1 /*
2 * Copyright (c) 1997, 2015, 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/interpreterGenerator.hpp"
30 #include "interpreter/interpreterRuntime.hpp"
31 #include "interpreter/interp_masm.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 CC_INTERP
51 #ifndef FAST_DISPATCH
52 #define FAST_DISPATCH 1
53 #endif
54 #undef FAST_DISPATCH
55
56
57 // Generation of Interpreter
58 //
59 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
60
61
62 #define __ _masm->
63
64
65 //----------------------------------------------------------------------------------------------------
66
67
68 void InterpreterGenerator::save_native_result(void) {
69 // result potentially in O0/O1: save it across calls
70 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
71
72 // result potentially in F0/F1: save it across calls
73 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
74
75 // save and restore any potential method result value around the unlocking operation
76 __ stf(FloatRegisterImpl::D, F0, d_tmp);
77 #ifdef _LP64
78 __ stx(O0, l_tmp);
79 #else
80 __ std(O0, l_tmp);
81 #endif
82 }
83
84 void InterpreterGenerator::restore_native_result(void) {
85 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
86 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
87
88 // Restore any method result value
89 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
90 #ifdef _LP64
91 __ ldx(l_tmp, O0);
92 #else
93 __ ldd(l_tmp, O0);
94 #endif
95 }
96
97 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
98 assert(!pass_oop || message == NULL, "either oop or message but not both");
99 address entry = __ pc();
100 // expression stack must be empty before entering the VM if an exception happened
101 __ empty_expression_stack();
102 // load exception object
103 __ set((intptr_t)name, G3_scratch);
104 if (pass_oop) {
105 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
106 } else {
107 __ set((intptr_t)message, G4_scratch);
108 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
109 }
110 // throw exception
111 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
112 AddressLiteral thrower(Interpreter::throw_exception_entry());
113 __ jump_to(thrower, G3_scratch);
114 __ delayed()->nop();
115 return entry;
116 }
117
118 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
119 address entry = __ pc();
120 // expression stack must be empty before entering the VM if an exception
121 // happened
122 __ empty_expression_stack();
123 // load exception object
124 __ call_VM(Oexception,
125 CAST_FROM_FN_PTR(address,
126 InterpreterRuntime::throw_ClassCastException),
127 Otos_i);
128 __ should_not_reach_here();
129 return entry;
130 }
131
132
133 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
134 address entry = __ pc();
135 // expression stack must be empty before entering the VM if an exception happened
136 __ empty_expression_stack();
137 // convention: expect aberrant index in register G3_scratch, then shuffle the
138 // index to G4_scratch for the VM call
139 __ mov(G3_scratch, G4_scratch);
140 __ set((intptr_t)name, G3_scratch);
141 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
142 __ should_not_reach_here();
143 return entry;
144 }
145
146
147 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
148 address entry = __ pc();
149 // expression stack must be empty before entering the VM if an exception happened
150 __ empty_expression_stack();
151 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
152 __ should_not_reach_here();
153 return entry;
154 }
155
156
157 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
158 address entry = __ pc();
159
160 if (state == atos) {
161 __ profile_return_type(O0, G3_scratch, G1_scratch);
162 }
163
164 #if !defined(_LP64) && defined(COMPILER2)
165 // All return values are where we want them, except for Longs. C2 returns
166 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
167 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
168 // build even if we are returning from interpreted we just do a little
169 // stupid shuffing.
170 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
171 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
172 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
173
174 if (state == ltos) {
175 __ srl (G1, 0, O1);
176 __ srlx(G1, 32, O0);
177 }
178 #endif // !_LP64 && COMPILER2
179
180 // The callee returns with the stack possibly adjusted by adapter transition
181 // We remove that possible adjustment here.
182 // All interpreter local registers are untouched. Any result is passed back
183 // in the O0/O1 or float registers. Before continuing, the arguments must be
184 // popped from the java expression stack; i.e., Lesp must be adjusted.
185
186 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
187
188 const Register cache = G3_scratch;
189 const Register index = G1_scratch;
190 __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
191
192 const Register flags = cache;
193 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
194 const Register parameter_size = flags;
195 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words
196 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes
197 __ add(Lesp, parameter_size, Lesp); // pop arguments
198 __ dispatch_next(state, step);
199
200 return entry;
201 }
202
203
204 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
205 address entry = __ pc();
206 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
207 #if INCLUDE_JVMCI
208 // Check if we need to take lock at entry of synchronized method.
209 if (UseJVMCICompiler) {
210 Label L;
211 Address pending_monitor_enter_addr(G2_thread, JavaThread::pending_monitorenter_offset());
212 __ ldbool(pending_monitor_enter_addr, Gtemp); // Load if pending monitor enter
213 __ cmp_and_br_short(Gtemp, G0, Assembler::equal, Assembler::pn, L);
214 // Clear flag.
215 __ stbool(G0, pending_monitor_enter_addr);
216 // Take lock.
217 lock_method();
218 __ bind(L);
219 }
220 #endif
221 { Label L;
222 Address exception_addr(G2_thread, Thread::pending_exception_offset());
223 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
224 __ br_null_short(Gtemp, Assembler::pt, L);
225 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
226 __ should_not_reach_here();
227 __ bind(L);
228 }
229 __ dispatch_next(state, step);
230 return entry;
231 }
232
233 // A result handler converts/unboxes a native call result into
234 // a java interpreter/compiler result. The current frame is an
235 // interpreter frame. The activation frame unwind code must be
236 // consistent with that of TemplateTable::_return(...). In the
237 // case of native methods, the caller's SP was not modified.
238 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
239 address entry = __ pc();
240 Register Itos_i = Otos_i ->after_save();
241 Register Itos_l = Otos_l ->after_save();
242 Register Itos_l1 = Otos_l1->after_save();
243 Register Itos_l2 = Otos_l2->after_save();
244 switch (type) {
245 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
246 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
247 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
248 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
249 case T_LONG :
250 #ifndef _LP64
251 __ mov(O1, Itos_l2); // move other half of long
252 #endif // ifdef or no ifdef, fall through to the T_INT case
253 case T_INT : __ mov(O0, Itos_i); break;
254 case T_VOID : /* nothing to do */ break;
255 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
256 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
257 case T_OBJECT :
258 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
259 __ verify_oop(Itos_i);
260 break;
261 default : ShouldNotReachHere();
262 }
263 __ ret(); // return from interpreter activation
264 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
265 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
266 return entry;
267 }
268
269 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
270 address entry = __ pc();
271 __ push(state);
272 __ call_VM(noreg, runtime_entry);
273 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
274 return entry;
275 }
276
277
278 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
279 address entry = __ pc();
280 __ dispatch_next(state);
281 return entry;
282 }
283
284 //
285 // Helpers for commoning out cases in the various type of method entries.
286 //
287
288 // increment invocation count & check for overflow
289 //
290 // Note: checking for negative value instead of overflow
291 // so we have a 'sticky' overflow test
292 //
293 // Lmethod: method
294 // ??: invocation counter
295 //
296 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
297 // Note: In tiered we increment either counters in MethodCounters* or in
298 // MDO depending if we're profiling or not.
299 const Register G3_method_counters = G3_scratch;
300 Label done;
301
302 if (TieredCompilation) {
303 const int increment = InvocationCounter::count_increment;
304 Label no_mdo;
305 if (ProfileInterpreter) {
306 // If no method data exists, go to profile_continue.
