Wed, 27 Aug 2008 00:21:55 -0700
5108146: Merge i486 and amd64 cpu directories
6459804: Want client (c1) compiler for x86_64 (amd64) for faster start-up
Reviewed-by: kvn
1 /*
2 * Copyright 2003-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_sharedRuntime_sparc.cpp.incl"
28 #define __ masm->
30 #ifdef COMPILER2
31 UncommonTrapBlob* SharedRuntime::_uncommon_trap_blob;
32 #endif // COMPILER2
34 DeoptimizationBlob* SharedRuntime::_deopt_blob;
35 SafepointBlob* SharedRuntime::_polling_page_safepoint_handler_blob;
36 SafepointBlob* SharedRuntime::_polling_page_return_handler_blob;
37 RuntimeStub* SharedRuntime::_wrong_method_blob;
38 RuntimeStub* SharedRuntime::_ic_miss_blob;
39 RuntimeStub* SharedRuntime::_resolve_opt_virtual_call_blob;
40 RuntimeStub* SharedRuntime::_resolve_virtual_call_blob;
41 RuntimeStub* SharedRuntime::_resolve_static_call_blob;
43 class RegisterSaver {
45 // Used for saving volatile registers. This is Gregs, Fregs, I/L/O.
46 // The Oregs are problematic. In the 32bit build the compiler can
47 // have O registers live with 64 bit quantities. A window save will
48 // cut the heads off of the registers. We have to do a very extensive
49 // stack dance to save and restore these properly.
51 // Note that the Oregs problem only exists if we block at either a polling
52 // page exception a compiled code safepoint that was not originally a call
53 // or deoptimize following one of these kinds of safepoints.
55 // Lots of registers to save. For all builds, a window save will preserve
56 // the %i and %l registers. For the 32-bit longs-in-two entries and 64-bit
57 // builds a window-save will preserve the %o registers. In the LION build
58 // we need to save the 64-bit %o registers which requires we save them
59 // before the window-save (as then they become %i registers and get their
60 // heads chopped off on interrupt). We have to save some %g registers here
61 // as well.
62 enum {
63 // This frame's save area. Includes extra space for the native call:
64 // vararg's layout space and the like. Briefly holds the caller's
65 // register save area.
66 call_args_area = frame::register_save_words_sp_offset +
67 frame::memory_parameter_word_sp_offset*wordSize,
68 // Make sure save locations are always 8 byte aligned.
69 // can't use round_to because it doesn't produce compile time constant
70 start_of_extra_save_area = ((call_args_area + 7) & ~7),
71 g1_offset = start_of_extra_save_area, // g-regs needing saving
72 g3_offset = g1_offset+8,
73 g4_offset = g3_offset+8,
74 g5_offset = g4_offset+8,
75 o0_offset = g5_offset+8,
76 o1_offset = o0_offset+8,
77 o2_offset = o1_offset+8,
78 o3_offset = o2_offset+8,
79 o4_offset = o3_offset+8,
80 o5_offset = o4_offset+8,
81 start_of_flags_save_area = o5_offset+8,
82 ccr_offset = start_of_flags_save_area,
83 fsr_offset = ccr_offset + 8,
84 d00_offset = fsr_offset+8, // Start of float save area
85 register_save_size = d00_offset+8*32
86 };
89 public:
91 static int Oexception_offset() { return o0_offset; };
92 static int G3_offset() { return g3_offset; };
93 static int G5_offset() { return g5_offset; };
94 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
95 static void restore_live_registers(MacroAssembler* masm);
97 // During deoptimization only the result register need to be restored
98 // all the other values have already been extracted.
100 static void restore_result_registers(MacroAssembler* masm);
101 };
103 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
104 // Record volatile registers as callee-save values in an OopMap so their save locations will be
105 // propagated to the caller frame's RegisterMap during StackFrameStream construction (needed for
106 // deoptimization; see compiledVFrame::create_stack_value). The caller's I, L and O registers
107 // are saved in register windows - I's and L's in the caller's frame and O's in the stub frame
108 // (as the stub's I's) when the runtime routine called by the stub creates its frame.
109 int i;
110 // Always make the frame size 16 bytr aligned.
111 int frame_size = round_to(additional_frame_words + register_save_size, 16);
112 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words
113 int frame_size_in_slots = frame_size / sizeof(jint);
114 // CodeBlob frame size is in words.
115 *total_frame_words = frame_size / wordSize;
116 // OopMap* map = new OopMap(*total_frame_words, 0);
117 OopMap* map = new OopMap(frame_size_in_slots, 0);
119 #if !defined(_LP64)
121 // Save 64-bit O registers; they will get their heads chopped off on a 'save'.
122 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
123 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
124 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
125 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
126 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
127 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
128 #endif /* _LP64 */
130 __ save(SP, -frame_size, SP);
132 #ifndef _LP64
133 // Reload the 64 bit Oregs. Although they are now Iregs we load them
134 // to Oregs here to avoid interrupts cutting off their heads
136 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
137 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
138 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
139 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
140 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
141 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
143 __ stx(O0, SP, o0_offset+STACK_BIAS);
144 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset + 4)>>2), O0->as_VMReg());
146 __ stx(O1, SP, o1_offset+STACK_BIAS);
148 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset + 4)>>2), O1->as_VMReg());
150 __ stx(O2, SP, o2_offset+STACK_BIAS);
151 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset + 4)>>2), O2->as_VMReg());
153 __ stx(O3, SP, o3_offset+STACK_BIAS);
154 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset + 4)>>2), O3->as_VMReg());
156 __ stx(O4, SP, o4_offset+STACK_BIAS);
157 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset + 4)>>2), O4->as_VMReg());
159 __ stx(O5, SP, o5_offset+STACK_BIAS);
160 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset + 4)>>2), O5->as_VMReg());
161 #endif /* _LP64 */
164 #ifdef _LP64
165 int debug_offset = 0;
166 #else
167 int debug_offset = 4;
168 #endif
169 // Save the G's
170 __ stx(G1, SP, g1_offset+STACK_BIAS);
171 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset + debug_offset)>>2), G1->as_VMReg());
173 __ stx(G3, SP, g3_offset+STACK_BIAS);
174 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset + debug_offset)>>2), G3->as_VMReg());
176 __ stx(G4, SP, g4_offset+STACK_BIAS);
177 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset + debug_offset)>>2), G4->as_VMReg());
179 __ stx(G5, SP, g5_offset+STACK_BIAS);
180 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset + debug_offset)>>2), G5->as_VMReg());
182 // This is really a waste but we'll keep things as they were for now
183 if (true) {
184 #ifndef _LP64
185 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset)>>2), O0->as_VMReg()->next());
186 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset)>>2), O1->as_VMReg()->next());
187 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset)>>2), O2->as_VMReg()->next());
188 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset)>>2), O3->as_VMReg()->next());
189 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset)>>2), O4->as_VMReg()->next());
190 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset)>>2), O5->as_VMReg()->next());
191 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset)>>2), G1->as_VMReg()->next());
192 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset)>>2), G3->as_VMReg()->next());
193 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset)>>2), G4->as_VMReg()->next());
194 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset)>>2), G5->as_VMReg()->next());
195 #endif /* _LP64 */
196 }
199 // Save the flags
200 __ rdccr( G5 );
201 __ stx(G5, SP, ccr_offset+STACK_BIAS);
202 __ stxfsr(SP, fsr_offset+STACK_BIAS);
204 // Save all the FP registers
205 int offset = d00_offset;
206 for( int i=0; i<64; i+=2 ) {
207 FloatRegister f = as_FloatRegister(i);
208 __ stf(FloatRegisterImpl::D, f, SP, offset+STACK_BIAS);
209 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), f->as_VMReg());
210 if (true) {
211 map->set_callee_saved(VMRegImpl::stack2reg((offset + sizeof(float))>>2), f->as_VMReg()->next());
212 }
213 offset += sizeof(double);
214 }
216 // And we're done.
218 return map;
219 }
222 // Pop the current frame and restore all the registers that we
223 // saved.
224 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
226 // Restore all the FP registers
227 for( int i=0; i<64; i+=2 ) {
228 __ ldf(FloatRegisterImpl::D, SP, d00_offset+i*sizeof(float)+STACK_BIAS, as_FloatRegister(i));
229 }
231 __ ldx(SP, ccr_offset+STACK_BIAS, G1);
232 __ wrccr (G1) ;
234 // Restore the G's
235 // Note that G2 (AKA GThread) must be saved and restored separately.
236 // TODO-FIXME: save and restore some of the other ASRs, viz., %asi and %gsr.
238 __ ldx(SP, g1_offset+STACK_BIAS, G1);
239 __ ldx(SP, g3_offset+STACK_BIAS, G3);
240 __ ldx(SP, g4_offset+STACK_BIAS, G4);
241 __ ldx(SP, g5_offset+STACK_BIAS, G5);
244 #if !defined(_LP64)
245 // Restore the 64-bit O's.
246 __ ldx(SP, o0_offset+STACK_BIAS, O0);
247 __ ldx(SP, o1_offset+STACK_BIAS, O1);
248 __ ldx(SP, o2_offset+STACK_BIAS, O2);
249 __ ldx(SP, o3_offset+STACK_BIAS, O3);
250 __ ldx(SP, o4_offset+STACK_BIAS, O4);
251 __ ldx(SP, o5_offset+STACK_BIAS, O5);
253 // And temporarily place them in TLS
255 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
256 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
257 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
258 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
259 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
260 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
261 #endif /* _LP64 */
263 // Restore flags
265 __ ldxfsr(SP, fsr_offset+STACK_BIAS);
267 __ restore();
269 #if !defined(_LP64)
270 // Now reload the 64bit Oregs after we've restore the window.
271 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
272 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
273 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
274 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
275 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
276 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
277 #endif /* _LP64 */
279 }
281 // Pop the current frame and restore the registers that might be holding
282 // a result.
283 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
285 #if !defined(_LP64)
286 // 32bit build returns longs in G1
287 __ ldx(SP, g1_offset+STACK_BIAS, G1);
289 // Retrieve the 64-bit O's.
290 __ ldx(SP, o0_offset+STACK_BIAS, O0);
291 __ ldx(SP, o1_offset+STACK_BIAS, O1);
292 // and save to TLS
293 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
294 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
295 #endif /* _LP64 */
297 __ ldf(FloatRegisterImpl::D, SP, d00_offset+STACK_BIAS, as_FloatRegister(0));
299 __ restore();
301 #if !defined(_LP64)
302 // Now reload the 64bit Oregs after we've restore the window.
303 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
304 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
305 #endif /* _LP64 */
307 }
309 // The java_calling_convention describes stack locations as ideal slots on
310 // a frame with no abi restrictions. Since we must observe abi restrictions
311 // (like the placement of the register window) the slots must be biased by
312 // the following value.
313 static int reg2offset(VMReg r) {
314 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
315 }
317 // ---------------------------------------------------------------------------
318 // Read the array of BasicTypes from a signature, and compute where the
319 // arguments should go. Values in the VMRegPair regs array refer to 4-byte (VMRegImpl::stack_slot_size)
320 // quantities. Values less than VMRegImpl::stack0 are registers, those above
321 // refer to 4-byte stack slots. All stack slots are based off of the window
322 // top. VMRegImpl::stack0 refers to the first slot past the 16-word window,
323 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register
324 // values 0-63 (up to RegisterImpl::number_of_registers) are the 64-bit
325 // integer registers. Values 64-95 are the (32-bit only) float registers.
326 // Each 32-bit quantity is given its own number, so the integer registers
327 // (in either 32- or 64-bit builds) use 2 numbers. For example, there is
328 // an O0-low and an O0-high. Essentially, all int register numbers are doubled.
330 // Register results are passed in O0-O5, for outgoing call arguments. To
331 // convert to incoming arguments, convert all O's to I's. The regs array
332 // refer to the low and hi 32-bit words of 64-bit registers or stack slots.
333 // If the regs[].second() field is set to VMRegImpl::Bad(), it means it's unused (a
334 // 32-bit value was passed). If both are VMRegImpl::Bad(), it means no value was
335 // passed (used as a placeholder for the other half of longs and doubles in
336 // the 64-bit build). regs[].second() is either VMRegImpl::Bad() or regs[].second() is
337 // regs[].first()+1 (regs[].first() may be misaligned in the C calling convention).
338 // Sparc never passes a value in regs[].second() but not regs[].first() (regs[].first()
339 // == VMRegImpl::Bad() && regs[].second() != VMRegImpl::Bad()) nor unrelated values in the
340 // same VMRegPair.
342 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
343 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
344 // units regardless of build.
347 // ---------------------------------------------------------------------------
348 // The compiled Java calling convention. The Java convention always passes
349 // 64-bit values in adjacent aligned locations (either registers or stack),
350 // floats in float registers and doubles in aligned float pairs. Values are
351 // packed in the registers. There is no backing varargs store for values in
352 // registers. In the 32-bit build, longs are passed in G1 and G4 (cannot be
353 // passed in I's, because longs in I's get their heads chopped off at
354 // interrupt).
355 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
356 VMRegPair *regs,
357 int total_args_passed,
358 int is_outgoing) {
359 assert(F31->as_VMReg()->is_reg(), "overlapping stack/register numbers");
361 // Convention is to pack the first 6 int/oop args into the first 6 registers
362 // (I0-I5), extras spill to the stack. Then pack the first 8 float args
363 // into F0-F7, extras spill to the stack. Then pad all register sets to
364 // align. Then put longs and doubles into the same registers as they fit,
365 // else spill to the stack.
366 const int int_reg_max = SPARC_ARGS_IN_REGS_NUM;
367 const int flt_reg_max = 8;
368 //
369 // Where 32-bit 1-reg longs start being passed
370 // In tiered we must pass on stack because c1 can't use a "pair" in a single reg.
371 // So make it look like we've filled all the G regs that c2 wants to use.
372 Register g_reg = TieredCompilation ? noreg : G1;
374 // Count int/oop and float args. See how many stack slots we'll need and
375 // where the longs & doubles will go.
376 int int_reg_cnt = 0;
377 int flt_reg_cnt = 0;
378 // int stk_reg_pairs = frame::register_save_words*(wordSize>>2);
379 // int stk_reg_pairs = SharedRuntime::out_preserve_stack_slots();
380 int stk_reg_pairs = 0;
381 for (int i = 0; i < total_args_passed; i++) {
382 switch (sig_bt[i]) {
383 case T_LONG: // LP64, longs compete with int args
384 assert(sig_bt[i+1] == T_VOID, "");
385 #ifdef _LP64
386 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
387 #endif
388 break;
389 case T_OBJECT:
390 case T_ARRAY:
391 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
392 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
393 #ifndef _LP64
394 else stk_reg_pairs++;
395 #endif
396 break;
397 case T_INT:
398 case T_SHORT:
399 case T_CHAR:
400 case T_BYTE:
401 case T_BOOLEAN:
402 if (int_reg_cnt < int_reg_max) int_reg_cnt++;
403 else stk_reg_pairs++;
404 break;
405 case T_FLOAT:
406 if (flt_reg_cnt < flt_reg_max) flt_reg_cnt++;
407 else stk_reg_pairs++;
408 break;
409 case T_DOUBLE:
410 assert(sig_bt[i+1] == T_VOID, "");
411 break;
412 case T_VOID:
413 break;
414 default:
415 ShouldNotReachHere();
416 }
417 }
419 // This is where the longs/doubles start on the stack.
