1 /*
2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "interpreter/bytecodeHistogram.hpp"
28 #include "interpreter/interpreter.hpp"
29 #include "interpreter/interpreterRuntime.hpp"
30 #include "interpreter/interp_masm.hpp"
31 #include "interpreter/templateInterpreterGenerator.hpp"
32 #include "interpreter/templateTable.hpp"
33 #include "oops/arrayOop.hpp"
34 #include "oops/methodData.hpp"
35 #include "oops/method.hpp"
36 #include "oops/oop.inline.hpp"
37 #include "prims/jvmtiExport.hpp"
38 #include "prims/jvmtiThreadState.hpp"
39 #include "runtime/arguments.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/frame.inline.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "runtime/stubRoutines.hpp"
44 #include "runtime/synchronizer.hpp"
45 #include "runtime/timer.hpp"
46 #include "runtime/vframeArray.hpp"
47 #include "utilities/debug.hpp"
48 #include "utilities/macros.hpp"
49
50 #ifndef FAST_DISPATCH
51 #define FAST_DISPATCH 1
52 #endif
53 #undef FAST_DISPATCH
54
55 // Size of interpreter code. Increase if too small. Interpreter will
56 // fail with a guarantee ("not enough space for interpreter generation");
57 // if too small.
58 // Run with +PrintInterpreter to get the VM to print out the size.
59 // Max size with JVMTI
60 #ifdef _LP64
61 // The sethi() instruction generates lots more instructions when shell
62 // stack limit is unlimited, so that's why this is much bigger.
63 int TemplateInterpreter::InterpreterCodeSize = 260 * K;
64 #else
65 int TemplateInterpreter::InterpreterCodeSize = 230 * K;
66 #endif
67
68 // Generation of Interpreter
69 //
70 // The TemplateInterpreterGenerator generates the interpreter into Interpreter::_code.
71
72
73 #define __ _masm->
74
75
76 //----------------------------------------------------------------------------------------------------
77
78 #ifndef _LP64
79 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
80 address entry = __ pc();
81 Argument argv(0, true);
82
83 // We are in the jni transition frame. Save the last_java_frame corresponding to the
84 // outer interpreter frame
85 //
86 __ set_last_Java_frame(FP, noreg);
87 // make sure the interpreter frame we've pushed has a valid return pc
88 __ mov(O7, I7);
89 __ mov(Lmethod, G3_scratch);
90 __ mov(Llocals, G4_scratch);
91 __ save_frame(0);
92 __ mov(G2_thread, L7_thread_cache);
93 __ add(argv.address_in_frame(), O3);
94 __ mov(G2_thread, O0);
95 __ mov(G3_scratch, O1);
96 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type);
97 __ delayed()->mov(G4_scratch, O2);
98 __ mov(L7_thread_cache, G2_thread);
99 __ reset_last_Java_frame();
100
101 // load the register arguments (the C code packed them as varargs)
102 for (Argument ldarg = argv.successor(); ldarg.is_register(); ldarg = ldarg.successor()) {
103 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register());
104 }
105 __ ret();
106 __ delayed()->
107 restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler
108 return entry;
109 }
110
111
112 #else
113 // LP64 passes floating point arguments in F1, F3, F5, etc. instead of
114 // O0, O1, O2 etc..
115 // Doubles are passed in D0, D2, D4
116 // We store the signature of the first 16 arguments in the first argument
117 // slot because it will be overwritten prior to calling the native
118 // function, with the pointer to the JNIEnv.
119 // If LP64 there can be up to 16 floating point arguments in registers
120 // or 6 integer registers.
121 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
122
123 enum {
124 non_float = 0,
125 float_sig = 1,
126 double_sig = 2,
127 sig_mask = 3
128 };
129
130 address entry = __ pc();
131 Argument argv(0, true);
132
133 // We are in the jni transition frame. Save the last_java_frame corresponding to the
134 // outer interpreter frame
135 //
136 __ set_last_Java_frame(FP, noreg);
137 // make sure the interpreter frame we've pushed has a valid return pc
138 __ mov(O7, I7);
139 __ mov(Lmethod, G3_scratch);
140 __ mov(Llocals, G4_scratch);
141 __ save_frame(0);
142 __ mov(G2_thread, L7_thread_cache);
143 __ add(argv.address_in_frame(), O3);
144 __ mov(G2_thread, O0);
145 __ mov(G3_scratch, O1);
146 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type);
147 __ delayed()->mov(G4_scratch, O2);
148 __ mov(L7_thread_cache, G2_thread);
149 __ reset_last_Java_frame();
150
151
152 // load the register arguments (the C code packed them as varargs)
153 Address Sig = argv.address_in_frame(); // Argument 0 holds the signature
154 __ ld_ptr( Sig, G3_scratch ); // Get register argument signature word into G3_scratch
155 __ mov( G3_scratch, G4_scratch);
156 __ srl( G4_scratch, 2, G4_scratch); // Skip Arg 0
157 Label done;
158 for (Argument ldarg = argv.successor(); ldarg.is_float_register(); ldarg = ldarg.successor()) {
159 Label NonFloatArg;
160 Label LoadFloatArg;
161 Label LoadDoubleArg;
162 Label NextArg;
163 Address a = ldarg.address_in_frame();
164 __ andcc(G4_scratch, sig_mask, G3_scratch);
165 __ br(Assembler::zero, false, Assembler::pt, NonFloatArg);
166 __ delayed()->nop();
167
168 __ cmp(G3_scratch, float_sig );
169 __ br(Assembler::equal, false, Assembler::pt, LoadFloatArg);
170 __ delayed()->nop();
171
172 __ cmp(G3_scratch, double_sig );
173 __ br(Assembler::equal, false, Assembler::pt, LoadDoubleArg);
174 __ delayed()->nop();
175
176 __ bind(NonFloatArg);
177 // There are only 6 integer register arguments!
178 if ( ldarg.is_register() )
179 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register());
180 else {
181 // Optimization, see if there are any more args and get out prior to checking
182 // all 16 float registers. My guess is that this is rare.
183 // If is_register is false, then we are done the first six integer args.
184 __ br_null_short(G4_scratch, Assembler::pt, done);
185 }
186 __ ba(NextArg);
187 __ delayed()->srl( G4_scratch, 2, G4_scratch );
188
189 __ bind(LoadFloatArg);
190 __ ldf( FloatRegisterImpl::S, a, ldarg.as_float_register(), 4);
191 __ ba(NextArg);
192 __ delayed()->srl( G4_scratch, 2, G4_scratch );
193
194 __ bind(LoadDoubleArg);
195 __ ldf( FloatRegisterImpl::D, a, ldarg.as_double_register() );
196 __ ba(NextArg);
197 __ delayed()->srl( G4_scratch, 2, G4_scratch );
198
199 __ bind(NextArg);
200
201 }
202
203 __ bind(done);
204 __ ret();
205 __ delayed()->
206 restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler
207 return entry;
208 }
209 #endif
210
211 void TemplateInterpreterGenerator::generate_counter_overflow(Label& Lcontinue) {
212
213 // Generate code to initiate compilation on the counter overflow.
214
215 // InterpreterRuntime::frequency_counter_overflow takes two arguments,
216 // the first indicates if the counter overflow occurs at a backwards branch (NULL bcp)
217 // and the second is only used when the first is true. We pass zero for both.
218 // The call returns the address of the verified entry point for the method or NULL
219 // if the compilation did not complete (either went background or bailed out).
220 __ set((int)false, O2);
221 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), O2, O2, true);
222 // returns verified_entry_point or NULL
223 // we ignore it in any case
224 __ ba_short(Lcontinue);
225
226 }
227
228
229 // End of helpers
230
231 // Various method entries
232
233 // Abstract method entry
234 // Attempt to execute abstract method. Throw exception
235 //
236 address TemplateInterpreterGenerator::generate_abstract_entry(void) {
237 address entry = __ pc();
238 // abstract method entry
239 // throw exception
240 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError));
241 // the call_VM checks for exception, so we should never return here.
