Sun, 13 Apr 2008 17:43:42 -0400
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
Summary: Compressed oops in instances, arrays, and headers. Code contributors are coleenp, phh, never, swamyv
Reviewed-by: jmasa, kamg, acorn, tbell, kvn, rasbold
duke@435 | 1 | /* |
duke@435 | 2 | * Copyright 2003-2007 Sun Microsystems, Inc. All Rights Reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | #include "incls/_precompiled.incl" |
duke@435 | 26 | #include "incls/_sharedRuntime_sparc.cpp.incl" |
duke@435 | 27 | |
duke@435 | 28 | #define __ masm-> |
duke@435 | 29 | |
duke@435 | 30 | #ifdef COMPILER2 |
duke@435 | 31 | UncommonTrapBlob* SharedRuntime::_uncommon_trap_blob; |
duke@435 | 32 | #endif // COMPILER2 |
duke@435 | 33 | |
duke@435 | 34 | DeoptimizationBlob* SharedRuntime::_deopt_blob; |
duke@435 | 35 | SafepointBlob* SharedRuntime::_polling_page_safepoint_handler_blob; |
duke@435 | 36 | SafepointBlob* SharedRuntime::_polling_page_return_handler_blob; |
duke@435 | 37 | RuntimeStub* SharedRuntime::_wrong_method_blob; |
duke@435 | 38 | RuntimeStub* SharedRuntime::_ic_miss_blob; |
duke@435 | 39 | RuntimeStub* SharedRuntime::_resolve_opt_virtual_call_blob; |
duke@435 | 40 | RuntimeStub* SharedRuntime::_resolve_virtual_call_blob; |
duke@435 | 41 | RuntimeStub* SharedRuntime::_resolve_static_call_blob; |
duke@435 | 42 | |
duke@435 | 43 | class RegisterSaver { |
duke@435 | 44 | |
duke@435 | 45 | // Used for saving volatile registers. This is Gregs, Fregs, I/L/O. |
duke@435 | 46 | // The Oregs are problematic. In the 32bit build the compiler can |
duke@435 | 47 | // have O registers live with 64 bit quantities. A window save will |
duke@435 | 48 | // cut the heads off of the registers. We have to do a very extensive |
duke@435 | 49 | // stack dance to save and restore these properly. |
duke@435 | 50 | |
duke@435 | 51 | // Note that the Oregs problem only exists if we block at either a polling |
duke@435 | 52 | // page exception a compiled code safepoint that was not originally a call |
duke@435 | 53 | // or deoptimize following one of these kinds of safepoints. |
duke@435 | 54 | |
duke@435 | 55 | // Lots of registers to save. For all builds, a window save will preserve |
duke@435 | 56 | // the %i and %l registers. For the 32-bit longs-in-two entries and 64-bit |
duke@435 | 57 | // builds a window-save will preserve the %o registers. In the LION build |
duke@435 | 58 | // we need to save the 64-bit %o registers which requires we save them |
duke@435 | 59 | // before the window-save (as then they become %i registers and get their |
duke@435 | 60 | // heads chopped off on interrupt). We have to save some %g registers here |
duke@435 | 61 | // as well. |
duke@435 | 62 | enum { |
duke@435 | 63 | // This frame's save area. Includes extra space for the native call: |
duke@435 | 64 | // vararg's layout space and the like. Briefly holds the caller's |
duke@435 | 65 | // register save area. |
duke@435 | 66 | call_args_area = frame::register_save_words_sp_offset + |
duke@435 | 67 | frame::memory_parameter_word_sp_offset*wordSize, |
duke@435 | 68 | // Make sure save locations are always 8 byte aligned. |
duke@435 | 69 | // can't use round_to because it doesn't produce compile time constant |
duke@435 | 70 | start_of_extra_save_area = ((call_args_area + 7) & ~7), |
duke@435 | 71 | g1_offset = start_of_extra_save_area, // g-regs needing saving |
duke@435 | 72 | g3_offset = g1_offset+8, |
duke@435 | 73 | g4_offset = g3_offset+8, |
duke@435 | 74 | g5_offset = g4_offset+8, |
duke@435 | 75 | o0_offset = g5_offset+8, |
duke@435 | 76 | o1_offset = o0_offset+8, |
duke@435 | 77 | o2_offset = o1_offset+8, |
duke@435 | 78 | o3_offset = o2_offset+8, |
duke@435 | 79 | o4_offset = o3_offset+8, |
duke@435 | 80 | o5_offset = o4_offset+8, |
duke@435 | 81 | start_of_flags_save_area = o5_offset+8, |
duke@435 | 82 | ccr_offset = start_of_flags_save_area, |
duke@435 | 83 | fsr_offset = ccr_offset + 8, |
duke@435 | 84 | d00_offset = fsr_offset+8, // Start of float save area |
duke@435 | 85 | register_save_size = d00_offset+8*32 |
duke@435 | 86 | }; |
duke@435 | 87 | |
duke@435 | 88 | |
duke@435 | 89 | public: |
duke@435 | 90 | |
duke@435 | 91 | static int Oexception_offset() { return o0_offset; }; |
duke@435 | 92 | static int G3_offset() { return g3_offset; }; |
duke@435 | 93 | static int G5_offset() { return g5_offset; }; |
duke@435 | 94 | static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words); |
duke@435 | 95 | static void restore_live_registers(MacroAssembler* masm); |
duke@435 | 96 | |
duke@435 | 97 | // During deoptimization only the result register need to be restored |
duke@435 | 98 | // all the other values have already been extracted. |
duke@435 | 99 | |
duke@435 | 100 | static void restore_result_registers(MacroAssembler* masm); |
duke@435 | 101 | }; |
duke@435 | 102 | |
duke@435 | 103 | OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) { |
duke@435 | 104 | // Record volatile registers as callee-save values in an OopMap so their save locations will be |
duke@435 | 105 | // propagated to the caller frame's RegisterMap during StackFrameStream construction (needed for |
duke@435 | 106 | // deoptimization; see compiledVFrame::create_stack_value). The caller's I, L and O registers |
duke@435 | 107 | // are saved in register windows - I's and L's in the caller's frame and O's in the stub frame |
duke@435 | 108 | // (as the stub's I's) when the runtime routine called by the stub creates its frame. |
duke@435 | 109 | int i; |
duke@435 | 110 | // Always make the frame size 16 bytr aligned. |
duke@435 | 111 | int frame_size = round_to(additional_frame_words + register_save_size, 16); |
duke@435 | 112 | // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words |
duke@435 | 113 | int frame_size_in_slots = frame_size / sizeof(jint); |
duke@435 | 114 | // CodeBlob frame size is in words. |
duke@435 | 115 | *total_frame_words = frame_size / wordSize; |
duke@435 | 116 | // OopMap* map = new OopMap(*total_frame_words, 0); |
duke@435 | 117 | OopMap* map = new OopMap(frame_size_in_slots, 0); |
duke@435 | 118 | |
duke@435 | 119 | #if !defined(_LP64) |
duke@435 | 120 | |
duke@435 | 121 | // Save 64-bit O registers; they will get their heads chopped off on a 'save'. |
duke@435 | 122 | __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8); |
duke@435 | 123 | __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8); |
duke@435 | 124 | __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8); |
duke@435 | 125 | __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8); |
duke@435 | 126 | __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8); |
duke@435 | 127 | __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8); |
duke@435 | 128 | #endif /* _LP64 */ |
duke@435 | 129 | |
duke@435 | 130 | __ save(SP, -frame_size, SP); |
duke@435 | 131 | |
duke@435 | 132 | #ifndef _LP64 |
duke@435 | 133 | // Reload the 64 bit Oregs. Although they are now Iregs we load them |
duke@435 | 134 | // to Oregs here to avoid interrupts cutting off their heads |
duke@435 | 135 | |
duke@435 | 136 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0); |
duke@435 | 137 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1); |
duke@435 | 138 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2); |
duke@435 | 139 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3); |
duke@435 | 140 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4); |
duke@435 | 141 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5); |
duke@435 | 142 | |
duke@435 | 143 | __ stx(O0, SP, o0_offset+STACK_BIAS); |
duke@435 | 144 | map->set_callee_saved(VMRegImpl::stack2reg((o0_offset + 4)>>2), O0->as_VMReg()); |
duke@435 | 145 | |
duke@435 | 146 | __ stx(O1, SP, o1_offset+STACK_BIAS); |
duke@435 | 147 | |
duke@435 | 148 | map->set_callee_saved(VMRegImpl::stack2reg((o1_offset + 4)>>2), O1->as_VMReg()); |
duke@435 | 149 | |
duke@435 | 150 | __ stx(O2, SP, o2_offset+STACK_BIAS); |
duke@435 | 151 | map->set_callee_saved(VMRegImpl::stack2reg((o2_offset + 4)>>2), O2->as_VMReg()); |
duke@435 | 152 | |
duke@435 | 153 | __ stx(O3, SP, o3_offset+STACK_BIAS); |
duke@435 | 154 | map->set_callee_saved(VMRegImpl::stack2reg((o3_offset + 4)>>2), O3->as_VMReg()); |
duke@435 | 155 | |
duke@435 | 156 | __ stx(O4, SP, o4_offset+STACK_BIAS); |
duke@435 | 157 | map->set_callee_saved(VMRegImpl::stack2reg((o4_offset + 4)>>2), O4->as_VMReg()); |
duke@435 | 158 | |
duke@435 | 159 | __ stx(O5, SP, o5_offset+STACK_BIAS); |
duke@435 | 160 | map->set_callee_saved(VMRegImpl::stack2reg((o5_offset + 4)>>2), O5->as_VMReg()); |
duke@435 | 161 | #endif /* _LP64 */ |
duke@435 | 162 | |
coleenp@548 | 163 | |
coleenp@548 | 164 | #ifdef _LP64 |
coleenp@548 | 165 | int debug_offset = 0; |
coleenp@548 | 166 | #else |
coleenp@548 | 167 | int debug_offset = 4; |
coleenp@548 | 168 | #endif |
duke@435 | 169 | // Save the G's |
duke@435 | 170 | __ stx(G1, SP, g1_offset+STACK_BIAS); |
coleenp@548 | 171 | map->set_callee_saved(VMRegImpl::stack2reg((g1_offset + debug_offset)>>2), G1->as_VMReg()); |
duke@435 | 172 | |
duke@435 | 173 | __ stx(G3, SP, g3_offset+STACK_BIAS); |
coleenp@548 | 174 | map->set_callee_saved(VMRegImpl::stack2reg((g3_offset + debug_offset)>>2), G3->as_VMReg()); |
duke@435 | 175 | |
duke@435 | 176 | __ stx(G4, SP, g4_offset+STACK_BIAS); |
coleenp@548 | 177 | map->set_callee_saved(VMRegImpl::stack2reg((g4_offset + debug_offset)>>2), G4->as_VMReg()); |
duke@435 | 178 | |
duke@435 | 179 | __ stx(G5, SP, g5_offset+STACK_BIAS); |
coleenp@548 | 180 | map->set_callee_saved(VMRegImpl::stack2reg((g5_offset + debug_offset)>>2), G5->as_VMReg()); |
duke@435 | 181 | |
duke@435 | 182 | // This is really a waste but we'll keep things as they were for now |
duke@435 | 183 | if (true) { |
duke@435 | 184 | #ifndef _LP64 |
duke@435 | 185 | map->set_callee_saved(VMRegImpl::stack2reg((o0_offset)>>2), O0->as_VMReg()->next()); |
duke@435 | 186 | map->set_callee_saved(VMRegImpl::stack2reg((o1_offset)>>2), O1->as_VMReg()->next()); |
duke@435 | 187 | map->set_callee_saved(VMRegImpl::stack2reg((o2_offset)>>2), O2->as_VMReg()->next()); |
duke@435 | 188 | map->set_callee_saved(VMRegImpl::stack2reg((o3_offset)>>2), O3->as_VMReg()->next()); |
duke@435 | 189 | map->set_callee_saved(VMRegImpl::stack2reg((o4_offset)>>2), O4->as_VMReg()->next()); |
duke@435 | 190 | map->set_callee_saved(VMRegImpl::stack2reg((o5_offset)>>2), O5->as_VMReg()->next()); |
duke@435 | 191 | map->set_callee_saved(VMRegImpl::stack2reg((g1_offset)>>2), G1->as_VMReg()->next()); |
duke@435 | 192 | map->set_callee_saved(VMRegImpl::stack2reg((g3_offset)>>2), G3->as_VMReg()->next()); |
duke@435 | 193 | map->set_callee_saved(VMRegImpl::stack2reg((g4_offset)>>2), G4->as_VMReg()->next()); |
duke@435 | 194 | map->set_callee_saved(VMRegImpl::stack2reg((g5_offset)>>2), G5->as_VMReg()->next()); |
coleenp@548 | 195 | #endif /* _LP64 */ |
duke@435 | 196 | } |
duke@435 | 197 | |
duke@435 | 198 | |
duke@435 | 199 | // Save the flags |
duke@435 | 200 | __ rdccr( G5 ); |
duke@435 | 201 | __ stx(G5, SP, ccr_offset+STACK_BIAS); |
duke@435 | 202 | __ stxfsr(SP, fsr_offset+STACK_BIAS); |
duke@435 | 203 | |
duke@435 | 204 | // Save all the FP registers |
duke@435 | 205 | int offset = d00_offset; |
duke@435 | 206 | for( int i=0; i<64; i+=2 ) { |
duke@435 | 207 | FloatRegister f = as_FloatRegister(i); |
duke@435 | 208 | __ stf(FloatRegisterImpl::D, f, SP, offset+STACK_BIAS); |
duke@435 | 209 | map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), f->as_VMReg()); |
duke@435 | 210 | if (true) { |
duke@435 | 211 | map->set_callee_saved(VMRegImpl::stack2reg((offset + sizeof(float))>>2), f->as_VMReg()->next()); |
duke@435 | 212 | } |
duke@435 | 213 | offset += sizeof(double); |
duke@435 | 214 | } |
duke@435 | 215 | |
duke@435 | 216 | // And we're done. |
duke@435 | 217 | |
duke@435 | 218 | return map; |
duke@435 | 219 | } |
duke@435 | 220 | |
duke@435 | 221 | |
duke@435 | 222 | // Pop the current frame and restore all the registers that we |
duke@435 | 223 | // saved. |
duke@435 | 224 | void RegisterSaver::restore_live_registers(MacroAssembler* masm) { |
duke@435 | 225 | |
duke@435 | 226 | // Restore all the FP registers |
duke@435 | 227 | for( int i=0; i<64; i+=2 ) { |
duke@435 | 228 | __ ldf(FloatRegisterImpl::D, SP, d00_offset+i*sizeof(float)+STACK_BIAS, as_FloatRegister(i)); |
duke@435 | 229 | } |
duke@435 | 230 | |
duke@435 | 231 | __ ldx(SP, ccr_offset+STACK_BIAS, G1); |
duke@435 | 232 | __ wrccr (G1) ; |
duke@435 | 233 | |
duke@435 | 234 | // Restore the G's |
duke@435 | 235 | // Note that G2 (AKA GThread) must be saved and restored separately. |
duke@435 | 236 | // TODO-FIXME: save and restore some of the other ASRs, viz., %asi and %gsr. |
duke@435 | 237 | |
duke@435 | 238 | __ ldx(SP, g1_offset+STACK_BIAS, G1); |
duke@435 | 239 | __ ldx(SP, g3_offset+STACK_BIAS, G3); |
duke@435 | 240 | __ ldx(SP, g4_offset+STACK_BIAS, G4); |
duke@435 | 241 | __ ldx(SP, g5_offset+STACK_BIAS, G5); |
duke@435 | 242 | |
duke@435 | 243 | |
duke@435 | 244 | #if !defined(_LP64) |
duke@435 | 245 | // Restore the 64-bit O's. |
duke@435 | 246 | __ ldx(SP, o0_offset+STACK_BIAS, O0); |
duke@435 | 247 | __ ldx(SP, o1_offset+STACK_BIAS, O1); |
duke@435 | 248 | __ ldx(SP, o2_offset+STACK_BIAS, O2); |
duke@435 | 249 | __ ldx(SP, o3_offset+STACK_BIAS, O3); |
duke@435 | 250 | __ ldx(SP, o4_offset+STACK_BIAS, O4); |
duke@435 | 251 | __ ldx(SP, o5_offset+STACK_BIAS, O5); |
duke@435 | 252 | |
duke@435 | 253 | // And temporarily place them in TLS |
duke@435 | 254 | |
duke@435 | 255 | __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8); |
duke@435 | 256 | __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8); |
duke@435 | 257 | __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8); |
duke@435 | 258 | __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8); |
duke@435 | 259 | __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8); |
duke@435 | 260 | __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8); |
duke@435 | 261 | #endif /* _LP64 */ |
duke@435 | 262 | |
duke@435 | 263 | // Restore flags |
duke@435 | 264 | |
duke@435 | 265 | __ ldxfsr(SP, fsr_offset+STACK_BIAS); |
duke@435 | 266 | |
duke@435 | 267 | __ restore(); |
duke@435 | 268 | |
duke@435 | 269 | #if !defined(_LP64) |
duke@435 | 270 | // Now reload the 64bit Oregs after we've restore the window. |
duke@435 | 271 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0); |
duke@435 | 272 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1); |
duke@435 | 273 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2); |
duke@435 | 274 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3); |
duke@435 | 275 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4); |
duke@435 | 276 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5); |
duke@435 | 277 | #endif /* _LP64 */ |
duke@435 | 278 | |
duke@435 | 279 | } |
duke@435 | 280 | |
duke@435 | 281 | // Pop the current frame and restore the registers that might be holding |
duke@435 | 282 | // a result. |
duke@435 | 283 | void RegisterSaver::restore_result_registers(MacroAssembler* masm) { |
duke@435 | 284 | |
duke@435 | 285 | #if !defined(_LP64) |
duke@435 | 286 | // 32bit build returns longs in G1 |
duke@435 | 287 | __ ldx(SP, g1_offset+STACK_BIAS, G1); |
duke@435 | 288 | |
duke@435 | 289 | // Retrieve the 64-bit O's. |
duke@435 | 290 | __ ldx(SP, o0_offset+STACK_BIAS, O0); |
duke@435 | 291 | __ ldx(SP, o1_offset+STACK_BIAS, O1); |
duke@435 | 292 | // and save to TLS |
duke@435 | 293 | __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8); |
duke@435 | 294 | __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8); |
duke@435 | 295 | #endif /* _LP64 */ |
duke@435 | 296 | |
duke@435 | 297 | __ ldf(FloatRegisterImpl::D, SP, d00_offset+STACK_BIAS, as_FloatRegister(0)); |
duke@435 | 298 | |
duke@435 | 299 | __ restore(); |
duke@435 | 300 | |
duke@435 | 301 | #if !defined(_LP64) |
duke@435 | 302 | // Now reload the 64bit Oregs after we've restore the window. |
duke@435 | 303 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0); |
duke@435 | 304 | __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1); |
duke@435 | 305 | #endif /* _LP64 */ |
duke@435 | 306 | |
duke@435 | 307 | } |
duke@435 | 308 | |
duke@435 | 309 | // The java_calling_convention describes stack locations as ideal slots on |
duke@435 | 310 | // a frame with no abi restrictions. Since we must observe abi restrictions |
duke@435 | 311 | // (like the placement of the register window) the slots must be biased by |
duke@435 | 312 | // the following value. |
duke@435 | 313 | static int reg2offset(VMReg r) { |
duke@435 | 314 | return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size; |
duke@435 | 315 | } |
duke@435 | 316 | |
duke@435 | 317 | // --------------------------------------------------------------------------- |
duke@435 | 318 | // Read the array of BasicTypes from a signature, and compute where the |
duke@435 | 319 | // arguments should go. Values in the VMRegPair regs array refer to 4-byte (VMRegImpl::stack_slot_size) |
duke@435 | 320 | // quantities. Values less than VMRegImpl::stack0 are registers, those above |
duke@435 | 321 | // refer to 4-byte stack slots. All stack slots are based off of the window |
duke@435 | 322 | // top. VMRegImpl::stack0 refers to the first slot past the 16-word window, |
duke@435 | 323 | // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register |
duke@435 | 324 | // values 0-63 (up to RegisterImpl::number_of_registers) are the 64-bit |
duke@435 | 325 | // integer registers. Values 64-95 are the (32-bit only) float registers. |
duke@435 | 326 | // Each 32-bit quantity is given its own number, so the integer registers |
duke@435 | 327 | // (in either 32- or 64-bit builds) use 2 numbers. For example, there is |
duke@435 | 328 | // an O0-low and an O0-high. Essentially, all int register numbers are doubled. |
duke@435 | 329 | |
duke@435 | 330 | // Register results are passed in O0-O5, for outgoing call arguments. To |
duke@435 | 331 | // convert to incoming arguments, convert all O's to I's. The regs array |
duke@435 | 332 | // refer to the low and hi 32-bit words of 64-bit registers or stack slots. |
duke@435 | 333 | // If the regs[].second() field is set to VMRegImpl::Bad(), it means it's unused (a |
duke@435 | 334 | // 32-bit value was passed). If both are VMRegImpl::Bad(), it means no value was |
duke@435 | 335 | // passed (used as a placeholder for the other half of longs and doubles in |
duke@435 | 336 | // the 64-bit build). regs[].second() is either VMRegImpl::Bad() or regs[].second() is |
duke@435 | 337 | // regs[].first()+1 (regs[].first() may be misaligned in the C calling convention). |
duke@435 | 338 | // Sparc never passes a value in regs[].second() but not regs[].first() (regs[].first() |
duke@435 | 339 | // == VMRegImpl::Bad() && regs[].second() != VMRegImpl::Bad()) nor unrelated values in the |
duke@435 | 340 | // same VMRegPair. |
duke@435 | 341 | |
duke@435 | 342 | // Note: the INPUTS in sig_bt are in units of Java argument words, which are |
duke@435 | 343 | // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit |
duke@435 | 344 | // units regardless of build. |
duke@435 | 345 | |
duke@435 | 346 | |
duke@435 | 347 | // --------------------------------------------------------------------------- |
duke@435 | 348 | // The compiled Java calling convention. The Java convention always passes |
duke@435 | 349 | // 64-bit values in adjacent aligned locations (either registers or stack), |
duke@435 | 350 | // floats in float registers and doubles in aligned float pairs. Values are |
duke@435 | 351 | // packed in the registers. There is no backing varargs store for values in |
duke@435 | 352 | // registers. In the 32-bit build, longs are passed in G1 and G4 (cannot be |
duke@435 | 353 | // passed in I's, because longs in I's get their heads chopped off at |
duke@435 | 354 | // interrupt). |
duke@435 | 355 | int SharedRuntime::java_calling_convention(const BasicType *sig_bt, |
duke@435 | 356 | VMRegPair *regs, |
duke@435 | 357 | int total_args_passed, |
duke@435 | 358 | int is_outgoing) { |
duke@435 | 359 | assert(F31->as_VMReg()->is_reg(), "overlapping stack/register numbers"); |
duke@435 | 360 | |
duke@435 | 361 | // Convention is to pack the first 6 int/oop args into the first 6 registers |
duke@435 | 362 | // (I0-I5), extras spill to the stack. Then pack the first 8 float args |
duke@435 | 363 | // into F0-F7, extras spill to the stack. Then pad all register sets to |
duke@435 | 364 | // align. Then put longs and doubles into the same registers as they fit, |
duke@435 | 365 | // else spill to the stack. |
duke@435 | 366 | const int int_reg_max = SPARC_ARGS_IN_REGS_NUM; |
duke@435 | 367 | const int flt_reg_max = 8; |
duke@435 | 368 | // |
duke@435 | 369 | // Where 32-bit 1-reg longs start being passed |
duke@435 | 370 | // In tiered we must pass on stack because c1 can't use a "pair" in a single reg. |
duke@435 | 371 | // So make it look like we've filled all the G regs that c2 wants to use. |
duke@435 | 372 | Register g_reg = TieredCompilation ? noreg : G1; |
duke@435 | 373 | |
duke@435 | 374 | // Count int/oop and float args. See how many stack slots we'll need and |
