1 #ifdef USE_PRAGMA_IDENT_HDR
2 #pragma ident "@(#)cppInterpreter_sparc.cpp 1.1 07/08/29 13:42:16 JVM"
3 #endif
4 /*
5 * Copyright 2007 Sun Microsystems, Inc. All Rights Reserved.
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7 *
8 * This code is free software; you can redistribute it and/or modify it
9 * under the terms of the GNU General Public License version 2 only, as
10 * published by the Free Software Foundation.
11 *
12 * This code is distributed in the hope that it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 * version 2 for more details (a copy is included in the LICENSE file that
16 * accompanied this code).
17 *
18 * You should have received a copy of the GNU General Public License version
19 * 2 along with this work; if not, write to the Free Software Foundation,
20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
21 *
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
24 * have any questions.
25 *
26 */
27
28 #include "incls/_precompiled.incl"
29 #include "incls/_cppInterpreter_sparc.cpp.incl"
30
31 #ifdef CC_INTERP
32
33 // Routine exists to make tracebacks look decent in debugger
34 // while "shadow" interpreter frames are on stack. It is also
35 // used to distinguish interpreter frames.
36
37 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
38 ShouldNotReachHere();
39 }
40
41 bool CppInterpreter::contains(address pc) {
42 return ( _code->contains(pc) ||
43 ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
44 }
45
46 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
47 #define __ _masm->
48
49 Label frame_manager_entry;
50 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized
51 // c++ interpreter entry point this holds that entry point label.
52
53 static address unctrap_frame_manager_entry = NULL;
54
55 static address interpreter_return_address = NULL;
56 static address deopt_frame_manager_return_atos = NULL;
57 static address deopt_frame_manager_return_btos = NULL;
58 static address deopt_frame_manager_return_itos = NULL;
59 static address deopt_frame_manager_return_ltos = NULL;
60 static address deopt_frame_manager_return_ftos = NULL;
61 static address deopt_frame_manager_return_dtos = NULL;
62 static address deopt_frame_manager_return_vtos = NULL;
63
64 const Register prevState = G1_scratch;
65
66 void InterpreterGenerator::save_native_result(void) {
67 // result potentially in O0/O1: save it across calls
68 __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
69 #ifdef _LP64
70 __ stx(O0, STATE(_native_lresult));
71 #else
72 __ std(O0, STATE(_native_lresult));
73 #endif
74 }
75
76 void InterpreterGenerator::restore_native_result(void) {
77
78 // Restore any method result value
79 __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
80 #ifdef _LP64
81 __ ldx(STATE(_native_lresult), O0);
82 #else
83 __ ldd(STATE(_native_lresult), O0);
84 #endif
85 }
86
87 // A result handler converts/unboxes a native call result into
88 // a java interpreter/compiler result. The current frame is an
89 // interpreter frame. The activation frame unwind code must be
90 // consistent with that of TemplateTable::_return(...). In the
91 // case of native methods, the caller's SP was not modified.
92 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
93 address entry = __ pc();
94 Register Itos_i = Otos_i ->after_save();
95 Register Itos_l = Otos_l ->after_save();
96 Register Itos_l1 = Otos_l1->after_save();
97 Register Itos_l2 = Otos_l2->after_save();
98 switch (type) {
99 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
100 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
101 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
102 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
103 case T_LONG :
104 #ifndef _LP64
105 __ mov(O1, Itos_l2); // move other half of long
106 #endif // ifdef or no ifdef, fall through to the T_INT case
107 case T_INT : __ mov(O0, Itos_i); break;
108 case T_VOID : /* nothing to do */ break;
109 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
110 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
111 case T_OBJECT :
112 __ ld_ptr(STATE(_oop_temp), Itos_i);
113 __ verify_oop(Itos_i);
114 break;
115 default : ShouldNotReachHere();
116 }
117 __ ret(); // return from interpreter activation
118 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
119 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
120 return entry;
121 }
122
123 // tosca based result to c++ interpreter stack based result.
124 // Result goes to address in L1_scratch
125
126 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
127 // A result is in the native abi result register from a native method call.
128 // We need to return this result to the interpreter by pushing the result on the interpreter's
129 // stack. This is relatively simple the destination is in L1_scratch
130 // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
131 // adjust L1_scratch
132 address entry = __ pc();
133 switch (type) {
134 case T_BOOLEAN:
135 // !0 => true; 0 => false
136 __ subcc(G0, O0, G0);
137 __ addc(G0, 0, O0);
138 __ st(O0, L1_scratch, 0);
139 __ sub(L1_scratch, wordSize, L1_scratch);
140 break;
141
142 // cannot use and3, 0xFFFF too big as immediate value!
143 case T_CHAR :
144 __ sll(O0, 16, O0);
145 __ srl(O0, 16, O0);
146 __ st(O0, L1_scratch, 0);
147 __ sub(L1_scratch, wordSize, L1_scratch);
148 break;
149
150 case T_BYTE :
151 __ sll(O0, 24, O0);
152 __ sra(O0, 24, O0);
153 __ st(O0, L1_scratch, 0);
154 __ sub(L1_scratch, wordSize, L1_scratch);
155 break;
156
157 case T_SHORT :
158 __ sll(O0, 16, O0);
159 __ sra(O0, 16, O0);
160 __ st(O0, L1_scratch, 0);
161 __ sub(L1_scratch, wordSize, L1_scratch);
162 break;
163 case T_LONG :
164 #ifndef _LP64
165 #if !defined(_LP64) && defined(COMPILER2)
166 // All return values are where we want them, except for Longs. C2 returns
167 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
168 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
169 // build even if we are returning from interpreted we just do a little
170 // stupid shuffing.
171 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
172 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
173 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
174 __ stx(G1, L1_scratch, -wordSize);
175 #else
176 // native result is in O0, O1
177 __ st(O1, L1_scratch, 0); // Low order
178 __ st(O0, L1_scratch, -wordSize); // High order
179 #endif /* !_LP64 && COMPILER2 */
180 #else
181 __ stx(O0, L1_scratch, 0);
182 __ breakpoint_trap();
183 #endif
184 __ sub(L1_scratch, 2*wordSize, L1_scratch);
185 break;
186
187 case T_INT :
188 __ st(O0, L1_scratch, 0);
189 __ sub(L1_scratch, wordSize, L1_scratch);
190 break;
191
192 case T_VOID : /* nothing to do */
193 break;
194
195 case T_FLOAT :
196 __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
197 __ sub(L1_scratch, wordSize, L1_scratch);
198 break;
199
200 case T_DOUBLE :
201 // Every stack slot is aligned on 64 bit, However is this
202 // the correct stack slot on 64bit?? QQQ
203 __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
204 __ sub(L1_scratch, 2*wordSize, L1_scratch);
205 break;
206 case T_OBJECT :
207 __ verify_oop(O0);
208 __ st_ptr(O0, L1_scratch, 0);
209 __ sub(L1_scratch, wordSize, L1_scratch);
210 break;
211 default : ShouldNotReachHere();
212 }
213 __ retl(); // return from interpreter activation
214 __ delayed()->nop(); // schedule this better
215 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
216 return entry;
217 }
218
219 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
220 // A result is in the java expression stack of the interpreted method that has just
221 // returned. Place this result on the java expression stack of the caller.
222 //
223 // The current interpreter activation in Lstate is for the method just returning its
224 // result. So we know that the result of this method is on the top of the current
225 // execution stack (which is pre-pushed) and will be return to the top of the caller
226 // stack. The top of the callers stack is the bottom of the locals of the current
227 // activation.
228 // Because of the way activation are managed by the frame manager the value of esp is
229 // below both the stack top of the current activation and naturally the stack top
230 // of the calling activation. This enable this routine to leave the return address
231 // to the frame manager on the stack and do a vanilla return.
232 //
233 // On entry: O0 - points to source (callee stack top)
234 // O1 - points to destination (caller stack top [i.e. free location])
235 // destroys O2, O3
236 //
237
238 address entry = __ pc();
239 switch (type) {
240 case T_VOID: break;
241 break;
242 case T_FLOAT :
243 __ breakpoint_trap(Assembler::zero);
244 case T_BOOLEAN:
245 case T_CHAR :
246 case T_BYTE :
247 case T_SHORT :
248 case T_INT :
249 // 1 word result
250 __ ld(O0, 0, O2);
251 __ st(O2, O1, 0);
252 __ sub(O1, wordSize, O1);
253 break;
254 case T_DOUBLE :
255 case T_LONG :
256 // return top two words on current expression stack to caller's expression stack
257 // The caller's expression stack is adjacent to the current frame manager's intepretState
258 // except we allocated one extra word for this intepretState so we won't overwrite it
259 // when we return a two word result.
260 #ifdef _LP64
261 __ breakpoint_trap();
262 // Hmm now that longs are in one entry should "_ptr" really be "x"?
