1 /*
2 * Copyright (c) 2007, 2012, 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 #ifdef SHARK
49 #include "shark/shark_globals.hpp"
50 #endif
51
52 #ifdef CC_INTERP
53
54 // Routine exists to make tracebacks look decent in debugger
55 // while "shadow" interpreter frames are on stack. It is also
56 // used to distinguish interpreter frames.
57
58 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
59 ShouldNotReachHere();
60 }
61
62 bool CppInterpreter::contains(address pc) {
63 return ( _code->contains(pc) ||
64 ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
65 }
66
67 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
68 #define __ _masm->
69
70 Label frame_manager_entry;
71 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized
72 // c++ interpreter entry point this holds that entry point label.
73
74 static address unctrap_frame_manager_entry = NULL;
75
76 static address interpreter_return_address = NULL;
77 static address deopt_frame_manager_return_atos = NULL;
78 static address deopt_frame_manager_return_btos = NULL;
79 static address deopt_frame_manager_return_itos = NULL;
80 static address deopt_frame_manager_return_ltos = NULL;
81 static address deopt_frame_manager_return_ftos = NULL;
82 static address deopt_frame_manager_return_dtos = NULL;
83 static address deopt_frame_manager_return_vtos = NULL;
84
85 const Register prevState = G1_scratch;
86
87 void InterpreterGenerator::save_native_result(void) {
88 // result potentially in O0/O1: save it across calls
89 __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
90 #ifdef _LP64
91 __ stx(O0, STATE(_native_lresult));
92 #else
93 __ std(O0, STATE(_native_lresult));
94 #endif
95 }
96
97 void InterpreterGenerator::restore_native_result(void) {
98
99 // Restore any method result value
100 __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
101 #ifdef _LP64
102 __ ldx(STATE(_native_lresult), O0);
103 #else
104 __ ldd(STATE(_native_lresult), O0);
105 #endif
106 }
107
108 // A result handler converts/unboxes a native call result into
109 // a java interpreter/compiler result. The current frame is an
110 // interpreter frame. The activation frame unwind code must be
111 // consistent with that of TemplateTable::_return(...). In the
112 // case of native methods, the caller's SP was not modified.
113 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
114 address entry = __ pc();
115 Register Itos_i = Otos_i ->after_save();
116 Register Itos_l = Otos_l ->after_save();
117 Register Itos_l1 = Otos_l1->after_save();
118 Register Itos_l2 = Otos_l2->after_save();
119 switch (type) {
120 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
121 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
122 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
123 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
124 case T_LONG :
125 #ifndef _LP64
126 __ mov(O1, Itos_l2); // move other half of long
127 #endif // ifdef or no ifdef, fall through to the T_INT case
128 case T_INT : __ mov(O0, Itos_i); break;
129 case T_VOID : /* nothing to do */ break;
130 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
131 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
132 case T_OBJECT :
133 __ ld_ptr(STATE(_oop_temp), Itos_i);
134 __ verify_oop(Itos_i);
135 break;
136 default : ShouldNotReachHere();
137 }
138 __ ret(); // return from interpreter activation
139 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
140 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
141 return entry;
142 }
143
144 // tosca based result to c++ interpreter stack based result.
145 // Result goes to address in L1_scratch
146
147 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
148 // A result is in the native abi result register from a native method call.
149 // We need to return this result to the interpreter by pushing the result on the interpreter's
150 // stack. This is relatively simple the destination is in L1_scratch
151 // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
152 // adjust L1_scratch
153 address entry = __ pc();
154 switch (type) {
155 case T_BOOLEAN:
156 // !0 => true; 0 => false
157 __ subcc(G0, O0, G0);
158 __ addc(G0, 0, O0);
159 __ st(O0, L1_scratch, 0);
160 __ sub(L1_scratch, wordSize, L1_scratch);
161 break;
162
163 // cannot use and3, 0xFFFF too big as immediate value!
164 case T_CHAR :
165 __ sll(O0, 16, O0);
166 __ srl(O0, 16, O0);
167 __ st(O0, L1_scratch, 0);
168 __ sub(L1_scratch, wordSize, L1_scratch);
169 break;
170
171 case T_BYTE :
172 __ sll(O0, 24, O0);
173 __ sra(O0, 24, O0);
174 __ st(O0, L1_scratch, 0);
175 __ sub(L1_scratch, wordSize, L1_scratch);
176 break;
177
178 case T_SHORT :
179 __ sll(O0, 16, O0);
180 __ sra(O0, 16, O0);
181 __ st(O0, L1_scratch, 0);
182 __ sub(L1_scratch, wordSize, L1_scratch);
183 break;
184 case T_LONG :
185 #ifndef _LP64
186 #if defined(COMPILER2)
187 // All return values are where we want them, except for Longs. C2 returns
188 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
189 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
190 // build even if we are returning from interpreted we just do a little
191 // stupid shuffing.
192 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
193 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
194 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
195 __ stx(G1, L1_scratch, -wordSize);
196 #else
197 // native result is in O0, O1
198 __ st(O1, L1_scratch, 0); // Low order
199 __ st(O0, L1_scratch, -wordSize); // High order
200 #endif /* COMPILER2 */
201 #else
202 __ stx(O0, L1_scratch, -wordSize);
203 #endif
204 __ sub(L1_scratch, 2*wordSize, L1_scratch);
205 break;
206
207 case T_INT :
208 __ st(O0, L1_scratch, 0);
209 __ sub(L1_scratch, wordSize, L1_scratch);
210 break;
211
212 case T_VOID : /* nothing to do */
213 break;
214
215 case T_FLOAT :
216 __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
217 __ sub(L1_scratch, wordSize, L1_scratch);
218 break;
219
220 case T_DOUBLE :
221 // Every stack slot is aligned on 64 bit, However is this
222 // the correct stack slot on 64bit?? QQQ
223 __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
224 __ sub(L1_scratch, 2*wordSize, L1_scratch);
225 break;
226 case T_OBJECT :
227 __ verify_oop(O0);
228 __ st_ptr(O0, L1_scratch, 0);
229 __ sub(L1_scratch, wordSize, L1_scratch);
230 break;
231 default : ShouldNotReachHere();
232 }
233 __ retl(); // return from interpreter activation
234 __ delayed()->nop(); // schedule this better
235 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
236 return entry;
237 }
238
239 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
240 // A result is in the java expression stack of the interpreted method that has just
241 // returned. Place this result on the java expression stack of the caller.
242 //
243 // The current interpreter activation in Lstate is for the method just returning its
244 // result. So we know that the result of this method is on the top of the current
245 // execution stack (which is pre-pushed) and will be return to the top of the caller
246 // stack. The top of the callers stack is the bottom of the locals of the current
247 // activation.
248 // Because of the way activation are managed by the frame manager the value of esp is
249 // below both the stack top of the current activation and naturally the stack top
250 // of the calling activation. This enable this routine to leave the return address
251 // to the frame manager on the stack and do a vanilla return.
252 //
253 // On entry: O0 - points to source (callee stack top)
254 // O1 - points to destination (caller stack top [i.e. free location])
255 // destroys O2, O3
256 //
257
258 address entry = __ pc();
259 switch (type) {
260 case T_VOID: break;
261 break;
262 case T_FLOAT :
263 case T_BOOLEAN:
264 case T_CHAR :
265 case T_BYTE :
266 case T_SHORT :
267 case T_INT :
268 // 1 word result
269 __ ld(O0, 0, O2);
270 __ st(O2, O1, 0);
271 __ sub(O1, wordSize, O1);
272 break;
273 case T_DOUBLE :
274 case T_LONG :
275 // return top two words on current expression stack to caller's expression stack
276 // The caller's expression stack is adjacent to the current frame manager's intepretState
277 // except we allocated one extra word for this intepretState so we won't overwrite it
278 // when we return a two word result.
279 #ifdef _LP64
280 __ ld_ptr(O0, 0, O2);
281 __ st_ptr(O2, O1, -wordSize);
282 #else
283 __ ld(O0, 0, O2);
284 __ ld(O0, wordSize, O3);
285 __ st(O3, O1, 0);
286 __ st(O2, O1, -wordSize);
287 #endif
288 __ sub(O1, 2*wordSize, O1);
289 break;
290 case T_OBJECT :
291 __ ld_ptr(O0, 0, O2);
292 __ verify_oop(O2); // verify it
293 __ st_ptr(O2, O1, 0);
294 __ sub(O1, wordSize, O1);
295 break;
296 default : ShouldNotReachHere();
297 }
298 __ retl();
299 __ delayed()->nop(); // QQ schedule this better
300 return entry;
301 }
302
303 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
304 // A result is in the java expression stack of the interpreted method that has just
305 // returned. Place this result in the native abi that the caller expects.
306 // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
307 //
308 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
309 // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
310 // and so rather than return result onto caller's java expression stack we return the
311 // result in the expected location based on the native abi.