420 stk_reg_pairs = (stk_reg_pairs+1) & ~1; // Round
422 int int_reg_pairs = (int_reg_cnt+1) & ~1; // 32-bit 2-reg longs only
423 int flt_reg_pairs = (flt_reg_cnt+1) & ~1;
425 // int stk_reg = frame::register_save_words*(wordSize>>2);
426 // int stk_reg = SharedRuntime::out_preserve_stack_slots();
427 int stk_reg = 0;
428 int int_reg = 0;
429 int flt_reg = 0;
431 // Now do the signature layout
432 for (int i = 0; i < total_args_passed; i++) {
433 switch (sig_bt[i]) {
434 case T_INT:
435 case T_SHORT:
436 case T_CHAR:
437 case T_BYTE:
438 case T_BOOLEAN:
439 #ifndef _LP64
440 case T_OBJECT:
441 case T_ARRAY:
442 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
443 #endif // _LP64
444 if (int_reg < int_reg_max) {
445 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
446 regs[i].set1(r->as_VMReg());
447 } else {
448 regs[i].set1(VMRegImpl::stack2reg(stk_reg++));
449 }
450 break;
452 #ifdef _LP64
453 case T_OBJECT:
454 case T_ARRAY:
455 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
456 if (int_reg < int_reg_max) {
457 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
458 regs[i].set2(r->as_VMReg());
459 } else {
460 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
461 stk_reg_pairs += 2;
462 }
463 break;
464 #endif // _LP64
466 case T_LONG:
467 assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half");
468 #ifdef _LP64
469 if (int_reg < int_reg_max) {
470 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
471 regs[i].set2(r->as_VMReg());
472 } else {
473 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
474 stk_reg_pairs += 2;
475 }
476 #else
477 #ifdef COMPILER2
478 // For 32-bit build, can't pass longs in O-regs because they become
479 // I-regs and get trashed. Use G-regs instead. G1 and G4 are almost
480 // spare and available. This convention isn't used by the Sparc ABI or
481 // anywhere else. If we're tiered then we don't use G-regs because c1
482 // can't deal with them as a "pair". (Tiered makes this code think g's are filled)
483 // G0: zero
484 // G1: 1st Long arg
485 // G2: global allocated to TLS
486 // G3: used in inline cache check
487 // G4: 2nd Long arg
488 // G5: used in inline cache check
489 // G6: used by OS
490 // G7: used by OS
492 if (g_reg == G1) {
493 regs[i].set2(G1->as_VMReg()); // This long arg in G1
494 g_reg = G4; // Where the next arg goes
495 } else if (g_reg == G4) {
496 regs[i].set2(G4->as_VMReg()); // The 2nd long arg in G4
497 g_reg = noreg; // No more longs in registers
498 } else {
499 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
500 stk_reg_pairs += 2;
501 }
502 #else // COMPILER2
503 if (int_reg_pairs + 1 < int_reg_max) {
504 if (is_outgoing) {
505 regs[i].set_pair(as_oRegister(int_reg_pairs + 1)->as_VMReg(), as_oRegister(int_reg_pairs)->as_VMReg());
506 } else {
507 regs[i].set_pair(as_iRegister(int_reg_pairs + 1)->as_VMReg(), as_iRegister(int_reg_pairs)->as_VMReg());
508 }
509 int_reg_pairs += 2;
510 } else {
511 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
512 stk_reg_pairs += 2;
513 }
514 #endif // COMPILER2
515 #endif // _LP64
516 break;
518 case T_FLOAT:
519 if (flt_reg < flt_reg_max) regs[i].set1(as_FloatRegister(flt_reg++)->as_VMReg());
520 else regs[i].set1( VMRegImpl::stack2reg(stk_reg++));
521 break;
522 case T_DOUBLE:
523 assert(sig_bt[i+1] == T_VOID, "expecting half");
524 if (flt_reg_pairs + 1 < flt_reg_max) {
525 regs[i].set2(as_FloatRegister(flt_reg_pairs)->as_VMReg());
526 flt_reg_pairs += 2;
527 } else {
528 regs[i].set2(VMRegImpl::stack2reg(stk_reg_pairs));
529 stk_reg_pairs += 2;
530 }
531 break;
532 case T_VOID: regs[i].set_bad(); break; // Halves of longs & doubles
533 default:
534 ShouldNotReachHere();
535 }
536 }
538 // retun the amount of stack space these arguments will need.
539 return stk_reg_pairs;
541 }
543 // Helper class mostly to avoid passing masm everywhere, and handle store
544 // displacement overflow logic for LP64
545 class AdapterGenerator {
546 MacroAssembler *masm;
547 #ifdef _LP64
548 Register Rdisp;
549 void set_Rdisp(Register r) { Rdisp = r; }
550 #endif // _LP64
552 void patch_callers_callsite();
553 void tag_c2i_arg(frame::Tag t, Register base, int st_off, Register scratch);
555 // base+st_off points to top of argument
556 int arg_offset(const int st_off) { return st_off + Interpreter::value_offset_in_bytes(); }
557 int next_arg_offset(const int st_off) {
558 return st_off - Interpreter::stackElementSize() + Interpreter::value_offset_in_bytes();
559 }
561 #ifdef _LP64
562 // On _LP64 argument slot values are loaded first into a register
563 // because they might not fit into displacement.
564 Register arg_slot(const int st_off);
565 Register next_arg_slot(const int st_off);
566 #else
567 int arg_slot(const int st_off) { return arg_offset(st_off); }
568 int next_arg_slot(const int st_off) { return next_arg_offset(st_off); }
569 #endif // _LP64
571 // Stores long into offset pointed to by base
572 void store_c2i_long(Register r, Register base,
573 const int st_off, bool is_stack);
574 void store_c2i_object(Register r, Register base,
575 const int st_off);
576 void store_c2i_int(Register r, Register base,
577 const int st_off);
578 void store_c2i_double(VMReg r_2,
579 VMReg r_1, Register base, const int st_off);
580 void store_c2i_float(FloatRegister f, Register base,
581 const int st_off);
583 public:
584 void gen_c2i_adapter(int total_args_passed,
585 // VMReg max_arg,
586 int comp_args_on_stack, // VMRegStackSlots
587 const BasicType *sig_bt,
588 const VMRegPair *regs,
589 Label& skip_fixup);
590 void gen_i2c_adapter(int total_args_passed,
591 // VMReg max_arg,
592 int comp_args_on_stack, // VMRegStackSlots
593 const BasicType *sig_bt,
594 const VMRegPair *regs);
596 AdapterGenerator(MacroAssembler *_masm) : masm(_masm) {}
597 };
600 // Patch the callers callsite with entry to compiled code if it exists.
601 void AdapterGenerator::patch_callers_callsite() {
602 Label L;
603 __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch);
604 __ br_null(G3_scratch, false, __ pt, L);
605 // Schedule the branch target address early.
606 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
607 // Call into the VM to patch the caller, then jump to compiled callee
608 __ save_frame(4); // Args in compiled layout; do not blow them
610 // Must save all the live Gregs the list is:
611 // G1: 1st Long arg (32bit build)
612 // G2: global allocated to TLS
613 // G3: used in inline cache check (scratch)
614 // G4: 2nd Long arg (32bit build);
615 // G5: used in inline cache check (methodOop)
617 // The longs must go to the stack by hand since in the 32 bit build they can be trashed by window ops.
619 #ifdef _LP64
620 // mov(s,d)
621 __ mov(G1, L1);
622 __ mov(G4, L4);
623 __ mov(G5_method, L5);
624 __ mov(G5_method, O0); // VM needs target method
625 __ mov(I7, O1); // VM needs caller's callsite
626 // Must be a leaf call...
627 // can be very far once the blob has been relocated
628 Address dest(O7, CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
629 __ relocate(relocInfo::runtime_call_type);
630 __ jumpl_to(dest, O7);
631 __ delayed()->mov(G2_thread, L7_thread_cache);
632 __ mov(L7_thread_cache, G2_thread);
633 __ mov(L1, G1);
634 __ mov(L4, G4);
635 __ mov(L5, G5_method);
636 #else
637 __ stx(G1, FP, -8 + STACK_BIAS);
638 __ stx(G4, FP, -16 + STACK_BIAS);
639 __ mov(G5_method, L5);
640 __ mov(G5_method, O0); // VM needs target method
641 __ mov(I7, O1); // VM needs caller's callsite
642 // Must be a leaf call...
643 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), relocInfo::runtime_call_type);
644 __ delayed()->mov(G2_thread, L7_thread_cache);
645 __ mov(L7_thread_cache, G2_thread);
646 __ ldx(FP, -8 + STACK_BIAS, G1);
647 __ ldx(FP, -16 + STACK_BIAS, G4);
648 __ mov(L5, G5_method);
649 __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
650 #endif /* _LP64 */
652 __ restore(); // Restore args
653 __ bind(L);
654 }
656 void AdapterGenerator::tag_c2i_arg(frame::Tag t, Register base, int st_off,
657 Register scratch) {
658 if (TaggedStackInterpreter) {
659 int tag_off = st_off + Interpreter::tag_offset_in_bytes();
660 #ifdef _LP64
661 Register tag_slot = Rdisp;
662 __ set(tag_off, tag_slot);
663 #else
664 int tag_slot = tag_off;
665 #endif // _LP64
666 // have to store zero because local slots can be reused (rats!)
667 if (t == frame::TagValue) {
668 __ st_ptr(G0, base, tag_slot);
669 } else if (t == frame::TagCategory2) {
670 __ st_ptr(G0, base, tag_slot);
671 int next_tag_off = st_off - Interpreter::stackElementSize() +
672 Interpreter::tag_offset_in_bytes();
673 #ifdef _LP64
674 __ set(next_tag_off, tag_slot);
675 #else
676 tag_slot = next_tag_off;
677 #endif // _LP64
678 __ st_ptr(G0, base, tag_slot);
679 } else {
680 __ mov(t, scratch);
681 __ st_ptr(scratch, base, tag_slot);
682 }
683 }
684 }
686 #ifdef _LP64
687 Register AdapterGenerator::arg_slot(const int st_off) {
688 __ set( arg_offset(st_off), Rdisp);
689 return Rdisp;
690 }
692 Register AdapterGenerator::next_arg_slot(const int st_off){
693 __ set( next_arg_offset(st_off), Rdisp);
694 return Rdisp;
695 }
696 #endif // _LP64
698 // Stores long into offset pointed to by base
699 void AdapterGenerator::store_c2i_long(Register r, Register base,
700 const int st_off, bool is_stack) {
701 #ifdef _LP64
702 // In V9, longs are given 2 64-bit slots in the interpreter, but the
703 // data is passed in only 1 slot.
704 __ stx(r, base, next_arg_slot(st_off));
705 #else
706 #ifdef COMPILER2
707 // Misaligned store of 64-bit data
708 __ stw(r, base, arg_slot(st_off)); // lo bits
709 __ srlx(r, 32, r);
710 __ stw(r, base, next_arg_slot(st_off)); // hi bits
711 #else
712 if (is_stack) {
713 // Misaligned store of 64-bit data
714 __ stw(r, base, arg_slot(st_off)); // lo bits
715 __ srlx(r, 32, r);
716 __ stw(r, base, next_arg_slot(st_off)); // hi bits
717 } else {
718 __ stw(r->successor(), base, arg_slot(st_off) ); // lo bits
719 __ stw(r , base, next_arg_slot(st_off)); // hi bits
720 }
721 #endif // COMPILER2
722 #endif // _LP64
723 tag_c2i_arg(frame::TagCategory2, base, st_off, r);
724 }
726 void AdapterGenerator::store_c2i_object(Register r, Register base,
727 const int st_off) {
728 __ st_ptr (r, base, arg_slot(st_off));
729 tag_c2i_arg(frame::TagReference, base, st_off, r);
730 }
732 void AdapterGenerator::store_c2i_int(Register r, Register base,
733 const int st_off) {
734 __ st (r, base, arg_slot(st_off));
735 tag_c2i_arg(frame::TagValue, base, st_off, r);
736 }
738 // Stores into offset pointed to by base
739 void AdapterGenerator::store_c2i_double(VMReg r_2,
740 VMReg r_1, Register base, const int st_off) {
741 #ifdef _LP64
742 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
743 // data is passed in only 1 slot.
744 __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
745 #else
746 // Need to marshal 64-bit value from misaligned Lesp loads
747 __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
748 __ stf(FloatRegisterImpl::S, r_2->as_FloatRegister(), base, arg_slot(st_off) );
749 #endif
750 tag_c2i_arg(frame::TagCategory2, base, st_off, G1_scratch);
751 }
753 void AdapterGenerator::store_c2i_float(FloatRegister f, Register base,
754 const int st_off) {
755 __ stf(FloatRegisterImpl::S, f, base, arg_slot(st_off));
756 tag_c2i_arg(frame::TagValue, base, st_off, G1_scratch);
757 }
759 void AdapterGenerator::gen_c2i_adapter(
760 int total_args_passed,
761 // VMReg max_arg,
762 int comp_args_on_stack, // VMRegStackSlots
763 const BasicType *sig_bt,
764 const VMRegPair *regs,
765 Label& skip_fixup) {
767 // Before we get into the guts of the C2I adapter, see if we should be here
768 // at all. We've come from compiled code and are attempting to jump to the
769 // interpreter, which means the caller made a static call to get here
770 // (vcalls always get a compiled target if there is one). Check for a
771 // compiled target. If there is one, we need to patch the caller's call.
772 // However we will run interpreted if we come thru here. The next pass
773 // thru the call site will run compiled. If we ran compiled here then
774 // we can (theorectically) do endless i2c->c2i->i2c transitions during
775 // deopt/uncommon trap cycles. If we always go interpreted here then
776 // we can have at most one and don't need to play any tricks to keep
777 // from endlessly growing the stack.
778 //
779 // Actually if we detected that we had an i2c->c2i transition here we
780 // ought to be able to reset the world back to the state of the interpreted
781 // call and not bother building another interpreter arg area. We don't
782 // do that at this point.
784 patch_callers_callsite();
786 __ bind(skip_fixup);
788 // Since all args are passed on the stack, total_args_passed*wordSize is the
789 // space we need. Add in varargs area needed by the interpreter. Round up
790 // to stack alignment.
791 const int arg_size = total_args_passed * Interpreter::stackElementSize();
792 const int varargs_area =
793 (frame::varargs_offset - frame::register_save_words)*wordSize;
794 const int extraspace = round_to(arg_size + varargs_area, 2*wordSize);
796 int bias = STACK_BIAS;
797 const int interp_arg_offset = frame::varargs_offset*wordSize +
798 (total_args_passed-1)*Interpreter::stackElementSize();
800 Register base = SP;
802 #ifdef _LP64
803 // In the 64bit build because of wider slots and STACKBIAS we can run
804 // out of bits in the displacement to do loads and stores. Use g3 as
805 // temporary displacement.
806 if (! __ is_simm13(extraspace)) {
807 __ set(extraspace, G3_scratch);
808 __ sub(SP, G3_scratch, SP);
809 } else {
810 __ sub(SP, extraspace, SP);
811 }
812 set_Rdisp(G3_scratch);
813 #else
814 __ sub(SP, extraspace, SP);
815 #endif // _LP64
817 // First write G1 (if used) to where ever it must go
818 for (int i=0; i<total_args_passed; i++) {
819 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize()) + bias;
820 VMReg r_1 = regs[i].first();
821 VMReg r_2 = regs[i].second();
822 if (r_1 == G1_scratch->as_VMReg()) {
823 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) {
824 store_c2i_object(G1_scratch, base, st_off);
825 } else if (sig_bt[i] == T_LONG) {
826 assert(!TieredCompilation, "should not use register args for longs");
827 store_c2i_long(G1_scratch, base, st_off, false);
828 } else {
829 store_c2i_int(G1_scratch, base, st_off);
830 }
831 }
832 }
834 // Now write the args into the outgoing interpreter space
835 for (int i=0; i<total_args_passed; i++) {
836 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize()) + bias;
837 VMReg r_1 = regs[i].first();
838 VMReg r_2 = regs[i].second();
839 if (!r_1->is_valid()) {
840 assert(!r_2->is_valid(), "");
841 continue;
842 }
843 // Skip G1 if found as we did it first in order to free it up
844 if (r_1 == G1_scratch->as_VMReg()) {
845 continue;
846 }
847 #ifdef ASSERT
848 bool G1_forced = false;
849 #endif // ASSERT
850 if (r_1->is_stack()) { // Pretend stack targets are loaded into G1
851 #ifdef _LP64
852 Register ld_off = Rdisp;
853 __ set(reg2offset(r_1) + extraspace + bias, ld_off);
854 #else
855 int ld_off = reg2offset(r_1) + extraspace + bias;
856 #ifdef ASSERT
857 G1_forced = true;
858 #endif // ASSERT
859 #endif // _LP64
860 r_1 = G1_scratch->as_VMReg();// as part of the load/store shuffle
861 if (!r_2->is_valid()) __ ld (base, ld_off, G1_scratch);
862 else __ ldx(base, ld_off, G1_scratch);
863 }
865 if (r_1->is_Register()) {
866 Register r = r_1->as_Register()->after_restore();
867 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) {
868 store_c2i_object(r, base, st_off);
869 } else if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
870 if (TieredCompilation) {
871 assert(G1_forced || sig_bt[i] != T_LONG, "should not use register args for longs");
872 }
873 store_c2i_long(r, base, st_off, r_2->is_stack());
874 } else {
875 store_c2i_int(r, base, st_off);
876 }
877 } else {
878 assert(r_1->is_FloatRegister(), "");
879 if (sig_bt[i] == T_FLOAT) {
880 store_c2i_float(r_1->as_FloatRegister(), base, st_off);
881 } else {
882 assert(sig_bt[i] == T_DOUBLE, "wrong type");
883 store_c2i_double(r_2, r_1, base, st_off);
884 }
885 }
886 }
888 #ifdef _LP64
889 // Need to reload G3_scratch, used for temporary displacements.
890 __ ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
892 // Pass O5_savedSP as an argument to the interpreter.
893 // The interpreter will restore SP to this value before returning.
894 __ set(extraspace, G1);
895 __ add(SP, G1, O5_savedSP);
896 #else
897 // Pass O5_savedSP as an argument to the interpreter.
898 // The interpreter will restore SP to this value before returning.
899 __ add(SP, extraspace, O5_savedSP);
900 #endif // _LP64
902 __ mov((frame::varargs_offset)*wordSize -
903 1*Interpreter::stackElementSize()+bias+BytesPerWord, G1);
904 // Jump to the interpreter just as if interpreter was doing it.
905 __ jmpl(G3_scratch, 0, G0);
906 // Setup Lesp for the call. Cannot actually set Lesp as the current Lesp
907 // (really L0) is in use by the compiled frame as a generic temp. However,
908 // the interpreter does not know where its args are without some kind of
909 // arg pointer being passed in. Pass it in Gargs.
