1 /*
2 * Copyright (c) 2007, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "interpreter/bytecodeHistogram.hpp"
28 #include "interpreter/cppInterpreter.hpp"
29 #include "interpreter/interpreter.hpp"
30 #include "interpreter/interpreterGenerator.hpp"
31 #include "interpreter/interpreterRuntime.hpp"
32 #include "oops/arrayOop.hpp"
33 #include "oops/methodData.hpp"
34 #include "oops/method.hpp"
35 #include "oops/oop.inline.hpp"
36 #include "prims/jvmtiExport.hpp"
37 #include "prims/jvmtiThreadState.hpp"
38 #include "runtime/arguments.hpp"
39 #include "runtime/deoptimization.hpp"
40 #include "runtime/frame.inline.hpp"
41 #include "runtime/interfaceSupport.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "runtime/stubRoutines.hpp"
44 #include "runtime/synchronizer.hpp"
45 #include "runtime/timer.hpp"
46 #include "runtime/vframeArray.hpp"
47 #include "utilities/debug.hpp"
48 #include "utilities/macros.hpp"
49 #ifdef SHARK
50 #include "shark/shark_globals.hpp"
51 #endif
52
53 #ifdef CC_INTERP
54
55 // Routine exists to make tracebacks look decent in debugger
56 // while "shadow" interpreter frames are on stack. It is also
57 // used to distinguish interpreter frames.
58
RecursiveInterpreterActivation(interpreterState istate)59 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
60 ShouldNotReachHere();
61 }
62
contains(address pc)63 bool CppInterpreter::contains(address pc) {
64 return ( _code->contains(pc) ||
65 ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
66 }
67
68 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
69 #define __ _masm->
70
71 Label frame_manager_entry;
72 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized
73 // c++ interpreter entry point this holds that entry point label.
74
75 static address unctrap_frame_manager_entry = NULL;
76
77 static address interpreter_return_address = NULL;
78 static address deopt_frame_manager_return_atos = NULL;
79 static address deopt_frame_manager_return_btos = NULL;
80 static address deopt_frame_manager_return_itos = NULL;
81 static address deopt_frame_manager_return_ltos = NULL;
82 static address deopt_frame_manager_return_ftos = NULL;
83 static address deopt_frame_manager_return_dtos = NULL;
84 static address deopt_frame_manager_return_vtos = NULL;
85
86 const Register prevState = G1_scratch;
87
save_native_result(void)88 void InterpreterGenerator::save_native_result(void) {
89 // result potentially in O0/O1: save it across calls
90 __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
91 #ifdef _LP64
92 __ stx(O0, STATE(_native_lresult));
93 #else
94 __ std(O0, STATE(_native_lresult));
95 #endif
96 }
97
restore_native_result(void)98 void InterpreterGenerator::restore_native_result(void) {
99
100 // Restore any method result value
101 __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
102 #ifdef _LP64
103 __ ldx(STATE(_native_lresult), O0);
104 #else
105 __ ldd(STATE(_native_lresult), O0);
106 #endif
107 }
108
109 // A result handler converts/unboxes a native call result into
110 // a java interpreter/compiler result. The current frame is an
111 // interpreter frame. The activation frame unwind code must be
112 // consistent with that of TemplateTable::_return(...). In the
113 // case of native methods, the caller's SP was not modified.
generate_result_handler_for(BasicType type)114 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
115 address entry = __ pc();
116 Register Itos_i = Otos_i ->after_save();
117 Register Itos_l = Otos_l ->after_save();
118 Register Itos_l1 = Otos_l1->after_save();
119 Register Itos_l2 = Otos_l2->after_save();
120 switch (type) {
121 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
122 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
123 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
124 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
125 case T_LONG :
126 #ifndef _LP64
127 __ mov(O1, Itos_l2); // move other half of long
128 #endif // ifdef or no ifdef, fall through to the T_INT case
129 case T_INT : __ mov(O0, Itos_i); break;
130 case T_VOID : /* nothing to do */ break;
131 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
132 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
133 case T_OBJECT :
134 __ ld_ptr(STATE(_oop_temp), Itos_i);
135 __ verify_oop(Itos_i);
136 break;
137 default : ShouldNotReachHere();
138 }
139 __ ret(); // return from interpreter activation
140 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
141 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
142 return entry;
143 }
144
145 // tosca based result to c++ interpreter stack based result.
146 // Result goes to address in L1_scratch
147
generate_tosca_to_stack_converter(BasicType type)148 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
149 // A result is in the native abi result register from a native method call.
150 // We need to return this result to the interpreter by pushing the result on the interpreter's
151 // stack. This is relatively simple the destination is in L1_scratch
152 // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
153 // adjust L1_scratch
154 address entry = __ pc();
155 switch (type) {
156 case T_BOOLEAN:
157 // !0 => true; 0 => false
158 __ subcc(G0, O0, G0);
159 __ addc(G0, 0, O0);
160 __ st(O0, L1_scratch, 0);
161 __ sub(L1_scratch, wordSize, L1_scratch);
162 break;
163
164 // cannot use and3, 0xFFFF too big as immediate value!
165 case T_CHAR :
166 __ sll(O0, 16, O0);
167 __ srl(O0, 16, O0);
168 __ st(O0, L1_scratch, 0);
169 __ sub(L1_scratch, wordSize, L1_scratch);
170 break;
171
172 case T_BYTE :
173 __ sll(O0, 24, O0);
174 __ sra(O0, 24, O0);
175 __ st(O0, L1_scratch, 0);
176 __ sub(L1_scratch, wordSize, L1_scratch);
177 break;
178
179 case T_SHORT :
180 __ sll(O0, 16, O0);
181 __ sra(O0, 16, O0);
182 __ st(O0, L1_scratch, 0);
183 __ sub(L1_scratch, wordSize, L1_scratch);
184 break;
185 case T_LONG :
186 #ifndef _LP64
187 #if defined(COMPILER2)
188 // All return values are where we want them, except for Longs. C2 returns
189 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
190 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
191 // build even if we are returning from interpreted we just do a little
192 // stupid shuffing.
193 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
194 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
195 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
196 __ stx(G1, L1_scratch, -wordSize);
197 #else
198 // native result is in O0, O1
199 __ st(O1, L1_scratch, 0); // Low order
200 __ st(O0, L1_scratch, -wordSize); // High order
201 #endif /* COMPILER2 */
202 #else
203 __ stx(O0, L1_scratch, -wordSize);
204 #endif
205 __ sub(L1_scratch, 2*wordSize, L1_scratch);
206 break;
207
208 case T_INT :
209 __ st(O0, L1_scratch, 0);
210 __ sub(L1_scratch, wordSize, L1_scratch);
211 break;
212
213 case T_VOID : /* nothing to do */
214 break;
215
216 case T_FLOAT :
217 __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
218 __ sub(L1_scratch, wordSize, L1_scratch);
219 break;
220
221 case T_DOUBLE :
222 // Every stack slot is aligned on 64 bit, However is this
223 // the correct stack slot on 64bit?? QQQ
224 __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
225 __ sub(L1_scratch, 2*wordSize, L1_scratch);
226 break;
227 case T_OBJECT :
228 __ verify_oop(O0);
229 __ st_ptr(O0, L1_scratch, 0);
230 __ sub(L1_scratch, wordSize, L1_scratch);
231 break;
232 default : ShouldNotReachHere();
233 }
234 __ retl(); // return from interpreter activation
235 __ delayed()->nop(); // schedule this better
236 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
237 return entry;
238 }
239
generate_stack_to_stack_converter(BasicType type)240 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
241 // A result is in the java expression stack of the interpreted method that has just
242 // returned. Place this result on the java expression stack of the caller.
243 //
244 // The current interpreter activation in Lstate is for the method just returning its
245 // result. So we know that the result of this method is on the top of the current
246 // execution stack (which is pre-pushed) and will be return to the top of the caller
247 // stack. The top of the callers stack is the bottom of the locals of the current
248 // activation.
249 // Because of the way activation are managed by the frame manager the value of esp is
250 // below both the stack top of the current activation and naturally the stack top
251 // of the calling activation. This enable this routine to leave the return address
252 // to the frame manager on the stack and do a vanilla return.
253 //
254 // On entry: O0 - points to source (callee stack top)
255 // O1 - points to destination (caller stack top [i.e. free location])
256 // destroys O2, O3
257 //
258
259 address entry = __ pc();
260 switch (type) {
261 case T_VOID: break;
262 break;
263 case T_FLOAT :
264 case T_BOOLEAN:
265 case T_CHAR :
266 case T_BYTE :
267 case T_SHORT :
268 case T_INT :
269 // 1 word result
270 __ ld(O0, 0, O2);
271 __ st(O2, O1, 0);
272 __ sub(O1, wordSize, O1);
273 break;
274 case T_DOUBLE :
275 case T_LONG :
276 // return top two words on current expression stack to caller's expression stack
277 // The caller's expression stack is adjacent to the current frame manager's intepretState
278 // except we allocated one extra word for this intepretState so we won't overwrite it
279 // when we return a two word result.
280 #ifdef _LP64
281 __ ld_ptr(O0, 0, O2);
282 __ st_ptr(O2, O1, -wordSize);
283 #else
284 __ ld(O0, 0, O2);
285 __ ld(O0, wordSize, O3);
286 __ st(O3, O1, 0);
287 __ st(O2, O1, -wordSize);
288 #endif
289 __ sub(O1, 2*wordSize, O1);
290 break;
291 case T_OBJECT :
292 __ ld_ptr(O0, 0, O2);
293 __ verify_oop(O2); // verify it
294 __ st_ptr(O2, O1, 0);
295 __ sub(O1, wordSize, O1);
296 break;
297 default : ShouldNotReachHere();
298 }
299 __ retl();
300 __ delayed()->nop(); // QQ schedule this better
301 return entry;
302 }
303
generate_stack_to_native_abi_converter(BasicType type)304 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
305 // A result is in the java expression stack of the interpreted method that has just
306 // returned. Place this result in the native abi that the caller expects.
