1 /* 2 * Copyright (c) 2003, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.hpp" 27 #include "assembler_sparc.inline.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "code/icBuffer.hpp" 30 #include "code/vtableStubs.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "oops/compiledICHolderOop.hpp" 33 #include "prims/jvmtiRedefineClassesTrace.hpp" 34 #include "runtime/sharedRuntime.hpp" 35 #include "runtime/vframeArray.hpp" 36 #include "vmreg_sparc.inline.hpp" 37 #ifdef COMPILER1 38 #include "c1/c1_Runtime1.hpp" 39 #endif 40 #ifdef COMPILER2 41 #include "opto/runtime.hpp" 42 #endif 43 #ifdef SHARK 44 #include "compiler/compileBroker.hpp" 45 #include "shark/sharkCompiler.hpp" 46 #endif 47 48 #define __ masm-> 49 50 51 class RegisterSaver { 52 53 // Used for saving volatile registers. This is Gregs, Fregs, I/L/O. 54 // The Oregs are problematic. In the 32bit build the compiler can 55 // have O registers live with 64 bit quantities. A window save will 56 // cut the heads off of the registers. We have to do a very extensive 57 // stack dance to save and restore these properly. 58 59 // Note that the Oregs problem only exists if we block at either a polling 60 // page exception a compiled code safepoint that was not originally a call 61 // or deoptimize following one of these kinds of safepoints. 62 63 // Lots of registers to save. For all builds, a window save will preserve 64 // the %i and %l registers. For the 32-bit longs-in-two entries and 64-bit 65 // builds a window-save will preserve the %o registers. In the LION build 66 // we need to save the 64-bit %o registers which requires we save them 67 // before the window-save (as then they become %i registers and get their 68 // heads chopped off on interrupt). We have to save some %g registers here 69 // as well. 70 enum { 71 // This frame's save area. Includes extra space for the native call: 72 // vararg's layout space and the like. Briefly holds the caller's 73 // register save area. 74 call_args_area = frame::register_save_words_sp_offset + 75 frame::memory_parameter_word_sp_offset*wordSize, 76 // Make sure save locations are always 8 byte aligned. 77 // can't use round_to because it doesn't produce compile time constant 78 start_of_extra_save_area = ((call_args_area + 7) & ~7), 79 g1_offset = start_of_extra_save_area, // g-regs needing saving 80 g3_offset = g1_offset+8, 81 g4_offset = g3_offset+8, 82 g5_offset = g4_offset+8, 83 o0_offset = g5_offset+8, 84 o1_offset = o0_offset+8, 85 o2_offset = o1_offset+8, 86 o3_offset = o2_offset+8, 87 o4_offset = o3_offset+8, 88 o5_offset = o4_offset+8, 89 start_of_flags_save_area = o5_offset+8, 90 ccr_offset = start_of_flags_save_area, 91 fsr_offset = ccr_offset + 8, 92 d00_offset = fsr_offset+8, // Start of float save area 93 register_save_size = d00_offset+8*32 94 }; 95 96 97 public: 98 99 static int Oexception_offset() { return o0_offset; }; 100 static int G3_offset() { return g3_offset; }; 101 static int G5_offset() { return g5_offset; }; 102 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words); 103 static void restore_live_registers(MacroAssembler* masm); 104 105 // During deoptimization only the result register need to be restored 106 // all the other values have already been extracted. 107 108 static void restore_result_registers(MacroAssembler* masm); 109 }; 110 111 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) { 112 // Record volatile registers as callee-save values in an OopMap so their save locations will be 113 // propagated to the caller frame's RegisterMap during StackFrameStream construction (needed for 114 // deoptimization; see compiledVFrame::create_stack_value). The caller's I, L and O registers 115 // are saved in register windows - I's and L's in the caller's frame and O's in the stub frame 116 // (as the stub's I's) when the runtime routine called by the stub creates its frame. 117 int i; 118 // Always make the frame size 16 byte aligned. 119 int frame_size = round_to(additional_frame_words + register_save_size, 16); 120 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words 121 int frame_size_in_slots = frame_size / sizeof(jint); 122 // CodeBlob frame size is in words. 123 *total_frame_words = frame_size / wordSize; 124 // OopMap* map = new OopMap(*total_frame_words, 0); 125 OopMap* map = new OopMap(frame_size_in_slots, 0); 126 127 #if !defined(_LP64) 128 129 // Save 64-bit O registers; they will get their heads chopped off on a 'save'. 130 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8); 131 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8); 132 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8); 133 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8); 134 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8); 135 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8); 136 #endif /* _LP64 */ 137 138 __ save(SP, -frame_size, SP); 139 140 #ifndef _LP64 141 // Reload the 64 bit Oregs. Although they are now Iregs we load them 142 // to Oregs here to avoid interrupts cutting off their heads 143 144 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0); 145 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1); 146 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2); 147 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3); 148 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4); 149 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5); 150 151 __ stx(O0, SP, o0_offset+STACK_BIAS); 152 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset + 4)>>2), O0->as_VMReg()); 153 154 __ stx(O1, SP, o1_offset+STACK_BIAS); 155 156 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset + 4)>>2), O1->as_VMReg()); 157 158 __ stx(O2, SP, o2_offset+STACK_BIAS); 159 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset + 4)>>2), O2->as_VMReg()); 160 161 __ stx(O3, SP, o3_offset+STACK_BIAS); 162 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset + 4)>>2), O3->as_VMReg()); 163 164 __ stx(O4, SP, o4_offset+STACK_BIAS); 165 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset + 4)>>2), O4->as_VMReg()); 166 167 __ stx(O5, SP, o5_offset+STACK_BIAS); 168 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset + 4)>>2), O5->as_VMReg()); 169 #endif /* _LP64 */ 170 171 172 #ifdef _LP64 173 int debug_offset = 0; 174 #else 175 int debug_offset = 4; 176 #endif 177 // Save the G's 178 __ stx(G1, SP, g1_offset+STACK_BIAS); 179 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset + debug_offset)>>2), G1->as_VMReg()); 180 181 __ stx(G3, SP, g3_offset+STACK_BIAS); 182 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset + debug_offset)>>2), G3->as_VMReg()); 183 184 __ stx(G4, SP, g4_offset+STACK_BIAS); 185 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset + debug_offset)>>2), G4->as_VMReg()); 186 187 __ stx(G5, SP, g5_offset+STACK_BIAS); 188 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset + debug_offset)>>2), G5->as_VMReg()); 189 190 // This is really a waste but we'll keep things as they were for now 191 if (true) { 192 #ifndef _LP64 193 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset)>>2), O0->as_VMReg()->next()); 194 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset)>>2), O1->as_VMReg()->next()); 195 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset)>>2), O2->as_VMReg()->next()); 196 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset)>>2), O3->as_VMReg()->next()); 197 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset)>>2), O4->as_VMReg()->next()); 198 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset)>>2), O5->as_VMReg()->next()); 199 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset)>>2), G1->as_VMReg()->next()); 200 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset)>>2), G3->as_VMReg()->next()); 201 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset)>>2), G4->as_VMReg()->next()); 202 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset)>>2), G5->as_VMReg()->next()); 203 #endif /* _LP64 */ 204 } 205 206 207 // Save the flags 208 __ rdccr( G5 ); 209 __ stx(G5, SP, ccr_offset+STACK_BIAS); 210 __ stxfsr(SP, fsr_offset+STACK_BIAS); 211 212 // Save all the FP registers: 32 doubles (32 floats correspond to the 2 halves of the first 16 doubles) 213 int offset = d00_offset; 214 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) { 215 FloatRegister f = as_FloatRegister(i); 216 __ stf(FloatRegisterImpl::D, f, SP, offset+STACK_BIAS); 217 // Record as callee saved both halves of double registers (2 float registers). 218 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), f->as_VMReg()); 219 map->set_callee_saved(VMRegImpl::stack2reg((offset + sizeof(float))>>2), f->as_VMReg()->next()); 220 offset += sizeof(double); 221 } 222 223 // And we're done. 224 225 return map; 226 } 227 228 229 // Pop the current frame and restore all the registers that we 230 // saved. 231 void RegisterSaver::restore_live_registers(MacroAssembler* masm) { 232 233 // Restore all the FP registers 234 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) { 235 __ ldf(FloatRegisterImpl::D, SP, d00_offset+i*sizeof(float)+STACK_BIAS, as_FloatRegister(i)); 236 } 237 238 __ ldx(SP, ccr_offset+STACK_BIAS, G1); 239 __ wrccr (G1) ; 240 241 // Restore the G's 242 // Note that G2 (AKA GThread) must be saved and restored separately. 243 // TODO-FIXME: save and restore some of the other ASRs, viz., %asi and %gsr. 244 245 __ ldx(SP, g1_offset+STACK_BIAS, G1); 246 __ ldx(SP, g3_offset+STACK_BIAS, G3); 247 __ ldx(SP, g4_offset+STACK_BIAS, G4); 248 __ ldx(SP, g5_offset+STACK_BIAS, G5); 249 250 251 #if !defined(_LP64) 252 // Restore the 64-bit O's. 253 __ ldx(SP, o0_offset+STACK_BIAS, O0); 254 __ ldx(SP, o1_offset+STACK_BIAS, O1); 255 __ ldx(SP, o2_offset+STACK_BIAS, O2); 256 __ ldx(SP, o3_offset+STACK_BIAS, O3); 257 __ ldx(SP, o4_offset+STACK_BIAS, O4); 258 __ ldx(SP, o5_offset+STACK_BIAS, O5); 259 260 // And temporarily place them in TLS 261 262 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8); 263 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8); 264 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8); 265 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8); 266 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8); 267 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8); 268 #endif /* _LP64 */ 269 270 // Restore flags 271 272 __ ldxfsr(SP, fsr_offset+STACK_BIAS); 273 274 __ restore(); 275 276 #if !defined(_LP64) 277 // Now reload the 64bit Oregs after we've restore the window. 278 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0); 279 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1); 280 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2); 281 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3); 282 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4); 283 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5); 284 #endif /* _LP64 */ 285 286 } 287 288 // Pop the current frame and restore the registers that might be holding 289 // a result. 290 void RegisterSaver::restore_result_registers(MacroAssembler* masm) { 291 292 #if !defined(_LP64) 293 // 32bit build returns longs in G1 294 __ ldx(SP, g1_offset+STACK_BIAS, G1); 295 296 // Retrieve the 64-bit O's. 297 __ ldx(SP, o0_offset+STACK_BIAS, O0); 298 __ ldx(SP, o1_offset+STACK_BIAS, O1); 299 // and save to TLS 300 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8); 301 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8); 302 #endif /* _LP64 */ 303 304 __ ldf(FloatRegisterImpl::D, SP, d00_offset+STACK_BIAS, as_FloatRegister(0)); 305 306 __ restore(); 307 308 #if !defined(_LP64) 309 // Now reload the 64bit Oregs after we've restore the window. 310 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0); 311 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1); 312 #endif /* _LP64 */ 313 314 } 315 316 // The java_calling_convention describes stack locations as ideal slots on 317 // a frame with no abi restrictions. Since we must observe abi restrictions 318 // (like the placement of the register window) the slots must be biased by 319 // the following value. 320 static int reg2offset(VMReg r) { 321 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; 322 } 323 324 // --------------------------------------------------------------------------- 325 // Read the array of BasicTypes from a signature, and compute where the 326 // arguments should go. Values in the VMRegPair regs array refer to 4-byte (VMRegImpl::stack_slot_size) 327 // quantities. Values less than VMRegImpl::stack0 are registers, those above 328 // refer to 4-byte stack slots. All stack slots are based off of the window 329 // top. VMRegImpl::stack0 refers to the first slot past the 16-word window, 330 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register 331 // values 0-63 (up to RegisterImpl::number_of_registers) are the 64-bit 332 // integer registers. Values 64-95 are the (32-bit only) float registers. 333 // Each 32-bit quantity is given its own number, so the integer registers 334 // (in either 32- or 64-bit builds) use 2 numbers. For example, there is 335 // an O0-low and an O0-high. Essentially, all int register numbers are doubled. 336 337 // Register results are passed in O0-O5, for outgoing call arguments. To 338 // convert to incoming arguments, convert all O's to I's. The regs array 339 // refer to the low and hi 32-bit words of 64-bit registers or stack slots. 340 // If the regs[].second() field is set to VMRegImpl::Bad(), it means it's unused (a 341 // 32-bit value was passed). If both are VMRegImpl::Bad(), it means no value was 342 // passed (used as a placeholder for the other half of longs and doubles in 343 // the 64-bit build). regs[].second() is either VMRegImpl::Bad() or regs[].second() is 344 // regs[].first()+1 (regs[].first() may be misaligned in the C calling convention). 345 // Sparc never passes a value in regs[].second() but not regs[].first() (regs[].first() 346 // == VMRegImpl::Bad() && regs[].second() != VMRegImpl::Bad()) nor unrelated values in the 347 // same VMRegPair. 348 349 // Note: the INPUTS in sig_bt are in units of Java argument words, which are 350 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit 351 // units regardless of build. 352 353 354 // --------------------------------------------------------------------------- 355 // The compiled Java calling convention. The Java convention always passes 356 // 64-bit values in adjacent aligned locations (either registers or stack), 357 // floats in float registers and doubles in aligned float pairs. Values are 358 // packed in the registers. There is no backing varargs store for values in 359 // registers. In the 32-bit build, longs are passed in G1 and G4 (cannot be 360 // passed in I's, because longs in I's get their heads chopped off at 361 // interrupt). 362 int SharedRuntime::java_calling_convention(const BasicType *sig_bt, 363 VMRegPair *regs, 364 int total_args_passed, 365 int is_outgoing) { 366 assert(F31->as_VMReg()->is_reg(), "overlapping stack/register numbers"); 367 368 // Convention is to pack the first 6 int/oop args into the first 6 registers 369 // (I0-I5), extras spill to the stack. Then pack the first 8 float args 370 // into F0-F7, extras spill to the stack. Then pad all register sets to 371 // align. Then put longs and doubles into the same registers as they fit, 372 // else spill to the stack. 373 const int int_reg_max = SPARC_ARGS_IN_REGS_NUM; 374 const int flt_reg_max = 8; 375 // 376 // Where 32-bit 1-reg longs start being passed 377 // In tiered we must pass on stack because c1 can't use a "pair" in a single reg. 378 // So make it look like we've filled all the G regs that c2 wants to use. 379 Register g_reg = TieredCompilation ? noreg : G1; 380 381 // Count int/oop and float args. See how many stack slots we'll need and 382 // where the longs & doubles will go. 383 int int_reg_cnt = 0; 384 int flt_reg_cnt = 0; 385 // int stk_reg_pairs = frame::register_save_words*(wordSize>>2); 386 // int stk_reg_pairs = SharedRuntime::out_preserve_stack_slots(); 387 int stk_reg_pairs = 0; 388 for (int i = 0; i < total_args_passed; i++) { 389 switch (sig_bt[i]) { 390 case T_LONG: // LP64, longs compete with int args 391 assert(sig_bt[i+1] == T_VOID, ""); 392 #ifdef _LP64 393 if (int_reg_cnt < int_reg_max) int_reg_cnt++; 394 #endif 395 break; 396 case T_OBJECT: 397 case T_ARRAY: 398 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address 399 if (int_reg_cnt < int_reg_max) int_reg_cnt++; 400 #ifndef _LP64 401 else stk_reg_pairs++; 402 #endif 403 break; 404 case T_INT: 405 case T_SHORT: 406 case T_CHAR: 407 case T_BYTE: 408 case T_BOOLEAN: 409 if (int_reg_cnt < int_reg_max) int_reg_cnt++; 410 else stk_reg_pairs++; 411 break; 412 case T_FLOAT: 413 if (flt_reg_cnt < flt_reg_max) flt_reg_cnt++; 414 else stk_reg_pairs++; 415 break; 416 case T_DOUBLE: 417 assert(sig_bt[i+1] == T_VOID, ""); 418 break; 419 case T_VOID: 420 break; 421 default: 422 ShouldNotReachHere(); 423 } 424 } 425 426 // This is where the longs/doubles start on the stack. 