910 __ delayed()->add(SP, G1, Gargs);
911 }
913 void AdapterGenerator::gen_i2c_adapter(
914 int total_args_passed,
915 // VMReg max_arg,
916 int comp_args_on_stack, // VMRegStackSlots
917 const BasicType *sig_bt,
918 const VMRegPair *regs) {
920 // Generate an I2C adapter: adjust the I-frame to make space for the C-frame
921 // layout. Lesp was saved by the calling I-frame and will be restored on
922 // return. Meanwhile, outgoing arg space is all owned by the callee
923 // C-frame, so we can mangle it at will. After adjusting the frame size,
924 // hoist register arguments and repack other args according to the compiled
925 // code convention. Finally, end in a jump to the compiled code. The entry
926 // point address is the start of the buffer.
928 // We will only enter here from an interpreted frame and never from after
929 // passing thru a c2i. Azul allowed this but we do not. If we lose the
930 // race and use a c2i we will remain interpreted for the race loser(s).
931 // This removes all sorts of headaches on the x86 side and also eliminates
932 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
934 // As you can see from the list of inputs & outputs there are not a lot
935 // of temp registers to work with: mostly G1, G3 & G4.
937 // Inputs:
938 // G2_thread - TLS
939 // G5_method - Method oop
940 // O0 - Flag telling us to restore SP from O5
941 // O4_args - Pointer to interpreter's args
942 // O5 - Caller's saved SP, to be restored if needed
943 // O6 - Current SP!
944 // O7 - Valid return address
945 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
947 // Outputs:
948 // G2_thread - TLS
949 // G1, G4 - Outgoing long args in 32-bit build
950 // O0-O5 - Outgoing args in compiled layout
951 // O6 - Adjusted or restored SP
952 // O7 - Valid return address
953 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
954 // F0-F7 - more outgoing args
957 // O4 is about to get loaded up with compiled callee's args
958 __ sub(Gargs, BytesPerWord, Gargs);
960 #ifdef ASSERT
961 {
962 // on entry OsavedSP and SP should be equal
963 Label ok;
964 __ cmp(O5_savedSP, SP);
965 __ br(Assembler::equal, false, Assembler::pt, ok);
966 __ delayed()->nop();
967 __ stop("I5_savedSP not set");
968 __ should_not_reach_here();
969 __ bind(ok);
970 }
971 #endif
973 // ON ENTRY TO THE CODE WE ARE MAKING, WE HAVE AN INTERPRETED FRAME
974 // WITH O7 HOLDING A VALID RETURN PC
975 //
976 // | |
977 // : java stack :
978 // | |
979 // +--------------+ <--- start of outgoing args
980 // | receiver | |
981 // : rest of args : |---size is java-arg-words
982 // | | |
983 // +--------------+ <--- O4_args (misaligned) and Lesp if prior is not C2I
984 // | | |
985 // : unused : |---Space for max Java stack, plus stack alignment
986 // | | |
987 // +--------------+ <--- SP + 16*wordsize
988 // | |
989 // : window :
990 // | |
991 // +--------------+ <--- SP
993 // WE REPACK THE STACK. We use the common calling convention layout as
994 // discovered by calling SharedRuntime::calling_convention. We assume it
995 // causes an arbitrary shuffle of memory, which may require some register
996 // temps to do the shuffle. We hope for (and optimize for) the case where
997 // temps are not needed. We may have to resize the stack slightly, in case
998 // we need alignment padding (32-bit interpreter can pass longs & doubles
999 // misaligned, but the compilers expect them aligned).
1000 //
1001 // | |
1002 // : java stack :
1003 // | |
1004 // +--------------+ <--- start of outgoing args
1005 // | pad, align | |
1006 // +--------------+ |
1007 // | ints, floats | |---Outgoing stack args, packed low.
1008 // +--------------+ | First few args in registers.
1009 // : doubles : |
1010 // | longs | |
1011 // +--------------+ <--- SP' + 16*wordsize
1012 // | |
1013 // : window :
1014 // | |
1015 // +--------------+ <--- SP'
1017 // ON EXIT FROM THE CODE WE ARE MAKING, WE STILL HAVE AN INTERPRETED FRAME
1018 // WITH O7 HOLDING A VALID RETURN PC - ITS JUST THAT THE ARGS ARE NOW SETUP
1019 // FOR COMPILED CODE AND THE FRAME SLIGHTLY GROWN.
1021 // Cut-out for having no stack args. Since up to 6 args are passed
1022 // in registers, we will commonly have no stack args.
1023 if (comp_args_on_stack > 0) {
1025 // Convert VMReg stack slots to words.
1026 int comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1027 // Round up to miminum stack alignment, in wordSize
1028 comp_words_on_stack = round_to(comp_words_on_stack, 2);
1029 // Now compute the distance from Lesp to SP. This calculation does not
1030 // include the space for total_args_passed because Lesp has not yet popped
1031 // the arguments.
1032 __ sub(SP, (comp_words_on_stack)*wordSize, SP);
1033 }
1035 // Will jump to the compiled code just as if compiled code was doing it.
1036 // Pre-load the register-jump target early, to schedule it better.
1037 __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_compiled_offset()), G3);
1039 // Now generate the shuffle code. Pick up all register args and move the
1040 // rest through G1_scratch.
1041 for (int i=0; i<total_args_passed; i++) {
1042 if (sig_bt[i] == T_VOID) {
1043 // Longs and doubles are passed in native word order, but misaligned
1044 // in the 32-bit build.
1045 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
1046 continue;
1047 }
1049 // Pick up 0, 1 or 2 words from Lesp+offset. Assume mis-aligned in the
1050 // 32-bit build and aligned in the 64-bit build. Look for the obvious
1051 // ldx/lddf optimizations.
1053 // Load in argument order going down.
1054 const int ld_off = (total_args_passed-i)*Interpreter::stackElementSize();
1055 #ifdef _LP64
1056 set_Rdisp(G1_scratch);
1057 #endif // _LP64
1059 VMReg r_1 = regs[i].first();
1060 VMReg r_2 = regs[i].second();
1061 if (!r_1->is_valid()) {
1062 assert(!r_2->is_valid(), "");
1063 continue;
1064 }
1065 if (r_1->is_stack()) { // Pretend stack targets are loaded into F8/F9
1066 r_1 = F8->as_VMReg(); // as part of the load/store shuffle
1067 if (r_2->is_valid()) r_2 = r_1->next();
1068 }
1069 if (r_1->is_Register()) { // Register argument
1070 Register r = r_1->as_Register()->after_restore();
1071 if (!r_2->is_valid()) {
1072 __ ld(Gargs, arg_slot(ld_off), r);
1073 } else {
1074 #ifdef _LP64
1075 // In V9, longs are given 2 64-bit slots in the interpreter, but the
1076 // data is passed in only 1 slot.
1077 Register slot = (sig_bt[i]==T_LONG) ?
1078 next_arg_slot(ld_off) : arg_slot(ld_off);
1079 __ ldx(Gargs, slot, r);
1080 #else
1081 // Need to load a 64-bit value into G1/G4, but G1/G4 is being used in the
1082 // stack shuffle. Load the first 2 longs into G1/G4 later.
1083 #endif
1084 }
1085 } else {
1086 assert(r_1->is_FloatRegister(), "");
1087 if (!r_2->is_valid()) {
1088 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_1->as_FloatRegister());
1089 } else {
1090 #ifdef _LP64
1091 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
1092 // data is passed in only 1 slot. This code also handles longs that
1093 // are passed on the stack, but need a stack-to-stack move through a
1094 // spare float register.
1095 Register slot = (sig_bt[i]==T_LONG || sig_bt[i] == T_DOUBLE) ?
1096 next_arg_slot(ld_off) : arg_slot(ld_off);
1097 __ ldf(FloatRegisterImpl::D, Gargs, slot, r_1->as_FloatRegister());
1098 #else
1099 // Need to marshal 64-bit value from misaligned Lesp loads
1100 __ ldf(FloatRegisterImpl::S, Gargs, next_arg_slot(ld_off), r_1->as_FloatRegister());
1101 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_2->as_FloatRegister());
1102 #endif
1103 }
1104 }
1105 // Was the argument really intended to be on the stack, but was loaded
1106 // into F8/F9?
1107 if (regs[i].first()->is_stack()) {
1108 assert(r_1->as_FloatRegister() == F8, "fix this code");
1109 // Convert stack slot to an SP offset
1110 int st_off = reg2offset(regs[i].first()) + STACK_BIAS;
1111 // Store down the shuffled stack word. Target address _is_ aligned.
1112 if (!r_2->is_valid()) __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), SP, st_off);
1113 else __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), SP, st_off);
1114 }
1115 }
1116 bool made_space = false;
1117 #ifndef _LP64
1118 // May need to pick up a few long args in G1/G4
1119 bool g4_crushed = false;
1120 bool g3_crushed = false;
1121 for (int i=0; i<total_args_passed; i++) {
1122 if (regs[i].first()->is_Register() && regs[i].second()->is_valid()) {
1123 // Load in argument order going down
1124 int ld_off = (total_args_passed-i)*Interpreter::stackElementSize();
1125 // Need to marshal 64-bit value from misaligned Lesp loads
1126 Register r = regs[i].first()->as_Register()->after_restore();
1127 if (r == G1 || r == G4) {
1128 assert(!g4_crushed, "ordering problem");
1129 if (r == G4){
1130 g4_crushed = true;
1131 __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits
1132 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
1133 } else {
1134 // better schedule this way
1135 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
1136 __ lduw(Gargs, arg_slot(ld_off) , G3_scratch); // Load lo bits
1137 }
1138 g3_crushed = true;
1139 __ sllx(r, 32, r);
1140 __ or3(G3_scratch, r, r);
1141 } else {
1142 assert(r->is_out(), "longs passed in two O registers");
1143 __ ld (Gargs, arg_slot(ld_off) , r->successor()); // Load lo bits
1144 __ ld (Gargs, next_arg_slot(ld_off), r); // Load hi bits
1145 }
1146 }
1147 }
1148 #endif
1150 // Jump to the compiled code just as if compiled code was doing it.
1151 //
1152 #ifndef _LP64
1153 if (g3_crushed) {
1154 // Rats load was wasted, at least it is in cache...
1155 __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_compiled_offset()), G3);
1156 }
1157 #endif /* _LP64 */
1159 // 6243940 We might end up in handle_wrong_method if
1160 // the callee is deoptimized as we race thru here. If that
1161 // happens we don't want to take a safepoint because the
1162 // caller frame will look interpreted and arguments are now
1163 // "compiled" so it is much better to make this transition
1164 // invisible to the stack walking code. Unfortunately if
1165 // we try and find the callee by normal means a safepoint
1166 // is possible. So we stash the desired callee in the thread
1167 // and the vm will find there should this case occur.
1168 Address callee_target_addr(G2_thread, 0, in_bytes(JavaThread::callee_target_offset()));
1169 __ st_ptr(G5_method, callee_target_addr);
1171 if (StressNonEntrant) {
1172 // Open a big window for deopt failure
1173 __ save_frame(0);
1174 __ mov(G0, L0);
1175 Label loop;
1176 __ bind(loop);
1177 __ sub(L0, 1, L0);
1178 __ br_null(L0, false, Assembler::pt, loop);
1179 __ delayed()->nop();
1181 __ restore();
1182 }
1185 __ jmpl(G3, 0, G0);
1186 __ delayed()->nop();
1187 }
1189 // ---------------------------------------------------------------
1190 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1191 int total_args_passed,
1192 // VMReg max_arg,
1193 int comp_args_on_stack, // VMRegStackSlots
1194 const BasicType *sig_bt,
1195 const VMRegPair *regs) {
1196 address i2c_entry = __ pc();
1198 AdapterGenerator agen(masm);
1200 agen.gen_i2c_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs);
1203 // -------------------------------------------------------------------------
1204 // Generate a C2I adapter. On entry we know G5 holds the methodOop. The
1205 // args start out packed in the compiled layout. They need to be unpacked
1206 // into the interpreter layout. This will almost always require some stack
1207 // space. We grow the current (compiled) stack, then repack the args. We
1208 // finally end in a jump to the generic interpreter entry point. On exit
1209 // from the interpreter, the interpreter will restore our SP (lest the
1210 // compiled code, which relys solely on SP and not FP, get sick).
1212 address c2i_unverified_entry = __ pc();
1213 Label skip_fixup;
1214 {
1215 #if !defined(_LP64) && defined(COMPILER2)
1216 Register R_temp = L0; // another scratch register
1217 #else
1218 Register R_temp = G1; // another scratch register
1219 #endif
1221 Address ic_miss(G3_scratch, SharedRuntime::get_ic_miss_stub());
1223 __ verify_oop(O0);
1224 __ verify_oop(G5_method);
1225 __ load_klass(O0, G3_scratch);
1226 __ verify_oop(G3_scratch);
1228 #if !defined(_LP64) && defined(COMPILER2)
1229 __ save(SP, -frame::register_save_words*wordSize, SP);
1230 __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp);
1231 __ verify_oop(R_temp);
1232 __ cmp(G3_scratch, R_temp);
1233 __ restore();
1234 #else
1235 __ ld_ptr(G5_method, compiledICHolderOopDesc::holder_klass_offset(), R_temp);
1236 __ verify_oop(R_temp);
1237 __ cmp(G3_scratch, R_temp);
1238 #endif
1240 Label ok, ok2;
1241 __ brx(Assembler::equal, false, Assembler::pt, ok);
1242 __ delayed()->ld_ptr(G5_method, compiledICHolderOopDesc::holder_method_offset(), G5_method);
1243 __ jump_to(ic_miss);
1244 __ delayed()->nop();
1246 __ bind(ok);
1247 // Method might have been compiled since the call site was patched to
1248 // interpreted if that is the case treat it as a miss so we can get
1249 // the call site corrected.
1250 __ ld_ptr(G5_method, in_bytes(methodOopDesc::code_offset()), G3_scratch);
1251 __ bind(ok2);
1252 __ br_null(G3_scratch, false, __ pt, skip_fixup);
1253 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), G3_scratch);
1254 __ jump_to(ic_miss);
1255 __ delayed()->nop();
1257 }
1259 address c2i_entry = __ pc();
1261 agen.gen_c2i_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
1263 __ flush();
1264 return new AdapterHandlerEntry(i2c_entry, c2i_entry, c2i_unverified_entry);
1266 }
1268 // Helper function for native calling conventions
1269 static VMReg int_stk_helper( int i ) {
1270 // Bias any stack based VMReg we get by ignoring the window area
1271 // but not the register parameter save area.
1272 //
1273 // This is strange for the following reasons. We'd normally expect
1274 // the calling convention to return an VMReg for a stack slot
1275 // completely ignoring any abi reserved area. C2 thinks of that
1276 // abi area as only out_preserve_stack_slots. This does not include
1277 // the area allocated by the C abi to store down integer arguments
1278 // because the java calling convention does not use it. So
1279 // since c2 assumes that there are only out_preserve_stack_slots
1280 // to bias the optoregs (which impacts VMRegs) when actually referencing any actual stack
1281 // location the c calling convention must add in this bias amount
1282 // to make up for the fact that the out_preserve_stack_slots is
1283 // insufficient for C calls. What a mess. I sure hope those 6
1284 // stack words were worth it on every java call!
1286 // Another way of cleaning this up would be for out_preserve_stack_slots
1287 // to take a parameter to say whether it was C or java calling conventions.
1288 // Then things might look a little better (but not much).
1290 int mem_parm_offset = i - SPARC_ARGS_IN_REGS_NUM;
1291 if( mem_parm_offset < 0 ) {
1292 return as_oRegister(i)->as_VMReg();
1293 } else {
1294 int actual_offset = (mem_parm_offset + frame::memory_parameter_word_sp_offset) * VMRegImpl::slots_per_word;
1295 // Now return a biased offset that will be correct when out_preserve_slots is added back in
1296 return VMRegImpl::stack2reg(actual_offset - SharedRuntime::out_preserve_stack_slots());
1297 }
1298 }
1301 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1302 VMRegPair *regs,
1303 int total_args_passed) {
1305 // Return the number of VMReg stack_slots needed for the args.
1306 // This value does not include an abi space (like register window
1307 // save area).
1309 // The native convention is V8 if !LP64
1310 // The LP64 convention is the V9 convention which is slightly more sane.
1312 // We return the amount of VMReg stack slots we need to reserve for all
1313 // the arguments NOT counting out_preserve_stack_slots. Since we always
1314 // have space for storing at least 6 registers to memory we start with that.
1315 // See int_stk_helper for a further discussion.
1316 int max_stack_slots = (frame::varargs_offset * VMRegImpl::slots_per_word) - SharedRuntime::out_preserve_stack_slots();
1318 #ifdef _LP64
1319 // V9 convention: All things "as-if" on double-wide stack slots.
1320 // Hoist any int/ptr/long's in the first 6 to int regs.
1321 // Hoist any flt/dbl's in the first 16 dbl regs.
