Tue, 01 Apr 2014 09:36:49 +0200
8032410: compiler/uncommontrap/TestStackBangRbp.java times out on Solaris-Sparc V9
Summary: make compiled code bang the stack by the worst case size of the interpreter frame at deoptimization points.
Reviewed-by: twisti, kvn
1 /*
2 * Copyright (c) 1997, 2013, 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/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/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"
47 #include "utilities/macros.hpp"
49 #ifndef CC_INTERP
50 #ifndef FAST_DISPATCH
51 #define FAST_DISPATCH 1
52 #endif
53 #undef FAST_DISPATCH
56 // Generation of Interpreter
57 //
58 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
61 #define __ _masm->
64 //----------------------------------------------------------------------------------------------------
67 void InterpreterGenerator::save_native_result(void) {
68 // result potentially in O0/O1: save it across calls
69 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
71 // result potentially in F0/F1: save it across calls
72 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
74 // save and restore any potential method result value around the unlocking operation
75 __ stf(FloatRegisterImpl::D, F0, d_tmp);
76 #ifdef _LP64
77 __ stx(O0, l_tmp);
78 #else
79 __ std(O0, l_tmp);
80 #endif
81 }
83 void InterpreterGenerator::restore_native_result(void) {
84 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
85 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
87 // Restore any method result value
88 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
89 #ifdef _LP64
90 __ ldx(l_tmp, O0);
91 #else
92 __ ldd(l_tmp, O0);
93 #endif
94 }
96 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
97 assert(!pass_oop || message == NULL, "either oop or message but not both");
98 address entry = __ pc();
99 // expression stack must be empty before entering the VM if an exception happened
100 __ empty_expression_stack();
101 // load exception object
102 __ set((intptr_t)name, G3_scratch);
103 if (pass_oop) {
104 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
105 } else {
106 __ set((intptr_t)message, G4_scratch);
107 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
108 }
109 // throw exception
110 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
111 AddressLiteral thrower(Interpreter::throw_exception_entry());
112 __ jump_to(thrower, G3_scratch);
113 __ delayed()->nop();
114 return entry;
115 }
117 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
118 address entry = __ pc();
119 // expression stack must be empty before entering the VM if an exception
120 // happened
121 __ empty_expression_stack();
122 // load exception object
123 __ call_VM(Oexception,
124 CAST_FROM_FN_PTR(address,
125 InterpreterRuntime::throw_ClassCastException),
126 Otos_i);
127 __ should_not_reach_here();
128 return entry;
129 }
132 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
133 address entry = __ pc();
134 // expression stack must be empty before entering the VM if an exception happened
135 __ empty_expression_stack();
136 // convention: expect aberrant index in register G3_scratch, then shuffle the
137 // index to G4_scratch for the VM call
138 __ mov(G3_scratch, G4_scratch);
139 __ set((intptr_t)name, G3_scratch);
140 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
141 __ should_not_reach_here();
142 return entry;
143 }
146 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
147 address entry = __ pc();
148 // expression stack must be empty before entering the VM if an exception happened
149 __ empty_expression_stack();
150 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
151 __ should_not_reach_here();
152 return entry;
153 }
156 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
157 address entry = __ pc();
159 if (state == atos) {
160 __ profile_return_type(O0, G3_scratch, G1_scratch);
161 }
163 #if !defined(_LP64) && defined(COMPILER2)
164 // All return values are where we want them, except for Longs. C2 returns
165 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
166 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
167 // build even if we are returning from interpreted we just do a little
168 // stupid shuffing.
169 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
170 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
171 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
173 if (state == ltos) {
174 __ srl (G1, 0, O1);
175 __ srlx(G1, 32, O0);
176 }
177 #endif // !_LP64 && COMPILER2
179 // The callee returns with the stack possibly adjusted by adapter transition
180 // We remove that possible adjustment here.
181 // All interpreter local registers are untouched. Any result is passed back
182 // in the O0/O1 or float registers. Before continuing, the arguments must be
183 // popped from the java expression stack; i.e., Lesp must be adjusted.
185 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
187 const Register cache = G3_scratch;
188 const Register index = G1_scratch;
189 __ get_cache_and_index_at_bcp(cache, index, 1, index_size);
191 const Register flags = cache;
192 __ ld_ptr(cache, ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset(), flags);
193 const Register parameter_size = flags;
194 __ and3(flags, ConstantPoolCacheEntry::parameter_size_mask, parameter_size); // argument size in words
195 __ sll(parameter_size, Interpreter::logStackElementSize, parameter_size); // each argument size in bytes
196 __ add(Lesp, parameter_size, Lesp); // pop arguments
197 __ dispatch_next(state, step);
199 return entry;
200 }
203 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
204 address entry = __ pc();
205 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
206 { Label L;
207 Address exception_addr(G2_thread, Thread::pending_exception_offset());
208 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
209 __ br_null_short(Gtemp, Assembler::pt, L);
210 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
211 __ should_not_reach_here();
212 __ bind(L);
213 }
214 __ dispatch_next(state, step);
215 return entry;
216 }
218 // A result handler converts/unboxes a native call result into
219 // a java interpreter/compiler result. The current frame is an
220 // interpreter frame. The activation frame unwind code must be
221 // consistent with that of TemplateTable::_return(...). In the
222 // case of native methods, the caller's SP was not modified.
