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 Register unused) {
589 const int page_size = os::vm_page_size();
590 Label after_frame_check;
591
592 assert_different_registers(Rframe_size, Rscratch);
593
594 __ set(page_size, Rscratch);
595 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
596
597 // Get the stack overflow limit, and in debug, verify it is non-zero.
598 __ ld_ptr(G2_thread, JavaThread::stack_overflow_limit_offset(), Rscratch);
599 #ifdef ASSERT
600 Label limit_ok;
601 __ br_notnull_short(Rscratch, Assembler::pn, limit_ok);
602 __ stop("stack overflow limit is zero in generate_stack_overflow_check");
603 __ bind(limit_ok);
604 #endif
605
606 // Add in the size of the frame (which is the same as subtracting it from the
607 // SP, which would take another register.
608 __ add(Rscratch, Rframe_size, Rscratch);
609
610 // The frame is greater than one page in size, so check against
611 // the bottom of the stack.
612 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
613
614 // The stack will overflow, throw an exception.
615
616 // Note that SP is restored to sender's sp (in the delay slot). This
617 // is necessary if the sender's frame is an extended compiled frame
618 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
619 // adaptations.
620
621 // Note also that the restored frame is not necessarily interpreted.
622 // Use the shared runtime version of the StackOverflowError.
623 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
624 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
625 __ jump_to(stub, Rscratch);
626 __ delayed()->mov(O5_savedSP, SP);
627
628 // If you get to here, then there is enough stack space.
629 __ bind(after_frame_check);
630 }
631
632
633 //
634 // Generate a fixed interpreter frame. This is identical setup for interpreted
635 // methods and for native methods hence the shared code.
636
637
638 //----------------------------------------------------------------------------------------------------
639 // Stack frame layout
640 //
641 // When control flow reaches any of the entry types for the interpreter
642 // the following holds ->
643 //
644 // C2 Calling Conventions:
645 //
646 // The entry code below assumes that the following registers are set
647 // when coming in:
648 // G5_method: holds the Method* of the method to call
649 // Lesp: points to the TOS of the callers expression stack
650 // after having pushed all the parameters
651 //
652 // The entry code does the following to setup an interpreter frame
653 // pop parameters from the callers stack by adjusting Lesp
654 // set O0 to Lesp
655 // compute X = (max_locals - num_parameters)
656 // bump SP up by X to accomadate the extra locals
657 // compute X = max_expression_stack
658 // + vm_local_words
659 // + 16 words of register save area
660 // save frame doing a save sp, -X, sp growing towards lower addresses
661 // set Lbcp, Lmethod, LcpoolCache
662 // set Llocals to i0
663 // set Lmonitors to FP - rounded_vm_local_words
664 // set Lesp to Lmonitors - 4
665 //
666 // The frame has now been setup to do the rest of the entry code
667
668 // Try this optimization: Most method entries could live in a
669 // "one size fits all" stack frame without all the dynamic size
670 // calculations. It might be profitable to do all this calculation
671 // statically and approximately for "small enough" methods.
672
673 //-----------------------------------------------------------------------------------------------
674
675 // C1 Calling conventions
676 //
677 // Upon method entry, the following registers are setup:
678 //
679 // g2 G2_thread: current thread
680 // g5 G5_method: method to activate
681 // g4 Gargs : pointer to last argument
682 //
683 //
684 // Stack:
685 //
686 // +---------------+ <--- sp
687 // | |
688 // : reg save area :
689 // | |
690 // +---------------+ <--- sp + 0x40
691 // | |
692 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
693 // | |
694 // +---------------+ <--- sp + 0x5c
695 // | |
696 // : free :
697 // | |
698 // +---------------+ <--- Gargs
699 // | |
700 // : arguments :
701 // | |
702 // +---------------+
703 // | |
704 //
705 //
706 //
707 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
708 //
709 // +---------------+ <--- sp
710 // | |
711 // : reg save area :
712 // | |
713 // +---------------+ <--- sp + 0x40
714 // | |
715 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
716 // | |
717 // +---------------+ <--- sp + 0x5c
718 // | |
719 // : :
720 // | | <--- Lesp
721 // +---------------+ <--- Lmonitors (fp - 0x18)
722 // | VM locals |
723 // +---------------+ <--- fp
724 // | |
725 // : reg save area :
726 // | |
727 // +---------------+ <--- fp + 0x40
728 // | |
729 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
730 // | |
731 // +---------------+ <--- fp + 0x5c
732 // | |
733 // : free :
734 // | |
735 // +---------------+
736 // | |
737 // : nonarg locals :
738 // | |
739 // +---------------+
740 // | |
741 // : arguments :
742 // | | <--- Llocals
743 // +---------------+ <--- Gargs
744 // | |
745
746 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
747 //
748 //
749 // The entry code sets up a new interpreter frame in 4 steps:
750 //
751 // 1) Increase caller's SP by for the extra local space needed:
752 // (check for overflow)
753 // Efficient implementation of xload/xstore bytecodes requires
754 // that arguments and non-argument locals are in a contigously
755 // addressable memory block => non-argument locals must be
756 // allocated in the caller's frame.
757 //
758 // 2) Create a new stack frame and register window:
759 // The new stack frame must provide space for the standard
760 // register save area, the maximum java expression stack size,
761 // the monitor slots (0 slots initially), and some frame local
762 // scratch locations.
763 //
764 // 3) The following interpreter activation registers must be setup:
765 // Lesp : expression stack pointer
766 // Lbcp : bytecode pointer
767 // Lmethod : method
768 // Llocals : locals pointer
769 // Lmonitors : monitor pointer
770 // LcpoolCache: constant pool cache
771 //
772 // 4) Initialize the non-argument locals if necessary:
773 // Non-argument locals may need to be initialized to NULL
774 // for GC to work. If the oop-map information is accurate
775 // (in the absence of the JSR problem), no initialization
776 // is necessary.
