1 /* 2 * Copyright (c) 2003, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "code/icBuffer.hpp" 30 #include "code/vtableStubs.hpp" 31 #include "gc/shared/gcLocker.hpp" 32 #include "interpreter/interpreter.hpp" 33 #include "logging/log.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "oops/compiledICHolder.hpp" 36 #include "runtime/safepointMechanism.hpp" 37 #include "runtime/sharedRuntime.hpp" 38 #include "runtime/vframeArray.hpp" 39 #include "utilities/align.hpp" 40 #include "vmreg_x86.inline.hpp" 41 #ifdef COMPILER1 42 #include "c1/c1_Runtime1.hpp" 43 #endif 44 #ifdef COMPILER2 45 #include "opto/runtime.hpp" 46 #endif 47 #include "vm_version_x86.hpp" 48 49 #define __ masm-> 50 51 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 52 53 class RegisterSaver { 54 // Capture info about frame layout 55 #define DEF_XMM_OFFS(regnum) xmm ## regnum ## _off = xmm_off + (regnum)*16/BytesPerInt, xmm ## regnum ## H_off 56 enum layout { 57 fpu_state_off = 0, 58 fpu_state_end = fpu_state_off+FPUStateSizeInWords, 59 st0_off, st0H_off, 60 st1_off, st1H_off, 61 st2_off, st2H_off, 62 st3_off, st3H_off, 63 st4_off, st4H_off, 64 st5_off, st5H_off, 65 st6_off, st6H_off, 66 st7_off, st7H_off, 67 xmm_off, 68 DEF_XMM_OFFS(0), 69 DEF_XMM_OFFS(1), 70 DEF_XMM_OFFS(2), 71 DEF_XMM_OFFS(3), 72 DEF_XMM_OFFS(4), 73 DEF_XMM_OFFS(5), 74 DEF_XMM_OFFS(6), 75 DEF_XMM_OFFS(7), 76 flags_off = xmm7_off + 16/BytesPerInt + 1, // 16-byte stack alignment fill word 77 rdi_off, 78 rsi_off, 79 ignore_off, // extra copy of rbp, 80 rsp_off, 81 rbx_off, 82 rdx_off, 83 rcx_off, 84 rax_off, 85 // The frame sender code expects that rbp will be in the "natural" place and 86 // will override any oopMap setting for it. We must therefore force the layout 87 // so that it agrees with the frame sender code. 88 rbp_off, 89 return_off, // slot for return address 90 reg_save_size }; 91 enum { FPU_regs_live = flags_off - fpu_state_end }; 92 93 public: 94 95 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, 96 int* total_frame_words, bool verify_fpu = true, bool save_vectors = false); 97 static void restore_live_registers(MacroAssembler* masm, bool restore_vectors = false); 98 99 static int rax_offset() { return rax_off; } 100 static int rbx_offset() { return rbx_off; } 101 102 // Offsets into the register save area 103 // Used by deoptimization when it is managing result register 104 // values on its own 105 106 static int raxOffset(void) { return rax_off; } 107 static int rdxOffset(void) { return rdx_off; } 108 static int rbxOffset(void) { return rbx_off; } 109 static int xmm0Offset(void) { return xmm0_off; } 110 // This really returns a slot in the fp save area, which one is not important 111 static int fpResultOffset(void) { return st0_off; } 112 113 // During deoptimization only the result register need to be restored 114 // all the other values have already been extracted. 115 116 static void restore_result_registers(MacroAssembler* masm); 117 118 }; 119 120 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, 121 int* total_frame_words, bool verify_fpu, bool save_vectors) { 122 int num_xmm_regs = XMMRegisterImpl::number_of_registers; 123 int ymm_bytes = num_xmm_regs * 16; 124 int zmm_bytes = num_xmm_regs * 32; 125 #ifdef COMPILER2 126 if (save_vectors) { 127 assert(UseAVX > 0, "Vectors larger than 16 byte long are supported only with AVX"); 128 assert(MaxVectorSize <= 64, "Only up to 64 byte long vectors are supported"); 129 // Save upper half of YMM registers 130 int vect_bytes = ymm_bytes; 131 if (UseAVX > 2) { 132 // Save upper half of ZMM registers as well 133 vect_bytes += zmm_bytes; 134 } 135 additional_frame_words += vect_bytes / wordSize; 136 } 137 #else 138 assert(!save_vectors, "vectors are generated only by C2"); 139 #endif 140 int frame_size_in_bytes = (reg_save_size + additional_frame_words) * wordSize; 141 int frame_words = frame_size_in_bytes / wordSize; 142 *total_frame_words = frame_words; 143 144 assert(FPUStateSizeInWords == 27, "update stack layout"); 145 146 // save registers, fpu state, and flags 147 // We assume caller has already has return address slot on the stack 148 // We push epb twice in this sequence because we want the real rbp, 149 // to be under the return like a normal enter and we want to use pusha 150 // We push by hand instead of using push. 151 __ enter(); 152 __ pusha(); 153 __ pushf(); 154 __ subptr(rsp,FPU_regs_live*wordSize); // Push FPU registers space 155 __ push_FPU_state(); // Save FPU state & init 156 157 if (verify_fpu) { 158 // Some stubs may have non standard FPU control word settings so 159 // only check and reset the value when it required to be the 160 // standard value. The safepoint blob in particular can be used 161 // in methods which are using the 24 bit control word for 162 // optimized float math. 163 164 #ifdef ASSERT 165 // Make sure the control word has the expected value 166 Label ok; 167 __ cmpw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std()); 168 __ jccb(Assembler::equal, ok); 169 __ stop("corrupted control word detected"); 170 __ bind(ok); 171 #endif 172 173 // Reset the control word to guard against exceptions being unmasked 174 // since fstp_d can cause FPU stack underflow exceptions. Write it 175 // into the on stack copy and then reload that to make sure that the 176 // current and future values are correct. 177 __ movw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std()); 178 } 179 180 __ frstor(Address(rsp, 0)); 181 if (!verify_fpu) { 182 // Set the control word so that exceptions are masked for the 183 // following code. 184 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std())); 185 } 186 187 int off = st0_off; 188 int delta = st1_off - off; 189 190 // Save the FPU registers in de-opt-able form 191 for (int n = 0; n < FloatRegisterImpl::number_of_registers; n++) { 192 __ fstp_d(Address(rsp, off*wordSize)); 193 off += delta; 194 } 195 196 off = xmm0_off; 197 delta = xmm1_off - off; 198 if(UseSSE == 1) { 199 // Save the XMM state 200 for (int n = 0; n < num_xmm_regs; n++) { 201 __ movflt(Address(rsp, off*wordSize), as_XMMRegister(n)); 202 off += delta; 203 } 204 } else if(UseSSE >= 2) { 205 // Save whole 128bit (16 bytes) XMM registers 206 for (int n = 0; n < num_xmm_regs; n++) { 207 __ movdqu(Address(rsp, off*wordSize), as_XMMRegister(n)); 208 off += delta; 209 } 210 } 211 212 if (save_vectors) { 213 __ subptr(rsp, ymm_bytes); 214 // Save upper half of YMM registers 215 for (int n = 0; n < num_xmm_regs; n++) { 216 __ vextractf128_high(Address(rsp, n*16), as_XMMRegister(n)); 217 } 218 if (UseAVX > 2) { 219 __ subptr(rsp, zmm_bytes); 220 // Save upper half of ZMM registers 221 for (int n = 0; n < num_xmm_regs; n++) { 222 __ vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n)); 223 } 224 } 225 } 226 __ vzeroupper(); 227 228 // Set an oopmap for the call site. This oopmap will map all 229 // oop-registers and debug-info registers as callee-saved. This 230 // will allow deoptimization at this safepoint to find all possible 231 // debug-info recordings, as well as let GC find all oops. 232 233 OopMapSet *oop_maps = new OopMapSet(); 234 OopMap* map = new OopMap( frame_words, 0 ); 235 236 #define STACK_OFFSET(x) VMRegImpl::stack2reg((x) + additional_frame_words) 237 #define NEXTREG(x) (x)->as_VMReg()->next() 238 239 map->set_callee_saved(STACK_OFFSET(rax_off), rax->as_VMReg()); 240 map->set_callee_saved(STACK_OFFSET(rcx_off), rcx->as_VMReg()); 241 map->set_callee_saved(STACK_OFFSET(rdx_off), rdx->as_VMReg()); 242 map->set_callee_saved(STACK_OFFSET(rbx_off), rbx->as_VMReg()); 243 // rbp, location is known implicitly, no oopMap 244 map->set_callee_saved(STACK_OFFSET(rsi_off), rsi->as_VMReg()); 245 map->set_callee_saved(STACK_OFFSET(rdi_off), rdi->as_VMReg()); 246 // %%% This is really a waste but we'll keep things as they were for now for the upper component 247 off = st0_off; 248 delta = st1_off - off; 249 for (int n = 0; n < FloatRegisterImpl::number_of_registers; n++) { 250 FloatRegister freg_name = as_FloatRegister(n); 251 map->set_callee_saved(STACK_OFFSET(off), freg_name->as_VMReg()); 252 map->set_callee_saved(STACK_OFFSET(off+1), NEXTREG(freg_name)); 253 off += delta; 254 } 255 off = xmm0_off; 256 delta = xmm1_off - off; 257 for (int n = 0; n < num_xmm_regs; n++) { 258 XMMRegister xmm_name = as_XMMRegister(n); 259 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg()); 260 map->set_callee_saved(STACK_OFFSET(off+1), NEXTREG(xmm_name)); 261 off += delta; 262 } 263 #undef NEXTREG 264 #undef STACK_OFFSET 265 266 return map; 267 } 268 269 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_vectors) { 270 int num_xmm_regs = XMMRegisterImpl::number_of_registers; 271 int ymm_bytes = num_xmm_regs * 16; 272 int zmm_bytes = num_xmm_regs * 32; 273 // Recover XMM & FPU state 274 int additional_frame_bytes = 0; 275 #ifdef COMPILER2 276 if (restore_vectors) { 277 assert(UseAVX > 0, "Vectors larger than 16 byte long are supported only with AVX"); 278 assert(MaxVectorSize <= 64, "Only up to 64 byte long vectors are supported"); 279 // Save upper half of YMM registers 280 additional_frame_bytes = ymm_bytes; 281 if (UseAVX > 2) { 282 // Save upper half of ZMM registers as well 283 additional_frame_bytes += zmm_bytes; 284 } 285 } 286 #else 287 assert(!restore_vectors, "vectors are generated only by C2"); 288 #endif 289 290 int off = xmm0_off; 291 int delta = xmm1_off - off; 292 293 __ vzeroupper(); 294 295 if (UseSSE == 1) { 296 // Restore XMM registers 297 assert(additional_frame_bytes == 0, ""); 298 for (int n = 0; n < num_xmm_regs; n++) { 299 __ movflt(as_XMMRegister(n), Address(rsp, off*wordSize)); 300 off += delta; 301 } 302 } else if (UseSSE >= 2) { 303 // Restore whole 128bit (16 bytes) XMM registers. Do this before restoring YMM and 304 // ZMM because the movdqu instruction zeros the upper part of the XMM register. 305 for (int n = 0; n < num_xmm_regs; n++) { 306 __ movdqu(as_XMMRegister(n), Address(rsp, off*wordSize+additional_frame_bytes)); 307 off += delta; 308 } 309 } 310 311 if (restore_vectors) { 312 if (UseAVX > 2) { 313 // Restore upper half of ZMM registers. 314 for (int n = 0; n < num_xmm_regs; n++) { 315 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32)); 316 } 317 __ addptr(rsp, zmm_bytes); 318 } 319 // Restore upper half of YMM registers. 320 for (int n = 0; n < num_xmm_regs; n++) { 321 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16)); 322 } 323 __ addptr(rsp, ymm_bytes); 324 } 325 326 __ pop_FPU_state(); 327 __ addptr(rsp, FPU_regs_live*wordSize); // Pop FPU registers 328 329 __ popf(); 330 __ popa(); 331 // Get the rbp, described implicitly by the frame sender code (no oopMap) 332 __ pop(rbp); 333 } 334 335 void RegisterSaver::restore_result_registers(MacroAssembler* masm) { 336 337 // Just restore result register. Only used by deoptimization. By 338 // now any callee save register that needs to be restore to a c2 339 // caller of the deoptee has been extracted into the vframeArray 340 // and will be stuffed into the c2i adapter we create for later 341 // restoration so only result registers need to be restored here. 342 // 343 344 __ frstor(Address(rsp, 0)); // Restore fpu state 345 346 // Recover XMM & FPU state 347 if( UseSSE == 1 ) { 348 __ movflt(xmm0, Address(rsp, xmm0_off*wordSize)); 349 } else if( UseSSE >= 2 ) { 350 __ movdbl(xmm0, Address(rsp, xmm0_off*wordSize)); 351 } 352 __ movptr(rax, Address(rsp, rax_off*wordSize)); 353 __ movptr(rdx, Address(rsp, rdx_off*wordSize)); 354 // Pop all of the register save are off the stack except the return address 355 __ addptr(rsp, return_off * wordSize); 356 } 357 358 // Is vector's size (in bytes) bigger than a size saved by default? 359 // 16 bytes XMM registers are saved by default using SSE2 movdqu instructions. 360 // Note, MaxVectorSize == 0 with UseSSE < 2 and vectors are not generated. 361 bool SharedRuntime::is_wide_vector(int size) { 362 return size > 16; 363 } 364 365 size_t SharedRuntime::trampoline_size() { 366 return 16; 367 } 368 369 void SharedRuntime::generate_trampoline(MacroAssembler *masm, address destination) { 370 __ jump(RuntimeAddress(destination)); 371 } 372 373 // The java_calling_convention describes stack locations as ideal slots on 374 // a frame with no abi restrictions. Since we must observe abi restrictions 375 // (like the placement of the register window) the slots must be biased by 376 // the following value. 377 static int reg2offset_in(VMReg r) { 378 // Account for saved rbp, and return address 379 // This should really be in_preserve_stack_slots 380 return (r->reg2stack() + 2) * VMRegImpl::stack_slot_size; 381 } 382 383 static int reg2offset_out(VMReg r) { 384 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 385 } 386 387 // --------------------------------------------------------------------------- 388 // Read the array of BasicTypes from a signature, and compute where the 389 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 390 // quantities. Values less than SharedInfo::stack0 are registers, those above 391 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 392 // as framesizes are fixed. 393 // VMRegImpl::stack0 refers to the first slot 0(sp). 394 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register 395 // up to RegisterImpl::number_of_registers) are the 32-bit 396 // integer registers. 397 398 // Pass first two oop/int args in registers ECX and EDX. 399 // Pass first two float/double args in registers XMM0 and XMM1. 400 // Doubles have precedence, so if you pass a mix of floats and doubles 401 // the doubles will grab the registers before the floats will. 402 403 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 404 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 405 // units regardless of build. Of course for i486 there is no 64 bit build 406 407 408 // --------------------------------------------------------------------------- 409 // The compiled Java calling convention. 410 // Pass first two oop/int args in registers ECX and EDX. 411 // Pass first two float/double args in registers XMM0 and XMM1. 412 // Doubles have precedence, so if you pass a mix of floats and doubles 413 // the doubles will grab the registers before the floats will. 414 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 415 VMRegPair *regs, 416 int total_args_passed, 417 int is_outgoing) { 418 uint stack = 0; // Starting stack position for args on stack 419 420 421 // Pass first two oop/int args in registers ECX and EDX. 422 uint reg_arg0 = 9999; 423 uint reg_arg1 = 9999; 424 425 // Pass first two float/double args in registers XMM0 and XMM1. 