Print this page
Split |
Close |
Expand all |
Collapse all |
--- old/src/cpu/sparc/vm/templateInterpreter_sparc.cpp
+++ new/src/cpu/sparc/vm/templateInterpreter_sparc.cpp
1 1 /*
2 2 * Copyright 1997-2010 Sun Microsystems, Inc. All Rights Reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 21 * have any questions.
22 22 *
23 23 */
24 24
25 25 #include "incls/_precompiled.incl"
26 26 #include "incls/_templateInterpreter_sparc.cpp.incl"
27 27
28 28 #ifndef CC_INTERP
29 29 #ifndef FAST_DISPATCH
30 30 #define FAST_DISPATCH 1
31 31 #endif
32 32 #undef FAST_DISPATCH
33 33
34 34
35 35 // Generation of Interpreter
36 36 //
37 37 // The InterpreterGenerator generates the interpreter into Interpreter::_code.
38 38
39 39
40 40 #define __ _masm->
41 41
42 42
43 43 //----------------------------------------------------------------------------------------------------
44 44
45 45
46 46 void InterpreterGenerator::save_native_result(void) {
47 47 // result potentially in O0/O1: save it across calls
48 48 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
49 49
50 50 // result potentially in F0/F1: save it across calls
51 51 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
52 52
53 53 // save and restore any potential method result value around the unlocking operation
54 54 __ stf(FloatRegisterImpl::D, F0, d_tmp);
55 55 #ifdef _LP64
56 56 __ stx(O0, l_tmp);
57 57 #else
58 58 __ std(O0, l_tmp);
59 59 #endif
60 60 }
61 61
62 62 void InterpreterGenerator::restore_native_result(void) {
63 63 const Address& l_tmp = InterpreterMacroAssembler::l_tmp;
64 64 const Address& d_tmp = InterpreterMacroAssembler::d_tmp;
65 65
66 66 // Restore any method result value
67 67 __ ldf(FloatRegisterImpl::D, d_tmp, F0);
68 68 #ifdef _LP64
69 69 __ ldx(l_tmp, O0);
70 70 #else
71 71 __ ldd(l_tmp, O0);
72 72 #endif
73 73 }
74 74
75 75 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
76 76 assert(!pass_oop || message == NULL, "either oop or message but not both");
77 77 address entry = __ pc();
78 78 // expression stack must be empty before entering the VM if an exception happened
79 79 __ empty_expression_stack();
80 80 // load exception object
81 81 __ set((intptr_t)name, G3_scratch);
82 82 if (pass_oop) {
83 83 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), G3_scratch, Otos_i);
84 84 } else {
85 85 __ set((intptr_t)message, G4_scratch);
86 86 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), G3_scratch, G4_scratch);
87 87 }
88 88 // throw exception
89 89 assert(Interpreter::throw_exception_entry() != NULL, "generate it first");
90 90 AddressLiteral thrower(Interpreter::throw_exception_entry());
91 91 __ jump_to(thrower, G3_scratch);
92 92 __ delayed()->nop();
93 93 return entry;
94 94 }
95 95
96 96 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
97 97 address entry = __ pc();
98 98 // expression stack must be empty before entering the VM if an exception
99 99 // happened
100 100 __ empty_expression_stack();
101 101 // load exception object
102 102 __ call_VM(Oexception,
103 103 CAST_FROM_FN_PTR(address,
104 104 InterpreterRuntime::throw_ClassCastException),
105 105 Otos_i);
106 106 __ should_not_reach_here();
107 107 return entry;
108 108 }
109 109
110 110
111 111 // Arguments are: required type in G5_method_type, and
112 112 // failing object (or NULL) in G3_method_handle.
113 113 address TemplateInterpreterGenerator::generate_WrongMethodType_handler() {
114 114 address entry = __ pc();
115 115 // expression stack must be empty before entering the VM if an exception
116 116 // happened
117 117 __ empty_expression_stack();
118 118 // load exception object
119 119 __ call_VM(Oexception,
120 120 CAST_FROM_FN_PTR(address,
121 121 InterpreterRuntime::throw_WrongMethodTypeException),
122 122 G5_method_type, // required
123 123 G3_method_handle); // actual
124 124 __ should_not_reach_here();
125 125 return entry;
126 126 }
127 127
128 128
129 129 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
130 130 address entry = __ pc();
131 131 // expression stack must be empty before entering the VM if an exception happened
132 132 __ empty_expression_stack();
133 133 // convention: expect aberrant index in register G3_scratch, then shuffle the
134 134 // index to G4_scratch for the VM call
135 135 __ mov(G3_scratch, G4_scratch);
136 136 __ set((intptr_t)name, G3_scratch);
137 137 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), G3_scratch, G4_scratch);
138 138 __ should_not_reach_here();
139 139 return entry;
140 140 }
141 141
142 142
143 143 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
↓ open down ↓ |
143 lines elided |
↑ open up ↑ |
144 144 address entry = __ pc();
145 145 // expression stack must be empty before entering the VM if an exception happened
146 146 __ empty_expression_stack();
147 147 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
148 148 __ should_not_reach_here();
149 149 return entry;
150 150 }
151 151
152 152
153 153 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step) {
154 - address compiled_entry = __ pc();
154 + TosState incoming_state = state;
155 +
155 156 Label cont;
157 + address compiled_entry = __ pc();
156 158
157 159 address entry = __ pc();
158 160 #if !defined(_LP64) && defined(COMPILER2)
159 161 // All return values are where we want them, except for Longs. C2 returns
160 162 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
161 163 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
162 164 // build even if we are returning from interpreted we just do a little
163 165 // stupid shuffing.
164 166 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
165 167 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
166 168 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
167 169
168 - if( state == ltos ) {
169 - __ srl (G1, 0,O1);
170 - __ srlx(G1,32,O0);
170 + if (incoming_state == ltos) {
171 + __ srl (G1, 0, O1);
172 + __ srlx(G1, 32, O0);
171 173 }
172 -#endif /* !_LP64 && COMPILER2 */
173 -
174 +#endif // !_LP64 && COMPILER2
174 175
175 176 __ bind(cont);
176 177
177 178 // The callee returns with the stack possibly adjusted by adapter transition
178 179 // We remove that possible adjustment here.
179 180 // All interpreter local registers are untouched. Any result is passed back
180 181 // in the O0/O1 or float registers. Before continuing, the arguments must be
181 182 // popped from the java expression stack; i.e., Lesp must be adjusted.
182 183
183 184 __ mov(Llast_SP, SP); // Remove any adapter added stack space.
184 185
185 -
186 + Label L_got_cache, L_giant_index;
186 187 const Register cache = G3_scratch;
187 188 const Register size = G1_scratch;
189 + if (EnableInvokeDynamic) {
190 + __ ldub(Address(Lbcp, 0), G1_scratch); // Load current bytecode.
191 + __ cmp(G1_scratch, Bytecodes::_invokedynamic);
192 + __ br(Assembler::equal, false, Assembler::pn, L_giant_index);
193 + __ delayed()->nop();
194 + }
188 195 __ get_cache_and_index_at_bcp(cache, G1_scratch, 1);
196 + __ bind(L_got_cache);
189 197 __ ld_ptr(cache, constantPoolCacheOopDesc::base_offset() +
190 198 ConstantPoolCacheEntry::flags_offset(), size);
191 199 __ and3(size, 0xFF, size); // argument size in words
192 200 __ sll(size, Interpreter::logStackElementSize(), size); // each argument size in bytes
193 201 __ add(Lesp, size, Lesp); // pop arguments
194 202 __ dispatch_next(state, step);
195 203
204 + // out of the main line of code...
205 + if (EnableInvokeDynamic) {
206 + __ bind(L_giant_index);
207 + __ get_cache_and_index_at_bcp(cache, G1_scratch, 1, true);
208 + __ ba(false, L_got_cache);
209 + __ delayed()->nop();
210 + }
211 +
196 212 return entry;
197 213 }
198 214
199 215
200 216 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
201 217 address entry = __ pc();
202 218 __ get_constant_pool_cache(LcpoolCache); // load LcpoolCache
203 219 { Label L;
204 220 Address exception_addr(G2_thread, Thread::pending_exception_offset());
205 221 __ ld_ptr(exception_addr, Gtemp); // Load pending exception.
206 222 __ tst(Gtemp);
207 223 __ brx(Assembler::equal, false, Assembler::pt, L);
208 224 __ delayed()->nop();
209 225 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
210 226 __ should_not_reach_here();
211 227 __ bind(L);
212 228 }
213 229 __ dispatch_next(state, step);
214 230 return entry;
215 231 }
216 232
217 233 // A result handler converts/unboxes a native call result into
218 234 // a java interpreter/compiler result. The current frame is an
219 235 // interpreter frame. The activation frame unwind code must be
220 236 // consistent with that of TemplateTable::_return(...). In the
221 237 // case of native methods, the caller's SP was not modified.
222 238 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
223 239 address entry = __ pc();
224 240 Register Itos_i = Otos_i ->after_save();
225 241 Register Itos_l = Otos_l ->after_save();
226 242 Register Itos_l1 = Otos_l1->after_save();
227 243 Register Itos_l2 = Otos_l2->after_save();
228 244 switch (type) {
229 245 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
230 246 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
231 247 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
232 248 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
233 249 case T_LONG :
234 250 #ifndef _LP64
235 251 __ mov(O1, Itos_l2); // move other half of long
236 252 #endif // ifdef or no ifdef, fall through to the T_INT case
237 253 case T_INT : __ mov(O0, Itos_i); break;
238 254 case T_VOID : /* nothing to do */ break;
239 255 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
240 256 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
241 257 case T_OBJECT :
242 258 __ ld_ptr(FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS, Itos_i);
243 259 __ verify_oop(Itos_i);
244 260 break;
245 261 default : ShouldNotReachHere();
246 262 }
247 263 __ ret(); // return from interpreter activation
248 264 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
249 265 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
250 266 return entry;
251 267 }
252 268
253 269 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
254 270 address entry = __ pc();
255 271 __ push(state);
256 272 __ call_VM(noreg, runtime_entry);
257 273 __ dispatch_via(vtos, Interpreter::normal_table(vtos));
258 274 return entry;
259 275 }
260 276
261 277
262 278 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
263 279 address entry = __ pc();
264 280 __ dispatch_next(state);
265 281 return entry;
266 282 }
267 283
268 284 //
269 285 // Helpers for commoning out cases in the various type of method entries.
