1 /*
2 * Copyright (c) 1997, 2013, 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/macroAssembler.hpp"
27 #include "ci/ciMethod.hpp"
28 #include "interpreter/bytecodeHistogram.hpp"
29 #include "interpreter/interpreter.hpp"
30 #include "interpreter/interpreterGenerator.hpp"
31 #include "interpreter/interpreterRuntime.hpp"
32 #include "interpreter/templateTable.hpp"
33 #include "oops/arrayOop.hpp"
34 #include "oops/methodData.hpp"
35 #include "oops/method.hpp"
36 #include "oops/oop.inline.hpp"
37 #include "prims/jvmtiExport.hpp"
38 #include "prims/jvmtiThreadState.hpp"
39 #include "runtime/arguments.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/frame.inline.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "runtime/stubRoutines.hpp"
44 #include "runtime/synchronizer.hpp"
45 #include "runtime/timer.hpp"
46 #include "runtime/vframeArray.hpp"
47 #include "utilities/debug.hpp"
48 #include "utilities/macros.hpp"
49
50 #ifndef CC_INTERP
51 #ifndef FAST_DISPATCH
52 #define FAST_DISPATCH 1
53 #endif
54 #undef FAST_DISPATCH
55
56
57 // Generation of Interpreter
58 //
59 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
60
61
62 #define __ _masm->
63
64
65 //----------------------------------------------------------------------------------------------------
66
67
68 void InterpreterGenerator::save_native_result(void) {
69 // result potentially in O0/O1: save it across calls
70 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
71
72 // result potentially in F0/F1: save it across calls
73 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
74
75 // save and restore any potential method result value around the unlocking operation
76 __ stf(FloatRegisterImpl::D, F0, d_tmp);
77 #ifdef _LP64
78 __ stx(O0, l_tmp);
79 #else
80 __ std(O0, l_tmp);
81 #endif
82 }
83
84 void InterpreterGenerator::restore_native_result(void) {
85 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
86 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
87
88 // Restore any method result value
89 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
90 #ifdef _LP64
91 __ ldx(l_tmp, O0);
92 #else
93 __ ldd(l_tmp, O0);
94 #endif
95 }
96
97 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
98 assert(!pass_oop || message == NULL, "either oop or message but not both");
99 address entry = __ pc();
100 // expression stack must be empty before entering the VM if an exception happened
101 __ empty_expression_stack();
102 // load exception object
103 __ set((intptr_t)name, G3_scratch);
104 if (pass_oop) {
105 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
106 } else {
107 __ set((intptr_t)message, G4_scratch);
108 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
109 }
110 // throw exception
111 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
112 AddressLiteral thrower(Interpreter::throw_exception_entry());
113 __ jump_to(thrower, G3_scratch);
114 __ delayed()->nop();
115 return entry;
116 }
117
118 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
119 address entry = __ pc();
120 // expression stack must be empty before entering the VM if an exception
121 // happened
122 __ empty_expression_stack();
123 // load exception object
124 __ call_VM(Oexception,
125 CAST_FROM_FN_PTR(address,
126 InterpreterRuntime::throw_ClassCastException),
127 Otos_i);
128 __ should_not_reach_here();
129 return entry;
130 }
131
132
133 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
134 address entry = __ pc();
135 // expression stack must be empty before entering the VM if an exception happened
136 __ empty_expression_stack();
137 // convention: expect aberrant index in register G3_scratch, then shuffle the
138 // index to G4_scratch for the VM call
139 __ mov(G3_scratch, G4_scratch);
140 __ set((intptr_t)name, G3_scratch);
141 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
142 __ should_not_reach_here();
143 return entry;
144 }
145
146
147 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
148 address entry = __ pc();
149 // expression stack must be empty before entering the VM if an exception happened
150 __ empty_expression_stack();
151 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
152 __ should_not_reach_here();
153 return entry;
154 }
155
156
157 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
158 address entry = __ pc();
159
160 if (state == atos) {
161 __ profile_return_type(O0, G3_scratch, G1_scratch);
162 }
163
164 #if !defined(_LP64) && defined(COMPILER2)
165 // All return values are where we want them, except for Longs. C2 returns
166 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
167 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
168 // build even if we are returning from interpreted we just do a little
169 // stupid shuffing.
170 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
171 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
172 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
173
174 if (state == ltos) {
175 __ srl (G1, 0, O1);
176 __ srlx(G1, 32, O0);
177 }
178 #endif // !_LP64 && COMPILER2
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 const Register cache = G3_scratch;
189 const Register index = G1_scratch;
190 __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
191
192 const Register flags = cache;
193 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
194 const Register parameter_size = flags;
195 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words
196 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes
197 __ add(Lesp, parameter_size, Lesp); // pop arguments
198 __ dispatch_next(state, step);
199
200 return entry;
201 }
202
203
204 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
205 address entry = __ pc();
206 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
207 { Label L;
208 Address exception_addr(G2_thread, Thread::pending_exception_offset());
209 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
210 __ br_null_short(Gtemp, Assembler::pt, L);
211 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
212 __ should_not_reach_here();
213 __ bind(L);
214 }
215 __ dispatch_next(state, step);
216 return entry;
217 }
218
219 // A result handler converts/unboxes a native call result into
220 // a java interpreter/compiler result. The current frame is an
221 // interpreter frame. The activation frame unwind code must be
222 // consistent with that of TemplateTable::_return(...). In the
223 // case of native methods, the caller's SP was not modified.
224 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
225 address entry = __ pc();
226 Register Itos_i = Otos_i ->after_save();
227 Register Itos_l = Otos_l ->after_save();
228 Register Itos_l1 = Otos_l1->after_save();
229 Register Itos_l2 = Otos_l2->after_save();
230 switch (type) {
231 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
232 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
233 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
234 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
235 case T_LONG :
236 #ifndef _LP64
237 __ mov(O1, Itos_l2); // move other half of long
238 #endif // ifdef or no ifdef, fall through to the T_INT case
239 case T_INT : __ mov(O0, Itos_i); break;
240 case T_VOID : /* nothing to do */ break;
241 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
242 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
243 case T_OBJECT :
244 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
245 __ verify_oop(Itos_i);
246 break;
247 default : ShouldNotReachHere();
248 }
249 __ ret(); // return from interpreter activation
250 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
251 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
252 return entry;
253 }
254
255 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
256 address entry = __ pc();
257 __ push(state);
258 __ call_VM(noreg, runtime_entry);
259 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
260 return entry;
261 }
262
263
264 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
265 address entry = __ pc();
266 __ dispatch_next(state);
267 return entry;
268 }
269
270 //
271 // Helpers for commoning out cases in the various type of method entries.
272 //
273
274 // increment invocation count & check for overflow
275 //
276 // Note: checking for negative value instead of overflow
277 // so we have a 'sticky' overflow test
278 //
279 // Lmethod: method
280 // ??: invocation counter
281 //
282 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
283 // Note: In tiered we increment either counters in MethodCounters* or in
284 // MDO depending if we're profiling or not.
285 const Register Rcounters = G3_scratch;
286 Label done;
287
288 if (TieredCompilation) {
289 const int increment = InvocationCounter::count_increment;
290 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
291 Label no_mdo;
292 if (ProfileInterpreter) {
293 // If no method data exists, go to profile_continue.