307 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
308 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
309 // Increment counter
310 Address mdo_invocation_counter(G4_scratch,
311 in_bytes(MethodData::invocation_counter_offset()) +
312 in_bytes(InvocationCounter::counter_offset()));
313 Address mask(G4_scratch, in_bytes(MethodData::invoke_mask_offset()));
314 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
315 G3_scratch, Lscratch,
316 Assembler::zero, overflow);
317 __ ba_short(done);
318 }
319
320 // Increment counter in MethodCounters*
321 __ bind(no_mdo);
322 Address invocation_counter(G3_method_counters,
323 in_bytes(MethodCounters::invocation_counter_offset()) +
324 in_bytes(InvocationCounter::counter_offset()));
325 __ get_method_counters(Lmethod, G3_method_counters, done);
326 Address mask(G3_method_counters, in_bytes(MethodCounters::invoke_mask_offset()));
327 __ increment_mask_and_jump(invocation_counter, increment, mask,
328 G4_scratch, Lscratch,
329 Assembler::zero, overflow);
330 __ bind(done);
331 } else { // not TieredCompilation
332 // Update standard invocation counters
333 __ get_method_counters(Lmethod, G3_method_counters, done);
334 __ increment_invocation_counter(G3_method_counters, O0, G4_scratch);
335 if (ProfileInterpreter) {
336 Address interpreter_invocation_counter(G3_method_counters,
337 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
338 __ ld(interpreter_invocation_counter, G4_scratch);
339 __ inc(G4_scratch);
340 __ st(G4_scratch, interpreter_invocation_counter);
341 }
342
343 if (ProfileInterpreter && profile_method != NULL) {
344 // Test to see if we should create a method data oop
345 Address profile_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_profile_limit_offset()));
346 __ ld(profile_limit, G1_scratch);
347 __ cmp_and_br_short(O0, G1_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
348
349 // if no method data exists, go to profile_method
350 __ test_method_data_pointer(*profile_method);
351 }
352
353 Address invocation_limit(G3_method_counters, in_bytes(MethodCounters::interpreter_invocation_limit_offset()));
354 __ ld(invocation_limit, G3_scratch);
355 __ cmp(O0, G3_scratch);
356 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
357 __ delayed()->nop();
358 __ bind(done);
359 }
360
361 }
362
363 // Allocate monitor and lock method (asm interpreter)
364 // ebx - Method*
365 //
366 void TemplateInterpreterGenerator::lock_method() {
367 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
368
369 #ifdef ASSERT
370 { Label ok;
371 __ btst(JVM_ACC_SYNCHRONIZED, O0);
372 __ br( Assembler::notZero, false, Assembler::pt, ok);
373 __ delayed()->nop();
374 __ stop("method doesn't need synchronization");
375 __ bind(ok);
376 }
377 #endif // ASSERT
378
379 // get synchronization object to O0
380 { Label done;
381 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
382 __ btst(JVM_ACC_STATIC, O0);
383 __ br( Assembler::zero, true, Assembler::pt, done);
384 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
385
386 __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0);
387 __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0);
388 __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0);
389
390 // lock the mirror, not the Klass*
391 __ ld_ptr( O0, mirror_offset, O0);
392
393 #ifdef ASSERT
394 __ tst(O0);
395 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
396 #endif // ASSERT
397
398 __ bind(done);
399 }
400
401 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
402 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
403 // __ untested("lock_object from method entry");
404 __ lock_object(Lmonitors, O0);
405 }
406
407
408 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
409 Register Rscratch,
410 Register Rscratch2) {
411 const int page_size = os::vm_page_size();
412 Label after_frame_check;
413
414 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
415
416 __ set(page_size, Rscratch);
417 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
418
419 // get the stack base, and in debug, verify it is non-zero
420 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
421 #ifdef ASSERT
422 Label base_not_zero;
423 __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
424 __ stop("stack base is zero in generate_stack_overflow_check");
425 __ bind(base_not_zero);
426 #endif
427
428 // get the stack size, and in debug, verify it is non-zero
429 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
430 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
431 #ifdef ASSERT
432 Label size_not_zero;
433 __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
434 __ stop("stack size is zero in generate_stack_overflow_check");
435 __ bind(size_not_zero);
436 #endif
437
438 // compute the beginning of the protected zone minus the requested frame size
439 __ sub( Rscratch, Rscratch2, Rscratch );
440 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
441 __ add( Rscratch, Rscratch2, Rscratch );
442
443 // Add in the size of the frame (which is the same as subtracting it from the
444 // SP, which would take another register
445 __ add( Rscratch, Rframe_size, Rscratch );
446
447 // the frame is greater than one page in size, so check against
448 // the bottom of the stack
449 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
450
451 // the stack will overflow, throw an exception
452
453 // Note that SP is restored to sender's sp (in the delay slot). This
454 // is necessary if the sender's frame is an extended compiled frame
455 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
456 // adaptations.
457
458 // Note also that the restored frame is not necessarily interpreted.
459 // Use the shared runtime version of the StackOverflowError.
460 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
461 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
462 __ jump_to(stub, Rscratch);
463 __ delayed()->mov(O5_savedSP, SP);
464
465 // if you get to here, then there is enough stack space
466 __ bind( after_frame_check );
467 }
468
469
470 //
471 // Generate a fixed interpreter frame. This is identical setup for interpreted
472 // methods and for native methods hence the shared code.
473
474
475 //----------------------------------------------------------------------------------------------------
476 // Stack frame layout
477 //
478 // When control flow reaches any of the entry types for the interpreter
479 // the following holds ->
480 //
481 // C2 Calling Conventions:
482 //
483 // The entry code below assumes that the following registers are set
484 // when coming in:
485 // G5_method: holds the Method* of the method to call
486 // Lesp: points to the TOS of the callers expression stack
487 // after having pushed all the parameters
488 //
489 // The entry code does the following to setup an interpreter frame
490 // pop parameters from the callers stack by adjusting Lesp
491 // set O0 to Lesp
492 // compute X = (max_locals - num_parameters)
493 // bump SP up by X to accomadate the extra locals
494 // compute X = max_expression_stack
495 // + vm_local_words
496 // + 16 words of register save area
497 // save frame doing a save sp, -X, sp growing towards lower addresses
498 // set Lbcp, Lmethod, LcpoolCache
499 // set Llocals to i0
500 // set Lmonitors to FP - rounded_vm_local_words
501 // set Lesp to Lmonitors - 4
502 //
503 // The frame has now been setup to do the rest of the entry code
504
505 // Try this optimization: Most method entries could live in a
506 // "one size fits all" stack frame without all the dynamic size
507 // calculations. It might be profitable to do all this calculation
508 // statically and approximately for "small enough" methods.
509
510 //-----------------------------------------------------------------------------------------------
511
512 // C1 Calling conventions
513 //
514 // Upon method entry, the following registers are setup:
515 //
516 // g2 G2_thread: current thread
517 // g5 G5_method: method to activate
518 // g4 Gargs : pointer to last argument
519 //
520 //
521 // Stack:
522 //
523 // +---------------+ <--- sp
524 // | |
525 // : reg save area :
526 // | |
527 // +---------------+ <--- sp + 0x40
528 // | |
529 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
530 // | |
531 // +---------------+ <--- sp + 0x5c
532 // | |
533 // : free :
534 // | |
535 // +---------------+ <--- Gargs
536 // | |
537 // : arguments :
538 // | |
539 // +---------------+
540 // | |
541 //
542 //
543 //
544 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
545 //
546 // +---------------+ <--- sp
547 // | |
548 // : reg save area :
549 // | |
550 // +---------------+ <--- sp + 0x40
551 // | |
552 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
553 // | |
554 // +---------------+ <--- sp + 0x5c
555 // | |
556 // : :
557 // | | <--- Lesp
558 // +---------------+ <--- Lmonitors (fp - 0x18)
559 // | VM locals |
560 // +---------------+ <--- fp
561 // | |
562 // : reg save area :
563 // | |
564 // +---------------+ <--- fp + 0x40
565 // | |
566 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
567 // | |
568 // +---------------+ <--- fp + 0x5c
569 // | |
570 // : free :
571 // | |
572 // +---------------+
573 // | |
574 // : nonarg locals :
575 // | |
576 // +---------------+
577 // | |
578 // : arguments :
579 // | | <--- Llocals
580 // +---------------+ <--- Gargs
581 // | |
582
583 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
584 //
585 //
586 // The entry code sets up a new interpreter frame in 4 steps:
587 //
588 // 1) Increase caller's SP by for the extra local space needed:
589 // (check for overflow)
590 // Efficient implementation of xload/xstore bytecodes requires
591 // that arguments and non-argument locals are in a contigously
592 // addressable memory block => non-argument locals must be
593 // allocated in the caller's frame.
