1 /*
2 * Copyright 1997-2008 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 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_templateInterpreter_sparc.cpp.incl"
27
28 #ifndef CC_INTERP
29 #ifndef FAST_DISPATCH
30 #define FAST_DISPATCH 1
31 #endif
32 #undef FAST_DISPATCH
33
34
35 // Generation of Interpreter
36 //
37 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
38
39
40 #define __ _masm->
41
42
43 //----------------------------------------------------------------------------------------------------
44
45
46 void InterpreterGenerator::save_native_result(void) {
47 // result potentially in O0/O1: save it across calls
48 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
49
50 // result potentially in F0/F1: save it across calls
51 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
52
53 // save and restore any potential method result value around the unlocking operation
54 __ stf(FloatRegisterImpl::D, F0, d_tmp);
55 #ifdef _LP64
56 __ stx(O0, l_tmp);
57 #else
58 __ std(O0, l_tmp);
59 #endif
60 }
61
62 void InterpreterGenerator::restore_native_result(void) {
63 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
64 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
65
66 // Restore any method result value
67 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
68 #ifdef _LP64
69 __ ldx(l_tmp, O0);
70 #else
71 __ ldd(l_tmp, O0);
72 #endif
73 }
74
75 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
76 assert(!pass_oop || message == NULL, "either oop or message but not both");
77 address entry = __ pc();
78 // expression stack must be empty before entering the VM if an exception happened
79 __ empty_expression_stack();
80 // load exception object
81 __ set((intptr_t)name, G3_scratch);
82 if (pass_oop) {
83 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
84 } else {
85 __ set((intptr_t)message, G4_scratch);
86 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
87 }
88 // throw exception
89 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
90 Address thrower(G3_scratch, Interpreter::throw_exception_entry());
91 __ jump_to (thrower);
92 __ delayed()->nop();
93 return entry;
94 }
95
96 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
97 address entry = __ pc();
98 // expression stack must be empty before entering the VM if an exception
99 // happened
100 __ empty_expression_stack();
101 // load exception object
102 __ call_VM(Oexception,
103 CAST_FROM_FN_PTR(address,
104 InterpreterRuntime::throw_ClassCastException),
105 Otos_i);
106 __ should_not_reach_here();
107 return entry;
108 }
109
110
111 #ifdef ASSERT
112 address last_WrongMethodType_caller;
113 #endif //ASSERT
114
115 // Arguments are: required type in G5_method_type, and
116 // failing object (or NULL) in G3_method_handle.
117 // In the debug build, the caller should put his own PC in G1.
118 address TemplateInterpreterGenerator::generate_WrongMethodType_handler() {
119 address entry = __ pc();
120 #ifdef ASSERT
121 Address last_caller_addr(O3, (address)&last_WrongMethodType_caller);
122 __ sethi(last_caller_addr);
123 __ st_ptr(G1, last_caller_addr);
124 #endif //ASSERT
125 // expression stack must be empty before entering the VM if an exception
126 // happened
127 __ empty_expression_stack();
128 // load exception object
129 __ call_VM(Oexception,
130 CAST_FROM_FN_PTR(address,
131 InterpreterRuntime::throw_WrongMethodTypeException),
132 G5_method_type, // required
133 G3_method_handle); // actual
134 __ should_not_reach_here();
135 return entry;
136 }
137
138
139 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
140 address entry = __ pc();
141 // expression stack must be empty before entering the VM if an exception happened
142 __ empty_expression_stack();
143 // convention: expect aberrant index in register G3_scratch, then shuffle the
144 // index to G4_scratch for the VM call
145 __ mov(G3_scratch, G4_scratch);
146 __ set((intptr_t)name, G3_scratch);
147 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
148 __ should_not_reach_here();
149 return entry;
150 }
151
152
153 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
154 address entry = __ pc();
155 // expression stack must be empty before entering the VM if an exception happened
156 __ empty_expression_stack();
157 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
158 __ should_not_reach_here();
159 return entry;
160 }
161
162
163 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) {
164 address compiled_entry = __ pc();
165 Label cont;
166
167 address entry = __ pc();
168 #if !defined(_LP64) && defined(COMPILER2)
169 // All return values are where we want them, except for Longs. C2 returns
170 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
171 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
172 // build even if we are returning from interpreted we just do a little
173 // stupid shuffing.
174 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
175 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
176 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
177
178 if( state == ltos ) {
179 __ srl (G1, 0,O1);
180 __ srlx(G1,32,O0);
181 }
182 #endif /* !_LP64 && COMPILER2 */
183
184
185 __ bind(cont);
186
187 // The callee returns with the stack possibly adjusted by adapter transition
188 // We remove that possible adjustment here.
189 // All interpreter local registers are untouched. Any result is passed back
190 // in the O0/O1 or float registers. Before continuing, the arguments must be
191 // popped from the java expression stack; i.e., Lesp must be adjusted.
192
193 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
194
195
196 const Register cache = G3_scratch;
197 const Register size = G1_scratch;
198 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1);
199 __ ld_ptr(Address(cache, 0, in_bytes(constantPoolCacheOopDesc::base_offset()) +
200 in_bytes(ConstantPoolCacheEntry::flags_offset())), size);
201 __ and3(size, 0xFF, size); // argument size in words
202 __ sll(size, Interpreter::logStackElementSize(), size); // each argument size in bytes
203 __ add(Lesp, size, Lesp); // pop arguments
204 __ dispatch_next(state, step);
205
206 return entry;
207 }
208
209
210 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
211 address entry = __ pc();
212 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
213 { Label L;
214 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
215
216 __ ld_ptr(exception_addr, Gtemp);
217 __ tst(Gtemp);
218 __ brx(Assembler::equal, false, Assembler::pt, L);
219 __ delayed()->nop();
220 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
221 __ should_not_reach_here();
222 __ bind(L);
223 }
224 __ dispatch_next(state, step);
225 return entry;
226 }
227
228 // A result handler converts/unboxes a native call result into
229 // a java interpreter/compiler result. The current frame is an
230 // interpreter frame. The activation frame unwind code must be
231 // consistent with that of TemplateTable::_return(...). In the
232 // case of native methods, the caller's SP was not modified.
233 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
234 address entry = __ pc();
235 Register Itos_i = Otos_i ->after_save();
236 Register Itos_l = Otos_l ->after_save();
237 Register Itos_l1 = Otos_l1->after_save();
238 Register Itos_l2 = Otos_l2->after_save();
239 switch (type) {
240 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
241 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
242 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
243 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
244 case T_LONG :
245 #ifndef _LP64
246 __ mov(O1, Itos_l2); // move other half of long
247 #endif // ifdef or no ifdef, fall through to the T_INT case
248 case T_INT : __ mov(O0, Itos_i); break;
249 case T_VOID : /* nothing to do */ break;
250 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
251 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
252 case T_OBJECT :
253 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
254 __ verify_oop(Itos_i);
255 break;
256 default : ShouldNotReachHere();
257 }
258 __ ret(); // return from interpreter activation
259 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
260 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
261 return entry;
262 }
263
264 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
265 address entry = __ pc();
266 __ push(state);
267 __ call_VM(noreg, runtime_entry);
268 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
269 return entry;
270 }
271
272
273 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
274 address entry = __ pc();
275 __ dispatch_next(state);
276 return entry;
277 }
278
279 //
280 // Helpers for commoning out cases in the various type of method entries.
