 intptr_t oopmap_pc; 1945 1946 Label ic_miss; 1947 Label handle_pending_exception; 1948 1949 Register r_callers_sp = R21; 1950 Register r_temp_1 = R22; 1951 Register r_temp_2 = R23; 1952 Register r_temp_3 = R24; 1953 Register r_temp_4 = R25; 1954 Register r_temp_5 = R26; 1955 Register r_temp_6 = R27; 1956 Register r_return_pc = R28; 1957 1958 Register r_carg1_jnienv = noreg; 1959 Register r_carg2_classorobject = noreg; 1960 if (!is_critical_native) { 1961 r_carg1_jnienv = out_regs[0].first()->as_Register(); 1962 r_carg2_classorobject = out_regs[1].first()->as_Register(); 1963 } 1964 1965 1966 // Generate the Unverified Entry Point (UEP). 1967 // -------------------------------------------------------------------------- 1968 assert(start_pc == (intptr_t)__ pc(), "uep must be at start"); 1969 1970 // Check ic: object class == cached class? 1971 if (!method_is_static) { 1972 Register ic = as_Register(Matcher::inline_cache_reg_encode()); 1973 Register receiver_klass = r_temp_1; 1974 1975 __ cmpdi(CCR0, R3_ARG1, 0); 1976 __ beq(CCR0, ic_miss); 1977 __ verify_oop(R3_ARG1); 1978 __ load_klass(receiver_klass, R3_ARG1); 1979 1980 __ cmpd(CCR0, receiver_klass, ic); 1981 __ bne(CCR0, ic_miss); 1982 } 1983 1984 1985 // Generate the Verified Entry Point (VEP). 1986 // -------------------------------------------------------------------------- 1987 vep_start_pc = (intptr_t)__ pc(); 1988 1989 __ save_LR_CR(r_temp_1); 1990 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame. 1991 __ mr(r_callers_sp, R1_SP); // Remember frame pointer. 1992 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame. 1993 frame_done_pc = (intptr_t)__ pc(); 1994 1995 __ verify_thread(); 1996 1997 // Native nmethod wrappers never take possesion of the oop arguments. 1998 // So the caller will gc the arguments. 1999 // The only thing we need an oopMap for is if the call is static. 2000 // 2001 // An OopMap for lock (and class if static), and one for the VM call itself. 2002 OopMapSet *oop_maps = new OopMapSet(); 2003 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 2004 2005 if (is_critical_native) { 2006 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); 2007 } 2008 2009 // Move arguments from register/stack to register/stack. 2010 // -------------------------------------------------------------------------- 2011 // 2012 // We immediately shuffle the arguments so that for any vm call we have 2013 // to make from here on out (sync slow path, jvmti, etc.) we will have 2014 // captured the oops from our caller and have a valid oopMap for them. 2015 // 2016 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv* 2017 // (derived from JavaThread* which is in R16_thread) and, if static, 2018 // the class mirror instead of a receiver. This pretty much guarantees that 2019 // register layout will not match. We ignore these extra arguments during 2020 // the shuffle. The shuffle is described by the two calling convention 2021 // vectors we have in our possession. We simply walk the java vector to 2022 // get the source locations and the c vector to get the destinations. 2023 2024 // Record sp-based slot for receiver on stack for non-static methods. 2025 int receiver_offset = -1; 2026 2027 // We move the arguments backward because the floating point registers 2028 // destination will always be to a register with a greater or equal 2029 // register number or the stack. 2030 // in is the index of the incoming Java arguments 2031 // out is the index of the outgoing C arguments 2032 2033 #ifdef ASSERT 2034 bool reg_destroyed[RegisterImpl::number_of_registers]; 2035 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 2036 for (int r = 0 ; r < RegisterImpl::number_of_registers ; r++) { 2037 reg_destroyed[r] = false; 2038 } 2039 for (int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++) { 2040 freg_destroyed[f] = false; 2041 } 2042 #endif // ASSERT 2043 2044 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) { 2045 2046 #ifdef ASSERT 2047 if (in_regs[in].first()->is_Register()) { 2048 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!"); 2049 } else if (in_regs[in].first()->is_FloatRegister()) { 2050 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!"); 2051 } 2052 if (out_regs[out].first()->is_Register()) { 2053 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true; 2054 } else if (out_regs[out].first()->is_FloatRegister()) { 2055 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true; 2056 } 2057 if (out_regs2[out].first()->is_Register()) { 2058 reg_destroyed[out_regs2[out].first()->as_Register()->encoding()] = true; 2059 } else if (out_regs2[out].first()->is_FloatRegister()) { 2060 freg_destroyed[out_regs2[out].first()->as_FloatRegister()->encoding()] = true; 2061 } 2062 #endif // ASSERT 2063 2064 switch (in_sig_bt[in]) { 2065 case T_BOOLEAN: 2066 case T_CHAR: 2067 case T_BYTE: 2068 case T_SHORT: 2069 case T_INT: 2070 // Move int and do sign extension. 