1 /* 2 * Copyright (c) 2003, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #ifndef _WINDOWS 27 #include "alloca.h" 28 #endif 29 #include "asm/macroAssembler.hpp" 30 #include "asm/macroAssembler.inline.hpp" 31 #include "classfile/symbolTable.hpp" 32 #include "code/debugInfoRec.hpp" 33 #include "code/icBuffer.hpp" 34 #include "code/vtableStubs.hpp" 35 #include "interpreter/interpreter.hpp" 36 #include "logging/log.hpp" 37 #include "memory/resourceArea.hpp" 38 #include "oops/compiledICHolder.hpp" 39 #include "runtime/sharedRuntime.hpp" 40 #include "runtime/vframeArray.hpp" 41 #include "vmreg_x86.inline.hpp" 42 #ifdef COMPILER1 43 #include "c1/c1_Runtime1.hpp" 44 #endif 45 #ifdef COMPILER2 46 #include "opto/runtime.hpp" 47 #endif 48 #if INCLUDE_JVMCI 49 #include "jvmci/jvmciJavaClasses.hpp" 50 #endif 51 52 #define __ masm-> 53 54 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size; 55 56 class SimpleRuntimeFrame { 57 58 public: 59 60 // Most of the runtime stubs have this simple frame layout. 61 // This class exists to make the layout shared in one place. 62 // Offsets are for compiler stack slots, which are jints. 63 enum layout { 64 // The frame sender code expects that rbp will be in the "natural" place and 65 // will override any oopMap setting for it. We must therefore force the layout 66 // so that it agrees with the frame sender code. 67 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt, 68 rbp_off2, 69 return_off, return_off2, 70 framesize 71 }; 72 }; 73 74 class RegisterSaver { 75 // Capture info about frame layout. Layout offsets are in jint 76 // units because compiler frame slots are jints. 77 #define XSAVE_AREA_BEGIN 160 78 #define XSAVE_AREA_YMM_BEGIN 576 79 #define XSAVE_AREA_ZMM_BEGIN 1152 80 #define XSAVE_AREA_UPPERBANK 1664 81 #define DEF_XMM_OFFS(regnum) xmm ## regnum ## _off = xmm_off + (regnum)*16/BytesPerInt, xmm ## regnum ## H_off 82 #define DEF_YMM_OFFS(regnum) ymm ## regnum ## _off = ymm_off + (regnum)*16/BytesPerInt, ymm ## regnum ## H_off 83 #define DEF_ZMM_OFFS(regnum) zmm ## regnum ## _off = zmm_off + (regnum-16)*64/BytesPerInt, zmm ## regnum ## H_off 84 enum layout { 85 fpu_state_off = frame::arg_reg_save_area_bytes/BytesPerInt, // fxsave save area 86 xmm_off = fpu_state_off + XSAVE_AREA_BEGIN/BytesPerInt, // offset in fxsave save area 87 DEF_XMM_OFFS(0), 88 DEF_XMM_OFFS(1), 89 // 2..15 are implied in range usage 90 ymm_off = xmm_off + (XSAVE_AREA_YMM_BEGIN - XSAVE_AREA_BEGIN)/BytesPerInt, 91 DEF_YMM_OFFS(0), 92 DEF_YMM_OFFS(1), 93 // 2..15 are implied in range usage 94 zmm_high = xmm_off + (XSAVE_AREA_ZMM_BEGIN - XSAVE_AREA_BEGIN)/BytesPerInt, 95 zmm_off = xmm_off + (XSAVE_AREA_UPPERBANK - XSAVE_AREA_BEGIN)/BytesPerInt, 96 DEF_ZMM_OFFS(16), 97 DEF_ZMM_OFFS(17), 98 // 18..31 are implied in range usage 99 fpu_state_end = fpu_state_off + ((FPUStateSizeInWords-1)*wordSize / BytesPerInt), 100 fpu_stateH_end, 101 r15_off, r15H_off, 102 r14_off, r14H_off, 103 r13_off, r13H_off, 104 r12_off, r12H_off, 105 r11_off, r11H_off, 106 r10_off, r10H_off, 107 r9_off, r9H_off, 108 r8_off, r8H_off, 109 rdi_off, rdiH_off, 110 rsi_off, rsiH_off, 111 ignore_off, ignoreH_off, // extra copy of rbp 112 rsp_off, rspH_off, 113 rbx_off, rbxH_off, 114 rdx_off, rdxH_off, 115 rcx_off, rcxH_off, 116 rax_off, raxH_off, 117 // 16-byte stack alignment fill word: see MacroAssembler::push/pop_IU_state 118 align_off, alignH_off, 119 flags_off, flagsH_off, 120 // The frame sender code expects that rbp will be in the "natural" place and 121 // will override any oopMap setting for it. We must therefore force the layout 122 // so that it agrees with the frame sender code. 123 rbp_off, rbpH_off, // copy of rbp we will restore 124 return_off, returnH_off, // slot for return address 125 reg_save_size // size in compiler stack slots 126 }; 127 128 public: 129 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_vectors = false); 130 static void restore_live_registers(MacroAssembler* masm, bool restore_vectors = false); 131 132 // Offsets into the register save area 133 // Used by deoptimization when it is managing result register 134 // values on its own 135 136 static int rax_offset_in_bytes(void) { return BytesPerInt * rax_off; } 137 static int rdx_offset_in_bytes(void) { return BytesPerInt * rdx_off; } 138 static int rbx_offset_in_bytes(void) { return BytesPerInt * rbx_off; } 139 static int xmm0_offset_in_bytes(void) { return BytesPerInt * xmm0_off; } 140 static int return_offset_in_bytes(void) { return BytesPerInt * return_off; } 141 142 // During deoptimization only the result registers need to be restored, 143 // all the other values have already been extracted. 144 static void restore_result_registers(MacroAssembler* masm); 145 }; 146 147 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_vectors) { 148 int off = 0; 149 int num_xmm_regs = XMMRegisterImpl::number_of_registers; 150 if (UseAVX < 3) { 151 num_xmm_regs = num_xmm_regs/2; 152 } 153 #if defined(COMPILER2) || INCLUDE_JVMCI 154 if (save_vectors) { 155 assert(UseAVX > 0, "up to 512bit vectors are supported with EVEX"); 156 assert(MaxVectorSize <= 64, "up to 512bit vectors are supported now"); 157 } 158 #else 159 assert(!save_vectors, "vectors are generated only by C2 and JVMCI"); 160 #endif 161 162 // Always make the frame size 16-byte aligned, both vector and non vector stacks are always allocated 163 int frame_size_in_bytes = round_to(reg_save_size*BytesPerInt, num_xmm_regs); 164 // OopMap frame size is in compiler stack slots (jint's) not bytes or words 165 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt; 166 // CodeBlob frame size is in words. 167 int frame_size_in_words = frame_size_in_bytes / wordSize; 168 *total_frame_words = frame_size_in_words; 169 170 // Save registers, fpu state, and flags. 171 // We assume caller has already pushed the return address onto the 172 // stack, so rsp is 8-byte aligned here. 173 // We push rpb twice in this sequence because we want the real rbp 174 // to be under the return like a normal enter. 175 176 __ enter(); // rsp becomes 16-byte aligned here 177 __ push_CPU_state(); // Push a multiple of 16 bytes 178 179 // push cpu state handles this on EVEX enabled targets 180 if (save_vectors) { 181 // Save upper half of YMM registers(0..15) 182 int base_addr = XSAVE_AREA_YMM_BEGIN; 183 for (int n = 0; n < 16; n++) { 184 __ vextractf128_high(Address(rsp, base_addr+n*16), as_XMMRegister(n)); 185 } 186 if (VM_Version::supports_evex()) { 187 // Save upper half of ZMM registers(0..15) 188 base_addr = XSAVE_AREA_ZMM_BEGIN; 189 for (int n = 0; n < 16; n++) { 190 __ vextractf64x4_high(Address(rsp, base_addr+n*32), as_XMMRegister(n)); 191 } 192 // Save full ZMM registers(16..num_xmm_regs) 193 base_addr = XSAVE_AREA_UPPERBANK; 194 off = 0; 195 int vector_len = Assembler::AVX_512bit; 196 for (int n = 16; n < num_xmm_regs; n++) { 197 __ evmovdqul(Address(rsp, base_addr+(off++*64)), as_XMMRegister(n), vector_len); 198 } 199 } 200 } else { 201 if (VM_Version::supports_evex()) { 202 // Save upper bank of ZMM registers(16..31) for double/float usage 203 int base_addr = XSAVE_AREA_UPPERBANK; 204 off = 0; 205 for (int n = 16; n < num_xmm_regs; n++) { 206 __ movsd(Address(rsp, base_addr+(off++*64)), as_XMMRegister(n)); 207 } 208 } 209 } 210 if (frame::arg_reg_save_area_bytes != 0) { 211 // Allocate argument register save area 212 __ subptr(rsp, frame::arg_reg_save_area_bytes); 213 } 214 215 // Set an oopmap for the call site. This oopmap will map all 216 // oop-registers and debug-info registers as callee-saved. This 217 // will allow deoptimization at this safepoint to find all possible 218 // debug-info recordings, as well as let GC find all oops. 219 220 OopMapSet *oop_maps = new OopMapSet(); 221 OopMap* map = new OopMap(frame_size_in_slots, 0); 222 223 #define STACK_OFFSET(x) VMRegImpl::stack2reg((x)) 224 225 map->set_callee_saved(STACK_OFFSET( rax_off ), rax->as_VMReg()); 226 map->set_callee_saved(STACK_OFFSET( rcx_off ), rcx->as_VMReg()); 227 map->set_callee_saved(STACK_OFFSET( rdx_off ), rdx->as_VMReg()); 228 map->set_callee_saved(STACK_OFFSET( rbx_off ), rbx->as_VMReg()); 229 // rbp location is known implicitly by the frame sender code, needs no oopmap 230 // and the location where rbp was saved by is ignored 231 map->set_callee_saved(STACK_OFFSET( rsi_off ), rsi->as_VMReg()); 232 map->set_callee_saved(STACK_OFFSET( rdi_off ), rdi->as_VMReg()); 233 map->set_callee_saved(STACK_OFFSET( r8_off ), r8->as_VMReg()); 234 map->set_callee_saved(STACK_OFFSET( r9_off ), r9->as_VMReg()); 235 map->set_callee_saved(STACK_OFFSET( r10_off ), r10->as_VMReg()); 236 map->set_callee_saved(STACK_OFFSET( r11_off ), r11->as_VMReg()); 237 map->set_callee_saved(STACK_OFFSET( r12_off ), r12->as_VMReg()); 238 map->set_callee_saved(STACK_OFFSET( r13_off ), r13->as_VMReg()); 239 map->set_callee_saved(STACK_OFFSET( r14_off ), r14->as_VMReg()); 240 map->set_callee_saved(STACK_OFFSET( r15_off ), r15->as_VMReg()); 241 // For both AVX and EVEX we will use the legacy FXSAVE area for xmm0..xmm15, 242 // on EVEX enabled targets, we get it included in the xsave area 243 off = xmm0_off; 244 int delta = xmm1_off - off; 245 for (int n = 0; n < 16; n++) { 246 XMMRegister xmm_name = as_XMMRegister(n); 247 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg()); 248 off += delta; 249 } 250 if(UseAVX > 2) { 251 // Obtain xmm16..xmm31 from the XSAVE area on EVEX enabled targets 252 off = zmm16_off; 253 delta = zmm17_off - off; 254 for (int n = 16; n < num_xmm_regs; n++) { 255 XMMRegister zmm_name = as_XMMRegister(n); 256 map->set_callee_saved(STACK_OFFSET(off), zmm_name->as_VMReg()); 257 off += delta; 258 } 259 } 260 261 #if defined(COMPILER2) || INCLUDE_JVMCI 262 if (save_vectors) { 263 off = ymm0_off; 264 int delta = ymm1_off - off; 265 for (int n = 0; n < 16; n++) { 266 XMMRegister ymm_name = as_XMMRegister(n); 267 map->set_callee_saved(STACK_OFFSET(off), ymm_name->as_VMReg()->next(4)); 268 off += delta; 269 } 270 } 271 #endif // COMPILER2 || INCLUDE_JVMCI 272 273 // %%% These should all be a waste but we'll keep things as they were for now 274 if (true) { 275 map->set_callee_saved(STACK_OFFSET( raxH_off ), rax->as_VMReg()->next()); 276 map->set_callee_saved(STACK_OFFSET( rcxH_off ), rcx->as_VMReg()->next()); 277 map->set_callee_saved(STACK_OFFSET( rdxH_off ), rdx->as_VMReg()->next()); 278 map->set_callee_saved(STACK_OFFSET( rbxH_off ), rbx->as_VMReg()->next()); 279 // rbp location is known implicitly by the frame sender code, needs no oopmap 280 map->set_callee_saved(STACK_OFFSET( rsiH_off ), rsi->as_VMReg()->next()); 281 map->set_callee_saved(STACK_OFFSET( rdiH_off ), rdi->as_VMReg()->next()); 282 map->set_callee_saved(STACK_OFFSET( r8H_off ), r8->as_VMReg()->next()); 283 map->set_callee_saved(STACK_OFFSET( r9H_off ), r9->as_VMReg()->next()); 284 map->set_callee_saved(STACK_OFFSET( r10H_off ), r10->as_VMReg()->next()); 285 map->set_callee_saved(STACK_OFFSET( r11H_off ), r11->as_VMReg()->next()); 286 map->set_callee_saved(STACK_OFFSET( r12H_off ), r12->as_VMReg()->next()); 287 map->set_callee_saved(STACK_OFFSET( r13H_off ), r13->as_VMReg()->next()); 288 map->set_callee_saved(STACK_OFFSET( r14H_off ), r14->as_VMReg()->next()); 289 map->set_callee_saved(STACK_OFFSET( r15H_off ), r15->as_VMReg()->next()); 290 // For both AVX and EVEX we will use the legacy FXSAVE area for xmm0..xmm15, 291 // on EVEX enabled targets, we get it included in the xsave area 292 off = xmm0H_off; 293 delta = xmm1H_off - off; 294 for (int n = 0; n < 16; n++) { 295 XMMRegister xmm_name = as_XMMRegister(n); 296 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg()->next()); 297 off += delta; 298 } 299 if (UseAVX > 2) { 300 // Obtain xmm16..xmm31 from the XSAVE area on EVEX enabled targets 301 off = zmm16H_off; 302 delta = zmm17H_off - off; 303 for (int n = 16; n < num_xmm_regs; n++) { 304 XMMRegister zmm_name = as_XMMRegister(n); 305 map->set_callee_saved(STACK_OFFSET(off), zmm_name->as_VMReg()->next()); 306 off += delta; 307 } 308 } 309 } 310 311 return map; 312 } 313 314 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_vectors) { 315 int num_xmm_regs = XMMRegisterImpl::number_of_registers; 316 if (UseAVX < 3) { 317 num_xmm_regs = num_xmm_regs/2; 318 } 319 if (frame::arg_reg_save_area_bytes != 0) { 320 // Pop arg register save area 321 __ addptr(rsp, frame::arg_reg_save_area_bytes); 322 } 323 324 #if defined(COMPILER2) || INCLUDE_JVMCI 325 if (restore_vectors) { 326 assert(UseAVX > 0, "up to 512bit vectors are supported with EVEX"); 327 assert(MaxVectorSize <= 64, "up to 512bit vectors are supported now"); 328 } 329 #else 330 assert(!restore_vectors, "vectors are generated only by C2"); 331 #endif 332 333 // On EVEX enabled targets everything is handled in pop fpu state 334 if (restore_vectors) { 335 // Restore upper half of YMM registers (0..15) 336 int base_addr = XSAVE_AREA_YMM_BEGIN; 337 for (int n = 0; n < 16; n++) { 338 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, base_addr+n*16)); 339 } 340 if (VM_Version::supports_evex()) { 341 // Restore upper half of ZMM registers (0..15) 342 base_addr = XSAVE_AREA_ZMM_BEGIN; 343 for (int n = 0; n < 16; n++) { 344 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, base_addr+n*32)); 345 } 346 // Restore full ZMM registers(16..num_xmm_regs) 347 base_addr = XSAVE_AREA_UPPERBANK; 348 int vector_len = Assembler::AVX_512bit; 349 int off = 0; 350 for (int n = 16; n < num_xmm_regs; n++) { 351 __ evmovdqul(as_XMMRegister(n), Address(rsp, base_addr+(off++*64)), vector_len); 352 } 353 } 354 } else { 355 if (VM_Version::supports_evex()) { 356 // Restore upper bank of ZMM registers(16..31) for double/float usage 357 int base_addr = XSAVE_AREA_UPPERBANK; 358 int off = 0; 359 for (int n = 16; n < num_xmm_regs; n++) { 360 __ movsd(as_XMMRegister(n), Address(rsp, base_addr+(off++*64))); 361 } 362 } 363 } 364 365 // Recover CPU state 366 __ pop_CPU_state(); 367 // Get the rbp described implicitly by the calling convention (no oopMap) 368 __ pop(rbp); 369 } 370 371 void RegisterSaver::restore_result_registers(MacroAssembler* masm) { 372 373 // Just restore result register. Only used by deoptimization. By 374 // now any callee save register that needs to be restored to a c2 375 // caller of the deoptee has been extracted into the vframeArray 376 // and will be stuffed into the c2i adapter we create for later 377 // restoration so only result registers need to be restored here. 378 379 // Restore fp result register 380 __ movdbl(xmm0, Address(rsp, xmm0_offset_in_bytes())); 381 // Restore integer result register 382 __ movptr(rax, Address(rsp, rax_offset_in_bytes())); 383 __ movptr(rdx, Address(rsp, rdx_offset_in_bytes())); 384 385 // Pop all of the register save are off the stack except the return address 386 __ addptr(rsp, return_offset_in_bytes()); 387 } 388 389 // Is vector's size (in bytes) bigger than a size saved by default? 390 // 16 bytes XMM registers are saved by default using fxsave/fxrstor instructions. 391 bool SharedRuntime::is_wide_vector(int size) { 392 return size > 16; 393 } 394 395 size_t SharedRuntime::trampoline_size() { 396 return 16; 397 } 398 399 void SharedRuntime::generate_trampoline(MacroAssembler *masm, address destination) { 400 __ jump(RuntimeAddress(destination)); 401 } 402 403 // The java_calling_convention describes stack locations as ideal slots on 404 // a frame with no abi restrictions. Since we must observe abi restrictions 405 // (like the placement of the register window) the slots must be biased by 406 // the following value. 407 static int reg2offset_in(VMReg r) { 408 // Account for saved rbp and return address 409 // This should really be in_preserve_stack_slots 410 return (r->reg2stack() + 4) * VMRegImpl::stack_slot_size; 411 } 412 413 static int reg2offset_out(VMReg r) { 414 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 415 } 416 417 // --------------------------------------------------------------------------- 418 // Read the array of BasicTypes from a signature, and compute where the 419 // arguments should go. Values in the VMRegPair regs array refer to 4-byte 420 // quantities. Values less than VMRegImpl::stack0 are registers, those above 421 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer 422 // as framesizes are fixed. 423 // VMRegImpl::stack0 refers to the first slot 0(sp). 424 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register 425 // up to RegisterImpl::number_of_registers) are the 64-bit 426 // integer registers. 427 428 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 429 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 430 // units regardless of build. Of course for i486 there is no 64 bit build 431 432 // The Java calling convention is a "shifted" version of the C ABI. 433 // By skipping the first C ABI register we can call non-static jni methods 434 // with small numbers of arguments without having to shuffle the arguments 435 // at all. Since we control the java ABI we ought to at least get some 436 // advantage out of it. 437 438 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 439 VMRegPair *regs, 440 int total_args_passed, 441 int is_outgoing) { 442 443 // Create the mapping between argument positions and 444 // registers. 445 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = { 446 j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5 447 }; 448 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_j] = { 449 j_farg0, j_farg1, j_farg2, j_farg3, 450 j_farg4, j_farg5, j_farg6, j_farg7 451 }; 452 453 454 uint int_args = 0; 455 uint fp_args = 0; 456 uint stk_args = 0; // inc by 2 each time 457 458 for (int i = 0; i < total_args_passed; i++) { 459 switch (sig_bt[i]) { 460 case T_BOOLEAN: 461 case T_CHAR: 462 case T_BYTE: 463 case T_SHORT: 464 case T_INT: 465 if (int_args < Argument::n_int_register_parameters_j) { 466 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 467 } else { 468 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 469 stk_args += 2; 470 } 471 break; 472 case T_VOID: 473 // halves of T_LONG or T_DOUBLE 474 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 475 regs[i].set_bad(); 476 break; 477 case T_LONG: 478 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 479 // fall through 480 case T_OBJECT: 481 case T_ARRAY: 482 case T_ADDRESS: 483 case T_VALUETYPE: // just treat as ref for now 484 if (int_args < Argument::n_int_register_parameters_j) { 485 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 486 } else { 487 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 488 stk_args += 2; 489 } 490 break; 491 case T_FLOAT: 492 if (fp_args < Argument::n_float_register_parameters_j) { 493 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 494 } else { 495 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 496 stk_args += 2; 497 } 498 break; 499 case T_DOUBLE: 500 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 501 if (fp_args < Argument::n_float_register_parameters_j) { 502 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 503 } else { 504 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 505 stk_args += 2; 506 } 507 break; 508 default: 509 ShouldNotReachHere(); 510 break; 511 } 512 } 513 514 return round_to(stk_args, 2); 515 } 516 517 // Patch the callers callsite with entry to compiled code if it exists. 