Wed, 15 May 2013 11:05:09 +0200
6843347: Boundary values in some public GC options cause crashes
Summary: Setting some public integer options to specific values causes crashes or undefined GC behavior. This patchset adds the necessary argument checking for these options.
Reviewed-by: jmasa, brutisso
1 /*
2 * Copyright (c) 2007, 2013, 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 */
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "interpreter/bytecodeHistogram.hpp"
28 #include "interpreter/cppInterpreter.hpp"
29 #include "interpreter/interpreter.hpp"
30 #include "interpreter/interpreterGenerator.hpp"
31 #include "interpreter/interpreterRuntime.hpp"
32 #include "oops/arrayOop.hpp"
33 #include "oops/methodData.hpp"
34 #include "oops/method.hpp"
35 #include "oops/oop.inline.hpp"
36 #include "prims/jvmtiExport.hpp"
37 #include "prims/jvmtiThreadState.hpp"
38 #include "runtime/arguments.hpp"
39 #include "runtime/deoptimization.hpp"
40 #include "runtime/frame.inline.hpp"
41 #include "runtime/interfaceSupport.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "runtime/stubRoutines.hpp"
44 #include "runtime/synchronizer.hpp"
45 #include "runtime/timer.hpp"
46 #include "runtime/vframeArray.hpp"
47 #include "utilities/debug.hpp"
48 #include "utilities/macros.hpp"
49 #ifdef SHARK
50 #include "shark/shark_globals.hpp"
51 #endif
53 #ifdef CC_INTERP
55 // Routine exists to make tracebacks look decent in debugger
56 // while "shadow" interpreter frames are on stack. It is also
57 // used to distinguish interpreter frames.
59 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
60 ShouldNotReachHere();
61 }
63 bool CppInterpreter::contains(address pc) {
64 return ( _code->contains(pc) ||
65 ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
66 }
68 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
69 #define __ _masm->
71 Label frame_manager_entry;
72 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized
73 // c++ interpreter entry point this holds that entry point label.
75 static address unctrap_frame_manager_entry = NULL;
77 static address interpreter_return_address = NULL;
78 static address deopt_frame_manager_return_atos = NULL;
79 static address deopt_frame_manager_return_btos = NULL;
80 static address deopt_frame_manager_return_itos = NULL;
81 static address deopt_frame_manager_return_ltos = NULL;
82 static address deopt_frame_manager_return_ftos = NULL;
83 static address deopt_frame_manager_return_dtos = NULL;
84 static address deopt_frame_manager_return_vtos = NULL;
86 const Register prevState = G1_scratch;
88 void InterpreterGenerator::save_native_result(void) {
89 // result potentially in O0/O1: save it across calls
90 __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
91 #ifdef _LP64
92 __ stx(O0, STATE(_native_lresult));
93 #else
94 __ std(O0, STATE(_native_lresult));
95 #endif
96 }
98 void InterpreterGenerator::restore_native_result(void) {
100 // Restore any method result value
101 __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
102 #ifdef _LP64
103 __ ldx(STATE(_native_lresult), O0);
104 #else
105 __ ldd(STATE(_native_lresult), O0);
106 #endif
107 }
109 // A result handler converts/unboxes a native call result into
110 // a java interpreter/compiler result. The current frame is an
111 // interpreter frame. The activation frame unwind code must be
112 // consistent with that of TemplateTable::_return(...). In the
113 // case of native methods, the caller's SP was not modified.
114 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
115 address entry = __ pc();
116 Register Itos_i = Otos_i ->after_save();
117 Register Itos_l = Otos_l ->after_save();
118 Register Itos_l1 = Otos_l1->after_save();
119 Register Itos_l2 = Otos_l2->after_save();
120 switch (type) {
121 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
122 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
123 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
124 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
125 case T_LONG :
126 #ifndef _LP64
127 __ mov(O1, Itos_l2); // move other half of long
128 #endif // ifdef or no ifdef, fall through to the T_INT case
129 case T_INT : __ mov(O0, Itos_i); break;
130 case T_VOID : /* nothing to do */ break;
131 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
132 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
133 case T_OBJECT :
134 __ ld_ptr(STATE(_oop_temp), Itos_i);
135 __ verify_oop(Itos_i);
136 break;
137 default : ShouldNotReachHere();
138 }
139 __ ret(); // return from interpreter activation
140 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
141 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
142 return entry;
143 }
145 // tosca based result to c++ interpreter stack based result.
146 // Result goes to address in L1_scratch
148 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
149 // A result is in the native abi result register from a native method call.
150 // We need to return this result to the interpreter by pushing the result on the interpreter's
151 // stack. This is relatively simple the destination is in L1_scratch
152 // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
153 // adjust L1_scratch
154 address entry = __ pc();
155 switch (type) {
156 case T_BOOLEAN:
157 // !0 => true; 0 => false
158 __ subcc(G0, O0, G0);
159 __ addc(G0, 0, O0);
160 __ st(O0, L1_scratch, 0);
161 __ sub(L1_scratch, wordSize, L1_scratch);
162 break;
164 // cannot use and3, 0xFFFF too big as immediate value!
165 case T_CHAR :
166 __ sll(O0, 16, O0);
167 __ srl(O0, 16, O0);
168 __ st(O0, L1_scratch, 0);
169 __ sub(L1_scratch, wordSize, L1_scratch);
170 break;
172 case T_BYTE :
173 __ sll(O0, 24, O0);
174 __ sra(O0, 24, O0);
175 __ st(O0, L1_scratch, 0);
176 __ sub(L1_scratch, wordSize, L1_scratch);
177 break;
179 case T_SHORT :
180 __ sll(O0, 16, O0);
181 __ sra(O0, 16, O0);
182 __ st(O0, L1_scratch, 0);
183 __ sub(L1_scratch, wordSize, L1_scratch);
184 break;
185 case T_LONG :
186 #ifndef _LP64
187 #if defined(COMPILER2)
188 // All return values are where we want them, except for Longs. C2 returns
189 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
190 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
191 // build even if we are returning from interpreted we just do a little
192 // stupid shuffing.
193 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
194 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
195 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
196 __ stx(G1, L1_scratch, -wordSize);
197 #else
198 // native result is in O0, O1
199 __ st(O1, L1_scratch, 0); // Low order
200 __ st(O0, L1_scratch, -wordSize); // High order
201 #endif /* COMPILER2 */
202 #else
203 __ stx(O0, L1_scratch, -wordSize);
204 #endif
205 __ sub(L1_scratch, 2*wordSize, L1_scratch);
206 break;
208 case T_INT :
209 __ st(O0, L1_scratch, 0);
210 __ sub(L1_scratch, wordSize, L1_scratch);
211 break;
213 case T_VOID : /* nothing to do */
214 break;
216 case T_FLOAT :
217 __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
218 __ sub(L1_scratch, wordSize, L1_scratch);
219 break;
221 case T_DOUBLE :
222 // Every stack slot is aligned on 64 bit, However is this
223 // the correct stack slot on 64bit?? QQQ
224 __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
225 __ sub(L1_scratch, 2*wordSize, L1_scratch);
226 break;
227 case T_OBJECT :
228 __ verify_oop(O0);
229 __ st_ptr(O0, L1_scratch, 0);
230 __ sub(L1_scratch, wordSize, L1_scratch);
231 break;
232 default : ShouldNotReachHere();
233 }
234 __ retl(); // return from interpreter activation
235 __ delayed()->nop(); // schedule this better
236 NOT_PRODUCT(__ emit_int32(0);) // marker for disassembly
237 return entry;
238 }
240 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
241 // A result is in the java expression stack of the interpreted method that has just
242 // returned. Place this result on the java expression stack of the caller.
243 //
244 // The current interpreter activation in Lstate is for the method just returning its
245 // result. So we know that the result of this method is on the top of the current
246 // execution stack (which is pre-pushed) and will be return to the top of the caller
247 // stack. The top of the callers stack is the bottom of the locals of the current
248 // activation.