223 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
224 address entry = __ pc();
225 Register Itos_i = Otos_i ->after_save();
226 Register Itos_l = Otos_l ->after_save();
227 Register Itos_l1 = Otos_l1->after_save();
228 Register Itos_l2 = Otos_l2->after_save();
229 switch (type) {
230 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
231 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
232 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
233 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
234 case T_LONG :
235 #ifndef _LP64
236 __ mov(O1, Itos_l2); // move other half of long
237 #endif // ifdef or no ifdef, fall through to the T_INT case
238 case T_INT : __ mov(O0, Itos_i); break;
239 case T_VOID : /* nothing to do */ break;
240 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
241 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
242 case T_OBJECT :
243 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
244 __ verify_oop(Itos_i);
245 break;
246 default : ShouldNotReachHere();
247 }
248 __ ret(); // return from interpreter activation
249 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
250 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
251 return entry;
252 }
254 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
255 address entry = __ pc();
256 __ push(state);
257 __ call_VM(noreg, runtime_entry);
258 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
259 return entry;
260 }
263 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
264 address entry = __ pc();
265 __ dispatch_next(state);
266 return entry;
267 }
269 //
270 // Helpers for commoning out cases in the various type of method entries.
271 //
273 // increment invocation count & check for overflow
274 //
275 // Note: checking for negative value instead of overflow
276 // so we have a 'sticky' overflow test
277 //
278 // Lmethod: method
279 // ??: invocation counter
280 //
281 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
282 // Note: In tiered we increment either counters in MethodCounters* or in
283 // MDO depending if we're profiling or not.
284 const Register Rcounters = G3_scratch;
285 Label done;
287 if (TieredCompilation) {
288 const int increment = InvocationCounter::count_increment;
289 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
290 Label no_mdo;
291 if (ProfileInterpreter) {
292 // If no method data exists, go to profile_continue.
293 __ ld_ptr(Lmethod, Method::method_data_offset(), G4_scratch);
294 __ br_null_short(G4_scratch, Assembler::pn, no_mdo);
295 // Increment counter
296 Address mdo_invocation_counter(G4_scratch,
297 in_bytes(MethodData::invocation_counter_offset()) +
298 in_bytes(InvocationCounter::counter_offset()));
299 __ increment_mask_and_jump(mdo_invocation_counter, increment, mask,
300 G3_scratch, Lscratch,
301 Assembler::zero, overflow);
302 __ ba_short(done);
303 }
305 // Increment counter in MethodCounters*
306 __ bind(no_mdo);
307 Address invocation_counter(Rcounters,
308 in_bytes(MethodCounters::invocation_counter_offset()) +
309 in_bytes(InvocationCounter::counter_offset()));
310 __ get_method_counters(Lmethod, Rcounters, done);
311 __ increment_mask_and_jump(invocation_counter, increment, mask,
312 G4_scratch, Lscratch,
313 Assembler::zero, overflow);
314 __ bind(done);
315 } else {
316 // Update standard invocation counters
317 __ get_method_counters(Lmethod, Rcounters, done);
318 __ increment_invocation_counter(Rcounters, O0, G4_scratch);
319 if (ProfileInterpreter) {
320 Address interpreter_invocation_counter(Rcounters,
321 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
322 __ ld(interpreter_invocation_counter, G4_scratch);
323 __ inc(G4_scratch);
324 __ st(G4_scratch, interpreter_invocation_counter);
325 }
327 if (ProfileInterpreter && profile_method != NULL) {
328 // Test to see if we should create a method data oop
329 AddressLiteral profile_limit((address)&InvocationCounter::InterpreterProfileLimit);
330 __ load_contents(profile_limit, G3_scratch);
331 __ cmp_and_br_short(O0, G3_scratch, Assembler::lessUnsigned, Assembler::pn, *profile_method_continue);
333 // if no method data exists, go to profile_method
334 __ test_method_data_pointer(*profile_method);
335 }
337 AddressLiteral invocation_limit((address)&InvocationCounter::InterpreterInvocationLimit);
338 __ load_contents(invocation_limit, G3_scratch);
339 __ cmp(O0, G3_scratch);
340 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow); // Far distance
341 __ delayed()->nop();
342 __ bind(done);
343 }
345 }
347 // Allocate monitor and lock method (asm interpreter)
348 // ebx - Method*
349 //
350 void InterpreterGenerator::lock_method(void) {
351 __ ld(Lmethod, in_bytes(Method::access_flags_offset()), O0); // Load access flags.