312 // On entry: O0 - source (stack top)
313 // On exit result in expected output register
314 // QQQ schedule this better
315
316 address entry = __ pc();
317 switch (type) {
318 case T_VOID: break;
319 break;
320 case T_FLOAT :
321 __ ldf(FloatRegisterImpl::S, O0, 0, F0);
322 break;
323 case T_BOOLEAN:
324 case T_CHAR :
325 case T_BYTE :
326 case T_SHORT :
327 case T_INT :
328 // 1 word result
329 __ ld(O0, 0, O0->after_save());
330 break;
331 case T_DOUBLE :
332 __ ldf(FloatRegisterImpl::D, O0, 0, F0);
333 break;
334 case T_LONG :
335 // return top two words on current expression stack to caller's expression stack
336 // The caller's expression stack is adjacent to the current frame manager's interpretState
337 // except we allocated one extra word for this intepretState so we won't overwrite it
338 // when we return a two word result.
339 #ifdef _LP64
340 __ ld_ptr(O0, 0, O0->after_save());
341 #else
342 __ ld(O0, wordSize, O1->after_save());
343 __ ld(O0, 0, O0->after_save());
344 #endif
345 #if defined(COMPILER2) && !defined(_LP64)
346 // C2 expects long results in G1 we can't tell if we're returning to interpreted
347 // or compiled so just be safe use G1 and O0/O1
348
349 // Shift bits into high (msb) of G1
350 __ sllx(Otos_l1->after_save(), 32, G1);
351 // Zero extend low bits
352 __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
353 __ or3 (Otos_l2->after_save(), G1, G1);
354 #endif /* COMPILER2 */
355 break;
356 case T_OBJECT :
357 __ ld_ptr(O0, 0, O0->after_save());
358 __ verify_oop(O0->after_save()); // verify it
359 break;
360 default : ShouldNotReachHere();
361 }
362 __ retl();
363 __ delayed()->nop();
364 return entry;
365 }
366
367 address CppInterpreter::return_entry(TosState state, int length) {
368 // make it look good in the debugger
369 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
370 }
371
372 address CppInterpreter::deopt_entry(TosState state, int length) {
373 address ret = NULL;
374 if (length != 0) {
375 switch (state) {
376 case atos: ret = deopt_frame_manager_return_atos; break;
377 case btos: ret = deopt_frame_manager_return_btos; break;
378 case ctos:
379 case stos:
380 case itos: ret = deopt_frame_manager_return_itos; break;
381 case ltos: ret = deopt_frame_manager_return_ltos; break;
382 case ftos: ret = deopt_frame_manager_return_ftos; break;
383 case dtos: ret = deopt_frame_manager_return_dtos; break;
384 case vtos: ret = deopt_frame_manager_return_vtos; break;
385 }
386 } else {
387 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap)
388 }
389 assert(ret != NULL, "Not initialized");
390 return ret;
391 }
392
393 //
394 // Helpers for commoning out cases in the various type of method entries.
395 //
396
397 // increment invocation count & check for overflow
398 //
399 // Note: checking for negative value instead of overflow
400 // so we have a 'sticky' overflow test
401 //
402 // Lmethod: method
403 // ??: invocation counter
404 //
405 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
406 // Update standard invocation counters
407 __ increment_invocation_counter(O0, G3_scratch);
408 if (ProfileInterpreter) { // %%% Merge this into MethodData*
409 __ ld_ptr(STATE(_method), G3_scratch);
410 Address interpreter_invocation_counter(G3_scratch, 0, in_bytes(Method::interpreter_invocation_counter_offset()));
411 __ ld(interpreter_invocation_counter, G3_scratch);
412 __ inc(G3_scratch);
413 __ st(G3_scratch, interpreter_invocation_counter);
414 }
415
416 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
417 __ sethi(invocation_limit);
418 __ ld(invocation_limit, G3_scratch);
419 __ cmp(O0, G3_scratch);
420 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
421 __ delayed()->nop();
422
423 }
424
425 address InterpreterGenerator::generate_empty_entry(void) {
426
427 // A method that does nothing but return...
428
429 address entry = __ pc();
430 Label slow_path;
431
432 // do nothing for empty methods (do not even increment invocation counter)
433 if ( UseFastEmptyMethods) {
434 // If we need a safepoint check, generate full interpreter entry.
435 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
436 __ load_contents(sync_state, G3_scratch);
437 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
438 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
439 __ delayed()->nop();
440
441 // Code: _return
442 __ retl();
443 __ delayed()->mov(O5_savedSP, SP);
444 return entry;
445 }
446 return NULL;
447 }
448
449 // Call an accessor method (assuming it is resolved, otherwise drop into
450 // vanilla (slow path) entry
451
452 // Generates code to elide accessor methods
453 // Uses G3_scratch and G1_scratch as scratch
454 address InterpreterGenerator::generate_accessor_entry(void) {
455
456 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
457 // parameter size = 1
458 // Note: We can only use this code if the getfield has been resolved
459 // and if we don't have a null-pointer exception => check for
460 // these conditions first and use slow path if necessary.
461 address entry = __ pc();
462 Label slow_path;
463
464 if ( UseFastAccessorMethods) {
465 // Check if we need to reach a safepoint and generate full interpreter
466 // frame if so.
467 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
468 __ load_contents(sync_state, G3_scratch);
469 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
470 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
471 __ delayed()->nop();
472
473 // Check if local 0 != NULL
474 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
475 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
476 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
477 __ delayed()->nop();
478
479
480 // read first instruction word and extract bytecode @ 1 and index @ 2
481 // get first 4 bytes of the bytecodes (big endian!)
482 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::const_offset())), G1_scratch);
483 __ ld(Address(G1_scratch, 0, in_bytes(ConstMethod::codes_offset())), G1_scratch);
484
485 // move index @ 2 far left then to the right most two bytes.
486 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
487 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
488 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
489
490 // get constant pool cache
491 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), G3_scratch);
492 __ ld_ptr(G3_scratch, in_bytes(ConstMethod::constants_offset()), G3_scratch);
493 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
494
495 // get specific constant pool cache entry
496 __ add(G3_scratch, G1_scratch, G3_scratch);
497
498 // Check the constant Pool cache entry to see if it has been resolved.
499 // If not, need the slow path.
500 ByteSize cp_base_offset = ConstantPoolCache::base_offset();
501 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
502 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
503 __ and3(G1_scratch, 0xFF, G1_scratch);
504 __ cmp(G1_scratch, Bytecodes::_getfield);
505 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
506 __ delayed()->nop();
507
508 // Get the type and return field offset from the constant pool cache
509 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
510 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
511
512 Label xreturn_path;
513 // Need to differentiate between igetfield, agetfield, bgetfield etc.
514 // because they are different sizes.
515 // Get the type from the constant pool cache
516 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
517 // Make sure we don't need to mask G1_scratch after the above shift
518 ConstantPoolCacheEntry::verify_tos_state_shift();
519 __ cmp(G1_scratch, atos );
520 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
521 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
522 __ cmp(G1_scratch, itos);
523 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
524 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
525 __ cmp(G1_scratch, stos);
526 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
527 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
528 __ cmp(G1_scratch, ctos);
529 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
530 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
531 #ifdef ASSERT
532 __ cmp(G1_scratch, btos);
533 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
534 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
535 __ should_not_reach_here();
536 #endif
537 __ ldsb(Otos_i, G3_scratch, Otos_i);
538 __ bind(xreturn_path);
539
540 // _ireturn/_areturn
541 __ retl(); // return from leaf routine
542 __ delayed()->mov(O5_savedSP, SP);
543
544 // Generate regular method entry
545 __ bind(slow_path);
546 __ ba(fast_accessor_slow_entry_path);
547 __ delayed()->nop();
548 return entry;
549 }
550 return NULL;
551 }
552
553 address InterpreterGenerator::generate_Reference_get_entry(void) {
554 #ifndef SERIALGC
555 if (UseG1GC) {
556 // We need to generate have a routine that generates code to:
557 // * load the value in the referent field
558 // * passes that value to the pre-barrier.
559 //
560 // In the case of G1 this will record the value of the
561 // referent in an SATB buffer if marking is active.
562 // This will cause concurrent marking to mark the referent
563 // field as live.
564 Unimplemented();
565 }
566 #endif // SERIALGC
567
568 // If G1 is not enabled then attempt to go through the accessor entry point
569 // Reference.get is an accessor
570 return generate_accessor_entry();
571 }
572
573 //
574 // Interpreter stub for calling a native method. (C++ interpreter)
575 // This sets up a somewhat different looking stack for calling the native method
576 // than the typical interpreter frame setup.
577 //
578
579 address InterpreterGenerator::generate_native_entry(bool synchronized) {
580 address entry = __ pc();
581
582 // the following temporary registers are used during frame creation
583 const Register Gtmp1 = G3_scratch ;
584 const Register Gtmp2 = G1_scratch;
585 const Address size_of_parameters(G5_method, 0, in_bytes(Method::size_of_parameters_offset()));
586
587 bool inc_counter = UseCompiler || CountCompiledCalls;
588
589 // make sure registers are different!