777 //
778 // (gri - 2/25/2000)
779
780
781 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
782
783 const int extra_space =
784 rounded_vm_local_words + // frame local scratch space
785 Method::extra_stack_entries() + // extra stack for jsr 292
786 frame::memory_parameter_word_sp_offset + // register save area
787 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
788
789 const Register Glocals_size = G3;
790 const Register RconstMethod = Glocals_size;
791 const Register Otmp1 = O3;
792 const Register Otmp2 = O4;
793 // Lscratch can't be used as a temporary because the call_stub uses
794 // it to assert that the stack frame was setup correctly.
795 const Address constMethod (G5_method, Method::const_offset());
796 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
797
798 __ ld_ptr( constMethod, RconstMethod );
799 __ lduh( size_of_parameters, Glocals_size);
800
801 // Gargs points to first local + BytesPerWord
802 // Set the saved SP after the register window save
803 //
804 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
805 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
806 __ add(Gargs, Otmp1, Gargs);
807
808 if (native_call) {
809 __ calc_mem_param_words( Glocals_size, Gframe_size );
810 __ add( Gframe_size, extra_space, Gframe_size);
811 __ round_to( Gframe_size, WordsPerLong );
812 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
813
814 // Native calls don't need the stack size check since they have no
815 // expression stack and the arguments are already on the stack and
816 // we only add a handful of words to the stack.
817 } else {
818
819 //
820 // Compute number of locals in method apart from incoming parameters
821 //
822 const Address size_of_locals(Otmp1, ConstMethod::size_of_locals_offset());
823 __ ld_ptr(constMethod, Otmp1);
824 __ lduh(size_of_locals, Otmp1);
825 __ sub(Otmp1, Glocals_size, Glocals_size);
826 __ round_to(Glocals_size, WordsPerLong);
827 __ sll(Glocals_size, Interpreter::logStackElementSize, Glocals_size);
828
829 // See if the frame is greater than one page in size. If so,
830 // then we need to verify there is enough stack space remaining.
831 // Frame_size = (max_stack + extra_space) * BytesPerWord;
832 __ ld_ptr(constMethod, Gframe_size);
833 __ lduh(Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size);
834 __ add(Gframe_size, extra_space, Gframe_size);
835 __ round_to(Gframe_size, WordsPerLong);
836 __ sll(Gframe_size, Interpreter::logStackElementSize, Gframe_size);
837
838 // Add in java locals size for stack overflow check only
839 __ add(Gframe_size, Glocals_size, Gframe_size);
840
841 const Register Otmp2 = O4;
842 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
843 generate_stack_overflow_check(Gframe_size, Otmp1);
844
845 __ sub(Gframe_size, Glocals_size, Gframe_size);
846
847 //
848 // bump SP to accomodate the extra locals
849 //
850 __ sub(SP, Glocals_size, SP);
851 }
852
853 //
854 // now set up a stack frame with the size computed above
855 //
856 __ neg( Gframe_size );
857 __ save( SP, Gframe_size, SP );
858
859 //
860 // now set up all the local cache registers
861 //
862 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
863 // that all present references to Lbyte_code initialize the register
864 // immediately before use
865 if (native_call) {
866 __ mov(G0, Lbcp);
867 } else {
868 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
869 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
870 }
871 __ mov( G5_method, Lmethod); // set Lmethod
872 // Get mirror and store it in the frame as GC root for this Method*
873 Register mirror = LcpoolCache;
874 __ load_mirror(mirror, Lmethod);
875 __ st_ptr(mirror, FP, (frame::interpreter_frame_mirror_offset * wordSize) + STACK_BIAS);
876 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
877 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
878 #ifdef _LP64
879 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
880 #endif
881 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
882
883 // setup interpreter activation registers
884 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
885
886 if (ProfileInterpreter) {
887 #ifdef FAST_DISPATCH
888 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
889 // they both use I2.
890 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
891 #endif // FAST_DISPATCH
892 __ set_method_data_pointer();
893 }
894
895 }
896
897 // Method entry for java.lang.ref.Reference.get.
898 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
899 #if INCLUDE_ALL_GCS
900 // Code: _aload_0, _getfield, _areturn
901 // parameter size = 1
902 //
903 // The code that gets generated by this routine is split into 2 parts:
904 // 1. The "intrinsified" code for G1 (or any SATB based GC),
905 // 2. The slow path - which is an expansion of the regular method entry.
906 //
907 // Notes:-
908 // * In the G1 code we do not check whether we need to block for
909 // a safepoint. If G1 is enabled then we must execute the specialized
910 // code for Reference.get (except when the Reference object is null)
911 // so that we can log the value in the referent field with an SATB
912 // update buffer.
913 // If the code for the getfield template is modified so that the
914 // G1 pre-barrier code is executed when the current method is
915 // Reference.get() then going through the normal method entry
916 // will be fine.
917 // * The G1 code can, however, check the receiver object (the instance
918 // of java.lang.Reference) and jump to the slow path if null. If the
919 // Reference object is null then we obviously cannot fetch the referent
920 // and so we don't need to call the G1 pre-barrier. Thus we can use the
921 // regular method entry code to generate the NPE.
922 //
923 // This code is based on generate_accessor_enty.
924
925 address entry = __ pc();
926
927 const int referent_offset = java_lang_ref_Reference::referent_offset;
928 guarantee(referent_offset > 0, "referent offset not initialized");
929
930 if (UseG1GC) {
931 Label slow_path;
932
933 // In the G1 code we don't check if we need to reach a safepoint. We
934 // continue and the thread will safepoint at the next bytecode dispatch.
935
936 // Check if local 0 != NULL
937 // If the receiver is null then it is OK to jump to the slow path.
938 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
939 // check if local 0 == NULL and go the slow path
940 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
941
942
943 // Load the value of the referent field.