427 stk_reg_pairs = (stk_reg_pairs+1) & ~1; // Round 428 429 int int_reg_pairs = (int_reg_cnt+1) & ~1; // 32-bit 2-reg longs only 430 int flt_reg_pairs = (flt_reg_cnt+1) & ~1; 431 432 // int stk_reg = frame::register_save_words*(wordSize>>2); 433 // int stk_reg = SharedRuntime::out_preserve_stack_slots(); 434 int stk_reg = 0; 435 int int_reg = 0; 436 int flt_reg = 0; 437 438 // Now do the signature layout 439 for (int i = 0; i < total_args_passed; i++) { 440 switch (sig_bt[i]) { 441 case T_INT: 442 case T_SHORT: 443 case T_CHAR: 444 case T_BYTE: 445 case T_BOOLEAN: 446 #ifndef _LP64 447 case T_OBJECT: 448 case T_ARRAY: 449 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address 450 #endif // _LP64 451 if (int_reg < int_reg_max) { 452 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++); 453 regs[i].set1(r->as_VMReg()); 454 } else { 455 regs[i].set1(VMRegImpl::stack2reg(stk_reg++)); 456 } 457 break; 458 459 #ifdef _LP64 460 case T_OBJECT: 461 case T_ARRAY: 462 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address 463 if (int_reg < int_reg_max) { 464 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++); 465 regs[i].set2(r->as_VMReg()); 466 } else { 467 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); 468 stk_reg_pairs += 2; 469 } 470 break; 471 #endif // _LP64 472 473 case T_LONG: 474 assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half"); 475 #ifdef _LP64 476 if (int_reg < int_reg_max) { 477 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++); 478 regs[i].set2(r->as_VMReg()); 479 } else { 480 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); 481 stk_reg_pairs += 2; 482 } 483 #else 484 #ifdef COMPILER2 485 // For 32-bit build, can't pass longs in O-regs because they become 486 // I-regs and get trashed. Use G-regs instead. G1 and G4 are almost 487 // spare and available. This convention isn't used by the Sparc ABI or 488 // anywhere else. If we're tiered then we don't use G-regs because c1 489 // can't deal with them as a "pair". (Tiered makes this code think g's are filled) 490 // G0: zero 491 // G1: 1st Long arg 492 // G2: global allocated to TLS 493 // G3: used in inline cache check 494 // G4: 2nd Long arg 495 // G5: used in inline cache check 496 // G6: used by OS 497 // G7: used by OS 498 499 if (g_reg == G1) { 500 regs[i].set2(G1->as_VMReg()); // This long arg in G1 501 g_reg = G4; // Where the next arg goes 502 } else if (g_reg == G4) { 503 regs[i].set2(G4->as_VMReg()); // The 2nd long arg in G4 504 g_reg = noreg; // No more longs in registers 505 } else { 506 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); 507 stk_reg_pairs += 2; 508 } 509 #else // COMPILER2 510 if (int_reg_pairs + 1 < int_reg_max) { 511 if (is_outgoing) { 512 regs[i].set_pair(as_oRegister(int_reg_pairs + 1)->as_VMReg(), as_oRegister(int_reg_pairs)->as_VMReg()); 513 } else { 514 regs[i].set_pair(as_iRegister(int_reg_pairs + 1)->as_VMReg(), as_iRegister(int_reg_pairs)->as_VMReg()); 515 } 516 int_reg_pairs += 2; 517 } else { 518 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); 519 stk_reg_pairs += 2; 520 } 521 #endif // COMPILER2 522 #endif // _LP64 523 break; 524 525 case T_FLOAT: 526 if (flt_reg < flt_reg_max) regs[i].set1(as_FloatRegister(flt_reg++)->as_VMReg()); 527 else regs[i].set1( VMRegImpl::stack2reg(stk_reg++)); 528 break; 529 case T_DOUBLE: 530 assert(sig_bt[i+1] == T_VOID, "expecting half"); 531 if (flt_reg_pairs + 1 < flt_reg_max) { 532 regs[i].set2(as_FloatRegister(flt_reg_pairs)->as_VMReg()); 533 flt_reg_pairs += 2; 534 } else { 535 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); 536 stk_reg_pairs += 2; 537 } 538 break; 539 case T_VOID: regs[i].set_bad(); break; // Halves of longs & doubles 540 default: 541 ShouldNotReachHere(); 542 } 543 } 544 545 // retun the amount of stack space these arguments will need. 546 return stk_reg_pairs; 547 548 } 549 550 // Helper class mostly to avoid passing masm everywhere, and handle 551 // store displacement overflow logic. 552 class AdapterGenerator { 553 MacroAssembler *masm; 554 Register Rdisp; 555 void set_Rdisp(Register r) { Rdisp = r; } 556 557 void patch_callers_callsite(); 558 559 // base+st_off points to top of argument 560 int arg_offset(const int st_off) { return st_off; } 561 int next_arg_offset(const int st_off) { 562 return st_off - Interpreter::stackElementSize; 563 } 564 565 // Argument slot values may be loaded first into a register because 566 // they might not fit into displacement. 567 RegisterOrConstant arg_slot(const int st_off); 568 RegisterOrConstant next_arg_slot(const int st_off); 569 570 // Stores long into offset pointed to by base 571 void store_c2i_long(Register r, Register base, 572 const int st_off, bool is_stack); 573 void store_c2i_object(Register r, Register base, 574 const int st_off); 575 void store_c2i_int(Register r, Register base, 576 const int st_off); 577 void store_c2i_double(VMReg r_2, 578 VMReg r_1, Register base, const int st_off); 579 void store_c2i_float(FloatRegister f, Register base, 580 const int st_off); 581 582 public: 583 void gen_c2i_adapter(int total_args_passed, 584 // VMReg max_arg, 585 int comp_args_on_stack, // VMRegStackSlots 586 const BasicType *sig_bt, 587 const VMRegPair *regs, 588 Label& skip_fixup); 589 void gen_i2c_adapter(int total_args_passed, 590 // VMReg max_arg, 591 int comp_args_on_stack, // VMRegStackSlots 592 const BasicType *sig_bt, 593 const VMRegPair *regs); 594 595 AdapterGenerator(MacroAssembler *_masm) : masm(_masm) {} 596 }; 597 598 599 // Patch the callers callsite with entry to compiled code if it exists. 600 void AdapterGenerator::patch_callers_callsite() { 601 Label L; 602 __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch); 603 __ br_null(G3_scratch, false, Assembler::pt, L); 604 // Schedule the branch target address early. 605 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch); 606 // Call into the VM to patch the caller, then jump to compiled callee 607 __ save_frame(4); // Args in compiled layout; do not blow them 608 609 // Must save all the live Gregs the list is: 610 // G1: 1st Long arg (32bit build) 611 // G2: global allocated to TLS 612 // G3: used in inline cache check (scratch) 613 // G4: 2nd Long arg (32bit build); 614 // G5: used in inline cache check (methodOop) 615 616 // The longs must go to the stack by hand since in the 32 bit build they can be trashed by window ops. 617 618 #ifdef _LP64 619 // mov(s,d) 620 __ mov(G1, L1); 621 __ mov(G4, L4); 622 __ mov(G5_method, L5); 623 __ mov(G5_method, O0); // VM needs target method 624 __ mov(I7, O1); // VM needs caller's callsite 625 // Must be a leaf call... 626 // can be very far once the blob has been relocated 627 AddressLiteral dest(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)); 628 __ relocate(relocInfo::runtime_call_type); 629 __ jumpl_to(dest, O7, O7); 630 __ delayed()->mov(G2_thread, L7_thread_cache); 631 __ mov(L7_thread_cache, G2_thread); 632 __ mov(L1, G1); 633 __ mov(L4, G4); 634 __ mov(L5, G5_method); 635 #else 636 __ stx(G1, FP, -8 + STACK_BIAS); 637 __ stx(G4, FP, -16 + STACK_BIAS); 638 __ mov(G5_method, L5); 639 __ mov(G5_method, O0); // VM needs target method 640 __ mov(I7, O1); // VM needs caller's callsite 641 // Must be a leaf call... 642 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), relocInfo::runtime_call_type); 643 __ delayed()->mov(G2_thread, L7_thread_cache); 644 __ mov(L7_thread_cache, G2_thread); 645 __ ldx(FP, -8 + STACK_BIAS, G1); 646 __ ldx(FP, -16 + STACK_BIAS, G4); 647 __ mov(L5, G5_method); 648 __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch); 649 #endif /* _LP64 */ 650 651 __ restore(); // Restore args 652 __ bind(L); 653 } 654 655 656 RegisterOrConstant AdapterGenerator::arg_slot(const int st_off) { 657 RegisterOrConstant roc(arg_offset(st_off)); 658 return __ ensure_simm13_or_reg(roc, Rdisp); 659 } 660 661 RegisterOrConstant AdapterGenerator::next_arg_slot(const int st_off) { 662 RegisterOrConstant roc(next_arg_offset(st_off)); 663 return __ ensure_simm13_or_reg(roc, Rdisp); 664 } 665 666 667 // Stores long into offset pointed to by base 668 void AdapterGenerator::store_c2i_long(Register r, Register base, 669 const int st_off, bool is_stack) { 670 #ifdef _LP64 671 // In V9, longs are given 2 64-bit slots in the interpreter, but the 672 // data is passed in only 1 slot. 673 __ stx(r, base, next_arg_slot(st_off)); 674 #else 675 #ifdef COMPILER2 676 // Misaligned store of 64-bit data 677 __ stw(r, base, arg_slot(st_off)); // lo bits 678 __ srlx(r, 32, r); 679 __ stw(r, base, next_arg_slot(st_off)); // hi bits 680 #else 681 if (is_stack) { 682 // Misaligned store of 64-bit data 683 __ stw(r, base, arg_slot(st_off)); // lo bits 684 __ srlx(r, 32, r); 685 __ stw(r, base, next_arg_slot(st_off)); // hi bits 686 } else { 687 __ stw(r->successor(), base, arg_slot(st_off) ); // lo bits 688 __ stw(r , base, next_arg_slot(st_off)); // hi bits 689 } 690 #endif // COMPILER2 691 #endif // _LP64 692 } 693 694 void AdapterGenerator::store_c2i_object(Register r, Register base, 695 const int st_off) { 696 __ st_ptr (r, base, arg_slot(st_off)); 697 } 698 699 void AdapterGenerator::store_c2i_int(Register r, Register base, 700 const int st_off) { 701 __ st (r, base, arg_slot(st_off)); 702 } 703 704 // Stores into offset pointed to by base 705 void AdapterGenerator::store_c2i_double(VMReg r_2, 706 VMReg r_1, Register base, const int st_off) { 707 #ifdef _LP64 708 // In V9, doubles are given 2 64-bit slots in the interpreter, but the 709 // data is passed in only 1 slot. 710 __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), base, next_arg_slot(st_off)); 711 #else 712 // Need to marshal 64-bit value from misaligned Lesp loads 713 __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), base, next_arg_slot(st_off)); 714 __ stf(FloatRegisterImpl::S, r_2->as_FloatRegister(), base, arg_slot(st_off) ); 715 #endif 716 } 717 718 void AdapterGenerator::store_c2i_float(FloatRegister f, Register base, 719 const int st_off) { 720 __ stf(FloatRegisterImpl::S, f, base, arg_slot(st_off)); 721 } 722 723 void AdapterGenerator::gen_c2i_adapter( 724 int total_args_passed, 725 // VMReg max_arg, 726 int comp_args_on_stack, // VMRegStackSlots 727 const BasicType *sig_bt, 728 const VMRegPair *regs, 729 Label& skip_fixup) { 730 731 // Before we get into the guts of the C2I adapter, see if we should be here 732 // at all. We've come from compiled code and are attempting to jump to the 733 // interpreter, which means the caller made a static call to get here 734 // (vcalls always get a compiled target if there is one). Check for a 735 // compiled target. If there is one, we need to patch the caller's call. 736 // However we will run interpreted if we come thru here. The next pass 737 // thru the call site will run compiled. If we ran compiled here then 738 // we can (theorectically) do endless i2c->c2i->i2c transitions during 739 // deopt/uncommon trap cycles. If we always go interpreted here then 740 // we can have at most one and don't need to play any tricks to keep 741 // from endlessly growing the stack. 742 // 743 // Actually if we detected that we had an i2c->c2i transition here we 744 // ought to be able to reset the world back to the state of the interpreted 745 // call and not bother building another interpreter arg area. We don't 746 // do that at this point. 747 748 patch_callers_callsite(); 749 750 __ bind(skip_fixup); 751 752 // Since all args are passed on the stack, total_args_passed*wordSize is the 753 // space we need. Add in varargs area needed by the interpreter. Round up 754 // to stack alignment. 755 const int arg_size = total_args_passed * Interpreter::stackElementSize; 756 const int varargs_area = 757 (frame::varargs_offset - frame::register_save_words)*wordSize; 758 const int extraspace = round_to(arg_size + varargs_area, 2*wordSize); 759 760 int bias = STACK_BIAS; 761 const int interp_arg_offset = frame::varargs_offset*wordSize + 762 (total_args_passed-1)*Interpreter::stackElementSize; 763 764 Register base = SP; 765 766 #ifdef _LP64 767 // In the 64bit build because of wider slots and STACKBIAS we can run 768 // out of bits in the displacement to do loads and stores. Use g3 as 769 // temporary displacement. 770 if (! __ is_simm13(extraspace)) { 771 __ set(extraspace, G3_scratch); 772 __ sub(SP, G3_scratch, SP); 773 } else { 774 __ sub(SP, extraspace, SP); 775 } 776 set_Rdisp(G3_scratch); 777 #else 778 __ sub(SP, extraspace, SP); 779 #endif // _LP64 780 781 // First write G1 (if used) to where ever it must go 782 for (int i=0; i<total_args_passed; i++) { 783 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize) + bias; 784 VMReg r_1 = regs[i].first(); 785 VMReg r_2 = regs[i].second(); 786 if (r_1 == G1_scratch->as_VMReg()) { 787 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) { 788 store_c2i_object(G1_scratch, base, st_off); 789 } else if (sig_bt[i] == T_LONG) { 790 assert(!TieredCompilation, "should not use register args for longs"); 791 store_c2i_long(G1_scratch, base, st_off, false); 792 } else { 793 store_c2i_int(G1_scratch, base, st_off); 794 } 795 } 796 } 797 798 // Now write the args into the outgoing interpreter space 799 for (int i=0; i<total_args_passed; i++) { 800 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize) + bias; 801 VMReg r_1 = regs[i].first(); 802 VMReg r_2 = regs[i].second(); 803 if (!r_1->is_valid()) { 804 assert(!r_2->is_valid(), ""); 805 continue; 806 } 807 // Skip G1 if found as we did it first in order to free it up 808 if (r_1 == G1_scratch->as_VMReg()) { 809 continue; 810 } 811 #ifdef ASSERT 812 bool G1_forced = false; 813 #endif // ASSERT 814 if (r_1->is_stack()) { // Pretend stack targets are loaded into G1 815 #ifdef _LP64 816 Register ld_off = Rdisp; 817 __ set(reg2offset(r_1) + extraspace + bias, ld_off); 818 #else 819 int ld_off = reg2offset(r_1) + extraspace + bias; 820 #endif // _LP64 821 #ifdef ASSERT 822 G1_forced = true; 823 #endif // ASSERT 824 r_1 = G1_scratch->as_VMReg();// as part of the load/store shuffle 825 if (!r_2->is_valid()) __ ld (base, ld_off, G1_scratch); 826 else __ ldx(base, ld_off, G1_scratch); 827 } 828 829 if (r_1->is_Register()) { 830 Register r = r_1->as_Register()->after_restore(); 831 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) { 832 store_c2i_object(r, base, st_off); 833 } else if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { 834 #ifndef _LP64 835 if (TieredCompilation) { 836 assert(G1_forced || sig_bt[i] != T_LONG, "should not use register args for longs"); 837 } 838 #endif // _LP64 839 store_c2i_long(r, base, st_off, r_2->is_stack()); 840 } else { 841 store_c2i_int(r, base, st_off); 842 } 843 } else { 844 assert(r_1->is_FloatRegister(), ""); 845 if (sig_bt[i] == T_FLOAT) { 846 store_c2i_float(r_1->as_FloatRegister(), base, st_off); 847 } else { 848 assert(sig_bt[i] == T_DOUBLE, "wrong type"); 849 store_c2i_double(r_2, r_1, base, st_off); 850 } 851 } 852 } 853 854 #ifdef _LP64 855 // Need to reload G3_scratch, used for temporary displacements. 856 __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch); 857 858 // Pass O5_savedSP as an argument to the interpreter. 859 // The interpreter will restore SP to this value before returning. 860 __ set(extraspace, G1); 861 __ add(SP, G1, O5_savedSP); 862 #else 863 // Pass O5_savedSP as an argument to the interpreter. 864 // The interpreter will restore SP to this value before returning. 865 __ add(SP, extraspace, O5_savedSP); 866 #endif // _LP64 867 868 __ mov((frame::varargs_offset)*wordSize - 869 1*Interpreter::stackElementSize+bias+BytesPerWord, G1); 870 // Jump to the interpreter just as if interpreter was doing it. 871 __ jmpl(G3_scratch, 0, G0); 872 // Setup Lesp for the call. Cannot actually set Lesp as the current Lesp 873 // (really L0) is in use by the compiled frame as a generic temp. However, 874 // the interpreter does not know where its args are without some kind of 875 // arg pointer being passed in. Pass it in Gargs. 876 __ delayed()->add(SP, G1, Gargs); 877 } 878 879 void AdapterGenerator::gen_i2c_adapter( 880 int total_args_passed, 881 // VMReg max_arg, 882 int comp_args_on_stack, // VMRegStackSlots 883 const BasicType *sig_bt, 884 const VMRegPair *regs) { 885 886 // Generate an I2C adapter: adjust the I-frame to make space for the C-frame 887 // layout. Lesp was saved by the calling I-frame and will be restored on 888 // return. Meanwhile, outgoing arg space is all owned by the callee 889 // C-frame, so we can mangle it at will. After adjusting the frame size, 890 // hoist register arguments and repack other args according to the compiled 891 // code convention. Finally, end in a jump to the compiled code. The entry 892 // point address is the start of the buffer. 893 894 // We will only enter here from an interpreted frame and never from after 895 // passing thru a c2i. Azul allowed this but we do not. If we lose the 896 // race and use a c2i we will remain interpreted for the race loser(s). 