1322 int j = 0; // Count of actual args, not HALVES
1323 for( int i=0; i<total_args_passed; i++, j++ ) {
1324 switch( sig_bt[i] ) {
1325 case T_BOOLEAN:
1326 case T_BYTE:
1327 case T_CHAR:
1328 case T_INT:
1329 case T_SHORT:
1330 regs[i].set1( int_stk_helper( j ) ); break;
1331 case T_LONG:
1332 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1333 case T_ADDRESS: // raw pointers, like current thread, for VM calls
1334 case T_ARRAY:
1335 case T_OBJECT:
1336 regs[i].set2( int_stk_helper( j ) );
1337 break;
1338 case T_FLOAT:
1339 if ( j < 16 ) {
1340 // V9ism: floats go in ODD registers
1341 regs[i].set1(as_FloatRegister(1 + (j<<1))->as_VMReg());
1342 } else {
1343 // V9ism: floats go in ODD stack slot
1344 regs[i].set1(VMRegImpl::stack2reg(1 + (j<<1)));
1345 }
1346 break;
1347 case T_DOUBLE:
1348 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1349 if ( j < 16 ) {
1350 // V9ism: doubles go in EVEN/ODD regs
1351 regs[i].set2(as_FloatRegister(j<<1)->as_VMReg());
1352 } else {
1353 // V9ism: doubles go in EVEN/ODD stack slots
1354 regs[i].set2(VMRegImpl::stack2reg(j<<1));
1355 }
1356 break;
1357 case T_VOID: regs[i].set_bad(); j--; break; // Do not count HALVES
1358 default:
1359 ShouldNotReachHere();
1360 }
1361 if (regs[i].first()->is_stack()) {
1362 int off = regs[i].first()->reg2stack();
1363 if (off > max_stack_slots) max_stack_slots = off;
1364 }
1365 if (regs[i].second()->is_stack()) {
1366 int off = regs[i].second()->reg2stack();
1367 if (off > max_stack_slots) max_stack_slots = off;
1368 }
1369 }
1371 #else // _LP64
1372 // V8 convention: first 6 things in O-regs, rest on stack.
1373 // Alignment is willy-nilly.
1374 for( int i=0; i<total_args_passed; i++ ) {
1375 switch( sig_bt[i] ) {
1376 case T_ADDRESS: // raw pointers, like current thread, for VM calls
1377 case T_ARRAY:
1378 case T_BOOLEAN:
1379 case T_BYTE:
1380 case T_CHAR:
1381 case T_FLOAT:
1382 case T_INT:
1383 case T_OBJECT:
1384 case T_SHORT:
1385 regs[i].set1( int_stk_helper( i ) );
1386 break;
1387 case T_DOUBLE:
1388 case T_LONG:
1389 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1390 regs[i].set_pair( int_stk_helper( i+1 ), int_stk_helper( i ) );
1391 break;
1392 case T_VOID: regs[i].set_bad(); break;
1393 default:
1394 ShouldNotReachHere();
1395 }
1396 if (regs[i].first()->is_stack()) {
1397 int off = regs[i].first()->reg2stack();
1398 if (off > max_stack_slots) max_stack_slots = off;
1399 }
1400 if (regs[i].second()->is_stack()) {
1401 int off = regs[i].second()->reg2stack();
1402 if (off > max_stack_slots) max_stack_slots = off;
1403 }
1404 }
1405 #endif // _LP64
1407 return round_to(max_stack_slots + 1, 2);
1409 }
1412 // ---------------------------------------------------------------------------
1413 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1414 switch (ret_type) {
1415 case T_FLOAT:
1416 __ stf(FloatRegisterImpl::S, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS);
1417 break;
1418 case T_DOUBLE:
1419 __ stf(FloatRegisterImpl::D, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS);
1420 break;
1421 }
1422 }
1424 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1425 switch (ret_type) {
1426 case T_FLOAT:
1427 __ ldf(FloatRegisterImpl::S, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS, F0);
1428 break;
1429 case T_DOUBLE:
1430 __ ldf(FloatRegisterImpl::D, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS, F0);
1431 break;
1432 }
1433 }
1435 // Check and forward and pending exception. Thread is stored in
1436 // L7_thread_cache and possibly NOT in G2_thread. Since this is a native call, there
1437 // is no exception handler. We merely pop this frame off and throw the
1438 // exception in the caller's frame.
1439 static void check_forward_pending_exception(MacroAssembler *masm, Register Rex_oop) {
1440 Label L;
1441 __ br_null(Rex_oop, false, Assembler::pt, L);
1442 __ delayed()->mov(L7_thread_cache, G2_thread); // restore in case we have exception
1443 // Since this is a native call, we *know* the proper exception handler
1444 // without calling into the VM: it's the empty function. Just pop this
1445 // frame and then jump to forward_exception_entry; O7 will contain the
1446 // native caller's return PC.
1447 Address exception_entry(G3_scratch, StubRoutines::forward_exception_entry());
1448 __ jump_to(exception_entry);
1449 __ delayed()->restore(); // Pop this frame off.
1450 __ bind(L);
1451 }
1453 // A simple move of integer like type
1454 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1455 if (src.first()->is_stack()) {
1456 if (dst.first()->is_stack()) {
1457 // stack to stack
1458 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1459 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1460 } else {
1461 // stack to reg
1462 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1463 }
1464 } else if (dst.first()->is_stack()) {
1465 // reg to stack
1466 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1467 } else {
1468 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1469 }
1470 }
1472 // On 64 bit we will store integer like items to the stack as
1473 // 64 bits items (sparc abi) even though java would only store
1474 // 32bits for a parameter. On 32bit it will simply be 32 bits
1475 // So this routine will do 32->32 on 32bit and 32->64 on 64bit
1476 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1477 if (src.first()->is_stack()) {
1478 if (dst.first()->is_stack()) {
1479 // stack to stack
1480 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1481 __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1482 } else {
1483 // stack to reg
1484 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1485 }
1486 } else if (dst.first()->is_stack()) {
1487 // reg to stack
1488 __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1489 } else {
1490 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1491 }
1492 }
1495 // An oop arg. Must pass a handle not the oop itself
1496 static void object_move(MacroAssembler* masm,
1497 OopMap* map,
1498 int oop_handle_offset,
1499 int framesize_in_slots,
1500 VMRegPair src,
1501 VMRegPair dst,
1502 bool is_receiver,
1503 int* receiver_offset) {
1505 // must pass a handle. First figure out the location we use as a handle
1507 if (src.first()->is_stack()) {
1508 // Oop is already on the stack
1509 Register rHandle = dst.first()->is_stack() ? L5 : dst.first()->as_Register();
1510 __ add(FP, reg2offset(src.first()) + STACK_BIAS, rHandle);
1511 __ ld_ptr(rHandle, 0, L4);
1512 #ifdef _LP64
1513 __ movr( Assembler::rc_z, L4, G0, rHandle );
1514 #else
1515 __ tst( L4 );
1516 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
1517 #endif
1518 if (dst.first()->is_stack()) {
1519 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
1520 }
1521 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1522 if (is_receiver) {
1523 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
1524 }
1525 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
1526 } else {
1527 // Oop is in an input register pass we must flush it to the stack
1528 const Register rOop = src.first()->as_Register();
1529 const Register rHandle = L5;
1530 int oop_slot = rOop->input_number() * VMRegImpl::slots_per_word + oop_handle_offset;
1531 int offset = oop_slot*VMRegImpl::stack_slot_size;
1532 Label skip;
1533 __ st_ptr(rOop, SP, offset + STACK_BIAS);
1534 if (is_receiver) {
1535 *receiver_offset = oop_slot * VMRegImpl::stack_slot_size;
1536 }
1537 map->set_oop(VMRegImpl::stack2reg(oop_slot));
1538 __ add(SP, offset + STACK_BIAS, rHandle);
1539 #ifdef _LP64
1540 __ movr( Assembler::rc_z, rOop, G0, rHandle );
1541 #else
1542 __ tst( rOop );
1543 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
1544 #endif
1546 if (dst.first()->is_stack()) {
1547 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
1548 } else {
1549 __ mov(rHandle, dst.first()->as_Register());
1550 }
1551 }
1552 }
1554 // A float arg may have to do float reg int reg conversion
1555 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1556 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move");
1558 if (src.first()->is_stack()) {
1559 if (dst.first()->is_stack()) {
1560 // stack to stack the easiest of the bunch
1561 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1562 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1563 } else {
1564 // stack to reg
1565 if (dst.first()->is_Register()) {
1566 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1567 } else {
1568 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1569 }
1570 }
1571 } else if (dst.first()->is_stack()) {
1572 // reg to stack
1573 if (src.first()->is_Register()) {
1574 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1575 } else {
1576 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1577 }
1578 } else {
1579 // reg to reg
1580 if (src.first()->is_Register()) {
1581 if (dst.first()->is_Register()) {
1582 // gpr -> gpr
1583 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1584 } else {
1585 // gpr -> fpr
1586 __ st(src.first()->as_Register(), FP, -4 + STACK_BIAS);
1587 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.first()->as_FloatRegister());
1588 }
1589 } else if (dst.first()->is_Register()) {
1590 // fpr -> gpr
1591 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), FP, -4 + STACK_BIAS);
1592 __ ld(FP, -4 + STACK_BIAS, dst.first()->as_Register());
1593 } else {
1594 // fpr -> fpr
1595 // In theory these overlap but the ordering is such that this is likely a nop
1596 if ( src.first() != dst.first()) {
1597 __ fmov(FloatRegisterImpl::S, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
1598 }
1599 }
1600 }
1601 }
1603 static void split_long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1604 VMRegPair src_lo(src.first());
1605 VMRegPair src_hi(src.second());
1606 VMRegPair dst_lo(dst.first());
1607 VMRegPair dst_hi(dst.second());
1608 simple_move32(masm, src_lo, dst_lo);
1609 simple_move32(masm, src_hi, dst_hi);
1610 }
1612 // A long move
1613 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1615 // Do the simple ones here else do two int moves
1616 if (src.is_single_phys_reg() ) {
1617 if (dst.is_single_phys_reg()) {
1618 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1619 } else {
1620 // split src into two separate registers
1621 // Remember hi means hi address or lsw on sparc
1622 // Move msw to lsw
1623 if (dst.second()->is_reg()) {
1624 // MSW -> MSW
1625 __ srax(src.first()->as_Register(), 32, dst.first()->as_Register());
1626 // Now LSW -> LSW
1627 // this will only move lo -> lo and ignore hi
1628 VMRegPair split(dst.second());
1629 simple_move32(masm, src, split);
1630 } else {
1631 VMRegPair split(src.first(), L4->as_VMReg());
1632 // MSW -> MSW (lo ie. first word)
1633 __ srax(src.first()->as_Register(), 32, L4);
1634 split_long_move(masm, split, dst);
1635 }
1636 }
1637 } else if (dst.is_single_phys_reg()) {
1638 if (src.is_adjacent_aligned_on_stack(2)) {
1639 __ ldx(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1640 } else {
1641 // dst is a single reg.
1642 // Remember lo is low address not msb for stack slots
1643 // and lo is the "real" register for registers
1644 // src is
1646 VMRegPair split;
1648 if (src.first()->is_reg()) {
1649 // src.lo (msw) is a reg, src.hi is stk/reg
1650 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> src.lo [the MSW is in the LSW of the reg]
1651 split.set_pair(dst.first(), src.first());
1652 } else {
1653 // msw is stack move to L5
1654 // lsw is stack move to dst.lo (real reg)
1655 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> L5
1656 split.set_pair(dst.first(), L5->as_VMReg());
1657 }
1659 // src.lo -> src.lo/L5, src.hi -> dst.lo (the real reg)
1660 // msw -> src.lo/L5, lsw -> dst.lo
1661 split_long_move(masm, src, split);
1663 // So dst now has the low order correct position the
1664 // msw half
1665 __ sllx(split.first()->as_Register(), 32, L5);
1667 const Register d = dst.first()->as_Register();
1668 __ or3(L5, d, d);
1669 }
1670 } else {
1671 // For LP64 we can probably do better.
1672 split_long_move(masm, src, dst);
1673 }
1674 }
1676 // A double move
1677 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1679 // The painful thing here is that like long_move a VMRegPair might be
1680 // 1: a single physical register
1681 // 2: two physical registers (v8)
1682 // 3: a physical reg [lo] and a stack slot [hi] (v8)
1683 // 4: two stack slots
1685 // Since src is always a java calling convention we know that the src pair
1686 // is always either all registers or all stack (and aligned?)
1688 // in a register [lo] and a stack slot [hi]
1689 if (src.first()->is_stack()) {
1690 if (dst.first()->is_stack()) {
1691 // stack to stack the easiest of the bunch
1692 // ought to be a way to do this where if alignment is ok we use ldd/std when possible
1693 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1694 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1695 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1696 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1697 } else {
1698 // stack to reg
1699 if (dst.second()->is_stack()) {
1700 // stack -> reg, stack -> stack
1701 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1702 if (dst.first()->is_Register()) {
1703 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1704 } else {
1705 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1706 }
1707 // This was missing. (very rare case)
1708 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1709 } else {
1710 // stack -> reg
1711 // Eventually optimize for alignment QQQ
1712 if (dst.first()->is_Register()) {
1713 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1714 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_Register());
1715 } else {
1716 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1717 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_FloatRegister());
1718 }
1719 }
1720 }
1721 } else if (dst.first()->is_stack()) {
1722 // reg to stack
1723 if (src.first()->is_Register()) {
1724 // Eventually optimize for alignment QQQ
1725 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1726 if (src.second()->is_stack()) {
1727 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1728 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1729 } else {
1730 __ st(src.second()->as_Register(), SP, reg2offset(dst.second()) + STACK_BIAS);
1731 }
1732 } else {
1733 // fpr to stack
1734 if (src.second()->is_stack()) {
1735 ShouldNotReachHere();
1736 } else {
1737 // Is the stack aligned?
1738 if (reg2offset(dst.first()) & 0x7) {
1739 // No do as pairs
1740 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1741 __ stf(FloatRegisterImpl::S, src.second()->as_FloatRegister(), SP, reg2offset(dst.second()) + STACK_BIAS);
1742 } else {
1743 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1744 }
1745 }
1746 }
1747 } else {
1748 // reg to reg
1749 if (src.first()->is_Register()) {
1750 if (dst.first()->is_Register()) {
1751 // gpr -> gpr
1752 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1753 __ mov(src.second()->as_Register(), dst.second()->as_Register());
1754 } else {
1755 // gpr -> fpr
1756 // ought to be able to do a single store
1757 __ stx(src.first()->as_Register(), FP, -8 + STACK_BIAS);
1758 __ stx(src.second()->as_Register(), FP, -4 + STACK_BIAS);
1759 // ought to be able to do a single load
1760 __ ldf(FloatRegisterImpl::S, FP, -8 + STACK_BIAS, dst.first()->as_FloatRegister());
1761 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.second()->as_FloatRegister());
1762 }
1763 } else if (dst.first()->is_Register()) {
1764 // fpr -> gpr
1765 // ought to be able to do a single store
1766 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), FP, -8 + STACK_BIAS);
1767 // ought to be able to do a single load
1768 // REMEMBER first() is low address not LSB
1769 __ ld(FP, -8 + STACK_BIAS, dst.first()->as_Register());
1770 if (dst.second()->is_Register()) {
1771 __ ld(FP, -4 + STACK_BIAS, dst.second()->as_Register());
1772 } else {
1773 __ ld(FP, -4 + STACK_BIAS, L4);
1774 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1775 }
1776 } else {
1777 // fpr -> fpr
1778 // In theory these overlap but the ordering is such that this is likely a nop
1779 if ( src.first() != dst.first()) {
1780 __ fmov(FloatRegisterImpl::D, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
1781 }
1782 }
1783 }
1784 }
1786 // Creates an inner frame if one hasn't already been created, and
1787 // saves a copy of the thread in L7_thread_cache
1788 static void create_inner_frame(MacroAssembler* masm, bool* already_created) {
1789 if (!*already_created) {
1790 __ save_frame(0);
1791 // Save thread in L7 (INNER FRAME); it crosses a bunch of VM calls below
1792 // Don't use save_thread because it smashes G2 and we merely want to save a
1793 // copy
1794 __ mov(G2_thread, L7_thread_cache);
1795 *already_created = true;
1796 }
1797 }
1799 // ---------------------------------------------------------------------------
1800 // Generate a native wrapper for a given method. The method takes arguments
1801 // in the Java compiled code convention, marshals them to the native
1802 // convention (handlizes oops, etc), transitions to native, makes the call,
1803 // returns to java state (possibly blocking), unhandlizes any result and
1804 // returns.
1805 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1806 methodHandle method,
1807 int total_in_args,
1808 int comp_args_on_stack, // in VMRegStackSlots
1809 BasicType *in_sig_bt,
1810 VMRegPair *in_regs,
1811 BasicType ret_type) {
1813 // Native nmethod wrappers never take possesion of the oop arguments.
1814 // So the caller will gc the arguments. The only thing we need an
1815 // oopMap for is if the call is static
1816 //
1817 // An OopMap for lock (and class if static), and one for the VM call itself
1818 OopMapSet *oop_maps = new OopMapSet();
1819 intptr_t start = (intptr_t)__ pc();
1821 // First thing make an ic check to see if we should even be here
1822 {
1823 Label L;
1824 const Register temp_reg = G3_scratch;
1825 Address ic_miss(temp_reg, SharedRuntime::get_ic_miss_stub());
1826 __ verify_oop(O0);
1827 __ load_klass(O0, temp_reg);
1828 __ cmp(temp_reg, G5_inline_cache_reg);
1829 __ brx(Assembler::equal, true, Assembler::pt, L);
1830 __ delayed()->nop();
1832 __ jump_to(ic_miss, 0);
1833 __ delayed()->nop();
1834 __ align(CodeEntryAlignment);
1835 __ bind(L);
1836 }
1838 int vep_offset = ((intptr_t)__ pc()) - start;
1840 #ifdef COMPILER1
1841 if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) {
1842 // Object.hashCode can pull the hashCode from the header word
1843 // instead of doing a full VM transition once it's been computed.