944 if (Assembler::is_simm13(referent_offset)) {
945 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
946 } else {
947 __ set(referent_offset, G3_scratch);
948 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
949 }
950
951 // Generate the G1 pre-barrier code to log the value of
952 // the referent field in an SATB buffer. Note with
953 // these parameters the pre-barrier does not generate
954 // the load of the previous value
955
956 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
957 Otos_i /* pre_val */,
958 G3_scratch /* tmp */,
959 true /* preserve_o_regs */);
960
961 // _areturn
962 __ retl(); // return from leaf routine
963 __ delayed()->mov(O5_savedSP, SP);
964
965 // Generate regular method entry
966 __ bind(slow_path);
967 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
968 return entry;
969 }
970 #endif // INCLUDE_ALL_GCS
971
972 // If G1 is not enabled then attempt to go through the accessor entry point
973 // Reference.get is an accessor
974 return NULL;
975 }
976
977 /**
978 * Method entry for static native methods:
979 * int java.util.zip.CRC32.update(int crc, int b)
980 */
981 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
982
983 if (UseCRC32Intrinsics) {
984 address entry = __ pc();
985
986 Label L_slow_path;
987 // If we need a safepoint check, generate full interpreter entry.
988 ExternalAddress state(SafepointSynchronize::address_of_state());
989 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
990 __ set(SafepointSynchronize::_not_synchronized, O3);
991 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
992
993 // Load parameters
994 const Register crc = O0; // initial crc
995 const Register val = O1; // byte to update with
996 const Register table = O2; // address of 256-entry lookup table
997
998 __ ldub(Gargs, 3, val);
999 __ lduw(Gargs, 8, crc);
1000
1001 __ set(ExternalAddress(StubRoutines::crc_table_addr()), table);
1002
1003 __ not1(crc); // ~crc
1004 __ clruwu(crc);
1005 __ update_byte_crc32(crc, val, table);
1006 __ not1(crc); // ~crc
1007
1008 // result in O0
1009 __ retl();
1010 __ delayed()->nop();
1011
1012 // generate a vanilla native entry as the slow path
1013 __ bind(L_slow_path);
1014 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
1015 return entry;
1016 }
1017 return NULL;
1018 }
1019
1020 /**
1021 * Method entry for static native methods:
1022 * int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len)
1023 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
1024 */
1025 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1026
1027 if (UseCRC32Intrinsics) {
1028 address entry = __ pc();
1029
1030 Label L_slow_path;
1031 // If we need a safepoint check, generate full interpreter entry.
1032 ExternalAddress state(SafepointSynchronize::address_of_state());
1033 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
1034 __ set(SafepointSynchronize::_not_synchronized, O3);
1035 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
1036
1037 // Load parameters from the stack
1038 const Register crc = O0; // initial crc
1039 const Register buf = O1; // source java byte array address
1040 const Register len = O2; // len
1041 const Register offset = O3; // offset
1042
1043 // Arguments are reversed on java expression stack
1044 // Calculate address of start element
1045 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) {
1046 __ lduw(Gargs, 0, len);
1047 __ lduw(Gargs, 8, offset);
1048 __ ldx( Gargs, 16, buf);
1049 __ lduw(Gargs, 32, crc);
1050 __ add(buf, offset, buf);
1051 } else {
1052 __ lduw(Gargs, 0, len);
1053 __ lduw(Gargs, 8, offset);
1054 __ ldx( Gargs, 16, buf);
1055 __ lduw(Gargs, 24, crc);
1056 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1057 __ add(buf ,offset, buf);
1058 }
1059
1060 // Call the crc32 kernel
1061 __ MacroAssembler::save_thread(L7_thread_cache);
1062 __ kernel_crc32(crc, buf, len, O3);
1063 __ MacroAssembler::restore_thread(L7_thread_cache);
1064
1065 // result in O0
1066 __ retl();
1067 __ delayed()->nop();
1068
1069 // generate a vanilla native entry as the slow path
1070 __ bind(L_slow_path);
1071 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
1072 return entry;
1073 }
1074 return NULL;
1075 }
1076
1077 /**
1078 * Method entry for intrinsic-candidate (non-native) methods:
1079 * int java.util.zip.CRC32C.updateBytes(int crc, byte[] b, int off, int end)
1080 * int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
1081 * Unlike CRC32, CRC32C does not have any methods marked as native
1082 * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
1083 */
1084 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1085
1086 if (UseCRC32CIntrinsics) {
1087 address entry = __ pc();
1088
1089 // Load parameters from the stack
1090 const Register crc = O0; // initial crc
1091 const Register buf = O1; // source java byte array address
1092 const Register offset = O2; // offset
1093 const Register end = O3; // index of last element to process
1094 const Register len = O2; // len argument to the kernel
1095 const Register table = O3; // crc32c lookup table address
1096
1097 // Arguments are reversed on java expression stack
1098 // Calculate address of start element
1099 if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) {
1100 __ lduw(Gargs, 0, end);
1101 __ lduw(Gargs, 8, offset);
1102 __ ldx( Gargs, 16, buf);
1103 __ lduw(Gargs, 32, crc);
1104 __ add(buf, offset, buf);
1105 __ sub(end, offset, len);
1106 } else {
1107 __ lduw(Gargs, 0, end);
1108 __ lduw(Gargs, 8, offset);
1109 __ ldx( Gargs, 16, buf);
1110 __ lduw(Gargs, 24, crc);
1111 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1112 __ add(buf, offset, buf);
1113 __ sub(end, offset, len);
1114 }
1115
1116 // Call the crc32c kernel
1117 __ MacroAssembler::save_thread(L7_thread_cache);
1118 __ kernel_crc32c(crc, buf, len, table);
1119 __ MacroAssembler::restore_thread(L7_thread_cache);
1120
1121 // result in O0
1122 __ retl();
1123 __ delayed()->nop();
1124
1125 return entry;
1126 }
1127 return NULL;
1128 }
1129
1130 // Not supported
1131 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
1132 return NULL;
1133 }
1134
1135 // TODO: rather than touching all pages, check against stack_overflow_limit and bang yellow page to
1136 // generate exception
1137 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1138 // Quick & dirty stack overflow checking: bang the stack & handle trap.
1139 // Note that we do the banging after the frame is setup, since the exception
1140 // handling code expects to find a valid interpreter frame on the stack.
1141 // Doing the banging earlier fails if the caller frame is not an interpreter
1142 // frame.