263 __ ld_ptr(O0, 0, O2);
264 __ ld_ptr(O0, wordSize, O3);
265 __ st_ptr(O3, O1, 0);
266 __ st_ptr(O2, O1, -wordSize);
267 #else
268 __ ld(O0, 0, O2);
269 __ ld(O0, wordSize, O3);
270 __ st(O3, O1, 0);
271 __ st(O2, O1, -wordSize);
272 #endif
273 __ sub(O1, 2*wordSize, O1);
274 break;
275 case T_OBJECT :
276 __ ld_ptr(O0, 0, O2);
277 __ verify_oop(O2); // verify it
278 __ st_ptr(O2, O1, 0);
279 __ sub(O1, wordSize, O1);
280 break;
281 default : ShouldNotReachHere();
282 }
283 __ retl();
284 __ delayed()->nop(); // QQ schedule this better
285 return entry;
286 }
287
288 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
289 // A result is in the java expression stack of the interpreted method that has just
290 // returned. Place this result in the native abi that the caller expects.
291 // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
292 //
293 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
294 // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
295 // and so rather than return result onto caller's java expression stack we return the
296 // result in the expected location based on the native abi.
297 // On entry: O0 - source (stack top)
298 // On exit result in expected output register
299 // QQQ schedule this better
300
301 address entry = __ pc();
302 switch (type) {
303 case T_VOID: break;
304 break;
305 case T_FLOAT :
306 __ ldf(FloatRegisterImpl::S, O0, 0, F0);
307 break;
308 case T_BOOLEAN:
309 case T_CHAR :
310 case T_BYTE :
311 case T_SHORT :
312 case T_INT :
313 // 1 word result
314 __ ld(O0, 0, O0->after_save());
315 break;
316 case T_DOUBLE :
317 __ ldf(FloatRegisterImpl::D, O0, 0, F0);
318 break;
319 case T_LONG :
320 // return top two words on current expression stack to caller's expression stack
321 // The caller's expression stack is adjacent to the current frame manager's interpretState
322 // except we allocated one extra word for this intepretState so we won't overwrite it
323 // when we return a two word result.
324 #ifdef _LP64
325 __ breakpoint_trap();
326 // Hmm now that longs are in one entry should "_ptr" really be "x"?
327 __ ld_ptr(O0, 0, O0->after_save());
328 __ ld_ptr(O0, wordSize, O1->after_save());
329 #else
330 __ ld(O0, wordSize, O1->after_save());
331 __ ld(O0, 0, O0->after_save());
332 #endif
333 #if defined(COMPILER2) && !defined(_LP64)
334 // C2 expects long results in G1 we can't tell if we're returning to interpreted
335 // or compiled so just be safe use G1 and O0/O1
336
337 // Shift bits into high (msb) of G1
338 __ sllx(Otos_l1->after_save(), 32, G1);
339 // Zero extend low bits
340 __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
341 __ or3 (Otos_l2->after_save(), G1, G1);
342 #endif /* COMPILER2 */
343 break;
344 case T_OBJECT :
345 __ ld_ptr(O0, 0, O0->after_save());
346 __ verify_oop(O0->after_save()); // verify it
347 break;
348 default : ShouldNotReachHere();
349 }
350 __ retl();
351 __ delayed()->nop();
352 return entry;
353 }
354
355 address CppInterpreter::return_entry(TosState state, int length) {
356 // make it look good in the debugger
357 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
358 }
359
360 address CppInterpreter::deopt_entry(TosState state, int length) {
361 address ret = NULL;
362 if (length != 0) {
363 switch (state) {
364 case atos: ret = deopt_frame_manager_return_atos; break;
365 case btos: ret = deopt_frame_manager_return_btos; break;
366 case ctos:
367 case stos:
368 case itos: ret = deopt_frame_manager_return_itos; break;
369 case ltos: ret = deopt_frame_manager_return_ltos; break;
370 case ftos: ret = deopt_frame_manager_return_ftos; break;
371 case dtos: ret = deopt_frame_manager_return_dtos; break;
372 case vtos: ret = deopt_frame_manager_return_vtos; break;
373 }
374 } else {
375 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap)
376 }
377 assert(ret != NULL, "Not initialized");
378 return ret;
379 }
380
381 //
382 // Helpers for commoning out cases in the various type of method entries.
383 //
384
385 // increment invocation count & check for overflow
386 //
387 // Note: checking for negative value instead of overflow
388 // so we have a 'sticky' overflow test
389 //
390 // Lmethod: method
391 // ??: invocation counter
392 //
393 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
394 // Update standard invocation counters
395 __ increment_invocation_counter(O0, G3_scratch);
396 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
397 __ ld_ptr(STATE(_method), G3_scratch);
398 Address interpreter_invocation_counter(G3_scratch, 0, in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
399 __ ld(interpreter_invocation_counter, G3_scratch);
400 __ inc(G3_scratch);
401 __ st(G3_scratch, interpreter_invocation_counter);
402 }
403
404 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
405 __ sethi(invocation_limit);
406 __ ld(invocation_limit, G3_scratch);
407 __ cmp(O0, G3_scratch);
408 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
409 __ delayed()->nop();
410
411 }
412
413 address InterpreterGenerator::generate_empty_entry(void) {
414
415 // A method that does nothing but return...
416
417 address entry = __ pc();
418 Label slow_path;
419
420 __ verify_oop(G5_method);
421
422 // do nothing for empty methods (do not even increment invocation counter)
423 if ( UseFastEmptyMethods) {
424 // If we need a safepoint check, generate full interpreter entry.
425 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
426 __ load_contents(sync_state, G3_scratch);
427 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
428 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
429 __ delayed()->nop();
430
431 // Code: _return
432 __ retl();
433 __ delayed()->mov(O5_savedSP, SP);
434 return entry;
435 }
436 return NULL;
437 }
438
439 // Call an accessor method (assuming it is resolved, otherwise drop into
440 // vanilla (slow path) entry
441
442 // Generates code to elide accessor methods
443 // Uses G3_scratch and G1_scratch as scratch
444 address InterpreterGenerator::generate_accessor_entry(void) {
445
446 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
447 // parameter size = 1
448 // Note: We can only use this code if the getfield has been resolved
449 // and if we don't have a null-pointer exception => check for
450 // these conditions first and use slow path if necessary.
451 address entry = __ pc();
452 Label slow_path;
453
454 if ( UseFastAccessorMethods) {
455 // Check if we need to reach a safepoint and generate full interpreter
456 // frame if so.
457 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
458 __ load_contents(sync_state, G3_scratch);
459 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
460 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
461 __ delayed()->nop();
462
463 // Check if local 0 != NULL
464 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
465 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
466 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
467 __ delayed()->nop();
468
469
470 // read first instruction word and extract bytecode @ 1 and index @ 2
471 // get first 4 bytes of the bytecodes (big endian!)
472 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), G1_scratch);
473 __ ld(Address(G1_scratch, 0, in_bytes(constMethodOopDesc::codes_offset())), G1_scratch);
474
475 // move index @ 2 far left then to the right most two bytes.
476 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
477 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
478 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
479
480 // get constant pool cache
481 __ ld_ptr(G5_method, in_bytes(methodOopDesc::constants_offset()), G3_scratch);
482 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
483
484 // get specific constant pool cache entry
485 __ add(G3_scratch, G1_scratch, G3_scratch);
486
487 // Check the constant Pool cache entry to see if it has been resolved.
488 // If not, need the slow path.
489 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
490 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
491 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
492 __ and3(G1_scratch, 0xFF, G1_scratch);
493 __ cmp(G1_scratch, Bytecodes::_getfield);
494 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
495 __ delayed()->nop();
496
497 // Get the type and return field offset from the constant pool cache
498 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
499 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
500
501 Label xreturn_path;
502 // Need to differentiate between igetfield, agetfield, bgetfield etc.
503 // because they are different sizes.
504 // Get the type from the constant pool cache
505 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
506 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
507 ConstantPoolCacheEntry::verify_tosBits();
508 __ cmp(G1_scratch, atos );
509 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
510 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
511 __ cmp(G1_scratch, itos);
512 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
513 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
514 __ cmp(G1_scratch, stos);
515 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
516 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
517 __ cmp(G1_scratch, ctos);
518 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
519 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
520 #ifdef ASSERT
521 __ cmp(G1_scratch, btos);
522 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
523 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
524 __ should_not_reach_here();
525 #endif
526 __ ldsb(Otos_i, G3_scratch, Otos_i);
527 __ bind(xreturn_path);
528
529 // _ireturn/_areturn
530 __ retl(); // return from leaf routine
531 __ delayed()->mov(O5_savedSP, SP);
532
533 // Generate regular method entry
534 __ bind(slow_path);
535 __ ba(false, fast_accessor_slow_entry_path);
536 __ delayed()->nop();
537 return entry;
538 }
539 return NULL;
540 }
541
542 //
543 // Interpreter stub for calling a native method. (C++ interpreter)
544 // This sets up a somewhat different looking stack for calling the native method
545 // than the typical interpreter frame setup.