2071 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2072 break; 2073 case T_LONG: 2074 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2075 break; 2076 case T_ARRAY: 2077 if (is_critical_native) { 2078 int body_arg = out; 2079 out -= 1; // Point to length arg. 2080 unpack_array_argument(masm, in_regs[in], in_elem_bt[in], out_regs[body_arg], out_regs[out], 2081 r_callers_sp, r_temp_1, r_temp_2); 2082 break; 2083 } 2084 case T_OBJECT: 2085 assert(!is_critical_native, "no oop arguments"); 2086 object_move(masm, stack_slots, 2087 oop_map, oop_handle_slot_offset, 2088 ((in == 0) && (!method_is_static)), &receiver_offset, 2089 in_regs[in], out_regs[out], 2090 r_callers_sp, r_temp_1, r_temp_2); 2091 break; 2092 case T_VOID: 2093 break; 2094 case T_FLOAT: 2095 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2096 if (out_regs2[out].first()->is_valid()) { 2097 float_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1); 2098 } 2099 break; 2100 case T_DOUBLE: 2101 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1); 2102 if (out_regs2[out].first()->is_valid()) { 2103 double_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1); 2104 } 2105 break; 2106 case T_ADDRESS: 2107 fatal("found type (T_ADDRESS) in java args"); 2108 break; 2109 default: 2110 ShouldNotReachHere(); 2111 break; 2112 } 2113 } 2114 2115 // Pre-load a static method's oop into ARG2. 2116 // Used both by locking code and the normal JNI call code. 2117 if (method_is_static && !is_critical_native) { 2118 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()), 2119 r_carg2_classorobject); 2120 2121 // Now handlize the static class mirror in carg2. It's known not-null. 2122 __ std(r_carg2_classorobject, klass_offset, R1_SP); 2123 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 2124 __ addi(r_carg2_classorobject, R1_SP, klass_offset); 2125 } 2126 2127 // Get JNIEnv* which is first argument to native. 2128 if (!is_critical_native) { 2129 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset())); 2130 } 2131 2132 // NOTE: 2133 // 2134 // We have all of the arguments setup at this point. 2135 // We MUST NOT touch any outgoing regs from this point on. 2136 // So if we must call out we must push a new frame. 2137 2138 // Get current pc for oopmap, and load it patchable relative to global toc. 2139 oopmap_pc = (intptr_t) __ pc(); 2140 __ calculate_address_from_global_toc(r_return_pc, (address)oopmap_pc, true, true, true, true); 2141 2142 // We use the same pc/oopMap repeatedly when we call out. 2143 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map); 2144 2145 // r_return_pc now has the pc loaded that we will use when we finally call 2146 // to native. 2147 2148 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below. 2149 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register"); 2150 2151 # if 0 2152 // DTrace method entry 2153 # endif 2154 2155 // Lock a synchronized method. 2156 // -------------------------------------------------------------------------- 2157 2158 if (method->is_synchronized()) { 2159 assert(!is_critical_native, "unhandled"); 2160 ConditionRegister r_flag = CCR1; 2161 Register r_oop = r_temp_4; 2162 const Register r_box = r_temp_5; 2163 Label done, locked; 2164 2165 // Load the oop for the object or class. r_carg2_classorobject contains 2166 // either the handlized oop from the incoming arguments or the handlized 2167 // class mirror (if the method is static). 2168 __ ld(r_oop, 0, r_carg2_classorobject); 2169 2170 // Get the lock box slot's address. 2171 __ addi(r_box, R1_SP, lock_offset); 2172 2173 # ifdef ASSERT 2174 if (UseBiasedLocking) { 2175 // Making the box point to itself will make it clear it went unused 2176 // but also be obviously invalid. 2177 __ std(r_box, 0, r_box); 2178 } 2179 # endif // ASSERT 2180 2181 // Try fastpath for locking. 2182 // fast_lock kills r_temp_1, r_temp_2, r_temp_3. 2183 __ compiler_fast_lock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3); 2184 __ beq(r_flag, locked); 2185 2186 // None of the above fast optimizations worked so we have to get into the 2187 // slow case of monitor enter. Inline a special case of call_VM that 2188 // disallows any pending_exception. 2189 2190 // Save argument registers and leave room for C-compatible ABI_REG_ARGS. 2191 int frame_size = frame::abi_reg_args_size + 2192 round_to(total_c_args * wordSize, frame::alignment_in_bytes); 2193 __ mr(R11_scratch1, R1_SP); 2194 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs, out_regs2); 2195 2196 // Do the call. 2197 __ set_last_Java_frame(R11_scratch1, r_return_pc); 2198 assert(r_return_pc->is_nonvolatile(), "expecting return pc to be in non-volatile register"); 2199 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread); 2200 __ reset_last_Java_frame(); 2201 2202 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs, out_regs2); 2203 2204 __ asm_assert_mem8_is_zero(thread_(pending_exception), 2205 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C", 0); 2206 2207 __ bind(locked); 2208 } 2209 2210 2211 // Publish thread state 2212 // -------------------------------------------------------------------------- 2213 2214 // Use that pc we placed in r_return_pc a while back as the current frame anchor. 2215 __ set_last_Java_frame(R1_SP, r_return_pc); 2216 2217 // Transition from _thread_in_Java to _thread_in_native. 2218 __ li(R0, _thread_in_native); 2219 __ release(); 2220 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2221 __ stw(R0, thread_(thread_state)); 2222 if (UseMembar) { 2223 __ fence(); 2224 } 2225 2226 2227 // The JNI call 2228 // -------------------------------------------------------------------------- 2229 #if defined(ABI_ELFv2) 2230 __ call_c(native_func, relocInfo::runtime_call_type); 2231 #else 2232 FunctionDescriptor* fd_native_method = (FunctionDescriptor*) native_func; 2233 __ call_c(fd_native_method, relocInfo::runtime_call_type); 2234 #endif 2235 2236 2237 // Now, we are back from the native code. 2238 2239 2240 // Unpack the native result. 