294 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
295 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
296 // Increment counter
297 Address mdo_invocation_counter(G4_scratch,
298 in_bytes(MethodData::invocation_counter_offset()) +
299 in_bytes(InvocationCounter::counter_offset()));
300 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
301 G3_scratch, Lscratch,
302 Assembler::zero, overflow);
303 __ ba_short(done);
304 }
305
306 // Increment counter in MethodCounters*
307 __ bind(no_mdo);
308 Address invocation_counter(Rcounters,
309 in_bytes(MethodCounters::invocation_counter_offset()) +
310 in_bytes(InvocationCounter::counter_offset()));
311 __ get_method_counters(Lmethod, Rcounters, done);
312 __ increment_mask_and_jump(invocation_counter, increment, mask,
313 G4_scratch, Lscratch,
314 Assembler::zero, overflow);
315 __ bind(done);
316 } else {
317 // Update standard invocation counters
318 __ get_method_counters(Lmethod, Rcounters, done);
319 __ increment_invocation_counter(Rcounters, O0, G4_scratch);
320 if (ProfileInterpreter) {
321 Address interpreter_invocation_counter(Rcounters,
322 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
323 __ ld(interpreter_invocation_counter, G4_scratch);
324 __ inc(G4_scratch);
325 __ st(G4_scratch, interpreter_invocation_counter);
326 }
327
328 if (ProfileInterpreter && profile_method != NULL) {
329 // Test to see if we should create a method data oop
330 AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
331 __ load_contents(profile_limit, G3_scratch);
332 __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
333
334 // if no method data exists, go to profile_method
335 __ test_method_data_pointer(*profile_method);
336 }
337
338 AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
339 __ load_contents(invocation_limit, G3_scratch);
340 __ cmp(O0, G3_scratch);
341 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
342 __ delayed()->nop();
343 __ bind(done);
344 }
345
346 }
347
348 // Allocate monitor and lock method (asm interpreter)
349 // ebx - Method*
350 //
351 void InterpreterGenerator::lock_method(void) {
352 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
353
354 #ifdef ASSERT
355 { Label ok;
356 __ btst(JVM_ACC_SYNCHRONIZED, O0);
357 __ br( Assembler::notZero, false, Assembler::pt, ok);
358 __ delayed()->nop();
359 __ stop("method doesn't need synchronization");
360 __ bind(ok);
361 }
362 #endif // ASSERT
363
364 // get synchronization object to O0
365 { Label done;
366 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
367 __ btst(JVM_ACC_STATIC, O0);
368 __ br( Assembler::zero, true, Assembler::pt, done);
369 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
370
371 __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0);
372 __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0);
373 __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0);
374
375 // lock the mirror, not the Klass*
376 __ ld_ptr( O0, mirror_offset, O0);
377
378 #ifdef ASSERT
379 __ tst(O0);
380 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
381 #endif // ASSERT
382
383 __ bind(done);
384 }
385
386 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
387 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
388 // __ untested("lock_object from method entry");
389 __ lock_object(Lmonitors, O0);
390 }
391
392
393 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
394 Register Rscratch,
395 Register Rscratch2) {
396 const int page_size = os::vm_page_size();
397 Label after_frame_check;
398
399 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
400
401 __ set(page_size, Rscratch);
402 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
403
404 // get the stack base, and in debug, verify it is non-zero
405 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
406 #ifdef ASSERT
407 Label base_not_zero;
408 __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
409 __ stop("stack base is zero in generate_stack_overflow_check");
410 __ bind(base_not_zero);
411 #endif
412
413 // get the stack size, and in debug, verify it is non-zero
414 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
415 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
416 #ifdef ASSERT
417 Label size_not_zero;
418 __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
419 __ stop("stack size is zero in generate_stack_overflow_check");
420 __ bind(size_not_zero);
421 #endif
422
423 // compute the beginning of the protected zone minus the requested frame size
424 __ sub( Rscratch, Rscratch2, Rscratch );
425 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
426 __ add( Rscratch, Rscratch2, Rscratch );
427
428 // Add in the size of the frame (which is the same as subtracting it from the
429 // SP, which would take another register
430 __ add( Rscratch, Rframe_size, Rscratch );
431
432 // the frame is greater than one page in size, so check against
433 // the bottom of the stack
434 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
435
436 // the stack will overflow, throw an exception
437
438 // Note that SP is restored to sender's sp (in the delay slot). This
439 // is necessary if the sender's frame is an extended compiled frame
440 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
441 // adaptations.
442
443 // Note also that the restored frame is not necessarily interpreted.
444 // Use the shared runtime version of the StackOverflowError.
445 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
446 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
447 __ jump_to(stub, Rscratch);
448 __ delayed()->mov(O5_savedSP, SP);
449
450 // if you get to here, then there is enough stack space
451 __ bind( after_frame_check );
452 }
453
454
455 //
456 // Generate a fixed interpreter frame. This is identical setup for interpreted
457 // methods and for native methods hence the shared code.
458
459 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
460 //
461 //
462 // The entry code sets up a new interpreter frame in 4 steps:
463 //
464 // 1) Increase caller's SP by for the extra local space needed:
465 // (check for overflow)
466 // Efficient implementation of xload/xstore bytecodes requires
467 // that arguments and non-argument locals are in a contigously
468 // addressable memory block => non-argument locals must be
469 // allocated in the caller's frame.
470 //
471 // 2) Create a new stack frame and register window:
472 // The new stack frame must provide space for the standard
473 // register save area, the maximum java expression stack size,
474 // the monitor slots (0 slots initially), and some frame local
475 // scratch locations.
476 //
477 // 3) The following interpreter activation registers must be setup:
478 // Lesp : expression stack pointer
479 // Lbcp : bytecode pointer
480 // Lmethod : method
481 // Llocals : locals pointer
482 // Lmonitors : monitor pointer
483 // LcpoolCache: constant pool cache
484 //
485 // 4) Initialize the non-argument locals if necessary:
486 // Non-argument locals may need to be initialized to NULL
487 // for GC to work. If the oop-map information is accurate
488 // (in the absence of the JSR problem), no initialization
489 // is necessary.
490 //
491 // (gri - 2/25/2000)
492
493
494 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
495
496 const int extra_space =
497 rounded_vm_local_words + // frame local scratch space
498 Method::extra_stack_entries() + // extra stack for jsr 292
499 frame::memory_parameter_word_sp_offset + // register save area
500 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
501
502 const Register Glocals_size = G3;
503 const Register RconstMethod = Glocals_size;
504 const Register Otmp1 = O3;
505 const Register Otmp2 = O4;
506 // Lscratch can't be used as a temporary because the call_stub uses
507 // it to assert that the stack frame was setup correctly.
508 const Address constMethod (G5_method, Method::const_offset());
509 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
510
511 __ ld_ptr( constMethod, RconstMethod );
512 __ lduh( size_of_parameters, Glocals_size);
513
514 // Gargs points to first local + BytesPerWord
515 // Set the saved SP after the register window save
516 //
517 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
518 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
519 __ add(Gargs, Otmp1, Gargs);
520
521 if (native_call) {
522 __ calc_mem_param_words( Glocals_size, Gframe_size );
523 __ add( Gframe_size, extra_space, Gframe_size);
524 __ round_to( Gframe_size, WordsPerLong );
525 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
526 } else {
527
528 //
529 // Compute number of locals in method apart from incoming parameters
530 //
531 const Address size_of_locals (Otmp1, ConstMethod::size_of_locals_offset());
532 __ ld_ptr( constMethod, Otmp1 );
533 __ lduh( size_of_locals, Otmp1 );
534 __ sub( Otmp1, Glocals_size, Glocals_size );
535 __ round_to( Glocals_size, WordsPerLong );
536 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
537
538 // see if the frame is greater than one page in size. If so,
539 // then we need to verify there is enough stack space remaining
540 // Frame_size = (max_stack + extra_space) * BytesPerWord;
541 __ ld_ptr( constMethod, Gframe_size );
542 __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size );
543 __ add( Gframe_size, extra_space, Gframe_size );
544 __ round_to( Gframe_size, WordsPerLong );
545 __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
546
547 // Add in java locals size for stack overflow check only
548 __ add( Gframe_size, Glocals_size, Gframe_size );
549
550 const Register Otmp2 = O4;
551 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
552 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
553
554 __ sub( Gframe_size, Glocals_size, Gframe_size);
555
556 //
557 // bump SP to accomodate the extra locals
558 //
559 __ sub( SP, Glocals_size, SP );
560 }
561
562 //
563 // now set up a stack frame with the size computed above
564 //
565 __ neg( Gframe_size );
566 __ save( SP, Gframe_size, SP );
567
568 //
569 // now set up all the local cache registers
570 //
571 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
572 // that all present references to Lbyte_code initialize the register
573 // immediately before use
574 if (native_call) {
575 __ mov(G0, Lbcp);
576 } else {
577 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
578 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
579 }
580 __ mov( G5_method, Lmethod); // set Lmethod
581 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
582 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
583 #ifdef _LP64
584 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
585 #endif
586 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
587
588 // setup interpreter activation registers
589 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
590
591 if (ProfileInterpreter) {
592 #ifdef FAST_DISPATCH
593 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
594 // they both use I2.