duke@435 | 375 | // where the longs & doubles will go. |
duke@435 | 376 | int int_reg_cnt = 0; |
duke@435 | 377 | int flt_reg_cnt = 0; |
duke@435 | 378 | // int stk_reg_pairs = frame::register_save_words*(wordSize>>2); |
duke@435 | 379 | // int stk_reg_pairs = SharedRuntime::out_preserve_stack_slots(); |
duke@435 | 380 | int stk_reg_pairs = 0; |
duke@435 | 381 | for (int i = 0; i < total_args_passed; i++) { |
duke@435 | 382 | switch (sig_bt[i]) { |
duke@435 | 383 | case T_LONG: // LP64, longs compete with int args |
duke@435 | 384 | assert(sig_bt[i+1] == T_VOID, ""); |
duke@435 | 385 | #ifdef _LP64 |
duke@435 | 386 | if (int_reg_cnt < int_reg_max) int_reg_cnt++; |
duke@435 | 387 | #endif |
duke@435 | 388 | break; |
duke@435 | 389 | case T_OBJECT: |
duke@435 | 390 | case T_ARRAY: |
duke@435 | 391 | case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address |
duke@435 | 392 | if (int_reg_cnt < int_reg_max) int_reg_cnt++; |
duke@435 | 393 | #ifndef _LP64 |
duke@435 | 394 | else stk_reg_pairs++; |
duke@435 | 395 | #endif |
duke@435 | 396 | break; |
duke@435 | 397 | case T_INT: |
duke@435 | 398 | case T_SHORT: |
duke@435 | 399 | case T_CHAR: |
duke@435 | 400 | case T_BYTE: |
duke@435 | 401 | case T_BOOLEAN: |
duke@435 | 402 | if (int_reg_cnt < int_reg_max) int_reg_cnt++; |
duke@435 | 403 | else stk_reg_pairs++; |
duke@435 | 404 | break; |
duke@435 | 405 | case T_FLOAT: |
duke@435 | 406 | if (flt_reg_cnt < flt_reg_max) flt_reg_cnt++; |
duke@435 | 407 | else stk_reg_pairs++; |
duke@435 | 408 | break; |
duke@435 | 409 | case T_DOUBLE: |
duke@435 | 410 | assert(sig_bt[i+1] == T_VOID, ""); |
duke@435 | 411 | break; |
duke@435 | 412 | case T_VOID: |
duke@435 | 413 | break; |
duke@435 | 414 | default: |
duke@435 | 415 | ShouldNotReachHere(); |
duke@435 | 416 | } |
duke@435 | 417 | } |
duke@435 | 418 | |
duke@435 | 419 | // This is where the longs/doubles start on the stack. |
duke@435 | 420 | stk_reg_pairs = (stk_reg_pairs+1) & ~1; // Round |
duke@435 | 421 | |
duke@435 | 422 | int int_reg_pairs = (int_reg_cnt+1) & ~1; // 32-bit 2-reg longs only |
duke@435 | 423 | int flt_reg_pairs = (flt_reg_cnt+1) & ~1; |
duke@435 | 424 | |
duke@435 | 425 | // int stk_reg = frame::register_save_words*(wordSize>>2); |
duke@435 | 426 | // int stk_reg = SharedRuntime::out_preserve_stack_slots(); |
duke@435 | 427 | int stk_reg = 0; |
duke@435 | 428 | int int_reg = 0; |
duke@435 | 429 | int flt_reg = 0; |
duke@435 | 430 | |
duke@435 | 431 | // Now do the signature layout |
duke@435 | 432 | for (int i = 0; i < total_args_passed; i++) { |
duke@435 | 433 | switch (sig_bt[i]) { |
duke@435 | 434 | case T_INT: |
duke@435 | 435 | case T_SHORT: |
duke@435 | 436 | case T_CHAR: |
duke@435 | 437 | case T_BYTE: |
duke@435 | 438 | case T_BOOLEAN: |
duke@435 | 439 | #ifndef _LP64 |
duke@435 | 440 | case T_OBJECT: |
duke@435 | 441 | case T_ARRAY: |
duke@435 | 442 | case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address |
duke@435 | 443 | #endif // _LP64 |
duke@435 | 444 | if (int_reg < int_reg_max) { |
duke@435 | 445 | Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++); |
duke@435 | 446 | regs[i].set1(r->as_VMReg()); |
duke@435 | 447 | } else { |
duke@435 | 448 | regs[i].set1(VMRegImpl::stack2reg(stk_reg++)); |
duke@435 | 449 | } |
duke@435 | 450 | break; |
duke@435 | 451 | |
duke@435 | 452 | #ifdef _LP64 |
duke@435 | 453 | case T_OBJECT: |
duke@435 | 454 | case T_ARRAY: |
duke@435 | 455 | case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address |
duke@435 | 456 | if (int_reg < int_reg_max) { |
duke@435 | 457 | Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++); |
duke@435 | 458 | regs[i].set2(r->as_VMReg()); |
duke@435 | 459 | } else { |
duke@435 | 460 | regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); |
duke@435 | 461 | stk_reg_pairs += 2; |
duke@435 | 462 | } |
duke@435 | 463 | break; |
duke@435 | 464 | #endif // _LP64 |
duke@435 | 465 | |
duke@435 | 466 | case T_LONG: |
duke@435 | 467 | assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half"); |
duke@435 | 468 | #ifdef COMPILER2 |
duke@435 | 469 | #ifdef _LP64 |
duke@435 | 470 | // Can't be tiered (yet) |
duke@435 | 471 | if (int_reg < int_reg_max) { |
duke@435 | 472 | Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++); |
duke@435 | 473 | regs[i].set2(r->as_VMReg()); |
duke@435 | 474 | } else { |
duke@435 | 475 | regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); |
duke@435 | 476 | stk_reg_pairs += 2; |
duke@435 | 477 | } |
duke@435 | 478 | #else |
duke@435 | 479 | // For 32-bit build, can't pass longs in O-regs because they become |
duke@435 | 480 | // I-regs and get trashed. Use G-regs instead. G1 and G4 are almost |
duke@435 | 481 | // spare and available. This convention isn't used by the Sparc ABI or |
duke@435 | 482 | // anywhere else. If we're tiered then we don't use G-regs because c1 |
duke@435 | 483 | // can't deal with them as a "pair". |
duke@435 | 484 | // G0: zero |
duke@435 | 485 | // G1: 1st Long arg |
duke@435 | 486 | // G2: global allocated to TLS |
duke@435 | 487 | // G3: used in inline cache check |
duke@435 | 488 | // G4: 2nd Long arg |
duke@435 | 489 | // G5: used in inline cache check |
duke@435 | 490 | // G6: used by OS |
duke@435 | 491 | // G7: used by OS |
duke@435 | 492 | |
duke@435 | 493 | if (g_reg == G1) { |
duke@435 | 494 | regs[i].set2(G1->as_VMReg()); // This long arg in G1 |
duke@435 | 495 | g_reg = G4; // Where the next arg goes |
duke@435 | 496 | } else if (g_reg == G4) { |
duke@435 | 497 | regs[i].set2(G4->as_VMReg()); // The 2nd long arg in G4 |
duke@435 | 498 | g_reg = noreg; // No more longs in registers |
duke@435 | 499 | } else { |
duke@435 | 500 | regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); |
duke@435 | 501 | stk_reg_pairs += 2; |
duke@435 | 502 | } |
duke@435 | 503 | #endif // _LP64 |
duke@435 | 504 | #else // COMPILER2 |
duke@435 | 505 | if (int_reg_pairs + 1 < int_reg_max) { |
duke@435 | 506 | if (is_outgoing) { |
duke@435 | 507 | regs[i].set_pair(as_oRegister(int_reg_pairs + 1)->as_VMReg(), as_oRegister(int_reg_pairs)->as_VMReg()); |
duke@435 | 508 | } else { |
duke@435 | 509 | regs[i].set_pair(as_iRegister(int_reg_pairs + 1)->as_VMReg(), as_iRegister(int_reg_pairs)->as_VMReg()); |
duke@435 | 510 | } |
duke@435 | 511 | int_reg_pairs += 2; |
duke@435 | 512 | } else { |
duke@435 | 513 | regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); |
duke@435 | 514 | stk_reg_pairs += 2; |
duke@435 | 515 | } |
duke@435 | 516 | #endif // COMPILER2 |
duke@435 | 517 | break; |
duke@435 | 518 | |
duke@435 | 519 | case T_FLOAT: |
duke@435 | 520 | if (flt_reg < flt_reg_max) regs[i].set1(as_FloatRegister(flt_reg++)->as_VMReg()); |
duke@435 | 521 | else regs[i].set1( VMRegImpl::stack2reg(stk_reg++)); |
duke@435 | 522 | break; |
duke@435 | 523 | case T_DOUBLE: |
duke@435 | 524 | assert(sig_bt[i+1] == T_VOID, "expecting half"); |
duke@435 | 525 | if (flt_reg_pairs + 1 < flt_reg_max) { |
duke@435 | 526 | regs[i].set2(as_FloatRegister(flt_reg_pairs)->as_VMReg()); |
duke@435 | 527 | flt_reg_pairs += 2; |
duke@435 | 528 | } else { |
duke@435 | 529 | regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs)); |
duke@435 | 530 | stk_reg_pairs += 2; |
duke@435 | 531 | } |
duke@435 | 532 | break; |
duke@435 | 533 | case T_VOID: regs[i].set_bad(); break; // Halves of longs & doubles |
duke@435 | 534 | default: |
duke@435 | 535 | ShouldNotReachHere(); |
duke@435 | 536 | } |
duke@435 | 537 | } |
duke@435 | 538 | |
duke@435 | 539 | // retun the amount of stack space these arguments will need. |
duke@435 | 540 | return stk_reg_pairs; |
duke@435 | 541 | |
duke@435 | 542 | } |
duke@435 | 543 | |
duke@435 | 544 | // Helper class mostly to avoid passing masm everywhere, and handle store |
duke@435 | 545 | // displacement overflow logic for LP64 |
duke@435 | 546 | class AdapterGenerator { |
duke@435 | 547 | MacroAssembler *masm; |
duke@435 | 548 | #ifdef _LP64 |
duke@435 | 549 | Register Rdisp; |
duke@435 | 550 | void set_Rdisp(Register r) { Rdisp = r; } |
duke@435 | 551 | #endif // _LP64 |
duke@435 | 552 | |
duke@435 | 553 | void patch_callers_callsite(); |
duke@435 | 554 | void tag_c2i_arg(frame::Tag t, Register base, int st_off, Register scratch); |
duke@435 | 555 | |
duke@435 | 556 | // base+st_off points to top of argument |
duke@435 | 557 | int arg_offset(const int st_off) { return st_off + Interpreter::value_offset_in_bytes(); } |
duke@435 | 558 | int next_arg_offset(const int st_off) { |
duke@435 | 559 | return st_off - Interpreter::stackElementSize() + Interpreter::value_offset_in_bytes(); |
duke@435 | 560 | } |
duke@435 | 561 | |
duke@435 | 562 | #ifdef _LP64 |
duke@435 | 563 | // On _LP64 argument slot values are loaded first into a register |
duke@435 | 564 | // because they might not fit into displacement. |
duke@435 | 565 | Register arg_slot(const int st_off); |
duke@435 | 566 | Register next_arg_slot(const int st_off); |
duke@435 | 567 | #else |
duke@435 | 568 | int arg_slot(const int st_off) { return arg_offset(st_off); } |
duke@435 | 569 | int next_arg_slot(const int st_off) { return next_arg_offset(st_off); } |
duke@435 | 570 | #endif // _LP64 |
duke@435 | 571 | |
duke@435 | 572 | // Stores long into offset pointed to by base |
duke@435 | 573 | void store_c2i_long(Register r, Register base, |
duke@435 | 574 | const int st_off, bool is_stack); |
duke@435 | 575 | void store_c2i_object(Register r, Register base, |
duke@435 | 576 | const int st_off); |
duke@435 | 577 | void store_c2i_int(Register r, Register base, |
duke@435 | 578 | const int st_off); |
duke@435 | 579 | void store_c2i_double(VMReg r_2, |
duke@435 | 580 | VMReg r_1, Register base, const int st_off); |
duke@435 | 581 | void store_c2i_float(FloatRegister f, Register base, |
duke@435 | 582 | const int st_off); |
duke@435 | 583 | |
duke@435 | 584 | public: |
duke@435 | 585 | void gen_c2i_adapter(int total_args_passed, |
duke@435 | 586 | // VMReg max_arg, |
duke@435 | 587 | int comp_args_on_stack, // VMRegStackSlots |
duke@435 | 588 | const BasicType *sig_bt, |
duke@435 | 589 | const VMRegPair *regs, |
duke@435 | 590 | Label& skip_fixup); |
duke@435 | 591 | void gen_i2c_adapter(int total_args_passed, |
duke@435 | 592 | // VMReg max_arg, |
duke@435 | 593 | int comp_args_on_stack, // VMRegStackSlots |
duke@435 | 594 | const BasicType *sig_bt, |
duke@435 | 595 | const VMRegPair *regs); |
duke@435 | 596 | |
duke@435 | 597 | AdapterGenerator(MacroAssembler *_masm) : masm(_masm) {} |
duke@435 | 598 | }; |
duke@435 | 599 | |
duke@435 | 600 | |
duke@435 | 601 | // Patch the callers callsite with entry to compiled code if it exists. |
duke@435 | 602 | void AdapterGenerator::patch_callers_callsite() { |
duke@435 | 603 | Label L; |
duke@435 | 604 | __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch); |
duke@435 | 605 | __ br_null(G3_scratch, false, __ pt, L); |
duke@435 | 606 | // Schedule the branch target address early. |
duke@435 | 607 | __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch); |
duke@435 | 608 | // Call into the VM to patch the caller, then jump to compiled callee |
duke@435 | 609 | __ save_frame(4); // Args in compiled layout; do not blow them |
duke@435 | 610 | |
duke@435 | 611 | // Must save all the live Gregs the list is: |
duke@435 | 612 | // G1: 1st Long arg (32bit build) |
duke@435 | 613 | // G2: global allocated to TLS |
duke@435 | 614 | // G3: used in inline cache check (scratch) |
duke@435 | 615 | // G4: 2nd Long arg (32bit build); |
duke@435 | 616 | // G5: used in inline cache check (methodOop) |
duke@435 | 617 | |
duke@435 | 618 | // The longs must go to the stack by hand since in the 32 bit build they can be trashed by window ops. |
duke@435 | 619 | |
duke@435 | 620 | #ifdef _LP64 |
duke@435 | 621 | // mov(s,d) |
duke@435 | 622 | __ mov(G1, L1); |
duke@435 | 623 | __ mov(G4, L4); |
duke@435 | 624 | __ mov(G5_method, L5); |
duke@435 | 625 | __ mov(G5_method, O0); // VM needs target method |
duke@435 | 626 | __ mov(I7, O1); // VM needs caller's callsite |
duke@435 | 627 | // Must be a leaf call... |
duke@435 | 628 | // can be very far once the blob has been relocated |
duke@435 | 629 | Address dest(O7, CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)); |
duke@435 | 630 | __ relocate(relocInfo::runtime_call_type); |
duke@435 | 631 | __ jumpl_to(dest, O7); |
duke@435 | 632 | __ delayed()->mov(G2_thread, L7_thread_cache); |
duke@435 | 633 | __ mov(L7_thread_cache, G2_thread); |
duke@435 | 634 | __ mov(L1, G1); |
duke@435 | 635 | __ mov(L4, G4); |
duke@435 | 636 | __ mov(L5, G5_method); |
duke@435 | 637 | #else |
duke@435 | 638 | __ stx(G1, FP, -8 + STACK_BIAS); |
duke@435 | 639 | __ stx(G4, FP, -16 + STACK_BIAS); |
duke@435 | 640 | __ mov(G5_method, L5); |
duke@435 | 641 | __ mov(G5_method, O0); // VM needs target method |
duke@435 | 642 | __ mov(I7, O1); // VM needs caller's callsite |
duke@435 | 643 | // Must be a leaf call... |
duke@435 | 644 | __ call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), relocInfo::runtime_call_type); |
duke@435 | 645 | __ delayed()->mov(G2_thread, L7_thread_cache); |
duke@435 | 646 | __ mov(L7_thread_cache, G2_thread); |
duke@435 | 647 | __ ldx(FP, -8 + STACK_BIAS, G1); |
duke@435 | 648 | __ ldx(FP, -16 + STACK_BIAS, G4); |
duke@435 | 649 | __ mov(L5, G5_method); |
duke@435 | 650 | __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch); |
duke@435 | 651 | #endif /* _LP64 */ |
duke@435 | 652 | |
duke@435 | 653 | __ restore(); // Restore args |
duke@435 | 654 | __ bind(L); |
duke@435 | 655 | } |
duke@435 | 656 | |
duke@435 | 657 | void AdapterGenerator::tag_c2i_arg(frame::Tag t, Register base, int st_off, |
duke@435 | 658 | Register scratch) { |
duke@435 | 659 | if (TaggedStackInterpreter) { |
duke@435 | 660 | int tag_off = st_off + Interpreter::tag_offset_in_bytes(); |
duke@435 | 661 | #ifdef _LP64 |
duke@435 | 662 | Register tag_slot = Rdisp; |
duke@435 | 663 | __ set(tag_off, tag_slot); |
duke@435 | 664 | #else |
duke@435 | 665 | int tag_slot = tag_off; |
duke@435 | 666 | #endif // _LP64 |
duke@435 | 667 | // have to store zero because local slots can be reused (rats!) |
duke@435 | 668 | if (t == frame::TagValue) { |
duke@435 | 669 | __ st_ptr(G0, base, tag_slot); |
duke@435 | 670 | } else if (t == frame::TagCategory2) { |
duke@435 | 671 | __ st_ptr(G0, base, tag_slot); |
duke@435 | 672 | int next_tag_off = st_off - Interpreter::stackElementSize() + |
duke@435 | 673 | Interpreter::tag_offset_in_bytes(); |
duke@435 | 674 | #ifdef _LP64 |
duke@435 | 675 | __ set(next_tag_off, tag_slot); |
duke@435 | 676 | #else |
duke@435 | 677 | tag_slot = next_tag_off; |
duke@435 | 678 | #endif // _LP64 |
duke@435 | 679 | __ st_ptr(G0, base, tag_slot); |
duke@435 | 680 | } else { |
duke@435 | 681 | __ mov(t, scratch); |
duke@435 | 682 | __ st_ptr(scratch, base, tag_slot); |
duke@435 | 683 | } |
duke@435 | 684 | } |
duke@435 | 685 | } |
duke@435 | 686 | |
duke@435 | 687 | #ifdef _LP64 |
duke@435 | 688 | Register AdapterGenerator::arg_slot(const int st_off) { |
duke@435 | 689 | __ set( arg_offset(st_off), Rdisp); |
duke@435 | 690 | return Rdisp; |
duke@435 | 691 | } |
duke@435 | 692 | |
duke@435 | 693 | Register AdapterGenerator::next_arg_slot(const int st_off){ |
duke@435 | 694 | __ set( next_arg_offset(st_off), Rdisp); |
duke@435 | 695 | return Rdisp; |
duke@435 | 696 | } |
duke@435 | 697 | #endif // _LP64 |
duke@435 | 698 | |
duke@435 | 699 | // Stores long into offset pointed to by base |
duke@435 | 700 | void AdapterGenerator::store_c2i_long(Register r, Register base, |
duke@435 | 701 | const int st_off, bool is_stack) { |
duke@435 | 702 | #ifdef COMPILER2 |
duke@435 | 703 | #ifdef _LP64 |
duke@435 | 704 | // In V9, longs are given 2 64-bit slots in the interpreter, but the |
duke@435 | 705 | // data is passed in only 1 slot. |
duke@435 | 706 | __ stx(r, base, next_arg_slot(st_off)); |
duke@435 | 707 | #else |
duke@435 | 708 | // Misaligned store of 64-bit data |
duke@435 | 709 | __ stw(r, base, arg_slot(st_off)); // lo bits |
duke@435 | 710 | __ srlx(r, 32, r); |
duke@435 | 711 | __ stw(r, base, next_arg_slot(st_off)); // hi bits |
duke@435 | 712 | #endif // _LP64 |
duke@435 | 713 | #else |
duke@435 | 714 | if (is_stack) { |
duke@435 | 715 | // Misaligned store of 64-bit data |
duke@435 | 716 | __ stw(r, base, arg_slot(st_off)); // lo bits |
duke@435 | 717 | __ srlx(r, 32, r); |
duke@435 | 718 | __ stw(r, base, next_arg_slot(st_off)); // hi bits |
duke@435 | 719 | } else { |
duke@435 | 720 | __ stw(r->successor(), base, arg_slot(st_off) ); // lo bits |
duke@435 | 721 | __ stw(r , base, next_arg_slot(st_off)); // hi bits |
duke@435 | 722 | } |
duke@435 | 723 | #endif // COMPILER2 |
duke@435 | 724 | tag_c2i_arg(frame::TagCategory2, base, st_off, r); |
duke@435 | 725 | } |
duke@435 | 726 | |
duke@435 | 727 | void AdapterGenerator::store_c2i_object(Register r, Register base, |
duke@435 | 728 | const int st_off) { |
duke@435 | 729 | __ st_ptr (r, base, arg_slot(st_off)); |
duke@435 | 730 | tag_c2i_arg(frame::TagReference, base, st_off, r); |
duke@435 | 731 | } |
duke@435 | 732 | |
duke@435 | 733 | void AdapterGenerator::store_c2i_int(Register r, Register base, |
duke@435 | 734 | const int st_off) { |
duke@435 | 735 | __ st (r, base, arg_slot(st_off)); |
duke@435 | 736 | tag_c2i_arg(frame::TagValue, base, st_off, r); |
duke@435 | 737 | } |
duke@435 | 738 | |
duke@435 | 739 | // Stores into offset pointed to by base |
duke@435 | 740 | void AdapterGenerator::store_c2i_double(VMReg r_2, |
duke@435 | 741 | VMReg r_1, Register base, const int st_off) { |
duke@435 | 742 | #ifdef _LP64 |
duke@435 | 743 | // In V9, doubles are given 2 64-bit slots in the interpreter, but the |
duke@435 | 744 | // data is passed in only 1 slot. |
duke@435 | 745 | __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), base, next_arg_slot(st_off)); |
duke@435 | 746 | #else |
duke@435 | 747 | // Need to marshal 64-bit value from misaligned Lesp loads |
duke@435 | 748 | __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), base, next_arg_slot(st_off)); |
duke@435 | 749 | __ stf(FloatRegisterImpl::S, r_2->as_FloatRegister(), base, arg_slot(st_off) ); |
duke@435 | 750 | #endif |
duke@435 | 751 | tag_c2i_arg(frame::TagCategory2, base, st_off, G1_scratch); |
duke@435 | 752 | } |
duke@435 | 753 | |
duke@435 | 754 | void AdapterGenerator::store_c2i_float(FloatRegister f, Register base, |
duke@435 | 755 | const int st_off) { |
duke@435 | 756 | __ stf(FloatRegisterImpl::S, f, base, arg_slot(st_off)); |
duke@435 | 757 | tag_c2i_arg(frame::TagValue, base, st_off, G1_scratch); |
duke@435 | 758 | } |
duke@435 | 759 | |
duke@435 | 760 | void AdapterGenerator::gen_c2i_adapter( |
duke@435 | 761 | int total_args_passed, |
duke@435 | 762 | // VMReg max_arg, |
duke@435 | 763 | int comp_args_on_stack, // VMRegStackSlots |
duke@435 | 764 | const BasicType *sig_bt, |
duke@435 | 765 | const VMRegPair *regs, |
duke@435 | 766 | Label& skip_fixup) { |
duke@435 | 767 | |
duke@435 | 768 | // Before we get into the guts of the C2I adapter, see if we should be here |
duke@435 | 769 | // at all. We've come from compiled code and are attempting to jump to the |
duke@435 | 770 | // interpreter, which means the caller made a static call to get here |
duke@435 | 771 | // (vcalls always get a compiled target if there is one). Check for a |
duke@435 | 772 | // compiled target. If there is one, we need to patch the caller's call. |
duke@435 | 773 | // However we will run interpreted if we come thru here. The next pass |
duke@435 | 774 | // thru the call site will run compiled. If we ran compiled here then |
duke@435 | 775 | // we can (theorectically) do endless i2c->c2i->i2c transitions during |
duke@435 | 776 | // deopt/uncommon trap cycles. If we always go interpreted here then |
duke@435 | 777 | // we can have at most one and don't need to play any tricks to keep |
duke@435 | 778 | // from endlessly growing the stack. |
duke@435 | 779 | // |
duke@435 | 780 | // Actually if we detected that we had an i2c->c2i transition here we |
duke@435 | 781 | // ought to be able to reset the world back to the state of the interpreted |
duke@435 | 782 | // call and not bother building another interpreter arg area. We don't |
duke@435 | 783 | // do that at this point. |
duke@435 | 784 | |
duke@435 | 785 | patch_callers_callsite(); |
duke@435 | 786 | |
duke@435 | 787 | __ bind(skip_fixup); |
duke@435 | 788 | |
duke@435 | 789 | // Since all args are passed on the stack, total_args_passed*wordSize is the |
duke@435 | 790 | // space we need. Add in varargs area needed by the interpreter. Round up |
duke@435 | 791 | // to stack alignment. |
duke@435 | 792 | const int arg_size = total_args_passed * Interpreter::stackElementSize(); |
duke@435 | 793 | const int varargs_area = |
duke@435 | 794 | (frame::varargs_offset - frame::register_save_words)*wordSize; |
duke@435 | 795 | const int extraspace = round_to(arg_size + varargs_area, 2*wordSize); |
duke@435 | 796 | |
duke@435 | 797 | int bias = STACK_BIAS; |
duke@435 | 798 | const int interp_arg_offset = frame::varargs_offset*wordSize + |
duke@435 | 799 | (total_args_passed-1)*Interpreter::stackElementSize(); |
duke@435 | 800 | |
duke@435 | 801 | Register base = SP; |
duke@435 | 802 | |
duke@435 | 803 | #ifdef _LP64 |