546 //
547
548 address InterpreterGenerator::generate_native_entry(bool synchronized) {
549 address entry = __ pc();
550
551 // the following temporary registers are used during frame creation
552 const Register Gtmp1 = G3_scratch ;
553 const Register Gtmp2 = G1_scratch;
554 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
555
556 bool inc_counter = UseCompiler || CountCompiledCalls;
557
558 // make sure registers are different!
559 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
560
561 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
562
563 Label Lentry;
564 __ bind(Lentry);
565
566 __ verify_oop(G5_method);
567
568 const Register Glocals_size = G3;
569 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
570
571 // make sure method is native & not abstract
572 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
573 #ifdef ASSERT
574 __ ld(access_flags, Gtmp1);
575 {
576 Label L;
577 __ btst(JVM_ACC_NATIVE, Gtmp1);
578 __ br(Assembler::notZero, false, Assembler::pt, L);
579 __ delayed()->nop();
580 __ stop("tried to execute non-native method as native");
581 __ bind(L);
582 }
583 { Label L;
584 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
585 __ br(Assembler::zero, false, Assembler::pt, L);
586 __ delayed()->nop();
587 __ stop("tried to execute abstract method as non-abstract");
588 __ bind(L);
589 }
590 #endif // ASSERT
591
592 __ lduh(size_of_parameters, Gtmp1);
593 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes
594 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord
595 // NEW
596 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
597 // generate the code to allocate the interpreter stack frame
598 // NEW FRAME ALLOCATED HERE
599 // save callers original sp
600 // __ mov(SP, I5_savedSP->after_restore());
601
602 generate_compute_interpreter_state(Lstate, G0, true);
603
604 // At this point Lstate points to new interpreter state
605 //
606
607 const Address do_not_unlock_if_synchronized(G2_thread, 0,
608 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
609 // Since at this point in the method invocation the exception handler
610 // would try to exit the monitor of synchronized methods which hasn't
611 // been entered yet, we set the thread local variable
612 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
613 // runtime, exception handling i.e. unlock_if_synchronized_method will
614 // check this thread local flag.
615 // This flag has two effects, one is to force an unwind in the topmost
616 // interpreter frame and not perform an unlock while doing so.
617
618 __ movbool(true, G3_scratch);
619 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
620
621
622 // increment invocation counter and check for overflow
623 //
624 // Note: checking for negative value instead of overflow
625 // so we have a 'sticky' overflow test (may be of
626 // importance as soon as we have true MT/MP)
627 Label invocation_counter_overflow;
628 if (inc_counter) {
629 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
630 }
631 Label Lcontinue;
632 __ bind(Lcontinue);
633
634 bang_stack_shadow_pages(true);
635 // reset the _do_not_unlock_if_synchronized flag
636 __ stbool(G0, do_not_unlock_if_synchronized);
637
638 // check for synchronized methods
639 // Must happen AFTER invocation_counter check, so method is not locked
640 // if counter overflows.
641
642 if (synchronized) {
643 lock_method();
644 // Don't see how G2_thread is preserved here...
645 // __ verify_thread(); QQQ destroys L0,L1 can't use
646 } else {
647 #ifdef ASSERT
648 { Label ok;
649 __ ld_ptr(STATE(_method), G5_method);
650 __ ld(access_flags, O0);
651 __ btst(JVM_ACC_SYNCHRONIZED, O0);
652 __ br( Assembler::zero, false, Assembler::pt, ok);
653 __ delayed()->nop();
654 __ stop("method needs synchronization");
655 __ bind(ok);
656 }
657 #endif // ASSERT
658 }
659
660 // start execution
661
662 // __ verify_thread(); kills L1,L2 can't use at the moment
663
664 // jvmti/jvmpi support
665 __ notify_method_entry();
666
667 // native call
668
669 // (note that O0 is never an oop--at most it is a handle)
670 // It is important not to smash any handles created by this call,
671 // until any oop handle in O0 is dereferenced.
672
673 // (note that the space for outgoing params is preallocated)
674
675 // get signature handler
676
677 Label pending_exception_present;
678
679 { Label L;
680 __ ld_ptr(STATE(_method), G5_method);
681 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
682 __ tst(G3_scratch);
683 __ brx(Assembler::notZero, false, Assembler::pt, L);
684 __ delayed()->nop();
685 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
686 __ ld_ptr(STATE(_method), G5_method);
687
688 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
689 __ ld_ptr(exception_addr, G3_scratch);
690 __ br_notnull(G3_scratch, false, Assembler::pn, pending_exception_present);
691 __ delayed()->nop();
692 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
693 __ bind(L);
694 }
695
696 // Push a new frame so that the args will really be stored in
697 // Copy a few locals across so the new frame has the variables
698 // we need but these values will be dead at the jni call and
699 // therefore not gc volatile like the values in the current
700 // frame (Lstate in particular)
701
702 // Flush the state pointer to the register save area
703 // Which is the only register we need for a stack walk.
704 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
705
706 __ mov(Lstate, O1); // Need to pass the state pointer across the frame
707
708 // Calculate current frame size
709 __ sub(SP, FP, O3); // Calculate negative of current frame size
710 __ save(SP, O3, SP); // Allocate an identical sized frame
711
712 __ mov(I1, Lstate); // In the "natural" register.
713
714 // Note I7 has leftover trash. Slow signature handler will fill it in
715 // should we get there. Normal jni call will set reasonable last_Java_pc
716 // below (and fix I7 so the stack trace doesn't have a meaningless frame
717 // in it).
718
719
720 // call signature handler
721 __ ld_ptr(STATE(_method), Lmethod);
722 __ ld_ptr(STATE(_locals), Llocals);
723
724 __ callr(G3_scratch, 0);
725 __ delayed()->nop();
726 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed)
727
728 { Label not_static;
729
730 __ ld_ptr(STATE(_method), G5_method);
731 __ ld(access_flags, O0);
732 __ btst(JVM_ACC_STATIC, O0);
733 __ br( Assembler::zero, false, Assembler::pt, not_static);
734 __ delayed()->
735 // get native function entry point(O0 is a good temp until the very end)
736 ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::native_function_offset())), O0);
737 // for static methods insert the mirror argument
738 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
739
740 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc:: constants_offset())), O1);
741 __ ld_ptr(Address(O1, 0, constantPoolOopDesc::pool_holder_offset_in_bytes()), O1);
742 __ ld_ptr(O1, mirror_offset, O1);
743 // where the mirror handle body is allocated:
744 #ifdef ASSERT
745 if (!PrintSignatureHandlers) // do not dirty the output with this
746 { Label L;
747 __ tst(O1);
748 __ brx(Assembler::notZero, false, Assembler::pt, L);
749 __ delayed()->nop();
750 __ stop("mirror is missing");
751 __ bind(L);
752 }
753 #endif // ASSERT
754 __ st_ptr(O1, STATE(_oop_temp));
755 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add
756 __ bind(not_static);
757 }
758
759 // At this point, arguments have been copied off of stack into
760 // their JNI positions, which are O1..O5 and SP[68..].
761 // Oops are boxed in-place on the stack, with handles copied to arguments.
762 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
763
764 #ifdef ASSERT
765 { Label L;
766 __ tst(O0);
767 __ brx(Assembler::notZero, false, Assembler::pt, L);
768 __ delayed()->nop();
769 __ stop("native entry point is missing");
770 __ bind(L);
771 }
772 #endif // ASSERT
773
774 //
775 // setup the java frame anchor
776 //
777 // The scavenge function only needs to know that the PC of this frame is
778 // in the interpreter method entry code, it doesn't need to know the exact
779 // PC and hence we can use O7 which points to the return address from the
780 // previous call in the code stream (signature handler function)
781 //
782 // The other trick is we set last_Java_sp to FP instead of the usual SP because
783 // we have pushed the extra frame in order to protect the volatile register(s)
784 // in that frame when we return from the jni call
785 //
786
787
788 __ set_last_Java_frame(FP, O7);
789 __ mov(O7, I7); // make dummy interpreter frame look like one above,
790 // not meaningless information that'll confuse me.
791
792 // flush the windows now. We don't care about the current (protection) frame
793 // only the outer frames
794
795 __ flush_windows();
796
797 // mark windows as flushed
798 Address flags(G2_thread,
799 0,
800 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
801 __ set(JavaFrameAnchor::flushed, G3_scratch);
802 __ st(G3_scratch, flags);
803
804 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
805
806 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
807 #ifdef ASSERT
808 { Label L;
809 __ ld(thread_state, G3_scratch);
810 __ cmp(G3_scratch, _thread_in_Java);
811 __ br(Assembler::equal, false, Assembler::pt, L);
812 __ delayed()->nop();
813 __ stop("Wrong thread state in native stub");
814 __ bind(L);
815 }
816 #endif // ASSERT
817 __ set(_thread_in_native, G3_scratch);
818 __ st(G3_scratch, thread_state);
819
820 // Call the jni method, using the delay slot to set the JNIEnv* argument.
821 __ callr(O0, 0);
822 __ delayed()->
823 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
824 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
825
826 // must we block?
