1 /*
2 * Copyright (c) 1997, 2017, 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 "gc/shared/barrierSetCodeGen.hpp"
28 #include "interpreter/bytecodeHistogram.hpp"
29 #include "interpreter/interpreter.hpp"
30 #include "interpreter/interpreterRuntime.hpp"
31 #include "interpreter/interp_masm.hpp"
32 #include "interpreter/templateInterpreterGenerator.hpp"
33 #include "interpreter/templateTable.hpp"
34 #include "oops/arrayOop.hpp"
35 #include "oops/methodData.hpp"
36 #include "oops/method.hpp"
37 #include "oops/oop.inline.hpp"
38 #include "prims/jvmtiExport.hpp"
39 #include "prims/jvmtiThreadState.hpp"
40 #include "runtime/arguments.hpp"
41 #include "runtime/deoptimization.hpp"
42 #include "runtime/frame.inline.hpp"
43 #include "runtime/sharedRuntime.hpp"
44 #include "runtime/stubRoutines.hpp"
45 #include "runtime/synchronizer.hpp"
46 #include "runtime/timer.hpp"
47 #include "runtime/vframeArray.hpp"
48 #include "utilities/debug.hpp"
49 #include "utilities/macros.hpp"
50
51 #ifndef FAST_DISPATCH
52 #define FAST_DISPATCH 1
53 #endif
54 #undef FAST_DISPATCH
55
56 // Size of interpreter code. Increase if too small. Interpreter will
57 // fail with a guarantee ("not enough space for interpreter generation");
58 // if too small.
59 // Run with +PrintInterpreter to get the VM to print out the size.
60 // Max size with JVMTI
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
65 // Generation of Interpreter
66 //
67 // The TemplateInterpreterGenerator generates the interpreter into Interpreter::_code.
68
69
70 #define __ _masm->
71
72
73 //----------------------------------------------------------------------------------------------------
74
75 // LP64 passes floating point arguments in F1, F3, F5, etc. instead of
76 // O0, O1, O2 etc..
77 // Doubles are passed in D0, D2, D4
78 // We store the signature of the first 16 arguments in the first argument
79 // slot because it will be overwritten prior to calling the native
80 // function, with the pointer to the JNIEnv.
81 // If LP64 there can be up to 16 floating point arguments in registers
82 // or 6 integer registers.
83 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
84
85 enum {
86 non_float = 0,
87 float_sig = 1,
88 double_sig = 2,
89 sig_mask = 3
90 };
91
92 address entry = __ pc();
93 Argument argv(0, true);
94
95 // We are in the jni transition frame. Save the last_java_frame corresponding to the
96 // outer interpreter frame
97 //
98 __ set_last_Java_frame(FP, noreg);
99 // make sure the interpreter frame we've pushed has a valid return pc
100 __ mov(O7, I7);
101 __ mov(Lmethod, G3_scratch);
102 __ mov(Llocals, G4_scratch);
103 __ save_frame(0);
104 __ mov(G2_thread, L7_thread_cache);
105 __ add(argv.address_in_frame(), O3);
106 __ mov(G2_thread, O0);
107 __ mov(G3_scratch, O1);
108 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type);
109 __ delayed()->mov(G4_scratch, O2);
110 __ mov(L7_thread_cache, G2_thread);
111 __ reset_last_Java_frame();
112
113
114 // load the register arguments (the C code packed them as varargs)
115 Address Sig = argv.address_in_frame(); // Argument 0 holds the signature
116 __ ld_ptr( Sig, G3_scratch ); // Get register argument signature word into G3_scratch
117 __ mov( G3_scratch, G4_scratch);
118 __ srl( G4_scratch, 2, G4_scratch); // Skip Arg 0
119 Label done;
120 for (Argument ldarg = argv.successor(); ldarg.is_float_register(); ldarg = ldarg.successor()) {
121 Label NonFloatArg;
122 Label LoadFloatArg;
123 Label LoadDoubleArg;
124 Label NextArg;
125 Address a = ldarg.address_in_frame();
126 __ andcc(G4_scratch, sig_mask, G3_scratch);
127 __ br(Assembler::zero, false, Assembler::pt, NonFloatArg);
128 __ delayed()->nop();
129
130 __ cmp(G3_scratch, float_sig );
131 __ br(Assembler::equal, false, Assembler::pt, LoadFloatArg);
132 __ delayed()->nop();
133
134 __ cmp(G3_scratch, double_sig );
135 __ br(Assembler::equal, false, Assembler::pt, LoadDoubleArg);
136 __ delayed()->nop();
137
138 __ bind(NonFloatArg);
139 // There are only 6 integer register arguments!
140 if ( ldarg.is_register() )
141 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register());
142 else {
143 // Optimization, see if there are any more args and get out prior to checking
144 // all 16 float registers. My guess is that this is rare.
145 // If is_register is false, then we are done the first six integer args.
146 __ br_null_short(G4_scratch, Assembler::pt, done);
147 }
148 __ ba(NextArg);
149 __ delayed()->srl( G4_scratch, 2, G4_scratch );
150
151 __ bind(LoadFloatArg);
152 __ ldf( FloatRegisterImpl::S, a, ldarg.as_float_register(), 4);
153 __ ba(NextArg);
154 __ delayed()->srl( G4_scratch, 2, G4_scratch );
155
156 __ bind(LoadDoubleArg);
157 __ ldf( FloatRegisterImpl::D, a, ldarg.as_double_register() );
158 __ ba(NextArg);
159 __ delayed()->srl( G4_scratch, 2, G4_scratch );
160
161 __ bind(NextArg);
162
163 }
164
165 __ bind(done);
166 __ ret();
167 __ delayed()->
168 restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler
169 return entry;
170 }
171
172 void TemplateInterpreterGenerator::generate_counter_overflow(Label& Lcontinue) {
173
174 // Generate code to initiate compilation on the counter overflow.
175
176 // InterpreterRuntime::frequency_counter_overflow takes two arguments,
177 // the first indicates if the counter overflow occurs at a backwards branch (NULL bcp)
178 // and the second is only used when the first is true. We pass zero for both.
179 // The call returns the address of the verified entry point for the method or NULL
180 // if the compilation did not complete (either went background or bailed out).
181 __ set((int)false, O2);
182 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), O2, O2, true);
183 // returns verified_entry_point or NULL
184 // we ignore it in any case
185 __ ba_short(Lcontinue);
186
187 }
188
189
190 // End of helpers
191
192 // Various method entries
193
194 // Abstract method entry
195 // Attempt to execute abstract method. Throw exception
196 //
197 address TemplateInterpreterGenerator::generate_abstract_entry(void) {
198 address entry = __ pc();
199 // abstract method entry
200 // throw exception
201 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError));
202 // the call_VM checks for exception, so we should never return here.