353 #ifdef ASSERT
354 { Label ok;
355 __ btst(JVM_ACC_SYNCHRONIZED, O0);
356 __ br( Assembler::notZero, false, Assembler::pt, ok);
357 __ delayed()->nop();
358 __ stop("method doesn't need synchronization");
359 __ bind(ok);
360 }
361 #endif // ASSERT
363 // get synchronization object to O0
364 { Label done;
365 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
366 __ btst(JVM_ACC_STATIC, O0);
367 __ br( Assembler::zero, true, Assembler::pt, done);
368 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
370 __ ld_ptr( Lmethod, in_bytes(Method::const_offset()), O0);
371 __ ld_ptr( O0, in_bytes(ConstMethod::constants_offset()), O0);
372 __ ld_ptr( O0, ConstantPool::pool_holder_offset_in_bytes(), O0);
374 // lock the mirror, not the Klass*
375 __ ld_ptr( O0, mirror_offset, O0);
377 #ifdef ASSERT
378 __ tst(O0);
379 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
380 #endif // ASSERT
382 __ bind(done);
383 }
385 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
386 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
387 // __ untested("lock_object from method entry");
388 __ lock_object(Lmonitors, O0);
389 }
392 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
393 Register Rscratch,
394 Register Rscratch2) {
395 const int page_size = os::vm_page_size();
396 Label after_frame_check;
398 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
400 __ set(page_size, Rscratch);
401 __ cmp_and_br_short(Rframe_size, Rscratch, Assembler::lessEqual, Assembler::pt, after_frame_check);
403 // get the stack base, and in debug, verify it is non-zero
404 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
405 #ifdef ASSERT
406 Label base_not_zero;
407 __ br_notnull_short(Rscratch, Assembler::pn, base_not_zero);
408 __ stop("stack base is zero in generate_stack_overflow_check");
409 __ bind(base_not_zero);
410 #endif
412 // get the stack size, and in debug, verify it is non-zero
413 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
414 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
415 #ifdef ASSERT
416 Label size_not_zero;
417 __ br_notnull_short(Rscratch2, Assembler::pn, size_not_zero);
418 __ stop("stack size is zero in generate_stack_overflow_check");
419 __ bind(size_not_zero);
420 #endif
422 // compute the beginning of the protected zone minus the requested frame size
423 __ sub( Rscratch, Rscratch2, Rscratch );
424 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
425 __ add( Rscratch, Rscratch2, Rscratch );
427 // Add in the size of the frame (which is the same as subtracting it from the
428 // SP, which would take another register
429 __ add( Rscratch, Rframe_size, Rscratch );
431 // the frame is greater than one page in size, so check against
432 // the bottom of the stack
433 __ cmp_and_brx_short(SP, Rscratch, Assembler::greaterUnsigned, Assembler::pt, after_frame_check);
435 // the stack will overflow, throw an exception
437 // Note that SP is restored to sender's sp (in the delay slot). This
438 // is necessary if the sender's frame is an extended compiled frame
439 // (see gen_c2i_adapter()) and safer anyway in case of JSR292
440 // adaptations.
442 // Note also that the restored frame is not necessarily interpreted.
443 // Use the shared runtime version of the StackOverflowError.
444 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "stub not yet generated");
445 AddressLiteral stub(StubRoutines::throw_StackOverflowError_entry());
446 __ jump_to(stub, Rscratch);
447 __ delayed()->mov(O5_savedSP, SP);
449 // if you get to here, then there is enough stack space
450 __ bind( after_frame_check );
451 }
454 //
455 // Generate a fixed interpreter frame. This is identical setup for interpreted
456 // methods and for native methods hence the shared code.
458 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
459 //
460 //
461 // The entry code sets up a new interpreter frame in 4 steps:
462 //
463 // 1) Increase caller's SP by for the extra local space needed:
464 // (check for overflow)
465 // Efficient implementation of xload/xstore bytecodes requires
466 // that arguments and non-argument locals are in a contigously
467 // addressable memory block => non-argument locals must be
468 // allocated in the caller's frame.
469 //
470 // 2) Create a new stack frame and register window:
471 // The new stack frame must provide space for the standard
472 // register save area, the maximum java expression stack size,
473 // the monitor slots (0 slots initially), and some frame local
474 // scratch locations.
475 //
476 // 3) The following interpreter activation registers must be setup:
477 // Lesp : expression stack pointer
478 // Lbcp : bytecode pointer
479 // Lmethod : method
480 // Llocals : locals pointer
481 // Lmonitors : monitor pointer
482 // LcpoolCache: constant pool cache
483 //
484 // 4) Initialize the non-argument locals if necessary:
485 // Non-argument locals may need to be initialized to NULL
486 // for GC to work. If the oop-map information is accurate
487 // (in the absence of the JSR problem), no initialization
488 // is necessary.
489 //
490 // (gri - 2/25/2000)
493 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
495 const int extra_space =
496 rounded_vm_local_words + // frame local scratch space
497 Method::extra_stack_entries() + // extra stack for jsr 292
498 frame::memory_parameter_word_sp_offset + // register save area
499 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
501 const Register Glocals_size = G3;
502 const Register RconstMethod = Glocals_size;
503 const Register Otmp1 = O3;
504 const Register Otmp2 = O4;
505 // Lscratch can't be used as a temporary because the call_stub uses
506 // it to assert that the stack frame was setup correctly.
507 const Address constMethod (G5_method, Method::const_offset());
508 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
510 __ ld_ptr( constMethod, RconstMethod );
511 __ lduh( size_of_parameters, Glocals_size);
513 // Gargs points to first local + BytesPerWord
514 // Set the saved SP after the register window save
515 //
516 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
517 __ sll(Glocals_size, Interpreter::logStackElementSize, Otmp1);
518 __ add(Gargs, Otmp1, Gargs);
520 if (native_call) {
521 __ calc_mem_param_words( Glocals_size, Gframe_size );
522 __ add( Gframe_size, extra_space, Gframe_size);
523 __ round_to( Gframe_size, WordsPerLong );
524 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
525 } else {
527 //
528 // Compute number of locals in method apart from incoming parameters
529 //
530 const Address size_of_locals (Otmp1, ConstMethod::size_of_locals_offset());
531 __ ld_ptr( constMethod, Otmp1 );
532 __ lduh( size_of_locals, Otmp1 );
533 __ sub( Otmp1, Glocals_size, Glocals_size );
534 __ round_to( Glocals_size, WordsPerLong );
535 __ sll( Glocals_size, Interpreter::logStackElementSize, Glocals_size );
537 // see if the frame is greater than one page in size. If so,
538 // then we need to verify there is enough stack space remaining
539 // Frame_size = (max_stack + extra_space) * BytesPerWord;
540 __ ld_ptr( constMethod, Gframe_size );
541 __ lduh( Gframe_size, in_bytes(ConstMethod::max_stack_offset()), 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 #if INCLUDE_ALL_GCS
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 // INCLUDE_ALL_GCS
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 __ flushw();
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(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 constMethod (G5_method, Method::const_offset());
1259 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1260 // and use in the asserts.
