0
|
1 /*
|
|
2 * Copyright 1997-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
|
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or
|
|
21 * have any questions.
|
|
22 *
|
|
23 */
|
|
24
|
|
25 #include "incls/_precompiled.incl"
|
|
26 #include "incls/_templateInterpreter_sparc.cpp.incl"
|
|
27
|
|
28 #ifndef CC_INTERP
|
|
29 #ifndef FAST_DISPATCH
|
|
30 #define FAST_DISPATCH 1
|
|
31 #endif
|
|
32 #undef FAST_DISPATCH
|
|
33
|
|
34
|
|
35 // Generation of Interpreter
|
|
36 //
|
|
37 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
|
|
38
|
|
39
|
|
40 #define __ _masm->
|
|
41
|
|
42
|
|
43 //----------------------------------------------------------------------------------------------------
|
|
44
|
|
45
|
|
46 void InterpreterGenerator::save_native_result(void) {
|
|
47 // result potentially in O0/O1: save it across calls
|
|
48 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
|
|
49
|
|
50 // result potentially in F0/F1: save it across calls
|
|
51 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
|
|
52
|
|
53 // save and restore any potential method result value around the unlocking operation
|
|
54 __ stf(FloatRegisterImpl::D, F0, d_tmp);
|
|
55 #ifdef _LP64
|
|
56 __ stx(O0, l_tmp);
|
|
57 #else
|
|
58 __ std(O0, l_tmp);
|
|
59 #endif
|
|
60 }
|
|
61
|
|
62 void InterpreterGenerator::restore_native_result(void) {
|
|
63 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
|
|
64 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
|
|
65
|
|
66 // Restore any method result value
|
|
67 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
|
|
68 #ifdef _LP64
|
|
69 __ ldx(l_tmp, O0);
|
|
70 #else
|
|
71 __ ldd(l_tmp, O0);
|
|
72 #endif
|
|
73 }
|
|
74
|
|
75 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
|
|
76 assert(!pass_oop || message == NULL, "either oop or message but not both");
|
|
77 address entry = __ pc();
|
|
78 // expression stack must be empty before entering the VM if an exception happened
|
|
79 __ empty_expression_stack();
|
|
80 // load exception object
|
|
81 __ set((intptr_t)name, G3_scratch);
|
|
82 if (pass_oop) {
|
|
83 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
|
|
84 } else {
|
|
85 __ set((intptr_t)message, G4_scratch);
|
|
86 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
|
|
87 }
|
|
88 // throw exception
|
|
89 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
|
|
90 Address thrower(G3_scratch, Interpreter::throw_exception_entry());
|
|
91 __ jump_to (thrower);
|
|
92 __ delayed()->nop();
|
|
93 return entry;
|
|
94 }
|
|
95
|
|
96 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
|
|
97 address entry = __ pc();
|
|
98 // expression stack must be empty before entering the VM if an exception
|
|
99 // happened
|
|
100 __ empty_expression_stack();
|
|
101 // load exception object
|
|
102 __ call_VM(Oexception,
|
|
103 CAST_FROM_FN_PTR(address,
|
|
104 InterpreterRuntime::throw_ClassCastException),
|
|
105 Otos_i);
|
|
106 __ should_not_reach_here();
|
|
107 return entry;
|
|
108 }
|
|
109
|
|
110
|
|
111 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
|
|
112 address entry = __ pc();
|
|
113 // expression stack must be empty before entering the VM if an exception happened
|
|
114 __ empty_expression_stack();
|
|
115 // convention: expect aberrant index in register G3_scratch, then shuffle the
|
|
116 // index to G4_scratch for the VM call
|
|
117 __ mov(G3_scratch, G4_scratch);
|
|
118 __ set((intptr_t)name, G3_scratch);
|
|
119 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
|
|
120 __ should_not_reach_here();
|
|
121 return entry;
|
|
122 }
|
|
123
|
|
124
|
|
125 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
|
|
126 address entry = __ pc();
|
|
127 // expression stack must be empty before entering the VM if an exception happened
|
|
128 __ empty_expression_stack();
|
|
129 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
|
|
130 __ should_not_reach_here();
|
|
131 return entry;
|
|
132 }
|
|
133
|
|
134
|
|
135 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) {
|
|
136 address compiled_entry = __ pc();
|
|
137 Label cont;
|
|
138
|
|
139 address entry = __ pc();
|
|
140 #if !defined(_LP64) && defined(COMPILER2)
|
|
141 // All return values are where we want them, except for Longs. C2 returns
|
|
142 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
|
|
143 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
|
|
144 // build even if we are returning from interpreted we just do a little
|
|
145 // stupid shuffing.
|
|
146 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
|
|
147 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
|
|
148 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
|
|
149
|
|
150 if( state == ltos ) {
|
|
151 __ srl (G1, 0,O1);
|
|
152 __ srlx(G1,32,O0);
|
|
153 }
|
|
154 #endif /* !_LP64 && COMPILER2 */
|
|
155
|
|
156
|
|
157 __ bind(cont);
|
|
158
|
|
159 // The callee returns with the stack possibly adjusted by adapter transition
|
|
160 // We remove that possible adjustment here.
|
|
161 // All interpreter local registers are untouched. Any result is passed back
|
|
162 // in the O0/O1 or float registers. Before continuing, the arguments must be
|
|
163 // popped from the java expression stack; i.e., Lesp must be adjusted.
|
|
164
|
|
165 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
|
|
166
|
|
167
|
|
168 const Register cache = G3_scratch;
|
|
169 const Register size = G1_scratch;
|
|
170 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1);
|
|
171 __ ld_ptr(Address(cache, 0, in_bytes(constantPoolCacheOopDesc::base_offset()) +
|
|
172 in_bytes(ConstantPoolCacheEntry::flags_offset())), size);
|
|
173 __ and3(size, 0xFF, size); // argument size in words
|
|
174 __ sll(size, Interpreter::logStackElementSize(), size); // each argument size in bytes
|
|
175 __ add(Lesp, size, Lesp); // pop arguments
|
|
176 __ dispatch_next(state, step);
|
|
177
|
|
178 return entry;
|
|
179 }
|
|
180
|
|
181
|
|
182 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
|
|
183 address entry = __ pc();
|
|
184 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
|
|
185 { Label L;
|
|
186 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
|
|
187
|
|
188 __ ld_ptr(exception_addr, Gtemp);
|
|
189 __ tst(Gtemp);
|
|
190 __ brx(Assembler::equal, false, Assembler::pt, L);
|
|
191 __ delayed()->nop();
|
|
192 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
|
|
193 __ should_not_reach_here();
|
|
194 __ bind(L);
|
|
195 }
|
|
196 __ dispatch_next(state, step);
|
|
197 return entry;
|
|
198 }
|
|
199
|
|
200 // A result handler converts/unboxes a native call result into
|
|
201 // a java interpreter/compiler result. The current frame is an
|
|
202 // interpreter frame. The activation frame unwind code must be
|
|
203 // consistent with that of TemplateTable::_return(...). In the
|
|
204 // case of native methods, the caller's SP was not modified.
|
|
205 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
|
|
206 address entry = __ pc();
|
|
207 Register Itos_i = Otos_i ->after_save();
|
|
208 Register Itos_l = Otos_l ->after_save();
|
|
209 Register Itos_l1 = Otos_l1->after_save();
|
|
210 Register Itos_l2 = Otos_l2->after_save();
|
|
211 switch (type) {
|
|
212 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
|
|
213 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
|
|
214 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
|
|
215 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
|
|
216 case T_LONG :
|
|
217 #ifndef _LP64
|
|
218 __ mov(O1, Itos_l2); // move other half of long
|
|
219 #endif // ifdef or no ifdef, fall through to the T_INT case
|
|
220 case T_INT : __ mov(O0, Itos_i); break;
|
|
221 case T_VOID : /* nothing to do */ break;
|
|
222 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
|
|
223 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
|
|
224 case T_OBJECT :
|
|
225 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
|
|
226 __ verify_oop(Itos_i);
|
|
227 break;
|
|
228 default : ShouldNotReachHere();
|
|
229 }
|
|
230 __ ret(); // return from interpreter activation
|
|
231 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
|
|
232 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
|
|
233 return entry;
|
|
234 }
|
|
235
|
|
236 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
|
|
237 address entry = __ pc();
|
|
238 __ push(state);
|
|
239 __ call_VM(noreg, runtime_entry);
|
|
240 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
|
|
241 return entry;
|
|
242 }
|
|
243
|
|
244
|
|
245 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
|
|
246 address entry = __ pc();
|
|
247 __ dispatch_next(state);
|
|
248 return entry;
|
|
249 }
|
|
250
|
|
251 //
|
|
252 // Helpers for commoning out cases in the various type of method entries.