281 //
282
283 // increment invocation count & check for overflow
284 //
285 // Note: checking for negative value instead of overflow
286 // so we have a 'sticky' overflow test
287 //
288 // Lmethod: method
289 // ??: invocation counter
290 //
291 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
292 // Update standard invocation counters
293 __ increment_invocation_counter(O0, G3_scratch);
294 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
295 Address interpreter_invocation_counter(Lmethod, 0, in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
296 __ ld(interpreter_invocation_counter, G3_scratch);
297 __ inc(G3_scratch);
298 __ st(G3_scratch, interpreter_invocation_counter);
299 }
300
301 if (ProfileInterpreter && profile_method != NULL) {
302 // Test to see if we should create a method data oop
303 Address profile_limit(G3_scratch, (address)&InvocationCounter::InterpreterProfileLimit);
304 __ sethi(profile_limit);
305 __ ld(profile_limit, G3_scratch);
306 __ cmp(O0, G3_scratch);
307 __ br(Assembler::lessUnsigned, false, Assembler::pn, *profile_method_continue);
308 __ delayed()->nop();
309
310 // if no method data exists, go to profile_method
311 __ test_method_data_pointer(*profile_method);
312 }
313
314 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
315 __ sethi(invocation_limit);
316 __ ld(invocation_limit, G3_scratch);
317 __ cmp(O0, G3_scratch);
318 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
319 __ delayed()->nop();
320
321 }
322
323 // Allocate monitor and lock method (asm interpreter)
324 // ebx - methodOop
325 //
326 void InterpreterGenerator::lock_method(void) {
327 const Address access_flags (Lmethod, 0, in_bytes(methodOopDesc::access_flags_offset()));
328 __ ld(access_flags, O0);
329
330 #ifdef ASSERT
331 { Label ok;
332 __ btst(JVM_ACC_SYNCHRONIZED, O0);
333 __ br( Assembler::notZero, false, Assembler::pt, ok);
334 __ delayed()->nop();
335 __ stop("method doesn't need synchronization");
336 __ bind(ok);
337 }
338 #endif // ASSERT
339
340 // get synchronization object to O0
341 { Label done;
342 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
343 __ btst(JVM_ACC_STATIC, O0);
344 __ br( Assembler::zero, true, Assembler::pt, done);
345 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
346
347 __ ld_ptr( Lmethod, in_bytes(methodOopDesc::constants_offset()), O0);
348 __ ld_ptr( O0, constantPoolOopDesc::pool_holder_offset_in_bytes(), O0);
349
350 // lock the mirror, not the klassOop
351 __ ld_ptr( O0, mirror_offset, O0);
352
353 #ifdef ASSERT
354 __ tst(O0);
355 __ breakpoint_trap(Assembler::zero);
356 #endif // ASSERT
357
358 __ bind(done);
359 }
360
361 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
362 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
363 // __ untested("lock_object from method entry");
364 __ lock_object(Lmonitors, O0);
365 }
366
367
368 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
369 Register Rscratch,
370 Register Rscratch2) {
371 const int page_size = os::vm_page_size();
372 Address saved_exception_pc(G2_thread, 0,
373 in_bytes(JavaThread::saved_exception_pc_offset()));
374 Label after_frame_check;
375
376 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
377
378 __ set( page_size, Rscratch );
379 __ cmp( Rframe_size, Rscratch );
380
381 __ br( Assembler::lessEqual, false, Assembler::pt, after_frame_check );
382 __ delayed()->nop();
383
384 // get the stack base, and in debug, verify it is non-zero
385 __ ld_ptr( G2_thread, in_bytes(Thread::stack_base_offset()), Rscratch );
386 #ifdef ASSERT
387 Label base_not_zero;
388 __ cmp( Rscratch, G0 );
389 __ brx( Assembler::notEqual, false, Assembler::pn, base_not_zero );
390 __ delayed()->nop();
391 __ stop("stack base is zero in generate_stack_overflow_check");
392 __ bind(base_not_zero);
393 #endif
394
395 // get the stack size, and in debug, verify it is non-zero
396 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
397 __ ld_ptr( G2_thread, in_bytes(Thread::stack_size_offset()), Rscratch2 );
398 #ifdef ASSERT
399 Label size_not_zero;
400 __ cmp( Rscratch2, G0 );
401 __ brx( Assembler::notEqual, false, Assembler::pn, size_not_zero );
402 __ delayed()->nop();
403 __ stop("stack size is zero in generate_stack_overflow_check");
404 __ bind(size_not_zero);
405 #endif
406
407 // compute the beginning of the protected zone minus the requested frame size
408 __ sub( Rscratch, Rscratch2, Rscratch );
409 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
410 __ add( Rscratch, Rscratch2, Rscratch );
411
412 // Add in the size of the frame (which is the same as subtracting it from the
413 // SP, which would take another register
414 __ add( Rscratch, Rframe_size, Rscratch );
415
416 // the frame is greater than one page in size, so check against
417 // the bottom of the stack
418 __ cmp( SP, Rscratch );
419 __ brx( Assembler::greater, false, Assembler::pt, after_frame_check );
420 __ delayed()->nop();
421
422 // Save the return address as the exception pc
423 __ st_ptr(O7, saved_exception_pc);
424
425 // the stack will overflow, throw an exception
426 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
427
428 // if you get to here, then there is enough stack space
429 __ bind( after_frame_check );
430 }
431
432
433 //
434 // Generate a fixed interpreter frame. This is identical setup for interpreted
435 // methods and for native methods hence the shared code.
436
437 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
438 //
439 //
440 // The entry code sets up a new interpreter frame in 4 steps:
441 //
442 // 1) Increase caller's SP by for the extra local space needed:
443 // (check for overflow)
444 // Efficient implementation of xload/xstore bytecodes requires
445 // that arguments and non-argument locals are in a contigously
446 // addressable memory block => non-argument locals must be
447 // allocated in the caller's frame.
448 //
449 // 2) Create a new stack frame and register window:
450 // The new stack frame must provide space for the standard
451 // register save area, the maximum java expression stack size,
452 // the monitor slots (0 slots initially), and some frame local
453 // scratch locations.
454 //
455 // 3) The following interpreter activation registers must be setup:
456 // Lesp : expression stack pointer
457 // Lbcp : bytecode pointer
458 // Lmethod : method
459 // Llocals : locals pointer
460 // Lmonitors : monitor pointer
461 // LcpoolCache: constant pool cache
462 //
463 // 4) Initialize the non-argument locals if necessary:
464 // Non-argument locals may need to be initialized to NULL
465 // for GC to work. If the oop-map information is accurate
466 // (in the absence of the JSR problem), no initialization
467 // is necessary.
468 //
469 // (gri - 2/25/2000)
470
471
472 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
473 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
474 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
475 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
476
477 const int extra_space =
478 rounded_vm_local_words + // frame local scratch space
479 methodOopDesc::extra_stack() + // extra push slot for MH insertion
480 frame::memory_parameter_word_sp_offset + // register save area
481 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
482
483 const Register Glocals_size = G3;
484 const Register Otmp1 = O3;
485 const Register Otmp2 = O4;
486 // Lscratch can't be used as a temporary because the call_stub uses
487 // it to assert that the stack frame was setup correctly.
488
489 __ lduh( size_of_parameters, Glocals_size);
490
491 // Gargs points to first local + BytesPerWord
492 // Set the saved SP after the register window save
493 //
494 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
495 __ sll(Glocals_size, Interpreter::logStackElementSize(), Otmp1);
496 __ add(Gargs, Otmp1, Gargs);
497
498 if (native_call) {
499 __ calc_mem_param_words( Glocals_size, Gframe_size );
500 __ add( Gframe_size, extra_space, Gframe_size);
501 __ round_to( Gframe_size, WordsPerLong );
502 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
503 } else {
504
505 //
506 // Compute number of locals in method apart from incoming parameters
507 //
508 __ lduh( size_of_locals, Otmp1 );
509 __ sub( Otmp1, Glocals_size, Glocals_size );
510 __ round_to( Glocals_size, WordsPerLong );
511 __ sll( Glocals_size, Interpreter::logStackElementSize(), Glocals_size );
512
513 // see if the frame is greater than one page in size. If so,
514 // then we need to verify there is enough stack space remaining
515 // Frame_size = (max_stack + extra_space) * BytesPerWord;
516 __ lduh( max_stack, Gframe_size );
517 __ add( Gframe_size, extra_space, Gframe_size );
518 __ round_to( Gframe_size, WordsPerLong );
519 __ sll( Gframe_size, Interpreter::logStackElementSize(), Gframe_size);
520
521 // Add in java locals size for stack overflow check only
522 __ add( Gframe_size, Glocals_size, Gframe_size );
523
524 const Register Otmp2 = O4;
525 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
526 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
527
528 __ sub( Gframe_size, Glocals_size, Gframe_size);
529
530 //
531 // bump SP to accomodate the extra locals
532 //
533 __ sub( SP, Glocals_size, SP );
534 }
535
536 //
537 // now set up a stack frame with the size computed above
538 //
539 __ neg( Gframe_size );
540 __ save( SP, Gframe_size, SP );
541
542 //
543 // now set up all the local cache registers
544 //
545 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
546 // that all present references to Lbyte_code initialize the register
547 // immediately before use
548 if (native_call) {
549 __ mov(G0, Lbcp);
550 } else {
551 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), Lbcp );
552 __ add(Address(Lbcp, 0, in_bytes(constMethodOopDesc::codes_offset())), Lbcp );
553 }
554 __ mov( G5_method, Lmethod); // set Lmethod
555 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
556 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
557 #ifdef _LP64
558 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
559 #endif
560 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
561
562 // setup interpreter activation registers
563 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
564
565 if (ProfileInterpreter) {
566 #ifdef FAST_DISPATCH
567 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
568 // they both use I2.