595 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
596 #endif // FAST_DISPATCH
597 __ set_method_data_pointer();
598 }
599
600 }
601
602 // Empty method, generate a very fast return.
603
604 address InterpreterGenerator::generate_empty_entry(void) {
605
606 // A method that does nother but return...
607
608 address entry = __ pc();
609 Label slow_path;
610
611 // do nothing for empty methods (do not even increment invocation counter)
612 if ( UseFastEmptyMethods) {
613 // If we need a safepoint check, generate full interpreter entry.
614 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
615 __ set(sync_state, G3_scratch);
616 __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
617
618 // Code: _return
619 __ retl();
620 __ delayed()->mov(O5_savedSP, SP);
621
622 __ bind(slow_path);
623 (void) generate_normal_entry(false);
624
625 return entry;
626 }
627 return NULL;
628 }
629
630 // Call an accessor method (assuming it is resolved, otherwise drop into
631 // vanilla (slow path) entry
632
633 // Generates code to elide accessor methods
634 // Uses G3_scratch and G1_scratch as scratch
635 address InterpreterGenerator::generate_accessor_entry(void) {
636
637 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
638 // parameter size = 1
639 // Note: We can only use this code if the getfield has been resolved
640 // and if we don't have a null-pointer exception => check for
641 // these conditions first and use slow path if necessary.
642 address entry = __ pc();
643 Label slow_path;
644
645
646 // XXX: for compressed oops pointer loading and decoding doesn't fit in
647 // delay slot and damages G1
648 if ( UseFastAccessorMethods && !UseCompressedOops ) {
649 // Check if we need to reach a safepoint and generate full interpreter
650 // frame if so.
651 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
652 __ load_contents(sync_state, G3_scratch);
653 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
654 __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
655
656 // Check if local 0 != NULL
657 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
658 // check if local 0 == NULL and go the slow path
659 __ br_null_short(Otos_i, Assembler::pn, slow_path);
660
661
662 // read first instruction word and extract bytecode @ 1 and index @ 2
663 // get first 4 bytes of the bytecodes (big endian!)
664 __ ld_ptr(G5_method, Method::const_offset(), G1_scratch);
665 __ ld(G1_scratch, ConstMethod::codes_offset(), G1_scratch);
666
667 // move index @ 2 far left then to the right most two bytes.
668 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
669 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
670 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
671
672 // get constant pool cache
673 __ ld_ptr(G5_method, Method::const_offset(), G3_scratch);
674 __ ld_ptr(G3_scratch, ConstMethod::constants_offset(), G3_scratch);
675 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
676
677 // get specific constant pool cache entry
678 __ add(G3_scratch, G1_scratch, G3_scratch);
679
680 // Check the constant Pool cache entry to see if it has been resolved.
681 // If not, need the slow path.
682 ByteSize cp_base_offset = ConstantPoolCache::base_offset();
683 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
684 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
685 __ and3(G1_scratch, 0xFF, G1_scratch);
686 __ cmp_and_br_short(G1_scratch, Bytecodes::_getfield, Assembler::notEqual, Assembler::pn, slow_path);
687
688 // Get the type and return field offset from the constant pool cache
689 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
690 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
691
692 Label xreturn_path;
693 // Need to differentiate between igetfield, agetfield, bgetfield etc.
694 // because they are different sizes.
695 // Get the type from the constant pool cache
696 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
697 // Make sure we don't need to mask G1_scratch after the above shift
698 ConstantPoolCacheEntry::verify_tos_state_shift();
699 __ cmp(G1_scratch, atos );
700 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
701 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
702 __ cmp(G1_scratch, itos);
703 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
704 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
705 __ cmp(G1_scratch, stos);
706 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
707 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
708 __ cmp(G1_scratch, ctos);
709 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
710 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
711 #ifdef ASSERT
712 __ cmp(G1_scratch, btos);
713 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
714 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
715 __ should_not_reach_here();
716 #endif
717 __ ldsb(Otos_i, G3_scratch, Otos_i);
718 __ bind(xreturn_path);
719
720 // _ireturn/_areturn
721 __ retl(); // return from leaf routine
722 __ delayed()->mov(O5_savedSP, SP);
723
724 // Generate regular method entry
725 __ bind(slow_path);
726 (void) generate_normal_entry(false);
727 return entry;
728 }
729 return NULL;
730 }
731
732 // Method entry for java.lang.ref.Reference.get.
733 address InterpreterGenerator::generate_Reference_get_entry(void) {
734 #if INCLUDE_ALL_GCS
735 // Code: _aload_0, _getfield, _areturn
736 // parameter size = 1
737 //
738 // The code that gets generated by this routine is split into 2 parts:
739 // 1. The "intrinsified" code for G1 (or any SATB based GC),
740 // 2. The slow path - which is an expansion of the regular method entry.
741 //
742 // Notes:-
743 // * In the G1 code we do not check whether we need to block for
744 // a safepoint. If G1 is enabled then we must execute the specialized
745 // code for Reference.get (except when the Reference object is null)
746 // so that we can log the value in the referent field with an SATB
747 // update buffer.
748 // If the code for the getfield template is modified so that the
749 // G1 pre-barrier code is executed when the current method is
750 // Reference.get() then going through the normal method entry
751 // will be fine.
752 // * The G1 code can, however, check the receiver object (the instance
753 // of java.lang.Reference) and jump to the slow path if null. If the
754 // Reference object is null then we obviously cannot fetch the referent
755 // and so we don't need to call the G1 pre-barrier. Thus we can use the
756 // regular method entry code to generate the NPE.
757 //
758 // This code is based on generate_accessor_enty.
759
760 address entry = __ pc();
761
762 const int referent_offset = java_lang_ref_Reference::referent_offset;
763 guarantee(referent_offset > 0, "referent offset not initialized");
764
765 if (UseG1GC) {
766 Label slow_path;
767
768 // In the G1 code we don't check if we need to reach a safepoint. We
769 // continue and the thread will safepoint at the next bytecode dispatch.
770
771 // Check if local 0 != NULL
772 // If the receiver is null then it is OK to jump to the slow path.
773 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
774 // check if local 0 == NULL and go the slow path
775 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
776
777
778 // Load the value of the referent field.
779 if (Assembler::is_simm13(referent_offset)) {
780 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
781 } else {
782 __ set(referent_offset, G3_scratch);
783 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
784 }
785
786 // Generate the G1 pre-barrier code to log the value of
787 // the referent field in an SATB buffer. Note with
788 // these parameters the pre-barrier does not generate
789 // the load of the previous value
790
791 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
792 Otos_i /* pre_val */,
793 G3_scratch /* tmp */,
794 true /* preserve_o_regs */);
795
796 // _areturn
797 __ retl(); // return from leaf routine
798 __ delayed()->mov(O5_savedSP, SP);
799
800 // Generate regular method entry
801 __ bind(slow_path);
802 (void) generate_normal_entry(false);
803 return entry;
804 }
805 #endif // INCLUDE_ALL_GCS
806
807 // If G1 is not enabled then attempt to go through the accessor entry point
808 // Reference.get is an accessor
809 return generate_accessor_entry();
810 }
811
812 //
813 // Interpreter stub for calling a native method. (asm interpreter)
814 // This sets up a somewhat different looking stack for calling the native method
815 // than the typical interpreter frame setup.
816 //
817
818 address InterpreterGenerator::generate_native_entry(bool synchronized) {
819 address entry = __ pc();
820
821 // the following temporary registers are used during frame creation
822 const Register Gtmp1 = G3_scratch ;
823 const Register Gtmp2 = G1_scratch;
824 bool inc_counter = UseCompiler || CountCompiledCalls;
825
826 // make sure registers are different!