duke@435 | 804 | // In the 64bit build because of wider slots and STACKBIAS we can run |
duke@435 | 805 | // out of bits in the displacement to do loads and stores. Use g3 as |
duke@435 | 806 | // temporary displacement. |
duke@435 | 807 | if (! __ is_simm13(extraspace)) { |
duke@435 | 808 | __ set(extraspace, G3_scratch); |
duke@435 | 809 | __ sub(SP, G3_scratch, SP); |
duke@435 | 810 | } else { |
duke@435 | 811 | __ sub(SP, extraspace, SP); |
duke@435 | 812 | } |
duke@435 | 813 | set_Rdisp(G3_scratch); |
duke@435 | 814 | #else |
duke@435 | 815 | __ sub(SP, extraspace, SP); |
duke@435 | 816 | #endif // _LP64 |
duke@435 | 817 | |
duke@435 | 818 | // First write G1 (if used) to where ever it must go |
duke@435 | 819 | for (int i=0; i<total_args_passed; i++) { |
duke@435 | 820 | const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize()) + bias; |
duke@435 | 821 | VMReg r_1 = regs[i].first(); |
duke@435 | 822 | VMReg r_2 = regs[i].second(); |
duke@435 | 823 | if (r_1 == G1_scratch->as_VMReg()) { |
duke@435 | 824 | if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) { |
duke@435 | 825 | store_c2i_object(G1_scratch, base, st_off); |
duke@435 | 826 | } else if (sig_bt[i] == T_LONG) { |
duke@435 | 827 | assert(!TieredCompilation, "should not use register args for longs"); |
duke@435 | 828 | store_c2i_long(G1_scratch, base, st_off, false); |
duke@435 | 829 | } else { |
duke@435 | 830 | store_c2i_int(G1_scratch, base, st_off); |
duke@435 | 831 | } |
duke@435 | 832 | } |
duke@435 | 833 | } |
duke@435 | 834 | |
duke@435 | 835 | // Now write the args into the outgoing interpreter space |
duke@435 | 836 | for (int i=0; i<total_args_passed; i++) { |
duke@435 | 837 | const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize()) + bias; |
duke@435 | 838 | VMReg r_1 = regs[i].first(); |
duke@435 | 839 | VMReg r_2 = regs[i].second(); |
duke@435 | 840 | if (!r_1->is_valid()) { |
duke@435 | 841 | assert(!r_2->is_valid(), ""); |
duke@435 | 842 | continue; |
duke@435 | 843 | } |
duke@435 | 844 | // Skip G1 if found as we did it first in order to free it up |
duke@435 | 845 | if (r_1 == G1_scratch->as_VMReg()) { |
duke@435 | 846 | continue; |
duke@435 | 847 | } |
duke@435 | 848 | #ifdef ASSERT |
duke@435 | 849 | bool G1_forced = false; |
duke@435 | 850 | #endif // ASSERT |
duke@435 | 851 | if (r_1->is_stack()) { // Pretend stack targets are loaded into G1 |
duke@435 | 852 | #ifdef _LP64 |
duke@435 | 853 | Register ld_off = Rdisp; |
duke@435 | 854 | __ set(reg2offset(r_1) + extraspace + bias, ld_off); |
duke@435 | 855 | #else |
duke@435 | 856 | int ld_off = reg2offset(r_1) + extraspace + bias; |
duke@435 | 857 | #ifdef ASSERT |
duke@435 | 858 | G1_forced = true; |
duke@435 | 859 | #endif // ASSERT |
duke@435 | 860 | #endif // _LP64 |
duke@435 | 861 | r_1 = G1_scratch->as_VMReg();// as part of the load/store shuffle |
duke@435 | 862 | if (!r_2->is_valid()) __ ld (base, ld_off, G1_scratch); |
duke@435 | 863 | else __ ldx(base, ld_off, G1_scratch); |
duke@435 | 864 | } |
duke@435 | 865 | |
duke@435 | 866 | if (r_1->is_Register()) { |
duke@435 | 867 | Register r = r_1->as_Register()->after_restore(); |
duke@435 | 868 | if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) { |
duke@435 | 869 | store_c2i_object(r, base, st_off); |
duke@435 | 870 | } else if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) { |
duke@435 | 871 | if (TieredCompilation) { |
duke@435 | 872 | assert(G1_forced || sig_bt[i] != T_LONG, "should not use register args for longs"); |
duke@435 | 873 | } |
duke@435 | 874 | store_c2i_long(r, base, st_off, r_2->is_stack()); |
duke@435 | 875 | } else { |
duke@435 | 876 | store_c2i_int(r, base, st_off); |
duke@435 | 877 | } |
duke@435 | 878 | } else { |
duke@435 | 879 | assert(r_1->is_FloatRegister(), ""); |
duke@435 | 880 | if (sig_bt[i] == T_FLOAT) { |
duke@435 | 881 | store_c2i_float(r_1->as_FloatRegister(), base, st_off); |
duke@435 | 882 | } else { |
duke@435 | 883 | assert(sig_bt[i] == T_DOUBLE, "wrong type"); |
duke@435 | 884 | store_c2i_double(r_2, r_1, base, st_off); |
duke@435 | 885 | } |
duke@435 | 886 | } |
duke@435 | 887 | } |
duke@435 | 888 | |
duke@435 | 889 | #ifdef _LP64 |
duke@435 | 890 | // Need to reload G3_scratch, used for temporary displacements. |
duke@435 | 891 | __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch); |
duke@435 | 892 | |
duke@435 | 893 | // Pass O5_savedSP as an argument to the interpreter. |
duke@435 | 894 | // The interpreter will restore SP to this value before returning. |
duke@435 | 895 | __ set(extraspace, G1); |
duke@435 | 896 | __ add(SP, G1, O5_savedSP); |
duke@435 | 897 | #else |
duke@435 | 898 | // Pass O5_savedSP as an argument to the interpreter. |
duke@435 | 899 | // The interpreter will restore SP to this value before returning. |
duke@435 | 900 | __ add(SP, extraspace, O5_savedSP); |
duke@435 | 901 | #endif // _LP64 |
duke@435 | 902 | |
duke@435 | 903 | __ mov((frame::varargs_offset)*wordSize - |
duke@435 | 904 | 1*Interpreter::stackElementSize()+bias+BytesPerWord, G1); |
duke@435 | 905 | // Jump to the interpreter just as if interpreter was doing it. |
duke@435 | 906 | __ jmpl(G3_scratch, 0, G0); |
duke@435 | 907 | // Setup Lesp for the call. Cannot actually set Lesp as the current Lesp |
duke@435 | 908 | // (really L0) is in use by the compiled frame as a generic temp. However, |
duke@435 | 909 | // the interpreter does not know where its args are without some kind of |
duke@435 | 910 | // arg pointer being passed in. Pass it in Gargs. |
duke@435 | 911 | __ delayed()->add(SP, G1, Gargs); |
duke@435 | 912 | } |
duke@435 | 913 | |
duke@435 | 914 | void AdapterGenerator::gen_i2c_adapter( |
duke@435 | 915 | int total_args_passed, |
duke@435 | 916 | // VMReg max_arg, |
duke@435 | 917 | int comp_args_on_stack, // VMRegStackSlots |
duke@435 | 918 | const BasicType *sig_bt, |
duke@435 | 919 | const VMRegPair *regs) { |
duke@435 | 920 | |
duke@435 | 921 | // Generate an I2C adapter: adjust the I-frame to make space for the C-frame |
duke@435 | 922 | // layout. Lesp was saved by the calling I-frame and will be restored on |
duke@435 | 923 | // return. Meanwhile, outgoing arg space is all owned by the callee |
duke@435 | 924 | // C-frame, so we can mangle it at will. After adjusting the frame size, |
duke@435 | 925 | // hoist register arguments and repack other args according to the compiled |
duke@435 | 926 | // code convention. Finally, end in a jump to the compiled code. The entry |
duke@435 | 927 | // point address is the start of the buffer. |
duke@435 | 928 | |
duke@435 | 929 | // We will only enter here from an interpreted frame and never from after |
duke@435 | 930 | // passing thru a c2i. Azul allowed this but we do not. If we lose the |
duke@435 | 931 | // race and use a c2i we will remain interpreted for the race loser(s). |
duke@435 | 932 | // This removes all sorts of headaches on the x86 side and also eliminates |
duke@435 | 933 | // the possibility of having c2i -> i2c -> c2i -> ... endless transitions. |
duke@435 | 934 | |
duke@435 | 935 | // As you can see from the list of inputs & outputs there are not a lot |
duke@435 | 936 | // of temp registers to work with: mostly G1, G3 & G4. |
duke@435 | 937 | |
duke@435 | 938 | // Inputs: |
duke@435 | 939 | // G2_thread - TLS |
duke@435 | 940 | // G5_method - Method oop |
duke@435 | 941 | // O0 - Flag telling us to restore SP from O5 |
duke@435 | 942 | // O4_args - Pointer to interpreter's args |
duke@435 | 943 | // O5 - Caller's saved SP, to be restored if needed |
duke@435 | 944 | // O6 - Current SP! |
duke@435 | 945 | // O7 - Valid return address |
duke@435 | 946 | // L0-L7, I0-I7 - Caller's temps (no frame pushed yet) |
duke@435 | 947 | |
duke@435 | 948 | // Outputs: |
duke@435 | 949 | // G2_thread - TLS |
duke@435 | 950 | // G1, G4 - Outgoing long args in 32-bit build |
duke@435 | 951 | // O0-O5 - Outgoing args in compiled layout |
duke@435 | 952 | // O6 - Adjusted or restored SP |
duke@435 | 953 | // O7 - Valid return address |
duke@435 | 954 | // L0-L7, I0-I7 - Caller's temps (no frame pushed yet) |
duke@435 | 955 | // F0-F7 - more outgoing args |
duke@435 | 956 | |
duke@435 | 957 | |
duke@435 | 958 | // O4 is about to get loaded up with compiled callee's args |
duke@435 | 959 | __ sub(Gargs, BytesPerWord, Gargs); |
duke@435 | 960 | |
duke@435 | 961 | #ifdef ASSERT |
duke@435 | 962 | { |
duke@435 | 963 | // on entry OsavedSP and SP should be equal |
duke@435 | 964 | Label ok; |
duke@435 | 965 | __ cmp(O5_savedSP, SP); |
duke@435 | 966 | __ br(Assembler::equal, false, Assembler::pt, ok); |
duke@435 | 967 | __ delayed()->nop(); |
duke@435 | 968 | __ stop("I5_savedSP not set"); |
duke@435 | 969 | __ should_not_reach_here(); |
duke@435 | 970 | __ bind(ok); |
duke@435 | 971 | } |
duke@435 | 972 | #endif |
duke@435 | 973 | |
duke@435 | 974 | // ON ENTRY TO THE CODE WE ARE MAKING, WE HAVE AN INTERPRETED FRAME |
duke@435 | 975 | // WITH O7 HOLDING A VALID RETURN PC |
duke@435 | 976 | // |
duke@435 | 977 | // | | |
duke@435 | 978 | // : java stack : |
duke@435 | 979 | // | | |
duke@435 | 980 | // +--------------+ <--- start of outgoing args |
duke@435 | 981 | // | receiver | | |
duke@435 | 982 | // : rest of args : |---size is java-arg-words |
duke@435 | 983 | // | | | |
duke@435 | 984 | // +--------------+ <--- O4_args (misaligned) and Lesp if prior is not C2I |
duke@435 | 985 | // | | | |
duke@435 | 986 | // : unused : |---Space for max Java stack, plus stack alignment |
duke@435 | 987 | // | | | |
duke@435 | 988 | // +--------------+ <--- SP + 16*wordsize |
duke@435 | 989 | // | | |
duke@435 | 990 | // : window : |
duke@435 | 991 | // | | |
duke@435 | 992 | // +--------------+ <--- SP |
duke@435 | 993 | |
duke@435 | 994 | // WE REPACK THE STACK. We use the common calling convention layout as |
duke@435 | 995 | // discovered by calling SharedRuntime::calling_convention. We assume it |
duke@435 | 996 | // causes an arbitrary shuffle of memory, which may require some register |
duke@435 | 997 | // temps to do the shuffle. We hope for (and optimize for) the case where |
duke@435 | 998 | // temps are not needed. We may have to resize the stack slightly, in case |
duke@435 | 999 | // we need alignment padding (32-bit interpreter can pass longs & doubles |
duke@435 | 1000 | // misaligned, but the compilers expect them aligned). |
duke@435 | 1001 | // |
duke@435 | 1002 | // | | |
duke@435 | 1003 | // : java stack : |
duke@435 | 1004 | // | | |
duke@435 | 1005 | // +--------------+ <--- start of outgoing args |
duke@435 | 1006 | // | pad, align | | |
duke@435 | 1007 | // +--------------+ | |
duke@435 | 1008 | // | ints, floats | |---Outgoing stack args, packed low. |
duke@435 | 1009 | // +--------------+ | First few args in registers. |
duke@435 | 1010 | // : doubles : | |
duke@435 | 1011 | // | longs | | |
duke@435 | 1012 | // +--------------+ <--- SP' + 16*wordsize |
duke@435 | 1013 | // | | |
duke@435 | 1014 | // : window : |
duke@435 | 1015 | // | | |
duke@435 | 1016 | // +--------------+ <--- SP' |
duke@435 | 1017 | |
duke@435 | 1018 | // ON EXIT FROM THE CODE WE ARE MAKING, WE STILL HAVE AN INTERPRETED FRAME |
duke@435 | 1019 | // WITH O7 HOLDING A VALID RETURN PC - ITS JUST THAT THE ARGS ARE NOW SETUP |
duke@435 | 1020 | // FOR COMPILED CODE AND THE FRAME SLIGHTLY GROWN. |
duke@435 | 1021 | |
duke@435 | 1022 | // Cut-out for having no stack args. Since up to 6 args are passed |
duke@435 | 1023 | // in registers, we will commonly have no stack args. |
duke@435 | 1024 | if (comp_args_on_stack > 0) { |
duke@435 | 1025 | |
duke@435 | 1026 | // Convert VMReg stack slots to words. |
duke@435 | 1027 | int comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord; |
duke@435 | 1028 | // Round up to miminum stack alignment, in wordSize |
duke@435 | 1029 | comp_words_on_stack = round_to(comp_words_on_stack, 2); |
duke@435 | 1030 | // Now compute the distance from Lesp to SP. This calculation does not |
duke@435 | 1031 | // include the space for total_args_passed because Lesp has not yet popped |
duke@435 | 1032 | // the arguments. |
duke@435 | 1033 | __ sub(SP, (comp_words_on_stack)*wordSize, SP); |
duke@435 | 1034 | } |
duke@435 | 1035 | |
duke@435 | 1036 | // Will jump to the compiled code just as if compiled code was doing it. |
duke@435 | 1037 | // Pre-load the register-jump target early, to schedule it better. |
duke@435 | 1038 | __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_compiled_offset()), G3); |
duke@435 | 1039 | |
duke@435 | 1040 | // Now generate the shuffle code. Pick up all register args and move the |
duke@435 | 1041 | // rest through G1_scratch. |
duke@435 | 1042 | for (int i=0; i<total_args_passed; i++) { |
duke@435 | 1043 | if (sig_bt[i] == T_VOID) { |
duke@435 | 1044 | // Longs and doubles are passed in native word order, but misaligned |
duke@435 | 1045 | // in the 32-bit build. |
duke@435 | 1046 | assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half"); |
duke@435 | 1047 | continue; |
duke@435 | 1048 | } |
duke@435 | 1049 | |
duke@435 | 1050 | // Pick up 0, 1 or 2 words from Lesp+offset. Assume mis-aligned in the |
duke@435 | 1051 | // 32-bit build and aligned in the 64-bit build. Look for the obvious |
duke@435 | 1052 | // ldx/lddf optimizations. |
duke@435 | 1053 | |
duke@435 | 1054 | // Load in argument order going down. |
duke@435 | 1055 | const int ld_off = (total_args_passed-i)*Interpreter::stackElementSize(); |
duke@435 | 1056 | #ifdef _LP64 |
duke@435 | 1057 | set_Rdisp(G1_scratch); |
duke@435 | 1058 | #endif // _LP64 |
duke@435 | 1059 | |
duke@435 | 1060 | VMReg r_1 = regs[i].first(); |
duke@435 | 1061 | VMReg r_2 = regs[i].second(); |
duke@435 | 1062 | if (!r_1->is_valid()) { |
duke@435 | 1063 | assert(!r_2->is_valid(), ""); |
duke@435 | 1064 | continue; |
duke@435 | 1065 | } |
duke@435 | 1066 | if (r_1->is_stack()) { // Pretend stack targets are loaded into F8/F9 |
duke@435 | 1067 | r_1 = F8->as_VMReg(); // as part of the load/store shuffle |
duke@435 | 1068 | if (r_2->is_valid()) r_2 = r_1->next(); |
duke@435 | 1069 | } |
duke@435 | 1070 | if (r_1->is_Register()) { // Register argument |
duke@435 | 1071 | Register r = r_1->as_Register()->after_restore(); |
duke@435 | 1072 | if (!r_2->is_valid()) { |
duke@435 | 1073 | __ ld(Gargs, arg_slot(ld_off), r); |
duke@435 | 1074 | } else { |
duke@435 | 1075 | #ifdef _LP64 |
duke@435 | 1076 | // In V9, longs are given 2 64-bit slots in the interpreter, but the |
duke@435 | 1077 | // data is passed in only 1 slot. |
duke@435 | 1078 | Register slot = (sig_bt[i]==T_LONG) ? |
duke@435 | 1079 | next_arg_slot(ld_off) : arg_slot(ld_off); |
duke@435 | 1080 | __ ldx(Gargs, slot, r); |
duke@435 | 1081 | #else |
duke@435 | 1082 | // Need to load a 64-bit value into G1/G4, but G1/G4 is being used in the |
duke@435 | 1083 | // stack shuffle. Load the first 2 longs into G1/G4 later. |
duke@435 | 1084 | #endif |
duke@435 | 1085 | } |
duke@435 | 1086 | } else { |
duke@435 | 1087 | assert(r_1->is_FloatRegister(), ""); |
duke@435 | 1088 | if (!r_2->is_valid()) { |
duke@435 | 1089 | __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_1->as_FloatRegister()); |
duke@435 | 1090 | } else { |
duke@435 | 1091 | #ifdef _LP64 |
duke@435 | 1092 | // In V9, doubles are given 2 64-bit slots in the interpreter, but the |
duke@435 | 1093 | // data is passed in only 1 slot. This code also handles longs that |
duke@435 | 1094 | // are passed on the stack, but need a stack-to-stack move through a |
duke@435 | 1095 | // spare float register. |
duke@435 | 1096 | Register slot = (sig_bt[i]==T_LONG || sig_bt[i] == T_DOUBLE) ? |
duke@435 | 1097 | next_arg_slot(ld_off) : arg_slot(ld_off); |
duke@435 | 1098 | __ ldf(FloatRegisterImpl::D, Gargs, slot, r_1->as_FloatRegister()); |
duke@435 | 1099 | #else |
duke@435 | 1100 | // Need to marshal 64-bit value from misaligned Lesp loads |
duke@435 | 1101 | __ ldf(FloatRegisterImpl::S, Gargs, next_arg_slot(ld_off), r_1->as_FloatRegister()); |
duke@435 | 1102 | __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_2->as_FloatRegister()); |
duke@435 | 1103 | #endif |
duke@435 | 1104 | } |
duke@435 | 1105 | } |
duke@435 | 1106 | // Was the argument really intended to be on the stack, but was loaded |
duke@435 | 1107 | // into F8/F9? |
duke@435 | 1108 | if (regs[i].first()->is_stack()) { |
duke@435 | 1109 | assert(r_1->as_FloatRegister() == F8, "fix this code"); |
duke@435 | 1110 | // Convert stack slot to an SP offset |
duke@435 | 1111 | int st_off = reg2offset(regs[i].first()) + STACK_BIAS; |
duke@435 | 1112 | // Store down the shuffled stack word. Target address _is_ aligned. |
duke@435 | 1113 | if (!r_2->is_valid()) __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), SP, st_off); |
duke@435 | 1114 | else __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), SP, st_off); |
duke@435 | 1115 | } |
duke@435 | 1116 | } |
duke@435 | 1117 | bool made_space = false; |
duke@435 | 1118 | #ifndef _LP64 |
duke@435 | 1119 | // May need to pick up a few long args in G1/G4 |
duke@435 | 1120 | bool g4_crushed = false; |
duke@435 | 1121 | bool g3_crushed = false; |
duke@435 | 1122 | for (int i=0; i<total_args_passed; i++) { |
duke@435 | 1123 | if (regs[i].first()->is_Register() && regs[i].second()->is_valid()) { |
duke@435 | 1124 | // Load in argument order going down |
duke@435 | 1125 | int ld_off = (total_args_passed-i)*Interpreter::stackElementSize(); |
duke@435 | 1126 | // Need to marshal 64-bit value from misaligned Lesp loads |
duke@435 | 1127 | Register r = regs[i].first()->as_Register()->after_restore(); |
duke@435 | 1128 | if (r == G1 || r == G4) { |
duke@435 | 1129 | assert(!g4_crushed, "ordering problem"); |
duke@435 | 1130 | if (r == G4){ |
duke@435 | 1131 | g4_crushed = true; |
duke@435 | 1132 | __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits |
duke@435 | 1133 | __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits |
duke@435 | 1134 | } else { |
duke@435 | 1135 | // better schedule this way |
duke@435 | 1136 | __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits |
duke@435 | 1137 | __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits |
duke@435 | 1138 | } |
duke@435 | 1139 | g3_crushed = true; |
duke@435 | 1140 | __ sllx(r, 32, r); |
duke@435 | 1141 | __ or3(G3_scratch, r, r); |
duke@435 | 1142 | } else { |
duke@435 | 1143 | assert(r->is_out(), "longs passed in two O registers"); |
duke@435 | 1144 | __ ld (Gargs, arg_slot(ld_off) , r->successor()); // Load lo bits |
duke@435 | 1145 | __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits |
duke@435 | 1146 | } |
duke@435 | 1147 | } |
duke@435 | 1148 | } |
duke@435 | 1149 | #endif |
duke@435 | 1150 | |
duke@435 | 1151 | // Jump to the compiled code just as if compiled code was doing it. |
duke@435 | 1152 | // |
duke@435 | 1153 | #ifndef _LP64 |
duke@435 | 1154 | if (g3_crushed) { |
duke@435 | 1155 | // Rats load was wasted, at least it is in cache... |
duke@435 | 1156 | __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_compiled_offset()), G3); |
duke@435 | 1157 | } |
duke@435 | 1158 | #endif /* _LP64 */ |
duke@435 | 1159 | |
duke@435 | 1160 | // 6243940 We might end up in handle_wrong_method if |
duke@435 | 1161 | // the callee is deoptimized as we race thru here. If that |
duke@435 | 1162 | // happens we don't want to take a safepoint because the |
duke@435 | 1163 | // caller frame will look interpreted and arguments are now |
duke@435 | 1164 | // "compiled" so it is much better to make this transition |
duke@435 | 1165 | // invisible to the stack walking code. Unfortunately if |
duke@435 | 1166 | // we try and find the callee by normal means a safepoint |
duke@435 | 1167 | // is possible. So we stash the desired callee in the thread |
duke@435 | 1168 | // and the vm will find there should this case occur. |
duke@435 | 1169 | Address callee_target_addr(G2_thread, 0, in_bytes(JavaThread::callee_target_offset())); |
duke@435 | 1170 | __ st_ptr(G5_method, callee_target_addr); |
duke@435 | 1171 | |
duke@435 | 1172 | if (StressNonEntrant) { |
duke@435 | 1173 | // Open a big window for deopt failure |
duke@435 | 1174 | __ save_frame(0); |
duke@435 | 1175 | __ mov(G0, L0); |
duke@435 | 1176 | Label loop; |
duke@435 | 1177 | __ bind(loop); |
duke@435 | 1178 | __ sub(L0, 1, L0); |
duke@435 | 1179 | __ br_null(L0, false, Assembler::pt, loop); |
duke@435 | 1180 | __ delayed()->nop(); |
duke@435 | 1181 | |
duke@435 | 1182 | __ restore(); |
duke@435 | 1183 | } |
duke@435 | 1184 | |
duke@435 | 1185 | |
duke@435 | 1186 | __ jmpl(G3, 0, G0); |
duke@435 | 1187 | __ delayed()->nop(); |
duke@435 | 1188 | } |
duke@435 | 1189 | |
duke@435 | 1190 | // --------------------------------------------------------------- |
duke@435 | 1191 | AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm, |
duke@435 | 1192 | int total_args_passed, |
duke@435 | 1193 | // VMReg max_arg, |
duke@435 | 1194 | int comp_args_on_stack, // VMRegStackSlots |
duke@435 | 1195 | const BasicType *sig_bt, |
duke@435 | 1196 | const VMRegPair *regs) { |
duke@435 | 1197 | address i2c_entry = __ pc(); |
duke@435 | 1198 | |
duke@435 | 1199 | AdapterGenerator agen(masm); |
duke@435 | 1200 | |
duke@435 | 1201 | agen.gen_i2c_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs); |
duke@435 | 1202 | |