307 // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
308 //
309 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
310 // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
311 // and so rather than return result onto caller's java expression stack we return the
312 // result in the expected location based on the native abi.
313 // On entry: O0 - source (stack top)
314 // On exit result in expected output register
315 // QQQ schedule this better
316
317 address entry = __ pc();
318 switch (type) {
319 case T_VOID: break;
320 break;
321 case T_FLOAT :
322 __ ldf(FloatRegisterImpl::S, O0, 0, F0);
323 break;
324 case T_BOOLEAN:
325 case T_CHAR :
326 case T_BYTE :
327 case T_SHORT :
328 case T_INT :
329 // 1 word result
330 __ ld(O0, 0, O0->after_save());
331 break;
332 case T_DOUBLE :
333 __ ldf(FloatRegisterImpl::D, O0, 0, F0);
334 break;
335 case T_LONG :
336 // return top two words on current expression stack to caller's expression stack
337 // The caller's expression stack is adjacent to the current frame manager's interpretState
338 // except we allocated one extra word for this intepretState so we won't overwrite it
339 // when we return a two word result.
340 #ifdef _LP64
341 __ ld_ptr(O0, 0, O0->after_save());
342 #else
343 __ ld(O0, wordSize, O1->after_save());
344 __ ld(O0, 0, O0->after_save());
345 #endif
346 #if defined(COMPILER2) && !defined(_LP64)
347 // C2 expects long results in G1 we can't tell if we're returning to interpreted
348 // or compiled so just be safe use G1 and O0/O1
349
350 // Shift bits into high (msb) of G1
351 __ sllx(Otos_l1->after_save(), 32, G1);
352 // Zero extend low bits
353 __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
354 __ or3 (Otos_l2->after_save(), G1, G1);
355 #endif /* COMPILER2 */
356 break;
357 case T_OBJECT :
358 __ ld_ptr(O0, 0, O0->after_save());
359 __ verify_oop(O0->after_save()); // verify it
360 break;
361 default : ShouldNotReachHere();
362 }
363 __ retl();
364 __ delayed()->nop();
365 return entry;
366 }
367
return_entry(TosState state,int length,Bytecodes::Code code)368 address CppInterpreter::return_entry(TosState state, int length, Bytecodes::Code code) {
369 // make it look good in the debugger
370 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
371 }
372
deopt_entry(TosState state,int length)373 address CppInterpreter::deopt_entry(TosState state, int length) {
374 address ret = NULL;
375 if (length != 0) {
376 switch (state) {
377 case atos: ret = deopt_frame_manager_return_atos; break;
378 case btos: ret = deopt_frame_manager_return_btos; break;
379 case ctos:
380 case stos:
381 case itos: ret = deopt_frame_manager_return_itos; break;
382 case ltos: ret = deopt_frame_manager_return_ltos; break;
383 case ftos: ret = deopt_frame_manager_return_ftos; break;
384 case dtos: ret = deopt_frame_manager_return_dtos; break;
385 case vtos: ret = deopt_frame_manager_return_vtos; break;
386 }
387 } else {
388 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap)
389 }
390 assert(ret != NULL, "Not initialized");
391 return ret;
392 }
393
394 //
395 // Helpers for commoning out cases in the various type of method entries.
396 //
397
398 // increment invocation count & check for overflow
399 //
400 // Note: checking for negative value instead of overflow
401 // so we have a 'sticky' overflow test
402 //
403 // Lmethod: method
404 // ??: invocation counter
405 //
generate_counter_incr(Label * overflow,Label * profile_method,Label * profile_method_continue)406 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
407 Label done;
408 const Register Rcounters = G3_scratch;
409
410 __ ld_ptr(STATE(_method), G5_method);
411 __ get_method_counters(G5_method, Rcounters, done);
412
413 // Update standard invocation counters
414 __ increment_invocation_counter(Rcounters, O0, G4_scratch);
415 if (ProfileInterpreter) {
416 Address interpreter_invocation_counter(Rcounters, 0,
417 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
418 __ ld(interpreter_invocation_counter, G4_scratch);
419 __ inc(G4_scratch);
420 __ st(G4_scratch, interpreter_invocation_counter);
421 }
422
423 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
424 __ sethi(invocation_limit);
425 __ ld(invocation_limit, G3_scratch);
426 __ cmp(O0, G3_scratch);
427 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
428 __ delayed()->nop();
429 __ bind(done);
430 }
431
generate_empty_entry(void)432 address InterpreterGenerator::generate_empty_entry(void) {
433
434 // A method that does nothing but return...
435
436 address entry = __ pc();
437 Label slow_path;
438
439 // do nothing for empty methods (do not even increment invocation counter)
440 if ( UseFastEmptyMethods) {
441 // If we need a safepoint check, generate full interpreter entry.
442 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
443 __ load_contents(sync_state, G3_scratch);
444 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
445 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
446 __ delayed()->nop();
447
448 // Code: _return
449 __ retl();
450 __ delayed()->mov(O5_savedSP, SP);
451 return entry;
452 }
453 return NULL;
454 }
455
456 // Call an accessor method (assuming it is resolved, otherwise drop into
457 // vanilla (slow path) entry
458
459 // Generates code to elide accessor methods
460 // Uses G3_scratch and G1_scratch as scratch
generate_accessor_entry(void)461 address InterpreterGenerator::generate_accessor_entry(void) {
462
463 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
464 // parameter size = 1
465 // Note: We can only use this code if the getfield has been resolved
466 // and if we don't have a null-pointer exception => check for
467 // these conditions first and use slow path if necessary.
468 address entry = __ pc();
469 Label slow_path;
470
471 if ( UseFastAccessorMethods) {
472 // Check if we need to reach a safepoint and generate full interpreter
473 // frame if so.
474 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
475 __ load_contents(sync_state, G3_scratch);
476 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
477 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
478 __ delayed()->nop();
479
480 // Check if local 0 != NULL
481 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
482 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
483 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
484 __ delayed()->nop();
485
486
487 // read first instruction word and extract bytecode @ 1 and index @ 2
488 // get first 4 bytes of the bytecodes (big endian!)
489 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::const_offset())), G1_scratch);
490 __ ld(Address(G1_scratch, 0, in_bytes(ConstMethod::codes_offset())), G1_scratch);
491
492 // move index @ 2 far left then to the right most two bytes.
493 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
494 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
495 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
496
497 // get constant pool cache
498 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), G3_scratch);
499 __ ld_ptr(G3_scratch, in_bytes(ConstMethod::constants_offset()), G3_scratch);
500 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
501
502 // get specific constant pool cache entry
503 __ add(G3_scratch, G1_scratch, G3_scratch);
504
505 // Check the constant Pool cache entry to see if it has been resolved.
506 // If not, need the slow path.
507 ByteSize cp_base_offset = ConstantPoolCache::base_offset();
508 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
509 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
510 __ and3(G1_scratch, 0xFF, G1_scratch);
511 __ cmp(G1_scratch, Bytecodes::_getfield);
512 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
513 __ delayed()->nop();
514
515 // Get the type and return field offset from the constant pool cache
516 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
517 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
518
519 Label xreturn_path;
520 // Need to differentiate between igetfield, agetfield, bgetfield etc.
521 // because they are different sizes.
522 // Get the type from the constant pool cache
523 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
524 // Make sure we don't need to mask G1_scratch after the above shift
525 ConstantPoolCacheEntry::verify_tos_state_shift();
526 __ cmp(G1_scratch, atos );
527 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
528 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
529 __ cmp(G1_scratch, itos);
530 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
531 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
532 __ cmp(G1_scratch, stos);
533 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
534 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
535 __ cmp(G1_scratch, ctos);
536 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
537 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
538 #ifdef ASSERT
539 __ cmp(G1_scratch, btos);
540 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
541 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
542 __ cmp(G1_scratch, ztos);
543 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
544 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
545 __ should_not_reach_here();
546 #endif
547 __ ldsb(Otos_i, G3_scratch, Otos_i);
548 __ bind(xreturn_path);
549
550 // _ireturn/_areturn
551 __ retl(); // return from leaf routine
552 __ delayed()->mov(O5_savedSP, SP);
553
554 // Generate regular method entry
555 __ bind(slow_path);
556 __ ba(fast_accessor_slow_entry_path);
557 __ delayed()->nop();
558 return entry;
559 }
560 return NULL;
561 }
562
generate_Reference_get_entry(void)563 address InterpreterGenerator::generate_Reference_get_entry(void) {
564 #if INCLUDE_ALL_GCS
565 if (UseG1GC) {
566 // We need to generate have a routine that generates code to:
567 // * load the value in the referent field
568 // * passes that value to the pre-barrier.
569 //
570 // In the case of G1 this will record the value of the
571 // referent in an SATB buffer if marking is active.
572 // This will cause concurrent marking to mark the referent
573 // field as live.
574 Unimplemented();
575 }
576 #endif // INCLUDE_ALL_GCS
577
578 // If G1 is not enabled then attempt to go through the accessor entry point
579 // Reference.get is an accessor
580 return generate_accessor_entry();
581 }
582
583 //
584 // Interpreter stub for calling a native method. (C++ interpreter)
585 // This sets up a somewhat different looking stack for calling the native method
586 // than the typical interpreter frame setup.