827 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
828
829 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
830
831 const Register Glocals_size = G3;
832 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
833
834 // make sure method is native & not abstract
835 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
836 #ifdef ASSERT
837 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
838 {
839 Label L;
840 __ btst(JVM_ACC_NATIVE, Gtmp1);
841 __ br(Assembler::notZero, false, Assembler::pt, L);
842 __ delayed()->nop();
843 __ stop("tried to execute non-native method as native");
844 __ bind(L);
845 }
846 { Label L;
847 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
848 __ br(Assembler::zero, false, Assembler::pt, L);
849 __ delayed()->nop();
850 __ stop("tried to execute abstract method as non-abstract");
851 __ bind(L);
852 }
853 #endif // ASSERT
854
855 // generate the code to allocate the interpreter stack frame
856 generate_fixed_frame(true);
857
858 //
859 // No locals to initialize for native method
860 //
861
862 // this slot will be set later, we initialize it to null here just in
863 // case we get a GC before the actual value is stored later
864 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
865
866 const Address do_not_unlock_if_synchronized(G2_thread,
867 JavaThread::do_not_unlock_if_synchronized_offset());
868 // Since at this point in the method invocation the exception handler
869 // would try to exit the monitor of synchronized methods which hasn't
870 // been entered yet, we set the thread local variable
871 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
872 // runtime, exception handling i.e. unlock_if_synchronized_method will
873 // check this thread local flag.
874 // This flag has two effects, one is to force an unwind in the topmost
875 // interpreter frame and not perform an unlock while doing so.
876
877 __ movbool(true, G3_scratch);
878 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
879
880 // increment invocation counter and check for overflow
881 //
882 // Note: checking for negative value instead of overflow
883 // so we have a 'sticky' overflow test (may be of
884 // importance as soon as we have true MT/MP)
885 Label invocation_counter_overflow;
886 Label Lcontinue;
887 if (inc_counter) {
888 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
889
890 }
891 __ bind(Lcontinue);
892
893 bang_stack_shadow_pages(true);
894
895 // reset the _do_not_unlock_if_synchronized flag
896 __ stbool(G0, do_not_unlock_if_synchronized);
897
898 // check for synchronized methods
899 // Must happen AFTER invocation_counter check and stack overflow check,
900 // so method is not locked if overflows.
901
902 if (synchronized) {
903 lock_method();
904 } else {
905 #ifdef ASSERT
906 { Label ok;
907 __ ld(Laccess_flags, O0);
908 __ btst(JVM_ACC_SYNCHRONIZED, O0);
909 __ br( Assembler::zero, false, Assembler::pt, ok);
910 __ delayed()->nop();
911 __ stop("method needs synchronization");
912 __ bind(ok);
913 }
914 #endif // ASSERT
915 }
916
917
918 // start execution
919 __ verify_thread();
920
921 // JVMTI support
922 __ notify_method_entry();
923
924 // native call
925
926 // (note that O0 is never an oop--at most it is a handle)
927 // It is important not to smash any handles created by this call,
928 // until any oop handle in O0 is dereferenced.
929
930 // (note that the space for outgoing params is preallocated)
931
932 // get signature handler
933 { Label L;
934 Address signature_handler(Lmethod, Method::signature_handler_offset());
935 __ ld_ptr(signature_handler, G3_scratch);
936 __ br_notnull_short(G3_scratch, Assembler::pt, L);
937 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
938 __ ld_ptr(signature_handler, G3_scratch);
939 __ bind(L);
940 }
941
942 // Push a new frame so that the args will really be stored in
943 // Copy a few locals across so the new frame has the variables
944 // we need but these values will be dead at the jni call and
945 // therefore not gc volatile like the values in the current
946 // frame (Lmethod in particular)
947
948 // Flush the method pointer to the register save area
949 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
950 __ mov(Llocals, O1);
951
952 // calculate where the mirror handle body is allocated in the interpreter frame:
953 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
954
955 // Calculate current frame size
956 __ sub(SP, FP, O3); // Calculate negative of current frame size
957 __ save(SP, O3, SP); // Allocate an identical sized frame
958
959 // Note I7 has leftover trash. Slow signature handler will fill it in
960 // should we get there. Normal jni call will set reasonable last_Java_pc
961 // below (and fix I7 so the stack trace doesn't have a meaningless frame
962 // in it).
963
964 // Load interpreter frame's Lmethod into same register here
965
966 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
967
968 __ mov(I1, Llocals);
969 __ mov(I2, Lscratch2); // save the address of the mirror
970
971
972 // ONLY Lmethod and Llocals are valid here!
973
974 // call signature handler, It will move the arg properly since Llocals in current frame
975 // matches that in outer frame
976
977 __ callr(G3_scratch, 0);
978 __ delayed()->nop();
979
980 // Result handler is in Lscratch
981
982 // Reload interpreter frame's Lmethod since slow signature handler may block
983 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
984
985 { Label not_static;
986
987 __ ld(Laccess_flags, O0);
988 __ btst(JVM_ACC_STATIC, O0);
989 __ br( Assembler::zero, false, Assembler::pt, not_static);
990 // get native function entry point(O0 is a good temp until the very end)
991 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
992 // for static methods insert the mirror argument
993 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
994
995 __ ld_ptr(Lmethod, Method:: const_offset(), O1);
996 __ ld_ptr(O1, ConstMethod::constants_offset(), O1);
997 __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
998 __ ld_ptr(O1, mirror_offset, O1);
999 #ifdef ASSERT
1000 if (!PrintSignatureHandlers) // do not dirty the output with this
1001 { Label L;
1002 __ br_notnull_short(O1, Assembler::pt, L);
1003 __ stop("mirror is missing");
1004 __ bind(L);
1005 }
1006 #endif // ASSERT
1007 __ st_ptr(O1, Lscratch2, 0);
1008 __ mov(Lscratch2, O1);
1009 __ bind(not_static);
1010 }
1011
1012 // At this point, arguments have been copied off of stack into
1013 // their JNI positions, which are O1..O5 and SP[68..].
1014 // Oops are boxed in-place on the stack, with handles copied to arguments.
1015 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1016
1017 #ifdef ASSERT
1018 { Label L;
1019 __ br_notnull_short(O0, Assembler::pt, L);
1020 __ stop("native entry point is missing");
1021 __ bind(L);
1022 }
1023 #endif // ASSERT
1024
1025 //
1026 // setup the frame anchor
1027 //
1028 // The scavenge function only needs to know that the PC of this frame is
1029 // in the interpreter method entry code, it doesn't need to know the exact
1030 // PC and hence we can use O7 which points to the return address from the
1031 // previous call in the code stream (signature handler function)
1032 //
1033 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1034 // we have pushed the extra frame in order to protect the volatile register(s)
1035 // in that frame when we return from the jni call
1036 //
1037
1038 __ set_last_Java_frame(FP, O7);
1039 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1040 // not meaningless information that'll confuse me.
1041
1042 // flush the windows now. We don't care about the current (protection) frame
1043 // only the outer frames
1044
1045 __ flushw();
1046
1047 // mark windows as flushed
1048 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1049 __ set(JavaFrameAnchor::flushed, G3_scratch);
1050 __ st(G3_scratch, flags);
1051
1052 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1053
1054 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1055 #ifdef ASSERT
1056 { Label L;
1057 __ ld(thread_state, G3_scratch);
1058 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1059 __ stop("Wrong thread state in native stub");
1060 __ bind(L);
1061 }
1062 #endif // ASSERT
1063 __ set(_thread_in_native, G3_scratch);
1064 __ st(G3_scratch, thread_state);
1065
1066 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1067 __ save_thread(L7_thread_cache); // save Gthread
1068 __ callr(O0, 0);
1069 __ delayed()->
1070 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1071
1072 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1073
1074 __ restore_thread(L7_thread_cache); // restore G2_thread
1075 __ reinit_heapbase();
1076
1077 // must we block?
1078
1079 // Block, if necessary, before resuming in _thread_in_Java state.
1080 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1081 { Label no_block;
1082 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1083
1084 // Switch thread to "native transition" state before reading the synchronization state.
1085 // This additional state is necessary because reading and testing the synchronization
1086 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1087 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1088 // VM thread changes sync state to synchronizing and suspends threads for GC.
1089 // Thread A is resumed to finish this native method, but doesn't block here since it
1090 // didn't see any synchronization is progress, and escapes.
1091 __ set(_thread_in_native_trans, G3_scratch);
1092 __ st(G3_scratch, thread_state);
1093 if(os::is_MP()) {
1094 if (UseMembar) {
1095 // Force this write out before the read below
1096 __ membar(Assembler::StoreLoad);
1097 } else {
1098 // Write serialization page so VM thread can do a pseudo remote membar.
1099 // We use the current thread pointer to calculate a thread specific
1100 // offset to write to within the page. This minimizes bus traffic
1101 // due to cache line collision.
