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/align.hpp"
48 #include "utilities/debug.hpp"
49 #include "utilities/macros.hpp"
50
51 // Size of interpreter code. Increase if too small. Interpreter will
52 // fail with a guarantee ("not enough space for interpreter generation");
53 // if too small.
54 // Run with +PrintInterpreter to get the VM to print out the size.
55 // Max size with JVMTI
56 // The sethi() instruction generates lots more instructions when shell
57 // stack limit is unlimited, so that's why this is much bigger.
58 int TemplateInterpreter::InterpreterCodeSize = 260 * K;
59
60 // Generation of Interpreter
61 //
62 // The TemplateInterpreterGenerator generates the interpreter into Interpreter::_code.
63
64
65 #define __ _masm->
66
67
68 //----------------------------------------------------------------------------------------------------
69
70 // LP64 passes floating point arguments in F1, F3, F5, etc. instead of
71 // O0, O1, O2 etc..
72 // Doubles are passed in D0, D2, D4
73 // We store the signature of the first 16 arguments in the first argument
74 // slot because it will be overwritten prior to calling the native
75 // function, with the pointer to the JNIEnv.
76 // If LP64 there can be up to 16 floating point arguments in registers
77 // or 6 integer registers.
78 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
79
80 enum {
81 non_float = 0,
82 float_sig = 1,
83 double_sig = 2,
84 sig_mask = 3
85 };
86
87 address entry = __ pc();
88 Argument argv(0, true);
89
90 // We are in the jni transition frame. Save the last_java_frame corresponding to the
91 // outer interpreter frame
92 //
93 __ set_last_Java_frame(FP, noreg);
94 // make sure the interpreter frame we've pushed has a valid return pc
95 __ mov(O7, I7);
96 __ mov(Lmethod, G3_scratch);
97 __ mov(Llocals, G4_scratch);
98 __ save_frame(0);
99 __ mov(G2_thread, L7_thread_cache);
100 __ add(argv.address_in_frame(), O3);
101 __ mov(G2_thread, O0);
102 __ mov(G3_scratch, O1);
103 __ call(CAST_FROM_FN_PTR(address, InterpreterRuntime::slow_signature_handler), relocInfo::runtime_call_type);
104 __ delayed()->mov(G4_scratch, O2);
105 __ mov(L7_thread_cache, G2_thread);
106 __ reset_last_Java_frame();
107
108
109 // load the register arguments (the C code packed them as varargs)
110 Address Sig = argv.address_in_frame(); // Argument 0 holds the signature
111 __ ld_ptr( Sig, G3_scratch ); // Get register argument signature word into G3_scratch
112 __ mov( G3_scratch, G4_scratch);
113 __ srl( G4_scratch, 2, G4_scratch); // Skip Arg 0
114 Label done;
115 for (Argument ldarg = argv.successor(); ldarg.is_float_register(); ldarg = ldarg.successor()) {
116 Label NonFloatArg;
117 Label LoadFloatArg;
118 Label LoadDoubleArg;
119 Label NextArg;
120 Address a = ldarg.address_in_frame();
121 __ andcc(G4_scratch, sig_mask, G3_scratch);
122 __ br(Assembler::zero, false, Assembler::pt, NonFloatArg);
123 __ delayed()->nop();
124
125 __ cmp(G3_scratch, float_sig );
126 __ br(Assembler::equal, false, Assembler::pt, LoadFloatArg);
127 __ delayed()->nop();
128
129 __ cmp(G3_scratch, double_sig );
130 __ br(Assembler::equal, false, Assembler::pt, LoadDoubleArg);
131 __ delayed()->nop();
132
133 __ bind(NonFloatArg);
134 // There are only 6 integer register arguments!
135 if ( ldarg.is_register() )
136 __ ld_ptr(ldarg.address_in_frame(), ldarg.as_register());
137 else {
138 // Optimization, see if there are any more args and get out prior to checking
139 // all 16 float registers. My guess is that this is rare.
140 // If is_register is false, then we are done the first six integer args.
141 __ br_null_short(G4_scratch, Assembler::pt, done);
142 }
143 __ ba(NextArg);
144 __ delayed()->srl( G4_scratch, 2, G4_scratch );
145
146 __ bind(LoadFloatArg);
147 __ ldf( FloatRegisterImpl::S, a, ldarg.as_float_register(), 4);
148 __ ba(NextArg);
149 __ delayed()->srl( G4_scratch, 2, G4_scratch );
150
151 __ bind(LoadDoubleArg);
152 __ ldf( FloatRegisterImpl::D, a, ldarg.as_double_register() );
153 __ ba(NextArg);
154 __ delayed()->srl( G4_scratch, 2, G4_scratch );
155
156 __ bind(NextArg);
157 }
158
159 __ bind(done);
160 __ ret();
161 __ delayed()->restore(O0, 0, Lscratch); // caller's Lscratch gets the result handler
162
163 return entry;
164 }
165
166 void TemplateInterpreterGenerator::generate_counter_overflow(Label& Lcontinue) {
167
168 // Generate code to initiate compilation on the counter overflow.
169
170 // InterpreterRuntime::frequency_counter_overflow takes two arguments,
171 // the first indicates if the counter overflow occurs at a backwards branch (NULL bcp)
172 // and the second is only used when the first is true. We pass zero for both.
173 // The call returns the address of the verified entry point for the method or NULL
174 // if the compilation did not complete (either went background or bailed out).
175 __ set((int)false, O2);
176 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), O2, O2, true);
177 // returns verified_entry_point or NULL
178 // we ignore it in any case
179 __ ba_short(Lcontinue);
180 }
181
182
183 // End of helpers
184
185 // Various method entries
186
187 // Abstract method entry
188 // Attempt to execute abstract method. Throw exception
189 //
190 address TemplateInterpreterGenerator::generate_abstract_entry(void) {
191 address entry = __ pc();
192 // abstract method entry
193 // throw exception
194 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError));
195 // the call_VM checks for exception, so we should never return here.
196 __ should_not_reach_here();
197 return entry;
198 }
199
200 void TemplateInterpreterGenerator::save_native_result(void) {
201 // result potentially in O0/O1: save it across calls
202 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
203
204 // result potentially in F0/F1: save it across calls
205 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
206
207 // save and restore any potential method result value around the unlocking operation
208 __ stf(FloatRegisterImpl::D, F0, d_tmp);
209 __ stx(O0, l_tmp);
210 }
211
212 void TemplateInterpreterGenerator::restore_native_result(void) {
213 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
214 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
215
216 // Restore any method result value
217 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
218 __ ldx(l_tmp, O0);
219 }
220
221 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
222 assert(!pass_oop || message == NULL, "either oop or message but not both");
223 address entry = __ pc();
224 // expression stack must be empty before entering the VM if an exception happened
225 __ empty_expression_stack();
226 // load exception object
227 __ set((intptr_t)name, G3_scratch);
228 if (pass_oop) {
229 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
230 } else {
231 __ set((intptr_t)message, G4_scratch);
232 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
233 }
234 // throw exception
235 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
236 AddressLiteral thrower(Interpreter::throw_exception_entry());
237 __ jump_to(thrower, G3_scratch);
238 __ delayed()->nop();
239 return entry;
240 }
241
242 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
243 address entry = __ pc();
244 // expression stack must be empty before entering the VM if an exception
245 // happened
246 __ empty_expression_stack();
247 // load exception object
248 __ call_VM(Oexception,
249 CAST_FROM_FN_PTR(address,
250 InterpreterRuntime::throw_ClassCastException),
251 Otos_i);
252 __ should_not_reach_here();
253 return entry;
254 }
255
256
257 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
258 address entry = __ pc();
259 // expression stack must be empty before entering the VM if an exception happened
260 __ empty_expression_stack();
261 // convention: expect aberrant index in register G3_scratch, then shuffle the
262 // index to G4_scratch for the VM call
263 __ mov(G3_scratch, G4_scratch);
264 __ set((intptr_t)name, G3_scratch);
265 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
266 __ should_not_reach_here();
267 return entry;
268 }
269
270
271 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
272 address entry = __ pc();
273 // expression stack must be empty before entering the VM if an exception happened
274 __ empty_expression_stack();
275 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
276 __ should_not_reach_here();
277 return entry;
278 }
279
280
281 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
282 address entry = __ pc();
283
284 if (state == atos) {
285 __ profile_return_type(O0, G3_scratch, G1_scratch);
286 }
287
288 // The callee returns with the stack possibly adjusted by adapter transition
289 // We remove that possible adjustment here.