242 __ should_not_reach_here();
243 return entry;
244
245 }
246
247 void TemplateInterpreterGenerator::save_native_result(void) {
248 // result potentially in O0/O1: save it across calls
249 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
250
251 // result potentially in F0/F1: save it across calls
252 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
253
254 // save and restore any potential method result value around the unlocking operation
255 __ stf(FloatRegisterImpl::D, F0, d_tmp);
256 #ifdef _LP64
257 __ stx(O0, l_tmp);
258 #else
259 __ std(O0, l_tmp);
260 #endif
261 }
262
263 void TemplateInterpreterGenerator::restore_native_result(void) {
264 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
265 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
266
267 // Restore any method result value
268 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
269 #ifdef _LP64
270 __ ldx(l_tmp, O0);
271 #else
272 __ ldd(l_tmp, O0);
273 #endif
274 }
275
276 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
277 assert(!pass_oop || message == NULL, "either oop or message but not both");
278 address entry = __ pc();
279 // expression stack must be empty before entering the VM if an exception happened
280 __ empty_expression_stack();
281 // load exception object
282 __ set((intptr_t)name, G3_scratch);
283 if (pass_oop) {
284 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
285 } else {
286 __ set((intptr_t)message, G4_scratch);
287 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
288 }
289 // throw exception
290 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
291 AddressLiteral thrower(Interpreter::throw_exception_entry());
292 __ jump_to(thrower, G3_scratch);
293 __ delayed()->nop();
294 return entry;
295 }
296
297 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
298 address entry = __ pc();
299 // expression stack must be empty before entering the VM if an exception
300 // happened
301 __ empty_expression_stack();
302 // load exception object
303 __ call_VM(Oexception,
304 CAST_FROM_FN_PTR(address,
305 InterpreterRuntime::throw_ClassCastException),
306 Otos_i);
307 __ should_not_reach_here();
308 return entry;
309 }
310
311
312 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
313 address entry = __ pc();
314 // expression stack must be empty before entering the VM if an exception happened
315 __ empty_expression_stack();
316 // convention: expect aberrant index in register G3_scratch, then shuffle the
317 // index to G4_scratch for the VM call
318 __ mov(G3_scratch, G4_scratch);
319 __ set((intptr_t)name, G3_scratch);
320 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
321 __ should_not_reach_here();
322 return entry;
323 }
324
325
326 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
327 address entry = __ pc();
328 // expression stack must be empty before entering the VM if an exception happened
329 __ empty_expression_stack();
330 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
331 __ should_not_reach_here();
332 return entry;
333 }
334
335
336 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
337 address entry = __ pc();
338
339 if (state == atos) {
340 __ profile_return_type(O0, G3_scratch, G1_scratch);
341 }
342
343 #if !defined(_LP64) && defined(COMPILER2)
344 // All return values are where we want them, except for Longs. C2 returns
345 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
346 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
347 // build even if we are returning from interpreted we just do a little
348 // stupid shuffing.
349 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
350 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
351 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
352
353 if (state == ltos) {
354 __ srl (G1, 0, O1);
355 __ srlx(G1, 32, O0);
356 }
357 #endif // !_LP64 && COMPILER2
358
359 // The callee returns with the stack possibly adjusted by adapter transition
360 // We remove that possible adjustment here.
361 // All interpreter local registers are untouched. Any result is passed back
362 // in the O0/O1 or float registers. Before continuing, the arguments must be
363 // popped from the java expression stack; i.e., Lesp must be adjusted.
364
365 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
366
367 const Register cache = G3_scratch;
368 const Register index = G1_scratch;
369 __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
370
371 const Register flags = cache;
372 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
373 const Register parameter_size = flags;
374 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words
375 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes
376 __ add(Lesp, parameter_size, Lesp); // pop arguments
377 __ dispatch_next(state, step);
378
379 return entry;
380 }
381
382
383 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
384 address entry = __ pc();
385 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
386 #if INCLUDE_JVMCI
387 // Check if we need to take lock at entry of synchronized method. 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 //
470 // Helpers for commoning out cases in the various type of method entries.
471 //
472
473 // increment invocation count & check for overflow
474 //
475 // Note: checking for negative value instead of overflow
476 // so we have a 'sticky' overflow test
477 //
478 // Lmethod: method
479 // ??: invocation counter
480 //
481 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
482 // Note: In tiered we increment either counters in MethodCounters* or in
483 // MDO depending if we're profiling or not.
484 const Register G3_method_counters = G3_scratch;
485 Label done;
486
487 if (TieredCompilation) {
488 const int increment = InvocationCounter::count_increment;
489 Label no_mdo;
490 if (ProfileInterpreter) {
491 // If no method data exists, go to profile_continue.
492 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
493 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
494 // Increment counter
495 Address mdo_invocation_counter(G4_scratch,
496 in_bytes(MethodData::invocation_counter_offset()) +
497 in_bytes(InvocationCounter::counter_offset()));
498 Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset()));
499 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
500 G3_scratch, Lscratch,
501 Assembler::zero, overflow);
502 __ ba_short(done);
503 }
504
505 // Increment counter in MethodCounters*
506 __ bind(no_mdo);
507 Address invocation_counter(G3_method_counters,
508 in_bytes(MethodCounters::invocation_counter_offset()) +
509 in_bytes(InvocationCounter::counter_offset()));
510 __ get_method_counters(Lmethod, G3_method_counters, done);
511 Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset()));
512 __ increment_mask_and_jump(invocation_counter, increment, mask,
513 G4_scratch, Lscratch,
514 Assembler::zero, overflow);
515 __ bind(done);
516 } else { // not TieredCompilation
517 // Update standard invocation counters
518 __ get_method_counters(Lmethod, G3_method_counters, done);
519 __ increment_invocation_counter(G3_method_counters, O0, G4_scratch);
520 if (ProfileInterpreter) {
521 Address interpreter_invocation_counter(G3_method_counters,
522 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
523 __ ld(interpreter_invocation_counter, G4_scratch);
524 __ inc(G4_scratch);
525 __ st(G4_scratch, interpreter_invocation_counter);
526 }
527
528 if (ProfileInterpreter && profile_method != NULL) {
529 // Test to see if we should create a method data oop
530 Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset()));
531 __ ld(profile_limit, G1_scratch);
532 __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
533
534 // if no method data exists, go to profile_method
535 __ test_method_data_pointer(*profile_method);
536 }
537
538 Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset()));
539 __ ld(invocation_limit, G3_scratch);
540 __ cmp(O0, G3_scratch);
541 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
542 __ delayed()->nop();
543 __ bind(done);
544 }
545
546 }
547
548 // Allocate monitor and lock method (asm interpreter)
549 // ebx - Method*
550 //
551 void TemplateInterpreterGenerator::lock_method() {
552 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
553
554 #ifdef ASSERT
555 { Label ok;
556 __ btst(JVM_ACC_SYNCHRONIZED, O0);
557 __ br( Assembler::notZero, false, Assembler::pt, ok);
558 __ delayed()->nop();
559 __ stop("method doesn't need synchronization");
560 __ bind(ok);
561 }
562 #endif // ASSERT
563
564 // get synchronization object to O0
565 { Label done;
566 __ btst(JVM_ACC_STATIC, O0);
567 __ br( Assembler::zero, true, Assembler::pt, done);
568 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
569
570 // lock the mirror, not the Klass*
571 __ load_mirror(O0, Lmethod);
572
573 #ifdef ASSERT
574 __ tst(O0);
575 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
576 #endif // ASSERT
577
578 __ bind(done);
579 }
580
581 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
582 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
583 // __ untested("lock_object from method entry");
584 __ lock_object(Lmonitors, O0);
585 }
586
587 // See if we've got enough room on the stack for locals plus overhead below
588 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
589 // without going through the signal handler, i.e., reserved and yellow zones
590 // will not be made usable. The shadow zone must suffice to handle the
591 // overflow.