897 // This removes all sorts of headaches on the x86 side and also eliminates 898 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions. 899 900 // As you can see from the list of inputs & outputs there are not a lot 901 // of temp registers to work with: mostly G1, G3 & G4. 902 903 // Inputs: 904 // G2_thread - TLS 905 // G5_method - Method oop 906 // G4 (Gargs) - Pointer to interpreter's args 907 // O0..O4 - free for scratch 908 // O5_savedSP - Caller's saved SP, to be restored if needed 909 // O6 - Current SP! 910 // O7 - Valid return address 911 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet) 912 913 // Outputs: 914 // G2_thread - TLS 915 // G1, G4 - Outgoing long args in 32-bit build 916 // O0-O5 - Outgoing args in compiled layout 917 // O6 - Adjusted or restored SP 918 // O7 - Valid return address 919 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet) 920 // F0-F7 - more outgoing args 921 922 923 // Gargs is the incoming argument base, and also an outgoing argument. 924 __ sub(Gargs, BytesPerWord, Gargs); 925 926 // ON ENTRY TO THE CODE WE ARE MAKING, WE HAVE AN INTERPRETED FRAME 927 // WITH O7 HOLDING A VALID RETURN PC 928 // 929 // | | 930 // : java stack : 931 // | | 932 // +--------------+ <--- start of outgoing args 933 // | receiver | | 934 // : rest of args : |---size is java-arg-words 935 // | | | 936 // +--------------+ <--- O4_args (misaligned) and Lesp if prior is not C2I 937 // | | | 938 // : unused : |---Space for max Java stack, plus stack alignment 939 // | | | 940 // +--------------+ <--- SP + 16*wordsize 941 // | | 942 // : window : 943 // | | 944 // +--------------+ <--- SP 945 946 // WE REPACK THE STACK. We use the common calling convention layout as 947 // discovered by calling SharedRuntime::calling_convention. We assume it 948 // causes an arbitrary shuffle of memory, which may require some register 949 // temps to do the shuffle. We hope for (and optimize for) the case where 950 // temps are not needed. We may have to resize the stack slightly, in case 951 // we need alignment padding (32-bit interpreter can pass longs & doubles 952 // misaligned, but the compilers expect them aligned). 953 // 954 // | | 955 // : java stack : 956 // | | 957 // +--------------+ <--- start of outgoing args 958 // | pad, align | | 959 // +--------------+ | 960 // | ints, floats | |---Outgoing stack args, packed low. 961 // +--------------+ | First few args in registers. 962 // : doubles : | 963 // | longs | | 964 // +--------------+ <--- SP' + 16*wordsize 965 // | | 966 // : window : 967 // | | 968 // +--------------+ <--- SP' 969 970 // ON EXIT FROM THE CODE WE ARE MAKING, WE STILL HAVE AN INTERPRETED FRAME 971 // WITH O7 HOLDING A VALID RETURN PC - ITS JUST THAT THE ARGS ARE NOW SETUP 972 // FOR COMPILED CODE AND THE FRAME SLIGHTLY GROWN. 973 974 // Cut-out for having no stack args. Since up to 6 args are passed 975 // in registers, we will commonly have no stack args. 976 if (comp_args_on_stack > 0) { 977 978 // Convert VMReg stack slots to words. 979 int 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 // Now compute the distance from Lesp to SP. This calculation does not 983 // include the space for total_args_passed because Lesp has not yet popped 984 // the arguments. 985 __ sub(SP, (comp_words_on_stack)*wordSize, SP); 986 } 987 988 // Will jump to the compiled code just as if compiled code was doing it. 989 // Pre-load the register-jump target early, to schedule it better. 990 __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_compiled_offset()), G3); 991 992 // Now generate the shuffle code. Pick up all register args and move the 993 // rest through G1_scratch. 994 for (int i=0; i<total_args_passed; i++) { 995 if (sig_bt[i] == T_VOID) { 996 // Longs and doubles are passed in native word order, but misaligned 997 // in the 32-bit build. 998 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); 999 continue; 1000 } 1001 1002 // Pick up 0, 1 or 2 words from Lesp+offset. Assume mis-aligned in the 1003 // 32-bit build and aligned in the 64-bit build. Look for the obvious 1004 // ldx/lddf optimizations. 1005 1006 // Load in argument order going down. 1007 const int ld_off = (total_args_passed-i)*Interpreter::stackElementSize; 1008 set_Rdisp(G1_scratch); 1009 1010 VMReg r_1 = regs[i].first(); 1011 VMReg r_2 = regs[i].second(); 1012 if (!r_1->is_valid()) { 1013 assert(!r_2->is_valid(), ""); 1014 continue; 1015 } 1016 if (r_1->is_stack()) { // Pretend stack targets are loaded into F8/F9 1017 r_1 = F8->as_VMReg(); // as part of the load/store shuffle 1018 if (r_2->is_valid()) r_2 = r_1->next(); 1019 } 1020 if (r_1->is_Register()) { // Register argument 1021 Register r = r_1->as_Register()->after_restore(); 1022 if (!r_2->is_valid()) { 1023 __ ld(Gargs, arg_slot(ld_off), r); 1024 } else { 1025 #ifdef _LP64 1026 // In V9, longs are given 2 64-bit slots in the interpreter, but the 1027 // data is passed in only 1 slot. 1028 RegisterOrConstant slot = (sig_bt[i] == T_LONG) ? 1029 next_arg_slot(ld_off) : arg_slot(ld_off); 1030 __ ldx(Gargs, slot, r); 1031 #else 1032 // Need to load a 64-bit value into G1/G4, but G1/G4 is being used in the 1033 // stack shuffle. Load the first 2 longs into G1/G4 later. 1034 #endif 1035 } 1036 } else { 1037 assert(r_1->is_FloatRegister(), ""); 1038 if (!r_2->is_valid()) { 1039 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_1->as_FloatRegister()); 1040 } else { 1041 #ifdef _LP64 1042 // In V9, doubles are given 2 64-bit slots in the interpreter, but the 1043 // data is passed in only 1 slot. This code also handles longs that 1044 // are passed on the stack, but need a stack-to-stack move through a 1045 // spare float register. 1046 RegisterOrConstant slot = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ? 1047 next_arg_slot(ld_off) : arg_slot(ld_off); 1048 __ ldf(FloatRegisterImpl::D, Gargs, slot, r_1->as_FloatRegister()); 1049 #else 1050 // Need to marshal 64-bit value from misaligned Lesp loads 1051 __ ldf(FloatRegisterImpl::S, Gargs, next_arg_slot(ld_off), r_1->as_FloatRegister()); 1052 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_2->as_FloatRegister()); 1053 #endif 1054 } 1055 } 1056 // Was the argument really intended to be on the stack, but was loaded 1057 // into F8/F9? 1058 if (regs[i].first()->is_stack()) { 1059 assert(r_1->as_FloatRegister() == F8, "fix this code"); 1060 // Convert stack slot to an SP offset 1061 int st_off = reg2offset(regs[i].first()) + STACK_BIAS; 1062 // Store down the shuffled stack word. Target address _is_ aligned. 1063 RegisterOrConstant slot = __ ensure_simm13_or_reg(st_off, Rdisp); 1064 if (!r_2->is_valid()) __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), SP, slot); 1065 else __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), SP, slot); 1066 } 1067 } 1068 bool made_space = false; 1069 #ifndef _LP64 1070 // May need to pick up a few long args in G1/G4 1071 bool g4_crushed = false; 1072 bool g3_crushed = false; 1073 for (int i=0; i<total_args_passed; i++) { 1074 if (regs[i].first()->is_Register() && regs[i].second()->is_valid()) { 1075 // Load in argument order going down 1076 int ld_off = (total_args_passed-i)*Interpreter::stackElementSize; 1077 // Need to marshal 64-bit value from misaligned Lesp loads 1078 Register r = regs[i].first()->as_Register()->after_restore(); 1079 if (r == G1 || r == G4) { 1080 assert(!g4_crushed, "ordering problem"); 1081 if (r == G4){ 1082 g4_crushed = true; 1083 __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits 1084 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits 1085 } else { 1086 // better schedule this way 1087 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits 1088 __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits 1089 } 1090 g3_crushed = true; 1091 __ sllx(r, 32, r); 1092 __ or3(G3_scratch, r, r); 1093 } else { 1094 assert(r->is_out(), "longs passed in two O registers"); 1095 __ ld (Gargs, arg_slot(ld_off) , r->successor()); // Load lo bits 1096 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits 1097 } 1098 } 1099 } 1100 #endif 1101 1102 // Jump to the compiled code just as if compiled code was doing it. 1103 // 1104 #ifndef _LP64 1105 if (g3_crushed) { 1106 // Rats load was wasted, at least it is in cache... 1107 __ ld_ptr(G5_method, methodOopDesc::from_compiled_offset(), G3); 1108 } 1109 #endif /* _LP64 */ 1110 1111 // 6243940 We might end up in handle_wrong_method if 1112 // the callee is deoptimized as we race thru here. If that 1113 // happens we don't want to take a safepoint because the 1114 // caller frame will look interpreted and arguments are now 1115 // "compiled" so it is much better to make this transition 1116 // invisible to the stack walking code. Unfortunately if 1117 // we try and find the callee by normal means a safepoint 1118 // is possible. So we stash the desired callee in the thread 1119 // and the vm will find there should this case occur. 1120 Address callee_target_addr(G2_thread, JavaThread::callee_target_offset()); 1121 __ st_ptr(G5_method, callee_target_addr); 1122 1123 if (StressNonEntrant) { 1124 // Open a big window for deopt failure 1125 __ save_frame(0); 1126 __ mov(G0, L0); 1127 Label loop; 1128 __ bind(loop); 1129 __ sub(L0, 1, L0); 1130 __ br_null_short(L0, Assembler::pt, loop); 1131 1132 __ restore(); 1133 } 1134 1135 1136 __ jmpl(G3, 0, G0); 1137 __ delayed()->nop(); 1138 } 1139 1140 // --------------------------------------------------------------- 1141 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, 1142 int total_args_passed, 1143 // VMReg max_arg, 1144 int comp_args_on_stack, // VMRegStackSlots 1145 const BasicType *sig_bt, 1146 const VMRegPair *regs, 1147 AdapterFingerPrint* fingerprint) { 1148 address i2c_entry = __ pc(); 1149 1150 AdapterGenerator agen(masm); 1151 1152 agen.gen_i2c_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs); 1153 1154 1155 // ------------------------------------------------------------------------- 1156 // Generate a C2I adapter. On entry we know G5 holds the methodOop. The 1157 // args start out packed in the compiled layout. They need to be unpacked 1158 // into the interpreter layout. This will almost always require some stack 1159 // space. We grow the current (compiled) stack, then repack the args. We 1160 // finally end in a jump to the generic interpreter entry point. On exit 1161 // from the interpreter, the interpreter will restore our SP (lest the 1162 // compiled code, which relys solely on SP and not FP, get sick). 1163 1164 address c2i_unverified_entry = __ pc(); 1165 Label skip_fixup; 1166 { 1167 #if !defined(_LP64) && defined(COMPILER2) 1168 Register R_temp = L0; // another scratch register 1169 #else 1170 Register R_temp = G1; // another scratch register 1171 #endif 1172 1173 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub()); 1174 1175 __ verify_oop(O0); 1176 __ verify_oop(G5_method); 1177 __ load_klass(O0, G3_scratch); 1178 __ verify_oop(G3_scratch); 1179 1180 #if !defined(_LP64) && defined(COMPILER2) 1181 __ save(SP, -frame::register_save_words*wordSize, SP); 1182 __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp); 1183 __ verify_oop(R_temp); 1184 __ cmp(G3_scratch, R_temp); 1185 __ restore(); 1186 #else 1187 __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp); 1188 __ verify_oop(R_temp); 1189 __ cmp(G3_scratch, R_temp); 1190 #endif 1191 1192 Label ok, ok2; 1193 __ brx(Assembler::equal, false, Assembler::pt, ok); 1194 __ delayed()->ld_ptr(G5_method, compiledICHolderOopDesc::holder_method_offset(), G5_method); 1195 __ jump_to(ic_miss, G3_scratch); 1196 __ delayed()->nop(); 1197 1198 __ bind(ok); 1199 // Method might have been compiled since the call site was patched to 1200 // interpreted if that is the case treat it as a miss so we can get 1201 // the call site corrected. 1202 __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch); 1203 __ bind(ok2); 1204 __ br_null(G3_scratch, false, Assembler::pt, skip_fixup); 1205 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch); 1206 __ jump_to(ic_miss, G3_scratch); 1207 __ delayed()->nop(); 1208 1209 } 1210 1211 address c2i_entry = __ pc(); 1212 1213 agen.gen_c2i_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); 1214 1215 __ flush(); 1216 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry); 1217 1218 } 1219 1220 // Helper function for native calling conventions 1221 static VMReg int_stk_helper( int i ) { 1222 // Bias any stack based VMReg we get by ignoring the window area 1223 // but not the register parameter save area. 1224 // 1225 // This is strange for the following reasons. We'd normally expect 1226 // the calling convention to return an VMReg for a stack slot 1227 // completely ignoring any abi reserved area. C2 thinks of that 1228 // abi area as only out_preserve_stack_slots. This does not include 1229 // the area allocated by the C abi to store down integer arguments 1230 // because the java calling convention does not use it. So 1231 // since c2 assumes that there are only out_preserve_stack_slots 1232 // to bias the optoregs (which impacts VMRegs) when actually referencing any actual stack 1233 // location the c calling convention must add in this bias amount 1234 // to make up for the fact that the out_preserve_stack_slots is 1235 // insufficient for C calls. What a mess. I sure hope those 6 1236 // stack words were worth it on every java call! 1237 1238 // Another way of cleaning this up would be for out_preserve_stack_slots 1239 // to take a parameter to say whether it was C or java calling conventions. 1240 // Then things might look a little better (but not much). 1241 1242 int mem_parm_offset = i - SPARC_ARGS_IN_REGS_NUM; 1243 if( mem_parm_offset < 0 ) { 1244 return as_oRegister(i)->as_VMReg(); 1245 } else { 1246 int actual_offset = (mem_parm_offset + frame::memory_parameter_word_sp_offset) * VMRegImpl::slots_per_word; 1247 // Now return a biased offset that will be correct when out_preserve_slots is added back in 1248 return VMRegImpl::stack2reg(actual_offset - SharedRuntime::out_preserve_stack_slots()); 1249 } 1250 } 1251 1252 1253 int SharedRuntime::c_calling_convention(const BasicType *sig_bt, 1254 VMRegPair *regs, 1255 int total_args_passed) { 1256 1257 // Return the number of VMReg stack_slots needed for the args. 1258 // This value does not include an abi space (like register window 1259 // save area). 1260 1261 // The native convention is V8 if !LP64 1262 // The LP64 convention is the V9 convention which is slightly more sane. 1263 1264 // We return the amount of VMReg stack slots we need to reserve for all 1265 // the arguments NOT counting out_preserve_stack_slots. Since we always 1266 // have space for storing at least 6 registers to memory we start with that. 1267 // See int_stk_helper for a further discussion. 1268 int max_stack_slots = (frame::varargs_offset * VMRegImpl::slots_per_word) - SharedRuntime::out_preserve_stack_slots(); 1269 1270 #ifdef _LP64 1271 // V9 convention: All things "as-if" on double-wide stack slots. 1272 // Hoist any int/ptr/long's in the first 6 to int regs. 1273 // Hoist any flt/dbl's in the first 16 dbl regs. 1274 int j = 0; // Count of actual args, not HALVES 1275 for( int i=0; i<total_args_passed; i++, j++ ) { 1276 switch( sig_bt[i] ) { 1277 case T_BOOLEAN: 1278 case T_BYTE: 1279 case T_CHAR: 1280 case T_INT: 1281 case T_SHORT: 1282 regs[i].set1( int_stk_helper( j ) ); break; 1283 case T_LONG: 1284 assert( sig_bt[i+1] == T_VOID, "expecting half" ); 1285 case T_ADDRESS: // raw pointers, like current thread, for VM calls 1286 case T_ARRAY: 1287 case T_OBJECT: 1288 regs[i].set2( int_stk_helper( j ) ); 1289 break; 1290 case T_FLOAT: 1291 if ( j < 16 ) { 1292 // V9ism: floats go in ODD registers 1293 regs[i].set1(as_FloatRegister(1 + (j<<1))->as_VMReg()); 1294 } else { 1295 // V9ism: floats go in ODD stack slot 1296 regs[i].set1(VMRegImpl::stack2reg(1 + (j<<1))); 1297 } 1298 break; 1299 case T_DOUBLE: 1300 assert( sig_bt[i+1] == T_VOID, "expecting half" ); 1301 if ( j < 16 ) { 1302 // V9ism: doubles go in EVEN/ODD regs 1303 regs[i].set2(as_FloatRegister(j<<1)->as_VMReg()); 1304 } else { 1305 // V9ism: doubles go in EVEN/ODD stack slots 1306 regs[i].set2(VMRegImpl::stack2reg(j<<1)); 1307 } 1308 break; 1309 case T_VOID: regs[i].set_bad(); j--; break; // Do not count HALVES 1310 default: 1311 ShouldNotReachHere(); 1312 } 1313 if (regs[i].first()->is_stack()) { 1314 int off = regs[i].first()->reg2stack(); 1315 if (off > max_stack_slots) max_stack_slots = off; 1316 } 1317 if (regs[i].second()->is_stack()) { 1318 int off = regs[i].second()->reg2stack(); 1319 if (off > max_stack_slots) max_stack_slots = off; 1320 } 1321 } 1322 1323 #else // _LP64 1324 // V8 convention: first 6 things in O-regs, rest on stack. 1325 // Alignment is willy-nilly. 