203 __ should_not_reach_here();
204 return entry;
205
206 }
207
208 void TemplateInterpreterGenerator::save_native_result(void) {
209 // result potentially in O0/O1: save it across calls
210 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
211
212 // result potentially in F0/F1: save it across calls
213 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
214
215 // save and restore any potential method result value around the unlocking operation
216 __ stf(FloatRegisterImpl::D, F0, d_tmp);
217 __ stx(O0, l_tmp);
218 }
219
220 void TemplateInterpreterGenerator::restore_native_result(void) {
221 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
222 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
223
224 // Restore any method result value
225 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
226 __ ldx(l_tmp, O0);
227 }
228
229 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
230 assert(!pass_oop || message == NULL, "either oop or message but not both");
231 address entry = __ pc();
232 // expression stack must be empty before entering the VM if an exception happened
233 __ empty_expression_stack();
234 // load exception object
235 __ set((intptr_t)name, G3_scratch);
236 if (pass_oop) {
237 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
238 } else {
239 __ set((intptr_t)message, G4_scratch);
240 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
241 }
242 // throw exception
243 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
244 AddressLiteral thrower(Interpreter::throw_exception_entry());
245 __ jump_to(thrower, G3_scratch);
246 __ delayed()->nop();
247 return entry;
248 }
249
250 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
251 address entry = __ pc();
252 // expression stack must be empty before entering the VM if an exception
253 // happened
254 __ empty_expression_stack();
255 // load exception object
256 __ call_VM(Oexception,
257 CAST_FROM_FN_PTR(address,
258 InterpreterRuntime::throw_ClassCastException),
259 Otos_i);
260 __ should_not_reach_here();
261 return entry;
262 }
263
264
265 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
266 address entry = __ pc();
267 // expression stack must be empty before entering the VM if an exception happened
268 __ empty_expression_stack();
269 // convention: expect aberrant index in register G3_scratch, then shuffle the
270 // index to G4_scratch for the VM call
271 __ mov(G3_scratch, G4_scratch);
272 __ set((intptr_t)name, G3_scratch);
273 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
274 __ should_not_reach_here();
275 return entry;
276 }
277
278
279 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
280 address entry = __ pc();
281 // expression stack must be empty before entering the VM if an exception happened
282 __ empty_expression_stack();
283 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
284 __ should_not_reach_here();
285 return entry;
286 }
287
288
289 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
290 address entry = __ pc();
291
292 if (state == atos) {
293 __ profile_return_type(O0, G3_scratch, G1_scratch);
294 }
295
296 // The callee returns with the stack possibly adjusted by adapter transition
297 // We remove that possible adjustment here.
298 // All interpreter local registers are untouched. Any result is passed back
299 // in the O0/O1 or float registers. Before continuing, the arguments must be
300 // popped from the java expression stack; i.e., Lesp must be adjusted.
301
302 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
303
304 const Register cache = G3_scratch;
305 const Register index = G1_scratch;
306 __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
307
308 const Register flags = cache;
309 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
310 const Register parameter_size = flags;
311 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words
312 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes
313 __ add(Lesp, parameter_size, Lesp); // pop arguments
314
315 __ check_and_handle_popframe(Gtemp);
316 __ check_and_handle_earlyret(Gtemp);
317
318 __ dispatch_next(state, step);
319
320 return entry;
321 }
322
323
324 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
325 address entry = __ pc();
326 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
327 #if INCLUDE_JVMCI
328 // Check if we need to take lock at entry of synchronized method. This can
329 // only occur on method entry so emit it only for vtos with step 0.
330 if (UseJVMCICompiler && state == vtos && step == 0) {
331 Label L;
332 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
333 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter
334 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
335 // Clear flag.
336 __ stbool(G0, pending_monitor_enter_addr);
337 // Take lock.
338 lock_method();
339 __ bind(L);
340 } else {
341 #ifdef ASSERT
342 if (UseJVMCICompiler) {
343 Label L;
344 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
345 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter
346 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
347 __ stop("unexpected pending monitor in deopt entry");
348 __ bind(L);
349 }
350 #endif
351 }
352 #endif
353 { Label L;
354 Address exception_addr(G2_thread, Thread::pending_exception_offset());
355 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
356 __ br_null_short(Gtemp, Assembler::pt, L);
357 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
358 __ should_not_reach_here();
359 __ bind(L);
360 }
361 __ dispatch_next(state, step);
362 return entry;
363 }
364
365 // A result handler converts/unboxes a native call result into
366 // a java interpreter/compiler result. The current frame is an
367 // interpreter frame. The activation frame unwind code must be
368 // consistent with that of TemplateTable::_return(...). In the
369 // case of native methods, the caller's SP was not modified.
370 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
371 address entry = __ pc();
372 Register Itos_i = Otos_i ->after_save();
373 Register Itos_l = Otos_l ->after_save();
374 Register Itos_l1 = Otos_l1->after_save();
375 Register Itos_l2 = Otos_l2->after_save();
376 switch (type) {
377 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
378 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
379 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
380 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
381 case T_LONG :
382 case T_INT : __ mov(O0, Itos_i); break;
383 case T_VOID : /* nothing to do */ break;
384 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
385 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
386 case T_OBJECT :
387 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
388 __ verify_oop(Itos_i);
389 break;
390 default : ShouldNotReachHere();
391 }
392 __ ret(); // return from interpreter activation
393 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
394 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
395 return entry;
396 }
397
398 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
399 address entry = __ pc();
400 __ push(state);
401 __ call_VM(noreg, runtime_entry);
402 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
403 return entry;
404 }
405
406
407 //
408 // Helpers for commoning out cases in the various type of method entries.
409 //
410
411 // increment invocation count & check for overflow
412 //
413 // Note: checking for negative value instead of overflow
414 // so we have a 'sticky' overflow test
415 //
416 // Lmethod: method
417 // ??: invocation counter
418 //
419 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
420 // Note: In tiered we increment either counters in MethodCounters* or in
421 // MDO depending if we're profiling or not.
422 const Register G3_method_counters = G3_scratch;
423 Label done;
424
425 if (TieredCompilation) {
426 const int increment = InvocationCounter::count_increment;
427 Label no_mdo;
428 if (ProfileInterpreter) {
429 // If no method data exists, go to profile_continue.
430 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
431 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
432 // Increment counter
433 Address mdo_invocation_counter(G4_scratch,
434 in_bytes(MethodData::invocation_counter_offset()) +
435 in_bytes(InvocationCounter::counter_offset()));
436 Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset()));
437 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
438 G3_scratch, Lscratch,
439 Assembler::zero, overflow);
440 __ ba_short(done);
441 }
442
443 // Increment counter in MethodCounters*
444 __ bind(no_mdo);
445 Address invocation_counter(G3_method_counters,
446 in_bytes(MethodCounters::invocation_counter_offset()) +
447 in_bytes(InvocationCounter::counter_offset()));
448 __ get_method_counters(Lmethod, G3_method_counters, done);
449 Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset()));
450 __ increment_mask_and_jump(invocation_counter, increment, mask,
451 G4_scratch, Lscratch,
452 Assembler::zero, overflow);
453 __ bind(done);
454 } else { // not TieredCompilation
455 // Update standard invocation counters
456 __ get_method_counters(Lmethod, G3_method_counters, done);
457 __ increment_invocation_counter(G3_method_counters, O0, G4_scratch);
458 if (ProfileInterpreter) {
459 Address interpreter_invocation_counter(G3_method_counters,
460 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
461 __ ld(interpreter_invocation_counter, G4_scratch);
462 __ inc(G4_scratch);
463 __ st(G4_scratch, interpreter_invocation_counter);
464 }
465
466 if (ProfileInterpreter && profile_method != NULL) {
467 // Test to see if we should create a method data oop
468 Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset()));
469 __ ld(profile_limit, G1_scratch);
470 __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
471
472 // if no method data exists, go to profile_method
473 __ test_method_data_pointer(*profile_method);
474 }
475
476 Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset()));
477 __ ld(invocation_limit, G3_scratch);
478 __ cmp(O0, G3_scratch);
479 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
480 __ delayed()->nop();
481 __ bind(done);
482 }
483
484 }
485
486 // Allocate monitor and lock method (asm interpreter)
487 // ebx - Method*
488 //
489 void TemplateInterpreterGenerator::lock_method() {
490 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
491
492 #ifdef ASSERT
493 { Label ok;
494 __ btst(JVM_ACC_SYNCHRONIZED, O0);
495 __ br( Assembler::notZero, false, Assembler::pt, ok);
496 __ delayed()->nop();
497 __ stop("method doesn't need synchronization");
498 __ bind(ok);
499 }
500 #endif // ASSERT
501
502 // get synchronization object to O0
503 { Label done;
504 __ btst(JVM_ACC_STATIC, O0);
505 __ br( Assembler::zero, true, Assembler::pt, done);
506 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
507
508 // lock the mirror, not the Klass*
509 __ load_mirror(O0, Lmethod);
510
511 #ifdef ASSERT
512 __ tst(O0);
513 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
514 #endif // ASSERT
515
516 __ bind(done);
517 }
518
519 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
520 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
521 // __ untested("lock_object from method entry");
522 __ lock_object(Lmonitors, O0);
523 }
524
525 // See if we've got enough room on the stack for locals plus overhead below
526 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
527 // without going through the signal handler, i.e., reserved and yellow zones
528 // will not be made usable. The shadow zone must suffice to handle the
529 // overflow.
