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 "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. This can
388 // only occur on method entry so emit it only for vtos with step 0.
389 if (UseJVMCICompiler && state == vtos && step == 0) {
390 Label L;
391 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
392 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter
393 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
394 // Clear flag.
395 __ stbool(G0, pending_monitor_enter_addr);
396 // Take lock.
397 lock_method();
398 __ bind(L);
399 } else {
400 #ifdef ASSERT
401 if (UseJVMCICompiler) {
402 Label L;
403 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
404 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter
405 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
406 __ stop("unexpected pending monitor in deopt entry");
407 __ bind(L);
408 }
409 #endif
410 }
411 #endif
412 { Label L;
413 Address exception_addr(G2_thread, Thread::pending_exception_offset());
414 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
415 __ br_null_short(Gtemp, Assembler::pt, L);
416 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
417 __ should_not_reach_here();
418 __ bind(L);
419 }
420 __ dispatch_next(state, step);
421 return entry;
422 }
423
424 // A result handler converts/unboxes a native call result into
425 // a java interpreter/compiler result. The current frame is an
426 // interpreter frame. The activation frame unwind code must be
427 // consistent with that of TemplateTable::_return(...). In the
428 // case of native methods, the caller's SP was not modified.
429 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
430 address entry = __ pc();
431 Register Itos_i = Otos_i ->after_save();
432 Register Itos_l = Otos_l ->after_save();
433 Register Itos_l1 = Otos_l1->after_save();
434 Register Itos_l2 = Otos_l2->after_save();
435 switch (type) {
436 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
437 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
438 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
439 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
440 case T_LONG :
441 #ifndef _LP64
442 __ mov(O1, Itos_l2); // move other half of long
443 #endif // ifdef or no ifdef, fall through to the T_INT case
444 case T_INT : __ mov(O0, Itos_i); break;
445 case T_VOID : /* nothing to do */ break;
446 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
447 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
448 case T_OBJECT :
449 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
450 __ verify_oop(Itos_i);
451 break;
452 default : ShouldNotReachHere();
453 }
454 __ ret(); // return from interpreter activation
455 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
456 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
457 return entry;
458 }
459
460 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
461 address entry = __ pc();
462 __ push(state);
463 __ call_VM(noreg, runtime_entry);
464 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
465 return entry;
466 }
467
468
469 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
470 address entry = __ pc();
471 __ dispatch_next(state);
472 return entry;
473 }
474
475 //
476 // Helpers for commoning out cases in the various type of method entries.
477 //
478
479 // increment invocation count & check for overflow
480 //
481 // Note: checking for negative value instead of overflow
482 // so we have a 'sticky' overflow test
483 //
484 // Lmethod: method
485 // ??: invocation counter
486 //
487 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
488 // Note: In tiered we increment either counters in MethodCounters* or in
489 // MDO depending if we're profiling or not.
490 const Register G3_method_counters = G3_scratch;
491 Label done;
492
493 if (TieredCompilation) {
494 const int increment = InvocationCounter::count_increment;
495 Label no_mdo;
496 if (ProfileInterpreter) {
497 // If no method data exists, go to profile_continue.
498 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
499 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
500 // Increment counter
501 Address mdo_invocation_counter(G4_scratch,
502 in_bytes(MethodData::invocation_counter_offset()) +
503 in_bytes(InvocationCounter::counter_offset()));
504 Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset()));
505 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
506 G3_scratch, Lscratch,
507 Assembler::zero, overflow);
508 __ ba_short(done);
509 }
510
511 // Increment counter in MethodCounters*
512 __ bind(no_mdo);
513 Address invocation_counter(G3_method_counters,
514 in_bytes(MethodCounters::invocation_counter_offset()) +
515 in_bytes(InvocationCounter::counter_offset()));
516 __ get_method_counters(Lmethod, G3_method_counters, done);
517 Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset()));
518 __ increment_mask_and_jump(invocation_counter, increment, mask,
519 G4_scratch, Lscratch,
520 Assembler::zero, overflow);
521 __ bind(done);
522 } else { // not TieredCompilation
523 // Update standard invocation counters
524 __ get_method_counters(Lmethod, G3_method_counters, done);
525 __ increment_invocation_counter(G3_method_counters, O0, G4_scratch);
526 if (ProfileInterpreter) {
527 Address interpreter_invocation_counter(G3_method_counters,
528 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
529 __ ld(interpreter_invocation_counter, G4_scratch);
530 __ inc(G4_scratch);
531 __ st(G4_scratch, interpreter_invocation_counter);
532 }
533
534 if (ProfileInterpreter && profile_method != NULL) {
535 // Test to see if we should create a method data oop
536 Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset()));
537 __ ld(profile_limit, G1_scratch);
538 __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
539
540 // if no method data exists, go to profile_method
541 __ test_method_data_pointer(*profile_method);
542 }
543
544 Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset()));
545 __ ld(invocation_limit, G3_scratch);
546 __ cmp(O0, G3_scratch);
547 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
548 __ delayed()->nop();
549 __ bind(done);
550 }
551
552 }
553
554 // Allocate monitor and lock method (asm interpreter)
555 // ebx - Method*
556 //
557 void TemplateInterpreterGenerator::lock_method() {
558 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
559
560 #ifdef ASSERT
561 { Label ok;
562 __ btst(JVM_ACC_SYNCHRONIZED, O0);
563 __ br( Assembler::notZero, false, Assembler::pt, ok);
564 __ delayed()->nop();
565 __ stop("method doesn't need synchronization");
566 __ bind(ok);
567 }
568 #endif // ASSERT
569
570 // get synchronization object to O0
571 { Label done;
572 __ btst(JVM_ACC_STATIC, O0);
573 __ br( Assembler::zero, true, Assembler::pt, done);
574 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
575
576 // lock the mirror, not the Klass*
577 __ load_mirror(O0, Lmethod);
578
579 #ifdef ASSERT
580 __ tst(O0);
581 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
582 #endif // ASSERT
583
584 __ bind(done);
585 }
586
587 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
588 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
589 // __ untested("lock_object from method entry");
590 __ lock_object(Lmonitors, O0);
591 }
592
593 // See if we've got enough room on the stack for locals plus overhead below
594 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
595 // without going through the signal handler, i.e., reserved and yellow zones
596 // will not be made usable. The shadow zone must suffice to handle the
597 // overflow.
598 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
599 Register Rscratch) {
600 const int page_size = os::vm_page_size();
601 Label after_frame_check;
602
603 assert_different_registers(Rframe_size, Rscratch);
604
605 __ set(page_size, Rscratch);
606 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
607
608 // Get the stack overflow limit, and in debug, verify it is non-zero.
609 __ ld_ptr(G2_thread, JavaThread::stack_overflow_limit_offset(), Rscratch);
610 #ifdef ASSERT
611 Label limit_ok;
612 __ br_notnull_short(Rscratch, Assembler::pn, limit_ok);
613 __ stop("stack overflow limit is zero in generate_stack_overflow_check");
614 __ bind(limit_ok);
615 #endif
616
617 // Add in the size of the frame (which is the same as subtracting it from the
618 // SP, which would take another register.