590 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
591
592 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
593
594 Label Lentry;
595 __ bind(Lentry);
596
597 const Register Glocals_size = G3;
598 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
599
600 // make sure method is native & not abstract
601 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
602 #ifdef ASSERT
603 __ ld(access_flags, Gtmp1);
604 {
605 Label L;
606 __ btst(JVM_ACC_NATIVE, Gtmp1);
607 __ br(Assembler::notZero, false, Assembler::pt, L);
608 __ delayed()->nop();
609 __ stop("tried to execute non-native method as native");
610 __ bind(L);
611 }
612 { Label L;
613 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
614 __ br(Assembler::zero, false, Assembler::pt, L);
615 __ delayed()->nop();
616 __ stop("tried to execute abstract method as non-abstract");
617 __ bind(L);
618 }
619 #endif // ASSERT
620
621 __ lduh(size_of_parameters, Gtmp1);
622 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes
623 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord
624 // NEW
625 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
626 // generate the code to allocate the interpreter stack frame
627 // NEW FRAME ALLOCATED HERE
628 // save callers original sp
629 // __ mov(SP, I5_savedSP->after_restore());
630
631 generate_compute_interpreter_state(Lstate, G0, true);
632
633 // At this point Lstate points to new interpreter state
634 //
635
636 const Address do_not_unlock_if_synchronized(G2_thread, 0,
637 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
638 // Since at this point in the method invocation the exception handler
639 // would try to exit the monitor of synchronized methods which hasn't
640 // been entered yet, we set the thread local variable
641 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
642 // runtime, exception handling i.e. unlock_if_synchronized_method will
643 // check this thread local flag.
644 // This flag has two effects, one is to force an unwind in the topmost
645 // interpreter frame and not perform an unlock while doing so.
646
647 __ movbool(true, G3_scratch);
648 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
649
650
651 // increment invocation counter and check for overflow
652 //
653 // Note: checking for negative value instead of overflow
654 // so we have a 'sticky' overflow test (may be of
655 // importance as soon as we have true MT/MP)
656 Label invocation_counter_overflow;
657 if (inc_counter) {
658 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
659 }
660 Label Lcontinue;
661 __ bind(Lcontinue);
662
663 bang_stack_shadow_pages(true);
664 // reset the _do_not_unlock_if_synchronized flag
665 __ stbool(G0, do_not_unlock_if_synchronized);
666
667 // check for synchronized methods
668 // Must happen AFTER invocation_counter check, so method is not locked
669 // if counter overflows.
670
671 if (synchronized) {
672 lock_method();
673 // Don't see how G2_thread is preserved here...
674 // __ verify_thread(); QQQ destroys L0,L1 can't use
675 } else {
676 #ifdef ASSERT
677 { Label ok;
678 __ ld_ptr(STATE(_method), G5_method);
679 __ ld(access_flags, O0);
680 __ btst(JVM_ACC_SYNCHRONIZED, O0);
681 __ br( Assembler::zero, false, Assembler::pt, ok);
682 __ delayed()->nop();
683 __ stop("method needs synchronization");
684 __ bind(ok);
685 }
686 #endif // ASSERT
687 }
688
689 // start execution
690
691 // __ verify_thread(); kills L1,L2 can't use at the moment
692
693 // jvmti/jvmpi support
694 __ notify_method_entry();
695
696 // native call
697
698 // (note that O0 is never an oop--at most it is a handle)
699 // It is important not to smash any handles created by this call,
700 // until any oop handle in O0 is dereferenced.
701
702 // (note that the space for outgoing params is preallocated)
703
704 // get signature handler
705
706 Label pending_exception_present;
707
708 { Label L;
709 __ ld_ptr(STATE(_method), G5_method);
710 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
711 __ tst(G3_scratch);
712 __ brx(Assembler::notZero, false, Assembler::pt, L);
713 __ delayed()->nop();
714 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
715 __ ld_ptr(STATE(_method), G5_method);
716
717 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
718 __ ld_ptr(exception_addr, G3_scratch);
719 __ br_notnull_short(G3_scratch, Assembler::pn, pending_exception_present);
720 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
721 __ bind(L);
722 }
723
724 // Push a new frame so that the args will really be stored in
725 // Copy a few locals across so the new frame has the variables
726 // we need but these values will be dead at the jni call and
727 // therefore not gc volatile like the values in the current
728 // frame (Lstate in particular)
729
730 // Flush the state pointer to the register save area
731 // Which is the only register we need for a stack walk.
732 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
733
734 __ mov(Lstate, O1); // Need to pass the state pointer across the frame
735
736 // Calculate current frame size
737 __ sub(SP, FP, O3); // Calculate negative of current frame size
738 __ save(SP, O3, SP); // Allocate an identical sized frame
739
740 __ mov(I1, Lstate); // In the "natural" register.
741
742 // Note I7 has leftover trash. Slow signature handler will fill it in
743 // should we get there. Normal jni call will set reasonable last_Java_pc
744 // below (and fix I7 so the stack trace doesn't have a meaningless frame
745 // in it).
746
747
748 // call signature handler
749 __ ld_ptr(STATE(_method), Lmethod);
750 __ ld_ptr(STATE(_locals), Llocals);
751
752 __ callr(G3_scratch, 0);
753 __ delayed()->nop();
754 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed)
755
756 { Label not_static;
757
758 __ ld_ptr(STATE(_method), G5_method);
759 __ ld(access_flags, O0);
760 __ btst(JVM_ACC_STATIC, O0);
761 __ br( Assembler::zero, false, Assembler::pt, not_static);
762 __ delayed()->
763 // get native function entry point(O0 is a good temp until the very end)
764 ld_ptr(Address(G5_method, 0, in_bytes(Method::native_function_offset())), O0);
765 // for static methods insert the mirror argument
766 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
767
768 __ ld_ptr(Address(G5_method, 0, in_bytes(Method:: const_offset())), O1);
769 __ ld_ptr(Address(O1, 0, in_bytes(ConstMethod::constants_offset())), O1);
770 __ ld_ptr(Address(O1, 0, ConstantPool::pool_holder_offset_in_bytes()), O1);
771 __ ld_ptr(O1, mirror_offset, O1);
772 // where the mirror handle body is allocated:
773 #ifdef ASSERT
774 if (!PrintSignatureHandlers) // do not dirty the output with this
775 { Label L;
776 __ tst(O1);
777 __ brx(Assembler::notZero, false, Assembler::pt, L);
778 __ delayed()->nop();
779 __ stop("mirror is missing");
780 __ bind(L);
781 }
782 #endif // ASSERT
783 __ st_ptr(O1, STATE(_oop_temp));
784 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add
785 __ bind(not_static);
786 }
787
788 // At this point, arguments have been copied off of stack into
789 // their JNI positions, which are O1..O5 and SP[68..].
790 // Oops are boxed in-place on the stack, with handles copied to arguments.
791 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
792
793 #ifdef ASSERT
794 { Label L;
795 __ tst(O0);
796 __ brx(Assembler::notZero, false, Assembler::pt, L);
797 __ delayed()->nop();
798 __ stop("native entry point is missing");
799 __ bind(L);
800 }
801 #endif // ASSERT
802
803 //
804 // setup the java frame anchor
805 //
806 // The scavenge function only needs to know that the PC of this frame is
807 // in the interpreter method entry code, it doesn't need to know the exact
808 // PC and hence we can use O7 which points to the return address from the
809 // previous call in the code stream (signature handler function)
810 //
811 // The other trick is we set last_Java_sp to FP instead of the usual SP because
812 // we have pushed the extra frame in order to protect the volatile register(s)
813 // in that frame when we return from the jni call
814 //
815
816
817 __ set_last_Java_frame(FP, O7);
818 __ mov(O7, I7); // make dummy interpreter frame look like one above,
819 // not meaningless information that'll confuse me.
820
821 // flush the windows now. We don't care about the current (protection) frame
822 // only the outer frames
823
824 __ flush_windows();
825
826 // mark windows as flushed
827 Address flags(G2_thread,
828 0,
829 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
830 __ set(JavaFrameAnchor::flushed, G3_scratch);
831 __ st(G3_scratch, flags);
832
833 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
834
835 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
836 #ifdef ASSERT
837 { Label L;
838 __ ld(thread_state, G3_scratch);
839 __ cmp(G3_scratch, _thread_in_Java);
840 __ br(Assembler::equal, false, Assembler::pt, L);
841 __ delayed()->nop();
842 __ stop("Wrong thread state in native stub");
843 __ bind(L);
844 }
845 #endif // ASSERT
846 __ set(_thread_in_native, G3_scratch);
847 __ st(G3_scratch, thread_state);
848
849 // Call the jni method, using the delay slot to set the JNIEnv* argument.
850 __ callr(O0, 0);
851 __ delayed()->
852 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
853 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
854
855 // must we block?
856
857 // Block, if necessary, before resuming in _thread_in_Java state.
858 // In order for GC to work, don't clear the last_Java_sp until after blocking.
859 { Label no_block;
860 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
861
862 // Switch thread to "native transition" state before reading the synchronization state.
863 // This additional state is necessary because reading and testing the synchronization
864 // state is not atomic w.r.t. GC, as this scenario demonstrates:
865 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
866 // VM thread changes sync state to synchronizing and suspends threads for GC.
867 // Thread A is resumed to finish this native method, but doesn't block here since it
868 // didn't see any synchronization is progress, and escapes.