587 //
588
generate_native_entry(bool synchronized)589 address InterpreterGenerator::generate_native_entry(bool synchronized) {
590 address entry = __ pc();
591
592 // the following temporary registers are used during frame creation
593 const Register Gtmp1 = G3_scratch ;
594 const Register Gtmp2 = G1_scratch;
595 const Register RconstMethod = Gtmp1;
596 const Address constMethod(G5_method, 0, in_bytes(Method::const_offset()));
597 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
598
599 bool inc_counter = UseCompiler || CountCompiledCalls;
600
601 // make sure registers are different!
602 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
603
604 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
605
606 Label Lentry;
607 __ bind(Lentry);
608
609 const Register Glocals_size = G3;
610 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
611
612 // make sure method is native & not abstract
613 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
614 #ifdef ASSERT
615 __ ld(access_flags, Gtmp1);
616 {
617 Label L;
618 __ btst(JVM_ACC_NATIVE, Gtmp1);
619 __ br(Assembler::notZero, false, Assembler::pt, L);
620 __ delayed()->nop();
621 __ stop("tried to execute non-native method as native");
622 __ bind(L);
623 }
624 { Label L;
625 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
626 __ br(Assembler::zero, false, Assembler::pt, L);
627 __ delayed()->nop();
628 __ stop("tried to execute abstract method as non-abstract");
629 __ bind(L);
630 }
631 #endif // ASSERT
632
633 __ ld_ptr(constMethod, RconstMethod);
634 __ lduh(size_of_parameters, Gtmp1);
635 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes
636 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord
637 // NEW
638 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
639 // generate the code to allocate the interpreter stack frame
640 // NEW FRAME ALLOCATED HERE
641 // save callers original sp
642 // __ mov(SP, I5_savedSP->after_restore());
643
644 generate_compute_interpreter_state(Lstate, G0, true);
645
646 // At this point Lstate points to new interpreter state
647 //
648
649 const Address do_not_unlock_if_synchronized(G2_thread, 0,
650 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
651 // Since at this point in the method invocation the exception handler
652 // would try to exit the monitor of synchronized methods which hasn't
653 // been entered yet, we set the thread local variable
654 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
655 // runtime, exception handling i.e. unlock_if_synchronized_method will
656 // check this thread local flag.
657 // This flag has two effects, one is to force an unwind in the topmost
658 // interpreter frame and not perform an unlock while doing so.
659
660 __ movbool(true, G3_scratch);
661 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
662
663
664 // increment invocation counter and check for overflow
665 //
666 // Note: checking for negative value instead of overflow
667 // so we have a 'sticky' overflow test (may be of
668 // importance as soon as we have true MT/MP)
669 Label invocation_counter_overflow;
670 if (inc_counter) {
671 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
672 }
673 Label Lcontinue;
674 __ bind(Lcontinue);
675
676 bang_stack_shadow_pages(true);
677 // reset the _do_not_unlock_if_synchronized flag
678 __ stbool(G0, do_not_unlock_if_synchronized);
679
680 // check for synchronized methods
681 // Must happen AFTER invocation_counter check, so method is not locked
682 // if counter overflows.
683
684 if (synchronized) {
685 lock_method();
686 // Don't see how G2_thread is preserved here...
687 // __ verify_thread(); QQQ destroys L0,L1 can't use
688 } else {
689 #ifdef ASSERT
690 { Label ok;
691 __ ld_ptr(STATE(_method), G5_method);
692 __ ld(access_flags, O0);
693 __ btst(JVM_ACC_SYNCHRONIZED, O0);
694 __ br( Assembler::zero, false, Assembler::pt, ok);
695 __ delayed()->nop();
696 __ stop("method needs synchronization");
697 __ bind(ok);
698 }
699 #endif // ASSERT
700 }
701
702 // start execution
703
704 // __ verify_thread(); kills L1,L2 can't use at the moment
705
706 // jvmti/jvmpi support
707 __ notify_method_entry();
708
709 // native call
710
711 // (note that O0 is never an oop--at most it is a handle)
712 // It is important not to smash any handles created by this call,
713 // until any oop handle in O0 is dereferenced.
714
715 // (note that the space for outgoing params is preallocated)
716
717 // get signature handler
718
719 Label pending_exception_present;
720
721 { Label L;
722 __ ld_ptr(STATE(_method), G5_method);
723 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
724 __ tst(G3_scratch);
725 __ brx(Assembler::notZero, false, Assembler::pt, L);
726 __ delayed()->nop();
727 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
728 __ ld_ptr(STATE(_method), G5_method);
729
730 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
731 __ ld_ptr(exception_addr, G3_scratch);
732 __ br_notnull_short(G3_scratch, Assembler::pn, pending_exception_present);
733 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
734 __ bind(L);
735 }
736
737 // Push a new frame so that the args will really be stored in
738 // Copy a few locals across so the new frame has the variables
739 // we need but these values will be dead at the jni call and
740 // therefore not gc volatile like the values in the current
741 // frame (Lstate in particular)
742
743 // Flush the state pointer to the register save area
744 // Which is the only register we need for a stack walk.
745 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
746
747 __ mov(Lstate, O1); // Need to pass the state pointer across the frame
748
749 // Calculate current frame size
750 __ sub(SP, FP, O3); // Calculate negative of current frame size
751 __ save(SP, O3, SP); // Allocate an identical sized frame
752
753 __ mov(I1, Lstate); // In the "natural" register.
754
755 // Note I7 has leftover trash. Slow signature handler will fill it in
756 // should we get there. Normal jni call will set reasonable last_Java_pc
757 // below (and fix I7 so the stack trace doesn't have a meaningless frame
758 // in it).
759
760
761 // call signature handler
762 __ ld_ptr(STATE(_method), Lmethod);
763 __ ld_ptr(STATE(_locals), Llocals);
764
765 __ callr(G3_scratch, 0);
766 __ delayed()->nop();
767 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed)
768
769 { Label not_static;
770
771 __ ld_ptr(STATE(_method), G5_method);
772 __ ld(access_flags, O0);
773 __ btst(JVM_ACC_STATIC, O0);
774 __ br( Assembler::zero, false, Assembler::pt, not_static);
775 __ delayed()->
776 // get native function entry point(O0 is a good temp until the very end)
777 ld_ptr(Address(G5_method, 0, in_bytes(Method::native_function_offset())), O0);
778 // for static methods insert the mirror argument
779 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
780
781 __ ld_ptr(Address(G5_method, 0, in_bytes(Method:: const_offset())), O1);
782 __ ld_ptr(Address(O1, 0, in_bytes(ConstMethod::constants_offset())), O1);
783 __ ld_ptr(Address(O1, 0, ConstantPool::pool_holder_offset_in_bytes()), O1);
784 __ ld_ptr(O1, mirror_offset, O1);
785 // where the mirror handle body is allocated:
786 #ifdef ASSERT
787 if (!PrintSignatureHandlers) // do not dirty the output with this
788 { Label L;
789 __ tst(O1);
790 __ brx(Assembler::notZero, false, Assembler::pt, L);
791 __ delayed()->nop();
792 __ stop("mirror is missing");
793 __ bind(L);
794 }
795 #endif // ASSERT
796 __ st_ptr(O1, STATE(_oop_temp));
797 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add
798 __ bind(not_static);
799 }
800
801 // At this point, arguments have been copied off of stack into
802 // their JNI positions, which are O1..O5 and SP[68..].
803 // Oops are boxed in-place on the stack, with handles copied to arguments.
804 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
805
806 #ifdef ASSERT
807 { Label L;
808 __ tst(O0);
809 __ brx(Assembler::notZero, false, Assembler::pt, L);
810 __ delayed()->nop();
811 __ stop("native entry point is missing");
812 __ bind(L);
813 }
814 #endif // ASSERT
815
816 //
817 // setup the java frame anchor
818 //
819 // The scavenge function only needs to know that the PC of this frame is
820 // in the interpreter method entry code, it doesn't need to know the exact
821 // PC and hence we can use O7 which points to the return address from the
822 // previous call in the code stream (signature handler function)
823 //
824 // The other trick is we set last_Java_sp to FP instead of the usual SP because
825 // we have pushed the extra frame in order to protect the volatile register(s)
826 // in that frame when we return from the jni call
827 //
828
829
830 __ set_last_Java_frame(FP, O7);
831 __ mov(O7, I7); // make dummy interpreter frame look like one above,
832 // not meaningless information that'll confuse me.
833
834 // flush the windows now. We don't care about the current (protection) frame
835 // only the outer frames
836
837 __ flush_windows();
838
839 // mark windows as flushed
840 Address flags(G2_thread,
841 0,
842 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
843 __ set(JavaFrameAnchor::flushed, G3_scratch);
844 __ st(G3_scratch, flags);
845
846 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
847
848 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
849 #ifdef ASSERT
850 { Label L;
851 __ ld(thread_state, G3_scratch);
852 __ cmp(G3_scratch, _thread_in_Java);
853 __ br(Assembler::equal, false, Assembler::pt, L);
854 __ delayed()->nop();
855 __ stop("Wrong thread state in native stub");
856 __ bind(L);
857 }
858 #endif // ASSERT
859 __ set(_thread_in_native, G3_scratch);
860 __ st(G3_scratch, thread_state);
861
862 // Call the jni method, using the delay slot to set the JNIEnv* argument.
863 __ callr(O0, 0);
864 __ delayed()->
865 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
866 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
867
868 // must we block?
869
870 // Block, if necessary, before resuming in _thread_in_Java state.
871 // In order for GC to work, don't clear the last_Java_sp until after blocking.
872 { Label no_block;
873 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
874
875 // Switch thread to "native transition" state before reading the synchronization state.