619 __ add(Rscratch, Rframe_size, Rscratch);
620
621 // The frame is greater than one page in size, so check against
622 // the bottom of the stack.
623 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
624
625 // The stack will overflow, throw an exception.
626
627 // Note that SP is restored to sender's sp (in the delay slot). This
628 // is necessary if the sender's frame is an extended compiled frame
629 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
630 // adaptations.
631
632 // Note also that the restored frame is not necessarily interpreted.
633 // Use the shared runtime version of the StackOverflowError.
634 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
635 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
636 __ jump_to(stub, Rscratch);
637 __ delayed()->mov(O5_savedSP, SP);
638
639 // If you get to here, then there is enough stack space.
640 __ bind(after_frame_check);
641 }
642
643
644 //
645 // Generate a fixed interpreter frame. This is identical setup for interpreted
646 // methods and for native methods hence the shared code.
647
648
649 //----------------------------------------------------------------------------------------------------
650 // Stack frame layout
651 //
652 // When control flow reaches any of the entry types for the interpreter
653 // the following holds ->
654 //
655 // C2 Calling Conventions:
656 //
657 // The entry code below assumes that the following registers are set
658 // when coming in:
659 // G5_method: holds the Method* of the method to call
660 // Lesp: points to the TOS of the callers expression stack
661 // after having pushed all the parameters
662 //
663 // The entry code does the following to setup an interpreter frame
664 // pop parameters from the callers stack by adjusting Lesp
665 // set O0 to Lesp
666 // compute X = (max_locals - num_parameters)
667 // bump SP up by X to accomadate the extra locals
668 // compute X = max_expression_stack
669 // + vm_local_words
670 // + 16 words of register save area
671 // save frame doing a save sp, -X, sp growing towards lower addresses
672 // set Lbcp, Lmethod, LcpoolCache
673 // set Llocals to i0
674 // set Lmonitors to FP - rounded_vm_local_words
675 // set Lesp to Lmonitors - 4
676 //
677 // The frame has now been setup to do the rest of the entry code
678
679 // Try this optimization: Most method entries could live in a
680 // "one size fits all" stack frame without all the dynamic size
681 // calculations. It might be profitable to do all this calculation
682 // statically and approximately for "small enough" methods.
683
684 //-----------------------------------------------------------------------------------------------
685
686 // C1 Calling conventions
687 //
688 // Upon method entry, the following registers are setup:
689 //
690 // g2 G2_thread: current thread
691 // g5 G5_method: method to activate
692 // g4 Gargs : pointer to last argument
693 //
694 //
695 // Stack:
696 //
697 // +---------------+ <--- sp
698 // | |
699 // : reg save area :
700 // | |
701 // +---------------+ <--- sp + 0x40
702 // | |
703 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
704 // | |
705 // +---------------+ <--- sp + 0x5c
706 // | |
707 // : free :
708 // | |
709 // +---------------+ <--- Gargs
710 // | |
711 // : arguments :
712 // | |
713 // +---------------+
714 // | |
715 //
716 //
717 //
718 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
719 //
720 // +---------------+ <--- sp
721 // | |
722 // : reg save area :
723 // | |
724 // +---------------+ <--- sp + 0x40
725 // | |
726 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
727 // | |
728 // +---------------+ <--- sp + 0x5c
729 // | |
730 // : :
731 // | | <--- Lesp
732 // +---------------+ <--- Lmonitors (fp - 0x18)
733 // | VM locals |
734 // +---------------+ <--- fp
735 // | |
736 // : reg save area :
737 // | |
738 // +---------------+ <--- fp + 0x40
739 // | |
740 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
741 // | |
742 // +---------------+ <--- fp + 0x5c
743 // | |
744 // : free :
745 // | |
746 // +---------------+
747 // | |
748 // : nonarg locals :
749 // | |
750 // +---------------+
751 // | |
752 // : arguments :
753 // | | <--- Llocals
754 // +---------------+ <--- Gargs
755 // | |
756
757 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
758 //
759 //
760 // The entry code sets up a new interpreter frame in 4 steps:
761 //
762 // 1) Increase caller's SP by for the extra local space needed:
763 // (check for overflow)
764 // Efficient implementation of xload/xstore bytecodes requires
765 // that arguments and non-argument locals are in a contigously
766 // addressable memory block => non-argument locals must be
767 // allocated in the caller's frame.
768 //
769 // 2) Create a new stack frame and register window:
770 // The new stack frame must provide space for the standard
771 // register save area, the maximum java expression stack size,
772 // the monitor slots (0 slots initially), and some frame local
773 // scratch locations.
774 //
775 // 3) The following interpreter activation registers must be setup:
776 // Lesp : expression stack pointer
777 // Lbcp : bytecode pointer
778 // Lmethod : method
779 // Llocals : locals pointer
780 // Lmonitors : monitor pointer
781 // LcpoolCache: constant pool cache
782 //
783 // 4) Initialize the non-argument locals if necessary:
784 // Non-argument locals may need to be initialized to NULL
785 // for GC to work. If the oop-map information is accurate
786 // (in the absence of the JSR problem), no initialization
787 // is necessary.
788 //
789 // (gri - 2/25/2000)
790
791
792 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
793
794 const int extra_space =
795 rounded_vm_local_words + // frame local scratch space
796 Method::extra_stack_entries() + // extra stack for jsr 292
797 frame::memory_parameter_word_sp_offset + // register save area
798 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
799
800 const Register Glocals_size = G3;
801 const Register RconstMethod = Glocals_size;
802 const Register Otmp1 = O3;
803 const Register Otmp2 = O4;
804 // Lscratch can't be used as a temporary because the call_stub uses
805 // it to assert that the stack frame was setup correctly.
806 const Address constMethod (G5_method, Method::const_offset());
807 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
808
809 __ ld_ptr( constMethod, RconstMethod );
810 __ lduh( size_of_parameters, Glocals_size);
811
812 // Gargs points to first local + BytesPerWord
813 // Set the saved SP after the register window save
814 //
815 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
816 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
817 __ add(Gargs, Otmp1, Gargs);
818
819 if (native_call) {
820 __ calc_mem_param_words( Glocals_size, Gframe_size );
821 __ add( Gframe_size, extra_space, Gframe_size);
822 __ round_to( Gframe_size, WordsPerLong );
823 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
824
825 // Native calls don't need the stack size check since they have no
826 // expression stack and the arguments are already on the stack and
827 // we only add a handful of words to the stack.
828 } else {
829
830 //
831 // Compute number of locals in method apart from incoming parameters
832 //
833 const Address size_of_locals(Otmp1, ConstMethod::size_of_locals_offset());
834 __ ld_ptr(constMethod, Otmp1);
835 __ lduh(size_of_locals, Otmp1);
836 __ sub(Otmp1, Glocals_size, Glocals_size);
837 __ round_to(Glocals_size, WordsPerLong);
838 __ sll(Glocals_size, Interpreter::logStackElementSize, Glocals_size);
839
840 // See if the frame is greater than one page in size. If so,
841 // then we need to verify there is enough stack space remaining.