290 // All interpreter local registers are untouched. Any result is passed back
291 // in the O0/O1 or float registers. Before continuing, the arguments must be
292 // popped from the java expression stack; i.e., Lesp must be adjusted.
293
294 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
295
296 const Register cache = G3_scratch;
297 const Register index = G1_scratch;
298 __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
299
300 const Register flags = cache;
301 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
302 const Register parameter_size = flags;
303 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words
304 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes
305 __ add(Lesp, parameter_size, Lesp); // pop arguments
306
307 __ check_and_handle_popframe(Gtemp);
308 __ check_and_handle_earlyret(Gtemp);
309
310 __ dispatch_next(state, step);
311
312 return entry;
313 }
314
315
316 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
317 address entry = __ pc();
318 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
319 #if INCLUDE_JVMCI
320 // Check if we need to take lock at entry of synchronized method. This can
321 // only occur on method entry so emit it only for vtos with step 0.
322 if (UseJVMCICompiler && state == vtos && step == 0) {
323 Label L;
324 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
325 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter
326 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
327 // Clear flag.
328 __ stbool(G0, pending_monitor_enter_addr);
329 // Take lock.
330 lock_method();
331 __ bind(L);
332 } else {
333 #ifdef ASSERT
334 if (UseJVMCICompiler) {
335 Label L;
336 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
337 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter
338 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
339 __ stop("unexpected pending monitor in deopt entry");
340 __ bind(L);
341 }
342 #endif
343 }
344 #endif
345 { Label L;
346 Address exception_addr(G2_thread, Thread::pending_exception_offset());
347 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
348 __ br_null_short(Gtemp, Assembler::pt, L);
349 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
350 __ should_not_reach_here();
351 __ bind(L);
352 }
353 __ dispatch_next(state, step);
354 return entry;
355 }
356
357 // A result handler converts/unboxes a native call result into
358 // a java interpreter/compiler result. The current frame is an
359 // interpreter frame. The activation frame unwind code must be
360 // consistent with that of TemplateTable::_return(...). In the
361 // case of native methods, the caller's SP was not modified.
362 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
363 address entry = __ pc();
364 Register Itos_i = Otos_i ->after_save();
365 Register Itos_l = Otos_l ->after_save();
366 Register Itos_l1 = Otos_l1->after_save();
367 Register Itos_l2 = Otos_l2->after_save();
368 switch (type) {
369 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
370 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
371 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
372 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
373 case T_LONG :
374 case T_INT : __ mov(O0, Itos_i); break;
375 case T_VOID : /* nothing to do */ break;
376 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
377 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
378 case T_OBJECT :
379 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
380 __ verify_oop(Itos_i);
381 break;
382 default : ShouldNotReachHere();
383 }
384 __ ret(); // return from interpreter activation
385 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
386 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
387 return entry;
388 }
389
390 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
391 address entry = __ pc();
392 __ push(state);
393 __ call_VM(noreg, runtime_entry);
394 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
395 return entry;
396 }
397
398
399 //
400 // Helpers for commoning out cases in the various type of method entries.
401 //
402
403 // increment invocation count & check for overflow
404 //
405 // Note: checking for negative value instead of overflow
406 // so we have a 'sticky' overflow test
407 //
408 // Lmethod: method
409 // ??: invocation counter
410 //
411 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
412 // Note: In tiered we increment either counters in MethodCounters* or in
413 // MDO depending if we're profiling or not.
414 const Register G3_method_counters = G3_scratch;
415 Label done;
416
417 if (TieredCompilation) {
418 const int increment = InvocationCounter::count_increment;
419 Label no_mdo;
420 if (ProfileInterpreter) {
421 // If no method data exists, go to profile_continue.
422 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
423 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
424 // Increment counter
425 Address mdo_invocation_counter(G4_scratch,
426 in_bytes(MethodData::invocation_counter_offset()) +
427 in_bytes(InvocationCounter::counter_offset()));
428 Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset()));
429 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
430 G3_scratch, Lscratch,
431 Assembler::zero, overflow);
432 __ ba_short(done);
433 }
434
435 // Increment counter in MethodCounters*
436 __ bind(no_mdo);
437 Address invocation_counter(G3_method_counters,
438 in_bytes(MethodCounters::invocation_counter_offset()) +
439 in_bytes(InvocationCounter::counter_offset()));
440 __ get_method_counters(Lmethod, G3_method_counters, done);
441 Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset()));
442 __ increment_mask_and_jump(invocation_counter, increment, mask,
443 G4_scratch, Lscratch,
444 Assembler::zero, overflow);
445 __ bind(done);
446 } else { // not TieredCompilation
447 // Update standard invocation counters
448 __ get_method_counters(Lmethod, G3_method_counters, done);
449 __ increment_invocation_counter(G3_method_counters, O0, G4_scratch);
450 if (ProfileInterpreter) {
451 Address interpreter_invocation_counter(G3_method_counters,
452 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
453 __ ld(interpreter_invocation_counter, G4_scratch);
454 __ inc(G4_scratch);
455 __ st(G4_scratch, interpreter_invocation_counter);
456 }
457
458 if (ProfileInterpreter && profile_method != NULL) {
459 // Test to see if we should create a method data oop
460 Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset()));
461 __ ld(profile_limit, G1_scratch);
462 __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
463
464 // if no method data exists, go to profile_method
465 __ test_method_data_pointer(*profile_method);
466 }
467
468 Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset()));
469 __ ld(invocation_limit, G3_scratch);
470 __ cmp(O0, G3_scratch);
471 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
472 __ delayed()->nop();
473 __ bind(done);
474 }
475 }
476
477 // Allocate monitor and lock method (asm interpreter)
478 // ebx - Method*
479 //
480 void TemplateInterpreterGenerator::lock_method() {
481 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
482
483 #ifdef ASSERT
484 { Label ok;
485 __ btst(JVM_ACC_SYNCHRONIZED, O0);
486 __ br( Assembler::notZero, false, Assembler::pt, ok);
487 __ delayed()->nop();
488 __ stop("method doesn't need synchronization");
489 __ bind(ok);
490 }
491 #endif // ASSERT
492
493 // get synchronization object to O0
494 { Label done;
495 __ btst(JVM_ACC_STATIC, O0);
496 __ br( Assembler::zero, true, Assembler::pt, done);
497 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
498
499 // lock the mirror, not the Klass*
500 __ load_mirror(O0, Lmethod);
501
502 #ifdef ASSERT
503 __ tst(O0);
504 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
505 #endif // ASSERT
506
507 __ bind(done);
508 }
509
510 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
511 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
512 // __ untested("lock_object from method entry");
513 __ lock_object(Lmonitors, O0);
514 }
515
516 // See if we've got enough room on the stack for locals plus overhead below
517 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
518 // without going through the signal handler, i.e., reserved and yellow zones
519 // will not be made usable. The shadow zone must suffice to handle the
520 // overflow.
521 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
522 Register Rscratch) {
523 const int page_size = os::vm_page_size();
524 Label after_frame_check;
525
526 assert_different_registers(Rframe_size, Rscratch);
527
528 __ set(page_size, Rscratch);
529 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
530
531 // Get the stack overflow limit, and in debug, verify it is non-zero.
532 __ ld_ptr(G2_thread, JavaThread::stack_overflow_limit_offset(), Rscratch);
533 #ifdef ASSERT
534 Label limit_ok;
535 __ br_notnull_short(Rscratch, Assembler::pn, limit_ok);
536 __ stop("stack overflow limit is zero in generate_stack_overflow_check");
537 __ bind(limit_ok);
538 #endif
539
540 // Add in the size of the frame (which is the same as subtracting it from the
541 // SP, which would take another register.
542 __ add(Rscratch, Rframe_size, Rscratch);
543
544 // The frame is greater than one page in size, so check against
545 // the bottom of the stack.
546 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
547
548 // The stack will overflow, throw an exception.