592 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
593 Register Rscratch) {
594 const int page_size = os::vm_page_size();
595 Label after_frame_check;
596
597 assert_different_registers(Rframe_size, Rscratch);
598
599 __ set(page_size, Rscratch);
600 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
601
602 // Get the stack overflow limit, and in debug, verify it is non-zero.
603 __ ld_ptr(G2_thread, JavaThread::stack_overflow_limit_offset(), Rscratch);
604 #ifdef ASSERT
605 Label limit_ok;
606 __ br_notnull_short(Rscratch, Assembler::pn, limit_ok);
607 __ stop("stack overflow limit is zero in generate_stack_overflow_check");
608 __ bind(limit_ok);
609 #endif
610
611 // Add in the size of the frame (which is the same as subtracting it from the
612 // SP, which would take another register.
613 __ add(Rscratch, Rframe_size, Rscratch);
614
615 // The frame is greater than one page in size, so check against
616 // the bottom of the stack.
617 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
618
619 // The stack will overflow, throw an exception.
620
621 // Note that SP is restored to sender's sp (in the delay slot). This
622 // is necessary if the sender's frame is an extended compiled frame
623 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
624 // adaptations.
625
626 // Note also that the restored frame is not necessarily interpreted.
627 // Use the shared runtime version of the StackOverflowError.
628 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
629 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
630 __ jump_to(stub, Rscratch);
631 __ delayed()->mov(O5_savedSP, SP);
632
633 // If you get to here, then there is enough stack space.
634 __ bind(after_frame_check);
635 }
636
637
638 //
639 // Generate a fixed interpreter frame. This is identical setup for interpreted
640 // methods and for native methods hence the shared code.
641
642
643 //----------------------------------------------------------------------------------------------------
644 // Stack frame layout
645 //
646 // When control flow reaches any of the entry types for the interpreter
647 // the following holds ->
648 //
649 // C2 Calling Conventions:
650 //
651 // The entry code below assumes that the following registers are set
652 // when coming in:
653 // G5_method: holds the Method* of the method to call
654 // Lesp: points to the TOS of the callers expression stack
655 // after having pushed all the parameters
656 //
657 // The entry code does the following to setup an interpreter frame
658 // pop parameters from the callers stack by adjusting Lesp
659 // set O0 to Lesp
660 // compute X = (max_locals - num_parameters)
661 // bump SP up by X to accomadate the extra locals
662 // compute X = max_expression_stack
663 // + vm_local_words
664 // + 16 words of register save area
665 // save frame doing a save sp, -X, sp growing towards lower addresses
666 // set Lbcp, Lmethod, LcpoolCache
667 // set Llocals to i0
668 // set Lmonitors to FP - rounded_vm_local_words
669 // set Lesp to Lmonitors - 4
670 //
671 // The frame has now been setup to do the rest of the entry code
672
673 // Try this optimization: Most method entries could live in a
674 // "one size fits all" stack frame without all the dynamic size
675 // calculations. It might be profitable to do all this calculation
676 // statically and approximately for "small enough" methods.
677
678 //-----------------------------------------------------------------------------------------------
679
680 // C1 Calling conventions
681 //
682 // Upon method entry, the following registers are setup:
683 //
684 // g2 G2_thread: current thread
685 // g5 G5_method: method to activate
686 // g4 Gargs : pointer to last argument
687 //
688 //
689 // Stack:
690 //
691 // +---------------+ <--- sp
692 // | |
693 // : reg save area :
694 // | |
695 // +---------------+ <--- sp + 0x40
696 // | |
697 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
698 // | |
699 // +---------------+ <--- sp + 0x5c
700 // | |
701 // : free :
702 // | |
703 // +---------------+ <--- Gargs
704 // | |
705 // : arguments :
706 // | |
707 // +---------------+
708 // | |
709 //
710 //
711 //
712 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
713 //
714 // +---------------+ <--- sp
715 // | |
716 // : reg save area :
717 // | |
718 // +---------------+ <--- sp + 0x40
719 // | |
720 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
721 // | |
722 // +---------------+ <--- sp + 0x5c
723 // | |
724 // : :
725 // | | <--- Lesp
726 // +---------------+ <--- Lmonitors (fp - 0x18)
727 // | VM locals |
728 // +---------------+ <--- fp
729 // | |
730 // : reg save area :
731 // | |
732 // +---------------+ <--- fp + 0x40
733 // | |
734 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
735 // | |
736 // +---------------+ <--- fp + 0x5c
737 // | |
738 // : free :
739 // | |
740 // +---------------+
741 // | |
742 // : nonarg locals :
743 // | |
744 // +---------------+
745 // | |
746 // : arguments :
747 // | | <--- Llocals
748 // +---------------+ <--- Gargs
749 // | |
750
751 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
752 //
753 //
754 // The entry code sets up a new interpreter frame in 4 steps:
755 //
756 // 1) Increase caller's SP by for the extra local space needed:
757 // (check for overflow)
758 // Efficient implementation of xload/xstore bytecodes requires
759 // that arguments and non-argument locals are in a contigously
760 // addressable memory block => non-argument locals must be
761 // allocated in the caller's frame.
762 //
763 // 2) Create a new stack frame and register window:
764 // The new stack frame must provide space for the standard
765 // register save area, the maximum java expression stack size,
766 // the monitor slots (0 slots initially), and some frame local
767 // scratch locations.
768 //
769 // 3) The following interpreter activation registers must be setup:
770 // Lesp : expression stack pointer
771 // Lbcp : bytecode pointer
772 // Lmethod : method
773 // Llocals : locals pointer
774 // Lmonitors : monitor pointer
775 // LcpoolCache: constant pool cache
776 //
777 // 4) Initialize the non-argument locals if necessary:
778 // Non-argument locals may need to be initialized to NULL
779 // for GC to work. If the oop-map information is accurate
780 // (in the absence of the JSR problem), no initialization
781 // is necessary.
782 //
783 // (gri - 2/25/2000)
784
785
786 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
787
788 const int extra_space =
789 rounded_vm_local_words + // frame local scratch space
790 Method::extra_stack_entries() + // extra stack for jsr 292
791 frame::memory_parameter_word_sp_offset + // register save area
792 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
793
794 const Register Glocals_size = G3;
795 const Register RconstMethod = Glocals_size;
796 const Register Otmp1 = O3;
797 const Register Otmp2 = O4;
798 // Lscratch can't be used as a temporary because the call_stub uses
799 // it to assert that the stack frame was setup correctly.
800 const Address constMethod (G5_method, Method::const_offset());
801 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
802
803 __ ld_ptr( constMethod, RconstMethod );
804 __ lduh( size_of_parameters, Glocals_size);
805
806 // Gargs points to first local + BytesPerWord
807 // Set the saved SP after the register window save
808 //
809 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
810 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
811 __ add(Gargs, Otmp1, Gargs);
812
813 if (native_call) {
814 __ calc_mem_param_words( Glocals_size, Gframe_size );
815 __ add( Gframe_size, extra_space, Gframe_size);
816 __ round_to( Gframe_size, WordsPerLong );
817 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
818
819 // Native calls don't need the stack size check since they have no
820 // expression stack and the arguments are already on the stack and
821 // we only add a handful of words to the stack.
822 } else {
823
824 //
825 // Compute number of locals in method apart from incoming parameters
826 //
827 const Address size_of_locals(Otmp1, ConstMethod::size_of_locals_offset());
828 __ ld_ptr(constMethod, Otmp1);
829 __ lduh(size_of_locals, Otmp1);
830 __ sub(Otmp1, Glocals_size, Glocals_size);
831 __ round_to(Glocals_size, WordsPerLong);
832 __ sll(Glocals_size, Interpreter::logStackElementSize, Glocals_size);
833
834 // See if the frame is greater than one page in size. If so,
835 // then we need to verify there is enough stack space remaining.
