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