426 // Doubles have precedence, so if you pass a mix of floats and doubles 427 // the doubles will grab the registers before the floats will. 428 // CNC - TURNED OFF FOR non-SSE. 429 // On Intel we have to round all doubles (and most floats) at 430 // call sites by storing to the stack in any case. 431 // UseSSE=0 ==> Don't Use ==> 9999+0 432 // UseSSE=1 ==> Floats only ==> 9999+1 433 // UseSSE>=2 ==> Floats or doubles ==> 9999+2 434 enum { fltarg_dontuse = 9999+0, fltarg_float_only = 9999+1, fltarg_flt_dbl = 9999+2 }; 435 uint fargs = (UseSSE>=2) ? 2 : UseSSE; 436 uint freg_arg0 = 9999+fargs; 437 uint freg_arg1 = 9999+fargs; 438 439 // Pass doubles & longs aligned on the stack. First count stack slots for doubles 440 int i; 441 for( i = 0; i < total_args_passed; i++) { 442 if( sig_bt[i] == T_DOUBLE ) { 443 // first 2 doubles go in registers 444 if( freg_arg0 == fltarg_flt_dbl ) freg_arg0 = i; 445 else if( freg_arg1 == fltarg_flt_dbl ) freg_arg1 = i; 446 else // Else double is passed low on the stack to be aligned. 447 stack += 2; 448 } else if( sig_bt[i] == T_LONG ) { 449 stack += 2; 450 } 451 } 452 int dstack = 0; // Separate counter for placing doubles 453 454 // Now pick where all else goes. 455 for( i = 0; i < total_args_passed; i++) { 456 // From the type and the argument number (count) compute the location 457 switch( sig_bt[i] ) { 458 case T_SHORT: 459 case T_CHAR: 460 case T_BYTE: 461 case T_BOOLEAN: 462 case T_INT: 463 case T_ARRAY: 464 case T_OBJECT: 465 case T_ADDRESS: 466 if( reg_arg0 == 9999 ) { 467 reg_arg0 = i; 468 regs[i].set1(rcx->as_VMReg()); 469 } else if( reg_arg1 == 9999 ) { 470 reg_arg1 = i; 471 regs[i].set1(rdx->as_VMReg()); 472 } else { 473 regs[i].set1(VMRegImpl::stack2reg(stack++)); 474 } 475 break; 476 case T_FLOAT: 477 if( freg_arg0 == fltarg_flt_dbl || freg_arg0 == fltarg_float_only ) { 478 freg_arg0 = i; 479 regs[i].set1(xmm0->as_VMReg()); 480 } else if( freg_arg1 == fltarg_flt_dbl || freg_arg1 == fltarg_float_only ) { 481 freg_arg1 = i; 482 regs[i].set1(xmm1->as_VMReg()); 483 } else { 484 regs[i].set1(VMRegImpl::stack2reg(stack++)); 485 } 486 break; 487 case T_LONG: 488 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "missing Half" ); 489 regs[i].set2(VMRegImpl::stack2reg(dstack)); 490 dstack += 2; 491 break; 492 case T_DOUBLE: 493 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "missing Half" ); 494 if( freg_arg0 == (uint)i ) { 495 regs[i].set2(xmm0->as_VMReg()); 496 } else if( freg_arg1 == (uint)i ) { 497 regs[i].set2(xmm1->as_VMReg()); 498 } else { 499 regs[i].set2(VMRegImpl::stack2reg(dstack)); 500 dstack += 2; 501 } 502 break; 503 case T_VOID: regs[i].set_bad(); break; 504 break; 505 default: 506 ShouldNotReachHere(); 507 break; 508 } 509 } 510 511 // return value can be odd number of VMRegImpl stack slots make multiple of 2 512 return align_up(stack, 2); 513 } 514 515 // Patch the callers callsite with entry to compiled code if it exists. 516 static void patch_callers_callsite(MacroAssembler *masm) { 517 Label L; 518 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD); 519 __ jcc(Assembler::equal, L); 520 // Schedule the branch target address early. 521 // Call into the VM to patch the caller, then jump to compiled callee 522 // rax, isn't live so capture return address while we easily can 523 __ movptr(rax, Address(rsp, 0)); 524 __ pusha(); 525 __ pushf(); 526 527 if (UseSSE == 1) { 528 __ subptr(rsp, 2*wordSize); 529 __ movflt(Address(rsp, 0), xmm0); 530 __ movflt(Address(rsp, wordSize), xmm1); 531 } 532 if (UseSSE >= 2) { 533 __ subptr(rsp, 4*wordSize); 534 __ movdbl(Address(rsp, 0), xmm0); 535 __ movdbl(Address(rsp, 2*wordSize), xmm1); 536 } 537 #ifdef COMPILER2 538 // C2 may leave the stack dirty if not in SSE2+ mode 539 if (UseSSE >= 2) { 540 __ verify_FPU(0, "c2i transition should have clean FPU stack"); 541 } else { 542 __ empty_FPU_stack(); 543 } 544 #endif /* COMPILER2 */ 545 546 // VM needs caller's callsite 547 __ push(rax); 548 // VM needs target method 549 __ push(rbx); 550 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite))); 551 __ addptr(rsp, 2*wordSize); 552 553 if (UseSSE == 1) { 554 __ movflt(xmm0, Address(rsp, 0)); 555 __ movflt(xmm1, Address(rsp, wordSize)); 556 __ addptr(rsp, 2*wordSize); 557 } 558 if (UseSSE >= 2) { 559 __ movdbl(xmm0, Address(rsp, 0)); 560 __ movdbl(xmm1, Address(rsp, 2*wordSize)); 561 __ addptr(rsp, 4*wordSize); 562 } 563 564 __ popf(); 565 __ popa(); 566 __ bind(L); 567 } 568 569 570 static void move_c2i_double(MacroAssembler *masm, XMMRegister r, int st_off) { 571 int next_off = st_off - Interpreter::stackElementSize; 572 __ movdbl(Address(rsp, next_off), r); 573 } 574 575 static void gen_c2i_adapter(MacroAssembler *masm, 576 int total_args_passed, 577 int comp_args_on_stack, 578 const BasicType *sig_bt, 579 const VMRegPair *regs, 580 Label& skip_fixup) { 581 // Before we get into the guts of the C2I adapter, see if we should be here 582 // at all. We've come from compiled code and are attempting to jump to the 583 // interpreter, which means the caller made a static call to get here 584 // (vcalls always get a compiled target if there is one). Check for a 585 // compiled target. If there is one, we need to patch the caller's call. 586 patch_callers_callsite(masm); 587 588 __ bind(skip_fixup); 589 590 #ifdef COMPILER2 591 // C2 may leave the stack dirty if not in SSE2+ mode 592 if (UseSSE >= 2) { 593 __ verify_FPU(0, "c2i transition should have clean FPU stack"); 594 } else { 595 __ empty_FPU_stack(); 596 } 597 #endif /* COMPILER2 */ 598 599 // Since all args are passed on the stack, total_args_passed * interpreter_ 600 // stack_element_size is the 601 // space we need. 602 int extraspace = total_args_passed * Interpreter::stackElementSize; 603 604 // Get return address 605 __ pop(rax); 606 607 // set senderSP value 608 __ movptr(rsi, rsp); 609 610 __ subptr(rsp, extraspace); 611 612 // Now write the args into the outgoing interpreter space 613 for (int i = 0; i < total_args_passed; i++) { 614 if (sig_bt[i] == T_VOID) { 615 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 616 continue; 617 } 618 619 // st_off points to lowest address on stack. 620 int st_off = ((total_args_passed - 1) - i) * Interpreter::stackElementSize; 621 int next_off = st_off - Interpreter::stackElementSize; 622 623 // Say 4 args: 624 // i st_off 625 // 0 12 T_LONG 626 // 1 8 T_VOID 627 // 2 4 T_OBJECT 628 // 3 0 T_BOOL 629 VMReg r_1 = regs[i].first(); 630 VMReg r_2 = regs[i].second(); 631 if (!r_1->is_valid()) { 632 assert(!r_2->is_valid(), ""); 633 continue; 634 } 635 636 if (r_1->is_stack()) { 637 // memory to memory use fpu stack top 638 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace; 639 640 if (!r_2->is_valid()) { 641 __ movl(rdi, Address(rsp, ld_off)); 642 __ movptr(Address(rsp, st_off), rdi); 643 } else { 644 645 // ld_off == LSW, ld_off+VMRegImpl::stack_slot_size == MSW 646 // st_off == MSW, st_off-wordSize == LSW 647 648 __ movptr(rdi, Address(rsp, ld_off)); 649 __ movptr(Address(rsp, next_off), rdi); 650 #ifndef _LP64 651 __ movptr(rdi, Address(rsp, ld_off + wordSize)); 652 __ movptr(Address(rsp, st_off), rdi); 653 #else 654 #ifdef ASSERT 655 // Overwrite the unused slot with known junk 656 __ mov64(rax, CONST64(0xdeadffffdeadaaaa)); 657 __ movptr(Address(rsp, st_off), rax); 658 #endif /* ASSERT */ 659 #endif // _LP64 660 } 661 } else if (r_1->is_Register()) { 662 Register r = r_1->as_Register(); 663 if (!r_2->is_valid()) { 664 __ movl(Address(rsp, st_off), r); 665 } else { 666 // long/double in gpr 667 NOT_LP64(ShouldNotReachHere()); 668 // Two VMRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 669 // T_DOUBLE and T_LONG use two slots in the interpreter 670 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 671 // long/double in gpr 672 #ifdef ASSERT 673 // Overwrite the unused slot with known junk 674 LP64_ONLY(__ mov64(rax, CONST64(0xdeadffffdeadaaab))); 675 __ movptr(Address(rsp, st_off), rax); 676 #endif /* ASSERT */ 677 __ movptr(Address(rsp, next_off), r); 678 } else { 679 __ movptr(Address(rsp, st_off), r); 680 } 681 } 682 } else { 683 assert(r_1->is_XMMRegister(), ""); 684 if (!r_2->is_valid()) { 685 __ movflt(Address(rsp, st_off), r_1->as_XMMRegister()); 686 } else { 687 assert(sig_bt[i] == T_DOUBLE || sig_bt[i] == T_LONG, "wrong type"); 688 move_c2i_double(masm, r_1->as_XMMRegister(), st_off); 689 } 690 } 691 } 692 693 // Schedule the branch target address early. 694 __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset()))); 695 // And repush original return address 696 __ push(rax); 697 __ jmp(rcx); 698 } 699 700 701 static void move_i2c_double(MacroAssembler *masm, XMMRegister r, Register saved_sp, int ld_off) { 702 int next_val_off = ld_off - Interpreter::stackElementSize; 703 __ movdbl(r, Address(saved_sp, next_val_off)); 704 } 705 706 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg, 707 address code_start, address code_end, 708 Label& L_ok) { 709 Label L_fail; 710 __ lea(temp_reg, ExternalAddress(code_start)); 711 __ cmpptr(pc_reg, temp_reg); 712 __ jcc(Assembler::belowEqual, L_fail); 713 __ lea(temp_reg, ExternalAddress(code_end)); 714 __ cmpptr(pc_reg, temp_reg); 715 __ jcc(Assembler::below, L_ok); 716 __ bind(L_fail); 717 } 718 719 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 720 int total_args_passed, 721 int comp_args_on_stack, 722 const BasicType *sig_bt, 723 const VMRegPair *regs) { 724 // Note: rsi contains the senderSP on entry. We must preserve it since 725 // we may do a i2c -> c2i transition if we lose a race where compiled 726 // code goes non-entrant while we get args ready. 727 728 // Adapters can be frameless because they do not require the caller 729 // to perform additional cleanup work, such as correcting the stack pointer. 730 // An i2c adapter is frameless because the *caller* frame, which is interpreted, 731 // routinely repairs its own stack pointer (from interpreter_frame_last_sp), 732 // even if a callee has modified the stack pointer. 733 // A c2i adapter is frameless because the *callee* frame, which is interpreted, 734 // routinely repairs its caller's stack pointer (from sender_sp, which is set 735 // up via the senderSP register). 736 // In other words, if *either* the caller or callee is interpreted, we can 737 // get the stack pointer repaired after a call. 738 // This is why c2i and i2c adapters cannot be indefinitely composed. 739 // In particular, if a c2i adapter were to somehow call an i2c adapter, 740 // both caller and callee would be compiled methods, and neither would 741 // clean up the stack pointer changes performed by the two adapters. 742 // If this happens, control eventually transfers back to the compiled 743 // caller, but with an uncorrected stack, causing delayed havoc. 744 745 // Pick up the return address 746 __ movptr(rax, Address(rsp, 0)); 747 748 if (VerifyAdapterCalls && 749 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) { 750 // So, let's test for cascading c2i/i2c adapters right now. 751 // assert(Interpreter::contains($return_addr) || 752 // StubRoutines::contains($return_addr), 753 // "i2c adapter must return to an interpreter frame"); 754 __ block_comment("verify_i2c { "); 755 Label L_ok; 756 if (Interpreter::code() != NULL) 757 range_check(masm, rax, rdi, 758 Interpreter::code()->code_start(), Interpreter::code()->code_end(), 759 L_ok); 760 if (StubRoutines::code1() != NULL) 761 range_check(masm, rax, rdi, 762 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(), 763 L_ok); 764 if (StubRoutines::code2() != NULL) 765 range_check(masm, rax, rdi, 766 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(), 767 L_ok); 768 const char* msg = "i2c adapter must return to an interpreter frame"; 769 __ block_comment(msg); 770 __ stop(msg); 771 __ bind(L_ok); 772 __ block_comment("} verify_i2ce "); 773 } 774 775 // Must preserve original SP for loading incoming arguments because 776 // we need to align the outgoing SP for compiled code. 777 __ movptr(rdi, rsp); 778 779 // Cut-out for having no stack args. Since up to 2 int/oop args are passed 780 // in registers, we will occasionally have no stack args. 781 int comp_words_on_stack = 0; 782 if (comp_args_on_stack) { 783 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in 784 // registers are below. By subtracting stack0, we either get a negative 785 // number (all values in registers) or the maximum stack slot accessed. 786 // int comp_args_on_stack = VMRegImpl::reg2stack(max_arg); 787 // Convert 4-byte stack slots to words. 788 comp_words_on_stack = align_up(comp_args_on_stack*4, wordSize)>>LogBytesPerWord; 789 // Round up to miminum stack alignment, in wordSize 790 comp_words_on_stack = align_up(comp_words_on_stack, 2); 791 __ subptr(rsp, comp_words_on_stack * wordSize); 792 } 793 794 // Align the outgoing SP 795 __ andptr(rsp, -(StackAlignmentInBytes)); 796 797 // push the return address on the stack (note that pushing, rather 798 // than storing it, yields the correct frame alignment for the callee) 799 __ push(rax); 800 801 // Put saved SP in another register 802 const Register saved_sp = rax; 803 __ movptr(saved_sp, rdi); 804 805 806 // Will jump to the compiled code just as if compiled code was doing it. 807 // Pre-load the register-jump target early, to schedule it better. 808 __ movptr(rdi, Address(rbx, in_bytes(Method::from_compiled_offset()))); 809 810 // Now generate the shuffle code. Pick up all register args and move the 811 // rest through the floating point stack top. 812 for (int i = 0; i < total_args_passed; i++) { 813 if (sig_bt[i] == T_VOID) { 814 // Longs and doubles are passed in native word order, but misaligned 815 // in the 32-bit build. 816 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 817 continue; 818 } 819 820 // Pick up 0, 1 or 2 words from SP+offset. 