1 /*
2 * Copyright (c) 1997, 2016, 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 "interpreter/bytecodeHistogram.hpp"
28 #include "interpreter/interpreter.hpp"
29 #include "interpreter/interpreterRuntime.hpp"
30 #include "interpreter/interp_masm.hpp"
31 #include "interpreter/templateInterpreterGenerator.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 FAST_DISPATCH
51 #define FAST_DISPATCH 1
52 #endif
53 #undef FAST_DISPATCH
54
55 // Size of interpreter code. Increase if too small. Interpreter will
56 // fail with a guarantee ("not enough space for interpreter generation");
57 // if too small.
58 // Run with +PrintInterpreter to get the VM to print out the size.
59 // Max size with JVMTI
60 #ifdef _LP64
61 // The sethi() instruction generates lots more instructions when shell
62 // stack limit is unlimited, so that's why this is much bigger.
63 int TemplateInterpreter::InterpreterCodeSize = 260 * K;
64 #else
65 int TemplateInterpreter::InterpreterCodeSize = 230 * K;
66 #endif
67
68 // Generation of Interpreter
69 //
70 // The TemplateInterpreterGenerator generates the interpreter into Interpreter::_code.
71
72
73 #define __ _masm->
74
75
76 //----------------------------------------------------------------------------------------------------
77
78 #ifndef _LP64
79 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
80 address entry = __ pc();
81 Argument argv(0, true);
82
83 // We are in the jni transition frame. Save the last_java_frame corresponding to the
84 // outer interpreter frame
85 //
86 __ set_last_Java_frame(FP, noreg);
87 // make sure the interpreter frame we've pushed has a valid return pc
88 __ mov(O7, I7);
89 __ mov(Lmethod, G3_scratch);
90 __ mov(Llocals, G4_scratch);
91 __ save_frame(0);
92 __ mov(G2_thread, L7_thread_cache);
93 __ add(argv.address_in_frame(), O3);
94 __ mov(G2_thread, O0);
95 __ mov(G3_scratch, O1);
96 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type);
97 __ delayed()->mov(G4_scratch, O2);
98 __ mov(L7_thread_cache, G2_thread);
99 __ reset_last_Java_frame();
100
101 // load the register arguments (the C code packed them as varargs)
102 for (Argument ldarg = argv.successor(); ldarg.is_register(); ldarg = ldarg.successor()) {
103 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register());
104 }
105 __ ret();
106 __ delayed()->
107 restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler
108 return entry;
109 }
110
111
112 #else
113 // LP64 passes floating point arguments in F1, F3, F5, etc. instead of
114 // O0, O1, O2 etc..
115 // Doubles are passed in D0, D2, D4
116 // We store the signature of the first 16 arguments in the first argument
117 // slot because it will be overwritten prior to calling the native
118 // function, with the pointer to the JNIEnv.
119 // If LP64 there can be up to 16 floating point arguments in registers
120 // or 6 integer registers.
121 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
122
123 enum {
124 non_float = 0,
125 float_sig = 1,
126 double_sig = 2,
127 sig_mask = 3
128 };
129
130 address entry = __ pc();
131 Argument argv(0, true);
132
133 // We are in the jni transition frame. Save the last_java_frame corresponding to the
134 // outer interpreter frame
135 //
136 __ set_last_Java_frame(FP, noreg);
137 // make sure the interpreter frame we've pushed has a valid return pc
138 __ mov(O7, I7);
139 __ mov(Lmethod, G3_scratch);
140 __ mov(Llocals, G4_scratch);
141 __ save_frame(0);
142 __ mov(G2_thread, L7_thread_cache);
143 __ add(argv.address_in_frame(), O3);
144 __ mov(G2_thread, O0);
145 __ mov(G3_scratch, O1);
146 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type);
147 __ delayed()->mov(G4_scratch, O2);
148 __ mov(L7_thread_cache, G2_thread);
149 __ reset_last_Java_frame();
150
151
152 // load the register arguments (the C code packed them as varargs)
153 Address Sig = argv.address_in_frame(); // Argument 0 holds the signature
154 __ ld_ptr( Sig, G3_scratch ); // Get register argument signature word into G3_scratch
155 __ mov( G3_scratch, G4_scratch);
156 __ srl( G4_scratch, 2, G4_scratch); // Skip Arg 0
157 Label done;
158 for (Argument ldarg = argv.successor(); ldarg.is_float_register(); ldarg = ldarg.successor()) {
159 Label NonFloatArg;
160 Label LoadFloatArg;
161 Label LoadDoubleArg;
162 Label NextArg;
163 Address a = ldarg.address_in_frame();
164 __ andcc(G4_scratch, sig_mask, G3_scratch);
165 __ br(Assembler::zero, false, Assembler::pt, NonFloatArg);
166 __ delayed()->nop();
167
168 __ cmp(G3_scratch, float_sig );
169 __ br(Assembler::equal, false, Assembler::pt, LoadFloatArg);
170 __ delayed()->nop();
171
172 __ cmp(G3_scratch, double_sig );
173 __ br(Assembler::equal, false, Assembler::pt, LoadDoubleArg);
174 __ delayed()->nop();
175
176 __ bind(NonFloatArg);
177 // There are only 6 integer register arguments!
178 if ( ldarg.is_register() )
179 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register());
180 else {
181 // Optimization, see if there are any more args and get out prior to checking
182 // all 16 float registers. My guess is that this is rare.
183 // If is_register is false, then we are done the first six integer args.
184 __ br_null_short(G4_scratch, Assembler::pt, done);
185 }
186 __ ba(NextArg);
187 __ delayed()->srl( G4_scratch, 2, G4_scratch );
188
189 __ bind(LoadFloatArg);
190 __ ldf( FloatRegisterImpl::S, a, ldarg.as_float_register(), 4);
191 __ ba(NextArg);
192 __ delayed()->srl( G4_scratch, 2, G4_scratch );
193
194 __ bind(LoadDoubleArg);
195 __ ldf( FloatRegisterImpl::D, a, ldarg.as_double_register() );
196 __ ba(NextArg);
197 __ delayed()->srl( G4_scratch, 2, G4_scratch );
198
199 __ bind(NextArg);
200
201 }
202
203 __ bind(done);
204 __ ret();
205 __ delayed()->
206 restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler
207 return entry;
208 }
209 #endif
210
211 void TemplateInterpreterGenerator::generate_counter_overflow(Label& Lcontinue) {
212
213 // Generate code to initiate compilation on the counter overflow.
214
215 // InterpreterRuntime::frequency_counter_overflow takes two arguments,
216 // the first indicates if the counter overflow occurs at a backwards branch (NULL bcp)
217 // and the second is only used when the first is true. We pass zero for both.
218 // The call returns the address of the verified entry point for the method or NULL
219 // if the compilation did not complete (either went background or bailed out).
220 __ set((int)false, O2);
221 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), O2, O2, true);
222 // returns verified_entry_point or NULL
223 // we ignore it in any case
224 __ ba_short(Lcontinue);
225
226 }
227
228
229 // End of helpers
230
231 // Various method entries
232
233 // Abstract method entry
234 // Attempt to execute abstract method. Throw exception
235 //
236 address TemplateInterpreterGenerator::generate_abstract_entry(void) {
237 address entry = __ pc();
238 // abstract method entry
239 // throw exception
240 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError));
241 // the call_VM checks for exception, so we should never return here.