549
550 // Note that SP is restored to sender's sp (in the delay slot). This
551 // is necessary if the sender's frame is an extended compiled frame
552 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
553 // adaptations.
554
555 // Note also that the restored frame is not necessarily interpreted.
556 // Use the shared runtime version of the StackOverflowError.
557 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
558 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
559 __ jump_to(stub, Rscratch);
560 __ delayed()->mov(O5_savedSP, SP);
561
562 // If you get to here, then there is enough stack space.
563 __ bind(after_frame_check);
564 }
565
566
567 //
568 // Generate a fixed interpreter frame. This is identical setup for interpreted
569 // methods and for native methods hence the shared code.
570
571
572 //----------------------------------------------------------------------------------------------------
573 // Stack frame layout
574 //
575 // When control flow reaches any of the entry types for the interpreter
576 // the following holds ->
577 //
578 // C2 Calling Conventions:
579 //
580 // The entry code below assumes that the following registers are set
581 // when coming in:
582 // G5_method: holds the Method* of the method to call
583 // Lesp: points to the TOS of the callers expression stack
584 // after having pushed all the parameters
585 //
586 // The entry code does the following to setup an interpreter frame
587 // pop parameters from the callers stack by adjusting Lesp
588 // set O0 to Lesp
589 // compute X = (max_locals - num_parameters)
590 // bump SP up by X to accommodate the extra locals
591 // compute X = max_expression_stack
592 // + vm_local_words
593 // + 16 words of register save area
594 // save frame doing a save sp, -X, sp growing towards lower addresses
595 // set Lbcp, Lmethod, LcpoolCache
596 // set Llocals to i0
597 // set Lmonitors to FP - rounded_vm_local_words
598 // set Lesp to Lmonitors - 4
599 //
600 // The frame has now been setup to do the rest of the entry code
601
602 // Try this optimization: Most method entries could live in a
603 // "one size fits all" stack frame without all the dynamic size
604 // calculations. It might be profitable to do all this calculation
605 // statically and approximately for "small enough" methods.
606
607 //-----------------------------------------------------------------------------------------------
608
609 // C1 Calling conventions
610 //
611 // Upon method entry, the following registers are setup:
612 //
613 // g2 G2_thread: current thread
614 // g5 G5_method: method to activate
615 // g4 Gargs : pointer to last argument
616 //
617 //
618 // Stack:
619 //
620 // +---------------+ <--- sp
621 // | |
622 // : reg save area :
623 // | |
624 // +---------------+ <--- sp + 0x40
625 // | |
626 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
627 // | |
628 // +---------------+ <--- sp + 0x5c
629 // | |
630 // : free :
631 // | |
632 // +---------------+ <--- Gargs
633 // | |
634 // : arguments :
635 // | |
636 // +---------------+
637 // | |
638 //
639 //
640 //
641 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
642 //
643 // +---------------+ <--- sp
644 // | |
645 // : reg save area :
646 // | |
647 // +---------------+ <--- sp + 0x40
648 // | |
649 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
650 // | |
651 // +---------------+ <--- sp + 0x5c
652 // | |
653 // : :
654 // | | <--- Lesp
655 // +---------------+ <--- Lmonitors (fp - 0x18)
656 // | VM locals |
657 // +---------------+ <--- fp
658 // | |
659 // : reg save area :
660 // | |
661 // +---------------+ <--- fp + 0x40
662 // | |
663 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
664 // | |
665 // +---------------+ <--- fp + 0x5c
666 // | |
667 // : free :
668 // | |
669 // +---------------+
670 // | |
671 // : nonarg locals :
672 // | |
673 // +---------------+
674 // | |
675 // : arguments :
676 // | | <--- Llocals
677 // +---------------+ <--- Gargs
678 // | |
679
680 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
681 //
682 //
683 // The entry code sets up a new interpreter frame in 4 steps:
684 //
685 // 1) Increase caller's SP by for the extra local space needed:
686 // (check for overflow)
687 // Efficient implementation of xload/xstore bytecodes requires
688 // that arguments and non-argument locals are in a contiguously
689 // addressable memory block => non-argument locals must be
690 // allocated in the caller's frame.
691 //
692 // 2) Create a new stack frame and register window:
693 // The new stack frame must provide space for the standard
694 // register save area, the maximum java expression stack size,
695 // the monitor slots (0 slots initially), and some frame local
696 // scratch locations.
697 //
698 // 3) The following interpreter activation registers must be setup:
699 // Lesp : expression stack pointer
700 // Lbcp : bytecode pointer
701 // Lmethod : method
702 // Llocals : locals pointer
703 // Lmonitors : monitor pointer
704 // LcpoolCache: constant pool cache
705 //
706 // 4) Initialize the non-argument locals if necessary:
707 // Non-argument locals may need to be initialized to NULL
708 // for GC to work. If the oop-map information is accurate
709 // (in the absence of the JSR problem), no initialization
710 // is necessary.
711 //
712 // (gri - 2/25/2000)
713
714
715 int rounded_vm_local_words = align_up((int)frame::interpreter_frame_vm_local_words, WordsPerLong );
716
717 const int extra_space =
718 rounded_vm_local_words + // frame local scratch space
719 Method::extra_stack_entries() + // extra stack for jsr 292
720 frame::memory_parameter_word_sp_offset + // register save area
721 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
722
723 const Register Glocals_size = G3;
724 const Register RconstMethod = Glocals_size;
725 const Register Otmp1 = O3;
726 const Register Otmp2 = O4;
727 // Lscratch can't be used as a temporary because the call_stub uses
728 // it to assert that the stack frame was setup correctly.
729 const Address constMethod (G5_method, Method::const_offset());
730 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
731
732 __ ld_ptr( constMethod, RconstMethod );
733 __ lduh( size_of_parameters, Glocals_size);
734
735 // Gargs points to first local + BytesPerWord
736 // Set the saved SP after the register window save
737 //
738 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
739 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
740 __ add(Gargs, Otmp1, Gargs);
741
742 if (native_call) {
743 __ calc_mem_param_words( Glocals_size, Gframe_size );
744 __ add( Gframe_size, extra_space, Gframe_size);
745 __ round_to( Gframe_size, WordsPerLong );
746 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
747
748 // Native calls don't need the stack size check since they have no
749 // expression stack and the arguments are already on the stack and
750 // we only add a handful of words to the stack.
751 } else {
752
753 //
754 // Compute number of locals in method apart from incoming parameters
755 //
756 const Address size_of_locals(Otmp1, ConstMethod::size_of_locals_offset());
757 __ ld_ptr(constMethod, Otmp1);
758 __ lduh(size_of_locals, Otmp1);
759 __ sub(Otmp1, Glocals_size, Glocals_size);
760 __ round_to(Glocals_size, WordsPerLong);
761 __ sll(Glocals_size, Interpreter::logStackElementSize, Glocals_size);
762
763 // See if the frame is greater than one page in size. If so,
764 // then we need to verify there is enough stack space remaining.
765 // Frame_size = (max_stack + extra_space) * BytesPerWord;
766 __ ld_ptr(constMethod, Gframe_size);
767 __ lduh(Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size);
768 __ add(Gframe_size, extra_space, Gframe_size);
769 __ round_to(Gframe_size, WordsPerLong);
770 __ sll(Gframe_size, Interpreter::logStackElementSize, Gframe_size);
771
772 // Add in java locals size for stack overflow check only
773 __ add(Gframe_size, Glocals_size, Gframe_size);
774
775 const Register Otmp2 = O4;
776 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
777 generate_stack_overflow_check(Gframe_size, Otmp1);
778
779 __ sub(Gframe_size, Glocals_size, Gframe_size);
780
781 //
782 // bump SP to accommodate the extra locals
783 //
784 __ sub(SP, Glocals_size, SP);
785 }
786
787 //
788 // now set up a stack frame with the size computed above
789 //
790 __ neg( Gframe_size );
791 __ save( SP, Gframe_size, SP );
792
793 //
794 // now set up all the local cache registers
795 //
796 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
797 // that all present references to Lbyte_code initialize the register
798 // immediately before use
799 if (native_call) {
800 __ mov(G0, Lbcp);
801 } else {
802 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
803 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
804 }
805 __ mov( G5_method, Lmethod); // set Lmethod
806 // Get mirror and store it in the frame as GC root for this Method*
807 Register mirror = LcpoolCache;
808 __ load_mirror(mirror, Lmethod);
809 __ st_ptr(mirror, FP, (frame::interpreter_frame_mirror_offset * wordSize) + STACK_BIAS);
810 __ get_constant_pool_cache(LcpoolCache); // set LcpoolCache
811 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
812 __ add(Lmonitors, STACK_BIAS, Lmonitors); // Account for 64 bit stack bias
813 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
814
815 // setup interpreter activation registers
816 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
817
818 if (ProfileInterpreter) {
819 __ set_method_data_pointer();
820 }
821
822 }
823
824 // Method entry for java.lang.ref.Reference.get.
