1 /*
2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "interpreter/bytecodeHistogram.hpp"
28 #include "interpreter/interpreter.hpp"
29 #include "interpreter/interpreterGenerator.hpp"
30 #include "interpreter/interpreterRuntime.hpp"
31 #include "interpreter/templateTable.hpp"
32 #include "oops/arrayOop.hpp"
33 #include "oops/methodDataOop.hpp"
34 #include "oops/methodOop.hpp"
35 #include "oops/oop.inline.hpp"
36 #include "prims/jvmtiExport.hpp"
37 #include "prims/jvmtiThreadState.hpp"
38 #include "runtime/arguments.hpp"
39 #include "runtime/deoptimization.hpp"
40 #include "runtime/frame.inline.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "runtime/stubRoutines.hpp"
43 #include "runtime/synchronizer.hpp"
44 #include "runtime/timer.hpp"
45 #include "runtime/vframeArray.hpp"
46 #include "utilities/debug.hpp"
47
48 #ifndef CC_INTERP
49 #ifndef FAST_DISPATCH
50 #define FAST_DISPATCH 1
51 #endif
52 #undef FAST_DISPATCH
53
54
55 // Generation of Interpreter
56 //
57 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
58
59
60 #define __ _masm->
61
62
63 //----------------------------------------------------------------------------------------------------
64
65
66 void InterpreterGenerator::save_native_result(void) {
67 // result potentially in O0/O1: save it across calls
68 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
69
70 // result potentially in F0/F1: save it across calls
71 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
72
73 // save and restore any potential method result value around the unlocking operation
74 __ stf(FloatRegisterImpl::D, F0, d_tmp);
75 #ifdef _LP64
76 __ stx(O0, l_tmp);
77 #else
78 __ std(O0, l_tmp);
79 #endif
80 }
81
82 void InterpreterGenerator::restore_native_result(void) {
83 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
84 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
85
86 // Restore any method result value
87 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
88 #ifdef _LP64
89 __ ldx(l_tmp, O0);
90 #else
91 __ ldd(l_tmp, O0);
92 #endif
93 }
94
95 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
96 assert(!pass_oop || message == NULL, "either oop or message but not both");
97 address entry = __ pc();
98 // expression stack must be empty before entering the VM if an exception happened
99 __ empty_expression_stack();
100 // load exception object
101 __ set((intptr_t)name, G3_scratch);
102 if (pass_oop) {
103 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
104 } else {
105 __ set((intptr_t)message, G4_scratch);
106 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
107 }
108 // throw exception
109 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
110 AddressLiteral thrower(Interpreter::throw_exception_entry());
111 __ jump_to(thrower, G3_scratch);
112 __ delayed()->nop();
113 return entry;
114 }
115
116 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
117 address entry = __ pc();
118 // expression stack must be empty before entering the VM if an exception
119 // happened
120 __ empty_expression_stack();
121 // load exception object
122 __ call_VM(Oexception,
123 CAST_FROM_FN_PTR(address,
124 InterpreterRuntime::throw_ClassCastException),
125 Otos_i);
126 __ should_not_reach_here();
127 return entry;
128 }
129
130
131 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
132 address entry = __ pc();
133 // expression stack must be empty before entering the VM if an exception happened
134 __ empty_expression_stack();
135 // convention: expect aberrant index in register G3_scratch, then shuffle the
136 // index to G4_scratch for the VM call
137 __ mov(G3_scratch, G4_scratch);
138 __ set((intptr_t)name, G3_scratch);
139 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
140 __ should_not_reach_here();
141 return entry;
142 }
143
144
145 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
146 address entry = __ pc();
147 // expression stack must be empty before entering the VM if an exception happened
148 __ empty_expression_stack();
149 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
150 __ should_not_reach_here();
151 return entry;
152 }
153
154
155 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) {
156 TosState incoming_state = state;
157
158 Label cont;
159 address compiled_entry = __ pc();
160
161 address entry = __ pc();
162 #if !defined(_LP64) && defined(COMPILER2)
163 // All return values are where we want them, except for Longs. C2 returns
164 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
165 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
166 // build even if we are returning from interpreted we just do a little
167 // stupid shuffing.
168 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
169 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
170 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
171
172 if (incoming_state == ltos) {
173 __ srl (G1, 0, O1);
174 __ srlx(G1, 32, O0);
175 }
176 #endif // !_LP64 && COMPILER2
177
178 __ bind(cont);
179
180 // The callee returns with the stack possibly adjusted by adapter transition
181 // We remove that possible adjustment here.
182 // All interpreter local registers are untouched. Any result is passed back
183 // in the O0/O1 or float registers. Before continuing, the arguments must be
184 // popped from the java expression stack; i.e., Lesp must be adjusted.
185
186 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
187
188 Label L_got_cache, L_giant_index;
189 const Register cache = G3_scratch;
190 const Register size = G1_scratch;
191 if (EnableInvokeDynamic) {
192 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode.
193 __ cmp(G1_scratch, Bytecodes::_invokedynamic);
194 __ br(Assembler::equal, false, Assembler::pn, L_giant_index);
195 __ delayed()->nop();
196 }
197 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1);
198 __ bind(L_got_cache);
199 __ ld_ptr(cache, constantPoolCacheOopDesc::base_offset() +
200 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 // out of the main line of code...
207 if (EnableInvokeDynamic) {
208 __ bind(L_giant_index);
209 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1, sizeof(u4));
210 __ ba(false, L_got_cache);
211 __ delayed()->nop();
212 }
213
214 return entry;
215 }
216
217
218 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
219 address entry = __ pc();
220 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
221 { Label L;
222 Address exception_addr(G2_thread, Thread::pending_exception_offset());
223 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
224 __ tst(Gtemp);
225 __ brx(Assembler::equal, false, Assembler::pt, L);
226 __ delayed()->nop();
227 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
228 __ should_not_reach_here();
229 __ bind(L);
230 }
231 __ dispatch_next(state, step);
232 return entry;
233 }
234
235 // A result handler converts/unboxes a native call result into
236 // a java interpreter/compiler result. The current frame is an
237 // interpreter frame. The activation frame unwind code must be
238 // consistent with that of TemplateTable::_return(...). In the
239 // case of native methods, the caller's SP was not modified.
240 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
241 address entry = __ pc();
242 Register Itos_i = Otos_i ->after_save();
243 Register Itos_l = Otos_l ->after_save();
244 Register Itos_l1 = Otos_l1->after_save();
245 Register Itos_l2 = Otos_l2->after_save();
246 switch (type) {
247 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
248 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
249 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
250 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
251 case T_LONG :
252 #ifndef _LP64
253 __ mov(O1, Itos_l2); // move other half of long
254 #endif // ifdef or no ifdef, fall through to the T_INT case
255 case T_INT : __ mov(O0, Itos_i); break;
256 case T_VOID : /* nothing to do */ break;
257 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
258 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
259 case T_OBJECT :
260 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
261 __ verify_oop(Itos_i);
262 break;
263 default : ShouldNotReachHere();
264 }
265 __ ret(); // return from interpreter activation
266 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
267 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
268 return entry;
269 }
270
271 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
272 address entry = __ pc();
273 __ push(state);
274 __ call_VM(noreg, runtime_entry);
275 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
276 return entry;
277 }
278
279
280 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
281 address entry = __ pc();
282 __ dispatch_next(state);
283 return entry;
284 }
285
286 //
287 // Helpers for commoning out cases in the various type of method entries.
288 //
289
290 // increment invocation count & check for overflow
291 //
292 // Note: checking for negative value instead of overflow
293 // so we have a 'sticky' overflow test
294 //
295 // Lmethod: method
296 // ??: invocation counter
297 //
298 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
299 // Note: In tiered we increment either counters in methodOop or in MDO depending if we're profiling or not.
300 if (TieredCompilation) {
301 const int increment = InvocationCounter::count_increment;
302 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
303 Label no_mdo, done;
304 if (ProfileInterpreter) {
305 // If no method data exists, go to profile_continue.