569 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
570 #endif // FAST_DISPATCH
571 __ set_method_data_pointer();
572 }
573
574 }
575
576 // Empty method, generate a very fast return.
577
578 address InterpreterGenerator::generate_empty_entry(void) {
579
580 // A method that does nother but return...
581
582 address entry = __ pc();
583 Label slow_path;
584
585 __ verify_oop(G5_method);
586
587 // do nothing for empty methods (do not even increment invocation counter)
588 if ( UseFastEmptyMethods) {
589 // If we need a safepoint check, generate full interpreter entry.
590 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
591 __ load_contents(sync_state, G3_scratch);
592 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
593 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
594 __ delayed()->nop();
595
596 // Code: _return
597 __ retl();
598 __ delayed()->mov(O5_savedSP, SP);
599
600 __ bind(slow_path);
601 (void) generate_normal_entry(false);
602
603 return entry;
604 }
605 return NULL;
606 }
607
608 // Call an accessor method (assuming it is resolved, otherwise drop into
609 // vanilla (slow path) entry
610
611 // Generates code to elide accessor methods
612 // Uses G3_scratch and G1_scratch as scratch
613 address InterpreterGenerator::generate_accessor_entry(void) {
614
615 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
616 // parameter size = 1
617 // Note: We can only use this code if the getfield has been resolved
618 // and if we don't have a null-pointer exception => check for
619 // these conditions first and use slow path if necessary.
620 address entry = __ pc();
621 Label slow_path;
622
623
624 // XXX: for compressed oops pointer loading and decoding doesn't fit in
625 // delay slot and damages G1
626 if ( UseFastAccessorMethods && !UseCompressedOops ) {
627 // Check if we need to reach a safepoint and generate full interpreter
628 // frame if so.
629 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
630 __ load_contents(sync_state, G3_scratch);
631 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
632 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
633 __ delayed()->nop();
634
635 // Check if local 0 != NULL
636 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
637 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
638 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
639 __ delayed()->nop();
640
641
642 // read first instruction word and extract bytecode @ 1 and index @ 2
643 // get first 4 bytes of the bytecodes (big endian!)
644 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), G1_scratch);
645 __ ld(Address(G1_scratch, 0, in_bytes(constMethodOopDesc::codes_offset())), G1_scratch);
646
647 // move index @ 2 far left then to the right most two bytes.
648 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
649 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
650 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
651
652 // get constant pool cache
653 __ ld_ptr(G5_method, in_bytes(methodOopDesc::constants_offset()), G3_scratch);
654 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
655
656 // get specific constant pool cache entry
657 __ add(G3_scratch, G1_scratch, G3_scratch);
658
659 // Check the constant Pool cache entry to see if it has been resolved.
660 // If not, need the slow path.
661 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
662 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
663 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
664 __ and3(G1_scratch, 0xFF, G1_scratch);
665 __ cmp(G1_scratch, Bytecodes::_getfield);
666 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
667 __ delayed()->nop();
668
669 // Get the type and return field offset from the constant pool cache
670 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
671 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
672
673 Label xreturn_path;
674 // Need to differentiate between igetfield, agetfield, bgetfield etc.
675 // because they are different sizes.
676 // Get the type from the constant pool cache
677 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
678 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
679 ConstantPoolCacheEntry::verify_tosBits();
680 __ cmp(G1_scratch, atos );
681 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
682 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
683 __ cmp(G1_scratch, itos);
684 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
685 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
686 __ cmp(G1_scratch, stos);
687 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
688 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
689 __ cmp(G1_scratch, ctos);
690 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
691 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
692 #ifdef ASSERT
693 __ cmp(G1_scratch, btos);
694 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
695 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
696 __ should_not_reach_here();
697 #endif
698 __ ldsb(Otos_i, G3_scratch, Otos_i);
699 __ bind(xreturn_path);
700
701 // _ireturn/_areturn
702 __ retl(); // return from leaf routine
703 __ delayed()->mov(O5_savedSP, SP);
704
705 // Generate regular method entry
706 __ bind(slow_path);
707 (void) generate_normal_entry(false);
708 return entry;
709 }
710 return NULL;
711 }
712
713 //
714 // Interpreter stub for calling a native method. (asm interpreter)
715 // This sets up a somewhat different looking stack for calling the native method
716 // than the typical interpreter frame setup.
717 //
718
719 address InterpreterGenerator::generate_native_entry(bool synchronized) {
720 address entry = __ pc();
721
722 // the following temporary registers are used during frame creation
723 const Register Gtmp1 = G3_scratch ;
724 const Register Gtmp2 = G1_scratch;
725 bool inc_counter = UseCompiler || CountCompiledCalls;
726
727 // make sure registers are different!
728 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
729
730 const Address Laccess_flags (Lmethod, 0, in_bytes(methodOopDesc::access_flags_offset()));
731
732 __ verify_oop(G5_method);
733
734 const Register Glocals_size = G3;
735 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
736
737 // make sure method is native & not abstract
738 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
739 #ifdef ASSERT
740 __ ld(G5_method, in_bytes(methodOopDesc::access_flags_offset()), Gtmp1);
741 {
742 Label L;
743 __ btst(JVM_ACC_NATIVE, Gtmp1);
744 __ br(Assembler::notZero, false, Assembler::pt, L);
745 __ delayed()->nop();
746 __ stop("tried to execute non-native method as native");
747 __ bind(L);
748 }
749 { Label L;
750 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
751 __ br(Assembler::zero, false, Assembler::pt, L);
752 __ delayed()->nop();
753 __ stop("tried to execute abstract method as non-abstract");
754 __ bind(L);
755 }
756 #endif // ASSERT
757
758 // generate the code to allocate the interpreter stack frame
759 generate_fixed_frame(true);
760
761 //
762 // No locals to initialize for native method
763 //
764
765 // this slot will be set later, we initialize it to null here just in
766 // case we get a GC before the actual value is stored later
767 __ st_ptr(G0, Address(FP, 0, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS));
768
769 const Address do_not_unlock_if_synchronized(G2_thread, 0,
770 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
771 // Since at this point in the method invocation the exception handler
772 // would try to exit the monitor of synchronized methods which hasn't
773 // been entered yet, we set the thread local variable
774 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
775 // runtime, exception handling i.e. unlock_if_synchronized_method will
776 // check this thread local flag.
777 // This flag has two effects, one is to force an unwind in the topmost
778 // interpreter frame and not perform an unlock while doing so.
779
780 __ movbool(true, G3_scratch);
781 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
782
783 // increment invocation counter and check for overflow
784 //
785 // Note: checking for negative value instead of overflow
786 // so we have a 'sticky' overflow test (may be of
787 // importance as soon as we have true MT/MP)
788 Label invocation_counter_overflow;
789 Label Lcontinue;
790 if (inc_counter) {
791 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
792
793 }
794 __ bind(Lcontinue);
795
796 bang_stack_shadow_pages(true);
797
798 // reset the _do_not_unlock_if_synchronized flag
799 __ stbool(G0, do_not_unlock_if_synchronized);
800
801 // check for synchronized methods
802 // Must happen AFTER invocation_counter check and stack overflow check,
803 // so method is not locked if overflows.