869 __ set(_thread_in_native_trans, G3_scratch);
870 __ st(G3_scratch, thread_state);
871 if(os::is_MP()) {
872 // Write serialization page so VM thread can do a pseudo remote membar.
873 // We use the current thread pointer to calculate a thread specific
874 // offset to write to within the page. This minimizes bus traffic
875 // due to cache line collision.
876 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
877 }
878 __ load_contents(sync_state, G3_scratch);
879 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
880
881
882 Label L;
883 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
884 __ br(Assembler::notEqual, false, Assembler::pn, L);
885 __ delayed()->
886 ld(suspend_state, G3_scratch);
887 __ cmp(G3_scratch, 0);
888 __ br(Assembler::equal, false, Assembler::pt, no_block);
889 __ delayed()->nop();
890 __ bind(L);
891
892 // Block. Save any potential method result value before the operation and
893 // use a leaf call to leave the last_Java_frame setup undisturbed.
894 save_native_result();
895 __ call_VM_leaf(noreg,
896 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
897 G2_thread);
898 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
899 // Restore any method result value
900 restore_native_result();
901 __ bind(no_block);
902 }
903
904 // Clear the frame anchor now
905
906 __ reset_last_Java_frame();
907
908 // Move the result handler address
909 __ mov(Lscratch, G3_scratch);
910 // return possible result to the outer frame
911 #ifndef __LP64
912 __ mov(O0, I0);
913 __ restore(O1, G0, O1);
914 #else
915 __ restore(O0, G0, O0);
916 #endif /* __LP64 */
917
918 // Move result handler to expected register
919 __ mov(G3_scratch, Lscratch);
920
921
922 // thread state is thread_in_native_trans. Any safepoint blocking has
923 // happened in the trampoline we are ready to switch to thread_in_Java.
924
925 __ set(_thread_in_Java, G3_scratch);
926 __ st(G3_scratch, thread_state);
927
928 // If we have an oop result store it where it will be safe for any further gc
929 // until we return now that we've released the handle it might be protected by
930
931 {
932 Label no_oop, store_result;
933
934 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
935 __ cmp(G3_scratch, Lscratch);
936 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
937 __ delayed()->nop();
938 __ addcc(G0, O0, O0);
939 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
940 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
941 __ mov(G0, O0);
942
943 __ bind(store_result);
944 // Store it where gc will look for it and result handler expects it.
945 __ st_ptr(O0, STATE(_oop_temp));
946
947 __ bind(no_oop);
948
949 }
950
951 // reset handle block
952 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
953 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
954
955
956 // handle exceptions (exception handling will handle unlocking!)
957 { Label L;
958 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
959
960 __ ld_ptr(exception_addr, Gtemp);
961 __ tst(Gtemp);
962 __ brx(Assembler::equal, false, Assembler::pt, L);
963 __ delayed()->nop();
964 __ bind(pending_exception_present);
965 // With c++ interpreter we just leave it pending caller will do the correct thing. However...
966 // Like x86 we ignore the result of the native call and leave the method locked. This
967 // seems wrong to leave things locked.
968
969 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
970 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
971
972 __ bind(L);
973 }
974
975 // jvmdi/jvmpi support (preserves thread register)
976 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
977
978 if (synchronized) {
979 // save and restore any potential method result value around the unlocking operation
980 save_native_result();
981
982 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
983 // Get the initial monitor we allocated
984 __ sub(Lstate, entry_size, O1); // initial monitor
985 __ unlock_object(O1);
986 restore_native_result();
987 }
988
989 #if defined(COMPILER2) && !defined(_LP64)
990
991 // C2 expects long results in G1 we can't tell if we're returning to interpreted
992 // or compiled so just be safe.
993
994 __ sllx(O0, 32, G1); // Shift bits into high G1
995 __ srl (O1, 0, O1); // Zero extend O1
996 __ or3 (O1, G1, G1); // OR 64 bits into G1
997
998 #endif /* COMPILER2 && !_LP64 */
999
1000 #ifdef ASSERT
1001 {
1002 Label ok;
1003 __ cmp(I5_savedSP, FP);
1004 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1005 __ delayed()->nop();
1006 __ stop("bad I5_savedSP value");
1007 __ should_not_reach_here();
1008 __ bind(ok);
1009 }
1010 #endif
1011 // Calls result handler which POPS FRAME
1012 if (TraceJumps) {
1013 // Move target to register that is recordable
1014 __ mov(Lscratch, G3_scratch);
1015 __ JMP(G3_scratch, 0);
1016 } else {
1017 __ jmp(Lscratch, 0);
1018 }
1019 __ delayed()->nop();
1020
1021 if (inc_counter) {
1022 // handle invocation counter overflow
1023 __ bind(invocation_counter_overflow);
1024 generate_counter_overflow(Lcontinue);
1025 }
1026
1027
1028 return entry;
1029 }
1030
1031 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1032 const Register prev_state,
1033 bool native) {
1034
1035 // On entry
1036 // G5_method - caller's method
1037 // Gargs - points to initial parameters (i.e. locals[0])
1038 // G2_thread - valid? (C1 only??)
1039 // "prev_state" - contains any previous frame manager state which we must save a link
1040 //
1041 // On return
1042 // "state" is a pointer to the newly allocated state object. We must allocate and initialize
1043 // a new interpretState object and the method expression stack.
1044
1045 assert_different_registers(state, prev_state);
1046 assert_different_registers(prev_state, G3_scratch);
1047 const Register Gtmp = G3_scratch;
1048 const Address constMethod (G5_method, 0, in_bytes(Method::const_offset()));
1049 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
1050 const Address size_of_parameters(G5_method, 0, in_bytes(Method::size_of_parameters_offset()));
1051 const Address size_of_locals (G5_method, 0, in_bytes(Method::size_of_locals_offset()));
1052
1053 // slop factor is two extra slots on the expression stack so that
1054 // we always have room to store a result when returning from a call without parameters
1055 // that returns a result.
1056
1057 const int slop_factor = 2*wordSize;
1058
1059 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1060 //6815692//Method::extra_stack_words() + // extra push slots for MH adapters
1061 frame::memory_parameter_word_sp_offset + // register save area + param window
1062 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1063
1064 // XXX G5_method valid
1065
1066 // Now compute new frame size
1067
1068 if (native) {
1069 __ lduh( size_of_parameters, Gtmp );
1070 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words
1071 } else {
1072 // Full size expression stack
1073 __ ld_ptr(constMethod, Gtmp);
1074 __ lduh(Gtmp, in_bytes(ConstMethod::max_stack_offset()), Gtmp);
1075 }
1076 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1077
1078 __ neg(Gtmp); // negative space for stack/parameters in words
1079 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1080 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1081
1082 // Need to do stack size check here before we fault on large frames
1083
1084 Label stack_ok;
1085
1086 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1087 (StackRedPages+StackYellowPages);
1088
1089
1090 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1091 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1092 // compute stack bottom
1093 __ sub(O0, O1, O0);
1094
1095 // Avoid touching the guard pages
1096 // Also a fudge for frame size of BytecodeInterpreter::run
1097 // It varies from 1k->4k depending on build type
1098 const int fudge = 6 * K;
1099
1100 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1101
1102 __ add(O0, O1, O0);
1103 __ sub(O0, Gtmp, O0);
1104 __ cmp(SP, O0);
1105 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1106 __ delayed()->nop();
1107
1108 // throw exception return address becomes throwing pc
1109
1110 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1111 __ stop("never reached");
1112
1113 __ bind(stack_ok);
1114
1115 __ save(SP, Gtmp, SP); // setup new frame and register window
1116
1117 // New window I7 call_stub or previous activation
1118 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1119 //
1120 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1121 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1122
1123 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1124
1125 // Initialize a new Interpreter state
1126 // orig_sp - caller's original sp
1127 // G2_thread - thread
1128 // Gargs - &locals[0] (unbiased?)
1129 // G5_method - method
1130 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1131
1132
1133 __ set(0xdead0004, O1);
1134
1135
1136 __ st_ptr(Gargs, XXX_STATE(_locals));
1137 __ st_ptr(G0, XXX_STATE(_oop_temp));
1138
1139 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1140 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1141 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1142
1143 if (native) {
1144 __ st_ptr(G0, XXX_STATE(_bcp));
1145 } else {
1146 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), O2); // get ConstMethod*
1147 __ add(O2, in_bytes(ConstMethod::codes_offset()), O2); // get bcp
1148 __ st_ptr(O2, XXX_STATE(_bcp));
1149 }
1150
1151 __ st_ptr(G0, XXX_STATE(_mdx));
1152 __ st_ptr(G5_method, XXX_STATE(_method));
1153
1154 __ set((int) BytecodeInterpreter::method_entry, O1);
1155 __ st(O1, XXX_STATE(_msg));
1156
1157 __ ld_ptr(constMethod, O3);
1158 __ ld_ptr(O3, in_bytes(ConstMethod::constants_offset()), O3);
1159 __ ld_ptr(O3, ConstantPool::cache_offset_in_bytes(), O2);
1160 __ st_ptr(O2, XXX_STATE(_constants));
1161
1162 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1163
1164 // Monitor base is just start of BytecodeInterpreter object;
1165 __ mov(state, O2);
1166 __ st_ptr(O2, XXX_STATE(_monitor_base));
1167
1168 // Do we need a monitor for synchonized method?