249 // Because of the way activation are managed by the frame manager the value of esp is
250 // below both the stack top of the current activation and naturally the stack top
251 // of the calling activation. This enable this routine to leave the return address
252 // to the frame manager on the stack and do a vanilla return.
253 //
254 // On entry: O0 - points to source (callee stack top)
255 // O1 - points to destination (caller stack top [i.e. free location])
256 // destroys O2, O3
257 //
259 address entry = __ pc();
260 switch (type) {
261 case T_VOID: break;
262 break;
263 case T_FLOAT :
264 case T_BOOLEAN:
265 case T_CHAR :
266 case T_BYTE :
267 case T_SHORT :
268 case T_INT :
269 // 1 word result
270 __ ld(O0, 0, O2);
271 __ st(O2, O1, 0);
272 __ sub(O1, wordSize, O1);
273 break;
274 case T_DOUBLE :
275 case T_LONG :
276 // return top two words on current expression stack to caller's expression stack
277 // The caller's expression stack is adjacent to the current frame manager's intepretState
278 // except we allocated one extra word for this intepretState so we won't overwrite it
279 // when we return a two word result.
280 #ifdef _LP64
281 __ ld_ptr(O0, 0, O2);
282 __ st_ptr(O2, O1, -wordSize);
283 #else
284 __ ld(O0, 0, O2);
285 __ ld(O0, wordSize, O3);
286 __ st(O3, O1, 0);
287 __ st(O2, O1, -wordSize);
288 #endif
289 __ sub(O1, 2*wordSize, O1);
290 break;
291 case T_OBJECT :
292 __ ld_ptr(O0, 0, O2);
293 __ verify_oop(O2); // verify it
294 __ st_ptr(O2, O1, 0);
295 __ sub(O1, wordSize, O1);
296 break;
297 default : ShouldNotReachHere();
298 }
299 __ retl();
300 __ delayed()->nop(); // QQ schedule this better
301 return entry;
302 }
304 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
305 // A result is in the java expression stack of the interpreted method that has just
306 // returned. Place this result in the native abi that the caller expects.
307 // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
308 //
309 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
310 // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
311 // and so rather than return result onto caller's java expression stack we return the
312 // result in the expected location based on the native abi.
313 // On entry: O0 - source (stack top)
314 // On exit result in expected output register
315 // QQQ schedule this better
317 address entry = __ pc();
318 switch (type) {
319 case T_VOID: break;
320 break;
321 case T_FLOAT :
322 __ ldf(FloatRegisterImpl::S, O0, 0, F0);
323 break;
324 case T_BOOLEAN:
325 case T_CHAR :
326 case T_BYTE :
327 case T_SHORT :
328 case T_INT :
329 // 1 word result
330 __ ld(O0, 0, O0->after_save());
331 break;
332 case T_DOUBLE :
333 __ ldf(FloatRegisterImpl::D, O0, 0, F0);
334 break;
335 case T_LONG :
336 // return top two words on current expression stack to caller's expression stack
337 // The caller's expression stack is adjacent to the current frame manager's interpretState
338 // except we allocated one extra word for this intepretState so we won't overwrite it
339 // when we return a two word result.
340 #ifdef _LP64
341 __ ld_ptr(O0, 0, O0->after_save());
342 #else
343 __ ld(O0, wordSize, O1->after_save());
344 __ ld(O0, 0, O0->after_save());
345 #endif
346 #if defined(COMPILER2) && !defined(_LP64)
347 // C2 expects long results in G1 we can't tell if we're returning to interpreted
348 // or compiled so just be safe use G1 and O0/O1
350 // Shift bits into high (msb) of G1
351 __ sllx(Otos_l1->after_save(), 32, G1);
352 // Zero extend low bits
353 __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
354 __ or3 (Otos_l2->after_save(), G1, G1);
355 #endif /* COMPILER2 */
356 break;
357 case T_OBJECT :
358 __ ld_ptr(O0, 0, O0->after_save());
359 __ verify_oop(O0->after_save()); // verify it
360 break;
361 default : ShouldNotReachHere();
362 }
363 __ retl();
364 __ delayed()->nop();
365 return entry;
366 }
368 address CppInterpreter::return_entry(TosState state, int length) {
369 // make it look good in the debugger
370 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
371 }
373 address CppInterpreter::deopt_entry(TosState state, int length) {
374 address ret = NULL;
375 if (length != 0) {
376 switch (state) {
377 case atos: ret = deopt_frame_manager_return_atos; break;
378 case btos: ret = deopt_frame_manager_return_btos; break;
379 case ctos:
380 case stos:
381 case itos: ret = deopt_frame_manager_return_itos; break;
382 case ltos: ret = deopt_frame_manager_return_ltos; break;
383 case ftos: ret = deopt_frame_manager_return_ftos; break;
384 case dtos: ret = deopt_frame_manager_return_dtos; break;
385 case vtos: ret = deopt_frame_manager_return_vtos; break;
386 }
387 } else {
388 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap)
389 }
390 assert(ret != NULL, "Not initialized");
391 return ret;
392 }
394 //
395 // Helpers for commoning out cases in the various type of method entries.
396 //
398 // increment invocation count & check for overflow
399 //
400 // Note: checking for negative value instead of overflow
401 // so we have a 'sticky' overflow test
402 //
403 // Lmethod: method
404 // ??: invocation counter
405 //
406 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
407 Label done;
408 const Register Rcounters = G3_scratch;
410 __ ld_ptr(STATE(_method), G5_method);
411 __ get_method_counters(G5_method, Rcounters, done);
413 // Update standard invocation counters
414 __ increment_invocation_counter(Rcounters, O0, G4_scratch);
415 if (ProfileInterpreter) {
416 Address interpreter_invocation_counter(Rcounters, 0,
417 in_bytes(MethodCounters::interpreter_invocation_counter_offset()));
418 __ ld(interpreter_invocation_counter, G4_scratch);
419 __ inc(G4_scratch);
420 __ st(G4_scratch, interpreter_invocation_counter);
421 }
423 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
424 __ sethi(invocation_limit);
425 __ ld(invocation_limit, G3_scratch);
426 __ cmp(O0, G3_scratch);
427 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
428 __ delayed()->nop();
429 __ bind(done);
430 }
432 address InterpreterGenerator::generate_empty_entry(void) {
434 // A method that does nothing but return...
436 address entry = __ pc();
437 Label slow_path;
439 // do nothing for empty methods (do not even increment invocation counter)
440 if ( UseFastEmptyMethods) {
441 // If we need a safepoint check, generate full interpreter entry.
442 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
443 __ load_contents(sync_state, G3_scratch);
444 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
445 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
446 __ delayed()->nop();
448 // Code: _return
449 __ retl();
450 __ delayed()->mov(O5_savedSP, SP);
451 return entry;
452 }
453 return NULL;
454 }
456 // Call an accessor method (assuming it is resolved, otherwise drop into
457 // vanilla (slow path) entry
459 // Generates code to elide accessor methods
460 // Uses G3_scratch and G1_scratch as scratch
461 address InterpreterGenerator::generate_accessor_entry(void) {
463 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
464 // parameter size = 1
465 // Note: We can only use this code if the getfield has been resolved
466 // and if we don't have a null-pointer exception => check for
467 // these conditions first and use slow path if necessary.
468 address entry = __ pc();
469 Label slow_path;
471 if ( UseFastAccessorMethods) {
472 // Check if we need to reach a safepoint and generate full interpreter
473 // frame if so.
474 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
475 __ load_contents(sync_state, G3_scratch);
476 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
477 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
478 __ delayed()->nop();
480 // Check if local 0 != NULL
481 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
482 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
483 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
484 __ delayed()->nop();
487 // read first instruction word and extract bytecode @ 1 and index @ 2
488 // get first 4 bytes of the bytecodes (big endian!)
