annotate src/cpu/sparc/vm/sharedRuntime_sparc.cpp @ 1869:f95d63e2154a

6989984: Use standard include model for Hospot Summary: Replaced MakeDeps and the includeDB files with more standardized solutions. Reviewed-by: coleenp, kvn, kamg
author stefank
date Tue, 23 Nov 2010 13:22:55 -0800
parents d5d065957597
children 3582bf76420e
rev   line source
duke@0 1 /*
trims@1489 2 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
trims@1489 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1489 20 * or visit www.oracle.com if you need additional information or have any
trims@1489 21 * questions.
duke@0 22 *
duke@0 23 */
duke@0 24
stefank@1869 25 #include "precompiled.hpp"
stefank@1869 26 #include "asm/assembler.hpp"
stefank@1869 27 #include "assembler_sparc.inline.hpp"
stefank@1869 28 #include "code/debugInfoRec.hpp"
stefank@1869 29 #include "code/icBuffer.hpp"
stefank@1869 30 #include "code/vtableStubs.hpp"
stefank@1869 31 #include "interpreter/interpreter.hpp"
stefank@1869 32 #include "oops/compiledICHolderOop.hpp"
stefank@1869 33 #include "prims/jvmtiRedefineClassesTrace.hpp"
stefank@1869 34 #include "runtime/sharedRuntime.hpp"
stefank@1869 35 #include "runtime/vframeArray.hpp"
stefank@1869 36 #include "vmreg_sparc.inline.hpp"
stefank@1869 37 #ifdef COMPILER1
stefank@1869 38 #include "c1/c1_Runtime1.hpp"
stefank@1869 39 #endif
stefank@1869 40 #ifdef COMPILER2
stefank@1869 41 #include "opto/runtime.hpp"
stefank@1869 42 #endif
stefank@1869 43 #ifdef SHARK
stefank@1869 44 #include "compiler/compileBroker.hpp"
stefank@1869 45 #include "shark/sharkCompiler.hpp"
stefank@1869 46 #endif
duke@0 47
duke@0 48 #define __ masm->
duke@0 49
duke@0 50 #ifdef COMPILER2
duke@0 51 UncommonTrapBlob* SharedRuntime::_uncommon_trap_blob;
duke@0 52 #endif // COMPILER2
duke@0 53
duke@0 54 DeoptimizationBlob* SharedRuntime::_deopt_blob;
duke@0 55 SafepointBlob* SharedRuntime::_polling_page_safepoint_handler_blob;
duke@0 56 SafepointBlob* SharedRuntime::_polling_page_return_handler_blob;
duke@0 57 RuntimeStub* SharedRuntime::_wrong_method_blob;
duke@0 58 RuntimeStub* SharedRuntime::_ic_miss_blob;
duke@0 59 RuntimeStub* SharedRuntime::_resolve_opt_virtual_call_blob;
duke@0 60 RuntimeStub* SharedRuntime::_resolve_virtual_call_blob;
duke@0 61 RuntimeStub* SharedRuntime::_resolve_static_call_blob;
duke@0 62
duke@0 63 class RegisterSaver {
duke@0 64
duke@0 65 // Used for saving volatile registers. This is Gregs, Fregs, I/L/O.
duke@0 66 // The Oregs are problematic. In the 32bit build the compiler can
duke@0 67 // have O registers live with 64 bit quantities. A window save will
duke@0 68 // cut the heads off of the registers. We have to do a very extensive
duke@0 69 // stack dance to save and restore these properly.
duke@0 70
duke@0 71 // Note that the Oregs problem only exists if we block at either a polling
duke@0 72 // page exception a compiled code safepoint that was not originally a call
duke@0 73 // or deoptimize following one of these kinds of safepoints.
duke@0 74
duke@0 75 // Lots of registers to save. For all builds, a window save will preserve
duke@0 76 // the %i and %l registers. For the 32-bit longs-in-two entries and 64-bit
duke@0 77 // builds a window-save will preserve the %o registers. In the LION build
duke@0 78 // we need to save the 64-bit %o registers which requires we save them
duke@0 79 // before the window-save (as then they become %i registers and get their
duke@0 80 // heads chopped off on interrupt). We have to save some %g registers here
duke@0 81 // as well.
duke@0 82 enum {
duke@0 83 // This frame's save area. Includes extra space for the native call:
duke@0 84 // vararg's layout space and the like. Briefly holds the caller's
duke@0 85 // register save area.
duke@0 86 call_args_area = frame::register_save_words_sp_offset +
duke@0 87 frame::memory_parameter_word_sp_offset*wordSize,
duke@0 88 // Make sure save locations are always 8 byte aligned.
duke@0 89 // can't use round_to because it doesn't produce compile time constant
duke@0 90 start_of_extra_save_area = ((call_args_area + 7) & ~7),
duke@0 91 g1_offset = start_of_extra_save_area, // g-regs needing saving
duke@0 92 g3_offset = g1_offset+8,
duke@0 93 g4_offset = g3_offset+8,
duke@0 94 g5_offset = g4_offset+8,
duke@0 95 o0_offset = g5_offset+8,
duke@0 96 o1_offset = o0_offset+8,
duke@0 97 o2_offset = o1_offset+8,
duke@0 98 o3_offset = o2_offset+8,
duke@0 99 o4_offset = o3_offset+8,
duke@0 100 o5_offset = o4_offset+8,
duke@0 101 start_of_flags_save_area = o5_offset+8,
duke@0 102 ccr_offset = start_of_flags_save_area,
duke@0 103 fsr_offset = ccr_offset + 8,
duke@0 104 d00_offset = fsr_offset+8, // Start of float save area
duke@0 105 register_save_size = d00_offset+8*32
duke@0 106 };
duke@0 107
duke@0 108
duke@0 109 public:
duke@0 110
duke@0 111 static int Oexception_offset() { return o0_offset; };
duke@0 112 static int G3_offset() { return g3_offset; };
duke@0 113 static int G5_offset() { return g5_offset; };
duke@0 114 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
duke@0 115 static void restore_live_registers(MacroAssembler* masm);
duke@0 116
duke@0 117 // During deoptimization only the result register need to be restored
duke@0 118 // all the other values have already been extracted.
duke@0 119
duke@0 120 static void restore_result_registers(MacroAssembler* masm);
duke@0 121 };
duke@0 122
duke@0 123 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
duke@0 124 // Record volatile registers as callee-save values in an OopMap so their save locations will be
duke@0 125 // propagated to the caller frame's RegisterMap during StackFrameStream construction (needed for
duke@0 126 // deoptimization; see compiledVFrame::create_stack_value). The caller's I, L and O registers
duke@0 127 // are saved in register windows - I's and L's in the caller's frame and O's in the stub frame
duke@0 128 // (as the stub's I's) when the runtime routine called by the stub creates its frame.
duke@0 129 int i;
kvn@992 130 // Always make the frame size 16 byte aligned.
duke@0 131 int frame_size = round_to(additional_frame_words + register_save_size, 16);
duke@0 132 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words
duke@0 133 int frame_size_in_slots = frame_size / sizeof(jint);
duke@0 134 // CodeBlob frame size is in words.
duke@0 135 *total_frame_words = frame_size / wordSize;
duke@0 136 // OopMap* map = new OopMap(*total_frame_words, 0);
duke@0 137 OopMap* map = new OopMap(frame_size_in_slots, 0);
duke@0 138
duke@0 139 #if !defined(_LP64)
duke@0 140
duke@0 141 // Save 64-bit O registers; they will get their heads chopped off on a 'save'.
duke@0 142 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
duke@0 143 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
duke@0 144 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
duke@0 145 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
duke@0 146 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
duke@0 147 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
duke@0 148 #endif /* _LP64 */
duke@0 149
duke@0 150 __ save(SP, -frame_size, SP);
duke@0 151
duke@0 152 #ifndef _LP64
duke@0 153 // Reload the 64 bit Oregs. Although they are now Iregs we load them
duke@0 154 // to Oregs here to avoid interrupts cutting off their heads
duke@0 155
duke@0 156 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
duke@0 157 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
duke@0 158 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
duke@0 159 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
duke@0 160 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
duke@0 161 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
duke@0 162
duke@0 163 __ stx(O0, SP, o0_offset+STACK_BIAS);
duke@0 164 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset + 4)>>2), O0->as_VMReg());
duke@0 165
duke@0 166 __ stx(O1, SP, o1_offset+STACK_BIAS);
duke@0 167
duke@0 168 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset + 4)>>2), O1->as_VMReg());
duke@0 169
duke@0 170 __ stx(O2, SP, o2_offset+STACK_BIAS);
duke@0 171 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset + 4)>>2), O2->as_VMReg());
duke@0 172
duke@0 173 __ stx(O3, SP, o3_offset+STACK_BIAS);
duke@0 174 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset + 4)>>2), O3->as_VMReg());
duke@0 175
duke@0 176 __ stx(O4, SP, o4_offset+STACK_BIAS);
duke@0 177 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset + 4)>>2), O4->as_VMReg());
duke@0 178
duke@0 179 __ stx(O5, SP, o5_offset+STACK_BIAS);
duke@0 180 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset + 4)>>2), O5->as_VMReg());
duke@0 181 #endif /* _LP64 */
duke@0 182
coleenp@108 183
coleenp@108 184 #ifdef _LP64
coleenp@108 185 int debug_offset = 0;
coleenp@108 186 #else
coleenp@108 187 int debug_offset = 4;
coleenp@108 188 #endif
duke@0 189 // Save the G's
duke@0 190 __ stx(G1, SP, g1_offset+STACK_BIAS);
coleenp@108 191 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset + debug_offset)>>2), G1->as_VMReg());
duke@0 192
duke@0 193 __ stx(G3, SP, g3_offset+STACK_BIAS);
coleenp@108 194 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset + debug_offset)>>2), G3->as_VMReg());
duke@0 195
duke@0 196 __ stx(G4, SP, g4_offset+STACK_BIAS);
coleenp@108 197 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset + debug_offset)>>2), G4->as_VMReg());
duke@0 198
duke@0 199 __ stx(G5, SP, g5_offset+STACK_BIAS);
coleenp@108 200 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset + debug_offset)>>2), G5->as_VMReg());
duke@0 201
duke@0 202 // This is really a waste but we'll keep things as they were for now
duke@0 203 if (true) {
duke@0 204 #ifndef _LP64
duke@0 205 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset)>>2), O0->as_VMReg()->next());
duke@0 206 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset)>>2), O1->as_VMReg()->next());
duke@0 207 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset)>>2), O2->as_VMReg()->next());
duke@0 208 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset)>>2), O3->as_VMReg()->next());
duke@0 209 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset)>>2), O4->as_VMReg()->next());
duke@0 210 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset)>>2), O5->as_VMReg()->next());
duke@0 211 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset)>>2), G1->as_VMReg()->next());
duke@0 212 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset)>>2), G3->as_VMReg()->next());
duke@0 213 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset)>>2), G4->as_VMReg()->next());
duke@0 214 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset)>>2), G5->as_VMReg()->next());
coleenp@108 215 #endif /* _LP64 */
duke@0 216 }
duke@0 217
duke@0 218
duke@0 219 // Save the flags
duke@0 220 __ rdccr( G5 );
duke@0 221 __ stx(G5, SP, ccr_offset+STACK_BIAS);
duke@0 222 __ stxfsr(SP, fsr_offset+STACK_BIAS);
duke@0 223
kvn@992 224 // Save all the FP registers: 32 doubles (32 floats correspond to the 2 halves of the first 16 doubles)
duke@0 225 int offset = d00_offset;
kvn@992 226 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) {
duke@0 227 FloatRegister f = as_FloatRegister(i);
duke@0 228 __ stf(FloatRegisterImpl::D, f, SP, offset+STACK_BIAS);
kvn@992 229 // Record as callee saved both halves of double registers (2 float registers).
duke@0 230 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), f->as_VMReg());
kvn@992 231 map->set_callee_saved(VMRegImpl::stack2reg((offset + sizeof(float))>>2), f->as_VMReg()->next());
duke@0 232 offset += sizeof(double);
duke@0 233 }
duke@0 234
duke@0 235 // And we're done.
duke@0 236
duke@0 237 return map;
duke@0 238 }
duke@0 239
duke@0 240
duke@0 241 // Pop the current frame and restore all the registers that we
duke@0 242 // saved.
duke@0 243 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
duke@0 244
duke@0 245 // Restore all the FP registers
kvn@992 246 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) {
duke@0 247 __ ldf(FloatRegisterImpl::D, SP, d00_offset+i*sizeof(float)+STACK_BIAS, as_FloatRegister(i));
duke@0 248 }
duke@0 249
duke@0 250 __ ldx(SP, ccr_offset+STACK_BIAS, G1);
duke@0 251 __ wrccr (G1) ;
duke@0 252
duke@0 253 // Restore the G's
duke@0 254 // Note that G2 (AKA GThread) must be saved and restored separately.
duke@0 255 // TODO-FIXME: save and restore some of the other ASRs, viz., %asi and %gsr.
duke@0 256
duke@0 257 __ ldx(SP, g1_offset+STACK_BIAS, G1);
duke@0 258 __ ldx(SP, g3_offset+STACK_BIAS, G3);
duke@0 259 __ ldx(SP, g4_offset+STACK_BIAS, G4);
duke@0 260 __ ldx(SP, g5_offset+STACK_BIAS, G5);
duke@0 261
duke@0 262
duke@0 263 #if !defined(_LP64)
duke@0 264 // Restore the 64-bit O's.
duke@0 265 __ ldx(SP, o0_offset+STACK_BIAS, O0);
duke@0 266 __ ldx(SP, o1_offset+STACK_BIAS, O1);
duke@0 267 __ ldx(SP, o2_offset+STACK_BIAS, O2);
duke@0 268 __ ldx(SP, o3_offset+STACK_BIAS, O3);
duke@0 269 __ ldx(SP, o4_offset+STACK_BIAS, O4);
duke@0 270 __ ldx(SP, o5_offset+STACK_BIAS, O5);
duke@0 271
duke@0 272 // And temporarily place them in TLS
duke@0 273
duke@0 274 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
duke@0 275 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
duke@0 276 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
duke@0 277 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
duke@0 278 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
duke@0 279 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
duke@0 280 #endif /* _LP64 */
duke@0 281
duke@0 282 // Restore flags
duke@0 283
duke@0 284 __ ldxfsr(SP, fsr_offset+STACK_BIAS);
duke@0 285
duke@0 286 __ restore();
duke@0 287
duke@0 288 #if !defined(_LP64)
duke@0 289 // Now reload the 64bit Oregs after we've restore the window.
duke@0 290 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
duke@0 291 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
duke@0 292 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
duke@0 293 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
duke@0 294 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
duke@0 295 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
duke@0 296 #endif /* _LP64 */
duke@0 297
duke@0 298 }
duke@0 299
duke@0 300 // Pop the current frame and restore the registers that might be holding
duke@0 301 // a result.
duke@0 302 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
duke@0 303
duke@0 304 #if !defined(_LP64)
duke@0 305 // 32bit build returns longs in G1
duke@0 306 __ ldx(SP, g1_offset+STACK_BIAS, G1);
duke@0 307
duke@0 308 // Retrieve the 64-bit O's.
duke@0 309 __ ldx(SP, o0_offset+STACK_BIAS, O0);
duke@0 310 __ ldx(SP, o1_offset+STACK_BIAS, O1);
duke@0 311 // and save to TLS
duke@0 312 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
duke@0 313 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
duke@0 314 #endif /* _LP64 */
duke@0 315
duke@0 316 __ ldf(FloatRegisterImpl::D, SP, d00_offset+STACK_BIAS, as_FloatRegister(0));
duke@0 317
duke@0 318 __ restore();
duke@0 319
duke@0 320 #if !defined(_LP64)
duke@0 321 // Now reload the 64bit Oregs after we've restore the window.
duke@0 322 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
duke@0 323 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
duke@0 324 #endif /* _LP64 */
duke@0 325
duke@0 326 }
duke@0 327
duke@0 328 // The java_calling_convention describes stack locations as ideal slots on
duke@0 329 // a frame with no abi restrictions. Since we must observe abi restrictions
duke@0 330 // (like the placement of the register window) the slots must be biased by
duke@0 331 // the following value.
duke@0 332 static int reg2offset(VMReg r) {
duke@0 333 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
duke@0 334 }
duke@0 335
duke@0 336 // ---------------------------------------------------------------------------
duke@0 337 // Read the array of BasicTypes from a signature, and compute where the
duke@0 338 // arguments should go. Values in the VMRegPair regs array refer to 4-byte (VMRegImpl::stack_slot_size)
duke@0 339 // quantities. Values less than VMRegImpl::stack0 are registers, those above
duke@0 340 // refer to 4-byte stack slots. All stack slots are based off of the window
duke@0 341 // top. VMRegImpl::stack0 refers to the first slot past the 16-word window,
duke@0 342 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register
duke@0 343 // values 0-63 (up to RegisterImpl::number_of_registers) are the 64-bit
duke@0 344 // integer registers. Values 64-95 are the (32-bit only) float registers.
duke@0 345 // Each 32-bit quantity is given its own number, so the integer registers
duke@0 346 // (in either 32- or 64-bit builds) use 2 numbers. For example, there is
duke@0 347 // an O0-low and an O0-high. Essentially, all int register numbers are doubled.
duke@0 348
duke@0 349 // Register results are passed in O0-O5, for outgoing call arguments. To
duke@0 350 // convert to incoming arguments, convert all O's to I's. The regs array
duke@0 351 // refer to the low and hi 32-bit words of 64-bit registers or stack slots.
duke@0 352 // If the regs[].second() field is set to VMRegImpl::Bad(), it means it's unused (a
duke@0 353 // 32-bit value was passed). If both are VMRegImpl::Bad(), it means no value was
duke@0 354 // passed (used as a placeholder for the other half of longs and doubles in
duke@0 355 // the 64-bit build). regs[].second() is either VMRegImpl::Bad() or regs[].second() is
duke@0 356 // regs[].first()+1 (regs[].first() may be misaligned in the C calling convention).
duke@0 357 // Sparc never passes a value in regs[].second() but not regs[].first() (regs[].first()
duke@0 358 // == VMRegImpl::Bad() && regs[].second() != VMRegImpl::Bad()) nor unrelated values in the
duke@0 359 // same VMRegPair.
duke@0 360
duke@0 361 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
duke@0 362 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
duke@0 363 // units regardless of build.
duke@0 364
duke@0 365
duke@0 366 // ---------------------------------------------------------------------------
duke@0 367 // The compiled Java calling convention. The Java convention always passes
duke@0 368 // 64-bit values in adjacent aligned locations (either registers or stack),
duke@0 369 // floats in float registers and doubles in aligned float pairs. Values are
duke@0 370 // packed in the registers. There is no backing varargs store for values in
duke@0 371 // registers. In the 32-bit build, longs are passed in G1 and G4 (cannot be
duke@0 372 // passed in I's, because longs in I's get their heads chopped off at
duke@0 373 // interrupt).
duke@0 374 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
duke@0 375 VMRegPair *regs,
duke@0 376 int total_args_passed,
duke@0 377 int is_outgoing) {
duke@0 378 assert(F31->as_VMReg()->is_reg(), "overlapping stack/register numbers");
duke@0 379
duke@0 380 // Convention is to pack the first 6 int/oop args into the first 6 registers
duke@0 381 // (I0-I5), extras spill to the stack. Then pack the first 8 float args
duke@0 382 // into F0-F7, extras spill to the stack. Then pad all register sets to
duke@0 383 // align. Then put longs and doubles into the same registers as they fit,
duke@0 384 // else spill to the stack.
duke@0 385 const int int_reg_max = SPARC_ARGS_IN_REGS_NUM;
duke@0 386 const int flt_reg_max = 8;
duke@0 387 //
duke@0 388 // Where 32-bit 1-reg longs start being passed
duke@0 389 // In tiered we must pass on stack because c1 can't use a "pair" in a single reg.
duke@0 390 // So make it look like we've filled all the G regs that c2 wants to use.
duke@0 391 Register g_reg = TieredCompilation ? noreg : G1;
duke@0 392
duke@0 393 // Count int/oop and float args. See how many stack slots we'll need and
duke@0 394 // where the longs & doubles will go.
duke@0 395 int int_reg_cnt = 0;
duke@0 396 int flt_reg_cnt = 0;
duke@0 397 // int stk_reg_pairs = frame::register_save_words*(wordSize>>2);
duke@0 398 // int stk_reg_pairs = SharedRuntime::out_preserve_stack_slots();
duke@0 399 int stk_reg_pairs = 0;
duke@0 400 for (int i = 0; i < total_args_passed; i++) {
duke@0 401 switch (sig_bt[i]) {
duke@0 402 case T_LONG: // LP64, longs compete with int args
duke@0 403 assert(sig_bt[i+1] == T_VOID, "");
duke@0 404 #ifdef _LP64
duke@0 405 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
duke@0 406 #endif
duke@0 407 break;
duke@0 408 case T_OBJECT:
duke@0 409 case T_ARRAY:
duke@0 410 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
duke@0 411 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
duke@0 412 #ifndef _LP64
duke@0 413 else stk_reg_pairs++;
duke@0 414 #endif
duke@0 415 break;
duke@0 416 case T_INT:
duke@0 417 case T_SHORT:
duke@0 418 case T_CHAR:
duke@0 419 case T_BYTE:
duke@0 420 case T_BOOLEAN:
duke@0 421 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
duke@0 422 else stk_reg_pairs++;
duke@0 423 break;
duke@0 424 case T_FLOAT:
duke@0 425 if (flt_reg_cnt < flt_reg_max) flt_reg_cnt++;
duke@0 426 else stk_reg_pairs++;
duke@0 427 break;
duke@0 428 case T_DOUBLE:
duke@0 429 assert(sig_bt[i+1] == T_VOID, "");
duke@0 430 break;
duke@0 431 case T_VOID:
duke@0 432 break;
duke@0 433 default:
duke@0 434 ShouldNotReachHere();
duke@0 435 }
duke@0 436 }
duke@0 437
duke@0 438 // This is where the longs/doubles start on the stack.