518 static void patch_callers_callsite(MacroAssembler *masm) { 519 Label L; 520 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD); 521 __ jcc(Assembler::equal, L); 522 523 // Save the current stack pointer 524 __ mov(r13, rsp); 525 // Schedule the branch target address early. 526 // Call into the VM to patch the caller, then jump to compiled callee 527 // rax isn't live so capture return address while we easily can 528 __ movptr(rax, Address(rsp, 0)); 529 530 // align stack so push_CPU_state doesn't fault 531 __ andptr(rsp, -(StackAlignmentInBytes)); 532 __ push_CPU_state(); 533 534 // VM needs caller's callsite 535 // VM needs target method 536 // This needs to be a long call since we will relocate this adapter to 537 // the codeBuffer and it may not reach 538 539 // Allocate argument register save area 540 if (frame::arg_reg_save_area_bytes != 0) { 541 __ subptr(rsp, frame::arg_reg_save_area_bytes); 542 } 543 __ mov(c_rarg0, rbx); 544 __ mov(c_rarg1, rax); 545 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite))); 546 547 // De-allocate argument register save area 548 if (frame::arg_reg_save_area_bytes != 0) { 549 __ addptr(rsp, frame::arg_reg_save_area_bytes); 550 } 551 552 __ pop_CPU_state(); 553 // restore sp 554 __ mov(rsp, r13); 555 __ bind(L); 556 } 557 558 // For each value type argument, sig includes the list of fields of 559 // the value type. This utility function computes the number of 560 // arguments for the call if value types are passed by reference (the 561 // calling convention the interpreter expects). 562 static int compute_total_args_passed_int(const GrowableArray<SigEntry>& sig_extended) { 563 int total_args_passed = 0; 564 if (ValueTypePassFieldsAsArgs) { 565 for (int i = 0; i < sig_extended.length(); i++) { 566 BasicType bt = sig_extended.at(i)._bt; 567 if (bt == T_VALUETYPE) { 568 // In sig_extended, a value type argument starts with: 569 // T_VALUETYPE, followed by the types of the fields of the 570 // value type and T_VOID to mark the end of the value 571 // type. Value types are flattened so, for instance, in the 572 // case of a value type with an int field and a value type 573 // field that itself has 2 fields, an int and a long: 574 // T_VALUETYPE T_INT T_VALUETYPE T_INT T_LONG T_VOID (second 575 // slot for the T_LONG) T_VOID (inner T_VALUETYPE) T_VOID 576 // (outer T_VALUETYPE) 577 total_args_passed++; 578 int vt = 1; 579 do { 580 i++; 581 BasicType bt = sig_extended.at(i)._bt; 582 BasicType prev_bt = sig_extended.at(i-1)._bt; 583 if (bt == T_VALUETYPE) { 584 vt++; 585 } else if (bt == T_VOID && 586 prev_bt != T_LONG && 587 prev_bt != T_DOUBLE) { 588 vt--; 589 } 590 } while (vt != 0); 591 } else { 592 total_args_passed++; 593 } 594 } 595 } else { 596 total_args_passed = sig_extended.length(); 597 } 598 return total_args_passed; 599 } 600 601 602 static void gen_c2i_adapter_helper(MacroAssembler* masm, 603 BasicType bt, 604 BasicType prev_bt, 605 size_t size_in_bytes, 606 const VMRegPair& reg_pair, 607 const Address& to, 608 int extraspace, 609 bool is_oop) { 610 assert(bt != T_VALUETYPE || !ValueTypePassFieldsAsArgs, "no value type here"); 611 if (bt == T_VOID) { 612 assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half"); 613 return; 614 } 615 616 // Say 4 args: 617 // i st_off 618 // 0 32 T_LONG 619 // 1 24 T_VOID 620 // 2 16 T_OBJECT 621 // 3 8 T_BOOL 622 // - 0 return address 623 // 624 // However to make thing extra confusing. Because we can fit a long/double in 625 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter 626 // leaves one slot empty and only stores to a single slot. In this case the 627 // slot that is occupied is the T_VOID slot. See I said it was confusing. 628 629 bool wide = (size_in_bytes == wordSize); 630 VMReg r_1 = reg_pair.first(); 631 VMReg r_2 = reg_pair.second(); 632 assert(r_2->is_valid() == wide, "invalid size"); 633 if (!r_1->is_valid()) { 634 assert(!r_2->is_valid(), "must be invalid"); 635 return; 636 } 637 638 if (!r_1->is_XMMRegister()) { 639 Register val = rax; 640 assert_different_registers(to.base(), val); 641 if(r_1->is_stack()) { 642 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace; 643 __ load_sized_value(val, Address(rsp, ld_off), size_in_bytes, /* is_signed */ false); 644 } else { 645 val = r_1->as_Register(); 646 } 647 if (is_oop) { 648 __ store_heap_oop(to, val); 649 } else { 650 __ store_sized_value(to, val, size_in_bytes); 651 } 652 } else { 653 if (wide) { 654 __ movdbl(to, r_1->as_XMMRegister()); 655 } else { 656 __ movflt(to, r_1->as_XMMRegister()); 657 } 658 } 659 } 660 661 static void gen_c2i_adapter(MacroAssembler *masm, 662 const GrowableArray<SigEntry>& sig_extended, 663 const VMRegPair *regs, 664 Label& skip_fixup, 665 address start, 666 OopMapSet*& oop_maps, 667 int& frame_complete, 668 int& frame_size_in_words) { 669 // Before we get into the guts of the C2I adapter, see if we should be here 670 // at all. We've come from compiled code and are attempting to jump to the 671 // interpreter, which means the caller made a static call to get here 672 // (vcalls always get a compiled target if there is one). Check for a 673 // compiled target. If there is one, we need to patch the caller's call. 674 patch_callers_callsite(masm); 675 676 __ bind(skip_fixup); 677 678 bool has_value_argument = false; 679 if (ValueTypePassFieldsAsArgs) { 680 // Is there a value type argument? 681 for (int i = 0; i < sig_extended.length() && !has_value_argument; i++) { 682 has_value_argument = (sig_extended.at(i)._bt == T_VALUETYPE); 683 } 684 if (has_value_argument) { 685 // There is at least a value type argument: we're coming from 686 // compiled code so we have no buffers to back the value 687 // types. Allocate the buffers here with a runtime call. 688 oop_maps = new OopMapSet(); 689 OopMap* map = NULL; 690 691 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 692 693 frame_complete = __ offset(); 694 695 __ set_last_Java_frame(noreg, noreg, NULL); 696 697 __ mov(c_rarg0, r15_thread); 698 __ mov(c_rarg1, rbx); 699 700 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_value_types))); 701 702 oop_maps->add_gc_map((int)(__ pc() - start), map); 703 __ reset_last_Java_frame(false); 704 705 RegisterSaver::restore_live_registers(masm); 706 707 Label no_exception; 708 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 709 __ jcc(Assembler::equal, no_exception); 710 711 __ movptr(Address(r15_thread, JavaThread::vm_result_offset()), (int)NULL_WORD); 712 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset())); 713 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 714 715 __ bind(no_exception); 716 717 // We get an array of objects from the runtime call 718 __ get_vm_result(r13, r15_thread); // Use r13 as temporary because r10 is trashed by movptr() 719 __ get_vm_result_2(rbx, r15_thread); // TODO: required to keep the callee Method live? 720 __ mov(r10, r13); 721 } 722 } 723 724 // Since all args are passed on the stack, total_args_passed * 725 // Interpreter::stackElementSize is the space we need. Plus 1 because 726 // we also account for the return address location since 727 // we store it first rather than hold it in rax across all the shuffling 728 int total_args_passed = compute_total_args_passed_int(sig_extended); 729 int extraspace = (total_args_passed * Interpreter::stackElementSize) + wordSize; 730 731 // stack is aligned, keep it that way 732 extraspace = round_to(extraspace, 2*wordSize); 733 734 // Get return address 735 __ pop(rax); 736 737 // set senderSP value 738 __ mov(r13, rsp); 739 740 __ subptr(rsp, extraspace); 741 742 // Store the return address in the expected location 743 __ movptr(Address(rsp, 0), rax); 744 745 // Now write the args into the outgoing interpreter space 746 747 // next_arg_comp is the next argument from the compiler point of 748 // view (value type fields are passed in registers/on the stack). In 749 // sig_extended, a value type argument starts with: T_VALUETYPE, 750 // followed by the types of the fields of the value type and T_VOID 751 // to mark the end of the value type. ignored counts the number of 752 // T_VALUETYPE/T_VOID. next_vt_arg is the next value type argument: 753 // used to get the buffer for that argument from the pool of buffers 754 // we allocated above and want to pass to the 755 // interpreter. next_arg_int is the next argument from the 756 // interpreter point of view (value types are passed by reference). 757 bool has_oop_field = false; 758 for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0; 759 next_arg_comp < sig_extended.length(); next_arg_comp++) { 760 assert(ignored <= next_arg_comp, "shouldn't skip over more slot than there are arguments"); 761 assert(next_arg_int < total_args_passed, "more arguments for the interpreter than expected?"); 762 BasicType bt = sig_extended.at(next_arg_comp)._bt; 763 int st_off = (total_args_passed - next_arg_int) * Interpreter::stackElementSize; 764 if (!ValueTypePassFieldsAsArgs || bt != T_VALUETYPE) { 765 int next_off = st_off - Interpreter::stackElementSize; 766 const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off; 767 const VMRegPair reg_pair = regs[next_arg_comp-ignored]; 768 size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4; 769 gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended.at(next_arg_comp-1)._bt : T_ILLEGAL, 770 size_in_bytes, reg_pair, Address(rsp, offset), extraspace, false); 771 next_arg_int++; 772 #ifdef ASSERT 773 if (bt == T_LONG || bt == T_DOUBLE) { 774 // Overwrite the unused slot with known junk 775 __ mov64(rax, CONST64(0xdeadffffdeadaaaa)); 776 __ movptr(Address(rsp, st_off), rax); 777 } 778 #endif /* ASSERT */ 779 } else { 780 ignored++; 781 // get the buffer from the just allocated pool of buffers 782 int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_VALUETYPE); 783 __ load_heap_oop(r11, Address(r10, index)); 784 next_vt_arg++; next_arg_int++; 785 int vt = 1; 786 // write fields we get from compiled code in registers/stack 787 // slots to the buffer: we know we are done with that value type 788 // argument when we hit the T_VOID that acts as an end of value 789 // type delimiter for this value type. Value types are flattened 790 // so we might encounter embedded value types. Each entry in 791 // sig_extended contains a field offset in the buffer. 792 do { 793 next_arg_comp++; 794 BasicType bt = sig_extended.at(next_arg_comp)._bt; 795 BasicType prev_bt = sig_extended.at(next_arg_comp-1)._bt; 796 if (bt == T_VALUETYPE) { 797 vt++; 798 ignored++; 799 } else if (bt == T_VOID && 800 prev_bt != T_LONG && 801 prev_bt != T_DOUBLE) { 802 vt--; 803 ignored++; 804 } else { 805 int off = sig_extended.at(next_arg_comp)._offset; 806 assert(off > 0, "offset in object should be positive"); 807 size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize; 808 bool is_oop = (bt == T_OBJECT || bt == T_ARRAY); 809 has_oop_field = has_oop_field || is_oop; 810 gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended.at(next_arg_comp-1)._bt : T_ILLEGAL, 811 size_in_bytes, regs[next_arg_comp-ignored], Address(r11, off), extraspace, is_oop); 812 } 813 } while (vt != 0); 814 // pass the buffer to the interpreter 815 __ movptr(Address(rsp, st_off), r11); 816 } 817 } 818 819 // If a value type was allocated and initialized, apply post barrier to all oop fields 820 if (has_value_argument && has_oop_field) { 821 __ push(r13); // save senderSP 822 __ push(rbx); // save callee 823 // Allocate argument register save area 824 if (frame::arg_reg_save_area_bytes != 0) { 825 __ subptr(rsp, frame::arg_reg_save_area_bytes); 826 } 827 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::apply_post_barriers), r15_thread, r10); 828 // De-allocate argument register save area 829 if (frame::arg_reg_save_area_bytes != 0) { 830 __ addptr(rsp, frame::arg_reg_save_area_bytes); 831 } 832 __ pop(rbx); // restore callee 833 __ pop(r13); // restore sender SP 834 } 835 836 // Schedule the branch target address early. 837 __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset()))); 838 __ jmp(rcx); 839 } 840 841 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg, 842 address code_start, address code_end, 843 Label& L_ok) { 844 Label L_fail; 845 __ lea(temp_reg, ExternalAddress(code_start)); 846 __ cmpptr(pc_reg, temp_reg); 847 __ jcc(Assembler::belowEqual, L_fail); 848 __ lea(temp_reg, ExternalAddress(code_end)); 849 __ cmpptr(pc_reg, temp_reg); 850 __ jcc(Assembler::below, L_ok); 851 __ bind(L_fail); 852 } 853 854 static void gen_i2c_adapter_helper(MacroAssembler* masm, 855 BasicType bt, 856 BasicType prev_bt, 857 size_t size_in_bytes, 858 const VMRegPair& reg_pair, 859 const Address& from, 860 bool is_oop) { 861 assert(bt != T_VALUETYPE || !ValueTypePassFieldsAsArgs, "no value type here"); 862 if (bt == T_VOID) { 863 // Longs and doubles are passed in native word order, but misaligned 864 // in the 32-bit build. 865 assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half"); 866 return; 867 } 868 assert(!reg_pair.second()->is_valid() || reg_pair.first()->next() == reg_pair.second(), 869 "scrambled load targets?"); 870 871 bool wide = (size_in_bytes == wordSize); 872 VMReg r_1 = reg_pair.first(); 873 VMReg r_2 = reg_pair.second(); 874 assert(r_2->is_valid() == wide, "invalid size"); 875 if (!r_1->is_valid()) { 876 assert(!r_2->is_valid(), "must be invalid"); 877 return; 878 } 879 880 bool is_signed = (bt != T_CHAR) && (bt != T_BOOLEAN); 881 if (!r_1->is_XMMRegister()) { 882 // We can use r13 as a temp here because compiled code doesn't need r13 as an input 883 // and if we end up going thru a c2i because of a miss a reasonable value of r13 884 // will be generated. 885 Register dst = r_1->is_stack() ? r13 : r_1->as_Register(); 886 if (is_oop) { 887 __ load_heap_oop(dst, from); 888 } else { 889 __ load_sized_value(dst, from, size_in_bytes, is_signed); 890 } 891 if (r_1->is_stack()) { 892 // Convert stack slot to an SP offset (+ wordSize to account for return address) 893 int st_off = reg_pair.first()->reg2stack() * VMRegImpl::stack_slot_size + wordSize; 894 __ movq(Address(rsp, st_off), dst); 895 } 896 } else { 897 if (wide) { 898 __ movdbl(r_1->as_XMMRegister(), from); 899 } else { 900 __ movflt(r_1->as_XMMRegister(), from); 901 } 902 } 903 } 904 905 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, 906 int comp_args_on_stack, 907 const GrowableArray<SigEntry>& sig_extended, 908 const VMRegPair *regs) { 909 910 // Note: r13 contains the senderSP on entry. We must preserve it since 911 // we may do a i2c -> c2i transition if we lose a race where compiled 912 // code goes non-entrant while we get args ready. 913 // In addition we use r13 to locate all the interpreter args as 914 // we must align the stack to 16 bytes on an i2c entry else we 915 // lose alignment we expect in all compiled code and register 916 // save code can segv when fxsave instructions find improperly 917 // aligned stack pointer. 918 919 // Adapters can be frameless because they do not require the caller 920 // to perform additional cleanup work, such as correcting the stack pointer. 921 // An i2c adapter is frameless because the *caller* frame, which is interpreted, 922 // routinely repairs its own stack pointer (from interpreter_frame_last_sp), 923 // even if a callee has modified the stack pointer. 924 // A c2i adapter is frameless because the *callee* frame, which is interpreted, 925 // routinely repairs its caller's stack pointer (from sender_sp, which is set 926 // up via the senderSP register). 927 // In other words, if *either* the caller or callee is interpreted, we can 928 // get the stack pointer repaired after a call. 929 // This is why c2i and i2c adapters cannot be indefinitely composed. 930 // In particular, if a c2i adapter were to somehow call an i2c adapter, 931 // both caller and callee would be compiled methods, and neither would 932 // clean up the stack pointer changes performed by the two adapters. 933 // If this happens, control eventually transfers back to the compiled 934 // caller, but with an uncorrected stack, causing delayed havoc. 935 936 // Pick up the return address 937 __ movptr(rax, Address(rsp, 0)); 938 939 if (VerifyAdapterCalls && 940 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) { 941 // So, let's test for cascading c2i/i2c adapters right now. 942 // assert(Interpreter::contains($return_addr) || 943 // StubRoutines::contains($return_addr), 944 // "i2c adapter must return to an interpreter frame"); 945 __ block_comment("verify_i2c { "); 946 Label L_ok; 947 if (Interpreter::code() != NULL) 948 range_check(masm, rax, r11, 949 Interpreter::code()->code_start(), Interpreter::code()->code_end(), 950 L_ok); 951 if (StubRoutines::code1() != NULL) 952 range_check(masm, rax, r11, 953 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(), 954 L_ok); 955 if (StubRoutines::code2() != NULL) 956 range_check(masm, rax, r11, 957 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(), 958 L_ok); 959 const char* msg = "i2c adapter must return to an interpreter frame"; 960 __ block_comment(msg); 961 __ stop(msg); 962 __ bind(L_ok); 963 __ block_comment("} verify_i2ce "); 964 } 965 966 // Must preserve original SP for loading incoming arguments because 967 // we need to align the outgoing SP for compiled code. 968 __ movptr(r11, rsp); 969 970 // Cut-out for having no stack args. Since up to 2 int/oop args are passed 971 // in registers, we will occasionally have no stack args. 972 int comp_words_on_stack = 0; 973 if (comp_args_on_stack) { 974 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in 975 // registers are below. By subtracting stack0, we either get a negative 976 // number (all values in registers) or the maximum stack slot accessed. 977 978 // Convert 4-byte c2 stack slots to words. 