825 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
826 #if INCLUDE_ALL_GCS
827 // Code: _aload_0, _getfield, _areturn
828 // parameter size = 1
829 //
830 // The code that gets generated by this routine is split into 2 parts:
831 // 1. The "intrinsified" code for G1 (or any SATB based GC),
832 // 2. The slow path - which is an expansion of the regular method entry.
833 //
834 // Notes:-
835 // * In the G1 code we do not check whether we need to block for
836 // a safepoint. If G1 is enabled then we must execute the specialized
837 // code for Reference.get (except when the Reference object is null)
838 // so that we can log the value in the referent field with an SATB
839 // update buffer.
840 // If the code for the getfield template is modified so that the
841 // G1 pre-barrier code is executed when the current method is
842 // Reference.get() then going through the normal method entry
843 // will be fine.
844 // * The G1 code can, however, check the receiver object (the instance
845 // of java.lang.Reference) and jump to the slow path if null. If the
846 // Reference object is null then we obviously cannot fetch the referent
847 // and so we don't need to call the G1 pre-barrier. Thus we can use the
848 // regular method entry code to generate the NPE.
849 //
850 // This code is based on generate_accessor_enty.
851
852 address entry = __ pc();
853
854 const int referent_offset = java_lang_ref_Reference::referent_offset;
855 guarantee(referent_offset > 0, "referent offset not initialized");
856
857 if (UseG1GC) {
858 Label slow_path;
859
860 // In the G1 code we don't check if we need to reach a safepoint. We
861 // continue and the thread will safepoint at the next bytecode dispatch.
862
863 // Check if local 0 != NULL
864 // If the receiver is null then it is OK to jump to the slow path.
865 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
866 // check if local 0 == NULL and go the slow path
867 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
868
869
870 // Load the value of the referent field.
871 if (Assembler::is_simm13(referent_offset)) {
872 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
873 } else {
874 __ set(referent_offset, G3_scratch);
875 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
876 }
877
878 // Generate the G1 pre-barrier code to log the value of
879 // the referent field in an SATB buffer. Note with
880 // these parameters the pre-barrier does not generate
881 // the load of the previous value
882
883 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
884 Otos_i /* pre_val */,
885 G3_scratch /* tmp */,
886 true /* preserve_o_regs */);
887
888 // _areturn
889 __ retl(); // return from leaf routine
890 __ delayed()->mov(O5_savedSP, SP);
891
892 // Generate regular method entry
893 __ bind(slow_path);
894 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
895 return entry;
896 }
897 #endif // INCLUDE_ALL_GCS
898
899 // If G1 is not enabled then attempt to go through the accessor entry point
900 // Reference.get is an accessor
901 return NULL;
902 }
903
904 /**
905 * Method entry for static native methods:
906 * int java.util.zip.CRC32.update(int crc, int b)
907 */
908 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
909
910 if (UseCRC32Intrinsics) {
911 address entry = __ pc();
912
913 Label L_slow_path;
914 // If we need a safepoint check, generate full interpreter entry.
915 __ safepoint_poll(L_slow_path, false, G2_thread, O2);
916 __ delayed()->nop();
917
918 // Load parameters
919 const Register crc = O0; // initial crc
920 const Register val = O1; // byte to update with
921 const Register table = O2; // address of 256-entry lookup table
922
923 __ ldub(Gargs, 3, val);
924 __ lduw(Gargs, 8, crc);
925
926 __ set(ExternalAddress(StubRoutines::crc_table_addr()), table);
927
928 __ not1(crc); // ~crc
929 __ clruwu(crc);
930 __ update_byte_crc32(crc, val, table);
931 __ not1(crc); // ~crc
932
933 // result in O0
934 __ retl();
935 __ delayed()->nop();
936
937 // generate a vanilla native entry as the slow path
938 __ bind(L_slow_path);
939 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
940 return entry;
941 }
942 return NULL;
943 }
944
945 /**
946 * Method entry for static native methods:
947 * int java.util.zip.CRC32.updateBytes(int crc, byte[] b, int off, int len)
948 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
949 */
950 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
951
952 if (UseCRC32Intrinsics) {
953 address entry = __ pc();
954
955 Label L_slow_path;
956 // If we need a safepoint check, generate full interpreter entry.
957
958 __ safepoint_poll(L_slow_path, false, G2_thread, O2);
959 __ delayed()->nop();
960
961 // Load parameters from the stack
962 const Register crc = O0; // initial crc
963 const Register buf = O1; // source java byte array address
964 const Register len = O2; // len
965 const Register offset = O3; // offset
966
967 // Arguments are reversed on java expression stack
968 // Calculate address of start element
969 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) {
970 __ lduw(Gargs, 0, len);
971 __ lduw(Gargs, 8, offset);
972 __ ldx( Gargs, 16, buf);
973 __ lduw(Gargs, 32, crc);
974 __ add(buf, offset, buf);
975 } else {
976 __ lduw(Gargs, 0, len);
977 __ lduw(Gargs, 8, offset);
978 __ ldx( Gargs, 16, buf);
979 __ lduw(Gargs, 24, crc);
980 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
981 __ add(buf, offset, buf);
982 }
983
984 // Call the crc32 kernel
985 __ MacroAssembler::save_thread(L7_thread_cache);
986 __ kernel_crc32(crc, buf, len, O3);
987 __ MacroAssembler::restore_thread(L7_thread_cache);
988
989 // result in O0
990 __ retl();
991 __ delayed()->nop();
992
993 // generate a vanilla native entry as the slow path
994 __ bind(L_slow_path);
995 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native));
996 return entry;
997 }
998 return NULL;
999 }
1000
1001 /**
1002 * Method entry for intrinsic-candidate (non-native) methods:
1003 * int java.util.zip.CRC32C.updateBytes(int crc, byte[] b, int off, int end)
1004 * int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
1005 * Unlike CRC32, CRC32C does not have any methods marked as native
1006 * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
1007 */
1008 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1009
1010 if (UseCRC32CIntrinsics) {
1011 address entry = __ pc();
1012
1013 // Load parameters from the stack
1014 const Register crc = O0; // initial crc
1015 const Register buf = O1; // source java byte array address
1016 const Register offset = O2; // offset
1017 const Register end = O3; // index of last element to process
1018 const Register len = O2; // len argument to the kernel
1019 const Register table = O3; // crc32c lookup table address
1020
1021 // Arguments are reversed on java expression stack
1022 // Calculate address of start element
1023 if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) {
1024 __ lduw(Gargs, 0, end);
1025 __ lduw(Gargs, 8, offset);
1026 __ ldx( Gargs, 16, buf);
1027 __ lduw(Gargs, 32, crc);
1028 __ add(buf, offset, buf);
1029 __ sub(end, offset, len);
1030 } else {
1031 __ lduw(Gargs, 0, end);
1032 __ lduw(Gargs, 8, offset);
1033 __ ldx( Gargs, 16, buf);
1034 __ lduw(Gargs, 24, crc);
1035 __ add(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE), buf); // account for the header size
1036 __ add(buf, offset, buf);
1037 __ sub(end, offset, len);
1038 }
1039
1040 // Call the crc32c kernel
1041 __ MacroAssembler::save_thread(L7_thread_cache);
1042 __ kernel_crc32c(crc, buf, len, table);
1043 __ MacroAssembler::restore_thread(L7_thread_cache);
1044
1045 // result in O0
1046 __ retl();
1047 __ delayed()->nop();
1048
1049 return entry;
1050 }
1051 return NULL;
1052 }
1053
1054 /* Math routines only partially supported.