594 //
595 // 2) Create a new stack frame and register window:
596 // The new stack frame must provide space for the standard
597 // register save area, the maximum java expression stack size,
598 // the monitor slots (0 slots initially), and some frame local
599 // scratch locations.
600 //
601 // 3) The following interpreter activation registers must be setup:
602 // Lesp : expression stack pointer
603 // Lbcp : bytecode pointer
604 // Lmethod : method
605 // Llocals : locals pointer
606 // Lmonitors : monitor pointer
607 // LcpoolCache: constant pool cache
608 //
609 // 4) Initialize the non-argument locals if necessary:
610 // Non-argument locals may need to be initialized to NULL
611 // for GC to work. If the oop-map information is accurate
612 // (in the absence of the JSR problem), no initialization
613 // is necessary.
614 //
615 // (gri - 2/25/2000)
616
617
618 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
619
620 const int extra_space =
621 rounded_vm_local_words + // frame local scratch space
622 Method::extra_stack_entries() + // extra stack for jsr 292
623 frame::memory_parameter_word_sp_offset + // register save area
624 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
625
626 const Register Glocals_size = G3;
627 const Register RconstMethod = Glocals_size;
628 const Register Otmp1 = O3;
629 const Register Otmp2 = O4;
630 // Lscratch can't be used as a temporary because the call_stub uses
631 // it to assert that the stack frame was setup correctly.
632 const Address constMethod (G5_method, Method::const_offset());
633 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
634
635 __ ld_ptr( constMethod, RconstMethod );
636 __ lduh( size_of_parameters, Glocals_size);
637
638 // Gargs points to first local + BytesPerWord
639 // Set the saved SP after the register window save
640 //
641 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
642 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
643 __ add(Gargs, Otmp1, Gargs);
644
645 if (native_call) {
646 __ calc_mem_param_words( Glocals_size, Gframe_size );
647 __ add( Gframe_size, extra_space, Gframe_size);
648 __ round_to( Gframe_size, WordsPerLong );
649 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
650 } else {
651
652 //
653 // Compute number of locals in method apart from incoming parameters
654 //
655 const Address size_of_locals (Otmp1, ConstMethod::size_of_locals_offset());
656 __ ld_ptr( constMethod, Otmp1 );
657 __ lduh( size_of_locals, Otmp1 );
658 __ sub( Otmp1, Glocals_size, Glocals_size );
659 __ round_to( Glocals_size, WordsPerLong );
660 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
661
662 // see if the frame is greater than one page in size. If so,
663 // then we need to verify there is enough stack space remaining
664 // Frame_size = (max_stack + extra_space) * BytesPerWord;
665 __ ld_ptr( constMethod, Gframe_size );
666 __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), Gframe_size );
667 __ add( Gframe_size, extra_space, Gframe_size );
668 __ round_to( Gframe_size, WordsPerLong );
669 __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
670
671 // Add in java locals size for stack overflow check only
672 __ add( Gframe_size, Glocals_size, Gframe_size );
673
674 const Register Otmp2 = O4;
675 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
676 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
677
678 __ sub( Gframe_size, Glocals_size, Gframe_size);
679
680 //
681 // bump SP to accomodate the extra locals
682 //
683 __ sub( SP, Glocals_size, SP );
684 }
685
686 //
687 // now set up a stack frame with the size computed above
688 //
689 __ neg( Gframe_size );
690 __ save( SP, Gframe_size, SP );
691
692 //
693 // now set up all the local cache registers
694 //
695 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
696 // that all present references to Lbyte_code initialize the register
697 // immediately before use
698 if (native_call) {
699 __ mov(G0, Lbcp);
700 } else {
701 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
702 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
703 }
704 __ mov( G5_method, Lmethod); // set Lmethod
705 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
706 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
707 #ifdef _LP64
708 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
709 #endif
710 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
711
712 // setup interpreter activation registers
713 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
714
715 if (ProfileInterpreter) {
716 #ifdef FAST_DISPATCH
717 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
718 // they both use I2.
719 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
720 #endif // FAST_DISPATCH
721 __ set_method_data_pointer();
722 }
723
724 }
725
726 // Method entry for java.lang.ref.Reference.get.
727 address InterpreterGenerator::generate_Reference_get_entry(void) {
728 #if INCLUDE_ALL_GCS
729 // Code: _aload_0, _getfield, _areturn
730 // parameter size = 1
731 //
732 // The code that gets generated by this routine is split into 2 parts:
733 // 1. The "intrinsified" code for G1 (or any SATB based GC),
734 // 2. The slow path - which is an expansion of the regular method entry.
735 //
736 // Notes:-
737 // * In the G1 code we do not check whether we need to block for
738 // a safepoint. If G1 is enabled then we must execute the specialized
739 // code for Reference.get (except when the Reference object is null)
740 // so that we can log the value in the referent field with an SATB
741 // update buffer.
742 // If the code for the getfield template is modified so that the
743 // G1 pre-barrier code is executed when the current method is
744 // Reference.get() then going through the normal method entry
745 // will be fine.
746 // * The G1 code can, however, check the receiver object (the instance
747 // of java.lang.Reference) and jump to the slow path if null. If the
748 // Reference object is null then we obviously cannot fetch the referent
749 // and so we don't need to call the G1 pre-barrier. Thus we can use the
750 // regular method entry code to generate the NPE.
751 //
752 // This code is based on generate_accessor_enty.
753
754 address entry = __ pc();
755
756 const int referent_offset = java_lang_ref_Reference::referent_offset;
757 guarantee(referent_offset > 0, "referent offset not initialized");
758
759 if (UseG1GC) {
760 Label slow_path;
761
762 // In the G1 code we don't check if we need to reach a safepoint. We
763 // continue and the thread will safepoint at the next bytecode dispatch.
764
765 // Check if local 0 != NULL
766 // If the receiver is null then it is OK to jump to the slow path.
767 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
768 // check if local 0 == NULL and go the slow path
769 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
770
771
772 // Load the value of the referent field.
773 if (Assembler::is_simm13(referent_offset)) {
774 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
775 } else {
776 __ set(referent_offset, G3_scratch);
777 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
778 }
779
780 // Generate the G1 pre-barrier code to log the value of
781 // the referent field in an SATB buffer. Note with
782 // these parameters the pre-barrier does not generate
783 // the load of the previous value
784
785 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
786 Otos_i /* pre_val */,
787 G3_scratch /* tmp */,
788 true /* preserve_o_regs */);
789
790 // _areturn
791 __ retl(); // return from leaf routine
792 __ delayed()->mov(O5_savedSP, SP);
793
794 // Generate regular method entry
795 __ bind(slow_path);
796 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals));
797 return entry;
798 }
799 #endif // INCLUDE_ALL_GCS
800
801 // If G1 is not enabled then attempt to go through the accessor entry point
802 // Reference.get is an accessor
803 return NULL;
804 }
805
806 //
807 // Interpreter stub for calling a native method. (asm interpreter)
808 // This sets up a somewhat different looking stack for calling the native method
809 // than the typical interpreter frame setup.
