Tue, 09 Oct 2012 07:41:27 +0200
Merge
1 /*
2 * Copyright (c) 1997, 2012, 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 */
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "interpreter/bytecodeHistogram.hpp"
28 #include "interpreter/interpreter.hpp"
29 #include "interpreter/interpreterGenerator.hpp"
30 #include "interpreter/interpreterRuntime.hpp"
31 #include "interpreter/templateTable.hpp"
32 #include "oops/arrayOop.hpp"
33 #include "oops/methodData.hpp"
34 #include "oops/method.hpp"
35 #include "oops/oop.inline.hpp"
36 #include "prims/jvmtiExport.hpp"
37 #include "prims/jvmtiThreadState.hpp"
38 #include "runtime/arguments.hpp"
39 #include "runtime/deoptimization.hpp"
40 #include "runtime/frame.inline.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "runtime/stubRoutines.hpp"
43 #include "runtime/synchronizer.hpp"
44 #include "runtime/timer.hpp"
45 #include "runtime/vframeArray.hpp"
46 #include "utilities/debug.hpp"
48 #ifndef CC_INTERP
49 #ifndef FAST_DISPATCH
50 #define FAST_DISPATCH 1
51 #endif
52 #undef FAST_DISPATCH
55 // Generation of Interpreter
56 //
57 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
60 #define __ _masm->
63 //----------------------------------------------------------------------------------------------------
66 void InterpreterGenerator::save_native_result(void) {
67 // result potentially in O0/O1: save it across calls
68 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
70 // result potentially in F0/F1: save it across calls
71 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
73 // save and restore any potential method result value around the unlocking operation
74 __ stf(FloatRegisterImpl::D, F0, d_tmp);
75 #ifdef _LP64
76 __ stx(O0, l_tmp);
77 #else
78 __ std(O0, l_tmp);
79 #endif
80 }
82 void InterpreterGenerator::restore_native_result(void) {
83 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
84 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
86 // Restore any method result value
87 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
88 #ifdef _LP64
89 __ ldx(l_tmp, O0);
90 #else
91 __ ldd(l_tmp, O0);
92 #endif
93 }
95 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
96 assert(!pass_oop || message == NULL, "either oop or message but not both");
97 address entry = __ pc();
98 // expression stack must be empty before entering the VM if an exception happened
99 __ empty_expression_stack();
100 // load exception object
101 __ set((intptr_t)name, G3_scratch);
102 if (pass_oop) {
103 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
104 } else {
105 __ set((intptr_t)message, G4_scratch);
106 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
107 }
108 // throw exception
109 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
110 AddressLiteral thrower(Interpreter::throw_exception_entry());
111 __ jump_to(thrower, G3_scratch);
112 __ delayed()->nop();
113 return entry;
114 }
116 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
117 address entry = __ pc();
118 // expression stack must be empty before entering the VM if an exception
119 // happened
120 __ empty_expression_stack();
121 // load exception object
122 __ call_VM(Oexception,
123 CAST_FROM_FN_PTR(address,
124 InterpreterRuntime::throw_ClassCastException),
125 Otos_i);
126 __ should_not_reach_here();
127 return entry;
128 }
131 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
132 address entry = __ pc();
133 // expression stack must be empty before entering the VM if an exception happened
134 __ empty_expression_stack();
135 // convention: expect aberrant index in register G3_scratch, then shuffle the
136 // index to G4_scratch for the VM call
137 __ mov(G3_scratch, G4_scratch);
138 __ set((intptr_t)name, G3_scratch);
139 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
140 __ should_not_reach_here();
141 return entry;
142 }
145 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
146 address entry = __ pc();
147 // expression stack must be empty before entering the VM if an exception happened
148 __ empty_expression_stack();
149 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
150 __ should_not_reach_here();
151 return entry;
152 }
155 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) {
156 TosState incoming_state = state;
158 Label cont;
159 address compiled_entry = __ pc();
161 address entry = __ pc();
162 #if !defined(_LP64) && defined(COMPILER2)
163 // All return values are where we want them, except for Longs. C2 returns
164 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
165 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
166 // build even if we are returning from interpreted we just do a little
167 // stupid shuffing.
168 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
169 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
170 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
172 if (incoming_state == ltos) {
173 __ srl (G1, 0, O1);
174 __ srlx(G1, 32, O0);
175 }
176 #endif // !_LP64 && COMPILER2
178 __ bind(cont);
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.
186 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
188 Label L_got_cache, L_giant_index;
189 const Register cache = G3_scratch;
190 const Register size = G1_scratch;
191 if (EnableInvokeDynamic) {
192 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode.
193 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokedynamic, Assembler::equal, Assembler::pn, L_giant_index);
194 }
195 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1);
196 __ bind(L_got_cache);
197 __ ld_ptr(cache, ConstantPoolCache::base_offset() +
198 ConstantPoolCacheEntry::flags_offset(), size);
199 __ and3(size, 0xFF, size); // argument size in words
200 __ sll(size, Interpreter::logStackElementSize, size); // each argument size in bytes
201 __ add(Lesp, size, Lesp); // pop arguments
202 __ dispatch_next(state, step);
204 // out of the main line of code...
205 if (EnableInvokeDynamic) {
206 __ bind(L_giant_index);
207 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1, sizeof(u4));
208 __ ba_short(L_got_cache);
209 }
211 return entry;
212 }
215 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
216 address entry = __ pc();
217 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
218 { Label L;
219 Address exception_addr(G2_thread, Thread::pending_exception_offset());
220 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
221 __ br_null_short(Gtemp, Assembler::pt, L);
222 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
223 __ should_not_reach_here();
224 __ bind(L);
225 }
226 __ dispatch_next(state, step);
227 return entry;
228 }
230 // A result handler converts/unboxes a native call result into
231 // a java interpreter/compiler result. The current frame is an
232 // interpreter frame. The activation frame unwind code must be
233 // consistent with that of TemplateTable::_return(...). In the
234 // case of native methods, the caller's SP was not modified.
235 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
236 address entry = __ pc();
237 Register Itos_i = Otos_i ->after_save();
238 Register Itos_l = Otos_l ->after_save();
239 Register Itos_l1 = Otos_l1->after_save();
240 Register Itos_l2 = Otos_l2->after_save();
241 switch (type) {
242 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
243 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
244 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
245 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
246 case T_LONG :
247 #ifndef _LP64
248 __ mov(O1, Itos_l2); // move other half of long
249 #endif // ifdef or no ifdef, fall through to the T_INT case
250 case T_INT : __ mov(O0, Itos_i); break;
251 case T_VOID : /* nothing to do */ break;
252 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
253 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
254 case T_OBJECT :
255 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
256 __ verify_oop(Itos_i);
257 break;
258 default : ShouldNotReachHere();
259 }
260 __ ret(); // return from interpreter activation
261 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
262 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
263 return entry;
264 }
266 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
267 address entry = __ pc();
268 __ push(state);
269 __ call_VM(noreg, runtime_entry);
270 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
271 return entry;
272 }
275 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
276 address entry = __ pc();
277 __ dispatch_next(state);
278 return entry;
279 }
281 //
282 // Helpers for commoning out cases in the various type of method entries.
283 //
285 // increment invocation count & check for overflow
286 //
287 // Note: checking for negative value instead of overflow
288 // so we have a 'sticky' overflow test
289 //
290 // Lmethod: method
291 // ??: invocation counter
292 //
293 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
294 // Note: In tiered we increment either counters in Method* or in MDO depending if we're profiling or not.
295 if (TieredCompilation) {
296 const int increment = InvocationCounter::count_increment;
297 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
298 Label no_mdo, done;
299 if (ProfileInterpreter) {
300 // If no method data exists, go to profile_continue.
