1 /* 2 * Copyright (c) 2003, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "code/icBuffer.hpp" 30 #include "code/vtableStubs.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "oops/compiledICHolder.hpp" 33 #include "prims/jvmtiRedefineClassesTrace.hpp" 34 #include "runtime/sharedRuntime.hpp" 35 #include "runtime/vframeArray.hpp" 36 #include "vmreg_x86.inline.hpp" 37 #ifdef COMPILER1 38 #include "c1/c1_Runtime1.hpp" 39 #endif 40 #ifdef COMPILER2 41 #include "opto/runtime.hpp" 42 #endif 43 44 #define __ masm-> 45 46 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 47 48 class RegisterSaver { 49 // Capture info about frame layout 50 #define DEF_XMM_OFFS(regnum) xmm ## regnum ## _off = xmm_off + (regnum)*16/BytesPerInt, xmm ## regnum ## H_off 51 enum layout { 52 fpu_state_off = 0, 53 fpu_state_end = fpu_state_off+FPUStateSizeInWords, 54 st0_off, st0H_off, 55 st1_off, st1H_off, 56 st2_off, st2H_off, 57 st3_off, st3H_off, 58 st4_off, st4H_off, 59 st5_off, st5H_off, 60 st6_off, st6H_off, 61 st7_off, st7H_off, 62 xmm_off, 63 DEF_XMM_OFFS(0), 64 DEF_XMM_OFFS(1), 65 DEF_XMM_OFFS(2), 66 DEF_XMM_OFFS(3), 67 DEF_XMM_OFFS(4), 68 DEF_XMM_OFFS(5), 69 DEF_XMM_OFFS(6), 70 DEF_XMM_OFFS(7), 71 flags_off = xmm7_off + 16/BytesPerInt + 1, // 16-byte stack alignment fill word 72 rdi_off, 73 rsi_off, 74 ignore_off, // extra copy of rbp, 75 rsp_off, 76 rbx_off, 77 rdx_off, 78 rcx_off, 79 rax_off, 80 // The frame sender code expects that rbp will be in the "natural" place and 81 // will override any oopMap setting for it. We must therefore force the layout 82 // so that it agrees with the frame sender code. 83 rbp_off, 84 return_off, // slot for return address 85 reg_save_size }; 86 enum { FPU_regs_live = flags_off - fpu_state_end }; 87 88 public: 89 90 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, 91 int* total_frame_words, bool verify_fpu = true, bool save_vectors = false); 92 static void restore_live_registers(MacroAssembler* masm, bool restore_vectors = false); 93 94 static int rax_offset() { return rax_off; } 95 static int rbx_offset() { return rbx_off; } 96 97 // Offsets into the register save area 98 // Used by deoptimization when it is managing result register 99 // values on its own 100 101 static int raxOffset(void) { return rax_off; } 102 static int rdxOffset(void) { return rdx_off; } 103 static int rbxOffset(void) { return rbx_off; } 104 static int xmm0Offset(void) { return xmm0_off; } 105 // This really returns a slot in the fp save area, which one is not important 106 static int fpResultOffset(void) { return st0_off; } 107 108 // During deoptimization only the result register need to be restored 109 // all the other values have already been extracted. 110 111 static void restore_result_registers(MacroAssembler* masm); 112 113 }; 114 115 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, 116 int* total_frame_words, bool verify_fpu, bool save_vectors) { 117 int num_xmm_regs = XMMRegisterImpl::number_of_registers; 118 int ymm_bytes = num_xmm_regs * 16; 119 int zmm_bytes = num_xmm_regs * 32; 120 #ifdef COMPILER2 121 if (save_vectors) { 122 assert(UseAVX > 0, "up to 512bit vectors are supported with EVEX"); 123 assert(MaxVectorSize <= 64, "up to 512bit vectors are supported now"); 124 // Save upper half of YMM registers 125 int vect_bytes = ymm_bytes; 126 if (UseAVX > 2) { 127 // Save upper half of ZMM registers as well 128 vect_bytes += zmm_bytes; 129 } 130 additional_frame_words += vect_bytes / wordSize; 131 } 132 #else 133 assert(!save_vectors, "vectors are generated only by C2"); 134 #endif 135 int frame_size_in_bytes = (reg_save_size + additional_frame_words) * wordSize; 136 int frame_words = frame_size_in_bytes / wordSize; 137 *total_frame_words = frame_words; 138 139 assert(FPUStateSizeInWords == 27, "update stack layout"); 140 141 // save registers, fpu state, and flags 142 // We assume caller has already has return address slot on the stack 143 // We push epb twice in this sequence because we want the real rbp, 144 // to be under the return like a normal enter and we want to use pusha 145 // We push by hand instead of pusing push 146 __ enter(); 147 __ pusha(); 148 __ pushf(); 149 __ subptr(rsp,FPU_regs_live*wordSize); // Push FPU registers space 150 __ push_FPU_state(); // Save FPU state & init 151 152 if (verify_fpu) { 153 // Some stubs may have non standard FPU control word settings so 154 // only check and reset the value when it required to be the 155 // standard value. The safepoint blob in particular can be used 156 // in methods which are using the 24 bit control word for 157 // optimized float math. 158 159 #ifdef ASSERT 160 // Make sure the control word has the expected value 161 Label ok; 162 __ cmpw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std()); 163 __ jccb(Assembler::equal, ok); 164 __ stop("corrupted control word detected"); 165 __ bind(ok); 166 #endif 167 168 // Reset the control word to guard against exceptions being unmasked 169 // since fstp_d can cause FPU stack underflow exceptions. Write it 170 // into the on stack copy and then reload that to make sure that the 171 // current and future values are correct. 172 __ movw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std()); 173 } 174 175 __ frstor(Address(rsp, 0)); 176 if (!verify_fpu) { 177 // Set the control word so that exceptions are masked for the 178 // following code. 179 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std())); 180 } 181 182 int off = st0_off; 183 int delta = st1_off - off; 184 185 // Save the FPU registers in de-opt-able form 186 for (int n = 0; n < FloatRegisterImpl::number_of_registers; n++) { 187 __ fstp_d(Address(rsp, off*wordSize)); 188 off += delta; 189 } 190 191 off = xmm0_off; 192 delta = xmm1_off - off; 193 if(UseSSE == 1) { 194 // Save the XMM state 195 for (int n = 0; n < num_xmm_regs; n++) { 196 __ movflt(Address(rsp, off*wordSize), as_XMMRegister(n)); 197 off += delta; 198 } 199 } else if(UseSSE >= 2) { 200 // Save whole 128bit (16 bytes) XMM registers 201 for (int n = 0; n < num_xmm_regs; n++) { 202 __ movdqu(Address(rsp, off*wordSize), as_XMMRegister(n)); 203 off += delta; 204 } 205 } 206 207 if (save_vectors) { 208 __ subptr(rsp, ymm_bytes); 209 // Save upper half of YMM registers 210 for (int n = 0; n < num_xmm_regs; n++) { 211 __ vextractf128h(Address(rsp, n*16), as_XMMRegister(n)); 212 } 213 if (UseAVX > 2) { 214 __ subptr(rsp, zmm_bytes); 215 // Save upper half of ZMM registers 216 for (int n = 0; n < num_xmm_regs; n++) { 217 __ vextractf64x4h(Address(rsp, n*32), as_XMMRegister(n), 1); 218 } 219 } 220 } 221 222 // Set an oopmap for the call site. This oopmap will map all 223 // oop-registers and debug-info registers as callee-saved. This 224 // will allow deoptimization at this safepoint to find all possible 225 // debug-info recordings, as well as let GC find all oops. 226 227 OopMapSet *oop_maps = new OopMapSet(); 228 OopMap* map = new OopMap( frame_words, 0 ); 229 230 #define STACK_OFFSET(x) VMRegImpl::stack2reg((x) + additional_frame_words) 231 #define NEXTREG(x) (x)->as_VMReg()->next() 232 233 map->set_callee_saved(STACK_OFFSET(rax_off), rax->as_VMReg()); 234 map->set_callee_saved(STACK_OFFSET(rcx_off), rcx->as_VMReg()); 235 map->set_callee_saved(STACK_OFFSET(rdx_off), rdx->as_VMReg()); 236 map->set_callee_saved(STACK_OFFSET(rbx_off), rbx->as_VMReg()); 237 // rbp, location is known implicitly, no oopMap 238 map->set_callee_saved(STACK_OFFSET(rsi_off), rsi->as_VMReg()); 239 map->set_callee_saved(STACK_OFFSET(rdi_off), rdi->as_VMReg()); 240 // %%% This is really a waste but we'll keep things as they were for now for the upper component 241 off = st0_off; 242 delta = st1_off - off; 243 for (int n = 0; n < FloatRegisterImpl::number_of_registers; n++) { 244 FloatRegister freg_name = as_FloatRegister(n); 245 map->set_callee_saved(STACK_OFFSET(off), freg_name->as_VMReg()); 246 map->set_callee_saved(STACK_OFFSET(off+1), NEXTREG(freg_name)); 247 off += delta; 248 } 249 off = xmm0_off; 250 delta = xmm1_off - off; 251 for (int n = 0; n < num_xmm_regs; n++) { 252 XMMRegister xmm_name = as_XMMRegister(n); 253 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg()); 254 map->set_callee_saved(STACK_OFFSET(off+1), NEXTREG(xmm_name)); 255 off += delta; 256 } 257 #undef NEXTREG 258 #undef STACK_OFFSET 259 260 return map; 261 } 262 263 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_vectors) { 264 int num_xmm_regs = XMMRegisterImpl::number_of_registers; 265 int ymm_bytes = num_xmm_regs * 16; 266 int zmm_bytes = num_xmm_regs * 32; 267 // Recover XMM & FPU state 268 int additional_frame_bytes = 0; 269 #ifdef COMPILER2 270 if (restore_vectors) { 271 assert(UseAVX > 0, "up to 512bit vectors are supported with EVEX"); 272 assert(MaxVectorSize <= 64, "up to 512bit vectors are supported now"); 273 // Save upper half of YMM registers 274 additional_frame_bytes = ymm_bytes; 275 if (UseAVX > 2) { 276 // Save upper half of ZMM registers as well 277 additional_frame_bytes += zmm_bytes; 278 } 279 } 280 #else 281 assert(!restore_vectors, "vectors are generated only by C2"); 282 #endif 283 284 int off = xmm0_off; 285 int delta = xmm1_off - off; 286 287 if (UseSSE == 1) { 288 // Restore XMM registers 289 assert(additional_frame_bytes == 0, ""); 290 for (int n = 0; n < num_xmm_regs; n++) { 291 __ movflt(as_XMMRegister(n), Address(rsp, off*wordSize)); 292 off += delta; 293 } 294 } else if (UseSSE >= 2) { 295 // Restore whole 128bit (16 bytes) XMM registers. Do this before restoring YMM and 296 // ZMM because the movdqu instruction zeros the upper part of the XMM register. 297 for (int n = 0; n < num_xmm_regs; n++) { 298 __ movdqu(as_XMMRegister(n), Address(rsp, off*wordSize+additional_frame_bytes)); 299 off += delta; 300 } 301 } 302 303 if (restore_vectors) { 304 if (UseAVX > 2) { 305 // Restore upper half of ZMM registers. 306 for (int n = 0; n < num_xmm_regs; n++) { 307 __ vinsertf64x4h(as_XMMRegister(n), Address(rsp, n*32), 1); 308 } 309 __ addptr(rsp, zmm_bytes); 310 } 311 // Restore upper half of YMM registers. 312 for (int n = 0; n < num_xmm_regs; n++) { 313 __ vinsertf128h(as_XMMRegister(n), Address(rsp, n*16)); 314 } 315 __ addptr(rsp, ymm_bytes); 316 } 317 318 __ pop_FPU_state(); 319 __ addptr(rsp, FPU_regs_live*wordSize); // Pop FPU registers 320 321 __ popf(); 322 __ popa(); 323 // Get the rbp, described implicitly by the frame sender code (no oopMap) 324 __ pop(rbp); 325 } 326 327 void RegisterSaver::restore_result_registers(MacroAssembler* masm) { 328 329 // Just restore result register. Only used by deoptimization. By 330 // now any callee save register that needs to be restore to a c2 331 // caller of the deoptee has been extracted into the vframeArray 332 // and will be stuffed into the c2i adapter we create for later 333 // restoration so only result registers need to be restored here. 334 // 335 336 __ frstor(Address(rsp, 0)); // Restore fpu state 337 338 // Recover XMM & FPU state 339 if( UseSSE == 1 ) { 340 __ movflt(xmm0, Address(rsp, xmm0_off*wordSize)); 341 } else if( UseSSE >= 2 ) { 342 __ movdbl(xmm0, Address(rsp, xmm0_off*wordSize)); 343 } 344 __ movptr(rax, Address(rsp, rax_off*wordSize)); 345 __ movptr(rdx, Address(rsp, rdx_off*wordSize)); 346 // Pop all of the register save are off the stack except the return address 347 __ addptr(rsp, return_off * wordSize); 348 } 349 350 // Is vector's size (in bytes) bigger than a size saved by default? 351 // 16 bytes XMM registers are saved by default using SSE2 movdqu instructions. 352 // Note, MaxVectorSize == 0 with UseSSE < 2 and vectors are not generated. 353 bool SharedRuntime::is_wide_vector(int size) { 354 return size > 16; 355 } 356 357 // The java_calling_convention describes stack locations as ideal slots on 358 // a frame with no abi restrictions. Since we must observe abi restrictions 359 // (like the placement of the register window) the slots must be biased by 360 // the following value. 361 static int reg2offset_in(VMReg r) { 362 // Account for saved rbp, and return address 363 // This should really be in_preserve_stack_slots 364 return (r->reg2stack() + 2) * VMRegImpl::stack_slot_size; 365 } 366 367 static int reg2offset_out(VMReg r) { 368 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 369 } 370 371 // --------------------------------------------------------------------------- 372 // Read the array of BasicTypes from a signature, and compute where the 373 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 374 // quantities. Values less than SharedInfo::stack0 are registers, those above 375 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 376 // as framesizes are fixed. 377 // VMRegImpl::stack0 refers to the first slot 0(sp). 378 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register 379 // up to RegisterImpl::number_of_registers) are the 32-bit 380 // integer registers. 381 382 // Pass first two oop/int args in registers ECX and EDX. 383 // Pass first two float/double args in registers XMM0 and XMM1. 384 // Doubles have precedence, so if you pass a mix of floats and doubles 385 // the doubles will grab the registers before the floats will. 386 387 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 388 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 389 // units regardless of build. Of course for i486 there is no 64 bit build 390 391 392 // --------------------------------------------------------------------------- 393 // The compiled Java calling convention. 394 // Pass first two oop/int args in registers ECX and EDX. 395 // Pass first two float/double args in registers XMM0 and XMM1. 396 // Doubles have precedence, so if you pass a mix of floats and doubles 397 // the doubles will grab the registers before the floats will. 398 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 399 VMRegPair *regs, 400 int total_args_passed, 401 int is_outgoing) { 402 uint stack = 0; // Starting stack position for args on stack 403 404 405 // Pass first two oop/int args in registers ECX and EDX. 406 uint reg_arg0 = 9999; 407 uint reg_arg1 = 9999; 408 409 // Pass first two float/double args in registers XMM0 and XMM1. 