1 /*
2 * Copyright (c) 2003, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "code/debugInfoRec.hpp"
28 #include "code/icBuffer.hpp"
29 #include "code/vtableStubs.hpp"
30 #include "interpreter/interpreter.hpp"
31 #include "oops/compiledICHolder.hpp"
32 #include "prims/jvmtiRedefineClassesTrace.hpp"
33 #include "runtime/sharedRuntime.hpp"
34 #include "runtime/vframeArray.hpp"
35 #include "vmreg_sparc.inline.hpp"
36 #ifdef COMPILER1
37 #include "c1/c1_Runtime1.hpp"
38 #endif
39 #ifdef COMPILER2
40 #include "opto/runtime.hpp"
41 #endif
42 #ifdef SHARK
43 #include "compiler/compileBroker.hpp"
44 #include "shark/sharkCompiler.hpp"
45 #endif
46
47 #define __ masm->
48
49
50 class RegisterSaver {
51
52 // Used for saving volatile registers. This is Gregs, Fregs, I/L/O.
53 // The Oregs are problematic. In the 32bit build the compiler can
54 // have O registers live with 64 bit quantities. A window save will
55 // cut the heads off of the registers. We have to do a very extensive
56 // stack dance to save and restore these properly.
57
58 // Note that the Oregs problem only exists if we block at either a polling
59 // page exception a compiled code safepoint that was not originally a call
60 // or deoptimize following one of these kinds of safepoints.
61
62 // Lots of registers to save. For all builds, a window save will preserve
63 // the %i and %l registers. For the 32-bit longs-in-two entries and 64-bit
64 // builds a window-save will preserve the %o registers. In the LION build
65 // we need to save the 64-bit %o registers which requires we save them
66 // before the window-save (as then they become %i registers and get their
67 // heads chopped off on interrupt). We have to save some %g registers here
68 // as well.
69 enum {
70 // This frame's save area. Includes extra space for the native call:
71 // vararg's layout space and the like. Briefly holds the caller's
72 // register save area.
73 call_args_area = frame::register_save_words_sp_offset +
74 frame::memory_parameter_word_sp_offset*wordSize,
75 // Make sure save locations are always 8 byte aligned.
76 // can't use round_to because it doesn't produce compile time constant
77 start_of_extra_save_area = ((call_args_area + 7) & ~7),
78 g1_offset = start_of_extra_save_area, // g-regs needing saving
79 g3_offset = g1_offset+8,
80 g4_offset = g3_offset+8,
81 g5_offset = g4_offset+8,
82 o0_offset = g5_offset+8,
83 o1_offset = o0_offset+8,
84 o2_offset = o1_offset+8,
85 o3_offset = o2_offset+8,
86 o4_offset = o3_offset+8,
87 o5_offset = o4_offset+8,
88 start_of_flags_save_area = o5_offset+8,
89 ccr_offset = start_of_flags_save_area,
90 fsr_offset = ccr_offset + 8,
91 d00_offset = fsr_offset+8, // Start of float save area
92 register_save_size = d00_offset+8*32
93 };
94
95
96 public:
97
98 static int Oexception_offset() { return o0_offset; };
99 static int G3_offset() { return g3_offset; };
100 static int G5_offset() { return g5_offset; };
101 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
102 static void restore_live_registers(MacroAssembler* masm);
103
104 // During deoptimization only the result register need to be restored
105 // all the other values have already been extracted.
106
107 static void restore_result_registers(MacroAssembler* masm);
108 };
109
110 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
111 // Record volatile registers as callee-save values in an OopMap so their save locations will be
112 // propagated to the caller frame's RegisterMap during StackFrameStream construction (needed for
113 // deoptimization; see compiledVFrame::create_stack_value). The caller's I, L and O registers
114 // are saved in register windows - I's and L's in the caller's frame and O's in the stub frame
115 // (as the stub's I's) when the runtime routine called by the stub creates its frame.
116 int i;
117 // Always make the frame size 16 byte aligned.
118 int frame_size = round_to(additional_frame_words + register_save_size, 16);
119 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words
120 int frame_size_in_slots = frame_size / sizeof(jint);
121 // CodeBlob frame size is in words.
122 *total_frame_words = frame_size / wordSize;
123 // OopMap* map = new OopMap(*total_frame_words, 0);
124 OopMap* map = new OopMap(frame_size_in_slots, 0);
125
126 #if !defined(_LP64)
127
128 // Save 64-bit O registers; they will get their heads chopped off on a 'save'.
129 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
130 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
131 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
132 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
133 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
134 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
135 #endif /* _LP64 */
136
137 __ save(SP, -frame_size, SP);
138
139 #ifndef _LP64
140 // Reload the 64 bit Oregs. Although they are now Iregs we load them
141 // to Oregs here to avoid interrupts cutting off their heads
142
143 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
144 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
145 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
146 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
147 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
148 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
149
150 __ stx(O0, SP, o0_offset+STACK_BIAS);
151 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset + 4)>>2), O0->as_VMReg());
152
153 __ stx(O1, SP, o1_offset+STACK_BIAS);
154
155 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset + 4)>>2), O1->as_VMReg());
156
157 __ stx(O2, SP, o2_offset+STACK_BIAS);
158 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset + 4)>>2), O2->as_VMReg());
159
160 __ stx(O3, SP, o3_offset+STACK_BIAS);
161 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset + 4)>>2), O3->as_VMReg());
162
163 __ stx(O4, SP, o4_offset+STACK_BIAS);
164 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset + 4)>>2), O4->as_VMReg());
165
166 __ stx(O5, SP, o5_offset+STACK_BIAS);
167 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset + 4)>>2), O5->as_VMReg());
168 #endif /* _LP64 */
169
170
171 #ifdef _LP64
172 int debug_offset = 0;
173 #else
174 int debug_offset = 4;
175 #endif
176 // Save the G's
177 __ stx(G1, SP, g1_offset+STACK_BIAS);
178 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset + debug_offset)>>2), G1->as_VMReg());
179
180 __ stx(G3, SP, g3_offset+STACK_BIAS);
181 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset + debug_offset)>>2), G3->as_VMReg());
182
183 __ stx(G4, SP, g4_offset+STACK_BIAS);
184 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset + debug_offset)>>2), G4->as_VMReg());
185
186 __ stx(G5, SP, g5_offset+STACK_BIAS);
187 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset + debug_offset)>>2), G5->as_VMReg());
188
189 // This is really a waste but we'll keep things as they were for now
190 if (true) {
191 #ifndef _LP64
192 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset)>>2), O0->as_VMReg()->next());
193 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset)>>2), O1->as_VMReg()->next());
194 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset)>>2), O2->as_VMReg()->next());
195 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset)>>2), O3->as_VMReg()->next());
196 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset)>>2), O4->as_VMReg()->next());
197 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset)>>2), O5->as_VMReg()->next());
198 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset)>>2), G1->as_VMReg()->next());
199 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset)>>2), G3->as_VMReg()->next());
200 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset)>>2), G4->as_VMReg()->next());
201 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset)>>2), G5->as_VMReg()->next());
202 #endif /* _LP64 */
203 }
204
205
206 // Save the flags
207 __ rdccr( G5 );
208 __ stx(G5, SP, ccr_offset+STACK_BIAS);
209 __ stxfsr(SP, fsr_offset+STACK_BIAS);
210
211 // Save all the FP registers: 32 doubles (32 floats correspond to the 2 halves of the first 16 doubles)
212 int offset = d00_offset;
213 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) {
214 FloatRegister f = as_FloatRegister(i);
215 __ stf(FloatRegisterImpl::D, f, SP, offset+STACK_BIAS);
216 // Record as callee saved both halves of double registers (2 float registers).
217 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), f->as_VMReg());
218 map->set_callee_saved(VMRegImpl::stack2reg((offset + sizeof(float))>>2), f->as_VMReg()->next());
219 offset += sizeof(double);
220 }
221
222 // And we're done.
223
224 return map;
225 }
226
227
228 // Pop the current frame and restore all the registers that we
229 // saved.
230 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
231
232 // Restore all the FP registers
233 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) {
234 __ ldf(FloatRegisterImpl::D, SP, d00_offset+i*sizeof(float)+STACK_BIAS, as_FloatRegister(i));
235 }
236
237 __ ldx(SP, ccr_offset+STACK_BIAS, G1);
238 __ wrccr (G1) ;
239
240 // Restore the G's
241 // Note that G2 (AKA GThread) must be saved and restored separately.
242 // TODO-FIXME: save and restore some of the other ASRs, viz., %asi and %gsr.
243
244 __ ldx(SP, g1_offset+STACK_BIAS, G1);
245 __ ldx(SP, g3_offset+STACK_BIAS, G3);
246 __ ldx(SP, g4_offset+STACK_BIAS, G4);
247 __ ldx(SP, g5_offset+STACK_BIAS, G5);
248
249
250 #if !defined(_LP64)
251 // Restore the 64-bit O's.
252 __ ldx(SP, o0_offset+STACK_BIAS, O0);
253 __ ldx(SP, o1_offset+STACK_BIAS, O1);
254 __ ldx(SP, o2_offset+STACK_BIAS, O2);
255 __ ldx(SP, o3_offset+STACK_BIAS, O3);
256 __ ldx(SP, o4_offset+STACK_BIAS, O4);
257 __ ldx(SP, o5_offset+STACK_BIAS, O5);
258
259 // And temporarily place them in TLS
260
261 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
262 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
263 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
264 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
265 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
266 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
267 #endif /* _LP64 */
268
269 // Restore flags
270
271 __ ldxfsr(SP, fsr_offset+STACK_BIAS);
272
273 __ restore();
274
275 #if !defined(_LP64)
276 // Now reload the 64bit Oregs after we've restore the window.
277 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
278 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
279 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
280 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
281 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
282 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
283 #endif /* _LP64 */
284
285 }
286
287 // Pop the current frame and restore the registers that might be holding
288 // a result.
289 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
290
291 #if !defined(_LP64)
292 // 32bit build returns longs in G1
293 __ ldx(SP, g1_offset+STACK_BIAS, G1);
294
295 // Retrieve the 64-bit O's.
296 __ ldx(SP, o0_offset+STACK_BIAS, O0);
297 __ ldx(SP, o1_offset+STACK_BIAS, O1);
298 // and save to TLS
299 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
300 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
301 #endif /* _LP64 */
302
303 __ ldf(FloatRegisterImpl::D, SP, d00_offset+STACK_BIAS, as_FloatRegister(0));
304
305 __ restore();
306
307 #if !defined(_LP64)
308 // Now reload the 64bit Oregs after we've restore the window.
309 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
310 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
311 #endif /* _LP64 */
312
313 }
314
315 // Is vector's size (in bytes) bigger than a size saved by default?
316 // 8 bytes FP registers are saved by default on SPARC.
317 bool SharedRuntime::is_wide_vector(int size) {
318 // Note, MaxVectorSize == 8 on SPARC.
319 assert(size <= 8, err_msg_res("%d bytes vectors are not supported", size));
320 return size > 8;
321 }
322
323 // The java_calling_convention describes stack locations as ideal slots on
324 // a frame with no abi restrictions. Since we must observe abi restrictions
325 // (like the placement of the register window) the slots must be biased by
326 // the following value.
327 static int reg2offset(VMReg r) {
328 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
329 }
330
331 static VMRegPair reg64_to_VMRegPair(Register r) {
332 VMRegPair ret;
333 if (wordSize == 8) {
334 ret.set2(r->as_VMReg());
335 } else {
336 ret.set_pair(r->successor()->as_VMReg(), r->as_VMReg());
337 }
338 return ret;
339 }
340
341 // ---------------------------------------------------------------------------
342 // Read the array of BasicTypes from a signature, and compute where the
343 // arguments should go. Values in the VMRegPair regs array refer to 4-byte (VMRegImpl::stack_slot_size)
344 // quantities. Values less than VMRegImpl::stack0 are registers, those above
345 // refer to 4-byte stack slots. All stack slots are based off of the window
346 // top. VMRegImpl::stack0 refers to the first slot past the 16-word window,
347 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register
348 // values 0-63 (up to RegisterImpl::number_of_registers) are the 64-bit
349 // integer registers. Values 64-95 are the (32-bit only) float registers.
350 // Each 32-bit quantity is given its own number, so the integer registers
351 // (in either 32- or 64-bit builds) use 2 numbers. For example, there is
352 // an O0-low and an O0-high. Essentially, all int register numbers are doubled.
353
354 // Register results are passed in O0-O5, for outgoing call arguments. To
355 // convert to incoming arguments, convert all O's to I's. The regs array
356 // refer to the low and hi 32-bit words of 64-bit registers or stack slots.
357 // If the regs[].second() field is set to VMRegImpl::Bad(), it means it's unused (a
358 // 32-bit value was passed). If both are VMRegImpl::Bad(), it means no value was
359 // passed (used as a placeholder for the other half of longs and doubles in
360 // the 64-bit build). regs[].second() is either VMRegImpl::Bad() or regs[].second() is
361 // regs[].first()+1 (regs[].first() may be misaligned in the C calling convention).
362 // Sparc never passes a value in regs[].second() but not regs[].first() (regs[].first()
363 // == VMRegImpl::Bad() && regs[].second() != VMRegImpl::Bad()) nor unrelated values in the
364 // same VMRegPair.
365
366 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
367 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
368 // units regardless of build.
369
370
371 // ---------------------------------------------------------------------------
372 // The compiled Java calling convention. The Java convention always passes
373 // 64-bit values in adjacent aligned locations (either registers or stack),
374 // floats in float registers and doubles in aligned float pairs. There is
375 // no backing varargs store for values in registers.
376 // In the 32-bit build, longs are passed on the stack (cannot be
377 // passed in I's, because longs in I's get their heads chopped off at
378 // interrupt).
379 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
380 VMRegPair *regs,
381 int total_args_passed,
382 int is_outgoing) {
383 assert(F31->as_VMReg()->is_reg(), "overlapping stack/register numbers");
384
385 const int int_reg_max = SPARC_ARGS_IN_REGS_NUM;
386 const int flt_reg_max = 8;
387
388 int int_reg = 0;
389 int flt_reg = 0;
390 int slot = 0;
391
392 for (int i = 0; i < total_args_passed; i++) {
393 switch (sig_bt[i]) {
394 case T_INT:
395 case T_SHORT:
396 case T_CHAR:
397 case T_BYTE:
398 case T_BOOLEAN:
399 #ifndef _LP64
400 case T_OBJECT:
401 case T_ARRAY:
402 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
403 #endif // _LP64
404 if (int_reg < int_reg_max) {
405 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
406 regs[i].set1(r->as_VMReg());
407 } else {
408 regs[i].set1(VMRegImpl::stack2reg(slot++));
409 }
410 break;
411
412 #ifdef _LP64
413 case T_LONG:
414 assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half");
415 // fall-through
416 case T_OBJECT:
417 case T_ARRAY:
418 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
419 if (int_reg < int_reg_max) {
420 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
421 regs[i].set2(r->as_VMReg());
422 } else {
423 slot = round_to(slot, 2); // align
424 regs[i].set2(VMRegImpl::stack2reg(slot));
425 slot += 2;
426 }
427 break;
428 #else
429 case T_LONG:
430 assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half");
431 // On 32-bit SPARC put longs always on the stack to keep the pressure off
432 // integer argument registers. They should be used for oops.
433 slot = round_to(slot, 2); // align
434 regs[i].set2(VMRegImpl::stack2reg(slot));
435 slot += 2;
436 #endif
437 break;
438
439 case T_FLOAT:
440 if (flt_reg < flt_reg_max) {
441 FloatRegister r = as_FloatRegister(flt_reg++);
442 regs[i].set1(r->as_VMReg());
443 } else {
444 regs[i].set1(VMRegImpl::stack2reg(slot++));
445 }
446 break;
447
448 case T_DOUBLE:
449 assert(sig_bt[i+1] == T_VOID, "expecting half");
450 if (round_to(flt_reg, 2) + 1 < flt_reg_max) {
451 flt_reg = round_to(flt_reg, 2); // align
452 FloatRegister r = as_FloatRegister(flt_reg);
453 regs[i].set2(r->as_VMReg());
454 flt_reg += 2;
455 } else {
456 slot = round_to(slot, 2); // align
457 regs[i].set2(VMRegImpl::stack2reg(slot));
458 slot += 2;
459 }
460 break;
461
462 case T_VOID:
463 regs[i].set_bad(); // Halves of longs & doubles
464 break;
465
466 default:
467 fatal(err_msg_res("unknown basic type %d", sig_bt[i]));
468 break;
469 }
470 }
471
472 // retun the amount of stack space these arguments will need.
473 return slot;
474 }
475
476 // Helper class mostly to avoid passing masm everywhere, and handle
477 // store displacement overflow logic.
478 class AdapterGenerator {
479 MacroAssembler *masm;
480 Register Rdisp;
481 void set_Rdisp(Register r) { Rdisp = r; }
482
483 void patch_callers_callsite();
484
485 // base+st_off points to top of argument
486 int arg_offset(const int st_off) { return st_off; }
487 int next_arg_offset(const int st_off) {
488 return st_off - Interpreter::stackElementSize;
489 }
490
491 // Argument slot values may be loaded first into a register because
492 // they might not fit into displacement.
493 RegisterOrConstant arg_slot(const int st_off);
494 RegisterOrConstant next_arg_slot(const int st_off);
495
496 // Stores long into offset pointed to by base
497 void store_c2i_long(Register r, Register base,
498 const int st_off, bool is_stack);
499 void store_c2i_object(Register r, Register base,
500 const int st_off);
501 void store_c2i_int(Register r, Register base,
502 const int st_off);
503 void store_c2i_double(VMReg r_2,
504 VMReg r_1, Register base, const int st_off);
505 void store_c2i_float(FloatRegister f, Register base,
506 const int st_off);
507
508 public:
509 void gen_c2i_adapter(int total_args_passed,
510 // VMReg max_arg,
511 int comp_args_on_stack, // VMRegStackSlots
512 const BasicType *sig_bt,
513 const VMRegPair *regs,
514 Label& skip_fixup);
515 void gen_i2c_adapter(int total_args_passed,
516 // VMReg max_arg,
517 int comp_args_on_stack, // VMRegStackSlots
518 const BasicType *sig_bt,
519 const VMRegPair *regs);
520
521 AdapterGenerator(MacroAssembler *_masm) : masm(_masm) {}
522 };
523
524
525 // Patch the callers callsite with entry to compiled code if it exists.