827
828 // Block, if necessary, before resuming in _thread_in_Java state.
829 // In order for GC to work, don't clear the last_Java_sp until after blocking.
830 { Label no_block;
831 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
832
833 // Switch thread to "native transition" state before reading the synchronization state.
834 // This additional state is necessary because reading and testing the synchronization
835 // state is not atomic w.r.t. GC, as this scenario demonstrates:
836 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
837 // VM thread changes sync state to synchronizing and suspends threads for GC.
838 // Thread A is resumed to finish this native method, but doesn't block here since it
839 // didn't see any synchronization is progress, and escapes.
840 __ set(_thread_in_native_trans, G3_scratch);
841 __ st(G3_scratch, thread_state);
842 if(os::is_MP()) {
843 // Write serialization page so VM thread can do a pseudo remote membar.
844 // We use the current thread pointer to calculate a thread specific
845 // offset to write to within the page. This minimizes bus traffic
846 // due to cache line collision.
847 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
848 }
849 __ load_contents(sync_state, G3_scratch);
850 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
851
852
853 Label L;
854 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
855 __ br(Assembler::notEqual, false, Assembler::pn, L);
856 __ delayed()->
857 ld(suspend_state, G3_scratch);
858 __ cmp(G3_scratch, 0);
859 __ br(Assembler::equal, false, Assembler::pt, no_block);
860 __ delayed()->nop();
861 __ bind(L);
862
863 // Block. Save any potential method result value before the operation and
864 // use a leaf call to leave the last_Java_frame setup undisturbed.
865 save_native_result();
866 __ call_VM_leaf(noreg,
867 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
868 G2_thread);
869 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
870 // Restore any method result value
871 restore_native_result();
872 __ bind(no_block);
873 }
874
875 // Clear the frame anchor now
876
877 __ reset_last_Java_frame();
878
879 // Move the result handler address
880 __ mov(Lscratch, G3_scratch);
881 // return possible result to the outer frame
882 #ifndef __LP64
883 __ mov(O0, I0);
884 __ restore(O1, G0, O1);
885 #else
886 __ restore(O0, G0, O0);
887 #endif /* __LP64 */
888
889 // Move result handler to expected register
890 __ mov(G3_scratch, Lscratch);
891
892
893 // thread state is thread_in_native_trans. Any safepoint blocking has
894 // happened in the trampoline we are ready to switch to thread_in_Java.
895
896 __ set(_thread_in_Java, G3_scratch);
897 __ st(G3_scratch, thread_state);
898
899 // If we have an oop result store it where it will be safe for any further gc
900 // until we return now that we've released the handle it might be protected by
901
902 {
903 Label no_oop, store_result;
904
905 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
906 __ cmp(G3_scratch, Lscratch);
907 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
908 __ delayed()->nop();
909 __ addcc(G0, O0, O0);
910 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
911 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
912 __ mov(G0, O0);
913
914 __ bind(store_result);
915 // Store it where gc will look for it and result handler expects it.
916 __ st_ptr(O0, STATE(_oop_temp));
917
918 __ bind(no_oop);
919
920 }
921
922 // reset handle block
923 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
924 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
925
926
927 // handle exceptions (exception handling will handle unlocking!)
928 { Label L;
929 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
930
931 __ ld_ptr(exception_addr, Gtemp);
932 __ tst(Gtemp);
933 __ brx(Assembler::equal, false, Assembler::pt, L);
934 __ delayed()->nop();
935 __ bind(pending_exception_present);
936 // With c++ interpreter we just leave it pending caller will do the correct thing. However...
937 // Like x86 we ignore the result of the native call and leave the method locked. This
938 // seems wrong to leave things locked.
939
940 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
941 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
942
943 __ bind(L);
944 }
945
946 // jvmdi/jvmpi support (preserves thread register)
947 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
948
949 if (synchronized) {
950 // save and restore any potential method result value around the unlocking operation
951 save_native_result();
952
953 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
954 // Get the initial monitor we allocated
955 __ sub(Lstate, entry_size, O1); // initial monitor
956 __ unlock_object(O1);
957 restore_native_result();
958 }
959
960 #if defined(COMPILER2) && !defined(_LP64)
961
962 // C2 expects long results in G1 we can't tell if we're returning to interpreted
963 // or compiled so just be safe.
964
965 __ sllx(O0, 32, G1); // Shift bits into high G1
966 __ srl (O1, 0, O1); // Zero extend O1
967 __ or3 (O1, G1, G1); // OR 64 bits into G1
968
969 #endif /* COMPILER2 && !_LP64 */
970
971 #ifdef ASSERT
972 {
973 Label ok;
974 __ cmp(I5_savedSP, FP);
975 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
976 __ delayed()->nop();
977 __ stop("bad I5_savedSP value");
978 __ should_not_reach_here();
979 __ bind(ok);
980 }
981 #endif
982 // Calls result handler which POPS FRAME
983 if (TraceJumps) {
984 // Move target to register that is recordable
985 __ mov(Lscratch, G3_scratch);
986 __ JMP(G3_scratch, 0);
987 } else {
988 __ jmp(Lscratch, 0);
989 }
990 __ delayed()->nop();
991
992 if (inc_counter) {
993 // handle invocation counter overflow
994 __ bind(invocation_counter_overflow);
995 generate_counter_overflow(Lcontinue);
996 }
997
998
999 return entry;
1000 }
1001
1002 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1003 const Register prev_state,
1004 bool native) {
1005
1006 // On entry
1007 // G5_method - caller's method
1008 // Gargs - points to initial parameters (i.e. locals[0])
1009 // G2_thread - valid? (C1 only??)
1010 // "prev_state" - contains any previous frame manager state which we must save a link
1011 //
1012 // On return
1013 // "state" is a pointer to the newly allocated state object. We must allocate and initialize
1014 // a new interpretState object and the method expression stack.
1015
1016 assert_different_registers(state, prev_state);
1017 assert_different_registers(prev_state, G3_scratch);
1018 const Register Gtmp = G3_scratch;
1019 const Address constants (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1020 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1021 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1022 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1023 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1024
1025 // slop factor is two extra slots on the expression stack so that
1026 // we always have room to store a result when returning from a call without parameters
1027 // that returns a result.
1028
1029 const int slop_factor = 2*wordSize;
1030
1031 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1032 frame::memory_parameter_word_sp_offset + // register save area + param window
1033 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1034
1035 // XXX G5_method valid
1036
1037 // Now compute new frame size
1038
1039 if (native) {
1040 __ lduh( size_of_parameters, Gtmp );
1041 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words
1042 } else {
1043 __ lduh(max_stack, Gtmp); // Full size expression stack
1044 }
1045 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1046
1047 __ neg(Gtmp); // negative space for stack/parameters in words
1048 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1049 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1050
1051 // Need to do stack size check here before we fault on large frames
1052
1053 Label stack_ok;
1054
1055 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1056 (StackRedPages+StackYellowPages);
1057
1058
1059 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1060 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1061 // compute stack bottom
1062 __ sub(O0, O1, O0);
1063
1064 // Avoid touching the guard pages
1065 // Also a fudge for frame size of BytecodeInterpreter::run
1066 // It varies from 1k->4k depending on build type
1067 const int fudge = 6 * K;
1068
1069 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1070
1071 __ add(O0, O1, O0);
1072 __ sub(O0, Gtmp, O0);
1073 __ cmp(SP, O0);
1074 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1075 __ delayed()->nop();
1076
1077 // throw exception return address becomes throwing pc
1078
1079 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1080 __ stop("never reached");
1081
1082 __ bind(stack_ok);
1083
1084 __ save(SP, Gtmp, SP); // setup new frame and register window
1085
1086 // New window I7 call_stub or previous activation
1087 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1088 //
1089 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1090 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1091
1092 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1093
1094 // Initialize a new Interpreter state
1095 // orig_sp - caller's original sp
1096 // G2_thread - thread
1097 // Gargs - &locals[0] (unbiased?)
1098 // G5_method - method
1099 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1100
1101
1102 __ set(0xdead0004, O1);
1103
1104
1105 __ st_ptr(Gargs, XXX_STATE(_locals));
1106 __ st_ptr(G0, XXX_STATE(_oop_temp));
1107
1108 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1109 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1110 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1111
1112 if (native) {
1113 __ st_ptr(G0, XXX_STATE(_bcp));
1114 } else {
1115 __ ld_ptr(G5_method, in_bytes(methodOopDesc::const_offset()), O2); // get constMethodOop
1116 __ add(O2, in_bytes(constMethodOopDesc::codes_offset()), O2); // get bcp
1117 __ st_ptr(O2, XXX_STATE(_bcp));
1118 }
1119
1120 __ st_ptr(G0, XXX_STATE(_mdx));
1121 __ st_ptr(G5_method, XXX_STATE(_method));
1122
1123 __ set((int) BytecodeInterpreter::method_entry, O1);
1124 __ st(O1, XXX_STATE(_msg));
1125
1126 __ ld_ptr(constants, O3);
1127 __ ld_ptr(O3, constantPoolOopDesc::cache_offset_in_bytes(), O2);
1128 __ st_ptr(O2, XXX_STATE(_constants));
1129
1130 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1131
1132 // Monitor base is just start of BytecodeInterpreter object;
1133 __ mov(state, O2);
1134 __ st_ptr(O2, XXX_STATE(_monitor_base));
1135
1136 // Do we need a monitor for synchonized method?