842 // Frame_size = (max_stack + extra_space) * BytesPerWord;
843 __ ld_ptr(constMethod, Gframe_size);
844 __ lduh(Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size);
845 __ add(Gframe_size, extra_space, Gframe_size);
846 __ round_to(Gframe_size, WordsPerLong);
847 __ sll(Gframe_size, Interpreter::logStackElementSize, Gframe_size);
848
849 // Add in java locals size for stack overflow check only
850 __ add(Gframe_size, Glocals_size, Gframe_size);
851
852 const Register Otmp2 = O4;
853 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
854 generate_stack_overflow_check(Gframe_size, Otmp1);
855
856 __ sub(Gframe_size, Glocals_size, Gframe_size);
857
858 //
859 // bump SP to accomodate the extra locals
860 //
861 __ sub(SP, Glocals_size, SP);
862 }
863
864 //
865 // now set up a stack frame with the size computed above
866 //
867 __ neg( Gframe_size );
868 __ save( SP, Gframe_size, SP );
869
870 //
871 // now set up all the local cache registers
872 //
873 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
874 // that all present references to Lbyte_code initialize the register
875 // immediately before use
876 if (native_call) {
877 __ mov(G0, Lbcp);
878 } else {
879 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
880 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
881 }
882 __ mov( G5_method, Lmethod); // set Lmethod
883 // Get mirror and store it in the frame as GC root for this Method*
884 Register mirror = LcpoolCache;
885 __ load_mirror(mirror, Lmethod);
886 __ st_ptr(mirror, FP, (frame::interpreter_frame_mirror_offset * wordSize) + STACK_BIAS);
887 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
888 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
889 #ifdef _LP64
890 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
891 #endif
892 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
893
894 // setup interpreter activation registers
895 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
896
897 if (ProfileInterpreter) {
898 #ifdef FAST_DISPATCH
899 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
900 // they both use I2.
901 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
902 #endif // FAST_DISPATCH
903 __ set_method_data_pointer();
904 }
905
906 }
907
908 // Method entry for java.lang.ref.Reference.get.
909 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
910 #if INCLUDE_ALL_GCS
911 // Code: _aload_0, _getfield, _areturn
912 // parameter size = 1
913 //
914 // The code that gets generated by this routine is split into 2 parts:
915 // 1. The "intrinsified" code for G1 (or any SATB based GC),
916 // 2. The slow path - which is an expansion of the regular method entry.
917 //
918 // Notes:-
919 // * In the G1 code we do not check whether we need to block for
920 // a safepoint. If G1 is enabled then we must execute the specialized
921 // code for Reference.get (except when the Reference object is null)
922 // so that we can log the value in the referent field with an SATB
923 // update buffer.
924 // If the code for the getfield template is modified so that the
925 // G1 pre-barrier code is executed when the current method is
926 // Reference.get() then going through the normal method entry
927 // will be fine.
928 // * The G1 code can, however, check the receiver object (the instance
929 // of java.lang.Reference) and jump to the slow path if null. If the
930 // Reference object is null then we obviously cannot fetch the referent
931 // and so we don't need to call the G1 pre-barrier. Thus we can use the
932 // regular method entry code to generate the NPE.
933 //
934 // This code is based on generate_accessor_enty.
935
936 address entry = __ pc();
937
938 const int referent_offset = java_lang_ref_Reference::referent_offset;
939 guarantee(referent_offset > 0, "referent offset not initialized");
940
941 if (UseG1GC) {
942 Label slow_path;
943
944 // In the G1 code we don't check if we need to reach a safepoint. We
945 // continue and the thread will safepoint at the next bytecode dispatch.
946
947 // Check if local 0 != NULL
948 // If the receiver is null then it is OK to jump to the slow path.
949 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
950 // check if local 0 == NULL and go the slow path
951 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
952
953
954 // Load the value of the referent field.
955 if (Assembler::is_simm13(referent_offset)) {
956 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
957 } else {
958 __ set(referent_offset, G3_scratch);
959 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
960 }
961
962 // Generate the G1 pre-barrier code to log the value of
963 // the referent field in an SATB buffer. Note with
964 // these parameters the pre-barrier does not generate
965 // the load of the previous value
966
967 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
968 Otos_i /* pre_val */,
969 G3_scratch /* tmp */,
970 true /* preserve_o_regs */);
971
972 // _areturn
973 __ retl(); // return from leaf routine
974 __ delayed()->mov(O5_savedSP, SP);
975
976 // Generate regular method entry
977 __ bind(slow_path);
978 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
979 return entry;
980 }
981 #endif // INCLUDE_ALL_GCS
982
983 // If G1 is not enabled then attempt to go through the accessor entry point
984 // Reference.get is an accessor
985 return NULL;
986 }
987
988 /**
989 * Method entry for static native methods:
990 * int java.util.zip.CRC32.update(int crc, int b)
991 */
992 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
993
994 if (UseCRC32Intrinsics) {
995 address entry = __ pc();
996
997 Label L_slow_path;
998 // If we need a safepoint check, generate full interpreter entry.
999 ExternalAddress state(SafepointSynchronize::address_of_state());
1000 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
1001 __ set(SafepointSynchronize::_not_synchronized, O3);
1002 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
1003
1004 // Load parameters
1005 const Register crc = O0; // initial crc
1006 const Register val = O1; // byte to update with
1007 const Register table = O2; // address of 256-entry lookup table
1008
1009 __ ldub(Gargs, 3, val);
1010 __ lduw(Gargs, 8, crc);
1011
1012 __ set(ExternalAddress(StubRoutines::crc_table_addr()), table);
1013
1014 __ not1(crc); // ~crc
1015 __ clruwu(crc);
1016 __ update_byte_crc32(crc, val, table);
1017 __ not1(crc); // ~crc
1018
1019 // result in O0
1020 __ retl();
1021 __ delayed()->nop();
1022
1023 // generate a vanilla native entry as the slow path
1024 __ bind(L_slow_path);
1025 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
1026 return entry;
1027 }
1028 return NULL;
1029 }
1030
1031 /**
1032 * Method entry for static native methods:
1033 * int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len)
1034 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
1035 */
1036 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1037
1038 if (UseCRC32Intrinsics) {
1039 address entry = __ pc();
1040
1041 Label L_slow_path;
1042 // If we need a safepoint check, generate full interpreter entry.
