1 /*
2 * Copyright (c) 2003, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "assembler_sparc.inline.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/icBuffer.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "oops/compiledICHolderOop.hpp"
33 #include "prims/jvmtiRedefineClassesTrace.hpp"
34 #include "runtime/sharedRuntime.hpp"
35 #include "runtime/vframeArray.hpp"
36 #include "vmreg_sparc.inline.hpp"
37 #ifdef COMPILER1
38 #include "c1/c1_Runtime1.hpp"
39 #endif
40 #ifdef COMPILER2
41 #include "opto/runtime.hpp"
42 #endif
43 #ifdef SHARK
44 #include "compiler/compileBroker.hpp"
45 #include "shark/sharkCompiler.hpp"
46 #endif
47
48 #define __ masm->
49
50
51 class RegisterSaver {
52
53 // Used for saving volatile registers. This is Gregs, Fregs, I/L/O.
54 // The Oregs are problematic. In the 32bit build the compiler can
55 // have O registers live with 64 bit quantities. A window save will
56 // cut the heads off of the registers. We have to do a very extensive
57 // stack dance to save and restore these properly.
58
59 // Note that the Oregs problem only exists if we block at either a polling
60 // page exception a compiled code safepoint that was not originally a call
61 // or deoptimize following one of these kinds of safepoints.
62
63 // Lots of registers to save. For all builds, a window save will preserve
64 // the %i and %l registers. For the 32-bit longs-in-two entries and 64-bit
65 // builds a window-save will preserve the %o registers. In the LION build
66 // we need to save the 64-bit %o registers which requires we save them
67 // before the window-save (as then they become %i registers and get their
68 // heads chopped off on interrupt). We have to save some %g registers here
69 // as well.
70 enum {
71 // This frame's save area. Includes extra space for the native call:
72 // vararg's layout space and the like. Briefly holds the caller's
73 // register save area.
74 call_args_area = frame::register_save_words_sp_offset +
75 frame::memory_parameter_word_sp_offset*wordSize,
76 // Make sure save locations are always 8 byte aligned.
77 // can't use round_to because it doesn't produce compile time constant
78 start_of_extra_save_area = ((call_args_area + 7) & ~7),
79 g1_offset = start_of_extra_save_area, // g-regs needing saving
80 g3_offset = g1_offset+8,
81 g4_offset = g3_offset+8,
82 g5_offset = g4_offset+8,
83 o0_offset = g5_offset+8,
84 o1_offset = o0_offset+8,
85 o2_offset = o1_offset+8,
86 o3_offset = o2_offset+8,
87 o4_offset = o3_offset+8,
88 o5_offset = o4_offset+8,
89 start_of_flags_save_area = o5_offset+8,
90 ccr_offset = start_of_flags_save_area,
91 fsr_offset = ccr_offset + 8,
92 d00_offset = fsr_offset+8, // Start of float save area
93 register_save_size = d00_offset+8*32
94 };
95
96
97 public:
98
99 static int Oexception_offset() { return o0_offset; };
100 static int G3_offset() { return g3_offset; };
101 static int G5_offset() { return g5_offset; };
102 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
103 static void restore_live_registers(MacroAssembler* masm);
104
105 // During deoptimization only the result register need to be restored
106 // all the other values have already been extracted.
107
108 static void restore_result_registers(MacroAssembler* masm);
109 };
110
111 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
112 // Record volatile registers as callee-save values in an OopMap so their save locations will be
113 // propagated to the caller frame's RegisterMap during StackFrameStream construction (needed for
114 // deoptimization; see compiledVFrame::create_stack_value). The caller's I, L and O registers
115 // are saved in register windows - I's and L's in the caller's frame and O's in the stub frame
116 // (as the stub's I's) when the runtime routine called by the stub creates its frame.
117 int i;
118 // Always make the frame size 16 byte aligned.
119 int frame_size = round_to(additional_frame_words + register_save_size, 16);
120 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words
121 int frame_size_in_slots = frame_size / sizeof(jint);
122 // CodeBlob frame size is in words.
123 *total_frame_words = frame_size / wordSize;
124 // OopMap* map = new OopMap(*total_frame_words, 0);
125 OopMap* map = new OopMap(frame_size_in_slots, 0);
126
127 #if !defined(_LP64)
128
129 // Save 64-bit O registers; they will get their heads chopped off on a 'save'.
130 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
131 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
132 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
133 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
134 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
135 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
136 #endif /* _LP64 */
137
138 __ save(SP, -frame_size, SP);
139
140 #ifndef _LP64
141 // Reload the 64 bit Oregs. Although they are now Iregs we load them
142 // to Oregs here to avoid interrupts cutting off their heads
143
144 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
145 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
146 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
147 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
148 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
149 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
150
151 __ stx(O0, SP, o0_offset+STACK_BIAS);
152 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset + 4)>>2), O0->as_VMReg());
153
154 __ stx(O1, SP, o1_offset+STACK_BIAS);
155
156 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset + 4)>>2), O1->as_VMReg());
157
158 __ stx(O2, SP, o2_offset+STACK_BIAS);
159 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset + 4)>>2), O2->as_VMReg());
160
161 __ stx(O3, SP, o3_offset+STACK_BIAS);
162 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset + 4)>>2), O3->as_VMReg());
163
164 __ stx(O4, SP, o4_offset+STACK_BIAS);
165 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset + 4)>>2), O4->as_VMReg());
166
167 __ stx(O5, SP, o5_offset+STACK_BIAS);
168 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset + 4)>>2), O5->as_VMReg());
169 #endif /* _LP64 */
170
171
172 #ifdef _LP64
173 int debug_offset = 0;
174 #else
175 int debug_offset = 4;
176 #endif
177 // Save the G's
178 __ stx(G1, SP, g1_offset+STACK_BIAS);
179 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset + debug_offset)>>2), G1->as_VMReg());
180
181 __ stx(G3, SP, g3_offset+STACK_BIAS);
182 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset + debug_offset)>>2), G3->as_VMReg());
183
184 __ stx(G4, SP, g4_offset+STACK_BIAS);
185 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset + debug_offset)>>2), G4->as_VMReg());
186
187 __ stx(G5, SP, g5_offset+STACK_BIAS);
188 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset + debug_offset)>>2), G5->as_VMReg());
189
190 // This is really a waste but we'll keep things as they were for now
191 if (true) {
192 #ifndef _LP64
193 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset)>>2), O0->as_VMReg()->next());
194 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset)>>2), O1->as_VMReg()->next());
195 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset)>>2), O2->as_VMReg()->next());
196 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset)>>2), O3->as_VMReg()->next());
197 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset)>>2), O4->as_VMReg()->next());
198 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset)>>2), O5->as_VMReg()->next());
199 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset)>>2), G1->as_VMReg()->next());
200 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset)>>2), G3->as_VMReg()->next());
201 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset)>>2), G4->as_VMReg()->next());
202 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset)>>2), G5->as_VMReg()->next());
203 #endif /* _LP64 */
204 }
205
206
207 // Save the flags
208 __ rdccr( G5 );
209 __ stx(G5, SP, ccr_offset+STACK_BIAS);
210 __ stxfsr(SP, fsr_offset+STACK_BIAS);
211
212 // Save all the FP registers: 32 doubles (32 floats correspond to the 2 halves of the first 16 doubles)
213 int offset = d00_offset;
214 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) {
215 FloatRegister f = as_FloatRegister(i);
216 __ stf(FloatRegisterImpl::D, f, SP, offset+STACK_BIAS);
217 // Record as callee saved both halves of double registers (2 float registers).
218 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), f->as_VMReg());
219 map->set_callee_saved(VMRegImpl::stack2reg((offset + sizeof(float))>>2), f->as_VMReg()->next());
220 offset += sizeof(double);
221 }
222
223 // And we're done.
224
225 return map;
226 }
227
228
229 // Pop the current frame and restore all the registers that we
230 // saved.
231 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
232
233 // Restore all the FP registers
234 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) {
235 __ ldf(FloatRegisterImpl::D, SP, d00_offset+i*sizeof(float)+STACK_BIAS, as_FloatRegister(i));
236 }
237
238 __ ldx(SP, ccr_offset+STACK_BIAS, G1);
239 __ wrccr (G1) ;
240
241 // Restore the G's
242 // Note that G2 (AKA GThread) must be saved and restored separately.
243 // TODO-FIXME: save and restore some of the other ASRs, viz., %asi and %gsr.
244
245 __ ldx(SP, g1_offset+STACK_BIAS, G1);
246 __ ldx(SP, g3_offset+STACK_BIAS, G3);
247 __ ldx(SP, g4_offset+STACK_BIAS, G4);
248 __ ldx(SP, g5_offset+STACK_BIAS, G5);
249
250
251 #if !defined(_LP64)
252 // Restore the 64-bit O's.
253 __ ldx(SP, o0_offset+STACK_BIAS, O0);
254 __ ldx(SP, o1_offset+STACK_BIAS, O1);
255 __ ldx(SP, o2_offset+STACK_BIAS, O2);
256 __ ldx(SP, o3_offset+STACK_BIAS, O3);
257 __ ldx(SP, o4_offset+STACK_BIAS, O4);
258 __ ldx(SP, o5_offset+STACK_BIAS, O5);
259
260 // And temporarily place them in TLS
261
262 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
263 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
264 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
265 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
266 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
267 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
268 #endif /* _LP64 */
269
270 // Restore flags
271
272 __ ldxfsr(SP, fsr_offset+STACK_BIAS);
273
274 __ restore();
275
276 #if !defined(_LP64)
277 // Now reload the 64bit Oregs after we've restore the window.
278 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
279 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
280 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
281 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
282 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
283 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
284 #endif /* _LP64 */
285
286 }
287
288 // Pop the current frame and restore the registers that might be holding
289 // a result.
290 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
291
292 #if !defined(_LP64)
293 // 32bit build returns longs in G1
294 __ ldx(SP, g1_offset+STACK_BIAS, G1);
295
296 // Retrieve the 64-bit O's.
297 __ ldx(SP, o0_offset+STACK_BIAS, O0);
298 __ ldx(SP, o1_offset+STACK_BIAS, O1);
299 // and save to TLS
300 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
301 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
302 #endif /* _LP64 */
303
304 __ ldf(FloatRegisterImpl::D, SP, d00_offset+STACK_BIAS, as_FloatRegister(0));
305
306 __ restore();
307
308 #if !defined(_LP64)
309 // Now reload the 64bit Oregs after we've restore the window.
310 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
311 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
312 #endif /* _LP64 */
313
314 }
315
316 // The java_calling_convention describes stack locations as ideal slots on
317 // a frame with no abi restrictions. Since we must observe abi restrictions
318 // (like the placement of the register window) the slots must be biased by
319 // the following value.
320 static int reg2offset(VMReg r) {
321 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
322 }
323
324 // ---------------------------------------------------------------------------
325 // Read the array of BasicTypes from a signature, and compute where the
326 // arguments should go. Values in the VMRegPair regs array refer to 4-byte (VMRegImpl::stack_slot_size)
327 // quantities. Values less than VMRegImpl::stack0 are registers, those above
328 // refer to 4-byte stack slots. All stack slots are based off of the window
329 // top. VMRegImpl::stack0 refers to the first slot past the 16-word window,
330 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register
331 // values 0-63 (up to RegisterImpl::number_of_registers) are the 64-bit
332 // integer registers. Values 64-95 are the (32-bit only) float registers.
333 // Each 32-bit quantity is given its own number, so the integer registers
334 // (in either 32- or 64-bit builds) use 2 numbers. For example, there is
335 // an O0-low and an O0-high. Essentially, all int register numbers are doubled.
336
337 // Register results are passed in O0-O5, for outgoing call arguments. To
338 // convert to incoming arguments, convert all O's to I's. The regs array
339 // refer to the low and hi 32-bit words of 64-bit registers or stack slots.
340 // If the regs[].second() field is set to VMRegImpl::Bad(), it means it's unused (a
341 // 32-bit value was passed). If both are VMRegImpl::Bad(), it means no value was
342 // passed (used as a placeholder for the other half of longs and doubles in
343 // the 64-bit build). regs[].second() is either VMRegImpl::Bad() or regs[].second() is
344 // regs[].first()+1 (regs[].first() may be misaligned in the C calling convention).
345 // Sparc never passes a value in regs[].second() but not regs[].first() (regs[].first()
346 // == VMRegImpl::Bad() && regs[].second() != VMRegImpl::Bad()) nor unrelated values in the
347 // same VMRegPair.
348
349 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
350 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
351 // units regardless of build.
352
353
354 // ---------------------------------------------------------------------------
355 // The compiled Java calling convention. The Java convention always passes
356 // 64-bit values in adjacent aligned locations (either registers or stack),
357 // floats in float registers and doubles in aligned float pairs. Values are
358 // packed in the registers. There is no backing varargs store for values in
359 // registers. In the 32-bit build, longs are passed in G1 and G4 (cannot be
360 // passed in I's, because longs in I's get their heads chopped off at
361 // interrupt).
362 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
363 VMRegPair *regs,
364 int total_args_passed,
365 int is_outgoing) {
366 assert(F31->as_VMReg()->is_reg(), "overlapping stack/register numbers");
367
368 // Convention is to pack the first 6 int/oop args into the first 6 registers
369 // (I0-I5), extras spill to the stack. Then pack the first 8 float args
370 // into F0-F7, extras spill to the stack. Then pad all register sets to
371 // align. Then put longs and doubles into the same registers as they fit,
372 // else spill to the stack.
373 const int int_reg_max = SPARC_ARGS_IN_REGS_NUM;
374 const int flt_reg_max = 8;
375 //
376 // Where 32-bit 1-reg longs start being passed
377 // In tiered we must pass on stack because c1 can't use a "pair" in a single reg.
378 // So make it look like we've filled all the G regs that c2 wants to use.
379 Register g_reg = TieredCompilation ? noreg : G1;
380
381 // Count int/oop and float args. See how many stack slots we'll need and
382 // where the longs & doubles will go.
383 int int_reg_cnt = 0;
384 int flt_reg_cnt = 0;
385 // int stk_reg_pairs = frame::register_save_words*(wordSize>>2);
386 // int stk_reg_pairs = SharedRuntime::out_preserve_stack_slots();
387 int stk_reg_pairs = 0;
388 for (int i = 0; i < total_args_passed; i++) {
389 switch (sig_bt[i]) {
390 case T_LONG: // LP64, longs compete with int args
391 assert(sig_bt[i+1] == T_VOID, "");
392 #ifdef _LP64
393 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
394 #endif
395 break;
396 case T_OBJECT:
397 case T_ARRAY:
398 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
399 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
400 #ifndef _LP64
401 else stk_reg_pairs++;
402 #endif
403 break;
404 case T_INT:
405 case T_SHORT:
406 case T_CHAR:
407 case T_BYTE:
408 case T_BOOLEAN:
409 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
410 else stk_reg_pairs++;
411 break;
412 case T_FLOAT:
413 if (flt_reg_cnt < flt_reg_max) flt_reg_cnt++;
414 else stk_reg_pairs++;
415 break;
416 case T_DOUBLE:
417 assert(sig_bt[i+1] == T_VOID, "");
418 break;
419 case T_VOID:
420 break;
421 default:
422 ShouldNotReachHere();
423 }
424 }
425
426 // This is where the longs/doubles start on the stack.