270 286 //
271 287
272 288 // increment invocation count & check for overflow
273 289 //
274 290 // Note: checking for negative value instead of overflow
275 291 // so we have a 'sticky' overflow test
276 292 //
277 293 // Lmethod: method
278 294 // ??: invocation counter
279 295 //
280 296 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
281 297 // Update standard invocation counters
282 298 __ increment_invocation_counter(O0, G3_scratch);
283 299 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
284 300 Address interpreter_invocation_counter(Lmethod, methodOopDesc::interpreter_invocation_counter_offset());
285 301 __ ld(interpreter_invocation_counter, G3_scratch);
286 302 __ inc(G3_scratch);
287 303 __ st(G3_scratch, interpreter_invocation_counter);
288 304 }
289 305
290 306 if (ProfileInterpreter && profile_method != NULL) {
291 307 // Test to see if we should create a method data oop
292 308 AddressLiteral profile_limit(&InvocationCounter::InterpreterProfileLimit);
293 309 __ sethi(profile_limit, G3_scratch);
294 310 __ ld(G3_scratch, profile_limit.low10(), G3_scratch);
295 311 __ cmp(O0, G3_scratch);
296 312 __ br(Assembler::lessUnsigned, false, Assembler::pn, *profile_method_continue);
297 313 __ delayed()->nop();
298 314
299 315 // if no method data exists, go to profile_method
300 316 __ test_method_data_pointer(*profile_method);
301 317 }
302 318
303 319 AddressLiteral invocation_limit(&InvocationCounter::InterpreterInvocationLimit);
304 320 __ sethi(invocation_limit, G3_scratch);
305 321 __ ld(G3_scratch, invocation_limit.low10(), G3_scratch);
306 322 __ cmp(O0, G3_scratch);
307 323 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
308 324 __ delayed()->nop();
309 325
310 326 }
311 327
312 328 // Allocate monitor and lock method (asm interpreter)
313 329 // ebx - methodOop
314 330 //
315 331 void InterpreterGenerator::lock_method(void) {
316 332 __ ld(Lmethod, in_bytes(methodOopDesc::access_flags_offset()), O0); // Load access flags.
317 333
318 334 #ifdef ASSERT
319 335 { Label ok;
320 336 __ btst(JVM_ACC_SYNCHRONIZED, O0);
321 337 __ br( Assembler::notZero, false, Assembler::pt, ok);
322 338 __ delayed()->nop();
323 339 __ stop("method doesn't need synchronization");
324 340 __ bind(ok);
325 341 }
326 342 #endif // ASSERT
327 343
328 344 // get synchronization object to O0
329 345 { Label done;
330 346 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
331 347 __ btst(JVM_ACC_STATIC, O0);
332 348 __ br( Assembler::zero, true, Assembler::pt, done);
333 349 __ delayed()->ld_ptr(Llocals, Interpreter::local_offset_in_bytes(0), O0); // get receiver for not-static case
334 350
335 351 __ ld_ptr( Lmethod, in_bytes(methodOopDesc::constants_offset()), O0);
336 352 __ ld_ptr( O0, constantPoolOopDesc::pool_holder_offset_in_bytes(), O0);
337 353
338 354 // lock the mirror, not the klassOop
339 355 __ ld_ptr( O0, mirror_offset, O0);
340 356
341 357 #ifdef ASSERT
342 358 __ tst(O0);
343 359 __ breakpoint_trap(Assembler::zero);
344 360 #endif // ASSERT
345 361
346 362 __ bind(done);
347 363 }
348 364
349 365 __ add_monitor_to_stack(true, noreg, noreg); // allocate monitor elem
350 366 __ st_ptr( O0, Lmonitors, BasicObjectLock::obj_offset_in_bytes()); // store object
351 367 // __ untested("lock_object from method entry");
352 368 __ lock_object(Lmonitors, O0);
353 369 }
354 370
355 371
356 372 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rframe_size,
357 373 Register Rscratch,
358 374 Register Rscratch2) {
359 375 const int page_size = os::vm_page_size();
360 376 Address saved_exception_pc(G2_thread, JavaThread::saved_exception_pc_offset());
361 377 Label after_frame_check;
362 378
363 379 assert_different_registers(Rframe_size, Rscratch, Rscratch2);
364 380
365 381 __ set( page_size, Rscratch );
366 382 __ cmp( Rframe_size, Rscratch );
367 383
368 384 __ br( Assembler::lessEqual, false, Assembler::pt, after_frame_check );
369 385 __ delayed()->nop();
370 386
371 387 // get the stack base, and in debug, verify it is non-zero
372 388 __ ld_ptr( G2_thread, Thread::stack_base_offset(), Rscratch );
373 389 #ifdef ASSERT
374 390 Label base_not_zero;
375 391 __ cmp( Rscratch, G0 );
376 392 __ brx( Assembler::notEqual, false, Assembler::pn, base_not_zero );
377 393 __ delayed()->nop();
378 394 __ stop("stack base is zero in generate_stack_overflow_check");
379 395 __ bind(base_not_zero);
380 396 #endif
381 397
382 398 // get the stack size, and in debug, verify it is non-zero
383 399 assert( sizeof(size_t) == sizeof(intptr_t), "wrong load size" );
384 400 __ ld_ptr( G2_thread, Thread::stack_size_offset(), Rscratch2 );
385 401 #ifdef ASSERT
386 402 Label size_not_zero;
387 403 __ cmp( Rscratch2, G0 );
388 404 __ brx( Assembler::notEqual, false, Assembler::pn, size_not_zero );
389 405 __ delayed()->nop();
390 406 __ stop("stack size is zero in generate_stack_overflow_check");
391 407 __ bind(size_not_zero);
392 408 #endif
393 409
394 410 // compute the beginning of the protected zone minus the requested frame size
395 411 __ sub( Rscratch, Rscratch2, Rscratch );
396 412 __ set( (StackRedPages+StackYellowPages) * page_size, Rscratch2 );
397 413 __ add( Rscratch, Rscratch2, Rscratch );
398 414
399 415 // Add in the size of the frame (which is the same as subtracting it from the
400 416 // SP, which would take another register
401 417 __ add( Rscratch, Rframe_size, Rscratch );
402 418
403 419 // the frame is greater than one page in size, so check against
404 420 // the bottom of the stack
405 421 __ cmp( SP, Rscratch );
406 422 __ brx( Assembler::greater, false, Assembler::pt, after_frame_check );
407 423 __ delayed()->nop();
408 424
409 425 // Save the return address as the exception pc
410 426 __ st_ptr(O7, saved_exception_pc);
411 427
412 428 // the stack will overflow, throw an exception
413 429 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
414 430
415 431 // if you get to here, then there is enough stack space
416 432 __ bind( after_frame_check );
417 433 }
418 434
419 435
420 436 //
421 437 // Generate a fixed interpreter frame. This is identical setup for interpreted
422 438 // methods and for native methods hence the shared code.
423 439
424 440 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call) {
425 441 //
426 442 //
427 443 // The entry code sets up a new interpreter frame in 4 steps:
428 444 //
429 445 // 1) Increase caller's SP by for the extra local space needed:
430 446 // (check for overflow)
431 447 // Efficient implementation of xload/xstore bytecodes requires
432 448 // that arguments and non-argument locals are in a contigously
433 449 // addressable memory block => non-argument locals must be
434 450 // allocated in the caller's frame.
435 451 //
436 452 // 2) Create a new stack frame and register window:
437 453 // The new stack frame must provide space for the standard
438 454 // register save area, the maximum java expression stack size,
439 455 // the monitor slots (0 slots initially), and some frame local
440 456 // scratch locations.
441 457 //
442 458 // 3) The following interpreter activation registers must be setup:
443 459 // Lesp : expression stack pointer
444 460 // Lbcp : bytecode pointer
445 461 // Lmethod : method
446 462 // Llocals : locals pointer
447 463 // Lmonitors : monitor pointer
448 464 // LcpoolCache: constant pool cache
449 465 //
450 466 // 4) Initialize the non-argument locals if necessary:
451 467 // Non-argument locals may need to be initialized to NULL
452 468 // for GC to work. If the oop-map information is accurate
453 469 // (in the absence of the JSR problem), no initialization
454 470 // is necessary.
455 471 //
456 472 // (gri - 2/25/2000)
457 473
458 474
459 475 const Address size_of_parameters(G5_method, methodOopDesc::size_of_parameters_offset());
460 476 const Address size_of_locals (G5_method, methodOopDesc::size_of_locals_offset());
461 477 const Address max_stack (G5_method, methodOopDesc::max_stack_offset());
462 478 int rounded_vm_local_words = round_to( frame::interpreter_frame_vm_local_words, WordsPerLong );
463 479
464 480 const int extra_space =
465 481 rounded_vm_local_words + // frame local scratch space
466 482 //6815692//methodOopDesc::extra_stack_words() + // extra push slots for MH adapters
467 483 frame::memory_parameter_word_sp_offset + // register save area
468 484 (native_call ? frame::interpreter_frame_extra_outgoing_argument_words : 0);
469 485
470 486 const Register Glocals_size = G3;
471 487 const Register Otmp1 = O3;
472 488 const Register Otmp2 = O4;
473 489 // Lscratch can't be used as a temporary because the call_stub uses
474 490 // it to assert that the stack frame was setup correctly.
475 491
476 492 __ lduh( size_of_parameters, Glocals_size);
477 493
478 494 // Gargs points to first local + BytesPerWord
479 495 // Set the saved SP after the register window save
480 496 //
481 497 assert_different_registers(Gargs, Glocals_size, Gframe_size, O5_savedSP);
482 498 __ sll(Glocals_size, Interpreter::logStackElementSize(), Otmp1);
483 499 __ add(Gargs, Otmp1, Gargs);
484 500
485 501 if (native_call) {
486 502 __ calc_mem_param_words( Glocals_size, Gframe_size );
487 503 __ add( Gframe_size, extra_space, Gframe_size);
488 504 __ round_to( Gframe_size, WordsPerLong );
489 505 __ sll( Gframe_size, LogBytesPerWord, Gframe_size );
490 506 } else {
491 507
492 508 //
493 509 // Compute number of locals in method apart from incoming parameters
494 510 //
495 511 __ lduh( size_of_locals, Otmp1 );
496 512 __ sub( Otmp1, Glocals_size, Glocals_size );
497 513 __ round_to( Glocals_size, WordsPerLong );
498 514 __ sll( Glocals_size, Interpreter::logStackElementSize(), Glocals_size );
499 515
500 516 // see if the frame is greater than one page in size. If so,
501 517 // then we need to verify there is enough stack space remaining
502 518 // Frame_size = (max_stack + extra_space) * BytesPerWord;
503 519 __ lduh( max_stack, Gframe_size );
504 520 __ add( Gframe_size, extra_space, Gframe_size );
505 521 __ round_to( Gframe_size, WordsPerLong );
506 522 __ sll( Gframe_size, Interpreter::logStackElementSize(), Gframe_size);
507 523
508 524 // Add in java locals size for stack overflow check only
509 525 __ add( Gframe_size, Glocals_size, Gframe_size );
510 526
511 527 const Register Otmp2 = O4;
512 528 assert_different_registers(Otmp1, Otmp2, O5_savedSP);
513 529 generate_stack_overflow_check(Gframe_size, Otmp1, Otmp2);
514 530
515 531 __ sub( Gframe_size, Glocals_size, Gframe_size);
516 532
517 533 //
518 534 // bump SP to accomodate the extra locals
519 535 //
520 536 __ sub( SP, Glocals_size, SP );
521 537 }
522 538
523 539 //
524 540 // now set up a stack frame with the size computed above
525 541 //
526 542 __ neg( Gframe_size );
527 543 __ save( SP, Gframe_size, SP );
528 544
529 545 //
530 546 // now set up all the local cache registers
531 547 //
532 548 // NOTE: At this point, Lbyte_code/Lscratch has been modified. Note
533 549 // that all present references to Lbyte_code initialize the register
534 550 // immediately before use
535 551 if (native_call) {
536 552 __ mov(G0, Lbcp);
537 553 } else {
538 554 __ ld_ptr(G5_method, methodOopDesc::const_offset(), Lbcp);
539 555 __ add(Lbcp, in_bytes(constMethodOopDesc::codes_offset()), Lbcp);
540 556 }
541 557 __ mov( G5_method, Lmethod); // set Lmethod
542 558 __ get_constant_pool_cache( LcpoolCache ); // set LcpoolCache
543 559 __ sub(FP, rounded_vm_local_words * BytesPerWord, Lmonitors ); // set Lmonitors
544 560 #ifdef _LP64
545 561 __ add( Lmonitors, STACK_BIAS, Lmonitors ); // Account for 64 bit stack bias
546 562 #endif
547 563 __ sub(Lmonitors, BytesPerWord, Lesp); // set Lesp
548 564
549 565 // setup interpreter activation registers
550 566 __ sub(Gargs, BytesPerWord, Llocals); // set Llocals
551 567
552 568 if (ProfileInterpreter) {
553 569 #ifdef FAST_DISPATCH
554 570 // FAST_DISPATCH and ProfileInterpreter are mutually exclusive since
555 571 // they both use I2.