979 comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; 980 // Round up to miminum stack alignment, in wordSize 981 comp_words_on_stack = round_to(comp_words_on_stack, 2); 982 __ subptr(rsp, comp_words_on_stack * wordSize); 983 } 984 985 986 // Ensure compiled code always sees stack at proper alignment 987 __ andptr(rsp, -16); 988 989 // push the return address and misalign the stack that youngest frame always sees 990 // as far as the placement of the call instruction 991 __ push(rax); 992 993 // Put saved SP in another register 994 const Register saved_sp = rax; 995 __ movptr(saved_sp, r11); 996 997 // Will jump to the compiled code just as if compiled code was doing it. 998 // Pre-load the register-jump target early, to schedule it better. 999 __ movptr(r11, Address(rbx, in_bytes(Method::from_compiled_offset()))); 1000 1001 #if INCLUDE_JVMCI 1002 if (EnableJVMCI || UseAOT) { 1003 // check if this call should be routed towards a specific entry point 1004 __ cmpptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0); 1005 Label no_alternative_target; 1006 __ jcc(Assembler::equal, no_alternative_target); 1007 __ movptr(r11, Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset()))); 1008 __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0); 1009 __ bind(no_alternative_target); 1010 } 1011 #endif // INCLUDE_JVMCI 1012 1013 int total_args_passed = compute_total_args_passed_int(sig_extended); 1014 // Now generate the shuffle code. Pick up all register args and move the 1015 // rest through the floating point stack top. 1016 1017 // next_arg_comp is the next argument from the compiler point of 1018 // view (value type fields are passed in registers/on the stack). In 1019 // sig_extended, a value type argument starts with: T_VALUETYPE, 1020 // followed by the types of the fields of the value type and T_VOID 1021 // to mark the end of the value type. ignored counts the number of 1022 // T_VALUETYPE/T_VOID. next_arg_int is the next argument from the 1023 // interpreter point of view (value types are passed by reference). 1024 for (int next_arg_comp = 0, ignored = 0, next_arg_int = 0; next_arg_comp < sig_extended.length(); next_arg_comp++) { 1025 assert(ignored <= next_arg_comp, "shouldn't skip over more slot than there are arguments"); 1026 assert(next_arg_int < total_args_passed, "more arguments from the interpreter than expected?"); 1027 BasicType bt = sig_extended.at(next_arg_comp)._bt; 1028 int ld_off = (total_args_passed - next_arg_int)*Interpreter::stackElementSize; 1029 if (!ValueTypePassFieldsAsArgs || bt != T_VALUETYPE) { 1030 // Load in argument order going down. 1031 // Point to interpreter value (vs. tag) 1032 int next_off = ld_off - Interpreter::stackElementSize; 1033 int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off; 1034 const VMRegPair reg_pair = regs[next_arg_comp-ignored]; 1035 size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4; 1036 gen_i2c_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended.at(next_arg_comp-1)._bt : T_ILLEGAL, 1037 size_in_bytes, reg_pair, Address(saved_sp, offset), false); 1038 next_arg_int++; 1039 } else { 1040 next_arg_int++; 1041 ignored++; 1042 // get the buffer for that value type 1043 __ movptr(r10, Address(saved_sp, ld_off)); 1044 int vt = 1; 1045 // load fields to registers/stack slots from the buffer: we know 1046 // we are done with that value type argument when we hit the 1047 // T_VOID that acts as an end of value type delimiter for this 1048 // value type. Value types are flattened so we might encounter 1049 // embedded value types. Each entry in sig_extended contains a 1050 // field offset in the buffer. 1051 do { 1052 next_arg_comp++; 1053 BasicType bt = sig_extended.at(next_arg_comp)._bt; 1054 BasicType prev_bt = sig_extended.at(next_arg_comp-1)._bt; 1055 if (bt == T_VALUETYPE) { 1056 vt++; 1057 ignored++; 1058 } else if (bt == T_VOID && 1059 prev_bt != T_LONG && 1060 prev_bt != T_DOUBLE) { 1061 vt--; 1062 ignored++; 1063 } else { 1064 int off = sig_extended.at(next_arg_comp)._offset; 1065 assert(off > 0, "offset in object should be positive"); 1066 size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize; 1067 bool is_oop = (bt == T_OBJECT || bt == T_ARRAY); 1068 gen_i2c_adapter_helper(masm, bt, prev_bt, size_in_bytes, regs[next_arg_comp - ignored], Address(r10, off), is_oop); 1069 } 1070 } while (vt != 0); 1071 } 1072 } 1073 1074 // 6243940 We might end up in handle_wrong_method if 1075 // the callee is deoptimized as we race thru here. If that 1076 // happens we don't want to take a safepoint because the 1077 // caller frame will look interpreted and arguments are now 1078 // "compiled" so it is much better to make this transition 1079 // invisible to the stack walking code. Unfortunately if 1080 // we try and find the callee by normal means a safepoint 1081 // is possible. So we stash the desired callee in the thread 1082 // and the vm will find there should this case occur. 1083 1084 __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx); 1085 1086 // put Method* where a c2i would expect should we end up there 1087 // only needed because of c2 resolve stubs return Method* as a result in 1088 // rax 1089 __ mov(rax, rbx); 1090 __ jmp(r11); 1091 } 1092 1093 // --------------------------------------------------------------- 1094 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 1095 int comp_args_on_stack, 1096 const GrowableArray<SigEntry>& sig_extended, 1097 const VMRegPair *regs, 1098 AdapterFingerPrint* fingerprint, 1099 AdapterBlob*& new_adapter) { 1100 address i2c_entry = __ pc(); 1101 1102 gen_i2c_adapter(masm, comp_args_on_stack, sig_extended, regs); 1103 1104 // ------------------------------------------------------------------------- 1105 // Generate a C2I adapter. On entry we know rbx holds the Method* during calls 1106 // to the interpreter. The args start out packed in the compiled layout. They 1107 // need to be unpacked into the interpreter layout. This will almost always 1108 // require some stack space. We grow the current (compiled) stack, then repack 1109 // the args. We finally end in a jump to the generic interpreter entry point. 1110 // On exit from the interpreter, the interpreter will restore our SP (lest the 1111 // compiled code, which relys solely on SP and not RBP, get sick). 1112 1113 address c2i_unverified_entry = __ pc(); 1114 Label skip_fixup; 1115 Label ok; 1116 1117 Register holder = rax; 1118 Register receiver = j_rarg0; 1119 Register temp = rbx; 1120 1121 { 1122 __ load_klass(temp, receiver); 1123 __ cmpptr(temp, Address(holder, CompiledICHolder::holder_klass_offset())); 1124 __ movptr(rbx, Address(holder, CompiledICHolder::holder_method_offset())); 1125 __ jcc(Assembler::equal, ok); 1126 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1127 1128 __ bind(ok); 1129 // Method might have been compiled since the call site was patched to 1130 // interpreted if that is the case treat it as a miss so we can get 1131 // the call site corrected. 1132 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD); 1133 __ jcc(Assembler::equal, skip_fixup); 1134 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 1135 } 1136 1137 address c2i_entry = __ pc(); 1138 1139 OopMapSet* oop_maps = NULL; 1140 int frame_complete = CodeOffsets::frame_never_safe; 1141 int frame_size_in_words = 0; 1142 gen_c2i_adapter(masm, sig_extended, regs, skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words); 1143 1144 __ flush(); 1145 new_adapter = AdapterBlob::create(masm->code(), frame_complete, frame_size_in_words, oop_maps); 1146 1147 // If value types are passed as fields, save the extended signature as symbol in 1148 // the AdapterHandlerEntry to be used by nmethod::preserve_callee_argument_oops(). 1149 Symbol* extended_signature = NULL; 1150 if (ValueTypePassFieldsAsArgs) { 1151 bool has_value_argument = false; 1152 Thread* THREAD = Thread::current(); 1153 ResourceMark rm(THREAD); 1154 int length = sig_extended.length(); 1155 char* sig_str = NEW_RESOURCE_ARRAY(char, 2*length + 3); 1156 int idx = 0; 1157 sig_str[idx++] = '('; 1158 for (int index = 0; index < length; index++) { 1159 BasicType bt = sig_extended.at(index)._bt; 1160 if (bt == T_VALUETYPE || bt == T_VOID) { 1161 has_value_argument = true; 1162 continue; // Ignore wrapper types 1163 } 1164 sig_str[idx++] = type2char(bt); 1165 if (bt == T_OBJECT) { 1166 sig_str[idx++] = ';'; 1167 } else if (bt == T_ARRAY) { 1168 // We don't know the array element type, put void as placeholder 1169 sig_str[idx++] = 'V'; 1170 } 1171 } 1172 sig_str[idx++] = ')'; 1173 sig_str[idx++] = '\0'; 1174 if (has_value_argument) { 1175 // Extended signature is only required if a value type argument is passed 1176 extended_signature = SymbolTable::new_permanent_symbol(sig_str, THREAD); 1177 } 1178 } 1179 1180 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, extended_signature); 1181 } 1182 1183 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 1184 VMRegPair *regs, 1185 VMRegPair *regs2, 1186 int total_args_passed) { 1187 assert(regs2 == NULL, "not needed on x86"); 1188 // We return the amount of VMRegImpl stack slots we need to reserve for all 1189 // the arguments NOT counting out_preserve_stack_slots. 1190 1191 // NOTE: These arrays will have to change when c1 is ported 1192 #ifdef _WIN64 1193 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 1194 c_rarg0, c_rarg1, c_rarg2, c_rarg3 1195 }; 1196 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = { 1197 c_farg0, c_farg1, c_farg2, c_farg3 1198 }; 1199 #else 1200 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = { 1201 c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5 1202 }; 1203 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = { 1204 c_farg0, c_farg1, c_farg2, c_farg3, 1205 c_farg4, c_farg5, c_farg6, c_farg7 1206 }; 1207 #endif // _WIN64 1208 1209 1210 uint int_args = 0; 1211 uint fp_args = 0; 1212 uint stk_args = 0; // inc by 2 each time 1213 1214 for (int i = 0; i < total_args_passed; i++) { 1215 switch (sig_bt[i]) { 1216 case T_BOOLEAN: 1217 case T_CHAR: 1218 case T_BYTE: 1219 case T_SHORT: 1220 case T_INT: 1221 if (int_args < Argument::n_int_register_parameters_c) { 1222 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg()); 1223 #ifdef _WIN64 1224 fp_args++; 1225 // Allocate slots for callee to stuff register args the stack. 1226 stk_args += 2; 1227 #endif 1228 } else { 1229 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 1230 stk_args += 2; 1231 } 1232 break; 1233 case T_LONG: 1234 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 1235 // fall through 1236 case T_OBJECT: 1237 case T_ARRAY: 1238 case T_ADDRESS: 1239 case T_METADATA: 1240 if (int_args < Argument::n_int_register_parameters_c) { 1241 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg()); 1242 #ifdef _WIN64 1243 fp_args++; 1244 stk_args += 2; 1245 #endif 1246 } else { 1247 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 1248 stk_args += 2; 1249 } 1250 break; 1251 case T_FLOAT: 1252 if (fp_args < Argument::n_float_register_parameters_c) { 1253 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg()); 1254 #ifdef _WIN64 1255 int_args++; 1256 // Allocate slots for callee to stuff register args the stack. 1257 stk_args += 2; 1258 #endif 1259 } else { 1260 regs[i].set1(VMRegImpl::stack2reg(stk_args)); 1261 stk_args += 2; 1262 } 1263 break; 1264 case T_DOUBLE: 1265 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half"); 1266 if (fp_args < Argument::n_float_register_parameters_c) { 1267 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg()); 1268 #ifdef _WIN64 1269 int_args++; 1270 // Allocate slots for callee to stuff register args the stack. 1271 stk_args += 2; 1272 #endif 1273 } else { 1274 regs[i].set2(VMRegImpl::stack2reg(stk_args)); 1275 stk_args += 2; 1276 } 1277 break; 1278 case T_VOID: // Halves of longs and doubles 1279 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 1280 regs[i].set_bad(); 1281 break; 1282 default: 1283 ShouldNotReachHere(); 1284 break; 1285 } 1286 } 1287 #ifdef _WIN64 1288 // windows abi requires that we always allocate enough stack space 1289 // for 4 64bit registers to be stored down. 1290 if (stk_args < 8) { 1291 stk_args = 8; 1292 } 1293 #endif // _WIN64 1294 1295 return stk_args; 1296 } 1297 1298 // On 64 bit we will store integer like items to the stack as 1299 // 64 bits items (sparc abi) even though java would only store 1300 // 32bits for a parameter. On 32bit it will simply be 32 bits 1301 // So this routine will do 32->32 on 32bit and 32->64 on 64bit 1302 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1303 if (src.first()->is_stack()) { 1304 if (dst.first()->is_stack()) { 1305 // stack to stack 1306 __ movslq(rax, Address(rbp, reg2offset_in(src.first()))); 1307 __ movq(Address(rsp, reg2offset_out(dst.first())), rax); 1308 } else { 1309 // stack to reg 1310 __ movslq(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first()))); 1311 } 1312 } else if (dst.first()->is_stack()) { 1313 // reg to stack 1314 // Do we really have to sign extend??? 1315 // __ movslq(src.first()->as_Register(), src.first()->as_Register()); 1316 __ movq(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register()); 1317 } else { 1318 // Do we really have to sign extend??? 1319 // __ movslq(dst.first()->as_Register(), src.first()->as_Register()); 1320 if (dst.first() != src.first()) { 1321 __ movq(dst.first()->as_Register(), src.first()->as_Register()); 1322 } 1323 } 1324 } 1325 1326 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1327 if (src.first()->is_stack()) { 1328 if (dst.first()->is_stack()) { 1329 // stack to stack 1330 __ movq(rax, Address(rbp, reg2offset_in(src.first()))); 1331 __ movq(Address(rsp, reg2offset_out(dst.first())), rax); 1332 } else { 1333 // stack to reg 1334 __ movq(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first()))); 1335 } 1336 } else if (dst.first()->is_stack()) { 1337 // reg to stack 1338 __ movq(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register()); 1339 } else { 1340 if (dst.first() != src.first()) { 1341 __ movq(dst.first()->as_Register(), src.first()->as_Register()); 1342 } 1343 } 1344 } 1345 1346 // An oop arg. Must pass a handle not the oop itself 1347 static void object_move(MacroAssembler* masm, 1348 OopMap* map, 1349 int oop_handle_offset, 1350 int framesize_in_slots, 1351 VMRegPair src, 1352 VMRegPair dst, 1353 bool is_receiver, 1354 int* receiver_offset) { 1355 1356 // must pass a handle. First figure out the location we use as a handle 1357 1358 Register rHandle = dst.first()->is_stack() ? rax : dst.first()->as_Register(); 1359 1360 // See if oop is NULL if it is we need no handle 1361 1362 if (src.first()->is_stack()) { 1363 1364 // Oop is already on the stack as an argument 1365 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1366 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 1367 if (is_receiver) { 1368 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 1369 } 1370 1371 __ cmpptr(Address(rbp, reg2offset_in(src.first())), (int32_t)NULL_WORD); 1372 __ lea(rHandle, Address(rbp, reg2offset_in(src.first()))); 1373 // conditionally move a NULL 1374 __ cmovptr(Assembler::equal, rHandle, Address(rbp, reg2offset_in(src.first()))); 1375 } else { 1376 1377 // Oop is in an a register we must store it to the space we reserve 1378 // on the stack for oop_handles and pass a handle if oop is non-NULL 1379 1380 const Register rOop = src.first()->as_Register(); 1381 int oop_slot; 1382 if (rOop == j_rarg0) 1383 oop_slot = 0; 1384 else if (rOop == j_rarg1) 1385 oop_slot = 1; 1386 else if (rOop == j_rarg2) 1387 oop_slot = 2; 1388 else if (rOop == j_rarg3) 1389 oop_slot = 3; 1390 else if (rOop == j_rarg4) 1391 oop_slot = 4; 1392 else { 1393 assert(rOop == j_rarg5, "wrong register"); 1394 oop_slot = 5; 1395 } 1396 1397 oop_slot = oop_slot * VMRegImpl::slots_per_word + oop_handle_offset; 1398 int offset = oop_slot*VMRegImpl::stack_slot_size; 1399 1400 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1401 // Store oop in handle area, may be NULL 1402 __ movptr(Address(rsp, offset), rOop); 1403 if (is_receiver) { 1404 *receiver_offset = offset; 1405 } 1406 1407 __ cmpptr(rOop, (int32_t)NULL_WORD); 1408 __ lea(rHandle, Address(rsp, offset)); 1409 // conditionally move a NULL from the handle area where it was just stored 1410 __ cmovptr(Assembler::equal, rHandle, Address(rsp, offset)); 1411 } 1412 1413 // If arg is on the stack then place it otherwise it is already in correct reg. 1414 if (dst.first()->is_stack()) { 1415 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle); 1416 } 1417 } 1418 1419 // A float arg may have to do float reg int reg conversion 1420 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1421 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move"); 1422 1423 // The calling conventions assures us that each VMregpair is either 1424 // all really one physical register or adjacent stack slots. 1425 // This greatly simplifies the cases here compared to sparc. 1426 1427 if (src.first()->is_stack()) { 1428 if (dst.first()->is_stack()) { 1429 __ movl(rax, Address(rbp, reg2offset_in(src.first()))); 1430 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax); 1431 } else { 1432 // stack to reg 1433 assert(dst.first()->is_XMMRegister(), "only expect xmm registers as parameters"); 1434 __ movflt(dst.first()->as_XMMRegister(), Address(rbp, reg2offset_in(src.first()))); 1435 } 1436 } else if (dst.first()->is_stack()) { 1437 // reg to stack 1438 assert(src.first()->is_XMMRegister(), "only expect xmm registers as parameters"); 1439 __ movflt(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister()); 1440 } else { 1441 // reg to reg 1442 // In theory these overlap but the ordering is such that this is likely a nop 1443 if ( src.first() != dst.first()) { 1444 __ movdbl(dst.first()->as_XMMRegister(), src.first()->as_XMMRegister()); 1445 } 1446 } 1447 } 1448 1449 // A long move 1450 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1451 1452 // The calling conventions assures us that each VMregpair is either 1453 // all really one physical register or adjacent stack slots. 1454 // This greatly simplifies the cases here compared to sparc. 1455 1456 if (src.is_single_phys_reg() ) { 1457 if (dst.is_single_phys_reg()) { 1458 if (dst.first() != src.first()) { 1459 __ mov(dst.first()->as_Register(), src.first()->as_Register()); 1460 } 1461 } else { 1462 assert(dst.is_single_reg(), "not a stack pair"); 1463 __ movq(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register()); 1464 } 1465 } else if (dst.is_single_phys_reg()) { 1466 assert(src.is_single_reg(), "not a stack pair"); 1467 __ movq(dst.first()->as_Register(), Address(rbp, reg2offset_out(src.first()))); 1468 } else { 1469 assert(src.is_single_reg() && dst.is_single_reg(), "not stack pairs"); 1470 __ movq(rax, Address(rbp, reg2offset_in(src.first()))); 1471 __ movq(Address(rsp, reg2offset_out(dst.first())), rax); 1472 } 1473 } 1474 1475 // A double move 1476 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1477 1478 // The calling conventions assures us that each VMregpair is either 1479 // all really one physical register or adjacent stack slots. 