1 /*
2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "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 AddressLiteral thrower(Interpreter::throw_exception_entry());
91 __ jump_to(thrower, G3_scratch);
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 // Arguments are: required type in G5_method_type, and
112 // failing object (or NULL) in G3_method_handle.
113 address TemplateInterpreterGenerator::generate_WrongMethodType_handler() {
114 address entry = __ pc();
115 // expression stack must be empty before entering the VM if an exception
116 // happened
117 __ empty_expression_stack();
118 // load exception object
119 __ call_VM(Oexception,
120 CAST_FROM_FN_PTR(address,
121 InterpreterRuntime::throw_WrongMethodTypeException),
122 G5_method_type, // required
123 G3_method_handle); // actual
124 __ should_not_reach_here();
125 return entry;
126 }
127
128
129 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
130 address entry = __ pc();
131 // expression stack must be empty before entering the VM if an exception happened
132 __ empty_expression_stack();
133 // convention: expect aberrant index in register G3_scratch, then shuffle the
134 // index to G4_scratch for the VM call
135 __ mov(G3_scratch, G4_scratch);
136 __ set((intptr_t)name, G3_scratch);
137 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
138 __ should_not_reach_here();
139 return entry;
140 }
141
142
143 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
144 address entry = __ pc();
145 // expression stack must be empty before entering the VM if an exception happened
146 __ empty_expression_stack();
147 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
148 __ should_not_reach_here();
149 return entry;
150 }
151
152
153 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) {
154 TosState incoming_state = state;
155
156 Label cont;
157 address compiled_entry = __ pc();
158
159 address entry = __ pc();
160 #if !defined(_LP64) && defined(COMPILER2)
161 // All return values are where we want them, except for Longs. C2 returns
162 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
163 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
164 // build even if we are returning from interpreted we just do a little
165 // stupid shuffing.
166 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
167 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
168 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
169
170 if (incoming_state == ltos) {
171 __ srl (G1, 0, O1);
172 __ srlx(G1, 32, O0);
173 }
174 #endif // !_LP64 && COMPILER2
175
176 __ bind(cont);
177
178 // The callee returns with the stack possibly adjusted by adapter transition
179 // We remove that possible adjustment here.
180 // All interpreter local registers are untouched. Any result is passed back
181 // in the O0/O1 or float registers. Before continuing, the arguments must be
182 // popped from the java expression stack; i.e., Lesp must be adjusted.
183
184 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
185
186 Label L_got_cache, L_giant_index;
187 const Register cache = G3_scratch;
188 const Register size = G1_scratch;
189 if (EnableInvokeDynamic) {
190 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode.
191 __ cmp(G1_scratch, Bytecodes::_invokedynamic);
192 __ br(Assembler::equal, false, Assembler::pn, L_giant_index);
193 __ delayed()->nop();
194 }
195 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1);
196 __ bind(L_got_cache);
197 __ ld_ptr(cache, constantPoolCacheOopDesc::base_offset() +
198 ConstantPoolCacheEntry::flags_offset(), size);
199 __ and3(size, 0xFF, size); // argument size in words
200 __ sll(size, Interpreter::logStackElementSize, size); // each argument size in bytes
201 __ add(Lesp, size, Lesp); // pop arguments
202 __ dispatch_next(state, step);
203
204 // out of the main line of code...
205 if (EnableInvokeDynamic) {
206 __ bind(L_giant_index);
207 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1, sizeof(u4));
208 __ ba(false, L_got_cache);
209 __ delayed()->nop();
210 }
211
212 return entry;
213 }
214
215
216 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
217 address entry = __ pc();
218 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
219 { Label L;
220 Address exception_addr(G2_thread, Thread::pending_exception_offset());
221 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
222 __ tst(Gtemp);
223 __ brx(Assembler::equal, false, Assembler::pt, L);
224 __ delayed()->nop();
225 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
226 __ should_not_reach_here();
227 __ bind(L);
228 }
229 __ dispatch_next(state, step);
230 return entry;
231 }
232
233 // A result handler converts/unboxes a native call result into
234 // a java interpreter/compiler result. The current frame is an
235 // interpreter frame. The activation frame unwind code must be
236 // consistent with that of TemplateTable::_return(...). In the
237 // case of native methods, the caller's SP was not modified.
238 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
239 address entry = __ pc();
240 Register Itos_i = Otos_i ->after_save();
241 Register Itos_l = Otos_l ->after_save();
242 Register Itos_l1 = Otos_l1->after_save();
243 Register Itos_l2 = Otos_l2->after_save();
244 switch (type) {
245 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
246 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
247 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
248 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
249 case T_LONG :
250 #ifndef _LP64
251 __ mov(O1, Itos_l2); // move other half of long
252 #endif // ifdef or no ifdef, fall through to the T_INT case
253 case T_INT : __ mov(O0, Itos_i); break;
254 case T_VOID : /* nothing to do */ break;
255 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
256 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
257 case T_OBJECT :
258 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
259 __ verify_oop(Itos_i);
260 break;
261 default : ShouldNotReachHere();
262 }
263 __ ret(); // return from interpreter activation
264 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
265 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
266 return entry;
267 }
268
269 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
270 address entry = __ pc();
271 __ push(state);
272 __ call_VM(noreg, runtime_entry);
273 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
274 return entry;
275 }
276
277
278 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
279 address entry = __ pc();
280 __ dispatch_next(state);
281 return entry;
282 }
283
284 //
285 // Helpers for commoning out cases in the various type of method entries.
286 //
287
288 // increment invocation count & check for overflow
289 //
290 // Note: checking for negative value instead of overflow
291 // so we have a 'sticky' overflow test
292 //
293 // Lmethod: method
294 // ??: invocation counter
295 //
296 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
297 // Note: In tiered we increment either counters in methodOop or in MDO depending if we're profiling or not.
298 if (TieredCompilation) {
299 const int increment = InvocationCounter::count_increment;
300 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
301 Label no_mdo, done;
302 if (ProfileInterpreter) {
303 // If no method data exists, go to profile_continue.
304 __ ld_ptr(Lmethod, methodOopDesc::method_data_offset(), G4_scratch);
305 __ br_null(G4_scratch, false, Assembler::pn, no_mdo);
306 __ delayed()->nop();
307 // Increment counter
308 Address mdo_invocation_counter(G4_scratch,
309 in_bytes(methodDataOopDesc::invocation_counter_offset()) +
310 in_bytes(InvocationCounter::counter_offset()));
311 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
312 G3_scratch, Lscratch,
313 Assembler::zero, overflow);
314 __ ba(false, done);
315 __ delayed()->nop();
316 }
317
318 // Increment counter in methodOop
319 __ bind(no_mdo);
320 Address invocation_counter(Lmethod,
321 in_bytes(methodOopDesc::invocation_counter_offset()) +
322 in_bytes(InvocationCounter::counter_offset()));
323 __ increment_mask_and_jump(invocation_counter, increment, mask,
324 G3_scratch, Lscratch,
325 Assembler::zero, overflow);
326 __ bind(done);
327 } else {
328 // Update standard invocation counters
329 __ increment_invocation_counter(O0, G3_scratch);
330 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
331 Address interpreter_invocation_counter(Lmethod,in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
332 __ ld(interpreter_invocation_counter, G3_scratch);
333 __ inc(G3_scratch);
334 __ st(G3_scratch, interpreter_invocation_counter);
335 }
336
337 if (ProfileInterpreter && profile_method != NULL) {
338 // Test to see if we should create a method data oop
339 AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
340 __ load_contents(profile_limit, G3_scratch);
341 __ cmp(O0, G3_scratch);
342 __ br(Assembler::lessUnsigned, false, Assembler::pn, *profile_method_continue);
343 __ delayed()->nop();
344
345 // if no method data exists, go to profile_method
346 __ test_method_data_pointer(*profile_method);
347 }
348
349 AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
350 __ load_contents(invocation_limit, G3_scratch);
351 __ cmp(O0, G3_scratch);
352 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
353 __ delayed()->nop();
354 }
355
356 }
357
358 // Allocate monitor and lock method (asm interpreter)
359 // ebx - methodOop
360 //
361 void InterpreterGenerator::lock_method(void) {
362 __ ld(Lmethod, in_bytes(methodOopDesc::access_flags_offset()), O0); // Load access flags.