427 stk_reg_pairs = (stk_reg_pairs+1) & ~1; // Round
428
429 int int_reg_pairs = (int_reg_cnt+1) & ~1; // 32-bit 2-reg longs only
430 int flt_reg_pairs = (flt_reg_cnt+1) & ~1;
431
432 // int stk_reg = frame::register_save_words*(wordSize>>2);
433 // int stk_reg = SharedRuntime::out_preserve_stack_slots();
434 int stk_reg = 0;
435 int int_reg = 0;
436 int flt_reg = 0;
437
438 // Now do the signature layout
439 for (int i = 0; i < total_args_passed; i++) {
440 switch (sig_bt[i]) {
441 case T_INT:
442 case T_SHORT:
443 case T_CHAR:
444 case T_BYTE:
445 case T_BOOLEAN:
446 #ifndef _LP64
447 case T_OBJECT:
448 case T_ARRAY:
449 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
450 #endif // _LP64
451 if (int_reg < int_reg_max) {
452 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
453 regs[i].set1(r->as_VMReg());
454 } else {
455 regs[i].set1(VMRegImpl::stack2reg(stk_reg++));
456 }
457 break;
458
459 #ifdef _LP64
460 case T_OBJECT:
461 case T_ARRAY:
462 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
463 if (int_reg < int_reg_max) {
464 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
465 regs[i].set2(r->as_VMReg());
466 } else {
467 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
468 stk_reg_pairs += 2;
469 }
470 break;
471 #endif // _LP64
472
473 case T_LONG:
474 assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half");
475 #ifdef _LP64
476 if (int_reg < int_reg_max) {
477 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
478 regs[i].set2(r->as_VMReg());
479 } else {
480 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
481 stk_reg_pairs += 2;
482 }
483 #else
484 #ifdef COMPILER2
485 // For 32-bit build, can't pass longs in O-regs because they become
486 // I-regs and get trashed. Use G-regs instead. G1 and G4 are almost
487 // spare and available. This convention isn't used by the Sparc ABI or
488 // anywhere else. If we're tiered then we don't use G-regs because c1
489 // can't deal with them as a "pair". (Tiered makes this code think g's are filled)
490 // G0: zero
491 // G1: 1st Long arg
492 // G2: global allocated to TLS
493 // G3: used in inline cache check
494 // G4: 2nd Long arg
495 // G5: used in inline cache check
496 // G6: used by OS
497 // G7: used by OS
498
499 if (g_reg == G1) {
500 regs[i].set2(G1->as_VMReg()); // This long arg in G1
501 g_reg = G4; // Where the next arg goes
502 } else if (g_reg == G4) {
503 regs[i].set2(G4->as_VMReg()); // The 2nd long arg in G4
504 g_reg = noreg; // No more longs in registers
505 } else {
506 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
507 stk_reg_pairs += 2;
508 }
509 #else // COMPILER2
510 if (int_reg_pairs + 1 < int_reg_max) {
511 if (is_outgoing) {
512 regs[i].set_pair(as_oRegister(int_reg_pairs + 1)->as_VMReg(), as_oRegister(int_reg_pairs)->as_VMReg());
513 } else {
514 regs[i].set_pair(as_iRegister(int_reg_pairs + 1)->as_VMReg(), as_iRegister(int_reg_pairs)->as_VMReg());
515 }
516 int_reg_pairs += 2;
517 } else {
518 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
519 stk_reg_pairs += 2;
520 }
521 #endif // COMPILER2
522 #endif // _LP64
523 break;
524
525 case T_FLOAT:
526 if (flt_reg < flt_reg_max) regs[i].set1(as_FloatRegister(flt_reg++)->as_VMReg());
527 else regs[i].set1( VMRegImpl::stack2reg(stk_reg++));
528 break;
529 case T_DOUBLE:
530 assert(sig_bt[i+1] == T_VOID, "expecting half");
531 if (flt_reg_pairs + 1 < flt_reg_max) {
532 regs[i].set2(as_FloatRegister(flt_reg_pairs)->as_VMReg());
533 flt_reg_pairs += 2;
534 } else {
535 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
536 stk_reg_pairs += 2;
537 }
538 break;
539 case T_VOID: regs[i].set_bad(); break; // Halves of longs & doubles
540 default:
541 ShouldNotReachHere();
542 }
543 }
544
545 // retun the amount of stack space these arguments will need.
546 return stk_reg_pairs;
547
548 }
549
550 // Helper class mostly to avoid passing masm everywhere, and handle
551 // store displacement overflow logic.
552 class AdapterGenerator {
553 MacroAssembler *masm;
554 Register Rdisp;
555 void set_Rdisp(Register r) { Rdisp = r; }
556
557 void patch_callers_callsite();
558
559 // base+st_off points to top of argument
560 int arg_offset(const int st_off) { return st_off; }
561 int next_arg_offset(const int st_off) {
562 return st_off - Interpreter::stackElementSize;
563 }
564
565 // Argument slot values may be loaded first into a register because
566 // they might not fit into displacement.
567 RegisterOrConstant arg_slot(const int st_off);
568 RegisterOrConstant next_arg_slot(const int st_off);
569
570 // Stores long into offset pointed to by base
571 void store_c2i_long(Register r, Register base,
572 const int st_off, bool is_stack);
573 void store_c2i_object(Register r, Register base,
574 const int st_off);
575 void store_c2i_int(Register r, Register base,
576 const int st_off);
577 void store_c2i_double(VMReg r_2,
578 VMReg r_1, Register base, const int st_off);
579 void store_c2i_float(FloatRegister f, Register base,
580 const int st_off);
581
582 public:
583 void gen_c2i_adapter(int total_args_passed,
584 // VMReg max_arg,
585 int comp_args_on_stack, // VMRegStackSlots
586 const BasicType *sig_bt,
587 const VMRegPair *regs,
588 Label& skip_fixup);
589 void gen_i2c_adapter(int total_args_passed,
590 // VMReg max_arg,
591 int comp_args_on_stack, // VMRegStackSlots
592 const BasicType *sig_bt,
593 const VMRegPair *regs);
594
595 AdapterGenerator(MacroAssembler *_masm) : masm(_masm) {}
596 };
597
598
599 // Patch the callers callsite with entry to compiled code if it exists.
600 void AdapterGenerator::patch_callers_callsite() {
601 Label L;
602 __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch);
603 __ br_null(G3_scratch, false, __ pt, L);
604 // Schedule the branch target address early.
605 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
606 // Call into the VM to patch the caller, then jump to compiled callee
607 __ save_frame(4); // Args in compiled layout; do not blow them
608
609 // Must save all the live Gregs the list is:
610 // G1: 1st Long arg (32bit build)
611 // G2: global allocated to TLS
612 // G3: used in inline cache check (scratch)
613 // G4: 2nd Long arg (32bit build);
614 // G5: used in inline cache check (methodOop)
615
616 // The longs must go to the stack by hand since in the 32 bit build they can be trashed by window ops.
617
618 #ifdef _LP64
619 // mov(s,d)
620 __ mov(G1, L1);
621 __ mov(G4, L4);
622 __ mov(G5_method, L5);
623 __ mov(G5_method, O0); // VM needs target method
624 __ mov(I7, O1); // VM needs caller's callsite
625 // Must be a leaf call...
626 // can be very far once the blob has been relocated
627 AddressLiteral dest(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
628 __ relocate(relocInfo::runtime_call_type);
629 __ jumpl_to(dest, O7, O7);
630 __ delayed()->mov(G2_thread, L7_thread_cache);
631 __ mov(L7_thread_cache, G2_thread);
632 __ mov(L1, G1);
633 __ mov(L4, G4);
634 __ mov(L5, G5_method);
635 #else
636 __ stx(G1, FP, -8 + STACK_BIAS);
637 __ stx(G4, FP, -16 + STACK_BIAS);
638 __ mov(G5_method, L5);
639 __ mov(G5_method, O0); // VM needs target method
640 __ mov(I7, O1); // VM needs caller's callsite
641 // Must be a leaf call...
642 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), relocInfo::runtime_call_type);
643 __ delayed()->mov(G2_thread, L7_thread_cache);
644 __ mov(L7_thread_cache, G2_thread);
645 __ ldx(FP, -8 + STACK_BIAS, G1);
646 __ ldx(FP, -16 + STACK_BIAS, G4);
647 __ mov(L5, G5_method);
648 __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
649 #endif /* _LP64 */
650
651 __ restore(); // Restore args
652 __ bind(L);
653 }
654
655
656 RegisterOrConstant AdapterGenerator::arg_slot(const int st_off) {
657 RegisterOrConstant roc(arg_offset(st_off));
658 return __ ensure_simm13_or_reg(roc, Rdisp);
659 }
660
661 RegisterOrConstant AdapterGenerator::next_arg_slot(const int st_off) {
662 RegisterOrConstant roc(next_arg_offset(st_off));
663 return __ ensure_simm13_or_reg(roc, Rdisp);
664 }
665
666
667 // Stores long into offset pointed to by base
668 void AdapterGenerator::store_c2i_long(Register r, Register base,
669 const int st_off, bool is_stack) {
670 #ifdef _LP64
671 // In V9, longs are given 2 64-bit slots in the interpreter, but the
672 // data is passed in only 1 slot.
673 __ stx(r, base, next_arg_slot(st_off));
674 #else
675 #ifdef COMPILER2
676 // Misaligned store of 64-bit data
677 __ stw(r, base, arg_slot(st_off)); // lo bits
678 __ srlx(r, 32, r);
679 __ stw(r, base, next_arg_slot(st_off)); // hi bits
680 #else
681 if (is_stack) {
682 // Misaligned store of 64-bit data
683 __ stw(r, base, arg_slot(st_off)); // lo bits
684 __ srlx(r, 32, r);
685 __ stw(r, base, next_arg_slot(st_off)); // hi bits
686 } else {
687 __ stw(r->successor(), base, arg_slot(st_off) ); // lo bits
688 __ stw(r , base, next_arg_slot(st_off)); // hi bits
689 }
690 #endif // COMPILER2
691 #endif // _LP64
692 }
693
694 void AdapterGenerator::store_c2i_object(Register r, Register base,
695 const int st_off) {
696 __ st_ptr (r, base, arg_slot(st_off));
697 }
698
699 void AdapterGenerator::store_c2i_int(Register r, Register base,
700 const int st_off) {
701 __ st (r, base, arg_slot(st_off));
702 }
703
704 // Stores into offset pointed to by base
705 void AdapterGenerator::store_c2i_double(VMReg r_2,
706 VMReg r_1, Register base, const int st_off) {
707 #ifdef _LP64
708 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
709 // data is passed in only 1 slot.
710 __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
711 #else
712 // Need to marshal 64-bit value from misaligned Lesp loads
713 __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
714 __ stf(FloatRegisterImpl::S, r_2->as_FloatRegister(), base, arg_slot(st_off) );
715 #endif
716 }
717
718 void AdapterGenerator::store_c2i_float(FloatRegister f, Register base,
719 const int st_off) {
720 __ stf(FloatRegisterImpl::S, f, base, arg_slot(st_off));
721 }
722
723 void AdapterGenerator::gen_c2i_adapter(
724 int total_args_passed,
725 // VMReg max_arg,
726 int comp_args_on_stack, // VMRegStackSlots
727 const BasicType *sig_bt,
728 const VMRegPair *regs,
729 Label& skip_fixup) {
730
731 // Before we get into the guts of the C2I adapter, see if we should be here
732 // at all. We've come from compiled code and are attempting to jump to the
733 // interpreter, which means the caller made a static call to get here
734 // (vcalls always get a compiled target if there is one). Check for a
735 // compiled target. If there is one, we need to patch the caller's call.
736 // However we will run interpreted if we come thru here. The next pass
737 // thru the call site will run compiled. If we ran compiled here then
738 // we can (theorectically) do endless i2c->c2i->i2c transitions during
739 // deopt/uncommon trap cycles. If we always go interpreted here then
740 // we can have at most one and don't need to play any tricks to keep
741 // from endlessly growing the stack.
742 //
743 // Actually if we detected that we had an i2c->c2i transition here we
744 // ought to be able to reset the world back to the state of the interpreted
745 // call and not bother building another interpreter arg area. We don't
746 // do that at this point.
747
748 patch_callers_callsite();
749
750 __ bind(skip_fixup);
751
752 // Since all args are passed on the stack, total_args_passed*wordSize is the
753 // space we need. Add in varargs area needed by the interpreter. Round up
754 // to stack alignment.
755 const int arg_size = total_args_passed * Interpreter::stackElementSize;
756 const int varargs_area =
757 (frame::varargs_offset - frame::register_save_words)*wordSize;
758 const int extraspace = round_to(arg_size + varargs_area, 2*wordSize);
759
760 int bias = STACK_BIAS;
761 const int interp_arg_offset = frame::varargs_offset*wordSize +
762 (total_args_passed-1)*Interpreter::stackElementSize;
763
764 Register base = SP;
765
766 #ifdef _LP64
767 // In the 64bit build because of wider slots and STACKBIAS we can run
768 // out of bits in the displacement to do loads and stores. Use g3 as
769 // temporary displacement.
770 if (! __ is_simm13(extraspace)) {
771 __ set(extraspace, G3_scratch);
772 __ sub(SP, G3_scratch, SP);
773 } else {
774 __ sub(SP, extraspace, SP);
775 }
776 set_Rdisp(G3_scratch);
777 #else
778 __ sub(SP, extraspace, SP);
779 #endif // _LP64
780
781 // First write G1 (if used) to where ever it must go
782 for (int i=0; i<total_args_passed; i++) {
783 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize) + bias;
784 VMReg r_1 = regs[i].first();
785 VMReg r_2 = regs[i].second();
786 if (r_1 == G1_scratch->as_VMReg()) {
787 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) {
788 store_c2i_object(G1_scratch, base, st_off);
789 } else if (sig_bt[i] == T_LONG) {
790 assert(!TieredCompilation, "should not use register args for longs");
791 store_c2i_long(G1_scratch, base, st_off, false);
792 } else {
793 store_c2i_int(G1_scratch, base, st_off);
794 }
795 }
796 }
797
798 // Now write the args into the outgoing interpreter space
799 for (int i=0; i<total_args_passed; i++) {
800 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize) + bias;
801 VMReg r_1 = regs[i].first();
802 VMReg r_2 = regs[i].second();
803 if (!r_1->is_valid()) {
804 assert(!r_2->is_valid(), "");
805 continue;
806 }
807 // Skip G1 if found as we did it first in order to free it up
808 if (r_1 == G1_scratch->as_VMReg()) {
809 continue;
810 }
811 #ifdef ASSERT
812 bool G1_forced = false;
813 #endif // ASSERT
814 if (r_1->is_stack()) { // Pretend stack targets are loaded into G1
815 #ifdef _LP64
816 Register ld_off = Rdisp;
817 __ set(reg2offset(r_1) + extraspace + bias, ld_off);
818 #else
819 int ld_off = reg2offset(r_1) + extraspace + bias;
820 #endif // _LP64
821 #ifdef ASSERT
822 G1_forced = true;
823 #endif // ASSERT
824 r_1 = G1_scratch->as_VMReg();// as part of the load/store shuffle
825 if (!r_2->is_valid()) __ ld (base, ld_off, G1_scratch);
826 else __ ldx(base, ld_off, G1_scratch);
827 }
828
829 if (r_1->is_Register()) {
830 Register r = r_1->as_Register()->after_restore();
831 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) {
832 store_c2i_object(r, base, st_off);
833 } else if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
834 #ifndef _LP64
835 if (TieredCompilation) {
836 assert(G1_forced || sig_bt[i] != T_LONG, "should not use register args for longs");
837 }
838 #endif // _LP64
839 store_c2i_long(r, base, st_off, r_2->is_stack());
840 } else {
841 store_c2i_int(r, base, st_off);
842 }
843 } else {
844 assert(r_1->is_FloatRegister(), "");
845 if (sig_bt[i] == T_FLOAT) {
846 store_c2i_float(r_1->as_FloatRegister(), base, st_off);
847 } else {
848 assert(sig_bt[i] == T_DOUBLE, "wrong type");
849 store_c2i_double(r_2, r_1, base, st_off);
850 }
851 }
852 }
853
854 #ifdef _LP64
855 // Need to reload G3_scratch, used for temporary displacements.
856 __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
857
858 // Pass O5_savedSP as an argument to the interpreter.
859 // The interpreter will restore SP to this value before returning.
860 __ set(extraspace, G1);
861 __ add(SP, G1, O5_savedSP);
862 #else
863 // Pass O5_savedSP as an argument to the interpreter.
864 // The interpreter will restore SP to this value before returning.
865 __ add(SP, extraspace, O5_savedSP);
866 #endif // _LP64
867
868 __ mov((frame::varargs_offset)*wordSize -
869 1*Interpreter::stackElementSize+bias+BytesPerWord, G1);
870 // Jump to the interpreter just as if interpreter was doing it.
871 __ jmpl(G3_scratch, 0, G0);
872 // Setup Lesp for the call. Cannot actually set Lesp as the current Lesp
873 // (really L0) is in use by the compiled frame as a generic temp. However,
874 // the interpreter does not know where its args are without some kind of
875 // arg pointer being passed in. Pass it in Gargs.
876 __ delayed()->add(SP, G1, Gargs);
877 }
878
879 void AdapterGenerator::gen_i2c_adapter(
880 int total_args_passed,
881 // VMReg max_arg,
882 int comp_args_on_stack, // VMRegStackSlots
883 const BasicType *sig_bt,
884 const VMRegPair *regs) {
885
886 // Generate an I2C adapter: adjust the I-frame to make space for the C-frame
887 // layout. Lesp was saved by the calling I-frame and will be restored on
888 // return. Meanwhile, outgoing arg space is all owned by the callee
889 // C-frame, so we can mangle it at will. After adjusting the frame size,
890 // hoist register arguments and repack other args according to the compiled
891 // code convention. Finally, end in a jump to the compiled code. The entry
892 // point address is the start of the buffer.
893
894 // We will only enter here from an interpreted frame and never from after
895 // passing thru a c2i. Azul allowed this but we do not. If we lose the
896 // race and use a c2i we will remain interpreted for the race loser(s).
897 // This removes all sorts of headaches on the x86 side and also eliminates
898 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
899
900 // As you can see from the list of inputs & outputs there are not a lot
901 // of temp registers to work with: mostly G1, G3 & G4.
902
903 // Inputs:
904 // G2_thread - TLS
905 // G5_method - Method oop
906 // G4 (Gargs) - Pointer to interpreter's args
907 // O0..O4 - free for scratch
908 // O5_savedSP - Caller's saved SP, to be restored if needed
909 // O6 - Current SP!