836 // Frame_size = (max_stack + extra_space) * BytesPerWord;
837 __ ld_ptr(constMethod, Gframe_size);
838 __ lduh(Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size);
839 __ add(Gframe_size, extra_space, Gframe_size);
840 __ round_to(Gframe_size, WordsPerLong);
841 __ sll(Gframe_size, Interpreter::logStackElementSize, Gframe_size);
842
843 // Add in java locals size for stack overflow check only
844 __ add(Gframe_size, Glocals_size, Gframe_size);
845
846 const Register Otmp2 = O4;
847 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
848 generate_stack_overflow_check(Gframe_size, Otmp1);
849
850 __ sub(Gframe_size, Glocals_size, Gframe_size);
851
852 //
853 // bump SP to accomodate the extra locals
854 //
855 __ sub(SP, Glocals_size, SP);
856 }
857
858 //
859 // now set up a stack frame with the size computed above
860 //
861 __ neg( Gframe_size );
862 __ save( SP, Gframe_size, SP );
863
864 //
865 // now set up all the local cache registers
866 //
867 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
868 // that all present references to Lbyte_code initialize the register
869 // immediately before use
870 if (native_call) {
871 __ mov(G0, Lbcp);
872 } else {
873 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
874 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
875 }
876 __ mov( G5_method, Lmethod); // set Lmethod
877 // Get mirror and store it in the frame as GC root for this Method*
878 Register mirror = LcpoolCache;
879 __ load_mirror(mirror, Lmethod);
880 __ st_ptr(mirror, FP, (frame::interpreter_frame_mirror_offset * wordSize) + STACK_BIAS);
881 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
882 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
883 #ifdef _LP64
884 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
885 #endif
886 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
887
888 // setup interpreter activation registers
889 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
890
891 if (ProfileInterpreter) {
892 #ifdef FAST_DISPATCH
893 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
894 // they both use I2.
895 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
896 #endif // FAST_DISPATCH
897 __ set_method_data_pointer();
898 }
899
900 }
901
902 // Method entry for java.lang.ref.Reference.get.
903 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
904 #if INCLUDE_ALL_GCS
905 // Code: _aload_0, _getfield, _areturn
906 // parameter size = 1
907 //
908 // The code that gets generated by this routine is split into 2 parts:
909 // 1. The "intrinsified" code for G1 (or any SATB based GC),
910 // 2. The slow path - which is an expansion of the regular method entry.
911 //
912 // Notes:-
913 // * In the G1 code we do not check whether we need to block for
914 // a safepoint. If G1 is enabled then we must execute the specialized
915 // code for Reference.get (except when the Reference object is null)
916 // so that we can log the value in the referent field with an SATB
917 // update buffer.
918 // If the code for the getfield template is modified so that the
919 // G1 pre-barrier code is executed when the current method is
920 // Reference.get() then going through the normal method entry
921 // will be fine.
922 // * The G1 code can, however, check the receiver object (the instance
923 // of java.lang.Reference) and jump to the slow path if null. If the
924 // Reference object is null then we obviously cannot fetch the referent
925 // and so we don't need to call the G1 pre-barrier. Thus we can use the
926 // regular method entry code to generate the NPE.
927 //
928 // This code is based on generate_accessor_enty.
929
930 address entry = __ pc();
931
932 const int referent_offset = java_lang_ref_Reference::referent_offset;
933 guarantee(referent_offset > 0, "referent offset not initialized");
934
935 if (UseG1GC) {
936 Label slow_path;
937
938 // In the G1 code we don't check if we need to reach a safepoint. We
939 // continue and the thread will safepoint at the next bytecode dispatch.
940
941 // Check if local 0 != NULL
942 // If the receiver is null then it is OK to jump to the slow path.
943 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
944 // check if local 0 == NULL and go the slow path
945 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
946
947
948 // Load the value of the referent field.
949 if (Assembler::is_simm13(referent_offset)) {
950 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
951 } else {
952 __ set(referent_offset, G3_scratch);
953 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
954 }
955
956 // Generate the G1 pre-barrier code to log the value of
957 // the referent field in an SATB buffer. Note with
958 // these parameters the pre-barrier does not generate
959 // the load of the previous value
960
961 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
962 Otos_i /* pre_val */,
963 G3_scratch /* tmp */,
964 true /* preserve_o_regs */);
965
966 // _areturn
967 __ retl(); // return from leaf routine
968 __ delayed()->mov(O5_savedSP, SP);
969
970 // Generate regular method entry
971 __ bind(slow_path);
972 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
973 return entry;
974 }
975 #endif // INCLUDE_ALL_GCS
976
977 // If G1 is not enabled then attempt to go through the accessor entry point
978 // Reference.get is an accessor
979 return NULL;
980 }
981
982 /**
983 * Method entry for static native methods:
984 * int java.util.zip.CRC32.update(int crc, int b)
985 */
986 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
987
988 if (UseCRC32Intrinsics) {
989 address entry = __ pc();
990
991 Label L_slow_path;
992 // If we need a safepoint check, generate full interpreter entry.
993 ExternalAddress state(SafepointSynchronize::address_of_state());
994 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
995 __ set(SafepointSynchronize::_not_synchronized, O3);
996 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
997
998 // Load parameters
999 const Register crc = O0; // initial crc
1000 const Register val = O1; // byte to update with
1001 const Register table = O2; // address of 256-entry lookup table
1002
1003 __ ldub(Gargs, 3, val);
1004 __ lduw(Gargs, 8, crc);
1005
1006 __ set(ExternalAddress(StubRoutines::crc_table_addr()), table);
1007
1008 __ not1(crc); // ~crc
1009 __ clruwu(crc);
1010 __ update_byte_crc32(crc, val, table);
1011 __ not1(crc); // ~crc
1012
1013 // result in O0
1014 __ retl();
1015 __ delayed()->nop();
1016
1017 // generate a vanilla native entry as the slow path
1018 __ bind(L_slow_path);
1019 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
1020 return entry;
1021 }
1022 return NULL;
1023 }
1024
1025 /**
1026 * Method entry for static native methods:
1027 * int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len)
1028 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
1029 */
1030 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1031
1032 if (UseCRC32Intrinsics) {
1033 address entry = __ pc();
1034
1035 Label L_slow_path;
1036 // If we need a safepoint check, generate full interpreter entry.
1037 ExternalAddress state(SafepointSynchronize::address_of_state());
1038 __ set(ExternalAddress(SafepointSynchronize::address_of_state()), O2);
1039 __ set(SafepointSynchronize::_not_synchronized, O3);
1040 __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pt, L_slow_path);
1041
1042 // Load parameters from the stack
1043 const Register crc = O0; // initial crc
1044 const Register buf = O1; // source java byte array address
1045 const Register len = O2; // len
1046 const Register offset = O3; // offset
1047
1048 // Arguments are reversed on java expression stack
1049 // Calculate address of start element
1050 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) {
1051 __ lduw(Gargs, 0, len);
1052 __ lduw(Gargs, 8, offset);
1053 __ ldx( Gargs, 16, buf);
1054 __ lduw(Gargs, 32, crc);
1055 __ add(buf, offset, buf);
1056 } else {
1057 __ lduw(Gargs, 0, len);
1058 __ lduw(Gargs, 8, offset);
1059 __ ldx( Gargs, 16, buf);
1060 __ lduw(Gargs, 24, crc);
1061 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1062 __ add(buf ,offset, buf);
1063 }
1064
1065 // Call the crc32 kernel
1066 __ MacroAssembler::save_thread(L7_thread_cache);
1067 __ kernel_crc32(crc, buf, len, O3);
1068 __ MacroAssembler::restore_thread(L7_thread_cache);
1069
1070 // result in O0
1071 __ retl();
1072 __ delayed()->nop();
1073
1074 // generate a vanilla native entry as the slow path
1075 __ bind(L_slow_path);
1076 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
1077 return entry;
1078 }
1079 return NULL;
1080 }
1081
1082 /**
1083 * Method entry for intrinsic-candidate (non-native) methods:
1084 * int java.util.zip.CRC32C.updateBytes(int crc, byte[] b, int off, int end)
1085 * int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
1086 * Unlike CRC32, CRC32C does not have any methods marked as native
1087 * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
1088 */
1089 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1090
1091 if (UseCRC32CIntrinsics) {
1092 address entry = __ pc();
1093
1094 // Load parameters from the stack
1095 const Register crc = O0; // initial crc
1096 const Register buf = O1; // source java byte array address
1097 const Register offset = O2; // offset
1098 const Register end = O3; // index of last element to process
1099 const Register len = O2; // len argument to the kernel
1100 const Register table = O3; // crc32c lookup table address
1101
1102 // Arguments are reversed on java expression stack
1103 // Calculate address of start element
1104 if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) {
1105 __ lduw(Gargs, 0, end);
1106 __ lduw(Gargs, 8, offset);
1107 __ ldx( Gargs, 16, buf);
1108 __ lduw(Gargs, 32, crc);
1109 __ add(buf, offset, buf);
1110 __ sub(end, offset, len);
1111 } else {
1112 __ lduw(Gargs, 0, end);
1113 __ lduw(Gargs, 8, offset);
1114 __ ldx( Gargs, 16, buf);
1115 __ lduw(Gargs, 24, crc);
1116 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1117 __ add(buf, offset, buf);
1118 __ sub(end, offset, len);
1119 }
1120
1121 // Call the crc32c kernel
1122 __ MacroAssembler::save_thread(L7_thread_cache);
1123 __ kernel_crc32c(crc, buf, len, table);
1124 __ MacroAssembler::restore_thread(L7_thread_cache);
1125
1126 // result in O0
1127 __ retl();
1128 __ delayed()->nop();
1129
1130 return entry;
1131 }
1132 return NULL;
1133 }
1134
1135 // Not supported
1136 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
1137 return NULL;
1138 }
1139
1140 // TODO: rather than touching all pages, check against stack_overflow_limit and bang yellow page to
1141 // generate exception
1142 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1143 // Quick & dirty stack overflow checking: bang the stack & handle trap.