489 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::const_offset())), G1_scratch);
490 __ ld(Address(G1_scratch, 0, in_bytes(ConstMethod::codes_offset())), G1_scratch);
492 // move index @ 2 far left then to the right most two bytes.
493 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
494 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
495 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
497 // get constant pool cache
498 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), G3_scratch);
499 __ ld_ptr(G3_scratch, in_bytes(ConstMethod::constants_offset()), G3_scratch);
500 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
502 // get specific constant pool cache entry
503 __ add(G3_scratch, G1_scratch, G3_scratch);
505 // Check the constant Pool cache entry to see if it has been resolved.
506 // If not, need the slow path.
507 ByteSize cp_base_offset = ConstantPoolCache::base_offset();
508 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
509 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
510 __ and3(G1_scratch, 0xFF, G1_scratch);
511 __ cmp(G1_scratch, Bytecodes::_getfield);
512 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
513 __ delayed()->nop();
515 // Get the type and return field offset from the constant pool cache
516 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
517 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
519 Label xreturn_path;
520 // Need to differentiate between igetfield, agetfield, bgetfield etc.
521 // because they are different sizes.
522 // Get the type from the constant pool cache
523 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
524 // Make sure we don't need to mask G1_scratch after the above shift
525 ConstantPoolCacheEntry::verify_tos_state_shift();
526 __ cmp(G1_scratch, atos );
527 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
528 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
529 __ cmp(G1_scratch, itos);
530 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
531 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
532 __ cmp(G1_scratch, stos);
533 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
534 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
535 __ cmp(G1_scratch, ctos);
536 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
537 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
538 #ifdef ASSERT
539 __ cmp(G1_scratch, btos);
540 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
541 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
542 __ should_not_reach_here();
543 #endif
544 __ ldsb(Otos_i, G3_scratch, Otos_i);
545 __ bind(xreturn_path);
547 // _ireturn/_areturn
548 __ retl(); // return from leaf routine
549 __ delayed()->mov(O5_savedSP, SP);
551 // Generate regular method entry
552 __ bind(slow_path);
553 __ ba(fast_accessor_slow_entry_path);
554 __ delayed()->nop();
555 return entry;
556 }
557 return NULL;
558 }
560 address InterpreterGenerator::generate_Reference_get_entry(void) {
561 #if INCLUDE_ALL_GCS
562 if (UseG1GC) {
563 // We need to generate have a routine that generates code to:
564 // * load the value in the referent field
565 // * passes that value to the pre-barrier.
566 //
567 // In the case of G1 this will record the value of the
568 // referent in an SATB buffer if marking is active.
569 // This will cause concurrent marking to mark the referent
570 // field as live.
571 Unimplemented();
572 }
573 #endif // INCLUDE_ALL_GCS
575 // If G1 is not enabled then attempt to go through the accessor entry point
576 // Reference.get is an accessor
577 return generate_accessor_entry();
578 }
580 //
581 // Interpreter stub for calling a native method. (C++ interpreter)
582 // This sets up a somewhat different looking stack for calling the native method
583 // than the typical interpreter frame setup.
584 //
586 address InterpreterGenerator::generate_native_entry(bool synchronized) {
587 address entry = __ pc();
589 // the following temporary registers are used during frame creation
590 const Register Gtmp1 = G3_scratch ;
591 const Register Gtmp2 = G1_scratch;
592 const Register RconstMethod = Gtmp1;
593 const Address constMethod(G5_method, 0, in_bytes(Method::const_offset()));
594 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
596 bool inc_counter = UseCompiler || CountCompiledCalls;
598 // make sure registers are different!
599 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
601 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
603 Label Lentry;
604 __ bind(Lentry);
606 const Register Glocals_size = G3;
607 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
609 // make sure method is native & not abstract
610 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
611 #ifdef ASSERT
612 __ ld(access_flags, Gtmp1);
613 {
614 Label L;
615 __ btst(JVM_ACC_NATIVE, Gtmp1);
616 __ br(Assembler::notZero, false, Assembler::pt, L);
617 __ delayed()->nop();
618 __ stop("tried to execute non-native method as native");
619 __ bind(L);
620 }
621 { Label L;
622 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
623 __ br(Assembler::zero, false, Assembler::pt, L);
624 __ delayed()->nop();
625 __ stop("tried to execute abstract method as non-abstract");
626 __ bind(L);
627 }
628 #endif // ASSERT
630 __ ld_ptr(constMethod, RconstMethod);
631 __ lduh(size_of_parameters, Gtmp1);
632 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes
633 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord
634 // NEW
635 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
636 // generate the code to allocate the interpreter stack frame
637 // NEW FRAME ALLOCATED HERE
638 // save callers original sp
639 // __ mov(SP, I5_savedSP->after_restore());
641 generate_compute_interpreter_state(Lstate, G0, true);
643 // At this point Lstate points to new interpreter state
644 //
646 const Address do_not_unlock_if_synchronized(G2_thread, 0,
647 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
648 // Since at this point in the method invocation the exception handler
649 // would try to exit the monitor of synchronized methods which hasn't
650 // been entered yet, we set the thread local variable
651 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
652 // runtime, exception handling i.e. unlock_if_synchronized_method will
653 // check this thread local flag.
654 // This flag has two effects, one is to force an unwind in the topmost
655 // interpreter frame and not perform an unlock while doing so.
657 __ movbool(true, G3_scratch);
658 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
661 // increment invocation counter and check for overflow
662 //
663 // Note: checking for negative value instead of overflow
664 // so we have a 'sticky' overflow test (may be of
665 // importance as soon as we have true MT/MP)
666 Label invocation_counter_overflow;
667 if (inc_counter) {
668 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
669 }
670 Label Lcontinue;
671 __ bind(Lcontinue);
673 bang_stack_shadow_pages(true);
674 // reset the _do_not_unlock_if_synchronized flag
675 __ stbool(G0, do_not_unlock_if_synchronized);
677 // check for synchronized methods
678 // Must happen AFTER invocation_counter check, so method is not locked
679 // if counter overflows.
681 if (synchronized) {
682 lock_method();
683 // Don't see how G2_thread is preserved here...
684 // __ verify_thread(); QQQ destroys L0,L1 can't use
685 } else {
686 #ifdef ASSERT
687 { Label ok;
688 __ ld_ptr(STATE(_method), G5_method);
689 __ ld(access_flags, O0);
690 __ btst(JVM_ACC_SYNCHRONIZED, O0);
691 __ br( Assembler::zero, false, Assembler::pt, ok);
692 __ delayed()->nop();
693 __ stop("method needs synchronization");
694 __ bind(ok);
695 }
696 #endif // ASSERT
697 }
699 // start execution
701 // __ verify_thread(); kills L1,L2 can't use at the moment
703 // jvmti/jvmpi support
704 __ notify_method_entry();
706 // native call
708 // (note that O0 is never an oop--at most it is a handle)
709 // It is important not to smash any handles created by this call,
710 // until any oop handle in O0 is dereferenced.
712 // (note that the space for outgoing params is preallocated)
714 // get signature handler
716 Label pending_exception_present;
718 { Label L;
719 __ ld_ptr(STATE(_method), G5_method);
720 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
721 __ tst(G3_scratch);
722 __ brx(Assembler::notZero, false, Assembler::pt, L);
723 __ delayed()->nop();
724 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
725 __ ld_ptr(STATE(_method), G5_method);
727 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
728 __ ld_ptr(exception_addr, G3_scratch);
729 __ br_notnull_short(G3_scratch, Assembler::pn, pending_exception_present);
730 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
731 __ bind(L);
732 }
734 // Push a new frame so that the args will really be stored in
735 // Copy a few locals across so the new frame has the variables
736 // we need but these values will be dead at the jni call and
737 // therefore not gc volatile like the values in the current
738 // frame (Lstate in particular)
740 // Flush the state pointer to the register save area
741 // Which is the only register we need for a stack walk.