242 __ should_not_reach_here();
243 return entry;
244
245 }
246
247 void TemplateInterpreterGenerator::save_native_result(void) {
248 // result potentially in O0/O1: save it across calls
249 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
250
251 // result potentially in F0/F1: save it across calls
252 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
253
254 // save and restore any potential method result value around the unlocking operation
255 __ stf(FloatRegisterImpl::D, F0, d_tmp);
256 #ifdef _LP64
257 __ stx(O0, l_tmp);
258 #else
259 __ std(O0, l_tmp);
260 #endif
261 }
262
263 void TemplateInterpreterGenerator::restore_native_result(void) {
264 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
265 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
266
267 // Restore any method result value
268 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
269 #ifdef _LP64
270 __ ldx(l_tmp, O0);
271 #else
272 __ ldd(l_tmp, O0);
273 #endif
274 }
275
276 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
277 assert(!pass_oop || message == NULL, "either oop or message but not both");
278 address entry = __ pc();
279 // expression stack must be empty before entering the VM if an exception happened
280 __ empty_expression_stack();
281 // load exception object
282 __ set((intptr_t)name, G3_scratch);
283 if (pass_oop) {
284 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
285 } else {
286 __ set((intptr_t)message, G4_scratch);
287 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
288 }
289 // throw exception
290 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
291 AddressLiteral thrower(Interpreter::throw_exception_entry());
292 __ jump_to(thrower, G3_scratch);
293 __ delayed()->nop();
294 return entry;
295 }
296
297 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
298 address entry = __ pc();
299 // expression stack must be empty before entering the VM if an exception
300 // happened
301 __ empty_expression_stack();
302 // load exception object
303 __ call_VM(Oexception,
304 CAST_FROM_FN_PTR(address,
305 InterpreterRuntime::throw_ClassCastException),
306 Otos_i);
307 __ should_not_reach_here();
308 return entry;
309 }
310
311
312 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
313 address entry = __ pc();
314 // expression stack must be empty before entering the VM if an exception happened
315 __ empty_expression_stack();
316 // convention: expect aberrant index in register G3_scratch, then shuffle the
317 // index to G4_scratch for the VM call
318 __ mov(G3_scratch, G4_scratch);
319 __ set((intptr_t)name, G3_scratch);
320 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
321 __ should_not_reach_here();
322 return entry;
323 }
324
325
326 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
327 address entry = __ pc();
328 // expression stack must be empty before entering the VM if an exception happened
329 __ empty_expression_stack();
330 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
331 __ should_not_reach_here();
332 return entry;
333 }
334
335
336 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
337 address entry = __ pc();
338
339 if (state == atos) {
340 __ profile_return_type(O0, G3_scratch, G1_scratch);
341 }
342
343 #if !defined(_LP64) && defined(COMPILER2)
344 // All return values are where we want them, except for Longs. C2 returns
345 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
346 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
347 // build even if we are returning from interpreted we just do a little
348 // stupid shuffing.
349 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
350 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
351 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
352
353 if (state == ltos) {
354 __ srl (G1, 0, O1);
355 __ srlx(G1, 32, O0);
356 }
357 #endif // !_LP64 && COMPILER2
358
359 // The callee returns with the stack possibly adjusted by adapter transition
360 // We remove that possible adjustment here.
361 // All interpreter local registers are untouched. Any result is passed back
362 // in the O0/O1 or float registers. Before continuing, the arguments must be
363 // popped from the java expression stack; i.e., Lesp must be adjusted.
364
365 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
366
367 const Register cache = G3_scratch;
368 const Register index = G1_scratch;
369 __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
370
371 const Register flags = cache;
372 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
373 const Register parameter_size = flags;
374 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words
375 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes
376 __ add(Lesp, parameter_size, Lesp); // pop arguments
377 __ dispatch_next(state, step);
378
379 return entry;
380 }
381
382
383 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
384 address entry = __ pc();
385 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
386 #if INCLUDE_JVMCI
387 // Check if we need to take lock at entry of synchronized method.
388 if (UseJVMCICompiler) {
389 Label L;
390 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
391 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter
392 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
393 // Clear flag.
394 __ stbool(G0, pending_monitor_enter_addr);
395 // Take lock.
396 lock_method();
397 __ bind(L);
398 }
399 #endif
400 { Label L;
401 Address exception_addr(G2_thread, Thread::pending_exception_offset());
402 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
403 __ br_null_short(Gtemp, Assembler::pt, L);
404 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
405 __ should_not_reach_here();
406 __ bind(L);
407 }
408 __ dispatch_next(state, step);
409 return entry;
410 }
411
412 // A result handler converts/unboxes a native call result into
413 // a java interpreter/compiler result. The current frame is an
414 // interpreter frame. The activation frame unwind code must be
415 // consistent with that of TemplateTable::_return(...). In the
416 // case of native methods, the caller's SP was not modified.
417 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
418 address entry = __ pc();
419 Register Itos_i = Otos_i ->after_save();
420 Register Itos_l = Otos_l ->after_save();
421 Register Itos_l1 = Otos_l1->after_save();
422 Register Itos_l2 = Otos_l2->after_save();
423 switch (type) {
424 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
425 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
426 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
427 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
428 case T_LONG :
429 #ifndef _LP64
430 __ mov(O1, Itos_l2); // move other half of long
431 #endif // ifdef or no ifdef, fall through to the T_INT case
432 case T_INT : __ mov(O0, Itos_i); break;
433 case T_VOID : /* nothing to do */ break;
434 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
435 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
436 case T_OBJECT :
437 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
438 __ verify_oop(Itos_i);
439 break;
440 default : ShouldNotReachHere();
441 }
442 __ ret(); // return from interpreter activation
443 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
444 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
445 return entry;
446 }
447
448 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
449 address entry = __ pc();
450 __ push(state);
451 __ call_VM(noreg, runtime_entry);
452 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
453 return entry;
454 }
455
456
457 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
458 address entry = __ pc();
459 __ dispatch_next(state);
460 return entry;
461 }
462
463 //
464 // Helpers for commoning out cases in the various type of method entries.
465 //
466
467 // increment invocation count & check for overflow
468 //
469 // Note: checking for negative value instead of overflow
470 // so we have a 'sticky' overflow test
471 //
472 // Lmethod: method
473 // ??: invocation counter
474 //
475 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
476 // Note: In tiered we increment either counters in MethodCounters* or in
477 // MDO depending if we're profiling or not.
478 const Register G3_method_counters = G3_scratch;
479 Label done;
480
481 if (TieredCompilation) {
482 const int increment = InvocationCounter::count_increment;
483 Label no_mdo;
484 if (ProfileInterpreter) {
485 // If no method data exists, go to profile_continue.
486 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
487 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
488 // Increment counter
489 Address mdo_invocation_counter(G4_scratch,
490 in_bytes(MethodData::invocation_counter_offset()) +
491 in_bytes(InvocationCounter::counter_offset()));
492 Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset()));
493 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
494 G3_scratch, Lscratch,
495 Assembler::zero, overflow);
496 __ ba_short(done);
497 }
498
499 // Increment counter in MethodCounters*
500 __ bind(no_mdo);
501 Address invocation_counter(G3_method_counters,
502 in_bytes(MethodCounters::invocation_counter_offset()) +
503 in_bytes(InvocationCounter::counter_offset()));
504 __ get_method_counters(Lmethod, G3_method_counters, done);
505 Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset()));
506 __ increment_mask_and_jump(invocation_counter, increment, mask,
507 G4_scratch, Lscratch,
508 Assembler::zero, overflow);
509 __ bind(done);
510 } else { // not TieredCompilation
511 // Update standard invocation counters
512 __ get_method_counters(Lmethod, G3_method_counters, done);
513 __ increment_invocation_counter(G3_method_counters, O0, G4_scratch);
514 if (ProfileInterpreter) {
515 Address interpreter_invocation_counter(G3_method_counters,
516 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
517 __ ld(interpreter_invocation_counter, G4_scratch);
518 __ inc(G4_scratch);
519 __ st(G4_scratch, interpreter_invocation_counter);
520 }
521
522 if (ProfileInterpreter && profile_method != NULL) {
523 // Test to see if we should create a method data oop
524 Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset()));
525 __ ld(profile_limit, G1_scratch);
526 __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
527
528 // if no method data exists, go to profile_method
529 __ test_method_data_pointer(*profile_method);
530 }
531
532 Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset()));
533 __ ld(invocation_limit, G3_scratch);
534 __ cmp(O0, G3_scratch);
535 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
536 __ delayed()->nop();
537 __ bind(done);
538 }
539
540 }
541
542 // Allocate monitor and lock method (asm interpreter)
543 // ebx - Method*
544 //
545 void TemplateInterpreterGenerator::lock_method() {
546 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
547
548 #ifdef ASSERT
549 { Label ok;
550 __ btst(JVM_ACC_SYNCHRONIZED, O0);
551 __ br( Assembler::notZero, false, Assembler::pt, ok);
552 __ delayed()->nop();
553 __ stop("method doesn't need synchronization");
554 __ bind(ok);
555 }
556 #endif // ASSERT
557
558 // get synchronization object to O0
559 { Label done;
560 __ btst(JVM_ACC_STATIC, O0);
561 __ br( Assembler::zero, true, Assembler::pt, done);
562 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
563
564 // lock the mirror, not the Klass*
565 __ load_mirror(O0, Lmethod);
566
567 #ifdef ASSERT
568 __ tst(O0);
569 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
570 #endif // ASSERT
571
572 __ bind(done);
573 }
574
575 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
576 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
577 // __ untested("lock_object from method entry");
578 __ lock_object(Lmonitors, O0);
579 }
580
581 // See if we've got enough room on the stack for locals plus overhead below
582 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
583 // without going through the signal handler, i.e., reserved and yellow zones
584 // will not be made usable. The shadow zone must suffice to handle the
585 // overflow.