duke@435 | 1203 | |
duke@435 | 1204 | // ------------------------------------------------------------------------- |
duke@435 | 1205 | // Generate a C2I adapter. On entry we know G5 holds the methodOop. The |
duke@435 | 1206 | // args start out packed in the compiled layout. They need to be unpacked |
duke@435 | 1207 | // into the interpreter layout. This will almost always require some stack |
duke@435 | 1208 | // space. We grow the current (compiled) stack, then repack the args. We |
duke@435 | 1209 | // finally end in a jump to the generic interpreter entry point. On exit |
duke@435 | 1210 | // from the interpreter, the interpreter will restore our SP (lest the |
duke@435 | 1211 | // compiled code, which relys solely on SP and not FP, get sick). |
duke@435 | 1212 | |
duke@435 | 1213 | address c2i_unverified_entry = __ pc(); |
duke@435 | 1214 | Label skip_fixup; |
duke@435 | 1215 | { |
duke@435 | 1216 | #if !defined(_LP64) && defined(COMPILER2) |
duke@435 | 1217 | Register R_temp = L0; // another scratch register |
duke@435 | 1218 | #else |
duke@435 | 1219 | Register R_temp = G1; // another scratch register |
duke@435 | 1220 | #endif |
duke@435 | 1221 | |
duke@435 | 1222 | Address ic_miss(G3_scratch, SharedRuntime::get_ic_miss_stub()); |
duke@435 | 1223 | |
duke@435 | 1224 | __ verify_oop(O0); |
duke@435 | 1225 | __ verify_oop(G5_method); |
coleenp@548 | 1226 | __ load_klass(O0, G3_scratch); |
duke@435 | 1227 | __ verify_oop(G3_scratch); |
duke@435 | 1228 | |
duke@435 | 1229 | #if !defined(_LP64) && defined(COMPILER2) |
duke@435 | 1230 | __ save(SP, -frame::register_save_words*wordSize, SP); |
duke@435 | 1231 | __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp); |
duke@435 | 1232 | __ verify_oop(R_temp); |
duke@435 | 1233 | __ cmp(G3_scratch, R_temp); |
duke@435 | 1234 | __ restore(); |
duke@435 | 1235 | #else |
duke@435 | 1236 | __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp); |
duke@435 | 1237 | __ verify_oop(R_temp); |
duke@435 | 1238 | __ cmp(G3_scratch, R_temp); |
duke@435 | 1239 | #endif |
duke@435 | 1240 | |
duke@435 | 1241 | Label ok, ok2; |
duke@435 | 1242 | __ brx(Assembler::equal, false, Assembler::pt, ok); |
duke@435 | 1243 | __ delayed()->ld_ptr(G5_method, compiledICHolderOopDesc::holder_method_offset(), G5_method); |
duke@435 | 1244 | __ jump_to(ic_miss); |
duke@435 | 1245 | __ delayed()->nop(); |
duke@435 | 1246 | |
duke@435 | 1247 | __ bind(ok); |
duke@435 | 1248 | // Method might have been compiled since the call site was patched to |
duke@435 | 1249 | // interpreted if that is the case treat it as a miss so we can get |
duke@435 | 1250 | // the call site corrected. |
duke@435 | 1251 | __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch); |
duke@435 | 1252 | __ bind(ok2); |
duke@435 | 1253 | __ br_null(G3_scratch, false, __ pt, skip_fixup); |
duke@435 | 1254 | __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch); |
duke@435 | 1255 | __ jump_to(ic_miss); |
duke@435 | 1256 | __ delayed()->nop(); |
duke@435 | 1257 | |
duke@435 | 1258 | } |
duke@435 | 1259 | |
duke@435 | 1260 | address c2i_entry = __ pc(); |
duke@435 | 1261 | |
duke@435 | 1262 | agen.gen_c2i_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup); |
duke@435 | 1263 | |
duke@435 | 1264 | __ flush(); |
duke@435 | 1265 | return new AdapterHandlerEntry(i2c_entry, c2i_entry, c2i_unverified_entry); |
duke@435 | 1266 | |
duke@435 | 1267 | } |
duke@435 | 1268 | |
duke@435 | 1269 | // Helper function for native calling conventions |
duke@435 | 1270 | static VMReg int_stk_helper( int i ) { |
duke@435 | 1271 | // Bias any stack based VMReg we get by ignoring the window area |
duke@435 | 1272 | // but not the register parameter save area. |
duke@435 | 1273 | // |
duke@435 | 1274 | // This is strange for the following reasons. We'd normally expect |
duke@435 | 1275 | // the calling convention to return an VMReg for a stack slot |
duke@435 | 1276 | // completely ignoring any abi reserved area. C2 thinks of that |
duke@435 | 1277 | // abi area as only out_preserve_stack_slots. This does not include |
duke@435 | 1278 | // the area allocated by the C abi to store down integer arguments |
duke@435 | 1279 | // because the java calling convention does not use it. So |
duke@435 | 1280 | // since c2 assumes that there are only out_preserve_stack_slots |
duke@435 | 1281 | // to bias the optoregs (which impacts VMRegs) when actually referencing any actual stack |
duke@435 | 1282 | // location the c calling convention must add in this bias amount |
duke@435 | 1283 | // to make up for the fact that the out_preserve_stack_slots is |
duke@435 | 1284 | // insufficient for C calls. What a mess. I sure hope those 6 |
duke@435 | 1285 | // stack words were worth it on every java call! |
duke@435 | 1286 | |
duke@435 | 1287 | // Another way of cleaning this up would be for out_preserve_stack_slots |
duke@435 | 1288 | // to take a parameter to say whether it was C or java calling conventions. |
duke@435 | 1289 | // Then things might look a little better (but not much). |
duke@435 | 1290 | |
duke@435 | 1291 | int mem_parm_offset = i - SPARC_ARGS_IN_REGS_NUM; |
duke@435 | 1292 | if( mem_parm_offset < 0 ) { |
duke@435 | 1293 | return as_oRegister(i)->as_VMReg(); |
duke@435 | 1294 | } else { |
duke@435 | 1295 | int actual_offset = (mem_parm_offset + frame::memory_parameter_word_sp_offset) * VMRegImpl::slots_per_word; |
duke@435 | 1296 | // Now return a biased offset that will be correct when out_preserve_slots is added back in |
duke@435 | 1297 | return VMRegImpl::stack2reg(actual_offset - SharedRuntime::out_preserve_stack_slots()); |
duke@435 | 1298 | } |
duke@435 | 1299 | } |
duke@435 | 1300 | |
duke@435 | 1301 | |
duke@435 | 1302 | int SharedRuntime::c_calling_convention(const BasicType *sig_bt, |
duke@435 | 1303 | VMRegPair *regs, |
duke@435 | 1304 | int total_args_passed) { |
duke@435 | 1305 | |
duke@435 | 1306 | // Return the number of VMReg stack_slots needed for the args. |
duke@435 | 1307 | // This value does not include an abi space (like register window |
duke@435 | 1308 | // save area). |
duke@435 | 1309 | |
duke@435 | 1310 | // The native convention is V8 if !LP64 |
duke@435 | 1311 | // The LP64 convention is the V9 convention which is slightly more sane. |
duke@435 | 1312 | |
duke@435 | 1313 | // We return the amount of VMReg stack slots we need to reserve for all |
duke@435 | 1314 | // the arguments NOT counting out_preserve_stack_slots. Since we always |
duke@435 | 1315 | // have space for storing at least 6 registers to memory we start with that. |
duke@435 | 1316 | // See int_stk_helper for a further discussion. |
duke@435 | 1317 | int max_stack_slots = (frame::varargs_offset * VMRegImpl::slots_per_word) - SharedRuntime::out_preserve_stack_slots(); |
duke@435 | 1318 | |
duke@435 | 1319 | #ifdef _LP64 |
duke@435 | 1320 | // V9 convention: All things "as-if" on double-wide stack slots. |
duke@435 | 1321 | // Hoist any int/ptr/long's in the first 6 to int regs. |
duke@435 | 1322 | // Hoist any flt/dbl's in the first 16 dbl regs. |
duke@435 | 1323 | int j = 0; // Count of actual args, not HALVES |
duke@435 | 1324 | for( int i=0; i<total_args_passed; i++, j++ ) { |
duke@435 | 1325 | switch( sig_bt[i] ) { |
duke@435 | 1326 | case T_BOOLEAN: |
duke@435 | 1327 | case T_BYTE: |
duke@435 | 1328 | case T_CHAR: |
duke@435 | 1329 | case T_INT: |
duke@435 | 1330 | case T_SHORT: |
duke@435 | 1331 | regs[i].set1( int_stk_helper( j ) ); break; |
duke@435 | 1332 | case T_LONG: |
duke@435 | 1333 | assert( sig_bt[i+1] == T_VOID, "expecting half" ); |
duke@435 | 1334 | case T_ADDRESS: // raw pointers, like current thread, for VM calls |
duke@435 | 1335 | case T_ARRAY: |
duke@435 | 1336 | case T_OBJECT: |
duke@435 | 1337 | regs[i].set2( int_stk_helper( j ) ); |
duke@435 | 1338 | break; |
duke@435 | 1339 | case T_FLOAT: |
duke@435 | 1340 | if ( j < 16 ) { |
duke@435 | 1341 | // V9ism: floats go in ODD registers |
duke@435 | 1342 | regs[i].set1(as_FloatRegister(1 + (j<<1))->as_VMReg()); |
duke@435 | 1343 | } else { |
duke@435 | 1344 | // V9ism: floats go in ODD stack slot |
duke@435 | 1345 | regs[i].set1(VMRegImpl::stack2reg(1 + (j<<1))); |
duke@435 | 1346 | } |
duke@435 | 1347 | break; |
duke@435 | 1348 | case T_DOUBLE: |
duke@435 | 1349 | assert( sig_bt[i+1] == T_VOID, "expecting half" ); |
duke@435 | 1350 | if ( j < 16 ) { |
duke@435 | 1351 | // V9ism: doubles go in EVEN/ODD regs |
duke@435 | 1352 | regs[i].set2(as_FloatRegister(j<<1)->as_VMReg()); |
duke@435 | 1353 | } else { |
duke@435 | 1354 | // V9ism: doubles go in EVEN/ODD stack slots |
duke@435 | 1355 | regs[i].set2(VMRegImpl::stack2reg(j<<1)); |
duke@435 | 1356 | } |
duke@435 | 1357 | break; |
duke@435 | 1358 | case T_VOID: regs[i].set_bad(); j--; break; // Do not count HALVES |
duke@435 | 1359 | default: |
duke@435 | 1360 | ShouldNotReachHere(); |
duke@435 | 1361 | } |
duke@435 | 1362 | if (regs[i].first()->is_stack()) { |
duke@435 | 1363 | int off = regs[i].first()->reg2stack(); |
duke@435 | 1364 | if (off > max_stack_slots) max_stack_slots = off; |
duke@435 | 1365 | } |
duke@435 | 1366 | if (regs[i].second()->is_stack()) { |
duke@435 | 1367 | int off = regs[i].second()->reg2stack(); |
duke@435 | 1368 | if (off > max_stack_slots) max_stack_slots = off; |
duke@435 | 1369 | } |
duke@435 | 1370 | } |
duke@435 | 1371 | |
duke@435 | 1372 | #else // _LP64 |
duke@435 | 1373 | // V8 convention: first 6 things in O-regs, rest on stack. |
duke@435 | 1374 | // Alignment is willy-nilly. |
duke@435 | 1375 | for( int i=0; i<total_args_passed; i++ ) { |
duke@435 | 1376 | switch( sig_bt[i] ) { |
duke@435 | 1377 | case T_ADDRESS: // raw pointers, like current thread, for VM calls |
duke@435 | 1378 | case T_ARRAY: |
duke@435 | 1379 | case T_BOOLEAN: |
duke@435 | 1380 | case T_BYTE: |
duke@435 | 1381 | case T_CHAR: |
duke@435 | 1382 | case T_FLOAT: |
duke@435 | 1383 | case T_INT: |
duke@435 | 1384 | case T_OBJECT: |
duke@435 | 1385 | case T_SHORT: |
duke@435 | 1386 | regs[i].set1( int_stk_helper( i ) ); |
duke@435 | 1387 | break; |
duke@435 | 1388 | case T_DOUBLE: |
duke@435 | 1389 | case T_LONG: |
duke@435 | 1390 | assert( sig_bt[i+1] == T_VOID, "expecting half" ); |
duke@435 | 1391 | regs[i].set_pair( int_stk_helper( i+1 ), int_stk_helper( i ) ); |
duke@435 | 1392 | break; |
duke@435 | 1393 | case T_VOID: regs[i].set_bad(); break; |
duke@435 | 1394 | default: |
duke@435 | 1395 | ShouldNotReachHere(); |
duke@435 | 1396 | } |
duke@435 | 1397 | if (regs[i].first()->is_stack()) { |
duke@435 | 1398 | int off = regs[i].first()->reg2stack(); |
duke@435 | 1399 | if (off > max_stack_slots) max_stack_slots = off; |
duke@435 | 1400 | } |
duke@435 | 1401 | if (regs[i].second()->is_stack()) { |
duke@435 | 1402 | int off = regs[i].second()->reg2stack(); |
duke@435 | 1403 | if (off > max_stack_slots) max_stack_slots = off; |
duke@435 | 1404 | } |
duke@435 | 1405 | } |
duke@435 | 1406 | #endif // _LP64 |
duke@435 | 1407 | |
duke@435 | 1408 | return round_to(max_stack_slots + 1, 2); |
duke@435 | 1409 | |
duke@435 | 1410 | } |
duke@435 | 1411 | |
duke@435 | 1412 | |
duke@435 | 1413 | // --------------------------------------------------------------------------- |
duke@435 | 1414 | void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { |
duke@435 | 1415 | switch (ret_type) { |
duke@435 | 1416 | case T_FLOAT: |
duke@435 | 1417 | __ stf(FloatRegisterImpl::S, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS); |
duke@435 | 1418 | break; |
duke@435 | 1419 | case T_DOUBLE: |
duke@435 | 1420 | __ stf(FloatRegisterImpl::D, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS); |
duke@435 | 1421 | break; |
duke@435 | 1422 | } |
duke@435 | 1423 | } |
duke@435 | 1424 | |
duke@435 | 1425 | void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) { |
duke@435 | 1426 | switch (ret_type) { |
duke@435 | 1427 | case T_FLOAT: |
duke@435 | 1428 | __ ldf(FloatRegisterImpl::S, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS, F0); |
duke@435 | 1429 | break; |
duke@435 | 1430 | case T_DOUBLE: |
duke@435 | 1431 | __ ldf(FloatRegisterImpl::D, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS, F0); |
duke@435 | 1432 | break; |
duke@435 | 1433 | } |
duke@435 | 1434 | } |
duke@435 | 1435 | |
duke@435 | 1436 | // Check and forward and pending exception. Thread is stored in |
duke@435 | 1437 | // L7_thread_cache and possibly NOT in G2_thread. Since this is a native call, there |
duke@435 | 1438 | // is no exception handler. We merely pop this frame off and throw the |
duke@435 | 1439 | // exception in the caller's frame. |
duke@435 | 1440 | static void check_forward_pending_exception(MacroAssembler *masm, Register Rex_oop) { |
duke@435 | 1441 | Label L; |
duke@435 | 1442 | __ br_null(Rex_oop, false, Assembler::pt, L); |
duke@435 | 1443 | __ delayed()->mov(L7_thread_cache, G2_thread); // restore in case we have exception |
duke@435 | 1444 | // Since this is a native call, we *know* the proper exception handler |
duke@435 | 1445 | // without calling into the VM: it's the empty function. Just pop this |
duke@435 | 1446 | // frame and then jump to forward_exception_entry; O7 will contain the |
duke@435 | 1447 | // native caller's return PC. |
duke@435 | 1448 | Address exception_entry(G3_scratch, StubRoutines::forward_exception_entry()); |
duke@435 | 1449 | __ jump_to(exception_entry); |
duke@435 | 1450 | __ delayed()->restore(); // Pop this frame off. |
duke@435 | 1451 | __ bind(L); |
duke@435 | 1452 | } |
duke@435 | 1453 | |
duke@435 | 1454 | // A simple move of integer like type |
duke@435 | 1455 | static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { |
duke@435 | 1456 | if (src.first()->is_stack()) { |
duke@435 | 1457 | if (dst.first()->is_stack()) { |
duke@435 | 1458 | // stack to stack |
duke@435 | 1459 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); |
duke@435 | 1460 | __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1461 | } else { |
duke@435 | 1462 | // stack to reg |
duke@435 | 1463 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); |
duke@435 | 1464 | } |
duke@435 | 1465 | } else if (dst.first()->is_stack()) { |
duke@435 | 1466 | // reg to stack |
duke@435 | 1467 | __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1468 | } else { |
duke@435 | 1469 | __ mov(src.first()->as_Register(), dst.first()->as_Register()); |
duke@435 | 1470 | } |
duke@435 | 1471 | } |
duke@435 | 1472 | |
duke@435 | 1473 | // On 64 bit we will store integer like items to the stack as |
duke@435 | 1474 | // 64 bits items (sparc abi) even though java would only store |
duke@435 | 1475 | // 32bits for a parameter. On 32bit it will simply be 32 bits |
duke@435 | 1476 | // So this routine will do 32->32 on 32bit and 32->64 on 64bit |
duke@435 | 1477 | static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { |
duke@435 | 1478 | if (src.first()->is_stack()) { |
duke@435 | 1479 | if (dst.first()->is_stack()) { |
duke@435 | 1480 | // stack to stack |
duke@435 | 1481 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); |
duke@435 | 1482 | __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1483 | } else { |
duke@435 | 1484 | // stack to reg |
duke@435 | 1485 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); |
duke@435 | 1486 | } |
duke@435 | 1487 | } else if (dst.first()->is_stack()) { |
duke@435 | 1488 | // reg to stack |
duke@435 | 1489 | __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1490 | } else { |
duke@435 | 1491 | __ mov(src.first()->as_Register(), dst.first()->as_Register()); |
duke@435 | 1492 | } |
duke@435 | 1493 | } |
duke@435 | 1494 | |
duke@435 | 1495 | |
duke@435 | 1496 | // An oop arg. Must pass a handle not the oop itself |
duke@435 | 1497 | static void object_move(MacroAssembler* masm, |
duke@435 | 1498 | OopMap* map, |
duke@435 | 1499 | int oop_handle_offset, |
duke@435 | 1500 | int framesize_in_slots, |
duke@435 | 1501 | VMRegPair src, |
duke@435 | 1502 | VMRegPair dst, |
duke@435 | 1503 | bool is_receiver, |
duke@435 | 1504 | int* receiver_offset) { |
duke@435 | 1505 | |
duke@435 | 1506 | // must pass a handle. First figure out the location we use as a handle |
duke@435 | 1507 | |
duke@435 | 1508 | if (src.first()->is_stack()) { |
duke@435 | 1509 | // Oop is already on the stack |
duke@435 | 1510 | Register rHandle = dst.first()->is_stack() ? L5 : dst.first()->as_Register(); |
duke@435 | 1511 | __ add(FP, reg2offset(src.first()) + STACK_BIAS, rHandle); |
duke@435 | 1512 | __ ld_ptr(rHandle, 0, L4); |
duke@435 | 1513 | #ifdef _LP64 |
duke@435 | 1514 | __ movr( Assembler::rc_z, L4, G0, rHandle ); |
duke@435 | 1515 | #else |
duke@435 | 1516 | __ tst( L4 ); |
duke@435 | 1517 | __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle ); |
duke@435 | 1518 | #endif |
duke@435 | 1519 | if (dst.first()->is_stack()) { |
duke@435 | 1520 | __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1521 | } |
duke@435 | 1522 | int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots(); |
duke@435 | 1523 | if (is_receiver) { |
duke@435 | 1524 | *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size; |
duke@435 | 1525 | } |
duke@435 | 1526 | map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots)); |
duke@435 | 1527 | } else { |
duke@435 | 1528 | // Oop is in an input register pass we must flush it to the stack |
duke@435 | 1529 | const Register rOop = src.first()->as_Register(); |
duke@435 | 1530 | const Register rHandle = L5; |
duke@435 | 1531 | int oop_slot = rOop->input_number() * VMRegImpl::slots_per_word + oop_handle_offset; |
duke@435 | 1532 | int offset = oop_slot*VMRegImpl::stack_slot_size; |
duke@435 | 1533 | Label skip; |
duke@435 | 1534 | __ st_ptr(rOop, SP, offset + STACK_BIAS); |
duke@435 | 1535 | if (is_receiver) { |
duke@435 | 1536 | *receiver_offset = oop_slot * VMRegImpl::stack_slot_size; |
duke@435 | 1537 | } |
duke@435 | 1538 | map->set_oop(VMRegImpl::stack2reg(oop_slot)); |
duke@435 | 1539 | __ add(SP, offset + STACK_BIAS, rHandle); |
duke@435 | 1540 | #ifdef _LP64 |
duke@435 | 1541 | __ movr( Assembler::rc_z, rOop, G0, rHandle ); |
duke@435 | 1542 | #else |
duke@435 | 1543 | __ tst( rOop ); |
duke@435 | 1544 | __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle ); |
duke@435 | 1545 | #endif |
duke@435 | 1546 | |
duke@435 | 1547 | if (dst.first()->is_stack()) { |
duke@435 | 1548 | __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1549 | } else { |
duke@435 | 1550 | __ mov(rHandle, dst.first()->as_Register()); |
duke@435 | 1551 | } |
duke@435 | 1552 | } |
duke@435 | 1553 | } |
duke@435 | 1554 | |
duke@435 | 1555 | // A float arg may have to do float reg int reg conversion |
duke@435 | 1556 | static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { |
duke@435 | 1557 | assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move"); |
duke@435 | 1558 | |
duke@435 | 1559 | if (src.first()->is_stack()) { |
duke@435 | 1560 | if (dst.first()->is_stack()) { |
duke@435 | 1561 | // stack to stack the easiest of the bunch |
duke@435 | 1562 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); |
duke@435 | 1563 | __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1564 | } else { |
duke@435 | 1565 | // stack to reg |
duke@435 | 1566 | if (dst.first()->is_Register()) { |
duke@435 | 1567 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); |
duke@435 | 1568 | } else { |
duke@435 | 1569 | __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister()); |
duke@435 | 1570 | } |
duke@435 | 1571 | } |
duke@435 | 1572 | } else if (dst.first()->is_stack()) { |
duke@435 | 1573 | // reg to stack |
duke@435 | 1574 | if (src.first()->is_Register()) { |
duke@435 | 1575 | __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1576 | } else { |
duke@435 | 1577 | __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1578 | } |
duke@435 | 1579 | } else { |
duke@435 | 1580 | // reg to reg |
duke@435 | 1581 | if (src.first()->is_Register()) { |
duke@435 | 1582 | if (dst.first()->is_Register()) { |
duke@435 | 1583 | // gpr -> gpr |
duke@435 | 1584 | __ mov(src.first()->as_Register(), dst.first()->as_Register()); |
duke@435 | 1585 | } else { |
duke@435 | 1586 | // gpr -> fpr |
duke@435 | 1587 | __ st(src.first()->as_Register(), FP, -4 + STACK_BIAS); |
duke@435 | 1588 | __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.first()->as_FloatRegister()); |
duke@435 | 1589 | } |
duke@435 | 1590 | } else if (dst.first()->is_Register()) { |
duke@435 | 1591 | // fpr -> gpr |
duke@435 | 1592 | __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), FP, -4 + STACK_BIAS); |
duke@435 | 1593 | __ ld(FP, -4 + STACK_BIAS, dst.first()->as_Register()); |
duke@435 | 1594 | } else { |
duke@435 | 1595 | // fpr -> fpr |
duke@435 | 1596 | // In theory these overlap but the ordering is such that this is likely a nop |
duke@435 | 1597 | if ( src.first() != dst.first()) { |
duke@435 | 1598 | __ fmov(FloatRegisterImpl::S, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister()); |
duke@435 | 1599 | } |
duke@435 | 1600 | } |
duke@435 | 1601 | } |
duke@435 | 1602 | } |
duke@435 | 1603 | |
duke@435 | 1604 | static void split_long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { |
duke@435 | 1605 | VMRegPair src_lo(src.first()); |
duke@435 | 1606 | VMRegPair src_hi(src.second()); |