1261 const Address access_flags (Lmethod, Method::access_flags_offset());
1263 const Register Glocals_size = G3;
1264 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1266 // make sure method is not native & not abstract
1267 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1268 #ifdef ASSERT
1269 __ ld(G5_method, Method::access_flags_offset(), Gtmp1);
1270 {
1271 Label L;
1272 __ btst(JVM_ACC_NATIVE, Gtmp1);
1273 __ br(Assembler::zero, false, Assembler::pt, L);
1274 __ delayed()->nop();
1275 __ stop("tried to execute native method as non-native");
1276 __ bind(L);
1277 }
1278 { Label L;
1279 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1280 __ br(Assembler::zero, false, Assembler::pt, L);
1281 __ delayed()->nop();
1282 __ stop("tried to execute abstract method as non-abstract");
1283 __ bind(L);
1284 }
1285 #endif // ASSERT
1287 // generate the code to allocate the interpreter stack frame
1289 generate_fixed_frame(false);
1291 #ifdef FAST_DISPATCH
1292 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1293 // set bytecode dispatch table base
1294 #endif
1296 //
1297 // Code to initialize the extra (i.e. non-parm) locals
1298 //
1299 Register init_value = noreg; // will be G0 if we must clear locals
1300 // The way the code was setup before zerolocals was always true for vanilla java entries.
1301 // It could only be false for the specialized entries like accessor or empty which have
1302 // no extra locals so the testing was a waste of time and the extra locals were always
1303 // initialized. We removed this extra complication to already over complicated code.
1305 init_value = G0;
1306 Label clear_loop;
1308 const Register RconstMethod = O1;
1309 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1310 const Address size_of_locals (RconstMethod, ConstMethod::size_of_locals_offset());
1312 // NOTE: If you change the frame layout, this code will need to
1313 // be updated!
1314 __ ld_ptr( constMethod, RconstMethod );
1315 __ lduh( size_of_locals, O2 );
1316 __ lduh( size_of_parameters, O1 );
1317 __ sll( O2, Interpreter::logStackElementSize, O2);
1318 __ sll( O1, Interpreter::logStackElementSize, O1 );
1319 __ sub( Llocals, O2, O2 );
1320 __ sub( Llocals, O1, O1 );
1322 __ bind( clear_loop );
1323 __ inc( O2, wordSize );
1325 __ cmp( O2, O1 );
1326 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1327 __ delayed()->st_ptr( init_value, O2, 0 );
1329 const Address do_not_unlock_if_synchronized(G2_thread,
1330 JavaThread::do_not_unlock_if_synchronized_offset());
1331 // Since at this point in the method invocation the exception handler
1332 // would try to exit the monitor of synchronized methods which hasn't
1333 // been entered yet, we set the thread local variable
1334 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1335 // runtime, exception handling i.e. unlock_if_synchronized_method will
1336 // check this thread local flag.
1337 __ movbool(true, G3_scratch);
1338 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1340 __ profile_parameters_type(G1_scratch, G3_scratch, G4_scratch, Lscratch);
1341 // increment invocation counter and check for overflow
1342 //
1343 // Note: checking for negative value instead of overflow
1344 // so we have a 'sticky' overflow test (may be of
1345 // importance as soon as we have true MT/MP)
1346 Label invocation_counter_overflow;
1347 Label profile_method;
1348 Label profile_method_continue;
1349 Label Lcontinue;
1350 if (inc_counter) {
1351 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1352 if (ProfileInterpreter) {
1353 __ bind(profile_method_continue);
1354 }
1355 }
1356 __ bind(Lcontinue);
1358 bang_stack_shadow_pages(false);
1360 // reset the _do_not_unlock_if_synchronized flag
1361 __ stbool(G0, do_not_unlock_if_synchronized);
1363 // check for synchronized methods
1364 // Must happen AFTER invocation_counter check and stack overflow check,
1365 // so method is not locked if overflows.
1367 if (synchronized) {
1368 lock_method();
1369 } else {
1370 #ifdef ASSERT
1371 { Label ok;
1372 __ ld(access_flags, O0);
1373 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1374 __ br( Assembler::zero, false, Assembler::pt, ok);
1375 __ delayed()->nop();
1376 __ stop("method needs synchronization");
1377 __ bind(ok);
1378 }
1379 #endif // ASSERT
1380 }
1382 // start execution
1384 __ verify_thread();
1386 // jvmti support
1387 __ notify_method_entry();
1389 // start executing instructions
1390 __ dispatch_next(vtos);
1393 if (inc_counter) {
1394 if (ProfileInterpreter) {
1395 // We have decided to profile this method in the interpreter
1396 __ bind(profile_method);
1398 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1399 __ set_method_data_pointer_for_bcp();
1400 __ ba_short(profile_method_continue);
1401 }
1403 // handle invocation counter overflow
1404 __ bind(invocation_counter_overflow);
1405 generate_counter_overflow(Lcontinue);
1406 }
1409 return entry;
1410 }
1413 //----------------------------------------------------------------------------------------------------
1414 // Entry points & stack frame layout
1415 //
1416 // Here we generate the various kind of entries into the interpreter.
1417 // The two main entry type are generic bytecode methods and native call method.
1418 // These both come in synchronized and non-synchronized versions but the
1419 // frame layout they create is very similar. The other method entry
1420 // types are really just special purpose entries that are really entry
1421 // and interpretation all in one. These are for trivial methods like
1422 // accessor, empty, or special math methods.