556 572 assert(0, "FAST_DISPATCH and +ProfileInterpreter are mutually exclusive");
557 573 #endif // FAST_DISPATCH
558 574 __ set_method_data_pointer();
559 575 }
560 576
561 577 }
562 578
563 579 // Empty method, generate a very fast return.
564 580
565 581 address InterpreterGenerator::generate_empty_entry(void) {
566 582
567 583 // A method that does nother but return...
568 584
569 585 address entry = __ pc();
570 586 Label slow_path;
571 587
572 588 __ verify_oop(G5_method);
573 589
574 590 // do nothing for empty methods (do not even increment invocation counter)
575 591 if ( UseFastEmptyMethods) {
576 592 // If we need a safepoint check, generate full interpreter entry.
577 593 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
578 594 __ set(sync_state, G3_scratch);
579 595 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
580 596 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
581 597 __ delayed()->nop();
582 598
583 599 // Code: _return
584 600 __ retl();
585 601 __ delayed()->mov(O5_savedSP, SP);
586 602
587 603 __ bind(slow_path);
588 604 (void) generate_normal_entry(false);
589 605
590 606 return entry;
591 607 }
592 608 return NULL;
593 609 }
594 610
595 611 // Call an accessor method (assuming it is resolved, otherwise drop into
596 612 // vanilla (slow path) entry
597 613
598 614 // Generates code to elide accessor methods
599 615 // Uses G3_scratch and G1_scratch as scratch
600 616 address InterpreterGenerator::generate_accessor_entry(void) {
601 617
602 618 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
603 619 // parameter size = 1
604 620 // Note: We can only use this code if the getfield has been resolved
605 621 // and if we don't have a null-pointer exception => check for
606 622 // these conditions first and use slow path if necessary.
607 623 address entry = __ pc();
608 624 Label slow_path;
609 625
610 626
611 627 // XXX: for compressed oops pointer loading and decoding doesn't fit in
612 628 // delay slot and damages G1
613 629 if ( UseFastAccessorMethods && !UseCompressedOops ) {
614 630 // Check if we need to reach a safepoint and generate full interpreter
615 631 // frame if so.
616 632 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
617 633 __ load_contents(sync_state, G3_scratch);
618 634 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
619 635 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
620 636 __ delayed()->nop();
621 637
622 638 // Check if local 0 != NULL
623 639 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
624 640 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
625 641 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
626 642 __ delayed()->nop();
627 643
628 644
629 645 // read first instruction word and extract bytecode @ 1 and index @ 2
630 646 // get first 4 bytes of the bytecodes (big endian!)
631 647 __ ld_ptr(G5_method, methodOopDesc::const_offset(), G1_scratch);
632 648 __ ld(G1_scratch, constMethodOopDesc::codes_offset(), G1_scratch);
633 649
634 650 // move index @ 2 far left then to the right most two bytes.
635 651 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
636 652 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
637 653 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
638 654
639 655 // get constant pool cache
640 656 __ ld_ptr(G5_method, methodOopDesc::constants_offset(), G3_scratch);
641 657 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
642 658
643 659 // get specific constant pool cache entry
644 660 __ add(G3_scratch, G1_scratch, G3_scratch);
645 661
646 662 // Check the constant Pool cache entry to see if it has been resolved.
647 663 // If not, need the slow path.
648 664 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
649 665 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::indices_offset(), G1_scratch);
650 666 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
651 667 __ and3(G1_scratch, 0xFF, G1_scratch);
652 668 __ cmp(G1_scratch, Bytecodes::_getfield);
653 669 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
654 670 __ delayed()->nop();
655 671
656 672 // Get the type and return field offset from the constant pool cache
657 673 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::flags_offset(), G1_scratch);
658 674 __ ld_ptr(G3_scratch, cp_base_offset + ConstantPoolCacheEntry::f2_offset(), G3_scratch);
659 675
660 676 Label xreturn_path;
661 677 // Need to differentiate between igetfield, agetfield, bgetfield etc.
662 678 // because they are different sizes.
663 679 // Get the type from the constant pool cache
664 680 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
665 681 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
666 682 ConstantPoolCacheEntry::verify_tosBits();
667 683 __ cmp(G1_scratch, atos );
668 684 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
669 685 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
670 686 __ cmp(G1_scratch, itos);
671 687 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
672 688 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
673 689 __ cmp(G1_scratch, stos);
674 690 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
675 691 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
676 692 __ cmp(G1_scratch, ctos);
677 693 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
678 694 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
679 695 #ifdef ASSERT
680 696 __ cmp(G1_scratch, btos);
681 697 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
682 698 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
683 699 __ should_not_reach_here();
684 700 #endif
685 701 __ ldsb(Otos_i, G3_scratch, Otos_i);
686 702 __ bind(xreturn_path);
687 703
688 704 // _ireturn/_areturn
689 705 __ retl(); // return from leaf routine
690 706 __ delayed()->mov(O5_savedSP, SP);
691 707
692 708 // Generate regular method entry
693 709 __ bind(slow_path);
694 710 (void) generate_normal_entry(false);
695 711 return entry;
696 712 }
697 713 return NULL;
698 714 }
699 715
700 716 //
701 717 // Interpreter stub for calling a native method. (asm interpreter)
702 718 // This sets up a somewhat different looking stack for calling the native method
703 719 // than the typical interpreter frame setup.
704 720 //
705 721
706 722 address InterpreterGenerator::generate_native_entry(bool synchronized) {
707 723 address entry = __ pc();
708 724
709 725 // the following temporary registers are used during frame creation
710 726 const Register Gtmp1 = G3_scratch ;
711 727 const Register Gtmp2 = G1_scratch;
712 728 bool inc_counter = UseCompiler || CountCompiledCalls;
713 729
714 730 // make sure registers are different!
715 731 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
716 732
717 733 const Address Laccess_flags(Lmethod, methodOopDesc::access_flags_offset());
718 734
719 735 __ verify_oop(G5_method);
720 736
721 737 const Register Glocals_size = G3;
722 738 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
723 739
724 740 // make sure method is native & not abstract
725 741 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
726 742 #ifdef ASSERT
727 743 __ ld(G5_method, methodOopDesc::access_flags_offset(), Gtmp1);
728 744 {
729 745 Label L;
730 746 __ btst(JVM_ACC_NATIVE, Gtmp1);
731 747 __ br(Assembler::notZero, false, Assembler::pt, L);
732 748 __ delayed()->nop();
733 749 __ stop("tried to execute non-native method as native");
734 750 __ bind(L);
735 751 }
736 752 { Label L;
737 753 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
738 754 __ br(Assembler::zero, false, Assembler::pt, L);
739 755 __ delayed()->nop();
740 756 __ stop("tried to execute abstract method as non-abstract");
741 757 __ bind(L);
742 758 }
743 759 #endif // ASSERT
744 760
745 761 // generate the code to allocate the interpreter stack frame
746 762 generate_fixed_frame(true);
747 763
748 764 //
749 765 // No locals to initialize for native method
750 766 //
751 767
752 768 // this slot will be set later, we initialize it to null here just in
753 769 // case we get a GC before the actual value is stored later
754 770 __ st_ptr(G0, FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS);
755 771
756 772 const Address do_not_unlock_if_synchronized(G2_thread,
757 773 JavaThread::do_not_unlock_if_synchronized_offset());
758 774 // Since at this point in the method invocation the exception handler
759 775 // would try to exit the monitor of synchronized methods which hasn't
760 776 // been entered yet, we set the thread local variable
761 777 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
762 778 // runtime, exception handling i.e. unlock_if_synchronized_method will
763 779 // check this thread local flag.
764 780 // This flag has two effects, one is to force an unwind in the topmost
765 781 // interpreter frame and not perform an unlock while doing so.
766 782
767 783 __ movbool(true, G3_scratch);
768 784 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
769 785
770 786 // increment invocation counter and check for overflow
771 787 //
772 788 // Note: checking for negative value instead of overflow
773 789 // so we have a 'sticky' overflow test (may be of
774 790 // importance as soon as we have true MT/MP)
775 791 Label invocation_counter_overflow;
776 792 Label Lcontinue;
777 793 if (inc_counter) {
778 794 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
779 795
780 796 }
781 797 __ bind(Lcontinue);
782 798
783 799 bang_stack_shadow_pages(true);
784 800
785 801 // reset the _do_not_unlock_if_synchronized flag
786 802 __ stbool(G0, do_not_unlock_if_synchronized);
787 803
788 804 // check for synchronized methods
789 805 // Must happen AFTER invocation_counter check and stack overflow check,
790 806 // so method is not locked if overflows.
791 807
792 808 if (synchronized) {
793 809 lock_method();
794 810 } else {
795 811 #ifdef ASSERT
796 812 { Label ok;
797 813 __ ld(Laccess_flags, O0);
798 814 __ btst(JVM_ACC_SYNCHRONIZED, O0);
799 815 __ br( Assembler::zero, false, Assembler::pt, ok);
800 816 __ delayed()->nop();
801 817 __ stop("method needs synchronization");
802 818 __ bind(ok);
803 819 }
804 820 #endif // ASSERT
805 821 }
806 822
807 823
808 824 // start execution
809 825 __ verify_thread();
810 826
811 827 // JVMTI support
812 828 __ notify_method_entry();
813 829
814 830 // native call
815 831
816 832 // (note that O0 is never an oop--at most it is a handle)
817 833 // It is important not to smash any handles created by this call,
818 834 // until any oop handle in O0 is dereferenced.