876 // This additional state is necessary because reading and testing the synchronization
877 // state is not atomic w.r.t. GC, as this scenario demonstrates:
878 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
879 // VM thread changes sync state to synchronizing and suspends threads for GC.
880 // Thread A is resumed to finish this native method, but doesn't block here since it
881 // didn't see any synchronization is progress, and escapes.
882 __ set(_thread_in_native_trans, G3_scratch);
883 __ st(G3_scratch, thread_state);
884 if(os::is_MP()) {
885 // Write serialization page so VM thread can do a pseudo remote membar.
886 // We use the current thread pointer to calculate a thread specific
887 // offset to write to within the page. This minimizes bus traffic
888 // due to cache line collision.
889 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
890 }
891 __ load_contents(sync_state, G3_scratch);
892 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
893
894
895 Label L;
896 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
897 __ br(Assembler::notEqual, false, Assembler::pn, L);
898 __ delayed()->
899 ld(suspend_state, G3_scratch);
900 __ cmp(G3_scratch, 0);
901 __ br(Assembler::equal, false, Assembler::pt, no_block);
902 __ delayed()->nop();
903 __ bind(L);
904
905 // Block. Save any potential method result value before the operation and
906 // use a leaf call to leave the last_Java_frame setup undisturbed.
907 save_native_result();
908 __ call_VM_leaf(noreg,
909 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
910 G2_thread);
911 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
912 // Restore any method result value
913 restore_native_result();
914 __ bind(no_block);
915 }
916
917 // Clear the frame anchor now
918
919 __ reset_last_Java_frame();
920
921 // Move the result handler address
922 __ mov(Lscratch, G3_scratch);
923 // return possible result to the outer frame
924 #ifndef __LP64
925 __ mov(O0, I0);
926 __ restore(O1, G0, O1);
927 #else
928 __ restore(O0, G0, O0);
929 #endif /* __LP64 */
930
931 // Move result handler to expected register
932 __ mov(G3_scratch, Lscratch);
933
934
935 // thread state is thread_in_native_trans. Any safepoint blocking has
936 // happened in the trampoline we are ready to switch to thread_in_Java.
937
938 __ set(_thread_in_Java, G3_scratch);
939 __ st(G3_scratch, thread_state);
940
941 // If we have an oop result store it where it will be safe for any further gc
942 // until we return now that we've released the handle it might be protected by
943
944 {
945 Label no_oop, store_result;
946
947 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
948 __ cmp(G3_scratch, Lscratch);
949 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
950 __ delayed()->nop();
951 __ addcc(G0, O0, O0);
952 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
953 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
954 __ mov(G0, O0);
955
956 __ bind(store_result);
957 // Store it where gc will look for it and result handler expects it.
958 __ st_ptr(O0, STATE(_oop_temp));
959
960 __ bind(no_oop);
961
962 }
963
964 // reset handle block
965 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
966 __ st(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
967
968
969 // handle exceptions (exception handling will handle unlocking!)
970 { Label L;
971 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
972
973 __ ld_ptr(exception_addr, Gtemp);
974 __ tst(Gtemp);
975 __ brx(Assembler::equal, false, Assembler::pt, L);
976 __ delayed()->nop();
977 __ bind(pending_exception_present);
978 // With c++ interpreter we just leave it pending caller will do the correct thing. However...
979 // Like x86 we ignore the result of the native call and leave the method locked. This
980 // seems wrong to leave things locked.
981
982 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
983 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
984
985 __ bind(L);
986 }
987
988 // jvmdi/jvmpi support (preserves thread register)
989 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
990
991 if (synchronized) {
992 // save and restore any potential method result value around the unlocking operation
993 save_native_result();
994
995 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
996 // Get the initial monitor we allocated
997 __ sub(Lstate, entry_size, O1); // initial monitor
998 __ unlock_object(O1);
999 restore_native_result();
1000 }
1001
1002 #if defined(COMPILER2) && !defined(_LP64)
1003
1004 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1005 // or compiled so just be safe.
1006
1007 __ sllx(O0, 32, G1); // Shift bits into high G1
1008 __ srl (O1, 0, O1); // Zero extend O1
1009 __ or3 (O1, G1, G1); // OR 64 bits into G1
1010
1011 #endif /* COMPILER2 && !_LP64 */
1012
1013 #ifdef ASSERT
1014 {
1015 Label ok;
1016 __ cmp(I5_savedSP, FP);
1017 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1018 __ delayed()->nop();
1019 __ stop("bad I5_savedSP value");
1020 __ should_not_reach_here();
1021 __ bind(ok);
1022 }
1023 #endif
1024 // Calls result handler which POPS FRAME
1025 if (TraceJumps) {
1026 // Move target to register that is recordable
1027 __ mov(Lscratch, G3_scratch);
1028 __ JMP(G3_scratch, 0);
1029 } else {
1030 __ jmp(Lscratch, 0);
1031 }
1032 __ delayed()->nop();
1033
1034 if (inc_counter) {
1035 // handle invocation counter overflow
1036 __ bind(invocation_counter_overflow);
1037 generate_counter_overflow(Lcontinue);
1038 }
1039
1040
1041 return entry;
1042 }
1043
generate_compute_interpreter_state(const Register state,const Register prev_state,bool native)1044 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1045 const Register prev_state,
1046 bool native) {
1047
1048 // On entry
1049 // G5_method - caller's method
1050 // Gargs - points to initial parameters (i.e. locals[0])
1051 // G2_thread - valid? (C1 only??)
1052 // "prev_state" - contains any previous frame manager state which we must save a link
1053 //
1054 // On return
1055 // "state" is a pointer to the newly allocated state object. We must allocate and initialize
1056 // a new interpretState object and the method expression stack.
1057
1058 assert_different_registers(state, prev_state);
1059 assert_different_registers(prev_state, G3_scratch);
1060 const Register Gtmp = G3_scratch;
1061 const Address constMethod (G5_method, 0, in_bytes(Method::const_offset()));
1062 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
1063
1064 // slop factor is two extra slots on the expression stack so that
1065 // we always have room to store a result when returning from a call without parameters
1066 // that returns a result.
1067
1068 const int slop_factor = 2*wordSize;
1069
1070 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1071 Method::extra_stack_entries() + // extra stack for jsr 292
1072 frame::memory_parameter_word_sp_offset + // register save area + param window
1073 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1074
1075 // XXX G5_method valid
1076
1077 // Now compute new frame size
1078
1079 if (native) {
1080 const Register RconstMethod = Gtmp;
1081 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1082 __ ld_ptr(constMethod, RconstMethod);
1083 __ lduh( size_of_parameters, Gtmp );
1084 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words
1085 } else {
1086 // Full size expression stack
1087 __ ld_ptr(constMethod, Gtmp);
1088 __ lduh(Gtmp, in_bytes(ConstMethod::max_stack_offset()), Gtmp);
1089 }
1090 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1091
1092 __ neg(Gtmp); // negative space for stack/parameters in words
1093 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1094 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1095
1096 // Need to do stack size check here before we fault on large frames
1097
1098 Label stack_ok;
1099
1100 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1101 (StackRedPages+StackYellowPages);
1102
1103
1104 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1105 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1106 // compute stack bottom
1107 __ sub(O0, O1, O0);
1108
1109 // Avoid touching the guard pages
1110 // Also a fudge for frame size of BytecodeInterpreter::run
1111 // It varies from 1k->4k depending on build type
1112 const int fudge = 6 * K;
1113
1114 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1115
1116 __ add(O0, O1, O0);
1117 __ sub(O0, Gtmp, O0);
1118 __ cmp(SP, O0);
1119 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1120 __ delayed()->nop();
1121
1122 // throw exception return address becomes throwing pc
1123
1124 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1125 __ stop("never reached");
1126
1127 __ bind(stack_ok);
1128
1129 __ save(SP, Gtmp, SP); // setup new frame and register window
1130
1131 // New window I7 call_stub or previous activation
1132 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1133 //
1134 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1135 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1136
1137 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1138
1139 // Initialize a new Interpreter state
1140 // orig_sp - caller's original sp
1141 // G2_thread - thread
1142 // Gargs - &locals[0] (unbiased?)
1143 // G5_method - method
1144 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1145
1146
1147 __ set(0xdead0004, O1);
1148
1149
1150 __ st_ptr(Gargs, XXX_STATE(_locals));
1151 __ st_ptr(G0, XXX_STATE(_oop_temp));
1152
1153 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1154 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1155 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1156
1157 if (native) {
1158 __ st_ptr(G0, XXX_STATE(_bcp));
1159 } else {
1160 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), O2); // get ConstMethod*
1161 __ add(O2, in_bytes(ConstMethod::codes_offset()), O2); // get bcp
1162 __ st_ptr(O2, XXX_STATE(_bcp));
1163 }
1164
1165 __ st_ptr(G0, XXX_STATE(_mdx));
1166 __ st_ptr(G5_method, XXX_STATE(_method));
1167
1168 __ set((int) BytecodeInterpreter::method_entry, O1);
1169 __ st(O1, XXX_STATE(_msg));
1170
1171 __ ld_ptr(constMethod, O3);
1172 __ ld_ptr(O3, in_bytes(ConstMethod::constants_offset()), O3);
1173 __ ld_ptr(O3, ConstantPool::cache_offset_in_bytes(), O2);
1174 __ st_ptr(O2, XXX_STATE(_constants));
1175
1176 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1177
1178 // Monitor base is just start of BytecodeInterpreter object;
1179 __ mov(state, O2);
1180 __ st_ptr(O2, XXX_STATE(_monitor_base));
1181
1182 // Do we need a monitor for synchonized method?