1423 //
1424 // When control flow reaches any of the entry types for the interpreter
1425 // the following holds ->
1426 //
1427 // C2 Calling Conventions:
1428 //
1429 // The entry code below assumes that the following registers are set
1430 // when coming in:
1431 // G5_method: holds the Method* of the method to call
1432 // Lesp: points to the TOS of the callers expression stack
1433 // after having pushed all the parameters
1434 //
1435 // The entry code does the following to setup an interpreter frame
1436 // pop parameters from the callers stack by adjusting Lesp
1437 // set O0 to Lesp
1438 // compute X = (max_locals - num_parameters)
1439 // bump SP up by X to accomadate the extra locals
1440 // compute X = max_expression_stack
1441 // + vm_local_words
1442 // + 16 words of register save area
1443 // save frame doing a save sp, -X, sp growing towards lower addresses
1444 // set Lbcp, Lmethod, LcpoolCache
1445 // set Llocals to i0
1446 // set Lmonitors to FP - rounded_vm_local_words
1447 // set Lesp to Lmonitors - 4
1448 //
1449 // The frame has now been setup to do the rest of the entry code
1451 // Try this optimization: Most method entries could live in a
1452 // "one size fits all" stack frame without all the dynamic size
1453 // calculations. It might be profitable to do all this calculation
1454 // statically and approximately for "small enough" methods.
1456 //-----------------------------------------------------------------------------------------------
1458 // C1 Calling conventions
1459 //
1460 // Upon method entry, the following registers are setup:
1461 //
1462 // g2 G2_thread: current thread
1463 // g5 G5_method: method to activate
1464 // g4 Gargs : pointer to last argument
1465 //
1466 //
1467 // Stack:
1468 //
1469 // +---------------+ <--- sp
1470 // | |
1471 // : reg save area :
1472 // | |
1473 // +---------------+ <--- sp + 0x40
1474 // | |
1475 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1476 // | |
1477 // +---------------+ <--- sp + 0x5c
1478 // | |
1479 // : free :
1480 // | |
1481 // +---------------+ <--- Gargs
1482 // | |
1483 // : arguments :
1484 // | |
1485 // +---------------+
1486 // | |
1487 //
1488 //
1489 //
1490 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1491 //
1492 // +---------------+ <--- sp
1493 // | |
1494 // : reg save area :
1495 // | |
1496 // +---------------+ <--- sp + 0x40
1497 // | |
1498 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1499 // | |
1500 // +---------------+ <--- sp + 0x5c
1501 // | |
1502 // : :
1503 // | | <--- Lesp
1504 // +---------------+ <--- Lmonitors (fp - 0x18)
1505 // | VM locals |
1506 // +---------------+ <--- fp
1507 // | |
1508 // : reg save area :
1509 // | |
1510 // +---------------+ <--- fp + 0x40
1511 // | |
1512 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1513 // | |
1514 // +---------------+ <--- fp + 0x5c
1515 // | |
1516 // : free :
1517 // | |
1518 // +---------------+
1519 // | |
1520 // : nonarg locals :
1521 // | |
1522 // +---------------+
1523 // | |
1524 // : arguments :
1525 // | | <--- Llocals
1526 // +---------------+ <--- Gargs
1527 // | |
1529 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1531 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1532 // expression stack, the callee will have callee_extra_locals (so we can account for
1533 // frame extension) and monitor_size for monitors. Basically we need to calculate
1534 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1535 //
1536 //
1537 // The big complicating thing here is that we must ensure that the stack stays properly
1538 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1539 // needs to be aligned for). We are given that the sp (fp) is already aligned by
1540 // the caller so we must ensure that it is properly aligned for our callee.
1541 //
1542 const int rounded_vm_local_words =
1543 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1544 // callee_locals and max_stack are counts, not the size in frame.
1545 const int locals_size =
1546 round_to(callee_extra_locals * Interpreter::stackElementWords, WordsPerLong);
1547 const int max_stack_words = max_stack * Interpreter::stackElementWords;
1548 return (round_to((max_stack_words
1549 + rounded_vm_local_words
1550 + frame::memory_parameter_word_sp_offset), WordsPerLong)
1551 // already rounded
1552 + locals_size + monitor_size);
1553 }
1555 // How much stack a method top interpreter activation needs in words.
1556 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1558 // See call_stub code
1559 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
1560 WordsPerLong); // 7 + register save area
1562 // Save space for one monitor to get into the interpreted method in case
1563 // the method is synchronized
1564 int monitor_size = method->is_synchronized() ?
1565 1*frame::interpreter_frame_monitor_size() : 0;
1566 return size_activation_helper(method->max_locals(), method->max_stack(),
1567 monitor_size) + call_stub_size;
1568 }
1570 int AbstractInterpreter::size_activation(int max_stack,
1571 int temps,
1572 int extra_args,
1573 int monitors,
1574 int callee_params,
1575 int callee_locals,
1576 bool is_top_frame) {
1577 // Note: This calculation must exactly parallel the frame setup
1578 // in InterpreterGenerator::generate_fixed_frame.
1580 int monitor_size = monitors * frame::interpreter_frame_monitor_size();
1582 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1584 //
1585 // Note: if you look closely this appears to be doing something much different
1586 // than generate_fixed_frame. What is happening is this. On sparc we have to do
1587 // this dance with interpreter_sp_adjustment because the window save area would
1588 // appear just below the bottom (tos) of the caller's java expression stack. Because
1589 // the interpreter want to have the locals completely contiguous generate_fixed_frame
1590 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1591 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1592 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1593 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1594 // because the oldest frame would have adjust its callers frame and yet that frame
1595 // already exists and isn't part of this array of frames we are unpacking. So at first
1596 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1597 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1598 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1599 // add up. It does seem like it simpler to account for the adjustment here (and remove the
1600 // callee... parameters here). However this would mean that this routine would have to take
1601 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1602 // and run the calling loop in the reverse order. This would also would appear to mean making
1603 // this code aware of what the interactions are when that initial caller fram was an osr or
1604 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
1605 // there is no sense in messing working code.