530 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
531 Register Rscratch) {
532 const int page_size = os::vm_page_size();
533 Label after_frame_check;
534
535 assert_different_registers(Rframe_size, Rscratch);
536
537 __ set(page_size, Rscratch);
538 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
539
540 // Get the stack overflow limit, and in debug, verify it is non-zero.
541 __ ld_ptr(G2_thread, JavaThread::stack_overflow_limit_offset(), Rscratch);
542 #ifdef ASSERT
543 Label limit_ok;
544 __ br_notnull_short(Rscratch, Assembler::pn, limit_ok);
545 __ stop("stack overflow limit is zero in generate_stack_overflow_check");
546 __ bind(limit_ok);
547 #endif
548
549 // Add in the size of the frame (which is the same as subtracting it from the
550 // SP, which would take another register.
551 __ add(Rscratch, Rframe_size, Rscratch);
552
553 // The frame is greater than one page in size, so check against
554 // the bottom of the stack.
555 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
556
557 // The stack will overflow, throw an exception.
558
559 // Note that SP is restored to sender's sp (in the delay slot). This
560 // is necessary if the sender's frame is an extended compiled frame
561 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
562 // adaptations.
563
564 // Note also that the restored frame is not necessarily interpreted.
565 // Use the shared runtime version of the StackOverflowError.
566 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
567 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
568 __ jump_to(stub, Rscratch);
569 __ delayed()->mov(O5_savedSP, SP);
570
571 // If you get to here, then there is enough stack space.
572 __ bind(after_frame_check);
573 }
574
575
576 //
577 // Generate a fixed interpreter frame. This is identical setup for interpreted
578 // methods and for native methods hence the shared code.
579
580
581 //----------------------------------------------------------------------------------------------------
582 // Stack frame layout
583 //
584 // When control flow reaches any of the entry types for the interpreter
585 // the following holds ->
586 //
587 // C2 Calling Conventions:
588 //
589 // The entry code below assumes that the following registers are set
590 // when coming in:
591 // G5_method: holds the Method* of the method to call
592 // Lesp: points to the TOS of the callers expression stack
593 // after having pushed all the parameters
594 //
595 // The entry code does the following to setup an interpreter frame
596 // pop parameters from the callers stack by adjusting Lesp
597 // set O0 to Lesp
598 // compute X = (max_locals - num_parameters)
599 // bump SP up by X to accomadate the extra locals
600 // compute X = max_expression_stack
601 // + vm_local_words
602 // + 16 words of register save area
603 // save frame doing a save sp, -X, sp growing towards lower addresses
604 // set Lbcp, Lmethod, LcpoolCache
605 // set Llocals to i0
606 // set Lmonitors to FP - rounded_vm_local_words
607 // set Lesp to Lmonitors - 4
608 //
609 // The frame has now been setup to do the rest of the entry code
610
611 // Try this optimization: Most method entries could live in a
612 // "one size fits all" stack frame without all the dynamic size
613 // calculations. It might be profitable to do all this calculation
614 // statically and approximately for "small enough" methods.
615
616 //-----------------------------------------------------------------------------------------------
617
618 // C1 Calling conventions
619 //
620 // Upon method entry, the following registers are setup:
621 //
622 // g2 G2_thread: current thread
623 // g5 G5_method: method to activate
624 // g4 Gargs : pointer to last argument
625 //
626 //
627 // Stack:
628 //
629 // +---------------+ <--- sp
630 // | |
631 // : reg save area :
632 // | |
633 // +---------------+ <--- sp + 0x40
634 // | |
635 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
636 // | |
637 // +---------------+ <--- sp + 0x5c
638 // | |
639 // : free :
640 // | |
641 // +---------------+ <--- Gargs
642 // | |
643 // : arguments :
644 // | |
645 // +---------------+
646 // | |
647 //
648 //
649 //
650 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
651 //
652 // +---------------+ <--- sp
653 // | |
654 // : reg save area :
655 // | |
656 // +---------------+ <--- sp + 0x40
657 // | |
658 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
659 // | |
660 // +---------------+ <--- sp + 0x5c
661 // | |
662 // : :
663 // | | <--- Lesp
664 // +---------------+ <--- Lmonitors (fp - 0x18)
665 // | VM locals |
666 // +---------------+ <--- fp
667 // | |
668 // : reg save area :
669 // | |
670 // +---------------+ <--- fp + 0x40
671 // | |
672 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
673 // | |
674 // +---------------+ <--- fp + 0x5c
675 // | |
676 // : free :
677 // | |
678 // +---------------+
679 // | |
680 // : nonarg locals :
681 // | |
682 // +---------------+
683 // | |
684 // : arguments :
685 // | | <--- Llocals
686 // +---------------+ <--- Gargs
687 // | |
688
689 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
690 //
691 //
692 // The entry code sets up a new interpreter frame in 4 steps:
693 //
694 // 1) Increase caller's SP by for the extra local space needed:
695 // (check for overflow)
696 // Efficient implementation of xload/xstore bytecodes requires
697 // that arguments and non-argument locals are in a contigously
698 // addressable memory block => non-argument locals must be
699 // allocated in the caller's frame.
700 //
701 // 2) Create a new stack frame and register window:
702 // The new stack frame must provide space for the standard
703 // register save area, the maximum java expression stack size,
704 // the monitor slots (0 slots initially), and some frame local
705 // scratch locations.
706 //
707 // 3) The following interpreter activation registers must be setup:
708 // Lesp : expression stack pointer
709 // Lbcp : bytecode pointer
710 // Lmethod : method
711 // Llocals : locals pointer
712 // Lmonitors : monitor pointer
713 // LcpoolCache: constant pool cache
714 //
715 // 4) Initialize the non-argument locals if necessary:
716 // Non-argument locals may need to be initialized to NULL
717 // for GC to work. If the oop-map information is accurate
718 // (in the absence of the JSR problem), no initialization
719 // is necessary.
720 //
721 // (gri - 2/25/2000)
722
723
724 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
725
726 const int extra_space =
727 rounded_vm_local_words + // frame local scratch space
728 Method::extra_stack_entries() + // extra stack for jsr 292
729 frame::memory_parameter_word_sp_offset + // register save area
730 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
731
732 const Register Glocals_size = G3;
733 const Register RconstMethod = Glocals_size;
734 const Register Otmp1 = O3;
735 const Register Otmp2 = O4;
736 // Lscratch can't be used as a temporary because the call_stub uses
737 // it to assert that the stack frame was setup correctly.
738 const Address constMethod (G5_method, Method::const_offset());
739 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
740
741 __ ld_ptr( constMethod, RconstMethod );
742 __ lduh( size_of_parameters, Glocals_size);
743
744 // Gargs points to first local + BytesPerWord
745 // Set the saved SP after the register window save
746 //
747 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
748 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
749 __ add(Gargs, Otmp1, Gargs);
750
751 if (native_call) {
752 __ calc_mem_param_words( Glocals_size, Gframe_size );
753 __ add( Gframe_size, extra_space, Gframe_size);
754 __ round_to( Gframe_size, WordsPerLong );
755 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
756
757 // Native calls don't need the stack size check since they have no
758 // expression stack and the arguments are already on the stack and
759 // we only add a handful of words to the stack.