363
364 #ifdef ASSERT
365 { Label ok;
366 __ btst(JVM_ACC_SYNCHRONIZED, O0);
367 __ br( Assembler::notZero, false, Assembler::pt, ok);
368 __ delayed()->nop();
369 __ stop("method doesn't need synchronization");
370 __ bind(ok);
371 }
372 #endif // ASSERT
373
374 // get synchronization object to O0
375 { Label done;
376 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
377 __ btst(JVM_ACC_STATIC, O0);
378 __ br( Assembler::zero, true, Assembler::pt, done);
379 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
380
381 __ ld_ptr( Lmethod, in_bytes(methodOopDesc::constants_offset()), O0);
382 __ ld_ptr( O0, constantPoolOopDesc::pool_holder_offset_in_bytes(), O0);
383
384 // lock the mirror, not the klassOop
385 __ ld_ptr( O0, mirror_offset, O0);
386
387 #ifdef ASSERT
388 __ tst(O0);
389 __ breakpoint_trap(Assembler::zero);
390 #endif // ASSERT
391
392 __ bind(done);
393 }
394
395 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
396 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
397 // __ untested("lock_object from method entry");
398 __ lock_object(Lmonitors, O0);
399 }
400
401
402 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
403 Register Rscratch,
404 Register Rscratch2) {
405 const int page_size = os::vm_page_size();
406 Address saved_exception_pc(G2_thread, JavaThread::saved_exception_pc_offset());
407 Label after_frame_check;
408
409 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
410
411 __ set( page_size, Rscratch );
412 __ cmp( Rframe_size, Rscratch );
413
414 __ br( Assembler::lessEqual, false, Assembler::pt, after_frame_check );
415 __ delayed()->nop();
416
417 // get the stack base, and in debug, verify it is non-zero
418 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
419 #ifdef ASSERT
420 Label base_not_zero;
421 __ cmp( Rscratch, G0 );
422 __ brx( Assembler::notEqual, false, Assembler::pn, base_not_zero );
423 __ delayed()->nop();
424 __ stop("stack base is zero in generate_stack_overflow_check");
425 __ bind(base_not_zero);
426 #endif
427
428 // get the stack size, and in debug, verify it is non-zero
429 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
430 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
431 #ifdef ASSERT
432 Label size_not_zero;
433 __ cmp( Rscratch2, G0 );
434 __ brx( Assembler::notEqual, false, Assembler::pn, size_not_zero );
435 __ delayed()->nop();
436 __ stop("stack size is zero in generate_stack_overflow_check");
437 __ bind(size_not_zero);
438 #endif
439
440 // compute the beginning of the protected zone minus the requested frame size
441 __ sub( Rscratch, Rscratch2, Rscratch );
442 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
443 __ add( Rscratch, Rscratch2, Rscratch );
444
445 // Add in the size of the frame (which is the same as subtracting it from the
446 // SP, which would take another register
447 __ add( Rscratch, Rframe_size, Rscratch );
448
449 // the frame is greater than one page in size, so check against
450 // the bottom of the stack
451 __ cmp( SP, Rscratch );
452 __ brx( Assembler::greater, false, Assembler::pt, after_frame_check );
453 __ delayed()->nop();
454
455 // Save the return address as the exception pc
456 __ st_ptr(O7, saved_exception_pc);
457
458 // the stack will overflow, throw an exception
459 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
460
461 // if you get to here, then there is enough stack space
462 __ bind( after_frame_check );
463 }
464
465
466 //
467 // Generate a fixed interpreter frame. This is identical setup for interpreted
468 // methods and for native methods hence the shared code.
469
470 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
471 //
472 //
473 // The entry code sets up a new interpreter frame in 4 steps:
474 //
475 // 1) Increase caller's SP by for the extra local space needed:
476 // (check for overflow)
477 // Efficient implementation of xload/xstore bytecodes requires
478 // that arguments and non-argument locals are in a contigously
479 // addressable memory block => non-argument locals must be
480 // allocated in the caller's frame.
481 //
482 // 2) Create a new stack frame and register window:
483 // The new stack frame must provide space for the standard
484 // register save area, the maximum java expression stack size,
485 // the monitor slots (0 slots initially), and some frame local
486 // scratch locations.
487 //
488 // 3) The following interpreter activation registers must be setup:
489 // Lesp : expression stack pointer
490 // Lbcp : bytecode pointer
491 // Lmethod : method
492 // Llocals : locals pointer
493 // Lmonitors : monitor pointer
494 // LcpoolCache: constant pool cache
495 //
496 // 4) Initialize the non-argument locals if necessary:
497 // Non-argument locals may need to be initialized to NULL
498 // for GC to work. If the oop-map information is accurate
499 // (in the absence of the JSR problem), no initialization
500 // is necessary.
501 //
502 // (gri - 2/25/2000)
503
504
505 const Address size_of_parameters(G5_method, methodOopDesc::size_of_parameters_offset());
506 const Address size_of_locals (G5_method, methodOopDesc::size_of_locals_offset());
507 const Address max_stack (G5_method, methodOopDesc::max_stack_offset());
508 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
509
510 const int extra_space =
511 rounded_vm_local_words + // frame local scratch space
512 //6815692//methodOopDesc::extra_stack_words() + // extra push slots for MH adapters
513 frame::memory_parameter_word_sp_offset + // register save area
514 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
515
516 const Register Glocals_size = G3;
517 const Register Otmp1 = O3;
518 const Register Otmp2 = O4;
519 // Lscratch can't be used as a temporary because the call_stub uses
520 // it to assert that the stack frame was setup correctly.
521
522 __ lduh( size_of_parameters, Glocals_size);
523
524 // Gargs points to first local + BytesPerWord
525 // Set the saved SP after the register window save
526 //
527 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
528 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
529 __ add(Gargs, Otmp1, Gargs);
530
531 if (native_call) {
532 __ calc_mem_param_words( Glocals_size, Gframe_size );
533 __ add( Gframe_size, extra_space, Gframe_size);
534 __ round_to( Gframe_size, WordsPerLong );
535 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
536 } else {
537
538 //
539 // Compute number of locals in method apart from incoming parameters
540 //
541 __ lduh( size_of_locals, Otmp1 );
542 __ sub( Otmp1, Glocals_size, Glocals_size );
543 __ round_to( Glocals_size, WordsPerLong );
544 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
545
546 // see if the frame is greater than one page in size. If so,
547 // then we need to verify there is enough stack space remaining
548 // Frame_size = (max_stack + extra_space) * BytesPerWord;
549 __ lduh( max_stack, Gframe_size );
550 __ add( Gframe_size, extra_space, Gframe_size );
551 __ round_to( Gframe_size, WordsPerLong );
552 __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
553
554 // Add in java locals size for stack overflow check only
555 __ add( Gframe_size, Glocals_size, Gframe_size );
556
557 const Register Otmp2 = O4;
558 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
559 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
560
561 __ sub( Gframe_size, Glocals_size, Gframe_size);
562
563 //
564 // bump SP to accomodate the extra locals
565 //
566 __ sub( SP, Glocals_size, SP );
567 }
568
569 //
570 // now set up a stack frame with the size computed above
571 //
572 __ neg( Gframe_size );
573 __ save( SP, Gframe_size, SP );
574
575 //
576 // now set up all the local cache registers
577 //
578 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
579 // that all present references to Lbyte_code initialize the register
580 // immediately before use
581 if (native_call) {
582 __ mov(G0, Lbcp);
583 } else {
584 __ ld_ptr(G5_method, methodOopDesc::const_offset(), Lbcp);
585 __ add(Lbcp, in_bytes(constMethodOopDesc::codes_offset()), Lbcp);
586 }
587 __ mov( G5_method, Lmethod); // set Lmethod
588 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
589 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
590 #ifdef _LP64
591 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
592 #endif
593 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
594
595 // setup interpreter activation registers
596 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
597
598 if (ProfileInterpreter) {
599 #ifdef FAST_DISPATCH
600 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
601 // they both use I2.