duke@0 439 stk_reg_pairs = (stk_reg_pairs+1) & ~1; // Round
duke@0 440
duke@0 441 int int_reg_pairs = (int_reg_cnt+1) & ~1; // 32-bit 2-reg longs only
duke@0 442 int flt_reg_pairs = (flt_reg_cnt+1) & ~1;
duke@0 443
duke@0 444 // int stk_reg = frame::register_save_words*(wordSize>>2);
duke@0 445 // int stk_reg = SharedRuntime::out_preserve_stack_slots();
duke@0 446 int stk_reg = 0;
duke@0 447 int int_reg = 0;
duke@0 448 int flt_reg = 0;
duke@0 449
duke@0 450 // Now do the signature layout
duke@0 451 for (int i = 0; i < total_args_passed; i++) {
duke@0 452 switch (sig_bt[i]) {
duke@0 453 case T_INT:
duke@0 454 case T_SHORT:
duke@0 455 case T_CHAR:
duke@0 456 case T_BYTE:
duke@0 457 case T_BOOLEAN:
duke@0 458 #ifndef _LP64
duke@0 459 case T_OBJECT:
duke@0 460 case T_ARRAY:
duke@0 461 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
duke@0 462 #endif // _LP64
duke@0 463 if (int_reg < int_reg_max) {
duke@0 464 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
duke@0 465 regs[i].set1(r->as_VMReg());
duke@0 466 } else {
duke@0 467 regs[i].set1(VMRegImpl::stack2reg(stk_reg++));
duke@0 468 }
duke@0 469 break;
duke@0 470
duke@0 471 #ifdef _LP64
duke@0 472 case T_OBJECT:
duke@0 473 case T_ARRAY:
duke@0 474 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
duke@0 475 if (int_reg < int_reg_max) {
duke@0 476 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
duke@0 477 regs[i].set2(r->as_VMReg());
duke@0 478 } else {
duke@0 479 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
duke@0 480 stk_reg_pairs += 2;
duke@0 481 }
duke@0 482 break;
duke@0 483 #endif // _LP64
duke@0 484
duke@0 485 case T_LONG:
duke@0 486 assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half");
duke@0 487 #ifdef _LP64
duke@0 488 if (int_reg < int_reg_max) {
duke@0 489 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
duke@0 490 regs[i].set2(r->as_VMReg());
duke@0 491 } else {
duke@0 492 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
duke@0 493 stk_reg_pairs += 2;
duke@0 494 }
duke@0 495 #else
never@297 496 #ifdef COMPILER2
duke@0 497 // For 32-bit build, can't pass longs in O-regs because they become
duke@0 498 // I-regs and get trashed. Use G-regs instead. G1 and G4 are almost
duke@0 499 // spare and available. This convention isn't used by the Sparc ABI or
duke@0 500 // anywhere else. If we're tiered then we don't use G-regs because c1
never@297 501 // can't deal with them as a "pair". (Tiered makes this code think g's are filled)
duke@0 502 // G0: zero
duke@0 503 // G1: 1st Long arg
duke@0 504 // G2: global allocated to TLS
duke@0 505 // G3: used in inline cache check
duke@0 506 // G4: 2nd Long arg
duke@0 507 // G5: used in inline cache check
duke@0 508 // G6: used by OS
duke@0 509 // G7: used by OS
duke@0 510
duke@0 511 if (g_reg == G1) {
duke@0 512 regs[i].set2(G1->as_VMReg()); // This long arg in G1
duke@0 513 g_reg = G4; // Where the next arg goes
duke@0 514 } else if (g_reg == G4) {
duke@0 515 regs[i].set2(G4->as_VMReg()); // The 2nd long arg in G4
duke@0 516 g_reg = noreg; // No more longs in registers
duke@0 517 } else {
duke@0 518 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
duke@0 519 stk_reg_pairs += 2;
duke@0 520 }
duke@0 521 #else // COMPILER2
duke@0 522 if (int_reg_pairs + 1 < int_reg_max) {
duke@0 523 if (is_outgoing) {
duke@0 524 regs[i].set_pair(as_oRegister(int_reg_pairs + 1)->as_VMReg(), as_oRegister(int_reg_pairs)->as_VMReg());
duke@0 525 } else {
duke@0 526 regs[i].set_pair(as_iRegister(int_reg_pairs + 1)->as_VMReg(), as_iRegister(int_reg_pairs)->as_VMReg());
duke@0 527 }
duke@0 528 int_reg_pairs += 2;
duke@0 529 } else {
duke@0 530 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
duke@0 531 stk_reg_pairs += 2;
duke@0 532 }
duke@0 533 #endif // COMPILER2
never@297 534 #endif // _LP64
duke@0 535 break;
duke@0 536
duke@0 537 case T_FLOAT:
duke@0 538 if (flt_reg < flt_reg_max) regs[i].set1(as_FloatRegister(flt_reg++)->as_VMReg());
duke@0 539 else regs[i].set1( VMRegImpl::stack2reg(stk_reg++));
duke@0 540 break;
duke@0 541 case T_DOUBLE:
duke@0 542 assert(sig_bt[i+1] == T_VOID, "expecting half");
duke@0 543 if (flt_reg_pairs + 1 < flt_reg_max) {
duke@0 544 regs[i].set2(as_FloatRegister(flt_reg_pairs)->as_VMReg());
duke@0 545 flt_reg_pairs += 2;
duke@0 546 } else {
duke@0 547 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
duke@0 548 stk_reg_pairs += 2;
duke@0 549 }
duke@0 550 break;
duke@0 551 case T_VOID: regs[i].set_bad(); break; // Halves of longs & doubles
duke@0 552 default:
duke@0 553 ShouldNotReachHere();
duke@0 554 }
duke@0 555 }
duke@0 556
duke@0 557 // retun the amount of stack space these arguments will need.
duke@0 558 return stk_reg_pairs;
duke@0 559
duke@0 560 }
duke@0 561
twisti@991 562 // Helper class mostly to avoid passing masm everywhere, and handle
twisti@991 563 // store displacement overflow logic.
duke@0 564 class AdapterGenerator {
duke@0 565 MacroAssembler *masm;
duke@0 566 Register Rdisp;
duke@0 567 void set_Rdisp(Register r) { Rdisp = r; }
duke@0 568
duke@0 569 void patch_callers_callsite();
duke@0 570
duke@0 571 // base+st_off points to top of argument
twisti@1401 572 int arg_offset(const int st_off) { return st_off; }
duke@0 573 int next_arg_offset(const int st_off) {
twisti@1401 574 return st_off - Interpreter::stackElementSize;
twisti@991 575 }
twisti@991 576
twisti@991 577 // Argument slot values may be loaded first into a register because
twisti@991 578 // they might not fit into displacement.
twisti@991 579 RegisterOrConstant arg_slot(const int st_off);
twisti@991 580 RegisterOrConstant next_arg_slot(const int st_off);
twisti@991 581
duke@0 582 // Stores long into offset pointed to by base
duke@0 583 void store_c2i_long(Register r, Register base,
duke@0 584 const int st_off, bool is_stack);
duke@0 585 void store_c2i_object(Register r, Register base,
duke@0 586 const int st_off);
duke@0 587 void store_c2i_int(Register r, Register base,
duke@0 588 const int st_off);
duke@0 589 void store_c2i_double(VMReg r_2,
duke@0 590 VMReg r_1, Register base, const int st_off);
duke@0 591 void store_c2i_float(FloatRegister f, Register base,
duke@0 592 const int st_off);
duke@0 593
duke@0 594 public:
duke@0 595 void gen_c2i_adapter(int total_args_passed,
duke@0 596 // VMReg max_arg,
duke@0 597 int comp_args_on_stack, // VMRegStackSlots
duke@0 598 const BasicType *sig_bt,
duke@0 599 const VMRegPair *regs,
duke@0 600 Label& skip_fixup);
duke@0 601 void gen_i2c_adapter(int total_args_passed,
duke@0 602 // VMReg max_arg,
duke@0 603 int comp_args_on_stack, // VMRegStackSlots
duke@0 604 const BasicType *sig_bt,
duke@0 605 const VMRegPair *regs);
duke@0 606
duke@0 607 AdapterGenerator(MacroAssembler *_masm) : masm(_masm) {}
duke@0 608 };
duke@0 609
duke@0 610
duke@0 611 // Patch the callers callsite with entry to compiled code if it exists.
duke@0 612 void AdapterGenerator::patch_callers_callsite() {
duke@0 613 Label L;
duke@0 614 __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch);
duke@0 615 __ br_null(G3_scratch, false, __ pt, L);
duke@0 616 // Schedule the branch target address early.
duke@0 617 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
duke@0 618 // Call into the VM to patch the caller, then jump to compiled callee
duke@0 619 __ save_frame(4); // Args in compiled layout; do not blow them
duke@0 620
duke@0 621 // Must save all the live Gregs the list is:
duke@0 622 // G1: 1st Long arg (32bit build)
duke@0 623 // G2: global allocated to TLS
duke@0 624 // G3: used in inline cache check (scratch)
duke@0 625 // G4: 2nd Long arg (32bit build);
duke@0 626 // G5: used in inline cache check (methodOop)
duke@0 627
duke@0 628 // The longs must go to the stack by hand since in the 32 bit build they can be trashed by window ops.
duke@0 629
duke@0 630 #ifdef _LP64
duke@0 631 // mov(s,d)
duke@0 632 __ mov(G1, L1);
duke@0 633 __ mov(G4, L4);
duke@0 634 __ mov(G5_method, L5);
duke@0 635 __ mov(G5_method, O0); // VM needs target method
duke@0 636 __ mov(I7, O1); // VM needs caller's callsite
duke@0 637 // Must be a leaf call...
duke@0 638 // can be very far once the blob has been relocated
twisti@720 639 AddressLiteral dest(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
duke@0 640 __ relocate(relocInfo::runtime_call_type);
twisti@720 641 __ jumpl_to(dest, O7, O7);
duke@0 642 __ delayed()->mov(G2_thread, L7_thread_cache);
duke@0 643 __ mov(L7_thread_cache, G2_thread);
duke@0 644 __ mov(L1, G1);
duke@0 645 __ mov(L4, G4);
duke@0 646 __ mov(L5, G5_method);
duke@0 647 #else
duke@0 648 __ stx(G1, FP, -8 + STACK_BIAS);
duke@0 649 __ stx(G4, FP, -16 + STACK_BIAS);
duke@0 650 __ mov(G5_method, L5);
duke@0 651 __ mov(G5_method, O0); // VM needs target method
duke@0 652 __ mov(I7, O1); // VM needs caller's callsite
duke@0 653 // Must be a leaf call...
duke@0 654 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), relocInfo::runtime_call_type);
duke@0 655 __ delayed()->mov(G2_thread, L7_thread_cache);
duke@0 656 __ mov(L7_thread_cache, G2_thread);
duke@0 657 __ ldx(FP, -8 + STACK_BIAS, G1);
duke@0 658 __ ldx(FP, -16 + STACK_BIAS, G4);
duke@0 659 __ mov(L5, G5_method);
duke@0 660 __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
duke@0 661 #endif /* _LP64 */
duke@0 662
duke@0 663 __ restore(); // Restore args
duke@0 664 __ bind(L);
duke@0 665 }
duke@0 666
twisti@991 667
twisti@991 668 RegisterOrConstant AdapterGenerator::arg_slot(const int st_off) {
twisti@991 669 RegisterOrConstant roc(arg_offset(st_off));
twisti@991 670 return __ ensure_simm13_or_reg(roc, Rdisp);
duke@0 671 }
duke@0 672
twisti@991 673 RegisterOrConstant AdapterGenerator::next_arg_slot(const int st_off) {
twisti@991 674 RegisterOrConstant roc(next_arg_offset(st_off));
twisti@991 675 return __ ensure_simm13_or_reg(roc, Rdisp);
duke@0 676 }
twisti@991 677
twisti@991 678
duke@0 679 // Stores long into offset pointed to by base
duke@0 680 void AdapterGenerator::store_c2i_long(Register r, Register base,
duke@0 681 const int st_off, bool is_stack) {
duke@0 682 #ifdef _LP64
duke@0 683 // In V9, longs are given 2 64-bit slots in the interpreter, but the
duke@0 684 // data is passed in only 1 slot.
duke@0 685 __ stx(r, base, next_arg_slot(st_off));
duke@0 686 #else
ysr@344 687 #ifdef COMPILER2
duke@0 688 // Misaligned store of 64-bit data
duke@0 689 __ stw(r, base, arg_slot(st_off)); // lo bits
duke@0 690 __ srlx(r, 32, r);
duke@0 691 __ stw(r, base, next_arg_slot(st_off)); // hi bits
duke@0 692 #else
duke@0 693 if (is_stack) {
duke@0 694 // Misaligned store of 64-bit data
duke@0 695 __ stw(r, base, arg_slot(st_off)); // lo bits
duke@0 696 __ srlx(r, 32, r);
duke@0 697 __ stw(r, base, next_arg_slot(st_off)); // hi bits
duke@0 698 } else {
duke@0 699 __ stw(r->successor(), base, arg_slot(st_off) ); // lo bits
duke@0 700 __ stw(r , base, next_arg_slot(st_off)); // hi bits
duke@0 701 }
duke@0 702 #endif // COMPILER2
ysr@344 703 #endif // _LP64
duke@0 704 }
duke@0 705
duke@0 706 void AdapterGenerator::store_c2i_object(Register r, Register base,
duke@0 707 const int st_off) {
duke@0 708 __ st_ptr (r, base, arg_slot(st_off));
duke@0 709 }
duke@0 710
duke@0 711 void AdapterGenerator::store_c2i_int(Register r, Register base,
duke@0 712 const int st_off) {
duke@0 713 __ st (r, base, arg_slot(st_off));
duke@0 714 }
duke@0 715
duke@0 716 // Stores into offset pointed to by base
duke@0 717 void AdapterGenerator::store_c2i_double(VMReg r_2,
duke@0 718 VMReg r_1, Register base, const int st_off) {
duke@0 719 #ifdef _LP64
duke@0 720 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
duke@0 721 // data is passed in only 1 slot.
duke@0 722 __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
duke@0 723 #else
duke@0 724 // Need to marshal 64-bit value from misaligned Lesp loads
duke@0 725 __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
duke@0 726 __ stf(FloatRegisterImpl::S, r_2->as_FloatRegister(), base, arg_slot(st_off) );
duke@0 727 #endif
duke@0 728 }
duke@0 729
duke@0 730 void AdapterGenerator::store_c2i_float(FloatRegister f, Register base,
duke@0 731 const int st_off) {
duke@0 732 __ stf(FloatRegisterImpl::S, f, base, arg_slot(st_off));
duke@0 733 }
duke@0 734
duke@0 735 void AdapterGenerator::gen_c2i_adapter(
duke@0 736 int total_args_passed,
duke@0 737 // VMReg max_arg,
duke@0 738 int comp_args_on_stack, // VMRegStackSlots
duke@0 739 const BasicType *sig_bt,
duke@0 740 const VMRegPair *regs,
duke@0 741 Label& skip_fixup) {
duke@0 742
duke@0 743 // Before we get into the guts of the C2I adapter, see if we should be here
duke@0 744 // at all. We've come from compiled code and are attempting to jump to the
duke@0 745 // interpreter, which means the caller made a static call to get here
duke@0 746 // (vcalls always get a compiled target if there is one). Check for a
duke@0 747 // compiled target. If there is one, we need to patch the caller's call.
duke@0 748 // However we will run interpreted if we come thru here. The next pass
duke@0 749 // thru the call site will run compiled. If we ran compiled here then
duke@0 750 // we can (theorectically) do endless i2c->c2i->i2c transitions during
duke@0 751 // deopt/uncommon trap cycles. If we always go interpreted here then
duke@0 752 // we can have at most one and don't need to play any tricks to keep
duke@0 753 // from endlessly growing the stack.
duke@0 754 //
duke@0 755 // Actually if we detected that we had an i2c->c2i transition here we
duke@0 756 // ought to be able to reset the world back to the state of the interpreted
duke@0 757 // call and not bother building another interpreter arg area. We don't
duke@0 758 // do that at this point.
duke@0 759
duke@0 760 patch_callers_callsite();
duke@0 761
duke@0 762 __ bind(skip_fixup);
duke@0 763
duke@0 764 // Since all args are passed on the stack, total_args_passed*wordSize is the
duke@0 765 // space we need. Add in varargs area needed by the interpreter. Round up
duke@0 766 // to stack alignment.
twisti@1401 767 const int arg_size = total_args_passed * Interpreter::stackElementSize;
duke@0 768 const int varargs_area =
duke@0 769 (frame::varargs_offset - frame::register_save_words)*wordSize;
duke@0 770 const int extraspace = round_to(arg_size + varargs_area, 2*wordSize);
duke@0 771
duke@0 772 int bias = STACK_BIAS;
duke@0 773 const int interp_arg_offset = frame::varargs_offset*wordSize +
twisti@1401 774 (total_args_passed-1)*Interpreter::stackElementSize;
duke@0 775
duke@0 776 Register base = SP;
duke@0 777
duke@0 778 #ifdef _LP64
duke@0 779 // In the 64bit build because of wider slots and STACKBIAS we can run
duke@0 780 // out of bits in the displacement to do loads and stores. Use g3 as
duke@0 781 // temporary displacement.
duke@0 782 if (! __ is_simm13(extraspace)) {
duke@0 783 __ set(extraspace, G3_scratch);
duke@0 784 __ sub(SP, G3_scratch, SP);
duke@0 785 } else {
duke@0 786 __ sub(SP, extraspace, SP);
duke@0 787 }
duke@0 788 set_Rdisp(G3_scratch);
duke@0 789 #else
duke@0 790 __ sub(SP, extraspace, SP);
duke@0 791 #endif // _LP64
duke@0 792
duke@0 793 // First write G1 (if used) to where ever it must go
duke@0 794 for (int i=0; i<total_args_passed; i++) {
twisti@1401 795 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize) + bias;
duke@0 796 VMReg r_1 = regs[i].first();
duke@0 797 VMReg r_2 = regs[i].second();
duke@0 798 if (r_1 == G1_scratch->as_VMReg()) {
duke@0 799 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) {
duke@0 800 store_c2i_object(G1_scratch, base, st_off);
duke@0 801 } else if (sig_bt[i] == T_LONG) {
duke@0 802 assert(!TieredCompilation, "should not use register args for longs");
duke@0 803 store_c2i_long(G1_scratch, base, st_off, false);
duke@0 804 } else {
duke@0 805 store_c2i_int(G1_scratch, base, st_off);
duke@0 806 }
duke@0 807 }
duke@0 808 }
duke@0 809
duke@0 810 // Now write the args into the outgoing interpreter space
duke@0 811 for (int i=0; i<total_args_passed; i++) {
twisti@1401 812 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize) + bias;
duke@0 813 VMReg r_1 = regs[i].first();
duke@0 814 VMReg r_2 = regs[i].second();
duke@0 815 if (!r_1->is_valid()) {
duke@0 816 assert(!r_2->is_valid(), "");
duke@0 817 continue;
duke@0 818 }
duke@0 819 // Skip G1 if found as we did it first in order to free it up
duke@0 820 if (r_1 == G1_scratch->as_VMReg()) {
duke@0 821 continue;
duke@0 822 }
duke@0 823 #ifdef ASSERT
duke@0 824 bool G1_forced = false;
duke@0 825 #endif // ASSERT
duke@0 826 if (r_1->is_stack()) { // Pretend stack targets are loaded into G1
duke@0 827 #ifdef _LP64
duke@0 828 Register ld_off = Rdisp;
duke@0 829 __ set(reg2offset(r_1) + extraspace + bias, ld_off);
duke@0 830 #else
duke@0 831 int ld_off = reg2offset(r_1) + extraspace + bias;
kvn@1209 832 #endif // _LP64
duke@0 833 #ifdef ASSERT
duke@0 834 G1_forced = true;
duke@0 835 #endif // ASSERT
duke@0 836 r_1 = G1_scratch->as_VMReg();// as part of the load/store shuffle
duke@0 837 if (!r_2->is_valid()) __ ld (base, ld_off, G1_scratch);
duke@0 838 else __ ldx(base, ld_off, G1_scratch);
duke@0 839 }
duke@0 840
duke@0 841 if (r_1->is_Register()) {
duke@0 842 Register r = r_1->as_Register()->after_restore();
duke@0 843 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) {
duke@0 844 store_c2i_object(r, base, st_off);
duke@0 845 } else if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
kvn@1209 846 #ifndef _LP64
duke@0 847 if (TieredCompilation) {
duke@0 848 assert(G1_forced || sig_bt[i] != T_LONG, "should not use register args for longs");
duke@0 849 }
kvn@1209 850 #endif // _LP64
duke@0 851 store_c2i_long(r, base, st_off, r_2->is_stack());
duke@0 852 } else {
duke@0 853 store_c2i_int(r, base, st_off);
duke@0 854 }
duke@0 855 } else {
duke@0 856 assert(r_1->is_FloatRegister(), "");
duke@0 857 if (sig_bt[i] == T_FLOAT) {
duke@0 858 store_c2i_float(r_1->as_FloatRegister(), base, st_off);
duke@0 859 } else {
duke@0 860 assert(sig_bt[i] == T_DOUBLE, "wrong type");
duke@0 861 store_c2i_double(r_2, r_1, base, st_off);
duke@0 862 }
duke@0 863 }
duke@0 864 }
duke@0 865
duke@0 866 #ifdef _LP64
duke@0 867 // Need to reload G3_scratch, used for temporary displacements.
duke@0 868 __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
duke@0 869
duke@0 870 // Pass O5_savedSP as an argument to the interpreter.
duke@0 871 // The interpreter will restore SP to this value before returning.
duke@0 872 __ set(extraspace, G1);
duke@0 873 __ add(SP, G1, O5_savedSP);
duke@0 874 #else
duke@0 875 // Pass O5_savedSP as an argument to the interpreter.
duke@0 876 // The interpreter will restore SP to this value before returning.
duke@0 877 __ add(SP, extraspace, O5_savedSP);
duke@0 878 #endif // _LP64
duke@0 879
duke@0 880 __ mov((frame::varargs_offset)*wordSize -
twisti@1401 881 1*Interpreter::stackElementSize+bias+BytesPerWord, G1);
duke@0 882 // Jump to the interpreter just as if interpreter was doing it.