2241 // -------------------------------------------------------------------------- 2242 2243 // For int-types, we do any needed sign-extension required. 2244 // Care must be taken that the return values (R3_RET and F1_RET) 2245 // will survive any VM calls for blocking or unlocking. 2246 // An OOP result (handle) is done specially in the slow-path code. 2247 2248 switch (ret_type) { 2249 case T_VOID: break; // Nothing to do! 2250 case T_FLOAT: break; // Got it where we want it (unless slow-path). 2251 case T_DOUBLE: break; // Got it where we want it (unless slow-path). 2252 case T_LONG: break; // Got it where we want it (unless slow-path). 2253 case T_OBJECT: break; // Really a handle. 2254 // Cannot de-handlize until after reclaiming jvm_lock. 2255 case T_ARRAY: break; 2256 2257 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1) 2258 Label skip_modify; 2259 __ cmpwi(CCR0, R3_RET, 0); 2260 __ beq(CCR0, skip_modify); 2261 __ li(R3_RET, 1); 2262 __ bind(skip_modify); 2263 break; 2264 } 2265 case T_BYTE: { // sign extension 2266 __ extsb(R3_RET, R3_RET); 2267 break; 2268 } 2269 case T_CHAR: { // unsigned result 2270 __ andi(R3_RET, R3_RET, 0xffff); 2271 break; 2272 } 2273 case T_SHORT: { // sign extension 2274 __ extsh(R3_RET, R3_RET); 2275 break; 2276 } 2277 case T_INT: // nothing to do 2278 break; 2279 default: 2280 ShouldNotReachHere(); 2281 break; 2282 } 2283 2284 2285 // Publish thread state 2286 // -------------------------------------------------------------------------- 2287 2288 // Switch thread to "native transition" state before reading the 2289 // synchronization state. This additional state is necessary because reading 2290 // and testing the synchronization state is not atomic w.r.t. GC, as this 2291 // scenario demonstrates: 2292 // - Java thread A, in _thread_in_native state, loads _not_synchronized 2293 // and is preempted. 2294 // - VM thread changes sync state to synchronizing and suspends threads 2295 // for GC. 2296 // - Thread A is resumed to finish this native method, but doesn't block 2297 // here since it didn't see any synchronization in progress, and escapes. 2298 2299 // Transition from _thread_in_native to _thread_in_native_trans. 2300 __ li(R0, _thread_in_native_trans); 2301 __ release(); 2302 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2303 __ stw(R0, thread_(thread_state)); 2304 2305 2306 // Must we block? 2307 // -------------------------------------------------------------------------- 2308 2309 // Block, if necessary, before resuming in _thread_in_Java state. 2310 // In order for GC to work, don't clear the last_Java_sp until after blocking. 2311 Label after_transition; 2312 { 2313 Label no_block, sync; 2314 2315 if (os::is_MP()) { 2316 if (UseMembar) { 2317 // Force this write out before the read below. 2318 __ fence(); 2319 } else { 2320 // Write serialization page so VM thread can do a pseudo remote membar. 2321 // We use the current thread pointer to calculate a thread specific 2322 // offset to write to within the page. This minimizes bus traffic 2323 // due to cache line collision. 2324 __ serialize_memory(R16_thread, r_temp_4, r_temp_5); 2325 } 2326 } 2327 2328 Register sync_state_addr = r_temp_4; 2329 Register sync_state = r_temp_5; 2330 Register suspend_flags = r_temp_6; 2331 2332 __ load_const(sync_state_addr, SafepointSynchronize::address_of_state(), /*temp*/ sync_state); 2333 2334 // TODO: PPC port assert(4 == SafepointSynchronize::sz_state(), "unexpected field size"); 2335 __ lwz(sync_state, 0, sync_state_addr); 2336 2337 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size"); 2338 __ lwz(suspend_flags, thread_(suspend_flags)); 2339 2340 __ acquire(); 2341 2342 Label do_safepoint; 2343 // No synchronization in progress nor yet synchronized. 2344 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized); 2345 // Not suspended. 2346 __ cmpwi(CCR1, suspend_flags, 0); 2347 2348 __ bne(CCR0, sync); 2349 __ beq(CCR1, no_block); 2350 2351 // Block. Save any potential method result value before the operation and 2352 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this 2353 // lets us share the oopMap we used when we went native rather than create 2354 // a distinct one for this pc. 2355 __ bind(sync); 2356 2357 address entry_point = is_critical_native 2358 ? CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition) 2359 : CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans); 2360 save_native_result(masm, ret_type, workspace_slot_offset); 2361 __ call_VM_leaf(entry_point, R16_thread); 2362 restore_native_result(masm, ret_type, workspace_slot_offset); 2363 2364 if (is_critical_native) { 2365 __ b(after_transition); // No thread state transition here. 2366 } 2367 __ bind(no_block); 2368 } 2369 2370 // Publish thread state. 