742 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
744 __ mov(Lstate, O1); // Need to pass the state pointer across the frame
746 // Calculate current frame size
747 __ sub(SP, FP, O3); // Calculate negative of current frame size
748 __ save(SP, O3, SP); // Allocate an identical sized frame
750 __ mov(I1, Lstate); // In the "natural" register.
752 // Note I7 has leftover trash. Slow signature handler will fill it in
753 // should we get there. Normal jni call will set reasonable last_Java_pc
754 // below (and fix I7 so the stack trace doesn't have a meaningless frame
755 // in it).
758 // call signature handler
759 __ ld_ptr(STATE(_method), Lmethod);
760 __ ld_ptr(STATE(_locals), Llocals);
762 __ callr(G3_scratch, 0);
763 __ delayed()->nop();
764 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed)
766 { Label not_static;
768 __ ld_ptr(STATE(_method), G5_method);
769 __ ld(access_flags, O0);
770 __ btst(JVM_ACC_STATIC, O0);
771 __ br( Assembler::zero, false, Assembler::pt, not_static);
772 __ delayed()->
773 // get native function entry point(O0 is a good temp until the very end)
774 ld_ptr(Address(G5_method, 0, in_bytes(Method::native_function_offset())), O0);
775 // for static methods insert the mirror argument
776 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
778 __ ld_ptr(Address(G5_method, 0, in_bytes(Method:: const_offset())), O1);
779 __ ld_ptr(Address(O1, 0, in_bytes(ConstMethod::constants_offset())), O1);
780 __ ld_ptr(Address(O1, 0, ConstantPool::pool_holder_offset_in_bytes()), O1);
781 __ ld_ptr(O1, mirror_offset, O1);
782 // where the mirror handle body is allocated:
783 #ifdef ASSERT
784 if (!PrintSignatureHandlers) // do not dirty the output with this
785 { Label L;
786 __ tst(O1);
787 __ brx(Assembler::notZero, false, Assembler::pt, L);
788 __ delayed()->nop();
789 __ stop("mirror is missing");
790 __ bind(L);
791 }
792 #endif // ASSERT
793 __ st_ptr(O1, STATE(_oop_temp));
794 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add
795 __ bind(not_static);
796 }
798 // At this point, arguments have been copied off of stack into
799 // their JNI positions, which are O1..O5 and SP[68..].
800 // Oops are boxed in-place on the stack, with handles copied to arguments.
801 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
803 #ifdef ASSERT
804 { Label L;
805 __ tst(O0);
806 __ brx(Assembler::notZero, false, Assembler::pt, L);
807 __ delayed()->nop();
808 __ stop("native entry point is missing");
809 __ bind(L);
810 }
811 #endif // ASSERT
813 //
814 // setup the java frame anchor
815 //
816 // The scavenge function only needs to know that the PC of this frame is
817 // in the interpreter method entry code, it doesn't need to know the exact
818 // PC and hence we can use O7 which points to the return address from the
819 // previous call in the code stream (signature handler function)
820 //
821 // The other trick is we set last_Java_sp to FP instead of the usual SP because
822 // we have pushed the extra frame in order to protect the volatile register(s)
823 // in that frame when we return from the jni call
824 //
827 __ set_last_Java_frame(FP, O7);
828 __ mov(O7, I7); // make dummy interpreter frame look like one above,
829 // not meaningless information that'll confuse me.
831 // flush the windows now. We don't care about the current (protection) frame
832 // only the outer frames
834 __ flush_windows();
836 // mark windows as flushed
837 Address flags(G2_thread,
838 0,
839 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
840 __ set(JavaFrameAnchor::flushed, G3_scratch);
841 __ st(G3_scratch, flags);
843 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
845 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
846 #ifdef ASSERT
847 { Label L;
848 __ ld(thread_state, G3_scratch);
849 __ cmp(G3_scratch, _thread_in_Java);
850 __ br(Assembler::equal, false, Assembler::pt, L);
851 __ delayed()->nop();
852 __ stop("Wrong thread state in native stub");
853 __ bind(L);
854 }
855 #endif // ASSERT
856 __ set(_thread_in_native, G3_scratch);
857 __ st(G3_scratch, thread_state);
859 // Call the jni method, using the delay slot to set the JNIEnv* argument.
860 __ callr(O0, 0);
861 __ delayed()->
862 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
863 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
865 // must we block?
867 // Block, if necessary, before resuming in _thread_in_Java state.
868 // In order for GC to work, don't clear the last_Java_sp until after blocking.
869 { Label no_block;
870 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
872 // Switch thread to "native transition" state before reading the synchronization state.
873 // This additional state is necessary because reading and testing the synchronization
874 // state is not atomic w.r.t. GC, as this scenario demonstrates:
875 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
876 // VM thread changes sync state to synchronizing and suspends threads for GC.
877 // Thread A is resumed to finish this native method, but doesn't block here since it
878 // didn't see any synchronization is progress, and escapes.
879 __ set(_thread_in_native_trans, G3_scratch);
880 __ st(G3_scratch, thread_state);
881 if(os::is_MP()) {
882 // Write serialization page so VM thread can do a pseudo remote membar.
883 // We use the current thread pointer to calculate a thread specific
884 // offset to write to within the page. This minimizes bus traffic
885 // due to cache line collision.
886 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
887 }
888 __ load_contents(sync_state, G3_scratch);
889 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
892 Label L;
893 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
894 __ br(Assembler::notEqual, false, Assembler::pn, L);
895 __ delayed()->
896 ld(suspend_state, G3_scratch);
897 __ cmp(G3_scratch, 0);
898 __ br(Assembler::equal, false, Assembler::pt, no_block);
899 __ delayed()->nop();
900 __ bind(L);
902 // Block. Save any potential method result value before the operation and
903 // use a leaf call to leave the last_Java_frame setup undisturbed.
904 save_native_result();
905 __ call_VM_leaf(noreg,
906 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
907 G2_thread);
908 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
909 // Restore any method result value
910 restore_native_result();
911 __ bind(no_block);
912 }
914 // Clear the frame anchor now
916 __ reset_last_Java_frame();
918 // Move the result handler address
919 __ mov(Lscratch, G3_scratch);
920 // return possible result to the outer frame
921 #ifndef __LP64
922 __ mov(O0, I0);
923 __ restore(O1, G0, O1);
924 #else
925 __ restore(O0, G0, O0);
926 #endif /* __LP64 */
928 // Move result handler to expected register
929 __ mov(G3_scratch, Lscratch);
932 // thread state is thread_in_native_trans. Any safepoint blocking has
933 // happened in the trampoline we are ready to switch to thread_in_Java.
935 __ set(_thread_in_Java, G3_scratch);
936 __ st(G3_scratch, thread_state);
938 // If we have an oop result store it where it will be safe for any further gc
939 // until we return now that we've released the handle it might be protected by
941 {
942 Label no_oop, store_result;
944 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
945 __ cmp(G3_scratch, Lscratch);
946 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
947 __ delayed()->nop();
948 __ addcc(G0, O0, O0);
949 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
950 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
951 __ mov(G0, O0);
953 __ bind(store_result);
954 // Store it where gc will look for it and result handler expects it.
955 __ st_ptr(O0, STATE(_oop_temp));
957 __ bind(no_oop);
959 }
961 // reset handle block
962 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
963 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
966 // handle exceptions (exception handling will handle unlocking!)
967 { Label L;
968 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
970 __ ld_ptr(exception_addr, Gtemp);
971 __ tst(Gtemp);
972 __ brx(Assembler::equal, false, Assembler::pt, L);
973 __ delayed()->nop();
974 __ bind(pending_exception_present);
975 // With c++ interpreter we just leave it pending caller will do the correct thing. However...
976 // Like x86 we ignore the result of the native call and leave the method locked. This
977 // seems wrong to leave things locked.