410 // Doubles have precedence, so if you pass a mix of floats and doubles 411 // the doubles will grab the registers before the floats will. 412 // CNC - TURNED OFF FOR non-SSE. 413 // On Intel we have to round all doubles (and most floats) at 414 // call sites by storing to the stack in any case. 415 // UseSSE=0 ==> Don't Use ==> 9999+0 416 // UseSSE=1 ==> Floats only ==> 9999+1 417 // UseSSE>=2 ==> Floats or doubles ==> 9999+2 418 enum { fltarg_dontuse = 9999+0, fltarg_float_only = 9999+1, fltarg_flt_dbl = 9999+2 }; 419 uint fargs = (UseSSE>=2) ? 2 : UseSSE; 420 uint freg_arg0 = 9999+fargs; 421 uint freg_arg1 = 9999+fargs; 422 423 // Pass doubles & longs aligned on the stack. First count stack slots for doubles 424 int i; 425 for( i = 0; i < total_args_passed; i++) { 426 if( sig_bt[i] == T_DOUBLE ) { 427 // first 2 doubles go in registers 428 if( freg_arg0 == fltarg_flt_dbl ) freg_arg0 = i; 429 else if( freg_arg1 == fltarg_flt_dbl ) freg_arg1 = i; 430 else // Else double is passed low on the stack to be aligned. 431 stack += 2; 432 } else if( sig_bt[i] == T_LONG ) { 433 stack += 2; 434 } 435 } 436 int dstack = 0; // Separate counter for placing doubles 437 438 // Now pick where all else goes. 439 for( i = 0; i < total_args_passed; i++) { 440 // From the type and the argument number (count) compute the location 441 switch( sig_bt[i] ) { 442 case T_SHORT: 443 case T_CHAR: 444 case T_BYTE: 445 case T_BOOLEAN: 446 case T_INT: 447 case T_ARRAY: 448 case T_OBJECT: 449 case T_ADDRESS: 450 if( reg_arg0 == 9999 ) { 451 reg_arg0 = i; 452 regs[i].set1(rcx->as_VMReg()); 453 } else if( reg_arg1 == 9999 ) { 454 reg_arg1 = i; 455 regs[i].set1(rdx->as_VMReg()); 456 } else { 457 regs[i].set1(VMRegImpl::stack2reg(stack++)); 458 } 459 break; 460 case T_FLOAT: 461 if( freg_arg0 == fltarg_flt_dbl || freg_arg0 == fltarg_float_only ) { 462 freg_arg0 = i; 463 regs[i].set1(xmm0->as_VMReg()); 464 } else if( freg_arg1 == fltarg_flt_dbl || freg_arg1 == fltarg_float_only ) { 465 freg_arg1 = i; 466 regs[i].set1(xmm1->as_VMReg()); 467 } else { 468 regs[i].set1(VMRegImpl::stack2reg(stack++)); 469 } 470 break; 471 case T_LONG: 472 assert(sig_bt[i+1] == T_VOID, "missing Half" ); 473 regs[i].set2(VMRegImpl::stack2reg(dstack)); 474 dstack += 2; 475 break; 476 case T_DOUBLE: 477 assert(sig_bt[i+1] == T_VOID, "missing Half" ); 478 if( freg_arg0 == (uint)i ) { 479 regs[i].set2(xmm0->as_VMReg()); 480 } else if( freg_arg1 == (uint)i ) { 481 regs[i].set2(xmm1->as_VMReg()); 482 } else { 483 regs[i].set2(VMRegImpl::stack2reg(dstack)); 484 dstack += 2; 485 } 486 break; 487 case T_VOID: regs[i].set_bad(); break; 488 break; 489 default: 490 ShouldNotReachHere(); 491 break; 492 } 493 } 494 495 // return value can be odd number of VMRegImpl stack slots make multiple of 2 496 return round_to(stack, 2); 497 } 498 499 // Patch the callers callsite with entry to compiled code if it exists. 500 static void patch_callers_callsite(MacroAssembler *masm) { 501 Label L; 502 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD); 503 __ jcc(Assembler::equal, L); 504 // Schedule the branch target address early. 505 // Call into the VM to patch the caller, then jump to compiled callee 506 // rax, isn't live so capture return address while we easily can 507 __ movptr(rax, Address(rsp, 0)); 508 __ pusha(); 509 __ pushf(); 510 511 if (UseSSE == 1) { 512 __ subptr(rsp, 2*wordSize); 513 __ movflt(Address(rsp, 0), xmm0); 514 __ movflt(Address(rsp, wordSize), xmm1); 515 } 516 if (UseSSE >= 2) { 517 __ subptr(rsp, 4*wordSize); 518 __ movdbl(Address(rsp, 0), xmm0); 519 __ movdbl(Address(rsp, 2*wordSize), xmm1); 520 } 521 #ifdef COMPILER2 522 // C2 may leave the stack dirty if not in SSE2+ mode 523 if (UseSSE >= 2) { 524 __ verify_FPU(0, "c2i transition should have clean FPU stack"); 525 } else { 526 __ empty_FPU_stack(); 527 } 528 #endif /* COMPILER2 */ 529 530 // VM needs caller's callsite 531 __ push(rax); 532 // VM needs target method 533 __ push(rbx); 534 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite))); 535 __ addptr(rsp, 2*wordSize); 536 537 if (UseSSE == 1) { 538 __ movflt(xmm0, Address(rsp, 0)); 539 __ movflt(xmm1, Address(rsp, wordSize)); 540 __ addptr(rsp, 2*wordSize); 541 } 542 if (UseSSE >= 2) { 543 __ movdbl(xmm0, Address(rsp, 0)); 544 __ movdbl(xmm1, Address(rsp, 2*wordSize)); 545 __ addptr(rsp, 4*wordSize); 546 } 547 548 __ popf(); 549 __ popa(); 550 __ bind(L); 551 } 552 553 554 static void move_c2i_double(MacroAssembler *masm, XMMRegister r, int st_off) { 555 int next_off = st_off - Interpreter::stackElementSize; 556 __ movdbl(Address(rsp, next_off), r); 557 } 558 559 static void gen_c2i_adapter(MacroAssembler *masm, 560 const GrowableArray<SigEntry>& sig_extended, 561 const VMRegPair *regs, 562 Label& skip_fixup, 563 address start, 564 OopMapSet*& oop_maps, 565 int& frame_complete, 566 int& frame_size_in_words) { 567 // Before we get into the guts of the C2I adapter, see if we should be here 568 // at all. We've come from compiled code and are attempting to jump to the 569 // interpreter, which means the caller made a static call to get here 570 // (vcalls always get a compiled target if there is one). Check for a 571 // compiled target. If there is one, we need to patch the caller's call. 572 patch_callers_callsite(masm); 573 574 __ bind(skip_fixup); 575 576 #ifdef COMPILER2 577 // C2 may leave the stack dirty if not in SSE2+ mode 578 if (UseSSE >= 2) { 579 __ verify_FPU(0, "c2i transition should have clean FPU stack"); 580 } else { 581 __ empty_FPU_stack(); 582 } 583 #endif /* COMPILER2 */ 584 585 // Since all args are passed on the stack, total_args_passed * interpreter_ 586 // stack_element_size is the 587 // space we need. 588 int extraspace = sig_extended.length() * Interpreter::stackElementSize; 589 590 // Get return address 591 __ pop(rax); 592 593 // set senderSP value 594 __ movptr(rsi, rsp); 595 596 __ subptr(rsp, extraspace); 597 598 // Now write the args into the outgoing interpreter space 599 for (int i = 0; i < sig_extended.length(); i++) { 600 if (sig_extended.at(i)._bt == T_VOID) { 601 assert(i > 0 && (sig_extended.at(i-1)._bt == T_LONG || sig_extended.at(i-1)._bt == T_DOUBLE), "missing half"); 602 continue; 603 } 604 605 // st_off points to lowest address on stack. 606 int st_off = ((sig_extended.length() - 1) - i) * Interpreter::stackElementSize; 607 int next_off = st_off - Interpreter::stackElementSize; 608 609 // Say 4 args: 610 // i st_off 611 // 0 12 T_LONG 612 // 1 8 T_VOID 613 // 2 4 T_OBJECT 614 // 3 0 T_BOOL 615 VMReg r_1 = regs[i].first(); 616 VMReg r_2 = regs[i].second(); 617 if (!r_1->is_valid()) { 618 assert(!r_2->is_valid(), ""); 619 continue; 620 } 621 622 if (r_1->is_stack()) { 623 // memory to memory use fpu stack top 624 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace; 625 626 if (!r_2->is_valid()) { 627 __ movl(rdi, Address(rsp, ld_off)); 628 __ movptr(Address(rsp, st_off), rdi); 629 } else { 630 631 // ld_off == LSW, ld_off+VMRegImpl::stack_slot_size == MSW 632 // st_off == MSW, st_off-wordSize == LSW 633 634 __ movptr(rdi, Address(rsp, ld_off)); 635 __ movptr(Address(rsp, next_off), rdi); 636 #ifndef _LP64 637 __ movptr(rdi, Address(rsp, ld_off + wordSize)); 638 __ movptr(Address(rsp, st_off), rdi); 639 #else 640 #ifdef ASSERT 641 // Overwrite the unused slot with known junk 642 __ mov64(rax, CONST64(0xdeadffffdeadaaaa)); 643 __ movptr(Address(rsp, st_off), rax); 644 #endif /* ASSERT */ 645 #endif // _LP64 646 } 647 } else if (r_1->is_Register()) { 648 Register r = r_1->as_Register(); 649 if (!r_2->is_valid()) { 650 __ movl(Address(rsp, st_off), r); 651 } else { 652 // long/double in gpr 653 NOT_LP64(ShouldNotReachHere()); 654 // Two VMRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 655 // T_DOUBLE and T_LONG use two slots in the interpreter 656 if ( sig_extended.at(i)._bt == T_LONG || sig_extended.at(i)._bt == T_DOUBLE) { 657 // long/double in gpr 658 #ifdef ASSERT 659 // Overwrite the unused slot with known junk 660 LP64_ONLY(__ mov64(rax, CONST64(0xdeadffffdeadaaab))); 661 __ movptr(Address(rsp, st_off), rax); 662 #endif /* ASSERT */ 663 __ movptr(Address(rsp, next_off), r); 664 } else { 665 __ movptr(Address(rsp, st_off), r); 666 } 667 } 668 } else { 669 assert(r_1->is_XMMRegister(), ""); 670 if (!r_2->is_valid()) { 671 __ movflt(Address(rsp, st_off), r_1->as_XMMRegister()); 672 } else { 673 assert(sig_extended.at(i)._bt == T_DOUBLE || sig_extended.at(i)._bt == T_LONG, "wrong type"); 674 move_c2i_double(masm, r_1->as_XMMRegister(), st_off); 675 } 676 } 677 } 678 679 // Schedule the branch target address early. 680 __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset()))); 681 // And repush original return address 682 __ push(rax); 683 __ jmp(rcx); 684 } 685 686 687 static void move_i2c_double(MacroAssembler *masm, XMMRegister r, Register saved_sp, int ld_off) { 688 int next_val_off = ld_off - Interpreter::stackElementSize; 689 __ movdbl(r, Address(saved_sp, next_val_off)); 690 } 691 692 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg, 693 address code_start, address code_end, 694 Label& L_ok) { 695 Label L_fail; 696 __ lea(temp_reg, ExternalAddress(code_start)); 697 __ cmpptr(pc_reg, temp_reg); 698 __ jcc(Assembler::belowEqual, L_fail); 699 __ lea(temp_reg, ExternalAddress(code_end)); 700 __ cmpptr(pc_reg, temp_reg); 701 __ jcc(Assembler::below, L_ok); 702 __ bind(L_fail); 703 } 704 705 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 706 int comp_args_on_stack, 707 const GrowableArray<SigEntry>& sig_extended, 708 const VMRegPair *regs) { 709 710 // Note: rsi contains the senderSP on entry. We must preserve it since 711 // we may do a i2c -> c2i transition if we lose a race where compiled 712 // code goes non-entrant while we get args ready. 713 714 // Adapters can be frameless because they do not require the caller 715 // to perform additional cleanup work, such as correcting the stack pointer. 716 // An i2c adapter is frameless because the *caller* frame, which is interpreted, 717 // routinely repairs its own stack pointer (from interpreter_frame_last_sp), 718 // even if a callee has modified the stack pointer. 719 // A c2i adapter is frameless because the *callee* frame, which is interpreted, 720 // routinely repairs its caller's stack pointer (from sender_sp, which is set 721 // up via the senderSP register). 722 // In other words, if *either* the caller or callee is interpreted, we can 723 // get the stack pointer repaired after a call. 724 // This is why c2i and i2c adapters cannot be indefinitely composed. 725 // In particular, if a c2i adapter were to somehow call an i2c adapter, 726 // both caller and callee would be compiled methods, and neither would 727 // clean up the stack pointer changes performed by the two adapters. 728 // If this happens, control eventually transfers back to the compiled 729 // caller, but with an uncorrected stack, causing delayed havoc. 730 731 // Pick up the return address 732 __ movptr(rax, Address(rsp, 0)); 733 734 if (VerifyAdapterCalls && 735 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) { 736 // So, let's test for cascading c2i/i2c adapters right now. 737 // assert(Interpreter::contains($return_addr) || 738 // StubRoutines::contains($return_addr), 739 // "i2c adapter must return to an interpreter frame"); 740 __ block_comment("verify_i2c { "); 741 Label L_ok; 742 if (Interpreter::code() != NULL) 743 range_check(masm, rax, rdi, 744 Interpreter::code()->code_start(), Interpreter::code()->code_end(), 745 L_ok); 746 if (StubRoutines::code1() != NULL) 747 range_check(masm, rax, rdi, 748 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(), 749 L_ok); 750 if (StubRoutines::code2() != NULL) 751 range_check(masm, rax, rdi, 752 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(), 753 L_ok); 754 const char* msg = "i2c adapter must return to an interpreter frame"; 755 __ block_comment(msg); 756 __ stop(msg); 757 __ bind(L_ok); 758 __ block_comment("} verify_i2ce "); 759 } 760 761 // Must preserve original SP for loading incoming arguments because 762 // we need to align the outgoing SP for compiled code. 763 __ movptr(rdi, rsp); 764 765 // Cut-out for having no stack args. Since up to 2 int/oop args are passed 766 // in registers, we will occasionally have no stack args. 767 int comp_words_on_stack = 0; 768 if (comp_args_on_stack) { 769 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in 770 // registers are below. By subtracting stack0, we either get a negative 771 // number (all values in registers) or the maximum stack slot accessed. 772 // int comp_args_on_stack = VMRegImpl::reg2stack(max_arg); 773 // Convert 4-byte stack slots to words. 774 comp_words_on_stack = round_to(comp_args_on_stack*4, wordSize)>>LogBytesPerWord; 775 // Round up to miminum stack alignment, in wordSize 776 comp_words_on_stack = round_to(comp_words_on_stack, 2); 777 __ subptr(rsp, comp_words_on_stack * wordSize); 778 } 779 780 // Align the outgoing SP 781 __ andptr(rsp, -(StackAlignmentInBytes)); 782 783 // push the return address on the stack (note that pushing, rather 784 // than storing it, yields the correct frame alignment for the callee) 785 __ push(rax); 786 787 // Put saved SP in another register 788 const Register saved_sp = rax; 789 __ movptr(saved_sp, rdi); 790 791 792 // Will jump to the compiled code just as if compiled code was doing it. 793 // Pre-load the register-jump target early, to schedule it better. 794 __ movptr(rdi, Address(rbx, in_bytes(Method::from_compiled_offset()))); 795 796 // Now generate the shuffle code. Pick up all register args and move the 797 // rest through the floating point stack top. 798 for (int i = 0; i < sig_extended.length(); i++) { 799 if (sig_extended.at(i)._bt == T_VOID) { 800 // Longs and doubles are passed in native word order, but misaligned 801 // in the 32-bit build. 802 assert(i > 0 && (sig_extended.at(i-1)._bt == T_LONG || sig_extended.at(i-1)._bt == T_DOUBLE), "missing half"); 803 continue; 804 } 805 806 // Pick up 0, 1 or 2 words from SP+offset. 