306 __ ld_ptr(Lmethod, methodOopDesc::method_data_offset(), G4_scratch);
307 __ br_null(G4_scratch, false, Assembler::pn, no_mdo);
308 __ delayed()->nop();
309 // Increment counter
310 Address mdo_invocation_counter(G4_scratch,
311 in_bytes(methodDataOopDesc::invocation_counter_offset()) +
312 in_bytes(InvocationCounter::counter_offset()));
313 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
314 G3_scratch, Lscratch,
315 Assembler::zero, overflow);
316 __ ba(false, done);
317 __ delayed()->nop();
318 }
319
320 // Increment counter in methodOop
321 __ bind(no_mdo);
322 Address invocation_counter(Lmethod,
323 in_bytes(methodOopDesc::invocation_counter_offset()) +
324 in_bytes(InvocationCounter::counter_offset()));
325 __ increment_mask_and_jump(invocation_counter, increment, mask,
326 G3_scratch, Lscratch,
327 Assembler::zero, overflow);
328 __ bind(done);
329 } else {
330 // Update standard invocation counters
331 __ increment_invocation_counter(O0, G3_scratch);
332 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
333 Address interpreter_invocation_counter(Lmethod,in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
334 __ ld(interpreter_invocation_counter, G3_scratch);
335 __ inc(G3_scratch);
336 __ st(G3_scratch, interpreter_invocation_counter);
337 }
338
339 if (ProfileInterpreter && profile_method != NULL) {
340 // Test to see if we should create a method data oop
341 AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
342 __ load_contents(profile_limit, G3_scratch);
343 __ cmp(O0, G3_scratch);
344 __ br(Assembler::lessUnsigned, false, Assembler::pn, *profile_method_continue);
345 __ delayed()->nop();
346
347 // if no method data exists, go to profile_method
348 __ test_method_data_pointer(*profile_method);
349 }
350
351 AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
352 __ load_contents(invocation_limit, G3_scratch);
353 __ cmp(O0, G3_scratch);
354 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
355 __ delayed()->nop();
356 }
357
358 }
359
360 // Allocate monitor and lock method (asm interpreter)
361 // ebx - methodOop
362 //
363 void InterpreterGenerator::lock_method(void) {
364 __ ld(Lmethod, in_bytes(methodOopDesc::access_flags_offset()), O0); // Load access flags.
365
366 #ifdef ASSERT
367 { Label ok;
368 __ btst(JVM_ACC_SYNCHRONIZED, O0);
369 __ br( Assembler::notZero, false, Assembler::pt, ok);
370 __ delayed()->nop();
371 __ stop("method doesn't need synchronization");
372 __ bind(ok);
373 }
374 #endif // ASSERT
375
376 // get synchronization object to O0
377 { Label done;
378 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
379 __ btst(JVM_ACC_STATIC, O0);
380 __ br( Assembler::zero, true, Assembler::pt, done);
381 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
382
383 __ ld_ptr( Lmethod, in_bytes(methodOopDesc::constants_offset()), O0);
384 __ ld_ptr( O0, constantPoolOopDesc::pool_holder_offset_in_bytes(), O0);
385
386 // lock the mirror, not the klassOop
387 __ ld_ptr( O0, mirror_offset, O0);
388
389 #ifdef ASSERT
390 __ tst(O0);
391 __ breakpoint_trap(Assembler::zero);
392 #endif // ASSERT
393
394 __ bind(done);
395 }
396
397 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
398 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
399 // __ untested("lock_object from method entry");
400 __ lock_object(Lmonitors, O0);
401 }
402
403
404 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
405 Register Rscratch,
406 Register Rscratch2) {
407 const int page_size = os::vm_page_size();
408 Address saved_exception_pc(G2_thread, JavaThread::saved_exception_pc_offset());
409 Label after_frame_check;
410
411 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
412
413 __ set( page_size, Rscratch );
414 __ cmp( Rframe_size, Rscratch );
415
416 __ br( Assembler::lessEqual, false, Assembler::pt, after_frame_check );
417 __ delayed()->nop();
418
419 // get the stack base, and in debug, verify it is non-zero
420 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
421 #ifdef ASSERT
422 Label base_not_zero;
423 __ cmp( Rscratch, G0 );
424 __ brx( Assembler::notEqual, false, Assembler::pn, base_not_zero );
425 __ delayed()->nop();
426 __ stop("stack base is zero in generate_stack_overflow_check");
427 __ bind(base_not_zero);
428 #endif
429
430 // get the stack size, and in debug, verify it is non-zero
431 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
432 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
433 #ifdef ASSERT
434 Label size_not_zero;
435 __ cmp( Rscratch2, G0 );
436 __ brx( Assembler::notEqual, false, Assembler::pn, size_not_zero );
437 __ delayed()->nop();
438 __ stop("stack size is zero in generate_stack_overflow_check");
439 __ bind(size_not_zero);
440 #endif
441
442 // compute the beginning of the protected zone minus the requested frame size
443 __ sub( Rscratch, Rscratch2, Rscratch );
444 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
445 __ add( Rscratch, Rscratch2, Rscratch );
446
447 // Add in the size of the frame (which is the same as subtracting it from the
448 // SP, which would take another register
449 __ add( Rscratch, Rframe_size, Rscratch );
450
451 // the frame is greater than one page in size, so check against
452 // the bottom of the stack
453 __ cmp( SP, Rscratch );
454 __ brx( Assembler::greater, false, Assembler::pt, after_frame_check );
455 __ delayed()->nop();
456
457 // Save the return address as the exception pc
458 __ st_ptr(O7, saved_exception_pc);
459
460 // the stack will overflow, throw an exception
461 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
462
463 // if you get to here, then there is enough stack space
464 __ bind( after_frame_check );
465 }
466
467
468 //
469 // Generate a fixed interpreter frame. This is identical setup for interpreted
470 // methods and for native methods hence the shared code.
471
472 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
473 //
474 //
475 // The entry code sets up a new interpreter frame in 4 steps:
476 //
477 // 1) Increase caller's SP by for the extra local space needed:
478 // (check for overflow)
479 // Efficient implementation of xload/xstore bytecodes requires
480 // that arguments and non-argument locals are in a contigously
481 // addressable memory block => non-argument locals must be
482 // allocated in the caller's frame.
483 //
484 // 2) Create a new stack frame and register window:
485 // The new stack frame must provide space for the standard
486 // register save area, the maximum java expression stack size,
487 // the monitor slots (0 slots initially), and some frame local
488 // scratch locations.
489 //
490 // 3) The following interpreter activation registers must be setup:
491 // Lesp : expression stack pointer
492 // Lbcp : bytecode pointer
493 // Lmethod : method
494 // Llocals : locals pointer
495 // Lmonitors : monitor pointer
496 // LcpoolCache: constant pool cache
497 //
498 // 4) Initialize the non-argument locals if necessary:
499 // Non-argument locals may need to be initialized to NULL
500 // for GC to work. If the oop-map information is accurate
501 // (in the absence of the JSR problem), no initialization
502 // is necessary.
503 //
504 // (gri - 2/25/2000)
505
506
507 const Address size_of_parameters(G5_method, methodOopDesc::size_of_parameters_offset());
508 const Address size_of_locals (G5_method, methodOopDesc::size_of_locals_offset());
509 const Address max_stack (G5_method, methodOopDesc::max_stack_offset());
510 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
511
512 const int extra_space =
513 rounded_vm_local_words + // frame local scratch space
514 //6815692//methodOopDesc::extra_stack_words() + // extra push slots for MH adapters
515 frame::memory_parameter_word_sp_offset + // register save area
516 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
517
518 const Register Glocals_size = G3;
519 const Register Otmp1 = O3;
520 const Register Otmp2 = O4;
521 // Lscratch can't be used as a temporary because the call_stub uses
522 // it to assert that the stack frame was setup correctly.