1480 // This greatly simplifies the cases here compared to sparc. 1481 1482 if (src.is_single_phys_reg() ) { 1483 if (dst.is_single_phys_reg()) { 1484 // In theory these overlap but the ordering is such that this is likely a nop 1485 if ( src.first() != dst.first()) { 1486 __ movdbl(dst.first()->as_XMMRegister(), src.first()->as_XMMRegister()); 1487 } 1488 } else { 1489 assert(dst.is_single_reg(), "not a stack pair"); 1490 __ movdbl(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister()); 1491 } 1492 } else if (dst.is_single_phys_reg()) { 1493 assert(src.is_single_reg(), "not a stack pair"); 1494 __ movdbl(dst.first()->as_XMMRegister(), Address(rbp, reg2offset_out(src.first()))); 1495 } else { 1496 assert(src.is_single_reg() && dst.is_single_reg(), "not stack pairs"); 1497 __ movq(rax, Address(rbp, reg2offset_in(src.first()))); 1498 __ movq(Address(rsp, reg2offset_out(dst.first())), rax); 1499 } 1500 } 1501 1502 1503 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1504 // We always ignore the frame_slots arg and just use the space just below frame pointer 1505 // which by this time is free to use 1506 switch (ret_type) { 1507 case T_FLOAT: 1508 __ movflt(Address(rbp, -wordSize), xmm0); 1509 break; 1510 case T_DOUBLE: 1511 __ movdbl(Address(rbp, -wordSize), xmm0); 1512 break; 1513 case T_VOID: break; 1514 default: { 1515 __ movptr(Address(rbp, -wordSize), rax); 1516 } 1517 } 1518 } 1519 1520 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1521 // We always ignore the frame_slots arg and just use the space just below frame pointer 1522 // which by this time is free to use 1523 switch (ret_type) { 1524 case T_FLOAT: 1525 __ movflt(xmm0, Address(rbp, -wordSize)); 1526 break; 1527 case T_DOUBLE: 1528 __ movdbl(xmm0, Address(rbp, -wordSize)); 1529 break; 1530 case T_VOID: break; 1531 default: { 1532 __ movptr(rax, Address(rbp, -wordSize)); 1533 } 1534 } 1535 } 1536 1537 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1538 for ( int i = first_arg ; i < arg_count ; i++ ) { 1539 if (args[i].first()->is_Register()) { 1540 __ push(args[i].first()->as_Register()); 1541 } else if (args[i].first()->is_XMMRegister()) { 1542 __ subptr(rsp, 2*wordSize); 1543 __ movdbl(Address(rsp, 0), args[i].first()->as_XMMRegister()); 1544 } 1545 } 1546 } 1547 1548 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) { 1549 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) { 1550 if (args[i].first()->is_Register()) { 1551 __ pop(args[i].first()->as_Register()); 1552 } else if (args[i].first()->is_XMMRegister()) { 1553 __ movdbl(args[i].first()->as_XMMRegister(), Address(rsp, 0)); 1554 __ addptr(rsp, 2*wordSize); 1555 } 1556 } 1557 } 1558 1559 1560 static void save_or_restore_arguments(MacroAssembler* masm, 1561 const int stack_slots, 1562 const int total_in_args, 1563 const int arg_save_area, 1564 OopMap* map, 1565 VMRegPair* in_regs, 1566 BasicType* in_sig_bt) { 1567 // if map is non-NULL then the code should store the values, 1568 // otherwise it should load them. 1569 int slot = arg_save_area; 1570 // Save down double word first 1571 for ( int i = 0; i < total_in_args; i++) { 1572 if (in_regs[i].first()->is_XMMRegister() && in_sig_bt[i] == T_DOUBLE) { 1573 int offset = slot * VMRegImpl::stack_slot_size; 1574 slot += VMRegImpl::slots_per_word; 1575 assert(slot <= stack_slots, "overflow"); 1576 if (map != NULL) { 1577 __ movdbl(Address(rsp, offset), in_regs[i].first()->as_XMMRegister()); 1578 } else { 1579 __ movdbl(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset)); 1580 } 1581 } 1582 if (in_regs[i].first()->is_Register() && 1583 (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_ARRAY)) { 1584 int offset = slot * VMRegImpl::stack_slot_size; 1585 if (map != NULL) { 1586 __ movq(Address(rsp, offset), in_regs[i].first()->as_Register()); 1587 if (in_sig_bt[i] == T_ARRAY) { 1588 map->set_oop(VMRegImpl::stack2reg(slot));; 1589 } 1590 } else { 1591 __ movq(in_regs[i].first()->as_Register(), Address(rsp, offset)); 1592 } 1593 slot += VMRegImpl::slots_per_word; 1594 } 1595 } 1596 // Save or restore single word registers 1597 for ( int i = 0; i < total_in_args; i++) { 1598 if (in_regs[i].first()->is_Register()) { 1599 int offset = slot * VMRegImpl::stack_slot_size; 1600 slot++; 1601 assert(slot <= stack_slots, "overflow"); 1602 1603 // Value is in an input register pass we must flush it to the stack 1604 const Register reg = in_regs[i].first()->as_Register(); 1605 switch (in_sig_bt[i]) { 1606 case T_BOOLEAN: 1607 case T_CHAR: 1608 case T_BYTE: 1609 case T_SHORT: 1610 case T_INT: 1611 if (map != NULL) { 1612 __ movl(Address(rsp, offset), reg); 1613 } else { 1614 __ movl(reg, Address(rsp, offset)); 1615 } 1616 break; 1617 case T_ARRAY: 1618 case T_LONG: 1619 // handled above 1620 break; 1621 case T_OBJECT: 1622 default: ShouldNotReachHere(); 1623 } 1624 } else if (in_regs[i].first()->is_XMMRegister()) { 1625 if (in_sig_bt[i] == T_FLOAT) { 1626 int offset = slot * VMRegImpl::stack_slot_size; 1627 slot++; 1628 assert(slot <= stack_slots, "overflow"); 1629 if (map != NULL) { 1630 __ movflt(Address(rsp, offset), in_regs[i].first()->as_XMMRegister()); 1631 } else { 1632 __ movflt(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset)); 1633 } 1634 } 1635 } else if (in_regs[i].first()->is_stack()) { 1636 if (in_sig_bt[i] == T_ARRAY && map != NULL) { 1637 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1638 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots)); 1639 } 1640 } 1641 } 1642 } 1643 1644 1645 // Check GCLocker::needs_gc and enter the runtime if it's true. This 1646 // keeps a new JNI critical region from starting until a GC has been 1647 // forced. Save down any oops in registers and describe them in an 1648 // OopMap. 1649 static void check_needs_gc_for_critical_native(MacroAssembler* masm, 1650 int stack_slots, 1651 int total_c_args, 1652 int total_in_args, 1653 int arg_save_area, 1654 OopMapSet* oop_maps, 1655 VMRegPair* in_regs, 1656 BasicType* in_sig_bt) { 1657 __ block_comment("check GCLocker::needs_gc"); 1658 Label cont; 1659 __ cmp8(ExternalAddress((address)GCLocker::needs_gc_address()), false); 1660 __ jcc(Assembler::equal, cont); 1661 1662 // Save down any incoming oops and call into the runtime to halt for a GC 1663 1664 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1665 save_or_restore_arguments(masm, stack_slots, total_in_args, 1666 arg_save_area, map, in_regs, in_sig_bt); 1667 1668 address the_pc = __ pc(); 1669 oop_maps->add_gc_map( __ offset(), map); 1670 __ set_last_Java_frame(rsp, noreg, the_pc); 1671 1672 __ block_comment("block_for_jni_critical"); 1673 __ movptr(c_rarg0, r15_thread); 1674 __ mov(r12, rsp); // remember sp 1675 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 1676 __ andptr(rsp, -16); // align stack as required by ABI 1677 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical))); 1678 __ mov(rsp, r12); // restore sp 1679 __ reinit_heapbase(); 1680 1681 __ reset_last_Java_frame(false); 1682 1683 save_or_restore_arguments(masm, stack_slots, total_in_args, 1684 arg_save_area, NULL, in_regs, in_sig_bt); 1685 1686 __ bind(cont); 1687 #ifdef ASSERT 1688 if (StressCriticalJNINatives) { 1689 // Stress register saving 1690 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1691 save_or_restore_arguments(masm, stack_slots, total_in_args, 1692 arg_save_area, map, in_regs, in_sig_bt); 1693 // Destroy argument registers 1694 for (int i = 0; i < total_in_args - 1; i++) { 1695 if (in_regs[i].first()->is_Register()) { 1696 const Register reg = in_regs[i].first()->as_Register(); 1697 __ xorptr(reg, reg); 1698 } else if (in_regs[i].first()->is_XMMRegister()) { 1699 __ xorpd(in_regs[i].first()->as_XMMRegister(), in_regs[i].first()->as_XMMRegister()); 1700 } else if (in_regs[i].first()->is_FloatRegister()) { 1701 ShouldNotReachHere(); 1702 } else if (in_regs[i].first()->is_stack()) { 1703 // Nothing to do 1704 } else { 1705 ShouldNotReachHere(); 1706 } 1707 if (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_DOUBLE) { 1708 i++; 1709 } 1710 } 1711 1712 save_or_restore_arguments(masm, stack_slots, total_in_args, 1713 arg_save_area, NULL, in_regs, in_sig_bt); 1714 } 1715 #endif 1716 } 1717 1718 // Unpack an array argument into a pointer to the body and the length 1719 // if the array is non-null, otherwise pass 0 for both. 1720 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) { 1721 Register tmp_reg = rax; 1722 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg, 1723 "possible collision"); 1724 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg, 1725 "possible collision"); 1726 1727 __ block_comment("unpack_array_argument {"); 1728 1729 // Pass the length, ptr pair 1730 Label is_null, done; 1731 VMRegPair tmp; 1732 tmp.set_ptr(tmp_reg->as_VMReg()); 1733 if (reg.first()->is_stack()) { 1734 // Load the arg up from the stack 1735 move_ptr(masm, reg, tmp); 1736 reg = tmp; 1737 } 1738 __ testptr(reg.first()->as_Register(), reg.first()->as_Register()); 1739 __ jccb(Assembler::equal, is_null); 1740 __ lea(tmp_reg, Address(reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type))); 1741 move_ptr(masm, tmp, body_arg); 1742 // load the length relative to the body. 1743 __ movl(tmp_reg, Address(tmp_reg, arrayOopDesc::length_offset_in_bytes() - 1744 arrayOopDesc::base_offset_in_bytes(in_elem_type))); 1745 move32_64(masm, tmp, length_arg); 1746 __ jmpb(done); 1747 __ bind(is_null); 1748 // Pass zeros 1749 __ xorptr(tmp_reg, tmp_reg); 1750 move_ptr(masm, tmp, body_arg); 1751 move32_64(masm, tmp, length_arg); 1752 __ bind(done); 1753 1754 __ block_comment("} unpack_array_argument"); 1755 } 1756 1757 1758 // Different signatures may require very different orders for the move 1759 // to avoid clobbering other arguments. There's no simple way to 1760 // order them safely. Compute a safe order for issuing stores and 1761 // break any cycles in those stores. This code is fairly general but 1762 // it's not necessary on the other platforms so we keep it in the 1763 // platform dependent code instead of moving it into a shared file. 1764 // (See bugs 7013347 & 7145024.) 1765 // Note that this code is specific to LP64. 1766 class ComputeMoveOrder: public StackObj { 1767 class MoveOperation: public ResourceObj { 1768 friend class ComputeMoveOrder; 1769 private: 1770 VMRegPair _src; 1771 VMRegPair _dst; 1772 int _src_index; 1773 int _dst_index; 1774 bool _processed; 1775 MoveOperation* _next; 1776 MoveOperation* _prev; 1777 1778 static int get_id(VMRegPair r) { 1779 return r.first()->value(); 1780 } 1781 1782 public: 1783 MoveOperation(int src_index, VMRegPair src, int dst_index, VMRegPair dst): 1784 _src(src) 1785 , _src_index(src_index) 1786 , _dst(dst) 1787 , _dst_index(dst_index) 1788 , _next(NULL) 1789 , _prev(NULL) 1790 , _processed(false) { 1791 } 1792 1793 VMRegPair src() const { return _src; } 1794 int src_id() const { return get_id(src()); } 1795 int src_index() const { return _src_index; } 1796 VMRegPair dst() const { return _dst; } 1797 void set_dst(int i, VMRegPair dst) { _dst_index = i, _dst = dst; } 1798 int dst_index() const { return _dst_index; } 1799 int dst_id() const { return get_id(dst()); } 1800 MoveOperation* next() const { return _next; } 1801 MoveOperation* prev() const { return _prev; } 1802 void set_processed() { _processed = true; } 1803 bool is_processed() const { return _processed; } 1804 1805 // insert 1806 void break_cycle(VMRegPair temp_register) { 1807 // create a new store following the last store 1808 // to move from the temp_register to the original 1809 MoveOperation* new_store = new MoveOperation(-1, temp_register, dst_index(), dst()); 1810 1811 // break the cycle of links and insert new_store at the end 1812 // break the reverse link. 1813 MoveOperation* p = prev(); 1814 assert(p->next() == this, "must be"); 1815 _prev = NULL; 1816 p->_next = new_store; 1817 new_store->_prev = p; 1818 1819 // change the original store to save it's value in the temp. 1820 set_dst(-1, temp_register); 1821 } 1822 1823 void link(GrowableArray<MoveOperation*>& killer) { 1824 // link this store in front the store that it depends on 1825 MoveOperation* n = killer.at_grow(src_id(), NULL); 1826 if (n != NULL) { 1827 assert(_next == NULL && n->_prev == NULL, "shouldn't have been set yet"); 1828 _next = n; 1829 n->_prev = this; 1830 } 1831 } 1832 }; 1833 1834 private: 1835 GrowableArray<MoveOperation*> edges; 1836 1837 public: 1838 ComputeMoveOrder(int total_in_args, VMRegPair* in_regs, int total_c_args, VMRegPair* out_regs, 1839 BasicType* in_sig_bt, GrowableArray<int>& arg_order, VMRegPair tmp_vmreg) { 1840 // Move operations where the dest is the stack can all be 1841 // scheduled first since they can't interfere with the other moves. 1842 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 1843 if (in_sig_bt[i] == T_ARRAY) { 1844 c_arg--; 1845 if (out_regs[c_arg].first()->is_stack() && 1846 out_regs[c_arg + 1].first()->is_stack()) { 1847 arg_order.push(i); 1848 arg_order.push(c_arg); 1849 } else { 1850 if (out_regs[c_arg].first()->is_stack() || 1851 in_regs[i].first() == out_regs[c_arg].first()) { 1852 add_edge(i, in_regs[i].first(), c_arg, out_regs[c_arg + 1]); 1853 } else { 1854 add_edge(i, in_regs[i].first(), c_arg, out_regs[c_arg]); 1855 } 1856 } 1857 } else if (in_sig_bt[i] == T_VOID) { 1858 arg_order.push(i); 1859 arg_order.push(c_arg); 1860 } else { 1861 if (out_regs[c_arg].first()->is_stack() || 1862 in_regs[i].first() == out_regs[c_arg].first()) { 1863 arg_order.push(i); 1864 arg_order.push(c_arg); 1865 } else { 1866 add_edge(i, in_regs[i].first(), c_arg, out_regs[c_arg]); 1867 } 1868 } 1869 } 1870 // Break any cycles in the register moves and emit the in the 1871 // proper order. 1872 GrowableArray<MoveOperation*>* stores = get_store_order(tmp_vmreg); 1873 for (int i = 0; i < stores->length(); i++) { 1874 arg_order.push(stores->at(i)->src_index()); 1875 arg_order.push(stores->at(i)->dst_index()); 1876 } 1877 } 1878 1879 // Collected all the move operations 1880 void add_edge(int src_index, VMRegPair src, int dst_index, VMRegPair dst) { 1881 if (src.first() == dst.first()) return; 1882 edges.append(new MoveOperation(src_index, src, dst_index, dst)); 1883 } 1884 1885 // Walk the edges breaking cycles between moves. The result list 1886 // can be walked in order to produce the proper set of loads 1887 GrowableArray<MoveOperation*>* get_store_order(VMRegPair temp_register) { 1888 // Record which moves kill which values 1889 GrowableArray<MoveOperation*> killer; 1890 for (int i = 0; i < edges.length(); i++) { 1891 MoveOperation* s = edges.at(i); 1892 assert(killer.at_grow(s->dst_id(), NULL) == NULL, "only one killer"); 1893 killer.at_put_grow(s->dst_id(), s, NULL); 1894 } 1895 assert(killer.at_grow(MoveOperation::get_id(temp_register), NULL) == NULL, 1896 "make sure temp isn't in the registers that are killed"); 1897 1898 // create links between loads and stores 1899 for (int i = 0; i < edges.length(); i++) { 1900 edges.at(i)->link(killer); 1901 } 1902 1903 // at this point, all the move operations are chained together 1904 // in a doubly linked list. Processing it backwards finds 1905 // the beginning of the chain, forwards finds the end. If there's 1906 // a cycle it can be broken at any point, so pick an edge and walk 1907 // backward until the list ends or we end where we started. 1908 GrowableArray<MoveOperation*>* stores = new GrowableArray<MoveOperation*>(); 1909 for (int e = 0; e < edges.length(); e++) { 1910 MoveOperation* s = edges.at(e); 1911 if (!s->is_processed()) { 1912 MoveOperation* start = s; 1913 // search for the beginning of the chain or cycle 1914 while (start->prev() != NULL && start->prev() != s) { 1915 start = start->prev(); 1916 } 1917 if (start->prev() == s) { 1918 start->break_cycle(temp_register); 1919 } 1920 // walk the chain forward inserting to store list 1921 while (start != NULL) { 1922 stores->append(start); 1923 start->set_processed(); 1924 start = start->next(); 1925 } 1926 } 1927 } 1928 return stores; 1929 } 1930 }; 1931 1932 static void verify_oop_args(MacroAssembler* masm, 1933 const methodHandle& method, 1934 const BasicType* sig_bt, 1935 const VMRegPair* regs) { 1936 Register temp_reg = rbx; // not part of any compiled calling seq 1937 if (VerifyOops) { 1938 for (int i = 0; i < method->size_of_parameters(); i++) { 1939 if (sig_bt[i] == T_OBJECT || 1940 sig_bt[i] == T_ARRAY) { 1941 VMReg r = regs[i].first(); 1942 assert(r->is_valid(), "bad oop arg"); 1943 if (r->is_stack()) { 1944 __ movptr(temp_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1945 __ verify_oop(temp_reg); 1946 } else { 1947 __ verify_oop(r->as_Register()); 1948 } 1949 } 1950 } 1951 } 1952 } 1953 1954 static void gen_special_dispatch(MacroAssembler* masm, 1955 methodHandle method, 1956 const BasicType* sig_bt, 1957 const VMRegPair* regs) { 1958 verify_oop_args(masm, method, sig_bt, regs); 1959 vmIntrinsics::ID iid = method->intrinsic_id(); 1960 1961 // Now write the args into the outgoing interpreter space 1962 bool has_receiver = false; 1963 Register receiver_reg = noreg; 1964 int member_arg_pos = -1; 1965 Register member_reg = noreg; 1966 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid); 1967 if (ref_kind != 0) { 1968 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument 1969 member_reg = rbx; // known to be free at this point 1970 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind); 1971 } else if (iid == vmIntrinsics::_invokeBasic) { 1972 has_receiver = true; 1973 } else { 1974 fatal("unexpected intrinsic id %d", iid); 1975 } 1976 1977 if (member_reg != noreg) { 1978 // Load the member_arg into register, if necessary. 1979 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs); 1980 VMReg r = regs[member_arg_pos].first(); 1981 if (r->is_stack()) { 1982 __ movptr(member_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 1983 } else { 1984 // no data motion is needed 1985 member_reg = r->as_Register(); 1986 } 1987 } 1988 1989 if (has_receiver) { 1990 // Make sure the receiver is loaded into a register. 1991 assert(method->size_of_parameters() > 0, "oob"); 1992 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object"); 1993 VMReg r = regs[0].first(); 1994 assert(r->is_valid(), "bad receiver arg"); 1995 if (r->is_stack()) { 1996 // Porting note: This assumes that compiled calling conventions always 1997 // pass the receiver oop in a register. If this is not true on some 1998 // platform, pick a temp and load the receiver from stack. 