821 822 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 823 "scrambled load targets?"); 824 // Load in argument order going down. 825 int ld_off = (total_args_passed - i) * Interpreter::stackElementSize; 826 // Point to interpreter value (vs. tag) 827 int next_off = ld_off - Interpreter::stackElementSize; 828 // 829 // 830 // 831 VMReg r_1 = regs[i].first(); 832 VMReg r_2 = regs[i].second(); 833 if (!r_1->is_valid()) { 834 assert(!r_2->is_valid(), ""); 835 continue; 836 } 837 if (r_1->is_stack()) { 838 // Convert stack slot to an SP offset (+ wordSize to account for return address ) 839 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize; 840 841 // We can use rsi as a temp here because compiled code doesn't need rsi as an input 842 // and if we end up going thru a c2i because of a miss a reasonable value of rsi 843 // we be generated. 844 if (!r_2->is_valid()) { 845 // __ fld_s(Address(saved_sp, ld_off)); 846 // __ fstp_s(Address(rsp, st_off)); 847 __ movl(rsi, Address(saved_sp, ld_off)); 848 __ movptr(Address(rsp, st_off), rsi); 849 } else { 850 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 851 // are accessed as negative so LSW is at LOW address 852 853 // ld_off is MSW so get LSW 854 // st_off is LSW (i.e. reg.first()) 855 // __ fld_d(Address(saved_sp, next_off)); 856 // __ fstp_d(Address(rsp, st_off)); 857 // 858 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE 859 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case 860 // So we must adjust where to pick up the data to match the interpreter. 861 // 862 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 863 // are accessed as negative so LSW is at LOW address 864 865 // ld_off is MSW so get LSW 866 const int offset = (NOT_LP64(true ||) sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 867 next_off : ld_off; 868 __ movptr(rsi, Address(saved_sp, offset)); 869 __ movptr(Address(rsp, st_off), rsi); 870 #ifndef _LP64 871 __ movptr(rsi, Address(saved_sp, ld_off)); 872 __ movptr(Address(rsp, st_off + wordSize), rsi); 873 #endif // _LP64 874 } 875 } else if (r_1->is_Register()) { // Register argument 876 Register r = r_1->as_Register(); 877 assert(r != rax, "must be different"); 878 if (r_2->is_valid()) { 879 // 880 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE 881 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case 882 // So we must adjust where to pick up the data to match the interpreter. 883 884 const int offset = (NOT_LP64(true ||) sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 885 next_off : ld_off; 886 887 // this can be a misaligned move 888 __ movptr(r, Address(saved_sp, offset)); 889 #ifndef _LP64 890 assert(r_2->as_Register() != rax, "need another temporary register"); 891 // Remember r_1 is low address (and LSB on x86) 892 // So r_2 gets loaded from high address regardless of the platform 893 __ movptr(r_2->as_Register(), Address(saved_sp, ld_off)); 894 #endif // _LP64 895 } else { 896 __ movl(r, Address(saved_sp, ld_off)); 897 } 898 } else { 899 assert(r_1->is_XMMRegister(), ""); 900 if (!r_2->is_valid()) { 901 __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off)); 902 } else { 903 move_i2c_double(masm, r_1->as_XMMRegister(), saved_sp, ld_off); 904 } 905 } 906 } 907 908 // 6243940 We might end up in handle_wrong_method if 909 // the callee is deoptimized as we race thru here. If that 910 // happens we don't want to take a safepoint because the 911 // caller frame will look interpreted and arguments are now 912 // "compiled" so it is much better to make this transition 913 // invisible to the stack walking code. Unfortunately if 914 // we try and find the callee by normal means a safepoint 915 // is possible. So we stash the desired callee in the thread 916 // and the vm will find there should this case occur. 917 918 __ get_thread(rax); 919 __ movptr(Address(rax, JavaThread::callee_target_offset()), rbx); 920 921 // move Method* to rax, in case we end up in an c2i adapter. 922 // the c2i adapters expect Method* in rax, (c2) because c2's 923 // resolve stubs return the result (the method) in rax,. 924 // I'd love to fix this. 925 __ mov(rax, rbx); 926 927 __ jmp(rdi); 928 } 929 930 // --------------------------------------------------------------- 931 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 932 int total_args_passed, 933 int comp_args_on_stack, 934 const BasicType *sig_bt, 935 const VMRegPair *regs, 936 AdapterFingerPrint* fingerprint) { 937 address i2c_entry = __ pc(); 938 939 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 940 941 // ------------------------------------------------------------------------- 942 // Generate a C2I adapter. On entry we know rbx, holds the Method* during calls 943 // to the interpreter. The args start out packed in the compiled layout. They 944 // need to be unpacked into the interpreter layout. This will almost always 945 // require some stack space. We grow the current (compiled) stack, then repack 946 // the args. We finally end in a jump to the generic interpreter entry point. 947 // On exit from the interpreter, the interpreter will restore our SP (lest the 948 // compiled code, which relys solely on SP and not EBP, get sick). 949 950 address c2i_unverified_entry = __ pc(); 951 Label skip_fixup; 952 953 Register holder = rax; 954 Register receiver = rcx; 955 Register temp = rbx; 956 957 { 958 959 Label missed; 960 __ movptr(temp, Address(receiver, oopDesc::klass_offset_in_bytes())); 961 __ cmpptr(temp, Address(holder, CompiledICHolder::holder_klass_offset())); 962 __ movptr(rbx, Address(holder, CompiledICHolder::holder_metadata_offset())); 963 __ jcc(Assembler::notEqual, missed); 964 // Method might have been compiled since the call site was patched to 965 // interpreted if that is the case treat it as a miss so we can get 966 // the call site corrected. 967 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD); 968 __ jcc(Assembler::equal, skip_fixup); 969 970 __ bind(missed); 971 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 972 } 973 974 address c2i_entry = __ pc(); 975 976 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); 977 978 __ flush(); 979 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 980 } 981 982 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 983 VMRegPair *regs, 984 VMRegPair *regs2, 985 int total_args_passed) { 986 assert(regs2 == NULL, "not needed on x86"); 987 // We return the amount of VMRegImpl stack slots we need to reserve for all 988 // the arguments NOT counting out_preserve_stack_slots. 989 990 uint stack = 0; // All arguments on stack 991 992 for( int i = 0; i < total_args_passed; i++) { 993 // From the type and the argument number (count) compute the location 994 switch( sig_bt[i] ) { 995 case T_BOOLEAN: 996 case T_CHAR: 997 case T_FLOAT: 998 case T_BYTE: 999 case T_SHORT: 1000 case T_INT: 1001 case T_OBJECT: 1002 case T_ARRAY: 1003 case T_ADDRESS: 1004 case T_METADATA: 1005 regs[i].set1(VMRegImpl::stack2reg(stack++)); 1006 break; 1007 case T_LONG: 1008 case T_DOUBLE: // The stack numbering is reversed from Java 1009 // Since C arguments do not get reversed, the ordering for 1010 // doubles on the stack must be opposite the Java convention 1011 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "missing Half" ); 1012 regs[i].set2(VMRegImpl::stack2reg(stack)); 1013 stack += 2; 1014 break; 1015 case T_VOID: regs[i].set_bad(); break; 1016 default: 1017 ShouldNotReachHere(); 1018 break; 1019 } 1020 } 1021 return stack; 1022 } 1023 1024 // A simple move of integer like type 1025 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1026 if (src.first()->is_stack()) { 1027 if (dst.first()->is_stack()) { 1028 // stack to stack 1029 // __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); 1030 // __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS); 1031 __ movl2ptr(rax, Address(rbp, reg2offset_in(src.first()))); 1032 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1033 } else { 1034 // stack to reg 1035 __ movl2ptr(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first()))); 1036 } 1037 } else if (dst.first()->is_stack()) { 1038 // reg to stack 1039 // no need to sign extend on 64bit 1040 __ movptr(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register()); 1041 } else { 1042 if (dst.first() != src.first()) { 1043 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 1044 } 1045 } 1046 } 1047 1048 // An oop arg. Must pass a handle not the oop itself 1049 static void object_move(MacroAssembler* masm, 1050 OopMap* map, 1051 int oop_handle_offset, 1052 int framesize_in_slots, 1053 VMRegPair src, 1054 VMRegPair dst, 1055 bool is_receiver, 1056 int* receiver_offset) { 1057 1058 // Because of the calling conventions we know that src can be a 1059 // register or a stack location. dst can only be a stack location. 1060 1061 assert(dst.first()->is_stack(), "must be stack"); 1062 // must pass a handle. First figure out the location we use as a handle 1063 1064 if (src.first()->is_stack()) { 1065 // Oop is already on the stack as an argument 1066 Register rHandle = rax; 1067 Label nil; 1068 __ xorptr(rHandle, rHandle); 1069 __ cmpptr(Address(rbp, reg2offset_in(src.first())), (int32_t)NULL_WORD); 1070 __ jcc(Assembler::equal, nil); 1071 __ lea(rHandle, Address(rbp, reg2offset_in(src.first()))); 1072 __ bind(nil); 1073 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle); 1074 1075 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1076 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 1077 if (is_receiver) { 1078 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 1079 } 1080 } else { 1081 // Oop is in an a register we must store it to the space we reserve 1082 // on the stack for oop_handles 1083 const Register rOop = src.first()->as_Register(); 1084 const Register rHandle = rax; 1085 int oop_slot = (rOop == rcx ? 0 : 1) * VMRegImpl::slots_per_word + oop_handle_offset; 1086 int offset = oop_slot*VMRegImpl::stack_slot_size; 1087 Label skip; 1088 __ movptr(Address(rsp, offset), rOop); 1089 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1090 __ xorptr(rHandle, rHandle); 1091 __ cmpptr(rOop, (int32_t)NULL_WORD); 1092 __ jcc(Assembler::equal, skip); 1093 __ lea(rHandle, Address(rsp, offset)); 1094 __ bind(skip); 1095 // Store the handle parameter 1096 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle); 1097 if (is_receiver) { 1098 *receiver_offset = offset; 1099 } 1100 } 1101 } 1102 1103 // A float arg may have to do float reg int reg conversion 1104 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1105 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move"); 1106 1107 // Because of the calling convention we know that src is either a stack location 1108 // or an xmm register. dst can only be a stack location. 1109 1110 assert(dst.first()->is_stack() && ( src.first()->is_stack() || src.first()->is_XMMRegister()), "bad parameters"); 1111 1112 if (src.first()->is_stack()) { 1113 __ movl(rax, Address(rbp, reg2offset_in(src.first()))); 1114 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1115 } else { 1116 // reg to stack 1117 __ movflt(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister()); 1118 } 1119 } 1120 1121 // A long move 1122 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1123 1124 // The only legal possibility for a long_move VMRegPair is: 1125 // 1: two stack slots (possibly unaligned) 1126 // as neither the java or C calling convention will use registers 1127 // for longs. 1128 1129 if (src.first()->is_stack() && dst.first()->is_stack()) { 1130 assert(src.second()->is_stack() && dst.second()->is_stack(), "must be all stack"); 1131 __ movptr(rax, Address(rbp, reg2offset_in(src.first()))); 1132 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second())))); 1133 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1134 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx)); 1135 } else { 1136 ShouldNotReachHere(); 1137 } 1138 } 1139 1140 // A double move 1141 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1142 1143 // The only legal possibilities for a double_move VMRegPair are: 1144 // The painful thing here is that like long_move a VMRegPair might be 1145 1146 // Because of the calling convention we know that src is either 1147 // 1: a single physical register (xmm registers only) 1148 // 2: two stack slots (possibly unaligned) 1149 // dst can only be a pair of stack slots. 1150 1151 assert(dst.first()->is_stack() && (src.first()->is_XMMRegister() || src.first()->is_stack()), "bad args"); 1152 1153 if (src.first()->is_stack()) { 1154 // source is all stack 1155 __ movptr(rax, Address(rbp, reg2offset_in(src.first()))); 1156 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second())))); 1157 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1158 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx)); 1159 } else { 1160 // reg to stack 1161 // No worries about stack alignment 1162 __ movdbl(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister()); 1163 } 1164 } 1165 1166 1167 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1168 // We always ignore the frame_slots arg and just use the space just below frame pointer 1169 // which by this time is free to use 1170 switch (ret_type) { 1171 case T_FLOAT: 1172 __ fstp_s(Address(rbp, -wordSize)); 1173 break; 1174 case T_DOUBLE: 1175 __ fstp_d(Address(rbp, -2*wordSize)); 1176 break; 1177 case T_VOID: break; 1178 case T_LONG: 1179 __ movptr(Address(rbp, -wordSize), rax); 1180 NOT_LP64(__ movptr(Address(rbp, -2*wordSize), rdx)); 1181 break; 1182 default: { 1183 __ movptr(Address(rbp, -wordSize), rax); 1184 } 1185 } 1186 } 1187 1188 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1189 // We always ignore the frame_slots arg and just use the space just below frame pointer 1190 // which by this time is free to use 1191 switch (ret_type) { 1192 case T_FLOAT: 1193 __ fld_s(Address(rbp, -wordSize)); 1194 break; 1195 case T_DOUBLE: 1196 __ fld_d(Address(rbp, -2*wordSize)); 1197 break; 1198 case T_LONG: 1199 __ movptr(rax, Address(rbp, -wordSize)); 1200 NOT_LP64(__ movptr(rdx, Address(rbp, -2*wordSize))); 1201 break; 1202 case T_VOID: break; 1203 default: { 1204 __ movptr(rax, Address(rbp, -wordSize)); 1205 } 1206 } 1207 } 1208 1209 1210 static void save_or_restore_arguments(MacroAssembler* masm, 1211 const int stack_slots, 1212 const int total_in_args, 1213 const int arg_save_area, 1214 OopMap* map, 1215 VMRegPair* in_regs, 1216 BasicType* in_sig_bt) { 1217 // if map is non-NULL then the code should store the values, 1218 // otherwise it should load them. 1219 int handle_index = 0; 1220 // Save down double word first 1221 for ( int i = 0; i < total_in_args; i++) { 1222 