1844 // Since hashCode is usually polymorphic at call sites we can't do
1845 // this optimization at the call site without a lot of work.
1846 Label slowCase;
1847 Register receiver = O0;
1848 Register result = O0;
1849 Register header = G3_scratch;
1850 Register hash = G3_scratch; // overwrite header value with hash value
1851 Register mask = G1; // to get hash field from header
1853 // Read the header and build a mask to get its hash field. Give up if the object is not unlocked.
1854 // We depend on hash_mask being at most 32 bits and avoid the use of
1855 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
1856 // vm: see markOop.hpp.
1857 __ ld_ptr(receiver, oopDesc::mark_offset_in_bytes(), header);
1858 __ sethi(markOopDesc::hash_mask, mask);
1859 __ btst(markOopDesc::unlocked_value, header);
1860 __ br(Assembler::zero, false, Assembler::pn, slowCase);
1861 if (UseBiasedLocking) {
1862 // Check if biased and fall through to runtime if so
1863 __ delayed()->nop();
1864 __ btst(markOopDesc::biased_lock_bit_in_place, header);
1865 __ br(Assembler::notZero, false, Assembler::pn, slowCase);
1866 }
1867 __ delayed()->or3(mask, markOopDesc::hash_mask & 0x3ff, mask);
1869 // Check for a valid (non-zero) hash code and get its value.
1870 #ifdef _LP64
1871 __ srlx(header, markOopDesc::hash_shift, hash);
1872 #else
1873 __ srl(header, markOopDesc::hash_shift, hash);
1874 #endif
1875 __ andcc(hash, mask, hash);
1876 __ br(Assembler::equal, false, Assembler::pn, slowCase);
1877 __ delayed()->nop();
1879 // leaf return.
1880 __ retl();
1881 __ delayed()->mov(hash, result);
1882 __ bind(slowCase);
1883 }
1884 #endif // COMPILER1
1887 // We have received a description of where all the java arg are located
1888 // on entry to the wrapper. We need to convert these args to where
1889 // the jni function will expect them. To figure out where they go
1890 // we convert the java signature to a C signature by inserting
1891 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1893 int total_c_args = total_in_args + 1;
1894 if (method->is_static()) {
1895 total_c_args++;
1896 }
1898 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1899 VMRegPair * out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1901 int argc = 0;
1902 out_sig_bt[argc++] = T_ADDRESS;
1903 if (method->is_static()) {
1904 out_sig_bt[argc++] = T_OBJECT;
1905 }
1907 for (int i = 0; i < total_in_args ; i++ ) {
1908 out_sig_bt[argc++] = in_sig_bt[i];
1909 }
1911 // Now figure out where the args must be stored and how much stack space
1912 // they require (neglecting out_preserve_stack_slots but space for storing
1913 // the 1st six register arguments). It's weird see int_stk_helper.
1914 //
1915 int out_arg_slots;
1916 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1918 // Compute framesize for the wrapper. We need to handlize all oops in
1919 // registers. We must create space for them here that is disjoint from
1920 // the windowed save area because we have no control over when we might
1921 // flush the window again and overwrite values that gc has since modified.
1922 // (The live window race)
1923 //
1924 // We always just allocate 6 word for storing down these object. This allow
1925 // us to simply record the base and use the Ireg number to decide which
1926 // slot to use. (Note that the reg number is the inbound number not the
1927 // outbound number).
1928 // We must shuffle args to match the native convention, and include var-args space.
1930 // Calculate the total number of stack slots we will need.
1932 // First count the abi requirement plus all of the outgoing args
1933 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1935 // Now the space for the inbound oop handle area
1937 int oop_handle_offset = stack_slots;
1938 stack_slots += 6*VMRegImpl::slots_per_word;
1940 // Now any space we need for handlizing a klass if static method
1942 int oop_temp_slot_offset = 0;
1943 int klass_slot_offset = 0;
1944 int klass_offset = -1;
1945 int lock_slot_offset = 0;
1946 bool is_static = false;
1948 if (method->is_static()) {
1949 klass_slot_offset = stack_slots;
1950 stack_slots += VMRegImpl::slots_per_word;
1951 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1952 is_static = true;
1953 }
1955 // Plus a lock if needed
1957 if (method->is_synchronized()) {
1958 lock_slot_offset = stack_slots;
1959 stack_slots += VMRegImpl::slots_per_word;
1960 }
1962 // Now a place to save return value or as a temporary for any gpr -> fpr moves
1963 stack_slots += 2;
1965 // Ok The space we have allocated will look like:
1966 //
1967 //
1968 // FP-> | |
1969 // |---------------------|
1970 // | 2 slots for moves |
1971 // |---------------------|
1972 // | lock box (if sync) |
1973 // |---------------------| <- lock_slot_offset
1974 // | klass (if static) |
1975 // |---------------------| <- klass_slot_offset
1976 // | oopHandle area |
1977 // |---------------------| <- oop_handle_offset
1978 // | outbound memory |
1979 // | based arguments |
1980 // | |
1981 // |---------------------|
1982 // | vararg area |
1983 // |---------------------|
1984 // | |
1985 // SP-> | out_preserved_slots |
1986 //
1987 //
1990 // Now compute actual number of stack words we need rounding to make
1991 // stack properly aligned.
1992 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word);
1994 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1996 // Generate stack overflow check before creating frame
1997 __ generate_stack_overflow_check(stack_size);
1999 // Generate a new frame for the wrapper.
2000 __ save(SP, -stack_size, SP);
2002 int frame_complete = ((intptr_t)__ pc()) - start;
2004 __ verify_thread();
2007 //
2008 // We immediately shuffle the arguments so that any vm call we have to
2009 // make from here on out (sync slow path, jvmti, etc.) we will have
2010 // captured the oops from our caller and have a valid oopMap for
2011 // them.
2013 // -----------------
2014 // The Grand Shuffle
2015 //
2016 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2017 // (derived from JavaThread* which is in L7_thread_cache) and, if static,
2018 // the class mirror instead of a receiver. This pretty much guarantees that
2019 // register layout will not match. We ignore these extra arguments during
2020 // the shuffle. The shuffle is described by the two calling convention
2021 // vectors we have in our possession. We simply walk the java vector to
2022 // get the source locations and the c vector to get the destinations.
2023 // Because we have a new window and the argument registers are completely
2024 // disjoint ( I0 -> O1, I1 -> O2, ...) we have nothing to worry about
2025 // here.
2027 // This is a trick. We double the stack slots so we can claim
2028 // the oops in the caller's frame. Since we are sure to have
2029 // more args than the caller doubling is enough to make
2030 // sure we can capture all the incoming oop args from the
2031 // caller.
2032 //
2033 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2034 int c_arg = total_c_args - 1;
2035 // Record sp-based slot for receiver on stack for non-static methods
2036 int receiver_offset = -1;
2038 // We move the arguments backward because the floating point registers
2039 // destination will always be to a register with a greater or equal register
2040 // number or the stack.
2042 #ifdef ASSERT
2043 bool reg_destroyed[RegisterImpl::number_of_registers];
2044 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
2045 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
2046 reg_destroyed[r] = false;
2047 }
2048 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
2049 freg_destroyed[f] = false;
2050 }
2052 #endif /* ASSERT */
2054 for ( int i = total_in_args - 1; i >= 0 ; i--, c_arg-- ) {
2056 #ifdef ASSERT
2057 if (in_regs[i].first()->is_Register()) {
2058 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "ack!");
2059 } else if (in_regs[i].first()->is_FloatRegister()) {
2060 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)], "ack!");
2061 }
2062 if (out_regs[c_arg].first()->is_Register()) {
2063 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
2064 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
2065 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)] = true;
2066 }
2067 #endif /* ASSERT */
2069 switch (in_sig_bt[i]) {
2070 case T_ARRAY:
2071 case T_OBJECT:
2072 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
2073 ((i == 0) && (!is_static)),
2074 &receiver_offset);
2075 break;
2076 case T_VOID:
2077 break;
2079 case T_FLOAT:
2080 float_move(masm, in_regs[i], out_regs[c_arg]);
2081 break;
2083 case T_DOUBLE:
2084 assert( i + 1 < total_in_args &&
2085 in_sig_bt[i + 1] == T_VOID &&
2086 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
2087 double_move(masm, in_regs[i], out_regs[c_arg]);
2088 break;
2090 case T_LONG :
2091 long_move(masm, in_regs[i], out_regs[c_arg]);
2092 break;
2094 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2096 default:
2097 move32_64(masm, in_regs[i], out_regs[c_arg]);
2098 }
2099 }
2101 // Pre-load a static method's oop into O1. Used both by locking code and
2102 // the normal JNI call code.
2103 if (method->is_static()) {
2104 __ set_oop_constant(JNIHandles::make_local(Klass::cast(method->method_holder())->java_mirror()), O1);
2106 // Now handlize the static class mirror in O1. It's known not-null.
2107 __ st_ptr(O1, SP, klass_offset + STACK_BIAS);
2108 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2109 __ add(SP, klass_offset + STACK_BIAS, O1);
2110 }
2113 const Register L6_handle = L6;
2115 if (method->is_synchronized()) {
2116 __ mov(O1, L6_handle);
2117 }
2119 // We have all of the arguments setup at this point. We MUST NOT touch any Oregs
2120 // except O6/O7. So if we must call out we must push a new frame. We immediately
2121 // push a new frame and flush the windows.
2123 #ifdef _LP64
2124 intptr_t thepc = (intptr_t) __ pc();
2125 {
2126 address here = __ pc();
2127 // Call the next instruction
2128 __ call(here + 8, relocInfo::none);
2129 __ delayed()->nop();
2130 }
2131 #else
2132 intptr_t thepc = __ load_pc_address(O7, 0);
2133 #endif /* _LP64 */
2135 // We use the same pc/oopMap repeatedly when we call out
2136 oop_maps->add_gc_map(thepc - start, map);
2138 // O7 now has the pc loaded that we will use when we finally call to native.
2140 // Save thread in L7; it crosses a bunch of VM calls below
2141 // Don't use save_thread because it smashes G2 and we merely
2142 // want to save a copy
2143 __ mov(G2_thread, L7_thread_cache);
2146 // If we create an inner frame once is plenty
2147 // when we create it we must also save G2_thread
2148 bool inner_frame_created = false;
2150 // dtrace method entry support
2151 {
2152 SkipIfEqual skip_if(
2153 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
2154 // create inner frame
2155 __ save_frame(0);
2156 __ mov(G2_thread, L7_thread_cache);
2157 __ set_oop_constant(JNIHandles::make_local(method()), O1);
2158 __ call_VM_leaf(L7_thread_cache,
2159 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2160 G2_thread, O1);
2161 __ restore();
2162 }
2164 // We are in the jni frame unless saved_frame is true in which case
2165 // we are in one frame deeper (the "inner" frame). If we are in the
2166 // "inner" frames the args are in the Iregs and if the jni frame then
2167 // they are in the Oregs.
2168 // If we ever need to go to the VM (for locking, jvmti) then
2169 // we will always be in the "inner" frame.
2171 // Lock a synchronized method
2172 int lock_offset = -1; // Set if locked
2173 if (method->is_synchronized()) {
2174 Register Roop = O1;
2175 const Register L3_box = L3;
2177 create_inner_frame(masm, &inner_frame_created);
2179 __ ld_ptr(I1, 0, O1);
2180 Label done;
2182 lock_offset = (lock_slot_offset * VMRegImpl::stack_slot_size);
2183 __ add(FP, lock_offset+STACK_BIAS, L3_box);
2184 #ifdef ASSERT
2185 if (UseBiasedLocking) {
2186 // making the box point to itself will make it clear it went unused
2187 // but also be obviously invalid
2188 __ st_ptr(L3_box, L3_box, 0);
2189 }
2190 #endif // ASSERT
2191 //
2192 // Compiler_lock_object (Roop, Rmark, Rbox, Rscratch) -- kills Rmark, Rbox, Rscratch
2193 //
2194 __ compiler_lock_object(Roop, L1, L3_box, L2);
2195 __ br(Assembler::equal, false, Assembler::pt, done);
2196 __ delayed() -> add(FP, lock_offset+STACK_BIAS, L3_box);
2199 // None of the above fast optimizations worked so we have to get into the
2200 // slow case of monitor enter. Inline a special case of call_VM that
2201 // disallows any pending_exception.
2202 __ mov(Roop, O0); // Need oop in O0
2203 __ mov(L3_box, O1);
2205 // Record last_Java_sp, in case the VM code releases the JVM lock.
2207 __ set_last_Java_frame(FP, I7);
2209 // do the call
2210 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), relocInfo::runtime_call_type);
2211 __ delayed()->mov(L7_thread_cache, O2);
2213 __ restore_thread(L7_thread_cache); // restore G2_thread
2214 __ reset_last_Java_frame();
2216 #ifdef ASSERT
2217 { Label L;
2218 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
2219 __ br_null(O0, false, Assembler::pt, L);
2220 __ delayed()->nop();
2221 __ stop("no pending exception allowed on exit from IR::monitorenter");
2222 __ bind(L);
2223 }
2224 #endif
2225 __ bind(done);
2226 }
2229 // Finally just about ready to make the JNI call
2231 __ flush_windows();
2232 if (inner_frame_created) {
2233 __ restore();
2234 } else {
2235 // Store only what we need from this frame
2236 // QQQ I think that non-v9 (like we care) we don't need these saves
2237 // either as the flush traps and the current window goes too.
2238 __ st_ptr(FP, SP, FP->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
2239 __ st_ptr(I7, SP, I7->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
2240 }
2242 // get JNIEnv* which is first argument to native
2244 __ add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
2246 // Use that pc we placed in O7 a while back as the current frame anchor
2248 __ set_last_Java_frame(SP, O7);
2250 // Transition from _thread_in_Java to _thread_in_native.
2251 __ set(_thread_in_native, G3_scratch);
2252 __ st(G3_scratch, G2_thread, in_bytes(JavaThread::thread_state_offset()));
2254 // We flushed the windows ages ago now mark them as flushed
2256 // mark windows as flushed
2257 __ set(JavaFrameAnchor::flushed, G3_scratch);
2259 Address flags(G2_thread,
2260 0,
2261 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
2263 #ifdef _LP64
2264 Address dest(O7, method->native_function());
2265 __ relocate(relocInfo::runtime_call_type);
2266 __ jumpl_to(dest, O7);
2267 #else
2268 __ call(method->native_function(), relocInfo::runtime_call_type);
2269 #endif
2270 __ delayed()->st(G3_scratch, flags);
2272 __ restore_thread(L7_thread_cache); // restore G2_thread
2274 // Unpack native results. For int-types, we do any needed sign-extension
2275 // and move things into I0. The return value there will survive any VM
2276 // calls for blocking or unlocking. An FP or OOP result (handle) is done
2277 // specially in the slow-path code.
2278 switch (ret_type) {
2279 case T_VOID: break; // Nothing to do!
2280 case T_FLOAT: break; // Got it where we want it (unless slow-path)
2281 case T_DOUBLE: break; // Got it where we want it (unless slow-path)
2282 // In 64 bits build result is in O0, in O0, O1 in 32bit build
2283 case T_LONG:
2284 #ifndef _LP64
2285 __ mov(O1, I1);
2286 #endif
2287 // Fall thru
2288 case T_OBJECT: // Really a handle
2289 case T_ARRAY:
2290 case T_INT:
2291 __ mov(O0, I0);
2292 break;
2293 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, I0); break; // !0 => true; 0 => false
2294 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, I0); break;
2295 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, I0); break; // cannot use and3, 0xFFFF too big as immediate value!
2296 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, I0); break;
2297 break; // Cannot de-handlize until after reclaiming jvm_lock
2298 default:
2299 ShouldNotReachHere();
2300 }
2302 // must we block?
2304 // Block, if necessary, before resuming in _thread_in_Java state.
2305 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2306 { Label no_block;
2307 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
2309 // Switch thread to "native transition" state before reading the synchronization state.
2310 // This additional state is necessary because reading and testing the synchronization
2311 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2312 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2313 // VM thread changes sync state to synchronizing and suspends threads for GC.
2314 // Thread A is resumed to finish this native method, but doesn't block here since it
2315 // didn't see any synchronization is progress, and escapes.
2316 __ set(_thread_in_native_trans, G3_scratch);
2317 __ st(G3_scratch, G2_thread, in_bytes(JavaThread::thread_state_offset()));
2318 if(os::is_MP()) {
2319 if (UseMembar) {
2320 // Force this write out before the read below
2321 __ membar(Assembler::StoreLoad);
2322 } else {
2323 // Write serialization page so VM thread can do a pseudo remote membar.
2324 // We use the current thread pointer to calculate a thread specific
2325 // offset to write to within the page. This minimizes bus traffic
2326 // due to cache line collision.