|
|
253 //
|
|
254
|
|
255 // increment invocation count & check for overflow
|
|
256 //
|
|
257 // Note: checking for negative value instead of overflow
|
|
258 // so we have a 'sticky' overflow test
|
|
259 //
|
|
260 // Lmethod: method
|
|
261 // ??: invocation counter
|
|
262 //
|
|
263 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
|
|
264 // Update standard invocation counters
|
|
265 __ increment_invocation_counter(O0, G3_scratch);
|
|
266 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
|
|
267 Address interpreter_invocation_counter(Lmethod, 0, in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
|
|
268 __ ld(interpreter_invocation_counter, G3_scratch);
|
|
269 __ inc(G3_scratch);
|
|
270 __ st(G3_scratch, interpreter_invocation_counter);
|
|
271 }
|
|
272
|
|
273 if (ProfileInterpreter && profile_method != NULL) {
|
|
274 // Test to see if we should create a method data oop
|
|
275 Address profile_limit(G3_scratch, (address)&InvocationCounter::InterpreterProfileLimit);
|
|
276 __ sethi(profile_limit);
|
|
277 __ ld(profile_limit, G3_scratch);
|
|
278 __ cmp(O0, G3_scratch);
|
|
279 __ br(Assembler::lessUnsigned, false, Assembler::pn, *profile_method_continue);
|
|
280 __ delayed()->nop();
|
|
281
|
|
282 // if no method data exists, go to profile_method
|
|
283 __ test_method_data_pointer(*profile_method);
|
|
284 }
|
|
285
|
|
286 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
|
|
287 __ sethi(invocation_limit);
|
|
288 __ ld(invocation_limit, G3_scratch);
|
|
289 __ cmp(O0, G3_scratch);
|
|
290 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
|
|
291 __ delayed()->nop();
|
|
292
|
|
293 }
|
|
294
|
|
295 // Allocate monitor and lock method (asm interpreter)
|
|
296 // ebx - methodOop
|
|
297 //
|
|
298 void InterpreterGenerator::lock_method(void) {
|
|
299 const Address access_flags (Lmethod, 0, in_bytes(methodOopDesc::access_flags_offset()));
|
|
300 __ ld(access_flags, O0);
|
|
301
|
|
302 #ifdef ASSERT
|
|
303 { Label ok;
|
|
304 __ btst(JVM_ACC_SYNCHRONIZED, O0);
|
|
305 __ br( Assembler::notZero, false, Assembler::pt, ok);
|
|
306 __ delayed()->nop();
|
|
307 __ stop("method doesn't need synchronization");
|
|
308 __ bind(ok);
|
|
309 }
|
|
310 #endif // ASSERT
|
|
311
|
|
312 // get synchronization object to O0
|
|
313 { Label done;
|
|
314 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
|
|
315 __ btst(JVM_ACC_STATIC, O0);
|
|
316 __ br( Assembler::zero, true, Assembler::pt, done);
|
|
317 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
|
|
318
|
|
319 __ ld_ptr( Lmethod, in_bytes(methodOopDesc::constants_offset()), O0);
|
|
320 __ ld_ptr( O0, constantPoolOopDesc::pool_holder_offset_in_bytes(), O0);
|
|
321
|
|
322 // lock the mirror, not the klassOop
|
|
323 __ ld_ptr( O0, mirror_offset, O0);
|
|
324
|
|
325 #ifdef ASSERT
|
|
326 __ tst(O0);
|
|
327 __ breakpoint_trap(Assembler::zero);
|
|
328 #endif // ASSERT
|
|
329
|
|
330 __ bind(done);
|
|
331 }
|
|
332
|
|
333 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
|
|
334 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
|
|
335 // __ untested("lock_object from method entry");
|
|
336 __ lock_object(Lmonitors, O0);
|
|
337 }
|
|
338
|
|
339
|
|
340 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
|
|
341 Register Rscratch,
|
|
342 Register Rscratch2) {
|
|
343 const int page_size = os::vm_page_size();
|
|
344 Address saved_exception_pc(G2_thread, 0,
|
|
345 in_bytes(JavaThread::saved_exception_pc_offset()));
|
|
346 Label after_frame_check;
|
|
347
|
|
348 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
|
|
349
|
|
350 __ set( page_size, Rscratch );
|
|
351 __ cmp( Rframe_size, Rscratch );
|
|
352
|
|
353 __ br( Assembler::lessEqual, false, Assembler::pt, after_frame_check );
|
|
354 __ delayed()->nop();
|
|
355
|
|
356 // get the stack base, and in debug, verify it is non-zero
|
|
357 __ ld_ptr( G2_thread, in_bytes(Thread::stack_base_offset()), Rscratch );
|
|
358 #ifdef ASSERT
|
|
359 Label base_not_zero;
|
|
360 __ cmp( Rscratch, G0 );
|
|
361 __ brx( Assembler::notEqual, false, Assembler::pn, base_not_zero );
|
|
362 __ delayed()->nop();
|
|
363 __ stop("stack base is zero in generate_stack_overflow_check");
|
|
364 __ bind(base_not_zero);
|
|
365 #endif
|
|
366
|
|
367 // get the stack size, and in debug, verify it is non-zero
|
|
368 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
|
|
369 __ ld_ptr( G2_thread, in_bytes(Thread::stack_size_offset()), Rscratch2 );
|
|
370 #ifdef ASSERT
|
|
371 Label size_not_zero;
|
|
372 __ cmp( Rscratch2, G0 );
|
|
373 __ brx( Assembler::notEqual, false, Assembler::pn, size_not_zero );
|
|
374 __ delayed()->nop();
|
|
375 __ stop("stack size is zero in generate_stack_overflow_check");
|
|
376 __ bind(size_not_zero);
|
|
377 #endif
|
|
378
|
|
379 // compute the beginning of the protected zone minus the requested frame size
|
|
380 __ sub( Rscratch, Rscratch2, Rscratch );
|
|
381 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
|
|
382 __ add( Rscratch, Rscratch2, Rscratch );
|
|
383
|
|
384 // Add in the size of the frame (which is the same as subtracting it from the
|
|
385 // SP, which would take another register
|
|
386 __ add( Rscratch, Rframe_size, Rscratch );
|
|
387
|
|
388 // the frame is greater than one page in size, so check against
|
|
389 // the bottom of the stack
|
|
390 __ cmp( SP, Rscratch );
|
|
391 __ brx( Assembler::greater, false, Assembler::pt, after_frame_check );
|
|
392 __ delayed()->nop();
|
|
393
|
|
394 // Save the return address as the exception pc
|
|
395 __ st_ptr(O7, saved_exception_pc);
|
|
396
|
|
397 // the stack will overflow, throw an exception
|
|
398 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
|
|
399
|
|
400 // if you get to here, then there is enough stack space
|
|
401 __ bind( after_frame_check );
|
|
402 }
|
|
403
|
|
404
|
|
405 //
|
|
406 // Generate a fixed interpreter frame. This is identical setup for interpreted
|
|
407 // methods and for native methods hence the shared code.
|
|
408
|
|
409 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
|
|
410 //
|
|
411 //
|
|
412 // The entry code sets up a new interpreter frame in 4 steps:
|
|
413 //
|
|
414 // 1) Increase caller's SP by for the extra local space needed:
|
|
415 // (check for overflow)
|
|
416 // Efficient implementation of xload/xstore bytecodes requires
|
|
417 // that arguments and non-argument locals are in a contigously
|
|
418 // addressable memory block => non-argument locals must be
|
|
419 // allocated in the caller's frame.
|
|
420 //
|
|
421 // 2) Create a new stack frame and register window:
|
|
422 // The new stack frame must provide space for the standard
|
|
423 // register save area, the maximum java expression stack size,
|
|
424 // the monitor slots (0 slots initially), and some frame local
|
|
425 // scratch locations.
|
|
426 //
|
|
427 // 3) The following interpreter activation registers must be setup:
|
|
428 // Lesp : expression stack pointer
|
|
429 // Lbcp : bytecode pointer
|
|
430 // Lmethod : method
|
|
431 // Llocals : locals pointer
|
|
432 // Lmonitors : monitor pointer
|
|
433 // LcpoolCache: constant pool cache
|
|
434 //
|
|
435 // 4) Initialize the non-argument locals if necessary:
|
|
436 // Non-argument locals may need to be initialized to NULL
|
|
437 // for GC to work. If the oop-map information is accurate
|
|
438 // (in the absence of the JSR problem), no initialization
|
|
439 // is necessary.
|
|
440 //
|
|
441 // (gri - 2/25/2000)
|
|
442
|
|
443
|
|
444 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
|
|
445 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
|
|
446 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
|
|
447 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
|
|
448
|
|
449 const int extra_space =
|
|
450 rounded_vm_local_words + // frame local scratch space
|
|
451 frame::memory_parameter_word_sp_offset + // register save area
|
|
452 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
|
|
453
|
|
454 const Register Glocals_size = G3;
|
|
455 const Register Otmp1 = O3;
|
|
456 const Register Otmp2 = O4;
|
|
457 // Lscratch can't be used as a temporary because the call_stub uses
|
|
458 // it to assert that the stack frame was setup correctly.
|
|
459
|
|
460 __ lduh( size_of_parameters, Glocals_size);
|
|
461
|
|
462 // Gargs points to first local + BytesPerWord
|
|
463 // Set the saved SP after the register window save
|
|
464 //
|
|
465 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
|
|
466 __ sll(Glocals_size, Interpreter::logStackElementSize(), Otmp1);
|
|
467 __ add(Gargs, Otmp1, Gargs);
|
|
468
|
|
469 if (native_call) {
|
|
470 __ calc_mem_param_words( Glocals_size, Gframe_size );
|
|
471 __ add( Gframe_size, extra_space, Gframe_size);
|
|
472 __ round_to( Gframe_size, WordsPerLong );
|
|
473 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
|
|
474 } else {
|
|
475
|
|
476 //
|
|
477 // Compute number of locals in method apart from incoming parameters
|
|
478 //
|
|
479 __ lduh( size_of_locals, Otmp1 );
|
|
480 __ sub( Otmp1, Glocals_size, Glocals_size );
|
|
481 __ round_to( Glocals_size, WordsPerLong );
|
|
482 __ sll( Glocals_size, Interpreter::logStackElementSize(), Glocals_size );
|
|
483
|
|
484 // see if the frame is greater than one page in size. If so,
|
|
485 // then we need to verify there is enough stack space remaining
|
|
486 // Frame_size = (max_stack + extra_space) * BytesPerWord;
|
|
487 __ lduh( max_stack, Gframe_size );
|
|
488 __ add( Gframe_size, extra_space, Gframe_size );
|
|
489 __ round_to( Gframe_size, WordsPerLong );
|
|
490 __ sll( Gframe_size, Interpreter::logStackElementSize(), Gframe_size);
|
|
491
|
|
492 // Add in java locals size for stack overflow check only
|
|
493 __ add( Gframe_size, Glocals_size, Gframe_size );
|
|
494
|
|
495 const Register Otmp2 = O4;
|
|
496 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
|
|
497 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
|
|
498
|
|
499 __ sub( Gframe_size, Glocals_size, Gframe_size);
|
|
500
|
|
501 //
|
|
502 // bump SP to accomodate the extra locals
|
|
503 //
|
|
504 __ sub( SP, Glocals_size, SP );
|
|
505 }
|
|
506
|
|
507 //
|
|
508 // now set up a stack frame with the size computed above
|
|
509 //
|
|
510 __ neg( Gframe_size );
|
|
511 __ save( SP, Gframe_size, SP );
|
|
512
|
|
513 //
|
|
514 // now set up all the local cache registers
|
|
515 //
|
|
516 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
|
|
517 // that all present references to Lbyte_code initialize the register
|
|
518 // immediately before use
|
|
519 if (native_call) {
|
|
520 __ mov(G0, Lbcp);
|
|
521 } else {
|
|
522 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), Lbcp );
|
|
523 __ add(Address(Lbcp, 0, in_bytes(constMethodOopDesc::codes_offset())), Lbcp );
|
|
524 }
|
|
525 __ mov( G5_method, Lmethod); // set Lmethod
|
|
526 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
|
|
527 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
|
|
528 #ifdef _LP64
|
|
529 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
|
|
530 #endif
|
|
531 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
|
|
532
|
|
533 // setup interpreter activation registers
|
|
534 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
|
|
535
|
|
536 if (ProfileInterpreter) {
|
|
537 #ifdef FAST_DISPATCH
|
|
538 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
|
|
539 // they both use I2.