1183 {
1184 __ ld(access_flags, O1);
1185 Label done;
1186 Label got_obj;
1187 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1188 __ br( Assembler::zero, false, Assembler::pt, done);
1189
1190 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1191 __ delayed()->btst(JVM_ACC_STATIC, O1);
1192 __ ld_ptr(XXX_STATE(_locals), O1);
1193 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1194 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1195 __ ld_ptr(constMethod, O1);
1196 __ ld_ptr( O1, in_bytes(ConstMethod::constants_offset()), O1);
1197 __ ld_ptr( O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1198 // lock the mirror, not the Klass*
1199 __ ld_ptr( O1, mirror_offset, O1);
1200
1201 __ bind(got_obj);
1202
1203 #ifdef ASSERT
1204 __ tst(O1);
1205 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
1206 #endif // ASSERT
1207
1208 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1209 __ sub(SP, entry_size, SP); // account for initial monitor
1210 __ sub(O2, entry_size, O2); // initial monitor
1211 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1212 __ bind(done);
1213 }
1214
1215 // Remember initial frame bottom
1216
1217 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1218
1219 __ st_ptr(O2, XXX_STATE(_stack_base));
1220
1221 __ sub(O2, wordSize, O2); // prepush
1222 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1223
1224 // Full size expression stack
1225 __ ld_ptr(constMethod, O3);
1226 __ lduh(O3, in_bytes(ConstMethod::max_stack_offset()), O3);
1227 __ inc(O3, Method::extra_stack_entries());
1228 __ sll(O3, LogBytesPerWord, O3);
1229 __ sub(O2, O3, O3);
1230 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1231 __ st_ptr(O3, XXX_STATE(_stack_limit));
1232
1233 if (!native) {
1234 //
1235 // Code to initialize locals
1236 //
1237 Register init_value = noreg; // will be G0 if we must clear locals
1238 // Now zero locals
1239 if (true /* zerolocals */ || ClearInterpreterLocals) {
1240 // explicitly initialize locals
1241 init_value = G0;
1242 } else {
1243 #ifdef ASSERT
1244 // initialize locals to a garbage pattern for better debugging
1245 init_value = O3;
1246 __ set( 0x0F0F0F0F, init_value );
1247 #endif // ASSERT
1248 }
1249 if (init_value != noreg) {
1250 Label clear_loop;
1251 const Register RconstMethod = O1;
1252 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1253 const Address size_of_locals (RconstMethod, 0, in_bytes(ConstMethod::size_of_locals_offset()));
1254
1255 // NOTE: If you change the frame layout, this code will need to
1256 // be updated!
1257 __ ld_ptr( constMethod, RconstMethod );
1258 __ lduh( size_of_locals, O2 );
1259 __ lduh( size_of_parameters, O1 );
1260 __ sll( O2, LogBytesPerWord, O2);
1261 __ sll( O1, LogBytesPerWord, O1 );
1262 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1263 __ sub( L2_scratch, O2, O2 );
1264 __ sub( L2_scratch, O1, O1 );
1265
1266 __ bind( clear_loop );
1267 __ inc( O2, wordSize );
1268
1269 __ cmp( O2, O1 );
1270 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1271 __ delayed()->st_ptr( init_value, O2, 0 );
1272 }
1273 }
1274 }
1275 // Find preallocated monitor and lock method (C++ interpreter)
1276 //
lock_method(void)1277 void InterpreterGenerator::lock_method(void) {
1278 // Lock the current method.
1279 // Destroys registers L2_scratch, L3_scratch, O0
1280 //
1281 // Find everything relative to Lstate
1282
1283 #ifdef ASSERT
1284 __ ld_ptr(STATE(_method), L2_scratch);
1285 __ ld(L2_scratch, in_bytes(Method::access_flags_offset()), O0);
1286
1287 { Label ok;
1288 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1289 __ br( Assembler::notZero, false, Assembler::pt, ok);
1290 __ delayed()->nop();
1291 __ stop("method doesn't need synchronization");
1292 __ bind(ok);
1293 }
1294 #endif // ASSERT
1295
1296 // monitor is already allocated at stack base
1297 // and the lockee is already present
1298 __ ld_ptr(STATE(_stack_base), L2_scratch);
1299 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1300 __ lock_object(L2_scratch, O0);
1301
1302 }
1303
1304 // Generate code for handling resuming a deopted method
generate_deopt_handling()1305 void CppInterpreterGenerator::generate_deopt_handling() {
1306
1307 Label return_from_deopt_common;
1308
1309 // deopt needs to jump to here to enter the interpreter (return a result)
1310 deopt_frame_manager_return_atos = __ pc();
1311
1312 // O0/O1 live
1313 __ ba(return_from_deopt_common);
1314 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1315
1316
1317 // deopt needs to jump to here to enter the interpreter (return a result)
1318 deopt_frame_manager_return_btos = __ pc();
1319
1320 // O0/O1 live
1321 __ ba(return_from_deopt_common);
1322 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1323
1324 // deopt needs to jump to here to enter the interpreter (return a result)
1325 deopt_frame_manager_return_itos = __ pc();
1326
1327 // O0/O1 live
1328 __ ba(return_from_deopt_common);
1329 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1330
1331 // deopt needs to jump to here to enter the interpreter (return a result)
1332
1333 deopt_frame_manager_return_ltos = __ pc();
1334 #if !defined(_LP64) && defined(COMPILER2)
1335 // All return values are where we want them, except for Longs. C2 returns
1336 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1337 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1338 // build even if we are returning from interpreted we just do a little
1339 // stupid shuffing.
1340 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1341 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1342 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1343
1344 __ srl (G1, 0,O1);
1345 __ srlx(G1,32,O0);
1346 #endif /* !_LP64 && COMPILER2 */
1347 // O0/O1 live
1348 __ ba(return_from_deopt_common);
1349 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1350
1351 // deopt needs to jump to here to enter the interpreter (return a result)
1352
1353 deopt_frame_manager_return_ftos = __ pc();
1354 // O0/O1 live
1355 __ ba(return_from_deopt_common);
1356 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1357
1358 // deopt needs to jump to here to enter the interpreter (return a result)
1359 deopt_frame_manager_return_dtos = __ pc();
1360
1361 // O0/O1 live
1362 __ ba(return_from_deopt_common);
1363 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1364
1365 // deopt needs to jump to here to enter the interpreter (return a result)
1366 deopt_frame_manager_return_vtos = __ pc();
1367
1368 // O0/O1 live
1369 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1370
1371 // Deopt return common
1372 // an index is present that lets us move any possible result being
1373 // return to the interpreter's stack
1374 //
1375 __ bind(return_from_deopt_common);
1376
1377 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1378 // stack is in the state that the calling convention left it.
1379 // Copy the result from native abi result and place it on java expression stack.
1380
1381 // Current interpreter state is present in Lstate
1382
1383 // Get current pre-pushed top of interpreter stack
1384 // Any result (if any) is in native abi
1385 // result type index is in L3_scratch
1386
1387 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1388
1389 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1390 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1391 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1392 __ jmpl(Lscratch, G0, O7); // and convert it
1393 __ delayed()->nop();
1394
1395 // L1_scratch points to top of stack (prepushed)
1396 __ st_ptr(L1_scratch, STATE(_stack));
1397 }
1398
1399 // Generate the code to handle a more_monitors message from the c++ interpreter
generate_more_monitors()1400 void CppInterpreterGenerator::generate_more_monitors() {
1401
1402 Label entry, loop;
1403 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1404 // 1. compute new pointers // esp: old expression stack top
1405 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1406 __ sub(L4_scratch, entry_size, L4_scratch);
1407 __ st_ptr(L4_scratch, STATE(_stack_base));
1408
1409 __ sub(SP, entry_size, SP); // Grow stack
1410 __ st_ptr(SP, STATE(_frame_bottom));
1411
1412 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1413 __ sub(L2_scratch, entry_size, L2_scratch);
1414 __ st_ptr(L2_scratch, STATE(_stack_limit));
1415
1416 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1417 __ sub(L1_scratch, entry_size, L1_scratch);
1418 __ st_ptr(L1_scratch, STATE(_stack));
1419 __ ba(entry);
1420 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1421
1422 // 2. move expression stack
1423
1424 __ bind(loop);
1425 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1426 __ add(L1_scratch, wordSize, L1_scratch);
1427 __ bind(entry);
1428 __ cmp(L1_scratch, L4_scratch);
1429 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1430 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1431
1432 // now zero the slot so we can find it.
1433 __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1434
1435 }
1436
1437 // Initial entry to C++ interpreter from the call_stub.
1438 // This entry point is called the frame manager since it handles the generation
1439 // of interpreter activation frames via requests directly from the vm (via call_stub)
1440 // and via requests from the interpreter. The requests from the call_stub happen
1441 // directly thru the entry point. Requests from the interpreter happen via returning
1442 // from the interpreter and examining the message the interpreter has returned to
1443 // the frame manager. The frame manager can take the following requests:
1444
1445 // NO_REQUEST - error, should never happen.
1446 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1447 // allocate a new monitor.
1448 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1449 // happens during entry during the entry via the call stub.
1450 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1451 //
1452 // Arguments:
1453 //
1454 // ebx: Method*
1455 // ecx: receiver - unused (retrieved from stack as needed)
1456 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1457 //
1458 //
1459 // Stack layout at entry
1460 //
1461 // [ return address ] <--- esp
1462 // [ parameter n ]
1463 // ...
1464 // [ parameter 1 ]
1465 // [ expression stack ]
1466 //
1467 //
1468 // We are free to blow any registers we like because the call_stub which brought us here
1469 // initially has preserved the callee save registers already.