1102 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1103 }
1104 }
1105 __ load_contents(sync_state, G3_scratch);
1106 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1107
1108 Label L;
1109 __ br(Assembler::notEqual, false, Assembler::pn, L);
1110 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1111 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1112 __ bind(L);
1113
1114 // Block. Save any potential method result value before the operation and
1115 // use a leaf call to leave the last_Java_frame setup undisturbed.
1116 save_native_result();
1117 __ call_VM_leaf(L7_thread_cache,
1118 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1119 G2_thread);
1120
1121 // Restore any method result value
1122 restore_native_result();
1123 __ bind(no_block);
1124 }
1125
1126 // Clear the frame anchor now
1127
1128 __ reset_last_Java_frame();
1129
1130 // Move the result handler address
1131 __ mov(Lscratch, G3_scratch);
1132 // return possible result to the outer frame
1133 #ifndef __LP64
1134 __ mov(O0, I0);
1135 __ restore(O1, G0, O1);
1136 #else
1137 __ restore(O0, G0, O0);
1138 #endif /* __LP64 */
1139
1140 // Move result handler to expected register
1141 __ mov(G3_scratch, Lscratch);
1142
1143 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1144 // switch to thread_in_Java.
1145
1146 __ set(_thread_in_Java, G3_scratch);
1147 __ st(G3_scratch, thread_state);
1148
1149 // reset handle block
1150 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1151 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1152
1153 // If we have an oop result store it where it will be safe for any further gc
1154 // until we return now that we've released the handle it might be protected by
1155
1156 {
1157 Label no_oop, store_result;
1158
1159 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1160 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1161 __ addcc(G0, O0, O0);
1162 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1163 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1164 __ mov(G0, O0);
1165
1166 __ bind(store_result);
1167 // Store it where gc will look for it and result handler expects it.
1168 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1169
1170 __ bind(no_oop);
1171
1172 }
1173
1174
1175 // handle exceptions (exception handling will handle unlocking!)
1176 { Label L;
1177 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1178 __ ld_ptr(exception_addr, Gtemp);
1179 __ br_null_short(Gtemp, Assembler::pt, L);
1180 // Note: This could be handled more efficiently since we know that the native
1181 // method doesn't have an exception handler. We could directly return
1182 // to the exception handler for the caller.
1183 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1184 __ should_not_reach_here();
1185 __ bind(L);
1186 }
1187
1188 // JVMTI support (preserves thread register)
1189 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1190
1191 if (synchronized) {
1192 // save and restore any potential method result value around the unlocking operation
1193 save_native_result();
1194
1195 __ add( __ top_most_monitor(), O1);
1196 __ unlock_object(O1);
1197
1198 restore_native_result();
1199 }
1200
1201 #if defined(COMPILER2) && !defined(_LP64)
1202
1203 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1204 // or compiled so just be safe.
1205
1206 __ sllx(O0, 32, G1); // Shift bits into high G1
1207 __ srl (O1, 0, O1); // Zero extend O1
1208 __ or3 (O1, G1, G1); // OR 64 bits into G1
1209
1210 #endif /* COMPILER2 && !_LP64 */
1211
1212 // dispose of return address and remove activation
1213 #ifdef ASSERT
1214 {
1215 Label ok;
1216 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1217 __ stop("bad I5_savedSP value");
1218 __ should_not_reach_here();
1219 __ bind(ok);
1220 }
1221 #endif
1222 if (TraceJumps) {
1223 // Move target to register that is recordable
1224 __ mov(Lscratch, G3_scratch);
1225 __ JMP(G3_scratch, 0);
1226 } else {
1227 __ jmp(Lscratch, 0);
1228 }
1229 __ delayed()->nop();
1230
1231
1232 if (inc_counter) {
1233 // handle invocation counter overflow
1234 __ bind(invocation_counter_overflow);
1235 generate_counter_overflow(Lcontinue);
1236 }
1237
1238
1239
1240 return entry;
1241 }
1242
1243
1244 // Generic method entry to (asm) interpreter
1245 //------------------------------------------------------------------------------------------------------------------------
1246 //
1247 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1248 address entry = __ pc();
1249
1250 bool inc_counter = UseCompiler || CountCompiledCalls;
1251
1252 // the following temporary registers are used during frame creation
1253 const Register Gtmp1 = G3_scratch ;
1254 const Register Gtmp2 = G1_scratch;
1255
1256 // make sure registers are different!
1257 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1258
1259 const Address constMethod (G5_method, Method::const_offset());
1260 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1261 // and use in the asserts.
1262 const Address access_flags (Lmethod, Method::access_flags_offset());
1263
1264 const Register Glocals_size = G3;
1265 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1266
1267 // make sure method is not native & not abstract
1268 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1269 #ifdef ASSERT
1270 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1271 {
1272 Label L;
1273 __ btst(JVM_ACC_NATIVE, Gtmp1);
1274 __ br(Assembler::zero, false, Assembler::pt, L);
1275 __ delayed()->nop();
1276 __ stop("tried to execute native method as non-native");
1277 __ bind(L);
1278 }
1279 { Label L;
1280 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1281 __ br(Assembler::zero, false, Assembler::pt, L);
1282 __ delayed()->nop();
1283 __ stop("tried to execute abstract method as non-abstract");
1284 __ bind(L);
1285 }
1286 #endif // ASSERT
1287
1288 // generate the code to allocate the interpreter stack frame
1289
1290 generate_fixed_frame(false);
1291
1292 #ifdef FAST_DISPATCH
1293 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1294 // set bytecode dispatch table base
1295 #endif
1296
1297 //
1298 // Code to initialize the extra (i.e. non-parm) locals
1299 //
1300 Register init_value = noreg; // will be G0 if we must clear locals
1301 // The way the code was setup before zerolocals was always true for vanilla java entries.
1302 // It could only be false for the specialized entries like accessor or empty which have
1303 // no extra locals so the testing was a waste of time and the extra locals were always
1304 // initialized. We removed this extra complication to already over complicated code.
1305
1306 init_value = G0;
1307 Label clear_loop;
1308
1309 const Register RconstMethod = O1;
1310 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1311 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1312
1313 // NOTE: If you change the frame layout, this code will need to
1314 // be updated!
1315 __ ld_ptr( constMethod, RconstMethod );
1316 __ lduh( size_of_locals, O2 );
1317 __ lduh( size_of_parameters, O1 );
1318 __ sll( O2, Interpreter::logStackElementSize, O2);
1319 __ sll( O1, Interpreter::logStackElementSize, O1 );
1320 __ sub( Llocals, O2, O2 );
1321 __ sub( Llocals, O1, O1 );
1322
1323 __ bind( clear_loop );
1324 __ inc( O2, wordSize );
1325
1326 __ cmp( O2, O1 );
1327 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1328 __ delayed()->st_ptr( init_value, O2, 0 );
1329
1330 const Address do_not_unlock_if_synchronized(G2_thread,
1331 JavaThread::do_not_unlock_if_synchronized_offset());
1332 // Since at this point in the method invocation the exception handler
1333 // would try to exit the monitor of synchronized methods which hasn't
1334 // been entered yet, we set the thread local variable
1335 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1336 // runtime, exception handling i.e. unlock_if_synchronized_method will
1337 // check this thread local flag.
1338 __ movbool(true, G3_scratch);
1339 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1340
1341 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1342 // increment invocation counter and check for overflow
1343 //
1344 // Note: checking for negative value instead of overflow
1345 // so we have a 'sticky' overflow test (may be of
1346 // importance as soon as we have true MT/MP)
1347 Label invocation_counter_overflow;
1348 Label profile_method;
1349 Label profile_method_continue;
1350 Label Lcontinue;
1351 if (inc_counter) {
1352 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1353 if (ProfileInterpreter) {
1354 __ bind(profile_method_continue);
1355 }
1356 }
1357 __ bind(Lcontinue);
1358
1359 bang_stack_shadow_pages(false);
1360
1361 // reset the _do_not_unlock_if_synchronized flag
1362 __ stbool(G0, do_not_unlock_if_synchronized);
1363
1364 // check for synchronized methods
1365 // Must happen AFTER invocation_counter check and stack overflow check,
1366 // so method is not locked if overflows.