|
|
540 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
|
|
541 #endif // FAST_DISPATCH
|
|
542 __ set_method_data_pointer();
|
|
543 }
|
|
544
|
|
545 }
|
|
546
|
|
547 // Empty method, generate a very fast return.
|
|
548
|
|
549 address InterpreterGenerator::generate_empty_entry(void) {
|
|
550
|
|
551 // A method that does nother but return...
|
|
552
|
|
553 address entry = __ pc();
|
|
554 Label slow_path;
|
|
555
|
|
556 __ verify_oop(G5_method);
|
|
557
|
|
558 // do nothing for empty methods (do not even increment invocation counter)
|
|
559 if ( UseFastEmptyMethods) {
|
|
560 // If we need a safepoint check, generate full interpreter entry.
|
|
561 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
|
|
562 __ load_contents(sync_state, G3_scratch);
|
|
563 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
|
|
564 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
|
|
565 __ delayed()->nop();
|
|
566
|
|
567 // Code: _return
|
|
568 __ retl();
|
|
569 __ delayed()->mov(O5_savedSP, SP);
|
|
570
|
|
571 __ bind(slow_path);
|
|
572 (void) generate_normal_entry(false);
|
|
573
|
|
574 return entry;
|
|
575 }
|
|
576 return NULL;
|
|
577 }
|
|
578
|
|
579 // Call an accessor method (assuming it is resolved, otherwise drop into
|
|
580 // vanilla (slow path) entry
|
|
581
|
|
582 // Generates code to elide accessor methods
|
|
583 // Uses G3_scratch and G1_scratch as scratch
|
|
584 address InterpreterGenerator::generate_accessor_entry(void) {
|
|
585
|
|
586 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
|
|
587 // parameter size = 1
|
|
588 // Note: We can only use this code if the getfield has been resolved
|
|
589 // and if we don't have a null-pointer exception => check for
|
|
590 // these conditions first and use slow path if necessary.
|
|
591 address entry = __ pc();
|
|
592 Label slow_path;
|
|
593
|
|
594 if ( UseFastAccessorMethods) {
|
|
595 // Check if we need to reach a safepoint and generate full interpreter
|
|
596 // frame if so.
|
|
597 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
|
|
598 __ load_contents(sync_state, G3_scratch);
|
|
599 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
|
|
600 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
|
|
601 __ delayed()->nop();
|
|
602
|
|
603 // Check if local 0 != NULL
|
|
604 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
|
|
605 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
|
|
606 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
|
|
607 __ delayed()->nop();
|
|
608
|
|
609
|
|
610 // read first instruction word and extract bytecode @ 1 and index @ 2
|
|
611 // get first 4 bytes of the bytecodes (big endian!)
|
|
612 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), G1_scratch);
|
|
613 __ ld(Address(G1_scratch, 0, in_bytes(constMethodOopDesc::codes_offset())), G1_scratch);
|
|
614
|
|
615 // move index @ 2 far left then to the right most two bytes.
|
|
616 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
|
|
617 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
|
|
618 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
|
|
619
|
|
620 // get constant pool cache
|
|
621 __ ld_ptr(G5_method, in_bytes(methodOopDesc::constants_offset()), G3_scratch);
|
|
622 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
|
|
623
|
|
624 // get specific constant pool cache entry
|
|
625 __ add(G3_scratch, G1_scratch, G3_scratch);
|
|
626
|
|
627 // Check the constant Pool cache entry to see if it has been resolved.
|
|
628 // If not, need the slow path.
|
|
629 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
|
|
630 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
|
|
631 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
|
|
632 __ and3(G1_scratch, 0xFF, G1_scratch);
|
|
633 __ cmp(G1_scratch, Bytecodes::_getfield);
|
|
634 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
|
|
635 __ delayed()->nop();
|
|
636
|
|
637 // Get the type and return field offset from the constant pool cache
|
|
638 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
|
|
639 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
|
|
640
|
|
641 Label xreturn_path;
|
|
642 // Need to differentiate between igetfield, agetfield, bgetfield etc.
|
|
643 // because they are different sizes.
|
|
644 // Get the type from the constant pool cache
|
|
645 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
|
|
646 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
|
|
647 ConstantPoolCacheEntry::verify_tosBits();
|
|
648 __ cmp(G1_scratch, atos );
|
|
649 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
|
|
650 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
|
|
651 __ cmp(G1_scratch, itos);
|
|
652 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
|
|
653 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
|
|
654 __ cmp(G1_scratch, stos);
|
|
655 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
|
|
656 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
|
|
657 __ cmp(G1_scratch, ctos);
|
|
658 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
|
|
659 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
|
|
660 #ifdef ASSERT
|
|
661 __ cmp(G1_scratch, btos);
|
|
662 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
|
|
663 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
|
|
664 __ should_not_reach_here();
|
|
665 #endif
|
|
666 __ ldsb(Otos_i, G3_scratch, Otos_i);
|
|
667 __ bind(xreturn_path);
|
|
668
|
|
669 // _ireturn/_areturn
|
|
670 __ retl(); // return from leaf routine
|
|
671 __ delayed()->mov(O5_savedSP, SP);
|
|
672
|
|
673 // Generate regular method entry
|
|
674 __ bind(slow_path);
|
|
675 (void) generate_normal_entry(false);
|
|
676 return entry;
|
|
677 }
|
|
678 return NULL;
|
|
679 }
|
|
680
|
|
681 //
|
|
682 // Interpreter stub for calling a native method. (asm interpreter)
|
|
683 // This sets up a somewhat different looking stack for calling the native method
|
|
684 // than the typical interpreter frame setup.
|
|
685 //
|
|
686
|
|
687 address InterpreterGenerator::generate_native_entry(bool synchronized) {
|
|
688 address entry = __ pc();
|
|
689
|
|
690 // the following temporary registers are used during frame creation
|
|
691 const Register Gtmp1 = G3_scratch ;
|
|
692 const Register Gtmp2 = G1_scratch;
|
|
693 bool inc_counter = UseCompiler || CountCompiledCalls;
|
|
694
|
|
695 // make sure registers are different!
|
|
696 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
|
|
697
|
|
698 const Address Laccess_flags (Lmethod, 0, in_bytes(methodOopDesc::access_flags_offset()));
|
|
699
|
|
700 __ verify_oop(G5_method);
|
|
701
|
|
702 const Register Glocals_size = G3;
|
|
703 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
|
|
704
|
|
705 // make sure method is native & not abstract
|
|
706 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
|
|
707 #ifdef ASSERT
|
|
708 __ ld(G5_method, in_bytes(methodOopDesc::access_flags_offset()), Gtmp1);
|
|
709 {
|
|
710 Label L;
|
|
711 __ btst(JVM_ACC_NATIVE, Gtmp1);
|
|
712 __ br(Assembler::notZero, false, Assembler::pt, L);
|
|
713 __ delayed()->nop();
|
|
714 __ stop("tried to execute non-native method as native");
|
|
715 __ bind(L);
|
|
716 }
|
|
717 { Label L;
|
|
718 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
|
|
719 __ br(Assembler::zero, false, Assembler::pt, L);
|
|
720 __ delayed()->nop();
|
|
721 __ stop("tried to execute abstract method as non-abstract");
|
|
722 __ bind(L);
|
|
723 }
|
|
724 #endif // ASSERT
|
|
725
|
|
726 // generate the code to allocate the interpreter stack frame
|
|
727 generate_fixed_frame(true);
|
|
728
|
|
729 //
|
|
730 // No locals to initialize for native method
|
|
731 //
|
|
732
|
|
733 // this slot will be set later, we initialize it to null here just in
|
|
734 // case we get a GC before the actual value is stored later
|
|
735 __ st_ptr(G0, Address(FP, 0, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS));
|
|
736
|
|
737 const Address do_not_unlock_if_synchronized(G2_thread, 0,
|
|
738 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
|
|
739 // Since at this point in the method invocation the exception handler
|
|
740 // would try to exit the monitor of synchronized methods which hasn't
|
|
741 // been entered yet, we set the thread local variable
|
|
742 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
|
|
743 // runtime, exception handling i.e. unlock_if_synchronized_method will
|
|
744 // check this thread local flag.
|
|
745 // This flag has two effects, one is to force an unwind in the topmost
|
|
746 // interpreter frame and not perform an unlock while doing so.
|
|
747
|
|
748 __ movbool(true, G3_scratch);
|
|
749 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
|
|
750
|
|
751 // increment invocation counter and check for overflow
|
|
752 //
|
|
753 // Note: checking for negative value instead of overflow
|
|
754 // so we have a 'sticky' overflow test (may be of
|
|
755 // importance as soon as we have true MT/MP)
|
|
756 Label invocation_counter_overflow;
|
|
757 Label Lcontinue;
|
|
758 if (inc_counter) {
|
|
759 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
|
|
760
|
|
761 }
|
|
762 __ bind(Lcontinue);
|
|
763
|
|
764 bang_stack_shadow_pages(true);
|
|
765
|
|
766 // reset the _do_not_unlock_if_synchronized flag
|
|
767 __ stbool(G0, do_not_unlock_if_synchronized);
|
|
768
|
|
769 // check for synchronized methods
|
|
770 // Must happen AFTER invocation_counter check and stack overflow check,
|
|
771 // so method is not locked if overflows.
|
|
772
|
|
773 if (synchronized) {
|
|
774 lock_method();
|
|
775 } else {
|
|
776 #ifdef ASSERT
|
|
777 { Label ok;
|
|
778 __ ld(Laccess_flags, O0);
|
|
779 __ btst(JVM_ACC_SYNCHRONIZED, O0);
|
|
780 __ br( Assembler::zero, false, Assembler::pt, ok);
|
|
781 __ delayed()->nop();
|
|
782 __ stop("method needs synchronization");
|
|
783 __ bind(ok);
|
|
784 }
|
|
785 #endif // ASSERT
|
|
786 }
|
|
787
|
|
788
|
|
789 // start execution
|
|
790 __ verify_thread();
|
|
791
|
|
792 // JVMTI support
|
|
793 __ notify_method_entry();
|
|
794
|
|
795 // native call
|
|
796
|
|
797 // (note that O0 is never an oop--at most it is a handle)
|
|
798 // It is important not to smash any handles created by this call,
|
|
799 // until any oop handle in O0 is dereferenced.