819 835
820 836 // (note that the space for outgoing params is preallocated)
821 837
822 838 // get signature handler
823 839 { Label L;
824 840 Address signature_handler(Lmethod, methodOopDesc::signature_handler_offset());
825 841 __ ld_ptr(signature_handler, G3_scratch);
826 842 __ tst(G3_scratch);
827 843 __ brx(Assembler::notZero, false, Assembler::pt, L);
828 844 __ delayed()->nop();
829 845 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), Lmethod);
830 846 __ ld_ptr(signature_handler, G3_scratch);
831 847 __ bind(L);
832 848 }
833 849
834 850 // Push a new frame so that the args will really be stored in
835 851 // Copy a few locals across so the new frame has the variables
836 852 // we need but these values will be dead at the jni call and
837 853 // therefore not gc volatile like the values in the current
838 854 // frame (Lmethod in particular)
839 855
840 856 // Flush the method pointer to the register save area
841 857 __ st_ptr(Lmethod, SP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
842 858 __ mov(Llocals, O1);
843 859
844 860 // calculate where the mirror handle body is allocated in the interpreter frame:
845 861 __ add(FP, (frame::interpreter_frame_oop_temp_offset * wordSize) + STACK_BIAS, O2);
846 862
847 863 // Calculate current frame size
848 864 __ sub(SP, FP, O3); // Calculate negative of current frame size
849 865 __ save(SP, O3, SP); // Allocate an identical sized frame
850 866
851 867 // Note I7 has leftover trash. Slow signature handler will fill it in
852 868 // should we get there. Normal jni call will set reasonable last_Java_pc
853 869 // below (and fix I7 so the stack trace doesn't have a meaningless frame
854 870 // in it).
855 871
856 872 // Load interpreter frame's Lmethod into same register here
857 873
858 874 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
859 875
860 876 __ mov(I1, Llocals);
861 877 __ mov(I2, Lscratch2); // save the address of the mirror
862 878
863 879
864 880 // ONLY Lmethod and Llocals are valid here!
865 881
866 882 // call signature handler, It will move the arg properly since Llocals in current frame
867 883 // matches that in outer frame
868 884
869 885 __ callr(G3_scratch, 0);
870 886 __ delayed()->nop();
871 887
872 888 // Result handler is in Lscratch
873 889
874 890 // Reload interpreter frame's Lmethod since slow signature handler may block
875 891 __ ld_ptr(FP, (Lmethod->sp_offset_in_saved_window() * wordSize) + STACK_BIAS, Lmethod);
876 892
877 893 { Label not_static;
878 894
879 895 __ ld(Laccess_flags, O0);
880 896 __ btst(JVM_ACC_STATIC, O0);
881 897 __ br( Assembler::zero, false, Assembler::pt, not_static);
882 898 // get native function entry point(O0 is a good temp until the very end)
883 899 __ delayed()->ld_ptr(Lmethod, in_bytes(methodOopDesc::native_function_offset()), O0);
884 900 // for static methods insert the mirror argument
885 901 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
886 902
887 903 __ ld_ptr(Lmethod, methodOopDesc:: constants_offset(), O1);
888 904 __ ld_ptr(O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
889 905 __ ld_ptr(O1, mirror_offset, O1);
890 906 #ifdef ASSERT
891 907 if (!PrintSignatureHandlers) // do not dirty the output with this
892 908 { Label L;
893 909 __ tst(O1);
894 910 __ brx(Assembler::notZero, false, Assembler::pt, L);
895 911 __ delayed()->nop();
896 912 __ stop("mirror is missing");
897 913 __ bind(L);
898 914 }
899 915 #endif // ASSERT
900 916 __ st_ptr(O1, Lscratch2, 0);
901 917 __ mov(Lscratch2, O1);
902 918 __ bind(not_static);
903 919 }
904 920
905 921 // At this point, arguments have been copied off of stack into
906 922 // their JNI positions, which are O1..O5 and SP[68..].
907 923 // Oops are boxed in-place on the stack, with handles copied to arguments.
908 924 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
909 925
910 926 #ifdef ASSERT
911 927 { Label L;
912 928 __ tst(O0);
913 929 __ brx(Assembler::notZero, false, Assembler::pt, L);
914 930 __ delayed()->nop();
915 931 __ stop("native entry point is missing");
916 932 __ bind(L);
917 933 }
918 934 #endif // ASSERT
919 935
920 936 //
921 937 // setup the frame anchor
922 938 //
923 939 // The scavenge function only needs to know that the PC of this frame is
924 940 // in the interpreter method entry code, it doesn't need to know the exact
925 941 // PC and hence we can use O7 which points to the return address from the
926 942 // previous call in the code stream (signature handler function)
927 943 //
928 944 // The other trick is we set last_Java_sp to FP instead of the usual SP because
929 945 // we have pushed the extra frame in order to protect the volatile register(s)
930 946 // in that frame when we return from the jni call
931 947 //
932 948
933 949 __ set_last_Java_frame(FP, O7);
934 950 __ mov(O7, I7); // make dummy interpreter frame look like one above,
935 951 // not meaningless information that'll confuse me.
936 952
937 953 // flush the windows now. We don't care about the current (protection) frame
938 954 // only the outer frames
939 955
940 956 __ flush_windows();
941 957
942 958 // mark windows as flushed
943 959 Address flags(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
944 960 __ set(JavaFrameAnchor::flushed, G3_scratch);
945 961 __ st(G3_scratch, flags);
946 962
947 963 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
948 964
949 965 Address thread_state(G2_thread, JavaThread::thread_state_offset());
950 966 #ifdef ASSERT
951 967 { Label L;
952 968 __ ld(thread_state, G3_scratch);
953 969 __ cmp(G3_scratch, _thread_in_Java);
954 970 __ br(Assembler::equal, false, Assembler::pt, L);
955 971 __ delayed()->nop();
956 972 __ stop("Wrong thread state in native stub");
957 973 __ bind(L);
958 974 }
959 975 #endif // ASSERT
960 976 __ set(_thread_in_native, G3_scratch);
961 977 __ st(G3_scratch, thread_state);
962 978
963 979 // Call the jni method, using the delay slot to set the JNIEnv* argument.
964 980 __ save_thread(L7_thread_cache); // save Gthread
965 981 __ callr(O0, 0);
966 982 __ delayed()->
967 983 add(L7_thread_cache, in_bytes(JavaThread::jni_environment_offset()), O0);
968 984
969 985 // Back from jni method Lmethod in this frame is DEAD, DEAD, DEAD
970 986
971 987 __ restore_thread(L7_thread_cache); // restore G2_thread
972 988 __ reinit_heapbase();
973 989
974 990 // must we block?
975 991
976 992 // Block, if necessary, before resuming in _thread_in_Java state.
977 993 // In order for GC to work, don't clear the last_Java_sp until after blocking.
978 994 { Label no_block;
979 995 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
980 996
981 997 // Switch thread to "native transition" state before reading the synchronization state.
982 998 // This additional state is necessary because reading and testing the synchronization
983 999 // state is not atomic w.r.t. GC, as this scenario demonstrates:
984 1000 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
985 1001 // VM thread changes sync state to synchronizing and suspends threads for GC.
986 1002 // Thread A is resumed to finish this native method, but doesn't block here since it
987 1003 // didn't see any synchronization is progress, and escapes.
988 1004 __ set(_thread_in_native_trans, G3_scratch);
989 1005 __ st(G3_scratch, thread_state);
990 1006 if(os::is_MP()) {
991 1007 if (UseMembar) {
992 1008 // Force this write out before the read below
993 1009 __ membar(Assembler::StoreLoad);
994 1010 } else {
995 1011 // Write serialization page so VM thread can do a pseudo remote membar.
996 1012 // We use the current thread pointer to calculate a thread specific
997 1013 // offset to write to within the page. This minimizes bus traffic
998 1014 // due to cache line collision.
999 1015 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
1000 1016 }
1001 1017 }
1002 1018 __ load_contents(sync_state, G3_scratch);
1003 1019 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
1004 1020
1005 1021 Label L;
1006 1022 __ br(Assembler::notEqual, false, Assembler::pn, L);
1007 1023 __ delayed()->ld(G2_thread, JavaThread::suspend_flags_offset(), G3_scratch);
1008 1024 __ cmp(G3_scratch, 0);
1009 1025 __ br(Assembler::equal, false, Assembler::pt, no_block);
1010 1026 __ delayed()->nop();
1011 1027 __ bind(L);
1012 1028
1013 1029 // Block. Save any potential method result value before the operation and
1014 1030 // use a leaf call to leave the last_Java_frame setup undisturbed.
1015 1031 save_native_result();
1016 1032 __ call_VM_leaf(L7_thread_cache,
1017 1033 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1018 1034 G2_thread);
1019 1035
1020 1036 // Restore any method result value
1021 1037 restore_native_result();
1022 1038 __ bind(no_block);
1023 1039 }
1024 1040
1025 1041 // Clear the frame anchor now
1026 1042
1027 1043 __ reset_last_Java_frame();
1028 1044
1029 1045 // Move the result handler address
1030 1046 __ mov(Lscratch, G3_scratch);
1031 1047 // return possible result to the outer frame
1032 1048 #ifndef __LP64
1033 1049 __ mov(O0, I0);
1034 1050 __ restore(O1, G0, O1);
1035 1051 #else
1036 1052 __ restore(O0, G0, O0);
1037 1053 #endif /* __LP64 */
1038 1054
1039 1055 // Move result handler to expected register
1040 1056 __ mov(G3_scratch, Lscratch);
1041 1057
1042 1058 // Back in normal (native) interpreter frame. State is thread_in_native_trans
1043 1059 // switch to thread_in_Java.
1044 1060
1045 1061 __ set(_thread_in_Java, G3_scratch);
1046 1062 __ st(G3_scratch, thread_state);
1047 1063
1048 1064 // reset handle block
1049 1065 __ ld_ptr(G2_thread, JavaThread::active_handles_offset(), G3_scratch);
1050 1066 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
1051 1067
1052 1068 // If we have an oop result store it where it will be safe for any further gc
1053 1069 // until we return now that we've released the handle it might be protected by
1054 1070
1055 1071 {
1056 1072 Label no_oop, store_result;
1057 1073
1058 1074 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
1059 1075 __ cmp(G3_scratch, Lscratch);
1060 1076 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
1061 1077 __ delayed()->nop();
1062 1078 __ addcc(G0, O0, O0);
1063 1079 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
1064 1080 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
1065 1081 __ mov(G0, O0);
1066 1082
1067 1083 __ bind(store_result);
1068 1084 // Store it where gc will look for it and result handler expects it.