1326 for( int i=0; i<total_args_passed; i++ ) { 1327 switch( sig_bt[i] ) { 1328 case T_ADDRESS: // raw pointers, like current thread, for VM calls 1329 case T_ARRAY: 1330 case T_BOOLEAN: 1331 case T_BYTE: 1332 case T_CHAR: 1333 case T_FLOAT: 1334 case T_INT: 1335 case T_OBJECT: 1336 case T_SHORT: 1337 regs[i].set1( int_stk_helper( i ) ); 1338 break; 1339 case T_DOUBLE: 1340 case T_LONG: 1341 assert( sig_bt[i+1] == T_VOID, "expecting half" ); 1342 regs[i].set_pair( int_stk_helper( i+1 ), int_stk_helper( i ) ); 1343 break; 1344 case T_VOID: regs[i].set_bad(); break; 1345 default: 1346 ShouldNotReachHere(); 1347 } 1348 if (regs[i].first()->is_stack()) { 1349 int off = regs[i].first()->reg2stack(); 1350 if (off > max_stack_slots) max_stack_slots = off; 1351 } 1352 if (regs[i].second()->is_stack()) { 1353 int off = regs[i].second()->reg2stack(); 1354 if (off > max_stack_slots) max_stack_slots = off; 1355 } 1356 } 1357 #endif // _LP64 1358 1359 return round_to(max_stack_slots + 1, 2); 1360 1361 } 1362 1363 1364 // --------------------------------------------------------------------------- 1365 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1366 switch (ret_type) { 1367 case T_FLOAT: 1368 __ stf(FloatRegisterImpl::S, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS); 1369 break; 1370 case T_DOUBLE: 1371 __ stf(FloatRegisterImpl::D, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS); 1372 break; 1373 } 1374 } 1375 1376 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { 1377 switch (ret_type) { 1378 case T_FLOAT: 1379 __ ldf(FloatRegisterImpl::S, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS, F0); 1380 break; 1381 case T_DOUBLE: 1382 __ ldf(FloatRegisterImpl::D, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS, F0); 1383 break; 1384 } 1385 } 1386 1387 // Check and forward and pending exception. Thread is stored in 1388 // L7_thread_cache and possibly NOT in G2_thread. Since this is a native call, there 1389 // is no exception handler. We merely pop this frame off and throw the 1390 // exception in the caller's frame. 1391 static void check_forward_pending_exception(MacroAssembler *masm, Register Rex_oop) { 1392 Label L; 1393 __ br_null(Rex_oop, false, Assembler::pt, L, false); 1394 __ delayed()->mov(L7_thread_cache, G2_thread); // restore in case we have exception 1395 // Since this is a native call, we *know* the proper exception handler 1396 // without calling into the VM: it's the empty function. Just pop this 1397 // frame and then jump to forward_exception_entry; O7 will contain the 1398 // native caller's return PC. 1399 AddressLiteral exception_entry(StubRoutines::forward_exception_entry()); 1400 __ jump_to(exception_entry, G3_scratch); 1401 __ delayed()->restore(); // Pop this frame off. 1402 __ bind(L); 1403 } 1404 1405 // A simple move of integer like type 1406 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1407 if (src.first()->is_stack()) { 1408 if (dst.first()->is_stack()) { 1409 // stack to stack 1410 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); 1411 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS); 1412 } else { 1413 // stack to reg 1414 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); 1415 } 1416 } else if (dst.first()->is_stack()) { 1417 // reg to stack 1418 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS); 1419 } else { 1420 __ mov(src.first()->as_Register(), dst.first()->as_Register()); 1421 } 1422 } 1423 1424 // On 64 bit we will store integer like items to the stack as 1425 // 64 bits items (sparc abi) even though java would only store 1426 // 32bits for a parameter. On 32bit it will simply be 32 bits 1427 // So this routine will do 32->32 on 32bit and 32->64 on 64bit 1428 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1429 if (src.first()->is_stack()) { 1430 if (dst.first()->is_stack()) { 1431 // stack to stack 1432 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); 1433 __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS); 1434 } else { 1435 // stack to reg 1436 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); 1437 } 1438 } else if (dst.first()->is_stack()) { 1439 // reg to stack 1440 __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS); 1441 } else { 1442 __ mov(src.first()->as_Register(), dst.first()->as_Register()); 1443 } 1444 } 1445 1446 1447 // An oop arg. Must pass a handle not the oop itself 1448 static void object_move(MacroAssembler* masm, 1449 OopMap* map, 1450 int oop_handle_offset, 1451 int framesize_in_slots, 1452 VMRegPair src, 1453 VMRegPair dst, 1454 bool is_receiver, 1455 int* receiver_offset) { 1456 1457 // must pass a handle. First figure out the location we use as a handle 1458 1459 if (src.first()->is_stack()) { 1460 // Oop is already on the stack 1461 Register rHandle = dst.first()->is_stack() ? L5 : dst.first()->as_Register(); 1462 __ add(FP, reg2offset(src.first()) + STACK_BIAS, rHandle); 1463 __ ld_ptr(rHandle, 0, L4); 1464 #ifdef _LP64 1465 __ movr( Assembler::rc_z, L4, G0, rHandle ); 1466 #else 1467 __ tst( L4 ); 1468 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle ); 1469 #endif 1470 if (dst.first()->is_stack()) { 1471 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS); 1472 } 1473 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); 1474 if (is_receiver) { 1475 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; 1476 } 1477 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); 1478 } else { 1479 // Oop is in an input register pass we must flush it to the stack 1480 const Register rOop = src.first()->as_Register(); 1481 const Register rHandle = L5; 1482 int oop_slot = rOop->input_number() * VMRegImpl::slots_per_word + oop_handle_offset; 1483 int offset = oop_slot*VMRegImpl::stack_slot_size; 1484 Label skip; 1485 __ st_ptr(rOop, SP, offset + STACK_BIAS); 1486 if (is_receiver) { 1487 *receiver_offset = oop_slot * VMRegImpl::stack_slot_size; 1488 } 1489 map->set_oop(VMRegImpl::stack2reg(oop_slot)); 1490 __ add(SP, offset + STACK_BIAS, rHandle); 1491 #ifdef _LP64 1492 __ movr( Assembler::rc_z, rOop, G0, rHandle ); 1493 #else 1494 __ tst( rOop ); 1495 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle ); 1496 #endif 1497 1498 if (dst.first()->is_stack()) { 1499 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS); 1500 } else { 1501 __ mov(rHandle, dst.first()->as_Register()); 1502 } 1503 } 1504 } 1505 1506 // A float arg may have to do float reg int reg conversion 1507 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1508 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move"); 1509 1510 if (src.first()->is_stack()) { 1511 if (dst.first()->is_stack()) { 1512 // stack to stack the easiest of the bunch 1513 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); 1514 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS); 1515 } else { 1516 // stack to reg 1517 if (dst.first()->is_Register()) { 1518 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); 1519 } else { 1520 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister()); 1521 } 1522 } 1523 } else if (dst.first()->is_stack()) { 1524 // reg to stack 1525 if (src.first()->is_Register()) { 1526 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS); 1527 } else { 1528 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS); 1529 } 1530 } else { 1531 // reg to reg 1532 if (src.first()->is_Register()) { 1533 if (dst.first()->is_Register()) { 1534 // gpr -> gpr 1535 __ mov(src.first()->as_Register(), dst.first()->as_Register()); 1536 } else { 1537 // gpr -> fpr 1538 __ st(src.first()->as_Register(), FP, -4 + STACK_BIAS); 1539 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.first()->as_FloatRegister()); 1540 } 1541 } else if (dst.first()->is_Register()) { 1542 // fpr -> gpr 1543 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), FP, -4 + STACK_BIAS); 1544 __ ld(FP, -4 + STACK_BIAS, dst.first()->as_Register()); 1545 } else { 1546 // fpr -> fpr 1547 // In theory these overlap but the ordering is such that this is likely a nop 1548 if ( src.first() != dst.first()) { 1549 __ fmov(FloatRegisterImpl::S, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister()); 1550 } 1551 } 1552 } 1553 } 1554 1555 static void split_long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1556 VMRegPair src_lo(src.first()); 1557 VMRegPair src_hi(src.second()); 1558 VMRegPair dst_lo(dst.first()); 1559 VMRegPair dst_hi(dst.second()); 1560 simple_move32(masm, src_lo, dst_lo); 1561 simple_move32(masm, src_hi, dst_hi); 1562 } 1563 1564 // A long move 1565 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1566 1567 // Do the simple ones here else do two int moves 1568 if (src.is_single_phys_reg() ) { 1569 if (dst.is_single_phys_reg()) { 1570 __ mov(src.first()->as_Register(), dst.first()->as_Register()); 1571 } else { 1572 // split src into two separate registers 1573 // Remember hi means hi address or lsw on sparc 1574 // Move msw to lsw 1575 if (dst.second()->is_reg()) { 1576 // MSW -> MSW 1577 __ srax(src.first()->as_Register(), 32, dst.first()->as_Register()); 1578 // Now LSW -> LSW 1579 // this will only move lo -> lo and ignore hi 1580 VMRegPair split(dst.second()); 1581 simple_move32(masm, src, split); 1582 } else { 1583 VMRegPair split(src.first(), L4->as_VMReg()); 1584 // MSW -> MSW (lo ie. first word) 1585 __ srax(src.first()->as_Register(), 32, L4); 1586 split_long_move(masm, split, dst); 1587 } 1588 } 1589 } else if (dst.is_single_phys_reg()) { 1590 if (src.is_adjacent_aligned_on_stack(2)) { 1591 __ ldx(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); 1592 } else { 1593 // dst is a single reg. 1594 // Remember lo is low address not msb for stack slots 1595 // and lo is the "real" register for registers 1596 // src is 1597 1598 VMRegPair split; 1599 1600 if (src.first()->is_reg()) { 1601 // src.lo (msw) is a reg, src.hi is stk/reg 1602 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> src.lo [the MSW is in the LSW of the reg] 1603 split.set_pair(dst.first(), src.first()); 1604 } else { 1605 // msw is stack move to L5 1606 // lsw is stack move to dst.lo (real reg) 1607 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> L5 1608 split.set_pair(dst.first(), L5->as_VMReg()); 1609 } 1610 1611 // src.lo -> src.lo/L5, src.hi -> dst.lo (the real reg) 1612 // msw -> src.lo/L5, lsw -> dst.lo 1613 split_long_move(masm, src, split); 1614 1615 // So dst now has the low order correct position the 1616 // msw half 1617 __ sllx(split.first()->as_Register(), 32, L5); 1618 1619 const Register d = dst.first()->as_Register(); 1620 __ or3(L5, d, d); 1621 } 1622 } else { 1623 // For LP64 we can probably do better. 1624 split_long_move(masm, src, dst); 1625 } 1626 } 1627 1628 // A double move 1629 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { 1630 1631 // The painful thing here is that like long_move a VMRegPair might be 1632 // 1: a single physical register 1633 // 2: two physical registers (v8) 1634 // 3: a physical reg [lo] and a stack slot [hi] (v8) 1635 // 4: two stack slots 1636 1637 // Since src is always a java calling convention we know that the src pair 1638 // is always either all registers or all stack (and aligned?) 1639 1640 // in a register [lo] and a stack slot [hi] 1641 if (src.first()->is_stack()) { 1642 if (dst.first()->is_stack()) { 1643 // stack to stack the easiest of the bunch 1644 // ought to be a way to do this where if alignment is ok we use ldd/std when possible 1645 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); 1646 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4); 1647 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS); 1648 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS); 1649 } else { 1650 // stack to reg 1651 if (dst.second()->is_stack()) { 1652 // stack -> reg, stack -> stack 1653 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4); 1654 if (dst.first()->is_Register()) { 1655 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); 1656 } else { 1657 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister()); 1658 } 1659 // This was missing. (very rare case) 1660 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS); 1661 } else { 1662 // stack -> reg 1663 // Eventually optimize for alignment QQQ 1664 if (dst.first()->is_Register()) { 1665 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); 1666 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_Register()); 1667 } else { 1668 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister()); 1669 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_FloatRegister()); 1670 } 1671 } 1672 } 1673 } else if (dst.first()->is_stack()) { 1674 // reg to stack 1675 if (src.first()->is_Register()) { 1676 // Eventually optimize for alignment QQQ 1677 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS); 1678 if (src.second()->is_stack()) { 1679 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4); 1680 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS); 1681 } else { 1682 __ st(src.second()->as_Register(), SP, reg2offset(dst.second()) + STACK_BIAS); 1683 } 1684 } else { 1685 // fpr to stack 1686 if (src.second()->is_stack()) { 1687 ShouldNotReachHere(); 1688 } else { 1689 // Is the stack aligned? 1690 if (reg2offset(dst.first()) & 0x7) { 1691 // No do as pairs 1692 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS); 1693 __ stf(FloatRegisterImpl::S, src.second()->as_FloatRegister(), SP, reg2offset(dst.second()) + STACK_BIAS); 1694 } else { 1695 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS); 1696 } 1697 } 1698 } 1699 } else { 1700 // reg to reg 1701 if (src.first()->is_Register()) { 1702 if (dst.first()->is_Register()) { 1703 // gpr -> gpr 1704 __ mov(src.first()->as_Register(), dst.first()->as_Register()); 1705 __ mov(src.second()->as_Register(), dst.second()->as_Register()); 1706 } else { 1707 // gpr -> fpr 1708 // ought to be able to do a single store 1709 __ stx(src.first()->as_Register(), FP, -8 + STACK_BIAS); 1710 __ stx(src.second()->as_Register(), FP, -4 + STACK_BIAS); 1711 // ought to be able to do a single load 1712 __ ldf(FloatRegisterImpl::S, FP, -8 + STACK_BIAS, dst.first()->as_FloatRegister()); 1713 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.second()->as_FloatRegister()); 1714 } 1715 } else if (dst.first()->is_Register()) { 1716 // fpr -> gpr 1717 // ought to be able to do a single store 1718 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), FP, -8 + STACK_BIAS); 1719 // ought to be able to do a single load 1720 // REMEMBER first() is low address not LSB 1721 __ ld(FP, -8 + STACK_BIAS, dst.first()->as_Register()); 1722 if (dst.second()->is_Register()) { 1723 __ ld(FP, -4 + STACK_BIAS, dst.second()->as_Register()); 1724 } else { 1725 __ ld(FP, -4 + STACK_BIAS, L4); 1726 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS); 1727 } 1728 } else { 1729 // fpr -> fpr 1730 // In theory these overlap but the ordering is such that this is likely a nop 1731 if ( src.first() != dst.first()) { 1732 __ fmov(FloatRegisterImpl::D, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister()); 1733 } 1734 } 1735 } 1736 } 1737 1738 // Creates an inner frame if one hasn't already been created, and 1739 // saves a copy of the thread in L7_thread_cache 1740 static void create_inner_frame(MacroAssembler* masm, bool* already_created) { 1741 if (!*already_created) { 1742 __ save_frame(0); 1743 // Save thread in L7 (INNER FRAME); it crosses a bunch of VM calls below 1744 // Don't use save_thread because it smashes G2 and we merely want to save a 1745 // copy 1746 __ mov(G2_thread, L7_thread_cache); 1747 *already_created = true; 1748 } 1749 } 1750 1751 // --------------------------------------------------------------------------- 1752 // Generate a native wrapper for a given method. The method takes arguments 1753 // in the Java compiled code convention, marshals them to the native 1754 // convention (handlizes oops, etc), transitions to native, makes the call, 1755 // returns to java state (possibly blocking), unhandlizes any result and 1756 // returns. 1757 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler* masm, 1758 methodHandle method, 1759 int compile_id, 1760 int total_in_args, 1761 int comp_args_on_stack, // in VMRegStackSlots 1762 BasicType *in_sig_bt, 1763 VMRegPair *in_regs, 1764 BasicType ret_type) { 1765 1766 // Native nmethod wrappers never take possesion of the oop arguments. 1767 // So the caller will gc the arguments. The only thing we need an 1768 // oopMap for is if the call is static 1769 // 1770 // An OopMap for lock (and class if static), and one for the VM call itself 1771 OopMapSet *oop_maps = new OopMapSet(); 1772 intptr_t start = (intptr_t)__ pc(); 1773 1774 // First thing make an ic check to see if we should even be here 1775 { 1776 Label L; 1777 const Register temp_reg = G3_scratch; 1778 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub()); 1779 __ verify_oop(O0); 1780 __ load_klass(O0, temp_reg); 1781 __ cmp_and_brx_short(temp_reg, G5_inline_cache_reg, Assembler::equal, Assembler::pt, L); 1782 1783 __ jump_to(ic_miss, temp_reg); 1784 __ delayed()->nop(); 1785 __ align(CodeEntryAlignment); 1786 __ bind(L); 1787 } 1788 1789 int vep_offset = ((intptr_t)__ pc()) - start; 1790 1791 #ifdef COMPILER1 1792 if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) { 1793 // Object.hashCode can pull the hashCode from the header word 1794 // instead of doing a full VM transition once it's been computed. 