301 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
302 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
303 // Increment counter
304 Address mdo_invocation_counter(G4_scratch,
305 in_bytes(MethodData::invocation_counter_offset()) +
306 in_bytes(InvocationCounter::counter_offset()));
307 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
308 G3_scratch, Lscratch,
309 Assembler::zero, overflow);
310 __ ba_short(done);
311 }
313 // Increment counter in Method*
314 __ bind(no_mdo);
315 Address invocation_counter(Lmethod,
316 in_bytes(Method::invocation_counter_offset()) +
317 in_bytes(InvocationCounter::counter_offset()));
318 __ increment_mask_and_jump(invocation_counter, increment, mask,
319 G3_scratch, Lscratch,
320 Assembler::zero, overflow);
321 __ bind(done);
322 } else {
323 // Update standard invocation counters
324 __ increment_invocation_counter(O0, G3_scratch);
325 if (ProfileInterpreter) { // %%% Merge this into MethodData*
326 Address interpreter_invocation_counter(Lmethod,in_bytes(Method::interpreter_invocation_counter_offset()));
327 __ ld(interpreter_invocation_counter, G3_scratch);
328 __ inc(G3_scratch);
329 __ st(G3_scratch, interpreter_invocation_counter);
330 }
332 if (ProfileInterpreter && profile_method != NULL) {
333 // Test to see if we should create a method data oop
334 AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
335 __ load_contents(profile_limit, G3_scratch);
336 __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
338 // if no method data exists, go to profile_method
339 __ test_method_data_pointer(*profile_method);
340 }
342 AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
343 __ load_contents(invocation_limit, G3_scratch);
344 __ cmp(O0, G3_scratch);
345 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
346 __ delayed()->nop();
347 }
349 }
351 // Allocate monitor and lock method (asm interpreter)
352 // ebx - Method*
353 //
354 void InterpreterGenerator::lock_method(void) {
355 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
357 #ifdef ASSERT
358 { Label ok;
359 __ btst(JVM_ACC_SYNCHRONIZED, O0);
360 __ br( Assembler::notZero, false, Assembler::pt, ok);
361 __ delayed()->nop();
362 __ stop("method doesn't need synchronization");
363 __ bind(ok);
364 }
365 #endif // ASSERT
367 // get synchronization object to O0
368 { Label done;
369 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
370 __ btst(JVM_ACC_STATIC, O0);
371 __ br( Assembler::zero, true, Assembler::pt, done);
372 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
374 __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0);
375 __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0);
376 __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0);
378 // lock the mirror, not the Klass*
379 __ ld_ptr( O0, mirror_offset, O0);
381 #ifdef ASSERT
382 __ tst(O0);
383 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
384 #endif // ASSERT
386 __ bind(done);
387 }
389 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
390 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
391 // __ untested("lock_object from method entry");
392 __ lock_object(Lmonitors, O0);
393 }
396 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
397 Register Rscratch,
398 Register Rscratch2) {
399 const int page_size = os::vm_page_size();
400 Label after_frame_check;
402 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
404 __ set(page_size, Rscratch);
405 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
407 // get the stack base, and in debug, verify it is non-zero
408 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
409 #ifdef ASSERT
410 Label base_not_zero;
411 __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
412 __ stop("stack base is zero in generate_stack_overflow_check");
413 __ bind(base_not_zero);
414 #endif
416 // get the stack size, and in debug, verify it is non-zero
417 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
418 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
419 #ifdef ASSERT
420 Label size_not_zero;
421 __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
422 __ stop("stack size is zero in generate_stack_overflow_check");
423 __ bind(size_not_zero);
424 #endif
426 // compute the beginning of the protected zone minus the requested frame size
427 __ sub( Rscratch, Rscratch2, Rscratch );
428 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
429 __ add( Rscratch, Rscratch2, Rscratch );
431 // Add in the size of the frame (which is the same as subtracting it from the
432 // SP, which would take another register
433 __ add( Rscratch, Rframe_size, Rscratch );
435 // the frame is greater than one page in size, so check against
436 // the bottom of the stack
437 __ cmp_and_brx_short(SP, Rscratch, Assembler::greater, Assembler::pt, after_frame_check);
439 // the stack will overflow, throw an exception
441 // Note that SP is restored to sender's sp (in the delay slot). This
442 // is necessary if the sender's frame is an extended compiled frame
443 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
444 // adaptations.
446 // Note also that the restored frame is not necessarily interpreted.
447 // Use the shared runtime version of the StackOverflowError.
448 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
449 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
450 __ jump_to(stub, Rscratch);
451 __ delayed()->mov(O5_savedSP, SP);
453 // if you get to here, then there is enough stack space
454 __ bind( after_frame_check );
455 }
458 //
459 // Generate a fixed interpreter frame. This is identical setup for interpreted
460 // methods and for native methods hence the shared code.
462 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
463 //
464 //
465 // The entry code sets up a new interpreter frame in 4 steps:
466 //
467 // 1) Increase caller's SP by for the extra local space needed:
468 // (check for overflow)
469 // Efficient implementation of xload/xstore bytecodes requires
470 // that arguments and non-argument locals are in a contigously
471 // addressable memory block => non-argument locals must be
472 // allocated in the caller's frame.
473 //
474 // 2) Create a new stack frame and register window:
475 // The new stack frame must provide space for the standard
476 // register save area, the maximum java expression stack size,
477 // the monitor slots (0 slots initially), and some frame local
478 // scratch locations.
479 //
480 // 3) The following interpreter activation registers must be setup:
481 // Lesp : expression stack pointer
482 // Lbcp : bytecode pointer
483 // Lmethod : method
484 // Llocals : locals pointer
485 // Lmonitors : monitor pointer
486 // LcpoolCache: constant pool cache
487 //
488 // 4) Initialize the non-argument locals if necessary:
489 // Non-argument locals may need to be initialized to NULL
490 // for GC to work. If the oop-map information is accurate
491 // (in the absence of the JSR problem), no initialization
492 // is necessary.
493 //
494 // (gri - 2/25/2000)
497 const Address size_of_parameters(G5_method, Method::size_of_parameters_offset());
498 const Address size_of_locals (G5_method, Method::size_of_locals_offset());
499 const Address max_stack (G5_method, Method::max_stack_offset());
500 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
502 const int extra_space =
503 rounded_vm_local_words + // frame local scratch space
504 //6815692//Method::extra_stack_words() + // extra push slots for MH adapters
505 frame::memory_parameter_word_sp_offset + // register save area
506 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
508 const Register Glocals_size = G3;
509 const Register Otmp1 = O3;
510 const Register Otmp2 = O4;
511 // Lscratch can't be used as a temporary because the call_stub uses
512 // it to assert that the stack frame was setup correctly.
514 __ lduh( size_of_parameters, Glocals_size);
516 // Gargs points to first local + BytesPerWord
517 // Set the saved SP after the register window save
518 //
519 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
520 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
521 __ add(Gargs, Otmp1, Gargs);
523 if (native_call) {
524 __ calc_mem_param_words( Glocals_size, Gframe_size );
525 __ add( Gframe_size, extra_space, Gframe_size);
526 __ round_to( Gframe_size, WordsPerLong );
527 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
528 } else {
530 //
531 // Compute number of locals in method apart from incoming parameters
532 //
533 __ lduh( size_of_locals, Otmp1 );
534 __ sub( Otmp1, Glocals_size, Glocals_size );
535 __ round_to( Glocals_size, WordsPerLong );
536 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
538 // see if the frame is greater than one page in size. If so,
539 // then we need to verify there is enough stack space remaining
540 // Frame_size = (max_stack + extra_space) * BytesPerWord;
541 __ lduh( max_stack, Gframe_size );
542 __ add( Gframe_size, extra_space, Gframe_size );
543 __ round_to( Gframe_size, WordsPerLong );
544 __ sll( Gframe_size, Interpreter::logStackElementSize, Gframe_size);
546 // Add in java locals size for stack overflow check only
547 __ add( Gframe_size, Glocals_size, Gframe_size );
549 const Register Otmp2 = O4;
550 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
551 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
553 __ sub( Gframe_size, Glocals_size, Gframe_size);
555 //
556 // bump SP to accomodate the extra locals
557 //
558 __ sub( SP, Glocals_size, SP );
559 }
561 //
562 // now set up a stack frame with the size computed above
563 //
564 __ neg( Gframe_size );
565 __ save( SP, Gframe_size, SP );
567 //
568 // now set up all the local cache registers
569 //
570 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
571 // that all present references to Lbyte_code initialize the register
572 // immediately before use
573 if (native_call) {
574 __ mov(G0, Lbcp);
575 } else {
576 __ ld_ptr(G5_method, Method::const_offset(), Lbcp);
577 __ add(Lbcp, in_bytes(ConstMethod::codes_offset()), Lbcp);
578 }
579 __ mov( G5_method, Lmethod); // set Lmethod
580 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
581 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
582 #ifdef _LP64
583 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
584 #endif
585 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
587 // setup interpreter activation registers
588 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
590 if (ProfileInterpreter) {
591 #ifdef FAST_DISPATCH
592 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
593 // they both use I2.