810 //
811
812 address InterpreterGenerator::generate_native_entry(bool synchronized) {
813 address entry = __ pc();
814
815 // the following temporary registers are used during frame creation
816 const Register Gtmp1 = G3_scratch ;
817 const Register Gtmp2 = G1_scratch;
818 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
819
820 // make sure registers are different!
821 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
822
823 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
824
825 const Register Glocals_size = G3;
826 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
827
828 // make sure method is native & not abstract
829 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
830 #ifdef ASSERT
831 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
832 {
833 Label L;
834 __ btst(JVM_ACC_NATIVE, Gtmp1);
835 __ br(Assembler::notZero, false, Assembler::pt, L);
836 __ delayed()->nop();
837 __ stop("tried to execute non-native method as native");
838 __ bind(L);
839 }
840 { Label L;
841 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
842 __ br(Assembler::zero, false, Assembler::pt, L);
843 __ delayed()->nop();
844 __ stop("tried to execute abstract method as non-abstract");
845 __ bind(L);
846 }
847 #endif // ASSERT
848
849 // generate the code to allocate the interpreter stack frame
850 generate_fixed_frame(true);
851
852 //
853 // No locals to initialize for native method
854 //
855
856 // this slot will be set later, we initialize it to null here just in
857 // case we get a GC before the actual value is stored later
858 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
859
860 const Address do_not_unlock_if_synchronized(G2_thread,
861 JavaThread::do_not_unlock_if_synchronized_offset());
862 // Since at this point in the method invocation the exception handler
863 // would try to exit the monitor of synchronized methods which hasn't
864 // been entered yet, we set the thread local variable
865 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
866 // runtime, exception handling i.e. unlock_if_synchronized_method will
867 // check this thread local flag.
868 // This flag has two effects, one is to force an unwind in the topmost
869 // interpreter frame and not perform an unlock while doing so.
870
871 __ movbool(true, G3_scratch);
872 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
873
874 // increment invocation counter and check for overflow
875 //
876 // Note: checking for negative value instead of overflow
877 // so we have a 'sticky' overflow test (may be of
878 // importance as soon as we have true MT/MP)
879 Label invocation_counter_overflow;
880 Label Lcontinue;
881 if (inc_counter) {
882 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
883
884 }
885 __ bind(Lcontinue);
886
887 bang_stack_shadow_pages(true);
888
889 // reset the _do_not_unlock_if_synchronized flag
890 __ stbool(G0, do_not_unlock_if_synchronized);
891
892 // check for synchronized methods
893 // Must happen AFTER invocation_counter check and stack overflow check,
894 // so method is not locked if overflows.
895
896 if (synchronized) {
897 lock_method();
898 } else {
899 #ifdef ASSERT
900 { Label ok;
901 __ ld(Laccess_flags, O0);
902 __ btst(JVM_ACC_SYNCHRONIZED, O0);
903 __ br( Assembler::zero, false, Assembler::pt, ok);
904 __ delayed()->nop();
905 __ stop("method needs synchronization");
906 __ bind(ok);
907 }
908 #endif // ASSERT
909 }
910
911
912 // start execution
913 __ verify_thread();
914
915 // JVMTI support
916 __ notify_method_entry();
917
918 // native call
919
920 // (note that O0 is never an oop--at most it is a handle)
921 // It is important not to smash any handles created by this call,
922 // until any oop handle in O0 is dereferenced.
923
924 // (note that the space for outgoing params is preallocated)
925
926 // get signature handler
927 { Label L;
928 Address signature_handler(Lmethod, Method::signature_handler_offset());
929 __ ld_ptr(signature_handler, G3_scratch);
930 __ br_notnull_short(G3_scratch, Assembler::pt, L);
931 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
932 __ ld_ptr(signature_handler, G3_scratch);
933 __ bind(L);
934 }
935
936 // Push a new frame so that the args will really be stored in
937 // Copy a few locals across so the new frame has the variables
938 // we need but these values will be dead at the jni call and
939 // therefore not gc volatile like the values in the current
940 // frame (Lmethod in particular)
941
942 // Flush the method pointer to the register save area
943 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
944 __ mov(Llocals, O1);
945
946 // calculate where the mirror handle body is allocated in the interpreter frame:
947 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
948
949 // Calculate current frame size
950 __ sub(SP, FP, O3); // Calculate negative of current frame size
951 __ save(SP, O3, SP); // Allocate an identical sized frame
952
953 // Note I7 has leftover trash. Slow signature handler will fill it in
954 // should we get there. Normal jni call will set reasonable last_Java_pc
955 // below (and fix I7 so the stack trace doesn't have a meaningless frame
956 // in it).
957
958 // Load interpreter frame's Lmethod into same register here
959
960 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
961
962 __ mov(I1, Llocals);
963 __ mov(I2, Lscratch2); // save the address of the mirror
964
965
966 // ONLY Lmethod and Llocals are valid here!
967
968 // call signature handler, It will move the arg properly since Llocals in current frame
969 // matches that in outer frame
970
971 __ callr(G3_scratch, 0);
972 __ delayed()->nop();
973
974 // Result handler is in Lscratch
975
976 // Reload interpreter frame's Lmethod since slow signature handler may block
977 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
978
979 { Label not_static;
980
981 __ ld(Laccess_flags, O0);
982 __ btst(JVM_ACC_STATIC, O0);
983 __ br( Assembler::zero, false, Assembler::pt, not_static);
984 // get native function entry point(O0 is a good temp until the very end)
985 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
986 // for static methods insert the mirror argument
987 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
988
989 __ ld_ptr(Lmethod, Method:: const_offset(), O1);
990 __ ld_ptr(O1, ConstMethod::constants_offset(), O1);
991 __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
992 __ ld_ptr(O1, mirror_offset, O1);
993 #ifdef ASSERT
994 if (!PrintSignatureHandlers) // do not dirty the output with this
995 { Label L;
996 __ br_notnull_short(O1, Assembler::pt, L);
997 __ stop("mirror is missing");
998 __ bind(L);
999 }
1000 #endif // ASSERT
1001 __ st_ptr(O1, Lscratch2, 0);
1002 __ mov(Lscratch2, O1);
1003 __ bind(not_static);
1004 }
1005
1006 // At this point, arguments have been copied off of stack into
1007 // their JNI positions, which are O1..O5 and SP[68..].
1008 // Oops are boxed in-place on the stack, with handles copied to arguments.
1009 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1010
1011 #ifdef ASSERT
1012 { Label L;
1013 __ br_notnull_short(O0, Assembler::pt, L);
1014 __ stop("native entry point is missing");
1015 __ bind(L);
1016 }
1017 #endif // ASSERT
1018
1019 //
1020 // setup the frame anchor
1021 //
1022 // The scavenge function only needs to know that the PC of this frame is
1023 // in the interpreter method entry code, it doesn't need to know the exact
1024 // PC and hence we can use O7 which points to the return address from the
1025 // previous call in the code stream (signature handler function)
1026 //
1027 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1028 // we have pushed the extra frame in order to protect the volatile register(s)
1029 // in that frame when we return from the jni call
1030 //
1031
1032 __ set_last_Java_frame(FP, O7);
1033 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1034 // not meaningless information that'll confuse me.