586 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
587 Register Rscratch) {
588 const int page_size = os::vm_page_size();
589 Label after_frame_check;
590
591 assert_different_registers(Rframe_size, Rscratch);
592
593 __ set(page_size, Rscratch);
594 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
595
596 // Get the stack overflow limit, and in debug, verify it is non-zero.
597 __ ld_ptr(G2_thread, JavaThread::stack_overflow_limit_offset(), Rscratch);
598 #ifdef ASSERT
599 Label limit_ok;
600 __ br_notnull_short(Rscratch, Assembler::pn, limit_ok);
601 __ stop("stack overflow limit is zero in generate_stack_overflow_check");
602 __ bind(limit_ok);
603 #endif
604
605 // Add in the size of the frame (which is the same as subtracting it from the
606 // SP, which would take another register.
607 __ add(Rscratch, Rframe_size, Rscratch);
608
609 // The frame is greater than one page in size, so check against
610 // the bottom of the stack.
611 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
612
613 // The stack will overflow, throw an exception.
614
615 // Note that SP is restored to sender's sp (in the delay slot). This
616 // is necessary if the sender's frame is an extended compiled frame
617 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
618 // adaptations.
619
620 // Note also that the restored frame is not necessarily interpreted.
621 // Use the shared runtime version of the StackOverflowError.
622 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
623 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
624 __ jump_to(stub, Rscratch);
625 __ delayed()->mov(O5_savedSP, SP);
626
627 // If you get to here, then there is enough stack space.
628 __ bind(after_frame_check);
629 }
630
631
632 //
633 // Generate a fixed interpreter frame. This is identical setup for interpreted
634 // methods and for native methods hence the shared code.
635
636
637 //----------------------------------------------------------------------------------------------------
638 // Stack frame layout
639 //
640 // When control flow reaches any of the entry types for the interpreter
641 // the following holds ->
642 //
643 // C2 Calling Conventions:
644 //
645 // The entry code below assumes that the following registers are set
646 // when coming in:
647 // G5_method: holds the Method* of the method to call
648 // Lesp: points to the TOS of the callers expression stack
649 // after having pushed all the parameters
650 //
651 // The entry code does the following to setup an interpreter frame
652 // pop parameters from the callers stack by adjusting Lesp
653 // set O0 to Lesp
654 // compute X = (max_locals - num_parameters)
655 // bump SP up by X to accomadate the extra locals
656 // compute X = max_expression_stack
657 // + vm_local_words
658 // + 16 words of register save area
659 // save frame doing a save sp, -X, sp growing towards lower addresses
660 // set Lbcp, Lmethod, LcpoolCache
661 // set Llocals to i0
662 // set Lmonitors to FP - rounded_vm_local_words
663 // set Lesp to Lmonitors - 4
664 //
665 // The frame has now been setup to do the rest of the entry code
666
667 // Try this optimization: Most method entries could live in a
668 // "one size fits all" stack frame without all the dynamic size
669 // calculations. It might be profitable to do all this calculation
670 // statically and approximately for "small enough" methods.
671
672 //-----------------------------------------------------------------------------------------------
673
674 // C1 Calling conventions
675 //
676 // Upon method entry, the following registers are setup:
677 //
678 // g2 G2_thread: current thread
679 // g5 G5_method: method to activate
680 // g4 Gargs : pointer to last argument
681 //
682 //
683 // Stack:
684 //
685 // +---------------+ <--- sp
686 // | |
687 // : reg save area :
688 // | |
689 // +---------------+ <--- sp + 0x40
690 // | |
691 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
692 // | |
693 // +---------------+ <--- sp + 0x5c
694 // | |
695 // : free :
696 // | |
697 // +---------------+ <--- Gargs
698 // | |
699 // : arguments :
700 // | |
701 // +---------------+
702 // | |
703 //
704 //
705 //
706 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
707 //
708 // +---------------+ <--- sp
709 // | |
710 // : reg save area :
711 // | |
712 // +---------------+ <--- sp + 0x40
713 // | |
714 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
715 // | |
716 // +---------------+ <--- sp + 0x5c
717 // | |
718 // : :
719 // | | <--- Lesp
720 // +---------------+ <--- Lmonitors (fp - 0x18)
721 // | VM locals |
722 // +---------------+ <--- fp
723 // | |
724 // : reg save area :
725 // | |
726 // +---------------+ <--- fp + 0x40
727 // | |
728 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
729 // | |
730 // +---------------+ <--- fp + 0x5c
731 // | |
732 // : free :
733 // | |
734 // +---------------+
735 // | |
736 // : nonarg locals :
737 // | |
738 // +---------------+
739 // | |
740 // : arguments :
741 // | | <--- Llocals
742 // +---------------+ <--- Gargs
743 // | |
744
745 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
746 //
747 //
748 // The entry code sets up a new interpreter frame in 4 steps:
749 //
750 // 1) Increase caller's SP by for the extra local space needed:
751 // (check for overflow)
752 // Efficient implementation of xload/xstore bytecodes requires
753 // that arguments and non-argument locals are in a contigously
754 // addressable memory block => non-argument locals must be
755 // allocated in the caller's frame.
756 //
757 // 2) Create a new stack frame and register window:
758 // The new stack frame must provide space for the standard
759 // register save area, the maximum java expression stack size,
760 // the monitor slots (0 slots initially), and some frame local
761 // scratch locations.
762 //
763 // 3) The following interpreter activation registers must be setup:
764 // Lesp : expression stack pointer
765 // Lbcp : bytecode pointer
766 // Lmethod : method
767 // Llocals : locals pointer
768 // Lmonitors : monitor pointer
769 // LcpoolCache: constant pool cache
770 //
771 // 4) Initialize the non-argument locals if necessary:
772 // Non-argument locals may need to be initialized to NULL
773 // for GC to work. If the oop-map information is accurate
774 // (in the absence of the JSR problem), no initialization
775 // is necessary.
776 //
777 // (gri - 2/25/2000)
778
779
780 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
781
782 const int extra_space =
783 rounded_vm_local_words + // frame local scratch space
784 Method::extra_stack_entries() + // extra stack for jsr 292
785 frame::memory_parameter_word_sp_offset + // register save area
786 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
787
788 const Register Glocals_size = G3;
789 const Register RconstMethod = Glocals_size;
790 const Register Otmp1 = O3;
791 const Register Otmp2 = O4;
792 // Lscratch can't be used as a temporary because the call_stub uses
793 // it to assert that the stack frame was setup correctly.
794 const Address constMethod (G5_method, Method::const_offset());
795 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
796
797 __ ld_ptr( constMethod, RconstMethod );
798 __ lduh( size_of_parameters, Glocals_size);
799
800 // Gargs points to first local + BytesPerWord
801 // Set the saved SP after the register window save
802 //
803 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
804 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
805 __ add(Gargs, Otmp1, Gargs);
806
807 if (native_call) {
808 __ calc_mem_param_words( Glocals_size, Gframe_size );
809 __ add( Gframe_size, extra_space, Gframe_size);
810 __ round_to( Gframe_size, WordsPerLong );
811 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
812
813 // Native calls don't need the stack size check since they have no
814 // expression stack and the arguments are already on the stack and
815 // we only add a handful of words to the stack.