1169 {
1170 __ ld(access_flags, O1);
1171 Label done;
1172 Label got_obj;
1173 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1174 __ br( Assembler::zero, false, Assembler::pt, done);
1175
1176 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1177 __ delayed()->btst(JVM_ACC_STATIC, O1);
1178 __ ld_ptr(XXX_STATE(_locals), O1);
1179 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1180 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1181 __ ld_ptr(constMethod, O1);
1182 __ ld_ptr( O1, in_bytes(ConstMethod::constants_offset()), O1);
1183 __ ld_ptr( O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1184 // lock the mirror, not the Klass*
1185 __ ld_ptr( O1, mirror_offset, O1);
1186
1187 __ bind(got_obj);
1188
1189 #ifdef ASSERT
1190 __ tst(O1);
1191 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
1192 #endif // ASSERT
1193
1194 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1195 __ sub(SP, entry_size, SP); // account for initial monitor
1196 __ sub(O2, entry_size, O2); // initial monitor
1197 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1198 __ bind(done);
1199 }
1200
1201 // Remember initial frame bottom
1202
1203 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1204
1205 __ st_ptr(O2, XXX_STATE(_stack_base));
1206
1207 __ sub(O2, wordSize, O2); // prepush
1208 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1209
1210 // Full size expression stack
1211 __ ld_ptr(constMethod, O3);
1212 __ lduh(O3, in_bytes(ConstMethod::max_stack_offset()), O3);
1213 guarantee(!EnableInvokeDynamic, "no support yet for java.lang.invoke.MethodHandle"); //6815692
1214 //6815692//if (EnableInvokeDynamic)
1215 //6815692// __ inc(O3, Method::extra_stack_entries());
1216 __ sll(O3, LogBytesPerWord, O3);
1217 __ sub(O2, O3, O3);
1218 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1219 __ st_ptr(O3, XXX_STATE(_stack_limit));
1220
1221 if (!native) {
1222 //
1223 // Code to initialize locals
1224 //
1225 Register init_value = noreg; // will be G0 if we must clear locals
1226 // Now zero locals
1227 if (true /* zerolocals */ || ClearInterpreterLocals) {
1228 // explicitly initialize locals
1229 init_value = G0;
1230 } else {
1231 #ifdef ASSERT
1232 // initialize locals to a garbage pattern for better debugging
1233 init_value = O3;
1234 __ set( 0x0F0F0F0F, init_value );
1235 #endif // ASSERT
1236 }
1237 if (init_value != noreg) {
1238 Label clear_loop;
1239
1240 // NOTE: If you change the frame layout, this code will need to
1241 // be updated!
1242 __ lduh( size_of_locals, O2 );
1243 __ lduh( size_of_parameters, O1 );
1244 __ sll( O2, LogBytesPerWord, O2);
1245 __ sll( O1, LogBytesPerWord, O1 );
1246 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1247 __ sub( L2_scratch, O2, O2 );
1248 __ sub( L2_scratch, O1, O1 );
1249
1250 __ bind( clear_loop );
1251 __ inc( O2, wordSize );
1252
1253 __ cmp( O2, O1 );
1254 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1255 __ delayed()->st_ptr( init_value, O2, 0 );
1256 }
1257 }
1258 }
1259 // Find preallocated monitor and lock method (C++ interpreter)
1260 //
1261 void InterpreterGenerator::lock_method(void) {
1262 // Lock the current method.
1263 // Destroys registers L2_scratch, L3_scratch, O0
1264 //
1265 // Find everything relative to Lstate
1266
1267 #ifdef ASSERT
1268 __ ld_ptr(STATE(_method), L2_scratch);
1269 __ ld(L2_scratch, in_bytes(Method::access_flags_offset()), O0);
1270
1271 { Label ok;
1272 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1273 __ br( Assembler::notZero, false, Assembler::pt, ok);
1274 __ delayed()->nop();
1275 __ stop("method doesn't need synchronization");
1276 __ bind(ok);
1277 }
1278 #endif // ASSERT
1279
1280 // monitor is already allocated at stack base
1281 // and the lockee is already present
1282 __ ld_ptr(STATE(_stack_base), L2_scratch);
1283 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1284 __ lock_object(L2_scratch, O0);
1285
1286 }
1287
1288 // Generate code for handling resuming a deopted method
1289 void CppInterpreterGenerator::generate_deopt_handling() {
1290
1291 Label return_from_deopt_common;
1292
1293 // deopt needs to jump to here to enter the interpreter (return a result)
1294 deopt_frame_manager_return_atos = __ pc();
1295
1296 // O0/O1 live
1297 __ ba(return_from_deopt_common);
1298 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1299
1300
1301 // deopt needs to jump to here to enter the interpreter (return a result)
1302 deopt_frame_manager_return_btos = __ pc();
1303
1304 // O0/O1 live
1305 __ ba(return_from_deopt_common);
1306 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1307
1308 // deopt needs to jump to here to enter the interpreter (return a result)
1309 deopt_frame_manager_return_itos = __ pc();
1310
1311 // O0/O1 live
1312 __ ba(return_from_deopt_common);
1313 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1314
1315 // deopt needs to jump to here to enter the interpreter (return a result)
1316
1317 deopt_frame_manager_return_ltos = __ pc();
1318 #if !defined(_LP64) && defined(COMPILER2)
1319 // All return values are where we want them, except for Longs. C2 returns
1320 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1321 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1322 // build even if we are returning from interpreted we just do a little
1323 // stupid shuffing.
1324 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1325 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1326 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1327
1328 __ srl (G1, 0,O1);
1329 __ srlx(G1,32,O0);
1330 #endif /* !_LP64 && COMPILER2 */
1331 // O0/O1 live
1332 __ ba(return_from_deopt_common);
1333 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1334
1335 // deopt needs to jump to here to enter the interpreter (return a result)
1336
1337 deopt_frame_manager_return_ftos = __ pc();
1338 // O0/O1 live
1339 __ ba(return_from_deopt_common);
1340 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1341
1342 // deopt needs to jump to here to enter the interpreter (return a result)
1343 deopt_frame_manager_return_dtos = __ pc();
1344
1345 // O0/O1 live
1346 __ ba(return_from_deopt_common);
1347 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1348
1349 // deopt needs to jump to here to enter the interpreter (return a result)
1350 deopt_frame_manager_return_vtos = __ pc();
1351
1352 // O0/O1 live
1353 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1354
1355 // Deopt return common
1356 // an index is present that lets us move any possible result being
1357 // return to the interpreter's stack
1358 //
1359 __ bind(return_from_deopt_common);
1360
1361 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1362 // stack is in the state that the calling convention left it.
1363 // Copy the result from native abi result and place it on java expression stack.
1364
1365 // Current interpreter state is present in Lstate
1366
1367 // Get current pre-pushed top of interpreter stack
1368 // Any result (if any) is in native abi
1369 // result type index is in L3_scratch
1370
1371 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1372
1373 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1374 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1375 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1376 __ jmpl(Lscratch, G0, O7); // and convert it
1377 __ delayed()->nop();
1378
1379 // L1_scratch points to top of stack (prepushed)
1380 __ st_ptr(L1_scratch, STATE(_stack));
1381 }
1382
1383 // Generate the code to handle a more_monitors message from the c++ interpreter
1384 void CppInterpreterGenerator::generate_more_monitors() {
1385
1386 Label entry, loop;
1387 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1388 // 1. compute new pointers // esp: old expression stack top
1389 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1390 __ sub(L4_scratch, entry_size, L4_scratch);
1391 __ st_ptr(L4_scratch, STATE(_stack_base));
1392
1393 __ sub(SP, entry_size, SP); // Grow stack
1394 __ st_ptr(SP, STATE(_frame_bottom));
1395
1396 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1397 __ sub(L2_scratch, entry_size, L2_scratch);
1398 __ st_ptr(L2_scratch, STATE(_stack_limit));
1399
1400 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1401 __ sub(L1_scratch, entry_size, L1_scratch);
1402 __ st_ptr(L1_scratch, STATE(_stack));
1403 __ ba(entry);
1404 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1405
1406 // 2. move expression stack
1407
1408 __ bind(loop);
1409 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1410 __ add(L1_scratch, wordSize, L1_scratch);
1411 __ bind(entry);
1412 __ cmp(L1_scratch, L4_scratch);
1413 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1414 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1415
1416 // now zero the slot so we can find it.
1417 __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1418
1419 }
1420
1421 // Initial entry to C++ interpreter from the call_stub.
1422 // This entry point is called the frame manager since it handles the generation
1423 // of interpreter activation frames via requests directly from the vm (via call_stub)
1424 // and via requests from the interpreter. The requests from the call_stub happen
1425 // directly thru the entry point. Requests from the interpreter happen via returning
1426 // from the interpreter and examining the message the interpreter has returned to
1427 // the frame manager. The frame manager can take the following requests:
1428
1429 // NO_REQUEST - error, should never happen.
1430 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1431 // allocate a new monitor.
1432 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1433 // happens during entry during the entry via the call stub.