1143 // (Also, the exception throwing code expects to unlock any synchronized
1144 // method receiever, so do the banging after locking the receiver.)
1145
1146 // Bang each page in the shadow zone. We can't assume it's been done for
1147 // an interpreter frame with greater than a page of locals, so each page
1148 // needs to be checked. Only true for non-native.
1149 if (UseStackBanging) {
1150 const int page_size = os::vm_page_size();
1151 const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1152 const int start_page = native_call ? n_shadow_pages : 1;
1153 for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1154 __ bang_stack_with_offset(pages*page_size);
1155 }
1156 }
1157 }
1158
1159 //
1160 // Interpreter stub for calling a native method. (asm interpreter)
1161 // This sets up a somewhat different looking stack for calling the native method
1162 // than the typical interpreter frame setup.
1163 //
1164
1165 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1166 address entry = __ pc();
1167
1168 // the following temporary registers are used during frame creation
1169 const Register Gtmp1 = G3_scratch ;
1170 const Register Gtmp2 = G1_scratch;
1171 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1172
1173 // make sure registers are different!
1174 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1175
1176 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
1177
1178 const Register Glocals_size = G3;
1179 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1180
1181 // make sure method is native & not abstract
1182 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1183 #ifdef ASSERT
1184 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1185 {
1186 Label L;
1187 __ btst(JVM_ACC_NATIVE, Gtmp1);
1188 __ br(Assembler::notZero, false, Assembler::pt, L);
1189 __ delayed()->nop();
1190 __ stop("tried to execute non-native method as native");
1191 __ bind(L);
1192 }
1193 { Label L;
1194 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1195 __ br(Assembler::zero, false, Assembler::pt, L);
1196 __ delayed()->nop();
1197 __ stop("tried to execute abstract method as non-abstract");
1198 __ bind(L);
1199 }
1200 #endif // ASSERT
1201
1202 // generate the code to allocate the interpreter stack frame
1203 generate_fixed_frame(true);
1204
1205 //
1206 // No locals to initialize for native method
1207 //
1208
1209 // this slot will be set later, we initialize it to null here just in
1210 // case we get a GC before the actual value is stored later
1211 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
1212
1213 const Address do_not_unlock_if_synchronized(G2_thread,
1214 JavaThread::do_not_unlock_if_synchronized_offset());
1215 // Since at this point in the method invocation the exception handler
1216 // would try to exit the monitor of synchronized methods which hasn't
1217 // been entered yet, we set the thread local variable
1218 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1219 // runtime, exception handling i.e. unlock_if_synchronized_method will
1220 // check this thread local flag.
1221 // This flag has two effects, one is to force an unwind in the topmost
1222 // interpreter frame and not perform an unlock while doing so.
1223
1224 __ movbool(true, G3_scratch);
1225 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1226
1227 // increment invocation counter and check for overflow
1228 //
1229 // Note: checking for negative value instead of overflow
1230 // so we have a 'sticky' overflow test (may be of
1231 // importance as soon as we have true MT/MP)
1232 Label invocation_counter_overflow;
1233 Label Lcontinue;
1234 if (inc_counter) {
1235 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1236
1237 }
1238 __ bind(Lcontinue);
1239
1240 bang_stack_shadow_pages(true);
1241
1242 // reset the _do_not_unlock_if_synchronized flag
1243 __ stbool(G0, do_not_unlock_if_synchronized);
1244
1245 // check for synchronized methods
1246 // Must happen AFTER invocation_counter check and stack overflow check,
1247 // so method is not locked if overflows.
1248
1249 if (synchronized) {
1250 lock_method();
1251 } else {
1252 #ifdef ASSERT
1253 { Label ok;
1254 __ ld(Laccess_flags, O0);
1255 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1256 __ br( Assembler::zero, false, Assembler::pt, ok);
1257 __ delayed()->nop();
1258 __ stop("method needs synchronization");
1259 __ bind(ok);
1260 }
1261 #endif // ASSERT
1262 }
1263
1264
1265 // start execution
1266 __ verify_thread();
1267
1268 // JVMTI support
1269 __ notify_method_entry();
1270
1271 // native call
1272
1273 // (note that O0 is never an oop--at most it is a handle)
1274 // It is important not to smash any handles created by this call,
1275 // until any oop handle in O0 is dereferenced.
1276
1277 // (note that the space for outgoing params is preallocated)
1278
1279 // get signature handler
1280 { Label L;
1281 Address signature_handler(Lmethod, Method::signature_handler_offset());
1282 __ ld_ptr(signature_handler, G3_scratch);
1283 __ br_notnull_short(G3_scratch, Assembler::pt, L);
1284 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
1285 __ ld_ptr(signature_handler, G3_scratch);
1286 __ bind(L);
1287 }
1288
1289 // Push a new frame so that the args will really be stored in
1290 // Copy a few locals across so the new frame has the variables
1291 // we need but these values will be dead at the jni call and
1292 // therefore not gc volatile like the values in the current
1293 // frame (Lmethod in particular)
1294
1295 // Flush the method pointer to the register save area
1296 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
1297 __ mov(Llocals, O1);
1298
1299 // calculate where the mirror handle body is allocated in the interpreter frame:
1300 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
1301
1302 // Calculate current frame size
1303 __ sub(SP, FP, O3); // Calculate negative of current frame size
1304 __ save(SP, O3, SP); // Allocate an identical sized frame
1305
1306 // Note I7 has leftover trash. Slow signature handler will fill it in
1307 // should we get there. Normal jni call will set reasonable last_Java_pc
1308 // below (and fix I7 so the stack trace doesn't have a meaningless frame
1309 // in it).
1310
1311 // Load interpreter frame's Lmethod into same register here
1312
1313 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1314
1315 __ mov(I1, Llocals);
1316 __ mov(I2, Lscratch2); // save the address of the mirror
1317
1318
1319 // ONLY Lmethod and Llocals are valid here!