1367
1368 if (synchronized) {
1369 lock_method();
1370 } else {
1371 #ifdef ASSERT
1372 { Label ok;
1373 __ ld(access_flags, O0);
1374 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1375 __ br( Assembler::zero, false, Assembler::pt, ok);
1376 __ delayed()->nop();
1377 __ stop("method needs synchronization");
1378 __ bind(ok);
1379 }
1380 #endif // ASSERT
1381 }
1382
1383 // start execution
1384
1385 __ verify_thread();
1386
1387 // jvmti support
1388 __ notify_method_entry();
1389
1390 // start executing instructions
1391 __ dispatch_next(vtos);
1392
1393
1394 if (inc_counter) {
1395 if (ProfileInterpreter) {
1396 // We have decided to profile this method in the interpreter
1397 __ bind(profile_method);
1398
1399 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1400 __ set_method_data_pointer_for_bcp();
1401 __ ba_short(profile_method_continue);
1402 }
1403
1404 // handle invocation counter overflow
1405 __ bind(invocation_counter_overflow);
1406 generate_counter_overflow(Lcontinue);
1407 }
1408
1409
1410 return entry;
1411 }
1412
1413
1414 //----------------------------------------------------------------------------------------------------
1415 // Entry points & stack frame layout
1416 //
1417 // Here we generate the various kind of entries into the interpreter.
1418 // The two main entry type are generic bytecode methods and native call method.
1419 // These both come in synchronized and non-synchronized versions but the
1420 // frame layout they create is very similar. The other method entry
1421 // types are really just special purpose entries that are really entry
1422 // and interpretation all in one. These are for trivial methods like
1423 // accessor, empty, or special math methods.
1424 //
1425 // When control flow reaches any of the entry types for the interpreter
1426 // the following holds ->
1427 //
1428 // C2 Calling Conventions:
1429 //
1430 // The entry code below assumes that the following registers are set
1431 // when coming in:
1432 // G5_method: holds the Method* of the method to call
1433 // Lesp: points to the TOS of the callers expression stack
1434 // after having pushed all the parameters
1435 //
1436 // The entry code does the following to setup an interpreter frame
1437 // pop parameters from the callers stack by adjusting Lesp
1438 // set O0 to Lesp
1439 // compute X = (max_locals - num_parameters)
1440 // bump SP up by X to accomadate the extra locals
1441 // compute X = max_expression_stack
1442 // + vm_local_words
1443 // + 16 words of register save area
1444 // save frame doing a save sp, -X, sp growing towards lower addresses
1445 // set Lbcp, Lmethod, LcpoolCache
1446 // set Llocals to i0
1447 // set Lmonitors to FP - rounded_vm_local_words
1448 // set Lesp to Lmonitors - 4
1449 //
1450 // The frame has now been setup to do the rest of the entry code
1451
1452 // Try this optimization: Most method entries could live in a
1453 // "one size fits all" stack frame without all the dynamic size
1454 // calculations. It might be profitable to do all this calculation
1455 // statically and approximately for "small enough" methods.
1456
1457 //-----------------------------------------------------------------------------------------------
1458
1459 // C1 Calling conventions
1460 //
1461 // Upon method entry, the following registers are setup:
1462 //
1463 // g2 G2_thread: current thread
1464 // g5 G5_method: method to activate
1465 // g4 Gargs : pointer to last argument
1466 //
1467 //
1468 // Stack:
1469 //
1470 // +---------------+ <--- sp
1471 // | |
1472 // : reg save area :
1473 // | |
1474 // +---------------+ <--- sp + 0x40
1475 // | |
1476 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1477 // | |
1478 // +---------------+ <--- sp + 0x5c
1479 // | |
1480 // : free :
1481 // | |
1482 // +---------------+ <--- Gargs
1483 // | |
1484 // : arguments :
1485 // | |
1486 // +---------------+
1487 // | |
1488 //
1489 //
1490 //
1491 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1492 //
1493 // +---------------+ <--- sp
1494 // | |
1495 // : reg save area :
1496 // | |
1497 // +---------------+ <--- sp + 0x40
1498 // | |
1499 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1500 // | |
1501 // +---------------+ <--- sp + 0x5c
1502 // | |
1503 // : :
1504 // | | <--- Lesp
1505 // +---------------+ <--- Lmonitors (fp - 0x18)
1506 // | VM locals |
1507 // +---------------+ <--- fp
1508 // | |
1509 // : reg save area :
1510 // | |
1511 // +---------------+ <--- fp + 0x40
1512 // | |
1513 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1514 // | |
1515 // +---------------+ <--- fp + 0x5c
1516 // | |
1517 // : free :
1518 // | |
1519 // +---------------+
1520 // | |
1521 // : nonarg locals :
1522 // | |
1523 // +---------------+
1524 // | |
1525 // : arguments :
1526 // | | <--- Llocals
1527 // +---------------+ <--- Gargs
1528 // | |
1529
1530 template<class M> static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1531
1532 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1533 // expression stack, the callee will have callee_extra_locals (so we can account for
1534 // frame extension) and monitor_size for monitors. Basically we need to calculate
1535 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1536 //
1537 //
1538 // The big complicating thing here is that we must ensure that the stack stays properly
1539 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1540 // needs to be aligned for). We are given that the sp (fp) is already aligned by
1541 // the caller so we must ensure that it is properly aligned for our callee.
1542 //
1543 const int rounded_vm_local_words =
1544 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1545 // callee_locals and max_stack are counts, not the size in frame.
1546 const int locals_size =
1547 round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1548 const int max_stack_words = max_stack * Interpreter::stackElementWords;
1549 return (round_to((max_stack_words
1550 + rounded_vm_local_words
1551 + frame::memory_parameter_word_sp_offset), WordsPerLong)
1552 // already rounded
1553 + locals_size + monitor_size);
1554 }
1555
1556 // How much stack a method top interpreter activation needs in words.
1557 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1558
1559 // See call_stub code
1560 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
1561 WordsPerLong); // 7 + register save area
1562
1563 // Save space for one monitor to get into the interpreted method in case
1564 // the method is synchronized
1565 int monitor_size = method->is_synchronized() ?
1566 1*frame::interpreter_frame_monitor_size() : 0;
1567 return size_activation_helper<Method>(method->max_locals(), method->max_stack(),
1568 monitor_size) + call_stub_size;
1569 }
1570
1571 template<class M> int AbstractInterpreter::size_activation(M* method,
1572 int temps,
1573 int popframe_args,
1574 int monitors,
1575 int callee_params,
1576 int callee_locals,
1577 bool is_top_frame) {
1578 // Note: This calculation must exactly parallel the frame setup
1579 // in InterpreterGenerator::generate_fixed_frame.
1580
1581 int monitor_size = monitors * frame::interpreter_frame_monitor_size();
1582
1583 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1584
1585 //
1586 // Note: if you look closely this appears to be doing something much different
1587 // than generate_fixed_frame. What is happening is this. On sparc we have to do
1588 // this dance with interpreter_sp_adjustment because the window save area would
1589 // appear just below the bottom (tos) of the caller's java expression stack. Because
1590 // the interpreter want to have the locals completely contiguous generate_fixed_frame
1591 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1592 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1593 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1594 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1595 // because the oldest frame would have adjust its callers frame and yet that frame
1596 // already exists and isn't part of this array of frames we are unpacking. So at first
1597 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1598 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1599 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1600 // add up. It does seem like it simpler to account for the adjustment here (and remove the
1601 // callee... parameters here). However this would mean that this routine would have to take
1602 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1603 // and run the calling loop in the reverse order. This would also would appear to mean making
1604 // this code aware of what the interactions are when that initial caller fram was an osr or
1605 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
1606 // there is no sense in messing working code.