523
524 __ lduh( size_of_parameters, Glocals_size);
525
526 // Gargs points to first local + BytesPerWord
527 // Set the saved SP after the register window save
528 //
529 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
530 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
531 __ add(Gargs, Otmp1, Gargs);
532
533 if (native_call) {
534 __ calc_mem_param_words( Glocals_size, Gframe_size );
535 __ add( Gframe_size, extra_space, Gframe_size);
536 __ round_to( Gframe_size, WordsPerLong );
537 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
538 } else {
539
540 //
541 // Compute number of locals in method apart from incoming parameters
542 //
543 __ lduh( size_of_locals, Otmp1 );
544 __ sub( Otmp1, Glocals_size, Glocals_size );
545 __ round_to( Glocals_size, WordsPerLong );
546 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
547
548 // see if the frame is greater than one page in size. If so,
549 // then we need to verify there is enough stack space remaining
550 // Frame_size = (max_stack + extra_space) * BytesPerWord;
551 __ lduh( max_stack, Gframe_size );
552 __ add( Gframe_size, extra_space, Gframe_size );
553 __ round_to( Gframe_size, WordsPerLong );
554 __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
555
556 // Add in java locals size for stack overflow check only
557 __ add( Gframe_size, Glocals_size, Gframe_size );
558
559 const Register Otmp2 = O4;
560 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
561 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
562
563 __ sub( Gframe_size, Glocals_size, Gframe_size);
564
565 //
566 // bump SP to accomodate the extra locals
567 //
568 __ sub( SP, Glocals_size, SP );
569 }
570
571 //
572 // now set up a stack frame with the size computed above
573 //
574 __ neg( Gframe_size );
575 __ save( SP, Gframe_size, SP );
576
577 //
578 // now set up all the local cache registers
579 //
580 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
581 // that all present references to Lbyte_code initialize the register
582 // immediately before use
583 if (native_call) {
584 __ mov(G0, Lbcp);
585 } else {
586 __ ld_ptr(G5_method, methodOopDesc::const_offset(), Lbcp);
587 __ add(Lbcp, in_bytes(constMethodOopDesc::codes_offset()), Lbcp);
588 }
589 __ mov( G5_method, Lmethod); // set Lmethod
590 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
591 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
592 #ifdef _LP64
593 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
594 #endif
595 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
596
597 // setup interpreter activation registers
598 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
599
600 if (ProfileInterpreter) {
601 #ifdef FAST_DISPATCH
602 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
603 // they both use I2.
604 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
605 #endif // FAST_DISPATCH
606 __ set_method_data_pointer();
607 }
608
609 }
610
611 // Empty method, generate a very fast return.
612
613 address InterpreterGenerator::generate_empty_entry(void) {
614
615 // A method that does nother but return...
616
617 address entry = __ pc();
618 Label slow_path;
619
620 __ verify_oop(G5_method);
621
622 // do nothing for empty methods (do not even increment invocation counter)
623 if ( UseFastEmptyMethods) {
624 // If we need a safepoint check, generate full interpreter entry.
625 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
626 __ set(sync_state, G3_scratch);
627 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
628 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
629 __ delayed()->nop();
630
631 // Code: _return
632 __ retl();
633 __ delayed()->mov(O5_savedSP, SP);
634
635 __ bind(slow_path);
636 (void) generate_normal_entry(false);
637
638 return entry;
639 }
640 return NULL;
641 }
642
643 // Call an accessor method (assuming it is resolved, otherwise drop into
644 // vanilla (slow path) entry
645
646 // Generates code to elide accessor methods
647 // Uses G3_scratch and G1_scratch as scratch
648 address InterpreterGenerator::generate_accessor_entry(void) {
649
650 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
651 // parameter size = 1
652 // Note: We can only use this code if the getfield has been resolved
653 // and if we don't have a null-pointer exception => check for
654 // these conditions first and use slow path if necessary.
655 address entry = __ pc();
656 Label slow_path;
657
658
659 // XXX: for compressed oops pointer loading and decoding doesn't fit in
660 // delay slot and damages G1
661 if ( UseFastAccessorMethods && !UseCompressedOops ) {
662 // Check if we need to reach a safepoint and generate full interpreter
663 // frame if so.
664 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
665 __ load_contents(sync_state, G3_scratch);
666 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
667 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
668 __ delayed()->nop();
669
670 // Check if local 0 != NULL
671 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
672 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
673 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
674 __ delayed()->nop();
675
676
677 // read first instruction word and extract bytecode @ 1 and index @ 2
678 // get first 4 bytes of the bytecodes (big endian!)
679 __ ld_ptr(G5_method, methodOopDesc::const_offset(), G1_scratch);
680 __ ld(G1_scratch, constMethodOopDesc::codes_offset(), G1_scratch);
681
682 // move index @ 2 far left then to the right most two bytes.
683 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
684 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
685 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
686
687 // get constant pool cache
688 __ ld_ptr(G5_method, methodOopDesc::constants_offset(), G3_scratch);
689 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
690
691 // get specific constant pool cache entry
692 __ add(G3_scratch, G1_scratch, G3_scratch);
693
694 // Check the constant Pool cache entry to see if it has been resolved.
695 // If not, need the slow path.
696 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
697 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
698 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
699 __ and3(G1_scratch, 0xFF, G1_scratch);
700 __ cmp(G1_scratch, Bytecodes::_getfield);
701 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
702 __ delayed()->nop();
703
704 // Get the type and return field offset from the constant pool cache
705 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
706 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
707
708 Label xreturn_path;
709 // Need to differentiate between igetfield, agetfield, bgetfield etc.
710 // because they are different sizes.
711 // Get the type from the constant pool cache
712 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
713 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
714 ConstantPoolCacheEntry::verify_tosBits();
715 __ cmp(G1_scratch, atos );
716 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
717 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
718 __ cmp(G1_scratch, itos);
719 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
720 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
721 __ cmp(G1_scratch, stos);
722 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
723 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
724 __ cmp(G1_scratch, ctos);
725 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
726 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
727 #ifdef ASSERT
728 __ cmp(G1_scratch, btos);
729 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
730 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
731 __ should_not_reach_here();
732 #endif
733 __ ldsb(Otos_i, G3_scratch, Otos_i);
734 __ bind(xreturn_path);
735
736 // _ireturn/_areturn
737 __ retl(); // return from leaf routine
738 __ delayed()->mov(O5_savedSP, SP);
739
740 // Generate regular method entry
741 __ bind(slow_path);
742 (void) generate_normal_entry(false);
743 return entry;
744 }
745 return NULL;
746 }
747
748 // Method entry for java.lang.ref.Reference.get.
749 address InterpreterGenerator::generate_Reference_get_entry(void) {
750 #ifndef SERIALGC
751 // Code: _aload_0, _getfield, _areturn
752 // parameter size = 1
753 //
754 // The code that gets generated by this routine is split into 2 parts:
755 // 1. The "intrinsified" code for G1 (or any SATB based GC),
756 // 2. The slow path - which is an expansion of the regular method entry.
757 //
758 // Notes:-
759 // * In the G1 code we do not check whether we need to block for
760 // a safepoint. If G1 is enabled then we must execute the specialized
761 // code for Reference.get (except when the Reference object is null)
762 // so that we can log the value in the referent field with an SATB
763 // update buffer.
764 // If the code for the getfield template is modified so that the
765 // G1 pre-barrier code is executed when the current method is
766 // Reference.get() then going through the normal method entry
767 // will be fine.
768 // * The G1 code can, however, check the receiver object (the instance
769 // of java.lang.Reference) and jump to the slow path if null. If the
770 // Reference object is null then we obviously cannot fetch the referent
771 // and so we don't need to call the G1 pre-barrier. Thus we can use the
772 // regular method entry code to generate the NPE.
773 //
774 // This code is based on generate_accessor_enty.
775
776 address entry = __ pc();
777
778 const int referent_offset = java_lang_ref_Reference::referent_offset;
779 guarantee(referent_offset > 0, "referent offset not initialized");
780
781 if (UseG1GC) {
782 Label slow_path;
783
784 // In the G1 code we don't check if we need to reach a safepoint. We
785 // continue and the thread will safepoint at the next bytecode dispatch.
786
787 // Check if local 0 != NULL
788 // If the receiver is null then it is OK to jump to the slow path.
789 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
790 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
791 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
792 __ delayed()->nop();
793
794
795 // Load the value of the referent field.
796 if (Assembler::is_simm13(referent_offset)) {
797 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
798 } else {
799 __ set(referent_offset, G3_scratch);
800 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
801 }
802
803 // Generate the G1 pre-barrier code to log the value of
804 // the referent field in an SATB buffer. Note with
805 // these parameters the pre-barrier does not generate
806 // the load of the previous value
807
808 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
809 Otos_i /* pre_val */,
810 G3_scratch /* tmp */,
811 true /* preserve_o_regs */);
812
813 // _areturn
814 __ retl(); // return from leaf routine
815 __ delayed()->mov(O5_savedSP, SP);
816
817 // Generate regular method entry
818 __ bind(slow_path);
819 (void) generate_normal_entry(false);
820 return entry;
821 }
822 #endif // SERIALGC
823
824 // If G1 is not enabled then attempt to go through the accessor entry point
825 // Reference.get is an accessor
826 return generate_accessor_entry();
827 }
828
829 //
830 // Interpreter stub for calling a native method. (asm interpreter)
831 // This sets up a somewhat different looking stack for calling the native method
832 // than the typical interpreter frame setup.