1320
1321 // call signature handler, It will move the arg properly since Llocals in current frame
1322 // matches that in outer frame
1323
1324 __ callr(G3_scratch, 0);
1325 __ delayed()->nop();
1326
1327 // Result handler is in Lscratch
1328
1329 // Reload interpreter frame's Lmethod since slow signature handler may block
1330 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1331
1332 { Label not_static;
1333
1334 __ ld(Laccess_flags, O0);
1335 __ btst(JVM_ACC_STATIC, O0);
1336 __ br( Assembler::zero, false, Assembler::pt, not_static);
1337 // get native function entry point(O0 is a good temp until the very end)
1338 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
1339 // for static methods insert the mirror argument
1340 __ load_mirror(O1, Lmethod);
1341 #ifdef ASSERT
1342 if (!PrintSignatureHandlers) // do not dirty the output with this
1343 { Label L;
1344 __ br_notnull_short(O1, Assembler::pt, L);
1345 __ stop("mirror is missing");
1346 __ bind(L);
1347 }
1348 #endif // ASSERT
1349 __ st_ptr(O1, Lscratch2, 0);
1350 __ mov(Lscratch2, O1);
1351 __ bind(not_static);
1352 }
1353
1354 // At this point, arguments have been copied off of stack into
1355 // their JNI positions, which are O1..O5 and SP[68..].
1356 // Oops are boxed in-place on the stack, with handles copied to arguments.
1357 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1358
1359 #ifdef ASSERT
1360 { Label L;
1361 __ br_notnull_short(O0, Assembler::pt, L);
1362 __ stop("native entry point is missing");
1363 __ bind(L);
1364 }
1365 #endif // ASSERT
1366
1367 //
1368 // setup the frame anchor
1369 //
1370 // The scavenge function only needs to know that the PC of this frame is
1371 // in the interpreter method entry code, it doesn't need to know the exact
1372 // PC and hence we can use O7 which points to the return address from the
1373 // previous call in the code stream (signature handler function)
1374 //
1375 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1376 // we have pushed the extra frame in order to protect the volatile register(s)
1377 // in that frame when we return from the jni call
1378 //
1379
1380 __ set_last_Java_frame(FP, O7);
1381 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1382 // not meaningless information that'll confuse me.
1383
1384 // flush the windows now. We don't care about the current (protection) frame
1385 // only the outer frames
1386
1387 __ flushw();
1388
1389 // mark windows as flushed
1390 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1391 __ set(JavaFrameAnchor::flushed, G3_scratch);
1392 __ st(G3_scratch, flags);
1393
1394 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1395
1396 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1397 #ifdef ASSERT
1398 { Label L;
1399 __ ld(thread_state, G3_scratch);
1400 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1401 __ stop("Wrong thread state in native stub");
1402 __ bind(L);
1403 }
1404 #endif // ASSERT
1405 __ set(_thread_in_native, G3_scratch);
1406 __ st(G3_scratch, thread_state);
1407
1408 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1409 __ save_thread(L7_thread_cache); // save Gthread
1410 __ callr(O0, 0);
1411 __ delayed()->
1412 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1413
1414 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1415
1416 __ restore_thread(L7_thread_cache); // restore G2_thread
1417 __ reinit_heapbase();
1418
1419 // must we block?
1420
1421 // Block, if necessary, before resuming in _thread_in_Java state.
1422 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1423 { Label no_block;
1424 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1425
1426 // Switch thread to "native transition" state before reading the synchronization state.
1427 // This additional state is necessary because reading and testing the synchronization
1428 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1429 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1430 // VM thread changes sync state to synchronizing and suspends threads for GC.
1431 // Thread A is resumed to finish this native method, but doesn't block here since it
1432 // didn't see any synchronization is progress, and escapes.
1433 __ set(_thread_in_native_trans, G3_scratch);
1434 __ st(G3_scratch, thread_state);
1435 if(os::is_MP()) {
1436 if (UseMembar) {
1437 // Force this write out before the read below
1438 __ membar(Assembler::StoreLoad);
1439 } else {
1440 // Write serialization page so VM thread can do a pseudo remote membar.
1441 // We use the current thread pointer to calculate a thread specific
1442 // offset to write to within the page. This minimizes bus traffic
1443 // due to cache line collision.
1444 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1445 }
1446 }
1447 __ load_contents(sync_state, G3_scratch);
1448 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1449
1450 Label L;
1451 __ br(Assembler::notEqual, false, Assembler::pn, L);
1452 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1453 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1454 __ bind(L);
1455
1456 // Block. Save any potential method result value before the operation and
1457 // use a leaf call to leave the last_Java_frame setup undisturbed.
1458 save_native_result();
1459 __ call_VM_leaf(L7_thread_cache,
1460 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1461 G2_thread);
1462
1463 // Restore any method result value
1464 restore_native_result();
1465 __ bind(no_block);
1466 }
1467
1468 // Clear the frame anchor now
1469
1470 __ reset_last_Java_frame();
1471
1472 // Move the result handler address
1473 __ mov(Lscratch, G3_scratch);
1474 // return possible result to the outer frame
1475 #ifndef __LP64
1476 __ mov(O0, I0);
1477 __ restore(O1, G0, O1);
1478 #else
1479 __ restore(O0, G0, O0);
1480 #endif /* __LP64 */
1481
1482 // Move result handler to expected register
1483 __ mov(G3_scratch, Lscratch);
1484
1485 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1486 // switch to thread_in_Java.
1487
1488 __ set(_thread_in_Java, G3_scratch);
1489 __ st(G3_scratch, thread_state);
1490
1491 // reset handle block
1492 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1493 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1494
1495 // If we have an oop result store it where it will be safe for any further gc
1496 // until we return now that we've released the handle it might be protected by
1497
1498 {
1499 Label no_oop, store_result;
1500
1501 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1502 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1503 __ addcc(G0, O0, O0);
1504 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1505 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1506 __ mov(G0, O0);
1507
1508 __ bind(store_result);
1509 // Store it where gc will look for it and result handler expects it.
1510 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1511
1512 __ bind(no_oop);
1513
1514 }
1515
1516
1517 // handle exceptions (exception handling will handle unlocking!)