816 } else {
817
818 //
819 // Compute number of locals in method apart from incoming parameters
820 //
821 const Address size_of_locals(Otmp1, ConstMethod::size_of_locals_offset());
822 __ ld_ptr(constMethod, Otmp1);
823 __ lduh(size_of_locals, Otmp1);
824 __ sub(Otmp1, Glocals_size, Glocals_size);
825 __ round_to(Glocals_size, WordsPerLong);
826 __ sll(Glocals_size, Interpreter::logStackElementSize, Glocals_size);
827
828 // See if the frame is greater than one page in size. If so,
829 // then we need to verify there is enough stack space remaining.
830 // Frame_size = (max_stack + extra_space) * BytesPerWord;
831 __ ld_ptr(constMethod, Gframe_size);
832 __ lduh(Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size);
833 __ add(Gframe_size, extra_space, Gframe_size);
834 __ round_to(Gframe_size, WordsPerLong);
835 __ sll(Gframe_size, Interpreter::logStackElementSize, Gframe_size);
836
837 // Add in java locals size for stack overflow check only
838 __ add(Gframe_size, Glocals_size, Gframe_size);
839
840 const Register Otmp2 = O4;
841 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
842 generate_stack_overflow_check(Gframe_size, Otmp1);
843
844 __ sub(Gframe_size, Glocals_size, Gframe_size);
845
846 //
847 // bump SP to accomodate the extra locals
848 //
849 __ sub(SP, Glocals_size, SP);
850 }
851
852 //
853 // now set up a stack frame with the size computed above
854 //
855 __ neg( Gframe_size );
856 __ save( SP, Gframe_size, SP );
857
858 //
859 // now set up all the local cache registers
860 //
861 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
862 // that all present references to Lbyte_code initialize the register
863 // immediately before use
864 if (native_call) {
865 __ mov(G0, Lbcp);
866 } else {
867 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
868 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
869 }
870 __ mov( G5_method, Lmethod); // set Lmethod
871 // Get mirror and store it in the frame as GC root for this Method*
872 Register mirror = LcpoolCache;
873 __ load_mirror(mirror, Lmethod);
874 __ st_ptr(mirror, FP, (frame::interpreter_frame_mirror_offset * wordSize) + STACK_BIAS);
875 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
876 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
877 #ifdef _LP64
878 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
879 #endif
880 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
881
882 // setup interpreter activation registers
883 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
884
885 if (ProfileInterpreter) {
886 #ifdef FAST_DISPATCH
887 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
888 // they both use I2.
889 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
890 #endif // FAST_DISPATCH
891 __ set_method_data_pointer();
892 }
893
894 }
895
896 // Method entry for java.lang.ref.Reference.get.
897 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
898 #if INCLUDE_ALL_GCS
899 // Code: _aload_0, _getfield, _areturn
900 // parameter size = 1
901 //
902 // The code that gets generated by this routine is split into 2 parts:
903 // 1. The "intrinsified" code for G1 (or any SATB based GC),
904 // 2. The slow path - which is an expansion of the regular method entry.
905 //
906 // Notes:-
907 // * In the G1 code we do not check whether we need to block for
908 // a safepoint. If G1 is enabled then we must execute the specialized
909 // code for Reference.get (except when the Reference object is null)
910 // so that we can log the value in the referent field with an SATB
911 // update buffer.
912 // If the code for the getfield template is modified so that the
913 // G1 pre-barrier code is executed when the current method is
914 // Reference.get() then going through the normal method entry
915 // will be fine.
916 // * The G1 code can, however, check the receiver object (the instance
917 // of java.lang.Reference) and jump to the slow path if null. If the
918 // Reference object is null then we obviously cannot fetch the referent
919 // and so we don't need to call the G1 pre-barrier. Thus we can use the
920 // regular method entry code to generate the NPE.
921 //
922 // This code is based on generate_accessor_enty.
923
924 address entry = __ pc();
925
926 const int referent_offset = java_lang_ref_Reference::referent_offset;
927 guarantee(referent_offset > 0, "referent offset not initialized");
928
929 if (UseG1GC) {
930 Label slow_path;
931
932 // In the G1 code we don't check if we need to reach a safepoint. We
933 // continue and the thread will safepoint at the next bytecode dispatch.
934
935 // Check if local 0 != NULL
936 // If the receiver is null then it is OK to jump to the slow path.
937 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
938 // check if local 0 == NULL and go the slow path
939 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
940
941
942 // Load the value of the referent field.
943 if (Assembler::is_simm13(referent_offset)) {
944 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
945 } else {
946 __ set(referent_offset, G3_scratch);
947 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
948 }
949
950 // Generate the G1 pre-barrier code to log the value of
951 // the referent field in an SATB buffer. Note with
952 // these parameters the pre-barrier does not generate
953 // the load of the previous value
954
955 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
956 Otos_i /* pre_val */,
957 G3_scratch /* tmp */,
958 true /* preserve_o_regs */);
959
960 // _areturn
961 __ retl(); // return from leaf routine
962 __ delayed()->mov(O5_savedSP, SP);
963
964 // Generate regular method entry
965 __ bind(slow_path);
966 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
967 return entry;
968 }
969 #endif // INCLUDE_ALL_GCS
970
971 // If G1 is not enabled then attempt to go through the accessor entry point
972 // Reference.get is an accessor
973 return NULL;
974 }
975
976 /**
977 * Method entry for static native methods:
978 * int java.util.zip.CRC32.update(int crc, int b)
979 */
980 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
981
982 if (UseCRC32Intrinsics) {
983 address entry = __ pc();
984
985 Label L_slow_path;
986 // If we need a safepoint check, generate full interpreter entry.
987 ExternalAddress state(SafepointSynchronize::address_of_state());
988 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
989 __ set(SafepointSynchronize::_not_synchronized, O3);
990 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
991
992 // Load parameters
993 const Register crc = O0; // initial crc
994 const Register val = O1; // byte to update with
995 const Register table = O2; // address of 256-entry lookup table
996
997 __ ldub(Gargs, 3, val);
998 __ lduw(Gargs, 8, crc);
999
1000 __ set(ExternalAddress(StubRoutines::crc_table_addr()), table);
1001
1002 __ not1(crc); // ~crc
1003 __ clruwu(crc);
1004 __ update_byte_crc32(crc, val, table);
1005 __ not1(crc); // ~crc
1006
1007 // result in O0
1008 __ retl();
1009 __ delayed()->nop();
1010
1011 // generate a vanilla native entry as the slow path
1012 __ bind(L_slow_path);
1013 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
1014 return entry;
1015 }
1016 return NULL;
1017 }
1018
1019 /**
1020 * Method entry for static native methods:
1021 * int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len)
1022 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
1023 */
1024 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1025
1026 if (UseCRC32Intrinsics) {
1027 address entry = __ pc();
1028
1029 Label L_slow_path;
1030 // If we need a safepoint check, generate full interpreter entry.