804
805 if (synchronized) {
806 lock_method();
807 } else {
808 #ifdef ASSERT
809 { Label ok;
810 __ ld(Laccess_flags, O0);
811 __ btst(JVM_ACC_SYNCHRONIZED, O0);
812 __ br( Assembler::zero, false, Assembler::pt, ok);
813 __ delayed()->nop();
814 __ stop("method needs synchronization");
815 __ bind(ok);
816 }
817 #endif // ASSERT
818 }
819
820
821 // start execution
822 __ verify_thread();
823
824 // JVMTI support
825 __ notify_method_entry();
826
827 // native call
828
829 // (note that O0 is never an oop--at most it is a handle)
830 // It is important not to smash any handles created by this call,
831 // until any oop handle in O0 is dereferenced.
832
833 // (note that the space for outgoing params is preallocated)
834
835 // get signature handler
836 { Label L;
837 __ ld_ptr(Address(Lmethod, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
838 __ tst(G3_scratch);
839 __ brx(Assembler::notZero, false, Assembler::pt, L);
840 __ delayed()->nop();
841 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
842 __ ld_ptr(Address(Lmethod, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
843 __ bind(L);
844 }
845
846 // Push a new frame so that the args will really be stored in
847 // Copy a few locals across so the new frame has the variables
848 // we need but these values will be dead at the jni call and
849 // therefore not gc volatile like the values in the current
850 // frame (Lmethod in particular)
851
852 // Flush the method pointer to the register save area
853 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
854 __ mov(Llocals, O1);
855 // calculate where the mirror handle body is allocated in the interpreter frame:
856
857 Address mirror(FP, 0, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
858 __ add(mirror, O2);
859
860 // Calculate current frame size
861 __ sub(SP, FP, O3); // Calculate negative of current frame size
862 __ save(SP, O3, SP); // Allocate an identical sized frame
863
864 // Note I7 has leftover trash. Slow signature handler will fill it in
865 // should we get there. Normal jni call will set reasonable last_Java_pc
866 // below (and fix I7 so the stack trace doesn't have a meaningless frame
867 // in it).
868
869 // Load interpreter frame's Lmethod into same register here
870
871 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
872
873 __ mov(I1, Llocals);
874 __ mov(I2, Lscratch2); // save the address of the mirror
875
876
877 // ONLY Lmethod and Llocals are valid here!
878
879 // call signature handler, It will move the arg properly since Llocals in current frame
880 // matches that in outer frame
881
882 __ callr(G3_scratch, 0);
883 __ delayed()->nop();
884
885 // Result handler is in Lscratch
886
887 // Reload interpreter frame's Lmethod since slow signature handler may block
888 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
889
890 { Label not_static;
891
892 __ ld(Laccess_flags, O0);
893 __ btst(JVM_ACC_STATIC, O0);
894 __ br( Assembler::zero, false, Assembler::pt, not_static);
895 __ delayed()->
896 // get native function entry point(O0 is a good temp until the very end)
897 ld_ptr(Address(Lmethod, 0, in_bytes(methodOopDesc::native_function_offset())), O0);
898 // for static methods insert the mirror argument
899 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
900
901 __ ld_ptr(Address(Lmethod, 0, in_bytes(methodOopDesc:: constants_offset())), O1);
902 __ ld_ptr(Address(O1, 0, constantPoolOopDesc::pool_holder_offset_in_bytes()), O1);
903 __ ld_ptr(O1, mirror_offset, O1);
904 #ifdef ASSERT
905 if (!PrintSignatureHandlers) // do not dirty the output with this
906 { Label L;
907 __ tst(O1);
908 __ brx(Assembler::notZero, false, Assembler::pt, L);
909 __ delayed()->nop();
910 __ stop("mirror is missing");
911 __ bind(L);
912 }
913 #endif // ASSERT
914 __ st_ptr(O1, Lscratch2, 0);
915 __ mov(Lscratch2, O1);
916 __ bind(not_static);
917 }
918
919 // At this point, arguments have been copied off of stack into
920 // their JNI positions, which are O1..O5 and SP[68..].
921 // Oops are boxed in-place on the stack, with handles copied to arguments.
922 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
923
924 #ifdef ASSERT
925 { Label L;
926 __ tst(O0);
927 __ brx(Assembler::notZero, false, Assembler::pt, L);
928 __ delayed()->nop();
929 __ stop("native entry point is missing");
930 __ bind(L);
931 }
932 #endif // ASSERT
933
934 //
935 // setup the frame anchor
936 //
937 // The scavenge function only needs to know that the PC of this frame is
938 // in the interpreter method entry code, it doesn't need to know the exact
939 // PC and hence we can use O7 which points to the return address from the
940 // previous call in the code stream (signature handler function)
941 //
942 // The other trick is we set last_Java_sp to FP instead of the usual SP because
943 // we have pushed the extra frame in order to protect the volatile register(s)
944 // in that frame when we return from the jni call
945 //
946
947 __ set_last_Java_frame(FP, O7);
948 __ mov(O7, I7); // make dummy interpreter frame look like one above,
949 // not meaningless information that'll confuse me.
950
951 // flush the windows now. We don't care about the current (protection) frame
952 // only the outer frames
953
954 __ flush_windows();
955
956 // mark windows as flushed
957 Address flags(G2_thread,
958 0,
959 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
960 __ set(JavaFrameAnchor::flushed, G3_scratch);
961 __ st(G3_scratch, flags);
962
963 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
964
965 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
966 #ifdef ASSERT
967 { Label L;
968 __ ld(thread_state, G3_scratch);
969 __ cmp(G3_scratch, _thread_in_Java);
970 __ br(Assembler::equal, false, Assembler::pt, L);
971 __ delayed()->nop();
972 __ stop("Wrong thread state in native stub");
973 __ bind(L);
974 }
975 #endif // ASSERT
976 __ set(_thread_in_native, G3_scratch);
977 __ st(G3_scratch, thread_state);
978
979 // Call the jni method, using the delay slot to set the JNIEnv* argument.
980 __ save_thread(L7_thread_cache); // save Gthread
981 __ callr(O0, 0);
982 __ delayed()->
983 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
984
985 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
986
987 __ restore_thread(L7_thread_cache); // restore G2_thread
988 __ reinit_heapbase();
989
990 // must we block?
991
992 // Block, if necessary, before resuming in _thread_in_Java state.
993 // In order for GC to work, don't clear the last_Java_sp until after blocking.
994 { Label no_block;
995 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
996
997 // Switch thread to "native transition" state before reading the synchronization state.
998 // This additional state is necessary because reading and testing the synchronization
999 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1000 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1001 // VM thread changes sync state to synchronizing and suspends threads for GC.
1002 // Thread A is resumed to finish this native method, but doesn't block here since it
1003 // didn't see any synchronization is progress, and escapes.
1004 __ set(_thread_in_native_trans, G3_scratch);
1005 __ st(G3_scratch, thread_state);
1006 if(os::is_MP()) {
1007 if (UseMembar) {
1008 // Force this write out before the read below
1009 __ membar(Assembler::StoreLoad);
1010 } else {
1011 // Write serialization page so VM thread can do a pseudo remote membar.
1012 // We use the current thread pointer to calculate a thread specific
1013 // offset to write to within the page. This minimizes bus traffic
1014 // due to cache line collision.
1015 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1016 }
1017 }
1018 __ load_contents(sync_state, G3_scratch);
1019 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1020
1021 Label L;
1022 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
1023 __ br(Assembler::notEqual, false, Assembler::pn, L);
1024 __ delayed()->
1025 ld(suspend_state, G3_scratch);
1026 __ cmp(G3_scratch, 0);
1027 __ br(Assembler::equal, false, Assembler::pt, no_block);
1028 __ delayed()->nop();
1029 __ bind(L);
1030
1031 // Block. Save any potential method result value before the operation and
1032 // use a leaf call to leave the last_Java_frame setup undisturbed.