1035
1036 // flush the windows now. We don't care about the current (protection) frame
1037 // only the outer frames
1038
1039 __ flushw();
1040
1041 // mark windows as flushed
1042 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1043 __ set(JavaFrameAnchor::flushed, G3_scratch);
1044 __ st(G3_scratch, flags);
1045
1046 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1047
1048 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1049 #ifdef ASSERT
1050 { Label L;
1051 __ ld(thread_state, G3_scratch);
1052 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1053 __ stop("Wrong thread state in native stub");
1054 __ bind(L);
1055 }
1056 #endif // ASSERT
1057 __ set(_thread_in_native, G3_scratch);
1058 __ st(G3_scratch, thread_state);
1059
1060 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1061 __ save_thread(L7_thread_cache); // save Gthread
1062 __ callr(O0, 0);
1063 __ delayed()->
1064 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1065
1066 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1067
1068 __ restore_thread(L7_thread_cache); // restore G2_thread
1069 __ reinit_heapbase();
1070
1071 // must we block?
1072
1073 // Block, if necessary, before resuming in _thread_in_Java state.
1074 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1075 { Label no_block;
1076 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1077
1078 // Switch thread to "native transition" state before reading the synchronization state.
1079 // This additional state is necessary because reading and testing the synchronization
1080 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1081 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1082 // VM thread changes sync state to synchronizing and suspends threads for GC.
1083 // Thread A is resumed to finish this native method, but doesn't block here since it
1084 // didn't see any synchronization is progress, and escapes.
1085 __ set(_thread_in_native_trans, G3_scratch);
1086 __ st(G3_scratch, thread_state);
1087 if(os::is_MP()) {
1088 if (UseMembar) {
1089 // Force this write out before the read below
1090 __ membar(Assembler::StoreLoad);
1091 } else {
1092 // Write serialization page so VM thread can do a pseudo remote membar.
1093 // We use the current thread pointer to calculate a thread specific
1094 // offset to write to within the page. This minimizes bus traffic
1095 // due to cache line collision.
1096 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1097 }
1098 }
1099 __ load_contents(sync_state, G3_scratch);
1100 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1101
1102 Label L;
1103 __ br(Assembler::notEqual, false, Assembler::pn, L);
1104 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1105 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1106 __ bind(L);
1107
1108 // Block. Save any potential method result value before the operation and
1109 // use a leaf call to leave the last_Java_frame setup undisturbed.
1110 save_native_result();
1111 __ call_VM_leaf(L7_thread_cache,
1112 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1113 G2_thread);
1114
1115 // Restore any method result value
1116 restore_native_result();
1117 __ bind(no_block);
1118 }
1119
1120 // Clear the frame anchor now
1121
1122 __ reset_last_Java_frame();
1123
1124 // Move the result handler address
1125 __ mov(Lscratch, G3_scratch);
1126 // return possible result to the outer frame
1127 #ifndef __LP64
1128 __ mov(O0, I0);
1129 __ restore(O1, G0, O1);
1130 #else
1131 __ restore(O0, G0, O0);
1132 #endif /* __LP64 */
1133
1134 // Move result handler to expected register
1135 __ mov(G3_scratch, Lscratch);
1136
1137 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1138 // switch to thread_in_Java.
1139
1140 __ set(_thread_in_Java, G3_scratch);
1141 __ st(G3_scratch, thread_state);
1142
1143 // reset handle block
1144 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1145 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1146
1147 // If we have an oop result store it where it will be safe for any further gc
1148 // until we return now that we've released the handle it might be protected by
1149
1150 {
1151 Label no_oop, store_result;
1152
1153 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1154 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1155 __ addcc(G0, O0, O0);
1156 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1157 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1158 __ mov(G0, O0);
1159
1160 __ bind(store_result);
1161 // Store it where gc will look for it and result handler expects it.
1162 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1163
1164 __ bind(no_oop);
1165
1166 }
1167
1168
1169 // handle exceptions (exception handling will handle unlocking!)
1170 { Label L;
1171 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1172 __ ld_ptr(exception_addr, Gtemp);
1173 __ br_null_short(Gtemp, Assembler::pt, L);
1174 // Note: This could be handled more efficiently since we know that the native
1175 // method doesn't have an exception handler. We could directly return
1176 // to the exception handler for the caller.
1177 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1178 __ should_not_reach_here();
1179 __ bind(L);
1180 }
1181
1182 // JVMTI support (preserves thread register)
1183 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1184
1185 if (synchronized) {
1186 // save and restore any potential method result value around the unlocking operation
1187 save_native_result();
1188
1189 __ add( __ top_most_monitor(), O1);
1190 __ unlock_object(O1);
1191
1192 restore_native_result();
1193 }
1194
1195 #if defined(COMPILER2) && !defined(_LP64)
1196
1197 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1198 // or compiled so just be safe.
1199
1200 __ sllx(O0, 32, G1); // Shift bits into high G1
1201 __ srl (O1, 0, O1); // Zero extend O1
1202 __ or3 (O1, G1, G1); // OR 64 bits into G1
1203
1204 #endif /* COMPILER2 && !_LP64 */
1205
1206 // dispose of return address and remove activation
1207 #ifdef ASSERT
1208 {
1209 Label ok;
1210 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1211 __ stop("bad I5_savedSP value");
1212 __ should_not_reach_here();
1213 __ bind(ok);
1214 }
1215 #endif
1216 if (TraceJumps) {
1217 // Move target to register that is recordable
1218 __ mov(Lscratch, G3_scratch);
1219 __ JMP(G3_scratch, 0);
1220 } else {
1221 __ jmp(Lscratch, 0);
1222 }
1223 __ delayed()->nop();
1224
1225
1226 if (inc_counter) {
1227 // handle invocation counter overflow
1228 __ bind(invocation_counter_overflow);
1229 generate_counter_overflow(Lcontinue);
1230 }
1231
1232
1233
1234 return entry;
1235 }
1236
1237
1238 // Generic method entry to (asm) interpreter
1239 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1240 address entry = __ pc();
1241
1242 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1243
1244 // the following temporary registers are used during frame creation
1245 const Register Gtmp1 = G3_scratch ;
1246 const Register Gtmp2 = G1_scratch;
1247
1248 // make sure registers are different!
1249 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1250
1251 const Address constMethod (G5_method, Method::const_offset());
1252 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1253 // and use in the asserts.
1254 const Address access_flags (Lmethod, Method::access_flags_offset());
1255
1256 const Register Glocals_size = G3;
1257 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1258
1259 // make sure method is not native & not abstract
1260 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1261 #ifdef ASSERT
1262 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1263 {
1264 Label L;
1265 __ btst(JVM_ACC_NATIVE, Gtmp1);
1266 __ br(Assembler::zero, false, Assembler::pt, L);
1267 __ delayed()->nop();
1268 __ stop("tried to execute native method as non-native");
1269 __ bind(L);
1270 }
1271 { Label L;
1272 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1273 __ br(Assembler::zero, false, Assembler::pt, L);
1274 __ delayed()->nop();
1275 __ stop("tried to execute abstract method as non-abstract");
1276 __ bind(L);
1277 }
1278 #endif // ASSERT
1279
1280 // generate the code to allocate the interpreter stack frame
1281
1282 generate_fixed_frame(false);
1283
1284 #ifdef FAST_DISPATCH
1285 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1286 // set bytecode dispatch table base
1287 #endif
1288
1289 //
1290 // Code to initialize the extra (i.e. non-parm) locals
1291 //
1292 Register init_value = noreg; // will be G0 if we must clear locals
1293 // The way the code was setup before zerolocals was always true for vanilla java entries.
1294 // It could only be false for the specialized entries like accessor or empty which have
1295 // no extra locals so the testing was a waste of time and the extra locals were always
1296 // initialized. We removed this extra complication to already over complicated code.
1297
1298 init_value = G0;
1299 Label clear_loop;
1300
1301 const Register RconstMethod = O1;
1302 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1303 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1304
1305 // NOTE: If you change the frame layout, this code will need to
1306 // be updated!