760 } else {
761
762 //
763 // Compute number of locals in method apart from incoming parameters
764 //
765 const Address size_of_locals(Otmp1, ConstMethod::size_of_locals_offset());
766 __ ld_ptr(constMethod, Otmp1);
767 __ lduh(size_of_locals, Otmp1);
768 __ sub(Otmp1, Glocals_size, Glocals_size);
769 __ round_to(Glocals_size, WordsPerLong);
770 __ sll(Glocals_size, Interpreter::logStackElementSize, Glocals_size);
771
772 // See if the frame is greater than one page in size. If so,
773 // then we need to verify there is enough stack space remaining.
774 // Frame_size = (max_stack + extra_space) * BytesPerWord;
775 __ ld_ptr(constMethod, Gframe_size);
776 __ lduh(Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size);
777 __ add(Gframe_size, extra_space, Gframe_size);
778 __ round_to(Gframe_size, WordsPerLong);
779 __ sll(Gframe_size, Interpreter::logStackElementSize, Gframe_size);
780
781 // Add in java locals size for stack overflow check only
782 __ add(Gframe_size, Glocals_size, Gframe_size);
783
784 const Register Otmp2 = O4;
785 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
786 generate_stack_overflow_check(Gframe_size, Otmp1);
787
788 __ sub(Gframe_size, Glocals_size, Gframe_size);
789
790 //
791 // bump SP to accomodate the extra locals
792 //
793 __ sub(SP, Glocals_size, SP);
794 }
795
796 //
797 // now set up a stack frame with the size computed above
798 //
799 __ neg( Gframe_size );
800 __ save( SP, Gframe_size, SP );
801
802 //
803 // now set up all the local cache registers
804 //
805 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
806 // that all present references to Lbyte_code initialize the register
807 // immediately before use
808 if (native_call) {
809 __ mov(G0, Lbcp);
810 } else {
811 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
812 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
813 }
814 __ mov( G5_method, Lmethod); // set Lmethod
815 // Get mirror and store it in the frame as GC root for this Method*
816 Register mirror = LcpoolCache;
817 __ load_mirror(mirror, Lmethod);
818 __ st_ptr(mirror, FP, (frame::interpreter_frame_mirror_offset * wordSize) + STACK_BIAS);
819 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
820 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
821 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
822 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
823
824 // setup interpreter activation registers
825 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
826
827 if (ProfileInterpreter) {
828 #ifdef FAST_DISPATCH
829 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
830 // they both use I2.
831 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
832 #endif // FAST_DISPATCH
833 __ set_method_data_pointer();
834 }
835
836 }
837
838 // Method entry for java.lang.ref.Reference.get.
839 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
840 // Code: _aload_0, _getfield, _areturn
841 // parameter size = 1
842 //
843 // The code that gets generated by this routine is split into 2 parts:
844 // 1. The "intrinsified" code for G1 (or any SATB based GC),
845 // 2. The slow path - which is an expansion of the regular method entry.
846 //
847 // Notes:-
848 // * In the G1 code we do not check whether we need to block for
849 // a safepoint. If G1 is enabled then we must execute the specialized
850 // code for Reference.get (except when the Reference object is null)
851 // so that we can log the value in the referent field with an SATB
852 // update buffer.
853 // If the code for the getfield template is modified so that the
854 // G1 pre-barrier code is executed when the current method is
855 // Reference.get() then going through the normal method entry
856 // will be fine.
857 // * The G1 code can, however, check the receiver object (the instance
858 // of java.lang.Reference) and jump to the slow path if null. If the
859 // Reference object is null then we obviously cannot fetch the referent
860 // and so we don't need to call the G1 pre-barrier. Thus we can use the
861 // regular method entry code to generate the NPE.
862 //
863 // This code is based on generate_accessor_enty.
864
865 address entry = __ pc();
866
867 const int referent_offset = java_lang_ref_Reference::referent_offset;
868 guarantee(referent_offset > 0, "referent offset not initialized");
869
870 Label slow_path;
871
872 // In the G1 code we don't check if we need to reach a safepoint. We
873 // continue and the thread will safepoint at the next bytecode dispatch.
874
875 // Check if local 0 != NULL
876 // If the receiver is null then it is OK to jump to the slow path.
877 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
878 // check if local 0 == NULL and go the slow path
879 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
880
881 BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->code_gen();
882 code_gen->load_at(_masm, ACCESS_ON_HEAP | GC_ACCESS_ON_WEAK, T_OBJECT,
883 Otos_i, noreg, referent_offset, Otos_i, G3_scratch);
884
885 // _areturn
886 __ retl(); // return from leaf routine
887 __ delayed()->mov(O5_savedSP, SP);
888
889 // Generate regular method entry
890 __ bind(slow_path);
891 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
892 return entry;
893 }
894
895 /**
896 * Method entry for static native methods:
897 * int java.util.zip.CRC32.update(int crc, int b)
898 */
899 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
900
901 if (UseCRC32Intrinsics) {
902 address entry = __ pc();
903
904 Label L_slow_path;
905 // If we need a safepoint check, generate full interpreter entry.
906 ExternalAddress state(SafepointSynchronize::address_of_state());
907 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
908 __ set(SafepointSynchronize::_not_synchronized, O3);
909 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
910
911 // Load parameters
912 const Register crc = O0; // initial crc
913 const Register val = O1; // byte to update with
914 const Register table = O2; // address of 256-entry lookup table
915
916 __ ldub(Gargs, 3, val);
917 __ lduw(Gargs, 8, crc);
918
919 __ set(ExternalAddress(StubRoutines::crc_table_addr()), table);
920
921 __ not1(crc); // ~crc
922 __ clruwu(crc);
923 __ update_byte_crc32(crc, val, table);
924 __ not1(crc); // ~crc
925
926 // result in O0
927 __ retl();
928 __ delayed()->nop();
929
930 // generate a vanilla native entry as the slow path
931 __ bind(L_slow_path);
932 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
933 return entry;
934 }
935 return NULL;
936 }
937
938 /**
939 * Method entry for static native methods:
940 * int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len)
941 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
942 */
943 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
944
945 if (UseCRC32Intrinsics) {
946 address entry = __ pc();
947
948 Label L_slow_path;
949 // If we need a safepoint check, generate full interpreter entry.