910 // O7 - Valid return address
911 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
912
913 // Outputs:
914 // G2_thread - TLS
915 // G1, G4 - Outgoing long args in 32-bit build
916 // O0-O5 - Outgoing args in compiled layout
917 // O6 - Adjusted or restored SP
918 // O7 - Valid return address
919 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
920 // F0-F7 - more outgoing args
921
922
923 // Gargs is the incoming argument base, and also an outgoing argument.
924 __ sub(Gargs, BytesPerWord, Gargs);
925
926 // ON ENTRY TO THE CODE WE ARE MAKING, WE HAVE AN INTERPRETED FRAME
927 // WITH O7 HOLDING A VALID RETURN PC
928 //
929 // | |
930 // : java stack :
931 // | |
932 // +--------------+ <--- start of outgoing args
933 // | receiver | |
934 // : rest of args : |---size is java-arg-words
935 // | | |
936 // +--------------+ <--- O4_args (misaligned) and Lesp if prior is not C2I
937 // | | |
938 // : unused : |---Space for max Java stack, plus stack alignment
939 // | | |
940 // +--------------+ <--- SP + 16*wordsize
941 // | |
942 // : window :
943 // | |
944 // +--------------+ <--- SP
945
946 // WE REPACK THE STACK. We use the common calling convention layout as
947 // discovered by calling SharedRuntime::calling_convention. We assume it
948 // causes an arbitrary shuffle of memory, which may require some register
949 // temps to do the shuffle. We hope for (and optimize for) the case where
950 // temps are not needed. We may have to resize the stack slightly, in case
951 // we need alignment padding (32-bit interpreter can pass longs & doubles
952 // misaligned, but the compilers expect them aligned).
953 //
954 // | |
955 // : java stack :
956 // | |
957 // +--------------+ <--- start of outgoing args
958 // | pad, align | |
959 // +--------------+ |
960 // | ints, floats | |---Outgoing stack args, packed low.
961 // +--------------+ | First few args in registers.
962 // : doubles : |
963 // | longs | |
964 // +--------------+ <--- SP' + 16*wordsize
965 // | |
966 // : window :
967 // | |
968 // +--------------+ <--- SP'
969
970 // ON EXIT FROM THE CODE WE ARE MAKING, WE STILL HAVE AN INTERPRETED FRAME
971 // WITH O7 HOLDING A VALID RETURN PC - ITS JUST THAT THE ARGS ARE NOW SETUP
972 // FOR COMPILED CODE AND THE FRAME SLIGHTLY GROWN.
973
974 // Cut-out for having no stack args. Since up to 6 args are passed
975 // in registers, we will commonly have no stack args.
976 if (comp_args_on_stack > 0) {
977
978 // Convert VMReg stack slots to words.
979 int comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
980 // Round up to miminum stack alignment, in wordSize
981 comp_words_on_stack = round_to(comp_words_on_stack, 2);
982 // Now compute the distance from Lesp to SP. This calculation does not
983 // include the space for total_args_passed because Lesp has not yet popped
984 // the arguments.
985 __ sub(SP, (comp_words_on_stack)*wordSize, SP);
986 }
987
988 // Will jump to the compiled code just as if compiled code was doing it.
989 // Pre-load the register-jump target early, to schedule it better.
990 __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_compiled_offset()), G3);
991
992 // Now generate the shuffle code. Pick up all register args and move the
993 // rest through G1_scratch.
994 for (int i=0; i<total_args_passed; i++) {
995 if (sig_bt[i] == T_VOID) {
996 // Longs and doubles are passed in native word order, but misaligned
997 // in the 32-bit build.
998 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
999 continue;
1000 }
1001
1002 // Pick up 0, 1 or 2 words from Lesp+offset. Assume mis-aligned in the
1003 // 32-bit build and aligned in the 64-bit build. Look for the obvious
1004 // ldx/lddf optimizations.
1005
1006 // Load in argument order going down.
1007 const int ld_off = (total_args_passed-i)*Interpreter::stackElementSize;
1008 set_Rdisp(G1_scratch);
1009
1010 VMReg r_1 = regs[i].first();
1011 VMReg r_2 = regs[i].second();
1012 if (!r_1->is_valid()) {
1013 assert(!r_2->is_valid(), "");
1014 continue;
1015 }
1016 if (r_1->is_stack()) { // Pretend stack targets are loaded into F8/F9
1017 r_1 = F8->as_VMReg(); // as part of the load/store shuffle
1018 if (r_2->is_valid()) r_2 = r_1->next();
1019 }
1020 if (r_1->is_Register()) { // Register argument
1021 Register r = r_1->as_Register()->after_restore();
1022 if (!r_2->is_valid()) {
1023 __ ld(Gargs, arg_slot(ld_off), r);
1024 } else {
1025 #ifdef _LP64
1026 // In V9, longs are given 2 64-bit slots in the interpreter, but the
1027 // data is passed in only 1 slot.
1028 RegisterOrConstant slot = (sig_bt[i] == T_LONG) ?
1029 next_arg_slot(ld_off) : arg_slot(ld_off);
1030 __ ldx(Gargs, slot, r);
1031 #else
1032 // Need to load a 64-bit value into G1/G4, but G1/G4 is being used in the
1033 // stack shuffle. Load the first 2 longs into G1/G4 later.
1034 #endif
1035 }
1036 } else {
1037 assert(r_1->is_FloatRegister(), "");
1038 if (!r_2->is_valid()) {
1039 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_1->as_FloatRegister());
1040 } else {
1041 #ifdef _LP64
1042 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
1043 // data is passed in only 1 slot. This code also handles longs that
1044 // are passed on the stack, but need a stack-to-stack move through a
1045 // spare float register.
1046 RegisterOrConstant slot = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
1047 next_arg_slot(ld_off) : arg_slot(ld_off);
1048 __ ldf(FloatRegisterImpl::D, Gargs, slot, r_1->as_FloatRegister());
1049 #else
1050 // Need to marshal 64-bit value from misaligned Lesp loads
1051 __ ldf(FloatRegisterImpl::S, Gargs, next_arg_slot(ld_off), r_1->as_FloatRegister());
1052 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_2->as_FloatRegister());
1053 #endif
1054 }
1055 }
1056 // Was the argument really intended to be on the stack, but was loaded
1057 // into F8/F9?
1058 if (regs[i].first()->is_stack()) {
1059 assert(r_1->as_FloatRegister() == F8, "fix this code");
1060 // Convert stack slot to an SP offset
1061 int st_off = reg2offset(regs[i].first()) + STACK_BIAS;
1062 // Store down the shuffled stack word. Target address _is_ aligned.
1063 RegisterOrConstant slot = __ ensure_simm13_or_reg(st_off, Rdisp);
1064 if (!r_2->is_valid()) __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), SP, slot);
1065 else __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), SP, slot);
1066 }
1067 }
1068 bool made_space = false;
1069 #ifndef _LP64
1070 // May need to pick up a few long args in G1/G4
1071 bool g4_crushed = false;
1072 bool g3_crushed = false;
1073 for (int i=0; i<total_args_passed; i++) {
1074 if (regs[i].first()->is_Register() && regs[i].second()->is_valid()) {
1075 // Load in argument order going down
1076 int ld_off = (total_args_passed-i)*Interpreter::stackElementSize;
1077 // Need to marshal 64-bit value from misaligned Lesp loads
1078 Register r = regs[i].first()->as_Register()->after_restore();
1079 if (r == G1 || r == G4) {
1080 assert(!g4_crushed, "ordering problem");
1081 if (r == G4){
1082 g4_crushed = true;
1083 __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits
1084 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
1085 } else {
1086 // better schedule this way
1087 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
1088 __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits
1089 }
1090 g3_crushed = true;
1091 __ sllx(r, 32, r);
1092 __ or3(G3_scratch, r, r);
1093 } else {
1094 assert(r->is_out(), "longs passed in two O registers");
1095 __ ld (Gargs, arg_slot(ld_off) , r->successor()); // Load lo bits
1096 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
1097 }
1098 }
1099 }
1100 #endif
1101
1102 // Jump to the compiled code just as if compiled code was doing it.
1103 //
1104 #ifndef _LP64
1105 if (g3_crushed) {
1106 // Rats load was wasted, at least it is in cache...
1107 __ ld_ptr(G5_method, methodOopDesc::from_compiled_offset(), G3);
1108 }
1109 #endif /* _LP64 */
1110
1111 // 6243940 We might end up in handle_wrong_method if
1112 // the callee is deoptimized as we race thru here. If that
1113 // happens we don't want to take a safepoint because the
1114 // caller frame will look interpreted and arguments are now
1115 // "compiled" so it is much better to make this transition
1116 // invisible to the stack walking code. Unfortunately if
1117 // we try and find the callee by normal means a safepoint
1118 // is possible. So we stash the desired callee in the thread
1119 // and the vm will find there should this case occur.
1120 Address callee_target_addr(G2_thread, JavaThread::callee_target_offset());
1121 __ st_ptr(G5_method, callee_target_addr);
1122
1123 if (StressNonEntrant) {
1124 // Open a big window for deopt failure
1125 __ save_frame(0);
1126 __ mov(G0, L0);
1127 Label loop;
1128 __ bind(loop);
1129 __ sub(L0, 1, L0);
1130 __ br_null(L0, false, Assembler::pt, loop);
1131 __ delayed()->nop();
1132
1133 __ restore();
1134 }
1135
1136
1137 __ jmpl(G3, 0, G0);
1138 __ delayed()->nop();
1139 }
1140
1141 // ---------------------------------------------------------------
1142 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1143 int total_args_passed,
1144 // VMReg max_arg,
1145 int comp_args_on_stack, // VMRegStackSlots
1146 const BasicType *sig_bt,
1147 const VMRegPair *regs,
1148 AdapterFingerPrint* fingerprint) {
1149 address i2c_entry = __ pc();
1150
1151 AdapterGenerator agen(masm);
1152
1153 agen.gen_i2c_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs);
1154
1155
1156 // -------------------------------------------------------------------------
1157 // Generate a C2I adapter. On entry we know G5 holds the methodOop. The
1158 // args start out packed in the compiled layout. They need to be unpacked
1159 // into the interpreter layout. This will almost always require some stack
1160 // space. We grow the current (compiled) stack, then repack the args. We
1161 // finally end in a jump to the generic interpreter entry point. On exit
1162 // from the interpreter, the interpreter will restore our SP (lest the
1163 // compiled code, which relys solely on SP and not FP, get sick).
1164
1165 address c2i_unverified_entry = __ pc();
1166 Label skip_fixup;
1167 {
1168 #if !defined(_LP64) && defined(COMPILER2)
1169 Register R_temp = L0; // another scratch register
1170 #else
1171 Register R_temp = G1; // another scratch register
1172 #endif
1173
1174 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
1175
1176 __ verify_oop(O0);
1177 __ verify_oop(G5_method);
1178 __ load_klass(O0, G3_scratch);
1179 __ verify_oop(G3_scratch);
1180
1181 #if !defined(_LP64) && defined(COMPILER2)
1182 __ save(SP, -frame::register_save_words*wordSize, SP);
1183 __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp);
1184 __ verify_oop(R_temp);
1185 __ cmp(G3_scratch, R_temp);
1186 __ restore();
1187 #else
1188 __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp);
1189 __ verify_oop(R_temp);
1190 __ cmp(G3_scratch, R_temp);
1191 #endif
1192
1193 Label ok, ok2;
1194 __ brx(Assembler::equal, false, Assembler::pt, ok);
1195 __ delayed()->ld_ptr(G5_method, compiledICHolderOopDesc::holder_method_offset(), G5_method);
1196 __ jump_to(ic_miss, G3_scratch);
1197 __ delayed()->nop();
1198
1199 __ bind(ok);
1200 // Method might have been compiled since the call site was patched to
1201 // interpreted if that is the case treat it as a miss so we can get
1202 // the call site corrected.
1203 __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch);
1204 __ bind(ok2);
1205 __ br_null(G3_scratch, false, __ pt, skip_fixup);
1206 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
1207 __ jump_to(ic_miss, G3_scratch);
1208 __ delayed()->nop();
1209
1210 }
1211
1212 address c2i_entry = __ pc();
1213
1214 agen.gen_c2i_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
1215
1216 __ flush();
1217 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry);
1218
1219 }
1220
1221 // Helper function for native calling conventions
1222 static VMReg int_stk_helper( int i ) {
1223 // Bias any stack based VMReg we get by ignoring the window area
1224 // but not the register parameter save area.
1225 //
1226 // This is strange for the following reasons. We'd normally expect
1227 // the calling convention to return an VMReg for a stack slot
1228 // completely ignoring any abi reserved area. C2 thinks of that
1229 // abi area as only out_preserve_stack_slots. This does not include
1230 // the area allocated by the C abi to store down integer arguments
1231 // because the java calling convention does not use it. So
1232 // since c2 assumes that there are only out_preserve_stack_slots
1233 // to bias the optoregs (which impacts VMRegs) when actually referencing any actual stack
1234 // location the c calling convention must add in this bias amount
1235 // to make up for the fact that the out_preserve_stack_slots is
1236 // insufficient for C calls. What a mess. I sure hope those 6
1237 // stack words were worth it on every java call!
1238
1239 // Another way of cleaning this up would be for out_preserve_stack_slots
1240 // to take a parameter to say whether it was C or java calling conventions.
1241 // Then things might look a little better (but not much).
1242
1243 int mem_parm_offset = i - SPARC_ARGS_IN_REGS_NUM;
1244 if( mem_parm_offset < 0 ) {
1245 return as_oRegister(i)->as_VMReg();
1246 } else {
1247 int actual_offset = (mem_parm_offset + frame::memory_parameter_word_sp_offset) * VMRegImpl::slots_per_word;
1248 // Now return a biased offset that will be correct when out_preserve_slots is added back in
1249 return VMRegImpl::stack2reg(actual_offset - SharedRuntime::out_preserve_stack_slots());
1250 }
1251 }
1252
1253
1254 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1255 VMRegPair *regs,
1256 int total_args_passed) {
1257
1258 // Return the number of VMReg stack_slots needed for the args.
1259 // This value does not include an abi space (like register window
1260 // save area).
1261
1262 // The native convention is V8 if !LP64
1263 // The LP64 convention is the V9 convention which is slightly more sane.
1264
1265 // We return the amount of VMReg stack slots we need to reserve for all
1266 // the arguments NOT counting out_preserve_stack_slots. Since we always
1267 // have space for storing at least 6 registers to memory we start with that.
1268 // See int_stk_helper for a further discussion.
1269 int max_stack_slots = (frame::varargs_offset * VMRegImpl::slots_per_word) - SharedRuntime::out_preserve_stack_slots();
1270
1271 #ifdef _LP64
1272 // V9 convention: All things "as-if" on double-wide stack slots.
1273 // Hoist any int/ptr/long's in the first 6 to int regs.
1274 // Hoist any flt/dbl's in the first 16 dbl regs.
1275 int j = 0; // Count of actual args, not HALVES
1276 for( int i=0; i<total_args_passed; i++, j++ ) {
1277 switch( sig_bt[i] ) {
1278 case T_BOOLEAN:
1279 case T_BYTE:
1280 case T_CHAR:
1281 case T_INT:
1282 case T_SHORT:
1283 regs[i].set1( int_stk_helper( j ) ); break;
1284 case T_LONG:
1285 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1286 case T_ADDRESS: // raw pointers, like current thread, for VM calls
1287 case T_ARRAY:
1288 case T_OBJECT:
1289 regs[i].set2( int_stk_helper( j ) );
1290 break;
1291 case T_FLOAT:
1292 if ( j < 16 ) {
1293 // V9ism: floats go in ODD registers
1294 regs[i].set1(as_FloatRegister(1 + (j<<1))->as_VMReg());
1295 } else {
1296 // V9ism: floats go in ODD stack slot
1297 regs[i].set1(VMRegImpl::stack2reg(1 + (j<<1)));
1298 }
1299 break;
1300 case T_DOUBLE:
1301 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1302 if ( j < 16 ) {
1303 // V9ism: doubles go in EVEN/ODD regs
1304 regs[i].set2(as_FloatRegister(j<<1)->as_VMReg());
1305 } else {
1306 // V9ism: doubles go in EVEN/ODD stack slots
1307 regs[i].set2(VMRegImpl::stack2reg(j<<1));
1308 }
1309 break;
1310 case T_VOID: regs[i].set_bad(); j--; break; // Do not count HALVES
1311 default:
1312 ShouldNotReachHere();
1313 }
1314 if (regs[i].first()->is_stack()) {
1315 int off = regs[i].first()->reg2stack();
1316 if (off > max_stack_slots) max_stack_slots = off;
1317 }
1318 if (regs[i].second()->is_stack()) {
1319 int off = regs[i].second()->reg2stack();
1320 if (off > max_stack_slots) max_stack_slots = off;
1321 }
1322 }
1323
1324 #else // _LP64
1325 // V8 convention: first 6 things in O-regs, rest on stack.
1326 // Alignment is willy-nilly.