duke@0 883 __ jmpl(G3_scratch, 0, G0);
duke@0 884 // Setup Lesp for the call. Cannot actually set Lesp as the current Lesp
duke@0 885 // (really L0) is in use by the compiled frame as a generic temp. However,
duke@0 886 // the interpreter does not know where its args are without some kind of
duke@0 887 // arg pointer being passed in. Pass it in Gargs.
duke@0 888 __ delayed()->add(SP, G1, Gargs);
duke@0 889 }
duke@0 890
duke@0 891 void AdapterGenerator::gen_i2c_adapter(
duke@0 892 int total_args_passed,
duke@0 893 // VMReg max_arg,
duke@0 894 int comp_args_on_stack, // VMRegStackSlots
duke@0 895 const BasicType *sig_bt,
duke@0 896 const VMRegPair *regs) {
duke@0 897
duke@0 898 // Generate an I2C adapter: adjust the I-frame to make space for the C-frame
duke@0 899 // layout. Lesp was saved by the calling I-frame and will be restored on
duke@0 900 // return. Meanwhile, outgoing arg space is all owned by the callee
duke@0 901 // C-frame, so we can mangle it at will. After adjusting the frame size,
duke@0 902 // hoist register arguments and repack other args according to the compiled
duke@0 903 // code convention. Finally, end in a jump to the compiled code. The entry
duke@0 904 // point address is the start of the buffer.
duke@0 905
duke@0 906 // We will only enter here from an interpreted frame and never from after
duke@0 907 // passing thru a c2i. Azul allowed this but we do not. If we lose the
duke@0 908 // race and use a c2i we will remain interpreted for the race loser(s).
duke@0 909 // This removes all sorts of headaches on the x86 side and also eliminates
duke@0 910 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
duke@0 911
duke@0 912 // As you can see from the list of inputs & outputs there are not a lot
duke@0 913 // of temp registers to work with: mostly G1, G3 & G4.
duke@0 914
duke@0 915 // Inputs:
duke@0 916 // G2_thread - TLS
duke@0 917 // G5_method - Method oop
jrose@689 918 // G4 (Gargs) - Pointer to interpreter's args
jrose@689 919 // O0..O4 - free for scratch
jrose@689 920 // O5_savedSP - Caller's saved SP, to be restored if needed
duke@0 921 // O6 - Current SP!
duke@0 922 // O7 - Valid return address
jrose@689 923 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
duke@0 924
duke@0 925 // Outputs:
duke@0 926 // G2_thread - TLS
duke@0 927 // G1, G4 - Outgoing long args in 32-bit build
duke@0 928 // O0-O5 - Outgoing args in compiled layout
duke@0 929 // O6 - Adjusted or restored SP
duke@0 930 // O7 - Valid return address
twisti@1457 931 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
duke@0 932 // F0-F7 - more outgoing args
duke@0 933
duke@0 934
jrose@689 935 // Gargs is the incoming argument base, and also an outgoing argument.
duke@0 936 __ sub(Gargs, BytesPerWord, Gargs);
duke@0 937
duke@0 938 // ON ENTRY TO THE CODE WE ARE MAKING, WE HAVE AN INTERPRETED FRAME
duke@0 939 // WITH O7 HOLDING A VALID RETURN PC
duke@0 940 //
duke@0 941 // | |
duke@0 942 // : java stack :
duke@0 943 // | |
duke@0 944 // +--------------+ <--- start of outgoing args
duke@0 945 // | receiver | |
duke@0 946 // : rest of args : |---size is java-arg-words
duke@0 947 // | | |
duke@0 948 // +--------------+ <--- O4_args (misaligned) and Lesp if prior is not C2I
duke@0 949 // | | |
duke@0 950 // : unused : |---Space for max Java stack, plus stack alignment
duke@0 951 // | | |
duke@0 952 // +--------------+ <--- SP + 16*wordsize
duke@0 953 // | |
duke@0 954 // : window :
duke@0 955 // | |
duke@0 956 // +--------------+ <--- SP
duke@0 957
duke@0 958 // WE REPACK THE STACK. We use the common calling convention layout as
duke@0 959 // discovered by calling SharedRuntime::calling_convention. We assume it
duke@0 960 // causes an arbitrary shuffle of memory, which may require some register
duke@0 961 // temps to do the shuffle. We hope for (and optimize for) the case where
duke@0 962 // temps are not needed. We may have to resize the stack slightly, in case
duke@0 963 // we need alignment padding (32-bit interpreter can pass longs & doubles
duke@0 964 // misaligned, but the compilers expect them aligned).
duke@0 965 //
duke@0 966 // | |
duke@0 967 // : java stack :
duke@0 968 // | |
duke@0 969 // +--------------+ <--- start of outgoing args
duke@0 970 // | pad, align | |
duke@0 971 // +--------------+ |
duke@0 972 // | ints, floats | |---Outgoing stack args, packed low.
duke@0 973 // +--------------+ | First few args in registers.
duke@0 974 // : doubles : |
duke@0 975 // | longs | |
duke@0 976 // +--------------+ <--- SP' + 16*wordsize
duke@0 977 // | |
duke@0 978 // : window :
duke@0 979 // | |
duke@0 980 // +--------------+ <--- SP'
duke@0 981
duke@0 982 // ON EXIT FROM THE CODE WE ARE MAKING, WE STILL HAVE AN INTERPRETED FRAME
duke@0 983 // WITH O7 HOLDING A VALID RETURN PC - ITS JUST THAT THE ARGS ARE NOW SETUP
duke@0 984 // FOR COMPILED CODE AND THE FRAME SLIGHTLY GROWN.
duke@0 985
duke@0 986 // Cut-out for having no stack args. Since up to 6 args are passed
duke@0 987 // in registers, we will commonly have no stack args.
duke@0 988 if (comp_args_on_stack > 0) {
duke@0 989
duke@0 990 // Convert VMReg stack slots to words.
duke@0 991 int comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
duke@0 992 // Round up to miminum stack alignment, in wordSize
duke@0 993 comp_words_on_stack = round_to(comp_words_on_stack, 2);
duke@0 994 // Now compute the distance from Lesp to SP. This calculation does not
duke@0 995 // include the space for total_args_passed because Lesp has not yet popped
duke@0 996 // the arguments.
duke@0 997 __ sub(SP, (comp_words_on_stack)*wordSize, SP);
duke@0 998 }
duke@0 999
duke@0 1000 // Will jump to the compiled code just as if compiled code was doing it.
duke@0 1001 // Pre-load the register-jump target early, to schedule it better.
duke@0 1002 __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_compiled_offset()), G3);
duke@0 1003
duke@0 1004 // Now generate the shuffle code. Pick up all register args and move the
duke@0 1005 // rest through G1_scratch.
duke@0 1006 for (int i=0; i<total_args_passed; i++) {
duke@0 1007 if (sig_bt[i] == T_VOID) {
duke@0 1008 // Longs and doubles are passed in native word order, but misaligned
duke@0 1009 // in the 32-bit build.
duke@0 1010 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
duke@0 1011 continue;
duke@0 1012 }
duke@0 1013
duke@0 1014 // Pick up 0, 1 or 2 words from Lesp+offset. Assume mis-aligned in the
duke@0 1015 // 32-bit build and aligned in the 64-bit build. Look for the obvious
duke@0 1016 // ldx/lddf optimizations.
duke@0 1017
duke@0 1018 // Load in argument order going down.
twisti@1401 1019 const int ld_off = (total_args_passed-i)*Interpreter::stackElementSize;
duke@0 1020 set_Rdisp(G1_scratch);
duke@0 1021
duke@0 1022 VMReg r_1 = regs[i].first();
duke@0 1023 VMReg r_2 = regs[i].second();
duke@0 1024 if (!r_1->is_valid()) {
duke@0 1025 assert(!r_2->is_valid(), "");
duke@0 1026 continue;
duke@0 1027 }
duke@0 1028 if (r_1->is_stack()) { // Pretend stack targets are loaded into F8/F9
duke@0 1029 r_1 = F8->as_VMReg(); // as part of the load/store shuffle
duke@0 1030 if (r_2->is_valid()) r_2 = r_1->next();
duke@0 1031 }
duke@0 1032 if (r_1->is_Register()) { // Register argument
duke@0 1033 Register r = r_1->as_Register()->after_restore();
duke@0 1034 if (!r_2->is_valid()) {
duke@0 1035 __ ld(Gargs, arg_slot(ld_off), r);
duke@0 1036 } else {
duke@0 1037 #ifdef _LP64
duke@0 1038 // In V9, longs are given 2 64-bit slots in the interpreter, but the
duke@0 1039 // data is passed in only 1 slot.
twisti@991 1040 RegisterOrConstant slot = (sig_bt[i] == T_LONG) ?
duke@0 1041 next_arg_slot(ld_off) : arg_slot(ld_off);
duke@0 1042 __ ldx(Gargs, slot, r);
duke@0 1043 #else
duke@0 1044 // Need to load a 64-bit value into G1/G4, but G1/G4 is being used in the
duke@0 1045 // stack shuffle. Load the first 2 longs into G1/G4 later.
duke@0 1046 #endif
duke@0 1047 }
duke@0 1048 } else {
duke@0 1049 assert(r_1->is_FloatRegister(), "");
duke@0 1050 if (!r_2->is_valid()) {
duke@0 1051 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_1->as_FloatRegister());
duke@0 1052 } else {
duke@0 1053 #ifdef _LP64
duke@0 1054 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
duke@0 1055 // data is passed in only 1 slot. This code also handles longs that
duke@0 1056 // are passed on the stack, but need a stack-to-stack move through a
duke@0 1057 // spare float register.
twisti@991 1058 RegisterOrConstant slot = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
duke@0 1059 next_arg_slot(ld_off) : arg_slot(ld_off);
duke@0 1060 __ ldf(FloatRegisterImpl::D, Gargs, slot, r_1->as_FloatRegister());
duke@0 1061 #else
duke@0 1062 // Need to marshal 64-bit value from misaligned Lesp loads
duke@0 1063 __ ldf(FloatRegisterImpl::S, Gargs, next_arg_slot(ld_off), r_1->as_FloatRegister());
duke@0 1064 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_2->as_FloatRegister());
duke@0 1065 #endif
duke@0 1066 }
duke@0 1067 }
duke@0 1068 // Was the argument really intended to be on the stack, but was loaded
duke@0 1069 // into F8/F9?
duke@0 1070 if (regs[i].first()->is_stack()) {
duke@0 1071 assert(r_1->as_FloatRegister() == F8, "fix this code");
duke@0 1072 // Convert stack slot to an SP offset
duke@0 1073 int st_off = reg2offset(regs[i].first()) + STACK_BIAS;
duke@0 1074 // Store down the shuffled stack word. Target address _is_ aligned.
twisti@991 1075 RegisterOrConstant slot = __ ensure_simm13_or_reg(st_off, Rdisp);
twisti@991 1076 if (!r_2->is_valid()) __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), SP, slot);
twisti@991 1077 else __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), SP, slot);
duke@0 1078 }
duke@0 1079 }
duke@0 1080 bool made_space = false;
duke@0 1081 #ifndef _LP64
duke@0 1082 // May need to pick up a few long args in G1/G4
duke@0 1083 bool g4_crushed = false;
duke@0 1084 bool g3_crushed = false;
duke@0 1085 for (int i=0; i<total_args_passed; i++) {
duke@0 1086 if (regs[i].first()->is_Register() && regs[i].second()->is_valid()) {
duke@0 1087 // Load in argument order going down
twisti@1401 1088 int ld_off = (total_args_passed-i)*Interpreter::stackElementSize;
duke@0 1089 // Need to marshal 64-bit value from misaligned Lesp loads
duke@0 1090 Register r = regs[i].first()->as_Register()->after_restore();
duke@0 1091 if (r == G1 || r == G4) {
duke@0 1092 assert(!g4_crushed, "ordering problem");
duke@0 1093 if (r == G4){
duke@0 1094 g4_crushed = true;
duke@0 1095 __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits
duke@0 1096 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
duke@0 1097 } else {
duke@0 1098 // better schedule this way
duke@0 1099 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
duke@0 1100 __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits
duke@0 1101 }
duke@0 1102 g3_crushed = true;
duke@0 1103 __ sllx(r, 32, r);
duke@0 1104 __ or3(G3_scratch, r, r);
duke@0 1105 } else {
duke@0 1106 assert(r->is_out(), "longs passed in two O registers");
duke@0 1107 __ ld (Gargs, arg_slot(ld_off) , r->successor()); // Load lo bits
duke@0 1108 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
duke@0 1109 }
duke@0 1110 }
duke@0 1111 }
duke@0 1112 #endif
duke@0 1113
duke@0 1114 // Jump to the compiled code just as if compiled code was doing it.
duke@0 1115 //
duke@0 1116 #ifndef _LP64
duke@0 1117 if (g3_crushed) {
duke@0 1118 // Rats load was wasted, at least it is in cache...
twisti@720 1119 __ ld_ptr(G5_method, methodOopDesc::from_compiled_offset(), G3);
duke@0 1120 }
duke@0 1121 #endif /* _LP64 */
duke@0 1122
duke@0 1123 // 6243940 We might end up in handle_wrong_method if
duke@0 1124 // the callee is deoptimized as we race thru here. If that
duke@0 1125 // happens we don't want to take a safepoint because the
duke@0 1126 // caller frame will look interpreted and arguments are now
duke@0 1127 // "compiled" so it is much better to make this transition
duke@0 1128 // invisible to the stack walking code. Unfortunately if
duke@0 1129 // we try and find the callee by normal means a safepoint
duke@0 1130 // is possible. So we stash the desired callee in the thread
duke@0 1131 // and the vm will find there should this case occur.
twisti@720 1132 Address callee_target_addr(G2_thread, JavaThread::callee_target_offset());
duke@0 1133 __ st_ptr(G5_method, callee_target_addr);
duke@0 1134
duke@0 1135 if (StressNonEntrant) {
duke@0 1136 // Open a big window for deopt failure
duke@0 1137 __ save_frame(0);
duke@0 1138 __ mov(G0, L0);
duke@0 1139 Label loop;
duke@0 1140 __ bind(loop);
duke@0 1141 __ sub(L0, 1, L0);
duke@0 1142 __ br_null(L0, false, Assembler::pt, loop);
duke@0 1143 __ delayed()->nop();
duke@0 1144
duke@0 1145 __ restore();
duke@0 1146 }
duke@0 1147
duke@0 1148
duke@0 1149 __ jmpl(G3, 0, G0);
duke@0 1150 __ delayed()->nop();
duke@0 1151 }
duke@0 1152
duke@0 1153 // ---------------------------------------------------------------
duke@0 1154 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
duke@0 1155 int total_args_passed,
duke@0 1156 // VMReg max_arg,
duke@0 1157 int comp_args_on_stack, // VMRegStackSlots
duke@0 1158 const BasicType *sig_bt,
never@1179 1159 const VMRegPair *regs,
never@1179 1160 AdapterFingerPrint* fingerprint) {
duke@0 1161 address i2c_entry = __ pc();
duke@0 1162
duke@0 1163 AdapterGenerator agen(masm);
duke@0 1164
duke@0 1165 agen.gen_i2c_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs);
duke@0 1166
duke@0 1167
duke@0 1168 // -------------------------------------------------------------------------
duke@0 1169 // Generate a C2I adapter. On entry we know G5 holds the methodOop. The
duke@0 1170 // args start out packed in the compiled layout. They need to be unpacked
duke@0 1171 // into the interpreter layout. This will almost always require some stack
duke@0 1172 // space. We grow the current (compiled) stack, then repack the args. We
duke@0 1173 // finally end in a jump to the generic interpreter entry point. On exit
duke@0 1174 // from the interpreter, the interpreter will restore our SP (lest the
duke@0 1175 // compiled code, which relys solely on SP and not FP, get sick).
duke@0 1176
duke@0 1177 address c2i_unverified_entry = __ pc();
duke@0 1178 Label skip_fixup;
duke@0 1179 {
duke@0 1180 #if !defined(_LP64) && defined(COMPILER2)
duke@0 1181 Register R_temp = L0; // another scratch register
duke@0 1182 #else
duke@0 1183 Register R_temp = G1; // another scratch register
duke@0 1184 #endif
duke@0 1185
twisti@720 1186 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
duke@0 1187
duke@0 1188 __ verify_oop(O0);
duke@0 1189 __ verify_oop(G5_method);
coleenp@108 1190 __ load_klass(O0, G3_scratch);
duke@0 1191 __ verify_oop(G3_scratch);
duke@0 1192
duke@0 1193 #if !defined(_LP64) && defined(COMPILER2)
duke@0 1194 __ save(SP, -frame::register_save_words*wordSize, SP);
duke@0 1195 __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp);
duke@0 1196 __ verify_oop(R_temp);
duke@0 1197 __ cmp(G3_scratch, R_temp);
duke@0 1198 __ restore();
duke@0 1199 #else
duke@0 1200 __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp);
duke@0 1201 __ verify_oop(R_temp);
duke@0 1202 __ cmp(G3_scratch, R_temp);
duke@0 1203 #endif
duke@0 1204
duke@0 1205 Label ok, ok2;
duke@0 1206 __ brx(Assembler::equal, false, Assembler::pt, ok);
duke@0 1207 __ delayed()->ld_ptr(G5_method, compiledICHolderOopDesc::holder_method_offset(), G5_method);
twisti@720 1208 __ jump_to(ic_miss, G3_scratch);
duke@0 1209 __ delayed()->nop();
duke@0 1210
duke@0 1211 __ bind(ok);
duke@0 1212 // Method might have been compiled since the call site was patched to
duke@0 1213 // interpreted if that is the case treat it as a miss so we can get
duke@0 1214 // the call site corrected.
duke@0 1215 __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch);
duke@0 1216 __ bind(ok2);
duke@0 1217 __ br_null(G3_scratch, false, __ pt, skip_fixup);
duke@0 1218 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
twisti@720 1219 __ jump_to(ic_miss, G3_scratch);
duke@0 1220 __ delayed()->nop();
duke@0 1221
duke@0 1222 }
duke@0 1223
duke@0 1224 address c2i_entry = __ pc();
duke@0 1225
duke@0 1226 agen.gen_c2i_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
duke@0 1227
duke@0 1228 __ flush();
never@1179 1229 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry);
duke@0 1230
duke@0 1231 }
duke@0 1232
duke@0 1233 // Helper function for native calling conventions
duke@0 1234 static VMReg int_stk_helper( int i ) {
duke@0 1235 // Bias any stack based VMReg we get by ignoring the window area
duke@0 1236 // but not the register parameter save area.
duke@0 1237 //
duke@0 1238 // This is strange for the following reasons. We'd normally expect
duke@0 1239 // the calling convention to return an VMReg for a stack slot
duke@0 1240 // completely ignoring any abi reserved area. C2 thinks of that
duke@0 1241 // abi area as only out_preserve_stack_slots. This does not include
duke@0 1242 // the area allocated by the C abi to store down integer arguments
duke@0 1243 // because the java calling convention does not use it. So
duke@0 1244 // since c2 assumes that there are only out_preserve_stack_slots
duke@0 1245 // to bias the optoregs (which impacts VMRegs) when actually referencing any actual stack
duke@0 1246 // location the c calling convention must add in this bias amount
duke@0 1247 // to make up for the fact that the out_preserve_stack_slots is
duke@0 1248 // insufficient for C calls. What a mess. I sure hope those 6
duke@0 1249 // stack words were worth it on every java call!
duke@0 1250
duke@0 1251 // Another way of cleaning this up would be for out_preserve_stack_slots
duke@0 1252 // to take a parameter to say whether it was C or java calling conventions.
duke@0 1253 // Then things might look a little better (but not much).
duke@0 1254
duke@0 1255 int mem_parm_offset = i - SPARC_ARGS_IN_REGS_NUM;
duke@0 1256 if( mem_parm_offset < 0 ) {
duke@0 1257 return as_oRegister(i)->as_VMReg();
duke@0 1258 } else {
duke@0 1259 int actual_offset = (mem_parm_offset + frame::memory_parameter_word_sp_offset) * VMRegImpl::slots_per_word;
duke@0 1260 // Now return a biased offset that will be correct when out_preserve_slots is added back in
duke@0 1261 return VMRegImpl::stack2reg(actual_offset - SharedRuntime::out_preserve_stack_slots());
duke@0 1262 }
duke@0 1263 }
duke@0 1264
duke@0 1265
duke@0 1266 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
duke@0 1267 VMRegPair *regs,
duke@0 1268 int total_args_passed) {
duke@0 1269
duke@0 1270 // Return the number of VMReg stack_slots needed for the args.
duke@0 1271 // This value does not include an abi space (like register window
duke@0 1272 // save area).
duke@0 1273
duke@0 1274 // The native convention is V8 if !LP64
duke@0 1275 // The LP64 convention is the V9 convention which is slightly more sane.
duke@0 1276
duke@0 1277 // We return the amount of VMReg stack slots we need to reserve for all
duke@0 1278 // the arguments NOT counting out_preserve_stack_slots. Since we always
duke@0 1279 // have space for storing at least 6 registers to memory we start with that.
duke@0 1280 // See int_stk_helper for a further discussion.
duke@0 1281 int max_stack_slots = (frame::varargs_offset * VMRegImpl::slots_per_word) - SharedRuntime::out_preserve_stack_slots();
duke@0 1282
duke@0 1283 #ifdef _LP64
duke@0 1284 // V9 convention: All things "as-if" on double-wide stack slots.
duke@0 1285 // Hoist any int/ptr/long's in the first 6 to int regs.
duke@0 1286 // Hoist any flt/dbl's in the first 16 dbl regs.