950 ExternalAddress state(SafepointSynchronize::address_of_state());
951 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
952 __ set(SafepointSynchronize::_not_synchronized, O3);
953 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
954
955 // Load parameters from the stack
956 const Register crc = O0; // initial crc
957 const Register buf = O1; // source java byte array address
958 const Register len = O2; // len
959 const Register offset = O3; // offset
960
961 // Arguments are reversed on java expression stack
962 // Calculate address of start element
963 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) {
964 __ lduw(Gargs, 0, len);
965 __ lduw(Gargs, 8, offset);
966 __ ldx( Gargs, 16, buf);
967 __ lduw(Gargs, 32, crc);
968 __ add(buf, offset, buf);
969 } else {
970 __ lduw(Gargs, 0, len);
971 __ lduw(Gargs, 8, offset);
972 __ ldx( Gargs, 16, buf);
973 __ lduw(Gargs, 24, crc);
974 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
975 __ add(buf ,offset, buf);
976 }
977
978 // Call the crc32 kernel
979 __ MacroAssembler::save_thread(L7_thread_cache);
980 __ kernel_crc32(crc, buf, len, O3);
981 __ MacroAssembler::restore_thread(L7_thread_cache);
982
983 // result in O0
984 __ retl();
985 __ delayed()->nop();
986
987 // generate a vanilla native entry as the slow path
988 __ bind(L_slow_path);
989 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
990 return entry;
991 }
992 return NULL;
993 }
994
995 /**
996 * Method entry for intrinsic-candidate (non-native) methods:
997 * int java.util.zip.CRC32C.updateBytes(int crc, byte[] b, int off, int end)
998 * int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
999 * Unlike CRC32, CRC32C does not have any methods marked as native
1000 * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
1001 */
1002 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1003
1004 if (UseCRC32CIntrinsics) {
1005 address entry = __ pc();
1006
1007 // Load parameters from the stack
1008 const Register crc = O0; // initial crc
1009 const Register buf = O1; // source java byte array address
1010 const Register offset = O2; // offset
1011 const Register end = O3; // index of last element to process
1012 const Register len = O2; // len argument to the kernel
1013 const Register table = O3; // crc32c lookup table address
1014
1015 // Arguments are reversed on java expression stack
1016 // Calculate address of start element
1017 if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) {
1018 __ lduw(Gargs, 0, end);
1019 __ lduw(Gargs, 8, offset);
1020 __ ldx( Gargs, 16, buf);
1021 __ lduw(Gargs, 32, crc);
1022 __ add(buf, offset, buf);
1023 __ sub(end, offset, len);
1024 } else {
1025 __ lduw(Gargs, 0, end);
1026 __ lduw(Gargs, 8, offset);
1027 __ ldx( Gargs, 16, buf);
1028 __ lduw(Gargs, 24, crc);
1029 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1030 __ add(buf, offset, buf);
1031 __ sub(end, offset, len);
1032 }
1033
1034 // Call the crc32c kernel
1035 __ MacroAssembler::save_thread(L7_thread_cache);
1036 __ kernel_crc32c(crc, buf, len, table);
1037 __ MacroAssembler::restore_thread(L7_thread_cache);
1038
1039 // result in O0
1040 __ retl();
1041 __ delayed()->nop();
1042
1043 return entry;
1044 }
1045 return NULL;
1046 }
1047
1048 // Not supported
1049 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
1050 return NULL;
1051 }
1052
1053 // TODO: rather than touching all pages, check against stack_overflow_limit and bang yellow page to
1054 // generate exception
1055 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1056 // Quick & dirty stack overflow checking: bang the stack & handle trap.
1057 // Note that we do the banging after the frame is setup, since the exception
1058 // handling code expects to find a valid interpreter frame on the stack.
1059 // Doing the banging earlier fails if the caller frame is not an interpreter
1060 // frame.
1061 // (Also, the exception throwing code expects to unlock any synchronized
1062 // method receiever, so do the banging after locking the receiver.)
1063
1064 // Bang each page in the shadow zone. We can't assume it's been done for
1065 // an interpreter frame with greater than a page of locals, so each page
1066 // needs to be checked. Only true for non-native.
1067 if (UseStackBanging) {
1068 const int page_size = os::vm_page_size();
1069 const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1070 const int start_page = native_call ? n_shadow_pages : 1;
1071 for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1072 __ bang_stack_with_offset(pages*page_size);
1073 }
1074 }
1075 }
1076
1077 //
1078 // Interpreter stub for calling a native method. (asm interpreter)
1079 // This sets up a somewhat different looking stack for calling the native method
1080 // than the typical interpreter frame setup.
1081 //
1082
1083 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1084 address entry = __ pc();
1085
1086 // the following temporary registers are used during frame creation
1087 const Register Gtmp1 = G3_scratch ;
1088 const Register Gtmp2 = G1_scratch;
1089 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1090
1091 // make sure registers are different!
1092 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1093
1094 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
1095
1096 const Register Glocals_size = G3;
1097 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1098
1099 // make sure method is native & not abstract
1100 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1101 #ifdef ASSERT
1102 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1103 {
1104 Label L;
1105 __ btst(JVM_ACC_NATIVE, Gtmp1);
1106 __ br(Assembler::notZero, false, Assembler::pt, L);
1107 __ delayed()->nop();
1108 __ stop("tried to execute non-native method as native");
1109 __ bind(L);
1110 }
1111 { Label L;
1112 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1113 __ br(Assembler::zero, false, Assembler::pt, L);
1114 __ delayed()->nop();
1115 __ stop("tried to execute abstract method as non-abstract");
1116 __ bind(L);
1117 }
1118 #endif // ASSERT
1119
1120 // generate the code to allocate the interpreter stack frame
1121 generate_fixed_frame(true);
1122
1123 //
1124 // No locals to initialize for native method
1125 //
1126
1127 // this slot will be set later, we initialize it to null here just in
1128 // case we get a GC before the actual value is stored later
1129 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
1130
1131 const Address do_not_unlock_if_synchronized(G2_thread,
1132 JavaThread::do_not_unlock_if_synchronized_offset());
1133 // Since at this point in the method invocation the exception handler
1134 // would try to exit the monitor of synchronized methods which hasn't
1135 // been entered yet, we set the thread local variable
1136 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1137 // runtime, exception handling i.e. unlock_if_synchronized_method will
1138 // check this thread local flag.
1139 // This flag has two effects, one is to force an unwind in the topmost
1140 // interpreter frame and not perform an unlock while doing so.
1141
1142 __ movbool(true, G3_scratch);
1143 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1144
1145 // increment invocation counter and check for overflow
1146 //
1147 // Note: checking for negative value instead of overflow
1148 // so we have a 'sticky' overflow test (may be of
1149 // importance as soon as we have true MT/MP)
1150 Label invocation_counter_overflow;
1151 Label Lcontinue;
1152 if (inc_counter) {
1153 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1154
1155 }
1156 __ bind(Lcontinue);
1157
1158 bang_stack_shadow_pages(true);
1159
1160 // reset the _do_not_unlock_if_synchronized flag
1161 __ stbool(G0, do_not_unlock_if_synchronized);
1162
1163 // check for synchronized methods
1164 // Must happen AFTER invocation_counter check and stack overflow check,
1165 // so method is not locked if overflows.
1166
1167 if (synchronized) {
1168 lock_method();
1169 } else {
1170 #ifdef ASSERT
1171 { Label ok;
1172 __ ld(Laccess_flags, O0);
1173 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1174 __ br( Assembler::zero, false, Assembler::pt, ok);
1175 __ delayed()->nop();
1176 __ stop("method needs synchronization");
1177 __ bind(ok);
1178 }
1179 #endif // ASSERT
1180 }
1181
1182
1183 // start execution
1184 __ verify_thread();
1185
1186 // JVMTI support
1187 __ notify_method_entry();
1188
1189 // native call
1190
1191 // (note that O0 is never an oop--at most it is a handle)
1192 // It is important not to smash any handles created by this call,
1193 // until any oop handle in O0 is dereferenced.
1194
1195 // (note that the space for outgoing params is preallocated)
1196
1197 // get signature handler
1198 { Label L;
1199 Address signature_handler(Lmethod, Method::signature_handler_offset());
1200 __ ld_ptr(signature_handler, G3_scratch);
1201 __ br_notnull_short(G3_scratch, Assembler::pt, L);
1202 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
1203 __ ld_ptr(signature_handler, G3_scratch);
1204 __ bind(L);
1205 }
1206
1207 // Push a new frame so that the args will really be stored in
1208 // Copy a few locals across so the new frame has the variables
1209 // we need but these values will be dead at the jni call and
1210 // therefore not gc volatile like the values in the current
1211 // frame (Lmethod in particular)
1212
1213 // Flush the method pointer to the register save area
1214 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
1215 __ mov(Llocals, O1);
1216
1217 // calculate where the mirror handle body is allocated in the interpreter frame:
1218 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
1219
1220 // Calculate current frame size
1221 __ sub(SP, FP, O3); // Calculate negative of current frame size
1222 __ save(SP, O3, SP); // Allocate an identical sized frame
1223
1224 // Note I7 has leftover trash. Slow signature handler will fill it in
1225 // should we get there. Normal jni call will set reasonable last_Java_pc
1226 // below (and fix I7 so the stack trace doesn't have a meaningless frame
1227 // in it).
1228
1229 // Load interpreter frame's Lmethod into same register here
1230
1231 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1232
1233 __ mov(I1, Llocals);
1234 __ mov(I2, Lscratch2); // save the address of the mirror
1235
1236
1237 // ONLY Lmethod and Llocals are valid here!