1606 //
1608 int rounded_cls = round_to((callee_locals - callee_params), WordsPerLong);
1609 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1611 int raw_frame_size = size_activation_helper(rounded_cls, max_stack, monitor_size);
1613 return raw_frame_size;
1614 }
1616 void AbstractInterpreter::layout_activation(Method* method,
1617 int tempcount,
1618 int popframe_extra_args,
1619 int moncount,
1620 int caller_actual_parameters,
1621 int callee_param_count,
1622 int callee_local_count,
1623 frame* caller,
1624 frame* interpreter_frame,
1625 bool is_top_frame,
1626 bool is_bottom_frame) {
1627 // Set up the following variables:
1628 // - Lmethod
1629 // - Llocals
1630 // - Lmonitors (to the indicated number of monitors)
1631 // - Lesp (to the indicated number of temps)
1632 // The frame caller on entry is a description of the caller of the
1633 // frame we are about to layout. We are guaranteed that we will be
1634 // able to fill in a new interpreter frame as its callee (i.e. the
1635 // stack space is allocated and the amount was determined by an
1636 // earlier call to the size_activation() method). On return caller
1637 // while describe the interpreter frame we just layed out.
1639 // The skeleton frame must already look like an interpreter frame
1640 // even if not fully filled out.
1641 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1643 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1644 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
1645 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1647 intptr_t* fp = interpreter_frame->fp();
1649 JavaThread* thread = JavaThread::current();
1650 RegisterMap map(thread, false);
1651 // More verification that skeleton frame is properly walkable
1652 assert(fp == caller->sp(), "fp must match");
1654 intptr_t* montop = fp - rounded_vm_local_words;
1656 // preallocate monitors (cf. __ add_monitor_to_stack)
1657 intptr_t* monitors = montop - monitor_size;
1659 // preallocate stack space
1660 intptr_t* esp = monitors - 1 -
1661 (tempcount * Interpreter::stackElementWords) -
1662 popframe_extra_args;
1664 int local_words = method->max_locals() * Interpreter::stackElementWords;
1665 NEEDS_CLEANUP;
1666 intptr_t* locals;
1667 if (caller->is_interpreted_frame()) {
1668 // Can force the locals area to end up properly overlapping the top of the expression stack.
1669 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1670 // Note that this computation means we replace size_of_parameters() values from the caller
1671 // interpreter frame's expression stack with our argument locals
1672 int parm_words = caller_actual_parameters * Interpreter::stackElementWords;
1673 locals = Lesp_ptr + parm_words;
1674 int delta = local_words - parm_words;
1675 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1676 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1677 if (!is_bottom_frame) {
1678 // Llast_SP is set below for the current frame to SP (with the
1679 // extra space for the callee's locals). Here we adjust
1680 // Llast_SP for the caller's frame, removing the extra space
1681 // for the current method's locals.
1682 *caller->register_addr(Llast_SP) = *interpreter_frame->register_addr(I5_savedSP);
1683 } else {
1684 assert(*caller->register_addr(Llast_SP) >= *interpreter_frame->register_addr(I5_savedSP), "strange Llast_SP");
1685 }
1686 } else {
1687 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1688 // Don't have Lesp available; lay out locals block in the caller
1689 // adjacent to the register window save area.
1690 //
1691 // Compiled frames do not allocate a varargs area which is why this if
1692 // statement is needed.
1693 //
1694 if (caller->is_compiled_frame()) {
1695 locals = fp + frame::register_save_words + local_words - 1;
1696 } else {
1697 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1698 }
1699 if (!caller->is_entry_frame()) {
1700 // Caller wants his own SP back
1701 int caller_frame_size = caller->cb()->frame_size();
1702 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1703 }
1704 }
1705 if (TraceDeoptimization) {
1706 if (caller->is_entry_frame()) {
1707 // make sure I5_savedSP and the entry frames notion of saved SP
1708 // agree. This assertion duplicate a check in entry frame code
1709 // but catches the failure earlier.
1710 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1711 "would change callers SP");
1712 }
1713 if (caller->is_entry_frame()) {
1714 tty->print("entry ");
1715 }
1716 if (caller->is_compiled_frame()) {
1717 tty->print("compiled ");
1718 if (caller->is_deoptimized_frame()) {
1719 tty->print("(deopt) ");
1720 }
1721 }
1722 if (caller->is_interpreted_frame()) {
1723 tty->print("interpreted ");
1724 }
1725 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1726 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1727 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1728 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1729 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1730 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1731 tty->print_cr("Llocals = 0x%x", locals);
1732 tty->print_cr("Lesp = 0x%x", esp);
1733 tty->print_cr("Lmonitors = 0x%x", monitors);
1734 }
1736 if (method->max_locals() > 0) {
1737 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1738 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1739 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1740 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1741 }
1742 #ifdef _LP64
1743 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1744 #endif
1746 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
1747 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
1748 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
1749 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
1750 // Llast_SP will be same as SP as there is no adapter space
1751 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1752 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1753 #ifdef FAST_DISPATCH
1754 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1755 #endif
1758 #ifdef ASSERT
1759 BasicObjectLock* mp = (BasicObjectLock*)monitors;
1761 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1762 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize)), "locals match");
1763 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
1764 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1765 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1767 // check bounds
1768 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1769 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1770 assert(lo < monitors && montop <= hi, "monitors in bounds");
1771 assert(lo <= esp && esp < monitors, "esp in bounds");
1772 #endif // ASSERT
1773 }
1775 //----------------------------------------------------------------------------------------------------
1776 // Exceptions
1777 void TemplateInterpreterGenerator::generate_throw_exception() {
1779 // Entry point in previous activation (i.e., if the caller was interpreted)
1780 Interpreter::_rethrow_exception_entry = __ pc();
1781 // O0: exception
1783 // entry point for exceptions thrown within interpreter code
1784 Interpreter::_throw_exception_entry = __ pc();
1785 __ verify_thread();
1786 // expression stack is undefined here
1787 // O0: exception, i.e. Oexception
1788 // Lbcp: exception bcx
1789 __ verify_oop(Oexception);
1792 // expression stack must be empty before entering the VM in case of an exception
1793 __ empty_expression_stack();
1794 // find exception handler address and preserve exception oop
1795 // call C routine to find handler and jump to it
1796 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1797 __ push_ptr(O1); // push exception for exception handler bytecodes
1799 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1800 __ delayed()->nop();
1803 // if the exception is not handled in the current frame
1804 // the frame is removed and the exception is rethrown
1805 // (i.e. exception continuation is _rethrow_exception)
1806 //
1807 // Note: At this point the bci is still the bxi for the instruction which caused
1808 // the exception and the expression stack is empty. Thus, for any VM calls
1809 // at this point, GC will find a legal oop map (with empty expression stack).