526 void AdapterGenerator::patch_callers_callsite() {
527 Label L;
528 __ ld_ptr(G5_method, in_bytes(Method::code_offset()), G3_scratch);
529 __ br_null(G3_scratch, false, Assembler::pt, L);
530 __ delayed()->nop();
531 // Call into the VM to patch the caller, then jump to compiled callee
532 __ save_frame(4); // Args in compiled layout; do not blow them
533
534 // Must save all the live Gregs the list is:
535 // G1: 1st Long arg (32bit build)
536 // G2: global allocated to TLS
537 // G3: used in inline cache check (scratch)
538 // G4: 2nd Long arg (32bit build);
539 // G5: used in inline cache check (Method*)
540
541 // The longs must go to the stack by hand since in the 32 bit build they can be trashed by window ops.
542
543 #ifdef _LP64
544 // mov(s,d)
545 __ mov(G1, L1);
546 __ mov(G4, L4);
547 __ mov(G5_method, L5);
548 __ mov(G5_method, O0); // VM needs target method
549 __ mov(I7, O1); // VM needs caller's callsite
550 // Must be a leaf call...
551 // can be very far once the blob has been relocated
552 AddressLiteral dest(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
553 __ relocate(relocInfo::runtime_call_type);
554 __ jumpl_to(dest, O7, O7);
555 __ delayed()->mov(G2_thread, L7_thread_cache);
556 __ mov(L7_thread_cache, G2_thread);
557 __ mov(L1, G1);
558 __ mov(L4, G4);
559 __ mov(L5, G5_method);
560 #else
561 __ stx(G1, FP, -8 + STACK_BIAS);
562 __ stx(G4, FP, -16 + STACK_BIAS);
563 __ mov(G5_method, L5);
564 __ mov(G5_method, O0); // VM needs target method
565 __ mov(I7, O1); // VM needs caller's callsite
566 // Must be a leaf call...
567 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), relocInfo::runtime_call_type);
568 __ delayed()->mov(G2_thread, L7_thread_cache);
569 __ mov(L7_thread_cache, G2_thread);
570 __ ldx(FP, -8 + STACK_BIAS, G1);
571 __ ldx(FP, -16 + STACK_BIAS, G4);
572 __ mov(L5, G5_method);
573 #endif /* _LP64 */
574
575 __ restore(); // Restore args
576 __ bind(L);
577 }
578
579
580 RegisterOrConstant AdapterGenerator::arg_slot(const int st_off) {
581 RegisterOrConstant roc(arg_offset(st_off));
582 return __ ensure_simm13_or_reg(roc, Rdisp);
583 }
584
585 RegisterOrConstant AdapterGenerator::next_arg_slot(const int st_off) {
586 RegisterOrConstant roc(next_arg_offset(st_off));
587 return __ ensure_simm13_or_reg(roc, Rdisp);
588 }
589
590
591 // Stores long into offset pointed to by base
592 void AdapterGenerator::store_c2i_long(Register r, Register base,
593 const int st_off, bool is_stack) {
594 #ifdef _LP64
595 // In V9, longs are given 2 64-bit slots in the interpreter, but the
596 // data is passed in only 1 slot.
597 __ stx(r, base, next_arg_slot(st_off));
598 #else
599 #ifdef COMPILER2
600 // Misaligned store of 64-bit data
601 __ stw(r, base, arg_slot(st_off)); // lo bits
602 __ srlx(r, 32, r);
603 __ stw(r, base, next_arg_slot(st_off)); // hi bits
604 #else
605 if (is_stack) {
606 // Misaligned store of 64-bit data
607 __ stw(r, base, arg_slot(st_off)); // lo bits
608 __ srlx(r, 32, r);
609 __ stw(r, base, next_arg_slot(st_off)); // hi bits
610 } else {
611 __ stw(r->successor(), base, arg_slot(st_off) ); // lo bits
612 __ stw(r , base, next_arg_slot(st_off)); // hi bits
613 }
614 #endif // COMPILER2
615 #endif // _LP64
616 }
617
618 void AdapterGenerator::store_c2i_object(Register r, Register base,
619 const int st_off) {
620 __ st_ptr (r, base, arg_slot(st_off));
621 }
622
623 void AdapterGenerator::store_c2i_int(Register r, Register base,
624 const int st_off) {
625 __ st (r, base, arg_slot(st_off));
626 }
627
628 // Stores into offset pointed to by base
629 void AdapterGenerator::store_c2i_double(VMReg r_2,
630 VMReg r_1, Register base, const int st_off) {
631 #ifdef _LP64
632 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
633 // data is passed in only 1 slot.
634 __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
635 #else
636 // Need to marshal 64-bit value from misaligned Lesp loads
637 __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
638 __ stf(FloatRegisterImpl::S, r_2->as_FloatRegister(), base, arg_slot(st_off) );
639 #endif
640 }
641
642 void AdapterGenerator::store_c2i_float(FloatRegister f, Register base,
643 const int st_off) {
644 __ stf(FloatRegisterImpl::S, f, base, arg_slot(st_off));
645 }
646
647 void AdapterGenerator::gen_c2i_adapter(
648 int total_args_passed,
649 // VMReg max_arg,
650 int comp_args_on_stack, // VMRegStackSlots
651 const BasicType *sig_bt,
652 const VMRegPair *regs,
653 Label& L_skip_fixup) {
654
655 // Before we get into the guts of the C2I adapter, see if we should be here
656 // at all. We've come from compiled code and are attempting to jump to the
657 // interpreter, which means the caller made a static call to get here
658 // (vcalls always get a compiled target if there is one). Check for a
659 // compiled target. If there is one, we need to patch the caller's call.
660 // However we will run interpreted if we come thru here. The next pass
661 // thru the call site will run compiled. If we ran compiled here then
662 // we can (theorectically) do endless i2c->c2i->i2c transitions during
663 // deopt/uncommon trap cycles. If we always go interpreted here then
664 // we can have at most one and don't need to play any tricks to keep
665 // from endlessly growing the stack.
666 //
667 // Actually if we detected that we had an i2c->c2i transition here we
668 // ought to be able to reset the world back to the state of the interpreted
669 // call and not bother building another interpreter arg area. We don't
670 // do that at this point.
671
672 patch_callers_callsite();
673
674 __ bind(L_skip_fixup);
675
676 // Since all args are passed on the stack, total_args_passed*wordSize is the
677 // space we need. Add in varargs area needed by the interpreter. Round up
678 // to stack alignment.
679 const int arg_size = total_args_passed * Interpreter::stackElementSize;
680 const int varargs_area =
681 (frame::varargs_offset - frame::register_save_words)*wordSize;
682 const int extraspace = round_to(arg_size + varargs_area, 2*wordSize);
683
684 const int bias = STACK_BIAS;
685 const int interp_arg_offset = frame::varargs_offset*wordSize +
686 (total_args_passed-1)*Interpreter::stackElementSize;
687
688 const Register base = SP;
689
690 // Make some extra space on the stack.
691 __ sub(SP, __ ensure_simm13_or_reg(extraspace, G3_scratch), SP);
692 set_Rdisp(G3_scratch);
693
694 // Write the args into the outgoing interpreter space.
695 for (int i = 0; i < total_args_passed; i++) {
696 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize) + bias;
697 VMReg r_1 = regs[i].first();
698 VMReg r_2 = regs[i].second();
699 if (!r_1->is_valid()) {
700 assert(!r_2->is_valid(), "");
701 continue;
702 }
703 if (r_1->is_stack()) { // Pretend stack targets are loaded into G1
704 RegisterOrConstant ld_off = reg2offset(r_1) + extraspace + bias;
705 ld_off = __ ensure_simm13_or_reg(ld_off, Rdisp);
706 r_1 = G1_scratch->as_VMReg();// as part of the load/store shuffle
707 if (!r_2->is_valid()) __ ld (base, ld_off, G1_scratch);
708 else __ ldx(base, ld_off, G1_scratch);
709 }
710
711 if (r_1->is_Register()) {
712 Register r = r_1->as_Register()->after_restore();
713 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) {
714 store_c2i_object(r, base, st_off);
715 } else if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
716 store_c2i_long(r, base, st_off, r_2->is_stack());
717 } else {
718 store_c2i_int(r, base, st_off);
719 }
720 } else {
721 assert(r_1->is_FloatRegister(), "");
722 if (sig_bt[i] == T_FLOAT) {
723 store_c2i_float(r_1->as_FloatRegister(), base, st_off);
724 } else {
725 assert(sig_bt[i] == T_DOUBLE, "wrong type");
726 store_c2i_double(r_2, r_1, base, st_off);
727 }
728 }
729 }
730
731 // Load the interpreter entry point.
732 __ ld_ptr(G5_method, in_bytes(Method::interpreter_entry_offset()), G3_scratch);
733
734 // Pass O5_savedSP as an argument to the interpreter.
735 // The interpreter will restore SP to this value before returning.
736 __ add(SP, __ ensure_simm13_or_reg(extraspace, G1), O5_savedSP);
737
738 __ mov((frame::varargs_offset)*wordSize -
739 1*Interpreter::stackElementSize+bias+BytesPerWord, G1);
740 // Jump to the interpreter just as if interpreter was doing it.
741 __ jmpl(G3_scratch, 0, G0);
742 // Setup Lesp for the call. Cannot actually set Lesp as the current Lesp
743 // (really L0) is in use by the compiled frame as a generic temp. However,
744 // the interpreter does not know where its args are without some kind of
745 // arg pointer being passed in. Pass it in Gargs.
746 __ delayed()->add(SP, G1, Gargs);
747 }
748
749 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg, Register temp2_reg,
750 address code_start, address code_end,
751 Label& L_ok) {
752 Label L_fail;
753 __ set(ExternalAddress(code_start), temp_reg);
754 __ set(pointer_delta(code_end, code_start, 1), temp2_reg);
755 __ cmp(pc_reg, temp_reg);
756 __ brx(Assembler::lessEqualUnsigned, false, Assembler::pn, L_fail);
757 __ delayed()->add(temp_reg, temp2_reg, temp_reg);
758 __ cmp(pc_reg, temp_reg);
759 __ cmp_and_brx_short(pc_reg, temp_reg, Assembler::lessUnsigned, Assembler::pt, L_ok);
760 __ bind(L_fail);
761 }
762
763 void AdapterGenerator::gen_i2c_adapter(
764 int total_args_passed,
765 // VMReg max_arg,
766 int comp_args_on_stack, // VMRegStackSlots
767 const BasicType *sig_bt,
768 const VMRegPair *regs) {
769
770 // Generate an I2C adapter: adjust the I-frame to make space for the C-frame
771 // layout. Lesp was saved by the calling I-frame and will be restored on
772 // return. Meanwhile, outgoing arg space is all owned by the callee
773 // C-frame, so we can mangle it at will. After adjusting the frame size,
774 // hoist register arguments and repack other args according to the compiled
775 // code convention. Finally, end in a jump to the compiled code. The entry
776 // point address is the start of the buffer.
777
778 // We will only enter here from an interpreted frame and never from after
779 // passing thru a c2i. Azul allowed this but we do not. If we lose the
780 // race and use a c2i we will remain interpreted for the race loser(s).
781 // This removes all sorts of headaches on the x86 side and also eliminates
782 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
783
784 // More detail:
785 // Adapters can be frameless because they do not require the caller
786 // to perform additional cleanup work, such as correcting the stack pointer.
787 // An i2c adapter is frameless because the *caller* frame, which is interpreted,
788 // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
789 // even if a callee has modified the stack pointer.
790 // A c2i adapter is frameless because the *callee* frame, which is interpreted,
791 // routinely repairs its caller's stack pointer (from sender_sp, which is set
792 // up via the senderSP register).
793 // In other words, if *either* the caller or callee is interpreted, we can
794 // get the stack pointer repaired after a call.
795 // This is why c2i and i2c adapters cannot be indefinitely composed.
796 // In particular, if a c2i adapter were to somehow call an i2c adapter,
797 // both caller and callee would be compiled methods, and neither would
798 // clean up the stack pointer changes performed by the two adapters.
799 // If this happens, control eventually transfers back to the compiled
800 // caller, but with an uncorrected stack, causing delayed havoc.
801
802 if (VerifyAdapterCalls &&
803 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) {
804 // So, let's test for cascading c2i/i2c adapters right now.
805 // assert(Interpreter::contains($return_addr) ||
806 // StubRoutines::contains($return_addr),
807 // "i2c adapter must return to an interpreter frame");
808 __ block_comment("verify_i2c { ");
809 Label L_ok;
810 if (Interpreter::code() != NULL)
811 range_check(masm, O7, O0, O1,
812 Interpreter::code()->code_start(), Interpreter::code()->code_end(),
813 L_ok);
814 if (StubRoutines::code1() != NULL)
815 range_check(masm, O7, O0, O1,
816 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(),
817 L_ok);
818 if (StubRoutines::code2() != NULL)
819 range_check(masm, O7, O0, O1,
820 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(),
821 L_ok);
822 const char* msg = "i2c adapter must return to an interpreter frame";
823 __ block_comment(msg);
824 __ stop(msg);
825 __ bind(L_ok);
826 __ block_comment("} verify_i2ce ");
827 }
828
829 // As you can see from the list of inputs & outputs there are not a lot
830 // of temp registers to work with: mostly G1, G3 & G4.
831
832 // Inputs:
833 // G2_thread - TLS
834 // G5_method - Method oop
835 // G4 (Gargs) - Pointer to interpreter's args
836 // O0..O4 - free for scratch
837 // O5_savedSP - Caller's saved SP, to be restored if needed
838 // O6 - Current SP!
839 // O7 - Valid return address
840 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
841
842 // Outputs:
843 // G2_thread - TLS
844 // O0-O5 - Outgoing args in compiled layout
845 // O6 - Adjusted or restored SP
846 // O7 - Valid return address
847 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
848 // F0-F7 - more outgoing args
849
850
851 // Gargs is the incoming argument base, and also an outgoing argument.
852 __ sub(Gargs, BytesPerWord, Gargs);
853
854 // ON ENTRY TO THE CODE WE ARE MAKING, WE HAVE AN INTERPRETED FRAME
855 // WITH O7 HOLDING A VALID RETURN PC
856 //
857 // | |
858 // : java stack :
859 // | |
860 // +--------------+ <--- start of outgoing args
861 // | receiver | |
862 // : rest of args : |---size is java-arg-words
863 // | | |
864 // +--------------+ <--- O4_args (misaligned) and Lesp if prior is not C2I
865 // | | |
866 // : unused : |---Space for max Java stack, plus stack alignment
867 // | | |
868 // +--------------+ <--- SP + 16*wordsize
869 // | |
870 // : window :
871 // | |
872 // +--------------+ <--- SP
873
874 // WE REPACK THE STACK. We use the common calling convention layout as
875 // discovered by calling SharedRuntime::calling_convention. We assume it
876 // causes an arbitrary shuffle of memory, which may require some register
877 // temps to do the shuffle. We hope for (and optimize for) the case where
878 // temps are not needed. We may have to resize the stack slightly, in case
879 // we need alignment padding (32-bit interpreter can pass longs & doubles
880 // misaligned, but the compilers expect them aligned).
881 //
882 // | |
883 // : java stack :
884 // | |
885 // +--------------+ <--- start of outgoing args
886 // | pad, align | |
887 // +--------------+ |
888 // | ints, longs, | |
889 // | floats, | |---Outgoing stack args.
890 // : doubles : | First few args in registers.
891 // | | |
892 // +--------------+ <--- SP' + 16*wordsize
893 // | |
894 // : window :
895 // | |
896 // +--------------+ <--- SP'
897
898 // ON EXIT FROM THE CODE WE ARE MAKING, WE STILL HAVE AN INTERPRETED FRAME
899 // WITH O7 HOLDING A VALID RETURN PC - ITS JUST THAT THE ARGS ARE NOW SETUP
900 // FOR COMPILED CODE AND THE FRAME SLIGHTLY GROWN.
901
902 // Cut-out for having no stack args. Since up to 6 args are passed
903 // in registers, we will commonly have no stack args.
904 if (comp_args_on_stack > 0) {
905 // Convert VMReg stack slots to words.
906 int comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
907 // Round up to miminum stack alignment, in wordSize
908 comp_words_on_stack = round_to(comp_words_on_stack, 2);
909 // Now compute the distance from Lesp to SP. This calculation does not
910 // include the space for total_args_passed because Lesp has not yet popped
911 // the arguments.
912 __ sub(SP, (comp_words_on_stack)*wordSize, SP);
913 }
914
915 // Now generate the shuffle code. Pick up all register args and move the
916 // rest through G1_scratch.
917 for (int i = 0; i < total_args_passed; i++) {
918 if (sig_bt[i] == T_VOID) {
919 // Longs and doubles are passed in native word order, but misaligned
920 // in the 32-bit build.
921 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
922 continue;
923 }
924
925 // Pick up 0, 1 or 2 words from Lesp+offset. Assume mis-aligned in the
926 // 32-bit build and aligned in the 64-bit build. Look for the obvious
927 // ldx/lddf optimizations.
928
929 // Load in argument order going down.
930 const int ld_off = (total_args_passed-i)*Interpreter::stackElementSize;
931 set_Rdisp(G1_scratch);
932
933 VMReg r_1 = regs[i].first();
934 VMReg r_2 = regs[i].second();
935 if (!r_1->is_valid()) {
936 assert(!r_2->is_valid(), "");
937 continue;
938 }
939 if (r_1->is_stack()) { // Pretend stack targets are loaded into F8/F9
940 r_1 = F8->as_VMReg(); // as part of the load/store shuffle
941 if (r_2->is_valid()) r_2 = r_1->next();
942 }
943 if (r_1->is_Register()) { // Register argument
944 Register r = r_1->as_Register()->after_restore();
945 if (!r_2->is_valid()) {
946 __ ld(Gargs, arg_slot(ld_off), r);
947 } else {
948 #ifdef _LP64
949 // In V9, longs are given 2 64-bit slots in the interpreter, but the
950 // data is passed in only 1 slot.
951 RegisterOrConstant slot = (sig_bt[i] == T_LONG) ?
952 next_arg_slot(ld_off) : arg_slot(ld_off);
953 __ ldx(Gargs, slot, r);
954 #else
955 fatal("longs should be on stack");
956 #endif
957 }
958 } else {
959 assert(r_1->is_FloatRegister(), "");
960 if (!r_2->is_valid()) {
961 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_1->as_FloatRegister());
962 } else {
963 #ifdef _LP64
964 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
965 // data is passed in only 1 slot. This code also handles longs that
966 // are passed on the stack, but need a stack-to-stack move through a
967 // spare float register.