807 808 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 809 "scrambled load targets?"); 810 // Load in argument order going down. 811 int ld_off = (sig_extended.length() - i) * Interpreter::stackElementSize; 812 // Point to interpreter value (vs. tag) 813 int next_off = ld_off - Interpreter::stackElementSize; 814 // 815 // 816 // 817 VMReg r_1 = regs[i].first(); 818 VMReg r_2 = regs[i].second(); 819 if (!r_1->is_valid()) { 820 assert(!r_2->is_valid(), ""); 821 continue; 822 } 823 if (r_1->is_stack()) { 824 // Convert stack slot to an SP offset (+ wordSize to account for return address ) 825 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize; 826 827 // We can use rsi as a temp here because compiled code doesn't need rsi as an input 828 // and if we end up going thru a c2i because of a miss a reasonable value of rsi 829 // we be generated. 830 if (!r_2->is_valid()) { 831 // __ fld_s(Address(saved_sp, ld_off)); 832 // __ fstp_s(Address(rsp, st_off)); 833 __ movl(rsi, Address(saved_sp, ld_off)); 834 __ movptr(Address(rsp, st_off), rsi); 835 } else { 836 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 837 // are accessed as negative so LSW is at LOW address 838 839 // ld_off is MSW so get LSW 840 // st_off is LSW (i.e. reg.first()) 841 // __ fld_d(Address(saved_sp, next_off)); 842 // __ fstp_d(Address(rsp, st_off)); 843 // 844 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE 845 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case 846 // So we must adjust where to pick up the data to match the interpreter. 847 // 848 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 849 // are accessed as negative so LSW is at LOW address 850 851 // ld_off is MSW so get LSW 852 const int offset = (NOT_LP64(true ||) sig_extended.at(i)._bt==T_LONG||sig_extended.at(i)._bt==T_DOUBLE)? 853 next_off : ld_off; 854 __ movptr(rsi, Address(saved_sp, offset)); 855 __ movptr(Address(rsp, st_off), rsi); 856 #ifndef _LP64 857 __ movptr(rsi, Address(saved_sp, ld_off)); 858 __ movptr(Address(rsp, st_off + wordSize), rsi); 859 #endif // _LP64 860 } 861 } else if (r_1->is_Register()) { // Register argument 862 Register r = r_1->as_Register(); 863 assert(r != rax, "must be different"); 864 if (r_2->is_valid()) { 865 // 866 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE 867 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case 868 // So we must adjust where to pick up the data to match the interpreter. 869 870 const int offset = (NOT_LP64(true ||) sig_extended.at(i)._bt==T_LONG||sig_extended.at(i)._bt==T_DOUBLE)? 871 next_off : ld_off; 872 873 // this can be a misaligned move 874 __ movptr(r, Address(saved_sp, offset)); 875 #ifndef _LP64 876 assert(r_2->as_Register() != rax, "need another temporary register"); 877 // Remember r_1 is low address (and LSB on x86) 878 // So r_2 gets loaded from high address regardless of the platform 879 __ movptr(r_2->as_Register(), Address(saved_sp, ld_off)); 880 #endif // _LP64 881 } else { 882 __ movl(r, Address(saved_sp, ld_off)); 883 } 884 } else { 885 assert(r_1->is_XMMRegister(), ""); 886 if (!r_2->is_valid()) { 887 __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off)); 888 } else { 889 move_i2c_double(masm, r_1->as_XMMRegister(), saved_sp, ld_off); 890 } 891 } 892 } 893 894 // 6243940 We might end up in handle_wrong_method if 895 // the callee is deoptimized as we race thru here. If that 896 // happens we don't want to take a safepoint because the 897 // caller frame will look interpreted and arguments are now 898 // "compiled" so it is much better to make this transition 899 // invisible to the stack walking code. Unfortunately if 900 // we try and find the callee by normal means a safepoint 901 // is possible. So we stash the desired callee in the thread 902 // and the vm will find there should this case occur. 903 904 __ get_thread(rax); 905 __ movptr(Address(rax, JavaThread::callee_target_offset()), rbx); 906 907 // move Method* to rax, in case we end up in an c2i adapter. 908 // the c2i adapters expect Method* in rax, (c2) because c2's 909 // resolve stubs return the result (the method) in rax,. 910 // I'd love to fix this. 911 __ mov(rax, rbx); 912 913 __ jmp(rdi); 914 } 915 916 // --------------------------------------------------------------- 917 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 918 int comp_args_on_stack, 919 const GrowableArray<SigEntry>& sig_extended, 920 const VMRegPair *regs, 921 AdapterFingerPrint* fingerprint, 922 AdapterBlob*& new_adapter) { 923 address i2c_entry = __ pc(); 924 925 gen_i2c_adapter(masm, comp_args_on_stack, sig_extended, regs); 926 927 // ------------------------------------------------------------------------- 928 // Generate a C2I adapter. On entry we know rbx, holds the Method* during calls 929 // to the interpreter. The args start out packed in the compiled layout. They 930 // need to be unpacked into the interpreter layout. This will almost always 931 // require some stack space. We grow the current (compiled) stack, then repack 932 // the args. We finally end in a jump to the generic interpreter entry point. 933 // On exit from the interpreter, the interpreter will restore our SP (lest the 934 // compiled code, which relys solely on SP and not EBP, get sick). 935 936 address c2i_unverified_entry = __ pc(); 937 Label skip_fixup; 938 939 Register holder = rax; 940 Register receiver = rcx; 941 Register temp = rbx; 942 943 { 944 945 Label missed; 946 __ movptr(temp, Address(receiver, oopDesc::klass_offset_in_bytes())); 947 __ cmpptr(temp, Address(holder, CompiledICHolder::holder_klass_offset())); 948 __ movptr(rbx, Address(holder, CompiledICHolder::holder_method_offset())); 949 __ jcc(Assembler::notEqual, missed); 950 // Method might have been compiled since the call site was patched to 951 // interpreted if that is the case treat it as a miss so we can get 952 // the call site corrected. 953 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD); 954 __ jcc(Assembler::equal, skip_fixup); 955 956 __ bind(missed); 957 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 958 } 959 960 address c2i_entry = __ pc(); 961 962 OopMapSet* oop_maps = NULL; 963 int frame_complete = CodeOffsets::frame_never_safe; 964 int frame_size_in_words = 0; 965 gen_c2i_adapter(masm, sig_extended, regs, skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words); 966 967 __ flush(); 968 new_adapter = AdapterBlob::create(masm->code(), frame_complete, frame_size_in_words, oop_maps); 969 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 970 } 971 972 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 973 VMRegPair *regs, 974 VMRegPair *regs2, 975 int total_args_passed) { 976 assert(regs2 == NULL, "not needed on x86"); 977 // We return the amount of VMRegImpl stack slots we need to reserve for all 978 // the arguments NOT counting out_preserve_stack_slots. 979 980 uint stack = 0; // All arguments on stack 981 982 for( int i = 0; i < total_args_passed; i++) { 983 // From the type and the argument number (count) compute the location 984 switch( sig_bt[i] ) { 985 case T_BOOLEAN: 986 case T_CHAR: 987 case T_FLOAT: 988 case T_BYTE: 989 case T_SHORT: 990 case T_INT: 991 case T_OBJECT: 992 case T_ARRAY: 993 case T_ADDRESS: 994 case T_METADATA: 995 regs[i].set1(VMRegImpl::stack2reg(stack++)); 996 break; 997 case T_LONG: 998 case T_DOUBLE: // The stack numbering is reversed from Java 999 // Since C arguments do not get reversed, the ordering for 1000 // doubles on the stack must be opposite the Java convention 1001 assert(sig_bt[i+1] == T_VOID, "missing Half" ); 1002 regs[i].set2(VMRegImpl::stack2reg(stack)); 1003 stack += 2; 1004 break; 1005 case T_VOID: regs[i].set_bad(); break; 1006 default: 1007 ShouldNotReachHere(); 1008 break; 1009 } 1010 } 1011 return stack; 1012 } 1013 1014 // A simple move of integer like type 1015 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1016 if (src.first()->is_stack()) { 1017 if (dst.first()->is_stack()) { 1018 // stack to stack 1019 // __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); 1020 // __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS); 1021 __ movl2ptr(rax, Address(rbp, reg2offset_in(src.first()))); 1022 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1023 } else { 1024 // stack to reg 1025 __ movl2ptr(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first()))); 1026 } 1027 } else if (dst.first()->is_stack()) { 1028 // reg to stack 1029 // no need to sign extend on 64bit 1030 __ movptr(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register()); 1031 } else { 1032 if (dst.first() != src.first()) { 1033 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 1034 } 1035 } 1036 } 1037 1038 // An oop arg. Must pass a handle not the oop itself 1039 static void object_move(MacroAssembler* masm, 1040 OopMap* map, 1041 int oop_handle_offset, 1042 int framesize_in_slots, 1043 VMRegPair src, 1044 VMRegPair dst, 1045 bool is_receiver, 1046 int* receiver_offset) { 1047 1048 // Because of the calling conventions we know that src can be a 1049 // register or a stack location. dst can only be a stack location. 1050 1051 assert(dst.first()->is_stack(), "must be stack"); 1052 // must pass a handle. First figure out the location we use as a handle 1053 1054 if (src.first()->is_stack()) { 1055 // Oop is already on the stack as an argument 1056 Register rHandle = rax; 1057 Label nil; 1058 __ xorptr(rHandle, rHandle); 1059 __ cmpptr(Address(rbp, reg2offset_in(src.first())), (int32_t)NULL_WORD); 1060 __ jcc(Assembler::equal, nil); 1061 __ lea(rHandle, Address(rbp, reg2offset_in(src.first()))); 1062 __ bind(nil); 1063 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle); 1064 1065 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1066 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 1067 if (is_receiver) { 1068 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 1069 } 1070 } else { 1071 // Oop is in an a register we must store it to the space we reserve 1072 // on the stack for oop_handles 1073 const Register rOop = src.first()->as_Register(); 1074 const Register rHandle = rax; 1075 int oop_slot = (rOop == rcx ? 0 : 1) * VMRegImpl::slots_per_word + oop_handle_offset; 1076 int offset = oop_slot*VMRegImpl::stack_slot_size; 1077 Label skip; 1078 __ movptr(Address(rsp, offset), rOop); 1079 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1080 __ xorptr(rHandle, rHandle); 1081 __ cmpptr(rOop, (int32_t)NULL_WORD); 1082 __ jcc(Assembler::equal, skip); 1083 __ lea(rHandle, Address(rsp, offset)); 1084 __ bind(skip); 1085 // Store the handle parameter 1086 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle); 1087 if (is_receiver) { 1088 *receiver_offset = offset; 1089 } 1090 } 1091 } 1092 1093 // A float arg may have to do float reg int reg conversion 1094 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1095 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move"); 1096 1097 // Because of the calling convention we know that src is either a stack location 1098 // or an xmm register. dst can only be a stack location. 1099 1100 assert(dst.first()->is_stack() && ( src.first()->is_stack() || src.first()->is_XMMRegister()), "bad parameters"); 1101 1102 if (src.first()->is_stack()) { 1103 __ movl(rax, Address(rbp, reg2offset_in(src.first()))); 1104 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1105 } else { 1106 // reg to stack 1107 __ movflt(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister()); 1108 } 1109 } 1110 1111 // A long move 1112 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1113 1114 // The only legal possibility for a long_move VMRegPair is: 1115 // 1: two stack slots (possibly unaligned) 1116 // as neither the java or C calling convention will use registers 1117 // for longs. 1118 1119 if (src.first()->is_stack() && dst.first()->is_stack()) { 1120 assert(src.second()->is_stack() && dst.second()->is_stack(), "must be all stack"); 1121 __ movptr(rax, Address(rbp, reg2offset_in(src.first()))); 1122 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second())))); 1123 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1124 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx)); 1125 } else { 1126 ShouldNotReachHere(); 1127 } 1128 } 1129 1130 // A double move 1131 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1132 1133 // The only legal possibilities for a double_move VMRegPair are: 1134 // The painful thing here is that like long_move a VMRegPair might be 1135 1136 // Because of the calling convention we know that src is either 1137 // 1: a single physical register (xmm registers only) 1138 // 2: two stack slots (possibly unaligned) 1139 // dst can only be a pair of stack slots. 