2371 // -------------------------------------------------------------------------- 2372 2373 // Thread state is thread_in_native_trans. Any safepoint blocking has 2374 // already happened so we can now change state to _thread_in_Java. 2375 2376 // Transition from _thread_in_native_trans to _thread_in_Java. 2377 __ li(R0, _thread_in_Java); 2378 __ release(); 2379 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size"); 2380 __ stw(R0, thread_(thread_state)); 2381 if (UseMembar) { 2382 __ fence(); 2383 } 2384 __ bind(after_transition); 2385 2386 // Reguard any pages if necessary. 2387 // -------------------------------------------------------------------------- 2388 2389 Label no_reguard; 2390 __ lwz(r_temp_1, thread_(stack_guard_state)); 2391 __ cmpwi(CCR0, r_temp_1, JavaThread::stack_guard_yellow_reserved_disabled); 2392 __ bne(CCR0, no_reguard); 2393 2394 save_native_result(masm, ret_type, workspace_slot_offset); 2395 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); 2396 restore_native_result(masm, ret_type, workspace_slot_offset); 2397 2398 __ bind(no_reguard); 2399 2400 2401 // Unlock 2402 // -------------------------------------------------------------------------- 2403 2404 if (method->is_synchronized()) { 2405 2406 ConditionRegister r_flag = CCR1; 2407 const Register r_oop = r_temp_4; 2408 const Register r_box = r_temp_5; 2409 const Register r_exception = r_temp_6; 2410 Label done; 2411 2412 // Get oop and address of lock object box. 2413 if (method_is_static) { 2414 assert(klass_offset != -1, ""); 2415 __ ld(r_oop, klass_offset, R1_SP); 2416 } else { 2417 assert(receiver_offset != -1, ""); 2418 __ ld(r_oop, receiver_offset, R1_SP); 2419 } 2420 __ addi(r_box, R1_SP, lock_offset); 2421 2422 // Try fastpath for unlocking. 2423 __ compiler_fast_unlock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3); 2424 __ beq(r_flag, done); 2425 2426 // Save and restore any potential method result value around the unlocking operation. 2427 save_native_result(masm, ret_type, workspace_slot_offset); 2428 2429 // Must save pending exception around the slow-path VM call. Since it's a 2430 // leaf call, the pending exception (if any) can be kept in a register. 2431 __ ld(r_exception, thread_(pending_exception)); 2432 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile"); 2433 __ li(R0, 0); 2434 __ std(R0, thread_(pending_exception)); 2435 2436 // Slow case of monitor enter. 2437 // Inline a special case of call_VM that disallows any pending_exception. 2438 // Arguments are (oop obj, BasicLock* lock, JavaThread* thread). 2439 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box, R16_thread); 2440 2441 __ asm_assert_mem8_is_zero(thread_(pending_exception), 2442 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C", 0); 2443 2444 restore_native_result(masm, ret_type, workspace_slot_offset); 2445 2446 // Check_forward_pending_exception jump to forward_exception if any pending 2447 // exception is set. The forward_exception routine expects to see the 2448 // exception in pending_exception and not in a register. Kind of clumsy, 2449 // since all folks who branch to forward_exception must have tested 2450 // pending_exception first and hence have it in a register already. 2451 __ std(r_exception, thread_(pending_exception)); 2452 2453 __ bind(done); 2454 } 2455 2456 # if 0 2457 // DTrace method exit 2458 # endif 2459 2460 // Clear "last Java frame" SP and PC. 2461 // -------------------------------------------------------------------------- 2462 2463 __ reset_last_Java_frame(); 2464 2465 // Unpack oop result. 2466 // -------------------------------------------------------------------------- 2467 2468 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2469 Label skip_unboxing; 2470 __ cmpdi(CCR0, R3_RET, 0); 2471 __ beq(CCR0, skip_unboxing); 2472 __ ld(R3_RET, 0, R3_RET); 2473 __ bind(skip_unboxing); 2474 __ verify_oop(R3_RET); 2475 } 2476 2477 2478 // Reset handle block. 2479 // -------------------------------------------------------------------------- 2480 if (!is_critical_native) { 2481 __ ld(r_temp_1, thread_(active_handles)); 2482 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size"); 2483 __ li(r_temp_2, 0); 2484 __ stw(r_temp_2, JNIHandleBlock::top_offset_in_bytes(), r_temp_1); 2485 2486 2487 // Check for pending exceptions. 2488 // -------------------------------------------------------------------------- 2489 __ ld(r_temp_2, thread_(pending_exception)); 2490 __ cmpdi(CCR0, r_temp_2, 0); 2491 __ bne(CCR0, handle_pending_exception); 2492 } 2493 2494 // Return 2495 // -------------------------------------------------------------------------- 2496 2497 __ pop_frame(); 2498 __ restore_LR_CR(R11); 2499 __ blr(); 2500 2501 2502 // Handler for pending exceptions (out-of-line). 2503 // -------------------------------------------------------------------------- 2504 2505 // Since this is a native call, we know the proper exception handler 2506 // is the empty function. We just pop this frame and then jump to 2507 // forward_exception_entry. 2508 if (!is_critical_native) { 2509 __ align(InteriorEntryAlignment); 2510 __ bind(handle_pending_exception); 2511 2512 __ pop_frame(); 2513 __ restore_LR_CR(R11); 2514 __ b64_patchable((address)StubRoutines::forward_exception_entry(), 2515 relocInfo::runtime_call_type); 2516 } 2517 2518 // Handler for a cache miss (out-of-line). 