1470 //
1471 //
1472
1473 static address interpreter_frame_manager = NULL;
1474
1475 #ifdef ASSERT
1476 #define VALIDATE_STATE(scratch, marker) \
1477 { \
1478 Label skip; \
1479 __ ld_ptr(STATE(_self_link), scratch); \
1480 __ cmp(Lstate, scratch); \
1481 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1482 __ delayed()->nop(); \
1483 __ breakpoint_trap(); \
1484 __ emit_int32(marker); \
1485 __ bind(skip); \
1486 }
1487 #else
1488 #define VALIDATE_STATE(scratch, marker)
1489 #endif /* ASSERT */
1490
adjust_callers_stack(Register args)1491 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1492 //
1493 // Adjust caller's stack so that all the locals can be contiguous with
1494 // the parameters.
1495 // Worries about stack overflow make this a pain.
1496 //
1497 // Destroys args, G3_scratch, G3_scratch
1498 // In/Out O5_savedSP (sender's original SP)
1499 //
1500 // assert_different_registers(state, prev_state);
1501 const Register Gtmp = G3_scratch;
1502 const RconstMethod = G3_scratch;
1503 const Register tmp = O2;
1504 const Address constMethod(G5_method, 0, in_bytes(Method::const_offset()));
1505 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1506 const Address size_of_locals (RconstMethod, 0, in_bytes(ConstMethod::size_of_locals_offset()));
1507
1508 __ ld_ptr(constMethod, RconstMethod);
1509 __ lduh(size_of_parameters, tmp);
1510 __ sll(tmp, LogBytesPerWord, Gargs); // parameter size in bytes
1511 __ add(args, Gargs, Gargs); // points to first local + BytesPerWord
1512 // NEW
1513 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1514 // determine extra space for non-argument locals & adjust caller's SP
1515 // Gtmp1: parameter size in words
1516 __ lduh(size_of_locals, Gtmp);
1517 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1518
1519 #if 1
1520 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1521 // the call_stub will place the final interpreter argument at
1522 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1523 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1524 // and try to make it look good in the debugger we will store the argument to
1525 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1526 // extra space for the compiler this will overwrite locals in the local array of the
1527 // interpreter.
1528 // QQQ still needed with frameless adapters???
1529
1530 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1531
1532 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1533 #endif // 1
1534
1535
1536 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1537 }
1538
generate_normal_entry(bool synchronized)1539 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1540
1541 // G5_method: Method*
1542 // G2_thread: thread (unused)
1543 // Gargs: bottom of args (sender_sp)
1544 // O5: sender's sp
1545
1546 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1547 // the code is pretty much ready. Would need to change the test below and for good measure
1548 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1549 // routines. Not clear this is worth it yet.
1550
1551 if (interpreter_frame_manager) {
1552 return interpreter_frame_manager;
1553 }
1554
1555 __ bind(frame_manager_entry);
1556
1557 // the following temporary registers are used during frame creation
1558 const Register Gtmp1 = G3_scratch;
1559 // const Register Lmirror = L1; // native mirror (native calls only)
1560
1561 const Address constMethod (G5_method, 0, in_bytes(Method::const_offset()));
1562 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
1563
1564 address entry_point = __ pc();
1565 __ mov(G0, prevState); // no current activation
1566
1567
1568 Label re_dispatch;
1569
1570 __ bind(re_dispatch);
1571
1572 // Interpreter needs to have locals completely contiguous. In order to do that
1573 // We must adjust the caller's stack pointer for any locals beyond just the
1574 // parameters
1575 adjust_callers_stack(Gargs);
1576
1577 // O5_savedSP still contains sender's sp
1578
1579 // NEW FRAME
1580
1581 generate_compute_interpreter_state(Lstate, prevState, false);
1582
1583 // At this point a new interpreter frame and state object are created and initialized
1584 // Lstate has the pointer to the new activation
1585 // Any stack banging or limit check should already be done.
1586
1587 Label call_interpreter;
1588
1589 __ bind(call_interpreter);
1590
1591
1592 #if 1
1593 __ set(0xdead002, Lmirror);
1594 __ set(0xdead002, L2_scratch);
1595 __ set(0xdead003, L3_scratch);
1596 __ set(0xdead004, L4_scratch);
1597 __ set(0xdead005, Lscratch);
1598 __ set(0xdead006, Lscratch2);
1599 __ set(0xdead007, L7_scratch);
1600
1601 __ set(0xdeaf002, O2);
1602 __ set(0xdeaf003, O3);
1603 __ set(0xdeaf004, O4);
1604 __ set(0xdeaf005, O5);
1605 #endif
1606
1607 // Call interpreter (stack bang complete) enter here if message is
1608 // set and we know stack size is valid
1609
1610 Label call_interpreter_2;
1611
1612 __ bind(call_interpreter_2);
1613
1614 #ifdef ASSERT
1615 {
1616 Label skip;
1617 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1618 __ cmp(G3_scratch, SP);
1619 __ brx(Assembler::equal, false, Assembler::pt, skip);
1620 __ delayed()->nop();
1621 __ stop("SP not restored to frame bottom");
1622 __ bind(skip);
1623 }
1624 #endif
1625
1626 VALIDATE_STATE(G3_scratch, 4);
1627 __ set_last_Java_frame(SP, noreg);
1628 __ mov(Lstate, O0); // (arg) pointer to current state
1629
1630 __ call(CAST_FROM_FN_PTR(address,
1631 JvmtiExport::can_post_interpreter_events() ?
1632 BytecodeInterpreter::runWithChecks
1633 : BytecodeInterpreter::run),
1634 relocInfo::runtime_call_type);
1635
1636 __ delayed()->nop();
1637
1638 __ ld_ptr(STATE(_thread), G2_thread);
1639 __ reset_last_Java_frame();
1640
1641 // examine msg from interpreter to determine next action
1642 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1643
1644 __ ld(STATE(_msg), L1_scratch); // Get new message
1645
1646 Label call_method;
1647 Label return_from_interpreted_method;
1648 Label throw_exception;
1649 Label do_OSR;
1650 Label bad_msg;
1651 Label resume_interpreter;
1652
1653 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1654 __ br(Assembler::equal, false, Assembler::pt, call_method);
1655 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1656 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1657 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1658 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1659 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1660 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1661 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1662 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1663
1664 // Allocate more monitor space, shuffle expression stack....
1665
1666 generate_more_monitors();
1667
1668 // new monitor slot allocated, resume the interpreter.
1669
1670 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1671 VALIDATE_STATE(G3_scratch, 5);
1672 __ ba(call_interpreter);
1673 __ delayed()->st(L1_scratch, STATE(_msg));
1674
1675 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1676 unctrap_frame_manager_entry = __ pc();
1677
1678 // QQQ what message do we send
1679
1680 __ ba(call_interpreter);
1681 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1682
1683 //=============================================================================
1684 // Returning from a compiled method into a deopted method. The bytecode at the
1685 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1686 // for the template based interpreter). Any stack space that was used by the
1687 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1688 // so all that we have to do is place any pending result on the expression stack
1689 // and resume execution on the next bytecode.
1690
1691 generate_deopt_handling();
1692
1693 // ready to resume the interpreter
1694
1695 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1696 __ ba(call_interpreter);
1697 __ delayed()->st(L1_scratch, STATE(_msg));
1698
1699 // Current frame has caught an exception we need to dispatch to the
1700 // handler. We can get here because a native interpreter frame caught
1701 // an exception in which case there is no handler and we must rethrow
1702 // If it is a vanilla interpreted frame the we simply drop into the
1703 // interpreter and let it do the lookup.
1704
1705 Interpreter::_rethrow_exception_entry = __ pc();
1706
1707 Label return_with_exception;
1708 Label unwind_and_forward;
1709
1710 // O0: exception
1711 // O7: throwing pc
1712
1713 // We want exception in the thread no matter what we ultimately decide about frame type.
1714
1715 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1716 __ verify_thread();
1717 __ st_ptr(O0, exception_addr);
1718
1719 // get the Method*
1720 __ ld_ptr(STATE(_method), G5_method);
1721
1722 // if this current frame vanilla or native?
1723
1724 __ ld(access_flags, Gtmp1);
1725 __ btst(JVM_ACC_NATIVE, Gtmp1);
1726 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1727 __ delayed()->nop();
1728
1729 // We drop thru to unwind a native interpreted frame with a pending exception
1730 // We jump here for the initial interpreter frame with exception pending
1731 // We unwind the current acivation and forward it to our caller.
1732
1733 __ bind(unwind_and_forward);
1734
1735 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1736 // as expected by forward_exception.
1737
1738 __ restore(FP, G0, SP); // unwind interpreter state frame
1739 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1740 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1741
1742 // Return point from a call which returns a result in the native abi
1743 // (c1/c2/jni-native). This result must be processed onto the java
1744 // expression stack.
1745 //
1746 // A pending exception may be present in which case there is no result present
1747
1748 address return_from_native_method = __ pc();
1749
1750 VALIDATE_STATE(G3_scratch, 6);
1751
1752 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1753 // stack is in the state that the calling convention left it.
1754 // Copy the result from native abi result and place it on java expression stack.
1755
1756 // Current interpreter state is present in Lstate
1757
1758 // Exception pending?
1759
1760 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1761 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1762 __ tst(Lscratch); // exception pending?