979 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
980 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
982 __ bind(L);
983 }
985 // jvmdi/jvmpi support (preserves thread register)
986 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
988 if (synchronized) {
989 // save and restore any potential method result value around the unlocking operation
990 save_native_result();
992 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
993 // Get the initial monitor we allocated
994 __ sub(Lstate, entry_size, O1); // initial monitor
995 __ unlock_object(O1);
996 restore_native_result();
997 }
999 #if defined(COMPILER2) && !defined(_LP64)
1001 // C2 expects long results in G1 we can't tell if we're returning to interpreted
1002 // or compiled so just be safe.
1004 __ sllx(O0, 32, G1); // Shift bits into high G1
1005 __ srl (O1, 0, O1); // Zero extend O1
1006 __ or3 (O1, G1, G1); // OR 64 bits into G1
1008 #endif /* COMPILER2 && !_LP64 */
1010 #ifdef ASSERT
1011 {
1012 Label ok;
1013 __ cmp(I5_savedSP, FP);
1014 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1015 __ delayed()->nop();
1016 __ stop("bad I5_savedSP value");
1017 __ should_not_reach_here();
1018 __ bind(ok);
1019 }
1020 #endif
1021 // Calls result handler which POPS FRAME
1022 if (TraceJumps) {
1023 // Move target to register that is recordable
1024 __ mov(Lscratch, G3_scratch);
1025 __ JMP(G3_scratch, 0);
1026 } else {
1027 __ jmp(Lscratch, 0);
1028 }
1029 __ delayed()->nop();
1031 if (inc_counter) {
1032 // handle invocation counter overflow
1033 __ bind(invocation_counter_overflow);
1034 generate_counter_overflow(Lcontinue);
1035 }
1038 return entry;
1039 }
1041 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1042 const Register prev_state,
1043 bool native) {
1045 // On entry
1046 // G5_method - caller's method
1047 // Gargs - points to initial parameters (i.e. locals[0])
1048 // G2_thread - valid? (C1 only??)
1049 // "prev_state" - contains any previous frame manager state which we must save a link
1050 //
1051 // On return
1052 // "state" is a pointer to the newly allocated state object. We must allocate and initialize
1053 // a new interpretState object and the method expression stack.
1055 assert_different_registers(state, prev_state);
1056 assert_different_registers(prev_state, G3_scratch);
1057 const Register Gtmp = G3_scratch;
1058 const Address constMethod (G5_method, 0, in_bytes(Method::const_offset()));
1059 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
1061 // slop factor is two extra slots on the expression stack so that
1062 // we always have room to store a result when returning from a call without parameters
1063 // that returns a result.
1065 const int slop_factor = 2*wordSize;
1067 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1068 //6815692//Method::extra_stack_words() + // extra push slots for MH adapters
1069 frame::memory_parameter_word_sp_offset + // register save area + param window
1070 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1072 // XXX G5_method valid
1074 // Now compute new frame size
1076 if (native) {
1077 const Register RconstMethod = Gtmp;
1078 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1079 __ ld_ptr(constMethod, RconstMethod);
1080 __ lduh( size_of_parameters, Gtmp );
1081 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words
1082 } else {
1083 // Full size expression stack
1084 __ ld_ptr(constMethod, Gtmp);
1085 __ lduh(Gtmp, in_bytes(ConstMethod::max_stack_offset()), Gtmp);
1086 }
1087 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1089 __ neg(Gtmp); // negative space for stack/parameters in words
1090 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1091 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1093 // Need to do stack size check here before we fault on large frames
1095 Label stack_ok;
1097 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1098 (StackRedPages+StackYellowPages);
1101 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1102 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1103 // compute stack bottom
1104 __ sub(O0, O1, O0);
1106 // Avoid touching the guard pages
1107 // Also a fudge for frame size of BytecodeInterpreter::run
1108 // It varies from 1k->4k depending on build type
1109 const int fudge = 6 * K;
1111 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1113 __ add(O0, O1, O0);
1114 __ sub(O0, Gtmp, O0);
1115 __ cmp(SP, O0);
1116 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1117 __ delayed()->nop();
1119 // throw exception return address becomes throwing pc
1121 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1122 __ stop("never reached");
1124 __ bind(stack_ok);
1126 __ save(SP, Gtmp, SP); // setup new frame and register window
1128 // New window I7 call_stub or previous activation
1129 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1130 //
1131 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1132 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1134 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1136 // Initialize a new Interpreter state
1137 // orig_sp - caller's original sp
1138 // G2_thread - thread
1139 // Gargs - &locals[0] (unbiased?)
1140 // G5_method - method
1141 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1144 __ set(0xdead0004, O1);
1147 __ st_ptr(Gargs, XXX_STATE(_locals));
1148 __ st_ptr(G0, XXX_STATE(_oop_temp));
1150 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1151 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1152 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1154 if (native) {
1155 __ st_ptr(G0, XXX_STATE(_bcp));
1156 } else {
1157 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), O2); // get ConstMethod*
1158 __ add(O2, in_bytes(ConstMethod::codes_offset()), O2); // get bcp
1159 __ st_ptr(O2, XXX_STATE(_bcp));
1160 }
1162 __ st_ptr(G0, XXX_STATE(_mdx));
1163 __ st_ptr(G5_method, XXX_STATE(_method));
1165 __ set((int) BytecodeInterpreter::method_entry, O1);
1166 __ st(O1, XXX_STATE(_msg));
1168 __ ld_ptr(constMethod, O3);
1169 __ ld_ptr(O3, in_bytes(ConstMethod::constants_offset()), O3);
1170 __ ld_ptr(O3, ConstantPool::cache_offset_in_bytes(), O2);
1171 __ st_ptr(O2, XXX_STATE(_constants));
1173 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1175 // Monitor base is just start of BytecodeInterpreter object;
1176 __ mov(state, O2);
1177 __ st_ptr(O2, XXX_STATE(_monitor_base));
1179 // Do we need a monitor for synchonized method?
1180 {
1181 __ ld(access_flags, O1);
1182 Label done;
1183 Label got_obj;
1184 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1185 __ br( Assembler::zero, false, Assembler::pt, done);
1187 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1188 __ delayed()->btst(JVM_ACC_STATIC, O1);
1189 __ ld_ptr(XXX_STATE(_locals), O1);
1190 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1191 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1192 __ ld_ptr(constMethod, O1);
1193 __ ld_ptr( O1, in_bytes(ConstMethod::constants_offset()), O1);
1194 __ ld_ptr( O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1195 // lock the mirror, not the Klass*
1196 __ ld_ptr( O1, mirror_offset, O1);
1198 __ bind(got_obj);
1200 #ifdef ASSERT
1201 __ tst(O1);
1202 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
1203 #endif // ASSERT
1205 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1206 __ sub(SP, entry_size, SP); // account for initial monitor
1207 __ sub(O2, entry_size, O2); // initial monitor
1208 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1209 __ bind(done);
1210 }
1212 // Remember initial frame bottom
1214 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1216 __ st_ptr(O2, XXX_STATE(_stack_base));
1218 __ sub(O2, wordSize, O2); // prepush
1219 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1221 // Full size expression stack
1222 __ ld_ptr(constMethod, O3);
1223 __ lduh(O3, in_bytes(ConstMethod::max_stack_offset()), O3);
1224 guarantee(!EnableInvokeDynamic, "no support yet for java.lang.invoke.MethodHandle"); //6815692
1225 //6815692//if (EnableInvokeDynamic)
1226 //6815692// __ inc(O3, Method::extra_stack_entries());
1227 __ sll(O3, LogBytesPerWord, O3);
1228 __ sub(O2, O3, O3);
1229 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1230 __ st_ptr(O3, XXX_STATE(_stack_limit));
1232 if (!native) {
1233 //
1234 // Code to initialize locals
1235 //
1236 Register init_value = noreg; // will be G0 if we must clear locals
1237 // Now zero locals
1238 if (true /* zerolocals */ || ClearInterpreterLocals) {
1239 // explicitly initialize locals
1240 init_value = G0;
1241 } else {
1242 #ifdef ASSERT
1243 // initialize locals to a garbage pattern for better debugging
1244 init_value = O3;
1245 __ set( 0x0F0F0F0F, init_value );
1246 #endif // ASSERT
1247 }
1248 if (init_value != noreg) {
1249 Label clear_loop;
1250 const Register RconstMethod = O1;
1251 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1252 const Address size_of_locals (RconstMethod, 0, in_bytes(ConstMethod::size_of_locals_offset()));
1254 // NOTE: If you change the frame layout, this code will need to
1255 // be updated!