1043 ExternalAddress state(SafepointSynchronize::address_of_state());
1044 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
1045 __ set(SafepointSynchronize::_not_synchronized, O3);
1046 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
1047
1048 // Load parameters from the stack
1049 const Register crc = O0; // initial crc
1050 const Register buf = O1; // source java byte array address
1051 const Register len = O2; // len
1052 const Register offset = O3; // offset
1053
1054 // Arguments are reversed on java expression stack
1055 // Calculate address of start element
1056 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) {
1057 __ lduw(Gargs, 0, len);
1058 __ lduw(Gargs, 8, offset);
1059 __ ldx( Gargs, 16, buf);
1060 __ lduw(Gargs, 32, crc);
1061 __ add(buf, offset, buf);
1062 } else {
1063 __ lduw(Gargs, 0, len);
1064 __ lduw(Gargs, 8, offset);
1065 __ ldx( Gargs, 16, buf);
1066 __ lduw(Gargs, 24, crc);
1067 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1068 __ add(buf ,offset, buf);
1069 }
1070
1071 // Call the crc32 kernel
1072 __ MacroAssembler::save_thread(L7_thread_cache);
1073 __ kernel_crc32(crc, buf, len, O3);
1074 __ MacroAssembler::restore_thread(L7_thread_cache);
1075
1076 // result in O0
1077 __ retl();
1078 __ delayed()->nop();
1079
1080 // generate a vanilla native entry as the slow path
1081 __ bind(L_slow_path);
1082 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
1083 return entry;
1084 }
1085 return NULL;
1086 }
1087
1088 /**
1089 * Method entry for intrinsic-candidate (non-native) methods:
1090 * int java.util.zip.CRC32C.updateBytes(int crc, byte[] b, int off, int end)
1091 * int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
1092 * Unlike CRC32, CRC32C does not have any methods marked as native
1093 * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
1094 */
1095 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1096
1097 if (UseCRC32CIntrinsics) {
1098 address entry = __ pc();
1099
1100 // Load parameters from the stack
1101 const Register crc = O0; // initial crc
1102 const Register buf = O1; // source java byte array address
1103 const Register offset = O2; // offset
1104 const Register end = O3; // index of last element to process
1105 const Register len = O2; // len argument to the kernel
1106 const Register table = O3; // crc32c lookup table address
1107
1108 // Arguments are reversed on java expression stack
1109 // Calculate address of start element
1110 if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) {
1111 __ lduw(Gargs, 0, end);
1112 __ lduw(Gargs, 8, offset);
1113 __ ldx( Gargs, 16, buf);
1114 __ lduw(Gargs, 32, crc);
1115 __ add(buf, offset, buf);
1116 __ sub(end, offset, len);
1117 } else {
1118 __ lduw(Gargs, 0, end);
1119 __ lduw(Gargs, 8, offset);
1120 __ ldx( Gargs, 16, buf);
1121 __ lduw(Gargs, 24, crc);
1122 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1123 __ add(buf, offset, buf);
1124 __ sub(end, offset, len);
1125 }
1126
1127 // Call the crc32c kernel
1128 __ MacroAssembler::save_thread(L7_thread_cache);
1129 __ kernel_crc32c(crc, buf, len, table);
1130 __ MacroAssembler::restore_thread(L7_thread_cache);
1131
1132 // result in O0
1133 __ retl();
1134 __ delayed()->nop();
1135
1136 return entry;
1137 }
1138 return NULL;
1139 }
1140
1141 // Not supported
1142 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
1143 return NULL;
1144 }
1145
1146 // TODO: rather than touching all pages, check against stack_overflow_limit and bang yellow page to
1147 // generate exception
1148 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1149 // Quick & dirty stack overflow checking: bang the stack & handle trap.
1150 // Note that we do the banging after the frame is setup, since the exception
1151 // handling code expects to find a valid interpreter frame on the stack.
1152 // Doing the banging earlier fails if the caller frame is not an interpreter
1153 // frame.
1154 // (Also, the exception throwing code expects to unlock any synchronized
1155 // method receiever, so do the banging after locking the receiver.)
1156
1157 // Bang each page in the shadow zone. We can't assume it's been done for
1158 // an interpreter frame with greater than a page of locals, so each page
1159 // needs to be checked. Only true for non-native.
1160 if (UseStackBanging) {
1161 const int page_size = os::vm_page_size();
1162 const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1163 const int start_page = native_call ? n_shadow_pages : 1;
1164 for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1165 __ bang_stack_with_offset(pages*page_size);
1166 }
1167 }
1168 }
1169
1170 //
1171 // Interpreter stub for calling a native method. (asm interpreter)
1172 // This sets up a somewhat different looking stack for calling the native method
1173 // than the typical interpreter frame setup.
1174 //
1175
1176 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1177 address entry = __ pc();
1178
1179 // the following temporary registers are used during frame creation
1180 const Register Gtmp1 = G3_scratch ;
1181 const Register Gtmp2 = G1_scratch;
1182 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1183
1184 // make sure registers are different!
1185 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1186
1187 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
1188
1189 const Register Glocals_size = G3;
1190 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1191
1192 // make sure method is native & not abstract
1193 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1194 #ifdef ASSERT
1195 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1196 {
1197 Label L;
1198 __ btst(JVM_ACC_NATIVE, Gtmp1);
1199 __ br(Assembler::notZero, false, Assembler::pt, L);
1200 __ delayed()->nop();
1201 __ stop("tried to execute non-native method as native");
1202 __ bind(L);
1203 }
1204 { Label L;
1205 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1206 __ br(Assembler::zero, false, Assembler::pt, L);
1207 __ delayed()->nop();
1208 __ stop("tried to execute abstract method as non-abstract");
1209 __ bind(L);
1210 }
1211 #endif // ASSERT
1212
1213 // generate the code to allocate the interpreter stack frame
1214 generate_fixed_frame(true);
1215
1216 //
1217 // No locals to initialize for native method
1218 //
1219
1220 // this slot will be set later, we initialize it to null here just in
1221 // case we get a GC before the actual value is stored later
1222 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
1223
1224 const Address do_not_unlock_if_synchronized(G2_thread,
1225 JavaThread::do_not_unlock_if_synchronized_offset());
1226 // Since at this point in the method invocation the exception handler
1227 // would try to exit the monitor of synchronized methods which hasn't
1228 // been entered yet, we set the thread local variable
1229 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1230 // runtime, exception handling i.e. unlock_if_synchronized_method will
1231 // check this thread local flag.
1232 // This flag has two effects, one is to force an unwind in the topmost
1233 // interpreter frame and not perform an unlock while doing so.
1234
1235 __ movbool(true, G3_scratch);
1236 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1237
1238 // increment invocation counter and check for overflow
1239 //
1240 // Note: checking for negative value instead of overflow
1241 // so we have a 'sticky' overflow test (may be of
1242 // importance as soon as we have true MT/MP)
1243 Label invocation_counter_overflow;
1244 Label Lcontinue;
1245 if (inc_counter) {
1246 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1247
1248 }
1249 __ bind(Lcontinue);
1250
1251 bang_stack_shadow_pages(true);
1252
1253 // reset the _do_not_unlock_if_synchronized flag
1254 __ stbool(G0, do_not_unlock_if_synchronized);
1255
1256 // check for synchronized methods
1257 // Must happen AFTER invocation_counter check and stack overflow check,
1258 // so method is not locked if overflows.