1434 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1435 //
1436 // Arguments:
1437 //
1438 // ebx: Method*
1439 // ecx: receiver - unused (retrieved from stack as needed)
1440 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1441 //
1442 //
1443 // Stack layout at entry
1444 //
1445 // [ return address ] <--- esp
1446 // [ parameter n ]
1447 // ...
1448 // [ parameter 1 ]
1449 // [ expression stack ]
1450 //
1451 //
1452 // We are free to blow any registers we like because the call_stub which brought us here
1453 // initially has preserved the callee save registers already.
1454 //
1455 //
1456
1457 static address interpreter_frame_manager = NULL;
1458
1459 #ifdef ASSERT
1460 #define VALIDATE_STATE(scratch, marker) \
1461 { \
1462 Label skip; \
1463 __ ld_ptr(STATE(_self_link), scratch); \
1464 __ cmp(Lstate, scratch); \
1465 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1466 __ delayed()->nop(); \
1467 __ breakpoint_trap(); \
1468 __ emit_long(marker); \
1469 __ bind(skip); \
1470 }
1471 #else
1472 #define VALIDATE_STATE(scratch, marker)
1473 #endif /* ASSERT */
1474
1475 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1476 //
1477 // Adjust caller's stack so that all the locals can be contiguous with
1478 // the parameters.
1479 // Worries about stack overflow make this a pain.
1480 //
1481 // Destroys args, G3_scratch, G3_scratch
1482 // In/Out O5_savedSP (sender's original SP)
1483 //
1484 // assert_different_registers(state, prev_state);
1485 const Register Gtmp = G3_scratch;
1486 const Register tmp = O2;
1487 const Address size_of_parameters(G5_method, 0, in_bytes(Method::size_of_parameters_offset()));
1488 const Address size_of_locals (G5_method, 0, in_bytes(Method::size_of_locals_offset()));
1489
1490 __ lduh(size_of_parameters, tmp);
1491 __ sll(tmp, LogBytesPerWord, Gtmp); // parameter size in bytes
1492 __ add(args, Gtmp, Gargs); // points to first local + BytesPerWord
1493 // NEW
1494 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1495 // determine extra space for non-argument locals & adjust caller's SP
1496 // Gtmp1: parameter size in words
1497 __ lduh(size_of_locals, Gtmp);
1498 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1499
1500 #if 1
1501 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1502 // the call_stub will place the final interpreter argument at
1503 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1504 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1505 // and try to make it look good in the debugger we will store the argument to
1506 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1507 // extra space for the compiler this will overwrite locals in the local array of the
1508 // interpreter.
1509 // QQQ still needed with frameless adapters???
1510
1511 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1512
1513 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1514 #endif // 1
1515
1516
1517 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1518 }
1519
1520 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1521
1522 // G5_method: Method*
1523 // G2_thread: thread (unused)
1524 // Gargs: bottom of args (sender_sp)
1525 // O5: sender's sp
1526
1527 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1528 // the code is pretty much ready. Would need to change the test below and for good measure
1529 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1530 // routines. Not clear this is worth it yet.
1531
1532 if (interpreter_frame_manager) {
1533 return interpreter_frame_manager;
1534 }
1535
1536 __ bind(frame_manager_entry);
1537
1538 // the following temporary registers are used during frame creation
1539 const Register Gtmp1 = G3_scratch;
1540 // const Register Lmirror = L1; // native mirror (native calls only)
1541
1542 const Address constMethod (G5_method, 0, in_bytes(Method::const_offset()));
1543 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
1544 const Address size_of_parameters(G5_method, 0, in_bytes(Method::size_of_parameters_offset()));
1545 const Address size_of_locals (G5_method, 0, in_bytes(Method::size_of_locals_offset()));
1546
1547 address entry_point = __ pc();
1548 __ mov(G0, prevState); // no current activation
1549
1550
1551 Label re_dispatch;
1552
1553 __ bind(re_dispatch);
1554
1555 // Interpreter needs to have locals completely contiguous. In order to do that
1556 // We must adjust the caller's stack pointer for any locals beyond just the
1557 // parameters
1558 adjust_callers_stack(Gargs);
1559
1560 // O5_savedSP still contains sender's sp
1561
1562 // NEW FRAME
1563
1564 generate_compute_interpreter_state(Lstate, prevState, false);
1565
1566 // At this point a new interpreter frame and state object are created and initialized
1567 // Lstate has the pointer to the new activation
1568 // Any stack banging or limit check should already be done.
1569
1570 Label call_interpreter;
1571
1572 __ bind(call_interpreter);
1573
1574
1575 #if 1
1576 __ set(0xdead002, Lmirror);
1577 __ set(0xdead002, L2_scratch);
1578 __ set(0xdead003, L3_scratch);
1579 __ set(0xdead004, L4_scratch);
1580 __ set(0xdead005, Lscratch);
1581 __ set(0xdead006, Lscratch2);
1582 __ set(0xdead007, L7_scratch);
1583
1584 __ set(0xdeaf002, O2);
1585 __ set(0xdeaf003, O3);
1586 __ set(0xdeaf004, O4);
1587 __ set(0xdeaf005, O5);
1588 #endif
1589
1590 // Call interpreter (stack bang complete) enter here if message is
1591 // set and we know stack size is valid
1592
1593 Label call_interpreter_2;
1594
1595 __ bind(call_interpreter_2);
1596
1597 #ifdef ASSERT
1598 {
1599 Label skip;
1600 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1601 __ cmp(G3_scratch, SP);
1602 __ brx(Assembler::equal, false, Assembler::pt, skip);
1603 __ delayed()->nop();
1604 __ stop("SP not restored to frame bottom");
1605 __ bind(skip);
1606 }
1607 #endif
1608
1609 VALIDATE_STATE(G3_scratch, 4);
1610 __ set_last_Java_frame(SP, noreg);
1611 __ mov(Lstate, O0); // (arg) pointer to current state
1612
1613 __ call(CAST_FROM_FN_PTR(address,
1614 JvmtiExport::can_post_interpreter_events() ?
1615 BytecodeInterpreter::runWithChecks
1616 : BytecodeInterpreter::run),
1617 relocInfo::runtime_call_type);
1618
1619 __ delayed()->nop();
1620
1621 __ ld_ptr(STATE(_thread), G2_thread);
1622 __ reset_last_Java_frame();
1623
1624 // examine msg from interpreter to determine next action
1625 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1626
1627 __ ld(STATE(_msg), L1_scratch); // Get new message
1628
1629 Label call_method;
1630 Label return_from_interpreted_method;
1631 Label throw_exception;
1632 Label do_OSR;
1633 Label bad_msg;
1634 Label resume_interpreter;
1635
1636 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1637 __ br(Assembler::equal, false, Assembler::pt, call_method);
1638 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1639 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1640 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1641 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1642 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1643 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1644 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1645 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1646
1647 // Allocate more monitor space, shuffle expression stack....
1648
1649 generate_more_monitors();
1650
1651 // new monitor slot allocated, resume the interpreter.
1652
1653 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1654 VALIDATE_STATE(G3_scratch, 5);
1655 __ ba(call_interpreter);
1656 __ delayed()->st(L1_scratch, STATE(_msg));
1657
1658 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1659 unctrap_frame_manager_entry = __ pc();
1660
1661 // QQQ what message do we send
1662
1663 __ ba(call_interpreter);
1664 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1665
1666 //=============================================================================
1667 // Returning from a compiled method into a deopted method. The bytecode at the
1668 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1669 // for the template based interpreter). Any stack space that was used by the
1670 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1671 // so all that we have to do is place any pending result on the expression stack
1672 // and resume execution on the next bytecode.
1673
1674 generate_deopt_handling();
1675
1676 // ready to resume the interpreter
1677
1678 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1679 __ ba(call_interpreter);
1680 __ delayed()->st(L1_scratch, STATE(_msg));
1681
1682 // Current frame has caught an exception we need to dispatch to the
1683 // handler. We can get here because a native interpreter frame caught
1684 // an exception in which case there is no handler and we must rethrow
1685 // If it is a vanilla interpreted frame the we simply drop into the
1686 // interpreter and let it do the lookup.
1687
1688 Interpreter::_rethrow_exception_entry = __ pc();
1689
1690 Label return_with_exception;
1691 Label unwind_and_forward;
1692
1693 // O0: exception
1694 // O7: throwing pc
1695
1696 // We want exception in the thread no matter what we ultimately decide about frame type.
1697
1698 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1699 __ verify_thread();
1700 __ st_ptr(O0, exception_addr);
1701
1702 // get the Method*
1703 __ ld_ptr(STATE(_method), G5_method);
1704
1705 // if this current frame vanilla or native?
1706
1707 __ ld(access_flags, Gtmp1);
1708 __ btst(JVM_ACC_NATIVE, Gtmp1);
1709 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1710 __ delayed()->nop();
1711
1712 // We drop thru to unwind a native interpreted frame with a pending exception
1713 // We jump here for the initial interpreter frame with exception pending
1714 // We unwind the current acivation and forward it to our caller.
1715
1716 __ bind(unwind_and_forward);
1717
1718 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1719 // as expected by forward_exception.