2519 // -------------------------------------------------------------------------- 2520 2521 if (!method_is_static) { 2522 __ align(InteriorEntryAlignment); 2523 __ bind(ic_miss); 2524 2525 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), 2526 relocInfo::runtime_call_type); 2527 } 2528 2529 // Done. 2530 // -------------------------------------------------------------------------- 2531 2532 __ flush(); 2533 2534 nmethod *nm = nmethod::new_native_nmethod(method, 2535 compile_id, 2536 masm->code(), 2537 vep_start_pc-start_pc, 2538 frame_done_pc-start_pc, 2539 stack_slots / VMRegImpl::slots_per_word, 2540 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2541 in_ByteSize(lock_offset), 2542 oop_maps); 2543 2544 if (is_critical_native) { 2545 nm->set_lazy_critical_native(true); 2546 } 2547 2548 return nm; 2549 #else 2550 ShouldNotReachHere(); 2551 return NULL; 2552 #endif // COMPILER2 2553 } 2554 2555 // This function returns the adjust size (in number of words) to a c2i adapter 2556 // activation for use during deoptimization. 2557 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 2558 return round_to((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::alignment_in_bytes); 2559 } 2560 2561 uint SharedRuntime::out_preserve_stack_slots() { 2562 #if defined(COMPILER1) || defined(COMPILER2) 2563 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size; 2564 #else 2565 return 0; 2566 #endif 2567 } 2568 2569 #ifdef COMPILER2 2570 // Frame generation for deopt and uncommon trap blobs. 2571 static void push_skeleton_frame(MacroAssembler* masm, bool deopt, 2572 /* Read */ 2573 Register unroll_block_reg, 2574 /* Update */ 2575 Register frame_sizes_reg, 2576 Register number_of_frames_reg, 2577 Register pcs_reg, 2578 /* Invalidate */ 2579 Register frame_size_reg, 2580 Register pc_reg) { 2581 2582 __ ld(pc_reg, 0, pcs_reg); 2583 __ ld(frame_size_reg, 0, frame_sizes_reg); 2584 __ std(pc_reg, _abi(lr), R1_SP); 2585 __ push_frame(frame_size_reg, R0/*tmp*/); 2586 #ifdef ASSERT 2587 __ load_const_optimized(pc_reg, 0x5afe); 2588 __ std(pc_reg, _ijava_state_neg(ijava_reserved), R1_SP); 2589 #endif 2590 __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP); 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 ASSERT 2663 __ load_const_optimized(pc_reg, 0x5afe); 2664 __ std(pc_reg, _ijava_state_neg(ijava_reserved), R1_SP); 2665 #endif 2666 __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP); 2667 2668 #ifdef ASSERT 2669 // Make sure that there is at least one entry in the array. 2670 __ cmpdi(CCR0, number_of_frames_reg, 0); 2671 __ asm_assert_ne("array_size must be > 0", 0x205); 2672 #endif 2673 2674 // Now push the new interpreter frames. 2675 // 2676 __ bind(loop); 2677 // Allocate a new frame, fill in the pc. 2678 push_skeleton_frame(masm, deopt, 2679 unroll_block_reg, 2680 frame_sizes_reg, 2681 number_of_frames_reg, 2682 pcs_reg, 2683 frame_size_reg, 2684 pc_reg); 2685 __ cmpdi(CCR0, number_of_frames_reg, 0); 2686 __ bne(CCR0, loop); 2687 2688 // Get the return address pointing into the frame manager. 2689 __ ld(R0, 0, pcs_reg); 2690 // Store it in the top interpreter frame. 2691 __ std(R0, _abi(lr), R1_SP); 2692 // Initialize frame_manager_lr of interpreter top frame. 2693 } 2694 #endif 2695 2696 void SharedRuntime::generate_deopt_blob() { 2697 // Allocate space for the code 2698 ResourceMark rm; 2699 // Setup code generation tools 2700 CodeBuffer buffer("deopt_blob", 2048, 1024); 2701 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer); 2702 Label exec_mode_initialized; 2703 int frame_size_in_words; 2704 OopMap* map = NULL; 2705 OopMapSet *oop_maps = new OopMapSet(); 2706 2707 // size of ABI112 plus spill slots for R3_RET and F1_RET. 2708 const int frame_size_in_bytes = frame::abi_reg_args_spill_size; 2709 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint); 2710 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info. 2711 2712 const Register exec_mode_reg = R21_tmp1; 2713 2714 const address start = __ pc(); 2715 2716 #ifdef COMPILER2 2717 // -------------------------------------------------------------------------- 2718 // Prolog for non exception case! 2719 2720 // We have been called from the deopt handler of the deoptee. 2721 // 2722 // deoptee: 2723 // ... 2724 // call X 2725 // ... 2726 // deopt_handler: call_deopt_stub 2727 // cur. return pc --> ... 2728 // 2729 // So currently SR_LR points behind the call in the deopt handler. 2730 // We adjust it such that it points to the start of the deopt handler. 2731 // The return_pc has been stored in the frame of the deoptee and 2732 // will replace the address of the deopt_handler in the call 2733 // to Deoptimization::fetch_unroll_info below. 2734 // We can't grab a free register here, because all registers may 2735 // contain live values, so let the RegisterSaver do the adjustment 2736 // of the return pc. 2737 const int return_pc_adjustment_no_exception = -HandlerImpl::size_deopt_handler(); 2738 2739 // Push the "unpack frame" 2740 // Save everything in sight. 2741 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 2742 &first_frame_size_in_bytes, 2743 /*generate_oop_map=*/ true, 2744 return_pc_adjustment_no_exception, 2745 RegisterSaver::return_pc_is_lr); 2746 assert(map != NULL, "OopMap must have been created"); 2747 2748 __ li(exec_mode_reg, Deoptimization::Unpack_deopt); 2749 // Save exec mode for unpack_frames. 2750 __ b(exec_mode_initialized); 2751 2752 // -------------------------------------------------------------------------- 2753 // Prolog for exception case 2754 2755 // An exception is pending. 