1259
1260 if (synchronized) {
1261 lock_method();
1262 } else {
1263 #ifdef ASSERT
1264 { Label ok;
1265 __ ld(Laccess_flags, O0);
1266 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1267 __ br( Assembler::zero, false, Assembler::pt, ok);
1268 __ delayed()->nop();
1269 __ stop("method needs synchronization");
1270 __ bind(ok);
1271 }
1272 #endif // ASSERT
1273 }
1274
1275
1276 // start execution
1277 __ verify_thread();
1278
1279 // JVMTI support
1280 __ notify_method_entry();
1281
1282 // native call
1283
1284 // (note that O0 is never an oop--at most it is a handle)
1285 // It is important not to smash any handles created by this call,
1286 // until any oop handle in O0 is dereferenced.
1287
1288 // (note that the space for outgoing params is preallocated)
1289
1290 // get signature handler
1291 { Label L;
1292 Address signature_handler(Lmethod, Method::signature_handler_offset());
1293 __ ld_ptr(signature_handler, G3_scratch);
1294 __ br_notnull_short(G3_scratch, Assembler::pt, L);
1295 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
1296 __ ld_ptr(signature_handler, G3_scratch);
1297 __ bind(L);
1298 }
1299
1300 // Push a new frame so that the args will really be stored in
1301 // Copy a few locals across so the new frame has the variables
1302 // we need but these values will be dead at the jni call and
1303 // therefore not gc volatile like the values in the current
1304 // frame (Lmethod in particular)
1305
1306 // Flush the method pointer to the register save area
1307 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
1308 __ mov(Llocals, O1);
1309
1310 // calculate where the mirror handle body is allocated in the interpreter frame:
1311 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
1312
1313 // Calculate current frame size
1314 __ sub(SP, FP, O3); // Calculate negative of current frame size
1315 __ save(SP, O3, SP); // Allocate an identical sized frame
1316
1317 // Note I7 has leftover trash. Slow signature handler will fill it in
1318 // should we get there. Normal jni call will set reasonable last_Java_pc
1319 // below (and fix I7 so the stack trace doesn't have a meaningless frame
1320 // in it).
1321
1322 // Load interpreter frame's Lmethod into same register here
1323
1324 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1325
1326 __ mov(I1, Llocals);
1327 __ mov(I2, Lscratch2); // save the address of the mirror
1328
1329
1330 // ONLY Lmethod and Llocals are valid here!
1331
1332 // call signature handler, It will move the arg properly since Llocals in current frame
1333 // matches that in outer frame
1334
1335 __ callr(G3_scratch, 0);
1336 __ delayed()->nop();
1337
1338 // Result handler is in Lscratch
1339
1340 // Reload interpreter frame's Lmethod since slow signature handler may block
1341 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1342
1343 { Label not_static;
1344
1345 __ ld(Laccess_flags, O0);
1346 __ btst(JVM_ACC_STATIC, O0);
1347 __ br( Assembler::zero, false, Assembler::pt, not_static);
1348 // get native function entry point(O0 is a good temp until the very end)
1349 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
1350 // for static methods insert the mirror argument
1351 __ load_mirror(O1, Lmethod);
1352 #ifdef ASSERT
1353 if (!PrintSignatureHandlers) // do not dirty the output with this
1354 { Label L;
1355 __ br_notnull_short(O1, Assembler::pt, L);
1356 __ stop("mirror is missing");
1357 __ bind(L);
1358 }
1359 #endif // ASSERT
1360 __ st_ptr(O1, Lscratch2, 0);
1361 __ mov(Lscratch2, O1);
1362 __ bind(not_static);
1363 }
1364
1365 // At this point, arguments have been copied off of stack into
1366 // their JNI positions, which are O1..O5 and SP[68..].
1367 // Oops are boxed in-place on the stack, with handles copied to arguments.
1368 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1369
1370 #ifdef ASSERT
1371 { Label L;
1372 __ br_notnull_short(O0, Assembler::pt, L);
1373 __ stop("native entry point is missing");
1374 __ bind(L);
1375 }
1376 #endif // ASSERT
1377
1378 //
1379 // setup the frame anchor
1380 //
1381 // The scavenge function only needs to know that the PC of this frame is
1382 // in the interpreter method entry code, it doesn't need to know the exact
1383 // PC and hence we can use O7 which points to the return address from the
1384 // previous call in the code stream (signature handler function)
1385 //
1386 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1387 // we have pushed the extra frame in order to protect the volatile register(s)
1388 // in that frame when we return from the jni call
1389 //
1390
1391 __ set_last_Java_frame(FP, O7);
1392 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1393 // not meaningless information that'll confuse me.
1394
1395 // flush the windows now. We don't care about the current (protection) frame
1396 // only the outer frames
1397
1398 __ flushw();
1399
1400 // mark windows as flushed
1401 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1402 __ set(JavaFrameAnchor::flushed, G3_scratch);
1403 __ st(G3_scratch, flags);
1404
1405 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1406
1407 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1408 #ifdef ASSERT
1409 { Label L;
1410 __ ld(thread_state, G3_scratch);
1411 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1412 __ stop("Wrong thread state in native stub");
1413 __ bind(L);
1414 }
1415 #endif // ASSERT
1416 __ set(_thread_in_native, G3_scratch);
1417 __ st(G3_scratch, thread_state);
1418
1419 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1420 __ save_thread(L7_thread_cache); // save Gthread
1421 __ callr(O0, 0);
1422 __ delayed()->
1423 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1424
1425 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1426
1427 __ restore_thread(L7_thread_cache); // restore G2_thread
1428 __ reinit_heapbase();
1429
1430 // must we block?
1431
1432 // Block, if necessary, before resuming in _thread_in_Java state.
1433 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1434 { Label no_block;
1435 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1436
1437 // Switch thread to "native transition" state before reading the synchronization state.
1438 // This additional state is necessary because reading and testing the synchronization
1439 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1440 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1441 // VM thread changes sync state to synchronizing and suspends threads for GC.
1442 // Thread A is resumed to finish this native method, but doesn't block here since it
1443 // didn't see any synchronization is progress, and escapes.
1444 __ set(_thread_in_native_trans, G3_scratch);
1445 __ st(G3_scratch, thread_state);
1446 if(os::is_MP()) {
1447 if (UseMembar) {
1448 // Force this write out before the read below
1449 __ membar(Assembler::StoreLoad);
1450 } else {
1451 // Write serialization page so VM thread can do a pseudo remote membar.
1452 // We use the current thread pointer to calculate a thread specific
1453 // offset to write to within the page. This minimizes bus traffic
1454 // due to cache line collision.