1140 1141 assert(dst.first()->is_stack() && (src.first()->is_XMMRegister() || src.first()->is_stack()), "bad args"); 1142 1143 if (src.first()->is_stack()) { 1144 // source is all stack 1145 __ movptr(rax, Address(rbp, reg2offset_in(src.first()))); 1146 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second())))); 1147 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1148 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx)); 1149 } else { 1150 // reg to stack 1151 // No worries about stack alignment 1152 __ movdbl(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister()); 1153 } 1154 } 1155 1156 1157 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1158 // We always ignore the frame_slots arg and just use the space just below frame pointer 1159 // which by this time is free to use 1160 switch (ret_type) { 1161 case T_FLOAT: 1162 __ fstp_s(Address(rbp, -wordSize)); 1163 break; 1164 case T_DOUBLE: 1165 __ fstp_d(Address(rbp, -2*wordSize)); 1166 break; 1167 case T_VOID: break; 1168 case T_LONG: 1169 __ movptr(Address(rbp, -wordSize), rax); 1170 NOT_LP64(__ movptr(Address(rbp, -2*wordSize), rdx)); 1171 break; 1172 default: { 1173 __ movptr(Address(rbp, -wordSize), rax); 1174 } 1175 } 1176 } 1177 1178 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1179 // We always ignore the frame_slots arg and just use the space just below frame pointer 1180 // which by this time is free to use 1181 switch (ret_type) { 1182 case T_FLOAT: 1183 __ fld_s(Address(rbp, -wordSize)); 1184 break; 1185 case T_DOUBLE: 1186 __ fld_d(Address(rbp, -2*wordSize)); 1187 break; 1188 case T_LONG: 1189 __ movptr(rax, Address(rbp, -wordSize)); 1190 NOT_LP64(__ movptr(rdx, Address(rbp, -2*wordSize))); 1191 break; 1192 case T_VOID: break; 1193 default: { 1194 __ movptr(rax, Address(rbp, -wordSize)); 1195 } 1196 } 1197 } 1198 1199 1200 static void save_or_restore_arguments(MacroAssembler* masm, 1201 const int stack_slots, 1202 const int total_in_args, 1203 const int arg_save_area, 1204 OopMap* map, 1205 VMRegPair* in_regs, 1206 BasicType* in_sig_bt) { 1207 // if map is non-NULL then the code should store the values, 1208 // otherwise it should load them. 1209 int handle_index = 0; 1210 // Save down double word first 1211 for ( int i = 0; i < total_in_args; i++) { 1212 if (in_regs[i].first()->is_XMMRegister() && in_sig_bt[i] == T_DOUBLE) { 1213 int slot = handle_index * VMRegImpl::slots_per_word + arg_save_area; 1214 int offset = slot * VMRegImpl::stack_slot_size; 1215 handle_index += 2; 1216 assert(handle_index <= stack_slots, "overflow"); 1217 if (map != NULL) { 1218 __ movdbl(Address(rsp, offset), in_regs[i].first()->as_XMMRegister()); 1219 } else { 1220 __ movdbl(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset)); 1221 } 1222 } 1223 if (in_regs[i].first()->is_Register() && in_sig_bt[i] == T_LONG) { 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 __ movl(Address(rsp, offset), in_regs[i].first()->as_Register()); 1230 if (in_regs[i].second()->is_Register()) { 1231 __ movl(Address(rsp, offset + 4), in_regs[i].second()->as_Register()); 1232 } 1233 } else { 1234 __ movl(in_regs[i].first()->as_Register(), Address(rsp, offset)); 1235 if (in_regs[i].second()->is_Register()) { 1236 __ movl(in_regs[i].second()->as_Register(), Address(rsp, offset + 4)); 1237 } 1238 } 1239 } 1240 } 1241 // Save or restore single word registers 1242 for ( int i = 0; i < total_in_args; i++) { 1243 if (in_regs[i].first()->is_Register()) { 1244 int slot = handle_index++ * VMRegImpl::slots_per_word + arg_save_area; 1245 int offset = slot * VMRegImpl::stack_slot_size; 1246 assert(handle_index <= stack_slots, "overflow"); 1247 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1248 map->set_oop(VMRegImpl::stack2reg(slot));; 1249 } 1250 1251 // Value is in an input register pass we must flush it to the stack 1252 const Register reg = in_regs[i].first()->as_Register(); 1253 switch (in_sig_bt[i]) { 1254 case T_ARRAY: 1255 if (map != NULL) { 1256 __ movptr(Address(rsp, offset), reg); 1257 } else { 1258 __ movptr(reg, Address(rsp, offset)); 1259 } 1260 break; 1261 case T_BOOLEAN: 1262 case T_CHAR: 1263 case T_BYTE: 1264 case T_SHORT: 1265 case T_INT: 1266 if (map != NULL) { 1267 __ movl(Address(rsp, offset), reg); 1268 } else { 1269 __ movl(reg, Address(rsp, offset)); 1270 } 1271 break; 1272 case T_OBJECT: 1273 default: ShouldNotReachHere(); 1274 } 1275 } else if (in_regs[i].first()->is_XMMRegister()) { 1276 if (in_sig_bt[i] == T_FLOAT) { 1277 int slot = handle_index++ * VMRegImpl::slots_per_word + arg_save_area; 1278 int offset = slot * VMRegImpl::stack_slot_size; 1279 assert(handle_index <= stack_slots, "overflow"); 1280 if (map != NULL) { 1281 __ movflt(Address(rsp, offset), in_regs[i].first()->as_XMMRegister()); 1282 } else { 1283 __ movflt(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset)); 1284 } 1285 } 1286 } else if (in_regs[i].first()->is_stack()) { 1287 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1288 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1289 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots)); 1290 } 1291 } 1292 } 1293 } 1294 1295 // Check GCLocker::needs_gc and enter the runtime if it's true. This 1296 // keeps a new JNI critical region from starting until a GC has been 1297 // forced. Save down any oops in registers and describe them in an 1298 // OopMap. 1299 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1300 Register thread, 1301 int stack_slots, 1302 int total_c_args, 1303 int total_in_args, 1304 int arg_save_area, 1305 OopMapSet* oop_maps, 1306 VMRegPair* in_regs, 1307 BasicType* in_sig_bt) { 1308 __ block_comment("check GCLocker::needs_gc"); 1309 Label cont; 1310 __ cmp8(ExternalAddress((address)GCLocker::needs_gc_address()), false); 1311 __ jcc(Assembler::equal, cont); 1312 1313 // Save down any incoming oops and call into the runtime to halt for a GC 1314 1315 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1316 1317 save_or_restore_arguments(masm, stack_slots, total_in_args, 1318 arg_save_area, map, in_regs, in_sig_bt); 1319 1320 address the_pc = __ pc(); 1321 oop_maps->add_gc_map( __ offset(), map); 1322 __ set_last_Java_frame(thread, rsp, noreg, the_pc); 1323 1324 __ block_comment("block_for_jni_critical"); 1325 __ push(thread); 1326 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical))); 1327 __ increment(rsp, wordSize); 1328 1329 __ get_thread(thread); 1330 __ reset_last_Java_frame(thread, false, true); 1331 1332 save_or_restore_arguments(masm, stack_slots, total_in_args, 1333 arg_save_area, NULL, in_regs, in_sig_bt); 1334 1335 __ bind(cont); 1336 #ifdef ASSERT 1337 if (StressCriticalJNINatives) { 1338 // Stress register saving 1339 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1340 save_or_restore_arguments(masm, stack_slots, total_in_args, 1341 arg_save_area, map, in_regs, in_sig_bt); 1342 // Destroy argument registers 1343 for (int i = 0; i < total_in_args - 1; i++) { 1344 if (in_regs[i].first()->is_Register()) { 1345 const Register reg = in_regs[i].first()->as_Register(); 1346 __ xorptr(reg, reg); 1347 } else if (in_regs[i].first()->is_XMMRegister()) { 1348 __ xorpd(in_regs[i].first()->as_XMMRegister(), in_regs[i].first()->as_XMMRegister()); 1349 } else if (in_regs[i].first()->is_FloatRegister()) { 1350 ShouldNotReachHere(); 1351 } else if (in_regs[i].first()->is_stack()) { 1352 // Nothing to do 1353 } else { 1354 ShouldNotReachHere(); 1355 } 1356 if (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_DOUBLE) { 1357 i++; 1358 } 1359 } 1360 1361 save_or_restore_arguments(masm, stack_slots, total_in_args, 1362 arg_save_area, NULL, in_regs, in_sig_bt); 1363 } 1364 #endif 1365 } 1366 1367 // Unpack an array argument into a pointer to the body and the length 1368 // if the array is non-null, otherwise pass 0 for both. 1369 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) { 1370 Register tmp_reg = rax; 1371 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg, 1372 "possible collision"); 1373 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg, 1374 "possible collision"); 1375 1376 // Pass the length, ptr pair 1377 Label is_null, done; 1378 VMRegPair tmp(tmp_reg->as_VMReg()); 1379 if (reg.first()->is_stack()) { 1380 // Load the arg up from the stack 1381 simple_move32(masm, reg, tmp); 1382 reg = tmp; 1383 } 1384 __ testptr(reg.first()->as_Register(), reg.first()->as_Register()); 1385 __ jccb(Assembler::equal, is_null); 1386 __ lea(tmp_reg, Address(reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type))); 1387 simple_move32(masm, tmp, body_arg); 1388 // load the length relative to the body. 1389 __ movl(tmp_reg, Address(tmp_reg, arrayOopDesc::length_offset_in_bytes() - 1390 arrayOopDesc::base_offset_in_bytes(in_elem_type))); 1391 simple_move32(masm, tmp, length_arg); 1392 __ jmpb(done); 1393 __ bind(is_null); 1394 // Pass zeros 1395 __ xorptr(tmp_reg, tmp_reg); 1396 simple_move32(masm, tmp, body_arg); 1397 simple_move32(masm, tmp, length_arg); 1398 __ bind(done); 1399 } 1400 1401 static void verify_oop_args(MacroAssembler* masm, 1402 methodHandle method, 1403 const BasicType* sig_bt, 1404 const VMRegPair* regs) { 1405 Register temp_reg = rbx; // not part of any compiled calling seq 1406 if (VerifyOops) { 1407 for (int i = 0; i < method->size_of_parameters(); i++) { 1408 if (sig_bt[i] == T_OBJECT || 1409 sig_bt[i] == T_ARRAY) { 1410 VMReg r = regs[i].first(); 1411 assert(r->is_valid(), "bad oop arg"); 1412 if (r->is_stack()) { 1413 __ movptr(temp_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1414 __ verify_oop(temp_reg); 1415 } else { 1416 __ verify_oop(r->as_Register()); 1417 } 1418 } 1419 } 1420 } 1421 } 1422 1423 static void gen_special_dispatch(MacroAssembler* masm, 1424 methodHandle method, 1425 const BasicType* sig_bt, 1426 const VMRegPair* regs) { 1427 verify_oop_args(masm, method, sig_bt, regs); 1428 vmIntrinsics::ID iid = method->intrinsic_id(); 1429 1430 // Now write the args into the outgoing interpreter space 1431 bool has_receiver = false; 1432 Register receiver_reg = noreg; 1433 int member_arg_pos = -1; 1434 Register member_reg = noreg; 1435 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1436 if (ref_kind != 0) { 1437 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1438 member_reg = rbx; // known to be free at this point 1439 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1440 } else if (iid == vmIntrinsics::_invokeBasic) { 1441 has_receiver = true; 1442 } else { 1443 fatal("unexpected intrinsic id %d", iid); 1444 } 1445 1446 if (member_reg != noreg) { 1447 // Load the member_arg into register, if necessary. 1448 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1449 VMReg r = regs[member_arg_pos].first(); 1450 if (r->is_stack()) { 1451 __ movptr(member_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1452 } else { 1453 // no data motion is needed 1454 member_reg = r->as_Register(); 1455 } 1456 } 1457 1458 if (has_receiver) { 1459 // Make sure the receiver is loaded into a register. 1460 assert(method->size_of_parameters() > 0, "oob"); 1461 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1462 VMReg r = regs[0].first(); 1463 assert(r->is_valid(), "bad receiver arg"); 1464 if (r->is_stack()) { 1465 // Porting note: This assumes that compiled calling conventions always 1466 // pass the receiver oop in a register. If this is not true on some 1467 // platform, pick a temp and load the receiver from stack. 1468 fatal("receiver always in a register"); 1469 receiver_reg = rcx; // known to be free at this point 1470 __ movptr(receiver_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1471 } else { 1472 // no data motion is needed 1473 receiver_reg = r->as_Register(); 1474 } 1475 } 1476 1477 // Figure out which address we are really jumping to: 1478 MethodHandles::generate_method_handle_dispatch(masm, iid, 1479 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1480 } 1481 1482 // --------------------------------------------------------------------------- 1483 // Generate a native wrapper for a given method. The method takes arguments 1484 // in the Java compiled code convention, marshals them to the native 1485 // convention (handlizes oops, etc), transitions to native, makes the call, 1486 // returns to java state (possibly blocking), unhandlizes any result and 1487 // returns. 1488 // 1489 // Critical native functions are a shorthand for the use of 1490 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1491 // functions. The wrapper is expected to unpack the arguments before 1492 // passing them to the callee and perform checks before and after the 1493 // native call to ensure that they GCLocker 1494 // lock_critical/unlock_critical semantics are followed. Some other 1495 // parts of JNI setup are skipped like the tear down of the JNI handle 1496 // block and the check for pending exceptions it's impossible for them 1497 // to be thrown. 1498 // 1499 // They are roughly structured like this: 1500 // if (GCLocker::needs_gc()) 1501 // SharedRuntime::block_for_jni_critical(); 1502 // tranistion to thread_in_native 1503 // unpack arrray arguments and call native entry point 1504 // check for safepoint in progress 1505 // check if any thread suspend flags are set 1506 // call into JVM and possible unlock the JNI critical 1507 // if a GC was suppressed while in the critical native. 1508 // transition back to thread_in_Java 1509 // return to caller 1510 // 1511 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1512 const methodHandle& method, 1513 int compile_id, 1514 BasicType* in_sig_bt, 1515 VMRegPair* in_regs, 1516 BasicType ret_type) { 1517 if (method->is_method_handle_intrinsic()) { 1518 vmIntrinsics::ID iid = method->intrinsic_id(); 1519 intptr_t start = (intptr_t)__ pc(); 1520 int vep_offset = ((intptr_t)__ pc()) - start; 1521 gen_special_dispatch(masm, 1522 method, 1523 in_sig_bt, 1524 in_regs); 1525 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1526 __ flush(); 1527 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1528 return nmethod::new_native_nmethod(method, 1529 compile_id, 1530 masm->code(), 1531 vep_offset, 1532 frame_complete, 1533 stack_slots / VMRegImpl::slots_per_word, 1534 in_ByteSize(-1), 1535 in_ByteSize(-1), 1536 (OopMapSet*)NULL); 1537 } 1538 bool is_critical_native = true; 1539 address native_func = method->critical_native_function(); 1540 if (native_func == NULL) { 1541 native_func = method->native_function(); 1542 is_critical_native = false; 1543 } 1544 assert(native_func != NULL, "must have function"); 1545 1546 // An OopMap for lock (and class if static) 1547 OopMapSet *oop_maps = new OopMapSet(); 1548 1549 // We have received a description of where all the java arg are located 1550 // on entry to the wrapper. We need to convert these args to where 1551 // the jni function will expect them. To figure out where they go 1552 // we convert the java signature to a C signature by inserting 1553 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1554 1555 const int total_in_args = method->size_of_parameters(); 1556 int total_c_args = total_in_args; 1557 if (!is_critical_native) { 1558 total_c_args += 1; 1559 if (method->is_static()) { 1560 total_c_args++; 1561 } 1562 } else { 1563 for (int i = 0; i < total_in_args; i++) { 1564 if (in_sig_bt[i] == T_ARRAY) { 1565 total_c_args++; 1566 } 1567 } 1568 } 1569 1570 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1571 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1572 BasicType* in_elem_bt = NULL; 1573 1574 int argc = 0; 1575 if (!is_critical_native) { 1576 out_sig_bt[argc++] = T_ADDRESS; 1577 if (method->is_static()) { 1578 out_sig_bt[argc++] = T_OBJECT; 1579 } 1580 1581 for (int i = 0; i < total_in_args ; i++ ) { 1582 out_sig_bt[argc++] = in_sig_bt[i]; 1583 } 1584 } else { 1585 Thread* THREAD = Thread::current(); 1586 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args); 1587 SignatureStream ss(method->signature()); 1588 for (int i = 0; i < total_in_args ; i++ ) { 1589 if (in_sig_bt[i] == T_ARRAY) { 1590 // Arrays are passed as int, elem* pair 1591 out_sig_bt[argc++] = T_INT; 1592 out_sig_bt[argc++] = T_ADDRESS; 1593 Symbol* atype = ss.as_symbol(CHECK_NULL); 1594 const char* at = atype->as_C_string(); 1595 if (strlen(at) == 2) { 1596 assert(at[0] == '[', "must be"); 1597 switch (at[1]) { 1598 case 'B': in_elem_bt[i] = T_BYTE; break; 1599 case 'C': in_elem_bt[i] = T_CHAR; break; 1600 case 'D': in_elem_bt[i] = T_DOUBLE; break; 1601 case 'F': in_elem_bt[i] = T_FLOAT; break; 1602 case 'I': in_elem_bt[i] = T_INT; break; 1603 case 'J': in_elem_bt[i] = T_LONG; break; 1604 case 'S': in_elem_bt[i] = T_SHORT; break; 1605 case 'Z': in_elem_bt[i] = T_BOOLEAN; break; 1606 default: ShouldNotReachHere(); 1607 } 1608 } 1609 } else { 1610 out_sig_bt[argc++] = in_sig_bt[i]; 1611 in_elem_bt[i] = T_VOID; 1612 } 1613 if (in_sig_bt[i] != T_VOID) { 1614 assert(in_sig_bt[i] == ss.type(), "must match"); 1615 ss.next(); 1616 } 1617 } 1618 } 1619 1620 // Now figure out where the args must be stored and how much stack space 1621 // they require. 