2756 // We have been called with a return (interpreter) or a jump (exception blob). 2757 // 2758 // - R3_ARG1: exception oop 2759 // - R4_ARG2: exception pc 2760 2761 int exception_offset = __ pc() - start; 2762 2763 BLOCK_COMMENT("Prolog for exception case"); 2764 2765 // The RegisterSaves doesn't need to adjust the return pc for this situation. 2766 const int return_pc_adjustment_exception = 0; 2767 2768 // Push the "unpack frame". 2769 // Save everything in sight. 2770 assert(R4 == R4_ARG2, "exception pc must be in r4"); 2771 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 2772 &first_frame_size_in_bytes, 2773 /*generate_oop_map=*/ false, 2774 return_pc_adjustment_exception, 2775 RegisterSaver::return_pc_is_r4); 2776 2777 // Deopt during an exception. Save exec mode for unpack_frames. 2778 __ li(exec_mode_reg, Deoptimization::Unpack_exception); 2779 2780 // Store exception oop and pc in thread (location known to GC). 2781 // This is needed since the call to "fetch_unroll_info()" may safepoint. 2782 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2783 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2784 2785 // fall through 2786 2787 // -------------------------------------------------------------------------- 2788 __ BIND(exec_mode_initialized); 2789 2790 { 2791 const Register unroll_block_reg = R22_tmp2; 2792 2793 // We need to set `last_Java_frame' because `fetch_unroll_info' will 2794 // call `last_Java_frame()'. The value of the pc in the frame is not 2795 // particularly important. It just needs to identify this blob. 2796 __ set_last_Java_frame(R1_SP, noreg); 2797 2798 // With EscapeAnalysis turned on, this call may safepoint! 2799 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread, exec_mode_reg); 2800 address calls_return_pc = __ last_calls_return_pc(); 2801 // Set an oopmap for the call site that describes all our saved registers. 2802 oop_maps->add_gc_map(calls_return_pc - start, map); 2803 2804 __ reset_last_Java_frame(); 2805 // Save the return value. 2806 __ mr(unroll_block_reg, R3_RET); 2807 2808 // Restore only the result registers that have been saved 2809 // by save_volatile_registers(...). 2810 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes); 2811 2812 // reload the exec mode from the UnrollBlock (it might have changed) 2813 __ lwz(exec_mode_reg, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes(), unroll_block_reg); 2814 // In excp_deopt_mode, restore and clear exception oop which we 2815 // stored in the thread during exception entry above. The exception 2816 // oop will be the return value of this stub. 2817 Label skip_restore_excp; 2818 __ cmpdi(CCR0, exec_mode_reg, Deoptimization::Unpack_exception); 2819 __ bne(CCR0, skip_restore_excp); 2820 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2821 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2822 __ li(R0, 0); 2823 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread); 2824 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread); 2825 __ BIND(skip_restore_excp); 2826 2827 __ pop_frame(); 2828 2829 // stack: (deoptee, optional i2c, caller of deoptee, ...). 2830 2831 // pop the deoptee's frame 2832 __ pop_frame(); 2833 2834 // stack: (caller_of_deoptee, ...). 2835 2836 // Loop through the `UnrollBlock' info and create interpreter frames. 2837 push_skeleton_frames(masm, true/*deopt*/, 2838 unroll_block_reg, 2839 R23_tmp3, 2840 R24_tmp4, 2841 R25_tmp5, 2842 R26_tmp6, 2843 R27_tmp7); 2844 2845 // stack: (skeletal interpreter frame, ..., optional skeletal 2846 // interpreter frame, optional c2i, caller of deoptee, ...). 2847 } 2848 2849 // push an `unpack_frame' taking care of float / int return values. 2850 __ push_frame(frame_size_in_bytes, R0/*tmp*/); 2851 2852 // stack: (unpack frame, skeletal interpreter frame, ..., optional 2853 // skeletal interpreter frame, optional c2i, caller of deoptee, 2854 // ...). 2855 2856 // Spill live volatile registers since we'll do a call. 2857 __ std( R3_RET, _abi_reg_args_spill(spill_ret), R1_SP); 2858 __ stfd(F1_RET, _abi_reg_args_spill(spill_fret), R1_SP); 2859 2860 // Let the unpacker layout information in the skeletal frames just 2861 // allocated. 2862 __ get_PC_trash_LR(R3_RET); 2863 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET); 2864 // This is a call to a LEAF method, so no oop map is required. 2865 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), 2866 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/); 2867 __ reset_last_Java_frame(); 2868 2869 // Restore the volatiles saved above. 2870 __ ld( R3_RET, _abi_reg_args_spill(spill_ret), R1_SP); 2871 __ lfd(F1_RET, _abi_reg_args_spill(spill_fret), R1_SP); 2872 2873 // Pop the unpack frame. 2874 __ pop_frame(); 2875 __ restore_LR_CR(R0); 2876 2877 // stack: (top interpreter frame, ..., optional interpreter frame, 2878 // optional c2i, caller of deoptee, ...). 2879 2880 // Initialize R14_state. 2881 __ restore_interpreter_state(R11_scratch1); 2882 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 2883 2884 // Return to the interpreter entry point. 2885 __ blr(); 2886 __ flush(); 2887 #else // COMPILER2 2888 __ unimplemented("deopt blob needed only with compiler"); 2889 int exception_offset = __ pc() - start; 2890 #endif // COMPILER2 2891 2892 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, 0, first_frame_size_in_bytes / wordSize); 2893 } 2894 2895 #ifdef COMPILER2 2896 void SharedRuntime::generate_uncommon_trap_blob() { 2897 // Allocate space for the code. 2898 ResourceMark rm; 2899 // Setup code generation tools. 