1307 __ ld_ptr( constMethod, RconstMethod );
1308 __ lduh( size_of_locals, O2 );
1309 __ lduh( size_of_parameters, O1 );
1310 __ sll( O2, Interpreter::logStackElementSize, O2);
1311 __ sll( O1, Interpreter::logStackElementSize, O1 );
1312 __ sub( Llocals, O2, O2 );
1313 __ sub( Llocals, O1, O1 );
1314
1315 __ bind( clear_loop );
1316 __ inc( O2, wordSize );
1317
1318 __ cmp( O2, O1 );
1319 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1320 __ delayed()->st_ptr( init_value, O2, 0 );
1321
1322 const Address do_not_unlock_if_synchronized(G2_thread,
1323 JavaThread::do_not_unlock_if_synchronized_offset());
1324 // Since at this point in the method invocation the exception handler
1325 // would try to exit the monitor of synchronized methods which hasn't
1326 // been entered yet, we set the thread local variable
1327 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1328 // runtime, exception handling i.e. unlock_if_synchronized_method will
1329 // check this thread local flag.
1330 __ movbool(true, G3_scratch);
1331 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1332
1333 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1334 // increment invocation counter and check for overflow
1335 //
1336 // Note: checking for negative value instead of overflow
1337 // so we have a 'sticky' overflow test (may be of
1338 // importance as soon as we have true MT/MP)
1339 Label invocation_counter_overflow;
1340 Label profile_method;
1341 Label profile_method_continue;
1342 Label Lcontinue;
1343 if (inc_counter) {
1344 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1345 if (ProfileInterpreter) {
1346 __ bind(profile_method_continue);
1347 }
1348 }
1349 __ bind(Lcontinue);
1350
1351 bang_stack_shadow_pages(false);
1352
1353 // reset the _do_not_unlock_if_synchronized flag
1354 __ stbool(G0, do_not_unlock_if_synchronized);
1355
1356 // check for synchronized methods
1357 // Must happen AFTER invocation_counter check and stack overflow check,
1358 // so method is not locked if overflows.
1359
1360 if (synchronized) {
1361 lock_method();
1362 } else {
1363 #ifdef ASSERT
1364 { Label ok;
1365 __ ld(access_flags, O0);
1366 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1367 __ br( Assembler::zero, false, Assembler::pt, ok);
1368 __ delayed()->nop();
1369 __ stop("method needs synchronization");
1370 __ bind(ok);
1371 }
1372 #endif // ASSERT
1373 }
1374
1375 // start execution
1376
1377 __ verify_thread();
1378
1379 // jvmti support
1380 __ notify_method_entry();
1381
1382 // start executing instructions
1383 __ dispatch_next(vtos);
1384
1385
1386 if (inc_counter) {
1387 if (ProfileInterpreter) {
1388 // We have decided to profile this method in the interpreter
1389 __ bind(profile_method);
1390
1391 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1392 __ set_method_data_pointer_for_bcp();
1393 __ ba_short(profile_method_continue);
1394 }
1395
1396 // handle invocation counter overflow
1397 __ bind(invocation_counter_overflow);
1398 generate_counter_overflow(Lcontinue);
1399 }
1400
1401
1402 return entry;
1403 }
1404
1405 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1406
1407 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1408 // expression stack, the callee will have callee_extra_locals (so we can account for
1409 // frame extension) and monitor_size for monitors. Basically we need to calculate
1410 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1411 //
1412 //
1413 // The big complicating thing here is that we must ensure that the stack stays properly
1414 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1415 // needs to be aligned for). We are given that the sp (fp) is already aligned by
1416 // the caller so we must ensure that it is properly aligned for our callee.
1417 //
1418 const int rounded_vm_local_words =
1419 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1420 // callee_locals and max_stack are counts, not the size in frame.
1421 const int locals_size =
1422 round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1423 const int max_stack_words = max_stack * Interpreter::stackElementWords;
1424 return (round_to((max_stack_words
1425 + rounded_vm_local_words
1426 + frame::memory_parameter_word_sp_offset), WordsPerLong)
1427 // already rounded
1428 + locals_size + monitor_size);
1429 }
1430
1431 // How much stack a method top interpreter activation needs in words.
1432 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1433
1434 // See call_stub code
1435 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
1436 WordsPerLong); // 7 + register save area
1437
1438 // Save space for one monitor to get into the interpreted method in case
1439 // the method is synchronized
1440 int monitor_size = method->is_synchronized() ?
1441 1*frame::interpreter_frame_monitor_size() : 0;
1442 return size_activation_helper(method->max_locals(), method->max_stack(),
1443 monitor_size) + call_stub_size;
1444 }
1445
1446 int AbstractInterpreter::size_activation(int max_stack,
1447 int temps,
1448 int extra_args,
1449 int monitors,
1450 int callee_params,
1451 int callee_locals,
1452 bool is_top_frame) {
1453 // Note: This calculation must exactly parallel the frame setup
1454 // in InterpreterGenerator::generate_fixed_frame.
1455
1456 int monitor_size = monitors * frame::interpreter_frame_monitor_size();
1457
1458 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1459
1460 //
1461 // Note: if you look closely this appears to be doing something much different
1462 // than generate_fixed_frame. What is happening is this. On sparc we have to do
1463 // this dance with interpreter_sp_adjustment because the window save area would
1464 // appear just below the bottom (tos) of the caller's java expression stack. Because
1465 // the interpreter want to have the locals completely contiguous generate_fixed_frame
1466 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1467 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1468 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1469 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1470 // because the oldest frame would have adjust its callers frame and yet that frame
1471 // already exists and isn't part of this array of frames we are unpacking. So at first
1472 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1473 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1474 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1475 // add up. It does seem like it simpler to account for the adjustment here (and remove the
1476 // callee... parameters here). However this would mean that this routine would have to take
1477 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1478 // and run the calling loop in the reverse order. This would also would appear to mean making
1479 // this code aware of what the interactions are when that initial caller fram was an osr or
1480 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
1481 // there is no sense in messing working code.
1482 //
1483
1484 int rounded_cls = round_to((callee_locals - callee_params), WordsPerLong);
1485 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1486
1487 int raw_frame_size = size_activation_helper(rounded_cls, max_stack, monitor_size);
1488
1489 return raw_frame_size;
1490 }
1491
1492 void AbstractInterpreter::layout_activation(Method* method,
1493 int tempcount,
1494 int popframe_extra_args,
1495 int moncount,
1496 int caller_actual_parameters,
1497 int callee_param_count,
1498 int callee_local_count,
1499 frame* caller,
1500 frame* interpreter_frame,
1501 bool is_top_frame,
1502 bool is_bottom_frame) {
1503 // Set up the following variables:
1504 // - Lmethod
1505 // - Llocals
1506 // - Lmonitors (to the indicated number of monitors)
1507 // - Lesp (to the indicated number of temps)
1508 // The frame caller on entry is a description of the caller of the
1509 // frame we are about to layout. We are guaranteed that we will be
1510 // able to fill in a new interpreter frame as its callee (i.e. the
1511 // stack space is allocated and the amount was determined by an
1512 // earlier call to the size_activation() method). On return caller
1513 // while describe the interpreter frame we just layed out.
1514
1515 // The skeleton frame must already look like an interpreter frame
1516 // even if not fully filled out.
1517 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1518
1519 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1520 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
1521 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1522
1523 intptr_t* fp = interpreter_frame->fp();
1524
1525 JavaThread* thread = JavaThread::current();
1526 RegisterMap map(thread, false);
1527 // More verification that skeleton frame is properly walkable
1528 assert(fp == caller->sp(), "fp must match");
1529
1530 intptr_t* montop = fp - rounded_vm_local_words;
1531
1532 // preallocate monitors (cf. __ add_monitor_to_stack)
1533 intptr_t* monitors = montop - monitor_size;
1534
1535 // preallocate stack space
1536 intptr_t* esp = monitors - 1 -
1537 (tempcount * Interpreter::stackElementWords) -
1538 popframe_extra_args;
1539
1540 int local_words = method->max_locals() * Interpreter::stackElementWords;
1541 NEEDS_CLEANUP;
1542 intptr_t* locals;
1543 if (caller->is_interpreted_frame()) {
1544 // Can force the locals area to end up properly overlapping the top of the expression stack.