1144 // Note that we do the banging after the frame is setup, since the exception
1145 // handling code expects to find a valid interpreter frame on the stack.
1146 // Doing the banging earlier fails if the caller frame is not an interpreter
1147 // frame.
1148 // (Also, the exception throwing code expects to unlock any synchronized
1149 // method receiever, so do the banging after locking the receiver.)
1150
1151 // Bang each page in the shadow zone. We can't assume it's been done for
1152 // an interpreter frame with greater than a page of locals, so each page
1153 // needs to be checked. Only true for non-native.
1154 if (UseStackBanging) {
1155 const int page_size = os::vm_page_size();
1156 const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1157 const int start_page = native_call ? n_shadow_pages : 1;
1158 for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1159 __ bang_stack_with_offset(pages*page_size);
1160 }
1161 }
1162 }
1163
1164 //
1165 // Interpreter stub for calling a native method. (asm interpreter)
1166 // This sets up a somewhat different looking stack for calling the native method
1167 // than the typical interpreter frame setup.
1168 //
1169
1170 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1171 address entry = __ pc();
1172
1173 // the following temporary registers are used during frame creation
1174 const Register Gtmp1 = G3_scratch ;
1175 const Register Gtmp2 = G1_scratch;
1176 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1177
1178 // make sure registers are different!
1179 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1180
1181 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
1182
1183 const Register Glocals_size = G3;
1184 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1185
1186 // make sure method is native & not abstract
1187 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1188 #ifdef ASSERT
1189 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1190 {
1191 Label L;
1192 __ btst(JVM_ACC_NATIVE, Gtmp1);
1193 __ br(Assembler::notZero, false, Assembler::pt, L);
1194 __ delayed()->nop();
1195 __ stop("tried to execute non-native method as native");
1196 __ bind(L);
1197 }
1198 { Label L;
1199 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1200 __ br(Assembler::zero, false, Assembler::pt, L);
1201 __ delayed()->nop();
1202 __ stop("tried to execute abstract method as non-abstract");
1203 __ bind(L);
1204 }
1205 #endif // ASSERT
1206
1207 // generate the code to allocate the interpreter stack frame
1208 generate_fixed_frame(true);
1209
1210 //
1211 // No locals to initialize for native method
1212 //
1213
1214 // this slot will be set later, we initialize it to null here just in
1215 // case we get a GC before the actual value is stored later
1216 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
1217
1218 const Address do_not_unlock_if_synchronized(G2_thread,
1219 JavaThread::do_not_unlock_if_synchronized_offset());
1220 // Since at this point in the method invocation the exception handler
1221 // would try to exit the monitor of synchronized methods which hasn't
1222 // been entered yet, we set the thread local variable
1223 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1224 // runtime, exception handling i.e. unlock_if_synchronized_method will
1225 // check this thread local flag.
1226 // This flag has two effects, one is to force an unwind in the topmost
1227 // interpreter frame and not perform an unlock while doing so.
1228
1229 __ movbool(true, G3_scratch);
1230 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1231
1232 // increment invocation counter and check for overflow
1233 //
1234 // Note: checking for negative value instead of overflow
1235 // so we have a 'sticky' overflow test (may be of
1236 // importance as soon as we have true MT/MP)
1237 Label invocation_counter_overflow;
1238 Label Lcontinue;
1239 if (inc_counter) {
1240 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1241
1242 }
1243 __ bind(Lcontinue);
1244
1245 bang_stack_shadow_pages(true);
1246
1247 // reset the _do_not_unlock_if_synchronized flag
1248 __ stbool(G0, do_not_unlock_if_synchronized);
1249
1250 // check for synchronized methods
1251 // Must happen AFTER invocation_counter check and stack overflow check,
1252 // so method is not locked if overflows.
1253
1254 if (synchronized) {
1255 lock_method();
1256 } else {
1257 #ifdef ASSERT
1258 { Label ok;
1259 __ ld(Laccess_flags, O0);
1260 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1261 __ br( Assembler::zero, false, Assembler::pt, ok);
1262 __ delayed()->nop();
1263 __ stop("method needs synchronization");
1264 __ bind(ok);
1265 }
1266 #endif // ASSERT
1267 }
1268
1269
1270 // start execution
1271 __ verify_thread();
1272
1273 // JVMTI support
1274 __ notify_method_entry();
1275
1276 // native call
1277
1278 // (note that O0 is never an oop--at most it is a handle)
1279 // It is important not to smash any handles created by this call,
1280 // until any oop handle in O0 is dereferenced.
1281
1282 // (note that the space for outgoing params is preallocated)
1283
1284 // get signature handler
1285 { Label L;
1286 Address signature_handler(Lmethod, Method::signature_handler_offset());
1287 __ ld_ptr(signature_handler, G3_scratch);
1288 __ br_notnull_short(G3_scratch, Assembler::pt, L);
1289 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
1290 __ ld_ptr(signature_handler, G3_scratch);
1291 __ bind(L);
1292 }
1293
1294 // Push a new frame so that the args will really be stored in
1295 // Copy a few locals across so the new frame has the variables
1296 // we need but these values will be dead at the jni call and
1297 // therefore not gc volatile like the values in the current
1298 // frame (Lmethod in particular)
1299
1300 // Flush the method pointer to the register save area
1301 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
1302 __ mov(Llocals, O1);
1303
1304 // calculate where the mirror handle body is allocated in the interpreter frame:
1305 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
1306
1307 // Calculate current frame size
1308 __ sub(SP, FP, O3); // Calculate negative of current frame size
1309 __ save(SP, O3, SP); // Allocate an identical sized frame
1310
1311 // Note I7 has leftover trash. Slow signature handler will fill it in
1312 // should we get there. Normal jni call will set reasonable last_Java_pc
1313 // below (and fix I7 so the stack trace doesn't have a meaningless frame
1314 // in it).
1315
1316 // Load interpreter frame's Lmethod into same register here
1317
1318 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1319
1320 __ mov(I1, Llocals);
1321 __ mov(I2, Lscratch2); // save the address of the mirror
1322
1323
1324 // ONLY Lmethod and Llocals are valid here!