|
|
800
|
|
801 // (note that the space for outgoing params is preallocated)
|
|
802
|
|
803 // get signature handler
|
|
804 { Label L;
|
|
805 __ ld_ptr(Address(Lmethod, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
|
|
806 __ tst(G3_scratch);
|
|
807 __ brx(Assembler::notZero, false, Assembler::pt, L);
|
|
808 __ delayed()->nop();
|
|
809 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
|
|
810 __ ld_ptr(Address(Lmethod, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
|
|
811 __ bind(L);
|
|
812 }
|
|
813
|
|
814 // Push a new frame so that the args will really be stored in
|
|
815 // Copy a few locals across so the new frame has the variables
|
|
816 // we need but these values will be dead at the jni call and
|
|
817 // therefore not gc volatile like the values in the current
|
|
818 // frame (Lmethod in particular)
|
|
819
|
|
820 // Flush the method pointer to the register save area
|
|
821 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
|
|
822 __ mov(Llocals, O1);
|
|
823 // calculate where the mirror handle body is allocated in the interpreter frame:
|
|
824
|
|
825 Address mirror(FP, 0, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
|
|
826 __ add(mirror, O2);
|
|
827
|
|
828 // Calculate current frame size
|
|
829 __ sub(SP, FP, O3); // Calculate negative of current frame size
|
|
830 __ save(SP, O3, SP); // Allocate an identical sized frame
|
|
831
|
|
832 // Note I7 has leftover trash. Slow signature handler will fill it in
|
|
833 // should we get there. Normal jni call will set reasonable last_Java_pc
|
|
834 // below (and fix I7 so the stack trace doesn't have a meaningless frame
|
|
835 // in it).
|
|
836
|
|
837 // Load interpreter frame's Lmethod into same register here
|
|
838
|
|
839 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
|
|
840
|
|
841 __ mov(I1, Llocals);
|
|
842 __ mov(I2, Lscratch2); // save the address of the mirror
|
|
843
|
|
844
|
|
845 // ONLY Lmethod and Llocals are valid here!
|
|
846
|
|
847 // call signature handler, It will move the arg properly since Llocals in current frame
|
|
848 // matches that in outer frame
|
|
849
|
|
850 __ callr(G3_scratch, 0);
|
|
851 __ delayed()->nop();
|
|
852
|
|
853 // Result handler is in Lscratch
|
|
854
|
|
855 // Reload interpreter frame's Lmethod since slow signature handler may block
|
|
856 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
|
|
857
|
|
858 { Label not_static;
|
|
859
|
|
860 __ ld(Laccess_flags, O0);
|
|
861 __ btst(JVM_ACC_STATIC, O0);
|
|
862 __ br( Assembler::zero, false, Assembler::pt, not_static);
|
|
863 __ delayed()->
|
|
864 // get native function entry point(O0 is a good temp until the very end)
|
|
865 ld_ptr(Address(Lmethod, 0, in_bytes(methodOopDesc::native_function_offset())), O0);
|
|
866 // for static methods insert the mirror argument
|
|
867 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
|
|
868
|
|
869 __ ld_ptr(Address(Lmethod, 0, in_bytes(methodOopDesc:: constants_offset())), O1);
|
|
870 __ ld_ptr(Address(O1, 0, constantPoolOopDesc::pool_holder_offset_in_bytes()), O1);
|
|
871 __ ld_ptr(O1, mirror_offset, O1);
|
|
872 #ifdef ASSERT
|
|
873 if (!PrintSignatureHandlers) // do not dirty the output with this
|
|
874 { Label L;
|
|
875 __ tst(O1);
|
|
876 __ brx(Assembler::notZero, false, Assembler::pt, L);
|
|
877 __ delayed()->nop();
|
|
878 __ stop("mirror is missing");
|
|
879 __ bind(L);
|
|
880 }
|
|
881 #endif // ASSERT
|
|
882 __ st_ptr(O1, Lscratch2, 0);
|
|
883 __ mov(Lscratch2, O1);
|
|
884 __ bind(not_static);
|
|
885 }
|
|
886
|
|
887 // At this point, arguments have been copied off of stack into
|
|
888 // their JNI positions, which are O1..O5 and SP[68..].
|
|
889 // Oops are boxed in-place on the stack, with handles copied to arguments.
|
|
890 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
|
|
891
|
|
892 #ifdef ASSERT
|
|
893 { Label L;
|
|
894 __ tst(O0);
|
|
895 __ brx(Assembler::notZero, false, Assembler::pt, L);
|
|
896 __ delayed()->nop();
|
|
897 __ stop("native entry point is missing");
|
|
898 __ bind(L);
|
|
899 }
|
|
900 #endif // ASSERT
|
|
901
|
|
902 //
|
|
903 // setup the frame anchor
|
|
904 //
|
|
905 // The scavenge function only needs to know that the PC of this frame is
|
|
906 // in the interpreter method entry code, it doesn't need to know the exact
|
|
907 // PC and hence we can use O7 which points to the return address from the
|
|
908 // previous call in the code stream (signature handler function)
|
|
909 //
|
|
910 // The other trick is we set last_Java_sp to FP instead of the usual SP because
|
|
911 // we have pushed the extra frame in order to protect the volatile register(s)
|
|
912 // in that frame when we return from the jni call
|
|
913 //
|
|
914
|
|
915 __ set_last_Java_frame(FP, O7);
|
|
916 __ mov(O7, I7); // make dummy interpreter frame look like one above,
|
|
917 // not meaningless information that'll confuse me.
|
|
918
|
|
919 // flush the windows now. We don't care about the current (protection) frame
|
|
920 // only the outer frames
|
|
921
|
|
922 __ flush_windows();
|
|
923
|
|
924 // mark windows as flushed
|
|
925 Address flags(G2_thread,
|
|
926 0,
|
|
927 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
|
|
928 __ set(JavaFrameAnchor::flushed, G3_scratch);
|
|
929 __ st(G3_scratch, flags);
|
|
930
|
|
931 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
|
|
932
|
|
933 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
|
|
934 #ifdef ASSERT
|
|
935 { Label L;
|
|
936 __ ld(thread_state, G3_scratch);
|
|
937 __ cmp(G3_scratch, _thread_in_Java);
|
|
938 __ br(Assembler::equal, false, Assembler::pt, L);
|
|
939 __ delayed()->nop();
|
|
940 __ stop("Wrong thread state in native stub");
|
|
941 __ bind(L);
|
|
942 }
|
|
943 #endif // ASSERT
|
|
944 __ set(_thread_in_native, G3_scratch);
|
|
945 __ st(G3_scratch, thread_state);
|
|
946
|
|
947 // Call the jni method, using the delay slot to set the JNIEnv* argument.
|
|
948 __ save_thread(L7_thread_cache); // save Gthread
|
|
949 __ callr(O0, 0);
|
|
950 __ delayed()->
|
|
951 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
|
|
952
|
|
953 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
|
|
954
|
|
955 __ restore_thread(L7_thread_cache); // restore G2_thread
|
|
956
|
|
957 // must we block?
|
|
958
|
|
959 // Block, if necessary, before resuming in _thread_in_Java state.
|
|
960 // In order for GC to work, don't clear the last_Java_sp until after blocking.
|
|
961 { Label no_block;
|
|
962 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
|
|
963
|
|
964 // Switch thread to "native transition" state before reading the synchronization state.
|
|
965 // This additional state is necessary because reading and testing the synchronization
|
|
966 // state is not atomic w.r.t. GC, as this scenario demonstrates:
|
|
967 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
|
|
968 // VM thread changes sync state to synchronizing and suspends threads for GC.
|
|
969 // Thread A is resumed to finish this native method, but doesn't block here since it
|
|
970 // didn't see any synchronization is progress, and escapes.
|
|
971 __ set(_thread_in_native_trans, G3_scratch);
|
|
972 __ st(G3_scratch, thread_state);
|
|
973 if(os::is_MP()) {
|
|
974 if (UseMembar) {
|
|
975 // Force this write out before the read below
|
|
976 __ membar(Assembler::StoreLoad);
|
|
977 } else {
|
|
978 // Write serialization page so VM thread can do a pseudo remote membar.
|
|
979 // We use the current thread pointer to calculate a thread specific
|
|
980 // offset to write to within the page. This minimizes bus traffic
|
|
981 // due to cache line collision.
|
|
982 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
|
|
983 }
|
|
984 }
|
|
985 __ load_contents(sync_state, G3_scratch);
|
|
986 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
|
|
987
|
|
988 Label L;
|
|
989 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
|
|
990 __ br(Assembler::notEqual, false, Assembler::pn, L);
|
|
991 __ delayed()->
|
|
992 ld(suspend_state, G3_scratch);
|
|
993 __ cmp(G3_scratch, 0);
|
|
994 __ br(Assembler::equal, false, Assembler::pt, no_block);
|
|
995 __ delayed()->nop();
|
|
996 __ bind(L);
|
|
997
|
|
998 // Block. Save any potential method result value before the operation and
|
|
999 // use a leaf call to leave the last_Java_frame setup undisturbed.
|
|
1000 save_native_result();
|
|
1001 __ call_VM_leaf(L7_thread_cache,
|
|
1002 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
|
|
1003 G2_thread);
|
|
1004
|
|
1005 // Restore any method result value
|
|
1006 restore_native_result();
|
|
1007 __ bind(no_block);
|
|
1008 }
|
|
1009
|
|
1010 // Clear the frame anchor now
|
|
1011
|
|
1012 __ reset_last_Java_frame();
|
|
1013
|
|
1014 // Move the result handler address
|
|
1015 __ mov(Lscratch, G3_scratch);
|
|
1016 // return possible result to the outer frame
|
|
1017 #ifndef __LP64
|
|
1018 __ mov(O0, I0);
|
|
1019 __ restore(O1, G0, O1);
|
|
1020 #else
|
|
1021 __ restore(O0, G0, O0);
|
|
1022 #endif /* __LP64 */
|
|
1023
|
|
1024 // Move result handler to expected register
|
|
1025 __ mov(G3_scratch, Lscratch);
|
|
1026
|
|
1027 // Back in normal (native) interpreter frame. State is thread_in_native_trans
|
|
1028 // switch to thread_in_Java.