1069 1085 __ st_ptr(O0, FP, (frame::interpreter_frame_oop_temp_offset*wordSize) + STACK_BIAS);
1070 1086
1071 1087 __ bind(no_oop);
1072 1088
1073 1089 }
1074 1090
1075 1091
1076 1092 // handle exceptions (exception handling will handle unlocking!)
1077 1093 { Label L;
1078 1094 Address exception_addr(G2_thread, Thread::pending_exception_offset());
1079 1095 __ ld_ptr(exception_addr, Gtemp);
1080 1096 __ tst(Gtemp);
1081 1097 __ brx(Assembler::equal, false, Assembler::pt, L);
1082 1098 __ delayed()->nop();
1083 1099 // Note: This could be handled more efficiently since we know that the native
1084 1100 // method doesn't have an exception handler. We could directly return
1085 1101 // to the exception handler for the caller.
1086 1102 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_pending_exception));
1087 1103 __ should_not_reach_here();
1088 1104 __ bind(L);
1089 1105 }
1090 1106
1091 1107 // JVMTI support (preserves thread register)
1092 1108 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
1093 1109
1094 1110 if (synchronized) {
1095 1111 // save and restore any potential method result value around the unlocking operation
1096 1112 save_native_result();
1097 1113
1098 1114 __ add( __ top_most_monitor(), O1);
1099 1115 __ unlock_object(O1);
1100 1116
1101 1117 restore_native_result();
1102 1118 }
1103 1119
1104 1120 #if defined(COMPILER2) && !defined(_LP64)
1105 1121
1106 1122 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1107 1123 // or compiled so just be safe.
1108 1124
1109 1125 __ sllx(O0, 32, G1); // Shift bits into high G1
1110 1126 __ srl (O1, 0, O1); // Zero extend O1
1111 1127 __ or3 (O1, G1, G1); // OR 64 bits into G1
1112 1128
1113 1129 #endif /* COMPILER2 && !_LP64 */
1114 1130
1115 1131 // dispose of return address and remove activation
1116 1132 #ifdef ASSERT
1117 1133 {
1118 1134 Label ok;
1119 1135 __ cmp(I5_savedSP, FP);
1120 1136 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1121 1137 __ delayed()->nop();
1122 1138 __ stop("bad I5_savedSP value");
1123 1139 __ should_not_reach_here();
1124 1140 __ bind(ok);
1125 1141 }
1126 1142 #endif
1127 1143 if (TraceJumps) {
1128 1144 // Move target to register that is recordable
1129 1145 __ mov(Lscratch, G3_scratch);
1130 1146 __ JMP(G3_scratch, 0);
1131 1147 } else {
1132 1148 __ jmp(Lscratch, 0);
1133 1149 }
1134 1150 __ delayed()->nop();
1135 1151
1136 1152
1137 1153 if (inc_counter) {
1138 1154 // handle invocation counter overflow
1139 1155 __ bind(invocation_counter_overflow);
1140 1156 generate_counter_overflow(Lcontinue);
1141 1157 }
1142 1158
1143 1159
1144 1160
1145 1161 return entry;
1146 1162 }
1147 1163
1148 1164
1149 1165 // Generic method entry to (asm) interpreter
1150 1166 //------------------------------------------------------------------------------------------------------------------------
1151 1167 //
1152 1168 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1153 1169 address entry = __ pc();
1154 1170
1155 1171 bool inc_counter = UseCompiler || CountCompiledCalls;
1156 1172
1157 1173 // the following temporary registers are used during frame creation
1158 1174 const Register Gtmp1 = G3_scratch ;
1159 1175 const Register Gtmp2 = G1_scratch;
1160 1176
1161 1177 // make sure registers are different!
1162 1178 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
1163 1179
1164 1180 const Address size_of_parameters(G5_method, methodOopDesc::size_of_parameters_offset());
1165 1181 const Address size_of_locals (G5_method, methodOopDesc::size_of_locals_offset());
1166 1182 // Seems like G5_method is live at the point this is used. So we could make this look consistent
1167 1183 // and use in the asserts.
1168 1184 const Address access_flags (Lmethod, methodOopDesc::access_flags_offset());
1169 1185
1170 1186 __ verify_oop(G5_method);
1171 1187
1172 1188 const Register Glocals_size = G3;
1173 1189 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
1174 1190
1175 1191 // make sure method is not native & not abstract
1176 1192 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
1177 1193 #ifdef ASSERT
1178 1194 __ ld(G5_method, methodOopDesc::access_flags_offset(), Gtmp1);
1179 1195 {
1180 1196 Label L;
1181 1197 __ btst(JVM_ACC_NATIVE, Gtmp1);
1182 1198 __ br(Assembler::zero, false, Assembler::pt, L);
1183 1199 __ delayed()->nop();
1184 1200 __ stop("tried to execute native method as non-native");
1185 1201 __ bind(L);
1186 1202 }
1187 1203 { Label L;
1188 1204 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
1189 1205 __ br(Assembler::zero, false, Assembler::pt, L);
1190 1206 __ delayed()->nop();
1191 1207 __ stop("tried to execute abstract method as non-abstract");
1192 1208 __ bind(L);
1193 1209 }
1194 1210 #endif // ASSERT
1195 1211
1196 1212 // generate the code to allocate the interpreter stack frame
1197 1213
1198 1214 generate_fixed_frame(false);
1199 1215
1200 1216 #ifdef FAST_DISPATCH
1201 1217 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1202 1218 // set bytecode dispatch table base
1203 1219 #endif
1204 1220
1205 1221 //
1206 1222 // Code to initialize the extra (i.e. non-parm) locals
1207 1223 //
1208 1224 Register init_value = noreg; // will be G0 if we must clear locals
1209 1225 // The way the code was setup before zerolocals was always true for vanilla java entries.
1210 1226 // It could only be false for the specialized entries like accessor or empty which have
1211 1227 // no extra locals so the testing was a waste of time and the extra locals were always
1212 1228 // initialized. We removed this extra complication to already over complicated code.
1213 1229
1214 1230 init_value = G0;
1215 1231 Label clear_loop;
1216 1232
1217 1233 // NOTE: If you change the frame layout, this code will need to
1218 1234 // be updated!
1219 1235 __ lduh( size_of_locals, O2 );
1220 1236 __ lduh( size_of_parameters, O1 );
1221 1237 __ sll( O2, Interpreter::logStackElementSize(), O2);
1222 1238 __ sll( O1, Interpreter::logStackElementSize(), O1 );
1223 1239 __ sub( Llocals, O2, O2 );
1224 1240 __ sub( Llocals, O1, O1 );
1225 1241
1226 1242 __ bind( clear_loop );
1227 1243 __ inc( O2, wordSize );
1228 1244
1229 1245 __ cmp( O2, O1 );
1230 1246 __ brx( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1231 1247 __ delayed()->st_ptr( init_value, O2, 0 );
1232 1248
1233 1249 const Address do_not_unlock_if_synchronized(G2_thread,
1234 1250 JavaThread::do_not_unlock_if_synchronized_offset());
1235 1251 // Since at this point in the method invocation the exception handler
1236 1252 // would try to exit the monitor of synchronized methods which hasn't
1237 1253 // been entered yet, we set the thread local variable
1238 1254 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1239 1255 // runtime, exception handling i.e. unlock_if_synchronized_method will
1240 1256 // check this thread local flag.
1241 1257 __ movbool(true, G3_scratch);
1242 1258 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
1243 1259
1244 1260 // increment invocation counter and check for overflow
1245 1261 //
1246 1262 // Note: checking for negative value instead of overflow
1247 1263 // so we have a 'sticky' overflow test (may be of
1248 1264 // importance as soon as we have true MT/MP)
1249 1265 Label invocation_counter_overflow;
1250 1266 Label profile_method;
1251 1267 Label profile_method_continue;
1252 1268 Label Lcontinue;
1253 1269 if (inc_counter) {
1254 1270 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1255 1271 if (ProfileInterpreter) {
1256 1272 __ bind(profile_method_continue);
1257 1273 }
1258 1274 }
1259 1275 __ bind(Lcontinue);
1260 1276
1261 1277 bang_stack_shadow_pages(false);
1262 1278
1263 1279 // reset the _do_not_unlock_if_synchronized flag
1264 1280 __ stbool(G0, do_not_unlock_if_synchronized);
1265 1281
1266 1282 // check for synchronized methods
1267 1283 // Must happen AFTER invocation_counter check and stack overflow check,
1268 1284 // so method is not locked if overflows.
1269 1285
1270 1286 if (synchronized) {
1271 1287 lock_method();
1272 1288 } else {
1273 1289 #ifdef ASSERT
1274 1290 { Label ok;
1275 1291 __ ld(access_flags, O0);
1276 1292 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1277 1293 __ br( Assembler::zero, false, Assembler::pt, ok);
1278 1294 __ delayed()->nop();
1279 1295 __ stop("method needs synchronization");
1280 1296 __ bind(ok);
1281 1297 }
1282 1298 #endif // ASSERT
1283 1299 }
1284 1300
1285 1301 // start execution
1286 1302
1287 1303 __ verify_thread();
1288 1304
1289 1305 // jvmti support
1290 1306 __ notify_method_entry();
1291 1307
1292 1308 // start executing instructions
1293 1309 __ dispatch_next(vtos);
1294 1310
1295 1311
1296 1312 if (inc_counter) {
1297 1313 if (ProfileInterpreter) {
1298 1314 // We have decided to profile this method in the interpreter
1299 1315 __ bind(profile_method);
1300 1316
1301 1317 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method), Lbcp, true);
1302 1318
1303 1319 #ifdef ASSERT
1304 1320 __ tst(O0);
1305 1321 __ breakpoint_trap(Assembler::notEqual);
1306 1322 #endif
1307 1323
1308 1324 __ set_method_data_pointer();
1309 1325
1310 1326 __ ba(false, profile_method_continue);
1311 1327 __ delayed()->nop();
1312 1328 }
1313 1329
1314 1330 // handle invocation counter overflow
1315 1331 __ bind(invocation_counter_overflow);
1316 1332 generate_counter_overflow(Lcontinue);
1317 1333 }
1318 1334
1319 1335
1320 1336 return entry;
1321 1337 }
1322 1338
1323 1339
1324 1340 //----------------------------------------------------------------------------------------------------
1325 1341 // Entry points & stack frame layout
1326 1342 //
1327 1343 // Here we generate the various kind of entries into the interpreter.
1328 1344 // The two main entry type are generic bytecode methods and native call method.
1329 1345 // These both come in synchronized and non-synchronized versions but the
1330 1346 // frame layout they create is very similar. The other method entry
1331 1347 // types are really just special purpose entries that are really entry
1332 1348 // and interpretation all in one. These are for trivial methods like
1333 1349 // accessor, empty, or special math methods.