duke@0 1287 int j = 0; // Count of actual args, not HALVES
duke@0 1288 for( int i=0; i<total_args_passed; i++, j++ ) {
duke@0 1289 switch( sig_bt[i] ) {
duke@0 1290 case T_BOOLEAN:
duke@0 1291 case T_BYTE:
duke@0 1292 case T_CHAR:
duke@0 1293 case T_INT:
duke@0 1294 case T_SHORT:
duke@0 1295 regs[i].set1( int_stk_helper( j ) ); break;
duke@0 1296 case T_LONG:
duke@0 1297 assert( sig_bt[i+1] == T_VOID, "expecting half" );
duke@0 1298 case T_ADDRESS: // raw pointers, like current thread, for VM calls
duke@0 1299 case T_ARRAY:
duke@0 1300 case T_OBJECT:
duke@0 1301 regs[i].set2( int_stk_helper( j ) );
duke@0 1302 break;
duke@0 1303 case T_FLOAT:
duke@0 1304 if ( j < 16 ) {
duke@0 1305 // V9ism: floats go in ODD registers
duke@0 1306 regs[i].set1(as_FloatRegister(1 + (j<<1))->as_VMReg());
duke@0 1307 } else {
duke@0 1308 // V9ism: floats go in ODD stack slot
duke@0 1309 regs[i].set1(VMRegImpl::stack2reg(1 + (j<<1)));
duke@0 1310 }
duke@0 1311 break;
duke@0 1312 case T_DOUBLE:
duke@0 1313 assert( sig_bt[i+1] == T_VOID, "expecting half" );
duke@0 1314 if ( j < 16 ) {
duke@0 1315 // V9ism: doubles go in EVEN/ODD regs
duke@0 1316 regs[i].set2(as_FloatRegister(j<<1)->as_VMReg());
duke@0 1317 } else {
duke@0 1318 // V9ism: doubles go in EVEN/ODD stack slots
duke@0 1319 regs[i].set2(VMRegImpl::stack2reg(j<<1));
duke@0 1320 }
duke@0 1321 break;
duke@0 1322 case T_VOID: regs[i].set_bad(); j--; break; // Do not count HALVES
duke@0 1323 default:
duke@0 1324 ShouldNotReachHere();
duke@0 1325 }
duke@0 1326 if (regs[i].first()->is_stack()) {
duke@0 1327 int off = regs[i].first()->reg2stack();
duke@0 1328 if (off > max_stack_slots) max_stack_slots = off;
duke@0 1329 }
duke@0 1330 if (regs[i].second()->is_stack()) {
duke@0 1331 int off = regs[i].second()->reg2stack();
duke@0 1332 if (off > max_stack_slots) max_stack_slots = off;
duke@0 1333 }
duke@0 1334 }
duke@0 1335
duke@0 1336 #else // _LP64
duke@0 1337 // V8 convention: first 6 things in O-regs, rest on stack.
duke@0 1338 // Alignment is willy-nilly.
duke@0 1339 for( int i=0; i<total_args_passed; i++ ) {
duke@0 1340 switch( sig_bt[i] ) {
duke@0 1341 case T_ADDRESS: // raw pointers, like current thread, for VM calls
duke@0 1342 case T_ARRAY:
duke@0 1343 case T_BOOLEAN:
duke@0 1344 case T_BYTE:
duke@0 1345 case T_CHAR:
duke@0 1346 case T_FLOAT:
duke@0 1347 case T_INT:
duke@0 1348 case T_OBJECT:
duke@0 1349 case T_SHORT:
duke@0 1350 regs[i].set1( int_stk_helper( i ) );
duke@0 1351 break;
duke@0 1352 case T_DOUBLE:
duke@0 1353 case T_LONG:
duke@0 1354 assert( sig_bt[i+1] == T_VOID, "expecting half" );
duke@0 1355 regs[i].set_pair( int_stk_helper( i+1 ), int_stk_helper( i ) );
duke@0 1356 break;
duke@0 1357 case T_VOID: regs[i].set_bad(); break;
duke@0 1358 default:
duke@0 1359 ShouldNotReachHere();
duke@0 1360 }
duke@0 1361 if (regs[i].first()->is_stack()) {
duke@0 1362 int off = regs[i].first()->reg2stack();
duke@0 1363 if (off > max_stack_slots) max_stack_slots = off;
duke@0 1364 }
duke@0 1365 if (regs[i].second()->is_stack()) {
duke@0 1366 int off = regs[i].second()->reg2stack();
duke@0 1367 if (off > max_stack_slots) max_stack_slots = off;
duke@0 1368 }
duke@0 1369 }
duke@0 1370 #endif // _LP64
duke@0 1371
duke@0 1372 return round_to(max_stack_slots + 1, 2);
duke@0 1373
duke@0 1374 }
duke@0 1375
duke@0 1376
duke@0 1377 // ---------------------------------------------------------------------------
duke@0 1378 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
duke@0 1379 switch (ret_type) {
duke@0 1380 case T_FLOAT:
duke@0 1381 __ stf(FloatRegisterImpl::S, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS);
duke@0 1382 break;
duke@0 1383 case T_DOUBLE:
duke@0 1384 __ stf(FloatRegisterImpl::D, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS);
duke@0 1385 break;
duke@0 1386 }
duke@0 1387 }
duke@0 1388
duke@0 1389 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
duke@0 1390 switch (ret_type) {
duke@0 1391 case T_FLOAT:
duke@0 1392 __ ldf(FloatRegisterImpl::S, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS, F0);
duke@0 1393 break;
duke@0 1394 case T_DOUBLE:
duke@0 1395 __ ldf(FloatRegisterImpl::D, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS, F0);
duke@0 1396 break;
duke@0 1397 }
duke@0 1398 }
duke@0 1399
duke@0 1400 // Check and forward and pending exception. Thread is stored in
duke@0 1401 // L7_thread_cache and possibly NOT in G2_thread. Since this is a native call, there
duke@0 1402 // is no exception handler. We merely pop this frame off and throw the
duke@0 1403 // exception in the caller's frame.
duke@0 1404 static void check_forward_pending_exception(MacroAssembler *masm, Register Rex_oop) {
duke@0 1405 Label L;
duke@0 1406 __ br_null(Rex_oop, false, Assembler::pt, L);
duke@0 1407 __ delayed()->mov(L7_thread_cache, G2_thread); // restore in case we have exception
duke@0 1408 // Since this is a native call, we *know* the proper exception handler
duke@0 1409 // without calling into the VM: it's the empty function. Just pop this
duke@0 1410 // frame and then jump to forward_exception_entry; O7 will contain the
duke@0 1411 // native caller's return PC.
twisti@720 1412 AddressLiteral exception_entry(StubRoutines::forward_exception_entry());
twisti@720 1413 __ jump_to(exception_entry, G3_scratch);
duke@0 1414 __ delayed()->restore(); // Pop this frame off.
duke@0 1415 __ bind(L);
duke@0 1416 }
duke@0 1417
duke@0 1418 // A simple move of integer like type
duke@0 1419 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
duke@0 1420 if (src.first()->is_stack()) {
duke@0 1421 if (dst.first()->is_stack()) {
duke@0 1422 // stack to stack
duke@0 1423 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
duke@0 1424 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1425 } else {
duke@0 1426 // stack to reg
duke@0 1427 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
duke@0 1428 }
duke@0 1429 } else if (dst.first()->is_stack()) {
duke@0 1430 // reg to stack
duke@0 1431 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1432 } else {
duke@0 1433 __ mov(src.first()->as_Register(), dst.first()->as_Register());
duke@0 1434 }
duke@0 1435 }
duke@0 1436
duke@0 1437 // On 64 bit we will store integer like items to the stack as
duke@0 1438 // 64 bits items (sparc abi) even though java would only store
duke@0 1439 // 32bits for a parameter. On 32bit it will simply be 32 bits
duke@0 1440 // So this routine will do 32->32 on 32bit and 32->64 on 64bit
duke@0 1441 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
duke@0 1442 if (src.first()->is_stack()) {
duke@0 1443 if (dst.first()->is_stack()) {
duke@0 1444 // stack to stack
duke@0 1445 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
duke@0 1446 __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1447 } else {
duke@0 1448 // stack to reg
duke@0 1449 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
duke@0 1450 }
duke@0 1451 } else if (dst.first()->is_stack()) {
duke@0 1452 // reg to stack
duke@0 1453 __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1454 } else {
duke@0 1455 __ mov(src.first()->as_Register(), dst.first()->as_Register());
duke@0 1456 }
duke@0 1457 }
duke@0 1458
duke@0 1459
duke@0 1460 // An oop arg. Must pass a handle not the oop itself
duke@0 1461 static void object_move(MacroAssembler* masm,
duke@0 1462 OopMap* map,
duke@0 1463 int oop_handle_offset,
duke@0 1464 int framesize_in_slots,
duke@0 1465 VMRegPair src,
duke@0 1466 VMRegPair dst,
duke@0 1467 bool is_receiver,
duke@0 1468 int* receiver_offset) {
duke@0 1469
duke@0 1470 // must pass a handle. First figure out the location we use as a handle
duke@0 1471
duke@0 1472 if (src.first()->is_stack()) {
duke@0 1473 // Oop is already on the stack
duke@0 1474 Register rHandle = dst.first()->is_stack() ? L5 : dst.first()->as_Register();
duke@0 1475 __ add(FP, reg2offset(src.first()) + STACK_BIAS, rHandle);
duke@0 1476 __ ld_ptr(rHandle, 0, L4);
duke@0 1477 #ifdef _LP64
duke@0 1478 __ movr( Assembler::rc_z, L4, G0, rHandle );
duke@0 1479 #else
duke@0 1480 __ tst( L4 );
duke@0 1481 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
duke@0 1482 #endif
duke@0 1483 if (dst.first()->is_stack()) {
duke@0 1484 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1485 }
duke@0 1486 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
duke@0 1487 if (is_receiver) {
duke@0 1488 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
duke@0 1489 }
duke@0 1490 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
duke@0 1491 } else {
duke@0 1492 // Oop is in an input register pass we must flush it to the stack
duke@0 1493 const Register rOop = src.first()->as_Register();
duke@0 1494 const Register rHandle = L5;
duke@0 1495 int oop_slot = rOop->input_number() * VMRegImpl::slots_per_word + oop_handle_offset;
duke@0 1496 int offset = oop_slot*VMRegImpl::stack_slot_size;
duke@0 1497 Label skip;
duke@0 1498 __ st_ptr(rOop, SP, offset + STACK_BIAS);
duke@0 1499 if (is_receiver) {
duke@0 1500 *receiver_offset = oop_slot * VMRegImpl::stack_slot_size;
duke@0 1501 }
duke@0 1502 map->set_oop(VMRegImpl::stack2reg(oop_slot));
duke@0 1503 __ add(SP, offset + STACK_BIAS, rHandle);
duke@0 1504 #ifdef _LP64
duke@0 1505 __ movr( Assembler::rc_z, rOop, G0, rHandle );
duke@0 1506 #else
duke@0 1507 __ tst( rOop );
duke@0 1508 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
duke@0 1509 #endif
duke@0 1510
duke@0 1511 if (dst.first()->is_stack()) {
duke@0 1512 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1513 } else {
duke@0 1514 __ mov(rHandle, dst.first()->as_Register());
duke@0 1515 }
duke@0 1516 }
duke@0 1517 }
duke@0 1518
duke@0 1519 // A float arg may have to do float reg int reg conversion
duke@0 1520 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
duke@0 1521 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move");
duke@0 1522
duke@0 1523 if (src.first()->is_stack()) {
duke@0 1524 if (dst.first()->is_stack()) {
duke@0 1525 // stack to stack the easiest of the bunch
duke@0 1526 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
duke@0 1527 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1528 } else {
duke@0 1529 // stack to reg
duke@0 1530 if (dst.first()->is_Register()) {
duke@0 1531 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
duke@0 1532 } else {
duke@0 1533 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
duke@0 1534 }
duke@0 1535 }
duke@0 1536 } else if (dst.first()->is_stack()) {
duke@0 1537 // reg to stack
duke@0 1538 if (src.first()->is_Register()) {
duke@0 1539 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1540 } else {
duke@0 1541 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1542 }
duke@0 1543 } else {
duke@0 1544 // reg to reg
duke@0 1545 if (src.first()->is_Register()) {
duke@0 1546 if (dst.first()->is_Register()) {
duke@0 1547 // gpr -> gpr
duke@0 1548 __ mov(src.first()->as_Register(), dst.first()->as_Register());
duke@0 1549 } else {
duke@0 1550 // gpr -> fpr
duke@0 1551 __ st(src.first()->as_Register(), FP, -4 + STACK_BIAS);
duke@0 1552 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.first()->as_FloatRegister());
duke@0 1553 }
duke@0 1554 } else if (dst.first()->is_Register()) {
duke@0 1555 // fpr -> gpr
duke@0 1556 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), FP, -4 + STACK_BIAS);
duke@0 1557 __ ld(FP, -4 + STACK_BIAS, dst.first()->as_Register());
duke@0 1558 } else {
duke@0 1559 // fpr -> fpr
duke@0 1560 // In theory these overlap but the ordering is such that this is likely a nop
duke@0 1561 if ( src.first() != dst.first()) {
duke@0 1562 __ fmov(FloatRegisterImpl::S, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
duke@0 1563 }
duke@0 1564 }
duke@0 1565 }
duke@0 1566 }
duke@0 1567
duke@0 1568 static void split_long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
duke@0 1569 VMRegPair src_lo(src.first());
duke@0 1570 VMRegPair src_hi(src.second());
duke@0 1571 VMRegPair dst_lo(dst.first());
duke@0 1572 VMRegPair dst_hi(dst.second());
duke@0 1573 simple_move32(masm, src_lo, dst_lo);
duke@0 1574 simple_move32(masm, src_hi, dst_hi);
duke@0 1575 }
duke@0 1576
duke@0 1577 // A long move
duke@0 1578 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
duke@0 1579
duke@0 1580 // Do the simple ones here else do two int moves
duke@0 1581 if (src.is_single_phys_reg() ) {
duke@0 1582 if (dst.is_single_phys_reg()) {
duke@0 1583 __ mov(src.first()->as_Register(), dst.first()->as_Register());
duke@0 1584 } else {
duke@0 1585 // split src into two separate registers
duke@0 1586 // Remember hi means hi address or lsw on sparc
duke@0 1587 // Move msw to lsw
duke@0 1588 if (dst.second()->is_reg()) {
duke@0 1589 // MSW -> MSW
duke@0 1590 __ srax(src.first()->as_Register(), 32, dst.first()->as_Register());
duke@0 1591 // Now LSW -> LSW
duke@0 1592 // this will only move lo -> lo and ignore hi
duke@0 1593 VMRegPair split(dst.second());
duke@0 1594 simple_move32(masm, src, split);
duke@0 1595 } else {
duke@0 1596 VMRegPair split(src.first(), L4->as_VMReg());
duke@0 1597 // MSW -> MSW (lo ie. first word)
duke@0 1598 __ srax(src.first()->as_Register(), 32, L4);
duke@0 1599 split_long_move(masm, split, dst);
duke@0 1600 }
duke@0 1601 }
duke@0 1602 } else if (dst.is_single_phys_reg()) {
duke@0 1603 if (src.is_adjacent_aligned_on_stack(2)) {
never@297 1604 __ ldx(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
duke@0 1605 } else {
duke@0 1606 // dst is a single reg.
duke@0 1607 // Remember lo is low address not msb for stack slots
duke@0 1608 // and lo is the "real" register for registers
duke@0 1609 // src is
duke@0 1610
duke@0 1611 VMRegPair split;
duke@0 1612
duke@0 1613 if (src.first()->is_reg()) {
duke@0 1614 // src.lo (msw) is a reg, src.hi is stk/reg
duke@0 1615 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> src.lo [the MSW is in the LSW of the reg]
duke@0 1616 split.set_pair(dst.first(), src.first());
duke@0 1617 } else {
duke@0 1618 // msw is stack move to L5
duke@0 1619 // lsw is stack move to dst.lo (real reg)
duke@0 1620 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> L5
duke@0 1621 split.set_pair(dst.first(), L5->as_VMReg());
duke@0 1622 }
duke@0 1623
duke@0 1624 // src.lo -> src.lo/L5, src.hi -> dst.lo (the real reg)
duke@0 1625 // msw -> src.lo/L5, lsw -> dst.lo
duke@0 1626 split_long_move(masm, src, split);
duke@0 1627
duke@0 1628 // So dst now has the low order correct position the
duke@0 1629 // msw half
duke@0 1630 __ sllx(split.first()->as_Register(), 32, L5);
duke@0 1631
duke@0 1632 const Register d = dst.first()->as_Register();
duke@0 1633 __ or3(L5, d, d);
duke@0 1634 }
duke@0 1635 } else {
duke@0 1636 // For LP64 we can probably do better.
duke@0 1637 split_long_move(masm, src, dst);
duke@0 1638 }
duke@0 1639 }
duke@0 1640
duke@0 1641 // A double move
duke@0 1642 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
duke@0 1643
duke@0 1644 // The painful thing here is that like long_move a VMRegPair might be
duke@0 1645 // 1: a single physical register
duke@0 1646 // 2: two physical registers (v8)
duke@0 1647 // 3: a physical reg [lo] and a stack slot [hi] (v8)
duke@0 1648 // 4: two stack slots
duke@0 1649
duke@0 1650 // Since src is always a java calling convention we know that the src pair
duke@0 1651 // is always either all registers or all stack (and aligned?)
duke@0 1652
duke@0 1653 // in a register [lo] and a stack slot [hi]
duke@0 1654 if (src.first()->is_stack()) {
duke@0 1655 if (dst.first()->is_stack()) {
duke@0 1656 // stack to stack the easiest of the bunch
duke@0 1657 // ought to be a way to do this where if alignment is ok we use ldd/std when possible
duke@0 1658 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
duke@0 1659 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
duke@0 1660 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1661 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
duke@0 1662 } else {
duke@0 1663 // stack to reg
duke@0 1664 if (dst.second()->is_stack()) {
duke@0 1665 // stack -> reg, stack -> stack
duke@0 1666 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
duke@0 1667 if (dst.first()->is_Register()) {
duke@0 1668 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
duke@0 1669 } else {
duke@0 1670 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
duke@0 1671 }
duke@0 1672 // This was missing. (very rare case)
duke@0 1673 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
duke@0 1674 } else {
duke@0 1675 // stack -> reg
duke@0 1676 // Eventually optimize for alignment QQQ
duke@0 1677 if (dst.first()->is_Register()) {
duke@0 1678 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
duke@0 1679 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_Register());
duke@0 1680 } else {
duke@0 1681 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
duke@0 1682 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_FloatRegister());
duke@0 1683 }
duke@0 1684 }
duke@0 1685 }
duke@0 1686 } else if (dst.first()->is_stack()) {
duke@0 1687 // reg to stack
duke@0 1688 if (src.first()->is_Register()) {
duke@0 1689 // Eventually optimize for alignment QQQ
duke@0 1690 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1691 if (src.second()->is_stack()) {
duke@0 1692 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
duke@0 1693 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
duke@0 1694 } else {
duke@0 1695 __ st(src.second()->as_Register(), SP, reg2offset(dst.second()) + STACK_BIAS);
duke@0 1696 }
duke@0 1697 } else {
duke@0 1698 // fpr to stack
duke@0 1699 if (src.second()->is_stack()) {
duke@0 1700 ShouldNotReachHere();
duke@0 1701 } else {
duke@0 1702 // Is the stack aligned?
duke@0 1703 if (reg2offset(dst.first()) & 0x7) {
duke@0 1704 // No do as pairs
duke@0 1705 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1706 __ stf(FloatRegisterImpl::S, src.second()->as_FloatRegister(), SP, reg2offset(dst.second()) + STACK_BIAS);
duke@0 1707 } else {
duke@0 1708 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
duke@0 1709 }
duke@0 1710 }
duke@0 1711 }
duke@0 1712 } else {
duke@0 1713 // reg to reg
duke@0 1714 if (src.first()->is_Register()) {
duke@0 1715 if (dst.first()->is_Register()) {
duke@0 1716 // gpr -> gpr
duke@0 1717 __ mov(src.first()->as_Register(), dst.first()->as_Register());
duke@0 1718 __ mov(src.second()->as_Register(), dst.second()->as_Register());
duke@0 1719 } else {
duke@0 1720 // gpr -> fpr
duke@0 1721 // ought to be able to do a single store
duke@0 1722 __ stx(src.first()->as_Register(), FP, -8 + STACK_BIAS);
duke@0 1723 __ stx(src.second()->as_Register(), FP, -4 + STACK_BIAS);
duke@0 1724 // ought to be able to do a single load
duke@0 1725 __ ldf(FloatRegisterImpl::S, FP, -8 + STACK_BIAS, dst.first()->as_FloatRegister());
duke@0 1726 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.second()->as_FloatRegister());
duke@0 1727 }
duke@0 1728 } else if (dst.first()->is_Register()) {
duke@0 1729 // fpr -> gpr
duke@0 1730 // ought to be able to do a single store
duke@0 1731 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), FP, -8 + STACK_BIAS);
duke@0 1732 // ought to be able to do a single load
duke@0 1733 // REMEMBER first() is low address not LSB
duke@0 1734 __ ld(FP, -8 + STACK_BIAS, dst.first()->as_Register());
duke@0 1735 if (dst.second()->is_Register()) {
duke@0 1736 __ ld(FP, -4 + STACK_BIAS, dst.second()->as_Register());
duke@0 1737 } else {
duke@0 1738 __ ld(FP, -4 + STACK_BIAS, L4);
duke@0 1739 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
duke@0 1740 }
duke@0 1741 } else {
duke@0 1742 // fpr -> fpr
duke@0 1743 // In theory these overlap but the ordering is such that this is likely a nop
duke@0 1744 if ( src.first() != dst.first()) {
duke@0 1745 __ fmov(FloatRegisterImpl::D, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
duke@0 1746 }
duke@0 1747 }
duke@0 1748 }
duke@0 1749 }
duke@0 1750
duke@0 1751 // Creates an inner frame if one hasn't already been created, and
duke@0 1752 // saves a copy of the thread in L7_thread_cache
duke@0 1753 static void create_inner_frame(MacroAssembler* masm, bool* already_created) {
duke@0 1754 if (!*already_created) {
duke@0 1755 __ save_frame(0);
duke@0 1756 // Save thread in L7 (INNER FRAME); it crosses a bunch of VM calls below
duke@0 1757 // Don't use save_thread because it smashes G2 and we merely want to save a
duke@0 1758 // copy
duke@0 1759 __ mov(G2_thread, L7_thread_cache);
duke@0 1760 *already_created = true;
duke@0 1761 }
duke@0 1762 }
duke@0 1763
duke@0 1764 // ---------------------------------------------------------------------------
duke@0 1765 // Generate a native wrapper for a given method. The method takes arguments
duke@0 1766 // in the Java compiled code convention, marshals them to the native
duke@0 1767 // convention (handlizes oops, etc), transitions to native, makes the call,
duke@0 1768 // returns to java state (possibly blocking), unhandlizes any result and
duke@0 1769 // returns.