1031 ExternalAddress state(SafepointSynchronize::address_of_state());
1032 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
1033 __ set(SafepointSynchronize::_not_synchronized, O3);
1034 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
1035
1036 // Load parameters from the stack
1037 const Register crc = O0; // initial crc
1038 const Register buf = O1; // source java byte array address
1039 const Register len = O2; // len
1040 const Register offset = O3; // offset
1041
1042 // Arguments are reversed on java expression stack
1043 // Calculate address of start element
1044 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) {
1045 __ lduw(Gargs, 0, len);
1046 __ lduw(Gargs, 8, offset);
1047 __ ldx( Gargs, 16, buf);
1048 __ lduw(Gargs, 32, crc);
1049 __ add(buf, offset, buf);
1050 } else {
1051 __ lduw(Gargs, 0, len);
1052 __ lduw(Gargs, 8, offset);
1053 __ ldx( Gargs, 16, buf);
1054 __ lduw(Gargs, 24, crc);
1055 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1056 __ add(buf ,offset, buf);
1057 }
1058
1059 // Call the crc32 kernel
1060 __ MacroAssembler::save_thread(L7_thread_cache);
1061 __ kernel_crc32(crc, buf, len, O3);
1062 __ MacroAssembler::restore_thread(L7_thread_cache);
1063
1064 // result in O0
1065 __ retl();
1066 __ delayed()->nop();
1067
1068 // generate a vanilla native entry as the slow path
1069 __ bind(L_slow_path);
1070 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
1071 return entry;
1072 }
1073 return NULL;
1074 }
1075
1076 /**
1077 * Method entry for intrinsic-candidate (non-native) methods:
1078 * int java.util.zip.CRC32C.updateBytes(int crc, byte[] b, int off, int end)
1079 * int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
1080 * Unlike CRC32, CRC32C does not have any methods marked as native
1081 * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
1082 */
1083 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1084
1085 if (UseCRC32CIntrinsics) {
1086 address entry = __ pc();
1087
1088 // Load parameters from the stack
1089 const Register crc = O0; // initial crc
1090 const Register buf = O1; // source java byte array address
1091 const Register offset = O2; // offset
1092 const Register end = O3; // index of last element to process
1093 const Register len = O2; // len argument to the kernel
1094 const Register table = O3; // crc32c lookup table address
1095
1096 // Arguments are reversed on java expression stack
1097 // Calculate address of start element
1098 if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) {
1099 __ lduw(Gargs, 0, end);
1100 __ lduw(Gargs, 8, offset);
1101 __ ldx( Gargs, 16, buf);
1102 __ lduw(Gargs, 32, crc);
1103 __ add(buf, offset, buf);
1104 __ sub(end, offset, len);
1105 } else {
1106 __ lduw(Gargs, 0, end);
1107 __ lduw(Gargs, 8, offset);
1108 __ ldx( Gargs, 16, buf);
1109 __ lduw(Gargs, 24, crc);
1110 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1111 __ add(buf, offset, buf);
1112 __ sub(end, offset, len);
1113 }
1114
1115 // Call the crc32c kernel
1116 __ MacroAssembler::save_thread(L7_thread_cache);
1117 __ kernel_crc32c(crc, buf, len, table);
1118 __ MacroAssembler::restore_thread(L7_thread_cache);
1119
1120 // result in O0
1121 __ retl();
1122 __ delayed()->nop();
1123
1124 return entry;
1125 }
1126 return NULL;
1127 }
1128
1129 // Not supported
1130 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
1131 return NULL;
1132 }
1133
1134 // TODO: rather than touching all pages, check against stack_overflow_limit and bang yellow page to
1135 // generate exception
1136 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1137 // Quick & dirty stack overflow checking: bang the stack & handle trap.
1138 // Note that we do the banging after the frame is setup, since the exception
1139 // handling code expects to find a valid interpreter frame on the stack.
1140 // Doing the banging earlier fails if the caller frame is not an interpreter
1141 // frame.
1142 // (Also, the exception throwing code expects to unlock any synchronized
1143 // method receiever, so do the banging after locking the receiver.)
1144
1145 // Bang each page in the shadow zone. We can't assume it's been done for
1146 // an interpreter frame with greater than a page of locals, so each page
1147 // needs to be checked. Only true for non-native.
1148 if (UseStackBanging) {
1149 const int page_size = os::vm_page_size();
1150 const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1151 const int start_page = native_call ? n_shadow_pages : 1;
1152 for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1153 __ bang_stack_with_offset(pages*page_size);
1154 }
1155 }
1156 }
1157
1158 //
1159 // Interpreter stub for calling a native method. (asm interpreter)
1160 // This sets up a somewhat different looking stack for calling the native method
1161 // than the typical interpreter frame setup.
1162 //
1163
1164 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1165 address entry = __ pc();
1166
1167 // the following temporary registers are used during frame creation
1168 const Register Gtmp1 = G3_scratch ;
1169 const Register Gtmp2 = G1_scratch;
1170 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1171
1172 // make sure registers are different!
1173 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1174
1175 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
1176
1177 const Register Glocals_size = G3;
1178 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1179
1180 // make sure method is native & not abstract
1181 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1182 #ifdef ASSERT
1183 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1184 {
1185 Label L;
1186 __ btst(JVM_ACC_NATIVE, Gtmp1);
1187 __ br(Assembler::notZero, false, Assembler::pt, L);
1188 __ delayed()->nop();
1189 __ stop("tried to execute non-native method as native");
1190 __ bind(L);
1191 }
1192 { Label L;
1193 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1194 __ br(Assembler::zero, false, Assembler::pt, L);
1195 __ delayed()->nop();
1196 __ stop("tried to execute abstract method as non-abstract");
1197 __ bind(L);
1198 }
1199 #endif // ASSERT
1200
1201 // generate the code to allocate the interpreter stack frame
1202 generate_fixed_frame(true);
1203
1204 //
1205 // No locals to initialize for native method
1206 //
1207
1208 // this slot will be set later, we initialize it to null here just in
1209 // case we get a GC before the actual value is stored later
1210 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
1211
1212 const Address do_not_unlock_if_synchronized(G2_thread,
1213 JavaThread::do_not_unlock_if_synchronized_offset());
1214 // Since at this point in the method invocation the exception handler
1215 // would try to exit the monitor of synchronized methods which hasn't
1216 // been entered yet, we set the thread local variable
1217 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1218 // runtime, exception handling i.e. unlock_if_synchronized_method will
1219 // check this thread local flag.
1220 // This flag has two effects, one is to force an unwind in the topmost
1221 // interpreter frame and not perform an unlock while doing so.
1222
1223 __ movbool(true, G3_scratch);
1224 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1225
1226 // increment invocation counter and check for overflow
1227 //
1228 // Note: checking for negative value instead of overflow
1229 // so we have a 'sticky' overflow test (may be of
1230 // importance as soon as we have true MT/MP)
1231 Label invocation_counter_overflow;
1232 Label Lcontinue;
1233 if (inc_counter) {
1234 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1235
1236 }
1237 __ bind(Lcontinue);
1238
1239 bang_stack_shadow_pages(true);
1240
1241 // reset the _do_not_unlock_if_synchronized flag
1242 __ stbool(G0, do_not_unlock_if_synchronized);
1243
1244 // check for synchronized methods
1245 // Must happen AFTER invocation_counter check and stack overflow check,
1246 // so method is not locked if overflows.
1247
1248 if (synchronized) {
1249 lock_method();
1250 } else {
1251 #ifdef ASSERT
1252 { Label ok;
1253 __ ld(Laccess_flags, O0);
1254 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1255 __ br( Assembler::zero, false, Assembler::pt, ok);
1256 __ delayed()->nop();
1257 __ stop("method needs synchronization");
1258 __ bind(ok);
1259 }
1260 #endif // ASSERT
1261 }
1262
1263
1264 // start execution
1265 __ verify_thread();
1266
1267 // JVMTI support
1268 __ notify_method_entry();
1269
1270 // native call
1271
1272 // (note that O0 is never an oop--at most it is a handle)
1273 // It is important not to smash any handles created by this call,
1274 // until any oop handle in O0 is dereferenced.
1275
1276 // (note that the space for outgoing params is preallocated)
1277
1278 // get signature handler
1279 { Label L;
1280 Address signature_handler(Lmethod, Method::signature_handler_offset());
1281 __ ld_ptr(signature_handler, G3_scratch);
1282 __ br_notnull_short(G3_scratch, Assembler::pt, L);
1283 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
1284 __ ld_ptr(signature_handler, G3_scratch);
1285 __ bind(L);
1286 }
1287
1288 // Push a new frame so that the args will really be stored in
1289 // Copy a few locals across so the new frame has the variables
1290 // we need but these values will be dead at the jni call and
1291 // therefore not gc volatile like the values in the current
1292 // frame (Lmethod in particular)
1293
1294 // Flush the method pointer to the register save area
1295 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
1296 __ mov(Llocals, O1);
1297
1298 // calculate where the mirror handle body is allocated in the interpreter frame:
1299 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
1300
1301 // Calculate current frame size
1302 __ sub(SP, FP, O3); // Calculate negative of current frame size
1303 __ save(SP, O3, SP); // Allocate an identical sized frame
1304
1305 // Note I7 has leftover trash. Slow signature handler will fill it in
1306 // should we get there. Normal jni call will set reasonable last_Java_pc
1307 // below (and fix I7 so the stack trace doesn't have a meaningless frame
1308 // in it).