1334 1350 //
1335 1351 // When control flow reaches any of the entry types for the interpreter
1336 1352 // the following holds ->
1337 1353 //
1338 1354 // C2 Calling Conventions:
1339 1355 //
1340 1356 // The entry code below assumes that the following registers are set
1341 1357 // when coming in:
1342 1358 // G5_method: holds the methodOop of the method to call
1343 1359 // Lesp: points to the TOS of the callers expression stack
1344 1360 // after having pushed all the parameters
1345 1361 //
1346 1362 // The entry code does the following to setup an interpreter frame
1347 1363 // pop parameters from the callers stack by adjusting Lesp
1348 1364 // set O0 to Lesp
1349 1365 // compute X = (max_locals - num_parameters)
1350 1366 // bump SP up by X to accomadate the extra locals
1351 1367 // compute X = max_expression_stack
1352 1368 // + vm_local_words
1353 1369 // + 16 words of register save area
1354 1370 // save frame doing a save sp, -X, sp growing towards lower addresses
1355 1371 // set Lbcp, Lmethod, LcpoolCache
1356 1372 // set Llocals to i0
1357 1373 // set Lmonitors to FP - rounded_vm_local_words
1358 1374 // set Lesp to Lmonitors - 4
1359 1375 //
1360 1376 // The frame has now been setup to do the rest of the entry code
1361 1377
1362 1378 // Try this optimization: Most method entries could live in a
1363 1379 // "one size fits all" stack frame without all the dynamic size
1364 1380 // calculations. It might be profitable to do all this calculation
1365 1381 // statically and approximately for "small enough" methods.
1366 1382
1367 1383 //-----------------------------------------------------------------------------------------------
1368 1384
1369 1385 // C1 Calling conventions
1370 1386 //
1371 1387 // Upon method entry, the following registers are setup:
1372 1388 //
1373 1389 // g2 G2_thread: current thread
1374 1390 // g5 G5_method: method to activate
1375 1391 // g4 Gargs : pointer to last argument
1376 1392 //
1377 1393 //
1378 1394 // Stack:
1379 1395 //
1380 1396 // +---------------+ <--- sp
1381 1397 // | |
1382 1398 // : reg save area :
1383 1399 // | |
1384 1400 // +---------------+ <--- sp + 0x40
1385 1401 // | |
1386 1402 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1387 1403 // | |
1388 1404 // +---------------+ <--- sp + 0x5c
1389 1405 // | |
1390 1406 // : free :
1391 1407 // | |
1392 1408 // +---------------+ <--- Gargs
1393 1409 // | |
1394 1410 // : arguments :
1395 1411 // | |
1396 1412 // +---------------+
1397 1413 // | |
1398 1414 //
1399 1415 //
1400 1416 //
1401 1417 // AFTER FRAME HAS BEEN SETUP for method interpretation the stack looks like:
1402 1418 //
1403 1419 // +---------------+ <--- sp
1404 1420 // | |
1405 1421 // : reg save area :
1406 1422 // | |
1407 1423 // +---------------+ <--- sp + 0x40
1408 1424 // | |
1409 1425 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1410 1426 // | |
1411 1427 // +---------------+ <--- sp + 0x5c
1412 1428 // | |
1413 1429 // : :
1414 1430 // | | <--- Lesp
1415 1431 // +---------------+ <--- Lmonitors (fp - 0x18)
1416 1432 // | VM locals |
1417 1433 // +---------------+ <--- fp
1418 1434 // | |
1419 1435 // : reg save area :
1420 1436 // | |
1421 1437 // +---------------+ <--- fp + 0x40
1422 1438 // | |
1423 1439 // : extra 7 slots : note: these slots are not really needed for the interpreter (fix later)
1424 1440 // | |
1425 1441 // +---------------+ <--- fp + 0x5c
1426 1442 // | |
1427 1443 // : free :
1428 1444 // | |
1429 1445 // +---------------+
1430 1446 // | |
1431 1447 // : nonarg locals :
1432 1448 // | |
1433 1449 // +---------------+
1434 1450 // | |
1435 1451 // : arguments :
1436 1452 // | | <--- Llocals
1437 1453 // +---------------+ <--- Gargs
1438 1454 // | |
1439 1455
1440 1456 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
1441 1457
1442 1458 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
1443 1459 // expression stack, the callee will have callee_extra_locals (so we can account for
1444 1460 // frame extension) and monitor_size for monitors. Basically we need to calculate
1445 1461 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
1446 1462 //
1447 1463 //
1448 1464 // The big complicating thing here is that we must ensure that the stack stays properly
1449 1465 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
1450 1466 // needs to be aligned for). We are given that the sp (fp) is already aligned by
1451 1467 // the caller so we must ensure that it is properly aligned for our callee.
1452 1468 //
1453 1469 const int rounded_vm_local_words =
1454 1470 round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1455 1471 // callee_locals and max_stack are counts, not the size in frame.
1456 1472 const int locals_size =
1457 1473 round_to(callee_extra_locals * Interpreter::stackElementWords(), WordsPerLong);
1458 1474 const int max_stack_words = max_stack * Interpreter::stackElementWords();
1459 1475 return (round_to((max_stack_words
1460 1476 //6815692//+ methodOopDesc::extra_stack_words()
1461 1477 + rounded_vm_local_words
1462 1478 + frame::memory_parameter_word_sp_offset), WordsPerLong)
1463 1479 // already rounded
1464 1480 + locals_size + monitor_size);
1465 1481 }
1466 1482
1467 1483 // How much stack a method top interpreter activation needs in words.
1468 1484 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
1469 1485
1470 1486 // See call_stub code
1471 1487 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
1472 1488 WordsPerLong); // 7 + register save area
1473 1489
1474 1490 // Save space for one monitor to get into the interpreted method in case
1475 1491 // the method is synchronized
1476 1492 int monitor_size = method->is_synchronized() ?
1477 1493 1*frame::interpreter_frame_monitor_size() : 0;
1478 1494 return size_activation_helper(method->max_locals(), method->max_stack(),
1479 1495 monitor_size) + call_stub_size;
1480 1496 }
1481 1497
1482 1498 int AbstractInterpreter::layout_activation(methodOop method,
1483 1499 int tempcount,
1484 1500 int popframe_extra_args,
1485 1501 int moncount,
1486 1502 int callee_param_count,
1487 1503 int callee_local_count,
1488 1504 frame* caller,
1489 1505 frame* interpreter_frame,
1490 1506 bool is_top_frame) {
1491 1507 // Note: This calculation must exactly parallel the frame setup
1492 1508 // in InterpreterGenerator::generate_fixed_frame.
1493 1509 // If f!=NULL, set up the following variables:
1494 1510 // - Lmethod
1495 1511 // - Llocals
1496 1512 // - Lmonitors (to the indicated number of monitors)
1497 1513 // - Lesp (to the indicated number of temps)
1498 1514 // The frame f (if not NULL) on entry is a description of the caller of the frame
1499 1515 // we are about to layout. We are guaranteed that we will be able to fill in a
1500 1516 // new interpreter frame as its callee (i.e. the stack space is allocated and
1501 1517 // the amount was determined by an earlier call to this method with f == NULL).
1502 1518 // On return f (if not NULL) while describe the interpreter frame we just layed out.
1503 1519
1504 1520 int monitor_size = moncount * frame::interpreter_frame_monitor_size();
1505 1521 int rounded_vm_local_words = round_to(frame::interpreter_frame_vm_local_words,WordsPerLong);
1506 1522
1507 1523 assert(monitor_size == round_to(monitor_size, WordsPerLong), "must align");
1508 1524 //
1509 1525 // Note: if you look closely this appears to be doing something much different
1510 1526 // than generate_fixed_frame. What is happening is this. On sparc we have to do
1511 1527 // this dance with interpreter_sp_adjustment because the window save area would
1512 1528 // appear just below the bottom (tos) of the caller's java expression stack. Because
1513 1529 // the interpreter want to have the locals completely contiguous generate_fixed_frame
1514 1530 // will adjust the caller's sp for the "extra locals" (max_locals - parameter_size).
1515 1531 // Now in generate_fixed_frame the extension of the caller's sp happens in the callee.
1516 1532 // In this code the opposite occurs the caller adjusts it's own stack base on the callee.
1517 1533 // This is mostly ok but it does cause a problem when we get to the initial frame (the oldest)
1518 1534 // because the oldest frame would have adjust its callers frame and yet that frame
1519 1535 // already exists and isn't part of this array of frames we are unpacking. So at first
1520 1536 // glance this would seem to mess up that frame. However Deoptimization::fetch_unroll_info_helper()
1521 1537 // will after it calculates all of the frame's on_stack_size()'s will then figure out the
1522 1538 // amount to adjust the caller of the initial (oldest) frame and the calculation will all
1523 1539 // add up. It does seem like it simpler to account for the adjustment here (and remove the
1524 1540 // callee... parameters here). However this would mean that this routine would have to take
1525 1541 // the caller frame as input so we could adjust its sp (and set it's interpreter_sp_adjustment)
1526 1542 // and run the calling loop in the reverse order. This would also would appear to mean making
1527 1543 // this code aware of what the interactions are when that initial caller fram was an osr or
1528 1544 // other adapter frame. deoptimization is complicated enough and hard enough to debug that
1529 1545 // there is no sense in messing working code.
1530 1546 //
1531 1547
1532 1548 int rounded_cls = round_to((callee_local_count - callee_param_count), WordsPerLong);
1533 1549 assert(rounded_cls == round_to(rounded_cls, WordsPerLong), "must align");
1534 1550
1535 1551 int raw_frame_size = size_activation_helper(rounded_cls, method->max_stack(),
1536 1552 monitor_size);
1537 1553
1538 1554 if (interpreter_frame != NULL) {
1539 1555 // The skeleton frame must already look like an interpreter frame
1540 1556 // even if not fully filled out.
1541 1557 assert(interpreter_frame->is_interpreted_frame(), "Must be interpreted frame");
1542 1558
1543 1559 intptr_t* fp = interpreter_frame->fp();
1544 1560
1545 1561 JavaThread* thread = JavaThread::current();
1546 1562 RegisterMap map(thread, false);
1547 1563 // More verification that skeleton frame is properly walkable
1548 1564 assert(fp == caller->sp(), "fp must match");
1549 1565
1550 1566 intptr_t* montop = fp - rounded_vm_local_words;
1551 1567
1552 1568 // preallocate monitors (cf. __ add_monitor_to_stack)
1553 1569 intptr_t* monitors = montop - monitor_size;
1554 1570
1555 1571 // preallocate stack space
1556 1572 intptr_t* esp = monitors - 1 -
1557 1573 (tempcount * Interpreter::stackElementWords()) -
1558 1574 popframe_extra_args;
1559 1575
1560 1576 int local_words = method->max_locals() * Interpreter::stackElementWords();
1561 1577 int parm_words = method->size_of_parameters() * Interpreter::stackElementWords();
1562 1578 NEEDS_CLEANUP;
1563 1579 intptr_t* locals;
1564 1580 if (caller->is_interpreted_frame()) {
1565 1581 // Can force the locals area to end up properly overlapping the top of the expression stack.