1795 // Since hashCode is usually polymorphic at call sites we can't do 1796 // this optimization at the call site without a lot of work. 1797 Label slowCase; 1798 Register receiver = O0; 1799 Register result = O0; 1800 Register header = G3_scratch; 1801 Register hash = G3_scratch; // overwrite header value with hash value 1802 Register mask = G1; // to get hash field from header 1803 1804 // Read the header and build a mask to get its hash field. Give up if the object is not unlocked. 1805 // We depend on hash_mask being at most 32 bits and avoid the use of 1806 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 1807 // vm: see markOop.hpp. 1808 __ ld_ptr(receiver, oopDesc::mark_offset_in_bytes(), header); 1809 __ sethi(markOopDesc::hash_mask, mask); 1810 __ btst(markOopDesc::unlocked_value, header); 1811 __ br(Assembler::zero, false, Assembler::pn, slowCase); 1812 if (UseBiasedLocking) { 1813 // Check if biased and fall through to runtime if so 1814 __ delayed()->nop(); 1815 __ btst(markOopDesc::biased_lock_bit_in_place, header); 1816 __ br(Assembler::notZero, false, Assembler::pn, slowCase); 1817 } 1818 __ delayed()->or3(mask, markOopDesc::hash_mask & 0x3ff, mask); 1819 1820 // Check for a valid (non-zero) hash code and get its value. 1821 #ifdef _LP64 1822 __ srlx(header, markOopDesc::hash_shift, hash); 1823 #else 1824 __ srl(header, markOopDesc::hash_shift, hash); 1825 #endif 1826 __ andcc(hash, mask, hash); 1827 __ br(Assembler::equal, false, Assembler::pn, slowCase); 1828 __ delayed()->nop(); 1829 1830 // leaf return. 1831 __ retl(); 1832 __ delayed()->mov(hash, result); 1833 __ bind(slowCase); 1834 } 1835 #endif // COMPILER1 1836 1837 1838 // We have received a description of where all the java arg are located 1839 // on entry to the wrapper. We need to convert these args to where 1840 // the jni function will expect them. To figure out where they go 1841 // we convert the java signature to a C signature by inserting 1842 // the hidden arguments as arg[0] and possibly arg[1] (static method) 1843 1844 int total_c_args = total_in_args + 1; 1845 if (method->is_static()) { 1846 total_c_args++; 1847 } 1848 1849 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); 1850 VMRegPair * out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); 1851 1852 int argc = 0; 1853 out_sig_bt[argc++] = T_ADDRESS; 1854 if (method->is_static()) { 1855 out_sig_bt[argc++] = T_OBJECT; 1856 } 1857 1858 for (int i = 0; i < total_in_args ; i++ ) { 1859 out_sig_bt[argc++] = in_sig_bt[i]; 1860 } 1861 1862 // Now figure out where the args must be stored and how much stack space 1863 // they require (neglecting out_preserve_stack_slots but space for storing 1864 // the 1st six register arguments). It's weird see int_stk_helper. 1865 // 1866 int out_arg_slots; 1867 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args); 1868 1869 // Compute framesize for the wrapper. We need to handlize all oops in 1870 // registers. We must create space for them here that is disjoint from 1871 // the windowed save area because we have no control over when we might 1872 // flush the window again and overwrite values that gc has since modified. 1873 // (The live window race) 1874 // 1875 // We always just allocate 6 word for storing down these object. This allow 1876 // us to simply record the base and use the Ireg number to decide which 1877 // slot to use. (Note that the reg number is the inbound number not the 1878 // outbound number). 1879 // We must shuffle args to match the native convention, and include var-args space. 1880 1881 // Calculate the total number of stack slots we will need. 1882 1883 // First count the abi requirement plus all of the outgoing args 1884 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 1885 1886 // Now the space for the inbound oop handle area 1887 1888 int oop_handle_offset = stack_slots; 1889 stack_slots += 6*VMRegImpl::slots_per_word; 1890 1891 // Now any space we need for handlizing a klass if static method 1892 1893 int oop_temp_slot_offset = 0; 1894 int klass_slot_offset = 0; 1895 int klass_offset = -1; 1896 int lock_slot_offset = 0; 1897 bool is_static = false; 1898 1899 if (method->is_static()) { 1900 klass_slot_offset = stack_slots; 1901 stack_slots += VMRegImpl::slots_per_word; 1902 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; 1903 is_static = true; 1904 } 1905 1906 // Plus a lock if needed 1907 1908 if (method->is_synchronized()) { 1909 lock_slot_offset = stack_slots; 1910 stack_slots += VMRegImpl::slots_per_word; 1911 } 1912 1913 // Now a place to save return value or as a temporary for any gpr -> fpr moves 1914 stack_slots += 2; 1915 1916 // Ok The space we have allocated will look like: 1917 // 1918 // 1919 // FP-> | | 1920 // |---------------------| 1921 // | 2 slots for moves | 1922 // |---------------------| 1923 // | lock box (if sync) | 1924 // |---------------------| <- lock_slot_offset 1925 // | klass (if static) | 1926 // |---------------------| <- klass_slot_offset 1927 // | oopHandle area | 1928 // |---------------------| <- oop_handle_offset 1929 // | outbound memory | 1930 // | based arguments | 1931 // | | 1932 // |---------------------| 1933 // | vararg area | 1934 // |---------------------| 1935 // | | 1936 // SP-> | out_preserved_slots | 1937 // 1938 // 1939 1940 1941 // Now compute actual number of stack words we need rounding to make 1942 // stack properly aligned. 1943 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word); 1944 1945 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 1946 1947 // Generate stack overflow check before creating frame 1948 __ generate_stack_overflow_check(stack_size); 1949 1950 // Generate a new frame for the wrapper. 1951 __ save(SP, -stack_size, SP); 1952 1953 int frame_complete = ((intptr_t)__ pc()) - start; 1954 1955 __ verify_thread(); 1956 1957 1958 // 1959 // We immediately shuffle the arguments so that any vm call we have to 1960 // make from here on out (sync slow path, jvmti, etc.) we will have 1961 // captured the oops from our caller and have a valid oopMap for 1962 // them. 1963 1964 // ----------------- 1965 // The Grand Shuffle 1966 // 1967 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv* 1968 // (derived from JavaThread* which is in L7_thread_cache) and, if static, 1969 // the class mirror instead of a receiver. This pretty much guarantees that 1970 // register layout will not match. We ignore these extra arguments during 1971 // the shuffle. The shuffle is described by the two calling convention 1972 // vectors we have in our possession. We simply walk the java vector to 1973 // get the source locations and the c vector to get the destinations. 1974 // Because we have a new window and the argument registers are completely 1975 // disjoint ( I0 -> O1, I1 -> O2, ...) we have nothing to worry about 1976 // here. 1977 1978 // This is a trick. We double the stack slots so we can claim 1979 // the oops in the caller's frame. Since we are sure to have 1980 // more args than the caller doubling is enough to make 1981 // sure we can capture all the incoming oop args from the 1982 // caller. 1983 // 1984 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); 1985 int c_arg = total_c_args - 1; 1986 // Record sp-based slot for receiver on stack for non-static methods 1987 int receiver_offset = -1; 1988 1989 // We move the arguments backward because the floating point registers 1990 // destination will always be to a register with a greater or equal register 1991 // number or the stack. 1992 1993 #ifdef ASSERT 1994 bool reg_destroyed[RegisterImpl::number_of_registers]; 1995 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 1996 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) { 1997 reg_destroyed[r] = false; 1998 } 1999 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) { 2000 freg_destroyed[f] = false; 2001 } 2002 2003 #endif /* ASSERT */ 2004 2005 for ( int i = total_in_args - 1; i >= 0 ; i--, c_arg-- ) { 2006 2007 #ifdef ASSERT 2008 if (in_regs[i].first()->is_Register()) { 2009 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "ack!"); 2010 } else if (in_regs[i].first()->is_FloatRegister()) { 2011 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)], "ack!"); 2012 } 2013 if (out_regs[c_arg].first()->is_Register()) { 2014 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; 2015 } else if (out_regs[c_arg].first()->is_FloatRegister()) { 2016 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)] = true; 2017 } 2018 #endif /* ASSERT */ 2019 2020 switch (in_sig_bt[i]) { 2021 case T_ARRAY: 2022 case T_OBJECT: 2023 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], 2024 ((i == 0) && (!is_static)), 2025 &receiver_offset); 2026 break; 2027 case T_VOID: 2028 break; 2029 2030 case T_FLOAT: 2031 float_move(masm, in_regs[i], out_regs[c_arg]); 2032 break; 2033 2034 case T_DOUBLE: 2035 assert( i + 1 < total_in_args && 2036 in_sig_bt[i + 1] == T_VOID && 2037 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 2038 double_move(masm, in_regs[i], out_regs[c_arg]); 2039 break; 2040 2041 case T_LONG : 2042 long_move(masm, in_regs[i], out_regs[c_arg]); 2043 break; 2044 2045 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 2046 2047 default: 2048 move32_64(masm, in_regs[i], out_regs[c_arg]); 2049 } 2050 } 2051 2052 // Pre-load a static method's oop into O1. Used both by locking code and 2053 // the normal JNI call code. 2054 if (method->is_static()) { 2055 __ set_oop_constant(JNIHandles::make_local(Klass::cast(method->method_holder())->java_mirror()), O1); 2056 2057 // Now handlize the static class mirror in O1. It's known not-null. 2058 __ st_ptr(O1, SP, klass_offset + STACK_BIAS); 2059 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); 2060 __ add(SP, klass_offset + STACK_BIAS, O1); 2061 } 2062 2063 2064 const Register L6_handle = L6; 2065 2066 if (method->is_synchronized()) { 2067 __ mov(O1, L6_handle); 2068 } 2069 2070 // We have all of the arguments setup at this point. We MUST NOT touch any Oregs 2071 // except O6/O7. So if we must call out we must push a new frame. We immediately 2072 // push a new frame and flush the windows. 2073 2074 #ifdef _LP64 2075 intptr_t thepc = (intptr_t) __ pc(); 2076 { 2077 address here = __ pc(); 2078 // Call the next instruction 2079 __ call(here + 8, relocInfo::none); 2080 __ delayed()->nop(); 2081 } 2082 #else 2083 intptr_t thepc = __ load_pc_address(O7, 0); 2084 #endif /* _LP64 */ 2085 2086 // We use the same pc/oopMap repeatedly when we call out 2087 oop_maps->add_gc_map(thepc - start, map); 2088 2089 // O7 now has the pc loaded that we will use when we finally call to native. 2090 2091 // Save thread in L7; it crosses a bunch of VM calls below 2092 // Don't use save_thread because it smashes G2 and we merely 2093 // want to save a copy 2094 __ mov(G2_thread, L7_thread_cache); 2095 2096 2097 // If we create an inner frame once is plenty 2098 // when we create it we must also save G2_thread 2099 bool inner_frame_created = false; 2100 2101 // dtrace method entry support 2102 { 2103 SkipIfEqual skip_if( 2104 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero); 2105 // create inner frame 2106 __ save_frame(0); 2107 __ mov(G2_thread, L7_thread_cache); 2108 __ set_oop_constant(JNIHandles::make_local(method()), O1); 2109 __ call_VM_leaf(L7_thread_cache, 2110 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), 2111 G2_thread, O1); 2112 __ restore(); 2113 } 2114 2115 // RedefineClasses() tracing support for obsolete method entry 2116 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) { 2117 // create inner frame 2118 __ save_frame(0); 2119 __ mov(G2_thread, L7_thread_cache); 2120 __ set_oop_constant(JNIHandles::make_local(method()), O1); 2121 __ call_VM_leaf(L7_thread_cache, 2122 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry), 2123 G2_thread, O1); 2124 __ restore(); 2125 } 2126 2127 // We are in the jni frame unless saved_frame is true in which case 2128 // we are in one frame deeper (the "inner" frame). If we are in the 2129 // "inner" frames the args are in the Iregs and if the jni frame then 2130 // they are in the Oregs. 2131 // If we ever need to go to the VM (for locking, jvmti) then 2132 // we will always be in the "inner" frame. 2133 2134 // Lock a synchronized method 2135 int lock_offset = -1; // Set if locked 2136 if (method->is_synchronized()) { 2137 Register Roop = O1; 2138 const Register L3_box = L3; 2139 2140 create_inner_frame(masm, &inner_frame_created); 2141 2142 __ ld_ptr(I1, 0, O1); 2143 Label done; 2144 2145 lock_offset = (lock_slot_offset * VMRegImpl::stack_slot_size); 2146 __ add(FP, lock_offset+STACK_BIAS, L3_box); 2147 #ifdef ASSERT 2148 if (UseBiasedLocking) { 2149 // making the box point to itself will make it clear it went unused 2150 // but also be obviously invalid 2151 __ st_ptr(L3_box, L3_box, 0); 2152 } 2153 #endif // ASSERT 2154 // 2155 // Compiler_lock_object (Roop, Rmark, Rbox, Rscratch) -- kills Rmark, Rbox, Rscratch 2156 // 2157 __ compiler_lock_object(Roop, L1, L3_box, L2); 2158 __ br(Assembler::equal, false, Assembler::pt, done); 2159 __ delayed() -> add(FP, lock_offset+STACK_BIAS, L3_box); 2160 2161 2162 // None of the above fast optimizations worked so we have to get into the 2163 // slow case of monitor enter. Inline a special case of call_VM that 2164 // disallows any pending_exception. 2165 __ mov(Roop, O0); // Need oop in O0 2166 __ mov(L3_box, O1); 2167 2168 // Record last_Java_sp, in case the VM code releases the JVM lock. 2169 2170 __ set_last_Java_frame(FP, I7); 2171 2172 // do the call 2173 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), relocInfo::runtime_call_type); 2174 __ delayed()->mov(L7_thread_cache, O2); 2175 2176 __ restore_thread(L7_thread_cache); // restore G2_thread 2177 __ reset_last_Java_frame(); 2178 2179 #ifdef ASSERT 2180 { Label L; 2181 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0); 2182 __ br_null_short(O0, Assembler::pt, L); 2183 __ stop("no pending exception allowed on exit from IR::monitorenter"); 2184 __ bind(L); 2185 } 2186 #endif 2187 __ bind(done); 2188 } 2189 2190 2191 // Finally just about ready to make the JNI call 2192 2193 __ flush_windows(); 2194 if (inner_frame_created) { 2195 __ restore(); 2196 } else { 2197 // Store only what we need from this frame 2198 // QQQ I think that non-v9 (like we care) we don't need these saves 2199 // either as the flush traps and the current window goes too. 2200 __ st_ptr(FP, SP, FP->sp_offset_in_saved_window()*wordSize + STACK_BIAS); 2201 __ st_ptr(I7, SP, I7->sp_offset_in_saved_window()*wordSize + STACK_BIAS); 2202 } 2203 2204 // get JNIEnv* which is first argument to native 2205 2206 __ add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0); 2207 2208 // Use that pc we placed in O7 a while back as the current frame anchor 2209 2210 __ set_last_Java_frame(SP, O7); 2211 2212 // Transition from _thread_in_Java to _thread_in_native. 2213 __ set(_thread_in_native, G3_scratch); 2214 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset()); 2215 2216 // We flushed the windows ages ago now mark them as flushed 2217 2218 // mark windows as flushed 2219 __ set(JavaFrameAnchor::flushed, G3_scratch); 2220 2221 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset()); 2222 2223 #ifdef _LP64 2224 AddressLiteral dest(method->native_function()); 2225 __ relocate(relocInfo::runtime_call_type); 2226 __ jumpl_to(dest, O7, O7); 2227 #else 2228 __ call(method->native_function(), relocInfo::runtime_call_type); 2229 #endif 2230 __ delayed()->st(G3_scratch, flags); 2231 2232 __ restore_thread(L7_thread_cache); // restore G2_thread 2233 2234 // Unpack native results. For int-types, we do any needed sign-extension 2235 // and move things into I0. The return value there will survive any VM 2236 // calls for blocking or unlocking. An FP or OOP result (handle) is done 2237 // specially in the slow-path code. 2238 switch (ret_type) { 2239 case T_VOID: break; // Nothing to do! 2240 case T_FLOAT: break; // Got it where we want it (unless slow-path) 2241 case T_DOUBLE: break; // Got it where we want it (unless slow-path) 2242 // In 64 bits build result is in O0, in O0, O1 in 32bit build 2243 case T_LONG: 2244 #ifndef _LP64 2245 __ mov(O1, I1); 2246 #endif 2247 // Fall thru 2248 case T_OBJECT: // Really a handle 2249 case T_ARRAY: 2250 case T_INT: 2251 __ mov(O0, I0); 2252 break; 2253 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, I0); break; // !0 => true; 0 => false 2254 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, I0); break; 2255 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, I0); break; // cannot use and3, 0xFFFF too big as immediate value! 