1238
1239 // call signature handler, It will move the arg properly since Llocals in current frame
1240 // matches that in outer frame
1241
1242 __ callr(G3_scratch, 0);
1243 __ delayed()->nop();
1244
1245 // Result handler is in Lscratch
1246
1247 // Reload interpreter frame's Lmethod since slow signature handler may block
1248 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1249
1250 { Label not_static;
1251
1252 __ ld(Laccess_flags, O0);
1253 __ btst(JVM_ACC_STATIC, O0);
1254 __ br( Assembler::zero, false, Assembler::pt, not_static);
1255 // get native function entry point(O0 is a good temp until the very end)
1256 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
1257 // for static methods insert the mirror argument
1258 __ load_mirror(O1, Lmethod);
1259 #ifdef ASSERT
1260 if (!PrintSignatureHandlers) // do not dirty the output with this
1261 { Label L;
1262 __ br_notnull_short(O1, Assembler::pt, L);
1263 __ stop("mirror is missing");
1264 __ bind(L);
1265 }
1266 #endif // ASSERT
1267 __ st_ptr(O1, Lscratch2, 0);
1268 __ mov(Lscratch2, O1);
1269 __ bind(not_static);
1270 }
1271
1272 // At this point, arguments have been copied off of stack into
1273 // their JNI positions, which are O1..O5 and SP[68..].
1274 // Oops are boxed in-place on the stack, with handles copied to arguments.
1275 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1276
1277 #ifdef ASSERT
1278 { Label L;
1279 __ br_notnull_short(O0, Assembler::pt, L);
1280 __ stop("native entry point is missing");
1281 __ bind(L);
1282 }
1283 #endif // ASSERT
1284
1285 //
1286 // setup the frame anchor
1287 //
1288 // The scavenge function only needs to know that the PC of this frame is
1289 // in the interpreter method entry code, it doesn't need to know the exact
1290 // PC and hence we can use O7 which points to the return address from the
1291 // previous call in the code stream (signature handler function)
1292 //
1293 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1294 // we have pushed the extra frame in order to protect the volatile register(s)
1295 // in that frame when we return from the jni call
1296 //
1297
1298 __ set_last_Java_frame(FP, O7);
1299 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1300 // not meaningless information that'll confuse me.
1301
1302 // flush the windows now. We don't care about the current (protection) frame
1303 // only the outer frames
1304
1305 __ flushw();
1306
1307 // mark windows as flushed
1308 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1309 __ set(JavaFrameAnchor::flushed, G3_scratch);
1310 __ st(G3_scratch, flags);
1311
1312 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1313
1314 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1315 #ifdef ASSERT
1316 { Label L;
1317 __ ld(thread_state, G3_scratch);
1318 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1319 __ stop("Wrong thread state in native stub");
1320 __ bind(L);
1321 }
1322 #endif // ASSERT
1323 __ set(_thread_in_native, G3_scratch);
1324 __ st(G3_scratch, thread_state);
1325
1326 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1327 __ save_thread(L7_thread_cache); // save Gthread
1328 __ callr(O0, 0);
1329 __ delayed()->
1330 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1331
1332 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1333
1334 __ restore_thread(L7_thread_cache); // restore G2_thread
1335 __ reinit_heapbase();
1336
1337 // must we block?
1338
1339 // Block, if necessary, before resuming in _thread_in_Java state.
1340 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1341 { Label no_block;
1342 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1343
1344 // Switch thread to "native transition" state before reading the synchronization state.
1345 // This additional state is necessary because reading and testing the synchronization
1346 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1347 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1348 // VM thread changes sync state to synchronizing and suspends threads for GC.
1349 // Thread A is resumed to finish this native method, but doesn't block here since it
1350 // didn't see any synchronization is progress, and escapes.
1351 __ set(_thread_in_native_trans, G3_scratch);
1352 __ st(G3_scratch, thread_state);
1353 if(os::is_MP()) {
1354 if (UseMembar) {
1355 // Force this write out before the read below
1356 __ membar(Assembler::StoreLoad);
1357 } else {
1358 // Write serialization page so VM thread can do a pseudo remote membar.
1359 // We use the current thread pointer to calculate a thread specific
1360 // offset to write to within the page. This minimizes bus traffic
1361 // due to cache line collision.
1362 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1363 }
1364 }
1365 __ load_contents(sync_state, G3_scratch);
1366 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1367
1368 Label L;
1369 __ br(Assembler::notEqual, false, Assembler::pn, L);
1370 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1371 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1372 __ bind(L);
1373
1374 // Block. Save any potential method result value before the operation and
1375 // use a leaf call to leave the last_Java_frame setup undisturbed.
1376 save_native_result();
1377 __ call_VM_leaf(L7_thread_cache,
1378 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1379 G2_thread);
1380
1381 // Restore any method result value
1382 restore_native_result();
1383 __ bind(no_block);
1384 }
1385
1386 // Clear the frame anchor now
1387
1388 __ reset_last_Java_frame();
1389
1390 // Move the result handler address
1391 __ mov(Lscratch, G3_scratch);
1392 // return possible result to the outer frame
1393 __ restore(O0, G0, O0);
1394
1395 // Move result handler to expected register
1396 __ mov(G3_scratch, Lscratch);
1397
1398 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1399 // switch to thread_in_Java.
1400
1401 __ set(_thread_in_Java, G3_scratch);
1402 __ st(G3_scratch, thread_state);
1403
1404 if (CheckJNICalls) {
1405 // clear_pending_jni_exception_check
1406 __ st_ptr(G0, G2_thread, JavaThread::pending_jni_exception_check_fn_offset());
1407 }
1408
1409 // reset handle block
1410 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1411 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1412
1413 // If we have an oop result store it where it will be safe for any further gc
1414 // until we return now that we've released the handle it might be protected by
1415
1416 {
1417 Label no_oop;
1418
1419 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1420 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1421 __ resolve_jobject(O0, G3_scratch);
1422 // Store it where gc will look for it and result handler expects it.
1423 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1424
1425 __ bind(no_oop);
1426 }
1427
1428
1429 // handle exceptions (exception handling will handle unlocking!)
1430 { Label L;
1431 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1432 __ ld_ptr(exception_addr, Gtemp);
1433 __ br_null_short(Gtemp, Assembler::pt, L);
1434 // Note: This could be handled more efficiently since we know that the native
1435 // method doesn't have an exception handler. We could directly return
1436 // to the exception handler for the caller.
1437 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1438 __ should_not_reach_here();
1439 __ bind(L);
1440 }
1441
1442 // JVMTI support (preserves thread register)
1443 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1444
1445 if (synchronized) {
1446 // save and restore any potential method result value around the unlocking operation
1447 save_native_result();
1448
1449 __ add( __ top_most_monitor(), O1);
1450 __ unlock_object(O1);
1451
1452 restore_native_result();
1453 }
1454
1455 // dispose of return address and remove activation
1456 #ifdef ASSERT
1457 {
1458 Label ok;
1459 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1460 __ stop("bad I5_savedSP value");
1461 __ should_not_reach_here();
1462 __ bind(ok);
1463 }
1464 #endif
1465 __ jmp(Lscratch, 0);
1466 __ delayed()->nop();
1467
1468
1469 if (inc_counter) {
1470 // handle invocation counter overflow
1471 __ bind(invocation_counter_overflow);
1472 generate_counter_overflow(Lcontinue);
1473 }
1474
1475
1476
1477 return entry;
1478 }
1479
1480
1481 // Generic method entry to (asm) interpreter
1482 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1483 address entry = __ pc();
1484
1485 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1486
1487 // the following temporary registers are used during frame creation
1488 const Register Gtmp1 = G3_scratch ;
1489 const Register Gtmp2 = G1_scratch;
1490
1491 // make sure registers are different!
1492 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1493
1494 const Address constMethod (G5_method, Method::const_offset());
1495 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1496 // and use in the asserts.