if (in_regs[i].first()->is_XMMRegister() && in_sig_bt[i] == T_DOUBLE) { 1223 int slot = handle_index * VMRegImpl::slots_per_word + arg_save_area; 1224 int offset = slot * VMRegImpl::stack_slot_size; 1225 handle_index += 2; 1226 assert(handle_index <= stack_slots, "overflow"); 1227 if (map != NULL) { 1228 __ movdbl(Address(rsp, offset), in_regs[i].first()->as_XMMRegister()); 1229 } else { 1230 __ movdbl(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset)); 1231 } 1232 } 1233 if (in_regs[i].first()->is_Register() && in_sig_bt[i] == T_LONG) { 1234 int slot = handle_index * VMRegImpl::slots_per_word + arg_save_area; 1235 int offset = slot * VMRegImpl::stack_slot_size; 1236 handle_index += 2; 1237 assert(handle_index <= stack_slots, "overflow"); 1238 if (map != NULL) { 1239 __ movl(Address(rsp, offset), in_regs[i].first()->as_Register()); 1240 if (in_regs[i].second()->is_Register()) { 1241 __ movl(Address(rsp, offset + 4), in_regs[i].second()->as_Register()); 1242 } 1243 } else { 1244 __ movl(in_regs[i].first()->as_Register(), Address(rsp, offset)); 1245 if (in_regs[i].second()->is_Register()) { 1246 __ movl(in_regs[i].second()->as_Register(), Address(rsp, offset + 4)); 1247 } 1248 } 1249 } 1250 } 1251 // Save or restore single word registers 1252 for ( int i = 0; i < total_in_args; i++) { 1253 if (in_regs[i].first()->is_Register()) { 1254 int slot = handle_index++ * VMRegImpl::slots_per_word + arg_save_area; 1255 int offset = slot * VMRegImpl::stack_slot_size; 1256 assert(handle_index <= stack_slots, "overflow"); 1257 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1258 map->set_oop(VMRegImpl::stack2reg(slot));; 1259 } 1260 1261 // Value is in an input register pass we must flush it to the stack 1262 const Register reg = in_regs[i].first()->as_Register(); 1263 switch (in_sig_bt[i]) { 1264 case T_ARRAY: 1265 if (map != NULL) { 1266 __ movptr(Address(rsp, offset), reg); 1267 } else { 1268 __ movptr(reg, Address(rsp, offset)); 1269 } 1270 break; 1271 case T_BOOLEAN: 1272 case T_CHAR: 1273 case T_BYTE: 1274 case T_SHORT: 1275 case T_INT: 1276 if (map != NULL) { 1277 __ movl(Address(rsp, offset), reg); 1278 } else { 1279 __ movl(reg, Address(rsp, offset)); 1280 } 1281 break; 1282 case T_OBJECT: 1283 default: ShouldNotReachHere(); 1284 } 1285 } else if (in_regs[i].first()->is_XMMRegister()) { 1286 if (in_sig_bt[i] == T_FLOAT) { 1287 int slot = handle_index++ * VMRegImpl::slots_per_word + arg_save_area; 1288 int offset = slot * VMRegImpl::stack_slot_size; 1289 assert(handle_index <= stack_slots, "overflow"); 1290 if (map != NULL) { 1291 __ movflt(Address(rsp, offset), in_regs[i].first()->as_XMMRegister()); 1292 } else { 1293 __ movflt(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset)); 1294 } 1295 } 1296 } else if (in_regs[i].first()->is_stack()) { 1297 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1298 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1299 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots)); 1300 } 1301 } 1302 } 1303 } 1304 1305 // Check GCLocker::needs_gc and enter the runtime if it's true. This 1306 // keeps a new JNI critical region from starting until a GC has been 1307 // forced. Save down any oops in registers and describe them in an 1308 // OopMap. 1309 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1310 Register thread, 1311 int stack_slots, 1312 int total_c_args, 1313 int total_in_args, 1314 int arg_save_area, 1315 OopMapSet* oop_maps, 1316 VMRegPair* in_regs, 1317 BasicType* in_sig_bt) { 1318 __ block_comment("check GCLocker::needs_gc"); 1319 Label cont; 1320 __ cmp8(ExternalAddress((address)GCLocker::needs_gc_address()), false); 1321 __ jcc(Assembler::equal, cont); 1322 1323 // Save down any incoming oops and call into the runtime to halt for a GC 1324 1325 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1326 1327 save_or_restore_arguments(masm, stack_slots, total_in_args, 1328 arg_save_area, map, in_regs, in_sig_bt); 1329 1330 address the_pc = __ pc(); 1331 oop_maps->add_gc_map( __ offset(), map); 1332 __ set_last_Java_frame(thread, rsp, noreg, the_pc); 1333 1334 __ block_comment("block_for_jni_critical"); 1335 __ push(thread); 1336 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical))); 1337 __ increment(rsp, wordSize); 1338 1339 __ get_thread(thread); 1340 __ reset_last_Java_frame(thread, false); 1341 1342 save_or_restore_arguments(masm, stack_slots, total_in_args, 1343 arg_save_area, NULL, in_regs, in_sig_bt); 1344 1345 __ bind(cont); 1346 #ifdef ASSERT 1347 if (StressCriticalJNINatives) { 1348 // Stress register saving 1349 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1350 save_or_restore_arguments(masm, stack_slots, total_in_args, 1351 arg_save_area, map, in_regs, in_sig_bt); 1352 // Destroy argument registers 1353 for (int i = 0; i < total_in_args - 1; i++) { 1354 if (in_regs[i].first()->is_Register()) { 1355 const Register reg = in_regs[i].first()->as_Register(); 1356 __ xorptr(reg, reg); 1357 } else if (in_regs[i].first()->is_XMMRegister()) { 1358 __ xorpd(in_regs[i].first()->as_XMMRegister(), in_regs[i].first()->as_XMMRegister()); 1359 } else if (in_regs[i].first()->is_FloatRegister()) { 1360 ShouldNotReachHere(); 1361 } else if (in_regs[i].first()->is_stack()) { 1362 // Nothing to do 1363 } else { 1364 ShouldNotReachHere(); 1365 } 1366 if (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_DOUBLE) { 1367 i++; 1368 } 1369 } 1370 1371 save_or_restore_arguments(masm, stack_slots, total_in_args, 1372 arg_save_area, NULL, in_regs, in_sig_bt); 1373 } 1374 #endif 1375 } 1376 1377 // Unpack an array argument into a pointer to the body and the length 1378 // if the array is non-null, otherwise pass 0 for both. 1379 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) { 1380 Register tmp_reg = rax; 1381 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg, 1382 "possible collision"); 1383 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg, 1384 "possible collision"); 1385 1386 // Pass the length, ptr pair 1387 Label is_null, done; 1388 VMRegPair tmp(tmp_reg->as_VMReg()); 1389 if (reg.first()->is_stack()) { 1390 // Load the arg up from the stack 1391 simple_move32(masm, reg, tmp); 1392 reg = tmp; 1393 } 1394 __ testptr(reg.first()->as_Register(), reg.first()->as_Register()); 1395 __ jccb(Assembler::equal, is_null); 1396 __ lea(tmp_reg, Address(reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type))); 1397 simple_move32(masm, tmp, body_arg); 1398 // load the length relative to the body. 1399 __ movl(tmp_reg, Address(tmp_reg, arrayOopDesc::length_offset_in_bytes() - 1400 arrayOopDesc::base_offset_in_bytes(in_elem_type))); 1401 simple_move32(masm, tmp, length_arg); 1402 __ jmpb(done); 1403 __ bind(is_null); 1404 // Pass zeros 1405 __ xorptr(tmp_reg, tmp_reg); 1406 simple_move32(masm, tmp, body_arg); 1407 simple_move32(masm, tmp, length_arg); 1408 __ bind(done); 1409 } 1410 1411 static void verify_oop_args(MacroAssembler* masm, 1412 const methodHandle& method, 1413 const BasicType* sig_bt, 1414 const VMRegPair* regs) { 1415 Register temp_reg = rbx; // not part of any compiled calling seq 1416 if (VerifyOops) { 1417 for (int i = 0; i < method->size_of_parameters(); i++) { 1418 if (sig_bt[i] == T_OBJECT || 1419 sig_bt[i] == T_ARRAY) { 1420 VMReg r = regs[i].first(); 1421 assert(r->is_valid(), "bad oop arg"); 1422 if (r->is_stack()) { 1423 __ movptr(temp_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1424 __ verify_oop(temp_reg); 1425 } else { 1426 __ verify_oop(r->as_Register()); 1427 } 1428 } 1429 } 1430 } 1431 } 1432 1433 static void gen_special_dispatch(MacroAssembler* masm, 1434 const methodHandle& method, 1435 const BasicType* sig_bt, 1436 const VMRegPair* regs) { 1437 verify_oop_args(masm, method, sig_bt, regs); 1438 vmIntrinsics::ID iid = method->intrinsic_id(); 1439 1440 // Now write the args into the outgoing interpreter space 1441 bool has_receiver = false; 1442 Register receiver_reg = noreg; 1443 int member_arg_pos = -1; 1444 Register member_reg = noreg; 1445 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1446 if (ref_kind != 0) { 1447 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1448 member_reg = rbx; // known to be free at this point 1449 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1450 } else if (iid == vmIntrinsics::_invokeBasic) { 1451 has_receiver = true; 1452 } else { 1453 fatal("unexpected intrinsic id %d", iid); 1454 } 1455 1456 if (member_reg != noreg) { 1457 // Load the member_arg into register, if necessary. 1458 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1459 VMReg r = regs[member_arg_pos].first(); 1460 if (r->is_stack()) { 1461 __ movptr(member_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1462 } else { 1463 // no data motion is needed 1464 member_reg = r->as_Register(); 1465 } 1466 } 1467 1468 if (has_receiver) { 1469 // Make sure the receiver is loaded into a register. 1470 assert(method->size_of_parameters() > 0, "oob"); 1471 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1472 VMReg r = regs[0].first(); 1473 assert(r->is_valid(), "bad receiver arg"); 1474 if (r->is_stack()) { 1475 // Porting note: This assumes that compiled calling conventions always 1476 // pass the receiver oop in a register. If this is not true on some 1477 // platform, pick a temp and load the receiver from stack. 1478 fatal("receiver always in a register"); 1479 receiver_reg = rcx; // known to be free at this point 1480 __ movptr(receiver_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1481 } else { 1482 // no data motion is needed 1483 receiver_reg = r->as_Register(); 1484 } 1485 } 1486 1487 // Figure out which address we are really jumping to: 1488 MethodHandles::generate_method_handle_dispatch(masm, iid, 1489 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1490 } 1491 1492 // --------------------------------------------------------------------------- 1493 // Generate a native wrapper for a given method. The method takes arguments 1494 // in the Java compiled code convention, marshals them to the native 1495 // convention (handlizes oops, etc), transitions to native, makes the call, 1496 // returns to java state (possibly blocking), unhandlizes any result and 1497 // returns. 1498 // 1499 // Critical native functions are a shorthand for the use of 1500 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1501 // functions. The wrapper is expected to unpack the arguments before 1502 // passing them to the callee and perform checks before and after the 1503 // native call to ensure that they GCLocker 1504 // lock_critical/unlock_critical semantics are followed. Some other 1505 // parts of JNI setup are skipped like the tear down of the JNI handle 1506 // block and the check for pending exceptions it's impossible for them 1507 // to be thrown. 1508 // 1509 // They are roughly structured like this: 1510 // if (GCLocker::needs_gc()) 1511 // SharedRuntime::block_for_jni_critical(); 1512 // tranistion to thread_in_native 1513 // unpack arrray arguments and call native entry point 1514 // check for safepoint in progress 1515 // check if any thread suspend flags are set 1516 // call into JVM and possible unlock the JNI critical 1517 // if a GC was suppressed while in the critical native. 1518 // transition back to thread_in_Java 1519 // return to caller 1520 // 1521 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1522 const methodHandle& method, 1523 int compile_id, 1524 BasicType* in_sig_bt, 1525 VMRegPair* in_regs, 1526 BasicType ret_type) { 1527 if (method->is_method_handle_intrinsic()) { 1528 vmIntrinsics::ID iid = method->intrinsic_id(); 1529 intptr_t start = (intptr_t)__ pc(); 1530 int vep_offset = ((intptr_t)__ pc()) - start; 1531 gen_special_dispatch(masm, 1532 method, 1533 in_sig_bt, 1534 in_regs); 1535 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1536 __ flush(); 1537 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1538 return nmethod::new_native_nmethod(method, 1539 compile_id, 1540 masm->code(), 1541 vep_offset, 1542 frame_complete, 1543 stack_slots / VMRegImpl::slots_per_word, 1544 in_ByteSize(-1), 1545 in_ByteSize(-1), 1546 (OopMapSet*)NULL); 1547 } 1548 bool is_critical_native = true; 1549 address native_func = method->critical_native_function(); 1550 if (native_func == NULL) { 1551 native_func = method->native_function(); 1552 is_critical_native = false; 1553 } 1554 assert(native_func != NULL, "must have function"); 1555 1556 // An OopMap for lock (and class if static) 1557 OopMapSet *oop_maps = new OopMapSet(); 1558 1559 // We have received a description of where all the java arg are located 1560 // on entry to the wrapper. We need to convert these args to where 1561 // the jni function will expect them. To figure out where they go 1562 // we convert the java signature to a C signature by inserting 1563 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1564 1565 const int total_in_args = method->size_of_parameters(); 1566 int total_c_args = total_in_args; 1567 if (!is_critical_native) { 1568 total_c_args += 1; 1569 if (method->is_static()) { 1570 total_c_args++; 1571 } 1572 } else { 1573 for (int i = 0; i < total_in_args; i++) { 1574 if (in_sig_bt[i] == T_ARRAY) { 1575 total_c_args++; 1576 } 1577 } 1578 } 1579 1580 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1581 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1582 BasicType* in_elem_bt = NULL; 1583 1584 int argc = 0; 1585 if (!is_critical_native) { 1586 out_sig_bt[argc++] = T_ADDRESS; 1587 if (method->is_static()) { 1588 out_sig_bt[argc++] = T_OBJECT; 1589 } 1590 1591 for (int i = 0; i < total_in_args ; i++ ) { 1592 out_sig_bt[argc++] = in_sig_bt[i]; 1593 } 1594 } else { 1595 Thread* THREAD = Thread::current(); 1596 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args); 1597 SignatureStream ss(method->signature()); 1598 for (int i = 0; i < total_in_args ; i++ ) { 1599 if (in_sig_bt[i] == T_ARRAY) { 1600 // Arrays are passed as int, elem* pair 1601 out_sig_bt[argc++] = T_INT; 1602 out_sig_bt[argc++] = T_ADDRESS; 1603 Symbol* atype = ss.as_symbol(CHECK_NULL); 1604 const char* at = atype->as_C_string(); 1605 if (strlen(at) == 2) { 1606 assert(at[0] == '[', "must be"); 1607 switch (at[1]) { 1608 case 'B': in_elem_bt[i] = T_BYTE; break; 1609 case 'C': in_elem_bt[i] = T_CHAR; break; 1610 case 'D': in_elem_bt[i] = T_DOUBLE; break; 1611 case 'F': in_elem_bt[i] = T_FLOAT; break; 1612 case 'I': in_elem_bt[i] = T_INT; break; 1613 case 'J': in_elem_bt[i] = T_LONG; break; 1614 case 'S': in_elem_bt[i] = T_SHORT; break; 1615 case 'Z': in_elem_bt[i] = T_BOOLEAN; break; 1616 default: ShouldNotReachHere(); 1617 } 1618 } 1619 } else { 1620 out_sig_bt[argc++] = in_sig_bt[i]; 1621 in_elem_bt[i] = T_VOID; 1622 } 1623 if (in_sig_bt[i] != T_VOID) { 1624 assert(in_sig_bt[i] == ss.type(), "must match"); 1625 ss.next(); 1626 } 1627 } 1628 } 1629 1630 // Now figure out where the args must be stored and how much stack space 1631 // they require. 