duke@435 | 1607 | VMRegPair dst_lo(dst.first()); |
duke@435 | 1608 | VMRegPair dst_hi(dst.second()); |
duke@435 | 1609 | simple_move32(masm, src_lo, dst_lo); |
duke@435 | 1610 | simple_move32(masm, src_hi, dst_hi); |
duke@435 | 1611 | } |
duke@435 | 1612 | |
duke@435 | 1613 | // A long move |
duke@435 | 1614 | static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { |
duke@435 | 1615 | |
duke@435 | 1616 | // Do the simple ones here else do two int moves |
duke@435 | 1617 | if (src.is_single_phys_reg() ) { |
duke@435 | 1618 | if (dst.is_single_phys_reg()) { |
duke@435 | 1619 | __ mov(src.first()->as_Register(), dst.first()->as_Register()); |
duke@435 | 1620 | } else { |
duke@435 | 1621 | // split src into two separate registers |
duke@435 | 1622 | // Remember hi means hi address or lsw on sparc |
duke@435 | 1623 | // Move msw to lsw |
duke@435 | 1624 | if (dst.second()->is_reg()) { |
duke@435 | 1625 | // MSW -> MSW |
duke@435 | 1626 | __ srax(src.first()->as_Register(), 32, dst.first()->as_Register()); |
duke@435 | 1627 | // Now LSW -> LSW |
duke@435 | 1628 | // this will only move lo -> lo and ignore hi |
duke@435 | 1629 | VMRegPair split(dst.second()); |
duke@435 | 1630 | simple_move32(masm, src, split); |
duke@435 | 1631 | } else { |
duke@435 | 1632 | VMRegPair split(src.first(), L4->as_VMReg()); |
duke@435 | 1633 | // MSW -> MSW (lo ie. first word) |
duke@435 | 1634 | __ srax(src.first()->as_Register(), 32, L4); |
duke@435 | 1635 | split_long_move(masm, split, dst); |
duke@435 | 1636 | } |
duke@435 | 1637 | } |
duke@435 | 1638 | } else if (dst.is_single_phys_reg()) { |
duke@435 | 1639 | if (src.is_adjacent_aligned_on_stack(2)) { |
duke@435 | 1640 | __ ldd(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); |
duke@435 | 1641 | } else { |
duke@435 | 1642 | // dst is a single reg. |
duke@435 | 1643 | // Remember lo is low address not msb for stack slots |
duke@435 | 1644 | // and lo is the "real" register for registers |
duke@435 | 1645 | // src is |
duke@435 | 1646 | |
duke@435 | 1647 | VMRegPair split; |
duke@435 | 1648 | |
duke@435 | 1649 | if (src.first()->is_reg()) { |
duke@435 | 1650 | // src.lo (msw) is a reg, src.hi is stk/reg |
duke@435 | 1651 | // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> src.lo [the MSW is in the LSW of the reg] |
duke@435 | 1652 | split.set_pair(dst.first(), src.first()); |
duke@435 | 1653 | } else { |
duke@435 | 1654 | // msw is stack move to L5 |
duke@435 | 1655 | // lsw is stack move to dst.lo (real reg) |
duke@435 | 1656 | // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> L5 |
duke@435 | 1657 | split.set_pair(dst.first(), L5->as_VMReg()); |
duke@435 | 1658 | } |
duke@435 | 1659 | |
duke@435 | 1660 | // src.lo -> src.lo/L5, src.hi -> dst.lo (the real reg) |
duke@435 | 1661 | // msw -> src.lo/L5, lsw -> dst.lo |
duke@435 | 1662 | split_long_move(masm, src, split); |
duke@435 | 1663 | |
duke@435 | 1664 | // So dst now has the low order correct position the |
duke@435 | 1665 | // msw half |
duke@435 | 1666 | __ sllx(split.first()->as_Register(), 32, L5); |
duke@435 | 1667 | |
duke@435 | 1668 | const Register d = dst.first()->as_Register(); |
duke@435 | 1669 | __ or3(L5, d, d); |
duke@435 | 1670 | } |
duke@435 | 1671 | } else { |
duke@435 | 1672 | // For LP64 we can probably do better. |
duke@435 | 1673 | split_long_move(masm, src, dst); |
duke@435 | 1674 | } |
duke@435 | 1675 | } |
duke@435 | 1676 | |
duke@435 | 1677 | // A double move |
duke@435 | 1678 | static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) { |
duke@435 | 1679 | |
duke@435 | 1680 | // The painful thing here is that like long_move a VMRegPair might be |
duke@435 | 1681 | // 1: a single physical register |
duke@435 | 1682 | // 2: two physical registers (v8) |
duke@435 | 1683 | // 3: a physical reg [lo] and a stack slot [hi] (v8) |
duke@435 | 1684 | // 4: two stack slots |
duke@435 | 1685 | |
duke@435 | 1686 | // Since src is always a java calling convention we know that the src pair |
duke@435 | 1687 | // is always either all registers or all stack (and aligned?) |
duke@435 | 1688 | |
duke@435 | 1689 | // in a register [lo] and a stack slot [hi] |
duke@435 | 1690 | if (src.first()->is_stack()) { |
duke@435 | 1691 | if (dst.first()->is_stack()) { |
duke@435 | 1692 | // stack to stack the easiest of the bunch |
duke@435 | 1693 | // ought to be a way to do this where if alignment is ok we use ldd/std when possible |
duke@435 | 1694 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5); |
duke@435 | 1695 | __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4); |
duke@435 | 1696 | __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1697 | __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS); |
duke@435 | 1698 | } else { |
duke@435 | 1699 | // stack to reg |
duke@435 | 1700 | if (dst.second()->is_stack()) { |
duke@435 | 1701 | // stack -> reg, stack -> stack |
duke@435 | 1702 | __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4); |
duke@435 | 1703 | if (dst.first()->is_Register()) { |
duke@435 | 1704 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); |
duke@435 | 1705 | } else { |
duke@435 | 1706 | __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister()); |
duke@435 | 1707 | } |
duke@435 | 1708 | // This was missing. (very rare case) |
duke@435 | 1709 | __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS); |
duke@435 | 1710 | } else { |
duke@435 | 1711 | // stack -> reg |
duke@435 | 1712 | // Eventually optimize for alignment QQQ |
duke@435 | 1713 | if (dst.first()->is_Register()) { |
duke@435 | 1714 | __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register()); |
duke@435 | 1715 | __ ld(FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_Register()); |
duke@435 | 1716 | } else { |
duke@435 | 1717 | __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister()); |
duke@435 | 1718 | __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_FloatRegister()); |
duke@435 | 1719 | } |
duke@435 | 1720 | } |
duke@435 | 1721 | } |
duke@435 | 1722 | } else if (dst.first()->is_stack()) { |
duke@435 | 1723 | // reg to stack |
duke@435 | 1724 | if (src.first()->is_Register()) { |
duke@435 | 1725 | // Eventually optimize for alignment QQQ |
duke@435 | 1726 | __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1727 | if (src.second()->is_stack()) { |
duke@435 | 1728 | __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4); |
duke@435 | 1729 | __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS); |
duke@435 | 1730 | } else { |
duke@435 | 1731 | __ st(src.second()->as_Register(), SP, reg2offset(dst.second()) + STACK_BIAS); |
duke@435 | 1732 | } |
duke@435 | 1733 | } else { |
duke@435 | 1734 | // fpr to stack |
duke@435 | 1735 | if (src.second()->is_stack()) { |
duke@435 | 1736 | ShouldNotReachHere(); |
duke@435 | 1737 | } else { |
duke@435 | 1738 | // Is the stack aligned? |
duke@435 | 1739 | if (reg2offset(dst.first()) & 0x7) { |
duke@435 | 1740 | // No do as pairs |
duke@435 | 1741 | __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1742 | __ stf(FloatRegisterImpl::S, src.second()->as_FloatRegister(), SP, reg2offset(dst.second()) + STACK_BIAS); |
duke@435 | 1743 | } else { |
duke@435 | 1744 | __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS); |
duke@435 | 1745 | } |
duke@435 | 1746 | } |
duke@435 | 1747 | } |
duke@435 | 1748 | } else { |
duke@435 | 1749 | // reg to reg |
duke@435 | 1750 | if (src.first()->is_Register()) { |
duke@435 | 1751 | if (dst.first()->is_Register()) { |
duke@435 | 1752 | // gpr -> gpr |
duke@435 | 1753 | __ mov(src.first()->as_Register(), dst.first()->as_Register()); |
duke@435 | 1754 | __ mov(src.second()->as_Register(), dst.second()->as_Register()); |
duke@435 | 1755 | } else { |
duke@435 | 1756 | // gpr -> fpr |
duke@435 | 1757 | // ought to be able to do a single store |
duke@435 | 1758 | __ stx(src.first()->as_Register(), FP, -8 + STACK_BIAS); |
duke@435 | 1759 | __ stx(src.second()->as_Register(), FP, -4 + STACK_BIAS); |
duke@435 | 1760 | // ought to be able to do a single load |
duke@435 | 1761 | __ ldf(FloatRegisterImpl::S, FP, -8 + STACK_BIAS, dst.first()->as_FloatRegister()); |
duke@435 | 1762 | __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.second()->as_FloatRegister()); |
duke@435 | 1763 | } |
duke@435 | 1764 | } else if (dst.first()->is_Register()) { |
duke@435 | 1765 | // fpr -> gpr |
duke@435 | 1766 | // ought to be able to do a single store |
duke@435 | 1767 | __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), FP, -8 + STACK_BIAS); |
duke@435 | 1768 | // ought to be able to do a single load |
duke@435 | 1769 | // REMEMBER first() is low address not LSB |
duke@435 | 1770 | __ ld(FP, -8 + STACK_BIAS, dst.first()->as_Register()); |
duke@435 | 1771 | if (dst.second()->is_Register()) { |
duke@435 | 1772 | __ ld(FP, -4 + STACK_BIAS, dst.second()->as_Register()); |
duke@435 | 1773 | } else { |
duke@435 | 1774 | __ ld(FP, -4 + STACK_BIAS, L4); |
duke@435 | 1775 | __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS); |
duke@435 | 1776 | } |
duke@435 | 1777 | } else { |
duke@435 | 1778 | // fpr -> fpr |
duke@435 | 1779 | // In theory these overlap but the ordering is such that this is likely a nop |
duke@435 | 1780 | if ( src.first() != dst.first()) { |
duke@435 | 1781 | __ fmov(FloatRegisterImpl::D, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister()); |
duke@435 | 1782 | } |
duke@435 | 1783 | } |
duke@435 | 1784 | } |
duke@435 | 1785 | } |
duke@435 | 1786 | |
duke@435 | 1787 | // Creates an inner frame if one hasn't already been created, and |
duke@435 | 1788 | // saves a copy of the thread in L7_thread_cache |
duke@435 | 1789 | static void create_inner_frame(MacroAssembler* masm, bool* already_created) { |
duke@435 | 1790 | if (!*already_created) { |
duke@435 | 1791 | __ save_frame(0); |
duke@435 | 1792 | // Save thread in L7 (INNER FRAME); it crosses a bunch of VM calls below |
duke@435 | 1793 | // Don't use save_thread because it smashes G2 and we merely want to save a |
duke@435 | 1794 | // copy |
duke@435 | 1795 | __ mov(G2_thread, L7_thread_cache); |
duke@435 | 1796 | *already_created = true; |
duke@435 | 1797 | } |
duke@435 | 1798 | } |
duke@435 | 1799 | |
duke@435 | 1800 | // --------------------------------------------------------------------------- |
duke@435 | 1801 | // Generate a native wrapper for a given method. The method takes arguments |
duke@435 | 1802 | // in the Java compiled code convention, marshals them to the native |
duke@435 | 1803 | // convention (handlizes oops, etc), transitions to native, makes the call, |
duke@435 | 1804 | // returns to java state (possibly blocking), unhandlizes any result and |
duke@435 | 1805 | // returns. |
duke@435 | 1806 | nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler* masm, |
duke@435 | 1807 | methodHandle method, |
duke@435 | 1808 | int total_in_args, |
duke@435 | 1809 | int comp_args_on_stack, // in VMRegStackSlots |
duke@435 | 1810 | BasicType *in_sig_bt, |
duke@435 | 1811 | VMRegPair *in_regs, |
duke@435 | 1812 | BasicType ret_type) { |
duke@435 | 1813 | |
duke@435 | 1814 | |
duke@435 | 1815 | // Native nmethod wrappers never take possesion of the oop arguments. |
duke@435 | 1816 | // So the caller will gc the arguments. The only thing we need an |
duke@435 | 1817 | // oopMap for is if the call is static |
duke@435 | 1818 | // |
duke@435 | 1819 | // An OopMap for lock (and class if static), and one for the VM call itself |
duke@435 | 1820 | OopMapSet *oop_maps = new OopMapSet(); |
duke@435 | 1821 | intptr_t start = (intptr_t)__ pc(); |
duke@435 | 1822 | |
duke@435 | 1823 | // First thing make an ic check to see if we should even be here |
duke@435 | 1824 | { |
duke@435 | 1825 | Label L; |
duke@435 | 1826 | const Register temp_reg = G3_scratch; |
duke@435 | 1827 | Address ic_miss(temp_reg, SharedRuntime::get_ic_miss_stub()); |
duke@435 | 1828 | __ verify_oop(O0); |
coleenp@548 | 1829 | __ load_klass(O0, temp_reg); |
duke@435 | 1830 | __ cmp(temp_reg, G5_inline_cache_reg); |
duke@435 | 1831 | __ brx(Assembler::equal, true, Assembler::pt, L); |
duke@435 | 1832 | __ delayed()->nop(); |
duke@435 | 1833 | |
duke@435 | 1834 | __ jump_to(ic_miss, 0); |
duke@435 | 1835 | __ delayed()->nop(); |
duke@435 | 1836 | __ align(CodeEntryAlignment); |
duke@435 | 1837 | __ bind(L); |
duke@435 | 1838 | } |
duke@435 | 1839 | |
duke@435 | 1840 | int vep_offset = ((intptr_t)__ pc()) - start; |
duke@435 | 1841 | |
duke@435 | 1842 | #ifdef COMPILER1 |
duke@435 | 1843 | if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) { |
duke@435 | 1844 | // Object.hashCode can pull the hashCode from the header word |
duke@435 | 1845 | // instead of doing a full VM transition once it's been computed. |
duke@435 | 1846 | // Since hashCode is usually polymorphic at call sites we can't do |
duke@435 | 1847 | // this optimization at the call site without a lot of work. |
duke@435 | 1848 | Label slowCase; |
duke@435 | 1849 | Register receiver = O0; |
duke@435 | 1850 | Register result = O0; |
duke@435 | 1851 | Register header = G3_scratch; |
duke@435 | 1852 | Register hash = G3_scratch; // overwrite header value with hash value |
duke@435 | 1853 | Register mask = G1; // to get hash field from header |
duke@435 | 1854 | |
duke@435 | 1855 | // Read the header and build a mask to get its hash field. Give up if the object is not unlocked. |
duke@435 | 1856 | // We depend on hash_mask being at most 32 bits and avoid the use of |
duke@435 | 1857 | // hash_mask_in_place because it could be larger than 32 bits in a 64-bit |
duke@435 | 1858 | // vm: see markOop.hpp. |
duke@435 | 1859 | __ ld_ptr(receiver, oopDesc::mark_offset_in_bytes(), header); |
duke@435 | 1860 | __ sethi(markOopDesc::hash_mask, mask); |
duke@435 | 1861 | __ btst(markOopDesc::unlocked_value, header); |
duke@435 | 1862 | __ br(Assembler::zero, false, Assembler::pn, slowCase); |
duke@435 | 1863 | if (UseBiasedLocking) { |
duke@435 | 1864 | // Check if biased and fall through to runtime if so |
duke@435 | 1865 | __ delayed()->nop(); |
duke@435 | 1866 | __ btst(markOopDesc::biased_lock_bit_in_place, header); |
duke@435 | 1867 | __ br(Assembler::notZero, false, Assembler::pn, slowCase); |
duke@435 | 1868 | } |
duke@435 | 1869 | __ delayed()->or3(mask, markOopDesc::hash_mask & 0x3ff, mask); |
duke@435 | 1870 | |
duke@435 | 1871 | // Check for a valid (non-zero) hash code and get its value. |
duke@435 | 1872 | #ifdef _LP64 |
duke@435 | 1873 | __ srlx(header, markOopDesc::hash_shift, hash); |
duke@435 | 1874 | #else |
duke@435 | 1875 | __ srl(header, markOopDesc::hash_shift, hash); |
duke@435 | 1876 | #endif |
duke@435 | 1877 | __ andcc(hash, mask, hash); |
duke@435 | 1878 | __ br(Assembler::equal, false, Assembler::pn, slowCase); |
duke@435 | 1879 | __ delayed()->nop(); |
duke@435 | 1880 | |
duke@435 | 1881 | // leaf return. |
duke@435 | 1882 | __ retl(); |
duke@435 | 1883 | __ delayed()->mov(hash, result); |
duke@435 | 1884 | __ bind(slowCase); |
duke@435 | 1885 | } |
duke@435 | 1886 | #endif // COMPILER1 |
duke@435 | 1887 | |
duke@435 | 1888 | |
duke@435 | 1889 | // We have received a description of where all the java arg are located |
duke@435 | 1890 | // on entry to the wrapper. We need to convert these args to where |
duke@435 | 1891 | // the jni function will expect them. To figure out where they go |
duke@435 | 1892 | // we convert the java signature to a C signature by inserting |
duke@435 | 1893 | // the hidden arguments as arg[0] and possibly arg[1] (static method) |
duke@435 | 1894 | |
duke@435 | 1895 | int total_c_args = total_in_args + 1; |
duke@435 | 1896 | if (method->is_static()) { |
duke@435 | 1897 | total_c_args++; |
duke@435 | 1898 | } |
duke@435 | 1899 | |
duke@435 | 1900 | BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args); |
duke@435 | 1901 | VMRegPair * out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args); |
duke@435 | 1902 | |
duke@435 | 1903 | int argc = 0; |
duke@435 | 1904 | out_sig_bt[argc++] = T_ADDRESS; |
duke@435 | 1905 | if (method->is_static()) { |
duke@435 | 1906 | out_sig_bt[argc++] = T_OBJECT; |
duke@435 | 1907 | } |
duke@435 | 1908 | |
duke@435 | 1909 | for (int i = 0; i < total_in_args ; i++ ) { |
duke@435 | 1910 | out_sig_bt[argc++] = in_sig_bt[i]; |
duke@435 | 1911 | } |
duke@435 | 1912 | |
duke@435 | 1913 | // Now figure out where the args must be stored and how much stack space |
duke@435 | 1914 | // they require (neglecting out_preserve_stack_slots but space for storing |
duke@435 | 1915 | // the 1st six register arguments). It's weird see int_stk_helper. |
duke@435 | 1916 | // |
duke@435 | 1917 | int out_arg_slots; |
duke@435 | 1918 | out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args); |
duke@435 | 1919 | |
duke@435 | 1920 | // Compute framesize for the wrapper. We need to handlize all oops in |
duke@435 | 1921 | // registers. We must create space for them here that is disjoint from |
duke@435 | 1922 | // the windowed save area because we have no control over when we might |
duke@435 | 1923 | // flush the window again and overwrite values that gc has since modified. |
duke@435 | 1924 | // (The live window race) |
duke@435 | 1925 | // |
duke@435 | 1926 | // We always just allocate 6 word for storing down these object. This allow |
duke@435 | 1927 | // us to simply record the base and use the Ireg number to decide which |
duke@435 | 1928 | // slot to use. (Note that the reg number is the inbound number not the |
duke@435 | 1929 | // outbound number). |
duke@435 | 1930 | // We must shuffle args to match the native convention, and include var-args space. |
duke@435 | 1931 | |
duke@435 | 1932 | // Calculate the total number of stack slots we will need. |
duke@435 | 1933 | |
duke@435 | 1934 | // First count the abi requirement plus all of the outgoing args |
duke@435 | 1935 | int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots; |
duke@435 | 1936 | |
duke@435 | 1937 | // Now the space for the inbound oop handle area |
duke@435 | 1938 | |
duke@435 | 1939 | int oop_handle_offset = stack_slots; |
duke@435 | 1940 | stack_slots += 6*VMRegImpl::slots_per_word; |
duke@435 | 1941 | |
duke@435 | 1942 | // Now any space we need for handlizing a klass if static method |
duke@435 | 1943 | |
duke@435 | 1944 | int oop_temp_slot_offset = 0; |
duke@435 | 1945 | int klass_slot_offset = 0; |
duke@435 | 1946 | int klass_offset = -1; |
duke@435 | 1947 | int lock_slot_offset = 0; |
duke@435 | 1948 | bool is_static = false; |
duke@435 | 1949 | |
duke@435 | 1950 | if (method->is_static()) { |
duke@435 | 1951 | klass_slot_offset = stack_slots; |
duke@435 | 1952 | stack_slots += VMRegImpl::slots_per_word; |
duke@435 | 1953 | klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size; |
duke@435 | 1954 | is_static = true; |
duke@435 | 1955 | } |
duke@435 | 1956 | |
duke@435 | 1957 | // Plus a lock if needed |
duke@435 | 1958 | |
duke@435 | 1959 | if (method->is_synchronized()) { |
duke@435 | 1960 | lock_slot_offset = stack_slots; |
duke@435 | 1961 | stack_slots += VMRegImpl::slots_per_word; |
duke@435 | 1962 | } |
duke@435 | 1963 | |
duke@435 | 1964 | // Now a place to save return value or as a temporary for any gpr -> fpr moves |
duke@435 | 1965 | stack_slots += 2; |
duke@435 | 1966 | |
duke@435 | 1967 | // Ok The space we have allocated will look like: |
duke@435 | 1968 | // |
duke@435 | 1969 | // |
duke@435 | 1970 | // FP-> | | |
duke@435 | 1971 | // |---------------------| |
duke@435 | 1972 | // | 2 slots for moves | |
duke@435 | 1973 | // |---------------------| |
duke@435 | 1974 | // | lock box (if sync) | |
duke@435 | 1975 | // |---------------------| <- lock_slot_offset |
duke@435 | 1976 | // | klass (if static) | |
duke@435 | 1977 | // |---------------------| <- klass_slot_offset |
duke@435 | 1978 | // | oopHandle area | |
duke@435 | 1979 | // |---------------------| <- oop_handle_offset |
duke@435 | 1980 | // | outbound memory | |
duke@435 | 1981 | // | based arguments | |
duke@435 | 1982 | // | | |
duke@435 | 1983 | // |---------------------| |
duke@435 | 1984 | // | vararg area | |
duke@435 | 1985 | // |---------------------| |
duke@435 | 1986 | // | | |
duke@435 | 1987 | // SP-> | out_preserved_slots | |
duke@435 | 1988 | // |
duke@435 | 1989 | // |
duke@435 | 1990 | |
duke@435 | 1991 | |
duke@435 | 1992 | // Now compute actual number of stack words we need rounding to make |
duke@435 | 1993 | // stack properly aligned. |
duke@435 | 1994 | stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word); |
duke@435 | 1995 | |
duke@435 | 1996 | int stack_size = stack_slots * VMRegImpl::stack_slot_size; |
duke@435 | 1997 | |
duke@435 | 1998 | // Generate stack overflow check before creating frame |
duke@435 | 1999 | __ generate_stack_overflow_check(stack_size); |
duke@435 | 2000 | |
duke@435 | 2001 | // Generate a new frame for the wrapper. |
duke@435 | 2002 | __ save(SP, -stack_size, SP); |
duke@435 | 2003 | |
duke@435 | 2004 | int frame_complete = ((intptr_t)__ pc()) - start; |
duke@435 | 2005 | |
duke@435 | 2006 | __ verify_thread(); |
duke@435 | 2007 | |
duke@435 | 2008 | |
duke@435 | 2009 | // |
duke@435 | 2010 | // We immediately shuffle the arguments so that any vm call we have to |
duke@435 | 2011 | // make from here on out (sync slow path, jvmti, etc.) we will have |