1055 *
1056 * Providing support for fma (float/double) only.
1057 */
1058 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind)
1059 {
1060 if (!InlineIntrinsics) return NULL; // Generate a vanilla entry
1061
1062 address entry = __ pc();
1063
1064 switch (kind) {
1065 case Interpreter::java_lang_math_fmaF:
1066 if (UseFMA) {
1067 // float .fma(float a, float b, float c)
1068 const FloatRegister ra = F1;
1069 const FloatRegister rb = F2;
1070 const FloatRegister rc = F3;
1071 const FloatRegister rd = F0; // Result.
1072
1073 __ ldf(FloatRegisterImpl::S, Gargs, 0, rc);
1074 __ ldf(FloatRegisterImpl::S, Gargs, 8, rb);
1075 __ ldf(FloatRegisterImpl::S, Gargs, 16, ra);
1076
1077 __ fmadd(FloatRegisterImpl::S, ra, rb, rc, rd);
1078 __ retl(); // Result in F0 (rd).
1079 __ delayed()->mov(O5_savedSP, SP);
1080
1081 return entry;
1082 }
1083 break;
1084 case Interpreter::java_lang_math_fmaD:
1085 if (UseFMA) {
1086 // double .fma(double a, double b, double c)
1087 const FloatRegister ra = F2; // D1
1088 const FloatRegister rb = F4; // D2
1089 const FloatRegister rc = F6; // D3
1090 const FloatRegister rd = F0; // D0 Result.
1091
1092 __ ldf(FloatRegisterImpl::D, Gargs, 0, rc);
1093 __ ldf(FloatRegisterImpl::D, Gargs, 16, rb);
1094 __ ldf(FloatRegisterImpl::D, Gargs, 32, ra);
1095
1096 __ fmadd(FloatRegisterImpl::D, ra, rb, rc, rd);
1097 __ retl(); // Result in D0 (rd).
1098 __ delayed()->mov(O5_savedSP, SP);
1099
1100 return entry;
1101 }
1102 break;
1103 default:
1104 break;
1105 }
1106 return NULL;
1107 }
1108
1109 // TODO: rather than touching all pages, check against stack_overflow_limit and bang yellow page to
1110 // generate exception
1111 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1112 // Quick & dirty stack overflow checking: bang the stack & handle trap.
1113 // Note that we do the banging after the frame is setup, since the exception
1114 // handling code expects to find a valid interpreter frame on the stack.
1115 // Doing the banging earlier fails if the caller frame is not an interpreter
1116 // frame.
1117 // (Also, the exception throwing code expects to unlock any synchronized
1118 // method receiver, so do the banging after locking the receiver.)
1119
1120 // Bang each page in the shadow zone. We can't assume it's been done for
1121 // an interpreter frame with greater than a page of locals, so each page
1122 // needs to be checked. Only true for non-native.
1123 if (UseStackBanging) {
1124 const int page_size = os::vm_page_size();
1125 const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1126 const int start_page = native_call ? n_shadow_pages : 1;
1127 for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1128 __ bang_stack_with_offset(pages*page_size);
1129 }
1130 }
1131 }
1132
1133 //
1134 // Interpreter stub for calling a native method. (asm interpreter)
1135 // This sets up a somewhat different looking stack for calling the native method
1136 // than the typical interpreter frame setup.
1137 //
1138
1139 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1140 address entry = __ pc();
1141
1142 // the following temporary registers are used during frame creation
1143 const Register Gtmp1 = G3_scratch ;
1144 const Register Gtmp2 = G1_scratch;
1145 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1146
1147 // make sure registers are different!
1148 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1149
1150 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
1151
1152 const Register Glocals_size = G3;
1153 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1154
1155 // make sure method is native & not abstract
1156 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1157 #ifdef ASSERT
1158 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1159 { Label L;
1160 __ btst(JVM_ACC_NATIVE, Gtmp1);
1161 __ br(Assembler::notZero, false, Assembler::pt, L);
1162 __ delayed()->nop();
1163 __ stop("tried to execute non-native method as native");
1164 __ bind(L);
1165 }
1166 { Label L;
1167 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1168 __ br(Assembler::zero, false, Assembler::pt, L);
1169 __ delayed()->nop();
1170 __ stop("tried to execute abstract method as non-abstract");
1171 __ bind(L);
1172 }
1173 #endif // ASSERT
1174
1175 // generate the code to allocate the interpreter stack frame
1176 generate_fixed_frame(true);
1177
1178 //
1179 // No locals to initialize for native method
1180 //
1181
1182 // this slot will be set later, we initialize it to null here just in
1183 // case we get a GC before the actual value is stored later
1184 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
1185
1186 const Address do_not_unlock_if_synchronized(G2_thread,
1187 JavaThread::do_not_unlock_if_synchronized_offset());
1188 // Since at this point in the method invocation the exception handler
1189 // would try to exit the monitor of synchronized methods which hasn't
1190 // been entered yet, we set the thread local variable
1191 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1192 // runtime, exception handling i.e. unlock_if_synchronized_method will
1193 // check this thread local flag.
1194 // This flag has two effects, one is to force an unwind in the topmost
1195 // interpreter frame and not perform an unlock while doing so.
1196
1197 __ movbool(true, G3_scratch);
1198 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1199
1200 // increment invocation counter and check for overflow
1201 //
1202 // Note: checking for negative value instead of overflow
1203 // so we have a 'sticky' overflow test (may be of
1204 // importance as soon as we have true MT/MP)
1205 Label invocation_counter_overflow;
1206 Label Lcontinue;
1207 if (inc_counter) {
1208 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1209
1210 }
1211 __ bind(Lcontinue);
1212
1213 bang_stack_shadow_pages(true);
1214
1215 // reset the _do_not_unlock_if_synchronized flag
1216 __ stbool(G0, do_not_unlock_if_synchronized);
1217
1218 // check for synchronized methods
1219 // Must happen AFTER invocation_counter check and stack overflow check,
1220 // so method is not locked if overflows.
1221
1222 if (synchronized) {
1223 lock_method();
1224 } else {
1225 #ifdef ASSERT
1226 { Label ok;
1227 __ ld(Laccess_flags, O0);
1228 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1229 __ br( Assembler::zero, false, Assembler::pt, ok);
1230 __ delayed()->nop();
1231 __ stop("method needs synchronization");
1232 __ bind(ok);
1233 }
1234 #endif // ASSERT
1235 }
1236
1237
1238 // start execution
1239 __ verify_thread();
1240
1241 // JVMTI support
1242 __ notify_method_entry();
1243
1244 // native call
1245
1246 // (note that O0 is never an oop--at most it is a handle)
1247 // It is important not to smash any handles created by this call,
1248 // until any oop handle in O0 is dereferenced.
1249
1250 // (note that the space for outgoing params is preallocated)
1251
1252 // get signature handler
1253 { Label L;
1254 Address signature_handler(Lmethod, Method::signature_handler_offset());
1255 __ ld_ptr(signature_handler, G3_scratch);
1256 __ br_notnull_short(G3_scratch, Assembler::pt, L);
1257 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
1258 __ ld_ptr(signature_handler, G3_scratch);
1259 __ bind(L);
1260 }
1261
1262 // Push a new frame so that the args will really be stored in
1263 // Copy a few locals across so the new frame has the variables
1264 // we need but these values will be dead at the jni call and
1265 // therefore not gc volatile like the values in the current
1266 // frame (Lmethod in particular)
1267
1268 // Flush the method pointer to the register save area
1269 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
1270 __ mov(Llocals, O1);
1271
1272 // calculate where the mirror handle body is allocated in the interpreter frame:
1273 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
1274
1275 // Calculate current frame size
1276 __ sub(SP, FP, O3); // Calculate negative of current frame size
1277 __ save(SP, O3, SP); // Allocate an identical sized frame
1278
1279 // Note I7 has leftover trash. Slow signature handler will fill it in
1280 // should we get there. Normal jni call will set reasonable last_Java_pc
1281 // below (and fix I7 so the stack trace doesn't have a meaningless frame
1282 // in it).