602 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
603 #endif // FAST_DISPATCH
604 __ set_method_data_pointer();
605 }
606
607 }
608
609 // Empty method, generate a very fast return.
610
611 address InterpreterGenerator::generate_empty_entry(void) {
612
613 // A method that does nother but return...
614
615 address entry = __ pc();
616 Label slow_path;
617
618 __ verify_oop(G5_method);
619
620 // do nothing for empty methods (do not even increment invocation counter)
621 if ( UseFastEmptyMethods) {
622 // If we need a safepoint check, generate full interpreter entry.
623 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
624 __ set(sync_state, G3_scratch);
625 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
626 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
627 __ delayed()->nop();
628
629 // Code: _return
630 __ retl();
631 __ delayed()->mov(O5_savedSP, SP);
632
633 __ bind(slow_path);
634 (void) generate_normal_entry(false);
635
636 return entry;
637 }
638 return NULL;
639 }
640
641 // Call an accessor method (assuming it is resolved, otherwise drop into
642 // vanilla (slow path) entry
643
644 // Generates code to elide accessor methods
645 // Uses G3_scratch and G1_scratch as scratch
646 address InterpreterGenerator::generate_accessor_entry(void) {
647
648 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
649 // parameter size = 1
650 // Note: We can only use this code if the getfield has been resolved
651 // and if we don't have a null-pointer exception => check for
652 // these conditions first and use slow path if necessary.
653 address entry = __ pc();
654 Label slow_path;
655
656
657 // XXX: for compressed oops pointer loading and decoding doesn't fit in
658 // delay slot and damages G1
659 if ( UseFastAccessorMethods && !UseCompressedOops ) {
660 // Check if we need to reach a safepoint and generate full interpreter
661 // frame if so.
662 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
663 __ load_contents(sync_state, G3_scratch);
664 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
665 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
666 __ delayed()->nop();
667
668 // Check if local 0 != NULL
669 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
670 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
671 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
672 __ delayed()->nop();
673
674
675 // read first instruction word and extract bytecode @ 1 and index @ 2
676 // get first 4 bytes of the bytecodes (big endian!)
677 __ ld_ptr(G5_method, methodOopDesc::const_offset(), G1_scratch);
678 __ ld(G1_scratch, constMethodOopDesc::codes_offset(), G1_scratch);
679
680 // move index @ 2 far left then to the right most two bytes.
681 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
682 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
683 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
684
685 // get constant pool cache
686 __ ld_ptr(G5_method, methodOopDesc::constants_offset(), G3_scratch);
687 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
688
689 // get specific constant pool cache entry
690 __ add(G3_scratch, G1_scratch, G3_scratch);
691
692 // Check the constant Pool cache entry to see if it has been resolved.
693 // If not, need the slow path.
694 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
695 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
696 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
697 __ and3(G1_scratch, 0xFF, G1_scratch);
698 __ cmp(G1_scratch, Bytecodes::_getfield);
699 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
700 __ delayed()->nop();
701
702 // Get the type and return field offset from the constant pool cache
703 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
704 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
705
706 Label xreturn_path;
707 // Need to differentiate between igetfield, agetfield, bgetfield etc.
708 // because they are different sizes.
709 // Get the type from the constant pool cache
710 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
711 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
712 ConstantPoolCacheEntry::verify_tosBits();
713 __ cmp(G1_scratch, atos );
714 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
715 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
716 __ cmp(G1_scratch, itos);
717 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
718 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
719 __ cmp(G1_scratch, stos);
720 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
721 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
722 __ cmp(G1_scratch, ctos);
723 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
724 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
725 #ifdef ASSERT
726 __ cmp(G1_scratch, btos);
727 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
728 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
729 __ should_not_reach_here();
730 #endif
731 __ ldsb(Otos_i, G3_scratch, Otos_i);
732 __ bind(xreturn_path);
733
734 // _ireturn/_areturn
735 __ retl(); // return from leaf routine
736 __ delayed()->mov(O5_savedSP, SP);
737
738 // Generate regular method entry
739 __ bind(slow_path);
740 (void) generate_normal_entry(false);
741 return entry;
742 }
743 return NULL;
744 }
745
746 //
747 // Interpreter stub for calling a native method. (asm interpreter)
748 // This sets up a somewhat different looking stack for calling the native method
749 // than the typical interpreter frame setup.
750 //
751
752 address InterpreterGenerator::generate_native_entry(bool synchronized) {
753 address entry = __ pc();
754
755 // the following temporary registers are used during frame creation
756 const Register Gtmp1 = G3_scratch ;
757 const Register Gtmp2 = G1_scratch;
758 bool inc_counter = UseCompiler || CountCompiledCalls;
759
760 // make sure registers are different!
761 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
762
763 const Address Laccess_flags(Lmethod, methodOopDesc::access_flags_offset());
764
765 __ verify_oop(G5_method);
766
767 const Register Glocals_size = G3;
768 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
769
770 // make sure method is native & not abstract
771 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
772 #ifdef ASSERT
773 __ ld(G5_method, methodOopDesc::access_flags_offset(), Gtmp1);
774 {
775 Label L;
776 __ btst(JVM_ACC_NATIVE, Gtmp1);
777 __ br(Assembler::notZero, false, Assembler::pt, L);
778 __ delayed()->nop();
779 __ stop("tried to execute non-native method as native");
780 __ bind(L);
781 }
782 { Label L;
783 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
784 __ br(Assembler::zero, false, Assembler::pt, L);
785 __ delayed()->nop();
786 __ stop("tried to execute abstract method as non-abstract");
787 __ bind(L);
788 }
789 #endif // ASSERT
790
791 // generate the code to allocate the interpreter stack frame
792 generate_fixed_frame(true);
793
794 //
795 // No locals to initialize for native method
796 //
797
798 // this slot will be set later, we initialize it to null here just in
799 // case we get a GC before the actual value is stored later
800 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
801
802 const Address do_not_unlock_if_synchronized(G2_thread,
803 JavaThread::do_not_unlock_if_synchronized_offset());
804 // Since at this point in the method invocation the exception handler
805 // would try to exit the monitor of synchronized methods which hasn't
806 // been entered yet, we set the thread local variable
807 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
808 // runtime, exception handling i.e. unlock_if_synchronized_method will
809 // check this thread local flag.
810 // This flag has two effects, one is to force an unwind in the topmost
811 // interpreter frame and not perform an unlock while doing so.
812
813 __ movbool(true, G3_scratch);
814 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
815
816 // increment invocation counter and check for overflow
817 //
818 // Note: checking for negative value instead of overflow
819 // so we have a 'sticky' overflow test (may be of
820 // importance as soon as we have true MT/MP)
821 Label invocation_counter_overflow;
822 Label Lcontinue;
823 if (inc_counter) {
824 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
825
826 }
827 __ bind(Lcontinue);
828
829 bang_stack_shadow_pages(true);
830
831 // reset the _do_not_unlock_if_synchronized flag
832 __ stbool(G0, do_not_unlock_if_synchronized);
833
834 // check for synchronized methods
835 // Must happen AFTER invocation_counter check and stack overflow check,
836 // so method is not locked if overflows.