1622 int out_arg_slots; 1623 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args); 1624 1625 // Compute framesize for the wrapper. We need to handlize all oops in 1626 // registers a max of 2 on x86. 1627 1628 // Calculate the total number of stack slots we will need. 1629 1630 // First count the abi requirement plus all of the outgoing args 1631 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 1632 1633 // Now the space for the inbound oop handle area 1634 int total_save_slots = 2 * VMRegImpl::slots_per_word; // 2 arguments passed in registers 1635 if (is_critical_native) { 1636 // Critical natives may have to call out so they need a save area 1637 // for register arguments. 1638 int double_slots = 0; 1639 int single_slots = 0; 1640 for ( int i = 0; i < total_in_args; i++) { 1641 if (in_regs[i].first()->is_Register()) { 1642 const Register reg = in_regs[i].first()->as_Register(); 1643 switch (in_sig_bt[i]) { 1644 case T_ARRAY: // critical array (uses 2 slots on LP64) 1645 case T_BOOLEAN: 1646 case T_BYTE: 1647 case T_SHORT: 1648 case T_CHAR: 1649 case T_INT: single_slots++; break; 1650 case T_LONG: double_slots++; break; 1651 default: ShouldNotReachHere(); 1652 } 1653 } else if (in_regs[i].first()->is_XMMRegister()) { 1654 switch (in_sig_bt[i]) { 1655 case T_FLOAT: single_slots++; break; 1656 case T_DOUBLE: double_slots++; break; 1657 default: ShouldNotReachHere(); 1658 } 1659 } else if (in_regs[i].first()->is_FloatRegister()) { 1660 ShouldNotReachHere(); 1661 } 1662 } 1663 total_save_slots = double_slots * 2 + single_slots; 1664 // align the save area 1665 if (double_slots != 0) { 1666 stack_slots = round_to(stack_slots, 2); 1667 } 1668 } 1669 1670 int oop_handle_offset = stack_slots; 1671 stack_slots += total_save_slots; 1672 1673 // Now any space we need for handlizing a klass if static method 1674 1675 int klass_slot_offset = 0; 1676 int klass_offset = -1; 1677 int lock_slot_offset = 0; 1678 bool is_static = false; 1679 1680 if (method->is_static()) { 1681 klass_slot_offset = stack_slots; 1682 stack_slots += VMRegImpl::slots_per_word; 1683 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1684 is_static = true; 1685 } 1686 1687 // Plus a lock if needed 1688 1689 if (method->is_synchronized()) { 1690 lock_slot_offset = stack_slots; 1691 stack_slots += VMRegImpl::slots_per_word; 1692 } 1693 1694 // Now a place (+2) to save return values or temp during shuffling 1695 // + 2 for return address (which we own) and saved rbp, 1696 stack_slots += 4; 1697 1698 // Ok The space we have allocated will look like: 1699 // 1700 // 1701 // FP-> | | 1702 // |---------------------| 1703 // | 2 slots for moves | 1704 // |---------------------| 1705 // | lock box (if sync) | 1706 // |---------------------| <- lock_slot_offset (-lock_slot_rbp_offset) 1707 // | klass (if static) | 1708 // |---------------------| <- klass_slot_offset 1709 // | oopHandle area | 1710 // |---------------------| <- oop_handle_offset (a max of 2 registers) 1711 // | outbound memory | 1712 // | based arguments | 1713 // | | 1714 // |---------------------| 1715 // | | 1716 // SP-> | out_preserved_slots | 1717 // 1718 // 1719 // **************************************************************************** 1720 // WARNING - on Windows Java Natives use pascal calling convention and pop the 1721 // arguments off of the stack after the jni call. Before the call we can use 1722 // instructions that are SP relative. After the jni call we switch to FP 1723 // relative instructions instead of re-adjusting the stack on windows. 1724 // **************************************************************************** 1725 1726 1727 // Now compute actual number of stack words we need rounding to make 1728 // stack properly aligned. 1729 stack_slots = round_to(stack_slots, StackAlignmentInSlots); 1730 1731 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 1732 1733 intptr_t start = (intptr_t)__ pc(); 1734 1735 // First thing make an ic check to see if we should even be here 1736 1737 // We are free to use all registers as temps without saving them and 1738 // restoring them except rbp. rbp is the only callee save register 1739 // as far as the interpreter and the compiler(s) are concerned. 1740 1741 1742 const Register ic_reg = rax; 1743 const Register receiver = rcx; 1744 Label hit; 1745 Label exception_pending; 1746 1747 __ verify_oop(receiver); 1748 __ cmpptr(ic_reg, Address(receiver, oopDesc::klass_offset_in_bytes())); 1749 __ jcc(Assembler::equal, hit); 1750 1751 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1752 1753 // verified entry must be aligned for code patching. 1754 // and the first 5 bytes must be in the same cache line 1755 // if we align at 8 then we will be sure 5 bytes are in the same line 1756 __ align(8); 1757 1758 __ bind(hit); 1759 1760 int vep_offset = ((intptr_t)__ pc()) - start; 1761 1762 #ifdef COMPILER1 1763 // For Object.hashCode, System.identityHashCode try to pull hashCode from object header if available. 1764 if ((InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) || (method->intrinsic_id() == vmIntrinsics::_identityHashCode)) { 1765 inline_check_hashcode_from_object_header(masm, method, rcx /*obj_reg*/, rax /*result*/); 1766 } 1767 #endif // COMPILER1 1768 1769 // The instruction at the verified entry point must be 5 bytes or longer 1770 // because it can be patched on the fly by make_non_entrant. The stack bang 1771 // instruction fits that requirement. 1772 1773 // Generate stack overflow check 1774 1775 if (UseStackBanging) { 1776 __ bang_stack_with_offset((int)JavaThread::stack_shadow_zone_size()); 1777 } else { 1778 // need a 5 byte instruction to allow MT safe patching to non-entrant 1779 __ fat_nop(); 1780 } 1781 1782 // Generate a new frame for the wrapper. 1783 __ enter(); 1784 // -2 because return address is already present and so is saved rbp 1785 __ subptr(rsp, stack_size - 2*wordSize); 1786 1787 // Frame is now completed as far as size and linkage. 1788 int frame_complete = ((intptr_t)__ pc()) - start; 1789 1790 if (UseRTMLocking) { 1791 // Abort RTM transaction before calling JNI 1792 // because critical section will be large and will be 1793 // aborted anyway. Also nmethod could be deoptimized. 1794 __ xabort(0); 1795 } 1796 1797 // Calculate the difference between rsp and rbp,. We need to know it 1798 // after the native call because on windows Java Natives will pop 1799 // the arguments and it is painful to do rsp relative addressing 1800 // in a platform independent way. So after the call we switch to 1801 // rbp, relative addressing. 1802 1803 int fp_adjustment = stack_size - 2*wordSize; 1804 1805 #ifdef COMPILER2 1806 // C2 may leave the stack dirty if not in SSE2+ mode 1807 if (UseSSE >= 2) { 1808 __ verify_FPU(0, "c2i transition should have clean FPU stack"); 1809 } else { 1810 __ empty_FPU_stack(); 1811 } 1812 #endif /* COMPILER2 */ 1813 1814 // Compute the rbp, offset for any slots used after the jni call 1815 1816 int lock_slot_rbp_offset = (lock_slot_offset*VMRegImpl::stack_slot_size) - fp_adjustment; 1817 1818 // We use rdi as a thread pointer because it is callee save and 1819 // if we load it once it is usable thru the entire wrapper 1820 const Register thread = rdi; 1821 1822 // We use rsi as the oop handle for the receiver/klass 1823 // It is callee save so it survives the call to native 1824 1825 const Register oop_handle_reg = rsi; 1826 1827 __ get_thread(thread); 1828 1829 if (is_critical_native) { 1830 check_needs_gc_for_critical_native(masm, thread, stack_slots, total_c_args, total_in_args, 1831 oop_handle_offset, oop_maps, in_regs, in_sig_bt); 1832 } 1833 1834 // 1835 // We immediately shuffle the arguments so that any vm call we have to 1836 // make from here on out (sync slow path, jvmti, etc.) we will have 1837 // captured the oops from our caller and have a valid oopMap for 1838 // them. 1839 1840 // ----------------- 1841 // The Grand Shuffle 1842 // 1843 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv* 1844 // and, if static, the class mirror instead of a receiver. This pretty much 1845 // guarantees that register layout will not match (and x86 doesn't use reg 1846 // parms though amd does). Since the native abi doesn't use register args 1847 // and the java conventions does we don't have to worry about collisions. 1848 // All of our moved are reg->stack or stack->stack. 1849 // We ignore the extra arguments during the shuffle and handle them at the 1850 // last moment. The shuffle is described by the two calling convention 1851 // vectors we have in our possession. We simply walk the java vector to 1852 // get the source locations and the c vector to get the destinations. 1853 1854 int c_arg = is_critical_native ? 0 : (method->is_static() ? 2 : 1 ); 1855 1856 // Record rsp-based slot for receiver on stack for non-static methods 1857 int receiver_offset = -1; 1858 1859 // This is a trick. We double the stack slots so we can claim 1860 // the oops in the caller's frame. Since we are sure to have 1861 // more args than the caller doubling is enough to make 1862 // sure we can capture all the incoming oop args from the 1863 // caller. 1864 // 1865 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1866 1867 // Mark location of rbp, 1868 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, rbp->as_VMReg()); 1869 1870 // We know that we only have args in at most two integer registers (rcx, rdx). So rax, rbx 1871 // Are free to temporaries if we have to do stack to steck moves. 1872 // All inbound args are referenced based on rbp, and all outbound args via rsp. 1873 1874 for (int i = 0; i < total_in_args ; i++, c_arg++ ) { 1875 switch (in_sig_bt[i]) { 1876 case T_ARRAY: 1877 if (is_critical_native) { 1878 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]); 1879 c_arg++; 1880 break; 1881 } 1882 case T_OBJECT: 1883 assert(!is_critical_native, "no oop arguments"); 1884 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 1885 ((i == 0) && (!is_static)), 1886 &receiver_offset); 1887 break; 1888 case T_VOID: 1889 break; 1890 1891 case T_FLOAT: 1892 float_move(masm, in_regs[i], out_regs[c_arg]); 1893 break; 1894 1895 case T_DOUBLE: 1896 assert( i + 1 < total_in_args && 1897 in_sig_bt[i + 1] == T_VOID && 1898 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 1899 double_move(masm, in_regs[i], out_regs[c_arg]); 1900 break; 1901 1902 case T_LONG : 1903 long_move(masm, in_regs[i], out_regs[c_arg]); 1904 break; 1905 1906 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 1907 1908 default: 1909 simple_move32(masm, in_regs[i], out_regs[c_arg]); 1910 } 1911 } 1912 1913 // Pre-load a static method's oop into rsi. Used both by locking code and 1914 // the normal JNI call code. 1915 if (method->is_static() && !is_critical_native) { 1916 1917 // load opp into a register 1918 __ movoop(oop_handle_reg, JNIHandles::make_local(method->method_holder()->java_mirror())); 1919 1920 // Now handlize the static class mirror it's known not-null. 1921 __ movptr(Address(rsp, klass_offset), oop_handle_reg); 1922 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 1923 1924 // Now get the handle 1925 __ lea(oop_handle_reg, Address(rsp, klass_offset)); 1926 // store the klass handle as second argument 1927 __ movptr(Address(rsp, wordSize), oop_handle_reg); 1928 } 1929 1930 // Change state to native (we save the return address in the thread, since it might not 1931 // be pushed on the stack when we do a a stack traversal). It is enough that the pc() 1932 // points into the right code segment. It does not have to be the correct return pc. 1933 // We use the same pc/oopMap repeatedly when we call out 1934 1935 intptr_t the_pc = (intptr_t) __ pc(); 1936 oop_maps->add_gc_map(the_pc - start, map); 1937 1938 __ set_last_Java_frame(thread, rsp, noreg, (address)the_pc); 1939 1940 1941 // We have all of the arguments setup at this point. We must not touch any register 1942 // argument registers at this point (what if we save/restore them there are no oop? 1943 1944 { 1945 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0); 1946 __ mov_metadata(rax, method()); 1947 __ call_VM_leaf( 1948 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 1949 thread, rax); 1950 } 1951 1952 // RedefineClasses() tracing support for obsolete method entry 1953 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) { 1954 __ mov_metadata(rax, method()); 1955 __ call_VM_leaf( 1956 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 1957 thread, rax); 1958 } 1959 1960 // These are register definitions we need for locking/unlocking 1961 const Register swap_reg = rax; // Must use rax, for cmpxchg instruction 1962 const Register obj_reg = rcx; // Will contain the oop 1963 const Register lock_reg = rdx; // Address of compiler lock object (BasicLock) 1964 1965 Label slow_path_lock; 1966 Label lock_done; 1967 1968 // Lock a synchronized method 1969 if (method->is_synchronized()) { 1970 assert(!is_critical_native, "unhandled"); 1971 1972 1973 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 1974 1975 // Get the handle (the 2nd argument) 1976 __ movptr(oop_handle_reg, Address(rsp, wordSize)); 1977 1978 // Get address of the box 1979 1980 __ lea(lock_reg, Address(rbp, lock_slot_rbp_offset)); 1981 1982 // Load the oop from the handle 1983 __ movptr(obj_reg, Address(oop_handle_reg, 0)); 1984 1985 if (UseBiasedLocking) { 1986 // Note that oop_handle_reg is trashed during this call 1987 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, oop_handle_reg, false, lock_done, &slow_path_lock); 1988 } 1989 1990 // Load immediate 1 into swap_reg %rax, 1991 __ movptr(swap_reg, 1); 1992 1993 // Load (object->mark() | 1) into swap_reg %rax, 1994 __ orptr(swap_reg, Address(obj_reg, 0)); 1995 1996 // Save (object->mark() | 1) into BasicLock's displaced header 1997 __ movptr(Address(lock_reg, mark_word_offset), swap_reg); 1998 1999 if (os::is_MP()) { 2000 __ lock(); 2001 } 2002 2003 // src -> dest iff dest == rax, else rax, <- dest 2004 // *obj_reg = lock_reg iff *obj_reg == rax, else rax, = *(obj_reg) 2005 __ cmpxchgptr(lock_reg, Address(obj_reg, 0)); 2006 __ jcc(Assembler::equal, lock_done); 2007 2008 // Test if the oopMark is an obvious stack pointer, i.e., 2009 // 1) (mark & 3) == 0, and 2010 // 2) rsp <= mark < mark + os::pagesize() 2011 // These 3 tests can be done by evaluating the following 2012 // expression: ((mark - rsp) & (3 - os::vm_page_size())), 2013 // assuming both stack pointer and pagesize have their 2014 // least significant 2 bits clear. 