1518 { Label L;
1519 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1520 __ ld_ptr(exception_addr, Gtemp);
1521 __ br_null_short(Gtemp, Assembler::pt, L);
1522 // Note: This could be handled more efficiently since we know that the native
1523 // method doesn't have an exception handler. We could directly return
1524 // to the exception handler for the caller.
1525 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1526 __ should_not_reach_here();
1527 __ bind(L);
1528 }
1529
1530 // JVMTI support (preserves thread register)
1531 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1532
1533 if (synchronized) {
1534 // save and restore any potential method result value around the unlocking operation
1535 save_native_result();
1536
1537 __ add( __ top_most_monitor(), O1);
1538 __ unlock_object(O1);
1539
1540 restore_native_result();
1541 }
1542
1543 #if defined(COMPILER2) && !defined(_LP64)
1544
1545 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1546 // or compiled so just be safe.
1547
1548 __ sllx(O0, 32, G1); // Shift bits into high G1
1549 __ srl (O1, 0, O1); // Zero extend O1
1550 __ or3 (O1, G1, G1); // OR 64 bits into G1
1551
1552 #endif /* COMPILER2 && !_LP64 */
1553
1554 // dispose of return address and remove activation
1555 #ifdef ASSERT
1556 {
1557 Label ok;
1558 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1559 __ stop("bad I5_savedSP value");
1560 __ should_not_reach_here();
1561 __ bind(ok);
1562 }
1563 #endif
1564 __ jmp(Lscratch, 0);
1565 __ delayed()->nop();
1566
1567
1568 if (inc_counter) {
1569 // handle invocation counter overflow
1570 __ bind(invocation_counter_overflow);
1571 generate_counter_overflow(Lcontinue);
1572 }
1573
1574
1575
1576 return entry;
1577 }
1578
1579
1580 // Generic method entry to (asm) interpreter
1581 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1582 address entry = __ pc();
1583
1584 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1585
1586 // the following temporary registers are used during frame creation
1587 const Register Gtmp1 = G3_scratch ;
1588 const Register Gtmp2 = G1_scratch;
1589
1590 // make sure registers are different!
1591 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1592
1593 const Address constMethod (G5_method, Method::const_offset());
1594 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1595 // and use in the asserts.
1596 const Address access_flags (Lmethod, Method::access_flags_offset());
1597
1598 const Register Glocals_size = G3;
1599 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1600
1601 // make sure method is not native & not abstract
1602 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1603 #ifdef ASSERT
1604 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1605 {
1606 Label L;
1607 __ btst(JVM_ACC_NATIVE, Gtmp1);
1608 __ br(Assembler::zero, false, Assembler::pt, L);
1609 __ delayed()->nop();
1610 __ stop("tried to execute native method as non-native");
1611 __ bind(L);
1612 }
1613 { Label L;
1614 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1615 __ br(Assembler::zero, false, Assembler::pt, L);
1616 __ delayed()->nop();
1617 __ stop("tried to execute abstract method as non-abstract");
1618 __ bind(L);
1619 }
1620 #endif // ASSERT
1621
1622 // generate the code to allocate the interpreter stack frame
1623
1624 generate_fixed_frame(false);
1625
1626 #ifdef FAST_DISPATCH
1627 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1628 // set bytecode dispatch table base
1629 #endif
1630
1631 //
1632 // Code to initialize the extra (i.e. non-parm) locals
1633 //
1634 Register init_value = noreg; // will be G0 if we must clear locals
1635 // The way the code was setup before zerolocals was always true for vanilla java entries.
1636 // It could only be false for the specialized entries like accessor or empty which have
1637 // no extra locals so the testing was a waste of time and the extra locals were always
1638 // initialized. We removed this extra complication to already over complicated code.
1639
1640 init_value = G0;
1641 Label clear_loop;
1642
1643 const Register RconstMethod = O1;
1644 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1645 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1646
1647 // NOTE: If you change the frame layout, this code will need to
1648 // be updated!
1649 __ ld_ptr( constMethod, RconstMethod );
1650 __ lduh( size_of_locals, O2 );
1651 __ lduh( size_of_parameters, O1 );
1652 __ sll( O2, Interpreter::logStackElementSize, O2);
1653 __ sll( O1, Interpreter::logStackElementSize, O1 );
1654 __ sub( Llocals, O2, O2 );
1655 __ sub( Llocals, O1, O1 );
1656
1657 __ bind( clear_loop );
1658 __ inc( O2, wordSize );
1659
1660 __ cmp( O2, O1 );
1661 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1662 __ delayed()->st_ptr( init_value, O2, 0 );
1663
1664 const Address do_not_unlock_if_synchronized(G2_thread,
1665 JavaThread::do_not_unlock_if_synchronized_offset());
1666 // Since at this point in the method invocation the exception handler
1667 // would try to exit the monitor of synchronized methods which hasn't
1668 // been entered yet, we set the thread local variable
1669 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1670 // runtime, exception handling i.e. unlock_if_synchronized_method will
1671 // check this thread local flag.
1672 __ movbool(true, G3_scratch);
1673 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1674
1675 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1676 // increment invocation counter and check for overflow
1677 //
1678 // Note: checking for negative value instead of overflow
1679 // so we have a 'sticky' overflow test (may be of
1680 // importance as soon as we have true MT/MP)
1681 Label invocation_counter_overflow;
1682 Label profile_method;
1683 Label profile_method_continue;
1684 Label Lcontinue;
1685 if (inc_counter) {
1686 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1687 if (ProfileInterpreter) {
1688 __ bind(profile_method_continue);
1689 }
1690 }
1691 __ bind(Lcontinue);
1692
1693 bang_stack_shadow_pages(false);
1694
1695 // reset the _do_not_unlock_if_synchronized flag
1696 __ stbool(G0, do_not_unlock_if_synchronized);
1697
1698 // check for synchronized methods
1699 // Must happen AFTER invocation_counter check and stack overflow check,
1700 // so method is not locked if overflows.