1545 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1546 // Note that this computation means we replace size_of_parameters() values from the caller
1547 // interpreter frame's expression stack with our argument locals
1548 int parm_words = caller_actual_parameters * Interpreter::stackElementWords;
1549 locals = Lesp_ptr + parm_words;
1550 int delta = local_words - parm_words;
1551 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1552 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1553 if (!is_bottom_frame) {
1554 // Llast_SP is set below for the current frame to SP (with the
1555 // extra space for the callee's locals). Here we adjust
1556 // Llast_SP for the caller's frame, removing the extra space
1557 // for the current method's locals.
1558 *caller->register_addr(Llast_SP) = *interpreter_frame->register_addr(I5_savedSP);
1559 } else {
1560 assert(*caller->register_addr(Llast_SP) >= *interpreter_frame->register_addr(I5_savedSP), "strange Llast_SP");
1561 }
1562 } else {
1563 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1564 // Don't have Lesp available; lay out locals block in the caller
1565 // adjacent to the register window save area.
1566 //
1567 // Compiled frames do not allocate a varargs area which is why this if
1568 // statement is needed.
1569 //
1570 if (caller->is_compiled_frame()) {
1571 locals = fp + frame::register_save_words + local_words - 1;
1572 } else {
1573 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1574 }
1575 if (!caller->is_entry_frame()) {
1576 // Caller wants his own SP back
1577 int caller_frame_size = caller->cb()->frame_size();
1578 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1579 }
1580 }
1581 if (TraceDeoptimization) {
1582 if (caller->is_entry_frame()) {
1583 // make sure I5_savedSP and the entry frames notion of saved SP
1584 // agree. This assertion duplicate a check in entry frame code
1585 // but catches the failure earlier.
1586 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1587 "would change callers SP");
1588 }
1589 if (caller->is_entry_frame()) {
1590 tty->print("entry ");
1591 }
1592 if (caller->is_compiled_frame()) {
1593 tty->print("compiled ");
1594 if (caller->is_deoptimized_frame()) {
1595 tty->print("(deopt) ");
1596 }
1597 }
1598 if (caller->is_interpreted_frame()) {
1599 tty->print("interpreted ");
1600 }
1601 tty->print_cr("caller fp=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(caller->fp()), p2i(caller->sp()));
1602 tty->print_cr("save area = " INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(caller->sp()), p2i(caller->sp() + 16));
1603 tty->print_cr("save area = " INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(caller->fp()), p2i(caller->fp() + 16));
1604 tty->print_cr("interpreter fp=" INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(interpreter_frame->fp()), p2i(interpreter_frame->sp()));
1605 tty->print_cr("save area = " INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(interpreter_frame->sp()), p2i(interpreter_frame->sp() + 16));
1606 tty->print_cr("save area = " INTPTR_FORMAT ", " INTPTR_FORMAT, p2i(interpreter_frame->fp()), p2i(interpreter_frame->fp() + 16));
1607 tty->print_cr("Llocals = " INTPTR_FORMAT, p2i(locals));
1608 tty->print_cr("Lesp = " INTPTR_FORMAT, p2i(esp));
1609 tty->print_cr("Lmonitors = " INTPTR_FORMAT, p2i(monitors));
1610 }
1611
1612 if (method->max_locals() > 0) {
1613 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1614 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1615 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1616 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1617 }
1618 #ifdef _LP64
1619 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1620 #endif
1621
1622 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
1623 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
1624 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
1625 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
1626 // Llast_SP will be same as SP as there is no adapter space
1627 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1628 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1629 #ifdef FAST_DISPATCH
1630 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1631 #endif
1632
1633
1634 #ifdef ASSERT
1635 BasicObjectLock* mp = (BasicObjectLock*)monitors;
1636
1637 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1638 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1639 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
1640 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1641 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1642
1643 // check bounds
1644 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1645 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1646 assert(lo < monitors && montop <= hi, "monitors in bounds");
1647 assert(lo <= esp && esp < monitors, "esp in bounds");
1648 #endif // ASSERT
1649 }
1650
1651 //----------------------------------------------------------------------------------------------------
1652 // Exceptions
1653 void TemplateInterpreterGenerator::generate_throw_exception() {
1654
1655 // Entry point in previous activation (i.e., if the caller was interpreted)
1656 Interpreter::_rethrow_exception_entry = __ pc();
1657 // O0: exception
1658
1659 // entry point for exceptions thrown within interpreter code
1660 Interpreter::_throw_exception_entry = __ pc();
1661 __ verify_thread();
1662 // expression stack is undefined here
1663 // O0: exception, i.e. Oexception
1664 // Lbcp: exception bcp
1665 __ verify_oop(Oexception);
1666
1667
1668 // expression stack must be empty before entering the VM in case of an exception
1669 __ empty_expression_stack();
1670 // find exception handler address and preserve exception oop
1671 // call C routine to find handler and jump to it
1672 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1673 __ push_ptr(O1); // push exception for exception handler bytecodes
1674
1675 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1676 __ delayed()->nop();
1677
1678
1679 // if the exception is not handled in the current frame
1680 // the frame is removed and the exception is rethrown
1681 // (i.e. exception continuation is _rethrow_exception)
1682 //
1683 // Note: At this point the bci is still the bxi for the instruction which caused
1684 // the exception and the expression stack is empty. Thus, for any VM calls
1685 // at this point, GC will find a legal oop map (with empty expression stack).
1686
1687 // in current activation
1688 // tos: exception
1689 // Lbcp: exception bcp
1690
1691 //
1692 // JVMTI PopFrame support
1693 //
1694
1695 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1696 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1697 // Set the popframe_processing bit in popframe_condition indicating that we are
1698 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1699 // popframe handling cycles.
1700
1701 __ ld(popframe_condition_addr, G3_scratch);
1702 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1703 __ stw(G3_scratch, popframe_condition_addr);
1704
1705 // Empty the expression stack, as in normal exception handling
1706 __ empty_expression_stack();
1707 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1708
1709 {
1710 // Check to see whether we are returning to a deoptimized frame.
1711 // (The PopFrame call ensures that the caller of the popped frame is
1712 // either interpreted or compiled and deoptimizes it if compiled.)
1713 // In this case, we can't call dispatch_next() after the frame is
1714 // popped, but instead must save the incoming arguments and restore
1715 // them after deoptimization has occurred.
1716 //
1717 // Note that we don't compare the return PC against the
1718 // deoptimization blob's unpack entry because of the presence of
1719 // adapter frames in C2.
1720 Label caller_not_deoptimized;
1721 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1722 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1723
1724 const Register Gtmp1 = G3_scratch;
1725 const Register Gtmp2 = G1_scratch;
1726 const Register RconstMethod = Gtmp1;
1727 const Address constMethod(Lmethod, Method::const_offset());
1728 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1729
1730 // Compute size of arguments for saving when returning to deoptimized caller
1731 __ ld_ptr(constMethod, RconstMethod);
1732 __ lduh(size_of_parameters, Gtmp1);
1733 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1734 __ sub(Llocals, Gtmp1, Gtmp2);
1735 __ add(Gtmp2, wordSize, Gtmp2);
1736 // Save these arguments
1737 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1738 // Inform deoptimization that it is responsible for restoring these arguments
1739 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1740 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1741 __ st(Gtmp1, popframe_condition_addr);
1742
1743 // Return from the current method
1744 // The caller's SP was adjusted upon method entry to accomodate
1745 // the callee's non-argument locals. Undo that adjustment.