968 RegisterOrConstant slot = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
969 next_arg_slot(ld_off) : arg_slot(ld_off);
970 __ ldf(FloatRegisterImpl::D, Gargs, slot, r_1->as_FloatRegister());
971 #else
972 // Need to marshal 64-bit value from misaligned Lesp loads
973 __ ldf(FloatRegisterImpl::S, Gargs, next_arg_slot(ld_off), r_1->as_FloatRegister());
974 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_2->as_FloatRegister());
975 #endif
976 }
977 }
978 // Was the argument really intended to be on the stack, but was loaded
979 // into F8/F9?
980 if (regs[i].first()->is_stack()) {
981 assert(r_1->as_FloatRegister() == F8, "fix this code");
982 // Convert stack slot to an SP offset
983 int st_off = reg2offset(regs[i].first()) + STACK_BIAS;
984 // Store down the shuffled stack word. Target address _is_ aligned.
985 RegisterOrConstant slot = __ ensure_simm13_or_reg(st_off, Rdisp);
986 if (!r_2->is_valid()) __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), SP, slot);
987 else __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), SP, slot);
988 }
989 }
990
991 // Jump to the compiled code just as if compiled code was doing it.
992 __ ld_ptr(G5_method, in_bytes(Method::from_compiled_offset()), G3);
993
994 // 6243940 We might end up in handle_wrong_method if
995 // the callee is deoptimized as we race thru here. If that
996 // happens we don't want to take a safepoint because the
997 // caller frame will look interpreted and arguments are now
998 // "compiled" so it is much better to make this transition
999 // invisible to the stack walking code. Unfortunately if
1000 // we try and find the callee by normal means a safepoint
1001 // is possible. So we stash the desired callee in the thread
1002 // and the vm will find there should this case occur.
1003 Address callee_target_addr(G2_thread, JavaThread::callee_target_offset());
1004 __ st_ptr(G5_method, callee_target_addr);
1005
1006 if (StressNonEntrant) {
1007 // Open a big window for deopt failure
1008 __ save_frame(0);
1009 __ mov(G0, L0);
1010 Label loop;
1011 __ bind(loop);
1012 __ sub(L0, 1, L0);
1013 __ br_null_short(L0, Assembler::pt, loop);
1014 __ restore();
1015 }
1016
1017 __ jmpl(G3, 0, G0);
1018 __ delayed()->nop();
1019 }
1020
1021 // ---------------------------------------------------------------
1022 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1023 int total_args_passed,
1024 // VMReg max_arg,
1025 int comp_args_on_stack, // VMRegStackSlots
1026 const BasicType *sig_bt,
1027 const VMRegPair *regs,
1028 AdapterFingerPrint* fingerprint) {
1029 address i2c_entry = __ pc();
1030
1031 AdapterGenerator agen(masm);
1032
1033 agen.gen_i2c_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs);
1034
1035
1036 // -------------------------------------------------------------------------
1037 // Generate a C2I adapter. On entry we know G5 holds the Method*. The
1038 // args start out packed in the compiled layout. They need to be unpacked
1039 // into the interpreter layout. This will almost always require some stack
1040 // space. We grow the current (compiled) stack, then repack the args. We
1041 // finally end in a jump to the generic interpreter entry point. On exit
1042 // from the interpreter, the interpreter will restore our SP (lest the
1043 // compiled code, which relys solely on SP and not FP, get sick).
1044
1045 address c2i_unverified_entry = __ pc();
1046 Label L_skip_fixup;
1047 {
1048 Register R_temp = G1; // another scratch register
1049
1050 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
1051
1052 __ verify_oop(O0);
1053 __ load_klass(O0, G3_scratch);
1054
1055 __ ld_ptr(G5_method, CompiledICHolder::holder_klass_offset(), R_temp);
1056 __ cmp(G3_scratch, R_temp);
1057
1058 Label ok, ok2;
1059 __ brx(Assembler::equal, false, Assembler::pt, ok);
1060 __ delayed()->ld_ptr(G5_method, CompiledICHolder::holder_method_offset(), G5_method);
1061 __ jump_to(ic_miss, G3_scratch);
1062 __ delayed()->nop();
1063
1064 __ bind(ok);
1065 // Method might have been compiled since the call site was patched to
1066 // interpreted if that is the case treat it as a miss so we can get
1067 // the call site corrected.
1068 __ ld_ptr(G5_method, in_bytes(Method::code_offset()), G3_scratch);
1069 __ bind(ok2);
1070 __ br_null(G3_scratch, false, Assembler::pt, L_skip_fixup);
1071 __ delayed()->nop();
1072 __ jump_to(ic_miss, G3_scratch);
1073 __ delayed()->nop();
1074
1075 }
1076
1077 address c2i_entry = __ pc();
1078
1079 agen.gen_c2i_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs, L_skip_fixup);
1080
1081 __ flush();
1082 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry);
1083
1084 }
1085
1086 // Helper function for native calling conventions
1087 static VMReg int_stk_helper( int i ) {
1088 // Bias any stack based VMReg we get by ignoring the window area
1089 // but not the register parameter save area.
1090 //
1091 // This is strange for the following reasons. We'd normally expect
1092 // the calling convention to return an VMReg for a stack slot
1093 // completely ignoring any abi reserved area. C2 thinks of that
1094 // abi area as only out_preserve_stack_slots. This does not include
1095 // the area allocated by the C abi to store down integer arguments
1096 // because the java calling convention does not use it. So
1097 // since c2 assumes that there are only out_preserve_stack_slots
1098 // to bias the optoregs (which impacts VMRegs) when actually referencing any actual stack
1099 // location the c calling convention must add in this bias amount
1100 // to make up for the fact that the out_preserve_stack_slots is
1101 // insufficient for C calls. What a mess. I sure hope those 6
1102 // stack words were worth it on every java call!
1103
1104 // Another way of cleaning this up would be for out_preserve_stack_slots
1105 // to take a parameter to say whether it was C or java calling conventions.
1106 // Then things might look a little better (but not much).
1107
1108 int mem_parm_offset = i - SPARC_ARGS_IN_REGS_NUM;
1109 if( mem_parm_offset < 0 ) {
1110 return as_oRegister(i)->as_VMReg();
1111 } else {
1112 int actual_offset = (mem_parm_offset + frame::memory_parameter_word_sp_offset) * VMRegImpl::slots_per_word;
1113 // Now return a biased offset that will be correct when out_preserve_slots is added back in
1114 return VMRegImpl::stack2reg(actual_offset - SharedRuntime::out_preserve_stack_slots());
1115 }
1116 }
1117
1118
1119 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1120 VMRegPair *regs,
1121 int total_args_passed) {
1122
1123 // Return the number of VMReg stack_slots needed for the args.
1124 // This value does not include an abi space (like register window
1125 // save area).
1126
1127 // The native convention is V8 if !LP64
1128 // The LP64 convention is the V9 convention which is slightly more sane.
1129
1130 // We return the amount of VMReg stack slots we need to reserve for all
1131 // the arguments NOT counting out_preserve_stack_slots. Since we always
1132 // have space for storing at least 6 registers to memory we start with that.
1133 // See int_stk_helper for a further discussion.
1134 int max_stack_slots = (frame::varargs_offset * VMRegImpl::slots_per_word) - SharedRuntime::out_preserve_stack_slots();
1135
1136 #ifdef _LP64
1137 // V9 convention: All things "as-if" on double-wide stack slots.
1138 // Hoist any int/ptr/long's in the first 6 to int regs.
1139 // Hoist any flt/dbl's in the first 16 dbl regs.
1140 int j = 0; // Count of actual args, not HALVES
1141 for( int i=0; i<total_args_passed; i++, j++ ) {
1142 switch( sig_bt[i] ) {
1143 case T_BOOLEAN:
1144 case T_BYTE:
1145 case T_CHAR:
1146 case T_INT:
1147 case T_SHORT:
1148 regs[i].set1( int_stk_helper( j ) ); break;
1149 case T_LONG:
1150 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1151 case T_ADDRESS: // raw pointers, like current thread, for VM calls
1152 case T_ARRAY:
1153 case T_OBJECT:
1154 case T_METADATA:
1155 regs[i].set2( int_stk_helper( j ) );
1156 break;
1157 case T_FLOAT:
1158 if ( j < 16 ) {
1159 // V9ism: floats go in ODD registers
1160 regs[i].set1(as_FloatRegister(1 + (j<<1))->as_VMReg());
1161 } else {
1162 // V9ism: floats go in ODD stack slot
1163 regs[i].set1(VMRegImpl::stack2reg(1 + (j<<1)));
1164 }
1165 break;
1166 case T_DOUBLE:
1167 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1168 if ( j < 16 ) {
1169 // V9ism: doubles go in EVEN/ODD regs
1170 regs[i].set2(as_FloatRegister(j<<1)->as_VMReg());
1171 } else {
1172 // V9ism: doubles go in EVEN/ODD stack slots
1173 regs[i].set2(VMRegImpl::stack2reg(j<<1));
1174 }
1175 break;
1176 case T_VOID: regs[i].set_bad(); j--; break; // Do not count HALVES
1177 default:
1178 ShouldNotReachHere();
1179 }
1180 if (regs[i].first()->is_stack()) {
1181 int off = regs[i].first()->reg2stack();
1182 if (off > max_stack_slots) max_stack_slots = off;
1183 }
1184 if (regs[i].second()->is_stack()) {
1185 int off = regs[i].second()->reg2stack();
1186 if (off > max_stack_slots) max_stack_slots = off;
1187 }
1188 }
1189
1190 #else // _LP64
1191 // V8 convention: first 6 things in O-regs, rest on stack.
1192 // Alignment is willy-nilly.
1193 for( int i=0; i<total_args_passed; i++ ) {
1194 switch( sig_bt[i] ) {
1195 case T_ADDRESS: // raw pointers, like current thread, for VM calls
1196 case T_ARRAY:
1197 case T_BOOLEAN:
1198 case T_BYTE:
1199 case T_CHAR:
1200 case T_FLOAT:
1201 case T_INT:
1202 case T_OBJECT:
1203 case T_METADATA:
1204 case T_SHORT:
1205 regs[i].set1( int_stk_helper( i ) );
1206 break;
1207 case T_DOUBLE:
1208 case T_LONG:
1209 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1210 regs[i].set_pair( int_stk_helper( i+1 ), int_stk_helper( i ) );
1211 break;
1212 case T_VOID: regs[i].set_bad(); break;
1213 default:
1214 ShouldNotReachHere();
1215 }
1216 if (regs[i].first()->is_stack()) {
1217 int off = regs[i].first()->reg2stack();
1218 if (off > max_stack_slots) max_stack_slots = off;
1219 }
1220 if (regs[i].second()->is_stack()) {
1221 int off = regs[i].second()->reg2stack();
1222 if (off > max_stack_slots) max_stack_slots = off;
1223 }
1224 }
1225 #endif // _LP64
1226
1227 return round_to(max_stack_slots + 1, 2);
1228
1229 }
1230
1231 // Do we need to convert ints to longs for c calls?
1232 bool SharedRuntime::c_calling_convention_requires_ints_as_longs() {
1233 return false;
1234 }
1235
1236 // ---------------------------------------------------------------------------
1237 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1238 switch (ret_type) {
1239 case T_FLOAT:
1240 __ stf(FloatRegisterImpl::S, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS);
1241 break;
1242 case T_DOUBLE:
1243 __ stf(FloatRegisterImpl::D, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS);
1244 break;
1245 }
1246 }
1247
1248 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1249 switch (ret_type) {
1250 case T_FLOAT:
1251 __ ldf(FloatRegisterImpl::S, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS, F0);
1252 break;
1253 case T_DOUBLE:
1254 __ ldf(FloatRegisterImpl::D, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS, F0);
1255 break;
1256 }
1257 }
1258
1259 // Check and forward and pending exception. Thread is stored in
1260 // L7_thread_cache and possibly NOT in G2_thread. Since this is a native call, there
1261 // is no exception handler. We merely pop this frame off and throw the
1262 // exception in the caller's frame.
1263 static void check_forward_pending_exception(MacroAssembler *masm, Register Rex_oop) {
1264 Label L;
1265 __ br_null(Rex_oop, false, Assembler::pt, L);
1266 __ delayed()->mov(L7_thread_cache, G2_thread); // restore in case we have exception
1267 // Since this is a native call, we *know* the proper exception handler
1268 // without calling into the VM: it's the empty function. Just pop this
1269 // frame and then jump to forward_exception_entry; O7 will contain the
1270 // native caller's return PC.
1271 AddressLiteral exception_entry(StubRoutines::forward_exception_entry());
1272 __ jump_to(exception_entry, G3_scratch);
1273 __ delayed()->restore(); // Pop this frame off.
1274 __ bind(L);
1275 }
1276
1277 // A simple move of integer like type
1278 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1279 if (src.first()->is_stack()) {
1280 if (dst.first()->is_stack()) {
1281 // stack to stack
1282 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1283 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1284 } else {
1285 // stack to reg
1286 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1287 }
1288 } else if (dst.first()->is_stack()) {
1289 // reg to stack
1290 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1291 } else {
1292 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1293 }
1294 }
1295
1296 // On 64 bit we will store integer like items to the stack as
1297 // 64 bits items (sparc abi) even though java would only store
1298 // 32bits for a parameter. On 32bit it will simply be 32 bits
1299 // So this routine will do 32->32 on 32bit and 32->64 on 64bit
1300 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1301 if (src.first()->is_stack()) {
1302 if (dst.first()->is_stack()) {
1303 // stack to stack
1304 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1305 __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1306 } else {
1307 // stack to reg
1308 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1309 }
1310 } else if (dst.first()->is_stack()) {
1311 // reg to stack
1312 __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1313 } else {
1314 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1315 }
1316 }
1317
1318
1319 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1320 if (src.first()->is_stack()) {
1321 if (dst.first()->is_stack()) {
1322 // stack to stack
1323 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1324 __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1325 } else {
1326 // stack to reg
1327 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1328 }
1329 } else if (dst.first()->is_stack()) {
1330 // reg to stack
1331 __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1332 } else {
1333 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1334 }
1335 }
1336
1337
1338 // An oop arg. Must pass a handle not the oop itself
1339 static void object_move(MacroAssembler* masm,
1340 OopMap* map,
1341 int oop_handle_offset,
1342 int framesize_in_slots,
1343 VMRegPair src,
1344 VMRegPair dst,
1345 bool is_receiver,
1346 int* receiver_offset) {
1347
1348 // must pass a handle. First figure out the location we use as a handle
1349
1350 if (src.first()->is_stack()) {
1351 // Oop is already on the stack
1352 Register rHandle = dst.first()->is_stack() ? L5 : dst.first()->as_Register();
1353 __ add(FP, reg2offset(src.first()) + STACK_BIAS, rHandle);
1354 __ ld_ptr(rHandle, 0, L4);
1355 #ifdef _LP64
1356 __ movr( Assembler::rc_z, L4, G0, rHandle );
1357 #else
1358 __ tst( L4 );
1359 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
1360 #endif
1361 if (dst.first()->is_stack()) {
1362 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
1363 }
1364 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1365 if (is_receiver) {
1366 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
1367 }
1368 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
1369 } else {
1370 // Oop is in an input register pass we must flush it to the stack
1371 const Register rOop = src.first()->as_Register();
1372 const Register rHandle = L5;
1373 int oop_slot = rOop->input_number() * VMRegImpl::slots_per_word + oop_handle_offset;
1374 int offset = oop_slot*VMRegImpl::stack_slot_size;
1375 Label skip;
1376 __ st_ptr(rOop, SP, offset + STACK_BIAS);
1377 if (is_receiver) {
1378 *receiver_offset = oop_slot * VMRegImpl::stack_slot_size;
1379 }
1380 map->set_oop(VMRegImpl::stack2reg(oop_slot));
1381 __ add(SP, offset + STACK_BIAS, rHandle);
1382 #ifdef _LP64
1383 __ movr( Assembler::rc_z, rOop, G0, rHandle );
1384 #else
1385 __ tst( rOop );
1386 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
1387 #endif
1388
1389 if (dst.first()->is_stack()) {
1390 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
1391 } else {
1392 __ mov(rHandle, dst.first()->as_Register());
1393 }
1394 }
1395 }
1396
1397 // A float arg may have to do float reg int reg conversion
1398 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1399 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move");
1400
1401 if (src.first()->is_stack()) {
1402 if (dst.first()->is_stack()) {
1403 // stack to stack the easiest of the bunch
1404 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1405 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1406 } else {
1407 // stack to reg
1408 if (dst.first()->is_Register()) {
1409 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1410 } else {
1411 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1412 }
1413 }
1414 } else if (dst.first()->is_stack()) {
1415 // reg to stack
1416 if (src.first()->is_Register()) {
1417 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1418 } else {
1419 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1420 }
1421 } else {
1422 // reg to reg
1423 if (src.first()->is_Register()) {
1424 if (dst.first()->is_Register()) {
1425 // gpr -> gpr
1426 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1427 } else {
1428 // gpr -> fpr
1429 __ st(src.first()->as_Register(), FP, -4 + STACK_BIAS);
1430 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.first()->as_FloatRegister());
1431 }
1432 } else if (dst.first()->is_Register()) {
1433 // fpr -> gpr
1434 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), FP, -4 + STACK_BIAS);
1435 __ ld(FP, -4 + STACK_BIAS, dst.first()->as_Register());
1436 } else {
1437 // fpr -> fpr
1438 // In theory these overlap but the ordering is such that this is likely a nop
1439 if ( src.first() != dst.first()) {
1440 __ fmov(FloatRegisterImpl::S, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
1441 }
1442 }
1443 }
1444 }
1445
1446 static void split_long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1447 VMRegPair src_lo(src.first());
1448 VMRegPair src_hi(src.second());
1449 VMRegPair dst_lo(dst.first());
1450 VMRegPair dst_hi(dst.second());
1451 simple_move32(masm, src_lo, dst_lo);
1452 simple_move32(masm, src_hi, dst_hi);
1453 }
1454
1455 // A long move
1456 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1457
1458 // Do the simple ones here else do two int moves
1459 if (src.is_single_phys_reg() ) {
1460 if (dst.is_single_phys_reg()) {
1461 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1462 } else {
1463 // split src into two separate registers
1464 // Remember hi means hi address or lsw on sparc
1465 // Move msw to lsw
1466 if (dst.second()->is_reg()) {
1467 // MSW -> MSW
1468 __ srax(src.first()->as_Register(), 32, dst.first()->as_Register());
1469 // Now LSW -> LSW
1470 // this will only move lo -> lo and ignore hi
1471 VMRegPair split(dst.second());
1472 simple_move32(masm, src, split);
1473 } else {
1474 VMRegPair split(src.first(), L4->as_VMReg());
1475 // MSW -> MSW (lo ie. first word)
1476 __ srax(src.first()->as_Register(), 32, L4);
1477 split_long_move(masm, split, dst);
1478 }
1479 }
1480 } else if (dst.is_single_phys_reg()) {
1481 if (src.is_adjacent_aligned_on_stack(2)) {
1482 __ ldx(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1483 } else {
1484 // dst is a single reg.