1325
1326 // call signature handler, It will move the arg properly since Llocals in current frame
1327 // matches that in outer frame
1328
1329 __ callr(G3_scratch, 0);
1330 __ delayed()->nop();
1331
1332 // Result handler is in Lscratch
1333
1334 // Reload interpreter frame's Lmethod since slow signature handler may block
1335 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1336
1337 { Label not_static;
1338
1339 __ ld(Laccess_flags, O0);
1340 __ btst(JVM_ACC_STATIC, O0);
1341 __ br( Assembler::zero, false, Assembler::pt, not_static);
1342 // get native function entry point(O0 is a good temp until the very end)
1343 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
1344 // for static methods insert the mirror argument
1345 __ load_mirror(O1, Lmethod);
1346 #ifdef ASSERT
1347 if (!PrintSignatureHandlers) // do not dirty the output with this
1348 { Label L;
1349 __ br_notnull_short(O1, Assembler::pt, L);
1350 __ stop("mirror is missing");
1351 __ bind(L);
1352 }
1353 #endif // ASSERT
1354 __ st_ptr(O1, Lscratch2, 0);
1355 __ mov(Lscratch2, O1);
1356 __ bind(not_static);
1357 }
1358
1359 // At this point, arguments have been copied off of stack into
1360 // their JNI positions, which are O1..O5 and SP[68..].
1361 // Oops are boxed in-place on the stack, with handles copied to arguments.
1362 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1363
1364 #ifdef ASSERT
1365 { Label L;
1366 __ br_notnull_short(O0, Assembler::pt, L);
1367 __ stop("native entry point is missing");
1368 __ bind(L);
1369 }
1370 #endif // ASSERT
1371
1372 //
1373 // setup the frame anchor
1374 //
1375 // The scavenge function only needs to know that the PC of this frame is
1376 // in the interpreter method entry code, it doesn't need to know the exact
1377 // PC and hence we can use O7 which points to the return address from the
1378 // previous call in the code stream (signature handler function)
1379 //
1380 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1381 // we have pushed the extra frame in order to protect the volatile register(s)
1382 // in that frame when we return from the jni call
1383 //
1384
1385 __ set_last_Java_frame(FP, O7);
1386 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1387 // not meaningless information that'll confuse me.
1388
1389 // flush the windows now. We don't care about the current (protection) frame
1390 // only the outer frames
1391
1392 __ flushw();
1393
1394 // mark windows as flushed
1395 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1396 __ set(JavaFrameAnchor::flushed, G3_scratch);
1397 __ st(G3_scratch, flags);
1398
1399 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1400
1401 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1402 #ifdef ASSERT
1403 { Label L;
1404 __ ld(thread_state, G3_scratch);
1405 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1406 __ stop("Wrong thread state in native stub");
1407 __ bind(L);
1408 }
1409 #endif // ASSERT
1410 __ set(_thread_in_native, G3_scratch);
1411 __ st(G3_scratch, thread_state);
1412
1413 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1414 __ save_thread(L7_thread_cache); // save Gthread
1415 __ callr(O0, 0);
1416 __ delayed()->
1417 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1418
1419 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1420
1421 __ restore_thread(L7_thread_cache); // restore G2_thread
1422 __ reinit_heapbase();
1423
1424 // must we block?
1425
1426 // Block, if necessary, before resuming in _thread_in_Java state.
1427 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1428 { Label no_block;
1429 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1430
1431 // Switch thread to "native transition" state before reading the synchronization state.
1432 // This additional state is necessary because reading and testing the synchronization
1433 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1434 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1435 // VM thread changes sync state to synchronizing and suspends threads for GC.
1436 // Thread A is resumed to finish this native method, but doesn't block here since it
1437 // didn't see any synchronization is progress, and escapes.
1438 __ set(_thread_in_native_trans, G3_scratch);
1439 __ st(G3_scratch, thread_state);
1440 if(os::is_MP()) {
1441 if (UseMembar) {
1442 // Force this write out before the read below
1443 __ membar(Assembler::StoreLoad);
1444 } else {
1445 // Write serialization page so VM thread can do a pseudo remote membar.
1446 // We use the current thread pointer to calculate a thread specific
1447 // offset to write to within the page. This minimizes bus traffic
1448 // due to cache line collision.
1449 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1450 }
1451 }
1452 __ load_contents(sync_state, G3_scratch);
1453 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1454
1455 Label L;
1456 __ br(Assembler::notEqual, false, Assembler::pn, L);
1457 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1458 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1459 __ bind(L);
1460
1461 // Block. Save any potential method result value before the operation and
1462 // use a leaf call to leave the last_Java_frame setup undisturbed.
1463 save_native_result();
1464 __ call_VM_leaf(L7_thread_cache,
1465 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1466 G2_thread);
1467
1468 // Restore any method result value
1469 restore_native_result();
1470 __ bind(no_block);
1471 }
1472
1473 // Clear the frame anchor now
1474
1475 __ reset_last_Java_frame();
1476
1477 // Move the result handler address
1478 __ mov(Lscratch, G3_scratch);
1479 // return possible result to the outer frame
1480 #ifndef __LP64
1481 __ mov(O0, I0);
1482 __ restore(O1, G0, O1);
1483 #else
1484 __ restore(O0, G0, O0);
1485 #endif /* __LP64 */
1486
1487 // Move result handler to expected register
1488 __ mov(G3_scratch, Lscratch);
1489
1490 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1491 // switch to thread_in_Java.
1492
1493 __ set(_thread_in_Java, G3_scratch);
1494 __ st(G3_scratch, thread_state);
1495
1496 if (CheckJNICalls) {
1497 // clear_pending_jni_exception_check
1498 __ st_ptr(G0, G2_thread, JavaThread::pending_jni_exception_check_fn_offset());
1499 }
1500
1501 // reset handle block
1502 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1503 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1504
1505 // If we have an oop result store it where it will be safe for any further gc
1506 // until we return now that we've released the handle it might be protected by
1507
1508 {
1509 Label no_oop, store_result;
1510
1511 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1512 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1513 __ addcc(G0, O0, O0);
1514 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1515 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1516 __ mov(G0, O0);
1517
1518 __ bind(store_result);
1519 // Store it where gc will look for it and result handler expects it.
1520 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1521
1522 __ bind(no_oop);
1523
1524 }
1525
1526
1527 // handle exceptions (exception handling will handle unlocking!)
1528 { Label L;
1529 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1530 __ ld_ptr(exception_addr, Gtemp);
1531 __ br_null_short(Gtemp, Assembler::pt, L);
1532 // Note: This could be handled more efficiently since we know that the native
1533 // method doesn't have an exception handler. We could directly return
1534 // to the exception handler for the caller.
1535 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1536 __ should_not_reach_here();
1537 __ bind(L);
1538 }
1539
1540 // JVMTI support (preserves thread register)
1541 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1542
1543 if (synchronized) {
1544 // save and restore any potential method result value around the unlocking operation
1545 save_native_result();
1546
1547 __ add( __ top_most_monitor(), O1);
1548 __ unlock_object(O1);
1549
1550 restore_native_result();
1551 }
1552
1553 #if defined(COMPILER2) && !defined(_LP64)
1554
1555 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1556 // or compiled so just be safe.
1557
1558 __ sllx(O0, 32, G1); // Shift bits into high G1
1559 __ srl (O1, 0, O1); // Zero extend O1
1560 __ or3 (O1, G1, G1); // OR 64 bits into G1
1561
1562 #endif /* COMPILER2 && !_LP64 */
1563
1564 // dispose of return address and remove activation
1565 #ifdef ASSERT
1566 {
1567 Label ok;
1568 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1569 __ stop("bad I5_savedSP value");
1570 __ should_not_reach_here();
1571 __ bind(ok);
1572 }
1573 #endif
1574 __ jmp(Lscratch, 0);
1575 __ delayed()->nop();
1576
1577
1578 if (inc_counter) {
1579 // handle invocation counter overflow
1580 __ bind(invocation_counter_overflow);
1581 generate_counter_overflow(Lcontinue);
1582 }
1583
1584
1585
1586 return entry;
1587 }
1588
1589
1590 // Generic method entry to (asm) interpreter
1591 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1592 address entry = __ pc();
1593
1594 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1595
1596 // the following temporary registers are used during frame creation
1597 const Register Gtmp1 = G3_scratch ;
1598 const Register Gtmp2 = G1_scratch;
1599
1600 // make sure registers are different!