1720
1721 __ restore(FP, G0, SP); // unwind interpreter state frame
1722 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1723 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1724
1725 // Return point from a call which returns a result in the native abi
1726 // (c1/c2/jni-native). This result must be processed onto the java
1727 // expression stack.
1728 //
1729 // A pending exception may be present in which case there is no result present
1730
1731 address return_from_native_method = __ pc();
1732
1733 VALIDATE_STATE(G3_scratch, 6);
1734
1735 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1736 // stack is in the state that the calling convention left it.
1737 // Copy the result from native abi result and place it on java expression stack.
1738
1739 // Current interpreter state is present in Lstate
1740
1741 // Exception pending?
1742
1743 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1744 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1745 __ tst(Lscratch); // exception pending?
1746 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1747 __ delayed()->nop();
1748
1749 // Process the native abi result to java expression stack
1750
1751 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1752 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1753 __ lduh(L4_scratch, in_bytes(Method::size_of_parameters_offset()), L2_scratch); // get parameter size
1754 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1755 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1756 __ ld(L4_scratch, in_bytes(Method::result_index_offset()), L3_scratch); // called method result type index
1757
1758 // tosca is really just native abi
1759 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1760 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1761 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1762 __ jmpl(Lscratch, G0, O7); // and convert it
1763 __ delayed()->nop();
1764
1765 // L1_scratch points to top of stack (prepushed)
1766
1767 __ ba(resume_interpreter);
1768 __ delayed()->mov(L1_scratch, O1);
1769
1770 // An exception is being caught on return to a vanilla interpreter frame.
1771 // Empty the stack and resume interpreter
1772
1773 __ bind(return_with_exception);
1774
1775 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1776 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1777 __ ba(resume_interpreter);
1778 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1779
1780 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1781 // interpreter call, or native) and unwind this interpreter activation.
1782 // All monitors should be unlocked.
1783
1784 __ bind(return_from_interpreted_method);
1785
1786 VALIDATE_STATE(G3_scratch, 7);
1787
1788 Label return_to_initial_caller;
1789
1790 // Interpreted result is on the top of the completed activation expression stack.
1791 // We must return it to the top of the callers stack if caller was interpreted
1792 // otherwise we convert to native abi result and return to call_stub/c1/c2
1793 // The caller's expression stack was truncated by the call however the current activation
1794 // has enough stuff on the stack that we have usable space there no matter what. The
1795 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1796 // for the current activation
1797
1798 __ ld_ptr(STATE(_prev_link), L1_scratch);
1799 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1800 __ ld(L2_scratch, in_bytes(Method::result_index_offset()), L2_scratch);
1801 __ tst(L1_scratch);
1802 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1803 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1804
1805 // Copy result to callers java stack
1806
1807 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1808 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1809 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1810 __ ld_ptr(STATE(_locals), O1); // stack destination
1811
1812 // O0 - will be source, O1 - will be destination (preserved)
1813 __ jmpl(Lscratch, G0, O7); // and convert it
1814 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1815
1816 // O1 == &locals[0]
1817
1818 // Result is now on caller's stack. Just unwind current activation and resume
1819
1820 Label unwind_recursive_activation;
1821
1822
1823 __ bind(unwind_recursive_activation);
1824
1825 // O1 == &locals[0] (really callers stacktop) for activation now returning
1826 // returning to interpreter method from "recursive" interpreter call
1827 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1828 // to. Now all we must do is unwind the state from the completed call
1829
1830 // Must restore stack
1831 VALIDATE_STATE(G3_scratch, 8);
1832
1833 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1834 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1835 // frame and tell interpreter to resume.
1836
1837
1838 __ mov(O1, I1); // pass back new stack top across activation
1839 // POP FRAME HERE ==================================
1840 __ restore(FP, G0, SP); // unwind interpreter state frame
1841 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1842
1843
1844 // Resume the interpreter. The current frame contains the current interpreter
1845 // state object.
1846 //
1847 // O1 == new java stack pointer
1848
1849 __ bind(resume_interpreter);
1850 VALIDATE_STATE(G3_scratch, 10);
1851
1852 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1853
1854 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1855 __ st(L1_scratch, STATE(_msg));
1856 __ ba(call_interpreter_2);
1857 __ delayed()->st_ptr(O1, STATE(_stack));
1858
1859
1860 // Fast accessor methods share this entry point.
1861 // This works because frame manager is in the same codelet
1862 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1863 // we need to do a little register fixup here once we distinguish the two of them
1864 if (UseFastAccessorMethods && !synchronized) {
1865 // Call stub_return address still in O7
1866 __ bind(fast_accessor_slow_entry_path);
1867 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1868 __ cmp(Gtmp1, O7); // returning to interpreter?
1869 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1870 __ delayed()->nop();
1871 __ ba(re_dispatch);
1872 __ delayed()->mov(G0, prevState); // initial entry
1873
1874 }
1875
1876 // interpreter returning to native code (call_stub/c1/c2)
1877 // convert result and unwind initial activation
1878 // L2_scratch - scaled result type index
1879
1880 __ bind(return_to_initial_caller);
1881
1882 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1883 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1884 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1885 __ jmpl(Lscratch, G0, O7); // and convert it
1886 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1887
1888 Label unwind_initial_activation;
1889 __ bind(unwind_initial_activation);
1890
1891 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1892 // we can return here with an exception that wasn't handled by interpreted code
1893 // how does c1/c2 see it on return?
1894
1895 // compute resulting sp before/after args popped depending upon calling convention
1896 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1897 //
1898 // POP FRAME HERE ==================================
1899 __ restore(FP, G0, SP);
1900 __ retl();
1901 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1902
1903 // OSR request, unwind the current frame and transfer to the OSR entry
1904 // and enter OSR nmethod
1905
1906 __ bind(do_OSR);
1907 Label remove_initial_frame;
1908 __ ld_ptr(STATE(_prev_link), L1_scratch);
1909 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1910
1911 // We are going to pop this frame. Is there another interpreter frame underneath
1912 // it or is it callstub/compiled?
1913
1914 __ tst(L1_scratch);
1915 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1916 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1917
1918 // Frame underneath is an interpreter frame simply unwind
1919 // POP FRAME HERE ==================================
1920 __ restore(FP, G0, SP); // unwind interpreter state frame
1921 __ mov(I5_savedSP->after_restore(), SP);
1922
1923 // Since we are now calling native need to change our "return address" from the
1924 // dummy RecursiveInterpreterActivation to a return from native
1925
1926 __ set((intptr_t)return_from_native_method - 8, O7);
1927
1928 __ jmpl(G3_scratch, G0, G0);
1929 __ delayed()->mov(G1_scratch, O0);
1930
1931 __ bind(remove_initial_frame);
1932
1933 // POP FRAME HERE ==================================
1934 __ restore(FP, G0, SP);
1935 __ mov(I5_savedSP->after_restore(), SP);
1936 __ jmpl(G3_scratch, G0, G0);
1937 __ delayed()->mov(G1_scratch, O0);
1938
1939 // Call a new method. All we do is (temporarily) trim the expression stack
1940 // push a return address to bring us back to here and leap to the new entry.
1941 // At this point we have a topmost frame that was allocated by the frame manager
1942 // which contains the current method interpreted state. We trim this frame
1943 // of excess java expression stack entries and then recurse.
1944
1945 __ bind(call_method);
1946
1947 // stack points to next free location and not top element on expression stack
1948 // method expects sp to be pointing to topmost element
1949
1950 __ ld_ptr(STATE(_thread), G2_thread);
1951 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1952
1953
1954 // SP already takes in to account the 2 extra words we use for slop
1955 // when we call a "static long no_params()" method. So if
1956 // we trim back sp by the amount of unused java expression stack
1957 // there will be automagically the 2 extra words we need.
1958 // We also have to worry about keeping SP aligned.
1959
1960 __ ld_ptr(STATE(_stack), Gargs);
1961 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1962
1963 // compute the unused java stack size
1964 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1965
1966 // Round down the unused space to that stack is always 16-byte aligned
1967 // by making the unused space a multiple of the size of two longs.
1968
1969 __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1970
1971 // Now trim the stack
1972 __ add(SP, L2_scratch, SP);
1973
1974
1975 // Now point to the final argument (account for prepush)
1976 __ add(Gargs, wordSize, Gargs);
1977 #ifdef ASSERT
1978 // Make sure we have space for the window
1979 __ sub(Gargs, SP, L1_scratch);
1980 __ cmp(L1_scratch, 16*wordSize);
1981 {
1982 Label skip;
1983 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1984 __ delayed()->nop();
1985 __ stop("killed stack");
1986 __ bind(skip);
1987 }
1988 #endif // ASSERT
1989
1990 // Create a new frame where we can store values that make it look like the interpreter
1991 // really recursed.
1992
1993 // prepare to recurse or call specialized entry
1994
1995 // First link the registers we need
1996
1997 // make the pc look good in debugger
1998 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
1999 // argument too
2000 __ mov(Lstate, I0);
2001
2002 // Record our sending SP
2003 __ mov(SP, O5_savedSP);
2004
2005 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
2006 __ set((intptr_t) entry_point, L1_scratch);
2007 __ cmp(L1_scratch, L2_scratch);
2008 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
2009 __ delayed()->mov(Lstate, prevState); // link activations
2010
2011 // method uses specialized entry, push a return so we look like call stub setup
2012 // this path will handle fact that result is returned in registers and not
2013 // on the java stack.