833 //
834
835 address InterpreterGenerator::generate_native_entry(bool synchronized) {
836 address entry = __ pc();
837
838 // the following temporary registers are used during frame creation
839 const Register Gtmp1 = G3_scratch ;
840 const Register Gtmp2 = G1_scratch;
841 bool inc_counter = UseCompiler || CountCompiledCalls;
842
843 // make sure registers are different!
844 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
845
846 const Address Laccess_flags(Lmethod, methodOopDesc::access_flags_offset());
847
848 __ verify_oop(G5_method);
849
850 const Register Glocals_size = G3;
851 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
852
853 // make sure method is native & not abstract
854 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
855 #ifdef ASSERT
856 __ ld(G5_method, methodOopDesc::access_flags_offset(), Gtmp1);
857 {
858 Label L;
859 __ btst(JVM_ACC_NATIVE, Gtmp1);
860 __ br(Assembler::notZero, false, Assembler::pt, L);
861 __ delayed()->nop();
862 __ stop("tried to execute non-native method as native");
863 __ bind(L);
864 }
865 { Label L;
866 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
867 __ br(Assembler::zero, false, Assembler::pt, L);
868 __ delayed()->nop();
869 __ stop("tried to execute abstract method as non-abstract");
870 __ bind(L);
871 }
872 #endif // ASSERT
873
874 // generate the code to allocate the interpreter stack frame
875 generate_fixed_frame(true);
876
877 //
878 // No locals to initialize for native method
879 //
880
881 // this slot will be set later, we initialize it to null here just in
882 // case we get a GC before the actual value is stored later
883 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
884
885 const Address do_not_unlock_if_synchronized(G2_thread,
886 JavaThread::do_not_unlock_if_synchronized_offset());
887 // Since at this point in the method invocation the exception handler
888 // would try to exit the monitor of synchronized methods which hasn't
889 // been entered yet, we set the thread local variable
890 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
891 // runtime, exception handling i.e. unlock_if_synchronized_method will
892 // check this thread local flag.
893 // This flag has two effects, one is to force an unwind in the topmost
894 // interpreter frame and not perform an unlock while doing so.
895
896 __ movbool(true, G3_scratch);
897 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
898
899 // increment invocation counter and check for overflow
900 //
901 // Note: checking for negative value instead of overflow
902 // so we have a 'sticky' overflow test (may be of
903 // importance as soon as we have true MT/MP)
904 Label invocation_counter_overflow;
905 Label Lcontinue;
906 if (inc_counter) {
907 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
908
909 }
910 __ bind(Lcontinue);
911
912 bang_stack_shadow_pages(true);
913
914 // reset the _do_not_unlock_if_synchronized flag
915 __ stbool(G0, do_not_unlock_if_synchronized);
916
917 // check for synchronized methods
918 // Must happen AFTER invocation_counter check and stack overflow check,
919 // so method is not locked if overflows.
920
921 if (synchronized) {
922 lock_method();
923 } else {
924 #ifdef ASSERT
925 { Label ok;
926 __ ld(Laccess_flags, O0);
927 __ btst(JVM_ACC_SYNCHRONIZED, O0);
928 __ br( Assembler::zero, false, Assembler::pt, ok);
929 __ delayed()->nop();
930 __ stop("method needs synchronization");
931 __ bind(ok);
932 }
933 #endif // ASSERT
934 }
935
936
937 // start execution
938 __ verify_thread();
939
940 // JVMTI support
941 __ notify_method_entry();
942
943 // native call
944
945 // (note that O0 is never an oop--at most it is a handle)
946 // It is important not to smash any handles created by this call,
947 // until any oop handle in O0 is dereferenced.
948
949 // (note that the space for outgoing params is preallocated)
950
951 // get signature handler
952 { Label L;
953 Address signature_handler(Lmethod, methodOopDesc::signature_handler_offset());
954 __ ld_ptr(signature_handler, G3_scratch);
955 __ tst(G3_scratch);
956 __ brx(Assembler::notZero, false, Assembler::pt, L);
957 __ delayed()->nop();
958 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
959 __ ld_ptr(signature_handler, G3_scratch);
960 __ bind(L);
961 }
962
963 // Push a new frame so that the args will really be stored in
964 // Copy a few locals across so the new frame has the variables
965 // we need but these values will be dead at the jni call and
966 // therefore not gc volatile like the values in the current
967 // frame (Lmethod in particular)
968
969 // Flush the method pointer to the register save area
970 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
971 __ mov(Llocals, O1);
972
973 // calculate where the mirror handle body is allocated in the interpreter frame:
974 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
975
976 // Calculate current frame size
977 __ sub(SP, FP, O3); // Calculate negative of current frame size
978 __ save(SP, O3, SP); // Allocate an identical sized frame
979
980 // Note I7 has leftover trash. Slow signature handler will fill it in
981 // should we get there. Normal jni call will set reasonable last_Java_pc
982 // below (and fix I7 so the stack trace doesn't have a meaningless frame
983 // in it).
984
985 // Load interpreter frame's Lmethod into same register here
986
987 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
988
989 __ mov(I1, Llocals);
990 __ mov(I2, Lscratch2); // save the address of the mirror
991
992
993 // ONLY Lmethod and Llocals are valid here!
994
995 // call signature handler, It will move the arg properly since Llocals in current frame
996 // matches that in outer frame
997
998 __ callr(G3_scratch, 0);
999 __ delayed()->nop();
1000
1001 // Result handler is in Lscratch
1002
1003 // Reload interpreter frame's Lmethod since slow signature handler may block
1004 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1005
1006 { Label not_static;
1007
1008 __ ld(Laccess_flags, O0);
1009 __ btst(JVM_ACC_STATIC, O0);
1010 __ br( Assembler::zero, false, Assembler::pt, not_static);
1011 // get native function entry point(O0 is a good temp until the very end)
1012 __ delayed()->ld_ptr(Lmethod, in_bytes(methodOopDesc::native_function_offset()), O0);
1013 // for static methods insert the mirror argument
1014 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1015
1016 __ ld_ptr(Lmethod, methodOopDesc:: constants_offset(), O1);
1017 __ ld_ptr(O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
1018 __ ld_ptr(O1, mirror_offset, O1);
1019 #ifdef ASSERT
1020 if (!PrintSignatureHandlers) // do not dirty the output with this
1021 { Label L;
1022 __ tst(O1);
1023 __ brx(Assembler::notZero, false, Assembler::pt, L);
1024 __ delayed()->nop();
1025 __ stop("mirror is missing");
1026 __ bind(L);
1027 }
1028 #endif // ASSERT
1029 __ st_ptr(O1, Lscratch2, 0);
1030 __ mov(Lscratch2, O1);
1031 __ bind(not_static);
1032 }
1033
1034 // At this point, arguments have been copied off of stack into
1035 // their JNI positions, which are O1..O5 and SP[68..].
1036 // Oops are boxed in-place on the stack, with handles copied to arguments.
1037 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1038
1039 #ifdef ASSERT
1040 { Label L;
1041 __ tst(O0);
1042 __ brx(Assembler::notZero, false, Assembler::pt, L);
1043 __ delayed()->nop();
1044 __ stop("native entry point is missing");
1045 __ bind(L);
1046 }
1047 #endif // ASSERT
1048
1049 //
1050 // setup the frame anchor
1051 //
1052 // The scavenge function only needs to know that the PC of this frame is
1053 // in the interpreter method entry code, it doesn't need to know the exact
1054 // PC and hence we can use O7 which points to the return address from the
1055 // previous call in the code stream (signature handler function)
1056 //
1057 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1058 // we have pushed the extra frame in order to protect the volatile register(s)
1059 // in that frame when we return from the jni call
1060 //
1061
1062 __ set_last_Java_frame(FP, O7);
1063 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1064 // not meaningless information that'll confuse me.