1811 // in current activation
1812 // tos: exception
1813 // Lbcp: exception bcp
1815 //
1816 // JVMTI PopFrame support
1817 //
1819 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1820 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1821 // Set the popframe_processing bit in popframe_condition indicating that we are
1822 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1823 // popframe handling cycles.
1825 __ ld(popframe_condition_addr, G3_scratch);
1826 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1827 __ stw(G3_scratch, popframe_condition_addr);
1829 // Empty the expression stack, as in normal exception handling
1830 __ empty_expression_stack();
1831 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1833 {
1834 // Check to see whether we are returning to a deoptimized frame.
1835 // (The PopFrame call ensures that the caller of the popped frame is
1836 // either interpreted or compiled and deoptimizes it if compiled.)
1837 // In this case, we can't call dispatch_next() after the frame is
1838 // popped, but instead must save the incoming arguments and restore
1839 // them after deoptimization has occurred.
1840 //
1841 // Note that we don't compare the return PC against the
1842 // deoptimization blob's unpack entry because of the presence of
1843 // adapter frames in C2.
1844 Label caller_not_deoptimized;
1845 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1846 __ br_notnull_short(O0, Assembler::pt, caller_not_deoptimized);
1848 const Register Gtmp1 = G3_scratch;
1849 const Register Gtmp2 = G1_scratch;
1850 const Register RconstMethod = Gtmp1;
1851 const Address constMethod(Lmethod, Method::const_offset());
1852 const Address size_of_parameters(RconstMethod, ConstMethod::size_of_parameters_offset());
1854 // Compute size of arguments for saving when returning to deoptimized caller
1855 __ ld_ptr(constMethod, RconstMethod);
1856 __ lduh(size_of_parameters, Gtmp1);
1857 __ sll(Gtmp1, Interpreter::logStackElementSize, Gtmp1);
1858 __ sub(Llocals, Gtmp1, Gtmp2);
1859 __ add(Gtmp2, wordSize, Gtmp2);
1860 // Save these arguments
1861 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1862 // Inform deoptimization that it is responsible for restoring these arguments
1863 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1864 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1865 __ st(Gtmp1, popframe_condition_addr);
1867 // Return from the current method
1868 // The caller's SP was adjusted upon method entry to accomodate
1869 // the callee's non-argument locals. Undo that adjustment.
1870 __ ret();
1871 __ delayed()->restore(I5_savedSP, G0, SP);
1873 __ bind(caller_not_deoptimized);
1874 }
1876 // Clear the popframe condition flag
1877 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1879 // Get out of the current method (how this is done depends on the particular compiler calling
1880 // convention that the interpreter currently follows)
1881 // The caller's SP was adjusted upon method entry to accomodate
1882 // the callee's non-argument locals. Undo that adjustment.
1883 __ restore(I5_savedSP, G0, SP);
1884 // The method data pointer was incremented already during
1885 // call profiling. We have to restore the mdp for the current bcp.
1886 if (ProfileInterpreter) {
1887 __ set_method_data_pointer_for_bcp();
1888 }
1890 #if INCLUDE_JVMTI
1891 if (EnableInvokeDynamic) {
1892 Label L_done;
1894 __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode
1895 __ cmp_and_br_short(G1_scratch, Bytecodes::_invokestatic, Assembler::notEqual, Assembler::pn, L_done);
1897 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
1898 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
1900 __ call_VM(G1_scratch, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), I0, Lmethod, Lbcp);
1902 __ br_null(G1_scratch, false, Assembler::pn, L_done);
1903 __ delayed()->nop();
1905 __ st_ptr(G1_scratch, Lesp, wordSize);
1906 __ bind(L_done);
1907 }
1908 #endif // INCLUDE_JVMTI
1910 // Resume bytecode interpretation at the current bcp
1911 __ dispatch_next(vtos);
1912 // end of JVMTI PopFrame support
1914 Interpreter::_remove_activation_entry = __ pc();
1916 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1917 __ pop_ptr(Oexception); // get exception
1919 // Intel has the following comment:
1920 //// remove the activation (without doing throws on illegalMonitorExceptions)
1921 // They remove the activation without checking for bad monitor state.
1922 // %%% We should make sure this is the right semantics before implementing.