1701
1702 if (synchronized) {
1703 lock_method();
1704 } else {
1705 #ifdef ASSERT
1706 { Label ok;
1707 __ ld(access_flags, O0);
1708 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1709 __ br( Assembler::zero, false, Assembler::pt, ok);
1710 __ delayed()->nop();
1711 __ stop("method needs synchronization");
1712 __ bind(ok);
1713 }
1714 #endif // ASSERT
1715 }
1716
1717 // start execution
1718
1719 __ verify_thread();
1720
1721 // jvmti support
1722 __ notify_method_entry();
1723
1724 // start executing instructions
1725 __ dispatch_next(vtos);
1726
1727
1728 if (inc_counter) {
1729 if (ProfileInterpreter) {
1730 // We have decided to profile this method in the interpreter
1731 __ bind(profile_method);
1732
1733 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1734 __ set_method_data_pointer_for_bcp();
1735 __ ba_short(profile_method_continue);
1736 }
1737
1738 // handle invocation counter overflow
1739 __ bind(invocation_counter_overflow);
1740 generate_counter_overflow(Lcontinue);
1741 }
1742
1743
1744 return entry;
1745 }
1746
1747 //----------------------------------------------------------------------------------------------------
1748 // Exceptions
1749 void TemplateInterpreterGenerator::generate_throw_exception() {
1750
1751 // Entry point in previous activation (i.e., if the caller was interpreted)
1752 Interpreter::_rethrow_exception_entry = __ pc();
1753 // O0: exception
1754
1755 // entry point for exceptions thrown within interpreter code
1756 Interpreter::_throw_exception_entry = __ pc();
1757 __ verify_thread();
1758 // expression stack is undefined here
1759 // O0: exception, i.e. Oexception
1760 // Lbcp: exception bcp
1761 __ verify_oop(Oexception);
1762
1763
1764 // expression stack must be empty before entering the VM in case of an exception
1765 __ empty_expression_stack();
1766 // find exception handler address and preserve exception oop
1767 // call C routine to find handler and jump to it
1768 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1769 __ push_ptr(O1); // push exception for exception handler bytecodes
1770
1771 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1772 __ delayed()->nop();
1773
1774
1775 // if the exception is not handled in the current frame
1776 // the frame is removed and the exception is rethrown
1777 // (i.e. exception continuation is _rethrow_exception)
1778 //
1779 // Note: At this point the bci is still the bxi for the instruction which caused
1780 // the exception and the expression stack is empty. Thus, for any VM calls
1781 // at this point, GC will find a legal oop map (with empty expression stack).
1782
1783 // in current activation
1784 // tos: exception
1785 // Lbcp: exception bcp
1786
1787 //
1788 // JVMTI PopFrame support
1789 //
1790
1791 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1792 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1793 // Set the popframe_processing bit in popframe_condition indicating that we are
1794 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1795 // popframe handling cycles.
1796
1797 __ ld(popframe_condition_addr, G3_scratch);
1798 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1799 __ stw(G3_scratch, popframe_condition_addr);
1800
1801 // Empty the expression stack, as in normal exception handling
1802 __ empty_expression_stack();
1803 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1804
1805 {
1806 // Check to see whether we are returning to a deoptimized frame.
1807 // (The PopFrame call ensures that the caller of the popped frame is
1808 // either interpreted or compiled and deoptimizes it if compiled.)
1809 // In this case, we can't call dispatch_next() after the frame is
1810 // popped, but instead must save the incoming arguments and restore
1811 // them after deoptimization has occurred.
1812 //
1813 // Note that we don't compare the return PC against the
1814 // deoptimization blob's unpack entry because of the presence of
1815 // adapter frames in C2.
1816 Label caller_not_deoptimized;
1817 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1818 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1819
1820 const Register Gtmp1 = G3_scratch;
1821 const Register Gtmp2 = G1_scratch;
1822 const Register RconstMethod = Gtmp1;
1823 const Address constMethod(Lmethod, Method::const_offset());
1824 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1825
1826 // Compute size of arguments for saving when returning to deoptimized caller
1827 __ ld_ptr(constMethod, RconstMethod);
1828 __ lduh(size_of_parameters, Gtmp1);
1829 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1830 __ sub(Llocals, Gtmp1, Gtmp2);
1831 __ add(Gtmp2, wordSize, Gtmp2);
1832 // Save these arguments
1833 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1834 // Inform deoptimization that it is responsible for restoring these arguments
1835 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1836 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1837 __ st(Gtmp1, popframe_condition_addr);
1838
1839 // Return from the current method
1840 // The caller's SP was adjusted upon method entry to accomodate
1841 // the callee's non-argument locals. Undo that adjustment.
1842 __ ret();
1843 __ delayed()->restore(I5_savedSP, G0, SP);
1844
1845 __ bind(caller_not_deoptimized);
1846 }
1847
1848 // Clear the popframe condition flag
1849 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1850
1851 // Get out of the current method (how this is done depends on the particular compiler calling
1852 // convention that the interpreter currently follows)
1853 // The caller's SP was adjusted upon method entry to accomodate
1854 // the callee's non-argument locals. Undo that adjustment.
1855 __ restore(I5_savedSP, G0, SP);
1856 // The method data pointer was incremented already during
1857 // call profiling. We have to restore the mdp for the current bcp.
1858 if (ProfileInterpreter) {
1859 __ set_method_data_pointer_for_bcp();
1860 }
1861
1862 #if INCLUDE_JVMTI
1863 {
1864 Label L_done;
1865
1866 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1867 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1868
1869 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1870 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1871
1872 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1873
1874 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1875 __ delayed()->nop();
1876
1877 __ st_ptr(G1_scratch, Lesp, wordSize);
1878 __ bind(L_done);
1879 }
1880 #endif // INCLUDE_JVMTI
1881
1882 // Resume bytecode interpretation at the current bcp
1883 __ dispatch_next(vtos);
1884 // end of JVMTI PopFrame support
1885
1886 Interpreter::_remove_activation_entry = __ pc();
1887
1888 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1889 __ pop_ptr(Oexception); // get exception
1890
1891 // Intel has the following comment:
1892 //// remove the activation (without doing throws on illegalMonitorExceptions)
1893 // They remove the activation without checking for bad monitor state.