837
838 if (synchronized) {
839 lock_method();
840 } else {
841 #ifdef ASSERT
842 { Label ok;
843 __ ld(Laccess_flags, O0);
844 __ btst(JVM_ACC_SYNCHRONIZED, O0);
845 __ br( Assembler::zero, false, Assembler::pt, ok);
846 __ delayed()->nop();
847 __ stop("method needs synchronization");
848 __ bind(ok);
849 }
850 #endif // ASSERT
851 }
852
853
854 // start execution
855 __ verify_thread();
856
857 // JVMTI support
858 __ notify_method_entry();
859
860 // native call
861
862 // (note that O0 is never an oop--at most it is a handle)
863 // It is important not to smash any handles created by this call,
864 // until any oop handle in O0 is dereferenced.
865
866 // (note that the space for outgoing params is preallocated)
867
868 // get signature handler
869 { Label L;
870 Address signature_handler(Lmethod, methodOopDesc::signature_handler_offset());
871 __ ld_ptr(signature_handler, G3_scratch);
872 __ tst(G3_scratch);
873 __ brx(Assembler::notZero, false, Assembler::pt, L);
874 __ delayed()->nop();
875 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
876 __ ld_ptr(signature_handler, G3_scratch);
877 __ bind(L);
878 }
879
880 // Push a new frame so that the args will really be stored in
881 // Copy a few locals across so the new frame has the variables
882 // we need but these values will be dead at the jni call and
883 // therefore not gc volatile like the values in the current
884 // frame (Lmethod in particular)
885
886 // Flush the method pointer to the register save area
887 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
888 __ mov(Llocals, O1);
889
890 // calculate where the mirror handle body is allocated in the interpreter frame:
891 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
892
893 // Calculate current frame size
894 __ sub(SP, FP, O3); // Calculate negative of current frame size
895 __ save(SP, O3, SP); // Allocate an identical sized frame
896
897 // Note I7 has leftover trash. Slow signature handler will fill it in
898 // should we get there. Normal jni call will set reasonable last_Java_pc
899 // below (and fix I7 so the stack trace doesn't have a meaningless frame
900 // in it).
901
902 // Load interpreter frame's Lmethod into same register here
903
904 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
905
906 __ mov(I1, Llocals);
907 __ mov(I2, Lscratch2); // save the address of the mirror
908
909
910 // ONLY Lmethod and Llocals are valid here!
911
912 // call signature handler, It will move the arg properly since Llocals in current frame
913 // matches that in outer frame
914
915 __ callr(G3_scratch, 0);
916 __ delayed()->nop();
917
918 // Result handler is in Lscratch
919
920 // Reload interpreter frame's Lmethod since slow signature handler may block
921 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
922
923 { Label not_static;
924
925 __ ld(Laccess_flags, O0);
926 __ btst(JVM_ACC_STATIC, O0);
927 __ br( Assembler::zero, false, Assembler::pt, not_static);
928 // get native function entry point(O0 is a good temp until the very end)
929 __ delayed()->ld_ptr(Lmethod, in_bytes(methodOopDesc::native_function_offset()), O0);
930 // for static methods insert the mirror argument
931 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
932
933 __ ld_ptr(Lmethod, methodOopDesc:: constants_offset(), O1);
934 __ ld_ptr(O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
935 __ ld_ptr(O1, mirror_offset, O1);
936 #ifdef ASSERT
937 if (!PrintSignatureHandlers) // do not dirty the output with this
938 { Label L;
939 __ tst(O1);
940 __ brx(Assembler::notZero, false, Assembler::pt, L);
941 __ delayed()->nop();
942 __ stop("mirror is missing");
943 __ bind(L);
944 }
945 #endif // ASSERT
946 __ st_ptr(O1, Lscratch2, 0);
947 __ mov(Lscratch2, O1);
948 __ bind(not_static);
949 }
950
951 // At this point, arguments have been copied off of stack into
952 // their JNI positions, which are O1..O5 and SP[68..].
953 // Oops are boxed in-place on the stack, with handles copied to arguments.
954 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
955
956 #ifdef ASSERT
957 { Label L;
958 __ tst(O0);
959 __ brx(Assembler::notZero, false, Assembler::pt, L);
960 __ delayed()->nop();
961 __ stop("native entry point is missing");
962 __ bind(L);
963 }
964 #endif // ASSERT
965
966 //
967 // setup the frame anchor
968 //
969 // The scavenge function only needs to know that the PC of this frame is
970 // in the interpreter method entry code, it doesn't need to know the exact
971 // PC and hence we can use O7 which points to the return address from the
972 // previous call in the code stream (signature handler function)
973 //
974 // The other trick is we set last_Java_sp to FP instead of the usual SP because
975 // we have pushed the extra frame in order to protect the volatile register(s)
976 // in that frame when we return from the jni call
977 //
978
979 __ set_last_Java_frame(FP, O7);
980 __ mov(O7, I7); // make dummy interpreter frame look like one above,
981 // not meaningless information that'll confuse me.
982
983 // flush the windows now. We don't care about the current (protection) frame
984 // only the outer frames
985
986 __ flush_windows();
987
988 // mark windows as flushed
989 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
990 __ set(JavaFrameAnchor::flushed, G3_scratch);
991 __ st(G3_scratch, flags);
992
993 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
994
995 Address thread_state(G2_thread, JavaThread::thread_state_offset());
996 #ifdef ASSERT
997 { Label L;
998 __ ld(thread_state, G3_scratch);
999 __ cmp(G3_scratch, _thread_in_Java);
1000 __ br(Assembler::equal, false, Assembler::pt, L);
1001 __ delayed()->nop();
1002 __ stop("Wrong thread state in native stub");
1003 __ bind(L);
1004 }
1005 #endif // ASSERT
1006 __ set(_thread_in_native, G3_scratch);
1007 __ st(G3_scratch, thread_state);
1008
1009 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1010 __ save_thread(L7_thread_cache); // save Gthread
1011 __ callr(O0, 0);
1012 __ delayed()->
1013 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1014
1015 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1016
1017 __ restore_thread(L7_thread_cache); // restore G2_thread
1018 __ reinit_heapbase();
1019
1020 // must we block?
1021
1022 // Block, if necessary, before resuming in _thread_in_Java state.
1023 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1024 { Label no_block;
1025 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1026
1027 // Switch thread to "native transition" state before reading the synchronization state.
1028 // This additional state is necessary because reading and testing the synchronization
1029 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1030 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1031 // VM thread changes sync state to synchronizing and suspends threads for GC.
1032 // Thread A is resumed to finish this native method, but doesn't block here since it
1033 // didn't see any synchronization is progress, and escapes.
1034 __ set(_thread_in_native_trans, G3_scratch);
1035 __ st(G3_scratch, thread_state);
1036 if(os::is_MP()) {
1037 if (UseMembar) {
1038 // Force this write out before the read below
1039 __ membar(Assembler::StoreLoad);
1040 } else {
1041 // Write serialization page so VM thread can do a pseudo remote membar.
1042 // We use the current thread pointer to calculate a thread specific
1043 // offset to write to within the page. This minimizes bus traffic
1044 // due to cache line collision.
1045 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1046 }
1047 }
1048 __ load_contents(sync_state, G3_scratch);
1049 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1050
1051 Label L;
1052 __ br(Assembler::notEqual, false, Assembler::pn, L);
1053 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1054 __ cmp(G3_scratch, 0);
1055 __ br(Assembler::equal, false, Assembler::pt, no_block);
1056 __ delayed()->nop();
1057 __ bind(L);
1058
1059 // Block. Save any potential method result value before the operation and
1060 // use a leaf call to leave the last_Java_frame setup undisturbed.