duke@435 | 2012 | // captured the oops from our caller and have a valid oopMap for |
duke@435 | 2013 | // them. |
duke@435 | 2014 | |
duke@435 | 2015 | // ----------------- |
duke@435 | 2016 | // The Grand Shuffle |
duke@435 | 2017 | // |
duke@435 | 2018 | // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv* |
duke@435 | 2019 | // (derived from JavaThread* which is in L7_thread_cache) and, if static, |
duke@435 | 2020 | // the class mirror instead of a receiver. This pretty much guarantees that |
duke@435 | 2021 | // register layout will not match. We ignore these extra arguments during |
duke@435 | 2022 | // the shuffle. The shuffle is described by the two calling convention |
duke@435 | 2023 | // vectors we have in our possession. We simply walk the java vector to |
duke@435 | 2024 | // get the source locations and the c vector to get the destinations. |
duke@435 | 2025 | // Because we have a new window and the argument registers are completely |
duke@435 | 2026 | // disjoint ( I0 -> O1, I1 -> O2, ...) we have nothing to worry about |
duke@435 | 2027 | // here. |
duke@435 | 2028 | |
duke@435 | 2029 | // This is a trick. We double the stack slots so we can claim |
duke@435 | 2030 | // the oops in the caller's frame. Since we are sure to have |
duke@435 | 2031 | // more args than the caller doubling is enough to make |
duke@435 | 2032 | // sure we can capture all the incoming oop args from the |
duke@435 | 2033 | // caller. |
duke@435 | 2034 | // |
duke@435 | 2035 | OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/); |
duke@435 | 2036 | int c_arg = total_c_args - 1; |
duke@435 | 2037 | // Record sp-based slot for receiver on stack for non-static methods |
duke@435 | 2038 | int receiver_offset = -1; |
duke@435 | 2039 | |
duke@435 | 2040 | // We move the arguments backward because the floating point registers |
duke@435 | 2041 | // destination will always be to a register with a greater or equal register |
duke@435 | 2042 | // number or the stack. |
duke@435 | 2043 | |
duke@435 | 2044 | #ifdef ASSERT |
duke@435 | 2045 | bool reg_destroyed[RegisterImpl::number_of_registers]; |
duke@435 | 2046 | bool freg_destroyed[FloatRegisterImpl::number_of_registers]; |
duke@435 | 2047 | for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) { |
duke@435 | 2048 | reg_destroyed[r] = false; |
duke@435 | 2049 | } |
duke@435 | 2050 | for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) { |
duke@435 | 2051 | freg_destroyed[f] = false; |
duke@435 | 2052 | } |
duke@435 | 2053 | |
duke@435 | 2054 | #endif /* ASSERT */ |
duke@435 | 2055 | |
duke@435 | 2056 | for ( int i = total_in_args - 1; i >= 0 ; i--, c_arg-- ) { |
duke@435 | 2057 | |
duke@435 | 2058 | #ifdef ASSERT |
duke@435 | 2059 | if (in_regs[i].first()->is_Register()) { |
duke@435 | 2060 | assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "ack!"); |
duke@435 | 2061 | } else if (in_regs[i].first()->is_FloatRegister()) { |
duke@435 | 2062 | assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)], "ack!"); |
duke@435 | 2063 | } |
duke@435 | 2064 | if (out_regs[c_arg].first()->is_Register()) { |
duke@435 | 2065 | reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true; |
duke@435 | 2066 | } else if (out_regs[c_arg].first()->is_FloatRegister()) { |
duke@435 | 2067 | freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)] = true; |
duke@435 | 2068 | } |
duke@435 | 2069 | #endif /* ASSERT */ |
duke@435 | 2070 | |
duke@435 | 2071 | switch (in_sig_bt[i]) { |
duke@435 | 2072 | case T_ARRAY: |
duke@435 | 2073 | case T_OBJECT: |
duke@435 | 2074 | object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg], |
duke@435 | 2075 | ((i == 0) && (!is_static)), |
duke@435 | 2076 | &receiver_offset); |
duke@435 | 2077 | break; |
duke@435 | 2078 | case T_VOID: |
duke@435 | 2079 | break; |
duke@435 | 2080 | |
duke@435 | 2081 | case T_FLOAT: |
duke@435 | 2082 | float_move(masm, in_regs[i], out_regs[c_arg]); |
duke@435 | 2083 | break; |
duke@435 | 2084 | |
duke@435 | 2085 | case T_DOUBLE: |
duke@435 | 2086 | assert( i + 1 < total_in_args && |
duke@435 | 2087 | in_sig_bt[i + 1] == T_VOID && |
duke@435 | 2088 | out_sig_bt[c_arg+1] == T_VOID, "bad arg list"); |
duke@435 | 2089 | double_move(masm, in_regs[i], out_regs[c_arg]); |
duke@435 | 2090 | break; |
duke@435 | 2091 | |
duke@435 | 2092 | case T_LONG : |
duke@435 | 2093 | long_move(masm, in_regs[i], out_regs[c_arg]); |
duke@435 | 2094 | break; |
duke@435 | 2095 | |
duke@435 | 2096 | case T_ADDRESS: assert(false, "found T_ADDRESS in java args"); |
duke@435 | 2097 | |
duke@435 | 2098 | default: |
duke@435 | 2099 | move32_64(masm, in_regs[i], out_regs[c_arg]); |
duke@435 | 2100 | } |
duke@435 | 2101 | } |
duke@435 | 2102 | |
duke@435 | 2103 | // Pre-load a static method's oop into O1. Used both by locking code and |
duke@435 | 2104 | // the normal JNI call code. |
duke@435 | 2105 | if (method->is_static()) { |
duke@435 | 2106 | __ set_oop_constant(JNIHandles::make_local(Klass::cast(method->method_holder())->java_mirror()), O1); |
duke@435 | 2107 | |
duke@435 | 2108 | // Now handlize the static class mirror in O1. It's known not-null. |
duke@435 | 2109 | __ st_ptr(O1, SP, klass_offset + STACK_BIAS); |
duke@435 | 2110 | map->set_oop(VMRegImpl::stack2reg(klass_slot_offset)); |
duke@435 | 2111 | __ add(SP, klass_offset + STACK_BIAS, O1); |
duke@435 | 2112 | } |
duke@435 | 2113 | |
duke@435 | 2114 | |
duke@435 | 2115 | const Register L6_handle = L6; |
duke@435 | 2116 | |
duke@435 | 2117 | if (method->is_synchronized()) { |
duke@435 | 2118 | __ mov(O1, L6_handle); |
duke@435 | 2119 | } |
duke@435 | 2120 | |
duke@435 | 2121 | // We have all of the arguments setup at this point. We MUST NOT touch any Oregs |
duke@435 | 2122 | // except O6/O7. So if we must call out we must push a new frame. We immediately |
duke@435 | 2123 | // push a new frame and flush the windows. |
duke@435 | 2124 | |
duke@435 | 2125 | #ifdef _LP64 |
duke@435 | 2126 | intptr_t thepc = (intptr_t) __ pc(); |
duke@435 | 2127 | { |
duke@435 | 2128 | address here = __ pc(); |
duke@435 | 2129 | // Call the next instruction |
duke@435 | 2130 | __ call(here + 8, relocInfo::none); |
duke@435 | 2131 | __ delayed()->nop(); |
duke@435 | 2132 | } |
duke@435 | 2133 | #else |
duke@435 | 2134 | intptr_t thepc = __ load_pc_address(O7, 0); |
duke@435 | 2135 | #endif /* _LP64 */ |
duke@435 | 2136 | |
duke@435 | 2137 | // We use the same pc/oopMap repeatedly when we call out |
duke@435 | 2138 | oop_maps->add_gc_map(thepc - start, map); |
duke@435 | 2139 | |
duke@435 | 2140 | // O7 now has the pc loaded that we will use when we finally call to native. |
duke@435 | 2141 | |
duke@435 | 2142 | // Save thread in L7; it crosses a bunch of VM calls below |
duke@435 | 2143 | // Don't use save_thread because it smashes G2 and we merely |
duke@435 | 2144 | // want to save a copy |
duke@435 | 2145 | __ mov(G2_thread, L7_thread_cache); |
duke@435 | 2146 | |
duke@435 | 2147 | |
duke@435 | 2148 | // If we create an inner frame once is plenty |
duke@435 | 2149 | // when we create it we must also save G2_thread |
duke@435 | 2150 | bool inner_frame_created = false; |
duke@435 | 2151 | |
duke@435 | 2152 | // dtrace method entry support |
duke@435 | 2153 | { |
duke@435 | 2154 | SkipIfEqual skip_if( |
duke@435 | 2155 | masm, G3_scratch, &DTraceMethodProbes, Assembler::zero); |
duke@435 | 2156 | // create inner frame |
duke@435 | 2157 | __ save_frame(0); |
duke@435 | 2158 | __ mov(G2_thread, L7_thread_cache); |
duke@435 | 2159 | __ set_oop_constant(JNIHandles::make_local(method()), O1); |
duke@435 | 2160 | __ call_VM_leaf(L7_thread_cache, |
duke@435 | 2161 | CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), |
duke@435 | 2162 | G2_thread, O1); |
duke@435 | 2163 | __ restore(); |
duke@435 | 2164 | } |
duke@435 | 2165 | |
duke@435 | 2166 | // We are in the jni frame unless saved_frame is true in which case |
duke@435 | 2167 | // we are in one frame deeper (the "inner" frame). If we are in the |
duke@435 | 2168 | // "inner" frames the args are in the Iregs and if the jni frame then |
duke@435 | 2169 | // they are in the Oregs. |
duke@435 | 2170 | // If we ever need to go to the VM (for locking, jvmti) then |
duke@435 | 2171 | // we will always be in the "inner" frame. |
duke@435 | 2172 | |
duke@435 | 2173 | // Lock a synchronized method |
duke@435 | 2174 | int lock_offset = -1; // Set if locked |
duke@435 | 2175 | if (method->is_synchronized()) { |
duke@435 | 2176 | Register Roop = O1; |
duke@435 | 2177 | const Register L3_box = L3; |
duke@435 | 2178 | |
duke@435 | 2179 | create_inner_frame(masm, &inner_frame_created); |
duke@435 | 2180 | |
duke@435 | 2181 | __ ld_ptr(I1, 0, O1); |
duke@435 | 2182 | Label done; |
duke@435 | 2183 | |
duke@435 | 2184 | lock_offset = (lock_slot_offset * VMRegImpl::stack_slot_size); |
duke@435 | 2185 | __ add(FP, lock_offset+STACK_BIAS, L3_box); |
duke@435 | 2186 | #ifdef ASSERT |
duke@435 | 2187 | if (UseBiasedLocking) { |
duke@435 | 2188 | // making the box point to itself will make it clear it went unused |
duke@435 | 2189 | // but also be obviously invalid |
duke@435 | 2190 | __ st_ptr(L3_box, L3_box, 0); |
duke@435 | 2191 | } |
duke@435 | 2192 | #endif // ASSERT |
duke@435 | 2193 | // |
duke@435 | 2194 | // Compiler_lock_object (Roop, Rmark, Rbox, Rscratch) -- kills Rmark, Rbox, Rscratch |
duke@435 | 2195 | // |
duke@435 | 2196 | __ compiler_lock_object(Roop, L1, L3_box, L2); |
duke@435 | 2197 | __ br(Assembler::equal, false, Assembler::pt, done); |
duke@435 | 2198 | __ delayed() -> add(FP, lock_offset+STACK_BIAS, L3_box); |
duke@435 | 2199 | |
duke@435 | 2200 | |
duke@435 | 2201 | // None of the above fast optimizations worked so we have to get into the |
duke@435 | 2202 | // slow case of monitor enter. Inline a special case of call_VM that |
duke@435 | 2203 | // disallows any pending_exception. |
duke@435 | 2204 | __ mov(Roop, O0); // Need oop in O0 |
duke@435 | 2205 | __ mov(L3_box, O1); |
duke@435 | 2206 | |
duke@435 | 2207 | // Record last_Java_sp, in case the VM code releases the JVM lock. |
duke@435 | 2208 | |
duke@435 | 2209 | __ set_last_Java_frame(FP, I7); |
duke@435 | 2210 | |
duke@435 | 2211 | // do the call |
duke@435 | 2212 | __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), relocInfo::runtime_call_type); |
duke@435 | 2213 | __ delayed()->mov(L7_thread_cache, O2); |
duke@435 | 2214 | |
duke@435 | 2215 | __ restore_thread(L7_thread_cache); // restore G2_thread |
duke@435 | 2216 | __ reset_last_Java_frame(); |
duke@435 | 2217 | |
duke@435 | 2218 | #ifdef ASSERT |
duke@435 | 2219 | { Label L; |
duke@435 | 2220 | __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0); |
duke@435 | 2221 | __ br_null(O0, false, Assembler::pt, L); |
duke@435 | 2222 | __ delayed()->nop(); |
duke@435 | 2223 | __ stop("no pending exception allowed on exit from IR::monitorenter"); |
duke@435 | 2224 | __ bind(L); |
duke@435 | 2225 | } |
duke@435 | 2226 | #endif |
duke@435 | 2227 | __ bind(done); |
duke@435 | 2228 | } |
duke@435 | 2229 | |
duke@435 | 2230 | |
duke@435 | 2231 | // Finally just about ready to make the JNI call |
duke@435 | 2232 | |
duke@435 | 2233 | __ flush_windows(); |
duke@435 | 2234 | if (inner_frame_created) { |
duke@435 | 2235 | __ restore(); |
duke@435 | 2236 | } else { |
duke@435 | 2237 | // Store only what we need from this frame |
duke@435 | 2238 | // QQQ I think that non-v9 (like we care) we don't need these saves |
duke@435 | 2239 | // either as the flush traps and the current window goes too. |
duke@435 | 2240 | __ st_ptr(FP, SP, FP->sp_offset_in_saved_window()*wordSize + STACK_BIAS); |
duke@435 | 2241 | __ st_ptr(I7, SP, I7->sp_offset_in_saved_window()*wordSize + STACK_BIAS); |
duke@435 | 2242 | } |
duke@435 | 2243 | |
duke@435 | 2244 | // get JNIEnv* which is first argument to native |
duke@435 | 2245 | |
duke@435 | 2246 | __ add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0); |
duke@435 | 2247 | |
duke@435 | 2248 | // Use that pc we placed in O7 a while back as the current frame anchor |
duke@435 | 2249 | |
duke@435 | 2250 | __ set_last_Java_frame(SP, O7); |
duke@435 | 2251 | |
duke@435 | 2252 | // Transition from _thread_in_Java to _thread_in_native. |
duke@435 | 2253 | __ set(_thread_in_native, G3_scratch); |
duke@435 | 2254 | __ st(G3_scratch, G2_thread, in_bytes(JavaThread::thread_state_offset())); |
duke@435 | 2255 | |
duke@435 | 2256 | // We flushed the windows ages ago now mark them as flushed |
duke@435 | 2257 | |
duke@435 | 2258 | // mark windows as flushed |
duke@435 | 2259 | __ set(JavaFrameAnchor::flushed, G3_scratch); |
duke@435 | 2260 | |
duke@435 | 2261 | Address flags(G2_thread, |
duke@435 | 2262 | 0, |
duke@435 | 2263 | in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset())); |
duke@435 | 2264 | |
duke@435 | 2265 | #ifdef _LP64 |
duke@435 | 2266 | Address dest(O7, method->native_function()); |
duke@435 | 2267 | __ relocate(relocInfo::runtime_call_type); |
duke@435 | 2268 | __ jumpl_to(dest, O7); |
duke@435 | 2269 | #else |
duke@435 | 2270 | __ call(method->native_function(), relocInfo::runtime_call_type); |
duke@435 | 2271 | #endif |
duke@435 | 2272 | __ delayed()->st(G3_scratch, flags); |
duke@435 | 2273 | |
duke@435 | 2274 | __ restore_thread(L7_thread_cache); // restore G2_thread |
duke@435 | 2275 | |
duke@435 | 2276 | // Unpack native results. For int-types, we do any needed sign-extension |
duke@435 | 2277 | // and move things into I0. The return value there will survive any VM |
duke@435 | 2278 | // calls for blocking or unlocking. An FP or OOP result (handle) is done |
duke@435 | 2279 | // specially in the slow-path code. |
duke@435 | 2280 | switch (ret_type) { |
duke@435 | 2281 | case T_VOID: break; // Nothing to do! |
duke@435 | 2282 | case T_FLOAT: break; // Got it where we want it (unless slow-path) |
duke@435 | 2283 | case T_DOUBLE: break; // Got it where we want it (unless slow-path) |
duke@435 | 2284 | // In 64 bits build result is in O0, in O0, O1 in 32bit build |
duke@435 | 2285 | case T_LONG: |
duke@435 | 2286 | #ifndef _LP64 |
duke@435 | 2287 | __ mov(O1, I1); |
duke@435 | 2288 | #endif |
duke@435 | 2289 | // Fall thru |
duke@435 | 2290 | case T_OBJECT: // Really a handle |
duke@435 | 2291 | case T_ARRAY: |
duke@435 | 2292 | case T_INT: |
duke@435 | 2293 | __ mov(O0, I0); |
duke@435 | 2294 | break; |
duke@435 | 2295 | case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, I0); break; // !0 => true; 0 => false |
duke@435 | 2296 | case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, I0); break; |
duke@435 | 2297 | case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, I0); break; // cannot use and3, 0xFFFF too big as immediate value! |
duke@435 | 2298 | case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, I0); break; |
duke@435 | 2299 | break; // Cannot de-handlize until after reclaiming jvm_lock |
duke@435 | 2300 | default: |
duke@435 | 2301 | ShouldNotReachHere(); |
duke@435 | 2302 | } |
duke@435 | 2303 | |
duke@435 | 2304 | // must we block? |
duke@435 | 2305 | |
duke@435 | 2306 | // Block, if necessary, before resuming in _thread_in_Java state. |
duke@435 | 2307 | // In order for GC to work, don't clear the last_Java_sp until after blocking. |
duke@435 | 2308 | { Label no_block; |
duke@435 | 2309 | Address sync_state(G3_scratch, SafepointSynchronize::address_of_state()); |
duke@435 | 2310 | |
duke@435 | 2311 | // Switch thread to "native transition" state before reading the synchronization state. |
duke@435 | 2312 | // This additional state is necessary because reading and testing the synchronization |
duke@435 | 2313 | // state is not atomic w.r.t. GC, as this scenario demonstrates: |
duke@435 | 2314 | // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted. |
duke@435 | 2315 | // VM thread changes sync state to synchronizing and suspends threads for GC. |
duke@435 | 2316 | // Thread A is resumed to finish this native method, but doesn't block here since it |
duke@435 | 2317 | // didn't see any synchronization is progress, and escapes. |
duke@435 | 2318 | __ set(_thread_in_native_trans, G3_scratch); |
duke@435 | 2319 | __ st(G3_scratch, G2_thread, in_bytes(JavaThread::thread_state_offset())); |
duke@435 | 2320 | if(os::is_MP()) { |
duke@435 | 2321 | if (UseMembar) { |
duke@435 | 2322 | // Force this write out before the read below |
duke@435 | 2323 | __ membar(Assembler::StoreLoad); |
duke@435 | 2324 | } else { |
duke@435 | 2325 | // Write serialization page so VM thread can do a pseudo remote membar. |
duke@435 | 2326 | // We use the current thread pointer to calculate a thread specific |
duke@435 | 2327 | // offset to write to within the page. This minimizes bus traffic |
duke@435 | 2328 | // due to cache line collision. |
duke@435 | 2329 | __ serialize_memory(G2_thread, G1_scratch, G3_scratch); |
duke@435 | 2330 | } |
duke@435 | 2331 | } |
duke@435 | 2332 | __ load_contents(sync_state, G3_scratch); |
duke@435 | 2333 | __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized); |
duke@435 | 2334 | |
duke@435 | 2335 | Label L; |
duke@435 | 2336 | Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset())); |
duke@435 | 2337 | __ br(Assembler::notEqual, false, Assembler::pn, L); |
duke@435 | 2338 | __ delayed()-> |
duke@435 | 2339 | ld(suspend_state, G3_scratch); |
duke@435 | 2340 | __ cmp(G3_scratch, 0); |
duke@435 | 2341 | __ br(Assembler::equal, false, Assembler::pt, no_block); |
duke@435 | 2342 | __ delayed()->nop(); |
duke@435 | 2343 | __ bind(L); |
duke@435 | 2344 | |
duke@435 | 2345 | // Block. Save any potential method result value before the operation and |
duke@435 | 2346 | // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this |
duke@435 | 2347 | // lets us share the oopMap we used when we went native rather the create |
duke@435 | 2348 | // a distinct one for this pc |
duke@435 | 2349 | // |
duke@435 | 2350 | save_native_result(masm, ret_type, stack_slots); |
duke@435 | 2351 | __ call_VM_leaf(L7_thread_cache, |
duke@435 | 2352 | CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans), |
duke@435 | 2353 | G2_thread); |
duke@435 | 2354 | |
duke@435 | 2355 | // Restore any method result value |
duke@435 | 2356 | restore_native_result(masm, ret_type, stack_slots); |
duke@435 | 2357 | __ bind(no_block); |
duke@435 | 2358 | } |
duke@435 | 2359 | |
duke@435 | 2360 | // thread state is thread_in_native_trans. Any safepoint blocking has already |
duke@435 | 2361 | // happened so we can now change state to _thread_in_Java. |
duke@435 | 2362 | |
duke@435 | 2363 | |
duke@435 | 2364 | __ set(_thread_in_Java, G3_scratch); |
duke@435 | 2365 | __ st(G3_scratch, G2_thread, in_bytes(JavaThread::thread_state_offset())); |
duke@435 | 2366 | |
duke@435 | 2367 | |
duke@435 | 2368 | Label no_reguard; |
duke@435 | 2369 | __ ld(G2_thread, in_bytes(JavaThread::stack_guard_state_offset()), G3_scratch); |
duke@435 | 2370 | __ cmp(G3_scratch, JavaThread::stack_guard_yellow_disabled); |
duke@435 | 2371 | __ br(Assembler::notEqual, false, Assembler::pt, no_reguard); |
duke@435 | 2372 | __ delayed()->nop(); |
duke@435 | 2373 | |
duke@435 | 2374 | save_native_result(masm, ret_type, stack_slots); |
duke@435 | 2375 | __ call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)); |
duke@435 | 2376 | __ delayed()->nop(); |
duke@435 | 2377 | |
duke@435 | 2378 | __ restore_thread(L7_thread_cache); // restore G2_thread |
duke@435 | 2379 | restore_native_result(masm, ret_type, stack_slots); |
duke@435 | 2380 | |
duke@435 | 2381 | __ bind(no_reguard); |
duke@435 | 2382 | |
duke@435 | 2383 | // Handle possible exception (will unlock if necessary) |
duke@435 | 2384 | |
duke@435 | 2385 | // native result if any is live in freg or I0 (and I1 if long and 32bit vm) |
duke@435 | 2386 | |
duke@435 | 2387 | // Unlock |
duke@435 | 2388 | if (method->is_synchronized()) { |
duke@435 | 2389 | Label done; |
duke@435 | 2390 | Register I2_ex_oop = I2; |
duke@435 | 2391 | const Register L3_box = L3; |
duke@435 | 2392 | // Get locked oop from the handle we passed to jni |
duke@435 | 2393 | __ ld_ptr(L6_handle, 0, L4); |
duke@435 | 2394 | __ add(SP, lock_offset+STACK_BIAS, L3_box); |
duke@435 | 2395 | // Must save pending exception around the slow-path VM call. Since it's a |
duke@435 | 2396 | // leaf call, the pending exception (if any) can be kept in a register. |
duke@435 | 2397 | __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), I2_ex_oop); |
duke@435 | 2398 | // Now unlock |
duke@435 | 2399 | // (Roop, Rmark, Rbox, Rscratch) |
duke@435 | 2400 | __ compiler_unlock_object(L4, L1, L3_box, L2); |
duke@435 | 2401 | __ br(Assembler::equal, false, Assembler::pt, done); |
duke@435 | 2402 | __ delayed()-> add(SP, lock_offset+STACK_BIAS, L3_box); |
duke@435 | 2403 | |
duke@435 | 2404 | // save and restore any potential method result value around the unlocking |
duke@435 | 2405 | // operation. Will save in I0 (or stack for FP returns). |
duke@435 | 2406 | save_native_result(masm, ret_type, stack_slots); |
duke@435 | 2407 | |
duke@435 | 2408 | // Must clear pending-exception before re-entering the VM. Since this is |
duke@435 | 2409 | // a leaf call, pending-exception-oop can be safely kept in a register. |
duke@435 | 2410 | __ st_ptr(G0, G2_thread, in_bytes(Thread::pending_exception_offset())); |