1746 __ ret();
1747 __ delayed()->restore(I5_savedSP, G0, SP);
1748
1749 __ bind(caller_not_deoptimized);
1750 }
1751
1752 // Clear the popframe condition flag
1753 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1754
1755 // Get out of the current method (how this is done depends on the particular compiler calling
1756 // convention that the interpreter currently follows)
1757 // The caller's SP was adjusted upon method entry to accomodate
1758 // the callee's non-argument locals. Undo that adjustment.
1759 __ restore(I5_savedSP, G0, SP);
1760 // The method data pointer was incremented already during
1761 // call profiling. We have to restore the mdp for the current bcp.
1762 if (ProfileInterpreter) {
1763 __ set_method_data_pointer_for_bcp();
1764 }
1765
1766 #if INCLUDE_JVMTI
1767 {
1768 Label L_done;
1769
1770 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1771 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1772
1773 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1774 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1775
1776 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1777
1778 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1779 __ delayed()->nop();
1780
1781 __ st_ptr(G1_scratch, Lesp, wordSize);
1782 __ bind(L_done);
1783 }
1784 #endif // INCLUDE_JVMTI
1785
1786 // Resume bytecode interpretation at the current bcp
1787 __ dispatch_next(vtos);
1788 // end of JVMTI PopFrame support
1789
1790 Interpreter::_remove_activation_entry = __ pc();
1791
1792 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1793 __ pop_ptr(Oexception); // get exception
1794
1795 // Intel has the following comment:
1796 //// remove the activation (without doing throws on illegalMonitorExceptions)
1797 // They remove the activation without checking for bad monitor state.
1798 // %%% We should make sure this is the right semantics before implementing.
1799
1800 __ set_vm_result(Oexception);
1801 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1802
1803 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1804
1805 __ get_vm_result(Oexception);
1806 __ verify_oop(Oexception);
1807
1808 const int return_reg_adjustment = frame::pc_return_offset;
1809 Address issuing_pc_addr(I7, return_reg_adjustment);
1810
1811 // We are done with this activation frame; find out where to go next.
1812 // The continuation point will be an exception handler, which expects
1813 // the following registers set up:
1814 //
1815 // Oexception: exception
1816 // Oissuing_pc: the local call that threw exception
1817 // Other On: garbage
1818 // In/Ln: the contents of the caller's register window
1819 //
1820 // We do the required restore at the last possible moment, because we
1821 // need to preserve some state across a runtime call.
1822 // (Remember that the caller activation is unknown--it might not be
1823 // interpreted, so things like Lscratch are useless in the caller.)
1824
1825 // Although the Intel version uses call_C, we can use the more
1826 // compact call_VM. (The only real difference on SPARC is a
1827 // harmlessly ignored [re]set_last_Java_frame, compared with
1828 // the Intel code which lacks this.)
1829 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1830 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1831 __ super_call_VM_leaf(L7_thread_cache,
1832 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1833 G2_thread, Oissuing_pc->after_save());
1834
1835 // The caller's SP was adjusted upon method entry to accomodate
1836 // the callee's non-argument locals. Undo that adjustment.
1837 __ JMP(O0, 0); // return exception handler in caller
1838 __ delayed()->restore(I5_savedSP, G0, SP);
1839
1840 // (same old exception object is already in Oexception; see above)
1841 // Note that an "issuing PC" is actually the next PC after the call
1842 }
1843
1844
1845 //
1846 // JVMTI ForceEarlyReturn support
1847 //
1848
1849 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1850 address entry = __ pc();
1851
1852 __ empty_expression_stack();
1853 __ load_earlyret_value(state);
1854
1855 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1856 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1857
1858 // Clear the earlyret state
1859 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1860
1861 __ remove_activation(state,
1862 /* throw_monitor_exception */ false,
1863 /* install_monitor_exception */ false);
1864
1865 // The caller's SP was adjusted upon method entry to accomodate
1866 // the callee's non-argument locals. Undo that adjustment.
1867 __ ret(); // return to caller
1868 __ delayed()->restore(I5_savedSP, G0, SP);
1869
1870 return entry;
1871 } // end of JVMTI ForceEarlyReturn support
1872
1873
1874 //------------------------------------------------------------------------------------------------------------------------
1875 // Helper for vtos entry point generation
1876
1877 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) {
1878 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1879 Label L;
1880 aep = __ pc(); __ push_ptr(); __ ba_short(L);
1881 fep = __ pc(); __ push_f(); __ ba_short(L);
1882 dep = __ pc(); __ push_d(); __ ba_short(L);
1883 lep = __ pc(); __ push_l(); __ ba_short(L);
1884 iep = __ pc(); __ push_i();
1885 bep = cep = sep = iep; // there aren't any
1886 vep = __ pc(); __ bind(L); // fall through
1887 generate_and_dispatch(t);
1888 }
1889
1890 // --------------------------------------------------------------------------------
1891
1892
1893 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1894 : TemplateInterpreterGenerator(code) {
1895 generate_all(); // down here so it can be "virtual"
1896 }
1897
1898 // --------------------------------------------------------------------------------
1899
1900 // Non-product code
1901 #ifndef PRODUCT
1902 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1903 address entry = __ pc();
1904
1905 __ push(state);
1906 __ mov(O7, Lscratch); // protect return address within interpreter
1907
1908 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1909 __ mov( Otos_l2, G3_scratch );
1910 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1911 __ mov(Lscratch, O7); // restore return address
1912 __ pop(state);
1913 __ retl();
1914 __ delayed()->nop();
1915
1916 return entry;
1917 }
1918
1919
1920 // helpers for generate_and_dispatch
1921
1922 void TemplateInterpreterGenerator::count_bytecode() {
1923 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1924 }
1925
1926
1927 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1928 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
1929 }
1930
1931
1932 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1933 AddressLiteral index (&BytecodePairHistogram::_index);
1934 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
1935
1936 // get index, shift out old bytecode, bring in new bytecode, and store it
1937 // _index = (_index >> log2_number_of_codes) |
1938 // (bytecode << log2_number_of_codes);
1939
1940 __ load_contents(index, G4_scratch);
1941 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
1942 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
1943 __ or3( G3_scratch, G4_scratch, G4_scratch );
1944 __ store_contents(G4_scratch, index, G3_scratch);
1945
1946 // bump bucket contents
1947 // _counters[_index] ++;
1948
1949 __ set(counters, G3_scratch); // loads into G3_scratch
1950 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
1951 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
1952 __ ld (G3_scratch, 0, G4_scratch);
1953 __ inc (G4_scratch);
1954 __ st (G4_scratch, 0, G3_scratch);
1955 }
1956
1957
1958 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
1959 // Call a little run-time stub to avoid blow-up for each bytecode.
1960 // The run-time runtime saves the right registers, depending on
1961 // the tosca in-state for the given template.
1962 address entry = Interpreter::trace_code(t->tos_in());
1963 guarantee(entry != NULL, "entry must have been generated");
1964 __ call(entry, relocInfo::none);
1965 __ delayed()->nop();
1966 }
1967
1968
1969 void TemplateInterpreterGenerator::stop_interpreter_at() {
1970 AddressLiteral counter(&BytecodeCounter::_counter_value);
1971 __ load_contents(counter, G3_scratch);
1972 AddressLiteral stop_at(&StopInterpreterAt);
1973 __ load_ptr_contents(stop_at, G4_scratch);
1974 __ cmp(G3_scratch, G4_scratch);
1975 __ breakpoint_trap(Assembler::equal, Assembler::icc);
1976 }
1977 #endif // not PRODUCT
1978 #endif // !CC_INTERP