1924 __ set_vm_result(Oexception);
1925 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1927 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1929 __ get_vm_result(Oexception);
1930 __ verify_oop(Oexception);
1932 const int return_reg_adjustment = frame::pc_return_offset;
1933 Address issuing_pc_addr(I7, return_reg_adjustment);
1935 // We are done with this activation frame; find out where to go next.
1936 // The continuation point will be an exception handler, which expects
1937 // the following registers set up:
1938 //
1939 // Oexception: exception
1940 // Oissuing_pc: the local call that threw exception
1941 // Other On: garbage
1942 // In/Ln: the contents of the caller's register window
1943 //
1944 // We do the required restore at the last possible moment, because we
1945 // need to preserve some state across a runtime call.
1946 // (Remember that the caller activation is unknown--it might not be
1947 // interpreted, so things like Lscratch are useless in the caller.)
1949 // Although the Intel version uses call_C, we can use the more
1950 // compact call_VM. (The only real difference on SPARC is a
1951 // harmlessly ignored [re]set_last_Java_frame, compared with
1952 // the Intel code which lacks this.)
1953 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1954 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1955 __ super_call_VM_leaf(L7_thread_cache,
1956 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1957 G2_thread, Oissuing_pc->after_save());
1959 // The caller's SP was adjusted upon method entry to accomodate
1960 // the callee's non-argument locals. Undo that adjustment.
1961 __ JMP(O0, 0); // return exception handler in caller
1962 __ delayed()->restore(I5_savedSP, G0, SP);
1964 // (same old exception object is already in Oexception; see above)
1965 // Note that an "issuing PC" is actually the next PC after the call
1966 }
1969 //
1970 // JVMTI ForceEarlyReturn support
1971 //
1973 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1974 address entry = __ pc();
1976 __ empty_expression_stack();
1977 __ load_earlyret_value(state);
1979 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1980 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1982 // Clear the earlyret state
1983 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1985 __ remove_activation(state,
1986 /* throw_monitor_exception */ false,
1987 /* install_monitor_exception */ false);
1989 // The caller's SP was adjusted upon method entry to accomodate
1990 // the callee's non-argument locals. Undo that adjustment.
1991 __ ret(); // return to caller
1992 __ delayed()->restore(I5_savedSP, G0, SP);
1994 return entry;
1995 } // end of JVMTI ForceEarlyReturn support
1998 //------------------------------------------------------------------------------------------------------------------------
1999 // Helper for vtos entry point generation
2001 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t, address& bep, address& cep, address& sep, address& aep, address& iep, address& lep, address& fep, address& dep, address& vep) {
2002 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
2003 Label L;
2004 aep = __ pc(); __ push_ptr(); __ ba_short(L);
2005 fep = __ pc(); __ push_f(); __ ba_short(L);
2006 dep = __ pc(); __ push_d(); __ ba_short(L);
2007 lep = __ pc(); __ push_l(); __ ba_short(L);
2008 iep = __ pc(); __ push_i();
2009 bep = cep = sep = iep; // there aren't any
2010 vep = __ pc(); __ bind(L); // fall through
2011 generate_and_dispatch(t);
2012 }
2014 // --------------------------------------------------------------------------------
2017 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2018 : TemplateInterpreterGenerator(code) {
2019 generate_all(); // down here so it can be "virtual"
2020 }
2022 // --------------------------------------------------------------------------------
2024 // Non-product code
2025 #ifndef PRODUCT
2026 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2027 address entry = __ pc();
2029 __ push(state);
2030 __ mov(O7, Lscratch); // protect return address within interpreter
2032 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
2033 __ mov( Otos_l2, G3_scratch );
2034 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
2035 __ mov(Lscratch, O7); // restore return address
2036 __ pop(state);
2037 __ retl();
2038 __ delayed()->nop();
2040 return entry;
2041 }
2044 // helpers for generate_and_dispatch
2046 void TemplateInterpreterGenerator::count_bytecode() {
2047 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
2048 }
2051 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2052 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
2053 }
2056 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2057 AddressLiteral index (&BytecodePairHistogram::_index);
2058 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
2060 // get index, shift out old bytecode, bring in new bytecode, and store it
2061 // _index = (_index >> log2_number_of_codes) |
2062 // (bytecode << log2_number_of_codes);
2064 __ load_contents(index, G4_scratch);
2065 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
2066 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
2067 __ or3( G3_scratch, G4_scratch, G4_scratch );
2068 __ store_contents(G4_scratch, index, G3_scratch);
2070 // bump bucket contents
2071 // _counters[_index] ++;
2073 __ set(counters, G3_scratch); // loads into G3_scratch
2074 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
2075 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
2076 __ ld (G3_scratch, 0, G4_scratch);
2077 __ inc (G4_scratch);
2078 __ st (G4_scratch, 0, G3_scratch);
2079 }
2082 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2083 // Call a little run-time stub to avoid blow-up for each bytecode.
2084 // The run-time runtime saves the right registers, depending on
2085 // the tosca in-state for the given template.
2086 address entry = Interpreter::trace_code(t->tos_in());
2087 guarantee(entry != NULL, "entry must have been generated");
2088 __ call(entry, relocInfo::none);
2089 __ delayed()->nop();
2090 }
2093 void TemplateInterpreterGenerator::stop_interpreter_at() {
2094 AddressLiteral counter(&BytecodeCounter::_counter_value);
2095 __ load_contents(counter, G3_scratch);
2096 AddressLiteral stop_at(&StopInterpreterAt);
2097 __ load_ptr_contents(stop_at, G4_scratch);
2098 __ cmp(G3_scratch, G4_scratch);
2099 __ breakpoint_trap(Assembler::equal, Assembler::icc);
2100 }
2101 #endif // not PRODUCT
2102 #endif // !CC_INTERP