2900 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 2901 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer); 2902 address start = __ pc(); 2903 2904 Register unroll_block_reg = R21_tmp1; 2905 Register klass_index_reg = R22_tmp2; 2906 Register unc_trap_reg = R23_tmp3; 2907 2908 OopMapSet* oop_maps = new OopMapSet(); 2909 int frame_size_in_bytes = frame::abi_reg_args_size; 2910 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0); 2911 2912 // stack: (deoptee, optional i2c, caller_of_deoptee, ...). 2913 2914 // Push a dummy `unpack_frame' and call 2915 // `Deoptimization::uncommon_trap' to pack the compiled frame into a 2916 // vframe array and return the `UnrollBlock' information. 2917 2918 // Save LR to compiled frame. 2919 __ save_LR_CR(R11_scratch1); 2920 2921 // Push an "uncommon_trap" frame. 2922 __ push_frame_reg_args(0, R11_scratch1); 2923 2924 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...). 2925 2926 // Set the `unpack_frame' as last_Java_frame. 2927 // `Deoptimization::uncommon_trap' expects it and considers its 2928 // sender frame as the deoptee frame. 2929 // Remember the offset of the instruction whose address will be 2930 // moved to R11_scratch1. 2931 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1); 2932 2933 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 2934 2935 __ mr(klass_index_reg, R3); 2936 __ li(R5_ARG3, Deoptimization::Unpack_uncommon_trap); 2937 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), 2938 R16_thread, klass_index_reg, R5_ARG3); 2939 2940 // Set an oopmap for the call site. 2941 oop_maps->add_gc_map(gc_map_pc - start, map); 2942 2943 __ reset_last_Java_frame(); 2944 2945 // Pop the `unpack frame'. 2946 __ pop_frame(); 2947 2948 // stack: (deoptee, optional i2c, caller_of_deoptee, ...). 2949 2950 // Save the return value. 2951 __ mr(unroll_block_reg, R3_RET); 2952 2953 // Pop the uncommon_trap frame. 2954 __ pop_frame(); 2955 2956 // stack: (caller_of_deoptee, ...). 2957 2958 #ifdef ASSERT 2959 __ lwz(R22_tmp2, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes(), unroll_block_reg); 2960 __ cmpdi(CCR0, R22_tmp2, (unsigned)Deoptimization::Unpack_uncommon_trap); 2961 __ asm_assert_eq("SharedRuntime::generate_deopt_blob: expected Unpack_uncommon_trap", 0); 2962 #endif 2963 2964 // Allocate new interpreter frame(s) and possibly a c2i adapter 2965 // frame. 2966 push_skeleton_frames(masm, false/*deopt*/, 2967 unroll_block_reg, 2968 R22_tmp2, 2969 R23_tmp3, 2970 R24_tmp4, 2971 R25_tmp5, 2972 R26_tmp6); 2973 2974 // stack: (skeletal interpreter frame, ..., optional skeletal 2975 // interpreter frame, optional c2i, caller of deoptee, ...). 2976 2977 // Push a dummy `unpack_frame' taking care of float return values. 2978 // Call `Deoptimization::unpack_frames' to layout information in the 2979 // interpreter frames just created. 2980 2981 // Push a simple "unpack frame" here. 2982 __ push_frame_reg_args(0, R11_scratch1); 2983 2984 // stack: (unpack frame, skeletal interpreter frame, ..., optional 2985 // skeletal interpreter frame, optional c2i, caller of deoptee, 2986 // ...). 2987 2988 // Set the "unpack_frame" as last_Java_frame. 2989 __ get_PC_trash_LR(R11_scratch1); 2990 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1); 2991 2992 // Indicate it is the uncommon trap case. 2993 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap); 2994 // Let the unpacker layout information in the skeletal frames just 2995 // allocated. 2996 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), 2997 R16_thread, unc_trap_reg); 2998 2999 __ reset_last_Java_frame(); 3000 // Pop the `unpack frame'. 3001 __ pop_frame(); 3002 // Restore LR from top interpreter frame. 3003 __ restore_LR_CR(R11_scratch1); 3004 3005 // stack: (top interpreter frame, ..., optional interpreter frame, 3006 // optional c2i, caller of deoptee, ...). 3007 3008 __ restore_interpreter_state(R11_scratch1); 3009 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1); 3010 3011 // Return to the interpreter entry point. 3012 __ blr(); 3013 3014 masm->flush(); 3015 3016 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize); 3017 } 3018 #endif // COMPILER2 3019 3020 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap. 3021 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 3022 assert(StubRoutines::forward_exception_entry() != NULL, 3023 "must be generated before"); 3024 3025 ResourceMark rm; 3026 OopMapSet *oop_maps = new OopMapSet(); 3027 OopMap* map; 3028 3029 // Allocate space for the code. Setup code generation tools. 3030 CodeBuffer buffer("handler_blob", 2048, 1024); 3031 MacroAssembler* masm = new MacroAssembler(&buffer); 3032 3033 address start = __ pc(); 3034 int frame_size_in_bytes = 0; 3035 3036 RegisterSaver::ReturnPCLocation return_pc_location; 3037 bool cause_return = (poll_type == POLL_AT_RETURN); 3038 if (cause_return) { 3039 // Nothing to do here. The frame has already been popped in MachEpilogNode. 3040 // Register LR already contains the return pc. 3041 return_pc_location = RegisterSaver::return_pc_is_lr; 3042 } else { 3043 // Use thread()->saved_exception_pc() as return pc. 3044 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc; 3045 } 3046 3047 // Save registers, fpu state, and flags. 