1601 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1602
1603 const Address constMethod (G5_method, Method::const_offset());
1604 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1605 // and use in the asserts.
1606 const Address access_flags (Lmethod, Method::access_flags_offset());
1607
1608 const Register Glocals_size = G3;
1609 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1610
1611 // make sure method is not native & not abstract
1612 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1613 #ifdef ASSERT
1614 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1615 {
1616 Label L;
1617 __ btst(JVM_ACC_NATIVE, Gtmp1);
1618 __ br(Assembler::zero, false, Assembler::pt, L);
1619 __ delayed()->nop();
1620 __ stop("tried to execute native method as non-native");
1621 __ bind(L);
1622 }
1623 { Label L;
1624 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1625 __ br(Assembler::zero, false, Assembler::pt, L);
1626 __ delayed()->nop();
1627 __ stop("tried to execute abstract method as non-abstract");
1628 __ bind(L);
1629 }
1630 #endif // ASSERT
1631
1632 // generate the code to allocate the interpreter stack frame
1633
1634 generate_fixed_frame(false);
1635
1636 #ifdef FAST_DISPATCH
1637 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1638 // set bytecode dispatch table base
1639 #endif
1640
1641 //
1642 // Code to initialize the extra (i.e. non-parm) locals
1643 //
1644 Register init_value = noreg; // will be G0 if we must clear locals
1645 // The way the code was setup before zerolocals was always true for vanilla java entries.
1646 // It could only be false for the specialized entries like accessor or empty which have
1647 // no extra locals so the testing was a waste of time and the extra locals were always
1648 // initialized. We removed this extra complication to already over complicated code.
1649
1650 init_value = G0;
1651 Label clear_loop;
1652
1653 const Register RconstMethod = O1;
1654 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1655 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1656
1657 // NOTE: If you change the frame layout, this code will need to
1658 // be updated!
1659 __ ld_ptr( constMethod, RconstMethod );
1660 __ lduh( size_of_locals, O2 );
1661 __ lduh( size_of_parameters, O1 );
1662 __ sll( O2, Interpreter::logStackElementSize, O2);
1663 __ sll( O1, Interpreter::logStackElementSize, O1 );
1664 __ sub( Llocals, O2, O2 );
1665 __ sub( Llocals, O1, O1 );
1666
1667 __ bind( clear_loop );
1668 __ inc( O2, wordSize );
1669
1670 __ cmp( O2, O1 );
1671 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1672 __ delayed()->st_ptr( init_value, O2, 0 );
1673
1674 const Address do_not_unlock_if_synchronized(G2_thread,
1675 JavaThread::do_not_unlock_if_synchronized_offset());
1676 // Since at this point in the method invocation the exception handler
1677 // would try to exit the monitor of synchronized methods which hasn't
1678 // been entered yet, we set the thread local variable
1679 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1680 // runtime, exception handling i.e. unlock_if_synchronized_method will
1681 // check this thread local flag.
1682 __ movbool(true, G3_scratch);
1683 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1684
1685 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1686 // increment invocation counter and check for overflow
1687 //
1688 // Note: checking for negative value instead of overflow
1689 // so we have a 'sticky' overflow test (may be of
1690 // importance as soon as we have true MT/MP)
1691 Label invocation_counter_overflow;
1692 Label profile_method;
1693 Label profile_method_continue;
1694 Label Lcontinue;
1695 if (inc_counter) {
1696 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1697 if (ProfileInterpreter) {
1698 __ bind(profile_method_continue);
1699 }
1700 }
1701 __ bind(Lcontinue);
1702
1703 bang_stack_shadow_pages(false);
1704
1705 // reset the _do_not_unlock_if_synchronized flag
1706 __ stbool(G0, do_not_unlock_if_synchronized);
1707
1708 // check for synchronized methods
1709 // Must happen AFTER invocation_counter check and stack overflow check,
1710 // so method is not locked if overflows.
1711
1712 if (synchronized) {
1713 lock_method();
1714 } else {
1715 #ifdef ASSERT
1716 { Label ok;
1717 __ ld(access_flags, O0);
1718 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1719 __ br( Assembler::zero, false, Assembler::pt, ok);
1720 __ delayed()->nop();
1721 __ stop("method needs synchronization");
1722 __ bind(ok);
1723 }
1724 #endif // ASSERT
1725 }
1726
1727 // start execution
1728
1729 __ verify_thread();
1730
1731 // jvmti support
1732 __ notify_method_entry();
1733
1734 // start executing instructions
1735 __ dispatch_next(vtos);
1736
1737
1738 if (inc_counter) {
1739 if (ProfileInterpreter) {
1740 // We have decided to profile this method in the interpreter
1741 __ bind(profile_method);
1742
1743 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1744 __ set_method_data_pointer_for_bcp();
1745 __ ba_short(profile_method_continue);
1746 }
1747
1748 // handle invocation counter overflow
1749 __ bind(invocation_counter_overflow);
1750 generate_counter_overflow(Lcontinue);
1751 }
1752
1753
1754 return entry;
1755 }
1756
1757 //----------------------------------------------------------------------------------------------------
1758 // Exceptions
1759 void TemplateInterpreterGenerator::generate_throw_exception() {
1760
1761 // Entry point in previous activation (i.e., if the caller was interpreted)
1762 Interpreter::_rethrow_exception_entry = __ pc();
1763 // O0: exception
1764
1765 // entry point for exceptions thrown within interpreter code
1766 Interpreter::_throw_exception_entry = __ pc();
1767 __ verify_thread();
1768 // expression stack is undefined here
1769 // O0: exception, i.e. Oexception
1770 // Lbcp: exception bcp
1771 __ verify_oop(Oexception);
1772
1773
1774 // expression stack must be empty before entering the VM in case of an exception
1775 __ empty_expression_stack();
1776 // find exception handler address and preserve exception oop
1777 // call C routine to find handler and jump to it
1778 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1779 __ push_ptr(O1); // push exception for exception handler bytecodes
1780
1781 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1782 __ delayed()->nop();
1783
1784
1785 // if the exception is not handled in the current frame
1786 // the frame is removed and the exception is rethrown
1787 // (i.e. exception continuation is _rethrow_exception)
1788 //
1789 // Note: At this point the bci is still the bxi for the instruction which caused
1790 // the exception and the expression stack is empty. Thus, for any VM calls
1791 // at this point, GC will find a legal oop map (with empty expression stack).
1792
1793 // in current activation
1794 // tos: exception
1795 // Lbcp: exception bcp
1796
1797 //
1798 // JVMTI PopFrame support
1799 //
1800
1801 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1802 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1803 // Set the popframe_processing bit in popframe_condition indicating that we are
1804 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1805 // popframe handling cycles.
1806
1807 __ ld(popframe_condition_addr, G3_scratch);
1808 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1809 __ stw(G3_scratch, popframe_condition_addr);
1810
1811 // Empty the expression stack, as in normal exception handling
1812 __ empty_expression_stack();
1813 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1814
1815 {
1816 // Check to see whether we are returning to a deoptimized frame.
1817 // (The PopFrame call ensures that the caller of the popped frame is
1818 // either interpreted or compiled and deoptimizes it if compiled.)
1819 // In this case, we can't call dispatch_next() after the frame is
1820 // popped, but instead must save the incoming arguments and restore
1821 // them after deoptimization has occurred.
1822 //
1823 // Note that we don't compare the return PC against the
1824 // deoptimization blob's unpack entry because of the presence of
1825 // adapter frames in C2.