1999 fatal("receiver always in a register"); 2000 receiver_reg = j_rarg0; // known to be free at this point 2001 __ movptr(receiver_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize)); 2002 } else { 2003 // no data motion is needed 2004 receiver_reg = r->as_Register(); 2005 } 2006 } 2007 2008 // Figure out which address we are really jumping to: 2009 MethodHandles::generate_method_handle_dispatch(masm, iid, 2010 receiver_reg, member_reg, /*for_compiler_entry:*/ true); 2011 } 2012 2013 // --------------------------------------------------------------------------- 2014 // Generate a native wrapper for a given method. The method takes arguments 2015 // in the Java compiled code convention, marshals them to the native 2016 // convention (handlizes oops, etc), transitions to native, makes the call, 2017 // returns to java state (possibly blocking), unhandlizes any result and 2018 // returns. 2019 // 2020 // Critical native functions are a shorthand for the use of 2021 // GetPrimtiveArrayCritical and disallow the use of any other JNI 2022 // functions. The wrapper is expected to unpack the arguments before 2023 // passing them to the callee and perform checks before and after the 2024 // native call to ensure that they GCLocker 2025 // lock_critical/unlock_critical semantics are followed. Some other 2026 // parts of JNI setup are skipped like the tear down of the JNI handle 2027 // block and the check for pending exceptions it's impossible for them 2028 // to be thrown. 2029 // 2030 // They are roughly structured like this: 2031 // if (GCLocker::needs_gc()) 2032 // SharedRuntime::block_for_jni_critical(); 2033 // tranistion to thread_in_native 2034 // unpack arrray arguments and call native entry point 2035 // check for safepoint in progress 2036 // check if any thread suspend flags are set 2037 // call into JVM and possible unlock the JNI critical 2038 // if a GC was suppressed while in the critical native. 2039 // transition back to thread_in_Java 2040 // return to caller 2041 // 2042 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 2043 const methodHandle& method, 2044 int compile_id, 2045 BasicType* in_sig_bt, 2046 VMRegPair* in_regs, 2047 BasicType ret_type) { 2048 if (method->is_method_handle_intrinsic()) { 2049 vmIntrinsics::ID iid = method->intrinsic_id(); 2050 intptr_t start = (intptr_t)__ pc(); 2051 int vep_offset = ((intptr_t)__ pc()) - start; 2052 gen_special_dispatch(masm, 2053 method, 2054 in_sig_bt, 2055 in_regs); 2056 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period 2057 __ flush(); 2058 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually 2059 return nmethod::new_native_nmethod(method, 2060 compile_id, 2061 masm->code(), 2062 vep_offset, 2063 frame_complete, 2064 stack_slots / VMRegImpl::slots_per_word, 2065 in_ByteSize(-1), 2066 in_ByteSize(-1), 2067 (OopMapSet*)NULL); 2068 } 2069 bool is_critical_native = true; 2070 address native_func = method->critical_native_function(); 2071 if (native_func == NULL) { 2072 native_func = method->native_function(); 2073 is_critical_native = false; 2074 } 2075 assert(native_func != NULL, "must have function"); 2076 2077 // An OopMap for lock (and class if static) 2078 OopMapSet *oop_maps = new OopMapSet(); 2079 intptr_t start = (intptr_t)__ pc(); 2080 2081 // We have received a description of where all the java arg are located 2082 // on entry to the wrapper. We need to convert these args to where 2083 // the jni function will expect them. To figure out where they go 2084 // we convert the java signature to a C signature by inserting 2085 // the hidden arguments as arg[0] and possibly arg[1] (static method) 2086 2087 const int total_in_args = method->size_of_parameters(); 2088 int total_c_args = total_in_args; 2089 if (!is_critical_native) { 2090 total_c_args += 1; 2091 if (method->is_static()) { 2092 total_c_args++; 2093 } 2094 } else { 2095 for (int i = 0; i < total_in_args; i++) { 2096 if (in_sig_bt[i] == T_ARRAY) { 2097 total_c_args++; 2098 } 2099 } 2100 } 2101 2102 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 2103 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 2104 BasicType* in_elem_bt = NULL; 2105 2106 int argc = 0; 2107 if (!is_critical_native) { 2108 out_sig_bt[argc++] = T_ADDRESS; 2109 if (method->is_static()) { 2110 out_sig_bt[argc++] = T_OBJECT; 2111 } 2112 2113 for (int i = 0; i < total_in_args ; i++ ) { 2114 out_sig_bt[argc++] = in_sig_bt[i]; 2115 } 2116 } else { 2117 Thread* THREAD = Thread::current(); 2118 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args); 2119 SignatureStream ss(method->signature()); 2120 for (int i = 0; i < total_in_args ; i++ ) { 2121 if (in_sig_bt[i] == T_ARRAY) { 2122 // Arrays are passed as int, elem* pair 2123 out_sig_bt[argc++] = T_INT; 2124 out_sig_bt[argc++] = T_ADDRESS; 2125 Symbol* atype = ss.as_symbol(CHECK_NULL); 2126 const char* at = atype->as_C_string(); 2127 if (strlen(at) == 2) { 2128 assert(at[0] == '[', "must be"); 2129 switch (at[1]) { 2130 case 'B': in_elem_bt[i] = T_BYTE; break; 2131 case 'C': in_elem_bt[i] = T_CHAR; break; 2132 case 'D': in_elem_bt[i] = T_DOUBLE; break; 2133 case 'F': in_elem_bt[i] = T_FLOAT; break; 2134 case 'I': in_elem_bt[i] = T_INT; break; 2135 case 'J': in_elem_bt[i] = T_LONG; break; 2136 case 'S': in_elem_bt[i] = T_SHORT; break; 2137 case 'Z': in_elem_bt[i] = T_BOOLEAN; break; 2138 default: ShouldNotReachHere(); 2139 } 2140 } 2141 } else { 2142 out_sig_bt[argc++] = in_sig_bt[i]; 2143 in_elem_bt[i] = T_VOID; 2144 } 2145 if (in_sig_bt[i] != T_VOID) { 2146 assert(in_sig_bt[i] == ss.type(), "must match"); 2147 ss.next(); 2148 } 2149 } 2150 } 2151 2152 // Now figure out where the args must be stored and how much stack space 2153 // they require. 2154 int out_arg_slots; 2155 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args); 2156 2157 // Compute framesize for the wrapper. We need to handlize all oops in 2158 // incoming registers 2159 2160 // Calculate the total number of stack slots we will need. 2161 2162 // First count the abi requirement plus all of the outgoing args 2163 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 2164 2165 // Now the space for the inbound oop handle area 2166 int total_save_slots = 6 * VMRegImpl::slots_per_word; // 6 arguments passed in registers 2167 if (is_critical_native) { 2168 // Critical natives may have to call out so they need a save area 2169 // for register arguments. 2170 int double_slots = 0; 2171 int single_slots = 0; 2172 for ( int i = 0; i < total_in_args; i++) { 2173 if (in_regs[i].first()->is_Register()) { 2174 const Register reg = in_regs[i].first()->as_Register(); 2175 switch (in_sig_bt[i]) { 2176 case T_BOOLEAN: 2177 case T_BYTE: 2178 case T_SHORT: 2179 case T_CHAR: 2180 case T_INT: single_slots++; break; 2181 case T_ARRAY: // specific to LP64 (7145024) 2182 case T_LONG: double_slots++; break; 2183 default: ShouldNotReachHere(); 2184 } 2185 } else if (in_regs[i].first()->is_XMMRegister()) { 2186 switch (in_sig_bt[i]) { 2187 case T_FLOAT: single_slots++; break; 2188 case T_DOUBLE: double_slots++; break; 2189 default: ShouldNotReachHere(); 2190 } 2191 } else if (in_regs[i].first()->is_FloatRegister()) { 2192 ShouldNotReachHere(); 2193 } 2194 } 2195 total_save_slots = double_slots * 2 + single_slots; 2196 // align the save area 2197 if (double_slots != 0) { 2198 stack_slots = round_to(stack_slots, 2); 2199 } 2200 } 2201 2202 int oop_handle_offset = stack_slots; 2203 stack_slots += total_save_slots; 2204 2205 // Now any space we need for handlizing a klass if static method 2206 2207 int klass_slot_offset = 0; 2208 int klass_offset = -1; 2209 int lock_slot_offset = 0; 2210 bool is_static = false; 2211 2212 if (method->is_static()) { 2213 klass_slot_offset = stack_slots; 2214 stack_slots += VMRegImpl::slots_per_word; 2215 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 2216 is_static = true; 2217 } 2218 2219 // Plus a lock if needed 2220 2221 if (method->is_synchronized()) { 2222 lock_slot_offset = stack_slots; 2223 stack_slots += VMRegImpl::slots_per_word; 2224 } 2225 2226 // Now a place (+2) to save return values or temp during shuffling 2227 // + 4 for return address (which we own) and saved rbp 2228 stack_slots += 6; 2229 2230 // Ok The space we have allocated will look like: 2231 // 2232 // 2233 // FP-> | | 2234 // |---------------------| 2235 // | 2 slots for moves | 2236 // |---------------------| 2237 // | lock box (if sync) | 2238 // |---------------------| <- lock_slot_offset 2239 // | klass (if static) | 2240 // |---------------------| <- klass_slot_offset 2241 // | oopHandle area | 2242 // |---------------------| <- oop_handle_offset (6 java arg registers) 2243 // | outbound memory | 2244 // | based arguments | 2245 // | | 2246 // |---------------------| 2247 // | | 2248 // SP-> | out_preserved_slots | 2249 // 2250 // 2251 2252 2253 // Now compute actual number of stack words we need rounding to make 2254 // stack properly aligned. 2255 stack_slots = round_to(stack_slots, StackAlignmentInSlots); 2256 2257 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 2258 2259 // First thing make an ic check to see if we should even be here 2260 2261 // We are free to use all registers as temps without saving them and 2262 // restoring them except rbp. rbp is the only callee save register 2263 // as far as the interpreter and the compiler(s) are concerned. 2264 2265 2266 const Register ic_reg = rax; 2267 const Register receiver = j_rarg0; 2268 2269 Label hit; 2270 Label exception_pending; 2271 2272 assert_different_registers(ic_reg, receiver, rscratch1); 2273 __ verify_oop(receiver); 2274 __ load_klass(rscratch1, receiver); 2275 __ cmpq(ic_reg, rscratch1); 2276 __ jcc(Assembler::equal, hit); 2277 2278 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub())); 2279 2280 // Verified entry point must be aligned 2281 __ align(8); 2282 2283 __ bind(hit); 2284 2285 int vep_offset = ((intptr_t)__ pc()) - start; 2286 2287 #ifdef COMPILER1 2288 // For Object.hashCode, System.identityHashCode try to pull hashCode from object header if available. 2289 if ((InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) || (method->intrinsic_id() == vmIntrinsics::_identityHashCode)) { 2290 inline_check_hashcode_from_object_header(masm, method, j_rarg0 /*obj_reg*/, rax /*result*/); 2291 } 2292 #endif // COMPILER1 2293 2294 // The instruction at the verified entry point must be 5 bytes or longer 2295 // because it can be patched on the fly by make_non_entrant. The stack bang 2296 // instruction fits that requirement. 2297 2298 // Generate stack overflow check 2299 2300 if (UseStackBanging) { 2301 __ bang_stack_with_offset((int)JavaThread::stack_shadow_zone_size()); 2302 } else { 2303 // need a 5 byte instruction to allow MT safe patching to non-entrant 2304 __ fat_nop(); 2305 } 2306 2307 // Generate a new frame for the wrapper. 2308 __ enter(); 2309 // -2 because return address is already present and so is saved rbp 2310 __ subptr(rsp, stack_size - 2*wordSize); 2311 2312 // Frame is now completed as far as size and linkage. 2313 int frame_complete = ((intptr_t)__ pc()) - start; 2314 2315 if (UseRTMLocking) { 2316 // Abort RTM transaction before calling JNI 2317 // because critical section will be large and will be 2318 // aborted anyway. Also nmethod could be deoptimized. 2319 __ xabort(0); 2320 } 2321 2322 #ifdef ASSERT 2323 { 2324 Label L; 2325 __ mov(rax, rsp); 2326 __ andptr(rax, -16); // must be 16 byte boundary (see amd64 ABI) 2327 __ cmpptr(rax, rsp); 2328 __ jcc(Assembler::equal, L); 2329 __ stop("improperly aligned stack"); 2330 __ bind(L); 2331 } 2332 #endif /* ASSERT */ 2333 2334 2335 // We use r14 as the oop handle for the receiver/klass 2336 // It is callee save so it survives the call to native 2337 2338 const Register oop_handle_reg = r14; 2339 2340 if (is_critical_native) { 2341 check_needs_gc_for_critical_native(masm, stack_slots, total_c_args, total_in_args, 2342 oop_handle_offset, oop_maps, in_regs, in_sig_bt); 2343 } 2344 2345 // 2346 // We immediately shuffle the arguments so that any vm call we have to 2347 // make from here on out (sync slow path, jvmti, etc.) we will have 2348 // captured the oops from our caller and have a valid oopMap for 2349 // them. 2350 2351 // ----------------- 2352 // The Grand Shuffle 2353 2354 // The Java calling convention is either equal (linux) or denser (win64) than the 2355 // c calling convention. However the because of the jni_env argument the c calling 2356 // convention always has at least one more (and two for static) arguments than Java. 2357 // Therefore if we move the args from java -> c backwards then we will never have 2358 // a register->register conflict and we don't have to build a dependency graph 2359 // and figure out how to break any cycles. 2360 // 2361 2362 // Record esp-based slot for receiver on stack for non-static methods 2363 int receiver_offset = -1; 2364 2365 // This is a trick. We double the stack slots so we can claim 2366 // the oops in the caller's frame. Since we are sure to have 2367 // more args than the caller doubling is enough to make 2368 // sure we can capture all the incoming oop args from the 2369 // caller. 2370 // 2371 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 2372 2373 // Mark location of rbp (someday) 2374 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rbp)); 2375 2376 // Use eax, ebx as temporaries during any memory-memory moves we have to do 2377 // All inbound args are referenced based on rbp and all outbound args via rsp. 2378 2379 2380 #ifdef ASSERT 2381 bool reg_destroyed[RegisterImpl::number_of_registers]; 2382 bool freg_destroyed[XMMRegisterImpl::number_of_registers]; 2383 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) { 2384 reg_destroyed[r] = false; 2385 } 2386 for ( int f = 0 ; f < XMMRegisterImpl::number_of_registers ; f++ ) { 2387 freg_destroyed[f] = false; 2388 } 2389 2390 #endif /* ASSERT */ 2391 2392 // This may iterate in two different directions depending on the 2393 // kind of native it is. The reason is that for regular JNI natives 2394 // the incoming and outgoing registers are offset upwards and for 2395 // critical natives they are offset down. 2396 GrowableArray<int> arg_order(2 * total_in_args); 2397 VMRegPair tmp_vmreg; 2398 tmp_vmreg.set1(rbx->as_VMReg()); 2399 2400 if (!is_critical_native) { 2401 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) { 2402 arg_order.push(i); 2403 arg_order.push(c_arg); 2404 } 2405 } else { 2406 // Compute a valid move order, using tmp_vmreg to break any cycles 2407 ComputeMoveOrder cmo(total_in_args, in_regs, total_c_args, out_regs, in_sig_bt, arg_order, tmp_vmreg); 2408 } 2409 2410 int temploc = -1; 2411 for (int ai = 0; ai < arg_order.length(); ai += 2) { 2412 int i = arg_order.at(ai); 2413 int c_arg = arg_order.at(ai + 1); 2414 __ block_comment(err_msg("move %d -> %d", i, c_arg)); 2415 if (c_arg == -1) { 2416 assert(is_critical_native, "should only be required for critical natives"); 2417 // This arg needs to be moved to a temporary 2418 __ mov(tmp_vmreg.first()->as_Register(), in_regs[i].first()->as_Register()); 2419 in_regs[i] = tmp_vmreg; 2420 temploc = i; 2421 continue; 2422 } else if (i == -1) { 2423 assert(is_critical_native, "should only be required for critical natives"); 2424 // Read from the temporary location 2425 assert(temploc != -1, "must be valid"); 2426 i = temploc; 2427 temploc = -1; 2428 } 2429 #ifdef ASSERT 2430 if (in_regs[i].first()->is_Register()) { 2431 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!"); 2432 } else if (in_regs[i].first()->is_XMMRegister()) { 2433 assert(!freg_destroyed[in_regs[i].first()->as_XMMRegister()->encoding()], "destroyed reg!"); 2434 } 2435 if (out_regs[c_arg].first()->is_Register()) { 2436 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 2437 } else if (out_regs[c_arg].first()->is_XMMRegister()) { 2438 freg_destroyed[out_regs[c_arg].first()->as_XMMRegister()->encoding()] = true; 2439 } 2440 #endif /* ASSERT */ 2441 switch (in_sig_bt[i]) { 2442 case T_ARRAY: 2443 if (is_critical_native) { 2444 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]); 2445 c_arg++; 2446 #ifdef ASSERT 2447 if (out_regs[c_arg].first()->is_Register()) { 2448 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 2449 } else if (out_regs[c_arg].first()->is_XMMRegister()) { 2450 freg_destroyed[out_regs[c_arg].first()->as_XMMRegister()->encoding()] = true; 2451 } 2452 #endif 2453 break; 2454 } 2455 case T_OBJECT: 2456 assert(!is_critical_native, "no oop arguments"); 2457 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 2458 ((i == 0) && (!is_static)), 2459 &receiver_offset); 2460 break; 2461 case T_VOID: 2462 break; 2463 2464 case T_FLOAT: 2465 float_move(masm, in_regs[i], out_regs[c_arg]); 2466 break; 2467 2468 case T_DOUBLE: 2469 assert( i + 1 < total_in_args && 2470 in_sig_bt[i + 1] == T_VOID && 2471 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 2472 double_move(masm, in_regs[i], out_regs[c_arg]); 2473 break; 2474 2475 case T_LONG : 2476 long_move(masm, in_regs[i], out_regs[c_arg]); 2477 break; 2478 2479 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 2480 2481 default: 2482 move32_64(masm, in_regs[i], out_regs[c_arg]); 2483 } 2484 } 2485 2486 int c_arg; 2487 2488 // Pre-load a static method's oop into r14. Used both by locking code and 2489 // the normal JNI call code. 2490 if (!is_critical_native) { 2491 // point c_arg at the first arg that is already loaded in case we 2492 // need to spill before we call out 2493 c_arg = total_c_args - total_in_args; 2494 2495 if (method->is_static()) { 2496 2497 // load oop into a register 2498 __ movoop(oop_handle_reg, JNIHandles::make_local(method->method_holder()->java_mirror())); 2499 2500 // Now handlize the static class mirror it's known not-null. 2501 __ movptr(Address(rsp, klass_offset), oop_handle_reg); 2502 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 2503 2504 // Now get the handle 2505 __ lea(oop_handle_reg, Address(rsp, klass_offset)); 2506 // store the klass handle as second argument 2507 __ movptr(c_rarg1, oop_handle_reg); 2508 // and protect the arg if we must spill 2509 c_arg--; 2510 } 2511 } else { 2512 // For JNI critical methods we need to save all registers in save_args. 2513 c_arg = 0; 2514 } 2515 2516 // Change state to native (we save the return address in the thread, since it might not 2517 // be pushed on the stack when we do a a stack traversal). It is enough that the pc() 2518 // points into the right code segment. It does not have to