1566 1582 intptr_t* Lesp_ptr = caller->interpreter_frame_tos_address() - 1;
1567 1583 // Note that this computation means we replace size_of_parameters() values from the caller
1568 1584 // interpreter frame's expression stack with our argument locals
1569 1585 locals = Lesp_ptr + parm_words;
1570 1586 int delta = local_words - parm_words;
1571 1587 int computed_sp_adjustment = (delta > 0) ? round_to(delta, WordsPerLong) : 0;
1572 1588 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t) (fp + computed_sp_adjustment) - STACK_BIAS;
1573 1589 } else {
1574 1590 assert(caller->is_compiled_frame() || caller->is_entry_frame(), "only possible cases");
1575 1591 // Don't have Lesp available; lay out locals block in the caller
1576 1592 // adjacent to the register window save area.
1577 1593 //
1578 1594 // Compiled frames do not allocate a varargs area which is why this if
1579 1595 // statement is needed.
1580 1596 //
1581 1597 if (caller->is_compiled_frame()) {
1582 1598 locals = fp + frame::register_save_words + local_words - 1;
1583 1599 } else {
1584 1600 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
1585 1601 }
1586 1602 if (!caller->is_entry_frame()) {
1587 1603 // Caller wants his own SP back
1588 1604 int caller_frame_size = caller->cb()->frame_size();
1589 1605 *interpreter_frame->register_addr(I5_savedSP) = (intptr_t)(caller->fp() - caller_frame_size) - STACK_BIAS;
1590 1606 }
1591 1607 }
1592 1608 if (TraceDeoptimization) {
1593 1609 if (caller->is_entry_frame()) {
1594 1610 // make sure I5_savedSP and the entry frames notion of saved SP
1595 1611 // agree. This assertion duplicate a check in entry frame code
1596 1612 // but catches the failure earlier.
1597 1613 assert(*caller->register_addr(Lscratch) == *interpreter_frame->register_addr(I5_savedSP),
1598 1614 "would change callers SP");
1599 1615 }
1600 1616 if (caller->is_entry_frame()) {
1601 1617 tty->print("entry ");
1602 1618 }
1603 1619 if (caller->is_compiled_frame()) {
1604 1620 tty->print("compiled ");
1605 1621 if (caller->is_deoptimized_frame()) {
1606 1622 tty->print("(deopt) ");
1607 1623 }
1608 1624 }
1609 1625 if (caller->is_interpreted_frame()) {
1610 1626 tty->print("interpreted ");
1611 1627 }
1612 1628 tty->print_cr("caller fp=0x%x sp=0x%x", caller->fp(), caller->sp());
1613 1629 tty->print_cr("save area = 0x%x, 0x%x", caller->sp(), caller->sp() + 16);
1614 1630 tty->print_cr("save area = 0x%x, 0x%x", caller->fp(), caller->fp() + 16);
1615 1631 tty->print_cr("interpreter fp=0x%x sp=0x%x", interpreter_frame->fp(), interpreter_frame->sp());
1616 1632 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->sp(), interpreter_frame->sp() + 16);
1617 1633 tty->print_cr("save area = 0x%x, 0x%x", interpreter_frame->fp(), interpreter_frame->fp() + 16);
1618 1634 tty->print_cr("Llocals = 0x%x", locals);
1619 1635 tty->print_cr("Lesp = 0x%x", esp);
1620 1636 tty->print_cr("Lmonitors = 0x%x", monitors);
1621 1637 }
1622 1638
1623 1639 if (method->max_locals() > 0) {
1624 1640 assert(locals < caller->sp() || locals >= (caller->sp() + 16), "locals in save area");
1625 1641 assert(locals < caller->fp() || locals > (caller->fp() + 16), "locals in save area");
1626 1642 assert(locals < interpreter_frame->sp() || locals > (interpreter_frame->sp() + 16), "locals in save area");
1627 1643 assert(locals < interpreter_frame->fp() || locals >= (interpreter_frame->fp() + 16), "locals in save area");
1628 1644 }
1629 1645 #ifdef _LP64
1630 1646 assert(*interpreter_frame->register_addr(I5_savedSP) & 1, "must be odd");
1631 1647 #endif
1632 1648
1633 1649 *interpreter_frame->register_addr(Lmethod) = (intptr_t) method;
1634 1650 *interpreter_frame->register_addr(Llocals) = (intptr_t) locals;
1635 1651 *interpreter_frame->register_addr(Lmonitors) = (intptr_t) monitors;
1636 1652 *interpreter_frame->register_addr(Lesp) = (intptr_t) esp;
1637 1653 // Llast_SP will be same as SP as there is no adapter space
1638 1654 *interpreter_frame->register_addr(Llast_SP) = (intptr_t) interpreter_frame->sp() - STACK_BIAS;
1639 1655 *interpreter_frame->register_addr(LcpoolCache) = (intptr_t) method->constants()->cache();
1640 1656 #ifdef FAST_DISPATCH
1641 1657 *interpreter_frame->register_addr(IdispatchTables) = (intptr_t) Interpreter::dispatch_table();
1642 1658 #endif
1643 1659
1644 1660
1645 1661 #ifdef ASSERT
1646 1662 BasicObjectLock* mp = (BasicObjectLock*)monitors;
1647 1663
1648 1664 assert(interpreter_frame->interpreter_frame_method() == method, "method matches");
1649 1665 assert(interpreter_frame->interpreter_frame_local_at(9) == (intptr_t *)((intptr_t)locals - (9 * Interpreter::stackElementSize())+Interpreter::value_offset_in_bytes()), "locals match");
1650 1666 assert(interpreter_frame->interpreter_frame_monitor_end() == mp, "monitor_end matches");
1651 1667 assert(((intptr_t *)interpreter_frame->interpreter_frame_monitor_begin()) == ((intptr_t *)mp)+monitor_size, "monitor_begin matches");
1652 1668 assert(interpreter_frame->interpreter_frame_tos_address()-1 == esp, "esp matches");
1653 1669
1654 1670 // check bounds
1655 1671 intptr_t* lo = interpreter_frame->sp() + (frame::memory_parameter_word_sp_offset - 1);
1656 1672 intptr_t* hi = interpreter_frame->fp() - rounded_vm_local_words;
1657 1673 assert(lo < monitors && montop <= hi, "monitors in bounds");
1658 1674 assert(lo <= esp && esp < monitors, "esp in bounds");
1659 1675 #endif // ASSERT
1660 1676 }
1661 1677
1662 1678 return raw_frame_size;
1663 1679 }
1664 1680
1665 1681 //----------------------------------------------------------------------------------------------------
1666 1682 // Exceptions
1667 1683 void TemplateInterpreterGenerator::generate_throw_exception() {
1668 1684
1669 1685 // Entry point in previous activation (i.e., if the caller was interpreted)
1670 1686 Interpreter::_rethrow_exception_entry = __ pc();
1671 1687 // O0: exception
1672 1688
1673 1689 // entry point for exceptions thrown within interpreter code
1674 1690 Interpreter::_throw_exception_entry = __ pc();
1675 1691 __ verify_thread();
1676 1692 // expression stack is undefined here
1677 1693 // O0: exception, i.e. Oexception
1678 1694 // Lbcp: exception bcx
1679 1695 __ verify_oop(Oexception);
1680 1696
1681 1697
1682 1698 // expression stack must be empty before entering the VM in case of an exception
1683 1699 __ empty_expression_stack();
1684 1700 // find exception handler address and preserve exception oop
1685 1701 // call C routine to find handler and jump to it
1686 1702 __ call_VM(O1, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Oexception);
1687 1703 __ push_ptr(O1); // push exception for exception handler bytecodes
1688 1704
1689 1705 __ JMP(O0, 0); // jump to exception handler (may be remove activation entry!)
1690 1706 __ delayed()->nop();
1691 1707
1692 1708
1693 1709 // if the exception is not handled in the current frame
1694 1710 // the frame is removed and the exception is rethrown
1695 1711 // (i.e. exception continuation is _rethrow_exception)
1696 1712 //
1697 1713 // Note: At this point the bci is still the bxi for the instruction which caused
1698 1714 // the exception and the expression stack is empty. Thus, for any VM calls
1699 1715 // at this point, GC will find a legal oop map (with empty expression stack).
1700 1716
1701 1717 // in current activation
1702 1718 // tos: exception
1703 1719 // Lbcp: exception bcp
1704 1720
1705 1721 //
1706 1722 // JVMTI PopFrame support
1707 1723 //
1708 1724
1709 1725 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1710 1726 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1711 1727 // Set the popframe_processing bit in popframe_condition indicating that we are
1712 1728 // currently handling popframe, so that call_VMs that may happen later do not trigger new
1713 1729 // popframe handling cycles.
1714 1730
1715 1731 __ ld(popframe_condition_addr, G3_scratch);
1716 1732 __ or3(G3_scratch, JavaThread::popframe_processing_bit, G3_scratch);
1717 1733 __ stw(G3_scratch, popframe_condition_addr);
1718 1734
1719 1735 // Empty the expression stack, as in normal exception handling
1720 1736 __ empty_expression_stack();
1721 1737 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1722 1738
1723 1739 {
1724 1740 // Check to see whether we are returning to a deoptimized frame.
1725 1741 // (The PopFrame call ensures that the caller of the popped frame is
1726 1742 // either interpreted or compiled and deoptimizes it if compiled.)
1727 1743 // In this case, we can't call dispatch_next() after the frame is
1728 1744 // popped, but instead must save the incoming arguments and restore
1729 1745 // them after deoptimization has occurred.
1730 1746 //
1731 1747 // Note that we don't compare the return PC against the
1732 1748 // deoptimization blob's unpack entry because of the presence of
1733 1749 // adapter frames in C2.
1734 1750 Label caller_not_deoptimized;
1735 1751 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), I7);
1736 1752 __ tst(O0);
1737 1753 __ brx(Assembler::notEqual, false, Assembler::pt, caller_not_deoptimized);
1738 1754 __ delayed()->nop();
1739 1755
1740 1756 const Register Gtmp1 = G3_scratch;
1741 1757 const Register Gtmp2 = G1_scratch;
1742 1758
1743 1759 // Compute size of arguments for saving when returning to deoptimized caller
1744 1760 __ lduh(Lmethod, in_bytes(methodOopDesc::size_of_parameters_offset()), Gtmp1);
1745 1761 __ sll(Gtmp1, Interpreter::logStackElementSize(), Gtmp1);
1746 1762 __ sub(Llocals, Gtmp1, Gtmp2);
1747 1763 __ add(Gtmp2, wordSize, Gtmp2);
1748 1764 // Save these arguments
1749 1765 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), G2_thread, Gtmp1, Gtmp2);
1750 1766 // Inform deoptimization that it is responsible for restoring these arguments
1751 1767 __ set(JavaThread::popframe_force_deopt_reexecution_bit, Gtmp1);
1752 1768 Address popframe_condition_addr(G2_thread, JavaThread::popframe_condition_offset());
1753 1769 __ st(Gtmp1, popframe_condition_addr);
1754 1770
1755 1771 // Return from the current method
1756 1772 // The caller's SP was adjusted upon method entry to accomodate
1757 1773 // the callee's non-argument locals. Undo that adjustment.