1455 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1456 }
1457 }
1458 __ load_contents(sync_state, G3_scratch);
1459 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1460
1461 Label L;
1462 __ br(Assembler::notEqual, false, Assembler::pn, L);
1463 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1464 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1465 __ bind(L);
1466
1467 // Block. Save any potential method result value before the operation and
1468 // use a leaf call to leave the last_Java_frame setup undisturbed.
1469 save_native_result();
1470 __ call_VM_leaf(L7_thread_cache,
1471 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1472 G2_thread);
1473
1474 // Restore any method result value
1475 restore_native_result();
1476 __ bind(no_block);
1477 }
1478
1479 // Clear the frame anchor now
1480
1481 __ reset_last_Java_frame();
1482
1483 // Move the result handler address
1484 __ mov(Lscratch, G3_scratch);
1485 // return possible result to the outer frame
1486 #ifndef __LP64
1487 __ mov(O0, I0);
1488 __ restore(O1, G0, O1);
1489 #else
1490 __ restore(O0, G0, O0);
1491 #endif /* __LP64 */
1492
1493 // Move result handler to expected register
1494 __ mov(G3_scratch, Lscratch);
1495
1496 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1497 // switch to thread_in_Java.
1498
1499 __ set(_thread_in_Java, G3_scratch);
1500 __ st(G3_scratch, thread_state);
1501
1502 if (CheckJNICalls) {
1503 // clear_pending_jni_exception_check
1504 __ st_ptr(G0, G2_thread, JavaThread::pending_jni_exception_check_fn_offset());
1505 }
1506
1507 // reset handle block
1508 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1509 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1510
1511 // If we have an oop result store it where it will be safe for any further gc
1512 // until we return now that we've released the handle it might be protected by
1513
1514 {
1515 Label no_oop, store_result;
1516
1517 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1518 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1519 // Unbox oop result, e.g. JNIHandles::resolve value in O0.
1520 __ br_null(O0, false, Assembler::pn, store_result); // Use NULL as-is.
1521 __ delayed()->andcc(O0, JNIHandles::weak_tag_mask, G0); // Test for jweak
1522 __ brx(Assembler::zero, true, Assembler::pt, store_result);
1523 __ delayed()->ld_ptr(O0, 0, O0); // Maybe resolve (untagged) jobject.
1524 // Resolve jweak.
1525 __ ld_ptr(O0, -JNIHandles::weak_tag_value, O0);
1526 #if INCLUDE_ALL_GCS
1527 if (UseG1GC) {
1528 __ g1_write_barrier_pre(noreg /* obj */,
1529 noreg /* index */,
1530 0 /* offset */,
1531 O0 /* pre_val */,
1532 G3_scratch /* tmp */,
1533 true /* preserve_o_regs */);
1534 }
1535 #endif // INCLUDE_ALL_GCS
1536 __ bind(store_result);
1537 // Store it where gc will look for it and result handler expects it.
1538 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1539
1540 __ bind(no_oop);
1541
1542 }
1543
1544
1545 // handle exceptions (exception handling will handle unlocking!)
1546 { Label L;
1547 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1548 __ ld_ptr(exception_addr, Gtemp);
1549 __ br_null_short(Gtemp, Assembler::pt, L);
1550 // Note: This could be handled more efficiently since we know that the native
1551 // method doesn't have an exception handler. We could directly return
1552 // to the exception handler for the caller.
1553 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1554 __ should_not_reach_here();
1555 __ bind(L);
1556 }
1557
1558 // JVMTI support (preserves thread register)
1559 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1560
1561 if (synchronized) {
1562 // save and restore any potential method result value around the unlocking operation
1563 save_native_result();
1564
1565 __ add( __ top_most_monitor(), O1);
1566 __ unlock_object(O1);
1567
1568 restore_native_result();
1569 }
1570
1571 #if defined(COMPILER2) && !defined(_LP64)
1572
1573 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1574 // or compiled so just be safe.
1575
1576 __ sllx(O0, 32, G1); // Shift bits into high G1
1577 __ srl (O1, 0, O1); // Zero extend O1
1578 __ or3 (O1, G1, G1); // OR 64 bits into G1
1579
1580 #endif /* COMPILER2 && !_LP64 */
1581
1582 // dispose of return address and remove activation
1583 #ifdef ASSERT
1584 {
1585 Label ok;
1586 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1587 __ stop("bad I5_savedSP value");
1588 __ should_not_reach_here();
1589 __ bind(ok);
1590 }
1591 #endif
1592 __ jmp(Lscratch, 0);
1593 __ delayed()->nop();
1594
1595
1596 if (inc_counter) {
1597 // handle invocation counter overflow
1598 __ bind(invocation_counter_overflow);
1599 generate_counter_overflow(Lcontinue);
1600 }
1601
1602
1603
1604 return entry;
1605 }
1606
1607
1608 // Generic method entry to (asm) interpreter
1609 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1610 address entry = __ pc();
1611
1612 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1613
1614 // the following temporary registers are used during frame creation
1615 const Register Gtmp1 = G3_scratch ;
1616 const Register Gtmp2 = G1_scratch;
1617
1618 // make sure registers are different!
1619 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1620
1621 const Address constMethod (G5_method, Method::const_offset());
1622 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1623 // and use in the asserts.
1624 const Address access_flags (Lmethod, Method::access_flags_offset());
1625
1626 const Register Glocals_size = G3;
1627 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1628
1629 // make sure method is not native & not abstract
1630 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1631 #ifdef ASSERT
1632 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1633 {
1634 Label L;
1635 __ btst(JVM_ACC_NATIVE, Gtmp1);
1636 __ br(Assembler::zero, false, Assembler::pt, L);
1637 __ delayed()->nop();
1638 __ stop("tried to execute native method as non-native");
1639 __ bind(L);
1640 }
1641 { Label L;
1642 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1643 __ br(Assembler::zero, false, Assembler::pt, L);
1644 __ delayed()->nop();
1645 __ stop("tried to execute abstract method as non-abstract");
1646 __ bind(L);
1647 }
1648 #endif // ASSERT
1649
1650 // generate the code to allocate the interpreter stack frame
1651
1652 generate_fixed_frame(false);
1653
1654 #ifdef FAST_DISPATCH
1655 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1656 // set bytecode dispatch table base
1657 #endif
1658
1659 //
1660 // Code to initialize the extra (i.e. non-parm) locals
1661 //
1662 Register init_value = noreg; // will be G0 if we must clear locals
1663 // The way the code was setup before zerolocals was always true for vanilla java entries.
1664 // It could only be false for the specialized entries like accessor or empty which have
1665 // no extra locals so the testing was a waste of time and the extra locals were always
1666 // initialized. We removed this extra complication to already over complicated code.
1667
1668 init_value = G0;
1669 Label clear_loop;
1670
1671 const Register RconstMethod = O1;
1672 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1673 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1674
1675 // NOTE: If you change the frame layout, this code will need to
1676 // be updated!