1061 save_native_result();
1062 __ call_VM_leaf(L7_thread_cache,
1063 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1064 G2_thread);
1065
1066 // Restore any method result value
1067 restore_native_result();
1068 __ bind(no_block);
1069 }
1070
1071 // Clear the frame anchor now
1072
1073 __ reset_last_Java_frame();
1074
1075 // Move the result handler address
1076 __ mov(Lscratch, G3_scratch);
1077 // return possible result to the outer frame
1078 #ifndef __LP64
1079 __ mov(O0, I0);
1080 __ restore(O1, G0, O1);
1081 #else
1082 __ restore(O0, G0, O0);
1083 #endif /* __LP64 */
1084
1085 // Move result handler to expected register
1086 __ mov(G3_scratch, Lscratch);
1087
1088 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1089 // switch to thread_in_Java.
1090
1091 __ set(_thread_in_Java, G3_scratch);
1092 __ st(G3_scratch, thread_state);
1093
1094 // reset handle block
1095 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1096 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1097
1098 // If we have an oop result store it where it will be safe for any further gc
1099 // until we return now that we've released the handle it might be protected by
1100
1101 {
1102 Label no_oop, store_result;
1103
1104 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1105 __ cmp(G3_scratch, Lscratch);
1106 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
1107 __ delayed()->nop();
1108 __ addcc(G0, O0, O0);
1109 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1110 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1111 __ mov(G0, O0);
1112
1113 __ bind(store_result);
1114 // Store it where gc will look for it and result handler expects it.
1115 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1116
1117 __ bind(no_oop);
1118
1119 }
1120
1121
1122 // handle exceptions (exception handling will handle unlocking!)
1123 { Label L;
1124 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1125 __ ld_ptr(exception_addr, Gtemp);
1126 __ tst(Gtemp);
1127 __ brx(Assembler::equal, false, Assembler::pt, L);
1128 __ delayed()->nop();
1129 // Note: This could be handled more efficiently since we know that the native
1130 // method doesn't have an exception handler. We could directly return
1131 // to the exception handler for the caller.
1132 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1133 __ should_not_reach_here();
1134 __ bind(L);
1135 }
1136
1137 // JVMTI support (preserves thread register)
1138 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1139
1140 if (synchronized) {
1141 // save and restore any potential method result value around the unlocking operation
1142 save_native_result();
1143
1144 __ add( __ top_most_monitor(), O1);
1145 __ unlock_object(O1);
1146
1147 restore_native_result();
1148 }
1149
1150 #if defined(COMPILER2) && !defined(_LP64)
1151
1152 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1153 // or compiled so just be safe.
1154
1155 __ sllx(O0, 32, G1); // Shift bits into high G1
1156 __ srl (O1, 0, O1); // Zero extend O1
1157 __ or3 (O1, G1, G1); // OR 64 bits into G1
1158
1159 #endif /* COMPILER2 && !_LP64 */
1160
1161 // dispose of return address and remove activation
1162 #ifdef ASSERT
1163 {
1164 Label ok;
1165 __ cmp(I5_savedSP, FP);
1166 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1167 __ delayed()->nop();
1168 __ stop("bad I5_savedSP value");
1169 __ should_not_reach_here();
1170 __ bind(ok);
1171 }
1172 #endif
1173 if (TraceJumps) {
1174 // Move target to register that is recordable
1175 __ mov(Lscratch, G3_scratch);
1176 __ JMP(G3_scratch, 0);
1177 } else {
1178 __ jmp(Lscratch, 0);
1179 }
1180 __ delayed()->nop();
1181
1182
1183 if (inc_counter) {
1184 // handle invocation counter overflow
1185 __ bind(invocation_counter_overflow);
1186 generate_counter_overflow(Lcontinue);
1187 }
1188
1189
1190
1191 return entry;
1192 }
1193
1194
1195 // Generic method entry to (asm) interpreter
1196 //------------------------------------------------------------------------------------------------------------------------
1197 //
1198 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1199 address entry = __ pc();
1200
1201 bool inc_counter = UseCompiler || CountCompiledCalls;
1202
1203 // the following temporary registers are used during frame creation
1204 const Register Gtmp1 = G3_scratch ;
1205 const Register Gtmp2 = G1_scratch;
1206
1207 // make sure registers are different!
1208 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1209
1210 const Address size_of_parameters(G5_method, methodOopDesc::size_of_parameters_offset());
1211 const Address size_of_locals (G5_method, methodOopDesc::size_of_locals_offset());
1212 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1213 // and use in the asserts.
1214 const Address access_flags (Lmethod, methodOopDesc::access_flags_offset());
1215
1216 __ verify_oop(G5_method);
1217
1218 const Register Glocals_size = G3;
1219 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1220
1221 // make sure method is not native & not abstract
1222 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1223 #ifdef ASSERT
1224 __ ld(G5_method, methodOopDesc::access_flags_offset(), Gtmp1);
1225 {
1226 Label L;
1227 __ btst(JVM_ACC_NATIVE, Gtmp1);
1228 __ br(Assembler::zero, false, Assembler::pt, L);
1229 __ delayed()->nop();
1230 __ stop("tried to execute native method as non-native");
1231 __ bind(L);
1232 }
1233 { Label L;
1234 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1235 __ br(Assembler::zero, false, Assembler::pt, L);
1236 __ delayed()->nop();
1237 __ stop("tried to execute abstract method as non-abstract");
1238 __ bind(L);
1239 }
1240 #endif // ASSERT
1241
1242 // generate the code to allocate the interpreter stack frame
1243
1244 generate_fixed_frame(false);
1245
1246 #ifdef FAST_DISPATCH
1247 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1248 // set bytecode dispatch table base
1249 #endif
1250
1251 //
1252 // Code to initialize the extra (i.e. non-parm) locals
1253 //
1254 Register init_value = noreg; // will be G0 if we must clear locals
1255 // The way the code was setup before zerolocals was always true for vanilla java entries.
1256 // It could only be false for the specialized entries like accessor or empty which have
1257 // no extra locals so the testing was a waste of time and the extra locals were always
1258 // initialized. We removed this extra complication to already over complicated code.
1259
1260 init_value = G0;
1261 Label clear_loop;
1262
1263 // NOTE: If you change the frame layout, this code will need to
1264 // be updated!
1265 __ lduh( size_of_locals, O2 );
1266 __ lduh( size_of_parameters, O1 );
1267 __ sll( O2, Interpreter::logStackElementSize, O2);
1268 __ sll( O1, Interpreter::logStackElementSize, O1 );
1269 __ sub( Llocals, O2, O2 );
1270 __ sub( Llocals, O1, O1 );
1271
1272 __ bind( clear_loop );
1273 __ inc( O2, wordSize );
1274
1275 __ cmp( O2, O1 );
1276 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1277 __ delayed()->st_ptr( init_value, O2, 0 );
1278
1279 const Address do_not_unlock_if_synchronized(G2_thread,
1280 JavaThread::do_not_unlock_if_synchronized_offset());
1281 // Since at this point in the method invocation the exception handler
1282 // would try to exit the monitor of synchronized methods which hasn't
1283 // been entered yet, we set the thread local variable
1284 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1285 // runtime, exception handling i.e. unlock_if_synchronized_method will
1286 // check this thread local flag.
1287 __ movbool(true, G3_scratch);
1288 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1289
1290 // increment invocation counter and check for overflow
1291 //
1292 // Note: checking for negative value instead of overflow
1293 // so we have a 'sticky' overflow test (may be of
1294 // importance as soon as we have true MT/MP)
1295 Label invocation_counter_overflow;
1296 Label profile_method;
1297 Label profile_method_continue;
1298 Label Lcontinue;
1299 if (inc_counter) {
1300 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1301 if (ProfileInterpreter) {
1302 __ bind(profile_method_continue);
1303 }
1304 }
1305 __ bind(Lcontinue);
1306
1307 bang_stack_shadow_pages(false);
1308
1309 // reset the _do_not_unlock_if_synchronized flag
1310 __ stbool(G0, do_not_unlock_if_synchronized);
1311
1312 // check for synchronized methods
1313 // Must happen AFTER invocation_counter check and stack overflow check,
1314 // so method is not locked if overflows.