594 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
595 #endif // FAST_DISPATCH
596 __ set_method_data_pointer();
597 }
599 }
601 // Empty method, generate a very fast return.
603 address InterpreterGenerator::generate_empty_entry(void) {
605 // A method that does nother but return...
607 address entry = __ pc();
608 Label slow_path;
610 // do nothing for empty methods (do not even increment invocation counter)
611 if ( UseFastEmptyMethods) {
612 // If we need a safepoint check, generate full interpreter entry.
613 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
614 __ set(sync_state, G3_scratch);
615 __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
617 // Code: _return
618 __ retl();
619 __ delayed()->mov(O5_savedSP, SP);
621 __ bind(slow_path);
622 (void) generate_normal_entry(false);
624 return entry;
625 }
626 return NULL;
627 }
629 // Call an accessor method (assuming it is resolved, otherwise drop into
630 // vanilla (slow path) entry
632 // Generates code to elide accessor methods
633 // Uses G3_scratch and G1_scratch as scratch
634 address InterpreterGenerator::generate_accessor_entry(void) {
636 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
637 // parameter size = 1
638 // Note: We can only use this code if the getfield has been resolved
639 // and if we don't have a null-pointer exception => check for
640 // these conditions first and use slow path if necessary.
641 address entry = __ pc();
642 Label slow_path;
645 // XXX: for compressed oops pointer loading and decoding doesn't fit in
646 // delay slot and damages G1
647 if ( UseFastAccessorMethods && !UseCompressedOops ) {
648 // Check if we need to reach a safepoint and generate full interpreter
649 // frame if so.
650 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
651 __ load_contents(sync_state, G3_scratch);
652 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
653 __ cmp_and_br_short(G3_scratch, SafepointSynchronize::_not_synchronized, Assembler::notEqual, Assembler::pn, slow_path);
655 // Check if local 0 != NULL
656 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
657 // check if local 0 == NULL and go the slow path
658 __ br_null_short(Otos_i, Assembler::pn, slow_path);
661 // read first instruction word and extract bytecode @ 1 and index @ 2
662 // get first 4 bytes of the bytecodes (big endian!)
663 __ ld_ptr(G5_method, Method::const_offset(), G1_scratch);
664 __ ld(G1_scratch, ConstMethod::codes_offset(), G1_scratch);
666 // move index @ 2 far left then to the right most two bytes.
667 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
668 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
669 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
671 // get constant pool cache
672 __ ld_ptr(G5_method, Method::const_offset(), G3_scratch);
673 __ ld_ptr(G3_scratch, ConstMethod::constants_offset(), G3_scratch);
674 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
676 // get specific constant pool cache entry
677 __ add(G3_scratch, G1_scratch, G3_scratch);
679 // Check the constant Pool cache entry to see if it has been resolved.
680 // If not, need the slow path.
681 ByteSize cp_base_offset = ConstantPoolCache::base_offset();
682 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
683 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
684 __ and3(G1_scratch, 0xFF, G1_scratch);
685 __ cmp_and_br_short(G1_scratch, Bytecodes::_getfield, Assembler::notEqual, Assembler::pn, slow_path);
687 // Get the type and return field offset from the constant pool cache
688 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
689 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
691 Label xreturn_path;
692 // Need to differentiate between igetfield, agetfield, bgetfield etc.
693 // because they are different sizes.
694 // Get the type from the constant pool cache
695 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
696 // Make sure we don't need to mask G1_scratch after the above shift
697 ConstantPoolCacheEntry::verify_tos_state_shift();
698 __ cmp(G1_scratch, atos );
699 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
700 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
701 __ cmp(G1_scratch, itos);
702 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
703 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
704 __ cmp(G1_scratch, stos);
705 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
706 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
707 __ cmp(G1_scratch, ctos);
708 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
709 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
710 #ifdef ASSERT
711 __ cmp(G1_scratch, btos);
712 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
713 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
714 __ should_not_reach_here();
715 #endif
716 __ ldsb(Otos_i, G3_scratch, Otos_i);
717 __ bind(xreturn_path);
719 // _ireturn/_areturn
720 __ retl(); // return from leaf routine
721 __ delayed()->mov(O5_savedSP, SP);
723 // Generate regular method entry
724 __ bind(slow_path);
725 (void) generate_normal_entry(false);
726 return entry;
727 }
728 return NULL;
729 }
731 // Method entry for java.lang.ref.Reference.get.
732 address InterpreterGenerator::generate_Reference_get_entry(void) {
733 #ifndef SERIALGC
734 // Code: _aload_0, _getfield, _areturn
735 // parameter size = 1
736 //
737 // The code that gets generated by this routine is split into 2 parts:
738 // 1. The "intrinsified" code for G1 (or any SATB based GC),
739 // 2. The slow path - which is an expansion of the regular method entry.
740 //
741 // Notes:-
742 // * In the G1 code we do not check whether we need to block for
743 // a safepoint. If G1 is enabled then we must execute the specialized
744 // code for Reference.get (except when the Reference object is null)
745 // so that we can log the value in the referent field with an SATB
746 // update buffer.
747 // If the code for the getfield template is modified so that the
748 // G1 pre-barrier code is executed when the current method is
749 // Reference.get() then going through the normal method entry
750 // will be fine.
751 // * The G1 code can, however, check the receiver object (the instance
752 // of java.lang.Reference) and jump to the slow path if null. If the
753 // Reference object is null then we obviously cannot fetch the referent
754 // and so we don't need to call the G1 pre-barrier. Thus we can use the
755 // regular method entry code to generate the NPE.
756 //
757 // This code is based on generate_accessor_enty.
759 address entry = __ pc();
761 const int referent_offset = java_lang_ref_Reference::referent_offset;
762 guarantee(referent_offset > 0, "referent offset not initialized");
764 if (UseG1GC) {
765 Label slow_path;
767 // In the G1 code we don't check if we need to reach a safepoint. We
768 // continue and the thread will safepoint at the next bytecode dispatch.
770 // Check if local 0 != NULL
771 // If the receiver is null then it is OK to jump to the slow path.
772 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
773 // check if local 0 == NULL and go the slow path
774 __ cmp_and_brx_short(Otos_i, 0, Assembler::equal, Assembler::pn, slow_path);
777 // Load the value of the referent field.
778 if (Assembler::is_simm13(referent_offset)) {
779 __ load_heap_oop(Otos_i, referent_offset, Otos_i);
780 } else {
781 __ set(referent_offset, G3_scratch);
782 __ load_heap_oop(Otos_i, G3_scratch, Otos_i);
783 }
785 // Generate the G1 pre-barrier code to log the value of
786 // the referent field in an SATB buffer. Note with
787 // these parameters the pre-barrier does not generate
788 // the load of the previous value
790 __ g1_write_barrier_pre(noreg /* obj */, noreg /* index */, 0 /* offset */,
791 Otos_i /* pre_val */,
792 G3_scratch /* tmp */,
793 true /* preserve_o_regs */);
795 // _areturn
796 __ retl(); // return from leaf routine
797 __ delayed()->mov(O5_savedSP, SP);
799 // Generate regular method entry
800 __ bind(slow_path);
801 (void) generate_normal_entry(false);
802 return entry;
803 }
804 #endif // SERIALGC
806 // If G1 is not enabled then attempt to go through the accessor entry point
807 // Reference.get is an accessor
808 return generate_accessor_entry();
809 }
811 //
812 // Interpreter stub for calling a native method. (asm interpreter)
813 // This sets up a somewhat different looking stack for calling the native method
814 // than the typical interpreter frame setup.