1309
1310 // Load interpreter frame's Lmethod into same register here
1311
1312 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1313
1314 __ mov(I1, Llocals);
1315 __ mov(I2, Lscratch2); // save the address of the mirror
1316
1317
1318 // ONLY Lmethod and Llocals are valid here!
1319
1320 // call signature handler, It will move the arg properly since Llocals in current frame
1321 // matches that in outer frame
1322
1323 __ callr(G3_scratch, 0);
1324 __ delayed()->nop();
1325
1326 // Result handler is in Lscratch
1327
1328 // Reload interpreter frame's Lmethod since slow signature handler may block
1329 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1330
1331 { Label not_static;
1332
1333 __ ld(Laccess_flags, O0);
1334 __ btst(JVM_ACC_STATIC, O0);
1335 __ br( Assembler::zero, false, Assembler::pt, not_static);
1336 // get native function entry point(O0 is a good temp until the very end)
1337 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
1338 // for static methods insert the mirror argument
1339 __ load_mirror(O1, Lmethod);
1340 #ifdef ASSERT
1341 if (!PrintSignatureHandlers) // do not dirty the output with this
1342 { Label L;
1343 __ br_notnull_short(O1, Assembler::pt, L);
1344 __ stop("mirror is missing");
1345 __ bind(L);
1346 }
1347 #endif // ASSERT
1348 __ st_ptr(O1, Lscratch2, 0);
1349 __ mov(Lscratch2, O1);
1350 __ bind(not_static);
1351 }
1352
1353 // At this point, arguments have been copied off of stack into
1354 // their JNI positions, which are O1..O5 and SP[68..].
1355 // Oops are boxed in-place on the stack, with handles copied to arguments.
1356 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1357
1358 #ifdef ASSERT
1359 { Label L;
1360 __ br_notnull_short(O0, Assembler::pt, L);
1361 __ stop("native entry point is missing");
1362 __ bind(L);
1363 }
1364 #endif // ASSERT
1365
1366 //
1367 // setup the frame anchor
1368 //
1369 // The scavenge function only needs to know that the PC of this frame is
1370 // in the interpreter method entry code, it doesn't need to know the exact
1371 // PC and hence we can use O7 which points to the return address from the
1372 // previous call in the code stream (signature handler function)
1373 //
1374 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1375 // we have pushed the extra frame in order to protect the volatile register(s)
1376 // in that frame when we return from the jni call
1377 //
1378
1379 __ set_last_Java_frame(FP, O7);
1380 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1381 // not meaningless information that'll confuse me.
1382
1383 // flush the windows now. We don't care about the current (protection) frame
1384 // only the outer frames
1385
1386 __ flushw();
1387
1388 // mark windows as flushed
1389 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1390 __ set(JavaFrameAnchor::flushed, G3_scratch);
1391 __ st(G3_scratch, flags);
1392
1393 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1394
1395 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1396 #ifdef ASSERT
1397 { Label L;
1398 __ ld(thread_state, G3_scratch);
1399 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1400 __ stop("Wrong thread state in native stub");
1401 __ bind(L);
1402 }
1403 #endif // ASSERT
1404 __ set(_thread_in_native, G3_scratch);
1405 __ st(G3_scratch, thread_state);
1406
1407 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1408 __ save_thread(L7_thread_cache); // save Gthread
1409 __ callr(O0, 0);
1410 __ delayed()->
1411 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1412
1413 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1414
1415 __ restore_thread(L7_thread_cache); // restore G2_thread
1416 __ reinit_heapbase();
1417
1418 // must we block?
1419
1420 // Block, if necessary, before resuming in _thread_in_Java state.
1421 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1422 { Label no_block;
1423 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1424
1425 // Switch thread to "native transition" state before reading the synchronization state.
1426 // This additional state is necessary because reading and testing the synchronization
1427 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1428 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1429 // VM thread changes sync state to synchronizing and suspends threads for GC.
1430 // Thread A is resumed to finish this native method, but doesn't block here since it
1431 // didn't see any synchronization is progress, and escapes.
1432 __ set(_thread_in_native_trans, G3_scratch);
1433 __ st(G3_scratch, thread_state);
1434 if(os::is_MP()) {
1435 if (UseMembar) {
1436 // Force this write out before the read below
1437 __ membar(Assembler::StoreLoad);
1438 } else {
1439 // Write serialization page so VM thread can do a pseudo remote membar.
1440 // We use the current thread pointer to calculate a thread specific
1441 // offset to write to within the page. This minimizes bus traffic
1442 // due to cache line collision.
1443 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1444 }
1445 }
1446 __ load_contents(sync_state, G3_scratch);
1447 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1448
1449 Label L;
1450 __ br(Assembler::notEqual, false, Assembler::pn, L);
1451 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1452 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1453 __ bind(L);
1454
1455 // Block. Save any potential method result value before the operation and
1456 // use a leaf call to leave the last_Java_frame setup undisturbed.
1457 save_native_result();
1458 __ call_VM_leaf(L7_thread_cache,
1459 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1460 G2_thread);
1461
1462 // Restore any method result value
1463 restore_native_result();
1464 __ bind(no_block);
1465 }
1466
1467 // Clear the frame anchor now
1468
1469 __ reset_last_Java_frame();
1470
1471 // Move the result handler address
1472 __ mov(Lscratch, G3_scratch);
1473 // return possible result to the outer frame
1474 #ifndef __LP64
1475 __ mov(O0, I0);
1476 __ restore(O1, G0, O1);
1477 #else
1478 __ restore(O0, G0, O0);
1479 #endif /* __LP64 */
1480
1481 // Move result handler to expected register
1482 __ mov(G3_scratch, Lscratch);
1483
1484 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1485 // switch to thread_in_Java.
1486
1487 __ set(_thread_in_Java, G3_scratch);
1488 __ st(G3_scratch, thread_state);
1489
1490 // reset handle block
1491 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1492 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1493
1494 // If we have an oop result store it where it will be safe for any further gc
1495 // until we return now that we've released the handle it might be protected by
1496
1497 {
1498 Label no_oop, store_result;
1499
1500 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1501 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1502 __ addcc(G0, O0, O0);
1503 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1504 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1505 __ mov(G0, O0);
1506
1507 __ bind(store_result);
1508 // Store it where gc will look for it and result handler expects it.
1509 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1510
1511 __ bind(no_oop);
1512
1513 }
1514
1515
1516 // handle exceptions (exception handling will handle unlocking!)
1517 { Label L;
1518 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1519 __ ld_ptr(exception_addr, Gtemp);
1520 __ br_null_short(Gtemp, Assembler::pt, L);
1521 // Note: This could be handled more efficiently since we know that the native
1522 // method doesn't have an exception handler. We could directly return
1523 // to the exception handler for the caller.
1524 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1525 __ should_not_reach_here();
1526 __ bind(L);
1527 }
1528
1529 // JVMTI support (preserves thread register)
1530 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1531
1532 if (synchronized) {
1533 // save and restore any potential method result value around the unlocking operation
1534 save_native_result();
1535
1536 __ add( __ top_most_monitor(), O1);
1537 __ unlock_object(O1);
1538
1539 restore_native_result();
1540 }
1541
1542 #if defined(COMPILER2) && !defined(_LP64)
1543
1544 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1545 // or compiled so just be safe.
1546
1547 __ sllx(O0, 32, G1); // Shift bits into high G1
1548 __ srl (O1, 0, O1); // Zero extend O1
1549 __ or3 (O1, G1, G1); // OR 64 bits into G1
1550
1551 #endif /* COMPILER2 && !_LP64 */
1552
1553 // dispose of return address and remove activation
1554 #ifdef ASSERT
1555 {
1556 Label ok;
1557 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1558 __ stop("bad I5_savedSP value");
1559 __ should_not_reach_here();
1560 __ bind(ok);
1561 }
1562 #endif
1563 __ jmp(Lscratch, 0);
1564 __ delayed()->nop();
1565
1566
1567 if (inc_counter) {
1568 // handle invocation counter overflow
1569 __ bind(invocation_counter_overflow);
1570 generate_counter_overflow(Lcontinue);
1571 }
1572
1573
1574
1575 return entry;
1576 }
1577
1578
1579 // Generic method entry to (asm) interpreter
1580 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1581 address entry = __ pc();
1582
1583 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1584
1585 // the following temporary registers are used during frame creation
1586 const Register Gtmp1 = G3_scratch ;
1587 const Register Gtmp2 = G1_scratch;
1588
1589 // make sure registers are different!