1283
1284 // Load interpreter frame's Lmethod into same register here
1285
1286 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1287
1288 __ mov(I1, Llocals);
1289 __ mov(I2, Lscratch2); // save the address of the mirror
1290
1291
1292 // ONLY Lmethod and Llocals are valid here!
1293
1294 // call signature handler, It will move the arg properly since Llocals in current frame
1295 // matches that in outer frame
1296
1297 __ callr(G3_scratch, 0);
1298 __ delayed()->nop();
1299
1300 // Result handler is in Lscratch
1301
1302 // Reload interpreter frame's Lmethod since slow signature handler may block
1303 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
1304
1305 { Label not_static;
1306
1307 __ ld(Laccess_flags, O0);
1308 __ btst(JVM_ACC_STATIC, O0);
1309 __ br( Assembler::zero, false, Assembler::pt, not_static);
1310 // get native function entry point(O0 is a good temp until the very end)
1311 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
1312 // for static methods insert the mirror argument
1313 __ load_mirror(O1, Lmethod);
1314 #ifdef ASSERT
1315 if (!PrintSignatureHandlers) // do not dirty the output with this
1316 { Label L;
1317 __ br_notnull_short(O1, Assembler::pt, L);
1318 __ stop("mirror is missing");
1319 __ bind(L);
1320 }
1321 #endif // ASSERT
1322 __ st_ptr(O1, Lscratch2, 0);
1323 __ mov(Lscratch2, O1);
1324 __ bind(not_static);
1325 }
1326
1327 // At this point, arguments have been copied off of stack into
1328 // their JNI positions, which are O1..O5 and SP[68..].
1329 // Oops are boxed in-place on the stack, with handles copied to arguments.
1330 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1331
1332 #ifdef ASSERT
1333 { Label L;
1334 __ br_notnull_short(O0, Assembler::pt, L);
1335 __ stop("native entry point is missing");
1336 __ bind(L);
1337 }
1338 #endif // ASSERT
1339
1340 //
1341 // setup the frame anchor
1342 //
1343 // The scavenge function only needs to know that the PC of this frame is
1344 // in the interpreter method entry code, it doesn't need to know the exact
1345 // PC and hence we can use O7 which points to the return address from the
1346 // previous call in the code stream (signature handler function)
1347 //
1348 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1349 // we have pushed the extra frame in order to protect the volatile register(s)
1350 // in that frame when we return from the jni call
1351 //
1352
1353 __ set_last_Java_frame(FP, O7);
1354 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1355 // not meaningless information that'll confuse me.
1356
1357 // flush the windows now. We don't care about the current (protection) frame
1358 // only the outer frames
1359
1360 __ flushw();
1361
1362 // mark windows as flushed
1363 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1364 __ set(JavaFrameAnchor::flushed, G3_scratch);
1365 __ st(G3_scratch, flags);
1366
1367 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1368
1369 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1370 #ifdef ASSERT
1371 { Label L;
1372 __ ld(thread_state, G3_scratch);
1373 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1374 __ stop("Wrong thread state in native stub");
1375 __ bind(L);
1376 }
1377 #endif // ASSERT
1378 __ set(_thread_in_native, G3_scratch);
1379 __ st(G3_scratch, thread_state);
1380
1381 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1382 __ save_thread(L7_thread_cache); // save Gthread
1383 __ callr(O0, 0);
1384 __ delayed()->
1385 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1386
1387 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1388
1389 __ restore_thread(L7_thread_cache); // restore G2_thread
1390 __ reinit_heapbase();
1391
1392 // must we block?
1393
1394 // Block, if necessary, before resuming in _thread_in_Java state.
1395 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1396 { Label no_block;
1397
1398 // Switch thread to "native transition" state before reading the synchronization state.
1399 // This additional state is necessary because reading and testing the synchronization
1400 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1401 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1402 // VM thread changes sync state to synchronizing and suspends threads for GC.
1403 // Thread A is resumed to finish this native method, but doesn't block here since it
1404 // didn't see any synchronization is progress, and escapes.
1405 __ set(_thread_in_native_trans, G3_scratch);
1406 __ st(G3_scratch, thread_state);
1407 if (os::is_MP()) {
1408 if (UseMembar) {
1409 // Force this write out before the read below
1410 __ membar(Assembler::StoreLoad);
1411 } else {
1412 // Write serialization page so VM thread can do a pseudo remote membar.
1413 // We use the current thread pointer to calculate a thread specific
1414 // offset to write to within the page. This minimizes bus traffic
1415 // due to cache line collision.
1416 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1417 }
1418 }
1419
1420 Label L;
1421 __ safepoint_poll(L, false, G2_thread, G3_scratch);
1422 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1423 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1424 __ bind(L);
1425
1426 // Block. Save any potential method result value before the operation and
1427 // use a leaf call to leave the last_Java_frame setup undisturbed.
1428 save_native_result();
1429 __ call_VM_leaf(L7_thread_cache,
1430 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1431 G2_thread);
1432
1433 // Restore any method result value
1434 restore_native_result();
1435 __ bind(no_block);
1436 }
1437
1438 // Clear the frame anchor now
1439
1440 __ reset_last_Java_frame();
1441
1442 // Move the result handler address
1443 __ mov(Lscratch, G3_scratch);
1444 // return possible result to the outer frame
1445 __ restore(O0, G0, O0);
1446
1447 // Move result handler to expected register
1448 __ mov(G3_scratch, Lscratch);
1449
1450 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1451 // switch to thread_in_Java.
1452
1453 __ set(_thread_in_Java, G3_scratch);
1454 __ st(G3_scratch, thread_state);
1455
1456 if (CheckJNICalls) {
1457 // clear_pending_jni_exception_check
1458 __ st_ptr(G0, G2_thread, JavaThread::pending_jni_exception_check_fn_offset());
1459 }
1460
1461 // reset handle block
1462 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1463 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1464
1465 // If we have an oop result store it where it will be safe for any further gc
1466 // until we return now that we've released the handle it might be protected by
1467
1468 { Label no_oop, store_result;
1469
1470 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1471 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1472 // Unbox oop result, e.g. JNIHandles::resolve value in O0.
1473 __ br_null(O0, false, Assembler::pn, store_result); // Use NULL as-is.
1474 __ delayed()->andcc(O0, JNIHandles::weak_tag_mask, G0); // Test for jweak
1475 __ brx(Assembler::zero, true, Assembler::pt, store_result);
1476 __ delayed()->ld_ptr(O0, 0, O0); // Maybe resolve (untagged) jobject.
1477 // Resolve jweak.
1478 __ ld_ptr(O0, -JNIHandles::weak_tag_value, O0);
1479 #if INCLUDE_ALL_GCS
1480 if (UseG1GC) {
1481 __ g1_write_barrier_pre(noreg /* obj */,
1482 noreg /* index */,
1483 0 /* offset */,
1484 O0 /* pre_val */,
1485 G3_scratch /* tmp */,
1486 true /* preserve_o_regs */);
1487 }
1488 #endif // INCLUDE_ALL_GCS
1489 __ bind(store_result);
1490 // Store it where gc will look for it and result handler expects it.
1491 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1492
1493 __ bind(no_oop);
1494
1495 }
1496
1497
1498 // handle exceptions (exception handling will handle unlocking!)
1499 { Label L;
1500 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1501 __ ld_ptr(exception_addr, Gtemp);
1502 __ br_null_short(Gtemp, Assembler::pt, L);
1503 // Note: This could be handled more efficiently since we know that the native
1504 // method doesn't have an exception handler. We could directly return
1505 // to the exception handler for the caller.