1065
1066 // flush the windows now. We don't care about the current (protection) frame
1067 // only the outer frames
1068
1069 __ flush_windows();
1070
1071 // mark windows as flushed
1072 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1073 __ set(JavaFrameAnchor::flushed, G3_scratch);
1074 __ st(G3_scratch, flags);
1075
1076 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1077
1078 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1079 #ifdef ASSERT
1080 { Label L;
1081 __ ld(thread_state, G3_scratch);
1082 __ cmp(G3_scratch, _thread_in_Java);
1083 __ br(Assembler::equal, false, Assembler::pt, L);
1084 __ delayed()->nop();
1085 __ stop("Wrong thread state in native stub");
1086 __ bind(L);
1087 }
1088 #endif // ASSERT
1089 __ set(_thread_in_native, G3_scratch);
1090 __ st(G3_scratch, thread_state);
1091
1092 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1093 __ save_thread(L7_thread_cache); // save Gthread
1094 __ callr(O0, 0);
1095 __ delayed()->
1096 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1097
1098 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1099
1100 __ restore_thread(L7_thread_cache); // restore G2_thread
1101 __ reinit_heapbase();
1102
1103 // must we block?
1104
1105 // Block, if necessary, before resuming in _thread_in_Java state.
1106 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1107 { Label no_block;
1108 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1109
1110 // Switch thread to "native transition" state before reading the synchronization state.
1111 // This additional state is necessary because reading and testing the synchronization
1112 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1113 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1114 // VM thread changes sync state to synchronizing and suspends threads for GC.
1115 // Thread A is resumed to finish this native method, but doesn't block here since it
1116 // didn't see any synchronization is progress, and escapes.
1117 __ set(_thread_in_native_trans, G3_scratch);
1118 __ st(G3_scratch, thread_state);
1119 if(os::is_MP()) {
1120 if (UseMembar) {
1121 // Force this write out before the read below
1122 __ membar(Assembler::StoreLoad);
1123 } else {
1124 // Write serialization page so VM thread can do a pseudo remote membar.
1125 // We use the current thread pointer to calculate a thread specific
1126 // offset to write to within the page. This minimizes bus traffic
1127 // due to cache line collision.
1128 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1129 }
1130 }
1131 __ load_contents(sync_state, G3_scratch);
1132 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1133
1134 Label L;
1135 __ br(Assembler::notEqual, false, Assembler::pn, L);
1136 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1137 __ cmp(G3_scratch, 0);
1138 __ br(Assembler::equal, false, Assembler::pt, no_block);
1139 __ delayed()->nop();
1140 __ bind(L);
1141
1142 // Block. Save any potential method result value before the operation and
1143 // use a leaf call to leave the last_Java_frame setup undisturbed.
1144 save_native_result();
1145 __ call_VM_leaf(L7_thread_cache,
1146 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1147 G2_thread);
1148
1149 // Restore any method result value
1150 restore_native_result();
1151 __ bind(no_block);
1152 }
1153
1154 // Clear the frame anchor now
1155
1156 __ reset_last_Java_frame();
1157
1158 // Move the result handler address
1159 __ mov(Lscratch, G3_scratch);
1160 // return possible result to the outer frame
1161 #ifndef __LP64
1162 __ mov(O0, I0);
1163 __ restore(O1, G0, O1);
1164 #else
1165 __ restore(O0, G0, O0);
1166 #endif /* __LP64 */
1167
1168 // Move result handler to expected register
1169 __ mov(G3_scratch, Lscratch);
1170
1171 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1172 // switch to thread_in_Java.
1173
1174 __ set(_thread_in_Java, G3_scratch);
1175 __ st(G3_scratch, thread_state);
1176
1177 // reset handle block
1178 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1179 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1180
1181 // If we have an oop result store it where it will be safe for any further gc
1182 // until we return now that we've released the handle it might be protected by
1183
1184 {
1185 Label no_oop, store_result;
1186
1187 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1188 __ cmp(G3_scratch, Lscratch);
1189 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
1190 __ delayed()->nop();
1191 __ addcc(G0, O0, O0);
1192 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1193 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1194 __ mov(G0, O0);
1195
1196 __ bind(store_result);
1197 // Store it where gc will look for it and result handler expects it.
1198 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1199
1200 __ bind(no_oop);
1201
1202 }
1203
1204
1205 // handle exceptions (exception handling will handle unlocking!)
1206 { Label L;
1207 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1208 __ ld_ptr(exception_addr, Gtemp);
1209 __ tst(Gtemp);
1210 __ brx(Assembler::equal, false, Assembler::pt, L);
1211 __ delayed()->nop();
1212 // Note: This could be handled more efficiently since we know that the native
1213 // method doesn't have an exception handler. We could directly return
1214 // to the exception handler for the caller.
1215 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1216 __ should_not_reach_here();
1217 __ bind(L);
1218 }
1219
1220 // JVMTI support (preserves thread register)
1221 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1222
1223 if (synchronized) {
1224 // save and restore any potential method result value around the unlocking operation
1225 save_native_result();
1226
1227 __ add( __ top_most_monitor(), O1);
1228 __ unlock_object(O1);
1229
1230 restore_native_result();
1231 }
1232
1233 #if defined(COMPILER2) && !defined(_LP64)
1234
1235 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1236 // or compiled so just be safe.
1237
1238 __ sllx(O0, 32, G1); // Shift bits into high G1
1239 __ srl (O1, 0, O1); // Zero extend O1
1240 __ or3 (O1, G1, G1); // OR 64 bits into G1
1241
1242 #endif /* COMPILER2 && !_LP64 */
1243
1244 // dispose of return address and remove activation
1245 #ifdef ASSERT
1246 {
1247 Label ok;
1248 __ cmp(I5_savedSP, FP);
1249 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1250 __ delayed()->nop();
1251 __ stop("bad I5_savedSP value");
1252 __ should_not_reach_here();
1253 __ bind(ok);
1254 }
1255 #endif
1256 if (TraceJumps) {
1257 // Move target to register that is recordable
1258 __ mov(Lscratch, G3_scratch);
1259 __ JMP(G3_scratch, 0);
1260 } else {
1261 __ jmp(Lscratch, 0);
1262 }
1263 __ delayed()->nop();
1264
1265
1266 if (inc_counter) {
1267 // handle invocation counter overflow
1268 __ bind(invocation_counter_overflow);
1269 generate_counter_overflow(Lcontinue);
1270 }
1271
1272
1273
1274 return entry;
1275 }
1276
1277
1278 // Generic method entry to (asm) interpreter
1279 //------------------------------------------------------------------------------------------------------------------------
1280 //
1281 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1282 address entry = __ pc();
1283
1284 bool inc_counter = UseCompiler || CountCompiledCalls;
1285
1286 // the following temporary registers are used during frame creation
1287 const Register Gtmp1 = G3_scratch ;
1288 const Register Gtmp2 = G1_scratch;
1289
1290 // make sure registers are different!
1291 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1292
1293 const Address size_of_parameters(G5_method, methodOopDesc::size_of_parameters_offset());
1294 const Address size_of_locals (G5_method, methodOopDesc::size_of_locals_offset());
1295 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1296 // and use in the asserts.
1297 const Address access_flags (Lmethod, methodOopDesc::access_flags_offset());
1298
1299 __ verify_oop(G5_method);
1300
1301 const Register Glocals_size = G3;
1302 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1303
1304 // make sure method is not native & not abstract
1305 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1306 #ifdef ASSERT
1307 __ ld(G5_method, methodOopDesc::access_flags_offset(), Gtmp1);
1308 {
1309 Label L;
1310 __ btst(JVM_ACC_NATIVE, Gtmp1);
1311 __ br(Assembler::zero, false, Assembler::pt, L);
1312 __ delayed()->nop();
1313 __ stop("tried to execute native method as non-native");
1314 __ bind(L);
1315 }
1316 { Label L;
1317 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1318 __ br(Assembler::zero, false, Assembler::pt, L);
1319 __ delayed()->nop();
1320 __ stop("tried to execute abstract method as non-abstract");
1321 __ bind(L);
1322 }
1323 #endif // ASSERT
1324
1325 // generate the code to allocate the interpreter stack frame
1326
1327 generate_fixed_frame(false);
1328
1329 #ifdef FAST_DISPATCH
1330 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1331 // set bytecode dispatch table base
1332 #endif
1333
1334 //
1335 // Code to initialize the extra (i.e. non-parm) locals
1336 //
1337 Register init_value = noreg; // will be G0 if we must clear locals
1338 // The way the code was setup before zerolocals was always true for vanilla java entries.
1339 // It could only be false for the specialized entries like accessor or empty which have
1340 // no extra locals so the testing was a waste of time and the extra locals were always
1341 // initialized. We removed this extra complication to already over complicated code.
1342
1343 init_value = G0;
1344 Label clear_loop;
1345
1346 // NOTE: If you change the frame layout, this code will need to
1347 // be updated!