duke@0 1770 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
duke@0 1771 methodHandle method,
duke@0 1772 int total_in_args,
duke@0 1773 int comp_args_on_stack, // in VMRegStackSlots
duke@0 1774 BasicType *in_sig_bt,
duke@0 1775 VMRegPair *in_regs,
duke@0 1776 BasicType ret_type) {
duke@0 1777
duke@0 1778 // Native nmethod wrappers never take possesion of the oop arguments.
duke@0 1779 // So the caller will gc the arguments. The only thing we need an
duke@0 1780 // oopMap for is if the call is static
duke@0 1781 //
duke@0 1782 // An OopMap for lock (and class if static), and one for the VM call itself
duke@0 1783 OopMapSet *oop_maps = new OopMapSet();
duke@0 1784 intptr_t start = (intptr_t)__ pc();
duke@0 1785
duke@0 1786 // First thing make an ic check to see if we should even be here
duke@0 1787 {
duke@0 1788 Label L;
duke@0 1789 const Register temp_reg = G3_scratch;
twisti@720 1790 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
duke@0 1791 __ verify_oop(O0);
coleenp@108 1792 __ load_klass(O0, temp_reg);
duke@0 1793 __ cmp(temp_reg, G5_inline_cache_reg);
duke@0 1794 __ brx(Assembler::equal, true, Assembler::pt, L);
duke@0 1795 __ delayed()->nop();
duke@0 1796
twisti@720 1797 __ jump_to(ic_miss, temp_reg);
duke@0 1798 __ delayed()->nop();
duke@0 1799 __ align(CodeEntryAlignment);
duke@0 1800 __ bind(L);
duke@0 1801 }
duke@0 1802
duke@0 1803 int vep_offset = ((intptr_t)__ pc()) - start;
duke@0 1804
duke@0 1805 #ifdef COMPILER1
duke@0 1806 if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) {
duke@0 1807 // Object.hashCode can pull the hashCode from the header word
duke@0 1808 // instead of doing a full VM transition once it's been computed.
duke@0 1809 // Since hashCode is usually polymorphic at call sites we can't do
duke@0 1810 // this optimization at the call site without a lot of work.
duke@0 1811 Label slowCase;
duke@0 1812 Register receiver = O0;
duke@0 1813 Register result = O0;
duke@0 1814 Register header = G3_scratch;
duke@0 1815 Register hash = G3_scratch; // overwrite header value with hash value
duke@0 1816 Register mask = G1; // to get hash field from header
duke@0 1817
duke@0 1818 // Read the header and build a mask to get its hash field. Give up if the object is not unlocked.
duke@0 1819 // We depend on hash_mask being at most 32 bits and avoid the use of
duke@0 1820 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
duke@0 1821 // vm: see markOop.hpp.
duke@0 1822 __ ld_ptr(receiver, oopDesc::mark_offset_in_bytes(), header);
duke@0 1823 __ sethi(markOopDesc::hash_mask, mask);
duke@0 1824 __ btst(markOopDesc::unlocked_value, header);
duke@0 1825 __ br(Assembler::zero, false, Assembler::pn, slowCase);
duke@0 1826 if (UseBiasedLocking) {
duke@0 1827 // Check if biased and fall through to runtime if so
duke@0 1828 __ delayed()->nop();
duke@0 1829 __ btst(markOopDesc::biased_lock_bit_in_place, header);
duke@0 1830 __ br(Assembler::notZero, false, Assembler::pn, slowCase);
duke@0 1831 }
duke@0 1832 __ delayed()->or3(mask, markOopDesc::hash_mask & 0x3ff, mask);
duke@0 1833
duke@0 1834 // Check for a valid (non-zero) hash code and get its value.
duke@0 1835 #ifdef _LP64
duke@0 1836 __ srlx(header, markOopDesc::hash_shift, hash);
duke@0 1837 #else
duke@0 1838 __ srl(header, markOopDesc::hash_shift, hash);
duke@0 1839 #endif
duke@0 1840 __ andcc(hash, mask, hash);
duke@0 1841 __ br(Assembler::equal, false, Assembler::pn, slowCase);
duke@0 1842 __ delayed()->nop();
duke@0 1843
duke@0 1844 // leaf return.
duke@0 1845 __ retl();
duke@0 1846 __ delayed()->mov(hash, result);
duke@0 1847 __ bind(slowCase);
duke@0 1848 }
duke@0 1849 #endif // COMPILER1
duke@0 1850
duke@0 1851
duke@0 1852 // We have received a description of where all the java arg are located
duke@0 1853 // on entry to the wrapper. We need to convert these args to where
duke@0 1854 // the jni function will expect them. To figure out where they go
duke@0 1855 // we convert the java signature to a C signature by inserting
duke@0 1856 // the hidden arguments as arg[0] and possibly arg[1] (static method)
duke@0 1857
duke@0 1858 int total_c_args = total_in_args + 1;
duke@0 1859 if (method->is_static()) {
duke@0 1860 total_c_args++;
duke@0 1861 }
duke@0 1862
duke@0 1863 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
duke@0 1864 VMRegPair * out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
duke@0 1865
duke@0 1866 int argc = 0;
duke@0 1867 out_sig_bt[argc++] = T_ADDRESS;
duke@0 1868 if (method->is_static()) {
duke@0 1869 out_sig_bt[argc++] = T_OBJECT;
duke@0 1870 }
duke@0 1871
duke@0 1872 for (int i = 0; i < total_in_args ; i++ ) {
duke@0 1873 out_sig_bt[argc++] = in_sig_bt[i];
duke@0 1874 }
duke@0 1875
duke@0 1876 // Now figure out where the args must be stored and how much stack space
duke@0 1877 // they require (neglecting out_preserve_stack_slots but space for storing
duke@0 1878 // the 1st six register arguments). It's weird see int_stk_helper.
duke@0 1879 //
duke@0 1880 int out_arg_slots;
duke@0 1881 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
duke@0 1882
duke@0 1883 // Compute framesize for the wrapper. We need to handlize all oops in
duke@0 1884 // registers. We must create space for them here that is disjoint from
duke@0 1885 // the windowed save area because we have no control over when we might
duke@0 1886 // flush the window again and overwrite values that gc has since modified.
duke@0 1887 // (The live window race)
duke@0 1888 //
duke@0 1889 // We always just allocate 6 word for storing down these object. This allow
duke@0 1890 // us to simply record the base and use the Ireg number to decide which
duke@0 1891 // slot to use. (Note that the reg number is the inbound number not the
duke@0 1892 // outbound number).
duke@0 1893 // We must shuffle args to match the native convention, and include var-args space.
duke@0 1894
duke@0 1895 // Calculate the total number of stack slots we will need.
duke@0 1896
duke@0 1897 // First count the abi requirement plus all of the outgoing args
duke@0 1898 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
duke@0 1899
duke@0 1900 // Now the space for the inbound oop handle area
duke@0 1901
duke@0 1902 int oop_handle_offset = stack_slots;
duke@0 1903 stack_slots += 6*VMRegImpl::slots_per_word;
duke@0 1904
duke@0 1905 // Now any space we need for handlizing a klass if static method
duke@0 1906
duke@0 1907 int oop_temp_slot_offset = 0;
duke@0 1908 int klass_slot_offset = 0;
duke@0 1909 int klass_offset = -1;
duke@0 1910 int lock_slot_offset = 0;
duke@0 1911 bool is_static = false;
duke@0 1912
duke@0 1913 if (method->is_static()) {
duke@0 1914 klass_slot_offset = stack_slots;
duke@0 1915 stack_slots += VMRegImpl::slots_per_word;
duke@0 1916 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
duke@0 1917 is_static = true;
duke@0 1918 }
duke@0 1919
duke@0 1920 // Plus a lock if needed
duke@0 1921
duke@0 1922 if (method->is_synchronized()) {
duke@0 1923 lock_slot_offset = stack_slots;
duke@0 1924 stack_slots += VMRegImpl::slots_per_word;
duke@0 1925 }
duke@0 1926
duke@0 1927 // Now a place to save return value or as a temporary for any gpr -> fpr moves
duke@0 1928 stack_slots += 2;
duke@0 1929
duke@0 1930 // Ok The space we have allocated will look like:
duke@0 1931 //
duke@0 1932 //
duke@0 1933 // FP-> | |
duke@0 1934 // |---------------------|
duke@0 1935 // | 2 slots for moves |
duke@0 1936 // |---------------------|
duke@0 1937 // | lock box (if sync) |
duke@0 1938 // |---------------------| <- lock_slot_offset
duke@0 1939 // | klass (if static) |
duke@0 1940 // |---------------------| <- klass_slot_offset
duke@0 1941 // | oopHandle area |
duke@0 1942 // |---------------------| <- oop_handle_offset
duke@0 1943 // | outbound memory |
duke@0 1944 // | based arguments |
duke@0 1945 // | |
duke@0 1946 // |---------------------|
duke@0 1947 // | vararg area |
duke@0 1948 // |---------------------|
duke@0 1949 // | |
duke@0 1950 // SP-> | out_preserved_slots |
duke@0 1951 //
duke@0 1952 //
duke@0 1953
duke@0 1954
duke@0 1955 // Now compute actual number of stack words we need rounding to make
duke@0 1956 // stack properly aligned.
duke@0 1957 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word);
duke@0 1958
duke@0 1959 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
duke@0 1960
duke@0 1961 // Generate stack overflow check before creating frame
duke@0 1962 __ generate_stack_overflow_check(stack_size);
duke@0 1963
duke@0 1964 // Generate a new frame for the wrapper.
duke@0 1965 __ save(SP, -stack_size, SP);
duke@0 1966
duke@0 1967 int frame_complete = ((intptr_t)__ pc()) - start;
duke@0 1968
duke@0 1969 __ verify_thread();
duke@0 1970
duke@0 1971
duke@0 1972 //
duke@0 1973 // We immediately shuffle the arguments so that any vm call we have to
duke@0 1974 // make from here on out (sync slow path, jvmti, etc.) we will have
duke@0 1975 // captured the oops from our caller and have a valid oopMap for
duke@0 1976 // them.
duke@0 1977
duke@0 1978 // -----------------
duke@0 1979 // The Grand Shuffle
duke@0 1980 //
duke@0 1981 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
duke@0 1982 // (derived from JavaThread* which is in L7_thread_cache) and, if static,
duke@0 1983 // the class mirror instead of a receiver. This pretty much guarantees that
duke@0 1984 // register layout will not match. We ignore these extra arguments during
duke@0 1985 // the shuffle. The shuffle is described by the two calling convention
duke@0 1986 // vectors we have in our possession. We simply walk the java vector to
duke@0 1987 // get the source locations and the c vector to get the destinations.
duke@0 1988 // Because we have a new window and the argument registers are completely
duke@0 1989 // disjoint ( I0 -> O1, I1 -> O2, ...) we have nothing to worry about
duke@0 1990 // here.
duke@0 1991
duke@0 1992 // This is a trick. We double the stack slots so we can claim
duke@0 1993 // the oops in the caller's frame. Since we are sure to have
duke@0 1994 // more args than the caller doubling is enough to make
duke@0 1995 // sure we can capture all the incoming oop args from the
duke@0 1996 // caller.
duke@0 1997 //
duke@0 1998 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
duke@0 1999 int c_arg = total_c_args - 1;
duke@0 2000 // Record sp-based slot for receiver on stack for non-static methods
duke@0 2001 int receiver_offset = -1;
duke@0 2002
duke@0 2003 // We move the arguments backward because the floating point registers
duke@0 2004 // destination will always be to a register with a greater or equal register
duke@0 2005 // number or the stack.
duke@0 2006
duke@0 2007 #ifdef ASSERT
duke@0 2008 bool reg_destroyed[RegisterImpl::number_of_registers];
duke@0 2009 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
duke@0 2010 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
duke@0 2011 reg_destroyed[r] = false;
duke@0 2012 }
duke@0 2013 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
duke@0 2014 freg_destroyed[f] = false;
duke@0 2015 }
duke@0 2016
duke@0 2017 #endif /* ASSERT */
duke@0 2018
duke@0 2019 for ( int i = total_in_args - 1; i >= 0 ; i--, c_arg-- ) {
duke@0 2020
duke@0 2021 #ifdef ASSERT
duke@0 2022 if (in_regs[i].first()->is_Register()) {
duke@0 2023 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "ack!");
duke@0 2024 } else if (in_regs[i].first()->is_FloatRegister()) {
duke@0 2025 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)], "ack!");
duke@0 2026 }
duke@0 2027 if (out_regs[c_arg].first()->is_Register()) {
duke@0 2028 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
duke@0 2029 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
duke@0 2030 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)] = true;
duke@0 2031 }
duke@0 2032 #endif /* ASSERT */
duke@0 2033
duke@0 2034 switch (in_sig_bt[i]) {
duke@0 2035 case T_ARRAY:
duke@0 2036 case T_OBJECT:
duke@0 2037 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
duke@0 2038 ((i == 0) && (!is_static)),
duke@0 2039 &receiver_offset);
duke@0 2040 break;
duke@0 2041 case T_VOID:
duke@0 2042 break;
duke@0 2043
duke@0 2044 case T_FLOAT:
duke@0 2045 float_move(masm, in_regs[i], out_regs[c_arg]);
duke@0 2046 break;
duke@0 2047
duke@0 2048 case T_DOUBLE:
duke@0 2049 assert( i + 1 < total_in_args &&
duke@0 2050 in_sig_bt[i + 1] == T_VOID &&
duke@0 2051 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
duke@0 2052 double_move(masm, in_regs[i], out_regs[c_arg]);
duke@0 2053 break;
duke@0 2054
duke@0 2055 case T_LONG :
duke@0 2056 long_move(masm, in_regs[i], out_regs[c_arg]);
duke@0 2057 break;
duke@0 2058
duke@0 2059 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
duke@0 2060
duke@0 2061 default:
duke@0 2062 move32_64(masm, in_regs[i], out_regs[c_arg]);
duke@0 2063 }
duke@0 2064 }
duke@0 2065
duke@0 2066 // Pre-load a static method's oop into O1. Used both by locking code and
duke@0 2067 // the normal JNI call code.
duke@0 2068 if (method->is_static()) {
duke@0 2069 __ set_oop_constant(JNIHandles::make_local(Klass::cast(method->method_holder())->java_mirror()), O1);
duke@0 2070
duke@0 2071 // Now handlize the static class mirror in O1. It's known not-null.
duke@0 2072 __ st_ptr(O1, SP, klass_offset + STACK_BIAS);
duke@0 2073 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
duke@0 2074 __ add(SP, klass_offset + STACK_BIAS, O1);
duke@0 2075 }
duke@0 2076
duke@0 2077
duke@0 2078 const Register L6_handle = L6;
duke@0 2079
duke@0 2080 if (method->is_synchronized()) {
duke@0 2081 __ mov(O1, L6_handle);
duke@0 2082 }
duke@0 2083
duke@0 2084 // We have all of the arguments setup at this point. We MUST NOT touch any Oregs
duke@0 2085 // except O6/O7. So if we must call out we must push a new frame. We immediately
duke@0 2086 // push a new frame and flush the windows.
duke@0 2087
duke@0 2088 #ifdef _LP64
duke@0 2089 intptr_t thepc = (intptr_t) __ pc();
duke@0 2090 {
duke@0 2091 address here = __ pc();
duke@0 2092 // Call the next instruction
duke@0 2093 __ call(here + 8, relocInfo::none);
duke@0 2094 __ delayed()->nop();
duke@0 2095 }
duke@0 2096 #else
duke@0 2097 intptr_t thepc = __ load_pc_address(O7, 0);
duke@0 2098 #endif /* _LP64 */
duke@0 2099
duke@0 2100 // We use the same pc/oopMap repeatedly when we call out
duke@0 2101 oop_maps->add_gc_map(thepc - start, map);
duke@0 2102
duke@0 2103 // O7 now has the pc loaded that we will use when we finally call to native.
duke@0 2104
duke@0 2105 // Save thread in L7; it crosses a bunch of VM calls below
duke@0 2106 // Don't use save_thread because it smashes G2 and we merely
duke@0 2107 // want to save a copy
duke@0 2108 __ mov(G2_thread, L7_thread_cache);
duke@0 2109
duke@0 2110
duke@0 2111 // If we create an inner frame once is plenty
duke@0 2112 // when we create it we must also save G2_thread
duke@0 2113 bool inner_frame_created = false;
duke@0 2114
duke@0 2115 // dtrace method entry support
duke@0 2116 {
duke@0 2117 SkipIfEqual skip_if(
duke@0 2118 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
duke@0 2119 // create inner frame
duke@0 2120 __ save_frame(0);
duke@0 2121 __ mov(G2_thread, L7_thread_cache);
duke@0 2122 __ set_oop_constant(JNIHandles::make_local(method()), O1);
duke@0 2123 __ call_VM_leaf(L7_thread_cache,
duke@0 2124 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
duke@0 2125 G2_thread, O1);
duke@0 2126 __ restore();
duke@0 2127 }
duke@0 2128
dcubed@606 2129 // RedefineClasses() tracing support for obsolete method entry
dcubed@606 2130 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) {
dcubed@606 2131 // create inner frame
dcubed@606 2132 __ save_frame(0);
dcubed@606 2133 __ mov(G2_thread, L7_thread_cache);
dcubed@606 2134 __ set_oop_constant(JNIHandles::make_local(method()), O1);
dcubed@606 2135 __ call_VM_leaf(L7_thread_cache,
dcubed@606 2136 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
dcubed@606 2137 G2_thread, O1);
dcubed@606 2138 __ restore();
dcubed@606 2139 }
dcubed@606 2140
duke@0 2141 // We are in the jni frame unless saved_frame is true in which case
duke@0 2142 // we are in one frame deeper (the "inner" frame). If we are in the
duke@0 2143 // "inner" frames the args are in the Iregs and if the jni frame then
duke@0 2144 // they are in the Oregs.
duke@0 2145 // If we ever need to go to the VM (for locking, jvmti) then
duke@0 2146 // we will always be in the "inner" frame.
duke@0 2147
duke@0 2148 // Lock a synchronized method
duke@0 2149 int lock_offset = -1; // Set if locked
duke@0 2150 if (method->is_synchronized()) {
duke@0 2151 Register Roop = O1;
duke@0 2152 const Register L3_box = L3;
duke@0 2153
duke@0 2154 create_inner_frame(masm, &inner_frame_created);
duke@0 2155
duke@0 2156 __ ld_ptr(I1, 0, O1);
duke@0 2157 Label done;
duke@0 2158
duke@0 2159 lock_offset = (lock_slot_offset * VMRegImpl::stack_slot_size);
duke@0 2160 __ add(FP, lock_offset+STACK_BIAS, L3_box);
duke@0 2161 #ifdef ASSERT
duke@0 2162 if (UseBiasedLocking) {
duke@0 2163 // making the box point to itself will make it clear it went unused
duke@0 2164 // but also be obviously invalid
duke@0 2165 __ st_ptr(L3_box, L3_box, 0);
duke@0 2166 }
duke@0 2167 #endif // ASSERT
duke@0 2168 //
duke@0 2169 // Compiler_lock_object (Roop, Rmark, Rbox, Rscratch) -- kills Rmark, Rbox, Rscratch
duke@0 2170 //
duke@0 2171 __ compiler_lock_object(Roop, L1, L3_box, L2);
duke@0 2172 __ br(Assembler::equal, false, Assembler::pt, done);
duke@0 2173 __ delayed() -> add(FP, lock_offset+STACK_BIAS, L3_box);
duke@0 2174
duke@0 2175
duke@0 2176 // None of the above fast optimizations worked so we have to get into the
duke@0 2177 // slow case of monitor enter. Inline a special case of call_VM that
duke@0 2178 // disallows any pending_exception.
duke@0 2179 __ mov(Roop, O0); // Need oop in O0
duke@0 2180 __ mov(L3_box, O1);
duke@0 2181
duke@0 2182 // Record last_Java_sp, in case the VM code releases the JVM lock.
duke@0 2183
duke@0 2184 __ set_last_Java_frame(FP, I7);
duke@0 2185
duke@0 2186 // do the call
duke@0 2187 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), relocInfo::runtime_call_type);
duke@0 2188 __ delayed()->mov(L7_thread_cache, O2);
duke@0 2189
duke@0 2190 __ restore_thread(L7_thread_cache); // restore G2_thread
duke@0 2191 __ reset_last_Java_frame();
duke@0 2192
duke@0 2193 #ifdef ASSERT
duke@0 2194 { Label L;
duke@0 2195 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
duke@0 2196 __ br_null(O0, false, Assembler::pt, L);
duke@0 2197 __ delayed()->nop();
duke@0 2198 __ stop("no pending exception allowed on exit from IR::monitorenter");
duke@0 2199 __ bind(L);
duke@0 2200 }
duke@0 2201 #endif
duke@0 2202 __ bind(done);
duke@0 2203 }
duke@0 2204
duke@0 2205
duke@0 2206 // Finally just about ready to make the JNI call
duke@0 2207
duke@0 2208 __ flush_windows();
duke@0 2209 if (inner_frame_created) {
duke@0 2210 __ restore();
duke@0 2211 } else {
duke@0 2212 // Store only what we need from this frame
duke@0 2213 // QQQ I think that non-v9 (like we care) we don't need these saves
duke@0 2214 // either as the flush traps and the current window goes too.