1763 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1764 __ delayed()->nop();
1765
1766 // Process the native abi result to java expression stack
1767
1768 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1769 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1770 // get parameter size
1771 __ ld_ptr(L4_scratch, in_bytes(Method::const_offset()), L2_scratch);
1772 __ lduh(L2_scratch, in_bytes(ConstMethod::size_of_parameters_offset()), L2_scratch);
1773 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1774 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1775 __ ld(L4_scratch, in_bytes(Method::result_index_offset()), L3_scratch); // called method result type index
1776
1777 // tosca is really just native abi
1778 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1779 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1780 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1781 __ jmpl(Lscratch, G0, O7); // and convert it
1782 __ delayed()->nop();
1783
1784 // L1_scratch points to top of stack (prepushed)
1785
1786 __ ba(resume_interpreter);
1787 __ delayed()->mov(L1_scratch, O1);
1788
1789 // An exception is being caught on return to a vanilla interpreter frame.
1790 // Empty the stack and resume interpreter
1791
1792 __ bind(return_with_exception);
1793
1794 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1795 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1796 __ ba(resume_interpreter);
1797 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1798
1799 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1800 // interpreter call, or native) and unwind this interpreter activation.
1801 // All monitors should be unlocked.
1802
1803 __ bind(return_from_interpreted_method);
1804
1805 VALIDATE_STATE(G3_scratch, 7);
1806
1807 Label return_to_initial_caller;
1808
1809 // Interpreted result is on the top of the completed activation expression stack.
1810 // We must return it to the top of the callers stack if caller was interpreted
1811 // otherwise we convert to native abi result and return to call_stub/c1/c2
1812 // The caller's expression stack was truncated by the call however the current activation
1813 // has enough stuff on the stack that we have usable space there no matter what. The
1814 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1815 // for the current activation
1816
1817 __ ld_ptr(STATE(_prev_link), L1_scratch);
1818 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1819 __ ld(L2_scratch, in_bytes(Method::result_index_offset()), L2_scratch);
1820 __ tst(L1_scratch);
1821 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1822 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1823
1824 // Copy result to callers java stack
1825
1826 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1827 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1828 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1829 __ ld_ptr(STATE(_locals), O1); // stack destination
1830
1831 // O0 - will be source, O1 - will be destination (preserved)
1832 __ jmpl(Lscratch, G0, O7); // and convert it
1833 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1834
1835 // O1 == &locals[0]
1836
1837 // Result is now on caller's stack. Just unwind current activation and resume
1838
1839 Label unwind_recursive_activation;
1840
1841
1842 __ bind(unwind_recursive_activation);
1843
1844 // O1 == &locals[0] (really callers stacktop) for activation now returning
1845 // returning to interpreter method from "recursive" interpreter call
1846 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1847 // to. Now all we must do is unwind the state from the completed call
1848
1849 // Must restore stack
1850 VALIDATE_STATE(G3_scratch, 8);
1851
1852 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1853 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1854 // frame and tell interpreter to resume.
1855
1856
1857 __ mov(O1, I1); // pass back new stack top across activation
1858 // POP FRAME HERE ==================================
1859 __ restore(FP, G0, SP); // unwind interpreter state frame
1860 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1861
1862
1863 // Resume the interpreter. The current frame contains the current interpreter
1864 // state object.
1865 //
1866 // O1 == new java stack pointer
1867
1868 __ bind(resume_interpreter);
1869 VALIDATE_STATE(G3_scratch, 10);
1870
1871 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1872
1873 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1874 __ st(L1_scratch, STATE(_msg));
1875 __ ba(call_interpreter_2);
1876 __ delayed()->st_ptr(O1, STATE(_stack));
1877
1878
1879 // Fast accessor methods share this entry point.
1880 // This works because frame manager is in the same codelet
1881 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1882 // we need to do a little register fixup here once we distinguish the two of them
1883 if (UseFastAccessorMethods && !synchronized) {
1884 // Call stub_return address still in O7
1885 __ bind(fast_accessor_slow_entry_path);
1886 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1887 __ cmp(Gtmp1, O7); // returning to interpreter?
1888 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1889 __ delayed()->nop();
1890 __ ba(re_dispatch);
1891 __ delayed()->mov(G0, prevState); // initial entry
1892
1893 }
1894
1895 // interpreter returning to native code (call_stub/c1/c2)
1896 // convert result and unwind initial activation
1897 // L2_scratch - scaled result type index
1898
1899 __ bind(return_to_initial_caller);
1900
1901 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1902 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1903 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1904 __ jmpl(Lscratch, G0, O7); // and convert it
1905 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1906
1907 Label unwind_initial_activation;
1908 __ bind(unwind_initial_activation);
1909
1910 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1911 // we can return here with an exception that wasn't handled by interpreted code
1912 // how does c1/c2 see it on return?
1913
1914 // compute resulting sp before/after args popped depending upon calling convention
1915 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1916 //
1917 // POP FRAME HERE ==================================
1918 __ restore(FP, G0, SP);
1919 __ retl();
1920 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1921
1922 // OSR request, unwind the current frame and transfer to the OSR entry
1923 // and enter OSR nmethod
1924
1925 __ bind(do_OSR);
1926 Label remove_initial_frame;
1927 __ ld_ptr(STATE(_prev_link), L1_scratch);
1928 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1929
1930 // We are going to pop this frame. Is there another interpreter frame underneath
1931 // it or is it callstub/compiled?
1932
1933 __ tst(L1_scratch);
1934 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1935 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1936
1937 // Frame underneath is an interpreter frame simply unwind
1938 // POP FRAME HERE ==================================
1939 __ restore(FP, G0, SP); // unwind interpreter state frame
1940 __ mov(I5_savedSP->after_restore(), SP);
1941
1942 // Since we are now calling native need to change our "return address" from the
1943 // dummy RecursiveInterpreterActivation to a return from native
1944
1945 __ set((intptr_t)return_from_native_method - 8, O7);
1946
1947 __ jmpl(G3_scratch, G0, G0);
1948 __ delayed()->mov(G1_scratch, O0);
1949
1950 __ bind(remove_initial_frame);
1951
1952 // POP FRAME HERE ==================================
1953 __ restore(FP, G0, SP);
1954 __ mov(I5_savedSP->after_restore(), SP);
1955 __ jmpl(G3_scratch, G0, G0);
1956 __ delayed()->mov(G1_scratch, O0);
1957
1958 // Call a new method. All we do is (temporarily) trim the expression stack
1959 // push a return address to bring us back to here and leap to the new entry.
1960 // At this point we have a topmost frame that was allocated by the frame manager
1961 // which contains the current method interpreted state. We trim this frame
1962 // of excess java expression stack entries and then recurse.
1963
1964 __ bind(call_method);
1965
1966 // stack points to next free location and not top element on expression stack
1967 // method expects sp to be pointing to topmost element
1968
1969 __ ld_ptr(STATE(_thread), G2_thread);
1970 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1971
1972
1973 // SP already takes in to account the 2 extra words we use for slop
1974 // when we call a "static long no_params()" method. So if
1975 // we trim back sp by the amount of unused java expression stack
1976 // there will be automagically the 2 extra words we need.
1977 // We also have to worry about keeping SP aligned.
1978
1979 __ ld_ptr(STATE(_stack), Gargs);
1980 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1981
1982 // compute the unused java stack size
1983 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1984
1985 // Round down the unused space to that stack is always 16-byte aligned
1986 // by making the unused space a multiple of the size of two longs.
1987
1988 __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1989
1990 // Now trim the stack
1991 __ add(SP, L2_scratch, SP);
1992
1993
1994 // Now point to the final argument (account for prepush)
1995 __ add(Gargs, wordSize, Gargs);
1996 #ifdef ASSERT
1997 // Make sure we have space for the window
1998 __ sub(Gargs, SP, L1_scratch);
1999 __ cmp(L1_scratch, 16*wordSize);
2000 {
2001 Label skip;
2002 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
2003 __ delayed()->nop();
2004 __ stop("killed stack");
2005 __ bind(skip);
2006 }
2007 #endif // ASSERT
2008
2009 // Create a new frame where we can store values that make it look like the interpreter
2010 // really recursed.
2011
2012 // prepare to recurse or call specialized entry
2013
2014 // First link the registers we need
2015
2016 // make the pc look good in debugger
2017 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
2018 // argument too
2019 __ mov(Lstate, I0);
2020
2021 // Record our sending SP
2022 __ mov(SP, O5_savedSP);
2023
2024 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
2025 __ set((intptr_t) entry_point, L1_scratch);
2026 __ cmp(L1_scratch, L2_scratch);
2027 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
2028 __ delayed()->mov(Lstate, prevState); // link activations
2029
2030 // method uses specialized entry, push a return so we look like call stub setup
2031 // this path will handle fact that result is returned in registers and not
2032 // on the java stack.
2033
2034 __ set((intptr_t)return_from_native_method - 8, O7);
2035 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
2036 __ delayed()->nop();
2037
2038 //
2039 // Bad Message from interpreter
2040 //
2041 __ bind(bad_msg);
2042 __ stop("Bad message from interpreter");
2043
2044 // Interpreted method "returned" with an exception pass it on...
2045 // Pass result, unwind activation and continue/return to interpreter/call_stub
2046 // We handle result (if any) differently based on return to interpreter or call_stub
2047
2048 __ bind(throw_exception);
2049 __ ld_ptr(STATE(_prev_link), L1_scratch);
2050 __ tst(L1_scratch);
2051 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2052 __ delayed()->nop();
2053
2054 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
2055 __ ba(unwind_recursive_activation);
2056 __ delayed()->nop();
2057
2058 interpreter_frame_manager = entry_point;
2059 return entry_point;
2060 }
2061
InterpreterGenerator(StubQueue * code)2062 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2063 : CppInterpreterGenerator(code) {
2064 generate_all(); // down here so it can be "virtual"
2065 }
2066
2067
size_activation_helper(int callee_extra_locals,int max_stack,int monitor_size)2068 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2069
2070 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2071 // expression stack, the callee will have callee_extra_locals (so we can account for
2072 // frame extension) and monitor_size for monitors. Basically we need to calculate
2073 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2074 //
2075 //
2076 // The big complicating thing here is that we must ensure that the stack stays properly
2077 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2078 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2079 // the caller so we must ensure that it is properly aligned for our callee.