1348 __ lduh( size_of_locals, O2 );
1349 __ lduh( size_of_parameters, O1 );
1350 __ sll( O2, Interpreter::logStackElementSize, O2);
1351 __ sll( O1, Interpreter::logStackElementSize, O1 );
1352 __ sub( Llocals, O2, O2 );
1353 __ sub( Llocals, O1, O1 );
1354
1355 __ bind( clear_loop );
1356 __ inc( O2, wordSize );
1357
1358 __ cmp( O2, O1 );
1359 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1360 __ delayed()->st_ptr( init_value, O2, 0 );
1361
1362 const Address do_not_unlock_if_synchronized(G2_thread,
1363 JavaThread::do_not_unlock_if_synchronized_offset());
1364 // Since at this point in the method invocation the exception handler
1365 // would try to exit the monitor of synchronized methods which hasn't
1366 // been entered yet, we set the thread local variable
1367 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1368 // runtime, exception handling i.e. unlock_if_synchronized_method will
1369 // check this thread local flag.
1370 __ movbool(true, G3_scratch);
1371 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1372
1373 // increment invocation counter and check for overflow
1374 //
1375 // Note: checking for negative value instead of overflow
1376 // so we have a 'sticky' overflow test (may be of
1377 // importance as soon as we have true MT/MP)
1378 Label invocation_counter_overflow;
1379 Label profile_method;
1380 Label profile_method_continue;
1381 Label Lcontinue;
1382 if (inc_counter) {
1383 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1384 if (ProfileInterpreter) {
1385 __ bind(profile_method_continue);
1386 }
1387 }
1388 __ bind(Lcontinue);
1389
1390 bang_stack_shadow_pages(false);
1391
1392 // reset the _do_not_unlock_if_synchronized flag
1393 __ stbool(G0, do_not_unlock_if_synchronized);
1394
1395 // check for synchronized methods
1396 // Must happen AFTER invocation_counter check and stack overflow check,
1397 // so method is not locked if overflows.
1398
1399 if (synchronized) {
1400 lock_method();
1401 } else {
1402 #ifdef ASSERT
1403 { Label ok;
1404 __ ld(access_flags, O0);
1405 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1406 __ br( Assembler::zero, false, Assembler::pt, ok);
1407 __ delayed()->nop();
1408 __ stop("method needs synchronization");
1409 __ bind(ok);
1410 }
1411 #endif // ASSERT
1412 }
1413
1414 // start execution
1415
1416 __ verify_thread();
1417
1418 // jvmti support
1419 __ notify_method_entry();
1420
1421 // start executing instructions
1422 __ dispatch_next(vtos);
1423
1424
1425 if (inc_counter) {
1426 if (ProfileInterpreter) {
1427 // We have decided to profile this method in the interpreter
1428 __ bind(profile_method);
1429
1430 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1431 __ set_method_data_pointer_for_bcp();
1432 __ ba(false, profile_method_continue);
1433 __ delayed()->nop();
1434 }
1435
1436 // handle invocation counter overflow
1437 __ bind(invocation_counter_overflow);
1438 generate_counter_overflow(Lcontinue);
1439 }
1440
1441
1442 return entry;
1443 }
1444
1445
1446 //----------------------------------------------------------------------------------------------------
1447 // Entry points & stack frame layout
1448 //
1449 // Here we generate the various kind of entries into the interpreter.
1450 // The two main entry type are generic bytecode methods and native call method.
1451 // These both come in synchronized and non-synchronized versions but the
1452 // frame layout they create is very similar. The other method entry
1453 // types are really just special purpose entries that are really entry
1454 // and interpretation all in one. These are for trivial methods like
1455 // accessor, empty, or special math methods.
1456 //
1457 // When control flow reaches any of the entry types for the interpreter
1458 // the following holds ->
1459 //
1460 // C2 Calling Conventions:
1461 //
1462 // The entry code below assumes that the following registers are set
1463 // when coming in:
1464 // G5_method: holds the methodOop of the method to call
1465 // Lesp: points to the TOS of the callers expression stack
1466 // after having pushed all the parameters
1467 //
1468 // The entry code does the following to setup an interpreter frame
1469 // pop parameters from the callers stack by adjusting Lesp
1470 // set O0 to Lesp
1471 // compute X = (max_locals - num_parameters)
1472 // bump SP up by X to accomadate the extra locals
1473 // compute X = max_expression_stack
1474 // + vm_local_words
1475 // + 16 words of register save area
1476 // save frame doing a save sp, -X, sp growing towards lower addresses
1477 // set Lbcp, Lmethod, LcpoolCache
1478 // set Llocals to i0
1479 // set Lmonitors to FP - rounded_vm_local_words
1480 // set Lesp to Lmonitors - 4
1481 //
1482 // The frame has now been setup to do the rest of the entry code
1483
1484 // Try this optimization: Most method entries could live in a
1485 // "one size fits all" stack frame without all the dynamic size
1486 // calculations. It might be profitable to do all this calculation
1487 // statically and approximately for "small enough" methods.
1488
1489 //-----------------------------------------------------------------------------------------------
1490
1491 // C1 Calling conventions
1492 //
1493 // Upon method entry, the following registers are setup:
1494 //
1495 // g2 G2_thread: current thread
1496 // g5 G5_method: method to activate
1497 // g4 Gargs : pointer to last argument
1498 //
1499 //
1500 // Stack:
1501 //
1502 // +---------------+ <--- sp
1503 // | |
1504 // : reg save area :
1505 // | |
1506 // +---------------+ <--- sp + 0x40
1507 // | |
1508 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1509 // | |
1510 // +---------------+ <--- sp + 0x5c
1511 // | |
1512 // : free :
1513 // | |
1514 // +---------------+ <--- Gargs
1515 // | |
1516 // : arguments :
1517 // | |
1518 // +---------------+
1519 // | |
1520 //
1521 //
1522 //
1523 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1524 //
1525 // +---------------+ <--- sp
1526 // | |
1527 // : reg save area :
1528 // | |
1529 // +---------------+ <--- sp + 0x40
1530 // | |
1531 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1532 // | |
1533 // +---------------+ <--- sp + 0x5c
1534 // | |
1535 // : :
1536 // | | <--- Lesp
1537 // +---------------+ <--- Lmonitors (fp - 0x18)
1538 // | VM locals |
1539 // +---------------+ <--- fp
1540 // | |
1541 // : reg save area :
1542 // | |
1543 // +---------------+ <--- fp + 0x40
1544 // | |
1545 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1546 // | |
1547 // +---------------+ <--- fp + 0x5c
1548 // | |
1549 // : free :
1550 // | |
1551 // +---------------+
1552 // | |
1553 // : nonarg locals :
1554 // | |
1555 // +---------------+
1556 // | |
1557 // : arguments :
1558 // | | <--- Llocals
1559 // +---------------+ <--- Gargs
1560 // | |
1561
1562 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1563
1564 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1565 // expression stack, the callee will have callee_extra_locals (so we can account for
1566 // frame extension) and monitor_size for monitors. Basically we need to calculate
1567 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1568 //
1569 //
1570 // The big complicating thing here is that we must ensure that the stack stays properly
1571 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1572 // needs to be aligned for). We are given that the sp (fp) is already aligned by
1573 // the caller so we must ensure that it is properly aligned for our callee.
1574 //
1575 const int rounded_vm_local_words =
1576 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1577 // callee_locals and max_stack are counts, not the size in frame.
1578 const int locals_size =
1579 round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1580 const int max_stack_words = max_stack * Interpreter::stackElementWords;
1581 return (round_to((max_stack_words
1582 //6815692//+ methodOopDesc::extra_stack_words()
1583 + rounded_vm_local_words
1584 + frame::memory_parameter_word_sp_offset), WordsPerLong)
1585 // already rounded
1586 + locals_size + monitor_size);
1587 }
1588
1589 // How much stack a method top interpreter activation needs in words.
1590 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
1591
1592 // See call_stub code
1593 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
1594 WordsPerLong); // 7 + register save area
1595
1596 // Save space for one monitor to get into the interpreted method in case
1597 // the method is synchronized
1598 int monitor_size = method->is_synchronized() ?