|
|
1029
|
|
1030 __ set(_thread_in_Java, G3_scratch);
|
|
1031 __ st(G3_scratch, thread_state);
|
|
1032
|
|
1033 // reset handle block
|
|
1034 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
|
|
1035 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
|
|
1036
|
|
1037 // If we have an oop result store it where it will be safe for any further gc
|
|
1038 // until we return now that we've released the handle it might be protected by
|
|
1039
|
|
1040 {
|
|
1041 Label no_oop, store_result;
|
|
1042
|
|
1043 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
|
|
1044 __ cmp(G3_scratch, Lscratch);
|
|
1045 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
|
|
1046 __ delayed()->nop();
|
|
1047 __ addcc(G0, O0, O0);
|
|
1048 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
|
|
1049 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
|
|
1050 __ mov(G0, O0);
|
|
1051
|
|
1052 __ bind(store_result);
|
|
1053 // Store it where gc will look for it and result handler expects it.
|
|
1054 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
|
|
1055
|
|
1056 __ bind(no_oop);
|
|
1057
|
|
1058 }
|
|
1059
|
|
1060
|
|
1061 // handle exceptions (exception handling will handle unlocking!)
|
|
1062 { Label L;
|
|
1063 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
|
|
1064
|
|
1065 __ ld_ptr(exception_addr, Gtemp);
|
|
1066 __ tst(Gtemp);
|
|
1067 __ brx(Assembler::equal, false, Assembler::pt, L);
|
|
1068 __ delayed()->nop();
|
|
1069 // Note: This could be handled more efficiently since we know that the native
|
|
1070 // method doesn't have an exception handler. We could directly return
|
|
1071 // to the exception handler for the caller.
|
|
1072 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
|
|
1073 __ should_not_reach_here();
|
|
1074 __ bind(L);
|
|
1075 }
|
|
1076
|
|
1077 // JVMTI support (preserves thread register)
|
|
1078 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
|
|
1079
|
|
1080 if (synchronized) {
|
|
1081 // save and restore any potential method result value around the unlocking operation
|
|
1082 save_native_result();
|
|
1083
|
|
1084 __ add( __ top_most_monitor(), O1);
|
|
1085 __ unlock_object(O1);
|
|
1086
|
|
1087 restore_native_result();
|
|
1088 }
|
|
1089
|
|
1090 #if defined(COMPILER2) && !defined(_LP64)
|
|
1091
|
|
1092 // C2 expects long results in G1 we can't tell if we're returning to interpreted
|
|
1093 // or compiled so just be safe.
|
|
1094
|
|
1095 __ sllx(O0, 32, G1); // Shift bits into high G1
|
|
1096 __ srl (O1, 0, O1); // Zero extend O1
|
|
1097 __ or3 (O1, G1, G1); // OR 64 bits into G1
|
|
1098
|
|
1099 #endif /* COMPILER2 && !_LP64 */
|
|
1100
|
|
1101 // dispose of return address and remove activation
|
|
1102 #ifdef ASSERT
|
|
1103 {
|
|
1104 Label ok;
|
|
1105 __ cmp(I5_savedSP, FP);
|
|
1106 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
|
|
1107 __ delayed()->nop();
|
|
1108 __ stop("bad I5_savedSP value");
|
|
1109 __ should_not_reach_here();
|
|
1110 __ bind(ok);
|
|
1111 }
|
|
1112 #endif
|
|
1113 if (TraceJumps) {
|
|
1114 // Move target to register that is recordable
|
|
1115 __ mov(Lscratch, G3_scratch);
|
|
1116 __ JMP(G3_scratch, 0);
|
|
1117 } else {
|
|
1118 __ jmp(Lscratch, 0);
|
|
1119 }
|
|
1120 __ delayed()->nop();
|
|
1121
|
|
1122
|
|
1123 if (inc_counter) {
|
|
1124 // handle invocation counter overflow
|
|
1125 __ bind(invocation_counter_overflow);
|
|
1126 generate_counter_overflow(Lcontinue);
|
|
1127 }
|
|
1128
|
|
1129
|
|
1130
|
|
1131 return entry;
|
|
1132 }
|
|
1133
|
|
1134
|
|
1135 // Generic method entry to (asm) interpreter
|
|
1136 //------------------------------------------------------------------------------------------------------------------------
|
|
1137 //
|
|
1138 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
|
|
1139 address entry = __ pc();
|
|
1140
|
|
1141 bool inc_counter = UseCompiler || CountCompiledCalls;
|
|
1142
|
|
1143 // the following temporary registers are used during frame creation
|
|
1144 const Register Gtmp1 = G3_scratch ;
|
|
1145 const Register Gtmp2 = G1_scratch;
|
|
1146
|
|
1147 // make sure registers are different!
|
|
1148 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
|
|
1149
|
|
1150 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
|
|
1151 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
|
|
1152 // Seems like G5_method is live at the point this is used. So we could make this look consistent
|
|
1153 // and use in the asserts.
|
|
1154 const Address access_flags (Lmethod, 0, in_bytes(methodOopDesc::access_flags_offset()));
|
|
1155
|
|
1156 __ verify_oop(G5_method);
|
|
1157
|
|
1158 const Register Glocals_size = G3;
|
|
1159 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
|
|
1160
|
|
1161 // make sure method is not native & not abstract
|
|
1162 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
|
|
1163 #ifdef ASSERT
|
|
1164 __ ld(G5_method, in_bytes(methodOopDesc::access_flags_offset()), Gtmp1);
|
|
1165 {
|
|
1166 Label L;
|
|
1167 __ btst(JVM_ACC_NATIVE, Gtmp1);
|
|
1168 __ br(Assembler::zero, false, Assembler::pt, L);
|
|
1169 __ delayed()->nop();
|
|
1170 __ stop("tried to execute native method as non-native");
|
|
1171 __ bind(L);
|
|
1172 }
|
|
1173 { Label L;
|
|
1174 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
|
|
1175 __ br(Assembler::zero, false, Assembler::pt, L);
|
|
1176 __ delayed()->nop();
|
|
1177 __ stop("tried to execute abstract method as non-abstract");
|
|
1178 __ bind(L);
|
|
1179 }
|
|
1180 #endif // ASSERT
|
|
1181
|
|
1182 // generate the code to allocate the interpreter stack frame
|
|
1183
|
|
1184 generate_fixed_frame(false);
|
|
1185
|
|
1186 #ifdef FAST_DISPATCH
|
|
1187 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
|
|
1188 // set bytecode dispatch table base
|
|
1189 #endif
|
|
1190
|
|
1191 //
|
|
1192 // Code to initialize the extra (i.e. non-parm) locals
|
|
1193 //
|
|
1194 Register init_value = noreg; // will be G0 if we must clear locals
|
|
1195 // The way the code was setup before zerolocals was always true for vanilla java entries.
|
|
1196 // It could only be false for the specialized entries like accessor or empty which have
|
|
1197 // no extra locals so the testing was a waste of time and the extra locals were always
|
|
1198 // initialized. We removed this extra complication to already over complicated code.
|
|
1199
|
|
1200 init_value = G0;
|
|
1201 Label clear_loop;
|
|
1202
|
|
1203 // NOTE: If you change the frame layout, this code will need to
|
|
1204 // be updated!
|
|
1205 __ lduh( size_of_locals, O2 );
|
|
1206 __ lduh( size_of_parameters, O1 );
|
|
1207 __ sll( O2, Interpreter::logStackElementSize(), O2);
|
|
1208 __ sll( O1, Interpreter::logStackElementSize(), O1 );
|
|
1209 __ sub( Llocals, O2, O2 );
|
|
1210 __ sub( Llocals, O1, O1 );
|
|
1211
|
|
1212 __ bind( clear_loop );
|
|
1213 __ inc( O2, wordSize );
|
|
1214
|
|
1215 __ cmp( O2, O1 );
|
|
1216 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
|
|
1217 __ delayed()->st_ptr( init_value, O2, 0 );
|
|
1218
|
|
1219 const Address do_not_unlock_if_synchronized(G2_thread, 0,
|
|
1220 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
|
|
1221 // Since at this point in the method invocation the exception handler
|
|
1222 // would try to exit the monitor of synchronized methods which hasn't
|
|
1223 // been entered yet, we set the thread local variable
|
|
1224 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
|
|
1225 // runtime, exception handling i.e. unlock_if_synchronized_method will
|
|
1226 // check this thread local flag.
|
|
1227 __ movbool(true, G3_scratch);
|
|
1228 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
|
|
1229
|
|
1230 // increment invocation counter and check for overflow
|
|
1231 //
|
|
1232 // Note: checking for negative value instead of overflow
|
|
1233 // so we have a 'sticky' overflow test (may be of
|
|
1234 // importance as soon as we have true MT/MP)
|
|
1235 Label invocation_counter_overflow;
|
|
1236 Label profile_method;
|
|
1237 Label profile_method_continue;
|
|
1238 Label Lcontinue;
|
|
1239 if (inc_counter) {
|
|
1240 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
|
|
1241 if (ProfileInterpreter) {
|
|
1242 __ bind(profile_method_continue);
|
|
1243 }
|
|
1244 }
|
|
1245 __ bind(Lcontinue);
|
|
1246
|
|
1247 bang_stack_shadow_pages(false);
|
|
1248
|
|
1249 // reset the _do_not_unlock_if_synchronized flag
|
|
1250 __ stbool(G0, do_not_unlock_if_synchronized);
|
|
1251
|
|
1252 // check for synchronized methods
|
|
1253 // Must happen AFTER invocation_counter check and stack overflow check,
|
|
1254 // so method is not locked if overflows.