2015 // NOTE: the oopMark is in swap_reg %rax, as the result of cmpxchg 2016 2017 __ subptr(swap_reg, rsp); 2018 __ andptr(swap_reg, 3 - os::vm_page_size()); 2019 2020 // Save the test result, for recursive case, the result is zero 2021 __ movptr(Address(lock_reg, mark_word_offset), swap_reg); 2022 __ jcc(Assembler::notEqual, slow_path_lock); 2023 // Slow path will re-enter here 2024 __ bind(lock_done); 2025 2026 if (UseBiasedLocking) { 2027 // Re-fetch oop_handle_reg as we trashed it above 2028 __ movptr(oop_handle_reg, Address(rsp, wordSize)); 2029 } 2030 } 2031 2032 2033 // Finally just about ready to make the JNI call 2034 2035 2036 // get JNIEnv* which is first argument to native 2037 if (!is_critical_native) { 2038 __ lea(rdx, Address(thread, in_bytes(JavaThread::jni_environment_offset()))); 2039 __ movptr(Address(rsp, 0), rdx); 2040 } 2041 2042 // Now set thread in native 2043 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native); 2044 2045 __ call(RuntimeAddress(native_func)); 2046 2047 // Verify or restore cpu control state after JNI call 2048 __ restore_cpu_control_state_after_jni(); 2049 2050 // WARNING - on Windows Java Natives use pascal calling convention and pop the 2051 // arguments off of the stack. We could just re-adjust the stack pointer here 2052 // and continue to do SP relative addressing but we instead switch to FP 2053 // relative addressing. 2054 2055 // Unpack native results. 2056 switch (ret_type) { 2057 case T_BOOLEAN: __ c2bool(rax); break; 2058 case T_CHAR : __ andptr(rax, 0xFFFF); break; 2059 case T_BYTE : __ sign_extend_byte (rax); break; 2060 case T_SHORT : __ sign_extend_short(rax); break; 2061 case T_INT : /* nothing to do */ break; 2062 case T_DOUBLE : 2063 case T_FLOAT : 2064 // Result is in st0 we'll save as needed 2065 break; 2066 case T_ARRAY: // Really a handle 2067 case T_OBJECT: // Really a handle 2068 break; // can't de-handlize until after safepoint check 2069 case T_VOID: break; 2070 case T_LONG: break; 2071 default : ShouldNotReachHere(); 2072 } 2073 2074 // Switch thread to "native transition" state before reading the synchronization state. 2075 // This additional state is necessary because reading and testing the synchronization 2076 // state is not atomic w.r.t. GC, as this scenario demonstrates: 2077 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 2078 // VM thread changes sync state to synchronizing and suspends threads for GC. 2079 // Thread A is resumed to finish this native method, but doesn't block here since it 2080 // didn't see any synchronization is progress, and escapes. 2081 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native_trans); 2082 2083 if(os::is_MP()) { 2084 if (UseMembar) { 2085 // Force this write out before the read below 2086 __ membar(Assembler::Membar_mask_bits( 2087 Assembler::LoadLoad | Assembler::LoadStore | 2088 Assembler::StoreLoad | Assembler::StoreStore)); 2089 } else { 2090 // Write serialization page so VM thread can do a pseudo remote membar. 2091 // We use the current thread pointer to calculate a thread specific 2092 // offset to write to within the page. This minimizes bus traffic 2093 // due to cache line collision. 2094 __ serialize_memory(thread, rcx); 2095 } 2096 } 2097 2098 if (AlwaysRestoreFPU) { 2099 // Make sure the control word is correct. 2100 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std())); 2101 } 2102 2103 Label after_transition; 2104 2105 // check for safepoint operation in progress and/or pending suspend requests 2106 { Label Continue; 2107 2108 __ cmp32(ExternalAddress((address)SafepointSynchronize::address_of_state()), 2109 SafepointSynchronize::_not_synchronized); 2110 2111 Label L; 2112 __ jcc(Assembler::notEqual, L); 2113 __ cmpl(Address(thread, JavaThread::suspend_flags_offset()), 0); 2114 __ jcc(Assembler::equal, Continue); 2115 __ bind(L); 2116 2117 // Don't use call_VM as it will see a possible pending exception and forward it 2118 // and never return here preventing us from clearing _last_native_pc down below. 2119 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are 2120 // preserved and correspond to the bcp/locals pointers. So we do a runtime call 2121 // by hand. 2122 // 2123 save_native_result(masm, ret_type, stack_slots); 2124 __ push(thread); 2125 if (!is_critical_native) { 2126 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, 2127 JavaThread::check_special_condition_for_native_trans))); 2128 } else { 2129 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, 2130 JavaThread::check_special_condition_for_native_trans_and_transition))); 2131 } 2132 __ increment(rsp, wordSize); 2133 // Restore any method result value 2134 restore_native_result(masm, ret_type, stack_slots); 2135 2136 if (is_critical_native) { 2137 // The call above performed the transition to thread_in_Java so 2138 // skip the transition logic below. 2139 __ jmpb(after_transition); 2140 } 2141 2142 __ bind(Continue); 2143 } 2144 2145 // change thread state 2146 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_Java); 2147 __ bind(after_transition); 2148 2149 Label reguard; 2150 Label reguard_done; 2151 __ cmpl(Address(thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_reserved_disabled); 2152 __ jcc(Assembler::equal, reguard); 2153 2154 // slow path reguard re-enters here 2155 __ bind(reguard_done); 2156 2157 // Handle possible exception (will unlock if necessary) 2158 2159 // native result if any is live 2160 2161 // Unlock 2162 Label slow_path_unlock; 2163 Label unlock_done; 2164 if (method->is_synchronized()) { 2165 2166 Label done; 2167 2168 // Get locked oop from the handle we passed to jni 2169 __ movptr(obj_reg, Address(oop_handle_reg, 0)); 2170 2171 if (UseBiasedLocking) { 2172 __ biased_locking_exit(obj_reg, rbx, done); 2173 } 2174 2175 // Simple recursive lock? 2176 2177 __ cmpptr(Address(rbp, lock_slot_rbp_offset), (int32_t)NULL_WORD); 2178 __ jcc(Assembler::equal, done); 2179 2180 // Must save rax, if if it is live now because cmpxchg must use it 2181 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2182 save_native_result(masm, ret_type, stack_slots); 2183 } 2184 2185 // get old displaced header 2186 __ movptr(rbx, Address(rbp, lock_slot_rbp_offset)); 2187 2188 // get address of the stack lock 2189 __ lea(rax, Address(rbp, lock_slot_rbp_offset)); 2190 2191 // Atomic swap old header if oop still contains the stack lock 2192 if (os::is_MP()) { 2193 __ lock(); 2194 } 2195 2196 // src -> dest iff dest == rax, else rax, <- dest 2197 // *obj_reg = rbx, iff *obj_reg == rax, else rax, = *(obj_reg) 2198 __ cmpxchgptr(rbx, Address(obj_reg, 0)); 2199 __ jcc(Assembler::notEqual, slow_path_unlock); 2200 2201 // slow path re-enters here 2202 __ bind(unlock_done); 2203 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2204 restore_native_result(masm, ret_type, stack_slots); 2205 } 2206 2207 __ bind(done); 2208 2209 } 2210 2211 { 2212 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0); 2213 // Tell dtrace about this method exit 2214 save_native_result(masm, ret_type, stack_slots); 2215 __ mov_metadata(rax, method()); 2216 __ call_VM_leaf( 2217 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2218 thread, rax); 2219 restore_native_result(masm, ret_type, stack_slots); 2220 } 2221 2222 // We can finally stop using that last_Java_frame we setup ages ago 2223 2224 __ reset_last_Java_frame(thread, false, true); 2225 2226 // Unpack oop result 2227 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2228 Label L; 2229 __ cmpptr(rax, (int32_t)NULL_WORD); 2230 __ jcc(Assembler::equal, L); 2231 __ movptr(rax, Address(rax, 0)); 2232 __ bind(L); 2233 __ verify_oop(rax); 2234 } 2235 2236 if (!is_critical_native) { 2237 // reset handle block 2238 __ movptr(rcx, Address(thread, JavaThread::active_handles_offset())); 2239 __ movl(Address(rcx, JNIHandleBlock::top_offset_in_bytes()), NULL_WORD); 2240 2241 // Any exception pending? 2242 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2243 __ jcc(Assembler::notEqual, exception_pending); 2244 } 2245 2246 // no exception, we're almost done 2247 2248 // check that only result value is on FPU stack 2249 __ verify_FPU(ret_type == T_FLOAT || ret_type == T_DOUBLE ? 1 : 0, "native_wrapper normal exit"); 2250 2251 // Fixup floating pointer results so that result looks like a return from a compiled method 2252 if (ret_type == T_FLOAT) { 2253 if (UseSSE >= 1) { 2254 // Pop st0 and store as float and reload into xmm register 2255 __ fstp_s(Address(rbp, -4)); 2256 __ movflt(xmm0, Address(rbp, -4)); 2257 } 2258 } else if (ret_type == T_DOUBLE) { 2259 if (UseSSE >= 2) { 2260 // Pop st0 and store as double and reload into xmm register 2261 __ fstp_d(Address(rbp, -8)); 2262 __ movdbl(xmm0, Address(rbp, -8)); 2263 } 2264 } 2265 2266 // Return 2267 2268 __ leave(); 2269 __ ret(0); 2270 2271 // Unexpected paths are out of line and go here 2272 2273 // Slow path locking & unlocking 2274 if (method->is_synchronized()) { 2275 2276 // BEGIN Slow path lock 2277 2278 __ bind(slow_path_lock); 2279 2280 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 2281 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2282 __ push(thread); 2283 __ push(lock_reg); 2284 __ push(obj_reg); 2285 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C))); 2286 __ addptr(rsp, 3*wordSize); 2287 2288 #ifdef ASSERT 2289 { Label L; 2290 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int)NULL_WORD); 2291 __ jcc(Assembler::equal, L); 2292 __ stop("no pending exception allowed on exit from monitorenter"); 2293 __ bind(L); 2294 } 2295 #endif 2296 __ jmp(lock_done); 2297 2298 // END Slow path lock 2299 2300 // BEGIN Slow path unlock 2301 __ bind(slow_path_unlock); 2302 2303 // Slow path unlock 2304 2305 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2306 save_native_result(masm, ret_type, stack_slots); 2307 } 2308 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 2309 2310 __ pushptr(Address(thread, in_bytes(Thread::pending_exception_offset()))); 2311 __ movptr(Address(thread, in_bytes(Thread::pending_exception_offset())), NULL_WORD); 2312 2313 2314 // should be a peal 2315 // +wordSize because of the push above 2316 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2317 __ push(thread); 2318 __ lea(rax, Address(rbp, lock_slot_rbp_offset)); 2319 __ push(rax); 2320 2321 __ push(obj_reg); 2322 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C))); 2323 __ addptr(rsp, 3*wordSize); 2324 #ifdef ASSERT 2325 { 2326 Label L; 2327 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2328 __ jcc(Assembler::equal, L); 2329 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 2330 __ bind(L); 2331 } 2332 #endif /* ASSERT */ 2333 2334 __ popptr(Address(thread, in_bytes(Thread::pending_exception_offset()))); 2335 2336 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2337 restore_native_result(masm, ret_type, stack_slots); 2338 } 2339 __ jmp(unlock_done); 2340 // END Slow path unlock 2341 2342 } 2343 2344 // SLOW PATH Reguard the stack if needed 2345 2346 __ bind(reguard); 2347 save_native_result(masm, ret_type, stack_slots); 2348 { 2349 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages))); 2350 } 2351 restore_native_result(masm, ret_type, stack_slots); 2352 __ jmp(reguard_done); 2353 2354 2355 // BEGIN EXCEPTION PROCESSING 2356 2357 if (!is_critical_native) { 2358 // Forward the exception 2359 __ bind(exception_pending); 2360 2361 // remove possible return value from FPU register stack 2362 __ empty_FPU_stack(); 2363 2364 // pop our frame 2365 __ leave(); 2366 // and forward the exception 2367 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2368 } 2369 2370 __ flush(); 2371 2372 nmethod *nm = nmethod::new_native_nmethod(method, 2373 compile_id, 2374 masm->code(), 2375 vep_offset, 2376 frame_complete, 2377 stack_slots / VMRegImpl::slots_per_word, 2378 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2379 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2380 oop_maps); 2381 2382 if (is_critical_native) { 2383 nm->set_lazy_critical_native(true); 2384 } 2385 2386 return nm; 2387 2388 } 2389 2390 // this function returns the adjust size (in number of words) to a c2i adapter 2391 // activation for use during deoptimization 2392 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals ) { 2393 return (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2394 } 2395 2396 2397 uint SharedRuntime::out_preserve_stack_slots() { 2398 return 0; 2399 } 2400 2401 //------------------------------generate_deopt_blob---------------------------- 2402 void SharedRuntime::generate_deopt_blob() { 2403 // allocate space for the code 2404 ResourceMark rm; 2405 // setup code generation tools 2406 // note: the buffer code size must account for StackShadowPages=50 2407 CodeBuffer buffer("deopt_blob", 1536, 1024); 2408 MacroAssembler* masm = new MacroAssembler(&buffer); 2409 int frame_size_in_words; 2410 OopMap* map = NULL; 2411 // Account for the extra args we place on the stack 2412 // by the time we call fetch_unroll_info 2413 const int additional_words = 2; // deopt kind, thread 2414 2415 OopMapSet *oop_maps = new OopMapSet(); 2416 2417 // ------------- 2418 // This code enters when returning to a de-optimized nmethod. A return 2419 // address has been pushed on the the stack, and return values are in 2420 // registers. 