duke@0 2215 __ st_ptr(FP, SP, FP->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
duke@0 2216 __ st_ptr(I7, SP, I7->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
duke@0 2217 }
duke@0 2218
duke@0 2219 // get JNIEnv* which is first argument to native
duke@0 2220
duke@0 2221 __ add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
duke@0 2222
duke@0 2223 // Use that pc we placed in O7 a while back as the current frame anchor
duke@0 2224
duke@0 2225 __ set_last_Java_frame(SP, O7);
duke@0 2226
duke@0 2227 // Transition from _thread_in_Java to _thread_in_native.
duke@0 2228 __ set(_thread_in_native, G3_scratch);
twisti@720 2229 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
duke@0 2230
duke@0 2231 // We flushed the windows ages ago now mark them as flushed
duke@0 2232
duke@0 2233 // mark windows as flushed
duke@0 2234 __ set(JavaFrameAnchor::flushed, G3_scratch);
duke@0 2235
twisti@720 2236 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
duke@0 2237
duke@0 2238 #ifdef _LP64
twisti@720 2239 AddressLiteral dest(method->native_function());
duke@0 2240 __ relocate(relocInfo::runtime_call_type);
twisti@720 2241 __ jumpl_to(dest, O7, O7);
duke@0 2242 #else
duke@0 2243 __ call(method->native_function(), relocInfo::runtime_call_type);
duke@0 2244 #endif
duke@0 2245 __ delayed()->st(G3_scratch, flags);
duke@0 2246
duke@0 2247 __ restore_thread(L7_thread_cache); // restore G2_thread
duke@0 2248
duke@0 2249 // Unpack native results. For int-types, we do any needed sign-extension
duke@0 2250 // and move things into I0. The return value there will survive any VM
duke@0 2251 // calls for blocking or unlocking. An FP or OOP result (handle) is done
duke@0 2252 // specially in the slow-path code.
duke@0 2253 switch (ret_type) {
duke@0 2254 case T_VOID: break; // Nothing to do!
duke@0 2255 case T_FLOAT: break; // Got it where we want it (unless slow-path)
duke@0 2256 case T_DOUBLE: break; // Got it where we want it (unless slow-path)
duke@0 2257 // In 64 bits build result is in O0, in O0, O1 in 32bit build
duke@0 2258 case T_LONG:
duke@0 2259 #ifndef _LP64
duke@0 2260 __ mov(O1, I1);
duke@0 2261 #endif
duke@0 2262 // Fall thru
duke@0 2263 case T_OBJECT: // Really a handle
duke@0 2264 case T_ARRAY:
duke@0 2265 case T_INT:
duke@0 2266 __ mov(O0, I0);
duke@0 2267 break;
duke@0 2268 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, I0); break; // !0 => true; 0 => false
duke@0 2269 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, I0); break;
duke@0 2270 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, I0); break; // cannot use and3, 0xFFFF too big as immediate value!
duke@0 2271 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, I0); break;
duke@0 2272 break; // Cannot de-handlize until after reclaiming jvm_lock
duke@0 2273 default:
duke@0 2274 ShouldNotReachHere();
duke@0 2275 }
duke@0 2276
duke@0 2277 // must we block?
duke@0 2278
duke@0 2279 // Block, if necessary, before resuming in _thread_in_Java state.
duke@0 2280 // In order for GC to work, don't clear the last_Java_sp until after blocking.
duke@0 2281 { Label no_block;
twisti@720 2282 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
duke@0 2283
duke@0 2284 // Switch thread to "native transition" state before reading the synchronization state.
duke@0 2285 // This additional state is necessary because reading and testing the synchronization
duke@0 2286 // state is not atomic w.r.t. GC, as this scenario demonstrates:
duke@0 2287 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
duke@0 2288 // VM thread changes sync state to synchronizing and suspends threads for GC.
duke@0 2289 // Thread A is resumed to finish this native method, but doesn't block here since it
duke@0 2290 // didn't see any synchronization is progress, and escapes.
duke@0 2291 __ set(_thread_in_native_trans, G3_scratch);
twisti@720 2292 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
duke@0 2293 if(os::is_MP()) {
duke@0 2294 if (UseMembar) {
duke@0 2295 // Force this write out before the read below
duke@0 2296 __ membar(Assembler::StoreLoad);
duke@0 2297 } else {
duke@0 2298 // Write serialization page so VM thread can do a pseudo remote membar.
duke@0 2299 // We use the current thread pointer to calculate a thread specific
duke@0 2300 // offset to write to within the page. This minimizes bus traffic
duke@0 2301 // due to cache line collision.
duke@0 2302 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
duke@0 2303 }
duke@0 2304 }
duke@0 2305 __ load_contents(sync_state, G3_scratch);
duke@0 2306 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
duke@0 2307
duke@0 2308 Label L;
twisti@720 2309 Address suspend_state(G2_thread, JavaThread::suspend_flags_offset());
duke@0 2310 __ br(Assembler::notEqual, false, Assembler::pn, L);
twisti@720 2311 __ delayed()->ld(suspend_state, G3_scratch);
duke@0 2312 __ cmp(G3_scratch, 0);
duke@0 2313 __ br(Assembler::equal, false, Assembler::pt, no_block);
duke@0 2314 __ delayed()->nop();
duke@0 2315 __ bind(L);
duke@0 2316
duke@0 2317 // Block. Save any potential method result value before the operation and
duke@0 2318 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
duke@0 2319 // lets us share the oopMap we used when we went native rather the create
duke@0 2320 // a distinct one for this pc
duke@0 2321 //
duke@0 2322 save_native_result(masm, ret_type, stack_slots);
duke@0 2323 __ call_VM_leaf(L7_thread_cache,
duke@0 2324 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
duke@0 2325 G2_thread);
duke@0 2326
duke@0 2327 // Restore any method result value
duke@0 2328 restore_native_result(masm, ret_type, stack_slots);
duke@0 2329 __ bind(no_block);
duke@0 2330 }
duke@0 2331
duke@0 2332 // thread state is thread_in_native_trans. Any safepoint blocking has already
duke@0 2333 // happened so we can now change state to _thread_in_Java.
duke@0 2334
duke@0 2335
duke@0 2336 __ set(_thread_in_Java, G3_scratch);
twisti@720 2337 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
duke@0 2338
duke@0 2339
duke@0 2340 Label no_reguard;
twisti@720 2341 __ ld(G2_thread, JavaThread::stack_guard_state_offset(), G3_scratch);
duke@0 2342 __ cmp(G3_scratch, JavaThread::stack_guard_yellow_disabled);
duke@0 2343 __ br(Assembler::notEqual, false, Assembler::pt, no_reguard);
duke@0 2344 __ delayed()->nop();
duke@0 2345
duke@0 2346 save_native_result(masm, ret_type, stack_slots);
duke@0 2347 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
duke@0 2348 __ delayed()->nop();
duke@0 2349
duke@0 2350 __ restore_thread(L7_thread_cache); // restore G2_thread
duke@0 2351 restore_native_result(masm, ret_type, stack_slots);
duke@0 2352
duke@0 2353 __ bind(no_reguard);
duke@0 2354
duke@0 2355 // Handle possible exception (will unlock if necessary)
duke@0 2356
duke@0 2357 // native result if any is live in freg or I0 (and I1 if long and 32bit vm)
duke@0 2358
duke@0 2359 // Unlock
duke@0 2360 if (method->is_synchronized()) {
duke@0 2361 Label done;
duke@0 2362 Register I2_ex_oop = I2;
duke@0 2363 const Register L3_box = L3;
duke@0 2364 // Get locked oop from the handle we passed to jni
duke@0 2365 __ ld_ptr(L6_handle, 0, L4);
duke@0 2366 __ add(SP, lock_offset+STACK_BIAS, L3_box);
duke@0 2367 // Must save pending exception around the slow-path VM call. Since it's a
duke@0 2368 // leaf call, the pending exception (if any) can be kept in a register.
duke@0 2369 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), I2_ex_oop);
duke@0 2370 // Now unlock
duke@0 2371 // (Roop, Rmark, Rbox, Rscratch)
duke@0 2372 __ compiler_unlock_object(L4, L1, L3_box, L2);
duke@0 2373 __ br(Assembler::equal, false, Assembler::pt, done);
duke@0 2374 __ delayed()-> add(SP, lock_offset+STACK_BIAS, L3_box);
duke@0 2375
duke@0 2376 // save and restore any potential method result value around the unlocking
duke@0 2377 // operation. Will save in I0 (or stack for FP returns).
duke@0 2378 save_native_result(masm, ret_type, stack_slots);
duke@0 2379
duke@0 2380 // Must clear pending-exception before re-entering the VM. Since this is
duke@0 2381 // a leaf call, pending-exception-oop can be safely kept in a register.
duke@0 2382 __ st_ptr(G0, G2_thread, in_bytes(Thread::pending_exception_offset()));
duke@0 2383
duke@0 2384 // slow case of monitor enter. Inline a special case of call_VM that
duke@0 2385 // disallows any pending_exception.
duke@0 2386 __ mov(L3_box, O1);
duke@0 2387
duke@0 2388 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), relocInfo::runtime_call_type);
duke@0 2389 __ delayed()->mov(L4, O0); // Need oop in O0
duke@0 2390
duke@0 2391 __ restore_thread(L7_thread_cache); // restore G2_thread
duke@0 2392
duke@0 2393 #ifdef ASSERT
duke@0 2394 { Label L;
duke@0 2395 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
duke@0 2396 __ br_null(O0, false, Assembler::pt, L);
duke@0 2397 __ delayed()->nop();
duke@0 2398 __ stop("no pending exception allowed on exit from IR::monitorexit");
duke@0 2399 __ bind(L);
duke@0 2400 }
duke@0 2401 #endif
duke@0 2402 restore_native_result(masm, ret_type, stack_slots);
duke@0 2403 // check_forward_pending_exception jump to forward_exception if any pending
duke@0 2404 // exception is set. The forward_exception routine expects to see the
duke@0 2405 // exception in pending_exception and not in a register. Kind of clumsy,
duke@0 2406 // since all folks who branch to forward_exception must have tested
duke@0 2407 // pending_exception first and hence have it in a register already.
duke@0 2408 __ st_ptr(I2_ex_oop, G2_thread, in_bytes(Thread::pending_exception_offset()));
duke@0 2409 __ bind(done);
duke@0 2410 }
duke@0 2411
duke@0 2412 // Tell dtrace about this method exit
duke@0 2413 {
duke@0 2414 SkipIfEqual skip_if(
duke@0 2415 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
duke@0 2416 save_native_result(masm, ret_type, stack_slots);
duke@0 2417 __ set_oop_constant(JNIHandles::make_local(method()), O1);
duke@0 2418 __ call_VM_leaf(L7_thread_cache,
duke@0 2419 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
duke@0 2420 G2_thread, O1);
duke@0 2421 restore_native_result(masm, ret_type, stack_slots);
duke@0 2422 }
duke@0 2423
duke@0 2424 // Clear "last Java frame" SP and PC.
duke@0 2425 __ verify_thread(); // G2_thread must be correct
duke@0 2426 __ reset_last_Java_frame();
duke@0 2427
duke@0 2428 // Unpack oop result
duke@0 2429 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
duke@0 2430 Label L;
duke@0 2431 __ addcc(G0, I0, G0);
duke@0 2432 __ brx(Assembler::notZero, true, Assembler::pt, L);
duke@0 2433 __ delayed()->ld_ptr(I0, 0, I0);
duke@0 2434 __ mov(G0, I0);
duke@0 2435 __ bind(L);
duke@0 2436 __ verify_oop(I0);
duke@0 2437 }
duke@0 2438
duke@0 2439 // reset handle block
duke@0 2440 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), L5);
duke@0 2441 __ st_ptr(G0, L5, JNIHandleBlock::top_offset_in_bytes());
duke@0 2442
duke@0 2443 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), G3_scratch);
duke@0 2444 check_forward_pending_exception(masm, G3_scratch);
duke@0 2445
duke@0 2446
duke@0 2447 // Return
duke@0 2448
duke@0 2449 #ifndef _LP64
duke@0 2450 if (ret_type == T_LONG) {
duke@0 2451
duke@0 2452 // Must leave proper result in O0,O1 and G1 (c2/tiered only)
duke@0 2453 __ sllx(I0, 32, G1); // Shift bits into high G1
duke@0 2454 __ srl (I1, 0, I1); // Zero extend O1 (harmless?)
duke@0 2455 __ or3 (I1, G1, G1); // OR 64 bits into G1
duke@0 2456 }
duke@0 2457 #endif
duke@0 2458
duke@0 2459 __ ret();
duke@0 2460 __ delayed()->restore();
duke@0 2461
duke@0 2462 __ flush();
duke@0 2463
duke@0 2464 nmethod *nm = nmethod::new_native_nmethod(method,
duke@0 2465 masm->code(),
duke@0 2466 vep_offset,
duke@0 2467 frame_complete,
duke@0 2468 stack_slots / VMRegImpl::slots_per_word,
duke@0 2469 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
duke@0 2470 in_ByteSize(lock_offset),
duke@0 2471 oop_maps);
duke@0 2472 return nm;
duke@0 2473
duke@0 2474 }
duke@0 2475
kamg@124 2476 #ifdef HAVE_DTRACE_H
kamg@124 2477 // ---------------------------------------------------------------------------
kamg@124 2478 // Generate a dtrace nmethod for a given signature. The method takes arguments
kamg@124 2479 // in the Java compiled code convention, marshals them to the native
kamg@124 2480 // abi and then leaves nops at the position you would expect to call a native
kamg@124 2481 // function. When the probe is enabled the nops are replaced with a trap
kamg@124 2482 // instruction that dtrace inserts and the trace will cause a notification
kamg@124 2483 // to dtrace.
kamg@124 2484 //
kamg@124 2485 // The probes are only able to take primitive types and java/lang/String as
kamg@124 2486 // arguments. No other java types are allowed. Strings are converted to utf8
kamg@124 2487 // strings so that from dtrace point of view java strings are converted to C
kamg@124 2488 // strings. There is an arbitrary fixed limit on the total space that a method
kamg@124 2489 // can use for converting the strings. (256 chars per string in the signature).
kamg@124 2490 // So any java string larger then this is truncated.
kamg@124 2491
kamg@124 2492 static int fp_offset[ConcreteRegisterImpl::number_of_registers] = { 0 };
kamg@124 2493 static bool offsets_initialized = false;
kamg@124 2494
kamg@124 2495 static VMRegPair reg64_to_VMRegPair(Register r) {
kamg@124 2496 VMRegPair ret;
kamg@124 2497 if (wordSize == 8) {
kamg@124 2498 ret.set2(r->as_VMReg());
kamg@124 2499 } else {
kamg@124 2500 ret.set_pair(r->successor()->as_VMReg(), r->as_VMReg());
kamg@124 2501 }
kamg@124 2502 return ret;
kamg@124 2503 }
kamg@124 2504
kamg@124 2505
kamg@124 2506 nmethod *SharedRuntime::generate_dtrace_nmethod(
kamg@124 2507 MacroAssembler *masm, methodHandle method) {
kamg@124 2508
kamg@124 2509
kamg@124 2510 // generate_dtrace_nmethod is guarded by a mutex so we are sure to
kamg@124 2511 // be single threaded in this method.
kamg@124 2512 assert(AdapterHandlerLibrary_lock->owned_by_self(), "must be");
kamg@124 2513
kamg@124 2514 // Fill in the signature array, for the calling-convention call.
kamg@124 2515 int total_args_passed = method->size_of_parameters();
kamg@124 2516
kamg@124 2517 BasicType* in_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed);
kamg@124 2518 VMRegPair *in_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed);
kamg@124 2519
kamg@124 2520 // The signature we are going to use for the trap that dtrace will see
kamg@124 2521 // java/lang/String is converted. We drop "this" and any other object
kamg@124 2522 // is converted to NULL. (A one-slot java/lang/Long object reference
kamg@124 2523 // is converted to a two-slot long, which is why we double the allocation).
kamg@124 2524 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed * 2);
kamg@124 2525 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed * 2);
kamg@124 2526
kamg@124 2527 int i=0;
kamg@124 2528 int total_strings = 0;
kamg@124 2529 int first_arg_to_pass = 0;
kamg@124 2530 int total_c_args = 0;
kamg@124 2531
kamg@124 2532 // Skip the receiver as dtrace doesn't want to see it
kamg@124 2533 if( !method->is_static() ) {
kamg@124 2534 in_sig_bt[i++] = T_OBJECT;
kamg@124 2535 first_arg_to_pass = 1;
kamg@124 2536 }
kamg@124 2537
kamg@124 2538 SignatureStream ss(method->signature());
kamg@124 2539 for ( ; !ss.at_return_type(); ss.next()) {
kamg@124 2540 BasicType bt = ss.type();
kamg@124 2541 in_sig_bt[i++] = bt; // Collect remaining bits of signature
kamg@124 2542 out_sig_bt[total_c_args++] = bt;
kamg@124 2543 if( bt == T_OBJECT) {
kamg@124 2544 symbolOop s = ss.as_symbol_or_null();
kamg@124 2545 if (s == vmSymbols::java_lang_String()) {
kamg@124 2546 total_strings++;
kamg@124 2547 out_sig_bt[total_c_args-1] = T_ADDRESS;
kamg@124 2548 } else if (s == vmSymbols::java_lang_Boolean() ||
kamg@124 2549 s == vmSymbols::java_lang_Byte()) {
kamg@124 2550 out_sig_bt[total_c_args-1] = T_BYTE;
kamg@124 2551 } else if (s == vmSymbols::java_lang_Character() ||
kamg@124 2552 s == vmSymbols::java_lang_Short()) {
kamg@124 2553 out_sig_bt[total_c_args-1] = T_SHORT;
kamg@124 2554 } else if (s == vmSymbols::java_lang_Integer() ||
kamg@124 2555 s == vmSymbols::java_lang_Float()) {
kamg@124 2556 out_sig_bt[total_c_args-1] = T_INT;
kamg@124 2557 } else if (s == vmSymbols::java_lang_Long() ||
kamg@124 2558 s == vmSymbols::java_lang_Double()) {
kamg@124 2559 out_sig_bt[total_c_args-1] = T_LONG;
kamg@124 2560 out_sig_bt[total_c_args++] = T_VOID;
kamg@124 2561 }
kamg@124 2562 } else if ( bt == T_LONG || bt == T_DOUBLE ) {
kamg@124 2563 in_sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots
kamg@124 2564 // We convert double to long
kamg@124 2565 out_sig_bt[total_c_args-1] = T_LONG;
kamg@124 2566 out_sig_bt[total_c_args++] = T_VOID;
kamg@124 2567 } else if ( bt == T_FLOAT) {
kamg@124 2568 // We convert float to int
kamg@124 2569 out_sig_bt[total_c_args-1] = T_INT;
kamg@124 2570 }
kamg@124 2571 }
kamg@124 2572
kamg@124 2573 assert(i==total_args_passed, "validly parsed signature");
kamg@124 2574
kamg@124 2575 // Now get the compiled-Java layout as input arguments
kamg@124 2576 int comp_args_on_stack;
kamg@124 2577 comp_args_on_stack = SharedRuntime::java_calling_convention(
kamg@124 2578 in_sig_bt, in_regs, total_args_passed, false);
kamg@124 2579
kamg@124 2580 // We have received a description of where all the java arg are located
kamg@124 2581 // on entry to the wrapper. We need to convert these args to where
kamg@124 2582 // the a native (non-jni) function would expect them. To figure out
kamg@124 2583 // where they go we convert the java signature to a C signature and remove
kamg@124 2584 // T_VOID for any long/double we might have received.
kamg@124 2585
kamg@124 2586
kamg@124 2587 // Now figure out where the args must be stored and how much stack space
kamg@124 2588 // they require (neglecting out_preserve_stack_slots but space for storing
kamg@124 2589 // the 1st six register arguments). It's weird see int_stk_helper.
kamg@124 2590 //
kamg@124 2591 int out_arg_slots;
kamg@124 2592 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
kamg@124 2593
kamg@124 2594 // Calculate the total number of stack slots we will need.
kamg@124 2595
kamg@124 2596 // First count the abi requirement plus all of the outgoing args
kamg@124 2597 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
kamg@124 2598
kamg@124 2599 // Plus a temp for possible converion of float/double/long register args
kamg@124 2600
kamg@124 2601 int conversion_temp = stack_slots;
kamg@124 2602 stack_slots += 2;
kamg@124 2603
kamg@124 2604
kamg@124 2605 // Now space for the string(s) we must convert
kamg@124 2606
kamg@124 2607 int string_locs = stack_slots;
kamg@124 2608 stack_slots += total_strings *
kamg@124 2609 (max_dtrace_string_size / VMRegImpl::stack_slot_size);
kamg@124 2610
kamg@124 2611 // Ok The space we have allocated will look like:
kamg@124 2612 //
kamg@124 2613 //
kamg@124 2614 // FP-> | |
kamg@124 2615 // |---------------------|
kamg@124 2616 // | string[n] |
kamg@124 2617 // |---------------------| <- string_locs[n]
kamg@124 2618 // | string[n-1] |
kamg@124 2619 // |---------------------| <- string_locs[n-1]
kamg@124 2620 // | ... |
kamg@124 2621 // | ... |
kamg@124 2622 // |---------------------| <- string_locs[1]
kamg@124 2623 // | string[0] |
kamg@124 2624 // |---------------------| <- string_locs[0]
kamg@124 2625 // | temp |
kamg@124 2626 // |---------------------| <- conversion_temp
kamg@124 2627 // | outbound memory |
kamg@124 2628 // | based arguments |
kamg@124 2629 // | |
kamg@124 2630 // |---------------------|
kamg@124 2631 // | |
kamg@124 2632 // SP-> | out_preserved_slots |
kamg@124 2633 //
kamg@124 2634 //
kamg@124 2635
kamg@124 2636 // Now compute actual number of stack words we need rounding to make
kamg@124 2637 // stack properly aligned.