1327 for( int i=0; i<total_args_passed; i++ ) {
1328 switch( sig_bt[i] ) {
1329 case T_ADDRESS: // raw pointers, like current thread, for VM calls
1330 case T_ARRAY:
1331 case T_BOOLEAN:
1332 case T_BYTE:
1333 case T_CHAR:
1334 case T_FLOAT:
1335 case T_INT:
1336 case T_OBJECT:
1337 case T_SHORT:
1338 regs[i].set1( int_stk_helper( i ) );
1339 break;
1340 case T_DOUBLE:
1341 case T_LONG:
1342 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1343 regs[i].set_pair( int_stk_helper( i+1 ), int_stk_helper( i ) );
1344 break;
1345 case T_VOID: regs[i].set_bad(); break;
1346 default:
1347 ShouldNotReachHere();
1348 }
1349 if (regs[i].first()->is_stack()) {
1350 int off = regs[i].first()->reg2stack();
1351 if (off > max_stack_slots) max_stack_slots = off;
1352 }
1353 if (regs[i].second()->is_stack()) {
1354 int off = regs[i].second()->reg2stack();
1355 if (off > max_stack_slots) max_stack_slots = off;
1356 }
1357 }
1358 #endif // _LP64
1359
1360 return round_to(max_stack_slots + 1, 2);
1361
1362 }
1363
1364
1365 // ---------------------------------------------------------------------------
1366 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1367 switch (ret_type) {
1368 case T_FLOAT:
1369 __ stf(FloatRegisterImpl::S, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS);
1370 break;
1371 case T_DOUBLE:
1372 __ stf(FloatRegisterImpl::D, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS);
1373 break;
1374 }
1375 }
1376
1377 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1378 switch (ret_type) {
1379 case T_FLOAT:
1380 __ ldf(FloatRegisterImpl::S, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS, F0);
1381 break;
1382 case T_DOUBLE:
1383 __ ldf(FloatRegisterImpl::D, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS, F0);
1384 break;
1385 }
1386 }
1387
1388 // Check and forward and pending exception. Thread is stored in
1389 // L7_thread_cache and possibly NOT in G2_thread. Since this is a native call, there
1390 // is no exception handler. We merely pop this frame off and throw the
1391 // exception in the caller's frame.
1392 static void check_forward_pending_exception(MacroAssembler *masm, Register Rex_oop) {
1393 Label L;
1394 __ br_null(Rex_oop, false, Assembler::pt, L);
1395 __ delayed()->mov(L7_thread_cache, G2_thread); // restore in case we have exception
1396 // Since this is a native call, we *know* the proper exception handler
1397 // without calling into the VM: it's the empty function. Just pop this
1398 // frame and then jump to forward_exception_entry; O7 will contain the
1399 // native caller's return PC.
1400 AddressLiteral exception_entry(StubRoutines::forward_exception_entry());
1401 __ jump_to(exception_entry, G3_scratch);
1402 __ delayed()->restore(); // Pop this frame off.
1403 __ bind(L);
1404 }
1405
1406 // A simple move of integer like type
1407 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1408 if (src.first()->is_stack()) {
1409 if (dst.first()->is_stack()) {
1410 // stack to stack
1411 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1412 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1413 } else {
1414 // stack to reg
1415 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1416 }
1417 } else if (dst.first()->is_stack()) {
1418 // reg to stack
1419 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1420 } else {
1421 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1422 }
1423 }
1424
1425 // On 64 bit we will store integer like items to the stack as
1426 // 64 bits items (sparc abi) even though java would only store
1427 // 32bits for a parameter. On 32bit it will simply be 32 bits
1428 // So this routine will do 32->32 on 32bit and 32->64 on 64bit
1429 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1430 if (src.first()->is_stack()) {
1431 if (dst.first()->is_stack()) {
1432 // stack to stack
1433 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1434 __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1435 } else {
1436 // stack to reg
1437 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1438 }
1439 } else if (dst.first()->is_stack()) {
1440 // reg to stack
1441 __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1442 } else {
1443 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1444 }
1445 }
1446
1447
1448 // An oop arg. Must pass a handle not the oop itself
1449 static void object_move(MacroAssembler* masm,
1450 OopMap* map,
1451 int oop_handle_offset,
1452 int framesize_in_slots,
1453 VMRegPair src,
1454 VMRegPair dst,
1455 bool is_receiver,
1456 int* receiver_offset) {
1457
1458 // must pass a handle. First figure out the location we use as a handle
1459
1460 if (src.first()->is_stack()) {
1461 // Oop is already on the stack
1462 Register rHandle = dst.first()->is_stack() ? L5 : dst.first()->as_Register();
1463 __ add(FP, reg2offset(src.first()) + STACK_BIAS, rHandle);
1464 __ ld_ptr(rHandle, 0, L4);
1465 #ifdef _LP64
1466 __ movr( Assembler::rc_z, L4, G0, rHandle );
1467 #else
1468 __ tst( L4 );
1469 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
1470 #endif
1471 if (dst.first()->is_stack()) {
1472 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
1473 }
1474 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1475 if (is_receiver) {
1476 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
1477 }
1478 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
1479 } else {
1480 // Oop is in an input register pass we must flush it to the stack
1481 const Register rOop = src.first()->as_Register();
1482 const Register rHandle = L5;
1483 int oop_slot = rOop->input_number() * VMRegImpl::slots_per_word + oop_handle_offset;
1484 int offset = oop_slot*VMRegImpl::stack_slot_size;
1485 Label skip;
1486 __ st_ptr(rOop, SP, offset + STACK_BIAS);
1487 if (is_receiver) {
1488 *receiver_offset = oop_slot * VMRegImpl::stack_slot_size;
1489 }
1490 map->set_oop(VMRegImpl::stack2reg(oop_slot));
1491 __ add(SP, offset + STACK_BIAS, rHandle);
1492 #ifdef _LP64
1493 __ movr( Assembler::rc_z, rOop, G0, rHandle );
1494 #else
1495 __ tst( rOop );
1496 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
1497 #endif
1498
1499 if (dst.first()->is_stack()) {
1500 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
1501 } else {
1502 __ mov(rHandle, dst.first()->as_Register());
1503 }
1504 }
1505 }
1506
1507 // A float arg may have to do float reg int reg conversion
1508 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1509 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move");
1510
1511 if (src.first()->is_stack()) {
1512 if (dst.first()->is_stack()) {
1513 // stack to stack the easiest of the bunch
1514 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1515 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1516 } else {
1517 // stack to reg
1518 if (dst.first()->is_Register()) {
1519 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1520 } else {
1521 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1522 }
1523 }
1524 } else if (dst.first()->is_stack()) {
1525 // reg to stack
1526 if (src.first()->is_Register()) {
1527 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1528 } else {
1529 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1530 }
1531 } else {
1532 // reg to reg
1533 if (src.first()->is_Register()) {
1534 if (dst.first()->is_Register()) {
1535 // gpr -> gpr
1536 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1537 } else {
1538 // gpr -> fpr
1539 __ st(src.first()->as_Register(), FP, -4 + STACK_BIAS);
1540 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.first()->as_FloatRegister());
1541 }
1542 } else if (dst.first()->is_Register()) {
1543 // fpr -> gpr
1544 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), FP, -4 + STACK_BIAS);
1545 __ ld(FP, -4 + STACK_BIAS, dst.first()->as_Register());
1546 } else {
1547 // fpr -> fpr
1548 // In theory these overlap but the ordering is such that this is likely a nop
1549 if ( src.first() != dst.first()) {
1550 __ fmov(FloatRegisterImpl::S, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
1551 }
1552 }
1553 }
1554 }
1555
1556 static void split_long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1557 VMRegPair src_lo(src.first());
1558 VMRegPair src_hi(src.second());
1559 VMRegPair dst_lo(dst.first());
1560 VMRegPair dst_hi(dst.second());
1561 simple_move32(masm, src_lo, dst_lo);
1562 simple_move32(masm, src_hi, dst_hi);
1563 }
1564
1565 // A long move
1566 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1567
1568 // Do the simple ones here else do two int moves
1569 if (src.is_single_phys_reg() ) {
1570 if (dst.is_single_phys_reg()) {
1571 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1572 } else {
1573 // split src into two separate registers
1574 // Remember hi means hi address or lsw on sparc
1575 // Move msw to lsw
1576 if (dst.second()->is_reg()) {
1577 // MSW -> MSW
1578 __ srax(src.first()->as_Register(), 32, dst.first()->as_Register());
1579 // Now LSW -> LSW
1580 // this will only move lo -> lo and ignore hi
1581 VMRegPair split(dst.second());
1582 simple_move32(masm, src, split);
1583 } else {
1584 VMRegPair split(src.first(), L4->as_VMReg());
1585 // MSW -> MSW (lo ie. first word)
1586 __ srax(src.first()->as_Register(), 32, L4);
1587 split_long_move(masm, split, dst);
1588 }
1589 }
1590 } else if (dst.is_single_phys_reg()) {
1591 if (src.is_adjacent_aligned_on_stack(2)) {
1592 __ ldx(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1593 } else {
1594 // dst is a single reg.
1595 // Remember lo is low address not msb for stack slots
1596 // and lo is the "real" register for registers
1597 // src is
1598
1599 VMRegPair split;
1600
1601 if (src.first()->is_reg()) {
1602 // src.lo (msw) is a reg, src.hi is stk/reg
1603 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> src.lo [the MSW is in the LSW of the reg]
1604 split.set_pair(dst.first(), src.first());
1605 } else {
1606 // msw is stack move to L5
1607 // lsw is stack move to dst.lo (real reg)
1608 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> L5
1609 split.set_pair(dst.first(), L5->as_VMReg());
1610 }
1611
1612 // src.lo -> src.lo/L5, src.hi -> dst.lo (the real reg)
1613 // msw -> src.lo/L5, lsw -> dst.lo
1614 split_long_move(masm, src, split);
1615
1616 // So dst now has the low order correct position the
1617 // msw half
1618 __ sllx(split.first()->as_Register(), 32, L5);
1619
1620 const Register d = dst.first()->as_Register();
1621 __ or3(L5, d, d);
1622 }
1623 } else {
1624 // For LP64 we can probably do better.
1625 split_long_move(masm, src, dst);
1626 }
1627 }
1628
1629 // A double move
1630 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1631
1632 // The painful thing here is that like long_move a VMRegPair might be
1633 // 1: a single physical register
1634 // 2: two physical registers (v8)
1635 // 3: a physical reg [lo] and a stack slot [hi] (v8)
1636 // 4: two stack slots
1637
1638 // Since src is always a java calling convention we know that the src pair
1639 // is always either all registers or all stack (and aligned?)
1640
1641 // in a register [lo] and a stack slot [hi]
1642 if (src.first()->is_stack()) {
1643 if (dst.first()->is_stack()) {
1644 // stack to stack the easiest of the bunch
1645 // ought to be a way to do this where if alignment is ok we use ldd/std when possible
1646 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1647 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1648 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1649 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1650 } else {
1651 // stack to reg
1652 if (dst.second()->is_stack()) {
1653 // stack -> reg, stack -> stack
1654 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1655 if (dst.first()->is_Register()) {
1656 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1657 } else {
1658 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1659 }
1660 // This was missing. (very rare case)
1661 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1662 } else {
1663 // stack -> reg
1664 // Eventually optimize for alignment QQQ
1665 if (dst.first()->is_Register()) {
1666 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1667 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_Register());
1668 } else {
1669 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1670 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_FloatRegister());
1671 }
1672 }
1673 }
1674 } else if (dst.first()->is_stack()) {
1675 // reg to stack
1676 if (src.first()->is_Register()) {
1677 // Eventually optimize for alignment QQQ
1678 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1679 if (src.second()->is_stack()) {
1680 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1681 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1682 } else {
1683 __ st(src.second()->as_Register(), SP, reg2offset(dst.second()) + STACK_BIAS);
1684 }
1685 } else {
1686 // fpr to stack
1687 if (src.second()->is_stack()) {
1688 ShouldNotReachHere();
1689 } else {
1690 // Is the stack aligned?
1691 if (reg2offset(dst.first()) & 0x7) {
1692 // No do as pairs
1693 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1694 __ stf(FloatRegisterImpl::S, src.second()->as_FloatRegister(), SP, reg2offset(dst.second()) + STACK_BIAS);
1695 } else {
1696 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1697 }
1698 }
1699 }
1700 } else {
1701 // reg to reg
1702 if (src.first()->is_Register()) {
1703 if (dst.first()->is_Register()) {
1704 // gpr -> gpr
1705 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1706 __ mov(src.second()->as_Register(), dst.second()->as_Register());
1707 } else {
1708 // gpr -> fpr
1709 // ought to be able to do a single store
1710 __ stx(src.first()->as_Register(), FP, -8 + STACK_BIAS);
1711 __ stx(src.second()->as_Register(), FP, -4 + STACK_BIAS);
1712 // ought to be able to do a single load
1713 __ ldf(FloatRegisterImpl::S, FP, -8 + STACK_BIAS, dst.first()->as_FloatRegister());
1714 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.second()->as_FloatRegister());
1715 }
1716 } else if (dst.first()->is_Register()) {
1717 // fpr -> gpr
1718 // ought to be able to do a single store
1719 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), FP, -8 + STACK_BIAS);
1720 // ought to be able to do a single load
1721 // REMEMBER first() is low address not LSB
1722 __ ld(FP, -8 + STACK_BIAS, dst.first()->as_Register());
1723 if (dst.second()->is_Register()) {
1724 __ ld(FP, -4 + STACK_BIAS, dst.second()->as_Register());
1725 } else {
1726 __ ld(FP, -4 + STACK_BIAS, L4);
1727 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1728 }
1729 } else {
1730 // fpr -> fpr
1731 // In theory these overlap but the ordering is such that this is likely a nop
1732 if ( src.first() != dst.first()) {
1733 __ fmov(FloatRegisterImpl::D, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
1734 }
1735 }
1736 }
1737 }
1738
1739 // Creates an inner frame if one hasn't already been created, and
1740 // saves a copy of the thread in L7_thread_cache
1741 static void create_inner_frame(MacroAssembler* masm, bool* already_created) {
1742 if (!*already_created) {
1743 __ save_frame(0);
1744 // Save thread in L7 (INNER FRAME); it crosses a bunch of VM calls below
1745 // Don't use save_thread because it smashes G2 and we merely want to save a
1746 // copy
1747 __ mov(G2_thread, L7_thread_cache);
1748 *already_created = true;
1749 }
1750 }
1751
1752 // ---------------------------------------------------------------------------
1753 // Generate a native wrapper for a given method. The method takes arguments
1754 // in the Java compiled code convention, marshals them to the native
1755 // convention (handlizes oops, etc), transitions to native, makes the call,
1756 // returns to java state (possibly blocking), unhandlizes any result and
1757 // returns.
1758 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1759 methodHandle method,
1760 int compile_id,
1761 int total_in_args,
1762 int comp_args_on_stack, // in VMRegStackSlots
1763 BasicType *in_sig_bt,
1764 VMRegPair *in_regs,
1765 BasicType ret_type) {
1766
1767 // Native nmethod wrappers never take possesion of the oop arguments.
1768 // So the caller will gc the arguments. The only thing we need an
1769 // oopMap for is if the call is static
1770 //
1771 // An OopMap for lock (and class if static), and one for the VM call itself
1772 OopMapSet *oop_maps = new OopMapSet();
1773 intptr_t start = (intptr_t)__ pc();
1774
1775 // First thing make an ic check to see if we should even be here
1776 {
1777 Label L;
1778 const Register temp_reg = G3_scratch;
1779 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
1780 __ verify_oop(O0);
1781 __ load_klass(O0, temp_reg);
1782 __ cmp(temp_reg, G5_inline_cache_reg);
1783 __ brx(Assembler::equal, true, Assembler::pt, L);
1784 __ delayed()->nop();
1785
1786 __ jump_to(ic_miss, temp_reg);
1787 __ delayed()->nop();
1788 __ align(CodeEntryAlignment);
1789 __ bind(L);
1790 }
1791
1792 int vep_offset = ((intptr_t)__ pc()) - start;
1793
1794 #ifdef COMPILER1
1795 if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) {
1796 // Object.hashCode can pull the hashCode from the header word
1797 // instead of doing a full VM transition once it's been computed.
1798 // Since hashCode is usually polymorphic at call sites we can't do
1799 // this optimization at the call site without a lot of work.
1800 Label slowCase;
1801 Register receiver = O0;
1802 Register result = O0;
1803 Register header = G3_scratch;
1804 Register hash = G3_scratch; // overwrite header value with hash value
1805 Register mask = G1; // to get hash field from header
1806
1807 // Read the header and build a mask to get its hash field. Give up if the object is not unlocked.
1808 // We depend on hash_mask being at most 32 bits and avoid the use of
1809 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
1810 // vm: see markOop.hpp.
1811 __ ld_ptr(receiver, oopDesc::mark_offset_in_bytes(), header);
1812 __ sethi(markOopDesc::hash_mask, mask);
1813 __ btst(markOopDesc::unlocked_value, header);
1814 __ br(Assembler::zero, false, Assembler::pn, slowCase);
1815 if (UseBiasedLocking) {
1816 // Check if biased and fall through to runtime if so
1817 __ delayed()->nop();
1818 __ btst(markOopDesc::biased_lock_bit_in_place, header);
1819 __ br(Assembler::notZero, false, Assembler::pn, slowCase);
1820 }
1821 __ delayed()->or3(mask, markOopDesc::hash_mask & 0x3ff, mask);
1822
1823 // Check for a valid (non-zero) hash code and get its value.