1599 1*frame::interpreter_frame_monitor_size() : 0;
1600 return size_activation_helper(method->max_locals(), method->max_stack(),
1601 monitor_size) + call_stub_size;
1602 }
1603
1604 int AbstractInterpreter::layout_activation(methodOop method,
1605 int tempcount,
1606 int popframe_extra_args,
1607 int moncount,
1608 int caller_actual_parameters,
1609 int callee_param_count,
1610 int callee_local_count,
1611 frame* caller,
1612 frame* interpreter_frame,
1613 bool is_top_frame) {
1614 // Note: This calculation must exactly parallel the frame setup
1615 // in InterpreterGenerator::generate_fixed_frame.
1616 // If f!=NULL, set up the following variables:
1617 // - Lmethod
1618 // - Llocals
1619 // - Lmonitors (to the indicated number of monitors)
1620 // - Lesp (to the indicated number of temps)
1621 // The frame f (if not NULL) on entry is a description of the caller of the frame
1622 // we are about to layout. We are guaranteed that we will be able to fill in a
1623 // new interpreter frame as its callee (i.e. the stack space is allocated and
1624 // the amount was determined by an earlier call to this method with f == NULL).
1625 // On return f (if not NULL) while describe the interpreter frame we just layed out.
1626
1627 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
1628 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1629
1630 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1631 //
1632 // Note: if you look closely this appears to be doing something much different
1633 // than generate_fixed_frame. What is happening is this. On sparc we have to do
1634 // this dance with interpreter_sp_adjustment because the window save area would
1635 // appear just below the bottom (tos) of the caller's java expression stack. Because
1636 // the interpreter want to have the locals completely contiguous generate_fixed_frame
1637 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1638 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1639 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1640 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1641 // because the oldest frame would have adjust its callers frame and yet that frame
1642 // already exists and isn't part of this array of frames we are unpacking. So at first
1643 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1644 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1645 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1646 // add up. It does seem like it simpler to account for the adjustment here (and remove the
1647 // callee... parameters here). However this would mean that this routine would have to take
1648 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1649 // and run the calling loop in the reverse order. This would also would appear to mean making
1650 // this code aware of what the interactions are when that initial caller fram was an osr or
1651 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
1652 // there is no sense in messing working code.
1653 //
1654
1655 int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
1656 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1657
1658 int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
1659 monitor_size);
1660
1661 if (interpreter_frame != NULL) {
1662 // The skeleton frame must already look like an interpreter frame
1663 // even if not fully filled out.
1664 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1665
1666 intptr_t* fp = interpreter_frame->fp();
1667
1668 JavaThread* thread = JavaThread::current();
1669 RegisterMap map(thread, false);
1670 // More verification that skeleton frame is properly walkable
1671 assert(fp == caller->sp(), "fp must match");
1672
1673 intptr_t* montop = fp - rounded_vm_local_words;
1674
1675 // preallocate monitors (cf. __ add_monitor_to_stack)
1676 intptr_t* monitors = montop - monitor_size;
1677
1678 // preallocate stack space
1679 intptr_t* esp = monitors - 1 -
1680 (tempcount * Interpreter::stackElementWords) -
1681 popframe_extra_args;
1682
1683 int local_words = method->max_locals() * Interpreter::stackElementWords;
1684 NEEDS_CLEANUP;
1685 intptr_t* locals;
1686 if (caller->is_interpreted_frame()) {
1687 // Can force the locals area to end up properly overlapping the top of the expression stack.
1688 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1689 // Note that this computation means we replace size_of_parameters() values from the caller
1690 // interpreter frame's expression stack with our argument locals
1691 int parm_words = caller_actual_parameters * Interpreter::stackElementWords;
1692 locals = Lesp_ptr + parm_words;
1693 int delta = local_words - parm_words;
1694 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1695 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1696 } else {
1697 assert(caller->is_compiled_frame() || caller->is_entry_frame() || caller->is_ricochet_frame(), "only possible cases");
1698 // Don't have Lesp available; lay out locals block in the caller
1699 // adjacent to the register window save area.
1700 //
1701 // Compiled frames do not allocate a varargs area which is why this if
1702 // statement is needed.
1703 //
1704 if (caller->is_compiled_frame()) {
1705 locals = fp + frame::register_save_words + local_words - 1;
1706 } else {
1707 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1708 }
1709 if (!caller->is_entry_frame()) {
1710 // Caller wants his own SP back
1711 int caller_frame_size = caller->cb()->frame_size();
1712 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1713 }
1714 }
1715 if (TraceDeoptimization) {
1716 if (caller->is_entry_frame()) {
1717 // make sure I5_savedSP and the entry frames notion of saved SP
1718 // agree. This assertion duplicate a check in entry frame code
1719 // but catches the failure earlier.
1720 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1721 "would change callers SP");
1722 }
1723 if (caller->is_entry_frame()) {
1724 tty->print("entry ");
1725 }
1726 if (caller->is_compiled_frame()) {
1727 tty->print("compiled ");
1728 if (caller->is_deoptimized_frame()) {
1729 tty->print("(deopt) ");
1730 }
1731 }
1732 if (caller->is_interpreted_frame()) {
1733 tty->print("interpreted ");
1734 }
1735 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1736 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1737 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1738 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1739 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1740 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1741 tty->print_cr("Llocals = 0x%x", locals);
1742 tty->print_cr("Lesp = 0x%x", esp);
1743 tty->print_cr("Lmonitors = 0x%x", monitors);
1744 }
1745
1746 if (method->max_locals() > 0) {
1747 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1748 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1749 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1750 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1751 }
1752 #ifdef _LP64
1753 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1754 #endif
1755
1756 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
1757 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
1758 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
1759 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
1760 // Llast_SP will be same as SP as there is no adapter space
1761 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1762 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1763 #ifdef FAST_DISPATCH
1764 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1765 #endif
1766
1767
1768 #ifdef ASSERT
1769 BasicObjectLock* mp = (BasicObjectLock*)monitors;
1770
1771 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1772 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1773 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
1774 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1775 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1776
1777 // check bounds
1778 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1779 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1780 assert(lo < monitors && montop <= hi, "monitors in bounds");
1781 assert(lo <= esp && esp < monitors, "esp in bounds");
1782 #endif // ASSERT
1783 }
1784
1785 return raw_frame_size;
1786 }
1787
1788 //----------------------------------------------------------------------------------------------------
1789 // Exceptions
1790 void TemplateInterpreterGenerator::generate_throw_exception() {
1791
1792 // Entry point in previous activation (i.e., if the caller was interpreted)
1793 Interpreter::_rethrow_exception_entry = __ pc();
1794 // O0: exception
1795
1796 // entry point for exceptions thrown within interpreter code
1797 Interpreter::_throw_exception_entry = __ pc();
1798 __ verify_thread();
1799 // expression stack is undefined here
1800 // O0: exception, i.e. Oexception
1801 // Lbcp: exception bcx
1802 __ verify_oop(Oexception);
1803
1804
1805 // expression stack must be empty before entering the VM in case of an exception
1806 __ empty_expression_stack();
1807 // find exception handler address and preserve exception oop
1808 // call C routine to find handler and jump to it
1809 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1810 __ push_ptr(O1); // push exception for exception handler bytecodes
1811
1812 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1813 __ delayed()->nop();
1814
1815
1816 // if the exception is not handled in the current frame
1817 // the frame is removed and the exception is rethrown
1818 // (i.e. exception continuation is _rethrow_exception)
1819 //
1820 // Note: At this point the bci is still the bxi for the instruction which caused
1821 // the exception and the expression stack is empty. Thus, for any VM calls
1822 // at this point, GC will find a legal oop map (with empty expression stack).
1823
1824 // in current activation
1825 // tos: exception
1826 // Lbcp: exception bcp
1827
1828 //
1829 // JVMTI PopFrame support
1830 //
1831
1832 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1833 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1834 // Set the popframe_processing bit in popframe_condition indicating that we are
1835 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1836 // popframe handling cycles.
1837
1838 __ ld(popframe_condition_addr, G3_scratch);
1839 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1840 __ stw(G3_scratch, popframe_condition_addr);
1841
1842 // Empty the expression stack, as in normal exception handling
1843 __ empty_expression_stack();
1844 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1845
1846 {
1847 // Check to see whether we are returning to a deoptimized frame.
1848 // (The PopFrame call ensures that the caller of the popped frame is
1849 // either interpreted or compiled and deoptimizes it if compiled.)
1850 // In this case, we can't call dispatch_next() after the frame is
1851 // popped, but instead must save the incoming arguments and restore
1852 // them after deoptimization has occurred.
1853 //
1854 // Note that we don't compare the return PC against the
1855 // deoptimization blob's unpack entry because of the presence of
1856 // adapter frames in C2.