1315
1316 if (synchronized) {
1317 lock_method();
1318 } else {
1319 #ifdef ASSERT
1320 { Label ok;
1321 __ ld(access_flags, O0);
1322 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1323 __ br( Assembler::zero, false, Assembler::pt, ok);
1324 __ delayed()->nop();
1325 __ stop("method needs synchronization");
1326 __ bind(ok);
1327 }
1328 #endif // ASSERT
1329 }
1330
1331 // start execution
1332
1333 __ verify_thread();
1334
1335 // jvmti support
1336 __ notify_method_entry();
1337
1338 // start executing instructions
1339 __ dispatch_next(vtos);
1340
1341
1342 if (inc_counter) {
1343 if (ProfileInterpreter) {
1344 // We have decided to profile this method in the interpreter
1345 __ bind(profile_method);
1346
1347 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method), Lbcp, true);
1348
1349 #ifdef ASSERT
1350 __ tst(O0);
1351 __ breakpoint_trap(Assembler::notEqual);
1352 #endif
1353
1354 __ set_method_data_pointer();
1355
1356 __ ba(false, profile_method_continue);
1357 __ delayed()->nop();
1358 }
1359
1360 // handle invocation counter overflow
1361 __ bind(invocation_counter_overflow);
1362 generate_counter_overflow(Lcontinue);
1363 }
1364
1365
1366 return entry;
1367 }
1368
1369
1370 //----------------------------------------------------------------------------------------------------
1371 // Entry points & stack frame layout
1372 //
1373 // Here we generate the various kind of entries into the interpreter.
1374 // The two main entry type are generic bytecode methods and native call method.
1375 // These both come in synchronized and non-synchronized versions but the
1376 // frame layout they create is very similar. The other method entry
1377 // types are really just special purpose entries that are really entry
1378 // and interpretation all in one. These are for trivial methods like
1379 // accessor, empty, or special math methods.
1380 //
1381 // When control flow reaches any of the entry types for the interpreter
1382 // the following holds ->
1383 //
1384 // C2 Calling Conventions:
1385 //
1386 // The entry code below assumes that the following registers are set
1387 // when coming in:
1388 // G5_method: holds the methodOop of the method to call
1389 // Lesp: points to the TOS of the callers expression stack
1390 // after having pushed all the parameters
1391 //
1392 // The entry code does the following to setup an interpreter frame
1393 // pop parameters from the callers stack by adjusting Lesp
1394 // set O0 to Lesp
1395 // compute X = (max_locals - num_parameters)
1396 // bump SP up by X to accomadate the extra locals
1397 // compute X = max_expression_stack
1398 // + vm_local_words
1399 // + 16 words of register save area
1400 // save frame doing a save sp, -X, sp growing towards lower addresses
1401 // set Lbcp, Lmethod, LcpoolCache
1402 // set Llocals to i0
1403 // set Lmonitors to FP - rounded_vm_local_words
1404 // set Lesp to Lmonitors - 4
1405 //
1406 // The frame has now been setup to do the rest of the entry code
1407
1408 // Try this optimization: Most method entries could live in a
1409 // "one size fits all" stack frame without all the dynamic size
1410 // calculations. It might be profitable to do all this calculation
1411 // statically and approximately for "small enough" methods.
1412
1413 //-----------------------------------------------------------------------------------------------
1414
1415 // C1 Calling conventions
1416 //
1417 // Upon method entry, the following registers are setup:
1418 //
1419 // g2 G2_thread: current thread
1420 // g5 G5_method: method to activate
1421 // g4 Gargs : pointer to last argument
1422 //
1423 //
1424 // Stack:
1425 //
1426 // +---------------+ <--- sp
1427 // | |
1428 // : reg save area :
1429 // | |
1430 // +---------------+ <--- sp + 0x40
1431 // | |
1432 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1433 // | |
1434 // +---------------+ <--- sp + 0x5c
1435 // | |
1436 // : free :
1437 // | |
1438 // +---------------+ <--- Gargs
1439 // | |
1440 // : arguments :
1441 // | |
1442 // +---------------+
1443 // | |
1444 //
1445 //
1446 //
1447 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1448 //
1449 // +---------------+ <--- sp
1450 // | |
1451 // : reg save area :
1452 // | |
1453 // +---------------+ <--- sp + 0x40
1454 // | |
1455 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1456 // | |
1457 // +---------------+ <--- sp + 0x5c
1458 // | |
1459 // : :
1460 // | | <--- Lesp
1461 // +---------------+ <--- Lmonitors (fp - 0x18)
1462 // | VM locals |
1463 // +---------------+ <--- fp
1464 // | |
1465 // : reg save area :
1466 // | |
1467 // +---------------+ <--- fp + 0x40
1468 // | |
1469 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1470 // | |
1471 // +---------------+ <--- fp + 0x5c
1472 // | |
1473 // : free :
1474 // | |
1475 // +---------------+
1476 // | |
1477 // : nonarg locals :
1478 // | |
1479 // +---------------+
1480 // | |
1481 // : arguments :
1482 // | | <--- Llocals
1483 // +---------------+ <--- Gargs
1484 // | |
1485
1486 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1487
1488 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1489 // expression stack, the callee will have callee_extra_locals (so we can account for
1490 // frame extension) and monitor_size for monitors. Basically we need to calculate
1491 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1492 //
1493 //
1494 // The big complicating thing here is that we must ensure that the stack stays properly
1495 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1496 // needs to be aligned for). We are given that the sp (fp) is already aligned by
1497 // the caller so we must ensure that it is properly aligned for our callee.
1498 //
1499 const int rounded_vm_local_words =
1500 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1501 // callee_locals and max_stack are counts, not the size in frame.
1502 const int locals_size =
1503 round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1504 const int max_stack_words = max_stack * Interpreter::stackElementWords;
1505 return (round_to((max_stack_words
1506 //6815692//+ methodOopDesc::extra_stack_words()
1507 + rounded_vm_local_words
1508 + frame::memory_parameter_word_sp_offset), WordsPerLong)
1509 // already rounded
1510 + locals_size + monitor_size);
1511 }
1512
1513 // How much stack a method top interpreter activation needs in words.
1514 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
1515
1516 // See call_stub code
1517 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
1518 WordsPerLong); // 7 + register save area
1519
1520 // Save space for one monitor to get into the interpreted method in case
1521 // the method is synchronized
1522 int monitor_size = method->is_synchronized() ?
1523 1*frame::interpreter_frame_monitor_size() : 0;
1524 return size_activation_helper(method->max_locals(), method->max_stack(),
1525 monitor_size) + call_stub_size;
1526 }
1527
1528 int AbstractInterpreter::layout_activation(methodOop method,
1529 int tempcount,
1530 int popframe_extra_args,
1531 int moncount,
1532 int callee_param_count,
1533 int callee_local_count,
1534 frame* caller,
1535 frame* interpreter_frame,
1536 bool is_top_frame) {
1537 // Note: This calculation must exactly parallel the frame setup
1538 // in InterpreterGenerator::generate_fixed_frame.
1539 // If f!=NULL, set up the following variables:
1540 // - Lmethod
1541 // - Llocals
1542 // - Lmonitors (to the indicated number of monitors)
1543 // - Lesp (to the indicated number of temps)
1544 // The frame f (if not NULL) on entry is a description of the caller of the frame
1545 // we are about to layout. We are guaranteed that we will be able to fill in a
1546 // new interpreter frame as its callee (i.e. the stack space is allocated and
1547 // the amount was determined by an earlier call to this method with f == NULL).
1548 // On return f (if not NULL) while describe the interpreter frame we just layed out.