1632 int out_arg_slots; 1633 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args); 1634 1635 // Compute framesize for the wrapper. We need to handlize all oops in 1636 // registers a max of 2 on x86. 1637 1638 // Calculate the total number of stack slots we will need. 1639 1640 // First count the abi requirement plus all of the outgoing args 1641 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 1642 1643 // Now the space for the inbound oop handle area 1644 int total_save_slots = 2 * VMRegImpl::slots_per_word; // 2 arguments passed in registers 1645 if (is_critical_native) { 1646 // Critical natives may have to call out so they need a save area 1647 // for register arguments. 1648 int double_slots = 0; 1649 int single_slots = 0; 1650 for ( int i = 0; i < total_in_args; i++) { 1651 if (in_regs[i].first()->is_Register()) { 1652 const Register reg = in_regs[i].first()->as_Register(); 1653 switch (in_sig_bt[i]) { 1654 case T_ARRAY: // critical array (uses 2 slots on LP64) 1655 case T_BOOLEAN: 1656 case T_BYTE: 1657 case T_SHORT: 1658 case T_CHAR: 1659 case T_INT: single_slots++; break; 1660 case T_LONG: double_slots++; break; 1661 default: ShouldNotReachHere(); 1662 } 1663 } else if (in_regs[i].first()->is_XMMRegister()) { 1664 switch (in_sig_bt[i]) { 1665 case T_FLOAT: single_slots++; break; 1666 case T_DOUBLE: double_slots++; break; 1667 default: ShouldNotReachHere(); 1668 } 1669 } else if (in_regs[i].first()->is_FloatRegister()) { 1670 ShouldNotReachHere(); 1671 } 1672 } 1673 total_save_slots = double_slots * 2 + single_slots; 1674 // align the save area 1675 if (double_slots != 0) { 1676 stack_slots = align_up(stack_slots, 2); 1677 } 1678 } 1679 1680 int oop_handle_offset = stack_slots; 1681 stack_slots += total_save_slots; 1682 1683 // Now any space we need for handlizing a klass if static method 1684 1685 int klass_slot_offset = 0; 1686 int klass_offset = -1; 1687 int lock_slot_offset = 0; 1688 bool is_static = false; 1689 1690 if (method->is_static()) { 1691 klass_slot_offset = stack_slots; 1692 stack_slots += VMRegImpl::slots_per_word; 1693 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1694 is_static = true; 1695 } 1696 1697 // Plus a lock if needed 1698 1699 if (method->is_synchronized()) { 1700 lock_slot_offset = stack_slots; 1701 stack_slots += VMRegImpl::slots_per_word; 1702 } 1703 1704 // Now a place (+2) to save return values or temp during shuffling 1705 // + 2 for return address (which we own) and saved rbp, 1706 stack_slots += 4; 1707 1708 // Ok The space we have allocated will look like: 1709 // 1710 // 1711 // FP-> | | 1712 // |---------------------| 1713 // | 2 slots for moves | 1714 // |---------------------| 1715 // | lock box (if sync) | 1716 // |---------------------| <- lock_slot_offset (-lock_slot_rbp_offset) 1717 // | klass (if static) | 1718 // |---------------------| <- klass_slot_offset 1719 // | oopHandle area | 1720 // |---------------------| <- oop_handle_offset (a max of 2 registers) 1721 // | outbound memory | 1722 // | based arguments | 1723 // | | 1724 // |---------------------| 1725 // | | 1726 // SP-> | out_preserved_slots | 1727 // 1728 // 1729 // **************************************************************************** 1730 // WARNING - on Windows Java Natives use pascal calling convention and pop the 1731 // arguments off of the stack after the jni call. Before the call we can use 1732 // instructions that are SP relative. After the jni call we switch to FP 1733 // relative instructions instead of re-adjusting the stack on windows. 1734 // **************************************************************************** 1735 1736 1737 // Now compute actual number of stack words we need rounding to make 1738 // stack properly aligned. 1739 stack_slots = align_up(stack_slots, StackAlignmentInSlots); 1740 1741 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 1742 1743 intptr_t start = (intptr_t)__ pc(); 1744 1745 // First thing make an ic check to see if we should even be here 1746 1747 // We are free to use all registers as temps without saving them and 1748 // restoring them except rbp. rbp is the only callee save register 1749 // as far as the interpreter and the compiler(s) are concerned. 1750 1751 1752 const Register ic_reg = rax; 1753 const Register receiver = rcx; 1754 Label hit; 1755 Label exception_pending; 1756 1757 __ verify_oop(receiver); 1758 __ cmpptr(ic_reg, Address(receiver, oopDesc::klass_offset_in_bytes())); 1759 __ jcc(Assembler::equal, hit); 1760 1761 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1762 1763 // verified entry must be aligned for code patching. 1764 // and the first 5 bytes must be in the same cache line 1765 // if we align at 8 then we will be sure 5 bytes are in the same line 1766 __ align(8); 1767 1768 __ bind(hit); 1769 1770 int vep_offset = ((intptr_t)__ pc()) - start; 1771 1772 #ifdef COMPILER1 1773 // For Object.hashCode, System.identityHashCode try to pull hashCode from object header if available. 1774 if ((InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) || (method->intrinsic_id() == vmIntrinsics::_identityHashCode)) { 1775 inline_check_hashcode_from_object_header(masm, method, rcx /*obj_reg*/, rax /*result*/); 1776 } 1777 #endif // COMPILER1 1778 1779 // The instruction at the verified entry point must be 5 bytes or longer 1780 // because it can be patched on the fly by make_non_entrant. The stack bang 1781 // instruction fits that requirement. 1782 1783 // Generate stack overflow check 1784 1785 if (UseStackBanging) { 1786 __ bang_stack_with_offset((int)JavaThread::stack_shadow_zone_size()); 1787 } else { 1788 // need a 5 byte instruction to allow MT safe patching to non-entrant 1789 __ fat_nop(); 1790 } 1791 1792 // Generate a new frame for the wrapper. 1793 __ enter(); 1794 // -2 because return address is already present and so is saved rbp 1795 __ subptr(rsp, stack_size - 2*wordSize); 1796 1797 // Frame is now completed as far as size and linkage. 1798 int frame_complete = ((intptr_t)__ pc()) - start; 1799 1800 if (UseRTMLocking) { 1801 // Abort RTM transaction before calling JNI 1802 // because critical section will be large and will be 1803 // aborted anyway. Also nmethod could be deoptimized. 1804 __ xabort(0); 1805 } 1806 1807 // Calculate the difference between rsp and rbp,. We need to know it 1808 // after the native call because on windows Java Natives will pop 1809 // the arguments and it is painful to do rsp relative addressing 1810 // in a platform independent way. So after the call we switch to 1811 // rbp, relative addressing. 1812 1813 int fp_adjustment = stack_size - 2*wordSize; 1814 1815 #ifdef COMPILER2 1816 // C2 may leave the stack dirty if not in SSE2+ mode 1817 if (UseSSE >= 2) { 1818 __ verify_FPU(0, "c2i transition should have clean FPU stack"); 1819 } else { 1820 __ empty_FPU_stack(); 1821 } 1822 #endif /* COMPILER2 */ 1823 1824 // Compute the rbp, offset for any slots used after the jni call 1825 1826 int lock_slot_rbp_offset = (lock_slot_offset*VMRegImpl::stack_slot_size) - fp_adjustment; 1827 1828 // We use rdi as a thread pointer because it is callee save and 1829 // if we load it once it is usable thru the entire wrapper 1830 const Register thread = rdi; 1831 1832 // We use rsi as the oop handle for the receiver/klass 1833 // It is callee save so it survives the call to native 1834 1835 const Register oop_handle_reg = rsi; 1836 1837 __ get_thread(thread); 1838 1839 if (is_critical_native) { 1840 check_needs_gc_for_critical_native(masm, thread, stack_slots, total_c_args, total_in_args, 1841 oop_handle_offset, oop_maps, in_regs, in_sig_bt); 1842 } 1843 1844 // 1845 // We immediately shuffle the arguments so that any vm call we have to 1846 // make from here on out (sync slow path, jvmti, etc.) we will have 1847 // captured the oops from our caller and have a valid oopMap for 1848 // them. 1849 1850 // ----------------- 1851 // The Grand Shuffle 1852 // 1853 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv* 1854 // and, if static, the class mirror instead of a receiver. This pretty much 1855 // guarantees that register layout will not match (and x86 doesn't use reg 1856 // parms though amd does). Since the native abi doesn't use register args 1857 // and the java conventions does we don't have to worry about collisions. 1858 // All of our moved are reg->stack or stack->stack. 1859 // We ignore the extra arguments during the shuffle and handle them at the 1860 // last moment. The shuffle is described by the two calling convention 1861 // vectors we have in our possession. We simply walk the java vector to 1862 // get the source locations and the c vector to get the destinations. 1863 1864 int c_arg = is_critical_native ? 0 : (method->is_static() ? 2 : 1 ); 1865 1866 // Record rsp-based slot for receiver on stack for non-static methods 1867 int receiver_offset = -1; 1868 1869 // This is a trick. We double the stack slots so we can claim 1870 // the oops in the caller's frame. Since we are sure to have 1871 // more args than the caller doubling is enough to make 1872 // sure we can capture all the incoming oop args from the 1873 // caller. 1874 // 1875 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1876 1877 // Mark location of rbp, 1878 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, rbp->as_VMReg()); 1879 1880 // We know that we only have args in at most two integer registers (rcx, rdx). So rax, rbx 1881 // Are free to temporaries if we have to do stack to steck moves. 1882 // All inbound args are referenced based on rbp, and all outbound args via rsp. 1883 1884 for (int i = 0; i < total_in_args ; i++, c_arg++ ) { 1885 switch (in_sig_bt[i]) { 1886 case T_ARRAY: 1887 if (is_critical_native) { 1888 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]); 1889 c_arg++; 1890 break; 1891 } 1892 case T_OBJECT: 1893 assert(!is_critical_native, "no oop arguments"); 1894 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 1895 ((i == 0) && (!is_static)), 1896 &receiver_offset); 1897 break; 1898 case T_VOID: 1899 break; 1900 1901 case T_FLOAT: 1902 float_move(masm, in_regs[i], out_regs[c_arg]); 1903 break; 1904 1905 case T_DOUBLE: 1906 assert( i + 1 < total_in_args && 1907 in_sig_bt[i + 1] == T_VOID && 1908 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 1909 double_move(masm, in_regs[i], out_regs[c_arg]); 1910 break; 1911 1912 case T_LONG : 1913 long_move(masm, in_regs[i], out_regs[c_arg]); 1914 break; 1915 1916 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 1917 1918 default: 1919 simple_move32(masm, in_regs[i], out_regs[c_arg]); 1920 } 1921 } 1922 1923 // Pre-load a static method's oop into rsi. Used both by locking code and 1924 // the normal JNI call code. 1925 if (method->is_static() && !is_critical_native) { 1926 1927 // load opp into a register 1928 __ movoop(oop_handle_reg, JNIHandles::make_local(method->method_holder()->java_mirror())); 1929 1930 // Now handlize the static class mirror it's known not-null. 1931 __ movptr(Address(rsp, klass_offset), oop_handle_reg); 1932 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 1933 1934 // Now get the handle 1935 __ lea(oop_handle_reg, Address(rsp, klass_offset)); 1936 // store the klass handle as second argument 1937 __ movptr(Address(rsp, wordSize), oop_handle_reg); 1938 } 1939 1940 // Change state to native (we save the return address in the thread, since it might not 1941 // be pushed on the stack when we do a a stack traversal). It is enough that the pc() 1942 // points into the right code segment. It does not have to be the correct return pc. 1943 // We use the same pc/oopMap repeatedly when we call out 1944 1945 intptr_t the_pc = (intptr_t) __ pc(); 1946 oop_maps->add_gc_map(the_pc - start, map); 1947 1948 __ set_last_Java_frame(thread, rsp, noreg, (address)the_pc); 1949 1950 1951 // We have all of the arguments setup at this point. We must not touch any register 1952 // argument registers at this point (what if we save/restore them there are no oop? 1953 1954 { 1955 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0); 1956 __ mov_metadata(rax, method()); 1957 __ call_VM_leaf( 1958 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 1959 thread, rax); 1960 } 1961 1962 // RedefineClasses() tracing support for obsolete method entry 1963 if (log_is_enabled(Trace, redefine, class, obsolete)) { 1964 __ mov_metadata(rax, method()); 1965 __ call_VM_leaf( 1966 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1967 thread, rax); 1968 } 1969 1970 // These are register definitions we need for locking/unlocking 1971 const Register swap_reg = rax; // Must use rax, for cmpxchg instruction 1972 const Register obj_reg = rcx; // Will contain the oop 1973 const Register lock_reg = rdx; // Address of compiler lock object (BasicLock) 1974 1975 Label slow_path_lock; 1976 Label lock_done; 1977 1978 // Lock a synchronized method 1979 if (method->is_synchronized()) { 1980 assert(!is_critical_native, "unhandled"); 1981 1982 1983 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 1984 1985 // Get the handle (the 2nd argument) 1986 __ movptr(oop_handle_reg, Address(rsp, wordSize)); 1987 1988 // Get address of the box 1989 1990 __ lea(lock_reg, Address(rbp, lock_slot_rbp_offset)); 1991 1992 // Load the oop from the handle 1993 __ movptr(obj_reg, Address(oop_handle_reg, 0)); 1994 1995 if (UseBiasedLocking) { 1996 // Note that oop_handle_reg is trashed during this call 1997 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, oop_handle_reg, false, lock_done, &slow_path_lock); 1998 } 1999 2000 // Load immediate 1 into swap_reg %rax, 2001 __ movptr(swap_reg, 1); 2002 2003 // Load (object->mark() | 1) into swap_reg %rax, 2004 __ orptr(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 2005 2006 // Save (object->mark() | 1) into BasicLock's displaced header 2007 __ movptr(Address(lock_reg, mark_word_offset), swap_reg); 2008 2009 if (os::is_MP()) { 2010 __ lock(); 2011 } 2012 2013 // src -> dest iff dest == rax, else rax, <- dest 2014 // *obj_reg = lock_reg iff *obj_reg == rax, else rax, = *(obj_reg) 2015 __ cmpxchgptr(lock_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 2016 __ jcc(Assembler::equal, lock_done); 2017 2018 // Test if the oopMark is an obvious stack pointer, i.e., 2019 // 1) (mark & 3) == 0, and 2020 // 2) rsp <= mark < mark + os::pagesize() 2021 // These 3 tests can be done by evaluating the following 2022 // expression: ((mark - rsp) & (3 - os::vm_page_size())), 2023 // assuming both stack pointer and pagesize have their 2024 // least significant 2 bits clear. 