2327 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
2328 }
2329 }
2330 __ load_contents(sync_state, G3_scratch);
2331 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
2333 Label L;
2334 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
2335 __ br(Assembler::notEqual, false, Assembler::pn, L);
2336 __ delayed()->
2337 ld(suspend_state, G3_scratch);
2338 __ cmp(G3_scratch, 0);
2339 __ br(Assembler::equal, false, Assembler::pt, no_block);
2340 __ delayed()->nop();
2341 __ bind(L);
2343 // Block. Save any potential method result value before the operation and
2344 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2345 // lets us share the oopMap we used when we went native rather the create
2346 // a distinct one for this pc
2347 //
2348 save_native_result(masm, ret_type, stack_slots);
2349 __ call_VM_leaf(L7_thread_cache,
2350 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
2351 G2_thread);
2353 // Restore any method result value
2354 restore_native_result(masm, ret_type, stack_slots);
2355 __ bind(no_block);
2356 }
2358 // thread state is thread_in_native_trans. Any safepoint blocking has already
2359 // happened so we can now change state to _thread_in_Java.
2362 __ set(_thread_in_Java, G3_scratch);
2363 __ st(G3_scratch, G2_thread, in_bytes(JavaThread::thread_state_offset()));
2366 Label no_reguard;
2367 __ ld(G2_thread, in_bytes(JavaThread::stack_guard_state_offset()), G3_scratch);
2368 __ cmp(G3_scratch, JavaThread::stack_guard_yellow_disabled);
2369 __ br(Assembler::notEqual, false, Assembler::pt, no_reguard);
2370 __ delayed()->nop();
2372 save_native_result(masm, ret_type, stack_slots);
2373 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2374 __ delayed()->nop();
2376 __ restore_thread(L7_thread_cache); // restore G2_thread
2377 restore_native_result(masm, ret_type, stack_slots);
2379 __ bind(no_reguard);
2381 // Handle possible exception (will unlock if necessary)
2383 // native result if any is live in freg or I0 (and I1 if long and 32bit vm)
2385 // Unlock
2386 if (method->is_synchronized()) {
2387 Label done;
2388 Register I2_ex_oop = I2;
2389 const Register L3_box = L3;
2390 // Get locked oop from the handle we passed to jni
2391 __ ld_ptr(L6_handle, 0, L4);
2392 __ add(SP, lock_offset+STACK_BIAS, L3_box);
2393 // Must save pending exception around the slow-path VM call. Since it's a
2394 // leaf call, the pending exception (if any) can be kept in a register.
2395 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), I2_ex_oop);
2396 // Now unlock
2397 // (Roop, Rmark, Rbox, Rscratch)
2398 __ compiler_unlock_object(L4, L1, L3_box, L2);
2399 __ br(Assembler::equal, false, Assembler::pt, done);
2400 __ delayed()-> add(SP, lock_offset+STACK_BIAS, L3_box);
2402 // save and restore any potential method result value around the unlocking
2403 // operation. Will save in I0 (or stack for FP returns).
2404 save_native_result(masm, ret_type, stack_slots);
2406 // Must clear pending-exception before re-entering the VM. Since this is
2407 // a leaf call, pending-exception-oop can be safely kept in a register.
2408 __ st_ptr(G0, G2_thread, in_bytes(Thread::pending_exception_offset()));
2410 // slow case of monitor enter. Inline a special case of call_VM that
2411 // disallows any pending_exception.
2412 __ mov(L3_box, O1);
2414 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), relocInfo::runtime_call_type);
2415 __ delayed()->mov(L4, O0); // Need oop in O0
2417 __ restore_thread(L7_thread_cache); // restore G2_thread
2419 #ifdef ASSERT
2420 { Label L;
2421 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
2422 __ br_null(O0, false, Assembler::pt, L);
2423 __ delayed()->nop();
2424 __ stop("no pending exception allowed on exit from IR::monitorexit");
2425 __ bind(L);
2426 }
2427 #endif
2428 restore_native_result(masm, ret_type, stack_slots);
2429 // check_forward_pending_exception jump to forward_exception if any pending
2430 // exception is set. The forward_exception routine expects to see the
2431 // exception in pending_exception and not in a register. Kind of clumsy,
2432 // since all folks who branch to forward_exception must have tested
2433 // pending_exception first and hence have it in a register already.
2434 __ st_ptr(I2_ex_oop, G2_thread, in_bytes(Thread::pending_exception_offset()));
2435 __ bind(done);
2436 }
2438 // Tell dtrace about this method exit
2439 {
2440 SkipIfEqual skip_if(
2441 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
2442 save_native_result(masm, ret_type, stack_slots);
2443 __ set_oop_constant(JNIHandles::make_local(method()), O1);
2444 __ call_VM_leaf(L7_thread_cache,
2445 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2446 G2_thread, O1);
2447 restore_native_result(masm, ret_type, stack_slots);
2448 }
2450 // Clear "last Java frame" SP and PC.
2451 __ verify_thread(); // G2_thread must be correct
2452 __ reset_last_Java_frame();
2454 // Unpack oop result
2455 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
2456 Label L;
2457 __ addcc(G0, I0, G0);
2458 __ brx(Assembler::notZero, true, Assembler::pt, L);
2459 __ delayed()->ld_ptr(I0, 0, I0);
2460 __ mov(G0, I0);
2461 __ bind(L);
2462 __ verify_oop(I0);
2463 }
2465 // reset handle block
2466 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), L5);
2467 __ st_ptr(G0, L5, JNIHandleBlock::top_offset_in_bytes());
2469 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), G3_scratch);
2470 check_forward_pending_exception(masm, G3_scratch);
2473 // Return
2475 #ifndef _LP64
2476 if (ret_type == T_LONG) {
2478 // Must leave proper result in O0,O1 and G1 (c2/tiered only)
2479 __ sllx(I0, 32, G1); // Shift bits into high G1
2480 __ srl (I1, 0, I1); // Zero extend O1 (harmless?)
2481 __ or3 (I1, G1, G1); // OR 64 bits into G1
2482 }
2483 #endif
2485 __ ret();
2486 __ delayed()->restore();
2488 __ flush();
2490 nmethod *nm = nmethod::new_native_nmethod(method,
2491 masm->code(),
2492 vep_offset,
2493 frame_complete,
2494 stack_slots / VMRegImpl::slots_per_word,
2495 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2496 in_ByteSize(lock_offset),
2497 oop_maps);
2498 return nm;
2500 }
2502 #ifdef HAVE_DTRACE_H
2503 // ---------------------------------------------------------------------------
2504 // Generate a dtrace nmethod for a given signature. The method takes arguments
2505 // in the Java compiled code convention, marshals them to the native
2506 // abi and then leaves nops at the position you would expect to call a native
2507 // function. When the probe is enabled the nops are replaced with a trap
2508 // instruction that dtrace inserts and the trace will cause a notification
2509 // to dtrace.
2510 //
2511 // The probes are only able to take primitive types and java/lang/String as
2512 // arguments. No other java types are allowed. Strings are converted to utf8
2513 // strings so that from dtrace point of view java strings are converted to C
2514 // strings. There is an arbitrary fixed limit on the total space that a method
2515 // can use for converting the strings. (256 chars per string in the signature).
2516 // So any java string larger then this is truncated.
2518 static int fp_offset[ConcreteRegisterImpl::number_of_registers] = { 0 };
2519 static bool offsets_initialized = false;
2521 static VMRegPair reg64_to_VMRegPair(Register r) {
2522 VMRegPair ret;
2523 if (wordSize == 8) {
2524 ret.set2(r->as_VMReg());
2525 } else {
2526 ret.set_pair(r->successor()->as_VMReg(), r->as_VMReg());
2527 }
2528 return ret;
2529 }
2532 nmethod *SharedRuntime::generate_dtrace_nmethod(
2533 MacroAssembler *masm, methodHandle method) {
2536 // generate_dtrace_nmethod is guarded by a mutex so we are sure to
2537 // be single threaded in this method.
2538 assert(AdapterHandlerLibrary_lock->owned_by_self(), "must be");
2540 // Fill in the signature array, for the calling-convention call.
2541 int total_args_passed = method->size_of_parameters();
2543 BasicType* in_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed);
2544 VMRegPair *in_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed);
2546 // The signature we are going to use for the trap that dtrace will see
2547 // java/lang/String is converted. We drop "this" and any other object
2548 // is converted to NULL. (A one-slot java/lang/Long object reference
2549 // is converted to a two-slot long, which is why we double the allocation).
2550 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed * 2);
2551 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed * 2);
2553 int i=0;
2554 int total_strings = 0;
2555 int first_arg_to_pass = 0;
2556 int total_c_args = 0;
2558 // Skip the receiver as dtrace doesn't want to see it
2559 if( !method->is_static() ) {
2560 in_sig_bt[i++] = T_OBJECT;
2561 first_arg_to_pass = 1;
2562 }
2564 SignatureStream ss(method->signature());
2565 for ( ; !ss.at_return_type(); ss.next()) {
2566 BasicType bt = ss.type();
2567 in_sig_bt[i++] = bt; // Collect remaining bits of signature
2568 out_sig_bt[total_c_args++] = bt;
2569 if( bt == T_OBJECT) {
2570 symbolOop s = ss.as_symbol_or_null();
2571 if (s == vmSymbols::java_lang_String()) {
2572 total_strings++;
2573 out_sig_bt[total_c_args-1] = T_ADDRESS;
2574 } else if (s == vmSymbols::java_lang_Boolean() ||
2575 s == vmSymbols::java_lang_Byte()) {
2576 out_sig_bt[total_c_args-1] = T_BYTE;
2577 } else if (s == vmSymbols::java_lang_Character() ||
2578 s == vmSymbols::java_lang_Short()) {
2579 out_sig_bt[total_c_args-1] = T_SHORT;
2580 } else if (s == vmSymbols::java_lang_Integer() ||
2581 s == vmSymbols::java_lang_Float()) {
2582 out_sig_bt[total_c_args-1] = T_INT;
2583 } else if (s == vmSymbols::java_lang_Long() ||
2584 s == vmSymbols::java_lang_Double()) {
2585 out_sig_bt[total_c_args-1] = T_LONG;
2586 out_sig_bt[total_c_args++] = T_VOID;
2587 }
2588 } else if ( bt == T_LONG || bt == T_DOUBLE ) {
2589 in_sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots
2590 // We convert double to long
2591 out_sig_bt[total_c_args-1] = T_LONG;
2592 out_sig_bt[total_c_args++] = T_VOID;
2593 } else if ( bt == T_FLOAT) {
2594 // We convert float to int
2595 out_sig_bt[total_c_args-1] = T_INT;
2596 }
2597 }
2599 assert(i==total_args_passed, "validly parsed signature");
2601 // Now get the compiled-Java layout as input arguments
2602 int comp_args_on_stack;
2603 comp_args_on_stack = SharedRuntime::java_calling_convention(
2604 in_sig_bt, in_regs, total_args_passed, false);
2606 // We have received a description of where all the java arg are located
2607 // on entry to the wrapper. We need to convert these args to where
2608 // the a native (non-jni) function would expect them. To figure out
2609 // where they go we convert the java signature to a C signature and remove
2610 // T_VOID for any long/double we might have received.
2613 // Now figure out where the args must be stored and how much stack space
2614 // they require (neglecting out_preserve_stack_slots but space for storing
2615 // the 1st six register arguments). It's weird see int_stk_helper.
2616 //
2617 int out_arg_slots;
2618 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
2620 // Calculate the total number of stack slots we will need.
2622 // First count the abi requirement plus all of the outgoing args
2623 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2625 // Plus a temp for possible converion of float/double/long register args
2627 int conversion_temp = stack_slots;
2628 stack_slots += 2;
2631 // Now space for the string(s) we must convert
2633 int string_locs = stack_slots;
2634 stack_slots += total_strings *
2635 (max_dtrace_string_size / VMRegImpl::stack_slot_size);
2637 // Ok The space we have allocated will look like:
2638 //
2639 //
2640 // FP-> | |
2641 // |---------------------|
2642 // | string[n] |
2643 // |---------------------| <- string_locs[n]
2644 // | string[n-1] |
2645 // |---------------------| <- string_locs[n-1]
2646 // | ... |
2647 // | ... |
2648 // |---------------------| <- string_locs[1]
2649 // | string[0] |
2650 // |---------------------| <- string_locs[0]
2651 // | temp |
2652 // |---------------------| <- conversion_temp
2653 // | outbound memory |
2654 // | based arguments |
2655 // | |
2656 // |---------------------|
2657 // | |
2658 // SP-> | out_preserved_slots |
2659 //
2660 //
2662 // Now compute actual number of stack words we need rounding to make
2663 // stack properly aligned.
2664 stack_slots = round_to(stack_slots, 4 * VMRegImpl::slots_per_word);
2666 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2668 intptr_t start = (intptr_t)__ pc();
2670 // First thing make an ic check to see if we should even be here
2672 {
2673 Label L;
2674 const Register temp_reg = G3_scratch;
2675 Address ic_miss(temp_reg, SharedRuntime::get_ic_miss_stub());
2676 __ verify_oop(O0);
2677 __ ld_ptr(O0, oopDesc::klass_offset_in_bytes(), temp_reg);
2678 __ cmp(temp_reg, G5_inline_cache_reg);
2679 __ brx(Assembler::equal, true, Assembler::pt, L);
2680 __ delayed()->nop();
2682 __ jump_to(ic_miss, 0);
2683 __ delayed()->nop();
2684 __ align(CodeEntryAlignment);
2685 __ bind(L);
2686 }
2688 int vep_offset = ((intptr_t)__ pc()) - start;
2691 // The instruction at the verified entry point must be 5 bytes or longer
2692 // because it can be patched on the fly by make_non_entrant. The stack bang
2693 // instruction fits that requirement.
2695 // Generate stack overflow check before creating frame
2696 __ generate_stack_overflow_check(stack_size);
2698 assert(((intptr_t)__ pc() - start - vep_offset) >= 5,
2699 "valid size for make_non_entrant");
2701 // Generate a new frame for the wrapper.
2702 __ save(SP, -stack_size, SP);
2704 // Frame is now completed as far a size and linkage.
2706 int frame_complete = ((intptr_t)__ pc()) - start;
2708 #ifdef ASSERT
2709 bool reg_destroyed[RegisterImpl::number_of_registers];
2710 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
2711 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
2712 reg_destroyed[r] = false;
2713 }
2714 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
2715 freg_destroyed[f] = false;
2716 }
2718 #endif /* ASSERT */
2720 VMRegPair zero;
2721 const Register g0 = G0; // without this we get a compiler warning (why??)
2722 zero.set2(g0->as_VMReg());
2724 int c_arg, j_arg;
2726 Register conversion_off = noreg;
2728 for (j_arg = first_arg_to_pass, c_arg = 0 ;
2729 j_arg < total_args_passed ; j_arg++, c_arg++ ) {
2731 VMRegPair src = in_regs[j_arg];
2732 VMRegPair dst = out_regs[c_arg];
2734 #ifdef ASSERT
2735 if (src.first()->is_Register()) {
2736 assert(!reg_destroyed[src.first()->as_Register()->encoding()], "ack!");
2737 } else if (src.first()->is_FloatRegister()) {
2738 assert(!freg_destroyed[src.first()->as_FloatRegister()->encoding(
2739 FloatRegisterImpl::S)], "ack!");
2740 }
2741 if (dst.first()->is_Register()) {
2742 reg_destroyed[dst.first()->as_Register()->encoding()] = true;
2743 } else if (dst.first()->is_FloatRegister()) {
2744 freg_destroyed[dst.first()->as_FloatRegister()->encoding(
2745 FloatRegisterImpl::S)] = true;
2746 }
2747 #endif /* ASSERT */
2749 switch (in_sig_bt[j_arg]) {
2750 case T_ARRAY:
2751 case T_OBJECT:
2752 {
2753 if (out_sig_bt[c_arg] == T_BYTE || out_sig_bt[c_arg] == T_SHORT ||
2754 out_sig_bt[c_arg] == T_INT || out_sig_bt[c_arg] == T_LONG) {
2755 // need to unbox a one-slot value
2756 Register in_reg = L0;
2757 Register tmp = L2;
2758 if ( src.first()->is_reg() ) {
2759 in_reg = src.first()->as_Register();
2760 } else {
2761 assert(Assembler::is_simm13(reg2offset(src.first()) + STACK_BIAS),
2762 "must be");
2763 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, in_reg);
2764 }
2765 // If the final destination is an acceptable register
2766 if ( dst.first()->is_reg() ) {
2767 if ( dst.is_single_phys_reg() || out_sig_bt[c_arg] != T_LONG ) {
2768 tmp = dst.first()->as_Register();
2769 }
2770 }
2772 Label skipUnbox;
2773 if ( wordSize == 4 && out_sig_bt[c_arg] == T_LONG ) {
2774 __ mov(G0, tmp->successor());
2775 }
2776 __ br_null(in_reg, true, Assembler::pn, skipUnbox);
2777 __ delayed()->mov(G0, tmp);
2779 BasicType bt = out_sig_bt[c_arg];
2780 int box_offset = java_lang_boxing_object::value_offset_in_bytes(bt);
2781 switch (bt) {
2782 case T_BYTE:
2783 __ ldub(in_reg, box_offset, tmp); break;
2784 case T_SHORT:
2785 __ lduh(in_reg, box_offset, tmp); break;
2786 case T_INT:
2787 __ ld(in_reg, box_offset, tmp); break;
2788 case T_LONG:
2789 __ ld_long(in_reg, box_offset, tmp); break;
2790 default: ShouldNotReachHere();
2791 }
2793 __ bind(skipUnbox);
2794 // If tmp wasn't final destination copy to final destination
2795 if (tmp == L2) {
2796 VMRegPair tmp_as_VM = reg64_to_VMRegPair(L2);
2797 if (out_sig_bt[c_arg] == T_LONG) {
2798 long_move(masm, tmp_as_VM, dst);
2799 } else {
2800 move32_64(masm, tmp_as_VM, out_regs[c_arg]);
2801 }
2802 }
2803 if (out_sig_bt[c_arg] == T_LONG) {
2804 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
2805 ++c_arg; // move over the T_VOID to keep the loop indices in sync
2806 }
2807 } else if (out_sig_bt[c_arg] == T_ADDRESS) {
2808 Register s =
2809 src.first()->is_reg() ? src.first()->as_Register() : L2;
2810 Register d =
2811 dst.first()->is_reg() ? dst.first()->as_Register() : L2;
2813 // We store the oop now so that the conversion pass can reach
2814 // while in the inner frame. This will be the only store if
2815 // the oop is NULL.