1256 __ ld_ptr( constMethod, RconstMethod );
1257 __ lduh( size_of_locals, O2 );
1258 __ lduh( size_of_parameters, O1 );
1259 __ sll( O2, LogBytesPerWord, O2);
1260 __ sll( O1, LogBytesPerWord, O1 );
1261 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1262 __ sub( L2_scratch, O2, O2 );
1263 __ sub( L2_scratch, O1, O1 );
1265 __ bind( clear_loop );
1266 __ inc( O2, wordSize );
1268 __ cmp( O2, O1 );
1269 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1270 __ delayed()->st_ptr( init_value, O2, 0 );
1271 }
1272 }
1273 }
1274 // Find preallocated monitor and lock method (C++ interpreter)
1275 //
1276 void InterpreterGenerator::lock_method(void) {
1277 // Lock the current method.
1278 // Destroys registers L2_scratch, L3_scratch, O0
1279 //
1280 // Find everything relative to Lstate
1282 #ifdef ASSERT
1283 __ ld_ptr(STATE(_method), L2_scratch);
1284 __ ld(L2_scratch, in_bytes(Method::access_flags_offset()), O0);
1286 { Label ok;
1287 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1288 __ br( Assembler::notZero, false, Assembler::pt, ok);
1289 __ delayed()->nop();
1290 __ stop("method doesn't need synchronization");
1291 __ bind(ok);
1292 }
1293 #endif // ASSERT
1295 // monitor is already allocated at stack base
1296 // and the lockee is already present
1297 __ ld_ptr(STATE(_stack_base), L2_scratch);
1298 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1299 __ lock_object(L2_scratch, O0);
1301 }
1303 // Generate code for handling resuming a deopted method
1304 void CppInterpreterGenerator::generate_deopt_handling() {
1306 Label return_from_deopt_common;
1308 // deopt needs to jump to here to enter the interpreter (return a result)
1309 deopt_frame_manager_return_atos = __ pc();
1311 // O0/O1 live
1312 __ ba(return_from_deopt_common);
1313 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1316 // deopt needs to jump to here to enter the interpreter (return a result)
1317 deopt_frame_manager_return_btos = __ pc();
1319 // O0/O1 live
1320 __ ba(return_from_deopt_common);
1321 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1323 // deopt needs to jump to here to enter the interpreter (return a result)
1324 deopt_frame_manager_return_itos = __ pc();
1326 // O0/O1 live
1327 __ ba(return_from_deopt_common);
1328 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1330 // deopt needs to jump to here to enter the interpreter (return a result)
1332 deopt_frame_manager_return_ltos = __ pc();
1333 #if !defined(_LP64) && defined(COMPILER2)
1334 // All return values are where we want them, except for Longs. C2 returns
1335 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1336 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1337 // build even if we are returning from interpreted we just do a little
1338 // stupid shuffing.
1339 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1340 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1341 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1343 __ srl (G1, 0,O1);
1344 __ srlx(G1,32,O0);
1345 #endif /* !_LP64 && COMPILER2 */
1346 // O0/O1 live
1347 __ ba(return_from_deopt_common);
1348 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1350 // deopt needs to jump to here to enter the interpreter (return a result)
1352 deopt_frame_manager_return_ftos = __ pc();
1353 // O0/O1 live
1354 __ ba(return_from_deopt_common);
1355 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1357 // deopt needs to jump to here to enter the interpreter (return a result)
1358 deopt_frame_manager_return_dtos = __ pc();
1360 // O0/O1 live
1361 __ ba(return_from_deopt_common);
1362 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1364 // deopt needs to jump to here to enter the interpreter (return a result)
1365 deopt_frame_manager_return_vtos = __ pc();
1367 // O0/O1 live
1368 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1370 // Deopt return common
1371 // an index is present that lets us move any possible result being
1372 // return to the interpreter's stack
1373 //
1374 __ bind(return_from_deopt_common);
1376 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1377 // stack is in the state that the calling convention left it.
1378 // Copy the result from native abi result and place it on java expression stack.
1380 // Current interpreter state is present in Lstate
1382 // Get current pre-pushed top of interpreter stack
1383 // Any result (if any) is in native abi
1384 // result type index is in L3_scratch
1386 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1388 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1389 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1390 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1391 __ jmpl(Lscratch, G0, O7); // and convert it
1392 __ delayed()->nop();
1394 // L1_scratch points to top of stack (prepushed)
1395 __ st_ptr(L1_scratch, STATE(_stack));
1396 }
1398 // Generate the code to handle a more_monitors message from the c++ interpreter
1399 void CppInterpreterGenerator::generate_more_monitors() {
1401 Label entry, loop;
1402 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1403 // 1. compute new pointers // esp: old expression stack top
1404 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1405 __ sub(L4_scratch, entry_size, L4_scratch);
1406 __ st_ptr(L4_scratch, STATE(_stack_base));
1408 __ sub(SP, entry_size, SP); // Grow stack
1409 __ st_ptr(SP, STATE(_frame_bottom));
1411 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1412 __ sub(L2_scratch, entry_size, L2_scratch);
1413 __ st_ptr(L2_scratch, STATE(_stack_limit));
1415 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1416 __ sub(L1_scratch, entry_size, L1_scratch);
1417 __ st_ptr(L1_scratch, STATE(_stack));
1418 __ ba(entry);
1419 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1421 // 2. move expression stack
1423 __ bind(loop);
1424 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1425 __ add(L1_scratch, wordSize, L1_scratch);
1426 __ bind(entry);
1427 __ cmp(L1_scratch, L4_scratch);
1428 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1429 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1431 // now zero the slot so we can find it.
1432 __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1434 }
1436 // Initial entry to C++ interpreter from the call_stub.
1437 // This entry point is called the frame manager since it handles the generation
1438 // of interpreter activation frames via requests directly from the vm (via call_stub)
1439 // and via requests from the interpreter. The requests from the call_stub happen
1440 // directly thru the entry point. Requests from the interpreter happen via returning
1441 // from the interpreter and examining the message the interpreter has returned to
1442 // the frame manager. The frame manager can take the following requests:
1444 // NO_REQUEST - error, should never happen.
1445 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1446 // allocate a new monitor.
1447 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1448 // happens during entry during the entry via the call stub.
1449 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1450 //
1451 // Arguments:
1452 //
1453 // ebx: Method*
1454 // ecx: receiver - unused (retrieved from stack as needed)
1455 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1456 //
1457 //
1458 // Stack layout at entry
1459 //
1460 // [ return address ] <--- esp
1461 // [ parameter n ]
1462 // ...
1463 // [ parameter 1 ]
1464 // [ expression stack ]
1465 //
1466 //
1467 // We are free to blow any registers we like because the call_stub which brought us here
1468 // initially has preserved the callee save registers already.
1469 //
1470 //
1472 static address interpreter_frame_manager = NULL;
1474 #ifdef ASSERT
1475 #define VALIDATE_STATE(scratch, marker) \
1476 { \
1477 Label skip; \
1478 __ ld_ptr(STATE(_self_link), scratch); \
1479 __ cmp(Lstate, scratch); \
1480 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1481 __ delayed()->nop(); \
1482 __ breakpoint_trap(); \
1483 __ emit_int32(marker); \
1484 __ bind(skip); \
1485 }
1486 #else
1487 #define VALIDATE_STATE(scratch, marker)
1488 #endif /* ASSERT */
1490 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1491 //
1492 // Adjust caller's stack so that all the locals can be contiguous with
1493 // the parameters.