1857 Label caller_not_deoptimized;
1858 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1859 __ tst(O0);
1860 __ brx(Assembler::notEqual, false, Assembler::pt, caller_not_deoptimized);
1861 __ delayed()->nop();
1862
1863 const Register Gtmp1 = G3_scratch;
1864 const Register Gtmp2 = G1_scratch;
1865
1866 // Compute size of arguments for saving when returning to deoptimized caller
1867 __ lduh(Lmethod, in_bytes(methodOopDesc::size_of_parameters_offset()), Gtmp1);
1868 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1869 __ sub(Llocals, Gtmp1, Gtmp2);
1870 __ add(Gtmp2, wordSize, Gtmp2);
1871 // Save these arguments
1872 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1873 // Inform deoptimization that it is responsible for restoring these arguments
1874 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1875 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1876 __ st(Gtmp1, popframe_condition_addr);
1877
1878 // Return from the current method
1879 // The caller's SP was adjusted upon method entry to accomodate
1880 // the callee's non-argument locals. Undo that adjustment.
1881 __ ret();
1882 __ delayed()->restore(I5_savedSP, G0, SP);
1883
1884 __ bind(caller_not_deoptimized);
1885 }
1886
1887 // Clear the popframe condition flag
1888 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1889
1890 // Get out of the current method (how this is done depends on the particular compiler calling
1891 // convention that the interpreter currently follows)
1892 // The caller's SP was adjusted upon method entry to accomodate
1893 // the callee's non-argument locals. Undo that adjustment.
1894 __ restore(I5_savedSP, G0, SP);
1895 // The method data pointer was incremented already during
1896 // call profiling. We have to restore the mdp for the current bcp.
1897 if (ProfileInterpreter) {
1898 __ set_method_data_pointer_for_bcp();
1899 }
1900 // Resume bytecode interpretation at the current bcp
1901 __ dispatch_next(vtos);
1902 // end of JVMTI PopFrame support
1903
1904 Interpreter::_remove_activation_entry = __ pc();
1905
1906 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1907 __ pop_ptr(Oexception); // get exception
1908
1909 // Intel has the following comment:
1910 //// remove the activation (without doing throws on illegalMonitorExceptions)
1911 // They remove the activation without checking for bad monitor state.
1912 // %%% We should make sure this is the right semantics before implementing.
1913
1914 // %%% changed set_vm_result_2 to set_vm_result and get_vm_result_2 to get_vm_result. Is there a bug here?
1915 __ set_vm_result(Oexception);
1916 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1917
1918 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1919
1920 __ get_vm_result(Oexception);
1921 __ verify_oop(Oexception);
1922
1923 const int return_reg_adjustment = frame::pc_return_offset;
1924 Address issuing_pc_addr(I7, return_reg_adjustment);
1925
1926 // We are done with this activation frame; find out where to go next.
1927 // The continuation point will be an exception handler, which expects
1928 // the following registers set up:
1929 //
1930 // Oexception: exception
1931 // Oissuing_pc: the local call that threw exception
1932 // Other On: garbage
1933 // In/Ln: the contents of the caller's register window
1934 //
1935 // We do the required restore at the last possible moment, because we
1936 // need to preserve some state across a runtime call.
1937 // (Remember that the caller activation is unknown--it might not be
1938 // interpreted, so things like Lscratch are useless in the caller.)
1939
1940 // Although the Intel version uses call_C, we can use the more
1941 // compact call_VM. (The only real difference on SPARC is a
1942 // harmlessly ignored [re]set_last_Java_frame, compared with
1943 // the Intel code which lacks this.)
1944 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1945 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1946 __ super_call_VM_leaf(L7_thread_cache,
1947 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1948 G2_thread, Oissuing_pc->after_save());
1949
1950 // The caller's SP was adjusted upon method entry to accomodate
1951 // the callee's non-argument locals. Undo that adjustment.
1952 __ JMP(O0, 0); // return exception handler in caller
1953 __ delayed()->restore(I5_savedSP, G0, SP);
1954
1955 // (same old exception object is already in Oexception; see above)
1956 // Note that an "issuing PC" is actually the next PC after the call
1957 }
1958
1959
1960 //
1961 // JVMTI ForceEarlyReturn support
1962 //
1963
1964 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1965 address entry = __ pc();
1966
1967 __ empty_expression_stack();
1968 __ load_earlyret_value(state);
1969
1970 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1971 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1972
1973 // Clear the earlyret state
1974 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1975
1976 __ remove_activation(state,
1977 /* throw_monitor_exception */ false,
1978 /* install_monitor_exception */ false);
1979
1980 // The caller's SP was adjusted upon method entry to accomodate
1981 // the callee's non-argument locals. Undo that adjustment.
1982 __ ret(); // return to caller
1983 __ delayed()->restore(I5_savedSP, G0, SP);
1984
1985 return entry;
1986 } // end of JVMTI ForceEarlyReturn support
1987
1988
1989 //------------------------------------------------------------------------------------------------------------------------
1990 // Helper for vtos entry point generation
1991
1992 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) {
1993 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1994 Label L;
1995 aep = __ pc(); __ push_ptr(); __ ba(false, L); __ delayed()->nop();
1996 fep = __ pc(); __ push_f(); __ ba(false, L); __ delayed()->nop();
1997 dep = __ pc(); __ push_d(); __ ba(false, L); __ delayed()->nop();
1998 lep = __ pc(); __ push_l(); __ ba(false, L); __ delayed()->nop();
1999 iep = __ pc(); __ push_i();
2000 bep = cep = sep = iep; // there aren't any
2001 vep = __ pc(); __ bind(L); // fall through
2002 generate_and_dispatch(t);
2003 }
2004
2005 // --------------------------------------------------------------------------------
2006
2007
2008 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2009 : TemplateInterpreterGenerator(code) {
2010 generate_all(); // down here so it can be "virtual"
2011 }
2012
2013 // --------------------------------------------------------------------------------
2014
2015 // Non-product code
2016 #ifndef PRODUCT
2017 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2018 address entry = __ pc();
2019
2020 __ push(state);
2021 __ mov(O7, Lscratch); // protect return address within interpreter
2022
2023 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
2024 __ mov( Otos_l2, G3_scratch );
2025 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
2026 __ mov(Lscratch, O7); // restore return address
2027 __ pop(state);
2028 __ retl();
2029 __ delayed()->nop();
2030
2031 return entry;
2032 }
2033
2034
2035 // helpers for generate_and_dispatch
2036
2037 void TemplateInterpreterGenerator::count_bytecode() {
2038 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2039 }
2040
2041
2042 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2043 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2044 }
2045
2046
2047 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2048 AddressLiteral index (&BytecodePairHistogram::_index);
2049 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2050
2051 // get index, shift out old bytecode, bring in new bytecode, and store it
2052 // _index = (_index >> log2_number_of_codes) |
2053 // (bytecode << log2_number_of_codes);
2054
2055 __ load_contents(index, G4_scratch);
2056 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2057 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
2058 __ or3( G3_scratch, G4_scratch, G4_scratch );
2059 __ store_contents(G4_scratch, index, G3_scratch);
2060
2061 // bump bucket contents
2062 // _counters[_index] ++;
2063
2064 __ set(counters, G3_scratch); // loads into G3_scratch
2065 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
2066 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
2067 __ ld (G3_scratch, 0, G4_scratch);
2068 __ inc (G4_scratch);
2069 __ st (G4_scratch, 0, G3_scratch);
2070 }
2071
2072
2073 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2074 // Call a little run-time stub to avoid blow-up for each bytecode.
2075 // The run-time runtime saves the right registers, depending on
2076 // the tosca in-state for the given template.
2077 address entry = Interpreter::trace_code(t->tos_in());
2078 guarantee(entry != NULL, "entry must have been generated");
2079 __ call(entry, relocInfo::none);
2080 __ delayed()->nop();
2081 }
2082
2083
2084 void TemplateInterpreterGenerator::stop_interpreter_at() {
2085 AddressLiteral counter(&BytecodeCounter::_counter_value);
2086 __ load_contents(counter, G3_scratch);
2087 AddressLiteral stop_at(&StopInterpreterAt);
2088 __ load_ptr_contents(stop_at, G4_scratch);
2089 __ cmp(G3_scratch, G4_scratch);
2090 __ breakpoint_trap(Assembler::equal);
2091 }
2092 #endif // not PRODUCT
2093 #endif // !CC_INTERP