2025 // NOTE: the oopMark is in swap_reg %rax, as the result of cmpxchg 2026 2027 __ subptr(swap_reg, rsp); 2028 __ andptr(swap_reg, 3 - os::vm_page_size()); 2029 2030 // Save the test result, for recursive case, the result is zero 2031 __ movptr(Address(lock_reg, mark_word_offset), swap_reg); 2032 __ jcc(Assembler::notEqual, slow_path_lock); 2033 // Slow path will re-enter here 2034 __ bind(lock_done); 2035 2036 if (UseBiasedLocking) { 2037 // Re-fetch oop_handle_reg as we trashed it above 2038 __ movptr(oop_handle_reg, Address(rsp, wordSize)); 2039 } 2040 } 2041 2042 2043 // Finally just about ready to make the JNI call 2044 2045 2046 // get JNIEnv* which is first argument to native 2047 if (!is_critical_native) { 2048 __ lea(rdx, Address(thread, in_bytes(JavaThread::jni_environment_offset()))); 2049 __ movptr(Address(rsp, 0), rdx); 2050 } 2051 2052 // Now set thread in native 2053 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native); 2054 2055 __ call(RuntimeAddress(native_func)); 2056 2057 // Verify or restore cpu control state after JNI call 2058 __ restore_cpu_control_state_after_jni(); 2059 2060 // WARNING - on Windows Java Natives use pascal calling convention and pop the 2061 // arguments off of the stack. We could just re-adjust the stack pointer here 2062 // and continue to do SP relative addressing but we instead switch to FP 2063 // relative addressing. 2064 2065 // Unpack native results. 2066 switch (ret_type) { 2067 case T_BOOLEAN: __ c2bool(rax); break; 2068 case T_CHAR : __ andptr(rax, 0xFFFF); break; 2069 case T_BYTE : __ sign_extend_byte (rax); break; 2070 case T_SHORT : __ sign_extend_short(rax); break; 2071 case T_INT : /* nothing to do */ break; 2072 case T_DOUBLE : 2073 case T_FLOAT : 2074 // Result is in st0 we'll save as needed 2075 break; 2076 case T_ARRAY: // Really a handle 2077 case T_OBJECT: // Really a handle 2078 break; // can't de-handlize until after safepoint check 2079 case T_VOID: break; 2080 case T_LONG: break; 2081 default : ShouldNotReachHere(); 2082 } 2083 2084 // Switch thread to "native transition" state before reading the synchronization state. 2085 // This additional state is necessary because reading and testing the synchronization 2086 // state is not atomic w.r.t. GC, as this scenario demonstrates: 2087 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 2088 // VM thread changes sync state to synchronizing and suspends threads for GC. 2089 // Thread A is resumed to finish this native method, but doesn't block here since it 2090 // didn't see any synchronization is progress, and escapes. 2091 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native_trans); 2092 2093 if(os::is_MP()) { 2094 if (UseMembar) { 2095 // Force this write out before the read below 2096 __ membar(Assembler::Membar_mask_bits( 2097 Assembler::LoadLoad | Assembler::LoadStore | 2098 Assembler::StoreLoad | Assembler::StoreStore)); 2099 } else { 2100 // Write serialization page so VM thread can do a pseudo remote membar. 2101 // We use the current thread pointer to calculate a thread specific 2102 // offset to write to within the page. This minimizes bus traffic 2103 // due to cache line collision. 2104 __ serialize_memory(thread, rcx); 2105 } 2106 } 2107 2108 if (AlwaysRestoreFPU) { 2109 // Make sure the control word is correct. 2110 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std())); 2111 } 2112 2113 Label after_transition; 2114 2115 // check for safepoint operation in progress and/or pending suspend requests 2116 { Label Continue, slow_path; 2117 2118 __ safepoint_poll(slow_path, thread, noreg); 2119 2120 __ cmpl(Address(thread, JavaThread::suspend_flags_offset()), 0); 2121 __ jcc(Assembler::equal, Continue); 2122 __ bind(slow_path); 2123 2124 // Don't use call_VM as it will see a possible pending exception and forward it 2125 // and never return here preventing us from clearing _last_native_pc down below. 2126 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are 2127 // preserved and correspond to the bcp/locals pointers. So we do a runtime call 2128 // by hand. 2129 // 2130 __ vzeroupper(); 2131 2132 save_native_result(masm, ret_type, stack_slots); 2133 __ push(thread); 2134 if (!is_critical_native) { 2135 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, 2136 JavaThread::check_special_condition_for_native_trans))); 2137 } else { 2138 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, 2139 JavaThread::check_special_condition_for_native_trans_and_transition))); 2140 } 2141 __ increment(rsp, wordSize); 2142 // Restore any method result value 2143 restore_native_result(masm, ret_type, stack_slots); 2144 2145 if (is_critical_native) { 2146 // The call above performed the transition to thread_in_Java so 2147 // skip the transition logic below. 2148 __ jmpb(after_transition); 2149 } 2150 2151 __ bind(Continue); 2152 } 2153 2154 // change thread state 2155 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_Java); 2156 __ bind(after_transition); 2157 2158 Label reguard; 2159 Label reguard_done; 2160 __ cmpl(Address(thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_reserved_disabled); 2161 __ jcc(Assembler::equal, reguard); 2162 2163 // slow path reguard re-enters here 2164 __ bind(reguard_done); 2165 2166 // Handle possible exception (will unlock if necessary) 2167 2168 // native result if any is live 2169 2170 // Unlock 2171 Label slow_path_unlock; 2172 Label unlock_done; 2173 if (method->is_synchronized()) { 2174 2175 Label done; 2176 2177 // Get locked oop from the handle we passed to jni 2178 __ movptr(obj_reg, Address(oop_handle_reg, 0)); 2179 2180 if (UseBiasedLocking) { 2181 __ biased_locking_exit(obj_reg, rbx, done); 2182 } 2183 2184 // Simple recursive lock? 2185 2186 __ cmpptr(Address(rbp, lock_slot_rbp_offset), (int32_t)NULL_WORD); 2187 __ jcc(Assembler::equal, done); 2188 2189 // Must save rax, if if it is live now because cmpxchg must use it 2190 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2191 save_native_result(masm, ret_type, stack_slots); 2192 } 2193 2194 // get old displaced header 2195 __ movptr(rbx, Address(rbp, lock_slot_rbp_offset)); 2196 2197 // get address of the stack lock 2198 __ lea(rax, Address(rbp, lock_slot_rbp_offset)); 2199 2200 // Atomic swap old header if oop still contains the stack lock 2201 if (os::is_MP()) { 2202 __ lock(); 2203 } 2204 2205 // src -> dest iff dest == rax, else rax, <- dest 2206 // *obj_reg = rbx, iff *obj_reg == rax, else rax, = *(obj_reg) 2207 __ cmpxchgptr(rbx, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 2208 __ jcc(Assembler::notEqual, slow_path_unlock); 2209 2210 // slow path re-enters here 2211 __ bind(unlock_done); 2212 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2213 restore_native_result(masm, ret_type, stack_slots); 2214 } 2215 2216 __ bind(done); 2217 2218 } 2219 2220 { 2221 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0); 2222 // Tell dtrace about this method exit 2223 save_native_result(masm, ret_type, stack_slots); 2224 __ mov_metadata(rax, method()); 2225 __ call_VM_leaf( 2226 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2227 thread, rax); 2228 restore_native_result(masm, ret_type, stack_slots); 2229 } 2230 2231 // We can finally stop using that last_Java_frame we setup ages ago 2232 2233 __ reset_last_Java_frame(thread, false); 2234 2235 // Unbox oop result, e.g. JNIHandles::resolve value. 2236 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2237 __ resolve_jobject(rax /* value */, 2238 thread /* thread */, 2239 rcx /* tmp */); 2240 } 2241 2242 if (CheckJNICalls) { 2243 // clear_pending_jni_exception_check 2244 __ movptr(Address(thread, JavaThread::pending_jni_exception_check_fn_offset()), NULL_WORD); 2245 } 2246 2247 if (!is_critical_native) { 2248 // reset handle block 2249 __ movptr(rcx, Address(thread, JavaThread::active_handles_offset())); 2250 __ movl(Address(rcx, JNIHandleBlock::top_offset_in_bytes()), NULL_WORD); 2251 2252 // Any exception pending? 2253 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2254 __ jcc(Assembler::notEqual, exception_pending); 2255 } 2256 2257 // no exception, we're almost done 2258 2259 // check that only result value is on FPU stack 2260 __ verify_FPU(ret_type == T_FLOAT || ret_type == T_DOUBLE ? 1 : 0, "native_wrapper normal exit"); 2261 2262 // Fixup floating pointer results so that result looks like a return from a compiled method 2263 if (ret_type == T_FLOAT) { 2264 if (UseSSE >= 1) { 2265 // Pop st0 and store as float and reload into xmm register 2266 __ fstp_s(Address(rbp, -4)); 2267 __ movflt(xmm0, Address(rbp, -4)); 2268 } 2269 } else if (ret_type == T_DOUBLE) { 2270 if (UseSSE >= 2) { 2271 // Pop st0 and store as double and reload into xmm register 2272 __ fstp_d(Address(rbp, -8)); 2273 __ movdbl(xmm0, Address(rbp, -8)); 2274 } 2275 } 2276 2277 // Return 2278 2279 __ leave(); 2280 __ ret(0); 2281 2282 // Unexpected paths are out of line and go here 2283 2284 // Slow path locking & unlocking 2285 if (method->is_synchronized()) { 2286 2287 // BEGIN Slow path lock 2288 2289 __ bind(slow_path_lock); 2290 2291 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 2292 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2293 __ push(thread); 2294 __ push(lock_reg); 2295 __ push(obj_reg); 2296 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C))); 2297 __ addptr(rsp, 3*wordSize); 2298 2299 #ifdef ASSERT 2300 { Label L; 2301 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int)NULL_WORD); 2302 __ jcc(Assembler::equal, L); 2303 __ stop("no pending exception allowed on exit from monitorenter"); 2304 __ bind(L); 2305 } 2306 #endif 2307 __ jmp(lock_done); 2308 2309 // END Slow path lock 2310 2311 // BEGIN Slow path unlock 2312 __ bind(slow_path_unlock); 2313 __ vzeroupper(); 2314 // Slow path unlock 2315 2316 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2317 save_native_result(masm, ret_type, stack_slots); 2318 } 2319 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 2320 2321 __ pushptr(Address(thread, in_bytes(Thread::pending_exception_offset()))); 2322 __ movptr(Address(thread, in_bytes(Thread::pending_exception_offset())), NULL_WORD); 2323 2324 2325 // should be a peal 2326 // +wordSize because of the push above 2327 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2328 __ push(thread); 2329 __ lea(rax, Address(rbp, lock_slot_rbp_offset)); 2330 __ push(rax); 2331 2332 __ push(obj_reg); 2333 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C))); 2334 __ addptr(rsp, 3*wordSize); 2335 #ifdef ASSERT 2336 { 2337 Label L; 2338 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2339 __ jcc(Assembler::equal, L); 2340 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 2341 __ bind(L); 2342 } 2343 #endif /* ASSERT */ 2344 2345 __ popptr(Address(thread, in_bytes(Thread::pending_exception_offset()))); 2346 2347 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2348 restore_native_result(masm, ret_type, stack_slots); 2349 } 2350 __ jmp(unlock_done); 2351 // END Slow path unlock 2352 2353 } 2354 2355 // SLOW PATH Reguard the stack if needed 2356 2357 __ bind(reguard); 2358 __ vzeroupper(); 2359 save_native_result(masm, ret_type, stack_slots); 2360 { 2361 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages))); 2362 } 2363 restore_native_result(masm, ret_type, stack_slots); 2364 __ jmp(reguard_done); 2365 2366 2367 // BEGIN EXCEPTION PROCESSING 2368 2369 if (!is_critical_native) { 2370 // Forward the exception 2371 __ bind(exception_pending); 2372 2373 // remove possible return value from FPU register stack 2374 __ empty_FPU_stack(); 2375 2376 // pop our frame 2377 __ leave(); 2378 // and forward the exception 2379 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2380 } 2381 2382 __ flush(); 2383 2384 nmethod *nm = nmethod::new_native_nmethod(method, 2385 compile_id, 2386 masm->code(), 2387 vep_offset, 2388 frame_complete, 2389 stack_slots / VMRegImpl::slots_per_word, 2390 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2391 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2392 oop_maps); 2393 2394 if (is_critical_native) { 2395 nm->set_lazy_critical_native(true); 2396 } 2397 2398 return nm; 2399 2400 } 2401 2402 // this function returns the adjust size (in number of words) to a c2i adapter 2403 // activation for use during deoptimization 2404 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals ) { 2405 return (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2406 } 2407 2408 2409 uint SharedRuntime::out_preserve_stack_slots() { 2410 return 0; 2411 } 2412 2413 //------------------------------generate_deopt_blob---------------------------- 2414 void SharedRuntime::generate_deopt_blob() { 2415 // allocate space for the code 2416 ResourceMark rm; 2417 // setup code generation tools 2418 // note: the buffer code size must account for StackShadowPages=50 2419 CodeBuffer buffer("deopt_blob", 1536, 1024); 2420 MacroAssembler* masm = new MacroAssembler(&buffer); 2421 int frame_size_in_words; 2422 OopMap* map = NULL; 2423 // Account for the extra args we place on the stack 2424 // by the time we call fetch_unroll_info 2425 const int additional_words = 2; // deopt kind, thread 2426 2427 OopMapSet *oop_maps = new OopMapSet(); 2428 2429 // ------------- 2430 // This code enters when returning to a de-optimized nmethod. A return 2431 // address has been pushed on the the stack, and return values are in 2432 // registers. 