1607 //
1608
1609 int rounded_cls = round_to((callee_locals - callee_params), WordsPerLong);
1610 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1611
1612 int raw_frame_size = size_activation_helper<M>(rounded_cls, method->max_stack(),
1613 monitor_size);
1614
1615 return raw_frame_size;
1616 }
1617
1618 template int AbstractInterpreter::size_activation<Method>(Method* method,
1619 int temps,
1620 int popframe_args,
1621 int monitors,
1622 int callee_params,
1623 int callee_locals,
1624 bool is_top_frame);
1625
1626 template int AbstractInterpreter::size_activation<ciMethod>(ciMethod* method,
1627 int temps,
1628 int popframe_args,
1629 int monitors,
1630 int callee_params,
1631 int callee_locals,
1632 bool is_top_frame);
1633
1634 void AbstractInterpreter::layout_activation(Method* method,
1635 int tempcount,
1636 int popframe_extra_args,
1637 int moncount,
1638 int caller_actual_parameters,
1639 int callee_param_count,
1640 int callee_local_count,
1641 frame* caller,
1642 frame* interpreter_frame,
1643 bool is_top_frame,
1644 bool is_bottom_frame) {
1645 // Set up the following variables:
1646 // - Lmethod
1647 // - Llocals
1648 // - Lmonitors (to the indicated number of monitors)
1649 // - Lesp (to the indicated number of temps)
1650 // The frame caller on entry is a description of the caller of the
1651 // frame we are about to layout. We are guaranteed that we will be
1652 // able to fill in a new interpreter frame as its callee (i.e. the
1653 // stack space is allocated and the amount was determined by an
1654 // earlier call to the size_activation() method). On return caller
1655 // while describe the interpreter frame we just layed out.
1656
1657 // The skeleton frame must already look like an interpreter frame
1658 // even if not fully filled out.
1659 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1660
1661 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1662 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
1663 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1664
1665 intptr_t* fp = interpreter_frame->fp();
1666
1667 JavaThread* thread = JavaThread::current();
1668 RegisterMap map(thread, false);
1669 // More verification that skeleton frame is properly walkable
1670 assert(fp == caller->sp(), "fp must match");
1671
1672 intptr_t* montop = fp - rounded_vm_local_words;
1673
1674 // preallocate monitors (cf. __ add_monitor_to_stack)
1675 intptr_t* monitors = montop - monitor_size;
1676
1677 // preallocate stack space
1678 intptr_t* esp = monitors - 1 -
1679 (tempcount * Interpreter::stackElementWords) -
1680 popframe_extra_args;
1681
1682 int local_words = method->max_locals() * Interpreter::stackElementWords;
1683 NEEDS_CLEANUP;
1684 intptr_t* locals;
1685 if (caller->is_interpreted_frame()) {
1686 // Can force the locals area to end up properly overlapping the top of the expression stack.
1687 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1688 // Note that this computation means we replace size_of_parameters() values from the caller
1689 // interpreter frame's expression stack with our argument locals
1690 int parm_words = caller_actual_parameters * Interpreter::stackElementWords;
1691 locals = Lesp_ptr + parm_words;
1692 int delta = local_words - parm_words;
1693 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1694 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1695 if (!is_bottom_frame) {
1696 // Llast_SP is set below for the current frame to SP (with the
1697 // extra space for the callee's locals). Here we adjust
1698 // Llast_SP for the caller's frame, removing the extra space
1699 // for the current method's locals.
1700 *caller->register_addr(Llast_SP) = *interpreter_frame->register_addr(I5_savedSP);
1701 } else {
1702 assert(*caller->register_addr(Llast_SP) >= *interpreter_frame->register_addr(I5_savedSP), "strange Llast_SP");
1703 }
1704 } else {
1705 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1706 // Don't have Lesp available; lay out locals block in the caller
1707 // adjacent to the register window save area.
1708 //
1709 // Compiled frames do not allocate a varargs area which is why this if
1710 // statement is needed.
1711 //
1712 if (caller->is_compiled_frame()) {
1713 locals = fp + frame::register_save_words + local_words - 1;
1714 } else {
1715 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1716 }
1717 if (!caller->is_entry_frame()) {
1718 // Caller wants his own SP back
1719 int caller_frame_size = caller->cb()->frame_size();
1720 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1721 }
1722 }
1723 if (TraceDeoptimization) {
1724 if (caller->is_entry_frame()) {
1725 // make sure I5_savedSP and the entry frames notion of saved SP
1726 // agree. This assertion duplicate a check in entry frame code
1727 // but catches the failure earlier.
1728 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1729 "would change callers SP");
1730 }
1731 if (caller->is_entry_frame()) {
1732 tty->print("entry ");
1733 }
1734 if (caller->is_compiled_frame()) {
1735 tty->print("compiled ");
1736 if (caller->is_deoptimized_frame()) {
1737 tty->print("(deopt) ");
1738 }
1739 }
1740 if (caller->is_interpreted_frame()) {
1741 tty->print("interpreted ");
1742 }
1743 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1744 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1745 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1746 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1747 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1748 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1749 tty->print_cr("Llocals = 0x%x", locals);
1750 tty->print_cr("Lesp = 0x%x", esp);
1751 tty->print_cr("Lmonitors = 0x%x", monitors);
1752 }
1753
1754 if (method->max_locals() > 0) {
1755 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1756 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1757 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1758 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1759 }
1760 #ifdef _LP64
1761 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1762 #endif
1763
1764 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
1765 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
1766 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
1767 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
1768 // Llast_SP will be same as SP as there is no adapter space
1769 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1770 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1771 #ifdef FAST_DISPATCH
1772 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1773 #endif
1774
1775
1776 #ifdef ASSERT
1777 BasicObjectLock* mp = (BasicObjectLock*)monitors;
1778
1779 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1780 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1781 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
1782 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1783 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1784
1785 // check bounds
1786 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1787 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1788 assert(lo < monitors && montop <= hi, "monitors in bounds");
1789 assert(lo <= esp && esp < monitors, "esp in bounds");
1790 #endif // ASSERT
1791 }
1792
1793 //----------------------------------------------------------------------------------------------------
1794 // Exceptions
1795 void TemplateInterpreterGenerator::generate_throw_exception() {
1796
1797 // Entry point in previous activation (i.e., if the caller was interpreted)
1798 Interpreter::_rethrow_exception_entry = __ pc();
1799 // O0: exception
1800
1801 // entry point for exceptions thrown within interpreter code
1802 Interpreter::_throw_exception_entry = __ pc();
1803 __ verify_thread();
1804 // expression stack is undefined here
1805 // O0: exception, i.e. Oexception
1806 // Lbcp: exception bcx
1807 __ verify_oop(Oexception);
1808
1809
1810 // expression stack must be empty before entering the VM in case of an exception
1811 __ empty_expression_stack();
1812 // find exception handler address and preserve exception oop
1813 // call C routine to find handler and jump to it
1814 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1815 __ push_ptr(O1); // push exception for exception handler bytecodes
1816
1817 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1818 __ delayed()->nop();
1819
1820
1821 // if the exception is not handled in the current frame
1822 // the frame is removed and the exception is rethrown
1823 // (i.e. exception continuation is _rethrow_exception)
1824 //
1825 // Note: At this point the bci is still the bxi for the instruction which caused
1826 // the exception and the expression stack is empty. Thus, for any VM calls
1827 // at this point, GC will find a legal oop map (with empty expression stack).
1828
1829 // in current activation
1830 // tos: exception
1831 // Lbcp: exception bcp
1832
1833 //
1834 // JVMTI PopFrame support
1835 //
1836
1837 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1838 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1839 // Set the popframe_processing bit in popframe_condition indicating that we are
1840 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1841 // popframe handling cycles.
1842
1843 __ ld(popframe_condition_addr, G3_scratch);
1844 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1845 __ stw(G3_scratch, popframe_condition_addr);
1846
1847 // Empty the expression stack, as in normal exception handling
1848 __ empty_expression_stack();
1849 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1850
1851 {
1852 // Check to see whether we are returning to a deoptimized frame.
1853 // (The PopFrame call ensures that the caller of the popped frame is
1854 // either interpreted or compiled and deoptimizes it if compiled.)
1855 // In this case, we can't call dispatch_next() after the frame is
1856 // popped, but instead must save the incoming arguments and restore
1857 // them after deoptimization has occurred.
1858 //
1859 // Note that we don't compare the return PC against the
1860 // deoptimization blob's unpack entry because of the presence of
1861 // adapter frames in C2.