duke@435 | 2411 | |
duke@435 | 2412 | // slow case of monitor enter. Inline a special case of call_VM that |
duke@435 | 2413 | // disallows any pending_exception. |
duke@435 | 2414 | __ mov(L3_box, O1); |
duke@435 | 2415 | |
duke@435 | 2416 | __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), relocInfo::runtime_call_type); |
duke@435 | 2417 | __ delayed()->mov(L4, O0); // Need oop in O0 |
duke@435 | 2418 | |
duke@435 | 2419 | __ restore_thread(L7_thread_cache); // restore G2_thread |
duke@435 | 2420 | |
duke@435 | 2421 | #ifdef ASSERT |
duke@435 | 2422 | { Label L; |
duke@435 | 2423 | __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0); |
duke@435 | 2424 | __ br_null(O0, false, Assembler::pt, L); |
duke@435 | 2425 | __ delayed()->nop(); |
duke@435 | 2426 | __ stop("no pending exception allowed on exit from IR::monitorexit"); |
duke@435 | 2427 | __ bind(L); |
duke@435 | 2428 | } |
duke@435 | 2429 | #endif |
duke@435 | 2430 | restore_native_result(masm, ret_type, stack_slots); |
duke@435 | 2431 | // check_forward_pending_exception jump to forward_exception if any pending |
duke@435 | 2432 | // exception is set. The forward_exception routine expects to see the |
duke@435 | 2433 | // exception in pending_exception and not in a register. Kind of clumsy, |
duke@435 | 2434 | // since all folks who branch to forward_exception must have tested |
duke@435 | 2435 | // pending_exception first and hence have it in a register already. |
duke@435 | 2436 | __ st_ptr(I2_ex_oop, G2_thread, in_bytes(Thread::pending_exception_offset())); |
duke@435 | 2437 | __ bind(done); |
duke@435 | 2438 | } |
duke@435 | 2439 | |
duke@435 | 2440 | // Tell dtrace about this method exit |
duke@435 | 2441 | { |
duke@435 | 2442 | SkipIfEqual skip_if( |
duke@435 | 2443 | masm, G3_scratch, &DTraceMethodProbes, Assembler::zero); |
duke@435 | 2444 | save_native_result(masm, ret_type, stack_slots); |
duke@435 | 2445 | __ set_oop_constant(JNIHandles::make_local(method()), O1); |
duke@435 | 2446 | __ call_VM_leaf(L7_thread_cache, |
duke@435 | 2447 | CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), |
duke@435 | 2448 | G2_thread, O1); |
duke@435 | 2449 | restore_native_result(masm, ret_type, stack_slots); |
duke@435 | 2450 | } |
duke@435 | 2451 | |
duke@435 | 2452 | // Clear "last Java frame" SP and PC. |
duke@435 | 2453 | __ verify_thread(); // G2_thread must be correct |
duke@435 | 2454 | __ reset_last_Java_frame(); |
duke@435 | 2455 | |
duke@435 | 2456 | // Unpack oop result |
duke@435 | 2457 | if (ret_type == T_OBJECT || ret_type == T_ARRAY) { |
duke@435 | 2458 | Label L; |
duke@435 | 2459 | __ addcc(G0, I0, G0); |
duke@435 | 2460 | __ brx(Assembler::notZero, true, Assembler::pt, L); |
duke@435 | 2461 | __ delayed()->ld_ptr(I0, 0, I0); |
duke@435 | 2462 | __ mov(G0, I0); |
duke@435 | 2463 | __ bind(L); |
duke@435 | 2464 | __ verify_oop(I0); |
duke@435 | 2465 | } |
duke@435 | 2466 | |
duke@435 | 2467 | // reset handle block |
duke@435 | 2468 | __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), L5); |
duke@435 | 2469 | __ st_ptr(G0, L5, JNIHandleBlock::top_offset_in_bytes()); |
duke@435 | 2470 | |
duke@435 | 2471 | __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), G3_scratch); |
duke@435 | 2472 | check_forward_pending_exception(masm, G3_scratch); |
duke@435 | 2473 | |
duke@435 | 2474 | |
duke@435 | 2475 | // Return |
duke@435 | 2476 | |
duke@435 | 2477 | #ifndef _LP64 |
duke@435 | 2478 | if (ret_type == T_LONG) { |
duke@435 | 2479 | |
duke@435 | 2480 | // Must leave proper result in O0,O1 and G1 (c2/tiered only) |
duke@435 | 2481 | __ sllx(I0, 32, G1); // Shift bits into high G1 |
duke@435 | 2482 | __ srl (I1, 0, I1); // Zero extend O1 (harmless?) |
duke@435 | 2483 | __ or3 (I1, G1, G1); // OR 64 bits into G1 |
duke@435 | 2484 | } |
duke@435 | 2485 | #endif |
duke@435 | 2486 | |
duke@435 | 2487 | __ ret(); |
duke@435 | 2488 | __ delayed()->restore(); |
duke@435 | 2489 | |
duke@435 | 2490 | __ flush(); |
duke@435 | 2491 | |
duke@435 | 2492 | nmethod *nm = nmethod::new_native_nmethod(method, |
duke@435 | 2493 | masm->code(), |
duke@435 | 2494 | vep_offset, |
duke@435 | 2495 | frame_complete, |
duke@435 | 2496 | stack_slots / VMRegImpl::slots_per_word, |
duke@435 | 2497 | (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)), |
duke@435 | 2498 | in_ByteSize(lock_offset), |
duke@435 | 2499 | oop_maps); |
duke@435 | 2500 | return nm; |
duke@435 | 2501 | |
duke@435 | 2502 | } |
duke@435 | 2503 | |
duke@435 | 2504 | // this function returns the adjust size (in number of words) to a c2i adapter |
duke@435 | 2505 | // activation for use during deoptimization |
duke@435 | 2506 | int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) { |
duke@435 | 2507 | assert(callee_locals >= callee_parameters, |
duke@435 | 2508 | "test and remove; got more parms than locals"); |
duke@435 | 2509 | if (callee_locals < callee_parameters) |
duke@435 | 2510 | return 0; // No adjustment for negative locals |
duke@435 | 2511 | int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords(); |
duke@435 | 2512 | return round_to(diff, WordsPerLong); |
duke@435 | 2513 | } |
duke@435 | 2514 | |
duke@435 | 2515 | // "Top of Stack" slots that may be unused by the calling convention but must |
duke@435 | 2516 | // otherwise be preserved. |
duke@435 | 2517 | // On Intel these are not necessary and the value can be zero. |
duke@435 | 2518 | // On Sparc this describes the words reserved for storing a register window |
duke@435 | 2519 | // when an interrupt occurs. |
duke@435 | 2520 | uint SharedRuntime::out_preserve_stack_slots() { |
duke@435 | 2521 | return frame::register_save_words * VMRegImpl::slots_per_word; |
duke@435 | 2522 | } |
duke@435 | 2523 | |
duke@435 | 2524 | static void gen_new_frame(MacroAssembler* masm, bool deopt) { |
duke@435 | 2525 | // |
duke@435 | 2526 | // Common out the new frame generation for deopt and uncommon trap |
duke@435 | 2527 | // |
duke@435 | 2528 | Register G3pcs = G3_scratch; // Array of new pcs (input) |
duke@435 | 2529 | Register Oreturn0 = O0; |
duke@435 | 2530 | Register Oreturn1 = O1; |
duke@435 | 2531 | Register O2UnrollBlock = O2; |
duke@435 | 2532 | Register O3array = O3; // Array of frame sizes (input) |
duke@435 | 2533 | Register O4array_size = O4; // number of frames (input) |
duke@435 | 2534 | Register O7frame_size = O7; // number of frames (input) |
duke@435 | 2535 | |
duke@435 | 2536 | __ ld_ptr(O3array, 0, O7frame_size); |
duke@435 | 2537 | __ sub(G0, O7frame_size, O7frame_size); |
duke@435 | 2538 | __ save(SP, O7frame_size, SP); |
duke@435 | 2539 | __ ld_ptr(G3pcs, 0, I7); // load frame's new pc |
duke@435 | 2540 | |
duke@435 | 2541 | #ifdef ASSERT |
duke@435 | 2542 | // make sure that the frames are aligned properly |
duke@435 | 2543 | #ifndef _LP64 |
duke@435 | 2544 | __ btst(wordSize*2-1, SP); |
duke@435 | 2545 | __ breakpoint_trap(Assembler::notZero); |
duke@435 | 2546 | #endif |
duke@435 | 2547 | #endif |
duke@435 | 2548 | |
duke@435 | 2549 | // Deopt needs to pass some extra live values from frame to frame |
duke@435 | 2550 | |
duke@435 | 2551 | if (deopt) { |
duke@435 | 2552 | __ mov(Oreturn0->after_save(), Oreturn0); |
duke@435 | 2553 | __ mov(Oreturn1->after_save(), Oreturn1); |
duke@435 | 2554 | } |
duke@435 | 2555 | |
duke@435 | 2556 | __ mov(O4array_size->after_save(), O4array_size); |
duke@435 | 2557 | __ sub(O4array_size, 1, O4array_size); |
duke@435 | 2558 | __ mov(O3array->after_save(), O3array); |
duke@435 | 2559 | __ mov(O2UnrollBlock->after_save(), O2UnrollBlock); |
duke@435 | 2560 | __ add(G3pcs, wordSize, G3pcs); // point to next pc value |
duke@435 | 2561 | |
duke@435 | 2562 | #ifdef ASSERT |
duke@435 | 2563 | // trash registers to show a clear pattern in backtraces |
duke@435 | 2564 | __ set(0xDEAD0000, I0); |
duke@435 | 2565 | __ add(I0, 2, I1); |
duke@435 | 2566 | __ add(I0, 4, I2); |
duke@435 | 2567 | __ add(I0, 6, I3); |
duke@435 | 2568 | __ add(I0, 8, I4); |
duke@435 | 2569 | // Don't touch I5 could have valuable savedSP |
duke@435 | 2570 | __ set(0xDEADBEEF, L0); |
duke@435 | 2571 | __ mov(L0, L1); |
duke@435 | 2572 | __ mov(L0, L2); |
duke@435 | 2573 | __ mov(L0, L3); |
duke@435 | 2574 | __ mov(L0, L4); |
duke@435 | 2575 | __ mov(L0, L5); |
duke@435 | 2576 | |
duke@435 | 2577 | // trash the return value as there is nothing to return yet |
duke@435 | 2578 | __ set(0xDEAD0001, O7); |
duke@435 | 2579 | #endif |
duke@435 | 2580 | |
duke@435 | 2581 | __ mov(SP, O5_savedSP); |
duke@435 | 2582 | } |
duke@435 | 2583 | |
duke@435 | 2584 | |
duke@435 | 2585 | static void make_new_frames(MacroAssembler* masm, bool deopt) { |
duke@435 | 2586 | // |
duke@435 | 2587 | // loop through the UnrollBlock info and create new frames |
duke@435 | 2588 | // |
duke@435 | 2589 | Register G3pcs = G3_scratch; |
duke@435 | 2590 | Register Oreturn0 = O0; |
duke@435 | 2591 | Register Oreturn1 = O1; |
duke@435 | 2592 | Register O2UnrollBlock = O2; |
duke@435 | 2593 | Register O3array = O3; |
duke@435 | 2594 | Register O4array_size = O4; |
duke@435 | 2595 | Label loop; |
duke@435 | 2596 | |
duke@435 | 2597 | // Before we make new frames, check to see if stack is available. |
duke@435 | 2598 | // Do this after the caller's return address is on top of stack |
duke@435 | 2599 | if (UseStackBanging) { |
duke@435 | 2600 | // Get total frame size for interpreted frames |
duke@435 | 2601 | __ ld(Address(O2UnrollBlock, 0, |
duke@435 | 2602 | Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes()), O4); |
duke@435 | 2603 | __ bang_stack_size(O4, O3, G3_scratch); |
duke@435 | 2604 | } |
duke@435 | 2605 | |
duke@435 | 2606 | __ ld(Address(O2UnrollBlock, 0, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()), O4array_size); |
duke@435 | 2607 | __ ld_ptr(Address(O2UnrollBlock, 0, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()), G3pcs); |
duke@435 | 2608 | |
duke@435 | 2609 | __ ld_ptr(Address(O2UnrollBlock, 0, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()), O3array); |
duke@435 | 2610 | |
duke@435 | 2611 | // Adjust old interpreter frame to make space for new frame's extra java locals |
duke@435 | 2612 | // |
duke@435 | 2613 | // We capture the original sp for the transition frame only because it is needed in |
duke@435 | 2614 | // order to properly calculate interpreter_sp_adjustment. Even though in real life |
duke@435 | 2615 | // every interpreter frame captures a savedSP it is only needed at the transition |
duke@435 | 2616 | // (fortunately). If we had to have it correct everywhere then we would need to |
duke@435 | 2617 | // be told the sp_adjustment for each frame we create. If the frame size array |
duke@435 | 2618 | // were to have twice the frame count entries then we could have pairs [sp_adjustment, frame_size] |
duke@435 | 2619 | // for each frame we create and keep up the illusion every where. |
duke@435 | 2620 | // |
duke@435 | 2621 | |
duke@435 | 2622 | __ ld(Address(O2UnrollBlock, 0, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()), O7); |
duke@435 | 2623 | __ mov(SP, O5_savedSP); // remember initial sender's original sp before adjustment |
duke@435 | 2624 | __ sub(SP, O7, SP); |
duke@435 | 2625 | |
duke@435 | 2626 | #ifdef ASSERT |
duke@435 | 2627 | // make sure that there is at least one entry in the array |
duke@435 | 2628 | __ tst(O4array_size); |
duke@435 | 2629 | __ breakpoint_trap(Assembler::zero); |
duke@435 | 2630 | #endif |
duke@435 | 2631 | |
duke@435 | 2632 | // Now push the new interpreter frames |
duke@435 | 2633 | __ bind(loop); |
duke@435 | 2634 | |
duke@435 | 2635 | // allocate a new frame, filling the registers |
duke@435 | 2636 | |
duke@435 | 2637 | gen_new_frame(masm, deopt); // allocate an interpreter frame |
duke@435 | 2638 | |
duke@435 | 2639 | __ tst(O4array_size); |
duke@435 | 2640 | __ br(Assembler::notZero, false, Assembler::pn, loop); |
duke@435 | 2641 | __ delayed()->add(O3array, wordSize, O3array); |
duke@435 | 2642 | __ ld_ptr(G3pcs, 0, O7); // load final frame new pc |
duke@435 | 2643 | |
duke@435 | 2644 | } |
duke@435 | 2645 | |
duke@435 | 2646 | //------------------------------generate_deopt_blob---------------------------- |
duke@435 | 2647 | // Ought to generate an ideal graph & compile, but here's some SPARC ASM |
duke@435 | 2648 | // instead. |
duke@435 | 2649 | void SharedRuntime::generate_deopt_blob() { |
duke@435 | 2650 | // allocate space for the code |
duke@435 | 2651 | ResourceMark rm; |
duke@435 | 2652 | // setup code generation tools |
duke@435 | 2653 | int pad = VerifyThread ? 512 : 0;// Extra slop space for more verify code |
duke@435 | 2654 | #ifdef _LP64 |
duke@435 | 2655 | CodeBuffer buffer("deopt_blob", 2100+pad, 512); |
duke@435 | 2656 | #else |
duke@435 | 2657 | // Measured 8/7/03 at 1212 in 32bit debug build (no VerifyThread) |
duke@435 | 2658 | // Measured 8/7/03 at 1396 in 32bit debug build (VerifyThread) |
duke@435 | 2659 | CodeBuffer buffer("deopt_blob", 1600+pad, 512); |
duke@435 | 2660 | #endif /* _LP64 */ |
duke@435 | 2661 | MacroAssembler* masm = new MacroAssembler(&buffer); |
duke@435 | 2662 | FloatRegister Freturn0 = F0; |
duke@435 | 2663 | Register Greturn1 = G1; |
duke@435 | 2664 | Register Oreturn0 = O0; |
duke@435 | 2665 | Register Oreturn1 = O1; |
duke@435 | 2666 | Register O2UnrollBlock = O2; |
duke@435 | 2667 | Register O3tmp = O3; |
duke@435 | 2668 | Register I5exception_tmp = I5; |
duke@435 | 2669 | Register G4exception_tmp = G4_scratch; |
duke@435 | 2670 | int frame_size_words; |
duke@435 | 2671 | Address saved_Freturn0_addr(FP, 0, -sizeof(double) + STACK_BIAS); |
duke@435 | 2672 | #if !defined(_LP64) && defined(COMPILER2) |
duke@435 | 2673 | Address saved_Greturn1_addr(FP, 0, -sizeof(double) -sizeof(jlong) + STACK_BIAS); |
duke@435 | 2674 | #endif |
duke@435 | 2675 | Label cont; |
duke@435 | 2676 | |
duke@435 | 2677 | OopMapSet *oop_maps = new OopMapSet(); |
duke@435 | 2678 | |
duke@435 | 2679 | // |
duke@435 | 2680 | // This is the entry point for code which is returning to a de-optimized |
duke@435 | 2681 | // frame. |
duke@435 | 2682 | // The steps taken by this frame are as follows: |
duke@435 | 2683 | // - push a dummy "register_save" and save the return values (O0, O1, F0/F1, G1) |
duke@435 | 2684 | // and all potentially live registers (at a pollpoint many registers can be live). |
duke@435 | 2685 | // |
duke@435 | 2686 | // - call the C routine: Deoptimization::fetch_unroll_info (this function |
duke@435 | 2687 | // returns information about the number and size of interpreter frames |
duke@435 | 2688 | // which are equivalent to the frame which is being deoptimized) |
duke@435 | 2689 | // - deallocate the unpack frame, restoring only results values. Other |
duke@435 | 2690 | // volatile registers will now be captured in the vframeArray as needed. |
duke@435 | 2691 | // - deallocate the deoptimization frame |
duke@435 | 2692 | // - in a loop using the information returned in the previous step |
duke@435 | 2693 | // push new interpreter frames (take care to propagate the return |
duke@435 | 2694 | // values through each new frame pushed) |
duke@435 | 2695 | // - create a dummy "unpack_frame" and save the return values (O0, O1, F0) |
duke@435 | 2696 | // - call the C routine: Deoptimization::unpack_frames (this function |
duke@435 | 2697 | // lays out values on the interpreter frame which was just created) |
duke@435 | 2698 | // - deallocate the dummy unpack_frame |
duke@435 | 2699 | // - ensure that all the return values are correctly set and then do |
duke@435 | 2700 | // a return to the interpreter entry point |
duke@435 | 2701 | // |
duke@435 | 2702 | // Refer to the following methods for more information: |
duke@435 | 2703 | // - Deoptimization::fetch_unroll_info |
duke@435 | 2704 | // - Deoptimization::unpack_frames |
duke@435 | 2705 | |
duke@435 | 2706 | OopMap* map = NULL; |
duke@435 | 2707 | |
duke@435 | 2708 | int start = __ offset(); |
duke@435 | 2709 | |
duke@435 | 2710 | // restore G2, the trampoline destroyed it |
duke@435 | 2711 | __ get_thread(); |
duke@435 | 2712 | |
duke@435 | 2713 | // On entry we have been called by the deoptimized nmethod with a call that |
duke@435 | 2714 | // replaced the original call (or safepoint polling location) so the deoptimizing |
duke@435 | 2715 | // pc is now in O7. Return values are still in the expected places |
duke@435 | 2716 | |
duke@435 | 2717 | map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words); |
duke@435 | 2718 | __ ba(false, cont); |
duke@435 | 2719 | __ delayed()->mov(Deoptimization::Unpack_deopt, I5exception_tmp); |
duke@435 | 2720 | |
duke@435 | 2721 | int exception_offset = __ offset() - start; |
duke@435 | 2722 | |
duke@435 | 2723 | // restore G2, the trampoline destroyed it |
duke@435 | 2724 | __ get_thread(); |
duke@435 | 2725 | |
duke@435 | 2726 | // On entry we have been jumped to by the exception handler (or exception_blob |
duke@435 | 2727 | // for server). O0 contains the exception oop and O7 contains the original |
duke@435 | 2728 | // exception pc. So if we push a frame here it will look to the |
duke@435 | 2729 | // stack walking code (fetch_unroll_info) just like a normal call so |
duke@435 | 2730 | // state will be extracted normally. |
duke@435 | 2731 | |
duke@435 | 2732 | // save exception oop in JavaThread and fall through into the |
duke@435 | 2733 | // exception_in_tls case since they are handled in same way except |
duke@435 | 2734 | // for where the pending exception is kept. |
duke@435 | 2735 | __ st_ptr(Oexception, G2_thread, in_bytes(JavaThread::exception_oop_offset())); |
duke@435 | 2736 | |
duke@435 | 2737 | // |
duke@435 | 2738 | // Vanilla deoptimization with an exception pending in exception_oop |
duke@435 | 2739 | // |
duke@435 | 2740 | int exception_in_tls_offset = __ offset() - start; |
duke@435 | 2741 | |
duke@435 | 2742 | // No need to update oop_map as each call to save_live_registers will produce identical oopmap |
duke@435 | 2743 | (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words); |
duke@435 | 2744 | |
duke@435 | 2745 | // Restore G2_thread |
duke@435 | 2746 | __ get_thread(); |
duke@435 | 2747 | |
duke@435 | 2748 | #ifdef ASSERT |
duke@435 | 2749 | { |
duke@435 | 2750 | // verify that there is really an exception oop in exception_oop |
duke@435 | 2751 | Label has_exception; |
duke@435 | 2752 | __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception); |
duke@435 | 2753 | __ br_notnull(Oexception, false, Assembler::pt, has_exception); |
duke@435 | 2754 | __ delayed()-> nop(); |
duke@435 | 2755 | __ stop("no exception in thread"); |
duke@435 | 2756 | __ bind(has_exception); |
duke@435 | 2757 | |
duke@435 | 2758 | // verify that there is no pending exception |
duke@435 | 2759 | Label no_pending_exception; |
duke@435 | 2760 | Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset())); |
duke@435 | 2761 | __ ld_ptr(exception_addr, Oexception); |
duke@435 | 2762 | __ br_null(Oexception, false, Assembler::pt, no_pending_exception); |
duke@435 | 2763 | __ delayed()->nop(); |
duke@435 | 2764 | __ stop("must not have pending exception here"); |
duke@435 | 2765 | __ bind(no_pending_exception); |
duke@435 | 2766 | } |
duke@435 | 2767 | #endif |
duke@435 | 2768 | |
duke@435 | 2769 | __ ba(false, cont); |
duke@435 | 2770 | __ delayed()->mov(Deoptimization::Unpack_exception, I5exception_tmp);; |
duke@435 | 2771 | |
duke@435 | 2772 | // |
duke@435 | 2773 | // Reexecute entry, similar to c2 uncommon trap |
duke@435 | 2774 | // |
duke@435 | 2775 | int reexecute_offset = __ offset() - start; |
duke@435 | 2776 | |
duke@435 | 2777 | // No need to update oop_map as each call to save_live_registers will produce identical oopmap |
duke@435 | 2778 | (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words); |
duke@435 | 2779 | |
duke@435 | 2780 | __ mov(Deoptimization::Unpack_reexecute, I5exception_tmp); |
duke@435 | 2781 | |
duke@435 | 2782 | __ bind(cont); |
duke@435 | 2783 | |
duke@435 | 2784 | __ set_last_Java_frame(SP, noreg); |
duke@435 | 2785 | |
duke@435 | 2786 | // do the call by hand so we can get the oopmap |
duke@435 | 2787 | |
duke@435 | 2788 | __ mov(G2_thread, L7_thread_cache); |
duke@435 | 2789 | __ call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), relocInfo::runtime_call_type); |
duke@435 | 2790 | __ delayed()->mov(G2_thread, O0); |
duke@435 | 2791 | |
duke@435 | 2792 | // Set an oopmap for the call site this describes all our saved volatile registers |
duke@435 | 2793 | |
duke@435 | 2794 | oop_maps->add_gc_map( __ offset()-start, map); |
duke@435 | 2795 | |
duke@435 | 2796 | __ mov(L7_thread_cache, G2_thread); |
duke@435 | 2797 | |
duke@435 | 2798 | __ reset_last_Java_frame(); |
duke@435 | 2799 | |
duke@435 | 2800 | // NOTE: we know that only O0/O1 will be reloaded by restore_result_registers |
duke@435 | 2801 | // so this move will survive |
duke@435 | 2802 | |
duke@435 | 2803 | __ mov(I5exception_tmp, G4exception_tmp); |
duke@435 | 2804 | |
duke@435 | 2805 | __ mov(O0, O2UnrollBlock->after_save()); |
duke@435 | 2806 | |
duke@435 | 2807 | RegisterSaver::restore_result_registers(masm); |
duke@435 | 2808 | |
duke@435 | 2809 | Label noException; |
duke@435 | 2810 | __ cmp(G4exception_tmp, Deoptimization::Unpack_exception); // Was exception pending? |