1824 #ifdef _LP64
1825 __ srlx(header, markOopDesc::hash_shift, hash);
1826 #else
1827 __ srl(header, markOopDesc::hash_shift, hash);
1828 #endif
1829 __ andcc(hash, mask, hash);
1830 __ br(Assembler::equal, false, Assembler::pn, slowCase);
1831 __ delayed()->nop();
1832
1833 // leaf return.
1834 __ retl();
1835 __ delayed()->mov(hash, result);
1836 __ bind(slowCase);
1837 }
1838 #endif // COMPILER1
1839
1840
1841 // We have received a description of where all the java arg are located
1842 // on entry to the wrapper. We need to convert these args to where
1843 // the jni function will expect them. To figure out where they go
1844 // we convert the java signature to a C signature by inserting
1845 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1846
1847 int total_c_args = total_in_args + 1;
1848 if (method->is_static()) {
1849 total_c_args++;
1850 }
1851
1852 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1853 VMRegPair * out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1854
1855 int argc = 0;
1856 out_sig_bt[argc++] = T_ADDRESS;
1857 if (method->is_static()) {
1858 out_sig_bt[argc++] = T_OBJECT;
1859 }
1860
1861 for (int i = 0; i < total_in_args ; i++ ) {
1862 out_sig_bt[argc++] = in_sig_bt[i];
1863 }
1864
1865 // Now figure out where the args must be stored and how much stack space
1866 // they require (neglecting out_preserve_stack_slots but space for storing
1867 // the 1st six register arguments). It's weird see int_stk_helper.
1868 //
1869 int out_arg_slots;
1870 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1871
1872 // Compute framesize for the wrapper. We need to handlize all oops in
1873 // registers. We must create space for them here that is disjoint from
1874 // the windowed save area because we have no control over when we might
1875 // flush the window again and overwrite values that gc has since modified.
1876 // (The live window race)
1877 //
1878 // We always just allocate 6 word for storing down these object. This allow
1879 // us to simply record the base and use the Ireg number to decide which
1880 // slot to use. (Note that the reg number is the inbound number not the
1881 // outbound number).
1882 // We must shuffle args to match the native convention, and include var-args space.
1883
1884 // Calculate the total number of stack slots we will need.
1885
1886 // First count the abi requirement plus all of the outgoing args
1887 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1888
1889 // Now the space for the inbound oop handle area
1890
1891 int oop_handle_offset = stack_slots;
1892 stack_slots += 6*VMRegImpl::slots_per_word;
1893
1894 // Now any space we need for handlizing a klass if static method
1895
1896 int oop_temp_slot_offset = 0;
1897 int klass_slot_offset = 0;
1898 int klass_offset = -1;
1899 int lock_slot_offset = 0;
1900 bool is_static = false;
1901
1902 if (method->is_static()) {
1903 klass_slot_offset = stack_slots;
1904 stack_slots += VMRegImpl::slots_per_word;
1905 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1906 is_static = true;
1907 }
1908
1909 // Plus a lock if needed
1910
1911 if (method->is_synchronized()) {
1912 lock_slot_offset = stack_slots;
1913 stack_slots += VMRegImpl::slots_per_word;
1914 }
1915
1916 // Now a place to save return value or as a temporary for any gpr -> fpr moves
1917 stack_slots += 2;
1918
1919 // Ok The space we have allocated will look like:
1920 //
1921 //
1922 // FP-> | |
1923 // |---------------------|
1924 // | 2 slots for moves |
1925 // |---------------------|
1926 // | lock box (if sync) |
1927 // |---------------------| <- lock_slot_offset
1928 // | klass (if static) |
1929 // |---------------------| <- klass_slot_offset
1930 // | oopHandle area |
1931 // |---------------------| <- oop_handle_offset
1932 // | outbound memory |
1933 // | based arguments |
1934 // | |
1935 // |---------------------|
1936 // | vararg area |
1937 // |---------------------|
1938 // | |
1939 // SP-> | out_preserved_slots |
1940 //
1941 //
1942
1943
1944 // Now compute actual number of stack words we need rounding to make
1945 // stack properly aligned.
1946 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word);
1947
1948 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1949
1950 // Generate stack overflow check before creating frame
1951 __ generate_stack_overflow_check(stack_size);
1952
1953 // Generate a new frame for the wrapper.
1954 __ save(SP, -stack_size, SP);
1955
1956 int frame_complete = ((intptr_t)__ pc()) - start;
1957
1958 __ verify_thread();
1959
1960
1961 //
1962 // We immediately shuffle the arguments so that any vm call we have to
1963 // make from here on out (sync slow path, jvmti, etc.) we will have
1964 // captured the oops from our caller and have a valid oopMap for
1965 // them.
1966
1967 // -----------------
1968 // The Grand Shuffle
1969 //
1970 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
1971 // (derived from JavaThread* which is in L7_thread_cache) and, if static,
1972 // the class mirror instead of a receiver. This pretty much guarantees that
1973 // register layout will not match. We ignore these extra arguments during
1974 // the shuffle. The shuffle is described by the two calling convention
1975 // vectors we have in our possession. We simply walk the java vector to
1976 // get the source locations and the c vector to get the destinations.
1977 // Because we have a new window and the argument registers are completely
1978 // disjoint ( I0 -> O1, I1 -> O2, ...) we have nothing to worry about
1979 // here.
1980
1981 // This is a trick. We double the stack slots so we can claim
1982 // the oops in the caller's frame. Since we are sure to have
1983 // more args than the caller doubling is enough to make
1984 // sure we can capture all the incoming oop args from the
1985 // caller.
1986 //
1987 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1988 int c_arg = total_c_args - 1;
1989 // Record sp-based slot for receiver on stack for non-static methods
1990 int receiver_offset = -1;
1991
1992 // We move the arguments backward because the floating point registers
1993 // destination will always be to a register with a greater or equal register
1994 // number or the stack.
1995
1996 #ifdef ASSERT
1997 bool reg_destroyed[RegisterImpl::number_of_registers];
1998 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
1999 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
2000 reg_destroyed[r] = false;
2001 }
2002 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
2003 freg_destroyed[f] = false;
2004 }
2005
2006 #endif /* ASSERT */
2007
2008 for ( int i = total_in_args - 1; i >= 0 ; i--, c_arg-- ) {
2009
2010 #ifdef ASSERT
2011 if (in_regs[i].first()->is_Register()) {
2012 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "ack!");
2013 } else if (in_regs[i].first()->is_FloatRegister()) {
2014 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)], "ack!");
2015 }
2016 if (out_regs[c_arg].first()->is_Register()) {
2017 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
2018 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
2019 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)] = true;
2020 }
2021 #endif /* ASSERT */
2022
2023 switch (in_sig_bt[i]) {
2024 case T_ARRAY:
2025 case T_OBJECT:
2026 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
2027 ((i == 0) && (!is_static)),
2028 &receiver_offset);
2029 break;
2030 case T_VOID:
2031 break;
2032
2033 case T_FLOAT:
2034 float_move(masm, in_regs[i], out_regs[c_arg]);
2035 break;
2036
2037 case T_DOUBLE:
2038 assert( i + 1 < total_in_args &&
2039 in_sig_bt[i + 1] == T_VOID &&
2040 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
2041 double_move(masm, in_regs[i], out_regs[c_arg]);
2042 break;
2043
2044 case T_LONG :
2045 long_move(masm, in_regs[i], out_regs[c_arg]);
2046 break;
2047
2048 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2049
2050 default:
2051 move32_64(masm, in_regs[i], out_regs[c_arg]);
2052 }
2053 }
2054
2055 // Pre-load a static method's oop into O1. Used both by locking code and
2056 // the normal JNI call code.
2057 if (method->is_static()) {
2058 __ set_oop_constant(JNIHandles::make_local(Klass::cast(method->method_holder())->java_mirror()), O1);
2059
2060 // Now handlize the static class mirror in O1. It's known not-null.
2061 __ st_ptr(O1, SP, klass_offset + STACK_BIAS);
2062 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2063 __ add(SP, klass_offset + STACK_BIAS, O1);
2064 }
2065
2066
2067 const Register L6_handle = L6;
2068
2069 if (method->is_synchronized()) {
2070 __ mov(O1, L6_handle);
2071 }
2072
2073 // We have all of the arguments setup at this point. We MUST NOT touch any Oregs
2074 // except O6/O7. So if we must call out we must push a new frame. We immediately
2075 // push a new frame and flush the windows.
2076
2077 #ifdef _LP64
2078 intptr_t thepc = (intptr_t) __ pc();
2079 {
2080 address here = __ pc();
2081 // Call the next instruction
2082 __ call(here + 8, relocInfo::none);
2083 __ delayed()->nop();
2084 }
2085 #else
2086 intptr_t thepc = __ load_pc_address(O7, 0);
2087 #endif /* _LP64 */
2088
2089 // We use the same pc/oopMap repeatedly when we call out
2090 oop_maps->add_gc_map(thepc - start, map);
2091
2092 // O7 now has the pc loaded that we will use when we finally call to native.
2093
2094 // Save thread in L7; it crosses a bunch of VM calls below
2095 // Don't use save_thread because it smashes G2 and we merely
2096 // want to save a copy
2097 __ mov(G2_thread, L7_thread_cache);
2098
2099
2100 // If we create an inner frame once is plenty
2101 // when we create it we must also save G2_thread
2102 bool inner_frame_created = false;
2103
2104 // dtrace method entry support
2105 {
2106 SkipIfEqual skip_if(
2107 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
2108 // create inner frame
2109 __ save_frame(0);
2110 __ mov(G2_thread, L7_thread_cache);
2111 __ set_oop_constant(JNIHandles::make_local(method()), O1);
2112 __ call_VM_leaf(L7_thread_cache,
2113 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2114 G2_thread, O1);
2115 __ restore();
2116 }
2117
2118 // RedefineClasses() tracing support for obsolete method entry
2119 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) {
2120 // create inner frame
2121 __ save_frame(0);
2122 __ mov(G2_thread, L7_thread_cache);
2123 __ set_oop_constant(JNIHandles::make_local(method()), O1);
2124 __ call_VM_leaf(L7_thread_cache,
2125 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
2126 G2_thread, O1);
2127 __ restore();
2128 }
2129
2130 // We are in the jni frame unless saved_frame is true in which case
2131 // we are in one frame deeper (the "inner" frame). If we are in the
2132 // "inner" frames the args are in the Iregs and if the jni frame then
2133 // they are in the Oregs.
2134 // If we ever need to go to the VM (for locking, jvmti) then
2135 // we will always be in the "inner" frame.
2136
2137 // Lock a synchronized method
2138 int lock_offset = -1; // Set if locked
2139 if (method->is_synchronized()) {
2140 Register Roop = O1;
2141 const Register L3_box = L3;
2142
2143 create_inner_frame(masm, &inner_frame_created);
2144
2145 __ ld_ptr(I1, 0, O1);
2146 Label done;
2147
2148 lock_offset = (lock_slot_offset * VMRegImpl::stack_slot_size);
2149 __ add(FP, lock_offset+STACK_BIAS, L3_box);
2150 #ifdef ASSERT
2151 if (UseBiasedLocking) {
2152 // making the box point to itself will make it clear it went unused
2153 // but also be obviously invalid
2154 __ st_ptr(L3_box, L3_box, 0);
2155 }
2156 #endif // ASSERT
2157 //
2158 // Compiler_lock_object (Roop, Rmark, Rbox, Rscratch) -- kills Rmark, Rbox, Rscratch
2159 //
2160 __ compiler_lock_object(Roop, L1, L3_box, L2);
2161 __ br(Assembler::equal, false, Assembler::pt, done);
2162 __ delayed() -> add(FP, lock_offset+STACK_BIAS, L3_box);
2163
2164
2165 // None of the above fast optimizations worked so we have to get into the
2166 // slow case of monitor enter. Inline a special case of call_VM that
2167 // disallows any pending_exception.
2168 __ mov(Roop, O0); // Need oop in O0
2169 __ mov(L3_box, O1);
2170
2171 // Record last_Java_sp, in case the VM code releases the JVM lock.
2172
2173 __ set_last_Java_frame(FP, I7);
2174
2175 // do the call
2176 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), relocInfo::runtime_call_type);
2177 __ delayed()->mov(L7_thread_cache, O2);
2178
2179 __ restore_thread(L7_thread_cache); // restore G2_thread
2180 __ reset_last_Java_frame();
2181
2182 #ifdef ASSERT
2183 { Label L;
2184 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
2185 __ br_null(O0, false, Assembler::pt, L);
2186 __ delayed()->nop();
2187 __ stop("no pending exception allowed on exit from IR::monitorenter");
2188 __ bind(L);
2189 }
2190 #endif
2191 __ bind(done);
2192 }
2193
2194
2195 // Finally just about ready to make the JNI call
2196
2197 __ flush_windows();
2198 if (inner_frame_created) {
2199 __ restore();
2200 } else {
2201 // Store only what we need from this frame
2202 // QQQ I think that non-v9 (like we care) we don't need these saves
2203 // either as the flush traps and the current window goes too.
2204 __ st_ptr(FP, SP, FP->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
2205 __ st_ptr(I7, SP, I7->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
2206 }
2207
2208 // get JNIEnv* which is first argument to native
2209
2210 __ add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
2211
2212 // Use that pc we placed in O7 a while back as the current frame anchor
2213
2214 __ set_last_Java_frame(SP, O7);
2215
2216 // Transition from _thread_in_Java to _thread_in_native.
2217 __ set(_thread_in_native, G3_scratch);
2218 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2219
2220 // We flushed the windows ages ago now mark them as flushed
2221
2222 // mark windows as flushed
2223 __ set(JavaFrameAnchor::flushed, G3_scratch);
2224
2225 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
2226
2227 #ifdef _LP64
2228 AddressLiteral dest(method->native_function());
2229 __ relocate(relocInfo::runtime_call_type);
2230 __ jumpl_to(dest, O7, O7);
2231 #else
2232 __ call(method->native_function(), relocInfo::runtime_call_type);
2233 #endif
2234 __ delayed()->st(G3_scratch, flags);
2235
2236 __ restore_thread(L7_thread_cache); // restore G2_thread
2237
2238 // Unpack native results. For int-types, we do any needed sign-extension
2239 // and move things into I0. The return value there will survive any VM
2240 // calls for blocking or unlocking. An FP or OOP result (handle) is done
2241 // specially in the slow-path code.
2242 switch (ret_type) {
2243 case T_VOID: break; // Nothing to do!
2244 case T_FLOAT: break; // Got it where we want it (unless slow-path)
2245 case T_DOUBLE: break; // Got it where we want it (unless slow-path)
2246 // In 64 bits build result is in O0, in O0, O1 in 32bit build
2247 case T_LONG:
2248 #ifndef _LP64
2249 __ mov(O1, I1);
2250 #endif
2251 // Fall thru
2252 case T_OBJECT: // Really a handle
2253 case T_ARRAY:
2254 case T_INT:
2255 __ mov(O0, I0);
2256 break;
2257 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, I0); break; // !0 => true; 0 => false
2258 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, I0); break;
2259 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, I0); break; // cannot use and3, 0xFFFF too big as immediate value!
2260 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, I0); break;
2261 break; // Cannot de-handlize until after reclaiming jvm_lock
2262 default:
2263 ShouldNotReachHere();
2264 }
2265
2266 // must we block?
2267
2268 // Block, if necessary, before resuming in _thread_in_Java state.
2269 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2270 { Label no_block;
2271 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
2272
2273 // Switch thread to "native transition" state before reading the synchronization state.
2274 // This additional state is necessary because reading and testing the synchronization
2275 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2276 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2277 // VM thread changes sync state to synchronizing and suspends threads for GC.
2278 // Thread A is resumed to finish this native method, but doesn't block here since it
2279 // didn't see any synchronization is progress, and escapes.
2280 __ set(_thread_in_native_trans, G3_scratch);
2281 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2282 if(os::is_MP()) {
2283 if (UseMembar) {
2284 // Force this write out before the read below
2285 __ membar(Assembler::StoreLoad);
2286 } else {
2287 // Write serialization page so VM thread can do a pseudo remote membar.
2288 // We use the current thread pointer to calculate a thread specific
2289 // offset to write to within the page. This minimizes bus traffic
2290 // due to cache line collision.
2291 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
2292 }
2293 }
2294 __ load_contents(sync_state, G3_scratch);
2295 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
2296
2297 Label L;
2298 Address suspend_state(G2_thread, JavaThread::suspend_flags_offset());
2299 __ br(Assembler::notEqual, false, Assembler::pn, L);
2300 __ delayed()->ld(suspend_state, G3_scratch);
2301 __ cmp(G3_scratch, 0);
2302 __ br(Assembler::equal, false, Assembler::pt, no_block);
2303 __ delayed()->nop();
2304 __ bind(L);
2305
2306 // Block. Save any potential method result value before the operation and
2307 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2308 // lets us share the oopMap we used when we went native rather the create
2309 // a distinct one for this pc
2310 //
2311 save_native_result(masm, ret_type, stack_slots);
2312 __ call_VM_leaf(L7_thread_cache,
2313 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
2314 G2_thread);
2315
2316 // Restore any method result value
2317 restore_native_result(masm, ret_type, stack_slots);
2318 __ bind(no_block);
2319 }
2320
2321 // thread state is thread_in_native_trans. Any safepoint blocking has already
2322 // happened so we can now change state to _thread_in_Java.