1758 1774 __ ret();
1759 1775 __ delayed()->restore(I5_savedSP, G0, SP);
1760 1776
1761 1777 __ bind(caller_not_deoptimized);
1762 1778 }
1763 1779
1764 1780 // Clear the popframe condition flag
1765 1781 __ stw(G0 /* popframe_inactive */, popframe_condition_addr);
1766 1782
1767 1783 // Get out of the current method (how this is done depends on the particular compiler calling
1768 1784 // convention that the interpreter currently follows)
1769 1785 // The caller's SP was adjusted upon method entry to accomodate
1770 1786 // the callee's non-argument locals. Undo that adjustment.
1771 1787 __ restore(I5_savedSP, G0, SP);
1772 1788 // The method data pointer was incremented already during
1773 1789 // call profiling. We have to restore the mdp for the current bcp.
1774 1790 if (ProfileInterpreter) {
1775 1791 __ set_method_data_pointer_for_bcp();
1776 1792 }
1777 1793 // Resume bytecode interpretation at the current bcp
1778 1794 __ dispatch_next(vtos);
1779 1795 // end of JVMTI PopFrame support
1780 1796
1781 1797 Interpreter::_remove_activation_entry = __ pc();
1782 1798
1783 1799 // preserve exception over this code sequence (remove activation calls the vm, but oopmaps are not correct here)
1784 1800 __ pop_ptr(Oexception); // get exception
1785 1801
1786 1802 // Intel has the following comment:
1787 1803 //// remove the activation (without doing throws on illegalMonitorExceptions)
1788 1804 // They remove the activation without checking for bad monitor state.
1789 1805 // %%% We should make sure this is the right semantics before implementing.
1790 1806
1791 1807 // %%% changed set_vm_result_2 to set_vm_result and get_vm_result_2 to get_vm_result. Is there a bug here?
1792 1808 __ set_vm_result(Oexception);
1793 1809 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false);
1794 1810
1795 1811 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI);
1796 1812
1797 1813 __ get_vm_result(Oexception);
1798 1814 __ verify_oop(Oexception);
1799 1815
1800 1816 const int return_reg_adjustment = frame::pc_return_offset;
1801 1817 Address issuing_pc_addr(I7, return_reg_adjustment);
1802 1818
1803 1819 // We are done with this activation frame; find out where to go next.
1804 1820 // The continuation point will be an exception handler, which expects
1805 1821 // the following registers set up:
1806 1822 //
1807 1823 // Oexception: exception
1808 1824 // Oissuing_pc: the local call that threw exception
1809 1825 // Other On: garbage
1810 1826 // In/Ln: the contents of the caller's register window
1811 1827 //
1812 1828 // We do the required restore at the last possible moment, because we
1813 1829 // need to preserve some state across a runtime call.
1814 1830 // (Remember that the caller activation is unknown--it might not be
1815 1831 // interpreted, so things like Lscratch are useless in the caller.)
1816 1832
1817 1833 // Although the Intel version uses call_C, we can use the more
1818 1834 // compact call_VM. (The only real difference on SPARC is a
1819 1835 // harmlessly ignored [re]set_last_Java_frame, compared with
1820 1836 // the Intel code which lacks this.)
1821 1837 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
1822 1838 __ add(issuing_pc_addr, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
1823 1839 __ super_call_VM_leaf(L7_thread_cache,
1824 1840 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1825 1841 G2_thread, Oissuing_pc->after_save());
1826 1842
1827 1843 // The caller's SP was adjusted upon method entry to accomodate
1828 1844 // the callee's non-argument locals. Undo that adjustment.
1829 1845 __ JMP(O0, 0); // return exception handler in caller
1830 1846 __ delayed()->restore(I5_savedSP, G0, SP);
1831 1847
1832 1848 // (same old exception object is already in Oexception; see above)
1833 1849 // Note that an "issuing PC" is actually the next PC after the call
1834 1850 }
1835 1851
1836 1852
1837 1853 //
1838 1854 // JVMTI ForceEarlyReturn support
1839 1855 //
1840 1856
1841 1857 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1842 1858 address entry = __ pc();
1843 1859
1844 1860 __ empty_expression_stack();
1845 1861 __ load_earlyret_value(state);
1846 1862
1847 1863 __ ld_ptr(G2_thread, JavaThread::jvmti_thread_state_offset(), G3_scratch);
1848 1864 Address cond_addr(G3_scratch, JvmtiThreadState::earlyret_state_offset());
1849 1865
1850 1866 // Clear the earlyret state
1851 1867 __ stw(G0 /* JvmtiThreadState::earlyret_inactive */, cond_addr);
1852 1868
1853 1869 __ remove_activation(state,
1854 1870 /* throw_monitor_exception */ false,
1855 1871 /* install_monitor_exception */ false);
1856 1872
1857 1873 // The caller's SP was adjusted upon method entry to accomodate
1858 1874 // the callee's non-argument locals. Undo that adjustment.
1859 1875 __ ret(); // return to caller
1860 1876 __ delayed()->restore(I5_savedSP, G0, SP);
1861 1877
1862 1878 return entry;
1863 1879 } // end of JVMTI ForceEarlyReturn support
1864 1880
1865 1881
1866 1882 //------------------------------------------------------------------------------------------------------------------------
1867 1883 // Helper for vtos entry point generation
1868 1884
1869 1885 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) {
1870 1886 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1871 1887 Label L;
1872 1888 aep = __ pc(); __ push_ptr(); __ ba(false, L); __ delayed()->nop();
1873 1889 fep = __ pc(); __ push_f(); __ ba(false, L); __ delayed()->nop();
1874 1890 dep = __ pc(); __ push_d(); __ ba(false, L); __ delayed()->nop();
1875 1891 lep = __ pc(); __ push_l(); __ ba(false, L); __ delayed()->nop();
1876 1892 iep = __ pc(); __ push_i();
1877 1893 bep = cep = sep = iep; // there aren't any
1878 1894 vep = __ pc(); __ bind(L); // fall through
1879 1895 generate_and_dispatch(t);
1880 1896 }
1881 1897
1882 1898 // --------------------------------------------------------------------------------
1883 1899
1884 1900
1885 1901 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1886 1902 : TemplateInterpreterGenerator(code) {
1887 1903 generate_all(); // down here so it can be "virtual"
1888 1904 }
1889 1905
1890 1906 // --------------------------------------------------------------------------------
1891 1907
1892 1908 // Non-product code
1893 1909 #ifndef PRODUCT
1894 1910 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1895 1911 address entry = __ pc();
1896 1912
1897 1913 __ push(state);
1898 1914 __ mov(O7, Lscratch); // protect return address within interpreter
1899 1915
1900 1916 // Pass a 0 (not used in sparc) and the top of stack to the bytecode tracer
1901 1917 __ mov( Otos_l2, G3_scratch );
1902 1918 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), G0, Otos_l1, G3_scratch);
1903 1919 __ mov(Lscratch, O7); // restore return address
1904 1920 __ pop(state);
1905 1921 __ retl();
1906 1922 __ delayed()->nop();
1907 1923
1908 1924 return entry;
1909 1925 }
1910 1926
1911 1927
1912 1928 // helpers for generate_and_dispatch
1913 1929
1914 1930 void TemplateInterpreterGenerator::count_bytecode() {
1915 1931 __ inc_counter(&BytecodeCounter::_counter_value, G3_scratch, G4_scratch);
1916 1932 }
1917 1933
1918 1934
1919 1935 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1920 1936 __ inc_counter(&BytecodeHistogram::_counters[t->bytecode()], G3_scratch, G4_scratch);
1921 1937 }
1922 1938
1923 1939
1924 1940 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1925 1941 AddressLiteral index (&BytecodePairHistogram::_index);
1926 1942 AddressLiteral counters((address) &BytecodePairHistogram::_counters);
1927 1943
1928 1944 // get index, shift out old bytecode, bring in new bytecode, and store it
1929 1945 // _index = (_index >> log2_number_of_codes) |
1930 1946 // (bytecode << log2_number_of_codes);
1931 1947
1932 1948 __ load_contents(index, G4_scratch);
1933 1949 __ srl( G4_scratch, BytecodePairHistogram::log2_number_of_codes, G4_scratch );
1934 1950 __ set( ((int)t->bytecode()) << BytecodePairHistogram::log2_number_of_codes, G3_scratch );
1935 1951 __ or3( G3_scratch, G4_scratch, G4_scratch );
1936 1952 __ store_contents(G4_scratch, index, G3_scratch);
1937 1953
1938 1954 // bump bucket contents
1939 1955 // _counters[_index] ++;
1940 1956
1941 1957 __ set(counters, G3_scratch); // loads into G3_scratch
1942 1958 __ sll( G4_scratch, LogBytesPerWord, G4_scratch ); // Index is word address
1943 1959 __ add (G3_scratch, G4_scratch, G3_scratch); // Add in index
1944 1960 __ ld (G3_scratch, 0, G4_scratch);
1945 1961 __ inc (G4_scratch);
1946 1962 __ st (G4_scratch, 0, G3_scratch);
1947 1963 }
1948 1964
1949 1965
1950 1966 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
1951 1967 // Call a little run-time stub to avoid blow-up for each bytecode.
1952 1968 // The run-time runtime saves the right registers, depending on
1953 1969 // the tosca in-state for the given template.
1954 1970 address entry = Interpreter::trace_code(t->tos_in());
1955 1971 guarantee(entry != NULL, "entry must have been generated");
1956 1972 __ call(entry, relocInfo::none);
1957 1973 __ delayed()->nop();
1958 1974 }
1959 1975
1960 1976
1961 1977 void TemplateInterpreterGenerator::stop_interpreter_at() {
1962 1978 AddressLiteral counter(&BytecodeCounter::_counter_value);
1963 1979 __ load_contents(counter, G3_scratch);
1964 1980 AddressLiteral stop_at(&StopInterpreterAt);
1965 1981 __ load_ptr_contents(stop_at, G4_scratch);
1966 1982 __ cmp(G3_scratch, G4_scratch);
1967 1983 __ breakpoint_trap(Assembler::equal);
1968 1984 }
1969 1985 #endif // not PRODUCT
1970 1986 #endif // !CC_INTERP
↓ open down ↓ |
1765 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX