1 /* 2 * Copyright (c) 2003, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #ifndef _WINDOWS 27 #include "alloca.h" 28 #endif 29 #include "asm/macroAssembler.hpp" 30 #include "asm/macroAssembler.inline.hpp" 31 #include "code/debugInfoRec.hpp" 32 #include "code/icBuffer.hpp" 33 #include "code/nativeInst.hpp" 34 #include "code/vtableStubs.hpp" 35 #include "gc/shared/gcLocker.hpp" 36 #include "interpreter/interpreter.hpp" 37 #include "logging/log.hpp" 38 #include "memory/resourceArea.hpp" 39 #include "oops/compiledICHolder.hpp" 40 #include "runtime/safepointMechanism.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 #include "runtime/vframeArray.hpp" 43 #include "utilities/align.hpp" 44 #include "utilities/formatBuffer.hpp" 45 #include "vm_version_x86.hpp" 46 #include "vmreg_x86.inline.hpp" 47 #ifdef COMPILER1 48 #include "c1/c1_Runtime1.hpp" 49 #endif 50 #ifdef COMPILER2 51 #include "opto/runtime.hpp" 52 #endif 53 #if INCLUDE_JVMCI 54 #include "jvmci/jvmciJavaClasses.hpp" 55 #endif 56 57 #define __ masm-> 58 59 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 60 61 class SimpleRuntimeFrame { 62 63 public: 64 65 // Most of the runtime stubs have this simple frame layout. 66 // This class exists to make the layout shared in one place. 67 // Offsets are for compiler stack slots, which are jints. 68 enum layout { 69 // The frame sender code expects that rbp will be in the "natural" place and 70 // will override any oopMap setting for it. We must therefore force the layout 71 // so that it agrees with the frame sender code. 72 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt, 73 rbp_off2, 74 return_off, return_off2, 75 framesize 76 }; 77 }; 78 79 class RegisterSaver { 80 // Capture info about frame layout. Layout offsets are in jint 81 // units because compiler frame slots are jints. 82 #define XSAVE_AREA_BEGIN 160 83 #define XSAVE_AREA_YMM_BEGIN 576 84 #define XSAVE_AREA_ZMM_BEGIN 1152 85 #define XSAVE_AREA_UPPERBANK 1664 86 #define DEF_XMM_OFFS(regnum) xmm ## regnum ## _off = xmm_off + (regnum)*16/BytesPerInt, xmm ## regnum ## H_off 87 #define DEF_YMM_OFFS(regnum) ymm ## regnum ## _off = ymm_off + (regnum)*16/BytesPerInt, ymm ## regnum ## H_off 88 #define DEF_ZMM_OFFS(regnum) zmm ## regnum ## _off = zmm_off + (regnum-16)*64/BytesPerInt, zmm ## regnum ## H_off 89 enum layout { 90 fpu_state_off = frame::arg_reg_save_area_bytes/BytesPerInt, // fxsave save area 91 xmm_off = fpu_state_off + XSAVE_AREA_BEGIN/BytesPerInt, // offset in fxsave save area 92 DEF_XMM_OFFS(0), 93 DEF_XMM_OFFS(1), 94 // 2..15 are implied in range usage 95 ymm_off = xmm_off + (XSAVE_AREA_YMM_BEGIN - XSAVE_AREA_BEGIN)/BytesPerInt, 96 DEF_YMM_OFFS(0), 97 DEF_YMM_OFFS(1), 98 // 2..15 are implied in range usage 99 zmm_high = xmm_off + (XSAVE_AREA_ZMM_BEGIN - XSAVE_AREA_BEGIN)/BytesPerInt, 100 zmm_off = xmm_off + (XSAVE_AREA_UPPERBANK - XSAVE_AREA_BEGIN)/BytesPerInt, 101 DEF_ZMM_OFFS(16), 102 DEF_ZMM_OFFS(17), 103 // 18..31 are implied in range usage 104 fpu_state_end = fpu_state_off + ((FPUStateSizeInWords-1)*wordSize / BytesPerInt), 105 fpu_stateH_end, 106 r15_off, r15H_off, 107 r14_off, r14H_off, 108 r13_off, r13H_off, 109 r12_off, r12H_off, 110 r11_off, r11H_off, 111 r10_off, r10H_off, 112 r9_off, r9H_off, 113 r8_off, r8H_off, 114 rdi_off, rdiH_off, 115 rsi_off, rsiH_off, 116 ignore_off, ignoreH_off, // extra copy of rbp 117 rsp_off, rspH_off, 118 rbx_off, rbxH_off, 119 rdx_off, rdxH_off, 120 rcx_off, rcxH_off, 121 rax_off, raxH_off, 122 // 16-byte stack alignment fill word: see MacroAssembler::push/pop_IU_state 123 align_off, alignH_off, 124 flags_off, flagsH_off, 125 // The frame sender code expects that rbp will be in the "natural" place and 126 // will override any oopMap setting for it. We must therefore force the layout 127 // so that it agrees with the frame sender code. 128 rbp_off, rbpH_off, // copy of rbp we will restore 129 return_off, returnH_off, // slot for return address 130 reg_save_size // size in compiler stack slots 131 }; 132 133 public: 134 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_vectors = false); 135 static void restore_live_registers(MacroAssembler* masm, bool restore_vectors = false); 136 137 // Offsets into the register save area 138 // Used by deoptimization when it is managing result register 139 // values on its own 140 141 static int rax_offset_in_bytes(void) { return BytesPerInt * rax_off; } 142 static int rdx_offset_in_bytes(void) { return BytesPerInt * rdx_off; } 143 static int rbx_offset_in_bytes(void) { return BytesPerInt * rbx_off; } 144 static int xmm0_offset_in_bytes(void) { return BytesPerInt * xmm0_off; } 145 static int return_offset_in_bytes(void) { return BytesPerInt * return_off; } 146 147 // During deoptimization only the result registers need to be restored, 148 // all the other values have already been extracted. 149 static void restore_result_registers(MacroAssembler* masm); 150 }; 151 152 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_vectors) { 153 int off = 0; 154 int num_xmm_regs = XMMRegisterImpl::number_of_registers; 155 if (UseAVX < 3) { 156 num_xmm_regs = num_xmm_regs/2; 157 } 158 #if COMPILER2_OR_JVMCI 159 if (save_vectors) { 160 assert(UseAVX > 0, "Vectors larger than 16 byte long are supported only with AVX"); 161 assert(MaxVectorSize <= 64, "Only up to 64 byte long vectors are supported"); 162 } 163 #else 164 assert(!save_vectors, "vectors are generated only by C2 and JVMCI"); 165 #endif 166 167 // Always make the frame size 16-byte aligned, both vector and non vector stacks are always allocated 168 int frame_size_in_bytes = align_up(reg_save_size*BytesPerInt, num_xmm_regs); 169 // OopMap frame size is in compiler stack slots (jint's) not bytes or words 170 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt; 171 // CodeBlob frame size is in words. 172 int frame_size_in_words = frame_size_in_bytes / wordSize; 173 *total_frame_words = frame_size_in_words; 174 175 // Save registers, fpu state, and flags. 176 // We assume caller has already pushed the return address onto the 177 // stack, so rsp is 8-byte aligned here. 178 // We push rpb twice in this sequence because we want the real rbp 179 // to be under the return like a normal enter. 180 181 __ enter(); // rsp becomes 16-byte aligned here 182 __ push_CPU_state(); // Push a multiple of 16 bytes 183 184 // push cpu state handles this on EVEX enabled targets 185 if (save_vectors) { 186 // Save upper half of YMM registers(0..15) 187 int base_addr = XSAVE_AREA_YMM_BEGIN; 188 for (int n = 0; n < 16; n++) { 189 __ vextractf128_high(Address(rsp, base_addr+n*16), as_XMMRegister(n)); 190 } 191 if (VM_Version::supports_evex()) { 192 // Save upper half of ZMM registers(0..15) 193 base_addr = XSAVE_AREA_ZMM_BEGIN; 194 for (int n = 0; n < 16; n++) { 195 __ vextractf64x4_high(Address(rsp, base_addr+n*32), as_XMMRegister(n)); 196 } 197 // Save full ZMM registers(16..num_xmm_regs) 198 base_addr = XSAVE_AREA_UPPERBANK; 199 off = 0; 200 int vector_len = Assembler::AVX_512bit; 201 for (int n = 16; n < num_xmm_regs; n++) { 202 __ evmovdqul(Address(rsp, base_addr+(off++*64)), as_XMMRegister(n), vector_len); 203 } 204 } 205 } else { 206 if (VM_Version::supports_evex()) { 207 // Save upper bank of ZMM registers(16..31) for double/float usage 208 int base_addr = XSAVE_AREA_UPPERBANK; 209 off = 0; 210 for (int n = 16; n < num_xmm_regs; n++) { 211 __ movsd(Address(rsp, base_addr+(off++*64)), as_XMMRegister(n)); 212 } 213 } 214 } 215 __ vzeroupper(); 216 if (frame::arg_reg_save_area_bytes != 0) { 217 // Allocate argument register save area 218 __ subptr(rsp, frame::arg_reg_save_area_bytes); 219 } 220 221 // Set an oopmap for the call site. This oopmap will map all 222 // oop-registers and debug-info registers as callee-saved. This 223 // will allow deoptimization at this safepoint to find all possible 224 // debug-info recordings, as well as let GC find all oops. 225 226 OopMapSet *oop_maps = new OopMapSet(); 227 OopMap* map = new OopMap(frame_size_in_slots, 0); 228 229 #define STACK_OFFSET(x) VMRegImpl::stack2reg((x)) 230 231 map->set_callee_saved(STACK_OFFSET( rax_off ), rax->as_VMReg()); 232 map->set_callee_saved(STACK_OFFSET( rcx_off ), rcx->as_VMReg()); 233 map->set_callee_saved(STACK_OFFSET( rdx_off ), rdx->as_VMReg()); 234 map->set_callee_saved(STACK_OFFSET( rbx_off ), rbx->as_VMReg()); 235 // rbp location is known implicitly by the frame sender code, needs no oopmap 236 // and the location where rbp was saved by is ignored 237 map->set_callee_saved(STACK_OFFSET( rsi_off ), rsi->as_VMReg()); 238 map->set_callee_saved(STACK_OFFSET( rdi_off ), rdi->as_VMReg()); 239 map->set_callee_saved(STACK_OFFSET( r8_off ), r8->as_VMReg()); 240 map->set_callee_saved(STACK_OFFSET( r9_off ), r9->as_VMReg()); 241 map->set_callee_saved(STACK_OFFSET( r10_off ), r10->as_VMReg()); 242 map->set_callee_saved(STACK_OFFSET( r11_off ), r11->as_VMReg()); 243 map->set_callee_saved(STACK_OFFSET( r12_off ), r12->as_VMReg()); 244 map->set_callee_saved(STACK_OFFSET( r13_off ), r13->as_VMReg()); 245 map->set_callee_saved(STACK_OFFSET( r14_off ), r14->as_VMReg()); 246 map->set_callee_saved(STACK_OFFSET( r15_off ), r15->as_VMReg()); 247 // For both AVX and EVEX we will use the legacy FXSAVE area for xmm0..xmm15, 248 // on EVEX enabled targets, we get it included in the xsave area 249 off = xmm0_off; 250 int delta = xmm1_off - off; 251 for (int n = 0; n < 16; n++) { 252 XMMRegister xmm_name = as_XMMRegister(n); 253 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg()); 254 off += delta; 255 } 256 if(UseAVX > 2) { 257 // Obtain xmm16..xmm31 from the XSAVE area on EVEX enabled targets 258 off = zmm16_off; 259 delta = zmm17_off - off; 260 for (int n = 16; n < num_xmm_regs; n++) { 261 XMMRegister zmm_name = as_XMMRegister(n); 262 map->set_callee_saved(STACK_OFFSET(off), zmm_name->as_VMReg()); 263 off += delta; 264 } 265 } 266 267 #if COMPILER2_OR_JVMCI 268 if (save_vectors) { 269 off = ymm0_off; 270 int delta = ymm1_off - off; 271 for (int n = 0; n < 16; n++) { 272 XMMRegister ymm_name = as_XMMRegister(n); 273 map->set_callee_saved(STACK_OFFSET(off), ymm_name->as_VMReg()->next(4)); 274 off += delta; 275 } 276 } 277 #endif // COMPILER2_OR_JVMCI 278 279 // %%% These should all be a waste but we'll keep things as they were for now 280 if (true) { 281 map->set_callee_saved(STACK_OFFSET( raxH_off ), rax->as_VMReg()->next()); 282 map->set_callee_saved(STACK_OFFSET( rcxH_off ), rcx->as_VMReg()->next()); 283 map->set_callee_saved(STACK_OFFSET( rdxH_off ), rdx->as_VMReg()->next()); 284 map->set_callee_saved(STACK_OFFSET( rbxH_off ), rbx->as_VMReg()->next()); 285 // rbp location is known implicitly by the frame sender code, needs no oopmap 286 map->set_callee_saved(STACK_OFFSET( rsiH_off ), rsi->as_VMReg()->next()); 287 map->set_callee_saved(STACK_OFFSET( rdiH_off ), rdi->as_VMReg()->next()); 288 map->set_callee_saved(STACK_OFFSET( r8H_off ), r8->as_VMReg()->next()); 289 map->set_callee_saved(STACK_OFFSET( r9H_off ), r9->as_VMReg()->next()); 290 map->set_callee_saved(STACK_OFFSET( r10H_off ), r10->as_VMReg()->next()); 291 map->set_callee_saved(STACK_OFFSET( r11H_off ), r11->as_VMReg()->next()); 292 map->set_callee_saved(STACK_OFFSET( r12H_off ), r12->as_VMReg()->next()); 293 map->set_callee_saved(STACK_OFFSET( r13H_off ), r13->as_VMReg()->next()); 294 map->set_callee_saved(STACK_OFFSET( r14H_off ), r14->as_VMReg()->next()); 295 map->set_callee_saved(STACK_OFFSET( r15H_off ), r15->as_VMReg()->next()); 296 // For both AVX and EVEX we will use the legacy FXSAVE area for xmm0..xmm15, 297 // on EVEX enabled targets, we get it included in the xsave area 298 off = xmm0H_off; 299 delta = xmm1H_off - off; 300 for (int n = 0; n < 16; n++) { 301 XMMRegister xmm_name = as_XMMRegister(n); 302 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg()->next()); 303 off += delta; 304 } 305 if (UseAVX > 2) { 306 // Obtain xmm16..xmm31 from the XSAVE area on EVEX enabled targets 307 off = zmm16H_off; 308 delta = zmm17H_off - off; 309 for (int n = 16; n < num_xmm_regs; n++) { 310 XMMRegister zmm_name = as_XMMRegister(n); 311 map->set_callee_saved(STACK_OFFSET(off), zmm_name->as_VMReg()->next()); 312 off += delta; 313 } 314 } 315 } 316 317 return map; 318 } 319 320 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_vectors) { 321 int num_xmm_regs = XMMRegisterImpl::number_of_registers; 322 if (UseAVX < 3) { 323 num_xmm_regs = num_xmm_regs/2; 324 } 325 if (frame::arg_reg_save_area_bytes != 0) { 326 // Pop arg register save area 327 __ addptr(rsp, frame::arg_reg_save_area_bytes); 328 } 329 330 #if COMPILER2_OR_JVMCI 331 if (restore_vectors) { 332 assert(UseAVX > 0, "Vectors larger than 16 byte long are supported only with AVX"); 333 assert(MaxVectorSize <= 64, "Only up to 64 byte long vectors are supported"); 334 } 335 #else 336 assert(!restore_vectors, "vectors are generated only by C2"); 337 #endif 338 339 __ vzeroupper(); 340 341 // On EVEX enabled targets everything is handled in pop fpu state 342 if (restore_vectors) { 343 // Restore upper half of YMM registers (0..15) 344 int base_addr = XSAVE_AREA_YMM_BEGIN; 345 for (int n = 0; n < 16; n++) { 346 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, base_addr+n*16)); 347 } 348 if (VM_Version::supports_evex()) { 349 // Restore upper half of ZMM registers (0..15) 350 base_addr = XSAVE_AREA_ZMM_BEGIN; 351 for (int n = 0; n < 16; n++) { 352 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, base_addr+n*32)); 353 } 354 // Restore full ZMM registers(16..num_xmm_regs) 355 base_addr = XSAVE_AREA_UPPERBANK; 356 int vector_len = Assembler::AVX_512bit; 357 int off = 0; 358 for (int n = 16; n < num_xmm_regs; n++) { 359 __ evmovdqul(as_XMMRegister(n), Address(rsp, base_addr+(off++*64)), vector_len); 360 } 361 } 362 } else { 363 if (VM_Version::supports_evex()) { 364 // Restore upper bank of ZMM registers(16..31) for double/float usage 365 int base_addr = XSAVE_AREA_UPPERBANK; 366 int off = 0; 367 for (int n = 16; n < num_xmm_regs; n++) { 368 __ movsd(as_XMMRegister(n), Address(rsp, base_addr+(off++*64))); 369 } 370 } 371 } 372 373 // Recover CPU state 374 __ pop_CPU_state(); 375 // Get the rbp described implicitly by the calling convention (no oopMap) 376 __ pop(rbp); 377 } 378 379 void RegisterSaver::restore_result_registers(MacroAssembler* masm) { 380 381 // Just restore result register. Only used by deoptimization. By 382 // now any callee save register that needs to be restored to a c2 383 // caller of the deoptee has been extracted into the vframeArray 384 // and will be stuffed into the c2i adapter we create for later 385 // restoration so only result registers need to be restored here. 386 387 // Restore fp result register 388 __ movdbl(xmm0, Address(rsp, xmm0_offset_in_bytes())); 389 // Restore integer result register 390 __ movptr(rax, Address(rsp, rax_offset_in_bytes())); 391 __ movptr(rdx, Address(rsp, rdx_offset_in_bytes())); 392 393 // Pop all of the register save are off the stack except the return address 394 __ addptr(rsp, return_offset_in_bytes()); 395 } 396 397 // Is vector's size (in bytes) bigger than a size saved by default? 398 // 16 bytes XMM registers are saved by default using fxsave/fxrstor instructions. 399 bool SharedRuntime::is_wide_vector(int size) { 400 return size > 16; 401 } 402 403 size_t SharedRuntime::trampoline_size() { 404 return 16; 405 } 406 407 void SharedRuntime::generate_trampoline(MacroAssembler *masm, address destination) { 408 __ jump(RuntimeAddress(destination)); 409 } 410 411 // The java_calling_convention describes stack locations as ideal slots on 412 // a frame with no abi restrictions. Since we must observe abi restrictions 413 // (like the placement of the register window) the slots must be biased by 414 // the following value. 415 static int reg2offset_in(VMReg r) { 416 // Account for saved rbp and return address 417 // This should really be in_preserve_stack_slots 418 return (r->reg2stack() + 4) * VMRegImpl::stack_slot_size; 419 } 420 421 static int reg2offset_out(VMReg r) { 422 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 423 } 424 425 // --------------------------------------------------------------------------- 426 // Read the array of BasicTypes from a signature, and compute where the 427 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 428 // quantities. Values less than VMRegImpl::stack0 are registers, those above 429 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 430 // as framesizes are fixed. 431 // VMRegImpl::stack0 refers to the first slot 0(sp). 432 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register 433 // up to RegisterImpl::number_of_registers) are the 64-bit 434 // integer registers. 435 436 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 437 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 438 // units regardless of build. Of course for i486 there is no 64 bit build 439 440 // The Java calling convention is a "shifted" version of the C ABI. 441 // By skipping the first C ABI register we can call non-static jni methods 442 // with small numbers of arguments without having to shuffle the arguments 443 // at all. Since we control the java ABI we ought to at least get some 444 // advantage out of it. 445 446 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 447 VMRegPair *regs, 448 int total_args_passed, 449 int is_outgoing) { 450 451 // Create the mapping between argument positions and 452 // registers. 453 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = { 454 j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5 455 }; 456 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_j] = { 457 j_farg0, j_farg1, j_farg2, j_farg3, 458 j_farg4, j_farg5, j_farg6, j_farg7 459 }; 460 461 462 uint int_args = 0; 463 uint fp_args = 0; 464 uint stk_args = 0; // inc by 2 each time 465 466 for (int i = 0; i < total_args_passed; i++) { 467 switch (sig_bt[i]) { 468 case T_BOOLEAN: 469 case T_CHAR: 470 case T_BYTE: 471 case T_SHORT: 472 case T_INT: 473 if (int_args < Argument::n_int_register_parameters_j) { 474 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 475 } else { 476 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 477 stk_args += 2; 478 } 479 break; 480 case T_VOID: 481 // halves of T_LONG or T_DOUBLE 482 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 483 regs[i].set_bad(); 484 break; 485 case T_LONG: 486 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 487 // fall through 488 case T_OBJECT: 489 case T_ARRAY: 490 case T_ADDRESS: 491 if (int_args < Argument::n_int_register_parameters_j) { 492 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 493 } else { 494 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 495 stk_args += 2; 496 } 497 break; 498 case T_FLOAT: 499 if (fp_args < Argument::n_float_register_parameters_j) { 500 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 501 } else { 502 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 503 stk_args += 2; 504 } 505 break; 506 case T_DOUBLE: 507 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 508 if (fp_args < Argument::n_float_register_parameters_j) { 509 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 510 } else { 511 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 512 stk_args += 2; 513 } 514 break; 515 default: 516 ShouldNotReachHere(); 517 break; 518 } 519 } 520 521 return align_up(stk_args, 2); 522 } 523 524 // Patch the callers callsite with entry to compiled code if it exists. 525 static void patch_callers_callsite(MacroAssembler *masm) { 526 Label L; 527 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD); 528 __ jcc(Assembler::equal, L); 529 530 // Save the current stack pointer 531 __ mov(r13, rsp); 532 // Schedule the branch target address early. 533 // Call into the VM to patch the caller, then jump to compiled callee 534 // rax isn't live so capture return address while we easily can 535 __ movptr(rax, Address(rsp, 0)); 536 537 // align stack so push_CPU_state doesn't fault 538 __ andptr(rsp, -(StackAlignmentInBytes)); 539 __ push_CPU_state(); 540 __ vzeroupper(); 541 // VM needs caller's callsite 542 // VM needs target method 543 // This needs to be a long call since we will relocate this adapter to 544 // the codeBuffer and it may not reach 545 546 // Allocate argument register save area 547 if (frame::arg_reg_save_area_bytes != 0) { 548 __ subptr(rsp, frame::arg_reg_save_area_bytes); 549 } 550 __ mov(c_rarg0, rbx); 551 __ mov(c_rarg1, rax); 552 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite))); 553 554 // De-allocate argument register save area 555 if (frame::arg_reg_save_area_bytes != 0) { 556 __ addptr(rsp, frame::arg_reg_save_area_bytes); 557 } 558 559 __ vzeroupper(); 560 __ pop_CPU_state(); 561 // restore sp 562 __ mov(rsp, r13); 563 __ bind(L); 564 } 565 566 567 static void gen_c2i_adapter(MacroAssembler *masm, 568 int total_args_passed, 569 int comp_args_on_stack, 570 const BasicType *sig_bt, 571 const VMRegPair *regs, 572 Label& skip_fixup) { 573 // Before we get into the guts of the C2I adapter, see if we should be here 574 // at all. We've come from compiled code and are attempting to jump to the 575 // interpreter, which means the caller made a static call to get here 576 // (vcalls always get a compiled target if there is one). Check for a 577 // compiled target. If there is one, we need to patch the caller's call. 578 patch_callers_callsite(masm); 579 580 __ bind(skip_fixup); 581 582 // Since all args are passed on the stack, total_args_passed * 583 // Interpreter::stackElementSize is the space we need. Plus 1 because 584 // we also account for the return address location since 585 // we store it first rather than hold it in rax across all the shuffling 586 587 int extraspace = (total_args_passed * Interpreter::stackElementSize) + wordSize; 588 589 // stack is aligned, keep it that way 590 extraspace = align_up(extraspace, 2*wordSize); 591 592 // Get return address 593 __ pop(rax); 594 595 // set senderSP value 596 __ mov(r13, rsp); 597 598 __ subptr(rsp, extraspace); 599 600 // Store the return address in the expected location 601 __ movptr(Address(rsp, 0), rax); 602 603 // Now write the args into the outgoing interpreter space 604 for (int i = 0; i < total_args_passed; i++) { 605 if (sig_bt[i] == T_VOID) { 606 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 607 continue; 608 } 609 610 // offset to start parameters 611 int st_off = (total_args_passed - i) * Interpreter::stackElementSize; 612 int next_off = st_off - Interpreter::stackElementSize; 613 614 // Say 4 args: 615 // i st_off 616 // 0 32 T_LONG 617 // 1 24 T_VOID 618 // 2 16 T_OBJECT 619 // 3 8 T_BOOL 620 // - 0 return address 621 // 622 // However to make thing extra confusing. Because we can fit a long/double in 623 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter 624 // leaves one slot empty and only stores to a single slot. In this case the 625 // slot that is occupied is the T_VOID slot. See I said it was confusing. 626 627 VMReg r_1 = regs[i].first(); 628 VMReg r_2 = regs[i].second(); 629 if (!r_1->is_valid()) { 630 assert(!r_2->is_valid(), ""); 631 continue; 632 } 633 if (r_1->is_stack()) { 634 // memory to memory use rax 635 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace; 636 if (!r_2->is_valid()) { 637 // sign extend?? 638 __ movl(rax, Address(rsp, ld_off)); 639 __ movptr(Address(rsp, st_off), rax); 640 641 } else { 642 643 __ movq(rax, Address(rsp, ld_off)); 644 645 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 646 // T_DOUBLE and T_LONG use two slots in the interpreter 647 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 648 // ld_off == LSW, ld_off+wordSize == MSW 649 // st_off == MSW, next_off == LSW 650 __ movq(Address(rsp, next_off), rax); 651 #ifdef ASSERT 652 // Overwrite the unused slot with known junk 653 __ mov64(rax, CONST64(0xdeadffffdeadaaaa)); 654 __ movptr(Address(rsp, st_off), rax); 655 #endif /* ASSERT */ 656 } else { 657 __ movq(Address(rsp, st_off), rax); 658 } 659 } 660 } else if (r_1->is_Register()) { 661 Register r = r_1->as_Register(); 662 if (!r_2->is_valid()) { 663 // must be only an int (or less ) so move only 32bits to slot 664 // why not sign extend?? 665 __ movl(Address(rsp, st_off), r); 666 } else { 667 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG 668 // T_DOUBLE and T_LONG use two slots in the interpreter 669 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 670 // long/double in gpr 671 #ifdef ASSERT 672 // Overwrite the unused slot with known junk 673 __ mov64(rax, CONST64(0xdeadffffdeadaaab)); 674 __ movptr(Address(rsp, st_off), rax); 675 #endif /* ASSERT */ 676 __ movq(Address(rsp, next_off), r); 677 } else { 678 __ movptr(Address(rsp, st_off), r); 679 } 680 } 681 } else { 682 assert(r_1->is_XMMRegister(), ""); 683 if (!r_2->is_valid()) { 684 // only a float use just part of the slot 685 __ movflt(Address(rsp, st_off), r_1->as_XMMRegister()); 686 } else { 687 #ifdef ASSERT 688 // Overwrite the unused slot with known junk 689 __ mov64(rax, CONST64(0xdeadffffdeadaaac)); 690 __ movptr(Address(rsp, st_off), rax); 691 #endif /* ASSERT */ 692 __ movdbl(Address(rsp, next_off), r_1->as_XMMRegister()); 693 } 694 } 695 } 696 697 // Schedule the branch target address early. 698 __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset()))); 699 __ jmp(rcx); 700 } 701 702 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg, 703 address code_start, address code_end, 704 Label& L_ok) { 705 Label L_fail; 706 __ lea(temp_reg, ExternalAddress(code_start)); 707 __ cmpptr(pc_reg, temp_reg); 708 __ jcc(Assembler::belowEqual, L_fail); 709 __ lea(temp_reg, ExternalAddress(code_end)); 710 __ cmpptr(pc_reg, temp_reg); 711 __ jcc(Assembler::below, L_ok); 712 __ bind(L_fail); 713 } 714 715 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 716 int total_args_passed, 717 int comp_args_on_stack, 718 const BasicType *sig_bt, 719 const VMRegPair *regs) { 720 721 // Note: r13 contains the senderSP on entry. We must preserve it since 722 // we may do a i2c -> c2i transition if we lose a race where compiled 723 // code goes non-entrant while we get args ready. 724 // In addition we use r13 to locate all the interpreter args as 725 // we must align the stack to 16 bytes on an i2c entry else we 726 // lose alignment we expect in all compiled code and register 727 // save code can segv when fxsave instructions find improperly 728 // aligned stack pointer. 729 730 // Adapters can be frameless because they do not require the caller 731 // to perform additional cleanup work, such as correcting the stack pointer. 732 // An i2c adapter is frameless because the *caller* frame, which is interpreted, 733 // routinely repairs its own stack pointer (from interpreter_frame_last_sp), 734 // even if a callee has modified the stack pointer. 735 // A c2i adapter is frameless because the *callee* frame, which is interpreted, 736 // routinely repairs its caller's stack pointer (from sender_sp, which is set 737 // up via the senderSP register). 738 // In other words, if *either* the caller or callee is interpreted, we can 739 // get the stack pointer repaired after a call. 740 // This is why c2i and i2c adapters cannot be indefinitely composed. 741 // In particular, if a c2i adapter were to somehow call an i2c adapter, 742 // both caller and callee would be compiled methods, and neither would 743 // clean up the stack pointer changes performed by the two adapters. 744 // If this happens, control eventually transfers back to the compiled 745 // caller, but with an uncorrected stack, causing delayed havoc. 746 747 // Pick up the return address 748 __ movptr(rax, Address(rsp, 0)); 749 750 if (VerifyAdapterCalls && 751 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) { 752 // So, let's test for cascading c2i/i2c adapters right now. 753 // assert(Interpreter::contains($return_addr) || 754 // StubRoutines::contains($return_addr), 755 // "i2c adapter must return to an interpreter frame"); 756 __ block_comment("verify_i2c { "); 757 Label L_ok; 758 if (Interpreter::code() != NULL) 759 range_check(masm, rax, r11, 760 Interpreter::code()->code_start(), Interpreter::code()->code_end(), 761 L_ok); 762 if (StubRoutines::code1() != NULL) 763 range_check(masm, rax, r11, 764 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(), 765 L_ok); 766 if (StubRoutines::code2() != NULL) 767 range_check(masm, rax, r11, 768 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(), 769 L_ok); 770 const char* msg = "i2c adapter must return to an interpreter frame"; 771 __ block_comment(msg); 772 __ stop(msg); 773 __ bind(L_ok); 774 __ block_comment("} verify_i2ce "); 775 } 776 777 // Must preserve original SP for loading incoming arguments because 778 // we need to align the outgoing SP for compiled code. 779 __ movptr(r11, rsp); 780 781 // Cut-out for having no stack args. Since up to 2 int/oop args are passed 782 // in registers, we will occasionally have no stack args. 783 int comp_words_on_stack = 0; 784 if (comp_args_on_stack) { 785 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in 786 // registers are below. By subtracting stack0, we either get a negative 787 // number (all values in registers) or the maximum stack slot accessed. 788 789 // Convert 4-byte c2 stack slots to words. 790 comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 791 // Round up to miminum stack alignment, in wordSize 792 comp_words_on_stack = align_up(comp_words_on_stack, 2); 793 __ subptr(rsp, comp_words_on_stack * wordSize); 794 } 795 796 797 // Ensure compiled code always sees stack at proper alignment 798 __ andptr(rsp, -16); 799 800 // push the return address and misalign the stack that youngest frame always sees 801 // as far as the placement of the call instruction 802 __ push(rax); 803 804 // Put saved SP in another register 805 const Register saved_sp = rax; 806 __ movptr(saved_sp, r11); 807 808 // Will jump to the compiled code just as if compiled code was doing it. 809 // Pre-load the register-jump target early, to schedule it better. 810 __ movptr(r11, Address(rbx, in_bytes(Method::from_compiled_offset()))); 811 812 #if INCLUDE_JVMCI 813 if (EnableJVMCI || UseAOT) { 814 // check if this call should be routed towards a specific entry point 815 __ cmpptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0); 816 Label no_alternative_target; 817 __ jcc(Assembler::equal, no_alternative_target); 818 __ movptr(r11, Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 819 __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0); 820 __ bind(no_alternative_target); 821 } 822 #endif // INCLUDE_JVMCI 823 824 // Now generate the shuffle code. Pick up all register args and move the 825 // rest through the floating point stack top. 826 for (int i = 0; i < total_args_passed; i++) { 827 if (sig_bt[i] == T_VOID) { 828 // Longs and doubles are passed in native word order, but misaligned 829 // in the 32-bit build. 830 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 831 continue; 832 } 833 834 // Pick up 0, 1 or 2 words from SP+offset. 835 836 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), 837 "scrambled load targets?"); 838 // Load in argument order going down. 839 int ld_off = (total_args_passed - i)*Interpreter::stackElementSize; 840 // Point to interpreter value (vs. tag) 841 int next_off = ld_off - Interpreter::stackElementSize; 842 // 843 // 844 // 845 VMReg r_1 = regs[i].first(); 846 VMReg r_2 = regs[i].second(); 847 if (!r_1->is_valid()) { 848 assert(!r_2->is_valid(), ""); 849 continue; 850 } 851 if (r_1->is_stack()) { 852 // Convert stack slot to an SP offset (+ wordSize to account for return address ) 853 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize; 854 855 // We can use r13 as a temp here because compiled code doesn't need r13 as an input 856 // and if we end up going thru a c2i because of a miss a reasonable value of r13 857 // will be generated. 858 if (!r_2->is_valid()) { 859 // sign extend??? 860 __ movl(r13, Address(saved_sp, ld_off)); 861 __ movptr(Address(rsp, st_off), r13); 862 } else { 863 // 864 // We are using two optoregs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE 865 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case 866 // So we must adjust where to pick up the data to match the interpreter. 867 // 868 // Interpreter local[n] == MSW, local[n+1] == LSW however locals 869 // are accessed as negative so LSW is at LOW address 870 871 // ld_off is MSW so get LSW 872 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 873 next_off : ld_off; 874 __ movq(r13, Address(saved_sp, offset)); 875 // st_off is LSW (i.e. reg.first()) 876 __ movq(Address(rsp, st_off), r13); 877 } 878 } else if (r_1->is_Register()) { // Register argument 879 Register r = r_1->as_Register(); 880 assert(r != rax, "must be different"); 881 if (r_2->is_valid()) { 882 // 883 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE 884 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case 885 // So we must adjust where to pick up the data to match the interpreter. 886 887 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)? 888 next_off : ld_off; 889 890 // this can be a misaligned move 891 __ movq(r, Address(saved_sp, offset)); 892 } else { 893 // sign extend and use a full word? 894 __ movl(r, Address(saved_sp, ld_off)); 895 } 896 } else { 897 if (!r_2->is_valid()) { 898 __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off)); 899 } else { 900 __ movdbl(r_1->as_XMMRegister(), Address(saved_sp, next_off)); 901 } 902 } 903 } 904 905 // 6243940 We might end up in handle_wrong_method if 906 // the callee is deoptimized as we race thru here. If that 907 // happens we don't want to take a safepoint because the 908 // caller frame will look interpreted and arguments are now 909 // "compiled" so it is much better to make this transition 910 // invisible to the stack walking code. Unfortunately if 911 // we try and find the callee by normal means a safepoint 912 // is possible. So we stash the desired callee in the thread 913 // and the vm will find there should this case occur. 914 915 __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx); 916 917 // put Method* where a c2i would expect should we end up there 918 // only needed becaus eof c2 resolve stubs return Method* as a result in 919 // rax 920 __ mov(rax, rbx); 921 __ jmp(r11); 922 } 923 924 // --------------------------------------------------------------- 925 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 926 int total_args_passed, 927 int comp_args_on_stack, 928 const BasicType *sig_bt, 929 const VMRegPair *regs, 930 AdapterFingerPrint* fingerprint) { 931 address i2c_entry = __ pc(); 932 933 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs); 934 935 // ------------------------------------------------------------------------- 936 // Generate a C2I adapter. On entry we know rbx holds the Method* during calls 937 // to the interpreter. The args start out packed in the compiled layout. They 938 // need to be unpacked into the interpreter layout. This will almost always 939 // require some stack space. We grow the current (compiled) stack, then repack 940 // the args. We finally end in a jump to the generic interpreter entry point. 941 // On exit from the interpreter, the interpreter will restore our SP (lest the 942 // compiled code, which relys solely on SP and not RBP, get sick). 943 944 address c2i_unverified_entry = __ pc(); 945 Label skip_fixup; 946 Label ok; 947 948 Register holder = rax; 949 Register receiver = j_rarg0; 950 Register temp = rbx; 951 952 { 953 __ load_klass(temp, receiver); 954 __ cmpptr(temp, Address(holder, CompiledICHolder::holder_klass_offset())); 955 __ movptr(rbx, Address(holder, CompiledICHolder::holder_metadata_offset())); 956 __ jcc(Assembler::equal, ok); 957 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 958 959 __ bind(ok); 960 // Method might have been compiled since the call site was patched to 961 // interpreted if that is the case treat it as a miss so we can get 962 // the call site corrected. 963 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD); 964 __ jcc(Assembler::equal, skip_fixup); 965 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 966 } 967 968 address c2i_entry = __ pc(); 969 970 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); 971 972 __ flush(); 973 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 974 } 975 976 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 977 VMRegPair *regs, 978 VMRegPair *regs2, 979 int total_args_passed) { 980 assert(regs2 == NULL, "not needed on x86"); 981 // We return the amount of VMRegImpl stack slots we need to reserve for all 982 // the arguments NOT counting out_preserve_stack_slots. 983 984 // NOTE: These arrays will have to change when c1 is ported 985 #ifdef _WIN64 986 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 987 c_rarg0, c_rarg1, c_rarg2, c_rarg3 988 }; 989 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = { 990 c_farg0, c_farg1, c_farg2, c_farg3 991 }; 992 #else 993 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 994 c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5 995 }; 996 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = { 997 c_farg0, c_farg1, c_farg2, c_farg3, 998 c_farg4, c_farg5, c_farg6, c_farg7 999 }; 1000 #endif // _WIN64 1001 1002 1003 uint int_args = 0; 1004 uint fp_args = 0; 1005 uint stk_args = 0; // inc by 2 each time 1006 1007 for (int i = 0; i < total_args_passed; i++) { 1008 switch (sig_bt[i]) { 1009 case T_BOOLEAN: 1010 case T_CHAR: 1011 case T_BYTE: 1012 case T_SHORT: 1013 case T_INT: 1014 if (int_args < Argument::n_int_register_parameters_c) { 1015 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 1016 #ifdef _WIN64 1017 fp_args++; 1018 // Allocate slots for callee to stuff register args the stack. 1019 stk_args += 2; 1020 #endif 1021 } else { 1022 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 1023 stk_args += 2; 1024 } 1025 break; 1026 case T_LONG: 1027 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 1028 // fall through 1029 case T_OBJECT: 1030 case T_ARRAY: 1031 case T_ADDRESS: 1032 case T_METADATA: 1033 if (int_args < Argument::n_int_register_parameters_c) { 1034 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 1035 #ifdef _WIN64 1036 fp_args++; 1037 stk_args += 2; 1038 #endif 1039 } else { 1040 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 1041 stk_args += 2; 1042 } 1043 break; 1044 case T_FLOAT: 1045 if (fp_args < Argument::n_float_register_parameters_c) { 1046 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 1047 #ifdef _WIN64 1048 int_args++; 1049 // Allocate slots for callee to stuff register args the stack. 1050 stk_args += 2; 1051 #endif 1052 } else { 1053 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 1054 stk_args += 2; 1055 } 1056 break; 1057 case T_DOUBLE: 1058 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 1059 if (fp_args < Argument::n_float_register_parameters_c) { 1060 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 1061 #ifdef _WIN64 1062 int_args++; 1063 // Allocate slots for callee to stuff register args the stack. 1064 stk_args += 2; 1065 #endif 1066 } else { 1067 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 1068 stk_args += 2; 1069 } 1070 break; 1071 case T_VOID: // Halves of longs and doubles 1072 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 1073 regs[i].set_bad(); 1074 break; 1075 default: 1076 ShouldNotReachHere(); 1077 break; 1078 } 1079 } 1080 #ifdef _WIN64 1081 // windows abi requires that we always allocate enough stack space 1082 // for 4 64bit registers to be stored down. 1083 if (stk_args < 8) { 1084 stk_args = 8; 1085 } 1086 #endif // _WIN64 1087 1088 return stk_args; 1089 } 1090 1091 // On 64 bit we will store integer like items to the stack as 1092 // 64 bits items (sparc abi) even though java would only store 1093 // 32bits for a parameter. On 32bit it will simply be 32 bits 1094 // So this routine will do 32->32 on 32bit and 32->64 on 64bit 1095 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1096 if (src.first()->is_stack()) { 1097 if (dst.first()->is_stack()) { 1098 // stack to stack 1099 __ movslq(rax, Address(rbp, reg2offset_in(src.first()))); 1100 __ movq(Address(rsp, reg2offset_out(dst.first())), rax); 1101 } else { 1102 // stack to reg 1103 __ movslq(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first()))); 1104 } 1105 } else if (dst.first()->is_stack()) { 1106 // reg to stack 1107 // Do we really have to sign extend??? 1108 // __ movslq(src.first()->as_Register(), src.first()->as_Register()); 1109 __ movq(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register()); 1110 } else { 1111 // Do we really have to sign extend??? 1112 // __ movslq(dst.first()->as_Register(), src.first()->as_Register()); 1113 if (dst.first() != src.first()) { 1114 __ movq(dst.first()->as_Register(), src.first()->as_Register()); 1115 } 1116 } 1117 } 1118 1119 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1120 if (src.first()->is_stack()) { 1121 if (dst.first()->is_stack()) { 1122 // stack to stack 1123 __ movq(rax, Address(rbp, reg2offset_in(src.first()))); 1124 __ movq(Address(rsp, reg2offset_out(dst.first())), rax); 1125 } else { 1126 // stack to reg 1127 __ movq(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first()))); 1128 } 1129 } else if (dst.first()->is_stack()) { 1130 // reg to stack 1131 __ movq(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register()); 1132 } else { 1133 if (dst.first() != src.first()) { 1134 __ movq(dst.first()->as_Register(), src.first()->as_Register()); 1135 } 1136 } 1137 } 1138 1139 // An oop arg. Must pass a handle not the oop itself 1140 static void object_move(MacroAssembler* masm, 1141 OopMap* map, 1142 int oop_handle_offset, 1143 int framesize_in_slots, 1144 VMRegPair src, 1145 VMRegPair dst, 1146 bool is_receiver, 1147 int* receiver_offset) { 1148 1149 // must pass a handle. First figure out the location we use as a handle 1150 1151 Register rHandle = dst.first()->is_stack() ? rax : dst.first()->as_Register(); 1152 1153 // See if oop is NULL if it is we need no handle 1154 1155 if (src.first()->is_stack()) { 1156 1157 // Oop is already on the stack as an argument 1158 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1159 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 1160 if (is_receiver) { 1161 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 1162 } 1163 1164 __ cmpptr(Address(rbp, reg2offset_in(src.first())), (int32_t)NULL_WORD); 1165 __ lea(rHandle, Address(rbp, reg2offset_in(src.first()))); 1166 // conditionally move a NULL 1167 __ cmovptr(Assembler::equal, rHandle, Address(rbp, reg2offset_in(src.first()))); 1168 } else { 1169 1170 // Oop is in an a register we must store it to the space we reserve 1171 // on the stack for oop_handles and pass a handle if oop is non-NULL 1172 1173 const Register rOop = src.first()->as_Register(); 1174 int oop_slot; 1175 if (rOop == j_rarg0) 1176 oop_slot = 0; 1177 else if (rOop == j_rarg1) 1178 oop_slot = 1; 1179 else if (rOop == j_rarg2) 1180 oop_slot = 2; 1181 else if (rOop == j_rarg3) 1182 oop_slot = 3; 1183 else if (rOop == j_rarg4) 1184 oop_slot = 4; 1185 else { 1186 assert(rOop == j_rarg5, "wrong register"); 1187 oop_slot = 5; 1188 } 1189 1190 oop_slot = oop_slot * VMRegImpl::slots_per_word + oop_handle_offset; 1191 int offset = oop_slot*VMRegImpl::stack_slot_size; 1192 1193 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1194 // Store oop in handle area, may be NULL 1195 __ movptr(Address(rsp, offset), rOop); 1196 if (is_receiver) { 1197 *receiver_offset = offset; 1198 } 1199 1200 __ cmpptr(rOop, (int32_t)NULL_WORD); 1201 __ lea(rHandle, Address(rsp, offset)); 1202 // conditionally move a NULL from the handle area where it was just stored 1203 __ cmovptr(Assembler::equal, rHandle, Address(rsp, offset)); 1204 } 1205 1206 // If arg is on the stack then place it otherwise it is already in correct reg. 1207 if (dst.first()->is_stack()) { 1208 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle); 1209 } 1210 } 1211 1212 // A float arg may have to do float reg int reg conversion 1213 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1214 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move"); 1215 1216 // The calling conventions assures us that each VMregpair is either 1217 // all really one physical register or adjacent stack slots. 1218 // This greatly simplifies the cases here compared to sparc. 1219 1220 if (src.first()->is_stack()) { 1221 if (dst.first()->is_stack()) { 1222 __ movl(rax, Address(rbp, reg2offset_in(src.first()))); 1223 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1224 } else { 1225 // stack to reg 1226 assert(dst.first()->is_XMMRegister(), "only expect xmm registers as parameters"); 1227 __ movflt(dst.first()->as_XMMRegister(), Address(rbp, reg2offset_in(src.first()))); 1228 } 1229 } else if (dst.first()->is_stack()) { 1230 // reg to stack 1231 assert(src.first()->is_XMMRegister(), "only expect xmm registers as parameters"); 1232 __ movflt(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister()); 1233 } else { 1234 // reg to reg 1235 // In theory these overlap but the ordering is such that this is likely a nop 1236 if ( src.first() != dst.first()) { 1237 __ movdbl(dst.first()->as_XMMRegister(), src.first()->as_XMMRegister()); 1238 } 1239 } 1240 } 1241 1242 // A long move 1243 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1244 1245 // The calling conventions assures us that each VMregpair is either 1246 // all really one physical register or adjacent stack slots. 1247 // This greatly simplifies the cases here compared to sparc. 1248 1249 if (src.is_single_phys_reg() ) { 1250 if (dst.is_single_phys_reg()) { 1251 if (dst.first() != src.first()) { 1252 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 1253 } 1254 } else { 1255 assert(dst.is_single_reg(), "not a stack pair"); 1256 __ movq(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register()); 1257 } 1258 } else if (dst.is_single_phys_reg()) { 1259 assert(src.is_single_reg(), "not a stack pair"); 1260 __ movq(dst.first()->as_Register(), Address(rbp, reg2offset_out(src.first()))); 1261 } else { 1262 assert(src.is_single_reg() && dst.is_single_reg(), "not stack pairs"); 1263 __ movq(rax, Address(rbp, reg2offset_in(src.first()))); 1264 __ movq(Address(rsp, reg2offset_out(dst.first())), rax); 1265 } 1266 } 1267 1268 // A double move 1269 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1270 1271 // The calling conventions assures us that each VMregpair is either 1272 // all really one physical register or adjacent stack slots. 1273 // This greatly simplifies the cases here compared to sparc. 1274 1275 if (src.is_single_phys_reg() ) { 1276 if (dst.is_single_phys_reg()) { 1277 // In theory these overlap but the ordering is such that this is likely a nop 1278 if ( src.first() != dst.first()) { 1279 __ movdbl(dst.first()->as_XMMRegister(), src.first()->as_XMMRegister()); 1280 } 1281 } else { 1282 assert(dst.is_single_reg(), "not a stack pair"); 1283 __ movdbl(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister()); 1284 } 1285 } else if (dst.is_single_phys_reg()) { 1286 assert(src.is_single_reg(), "not a stack pair"); 1287 __ movdbl(dst.first()->as_XMMRegister(), Address(rbp, reg2offset_out(src.first()))); 1288 } else { 1289 assert(src.is_single_reg() && dst.is_single_reg(), "not stack pairs"); 1290 __ movq(rax, Address(rbp, reg2offset_in(src.first()))); 1291 __ movq(Address(rsp, reg2offset_out(dst.first())), rax); 1292 } 1293 } 1294 1295 1296 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1297 // We always ignore the frame_slots arg and just use the space just below frame pointer 1298 // which by this time is free to use 1299 switch (ret_type) { 1300 case T_FLOAT: 1301 __ movflt(Address(rbp, -wordSize), xmm0); 1302 break; 1303 case T_DOUBLE: 1304 __ movdbl(Address(rbp, -wordSize), xmm0); 1305 break; 1306 case T_VOID: break; 1307 default: { 1308 __ movptr(Address(rbp, -wordSize), rax); 1309 } 1310 } 1311 } 1312 1313 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1314 // We always ignore the frame_slots arg and just use the space just below frame pointer 1315 // which by this time is free to use 1316 switch (ret_type) { 1317 case T_FLOAT: 1318 __ movflt(xmm0, Address(rbp, -wordSize)); 1319 break; 1320 case T_DOUBLE: 1321 __ movdbl(xmm0, Address(rbp, -wordSize)); 1322 break; 1323 case T_VOID: break; 1324 default: { 1325 __ movptr(rax, Address(rbp, -wordSize)); 1326 } 1327 } 1328 } 1329 1330 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1331 for ( int i = first_arg ; i < arg_count ; i++ ) { 1332 if (args[i].first()->is_Register()) { 1333 __ push(args[i].first()->as_Register()); 1334 } else if (args[i].first()->is_XMMRegister()) { 1335 __ subptr(rsp, 2*wordSize); 1336 __ movdbl(Address(rsp, 0), args[i].first()->as_XMMRegister()); 1337 } 1338 } 1339 } 1340 1341 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1342 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) { 1343 if (args[i].first()->is_Register()) { 1344 __ pop(args[i].first()->as_Register()); 1345 } else if (args[i].first()->is_XMMRegister()) { 1346 __ movdbl(args[i].first()->as_XMMRegister(), Address(rsp, 0)); 1347 __ addptr(rsp, 2*wordSize); 1348 } 1349 } 1350 } 1351 1352 1353 static void save_or_restore_arguments(MacroAssembler* masm, 1354 const int stack_slots, 1355 const int total_in_args, 1356 const int arg_save_area, 1357 OopMap* map, 1358 VMRegPair* in_regs, 1359 BasicType* in_sig_bt) { 1360 // if map is non-NULL then the code should store the values, 1361 // otherwise it should load them. 1362 int slot = arg_save_area; 1363 // Save down double word first 1364 for ( int i = 0; i < total_in_args; i++) { 1365 if (in_regs[i].first()->is_XMMRegister() && in_sig_bt[i] == T_DOUBLE) { 1366 int offset = slot * VMRegImpl::stack_slot_size; 1367 slot += VMRegImpl::slots_per_word; 1368 assert(slot <= stack_slots, "overflow"); 1369 if (map != NULL) { 1370 __ movdbl(Address(rsp, offset), in_regs[i].first()->as_XMMRegister()); 1371 } else { 1372 __ movdbl(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset)); 1373 } 1374 } 1375 if (in_regs[i].first()->is_Register() && 1376 (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_ARRAY)) { 1377 int offset = slot * VMRegImpl::stack_slot_size; 1378 if (map != NULL) { 1379 __ movq(Address(rsp, offset), in_regs[i].first()->as_Register()); 1380 if (in_sig_bt[i] == T_ARRAY) { 1381 map->set_oop(VMRegImpl::stack2reg(slot));; 1382 } 1383 } else { 1384 __ movq(in_regs[i].first()->as_Register(), Address(rsp, offset)); 1385 } 1386 slot += VMRegImpl::slots_per_word; 1387 } 1388 } 1389 // Save or restore single word registers 1390 for ( int i = 0; i < total_in_args; i++) { 1391 if (in_regs[i].first()->is_Register()) { 1392 int offset = slot * VMRegImpl::stack_slot_size; 1393 slot++; 1394 assert(slot <= stack_slots, "overflow"); 1395 1396 // Value is in an input register pass we must flush it to the stack 1397 const Register reg = in_regs[i].first()->as_Register(); 1398 switch (in_sig_bt[i]) { 1399 case T_BOOLEAN: 1400 case T_CHAR: 1401 case T_BYTE: 1402 case T_SHORT: 1403 case T_INT: 1404 if (map != NULL) { 1405 __ movl(Address(rsp, offset), reg); 1406 } else { 1407 __ movl(reg, Address(rsp, offset)); 1408 } 1409 break; 1410 case T_ARRAY: 1411 case T_LONG: 1412 // handled above 1413 break; 1414 case T_OBJECT: 1415 default: ShouldNotReachHere(); 1416 } 1417 } else if (in_regs[i].first()->is_XMMRegister()) { 1418 if (in_sig_bt[i] == T_FLOAT) { 1419 int offset = slot * VMRegImpl::stack_slot_size; 1420 slot++; 1421 assert(slot <= stack_slots, "overflow"); 1422 if (map != NULL) { 1423 __ movflt(Address(rsp, offset), in_regs[i].first()->as_XMMRegister()); 1424 } else { 1425 __ movflt(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset)); 1426 } 1427 } 1428 } else if (in_regs[i].first()->is_stack()) { 1429 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1430 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1431 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots)); 1432 } 1433 } 1434 } 1435 } 1436 1437 // Pin object, return pinned object or null in rax 1438 static void gen_pin_object(MacroAssembler* masm, 1439 VMRegPair reg) { 1440 __ block_comment("gen_pin_object {"); 1441 1442 // rax always contains oop, either incoming or 1443 // pinned. 1444 Register tmp_reg = rax; 1445 1446 Label is_null; 1447 VMRegPair tmp; 1448 VMRegPair in_reg = reg; 1449 1450 tmp.set_ptr(tmp_reg->as_VMReg()); 1451 if (reg.first()->is_stack()) { 1452 // Load the arg up from the stack 1453 move_ptr(masm, reg, tmp); 1454 reg = tmp; 1455 } else { 1456 __ movptr(rax, reg.first()->as_Register()); 1457 } 1458 __ testptr(reg.first()->as_Register(), reg.first()->as_Register()); 1459 __ jccb(Assembler::equal, is_null); 1460 1461 if (reg.first()->as_Register() != c_rarg1) { 1462 __ movptr(c_rarg1, reg.first()->as_Register()); 1463 } 1464 1465 __ call_VM_leaf( 1466 CAST_FROM_FN_PTR(address, SharedRuntime::pin_object), 1467 r15_thread, c_rarg1); 1468 1469 __ bind(is_null); 1470 __ block_comment("} gen_pin_object"); 1471 } 1472 1473 // Unpin object 1474 static void gen_unpin_object(MacroAssembler* masm, 1475 VMRegPair reg) { 1476 __ block_comment("gen_unpin_object {"); 1477 Label is_null; 1478 1479 if (reg.first()->is_stack()) { 1480 __ movptr(c_rarg1, Address(rbp, reg2offset_in(reg.first()))); 1481 } else if (reg.first()->as_Register() != c_rarg1) { 1482 __ movptr(c_rarg1, reg.first()->as_Register()); 1483 } 1484 1485 __ testptr(c_rarg1, c_rarg1); 1486 __ jccb(Assembler::equal, is_null); 1487 1488 __ call_VM_leaf( 1489 CAST_FROM_FN_PTR(address, SharedRuntime::unpin_object), 1490 r15_thread, c_rarg1); 1491 1492 __ bind(is_null); 1493 __ block_comment("} gen_unpin_object"); 1494 } 1495 1496 // Check GCLocker::needs_gc and enter the runtime if it's true. This 1497 // keeps a new JNI critical region from starting until a GC has been 1498 // forced. Save down any oops in registers and describe them in an 1499 // OopMap. 1500 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1501 int stack_slots, 1502 int total_c_args, 1503 int total_in_args, 1504 int arg_save_area, 1505 OopMapSet* oop_maps, 1506 VMRegPair* in_regs, 1507 BasicType* in_sig_bt) { 1508 __ block_comment("check GCLocker::needs_gc"); 1509 Label cont; 1510 __ cmp8(ExternalAddress((address)GCLocker::needs_gc_address()), false); 1511 __ jcc(Assembler::equal, cont); 1512 1513 // Save down any incoming oops and call into the runtime to halt for a GC 1514 1515 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1516 save_or_restore_arguments(masm, stack_slots, total_in_args, 1517 arg_save_area, map, in_regs, in_sig_bt); 1518 1519 address the_pc = __ pc(); 1520 oop_maps->add_gc_map( __ offset(), map); 1521 __ set_last_Java_frame(rsp, noreg, the_pc); 1522 1523 __ block_comment("block_for_jni_critical"); 1524 __ movptr(c_rarg0, r15_thread); 1525 __ mov(r12, rsp); // remember sp 1526 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 1527 __ andptr(rsp, -16); // align stack as required by ABI 1528 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical))); 1529 __ mov(rsp, r12); // restore sp 1530 __ reinit_heapbase(); 1531 1532 __ reset_last_Java_frame(false); 1533 1534 save_or_restore_arguments(masm, stack_slots, total_in_args, 1535 arg_save_area, NULL, in_regs, in_sig_bt); 1536 __ bind(cont); 1537 #ifdef ASSERT 1538 if (StressCriticalJNINatives) { 1539 // Stress register saving 1540 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1541 save_or_restore_arguments(masm, stack_slots, total_in_args, 1542 arg_save_area, map, in_regs, in_sig_bt); 1543 // Destroy argument registers 1544 for (int i = 0; i < total_in_args - 1; i++) { 1545 if (in_regs[i].first()->is_Register()) { 1546 const Register reg = in_regs[i].first()->as_Register(); 1547 __ xorptr(reg, reg); 1548 } else if (in_regs[i].first()->is_XMMRegister()) { 1549 __ xorpd(in_regs[i].first()->as_XMMRegister(), in_regs[i].first()->as_XMMRegister()); 1550 } else if (in_regs[i].first()->is_FloatRegister()) { 1551 ShouldNotReachHere(); 1552 } else if (in_regs[i].first()->is_stack()) { 1553 // Nothing to do 1554 } else { 1555 ShouldNotReachHere(); 1556 } 1557 if (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_DOUBLE) { 1558 i++; 1559 } 1560 } 1561 1562 save_or_restore_arguments(masm, stack_slots, total_in_args, 1563 arg_save_area, NULL, in_regs, in_sig_bt); 1564 } 1565 #endif 1566 } 1567 1568 // Unpack an array argument into a pointer to the body and the length 1569 // if the array is non-null, otherwise pass 0 for both. 1570 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) { 1571 Register tmp_reg = rax; 1572 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg, 1573 "possible collision"); 1574 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg, 1575 "possible collision"); 1576 1577 __ block_comment("unpack_array_argument {"); 1578 1579 // Pass the length, ptr pair 1580 Label is_null, done; 1581 VMRegPair tmp; 1582 tmp.set_ptr(tmp_reg->as_VMReg()); 1583 if (reg.first()->is_stack()) { 1584 // Load the arg up from the stack 1585 move_ptr(masm, reg, tmp); 1586 reg = tmp; 1587 } 1588 __ testptr(reg.first()->as_Register(), reg.first()->as_Register()); 1589 __ jccb(Assembler::equal, is_null); 1590 __ lea(tmp_reg, Address(reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type))); 1591 move_ptr(masm, tmp, body_arg); 1592 // load the length relative to the body. 1593 __ movl(tmp_reg, Address(tmp_reg, arrayOopDesc::length_offset_in_bytes() - 1594 arrayOopDesc::base_offset_in_bytes(in_elem_type))); 1595 move32_64(masm, tmp, length_arg); 1596 __ jmpb(done); 1597 __ bind(is_null); 1598 // Pass zeros 1599 __ xorptr(tmp_reg, tmp_reg); 1600 move_ptr(masm, tmp, body_arg); 1601 move32_64(masm, tmp, length_arg); 1602 __ bind(done); 1603 1604 __ block_comment("} unpack_array_argument"); 1605 } 1606 1607 1608 // Different signatures may require very different orders for the move 1609 // to avoid clobbering other arguments. There's no simple way to 1610 // order them safely. Compute a safe order for issuing stores and 1611 // break any cycles in those stores. This code is fairly general but 1612 // it's not necessary on the other platforms so we keep it in the 1613 // platform dependent code instead of moving it into a shared file. 1614 // (See bugs 7013347 & 7145024.) 1615 // Note that this code is specific to LP64. 1616 class ComputeMoveOrder: public StackObj { 1617 class MoveOperation: public ResourceObj { 1618 friend class ComputeMoveOrder; 1619 private: 1620 VMRegPair _src; 1621 VMRegPair _dst; 1622 int _src_index; 1623 int _dst_index; 1624 bool _processed; 1625 MoveOperation* _next; 1626 MoveOperation* _prev; 1627 1628 static int get_id(VMRegPair r) { 1629 return r.first()->value(); 1630 } 1631 1632 public: 1633 MoveOperation(int src_index, VMRegPair src, int dst_index, VMRegPair dst): 1634 _src(src) 1635 , _src_index(src_index) 1636 , _dst(dst) 1637 , _dst_index(dst_index) 1638 , _next(NULL) 1639 , _prev(NULL) 1640 , _processed(false) { 1641 } 1642 1643 VMRegPair src() const { return _src; } 1644 int src_id() const { return get_id(src()); } 1645 int src_index() const { return _src_index; } 1646 VMRegPair dst() const { return _dst; } 1647 void set_dst(int i, VMRegPair dst) { _dst_index = i, _dst = dst; } 1648 int dst_index() const { return _dst_index; } 1649 int dst_id() const { return get_id(dst()); } 1650 MoveOperation* next() const { return _next; } 1651 MoveOperation* prev() const { return _prev; } 1652 void set_processed() { _processed = true; } 1653 bool is_processed() const { return _processed; } 1654 1655 // insert 1656 void break_cycle(VMRegPair temp_register) { 1657 // create a new store following the last store 1658 // to move from the temp_register to the original 1659 MoveOperation* new_store = new MoveOperation(-1, temp_register, dst_index(), dst()); 1660 1661 // break the cycle of links and insert new_store at the end 1662 // break the reverse link. 1663 MoveOperation* p = prev(); 1664 assert(p->next() == this, "must be"); 1665 _prev = NULL; 1666 p->_next = new_store; 1667 new_store->_prev = p; 1668 1669 // change the original store to save it's value in the temp. 1670 set_dst(-1, temp_register); 1671 } 1672 1673 void link(GrowableArray<MoveOperation*>& killer) { 1674 // link this store in front the store that it depends on 1675 MoveOperation* n = killer.at_grow(src_id(), NULL); 1676 if (n != NULL) { 1677 assert(_next == NULL && n->_prev == NULL, "shouldn't have been set yet"); 1678 _next = n; 1679 n->_prev = this; 1680 } 1681 } 1682 }; 1683 1684 private: 1685 GrowableArray<MoveOperation*> edges; 1686 1687 public: 1688 ComputeMoveOrder(int total_in_args, VMRegPair* in_regs, int total_c_args, VMRegPair* out_regs, 1689 BasicType* in_sig_bt, GrowableArray<int>& arg_order, VMRegPair tmp_vmreg) { 1690 // Move operations where the dest is the stack can all be 1691 // scheduled first since they can't interfere with the other moves. 1692 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 1693 if (in_sig_bt[i] == T_ARRAY) { 1694 c_arg--; 1695 if (out_regs[c_arg].first()->is_stack() && 1696 out_regs[c_arg + 1].first()->is_stack()) { 1697 arg_order.push(i); 1698 arg_order.push(c_arg); 1699 } else { 1700 if (out_regs[c_arg].first()->is_stack() || 1701 in_regs[i].first() == out_regs[c_arg].first()) { 1702 add_edge(i, in_regs[i].first(), c_arg, out_regs[c_arg + 1]); 1703 } else { 1704 add_edge(i, in_regs[i].first(), c_arg, out_regs[c_arg]); 1705 } 1706 } 1707 } else if (in_sig_bt[i] == T_VOID) { 1708 arg_order.push(i); 1709 arg_order.push(c_arg); 1710 } else { 1711 if (out_regs[c_arg].first()->is_stack() || 1712 in_regs[i].first() == out_regs[c_arg].first()) { 1713 arg_order.push(i); 1714 arg_order.push(c_arg); 1715 } else { 1716 add_edge(i, in_regs[i].first(), c_arg, out_regs[c_arg]); 1717 } 1718 } 1719 } 1720 // Break any cycles in the register moves and emit the in the 1721 // proper order. 1722 GrowableArray<MoveOperation*>* stores = get_store_order(tmp_vmreg); 1723 for (int i = 0; i < stores->length(); i++) { 1724 arg_order.push(stores->at(i)->src_index()); 1725 arg_order.push(stores->at(i)->dst_index()); 1726 } 1727 } 1728 1729 // Collected all the move operations 1730 void add_edge(int src_index, VMRegPair src, int dst_index, VMRegPair dst) { 1731 if (src.first() == dst.first()) return; 1732 edges.append(new MoveOperation(src_index, src, dst_index, dst)); 1733 } 1734 1735 // Walk the edges breaking cycles between moves. The result list 1736 // can be walked in order to produce the proper set of loads 1737 GrowableArray<MoveOperation*>* get_store_order(VMRegPair temp_register) { 1738 // Record which moves kill which values 1739 GrowableArray<MoveOperation*> killer; 1740 for (int i = 0; i < edges.length(); i++) { 1741 MoveOperation* s = edges.at(i); 1742 assert(killer.at_grow(s->dst_id(), NULL) == NULL, "only one killer"); 1743 killer.at_put_grow(s->dst_id(), s, NULL); 1744 } 1745 assert(killer.at_grow(MoveOperation::get_id(temp_register), NULL) == NULL, 1746 "make sure temp isn't in the registers that are killed"); 1747 1748 // create links between loads and stores 1749 for (int i = 0; i < edges.length(); i++) { 1750 edges.at(i)->link(killer); 1751 } 1752 1753 // at this point, all the move operations are chained together 1754 // in a doubly linked list. Processing it backwards finds 1755 // the beginning of the chain, forwards finds the end. If there's 1756 // a cycle it can be broken at any point, so pick an edge and walk 1757 // backward until the list ends or we end where we started. 1758 GrowableArray<MoveOperation*>* stores = new GrowableArray<MoveOperation*>(); 1759 for (int e = 0; e < edges.length(); e++) { 1760 MoveOperation* s = edges.at(e); 1761 if (!s->is_processed()) { 1762 MoveOperation* start = s; 1763 // search for the beginning of the chain or cycle 1764 while (start->prev() != NULL && start->prev() != s) { 1765 start = start->prev(); 1766 } 1767 if (start->prev() == s) { 1768 start->break_cycle(temp_register); 1769 } 1770 // walk the chain forward inserting to store list 1771 while (start != NULL) { 1772 stores->append(start); 1773 start->set_processed(); 1774 start = start->next(); 1775 } 1776 } 1777 } 1778 return stores; 1779 } 1780 }; 1781 1782 static void verify_oop_args(MacroAssembler* masm, 1783 const methodHandle& method, 1784 const BasicType* sig_bt, 1785 const VMRegPair* regs) { 1786 Register temp_reg = rbx; // not part of any compiled calling seq 1787 if (VerifyOops) { 1788 for (int i = 0; i < method->size_of_parameters(); i++) { 1789 if (sig_bt[i] == T_OBJECT || 1790 sig_bt[i] == T_ARRAY) { 1791 VMReg r = regs[i].first(); 1792 assert(r->is_valid(), "bad oop arg"); 1793 if (r->is_stack()) { 1794 __ movptr(temp_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1795 __ verify_oop(temp_reg); 1796 } else { 1797 __ verify_oop(r->as_Register()); 1798 } 1799 } 1800 } 1801 } 1802 } 1803 1804 static void gen_special_dispatch(MacroAssembler* masm, 1805 const methodHandle& method, 1806 const BasicType* sig_bt, 1807 const VMRegPair* regs) { 1808 verify_oop_args(masm, method, sig_bt, regs); 1809 vmIntrinsics::ID iid = method->intrinsic_id(); 1810 1811 // Now write the args into the outgoing interpreter space 1812 bool has_receiver = false; 1813 Register receiver_reg = noreg; 1814 int member_arg_pos = -1; 1815 Register member_reg = noreg; 1816 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1817 if (ref_kind != 0) { 1818 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1819 member_reg = rbx; // known to be free at this point 1820 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1821 } else if (iid == vmIntrinsics::_invokeBasic) { 1822 has_receiver = true; 1823 } else { 1824 fatal("unexpected intrinsic id %d", iid); 1825 } 1826 1827 if (member_reg != noreg) { 1828 // Load the member_arg into register, if necessary. 1829 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1830 VMReg r = regs[member_arg_pos].first(); 1831 if (r->is_stack()) { 1832 __ movptr(member_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1833 } else { 1834 // no data motion is needed 1835 member_reg = r->as_Register(); 1836 } 1837 } 1838 1839 if (has_receiver) { 1840 // Make sure the receiver is loaded into a register. 1841 assert(method->size_of_parameters() > 0, "oob"); 1842 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1843 VMReg r = regs[0].first(); 1844 assert(r->is_valid(), "bad receiver arg"); 1845 if (r->is_stack()) { 1846 // Porting note: This assumes that compiled calling conventions always 1847 // pass the receiver oop in a register. If this is not true on some 1848 // platform, pick a temp and load the receiver from stack. 1849 fatal("receiver always in a register"); 1850 receiver_reg = j_rarg0; // known to be free at this point 1851 __ movptr(receiver_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1852 } else { 1853 // no data motion is needed 1854 receiver_reg = r->as_Register(); 1855 } 1856 } 1857 1858 // Figure out which address we are really jumping to: 1859 MethodHandles::generate_method_handle_dispatch(masm, iid, 1860 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 1861 } 1862 1863 // --------------------------------------------------------------------------- 1864 // Generate a native wrapper for a given method. The method takes arguments 1865 // in the Java compiled code convention, marshals them to the native 1866 // convention (handlizes oops, etc), transitions to native, makes the call, 1867 // returns to java state (possibly blocking), unhandlizes any result and 1868 // returns. 1869 // 1870 // Critical native functions are a shorthand for the use of 1871 // GetPrimtiveArrayCritical and disallow the use of any other JNI 1872 // functions. The wrapper is expected to unpack the arguments before 1873 // passing them to the callee and perform checks before and after the 1874 // native call to ensure that they GCLocker 1875 // lock_critical/unlock_critical semantics are followed. Some other 1876 // parts of JNI setup are skipped like the tear down of the JNI handle 1877 // block and the check for pending exceptions it's impossible for them 1878 // to be thrown. 1879 // 1880 // They are roughly structured like this: 1881 // if (GCLocker::needs_gc()) 1882 // SharedRuntime::block_for_jni_critical(); 1883 // tranistion to thread_in_native 1884 // unpack arrray arguments and call native entry point 1885 // check for safepoint in progress 1886 // check if any thread suspend flags are set 1887 // call into JVM and possible unlock the JNI critical 1888 // if a GC was suppressed while in the critical native. 1889 // transition back to thread_in_Java 1890 // return to caller 1891 // 1892 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1893 const methodHandle& method, 1894 int compile_id, 1895 BasicType* in_sig_bt, 1896 VMRegPair* in_regs, 1897 BasicType ret_type) { 1898 if (method->is_method_handle_intrinsic()) { 1899 vmIntrinsics::ID iid = method->intrinsic_id(); 1900 intptr_t start = (intptr_t)__ pc(); 1901 int vep_offset = ((intptr_t)__ pc()) - start; 1902 gen_special_dispatch(masm, 1903 method, 1904 in_sig_bt, 1905 in_regs); 1906 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 1907 __ flush(); 1908 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 1909 return nmethod::new_native_nmethod(method, 1910 compile_id, 1911 masm->code(), 1912 vep_offset, 1913 frame_complete, 1914 stack_slots / VMRegImpl::slots_per_word, 1915 in_ByteSize(-1), 1916 in_ByteSize(-1), 1917 (OopMapSet*)NULL); 1918 } 1919 bool is_critical_native = true; 1920 address native_func = method->critical_native_function(); 1921 if (native_func == NULL) { 1922 native_func = method->native_function(); 1923 is_critical_native = false; 1924 } 1925 assert(native_func != NULL, "must have function"); 1926 1927 // An OopMap for lock (and class if static) 1928 OopMapSet *oop_maps = new OopMapSet(); 1929 intptr_t start = (intptr_t)__ pc(); 1930 1931 // We have received a description of where all the java arg are located 1932 // on entry to the wrapper. We need to convert these args to where 1933 // the jni function will expect them. To figure out where they go 1934 // we convert the java signature to a C signature by inserting 1935 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1936 1937 const int total_in_args = method->size_of_parameters(); 1938 int total_c_args = total_in_args; 1939 if (!is_critical_native) { 1940 total_c_args += 1; 1941 if (method->is_static()) { 1942 total_c_args++; 1943 } 1944 } else { 1945 for (int i = 0; i < total_in_args; i++) { 1946 if (in_sig_bt[i] == T_ARRAY) { 1947 total_c_args++; 1948 } 1949 } 1950 } 1951 1952 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1953 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1954 BasicType* in_elem_bt = NULL; 1955 1956 int argc = 0; 1957 if (!is_critical_native) { 1958 out_sig_bt[argc++] = T_ADDRESS; 1959 if (method->is_static()) { 1960 out_sig_bt[argc++] = T_OBJECT; 1961 } 1962 1963 for (int i = 0; i < total_in_args ; i++ ) { 1964 out_sig_bt[argc++] = in_sig_bt[i]; 1965 } 1966 } else { 1967 Thread* THREAD = Thread::current(); 1968 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args); 1969 SignatureStream ss(method->signature()); 1970 for (int i = 0; i < total_in_args ; i++ ) { 1971 if (in_sig_bt[i] == T_ARRAY) { 1972 // Arrays are passed as int, elem* pair 1973 out_sig_bt[argc++] = T_INT; 1974 out_sig_bt[argc++] = T_ADDRESS; 1975 Symbol* atype = ss.as_symbol(CHECK_NULL); 1976 const char* at = atype->as_C_string(); 1977 if (strlen(at) == 2) { 1978 assert(at[0] == '[', "must be"); 1979 switch (at[1]) { 1980 case 'B': in_elem_bt[i] = T_BYTE; break; 1981 case 'C': in_elem_bt[i] = T_CHAR; break; 1982 case 'D': in_elem_bt[i] = T_DOUBLE; break; 1983 case 'F': in_elem_bt[i] = T_FLOAT; break; 1984 case 'I': in_elem_bt[i] = T_INT; break; 1985 case 'J': in_elem_bt[i] = T_LONG; break; 1986 case 'S': in_elem_bt[i] = T_SHORT; break; 1987 case 'Z': in_elem_bt[i] = T_BOOLEAN; break; 1988 default: ShouldNotReachHere(); 1989 } 1990 } 1991 } else { 1992 out_sig_bt[argc++] = in_sig_bt[i]; 1993 in_elem_bt[i] = T_VOID; 1994 } 1995 if (in_sig_bt[i] != T_VOID) { 1996 assert(in_sig_bt[i] == ss.type(), "must match"); 1997 ss.next(); 1998 } 1999 } 2000 } 2001 2002 // Now figure out where the args must be stored and how much stack space 2003 // they require. 2004 int out_arg_slots; 2005 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args); 2006 2007 // Compute framesize for the wrapper. We need to handlize all oops in 2008 // incoming registers 2009 2010 // Calculate the total number of stack slots we will need. 2011 2012 // First count the abi requirement plus all of the outgoing args 2013 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 2014 2015 // Now the space for the inbound oop handle area 2016 int total_save_slots = 6 * VMRegImpl::slots_per_word; // 6 arguments passed in registers 2017 if (is_critical_native) { 2018 // Critical natives may have to call out so they need a save area 2019 // for register arguments. 2020 int double_slots = 0; 2021 int single_slots = 0; 2022 for ( int i = 0; i < total_in_args; i++) { 2023 if (in_regs[i].first()->is_Register()) { 2024 const Register reg = in_regs[i].first()->as_Register(); 2025 switch (in_sig_bt[i]) { 2026 case T_BOOLEAN: 2027 case T_BYTE: 2028 case T_SHORT: 2029 case T_CHAR: 2030 case T_INT: single_slots++; break; 2031 case T_ARRAY: // specific to LP64 (7145024) 2032 case T_LONG: double_slots++; break; 2033 default: ShouldNotReachHere(); 2034 } 2035 } else if (in_regs[i].first()->is_XMMRegister()) { 2036 switch (in_sig_bt[i]) { 2037 case T_FLOAT: single_slots++; break; 2038 case T_DOUBLE: double_slots++; break; 2039 default: ShouldNotReachHere(); 2040 } 2041 } else if (in_regs[i].first()->is_FloatRegister()) { 2042 ShouldNotReachHere(); 2043 } 2044 } 2045 total_save_slots = double_slots * 2 + single_slots; 2046 // align the save area 2047 if (double_slots != 0) { 2048 stack_slots = align_up(stack_slots, 2); 2049 } 2050 } 2051 2052 int oop_handle_offset = stack_slots; 2053 stack_slots += total_save_slots; 2054 2055 // Now any space we need for handlizing a klass if static method 2056 2057 int klass_slot_offset = 0; 2058 int klass_offset = -1; 2059 int lock_slot_offset = 0; 2060 bool is_static = false; 2061 2062 if (method->is_static()) { 2063 klass_slot_offset = stack_slots; 2064 stack_slots += VMRegImpl::slots_per_word; 2065 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 2066 is_static = true; 2067 } 2068 2069 // Plus a lock if needed 2070 2071 if (method->is_synchronized()) { 2072 lock_slot_offset = stack_slots; 2073 stack_slots += VMRegImpl::slots_per_word; 2074 } 2075 2076 // Now a place (+2) to save return values or temp during shuffling 2077 // + 4 for return address (which we own) and saved rbp 2078 stack_slots += 6; 2079 2080 // Ok The space we have allocated will look like: 2081 // 2082 // 2083 // FP-> | | 2084 // |---------------------| 2085 // | 2 slots for moves | 2086 // |---------------------| 2087 // | lock box (if sync) | 2088 // |---------------------| <- lock_slot_offset 2089 // | klass (if static) | 2090 // |---------------------| <- klass_slot_offset 2091 // | oopHandle area | 2092 // |---------------------| <- oop_handle_offset (6 java arg registers) 2093 // | outbound memory | 2094 // | based arguments | 2095 // | | 2096 // |---------------------| 2097 // | | 2098 // SP-> | out_preserved_slots | 2099 // 2100 // 2101 2102 2103 // Now compute actual number of stack words we need rounding to make 2104 // stack properly aligned. 2105 stack_slots = align_up(stack_slots, StackAlignmentInSlots); 2106 2107 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 2108 2109 // First thing make an ic check to see if we should even be here 2110 2111 // We are free to use all registers as temps without saving them and 2112 // restoring them except rbp. rbp is the only callee save register 2113 // as far as the interpreter and the compiler(s) are concerned. 2114 2115 2116 const Register ic_reg = rax; 2117 const Register receiver = j_rarg0; 2118 2119 Label hit; 2120 Label exception_pending; 2121 2122 assert_different_registers(ic_reg, receiver, rscratch1); 2123 __ verify_oop(receiver); 2124 __ load_klass(rscratch1, receiver); 2125 __ cmpq(ic_reg, rscratch1); 2126 __ jcc(Assembler::equal, hit); 2127 2128 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 2129 2130 // Verified entry point must be aligned 2131 __ align(8); 2132 2133 __ bind(hit); 2134 2135 int vep_offset = ((intptr_t)__ pc()) - start; 2136 2137 #ifdef COMPILER1 2138 // For Object.hashCode, System.identityHashCode try to pull hashCode from object header if available. 2139 if ((InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) || (method->intrinsic_id() == vmIntrinsics::_identityHashCode)) { 2140 inline_check_hashcode_from_object_header(masm, method, j_rarg0 /*obj_reg*/, rax /*result*/); 2141 } 2142 #endif // COMPILER1 2143 2144 // The instruction at the verified entry point must be 5 bytes or longer 2145 // because it can be patched on the fly by make_non_entrant. The stack bang 2146 // instruction fits that requirement. 2147 2148 // Generate stack overflow check 2149 2150 if (UseStackBanging) { 2151 __ bang_stack_with_offset((int)JavaThread::stack_shadow_zone_size()); 2152 } else { 2153 // need a 5 byte instruction to allow MT safe patching to non-entrant 2154 __ fat_nop(); 2155 } 2156 2157 // Generate a new frame for the wrapper. 2158 __ enter(); 2159 // -2 because return address is already present and so is saved rbp 2160 __ subptr(rsp, stack_size - 2*wordSize); 2161 2162 // Frame is now completed as far as size and linkage. 2163 int frame_complete = ((intptr_t)__ pc()) - start; 2164 2165 if (UseRTMLocking) { 2166 // Abort RTM transaction before calling JNI 2167 // because critical section will be large and will be 2168 // aborted anyway. Also nmethod could be deoptimized. 2169 __ xabort(0); 2170 } 2171 2172 #ifdef ASSERT 2173 { 2174 Label L; 2175 __ mov(rax, rsp); 2176 __ andptr(rax, -16); // must be 16 byte boundary (see amd64 ABI) 2177 __ cmpptr(rax, rsp); 2178 __ jcc(Assembler::equal, L); 2179 __ stop("improperly aligned stack"); 2180 __ bind(L); 2181 } 2182 #endif /* ASSERT */ 2183 2184 2185 // We use r14 as the oop handle for the receiver/klass 2186 // It is callee save so it survives the call to native 2187 2188 const Register oop_handle_reg = r14; 2189 2190 if (is_critical_native && !Universe::heap()->supports_object_pinning()) { 2191 check_needs_gc_for_critical_native(masm, stack_slots, total_c_args, total_in_args, 2192 oop_handle_offset, oop_maps, in_regs, in_sig_bt); 2193 } 2194 2195 // 2196 // We immediately shuffle the arguments so that any vm call we have to 2197 // make from here on out (sync slow path, jvmti, etc.) we will have 2198 // captured the oops from our caller and have a valid oopMap for 2199 // them. 2200 2201 // ----------------- 2202 // The Grand Shuffle 2203 2204 // The Java calling convention is either equal (linux) or denser (win64) than the 2205 // c calling convention. However the because of the jni_env argument the c calling 2206 // convention always has at least one more (and two for static) arguments than Java. 2207 // Therefore if we move the args from java -> c backwards then we will never have 2208 // a register->register conflict and we don't have to build a dependency graph 2209 // and figure out how to break any cycles. 2210 // 2211 2212 // Record esp-based slot for receiver on stack for non-static methods 2213 int receiver_offset = -1; 2214 2215 // This is a trick. We double the stack slots so we can claim 2216 // the oops in the caller's frame. Since we are sure to have 2217 // more args than the caller doubling is enough to make 2218 // sure we can capture all the incoming oop args from the 2219 // caller. 2220 // 2221 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 2222 2223 // Mark location of rbp (someday) 2224 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rbp)); 2225 2226 // Use eax, ebx as temporaries during any memory-memory moves we have to do 2227 // All inbound args are referenced based on rbp and all outbound args via rsp. 2228 2229 2230 #ifdef ASSERT 2231 bool reg_destroyed[RegisterImpl::number_of_registers]; 2232 bool freg_destroyed[XMMRegisterImpl::number_of_registers]; 2233 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) { 2234 reg_destroyed[r] = false; 2235 } 2236 for ( int f = 0 ; f < XMMRegisterImpl::number_of_registers ; f++ ) { 2237 freg_destroyed[f] = false; 2238 } 2239 2240 #endif /* ASSERT */ 2241 2242 // This may iterate in two different directions depending on the 2243 // kind of native it is. The reason is that for regular JNI natives 2244 // the incoming and outgoing registers are offset upwards and for 2245 // critical natives they are offset down. 2246 GrowableArray<int> arg_order(2 * total_in_args); 2247 // Inbound arguments that need to be pinned for critical natives 2248 GrowableArray<int> pinned_args(total_in_args); 2249 // Current stack slot for storing register based array argument 2250 int pinned_slot = oop_handle_offset; 2251 2252 VMRegPair tmp_vmreg; 2253 tmp_vmreg.set2(rbx->as_VMReg()); 2254 2255 if (!is_critical_native) { 2256 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 2257 arg_order.push(i); 2258 arg_order.push(c_arg); 2259 } 2260 } else { 2261 // Compute a valid move order, using tmp_vmreg to break any cycles 2262 ComputeMoveOrder cmo(total_in_args, in_regs, total_c_args, out_regs, in_sig_bt, arg_order, tmp_vmreg); 2263 } 2264 2265 int temploc = -1; 2266 for (int ai = 0; ai < arg_order.length(); ai += 2) { 2267 int i = arg_order.at(ai); 2268 int c_arg = arg_order.at(ai + 1); 2269 __ block_comment(err_msg("move %d -> %d", i, c_arg)); 2270 if (c_arg == -1) { 2271 assert(is_critical_native, "should only be required for critical natives"); 2272 // This arg needs to be moved to a temporary 2273 __ mov(tmp_vmreg.first()->as_Register(), in_regs[i].first()->as_Register()); 2274 in_regs[i] = tmp_vmreg; 2275 temploc = i; 2276 continue; 2277 } else if (i == -1) { 2278 assert(is_critical_native, "should only be required for critical natives"); 2279 // Read from the temporary location 2280 assert(temploc != -1, "must be valid"); 2281 i = temploc; 2282 temploc = -1; 2283 } 2284 #ifdef ASSERT 2285 if (in_regs[i].first()->is_Register()) { 2286 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!"); 2287 } else if (in_regs[i].first()->is_XMMRegister()) { 2288 assert(!freg_destroyed[in_regs[i].first()->as_XMMRegister()->encoding()], "destroyed reg!"); 2289 } 2290 if (out_regs[c_arg].first()->is_Register()) { 2291 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 2292 } else if (out_regs[c_arg].first()->is_XMMRegister()) { 2293 freg_destroyed[out_regs[c_arg].first()->as_XMMRegister()->encoding()] = true; 2294 } 2295 #endif /* ASSERT */ 2296 switch (in_sig_bt[i]) { 2297 case T_ARRAY: 2298 if (is_critical_native) { 2299 // pin before unpack 2300 if (Universe::heap()->supports_object_pinning()) { 2301 save_args(masm, total_c_args, 0, out_regs); 2302 gen_pin_object(masm, in_regs[i]); 2303 pinned_args.append(i); 2304 restore_args(masm, total_c_args, 0, out_regs); 2305 2306 // rax has pinned array 2307 VMRegPair result_reg; 2308 result_reg.set_ptr(rax->as_VMReg()); 2309 move_ptr(masm, result_reg, in_regs[i]); 2310 if (!in_regs[i].first()->is_stack()) { 2311 assert(pinned_slot <= stack_slots, "overflow"); 2312 move_ptr(masm, result_reg, VMRegImpl::stack2reg(pinned_slot)); 2313 pinned_slot += VMRegImpl::slots_per_word; 2314 } 2315 } 2316 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]); 2317 c_arg++; 2318 #ifdef ASSERT 2319 if (out_regs[c_arg].first()->is_Register()) { 2320 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 2321 } else if (out_regs[c_arg].first()->is_XMMRegister()) { 2322 freg_destroyed[out_regs[c_arg].first()->as_XMMRegister()->encoding()] = true; 2323 } 2324 #endif 2325 break; 2326 } 2327 case T_OBJECT: 2328 assert(!is_critical_native, "no oop arguments"); 2329 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 2330 ((i == 0) && (!is_static)), 2331 &receiver_offset); 2332 break; 2333 case T_VOID: 2334 break; 2335 2336 case T_FLOAT: 2337 float_move(masm, in_regs[i], out_regs[c_arg]); 2338 break; 2339 2340 case T_DOUBLE: 2341 assert( i + 1 < total_in_args && 2342 in_sig_bt[i + 1] == T_VOID && 2343 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 2344 double_move(masm, in_regs[i], out_regs[c_arg]); 2345 break; 2346 2347 case T_LONG : 2348 long_move(masm, in_regs[i], out_regs[c_arg]); 2349 break; 2350 2351 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 2352 2353 default: 2354 move32_64(masm, in_regs[i], out_regs[c_arg]); 2355 } 2356 } 2357 2358 int c_arg; 2359 2360 // Pre-load a static method's oop into r14. Used both by locking code and 2361 // the normal JNI call code. 2362 if (!is_critical_native) { 2363 // point c_arg at the first arg that is already loaded in case we 2364 // need to spill before we call out 2365 c_arg = total_c_args - total_in_args; 2366 2367 if (method->is_static()) { 2368 2369 // load oop into a register 2370 __ movoop(oop_handle_reg, JNIHandles::make_local(method->method_holder()->java_mirror())); 2371 2372 // Now handlize the static class mirror it's known not-null. 2373 __ movptr(Address(rsp, klass_offset), oop_handle_reg); 2374 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 2375 2376 // Now get the handle 2377 __ lea(oop_handle_reg, Address(rsp, klass_offset)); 2378 // store the klass handle as second argument 2379 __ movptr(c_rarg1, oop_handle_reg); 2380 // and protect the arg if we must spill 2381 c_arg--; 2382 } 2383 } else { 2384 // For JNI critical methods we need to save all registers in save_args. 2385 c_arg = 0; 2386 } 2387 2388 // Change state to native (we save the return address in the thread, since it might not 2389 // be pushed on the stack when we do a a stack traversal). It is enough that the pc() 2390 // points into the right code segment. It does not have to be the correct return pc. 2391 // We use the same pc/oopMap repeatedly when we call out 2392 2393 intptr_t the_pc = (intptr_t) __ pc(); 2394 oop_maps->add_gc_map(the_pc - start, map); 2395 2396 __ set_last_Java_frame(rsp, noreg, (address)the_pc); 2397 2398 2399 // We have all of the arguments setup at this point. We must not touch any register 2400 // argument registers at this point (what if we save/restore them there are no oop? 2401 2402 { 2403 SkipIfEqual skip(masm, &DTraceMethodProbes, false); 2404 // protect the args we've loaded 2405 save_args(masm, total_c_args, c_arg, out_regs); 2406 __ mov_metadata(c_rarg1, method()); 2407 __ call_VM_leaf( 2408 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 2409 r15_thread, c_rarg1); 2410 restore_args(masm, total_c_args, c_arg, out_regs); 2411 } 2412 2413 // RedefineClasses() tracing support for obsolete method entry 2414 if (log_is_enabled(Trace, redefine, class, obsolete)) { 2415 // protect the args we've loaded 2416 save_args(masm, total_c_args, c_arg, out_regs); 2417 __ mov_metadata(c_rarg1, method()); 2418 __ call_VM_leaf( 2419 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 2420 r15_thread, c_rarg1); 2421 restore_args(masm, total_c_args, c_arg, out_regs); 2422 } 2423 2424 // Lock a synchronized method 2425 2426 // Register definitions used by locking and unlocking 2427 2428 const Register swap_reg = rax; // Must use rax for cmpxchg instruction 2429 const Register obj_reg = rbx; // Will contain the oop 2430 const Register lock_reg = r13; // Address of compiler lock object (BasicLock) 2431 const Register old_hdr = r13; // value of old header at unlock time 2432 2433 Label slow_path_lock; 2434 Label lock_done; 2435 2436 if (method->is_synchronized()) { 2437 assert(!is_critical_native, "unhandled"); 2438 2439 2440 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 2441 2442 // Get the handle (the 2nd argument) 2443 __ mov(oop_handle_reg, c_rarg1); 2444 2445 // Get address of the box 2446 2447 __ lea(lock_reg, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2448 2449 // Load the oop from the handle 2450 __ movptr(obj_reg, Address(oop_handle_reg, 0)); 2451 2452 if (UseBiasedLocking) { 2453 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, rscratch1, false, lock_done, &slow_path_lock); 2454 } 2455 2456 // Load immediate 1 into swap_reg %rax 2457 __ movl(swap_reg, 1); 2458 2459 // Load (object->mark() | 1) into swap_reg %rax 2460 __ orptr(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 2461 2462 // Save (object->mark() | 1) into BasicLock's displaced header 2463 __ movptr(Address(lock_reg, mark_word_offset), swap_reg); 2464 2465 if (os::is_MP()) { 2466 __ lock(); 2467 } 2468 2469 // src -> dest iff dest == rax else rax <- dest 2470 __ cmpxchgptr(lock_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 2471 __ jcc(Assembler::equal, lock_done); 2472 2473 // Hmm should this move to the slow path code area??? 2474 2475 // Test if the oopMark is an obvious stack pointer, i.e., 2476 // 1) (mark & 3) == 0, and 2477 // 2) rsp <= mark < mark + os::pagesize() 2478 // These 3 tests can be done by evaluating the following 2479 // expression: ((mark - rsp) & (3 - os::vm_page_size())), 2480 // assuming both stack pointer and pagesize have their 2481 // least significant 2 bits clear. 2482 // NOTE: the oopMark is in swap_reg %rax as the result of cmpxchg 2483 2484 __ subptr(swap_reg, rsp); 2485 __ andptr(swap_reg, 3 - os::vm_page_size()); 2486 2487 // Save the test result, for recursive case, the result is zero 2488 __ movptr(Address(lock_reg, mark_word_offset), swap_reg); 2489 __ jcc(Assembler::notEqual, slow_path_lock); 2490 2491 // Slow path will re-enter here 2492 2493 __ bind(lock_done); 2494 } 2495 2496 2497 // Finally just about ready to make the JNI call 2498 2499 2500 // get JNIEnv* which is first argument to native 2501 if (!is_critical_native) { 2502 __ lea(c_rarg0, Address(r15_thread, in_bytes(JavaThread::jni_environment_offset()))); 2503 } 2504 2505 // Now set thread in native 2506 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_native); 2507 2508 __ call(RuntimeAddress(native_func)); 2509 2510 // Verify or restore cpu control state after JNI call 2511 __ restore_cpu_control_state_after_jni(); 2512 2513 // Unpack native results. 2514 switch (ret_type) { 2515 case T_BOOLEAN: __ c2bool(rax); break; 2516 case T_CHAR : __ movzwl(rax, rax); break; 2517 case T_BYTE : __ sign_extend_byte (rax); break; 2518 case T_SHORT : __ sign_extend_short(rax); break; 2519 case T_INT : /* nothing to do */ break; 2520 case T_DOUBLE : 2521 case T_FLOAT : 2522 // Result is in xmm0 we'll save as needed 2523 break; 2524 case T_ARRAY: // Really a handle 2525 case T_OBJECT: // Really a handle 2526 break; // can't de-handlize until after safepoint check 2527 case T_VOID: break; 2528 case T_LONG: break; 2529 default : ShouldNotReachHere(); 2530 } 2531 2532 // unpin pinned arguments 2533 pinned_slot = oop_handle_offset; 2534 if (pinned_args.length() > 0) { 2535 // save return value that may be overwritten otherwise. 2536 save_native_result(masm, ret_type, stack_slots); 2537 for (int index = 0; index < pinned_args.length(); index ++) { 2538 int i = pinned_args.at(index); 2539 assert(pinned_slot <= stack_slots, "overflow"); 2540 if (!in_regs[i].first()->is_stack()) { 2541 int offset = pinned_slot * VMRegImpl::stack_slot_size; 2542 __ movq(in_regs[i].first()->as_Register(), Address(rsp, offset)); 2543 pinned_slot += VMRegImpl::slots_per_word; 2544 } 2545 gen_unpin_object(masm, in_regs[i]); 2546 } 2547 restore_native_result(masm, ret_type, stack_slots); 2548 } 2549 2550 // Switch thread to "native transition" state before reading the synchronization state. 2551 // This additional state is necessary because reading and testing the synchronization 2552 // state is not atomic w.r.t. GC, as this scenario demonstrates: 2553 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 2554 // VM thread changes sync state to synchronizing and suspends threads for GC. 2555 // Thread A is resumed to finish this native method, but doesn't block here since it 2556 // didn't see any synchronization is progress, and escapes. 2557 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_native_trans); 2558 2559 if(os::is_MP()) { 2560 if (UseMembar) { 2561 // Force this write out before the read below 2562 __ membar(Assembler::Membar_mask_bits( 2563 Assembler::LoadLoad | Assembler::LoadStore | 2564 Assembler::StoreLoad | Assembler::StoreStore)); 2565 } else { 2566 // Write serialization page so VM thread can do a pseudo remote membar. 2567 // We use the current thread pointer to calculate a thread specific 2568 // offset to write to within the page. This minimizes bus traffic 2569 // due to cache line collision. 2570 __ serialize_memory(r15_thread, rcx); 2571 } 2572 } 2573 2574 Label after_transition; 2575 2576 // check for safepoint operation in progress and/or pending suspend requests 2577 { 2578 Label Continue; 2579 Label slow_path; 2580 2581 __ safepoint_poll(slow_path, r15_thread, rscratch1); 2582 2583 __ cmpl(Address(r15_thread, JavaThread::suspend_flags_offset()), 0); 2584 __ jcc(Assembler::equal, Continue); 2585 __ bind(slow_path); 2586 2587 // Don't use call_VM as it will see a possible pending exception and forward it 2588 // and never return here preventing us from clearing _last_native_pc down below. 2589 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are 2590 // preserved and correspond to the bcp/locals pointers. So we do a runtime call 2591 // by hand. 2592 // 2593 __ vzeroupper(); 2594 save_native_result(masm, ret_type, stack_slots); 2595 __ mov(c_rarg0, r15_thread); 2596 __ mov(r12, rsp); // remember sp 2597 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 2598 __ andptr(rsp, -16); // align stack as required by ABI 2599 if (!is_critical_native) { 2600 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans))); 2601 } else { 2602 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition))); 2603 } 2604 __ mov(rsp, r12); // restore sp 2605 __ reinit_heapbase(); 2606 // Restore any method result value 2607 restore_native_result(masm, ret_type, stack_slots); 2608 2609 if (is_critical_native) { 2610 // The call above performed the transition to thread_in_Java so 2611 // skip the transition logic below. 2612 __ jmpb(after_transition); 2613 } 2614 2615 __ bind(Continue); 2616 } 2617 2618 // change thread state 2619 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_Java); 2620 __ bind(after_transition); 2621 2622 Label reguard; 2623 Label reguard_done; 2624 __ cmpl(Address(r15_thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_reserved_disabled); 2625 __ jcc(Assembler::equal, reguard); 2626 __ bind(reguard_done); 2627 2628 // native result if any is live 2629 2630 // Unlock 2631 Label unlock_done; 2632 Label slow_path_unlock; 2633 if (method->is_synchronized()) { 2634 2635 // Get locked oop from the handle we passed to jni 2636 __ movptr(obj_reg, Address(oop_handle_reg, 0)); 2637 2638 Label done; 2639 2640 if (UseBiasedLocking) { 2641 __ biased_locking_exit(obj_reg, old_hdr, done); 2642 } 2643 2644 // Simple recursive lock? 2645 2646 __ cmpptr(Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size), (int32_t)NULL_WORD); 2647 __ jcc(Assembler::equal, done); 2648 2649 // Must save rax if if it is live now because cmpxchg must use it 2650 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2651 save_native_result(masm, ret_type, stack_slots); 2652 } 2653 2654 2655 // get address of the stack lock 2656 __ lea(rax, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2657 // get old displaced header 2658 __ movptr(old_hdr, Address(rax, 0)); 2659 2660 // Atomic swap old header if oop still contains the stack lock 2661 if (os::is_MP()) { 2662 __ lock(); 2663 } 2664 __ cmpxchgptr(old_hdr, Address(obj_reg, oopDesc::mark_offset_in_bytes())); 2665 __ jcc(Assembler::notEqual, slow_path_unlock); 2666 2667 // slow path re-enters here 2668 __ bind(unlock_done); 2669 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2670 restore_native_result(masm, ret_type, stack_slots); 2671 } 2672 2673 __ bind(done); 2674 2675 } 2676 { 2677 SkipIfEqual skip(masm, &DTraceMethodProbes, false); 2678 save_native_result(masm, ret_type, stack_slots); 2679 __ mov_metadata(c_rarg1, method()); 2680 __ call_VM_leaf( 2681 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2682 r15_thread, c_rarg1); 2683 restore_native_result(masm, ret_type, stack_slots); 2684 } 2685 2686 __ reset_last_Java_frame(false); 2687 2688 // Unbox oop result, e.g. JNIHandles::resolve value. 2689 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2690 __ resolve_jobject(rax /* value */, 2691 r15_thread /* thread */, 2692 rcx /* tmp */); 2693 } 2694 2695 if (CheckJNICalls) { 2696 // clear_pending_jni_exception_check 2697 __ movptr(Address(r15_thread, JavaThread::pending_jni_exception_check_fn_offset()), NULL_WORD); 2698 } 2699 2700 if (!is_critical_native) { 2701 // reset handle block 2702 __ movptr(rcx, Address(r15_thread, JavaThread::active_handles_offset())); 2703 __ movl(Address(rcx, JNIHandleBlock::top_offset_in_bytes()), (int32_t)NULL_WORD); 2704 } 2705 2706 // pop our frame 2707 2708 __ leave(); 2709 2710 if (!is_critical_native) { 2711 // Any exception pending? 2712 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2713 __ jcc(Assembler::notEqual, exception_pending); 2714 } 2715 2716 // Return 2717 2718 __ ret(0); 2719 2720 // Unexpected paths are out of line and go here 2721 2722 if (!is_critical_native) { 2723 // forward the exception 2724 __ bind(exception_pending); 2725 2726 // and forward the exception 2727 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2728 } 2729 2730 // Slow path locking & unlocking 2731 if (method->is_synchronized()) { 2732 2733 // BEGIN Slow path lock 2734 __ bind(slow_path_lock); 2735 2736 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 2737 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2738 2739 // protect the args we've loaded 2740 save_args(masm, total_c_args, c_arg, out_regs); 2741 2742 __ mov(c_rarg0, obj_reg); 2743 __ mov(c_rarg1, lock_reg); 2744 __ mov(c_rarg2, r15_thread); 2745 2746 // Not a leaf but we have last_Java_frame setup as we want 2747 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3); 2748 restore_args(masm, total_c_args, c_arg, out_regs); 2749 2750 #ifdef ASSERT 2751 { Label L; 2752 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2753 __ jcc(Assembler::equal, L); 2754 __ stop("no pending exception allowed on exit from monitorenter"); 2755 __ bind(L); 2756 } 2757 #endif 2758 __ jmp(lock_done); 2759 2760 // END Slow path lock 2761 2762 // BEGIN Slow path unlock 2763 __ bind(slow_path_unlock); 2764 2765 // If we haven't already saved the native result we must save it now as xmm registers 2766 // are still exposed. 2767 __ vzeroupper(); 2768 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2769 save_native_result(masm, ret_type, stack_slots); 2770 } 2771 2772 __ lea(c_rarg1, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2773 2774 __ mov(c_rarg0, obj_reg); 2775 __ mov(c_rarg2, r15_thread); 2776 __ mov(r12, rsp); // remember sp 2777 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 2778 __ andptr(rsp, -16); // align stack as required by ABI 2779 2780 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 2781 // NOTE that obj_reg == rbx currently 2782 __ movptr(rbx, Address(r15_thread, in_bytes(Thread::pending_exception_offset()))); 2783 __ movptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2784 2785 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2786 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C))); 2787 __ mov(rsp, r12); // restore sp 2788 __ reinit_heapbase(); 2789 #ifdef ASSERT 2790 { 2791 Label L; 2792 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), (int)NULL_WORD); 2793 __ jcc(Assembler::equal, L); 2794 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 2795 __ bind(L); 2796 } 2797 #endif /* ASSERT */ 2798 2799 __ movptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), rbx); 2800 2801 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2802 restore_native_result(masm, ret_type, stack_slots); 2803 } 2804 __ jmp(unlock_done); 2805 2806 // END Slow path unlock 2807 2808 } // synchronized 2809 2810 // SLOW PATH Reguard the stack if needed 2811 2812 __ bind(reguard); 2813 __ vzeroupper(); 2814 save_native_result(masm, ret_type, stack_slots); 2815 __ mov(r12, rsp); // remember sp 2816 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 2817 __ andptr(rsp, -16); // align stack as required by ABI 2818 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages))); 2819 __ mov(rsp, r12); // restore sp 2820 __ reinit_heapbase(); 2821 restore_native_result(masm, ret_type, stack_slots); 2822 // and continue 2823 __ jmp(reguard_done); 2824 2825 2826 2827 __ flush(); 2828 2829 nmethod *nm = nmethod::new_native_nmethod(method, 2830 compile_id, 2831 masm->code(), 2832 vep_offset, 2833 frame_complete, 2834 stack_slots / VMRegImpl::slots_per_word, 2835 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2836 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2837 oop_maps); 2838 2839 if (is_critical_native) { 2840 nm->set_lazy_critical_native(true); 2841 } 2842 2843 return nm; 2844 2845 } 2846 2847 // this function returns the adjust size (in number of words) to a c2i adapter 2848 // activation for use during deoptimization 2849 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals ) { 2850 return (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2851 } 2852 2853 2854 uint SharedRuntime::out_preserve_stack_slots() { 2855 return 0; 2856 } 2857 2858 //------------------------------generate_deopt_blob---------------------------- 2859 void SharedRuntime::generate_deopt_blob() { 2860 // Allocate space for the code 2861 ResourceMark rm; 2862 // Setup code generation tools 2863 int pad = 0; 2864 #if INCLUDE_JVMCI 2865 if (EnableJVMCI || UseAOT) { 2866 pad += 512; // Increase the buffer size when compiling for JVMCI 2867 } 2868 #endif 2869 CodeBuffer buffer("deopt_blob", 2048+pad, 1024); 2870 MacroAssembler* masm = new MacroAssembler(&buffer); 2871 int frame_size_in_words; 2872 OopMap* map = NULL; 2873 OopMapSet *oop_maps = new OopMapSet(); 2874 2875 // ------------- 2876 // This code enters when returning to a de-optimized nmethod. A return 2877 // address has been pushed on the the stack, and return values are in 2878 // registers. 2879 // If we are doing a normal deopt then we were called from the patched 2880 // nmethod from the point we returned to the nmethod. So the return 2881 // address on the stack is wrong by NativeCall::instruction_size 2882 // We will adjust the value so it looks like we have the original return 2883 // address on the stack (like when we eagerly deoptimized). 2884 // In the case of an exception pending when deoptimizing, we enter 2885 // with a return address on the stack that points after the call we patched 2886 // into the exception handler. We have the following register state from, 2887 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp). 2888 // rax: exception oop 2889 // rbx: exception handler 2890 // rdx: throwing pc 2891 // So in this case we simply jam rdx into the useless return address and 2892 // the stack looks just like we want. 2893 // 2894 // At this point we need to de-opt. We save the argument return 2895 // registers. We call the first C routine, fetch_unroll_info(). This 2896 // routine captures the return values and returns a structure which 2897 // describes the current frame size and the sizes of all replacement frames. 2898 // The current frame is compiled code and may contain many inlined 2899 // functions, each with their own JVM state. We pop the current frame, then 2900 // push all the new frames. Then we call the C routine unpack_frames() to 2901 // populate these frames. Finally unpack_frames() returns us the new target 2902 // address. Notice that callee-save registers are BLOWN here; they have 2903 // already been captured in the vframeArray at the time the return PC was 2904 // patched. 2905 address start = __ pc(); 2906 Label cont; 2907 2908 // Prolog for non exception case! 2909 2910 // Save everything in sight. 2911 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2912 2913 // Normal deoptimization. Save exec mode for unpack_frames. 2914 __ movl(r14, Deoptimization::Unpack_deopt); // callee-saved 2915 __ jmp(cont); 2916 2917 int reexecute_offset = __ pc() - start; 2918 #if INCLUDE_JVMCI && !defined(COMPILER1) 2919 if (EnableJVMCI && UseJVMCICompiler) { 2920 // JVMCI does not use this kind of deoptimization 2921 __ should_not_reach_here(); 2922 } 2923 #endif 2924 2925 // Reexecute case 2926 // return address is the pc describes what bci to do re-execute at 2927 2928 // No need to update map as each call to save_live_registers will produce identical oopmap 2929 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2930 2931 __ movl(r14, Deoptimization::Unpack_reexecute); // callee-saved 2932 __ jmp(cont); 2933 2934 #if INCLUDE_JVMCI 2935 Label after_fetch_unroll_info_call; 2936 int implicit_exception_uncommon_trap_offset = 0; 2937 int uncommon_trap_offset = 0; 2938 2939 if (EnableJVMCI || UseAOT) { 2940 implicit_exception_uncommon_trap_offset = __ pc() - start; 2941 2942 __ pushptr(Address(r15_thread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 2943 __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())), (int32_t)NULL_WORD); 2944 2945 uncommon_trap_offset = __ pc() - start; 2946 2947 // Save everything in sight. 2948 RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2949 // fetch_unroll_info needs to call last_java_frame() 2950 __ set_last_Java_frame(noreg, noreg, NULL); 2951 2952 __ movl(c_rarg1, Address(r15_thread, in_bytes(JavaThread::pending_deoptimization_offset()))); 2953 __ movl(Address(r15_thread, in_bytes(JavaThread::pending_deoptimization_offset())), -1); 2954 2955 __ movl(r14, (int32_t)Deoptimization::Unpack_reexecute); 2956 __ mov(c_rarg0, r15_thread); 2957 __ movl(c_rarg2, r14); // exec mode 2958 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap))); 2959 oop_maps->add_gc_map( __ pc()-start, map->deep_copy()); 2960 2961 __ reset_last_Java_frame(false); 2962 2963 __ jmp(after_fetch_unroll_info_call); 2964 } // EnableJVMCI 2965 #endif // INCLUDE_JVMCI 2966 2967 int exception_offset = __ pc() - start; 2968 2969 // Prolog for exception case 2970 2971 // all registers are dead at this entry point, except for rax, and 2972 // rdx which contain the exception oop and exception pc 2973 // respectively. Set them in TLS and fall thru to the 2974 // unpack_with_exception_in_tls entry point. 2975 2976 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), rdx); 2977 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()), rax); 2978 2979 int exception_in_tls_offset = __ pc() - start; 2980 2981 // new implementation because exception oop is now passed in JavaThread 2982 2983 // Prolog for exception case 2984 // All registers must be preserved because they might be used by LinearScan 2985 // Exceptiop oop and throwing PC are passed in JavaThread 2986 // tos: stack at point of call to method that threw the exception (i.e. only 2987 // args are on the stack, no return address) 2988 2989 // make room on stack for the return address 2990 // It will be patched later with the throwing pc. The correct value is not 2991 // available now because loading it from memory would destroy registers. 2992 __ push(0); 2993 2994 // Save everything in sight. 2995 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 2996 2997 // Now it is safe to overwrite any register 2998 2999 // Deopt during an exception. Save exec mode for unpack_frames. 3000 __ movl(r14, Deoptimization::Unpack_exception); // callee-saved 3001 3002 // load throwing pc from JavaThread and patch it as the return address 3003 // of the current frame. Then clear the field in JavaThread 3004 3005 __ movptr(rdx, Address(r15_thread, JavaThread::exception_pc_offset())); 3006 __ movptr(Address(rbp, wordSize), rdx); 3007 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), (int32_t)NULL_WORD); 3008 3009 #ifdef ASSERT 3010 // verify that there is really an exception oop in JavaThread 3011 __ movptr(rax, Address(r15_thread, JavaThread::exception_oop_offset())); 3012 __ verify_oop(rax); 3013 3014 // verify that there is no pending exception 3015 Label no_pending_exception; 3016 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset())); 3017 __ testptr(rax, rax); 3018 __ jcc(Assembler::zero, no_pending_exception); 3019 __ stop("must not have pending exception here"); 3020 __ bind(no_pending_exception); 3021 #endif 3022 3023 __ bind(cont); 3024 3025 // Call C code. Need thread and this frame, but NOT official VM entry 3026 // crud. We cannot block on this call, no GC can happen. 3027 // 3028 // UnrollBlock* fetch_unroll_info(JavaThread* thread) 3029 3030 // fetch_unroll_info needs to call last_java_frame(). 3031 3032 __ set_last_Java_frame(noreg, noreg, NULL); 3033 #ifdef ASSERT 3034 { Label L; 3035 __ cmpptr(Address(r15_thread, 3036 JavaThread::last_Java_fp_offset()), 3037 (int32_t)0); 3038 __ jcc(Assembler::equal, L); 3039 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 3040 __ bind(L); 3041 } 3042 #endif // ASSERT 3043 __ mov(c_rarg0, r15_thread); 3044 __ movl(c_rarg1, r14); // exec_mode 3045 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info))); 3046 3047 // Need to have an oopmap that tells fetch_unroll_info where to 3048 // find any register it might need. 3049 oop_maps->add_gc_map(__ pc() - start, map); 3050 3051 __ reset_last_Java_frame(false); 3052 3053 #if INCLUDE_JVMCI 3054 if (EnableJVMCI || UseAOT) { 3055 __ bind(after_fetch_unroll_info_call); 3056 } 3057 #endif 3058 3059 // Load UnrollBlock* into rdi 3060 __ mov(rdi, rax); 3061 3062 __ movl(r14, Address(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes())); 3063 Label noException; 3064 __ cmpl(r14, Deoptimization::Unpack_exception); // Was exception pending? 3065 __ jcc(Assembler::notEqual, noException); 3066 __ movptr(rax, Address(r15_thread, JavaThread::exception_oop_offset())); 3067 // QQQ this is useless it was NULL above 3068 __ movptr(rdx, Address(r15_thread, JavaThread::exception_pc_offset())); 3069 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()), (int32_t)NULL_WORD); 3070 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), (int32_t)NULL_WORD); 3071 3072 __ verify_oop(rax); 3073 3074 // Overwrite the result registers with the exception results. 3075 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax); 3076 // I think this is useless 3077 __ movptr(Address(rsp, RegisterSaver::rdx_offset_in_bytes()), rdx); 3078 3079 __ bind(noException); 3080 3081 // Only register save data is on the stack. 3082 // Now restore the result registers. Everything else is either dead 3083 // or captured in the vframeArray. 3084 RegisterSaver::restore_result_registers(masm); 3085 3086 // All of the register save area has been popped of the stack. Only the 3087 // return address remains. 3088 3089 // Pop all the frames we must move/replace. 3090 // 3091 // Frame picture (youngest to oldest) 3092 // 1: self-frame (no frame link) 3093 // 2: deopting frame (no frame link) 3094 // 3: caller of deopting frame (could be compiled/interpreted). 3095 // 3096 // Note: by leaving the return address of self-frame on the stack 3097 // and using the size of frame 2 to adjust the stack 3098 // when we are done the return to frame 3 will still be on the stack. 3099 3100 // Pop deoptimized frame 3101 __ movl(rcx, Address(rdi, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 3102 __ addptr(rsp, rcx); 3103 3104 // rsp should be pointing at the return address to the caller (3) 3105 3106 // Pick up the initial fp we should save 3107 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved) 3108 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 3109 3110 #ifdef ASSERT 3111 // Compilers generate code that bang the stack by as much as the 3112 // interpreter would need. So this stack banging should never 3113 // trigger a fault. Verify that it does not on non product builds. 3114 if (UseStackBanging) { 3115 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 3116 __ bang_stack_size(rbx, rcx); 3117 } 3118 #endif 3119 3120 // Load address of array of frame pcs into rcx 3121 __ movptr(rcx, Address(rdi, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 3122 3123 // Trash the old pc 3124 __ addptr(rsp, wordSize); 3125 3126 // Load address of array of frame sizes into rsi 3127 __ movptr(rsi, Address(rdi, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 3128 3129 // Load counter into rdx 3130 __ movl(rdx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 3131 3132 // Now adjust the caller's stack to make up for the extra locals 3133 // but record the original sp so that we can save it in the skeletal interpreter 3134 // frame and the stack walking of interpreter_sender will get the unextended sp 3135 // value and not the "real" sp value. 3136 3137 const Register sender_sp = r8; 3138 3139 __ mov(sender_sp, rsp); 3140 __ movl(rbx, Address(rdi, 3141 Deoptimization::UnrollBlock:: 3142 caller_adjustment_offset_in_bytes())); 3143 __ subptr(rsp, rbx); 3144 3145 // Push interpreter frames in a loop 3146 Label loop; 3147 __ bind(loop); 3148 __ movptr(rbx, Address(rsi, 0)); // Load frame size 3149 __ subptr(rbx, 2*wordSize); // We'll push pc and ebp by hand 3150 __ pushptr(Address(rcx, 0)); // Save return address 3151 __ enter(); // Save old & set new ebp 3152 __ subptr(rsp, rbx); // Prolog 3153 // This value is corrected by layout_activation_impl 3154 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), (int32_t)NULL_WORD ); 3155 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), sender_sp); // Make it walkable 3156 __ mov(sender_sp, rsp); // Pass sender_sp to next frame 3157 __ addptr(rsi, wordSize); // Bump array pointer (sizes) 3158 __ addptr(rcx, wordSize); // Bump array pointer (pcs) 3159 __ decrementl(rdx); // Decrement counter 3160 __ jcc(Assembler::notZero, loop); 3161 __ pushptr(Address(rcx, 0)); // Save final return address 3162 3163 // Re-push self-frame 3164 __ enter(); // Save old & set new ebp 3165 3166 // Allocate a full sized register save area. 3167 // Return address and rbp are in place, so we allocate two less words. 3168 __ subptr(rsp, (frame_size_in_words - 2) * wordSize); 3169 3170 // Restore frame locals after moving the frame 3171 __ movdbl(Address(rsp, RegisterSaver::xmm0_offset_in_bytes()), xmm0); 3172 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax); 3173 3174 // Call C code. Need thread but NOT official VM entry 3175 // crud. We cannot block on this call, no GC can happen. Call should 3176 // restore return values to their stack-slots with the new SP. 3177 // 3178 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode) 3179 3180 // Use rbp because the frames look interpreted now 3181 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 3182 // Don't need the precise return PC here, just precise enough to point into this code blob. 3183 address the_pc = __ pc(); 3184 __ set_last_Java_frame(noreg, rbp, the_pc); 3185 3186 __ andptr(rsp, -(StackAlignmentInBytes)); // Fix stack alignment as required by ABI 3187 __ mov(c_rarg0, r15_thread); 3188 __ movl(c_rarg1, r14); // second arg: exec_mode 3189 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 3190 // Revert SP alignment after call since we're going to do some SP relative addressing below 3191 __ movptr(rsp, Address(r15_thread, JavaThread::last_Java_sp_offset())); 3192 3193 // Set an oopmap for the call site 3194 // Use the same PC we used for the last java frame 3195 oop_maps->add_gc_map(the_pc - start, 3196 new OopMap( frame_size_in_words, 0 )); 3197 3198 // Clear fp AND pc 3199 __ reset_last_Java_frame(true); 3200 3201 // Collect return values 3202 __ movdbl(xmm0, Address(rsp, RegisterSaver::xmm0_offset_in_bytes())); 3203 __ movptr(rax, Address(rsp, RegisterSaver::rax_offset_in_bytes())); 3204 // I think this is useless (throwing pc?) 3205 __ movptr(rdx, Address(rsp, RegisterSaver::rdx_offset_in_bytes())); 3206 3207 // Pop self-frame. 3208 __ leave(); // Epilog 3209 3210 // Jump to interpreter 3211 __ ret(0); 3212 3213 // Make sure all code is generated 3214 masm->flush(); 3215 3216 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 3217 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 3218 #if INCLUDE_JVMCI 3219 if (EnableJVMCI || UseAOT) { 3220 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset); 3221 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset); 3222 } 3223 #endif 3224 } 3225 3226 #ifdef COMPILER2 3227 //------------------------------generate_uncommon_trap_blob-------------------- 3228 void SharedRuntime::generate_uncommon_trap_blob() { 3229 // Allocate space for the code 3230 ResourceMark rm; 3231 // Setup code generation tools 3232 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 3233 MacroAssembler* masm = new MacroAssembler(&buffer); 3234 3235 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 3236 3237 address start = __ pc(); 3238 3239 if (UseRTMLocking) { 3240 // Abort RTM transaction before possible nmethod deoptimization. 3241 __ xabort(0); 3242 } 3243 3244 // Push self-frame. We get here with a return address on the 3245 // stack, so rsp is 8-byte aligned until we allocate our frame. 3246 __ subptr(rsp, SimpleRuntimeFrame::return_off << LogBytesPerInt); // Epilog! 3247 3248 // No callee saved registers. rbp is assumed implicitly saved 3249 __ movptr(Address(rsp, SimpleRuntimeFrame::rbp_off << LogBytesPerInt), rbp); 3250 3251 // compiler left unloaded_class_index in j_rarg0 move to where the 3252 // runtime expects it. 3253 __ movl(c_rarg1, j_rarg0); 3254 3255 __ set_last_Java_frame(noreg, noreg, NULL); 3256 3257 // Call C code. Need thread but NOT official VM entry 3258 // crud. We cannot block on this call, no GC can happen. Call should 3259 // capture callee-saved registers as well as return values. 3260 // Thread is in rdi already. 3261 // 3262 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index); 3263 3264 __ mov(c_rarg0, r15_thread); 3265 __ movl(c_rarg2, Deoptimization::Unpack_uncommon_trap); 3266 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap))); 3267 3268 // Set an oopmap for the call site 3269 OopMapSet* oop_maps = new OopMapSet(); 3270 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0); 3271 3272 // location of rbp is known implicitly by the frame sender code 3273 3274 oop_maps->add_gc_map(__ pc() - start, map); 3275 3276 __ reset_last_Java_frame(false); 3277 3278 // Load UnrollBlock* into rdi 3279 __ mov(rdi, rax); 3280 3281 #ifdef ASSERT 3282 { Label L; 3283 __ cmpptr(Address(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()), 3284 (int32_t)Deoptimization::Unpack_uncommon_trap); 3285 __ jcc(Assembler::equal, L); 3286 __ stop("SharedRuntime::generate_deopt_blob: expected Unpack_uncommon_trap"); 3287 __ bind(L); 3288 } 3289 #endif 3290 3291 // Pop all the frames we must move/replace. 3292 // 3293 // Frame picture (youngest to oldest) 3294 // 1: self-frame (no frame link) 3295 // 2: deopting frame (no frame link) 3296 // 3: caller of deopting frame (could be compiled/interpreted). 3297 3298 // Pop self-frame. We have no frame, and must rely only on rax and rsp. 3299 __ addptr(rsp, (SimpleRuntimeFrame::framesize - 2) << LogBytesPerInt); // Epilog! 3300 3301 // Pop deoptimized frame (int) 3302 __ movl(rcx, Address(rdi, 3303 Deoptimization::UnrollBlock:: 3304 size_of_deoptimized_frame_offset_in_bytes())); 3305 __ addptr(rsp, rcx); 3306 3307 // rsp should be pointing at the return address to the caller (3) 3308 3309 // Pick up the initial fp we should save 3310 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved) 3311 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 3312 3313 #ifdef ASSERT 3314 // Compilers generate code that bang the stack by as much as the 3315 // interpreter would need. So this stack banging should never 3316 // trigger a fault. Verify that it does not on non product builds. 3317 if (UseStackBanging) { 3318 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 3319 __ bang_stack_size(rbx, rcx); 3320 } 3321 #endif 3322 3323 // Load address of array of frame pcs into rcx (address*) 3324 __ movptr(rcx, Address(rdi, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 3325 3326 // Trash the return pc 3327 __ addptr(rsp, wordSize); 3328 3329 // Load address of array of frame sizes into rsi (intptr_t*) 3330 __ movptr(rsi, Address(rdi, Deoptimization::UnrollBlock:: frame_sizes_offset_in_bytes())); 3331 3332 // Counter 3333 __ movl(rdx, Address(rdi, Deoptimization::UnrollBlock:: number_of_frames_offset_in_bytes())); // (int) 3334 3335 // Now adjust the caller's stack to make up for the extra locals but 3336 // record the original sp so that we can save it in the skeletal 3337 // interpreter frame and the stack walking of interpreter_sender 3338 // will get the unextended sp value and not the "real" sp value. 3339 3340 const Register sender_sp = r8; 3341 3342 __ mov(sender_sp, rsp); 3343 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock:: caller_adjustment_offset_in_bytes())); // (int) 3344 __ subptr(rsp, rbx); 3345 3346 // Push interpreter frames in a loop 3347 Label loop; 3348 __ bind(loop); 3349 __ movptr(rbx, Address(rsi, 0)); // Load frame size 3350 __ subptr(rbx, 2 * wordSize); // We'll push pc and rbp by hand 3351 __ pushptr(Address(rcx, 0)); // Save return address 3352 __ enter(); // Save old & set new rbp 3353 __ subptr(rsp, rbx); // Prolog 3354 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), 3355 sender_sp); // Make it walkable 3356 // This value is corrected by layout_activation_impl 3357 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), (int32_t)NULL_WORD ); 3358 __ mov(sender_sp, rsp); // Pass sender_sp to next frame 3359 __ addptr(rsi, wordSize); // Bump array pointer (sizes) 3360 __ addptr(rcx, wordSize); // Bump array pointer (pcs) 3361 __ decrementl(rdx); // Decrement counter 3362 __ jcc(Assembler::notZero, loop); 3363 __ pushptr(Address(rcx, 0)); // Save final return address 3364 3365 // Re-push self-frame 3366 __ enter(); // Save old & set new rbp 3367 __ subptr(rsp, (SimpleRuntimeFrame::framesize - 4) << LogBytesPerInt); 3368 // Prolog 3369 3370 // Use rbp because the frames look interpreted now 3371 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 3372 // Don't need the precise return PC here, just precise enough to point into this code blob. 3373 address the_pc = __ pc(); 3374 __ set_last_Java_frame(noreg, rbp, the_pc); 3375 3376 // Call C code. Need thread but NOT official VM entry 3377 // crud. We cannot block on this call, no GC can happen. Call should 3378 // restore return values to their stack-slots with the new SP. 3379 // Thread is in rdi already. 3380 // 3381 // BasicType unpack_frames(JavaThread* thread, int exec_mode); 3382 3383 __ andptr(rsp, -(StackAlignmentInBytes)); // Align SP as required by ABI 3384 __ mov(c_rarg0, r15_thread); 3385 __ movl(c_rarg1, Deoptimization::Unpack_uncommon_trap); 3386 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 3387 3388 // Set an oopmap for the call site 3389 // Use the same PC we used for the last java frame 3390 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 3391 3392 // Clear fp AND pc 3393 __ reset_last_Java_frame(true); 3394 3395 // Pop self-frame. 3396 __ leave(); // Epilog 3397 3398 // Jump to interpreter 3399 __ ret(0); 3400 3401 // Make sure all code is generated 3402 masm->flush(); 3403 3404 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, 3405 SimpleRuntimeFrame::framesize >> 1); 3406 } 3407 #endif // COMPILER2 3408 3409 3410 //------------------------------generate_handler_blob------ 3411 // 3412 // Generate a special Compile2Runtime blob that saves all registers, 3413 // and setup oopmap. 3414 // 3415 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 3416 assert(StubRoutines::forward_exception_entry() != NULL, 3417 "must be generated before"); 3418 3419 ResourceMark rm; 3420 OopMapSet *oop_maps = new OopMapSet(); 3421 OopMap* map; 3422 3423 // Allocate space for the code. Setup code generation tools. 3424 CodeBuffer buffer("handler_blob", 2048, 1024); 3425 MacroAssembler* masm = new MacroAssembler(&buffer); 3426 3427 address start = __ pc(); 3428 address call_pc = NULL; 3429 int frame_size_in_words; 3430 bool cause_return = (poll_type == POLL_AT_RETURN); 3431 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP); 3432 3433 if (UseRTMLocking) { 3434 // Abort RTM transaction before calling runtime 3435 // because critical section will be large and will be 3436 // aborted anyway. Also nmethod could be deoptimized. 3437 __ xabort(0); 3438 } 3439 3440 // Make room for return address (or push it again) 3441 if (!cause_return) { 3442 __ push(rbx); 3443 } 3444 3445 // Save registers, fpu state, and flags 3446 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, save_vectors); 3447 3448 // The following is basically a call_VM. However, we need the precise 3449 // address of the call in order to generate an oopmap. Hence, we do all the 3450 // work outselves. 3451 3452 __ set_last_Java_frame(noreg, noreg, NULL); 3453 3454 // The return address must always be correct so that frame constructor never 3455 // sees an invalid pc. 3456 3457 if (!cause_return) { 3458 // Get the return pc saved by the signal handler and stash it in its appropriate place on the stack. 3459 // Additionally, rbx is a callee saved register and we can look at it later to determine 3460 // if someone changed the return address for us! 3461 __ movptr(rbx, Address(r15_thread, JavaThread::saved_exception_pc_offset())); 3462 __ movptr(Address(rbp, wordSize), rbx); 3463 } 3464 3465 // Do the call 3466 __ mov(c_rarg0, r15_thread); 3467 __ call(RuntimeAddress(call_ptr)); 3468 3469 // Set an oopmap for the call site. This oopmap will map all 3470 // oop-registers and debug-info registers as callee-saved. This 3471 // will allow deoptimization at this safepoint to find all possible 3472 // debug-info recordings, as well as let GC find all oops. 3473 3474 oop_maps->add_gc_map( __ pc() - start, map); 3475 3476 Label noException; 3477 3478 __ reset_last_Java_frame(false); 3479 3480 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 3481 __ jcc(Assembler::equal, noException); 3482 3483 // Exception pending 3484 3485 RegisterSaver::restore_live_registers(masm, save_vectors); 3486 3487 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3488 3489 // No exception case 3490 __ bind(noException); 3491 3492 Label no_adjust, bail, no_prefix, not_special; 3493 if (SafepointMechanism::uses_thread_local_poll() && !cause_return) { 3494 // If our stashed return pc was modified by the runtime we avoid touching it 3495 __ cmpptr(rbx, Address(rbp, wordSize)); 3496 __ jccb(Assembler::notEqual, no_adjust); 3497 3498 // Skip over the poll instruction. 3499 // See NativeInstruction::is_safepoint_poll() 3500 // Possible encodings: 3501 // 85 00 test %eax,(%rax) 3502 // 85 01 test %eax,(%rcx) 3503 // 85 02 test %eax,(%rdx) 3504 // 85 03 test %eax,(%rbx) 3505 // 85 06 test %eax,(%rsi) 3506 // 85 07 test %eax,(%rdi) 3507 // 3508 // 41 85 00 test %eax,(%r8) 3509 // 41 85 01 test %eax,(%r9) 3510 // 41 85 02 test %eax,(%r10) 3511 // 41 85 03 test %eax,(%r11) 3512 // 41 85 06 test %eax,(%r14) 3513 // 41 85 07 test %eax,(%r15) 3514 // 3515 // 85 04 24 test %eax,(%rsp) 3516 // 41 85 04 24 test %eax,(%r12) 3517 // 85 45 00 test %eax,0x0(%rbp) 3518 // 41 85 45 00 test %eax,0x0(%r13) 3519 3520 __ cmpb(Address(rbx, 0), NativeTstRegMem::instruction_rex_b_prefix); 3521 __ jcc(Assembler::notEqual, no_prefix); 3522 __ addptr(rbx, 1); 3523 __ bind(no_prefix); 3524 #ifdef ASSERT 3525 __ movptr(rax, rbx); // remember where 0x85 should be, for verification below 3526 #endif 3527 // r12/r13/rsp/rbp base encoding takes 3 bytes with the following register values: 3528 // r12/rsp 0x04 3529 // r13/rbp 0x05 3530 __ movzbq(rcx, Address(rbx, 1)); 3531 __ andptr(rcx, 0x07); // looking for 0x04 .. 0x05 3532 __ subptr(rcx, 4); // looking for 0x00 .. 0x01 3533 __ cmpptr(rcx, 1); 3534 __ jcc(Assembler::above, not_special); 3535 __ addptr(rbx, 1); 3536 __ bind(not_special); 3537 #ifdef ASSERT 3538 // Verify the correct encoding of the poll we're about to skip. 3539 __ cmpb(Address(rax, 0), NativeTstRegMem::instruction_code_memXregl); 3540 __ jcc(Assembler::notEqual, bail); 3541 // Mask out the modrm bits 3542 __ testb(Address(rax, 1), NativeTstRegMem::modrm_mask); 3543 // rax encodes to 0, so if the bits are nonzero it's incorrect 3544 __ jcc(Assembler::notZero, bail); 3545 #endif 3546 // Adjust return pc forward to step over the safepoint poll instruction 3547 __ addptr(rbx, 2); 3548 __ movptr(Address(rbp, wordSize), rbx); 3549 } 3550 3551 __ bind(no_adjust); 3552 // Normal exit, restore registers and exit. 3553 RegisterSaver::restore_live_registers(masm, save_vectors); 3554 __ ret(0); 3555 3556 #ifdef ASSERT 3557 __ bind(bail); 3558 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected"); 3559 #endif 3560 3561 // Make sure all code is generated 3562 masm->flush(); 3563 3564 // Fill-out other meta info 3565 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 3566 } 3567 3568 // 3569 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 3570 // 3571 // Generate a stub that calls into vm to find out the proper destination 3572 // of a java call. All the argument registers are live at this point 3573 // but since this is generic code we don't know what they are and the caller 3574 // must do any gc of the args. 3575 // 3576 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3577 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 3578 3579 // allocate space for the code 3580 ResourceMark rm; 3581 3582 CodeBuffer buffer(name, 1000, 512); 3583 MacroAssembler* masm = new MacroAssembler(&buffer); 3584 3585 int frame_size_in_words; 3586 3587 OopMapSet *oop_maps = new OopMapSet(); 3588 OopMap* map = NULL; 3589 3590 int start = __ offset(); 3591 3592 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 3593 3594 int frame_complete = __ offset(); 3595 3596 __ set_last_Java_frame(noreg, noreg, NULL); 3597 3598 __ mov(c_rarg0, r15_thread); 3599 3600 __ call(RuntimeAddress(destination)); 3601 3602 3603 // Set an oopmap for the call site. 3604 // We need this not only for callee-saved registers, but also for volatile 3605 // registers that the compiler might be keeping live across a safepoint. 3606 3607 oop_maps->add_gc_map( __ offset() - start, map); 3608 3609 // rax contains the address we are going to jump to assuming no exception got installed 3610 3611 // clear last_Java_sp 3612 __ reset_last_Java_frame(false); 3613 // check for pending exceptions 3614 Label pending; 3615 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 3616 __ jcc(Assembler::notEqual, pending); 3617 3618 // get the returned Method* 3619 __ get_vm_result_2(rbx, r15_thread); 3620 __ movptr(Address(rsp, RegisterSaver::rbx_offset_in_bytes()), rbx); 3621 3622 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax); 3623 3624 RegisterSaver::restore_live_registers(masm); 3625 3626 // We are back the the original state on entry and ready to go. 3627 3628 __ jmp(rax); 3629 3630 // Pending exception after the safepoint 3631 3632 __ bind(pending); 3633 3634 RegisterSaver::restore_live_registers(masm); 3635 3636 // exception pending => remove activation and forward to exception handler 3637 3638 __ movptr(Address(r15_thread, JavaThread::vm_result_offset()), (int)NULL_WORD); 3639 3640 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset())); 3641 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3642 3643 // ------------- 3644 // make sure all code is generated 3645 masm->flush(); 3646 3647 // return the blob 3648 // frame_size_words or bytes?? 3649 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true); 3650 } 3651 3652 3653 //------------------------------Montgomery multiplication------------------------ 3654 // 3655 3656 #ifndef _WINDOWS 3657 3658 #define ASM_SUBTRACT 3659 3660 #ifdef ASM_SUBTRACT 3661 // Subtract 0:b from carry:a. Return carry. 3662 static unsigned long 3663 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) { 3664 long i = 0, cnt = len; 3665 unsigned long tmp; 3666 asm volatile("clc; " 3667 "0: ; " 3668 "mov (%[b], %[i], 8), %[tmp]; " 3669 "sbb %[tmp], (%[a], %[i], 8); " 3670 "inc %[i]; dec %[cnt]; " 3671 "jne 0b; " 3672 "mov %[carry], %[tmp]; sbb $0, %[tmp]; " 3673 : [i]"+r"(i), [cnt]"+r"(cnt), [tmp]"=&r"(tmp) 3674 : [a]"r"(a), [b]"r"(b), [carry]"r"(carry) 3675 : "memory"); 3676 return tmp; 3677 } 3678 #else // ASM_SUBTRACT 3679 typedef int __attribute__((mode(TI))) int128; 3680 3681 // Subtract 0:b from carry:a. Return carry. 3682 static unsigned long 3683 sub(unsigned long a[], unsigned long b[], unsigned long carry, int len) { 3684 int128 tmp = 0; 3685 int i; 3686 for (i = 0; i < len; i++) { 3687 tmp += a[i]; 3688 tmp -= b[i]; 3689 a[i] = tmp; 3690 tmp >>= 64; 3691 assert(-1 <= tmp && tmp <= 0, "invariant"); 3692 } 3693 return tmp + carry; 3694 } 3695 #endif // ! ASM_SUBTRACT 3696 3697 // Multiply (unsigned) Long A by Long B, accumulating the double- 3698 // length result into the accumulator formed of T0, T1, and T2. 3699 #define MACC(A, B, T0, T1, T2) \ 3700 do { \ 3701 unsigned long hi, lo; \ 3702 __asm__ ("mul %5; add %%rax, %2; adc %%rdx, %3; adc $0, %4" \ 3703 : "=&d"(hi), "=a"(lo), "+r"(T0), "+r"(T1), "+g"(T2) \ 3704 : "r"(A), "a"(B) : "cc"); \ 3705 } while(0) 3706 3707 // As above, but add twice the double-length result into the 3708 // accumulator. 3709 #define MACC2(A, B, T0, T1, T2) \ 3710 do { \ 3711 unsigned long hi, lo; \ 3712 __asm__ ("mul %5; add %%rax, %2; adc %%rdx, %3; adc $0, %4; " \ 3713 "add %%rax, %2; adc %%rdx, %3; adc $0, %4" \ 3714 : "=&d"(hi), "=a"(lo), "+r"(T0), "+r"(T1), "+g"(T2) \ 3715 : "r"(A), "a"(B) : "cc"); \ 3716 } while(0) 3717 3718 // Fast Montgomery multiplication. The derivation of the algorithm is 3719 // in A Cryptographic Library for the Motorola DSP56000, 3720 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237. 3721 3722 static void __attribute__((noinline)) 3723 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[], 3724 unsigned long m[], unsigned long inv, int len) { 3725 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator 3726 int i; 3727 3728 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply"); 3729 3730 for (i = 0; i < len; i++) { 3731 int j; 3732 for (j = 0; j < i; j++) { 3733 MACC(a[j], b[i-j], t0, t1, t2); 3734 MACC(m[j], n[i-j], t0, t1, t2); 3735 } 3736 MACC(a[i], b[0], t0, t1, t2); 3737 m[i] = t0 * inv; 3738 MACC(m[i], n[0], t0, t1, t2); 3739 3740 assert(t0 == 0, "broken Montgomery multiply"); 3741 3742 t0 = t1; t1 = t2; t2 = 0; 3743 } 3744 3745 for (i = len; i < 2*len; i++) { 3746 int j; 3747 for (j = i-len+1; j < len; j++) { 3748 MACC(a[j], b[i-j], t0, t1, t2); 3749 MACC(m[j], n[i-j], t0, t1, t2); 3750 } 3751 m[i-len] = t0; 3752 t0 = t1; t1 = t2; t2 = 0; 3753 } 3754 3755 while (t0) 3756 t0 = sub(m, n, t0, len); 3757 } 3758 3759 // Fast Montgomery squaring. This uses asymptotically 25% fewer 3760 // multiplies so it should be up to 25% faster than Montgomery 3761 // multiplication. However, its loop control is more complex and it 3762 // may actually run slower on some machines. 3763 3764 static void __attribute__((noinline)) 3765 montgomery_square(unsigned long a[], unsigned long n[], 3766 unsigned long m[], unsigned long inv, int len) { 3767 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator 3768 int i; 3769 3770 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply"); 3771 3772 for (i = 0; i < len; i++) { 3773 int j; 3774 int end = (i+1)/2; 3775 for (j = 0; j < end; j++) { 3776 MACC2(a[j], a[i-j], t0, t1, t2); 3777 MACC(m[j], n[i-j], t0, t1, t2); 3778 } 3779 if ((i & 1) == 0) { 3780 MACC(a[j], a[j], t0, t1, t2); 3781 } 3782 for (; j < i; j++) { 3783 MACC(m[j], n[i-j], t0, t1, t2); 3784 } 3785 m[i] = t0 * inv; 3786 MACC(m[i], n[0], t0, t1, t2); 3787 3788 assert(t0 == 0, "broken Montgomery square"); 3789 3790 t0 = t1; t1 = t2; t2 = 0; 3791 } 3792 3793 for (i = len; i < 2*len; i++) { 3794 int start = i-len+1; 3795 int end = start + (len - start)/2; 3796 int j; 3797 for (j = start; j < end; j++) { 3798 MACC2(a[j], a[i-j], t0, t1, t2); 3799 MACC(m[j], n[i-j], t0, t1, t2); 3800 } 3801 if ((i & 1) == 0) { 3802 MACC(a[j], a[j], t0, t1, t2); 3803 } 3804 for (; j < len; j++) { 3805 MACC(m[j], n[i-j], t0, t1, t2); 3806 } 3807 m[i-len] = t0; 3808 t0 = t1; t1 = t2; t2 = 0; 3809 } 3810 3811 while (t0) 3812 t0 = sub(m, n, t0, len); 3813 } 3814 3815 // Swap words in a longword. 3816 static unsigned long swap(unsigned long x) { 3817 return (x << 32) | (x >> 32); 3818 } 3819 3820 // Copy len longwords from s to d, word-swapping as we go. The 3821 // destination array is reversed. 3822 static void reverse_words(unsigned long *s, unsigned long *d, int len) { 3823 d += len; 3824 while(len-- > 0) { 3825 d--; 3826 *d = swap(*s); 3827 s++; 3828 } 3829 } 3830 3831 // The threshold at which squaring is advantageous was determined 3832 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz. 3833 #define MONTGOMERY_SQUARING_THRESHOLD 64 3834 3835 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints, 3836 jint len, jlong inv, 3837 jint *m_ints) { 3838 assert(len % 2 == 0, "array length in montgomery_multiply must be even"); 3839 int longwords = len/2; 3840 3841 // Make very sure we don't use so much space that the stack might 3842 // overflow. 512 jints corresponds to an 16384-bit integer and 3843 // will use here a total of 8k bytes of stack space. 3844 int total_allocation = longwords * sizeof (unsigned long) * 4; 3845 guarantee(total_allocation <= 8192, "must be"); 3846 unsigned long *scratch = (unsigned long *)alloca(total_allocation); 3847 3848 // Local scratch arrays 3849 unsigned long 3850 *a = scratch + 0 * longwords, 3851 *b = scratch + 1 * longwords, 3852 *n = scratch + 2 * longwords, 3853 *m = scratch + 3 * longwords; 3854 3855 reverse_words((unsigned long *)a_ints, a, longwords); 3856 reverse_words((unsigned long *)b_ints, b, longwords); 3857 reverse_words((unsigned long *)n_ints, n, longwords); 3858 3859 ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords); 3860 3861 reverse_words(m, (unsigned long *)m_ints, longwords); 3862 } 3863 3864 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints, 3865 jint len, jlong inv, 3866 jint *m_ints) { 3867 assert(len % 2 == 0, "array length in montgomery_square must be even"); 3868 int longwords = len/2; 3869 3870 // Make very sure we don't use so much space that the stack might 3871 // overflow. 512 jints corresponds to an 16384-bit integer and 3872 // will use here a total of 6k bytes of stack space. 3873 int total_allocation = longwords * sizeof (unsigned long) * 3; 3874 guarantee(total_allocation <= 8192, "must be"); 3875 unsigned long *scratch = (unsigned long *)alloca(total_allocation); 3876 3877 // Local scratch arrays 3878 unsigned long 3879 *a = scratch + 0 * longwords, 3880 *n = scratch + 1 * longwords, 3881 *m = scratch + 2 * longwords; 3882 3883 reverse_words((unsigned long *)a_ints, a, longwords); 3884 reverse_words((unsigned long *)n_ints, n, longwords); 3885 3886 if (len >= MONTGOMERY_SQUARING_THRESHOLD) { 3887 ::montgomery_square(a, n, m, (unsigned long)inv, longwords); 3888 } else { 3889 ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords); 3890 } 3891 3892 reverse_words(m, (unsigned long *)m_ints, longwords); 3893 } 3894 3895 #endif // WINDOWS 3896 3897 #ifdef COMPILER2 3898 // This is here instead of runtime_x86_64.cpp because it uses SimpleRuntimeFrame 3899 // 3900 //------------------------------generate_exception_blob--------------------------- 3901 // creates exception blob at the end 3902 // Using exception blob, this code is jumped from a compiled method. 3903 // (see emit_exception_handler in x86_64.ad file) 3904 // 3905 // Given an exception pc at a call we call into the runtime for the 3906 // handler in this method. This handler might merely restore state 3907 // (i.e. callee save registers) unwind the frame and jump to the 3908 // exception handler for the nmethod if there is no Java level handler 3909 // for the nmethod. 3910 // 3911 // This code is entered with a jmp. 3912 // 3913 // Arguments: 3914 // rax: exception oop 3915 // rdx: exception pc 3916 // 3917 // Results: 3918 // rax: exception oop 3919 // rdx: exception pc in caller or ??? 3920 // destination: exception handler of caller 3921 // 3922 // Note: the exception pc MUST be at a call (precise debug information) 3923 // Registers rax, rdx, rcx, rsi, rdi, r8-r11 are not callee saved. 3924 // 3925 3926 void OptoRuntime::generate_exception_blob() { 3927 assert(!OptoRuntime::is_callee_saved_register(RDX_num), ""); 3928 assert(!OptoRuntime::is_callee_saved_register(RAX_num), ""); 3929 assert(!OptoRuntime::is_callee_saved_register(RCX_num), ""); 3930 3931 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 3932 3933 // Allocate space for the code 3934 ResourceMark rm; 3935 // Setup code generation tools 3936 CodeBuffer buffer("exception_blob", 2048, 1024); 3937 MacroAssembler* masm = new MacroAssembler(&buffer); 3938 3939 3940 address start = __ pc(); 3941 3942 // Exception pc is 'return address' for stack walker 3943 __ push(rdx); 3944 __ subptr(rsp, SimpleRuntimeFrame::return_off << LogBytesPerInt); // Prolog 3945 3946 // Save callee-saved registers. See x86_64.ad. 3947 3948 // rbp is an implicitly saved callee saved register (i.e., the calling 3949 // convention will save/restore it in the prolog/epilog). Other than that 3950 // there are no callee save registers now that adapter frames are gone. 3951 3952 __ movptr(Address(rsp, SimpleRuntimeFrame::rbp_off << LogBytesPerInt), rbp); 3953 3954 // Store exception in Thread object. We cannot pass any arguments to the 3955 // handle_exception call, since we do not want to make any assumption 3956 // about the size of the frame where the exception happened in. 3957 // c_rarg0 is either rdi (Linux) or rcx (Windows). 3958 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()),rax); 3959 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), rdx); 3960 3961 // This call does all the hard work. It checks if an exception handler 3962 // exists in the method. 3963 // If so, it returns the handler address. 3964 // If not, it prepares for stack-unwinding, restoring the callee-save 3965 // registers of the frame being removed. 3966 // 3967 // address OptoRuntime::handle_exception_C(JavaThread* thread) 3968 3969 // At a method handle call, the stack may not be properly aligned 3970 // when returning with an exception. 3971 address the_pc = __ pc(); 3972 __ set_last_Java_frame(noreg, noreg, the_pc); 3973 __ mov(c_rarg0, r15_thread); 3974 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack 3975 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C))); 3976 3977 // Set an oopmap for the call site. This oopmap will only be used if we 3978 // are unwinding the stack. Hence, all locations will be dead. 3979 // Callee-saved registers will be the same as the frame above (i.e., 3980 // handle_exception_stub), since they were restored when we got the 3981 // exception. 3982 3983 OopMapSet* oop_maps = new OopMapSet(); 3984 3985 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 3986 3987 __ reset_last_Java_frame(false); 3988 3989 // Restore callee-saved registers 3990 3991 // rbp is an implicitly saved callee-saved register (i.e., the calling 3992 // convention will save restore it in prolog/epilog) Other than that 3993 // there are no callee save registers now that adapter frames are gone. 3994 3995 __ movptr(rbp, Address(rsp, SimpleRuntimeFrame::rbp_off << LogBytesPerInt)); 3996 3997 __ addptr(rsp, SimpleRuntimeFrame::return_off << LogBytesPerInt); // Epilog 3998 __ pop(rdx); // No need for exception pc anymore 3999 4000 // rax: exception handler 4001 4002 // We have a handler in rax (could be deopt blob). 4003 __ mov(r8, rax); 4004 4005 // Get the exception oop 4006 __ movptr(rax, Address(r15_thread, JavaThread::exception_oop_offset())); 4007 // Get the exception pc in case we are deoptimized 4008 __ movptr(rdx, Address(r15_thread, JavaThread::exception_pc_offset())); 4009 #ifdef ASSERT 4010 __ movptr(Address(r15_thread, JavaThread::exception_handler_pc_offset()), (int)NULL_WORD); 4011 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), (int)NULL_WORD); 4012 #endif 4013 // Clear the exception oop so GC no longer processes it as a root. 4014 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()), (int)NULL_WORD); 4015 4016 // rax: exception oop 4017 // r8: exception handler 4018 // rdx: exception pc 4019 // Jump to handler 4020 4021 __ jmp(r8); 4022 4023 // Make sure all code is generated 4024 masm->flush(); 4025 4026 // Set exception blob 4027 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1); 4028 } 4029 #endif // COMPILER2