1485 // Remember lo is low address not msb for stack slots
1486 // and lo is the "real" register for registers
1487 // src is
1488
1489 VMRegPair split;
1490
1491 if (src.first()->is_reg()) {
1492 // src.lo (msw) is a reg, src.hi is stk/reg
1493 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> src.lo [the MSW is in the LSW of the reg]
1494 split.set_pair(dst.first(), src.first());
1495 } else {
1496 // msw is stack move to L5
1497 // lsw is stack move to dst.lo (real reg)
1498 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> L5
1499 split.set_pair(dst.first(), L5->as_VMReg());
1500 }
1501
1502 // src.lo -> src.lo/L5, src.hi -> dst.lo (the real reg)
1503 // msw -> src.lo/L5, lsw -> dst.lo
1504 split_long_move(masm, src, split);
1505
1506 // So dst now has the low order correct position the
1507 // msw half
1508 __ sllx(split.first()->as_Register(), 32, L5);
1509
1510 const Register d = dst.first()->as_Register();
1511 __ or3(L5, d, d);
1512 }
1513 } else {
1514 // For LP64 we can probably do better.
1515 split_long_move(masm, src, dst);
1516 }
1517 }
1518
1519 // A double move
1520 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1521
1522 // The painful thing here is that like long_move a VMRegPair might be
1523 // 1: a single physical register
1524 // 2: two physical registers (v8)
1525 // 3: a physical reg [lo] and a stack slot [hi] (v8)
1526 // 4: two stack slots
1527
1528 // Since src is always a java calling convention we know that the src pair
1529 // is always either all registers or all stack (and aligned?)
1530
1531 // in a register [lo] and a stack slot [hi]
1532 if (src.first()->is_stack()) {
1533 if (dst.first()->is_stack()) {
1534 // stack to stack the easiest of the bunch
1535 // ought to be a way to do this where if alignment is ok we use ldd/std when possible
1536 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1537 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1538 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1539 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1540 } else {
1541 // stack to reg
1542 if (dst.second()->is_stack()) {
1543 // stack -> reg, stack -> stack
1544 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1545 if (dst.first()->is_Register()) {
1546 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1547 } else {
1548 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1549 }
1550 // This was missing. (very rare case)
1551 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1552 } else {
1553 // stack -> reg
1554 // Eventually optimize for alignment QQQ
1555 if (dst.first()->is_Register()) {
1556 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1557 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_Register());
1558 } else {
1559 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1560 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_FloatRegister());
1561 }
1562 }
1563 }
1564 } else if (dst.first()->is_stack()) {
1565 // reg to stack
1566 if (src.first()->is_Register()) {
1567 // Eventually optimize for alignment QQQ
1568 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1569 if (src.second()->is_stack()) {
1570 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1571 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1572 } else {
1573 __ st(src.second()->as_Register(), SP, reg2offset(dst.second()) + STACK_BIAS);
1574 }
1575 } else {
1576 // fpr to stack
1577 if (src.second()->is_stack()) {
1578 ShouldNotReachHere();
1579 } else {
1580 // Is the stack aligned?
1581 if (reg2offset(dst.first()) & 0x7) {
1582 // No do as pairs
1583 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1584 __ stf(FloatRegisterImpl::S, src.second()->as_FloatRegister(), SP, reg2offset(dst.second()) + STACK_BIAS);
1585 } else {
1586 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1587 }
1588 }
1589 }
1590 } else {
1591 // reg to reg
1592 if (src.first()->is_Register()) {
1593 if (dst.first()->is_Register()) {
1594 // gpr -> gpr
1595 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1596 __ mov(src.second()->as_Register(), dst.second()->as_Register());
1597 } else {
1598 // gpr -> fpr
1599 // ought to be able to do a single store
1600 __ stx(src.first()->as_Register(), FP, -8 + STACK_BIAS);
1601 __ stx(src.second()->as_Register(), FP, -4 + STACK_BIAS);
1602 // ought to be able to do a single load
1603 __ ldf(FloatRegisterImpl::S, FP, -8 + STACK_BIAS, dst.first()->as_FloatRegister());
1604 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.second()->as_FloatRegister());
1605 }
1606 } else if (dst.first()->is_Register()) {
1607 // fpr -> gpr
1608 // ought to be able to do a single store
1609 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), FP, -8 + STACK_BIAS);
1610 // ought to be able to do a single load
1611 // REMEMBER first() is low address not LSB
1612 __ ld(FP, -8 + STACK_BIAS, dst.first()->as_Register());
1613 if (dst.second()->is_Register()) {
1614 __ ld(FP, -4 + STACK_BIAS, dst.second()->as_Register());
1615 } else {
1616 __ ld(FP, -4 + STACK_BIAS, L4);
1617 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1618 }
1619 } else {
1620 // fpr -> fpr
1621 // In theory these overlap but the ordering is such that this is likely a nop
1622 if ( src.first() != dst.first()) {
1623 __ fmov(FloatRegisterImpl::D, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
1624 }
1625 }
1626 }
1627 }
1628
1629 // Creates an inner frame if one hasn't already been created, and
1630 // saves a copy of the thread in L7_thread_cache
1631 static void create_inner_frame(MacroAssembler* masm, bool* already_created) {
1632 if (!*already_created) {
1633 __ save_frame(0);
1634 // Save thread in L7 (INNER FRAME); it crosses a bunch of VM calls below
1635 // Don't use save_thread because it smashes G2 and we merely want to save a
1636 // copy
1637 __ mov(G2_thread, L7_thread_cache);
1638 *already_created = true;
1639 }
1640 }
1641
1642
1643 static void save_or_restore_arguments(MacroAssembler* masm,
1644 const int stack_slots,
1645 const int total_in_args,
1646 const int arg_save_area,
1647 OopMap* map,
1648 VMRegPair* in_regs,
1649 BasicType* in_sig_bt) {
1650 // if map is non-NULL then the code should store the values,
1651 // otherwise it should load them.
1652 if (map != NULL) {
1653 // Fill in the map
1654 for (int i = 0; i < total_in_args; i++) {
1655 if (in_sig_bt[i] == T_ARRAY) {
1656 if (in_regs[i].first()->is_stack()) {
1657 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1658 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots));
1659 } else if (in_regs[i].first()->is_Register()) {
1660 map->set_oop(in_regs[i].first());
1661 } else {
1662 ShouldNotReachHere();
1663 }
1664 }
1665 }
1666 }
1667
1668 // Save or restore double word values
1669 int handle_index = 0;
1670 for (int i = 0; i < total_in_args; i++) {
1671 int slot = handle_index + arg_save_area;
1672 int offset = slot * VMRegImpl::stack_slot_size;
1673 if (in_sig_bt[i] == T_LONG && in_regs[i].first()->is_Register()) {
1674 const Register reg = in_regs[i].first()->as_Register();
1675 if (reg->is_global()) {
1676 handle_index += 2;
1677 assert(handle_index <= stack_slots, "overflow");
1678 if (map != NULL) {
1679 __ stx(reg, SP, offset + STACK_BIAS);
1680 } else {
1681 __ ldx(SP, offset + STACK_BIAS, reg);
1682 }
1683 }
1684 } else if (in_sig_bt[i] == T_DOUBLE && in_regs[i].first()->is_FloatRegister()) {
1685 handle_index += 2;
1686 assert(handle_index <= stack_slots, "overflow");
1687 if (map != NULL) {
1688 __ stf(FloatRegisterImpl::D, in_regs[i].first()->as_FloatRegister(), SP, offset + STACK_BIAS);
1689 } else {
1690 __ ldf(FloatRegisterImpl::D, SP, offset + STACK_BIAS, in_regs[i].first()->as_FloatRegister());
1691 }
1692 }
1693 }
1694 // Save floats
1695 for (int i = 0; i < total_in_args; i++) {
1696 int slot = handle_index + arg_save_area;
1697 int offset = slot * VMRegImpl::stack_slot_size;
1698 if (in_sig_bt[i] == T_FLOAT && in_regs[i].first()->is_FloatRegister()) {
1699 handle_index++;
1700 assert(handle_index <= stack_slots, "overflow");
1701 if (map != NULL) {
1702 __ stf(FloatRegisterImpl::S, in_regs[i].first()->as_FloatRegister(), SP, offset + STACK_BIAS);
1703 } else {
1704 __ ldf(FloatRegisterImpl::S, SP, offset + STACK_BIAS, in_regs[i].first()->as_FloatRegister());
1705 }
1706 }
1707 }
1708
1709 }
1710
1711
1712 // Check GC_locker::needs_gc and enter the runtime if it's true. This
1713 // keeps a new JNI critical region from starting until a GC has been
1714 // forced. Save down any oops in registers and describe them in an
1715 // OopMap.
1716 static void check_needs_gc_for_critical_native(MacroAssembler* masm,
1717 const int stack_slots,
1718 const int total_in_args,
1719 const int arg_save_area,
1720 OopMapSet* oop_maps,
1721 VMRegPair* in_regs,
1722 BasicType* in_sig_bt) {
1723 __ block_comment("check GC_locker::needs_gc");
1724 Label cont;
1725 AddressLiteral sync_state(GC_locker::needs_gc_address());
1726 __ load_bool_contents(sync_state, G3_scratch);
1727 __ cmp_zero_and_br(Assembler::equal, G3_scratch, cont);
1728 __ delayed()->nop();
1729
1730 // Save down any values that are live in registers and call into the
1731 // runtime to halt for a GC
1732 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1733 save_or_restore_arguments(masm, stack_slots, total_in_args,
1734 arg_save_area, map, in_regs, in_sig_bt);
1735
1736 __ mov(G2_thread, L7_thread_cache);
1737
1738 __ set_last_Java_frame(SP, noreg);
1739
1740 __ block_comment("block_for_jni_critical");
1741 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical), relocInfo::runtime_call_type);
1742 __ delayed()->mov(L7_thread_cache, O0);
1743 oop_maps->add_gc_map( __ offset(), map);
1744
1745 __ restore_thread(L7_thread_cache); // restore G2_thread
1746 __ reset_last_Java_frame();
1747
1748 // Reload all the register arguments
1749 save_or_restore_arguments(masm, stack_slots, total_in_args,
1750 arg_save_area, NULL, in_regs, in_sig_bt);
1751
1752 __ bind(cont);
1753 #ifdef ASSERT
1754 if (StressCriticalJNINatives) {
1755 // Stress register saving
1756 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1757 save_or_restore_arguments(masm, stack_slots, total_in_args,
1758 arg_save_area, map, in_regs, in_sig_bt);
1759 // Destroy argument registers
1760 for (int i = 0; i < total_in_args; i++) {
1761 if (in_regs[i].first()->is_Register()) {
1762 const Register reg = in_regs[i].first()->as_Register();
1763 if (reg->is_global()) {
1764 __ mov(G0, reg);
1765 }
1766 } else if (in_regs[i].first()->is_FloatRegister()) {
1767 __ fneg(FloatRegisterImpl::D, in_regs[i].first()->as_FloatRegister(), in_regs[i].first()->as_FloatRegister());
1768 }
1769 }
1770
1771 save_or_restore_arguments(masm, stack_slots, total_in_args,
1772 arg_save_area, NULL, in_regs, in_sig_bt);
1773 }
1774 #endif
1775 }
1776
1777 // Unpack an array argument into a pointer to the body and the length
1778 // if the array is non-null, otherwise pass 0 for both.
1779 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) {
1780 // Pass the length, ptr pair
1781 Label is_null, done;
1782 if (reg.first()->is_stack()) {
1783 VMRegPair tmp = reg64_to_VMRegPair(L2);
1784 // Load the arg up from the stack
1785 move_ptr(masm, reg, tmp);
1786 reg = tmp;
1787 }
1788 __ cmp(reg.first()->as_Register(), G0);
1789 __ brx(Assembler::equal, false, Assembler::pt, is_null);
1790 __ delayed()->add(reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type), L4);
1791 move_ptr(masm, reg64_to_VMRegPair(L4), body_arg);
1792 __ ld(reg.first()->as_Register(), arrayOopDesc::length_offset_in_bytes(), L4);
1793 move32_64(masm, reg64_to_VMRegPair(L4), length_arg);
1794 __ ba_short(done);
1795 __ bind(is_null);
1796 // Pass zeros
1797 move_ptr(masm, reg64_to_VMRegPair(G0), body_arg);
1798 move32_64(masm, reg64_to_VMRegPair(G0), length_arg);
1799 __ bind(done);
1800 }
1801
1802 static void verify_oop_args(MacroAssembler* masm,
1803 methodHandle method,
1804 const BasicType* sig_bt,
1805 const VMRegPair* regs) {
1806 Register temp_reg = G5_method; // not part of any compiled calling seq
1807 if (VerifyOops) {
1808 for (int i = 0; i < method->size_of_parameters(); i++) {
1809 if (sig_bt[i] == T_OBJECT ||
1810 sig_bt[i] == T_ARRAY) {
1811 VMReg r = regs[i].first();
1812 assert(r->is_valid(), "bad oop arg");
1813 if (r->is_stack()) {
1814 RegisterOrConstant ld_off = reg2offset(r) + STACK_BIAS;
1815 ld_off = __ ensure_simm13_or_reg(ld_off, temp_reg);
1816 __ ld_ptr(SP, ld_off, temp_reg);
1817 __ verify_oop(temp_reg);
1818 } else {
1819 __ verify_oop(r->as_Register());
1820 }
1821 }
1822 }
1823 }
1824 }
1825
1826 static void gen_special_dispatch(MacroAssembler* masm,
1827 methodHandle method,
1828 const BasicType* sig_bt,
1829 const VMRegPair* regs) {
1830 verify_oop_args(masm, method, sig_bt, regs);
1831 vmIntrinsics::ID iid = method->intrinsic_id();
1832
1833 // Now write the args into the outgoing interpreter space
1834 bool has_receiver = false;
1835 Register receiver_reg = noreg;
1836 int member_arg_pos = -1;
1837 Register member_reg = noreg;
1838 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1839 if (ref_kind != 0) {
1840 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1841 member_reg = G5_method; // known to be free at this point
1842 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1843 } else if (iid == vmIntrinsics::_invokeBasic) {
1844 has_receiver = true;
1845 } else {
1846 fatal(err_msg_res("unexpected intrinsic id %d", iid));
1847 }
1848
1849 if (member_reg != noreg) {
1850 // Load the member_arg into register, if necessary.
1851 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1852 VMReg r = regs[member_arg_pos].first();
1853 if (r->is_stack()) {
1854 RegisterOrConstant ld_off = reg2offset(r) + STACK_BIAS;
1855 ld_off = __ ensure_simm13_or_reg(ld_off, member_reg);
1856 __ ld_ptr(SP, ld_off, member_reg);
1857 } else {
1858 // no data motion is needed
1859 member_reg = r->as_Register();
1860 }
1861 }
1862
1863 if (has_receiver) {
1864 // Make sure the receiver is loaded into a register.
1865 assert(method->size_of_parameters() > 0, "oob");
1866 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1867 VMReg r = regs[0].first();
1868 assert(r->is_valid(), "bad receiver arg");
1869 if (r->is_stack()) {
1870 // Porting note: This assumes that compiled calling conventions always
1871 // pass the receiver oop in a register. If this is not true on some
1872 // platform, pick a temp and load the receiver from stack.
1873 fatal("receiver always in a register");
1874 receiver_reg = G3_scratch; // known to be free at this point
1875 RegisterOrConstant ld_off = reg2offset(r) + STACK_BIAS;
1876 ld_off = __ ensure_simm13_or_reg(ld_off, member_reg);
1877 __ ld_ptr(SP, ld_off, receiver_reg);
1878 } else {
1879 // no data motion is needed
1880 receiver_reg = r->as_Register();
1881 }
1882 }
1883
1884 // Figure out which address we are really jumping to:
1885 MethodHandles::generate_method_handle_dispatch(masm, iid,
1886 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1887 }
1888
1889 // ---------------------------------------------------------------------------
1890 // Generate a native wrapper for a given method. The method takes arguments
1891 // in the Java compiled code convention, marshals them to the native
1892 // convention (handlizes oops, etc), transitions to native, makes the call,
1893 // returns to java state (possibly blocking), unhandlizes any result and
1894 // returns.
1895 //
1896 // Critical native functions are a shorthand for the use of
1897 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1898 // functions. The wrapper is expected to unpack the arguments before
1899 // passing them to the callee and perform checks before and after the
1900 // native call to ensure that they GC_locker
1901 // lock_critical/unlock_critical semantics are followed. Some other
1902 // parts of JNI setup are skipped like the tear down of the JNI handle
1903 // block and the check for pending exceptions it's impossible for them
1904 // to be thrown.
1905 //
1906 // They are roughly structured like this:
1907 // if (GC_locker::needs_gc())
1908 // SharedRuntime::block_for_jni_critical();
1909 // tranistion to thread_in_native
1910 // unpack arrray arguments and call native entry point
1911 // check for safepoint in progress
1912 // check if any thread suspend flags are set
1913 // call into JVM and possible unlock the JNI critical
1914 // if a GC was suppressed while in the critical native.
1915 // transition back to thread_in_Java
1916 // return to caller
1917 //
1918 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1919 methodHandle method,
1920 int compile_id,
1921 BasicType* in_sig_bt,
1922 VMRegPair* in_regs,
1923 BasicType ret_type) {
1924 if (method->is_method_handle_intrinsic()) {
1925 vmIntrinsics::ID iid = method->intrinsic_id();
1926 intptr_t start = (intptr_t)__ pc();
1927 int vep_offset = ((intptr_t)__ pc()) - start;
1928 gen_special_dispatch(masm,
1929 method,
1930 in_sig_bt,
1931 in_regs);
1932 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1933 __ flush();
1934 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1935 return nmethod::new_native_nmethod(method,
1936 compile_id,
1937 masm->code(),
1938 vep_offset,
1939 frame_complete,
1940 stack_slots / VMRegImpl::slots_per_word,
1941 in_ByteSize(-1),
1942 in_ByteSize(-1),
1943 (OopMapSet*)NULL);
1944 }
1945 bool is_critical_native = true;
1946 address native_func = method->critical_native_function();
1947 if (native_func == NULL) {
1948 native_func = method->native_function();
1949 is_critical_native = false;
1950 }
1951 assert(native_func != NULL, "must have function");
1952
1953 // Native nmethod wrappers never take possesion of the oop arguments.