1494 // Worries about stack overflow make this a pain.
1495 //
1496 // Destroys args, G3_scratch, G3_scratch
1497 // In/Out O5_savedSP (sender's original SP)
1498 //
1499 // assert_different_registers(state, prev_state);
1500 const Register Gtmp = G3_scratch;
1501 const RconstMethod = G3_scratch;
1502 const Register tmp = O2;
1503 const Address constMethod(G5_method, 0, in_bytes(Method::const_offset()));
1504 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1505 const Address size_of_locals (RconstMethod, 0, in_bytes(ConstMethod::size_of_locals_offset()));
1507 __ ld_ptr(constMethod, RconstMethod);
1508 __ lduh(size_of_parameters, tmp);
1509 __ sll(tmp, LogBytesPerWord, Gargs); // parameter size in bytes
1510 __ add(args, Gargs, Gargs); // points to first local + BytesPerWord
1511 // NEW
1512 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1513 // determine extra space for non-argument locals & adjust caller's SP
1514 // Gtmp1: parameter size in words
1515 __ lduh(size_of_locals, Gtmp);
1516 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1518 #if 1
1519 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1520 // the call_stub will place the final interpreter argument at
1521 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1522 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1523 // and try to make it look good in the debugger we will store the argument to
1524 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1525 // extra space for the compiler this will overwrite locals in the local array of the
1526 // interpreter.
1527 // QQQ still needed with frameless adapters???
1529 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1531 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1532 #endif // 1
1535 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1536 }
1538 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1540 // G5_method: Method*
1541 // G2_thread: thread (unused)
1542 // Gargs: bottom of args (sender_sp)
1543 // O5: sender's sp
1545 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1546 // the code is pretty much ready. Would need to change the test below and for good measure
1547 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1548 // routines. Not clear this is worth it yet.
1550 if (interpreter_frame_manager) {
1551 return interpreter_frame_manager;
1552 }
1554 __ bind(frame_manager_entry);
1556 // the following temporary registers are used during frame creation
1557 const Register Gtmp1 = G3_scratch;
1558 // const Register Lmirror = L1; // native mirror (native calls only)
1560 const Address constMethod (G5_method, 0, in_bytes(Method::const_offset()));
1561 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
1563 address entry_point = __ pc();
1564 __ mov(G0, prevState); // no current activation
1567 Label re_dispatch;
1569 __ bind(re_dispatch);
1571 // Interpreter needs to have locals completely contiguous. In order to do that
1572 // We must adjust the caller's stack pointer for any locals beyond just the
1573 // parameters
1574 adjust_callers_stack(Gargs);
1576 // O5_savedSP still contains sender's sp
1578 // NEW FRAME
1580 generate_compute_interpreter_state(Lstate, prevState, false);
1582 // At this point a new interpreter frame and state object are created and initialized
1583 // Lstate has the pointer to the new activation
1584 // Any stack banging or limit check should already be done.
1586 Label call_interpreter;
1588 __ bind(call_interpreter);
1591 #if 1
1592 __ set(0xdead002, Lmirror);
1593 __ set(0xdead002, L2_scratch);
1594 __ set(0xdead003, L3_scratch);
1595 __ set(0xdead004, L4_scratch);
1596 __ set(0xdead005, Lscratch);
1597 __ set(0xdead006, Lscratch2);
1598 __ set(0xdead007, L7_scratch);
1600 __ set(0xdeaf002, O2);
1601 __ set(0xdeaf003, O3);
1602 __ set(0xdeaf004, O4);
1603 __ set(0xdeaf005, O5);
1604 #endif
1606 // Call interpreter (stack bang complete) enter here if message is
1607 // set and we know stack size is valid
1609 Label call_interpreter_2;
1611 __ bind(call_interpreter_2);
1613 #ifdef ASSERT
1614 {
1615 Label skip;
1616 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1617 __ cmp(G3_scratch, SP);
1618 __ brx(Assembler::equal, false, Assembler::pt, skip);
1619 __ delayed()->nop();
1620 __ stop("SP not restored to frame bottom");
1621 __ bind(skip);
1622 }
1623 #endif
1625 VALIDATE_STATE(G3_scratch, 4);
1626 __ set_last_Java_frame(SP, noreg);
1627 __ mov(Lstate, O0); // (arg) pointer to current state
1629 __ call(CAST_FROM_FN_PTR(address,
1630 JvmtiExport::can_post_interpreter_events() ?
1631 BytecodeInterpreter::runWithChecks
1632 : BytecodeInterpreter::run),
1633 relocInfo::runtime_call_type);
1635 __ delayed()->nop();
1637 __ ld_ptr(STATE(_thread), G2_thread);
1638 __ reset_last_Java_frame();
1640 // examine msg from interpreter to determine next action
1641 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1643 __ ld(STATE(_msg), L1_scratch); // Get new message
1645 Label call_method;
1646 Label return_from_interpreted_method;
1647 Label throw_exception;
1648 Label do_OSR;
1649 Label bad_msg;
1650 Label resume_interpreter;
1652 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1653 __ br(Assembler::equal, false, Assembler::pt, call_method);
1654 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1655 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1656 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1657 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1658 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1659 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1660 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1661 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1663 // Allocate more monitor space, shuffle expression stack....
1665 generate_more_monitors();
1667 // new monitor slot allocated, resume the interpreter.
1669 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1670 VALIDATE_STATE(G3_scratch, 5);
1671 __ ba(call_interpreter);
1672 __ delayed()->st(L1_scratch, STATE(_msg));
1674 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1675 unctrap_frame_manager_entry = __ pc();
1677 // QQQ what message do we send
1679 __ ba(call_interpreter);
1680 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1682 //=============================================================================
1683 // Returning from a compiled method into a deopted method. The bytecode at the
1684 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1685 // for the template based interpreter). Any stack space that was used by the
1686 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1687 // so all that we have to do is place any pending result on the expression stack
1688 // and resume execution on the next bytecode.
1690 generate_deopt_handling();
1692 // ready to resume the interpreter
1694 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1695 __ ba(call_interpreter);
1696 __ delayed()->st(L1_scratch, STATE(_msg));
1698 // Current frame has caught an exception we need to dispatch to the
1699 // handler. We can get here because a native interpreter frame caught
1700 // an exception in which case there is no handler and we must rethrow
1701 // If it is a vanilla interpreted frame the we simply drop into the
1702 // interpreter and let it do the lookup.
1704 Interpreter::_rethrow_exception_entry = __ pc();
1706 Label return_with_exception;
1707 Label unwind_and_forward;
1709 // O0: exception
1710 // O7: throwing pc
1712 // We want exception in the thread no matter what we ultimately decide about frame type.
1714 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1715 __ verify_thread();
1716 __ st_ptr(O0, exception_addr);
1718 // get the Method*
1719 __ ld_ptr(STATE(_method), G5_method);
1721 // if this current frame vanilla or native?
1723 __ ld(access_flags, Gtmp1);
1724 __ btst(JVM_ACC_NATIVE, Gtmp1);
1725 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1726 __ delayed()->nop();
1728 // We drop thru to unwind a native interpreted frame with a pending exception
1729 // We jump here for the initial interpreter frame with exception pending
1730 // We unwind the current acivation and forward it to our caller.
1732 __ bind(unwind_and_forward);
1734 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1735 // as expected by forward_exception.
1737 __ restore(FP, G0, SP); // unwind interpreter state frame
1738 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1739 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1741 // Return point from a call which returns a result in the native abi
1742 // (c1/c2/jni-native). This result must be processed onto the java
1743 // expression stack.