1033 save_native_result();
1034 __ call_VM_leaf(L7_thread_cache,
1035 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1036 G2_thread);
1037
1038 // Restore any method result value
1039 restore_native_result();
1040 __ bind(no_block);
1041 }
1042
1043 // Clear the frame anchor now
1044
1045 __ reset_last_Java_frame();
1046
1047 // Move the result handler address
1048 __ mov(Lscratch, G3_scratch);
1049 // return possible result to the outer frame
1050 #ifndef __LP64
1051 __ mov(O0, I0);
1052 __ restore(O1, G0, O1);
1053 #else
1054 __ restore(O0, G0, O0);
1055 #endif /* __LP64 */
1056
1057 // Move result handler to expected register
1058 __ mov(G3_scratch, Lscratch);
1059
1060 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1061 // switch to thread_in_Java.
1062
1063 __ set(_thread_in_Java, G3_scratch);
1064 __ st(G3_scratch, thread_state);
1065
1066 // reset handle block
1067 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
1068 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1069
1070 // If we have an oop result store it where it will be safe for any further gc
1071 // until we return now that we've released the handle it might be protected by
1072
1073 {
1074 Label no_oop, store_result;
1075
1076 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1077 __ cmp(G3_scratch, Lscratch);
1078 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
1079 __ delayed()->nop();
1080 __ addcc(G0, O0, O0);
1081 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1082 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1083 __ mov(G0, O0);
1084
1085 __ bind(store_result);
1086 // Store it where gc will look for it and result handler expects it.
1087 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1088
1089 __ bind(no_oop);
1090
1091 }
1092
1093
1094 // handle exceptions (exception handling will handle unlocking!)
1095 { Label L;
1096 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1097
1098 __ ld_ptr(exception_addr, Gtemp);
1099 __ tst(Gtemp);
1100 __ brx(Assembler::equal, false, Assembler::pt, L);
1101 __ delayed()->nop();
1102 // Note: This could be handled more efficiently since we know that the native
1103 // method doesn't have an exception handler. We could directly return
1104 // to the exception handler for the caller.
1105 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1106 __ should_not_reach_here();
1107 __ bind(L);
1108 }
1109
1110 // JVMTI support (preserves thread register)
1111 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1112
1113 if (synchronized) {
1114 // save and restore any potential method result value around the unlocking operation
1115 save_native_result();
1116
1117 __ add( __ top_most_monitor(), O1);
1118 __ unlock_object(O1);
1119
1120 restore_native_result();
1121 }
1122
1123 #if defined(COMPILER2) && !defined(_LP64)
1124
1125 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1126 // or compiled so just be safe.
1127
1128 __ sllx(O0, 32, G1); // Shift bits into high G1
1129 __ srl (O1, 0, O1); // Zero extend O1
1130 __ or3 (O1, G1, G1); // OR 64 bits into G1
1131
1132 #endif /* COMPILER2 && !_LP64 */
1133
1134 // dispose of return address and remove activation
1135 #ifdef ASSERT
1136 {
1137 Label ok;
1138 __ cmp(I5_savedSP, FP);
1139 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1140 __ delayed()->nop();
1141 __ stop("bad I5_savedSP value");
1142 __ should_not_reach_here();
1143 __ bind(ok);
1144 }
1145 #endif
1146 if (TraceJumps) {
1147 // Move target to register that is recordable
1148 __ mov(Lscratch, G3_scratch);
1149 __ JMP(G3_scratch, 0);
1150 } else {
1151 __ jmp(Lscratch, 0);
1152 }
1153 __ delayed()->nop();
1154
1155
1156 if (inc_counter) {
1157 // handle invocation counter overflow
1158 __ bind(invocation_counter_overflow);
1159 generate_counter_overflow(Lcontinue);
1160 }
1161
1162
1163
1164 return entry;
1165 }
1166
1167
1168 // Generic method entry to (asm) interpreter
1169 //------------------------------------------------------------------------------------------------------------------------
1170 //
1171 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1172 address entry = __ pc();
1173
1174 bool inc_counter = UseCompiler || CountCompiledCalls;
1175
1176 // the following temporary registers are used during frame creation
1177 const Register Gtmp1 = G3_scratch ;
1178 const Register Gtmp2 = G1_scratch;
1179
1180 // make sure registers are different!
1181 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1182
1183 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1184 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1185 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1186 // and use in the asserts.
1187 const Address access_flags (Lmethod, 0, in_bytes(methodOopDesc::access_flags_offset()));
1188
1189 __ verify_oop(G5_method);
1190
1191 const Register Glocals_size = G3;
1192 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1193
1194 // make sure method is not native & not abstract
1195 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1196 #ifdef ASSERT
1197 __ ld(G5_method, in_bytes(methodOopDesc::access_flags_offset()), Gtmp1);
1198 {
1199 Label L;
1200 __ btst(JVM_ACC_NATIVE, Gtmp1);
1201 __ br(Assembler::zero, false, Assembler::pt, L);
1202 __ delayed()->nop();
1203 __ stop("tried to execute native method as non-native");
1204 __ bind(L);
1205 }
1206 { Label L;
1207 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1208 __ br(Assembler::zero, false, Assembler::pt, L);
1209 __ delayed()->nop();
1210 __ stop("tried to execute abstract method as non-abstract");
1211 __ bind(L);
1212 }
1213 #endif // ASSERT
1214
1215 // generate the code to allocate the interpreter stack frame
1216
1217 generate_fixed_frame(false);
1218
1219 #ifdef FAST_DISPATCH
1220 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1221 // set bytecode dispatch table base
1222 #endif
1223
1224 //
1225 // Code to initialize the extra (i.e. non-parm) locals
1226 //
1227 Register init_value = noreg; // will be G0 if we must clear locals
1228 // The way the code was setup before zerolocals was always true for vanilla java entries.
1229 // It could only be false for the specialized entries like accessor or empty which have
1230 // no extra locals so the testing was a waste of time and the extra locals were always
1231 // initialized. We removed this extra complication to already over complicated code.
1232
1233 init_value = G0;
1234 Label clear_loop;
1235
1236 // NOTE: If you change the frame layout, this code will need to
1237 // be updated!
1238 __ lduh( size_of_locals, O2 );
1239 __ lduh( size_of_parameters, O1 );
1240 __ sll( O2, Interpreter::logStackElementSize(), O2);
1241 __ sll( O1, Interpreter::logStackElementSize(), O1 );
1242 __ sub( Llocals, O2, O2 );
1243 __ sub( Llocals, O1, O1 );
1244
1245 __ bind( clear_loop );
1246 __ inc( O2, wordSize );
1247
1248 __ cmp( O2, O1 );
1249 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1250 __ delayed()->st_ptr( init_value, O2, 0 );
1251
1252 const Address do_not_unlock_if_synchronized(G2_thread, 0,
1253 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
1254 // Since at this point in the method invocation the exception handler
1255 // would try to exit the monitor of synchronized methods which hasn't
1256 // been entered yet, we set the thread local variable
1257 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1258 // runtime, exception handling i.e. unlock_if_synchronized_method will
1259 // check this thread local flag.
1260 __ movbool(true, G3_scratch);
1261 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1262
1263 // increment invocation counter and check for overflow
1264 //
1265 // Note: checking for negative value instead of overflow
1266 // so we have a 'sticky' overflow test (may be of
1267 // importance as soon as we have true MT/MP)
1268 Label invocation_counter_overflow;
1269 Label profile_method;
1270 Label profile_method_continue;
1271 Label Lcontinue;
1272 if (inc_counter) {
1273 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1274 if (ProfileInterpreter) {
1275 __ bind(profile_method_continue);
1276 }
1277 }
1278 __ bind(Lcontinue);
1279
1280 bang_stack_shadow_pages(false);
1281
1282 // reset the _do_not_unlock_if_synchronized flag
1283 __ stbool(G0, do_not_unlock_if_synchronized);
1284
1285 // check for synchronized methods
1286 // Must happen AFTER invocation_counter check and stack overflow check,
1287 // so method is not locked if overflows.