2323
2324
2325 __ set(_thread_in_Java, G3_scratch);
2326 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2327
2328
2329 Label no_reguard;
2330 __ ld(G2_thread, JavaThread::stack_guard_state_offset(), G3_scratch);
2331 __ cmp(G3_scratch, JavaThread::stack_guard_yellow_disabled);
2332 __ br(Assembler::notEqual, false, Assembler::pt, no_reguard);
2333 __ delayed()->nop();
2334
2335 save_native_result(masm, ret_type, stack_slots);
2336 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2337 __ delayed()->nop();
2338
2339 __ restore_thread(L7_thread_cache); // restore G2_thread
2340 restore_native_result(masm, ret_type, stack_slots);
2341
2342 __ bind(no_reguard);
2343
2344 // Handle possible exception (will unlock if necessary)
2345
2346 // native result if any is live in freg or I0 (and I1 if long and 32bit vm)
2347
2348 // Unlock
2349 if (method->is_synchronized()) {
2350 Label done;
2351 Register I2_ex_oop = I2;
2352 const Register L3_box = L3;
2353 // Get locked oop from the handle we passed to jni
2354 __ ld_ptr(L6_handle, 0, L4);
2355 __ add(SP, lock_offset+STACK_BIAS, L3_box);
2356 // Must save pending exception around the slow-path VM call. Since it's a
2357 // leaf call, the pending exception (if any) can be kept in a register.
2358 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), I2_ex_oop);
2359 // Now unlock
2360 // (Roop, Rmark, Rbox, Rscratch)
2361 __ compiler_unlock_object(L4, L1, L3_box, L2);
2362 __ br(Assembler::equal, false, Assembler::pt, done);
2363 __ delayed()-> add(SP, lock_offset+STACK_BIAS, L3_box);
2364
2365 // save and restore any potential method result value around the unlocking
2366 // operation. Will save in I0 (or stack for FP returns).
2367 save_native_result(masm, ret_type, stack_slots);
2368
2369 // Must clear pending-exception before re-entering the VM. Since this is
2370 // a leaf call, pending-exception-oop can be safely kept in a register.
2371 __ st_ptr(G0, G2_thread, in_bytes(Thread::pending_exception_offset()));
2372
2373 // slow case of monitor enter. Inline a special case of call_VM that
2374 // disallows any pending_exception.
2375 __ mov(L3_box, O1);
2376
2377 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), relocInfo::runtime_call_type);
2378 __ delayed()->mov(L4, O0); // Need oop in O0
2379
2380 __ restore_thread(L7_thread_cache); // restore G2_thread
2381
2382 #ifdef ASSERT
2383 { Label L;
2384 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
2385 __ br_null(O0, false, Assembler::pt, L);
2386 __ delayed()->nop();
2387 __ stop("no pending exception allowed on exit from IR::monitorexit");
2388 __ bind(L);
2389 }
2390 #endif
2391 restore_native_result(masm, ret_type, stack_slots);
2392 // check_forward_pending_exception jump to forward_exception if any pending
2393 // exception is set. The forward_exception routine expects to see the
2394 // exception in pending_exception and not in a register. Kind of clumsy,
2395 // since all folks who branch to forward_exception must have tested
2396 // pending_exception first and hence have it in a register already.
2397 __ st_ptr(I2_ex_oop, G2_thread, in_bytes(Thread::pending_exception_offset()));
2398 __ bind(done);
2399 }
2400
2401 // Tell dtrace about this method exit
2402 {
2403 SkipIfEqual skip_if(
2404 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
2405 save_native_result(masm, ret_type, stack_slots);
2406 __ set_oop_constant(JNIHandles::make_local(method()), O1);
2407 __ call_VM_leaf(L7_thread_cache,
2408 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2409 G2_thread, O1);
2410 restore_native_result(masm, ret_type, stack_slots);
2411 }
2412
2413 // Clear "last Java frame" SP and PC.
2414 __ verify_thread(); // G2_thread must be correct
2415 __ reset_last_Java_frame();
2416
2417 // Unpack oop result
2418 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
2419 Label L;
2420 __ addcc(G0, I0, G0);
2421 __ brx(Assembler::notZero, true, Assembler::pt, L);
2422 __ delayed()->ld_ptr(I0, 0, I0);
2423 __ mov(G0, I0);
2424 __ bind(L);
2425 __ verify_oop(I0);
2426 }
2427
2428 // reset handle block
2429 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), L5);
2430 __ st_ptr(G0, L5, JNIHandleBlock::top_offset_in_bytes());
2431
2432 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), G3_scratch);
2433 check_forward_pending_exception(masm, G3_scratch);
2434
2435
2436 // Return
2437
2438 #ifndef _LP64
2439 if (ret_type == T_LONG) {
2440
2441 // Must leave proper result in O0,O1 and G1 (c2/tiered only)
2442 __ sllx(I0, 32, G1); // Shift bits into high G1
2443 __ srl (I1, 0, I1); // Zero extend O1 (harmless?)
2444 __ or3 (I1, G1, G1); // OR 64 bits into G1
2445 }
2446 #endif
2447
2448 __ ret();
2449 __ delayed()->restore();
2450
2451 __ flush();
2452
2453 nmethod *nm = nmethod::new_native_nmethod(method,
2454 compile_id,
2455 masm->code(),
2456 vep_offset,
2457 frame_complete,
2458 stack_slots / VMRegImpl::slots_per_word,
2459 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2460 in_ByteSize(lock_offset),
2461 oop_maps);
2462 return nm;
2463
2464 }
2465
2466 #ifdef HAVE_DTRACE_H
2467 // ---------------------------------------------------------------------------
2468 // Generate a dtrace nmethod for a given signature. The method takes arguments
2469 // in the Java compiled code convention, marshals them to the native
2470 // abi and then leaves nops at the position you would expect to call a native
2471 // function. When the probe is enabled the nops are replaced with a trap
2472 // instruction that dtrace inserts and the trace will cause a notification
2473 // to dtrace.
2474 //
2475 // The probes are only able to take primitive types and java/lang/String as
2476 // arguments. No other java types are allowed. Strings are converted to utf8
2477 // strings so that from dtrace point of view java strings are converted to C
2478 // strings. There is an arbitrary fixed limit on the total space that a method
2479 // can use for converting the strings. (256 chars per string in the signature).
2480 // So any java string larger then this is truncated.
2481
2482 static int fp_offset[ConcreteRegisterImpl::number_of_registers] = { 0 };
2483 static bool offsets_initialized = false;
2484
2485 static VMRegPair reg64_to_VMRegPair(Register r) {
2486 VMRegPair ret;
2487 if (wordSize == 8) {
2488 ret.set2(r->as_VMReg());
2489 } else {
2490 ret.set_pair(r->successor()->as_VMReg(), r->as_VMReg());
2491 }
2492 return ret;
2493 }
2494
2495
2496 nmethod *SharedRuntime::generate_dtrace_nmethod(
2497 MacroAssembler *masm, methodHandle method) {
2498
2499
2500 // generate_dtrace_nmethod is guarded by a mutex so we are sure to
2501 // be single threaded in this method.
2502 assert(AdapterHandlerLibrary_lock->owned_by_self(), "must be");
2503
2504 // Fill in the signature array, for the calling-convention call.
2505 int total_args_passed = method->size_of_parameters();
2506
2507 BasicType* in_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed);
2508 VMRegPair *in_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed);
2509
2510 // The signature we are going to use for the trap that dtrace will see
2511 // java/lang/String is converted. We drop "this" and any other object
2512 // is converted to NULL. (A one-slot java/lang/Long object reference
2513 // is converted to a two-slot long, which is why we double the allocation).
2514 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed * 2);
2515 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed * 2);
2516
2517 int i=0;
2518 int total_strings = 0;
2519 int first_arg_to_pass = 0;
2520 int total_c_args = 0;
2521
2522 // Skip the receiver as dtrace doesn't want to see it
2523 if( !method->is_static() ) {
2524 in_sig_bt[i++] = T_OBJECT;
2525 first_arg_to_pass = 1;
2526 }
2527
2528 SignatureStream ss(method->signature());
2529 for ( ; !ss.at_return_type(); ss.next()) {
2530 BasicType bt = ss.type();
2531 in_sig_bt[i++] = bt; // Collect remaining bits of signature
2532 out_sig_bt[total_c_args++] = bt;
2533 if( bt == T_OBJECT) {
2534 Symbol* s = ss.as_symbol_or_null();
2535 if (s == vmSymbols::java_lang_String()) {
2536 total_strings++;
2537 out_sig_bt[total_c_args-1] = T_ADDRESS;
2538 } else if (s == vmSymbols::java_lang_Boolean() ||
2539 s == vmSymbols::java_lang_Byte()) {
2540 out_sig_bt[total_c_args-1] = T_BYTE;
2541 } else if (s == vmSymbols::java_lang_Character() ||
2542 s == vmSymbols::java_lang_Short()) {
2543 out_sig_bt[total_c_args-1] = T_SHORT;
2544 } else if (s == vmSymbols::java_lang_Integer() ||
2545 s == vmSymbols::java_lang_Float()) {
2546 out_sig_bt[total_c_args-1] = T_INT;
2547 } else if (s == vmSymbols::java_lang_Long() ||
2548 s == vmSymbols::java_lang_Double()) {
2549 out_sig_bt[total_c_args-1] = T_LONG;
2550 out_sig_bt[total_c_args++] = T_VOID;
2551 }
2552 } else if ( bt == T_LONG || bt == T_DOUBLE ) {
2553 in_sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots
2554 // We convert double to long
2555 out_sig_bt[total_c_args-1] = T_LONG;
2556 out_sig_bt[total_c_args++] = T_VOID;
2557 } else if ( bt == T_FLOAT) {
2558 // We convert float to int
2559 out_sig_bt[total_c_args-1] = T_INT;
2560 }
2561 }
2562
2563 assert(i==total_args_passed, "validly parsed signature");
2564
2565 // Now get the compiled-Java layout as input arguments
2566 int comp_args_on_stack;
2567 comp_args_on_stack = SharedRuntime::java_calling_convention(
2568 in_sig_bt, in_regs, total_args_passed, false);
2569
2570 // We have received a description of where all the java arg are located
2571 // on entry to the wrapper. We need to convert these args to where
2572 // the a native (non-jni) function would expect them. To figure out
2573 // where they go we convert the java signature to a C signature and remove
2574 // T_VOID for any long/double we might have received.
2575
2576
2577 // Now figure out where the args must be stored and how much stack space
2578 // they require (neglecting out_preserve_stack_slots but space for storing
2579 // the 1st six register arguments). It's weird see int_stk_helper.
2580 //
2581 int out_arg_slots;
2582 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
2583
2584 // Calculate the total number of stack slots we will need.
2585
2586 // First count the abi requirement plus all of the outgoing args
2587 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2588
2589 // Plus a temp for possible converion of float/double/long register args
2590
2591 int conversion_temp = stack_slots;
2592 stack_slots += 2;
2593
2594
2595 // Now space for the string(s) we must convert
2596
2597 int string_locs = stack_slots;
2598 stack_slots += total_strings *
2599 (max_dtrace_string_size / VMRegImpl::stack_slot_size);
2600
2601 // Ok The space we have allocated will look like:
2602 //
2603 //
2604 // FP-> | |
2605 // |---------------------|
2606 // | string[n] |
2607 // |---------------------| <- string_locs[n]
2608 // | string[n-1] |
2609 // |---------------------| <- string_locs[n-1]
2610 // | ... |
2611 // | ... |
2612 // |---------------------| <- string_locs[1]
2613 // | string[0] |
2614 // |---------------------| <- string_locs[0]
2615 // | temp |
2616 // |---------------------| <- conversion_temp
2617 // | outbound memory |
2618 // | based arguments |
2619 // | |
2620 // |---------------------|
2621 // | |
2622 // SP-> | out_preserved_slots |
2623 //
2624 //
2625
2626 // Now compute actual number of stack words we need rounding to make
2627 // stack properly aligned.
2628 stack_slots = round_to(stack_slots, 4 * VMRegImpl::slots_per_word);
2629
2630 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2631
2632 intptr_t start = (intptr_t)__ pc();
2633
2634 // First thing make an ic check to see if we should even be here
2635
2636 {
2637 Label L;
2638 const Register temp_reg = G3_scratch;
2639 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
2640 __ verify_oop(O0);
2641 __ ld_ptr(O0, oopDesc::klass_offset_in_bytes(), temp_reg);
2642 __ cmp(temp_reg, G5_inline_cache_reg);
2643 __ brx(Assembler::equal, true, Assembler::pt, L);
2644 __ delayed()->nop();
2645
2646 __ jump_to(ic_miss, temp_reg);
2647 __ delayed()->nop();
2648 __ align(CodeEntryAlignment);
2649 __ bind(L);
2650 }
2651
2652 int vep_offset = ((intptr_t)__ pc()) - start;
2653
2654
2655 // The instruction at the verified entry point must be 5 bytes or longer
2656 // because it can be patched on the fly by make_non_entrant. The stack bang
2657 // instruction fits that requirement.
2658
2659 // Generate stack overflow check before creating frame
2660 __ generate_stack_overflow_check(stack_size);
2661
2662 assert(((intptr_t)__ pc() - start - vep_offset) >= 5,
2663 "valid size for make_non_entrant");
2664
2665 // Generate a new frame for the wrapper.
2666 __ save(SP, -stack_size, SP);
2667
2668 // Frame is now completed as far a size and linkage.
2669
2670 int frame_complete = ((intptr_t)__ pc()) - start;
2671
2672 #ifdef ASSERT
2673 bool reg_destroyed[RegisterImpl::number_of_registers];
2674 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
2675 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
2676 reg_destroyed[r] = false;
2677 }
2678 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
2679 freg_destroyed[f] = false;
2680 }
2681
2682 #endif /* ASSERT */
2683
2684 VMRegPair zero;
2685 const Register g0 = G0; // without this we get a compiler warning (why??)
2686 zero.set2(g0->as_VMReg());
2687
2688 int c_arg, j_arg;
2689
2690 Register conversion_off = noreg;
2691
2692 for (j_arg = first_arg_to_pass, c_arg = 0 ;
2693 j_arg < total_args_passed ; j_arg++, c_arg++ ) {
2694
2695 VMRegPair src = in_regs[j_arg];
2696 VMRegPair dst = out_regs[c_arg];
2697
2698 #ifdef ASSERT
2699 if (src.first()->is_Register()) {
2700 assert(!reg_destroyed[src.first()->as_Register()->encoding()], "ack!");
2701 } else if (src.first()->is_FloatRegister()) {
2702 assert(!freg_destroyed[src.first()->as_FloatRegister()->encoding(
2703 FloatRegisterImpl::S)], "ack!");
2704 }
2705 if (dst.first()->is_Register()) {
2706 reg_destroyed[dst.first()->as_Register()->encoding()] = true;
2707 } else if (dst.first()->is_FloatRegister()) {
2708 freg_destroyed[dst.first()->as_FloatRegister()->encoding(
2709 FloatRegisterImpl::S)] = true;
2710 }
2711 #endif /* ASSERT */
2712
2713 switch (in_sig_bt[j_arg]) {
2714 case T_ARRAY:
2715 case T_OBJECT:
2716 {
2717 if (out_sig_bt[c_arg] == T_BYTE || out_sig_bt[c_arg] == T_SHORT ||
2718 out_sig_bt[c_arg] == T_INT || out_sig_bt[c_arg] == T_LONG) {
2719 // need to unbox a one-slot value
2720 Register in_reg = L0;
2721 Register tmp = L2;
2722 if ( src.first()->is_reg() ) {
2723 in_reg = src.first()->as_Register();
2724 } else {
2725 assert(Assembler::is_simm13(reg2offset(src.first()) + STACK_BIAS),
2726 "must be");
2727 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, in_reg);
2728 }
2729 // If the final destination is an acceptable register
2730 if ( dst.first()->is_reg() ) {
2731 if ( dst.is_single_phys_reg() || out_sig_bt[c_arg] != T_LONG ) {
2732 tmp = dst.first()->as_Register();
2733 }
2734 }
2735
2736 Label skipUnbox;
2737 if ( wordSize == 4 && out_sig_bt[c_arg] == T_LONG ) {
2738 __ mov(G0, tmp->successor());
2739 }
2740 __ br_null(in_reg, true, Assembler::pn, skipUnbox);
2741 __ delayed()->mov(G0, tmp);
2742
2743 BasicType bt = out_sig_bt[c_arg];
2744 int box_offset = java_lang_boxing_object::value_offset_in_bytes(bt);
2745 switch (bt) {
2746 case T_BYTE:
2747 __ ldub(in_reg, box_offset, tmp); break;
2748 case T_SHORT:
2749 __ lduh(in_reg, box_offset, tmp); break;
2750 case T_INT:
2751 __ ld(in_reg, box_offset, tmp); break;
2752 case T_LONG:
2753 __ ld_long(in_reg, box_offset, tmp); break;
2754 default: ShouldNotReachHere();
2755 }
2756
2757 __ bind(skipUnbox);
2758 // If tmp wasn't final destination copy to final destination
2759 if (tmp == L2) {
2760 VMRegPair tmp_as_VM = reg64_to_VMRegPair(L2);
2761 if (out_sig_bt[c_arg] == T_LONG) {
2762 long_move(masm, tmp_as_VM, dst);
2763 } else {
2764 move32_64(masm, tmp_as_VM, out_regs[c_arg]);
2765 }
2766 }
2767 if (out_sig_bt[c_arg] == T_LONG) {
2768 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
2769 ++c_arg; // move over the T_VOID to keep the loop indices in sync
2770 }
2771 } else if (out_sig_bt[c_arg] == T_ADDRESS) {
2772 Register s =
2773 src.first()->is_reg() ? src.first()->as_Register() : L2;
2774 Register d =
2775 dst.first()->is_reg() ? dst.first()->as_Register() : L2;
2776
2777 // We store the oop now so that the conversion pass can reach
2778 // while in the inner frame. This will be the only store if
2779 // the oop is NULL.