kamg@124 2638 stack_slots = round_to(stack_slots, 4 * VMRegImpl::slots_per_word);
kamg@124 2639
kamg@124 2640 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
kamg@124 2641
kamg@124 2642 intptr_t start = (intptr_t)__ pc();
kamg@124 2643
kamg@124 2644 // First thing make an ic check to see if we should even be here
kamg@124 2645
kamg@124 2646 {
kamg@124 2647 Label L;
kamg@124 2648 const Register temp_reg = G3_scratch;
twisti@720 2649 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
kamg@124 2650 __ verify_oop(O0);
kamg@124 2651 __ ld_ptr(O0, oopDesc::klass_offset_in_bytes(), temp_reg);
kamg@124 2652 __ cmp(temp_reg, G5_inline_cache_reg);
kamg@124 2653 __ brx(Assembler::equal, true, Assembler::pt, L);
kamg@124 2654 __ delayed()->nop();
kamg@124 2655
twisti@720 2656 __ jump_to(ic_miss, temp_reg);
kamg@124 2657 __ delayed()->nop();
kamg@124 2658 __ align(CodeEntryAlignment);
kamg@124 2659 __ bind(L);
kamg@124 2660 }
kamg@124 2661
kamg@124 2662 int vep_offset = ((intptr_t)__ pc()) - start;
kamg@124 2663
kamg@124 2664
kamg@124 2665 // The instruction at the verified entry point must be 5 bytes or longer
kamg@124 2666 // because it can be patched on the fly by make_non_entrant. The stack bang
kamg@124 2667 // instruction fits that requirement.
kamg@124 2668
kamg@124 2669 // Generate stack overflow check before creating frame
kamg@124 2670 __ generate_stack_overflow_check(stack_size);
kamg@124 2671
kamg@124 2672 assert(((intptr_t)__ pc() - start - vep_offset) >= 5,
kamg@124 2673 "valid size for make_non_entrant");
kamg@124 2674
kamg@124 2675 // Generate a new frame for the wrapper.
kamg@124 2676 __ save(SP, -stack_size, SP);
kamg@124 2677
kamg@124 2678 // Frame is now completed as far a size and linkage.
kamg@124 2679
kamg@124 2680 int frame_complete = ((intptr_t)__ pc()) - start;
kamg@124 2681
kamg@124 2682 #ifdef ASSERT
kamg@124 2683 bool reg_destroyed[RegisterImpl::number_of_registers];
kamg@124 2684 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
kamg@124 2685 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
kamg@124 2686 reg_destroyed[r] = false;
kamg@124 2687 }
kamg@124 2688 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
kamg@124 2689 freg_destroyed[f] = false;
kamg@124 2690 }
kamg@124 2691
kamg@124 2692 #endif /* ASSERT */
kamg@124 2693
kamg@124 2694 VMRegPair zero;
kamg@182 2695 const Register g0 = G0; // without this we get a compiler warning (why??)
kamg@182 2696 zero.set2(g0->as_VMReg());
kamg@124 2697
kamg@124 2698 int c_arg, j_arg;
kamg@124 2699
kamg@124 2700 Register conversion_off = noreg;
kamg@124 2701
kamg@124 2702 for (j_arg = first_arg_to_pass, c_arg = 0 ;
kamg@124 2703 j_arg < total_args_passed ; j_arg++, c_arg++ ) {
kamg@124 2704
kamg@124 2705 VMRegPair src = in_regs[j_arg];
kamg@124 2706 VMRegPair dst = out_regs[c_arg];
kamg@124 2707
kamg@124 2708 #ifdef ASSERT
kamg@124 2709 if (src.first()->is_Register()) {
kamg@124 2710 assert(!reg_destroyed[src.first()->as_Register()->encoding()], "ack!");
kamg@124 2711 } else if (src.first()->is_FloatRegister()) {
kamg@124 2712 assert(!freg_destroyed[src.first()->as_FloatRegister()->encoding(
kamg@124 2713 FloatRegisterImpl::S)], "ack!");
kamg@124 2714 }
kamg@124 2715 if (dst.first()->is_Register()) {
kamg@124 2716 reg_destroyed[dst.first()->as_Register()->encoding()] = true;
kamg@124 2717 } else if (dst.first()->is_FloatRegister()) {
kamg@124 2718 freg_destroyed[dst.first()->as_FloatRegister()->encoding(
kamg@124 2719 FloatRegisterImpl::S)] = true;
kamg@124 2720 }
kamg@124 2721 #endif /* ASSERT */
kamg@124 2722
kamg@124 2723 switch (in_sig_bt[j_arg]) {
kamg@124 2724 case T_ARRAY:
kamg@124 2725 case T_OBJECT:
kamg@124 2726 {
kamg@124 2727 if (out_sig_bt[c_arg] == T_BYTE || out_sig_bt[c_arg] == T_SHORT ||
kamg@124 2728 out_sig_bt[c_arg] == T_INT || out_sig_bt[c_arg] == T_LONG) {
kamg@124 2729 // need to unbox a one-slot value
kamg@124 2730 Register in_reg = L0;
kamg@124 2731 Register tmp = L2;
kamg@124 2732 if ( src.first()->is_reg() ) {
kamg@124 2733 in_reg = src.first()->as_Register();
kamg@124 2734 } else {
kamg@124 2735 assert(Assembler::is_simm13(reg2offset(src.first()) + STACK_BIAS),
kamg@124 2736 "must be");
kamg@124 2737 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, in_reg);
kamg@124 2738 }
kamg@124 2739 // If the final destination is an acceptable register
kamg@124 2740 if ( dst.first()->is_reg() ) {
kamg@124 2741 if ( dst.is_single_phys_reg() || out_sig_bt[c_arg] != T_LONG ) {
kamg@124 2742 tmp = dst.first()->as_Register();
kamg@124 2743 }
kamg@124 2744 }
kamg@124 2745
kamg@124 2746 Label skipUnbox;
kamg@124 2747 if ( wordSize == 4 && out_sig_bt[c_arg] == T_LONG ) {
kamg@124 2748 __ mov(G0, tmp->successor());
kamg@124 2749 }
kamg@124 2750 __ br_null(in_reg, true, Assembler::pn, skipUnbox);
kamg@124 2751 __ delayed()->mov(G0, tmp);
kamg@124 2752
kvn@153 2753 BasicType bt = out_sig_bt[c_arg];
kvn@153 2754 int box_offset = java_lang_boxing_object::value_offset_in_bytes(bt);
kvn@153 2755 switch (bt) {
kamg@124 2756 case T_BYTE:
kamg@124 2757 __ ldub(in_reg, box_offset, tmp); break;
kamg@124 2758 case T_SHORT:
kamg@124 2759 __ lduh(in_reg, box_offset, tmp); break;
kamg@124 2760 case T_INT:
kamg@124 2761 __ ld(in_reg, box_offset, tmp); break;
kamg@124 2762 case T_LONG:
kamg@124 2763 __ ld_long(in_reg, box_offset, tmp); break;
kamg@124 2764 default: ShouldNotReachHere();
kamg@124 2765 }
kamg@124 2766
kamg@124 2767 __ bind(skipUnbox);
kamg@124 2768 // If tmp wasn't final destination copy to final destination
kamg@124 2769 if (tmp == L2) {
kamg@124 2770 VMRegPair tmp_as_VM = reg64_to_VMRegPair(L2);
kamg@124 2771 if (out_sig_bt[c_arg] == T_LONG) {
kamg@124 2772 long_move(masm, tmp_as_VM, dst);
kamg@124 2773 } else {
kamg@124 2774 move32_64(masm, tmp_as_VM, out_regs[c_arg]);
kamg@124 2775 }
kamg@124 2776 }
kamg@124 2777 if (out_sig_bt[c_arg] == T_LONG) {
kamg@124 2778 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
kamg@124 2779 ++c_arg; // move over the T_VOID to keep the loop indices in sync
kamg@124 2780 }
kamg@124 2781 } else if (out_sig_bt[c_arg] == T_ADDRESS) {
kamg@124 2782 Register s =
kamg@124 2783 src.first()->is_reg() ? src.first()->as_Register() : L2;
kamg@124 2784 Register d =
kamg@124 2785 dst.first()->is_reg() ? dst.first()->as_Register() : L2;
kamg@124 2786
kamg@124 2787 // We store the oop now so that the conversion pass can reach
kamg@124 2788 // while in the inner frame. This will be the only store if
kamg@124 2789 // the oop is NULL.
kamg@124 2790 if (s != L2) {
kamg@124 2791 // src is register
kamg@124 2792 if (d != L2) {
kamg@124 2793 // dst is register
kamg@124 2794 __ mov(s, d);
kamg@124 2795 } else {
kamg@124 2796 assert(Assembler::is_simm13(reg2offset(dst.first()) +
kamg@124 2797 STACK_BIAS), "must be");
kamg@124 2798 __ st_ptr(s, SP, reg2offset(dst.first()) + STACK_BIAS);
kamg@124 2799 }
kamg@124 2800 } else {
kamg@124 2801 // src not a register
kamg@124 2802 assert(Assembler::is_simm13(reg2offset(src.first()) +
kamg@124 2803 STACK_BIAS), "must be");
kamg@124 2804 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, d);
kamg@124 2805 if (d == L2) {
kamg@124 2806 assert(Assembler::is_simm13(reg2offset(dst.first()) +
kamg@124 2807 STACK_BIAS), "must be");
kamg@124 2808 __ st_ptr(d, SP, reg2offset(dst.first()) + STACK_BIAS);
kamg@124 2809 }
kamg@124 2810 }
kamg@124 2811 } else if (out_sig_bt[c_arg] != T_VOID) {
kamg@124 2812 // Convert the arg to NULL
kamg@124 2813 if (dst.first()->is_reg()) {
kamg@124 2814 __ mov(G0, dst.first()->as_Register());
kamg@124 2815 } else {
kamg@124 2816 assert(Assembler::is_simm13(reg2offset(dst.first()) +
kamg@124 2817 STACK_BIAS), "must be");
kamg@124 2818 __ st_ptr(G0, SP, reg2offset(dst.first()) + STACK_BIAS);
kamg@124 2819 }
kamg@124 2820 }
kamg@124 2821 }
kamg@124 2822 break;
kamg@124 2823 case T_VOID:
kamg@124 2824 break;
kamg@124 2825
kamg@124 2826 case T_FLOAT:
kamg@124 2827 if (src.first()->is_stack()) {
kamg@124 2828 // Stack to stack/reg is simple
kamg@124 2829 move32_64(masm, src, dst);
kamg@124 2830 } else {
kamg@124 2831 if (dst.first()->is_reg()) {
kamg@124 2832 // freg -> reg
kamg@124 2833 int off =
kamg@124 2834 STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
kamg@124 2835 Register d = dst.first()->as_Register();
kamg@124 2836 if (Assembler::is_simm13(off)) {
kamg@124 2837 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
kamg@124 2838 SP, off);
kamg@124 2839 __ ld(SP, off, d);
kamg@124 2840 } else {
kamg@124 2841 if (conversion_off == noreg) {
kamg@124 2842 __ set(off, L6);
kamg@124 2843 conversion_off = L6;
kamg@124 2844 }
kamg@124 2845 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
kamg@124 2846 SP, conversion_off);
kamg@124 2847 __ ld(SP, conversion_off , d);
kamg@124 2848 }
kamg@124 2849 } else {
kamg@124 2850 // freg -> mem
kamg@124 2851 int off = STACK_BIAS + reg2offset(dst.first());
kamg@124 2852 if (Assembler::is_simm13(off)) {
kamg@124 2853 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
kamg@124 2854 SP, off);
kamg@124 2855 } else {
kamg@124 2856 if (conversion_off == noreg) {
kamg@124 2857 __ set(off, L6);
kamg@124 2858 conversion_off = L6;
kamg@124 2859 }
kamg@124 2860 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
kamg@124 2861 SP, conversion_off);
kamg@124 2862 }
kamg@124 2863 }
kamg@124 2864 }
kamg@124 2865 break;
kamg@124 2866
kamg@124 2867 case T_DOUBLE:
kamg@124 2868 assert( j_arg + 1 < total_args_passed &&
kamg@124 2869 in_sig_bt[j_arg + 1] == T_VOID &&
kamg@124 2870 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
kamg@124 2871 if (src.first()->is_stack()) {
kamg@124 2872 // Stack to stack/reg is simple
kamg@124 2873 long_move(masm, src, dst);
kamg@124 2874 } else {
kamg@124 2875 Register d = dst.first()->is_reg() ? dst.first()->as_Register() : L2;
kamg@124 2876
kamg@124 2877 // Destination could be an odd reg on 32bit in which case
kamg@124 2878 // we can't load direct to the destination.
kamg@124 2879
kamg@124 2880 if (!d->is_even() && wordSize == 4) {
kamg@124 2881 d = L2;
kamg@124 2882 }
kamg@124 2883 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
kamg@124 2884 if (Assembler::is_simm13(off)) {
kamg@124 2885 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
kamg@124 2886 SP, off);
kamg@124 2887 __ ld_long(SP, off, d);
kamg@124 2888 } else {
kamg@124 2889 if (conversion_off == noreg) {
kamg@124 2890 __ set(off, L6);
kamg@124 2891 conversion_off = L6;
kamg@124 2892 }
kamg@124 2893 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
kamg@124 2894 SP, conversion_off);
kamg@124 2895 __ ld_long(SP, conversion_off, d);
kamg@124 2896 }
kamg@124 2897 if (d == L2) {
kamg@124 2898 long_move(masm, reg64_to_VMRegPair(L2), dst);
kamg@124 2899 }
kamg@124 2900 }
kamg@124 2901 break;
kamg@124 2902
kamg@124 2903 case T_LONG :
kamg@124 2904 // 32bit can't do a split move of something like g1 -> O0, O1
kamg@124 2905 // so use a memory temp
kamg@124 2906 if (src.is_single_phys_reg() && wordSize == 4) {
kamg@124 2907 Register tmp = L2;
kamg@124 2908 if (dst.first()->is_reg() &&
kamg@124 2909 (wordSize == 8 || dst.first()->as_Register()->is_even())) {
kamg@124 2910 tmp = dst.first()->as_Register();
kamg@124 2911 }
kamg@124 2912
kamg@124 2913 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
kamg@124 2914 if (Assembler::is_simm13(off)) {
kamg@124 2915 __ stx(src.first()->as_Register(), SP, off);
kamg@124 2916 __ ld_long(SP, off, tmp);
kamg@124 2917 } else {
kamg@124 2918 if (conversion_off == noreg) {
kamg@124 2919 __ set(off, L6);
kamg@124 2920 conversion_off = L6;
kamg@124 2921 }
kamg@124 2922 __ stx(src.first()->as_Register(), SP, conversion_off);
kamg@124 2923 __ ld_long(SP, conversion_off, tmp);
kamg@124 2924 }
kamg@124 2925
kamg@124 2926 if (tmp == L2) {
kamg@124 2927 long_move(masm, reg64_to_VMRegPair(L2), dst);
kamg@124 2928 }
kamg@124 2929 } else {
kamg@124 2930 long_move(masm, src, dst);
kamg@124 2931 }
kamg@124 2932 break;
kamg@124 2933
kamg@124 2934 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
kamg@124 2935
kamg@124 2936 default:
kamg@124 2937 move32_64(masm, src, dst);
kamg@124 2938 }
kamg@124 2939 }
kamg@124 2940
kamg@124 2941
kamg@124 2942 // If we have any strings we must store any register based arg to the stack
kamg@124 2943 // This includes any still live xmm registers too.
kamg@124 2944
kamg@124 2945 if (total_strings > 0 ) {
kamg@124 2946
kamg@124 2947 // protect all the arg registers
kamg@124 2948 __ save_frame(0);
kamg@124 2949 __ mov(G2_thread, L7_thread_cache);
kamg@124 2950 const Register L2_string_off = L2;
kamg@124 2951
kamg@124 2952 // Get first string offset
kamg@124 2953 __ set(string_locs * VMRegImpl::stack_slot_size, L2_string_off);
kamg@124 2954
kamg@124 2955 for (c_arg = 0 ; c_arg < total_c_args ; c_arg++ ) {
kamg@124 2956 if (out_sig_bt[c_arg] == T_ADDRESS) {
kamg@124 2957
kamg@124 2958 VMRegPair dst = out_regs[c_arg];
kamg@124 2959 const Register d = dst.first()->is_reg() ?
kamg@124 2960 dst.first()->as_Register()->after_save() : noreg;
kamg@124 2961
kamg@124 2962 // It's a string the oop and it was already copied to the out arg
kamg@124 2963 // position
kamg@124 2964 if (d != noreg) {
kamg@124 2965 __ mov(d, O0);
kamg@124 2966 } else {
kamg@124 2967 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
kamg@124 2968 "must be");
kamg@124 2969 __ ld_ptr(FP, reg2offset(dst.first()) + STACK_BIAS, O0);
kamg@124 2970 }
kamg@124 2971 Label skip;
kamg@124 2972
kamg@124 2973 __ br_null(O0, false, Assembler::pn, skip);
kamg@124 2974 __ delayed()->add(FP, L2_string_off, O1);
kamg@124 2975
kamg@124 2976 if (d != noreg) {
kamg@124 2977 __ mov(O1, d);
kamg@124 2978 } else {
kamg@124 2979 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
kamg@124 2980 "must be");
kamg@124 2981 __ st_ptr(O1, FP, reg2offset(dst.first()) + STACK_BIAS);
kamg@124 2982 }
kamg@124 2983
kamg@124 2984 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::get_utf),
kamg@124 2985 relocInfo::runtime_call_type);
kamg@124 2986 __ delayed()->add(L2_string_off, max_dtrace_string_size, L2_string_off);
kamg@124 2987
kamg@124 2988 __ bind(skip);
kamg@124 2989
kamg@124 2990 }
kamg@124 2991
kamg@124 2992 }
kamg@124 2993 __ mov(L7_thread_cache, G2_thread);
kamg@124 2994 __ restore();
kamg@124 2995
kamg@124 2996 }
kamg@124 2997
kamg@124 2998
kamg@124 2999 // Ok now we are done. Need to place the nop that dtrace wants in order to
kamg@124 3000 // patch in the trap
kamg@124 3001
kamg@124 3002 int patch_offset = ((intptr_t)__ pc()) - start;
kamg@124 3003
kamg@124 3004 __ nop();
kamg@124 3005
kamg@124 3006
kamg@124 3007 // Return
kamg@124 3008
kamg@124 3009 __ ret();
kamg@124 3010 __ delayed()->restore();
kamg@124 3011
kamg@124 3012 __ flush();
kamg@124 3013
kamg@124 3014 nmethod *nm = nmethod::new_dtrace_nmethod(
kamg@124 3015 method, masm->code(), vep_offset, patch_offset, frame_complete,
kamg@124 3016 stack_slots / VMRegImpl::slots_per_word);
kamg@124 3017 return nm;
kamg@124 3018
kamg@124 3019 }
kamg@124 3020
kamg@124 3021 #endif // HAVE_DTRACE_H
kamg@124 3022
duke@0 3023 // this function returns the adjust size (in number of words) to a c2i adapter
duke@0 3024 // activation for use during deoptimization
duke@0 3025 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
duke@0 3026 assert(callee_locals >= callee_parameters,
duke@0 3027 "test and remove; got more parms than locals");
duke@0 3028 if (callee_locals < callee_parameters)
duke@0 3029 return 0; // No adjustment for negative locals
twisti@1401 3030 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
duke@0 3031 return round_to(diff, WordsPerLong);
duke@0 3032 }
duke@0 3033
duke@0 3034 // "Top of Stack" slots that may be unused by the calling convention but must
duke@0 3035 // otherwise be preserved.
duke@0 3036 // On Intel these are not necessary and the value can be zero.
duke@0 3037 // On Sparc this describes the words reserved for storing a register window
duke@0 3038 // when an interrupt occurs.
duke@0 3039 uint SharedRuntime::out_preserve_stack_slots() {
duke@0 3040 return frame::register_save_words * VMRegImpl::slots_per_word;
duke@0 3041 }
duke@0 3042
duke@0 3043 static void gen_new_frame(MacroAssembler* masm, bool deopt) {
duke@0 3044 //
duke@0 3045 // Common out the new frame generation for deopt and uncommon trap
duke@0 3046 //
duke@0 3047 Register G3pcs = G3_scratch; // Array of new pcs (input)
duke@0 3048 Register Oreturn0 = O0;
duke@0 3049 Register Oreturn1 = O1;
duke@0 3050 Register O2UnrollBlock = O2;
duke@0 3051 Register O3array = O3; // Array of frame sizes (input)
duke@0 3052 Register O4array_size = O4; // number of frames (input)
duke@0 3053 Register O7frame_size = O7; // number of frames (input)
duke@0 3054
duke@0 3055 __ ld_ptr(O3array, 0, O7frame_size);
duke@0 3056 __ sub(G0, O7frame_size, O7frame_size);
duke@0 3057 __ save(SP, O7frame_size, SP);
duke@0 3058 __ ld_ptr(G3pcs, 0, I7); // load frame's new pc
duke@0 3059
duke@0 3060 #ifdef ASSERT
duke@0 3061 // make sure that the frames are aligned properly
duke@0 3062 #ifndef _LP64
duke@0 3063 __ btst(wordSize*2-1, SP);
duke@0 3064 __ breakpoint_trap(Assembler::notZero);
duke@0 3065 #endif
duke@0 3066 #endif
duke@0 3067
duke@0 3068 // Deopt needs to pass some extra live values from frame to frame
duke@0 3069
duke@0 3070 if (deopt) {
duke@0 3071 __ mov(Oreturn0->after_save(), Oreturn0);
duke@0 3072 __ mov(Oreturn1->after_save(), Oreturn1);
duke@0 3073 }
duke@0 3074
duke@0 3075 __ mov(O4array_size->after_save(), O4array_size);
duke@0 3076 __ sub(O4array_size, 1, O4array_size);
duke@0 3077 __ mov(O3array->after_save(), O3array);
duke@0 3078 __ mov(O2UnrollBlock->after_save(), O2UnrollBlock);
duke@0 3079 __ add(G3pcs, wordSize, G3pcs); // point to next pc value
duke@0 3080
duke@0 3081 #ifdef ASSERT
duke@0 3082 // trash registers to show a clear pattern in backtraces
duke@0 3083 __ set(0xDEAD0000, I0);
duke@0 3084 __ add(I0, 2, I1);
duke@0 3085 __ add(I0, 4, I2);
duke@0 3086 __ add(I0, 6, I3);
duke@0 3087 __ add(I0, 8, I4);
duke@0 3088 // Don't touch I5 could have valuable savedSP
duke@0 3089 __ set(0xDEADBEEF, L0);
duke@0 3090 __ mov(L0, L1);
duke@0 3091 __ mov(L0, L2);
duke@0 3092 __ mov(L0, L3);
duke@0 3093 __ mov(L0, L4);
duke@0 3094 __ mov(L0, L5);
duke@0 3095
duke@0 3096 // trash the return value as there is nothing to return yet
duke@0 3097 __ set(0xDEAD0001, O7);
duke@0 3098 #endif
duke@0 3099
duke@0 3100 __ mov(SP, O5_savedSP);
duke@0 3101 }
duke@0 3102
duke@0 3103
duke@0 3104 static void make_new_frames(MacroAssembler* masm, bool deopt) {
duke@0 3105 //
duke@0 3106 // loop through the UnrollBlock info and create new frames
duke@0 3107 //
duke@0 3108 Register G3pcs = G3_scratch;
duke@0 3109 Register Oreturn0 = O0;
duke@0 3110 Register Oreturn1 = O1;
duke@0 3111 Register O2UnrollBlock = O2;
duke@0 3112 Register O3array = O3;
duke@0 3113 Register O4array_size = O4;
duke@0 3114 Label loop;
duke@0 3115
duke@0 3116 // Before we make new frames, check to see if stack is available.