|
|
1255
|
|
1256 if (synchronized) {
|
|
1257 lock_method();
|
|
1258 } else {
|
|
1259 #ifdef ASSERT
|
|
1260 { Label ok;
|
|
1261 __ ld(access_flags, O0);
|
|
1262 __ btst(JVM_ACC_SYNCHRONIZED, O0);
|
|
1263 __ br( Assembler::zero, false, Assembler::pt, ok);
|
|
1264 __ delayed()->nop();
|
|
1265 __ stop("method needs synchronization");
|
|
1266 __ bind(ok);
|
|
1267 }
|
|
1268 #endif // ASSERT
|
|
1269 }
|
|
1270
|
|
1271 // start execution
|
|
1272
|
|
1273 __ verify_thread();
|
|
1274
|
|
1275 // jvmti support
|
|
1276 __ notify_method_entry();
|
|
1277
|
|
1278 // start executing instructions
|
|
1279 __ dispatch_next(vtos);
|
|
1280
|
|
1281
|
|
1282 if (inc_counter) {
|
|
1283 if (ProfileInterpreter) {
|
|
1284 // We have decided to profile this method in the interpreter
|
|
1285 __ bind(profile_method);
|
|
1286
|
|
1287 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method), Lbcp, true);
|
|
1288
|
|
1289 #ifdef ASSERT
|
|
1290 __ tst(O0);
|
|
1291 __ breakpoint_trap(Assembler::notEqual);
|
|
1292 #endif
|
|
1293
|
|
1294 __ set_method_data_pointer();
|
|
1295
|
|
1296 __ ba(false, profile_method_continue);
|
|
1297 __ delayed()->nop();
|
|
1298 }
|
|
1299
|
|
1300 // handle invocation counter overflow
|
|
1301 __ bind(invocation_counter_overflow);
|
|
1302 generate_counter_overflow(Lcontinue);
|
|
1303 }
|
|
1304
|
|
1305
|
|
1306 return entry;
|
|
1307 }
|
|
1308
|
|
1309
|
|
1310 //----------------------------------------------------------------------------------------------------
|
|
1311 // Entry points & stack frame layout
|
|
1312 //
|
|
1313 // Here we generate the various kind of entries into the interpreter.
|
|
1314 // The two main entry type are generic bytecode methods and native call method.
|
|
1315 // These both come in synchronized and non-synchronized versions but the
|
|
1316 // frame layout they create is very similar. The other method entry
|
|
1317 // types are really just special purpose entries that are really entry
|
|
1318 // and interpretation all in one. These are for trivial methods like
|
|
1319 // accessor, empty, or special math methods.
|
|
1320 //
|
|
1321 // When control flow reaches any of the entry types for the interpreter
|
|
1322 // the following holds ->
|
|
1323 //
|
|
1324 // C2 Calling Conventions:
|
|
1325 //
|
|
1326 // The entry code below assumes that the following registers are set
|
|
1327 // when coming in:
|
|
1328 // G5_method: holds the methodOop of the method to call
|
|
1329 // Lesp: points to the TOS of the callers expression stack
|
|
1330 // after having pushed all the parameters
|
|
1331 //
|
|
1332 // The entry code does the following to setup an interpreter frame
|
|
1333 // pop parameters from the callers stack by adjusting Lesp
|
|
1334 // set O0 to Lesp
|
|
1335 // compute X = (max_locals - num_parameters)
|
|
1336 // bump SP up by X to accomadate the extra locals
|
|
1337 // compute X = max_expression_stack
|
|
1338 // + vm_local_words
|
|
1339 // + 16 words of register save area
|
|
1340 // save frame doing a save sp, -X, sp growing towards lower addresses
|
|
1341 // set Lbcp, Lmethod, LcpoolCache
|
|
1342 // set Llocals to i0
|
|
1343 // set Lmonitors to FP - rounded_vm_local_words
|
|
1344 // set Lesp to Lmonitors - 4
|
|
1345 //
|
|
1346 // The frame has now been setup to do the rest of the entry code
|
|
1347
|
|
1348 // Try this optimization: Most method entries could live in a
|
|
1349 // "one size fits all" stack frame without all the dynamic size
|
|
1350 // calculations. It might be profitable to do all this calculation
|
|
1351 // statically and approximately for "small enough" methods.
|
|
1352
|
|
1353 //-----------------------------------------------------------------------------------------------
|
|
1354
|
|
1355 // C1 Calling conventions
|
|
1356 //
|
|
1357 // Upon method entry, the following registers are setup:
|
|
1358 //
|
|
1359 // g2 G2_thread: current thread
|
|
1360 // g5 G5_method: method to activate
|
|
1361 // g4 Gargs : pointer to last argument
|
|
1362 //
|
|
1363 //
|
|
1364 // Stack:
|
|
1365 //
|
|
1366 // +---------------+ <--- sp
|
|
1367 // | |
|
|
1368 // : reg save area :
|
|
1369 // | |
|
|
1370 // +---------------+ <--- sp + 0x40
|
|
1371 // | |
|
|
1372 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
|
|
1373 // | |
|
|
1374 // +---------------+ <--- sp + 0x5c
|
|
1375 // | |
|
|
1376 // : free :
|
|
1377 // | |
|
|
1378 // +---------------+ <--- Gargs
|
|
1379 // | |
|
|
1380 // : arguments :
|
|
1381 // | |
|
|
1382 // +---------------+
|
|
1383 // | |
|
|
1384 //
|
|
1385 //
|
|
1386 //
|
|
1387 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
|
|
1388 //
|
|
1389 // +---------------+ <--- sp
|
|
1390 // | |
|
|
1391 // : reg save area :
|
|
1392 // | |
|
|
1393 // +---------------+ <--- sp + 0x40
|
|
1394 // | |
|
|
1395 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
|
|
1396 // | |
|
|
1397 // +---------------+ <--- sp + 0x5c
|
|
1398 // | |
|
|
1399 // : :
|
|
1400 // | | <--- Lesp
|
|
1401 // +---------------+ <--- Lmonitors (fp - 0x18)
|
|
1402 // | VM locals |
|
|
1403 // +---------------+ <--- fp
|
|
1404 // | |
|
|
1405 // : reg save area :
|
|
1406 // | |
|
|
1407 // +---------------+ <--- fp + 0x40
|
|
1408 // | |
|
|
1409 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
|
|
1410 // | |
|
|
1411 // +---------------+ <--- fp + 0x5c
|
|
1412 // | |
|
|
1413 // : free :
|
|
1414 // | |
|
|
1415 // +---------------+
|
|
1416 // | |
|
|
1417 // : nonarg locals :
|
|
1418 // | |
|
|
1419 // +---------------+
|
|
1420 // | |
|
|
1421 // : arguments :
|
|
1422 // | | <--- Llocals
|
|
1423 // +---------------+ <--- Gargs
|
|
1424 // | |
|
|
1425
|
|
1426 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
|
|
1427
|
|
1428 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
|
|
1429 // expression stack, the callee will have callee_extra_locals (so we can account for
|
|
1430 // frame extension) and monitor_size for monitors. Basically we need to calculate
|
|
1431 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
|
|
1432 //
|
|
1433 //
|
|
1434 // The big complicating thing here is that we must ensure that the stack stays properly
|
|
1435 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
|
|
1436 // needs to be aligned for). We are given that the sp (fp) is already aligned by
|
|
1437 // the caller so we must ensure that it is properly aligned for our callee.
|
|
1438 //
|
|
1439 const int rounded_vm_local_words =
|
|
1440 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
|
|
1441 // callee_locals and max_stack are counts, not the size in frame.
|
|
1442 const int locals_size =
|
|
1443 round_to(callee_extra_locals * Interpreter::stackElementWords(), WordsPerLong);
|
|
1444 const int max_stack_words = max_stack * Interpreter::stackElementWords();
|
|
1445 return (round_to((max_stack_words
|
|
1446 + rounded_vm_local_words
|
|
1447 + frame::memory_parameter_word_sp_offset), WordsPerLong)
|
|
1448 // already rounded
|
|
1449 + locals_size + monitor_size);
|
|
1450 }
|
|
1451
|
|
1452 // How much stack a method top interpreter activation needs in words.
|
|
1453 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
|
|
1454
|
|
1455 // See call_stub code
|
|
1456 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
|
|
1457 WordsPerLong); // 7 + register save area
|
|
1458
|
|
1459 // Save space for one monitor to get into the interpreted method in case
|
|
1460 // the method is synchronized
|
|
1461 int monitor_size = method->is_synchronized() ?
|
|
1462 1*frame::interpreter_frame_monitor_size() : 0;
|
|
1463 return size_activation_helper(method->max_locals(), method->max_stack(),
|
|
1464 monitor_size) + call_stub_size;
|
|
1465 }
|
|
1466
|
|
1467 int AbstractInterpreter::layout_activation(methodOop method,
|
|
1468 int tempcount,
|
|
1469 int popframe_extra_args,
|
|
1470 int moncount,
|
|
1471 int callee_param_count,
|
|
1472 int callee_local_count,
|
|
1473 frame* caller,
|
|
1474 frame* interpreter_frame,
|
|
1475 bool is_top_frame) {
|
|
1476 // Note: This calculation must exactly parallel the frame setup
|
|
1477 // in InterpreterGenerator::generate_fixed_frame.
|
|
1478 // If f!=NULL, set up the following variables:
|
|
1479 // - Lmethod
|
|
1480 // - Llocals
|
|
1481 // - Lmonitors (to the indicated number of monitors)
|
|
1482 // - Lesp (to the indicated number of temps)
|
|
1483 // The frame f (if not NULL) on entry is a description of the caller of the frame
|
|
1484 // we are about to layout. We are guaranteed that we will be able to fill in a
|
|
1485 // new interpreter frame as its callee (i.e. the stack space is allocated and
|
|
1486 // the amount was determined by an earlier call to this method with f == NULL).
|
|
1487 // On return f (if not NULL) while describe the interpreter frame we just layed out.
|
|
1488
|
|
1489 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
|
|
1490 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
|
|
1491
|
|
1492 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
|
|
1493 //
|
|
1494 // Note: if you look closely this appears to be doing something much different
|
|
1495 // than generate_fixed_frame. What is happening is this. On sparc we have to do
|
|
1496 // this dance with interpreter_sp_adjustment because the window save area would
|
|
1497 // appear just below the bottom (tos) of the caller's java expression stack. Because
|
|
1498 // the interpreter want to have the locals completely contiguous generate_fixed_frame
|
|
1499 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
|
|
1500 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
|
|
1501 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
|
|
1502 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
|
|
1503 // because the oldest frame would have adjust its callers frame and yet that frame
|
|
1504 // already exists and isn't part of this array of frames we are unpacking. So at first
|
|
1505 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
|
|
1506 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
|
|
1507 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
|
|
1508 // add up. It does seem like it simpler to account for the adjustment here (and remove the
|
|
1509 // callee... parameters here). However this would mean that this routine would have to take
|
|
1510 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
|
|
1511 // and run the calling loop in the reverse order. This would also would appear to mean making
|
|
1512 // this code aware of what the interactions are when that initial caller fram was an osr or
|
|
1513 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
|
|
1514 // there is no sense in messing working code.