1590 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1591
1592 const Address constMethod (G5_method, Method::const_offset());
1593 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1594 // and use in the asserts.
1595 const Address access_flags (Lmethod, Method::access_flags_offset());
1596
1597 const Register Glocals_size = G3;
1598 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1599
1600 // make sure method is not native & not abstract
1601 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1602 #ifdef ASSERT
1603 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1604 {
1605 Label L;
1606 __ btst(JVM_ACC_NATIVE, Gtmp1);
1607 __ br(Assembler::zero, false, Assembler::pt, L);
1608 __ delayed()->nop();
1609 __ stop("tried to execute native method as non-native");
1610 __ bind(L);
1611 }
1612 { Label L;
1613 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1614 __ br(Assembler::zero, false, Assembler::pt, L);
1615 __ delayed()->nop();
1616 __ stop("tried to execute abstract method as non-abstract");
1617 __ bind(L);
1618 }
1619 #endif // ASSERT
1620
1621 // generate the code to allocate the interpreter stack frame
1622
1623 generate_fixed_frame(false);
1624
1625 #ifdef FAST_DISPATCH
1626 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1627 // set bytecode dispatch table base
1628 #endif
1629
1630 //
1631 // Code to initialize the extra (i.e. non-parm) locals
1632 //
1633 Register init_value = noreg; // will be G0 if we must clear locals
1634 // The way the code was setup before zerolocals was always true for vanilla java entries.
1635 // It could only be false for the specialized entries like accessor or empty which have
1636 // no extra locals so the testing was a waste of time and the extra locals were always
1637 // initialized. We removed this extra complication to already over complicated code.
1638
1639 init_value = G0;
1640 Label clear_loop;
1641
1642 const Register RconstMethod = O1;
1643 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1644 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1645
1646 // NOTE: If you change the frame layout, this code will need to
1647 // be updated!
1648 __ ld_ptr( constMethod, RconstMethod );
1649 __ lduh( size_of_locals, O2 );
1650 __ lduh( size_of_parameters, O1 );
1651 __ sll( O2, Interpreter::logStackElementSize, O2);
1652 __ sll( O1, Interpreter::logStackElementSize, O1 );
1653 __ sub( Llocals, O2, O2 );
1654 __ sub( Llocals, O1, O1 );
1655
1656 __ bind( clear_loop );
1657 __ inc( O2, wordSize );
1658
1659 __ cmp( O2, O1 );
1660 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1661 __ delayed()->st_ptr( init_value, O2, 0 );
1662
1663 const Address do_not_unlock_if_synchronized(G2_thread,
1664 JavaThread::do_not_unlock_if_synchronized_offset());
1665 // Since at this point in the method invocation the exception handler
1666 // would try to exit the monitor of synchronized methods which hasn't
1667 // been entered yet, we set the thread local variable
1668 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1669 // runtime, exception handling i.e. unlock_if_synchronized_method will
1670 // check this thread local flag.
1671 __ movbool(true, G3_scratch);
1672 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1673
1674 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1675 // increment invocation counter and check for overflow
1676 //
1677 // Note: checking for negative value instead of overflow
1678 // so we have a 'sticky' overflow test (may be of
1679 // importance as soon as we have true MT/MP)
1680 Label invocation_counter_overflow;
1681 Label profile_method;
1682 Label profile_method_continue;
1683 Label Lcontinue;
1684 if (inc_counter) {
1685 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1686 if (ProfileInterpreter) {
1687 __ bind(profile_method_continue);
1688 }
1689 }
1690 __ bind(Lcontinue);
1691
1692 bang_stack_shadow_pages(false);
1693
1694 // reset the _do_not_unlock_if_synchronized flag
1695 __ stbool(G0, do_not_unlock_if_synchronized);
1696
1697 // check for synchronized methods
1698 // Must happen AFTER invocation_counter check and stack overflow check,
1699 // so method is not locked if overflows.
1700
1701 if (synchronized) {
1702 lock_method();
1703 } else {
1704 #ifdef ASSERT
1705 { Label ok;
1706 __ ld(access_flags, O0);
1707 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1708 __ br( Assembler::zero, false, Assembler::pt, ok);
1709 __ delayed()->nop();
1710 __ stop("method needs synchronization");
1711 __ bind(ok);
1712 }
1713 #endif // ASSERT
1714 }
1715
1716 // start execution
1717
1718 __ verify_thread();
1719
1720 // jvmti support
1721 __ notify_method_entry();
1722
1723 // start executing instructions
1724 __ dispatch_next(vtos);
1725
1726
1727 if (inc_counter) {
1728 if (ProfileInterpreter) {
1729 // We have decided to profile this method in the interpreter
1730 __ bind(profile_method);
1731
1732 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1733 __ set_method_data_pointer_for_bcp();
1734 __ ba_short(profile_method_continue);
1735 }
1736
1737 // handle invocation counter overflow
1738 __ bind(invocation_counter_overflow);
1739 generate_counter_overflow(Lcontinue);
1740 }
1741
1742
1743 return entry;
1744 }
1745
1746 //----------------------------------------------------------------------------------------------------
1747 // Exceptions
1748 void TemplateInterpreterGenerator::generate_throw_exception() {
1749
1750 // Entry point in previous activation (i.e., if the caller was interpreted)
1751 Interpreter::_rethrow_exception_entry = __ pc();
1752 // O0: exception
1753
1754 // entry point for exceptions thrown within interpreter code
1755 Interpreter::_throw_exception_entry = __ pc();
1756 __ verify_thread();
1757 // expression stack is undefined here
1758 // O0: exception, i.e. Oexception
1759 // Lbcp: exception bcp
1760 __ verify_oop(Oexception);
1761
1762
1763 // expression stack must be empty before entering the VM in case of an exception
1764 __ empty_expression_stack();
1765 // find exception handler address and preserve exception oop
1766 // call C routine to find handler and jump to it
1767 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1768 __ push_ptr(O1); // push exception for exception handler bytecodes
1769
1770 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1771 __ delayed()->nop();
1772
1773
1774 // if the exception is not handled in the current frame
1775 // the frame is removed and the exception is rethrown
1776 // (i.e. exception continuation is _rethrow_exception)
1777 //
1778 // Note: At this point the bci is still the bxi for the instruction which caused
1779 // the exception and the expression stack is empty. Thus, for any VM calls
1780 // at this point, GC will find a legal oop map (with empty expression stack).
1781
1782 // in current activation
1783 // tos: exception
1784 // Lbcp: exception bcp
1785
1786 //
1787 // JVMTI PopFrame support
1788 //
1789
1790 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1791 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1792 // Set the popframe_processing bit in popframe_condition indicating that we are
1793 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1794 // popframe handling cycles.
1795
1796 __ ld(popframe_condition_addr, G3_scratch);
1797 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1798 __ stw(G3_scratch, popframe_condition_addr);
1799
1800 // Empty the expression stack, as in normal exception handling
1801 __ empty_expression_stack();
1802 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1803
1804 {
1805 // Check to see whether we are returning to a deoptimized frame.
1806 // (The PopFrame call ensures that the caller of the popped frame is
1807 // either interpreted or compiled and deoptimizes it if compiled.)
1808 // In this case, we can't call dispatch_next() after the frame is
1809 // popped, but instead must save the incoming arguments and restore
1810 // them after deoptimization has occurred.
1811 //
1812 // Note that we don't compare the return PC against the
1813 // deoptimization blob's unpack entry because of the presence of
1814 // adapter frames in C2.