1894 // %%% We should make sure this is the right semantics before implementing.
1895
1896 __ set_vm_result(Oexception);
1897 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1898
1899 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1900
1901 __ get_vm_result(Oexception);
1902 __ verify_oop(Oexception);
1903
1904 const int return_reg_adjustment = frame::pc_return_offset;
1905 Address issuing_pc_addr(I7, return_reg_adjustment);
1906
1907 // We are done with this activation frame; find out where to go next.
1908 // The continuation point will be an exception handler, which expects
1909 // the following registers set up:
1910 //
1911 // Oexception: exception
1912 // Oissuing_pc: the local call that threw exception
1913 // Other On: garbage
1914 // In/Ln: the contents of the caller's register window
1915 //
1916 // We do the required restore at the last possible moment, because we
1917 // need to preserve some state across a runtime call.
1918 // (Remember that the caller activation is unknown--it might not be
1919 // interpreted, so things like Lscratch are useless in the caller.)
1920
1921 // Although the Intel version uses call_C, we can use the more
1922 // compact call_VM. (The only real difference on SPARC is a
1923 // harmlessly ignored [re]set_last_Java_frame, compared with
1924 // the Intel code which lacks this.)
1925 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1926 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1927 __ super_call_VM_leaf(L7_thread_cache,
1928 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1929 G2_thread, Oissuing_pc->after_save());
1930
1931 // The caller's SP was adjusted upon method entry to accomodate
1932 // the callee's non-argument locals. Undo that adjustment.
1933 __ JMP(O0, 0); // return exception handler in caller
1934 __ delayed()->restore(I5_savedSP, G0, SP);
1935
1936 // (same old exception object is already in Oexception; see above)
1937 // Note that an "issuing PC" is actually the next PC after the call
1938 }
1939
1940
1941 //
1942 // JVMTI ForceEarlyReturn support
1943 //
1944
1945 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1946 address entry = __ pc();
1947
1948 __ empty_expression_stack();
1949 __ load_earlyret_value(state);
1950
1951 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1952 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1953
1954 // Clear the earlyret state
1955 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1956
1957 __ remove_activation(state,
1958 /* throw_monitor_exception */ false,
1959 /* install_monitor_exception */ false);
1960
1961 // The caller's SP was adjusted upon method entry to accomodate
1962 // the callee's non-argument locals. Undo that adjustment.
1963 __ ret(); // return to caller
1964 __ delayed()->restore(I5_savedSP, G0, SP);
1965
1966 return entry;
1967 } // end of JVMTI ForceEarlyReturn support
1968
1969
1970 //------------------------------------------------------------------------------------------------------------------------
1971 // Helper for vtos entry point generation
1972
1973 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) {
1974 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1975 Label L;
1976 aep = __ pc(); __ push_ptr(); __ ba_short(L);
1977 fep = __ pc(); __ push_f(); __ ba_short(L);
1978 dep = __ pc(); __ push_d(); __ ba_short(L);
1979 lep = __ pc(); __ push_l(); __ ba_short(L);
1980 iep = __ pc(); __ push_i();
1981 bep = cep = sep = iep; // there aren't any
1982 vep = __ pc(); __ bind(L); // fall through
1983 generate_and_dispatch(t);
1984 }
1985
1986 // --------------------------------------------------------------------------------
1987
1988 // Non-product code
1989 #ifndef PRODUCT
1990 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1991 address entry = __ pc();
1992
1993 __ push(state);
1994 __ mov(O7, Lscratch); // protect return address within interpreter
1995
1996 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1997 __ mov( Otos_l2, G3_scratch );
1998 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1999 __ mov(Lscratch, O7); // restore return address
2000 __ pop(state);
2001 __ retl();
2002 __ delayed()->nop();
2003
2004 return entry;
2005 }
2006
2007
2008 // helpers for generate_and_dispatch
2009
2010 void TemplateInterpreterGenerator::count_bytecode() {
2011 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2012 }
2013
2014
2015 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2016 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2017 }
2018
2019
2020 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2021 AddressLiteral index (&BytecodePairHistogram::_index);
2022 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2023
2024 // get index, shift out old bytecode, bring in new bytecode, and store it
2025 // _index = (_index >> log2_number_of_codes) |
2026 // (bytecode << log2_number_of_codes);
2027
2028 __ load_contents(index, G4_scratch);
2029 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2030 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
2031 __ or3( G3_scratch, G4_scratch, G4_scratch );
2032 __ store_contents(G4_scratch, index, G3_scratch);
2033
2034 // bump bucket contents
2035 // _counters[_index] ++;
2036
2037 __ set(counters, G3_scratch); // loads into G3_scratch
2038 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
2039 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
2040 __ ld (G3_scratch, 0, G4_scratch);
2041 __ inc (G4_scratch);
2042 __ st (G4_scratch, 0, G3_scratch);
2043 }
2044
2045
2046 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2047 // Call a little run-time stub to avoid blow-up for each bytecode.
2048 // The run-time runtime saves the right registers, depending on
2049 // the tosca in-state for the given template.
2050 address entry = Interpreter::trace_code(t->tos_in());
2051 guarantee(entry != NULL, "entry must have been generated");
2052 __ call(entry, relocInfo::none);
2053 __ delayed()->nop();
2054 }
2055
2056
2057 void TemplateInterpreterGenerator::stop_interpreter_at() {
2058 AddressLiteral counter(&BytecodeCounter::_counter_value);
2059 __ load_contents(counter, G3_scratch);
2060 AddressLiteral stop_at(&StopInterpreterAt);
2061 __ load_ptr_contents(stop_at, G4_scratch);
2062 __ cmp(G3_scratch, G4_scratch);
2063 __ breakpoint_trap(Assembler::equal, Assembler::icc);
2064 }
2065 #endif // not PRODUCT