1954 // So the caller will gc the arguments. The only thing we need an
1955 // oopMap for is if the call is static
1956 //
1957 // An OopMap for lock (and class if static), and one for the VM call itself
1958 OopMapSet *oop_maps = new OopMapSet();
1959 intptr_t start = (intptr_t)__ pc();
1960
1961 // First thing make an ic check to see if we should even be here
1962 {
1963 Label L;
1964 const Register temp_reg = G3_scratch;
1965 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
1966 __ verify_oop(O0);
1967 __ load_klass(O0, temp_reg);
1968 __ cmp_and_brx_short(temp_reg, G5_inline_cache_reg, Assembler::equal, Assembler::pt, L);
1969
1970 __ jump_to(ic_miss, temp_reg);
1971 __ delayed()->nop();
1972 __ align(CodeEntryAlignment);
1973 __ bind(L);
1974 }
1975
1976 int vep_offset = ((intptr_t)__ pc()) - start;
1977
1978 #ifdef COMPILER1
1979 if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) {
1980 // Object.hashCode can pull the hashCode from the header word
1981 // instead of doing a full VM transition once it's been computed.
1982 // Since hashCode is usually polymorphic at call sites we can't do
1983 // this optimization at the call site without a lot of work.
1984 Label slowCase;
1985 Register receiver = O0;
1986 Register result = O0;
1987 Register header = G3_scratch;
1988 Register hash = G3_scratch; // overwrite header value with hash value
1989 Register mask = G1; // to get hash field from header
1990
1991 // Read the header and build a mask to get its hash field. Give up if the object is not unlocked.
1992 // We depend on hash_mask being at most 32 bits and avoid the use of
1993 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
1994 // vm: see markOop.hpp.
1995 __ ld_ptr(receiver, oopDesc::mark_offset_in_bytes(), header);
1996 __ sethi(markOopDesc::hash_mask, mask);
1997 __ btst(markOopDesc::unlocked_value, header);
1998 __ br(Assembler::zero, false, Assembler::pn, slowCase);
1999 if (UseBiasedLocking) {
2000 // Check if biased and fall through to runtime if so
2001 __ delayed()->nop();
2002 __ btst(markOopDesc::biased_lock_bit_in_place, header);
2003 __ br(Assembler::notZero, false, Assembler::pn, slowCase);
2004 }
2005 __ delayed()->or3(mask, markOopDesc::hash_mask & 0x3ff, mask);
2006
2007 // Check for a valid (non-zero) hash code and get its value.
2008 #ifdef _LP64
2009 __ srlx(header, markOopDesc::hash_shift, hash);
2010 #else
2011 __ srl(header, markOopDesc::hash_shift, hash);
2012 #endif
2013 __ andcc(hash, mask, hash);
2014 __ br(Assembler::equal, false, Assembler::pn, slowCase);
2015 __ delayed()->nop();
2016
2017 // leaf return.
2018 __ retl();
2019 __ delayed()->mov(hash, result);
2020 __ bind(slowCase);
2021 }
2022 #endif // COMPILER1
2023
2024
2025 // We have received a description of where all the java arg are located
2026 // on entry to the wrapper. We need to convert these args to where
2027 // the jni function will expect them. To figure out where they go
2028 // we convert the java signature to a C signature by inserting
2029 // the hidden arguments as arg[0] and possibly arg[1] (static method)
2030
2031 const int total_in_args = method->size_of_parameters();
2032 int total_c_args = total_in_args;
2033 int total_save_slots = 6 * VMRegImpl::slots_per_word;
2034 if (!is_critical_native) {
2035 total_c_args += 1;
2036 if (method->is_static()) {
2037 total_c_args++;
2038 }
2039 } else {
2040 for (int i = 0; i < total_in_args; i++) {
2041 if (in_sig_bt[i] == T_ARRAY) {
2042 // These have to be saved and restored across the safepoint
2043 total_c_args++;
2044 }
2045 }
2046 }
2047
2048 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2049 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2050 BasicType* in_elem_bt = NULL;
2051
2052 int argc = 0;
2053 if (!is_critical_native) {
2054 out_sig_bt[argc++] = T_ADDRESS;
2055 if (method->is_static()) {
2056 out_sig_bt[argc++] = T_OBJECT;
2057 }
2058
2059 for (int i = 0; i < total_in_args ; i++ ) {
2060 out_sig_bt[argc++] = in_sig_bt[i];
2061 }
2062 } else {
2063 Thread* THREAD = Thread::current();
2064 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args);
2065 SignatureStream ss(method->signature());
2066 for (int i = 0; i < total_in_args ; i++ ) {
2067 if (in_sig_bt[i] == T_ARRAY) {
2068 // Arrays are passed as int, elem* pair
2069 out_sig_bt[argc++] = T_INT;
2070 out_sig_bt[argc++] = T_ADDRESS;
2071 Symbol* atype = ss.as_symbol(CHECK_NULL);
2072 const char* at = atype->as_C_string();
2073 if (strlen(at) == 2) {
2074 assert(at[0] == '[', "must be");
2075 switch (at[1]) {
2076 case 'B': in_elem_bt[i] = T_BYTE; break;
2077 case 'C': in_elem_bt[i] = T_CHAR; break;
2078 case 'D': in_elem_bt[i] = T_DOUBLE; break;
2079 case 'F': in_elem_bt[i] = T_FLOAT; break;
2080 case 'I': in_elem_bt[i] = T_INT; break;
2081 case 'J': in_elem_bt[i] = T_LONG; break;
2082 case 'S': in_elem_bt[i] = T_SHORT; break;
2083 case 'Z': in_elem_bt[i] = T_BOOLEAN; break;
2084 default: ShouldNotReachHere();
2085 }
2086 }
2087 } else {
2088 out_sig_bt[argc++] = in_sig_bt[i];
2089 in_elem_bt[i] = T_VOID;
2090 }
2091 if (in_sig_bt[i] != T_VOID) {
2092 assert(in_sig_bt[i] == ss.type(), "must match");
2093 ss.next();
2094 }
2095 }
2096 }
2097
2098 // Now figure out where the args must be stored and how much stack space
2099 // they require (neglecting out_preserve_stack_slots but space for storing
2100 // the 1st six register arguments). It's weird see int_stk_helper.
2101 //
2102 int out_arg_slots;
2103 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
2104
2105 if (is_critical_native) {
2106 // Critical natives may have to call out so they need a save area
2107 // for register arguments.
2108 int double_slots = 0;
2109 int single_slots = 0;
2110 for ( int i = 0; i < total_in_args; i++) {
2111 if (in_regs[i].first()->is_Register()) {
2112 const Register reg = in_regs[i].first()->as_Register();
2113 switch (in_sig_bt[i]) {
2114 case T_ARRAY:
2115 case T_BOOLEAN:
2116 case T_BYTE:
2117 case T_SHORT:
2118 case T_CHAR:
2119 case T_INT: assert(reg->is_in(), "don't need to save these"); break;
2120 case T_LONG: if (reg->is_global()) double_slots++; break;
2121 default: ShouldNotReachHere();
2122 }
2123 } else if (in_regs[i].first()->is_FloatRegister()) {
2124 switch (in_sig_bt[i]) {
2125 case T_FLOAT: single_slots++; break;
2126 case T_DOUBLE: double_slots++; break;
2127 default: ShouldNotReachHere();
2128 }
2129 }
2130 }
2131 total_save_slots = double_slots * 2 + single_slots;
2132 }
2133
2134 // Compute framesize for the wrapper. We need to handlize all oops in
2135 // registers. We must create space for them here that is disjoint from
2136 // the windowed save area because we have no control over when we might
2137 // flush the window again and overwrite values that gc has since modified.
2138 // (The live window race)
2139 //
2140 // We always just allocate 6 word for storing down these object. This allow
2141 // us to simply record the base and use the Ireg number to decide which
2142 // slot to use. (Note that the reg number is the inbound number not the
2143 // outbound number).
2144 // We must shuffle args to match the native convention, and include var-args space.
2145
2146 // Calculate the total number of stack slots we will need.
2147
2148 // First count the abi requirement plus all of the outgoing args
2149 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2150
2151 // Now the space for the inbound oop handle area
2152
2153 int oop_handle_offset = round_to(stack_slots, 2);
2154 stack_slots += total_save_slots;
2155
2156 // Now any space we need for handlizing a klass if static method
2157
2158 int klass_slot_offset = 0;
2159 int klass_offset = -1;
2160 int lock_slot_offset = 0;
2161 bool is_static = false;
2162
2163 if (method->is_static()) {
2164 klass_slot_offset = stack_slots;
2165 stack_slots += VMRegImpl::slots_per_word;
2166 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
2167 is_static = true;
2168 }
2169
2170 // Plus a lock if needed
2171
2172 if (method->is_synchronized()) {
2173 lock_slot_offset = stack_slots;
2174 stack_slots += VMRegImpl::slots_per_word;
2175 }
2176
2177 // Now a place to save return value or as a temporary for any gpr -> fpr moves
2178 stack_slots += 2;
2179
2180 // Ok The space we have allocated will look like:
2181 //
2182 //
2183 // FP-> | |
2184 // |---------------------|
2185 // | 2 slots for moves |
2186 // |---------------------|
2187 // | lock box (if sync) |
2188 // |---------------------| <- lock_slot_offset
2189 // | klass (if static) |
2190 // |---------------------| <- klass_slot_offset
2191 // | oopHandle area |
2192 // |---------------------| <- oop_handle_offset
2193 // | outbound memory |
2194 // | based arguments |
2195 // | |
2196 // |---------------------|
2197 // | vararg area |
2198 // |---------------------|
2199 // | |
2200 // SP-> | out_preserved_slots |
2201 //
2202 //
2203
2204
2205 // Now compute actual number of stack words we need rounding to make
2206 // stack properly aligned.
2207 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word);
2208
2209 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2210
2211 // Generate stack overflow check before creating frame
2212 __ generate_stack_overflow_check(stack_size);
2213
2214 // Generate a new frame for the wrapper.
2215 __ save(SP, -stack_size, SP);
2216
2217 int frame_complete = ((intptr_t)__ pc()) - start;
2218
2219 __ verify_thread();
2220
2221 if (is_critical_native) {
2222 check_needs_gc_for_critical_native(masm, stack_slots, total_in_args,
2223 oop_handle_offset, oop_maps, in_regs, in_sig_bt);
2224 }
2225
2226 //
2227 // We immediately shuffle the arguments so that any vm call we have to
2228 // make from here on out (sync slow path, jvmti, etc.) we will have
2229 // captured the oops from our caller and have a valid oopMap for
2230 // them.
2231
2232 // -----------------
2233 // The Grand Shuffle
2234 //
2235 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2236 // (derived from JavaThread* which is in L7_thread_cache) and, if static,
2237 // the class mirror instead of a receiver. This pretty much guarantees that
2238 // register layout will not match. We ignore these extra arguments during
2239 // the shuffle. The shuffle is described by the two calling convention
2240 // vectors we have in our possession. We simply walk the java vector to
2241 // get the source locations and the c vector to get the destinations.
2242 // Because we have a new window and the argument registers are completely
2243 // disjoint ( I0 -> O1, I1 -> O2, ...) we have nothing to worry about
2244 // here.
2245
2246 // This is a trick. We double the stack slots so we can claim
2247 // the oops in the caller's frame. Since we are sure to have
2248 // more args than the caller doubling is enough to make
2249 // sure we can capture all the incoming oop args from the
2250 // caller.
2251 //
2252 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2253 // Record sp-based slot for receiver on stack for non-static methods
2254 int receiver_offset = -1;
2255
2256 // We move the arguments backward because the floating point registers
2257 // destination will always be to a register with a greater or equal register
2258 // number or the stack.
2259
2260 #ifdef ASSERT
2261 bool reg_destroyed[RegisterImpl::number_of_registers];
2262 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
2263 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
2264 reg_destroyed[r] = false;
2265 }
2266 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
2267 freg_destroyed[f] = false;
2268 }
2269
2270 #endif /* ASSERT */
2271
2272 for ( int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0 ; i--, c_arg-- ) {
2273
2274 #ifdef ASSERT
2275 if (in_regs[i].first()->is_Register()) {
2276 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "ack!");
2277 } else if (in_regs[i].first()->is_FloatRegister()) {
2278 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)], "ack!");
2279 }
2280 if (out_regs[c_arg].first()->is_Register()) {
2281 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
2282 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
2283 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)] = true;
2284 }
2285 #endif /* ASSERT */
2286
2287 switch (in_sig_bt[i]) {
2288 case T_ARRAY:
2289 if (is_critical_native) {
2290 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg], out_regs[c_arg - 1]);
2291 c_arg--;
2292 break;
2293 }
2294 case T_OBJECT:
2295 assert(!is_critical_native, "no oop arguments");
2296 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
2297 ((i == 0) && (!is_static)),
2298 &receiver_offset);
2299 break;
2300 case T_VOID:
2301 break;
2302
2303 case T_FLOAT:
2304 float_move(masm, in_regs[i], out_regs[c_arg]);
2305 break;
2306
2307 case T_DOUBLE:
2308 assert( i + 1 < total_in_args &&
2309 in_sig_bt[i + 1] == T_VOID &&
2310 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
2311 double_move(masm, in_regs[i], out_regs[c_arg]);
2312 break;
2313
2314 case T_LONG :
2315 long_move(masm, in_regs[i], out_regs[c_arg]);
2316 break;
2317
2318 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2319
2320 default:
2321 move32_64(masm, in_regs[i], out_regs[c_arg]);
2322 }
2323 }
2324
2325 // Pre-load a static method's oop into O1. Used both by locking code and
2326 // the normal JNI call code.
2327 if (method->is_static() && !is_critical_native) {
2328 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()), O1);
2329
2330 // Now handlize the static class mirror in O1. It's known not-null.
2331 __ st_ptr(O1, SP, klass_offset + STACK_BIAS);
2332 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2333 __ add(SP, klass_offset + STACK_BIAS, O1);
2334 }
2335
2336
2337 const Register L6_handle = L6;
2338
2339 if (method->is_synchronized()) {
2340 assert(!is_critical_native, "unhandled");
2341 __ mov(O1, L6_handle);
2342 }
2343
2344 // We have all of the arguments setup at this point. We MUST NOT touch any Oregs
2345 // except O6/O7. So if we must call out we must push a new frame. We immediately
2346 // push a new frame and flush the windows.
2347 #ifdef _LP64
2348 intptr_t thepc = (intptr_t) __ pc();
2349 {
2350 address here = __ pc();
2351 // Call the next instruction
2352 __ call(here + 8, relocInfo::none);
2353 __ delayed()->nop();
2354 }
2355 #else
2356 intptr_t thepc = __ load_pc_address(O7, 0);
2357 #endif /* _LP64 */
2358
2359 // We use the same pc/oopMap repeatedly when we call out
2360 oop_maps->add_gc_map(thepc - start, map);
2361
2362 // O7 now has the pc loaded that we will use when we finally call to native.
2363
2364 // Save thread in L7; it crosses a bunch of VM calls below
2365 // Don't use save_thread because it smashes G2 and we merely
2366 // want to save a copy
2367 __ mov(G2_thread, L7_thread_cache);
2368
2369
2370 // If we create an inner frame once is plenty
2371 // when we create it we must also save G2_thread
2372 bool inner_frame_created = false;
2373
2374 // dtrace method entry support
2375 {
2376 SkipIfEqual skip_if(
2377 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
2378 // create inner frame
2379 __ save_frame(0);
2380 __ mov(G2_thread, L7_thread_cache);
2381 __ set_metadata_constant(method(), O1);
2382 __ call_VM_leaf(L7_thread_cache,
2383 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2384 G2_thread, O1);
2385 __ restore();
2386 }
2387
2388 // RedefineClasses() tracing support for obsolete method entry
2389 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) {
2390 // create inner frame
2391 __ save_frame(0);
2392 __ mov(G2_thread, L7_thread_cache);
2393 __ set_metadata_constant(method(), O1);
2394 __ call_VM_leaf(L7_thread_cache,
2395 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
2396 G2_thread, O1);
2397 __ restore();
2398 }
2399
2400 // We are in the jni frame unless saved_frame is true in which case
2401 // we are in one frame deeper (the "inner" frame). If we are in the
2402 // "inner" frames the args are in the Iregs and if the jni frame then
2403 // they are in the Oregs.
2404 // If we ever need to go to the VM (for locking, jvmti) then
2405 // we will always be in the "inner" frame.
2406
2407 // Lock a synchronized method
2408 int lock_offset = -1; // Set if locked
2409 if (method->is_synchronized()) {
2410 Register Roop = O1;
2411 const Register L3_box = L3;
2412
2413 create_inner_frame(masm, &inner_frame_created);
2414
2415 __ ld_ptr(I1, 0, O1);
2416 Label done;
2417
2418 lock_offset = (lock_slot_offset * VMRegImpl::stack_slot_size);
2419 __ add(FP, lock_offset+STACK_BIAS, L3_box);
2420 #ifdef ASSERT
2421 if (UseBiasedLocking) {
2422 // making the box point to itself will make it clear it went unused
2423 // but also be obviously invalid
2424 __ st_ptr(L3_box, L3_box, 0);
2425 }
2426 #endif // ASSERT
2427 //
2428 // Compiler_lock_object (Roop, Rmark, Rbox, Rscratch) -- kills Rmark, Rbox, Rscratch
2429 //
2430 __ compiler_lock_object(Roop, L1, L3_box, L2);
2431 __ br(Assembler::equal, false, Assembler::pt, done);
2432 __ delayed() -> add(FP, lock_offset+STACK_BIAS, L3_box);
2433
2434
2435 // None of the above fast optimizations worked so we have to get into the
2436 // slow case of monitor enter. Inline a special case of call_VM that
2437 // disallows any pending_exception.
2438 __ mov(Roop, O0); // Need oop in O0
2439 __ mov(L3_box, O1);
2440
2441 // Record last_Java_sp, in case the VM code releases the JVM lock.