duke@435 | 2811 | __ br(Assembler::notEqual, false, Assembler::pt, noException); |
duke@435 | 2812 | __ delayed()->nop(); |
duke@435 | 2813 | |
duke@435 | 2814 | // Move the pending exception from exception_oop to Oexception so |
duke@435 | 2815 | // the pending exception will be picked up the interpreter. |
duke@435 | 2816 | __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception); |
duke@435 | 2817 | __ st_ptr(G0, G2_thread, in_bytes(JavaThread::exception_oop_offset())); |
duke@435 | 2818 | __ bind(noException); |
duke@435 | 2819 | |
duke@435 | 2820 | // deallocate the deoptimization frame taking care to preserve the return values |
duke@435 | 2821 | __ mov(Oreturn0, Oreturn0->after_save()); |
duke@435 | 2822 | __ mov(Oreturn1, Oreturn1->after_save()); |
duke@435 | 2823 | __ mov(O2UnrollBlock, O2UnrollBlock->after_save()); |
duke@435 | 2824 | __ restore(); |
duke@435 | 2825 | |
duke@435 | 2826 | // Allocate new interpreter frame(s) and possible c2i adapter frame |
duke@435 | 2827 | |
duke@435 | 2828 | make_new_frames(masm, true); |
duke@435 | 2829 | |
duke@435 | 2830 | // push a dummy "unpack_frame" taking care of float return values and |
duke@435 | 2831 | // call Deoptimization::unpack_frames to have the unpacker layout |
duke@435 | 2832 | // information in the interpreter frames just created and then return |
duke@435 | 2833 | // to the interpreter entry point |
duke@435 | 2834 | __ save(SP, -frame_size_words*wordSize, SP); |
duke@435 | 2835 | __ stf(FloatRegisterImpl::D, Freturn0, saved_Freturn0_addr); |
duke@435 | 2836 | #if !defined(_LP64) |
duke@435 | 2837 | #if defined(COMPILER2) |
duke@435 | 2838 | if (!TieredCompilation) { |
duke@435 | 2839 | // 32-bit 1-register longs return longs in G1 |
duke@435 | 2840 | __ stx(Greturn1, saved_Greturn1_addr); |
duke@435 | 2841 | } |
duke@435 | 2842 | #endif |
duke@435 | 2843 | __ set_last_Java_frame(SP, noreg); |
duke@435 | 2844 | __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, G4exception_tmp); |
duke@435 | 2845 | #else |
duke@435 | 2846 | // LP64 uses g4 in set_last_Java_frame |
duke@435 | 2847 | __ mov(G4exception_tmp, O1); |
duke@435 | 2848 | __ set_last_Java_frame(SP, G0); |
duke@435 | 2849 | __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O1); |
duke@435 | 2850 | #endif |
duke@435 | 2851 | __ reset_last_Java_frame(); |
duke@435 | 2852 | __ ldf(FloatRegisterImpl::D, saved_Freturn0_addr, Freturn0); |
duke@435 | 2853 | |
duke@435 | 2854 | // In tiered we never use C2 to compile methods returning longs so |
duke@435 | 2855 | // the result is where we expect it already. |
duke@435 | 2856 | |
duke@435 | 2857 | #if !defined(_LP64) && defined(COMPILER2) |
duke@435 | 2858 | // In 32 bit, C2 returns longs in G1 so restore the saved G1 into |
duke@435 | 2859 | // I0/I1 if the return value is long. In the tiered world there is |
duke@435 | 2860 | // a mismatch between how C1 and C2 return longs compiles and so |
duke@435 | 2861 | // currently compilation of methods which return longs is disabled |
duke@435 | 2862 | // for C2 and so is this code. Eventually C1 and C2 will do the |
duke@435 | 2863 | // same thing for longs in the tiered world. |
duke@435 | 2864 | if (!TieredCompilation) { |
duke@435 | 2865 | Label not_long; |
duke@435 | 2866 | __ cmp(O0,T_LONG); |
duke@435 | 2867 | __ br(Assembler::notEqual, false, Assembler::pt, not_long); |
duke@435 | 2868 | __ delayed()->nop(); |
duke@435 | 2869 | __ ldd(saved_Greturn1_addr,I0); |
duke@435 | 2870 | __ bind(not_long); |
duke@435 | 2871 | } |
duke@435 | 2872 | #endif |
duke@435 | 2873 | __ ret(); |
duke@435 | 2874 | __ delayed()->restore(); |
duke@435 | 2875 | |
duke@435 | 2876 | masm->flush(); |
duke@435 | 2877 | _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_words); |
duke@435 | 2878 | _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset); |
duke@435 | 2879 | } |
duke@435 | 2880 | |
duke@435 | 2881 | #ifdef COMPILER2 |
duke@435 | 2882 | |
duke@435 | 2883 | //------------------------------generate_uncommon_trap_blob-------------------- |
duke@435 | 2884 | // Ought to generate an ideal graph & compile, but here's some SPARC ASM |
duke@435 | 2885 | // instead. |
duke@435 | 2886 | void SharedRuntime::generate_uncommon_trap_blob() { |
duke@435 | 2887 | // allocate space for the code |
duke@435 | 2888 | ResourceMark rm; |
duke@435 | 2889 | // setup code generation tools |
duke@435 | 2890 | int pad = VerifyThread ? 512 : 0; |
duke@435 | 2891 | #ifdef _LP64 |
duke@435 | 2892 | CodeBuffer buffer("uncommon_trap_blob", 2700+pad, 512); |
duke@435 | 2893 | #else |
duke@435 | 2894 | // Measured 8/7/03 at 660 in 32bit debug build (no VerifyThread) |
duke@435 | 2895 | // Measured 8/7/03 at 1028 in 32bit debug build (VerifyThread) |
duke@435 | 2896 | CodeBuffer buffer("uncommon_trap_blob", 2000+pad, 512); |
duke@435 | 2897 | #endif |
duke@435 | 2898 | MacroAssembler* masm = new MacroAssembler(&buffer); |
duke@435 | 2899 | Register O2UnrollBlock = O2; |
duke@435 | 2900 | Register O3tmp = O3; |
duke@435 | 2901 | Register O2klass_index = O2; |
duke@435 | 2902 | |
duke@435 | 2903 | // |
duke@435 | 2904 | // This is the entry point for all traps the compiler takes when it thinks |
duke@435 | 2905 | // it cannot handle further execution of compilation code. The frame is |
duke@435 | 2906 | // deoptimized in these cases and converted into interpreter frames for |
duke@435 | 2907 | // execution |
duke@435 | 2908 | // The steps taken by this frame are as follows: |
duke@435 | 2909 | // - push a fake "unpack_frame" |
duke@435 | 2910 | // - call the C routine Deoptimization::uncommon_trap (this function |
duke@435 | 2911 | // packs the current compiled frame into vframe arrays and returns |
duke@435 | 2912 | // information about the number and size of interpreter frames which |
duke@435 | 2913 | // are equivalent to the frame which is being deoptimized) |
duke@435 | 2914 | // - deallocate the "unpack_frame" |
duke@435 | 2915 | // - deallocate the deoptimization frame |
duke@435 | 2916 | // - in a loop using the information returned in the previous step |
duke@435 | 2917 | // push interpreter frames; |
duke@435 | 2918 | // - create a dummy "unpack_frame" |
duke@435 | 2919 | // - call the C routine: Deoptimization::unpack_frames (this function |
duke@435 | 2920 | // lays out values on the interpreter frame which was just created) |
duke@435 | 2921 | // - deallocate the dummy unpack_frame |
duke@435 | 2922 | // - return to the interpreter entry point |
duke@435 | 2923 | // |
duke@435 | 2924 | // Refer to the following methods for more information: |
duke@435 | 2925 | // - Deoptimization::uncommon_trap |
duke@435 | 2926 | // - Deoptimization::unpack_frame |
duke@435 | 2927 | |
duke@435 | 2928 | // the unloaded class index is in O0 (first parameter to this blob) |
duke@435 | 2929 | |
duke@435 | 2930 | // push a dummy "unpack_frame" |
duke@435 | 2931 | // and call Deoptimization::uncommon_trap to pack the compiled frame into |
duke@435 | 2932 | // vframe array and return the UnrollBlock information |
duke@435 | 2933 | __ save_frame(0); |
duke@435 | 2934 | __ set_last_Java_frame(SP, noreg); |
duke@435 | 2935 | __ mov(I0, O2klass_index); |
duke@435 | 2936 | __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), G2_thread, O2klass_index); |
duke@435 | 2937 | __ reset_last_Java_frame(); |
duke@435 | 2938 | __ mov(O0, O2UnrollBlock->after_save()); |
duke@435 | 2939 | __ restore(); |
duke@435 | 2940 | |
duke@435 | 2941 | // deallocate the deoptimized frame taking care to preserve the return values |
duke@435 | 2942 | __ mov(O2UnrollBlock, O2UnrollBlock->after_save()); |
duke@435 | 2943 | __ restore(); |
duke@435 | 2944 | |
duke@435 | 2945 | // Allocate new interpreter frame(s) and possible c2i adapter frame |
duke@435 | 2946 | |
duke@435 | 2947 | make_new_frames(masm, false); |
duke@435 | 2948 | |
duke@435 | 2949 | // push a dummy "unpack_frame" taking care of float return values and |
duke@435 | 2950 | // call Deoptimization::unpack_frames to have the unpacker layout |
duke@435 | 2951 | // information in the interpreter frames just created and then return |
duke@435 | 2952 | // to the interpreter entry point |
duke@435 | 2953 | __ save_frame(0); |
duke@435 | 2954 | __ set_last_Java_frame(SP, noreg); |
duke@435 | 2955 | __ mov(Deoptimization::Unpack_uncommon_trap, O3); // indicate it is the uncommon trap case |
duke@435 | 2956 | __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O3); |
duke@435 | 2957 | __ reset_last_Java_frame(); |
duke@435 | 2958 | __ ret(); |
duke@435 | 2959 | __ delayed()->restore(); |
duke@435 | 2960 | |
duke@435 | 2961 | masm->flush(); |
duke@435 | 2962 | _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, NULL, __ total_frame_size_in_bytes(0)/wordSize); |
duke@435 | 2963 | } |
duke@435 | 2964 | |
duke@435 | 2965 | #endif // COMPILER2 |
duke@435 | 2966 | |
duke@435 | 2967 | //------------------------------generate_handler_blob------------------- |
duke@435 | 2968 | // |
duke@435 | 2969 | // Generate a special Compile2Runtime blob that saves all registers, and sets |
duke@435 | 2970 | // up an OopMap. |
duke@435 | 2971 | // |
duke@435 | 2972 | // This blob is jumped to (via a breakpoint and the signal handler) from a |
duke@435 | 2973 | // safepoint in compiled code. On entry to this blob, O7 contains the |
duke@435 | 2974 | // address in the original nmethod at which we should resume normal execution. |
duke@435 | 2975 | // Thus, this blob looks like a subroutine which must preserve lots of |
duke@435 | 2976 | // registers and return normally. Note that O7 is never register-allocated, |
duke@435 | 2977 | // so it is guaranteed to be free here. |
duke@435 | 2978 | // |
duke@435 | 2979 | |
duke@435 | 2980 | // The hardest part of what this blob must do is to save the 64-bit %o |
duke@435 | 2981 | // registers in the 32-bit build. A simple 'save' turn the %o's to %i's and |
duke@435 | 2982 | // an interrupt will chop off their heads. Making space in the caller's frame |
duke@435 | 2983 | // first will let us save the 64-bit %o's before save'ing, but we cannot hand |
duke@435 | 2984 | // the adjusted FP off to the GC stack-crawler: this will modify the caller's |
duke@435 | 2985 | // SP and mess up HIS OopMaps. So we first adjust the caller's SP, then save |
duke@435 | 2986 | // the 64-bit %o's, then do a save, then fixup the caller's SP (our FP). |
duke@435 | 2987 | // Tricky, tricky, tricky... |
duke@435 | 2988 | |
duke@435 | 2989 | static SafepointBlob* generate_handler_blob(address call_ptr, bool cause_return) { |
duke@435 | 2990 | assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); |
duke@435 | 2991 | |
duke@435 | 2992 | // allocate space for the code |
duke@435 | 2993 | ResourceMark rm; |
duke@435 | 2994 | // setup code generation tools |
duke@435 | 2995 | // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread) |
duke@435 | 2996 | // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread) |
duke@435 | 2997 | // even larger with TraceJumps |
duke@435 | 2998 | int pad = TraceJumps ? 512 : 0; |
duke@435 | 2999 | CodeBuffer buffer("handler_blob", 1600 + pad, 512); |
duke@435 | 3000 | MacroAssembler* masm = new MacroAssembler(&buffer); |
duke@435 | 3001 | int frame_size_words; |
duke@435 | 3002 | OopMapSet *oop_maps = new OopMapSet(); |
duke@435 | 3003 | OopMap* map = NULL; |
duke@435 | 3004 | |
duke@435 | 3005 | int start = __ offset(); |
duke@435 | 3006 | |
duke@435 | 3007 | // If this causes a return before the processing, then do a "restore" |
duke@435 | 3008 | if (cause_return) { |
duke@435 | 3009 | __ restore(); |
duke@435 | 3010 | } else { |
duke@435 | 3011 | // Make it look like we were called via the poll |
duke@435 | 3012 | // so that frame constructor always sees a valid return address |
duke@435 | 3013 | __ ld_ptr(G2_thread, in_bytes(JavaThread::saved_exception_pc_offset()), O7); |
duke@435 | 3014 | __ sub(O7, frame::pc_return_offset, O7); |
duke@435 | 3015 | } |
duke@435 | 3016 | |
duke@435 | 3017 | map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words); |
duke@435 | 3018 | |
duke@435 | 3019 | // setup last_Java_sp (blows G4) |
duke@435 | 3020 | __ set_last_Java_frame(SP, noreg); |
duke@435 | 3021 | |
duke@435 | 3022 | // call into the runtime to handle illegal instructions exception |
duke@435 | 3023 | // Do not use call_VM_leaf, because we need to make a GC map at this call site. |
duke@435 | 3024 | __ mov(G2_thread, O0); |
duke@435 | 3025 | __ save_thread(L7_thread_cache); |
duke@435 | 3026 | __ call(call_ptr); |
duke@435 | 3027 | __ delayed()->nop(); |
duke@435 | 3028 | |
duke@435 | 3029 | // Set an oopmap for the call site. |
duke@435 | 3030 | // We need this not only for callee-saved registers, but also for volatile |
duke@435 | 3031 | // registers that the compiler might be keeping live across a safepoint. |
duke@435 | 3032 | |
duke@435 | 3033 | oop_maps->add_gc_map( __ offset() - start, map); |
duke@435 | 3034 | |
duke@435 | 3035 | __ restore_thread(L7_thread_cache); |
duke@435 | 3036 | // clear last_Java_sp |
duke@435 | 3037 | __ reset_last_Java_frame(); |
duke@435 | 3038 | |
duke@435 | 3039 | // Check for exceptions |
duke@435 | 3040 | Label pending; |
duke@435 | 3041 | |
duke@435 | 3042 | __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1); |
duke@435 | 3043 | __ tst(O1); |
duke@435 | 3044 | __ brx(Assembler::notEqual, true, Assembler::pn, pending); |
duke@435 | 3045 | __ delayed()->nop(); |
duke@435 | 3046 | |
duke@435 | 3047 | RegisterSaver::restore_live_registers(masm); |
duke@435 | 3048 | |
duke@435 | 3049 | // We are back the the original state on entry and ready to go. |
duke@435 | 3050 | |
duke@435 | 3051 | __ retl(); |
duke@435 | 3052 | __ delayed()->nop(); |
duke@435 | 3053 | |
duke@435 | 3054 | // Pending exception after the safepoint |
duke@435 | 3055 | |
duke@435 | 3056 | __ bind(pending); |
duke@435 | 3057 | |
duke@435 | 3058 | RegisterSaver::restore_live_registers(masm); |
duke@435 | 3059 | |
duke@435 | 3060 | // We are back the the original state on entry. |
duke@435 | 3061 | |
duke@435 | 3062 | // Tail-call forward_exception_entry, with the issuing PC in O7, |
duke@435 | 3063 | // so it looks like the original nmethod called forward_exception_entry. |
duke@435 | 3064 | __ set((intptr_t)StubRoutines::forward_exception_entry(), O0); |
duke@435 | 3065 | __ JMP(O0, 0); |
duke@435 | 3066 | __ delayed()->nop(); |
duke@435 | 3067 | |
duke@435 | 3068 | // ------------- |
duke@435 | 3069 | // make sure all code is generated |
duke@435 | 3070 | masm->flush(); |
duke@435 | 3071 | |
duke@435 | 3072 | // return exception blob |
duke@435 | 3073 | return SafepointBlob::create(&buffer, oop_maps, frame_size_words); |
duke@435 | 3074 | } |
duke@435 | 3075 | |
duke@435 | 3076 | // |
duke@435 | 3077 | // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss |
duke@435 | 3078 | // |
duke@435 | 3079 | // Generate a stub that calls into vm to find out the proper destination |
duke@435 | 3080 | // of a java call. All the argument registers are live at this point |
duke@435 | 3081 | // but since this is generic code we don't know what they are and the caller |
duke@435 | 3082 | // must do any gc of the args. |
duke@435 | 3083 | // |
duke@435 | 3084 | static RuntimeStub* generate_resolve_blob(address destination, const char* name) { |
duke@435 | 3085 | assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); |
duke@435 | 3086 | |
duke@435 | 3087 | // allocate space for the code |
duke@435 | 3088 | ResourceMark rm; |
duke@435 | 3089 | // setup code generation tools |
duke@435 | 3090 | // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread) |
duke@435 | 3091 | // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread) |
duke@435 | 3092 | // even larger with TraceJumps |
duke@435 | 3093 | int pad = TraceJumps ? 512 : 0; |
duke@435 | 3094 | CodeBuffer buffer(name, 1600 + pad, 512); |
duke@435 | 3095 | MacroAssembler* masm = new MacroAssembler(&buffer); |
duke@435 | 3096 | int frame_size_words; |
duke@435 | 3097 | OopMapSet *oop_maps = new OopMapSet(); |
duke@435 | 3098 | OopMap* map = NULL; |
duke@435 | 3099 | |
duke@435 | 3100 | int start = __ offset(); |
duke@435 | 3101 | |
duke@435 | 3102 | map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words); |
duke@435 | 3103 | |
duke@435 | 3104 | int frame_complete = __ offset(); |
duke@435 | 3105 | |
duke@435 | 3106 | // setup last_Java_sp (blows G4) |
duke@435 | 3107 | __ set_last_Java_frame(SP, noreg); |
duke@435 | 3108 | |
duke@435 | 3109 | // call into the runtime to handle illegal instructions exception |
duke@435 | 3110 | // Do not use call_VM_leaf, because we need to make a GC map at this call site. |
duke@435 | 3111 | __ mov(G2_thread, O0); |
duke@435 | 3112 | __ save_thread(L7_thread_cache); |
duke@435 | 3113 | __ call(destination, relocInfo::runtime_call_type); |
duke@435 | 3114 | __ delayed()->nop(); |
duke@435 | 3115 | |
duke@435 | 3116 | // O0 contains the address we are going to jump to assuming no exception got installed |
duke@435 | 3117 | |
duke@435 | 3118 | // Set an oopmap for the call site. |
duke@435 | 3119 | // We need this not only for callee-saved registers, but also for volatile |
duke@435 | 3120 | // registers that the compiler might be keeping live across a safepoint. |
duke@435 | 3121 | |
duke@435 | 3122 | oop_maps->add_gc_map( __ offset() - start, map); |
duke@435 | 3123 | |
duke@435 | 3124 | __ restore_thread(L7_thread_cache); |
duke@435 | 3125 | // clear last_Java_sp |
duke@435 | 3126 | __ reset_last_Java_frame(); |
duke@435 | 3127 | |
duke@435 | 3128 | // Check for exceptions |
duke@435 | 3129 | Label pending; |
duke@435 | 3130 | |
duke@435 | 3131 | __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1); |
duke@435 | 3132 | __ tst(O1); |
duke@435 | 3133 | __ brx(Assembler::notEqual, true, Assembler::pn, pending); |
duke@435 | 3134 | __ delayed()->nop(); |
duke@435 | 3135 | |
duke@435 | 3136 | // get the returned methodOop |
duke@435 | 3137 | |
duke@435 | 3138 | __ get_vm_result(G5_method); |
duke@435 | 3139 | __ stx(G5_method, SP, RegisterSaver::G5_offset()+STACK_BIAS); |
duke@435 | 3140 | |
duke@435 | 3141 | // O0 is where we want to jump, overwrite G3 which is saved and scratch |
duke@435 | 3142 | |
duke@435 | 3143 | __ stx(O0, SP, RegisterSaver::G3_offset()+STACK_BIAS); |
duke@435 | 3144 | |
duke@435 | 3145 | RegisterSaver::restore_live_registers(masm); |
duke@435 | 3146 | |
duke@435 | 3147 | // We are back the the original state on entry and ready to go. |
duke@435 | 3148 | |
duke@435 | 3149 | __ JMP(G3, 0); |
duke@435 | 3150 | __ delayed()->nop(); |
duke@435 | 3151 | |
duke@435 | 3152 | // Pending exception after the safepoint |
duke@435 | 3153 | |
duke@435 | 3154 | __ bind(pending); |
duke@435 | 3155 | |
duke@435 | 3156 | RegisterSaver::restore_live_registers(masm); |
duke@435 | 3157 | |
duke@435 | 3158 | // We are back the the original state on entry. |
duke@435 | 3159 | |
duke@435 | 3160 | // Tail-call forward_exception_entry, with the issuing PC in O7, |
duke@435 | 3161 | // so it looks like the original nmethod called forward_exception_entry. |
duke@435 | 3162 | __ set((intptr_t)StubRoutines::forward_exception_entry(), O0); |
duke@435 | 3163 | __ JMP(O0, 0); |
duke@435 | 3164 | __ delayed()->nop(); |
duke@435 | 3165 | |
duke@435 | 3166 | // ------------- |
duke@435 | 3167 | // make sure all code is generated |
duke@435 | 3168 | masm->flush(); |
duke@435 | 3169 | |
duke@435 | 3170 | // return the blob |
duke@435 | 3171 | // frame_size_words or bytes?? |
duke@435 | 3172 | return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true); |
duke@435 | 3173 | } |
duke@435 | 3174 | |
duke@435 | 3175 | void SharedRuntime::generate_stubs() { |
duke@435 | 3176 | |
duke@435 | 3177 | _wrong_method_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::handle_wrong_method), |
duke@435 | 3178 | "wrong_method_stub"); |
duke@435 | 3179 | |
duke@435 | 3180 | _ic_miss_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::handle_wrong_method_ic_miss), |
duke@435 | 3181 | "ic_miss_stub"); |
duke@435 | 3182 | |
duke@435 | 3183 | _resolve_opt_virtual_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_opt_virtual_call_C), |
duke@435 | 3184 | "resolve_opt_virtual_call"); |
duke@435 | 3185 | |
duke@435 | 3186 | _resolve_virtual_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_virtual_call_C), |
duke@435 | 3187 | "resolve_virtual_call"); |
duke@435 | 3188 | |
duke@435 | 3189 | _resolve_static_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_static_call_C), |
duke@435 | 3190 | "resolve_static_call"); |
duke@435 | 3191 | |
duke@435 | 3192 | _polling_page_safepoint_handler_blob = |
duke@435 | 3193 | generate_handler_blob(CAST_FROM_FN_PTR(address, |
duke@435 | 3194 | SafepointSynchronize::handle_polling_page_exception), false); |
duke@435 | 3195 | |
duke@435 | 3196 | _polling_page_return_handler_blob = |
duke@435 | 3197 | generate_handler_blob(CAST_FROM_FN_PTR(address, |
duke@435 | 3198 | SafepointSynchronize::handle_polling_page_exception), true); |
duke@435 | 3199 | |
duke@435 | 3200 | generate_deopt_blob(); |
duke@435 | 3201 | |
duke@435 | 3202 | #ifdef COMPILER2 |
duke@435 | 3203 | generate_uncommon_trap_blob(); |
duke@435 | 3204 | #endif // COMPILER2 |
duke@435 | 3205 | } |