1549
1550 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
1551 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1552
1553 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1554 //
1555 // Note: if you look closely this appears to be doing something much different
1556 // than generate_fixed_frame. What is happening is this. On sparc we have to do
1557 // this dance with interpreter_sp_adjustment because the window save area would
1558 // appear just below the bottom (tos) of the caller's java expression stack. Because
1559 // the interpreter want to have the locals completely contiguous generate_fixed_frame
1560 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1561 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1562 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1563 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1564 // because the oldest frame would have adjust its callers frame and yet that frame
1565 // already exists and isn't part of this array of frames we are unpacking. So at first
1566 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1567 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1568 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1569 // add up. It does seem like it simpler to account for the adjustment here (and remove the
1570 // callee... parameters here). However this would mean that this routine would have to take
1571 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1572 // and run the calling loop in the reverse order. This would also would appear to mean making
1573 // this code aware of what the interactions are when that initial caller fram was an osr or
1574 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
1575 // there is no sense in messing working code.
1576 //
1577
1578 int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
1579 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1580
1581 int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
1582 monitor_size);
1583
1584 if (interpreter_frame != NULL) {
1585 // The skeleton frame must already look like an interpreter frame
1586 // even if not fully filled out.
1587 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1588
1589 intptr_t* fp = interpreter_frame->fp();
1590
1591 JavaThread* thread = JavaThread::current();
1592 RegisterMap map(thread, false);
1593 // More verification that skeleton frame is properly walkable
1594 assert(fp == caller->sp(), "fp must match");
1595
1596 intptr_t* montop = fp - rounded_vm_local_words;
1597
1598 // preallocate monitors (cf. __ add_monitor_to_stack)
1599 intptr_t* monitors = montop - monitor_size;
1600
1601 // preallocate stack space
1602 intptr_t* esp = monitors - 1 -
1603 (tempcount * Interpreter::stackElementWords) -
1604 popframe_extra_args;
1605
1606 int local_words = method->max_locals() * Interpreter::stackElementWords;
1607 int parm_words = method->size_of_parameters() * Interpreter::stackElementWords;
1608 NEEDS_CLEANUP;
1609 intptr_t* locals;
1610 if (caller->is_interpreted_frame()) {
1611 // Can force the locals area to end up properly overlapping the top of the expression stack.
1612 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1613 // Note that this computation means we replace size_of_parameters() values from the caller
1614 // interpreter frame's expression stack with our argument locals
1615 locals = Lesp_ptr + parm_words;
1616 int delta = local_words - parm_words;
1617 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1618 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1619 } else {
1620 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1621 // Don't have Lesp available; lay out locals block in the caller
1622 // adjacent to the register window save area.
1623 //
1624 // Compiled frames do not allocate a varargs area which is why this if
1625 // statement is needed.
1626 //
1627 if (caller->is_compiled_frame()) {
1628 locals = fp + frame::register_save_words + local_words - 1;
1629 } else {
1630 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1631 }
1632 if (!caller->is_entry_frame()) {
1633 // Caller wants his own SP back
1634 int caller_frame_size = caller->cb()->frame_size();
1635 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1636 }
1637 }
1638 if (TraceDeoptimization) {
1639 if (caller->is_entry_frame()) {
1640 // make sure I5_savedSP and the entry frames notion of saved SP
1641 // agree. This assertion duplicate a check in entry frame code
1642 // but catches the failure earlier.
1643 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1644 "would change callers SP");
1645 }
1646 if (caller->is_entry_frame()) {
1647 tty->print("entry ");
1648 }
1649 if (caller->is_compiled_frame()) {
1650 tty->print("compiled ");
1651 if (caller->is_deoptimized_frame()) {
1652 tty->print("(deopt) ");
1653 }
1654 }
1655 if (caller->is_interpreted_frame()) {
1656 tty->print("interpreted ");
1657 }
1658 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1659 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1660 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1661 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1662 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1663 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1664 tty->print_cr("Llocals = 0x%x", locals);
1665 tty->print_cr("Lesp = 0x%x", esp);
1666 tty->print_cr("Lmonitors = 0x%x", monitors);
1667 }
1668
1669 if (method->max_locals() > 0) {
1670 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1671 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1672 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1673 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1674 }
1675 #ifdef _LP64
1676 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1677 #endif
1678
1679 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
1680 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
1681 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
1682 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
1683 // Llast_SP will be same as SP as there is no adapter space
1684 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1685 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1686 #ifdef FAST_DISPATCH
1687 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1688 #endif
1689
1690
1691 #ifdef ASSERT
1692 BasicObjectLock* mp = (BasicObjectLock*)monitors;
1693
1694 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1695 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1696 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
1697 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1698 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1699
1700 // check bounds
1701 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1702 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1703 assert(lo < monitors && montop <= hi, "monitors in bounds");
1704 assert(lo <= esp && esp < monitors, "esp in bounds");
1705 #endif // ASSERT
1706 }
1707
1708 return raw_frame_size;
1709 }
1710
1711 //----------------------------------------------------------------------------------------------------
1712 // Exceptions
1713 void TemplateInterpreterGenerator::generate_throw_exception() {
1714
1715 // Entry point in previous activation (i.e., if the caller was interpreted)
1716 Interpreter::_rethrow_exception_entry = __ pc();
1717 // O0: exception
1718
1719 // entry point for exceptions thrown within interpreter code
1720 Interpreter::_throw_exception_entry = __ pc();
1721 __ verify_thread();
1722 // expression stack is undefined here
1723 // O0: exception, i.e. Oexception
1724 // Lbcp: exception bcx
1725 __ verify_oop(Oexception);
1726
1727
1728 // expression stack must be empty before entering the VM in case of an exception
1729 __ empty_expression_stack();
1730 // find exception handler address and preserve exception oop
1731 // call C routine to find handler and jump to it
1732 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1733 __ push_ptr(O1); // push exception for exception handler bytecodes
1734
1735 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1736 __ delayed()->nop();
1737
1738
1739 // if the exception is not handled in the current frame
1740 // the frame is removed and the exception is rethrown
1741 // (i.e. exception continuation is _rethrow_exception)
1742 //
1743 // Note: At this point the bci is still the bxi for the instruction which caused
1744 // the exception and the expression stack is empty. Thus, for any VM calls
1745 // at this point, GC will find a legal oop map (with empty expression stack).
1746
1747 // in current activation
1748 // tos: exception
1749 // Lbcp: exception bcp
1750
1751 //
1752 // JVMTI PopFrame support
1753 //
1754
1755 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1756 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1757 // Set the popframe_processing bit in popframe_condition indicating that we are
1758 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1759 // popframe handling cycles.
1760
1761 __ ld(popframe_condition_addr, G3_scratch);
1762 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1763 __ stw(G3_scratch, popframe_condition_addr);
1764
1765 // Empty the expression stack, as in normal exception handling
1766 __ empty_expression_stack();
1767 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1768
1769 {
1770 // Check to see whether we are returning to a deoptimized frame.
1771 // (The PopFrame call ensures that the caller of the popped frame is
1772 // either interpreted or compiled and deoptimizes it if compiled.)
1773 // In this case, we can't call dispatch_next() after the frame is
1774 // popped, but instead must save the incoming arguments and restore
1775 // them after deoptimization has occurred.
1776 //
1777 // Note that we don't compare the return PC against the
1778 // deoptimization blob's unpack entry because of the presence of
1779 // adapter frames in C2.
1780 Label caller_not_deoptimized;
1781 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1782 __ tst(O0);
1783 __ brx(Assembler::notEqual, false, Assembler::pt, caller_not_deoptimized);
1784 __ delayed()->nop();
1785
1786 const Register Gtmp1 = G3_scratch;
1787 const Register Gtmp2 = G1_scratch;
1788
1789 // Compute size of arguments for saving when returning to deoptimized caller
1790 __ lduh(Lmethod, in_bytes(methodOopDesc::size_of_parameters_offset()), Gtmp1);
1791 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1792 __ sub(Llocals, Gtmp1, Gtmp2);
1793 __ add(Gtmp2, wordSize, Gtmp2);
1794 // Save these arguments
1795 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1796 // Inform deoptimization that it is responsible for restoring these arguments
1797 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1798 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1799 __ st(Gtmp1, popframe_condition_addr);
1800
1801 // Return from the current method
1802 // The caller's SP was adjusted upon method entry to accomodate
1803 // the callee's non-argument locals. Undo that adjustment.