1677 __ ld_ptr( constMethod, RconstMethod );
1678 __ lduh( size_of_locals, O2 );
1679 __ lduh( size_of_parameters, O1 );
1680 __ sll( O2, Interpreter::logStackElementSize, O2);
1681 __ sll( O1, Interpreter::logStackElementSize, O1 );
1682 __ sub( Llocals, O2, O2 );
1683 __ sub( Llocals, O1, O1 );
1684
1685 __ bind( clear_loop );
1686 __ inc( O2, wordSize );
1687
1688 __ cmp( O2, O1 );
1689 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1690 __ delayed()->st_ptr( init_value, O2, 0 );
1691
1692 const Address do_not_unlock_if_synchronized(G2_thread,
1693 JavaThread::do_not_unlock_if_synchronized_offset());
1694 // Since at this point in the method invocation the exception handler
1695 // would try to exit the monitor of synchronized methods which hasn't
1696 // been entered yet, we set the thread local variable
1697 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1698 // runtime, exception handling i.e. unlock_if_synchronized_method will
1699 // check this thread local flag.
1700 __ movbool(true, G3_scratch);
1701 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1702
1703 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1704 // increment invocation counter and check for overflow
1705 //
1706 // Note: checking for negative value instead of overflow
1707 // so we have a 'sticky' overflow test (may be of
1708 // importance as soon as we have true MT/MP)
1709 Label invocation_counter_overflow;
1710 Label profile_method;
1711 Label profile_method_continue;
1712 Label Lcontinue;
1713 if (inc_counter) {
1714 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1715 if (ProfileInterpreter) {
1716 __ bind(profile_method_continue);
1717 }
1718 }
1719 __ bind(Lcontinue);
1720
1721 bang_stack_shadow_pages(false);
1722
1723 // reset the _do_not_unlock_if_synchronized flag
1724 __ stbool(G0, do_not_unlock_if_synchronized);
1725
1726 // check for synchronized methods
1727 // Must happen AFTER invocation_counter check and stack overflow check,
1728 // so method is not locked if overflows.
1729
1730 if (synchronized) {
1731 lock_method();
1732 } else {
1733 #ifdef ASSERT
1734 { Label ok;
1735 __ ld(access_flags, O0);
1736 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1737 __ br( Assembler::zero, false, Assembler::pt, ok);
1738 __ delayed()->nop();
1739 __ stop("method needs synchronization");
1740 __ bind(ok);
1741 }
1742 #endif // ASSERT
1743 }
1744
1745 // start execution
1746
1747 __ verify_thread();
1748
1749 // jvmti support
1750 __ notify_method_entry();
1751
1752 // start executing instructions
1753 __ dispatch_next(vtos);
1754
1755
1756 if (inc_counter) {
1757 if (ProfileInterpreter) {
1758 // We have decided to profile this method in the interpreter
1759 __ bind(profile_method);
1760
1761 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1762 __ set_method_data_pointer_for_bcp();
1763 __ ba_short(profile_method_continue);
1764 }
1765
1766 // handle invocation counter overflow
1767 __ bind(invocation_counter_overflow);
1768 generate_counter_overflow(Lcontinue);
1769 }
1770
1771
1772 return entry;
1773 }
1774
1775 //----------------------------------------------------------------------------------------------------
1776 // Exceptions
1777 void TemplateInterpreterGenerator::generate_throw_exception() {
1778
1779 // Entry point in previous activation (i.e., if the caller was interpreted)
1780 Interpreter::_rethrow_exception_entry = __ pc();
1781 // O0: exception
1782
1783 // entry point for exceptions thrown within interpreter code
1784 Interpreter::_throw_exception_entry = __ pc();
1785 __ verify_thread();
1786 // expression stack is undefined here
1787 // O0: exception, i.e. Oexception
1788 // Lbcp: exception bcp
1789 __ verify_oop(Oexception);
1790
1791
1792 // expression stack must be empty before entering the VM in case of an exception
1793 __ empty_expression_stack();
1794 // find exception handler address and preserve exception oop
1795 // call C routine to find handler and jump to it
1796 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1797 __ push_ptr(O1); // push exception for exception handler bytecodes
1798
1799 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1800 __ delayed()->nop();
1801
1802
1803 // if the exception is not handled in the current frame
1804 // the frame is removed and the exception is rethrown
1805 // (i.e. exception continuation is _rethrow_exception)
1806 //
1807 // Note: At this point the bci is still the bxi for the instruction which caused
1808 // the exception and the expression stack is empty. Thus, for any VM calls
1809 // at this point, GC will find a legal oop map (with empty expression stack).
1810
1811 // in current activation
1812 // tos: exception
1813 // Lbcp: exception bcp
1814
1815 //
1816 // JVMTI PopFrame support
1817 //
1818
1819 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1820 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1821 // Set the popframe_processing bit in popframe_condition indicating that we are
1822 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1823 // popframe handling cycles.
1824
1825 __ ld(popframe_condition_addr, G3_scratch);
1826 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1827 __ stw(G3_scratch, popframe_condition_addr);
1828
1829 // Empty the expression stack, as in normal exception handling
1830 __ empty_expression_stack();
1831 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1832
1833 {
1834 // Check to see whether we are returning to a deoptimized frame.
1835 // (The PopFrame call ensures that the caller of the popped frame is
1836 // either interpreted or compiled and deoptimizes it if compiled.)
1837 // In this case, we can't call dispatch_next() after the frame is
1838 // popped, but instead must save the incoming arguments and restore
1839 // them after deoptimization has occurred.
1840 //
1841 // Note that we don't compare the return PC against the
1842 // deoptimization blob's unpack entry because of the presence of
1843 // adapter frames in C2.
1844 Label caller_not_deoptimized;
1845 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1846 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1847
1848 const Register Gtmp1 = G3_scratch;
1849 const Register Gtmp2 = G1_scratch;
1850 const Register RconstMethod = Gtmp1;
1851 const Address constMethod(Lmethod, Method::const_offset());
1852 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1853
1854 // Compute size of arguments for saving when returning to deoptimized caller
1855 __ ld_ptr(constMethod, RconstMethod);
1856 __ lduh(size_of_parameters, Gtmp1);
1857 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1858 __ sub(Llocals, Gtmp1, Gtmp2);
1859 __ add(Gtmp2, wordSize, Gtmp2);
1860 // Save these arguments
1861 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1862 // Inform deoptimization that it is responsible for restoring these arguments
1863 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1864 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1865 __ st(Gtmp1, popframe_condition_addr);
1866
1867 // Return from the current method
1868 // The caller's SP was adjusted upon method entry to accomodate
1869 // the callee's non-argument locals. Undo that adjustment.
1870 __ ret();
1871 __ delayed()->restore(I5_savedSP, G0, SP);
1872
1873 __ bind(caller_not_deoptimized);
1874 }
1875
1876 // Clear the popframe condition flag
1877 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1878
1879 // Get out of the current method (how this is done depends on the particular compiler calling
1880 // convention that the interpreter currently follows)
1881 // The caller's SP was adjusted upon method entry to accomodate
1882 // the callee's non-argument locals. Undo that adjustment.