2780 if (s != L2) {
2781 // src is register
2782 if (d != L2) {
2783 // dst is register
2784 __ mov(s, d);
2785 } else {
2786 assert(Assembler::is_simm13(reg2offset(dst.first()) +
2787 STACK_BIAS), "must be");
2788 __ st_ptr(s, SP, reg2offset(dst.first()) + STACK_BIAS);
2789 }
2790 } else {
2791 // src not a register
2792 assert(Assembler::is_simm13(reg2offset(src.first()) +
2793 STACK_BIAS), "must be");
2794 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, d);
2795 if (d == L2) {
2796 assert(Assembler::is_simm13(reg2offset(dst.first()) +
2797 STACK_BIAS), "must be");
2798 __ st_ptr(d, SP, reg2offset(dst.first()) + STACK_BIAS);
2799 }
2800 }
2801 } else if (out_sig_bt[c_arg] != T_VOID) {
2802 // Convert the arg to NULL
2803 if (dst.first()->is_reg()) {
2804 __ mov(G0, dst.first()->as_Register());
2805 } else {
2806 assert(Assembler::is_simm13(reg2offset(dst.first()) +
2807 STACK_BIAS), "must be");
2808 __ st_ptr(G0, SP, reg2offset(dst.first()) + STACK_BIAS);
2809 }
2810 }
2811 }
2812 break;
2813 case T_VOID:
2814 break;
2815
2816 case T_FLOAT:
2817 if (src.first()->is_stack()) {
2818 // Stack to stack/reg is simple
2819 move32_64(masm, src, dst);
2820 } else {
2821 if (dst.first()->is_reg()) {
2822 // freg -> reg
2823 int off =
2824 STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
2825 Register d = dst.first()->as_Register();
2826 if (Assembler::is_simm13(off)) {
2827 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
2828 SP, off);
2829 __ ld(SP, off, d);
2830 } else {
2831 if (conversion_off == noreg) {
2832 __ set(off, L6);
2833 conversion_off = L6;
2834 }
2835 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
2836 SP, conversion_off);
2837 __ ld(SP, conversion_off , d);
2838 }
2839 } else {
2840 // freg -> mem
2841 int off = STACK_BIAS + reg2offset(dst.first());
2842 if (Assembler::is_simm13(off)) {
2843 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
2844 SP, off);
2845 } else {
2846 if (conversion_off == noreg) {
2847 __ set(off, L6);
2848 conversion_off = L6;
2849 }
2850 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
2851 SP, conversion_off);
2852 }
2853 }
2854 }
2855 break;
2856
2857 case T_DOUBLE:
2858 assert( j_arg + 1 < total_args_passed &&
2859 in_sig_bt[j_arg + 1] == T_VOID &&
2860 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
2861 if (src.first()->is_stack()) {
2862 // Stack to stack/reg is simple
2863 long_move(masm, src, dst);
2864 } else {
2865 Register d = dst.first()->is_reg() ? dst.first()->as_Register() : L2;
2866
2867 // Destination could be an odd reg on 32bit in which case
2868 // we can't load direct to the destination.
2869
2870 if (!d->is_even() && wordSize == 4) {
2871 d = L2;
2872 }
2873 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
2874 if (Assembler::is_simm13(off)) {
2875 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
2876 SP, off);
2877 __ ld_long(SP, off, d);
2878 } else {
2879 if (conversion_off == noreg) {
2880 __ set(off, L6);
2881 conversion_off = L6;
2882 }
2883 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
2884 SP, conversion_off);
2885 __ ld_long(SP, conversion_off, d);
2886 }
2887 if (d == L2) {
2888 long_move(masm, reg64_to_VMRegPair(L2), dst);
2889 }
2890 }
2891 break;
2892
2893 case T_LONG :
2894 // 32bit can't do a split move of something like g1 -> O0, O1
2895 // so use a memory temp
2896 if (src.is_single_phys_reg() && wordSize == 4) {
2897 Register tmp = L2;
2898 if (dst.first()->is_reg() &&
2899 (wordSize == 8 || dst.first()->as_Register()->is_even())) {
2900 tmp = dst.first()->as_Register();
2901 }
2902
2903 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
2904 if (Assembler::is_simm13(off)) {
2905 __ stx(src.first()->as_Register(), SP, off);
2906 __ ld_long(SP, off, tmp);
2907 } else {
2908 if (conversion_off == noreg) {
2909 __ set(off, L6);
2910 conversion_off = L6;
2911 }
2912 __ stx(src.first()->as_Register(), SP, conversion_off);
2913 __ ld_long(SP, conversion_off, tmp);
2914 }
2915
2916 if (tmp == L2) {
2917 long_move(masm, reg64_to_VMRegPair(L2), dst);
2918 }
2919 } else {
2920 long_move(masm, src, dst);
2921 }
2922 break;
2923
2924 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2925
2926 default:
2927 move32_64(masm, src, dst);
2928 }
2929 }
2930
2931
2932 // If we have any strings we must store any register based arg to the stack
2933 // This includes any still live xmm registers too.
2934
2935 if (total_strings > 0 ) {
2936
2937 // protect all the arg registers
2938 __ save_frame(0);
2939 __ mov(G2_thread, L7_thread_cache);
2940 const Register L2_string_off = L2;
2941
2942 // Get first string offset
2943 __ set(string_locs * VMRegImpl::stack_slot_size, L2_string_off);
2944
2945 for (c_arg = 0 ; c_arg < total_c_args ; c_arg++ ) {
2946 if (out_sig_bt[c_arg] == T_ADDRESS) {
2947
2948 VMRegPair dst = out_regs[c_arg];
2949 const Register d = dst.first()->is_reg() ?
2950 dst.first()->as_Register()->after_save() : noreg;
2951
2952 // It's a string the oop and it was already copied to the out arg
2953 // position
2954 if (d != noreg) {
2955 __ mov(d, O0);
2956 } else {
2957 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
2958 "must be");
2959 __ ld_ptr(FP, reg2offset(dst.first()) + STACK_BIAS, O0);
2960 }
2961 Label skip;
2962
2963 __ br_null(O0, false, Assembler::pn, skip);
2964 __ delayed()->add(FP, L2_string_off, O1);
2965
2966 if (d != noreg) {
2967 __ mov(O1, d);
2968 } else {
2969 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
2970 "must be");
2971 __ st_ptr(O1, FP, reg2offset(dst.first()) + STACK_BIAS);
2972 }
2973
2974 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::get_utf),
2975 relocInfo::runtime_call_type);
2976 __ delayed()->add(L2_string_off, max_dtrace_string_size, L2_string_off);
2977
2978 __ bind(skip);
2979
2980 }
2981
2982 }
2983 __ mov(L7_thread_cache, G2_thread);
2984 __ restore();
2985
2986 }
2987
2988
2989 // Ok now we are done. Need to place the nop that dtrace wants in order to
2990 // patch in the trap
2991
2992 int patch_offset = ((intptr_t)__ pc()) - start;
2993
2994 __ nop();
2995
2996
2997 // Return
2998
2999 __ ret();
3000 __ delayed()->restore();
3001
3002 __ flush();
3003
3004 nmethod *nm = nmethod::new_dtrace_nmethod(
3005 method, masm->code(), vep_offset, patch_offset, frame_complete,
3006 stack_slots / VMRegImpl::slots_per_word);
3007 return nm;
3008
3009 }
3010
3011 #endif // HAVE_DTRACE_H
3012
3013 // this function returns the adjust size (in number of words) to a c2i adapter
3014 // activation for use during deoptimization
3015 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
3016 assert(callee_locals >= callee_parameters,
3017 "test and remove; got more parms than locals");
3018 if (callee_locals < callee_parameters)
3019 return 0; // No adjustment for negative locals
3020 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
3021 return round_to(diff, WordsPerLong);
3022 }
3023
3024 // "Top of Stack" slots that may be unused by the calling convention but must
3025 // otherwise be preserved.
3026 // On Intel these are not necessary and the value can be zero.
3027 // On Sparc this describes the words reserved for storing a register window
3028 // when an interrupt occurs.
3029 uint SharedRuntime::out_preserve_stack_slots() {
3030 return frame::register_save_words * VMRegImpl::slots_per_word;
3031 }
3032
3033 static void gen_new_frame(MacroAssembler* masm, bool deopt) {
3034 //
3035 // Common out the new frame generation for deopt and uncommon trap
3036 //
3037 Register G3pcs = G3_scratch; // Array of new pcs (input)
3038 Register Oreturn0 = O0;
3039 Register Oreturn1 = O1;
3040 Register O2UnrollBlock = O2;
3041 Register O3array = O3; // Array of frame sizes (input)
3042 Register O4array_size = O4; // number of frames (input)
3043 Register O7frame_size = O7; // number of frames (input)
3044
3045 __ ld_ptr(O3array, 0, O7frame_size);
3046 __ sub(G0, O7frame_size, O7frame_size);
3047 __ save(SP, O7frame_size, SP);
3048 __ ld_ptr(G3pcs, 0, I7); // load frame's new pc
3049
3050 #ifdef ASSERT
3051 // make sure that the frames are aligned properly
3052 #ifndef _LP64
3053 __ btst(wordSize*2-1, SP);
3054 __ breakpoint_trap(Assembler::notZero);
3055 #endif
3056 #endif
3057
3058 // Deopt needs to pass some extra live values from frame to frame
3059
3060 if (deopt) {
3061 __ mov(Oreturn0->after_save(), Oreturn0);
3062 __ mov(Oreturn1->after_save(), Oreturn1);
3063 }
3064
3065 __ mov(O4array_size->after_save(), O4array_size);
3066 __ sub(O4array_size, 1, O4array_size);
3067 __ mov(O3array->after_save(), O3array);
3068 __ mov(O2UnrollBlock->after_save(), O2UnrollBlock);
3069 __ add(G3pcs, wordSize, G3pcs); // point to next pc value
3070
3071 #ifdef ASSERT
3072 // trash registers to show a clear pattern in backtraces
3073 __ set(0xDEAD0000, I0);
3074 __ add(I0, 2, I1);
3075 __ add(I0, 4, I2);
3076 __ add(I0, 6, I3);
3077 __ add(I0, 8, I4);
3078 // Don't touch I5 could have valuable savedSP
3079 __ set(0xDEADBEEF, L0);
3080 __ mov(L0, L1);
3081 __ mov(L0, L2);
3082 __ mov(L0, L3);
3083 __ mov(L0, L4);
3084 __ mov(L0, L5);
3085
3086 // trash the return value as there is nothing to return yet
3087 __ set(0xDEAD0001, O7);
3088 #endif
3089
3090 __ mov(SP, O5_savedSP);
3091 }
3092
3093
3094 static void make_new_frames(MacroAssembler* masm, bool deopt) {
3095 //
3096 // loop through the UnrollBlock info and create new frames
3097 //
3098 Register G3pcs = G3_scratch;
3099 Register Oreturn0 = O0;
3100 Register Oreturn1 = O1;
3101 Register O2UnrollBlock = O2;
3102 Register O3array = O3;
3103 Register O4array_size = O4;
3104 Label loop;
3105
3106 // Before we make new frames, check to see if stack is available.
3107 // Do this after the caller's return address is on top of stack
3108 if (UseStackBanging) {
3109 // Get total frame size for interpreted frames
3110 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes(), O4);
3111 __ bang_stack_size(O4, O3, G3_scratch);
3112 }
3113
3114 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(), O4array_size);
3115 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(), G3pcs);
3116 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(), O3array);
3117
3118 // Adjust old interpreter frame to make space for new frame's extra java locals
3119 //
3120 // We capture the original sp for the transition frame only because it is needed in
3121 // order to properly calculate interpreter_sp_adjustment. Even though in real life
3122 // every interpreter frame captures a savedSP it is only needed at the transition
3123 // (fortunately). If we had to have it correct everywhere then we would need to
3124 // be told the sp_adjustment for each frame we create. If the frame size array
3125 // were to have twice the frame count entries then we could have pairs [sp_adjustment, frame_size]
3126 // for each frame we create and keep up the illusion every where.
3127 //
3128
3129 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(), O7);
3130 __ mov(SP, O5_savedSP); // remember initial sender's original sp before adjustment
3131 __ sub(SP, O7, SP);
3132
3133 #ifdef ASSERT
3134 // make sure that there is at least one entry in the array
3135 __ tst(O4array_size);
3136 __ breakpoint_trap(Assembler::zero);
3137 #endif
3138
3139 // Now push the new interpreter frames
3140 __ bind(loop);
3141
3142 // allocate a new frame, filling the registers
3143
3144 gen_new_frame(masm, deopt); // allocate an interpreter frame
3145
3146 __ tst(O4array_size);
3147 __ br(Assembler::notZero, false, Assembler::pn, loop);
3148 __ delayed()->add(O3array, wordSize, O3array);
3149 __ ld_ptr(G3pcs, 0, O7); // load final frame new pc
3150
3151 }
3152
3153 //------------------------------generate_deopt_blob----------------------------
3154 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
3155 // instead.
3156 void SharedRuntime::generate_deopt_blob() {
3157 // allocate space for the code
3158 ResourceMark rm;
3159 // setup code generation tools
3160 int pad = VerifyThread ? 512 : 0;// Extra slop space for more verify code
3161 #ifdef _LP64
3162 CodeBuffer buffer("deopt_blob", 2100+pad, 512);
3163 #else
3164 // Measured 8/7/03 at 1212 in 32bit debug build (no VerifyThread)
3165 // Measured 8/7/03 at 1396 in 32bit debug build (VerifyThread)
3166 CodeBuffer buffer("deopt_blob", 1600+pad, 512);
3167 #endif /* _LP64 */
3168 MacroAssembler* masm = new MacroAssembler(&buffer);
3169 FloatRegister Freturn0 = F0;
3170 Register Greturn1 = G1;
3171 Register Oreturn0 = O0;
3172 Register Oreturn1 = O1;
3173 Register O2UnrollBlock = O2;
3174 Register L0deopt_mode = L0;
3175 Register G4deopt_mode = G4_scratch;
3176 int frame_size_words;
3177 Address saved_Freturn0_addr(FP, -sizeof(double) + STACK_BIAS);
3178 #if !defined(_LP64) && defined(COMPILER2)
3179 Address saved_Greturn1_addr(FP, -sizeof(double) -sizeof(jlong) + STACK_BIAS);
3180 #endif
3181 Label cont;
3182
3183 OopMapSet *oop_maps = new OopMapSet();
3184
3185 //
3186 // This is the entry point for code which is returning to a de-optimized
3187 // frame.
3188 // The steps taken by this frame are as follows:
3189 // - push a dummy "register_save" and save the return values (O0, O1, F0/F1, G1)
3190 // and all potentially live registers (at a pollpoint many registers can be live).
3191 //
3192 // - call the C routine: Deoptimization::fetch_unroll_info (this function
3193 // returns information about the number and size of interpreter frames
3194 // which are equivalent to the frame which is being deoptimized)
3195 // - deallocate the unpack frame, restoring only results values. Other
3196 // volatile registers will now be captured in the vframeArray as needed.