be the correct return pc. 2519 // We use the same pc/oopMap repeatedly when we call out 2520 2521 intptr_t the_pc = (intptr_t) __ pc(); 2522 oop_maps->add_gc_map(the_pc - start, map); 2523 2524 __ set_last_Java_frame(rsp, noreg, (address)the_pc); 2525 2526 2527 // We have all of the arguments setup at this point. We must not touch any register 2528 // argument registers at this point (what if we save/restore them there are no oop? 2529 2530 { 2531 SkipIfEqual skip(masm, &DTraceMethodProbes, false); 2532 // protect the args we've loaded 2533 save_args(masm, total_c_args, c_arg, out_regs); 2534 __ mov_metadata(c_rarg1, method()); 2535 __ call_VM_leaf( 2536 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 2537 r15_thread, c_rarg1); 2538 restore_args(masm, total_c_args, c_arg, out_regs); 2539 } 2540 2541 // RedefineClasses() tracing support for obsolete method entry 2542 if (log_is_enabled(Trace, redefine, class, obsolete)) { 2543 // protect the args we've loaded 2544 save_args(masm, total_c_args, c_arg, out_regs); 2545 __ mov_metadata(c_rarg1, method()); 2546 __ call_VM_leaf( 2547 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 2548 r15_thread, c_rarg1); 2549 restore_args(masm, total_c_args, c_arg, out_regs); 2550 } 2551 2552 // Lock a synchronized method 2553 2554 // Register definitions used by locking and unlocking 2555 2556 const Register swap_reg = rax; // Must use rax for cmpxchg instruction 2557 const Register obj_reg = rbx; // Will contain the oop 2558 const Register lock_reg = r13; // Address of compiler lock object (BasicLock) 2559 const Register old_hdr = r13; // value of old header at unlock time 2560 2561 Label slow_path_lock; 2562 Label lock_done; 2563 2564 if (method->is_synchronized()) { 2565 assert(!is_critical_native, "unhandled"); 2566 2567 2568 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes(); 2569 2570 // Get the handle (the 2nd argument) 2571 __ mov(oop_handle_reg, c_rarg1); 2572 2573 // Get address of the box 2574 2575 __ lea(lock_reg, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2576 2577 // Load the oop from the handle 2578 __ movptr(obj_reg, Address(oop_handle_reg, 0)); 2579 2580 if (UseBiasedLocking) { 2581 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, rscratch1, false, lock_done, &slow_path_lock); 2582 } 2583 2584 // Load immediate 1 into swap_reg %rax 2585 __ movl(swap_reg, 1); 2586 2587 // Load (object->mark() | 1) into swap_reg %rax 2588 __ orptr(swap_reg, Address(obj_reg, 0)); 2589 2590 // Save (object->mark() | 1) into BasicLock's displaced header 2591 __ movptr(Address(lock_reg, mark_word_offset), swap_reg); 2592 2593 if (os::is_MP()) { 2594 __ lock(); 2595 } 2596 2597 // src -> dest iff dest == rax else rax <- dest 2598 __ cmpxchgptr(lock_reg, Address(obj_reg, 0)); 2599 __ jcc(Assembler::equal, lock_done); 2600 2601 // Hmm should this move to the slow path code area??? 2602 2603 // Test if the oopMark is an obvious stack pointer, i.e., 2604 // 1) (mark & 3) == 0, and 2605 // 2) rsp <= mark < mark + os::pagesize() 2606 // These 3 tests can be done by evaluating the following 2607 // expression: ((mark - rsp) & (3 - os::vm_page_size())), 2608 // assuming both stack pointer and pagesize have their 2609 // least significant 2 bits clear. 2610 // NOTE: the oopMark is in swap_reg %rax as the result of cmpxchg 2611 2612 __ subptr(swap_reg, rsp); 2613 __ andptr(swap_reg, 3 - os::vm_page_size()); 2614 2615 // Save the test result, for recursive case, the result is zero 2616 __ movptr(Address(lock_reg, mark_word_offset), swap_reg); 2617 __ jcc(Assembler::notEqual, slow_path_lock); 2618 2619 // Slow path will re-enter here 2620 2621 __ bind(lock_done); 2622 } 2623 2624 2625 // Finally just about ready to make the JNI call 2626 2627 2628 // get JNIEnv* which is first argument to native 2629 if (!is_critical_native) { 2630 __ lea(c_rarg0, Address(r15_thread, in_bytes(JavaThread::jni_environment_offset()))); 2631 } 2632 2633 // Now set thread in native 2634 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_native); 2635 2636 __ call(RuntimeAddress(native_func)); 2637 2638 // Verify or restore cpu control state after JNI call 2639 __ restore_cpu_control_state_after_jni(); 2640 2641 // Unpack native results. 2642 switch (ret_type) { 2643 case T_BOOLEAN: __ c2bool(rax); break; 2644 case T_CHAR : __ movzwl(rax, rax); break; 2645 case T_BYTE : __ sign_extend_byte (rax); break; 2646 case T_SHORT : __ sign_extend_short(rax); break; 2647 case T_INT : /* nothing to do */ break; 2648 case T_DOUBLE : 2649 case T_FLOAT : 2650 // Result is in xmm0 we'll save as needed 2651 break; 2652 case T_ARRAY: // Really a handle 2653 case T_OBJECT: // Really a handle 2654 break; // can't de-handlize until after safepoint check 2655 case T_VOID: break; 2656 case T_LONG: break; 2657 default : ShouldNotReachHere(); 2658 } 2659 2660 // Switch thread to "native transition" state before reading the synchronization state. 2661 // This additional state is necessary because reading and testing the synchronization 2662 // state is not atomic w.r.t. GC, as this scenario demonstrates: 2663 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 2664 // VM thread changes sync state to synchronizing and suspends threads for GC. 2665 // Thread A is resumed to finish this native method, but doesn't block here since it 2666 // didn't see any synchronization is progress, and escapes. 2667 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_native_trans); 2668 2669 if(os::is_MP()) { 2670 if (UseMembar) { 2671 // Force this write out before the read below 2672 __ membar(Assembler::Membar_mask_bits( 2673 Assembler::LoadLoad | Assembler::LoadStore | 2674 Assembler::StoreLoad | Assembler::StoreStore)); 2675 } else { 2676 // Write serialization page so VM thread can do a pseudo remote membar. 2677 // We use the current thread pointer to calculate a thread specific 2678 // offset to write to within the page. This minimizes bus traffic 2679 // due to cache line collision. 2680 __ serialize_memory(r15_thread, rcx); 2681 } 2682 } 2683 2684 Label after_transition; 2685 2686 // check for safepoint operation in progress and/or pending suspend requests 2687 { 2688 Label Continue; 2689 2690 __ cmp32(ExternalAddress((address)SafepointSynchronize::address_of_state()), 2691 SafepointSynchronize::_not_synchronized); 2692 2693 Label L; 2694 __ jcc(Assembler::notEqual, L); 2695 __ cmpl(Address(r15_thread, JavaThread::suspend_flags_offset()), 0); 2696 __ jcc(Assembler::equal, Continue); 2697 __ bind(L); 2698 2699 // Don't use call_VM as it will see a possible pending exception and forward it 2700 // and never return here preventing us from clearing _last_native_pc down below. 2701 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are 2702 // preserved and correspond to the bcp/locals pointers. So we do a runtime call 2703 // by hand. 2704 // 2705 save_native_result(masm, ret_type, stack_slots); 2706 __ mov(c_rarg0, r15_thread); 2707 __ mov(r12, rsp); // remember sp 2708 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 2709 __ andptr(rsp, -16); // align stack as required by ABI 2710 if (!is_critical_native) { 2711 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans))); 2712 } else { 2713 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition))); 2714 } 2715 __ mov(rsp, r12); // restore sp 2716 __ reinit_heapbase(); 2717 // Restore any method result value 2718 restore_native_result(masm, ret_type, stack_slots); 2719 2720 if (is_critical_native) { 2721 // The call above performed the transition to thread_in_Java so 2722 // skip the transition logic below. 2723 __ jmpb(after_transition); 2724 } 2725 2726 __ bind(Continue); 2727 } 2728 2729 // change thread state 2730 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_Java); 2731 __ bind(after_transition); 2732 2733 Label reguard; 2734 Label reguard_done; 2735 __ cmpl(Address(r15_thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_reserved_disabled); 2736 __ jcc(Assembler::equal, reguard); 2737 __ bind(reguard_done); 2738 2739 // native result if any is live 2740 2741 // Unlock 2742 Label unlock_done; 2743 Label slow_path_unlock; 2744 if (method->is_synchronized()) { 2745 2746 // Get locked oop from the handle we passed to jni 2747 __ movptr(obj_reg, Address(oop_handle_reg, 0)); 2748 2749 Label done; 2750 2751 if (UseBiasedLocking) { 2752 __ biased_locking_exit(obj_reg, old_hdr, done); 2753 } 2754 2755 // Simple recursive lock? 2756 2757 __ cmpptr(Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size), (int32_t)NULL_WORD); 2758 __ jcc(Assembler::equal, done); 2759 2760 // Must save rax if if it is live now because cmpxchg must use it 2761 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2762 save_native_result(masm, ret_type, stack_slots); 2763 } 2764 2765 2766 // get address of the stack lock 2767 __ lea(rax, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2768 // get old displaced header 2769 __ movptr(old_hdr, Address(rax, 0)); 2770 2771 // Atomic swap old header if oop still contains the stack lock 2772 if (os::is_MP()) { 2773 __ lock(); 2774 } 2775 __ cmpxchgptr(old_hdr, Address(obj_reg, 0)); 2776 __ jcc(Assembler::notEqual, slow_path_unlock); 2777 2778 // slow path re-enters here 2779 __ bind(unlock_done); 2780 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) { 2781 restore_native_result(masm, ret_type, stack_slots); 2782 } 2783 2784 __ bind(done); 2785 2786 } 2787 { 2788 SkipIfEqual skip(masm, &DTraceMethodProbes, false); 2789 save_native_result(masm, ret_type, stack_slots); 2790 __ mov_metadata(c_rarg1, method()); 2791 __ call_VM_leaf( 2792 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2793 r15_thread, c_rarg1); 2794 restore_native_result(masm, ret_type, stack_slots); 2795 } 2796 2797 __ reset_last_Java_frame(false); 2798 2799 // Unpack oop result 2800 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2801 Label L; 2802 __ testptr(rax, rax); 2803 __ jcc(Assembler::zero, L); 2804 __ movptr(rax, Address(rax, 0)); 2805 __ bind(L); 2806 __ verify_oop(rax); 2807 } 2808 2809 if (CheckJNICalls) { 2810 // clear_pending_jni_exception_check 2811 __ movptr(Address(r15_thread, JavaThread::pending_jni_exception_check_fn_offset()), NULL_WORD); 2812 } 2813 2814 if (!is_critical_native) { 2815 // reset handle block 2816 __ movptr(rcx, Address(r15_thread, JavaThread::active_handles_offset())); 2817 __ movl(Address(rcx, JNIHandleBlock::top_offset_in_bytes()), (int32_t)NULL_WORD); 2818 } 2819 2820 // pop our frame 2821 2822 __ leave(); 2823 2824 if (!is_critical_native) { 2825 // Any exception pending? 2826 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2827 __ jcc(Assembler::notEqual, exception_pending); 2828 } 2829 2830 // Return 2831 2832 __ ret(0); 2833 2834 // Unexpected paths are out of line and go here 2835 2836 if (!is_critical_native) { 2837 // forward the exception 2838 __ bind(exception_pending); 2839 2840 // and forward the exception 2841 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 2842 } 2843 2844 // Slow path locking & unlocking 2845 if (method->is_synchronized()) { 2846 2847 // BEGIN Slow path lock 2848 __ bind(slow_path_lock); 2849 2850 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM 2851 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2852 2853 // protect the args we've loaded 2854 save_args(masm, total_c_args, c_arg, out_regs); 2855 2856 __ mov(c_rarg0, obj_reg); 2857 __ mov(c_rarg1, lock_reg); 2858 __ mov(c_rarg2, r15_thread); 2859 2860 // Not a leaf but we have last_Java_frame setup as we want 2861 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3); 2862 restore_args(masm, total_c_args, c_arg, out_regs); 2863 2864 #ifdef ASSERT 2865 { Label L; 2866 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2867 __ jcc(Assembler::equal, L); 2868 __ stop("no pending exception allowed on exit from monitorenter"); 2869 __ bind(L); 2870 } 2871 #endif 2872 __ jmp(lock_done); 2873 2874 // END Slow path lock 2875 2876 // BEGIN Slow path unlock 2877 __ bind(slow_path_unlock); 2878 2879 // If we haven't already saved the native result we must save it now as xmm registers 2880 // are still exposed. 2881 2882 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2883 save_native_result(masm, ret_type, stack_slots); 2884 } 2885 2886 __ lea(c_rarg1, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size)); 2887 2888 __ mov(c_rarg0, obj_reg); 2889 __ mov(c_rarg2, r15_thread); 2890 __ mov(r12, rsp); // remember sp 2891 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 2892 __ andptr(rsp, -16); // align stack as required by ABI 2893 2894 // Save pending exception around call to VM (which contains an EXCEPTION_MARK) 2895 // NOTE that obj_reg == rbx currently 2896 __ movptr(rbx, Address(r15_thread, in_bytes(Thread::pending_exception_offset()))); 2897 __ movptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD); 2898 2899 // args are (oop obj, BasicLock* lock, JavaThread* thread) 2900 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C))); 2901 __ mov(rsp, r12); // restore sp 2902 __ reinit_heapbase(); 2903 #ifdef ASSERT 2904 { 2905 Label L; 2906 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), (int)NULL_WORD); 2907 __ jcc(Assembler::equal, L); 2908 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C"); 2909 __ bind(L); 2910 } 2911 #endif /* ASSERT */ 2912 2913 __ movptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), rbx); 2914 2915 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) { 2916 restore_native_result(masm, ret_type, stack_slots); 2917 } 2918 __ jmp(unlock_done); 2919 2920 // END Slow path unlock 2921 2922 } // synchronized 2923 2924 // SLOW PATH Reguard the stack if needed 2925 2926 __ bind(reguard); 2927 save_native_result(masm, ret_type, stack_slots); 2928 __ mov(r12, rsp); // remember sp 2929 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows 2930 __ andptr(rsp, -16); // align stack as required by ABI 2931 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages))); 2932 __ mov(rsp, r12); // restore sp 2933 __ reinit_heapbase(); 2934 restore_native_result(masm, ret_type, stack_slots); 2935 // and continue 2936 __ jmp(reguard_done); 2937 2938 2939 2940 __ flush(); 2941 2942 nmethod *nm = nmethod::new_native_nmethod(method, 2943 compile_id, 2944 masm->code(), 2945 vep_offset, 2946 frame_complete, 2947 stack_slots / VMRegImpl::slots_per_word, 2948 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2949 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size), 2950 oop_maps); 2951 2952 if (is_critical_native) { 2953 nm->set_lazy_critical_native(true); 2954 } 2955 2956 return nm; 2957 2958 } 2959 2960 // this function returns the adjust size (in number of words) to a c2i adapter 2961 // activation for use during deoptimization 2962 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals ) { 2963 return (callee_locals - callee_parameters) * Interpreter::stackElementWords; 2964 } 2965 2966 2967 uint SharedRuntime::out_preserve_stack_slots() { 2968 return 0; 2969 } 2970 2971 //------------------------------generate_deopt_blob---------------------------- 2972 void SharedRuntime::generate_deopt_blob() { 2973 // Allocate space for the code 2974 ResourceMark rm; 2975 // Setup code generation tools 2976 int pad = 0; 2977 #if INCLUDE_JVMCI 2978 if (EnableJVMCI || UseAOT) { 2979 pad += 512; // Increase the buffer size when compiling for JVMCI 2980 } 2981 #endif 2982 CodeBuffer buffer("deopt_blob", 2048+pad, 1024); 2983 MacroAssembler* masm = new MacroAssembler(&buffer); 2984 int frame_size_in_words; 2985 OopMap* map = NULL; 2986 OopMapSet *oop_maps = new OopMapSet(); 2987 2988 // ------------- 2989 // This code enters when returning to a de-optimized nmethod. A return 2990 // address has been pushed on the the stack, and return values are in 2991 // registers. 2992 // If we are doing a normal deopt then we were called from the patched 2993 // nmethod from the point we returned to the nmethod. So the return 2994 // address on the stack is wrong by NativeCall::instruction_size 2995 // We will adjust the value so it looks like we have the original return 2996 // address on the stack (like when we eagerly deoptimized). 2997 // In the case of an exception pending when deoptimizing, we enter 2998 // with a return address on the stack that points after the call we patched 2999 // into the exception handler. We have the following register state from, 3000 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp). 3001 // rax: exception oop 3002 // rbx: exception handler 3003 // rdx: throwing pc 3004 // So in this case we simply jam rdx into the useless return address and 3005 // the stack looks just like we want. 3006 // 3007 // At this point we need to de-opt. We save the argument return 3008 // registers. We call the first C routine, fetch_unroll_info(). This 3009 // routine captures the return values and returns a structure which 3010 // describes the current frame size and the sizes of all replacement frames. 3011 // The current frame is compiled code and may contain many inlined 3012 // functions, each with their own JVM state. We pop the current frame, then 3013 // push all the new frames. Then we call the C routine unpack_frames() to 3014 // populate these frames. Finally unpack_frames() returns us the new target 3015 // address. Notice that callee-save registers are BLOWN here; they have 3016 // already been captured in the vframeArray at the time the return PC was 3017 // patched. 3018 address start = __ pc(); 3019 Label cont; 3020 3021 // Prolog for non exception case! 3022 3023 // Save everything in sight. 3024 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 3025 3026 // Normal deoptimization. Save exec mode for unpack_frames. 3027 __ movl(r14, Deoptimization::Unpack_deopt); // callee-saved 3028 __ jmp(cont); 3029 3030 int reexecute_offset = __ pc() - start; 3031 #if INCLUDE_JVMCI && !defined(COMPILER1) 3032 if (EnableJVMCI && UseJVMCICompiler) { 3033 // JVMCI does not use this kind of deoptimization 3034 __ should_not_reach_here(); 3035 } 3036 #endif 3037 3038 // Reexecute case 3039 // return address is the pc describes what bci to do re-execute at 3040 3041 // No need to update map as each call to save_live_registers will produce identical oopmap 3042 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 3043 3044 __ movl(r14, Deoptimization::Unpack_reexecute); // callee-saved 3045 __ jmp(cont); 3046 3047 #if INCLUDE_JVMCI 3048 Label after_fetch_unroll_info_call; 3049 int implicit_exception_uncommon_trap_offset = 0; 3050 int uncommon_trap_offset = 0; 3051 3052 if (EnableJVMCI || UseAOT) { 3053 implicit_exception_uncommon_trap_offset = __ pc() - start; 3054 3055 __ pushptr(Address(r15_thread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset()))); 3056 __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())), (int32_t)NULL_WORD); 3057 3058 uncommon_trap_offset = __ pc() - start; 3059 3060 // Save everything in sight. 