1288
1289 if (synchronized) {
1290 lock_method();
1291 } else {
1292 #ifdef ASSERT
1293 { Label ok;
1294 __ ld(access_flags, O0);
1295 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1296 __ br( Assembler::zero, false, Assembler::pt, ok);
1297 __ delayed()->nop();
1298 __ stop("method needs synchronization");
1299 __ bind(ok);
1300 }
1301 #endif // ASSERT
1302 }
1303
1304 // start execution
1305
1306 __ verify_thread();
1307
1308 // jvmti support
1309 __ notify_method_entry();
1310
1311 // start executing instructions
1312 __ dispatch_next(vtos);
1313
1314
1315 if (inc_counter) {
1316 if (ProfileInterpreter) {
1317 // We have decided to profile this method in the interpreter
1318 __ bind(profile_method);
1319
1320 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method), Lbcp, true);
1321
1322 #ifdef ASSERT
1323 __ tst(O0);
1324 __ breakpoint_trap(Assembler::notEqual);
1325 #endif
1326
1327 __ set_method_data_pointer();
1328
1329 __ ba(false, profile_method_continue);
1330 __ delayed()->nop();
1331 }
1332
1333 // handle invocation counter overflow
1334 __ bind(invocation_counter_overflow);
1335 generate_counter_overflow(Lcontinue);
1336 }
1337
1338
1339 return entry;
1340 }
1341
1342
1343 //----------------------------------------------------------------------------------------------------
1344 // Entry points & stack frame layout
1345 //
1346 // Here we generate the various kind of entries into the interpreter.
1347 // The two main entry type are generic bytecode methods and native call method.
1348 // These both come in synchronized and non-synchronized versions but the
1349 // frame layout they create is very similar. The other method entry
1350 // types are really just special purpose entries that are really entry
1351 // and interpretation all in one. These are for trivial methods like
1352 // accessor, empty, or special math methods.
1353 //
1354 // When control flow reaches any of the entry types for the interpreter
1355 // the following holds ->
1356 //
1357 // C2 Calling Conventions:
1358 //
1359 // The entry code below assumes that the following registers are set
1360 // when coming in:
1361 // G5_method: holds the methodOop of the method to call
1362 // Lesp: points to the TOS of the callers expression stack
1363 // after having pushed all the parameters
1364 //
1365 // The entry code does the following to setup an interpreter frame
1366 // pop parameters from the callers stack by adjusting Lesp
1367 // set O0 to Lesp
1368 // compute X = (max_locals - num_parameters)
1369 // bump SP up by X to accomadate the extra locals
1370 // compute X = max_expression_stack
1371 // + vm_local_words
1372 // + 16 words of register save area
1373 // save frame doing a save sp, -X, sp growing towards lower addresses
1374 // set Lbcp, Lmethod, LcpoolCache
1375 // set Llocals to i0
1376 // set Lmonitors to FP - rounded_vm_local_words
1377 // set Lesp to Lmonitors - 4
1378 //
1379 // The frame has now been setup to do the rest of the entry code
1380
1381 // Try this optimization: Most method entries could live in a
1382 // "one size fits all" stack frame without all the dynamic size
1383 // calculations. It might be profitable to do all this calculation
1384 // statically and approximately for "small enough" methods.
1385
1386 //-----------------------------------------------------------------------------------------------
1387
1388 // C1 Calling conventions
1389 //
1390 // Upon method entry, the following registers are setup:
1391 //
1392 // g2 G2_thread: current thread
1393 // g5 G5_method: method to activate
1394 // g4 Gargs : pointer to last argument
1395 //
1396 //
1397 // Stack:
1398 //
1399 // +---------------+ <--- sp
1400 // | |
1401 // : reg save area :
1402 // | |
1403 // +---------------+ <--- sp + 0x40
1404 // | |
1405 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1406 // | |
1407 // +---------------+ <--- sp + 0x5c
1408 // | |
1409 // : free :
1410 // | |
1411 // +---------------+ <--- Gargs
1412 // | |
1413 // : arguments :
1414 // | |
1415 // +---------------+
1416 // | |
1417 //
1418 //
1419 //
1420 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1421 //
1422 // +---------------+ <--- sp
1423 // | |
1424 // : reg save area :
1425 // | |
1426 // +---------------+ <--- sp + 0x40
1427 // | |
1428 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1429 // | |
1430 // +---------------+ <--- sp + 0x5c
1431 // | |
1432 // : :
1433 // | | <--- Lesp
1434 // +---------------+ <--- Lmonitors (fp - 0x18)
1435 // | VM locals |
1436 // +---------------+ <--- fp
1437 // | |
1438 // : reg save area :
1439 // | |
1440 // +---------------+ <--- fp + 0x40
1441 // | |
1442 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1443 // | |
1444 // +---------------+ <--- fp + 0x5c
1445 // | |
1446 // : free :
1447 // | |
1448 // +---------------+
1449 // | |
1450 // : nonarg locals :
1451 // | |
1452 // +---------------+
1453 // | |
1454 // : arguments :
1455 // | | <--- Llocals
1456 // +---------------+ <--- Gargs
1457 // | |
1458
1459 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1460
1461 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1462 // expression stack, the callee will have callee_extra_locals (so we can account for
1463 // frame extension) and monitor_size for monitors. Basically we need to calculate
1464 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1465 //
1466 //
1467 // The big complicating thing here is that we must ensure that the stack stays properly
1468 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1469 // needs to be aligned for). We are given that the sp (fp) is already aligned by
1470 // the caller so we must ensure that it is properly aligned for our callee.
1471 //
1472 const int rounded_vm_local_words =
1473 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1474 // callee_locals and max_stack are counts, not the size in frame.
1475 const int locals_size =
1476 round_to(callee_extra_locals * Interpreter::stackElementWords(), WordsPerLong);
1477 const int extra_stack = methodOopDesc::extra_stack();
1478 const int max_stack_words = (max_stack + extra_stack) * Interpreter::stackElementWords();
1479 return (round_to((max_stack_words
1480 + rounded_vm_local_words
1481 + frame::memory_parameter_word_sp_offset), WordsPerLong)
1482 // already rounded
1483 + locals_size + monitor_size);
1484 }
1485
1486 // How much stack a method top interpreter activation needs in words.
1487 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
1488
1489 // See call_stub code
1490 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
1491 WordsPerLong); // 7 + register save area
1492
1493 // Save space for one monitor to get into the interpreted method in case
1494 // the method is synchronized
1495 int monitor_size = method->is_synchronized() ?
1496 1*frame::interpreter_frame_monitor_size() : 0;
1497 return size_activation_helper(method->max_locals(), method->max_stack(),
1498 monitor_size) + call_stub_size;
1499 }
1500
1501 int AbstractInterpreter::layout_activation(methodOop method,
1502 int tempcount,
1503 int popframe_extra_args,
1504 int moncount,
1505 int callee_param_count,
1506 int callee_local_count,
1507 frame* caller,
1508 frame* interpreter_frame,
1509 bool is_top_frame) {
1510 // Note: This calculation must exactly parallel the frame setup
1511 // in InterpreterGenerator::generate_fixed_frame.
1512 // If f!=NULL, set up the following variables:
1513 // - Lmethod
1514 // - Llocals
1515 // - Lmonitors (to the indicated number of monitors)
1516 // - Lesp (to the indicated number of temps)
1517 // The frame f (if not NULL) on entry is a description of the caller of the frame
1518 // we are about to layout. We are guaranteed that we will be able to fill in a
1519 // new interpreter frame as its callee (i.e. the stack space is allocated and
1520 // the amount was determined by an earlier call to this method with f == NULL).
1521 // On return f (if not NULL) while describe the interpreter frame we just layed out.
1522
1523 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
1524 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1525
1526 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1527 //
1528 // Note: if you look closely this appears to be doing something much different
1529 // than generate_fixed_frame. What is happening is this. On sparc we have to do
1530 // this dance with interpreter_sp_adjustment because the window save area would
1531 // appear just below the bottom (tos) of the caller's java expression stack. Because
1532 // the interpreter want to have the locals completely contiguous generate_fixed_frame
1533 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1534 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1535 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1536 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1537 // because the oldest frame would have adjust its callers frame and yet that frame
1538 // already exists and isn't part of this array of frames we are unpacking. So at first
1539 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1540 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1541 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1542 // add up. It does seem like it simpler to account for the adjustment here (and remove the
1543 // callee... parameters here). However this would mean that this routine would have to take
1544 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1545 // and run the calling loop in the reverse order. This would also would appear to mean making
1546 // this code aware of what the interactions are when that initial caller fram was an osr or
1547 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
1548 // there is no sense in messing working code.