2442
2443 __ set_last_Java_frame(FP, I7);
2444
2445 // do the call
2446 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), relocInfo::runtime_call_type);
2447 __ delayed()->mov(L7_thread_cache, O2);
2448
2449 __ restore_thread(L7_thread_cache); // restore G2_thread
2450 __ reset_last_Java_frame();
2451
2452 #ifdef ASSERT
2453 { Label L;
2454 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
2455 __ br_null_short(O0, Assembler::pt, L);
2456 __ stop("no pending exception allowed on exit from IR::monitorenter");
2457 __ bind(L);
2458 }
2459 #endif
2460 __ bind(done);
2461 }
2462
2463
2464 // Finally just about ready to make the JNI call
2465
2466 __ flushw();
2467 if (inner_frame_created) {
2468 __ restore();
2469 } else {
2470 // Store only what we need from this frame
2471 // QQQ I think that non-v9 (like we care) we don't need these saves
2472 // either as the flush traps and the current window goes too.
2473 __ st_ptr(FP, SP, FP->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
2474 __ st_ptr(I7, SP, I7->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
2475 }
2476
2477 // get JNIEnv* which is first argument to native
2478 if (!is_critical_native) {
2479 __ add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
2480 }
2481
2482 // Use that pc we placed in O7 a while back as the current frame anchor
2483 __ set_last_Java_frame(SP, O7);
2484
2485 // We flushed the windows ages ago now mark them as flushed before transitioning.
2486 __ set(JavaFrameAnchor::flushed, G3_scratch);
2487 __ st(G3_scratch, G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
2488
2489 // Transition from _thread_in_Java to _thread_in_native.
2490 __ set(_thread_in_native, G3_scratch);
2491
2492 #ifdef _LP64
2493 AddressLiteral dest(native_func);
2494 __ relocate(relocInfo::runtime_call_type);
2495 __ jumpl_to(dest, O7, O7);
2496 #else
2497 __ call(native_func, relocInfo::runtime_call_type);
2498 #endif
2499 __ delayed()->st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2500
2501 __ restore_thread(L7_thread_cache); // restore G2_thread
2502
2503 // Unpack native results. For int-types, we do any needed sign-extension
2504 // and move things into I0. The return value there will survive any VM
2505 // calls for blocking or unlocking. An FP or OOP result (handle) is done
2506 // specially in the slow-path code.
2507 switch (ret_type) {
2508 case T_VOID: break; // Nothing to do!
2509 case T_FLOAT: break; // Got it where we want it (unless slow-path)
2510 case T_DOUBLE: break; // Got it where we want it (unless slow-path)
2511 // In 64 bits build result is in O0, in O0, O1 in 32bit build
2512 case T_LONG:
2513 #ifndef _LP64
2514 __ mov(O1, I1);
2515 #endif
2516 // Fall thru
2517 case T_OBJECT: // Really a handle
2518 case T_ARRAY:
2519 case T_INT:
2520 __ mov(O0, I0);
2521 break;
2522 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, I0); break; // !0 => true; 0 => false
2523 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, I0); break;
2524 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, I0); break; // cannot use and3, 0xFFFF too big as immediate value!
2525 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, I0); break;
2526 break; // Cannot de-handlize until after reclaiming jvm_lock
2527 default:
2528 ShouldNotReachHere();
2529 }
2530
2531 Label after_transition;
2532 // must we block?
2533
2534 // Block, if necessary, before resuming in _thread_in_Java state.
2535 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2536 { Label no_block;
2537 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
2538
2539 // Switch thread to "native transition" state before reading the synchronization state.
2540 // This additional state is necessary because reading and testing the synchronization
2541 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2542 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2543 // VM thread changes sync state to synchronizing and suspends threads for GC.
2544 // Thread A is resumed to finish this native method, but doesn't block here since it
2545 // didn't see any synchronization is progress, and escapes.
2546 __ set(_thread_in_native_trans, G3_scratch);
2547 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2548 if(os::is_MP()) {
2549 if (UseMembar) {
2550 // Force this write out before the read below
2551 __ membar(Assembler::StoreLoad);
2552 } else {
2553 // Write serialization page so VM thread can do a pseudo remote membar.
2554 // We use the current thread pointer to calculate a thread specific
2555 // offset to write to within the page. This minimizes bus traffic
2556 // due to cache line collision.
2557 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
2558 }
2559 }
2560 __ load_contents(sync_state, G3_scratch);
2561 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
2562
2563 Label L;
2564 Address suspend_state(G2_thread, JavaThread::suspend_flags_offset());
2565 __ br(Assembler::notEqual, false, Assembler::pn, L);
2566 __ delayed()->ld(suspend_state, G3_scratch);
2567 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
2568 __ bind(L);
2569
2570 // Block. Save any potential method result value before the operation and
2571 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2572 // lets us share the oopMap we used when we went native rather the create
2573 // a distinct one for this pc
2574 //
2575 save_native_result(masm, ret_type, stack_slots);
2576 if (!is_critical_native) {
2577 __ call_VM_leaf(L7_thread_cache,
2578 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
2579 G2_thread);
2580 } else {
2581 __ call_VM_leaf(L7_thread_cache,
2582 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition),
2583 G2_thread);
2584 }
2585
2586 // Restore any method result value
2587 restore_native_result(masm, ret_type, stack_slots);
2588
2589 if (is_critical_native) {
2590 // The call above performed the transition to thread_in_Java so
2591 // skip the transition logic below.
2592 __ ba(after_transition);
2593 __ delayed()->nop();
2594 }
2595
2596 __ bind(no_block);
2597 }
2598
2599 // thread state is thread_in_native_trans. Any safepoint blocking has already
2600 // happened so we can now change state to _thread_in_Java.
2601 __ set(_thread_in_Java, G3_scratch);
2602 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2603 __ bind(after_transition);
2604
2605 Label no_reguard;
2606 __ ld(G2_thread, JavaThread::stack_guard_state_offset(), G3_scratch);
2607 __ cmp_and_br_short(G3_scratch, JavaThread::stack_guard_yellow_disabled, Assembler::notEqual, Assembler::pt, no_reguard);
2608
2609 save_native_result(masm, ret_type, stack_slots);
2610 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2611 __ delayed()->nop();
2612
2613 __ restore_thread(L7_thread_cache); // restore G2_thread
2614 restore_native_result(masm, ret_type, stack_slots);
2615
2616 __ bind(no_reguard);
2617
2618 // Handle possible exception (will unlock if necessary)
2619
2620 // native result if any is live in freg or I0 (and I1 if long and 32bit vm)
2621
2622 // Unlock
2623 if (method->is_synchronized()) {
2624 Label done;
2625 Register I2_ex_oop = I2;
2626 const Register L3_box = L3;
2627 // Get locked oop from the handle we passed to jni
2628 __ ld_ptr(L6_handle, 0, L4);
2629 __ add(SP, lock_offset+STACK_BIAS, L3_box);
2630 // Must save pending exception around the slow-path VM call. Since it's a
2631 // leaf call, the pending exception (if any) can be kept in a register.
2632 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), I2_ex_oop);
2633 // Now unlock
2634 // (Roop, Rmark, Rbox, Rscratch)
2635 __ compiler_unlock_object(L4, L1, L3_box, L2);
2636 __ br(Assembler::equal, false, Assembler::pt, done);
2637 __ delayed()-> add(SP, lock_offset+STACK_BIAS, L3_box);
2638
2639 // save and restore any potential method result value around the unlocking
2640 // operation. Will save in I0 (or stack for FP returns).
2641 save_native_result(masm, ret_type, stack_slots);
2642
2643 // Must clear pending-exception before re-entering the VM. Since this is
2644 // a leaf call, pending-exception-oop can be safely kept in a register.
2645 __ st_ptr(G0, G2_thread, in_bytes(Thread::pending_exception_offset()));
2646
2647 // slow case of monitor enter. Inline a special case of call_VM that
2648 // disallows any pending_exception.
2649 __ mov(L3_box, O1);
2650
2651 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), relocInfo::runtime_call_type);
2652 __ delayed()->mov(L4, O0); // Need oop in O0
2653
2654 __ restore_thread(L7_thread_cache); // restore G2_thread
2655
2656 #ifdef ASSERT
2657 { Label L;
2658 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
2659 __ br_null_short(O0, Assembler::pt, L);
2660 __ stop("no pending exception allowed on exit from IR::monitorexit");
2661 __ bind(L);
2662 }
2663 #endif
2664 restore_native_result(masm, ret_type, stack_slots);
2665 // check_forward_pending_exception jump to forward_exception if any pending
2666 // exception is set. The forward_exception routine expects to see the
2667 // exception in pending_exception and not in a register. Kind of clumsy,
2668 // since all folks who branch to forward_exception must have tested
2669 // pending_exception first and hence have it in a register already.
2670 __ st_ptr(I2_ex_oop, G2_thread, in_bytes(Thread::pending_exception_offset()));
2671 __ bind(done);
2672 }
2673
2674 // Tell dtrace about this method exit
2675 {
2676 SkipIfEqual skip_if(
2677 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
2678 save_native_result(masm, ret_type, stack_slots);
2679 __ set_metadata_constant(method(), O1);
2680 __ call_VM_leaf(L7_thread_cache,
2681 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2682 G2_thread, O1);
2683 restore_native_result(masm, ret_type, stack_slots);
2684 }
2685
2686 // Clear "last Java frame" SP and PC.
2687 __ verify_thread(); // G2_thread must be correct
2688 __ reset_last_Java_frame();
2689
2690 // Unpack oop result
2691 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
2692 Label L;
2693 __ addcc(G0, I0, G0);
2694 __ brx(Assembler::notZero, true, Assembler::pt, L);
2695 __ delayed()->ld_ptr(I0, 0, I0);
2696 __ mov(G0, I0);
2697 __ bind(L);
2698 __ verify_oop(I0);
2699 }
2700
2701 if (!is_critical_native) {
2702 // reset handle block
2703 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), L5);
2704 __ st_ptr(G0, L5, JNIHandleBlock::top_offset_in_bytes());
2705
2706 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), G3_scratch);
2707 check_forward_pending_exception(masm, G3_scratch);
2708 }
2709
2710
2711 // Return
2712
2713 #ifndef _LP64
2714 if (ret_type == T_LONG) {
2715
2716 // Must leave proper result in O0,O1 and G1 (c2/tiered only)
2717 __ sllx(I0, 32, G1); // Shift bits into high G1
2718 __ srl (I1, 0, I1); // Zero extend O1 (harmless?)
2719 __ or3 (I1, G1, G1); // OR 64 bits into G1
2720 }
2721 #endif
2722
2723 __ ret();
2724 __ delayed()->restore();
2725
2726 __ flush();
2727
2728 nmethod *nm = nmethod::new_native_nmethod(method,
2729 compile_id,
2730 masm->code(),
2731 vep_offset,
2732 frame_complete,
2733 stack_slots / VMRegImpl::slots_per_word,
2734 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2735 in_ByteSize(lock_offset),
2736 oop_maps);
2737
2738 if (is_critical_native) {
2739 nm->set_lazy_critical_native(true);
2740 }
2741 return nm;
2742
2743 }
2744
2745 #ifdef HAVE_DTRACE_H
2746 // ---------------------------------------------------------------------------
2747 // Generate a dtrace nmethod for a given signature. The method takes arguments
2748 // in the Java compiled code convention, marshals them to the native
2749 // abi and then leaves nops at the position you would expect to call a native
2750 // function. When the probe is enabled the nops are replaced with a trap
2751 // instruction that dtrace inserts and the trace will cause a notification
2752 // to dtrace.
2753 //
2754 // The probes are only able to take primitive types and java/lang/String as
2755 // arguments. No other java types are allowed. Strings are converted to utf8
2756 // strings so that from dtrace point of view java strings are converted to C
2757 // strings. There is an arbitrary fixed limit on the total space that a method
2758 // can use for converting the strings. (256 chars per string in the signature).
2759 // So any java string larger then this is truncated.
2760
2761 static int fp_offset[ConcreteRegisterImpl::number_of_registers] = { 0 };
2762 static bool offsets_initialized = false;
2763
2764 nmethod *SharedRuntime::generate_dtrace_nmethod(
2765 MacroAssembler *masm, methodHandle method) {
2766
2767
2768 // generate_dtrace_nmethod is guarded by a mutex so we are sure to
2769 // be single threaded in this method.
2770 assert(AdapterHandlerLibrary_lock->owned_by_self(), "must be");
2771
2772 // Fill in the signature array, for the calling-convention call.
2773 int total_args_passed = method->size_of_parameters();
2774
2775 BasicType* in_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed);
2776 VMRegPair *in_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed);
2777
2778 // The signature we are going to use for the trap that dtrace will see
2779 // java/lang/String is converted. We drop "this" and any other object
2780 // is converted to NULL. (A one-slot java/lang/Long object reference
2781 // is converted to a two-slot long, which is why we double the allocation).
2782 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed * 2);
2783 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed * 2);
2784
2785 int i=0;
2786 int total_strings = 0;
2787 int first_arg_to_pass = 0;
2788 int total_c_args = 0;
2789
2790 // Skip the receiver as dtrace doesn't want to see it
2791 if( !method->is_static() ) {
2792 in_sig_bt[i++] = T_OBJECT;
2793 first_arg_to_pass = 1;
2794 }
2795
2796 SignatureStream ss(method->signature());
2797 for ( ; !ss.at_return_type(); ss.next()) {
2798 BasicType bt = ss.type();
2799 in_sig_bt[i++] = bt; // Collect remaining bits of signature
2800 out_sig_bt[total_c_args++] = bt;
2801 if( bt == T_OBJECT) {
2802 Symbol* s = ss.as_symbol_or_null();
2803 if (s == vmSymbols::java_lang_String()) {
2804 total_strings++;
2805 out_sig_bt[total_c_args-1] = T_ADDRESS;
2806 } else if (s == vmSymbols::java_lang_Boolean() ||
2807 s == vmSymbols::java_lang_Byte()) {
2808 out_sig_bt[total_c_args-1] = T_BYTE;
2809 } else if (s == vmSymbols::java_lang_Character() ||
2810 s == vmSymbols::java_lang_Short()) {
2811 out_sig_bt[total_c_args-1] = T_SHORT;
2812 } else if (s == vmSymbols::java_lang_Integer() ||
2813 s == vmSymbols::java_lang_Float()) {
2814 out_sig_bt[total_c_args-1] = T_INT;
2815 } else if (s == vmSymbols::java_lang_Long() ||
2816 s == vmSymbols::java_lang_Double()) {
2817 out_sig_bt[total_c_args-1] = T_LONG;
2818 out_sig_bt[total_c_args++] = T_VOID;
2819 }
2820 } else if ( bt == T_LONG || bt == T_DOUBLE ) {
2821 in_sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots
2822 // We convert double to long
2823 out_sig_bt[total_c_args-1] = T_LONG;
2824 out_sig_bt[total_c_args++] = T_VOID;
2825 } else if ( bt == T_FLOAT) {
2826 // We convert float to int
2827 out_sig_bt[total_c_args-1] = T_INT;
2828 }
2829 }
2830
2831 assert(i==total_args_passed, "validly parsed signature");
2832
2833 // Now get the compiled-Java layout as input arguments
2834 int comp_args_on_stack;
2835 comp_args_on_stack = SharedRuntime::java_calling_convention(
2836 in_sig_bt, in_regs, total_args_passed, false);
2837
2838 // We have received a description of where all the java arg are located
2839 // on entry to the wrapper. We need to convert these args to where
2840 // the a native (non-jni) function would expect them. To figure out
2841 // where they go we convert the java signature to a C signature and remove
2842 // T_VOID for any long/double we might have received.
2843
2844
2845 // Now figure out where the args must be stored and how much stack space
2846 // they require (neglecting out_preserve_stack_slots but space for storing
2847 // the 1st six register arguments). It's weird see int_stk_helper.
2848 //
2849 int out_arg_slots;
2850 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
2851
2852 // Calculate the total number of stack slots we will need.
2853
2854 // First count the abi requirement plus all of the outgoing args
2855 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2856
2857 // Plus a temp for possible converion of float/double/long register args
2858
2859 int conversion_temp = stack_slots;
2860 stack_slots += 2;
2861
2862
2863 // Now space for the string(s) we must convert
2864
2865 int string_locs = stack_slots;
2866 stack_slots += total_strings *
2867 (max_dtrace_string_size / VMRegImpl::stack_slot_size);
2868
2869 // Ok The space we have allocated will look like:
2870 //
2871 //
2872 // FP-> | |
2873 // |---------------------|
2874 // | string[n] |
2875 // |---------------------| <- string_locs[n]
2876 // | string[n-1] |
2877 // |---------------------| <- string_locs[n-1]
2878 // | ... |
2879 // | ... |
2880 // |---------------------| <- string_locs[1]
2881 // | string[0] |
2882 // |---------------------| <- string_locs[0]
2883 // | temp |
2884 // |---------------------| <- conversion_temp
2885 // | outbound memory |
2886 // | based arguments |
2887 // | |
2888 // |---------------------|
2889 // | |
2890 // SP-> | out_preserved_slots |
2891 //
2892 //
2893
2894 // Now compute actual number of stack words we need rounding to make
2895 // stack properly aligned.
2896 stack_slots = round_to(stack_slots, 4 * VMRegImpl::slots_per_word);
2897
2898 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2899
2900 intptr_t start = (intptr_t)__ pc();
2901
2902 // First thing make an ic check to see if we should even be here
2903
2904 {
2905 Label L;
2906 const Register temp_reg = G3_scratch;
2907 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
2908 __ verify_oop(O0);
2909 __ ld_ptr(O0, oopDesc::klass_offset_in_bytes(), temp_reg);
2910 __ cmp_and_brx_short(temp_reg, G5_inline_cache_reg, Assembler::equal, Assembler::pt, L);
2911
2912 __ jump_to(ic_miss, temp_reg);
2913 __ delayed()->nop();
2914 __ align(CodeEntryAlignment);
2915 __ bind(L);
2916 }
2917
2918 int vep_offset = ((intptr_t)__ pc()) - start;
2919
2920
2921 // The instruction at the verified entry point must be 5 bytes or longer
2922 // because it can be patched on the fly by make_non_entrant. The stack bang
2923 // instruction fits that requirement.
2924
2925 // Generate stack overflow check before creating frame
2926 __ generate_stack_overflow_check(stack_size);
2927
2928 assert(((intptr_t)__ pc() - start - vep_offset) >= 5,
2929 "valid size for make_non_entrant");
2930
2931 // Generate a new frame for the wrapper.
2932 __ save(SP, -stack_size, SP);
2933
2934 // Frame is now completed as far a size and linkage.