1862 Label caller_not_deoptimized;
1863 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1864 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1865
1866 const Register Gtmp1 = G3_scratch;
1867 const Register Gtmp2 = G1_scratch;
1868 const Register RconstMethod = Gtmp1;
1869 const Address constMethod(Lmethod, Method::const_offset());
1870 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1871
1872 // Compute size of arguments for saving when returning to deoptimized caller
1873 __ ld_ptr(constMethod, RconstMethod);
1874 __ lduh(size_of_parameters, Gtmp1);
1875 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1876 __ sub(Llocals, Gtmp1, Gtmp2);
1877 __ add(Gtmp2, wordSize, Gtmp2);
1878 // Save these arguments
1879 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1880 // Inform deoptimization that it is responsible for restoring these arguments
1881 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1882 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1883 __ st(Gtmp1, popframe_condition_addr);
1884
1885 // Return from the current method
1886 // The caller's SP was adjusted upon method entry to accomodate
1887 // the callee's non-argument locals. Undo that adjustment.
1888 __ ret();
1889 __ delayed()->restore(I5_savedSP, G0, SP);
1890
1891 __ bind(caller_not_deoptimized);
1892 }
1893
1894 // Clear the popframe condition flag
1895 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1896
1897 // Get out of the current method (how this is done depends on the particular compiler calling
1898 // convention that the interpreter currently follows)
1899 // The caller's SP was adjusted upon method entry to accomodate
1900 // the callee's non-argument locals. Undo that adjustment.
1901 __ restore(I5_savedSP, G0, SP);
1902 // The method data pointer was incremented already during
1903 // call profiling. We have to restore the mdp for the current bcp.
1904 if (ProfileInterpreter) {
1905 __ set_method_data_pointer_for_bcp();
1906 }
1907
1908 #if INCLUDE_JVMTI
1909 if (EnableInvokeDynamic) {
1910 Label L_done;
1911
1912 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1913 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1914
1915 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1916 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1917
1918 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1919
1920 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1921 __ delayed()->nop();
1922
1923 __ st_ptr(G1_scratch, Lesp, wordSize);
1924 __ bind(L_done);
1925 }
1926 #endif // INCLUDE_JVMTI
1927
1928 // Resume bytecode interpretation at the current bcp
1929 __ dispatch_next(vtos);
1930 // end of JVMTI PopFrame support
1931
1932 Interpreter::_remove_activation_entry = __ pc();
1933
1934 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1935 __ pop_ptr(Oexception); // get exception
1936
1937 // Intel has the following comment:
1938 //// remove the activation (without doing throws on illegalMonitorExceptions)
1939 // They remove the activation without checking for bad monitor state.
1940 // %%% We should make sure this is the right semantics before implementing.
1941
1942 __ set_vm_result(Oexception);
1943 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1944
1945 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1946
1947 __ get_vm_result(Oexception);
1948 __ verify_oop(Oexception);
1949
1950 const int return_reg_adjustment = frame::pc_return_offset;
1951 Address issuing_pc_addr(I7, return_reg_adjustment);
1952
1953 // We are done with this activation frame; find out where to go next.
1954 // The continuation point will be an exception handler, which expects
1955 // the following registers set up:
1956 //
1957 // Oexception: exception
1958 // Oissuing_pc: the local call that threw exception
1959 // Other On: garbage
1960 // In/Ln: the contents of the caller's register window
1961 //
1962 // We do the required restore at the last possible moment, because we
1963 // need to preserve some state across a runtime call.
1964 // (Remember that the caller activation is unknown--it might not be
1965 // interpreted, so things like Lscratch are useless in the caller.)
1966
1967 // Although the Intel version uses call_C, we can use the more
1968 // compact call_VM. (The only real difference on SPARC is a
1969 // harmlessly ignored [re]set_last_Java_frame, compared with
1970 // the Intel code which lacks this.)
1971 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1972 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1973 __ super_call_VM_leaf(L7_thread_cache,
1974 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1975 G2_thread, Oissuing_pc->after_save());
1976
1977 // The caller's SP was adjusted upon method entry to accomodate
1978 // the callee's non-argument locals. Undo that adjustment.
1979 __ JMP(O0, 0); // return exception handler in caller
1980 __ delayed()->restore(I5_savedSP, G0, SP);
1981
1982 // (same old exception object is already in Oexception; see above)
1983 // Note that an "issuing PC" is actually the next PC after the call
1984 }
1985
1986
1987 //
1988 // JVMTI ForceEarlyReturn support
1989 //
1990
1991 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1992 address entry = __ pc();
1993
1994 __ empty_expression_stack();
1995 __ load_earlyret_value(state);
1996
1997 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1998 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1999
2000 // Clear the earlyret state
2001 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
2002
2003 __ remove_activation(state,
2004 /* throw_monitor_exception */ false,
2005 /* install_monitor_exception */ false);
2006
2007 // The caller's SP was adjusted upon method entry to accomodate
2008 // the callee's non-argument locals. Undo that adjustment.
2009 __ ret(); // return to caller
2010 __ delayed()->restore(I5_savedSP, G0, SP);
2011
2012 return entry;
2013 } // end of JVMTI ForceEarlyReturn support
2014
2015
2016 //------------------------------------------------------------------------------------------------------------------------
2017 // Helper for vtos entry point generation
2018
2019 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) {
2020 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
2021 Label L;
2022 aep = __ pc(); __ push_ptr(); __ ba_short(L);
2023 fep = __ pc(); __ push_f(); __ ba_short(L);
2024 dep = __ pc(); __ push_d(); __ ba_short(L);
2025 lep = __ pc(); __ push_l(); __ ba_short(L);
2026 iep = __ pc(); __ push_i();
2027 bep = cep = sep = iep; // there aren't any
2028 vep = __ pc(); __ bind(L); // fall through
2029 generate_and_dispatch(t);
2030 }
2031
2032 // --------------------------------------------------------------------------------
2033
2034
2035 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2036 : TemplateInterpreterGenerator(code) {
2037 generate_all(); // down here so it can be "virtual"
2038 }
2039
2040 // --------------------------------------------------------------------------------
2041
2042 // Non-product code
2043 #ifndef PRODUCT
2044 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2045 address entry = __ pc();
2046
2047 __ push(state);
2048 __ mov(O7, Lscratch); // protect return address within interpreter
2049
2050 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
2051 __ mov( Otos_l2, G3_scratch );
2052 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
2053 __ mov(Lscratch, O7); // restore return address
2054 __ pop(state);
2055 __ retl();
2056 __ delayed()->nop();
2057
2058 return entry;
2059 }
2060
2061
2062 // helpers for generate_and_dispatch
2063
2064 void TemplateInterpreterGenerator::count_bytecode() {
2065 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2066 }
2067
2068
2069 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2070 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2071 }
2072
2073
2074 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2075 AddressLiteral index (&BytecodePairHistogram::_index);
2076 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2077
2078 // get index, shift out old bytecode, bring in new bytecode, and store it
2079 // _index = (_index >> log2_number_of_codes) |
2080 // (bytecode << log2_number_of_codes);
2081
2082 __ load_contents(index, G4_scratch);
2083 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2084 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
2085 __ or3( G3_scratch, G4_scratch, G4_scratch );
2086 __ store_contents(G4_scratch, index, G3_scratch);
2087
2088 // bump bucket contents
2089 // _counters[_index] ++;
2090
2091 __ set(counters, G3_scratch); // loads into G3_scratch
2092 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
2093 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
2094 __ ld (G3_scratch, 0, G4_scratch);
2095 __ inc (G4_scratch);
2096 __ st (G4_scratch, 0, G3_scratch);
2097 }
2098
2099
2100 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2101 // Call a little run-time stub to avoid blow-up for each bytecode.
2102 // The run-time runtime saves the right registers, depending on
2103 // the tosca in-state for the given template.
2104 address entry = Interpreter::trace_code(t->tos_in());
2105 guarantee(entry != NULL, "entry must have been generated");
2106 __ call(entry, relocInfo::none);
2107 __ delayed()->nop();
2108 }
2109
2110
2111 void TemplateInterpreterGenerator::stop_interpreter_at() {
2112 AddressLiteral counter(&BytecodeCounter::_counter_value);
2113 __ load_contents(counter, G3_scratch);
2114 AddressLiteral stop_at(&StopInterpreterAt);
2115 __ load_ptr_contents(stop_at, G4_scratch);
2116 __ cmp(G3_scratch, G4_scratch);
2117 __ breakpoint_trap(Assembler::equal, Assembler::icc);
2118 }
2119 #endif // not PRODUCT
2120 #endif // !CC_INTERP