duke@0 3117 // Do this after the caller's return address is on top of stack
duke@0 3118 if (UseStackBanging) {
duke@0 3119 // Get total frame size for interpreted frames
twisti@720 3120 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes(), O4);
duke@0 3121 __ bang_stack_size(O4, O3, G3_scratch);
duke@0 3122 }
duke@0 3123
twisti@720 3124 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(), O4array_size);
twisti@720 3125 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(), G3pcs);
twisti@720 3126 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(), O3array);
duke@0 3127
duke@0 3128 // Adjust old interpreter frame to make space for new frame's extra java locals
duke@0 3129 //
duke@0 3130 // We capture the original sp for the transition frame only because it is needed in
duke@0 3131 // order to properly calculate interpreter_sp_adjustment. Even though in real life
duke@0 3132 // every interpreter frame captures a savedSP it is only needed at the transition
duke@0 3133 // (fortunately). If we had to have it correct everywhere then we would need to
duke@0 3134 // be told the sp_adjustment for each frame we create. If the frame size array
duke@0 3135 // were to have twice the frame count entries then we could have pairs [sp_adjustment, frame_size]
duke@0 3136 // for each frame we create and keep up the illusion every where.
duke@0 3137 //
duke@0 3138
twisti@720 3139 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(), O7);
duke@0 3140 __ mov(SP, O5_savedSP); // remember initial sender's original sp before adjustment
duke@0 3141 __ sub(SP, O7, SP);
duke@0 3142
duke@0 3143 #ifdef ASSERT
duke@0 3144 // make sure that there is at least one entry in the array
duke@0 3145 __ tst(O4array_size);
duke@0 3146 __ breakpoint_trap(Assembler::zero);
duke@0 3147 #endif
duke@0 3148
duke@0 3149 // Now push the new interpreter frames
duke@0 3150 __ bind(loop);
duke@0 3151
duke@0 3152 // allocate a new frame, filling the registers
duke@0 3153
duke@0 3154 gen_new_frame(masm, deopt); // allocate an interpreter frame
duke@0 3155
duke@0 3156 __ tst(O4array_size);
duke@0 3157 __ br(Assembler::notZero, false, Assembler::pn, loop);
duke@0 3158 __ delayed()->add(O3array, wordSize, O3array);
duke@0 3159 __ ld_ptr(G3pcs, 0, O7); // load final frame new pc
duke@0 3160
duke@0 3161 }
duke@0 3162
duke@0 3163 //------------------------------generate_deopt_blob----------------------------
duke@0 3164 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
duke@0 3165 // instead.
duke@0 3166 void SharedRuntime::generate_deopt_blob() {
duke@0 3167 // allocate space for the code
duke@0 3168 ResourceMark rm;
duke@0 3169 // setup code generation tools
duke@0 3170 int pad = VerifyThread ? 512 : 0;// Extra slop space for more verify code
duke@0 3171 #ifdef _LP64
duke@0 3172 CodeBuffer buffer("deopt_blob", 2100+pad, 512);
duke@0 3173 #else
duke@0 3174 // Measured 8/7/03 at 1212 in 32bit debug build (no VerifyThread)
duke@0 3175 // Measured 8/7/03 at 1396 in 32bit debug build (VerifyThread)
duke@0 3176 CodeBuffer buffer("deopt_blob", 1600+pad, 512);
duke@0 3177 #endif /* _LP64 */
duke@0 3178 MacroAssembler* masm = new MacroAssembler(&buffer);
duke@0 3179 FloatRegister Freturn0 = F0;
duke@0 3180 Register Greturn1 = G1;
duke@0 3181 Register Oreturn0 = O0;
duke@0 3182 Register Oreturn1 = O1;
duke@0 3183 Register O2UnrollBlock = O2;
never@1002 3184 Register L0deopt_mode = L0;
never@1002 3185 Register G4deopt_mode = G4_scratch;
duke@0 3186 int frame_size_words;
twisti@720 3187 Address saved_Freturn0_addr(FP, -sizeof(double) + STACK_BIAS);
duke@0 3188 #if !defined(_LP64) && defined(COMPILER2)
twisti@720 3189 Address saved_Greturn1_addr(FP, -sizeof(double) -sizeof(jlong) + STACK_BIAS);
duke@0 3190 #endif
duke@0 3191 Label cont;
duke@0 3192
duke@0 3193 OopMapSet *oop_maps = new OopMapSet();
duke@0 3194
duke@0 3195 //
duke@0 3196 // This is the entry point for code which is returning to a de-optimized
duke@0 3197 // frame.
duke@0 3198 // The steps taken by this frame are as follows:
duke@0 3199 // - push a dummy "register_save" and save the return values (O0, O1, F0/F1, G1)
duke@0 3200 // and all potentially live registers (at a pollpoint many registers can be live).
duke@0 3201 //
duke@0 3202 // - call the C routine: Deoptimization::fetch_unroll_info (this function
duke@0 3203 // returns information about the number and size of interpreter frames
duke@0 3204 // which are equivalent to the frame which is being deoptimized)
duke@0 3205 // - deallocate the unpack frame, restoring only results values. Other
duke@0 3206 // volatile registers will now be captured in the vframeArray as needed.
duke@0 3207 // - deallocate the deoptimization frame
duke@0 3208 // - in a loop using the information returned in the previous step
duke@0 3209 // push new interpreter frames (take care to propagate the return
duke@0 3210 // values through each new frame pushed)
duke@0 3211 // - create a dummy "unpack_frame" and save the return values (O0, O1, F0)
duke@0 3212 // - call the C routine: Deoptimization::unpack_frames (this function
duke@0 3213 // lays out values on the interpreter frame which was just created)
duke@0 3214 // - deallocate the dummy unpack_frame
duke@0 3215 // - ensure that all the return values are correctly set and then do
duke@0 3216 // a return to the interpreter entry point
duke@0 3217 //
duke@0 3218 // Refer to the following methods for more information:
duke@0 3219 // - Deoptimization::fetch_unroll_info
duke@0 3220 // - Deoptimization::unpack_frames
duke@0 3221
duke@0 3222 OopMap* map = NULL;
duke@0 3223
duke@0 3224 int start = __ offset();
duke@0 3225
duke@0 3226 // restore G2, the trampoline destroyed it
duke@0 3227 __ get_thread();
duke@0 3228
duke@0 3229 // On entry we have been called by the deoptimized nmethod with a call that
duke@0 3230 // replaced the original call (or safepoint polling location) so the deoptimizing
duke@0 3231 // pc is now in O7. Return values are still in the expected places
duke@0 3232
duke@0 3233 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
duke@0 3234 __ ba(false, cont);
never@1002 3235 __ delayed()->mov(Deoptimization::Unpack_deopt, L0deopt_mode);
duke@0 3236
duke@0 3237 int exception_offset = __ offset() - start;
duke@0 3238
duke@0 3239 // restore G2, the trampoline destroyed it
duke@0 3240 __ get_thread();
duke@0 3241
duke@0 3242 // On entry we have been jumped to by the exception handler (or exception_blob
duke@0 3243 // for server). O0 contains the exception oop and O7 contains the original
duke@0 3244 // exception pc. So if we push a frame here it will look to the
duke@0 3245 // stack walking code (fetch_unroll_info) just like a normal call so
duke@0 3246 // state will be extracted normally.
duke@0 3247
duke@0 3248 // save exception oop in JavaThread and fall through into the
duke@0 3249 // exception_in_tls case since they are handled in same way except
duke@0 3250 // for where the pending exception is kept.
twisti@720 3251 __ st_ptr(Oexception, G2_thread, JavaThread::exception_oop_offset());
duke@0 3252
duke@0 3253 //
duke@0 3254 // Vanilla deoptimization with an exception pending in exception_oop
duke@0 3255 //
duke@0 3256 int exception_in_tls_offset = __ offset() - start;
duke@0 3257
duke@0 3258 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
duke@0 3259 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
duke@0 3260
duke@0 3261 // Restore G2_thread
duke@0 3262 __ get_thread();
duke@0 3263
duke@0 3264 #ifdef ASSERT
duke@0 3265 {
duke@0 3266 // verify that there is really an exception oop in exception_oop
duke@0 3267 Label has_exception;
twisti@720 3268 __ ld_ptr(G2_thread, JavaThread::exception_oop_offset(), Oexception);
duke@0 3269 __ br_notnull(Oexception, false, Assembler::pt, has_exception);
duke@0 3270 __ delayed()-> nop();
duke@0 3271 __ stop("no exception in thread");
duke@0 3272 __ bind(has_exception);
duke@0 3273
duke@0 3274 // verify that there is no pending exception
duke@0 3275 Label no_pending_exception;
twisti@720 3276 Address exception_addr(G2_thread, Thread::pending_exception_offset());
duke@0 3277 __ ld_ptr(exception_addr, Oexception);
duke@0 3278 __ br_null(Oexception, false, Assembler::pt, no_pending_exception);
duke@0 3279 __ delayed()->nop();
duke@0 3280 __ stop("must not have pending exception here");
duke@0 3281 __ bind(no_pending_exception);
duke@0 3282 }
duke@0 3283 #endif
duke@0 3284
duke@0 3285 __ ba(false, cont);
never@1002 3286 __ delayed()->mov(Deoptimization::Unpack_exception, L0deopt_mode);;
duke@0 3287
duke@0 3288 //
duke@0 3289 // Reexecute entry, similar to c2 uncommon trap
duke@0 3290 //
duke@0 3291 int reexecute_offset = __ offset() - start;
duke@0 3292
duke@0 3293 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
duke@0 3294 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
duke@0 3295
never@1002 3296 __ mov(Deoptimization::Unpack_reexecute, L0deopt_mode);
duke@0 3297
duke@0 3298 __ bind(cont);
duke@0 3299
duke@0 3300 __ set_last_Java_frame(SP, noreg);
duke@0 3301
duke@0 3302 // do the call by hand so we can get the oopmap
duke@0 3303
duke@0 3304 __ mov(G2_thread, L7_thread_cache);
duke@0 3305 __ call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), relocInfo::runtime_call_type);
duke@0 3306 __ delayed()->mov(G2_thread, O0);
duke@0 3307
duke@0 3308 // Set an oopmap for the call site this describes all our saved volatile registers
duke@0 3309
duke@0 3310 oop_maps->add_gc_map( __ offset()-start, map);
duke@0 3311
duke@0 3312 __ mov(L7_thread_cache, G2_thread);
duke@0 3313
duke@0 3314 __ reset_last_Java_frame();
duke@0 3315
duke@0 3316 // NOTE: we know that only O0/O1 will be reloaded by restore_result_registers
duke@0 3317 // so this move will survive
duke@0 3318
never@1002 3319 __ mov(L0deopt_mode, G4deopt_mode);
duke@0 3320
duke@0 3321 __ mov(O0, O2UnrollBlock->after_save());
duke@0 3322
duke@0 3323 RegisterSaver::restore_result_registers(masm);
duke@0 3324
duke@0 3325 Label noException;
never@1002 3326 __ cmp(G4deopt_mode, Deoptimization::Unpack_exception); // Was exception pending?
duke@0 3327 __ br(Assembler::notEqual, false, Assembler::pt, noException);
duke@0 3328 __ delayed()->nop();
duke@0 3329
duke@0 3330 // Move the pending exception from exception_oop to Oexception so
duke@0 3331 // the pending exception will be picked up the interpreter.
duke@0 3332 __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception);
duke@0 3333 __ st_ptr(G0, G2_thread, in_bytes(JavaThread::exception_oop_offset()));
duke@0 3334 __ bind(noException);
duke@0 3335
duke@0 3336 // deallocate the deoptimization frame taking care to preserve the return values
duke@0 3337 __ mov(Oreturn0, Oreturn0->after_save());
duke@0 3338 __ mov(Oreturn1, Oreturn1->after_save());
duke@0 3339 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
duke@0 3340 __ restore();
duke@0 3341
duke@0 3342 // Allocate new interpreter frame(s) and possible c2i adapter frame
duke@0 3343
duke@0 3344 make_new_frames(masm, true);
duke@0 3345
duke@0 3346 // push a dummy "unpack_frame" taking care of float return values and
duke@0 3347 // call Deoptimization::unpack_frames to have the unpacker layout
duke@0 3348 // information in the interpreter frames just created and then return
duke@0 3349 // to the interpreter entry point
duke@0 3350 __ save(SP, -frame_size_words*wordSize, SP);
duke@0 3351 __ stf(FloatRegisterImpl::D, Freturn0, saved_Freturn0_addr);
duke@0 3352 #if !defined(_LP64)
duke@0 3353 #if defined(COMPILER2)
iveresov@1681 3354 // 32-bit 1-register longs return longs in G1
iveresov@1681 3355 __ stx(Greturn1, saved_Greturn1_addr);
duke@0 3356 #endif
duke@0 3357 __ set_last_Java_frame(SP, noreg);
never@1002 3358 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, G4deopt_mode);
duke@0 3359 #else
duke@0 3360 // LP64 uses g4 in set_last_Java_frame
never@1002 3361 __ mov(G4deopt_mode, O1);
duke@0 3362 __ set_last_Java_frame(SP, G0);
duke@0 3363 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O1);
duke@0 3364 #endif
duke@0 3365 __ reset_last_Java_frame();
duke@0 3366 __ ldf(FloatRegisterImpl::D, saved_Freturn0_addr, Freturn0);
duke@0 3367
duke@0 3368 #if !defined(_LP64) && defined(COMPILER2)
duke@0 3369 // In 32 bit, C2 returns longs in G1 so restore the saved G1 into
iveresov@1681 3370 // I0/I1 if the return value is long.
iveresov@1681 3371 Label not_long;
iveresov@1681 3372 __ cmp(O0,T_LONG);
iveresov@1681 3373 __ br(Assembler::notEqual, false, Assembler::pt, not_long);
iveresov@1681 3374 __ delayed()->nop();
iveresov@1681 3375 __ ldd(saved_Greturn1_addr,I0);
iveresov@1681 3376 __ bind(not_long);
duke@0 3377 #endif
duke@0 3378 __ ret();
duke@0 3379 __ delayed()->restore();
duke@0 3380
duke@0 3381 masm->flush();
duke@0 3382 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_words);
duke@0 3383 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
duke@0 3384 }
duke@0 3385
duke@0 3386 #ifdef COMPILER2
duke@0 3387
duke@0 3388 //------------------------------generate_uncommon_trap_blob--------------------
duke@0 3389 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
duke@0 3390 // instead.
duke@0 3391 void SharedRuntime::generate_uncommon_trap_blob() {
duke@0 3392 // allocate space for the code
duke@0 3393 ResourceMark rm;
duke@0 3394 // setup code generation tools
duke@0 3395 int pad = VerifyThread ? 512 : 0;
duke@0 3396 #ifdef _LP64
duke@0 3397 CodeBuffer buffer("uncommon_trap_blob", 2700+pad, 512);
duke@0 3398 #else
duke@0 3399 // Measured 8/7/03 at 660 in 32bit debug build (no VerifyThread)
duke@0 3400 // Measured 8/7/03 at 1028 in 32bit debug build (VerifyThread)
duke@0 3401 CodeBuffer buffer("uncommon_trap_blob", 2000+pad, 512);
duke@0 3402 #endif
duke@0 3403 MacroAssembler* masm = new MacroAssembler(&buffer);
duke@0 3404 Register O2UnrollBlock = O2;
duke@0 3405 Register O2klass_index = O2;
duke@0 3406
duke@0 3407 //
duke@0 3408 // This is the entry point for all traps the compiler takes when it thinks
duke@0 3409 // it cannot handle further execution of compilation code. The frame is
duke@0 3410 // deoptimized in these cases and converted into interpreter frames for
duke@0 3411 // execution
duke@0 3412 // The steps taken by this frame are as follows:
duke@0 3413 // - push a fake "unpack_frame"
duke@0 3414 // - call the C routine Deoptimization::uncommon_trap (this function
duke@0 3415 // packs the current compiled frame into vframe arrays and returns
duke@0 3416 // information about the number and size of interpreter frames which
duke@0 3417 // are equivalent to the frame which is being deoptimized)
duke@0 3418 // - deallocate the "unpack_frame"
duke@0 3419 // - deallocate the deoptimization frame
duke@0 3420 // - in a loop using the information returned in the previous step
duke@0 3421 // push interpreter frames;
duke@0 3422 // - create a dummy "unpack_frame"
duke@0 3423 // - call the C routine: Deoptimization::unpack_frames (this function
duke@0 3424 // lays out values on the interpreter frame which was just created)
duke@0 3425 // - deallocate the dummy unpack_frame
duke@0 3426 // - return to the interpreter entry point
duke@0 3427 //
duke@0 3428 // Refer to the following methods for more information:
duke@0 3429 // - Deoptimization::uncommon_trap
duke@0 3430 // - Deoptimization::unpack_frame
duke@0 3431
duke@0 3432 // the unloaded class index is in O0 (first parameter to this blob)
duke@0 3433
duke@0 3434 // push a dummy "unpack_frame"
duke@0 3435 // and call Deoptimization::uncommon_trap to pack the compiled frame into
duke@0 3436 // vframe array and return the UnrollBlock information
duke@0 3437 __ save_frame(0);
duke@0 3438 __ set_last_Java_frame(SP, noreg);
duke@0 3439 __ mov(I0, O2klass_index);
duke@0 3440 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), G2_thread, O2klass_index);
duke@0 3441 __ reset_last_Java_frame();
duke@0 3442 __ mov(O0, O2UnrollBlock->after_save());
duke@0 3443 __ restore();
duke@0 3444
duke@0 3445 // deallocate the deoptimized frame taking care to preserve the return values
duke@0 3446 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
duke@0 3447 __ restore();
duke@0 3448
duke@0 3449 // Allocate new interpreter frame(s) and possible c2i adapter frame
duke@0 3450
duke@0 3451 make_new_frames(masm, false);
duke@0 3452
duke@0 3453 // push a dummy "unpack_frame" taking care of float return values and
duke@0 3454 // call Deoptimization::unpack_frames to have the unpacker layout
duke@0 3455 // information in the interpreter frames just created and then return
duke@0 3456 // to the interpreter entry point
duke@0 3457 __ save_frame(0);
duke@0 3458 __ set_last_Java_frame(SP, noreg);
duke@0 3459 __ mov(Deoptimization::Unpack_uncommon_trap, O3); // indicate it is the uncommon trap case
duke@0 3460 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O3);
duke@0 3461 __ reset_last_Java_frame();
duke@0 3462 __ ret();
duke@0 3463 __ delayed()->restore();
duke@0 3464
duke@0 3465 masm->flush();
duke@0 3466 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, NULL, __ total_frame_size_in_bytes(0)/wordSize);
duke@0 3467 }
duke@0 3468
duke@0 3469 #endif // COMPILER2
duke@0 3470
duke@0 3471 //------------------------------generate_handler_blob-------------------
duke@0 3472 //
duke@0 3473 // Generate a special Compile2Runtime blob that saves all registers, and sets
duke@0 3474 // up an OopMap.
duke@0 3475 //
duke@0 3476 // This blob is jumped to (via a breakpoint and the signal handler) from a
duke@0 3477 // safepoint in compiled code. On entry to this blob, O7 contains the
duke@0 3478 // address in the original nmethod at which we should resume normal execution.
duke@0 3479 // Thus, this blob looks like a subroutine which must preserve lots of
duke@0 3480 // registers and return normally. Note that O7 is never register-allocated,
duke@0 3481 // so it is guaranteed to be free here.
duke@0 3482 //
duke@0 3483
duke@0 3484 // The hardest part of what this blob must do is to save the 64-bit %o
duke@0 3485 // registers in the 32-bit build. A simple 'save' turn the %o's to %i's and
duke@0 3486 // an interrupt will chop off their heads. Making space in the caller's frame
duke@0 3487 // first will let us save the 64-bit %o's before save'ing, but we cannot hand
duke@0 3488 // the adjusted FP off to the GC stack-crawler: this will modify the caller's
duke@0 3489 // SP and mess up HIS OopMaps. So we first adjust the caller's SP, then save
duke@0 3490 // the 64-bit %o's, then do a save, then fixup the caller's SP (our FP).
duke@0 3491 // Tricky, tricky, tricky...
duke@0 3492
duke@0 3493 static SafepointBlob* generate_handler_blob(address call_ptr, bool cause_return) {
duke@0 3494 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
duke@0 3495
duke@0 3496 // allocate space for the code
duke@0 3497 ResourceMark rm;
duke@0 3498 // setup code generation tools
duke@0 3499 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
duke@0 3500 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
duke@0 3501 // even larger with TraceJumps
duke@0 3502 int pad = TraceJumps ? 512 : 0;
duke@0 3503 CodeBuffer buffer("handler_blob", 1600 + pad, 512);
duke@0 3504 MacroAssembler* masm = new MacroAssembler(&buffer);