2080 //
2081 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2082 // stack at all times to deal with the "stack long no_params()" method issue. This
2083 // is "slop_factor" here.
2084 const int slop_factor = 2;
2085
2086 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2087 frame::memory_parameter_word_sp_offset; // register save area + param window
2088 return (round_to(max_stack +
2089 slop_factor +
2090 fixed_size +
2091 monitor_size +
2092 (callee_extra_locals * Interpreter::stackElementWords), WordsPerLong));
2093
2094 }
2095
size_top_interpreter_activation(Method * method)2096 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
2097
2098 // See call_stub code
2099 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2100 WordsPerLong); // 7 + register save area
2101
2102 // Save space for one monitor to get into the interpreted method in case
2103 // the method is synchronized
2104 int monitor_size = method->is_synchronized() ?
2105 1*frame::interpreter_frame_monitor_size() : 0;
2106 return size_activation_helper(method->max_locals(), method->max_stack(),
2107 monitor_size) + call_stub_size;
2108 }
2109
layout_interpreterState(interpreterState to_fill,frame * caller,frame * current,Method * method,intptr_t * locals,intptr_t * stack,intptr_t * stack_base,intptr_t * monitor_base,intptr_t * frame_bottom,bool is_top_frame)2110 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2111 frame* caller,
2112 frame* current,
2113 Method* method,
2114 intptr_t* locals,
2115 intptr_t* stack,
2116 intptr_t* stack_base,
2117 intptr_t* monitor_base,
2118 intptr_t* frame_bottom,
2119 bool is_top_frame
2120 )
2121 {
2122 // What about any vtable?
2123 //
2124 to_fill->_thread = JavaThread::current();
2125 // This gets filled in later but make it something recognizable for now
2126 to_fill->_bcp = method->code_base();
2127 to_fill->_locals = locals;
2128 to_fill->_constants = method->constants()->cache();
2129 to_fill->_method = method;
2130 to_fill->_mdx = NULL;
2131 to_fill->_stack = stack;
2132 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2133 to_fill->_msg = deopt_resume2;
2134 } else {
2135 to_fill->_msg = method_resume;
2136 }
2137 to_fill->_result._to_call._bcp_advance = 0;
2138 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2139 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2140 to_fill->_prev_link = NULL;
2141
2142 // Fill in the registers for the frame
2143
2144 // Need to install _sender_sp. Actually not too hard in C++!
2145 // When the skeletal frames are layed out we fill in a value
2146 // for _sender_sp. That value is only correct for the oldest
2147 // skeletal frame constructed (because there is only a single
2148 // entry for "caller_adjustment". While the skeletal frames
2149 // exist that is good enough. We correct that calculation
2150 // here and get all the frames correct.
2151
2152 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2153
2154 *current->register_addr(Lstate) = (intptr_t) to_fill;
2155 // skeletal already places a useful value here and this doesn't account
2156 // for alignment so don't bother.
2157 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2158
2159 if (caller->is_interpreted_frame()) {
2160 interpreterState prev = caller->get_interpreterState();
2161 to_fill->_prev_link = prev;
2162 // Make the prev callee look proper
2163 prev->_result._to_call._callee = method;
2164 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2165 prev->_result._to_call._bcp_advance = 5;
2166 } else {
2167 prev->_result._to_call._bcp_advance = 3;
2168 }
2169 }
2170 to_fill->_oop_temp = NULL;
2171 to_fill->_stack_base = stack_base;
2172 // Need +1 here because stack_base points to the word just above the first expr stack entry
2173 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2174 // See generate_compute_interpreter_state.
2175 to_fill->_stack_limit = stack_base - (method->max_stack() + 1);
2176 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2177
2178 // sparc specific
2179 to_fill->_frame_bottom = frame_bottom;
2180 to_fill->_self_link = to_fill;
2181 #ifdef ASSERT
2182 to_fill->_native_fresult = 123456.789;
2183 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2184 #endif
2185 }
2186
pd_layout_interpreterState(interpreterState istate,address last_Java_pc,intptr_t * last_Java_fp)2187 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2188 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2189 }
2190
frame_size_helper(int max_stack,int moncount,int callee_param_size,int callee_locals_size,bool is_top_frame,int & monitor_size,int & full_frame_words)2191 static int frame_size_helper(int max_stack,
2192 int moncount,
2193 int callee_param_size,
2194 int callee_locals_size,
2195 bool is_top_frame,
2196 int& monitor_size,
2197 int& full_frame_words) {
2198 int extra_locals_size = callee_locals_size - callee_param_size;
2199 monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2200 full_frame_words = size_activation_helper(extra_locals_size, max_stack, monitor_size);
2201 int short_frame_words = size_activation_helper(extra_locals_size, max_stack, monitor_size);
2202 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2203
2204 return frame_words;
2205 }
2206
size_activation(int max_stack,int tempcount,int extra_args,int moncount,int callee_param_size,int callee_locals_size,bool is_top_frame)2207 int AbstractInterpreter::size_activation(int max_stack,
2208 int tempcount,
2209 int extra_args,
2210 int moncount,
2211 int callee_param_size,
2212 int callee_locals_size,
2213 bool is_top_frame) {
2214 assert(extra_args == 0, "NEED TO FIX");
2215 // NOTE: return size is in words not bytes
2216 // Calculate the amount our frame will be adjust by the callee. For top frame
2217 // this is zero.
2218
2219 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2220 // calculates the extra locals based on itself. Not what the callee does
2221 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2222 // as getting sender_sp correct.
2223
2224 int unused_monitor_size = 0;
2225 int unused_full_frame_words = 0;
2226 return frame_size_helper(max_stack, moncount, callee_param_size, callee_locals_size, is_top_frame,
2227 unused_monitor_size, unused_full_frame_words);
2228 }
layout_activation(Method * method,int tempcount,int popframe_extra_args,int moncount,int caller_actual_parameters,int callee_param_size,int callee_locals_size,frame * caller,frame * interpreter_frame,bool is_top_frame,bool is_bottom_frame)2229 void AbstractInterpreter::layout_activation(Method* method,
2230 int tempcount, // Number of slots on java expression stack in use
2231 int popframe_extra_args,
2232 int moncount, // Number of active monitors
2233 int caller_actual_parameters,
2234 int callee_param_size,
2235 int callee_locals_size,
2236 frame* caller,
2237 frame* interpreter_frame,
2238 bool is_top_frame,
2239 bool is_bottom_frame) {
2240 assert(popframe_extra_args == 0, "NEED TO FIX");
2241 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2242 // does as far as allocating an interpreter frame.
2243 // Set up the method, locals, and monitors.
2244 // The frame interpreter_frame is guaranteed to be the right size,
2245 // as determined by a previous call to the size_activation() method.
2246 // It is also guaranteed to be walkable even though it is in a skeletal state
2247 // NOTE: tempcount is the current size of the java expression stack. For top most
2248 // frames we will allocate a full sized expression stack and not the curback
2249 // version that non-top frames have.
2250
2251 int monitor_size = 0;
2252 int full_frame_words = 0;
2253 int frame_words = frame_size_helper(method->max_stack(), moncount, callee_param_size, callee_locals_size,
2254 is_top_frame, monitor_size, full_frame_words);
2255
2256 /*
2257 We must now fill in all the pieces of the frame. This means both
2258 the interpreterState and the registers.
2259 */
2260
2261 // MUCHO HACK
2262
2263 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2264 // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2265 assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2266 frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2267
2268 /* Now fillin the interpreterState object */
2269
2270 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2271
2272
2273 intptr_t* locals;
2274
2275 // Calculate the postion of locals[0]. This is painful because of
2276 // stack alignment (same as ia64). The problem is that we can
2277 // not compute the location of locals from fp(). fp() will account
2278 // for the extra locals but it also accounts for aligning the stack
2279 // and we can't determine if the locals[0] was misaligned but max_locals
2280 // was enough to have the
2281 // calculate postion of locals. fp already accounts for extra locals.
2282 // +2 for the static long no_params() issue.
2283
2284 if (caller->is_interpreted_frame()) {
2285 // locals must agree with the caller because it will be used to set the
2286 // caller's tos when we return.
2287 interpreterState prev = caller->get_interpreterState();
2288 // stack() is prepushed.
2289 locals = prev->stack() + method->size_of_parameters();
2290 } else {
2291 // Lay out locals block in the caller adjacent to the register window save area.
2292 //
2293 // Compiled frames do not allocate a varargs area which is why this if
2294 // statement is needed.
2295 //
2296 intptr_t* fp = interpreter_frame->fp();
2297 int local_words = method->max_locals() * Interpreter::stackElementWords;
2298
2299 if (caller->is_compiled_frame()) {
2300 locals = fp + frame::register_save_words + local_words - 1;
2301 } else {
2302 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2303 }
2304
2305 }
2306 // END MUCHO HACK
2307
2308 intptr_t* monitor_base = (intptr_t*) cur_state;
2309 intptr_t* stack_base = monitor_base - monitor_size;
2310 /* +1 because stack is always prepushed */
2311 intptr_t* stack = stack_base - (tempcount + 1);
2312
2313
2314 BytecodeInterpreter::layout_interpreterState(cur_state,
2315 caller,
2316 interpreter_frame,
2317 method,
2318 locals,
2319 stack,
2320 stack_base,
2321 monitor_base,
2322 frame_bottom,
2323 is_top_frame);
2324
2325 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2326 }
2327
2328 #endif // CC_INTERP
2329