815 //
817 address InterpreterGenerator::generate_native_entry(bool synchronized) {
818 address entry = __ pc();
820 // the following temporary registers are used during frame creation
821 const Register Gtmp1 = G3_scratch ;
822 const Register Gtmp2 = G1_scratch;
823 bool inc_counter = UseCompiler || CountCompiledCalls;
825 // make sure registers are different!
826 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
828 const Address Laccess_flags(Lmethod, Method::access_flags_offset());
830 const Register Glocals_size = G3;
831 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
833 // make sure method is native & not abstract
834 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
835 #ifdef ASSERT
836 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
837 {
838 Label L;
839 __ btst(JVM_ACC_NATIVE, Gtmp1);
840 __ br(Assembler::notZero, false, Assembler::pt, L);
841 __ delayed()->nop();
842 __ stop("tried to execute non-native method as native");
843 __ bind(L);
844 }
845 { Label L;
846 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
847 __ br(Assembler::zero, false, Assembler::pt, L);
848 __ delayed()->nop();
849 __ stop("tried to execute abstract method as non-abstract");
850 __ bind(L);
851 }
852 #endif // ASSERT
854 // generate the code to allocate the interpreter stack frame
855 generate_fixed_frame(true);
857 //
858 // No locals to initialize for native method
859 //
861 // this slot will be set later, we initialize it to null here just in
862 // case we get a GC before the actual value is stored later
863 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
865 const Address do_not_unlock_if_synchronized(G2_thread,
866 JavaThread::do_not_unlock_if_synchronized_offset());
867 // Since at this point in the method invocation the exception handler
868 // would try to exit the monitor of synchronized methods which hasn't
869 // been entered yet, we set the thread local variable
870 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
871 // runtime, exception handling i.e. unlock_if_synchronized_method will
872 // check this thread local flag.
873 // This flag has two effects, one is to force an unwind in the topmost
874 // interpreter frame and not perform an unlock while doing so.
876 __ movbool(true, G3_scratch);
877 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
879 // increment invocation counter and check for overflow
880 //
881 // Note: checking for negative value instead of overflow
882 // so we have a 'sticky' overflow test (may be of
883 // importance as soon as we have true MT/MP)
884 Label invocation_counter_overflow;
885 Label Lcontinue;
886 if (inc_counter) {
887 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
889 }
890 __ bind(Lcontinue);
892 bang_stack_shadow_pages(true);
894 // reset the _do_not_unlock_if_synchronized flag
895 __ stbool(G0, do_not_unlock_if_synchronized);
897 // check for synchronized methods
898 // Must happen AFTER invocation_counter check and stack overflow check,
899 // so method is not locked if overflows.
901 if (synchronized) {
902 lock_method();
903 } else {
904 #ifdef ASSERT
905 { Label ok;
906 __ ld(Laccess_flags, O0);
907 __ btst(JVM_ACC_SYNCHRONIZED, O0);
908 __ br( Assembler::zero, false, Assembler::pt, ok);
909 __ delayed()->nop();
910 __ stop("method needs synchronization");
911 __ bind(ok);
912 }
913 #endif // ASSERT
914 }
917 // start execution
918 __ verify_thread();
920 // JVMTI support
921 __ notify_method_entry();
923 // native call
925 // (note that O0 is never an oop--at most it is a handle)
926 // It is important not to smash any handles created by this call,
927 // until any oop handle in O0 is dereferenced.
929 // (note that the space for outgoing params is preallocated)
931 // get signature handler
932 { Label L;
933 Address signature_handler(Lmethod, Method::signature_handler_offset());
934 __ ld_ptr(signature_handler, G3_scratch);
935 __ br_notnull_short(G3_scratch, Assembler::pt, L);
936 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
937 __ ld_ptr(signature_handler, G3_scratch);
938 __ bind(L);
939 }
941 // Push a new frame so that the args will really be stored in
942 // Copy a few locals across so the new frame has the variables
943 // we need but these values will be dead at the jni call and
944 // therefore not gc volatile like the values in the current
945 // frame (Lmethod in particular)
947 // Flush the method pointer to the register save area
948 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
949 __ mov(Llocals, O1);
951 // calculate where the mirror handle body is allocated in the interpreter frame:
952 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
954 // Calculate current frame size
955 __ sub(SP, FP, O3); // Calculate negative of current frame size
956 __ save(SP, O3, SP); // Allocate an identical sized frame
958 // Note I7 has leftover trash. Slow signature handler will fill it in
959 // should we get there. Normal jni call will set reasonable last_Java_pc
960 // below (and fix I7 so the stack trace doesn't have a meaningless frame
961 // in it).
963 // Load interpreter frame's Lmethod into same register here
965 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
967 __ mov(I1, Llocals);
968 __ mov(I2, Lscratch2); // save the address of the mirror
971 // ONLY Lmethod and Llocals are valid here!
973 // call signature handler, It will move the arg properly since Llocals in current frame
974 // matches that in outer frame
976 __ callr(G3_scratch, 0);
977 __ delayed()->nop();
979 // Result handler is in Lscratch
981 // Reload interpreter frame's Lmethod since slow signature handler may block
982 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
984 { Label not_static;
986 __ ld(Laccess_flags, O0);
987 __ btst(JVM_ACC_STATIC, O0);
988 __ br( Assembler::zero, false, Assembler::pt, not_static);
989 // get native function entry point(O0 is a good temp until the very end)
990 __ delayed()->ld_ptr(Lmethod, in_bytes(Method::native_function_offset()), O0);
991 // for static methods insert the mirror argument
992 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
994 __ ld_ptr(Lmethod, Method:: const_offset(), O1);
995 __ ld_ptr(O1, ConstMethod::constants_offset(), O1);
996 __ ld_ptr(O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
997 __ ld_ptr(O1, mirror_offset, O1);
998 #ifdef ASSERT
999 if (!PrintSignatureHandlers) // do not dirty the output with this
1000 { Label L;
1001 __ br_notnull_short(O1, Assembler::pt, L);
1002 __ stop("mirror is missing");
1003 __ bind(L);
1004 }
1005 #endif // ASSERT
1006 __ st_ptr(O1, Lscratch2, 0);
1007 __ mov(Lscratch2, O1);
1008 __ bind(not_static);
1009 }
1011 // At this point, arguments have been copied off of stack into
1012 // their JNI positions, which are O1..O5 and SP[68..].
1013 // Oops are boxed in-place on the stack, with handles copied to arguments.
1014 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
1016 #ifdef ASSERT
1017 { Label L;
1018 __ br_notnull_short(O0, Assembler::pt, L);
1019 __ stop("native entry point is missing");
1020 __ bind(L);
1021 }
1022 #endif // ASSERT
1024 //
1025 // setup the frame anchor
1026 //
1027 // The scavenge function only needs to know that the PC of this frame is
1028 // in the interpreter method entry code, it doesn't need to know the exact
1029 // PC and hence we can use O7 which points to the return address from the
1030 // previous call in the code stream (signature handler function)
1031 //
1032 // The other trick is we set last_Java_sp to FP instead of the usual SP because
1033 // we have pushed the extra frame in order to protect the volatile register(s)
1034 // in that frame when we return from the jni call
1035 //
1037 __ set_last_Java_frame(FP, O7);
1038 __ mov(O7, I7); // make dummy interpreter frame look like one above,
1039 // not meaningless information that'll confuse me.