2935
2936 int frame_complete = ((intptr_t)__ pc()) - start;
2937
2938 #ifdef ASSERT
2939 bool reg_destroyed[RegisterImpl::number_of_registers];
2940 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
2941 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
2942 reg_destroyed[r] = false;
2943 }
2944 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
2945 freg_destroyed[f] = false;
2946 }
2947
2948 #endif /* ASSERT */
2949
2950 VMRegPair zero;
2951 const Register g0 = G0; // without this we get a compiler warning (why??)
2952 zero.set2(g0->as_VMReg());
2953
2954 int c_arg, j_arg;
2955
2956 Register conversion_off = noreg;
2957
2958 for (j_arg = first_arg_to_pass, c_arg = 0 ;
2959 j_arg < total_args_passed ; j_arg++, c_arg++ ) {
2960
2961 VMRegPair src = in_regs[j_arg];
2962 VMRegPair dst = out_regs[c_arg];
2963
2964 #ifdef ASSERT
2965 if (src.first()->is_Register()) {
2966 assert(!reg_destroyed[src.first()->as_Register()->encoding()], "ack!");
2967 } else if (src.first()->is_FloatRegister()) {
2968 assert(!freg_destroyed[src.first()->as_FloatRegister()->encoding(
2969 FloatRegisterImpl::S)], "ack!");
2970 }
2971 if (dst.first()->is_Register()) {
2972 reg_destroyed[dst.first()->as_Register()->encoding()] = true;
2973 } else if (dst.first()->is_FloatRegister()) {
2974 freg_destroyed[dst.first()->as_FloatRegister()->encoding(
2975 FloatRegisterImpl::S)] = true;
2976 }
2977 #endif /* ASSERT */
2978
2979 switch (in_sig_bt[j_arg]) {
2980 case T_ARRAY:
2981 case T_OBJECT:
2982 {
2983 if (out_sig_bt[c_arg] == T_BYTE || out_sig_bt[c_arg] == T_SHORT ||
2984 out_sig_bt[c_arg] == T_INT || out_sig_bt[c_arg] == T_LONG) {
2985 // need to unbox a one-slot value
2986 Register in_reg = L0;
2987 Register tmp = L2;
2988 if ( src.first()->is_reg() ) {
2989 in_reg = src.first()->as_Register();
2990 } else {
2991 assert(Assembler::is_simm13(reg2offset(src.first()) + STACK_BIAS),
2992 "must be");
2993 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, in_reg);
2994 }
2995 // If the final destination is an acceptable register
2996 if ( dst.first()->is_reg() ) {
2997 if ( dst.is_single_phys_reg() || out_sig_bt[c_arg] != T_LONG ) {
2998 tmp = dst.first()->as_Register();
2999 }
3000 }
3001
3002 Label skipUnbox;
3003 if ( wordSize == 4 && out_sig_bt[c_arg] == T_LONG ) {
3004 __ mov(G0, tmp->successor());
3005 }
3006 __ br_null(in_reg, true, Assembler::pn, skipUnbox);
3007 __ delayed()->mov(G0, tmp);
3008
3009 BasicType bt = out_sig_bt[c_arg];
3010 int box_offset = java_lang_boxing_object::value_offset_in_bytes(bt);
3011 switch (bt) {
3012 case T_BYTE:
3013 __ ldub(in_reg, box_offset, tmp); break;
3014 case T_SHORT:
3015 __ lduh(in_reg, box_offset, tmp); break;
3016 case T_INT:
3017 __ ld(in_reg, box_offset, tmp); break;
3018 case T_LONG:
3019 __ ld_long(in_reg, box_offset, tmp); break;
3020 default: ShouldNotReachHere();
3021 }
3022
3023 __ bind(skipUnbox);
3024 // If tmp wasn't final destination copy to final destination
3025 if (tmp == L2) {
3026 VMRegPair tmp_as_VM = reg64_to_VMRegPair(L2);
3027 if (out_sig_bt[c_arg] == T_LONG) {
3028 long_move(masm, tmp_as_VM, dst);
3029 } else {
3030 move32_64(masm, tmp_as_VM, out_regs[c_arg]);
3031 }
3032 }
3033 if (out_sig_bt[c_arg] == T_LONG) {
3034 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
3035 ++c_arg; // move over the T_VOID to keep the loop indices in sync
3036 }
3037 } else if (out_sig_bt[c_arg] == T_ADDRESS) {
3038 Register s =
3039 src.first()->is_reg() ? src.first()->as_Register() : L2;
3040 Register d =
3041 dst.first()->is_reg() ? dst.first()->as_Register() : L2;
3042
3043 // We store the oop now so that the conversion pass can reach
3044 // while in the inner frame. This will be the only store if
3045 // the oop is NULL.
3046 if (s != L2) {
3047 // src is register
3048 if (d != L2) {
3049 // dst is register
3050 __ mov(s, d);
3051 } else {
3052 assert(Assembler::is_simm13(reg2offset(dst.first()) +
3053 STACK_BIAS), "must be");
3054 __ st_ptr(s, SP, reg2offset(dst.first()) + STACK_BIAS);
3055 }
3056 } else {
3057 // src not a register
3058 assert(Assembler::is_simm13(reg2offset(src.first()) +
3059 STACK_BIAS), "must be");
3060 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, d);
3061 if (d == L2) {
3062 assert(Assembler::is_simm13(reg2offset(dst.first()) +
3063 STACK_BIAS), "must be");
3064 __ st_ptr(d, SP, reg2offset(dst.first()) + STACK_BIAS);
3065 }
3066 }
3067 } else if (out_sig_bt[c_arg] != T_VOID) {
3068 // Convert the arg to NULL
3069 if (dst.first()->is_reg()) {
3070 __ mov(G0, dst.first()->as_Register());
3071 } else {
3072 assert(Assembler::is_simm13(reg2offset(dst.first()) +
3073 STACK_BIAS), "must be");
3074 __ st_ptr(G0, SP, reg2offset(dst.first()) + STACK_BIAS);
3075 }
3076 }
3077 }
3078 break;
3079 case T_VOID:
3080 break;
3081
3082 case T_FLOAT:
3083 if (src.first()->is_stack()) {
3084 // Stack to stack/reg is simple
3085 move32_64(masm, src, dst);
3086 } else {
3087 if (dst.first()->is_reg()) {
3088 // freg -> reg
3089 int off =
3090 STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
3091 Register d = dst.first()->as_Register();
3092 if (Assembler::is_simm13(off)) {
3093 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
3094 SP, off);
3095 __ ld(SP, off, d);
3096 } else {
3097 if (conversion_off == noreg) {
3098 __ set(off, L6);
3099 conversion_off = L6;
3100 }
3101 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
3102 SP, conversion_off);
3103 __ ld(SP, conversion_off , d);
3104 }
3105 } else {
3106 // freg -> mem
3107 int off = STACK_BIAS + reg2offset(dst.first());
3108 if (Assembler::is_simm13(off)) {
3109 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
3110 SP, off);
3111 } else {
3112 if (conversion_off == noreg) {
3113 __ set(off, L6);
3114 conversion_off = L6;
3115 }
3116 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
3117 SP, conversion_off);
3118 }
3119 }
3120 }
3121 break;
3122
3123 case T_DOUBLE:
3124 assert( j_arg + 1 < total_args_passed &&
3125 in_sig_bt[j_arg + 1] == T_VOID &&
3126 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
3127 if (src.first()->is_stack()) {
3128 // Stack to stack/reg is simple
3129 long_move(masm, src, dst);
3130 } else {
3131 Register d = dst.first()->is_reg() ? dst.first()->as_Register() : L2;
3132
3133 // Destination could be an odd reg on 32bit in which case
3134 // we can't load direct to the destination.
3135
3136 if (!d->is_even() && wordSize == 4) {
3137 d = L2;
3138 }
3139 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
3140 if (Assembler::is_simm13(off)) {
3141 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
3142 SP, off);
3143 __ ld_long(SP, off, d);
3144 } else {
3145 if (conversion_off == noreg) {
3146 __ set(off, L6);
3147 conversion_off = L6;
3148 }
3149 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
3150 SP, conversion_off);
3151 __ ld_long(SP, conversion_off, d);
3152 }
3153 if (d == L2) {
3154 long_move(masm, reg64_to_VMRegPair(L2), dst);
3155 }
3156 }
3157 break;
3158
3159 case T_LONG :
3160 // 32bit can't do a split move of something like g1 -> O0, O1
3161 // so use a memory temp
3162 if (src.is_single_phys_reg() && wordSize == 4) {
3163 Register tmp = L2;
3164 if (dst.first()->is_reg() &&
3165 (wordSize == 8 || dst.first()->as_Register()->is_even())) {
3166 tmp = dst.first()->as_Register();
3167 }
3168
3169 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
3170 if (Assembler::is_simm13(off)) {
3171 __ stx(src.first()->as_Register(), SP, off);
3172 __ ld_long(SP, off, tmp);
3173 } else {
3174 if (conversion_off == noreg) {
3175 __ set(off, L6);
3176 conversion_off = L6;
3177 }
3178 __ stx(src.first()->as_Register(), SP, conversion_off);
3179 __ ld_long(SP, conversion_off, tmp);
3180 }
3181
3182 if (tmp == L2) {
3183 long_move(masm, reg64_to_VMRegPair(L2), dst);
3184 }
3185 } else {
3186 long_move(masm, src, dst);
3187 }
3188 break;
3189
3190 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
3191
3192 default:
3193 move32_64(masm, src, dst);
3194 }
3195 }
3196
3197
3198 // If we have any strings we must store any register based arg to the stack
3199 // This includes any still live xmm registers too.
3200
3201 if (total_strings > 0 ) {
3202
3203 // protect all the arg registers
3204 __ save_frame(0);
3205 __ mov(G2_thread, L7_thread_cache);
3206 const Register L2_string_off = L2;
3207
3208 // Get first string offset
3209 __ set(string_locs * VMRegImpl::stack_slot_size, L2_string_off);
3210
3211 for (c_arg = 0 ; c_arg < total_c_args ; c_arg++ ) {
3212 if (out_sig_bt[c_arg] == T_ADDRESS) {
3213
3214 VMRegPair dst = out_regs[c_arg];
3215 const Register d = dst.first()->is_reg() ?
3216 dst.first()->as_Register()->after_save() : noreg;
3217
3218 // It's a string the oop and it was already copied to the out arg
3219 // position
3220 if (d != noreg) {
3221 __ mov(d, O0);
3222 } else {
3223 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
3224 "must be");
3225 __ ld_ptr(FP, reg2offset(dst.first()) + STACK_BIAS, O0);
3226 }
3227 Label skip;
3228
3229 __ br_null(O0, false, Assembler::pn, skip);
3230 __ delayed()->add(FP, L2_string_off, O1);
3231
3232 if (d != noreg) {
3233 __ mov(O1, d);
3234 } else {
3235 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
3236 "must be");
3237 __ st_ptr(O1, FP, reg2offset(dst.first()) + STACK_BIAS);
3238 }
3239
3240 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::get_utf),
3241 relocInfo::runtime_call_type);
3242 __ delayed()->add(L2_string_off, max_dtrace_string_size, L2_string_off);
3243
3244 __ bind(skip);
3245
3246 }
3247
3248 }
3249 __ mov(L7_thread_cache, G2_thread);
3250 __ restore();
3251
3252 }
3253
3254
3255 // Ok now we are done. Need to place the nop that dtrace wants in order to
3256 // patch in the trap
3257
3258 int patch_offset = ((intptr_t)__ pc()) - start;
3259
3260 __ nop();
3261
3262
3263 // Return
3264
3265 __ ret();
3266 __ delayed()->restore();
3267
3268 __ flush();
3269
3270 nmethod *nm = nmethod::new_dtrace_nmethod(
3271 method, masm->code(), vep_offset, patch_offset, frame_complete,
3272 stack_slots / VMRegImpl::slots_per_word);
3273 return nm;
3274
3275 }
3276
3277 #endif // HAVE_DTRACE_H
3278
3279 // this function returns the adjust size (in number of words) to a c2i adapter
3280 // activation for use during deoptimization
3281 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
3282 assert(callee_locals >= callee_parameters,
3283 "test and remove; got more parms than locals");
3284 if (callee_locals < callee_parameters)
3285 return 0; // No adjustment for negative locals
3286 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
3287 return round_to(diff, WordsPerLong);
3288 }
3289
3290 // "Top of Stack" slots that may be unused by the calling convention but must
3291 // otherwise be preserved.
3292 // On Intel these are not necessary and the value can be zero.
3293 // On Sparc this describes the words reserved for storing a register window
3294 // when an interrupt occurs.
3295 uint SharedRuntime::out_preserve_stack_slots() {
3296 return frame::register_save_words * VMRegImpl::slots_per_word;
3297 }
3298
3299 static void gen_new_frame(MacroAssembler* masm, bool deopt) {
3300 //
3301 // Common out the new frame generation for deopt and uncommon trap
3302 //
3303 Register G3pcs = G3_scratch; // Array of new pcs (input)
3304 Register Oreturn0 = O0;
3305 Register Oreturn1 = O1;
3306 Register O2UnrollBlock = O2;
3307 Register O3array = O3; // Array of frame sizes (input)
3308 Register O4array_size = O4; // number of frames (input)
3309 Register O7frame_size = O7; // number of frames (input)
3310
3311 __ ld_ptr(O3array, 0, O7frame_size);
3312 __ sub(G0, O7frame_size, O7frame_size);
3313 __ save(SP, O7frame_size, SP);
3314 __ ld_ptr(G3pcs, 0, I7); // load frame's new pc
3315
3316 #ifdef ASSERT
3317 // make sure that the frames are aligned properly
3318 #ifndef _LP64
3319 __ btst(wordSize*2-1, SP);
3320 __ breakpoint_trap(Assembler::notZero, Assembler::ptr_cc);
3321 #endif
3322 #endif
3323
3324 // Deopt needs to pass some extra live values from frame to frame
3325
3326 if (deopt) {
3327 __ mov(Oreturn0->after_save(), Oreturn0);
3328 __ mov(Oreturn1->after_save(), Oreturn1);
3329 }
3330
3331 __ mov(O4array_size->after_save(), O4array_size);
3332 __ sub(O4array_size, 1, O4array_size);
3333 __ mov(O3array->after_save(), O3array);
3334 __ mov(O2UnrollBlock->after_save(), O2UnrollBlock);
3335 __ add(G3pcs, wordSize, G3pcs); // point to next pc value
3336
3337 #ifdef ASSERT
3338 // trash registers to show a clear pattern in backtraces
3339 __ set(0xDEAD0000, I0);
3340 __ add(I0, 2, I1);
3341 __ add(I0, 4, I2);
3342 __ add(I0, 6, I3);
3343 __ add(I0, 8, I4);
3344 // Don't touch I5 could have valuable savedSP
3345 __ set(0xDEADBEEF, L0);
3346 __ mov(L0, L1);
3347 __ mov(L0, L2);
3348 __ mov(L0, L3);
3349 __ mov(L0, L4);
3350 __ mov(L0, L5);
3351
3352 // trash the return value as there is nothing to return yet
3353 __ set(0xDEAD0001, O7);
3354 #endif
3355
3356 __ mov(SP, O5_savedSP);
3357 }
3358
3359
3360 static void make_new_frames(MacroAssembler* masm, bool deopt) {
3361 //
3362 // loop through the UnrollBlock info and create new frames
3363 //
3364 Register G3pcs = G3_scratch;
3365 Register Oreturn0 = O0;
3366 Register Oreturn1 = O1;
3367 Register O2UnrollBlock = O2;
3368 Register O3array = O3;
3369 Register O4array_size = O4;
3370 Label loop;
3371
3372 // Before we make new frames, check to see if stack is available.
3373 // Do this after the caller's return address is on top of stack
3374 if (UseStackBanging) {
3375 // Get total frame size for interpreted frames
3376 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes(), O4);
3377 __ bang_stack_size(O4, O3, G3_scratch);
3378 }
3379
3380 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(), O4array_size);
3381 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(), G3pcs);
3382 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(), O3array);
3383
3384 // Adjust old interpreter frame to make space for new frame's extra java locals
3385 //
3386 // We capture the original sp for the transition frame only because it is needed in
3387 // order to properly calculate interpreter_sp_adjustment. Even though in real life
3388 // every interpreter frame captures a savedSP it is only needed at the transition
3389 // (fortunately). If we had to have it correct everywhere then we would need to
3390 // be told the sp_adjustment for each frame we create. If the frame size array
3391 // were to have twice the frame count entries then we could have pairs [sp_adjustment, frame_size]
3392 // for each frame we create and keep up the illusion every where.
3393 //
3394
3395 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(), O7);
3396 __ mov(SP, O5_savedSP); // remember initial sender's original sp before adjustment
3397 __ sub(SP, O7, SP);
3398
3399 #ifdef ASSERT
3400 // make sure that there is at least one entry in the array
3401 __ tst(O4array_size);
3402 __ breakpoint_trap(Assembler::zero, Assembler::icc);
3403 #endif
3404
3405 // Now push the new interpreter frames
3406 __ bind(loop);
3407
3408 // allocate a new frame, filling the registers
3409
3410 gen_new_frame(masm, deopt); // allocate an interpreter frame
3411
3412 __ cmp_zero_and_br(Assembler::notZero, O4array_size, loop);
3413 __ delayed()->add(O3array, wordSize, O3array);
3414 __ ld_ptr(G3pcs, 0, O7); // load final frame new pc
3415
3416 }
3417
3418 //------------------------------generate_deopt_blob----------------------------
3419 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
3420 // instead.
3421 void SharedRuntime::generate_deopt_blob() {
3422 // allocate space for the code
3423 ResourceMark rm;
3424 // setup code generation tools
3425 int pad = VerifyThread ? 512 : 0;// Extra slop space for more verify code
3426 if (UseStackBanging) {
3427 pad += StackShadowPages*16 + 32;
3428 }
3429 #ifdef _LP64
3430 CodeBuffer buffer("deopt_blob", 2100+pad, 512);
3431 #else
3432 // Measured 8/7/03 at 1212 in 32bit debug build (no VerifyThread)
3433 // Measured 8/7/03 at 1396 in 32bit debug build (VerifyThread)
3434 CodeBuffer buffer("deopt_blob", 1600+pad, 512);
3435 #endif /* _LP64 */
3436 MacroAssembler* masm = new MacroAssembler(&buffer);
3437 FloatRegister Freturn0 = F0;
3438 Register Greturn1 = G1;
3439 Register Oreturn0 = O0;
3440 Register Oreturn1 = O1;
3441 Register O2UnrollBlock = O2;
3442 Register L0deopt_mode = L0;
3443 Register G4deopt_mode = G4_scratch;
3444 int frame_size_words;
3445 Address saved_Freturn0_addr(FP, -sizeof(double) + STACK_BIAS);
3446 #if !defined(_LP64) && defined(COMPILER2)
3447 Address saved_Greturn1_addr(FP, -sizeof(double) -sizeof(jlong) + STACK_BIAS);
3448 #endif
3449 Label cont;
3450
3451 OopMapSet *oop_maps = new OopMapSet();
3452
3453 //
3454 // This is the entry point for code which is returning to a de-optimized
3455 // frame.