3061 RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 3062 // fetch_unroll_info needs to call last_java_frame() 3063 __ set_last_Java_frame(noreg, noreg, NULL); 3064 3065 __ movl(c_rarg1, Address(r15_thread, in_bytes(JavaThread::pending_deoptimization_offset()))); 3066 __ movl(Address(r15_thread, in_bytes(JavaThread::pending_deoptimization_offset())), -1); 3067 3068 __ movl(r14, (int32_t)Deoptimization::Unpack_reexecute); 3069 __ mov(c_rarg0, r15_thread); 3070 __ movl(c_rarg2, r14); // exec mode 3071 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap))); 3072 oop_maps->add_gc_map( __ pc()-start, map->deep_copy()); 3073 3074 __ reset_last_Java_frame(false); 3075 3076 __ jmp(after_fetch_unroll_info_call); 3077 } // EnableJVMCI 3078 #endif // INCLUDE_JVMCI 3079 3080 int exception_offset = __ pc() - start; 3081 3082 // Prolog for exception case 3083 3084 // all registers are dead at this entry point, except for rax, and 3085 // rdx which contain the exception oop and exception pc 3086 // respectively. Set them in TLS and fall thru to the 3087 // unpack_with_exception_in_tls entry point. 3088 3089 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), rdx); 3090 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()), rax); 3091 3092 int exception_in_tls_offset = __ pc() - start; 3093 3094 // new implementation because exception oop is now passed in JavaThread 3095 3096 // Prolog for exception case 3097 // All registers must be preserved because they might be used by LinearScan 3098 // Exceptiop oop and throwing PC are passed in JavaThread 3099 // tos: stack at point of call to method that threw the exception (i.e. only 3100 // args are on the stack, no return address) 3101 3102 // make room on stack for the return address 3103 // It will be patched later with the throwing pc. The correct value is not 3104 // available now because loading it from memory would destroy registers. 3105 __ push(0); 3106 3107 // Save everything in sight. 3108 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 3109 3110 // Now it is safe to overwrite any register 3111 3112 // Deopt during an exception. Save exec mode for unpack_frames. 3113 __ movl(r14, Deoptimization::Unpack_exception); // callee-saved 3114 3115 // load throwing pc from JavaThread and patch it as the return address 3116 // of the current frame. Then clear the field in JavaThread 3117 3118 __ movptr(rdx, Address(r15_thread, JavaThread::exception_pc_offset())); 3119 __ movptr(Address(rbp, wordSize), rdx); 3120 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), (int32_t)NULL_WORD); 3121 3122 #ifdef ASSERT 3123 // verify that there is really an exception oop in JavaThread 3124 __ movptr(rax, Address(r15_thread, JavaThread::exception_oop_offset())); 3125 __ verify_oop(rax); 3126 3127 // verify that there is no pending exception 3128 Label no_pending_exception; 3129 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset())); 3130 __ testptr(rax, rax); 3131 __ jcc(Assembler::zero, no_pending_exception); 3132 __ stop("must not have pending exception here"); 3133 __ bind(no_pending_exception); 3134 #endif 3135 3136 __ bind(cont); 3137 3138 // Call C code. Need thread and this frame, but NOT official VM entry 3139 // crud. We cannot block on this call, no GC can happen. 3140 // 3141 // UnrollBlock* fetch_unroll_info(JavaThread* thread) 3142 3143 // fetch_unroll_info needs to call last_java_frame(). 3144 3145 __ set_last_Java_frame(noreg, noreg, NULL); 3146 #ifdef ASSERT 3147 { Label L; 3148 __ cmpptr(Address(r15_thread, 3149 JavaThread::last_Java_fp_offset()), 3150 (int32_t)0); 3151 __ jcc(Assembler::equal, L); 3152 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared"); 3153 __ bind(L); 3154 } 3155 #endif // ASSERT 3156 __ mov(c_rarg0, r15_thread); 3157 __ movl(c_rarg1, r14); // exec_mode 3158 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info))); 3159 3160 // Need to have an oopmap that tells fetch_unroll_info where to 3161 // find any register it might need. 3162 oop_maps->add_gc_map(__ pc() - start, map); 3163 3164 __ reset_last_Java_frame(false); 3165 3166 #if INCLUDE_JVMCI 3167 if (EnableJVMCI || UseAOT) { 3168 __ bind(after_fetch_unroll_info_call); 3169 } 3170 #endif 3171 3172 // Load UnrollBlock* into rdi 3173 __ mov(rdi, rax); 3174 3175 __ movl(r14, Address(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes())); 3176 Label noException; 3177 __ cmpl(r14, Deoptimization::Unpack_exception); // Was exception pending? 3178 __ jcc(Assembler::notEqual, noException); 3179 __ movptr(rax, Address(r15_thread, JavaThread::exception_oop_offset())); 3180 // QQQ this is useless it was NULL above 3181 __ movptr(rdx, Address(r15_thread, JavaThread::exception_pc_offset())); 3182 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()), (int32_t)NULL_WORD); 3183 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), (int32_t)NULL_WORD); 3184 3185 __ verify_oop(rax); 3186 3187 // Overwrite the result registers with the exception results. 3188 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax); 3189 // I think this is useless 3190 __ movptr(Address(rsp, RegisterSaver::rdx_offset_in_bytes()), rdx); 3191 3192 __ bind(noException); 3193 3194 // Only register save data is on the stack. 3195 // Now restore the result registers. Everything else is either dead 3196 // or captured in the vframeArray. 3197 RegisterSaver::restore_result_registers(masm); 3198 3199 // All of the register save area has been popped of the stack. Only the 3200 // return address remains. 3201 3202 // Pop all the frames we must move/replace. 3203 // 3204 // Frame picture (youngest to oldest) 3205 // 1: self-frame (no frame link) 3206 // 2: deopting frame (no frame link) 3207 // 3: caller of deopting frame (could be compiled/interpreted). 3208 // 3209 // Note: by leaving the return address of self-frame on the stack 3210 // and using the size of frame 2 to adjust the stack 3211 // when we are done the return to frame 3 will still be on the stack. 3212 3213 // Pop deoptimized frame 3214 __ movl(rcx, Address(rdi, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes())); 3215 __ addptr(rsp, rcx); 3216 3217 // rsp should be pointing at the return address to the caller (3) 3218 3219 // Pick up the initial fp we should save 3220 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved) 3221 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 3222 3223 #ifdef ASSERT 3224 // Compilers generate code that bang the stack by as much as the 3225 // interpreter would need. So this stack banging should never 3226 // trigger a fault. Verify that it does not on non product builds. 3227 if (UseStackBanging) { 3228 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 3229 __ bang_stack_size(rbx, rcx); 3230 } 3231 #endif 3232 3233 // Load address of array of frame pcs into rcx 3234 __ movptr(rcx, Address(rdi, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 3235 3236 // Trash the old pc 3237 __ addptr(rsp, wordSize); 3238 3239 // Load address of array of frame sizes into rsi 3240 __ movptr(rsi, Address(rdi, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes())); 3241 3242 // Load counter into rdx 3243 __ movl(rdx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes())); 3244 3245 // Now adjust the caller's stack to make up for the extra locals 3246 // but record the original sp so that we can save it in the skeletal interpreter 3247 // frame and the stack walking of interpreter_sender will get the unextended sp 3248 // value and not the "real" sp value. 3249 3250 const Register sender_sp = r8; 3251 3252 __ mov(sender_sp, rsp); 3253 __ movl(rbx, Address(rdi, 3254 Deoptimization::UnrollBlock:: 3255 caller_adjustment_offset_in_bytes())); 3256 __ subptr(rsp, rbx); 3257 3258 // Push interpreter frames in a loop 3259 Label loop; 3260 __ bind(loop); 3261 __ movptr(rbx, Address(rsi, 0)); // Load frame size 3262 __ subptr(rbx, 2*wordSize); // We'll push pc and ebp by hand 3263 __ pushptr(Address(rcx, 0)); // Save return address 3264 __ enter(); // Save old & set new ebp 3265 __ subptr(rsp, rbx); // Prolog 3266 // This value is corrected by layout_activation_impl 3267 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), (int32_t)NULL_WORD ); 3268 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), sender_sp); // Make it walkable 3269 __ mov(sender_sp, rsp); // Pass sender_sp to next frame 3270 __ addptr(rsi, wordSize); // Bump array pointer (sizes) 3271 __ addptr(rcx, wordSize); // Bump array pointer (pcs) 3272 __ decrementl(rdx); // Decrement counter 3273 __ jcc(Assembler::notZero, loop); 3274 __ pushptr(Address(rcx, 0)); // Save final return address 3275 3276 // Re-push self-frame 3277 __ enter(); // Save old & set new ebp 3278 3279 // Allocate a full sized register save area. 3280 // Return address and rbp are in place, so we allocate two less words. 3281 __ subptr(rsp, (frame_size_in_words - 2) * wordSize); 3282 3283 // Restore frame locals after moving the frame 3284 __ movdbl(Address(rsp, RegisterSaver::xmm0_offset_in_bytes()), xmm0); 3285 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax); 3286 3287 // Call C code. Need thread but NOT official VM entry 3288 // crud. We cannot block on this call, no GC can happen. Call should 3289 // restore return values to their stack-slots with the new SP. 3290 // 3291 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode) 3292 3293 // Use rbp because the frames look interpreted now 3294 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 3295 // Don't need the precise return PC here, just precise enough to point into this code blob. 3296 address the_pc = __ pc(); 3297 __ set_last_Java_frame(noreg, rbp, the_pc); 3298 3299 __ andptr(rsp, -(StackAlignmentInBytes)); // Fix stack alignment as required by ABI 3300 __ mov(c_rarg0, r15_thread); 3301 __ movl(c_rarg1, r14); // second arg: exec_mode 3302 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 3303 // Revert SP alignment after call since we're going to do some SP relative addressing below 3304 __ movptr(rsp, Address(r15_thread, JavaThread::last_Java_sp_offset())); 3305 3306 // Set an oopmap for the call site 3307 // Use the same PC we used for the last java frame 3308 oop_maps->add_gc_map(the_pc - start, 3309 new OopMap( frame_size_in_words, 0 )); 3310 3311 // Clear fp AND pc 3312 __ reset_last_Java_frame(true); 3313 3314 // Collect return values 3315 __ movdbl(xmm0, Address(rsp, RegisterSaver::xmm0_offset_in_bytes())); 3316 __ movptr(rax, Address(rsp, RegisterSaver::rax_offset_in_bytes())); 3317 // I think this is useless (throwing pc?) 3318 __ movptr(rdx, Address(rsp, RegisterSaver::rdx_offset_in_bytes())); 3319 3320 // Pop self-frame. 3321 __ leave(); // Epilog 3322 3323 // Jump to interpreter 3324 __ ret(0); 3325 3326 // Make sure all code is generated 3327 masm->flush(); 3328 3329 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words); 3330 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 3331 #if INCLUDE_JVMCI 3332 if (EnableJVMCI || UseAOT) { 3333 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset); 3334 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset); 3335 } 3336 #endif 3337 } 3338 3339 #ifdef COMPILER2 3340 //------------------------------generate_uncommon_trap_blob-------------------- 3341 void SharedRuntime::generate_uncommon_trap_blob() { 3342 // Allocate space for the code 3343 ResourceMark rm; 3344 // Setup code generation tools 3345 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024); 3346 MacroAssembler* masm = new MacroAssembler(&buffer); 3347 3348 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 3349 3350 address start = __ pc(); 3351 3352 if (UseRTMLocking) { 3353 // Abort RTM transaction before possible nmethod deoptimization. 3354 __ xabort(0); 3355 } 3356 3357 // Push self-frame. We get here with a return address on the 3358 // stack, so rsp is 8-byte aligned until we allocate our frame. 3359 __ subptr(rsp, SimpleRuntimeFrame::return_off << LogBytesPerInt); // Epilog! 3360 3361 // No callee saved registers. rbp is assumed implicitly saved 3362 __ movptr(Address(rsp, SimpleRuntimeFrame::rbp_off << LogBytesPerInt), rbp); 3363 3364 // compiler left unloaded_class_index in j_rarg0 move to where the 3365 // runtime expects it. 3366 __ movl(c_rarg1, j_rarg0); 3367 3368 __ set_last_Java_frame(noreg, noreg, NULL); 3369 3370 // Call C code. Need thread but NOT official VM entry 3371 // crud. We cannot block on this call, no GC can happen. Call should 3372 // capture callee-saved registers as well as return values. 3373 // Thread is in rdi already. 3374 // 3375 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index); 3376 3377 __ mov(c_rarg0, r15_thread); 3378 __ movl(c_rarg2, Deoptimization::Unpack_uncommon_trap); 3379 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap))); 3380 3381 // Set an oopmap for the call site 3382 OopMapSet* oop_maps = new OopMapSet(); 3383 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0); 3384 3385 // location of rbp is known implicitly by the frame sender code 3386 3387 oop_maps->add_gc_map(__ pc() - start, map); 3388 3389 __ reset_last_Java_frame(false); 3390 3391 // Load UnrollBlock* into rdi 3392 __ mov(rdi, rax); 3393 3394 #ifdef ASSERT 3395 { Label L; 3396 __ cmpptr(Address(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()), 3397 (int32_t)Deoptimization::Unpack_uncommon_trap); 3398 __ jcc(Assembler::equal, L); 3399 __ stop("SharedRuntime::generate_deopt_blob: expected Unpack_uncommon_trap"); 3400 __ bind(L); 3401 } 3402 #endif 3403 3404 // Pop all the frames we must move/replace. 3405 // 3406 // Frame picture (youngest to oldest) 3407 // 1: self-frame (no frame link) 3408 // 2: deopting frame (no frame link) 3409 // 3: caller of deopting frame (could be compiled/interpreted). 3410 3411 // Pop self-frame. We have no frame, and must rely only on rax and rsp. 3412 __ addptr(rsp, (SimpleRuntimeFrame::framesize - 2) << LogBytesPerInt); // Epilog! 3413 3414 // Pop deoptimized frame (int) 3415 __ movl(rcx, Address(rdi, 3416 Deoptimization::UnrollBlock:: 3417 size_of_deoptimized_frame_offset_in_bytes())); 3418 __ addptr(rsp, rcx); 3419 3420 // rsp should be pointing at the return address to the caller (3) 3421 3422 // Pick up the initial fp we should save 3423 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved) 3424 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes())); 3425 3426 #ifdef ASSERT 3427 // Compilers generate code that bang the stack by as much as the 3428 // interpreter would need. So this stack banging should never 3429 // trigger a fault. Verify that it does not on non product builds. 3430 if (UseStackBanging) { 3431 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes())); 3432 __ bang_stack_size(rbx, rcx); 3433 } 3434 #endif 3435 3436 // Load address of array of frame pcs into rcx (address*) 3437 __ movptr(rcx, Address(rdi, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes())); 3438 3439 // Trash the return pc 3440 __ addptr(rsp, wordSize); 3441 3442 // Load address of array of frame sizes into rsi (intptr_t*) 3443 __ movptr(rsi, Address(rdi, Deoptimization::UnrollBlock:: frame_sizes_offset_in_bytes())); 3444 3445 // Counter 3446 __ movl(rdx, Address(rdi, Deoptimization::UnrollBlock:: number_of_frames_offset_in_bytes())); // (int) 3447 3448 // Now adjust the caller's stack to make up for the extra locals but 3449 // record the original sp so that we can save it in the skeletal 3450 // interpreter frame and the stack walking of interpreter_sender 3451 // will get the unextended sp value and not the "real" sp value. 3452 3453 const Register sender_sp = r8; 3454 3455 __ mov(sender_sp, rsp); 3456 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock:: caller_adjustment_offset_in_bytes())); // (int) 3457 __ subptr(rsp, rbx); 3458 3459 // Push interpreter frames in a loop 3460 Label loop; 3461 __ bind(loop); 3462 __ movptr(rbx, Address(rsi, 0)); // Load frame size 3463 __ subptr(rbx, 2 * wordSize); // We'll push pc and rbp by hand 3464 __ pushptr(Address(rcx, 0)); // Save return address 3465 __ enter(); // Save old & set new rbp 3466 __ subptr(rsp, rbx); // Prolog 3467 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), 3468 sender_sp); // Make it walkable 3469 // This value is corrected by layout_activation_impl 3470 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), (int32_t)NULL_WORD ); 3471 __ mov(sender_sp, rsp); // Pass sender_sp to next frame 3472 __ addptr(rsi, wordSize); // Bump array pointer (sizes) 3473 __ addptr(rcx, wordSize); // Bump array pointer (pcs) 3474 __ decrementl(rdx); // Decrement counter 3475 __ jcc(Assembler::notZero, loop); 3476 __ pushptr(Address(rcx, 0)); // Save final return address 3477 3478 // Re-push self-frame 3479 __ enter(); // Save old & set new rbp 3480 __ subptr(rsp, (SimpleRuntimeFrame::framesize - 4) << LogBytesPerInt); 3481 // Prolog 3482 3483 // Use rbp because the frames look interpreted now 3484 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP. 3485 // Don't need the precise return PC here, just precise enough to point into this code blob. 3486 address the_pc = __ pc(); 3487 __ set_last_Java_frame(noreg, rbp, the_pc); 3488 3489 // Call C code. Need thread but NOT official VM entry 3490 // crud. We cannot block on this call, no GC can happen. Call should 3491 // restore return values to their stack-slots with the new SP. 3492 // Thread is in rdi already. 3493 // 3494 // BasicType unpack_frames(JavaThread* thread, int exec_mode); 3495 3496 __ andptr(rsp, -(StackAlignmentInBytes)); // Align SP as required by ABI 3497 __ mov(c_rarg0, r15_thread); 3498 __ movl(c_rarg1, Deoptimization::Unpack_uncommon_trap); 3499 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames))); 3500 3501 // Set an oopmap for the call site 3502 // Use the same PC we used for the last java frame 3503 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 3504 3505 // Clear fp AND pc 3506 __ reset_last_Java_frame(true); 3507 3508 // Pop self-frame. 3509 __ leave(); // Epilog 3510 3511 // Jump to interpreter 3512 __ ret(0); 3513 3514 // Make sure all code is generated 3515 masm->flush(); 3516 3517 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, 3518 SimpleRuntimeFrame::framesize >> 1); 3519 } 3520 #endif // COMPILER2 3521 3522 3523 //------------------------------generate_handler_blob------ 3524 // 3525 // Generate a special Compile2Runtime blob that saves all registers, 3526 // and setup oopmap. 