1041 // flush the windows now. We don't care about the current (protection) frame
1042 // only the outer frames
1044 __ flush_windows();
1046 // mark windows as flushed
1047 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
1048 __ set(JavaFrameAnchor::flushed, G3_scratch);
1049 __ st(G3_scratch, flags);
1051 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
1053 Address thread_state(G2_thread, JavaThread::thread_state_offset());
1054 #ifdef ASSERT
1055 { Label L;
1056 __ ld(thread_state, G3_scratch);
1057 __ cmp_and_br_short(G3_scratch, _thread_in_Java, Assembler::equal, Assembler::pt, L);
1058 __ stop("Wrong thread state in native stub");
1059 __ bind(L);
1060 }
1061 #endif // ASSERT
1062 __ set(_thread_in_native, G3_scratch);
1063 __ st(G3_scratch, thread_state);
1065 // Call the jni method, using the delay slot to set the JNIEnv* argument.
1066 __ save_thread(L7_thread_cache); // save Gthread
1067 __ callr(O0, 0);
1068 __ delayed()->
1069 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
1071 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
1073 __ restore_thread(L7_thread_cache); // restore G2_thread
1074 __ reinit_heapbase();
1076 // must we block?
1078 // Block, if necessary, before resuming in _thread_in_Java state.
1079 // In order for GC to work, don't clear the last_Java_sp until after blocking.
1080 { Label no_block;
1081 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
1083 // Switch thread to "native transition" state before reading the synchronization state.
1084 // This additional state is necessary because reading and testing the synchronization
1085 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1086 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1087 // VM thread changes sync state to synchronizing and suspends threads for GC.
1088 // Thread A is resumed to finish this native method, but doesn't block here since it
1089 // didn't see any synchronization is progress, and escapes.
1090 __ set(_thread_in_native_trans, G3_scratch);
1091 __ st(G3_scratch, thread_state);
1092 if(os::is_MP()) {
1093 if (UseMembar) {
1094 // Force this write out before the read below
1095 __ membar(Assembler::StoreLoad);
1096 } else {
1097 // Write serialization page so VM thread can do a pseudo remote membar.
1098 // We use the current thread pointer to calculate a thread specific
1099 // offset to write to within the page. This minimizes bus traffic
1100 // due to cache line collision.
1101 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1102 }
1103 }
1104 __ load_contents(sync_state, G3_scratch);
1105 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1107 Label L;
1108 __ br(Assembler::notEqual, false, Assembler::pn, L);
1109 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1110 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
1111 __ bind(L);
1113 // Block. Save any potential method result value before the operation and
1114 // use a leaf call to leave the last_Java_frame setup undisturbed.
1115 save_native_result();
1116 __ call_VM_leaf(L7_thread_cache,
1117 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1118 G2_thread);
1120 // Restore any method result value
1121 restore_native_result();
1122 __ bind(no_block);
1123 }
1125 // Clear the frame anchor now
1127 __ reset_last_Java_frame();
1129 // Move the result handler address
1130 __ mov(Lscratch, G3_scratch);
1131 // return possible result to the outer frame
1132 #ifndef __LP64
1133 __ mov(O0, I0);
1134 __ restore(O1, G0, O1);
1135 #else
1136 __ restore(O0, G0, O0);
1137 #endif /* __LP64 */
1139 // Move result handler to expected register
1140 __ mov(G3_scratch, Lscratch);
1142 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1143 // switch to thread_in_Java.
1145 __ set(_thread_in_Java, G3_scratch);
1146 __ st(G3_scratch, thread_state);
1148 // reset handle block
1149 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1150 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1152 // If we have an oop result store it where it will be safe for any further gc
1153 // until we return now that we've released the handle it might be protected by
1155 {
1156 Label no_oop, store_result;
1158 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1159 __ cmp_and_brx_short(G3_scratch, Lscratch, Assembler::notEqual, Assembler::pt, no_oop);
1160 __ addcc(G0, O0, O0);
1161 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1162 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1163 __ mov(G0, O0);
1165 __ bind(store_result);
1166 // Store it where gc will look for it and result handler expects it.
1167 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1169 __ bind(no_oop);
1171 }
1174 // handle exceptions (exception handling will handle unlocking!)
1175 { Label L;
1176 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1177 __ ld_ptr(exception_addr, Gtemp);
1178 __ br_null_short(Gtemp, Assembler::pt, L);
1179 // Note: This could be handled more efficiently since we know that the native
1180 // method doesn't have an exception handler. We could directly return
1181 // to the exception handler for the caller.
1182 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1183 __ should_not_reach_here();
1184 __ bind(L);
1185 }
1187 // JVMTI support (preserves thread register)
1188 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1190 if (synchronized) {
1191 // save and restore any potential method result value around the unlocking operation
1192 save_native_result();
1194 __ add( __ top_most_monitor(), O1);
1195 __ unlock_object(O1);
1197 restore_native_result();
1198 }
1200 #if defined(COMPILER2) && !defined(_LP64)
1202 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1203 // or compiled so just be safe.
1205 __ sllx(O0, 32, G1); // Shift bits into high G1
1206 __ srl (O1, 0, O1); // Zero extend O1
1207 __ or3 (O1, G1, G1); // OR 64 bits into G1
1209 #endif /* COMPILER2 && !_LP64 */
1211 // dispose of return address and remove activation
1212 #ifdef ASSERT
1213 {
1214 Label ok;
1215 __ cmp_and_brx_short(I5_savedSP, FP, Assembler::greaterEqualUnsigned, Assembler::pt, ok);
1216 __ stop("bad I5_savedSP value");
1217 __ should_not_reach_here();
1218 __ bind(ok);
1219 }
1220 #endif
1221 if (TraceJumps) {
1222 // Move target to register that is recordable
1223 __ mov(Lscratch, G3_scratch);
1224 __ JMP(G3_scratch, 0);
1225 } else {
1226 __ jmp(Lscratch, 0);
1227 }
1228 __ delayed()->nop();
1231 if (inc_counter) {
1232 // handle invocation counter overflow
1233 __ bind(invocation_counter_overflow);
1234 generate_counter_overflow(Lcontinue);
1235 }
1239 return entry;
1240 }
1243 // Generic method entry to (asm) interpreter
1244 //------------------------------------------------------------------------------------------------------------------------
1245 //
1246 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1247 address entry = __ pc();
1249 bool inc_counter = UseCompiler || CountCompiledCalls;
1251 // the following temporary registers are used during frame creation
1252 const Register Gtmp1 = G3_scratch ;
1253 const Register Gtmp2 = G1_scratch;
1255 // make sure registers are different!
1256 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1258 const Address size_of_parameters(G5_method, Method::size_of_parameters_offset());
1259 const Address size_of_locals (G5_method, Method::size_of_locals_offset());
1260 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1261 // and use in the asserts.
1262 const Address access_flags (Lmethod, Method::access_flags_offset());
1264 const Register Glocals_size = G3;
1265 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1267 // make sure method is not native & not abstract
1268 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1269 #ifdef ASSERT
1270 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1271 {
1272 Label L;
1273 __ btst(JVM_ACC_NATIVE, Gtmp1);
1274 __ br(Assembler::zero, false, Assembler::pt, L);
1275 __ delayed()->nop();
1276 __ stop("tried to execute native method as non-native");
1277 __ bind(L);
1278 }
1279 { Label L;
1280 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1281 __ br(Assembler::zero, false, Assembler::pt, L);
1282 __ delayed()->nop();
1283 __ stop("tried to execute abstract method as non-abstract");
1284 __ bind(L);
1285 }
1286 #endif // ASSERT
1288 // generate the code to allocate the interpreter stack frame
1290 generate_fixed_frame(false);
1292 #ifdef FAST_DISPATCH
1293 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1294 // set bytecode dispatch table base
1295 #endif
1297 //
1298 // Code to initialize the extra (i.e. non-parm) locals
1299 //
1300 Register init_value = noreg; // will be G0 if we must clear locals
1301 // The way the code was setup before zerolocals was always true for vanilla java entries.
1302 // It could only be false for the specialized entries like accessor or empty which have
1303 // no extra locals so the testing was a waste of time and the extra locals were always
1304 // initialized. We removed this extra complication to already over complicated code.