1497 const Address access_flags (Lmethod, Method::access_flags_offset());
1498
1499 const Register Glocals_size = G3;
1500 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1501
1502 // make sure method is not native & not abstract
1503 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1504 #ifdef ASSERT
1505 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1506 {
1507 Label L;
1508 __ btst(JVM_ACC_NATIVE, Gtmp1);
1509 __ br(Assembler::zero, false, Assembler::pt, L);
1510 __ delayed()->nop();
1511 __ stop("tried to execute native method as non-native");
1512 __ bind(L);
1513 }
1514 { Label L;
1515 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1516 __ br(Assembler::zero, false, Assembler::pt, L);
1517 __ delayed()->nop();
1518 __ stop("tried to execute abstract method as non-abstract");
1519 __ bind(L);
1520 }
1521 #endif // ASSERT
1522
1523 // generate the code to allocate the interpreter stack frame
1524
1525 generate_fixed_frame(false);
1526
1527 #ifdef FAST_DISPATCH
1528 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1529 // set bytecode dispatch table base
1530 #endif
1531
1532 //
1533 // Code to initialize the extra (i.e. non-parm) locals
1534 //
1535 Register init_value = noreg; // will be G0 if we must clear locals
1536 // The way the code was setup before zerolocals was always true for vanilla java entries.
1537 // It could only be false for the specialized entries like accessor or empty which have
1538 // no extra locals so the testing was a waste of time and the extra locals were always
1539 // initialized. We removed this extra complication to already over complicated code.
1540
1541 init_value = G0;
1542 Label clear_loop;
1543
1544 const Register RconstMethod = O1;
1545 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1546 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1547
1548 // NOTE: If you change the frame layout, this code will need to
1549 // be updated!
1550 __ ld_ptr( constMethod, RconstMethod );
1551 __ lduh( size_of_locals, O2 );
1552 __ lduh( size_of_parameters, O1 );
1553 __ sll( O2, Interpreter::logStackElementSize, O2);
1554 __ sll( O1, Interpreter::logStackElementSize, O1 );
1555 __ sub( Llocals, O2, O2 );
1556 __ sub( Llocals, O1, O1 );
1557
1558 __ bind( clear_loop );
1559 __ inc( O2, wordSize );
1560
1561 __ cmp( O2, O1 );
1562 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1563 __ delayed()->st_ptr( init_value, O2, 0 );
1564
1565 const Address do_not_unlock_if_synchronized(G2_thread,
1566 JavaThread::do_not_unlock_if_synchronized_offset());
1567 // Since at this point in the method invocation the exception handler
1568 // would try to exit the monitor of synchronized methods which hasn't
1569 // been entered yet, we set the thread local variable
1570 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1571 // runtime, exception handling i.e. unlock_if_synchronized_method will
1572 // check this thread local flag.
1573 __ movbool(true, G3_scratch);
1574 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1575
1576 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1577 // increment invocation counter and check for overflow
1578 //
1579 // Note: checking for negative value instead of overflow
1580 // so we have a 'sticky' overflow test (may be of
1581 // importance as soon as we have true MT/MP)
1582 Label invocation_counter_overflow;
1583 Label profile_method;
1584 Label profile_method_continue;
1585 Label Lcontinue;
1586 if (inc_counter) {
1587 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1588 if (ProfileInterpreter) {
1589 __ bind(profile_method_continue);
1590 }
1591 }
1592 __ bind(Lcontinue);
1593
1594 bang_stack_shadow_pages(false);
1595
1596 // reset the _do_not_unlock_if_synchronized flag
1597 __ stbool(G0, do_not_unlock_if_synchronized);
1598
1599 // check for synchronized methods
1600 // Must happen AFTER invocation_counter check and stack overflow check,
1601 // so method is not locked if overflows.
1602
1603 if (synchronized) {
1604 lock_method();
1605 } else {
1606 #ifdef ASSERT
1607 { Label ok;
1608 __ ld(access_flags, O0);
1609 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1610 __ br( Assembler::zero, false, Assembler::pt, ok);
1611 __ delayed()->nop();
1612 __ stop("method needs synchronization");
1613 __ bind(ok);
1614 }
1615 #endif // ASSERT
1616 }
1617
1618 // start execution
1619
1620 __ verify_thread();
1621
1622 // jvmti support
1623 __ notify_method_entry();
1624
1625 // start executing instructions
1626 __ dispatch_next(vtos);
1627
1628
1629 if (inc_counter) {
1630 if (ProfileInterpreter) {
1631 // We have decided to profile this method in the interpreter
1632 __ bind(profile_method);
1633
1634 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1635 __ set_method_data_pointer_for_bcp();
1636 __ ba_short(profile_method_continue);
1637 }
1638
1639 // handle invocation counter overflow
1640 __ bind(invocation_counter_overflow);
1641 generate_counter_overflow(Lcontinue);
1642 }
1643
1644
1645 return entry;
1646 }
1647
1648 //----------------------------------------------------------------------------------------------------
1649 // Exceptions
1650 void TemplateInterpreterGenerator::generate_throw_exception() {
1651
1652 // Entry point in previous activation (i.e., if the caller was interpreted)
1653 Interpreter::_rethrow_exception_entry = __ pc();
1654 // O0: exception
1655
1656 // entry point for exceptions thrown within interpreter code
1657 Interpreter::_throw_exception_entry = __ pc();
1658 __ verify_thread();
1659 // expression stack is undefined here
1660 // O0: exception, i.e. Oexception
1661 // Lbcp: exception bcp
1662 __ verify_oop(Oexception);
1663
1664
1665 // expression stack must be empty before entering the VM in case of an exception
1666 __ empty_expression_stack();
1667 // find exception handler address and preserve exception oop
1668 // call C routine to find handler and jump to it
1669 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1670 __ push_ptr(O1); // push exception for exception handler bytecodes
1671
1672 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1673 __ delayed()->nop();
1674
1675
1676 // if the exception is not handled in the current frame
1677 // the frame is removed and the exception is rethrown
1678 // (i.e. exception continuation is _rethrow_exception)
1679 //
1680 // Note: At this point the bci is still the bxi for the instruction which caused
1681 // the exception and the expression stack is empty. Thus, for any VM calls
1682 // at this point, GC will find a legal oop map (with empty expression stack).
1683
1684 // in current activation
1685 // tos: exception
1686 // Lbcp: exception bcp
1687
1688 //
1689 // JVMTI PopFrame support
1690 //
1691
1692 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1693 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1694 // Set the popframe_processing bit in popframe_condition indicating that we are
1695 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1696 // popframe handling cycles.
1697
1698 __ ld(popframe_condition_addr, G3_scratch);
1699 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1700 __ stw(G3_scratch, popframe_condition_addr);
1701
1702 // Empty the expression stack, as in normal exception handling
1703 __ empty_expression_stack();
1704 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1705
1706 {
1707 // Check to see whether we are returning to a deoptimized frame.
1708 // (The PopFrame call ensures that the caller of the popped frame is
1709 // either interpreted or compiled and deoptimizes it if compiled.)
1710 // In this case, we can't call dispatch_next() after the frame is
1711 // popped, but instead must save the incoming arguments and restore
1712 // them after deoptimization has occurred.
1713 //
1714 // Note that we don't compare the return PC against the
1715 // deoptimization blob's unpack entry because of the presence of
1716 // adapter frames in C2.