1137 {
1138 __ ld(access_flags, O1);
1139 Label done;
1140 Label got_obj;
1141 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1142 __ br( Assembler::zero, false, Assembler::pt, done);
1143
1144 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1145 __ delayed()->btst(JVM_ACC_STATIC, O1);
1146 __ ld_ptr(XXX_STATE(_locals), O1);
1147 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1148 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1149 __ ld_ptr(constants, O1);
1150 __ ld_ptr( O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
1151 // lock the mirror, not the klassOop
1152 __ ld_ptr( O1, mirror_offset, O1);
1153
1154 __ bind(got_obj);
1155
1156 #ifdef ASSERT
1157 __ tst(O1);
1158 __ breakpoint_trap(Assembler::zero);
1159 #endif // ASSERT
1160
1161 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1162 __ sub(SP, entry_size, SP); // account for initial monitor
1163 __ sub(O2, entry_size, O2); // initial monitor
1164 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1165 __ bind(done);
1166 }
1167
1168 // Remember initial frame bottom
1169
1170 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1171
1172 __ st_ptr(O2, XXX_STATE(_stack_base));
1173
1174 __ sub(O2, wordSize, O2); // prepush
1175 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1176
1177 __ lduh(max_stack, O3); // Full size expression stack
1178 __ sll(O3, LogBytesPerWord, O3);
1179 __ sub(O2, O3, O3);
1180 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1181 __ st_ptr(O3, XXX_STATE(_stack_limit));
1182
1183 if (!native) {
1184 //
1185 // Code to initialize locals
1186 //
1187 Register init_value = noreg; // will be G0 if we must clear locals
1188 // Now zero locals
1189 if (true /* zerolocals */ || ClearInterpreterLocals) {
1190 // explicitly initialize locals
1191 init_value = G0;
1192 } else {
1193 #ifdef ASSERT
1194 // initialize locals to a garbage pattern for better debugging
1195 init_value = O3;
1196 __ set( 0x0F0F0F0F, init_value );
1197 #endif // ASSERT
1198 }
1199 if (init_value != noreg) {
1200 Label clear_loop;
1201
1202 // NOTE: If you change the frame layout, this code will need to
1203 // be updated!
1204 __ lduh( size_of_locals, O2 );
1205 __ lduh( size_of_parameters, O1 );
1206 __ sll( O2, LogBytesPerWord, O2);
1207 __ sll( O1, LogBytesPerWord, O1 );
1208 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1209 __ sub( L2_scratch, O2, O2 );
1210 __ sub( L2_scratch, O1, O1 );
1211
1212 __ bind( clear_loop );
1213 __ inc( O2, wordSize );
1214
1215 __ cmp( O2, O1 );
1216 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1217 __ delayed()->st_ptr( init_value, O2, 0 );
1218 }
1219 }
1220 }
1221 // Find preallocated monitor and lock method (C++ interpreter)
1222 //
1223 void InterpreterGenerator::lock_method(void) {
1224 // Lock the current method.
1225 // Destroys registers L2_scratch, L3_scratch, O0
1226 //
1227 // Find everything relative to Lstate
1228
1229 #ifdef ASSERT
1230 __ ld_ptr(STATE(_method), L2_scratch);
1231 __ ld(L2_scratch, in_bytes(methodOopDesc::access_flags_offset()), O0);
1232
1233 { Label ok;
1234 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1235 __ br( Assembler::notZero, false, Assembler::pt, ok);
1236 __ delayed()->nop();
1237 __ stop("method doesn't need synchronization");
1238 __ bind(ok);
1239 }
1240 #endif // ASSERT
1241
1242 // monitor is already allocated at stack base
1243 // and the lockee is already present
1244 __ ld_ptr(STATE(_stack_base), L2_scratch);
1245 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1246 __ lock_object(L2_scratch, O0);
1247
1248 }
1249
1250 // Generate code for handling resuming a deopted method
1251 void CppInterpreterGenerator::generate_deopt_handling() {
1252
1253 Label return_from_deopt_common;
1254
1255 // deopt needs to jump to here to enter the interpreter (return a result)
1256 deopt_frame_manager_return_atos = __ pc();
1257
1258 // O0/O1 live
1259 __ ba(false, return_from_deopt_common);
1260 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1261
1262
1263 // deopt needs to jump to here to enter the interpreter (return a result)
1264 deopt_frame_manager_return_btos = __ pc();
1265
1266 // O0/O1 live
1267 __ ba(false, return_from_deopt_common);
1268 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1269
1270 // deopt needs to jump to here to enter the interpreter (return a result)
1271 deopt_frame_manager_return_itos = __ pc();
1272
1273 // O0/O1 live
1274 __ ba(false, return_from_deopt_common);
1275 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1276
1277 // deopt needs to jump to here to enter the interpreter (return a result)
1278
1279 deopt_frame_manager_return_ltos = __ pc();
1280 #if !defined(_LP64) && defined(COMPILER2)
1281 // All return values are where we want them, except for Longs. C2 returns
1282 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1283 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1284 // build even if we are returning from interpreted we just do a little
1285 // stupid shuffing.
1286 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1287 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1288 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1289
1290 __ srl (G1, 0,O1);
1291 __ srlx(G1,32,O0);
1292 #endif /* !_LP64 && COMPILER2 */
1293 // O0/O1 live
1294 __ ba(false, return_from_deopt_common);
1295 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1296
1297 // deopt needs to jump to here to enter the interpreter (return a result)
1298
1299 deopt_frame_manager_return_ftos = __ pc();
1300 // O0/O1 live
1301 __ ba(false, return_from_deopt_common);
1302 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1303
1304 // deopt needs to jump to here to enter the interpreter (return a result)
1305 deopt_frame_manager_return_dtos = __ pc();
1306
1307 // O0/O1 live
1308 __ ba(false, return_from_deopt_common);
1309 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1310
1311 // deopt needs to jump to here to enter the interpreter (return a result)
1312 deopt_frame_manager_return_vtos = __ pc();
1313
1314 // O0/O1 live
1315 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1316
1317 // Deopt return common
1318 // an index is present that lets us move any possible result being
1319 // return to the interpreter's stack
1320 //
1321 __ bind(return_from_deopt_common);
1322
1323 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1324 // stack is in the state that the calling convention left it.
1325 // Copy the result from native abi result and place it on java expression stack.
1326
1327 // Current interpreter state is present in Lstate
1328
1329 // Get current pre-pushed top of interpreter stack
1330 // Any result (if any) is in native abi
1331 // result type index is in L3_scratch
1332
1333 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1334
1335 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1336 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1337 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1338 __ jmpl(Lscratch, G0, O7); // and convert it
1339 __ delayed()->nop();
1340
1341 // L1_scratch points to top of stack (prepushed)
1342 __ st_ptr(L1_scratch, STATE(_stack));
1343 }
1344
1345 // Generate the code to handle a more_monitors message from the c++ interpreter
1346 void CppInterpreterGenerator::generate_more_monitors() {
1347
1348 Label entry, loop;
1349 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1350 // 1. compute new pointers // esp: old expression stack top
1351 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1352 __ sub(L4_scratch, entry_size, L4_scratch);
1353 __ st_ptr(L4_scratch, STATE(_stack_base));
1354
1355 __ sub(SP, entry_size, SP); // Grow stack
1356 __ st_ptr(SP, STATE(_frame_bottom));
1357
1358 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1359 __ sub(L2_scratch, entry_size, L2_scratch);
1360 __ st_ptr(L2_scratch, STATE(_stack_limit));
1361
1362 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1363 __ sub(L1_scratch, entry_size, L1_scratch);
1364 __ st_ptr(L1_scratch, STATE(_stack));
1365 __ ba(false, entry);
1366 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1367
1368 // 2. move expression stack
1369
1370 __ bind(loop);
1371 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1372 __ add(L1_scratch, wordSize, L1_scratch);
1373 __ bind(entry);
1374 __ cmp(L1_scratch, L4_scratch);
1375 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1376 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1377
1378 // now zero the slot so we can find it.
1379 __ st(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1380
1381 }
1382
1383 // Initial entry to C++ interpreter from the call_stub.
1384 // This entry point is called the frame manager since it handles the generation
1385 // of interpreter activation frames via requests directly from the vm (via call_stub)
1386 // and via requests from the interpreter. The requests from the call_stub happen
1387 // directly thru the entry point. Requests from the interpreter happen via returning
1388 // from the interpreter and examining the message the interpreter has returned to
1389 // the frame manager. The frame manager can take the following requests:
1390
1391 // NO_REQUEST - error, should never happen.
1392 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1393 // allocate a new monitor.