2421 // If we are doing a normal deopt then we were called from the patched 2422 // nmethod from the point we returned to the nmethod. So the return 2423 // address on the stack is wrong by NativeCall::instruction_size 2424 // We will adjust the value to it looks like we have the original return 2425 // address on the stack (like when we eagerly deoptimized). 2426 // In the case of an exception pending with deoptimized then we enter 2427 // with a return address on the stack that points after the call we patched 2428 // into the exception handler. We have the following register state: 2429 // rax,: exception 2430 // rbx,: exception handler 2431 // rdx: throwing pc 2432 // So in this case we simply jam rdx into the useless return address and 2433 // the stack looks just like we want. 2434 // 2435 // At this point we need to de-opt. We save the argument return 2436 // registers. We call the first C routine, fetch_unroll_info(). This 2437 // routine captures the return values and returns a structure which 2438 // describes the current frame size and the sizes of all replacement frames. 2439 // The current frame is compiled code and may contain many inlined 2440 // functions, each with their own JVM state. We pop the current frame, then 2441 // push all the new frames. Then we call the C routine unpack_frames() to 2442 // populate these frames. Finally unpack_frames() returns us the new target 2443 // address. Notice that callee-save registers are BLOWN here; they have 2444 // already been captured in the vframeArray at the time the return PC was 2445 // patched. 2446 address start = __ pc(); 2447 Label cont; 2448 2449 // Prolog for non exception case! 2450 2451 // Save everything in sight. 2452 2453 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false); 2454 // Normal deoptimization 2455 __ push(Deoptimization::Unpack_deopt); 2456 __ jmp(cont); 2457 2458 int reexecute_offset = __ pc() - start; 2459 2460 // Reexecute case 2461 // return address is the pc describes what bci to do re-execute at 2462 2463 // No need to update map as each call to save_live_registers will produce identical oopmap 2464 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false); 2465 2466 __ push(Deoptimization::Unpack_reexecute); 2467 __ jmp(cont); 2468 2469 int exception_offset = __ pc() - start; 2470 2471 // Prolog for exception case 2472 2473 // all registers are dead at this entry point, except for rax, and 2474 // rdx which contain the exception oop and exception pc 2475 // respectively. Set them in TLS and fall thru to the 2476 // unpack_with_exception_in_tls entry point. 2477 2478 __ get_thread(rdi); 2479 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), rdx); 2480 __ movptr(Address(rdi, JavaThread::exception_oop_offset()), rax); 2481 2482 int exception_in_tls_offset = __ pc() - start; 2483 2484 // new implementation because exception oop is now passed in JavaThread 2485 2486 // Prolog for exception case 2487 // All registers must be preserved because they might be used by LinearScan 2488 // Exceptiop oop and throwing PC are passed in JavaThread 2489 // tos: stack at point of call to method that threw the exception (i.e. only 2490 // args are on the stack, no return address) 2491 2492 // make room on stack for the return address 2493 // It will be patched later with the throwing pc. The correct value is not 2494 // available now because loading it from memory would destroy registers. 2495 __ push(0); 2496 2497 // Save everything in sight. 2498 2499 // No need to update map as each call to save_live_registers will produce identical oopmap 2500 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false); 2501 2502 // Now it is safe to overwrite any register 2503 2504 // store the correct deoptimization type 2505 __ push(Deoptimization::Unpack_exception); 2506 2507 // load throwing pc from JavaThread and patch it as the return address 2508 // of the current frame. Then clear the field in JavaThread 2509 __ get_thread(rdi); 2510 __ movptr(rdx, Address(rdi, JavaThread::exception_pc_offset())); 2511 __ movptr(Address(rbp, wordSize), rdx); 2512 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), NULL_WORD); 2513 2514 #ifdef ASSERT 2515 // verify that there is really an exception oop in JavaThread 2516 __ movptr(rax, Address(rdi, JavaThread::exception_oop_offset())); 2517 __ verify_oop(rax); 2518 2519 // verify that there is no pending exception 2520 Label no_pending_exception; 2521 __ movptr(rax, Address(rdi, Thread::pending_exception_offset())); 2522 __ testptr(rax, rax); 2523 __ jcc(Assembler::zero, no_pending_exception); 2524 __ stop("must not have pending exception here"); 2525 __ bind(no_pending_exception); 2526 #endif 2527 2528 __ bind(cont); 2529 2530 // Compiled code leaves the floating point stack dirty, empty it. 2531 __ empty_FPU_stack(); 2532 2533 2534 // Call C code. Need thread and this frame, but NOT official VM entry 2535 // crud. We cannot block on this call, no GC can happen. 2536 __ get_thread(rcx); 2537 __ push(rcx); 2538 // fetch_unroll_info needs to call last_java_frame() 2539 __ set_last_Java_frame(rcx, noreg, noreg, NULL); 2540 2541 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info))); 2542 2543 // Need to have an oopmap that tells fetch_unroll_info where to 2544 // find any register it might need. 2545 2546 oop_maps->add_gc_map( __ pc()-start, map); 2547 2548 // Discard args to fetch_unroll_info 2549 __ pop(rcx); 2550 __ pop(rcx); 2551 2552 __ get_thread(rcx); 2553 __ reset_last_Java_frame(rcx, false, false); 2554 2555 // Load UnrollBlock into EDI 2556 __ mov(rdi, rax); 2557 2558 // Move the unpack kind to a safe place in the UnrollBlock because 2559 // we are very short of registers 2560 2561 Address unpack_kind(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()); 2562 // retrieve the deopt kind from the UnrollBlock. 2563 __ movl(rax, unpack_kind); 2564 2565 Label noException; 2566 __ cmpl(rax, Deoptimization::Unpack_exception); // Was exception pending? 2567 __ jcc(Assembler::notEqual, noException); 2568 __ movptr(rax, Address(rcx, JavaThread::exception_oop_offset())); 2569 __ movptr(rdx, Address(rcx, JavaThread::exception_pc_offset())); 2570 __ movptr(Address(rcx, JavaThread::exception_oop_offset()), NULL_WORD); 2571 __ movptr(Address(rcx, JavaThread::exception_pc_offset()), NULL_WORD); 2572 2573 __ verify_oop(rax); 2574 2575 // Overwrite the result registers with the exception results. 2576 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax); 2577 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx); 2578 2579 __ bind(noException); 2580 2581 // Stack is back to only having register save data on the stack. 2582 // Now restore the result registers. Everything else is either dead or captured 2583 // in the vframeArray. 2584 2585 RegisterSaver::restore_result_registers(masm); 2586 2587 // Non standard control word may be leaked out through a safepoint blob, and we can 2588 // deopt at a poll point with the non standard control word. However, we should make 2589 // sure the control word is correct after restore_result_registers. 2590 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std())); 2591 2592 // All of the register save area has been popped of the stack. Only the 2593 // return address remains. 2594 2595 // Pop all the frames we must move/replace. 2596 // 2597 // Frame picture (youngest to oldest) 2598 // 1: self-frame (no frame link) 2599 // 2: deopting frame (no frame link) 2600 // 3: caller of deopting frame (could be compiled/interpreted). 2601 // 2602 // Note: by leaving the return address of self-frame on the stack 2603 // and using the size of frame 2 to adjust the stack 2604 // when we are done the return to frame 3 will still be on the stack. 2605 2606 // Pop deoptimized frame 2607 __ addptr(rsp, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 2608 2609 // sp should be pointing at the return address to the caller (3) 2610 2611 // Pick up the initial fp we should save 2612 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved) 2613 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 2614 2615 #ifdef ASSERT 2616 // Compilers generate code that bang the stack by as much as the 2617 // interpreter would need. So this stack banging should never 2618 // trigger a fault. Verify that it does not on non product builds. 2619 if (UseStackBanging) { 2620 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 2621 __ bang_stack_size(rbx, rcx); 2622 } 2623 #endif 2624 2625 // Load array of frame pcs into ECX 2626 __ movptr(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2627 2628 __ pop(rsi); // trash the old pc 2629 2630 // Load array of frame sizes into ESI 2631 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 2632 2633 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes()); 2634 2635 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 2636 __ movl(counter, rbx); 2637 2638 // Now adjust the caller's stack to make up for the extra locals 2639 // but record the original sp so that we can save it in the skeletal interpreter 2640 // frame and the stack walking of interpreter_sender will get the unextended sp 2641 // value and not the "real" sp value. 2642 2643 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes()); 2644 __ movptr(sp_temp, rsp); 2645 __ movl2ptr(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes())); 2646 __ subptr(rsp, rbx); 2647 2648 // Push interpreter frames in a loop 2649 Label loop; 2650 __ bind(loop); 2651 __ movptr(rbx, Address(rsi, 0)); // Load frame size 2652 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand 2653 __ pushptr(Address(rcx, 0)); // save return address 2654 __ enter(); // save old & set new rbp, 2655 __ subptr(rsp, rbx); // Prolog! 2656 __ movptr(rbx, sp_temp); // sender's sp 2657 // This value is corrected by layout_activation_impl 2658 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD); 2659 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable 2660 __ movptr(sp_temp, rsp); // pass to next frame 2661 __ addptr(rsi, wordSize); // Bump array pointer (sizes) 2662 __ addptr(rcx, wordSize); // Bump array pointer (pcs) 2663 __ decrementl(counter); // decrement counter 2664 __ jcc(Assembler::notZero, loop); 2665 __ pushptr(Address(rcx, 0)); // save final return address 2666 2667 // Re-push self-frame 2668 __ enter(); // save old & set new rbp, 2669 2670 // Return address and rbp, are in place 2671 // We'll push additional args later. Just allocate a full sized 2672 // register save area 2673 __ subptr(rsp, (frame_size_in_words-additional_words - 2) * wordSize); 2674 2675 // Restore frame locals after moving the frame 2676 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax); 2677 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx); 2678 __ fstp_d(Address(rsp, RegisterSaver::fpResultOffset()*wordSize)); // Pop float stack and store in local 2679 if( UseSSE>=2 ) __ movdbl(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0); 2680 if( UseSSE==1 ) __ movflt(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0); 2681 2682 // Set up the args to unpack_frame 2683 2684 __ pushl(unpack_kind); // get the unpack_kind value 2685 __ get_thread(rcx); 2686 __ push(rcx); 2687 2688 // set last_Java_sp, last_Java_fp 2689 __ set_last_Java_frame(rcx, noreg, rbp, NULL); 2690 2691 // Call C code. Need thread but NOT official VM entry 2692 // crud. We cannot block on this call, no GC can happen. Call should 2693 // restore return values to their stack-slots with the new SP. 2694 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2695 // Set an oopmap for the call site 2696 oop_maps->add_gc_map( __ pc()-start, new OopMap( frame_size_in_words, 0 )); 2697 2698 // rax, contains the return result type 2699 __ push(rax); 2700 2701 __ get_thread(rcx); 2702 __ reset_last_Java_frame(rcx, false, false); 2703 2704 // Collect return values 2705 __ movptr(rax,Address(rsp, (RegisterSaver::raxOffset() + additional_words + 1)*wordSize)); 2706 __ movptr(rdx,Address(rsp, (RegisterSaver::rdxOffset() + additional_words + 1)*wordSize)); 2707 2708 // Clear floating point stack before returning to interpreter 2709 __ empty_FPU_stack(); 2710 2711 // Check if we should push the float or double return value. 2712 Label results_done, yes_double_value; 2713 __ cmpl(Address(rsp, 0), T_DOUBLE); 2714 __ jcc (Assembler::zero, yes_double_value); 2715 __ cmpl(Address(rsp, 0), T_FLOAT); 2716 __ jcc (Assembler::notZero, results_done); 2717 2718 // return float value as expected by interpreter 2719 if( UseSSE>=1 ) __ movflt(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize)); 2720 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize)); 2721 __ jmp(results_done); 2722 2723 // return double value as expected by interpreter 2724 __ bind(yes_double_value); 2725 if( UseSSE>=2 ) __ movdbl(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize)); 2726 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize)); 2727 2728 __ bind(results_done); 2729 2730 // Pop self-frame. 