2433 // If we are doing a normal deopt then we were called from the patched 2434 // nmethod from the point we returned to the nmethod. So the return 2435 // address on the stack is wrong by NativeCall::instruction_size 2436 // We will adjust the value to it looks like we have the original return 2437 // address on the stack (like when we eagerly deoptimized). 2438 // In the case of an exception pending with deoptimized then we enter 2439 // with a return address on the stack that points after the call we patched 2440 // into the exception handler. We have the following register state: 2441 // rax,: exception 2442 // rbx,: exception handler 2443 // rdx: throwing pc 2444 // So in this case we simply jam rdx into the useless return address and 2445 // the stack looks just like we want. 2446 // 2447 // At this point we need to de-opt. We save the argument return 2448 // registers. We call the first C routine, fetch_unroll_info(). This 2449 // routine captures the return values and returns a structure which 2450 // describes the current frame size and the sizes of all replacement frames. 2451 // The current frame is compiled code and may contain many inlined 2452 // functions, each with their own JVM state. We pop the current frame, then 2453 // push all the new frames. Then we call the C routine unpack_frames() to 2454 // populate these frames. Finally unpack_frames() returns us the new target 2455 // address. Notice that callee-save registers are BLOWN here; they have 2456 // already been captured in the vframeArray at the time the return PC was 2457 // patched. 2458 address start = __ pc(); 2459 Label cont; 2460 2461 // Prolog for non exception case! 2462 2463 // Save everything in sight. 2464 2465 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false); 2466 // Normal deoptimization 2467 __ push(Deoptimization::Unpack_deopt); 2468 __ jmp(cont); 2469 2470 int reexecute_offset = __ pc() - start; 2471 2472 // Reexecute case 2473 // return address is the pc describes what bci to do re-execute at 2474 2475 // No need to update map as each call to save_live_registers will produce identical oopmap 2476 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false); 2477 2478 __ push(Deoptimization::Unpack_reexecute); 2479 __ jmp(cont); 2480 2481 int exception_offset = __ pc() - start; 2482 2483 // Prolog for exception case 2484 2485 // all registers are dead at this entry point, except for rax, and 2486 // rdx which contain the exception oop and exception pc 2487 // respectively. Set them in TLS and fall thru to the 2488 // unpack_with_exception_in_tls entry point. 2489 2490 __ get_thread(rdi); 2491 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), rdx); 2492 __ movptr(Address(rdi, JavaThread::exception_oop_offset()), rax); 2493 2494 int exception_in_tls_offset = __ pc() - start; 2495 2496 // new implementation because exception oop is now passed in JavaThread 2497 2498 // Prolog for exception case 2499 // All registers must be preserved because they might be used by LinearScan 2500 // Exceptiop oop and throwing PC are passed in JavaThread 2501 // tos: stack at point of call to method that threw the exception (i.e. only 2502 // args are on the stack, no return address) 2503 2504 // make room on stack for the return address 2505 // It will be patched later with the throwing pc. The correct value is not 2506 // available now because loading it from memory would destroy registers. 2507 __ push(0); 2508 2509 // Save everything in sight. 2510 2511 // No need to update map as each call to save_live_registers will produce identical oopmap 2512 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false); 2513 2514 // Now it is safe to overwrite any register 2515 2516 // store the correct deoptimization type 2517 __ push(Deoptimization::Unpack_exception); 2518 2519 // load throwing pc from JavaThread and patch it as the return address 2520 // of the current frame. Then clear the field in JavaThread 2521 __ get_thread(rdi); 2522 __ movptr(rdx, Address(rdi, JavaThread::exception_pc_offset())); 2523 __ movptr(Address(rbp, wordSize), rdx); 2524 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), NULL_WORD); 2525 2526 #ifdef ASSERT 2527 // verify that there is really an exception oop in JavaThread 2528 __ movptr(rax, Address(rdi, JavaThread::exception_oop_offset())); 2529 __ verify_oop(rax); 2530 2531 // verify that there is no pending exception 2532 Label no_pending_exception; 2533 __ movptr(rax, Address(rdi, Thread::pending_exception_offset())); 2534 __ testptr(rax, rax); 2535 __ jcc(Assembler::zero, no_pending_exception); 2536 __ stop("must not have pending exception here"); 2537 __ bind(no_pending_exception); 2538 #endif 2539 2540 __ bind(cont); 2541 2542 // Compiled code leaves the floating point stack dirty, empty it. 2543 __ empty_FPU_stack(); 2544 2545 2546 // Call C code. Need thread and this frame, but NOT official VM entry 2547 // crud. We cannot block on this call, no GC can happen. 2548 __ get_thread(rcx); 2549 __ push(rcx); 2550 // fetch_unroll_info needs to call last_java_frame() 2551 __ set_last_Java_frame(rcx, noreg, noreg, NULL); 2552 2553 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info))); 2554 2555 // Need to have an oopmap that tells fetch_unroll_info where to 2556 // find any register it might need. 2557 2558 oop_maps->add_gc_map( __ pc()-start, map); 2559 2560 // Discard args to fetch_unroll_info 2561 __ pop(rcx); 2562 __ pop(rcx); 2563 2564 __ get_thread(rcx); 2565 __ reset_last_Java_frame(rcx, false); 2566 2567 // Load UnrollBlock into EDI 2568 __ mov(rdi, rax); 2569 2570 // Move the unpack kind to a safe place in the UnrollBlock because 2571 // we are very short of registers 2572 2573 Address unpack_kind(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()); 2574 // retrieve the deopt kind from the UnrollBlock. 2575 __ movl(rax, unpack_kind); 2576 2577 Label noException; 2578 __ cmpl(rax, Deoptimization::Unpack_exception); // Was exception pending? 2579 __ jcc(Assembler::notEqual, noException); 2580 __ movptr(rax, Address(rcx, JavaThread::exception_oop_offset())); 2581 __ movptr(rdx, Address(rcx, JavaThread::exception_pc_offset())); 2582 __ movptr(Address(rcx, JavaThread::exception_oop_offset()), NULL_WORD); 2583 __ movptr(Address(rcx, JavaThread::exception_pc_offset()), NULL_WORD); 2584 2585 __ verify_oop(rax); 2586 2587 // Overwrite the result registers with the exception results. 2588 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax); 2589 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx); 2590 2591 __ bind(noException); 2592 2593 // Stack is back to only having register save data on the stack. 2594 // Now restore the result registers. Everything else is either dead or captured 2595 // in the vframeArray. 2596 2597 RegisterSaver::restore_result_registers(masm); 2598 2599 // Non standard control word may be leaked out through a safepoint blob, and we can 2600 // deopt at a poll point with the non standard control word. However, we should make 2601 // sure the control word is correct after restore_result_registers. 2602 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std())); 2603 2604 // All of the register save area has been popped of the stack. Only the 2605 // return address remains. 2606 2607 // Pop all the frames we must move/replace. 2608 // 2609 // Frame picture (youngest to oldest) 2610 // 1: self-frame (no frame link) 2611 // 2: deopting frame (no frame link) 2612 // 3: caller of deopting frame (could be compiled/interpreted). 2613 // 2614 // Note: by leaving the return address of self-frame on the stack 2615 // and using the size of frame 2 to adjust the stack 2616 // when we are done the return to frame 3 will still be on the stack. 2617 2618 // Pop deoptimized frame 2619 __ addptr(rsp, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 2620 2621 // sp should be pointing at the return address to the caller (3) 2622 2623 // Pick up the initial fp we should save 2624 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved) 2625 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 2626 2627 #ifdef ASSERT 2628 // Compilers generate code that bang the stack by as much as the 2629 // interpreter would need. So this stack banging should never 2630 // trigger a fault. Verify that it does not on non product builds. 2631 if (UseStackBanging) { 2632 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 2633 __ bang_stack_size(rbx, rcx); 2634 } 2635 #endif 2636 2637 // Load array of frame pcs into ECX 2638 __ movptr(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2639 2640 __ pop(rsi); // trash the old pc 2641 2642 // Load array of frame sizes into ESI 2643 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 2644 2645 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes()); 2646 2647 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 2648 __ movl(counter, rbx); 2649 2650 // Now adjust the caller's stack to make up for the extra locals 2651 // but record the original sp so that we can save it in the skeletal interpreter 2652 // frame and the stack walking of interpreter_sender will get the unextended sp 2653 // value and not the "real" sp value. 2654 2655 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes()); 2656 __ movptr(sp_temp, rsp); 2657 __ movl2ptr(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes())); 2658 __ subptr(rsp, rbx); 2659 2660 // Push interpreter frames in a loop 2661 Label loop; 2662 __ bind(loop); 2663 __ movptr(rbx, Address(rsi, 0)); // Load frame size 2664 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand 2665 __ pushptr(Address(rcx, 0)); // save return address 2666 __ enter(); // save old & set new rbp, 2667 __ subptr(rsp, rbx); // Prolog! 2668 __ movptr(rbx, sp_temp); // sender's sp 2669 // This value is corrected by layout_activation_impl 2670 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD); 2671 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable 2672 __ movptr(sp_temp, rsp); // pass to next frame 2673 __ addptr(rsi, wordSize); // Bump array pointer (sizes) 2674 __ addptr(rcx, wordSize); // Bump array pointer (pcs) 2675 __ decrementl(counter); // decrement counter 2676 __ jcc(Assembler::notZero, loop); 2677 __ pushptr(Address(rcx, 0)); // save final return address 2678 2679 // Re-push self-frame 2680 __ enter(); // save old & set new rbp, 2681 2682 // Return address and rbp, are in place 2683 // We'll push additional args later. Just allocate a full sized 2684 // register save area 2685 __ subptr(rsp, (frame_size_in_words-additional_words - 2) * wordSize); 2686 2687 // Restore frame locals after moving the frame 2688 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax); 2689 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx); 2690 __ fstp_d(Address(rsp, RegisterSaver::fpResultOffset()*wordSize)); // Pop float stack and store in local 2691 if( UseSSE>=2 ) __ movdbl(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0); 2692 if( UseSSE==1 ) __ movflt(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0); 2693 2694 // Set up the args to unpack_frame 2695 2696 __ pushl(unpack_kind); // get the unpack_kind value 2697 __ get_thread(rcx); 2698 __ push(rcx); 2699 2700 // set last_Java_sp, last_Java_fp 2701 __ set_last_Java_frame(rcx, noreg, rbp, NULL); 2702 2703 // Call C code. Need thread but NOT official VM entry 2704 // crud. We cannot block on this call, no GC can happen. Call should 2705 // restore return values to their stack-slots with the new SP. 2706 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2707 // Set an oopmap for the call site 2708 oop_maps->add_gc_map( __ pc()-start, new OopMap( frame_size_in_words, 0 )); 2709 2710 // rax, contains the return result type 2711 __ push(rax); 2712 2713 __ get_thread(rcx); 2714 __ reset_last_Java_frame(rcx, false); 2715 2716 // Collect return values 2717 __ movptr(rax,Address(rsp, (RegisterSaver::raxOffset() + additional_words + 1)*wordSize)); 2718 __ movptr(rdx,Address(rsp, (RegisterSaver::rdxOffset() + additional_words + 1)*wordSize)); 2719 2720 // Clear floating point stack before returning to interpreter 2721 __ empty_FPU_stack(); 2722 2723 // Check if we should push the float or double return value. 2724 Label results_done, yes_double_value; 2725 __ cmpl(Address(rsp, 0), T_DOUBLE); 2726 __ jcc (Assembler::zero, yes_double_value); 2727 __ cmpl(Address(rsp, 0), T_FLOAT); 2728 __ jcc (Assembler::notZero, results_done); 2729 2730 // return float value as expected by interpreter 2731 if( UseSSE>=1 ) __ movflt(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize)); 2732 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize)); 2733 __ jmp(results_done); 2734 2735 // return double value as expected by interpreter 2736 __ bind(yes_double_value); 2737 if( UseSSE>=2 ) __ movdbl(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize)); 2738 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize)); 2739 2740 __ bind(results_done); 2741 2742 // Pop self-frame. 2743 __ leave(); // Epilog! 2744 2745 // Jump to interpreter 2746 __ ret(0); 2747 2748 // ------------- 2749 // make sure all code is generated 2750 masm->flush(); 2751 2752 _deopt_blob = DeoptimizationBlob::create( &buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 2753 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 2754 } 2755 2756 2757 #ifdef COMPILER2 2758 //------------------------------generate_uncommon_trap_blob-------------------- 2759 void SharedRuntime::generate_uncommon_trap_blob() { 2760 // allocate space for the code 2761 ResourceMark rm; 2762 // setup code generation tools 2763 CodeBuffer buffer("uncommon_trap_blob", 512, 512); 2764 MacroAssembler* masm = new MacroAssembler(&buffer); 2765 2766 enum frame_layout { 2767 arg0_off, // thread sp + 0 // Arg location for 2768 arg1_off, // unloaded_class_index sp + 1 // calling C 2769 arg2_off, // exec_mode sp + 2 2770 // The frame sender code expects that rbp will be in the "natural" place and 2771 // will override any oopMap setting for it. We must therefore force the layout 2772 // so that it agrees with the frame sender code. 2773 rbp_off, // callee saved register sp + 3 2774 return_off, // slot for return address sp + 4 2775 framesize 2776 }; 2777 2778 address start = __ pc(); 2779 2780 if (UseRTMLocking) { 2781 // Abort RTM transaction before possible nmethod deoptimization. 2782 __ xabort(0); 2783 } 2784 2785 // Push self-frame. 2786 __ subptr(rsp, return_off*wordSize); // Epilog! 