1394 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1395 // happens during entry during the entry via the call stub.
1396 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1397 //
1398 // Arguments:
1399 //
1400 // ebx: methodOop
1401 // ecx: receiver - unused (retrieved from stack as needed)
1402 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1403 //
1404 //
1405 // Stack layout at entry
1406 //
1407 // [ return address ] <--- esp
1408 // [ parameter n ]
1409 // ...
1410 // [ parameter 1 ]
1411 // [ expression stack ]
1412 //
1413 //
1414 // We are free to blow any registers we like because the call_stub which brought us here
1415 // initially has preserved the callee save registers already.
1416 //
1417 //
1418
1419 static address interpreter_frame_manager = NULL;
1420
1421 #ifdef ASSERT
1422 #define VALIDATE_STATE(scratch, marker) \
1423 { \
1424 Label skip; \
1425 __ ld_ptr(STATE(_self_link), scratch); \
1426 __ cmp(Lstate, scratch); \
1427 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1428 __ delayed()->nop(); \
1429 __ breakpoint_trap(); \
1430 __ emit_long(marker); \
1431 __ bind(skip); \
1432 }
1433 #else
1434 #define VALIDATE_STATE(scratch, marker)
1435 #endif /* ASSERT */
1436
1437 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1438 //
1439 // Adjust caller's stack so that all the locals can be contiguous with
1440 // the parameters.
1441 // Worries about stack overflow make this a pain.
1442 //
1443 // Destroys args, G3_scratch, G3_scratch
1444 // In/Out O5_savedSP (sender's original SP)
1445 //
1446 // assert_different_registers(state, prev_state);
1447 const Register Gtmp = G3_scratch;
1448 const Register tmp = O2;
1449 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1450 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1451
1452 __ lduh(size_of_parameters, tmp);
1453 __ sll(tmp, LogBytesPerWord, Gtmp); // parameter size in bytes
1454 __ add(args, Gtmp, Gargs); // points to first local + BytesPerWord
1455 // NEW
1456 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1457 // determine extra space for non-argument locals & adjust caller's SP
1458 // Gtmp1: parameter size in words
1459 __ lduh(size_of_locals, Gtmp);
1460 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1461
1462 #if 1
1463 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1464 // the call_stub will place the final interpreter argument at
1465 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1466 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1467 // and try to make it look good in the debugger we will store the argument to
1468 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1469 // extra space for the compiler this will overwrite locals in the local array of the
1470 // interpreter.
1471 // QQQ still needed with frameless adapters???
1472
1473 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1474
1475 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1476 #endif // 1
1477
1478
1479 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1480 }
1481
1482 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1483
1484 // G5_method: methodOop
1485 // G2_thread: thread (unused)
1486 // Gargs: bottom of args (sender_sp)
1487 // O5: sender's sp
1488
1489 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1490 // the code is pretty much ready. Would need to change the test below and for good measure
1491 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1492 // routines. Not clear this is worth it yet.
1493
1494 if (interpreter_frame_manager) {
1495 return interpreter_frame_manager;
1496 }
1497
1498 __ bind(frame_manager_entry);
1499
1500 // the following temporary registers are used during frame creation
1501 const Register Gtmp1 = G3_scratch;
1502 // const Register Lmirror = L1; // native mirror (native calls only)
1503
1504 const Address constants (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1505 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1506 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1507 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1508 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1509
1510 address entry_point = __ pc();
1511 __ mov(G0, prevState); // no current activation
1512
1513
1514 Label re_dispatch;
1515
1516 __ bind(re_dispatch);
1517
1518 // Interpreter needs to have locals completely contiguous. In order to do that
1519 // We must adjust the caller's stack pointer for any locals beyond just the
1520 // parameters
1521 adjust_callers_stack(Gargs);
1522
1523 // O5_savedSP still contains sender's sp
1524
1525 // NEW FRAME
1526
1527 generate_compute_interpreter_state(Lstate, prevState, false);
1528
1529 // At this point a new interpreter frame and state object are created and initialized
1530 // Lstate has the pointer to the new activation
1531 // Any stack banging or limit check should already be done.
1532
1533 Label call_interpreter;
1534
1535 __ bind(call_interpreter);
1536
1537
1538 #if 1
1539 __ set(0xdead002, Lmirror);
1540 __ set(0xdead002, L2_scratch);
1541 __ set(0xdead003, L3_scratch);
1542 __ set(0xdead004, L4_scratch);
1543 __ set(0xdead005, Lscratch);
1544 __ set(0xdead006, Lscratch2);
1545 __ set(0xdead007, L7_scratch);
1546
1547 __ set(0xdeaf002, O2);
1548 __ set(0xdeaf003, O3);
1549 __ set(0xdeaf004, O4);
1550 __ set(0xdeaf005, O5);
1551 #endif
1552
1553 // Call interpreter (stack bang complete) enter here if message is
1554 // set and we know stack size is valid
1555
1556 Label call_interpreter_2;
1557
1558 __ bind(call_interpreter_2);
1559
1560 #ifdef ASSERT
1561 {
1562 Label skip;
1563 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1564 __ cmp(G3_scratch, SP);
1565 __ brx(Assembler::equal, false, Assembler::pt, skip);
1566 __ delayed()->nop();
1567 __ stop("SP not restored to frame bottom");
1568 __ bind(skip);
1569 }
1570 #endif
1571
1572 VALIDATE_STATE(G3_scratch, 4);
1573 __ set_last_Java_frame(SP, noreg);
1574 __ mov(Lstate, O0); // (arg) pointer to current state
1575
1576 __ call(CAST_FROM_FN_PTR(address,
1577 JvmtiExport::can_post_interpreter_events() ?
1578 BytecodeInterpreter::runWithChecks
1579 : BytecodeInterpreter::run),
1580 relocInfo::runtime_call_type);
1581
1582 __ delayed()->nop();
1583
1584 __ ld_ptr(STATE(_thread), G2_thread);
1585 __ reset_last_Java_frame();
1586
1587 // examine msg from interpreter to determine next action
1588 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1589
1590 __ ld(STATE(_msg), L1_scratch); // Get new message
1591
1592 Label call_method;
1593 Label return_from_interpreted_method;
1594 Label throw_exception;
1595 Label do_OSR;
1596 Label bad_msg;
1597 Label resume_interpreter;
1598
1599 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1600 __ br(Assembler::equal, false, Assembler::pt, call_method);
1601 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1602 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1603 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1604 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1605 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1606 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1607 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1608 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1609
1610 // Allocate more monitor space, shuffle expression stack....
1611
1612 generate_more_monitors();
1613
1614 // new monitor slot allocated, resume the interpreter.
1615
1616 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1617 VALIDATE_STATE(G3_scratch, 5);
1618 __ ba(false, call_interpreter);
1619 __ delayed()->st(L1_scratch, STATE(_msg));
1620
1621 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1622 unctrap_frame_manager_entry = __ pc();
1623
1624 // QQQ what message do we send
1625
1626 __ ba(false, call_interpreter);
1627 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1628
1629 //=============================================================================
1630 // Returning from a compiled method into a deopted method. The bytecode at the
1631 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1632 // for the template based interpreter). Any stack space that was used by the
1633 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1634 // so all that we have to do is place any pending result on the expression stack
1635 // and resume execution on the next bytecode.
1636
1637 generate_deopt_handling();
1638
1639 // ready to resume the interpreter
1640
1641 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1642 __ ba(false, call_interpreter);
1643 __ delayed()->st(L1_scratch, STATE(_msg));
1644
1645 // Current frame has caught an exception we need to dispatch to the
1646 // handler. We can get here because a native interpreter frame caught
1647 // an exception in which case there is no handler and we must rethrow
1648 // If it is a vanilla interpreted frame the we simply drop into the
1649 // interpreter and let it do the lookup.
1650
1651 Interpreter::_rethrow_exception_entry = __ pc();
1652
1653 Label return_with_exception;
1654 Label unwind_and_forward;
1655
1656 // O0: exception
1657 // O7: throwing pc
1658
1659 // We want exception in the thread no matter what we ultimately decide about frame type.
1660
1661 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1662 __ verify_thread();
1663 __ st_ptr(O0, exception_addr);
1664
1665 // get the methodOop
1666 __ ld_ptr(STATE(_method), G5_method);
1667
1668 // if this current frame vanilla or native?
1669
1670 __ ld(access_flags, Gtmp1);
1671 __ btst(JVM_ACC_NATIVE, Gtmp1);
1672 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1673 __ delayed()->nop();
1674
1675 // We drop thru to unwind a native interpreted frame with a pending exception
1676 // We jump here for the initial interpreter frame with exception pending
1677 // We unwind the current acivation and forward it to our caller.
1678
1679 __ bind(unwind_and_forward);
1680
1681 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1682 // as expected by forward_exception.
1683
1684 __ restore(FP, G0, SP); // unwind interpreter state frame
1685 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1686 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1687
1688 // Return point from a call which returns a result in the native abi
1689 // (c1/c2/jni-native). This result must be processed onto the java
1690 // expression stack.