2731 __ leave(); // Epilog! 2732 2733 // Jump to interpreter 2734 __ ret(0); 2735 2736 // ------------- 2737 // make sure all code is generated 2738 masm->flush(); 2739 2740 _deopt_blob = DeoptimizationBlob::create( &buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 2741 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 2742 } 2743 2744 2745 #ifdef COMPILER2 2746 //------------------------------generate_uncommon_trap_blob-------------------- 2747 void SharedRuntime::generate_uncommon_trap_blob() { 2748 // allocate space for the code 2749 ResourceMark rm; 2750 // setup code generation tools 2751 CodeBuffer buffer("uncommon_trap_blob", 512, 512); 2752 MacroAssembler* masm = new MacroAssembler(&buffer); 2753 2754 enum frame_layout { 2755 arg0_off, // thread sp + 0 // Arg location for 2756 arg1_off, // unloaded_class_index sp + 1 // calling C 2757 arg2_off, // exec_mode sp + 2 2758 // The frame sender code expects that rbp will be in the "natural" place and 2759 // will override any oopMap setting for it. We must therefore force the layout 2760 // so that it agrees with the frame sender code. 2761 rbp_off, // callee saved register sp + 3 2762 return_off, // slot for return address sp + 4 2763 framesize 2764 }; 2765 2766 address start = __ pc(); 2767 2768 if (UseRTMLocking) { 2769 // Abort RTM transaction before possible nmethod deoptimization. 2770 __ xabort(0); 2771 } 2772 2773 // Push self-frame. 2774 __ subptr(rsp, return_off*wordSize); // Epilog! 2775 2776 // rbp, is an implicitly saved callee saved register (i.e. the calling 2777 // convention will save restore it in prolog/epilog) Other than that 2778 // there are no callee save registers no that adapter frames are gone. 2779 __ movptr(Address(rsp, rbp_off*wordSize), rbp); 2780 2781 // Clear the floating point exception stack 2782 __ empty_FPU_stack(); 2783 2784 // set last_Java_sp 2785 __ get_thread(rdx); 2786 __ set_last_Java_frame(rdx, noreg, noreg, NULL); 2787 2788 // Call C code. Need thread but NOT official VM entry 2789 // crud. We cannot block on this call, no GC can happen. Call should 2790 // capture callee-saved registers as well as return values. 2791 __ movptr(Address(rsp, arg0_off*wordSize), rdx); 2792 // argument already in ECX 2793 __ movl(Address(rsp, arg1_off*wordSize),rcx); 2794 __ movl(Address(rsp, arg2_off*wordSize), Deoptimization::Unpack_uncommon_trap); 2795 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap))); 2796 2797 // Set an oopmap for the call site 2798 OopMapSet *oop_maps = new OopMapSet(); 2799 OopMap* map = new OopMap( framesize, 0 ); 2800 // No oopMap for rbp, it is known implicitly 2801 2802 oop_maps->add_gc_map( __ pc()-start, map); 2803 2804 __ get_thread(rcx); 2805 2806 __ reset_last_Java_frame(rcx, false, false); 2807 2808 // Load UnrollBlock into EDI 2809 __ movptr(rdi, rax); 2810 2811 #ifdef ASSERT 2812 { Label L; 2813 __ cmpptr(Address(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()), 2814 (int32_t)Deoptimization::Unpack_uncommon_trap); 2815 __ jcc(Assembler::equal, L); 2816 __ stop("SharedRuntime::generate_deopt_blob: expected Unpack_uncommon_trap"); 2817 __ bind(L); 2818 } 2819 #endif 2820 2821 // Pop all the frames we must move/replace. 2822 // 2823 // Frame picture (youngest to oldest) 2824 // 1: self-frame (no frame link) 2825 // 2: deopting frame (no frame link) 2826 // 3: caller of deopting frame (could be compiled/interpreted). 2827 2828 // Pop self-frame. We have no frame, and must rely only on EAX and ESP. 2829 __ addptr(rsp,(framesize-1)*wordSize); // Epilog! 2830 2831 // Pop deoptimized frame 2832 __ movl2ptr(rcx, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 2833 __ addptr(rsp, rcx); 2834 2835 // sp should be pointing at the return address to the caller (3) 2836 2837 // Pick up the initial fp we should save 2838 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved) 2839 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 2840 2841 #ifdef ASSERT 2842 // Compilers generate code that bang the stack by as much as the 2843 // interpreter would need. So this stack banging should never 2844 // trigger a fault. Verify that it does not on non product builds. 2845 if (UseStackBanging) { 2846 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 2847 __ bang_stack_size(rbx, rcx); 2848 } 2849 #endif 2850 2851 // Load array of frame pcs into ECX 2852 __ movl(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 2853 2854 __ pop(rsi); // trash the pc 2855 2856 // Load array of frame sizes into ESI 2857 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 2858 2859 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes()); 2860 2861 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 2862 __ movl(counter, rbx); 2863 2864 // Now adjust the caller's stack to make up for the extra locals 2865 // but record the original sp so that we can save it in the skeletal interpreter 2866 // frame and the stack walking of interpreter_sender will get the unextended sp 2867 // value and not the "real" sp value. 2868 2869 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes()); 2870 __ movptr(sp_temp, rsp); 2871 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes())); 2872 __ subptr(rsp, rbx); 2873 2874 // Push interpreter frames in a loop 2875 Label loop; 2876 __ bind(loop); 2877 __ movptr(rbx, Address(rsi, 0)); // Load frame size 2878 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand 2879 __ pushptr(Address(rcx, 0)); // save return address 2880 __ enter(); // save old & set new rbp, 2881 __ subptr(rsp, rbx); // Prolog! 2882 __ movptr(rbx, sp_temp); // sender's sp 2883 // This value is corrected by layout_activation_impl 2884 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD ); 2885 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable 2886 __ movptr(sp_temp, rsp); // pass to next frame 2887 __ addptr(rsi, wordSize); // Bump array pointer (sizes) 2888 __ addptr(rcx, wordSize); // Bump array pointer (pcs) 2889 __ decrementl(counter); // decrement counter 2890 __ jcc(Assembler::notZero, loop); 2891 __ pushptr(Address(rcx, 0)); // save final return address 2892 2893 // Re-push self-frame 2894 __ enter(); // save old & set new rbp, 2895 __ subptr(rsp, (framesize-2) * wordSize); // Prolog! 2896 2897 2898 // set last_Java_sp, last_Java_fp 2899 __ get_thread(rdi); 2900 __ set_last_Java_frame(rdi, noreg, rbp, NULL); 2901 2902 // Call C code. Need thread but NOT official VM entry 2903 // crud. We cannot block on this call, no GC can happen. Call should 2904 // restore return values to their stack-slots with the new SP. 2905 __ movptr(Address(rsp,arg0_off*wordSize),rdi); 2906 __ movl(Address(rsp,arg1_off*wordSize), Deoptimization::Unpack_uncommon_trap); 2907 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 2908 // Set an oopmap for the call site 2909 oop_maps->add_gc_map( __ pc()-start, new OopMap( framesize, 0 ) ); 2910 2911 __ get_thread(rdi); 2912 __ reset_last_Java_frame(rdi, true, false); 2913 2914 // Pop self-frame. 2915 __ leave(); // Epilog! 2916 2917 // Jump to interpreter 2918 __ ret(0); 2919 2920 // ------------- 2921 // make sure all code is generated 2922 masm->flush(); 2923 2924 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, framesize); 2925 } 2926 #endif // COMPILER2 2927 2928 //------------------------------generate_handler_blob------ 2929 // 2930 // Generate a special Compile2Runtime blob that saves all registers, 2931 // setup oopmap, and calls safepoint code to stop the compiled code for 2932 // a safepoint. 2933 // 2934 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 2935 2936 // Account for thread arg in our frame 2937 const int additional_words = 1; 2938 int frame_size_in_words; 2939 2940 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 2941 2942 ResourceMark rm; 2943 OopMapSet *oop_maps = new OopMapSet(); 2944 OopMap* map; 2945 2946 // allocate space for the code 2947 // setup code generation tools 2948 CodeBuffer buffer("handler_blob", 1024, 512); 2949 MacroAssembler* masm = new MacroAssembler(&buffer); 2950 2951 const Register java_thread = rdi; // callee-saved for VC++ 2952 address start = __ pc(); 2953 address call_pc = NULL; 2954 bool cause_return = (poll_type == POLL_AT_RETURN); 2955 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP); 2956 2957 if (UseRTMLocking) { 2958 // Abort RTM transaction before calling runtime 2959 // because critical section will be large and will be 2960 // aborted anyway. Also nmethod could be deoptimized. 2961 __ xabort(0); 2962 } 2963 2964 // If cause_return is true we are at a poll_return and there is 2965 // the return address on the stack to the caller on the nmethod 2966 // that is safepoint. We can leave this return on the stack and 2967 // effectively complete the return and safepoint in the caller. 2968 // Otherwise we push space for a return address that the safepoint 2969 // handler will install later to make the stack walking sensible. 2970 if (!cause_return) 2971 __ push(rbx); // Make room for return address (or push it again) 2972 2973 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false, save_vectors); 2974 2975 // The following is basically a call_VM. However, we need the precise 2976 // address of the call in order to generate an oopmap. Hence, we do all the 2977 // work ourselves. 2978 2979 // Push thread argument and setup last_Java_sp 2980 __ get_thread(java_thread); 2981 __ push(java_thread); 2982 __ set_last_Java_frame(java_thread, noreg, noreg, NULL); 2983 2984 // if this was not a poll_return then we need to correct the return address now. 2985 if (!cause_return) { 2986 __ movptr(rax, Address(java_thread, JavaThread::saved_exception_pc_offset())); 2987 __ movptr(Address(rbp, wordSize), rax); 2988 } 2989 2990 // do the call 2991 __ call(RuntimeAddress(call_ptr)); 2992 2993 // Set an oopmap for the call site. This oopmap will map all 2994 // oop-registers and debug-info registers as callee-saved. This 2995 // will allow deoptimization at this safepoint to find all possible 2996 // debug-info recordings, as well as let GC find all oops. 2997 2998 oop_maps->add_gc_map( __ pc() - start, map); 2999 3000 // Discard arg 3001 __ pop(rcx); 3002 3003 Label noException; 3004 3005 // Clear last_Java_sp again 3006 __ get_thread(java_thread); 3007 __ reset_last_Java_frame(java_thread, false, false); 3008 3009 __ cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 3010 __ jcc(Assembler::equal, noException); 3011 3012 // Exception pending 3013 RegisterSaver::restore_live_registers(masm, save_vectors); 3014 3015 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3016 3017 __ bind(noException); 3018 3019 // Normal exit, register restoring and exit 3020 RegisterSaver::restore_live_registers(masm, save_vectors); 3021 3022 __ ret(0); 3023 3024 // make sure all code is generated 3025 masm->flush(); 3026 3027 // Fill-out other meta info 3028 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 3029 } 3030 3031 // 3032 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 3033 // 3034 // Generate a stub that calls into vm to find out the proper destination 3035 // of a java call. All the argument registers are live at this point 3036 // but since this is generic code we don't know what they are and the caller 3037 // must do any gc of the args. 3038 // 3039 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3040 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 3041 3042 // allocate space for the code 3043 ResourceMark rm; 3044 3045 CodeBuffer buffer(name, 1000, 512); 3046 MacroAssembler* masm = new MacroAssembler(&buffer); 3047 3048 int frame_size_words; 3049 enum frame_layout { 3050 thread_off, 3051 extra_words }; 3052 3053 OopMapSet *oop_maps = new OopMapSet(); 3054 OopMap* map = NULL; 3055 3056 int start = __ offset(); 3057 3058 map = RegisterSaver::save_live_registers(masm, extra_words, &frame_size_words); 3059 3060 int frame_complete = __ offset(); 3061 3062 const Register thread = rdi; 3063 __ get_thread(rdi); 3064 3065 __ push(thread); 3066 __ set_last_Java_frame(thread, noreg, rbp, NULL); 3067 3068 __ call(RuntimeAddress(destination)); 3069 3070 3071 // Set an oopmap for the call site. 3072 // We need this not only for callee-saved registers, but also for volatile 3073 // registers that the compiler might be keeping live across a safepoint. 3074 3075 oop_maps->add_gc_map( __ offset() - start, map); 3076 3077 // rax, contains the address we are going to jump to assuming no exception got installed 3078 3079 __ addptr(rsp, wordSize); 3080 3081 // clear last_Java_sp 3082 __ reset_last_Java_frame(thread, true, false); 3083 // check for pending exceptions 3084 Label pending; 3085 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 3086 __ jcc(Assembler::notEqual, pending); 3087 3088 // get the returned Method* 3089 __ get_vm_result_2(rbx, thread); 3090 __ movptr(Address(rsp, RegisterSaver::rbx_offset() * wordSize), rbx); 3091 3092 __ movptr(Address(rsp, RegisterSaver::rax_offset() * wordSize), rax); 3093 3094 RegisterSaver::restore_live_registers(masm); 3095 3096 // We are back the the original state on entry and ready to go. 3097 3098 __ jmp(rax); 3099 3100 // Pending exception after the safepoint 3101 3102 __ bind(pending); 3103 3104 RegisterSaver::restore_live_registers(masm); 3105 3106 // exception pending => remove activation and forward to exception handler 3107 3108 __ get_thread(thread); 3109 __ movptr(Address(thread, JavaThread::vm_result_offset()), NULL_WORD); 3110 __ movptr(rax, Address(thread, Thread::pending_exception_offset())); 3111 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3112 3113 // ------------- 3114 // make sure all code is generated 3115 masm->flush(); 3116 3117 // return the blob 3118 // frame_size_words or bytes?? 3119 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true); 3120 }