1826 Label caller_not_deoptimized;
1827 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1828 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1829
1830 const Register Gtmp1 = G3_scratch;
1831 const Register Gtmp2 = G1_scratch;
1832 const Register RconstMethod = Gtmp1;
1833 const Address constMethod(Lmethod, Method::const_offset());
1834 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1835
1836 // Compute size of arguments for saving when returning to deoptimized caller
1837 __ ld_ptr(constMethod, RconstMethod);
1838 __ lduh(size_of_parameters, Gtmp1);
1839 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1840 __ sub(Llocals, Gtmp1, Gtmp2);
1841 __ add(Gtmp2, wordSize, Gtmp2);
1842 // Save these arguments
1843 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1844 // Inform deoptimization that it is responsible for restoring these arguments
1845 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1846 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1847 __ st(Gtmp1, popframe_condition_addr);
1848
1849 // Return from the current method
1850 // The caller's SP was adjusted upon method entry to accomodate
1851 // the callee's non-argument locals. Undo that adjustment.
1852 __ ret();
1853 __ delayed()->restore(I5_savedSP, G0, SP);
1854
1855 __ bind(caller_not_deoptimized);
1856 }
1857
1858 // Clear the popframe condition flag
1859 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1860
1861 // Get out of the current method (how this is done depends on the particular compiler calling
1862 // convention that the interpreter currently follows)
1863 // The caller's SP was adjusted upon method entry to accomodate
1864 // the callee's non-argument locals. Undo that adjustment.
1865 __ restore(I5_savedSP, G0, SP);
1866 // The method data pointer was incremented already during
1867 // call profiling. We have to restore the mdp for the current bcp.
1868 if (ProfileInterpreter) {
1869 __ set_method_data_pointer_for_bcp();
1870 }
1871
1872 #if INCLUDE_JVMTI
1873 {
1874 Label L_done;
1875
1876 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1877 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1878
1879 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1880 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1881
1882 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1883
1884 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1885 __ delayed()->nop();
1886
1887 __ st_ptr(G1_scratch, Lesp, wordSize);
1888 __ bind(L_done);
1889 }
1890 #endif // INCLUDE_JVMTI
1891
1892 // Resume bytecode interpretation at the current bcp
1893 __ dispatch_next(vtos);
1894 // end of JVMTI PopFrame support
1895
1896 Interpreter::_remove_activation_entry = __ pc();
1897
1898 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1899 __ pop_ptr(Oexception); // get exception
1900
1901 // Intel has the following comment:
1902 //// remove the activation (without doing throws on illegalMonitorExceptions)
1903 // They remove the activation without checking for bad monitor state.
1904 // %%% We should make sure this is the right semantics before implementing.
1905
1906 __ set_vm_result(Oexception);
1907 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1908
1909 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1910
1911 __ get_vm_result(Oexception);
1912 __ verify_oop(Oexception);
1913
1914 const int return_reg_adjustment = frame::pc_return_offset;
1915 Address issuing_pc_addr(I7, return_reg_adjustment);
1916
1917 // We are done with this activation frame; find out where to go next.
1918 // The continuation point will be an exception handler, which expects
1919 // the following registers set up:
1920 //
1921 // Oexception: exception
1922 // Oissuing_pc: the local call that threw exception
1923 // Other On: garbage
1924 // In/Ln: the contents of the caller's register window
1925 //
1926 // We do the required restore at the last possible moment, because we
1927 // need to preserve some state across a runtime call.
1928 // (Remember that the caller activation is unknown--it might not be
1929 // interpreted, so things like Lscratch are useless in the caller.)
1930
1931 // Although the Intel version uses call_C, we can use the more
1932 // compact call_VM. (The only real difference on SPARC is a
1933 // harmlessly ignored [re]set_last_Java_frame, compared with
1934 // the Intel code which lacks this.)
1935 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1936 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1937 __ super_call_VM_leaf(L7_thread_cache,
1938 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1939 G2_thread, Oissuing_pc->after_save());
1940
1941 // The caller's SP was adjusted upon method entry to accomodate
1942 // the callee's non-argument locals. Undo that adjustment.
1943 __ JMP(O0, 0); // return exception handler in caller
1944 __ delayed()->restore(I5_savedSP, G0, SP);
1945
1946 // (same old exception object is already in Oexception; see above)
1947 // Note that an "issuing PC" is actually the next PC after the call
1948 }
1949
1950
1951 //
1952 // JVMTI ForceEarlyReturn support
1953 //
1954
1955 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1956 address entry = __ pc();
1957
1958 __ empty_expression_stack();
1959 __ load_earlyret_value(state);
1960
1961 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1962 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1963
1964 // Clear the earlyret state
1965 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1966
1967 __ remove_activation(state,
1968 /* throw_monitor_exception */ false,
1969 /* install_monitor_exception */ false);
1970
1971 // The caller's SP was adjusted upon method entry to accomodate
1972 // the callee's non-argument locals. Undo that adjustment.
1973 __ ret(); // return to caller
1974 __ delayed()->restore(I5_savedSP, G0, SP);
1975
1976 return entry;
1977 } // end of JVMTI ForceEarlyReturn support
1978
1979
1980 //------------------------------------------------------------------------------------------------------------------------
1981 // Helper for vtos entry point generation
1982
1983 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) {
1984 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1985 Label L;
1986 aep = __ pc(); __ push_ptr(); __ ba_short(L);
1987 fep = __ pc(); __ push_f(); __ ba_short(L);
1988 dep = __ pc(); __ push_d(); __ ba_short(L);
1989 lep = __ pc(); __ push_l(); __ ba_short(L);
1990 iep = __ pc(); __ push_i();
1991 bep = cep = sep = iep; // there aren't any
1992 vep = __ pc(); __ bind(L); // fall through
1993 generate_and_dispatch(t);
1994 }
1995
1996 // --------------------------------------------------------------------------------
1997
1998 // Non-product code
1999 #ifndef PRODUCT
2000 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2001 address entry = __ pc();
2002
2003 __ push(state);
2004 __ mov(O7, Lscratch); // protect return address within interpreter
2005
2006 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
2007 __ mov( Otos_l2, G3_scratch );
2008 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
2009 __ mov(Lscratch, O7); // restore return address
2010 __ pop(state);
2011 __ retl();
2012 __ delayed()->nop();
2013
2014 return entry;
2015 }
2016
2017
2018 // helpers for generate_and_dispatch
2019
2020 void TemplateInterpreterGenerator::count_bytecode() {
2021 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2022 }
2023
2024
2025 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2026 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2027 }
2028
2029
2030 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2031 AddressLiteral index (&BytecodePairHistogram::_index);
2032 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2033
2034 // get index, shift out old bytecode, bring in new bytecode, and store it
2035 // _index = (_index >> log2_number_of_codes) |
2036 // (bytecode << log2_number_of_codes);
2037
2038 __ load_contents(index, G4_scratch);
2039 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2040 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
2041 __ or3( G3_scratch, G4_scratch, G4_scratch );
2042 __ store_contents(G4_scratch, index, G3_scratch);
2043
2044 // bump bucket contents
2045 // _counters[_index] ++;
2046
2047 __ set(counters, G3_scratch); // loads into G3_scratch
2048 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
2049 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
2050 __ ld (G3_scratch, 0, G4_scratch);
2051 __ inc (G4_scratch);
2052 __ st (G4_scratch, 0, G3_scratch);
2053 }
2054
2055
2056 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2057 // Call a little run-time stub to avoid blow-up for each bytecode.
2058 // The run-time runtime saves the right registers, depending on
2059 // the tosca in-state for the given template.
2060 address entry = Interpreter::trace_code(t->tos_in());
2061 guarantee(entry != NULL, "entry must have been generated");
2062 __ call(entry, relocInfo::none);
2063 __ delayed()->nop();
2064 }
2065
2066
2067 void TemplateInterpreterGenerator::stop_interpreter_at() {
2068 AddressLiteral counter(&BytecodeCounter::_counter_value);
2069 __ load_contents(counter, G3_scratch);
2070 AddressLiteral stop_at(&StopInterpreterAt);
2071 __ load_ptr_contents(stop_at, G4_scratch);
2072 __ cmp(G3_scratch, G4_scratch);
2073 __ breakpoint_trap(Assembler::equal, Assembler::icc);
2074 }
2075 #endif // not PRODUCT