2014
2015 __ set((intptr_t)return_from_native_method - 8, O7);
2016 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
2017 __ delayed()->nop();
2018
2019 //
2020 // Bad Message from interpreter
2021 //
2022 __ bind(bad_msg);
2023 __ stop("Bad message from interpreter");
2024
2025 // Interpreted method "returned" with an exception pass it on...
2026 // Pass result, unwind activation and continue/return to interpreter/call_stub
2027 // We handle result (if any) differently based on return to interpreter or call_stub
2028
2029 __ bind(throw_exception);
2030 __ ld_ptr(STATE(_prev_link), L1_scratch);
2031 __ tst(L1_scratch);
2032 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2033 __ delayed()->nop();
2034
2035 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
2036 __ ba(unwind_recursive_activation);
2037 __ delayed()->nop();
2038
2039 interpreter_frame_manager = entry_point;
2040 return entry_point;
2041 }
2042
2043 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2044 : CppInterpreterGenerator(code) {
2045 generate_all(); // down here so it can be "virtual"
2046 }
2047
2048
2049 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2050
2051 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2052 // expression stack, the callee will have callee_extra_locals (so we can account for
2053 // frame extension) and monitor_size for monitors. Basically we need to calculate
2054 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2055 //
2056 //
2057 // The big complicating thing here is that we must ensure that the stack stays properly
2058 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2059 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2060 // the caller so we must ensure that it is properly aligned for our callee.
2061 //
2062 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2063 // stack at all times to deal with the "stack long no_params()" method issue. This
2064 // is "slop_factor" here.
2065 const int slop_factor = 2;
2066
2067 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2068 frame::memory_parameter_word_sp_offset; // register save area + param window
2069 const int extra_stack = 0; //6815692//Method::extra_stack_entries();
2070 return (round_to(max_stack +
2071 extra_stack +
2072 slop_factor +
2073 fixed_size +
2074 monitor_size +
2075 (callee_extra_locals * Interpreter::stackElementWords), WordsPerLong));
2076
2077 }
2078
2079 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
2080
2081 // See call_stub code
2082 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2083 WordsPerLong); // 7 + register save area
2084
2085 // Save space for one monitor to get into the interpreted method in case
2086 // the method is synchronized
2087 int monitor_size = method->is_synchronized() ?
2088 1*frame::interpreter_frame_monitor_size() : 0;
2089 return size_activation_helper(method->max_locals(), method->max_stack(),
2090 monitor_size) + call_stub_size;
2091 }
2092
2093 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2094 frame* caller,
2095 frame* current,
2096 Method* method,
2097 intptr_t* locals,
2098 intptr_t* stack,
2099 intptr_t* stack_base,
2100 intptr_t* monitor_base,
2101 intptr_t* frame_bottom,
2102 bool is_top_frame
2103 )
2104 {
2105 // What about any vtable?
2106 //
2107 to_fill->_thread = JavaThread::current();
2108 // This gets filled in later but make it something recognizable for now
2109 to_fill->_bcp = method->code_base();
2110 to_fill->_locals = locals;
2111 to_fill->_constants = method->constants()->cache();
2112 to_fill->_method = method;
2113 to_fill->_mdx = NULL;
2114 to_fill->_stack = stack;
2115 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2116 to_fill->_msg = deopt_resume2;
2117 } else {
2118 to_fill->_msg = method_resume;
2119 }
2120 to_fill->_result._to_call._bcp_advance = 0;
2121 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2122 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2123 to_fill->_prev_link = NULL;
2124
2125 // Fill in the registers for the frame
2126
2127 // Need to install _sender_sp. Actually not too hard in C++!
2128 // When the skeletal frames are layed out we fill in a value
2129 // for _sender_sp. That value is only correct for the oldest
2130 // skeletal frame constructed (because there is only a single
2131 // entry for "caller_adjustment". While the skeletal frames
2132 // exist that is good enough. We correct that calculation
2133 // here and get all the frames correct.
2134
2135 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2136
2137 *current->register_addr(Lstate) = (intptr_t) to_fill;
2138 // skeletal already places a useful value here and this doesn't account
2139 // for alignment so don't bother.
2140 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2141
2142 if (caller->is_interpreted_frame()) {
2143 interpreterState prev = caller->get_interpreterState();
2144 to_fill->_prev_link = prev;
2145 // Make the prev callee look proper
2146 prev->_result._to_call._callee = method;
2147 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2148 prev->_result._to_call._bcp_advance = 5;
2149 } else {
2150 prev->_result._to_call._bcp_advance = 3;
2151 }
2152 }
2153 to_fill->_oop_temp = NULL;
2154 to_fill->_stack_base = stack_base;
2155 // Need +1 here because stack_base points to the word just above the first expr stack entry
2156 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2157 // See generate_compute_interpreter_state.
2158 int extra_stack = 0; //6815692//Method::extra_stack_entries();
2159 to_fill->_stack_limit = stack_base - (method->max_stack() + 1 + extra_stack);
2160 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2161
2162 // sparc specific
2163 to_fill->_frame_bottom = frame_bottom;
2164 to_fill->_self_link = to_fill;
2165 #ifdef ASSERT
2166 to_fill->_native_fresult = 123456.789;
2167 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2168 #endif
2169 }
2170
2171 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2172 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2173 }
2174
2175
2176 int AbstractInterpreter::layout_activation(Method* method,
2177 int tempcount, // Number of slots on java expression stack in use
2178 int popframe_extra_args,
2179 int moncount, // Number of active monitors
2180 int caller_actual_parameters,
2181 int callee_param_size,
2182 int callee_locals_size,
2183 frame* caller,
2184 frame* interpreter_frame,
2185 bool is_top_frame) {
2186
2187 assert(popframe_extra_args == 0, "NEED TO FIX");
2188 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2189 // does as far as allocating an interpreter frame.
2190 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2191 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2192 // as determined by a previous call to this method.
2193 // It is also guaranteed to be walkable even though it is in a skeletal state
2194 // NOTE: return size is in words not bytes
2195 // NOTE: tempcount is the current size of the java expression stack. For top most
2196 // frames we will allocate a full sized expression stack and not the curback
2197 // version that non-top frames have.
2198
2199 // Calculate the amount our frame will be adjust by the callee. For top frame
2200 // this is zero.
2201
2202 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2203 // calculates the extra locals based on itself. Not what the callee does
2204 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2205 // as getting sender_sp correct.
2206
2207 int extra_locals_size = callee_locals_size - callee_param_size;
2208 int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2209 int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2210 int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2211 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2212
2213
2214 /*
2215 if we actually have a frame to layout we must now fill in all the pieces. This means both
2216 the interpreterState and the registers.
2217 */
2218 if (interpreter_frame != NULL) {
2219
2220 // MUCHO HACK
2221
2222 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2223 // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2224 assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2225 frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2226
2227 /* Now fillin the interpreterState object */
2228
2229 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2230
2231
2232 intptr_t* locals;
2233
2234 // Calculate the postion of locals[0]. This is painful because of
2235 // stack alignment (same as ia64). The problem is that we can
2236 // not compute the location of locals from fp(). fp() will account
2237 // for the extra locals but it also accounts for aligning the stack
2238 // and we can't determine if the locals[0] was misaligned but max_locals
2239 // was enough to have the
2240 // calculate postion of locals. fp already accounts for extra locals.
2241 // +2 for the static long no_params() issue.
2242
2243 if (caller->is_interpreted_frame()) {
2244 // locals must agree with the caller because it will be used to set the
2245 // caller's tos when we return.
2246 interpreterState prev = caller->get_interpreterState();
2247 // stack() is prepushed.
2248 locals = prev->stack() + method->size_of_parameters();
2249 } else {
2250 // Lay out locals block in the caller adjacent to the register window save area.
2251 //
2252 // Compiled frames do not allocate a varargs area which is why this if
2253 // statement is needed.
2254 //
2255 intptr_t* fp = interpreter_frame->fp();
2256 int local_words = method->max_locals() * Interpreter::stackElementWords;
2257
2258 if (caller->is_compiled_frame()) {
2259 locals = fp + frame::register_save_words + local_words - 1;
2260 } else {
2261 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2262 }
2263
2264 }
2265 // END MUCHO HACK
2266
2267 intptr_t* monitor_base = (intptr_t*) cur_state;
2268 intptr_t* stack_base = monitor_base - monitor_size;
2269 /* +1 because stack is always prepushed */
2270 intptr_t* stack = stack_base - (tempcount + 1);
2271
2272
2273 BytecodeInterpreter::layout_interpreterState(cur_state,
2274 caller,
2275 interpreter_frame,
2276 method,
2277 locals,
2278 stack,
2279 stack_base,
2280 monitor_base,
2281 frame_bottom,
2282 is_top_frame);
2283
2284 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2285
2286 }
2287 return frame_words;
2288 }
2289
2290 #endif // CC_INTERP