3197 // - deallocate the deoptimization frame
3198 // - in a loop using the information returned in the previous step
3199 // push new interpreter frames (take care to propagate the return
3200 // values through each new frame pushed)
3201 // - create a dummy "unpack_frame" and save the return values (O0, O1, F0)
3202 // - call the C routine: Deoptimization::unpack_frames (this function
3203 // lays out values on the interpreter frame which was just created)
3204 // - deallocate the dummy unpack_frame
3205 // - ensure that all the return values are correctly set and then do
3206 // a return to the interpreter entry point
3207 //
3208 // Refer to the following methods for more information:
3209 // - Deoptimization::fetch_unroll_info
3210 // - Deoptimization::unpack_frames
3211
3212 OopMap* map = NULL;
3213
3214 int start = __ offset();
3215
3216 // restore G2, the trampoline destroyed it
3217 __ get_thread();
3218
3219 // On entry we have been called by the deoptimized nmethod with a call that
3220 // replaced the original call (or safepoint polling location) so the deoptimizing
3221 // pc is now in O7. Return values are still in the expected places
3222
3223 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3224 __ ba(false, cont);
3225 __ delayed()->mov(Deoptimization::Unpack_deopt, L0deopt_mode);
3226
3227 int exception_offset = __ offset() - start;
3228
3229 // restore G2, the trampoline destroyed it
3230 __ get_thread();
3231
3232 // On entry we have been jumped to by the exception handler (or exception_blob
3233 // for server). O0 contains the exception oop and O7 contains the original
3234 // exception pc. So if we push a frame here it will look to the
3235 // stack walking code (fetch_unroll_info) just like a normal call so
3236 // state will be extracted normally.
3237
3238 // save exception oop in JavaThread and fall through into the
3239 // exception_in_tls case since they are handled in same way except
3240 // for where the pending exception is kept.
3241 __ st_ptr(Oexception, G2_thread, JavaThread::exception_oop_offset());
3242
3243 //
3244 // Vanilla deoptimization with an exception pending in exception_oop
3245 //
3246 int exception_in_tls_offset = __ offset() - start;
3247
3248 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
3249 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3250
3251 // Restore G2_thread
3252 __ get_thread();
3253
3254 #ifdef ASSERT
3255 {
3256 // verify that there is really an exception oop in exception_oop
3257 Label has_exception;
3258 __ ld_ptr(G2_thread, JavaThread::exception_oop_offset(), Oexception);
3259 __ br_notnull(Oexception, false, Assembler::pt, has_exception);
3260 __ delayed()-> nop();
3261 __ stop("no exception in thread");
3262 __ bind(has_exception);
3263
3264 // verify that there is no pending exception
3265 Label no_pending_exception;
3266 Address exception_addr(G2_thread, Thread::pending_exception_offset());
3267 __ ld_ptr(exception_addr, Oexception);
3268 __ br_null(Oexception, false, Assembler::pt, no_pending_exception);
3269 __ delayed()->nop();
3270 __ stop("must not have pending exception here");
3271 __ bind(no_pending_exception);
3272 }
3273 #endif
3274
3275 __ ba(false, cont);
3276 __ delayed()->mov(Deoptimization::Unpack_exception, L0deopt_mode);;
3277
3278 //
3279 // Reexecute entry, similar to c2 uncommon trap
3280 //
3281 int reexecute_offset = __ offset() - start;
3282
3283 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
3284 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3285
3286 __ mov(Deoptimization::Unpack_reexecute, L0deopt_mode);
3287
3288 __ bind(cont);
3289
3290 __ set_last_Java_frame(SP, noreg);
3291
3292 // do the call by hand so we can get the oopmap
3293
3294 __ mov(G2_thread, L7_thread_cache);
3295 __ call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), relocInfo::runtime_call_type);
3296 __ delayed()->mov(G2_thread, O0);
3297
3298 // Set an oopmap for the call site this describes all our saved volatile registers
3299
3300 oop_maps->add_gc_map( __ offset()-start, map);
3301
3302 __ mov(L7_thread_cache, G2_thread);
3303
3304 __ reset_last_Java_frame();
3305
3306 // NOTE: we know that only O0/O1 will be reloaded by restore_result_registers
3307 // so this move will survive
3308
3309 __ mov(L0deopt_mode, G4deopt_mode);
3310
3311 __ mov(O0, O2UnrollBlock->after_save());
3312
3313 RegisterSaver::restore_result_registers(masm);
3314
3315 Label noException;
3316 __ cmp(G4deopt_mode, Deoptimization::Unpack_exception); // Was exception pending?
3317 __ br(Assembler::notEqual, false, Assembler::pt, noException);
3318 __ delayed()->nop();
3319
3320 // Move the pending exception from exception_oop to Oexception so
3321 // the pending exception will be picked up the interpreter.
3322 __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception);
3323 __ st_ptr(G0, G2_thread, in_bytes(JavaThread::exception_oop_offset()));
3324 __ bind(noException);
3325
3326 // deallocate the deoptimization frame taking care to preserve the return values
3327 __ mov(Oreturn0, Oreturn0->after_save());
3328 __ mov(Oreturn1, Oreturn1->after_save());
3329 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
3330 __ restore();
3331
3332 // Allocate new interpreter frame(s) and possible c2i adapter frame
3333
3334 make_new_frames(masm, true);
3335
3336 // push a dummy "unpack_frame" taking care of float return values and
3337 // call Deoptimization::unpack_frames to have the unpacker layout
3338 // information in the interpreter frames just created and then return
3339 // to the interpreter entry point
3340 __ save(SP, -frame_size_words*wordSize, SP);
3341 __ stf(FloatRegisterImpl::D, Freturn0, saved_Freturn0_addr);
3342 #if !defined(_LP64)
3343 #if defined(COMPILER2)
3344 // 32-bit 1-register longs return longs in G1
3345 __ stx(Greturn1, saved_Greturn1_addr);
3346 #endif
3347 __ set_last_Java_frame(SP, noreg);
3348 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, G4deopt_mode);
3349 #else
3350 // LP64 uses g4 in set_last_Java_frame
3351 __ mov(G4deopt_mode, O1);
3352 __ set_last_Java_frame(SP, G0);
3353 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O1);
3354 #endif
3355 __ reset_last_Java_frame();
3356 __ ldf(FloatRegisterImpl::D, saved_Freturn0_addr, Freturn0);
3357
3358 #if !defined(_LP64) && defined(COMPILER2)
3359 // In 32 bit, C2 returns longs in G1 so restore the saved G1 into
3360 // I0/I1 if the return value is long.
3361 Label not_long;
3362 __ cmp(O0,T_LONG);
3363 __ br(Assembler::notEqual, false, Assembler::pt, not_long);
3364 __ delayed()->nop();
3365 __ ldd(saved_Greturn1_addr,I0);
3366 __ bind(not_long);
3367 #endif
3368 __ ret();
3369 __ delayed()->restore();
3370
3371 masm->flush();
3372 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_words);
3373 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3374 }
3375
3376 #ifdef COMPILER2
3377
3378 //------------------------------generate_uncommon_trap_blob--------------------
3379 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
3380 // instead.
3381 void SharedRuntime::generate_uncommon_trap_blob() {
3382 // allocate space for the code
3383 ResourceMark rm;
3384 // setup code generation tools
3385 int pad = VerifyThread ? 512 : 0;
3386 #ifdef _LP64
3387 CodeBuffer buffer("uncommon_trap_blob", 2700+pad, 512);
3388 #else
3389 // Measured 8/7/03 at 660 in 32bit debug build (no VerifyThread)
3390 // Measured 8/7/03 at 1028 in 32bit debug build (VerifyThread)
3391 CodeBuffer buffer("uncommon_trap_blob", 2000+pad, 512);
3392 #endif
3393 MacroAssembler* masm = new MacroAssembler(&buffer);
3394 Register O2UnrollBlock = O2;
3395 Register O2klass_index = O2;
3396
3397 //
3398 // This is the entry point for all traps the compiler takes when it thinks
3399 // it cannot handle further execution of compilation code. The frame is
3400 // deoptimized in these cases and converted into interpreter frames for
3401 // execution
3402 // The steps taken by this frame are as follows:
3403 // - push a fake "unpack_frame"
3404 // - call the C routine Deoptimization::uncommon_trap (this function
3405 // packs the current compiled frame into vframe arrays and returns
3406 // information about the number and size of interpreter frames which
3407 // are equivalent to the frame which is being deoptimized)
3408 // - deallocate the "unpack_frame"
3409 // - deallocate the deoptimization frame
3410 // - in a loop using the information returned in the previous step
3411 // push interpreter frames;
3412 // - create a dummy "unpack_frame"
3413 // - call the C routine: Deoptimization::unpack_frames (this function
3414 // lays out values on the interpreter frame which was just created)
3415 // - deallocate the dummy unpack_frame
3416 // - return to the interpreter entry point
3417 //
3418 // Refer to the following methods for more information:
3419 // - Deoptimization::uncommon_trap
3420 // - Deoptimization::unpack_frame
3421
3422 // the unloaded class index is in O0 (first parameter to this blob)
3423
3424 // push a dummy "unpack_frame"
3425 // and call Deoptimization::uncommon_trap to pack the compiled frame into
3426 // vframe array and return the UnrollBlock information
3427 __ save_frame(0);
3428 __ set_last_Java_frame(SP, noreg);
3429 __ mov(I0, O2klass_index);
3430 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), G2_thread, O2klass_index);
3431 __ reset_last_Java_frame();
3432 __ mov(O0, O2UnrollBlock->after_save());
3433 __ restore();
3434
3435 // deallocate the deoptimized frame taking care to preserve the return values
3436 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
3437 __ restore();
3438
3439 // Allocate new interpreter frame(s) and possible c2i adapter frame
3440
3441 make_new_frames(masm, false);
3442
3443 // push a dummy "unpack_frame" taking care of float return values and
3444 // call Deoptimization::unpack_frames to have the unpacker layout
3445 // information in the interpreter frames just created and then return
3446 // to the interpreter entry point
3447 __ save_frame(0);
3448 __ set_last_Java_frame(SP, noreg);
3449 __ mov(Deoptimization::Unpack_uncommon_trap, O3); // indicate it is the uncommon trap case
3450 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O3);
3451 __ reset_last_Java_frame();
3452 __ ret();
3453 __ delayed()->restore();
3454
3455 masm->flush();
3456 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, NULL, __ total_frame_size_in_bytes(0)/wordSize);
3457 }
3458
3459 #endif // COMPILER2
3460
3461 //------------------------------generate_handler_blob-------------------
3462 //
3463 // Generate a special Compile2Runtime blob that saves all registers, and sets
3464 // up an OopMap.
3465 //
3466 // This blob is jumped to (via a breakpoint and the signal handler) from a
3467 // safepoint in compiled code. On entry to this blob, O7 contains the
3468 // address in the original nmethod at which we should resume normal execution.
3469 // Thus, this blob looks like a subroutine which must preserve lots of
3470 // registers and return normally. Note that O7 is never register-allocated,
3471 // so it is guaranteed to be free here.
3472 //
3473
3474 // The hardest part of what this blob must do is to save the 64-bit %o
3475 // registers in the 32-bit build. A simple 'save' turn the %o's to %i's and
3476 // an interrupt will chop off their heads. Making space in the caller's frame
3477 // first will let us save the 64-bit %o's before save'ing, but we cannot hand
3478 // the adjusted FP off to the GC stack-crawler: this will modify the caller's
3479 // SP and mess up HIS OopMaps. So we first adjust the caller's SP, then save
3480 // the 64-bit %o's, then do a save, then fixup the caller's SP (our FP).
3481 // Tricky, tricky, tricky...
3482
3483 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, bool cause_return) {
3484 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3485
3486 // allocate space for the code
3487 ResourceMark rm;
3488 // setup code generation tools
3489 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
3490 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
3491 // even larger with TraceJumps
3492 int pad = TraceJumps ? 512 : 0;
3493 CodeBuffer buffer("handler_blob", 1600 + pad, 512);
3494 MacroAssembler* masm = new MacroAssembler(&buffer);
3495 int frame_size_words;
3496 OopMapSet *oop_maps = new OopMapSet();
3497 OopMap* map = NULL;
3498
3499 int start = __ offset();
3500
3501 // If this causes a return before the processing, then do a "restore"
3502 if (cause_return) {
3503 __ restore();
3504 } else {
3505 // Make it look like we were called via the poll
3506 // so that frame constructor always sees a valid return address
3507 __ ld_ptr(G2_thread, in_bytes(JavaThread::saved_exception_pc_offset()), O7);
3508 __ sub(O7, frame::pc_return_offset, O7);
3509 }
3510
3511 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3512
3513 // setup last_Java_sp (blows G4)
3514 __ set_last_Java_frame(SP, noreg);
3515
3516 // call into the runtime to handle illegal instructions exception
3517 // Do not use call_VM_leaf, because we need to make a GC map at this call site.
3518 __ mov(G2_thread, O0);
3519 __ save_thread(L7_thread_cache);
3520 __ call(call_ptr);
3521 __ delayed()->nop();
3522
3523 // Set an oopmap for the call site.
3524 // We need this not only for callee-saved registers, but also for volatile
3525 // registers that the compiler might be keeping live across a safepoint.
3526
3527 oop_maps->add_gc_map( __ offset() - start, map);
3528
3529 __ restore_thread(L7_thread_cache);
3530 // clear last_Java_sp
3531 __ reset_last_Java_frame();
3532
3533 // Check for exceptions
3534 Label pending;
3535
3536 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1);
3537 __ tst(O1);
3538 __ brx(Assembler::notEqual, true, Assembler::pn, pending);
3539 __ delayed()->nop();
3540
3541 RegisterSaver::restore_live_registers(masm);
3542
3543 // We are back the the original state on entry and ready to go.
3544
3545 __ retl();
3546 __ delayed()->nop();
3547
3548 // Pending exception after the safepoint
3549
3550 __ bind(pending);
3551
3552 RegisterSaver::restore_live_registers(masm);
3553
3554 // We are back the the original state on entry.
3555
3556 // Tail-call forward_exception_entry, with the issuing PC in O7,
3557 // so it looks like the original nmethod called forward_exception_entry.
3558 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0);
3559 __ JMP(O0, 0);
3560 __ delayed()->nop();
3561
3562 // -------------
3563 // make sure all code is generated
3564 masm->flush();
3565
3566 // return exception blob
3567 return SafepointBlob::create(&buffer, oop_maps, frame_size_words);
3568 }
3569
3570 //
3571 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3572 //
3573 // Generate a stub that calls into vm to find out the proper destination
3574 // of a java call. All the argument registers are live at this point
3575 // but since this is generic code we don't know what they are and the caller
3576 // must do any gc of the args.
3577 //
3578 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3579 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3580
3581 // allocate space for the code
3582 ResourceMark rm;
3583 // setup code generation tools
3584 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
3585 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
3586 // even larger with TraceJumps
3587 int pad = TraceJumps ? 512 : 0;
3588 CodeBuffer buffer(name, 1600 + pad, 512);
3589 MacroAssembler* masm = new MacroAssembler(&buffer);
3590 int frame_size_words;
3591 OopMapSet *oop_maps = new OopMapSet();
3592 OopMap* map = NULL;
3593
3594 int start = __ offset();
3595
3596 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3597
3598 int frame_complete = __ offset();
3599
3600 // setup last_Java_sp (blows G4)
3601 __ set_last_Java_frame(SP, noreg);
3602
3603 // call into the runtime to handle illegal instructions exception
3604 // Do not use call_VM_leaf, because we need to make a GC map at this call site.
3605 __ mov(G2_thread, O0);
3606 __ save_thread(L7_thread_cache);
3607 __ call(destination, relocInfo::runtime_call_type);
3608 __ delayed()->nop();
3609
3610 // O0 contains the address we are going to jump to assuming no exception got installed
3611
3612 // Set an oopmap for the call site.
3613 // We need this not only for callee-saved registers, but also for volatile
3614 // registers that the compiler might be keeping live across a safepoint.
3615
3616 oop_maps->add_gc_map( __ offset() - start, map);
3617
3618 __ restore_thread(L7_thread_cache);
3619 // clear last_Java_sp
3620 __ reset_last_Java_frame();
3621
3622 // Check for exceptions
3623 Label pending;
3624
3625 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1);
3626 __ tst(O1);
3627 __ brx(Assembler::notEqual, true, Assembler::pn, pending);
3628 __ delayed()->nop();
3629
3630 // get the returned methodOop
3631
3632 __ get_vm_result(G5_method);
3633 __ stx(G5_method, SP, RegisterSaver::G5_offset()+STACK_BIAS);
3634
3635 // O0 is where we want to jump, overwrite G3 which is saved and scratch
3636
3637 __ stx(O0, SP, RegisterSaver::G3_offset()+STACK_BIAS);
3638
3639 RegisterSaver::restore_live_registers(masm);
3640
3641 // We are back the the original state on entry and ready to go.
3642
3643 __ JMP(G3, 0);
3644 __ delayed()->nop();
3645
3646 // Pending exception after the safepoint
3647
3648 __ bind(pending);
3649
3650 RegisterSaver::restore_live_registers(masm);
3651
3652 // We are back the the original state on entry.
3653
3654 // Tail-call forward_exception_entry, with the issuing PC in O7,
3655 // so it looks like the original nmethod called forward_exception_entry.
3656 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0);
3657 __ JMP(O0, 0);
3658 __ delayed()->nop();
3659
3660 // -------------
3661 // make sure all code is generated
3662 masm->flush();
3663
3664 // return the blob
3665 // frame_size_words or bytes??
3666 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true);
3667 }