1306 init_value = G0;
1307 Label clear_loop;
1309 // NOTE: If you change the frame layout, this code will need to
1310 // be updated!
1311 __ lduh( size_of_locals, O2 );
1312 __ lduh( size_of_parameters, O1 );
1313 __ sll( O2, Interpreter::logStackElementSize, O2);
1314 __ sll( O1, Interpreter::logStackElementSize, O1 );
1315 __ sub( Llocals, O2, O2 );
1316 __ sub( Llocals, O1, O1 );
1318 __ bind( clear_loop );
1319 __ inc( O2, wordSize );
1321 __ cmp( O2, O1 );
1322 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1323 __ delayed()->st_ptr( init_value, O2, 0 );
1325 const Address do_not_unlock_if_synchronized(G2_thread,
1326 JavaThread::do_not_unlock_if_synchronized_offset());
1327 // Since at this point in the method invocation the exception handler
1328 // would try to exit the monitor of synchronized methods which hasn't
1329 // been entered yet, we set the thread local variable
1330 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1331 // runtime, exception handling i.e. unlock_if_synchronized_method will
1332 // check this thread local flag.
1333 __ movbool(true, G3_scratch);
1334 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1336 // increment invocation counter and check for overflow
1337 //
1338 // Note: checking for negative value instead of overflow
1339 // so we have a 'sticky' overflow test (may be of
1340 // importance as soon as we have true MT/MP)
1341 Label invocation_counter_overflow;
1342 Label profile_method;
1343 Label profile_method_continue;
1344 Label Lcontinue;
1345 if (inc_counter) {
1346 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1347 if (ProfileInterpreter) {
1348 __ bind(profile_method_continue);
1349 }
1350 }
1351 __ bind(Lcontinue);
1353 bang_stack_shadow_pages(false);
1355 // reset the _do_not_unlock_if_synchronized flag
1356 __ stbool(G0, do_not_unlock_if_synchronized);
1358 // check for synchronized methods
1359 // Must happen AFTER invocation_counter check and stack overflow check,
1360 // so method is not locked if overflows.
1362 if (synchronized) {
1363 lock_method();
1364 } else {
1365 #ifdef ASSERT
1366 { Label ok;
1367 __ ld(access_flags, O0);
1368 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1369 __ br( Assembler::zero, false, Assembler::pt, ok);
1370 __ delayed()->nop();
1371 __ stop("method needs synchronization");
1372 __ bind(ok);
1373 }
1374 #endif // ASSERT
1375 }
1377 // start execution
1379 __ verify_thread();
1381 // jvmti support
1382 __ notify_method_entry();
1384 // start executing instructions
1385 __ dispatch_next(vtos);
1388 if (inc_counter) {
1389 if (ProfileInterpreter) {
1390 // We have decided to profile this method in the interpreter
1391 __ bind(profile_method);
1393 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1394 __ set_method_data_pointer_for_bcp();
1395 __ ba_short(profile_method_continue);
1396 }
1398 // handle invocation counter overflow
1399 __ bind(invocation_counter_overflow);
1400 generate_counter_overflow(Lcontinue);
1401 }
1404 return entry;
1405 }
1408 //----------------------------------------------------------------------------------------------------
1409 // Entry points & stack frame layout
1410 //
1411 // Here we generate the various kind of entries into the interpreter.
1412 // The two main entry type are generic bytecode methods and native call method.
1413 // These both come in synchronized and non-synchronized versions but the
1414 // frame layout they create is very similar. The other method entry
1415 // types are really just special purpose entries that are really entry
1416 // and interpretation all in one. These are for trivial methods like
1417 // accessor, empty, or special math methods.
1418 //
1419 // When control flow reaches any of the entry types for the interpreter
1420 // the following holds ->
1421 //
1422 // C2 Calling Conventions:
1423 //
1424 // The entry code below assumes that the following registers are set
1425 // when coming in:
1426 // G5_method: holds the Method* of the method to call
1427 // Lesp: points to the TOS of the callers expression stack
1428 // after having pushed all the parameters
1429 //
1430 // The entry code does the following to setup an interpreter frame
1431 // pop parameters from the callers stack by adjusting Lesp
1432 // set O0 to Lesp
1433 // compute X = (max_locals - num_parameters)
1434 // bump SP up by X to accomadate the extra locals
1435 // compute X = max_expression_stack
1436 // + vm_local_words
1437 // + 16 words of register save area
1438 // save frame doing a save sp, -X, sp growing towards lower addresses
1439 // set Lbcp, Lmethod, LcpoolCache
1440 // set Llocals to i0
1441 // set Lmonitors to FP - rounded_vm_local_words
1442 // set Lesp to Lmonitors - 4
1443 //
1444 // The frame has now been setup to do the rest of the entry code
1446 // Try this optimization: Most method entries could live in a
1447 // "one size fits all" stack frame without all the dynamic size
1448 // calculations. It might be profitable to do all this calculation
1449 // statically and approximately for "small enough" methods.
1451 //-----------------------------------------------------------------------------------------------
1453 // C1 Calling conventions
1454 //
1455 // Upon method entry, the following registers are setup:
1456 //
1457 // g2 G2_thread: current thread
1458 // g5 G5_method: method to activate
1459 // g4 Gargs : pointer to last argument
1460 //
1461 //
1462 // Stack:
1463 //
1464 // +---------------+ <--- sp
1465 // | |
1466 // : reg save area :
1467 // | |
1468 // +---------------+ <--- sp + 0x40
1469 // | |
1470 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1471 // | |
1472 // +---------------+ <--- sp + 0x5c
1473 // | |
1474 // : free :
1475 // | |
1476 // +---------------+ <--- Gargs
1477 // | |
1478 // : arguments :
1479 // | |
1480 // +---------------+
1481 // | |
1482 //
1483 //
1484 //
1485 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1486 //
1487 // +---------------+ <--- sp
1488 // | |
1489 // : reg save area :
1490 // | |
1491 // +---------------+ <--- sp + 0x40
1492 // | |
1493 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1494 // | |
1495 // +---------------+ <--- sp + 0x5c
1496 // | |
1497 // : :
1498 // | | <--- Lesp
1499 // +---------------+ <--- Lmonitors (fp - 0x18)
1500 // | VM locals |
1501 // +---------------+ <--- fp
1502 // | |
1503 // : reg save area :
1504 // | |
1505 // +---------------+ <--- fp + 0x40
1506 // | |
1507 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1508 // | |
1509 // +---------------+ <--- fp + 0x5c
1510 // | |
1511 // : free :
1512 // | |
1513 // +---------------+
1514 // | |
1515 // : nonarg locals :
1516 // | |
1517 // +---------------+
1518 // | |
1519 // : arguments :
1520 // | | <--- Llocals
1521 // +---------------+ <--- Gargs
1522 // | |
1524 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1526 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1527 // expression stack, the callee will have callee_extra_locals (so we can account for
1528 // frame extension) and monitor_size for monitors. Basically we need to calculate
1529 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1530 //
1531 //
1532 // The big complicating thing here is that we must ensure that the stack stays properly
1533 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1534 // needs to be aligned for). We are given that the sp (fp) is already aligned by
1535 // the caller so we must ensure that it is properly aligned for our callee.
1536 //
1537 const int rounded_vm_local_words =
1538 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1539 // callee_locals and max_stack are counts, not the size in frame.
1540 const int locals_size =
1541 round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1542 const int max_stack_words = max_stack * Interpreter::stackElementWords;
1543 return (round_to((max_stack_words
1544 //6815692//+ Method::extra_stack_words()
1545 + rounded_vm_local_words
1546 + frame::memory_parameter_word_sp_offset), WordsPerLong)
1547 // already rounded
1548 + locals_size + monitor_size);
1549 }
1551 // How much stack a method top interpreter activation needs in words.
1552 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1554 // See call_stub code
1555 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
1556 WordsPerLong); // 7 + register save area
1558 // Save space for one monitor to get into the interpreted method in case
1559 // the method is synchronized
1560 int monitor_size = method->is_synchronized() ?