|
|
1515 //
|
|
1516
|
|
1517 int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
|
|
1518 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
|
|
1519
|
|
1520 int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
|
|
1521 monitor_size);
|
|
1522
|
|
1523 if (interpreter_frame != NULL) {
|
|
1524 // The skeleton frame must already look like an interpreter frame
|
|
1525 // even if not fully filled out.
|
|
1526 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
|
|
1527
|
|
1528 intptr_t* fp = interpreter_frame->fp();
|
|
1529
|
|
1530 JavaThread* thread = JavaThread::current();
|
|
1531 RegisterMap map(thread, false);
|
|
1532 // More verification that skeleton frame is properly walkable
|
|
1533 assert(fp == caller->sp(), "fp must match");
|
|
1534
|
|
1535 intptr_t* montop = fp - rounded_vm_local_words;
|
|
1536
|
|
1537 // preallocate monitors (cf. __ add_monitor_to_stack)
|
|
1538 intptr_t* monitors = montop - monitor_size;
|
|
1539
|
|
1540 // preallocate stack space
|
|
1541 intptr_t* esp = monitors - 1 -
|
|
1542 (tempcount * Interpreter::stackElementWords()) -
|
|
1543 popframe_extra_args;
|
|
1544
|
|
1545 int local_words = method->max_locals() * Interpreter::stackElementWords();
|
|
1546 int parm_words = method->size_of_parameters() * Interpreter::stackElementWords();
|
|
1547 NEEDS_CLEANUP;
|
|
1548 intptr_t* locals;
|
|
1549 if (caller->is_interpreted_frame()) {
|
|
1550 // Can force the locals area to end up properly overlapping the top of the expression stack.
|
|
1551 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
|
|
1552 // Note that this computation means we replace size_of_parameters() values from the caller
|
|
1553 // interpreter frame's expression stack with our argument locals
|
|
1554 locals = Lesp_ptr + parm_words;
|
|
1555 int delta = local_words - parm_words;
|
|
1556 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
|
|
1557 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
|
|
1558 } else {
|
|
1559 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
|
|
1560 // Don't have Lesp available; lay out locals block in the caller
|
|
1561 // adjacent to the register window save area.
|
|
1562 //
|
|
1563 // Compiled frames do not allocate a varargs area which is why this if
|
|
1564 // statement is needed.
|
|
1565 //
|
|
1566 if (caller->is_compiled_frame()) {
|
|
1567 locals = fp + frame::register_save_words + local_words - 1;
|
|
1568 } else {
|
|
1569 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
|
|
1570 }
|
|
1571 if (!caller->is_entry_frame()) {
|
|
1572 // Caller wants his own SP back
|
|
1573 int caller_frame_size = caller->cb()->frame_size();
|
|
1574 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
|
|
1575 }
|
|
1576 }
|
|
1577 if (TraceDeoptimization) {
|
|
1578 if (caller->is_entry_frame()) {
|
|
1579 // make sure I5_savedSP and the entry frames notion of saved SP
|
|
1580 // agree. This assertion duplicate a check in entry frame code
|
|
1581 // but catches the failure earlier.
|
|
1582 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
|
|
1583 "would change callers SP");
|
|
1584 }
|
|
1585 if (caller->is_entry_frame()) {
|
|
1586 tty->print("entry ");
|
|
1587 }
|
|
1588 if (caller->is_compiled_frame()) {
|
|
1589 tty->print("compiled ");
|
|
1590 if (caller->is_deoptimized_frame()) {
|
|
1591 tty->print("(deopt) ");
|
|
1592 }
|
|
1593 }
|
|
1594 if (caller->is_interpreted_frame()) {
|
|
1595 tty->print("interpreted ");
|
|
1596 }
|
|
1597 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
|
|
1598 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
|
|
1599 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
|
|
1600 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
|
|
1601 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
|
|
1602 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
|
|
1603 tty->print_cr("Llocals = 0x%x", locals);
|
|
1604 tty->print_cr("Lesp = 0x%x", esp);
|
|
1605 tty->print_cr("Lmonitors = 0x%x", monitors);
|
|
1606 }
|
|
1607
|
|
1608 if (method->max_locals() > 0) {
|
|
1609 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
|
|
1610 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
|
|
1611 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
|
|
1612 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
|
|
1613 }
|
|
1614 #ifdef _LP64
|
|
1615 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
|
|
1616 #endif
|
|
1617
|
|
1618 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
|
|
1619 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
|
|
1620 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
|
|
1621 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
|
|
1622 // Llast_SP will be same as SP as there is no adapter space
|
|
1623 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
|
|
1624 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
|
|
1625 #ifdef FAST_DISPATCH
|
|
1626 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
|
|
1627 #endif
|
|
1628
|
|
1629
|
|
1630 #ifdef ASSERT
|
|
1631 BasicObjectLock* mp = (BasicObjectLock*)monitors;
|
|
1632
|
|
1633 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
|
|
1634 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize())+Interpreter::value_offset_in_bytes()), "locals match");
|
|
1635 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
|
|
1636 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
|
|
1637 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
|
|
1638
|
|
1639 // check bounds
|
|
1640 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
|
|
1641 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
|
|
1642 assert(lo < monitors && montop <= hi, "monitors in bounds");
|
|
1643 assert(lo <= esp && esp < monitors, "esp in bounds");
|
|
1644 #endif // ASSERT
|
|
1645 }
|
|
1646
|
|
1647 return raw_frame_size;
|
|
1648 }
|
|
1649
|
|
1650 //----------------------------------------------------------------------------------------------------
|
|
1651 // Exceptions
|
|
1652 void TemplateInterpreterGenerator::generate_throw_exception() {
|
|
1653
|
|
1654 // Entry point in previous activation (i.e., if the caller was interpreted)
|
|
1655 Interpreter::_rethrow_exception_entry = __ pc();
|
|
1656 // O0: exception
|
|
1657
|
|
1658 // entry point for exceptions thrown within interpreter code
|
|
1659 Interpreter::_throw_exception_entry = __ pc();
|
|
1660 __ verify_thread();
|
|
1661 // expression stack is undefined here
|
|
1662 // O0: exception, i.e. Oexception
|
|
1663 // Lbcp: exception bcx
|
|
1664 __ verify_oop(Oexception);
|
|
1665
|
|
1666
|
|
1667 // expression stack must be empty before entering the VM in case of an exception
|
|
1668 __ empty_expression_stack();
|
|
1669 // find exception handler address and preserve exception oop
|
|
1670 // call C routine to find handler and jump to it
|
|
1671 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
|
|
1672 __ push_ptr(O1); // push exception for exception handler bytecodes
|
|
1673
|
|
1674 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
|
|
1675 __ delayed()->nop();
|
|
1676
|
|
1677
|
|
1678 // if the exception is not handled in the current frame
|
|
1679 // the frame is removed and the exception is rethrown
|
|
1680 // (i.e. exception continuation is _rethrow_exception)
|
|
1681 //
|
|
1682 // Note: At this point the bci is still the bxi for the instruction which caused
|
|
1683 // the exception and the expression stack is empty. Thus, for any VM calls
|
|
1684 // at this point, GC will find a legal oop map (with empty expression stack).
|
|
1685
|
|
1686 // in current activation
|
|
1687 // tos: exception
|
|
1688 // Lbcp: exception bcp
|
|
1689
|
|
1690 //
|
|
1691 // JVMTI PopFrame support
|
|
1692 //
|
|
1693
|
|
1694 Interpreter::_remove_activation_preserving_args_entry = __ pc();
|
|
1695 Address popframe_condition_addr (G2_thread, 0, in_bytes(JavaThread::popframe_condition_offset()));
|
|
1696 // Set the popframe_processing bit in popframe_condition indicating that we are
|
|
1697 // currently handling popframe, so that call_VMs that may happen later do not trigger new
|
|
1698 // popframe handling cycles.
|
|
1699
|
|
1700 __ ld(popframe_condition_addr, G3_scratch);
|
|
1701 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
|
|
1702 __ stw(G3_scratch, popframe_condition_addr);
|
|
1703
|
|
1704 // Empty the expression stack, as in normal exception handling
|
|
1705 __ empty_expression_stack();
|
|
1706 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
|
|
1707
|
|
1708 {
|
|
1709 // Check to see whether we are returning to a deoptimized frame.
|
|
1710 // (The PopFrame call ensures that the caller of the popped frame is
|
|
1711 // either interpreted or compiled and deoptimizes it if compiled.)
|
|
1712 // In this case, we can't call dispatch_next() after the frame is
|
|
1713 // popped, but instead must save the incoming arguments and restore
|
|
1714 // them after deoptimization has occurred.
|
|
1715 //
|
|
1716 // Note that we don't compare the return PC against the
|
|
1717 // deoptimization blob's unpack entry because of the presence of
|
|
1718 // adapter frames in C2.
|
|
1719 Label caller_not_deoptimized;
|
|
1720 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
|
|
1721 __ tst(O0);
|
|
1722 __ brx(Assembler::notEqual, false, Assembler::pt, caller_not_deoptimized);
|
|
1723 __ delayed()->nop();
|
|
1724
|
|
1725 const Register Gtmp1 = G3_scratch;
|
|
1726 const Register Gtmp2 = G1_scratch;
|
|
1727
|
|
1728 // Compute size of arguments for saving when returning to deoptimized caller
|
|
1729 __ lduh(Lmethod, in_bytes(methodOopDesc::size_of_parameters_offset()), Gtmp1);
|
|
1730 __ sll(Gtmp1, Interpreter::logStackElementSize(), Gtmp1);
|
|
1731 __ sub(Llocals, Gtmp1, Gtmp2);
|
|
1732 __ add(Gtmp2, wordSize, Gtmp2);
|
|
1733 // Save these arguments
|
|
1734 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
|
|
1735 // Inform deoptimization that it is responsible for restoring these arguments
|
|
1736 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
|
|
1737 Address popframe_condition_addr(G2_thread, 0, in_bytes(JavaThread::popframe_condition_offset()));
|
|
1738 __ st(Gtmp1, popframe_condition_addr);
|
|
1739
|
|
1740 // Return from the current method
|
|
1741 // The caller's SP was adjusted upon method entry to accomodate
|
|
1742 // the callee's non-argument locals. Undo that adjustment.