1549 //
1550
1551 int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
1552 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1553
1554 int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
1555 monitor_size);
1556
1557 if (interpreter_frame != NULL) {
1558 // The skeleton frame must already look like an interpreter frame
1559 // even if not fully filled out.
1560 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1561
1562 intptr_t* fp = interpreter_frame->fp();
1563
1564 JavaThread* thread = JavaThread::current();
1565 RegisterMap map(thread, false);
1566 // More verification that skeleton frame is properly walkable
1567 assert(fp == caller->sp(), "fp must match");
1568
1569 intptr_t* montop = fp - rounded_vm_local_words;
1570
1571 // preallocate monitors (cf. __ add_monitor_to_stack)
1572 intptr_t* monitors = montop - monitor_size;
1573
1574 // preallocate stack space
1575 intptr_t* esp = monitors - 1 -
1576 (tempcount * Interpreter::stackElementWords()) -
1577 popframe_extra_args;
1578
1579 int local_words = method->max_locals() * Interpreter::stackElementWords();
1580 int parm_words = method->size_of_parameters() * Interpreter::stackElementWords();
1581 NEEDS_CLEANUP;
1582 intptr_t* locals;
1583 if (caller->is_interpreted_frame()) {
1584 // Can force the locals area to end up properly overlapping the top of the expression stack.
1585 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1586 // Note that this computation means we replace size_of_parameters() values from the caller
1587 // interpreter frame's expression stack with our argument locals
1588 locals = Lesp_ptr + parm_words;
1589 int delta = local_words - parm_words;
1590 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1591 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1592 } else {
1593 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1594 // Don't have Lesp available; lay out locals block in the caller
1595 // adjacent to the register window save area.
1596 //
1597 // Compiled frames do not allocate a varargs area which is why this if
1598 // statement is needed.
1599 //
1600 if (caller->is_compiled_frame()) {
1601 locals = fp + frame::register_save_words + local_words - 1;
1602 } else {
1603 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1604 }
1605 if (!caller->is_entry_frame()) {
1606 // Caller wants his own SP back
1607 int caller_frame_size = caller->cb()->frame_size();
1608 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1609 }
1610 }
1611 if (TraceDeoptimization) {
1612 if (caller->is_entry_frame()) {
1613 // make sure I5_savedSP and the entry frames notion of saved SP
1614 // agree. This assertion duplicate a check in entry frame code
1615 // but catches the failure earlier.
1616 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1617 "would change callers SP");
1618 }
1619 if (caller->is_entry_frame()) {
1620 tty->print("entry ");
1621 }
1622 if (caller->is_compiled_frame()) {
1623 tty->print("compiled ");
1624 if (caller->is_deoptimized_frame()) {
1625 tty->print("(deopt) ");
1626 }
1627 }
1628 if (caller->is_interpreted_frame()) {
1629 tty->print("interpreted ");
1630 }
1631 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1632 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1633 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1634 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1635 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1636 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1637 tty->print_cr("Llocals = 0x%x", locals);
1638 tty->print_cr("Lesp = 0x%x", esp);
1639 tty->print_cr("Lmonitors = 0x%x", monitors);
1640 }
1641
1642 if (method->max_locals() > 0) {
1643 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1644 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1645 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1646 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1647 }
1648 #ifdef _LP64
1649 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1650 #endif
1651
1652 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
1653 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
1654 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
1655 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
1656 // Llast_SP will be same as SP as there is no adapter space
1657 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1658 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1659 #ifdef FAST_DISPATCH
1660 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1661 #endif
1662
1663
1664 #ifdef ASSERT
1665 BasicObjectLock* mp = (BasicObjectLock*)monitors;
1666
1667 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1668 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize())+Interpreter::value_offset_in_bytes()), "locals match");
1669 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
1670 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1671 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1672
1673 // check bounds
1674 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1675 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1676 assert(lo < monitors && montop <= hi, "monitors in bounds");
1677 assert(lo <= esp && esp < monitors, "esp in bounds");
1678 #endif // ASSERT
1679 }
1680
1681 return raw_frame_size;
1682 }
1683
1684 //----------------------------------------------------------------------------------------------------
1685 // Exceptions
1686 void TemplateInterpreterGenerator::generate_throw_exception() {
1687
1688 // Entry point in previous activation (i.e., if the caller was interpreted)
1689 Interpreter::_rethrow_exception_entry = __ pc();
1690 // O0: exception
1691
1692 // entry point for exceptions thrown within interpreter code
1693 Interpreter::_throw_exception_entry = __ pc();
1694 __ verify_thread();
1695 // expression stack is undefined here
1696 // O0: exception, i.e. Oexception
1697 // Lbcp: exception bcx
1698 __ verify_oop(Oexception);
1699
1700
1701 // expression stack must be empty before entering the VM in case of an exception
1702 __ empty_expression_stack();
1703 // find exception handler address and preserve exception oop
1704 // call C routine to find handler and jump to it
1705 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1706 __ push_ptr(O1); // push exception for exception handler bytecodes
1707
1708 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1709 __ delayed()->nop();
1710
1711
1712 // if the exception is not handled in the current frame
1713 // the frame is removed and the exception is rethrown
1714 // (i.e. exception continuation is _rethrow_exception)
1715 //
1716 // Note: At this point the bci is still the bxi for the instruction which caused
1717 // the exception and the expression stack is empty. Thus, for any VM calls
1718 // at this point, GC will find a legal oop map (with empty expression stack).
1719
1720 // in current activation
1721 // tos: exception
1722 // Lbcp: exception bcp
1723
1724 //
1725 // JVMTI PopFrame support
1726 //
1727
1728 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1729 Address popframe_condition_addr (G2_thread, 0, in_bytes(JavaThread::popframe_condition_offset()));
1730 // Set the popframe_processing bit in popframe_condition indicating that we are
1731 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1732 // popframe handling cycles.
1733
1734 __ ld(popframe_condition_addr, G3_scratch);
1735 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1736 __ stw(G3_scratch, popframe_condition_addr);
1737
1738 // Empty the expression stack, as in normal exception handling
1739 __ empty_expression_stack();
1740 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1741
1742 {
1743 // Check to see whether we are returning to a deoptimized frame.
1744 // (The PopFrame call ensures that the caller of the popped frame is
1745 // either interpreted or compiled and deoptimizes it if compiled.)
1746 // In this case, we can't call dispatch_next() after the frame is
1747 // popped, but instead must save the incoming arguments and restore
1748 // them after deoptimization has occurred.
1749 //
1750 // Note that we don't compare the return PC against the
1751 // deoptimization blob's unpack entry because of the presence of
1752 // adapter frames in C2.
1753 Label caller_not_deoptimized;
1754 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1755 __ tst(O0);
1756 __ brx(Assembler::notEqual, false, Assembler::pt, caller_not_deoptimized);
1757 __ delayed()->nop();
1758
1759 const Register Gtmp1 = G3_scratch;
1760 const Register Gtmp2 = G1_scratch;
1761
1762 // Compute size of arguments for saving when returning to deoptimized caller
1763 __ lduh(Lmethod, in_bytes(methodOopDesc::size_of_parameters_offset()), Gtmp1);
1764 __ sll(Gtmp1, Interpreter::logStackElementSize(), Gtmp1);
1765 __ sub(Llocals, Gtmp1, Gtmp2);
1766 __ add(Gtmp2, wordSize, Gtmp2);
1767 // Save these arguments
1768 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1769 // Inform deoptimization that it is responsible for restoring these arguments
1770 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1771 Address popframe_condition_addr(G2_thread, 0, in_bytes(JavaThread::popframe_condition_offset()));
1772 __ st(Gtmp1, popframe_condition_addr);
1773
1774 // Return from the current method
1775 // The caller's SP was adjusted upon method entry to accomodate
1776 // the callee's non-argument locals. Undo that adjustment.