2256 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, I0); break; 2257 break; // Cannot de-handlize until after reclaiming jvm_lock 2258 default: 2259 ShouldNotReachHere(); 2260 } 2261 2262 // must we block? 2263 2264 // Block, if necessary, before resuming in _thread_in_Java state. 2265 // In order for GC to work, don't clear the last_Java_sp until after blocking. 2266 { Label no_block; 2267 AddressLiteral sync_state(SafepointSynchronize::address_of_state()); 2268 2269 // Switch thread to "native transition" state before reading the synchronization state. 2270 // This additional state is necessary because reading and testing the synchronization 2271 // state is not atomic w.r.t. GC, as this scenario demonstrates: 2272 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. 2273 // VM thread changes sync state to synchronizing and suspends threads for GC. 2274 // Thread A is resumed to finish this native method, but doesn't block here since it 2275 // didn't see any synchronization is progress, and escapes. 2276 __ set(_thread_in_native_trans, G3_scratch); 2277 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset()); 2278 if(os::is_MP()) { 2279 if (UseMembar) { 2280 // Force this write out before the read below 2281 __ membar(Assembler::StoreLoad); 2282 } else { 2283 // Write serialization page so VM thread can do a pseudo remote membar. 2284 // We use the current thread pointer to calculate a thread specific 2285 // offset to write to within the page. This minimizes bus traffic 2286 // due to cache line collision. 2287 __ serialize_memory(G2_thread, G1_scratch, G3_scratch); 2288 } 2289 } 2290 __ load_contents(sync_state, G3_scratch); 2291 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); 2292 2293 Label L; 2294 Address suspend_state(G2_thread, JavaThread::suspend_flags_offset()); 2295 __ br(Assembler::notEqual, false, Assembler::pn, L); 2296 __ delayed()->ld(suspend_state, G3_scratch); 2297 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block); 2298 __ bind(L); 2299 2300 // Block. Save any potential method result value before the operation and 2301 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this 2302 // lets us share the oopMap we used when we went native rather the create 2303 // a distinct one for this pc 2304 // 2305 save_native_result(masm, ret_type, stack_slots); 2306 __ call_VM_leaf(L7_thread_cache, 2307 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), 2308 G2_thread); 2309 2310 // Restore any method result value 2311 restore_native_result(masm, ret_type, stack_slots); 2312 __ bind(no_block); 2313 } 2314 2315 // thread state is thread_in_native_trans. Any safepoint blocking has already 2316 // happened so we can now change state to _thread_in_Java. 2317 2318 2319 __ set(_thread_in_Java, G3_scratch); 2320 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset()); 2321 2322 2323 Label no_reguard; 2324 __ ld(G2_thread, JavaThread::stack_guard_state_offset(), G3_scratch); 2325 __ cmp_and_br_short(G3_scratch, JavaThread::stack_guard_yellow_disabled, Assembler::notEqual, Assembler::pt, no_reguard); 2326 2327 save_native_result(masm, ret_type, stack_slots); 2328 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); 2329 __ delayed()->nop(); 2330 2331 __ restore_thread(L7_thread_cache); // restore G2_thread 2332 restore_native_result(masm, ret_type, stack_slots); 2333 2334 __ bind(no_reguard); 2335 2336 // Handle possible exception (will unlock if necessary) 2337 2338 // native result if any is live in freg or I0 (and I1 if long and 32bit vm) 2339 2340 // Unlock 2341 if (method->is_synchronized()) { 2342 Label done; 2343 Register I2_ex_oop = I2; 2344 const Register L3_box = L3; 2345 // Get locked oop from the handle we passed to jni 2346 __ ld_ptr(L6_handle, 0, L4); 2347 __ add(SP, lock_offset+STACK_BIAS, L3_box); 2348 // Must save pending exception around the slow-path VM call. Since it's a 2349 // leaf call, the pending exception (if any) can be kept in a register. 2350 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), I2_ex_oop); 2351 // Now unlock 2352 // (Roop, Rmark, Rbox, Rscratch) 2353 __ compiler_unlock_object(L4, L1, L3_box, L2); 2354 __ br(Assembler::equal, false, Assembler::pt, done); 2355 __ delayed()-> add(SP, lock_offset+STACK_BIAS, L3_box); 2356 2357 // save and restore any potential method result value around the unlocking 2358 // operation. Will save in I0 (or stack for FP returns). 2359 save_native_result(masm, ret_type, stack_slots); 2360 2361 // Must clear pending-exception before re-entering the VM. Since this is 2362 // a leaf call, pending-exception-oop can be safely kept in a register. 2363 __ st_ptr(G0, G2_thread, in_bytes(Thread::pending_exception_offset())); 2364 2365 // slow case of monitor enter. Inline a special case of call_VM that 2366 // disallows any pending_exception. 2367 __ mov(L3_box, O1); 2368 2369 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), relocInfo::runtime_call_type); 2370 __ delayed()->mov(L4, O0); // Need oop in O0 2371 2372 __ restore_thread(L7_thread_cache); // restore G2_thread 2373 2374 #ifdef ASSERT 2375 { Label L; 2376 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0); 2377 __ br_null_short(O0, Assembler::pt, L); 2378 __ stop("no pending exception allowed on exit from IR::monitorexit"); 2379 __ bind(L); 2380 } 2381 #endif 2382 restore_native_result(masm, ret_type, stack_slots); 2383 // check_forward_pending_exception jump to forward_exception if any pending 2384 // exception is set. The forward_exception routine expects to see the 2385 // exception in pending_exception and not in a register. Kind of clumsy, 2386 // since all folks who branch to forward_exception must have tested 2387 // pending_exception first and hence have it in a register already. 2388 __ st_ptr(I2_ex_oop, G2_thread, in_bytes(Thread::pending_exception_offset())); 2389 __ bind(done); 2390 } 2391 2392 // Tell dtrace about this method exit 2393 { 2394 SkipIfEqual skip_if( 2395 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero); 2396 save_native_result(masm, ret_type, stack_slots); 2397 __ set_oop_constant(JNIHandles::make_local(method()), O1); 2398 __ call_VM_leaf(L7_thread_cache, 2399 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), 2400 G2_thread, O1); 2401 restore_native_result(masm, ret_type, stack_slots); 2402 } 2403 2404 // Clear "last Java frame" SP and PC. 2405 __ verify_thread(); // G2_thread must be correct 2406 __ reset_last_Java_frame(); 2407 2408 // Unpack oop result 2409 if (ret_type == T_OBJECT || ret_type == T_ARRAY) { 2410 Label L; 2411 __ addcc(G0, I0, G0); 2412 __ brx(Assembler::notZero, true, Assembler::pt, L); 2413 __ delayed()->ld_ptr(I0, 0, I0); 2414 __ mov(G0, I0); 2415 __ bind(L); 2416 __ verify_oop(I0); 2417 } 2418 2419 // reset handle block 2420 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), L5); 2421 __ st_ptr(G0, L5, JNIHandleBlock::top_offset_in_bytes()); 2422 2423 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), G3_scratch); 2424 check_forward_pending_exception(masm, G3_scratch); 2425 2426 2427 // Return 2428 2429 #ifndef _LP64 2430 if (ret_type == T_LONG) { 2431 2432 // Must leave proper result in O0,O1 and G1 (c2/tiered only) 2433 __ sllx(I0, 32, G1); // Shift bits into high G1 2434 __ srl (I1, 0, I1); // Zero extend O1 (harmless?) 2435 __ or3 (I1, G1, G1); // OR 64 bits into G1 2436 } 2437 #endif 2438 2439 __ ret(); 2440 __ delayed()->restore(); 2441 2442 __ flush(); 2443 2444 nmethod *nm = nmethod::new_native_nmethod(method, 2445 compile_id, 2446 masm->code(), 2447 vep_offset, 2448 frame_complete, 2449 stack_slots / VMRegImpl::slots_per_word, 2450 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), 2451 in_ByteSize(lock_offset), 2452 oop_maps); 2453 return nm; 2454 2455 } 2456 2457 #ifdef HAVE_DTRACE_H 2458 // --------------------------------------------------------------------------- 2459 // Generate a dtrace nmethod for a given signature. The method takes arguments 2460 // in the Java compiled code convention, marshals them to the native 2461 // abi and then leaves nops at the position you would expect to call a native 2462 // function. When the probe is enabled the nops are replaced with a trap 2463 // instruction that dtrace inserts and the trace will cause a notification 2464 // to dtrace. 2465 // 2466 // The probes are only able to take primitive types and java/lang/String as 2467 // arguments. No other java types are allowed. Strings are converted to utf8 2468 // strings so that from dtrace point of view java strings are converted to C 2469 // strings. There is an arbitrary fixed limit on the total space that a method 2470 // can use for converting the strings. (256 chars per string in the signature). 2471 // So any java string larger then this is truncated. 2472 2473 static int fp_offset[ConcreteRegisterImpl::number_of_registers] = { 0 }; 2474 static bool offsets_initialized = false; 2475 2476 static VMRegPair reg64_to_VMRegPair(Register r) { 2477 VMRegPair ret; 2478 if (wordSize == 8) { 2479 ret.set2(r->as_VMReg()); 2480 } else { 2481 ret.set_pair(r->successor()->as_VMReg(), r->as_VMReg()); 2482 } 2483 return ret; 2484 } 2485 2486 2487 nmethod *SharedRuntime::generate_dtrace_nmethod( 2488 MacroAssembler *masm, methodHandle method) { 2489 2490 2491 // generate_dtrace_nmethod is guarded by a mutex so we are sure to 2492 // be single threaded in this method. 2493 assert(AdapterHandlerLibrary_lock->owned_by_self(), "must be"); 2494 2495 // Fill in the signature array, for the calling-convention call. 2496 int total_args_passed = method->size_of_parameters(); 2497 2498 BasicType* in_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed); 2499 VMRegPair *in_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed); 2500 2501 // The signature we are going to use for the trap that dtrace will see 2502 // java/lang/String is converted. We drop "this" and any other object 2503 // is converted to NULL. (A one-slot java/lang/Long object reference 2504 // is converted to a two-slot long, which is why we double the allocation). 2505 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed * 2); 2506 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed * 2); 2507 2508 int i=0; 2509 int total_strings = 0; 2510 int first_arg_to_pass = 0; 2511 int total_c_args = 0; 2512 2513 // Skip the receiver as dtrace doesn't want to see it 2514 if( !method->is_static() ) { 2515 in_sig_bt[i++] = T_OBJECT; 2516 first_arg_to_pass = 1; 2517 } 2518 2519 SignatureStream ss(method->signature()); 2520 for ( ; !ss.at_return_type(); ss.next()) { 2521 BasicType bt = ss.type(); 2522 in_sig_bt[i++] = bt; // Collect remaining bits of signature 2523 out_sig_bt[total_c_args++] = bt; 2524 if( bt == T_OBJECT) { 2525 Symbol* s = ss.as_symbol_or_null(); 2526 if (s == vmSymbols::java_lang_String()) { 2527 total_strings++; 2528 out_sig_bt[total_c_args-1] = T_ADDRESS; 2529 } else if (s == vmSymbols::java_lang_Boolean() || 2530 s == vmSymbols::java_lang_Byte()) { 2531 out_sig_bt[total_c_args-1] = T_BYTE; 2532 } else if (s == vmSymbols::java_lang_Character() || 2533 s == vmSymbols::java_lang_Short()) { 2534 out_sig_bt[total_c_args-1] = T_SHORT; 2535 } else if (s == vmSymbols::java_lang_Integer() || 2536 s == vmSymbols::java_lang_Float()) { 2537 out_sig_bt[total_c_args-1] = T_INT; 2538 } else if (s == vmSymbols::java_lang_Long() || 2539 s == vmSymbols::java_lang_Double()) { 2540 out_sig_bt[total_c_args-1] = T_LONG; 2541 out_sig_bt[total_c_args++] = T_VOID; 2542 } 2543 } else if ( bt == T_LONG || bt == T_DOUBLE ) { 2544 in_sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots 2545 // We convert double to long 2546 out_sig_bt[total_c_args-1] = T_LONG; 2547 out_sig_bt[total_c_args++] = T_VOID; 2548 } else if ( bt == T_FLOAT) { 2549 // We convert float to int 2550 out_sig_bt[total_c_args-1] = T_INT; 2551 } 2552 } 2553 2554 assert(i==total_args_passed, "validly parsed signature"); 2555 2556 // Now get the compiled-Java layout as input arguments 2557 int comp_args_on_stack; 2558 comp_args_on_stack = SharedRuntime::java_calling_convention( 2559 in_sig_bt, in_regs, total_args_passed, false); 2560 2561 // We have received a description of where all the java arg are located 2562 // on entry to the wrapper. We need to convert these args to where 2563 // the a native (non-jni) function would expect them. To figure out 2564 // where they go we convert the java signature to a C signature and remove 2565 // T_VOID for any long/double we might have received. 2566 2567 2568 // Now figure out where the args must be stored and how much stack space 2569 // they require (neglecting out_preserve_stack_slots but space for storing 2570 // the 1st six register arguments). It's weird see int_stk_helper. 2571 // 2572 int out_arg_slots; 2573 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args); 2574 2575 // Calculate the total number of stack slots we will need. 2576 2577 // First count the abi requirement plus all of the outgoing args 2578 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; 2579 2580 // Plus a temp for possible converion of float/double/long register args 2581 2582 int conversion_temp = stack_slots; 2583 stack_slots += 2; 2584 2585 2586 // Now space for the string(s) we must convert 2587 2588 int string_locs = stack_slots; 2589 stack_slots += total_strings * 2590 (max_dtrace_string_size / VMRegImpl::stack_slot_size); 2591 2592 // Ok The space we have allocated will look like: 2593 // 2594 // 2595 // FP-> | | 2596 // |---------------------| 2597 // | string[n] | 2598 // |---------------------| <- string_locs[n] 2599 // | string[n-1] | 2600 // |---------------------| <- string_locs[n-1] 2601 // | ... | 2602 // | ... | 2603 // |---------------------| <- string_locs[1] 2604 // | string[0] | 2605 // |---------------------| <- string_locs[0] 2606 // | temp | 2607 // |---------------------| <- conversion_temp 2608 // | outbound memory | 2609 // | based arguments | 2610 // | | 2611 // |---------------------| 2612 // | | 2613 // SP-> | out_preserved_slots | 2614 // 2615 // 2616 2617 // Now compute actual number of stack words we need rounding to make 2618 // stack properly aligned. 2619 stack_slots = round_to(stack_slots, 4 * VMRegImpl::slots_per_word); 2620 2621 int stack_size = stack_slots * VMRegImpl::stack_slot_size; 2622 2623 intptr_t start = (intptr_t)__ pc(); 2624 2625 // First thing make an ic check to see if we should even be here 2626 2627 { 2628 Label L; 2629 const Register temp_reg = G3_scratch; 2630 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub()); 2631 __ verify_oop(O0); 2632 __ ld_ptr(O0, oopDesc::klass_offset_in_bytes(), temp_reg); 2633 __ cmp_and_brx_short(temp_reg, G5_inline_cache_reg, Assembler::equal, Assembler::pt, L); 2634 2635 __ jump_to(ic_miss, temp_reg); 2636 __ delayed()->nop(); 2637 __ align(CodeEntryAlignment); 2638 __ bind(L); 2639 } 2640 2641 int vep_offset = ((intptr_t)__ pc()) - start; 2642 2643 2644 // The instruction at the verified entry point must be 5 bytes or longer 2645 // because it can be patched on the fly by make_non_entrant. The stack bang 2646 // instruction fits that requirement. 2647 2648 // Generate stack overflow check before creating frame 2649 __ generate_stack_overflow_check(stack_size); 2650 2651 assert(((intptr_t)__ pc() - start - vep_offset) >= 5, 2652 "valid size for make_non_entrant"); 2653 2654 // Generate a new frame for the wrapper. 2655 __ save(SP, -stack_size, SP); 2656 2657 // Frame is now completed as far a size and linkage. 