|
|
1743 __ ret();
|
|
1744 __ delayed()->restore(I5_savedSP, G0, SP);
|
|
1745
|
|
1746 __ bind(caller_not_deoptimized);
|
|
1747 }
|
|
1748
|
|
1749 // Clear the popframe condition flag
|
|
1750 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
|
|
1751
|
|
1752 // Get out of the current method (how this is done depends on the particular compiler calling
|
|
1753 // convention that the interpreter currently follows)
|
|
1754 // The caller's SP was adjusted upon method entry to accomodate
|
|
1755 // the callee's non-argument locals. Undo that adjustment.
|
|
1756 __ restore(I5_savedSP, G0, SP);
|
|
1757 // The method data pointer was incremented already during
|
|
1758 // call profiling. We have to restore the mdp for the current bcp.
|
|
1759 if (ProfileInterpreter) {
|
|
1760 __ set_method_data_pointer_for_bcp();
|
|
1761 }
|
|
1762 // Resume bytecode interpretation at the current bcp
|
|
1763 __ dispatch_next(vtos);
|
|
1764 // end of JVMTI PopFrame support
|
|
1765
|
|
1766 Interpreter::_remove_activation_entry = __ pc();
|
|
1767
|
|
1768 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
|
|
1769 __ pop_ptr(Oexception); // get exception
|
|
1770
|
|
1771 // Intel has the following comment:
|
|
1772 //// remove the activation (without doing throws on illegalMonitorExceptions)
|
|
1773 // They remove the activation without checking for bad monitor state.
|
|
1774 // %%% We should make sure this is the right semantics before implementing.
|
|
1775
|
|
1776 // %%% changed set_vm_result_2 to set_vm_result and get_vm_result_2 to get_vm_result. Is there a bug here?
|
|
1777 __ set_vm_result(Oexception);
|
|
1778 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
|
|
1779
|
|
1780 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
|
|
1781
|
|
1782 __ get_vm_result(Oexception);
|
|
1783 __ verify_oop(Oexception);
|
|
1784
|
|
1785 const int return_reg_adjustment = frame::pc_return_offset;
|
|
1786 Address issuing_pc_addr(I7, 0, return_reg_adjustment);
|
|
1787
|
|
1788 // We are done with this activation frame; find out where to go next.
|
|
1789 // The continuation point will be an exception handler, which expects
|
|
1790 // the following registers set up:
|
|
1791 //
|
|
1792 // Oexception: exception
|
|
1793 // Oissuing_pc: the local call that threw exception
|
|
1794 // Other On: garbage
|
|
1795 // In/Ln: the contents of the caller's register window
|
|
1796 //
|
|
1797 // We do the required restore at the last possible moment, because we
|
|
1798 // need to preserve some state across a runtime call.
|
|
1799 // (Remember that the caller activation is unknown--it might not be
|
|
1800 // interpreted, so things like Lscratch are useless in the caller.)
|
|
1801
|
|
1802 // Although the Intel version uses call_C, we can use the more
|
|
1803 // compact call_VM. (The only real difference on SPARC is a
|
|
1804 // harmlessly ignored [re]set_last_Java_frame, compared with
|
|
1805 // the Intel code which lacks this.)
|
|
1806 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
|
|
1807 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
|
|
1808 __ super_call_VM_leaf(L7_thread_cache,
|
|
1809 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
|
|
1810 Oissuing_pc->after_save());
|
|
1811
|
|
1812 // The caller's SP was adjusted upon method entry to accomodate
|
|
1813 // the callee's non-argument locals. Undo that adjustment.
|
|
1814 __ JMP(O0, 0); // return exception handler in caller
|
|
1815 __ delayed()->restore(I5_savedSP, G0, SP);
|
|
1816
|
|
1817 // (same old exception object is already in Oexception; see above)
|
|
1818 // Note that an "issuing PC" is actually the next PC after the call
|
|
1819 }
|
|
1820
|
|
1821
|
|
1822 //
|
|
1823 // JVMTI ForceEarlyReturn support
|
|
1824 //
|
|
1825
|
|
1826 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
|
|
1827 address entry = __ pc();
|
|
1828
|
|
1829 __ empty_expression_stack();
|
|
1830 __ load_earlyret_value(state);
|
|
1831
|
|
1832 __ ld_ptr(Address(G2_thread, 0, in_bytes(JavaThread::jvmti_thread_state_offset())), G3_scratch);
|
|
1833 Address cond_addr(G3_scratch, 0, in_bytes(JvmtiThreadState::earlyret_state_offset()));
|
|
1834
|
|
1835 // Clear the earlyret state
|
|
1836 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
|
|
1837
|
|
1838 __ remove_activation(state,
|
|
1839 /* throw_monitor_exception */ false,
|
|
1840 /* install_monitor_exception */ false);
|
|
1841
|
|
1842 // The caller's SP was adjusted upon method entry to accomodate
|
|
1843 // the callee's non-argument locals. Undo that adjustment.
|
|
1844 __ ret(); // return to caller
|
|
1845 __ delayed()->restore(I5_savedSP, G0, SP);
|
|
1846
|
|
1847 return entry;
|
|
1848 } // end of JVMTI ForceEarlyReturn support
|
|
1849
|
|
1850
|
|
1851 //------------------------------------------------------------------------------------------------------------------------
|
|
1852 // Helper for vtos entry point generation
|
|
1853
|
|
1854 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) {
|
|
1855 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
|
|
1856 Label L;
|
|
1857 aep = __ pc(); __ push_ptr(); __ ba(false, L); __ delayed()->nop();
|
|
1858 fep = __ pc(); __ push_f(); __ ba(false, L); __ delayed()->nop();
|
|
1859 dep = __ pc(); __ push_d(); __ ba(false, L); __ delayed()->nop();
|
|
1860 lep = __ pc(); __ push_l(); __ ba(false, L); __ delayed()->nop();
|
|
1861 iep = __ pc(); __ push_i();
|
|
1862 bep = cep = sep = iep; // there aren't any
|
|
1863 vep = __ pc(); __ bind(L); // fall through
|
|
1864 generate_and_dispatch(t);
|
|
1865 }
|
|
1866
|
|
1867 // --------------------------------------------------------------------------------
|
|
1868
|
|
1869
|
|
1870 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
|
|
1871 : TemplateInterpreterGenerator(code) {
|
|
1872 generate_all(); // down here so it can be "virtual"
|
|
1873 }
|
|
1874
|
|
1875 // --------------------------------------------------------------------------------
|
|
1876
|
|
1877 // Non-product code
|
|
1878 #ifndef PRODUCT
|
|
1879 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
|
|
1880 address entry = __ pc();
|
|
1881
|
|
1882 __ push(state);
|
|
1883 __ mov(O7, Lscratch); // protect return address within interpreter
|
|
1884
|
|
1885 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
|
|
1886 __ mov( Otos_l2, G3_scratch );
|
|
1887 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
|
|
1888 __ mov(Lscratch, O7); // restore return address
|
|
1889 __ pop(state);
|
|
1890 __ retl();
|
|
1891 __ delayed()->nop();
|
|
1892
|
|
1893 return entry;
|
|
1894 }
|
|
1895
|
|
1896
|
|
1897 // helpers for generate_and_dispatch
|
|
1898
|
|
1899 void TemplateInterpreterGenerator::count_bytecode() {
|
|
1900 Address c(G3_scratch, (address)&BytecodeCounter::_counter_value);
|
|
1901 __ load_contents(c, G4_scratch);
|
|
1902 __ inc(G4_scratch);
|
|
1903 __ st(G4_scratch, c);
|
|
1904 }
|
|
1905
|
|
1906
|
|
1907 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
|
|
1908 Address bucket( G3_scratch, (address) &BytecodeHistogram::_counters[t->bytecode()] );
|
|
1909 __ load_contents(bucket, G4_scratch);
|
|
1910 __ inc(G4_scratch);
|
|
1911 __ st(G4_scratch, bucket);
|
|
1912 }
|
|
1913
|
|
1914
|
|
1915 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
|
|
1916 address index_addr = (address)&BytecodePairHistogram::_index;
|
|
1917 Address index(G3_scratch, index_addr);
|
|
1918
|
|
1919 address counters_addr = (address)&BytecodePairHistogram::_counters;
|
|
1920 Address counters(G3_scratch, counters_addr);
|
|
1921
|
|
1922 // get index, shift out old bytecode, bring in new bytecode, and store it
|
|
1923 // _index = (_index >> log2_number_of_codes) |
|
|
1924 // (bytecode << log2_number_of_codes);
|
|
1925
|
|
1926
|
|
1927 __ load_contents( index, G4_scratch );
|
|
1928 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
|
|
1929 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
|
|
1930 __ or3( G3_scratch, G4_scratch, G4_scratch );
|
|
1931 __ store_contents( G4_scratch, index );
|
|
1932
|
|
1933 // bump bucket contents
|
|
1934 // _counters[_index] ++;
|
|
1935
|
|
1936 __ load_address( counters ); // loads into G3_scratch
|
|
1937 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
|
|
1938 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
|
|
1939 __ ld (G3_scratch, 0, G4_scratch);
|
|
1940 __ inc (G4_scratch);
|
|
1941 __ st (G4_scratch, 0, G3_scratch);
|
|
1942 }
|
|
1943
|
|
1944
|
|
1945 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
|
|
1946 // Call a little run-time stub to avoid blow-up for each bytecode.
|
|
1947 // The run-time runtime saves the right registers, depending on
|
|
1948 // the tosca in-state for the given template.
|
|
1949 address entry = Interpreter::trace_code(t->tos_in());
|
|
1950 guarantee(entry != NULL, "entry must have been generated");
|
|
1951 __ call(entry, relocInfo::none);
|
|
1952 __ delayed()->nop();
|
|
1953 }
|
|
1954
|
|
1955
|
|
1956 void TemplateInterpreterGenerator::stop_interpreter_at() {
|
|
1957 Address counter(G3_scratch , (address)&BytecodeCounter::_counter_value);
|
|
1958 __ load_contents (counter, G3_scratch );
|
|
1959 Address stop_at(G4_scratch, (address)&StopInterpreterAt);
|
|
1960 __ load_ptr_contents(stop_at, G4_scratch);
|
|
1961 __ cmp(G3_scratch, G4_scratch);
|
|
1962 __ breakpoint_trap(Assembler::equal);
|
|
1963 }
|
|
1964 #endif // not PRODUCT
|
|
1965 #endif // !CC_INTERP
|