3048 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 3049 &frame_size_in_bytes, 3050 /*generate_oop_map=*/ true, 3051 /*return_pc_adjustment=*/0, 3052 return_pc_location); 3053 3054 // The following is basically a call_VM. However, we need the precise 3055 // address of the call in order to generate an oopmap. Hence, we do all the 3056 // work outselves. 3057 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg); 3058 3059 // The return address must always be correct so that the frame constructor 3060 // never sees an invalid pc. 3061 3062 // Do the call 3063 __ call_VM_leaf(call_ptr, R16_thread); 3064 address calls_return_pc = __ last_calls_return_pc(); 3065 3066 // Set an oopmap for the call site. This oopmap will map all 3067 // oop-registers and debug-info registers as callee-saved. This 3068 // will allow deoptimization at this safepoint to find all possible 3069 // debug-info recordings, as well as let GC find all oops. 3070 oop_maps->add_gc_map(calls_return_pc - start, map); 3071 3072 Label noException; 3073 3074 // Clear the last Java frame. 3075 __ reset_last_Java_frame(); 3076 3077 BLOCK_COMMENT(" Check pending exception."); 3078 const Register pending_exception = R0; 3079 __ ld(pending_exception, thread_(pending_exception)); 3080 __ cmpdi(CCR0, pending_exception, 0); 3081 __ beq(CCR0, noException); 3082 3083 // Exception pending 3084 RegisterSaver::restore_live_registers_and_pop_frame(masm, 3085 frame_size_in_bytes, 3086 /*restore_ctr=*/true); 3087 3088 BLOCK_COMMENT(" Jump to forward_exception_entry."); 3089 // Jump to forward_exception_entry, with the issuing PC in LR 3090 // so it looks like the original nmethod called forward_exception_entry. 3091 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 3092 3093 // No exception case. 3094 __ BIND(noException); 3095 3096 3097 // Normal exit, restore registers and exit. 3098 RegisterSaver::restore_live_registers_and_pop_frame(masm, 3099 frame_size_in_bytes, 3100 /*restore_ctr=*/true); 3101 3102 __ blr(); 3103 3104 // Make sure all code is generated 3105 masm->flush(); 3106 3107 // Fill-out other meta info 3108 // CodeBlob frame size is in words. 3109 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize); 3110 } 3111 3112 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss) 3113 // 3114 // Generate a stub that calls into the vm to find out the proper destination 3115 // of a java call. All the argument registers are live at this point 3116 // but since this is generic code we don't know what they are and the caller 3117 // must do any gc of the args. 3118 // 3119 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3120 3121 // allocate space for the code 3122 ResourceMark rm; 3123 3124 CodeBuffer buffer(name, 1000, 512); 3125 MacroAssembler* masm = new MacroAssembler(&buffer); 3126 3127 int frame_size_in_bytes; 3128 3129 OopMapSet *oop_maps = new OopMapSet(); 3130 OopMap* map = NULL; 3131 3132 address start = __ pc(); 3133 3134 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm, 3135 &frame_size_in_bytes, 3136 /*generate_oop_map*/ true, 3137 /*return_pc_adjustment*/ 0, 3138 RegisterSaver::return_pc_is_lr); 3139 3140 // Use noreg as last_Java_pc, the return pc will be reconstructed 3141 // from the physical frame. 3142 __ set_last_Java_frame(/*sp*/R1_SP, noreg); 3143 3144 int frame_complete = __ offset(); 3145 3146 // Pass R19_method as 2nd (optional) argument, used by 3147 // counter_overflow_stub. 3148 __ call_VM_leaf(destination, R16_thread, R19_method); 3149 address calls_return_pc = __ last_calls_return_pc(); 3150 // Set an oopmap for the call site. 3151 // We need this not only for callee-saved registers, but also for volatile 3152 // registers that the compiler might be keeping live across a safepoint. 3153 // Create the oopmap for the call's return pc. 3154 oop_maps->add_gc_map(calls_return_pc - start, map); 3155 3156 // R3_RET contains the address we are going to jump to assuming no exception got installed. 3157 3158 // clear last_Java_sp 3159 __ reset_last_Java_frame(); 3160 3161 // Check for pending exceptions. 3162 BLOCK_COMMENT("Check for pending exceptions."); 3163 Label pending; 3164 __ ld(R11_scratch1, thread_(pending_exception)); 3165 __ cmpdi(CCR0, R11_scratch1, 0); 3166 __ bne(CCR0, pending); 3167 3168 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame. 3169 3170 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false); 3171 3172 // Get the returned method. 3173 __ get_vm_result_2(R19_method); 3174 3175 __ bctr(); 3176 3177 3178 // Pending exception after the safepoint. 3179 __ BIND(pending); 3180 3181 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true); 3182 3183 // exception pending => remove activation and forward to exception handler 3184 3185 __ li(R11_scratch1, 0); 3186 __ ld(R3_ARG1, thread_(pending_exception)); 3187 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_offset()), R16_thread); 3188 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type); 3189 3190 // ------------- 3191 // Make sure all code is generated. 3192 masm->flush(); 3193 3194 // return the blob 3195 // frame_size_words or bytes?? 3196 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize, 3197 oop_maps, true); 3198 }