1691 //
1692 // A pending exception may be present in which case there is no result present
1693
1694 address return_from_native_method = __ pc();
1695
1696 VALIDATE_STATE(G3_scratch, 6);
1697
1698 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1699 // stack is in the state that the calling convention left it.
1700 // Copy the result from native abi result and place it on java expression stack.
1701
1702 // Current interpreter state is present in Lstate
1703
1704 // Exception pending?
1705
1706 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1707 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1708 __ tst(Lscratch); // exception pending?
1709 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1710 __ delayed()->nop();
1711
1712 // Process the native abi result to java expression stack
1713
1714 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1715 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1716 __ lduh(L4_scratch, in_bytes(methodOopDesc::size_of_parameters_offset()), L2_scratch); // get parameter size
1717 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1718 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1719 __ ld_ptr(L4_scratch, in_bytes(methodOopDesc::result_index_offset()), L3_scratch); // called method result type index
1720
1721 // tosca is really just native abi
1722 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1723 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1724 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1725 __ jmpl(Lscratch, G0, O7); // and convert it
1726 __ delayed()->nop();
1727
1728 // L1_scratch points to top of stack (prepushed)
1729
1730 __ ba(false, resume_interpreter);
1731 __ delayed()->mov(L1_scratch, O1);
1732
1733 // An exception is being caught on return to a vanilla interpreter frame.
1734 // Empty the stack and resume interpreter
1735
1736 __ bind(return_with_exception);
1737
1738 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1739 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1740 __ ba(false, resume_interpreter);
1741 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1742
1743 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1744 // interpreter call, or native) and unwind this interpreter activation.
1745 // All monitors should be unlocked.
1746
1747 __ bind(return_from_interpreted_method);
1748
1749 VALIDATE_STATE(G3_scratch, 7);
1750
1751 Label return_to_initial_caller;
1752
1753 // Interpreted result is on the top of the completed activation expression stack.
1754 // We must return it to the top of the callers stack if caller was interpreted
1755 // otherwise we convert to native abi result and return to call_stub/c1/c2
1756 // The caller's expression stack was truncated by the call however the current activation
1757 // has enough stuff on the stack that we have usable space there no matter what. The
1758 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1759 // for the current activation
1760
1761 __ ld_ptr(STATE(_prev_link), L1_scratch);
1762 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1763 __ ld_ptr(L2_scratch, in_bytes(methodOopDesc::result_index_offset()), L2_scratch);
1764 __ tst(L1_scratch);
1765 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1766 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1767
1768 // Copy result to callers java stack
1769
1770 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1771 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1772 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1773 __ ld_ptr(STATE(_locals), O1); // stack destination
1774
1775 // O0 - will be source, O1 - will be destination (preserved)
1776 __ jmpl(Lscratch, G0, O7); // and convert it
1777 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1778
1779 // O1 == &locals[0]
1780
1781 // Result is now on caller's stack. Just unwind current activation and resume
1782
1783 Label unwind_recursive_activation;
1784
1785
1786 __ bind(unwind_recursive_activation);
1787
1788 // O1 == &locals[0] (really callers stacktop) for activation now returning
1789 // returning to interpreter method from "recursive" interpreter call
1790 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1791 // to. Now all we must do is unwind the state from the completed call
1792
1793 // Must restore stack
1794 VALIDATE_STATE(G3_scratch, 8);
1795
1796 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1797 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1798 // frame and tell interpreter to resume.
1799
1800
1801 __ mov(O1, I1); // pass back new stack top across activation
1802 // POP FRAME HERE ==================================
1803 __ restore(FP, G0, SP); // unwind interpreter state frame
1804 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1805
1806
1807 // Resume the interpreter. The current frame contains the current interpreter
1808 // state object.
1809 //
1810 // O1 == new java stack pointer
1811
1812 __ bind(resume_interpreter);
1813 VALIDATE_STATE(G3_scratch, 10);
1814
1815 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1816
1817 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1818 __ st(L1_scratch, STATE(_msg));
1819 __ ba(false, call_interpreter_2);
1820 __ delayed()->st_ptr(O1, STATE(_stack));
1821
1822
1823 // Fast accessor methods share this entry point.
1824 // This works because frame manager is in the same codelet
1825 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1826 // we need to do a little register fixup here once we distinguish the two of them
1827 if (UseFastAccessorMethods && !synchronized) {
1828 // Call stub_return address still in O7
1829 __ bind(fast_accessor_slow_entry_path);
1830 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1831 __ cmp(Gtmp1, O7); // returning to interpreter?
1832 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1833 __ delayed()->nop();
1834 __ ba(false, re_dispatch);
1835 __ delayed()->mov(G0, prevState); // initial entry
1836
1837 }
1838
1839 // interpreter returning to native code (call_stub/c1/c2)
1840 // convert result and unwind initial activation
1841 // L2_scratch - scaled result type index
1842
1843 __ bind(return_to_initial_caller);
1844
1845 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1846 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1847 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1848 __ jmpl(Lscratch, G0, O7); // and convert it
1849 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1850
1851 Label unwind_initial_activation;
1852 __ bind(unwind_initial_activation);
1853
1854 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1855 // we can return here with an exception that wasn't handled by interpreted code
1856 // how does c1/c2 see it on return?
1857
1858 // compute resulting sp before/after args popped depending upon calling convention
1859 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1860 //
1861 // POP FRAME HERE ==================================
1862 __ restore(FP, G0, SP);
1863 __ retl();
1864 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1865
1866 // OSR request, unwind the current frame and transfer to the OSR entry
1867 // and enter OSR nmethod
1868
1869 __ bind(do_OSR);
1870 Label remove_initial_frame;
1871 __ ld_ptr(STATE(_prev_link), L1_scratch);
1872 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1873
1874 // We are going to pop this frame. Is there another interpreter frame underneath
1875 // it or is it callstub/compiled?
1876
1877 __ tst(L1_scratch);
1878 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1879 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1880
1881 // Frame underneath is an interpreter frame simply unwind
1882 // POP FRAME HERE ==================================
1883 __ restore(FP, G0, SP); // unwind interpreter state frame
1884 __ mov(I5_savedSP->after_restore(), SP);
1885
1886 // Since we are now calling native need to change our "return address" from the
1887 // dummy RecursiveInterpreterActivation to a return from native
1888
1889 __ set((intptr_t)return_from_native_method - 8, O7);
1890
1891 __ jmpl(G3_scratch, G0, G0);
1892 __ delayed()->mov(G1_scratch, O0);
1893
1894 __ bind(remove_initial_frame);
1895
1896 // POP FRAME HERE ==================================
1897 __ restore(FP, G0, SP);
1898 __ mov(I5_savedSP->after_restore(), SP);
1899 __ jmpl(G3_scratch, G0, G0);
1900 __ delayed()->mov(G1_scratch, O0);
1901
1902 // Call a new method. All we do is (temporarily) trim the expression stack
1903 // push a return address to bring us back to here and leap to the new entry.
1904 // At this point we have a topmost frame that was allocated by the frame manager
1905 // which contains the current method interpreted state. We trim this frame
1906 // of excess java expression stack entries and then recurse.
1907
1908 __ bind(call_method);
1909
1910 // stack points to next free location and not top element on expression stack
1911 // method expects sp to be pointing to topmost element
1912
1913 __ ld_ptr(STATE(_thread), G2_thread);
1914 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1915
1916
1917 // SP already takes in to account the 2 extra words we use for slop
1918 // when we call a "static long no_params()" method. So if
1919 // we trim back sp by the amount of unused java expression stack
1920 // there will be automagically the 2 extra words we need.
1921 // We also have to worry about keeping SP aligned.
1922
1923 __ ld_ptr(STATE(_stack), Gargs);
1924 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1925
1926 // compute the unused java stack size
1927 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1928
1929 // Round down the unused space to that stack is always aligned
1930 // by making the unused space a multiple of the size of a long.
1931
1932 __ and3(L2_scratch, -BytesPerLong, L2_scratch);
1933
1934 // Now trim the stack
1935 __ add(SP, L2_scratch, SP);
1936
1937
1938 // Now point to the final argument (account for prepush)
1939 __ add(Gargs, wordSize, Gargs);
1940 #ifdef ASSERT
1941 // Make sure we have space for the window
1942 __ sub(Gargs, SP, L1_scratch);
1943 __ cmp(L1_scratch, 16*wordSize);
1944 {
1945 Label skip;
1946 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1947 __ delayed()->nop();
1948 __ stop("killed stack");
1949 __ bind(skip);
1950 }
1951 #endif // ASSERT
1952
1953 // Create a new frame where we can store values that make it look like the interpreter
1954 // really recursed.