2787 2788 // rbp, is an implicitly saved callee saved register (i.e. the calling 2789 // convention will save restore it in prolog/epilog) Other than that 2790 // there are no callee save registers no that adapter frames are gone. 2791 __ movptr(Address(rsp, rbp_off*wordSize), rbp); 2792 2793 // Clear the floating point exception stack 2794 __ empty_FPU_stack(); 2795 2796 // set last_Java_sp 2797 __ get_thread(rdx); 2798 __ set_last_Java_frame(rdx, noreg, noreg, NULL); 2799 2800 // Call C code. Need thread but NOT official VM entry 2801 // crud. We cannot block on this call, no GC can happen. Call should 2802 // capture callee-saved registers as well as return values. 2803 __ movptr(Address(rsp, arg0_off*wordSize), rdx); 2804 // argument already in ECX 2805 __ movl(Address(rsp, arg1_off*wordSize),rcx); 2806 __ movl(Address(rsp, arg2_off*wordSize), Deoptimization::Unpack_uncommon_trap); 2807 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap))); 2808 2809 // Set an oopmap for the call site 2810 OopMapSet *oop_maps = new OopMapSet(); 2811 OopMap* map = new OopMap( framesize, 0 ); 2812 // No oopMap for rbp, it is known implicitly 2813 2814 oop_maps->add_gc_map( __ pc()-start, map); 2815 2816 __ get_thread(rcx); 2817 2818 __ reset_last_Java_frame(rcx, false); 2819 2820 // Load UnrollBlock into EDI 2821 __ movptr(rdi, rax); 2822 2823 #ifdef ASSERT 2824 { Label L; 2825 __ cmpptr(Address(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()), 2826 (int32_t)Deoptimization::Unpack_uncommon_trap); 2827 __ jcc(Assembler::equal, L); 2828 __ stop("SharedRuntime::generate_deopt_blob: expected Unpack_uncommon_trap"); 2829 __ bind(L); 2830 } 2831 #endif 2832 2833 // Pop all the frames we must move/replace. 2834 // 2835 // Frame picture (youngest to oldest) 2836 // 1: self-frame (no frame link) 2837 // 2: deopting frame (no frame link) 2838 // 3: caller of deopting frame (could be compiled/interpreted). 2839 2840 // Pop self-frame. We have no frame, and must rely only on EAX and ESP. 2841 __ addptr(rsp,(framesize-1)*wordSize); // Epilog! 2842 2843 // Pop deoptimized frame 2844 __ movl2ptr(rcx, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 2845 __ addptr(rsp, rcx); 2846 2847 // sp should be pointing at the return address to the caller (3) 2848 2849 // Pick up the initial fp we should save 2850 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved) 2851 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 2852 2853 #ifdef ASSERT 2854 // Compilers generate code that bang the stack by as much as the 2855 // interpreter would need. So this stack banging should never 2856 // trigger a fault. Verify that it does not on non product builds. 2857 if (UseStackBanging) { 2858 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 2859 __ bang_stack_size(rbx, rcx); 2860 } 2861 #endif 2862 2863 // Load array of frame pcs into ECX 2864 __ movl(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2865 2866 __ pop(rsi); // trash the pc 2867 2868 // Load array of frame sizes into ESI 2869 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 2870 2871 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes()); 2872 2873 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 2874 __ movl(counter, rbx); 2875 2876 // Now adjust the caller's stack to make up for the extra locals 2877 // but record the original sp so that we can save it in the skeletal interpreter 2878 // frame and the stack walking of interpreter_sender will get the unextended sp 2879 // value and not the "real" sp value. 2880 2881 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes()); 2882 __ movptr(sp_temp, rsp); 2883 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes())); 2884 __ subptr(rsp, rbx); 2885 2886 // Push interpreter frames in a loop 2887 Label loop; 2888 __ bind(loop); 2889 __ movptr(rbx, Address(rsi, 0)); // Load frame size 2890 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand 2891 __ pushptr(Address(rcx, 0)); // save return address 2892 __ enter(); // save old & set new rbp, 2893 __ subptr(rsp, rbx); // Prolog! 2894 __ movptr(rbx, sp_temp); // sender's sp 2895 // This value is corrected by layout_activation_impl 2896 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD ); 2897 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable 2898 __ movptr(sp_temp, rsp); // pass to next frame 2899 __ addptr(rsi, wordSize); // Bump array pointer (sizes) 2900 __ addptr(rcx, wordSize); // Bump array pointer (pcs) 2901 __ decrementl(counter); // decrement counter 2902 __ jcc(Assembler::notZero, loop); 2903 __ pushptr(Address(rcx, 0)); // save final return address 2904 2905 // Re-push self-frame 2906 __ enter(); // save old & set new rbp, 2907 __ subptr(rsp, (framesize-2) * wordSize); // Prolog! 2908 2909 2910 // set last_Java_sp, last_Java_fp 2911 __ get_thread(rdi); 2912 __ set_last_Java_frame(rdi, noreg, rbp, NULL); 2913 2914 // Call C code. Need thread but NOT official VM entry 2915 // crud. We cannot block on this call, no GC can happen. Call should 2916 // restore return values to their stack-slots with the new SP. 2917 __ movptr(Address(rsp,arg0_off*wordSize),rdi); 2918 __ movl(Address(rsp,arg1_off*wordSize), Deoptimization::Unpack_uncommon_trap); 2919 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2920 // Set an oopmap for the call site 2921 oop_maps->add_gc_map( __ pc()-start, new OopMap( framesize, 0 ) ); 2922 2923 __ get_thread(rdi); 2924 __ reset_last_Java_frame(rdi, true); 2925 2926 // Pop self-frame. 2927 __ leave(); // Epilog! 2928 2929 // Jump to interpreter 2930 __ ret(0); 2931 2932 // ------------- 2933 // make sure all code is generated 2934 masm->flush(); 2935 2936 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, framesize); 2937 } 2938 #endif // COMPILER2 2939 2940 //------------------------------generate_handler_blob------ 2941 // 2942 // Generate a special Compile2Runtime blob that saves all registers, 2943 // setup oopmap, and calls safepoint code to stop the compiled code for 2944 // a safepoint. 2945 // 2946 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 2947 2948 // Account for thread arg in our frame 2949 const int additional_words = 1; 2950 int frame_size_in_words; 2951 2952 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 2953 2954 ResourceMark rm; 2955 OopMapSet *oop_maps = new OopMapSet(); 2956 OopMap* map; 2957 2958 // allocate space for the code 2959 // setup code generation tools 2960 CodeBuffer buffer("handler_blob", 1024, 512); 2961 MacroAssembler* masm = new MacroAssembler(&buffer); 2962 2963 const Register java_thread = rdi; // callee-saved for VC++ 2964 address start = __ pc(); 2965 address call_pc = NULL; 2966 bool cause_return = (poll_type == POLL_AT_RETURN); 2967 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP); 2968 2969 if (UseRTMLocking) { 2970 // Abort RTM transaction before calling runtime 2971 // because critical section will be large and will be 2972 // aborted anyway. Also nmethod could be deoptimized. 2973 __ xabort(0); 2974 } 2975 2976 // If cause_return is true we are at a poll_return and there is 2977 // the return address on the stack to the caller on the nmethod 2978 // that is safepoint. We can leave this return on the stack and 2979 // effectively complete the return and safepoint in the caller. 2980 // Otherwise we push space for a return address that the safepoint 2981 // handler will install later to make the stack walking sensible. 2982 if (!cause_return) 2983 __ push(rbx); // Make room for return address (or push it again) 2984 2985 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false, save_vectors); 2986 2987 // The following is basically a call_VM. However, we need the precise 2988 // address of the call in order to generate an oopmap. Hence, we do all the 2989 // work ourselves. 2990 2991 // Push thread argument and setup last_Java_sp 2992 __ get_thread(java_thread); 2993 __ push(java_thread); 2994 __ set_last_Java_frame(java_thread, noreg, noreg, NULL); 2995 2996 // if this was not a poll_return then we need to correct the return address now. 2997 if (!cause_return) { 2998 // Get the return pc saved by the signal handler and stash it in its appropriate place on the stack. 2999 // Additionally, rbx is a callee saved register and we can look at it later to determine 3000 // if someone changed the return address for us! 3001 __ movptr(rbx, Address(java_thread, JavaThread::saved_exception_pc_offset())); 3002 __ movptr(Address(rbp, wordSize), rbx); 3003 } 3004 3005 // do the call 3006 __ call(RuntimeAddress(call_ptr)); 3007 3008 // Set an oopmap for the call site. This oopmap will map all 3009 // oop-registers and debug-info registers as callee-saved. This 3010 // will allow deoptimization at this safepoint to find all possible 3011 // debug-info recordings, as well as let GC find all oops. 3012 3013 oop_maps->add_gc_map( __ pc() - start, map); 3014 3015 // Discard arg 3016 __ pop(rcx); 3017 3018 Label noException; 3019 3020 // Clear last_Java_sp again 3021 __ get_thread(java_thread); 3022 __ reset_last_Java_frame(java_thread, false); 3023 3024 __ cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 3025 __ jcc(Assembler::equal, noException); 3026 3027 // Exception pending 3028 RegisterSaver::restore_live_registers(masm, save_vectors); 3029 3030 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3031 3032 __ bind(noException); 3033 3034 Label no_adjust, bail, not_special; 3035 if (SafepointMechanism::uses_thread_local_poll() && !cause_return) { 3036 // If our stashed return pc was modified by the runtime we avoid touching it 3037 __ cmpptr(rbx, Address(rbp, wordSize)); 3038 __ jccb(Assembler::notEqual, no_adjust); 3039 3040 // Skip over the poll instruction. 3041 // See NativeInstruction::is_safepoint_poll() 3042 // Possible encodings: 3043 // 85 00 test %eax,(%rax) 3044 // 85 01 test %eax,(%rcx) 3045 // 85 02 test %eax,(%rdx) 3046 // 85 03 test %eax,(%rbx) 3047 // 85 06 test %eax,(%rsi) 3048 // 85 07 test %eax,(%rdi) 3049 // 3050 // 85 04 24 test %eax,(%rsp) 3051 // 85 45 00 test %eax,0x0(%rbp) 3052 3053 #ifdef ASSERT 3054 __ movptr(rax, rbx); // remember where 0x85 should be, for verification below 3055 #endif 3056 // rsp/rbp base encoding takes 3 bytes with the following register values: 3057 // rsp 0x04 3058 // rbp 0x05 3059 __ movzbl(rcx, Address(rbx, 1)); 3060 __ andptr(rcx, 0x07); // looking for 0x04 .. 0x05 3061 __ subptr(rcx, 4); // looking for 0x00 .. 0x01 3062 __ cmpptr(rcx, 1); 3063 __ jcc(Assembler::above, not_special); 3064 __ addptr(rbx, 1); 3065 __ bind(not_special); 3066 #ifdef ASSERT 3067 // Verify the correct encoding of the poll we're about to skip. 3068 __ cmpb(Address(rax, 0), NativeTstRegMem::instruction_code_memXregl); 3069 __ jcc(Assembler::notEqual, bail); 3070 // Mask out the modrm bits 3071 __ testb(Address(rax, 1), NativeTstRegMem::modrm_mask); 3072 // rax encodes to 0, so if the bits are nonzero it's incorrect 3073 __ jcc(Assembler::notZero, bail); 3074 #endif 3075 // Adjust return pc forward to step over the safepoint poll instruction 3076 __ addptr(rbx, 2); 3077 __ movptr(Address(rbp, wordSize), rbx); 3078 } 3079 3080 __ bind(no_adjust); 3081 // Normal exit, register restoring and exit 3082 RegisterSaver::restore_live_registers(masm, save_vectors); 3083 3084 __ ret(0); 3085 3086 #ifdef ASSERT 3087 __ bind(bail); 3088 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected"); 3089 #endif 3090 3091 // make sure all code is generated 3092 masm->flush(); 3093 3094 // Fill-out other meta info 3095 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 3096 } 3097 3098 // 3099 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 3100 // 3101 // Generate a stub that calls into vm to find out the proper destination 3102 // of a java call. All the argument registers are live at this point 3103 // but since this is generic code we don't know what they are and the caller 3104 // must do any gc of the args. 3105 // 3106 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3107 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 3108 3109 // allocate space for the code 3110 ResourceMark rm; 3111 3112 CodeBuffer buffer(name, 1000, 512); 3113 MacroAssembler* masm = new MacroAssembler(&buffer); 3114 3115 int frame_size_words; 3116 enum frame_layout { 3117 thread_off, 3118 extra_words }; 3119 3120 OopMapSet *oop_maps = new OopMapSet(); 3121 OopMap* map = NULL; 3122 3123 int start = __ offset(); 3124 3125 map = RegisterSaver::save_live_registers(masm, extra_words, &frame_size_words); 3126 3127 int frame_complete = __ offset(); 3128 3129 const Register thread = rdi; 3130 __ get_thread(rdi); 3131 3132 __ push(thread); 3133 __ set_last_Java_frame(thread, noreg, rbp, NULL); 3134 3135 __ call(RuntimeAddress(destination)); 3136 3137 3138 // Set an oopmap for the call site. 3139 // We need this not only for callee-saved registers, but also for volatile 3140 // registers that the compiler might be keeping live across a safepoint. 3141 3142 oop_maps->add_gc_map( __ offset() - start, map); 3143 3144 // rax, contains the address we are going to jump to assuming no exception got installed 3145 3146 __ addptr(rsp, wordSize); 3147 3148 // clear last_Java_sp 3149 __ reset_last_Java_frame(thread, true); 3150 // check for pending exceptions 3151 Label pending; 3152 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 3153 __ jcc(Assembler::notEqual, pending); 3154 3155 // get the returned Method* 3156 __ get_vm_result_2(rbx, thread); 3157 __ movptr(Address(rsp, RegisterSaver::rbx_offset() * wordSize), rbx); 3158 3159 __ movptr(Address(rsp, RegisterSaver::rax_offset() * wordSize), rax); 3160 3161 RegisterSaver::restore_live_registers(masm); 3162 3163 // We are back the the original state on entry and ready to go. 3164 3165 __ jmp(rax); 3166 3167 // Pending exception after the safepoint 3168 3169 __ bind(pending); 3170 3171 RegisterSaver::restore_live_registers(masm); 3172 3173 // exception pending => remove activation and forward to exception handler 3174 3175 __ get_thread(thread); 3176 __ movptr(Address(thread, JavaThread::vm_result_offset()), NULL_WORD); 3177 __ movptr(rax, Address(thread, Thread::pending_exception_offset())); 3178 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3179 3180 // ------------- 3181 // make sure all code is generated 3182 masm->flush(); 3183 3184 // return the blob 3185 // frame_size_words or bytes?? 3186 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true); 3187 }