1815 Label caller_not_deoptimized;
1816 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1817 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1818
1819 const Register Gtmp1 = G3_scratch;
1820 const Register Gtmp2 = G1_scratch;
1821 const Register RconstMethod = Gtmp1;
1822 const Address constMethod(Lmethod, Method::const_offset());
1823 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1824
1825 // Compute size of arguments for saving when returning to deoptimized caller
1826 __ ld_ptr(constMethod, RconstMethod);
1827 __ lduh(size_of_parameters, Gtmp1);
1828 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1829 __ sub(Llocals, Gtmp1, Gtmp2);
1830 __ add(Gtmp2, wordSize, Gtmp2);
1831 // Save these arguments
1832 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1833 // Inform deoptimization that it is responsible for restoring these arguments
1834 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1835 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1836 __ st(Gtmp1, popframe_condition_addr);
1837
1838 // Return from the current method
1839 // The caller's SP was adjusted upon method entry to accomodate
1840 // the callee's non-argument locals. Undo that adjustment.
1841 __ ret();
1842 __ delayed()->restore(I5_savedSP, G0, SP);
1843
1844 __ bind(caller_not_deoptimized);
1845 }
1846
1847 // Clear the popframe condition flag
1848 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1849
1850 // Get out of the current method (how this is done depends on the particular compiler calling
1851 // convention that the interpreter currently follows)
1852 // The caller's SP was adjusted upon method entry to accomodate
1853 // the callee's non-argument locals. Undo that adjustment.
1854 __ restore(I5_savedSP, G0, SP);
1855 // The method data pointer was incremented already during
1856 // call profiling. We have to restore the mdp for the current bcp.
1857 if (ProfileInterpreter) {
1858 __ set_method_data_pointer_for_bcp();
1859 }
1860
1861 #if INCLUDE_JVMTI
1862 {
1863 Label L_done;
1864
1865 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1866 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1867
1868 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1869 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1870
1871 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1872
1873 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1874 __ delayed()->nop();
1875
1876 __ st_ptr(G1_scratch, Lesp, wordSize);
1877 __ bind(L_done);
1878 }
1879 #endif // INCLUDE_JVMTI
1880
1881 // Resume bytecode interpretation at the current bcp
1882 __ dispatch_next(vtos);
1883 // end of JVMTI PopFrame support
1884
1885 Interpreter::_remove_activation_entry = __ pc();
1886
1887 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1888 __ pop_ptr(Oexception); // get exception
1889
1890 // Intel has the following comment:
1891 //// remove the activation (without doing throws on illegalMonitorExceptions)
1892 // They remove the activation without checking for bad monitor state.
1893 // %%% We should make sure this is the right semantics before implementing.
1894
1895 __ set_vm_result(Oexception);
1896 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1897
1898 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1899
1900 __ get_vm_result(Oexception);
1901 __ verify_oop(Oexception);
1902
1903 const int return_reg_adjustment = frame::pc_return_offset;
1904 Address issuing_pc_addr(I7, return_reg_adjustment);
1905
1906 // We are done with this activation frame; find out where to go next.
1907 // The continuation point will be an exception handler, which expects
1908 // the following registers set up:
1909 //
1910 // Oexception: exception
1911 // Oissuing_pc: the local call that threw exception
1912 // Other On: garbage
1913 // In/Ln: the contents of the caller's register window
1914 //
1915 // We do the required restore at the last possible moment, because we
1916 // need to preserve some state across a runtime call.
1917 // (Remember that the caller activation is unknown--it might not be
1918 // interpreted, so things like Lscratch are useless in the caller.)
1919
1920 // Although the Intel version uses call_C, we can use the more
1921 // compact call_VM. (The only real difference on SPARC is a
1922 // harmlessly ignored [re]set_last_Java_frame, compared with
1923 // the Intel code which lacks this.)
1924 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1925 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1926 __ super_call_VM_leaf(L7_thread_cache,
1927 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1928 G2_thread, Oissuing_pc->after_save());
1929
1930 // The caller's SP was adjusted upon method entry to accomodate
1931 // the callee's non-argument locals. Undo that adjustment.
1932 __ JMP(O0, 0); // return exception handler in caller
1933 __ delayed()->restore(I5_savedSP, G0, SP);
1934
1935 // (same old exception object is already in Oexception; see above)
1936 // Note that an "issuing PC" is actually the next PC after the call
1937 }
1938
1939
1940 //
1941 // JVMTI ForceEarlyReturn support
1942 //
1943
1944 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1945 address entry = __ pc();
1946
1947 __ empty_expression_stack();
1948 __ load_earlyret_value(state);
1949
1950 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1951 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1952
1953 // Clear the earlyret state
1954 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1955
1956 __ remove_activation(state,
1957 /* throw_monitor_exception */ false,
1958 /* install_monitor_exception */ false);
1959
1960 // The caller's SP was adjusted upon method entry to accomodate
1961 // the callee's non-argument locals. Undo that adjustment.
1962 __ ret(); // return to caller
1963 __ delayed()->restore(I5_savedSP, G0, SP);
1964
1965 return entry;
1966 } // end of JVMTI ForceEarlyReturn support
1967
1968
1969 //------------------------------------------------------------------------------------------------------------------------
1970 // Helper for vtos entry point generation
1971
1972 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) {
1973 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1974 Label L;
1975 aep = __ pc(); __ push_ptr(); __ ba_short(L);
1976 fep = __ pc(); __ push_f(); __ ba_short(L);
1977 dep = __ pc(); __ push_d(); __ ba_short(L);
1978 lep = __ pc(); __ push_l(); __ ba_short(L);
1979 iep = __ pc(); __ push_i();
1980 bep = cep = sep = iep; // there aren't any
1981 vep = __ pc(); __ bind(L); // fall through
1982 generate_and_dispatch(t);
1983 }
1984
1985 // --------------------------------------------------------------------------------
1986
1987 // Non-product code
1988 #ifndef PRODUCT
1989 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1990 address entry = __ pc();
1991
1992 __ push(state);
1993 __ mov(O7, Lscratch); // protect return address within interpreter
1994
1995 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1996 __ mov( Otos_l2, G3_scratch );
1997 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1998 __ mov(Lscratch, O7); // restore return address
1999 __ pop(state);
2000 __ retl();
2001 __ delayed()->nop();
2002
2003 return entry;
2004 }
2005
2006
2007 // helpers for generate_and_dispatch
2008
2009 void TemplateInterpreterGenerator::count_bytecode() {
2010 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2011 }
2012
2013
2014 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2015 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2016 }
2017
2018
2019 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2020 AddressLiteral index (&BytecodePairHistogram::_index);
2021 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2022
2023 // get index, shift out old bytecode, bring in new bytecode, and store it
2024 // _index = (_index >> log2_number_of_codes) |
2025 // (bytecode << log2_number_of_codes);
2026
2027 __ load_contents(index, G4_scratch);
2028 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2029 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
2030 __ or3( G3_scratch, G4_scratch, G4_scratch );
2031 __ store_contents(G4_scratch, index, G3_scratch);
2032
2033 // bump bucket contents
2034 // _counters[_index] ++;
2035
2036 __ set(counters, G3_scratch); // loads into G3_scratch
2037 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
2038 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
2039 __ ld (G3_scratch, 0, G4_scratch);
2040 __ inc (G4_scratch);
2041 __ st (G4_scratch, 0, G3_scratch);
2042 }
2043
2044
2045 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2046 // Call a little run-time stub to avoid blow-up for each bytecode.
2047 // The run-time runtime saves the right registers, depending on
2048 // the tosca in-state for the given template.
2049 address entry = Interpreter::trace_code(t->tos_in());
2050 guarantee(entry != NULL, "entry must have been generated");
2051 __ call(entry, relocInfo::none);
2052 __ delayed()->nop();
2053 }
2054
2055
2056 void TemplateInterpreterGenerator::stop_interpreter_at() {
2057 AddressLiteral counter(&BytecodeCounter::_counter_value);
2058 __ load_contents(counter, G3_scratch);
2059 AddressLiteral stop_at(&StopInterpreterAt);
2060 __ load_ptr_contents(stop_at, G4_scratch);
2061 __ cmp(G3_scratch, G4_scratch);
2062 __ breakpoint_trap(Assembler::equal, Assembler::icc);
2063 }
2064 #endif // not PRODUCT