2658 2659 int frame_complete = ((intptr_t)__ pc()) - start; 2660 2661 #ifdef ASSERT 2662 bool reg_destroyed[RegisterImpl::number_of_registers]; 2663 bool freg_destroyed[FloatRegisterImpl::number_of_registers]; 2664 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) { 2665 reg_destroyed[r] = false; 2666 } 2667 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) { 2668 freg_destroyed[f] = false; 2669 } 2670 2671 #endif /* ASSERT */ 2672 2673 VMRegPair zero; 2674 const Register g0 = G0; // without this we get a compiler warning (why??) 2675 zero.set2(g0->as_VMReg()); 2676 2677 int c_arg, j_arg; 2678 2679 Register conversion_off = noreg; 2680 2681 for (j_arg = first_arg_to_pass, c_arg = 0 ; 2682 j_arg < total_args_passed ; j_arg++, c_arg++ ) { 2683 2684 VMRegPair src = in_regs[j_arg]; 2685 VMRegPair dst = out_regs[c_arg]; 2686 2687 #ifdef ASSERT 2688 if (src.first()->is_Register()) { 2689 assert(!reg_destroyed[src.first()->as_Register()->encoding()], "ack!"); 2690 } else if (src.first()->is_FloatRegister()) { 2691 assert(!freg_destroyed[src.first()->as_FloatRegister()->encoding( 2692 FloatRegisterImpl::S)], "ack!"); 2693 } 2694 if (dst.first()->is_Register()) { 2695 reg_destroyed[dst.first()->as_Register()->encoding()] = true; 2696 } else if (dst.first()->is_FloatRegister()) { 2697 freg_destroyed[dst.first()->as_FloatRegister()->encoding( 2698 FloatRegisterImpl::S)] = true; 2699 } 2700 #endif /* ASSERT */ 2701 2702 switch (in_sig_bt[j_arg]) { 2703 case T_ARRAY: 2704 case T_OBJECT: 2705 { 2706 if (out_sig_bt[c_arg] == T_BYTE || out_sig_bt[c_arg] == T_SHORT || 2707 out_sig_bt[c_arg] == T_INT || out_sig_bt[c_arg] == T_LONG) { 2708 // need to unbox a one-slot value 2709 Register in_reg = L0; 2710 Register tmp = L2; 2711 if ( src.first()->is_reg() ) { 2712 in_reg = src.first()->as_Register(); 2713 } else { 2714 assert(Assembler::is_simm13(reg2offset(src.first()) + STACK_BIAS), 2715 "must be"); 2716 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, in_reg); 2717 } 2718 // If the final destination is an acceptable register 2719 if ( dst.first()->is_reg() ) { 2720 if ( dst.is_single_phys_reg() || out_sig_bt[c_arg] != T_LONG ) { 2721 tmp = dst.first()->as_Register(); 2722 } 2723 } 2724 2725 Label skipUnbox; 2726 if ( wordSize == 4 && out_sig_bt[c_arg] == T_LONG ) { 2727 __ mov(G0, tmp->successor()); 2728 } 2729 __ br_null(in_reg, true, Assembler::pn, skipUnbox, false); 2730 __ delayed()->mov(G0, tmp); 2731 2732 BasicType bt = out_sig_bt[c_arg]; 2733 int box_offset = java_lang_boxing_object::value_offset_in_bytes(bt); 2734 switch (bt) { 2735 case T_BYTE: 2736 __ ldub(in_reg, box_offset, tmp); break; 2737 case T_SHORT: 2738 __ lduh(in_reg, box_offset, tmp); break; 2739 case T_INT: 2740 __ ld(in_reg, box_offset, tmp); break; 2741 case T_LONG: 2742 __ ld_long(in_reg, box_offset, tmp); break; 2743 default: ShouldNotReachHere(); 2744 } 2745 2746 __ bind(skipUnbox); 2747 // If tmp wasn't final destination copy to final destination 2748 if (tmp == L2) { 2749 VMRegPair tmp_as_VM = reg64_to_VMRegPair(L2); 2750 if (out_sig_bt[c_arg] == T_LONG) { 2751 long_move(masm, tmp_as_VM, dst); 2752 } else { 2753 move32_64(masm, tmp_as_VM, out_regs[c_arg]); 2754 } 2755 } 2756 if (out_sig_bt[c_arg] == T_LONG) { 2757 assert(out_sig_bt[c_arg+1] == T_VOID, "must be"); 2758 ++c_arg; // move over the T_VOID to keep the loop indices in sync 2759 } 2760 } else if (out_sig_bt[c_arg] == T_ADDRESS) { 2761 Register s = 2762 src.first()->is_reg() ? src.first()->as_Register() : L2; 2763 Register d = 2764 dst.first()->is_reg() ? dst.first()->as_Register() : L2; 2765 2766 // We store the oop now so that the conversion pass can reach 2767 // while in the inner frame. This will be the only store if 2768 // the oop is NULL. 2769 if (s != L2) { 2770 // src is register 2771 if (d != L2) { 2772 // dst is register 2773 __ mov(s, d); 2774 } else { 2775 assert(Assembler::is_simm13(reg2offset(dst.first()) + 2776 STACK_BIAS), "must be"); 2777 __ st_ptr(s, SP, reg2offset(dst.first()) + STACK_BIAS); 2778 } 2779 } else { 2780 // src not a register 2781 assert(Assembler::is_simm13(reg2offset(src.first()) + 2782 STACK_BIAS), "must be"); 2783 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, d); 2784 if (d == L2) { 2785 assert(Assembler::is_simm13(reg2offset(dst.first()) + 2786 STACK_BIAS), "must be"); 2787 __ st_ptr(d, SP, reg2offset(dst.first()) + STACK_BIAS); 2788 } 2789 } 2790 } else if (out_sig_bt[c_arg] != T_VOID) { 2791 // Convert the arg to NULL 2792 if (dst.first()->is_reg()) { 2793 __ mov(G0, dst.first()->as_Register()); 2794 } else { 2795 assert(Assembler::is_simm13(reg2offset(dst.first()) + 2796 STACK_BIAS), "must be"); 2797 __ st_ptr(G0, SP, reg2offset(dst.first()) + STACK_BIAS); 2798 } 2799 } 2800 } 2801 break; 2802 case T_VOID: 2803 break; 2804 2805 case T_FLOAT: 2806 if (src.first()->is_stack()) { 2807 // Stack to stack/reg is simple 2808 move32_64(masm, src, dst); 2809 } else { 2810 if (dst.first()->is_reg()) { 2811 // freg -> reg 2812 int off = 2813 STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size; 2814 Register d = dst.first()->as_Register(); 2815 if (Assembler::is_simm13(off)) { 2816 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), 2817 SP, off); 2818 __ ld(SP, off, d); 2819 } else { 2820 if (conversion_off == noreg) { 2821 __ set(off, L6); 2822 conversion_off = L6; 2823 } 2824 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), 2825 SP, conversion_off); 2826 __ ld(SP, conversion_off , d); 2827 } 2828 } else { 2829 // freg -> mem 2830 int off = STACK_BIAS + reg2offset(dst.first()); 2831 if (Assembler::is_simm13(off)) { 2832 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), 2833 SP, off); 2834 } else { 2835 if (conversion_off == noreg) { 2836 __ set(off, L6); 2837 conversion_off = L6; 2838 } 2839 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), 2840 SP, conversion_off); 2841 } 2842 } 2843 } 2844 break; 2845 2846 case T_DOUBLE: 2847 assert( j_arg + 1 < total_args_passed && 2848 in_sig_bt[j_arg + 1] == T_VOID && 2849 out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); 2850 if (src.first()->is_stack()) { 2851 // Stack to stack/reg is simple 2852 long_move(masm, src, dst); 2853 } else { 2854 Register d = dst.first()->is_reg() ? dst.first()->as_Register() : L2; 2855 2856 // Destination could be an odd reg on 32bit in which case 2857 // we can't load direct to the destination. 2858 2859 if (!d->is_even() && wordSize == 4) { 2860 d = L2; 2861 } 2862 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size; 2863 if (Assembler::is_simm13(off)) { 2864 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), 2865 SP, off); 2866 __ ld_long(SP, off, d); 2867 } else { 2868 if (conversion_off == noreg) { 2869 __ set(off, L6); 2870 conversion_off = L6; 2871 } 2872 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), 2873 SP, conversion_off); 2874 __ ld_long(SP, conversion_off, d); 2875 } 2876 if (d == L2) { 2877 long_move(masm, reg64_to_VMRegPair(L2), dst); 2878 } 2879 } 2880 break; 2881 2882 case T_LONG : 2883 // 32bit can't do a split move of something like g1 -> O0, O1 2884 // so use a memory temp 2885 if (src.is_single_phys_reg() && wordSize == 4) { 2886 Register tmp = L2; 2887 if (dst.first()->is_reg() && 2888 (wordSize == 8 || dst.first()->as_Register()->is_even())) { 2889 tmp = dst.first()->as_Register(); 2890 } 2891 2892 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size; 2893 if (Assembler::is_simm13(off)) { 2894 __ stx(src.first()->as_Register(), SP, off); 2895 __ ld_long(SP, off, tmp); 2896 } else { 2897 if (conversion_off == noreg) { 2898 __ set(off, L6); 2899 conversion_off = L6; 2900 } 2901 __ stx(src.first()->as_Register(), SP, conversion_off); 2902 __ ld_long(SP, conversion_off, tmp); 2903 } 2904 2905 if (tmp == L2) { 2906 long_move(masm, reg64_to_VMRegPair(L2), dst); 2907 } 2908 } else { 2909 long_move(masm, src, dst); 2910 } 2911 break; 2912 2913 case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); 2914 2915 default: 2916 move32_64(masm, src, dst); 2917 } 2918 } 2919 2920 2921 // If we have any strings we must store any register based arg to the stack 2922 // This includes any still live xmm registers too. 2923 2924 if (total_strings > 0 ) { 2925 2926 // protect all the arg registers 2927 __ save_frame(0); 2928 __ mov(G2_thread, L7_thread_cache); 2929 const Register L2_string_off = L2; 2930 2931 // Get first string offset 2932 __ set(string_locs * VMRegImpl::stack_slot_size, L2_string_off); 2933 2934 for (c_arg = 0 ; c_arg < total_c_args ; c_arg++ ) { 2935 if (out_sig_bt[c_arg] == T_ADDRESS) { 2936 2937 VMRegPair dst = out_regs[c_arg]; 2938 const Register d = dst.first()->is_reg() ? 2939 dst.first()->as_Register()->after_save() : noreg; 2940 2941 // It's a string the oop and it was already copied to the out arg 2942 // position 2943 if (d != noreg) { 2944 __ mov(d, O0); 2945 } else { 2946 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS), 2947 "must be"); 2948 __ ld_ptr(FP, reg2offset(dst.first()) + STACK_BIAS, O0); 2949 } 2950 Label skip; 2951 2952 __ br_null(O0, false, Assembler::pn, skip, false); 2953 __ delayed()->add(FP, L2_string_off, O1); 2954 2955 if (d != noreg) { 2956 __ mov(O1, d); 2957 } else { 2958 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS), 2959 "must be"); 2960 __ st_ptr(O1, FP, reg2offset(dst.first()) + STACK_BIAS); 2961 } 2962 2963 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::get_utf), 2964 relocInfo::runtime_call_type); 2965 __ delayed()->add(L2_string_off, max_dtrace_string_size, L2_string_off); 2966 2967 __ bind(skip); 2968 2969 } 2970 2971 } 2972 __ mov(L7_thread_cache, G2_thread); 2973 __ restore(); 2974 2975 } 2976 2977 2978 // Ok now we are done. Need to place the nop that dtrace wants in order to 2979 // patch in the trap 2980 2981 int patch_offset = ((intptr_t)__ pc()) - start; 2982 2983 __ nop(); 2984 2985 2986 // Return 2987 2988 __ ret(); 2989 __ delayed()->restore(); 2990 2991 __ flush(); 2992 2993 nmethod *nm = nmethod::new_dtrace_nmethod( 2994 method, masm->code(), vep_offset, patch_offset, frame_complete, 2995 stack_slots / VMRegImpl::slots_per_word); 2996 return nm; 2997 2998 } 2999 3000 #endif // HAVE_DTRACE_H 3001 3002 // this function returns the adjust size (in number of words) to a c2i adapter 3003 // activation for use during deoptimization 3004 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { 3005 assert(callee_locals >= callee_parameters, 3006 "test and remove; got more parms than locals"); 3007 if (callee_locals < callee_parameters) 3008 return 0; // No adjustment for negative locals 3009 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords; 3010 return round_to(diff, WordsPerLong); 3011 } 3012 3013 // "Top of Stack" slots that may be unused by the calling convention but must 3014 // otherwise be preserved. 3015 // On Intel these are not necessary and the value can be zero. 3016 // On Sparc this describes the words reserved for storing a register window 3017 // when an interrupt occurs. 3018 uint SharedRuntime::out_preserve_stack_slots() { 3019 return frame::register_save_words * VMRegImpl::slots_per_word; 3020 } 3021 3022 static void gen_new_frame(MacroAssembler* masm, bool deopt) { 3023 // 3024 // Common out the new frame generation for deopt and uncommon trap 3025 // 3026 Register G3pcs = G3_scratch; // Array of new pcs (input) 3027 Register Oreturn0 = O0; 3028 Register Oreturn1 = O1; 3029 Register O2UnrollBlock = O2; 3030 Register O3array = O3; // Array of frame sizes (input) 3031 Register O4array_size = O4; // number of frames (input) 3032 Register O7frame_size = O7; // number of frames (input) 3033 3034 __ ld_ptr(O3array, 0, O7frame_size); 3035 __ sub(G0, O7frame_size, O7frame_size); 3036 __ save(SP, O7frame_size, SP); 3037 __ ld_ptr(G3pcs, 0, I7); // load frame's new pc 3038 3039 #ifdef ASSERT 3040 // make sure that the frames are aligned properly 3041 #ifndef _LP64 3042 __ btst(wordSize*2-1, SP); 3043 __ breakpoint_trap(Assembler::notZero); 3044 #endif 3045 #endif 3046 3047 // Deopt needs to pass some extra live values from frame to frame 3048 3049 if (deopt) { 3050 __ mov(Oreturn0->after_save(), Oreturn0); 3051 __ mov(Oreturn1->after_save(), Oreturn1); 3052 } 3053 3054 __ mov(O4array_size->after_save(), O4array_size); 3055 __ sub(O4array_size, 1, O4array_size); 3056 __ mov(O3array->after_save(), O3array); 3057 __ mov(O2UnrollBlock->after_save(), O2UnrollBlock); 3058 __ add(G3pcs, wordSize, G3pcs); // point to next pc value 3059 3060 #ifdef ASSERT 3061 // trash registers to show a clear pattern in backtraces 3062 __ set(0xDEAD0000, I0); 3063 __ add(I0, 2, I1); 3064 __ add(I0, 4, I2); 3065 __ add(I0, 6, I3); 3066 __ add(I0, 8, I4); 3067 // Don't touch I5 could have valuable savedSP 3068 __ set(0xDEADBEEF, L0); 3069 __ mov(L0, L1); 3070 __ mov(L0, L2); 3071 __ mov(L0, L3); 3072 __ mov(L0, L4); 3073 __ mov(L0, L5); 3074 3075 // trash the return value as there is nothing to return yet 3076 __ set(0xDEAD0001, O7); 3077 #endif 3078 3079 __ mov(SP, O5_savedSP); 3080 } 3081 3082 3083 static void make_new_frames(MacroAssembler* masm, bool deopt) { 3084 // 3085 // loop through the UnrollBlock info and create new frames 3086 // 3087 Register G3pcs = G3_scratch; 3088 Register Oreturn0 = O0; 3089 Register Oreturn1 = O1; 3090 Register O2UnrollBlock = O2; 3091 Register O3array = O3; 3092 Register O4array_size = O4; 3093 Label loop; 3094 3095 // Before we make new frames, check to see if stack is available. 3096 // Do this after the caller's return address is on top of stack 3097 if (UseStackBanging) { 3098 // Get total frame size for interpreted frames 3099 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes(), O4); 3100 __ bang_stack_size(O4, O3, G3_scratch); 3101 } 3102 3103 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(), O4array_size); 3104 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(), G3pcs); 3105 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(), O3array); 3106 3107 // Adjust old interpreter frame to make space for new frame's extra java locals 3108 // 3109 // We capture the original sp for the transition frame only because it is needed in 3110 // order to properly calculate interpreter_sp_adjustment. Even though in real life 3111 // every interpreter frame captures a savedSP it is only needed at the transition 3112 // (fortunately). If we had to have it correct everywhere then we would need to 3113 // be told the sp_adjustment for each frame we create. If the frame size array 3114 // were to have twice the frame count entries then we could have pairs [sp_adjustment, frame_size] 3115 // for each frame we create and keep up the illusion every where. 3116 // 3117 3118 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(), O7); 3119 __ mov(SP, O5_savedSP); // remember initial sender's original sp before adjustment 3120 __ sub(SP, O7, SP); 3121 3122 #ifdef ASSERT 3123 // make sure that there is at least one entry in the array 3124 __ tst(O4array_size); 3125 __ breakpoint_trap(Assembler::zero); 3126 #endif 3127 3128 // Now push the new interpreter frames 3129 __ bind(loop); 3130 3131 // allocate a new frame, filling the registers 3132 3133 gen_new_frame(masm, deopt); // allocate an interpreter frame 3134 3135 __ cmp_zero_and_br(Assembler::notZero, O4array_size, loop); 3136 __ delayed()->add(O3array, wordSize, O3array); 3137 __ ld_ptr(G3pcs, 0, O7); // load final frame new pc 3138 3139 } 3140 3141 //------------------------------generate_deopt_blob---------------------------- 3142 // Ought to generate an ideal graph & compile, but here's some SPARC ASM 3143 // instead. 3144 void SharedRuntime::generate_deopt_blob() { 3145 // allocate space for the code 3146 ResourceMark rm; 3147 // setup code generation tools 3148 int pad = VerifyThread ? 512 : 0;// Extra slop space for more verify code 3149 #ifdef _LP64 3150 CodeBuffer buffer("deopt_blob", 2100+pad, 512); 3151 #else 3152 // Measured 8/7/03 at 1212 in 32bit debug build (no VerifyThread) 3153 // Measured 8/7/03 at 1396 in 32bit debug build (VerifyThread) 3154 CodeBuffer buffer("deopt_blob", 1600+pad, 512); 3155 #endif /* _LP64 */ 3156 MacroAssembler* masm = new MacroAssembler(&buffer); 3157 FloatRegister Freturn0 = F0; 3158 Register Greturn1 = G1; 3159 Register Oreturn0 = O0; 3160 Register Oreturn1 = O1; 3161 Register O2UnrollBlock = O2; 3162 Register L0deopt_mode = L0; 3163 Register G4deopt_mode = G4_scratch; 3164 int frame_size_words; 3165 Address saved_Freturn0_addr(FP, -sizeof(double) + STACK_BIAS); 3166 #if !defined(_LP64) && defined(COMPILER2) 3167 Address saved_Greturn1_addr(FP, -sizeof(double) -sizeof(jlong) + STACK_BIAS); 3168 #endif 3169 Label cont; 3170 3171 OopMapSet *oop_maps = new OopMapSet(); 3172 3173 // 3174 // This is the entry point for code which is returning to a de-optimized 3175 // frame. 3176 // The steps taken by this frame are as follows: 3177 // - push a dummy "register_save" and save the return values (O0, O1, F0/F1, G1) 3178 // and all potentially live registers (at a pollpoint many registers can be live). 3179 // 3180 // - call the C routine: Deoptimization::fetch_unroll_info (this function 3181 // returns information about the number and size of interpreter frames 3182 // which are equivalent to the frame which is being deoptimized) 3183 // - deallocate the unpack frame, restoring only results values. Other 3184 // volatile registers will now be captured in the vframeArray as needed. 3185 // - deallocate the deoptimization frame 3186 // - in a loop using the information returned in the previous step 3187 // push new interpreter frames (take care to propagate the return 3188 // values through each new frame pushed) 3189 // - create a dummy "unpack_frame" and save the return values (O0, O1, F0) 3190 // - call the C routine: Deoptimization::unpack_frames (this function 3191 // lays out values on the interpreter frame which was just created) 3192 // - deallocate the dummy unpack_frame 3193 // - ensure that all the return values are correctly set and then do 3194 // a return to the interpreter entry point 3195 // 3196 // Refer to the following methods for more information: 3197 // - Deoptimization::fetch_unroll_info 3198 // - Deoptimization::unpack_frames 3199 3200 OopMap* map = NULL; 3201 3202 int start = __ offset(); 3203 3204 // restore G2, the trampoline destroyed it 3205 __ get_thread(); 3206 3207 // On entry we have been called by the deoptimized nmethod with a call that 3208 // replaced the original call (or safepoint polling location) so the deoptimizing 3209 // pc is now in O7. Return values are still in the expected places 3210 3211 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words); 3212 __ ba(cont); 3213 __ delayed()->mov(Deoptimization::Unpack_deopt, L0deopt_mode); 3214 3215 int exception_offset = __ offset() - start; 3216 3217 // restore G2, the trampoline destroyed it 3218 __ get_thread(); 3219 3220 // On entry we have been jumped to by the exception handler (or exception_blob 3221 // for server). O0 contains the exception oop and O7 contains the original 3222 // exception pc. So if we push a frame here it will look to the 3223 // stack walking code (fetch_unroll_info) just like a normal call so 3224 // state will be extracted normally. 