1561 1*frame::interpreter_frame_monitor_size() : 0;
1562 return size_activation_helper(method->max_locals(), method->max_stack(),
1563 monitor_size) + call_stub_size;
1564 }
1566 int AbstractInterpreter::layout_activation(Method* method,
1567 int tempcount,
1568 int popframe_extra_args,
1569 int moncount,
1570 int caller_actual_parameters,
1571 int callee_param_count,
1572 int callee_local_count,
1573 frame* caller,
1574 frame* interpreter_frame,
1575 bool is_top_frame) {
1576 // Note: This calculation must exactly parallel the frame setup
1577 // in InterpreterGenerator::generate_fixed_frame.
1578 // If f!=NULL, set up the following variables:
1579 // - Lmethod
1580 // - Llocals
1581 // - Lmonitors (to the indicated number of monitors)
1582 // - Lesp (to the indicated number of temps)
1583 // The frame f (if not NULL) on entry is a description of the caller of the frame
1584 // we are about to layout. We are guaranteed that we will be able to fill in a
1585 // new interpreter frame as its callee (i.e. the stack space is allocated and
1586 // the amount was determined by an earlier call to this method with f == NULL).
1587 // On return f (if not NULL) while describe the interpreter frame we just layed out.
1589 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
1590 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1592 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1593 //
1594 // Note: if you look closely this appears to be doing something much different
1595 // than generate_fixed_frame. What is happening is this. On sparc we have to do
1596 // this dance with interpreter_sp_adjustment because the window save area would
1597 // appear just below the bottom (tos) of the caller's java expression stack. Because
1598 // the interpreter want to have the locals completely contiguous generate_fixed_frame
1599 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1600 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1601 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1602 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1603 // because the oldest frame would have adjust its callers frame and yet that frame
1604 // already exists and isn't part of this array of frames we are unpacking. So at first
1605 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1606 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1607 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1608 // add up. It does seem like it simpler to account for the adjustment here (and remove the
1609 // callee... parameters here). However this would mean that this routine would have to take
1610 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1611 // and run the calling loop in the reverse order. This would also would appear to mean making
1612 // this code aware of what the interactions are when that initial caller fram was an osr or
1613 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
1614 // there is no sense in messing working code.
1615 //
1617 int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
1618 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1620 int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
1621 monitor_size);
1623 if (interpreter_frame != NULL) {
1624 // The skeleton frame must already look like an interpreter frame
1625 // even if not fully filled out.
1626 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1628 intptr_t* fp = interpreter_frame->fp();
1630 JavaThread* thread = JavaThread::current();
1631 RegisterMap map(thread, false);
1632 // More verification that skeleton frame is properly walkable
1633 assert(fp == caller->sp(), "fp must match");
1635 intptr_t* montop = fp - rounded_vm_local_words;
1637 // preallocate monitors (cf. __ add_monitor_to_stack)
1638 intptr_t* monitors = montop - monitor_size;
1640 // preallocate stack space
1641 intptr_t* esp = monitors - 1 -
1642 (tempcount * Interpreter::stackElementWords) -
1643 popframe_extra_args;
1645 int local_words = method->max_locals() * Interpreter::stackElementWords;
1646 NEEDS_CLEANUP;
1647 intptr_t* locals;
1648 if (caller->is_interpreted_frame()) {
1649 // Can force the locals area to end up properly overlapping the top of the expression stack.
1650 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1651 // Note that this computation means we replace size_of_parameters() values from the caller
1652 // interpreter frame's expression stack with our argument locals
1653 int parm_words = caller_actual_parameters * Interpreter::stackElementWords;
1654 locals = Lesp_ptr + parm_words;
1655 int delta = local_words - parm_words;
1656 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1657 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1658 } else {
1659 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1660 // Don't have Lesp available; lay out locals block in the caller
1661 // adjacent to the register window save area.
1662 //
1663 // Compiled frames do not allocate a varargs area which is why this if
1664 // statement is needed.
1665 //
1666 if (caller->is_compiled_frame()) {
1667 locals = fp + frame::register_save_words + local_words - 1;
1668 } else {
1669 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1670 }
1671 if (!caller->is_entry_frame()) {
1672 // Caller wants his own SP back
1673 int caller_frame_size = caller->cb()->frame_size();
1674 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1675 }
1676 }
1677 if (TraceDeoptimization) {
1678 if (caller->is_entry_frame()) {
1679 // make sure I5_savedSP and the entry frames notion of saved SP
1680 // agree. This assertion duplicate a check in entry frame code
1681 // but catches the failure earlier.
1682 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1683 "would change callers SP");
1684 }
1685 if (caller->is_entry_frame()) {
1686 tty->print("entry ");
1687 }
1688 if (caller->is_compiled_frame()) {
1689 tty->print("compiled ");
1690 if (caller->is_deoptimized_frame()) {
1691 tty->print("(deopt) ");
1692 }
1693 }
1694 if (caller->is_interpreted_frame()) {
1695 tty->print("interpreted ");
1696 }
1697 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1698 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1699 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1700 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1701 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1702 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1703 tty->print_cr("Llocals = 0x%x", locals);
1704 tty->print_cr("Lesp = 0x%x", esp);
1705 tty->print_cr("Lmonitors = 0x%x", monitors);
1706 }
1708 if (method->max_locals() > 0) {
1709 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1710 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1711 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1712 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1713 }
1714 #ifdef _LP64
1715 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1716 #endif
1718 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
1719 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
1720 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
1721 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
1722 // Llast_SP will be same as SP as there is no adapter space
1723 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1724 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1725 #ifdef FAST_DISPATCH
1726 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1727 #endif
1730 #ifdef ASSERT
1731 BasicObjectLock* mp = (BasicObjectLock*)monitors;
1733 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1734 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1735 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
1736 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1737 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1739 // check bounds
1740 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1741 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1742 assert(lo < monitors && montop <= hi, "monitors in bounds");
1743 assert(lo <= esp && esp < monitors, "esp in bounds");
1744 #endif // ASSERT
1745 }
1747 return raw_frame_size;
1748 }
1750 //----------------------------------------------------------------------------------------------------
1751 // Exceptions
1752 void TemplateInterpreterGenerator::generate_throw_exception() {
1754 // Entry point in previous activation (i.e., if the caller was interpreted)
1755 Interpreter::_rethrow_exception_entry = __ pc();
1756 // O0: exception
1758 // entry point for exceptions thrown within interpreter code
1759 Interpreter::_throw_exception_entry = __ pc();
1760 __ verify_thread();
1761 // expression stack is undefined here
1762 // O0: exception, i.e. Oexception
1763 // Lbcp: exception bcx
1764 __ verify_oop(Oexception);
1767 // expression stack must be empty before entering the VM in case of an exception
1768 __ empty_expression_stack();
1769 // find exception handler address and preserve exception oop
1770 // call C routine to find handler and jump to it
1771 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1772 __ push_ptr(O1); // push exception for exception handler bytecodes
1774 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1775 __ delayed()->nop();
1778 // if the exception is not handled in the current frame
1779 // the frame is removed and the exception is rethrown
1780 // (i.e. exception continuation is _rethrow_exception)
1781 //
1782 // Note: At this point the bci is still the bxi for the instruction which caused
1783 // the exception and the expression stack is empty. Thus, for any VM calls
1784 // at this point, GC will find a legal oop map (with empty expression stack).
1786 // in current activation
1787 // tos: exception
1788 // Lbcp: exception bcp
1790 //
1791 // JVMTI PopFrame support
1792 //
1794 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1795 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1796 // Set the popframe_processing bit in popframe_condition indicating that we are
1797 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1798 // popframe handling cycles.
1800 __ ld(popframe_condition_addr, G3_scratch);
1801 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1802 __ stw(G3_scratch, popframe_condition_addr);
1804 // Empty the expression stack, as in normal exception handling
1805 __ empty_expression_stack();
1806 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1808 {
1809 // Check to see whether we are returning to a deoptimized frame.
1810 // (The PopFrame call ensures that the caller of the popped frame is
1811 // either interpreted or compiled and deoptimizes it if compiled.)