2816 if (s != L2) {
2817 // src is register
2818 if (d != L2) {
2819 // dst is register
2820 __ mov(s, d);
2821 } else {
2822 assert(Assembler::is_simm13(reg2offset(dst.first()) +
2823 STACK_BIAS), "must be");
2824 __ st_ptr(s, SP, reg2offset(dst.first()) + STACK_BIAS);
2825 }
2826 } else {
2827 // src not a register
2828 assert(Assembler::is_simm13(reg2offset(src.first()) +
2829 STACK_BIAS), "must be");
2830 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, d);
2831 if (d == L2) {
2832 assert(Assembler::is_simm13(reg2offset(dst.first()) +
2833 STACK_BIAS), "must be");
2834 __ st_ptr(d, SP, reg2offset(dst.first()) + STACK_BIAS);
2835 }
2836 }
2837 } else if (out_sig_bt[c_arg] != T_VOID) {
2838 // Convert the arg to NULL
2839 if (dst.first()->is_reg()) {
2840 __ mov(G0, dst.first()->as_Register());
2841 } else {
2842 assert(Assembler::is_simm13(reg2offset(dst.first()) +
2843 STACK_BIAS), "must be");
2844 __ st_ptr(G0, SP, reg2offset(dst.first()) + STACK_BIAS);
2845 }
2846 }
2847 }
2848 break;
2849 case T_VOID:
2850 break;
2852 case T_FLOAT:
2853 if (src.first()->is_stack()) {
2854 // Stack to stack/reg is simple
2855 move32_64(masm, src, dst);
2856 } else {
2857 if (dst.first()->is_reg()) {
2858 // freg -> reg
2859 int off =
2860 STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
2861 Register d = dst.first()->as_Register();
2862 if (Assembler::is_simm13(off)) {
2863 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
2864 SP, off);
2865 __ ld(SP, off, d);
2866 } else {
2867 if (conversion_off == noreg) {
2868 __ set(off, L6);
2869 conversion_off = L6;
2870 }
2871 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
2872 SP, conversion_off);
2873 __ ld(SP, conversion_off , d);
2874 }
2875 } else {
2876 // freg -> mem
2877 int off = STACK_BIAS + reg2offset(dst.first());
2878 if (Assembler::is_simm13(off)) {
2879 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
2880 SP, off);
2881 } else {
2882 if (conversion_off == noreg) {
2883 __ set(off, L6);
2884 conversion_off = L6;
2885 }
2886 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
2887 SP, conversion_off);
2888 }
2889 }
2890 }
2891 break;
2893 case T_DOUBLE:
2894 assert( j_arg + 1 < total_args_passed &&
2895 in_sig_bt[j_arg + 1] == T_VOID &&
2896 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
2897 if (src.first()->is_stack()) {
2898 // Stack to stack/reg is simple
2899 long_move(masm, src, dst);
2900 } else {
2901 Register d = dst.first()->is_reg() ? dst.first()->as_Register() : L2;
2903 // Destination could be an odd reg on 32bit in which case
2904 // we can't load direct to the destination.
2906 if (!d->is_even() && wordSize == 4) {
2907 d = L2;
2908 }
2909 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
2910 if (Assembler::is_simm13(off)) {
2911 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
2912 SP, off);
2913 __ ld_long(SP, off, d);
2914 } else {
2915 if (conversion_off == noreg) {
2916 __ set(off, L6);
2917 conversion_off = L6;
2918 }
2919 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
2920 SP, conversion_off);
2921 __ ld_long(SP, conversion_off, d);
2922 }
2923 if (d == L2) {
2924 long_move(masm, reg64_to_VMRegPair(L2), dst);
2925 }
2926 }
2927 break;
2929 case T_LONG :
2930 // 32bit can't do a split move of something like g1 -> O0, O1
2931 // so use a memory temp
2932 if (src.is_single_phys_reg() && wordSize == 4) {
2933 Register tmp = L2;
2934 if (dst.first()->is_reg() &&
2935 (wordSize == 8 || dst.first()->as_Register()->is_even())) {
2936 tmp = dst.first()->as_Register();
2937 }
2939 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
2940 if (Assembler::is_simm13(off)) {
2941 __ stx(src.first()->as_Register(), SP, off);
2942 __ ld_long(SP, off, tmp);
2943 } else {
2944 if (conversion_off == noreg) {
2945 __ set(off, L6);
2946 conversion_off = L6;
2947 }
2948 __ stx(src.first()->as_Register(), SP, conversion_off);
2949 __ ld_long(SP, conversion_off, tmp);
2950 }
2952 if (tmp == L2) {
2953 long_move(masm, reg64_to_VMRegPair(L2), dst);
2954 }
2955 } else {
2956 long_move(masm, src, dst);
2957 }
2958 break;
2960 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2962 default:
2963 move32_64(masm, src, dst);
2964 }
2965 }
2968 // If we have any strings we must store any register based arg to the stack
2969 // This includes any still live xmm registers too.
2971 if (total_strings > 0 ) {
2973 // protect all the arg registers
2974 __ save_frame(0);
2975 __ mov(G2_thread, L7_thread_cache);
2976 const Register L2_string_off = L2;
2978 // Get first string offset
2979 __ set(string_locs * VMRegImpl::stack_slot_size, L2_string_off);
2981 for (c_arg = 0 ; c_arg < total_c_args ; c_arg++ ) {
2982 if (out_sig_bt[c_arg] == T_ADDRESS) {
2984 VMRegPair dst = out_regs[c_arg];
2985 const Register d = dst.first()->is_reg() ?
2986 dst.first()->as_Register()->after_save() : noreg;
2988 // It's a string the oop and it was already copied to the out arg
2989 // position
2990 if (d != noreg) {
2991 __ mov(d, O0);
2992 } else {
2993 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
2994 "must be");
2995 __ ld_ptr(FP, reg2offset(dst.first()) + STACK_BIAS, O0);
2996 }
2997 Label skip;
2999 __ br_null(O0, false, Assembler::pn, skip);
3000 __ delayed()->add(FP, L2_string_off, O1);
3002 if (d != noreg) {
3003 __ mov(O1, d);
3004 } else {
3005 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
3006 "must be");
3007 __ st_ptr(O1, FP, reg2offset(dst.first()) + STACK_BIAS);
3008 }
3010 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::get_utf),
3011 relocInfo::runtime_call_type);
3012 __ delayed()->add(L2_string_off, max_dtrace_string_size, L2_string_off);
3014 __ bind(skip);
3016 }
3018 }
3019 __ mov(L7_thread_cache, G2_thread);
3020 __ restore();
3022 }
3025 // Ok now we are done. Need to place the nop that dtrace wants in order to
3026 // patch in the trap
3028 int patch_offset = ((intptr_t)__ pc()) - start;
3030 __ nop();
3033 // Return
3035 __ ret();
3036 __ delayed()->restore();
3038 __ flush();
3040 nmethod *nm = nmethod::new_dtrace_nmethod(
3041 method, masm->code(), vep_offset, patch_offset, frame_complete,
3042 stack_slots / VMRegImpl::slots_per_word);
3043 return nm;
3045 }
3047 #endif // HAVE_DTRACE_H
3049 // this function returns the adjust size (in number of words) to a c2i adapter
3050 // activation for use during deoptimization
3051 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
3052 assert(callee_locals >= callee_parameters,
3053 "test and remove; got more parms than locals");
3054 if (callee_locals < callee_parameters)
3055 return 0; // No adjustment for negative locals
3056 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords();
3057 return round_to(diff, WordsPerLong);
3058 }
3060 // "Top of Stack" slots that may be unused by the calling convention but must
3061 // otherwise be preserved.
3062 // On Intel these are not necessary and the value can be zero.
3063 // On Sparc this describes the words reserved for storing a register window
3064 // when an interrupt occurs.
3065 uint SharedRuntime::out_preserve_stack_slots() {
3066 return frame::register_save_words * VMRegImpl::slots_per_word;
3067 }
3069 static void gen_new_frame(MacroAssembler* masm, bool deopt) {
3070 //
3071 // Common out the new frame generation for deopt and uncommon trap
3072 //
3073 Register G3pcs = G3_scratch; // Array of new pcs (input)
3074 Register Oreturn0 = O0;
3075 Register Oreturn1 = O1;
3076 Register O2UnrollBlock = O2;
3077 Register O3array = O3; // Array of frame sizes (input)
3078 Register O4array_size = O4; // number of frames (input)
3079 Register O7frame_size = O7; // number of frames (input)
3081 __ ld_ptr(O3array, 0, O7frame_size);
3082 __ sub(G0, O7frame_size, O7frame_size);
3083 __ save(SP, O7frame_size, SP);
3084 __ ld_ptr(G3pcs, 0, I7); // load frame's new pc
3086 #ifdef ASSERT
3087 // make sure that the frames are aligned properly
3088 #ifndef _LP64
3089 __ btst(wordSize*2-1, SP);
3090 __ breakpoint_trap(Assembler::notZero);
3091 #endif
3092 #endif
3094 // Deopt needs to pass some extra live values from frame to frame
3096 if (deopt) {
3097 __ mov(Oreturn0->after_save(), Oreturn0);
3098 __ mov(Oreturn1->after_save(), Oreturn1);
3099 }
3101 __ mov(O4array_size->after_save(), O4array_size);
3102 __ sub(O4array_size, 1, O4array_size);
3103 __ mov(O3array->after_save(), O3array);
3104 __ mov(O2UnrollBlock->after_save(), O2UnrollBlock);
3105 __ add(G3pcs, wordSize, G3pcs); // point to next pc value
3107 #ifdef ASSERT
3108 // trash registers to show a clear pattern in backtraces
3109 __ set(0xDEAD0000, I0);
3110 __ add(I0, 2, I1);
3111 __ add(I0, 4, I2);
3112 __ add(I0, 6, I3);
3113 __ add(I0, 8, I4);
3114 // Don't touch I5 could have valuable savedSP
3115 __ set(0xDEADBEEF, L0);
3116 __ mov(L0, L1);
3117 __ mov(L0, L2);
3118 __ mov(L0, L3);
3119 __ mov(L0, L4);
3120 __ mov(L0, L5);
3122 // trash the return value as there is nothing to return yet
3123 __ set(0xDEAD0001, O7);
3124 #endif
3126 __ mov(SP, O5_savedSP);
3127 }
3130 static void make_new_frames(MacroAssembler* masm, bool deopt) {
3131 //
3132 // loop through the UnrollBlock info and create new frames
3133 //
3134 Register G3pcs = G3_scratch;
3135 Register Oreturn0 = O0;
3136 Register Oreturn1 = O1;
3137 Register O2UnrollBlock = O2;
3138 Register O3array = O3;
3139 Register O4array_size = O4;
3140 Label loop;
3142 // Before we make new frames, check to see if stack is available.
3143 // Do this after the caller's return address is on top of stack
3144 if (UseStackBanging) {
3145 // Get total frame size for interpreted frames
3146 __ ld(Address(O2UnrollBlock, 0,
3147 Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes()), O4);
3148 __ bang_stack_size(O4, O3, G3_scratch);
3149 }
3151 __ ld(Address(O2UnrollBlock, 0, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()), O4array_size);
3152 __ ld_ptr(Address(O2UnrollBlock, 0, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()), G3pcs);
3154 __ ld_ptr(Address(O2UnrollBlock, 0, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()), O3array);
3156 // Adjust old interpreter frame to make space for new frame's extra java locals
3157 //
3158 // We capture the original sp for the transition frame only because it is needed in
3159 // order to properly calculate interpreter_sp_adjustment. Even though in real life
3160 // every interpreter frame captures a savedSP it is only needed at the transition
3161 // (fortunately). If we had to have it correct everywhere then we would need to
3162 // be told the sp_adjustment for each frame we create. If the frame size array
3163 // were to have twice the frame count entries then we could have pairs [sp_adjustment, frame_size]
3164 // for each frame we create and keep up the illusion every where.
3165 //
3167 __ ld(Address(O2UnrollBlock, 0, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()), O7);
3168 __ mov(SP, O5_savedSP); // remember initial sender's original sp before adjustment
3169 __ sub(SP, O7, SP);
3171 #ifdef ASSERT
3172 // make sure that there is at least one entry in the array
3173 __ tst(O4array_size);
3174 __ breakpoint_trap(Assembler::zero);
3175 #endif
3177 // Now push the new interpreter frames
3178 __ bind(loop);
3180 // allocate a new frame, filling the registers
3182 gen_new_frame(masm, deopt); // allocate an interpreter frame
3184 __ tst(O4array_size);
3185 __ br(Assembler::notZero, false, Assembler::pn, loop);
3186 __ delayed()->add(O3array, wordSize, O3array);
3187 __ ld_ptr(G3pcs, 0, O7); // load final frame new pc
3189 }
3191 //------------------------------generate_deopt_blob----------------------------
3192 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
3193 // instead.
3194 void SharedRuntime::generate_deopt_blob() {
3195 // allocate space for the code
3196 ResourceMark rm;
3197 // setup code generation tools
3198 int pad = VerifyThread ? 512 : 0;// Extra slop space for more verify code
3199 #ifdef _LP64
3200 CodeBuffer buffer("deopt_blob", 2100+pad, 512);
3201 #else
3202 // Measured 8/7/03 at 1212 in 32bit debug build (no VerifyThread)
3203 // Measured 8/7/03 at 1396 in 32bit debug build (VerifyThread)
3204 CodeBuffer buffer("deopt_blob", 1600+pad, 512);
3205 #endif /* _LP64 */
3206 MacroAssembler* masm = new MacroAssembler(&buffer);
3207 FloatRegister Freturn0 = F0;
3208 Register Greturn1 = G1;
3209 Register Oreturn0 = O0;
3210 Register Oreturn1 = O1;
3211 Register O2UnrollBlock = O2;
3212 Register O3tmp = O3;
3213 Register I5exception_tmp = I5;
3214 Register G4exception_tmp = G4_scratch;
3215 int frame_size_words;
3216 Address saved_Freturn0_addr(FP, 0, -sizeof(double) + STACK_BIAS);
3217 #if !defined(_LP64) && defined(COMPILER2)
3218 Address saved_Greturn1_addr(FP, 0, -sizeof(double) -sizeof(jlong) + STACK_BIAS);
3219 #endif
3220 Label cont;
3222 OopMapSet *oop_maps = new OopMapSet();
3224 //
3225 // This is the entry point for code which is returning to a de-optimized
3226 // frame.
3227 // The steps taken by this frame are as follows:
3228 // - push a dummy "register_save" and save the return values (O0, O1, F0/F1, G1)
3229 // and all potentially live registers (at a pollpoint many registers can be live).
3230 //
3231 // - call the C routine: Deoptimization::fetch_unroll_info (this function
3232 // returns information about the number and size of interpreter frames
3233 // which are equivalent to the frame which is being deoptimized)
3234 // - deallocate the unpack frame, restoring only results values. Other
3235 // volatile registers will now be captured in the vframeArray as needed.
3236 // - deallocate the deoptimization frame
3237 // - in a loop using the information returned in the previous step
3238 // push new interpreter frames (take care to propagate the return
3239 // values through each new frame pushed)
3240 // - create a dummy "unpack_frame" and save the return values (O0, O1, F0)
3241 // - call the C routine: Deoptimization::unpack_frames (this function
3242 // lays out values on the interpreter frame which was just created)
3243 // - deallocate the dummy unpack_frame
3244 // - ensure that all the return values are correctly set and then do
3245 // a return to the interpreter entry point
3246 //
3247 // Refer to the following methods for more information:
3248 // - Deoptimization::fetch_unroll_info
3249 // - Deoptimization::unpack_frames
3251 OopMap* map = NULL;
3253 int start = __ offset();
3255 // restore G2, the trampoline destroyed it
3256 __ get_thread();
3258 // On entry we have been called by the deoptimized nmethod with a call that
3259 // replaced the original call (or safepoint polling location) so the deoptimizing
3260 // pc is now in O7. Return values are still in the expected places
3262 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3263 __ ba(false, cont);
3264 __ delayed()->mov(Deoptimization::Unpack_deopt, I5exception_tmp);
3266 int exception_offset = __ offset() - start;
3268 // restore G2, the trampoline destroyed it
3269 __ get_thread();
3271 // On entry we have been jumped to by the exception handler (or exception_blob
3272 // for server). O0 contains the exception oop and O7 contains the original
3273 // exception pc. So if we push a frame here it will look to the
3274 // stack walking code (fetch_unroll_info) just like a normal call so
3275 // state will be extracted normally.