1717 Label caller_not_deoptimized;
1718 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1719 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1720
1721 const Register Gtmp1 = G3_scratch;
1722 const Register Gtmp2 = G1_scratch;
1723 const Register RconstMethod = Gtmp1;
1724 const Address constMethod(Lmethod, Method::const_offset());
1725 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1726
1727 // Compute size of arguments for saving when returning to deoptimized caller
1728 __ ld_ptr(constMethod, RconstMethod);
1729 __ lduh(size_of_parameters, Gtmp1);
1730 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1731 __ sub(Llocals, Gtmp1, Gtmp2);
1732 __ add(Gtmp2, wordSize, Gtmp2);
1733 // Save these arguments
1734 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1735 // Inform deoptimization that it is responsible for restoring these arguments
1736 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1737 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1738 __ st(Gtmp1, popframe_condition_addr);
1739
1740 // Return from the current method
1741 // The caller's SP was adjusted upon method entry to accomodate
1742 // the callee's non-argument locals. Undo that adjustment.
1743 __ ret();
1744 __ delayed()->restore(I5_savedSP, G0, SP);
1745
1746 __ bind(caller_not_deoptimized);
1747 }
1748
1749 // Clear the popframe condition flag
1750 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1751
1752 // Get out of the current method (how this is done depends on the particular compiler calling
1753 // convention that the interpreter currently follows)
1754 // The caller's SP was adjusted upon method entry to accomodate
1755 // the callee's non-argument locals. Undo that adjustment.
1756 __ restore(I5_savedSP, G0, SP);
1757 // The method data pointer was incremented already during
1758 // call profiling. We have to restore the mdp for the current bcp.
1759 if (ProfileInterpreter) {
1760 __ set_method_data_pointer_for_bcp();
1761 }
1762
1763 #if INCLUDE_JVMTI
1764 {
1765 Label L_done;
1766
1767 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1768 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1769
1770 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1771 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1772
1773 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1774
1775 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1776 __ delayed()->nop();
1777
1778 __ st_ptr(G1_scratch, Lesp, wordSize);
1779 __ bind(L_done);
1780 }
1781 #endif // INCLUDE_JVMTI
1782
1783 // Resume bytecode interpretation at the current bcp
1784 __ dispatch_next(vtos);
1785 // end of JVMTI PopFrame support
1786
1787 Interpreter::_remove_activation_entry = __ pc();
1788
1789 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1790 __ pop_ptr(Oexception); // get exception
1791
1792 // Intel has the following comment:
1793 //// remove the activation (without doing throws on illegalMonitorExceptions)
1794 // They remove the activation without checking for bad monitor state.
1795 // %%% We should make sure this is the right semantics before implementing.
1796
1797 __ set_vm_result(Oexception);
1798 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1799
1800 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1801
1802 __ get_vm_result(Oexception);
1803 __ verify_oop(Oexception);
1804
1805 const int return_reg_adjustment = frame::pc_return_offset;
1806 Address issuing_pc_addr(I7, return_reg_adjustment);
1807
1808 // We are done with this activation frame; find out where to go next.
1809 // The continuation point will be an exception handler, which expects
1810 // the following registers set up:
1811 //
1812 // Oexception: exception
1813 // Oissuing_pc: the local call that threw exception
1814 // Other On: garbage
1815 // In/Ln: the contents of the caller's register window
1816 //
1817 // We do the required restore at the last possible moment, because we
1818 // need to preserve some state across a runtime call.
1819 // (Remember that the caller activation is unknown--it might not be
1820 // interpreted, so things like Lscratch are useless in the caller.)
1821
1822 // Although the Intel version uses call_C, we can use the more
1823 // compact call_VM. (The only real difference on SPARC is a
1824 // harmlessly ignored [re]set_last_Java_frame, compared with
1825 // the Intel code which lacks this.)
1826 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1827 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1828 __ super_call_VM_leaf(L7_thread_cache,
1829 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1830 G2_thread, Oissuing_pc->after_save());
1831
1832 // The caller's SP was adjusted upon method entry to accomodate
1833 // the callee's non-argument locals. Undo that adjustment.
1834 __ JMP(O0, 0); // return exception handler in caller
1835 __ delayed()->restore(I5_savedSP, G0, SP);
1836
1837 // (same old exception object is already in Oexception; see above)
1838 // Note that an "issuing PC" is actually the next PC after the call
1839 }
1840
1841
1842 //
1843 // JVMTI ForceEarlyReturn support
1844 //
1845
1846 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1847 address entry = __ pc();
1848
1849 __ empty_expression_stack();
1850 __ load_earlyret_value(state);
1851
1852 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1853 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1854
1855 // Clear the earlyret state
1856 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1857
1858 __ remove_activation(state,
1859 /* throw_monitor_exception */ false,
1860 /* install_monitor_exception */ false);
1861
1862 // The caller's SP was adjusted upon method entry to accomodate
1863 // the callee's non-argument locals. Undo that adjustment.
1864 __ ret(); // return to caller
1865 __ delayed()->restore(I5_savedSP, G0, SP);
1866
1867 return entry;
1868 } // end of JVMTI ForceEarlyReturn support
1869
1870
1871 //------------------------------------------------------------------------------------------------------------------------
1872 // Helper for vtos entry point generation
1873
1874 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) {
1875 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1876 Label L;
1877 aep = __ pc(); __ push_ptr(); __ ba_short(L);
1878 fep = __ pc(); __ push_f(); __ ba_short(L);
1879 dep = __ pc(); __ push_d(); __ ba_short(L);
1880 lep = __ pc(); __ push_l(); __ ba_short(L);
1881 iep = __ pc(); __ push_i();
1882 bep = cep = sep = iep; // there aren't any
1883 vep = __ pc(); __ bind(L); // fall through
1884 generate_and_dispatch(t);
1885 }
1886
1887 // --------------------------------------------------------------------------------
1888
1889 // Non-product code
1890 #ifndef PRODUCT
1891 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1892 address entry = __ pc();
1893
1894 __ push(state);
1895 __ mov(O7, Lscratch); // protect return address within interpreter
1896
1897 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1898 __ mov( Otos_l2, G3_scratch );
1899 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1900 __ mov(Lscratch, O7); // restore return address
1901 __ pop(state);
1902 __ retl();
1903 __ delayed()->nop();
1904
1905 return entry;
1906 }
1907
1908
1909 // helpers for generate_and_dispatch
1910
1911 void TemplateInterpreterGenerator::count_bytecode() {
1912 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1913 }
1914
1915
1916 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1917 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
1918 }
1919
1920
1921 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1922 AddressLiteral index (&BytecodePairHistogram::_index);
1923 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
1924
1925 // get index, shift out old bytecode, bring in new bytecode, and store it
1926 // _index = (_index >> log2_number_of_codes) |
1927 // (bytecode << log2_number_of_codes);
1928
1929 __ load_contents(index, G4_scratch);
1930 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
1931 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
1932 __ or3( G3_scratch, G4_scratch, G4_scratch );
1933 __ store_contents(G4_scratch, index, G3_scratch);
1934
1935 // bump bucket contents
1936 // _counters[_index] ++;
1937
1938 __ set(counters, G3_scratch); // loads into G3_scratch
1939 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
1940 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
1941 __ ld (G3_scratch, 0, G4_scratch);
1942 __ inc (G4_scratch);
1943 __ st (G4_scratch, 0, G3_scratch);
1944 }
1945
1946
1947 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
1948 // Call a little run-time stub to avoid blow-up for each bytecode.
1949 // The run-time runtime saves the right registers, depending on
1950 // the tosca in-state for the given template.
1951 address entry = Interpreter::trace_code(t->tos_in());
1952 guarantee(entry != NULL, "entry must have been generated");
1953 __ call(entry, relocInfo::none);
1954 __ delayed()->nop();
1955 }
1956
1957
1958 void TemplateInterpreterGenerator::stop_interpreter_at() {
1959 AddressLiteral counter(&BytecodeCounter::_counter_value);
1960 __ load_contents(counter, G3_scratch);
1961 AddressLiteral stop_at(&StopInterpreterAt);
1962 __ load_ptr_contents(stop_at, G4_scratch);
1963 __ cmp(G3_scratch, G4_scratch);
1964 __ breakpoint_trap(Assembler::equal, Assembler::icc);
1965 }
1966 #endif // not PRODUCT