1812 // In this case, we can't call dispatch_next() after the frame is
1813 // popped, but instead must save the incoming arguments and restore
1814 // them after deoptimization has occurred.
1815 //
1816 // Note that we don't compare the return PC against the
1817 // deoptimization blob's unpack entry because of the presence of
1818 // adapter frames in C2.
1819 Label caller_not_deoptimized;
1820 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1821 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1823 const Register Gtmp1 = G3_scratch;
1824 const Register Gtmp2 = G1_scratch;
1826 // Compute size of arguments for saving when returning to deoptimized caller
1827 __ lduh(Lmethod, in_bytes(Method::size_of_parameters_offset()), Gtmp1);
1828 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1829 __ sub(Llocals, Gtmp1, Gtmp2);
1830 __ add(Gtmp2, wordSize, Gtmp2);
1831 // Save these arguments
1832 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1833 // Inform deoptimization that it is responsible for restoring these arguments
1834 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1835 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1836 __ st(Gtmp1, popframe_condition_addr);
1838 // Return from the current method
1839 // The caller's SP was adjusted upon method entry to accomodate
1840 // the callee's non-argument locals. Undo that adjustment.
1841 __ ret();
1842 __ delayed()->restore(I5_savedSP, G0, SP);
1844 __ bind(caller_not_deoptimized);
1845 }
1847 // Clear the popframe condition flag
1848 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1850 // Get out of the current method (how this is done depends on the particular compiler calling
1851 // convention that the interpreter currently follows)
1852 // The caller's SP was adjusted upon method entry to accomodate
1853 // the callee's non-argument locals. Undo that adjustment.
1854 __ restore(I5_savedSP, G0, SP);
1855 // The method data pointer was incremented already during
1856 // call profiling. We have to restore the mdp for the current bcp.
1857 if (ProfileInterpreter) {
1858 __ set_method_data_pointer_for_bcp();
1859 }
1860 // Resume bytecode interpretation at the current bcp
1861 __ dispatch_next(vtos);
1862 // end of JVMTI PopFrame support
1864 Interpreter::_remove_activation_entry = __ pc();
1866 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1867 __ pop_ptr(Oexception); // get exception
1869 // Intel has the following comment:
1870 //// remove the activation (without doing throws on illegalMonitorExceptions)
1871 // They remove the activation without checking for bad monitor state.
1872 // %%% We should make sure this is the right semantics before implementing.
1874 __ set_vm_result(Oexception);
1875 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1877 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1879 __ get_vm_result(Oexception);
1880 __ verify_oop(Oexception);
1882 const int return_reg_adjustment = frame::pc_return_offset;
1883 Address issuing_pc_addr(I7, return_reg_adjustment);
1885 // We are done with this activation frame; find out where to go next.
1886 // The continuation point will be an exception handler, which expects
1887 // the following registers set up:
1888 //
1889 // Oexception: exception
1890 // Oissuing_pc: the local call that threw exception
1891 // Other On: garbage
1892 // In/Ln: the contents of the caller's register window
1893 //
1894 // We do the required restore at the last possible moment, because we
1895 // need to preserve some state across a runtime call.
1896 // (Remember that the caller activation is unknown--it might not be
1897 // interpreted, so things like Lscratch are useless in the caller.)
1899 // Although the Intel version uses call_C, we can use the more
1900 // compact call_VM. (The only real difference on SPARC is a
1901 // harmlessly ignored [re]set_last_Java_frame, compared with
1902 // the Intel code which lacks this.)
1903 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1904 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1905 __ super_call_VM_leaf(L7_thread_cache,
1906 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1907 G2_thread, Oissuing_pc->after_save());
1909 // The caller's SP was adjusted upon method entry to accomodate
1910 // the callee's non-argument locals. Undo that adjustment.
1911 __ JMP(O0, 0); // return exception handler in caller
1912 __ delayed()->restore(I5_savedSP, G0, SP);
1914 // (same old exception object is already in Oexception; see above)
1915 // Note that an "issuing PC" is actually the next PC after the call
1916 }
1919 //
1920 // JVMTI ForceEarlyReturn support
1921 //
1923 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1924 address entry = __ pc();
1926 __ empty_expression_stack();
1927 __ load_earlyret_value(state);
1929 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1930 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1932 // Clear the earlyret state
1933 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1935 __ remove_activation(state,
1936 /* throw_monitor_exception */ false,
1937 /* install_monitor_exception */ false);
1939 // The caller's SP was adjusted upon method entry to accomodate
1940 // the callee's non-argument locals. Undo that adjustment.
1941 __ ret(); // return to caller
1942 __ delayed()->restore(I5_savedSP, G0, SP);
1944 return entry;
1945 } // end of JVMTI ForceEarlyReturn support
1948 //------------------------------------------------------------------------------------------------------------------------
1949 // Helper for vtos entry point generation
1951 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) {
1952 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1953 Label L;
1954 aep = __ pc(); __ push_ptr(); __ ba_short(L);
1955 fep = __ pc(); __ push_f(); __ ba_short(L);
1956 dep = __ pc(); __ push_d(); __ ba_short(L);
1957 lep = __ pc(); __ push_l(); __ ba_short(L);
1958 iep = __ pc(); __ push_i();
1959 bep = cep = sep = iep; // there aren't any
1960 vep = __ pc(); __ bind(L); // fall through
1961 generate_and_dispatch(t);
1962 }
1964 // --------------------------------------------------------------------------------
1967 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1968 : TemplateInterpreterGenerator(code) {
1969 generate_all(); // down here so it can be "virtual"
1970 }
1972 // --------------------------------------------------------------------------------
1974 // Non-product code
1975 #ifndef PRODUCT
1976 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1977 address entry = __ pc();
1979 __ push(state);
1980 __ mov(O7, Lscratch); // protect return address within interpreter
1982 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1983 __ mov( Otos_l2, G3_scratch );
1984 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1985 __ mov(Lscratch, O7); // restore return address
1986 __ pop(state);
1987 __ retl();
1988 __ delayed()->nop();
1990 return entry;
1991 }
1994 // helpers for generate_and_dispatch
1996 void TemplateInterpreterGenerator::count_bytecode() {
1997 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1998 }
2001 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2002 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2003 }
2006 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2007 AddressLiteral index (&BytecodePairHistogram::_index);
2008 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2010 // get index, shift out old bytecode, bring in new bytecode, and store it
2011 // _index = (_index >> log2_number_of_codes) |
2012 // (bytecode << log2_number_of_codes);
2014 __ load_contents(index, G4_scratch);
2015 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2016 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
2017 __ or3( G3_scratch, G4_scratch, G4_scratch );
2018 __ store_contents(G4_scratch, index, G3_scratch);
2020 // bump bucket contents
2021 // _counters[_index] ++;
2023 __ set(counters, G3_scratch); // loads into G3_scratch
2024 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
2025 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
2026 __ ld (G3_scratch, 0, G4_scratch);
2027 __ inc (G4_scratch);
2028 __ st (G4_scratch, 0, G3_scratch);
2029 }
2032 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2033 // Call a little run-time stub to avoid blow-up for each bytecode.
2034 // The run-time runtime saves the right registers, depending on
2035 // the tosca in-state for the given template.
2036 address entry = Interpreter::trace_code(t->tos_in());
2037 guarantee(entry != NULL, "entry must have been generated");
2038 __ call(entry, relocInfo::none);
2039 __ delayed()->nop();
2040 }
2043 void TemplateInterpreterGenerator::stop_interpreter_at() {
2044 AddressLiteral counter(&BytecodeCounter::_counter_value);
2045 __ load_contents(counter, G3_scratch);
2046 AddressLiteral stop_at(&StopInterpreterAt);
2047 __ load_ptr_contents(stop_at, G4_scratch);
2048 __ cmp(G3_scratch, G4_scratch);
2049 __ breakpoint_trap(Assembler::equal, Assembler::icc);
2050 }
2051 #endif // not PRODUCT
2052 #endif // !CC_INTERP