3277 // save exception oop in JavaThread and fall through into the
3278 // exception_in_tls case since they are handled in same way except
3279 // for where the pending exception is kept.
3280 __ st_ptr(Oexception, G2_thread, in_bytes(JavaThread::exception_oop_offset()));
3282 //
3283 // Vanilla deoptimization with an exception pending in exception_oop
3284 //
3285 int exception_in_tls_offset = __ offset() - start;
3287 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
3288 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3290 // Restore G2_thread
3291 __ get_thread();
3293 #ifdef ASSERT
3294 {
3295 // verify that there is really an exception oop in exception_oop
3296 Label has_exception;
3297 __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception);
3298 __ br_notnull(Oexception, false, Assembler::pt, has_exception);
3299 __ delayed()-> nop();
3300 __ stop("no exception in thread");
3301 __ bind(has_exception);
3303 // verify that there is no pending exception
3304 Label no_pending_exception;
3305 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
3306 __ ld_ptr(exception_addr, Oexception);
3307 __ br_null(Oexception, false, Assembler::pt, no_pending_exception);
3308 __ delayed()->nop();
3309 __ stop("must not have pending exception here");
3310 __ bind(no_pending_exception);
3311 }
3312 #endif
3314 __ ba(false, cont);
3315 __ delayed()->mov(Deoptimization::Unpack_exception, I5exception_tmp);;
3317 //
3318 // Reexecute entry, similar to c2 uncommon trap
3319 //
3320 int reexecute_offset = __ offset() - start;
3322 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
3323 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3325 __ mov(Deoptimization::Unpack_reexecute, I5exception_tmp);
3327 __ bind(cont);
3329 __ set_last_Java_frame(SP, noreg);
3331 // do the call by hand so we can get the oopmap
3333 __ mov(G2_thread, L7_thread_cache);
3334 __ call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), relocInfo::runtime_call_type);
3335 __ delayed()->mov(G2_thread, O0);
3337 // Set an oopmap for the call site this describes all our saved volatile registers
3339 oop_maps->add_gc_map( __ offset()-start, map);
3341 __ mov(L7_thread_cache, G2_thread);
3343 __ reset_last_Java_frame();
3345 // NOTE: we know that only O0/O1 will be reloaded by restore_result_registers
3346 // so this move will survive
3348 __ mov(I5exception_tmp, G4exception_tmp);
3350 __ mov(O0, O2UnrollBlock->after_save());
3352 RegisterSaver::restore_result_registers(masm);
3354 Label noException;
3355 __ cmp(G4exception_tmp, Deoptimization::Unpack_exception); // Was exception pending?
3356 __ br(Assembler::notEqual, false, Assembler::pt, noException);
3357 __ delayed()->nop();
3359 // Move the pending exception from exception_oop to Oexception so
3360 // the pending exception will be picked up the interpreter.
3361 __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception);
3362 __ st_ptr(G0, G2_thread, in_bytes(JavaThread::exception_oop_offset()));
3363 __ bind(noException);
3365 // deallocate the deoptimization frame taking care to preserve the return values
3366 __ mov(Oreturn0, Oreturn0->after_save());
3367 __ mov(Oreturn1, Oreturn1->after_save());
3368 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
3369 __ restore();
3371 // Allocate new interpreter frame(s) and possible c2i adapter frame
3373 make_new_frames(masm, true);
3375 // push a dummy "unpack_frame" taking care of float return values and
3376 // call Deoptimization::unpack_frames to have the unpacker layout
3377 // information in the interpreter frames just created and then return
3378 // to the interpreter entry point
3379 __ save(SP, -frame_size_words*wordSize, SP);
3380 __ stf(FloatRegisterImpl::D, Freturn0, saved_Freturn0_addr);
3381 #if !defined(_LP64)
3382 #if defined(COMPILER2)
3383 if (!TieredCompilation) {
3384 // 32-bit 1-register longs return longs in G1
3385 __ stx(Greturn1, saved_Greturn1_addr);
3386 }
3387 #endif
3388 __ set_last_Java_frame(SP, noreg);
3389 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, G4exception_tmp);
3390 #else
3391 // LP64 uses g4 in set_last_Java_frame
3392 __ mov(G4exception_tmp, O1);
3393 __ set_last_Java_frame(SP, G0);
3394 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O1);
3395 #endif
3396 __ reset_last_Java_frame();
3397 __ ldf(FloatRegisterImpl::D, saved_Freturn0_addr, Freturn0);
3399 // In tiered we never use C2 to compile methods returning longs so
3400 // the result is where we expect it already.
3402 #if !defined(_LP64) && defined(COMPILER2)
3403 // In 32 bit, C2 returns longs in G1 so restore the saved G1 into
3404 // I0/I1 if the return value is long. In the tiered world there is
3405 // a mismatch between how C1 and C2 return longs compiles and so
3406 // currently compilation of methods which return longs is disabled
3407 // for C2 and so is this code. Eventually C1 and C2 will do the
3408 // same thing for longs in the tiered world.
3409 if (!TieredCompilation) {
3410 Label not_long;
3411 __ cmp(O0,T_LONG);
3412 __ br(Assembler::notEqual, false, Assembler::pt, not_long);
3413 __ delayed()->nop();
3414 __ ldd(saved_Greturn1_addr,I0);
3415 __ bind(not_long);
3416 }
3417 #endif
3418 __ ret();
3419 __ delayed()->restore();
3421 masm->flush();
3422 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_words);
3423 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3424 }
3426 #ifdef COMPILER2
3428 //------------------------------generate_uncommon_trap_blob--------------------
3429 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
3430 // instead.
3431 void SharedRuntime::generate_uncommon_trap_blob() {
3432 // allocate space for the code
3433 ResourceMark rm;
3434 // setup code generation tools
3435 int pad = VerifyThread ? 512 : 0;
3436 #ifdef _LP64
3437 CodeBuffer buffer("uncommon_trap_blob", 2700+pad, 512);
3438 #else
3439 // Measured 8/7/03 at 660 in 32bit debug build (no VerifyThread)
3440 // Measured 8/7/03 at 1028 in 32bit debug build (VerifyThread)
3441 CodeBuffer buffer("uncommon_trap_blob", 2000+pad, 512);
3442 #endif
3443 MacroAssembler* masm = new MacroAssembler(&buffer);
3444 Register O2UnrollBlock = O2;
3445 Register O3tmp = O3;
3446 Register O2klass_index = O2;
3448 //
3449 // This is the entry point for all traps the compiler takes when it thinks
3450 // it cannot handle further execution of compilation code. The frame is
3451 // deoptimized in these cases and converted into interpreter frames for
3452 // execution
3453 // The steps taken by this frame are as follows:
3454 // - push a fake "unpack_frame"
3455 // - call the C routine Deoptimization::uncommon_trap (this function
3456 // packs the current compiled frame into vframe arrays and returns
3457 // information about the number and size of interpreter frames which
3458 // are equivalent to the frame which is being deoptimized)
3459 // - deallocate the "unpack_frame"
3460 // - deallocate the deoptimization frame
3461 // - in a loop using the information returned in the previous step
3462 // push interpreter frames;
3463 // - create a dummy "unpack_frame"
3464 // - call the C routine: Deoptimization::unpack_frames (this function
3465 // lays out values on the interpreter frame which was just created)
3466 // - deallocate the dummy unpack_frame
3467 // - return to the interpreter entry point
3468 //
3469 // Refer to the following methods for more information:
3470 // - Deoptimization::uncommon_trap
3471 // - Deoptimization::unpack_frame
3473 // the unloaded class index is in O0 (first parameter to this blob)
3475 // push a dummy "unpack_frame"
3476 // and call Deoptimization::uncommon_trap to pack the compiled frame into
3477 // vframe array and return the UnrollBlock information
3478 __ save_frame(0);
3479 __ set_last_Java_frame(SP, noreg);
3480 __ mov(I0, O2klass_index);
3481 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), G2_thread, O2klass_index);
3482 __ reset_last_Java_frame();
3483 __ mov(O0, O2UnrollBlock->after_save());
3484 __ restore();
3486 // deallocate the deoptimized frame taking care to preserve the return values
3487 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
3488 __ restore();
3490 // Allocate new interpreter frame(s) and possible c2i adapter frame
3492 make_new_frames(masm, false);
3494 // push a dummy "unpack_frame" taking care of float return values and
3495 // call Deoptimization::unpack_frames to have the unpacker layout
3496 // information in the interpreter frames just created and then return
3497 // to the interpreter entry point
3498 __ save_frame(0);
3499 __ set_last_Java_frame(SP, noreg);
3500 __ mov(Deoptimization::Unpack_uncommon_trap, O3); // indicate it is the uncommon trap case
3501 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O3);
3502 __ reset_last_Java_frame();
3503 __ ret();
3504 __ delayed()->restore();
3506 masm->flush();
3507 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, NULL, __ total_frame_size_in_bytes(0)/wordSize);
3508 }
3510 #endif // COMPILER2
3512 //------------------------------generate_handler_blob-------------------
3513 //
3514 // Generate a special Compile2Runtime blob that saves all registers, and sets
3515 // up an OopMap.
3516 //
3517 // This blob is jumped to (via a breakpoint and the signal handler) from a
3518 // safepoint in compiled code. On entry to this blob, O7 contains the
3519 // address in the original nmethod at which we should resume normal execution.
3520 // Thus, this blob looks like a subroutine which must preserve lots of
3521 // registers and return normally. Note that O7 is never register-allocated,
3522 // so it is guaranteed to be free here.
3523 //
3525 // The hardest part of what this blob must do is to save the 64-bit %o
3526 // registers in the 32-bit build. A simple 'save' turn the %o's to %i's and
3527 // an interrupt will chop off their heads. Making space in the caller's frame
3528 // first will let us save the 64-bit %o's before save'ing, but we cannot hand
3529 // the adjusted FP off to the GC stack-crawler: this will modify the caller's
3530 // SP and mess up HIS OopMaps. So we first adjust the caller's SP, then save
3531 // the 64-bit %o's, then do a save, then fixup the caller's SP (our FP).
3532 // Tricky, tricky, tricky...
3534 static SafepointBlob* generate_handler_blob(address call_ptr, bool cause_return) {
3535 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3537 // allocate space for the code
3538 ResourceMark rm;
3539 // setup code generation tools
3540 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
3541 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
3542 // even larger with TraceJumps
3543 int pad = TraceJumps ? 512 : 0;
3544 CodeBuffer buffer("handler_blob", 1600 + pad, 512);
3545 MacroAssembler* masm = new MacroAssembler(&buffer);
3546 int frame_size_words;
3547 OopMapSet *oop_maps = new OopMapSet();
3548 OopMap* map = NULL;
3550 int start = __ offset();
3552 // If this causes a return before the processing, then do a "restore"
3553 if (cause_return) {
3554 __ restore();
3555 } else {
3556 // Make it look like we were called via the poll
3557 // so that frame constructor always sees a valid return address
3558 __ ld_ptr(G2_thread, in_bytes(JavaThread::saved_exception_pc_offset()), O7);
3559 __ sub(O7, frame::pc_return_offset, O7);
3560 }
3562 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3564 // setup last_Java_sp (blows G4)
3565 __ set_last_Java_frame(SP, noreg);
3567 // call into the runtime to handle illegal instructions exception
3568 // Do not use call_VM_leaf, because we need to make a GC map at this call site.
3569 __ mov(G2_thread, O0);
3570 __ save_thread(L7_thread_cache);
3571 __ call(call_ptr);
3572 __ delayed()->nop();
3574 // Set an oopmap for the call site.
3575 // We need this not only for callee-saved registers, but also for volatile
3576 // registers that the compiler might be keeping live across a safepoint.
3578 oop_maps->add_gc_map( __ offset() - start, map);
3580 __ restore_thread(L7_thread_cache);
3581 // clear last_Java_sp
3582 __ reset_last_Java_frame();
3584 // Check for exceptions
3585 Label pending;
3587 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1);
3588 __ tst(O1);
3589 __ brx(Assembler::notEqual, true, Assembler::pn, pending);
3590 __ delayed()->nop();
3592 RegisterSaver::restore_live_registers(masm);
3594 // We are back the the original state on entry and ready to go.
3596 __ retl();
3597 __ delayed()->nop();
3599 // Pending exception after the safepoint
3601 __ bind(pending);
3603 RegisterSaver::restore_live_registers(masm);
3605 // We are back the the original state on entry.
3607 // Tail-call forward_exception_entry, with the issuing PC in O7,
3608 // so it looks like the original nmethod called forward_exception_entry.
3609 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0);
3610 __ JMP(O0, 0);
3611 __ delayed()->nop();
3613 // -------------
3614 // make sure all code is generated
3615 masm->flush();
3617 // return exception blob
3618 return SafepointBlob::create(&buffer, oop_maps, frame_size_words);
3619 }
3621 //
3622 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3623 //
3624 // Generate a stub that calls into vm to find out the proper destination
3625 // of a java call. All the argument registers are live at this point
3626 // but since this is generic code we don't know what they are and the caller
3627 // must do any gc of the args.
3628 //
3629 static RuntimeStub* generate_resolve_blob(address destination, const char* name) {
3630 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3632 // allocate space for the code
3633 ResourceMark rm;
3634 // setup code generation tools
3635 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
3636 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
3637 // even larger with TraceJumps
3638 int pad = TraceJumps ? 512 : 0;
3639 CodeBuffer buffer(name, 1600 + pad, 512);
3640 MacroAssembler* masm = new MacroAssembler(&buffer);
3641 int frame_size_words;
3642 OopMapSet *oop_maps = new OopMapSet();
3643 OopMap* map = NULL;
3645 int start = __ offset();
3647 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3649 int frame_complete = __ offset();
3651 // setup last_Java_sp (blows G4)
3652 __ set_last_Java_frame(SP, noreg);
3654 // call into the runtime to handle illegal instructions exception
3655 // Do not use call_VM_leaf, because we need to make a GC map at this call site.
3656 __ mov(G2_thread, O0);
3657 __ save_thread(L7_thread_cache);
3658 __ call(destination, relocInfo::runtime_call_type);
3659 __ delayed()->nop();
3661 // O0 contains the address we are going to jump to assuming no exception got installed
3663 // Set an oopmap for the call site.
3664 // We need this not only for callee-saved registers, but also for volatile
3665 // registers that the compiler might be keeping live across a safepoint.
3667 oop_maps->add_gc_map( __ offset() - start, map);
3669 __ restore_thread(L7_thread_cache);
3670 // clear last_Java_sp
3671 __ reset_last_Java_frame();
3673 // Check for exceptions
3674 Label pending;
3676 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1);
3677 __ tst(O1);
3678 __ brx(Assembler::notEqual, true, Assembler::pn, pending);
3679 __ delayed()->nop();
3681 // get the returned methodOop
3683 __ get_vm_result(G5_method);
3684 __ stx(G5_method, SP, RegisterSaver::G5_offset()+STACK_BIAS);
3686 // O0 is where we want to jump, overwrite G3 which is saved and scratch
3688 __ stx(O0, SP, RegisterSaver::G3_offset()+STACK_BIAS);
3690 RegisterSaver::restore_live_registers(masm);
3692 // We are back the the original state on entry and ready to go.
3694 __ JMP(G3, 0);
3695 __ delayed()->nop();
3697 // Pending exception after the safepoint
3699 __ bind(pending);
3701 RegisterSaver::restore_live_registers(masm);
3703 // We are back the the original state on entry.
3705 // Tail-call forward_exception_entry, with the issuing PC in O7,
3706 // so it looks like the original nmethod called forward_exception_entry.
3707 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0);
3708 __ JMP(O0, 0);
3709 __ delayed()->nop();
3711 // -------------
3712 // make sure all code is generated
3713 masm->flush();
3715 // return the blob
3716 // frame_size_words or bytes??
3717 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true);
3718 }
3720 void SharedRuntime::generate_stubs() {
3722 _wrong_method_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::handle_wrong_method),
3723 "wrong_method_stub");
3725 _ic_miss_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::handle_wrong_method_ic_miss),
3726 "ic_miss_stub");
3728 _resolve_opt_virtual_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_opt_virtual_call_C),
3729 "resolve_opt_virtual_call");
3731 _resolve_virtual_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_virtual_call_C),
3732 "resolve_virtual_call");
3734 _resolve_static_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_static_call_C),
3735 "resolve_static_call");
3737 _polling_page_safepoint_handler_blob =
3738 generate_handler_blob(CAST_FROM_FN_PTR(address,
3739 SafepointSynchronize::handle_polling_page_exception), false);
3741 _polling_page_return_handler_blob =
3742 generate_handler_blob(CAST_FROM_FN_PTR(address,
3743 SafepointSynchronize::handle_polling_page_exception), true);
3745 generate_deopt_blob();
3747 #ifdef COMPILER2
3748 generate_uncommon_trap_blob();
3749 #endif // COMPILER2
3750 }