1506 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1507 __ should_not_reach_here();
1508 __ bind(L);
1509 }
1510
1511 // JVMTI support (preserves thread register)
1512 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1513
1514 if (synchronized) {
1515 // save and restore any potential method result value around the unlocking operation
1516 save_native_result();
1517
1518 __ add( __ top_most_monitor(), O1);
1519 __ unlock_object(O1);
1520
1521 restore_native_result();
1522 }
1523
1524 // dispose of return address and remove activation
1525 #ifdef ASSERT
1526 { Label ok;
1527 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1528 __ stop("bad I5_savedSP value");
1529 __ should_not_reach_here();
1530 __ bind(ok);
1531 }
1532 #endif
1533 __ jmp(Lscratch, 0);
1534 __ delayed()->nop();
1535
1536 if (inc_counter) {
1537 // handle invocation counter overflow
1538 __ bind(invocation_counter_overflow);
1539 generate_counter_overflow(Lcontinue);
1540 }
1541
1542 return entry;
1543 }
1544
1545
1546 // Generic method entry to (asm) interpreter
1547 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1548 address entry = __ pc();
1549
1550 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1551
1552 // the following temporary registers are used during frame creation
1553 const Register Gtmp1 = G3_scratch ;
1554 const Register Gtmp2 = G1_scratch;
1555
1556 // make sure registers are different!
1557 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1558
1559 const Address constMethod (G5_method, Method::const_offset());
1560 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1561 // and use in the asserts.
1562 const Address access_flags (Lmethod, Method::access_flags_offset());
1563
1564 const Register Glocals_size = G3;
1565 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1566
1567 // make sure method is not native & not abstract
1568 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1569 #ifdef ASSERT
1570 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1571 { Label L;
1572 __ btst(JVM_ACC_NATIVE, Gtmp1);
1573 __ br(Assembler::zero, false, Assembler::pt, L);
1574 __ delayed()->nop();
1575 __ stop("tried to execute native method as non-native");
1576 __ bind(L);
1577 }
1578 { Label L;
1579 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1580 __ br(Assembler::zero, false, Assembler::pt, L);
1581 __ delayed()->nop();
1582 __ stop("tried to execute abstract method as non-abstract");
1583 __ bind(L);
1584 }
1585 #endif // ASSERT
1586
1587 // generate the code to allocate the interpreter stack frame
1588
1589 generate_fixed_frame(false);
1590
1591 //
1592 // Code to initialize the extra (i.e. non-parm) locals
1593 //
1594 Register init_value = noreg; // will be G0 if we must clear locals
1595 // The way the code was setup before zerolocals was always true for vanilla java entries.
1596 // It could only be false for the specialized entries like accessor or empty which have
1597 // no extra locals so the testing was a waste of time and the extra locals were always
1598 // initialized. We removed this extra complication to already over complicated code.
1599
1600 init_value = G0;
1601 Label clear_loop;
1602
1603 const Register RconstMethod = O1;
1604 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1605 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1606
1607 // NOTE: If you change the frame layout, this code will need to
1608 // be updated!
1609 __ ld_ptr( constMethod, RconstMethod );
1610 __ lduh( size_of_locals, O2 );
1611 __ lduh( size_of_parameters, O1 );
1612 __ sll( O2, Interpreter::logStackElementSize, O2);
1613 __ sll( O1, Interpreter::logStackElementSize, O1 );
1614 __ sub( Llocals, O2, O2 );
1615 __ sub( Llocals, O1, O1 );
1616
1617 __ bind( clear_loop );
1618 __ inc( O2, wordSize );
1619
1620 __ cmp( O2, O1 );
1621 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1622 __ delayed()->st_ptr( init_value, O2, 0 );
1623
1624 const Address do_not_unlock_if_synchronized(G2_thread,
1625 JavaThread::do_not_unlock_if_synchronized_offset());
1626 // Since at this point in the method invocation the exception handler
1627 // would try to exit the monitor of synchronized methods which hasn't
1628 // been entered yet, we set the thread local variable
1629 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1630 // runtime, exception handling i.e. unlock_if_synchronized_method will
1631 // check this thread local flag.
1632 __ movbool(true, G3_scratch);
1633 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1634
1635 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1636 // increment invocation counter and check for overflow
1637 //
1638 // Note: checking for negative value instead of overflow
1639 // so we have a 'sticky' overflow test (may be of
1640 // importance as soon as we have true MT/MP)
1641 Label invocation_counter_overflow;
1642 Label profile_method;
1643 Label profile_method_continue;
1644 Label Lcontinue;
1645 if (inc_counter) {
1646 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1647 if (ProfileInterpreter) {
1648 __ bind(profile_method_continue);
1649 }
1650 }
1651 __ bind(Lcontinue);
1652
1653 bang_stack_shadow_pages(false);
1654
1655 // reset the _do_not_unlock_if_synchronized flag
1656 __ stbool(G0, do_not_unlock_if_synchronized);
1657
1658 // check for synchronized methods
1659 // Must happen AFTER invocation_counter check and stack overflow check,
1660 // so method is not locked if overflows.
1661
1662 if (synchronized) {
1663 lock_method();
1664 } else {
1665 #ifdef ASSERT
1666 { Label ok;
1667 __ ld(access_flags, O0);
1668 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1669 __ br( Assembler::zero, false, Assembler::pt, ok);
1670 __ delayed()->nop();
1671 __ stop("method needs synchronization");
1672 __ bind(ok);
1673 }
1674 #endif // ASSERT
1675 }
1676
1677 // start execution
1678
1679 __ verify_thread();
1680
1681 // jvmti support
1682 __ notify_method_entry();
1683
1684 // start executing instructions
1685 __ dispatch_next(vtos);
1686
1687
1688 if (inc_counter) {
1689 if (ProfileInterpreter) {
1690 // We have decided to profile this method in the interpreter
1691 __ bind(profile_method);
1692
1693 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1694 __ set_method_data_pointer_for_bcp();
1695 __ ba_short(profile_method_continue);
1696 }
1697
1698 // handle invocation counter overflow
1699 __ bind(invocation_counter_overflow);
1700 generate_counter_overflow(Lcontinue);
1701 }
1702
1703 return entry;
1704 }
1705
1706 //----------------------------------------------------------------------------------------------------
1707 // Exceptions
1708 void TemplateInterpreterGenerator::generate_throw_exception() {
1709
1710 // Entry point in previous activation (i.e., if the caller was interpreted)
1711 Interpreter::_rethrow_exception_entry = __ pc();
1712 // O0: exception
1713
1714 // entry point for exceptions thrown within interpreter code
1715 Interpreter::_throw_exception_entry = __ pc();
1716 __ verify_thread();
1717 // expression stack is undefined here
1718 // O0: exception, i.e. Oexception
1719 // Lbcp: exception bcp
1720 __ verify_oop(Oexception);
1721
1722
1723 // expression stack must be empty before entering the VM in case of an exception
1724 __ empty_expression_stack();
1725 // find exception handler address and preserve exception oop
1726 // call C routine to find handler and jump to it
1727 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1728 __ push_ptr(O1); // push exception for exception handler bytecodes
1729
1730 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1731 __ delayed()->nop();
1732
1733
1734 // if the exception is not handled in the current frame
1735 // the frame is removed and the exception is rethrown
1736 // (i.e. exception continuation is _rethrow_exception)
1737 //
1738 // Note: At this point the bci is still the bxi for the instruction which caused
1739 // the exception and the expression stack is empty. Thus, for any VM calls
1740 // at this point, GC will find a legal oop map (with empty expression stack).
1741
1742 // in current activation
1743 // tos: exception
1744 // Lbcp: exception bcp
1745
1746 //
1747 // JVMTI PopFrame support
1748 //
1749
1750 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1751 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1752 // Set the popframe_processing bit in popframe_condition indicating that we are
1753 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1754 // popframe handling cycles.
1755
1756 __ ld(popframe_condition_addr, G3_scratch);
1757 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1758 __ stw(G3_scratch, popframe_condition_addr);
1759
1760 // Empty the expression stack, as in normal exception handling
1761 __ empty_expression_stack();
1762 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1763
1764 {
1765 // Check to see whether we are returning to a deoptimized frame.
1766 // (The PopFrame call ensures that the caller of the popped frame is
1767 // either interpreted or compiled and deoptimizes it if compiled.)
1768 // In this case, we can't call dispatch_next() after the frame is
1769 // popped, but instead must save the incoming arguments and restore
1770 // them after deoptimization has occurred.
1771 //
1772 // Note that we don't compare the return PC against the
1773 // deoptimization blob's unpack entry because of the presence of
1774 // adapter frames in C2.