1804 __ ret();
1805 __ delayed()->restore(I5_savedSP, G0, SP);
1806
1807 __ bind(caller_not_deoptimized);
1808 }
1809
1810 // Clear the popframe condition flag
1811 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1812
1813 // Get out of the current method (how this is done depends on the particular compiler calling
1814 // convention that the interpreter currently follows)
1815 // The caller's SP was adjusted upon method entry to accomodate
1816 // the callee's non-argument locals. Undo that adjustment.
1817 __ restore(I5_savedSP, G0, SP);
1818 // The method data pointer was incremented already during
1819 // call profiling. We have to restore the mdp for the current bcp.
1820 if (ProfileInterpreter) {
1821 __ set_method_data_pointer_for_bcp();
1822 }
1823 // Resume bytecode interpretation at the current bcp
1824 __ dispatch_next(vtos);
1825 // end of JVMTI PopFrame support
1826
1827 Interpreter::_remove_activation_entry = __ pc();
1828
1829 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1830 __ pop_ptr(Oexception); // get exception
1831
1832 // Intel has the following comment:
1833 //// remove the activation (without doing throws on illegalMonitorExceptions)
1834 // They remove the activation without checking for bad monitor state.
1835 // %%% We should make sure this is the right semantics before implementing.
1836
1837 // %%% changed set_vm_result_2 to set_vm_result and get_vm_result_2 to get_vm_result. Is there a bug here?
1838 __ set_vm_result(Oexception);
1839 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1840
1841 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1842
1843 __ get_vm_result(Oexception);
1844 __ verify_oop(Oexception);
1845
1846 const int return_reg_adjustment = frame::pc_return_offset;
1847 Address issuing_pc_addr(I7, return_reg_adjustment);
1848
1849 // We are done with this activation frame; find out where to go next.
1850 // The continuation point will be an exception handler, which expects
1851 // the following registers set up:
1852 //
1853 // Oexception: exception
1854 // Oissuing_pc: the local call that threw exception
1855 // Other On: garbage
1856 // In/Ln: the contents of the caller's register window
1857 //
1858 // We do the required restore at the last possible moment, because we
1859 // need to preserve some state across a runtime call.
1860 // (Remember that the caller activation is unknown--it might not be
1861 // interpreted, so things like Lscratch are useless in the caller.)
1862
1863 // Although the Intel version uses call_C, we can use the more
1864 // compact call_VM. (The only real difference on SPARC is a
1865 // harmlessly ignored [re]set_last_Java_frame, compared with
1866 // the Intel code which lacks this.)
1867 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1868 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1869 __ super_call_VM_leaf(L7_thread_cache,
1870 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1871 G2_thread, Oissuing_pc->after_save());
1872
1873 // The caller's SP was adjusted upon method entry to accomodate
1874 // the callee's non-argument locals. Undo that adjustment.
1875 __ JMP(O0, 0); // return exception handler in caller
1876 __ delayed()->restore(I5_savedSP, G0, SP);
1877
1878 // (same old exception object is already in Oexception; see above)
1879 // Note that an "issuing PC" is actually the next PC after the call
1880 }
1881
1882
1883 //
1884 // JVMTI ForceEarlyReturn support
1885 //
1886
1887 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1888 address entry = __ pc();
1889
1890 __ empty_expression_stack();
1891 __ load_earlyret_value(state);
1892
1893 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1894 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1895
1896 // Clear the earlyret state
1897 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1898
1899 __ remove_activation(state,
1900 /* throw_monitor_exception */ false,
1901 /* install_monitor_exception */ false);
1902
1903 // The caller's SP was adjusted upon method entry to accomodate
1904 // the callee's non-argument locals. Undo that adjustment.
1905 __ ret(); // return to caller
1906 __ delayed()->restore(I5_savedSP, G0, SP);
1907
1908 return entry;
1909 } // end of JVMTI ForceEarlyReturn support
1910
1911
1912 //------------------------------------------------------------------------------------------------------------------------
1913 // Helper for vtos entry point generation
1914
1915 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) {
1916 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1917 Label L;
1918 aep = __ pc(); __ push_ptr(); __ ba(false, L); __ delayed()->nop();
1919 fep = __ pc(); __ push_f(); __ ba(false, L); __ delayed()->nop();
1920 dep = __ pc(); __ push_d(); __ ba(false, L); __ delayed()->nop();
1921 lep = __ pc(); __ push_l(); __ ba(false, L); __ delayed()->nop();
1922 iep = __ pc(); __ push_i();
1923 bep = cep = sep = iep; // there aren't any
1924 vep = __ pc(); __ bind(L); // fall through
1925 generate_and_dispatch(t);
1926 }
1927
1928 // --------------------------------------------------------------------------------
1929
1930
1931 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1932 : TemplateInterpreterGenerator(code) {
1933 generate_all(); // down here so it can be "virtual"
1934 }
1935
1936 // --------------------------------------------------------------------------------
1937
1938 // Non-product code
1939 #ifndef PRODUCT
1940 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1941 address entry = __ pc();
1942
1943 __ push(state);
1944 __ mov(O7, Lscratch); // protect return address within interpreter
1945
1946 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1947 __ mov( Otos_l2, G3_scratch );
1948 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1949 __ mov(Lscratch, O7); // restore return address
1950 __ pop(state);
1951 __ retl();
1952 __ delayed()->nop();
1953
1954 return entry;
1955 }
1956
1957
1958 // helpers for generate_and_dispatch
1959
1960 void TemplateInterpreterGenerator::count_bytecode() {
1961 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1962 }
1963
1964
1965 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1966 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
1967 }
1968
1969
1970 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1971 AddressLiteral index (&BytecodePairHistogram::_index);
1972 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
1973
1974 // get index, shift out old bytecode, bring in new bytecode, and store it
1975 // _index = (_index >> log2_number_of_codes) |
1976 // (bytecode << log2_number_of_codes);
1977
1978 __ load_contents(index, G4_scratch);
1979 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
1980 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
1981 __ or3( G3_scratch, G4_scratch, G4_scratch );
1982 __ store_contents(G4_scratch, index, G3_scratch);
1983
1984 // bump bucket contents
1985 // _counters[_index] ++;
1986
1987 __ set(counters, G3_scratch); // loads into G3_scratch
1988 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
1989 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
1990 __ ld (G3_scratch, 0, G4_scratch);
1991 __ inc (G4_scratch);
1992 __ st (G4_scratch, 0, G3_scratch);
1993 }
1994
1995
1996 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
1997 // Call a little run-time stub to avoid blow-up for each bytecode.
1998 // The run-time runtime saves the right registers, depending on
1999 // the tosca in-state for the given template.
2000 address entry = Interpreter::trace_code(t->tos_in());
2001 guarantee(entry != NULL, "entry must have been generated");
2002 __ call(entry, relocInfo::none);
2003 __ delayed()->nop();
2004 }
2005
2006
2007 void TemplateInterpreterGenerator::stop_interpreter_at() {
2008 AddressLiteral counter(&BytecodeCounter::_counter_value);
2009 __ load_contents(counter, G3_scratch);
2010 AddressLiteral stop_at(&StopInterpreterAt);
2011 __ load_ptr_contents(stop_at, G4_scratch);
2012 __ cmp(G3_scratch, G4_scratch);
2013 __ breakpoint_trap(Assembler::equal);
2014 }
2015 #endif // not PRODUCT
2016 #endif // !CC_INTERP