3225 3226 // save exception oop in JavaThread and fall through into the 3227 // exception_in_tls case since they are handled in same way except 3228 // for where the pending exception is kept. 3229 __ st_ptr(Oexception, G2_thread, JavaThread::exception_oop_offset()); 3230 3231 // 3232 // Vanilla deoptimization with an exception pending in exception_oop 3233 // 3234 int exception_in_tls_offset = __ offset() - start; 3235 3236 // No need to update oop_map as each call to save_live_registers will produce identical oopmap 3237 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words); 3238 3239 // Restore G2_thread 3240 __ get_thread(); 3241 3242 #ifdef ASSERT 3243 { 3244 // verify that there is really an exception oop in exception_oop 3245 Label has_exception; 3246 __ ld_ptr(G2_thread, JavaThread::exception_oop_offset(), Oexception); 3247 __ br_notnull_short(Oexception, Assembler::pt, has_exception); 3248 __ stop("no exception in thread"); 3249 __ bind(has_exception); 3250 3251 // verify that there is no pending exception 3252 Label no_pending_exception; 3253 Address exception_addr(G2_thread, Thread::pending_exception_offset()); 3254 __ ld_ptr(exception_addr, Oexception); 3255 __ br_null_short(Oexception, Assembler::pt, no_pending_exception); 3256 __ stop("must not have pending exception here"); 3257 __ bind(no_pending_exception); 3258 } 3259 #endif 3260 3261 __ ba(cont); 3262 __ delayed()->mov(Deoptimization::Unpack_exception, L0deopt_mode);; 3263 3264 // 3265 // Reexecute entry, similar to c2 uncommon trap 3266 // 3267 int reexecute_offset = __ offset() - start; 3268 3269 // No need to update oop_map as each call to save_live_registers will produce identical oopmap 3270 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words); 3271 3272 __ mov(Deoptimization::Unpack_reexecute, L0deopt_mode); 3273 3274 __ bind(cont); 3275 3276 __ set_last_Java_frame(SP, noreg); 3277 3278 // do the call by hand so we can get the oopmap 3279 3280 __ mov(G2_thread, L7_thread_cache); 3281 __ call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), relocInfo::runtime_call_type); 3282 __ delayed()->mov(G2_thread, O0); 3283 3284 // Set an oopmap for the call site this describes all our saved volatile registers 3285 3286 oop_maps->add_gc_map( __ offset()-start, map); 3287 3288 __ mov(L7_thread_cache, G2_thread); 3289 3290 __ reset_last_Java_frame(); 3291 3292 // NOTE: we know that only O0/O1 will be reloaded by restore_result_registers 3293 // so this move will survive 3294 3295 __ mov(L0deopt_mode, G4deopt_mode); 3296 3297 __ mov(O0, O2UnrollBlock->after_save()); 3298 3299 RegisterSaver::restore_result_registers(masm); 3300 3301 Label noException; 3302 __ cmp_and_br_short(G4deopt_mode, Deoptimization::Unpack_exception, Assembler::notEqual, Assembler::pt, noException); 3303 3304 // Move the pending exception from exception_oop to Oexception so 3305 // the pending exception will be picked up the interpreter. 3306 __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception); 3307 __ st_ptr(G0, G2_thread, in_bytes(JavaThread::exception_oop_offset())); 3308 __ bind(noException); 3309 3310 // deallocate the deoptimization frame taking care to preserve the return values 3311 __ mov(Oreturn0, Oreturn0->after_save()); 3312 __ mov(Oreturn1, Oreturn1->after_save()); 3313 __ mov(O2UnrollBlock, O2UnrollBlock->after_save()); 3314 __ restore(); 3315 3316 // Allocate new interpreter frame(s) and possible c2i adapter frame 3317 3318 make_new_frames(masm, true); 3319 3320 // push a dummy "unpack_frame" taking care of float return values and 3321 // call Deoptimization::unpack_frames to have the unpacker layout 3322 // information in the interpreter frames just created and then return 3323 // to the interpreter entry point 3324 __ save(SP, -frame_size_words*wordSize, SP); 3325 __ stf(FloatRegisterImpl::D, Freturn0, saved_Freturn0_addr); 3326 #if !defined(_LP64) 3327 #if defined(COMPILER2) 3328 // 32-bit 1-register longs return longs in G1 3329 __ stx(Greturn1, saved_Greturn1_addr); 3330 #endif 3331 __ set_last_Java_frame(SP, noreg); 3332 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, G4deopt_mode); 3333 #else 3334 // LP64 uses g4 in set_last_Java_frame 3335 __ mov(G4deopt_mode, O1); 3336 __ set_last_Java_frame(SP, G0); 3337 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O1); 3338 #endif 3339 __ reset_last_Java_frame(); 3340 __ ldf(FloatRegisterImpl::D, saved_Freturn0_addr, Freturn0); 3341 3342 #if !defined(_LP64) && defined(COMPILER2) 3343 // In 32 bit, C2 returns longs in G1 so restore the saved G1 into 3344 // I0/I1 if the return value is long. 3345 Label not_long; 3346 __ cmp_and_br_short(O0,T_LONG, Assembler::notEqual, Assembler::pt, not_long); 3347 __ ldd(saved_Greturn1_addr,I0); 3348 __ bind(not_long); 3349 #endif 3350 __ ret(); 3351 __ delayed()->restore(); 3352 3353 masm->flush(); 3354 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_words); 3355 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); 3356 } 3357 3358 #ifdef COMPILER2 3359 3360 //------------------------------generate_uncommon_trap_blob-------------------- 3361 // Ought to generate an ideal graph & compile, but here's some SPARC ASM 3362 // instead. 3363 void SharedRuntime::generate_uncommon_trap_blob() { 3364 // allocate space for the code 3365 ResourceMark rm; 3366 // setup code generation tools 3367 int pad = VerifyThread ? 512 : 0; 3368 #ifdef _LP64 3369 CodeBuffer buffer("uncommon_trap_blob", 2700+pad, 512); 3370 #else 3371 // Measured 8/7/03 at 660 in 32bit debug build (no VerifyThread) 3372 // Measured 8/7/03 at 1028 in 32bit debug build (VerifyThread) 3373 CodeBuffer buffer("uncommon_trap_blob", 2000+pad, 512); 3374 #endif 3375 MacroAssembler* masm = new MacroAssembler(&buffer); 3376 Register O2UnrollBlock = O2; 3377 Register O2klass_index = O2; 3378 3379 // 3380 // This is the entry point for all traps the compiler takes when it thinks 3381 // it cannot handle further execution of compilation code. The frame is 3382 // deoptimized in these cases and converted into interpreter frames for 3383 // execution 3384 // The steps taken by this frame are as follows: 3385 // - push a fake "unpack_frame" 3386 // - call the C routine Deoptimization::uncommon_trap (this function 3387 // packs the current compiled frame into vframe arrays and returns 3388 // information about the number and size of interpreter frames which 3389 // are equivalent to the frame which is being deoptimized) 3390 // - deallocate the "unpack_frame" 3391 // - deallocate the deoptimization frame 3392 // - in a loop using the information returned in the previous step 3393 // push interpreter frames; 3394 // - create a dummy "unpack_frame" 3395 // - call the C routine: Deoptimization::unpack_frames (this function 3396 // lays out values on the interpreter frame which was just created) 3397 // - deallocate the dummy unpack_frame 3398 // - return to the interpreter entry point 3399 // 3400 // Refer to the following methods for more information: 3401 // - Deoptimization::uncommon_trap 3402 // - Deoptimization::unpack_frame 3403 3404 // the unloaded class index is in O0 (first parameter to this blob) 3405 3406 // push a dummy "unpack_frame" 3407 // and call Deoptimization::uncommon_trap to pack the compiled frame into 3408 // vframe array and return the UnrollBlock information 3409 __ save_frame(0); 3410 __ set_last_Java_frame(SP, noreg); 3411 __ mov(I0, O2klass_index); 3412 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), G2_thread, O2klass_index); 3413 __ reset_last_Java_frame(); 3414 __ mov(O0, O2UnrollBlock->after_save()); 3415 __ restore(); 3416 3417 // deallocate the deoptimized frame taking care to preserve the return values 3418 __ mov(O2UnrollBlock, O2UnrollBlock->after_save()); 3419 __ restore(); 3420 3421 // Allocate new interpreter frame(s) and possible c2i adapter frame 3422 3423 make_new_frames(masm, false); 3424 3425 // push a dummy "unpack_frame" taking care of float return values and 3426 // call Deoptimization::unpack_frames to have the unpacker layout 3427 // information in the interpreter frames just created and then return 3428 // to the interpreter entry point 3429 __ save_frame(0); 3430 __ set_last_Java_frame(SP, noreg); 3431 __ mov(Deoptimization::Unpack_uncommon_trap, O3); // indicate it is the uncommon trap case 3432 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O3); 3433 __ reset_last_Java_frame(); 3434 __ ret(); 3435 __ delayed()->restore(); 3436 3437 masm->flush(); 3438 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, NULL, __ total_frame_size_in_bytes(0)/wordSize); 3439 } 3440 3441 #endif // COMPILER2 3442 3443 //------------------------------generate_handler_blob------------------- 3444 // 3445 // Generate a special Compile2Runtime blob that saves all registers, and sets 3446 // up an OopMap. 3447 // 3448 // This blob is jumped to (via a breakpoint and the signal handler) from a 3449 // safepoint in compiled code. On entry to this blob, O7 contains the 3450 // address in the original nmethod at which we should resume normal execution. 3451 // Thus, this blob looks like a subroutine which must preserve lots of 3452 // registers and return normally. Note that O7 is never register-allocated, 3453 // so it is guaranteed to be free here. 3454 // 3455 3456 // The hardest part of what this blob must do is to save the 64-bit %o 3457 // registers in the 32-bit build. A simple 'save' turn the %o's to %i's and 3458 // an interrupt will chop off their heads. Making space in the caller's frame 3459 // first will let us save the 64-bit %o's before save'ing, but we cannot hand 3460 // the adjusted FP off to the GC stack-crawler: this will modify the caller's 3461 // SP and mess up HIS OopMaps. So we first adjust the caller's SP, then save 3462 // the 64-bit %o's, then do a save, then fixup the caller's SP (our FP). 3463 // Tricky, tricky, tricky... 3464 3465 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, bool cause_return) { 3466 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 3467 3468 // allocate space for the code 3469 ResourceMark rm; 3470 // setup code generation tools 3471 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread) 3472 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread) 3473 // even larger with TraceJumps 3474 int pad = TraceJumps ? 512 : 0; 3475 CodeBuffer buffer("handler_blob", 1600 + pad, 512); 3476 MacroAssembler* masm = new MacroAssembler(&buffer); 3477 int frame_size_words; 3478 OopMapSet *oop_maps = new OopMapSet(); 3479 OopMap* map = NULL; 3480 3481 int start = __ offset(); 3482 3483 // If this causes a return before the processing, then do a "restore" 3484 if (cause_return) { 3485 __ restore(); 3486 } else { 3487 // Make it look like we were called via the poll 3488 // so that frame constructor always sees a valid return address 3489 __ ld_ptr(G2_thread, in_bytes(JavaThread::saved_exception_pc_offset()), O7); 3490 __ sub(O7, frame::pc_return_offset, O7); 3491 } 3492 3493 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words); 3494 3495 // setup last_Java_sp (blows G4) 3496 __ set_last_Java_frame(SP, noreg); 3497 3498 // call into the runtime to handle illegal instructions exception 3499 // Do not use call_VM_leaf, because we need to make a GC map at this call site. 3500 __ mov(G2_thread, O0); 3501 __ save_thread(L7_thread_cache); 3502 __ call(call_ptr); 3503 __ delayed()->nop(); 3504 3505 // Set an oopmap for the call site. 3506 // We need this not only for callee-saved registers, but also for volatile 3507 // registers that the compiler might be keeping live across a safepoint. 3508 3509 oop_maps->add_gc_map( __ offset() - start, map); 3510 3511 __ restore_thread(L7_thread_cache); 3512 // clear last_Java_sp 3513 __ reset_last_Java_frame(); 3514 3515 // Check for exceptions 3516 Label pending; 3517 3518 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1); 3519 __ br_notnull_short(O1, Assembler::pn, pending); 3520 3521 RegisterSaver::restore_live_registers(masm); 3522 3523 // We are back the the original state on entry and ready to go. 3524 3525 __ retl(); 3526 __ delayed()->nop(); 3527 3528 // Pending exception after the safepoint 3529 3530 __ bind(pending); 3531 3532 RegisterSaver::restore_live_registers(masm); 3533 3534 // We are back the the original state on entry. 3535 3536 // Tail-call forward_exception_entry, with the issuing PC in O7, 3537 // so it looks like the original nmethod called forward_exception_entry. 3538 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0); 3539 __ JMP(O0, 0); 3540 __ delayed()->nop(); 3541 3542 // ------------- 3543 // make sure all code is generated 3544 masm->flush(); 3545 3546 // return exception blob 3547 return SafepointBlob::create(&buffer, oop_maps, frame_size_words); 3548 } 3549 3550 // 3551 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss 3552 // 3553 // Generate a stub that calls into vm to find out the proper destination 3554 // of a java call. All the argument registers are live at this point 3555 // but since this is generic code we don't know what they are and the caller 3556 // must do any gc of the args. 3557 // 3558 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) { 3559 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); 3560 3561 // allocate space for the code 3562 ResourceMark rm; 3563 // setup code generation tools 3564 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread) 3565 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread) 3566 // even larger with TraceJumps 3567 int pad = TraceJumps ? 512 : 0; 3568 CodeBuffer buffer(name, 1600 + pad, 512); 3569 MacroAssembler* masm = new MacroAssembler(&buffer); 3570 int frame_size_words; 3571 OopMapSet *oop_maps = new OopMapSet(); 3572 OopMap* map = NULL; 3573 3574 int start = __ offset(); 3575 3576 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words); 3577 3578 int frame_complete = __ offset(); 3579 3580 // setup last_Java_sp (blows G4) 3581 __ set_last_Java_frame(SP, noreg); 3582 3583 // call into the runtime to handle illegal instructions exception 3584 // Do not use call_VM_leaf, because we need to make a GC map at this call site. 3585 __ mov(G2_thread, O0); 3586 __ save_thread(L7_thread_cache); 3587 __ call(destination, relocInfo::runtime_call_type); 3588 __ delayed()->nop(); 3589 3590 // O0 contains the address we are going to jump to assuming no exception got installed 3591 3592 // Set an oopmap for the call site. 3593 // We need this not only for callee-saved registers, but also for volatile 3594 // registers that the compiler might be keeping live across a safepoint. 3595 3596 oop_maps->add_gc_map( __ offset() - start, map); 3597 3598 __ restore_thread(L7_thread_cache); 3599 // clear last_Java_sp 3600 __ reset_last_Java_frame(); 3601 3602 // Check for exceptions 3603 Label pending; 3604 3605 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1); 3606 __ br_notnull_short(O1, Assembler::pn, pending); 3607 3608 // get the returned methodOop 3609 3610 __ get_vm_result(G5_method); 3611 __ stx(G5_method, SP, RegisterSaver::G5_offset()+STACK_BIAS); 3612 3613 // O0 is where we want to jump, overwrite G3 which is saved and scratch 3614 3615 __ stx(O0, SP, RegisterSaver::G3_offset()+STACK_BIAS); 3616 3617 RegisterSaver::restore_live_registers(masm); 3618 3619 // We are back the the original state on entry and ready to go. 3620 3621 __ JMP(G3, 0); 3622 __ delayed()->nop(); 3623 3624 // Pending exception after the safepoint 3625 3626 __ bind(pending); 3627 3628 RegisterSaver::restore_live_registers(masm); 3629 3630 // We are back the the original state on entry. 3631 3632 // Tail-call forward_exception_entry, with the issuing PC in O7, 3633 // so it looks like the original nmethod called forward_exception_entry. 3634 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0); 3635 __ JMP(O0, 0); 3636 __ delayed()->nop(); 3637 3638 // ------------- 3639 // make sure all code is generated 3640 masm->flush(); 3641 3642 // return the blob 3643 // frame_size_words or bytes?? 3644 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true); 3645 }