1883 __ restore(I5_savedSP, G0, SP);
1884 // The method data pointer was incremented already during
1885 // call profiling. We have to restore the mdp for the current bcp.
1886 if (ProfileInterpreter) {
1887 __ set_method_data_pointer_for_bcp();
1888 }
1889
1890 #if INCLUDE_JVMTI
1891 {
1892 Label L_done;
1893
1894 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1895 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1896
1897 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1898 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1899
1900 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1901
1902 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1903 __ delayed()->nop();
1904
1905 __ st_ptr(G1_scratch, Lesp, wordSize);
1906 __ bind(L_done);
1907 }
1908 #endif // INCLUDE_JVMTI
1909
1910 // Resume bytecode interpretation at the current bcp
1911 __ dispatch_next(vtos);
1912 // end of JVMTI PopFrame support
1913
1914 Interpreter::_remove_activation_entry = __ pc();
1915
1916 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1917 __ pop_ptr(Oexception); // get exception
1918
1919 // Intel has the following comment:
1920 //// remove the activation (without doing throws on illegalMonitorExceptions)
1921 // They remove the activation without checking for bad monitor state.
1922 // %%% We should make sure this is the right semantics before implementing.
1923
1924 __ set_vm_result(Oexception);
1925 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1926
1927 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1928
1929 __ get_vm_result(Oexception);
1930 __ verify_oop(Oexception);
1931
1932 const int return_reg_adjustment = frame::pc_return_offset;
1933 Address issuing_pc_addr(I7, return_reg_adjustment);
1934
1935 // We are done with this activation frame; find out where to go next.
1936 // The continuation point will be an exception handler, which expects
1937 // the following registers set up:
1938 //
1939 // Oexception: exception
1940 // Oissuing_pc: the local call that threw exception
1941 // Other On: garbage
1942 // In/Ln: the contents of the caller's register window
1943 //
1944 // We do the required restore at the last possible moment, because we
1945 // need to preserve some state across a runtime call.
1946 // (Remember that the caller activation is unknown--it might not be
1947 // interpreted, so things like Lscratch are useless in the caller.)
1948
1949 // Although the Intel version uses call_C, we can use the more
1950 // compact call_VM. (The only real difference on SPARC is a
1951 // harmlessly ignored [re]set_last_Java_frame, compared with
1952 // the Intel code which lacks this.)
1953 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1954 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1955 __ super_call_VM_leaf(L7_thread_cache,
1956 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1957 G2_thread, Oissuing_pc->after_save());
1958
1959 // The caller's SP was adjusted upon method entry to accomodate
1960 // the callee's non-argument locals. Undo that adjustment.
1961 __ JMP(O0, 0); // return exception handler in caller
1962 __ delayed()->restore(I5_savedSP, G0, SP);
1963
1964 // (same old exception object is already in Oexception; see above)
1965 // Note that an "issuing PC" is actually the next PC after the call
1966 }
1967
1968
1969 //
1970 // JVMTI ForceEarlyReturn support
1971 //
1972
1973 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1974 address entry = __ pc();
1975
1976 __ empty_expression_stack();
1977 __ load_earlyret_value(state);
1978
1979 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1980 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1981
1982 // Clear the earlyret state
1983 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1984
1985 __ remove_activation(state,
1986 /* throw_monitor_exception */ false,
1987 /* install_monitor_exception */ false);
1988
1989 // The caller's SP was adjusted upon method entry to accomodate
1990 // the callee's non-argument locals. Undo that adjustment.
1991 __ ret(); // return to caller
1992 __ delayed()->restore(I5_savedSP, G0, SP);
1993
1994 return entry;
1995 } // end of JVMTI ForceEarlyReturn support
1996
1997
1998 //------------------------------------------------------------------------------------------------------------------------
1999 // Helper for vtos entry point generation
2000
2001 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) {
2002 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
2003 Label L;
2004 aep = __ pc(); __ push_ptr(); __ ba_short(L);
2005 fep = __ pc(); __ push_f(); __ ba_short(L);
2006 dep = __ pc(); __ push_d(); __ ba_short(L);
2007 lep = __ pc(); __ push_l(); __ ba_short(L);
2008 iep = __ pc(); __ push_i();
2009 bep = cep = sep = iep; // there aren't any
2010 vep = __ pc(); __ bind(L); // fall through
2011 generate_and_dispatch(t);
2012 }
2013
2014 // --------------------------------------------------------------------------------
2015
2016 // Non-product code
2017 #ifndef PRODUCT
2018 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2019 address entry = __ pc();
2020
2021 __ push(state);
2022 __ mov(O7, Lscratch); // protect return address within interpreter
2023
2024 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
2025 __ mov( Otos_l2, G3_scratch );
2026 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
2027 __ mov(Lscratch, O7); // restore return address
2028 __ pop(state);
2029 __ retl();
2030 __ delayed()->nop();
2031
2032 return entry;
2033 }
2034
2035
2036 // helpers for generate_and_dispatch
2037
2038 void TemplateInterpreterGenerator::count_bytecode() {
2039 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2040 }
2041
2042
2043 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2044 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2045 }
2046
2047
2048 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2049 AddressLiteral index (&BytecodePairHistogram::_index);
2050 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2051
2052 // get index, shift out old bytecode, bring in new bytecode, and store it
2053 // _index = (_index >> log2_number_of_codes) |
2054 // (bytecode << log2_number_of_codes);
2055
2056 __ load_contents(index, G4_scratch);
2057 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2058 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
2059 __ or3( G3_scratch, G4_scratch, G4_scratch );
2060 __ store_contents(G4_scratch, index, G3_scratch);
2061
2062 // bump bucket contents
2063 // _counters[_index] ++;
2064
2065 __ set(counters, G3_scratch); // loads into G3_scratch
2066 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
2067 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
2068 __ ld (G3_scratch, 0, G4_scratch);
2069 __ inc (G4_scratch);
2070 __ st (G4_scratch, 0, G3_scratch);
2071 }
2072
2073
2074 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2075 // Call a little run-time stub to avoid blow-up for each bytecode.
2076 // The run-time runtime saves the right registers, depending on
2077 // the tosca in-state for the given template.
2078 address entry = Interpreter::trace_code(t->tos_in());
2079 guarantee(entry != NULL, "entry must have been generated");
2080 __ call(entry, relocInfo::none);
2081 __ delayed()->nop();
2082 }
2083
2084
2085 void TemplateInterpreterGenerator::stop_interpreter_at() {
2086 AddressLiteral counter(&BytecodeCounter::_counter_value);
2087 __ load_contents(counter, G3_scratch);
2088 AddressLiteral stop_at(&StopInterpreterAt);
2089 __ load_ptr_contents(stop_at, G4_scratch);
2090 __ cmp(G3_scratch, G4_scratch);
2091 __ breakpoint_trap(Assembler::equal, Assembler::icc);
2092 }
2093 #endif // not PRODUCT