1775 Label caller_not_deoptimized;
1776 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1777 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1778
1779 const Register Gtmp1 = G3_scratch;
1780 const Register Gtmp2 = G1_scratch;
1781 const Register RconstMethod = Gtmp1;
1782 const Address constMethod(Lmethod, Method::const_offset());
1783 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1784
1785 // Compute size of arguments for saving when returning to deoptimized caller
1786 __ ld_ptr(constMethod, RconstMethod);
1787 __ lduh(size_of_parameters, Gtmp1);
1788 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1789 __ sub(Llocals, Gtmp1, Gtmp2);
1790 __ add(Gtmp2, wordSize, Gtmp2);
1791 // Save these arguments
1792 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1793 // Inform deoptimization that it is responsible for restoring these arguments
1794 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1795 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1796 __ st(Gtmp1, popframe_condition_addr);
1797
1798 // Return from the current method
1799 // The caller's SP was adjusted upon method entry to accomodate
1800 // the callee's non-argument locals. Undo that adjustment.
1801 __ ret();
1802 __ delayed()->restore(I5_savedSP, G0, SP);
1803
1804 __ bind(caller_not_deoptimized);
1805 }
1806
1807 // Clear the popframe condition flag
1808 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1809
1810 // Get out of the current method (how this is done depends on the particular compiler calling
1811 // convention that the interpreter currently follows)
1812 // The caller's SP was adjusted upon method entry to accomodate
1813 // the callee's non-argument locals. Undo that adjustment.
1814 __ restore(I5_savedSP, G0, SP);
1815 // The method data pointer was incremented already during
1816 // call profiling. We have to restore the mdp for the current bcp.
1817 if (ProfileInterpreter) {
1818 __ set_method_data_pointer_for_bcp();
1819 }
1820
1821 #if INCLUDE_JVMTI
1822 { Label L_done;
1823
1824 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1825 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1826
1827 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1828 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1829
1830 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1831
1832 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1833 __ delayed()->nop();
1834
1835 __ st_ptr(G1_scratch, Lesp, wordSize);
1836 __ bind(L_done);
1837 }
1838 #endif // INCLUDE_JVMTI
1839
1840 // Resume bytecode interpretation at the current bcp
1841 __ dispatch_next(vtos);
1842 // end of JVMTI PopFrame support
1843
1844 Interpreter::_remove_activation_entry = __ pc();
1845
1846 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1847 __ pop_ptr(Oexception); // get exception
1848
1849 // Intel has the following comment:
1850 //// remove the activation (without doing throws on illegalMonitorExceptions)
1851 // They remove the activation without checking for bad monitor state.
1852 // %%% We should make sure this is the right semantics before implementing.
1853
1854 __ set_vm_result(Oexception);
1855 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1856
1857 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1858
1859 __ get_vm_result(Oexception);
1860 __ verify_oop(Oexception);
1861
1862 const int return_reg_adjustment = frame::pc_return_offset;
1863 Address issuing_pc_addr(I7, return_reg_adjustment);
1864
1865 // We are done with this activation frame; find out where to go next.
1866 // The continuation point will be an exception handler, which expects
1867 // the following registers set up:
1868 //
1869 // Oexception: exception
1870 // Oissuing_pc: the local call that threw exception
1871 // Other On: garbage
1872 // In/Ln: the contents of the caller's register window
1873 //
1874 // We do the required restore at the last possible moment, because we
1875 // need to preserve some state across a runtime call.
1876 // (Remember that the caller activation is unknown--it might not be
1877 // interpreted, so things like Lscratch are useless in the caller.)
1878
1879 // Although the Intel version uses call_C, we can use the more
1880 // compact call_VM. (The only real difference on SPARC is a
1881 // harmlessly ignored [re]set_last_Java_frame, compared with
1882 // the Intel code which lacks this.)
1883 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1884 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1885 __ super_call_VM_leaf(L7_thread_cache,
1886 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1887 G2_thread, Oissuing_pc->after_save());
1888
1889 // The caller's SP was adjusted upon method entry to accomodate
1890 // the callee's non-argument locals. Undo that adjustment.
1891 __ JMP(O0, 0); // return exception handler in caller
1892 __ delayed()->restore(I5_savedSP, G0, SP);
1893
1894 // (same old exception object is already in Oexception; see above)
1895 // Note that an "issuing PC" is actually the next PC after the call
1896 }
1897
1898
1899 //
1900 // JVMTI ForceEarlyReturn support
1901 //
1902
1903 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1904 address entry = __ pc();
1905
1906 __ empty_expression_stack();
1907 __ load_earlyret_value(state);
1908
1909 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1910 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1911
1912 // Clear the earlyret state
1913 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1914
1915 __ remove_activation(state,
1916 /* throw_monitor_exception */ false,
1917 /* install_monitor_exception */ false);
1918
1919 // The caller's SP was adjusted upon method entry to accomodate
1920 // the callee's non-argument locals. Undo that adjustment.
1921 __ ret(); // return to caller
1922 __ delayed()->restore(I5_savedSP, G0, SP);
1923
1924 return entry;
1925 } // end of JVMTI ForceEarlyReturn support
1926
1927
1928 //------------------------------------------------------------------------------------------------------------------------
1929 // Helper for vtos entry point generation
1930
1931 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) {
1932 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1933 Label L;
1934 aep = __ pc(); __ push_ptr(); __ ba_short(L);
1935 fep = __ pc(); __ push_f(); __ ba_short(L);
1936 dep = __ pc(); __ push_d(); __ ba_short(L);
1937 lep = __ pc(); __ push_l(); __ ba_short(L);
1938 iep = __ pc(); __ push_i();
1939 bep = cep = sep = iep; // there aren't any
1940 vep = __ pc(); __ bind(L); // fall through
1941 generate_and_dispatch(t);
1942 }
1943
1944 // --------------------------------------------------------------------------------
1945
1946 // Non-product code
1947 #ifndef PRODUCT
1948 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1949 address entry = __ pc();
1950
1951 __ push(state);
1952 __ mov(O7, Lscratch); // protect return address within interpreter
1953
1954 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1955 __ mov( Otos_l2, G3_scratch );
1956 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1957 __ mov(Lscratch, O7); // restore return address
1958 __ pop(state);
1959 __ retl();
1960 __ delayed()->nop();
1961
1962 return entry;
1963 }
1964
1965
1966 // helpers for generate_and_dispatch
1967
1968 void TemplateInterpreterGenerator::count_bytecode() {
1969 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1970 }
1971
1972
1973 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1974 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
1975 }
1976
1977
1978 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1979 AddressLiteral index (&BytecodePairHistogram::_index);
1980 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
1981
1982 // get index, shift out old bytecode, bring in new bytecode, and store it
1983 // _index = (_index >> log2_number_of_codes) |
1984 // (bytecode << log2_number_of_codes);
1985
1986 __ load_contents(index, G4_scratch);
1987 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
1988 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
1989 __ or3( G3_scratch, G4_scratch, G4_scratch );
1990 __ store_contents(G4_scratch, index, G3_scratch);
1991
1992 // bump bucket contents
1993 // _counters[_index] ++;
1994
1995 __ set(counters, G3_scratch); // loads into G3_scratch
1996 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
1997 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
1998 __ ld (G3_scratch, 0, G4_scratch);
1999 __ inc (G4_scratch);
2000 __ st (G4_scratch, 0, G3_scratch);
2001 }
2002
2003
2004 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2005 // Call a little run-time stub to avoid blow-up for each bytecode.
2006 // The run-time runtime saves the right registers, depending on
2007 // the tosca in-state for the given template.
2008 address entry = Interpreter::trace_code(t->tos_in());
2009 guarantee(entry != NULL, "entry must have been generated");
2010 __ call(entry, relocInfo::none);
2011 __ delayed()->nop();
2012 }
2013
2014
2015 void TemplateInterpreterGenerator::stop_interpreter_at() {
2016 AddressLiteral counter(&BytecodeCounter::_counter_value);
2017 __ load_contents(counter, G3_scratch);
2018 AddressLiteral stop_at(&StopInterpreterAt);
2019 __ load_ptr_contents(stop_at, G4_scratch);
2020 __ cmp(G3_scratch, G4_scratch);
2021 __ breakpoint_trap(Assembler::equal, Assembler::icc);
2022 }
2023 #endif // not PRODUCT