1955
1956 // prepare to recurse or call specialized entry
1957
1958 // First link the registers we need
1959
1960 // make the pc look good in debugger
1961 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
1962 // argument too
1963 __ mov(Lstate, I0);
1964
1965 // Record our sending SP
1966 __ mov(SP, O5_savedSP);
1967
1968 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
1969 __ set((intptr_t) entry_point, L1_scratch);
1970 __ cmp(L1_scratch, L2_scratch);
1971 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
1972 __ delayed()->mov(Lstate, prevState); // link activations
1973
1974 // method uses specialized entry, push a return so we look like call stub setup
1975 // this path will handle fact that result is returned in registers and not
1976 // on the java stack.
1977
1978 __ set((intptr_t)return_from_native_method - 8, O7);
1979 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
1980 __ delayed()->nop();
1981
1982 //
1983 // Bad Message from interpreter
1984 //
1985 __ bind(bad_msg);
1986 __ stop("Bad message from interpreter");
1987
1988 // Interpreted method "returned" with an exception pass it on...
1989 // Pass result, unwind activation and continue/return to interpreter/call_stub
1990 // We handle result (if any) differently based on return to interpreter or call_stub
1991
1992 __ bind(throw_exception);
1993 __ ld_ptr(STATE(_prev_link), L1_scratch);
1994 __ tst(L1_scratch);
1995 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
1996 __ delayed()->nop();
1997
1998 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
1999 __ ba(false, unwind_recursive_activation);
2000 __ delayed()->nop();
2001
2002 interpreter_frame_manager = entry_point;
2003 return entry_point;
2004 }
2005
2006 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2007 : CppInterpreterGenerator(code) {
2008 generate_all(); // down here so it can be "virtual"
2009 }
2010
2011
2012 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2013
2014 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2015 // expression stack, the callee will have callee_extra_locals (so we can account for
2016 // frame extension) and monitor_size for monitors. Basically we need to calculate
2017 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2018 //
2019 //
2020 // The big complicating thing here is that we must ensure that the stack stays properly
2021 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2022 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2023 // the caller so we must ensure that it is properly aligned for our callee.
2024 //
2025 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2026 // stack at all times to deal with the "stack long no_params()" method issue. This
2027 // is "slop_factor" here.
2028 const int slop_factor = 2;
2029
2030 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2031 frame::memory_parameter_word_sp_offset; // register save area + param window
2032 return (round_to(max_stack +
2033 slop_factor +
2034 fixed_size +
2035 monitor_size +
2036 (callee_extra_locals * Interpreter::stackElementWords()), WordsPerLong));
2037
2038 }
2039
2040 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
2041
2042 // See call_stub code
2043 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2044 WordsPerLong); // 7 + register save area
2045
2046 // Save space for one monitor to get into the interpreted method in case
2047 // the method is synchronized
2048 int monitor_size = method->is_synchronized() ?
2049 1*frame::interpreter_frame_monitor_size() : 0;
2050 return size_activation_helper(method->max_locals(), method->max_stack(),
2051 monitor_size) + call_stub_size;
2052 }
2053
2054 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2055 frame* caller,
2056 frame* current,
2057 methodOop method,
2058 intptr_t* locals,
2059 intptr_t* stack,
2060 intptr_t* stack_base,
2061 intptr_t* monitor_base,
2062 intptr_t* frame_bottom,
2063 bool is_top_frame
2064 )
2065 {
2066 // What about any vtable?
2067 //
2068 to_fill->_thread = JavaThread::current();
2069 // This gets filled in later but make it something recognizable for now
2070 to_fill->_bcp = method->code_base();
2071 to_fill->_locals = locals;
2072 to_fill->_constants = method->constants()->cache();
2073 to_fill->_method = method;
2074 to_fill->_mdx = NULL;
2075 to_fill->_stack = stack;
2076 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2077 to_fill->_msg = deopt_resume2;
2078 } else {
2079 to_fill->_msg = method_resume;
2080 }
2081 to_fill->_result._to_call._bcp_advance = 0;
2082 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2083 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2084 to_fill->_prev_link = NULL;
2085
2086 // Fill in the registers for the frame
2087
2088 // Need to install _sender_sp. Actually not too hard in C++!
2089 // When the skeletal frames are layed out we fill in a value
2090 // for _sender_sp. That value is only correct for the oldest
2091 // skeletal frame constructed (because there is only a single
2092 // entry for "caller_adjustment". While the skeletal frames
2093 // exist that is good enough. We correct that calculation
2094 // here and get all the frames correct.
2095
2096 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2097
2098 *current->register_addr(Lstate) = (intptr_t) to_fill;
2099 // skeletal already places a useful value here and this doesn't account
2100 // for alignment so don't bother.
2101 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2102
2103 if (caller->is_interpreted_frame()) {
2104 interpreterState prev = caller->get_interpreterState();
2105 to_fill->_prev_link = prev;
2106 // Make the prev callee look proper
2107 prev->_result._to_call._callee = method;
2108 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2109 prev->_result._to_call._bcp_advance = 5;
2110 } else {
2111 prev->_result._to_call._bcp_advance = 3;
2112 }
2113 }
2114 to_fill->_oop_temp = NULL;
2115 to_fill->_stack_base = stack_base;
2116 // Need +1 here because stack_base points to the word just above the first expr stack entry
2117 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2118 // See generate_compute_interpreter_state.
2119 to_fill->_stack_limit = stack_base - (method->max_stack() + 1);
2120 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2121
2122 // sparc specific
2123 to_fill->_frame_bottom = frame_bottom;
2124 to_fill->_self_link = to_fill;
2125 #ifdef ASSERT
2126 to_fill->_native_fresult = 123456.789;
2127 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2128 #endif
2129 }
2130
2131 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2132 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2133 }
2134
2135
2136 int AbstractInterpreter::layout_activation(methodOop method,
2137 int tempcount, // Number of slots on java expression stack in use
2138 int popframe_extra_args,
2139 int moncount, // Number of active monitors
2140 int callee_param_size,
2141 int callee_locals_size,
2142 frame* caller,
2143 frame* interpreter_frame,
2144 bool is_top_frame) {
2145
2146 assert(popframe_extra_args == 0, "NEED TO FIX");
2147 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2148 // does as far as allocating an interpreter frame.
2149 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2150 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2151 // as determined by a previous call to this method.
2152 // It is also guaranteed to be walkable even though it is in a skeletal state
2153 // NOTE: return size is in words not bytes
2154 // NOTE: tempcount is the current size of the java expression stack. For top most
2155 // frames we will allocate a full sized expression stack and not the curback
2156 // version that non-top frames have.
2157
2158 // Calculate the amount our frame will be adjust by the callee. For top frame
2159 // this is zero.
2160
2161 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2162 // calculates the extra locals based on itself. Not what the callee does
2163 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2164 // as getting sender_sp correct.
2165
2166 int extra_locals_size = callee_locals_size - callee_param_size;
2167 int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2168 int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2169 int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2170 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2171
2172
2173 /*
2174 if we actually have a frame to layout we must now fill in all the pieces. This means both
2175 the interpreterState and the registers.
2176 */
2177 if (interpreter_frame != NULL) {
2178
2179 // MUCHO HACK
2180
2181 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2182
2183 /* Now fillin the interpreterState object */
2184
2185 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2186
2187
2188 intptr_t* locals;
2189
2190 // Calculate the postion of locals[0]. This is painful because of
2191 // stack alignment (same as ia64). The problem is that we can
2192 // not compute the location of locals from fp(). fp() will account
2193 // for the extra locals but it also accounts for aligning the stack
2194 // and we can't determine if the locals[0] was misaligned but max_locals
2195 // was enough to have the
2196 // calculate postion of locals. fp already accounts for extra locals.
2197 // +2 for the static long no_params() issue.
2198
2199 if (caller->is_interpreted_frame()) {
2200 // locals must agree with the caller because it will be used to set the
2201 // caller's tos when we return.
2202 interpreterState prev = caller->get_interpreterState();
2203 // stack() is prepushed.
2204 locals = prev->stack() + method->size_of_parameters();
2205 } else {
2206 // Lay out locals block in the caller adjacent to the register window save area.
2207 //
2208 // Compiled frames do not allocate a varargs area which is why this if
2209 // statement is needed.
2210 //
2211 intptr_t* fp = interpreter_frame->fp();
2212 int local_words = method->max_locals() * Interpreter::stackElementWords();
2213
2214 if (caller->is_compiled_frame()) {
2215 locals = fp + frame::register_save_words + local_words - 1;
2216 } else {
2217 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2218 }
2219
2220 }
2221 // END MUCHO HACK
2222
2223 intptr_t* monitor_base = (intptr_t*) cur_state;
2224 intptr_t* stack_base = monitor_base - monitor_size;
2225 /* +1 because stack is always prepushed */
2226 intptr_t* stack = stack_base - (tempcount + 1);
2227
2228
2229 BytecodeInterpreter::layout_interpreterState(cur_state,
2230 caller,
2231 interpreter_frame,
2232 method,
2233 locals,
2234 stack,
2235 stack_base,
2236 monitor_base,
2237 frame_bottom,
2238 is_top_frame);
2239
2240 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2241
2242 }
2243 return frame_words;
2244 }
2245
2246 #endif // CC_INTERP
2247