3527 // 3528 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) { 3529 assert(StubRoutines::forward_exception_entry() != NULL, 3530 "must be generated before"); 3531 3532 ResourceMark rm; 3533 OopMapSet *oop_maps = new OopMapSet(); 3534 OopMap* map; 3535 3536 // Allocate space for the code. Setup code generation tools. 3537 CodeBuffer buffer("handler_blob", 2048, 1024); 3538 MacroAssembler* masm = new MacroAssembler(&buffer); 3539 3540 address start = __ pc(); 3541 address call_pc = NULL; 3542 int frame_size_in_words; 3543 bool cause_return = (poll_type == POLL_AT_RETURN); 3544 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP); 3545 3546 if (UseRTMLocking) { 3547 // Abort RTM transaction before calling runtime 3548 // because critical section will be large and will be 3549 // aborted anyway. Also nmethod could be deoptimized. 3550 __ xabort(0); 3551 } 3552 3553 // Make room for return address (or push it again) 3554 if (!cause_return) { 3555 __ push(rbx); 3556 } 3557 3558 // Save registers, fpu state, and flags 3559 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, save_vectors); 3560 3561 // The following is basically a call_VM. However, we need the precise 3562 // address of the call in order to generate an oopmap. Hence, we do all the 3563 // work outselves. 3564 3565 __ set_last_Java_frame(noreg, noreg, NULL); 3566 3567 // The return address must always be correct so that frame constructor never 3568 // sees an invalid pc. 3569 3570 if (!cause_return) { 3571 // overwrite the dummy value we pushed on entry 3572 __ movptr(c_rarg0, Address(r15_thread, JavaThread::saved_exception_pc_offset())); 3573 __ movptr(Address(rbp, wordSize), c_rarg0); 3574 } 3575 3576 // Do the call 3577 __ mov(c_rarg0, r15_thread); 3578 __ call(RuntimeAddress(call_ptr)); 3579 3580 // Set an oopmap for the call site. This oopmap will map all 3581 // oop-registers and debug-info registers as callee-saved. This 3582 // will allow deoptimization at this safepoint to find all possible 3583 // debug-info recordings, as well as let GC find all oops. 3584 3585 oop_maps->add_gc_map( __ pc() - start, map); 3586 3587 Label noException; 3588 3589 __ reset_last_Java_frame(false); 3590 3591 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 3592 __ jcc(Assembler::equal, noException); 3593 3594 // Exception pending 3595 3596 RegisterSaver::restore_live_registers(masm, save_vectors); 3597 3598 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3599 3600 // No exception case 3601 __ bind(noException); 3602 3603 // Normal exit, restore registers and exit. 3604 RegisterSaver::restore_live_registers(masm, save_vectors); 3605 3606 __ ret(0); 3607 3608 // Make sure all code is generated 3609 masm->flush(); 3610 3611 // Fill-out other meta info 3612 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words); 3613 } 3614 3615 // 3616 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 3617 // 3618 // Generate a stub that calls into vm to find out the proper destination 3619 // of a java call. All the argument registers are live at this point 3620 // but since this is generic code we don't know what they are and the caller 3621 // must do any gc of the args. 3622 // 3623 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3624 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 3625 3626 // allocate space for the code 3627 ResourceMark rm; 3628 3629 CodeBuffer buffer(name, 1000, 512); 3630 MacroAssembler* masm = new MacroAssembler(&buffer); 3631 3632 int frame_size_in_words; 3633 3634 OopMapSet *oop_maps = new OopMapSet(); 3635 OopMap* map = NULL; 3636 3637 int start = __ offset(); 3638 3639 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words); 3640 3641 int frame_complete = __ offset(); 3642 3643 __ set_last_Java_frame(noreg, noreg, NULL); 3644 3645 __ mov(c_rarg0, r15_thread); 3646 3647 __ call(RuntimeAddress(destination)); 3648 3649 3650 // Set an oopmap for the call site. 3651 // We need this not only for callee-saved registers, but also for volatile 3652 // registers that the compiler might be keeping live across a safepoint. 3653 3654 oop_maps->add_gc_map( __ offset() - start, map); 3655 3656 // rax contains the address we are going to jump to assuming no exception got installed 3657 3658 // clear last_Java_sp 3659 __ reset_last_Java_frame(false); 3660 // check for pending exceptions 3661 Label pending; 3662 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD); 3663 __ jcc(Assembler::notEqual, pending); 3664 3665 // get the returned Method* 3666 __ get_vm_result_2(rbx, r15_thread); 3667 __ movptr(Address(rsp, RegisterSaver::rbx_offset_in_bytes()), rbx); 3668 3669 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax); 3670 3671 RegisterSaver::restore_live_registers(masm); 3672 3673 // We are back the the original state on entry and ready to go. 3674 3675 __ jmp(rax); 3676 3677 // Pending exception after the safepoint 3678 3679 __ bind(pending); 3680 3681 RegisterSaver::restore_live_registers(masm); 3682 3683 // exception pending => remove activation and forward to exception handler 3684 3685 __ movptr(Address(r15_thread, JavaThread::vm_result_offset()), (int)NULL_WORD); 3686 3687 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset())); 3688 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry())); 3689 3690 // ------------- 3691 // make sure all code is generated 3692 masm->flush(); 3693 3694 // return the blob 3695 // frame_size_words or bytes?? 3696 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true); 3697 } 3698 3699 3700 //------------------------------Montgomery multiplication------------------------ 3701 // 3702 3703 #ifndef _WINDOWS 3704 3705 #define ASM_SUBTRACT 3706 3707 #ifdef ASM_SUBTRACT 3708 // Subtract 0:b from carry:a. Return carry. 3709 static unsigned long 3710 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) { 3711 long i = 0, cnt = len; 3712 unsigned long tmp; 3713 asm volatile("clc; " 3714 "0: ; " 3715 "mov (%[b], %[i], 8), %[tmp]; " 3716 "sbb %[tmp], (%[a], %[i], 8); " 3717 "inc %[i]; dec %[cnt]; " 3718 "jne 0b; " 3719 "mov %[carry], %[tmp]; sbb $0, %[tmp]; " 3720 : [i]"+r"(i), [cnt]"+r"(cnt), [tmp]"=&r"(tmp) 3721 : [a]"r"(a), [b]"r"(b), [carry]"r"(carry) 3722 : "memory"); 3723 return tmp; 3724 } 3725 #else // ASM_SUBTRACT 3726 typedef int __attribute__((mode(TI))) int128; 3727 3728 // Subtract 0:b from carry:a. Return carry. 3729 static unsigned long 3730 sub(unsigned long a[], unsigned long b[], unsigned long carry, int len) { 3731 int128 tmp = 0; 3732 int i; 3733 for (i = 0; i < len; i++) { 3734 tmp += a[i]; 3735 tmp -= b[i]; 3736 a[i] = tmp; 3737 tmp >>= 64; 3738 assert(-1 <= tmp && tmp <= 0, "invariant"); 3739 } 3740 return tmp + carry; 3741 } 3742 #endif // ! ASM_SUBTRACT 3743 3744 // Multiply (unsigned) Long A by Long B, accumulating the double- 3745 // length result into the accumulator formed of T0, T1, and T2. 3746 #define MACC(A, B, T0, T1, T2) \ 3747 do { \ 3748 unsigned long hi, lo; \ 3749 __asm__ ("mul %5; add %%rax, %2; adc %%rdx, %3; adc $0, %4" \ 3750 : "=&d"(hi), "=a"(lo), "+r"(T0), "+r"(T1), "+g"(T2) \ 3751 : "r"(A), "a"(B) : "cc"); \ 3752 } while(0) 3753 3754 // As above, but add twice the double-length result into the 3755 // accumulator. 3756 #define MACC2(A, B, T0, T1, T2) \ 3757 do { \ 3758 unsigned long hi, lo; \ 3759 __asm__ ("mul %5; add %%rax, %2; adc %%rdx, %3; adc $0, %4; " \ 3760 "add %%rax, %2; adc %%rdx, %3; adc $0, %4" \ 3761 : "=&d"(hi), "=a"(lo), "+r"(T0), "+r"(T1), "+g"(T2) \ 3762 : "r"(A), "a"(B) : "cc"); \ 3763 } while(0) 3764 3765 // Fast Montgomery multiplication. The derivation of the algorithm is 3766 // in A Cryptographic Library for the Motorola DSP56000, 3767 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237. 3768 3769 static void __attribute__((noinline)) 3770 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[], 3771 unsigned long m[], unsigned long inv, int len) { 3772 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator 3773 int i; 3774 3775 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply"); 3776 3777 for (i = 0; i < len; i++) { 3778 int j; 3779 for (j = 0; j < i; j++) { 3780 MACC(a[j], b[i-j], t0, t1, t2); 3781 MACC(m[j], n[i-j], t0, t1, t2); 3782 } 3783 MACC(a[i], b[0], t0, t1, t2); 3784 m[i] = t0 * inv; 3785 MACC(m[i], n[0], t0, t1, t2); 3786 3787 assert(t0 == 0, "broken Montgomery multiply"); 3788 3789 t0 = t1; t1 = t2; t2 = 0; 3790 } 3791 3792 for (i = len; i < 2*len; i++) { 3793 int j; 3794 for (j = i-len+1; j < len; j++) { 3795 MACC(a[j], b[i-j], t0, t1, t2); 3796 MACC(m[j], n[i-j], t0, t1, t2); 3797 } 3798 m[i-len] = t0; 3799 t0 = t1; t1 = t2; t2 = 0; 3800 } 3801 3802 while (t0) 3803 t0 = sub(m, n, t0, len); 3804 } 3805 3806 // Fast Montgomery squaring. This uses asymptotically 25% fewer 3807 // multiplies so it should be up to 25% faster than Montgomery 3808 // multiplication. However, its loop control is more complex and it 3809 // may actually run slower on some machines. 3810 3811 static void __attribute__((noinline)) 3812 montgomery_square(unsigned long a[], unsigned long n[], 3813 unsigned long m[], unsigned long inv, int len) { 3814 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator 3815 int i; 3816 3817 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply"); 3818 3819 for (i = 0; i < len; i++) { 3820 int j; 3821 int end = (i+1)/2; 3822 for (j = 0; j < end; j++) { 3823 MACC2(a[j], a[i-j], t0, t1, t2); 3824 MACC(m[j], n[i-j], t0, t1, t2); 3825 } 3826 if ((i & 1) == 0) { 3827 MACC(a[j], a[j], t0, t1, t2); 3828 } 3829 for (; j < i; j++) { 3830 MACC(m[j], n[i-j], t0, t1, t2); 3831 } 3832 m[i] = t0 * inv; 3833 MACC(m[i], n[0], t0, t1, t2); 3834 3835 assert(t0 == 0, "broken Montgomery square"); 3836 3837 t0 = t1; t1 = t2; t2 = 0; 3838 } 3839 3840 for (i = len; i < 2*len; i++) { 3841 int start = i-len+1; 3842 int end = start + (len - start)/2; 3843 int j; 3844 for (j = start; j < end; j++) { 3845 MACC2(a[j], a[i-j], t0, t1, t2); 3846 MACC(m[j], n[i-j], t0, t1, t2); 3847 } 3848 if ((i & 1) == 0) { 3849 MACC(a[j], a[j], t0, t1, t2); 3850 } 3851 for (; j < len; j++) { 3852 MACC(m[j], n[i-j], t0, t1, t2); 3853 } 3854 m[i-len] = t0; 3855 t0 = t1; t1 = t2; t2 = 0; 3856 } 3857 3858 while (t0) 3859 t0 = sub(m, n, t0, len); 3860 } 3861 3862 // Swap words in a longword. 3863 static unsigned long swap(unsigned long x) { 3864 return (x << 32) | (x >> 32); 3865 } 3866 3867 // Copy len longwords from s to d, word-swapping as we go. The 3868 // destination array is reversed. 3869 static void reverse_words(unsigned long *s, unsigned long *d, int len) { 3870 d += len; 3871 while(len-- > 0) { 3872 d--; 3873 *d = swap(*s); 3874 s++; 3875 } 3876 } 3877 3878 // The threshold at which squaring is advantageous was determined 3879 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz. 3880 #define MONTGOMERY_SQUARING_THRESHOLD 64 3881 3882 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints, 3883 jint len, jlong inv, 3884 jint *m_ints) { 3885 assert(len % 2 == 0, "array length in montgomery_multiply must be even"); 3886 int longwords = len/2; 3887 3888 // Make very sure we don't use so much space that the stack might 3889 // overflow. 512 jints corresponds to an 16384-bit integer and 3890 // will use here a total of 8k bytes of stack space. 3891 int total_allocation = longwords * sizeof (unsigned long) * 4; 3892 guarantee(total_allocation <= 8192, "must be"); 3893 unsigned long *scratch = (unsigned long *)alloca(total_allocation); 3894 3895 // Local scratch arrays 3896 unsigned long 3897 *a = scratch + 0 * longwords, 3898 *b = scratch + 1 * longwords, 3899 *n = scratch + 2 * longwords, 3900 *m = scratch + 3 * longwords; 3901 3902 reverse_words((unsigned long *)a_ints, a, longwords); 3903 reverse_words((unsigned long *)b_ints, b, longwords); 3904 reverse_words((unsigned long *)n_ints, n, longwords); 3905 3906 ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords); 3907 3908 reverse_words(m, (unsigned long *)m_ints, longwords); 3909 } 3910 3911 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints, 3912 jint len, jlong inv, 3913 jint *m_ints) { 3914 assert(len % 2 == 0, "array length in montgomery_square must be even"); 3915 int longwords = len/2; 3916 3917 // Make very sure we don't use so much space that the stack might 3918 // overflow. 512 jints corresponds to an 16384-bit integer and 3919 // will use here a total of 6k bytes of stack space. 3920 int total_allocation = longwords * sizeof (unsigned long) * 3; 3921 guarantee(total_allocation <= 8192, "must be"); 3922 unsigned long *scratch = (unsigned long *)alloca(total_allocation); 3923 3924 // Local scratch arrays 3925 unsigned long 3926 *a = scratch + 0 * longwords, 3927 *n = scratch + 1 * longwords, 3928 *m = scratch + 2 * longwords; 3929 3930 reverse_words((unsigned long *)a_ints, a, longwords); 3931 reverse_words((unsigned long *)n_ints, n, longwords); 3932 3933 if (len >= MONTGOMERY_SQUARING_THRESHOLD) { 3934 ::montgomery_square(a, n, m, (unsigned long)inv, longwords); 3935 } else { 3936 ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords); 3937 } 3938 3939 reverse_words(m, (unsigned long *)m_ints, longwords); 3940 } 3941 3942 #endif // WINDOWS 3943 3944 #ifdef COMPILER2 3945 // This is here instead of runtime_x86_64.cpp because it uses SimpleRuntimeFrame 3946 // 3947 //------------------------------generate_exception_blob--------------------------- 3948 // creates exception blob at the end 3949 // Using exception blob, this code is jumped from a compiled method. 3950 // (see emit_exception_handler in x86_64.ad file) 3951 // 3952 // Given an exception pc at a call we call into the runtime for the 3953 // handler in this method. This handler might merely restore state 3954 // (i.e. callee save registers) unwind the frame and jump to the 3955 // exception handler for the nmethod if there is no Java level handler 3956 // for the nmethod. 3957 // 3958 // This code is entered with a jmp. 3959 // 3960 // Arguments: 3961 // rax: exception oop 3962 // rdx: exception pc 3963 // 3964 // Results: 3965 // rax: exception oop 3966 // rdx: exception pc in caller or ??? 3967 // destination: exception handler of caller 3968 // 3969 // Note: the exception pc MUST be at a call (precise debug information) 3970 // Registers rax, rdx, rcx, rsi, rdi, r8-r11 are not callee saved. 3971 // 3972 3973 void OptoRuntime::generate_exception_blob() { 3974 assert(!OptoRuntime::is_callee_saved_register(RDX_num), ""); 3975 assert(!OptoRuntime::is_callee_saved_register(RAX_num), ""); 3976 assert(!OptoRuntime::is_callee_saved_register(RCX_num), ""); 3977 3978 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned"); 3979 3980 // Allocate space for the code 3981 ResourceMark rm; 3982 // Setup code generation tools 3983 CodeBuffer buffer("exception_blob", 2048, 1024); 3984 MacroAssembler* masm = new MacroAssembler(&buffer); 3985 3986 3987 address start = __ pc(); 3988 3989 // Exception pc is 'return address' for stack walker 3990 __ push(rdx); 3991 __ subptr(rsp, SimpleRuntimeFrame::return_off << LogBytesPerInt); // Prolog 3992 3993 // Save callee-saved registers. See x86_64.ad. 3994 3995 // rbp is an implicitly saved callee saved register (i.e., the calling 3996 // convention will save/restore it in the prolog/epilog). Other than that 3997 // there are no callee save registers now that adapter frames are gone. 3998 3999 __ movptr(Address(rsp, SimpleRuntimeFrame::rbp_off << LogBytesPerInt), rbp); 4000 4001 // Store exception in Thread object. We cannot pass any arguments to the 4002 // handle_exception call, since we do not want to make any assumption 4003 // about the size of the frame where the exception happened in. 4004 // c_rarg0 is either rdi (Linux) or rcx (Windows). 4005 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()),rax); 4006 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), rdx); 4007 4008 // This call does all the hard work. It checks if an exception handler 4009 // exists in the method. 4010 // If so, it returns the handler address. 4011 // If not, it prepares for stack-unwinding, restoring the callee-save 4012 // registers of the frame being removed. 4013 // 4014 // address OptoRuntime::handle_exception_C(JavaThread* thread) 4015 4016 // At a method handle call, the stack may not be properly aligned 4017 // when returning with an exception. 4018 address the_pc = __ pc(); 4019 __ set_last_Java_frame(noreg, noreg, the_pc); 4020 __ mov(c_rarg0, r15_thread); 4021 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack 4022 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C))); 4023 4024 // Set an oopmap for the call site. This oopmap will only be used if we 4025 // are unwinding the stack. Hence, all locations will be dead. 4026 // Callee-saved registers will be the same as the frame above (i.e., 4027 // handle_exception_stub), since they were restored when we got the 4028 // exception. 4029 4030 OopMapSet* oop_maps = new OopMapSet(); 4031 4032 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0)); 4033 4034 __ reset_last_Java_frame(false); 4035 4036 // Restore callee-saved registers 4037 4038 // rbp is an implicitly saved callee-saved register (i.e., the calling 4039 // convention will save restore it in prolog/epilog) Other than that 4040 // there are no callee save registers now that adapter frames are gone. 4041 4042 __ movptr(rbp, Address(rsp, SimpleRuntimeFrame::rbp_off << LogBytesPerInt)); 4043 4044 __ addptr(rsp, SimpleRuntimeFrame::return_off << LogBytesPerInt); // Epilog 4045 __ pop(rdx); // No need for exception pc anymore 4046 4047 // rax: exception handler 4048 4049 // We have a handler in rax (could be deopt blob). 4050 __ mov(r8, rax); 4051 4052 // Get the exception oop 4053 __ movptr(rax, Address(r15_thread, JavaThread::exception_oop_offset())); 4054 // Get the exception pc in case we are deoptimized 4055 __ movptr(rdx, Address(r15_thread, JavaThread::exception_pc_offset())); 4056 #ifdef ASSERT 4057 __ movptr(Address(r15_thread, JavaThread::exception_handler_pc_offset()), (int)NULL_WORD); 4058 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), (int)NULL_WORD); 4059 #endif 4060 // Clear the exception oop so GC no longer processes it as a root. 4061 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()), (int)NULL_WORD); 4062 4063 // rax: exception oop 4064 // r8: exception handler 4065 // rdx: exception pc 4066 // Jump to handler 4067 4068 __ jmp(r8); 4069 4070 // Make sure all code is generated 4071 masm->flush(); 4072 4073 // Set exception blob 4074 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1); 4075 } 4076 #endif // COMPILER2