3456 // The steps taken by this frame are as follows:
3457 // - push a dummy "register_save" and save the return values (O0, O1, F0/F1, G1)
3458 // and all potentially live registers (at a pollpoint many registers can be live).
3459 //
3460 // - call the C routine: Deoptimization::fetch_unroll_info (this function
3461 // returns information about the number and size of interpreter frames
3462 // which are equivalent to the frame which is being deoptimized)
3463 // - deallocate the unpack frame, restoring only results values. Other
3464 // volatile registers will now be captured in the vframeArray as needed.
3465 // - deallocate the deoptimization frame
3466 // - in a loop using the information returned in the previous step
3467 // push new interpreter frames (take care to propagate the return
3468 // values through each new frame pushed)
3469 // - create a dummy "unpack_frame" and save the return values (O0, O1, F0)
3470 // - call the C routine: Deoptimization::unpack_frames (this function
3471 // lays out values on the interpreter frame which was just created)
3472 // - deallocate the dummy unpack_frame
3473 // - ensure that all the return values are correctly set and then do
3474 // a return to the interpreter entry point
3475 //
3476 // Refer to the following methods for more information:
3477 // - Deoptimization::fetch_unroll_info
3478 // - Deoptimization::unpack_frames
3479
3480 OopMap* map = NULL;
3481
3482 int start = __ offset();
3483
3484 // restore G2, the trampoline destroyed it
3485 __ get_thread();
3486
3487 // On entry we have been called by the deoptimized nmethod with a call that
3488 // replaced the original call (or safepoint polling location) so the deoptimizing
3489 // pc is now in O7. Return values are still in the expected places
3490
3491 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3492 __ ba(cont);
3493 __ delayed()->mov(Deoptimization::Unpack_deopt, L0deopt_mode);
3494
3495 int exception_offset = __ offset() - start;
3496
3497 // restore G2, the trampoline destroyed it
3498 __ get_thread();
3499
3500 // On entry we have been jumped to by the exception handler (or exception_blob
3501 // for server). O0 contains the exception oop and O7 contains the original
3502 // exception pc. So if we push a frame here it will look to the
3503 // stack walking code (fetch_unroll_info) just like a normal call so
3504 // state will be extracted normally.
3505
3506 // save exception oop in JavaThread and fall through into the
3507 // exception_in_tls case since they are handled in same way except
3508 // for where the pending exception is kept.
3509 __ st_ptr(Oexception, G2_thread, JavaThread::exception_oop_offset());
3510
3511 //
3512 // Vanilla deoptimization with an exception pending in exception_oop
3513 //
3514 int exception_in_tls_offset = __ offset() - start;
3515
3516 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
3517 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3518
3519 // Restore G2_thread
3520 __ get_thread();
3521
3522 #ifdef ASSERT
3523 {
3524 // verify that there is really an exception oop in exception_oop
3525 Label has_exception;
3526 __ ld_ptr(G2_thread, JavaThread::exception_oop_offset(), Oexception);
3527 __ br_notnull_short(Oexception, Assembler::pt, has_exception);
3528 __ stop("no exception in thread");
3529 __ bind(has_exception);
3530
3531 // verify that there is no pending exception
3532 Label no_pending_exception;
3533 Address exception_addr(G2_thread, Thread::pending_exception_offset());
3534 __ ld_ptr(exception_addr, Oexception);
3535 __ br_null_short(Oexception, Assembler::pt, no_pending_exception);
3536 __ stop("must not have pending exception here");
3537 __ bind(no_pending_exception);
3538 }
3539 #endif
3540
3541 __ ba(cont);
3542 __ delayed()->mov(Deoptimization::Unpack_exception, L0deopt_mode);;
3543
3544 //
3545 // Reexecute entry, similar to c2 uncommon trap
3546 //
3547 int reexecute_offset = __ offset() - start;
3548
3549 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
3550 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3551
3552 __ mov(Deoptimization::Unpack_reexecute, L0deopt_mode);
3553
3554 __ bind(cont);
3555
3556 __ set_last_Java_frame(SP, noreg);
3557
3558 // do the call by hand so we can get the oopmap
3559
3560 __ mov(G2_thread, L7_thread_cache);
3561 __ call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), relocInfo::runtime_call_type);
3562 __ delayed()->mov(G2_thread, O0);
3563
3564 // Set an oopmap for the call site this describes all our saved volatile registers
3565
3566 oop_maps->add_gc_map( __ offset()-start, map);
3567
3568 __ mov(L7_thread_cache, G2_thread);
3569
3570 __ reset_last_Java_frame();
3571
3572 // NOTE: we know that only O0/O1 will be reloaded by restore_result_registers
3573 // so this move will survive
3574
3575 __ mov(L0deopt_mode, G4deopt_mode);
3576
3577 __ mov(O0, O2UnrollBlock->after_save());
3578
3579 RegisterSaver::restore_result_registers(masm);
3580
3581 Label noException;
3582 __ cmp_and_br_short(G4deopt_mode, Deoptimization::Unpack_exception, Assembler::notEqual, Assembler::pt, noException);
3583
3584 // Move the pending exception from exception_oop to Oexception so
3585 // the pending exception will be picked up the interpreter.
3586 __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception);
3587 __ st_ptr(G0, G2_thread, in_bytes(JavaThread::exception_oop_offset()));
3588 __ bind(noException);
3589
3590 // deallocate the deoptimization frame taking care to preserve the return values
3591 __ mov(Oreturn0, Oreturn0->after_save());
3592 __ mov(Oreturn1, Oreturn1->after_save());
3593 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
3594 __ restore();
3595
3596 // Allocate new interpreter frame(s) and possible c2i adapter frame
3597
3598 make_new_frames(masm, true);
3599
3600 // push a dummy "unpack_frame" taking care of float return values and
3601 // call Deoptimization::unpack_frames to have the unpacker layout
3602 // information in the interpreter frames just created and then return
3603 // to the interpreter entry point
3604 __ save(SP, -frame_size_words*wordSize, SP);
3605 __ stf(FloatRegisterImpl::D, Freturn0, saved_Freturn0_addr);
3606 #if !defined(_LP64)
3607 #if defined(COMPILER2)
3608 // 32-bit 1-register longs return longs in G1
3609 __ stx(Greturn1, saved_Greturn1_addr);
3610 #endif
3611 __ set_last_Java_frame(SP, noreg);
3612 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, G4deopt_mode);
3613 #else
3614 // LP64 uses g4 in set_last_Java_frame
3615 __ mov(G4deopt_mode, O1);
3616 __ set_last_Java_frame(SP, G0);
3617 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O1);
3618 #endif
3619 __ reset_last_Java_frame();
3620 __ ldf(FloatRegisterImpl::D, saved_Freturn0_addr, Freturn0);
3621
3622 #if !defined(_LP64) && defined(COMPILER2)
3623 // In 32 bit, C2 returns longs in G1 so restore the saved G1 into
3624 // I0/I1 if the return value is long.
3625 Label not_long;
3626 __ cmp_and_br_short(O0,T_LONG, Assembler::notEqual, Assembler::pt, not_long);
3627 __ ldd(saved_Greturn1_addr,I0);
3628 __ bind(not_long);
3629 #endif
3630 __ ret();
3631 __ delayed()->restore();
3632
3633 masm->flush();
3634 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_words);
3635 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3636 }
3637
3638 #ifdef COMPILER2
3639
3640 //------------------------------generate_uncommon_trap_blob--------------------
3641 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
3642 // instead.
3643 void SharedRuntime::generate_uncommon_trap_blob() {
3644 // allocate space for the code
3645 ResourceMark rm;
3646 // setup code generation tools
3647 int pad = VerifyThread ? 512 : 0;
3648 if (UseStackBanging) {
3649 pad += StackShadowPages*16 + 32;
3650 }
3651 #ifdef _LP64
3652 CodeBuffer buffer("uncommon_trap_blob", 2700+pad, 512);
3653 #else
3654 // Measured 8/7/03 at 660 in 32bit debug build (no VerifyThread)
3655 // Measured 8/7/03 at 1028 in 32bit debug build (VerifyThread)
3656 CodeBuffer buffer("uncommon_trap_blob", 2000+pad, 512);
3657 #endif
3658 MacroAssembler* masm = new MacroAssembler(&buffer);
3659 Register O2UnrollBlock = O2;
3660 Register O2klass_index = O2;
3661
3662 //
3663 // This is the entry point for all traps the compiler takes when it thinks
3664 // it cannot handle further execution of compilation code. The frame is
3665 // deoptimized in these cases and converted into interpreter frames for
3666 // execution
3667 // The steps taken by this frame are as follows:
3668 // - push a fake "unpack_frame"
3669 // - call the C routine Deoptimization::uncommon_trap (this function
3670 // packs the current compiled frame into vframe arrays and returns
3671 // information about the number and size of interpreter frames which
3672 // are equivalent to the frame which is being deoptimized)
3673 // - deallocate the "unpack_frame"
3674 // - deallocate the deoptimization frame
3675 // - in a loop using the information returned in the previous step
3676 // push interpreter frames;
3677 // - create a dummy "unpack_frame"
3678 // - call the C routine: Deoptimization::unpack_frames (this function
3679 // lays out values on the interpreter frame which was just created)
3680 // - deallocate the dummy unpack_frame
3681 // - return to the interpreter entry point
3682 //
3683 // Refer to the following methods for more information:
3684 // - Deoptimization::uncommon_trap
3685 // - Deoptimization::unpack_frame
3686
3687 // the unloaded class index is in O0 (first parameter to this blob)
3688
3689 // push a dummy "unpack_frame"
3690 // and call Deoptimization::uncommon_trap to pack the compiled frame into
3691 // vframe array and return the UnrollBlock information
3692 __ save_frame(0);
3693 __ set_last_Java_frame(SP, noreg);
3694 __ mov(I0, O2klass_index);
3695 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), G2_thread, O2klass_index);
3696 __ reset_last_Java_frame();
3697 __ mov(O0, O2UnrollBlock->after_save());
3698 __ restore();
3699
3700 // deallocate the deoptimized frame taking care to preserve the return values
3701 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
3702 __ restore();
3703
3704 // Allocate new interpreter frame(s) and possible c2i adapter frame
3705
3706 make_new_frames(masm, false);
3707
3708 // push a dummy "unpack_frame" taking care of float return values and
3709 // call Deoptimization::unpack_frames to have the unpacker layout
3710 // information in the interpreter frames just created and then return
3711 // to the interpreter entry point
3712 __ save_frame(0);
3713 __ set_last_Java_frame(SP, noreg);
3714 __ mov(Deoptimization::Unpack_uncommon_trap, O3); // indicate it is the uncommon trap case
3715 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O3);
3716 __ reset_last_Java_frame();
3717 __ ret();
3718 __ delayed()->restore();
3719
3720 masm->flush();
3721 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, NULL, __ total_frame_size_in_bytes(0)/wordSize);
3722 }
3723
3724 #endif // COMPILER2
3725
3726 //------------------------------generate_handler_blob-------------------
3727 //
3728 // Generate a special Compile2Runtime blob that saves all registers, and sets
3729 // up an OopMap.
3730 //
3731 // This blob is jumped to (via a breakpoint and the signal handler) from a
3732 // safepoint in compiled code. On entry to this blob, O7 contains the
3733 // address in the original nmethod at which we should resume normal execution.
3734 // Thus, this blob looks like a subroutine which must preserve lots of
3735 // registers and return normally. Note that O7 is never register-allocated,
3736 // so it is guaranteed to be free here.
3737 //
3738
3739 // The hardest part of what this blob must do is to save the 64-bit %o
3740 // registers in the 32-bit build. A simple 'save' turn the %o's to %i's and
3741 // an interrupt will chop off their heads. Making space in the caller's frame
3742 // first will let us save the 64-bit %o's before save'ing, but we cannot hand
3743 // the adjusted FP off to the GC stack-crawler: this will modify the caller's
3744 // SP and mess up HIS OopMaps. So we first adjust the caller's SP, then save
3745 // the 64-bit %o's, then do a save, then fixup the caller's SP (our FP).
3746 // Tricky, tricky, tricky...
3747
3748 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
3749 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3750
3751 // allocate space for the code
3752 ResourceMark rm;
3753 // setup code generation tools
3754 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
3755 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
3756 // even larger with TraceJumps
3757 int pad = TraceJumps ? 512 : 0;
3758 CodeBuffer buffer("handler_blob", 1600 + pad, 512);
3759 MacroAssembler* masm = new MacroAssembler(&buffer);
3760 int frame_size_words;
3761 OopMapSet *oop_maps = new OopMapSet();
3762 OopMap* map = NULL;
3763
3764 int start = __ offset();
3765
3766 bool cause_return = (poll_type == POLL_AT_RETURN);
3767 // If this causes a return before the processing, then do a "restore"
3768 if (cause_return) {
3769 __ restore();
3770 } else {
3771 // Make it look like we were called via the poll
3772 // so that frame constructor always sees a valid return address
3773 __ ld_ptr(G2_thread, in_bytes(JavaThread::saved_exception_pc_offset()), O7);
3774 __ sub(O7, frame::pc_return_offset, O7);
3775 }
3776
3777 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3778
3779 // setup last_Java_sp (blows G4)
3780 __ set_last_Java_frame(SP, noreg);
3781
3782 // call into the runtime to handle illegal instructions exception
3783 // Do not use call_VM_leaf, because we need to make a GC map at this call site.
3784 __ mov(G2_thread, O0);
3785 __ save_thread(L7_thread_cache);
3786 __ call(call_ptr);
3787 __ delayed()->nop();
3788
3789 // Set an oopmap for the call site.
3790 // We need this not only for callee-saved registers, but also for volatile
3791 // registers that the compiler might be keeping live across a safepoint.
3792
3793 oop_maps->add_gc_map( __ offset() - start, map);
3794
3795 __ restore_thread(L7_thread_cache);
3796 // clear last_Java_sp
3797 __ reset_last_Java_frame();
3798
3799 // Check for exceptions
3800 Label pending;
3801
3802 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1);
3803 __ br_notnull_short(O1, Assembler::pn, pending);
3804
3805 RegisterSaver::restore_live_registers(masm);
3806
3807 // We are back the the original state on entry and ready to go.
3808
3809 __ retl();
3810 __ delayed()->nop();
3811
3812 // Pending exception after the safepoint
3813
3814 __ bind(pending);
3815
3816 RegisterSaver::restore_live_registers(masm);
3817
3818 // We are back the the original state on entry.
3819
3820 // Tail-call forward_exception_entry, with the issuing PC in O7,
3821 // so it looks like the original nmethod called forward_exception_entry.
3822 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0);
3823 __ JMP(O0, 0);
3824 __ delayed()->nop();
3825
3826 // -------------
3827 // make sure all code is generated
3828 masm->flush();
3829
3830 // return exception blob
3831 return SafepointBlob::create(&buffer, oop_maps, frame_size_words);
3832 }
3833
3834 //
3835 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3836 //
3837 // Generate a stub that calls into vm to find out the proper destination
3838 // of a java call. All the argument registers are live at this point
3839 // but since this is generic code we don't know what they are and the caller
3840 // must do any gc of the args.
3841 //
3842 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3843 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3844
3845 // allocate space for the code
3846 ResourceMark rm;
3847 // setup code generation tools
3848 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
3849 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
3850 // even larger with TraceJumps
3851 int pad = TraceJumps ? 512 : 0;
3852 CodeBuffer buffer(name, 1600 + pad, 512);
3853 MacroAssembler* masm = new MacroAssembler(&buffer);
3854 int frame_size_words;
3855 OopMapSet *oop_maps = new OopMapSet();
3856 OopMap* map = NULL;
3857
3858 int start = __ offset();
3859
3860 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3861
3862 int frame_complete = __ offset();
3863
3864 // setup last_Java_sp (blows G4)
3865 __ set_last_Java_frame(SP, noreg);
3866
3867 // call into the runtime to handle illegal instructions exception
3868 // Do not use call_VM_leaf, because we need to make a GC map at this call site.
3869 __ mov(G2_thread, O0);
3870 __ save_thread(L7_thread_cache);
3871 __ call(destination, relocInfo::runtime_call_type);
3872 __ delayed()->nop();
3873
3874 // O0 contains the address we are going to jump to assuming no exception got installed
3875
3876 // Set an oopmap for the call site.
3877 // We need this not only for callee-saved registers, but also for volatile
3878 // registers that the compiler might be keeping live across a safepoint.
3879
3880 oop_maps->add_gc_map( __ offset() - start, map);
3881
3882 __ restore_thread(L7_thread_cache);
3883 // clear last_Java_sp
3884 __ reset_last_Java_frame();
3885
3886 // Check for exceptions
3887 Label pending;
3888
3889 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1);
3890 __ br_notnull_short(O1, Assembler::pn, pending);
3891
3892 // get the returned Method*
3893
3894 __ get_vm_result_2(G5_method);
3895 __ stx(G5_method, SP, RegisterSaver::G5_offset()+STACK_BIAS);
3896
3897 // O0 is where we want to jump, overwrite G3 which is saved and scratch
3898
3899 __ stx(O0, SP, RegisterSaver::G3_offset()+STACK_BIAS);
3900
3901 RegisterSaver::restore_live_registers(masm);
3902
3903 // We are back the the original state on entry and ready to go.
3904
3905 __ JMP(G3, 0);
3906 __ delayed()->nop();
3907
3908 // Pending exception after the safepoint
3909
3910 __ bind(pending);
3911
3912 RegisterSaver::restore_live_registers(masm);
3913
3914 // We are back the the original state on entry.
3915
3916 // Tail-call forward_exception_entry, with the issuing PC in O7,
3917 // so it looks like the original nmethod called forward_exception_entry.
3918 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0);
3919 __ JMP(O0, 0);
3920 __ delayed()->nop();
3921
3922 // -------------
3923 // make sure all code is generated
3924 masm->flush();
3925
3926 // return the blob
3927 // frame_size_words or bytes??
3928 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true);
3929 }