1777 __ ret();
1778 __ delayed()->restore(I5_savedSP, G0, SP);
1779
1780 __ bind(caller_not_deoptimized);
1781 }
1782
1783 // Clear the popframe condition flag
1784 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1785
1786 // Get out of the current method (how this is done depends on the particular compiler calling
1787 // convention that the interpreter currently follows)
1788 // The caller's SP was adjusted upon method entry to accomodate
1789 // the callee's non-argument locals. Undo that adjustment.
1790 __ restore(I5_savedSP, G0, SP);
1791 // The method data pointer was incremented already during
1792 // call profiling. We have to restore the mdp for the current bcp.
1793 if (ProfileInterpreter) {
1794 __ set_method_data_pointer_for_bcp();
1795 }
1796 // Resume bytecode interpretation at the current bcp
1797 __ dispatch_next(vtos);
1798 // end of JVMTI PopFrame support
1799
1800 Interpreter::_remove_activation_entry = __ pc();
1801
1802 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1803 __ pop_ptr(Oexception); // get exception
1804
1805 // Intel has the following comment:
1806 //// remove the activation (without doing throws on illegalMonitorExceptions)
1807 // They remove the activation without checking for bad monitor state.
1808 // %%% We should make sure this is the right semantics before implementing.
1809
1810 // %%% changed set_vm_result_2 to set_vm_result and get_vm_result_2 to get_vm_result. Is there a bug here?
1811 __ set_vm_result(Oexception);
1812 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1813
1814 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1815
1816 __ get_vm_result(Oexception);
1817 __ verify_oop(Oexception);
1818
1819 const int return_reg_adjustment = frame::pc_return_offset;
1820 Address issuing_pc_addr(I7, 0, return_reg_adjustment);
1821
1822 // We are done with this activation frame; find out where to go next.
1823 // The continuation point will be an exception handler, which expects
1824 // the following registers set up:
1825 //
1826 // Oexception: exception
1827 // Oissuing_pc: the local call that threw exception
1828 // Other On: garbage
1829 // In/Ln: the contents of the caller's register window
1830 //
1831 // We do the required restore at the last possible moment, because we
1832 // need to preserve some state across a runtime call.
1833 // (Remember that the caller activation is unknown--it might not be
1834 // interpreted, so things like Lscratch are useless in the caller.)
1835
1836 // Although the Intel version uses call_C, we can use the more
1837 // compact call_VM. (The only real difference on SPARC is a
1838 // harmlessly ignored [re]set_last_Java_frame, compared with
1839 // the Intel code which lacks this.)
1840 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1841 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1842 __ super_call_VM_leaf(L7_thread_cache,
1843 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1844 Oissuing_pc->after_save());
1845
1846 // The caller's SP was adjusted upon method entry to accomodate
1847 // the callee's non-argument locals. Undo that adjustment.
1848 __ JMP(O0, 0); // return exception handler in caller
1849 __ delayed()->restore(I5_savedSP, G0, SP);
1850
1851 // (same old exception object is already in Oexception; see above)
1852 // Note that an "issuing PC" is actually the next PC after the call
1853 }
1854
1855
1856 //
1857 // JVMTI ForceEarlyReturn support
1858 //
1859
1860 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1861 address entry = __ pc();
1862
1863 __ empty_expression_stack();
1864 __ load_earlyret_value(state);
1865
1866 __ ld_ptr(Address(G2_thread, 0, in_bytes(JavaThread::jvmti_thread_state_offset())), G3_scratch);
1867 Address cond_addr(G3_scratch, 0, in_bytes(JvmtiThreadState::earlyret_state_offset()));
1868
1869 // Clear the earlyret state
1870 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1871
1872 __ remove_activation(state,
1873 /* throw_monitor_exception */ false,
1874 /* install_monitor_exception */ false);
1875
1876 // The caller's SP was adjusted upon method entry to accomodate
1877 // the callee's non-argument locals. Undo that adjustment.
1878 __ ret(); // return to caller
1879 __ delayed()->restore(I5_savedSP, G0, SP);
1880
1881 return entry;
1882 } // end of JVMTI ForceEarlyReturn support
1883
1884
1885 //------------------------------------------------------------------------------------------------------------------------
1886 // Helper for vtos entry point generation
1887
1888 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t, address& bep, address& cep, address& sep, address& aep, address& iep, address& lep, address& fep, address& dep, address& vep) {
1889 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1890 Label L;
1891 aep = __ pc(); __ push_ptr(); __ ba(false, L); __ delayed()->nop();
1892 fep = __ pc(); __ push_f(); __ ba(false, L); __ delayed()->nop();
1893 dep = __ pc(); __ push_d(); __ ba(false, L); __ delayed()->nop();
1894 lep = __ pc(); __ push_l(); __ ba(false, L); __ delayed()->nop();
1895 iep = __ pc(); __ push_i();
1896 bep = cep = sep = iep; // there aren't any
1897 vep = __ pc(); __ bind(L); // fall through
1898 generate_and_dispatch(t);
1899 }
1900
1901 // --------------------------------------------------------------------------------
1902
1903
1904 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1905 : TemplateInterpreterGenerator(code) {
1906 generate_all(); // down here so it can be "virtual"
1907 }
1908
1909 // --------------------------------------------------------------------------------
1910
1911 // Non-product code
1912 #ifndef PRODUCT
1913 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1914 address entry = __ pc();
1915
1916 __ push(state);
1917 __ mov(O7, Lscratch); // protect return address within interpreter
1918
1919 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1920 __ mov( Otos_l2, G3_scratch );
1921 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1922 __ mov(Lscratch, O7); // restore return address
1923 __ pop(state);
1924 __ retl();
1925 __ delayed()->nop();
1926
1927 return entry;
1928 }
1929
1930
1931 // helpers for generate_and_dispatch
1932
1933 void TemplateInterpreterGenerator::count_bytecode() {
1934 Address c(G3_scratch, (address)&BytecodeCounter::_counter_value);
1935 __ load_contents(c, G4_scratch);
1936 __ inc(G4_scratch);
1937 __ st(G4_scratch, c);
1938 }
1939
1940
1941 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1942 Address bucket( G3_scratch, (address) &BytecodeHistogram::_counters[t->bytecode()] );
1943 __ load_contents(bucket, G4_scratch);
1944 __ inc(G4_scratch);
1945 __ st(G4_scratch, bucket);
1946 }
1947
1948
1949 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1950 address index_addr = (address)&BytecodePairHistogram::_index;
1951 Address index(G3_scratch, index_addr);
1952
1953 address counters_addr = (address)&BytecodePairHistogram::_counters;
1954 Address counters(G3_scratch, counters_addr);
1955
1956 // get index, shift out old bytecode, bring in new bytecode, and store it
1957 // _index = (_index >> log2_number_of_codes) |
1958 // (bytecode << log2_number_of_codes);
1959
1960
1961 __ load_contents( index, G4_scratch );
1962 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
1963 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
1964 __ or3( G3_scratch, G4_scratch, G4_scratch );
1965 __ store_contents( G4_scratch, index );
1966
1967 // bump bucket contents
1968 // _counters[_index] ++;
1969
1970 __ load_address( counters ); // loads into G3_scratch
1971 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
1972 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
1973 __ ld (G3_scratch, 0, G4_scratch);
1974 __ inc (G4_scratch);
1975 __ st (G4_scratch, 0, G3_scratch);
1976 }
1977
1978
1979 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
1980 // Call a little run-time stub to avoid blow-up for each bytecode.
1981 // The run-time runtime saves the right registers, depending on
1982 // the tosca in-state for the given template.
1983 address entry = Interpreter::trace_code(t->tos_in());
1984 guarantee(entry != NULL, "entry must have been generated");
1985 __ call(entry, relocInfo::none);
1986 __ delayed()->nop();
1987 }
1988
1989
1990 void TemplateInterpreterGenerator::stop_interpreter_at() {
1991 Address counter(G3_scratch , (address)&BytecodeCounter::_counter_value);
1992 __ load_contents (counter, G3_scratch );
1993 Address stop_at(G4_scratch, (address)&StopInterpreterAt);
1994 __ load_ptr_contents(stop_at, G4_scratch);
1995 __ cmp(G3_scratch, G4_scratch);
1996 __ breakpoint_trap(Assembler::equal);
1997 }
1998 #endif // not PRODUCT
1999 #endif // !CC_INTERP