1744 //
1745 // A pending exception may be present in which case there is no result present
1747 address return_from_native_method = __ pc();
1749 VALIDATE_STATE(G3_scratch, 6);
1751 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1752 // stack is in the state that the calling convention left it.
1753 // Copy the result from native abi result and place it on java expression stack.
1755 // Current interpreter state is present in Lstate
1757 // Exception pending?
1759 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1760 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1761 __ tst(Lscratch); // exception pending?
1762 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1763 __ delayed()->nop();
1765 // Process the native abi result to java expression stack
1767 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1768 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1769 // get parameter size
1770 __ ld_ptr(L4_scratch, in_bytes(Method::const_offset()), L2_scratch);
1771 __ lduh(L2_scratch, in_bytes(ConstMethod::size_of_parameters_offset()), L2_scratch);
1772 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1773 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1774 __ ld(L4_scratch, in_bytes(Method::result_index_offset()), L3_scratch); // called method result type index
1776 // tosca is really just native abi
1777 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1778 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1779 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1780 __ jmpl(Lscratch, G0, O7); // and convert it
1781 __ delayed()->nop();
1783 // L1_scratch points to top of stack (prepushed)
1785 __ ba(resume_interpreter);
1786 __ delayed()->mov(L1_scratch, O1);
1788 // An exception is being caught on return to a vanilla interpreter frame.
1789 // Empty the stack and resume interpreter
1791 __ bind(return_with_exception);
1793 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1794 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1795 __ ba(resume_interpreter);
1796 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1798 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1799 // interpreter call, or native) and unwind this interpreter activation.
1800 // All monitors should be unlocked.
1802 __ bind(return_from_interpreted_method);
1804 VALIDATE_STATE(G3_scratch, 7);
1806 Label return_to_initial_caller;
1808 // Interpreted result is on the top of the completed activation expression stack.
1809 // We must return it to the top of the callers stack if caller was interpreted
1810 // otherwise we convert to native abi result and return to call_stub/c1/c2
1811 // The caller's expression stack was truncated by the call however the current activation
1812 // has enough stuff on the stack that we have usable space there no matter what. The
1813 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1814 // for the current activation
1816 __ ld_ptr(STATE(_prev_link), L1_scratch);
1817 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1818 __ ld(L2_scratch, in_bytes(Method::result_index_offset()), L2_scratch);
1819 __ tst(L1_scratch);
1820 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1821 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1823 // Copy result to callers java stack
1825 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1826 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1827 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1828 __ ld_ptr(STATE(_locals), O1); // stack destination
1830 // O0 - will be source, O1 - will be destination (preserved)
1831 __ jmpl(Lscratch, G0, O7); // and convert it
1832 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1834 // O1 == &locals[0]
1836 // Result is now on caller's stack. Just unwind current activation and resume
1838 Label unwind_recursive_activation;
1841 __ bind(unwind_recursive_activation);
1843 // O1 == &locals[0] (really callers stacktop) for activation now returning
1844 // returning to interpreter method from "recursive" interpreter call
1845 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1846 // to. Now all we must do is unwind the state from the completed call
1848 // Must restore stack
1849 VALIDATE_STATE(G3_scratch, 8);
1851 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1852 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1853 // frame and tell interpreter to resume.
1856 __ mov(O1, I1); // pass back new stack top across activation
1857 // POP FRAME HERE ==================================
1858 __ restore(FP, G0, SP); // unwind interpreter state frame
1859 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1862 // Resume the interpreter. The current frame contains the current interpreter
1863 // state object.
1864 //
1865 // O1 == new java stack pointer
1867 __ bind(resume_interpreter);
1868 VALIDATE_STATE(G3_scratch, 10);
1870 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1872 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1873 __ st(L1_scratch, STATE(_msg));
1874 __ ba(call_interpreter_2);
1875 __ delayed()->st_ptr(O1, STATE(_stack));
1878 // Fast accessor methods share this entry point.
1879 // This works because frame manager is in the same codelet
1880 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1881 // we need to do a little register fixup here once we distinguish the two of them
1882 if (UseFastAccessorMethods && !synchronized) {
1883 // Call stub_return address still in O7
1884 __ bind(fast_accessor_slow_entry_path);
1885 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1886 __ cmp(Gtmp1, O7); // returning to interpreter?
1887 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1888 __ delayed()->nop();
1889 __ ba(re_dispatch);
1890 __ delayed()->mov(G0, prevState); // initial entry
1892 }
1894 // interpreter returning to native code (call_stub/c1/c2)
1895 // convert result and unwind initial activation
1896 // L2_scratch - scaled result type index
1898 __ bind(return_to_initial_caller);
1900 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1901 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1902 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1903 __ jmpl(Lscratch, G0, O7); // and convert it
1904 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1906 Label unwind_initial_activation;
1907 __ bind(unwind_initial_activation);
1909 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1910 // we can return here with an exception that wasn't handled by interpreted code
1911 // how does c1/c2 see it on return?
1913 // compute resulting sp before/after args popped depending upon calling convention
1914 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1915 //
1916 // POP FRAME HERE ==================================
1917 __ restore(FP, G0, SP);
1918 __ retl();
1919 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1921 // OSR request, unwind the current frame and transfer to the OSR entry
1922 // and enter OSR nmethod
1924 __ bind(do_OSR);
1925 Label remove_initial_frame;
1926 __ ld_ptr(STATE(_prev_link), L1_scratch);
1927 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1929 // We are going to pop this frame. Is there another interpreter frame underneath
1930 // it or is it callstub/compiled?
1932 __ tst(L1_scratch);
1933 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1934 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1936 // Frame underneath is an interpreter frame simply unwind
1937 // POP FRAME HERE ==================================
1938 __ restore(FP, G0, SP); // unwind interpreter state frame
1939 __ mov(I5_savedSP->after_restore(), SP);
1941 // Since we are now calling native need to change our "return address" from the
1942 // dummy RecursiveInterpreterActivation to a return from native
1944 __ set((intptr_t)return_from_native_method - 8, O7);
1946 __ jmpl(G3_scratch, G0, G0);
1947 __ delayed()->mov(G1_scratch, O0);
1949 __ bind(remove_initial_frame);
1951 // POP FRAME HERE ==================================
1952 __ restore(FP, G0, SP);
1953 __ mov(I5_savedSP->after_restore(), SP);
1954 __ jmpl(G3_scratch, G0, G0);
1955 __ delayed()->mov(G1_scratch, O0);
1957 // Call a new method. All we do is (temporarily) trim the expression stack
1958 // push a return address to bring us back to here and leap to the new entry.
1959 // At this point we have a topmost frame that was allocated by the frame manager
1960 // which contains the current method interpreted state. We trim this frame
1961 // of excess java expression stack entries and then recurse.
1963 __ bind(call_method);
1965 // stack points to next free location and not top element on expression stack
1966 // method expects sp to be pointing to topmost element
1968 __ ld_ptr(STATE(_thread), G2_thread);
1969 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1972 // SP already takes in to account the 2 extra words we use for slop
1973 // when we call a "static long no_params()" method. So if
1974 // we trim back sp by the amount of unused java expression stack
1975 // there will be automagically the 2 extra words we need.
1976 // We also have to worry about keeping SP aligned.
1978 __ ld_ptr(STATE(_stack), Gargs);
1979 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1981 // compute the unused java stack size
1982 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1984 // Round down the unused space to that stack is always 16-byte aligned
1985 // by making the unused space a multiple of the size of two longs.
1987 __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1989 // Now trim the stack
1990 __ add(SP, L2_scratch, SP);
1993 // Now point to the final argument (account for prepush)
1994 __ add(Gargs, wordSize, Gargs);
1995 #ifdef ASSERT
1996 // Make sure we have space for the window
1997 __ sub(Gargs, SP, L1_scratch);
1998 __ cmp(L1_scratch, 16*wordSize);
1999 {
2000 Label skip;
2001 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
2002 __ delayed()->nop();
2003 __ stop("killed stack");
2004 __ bind(skip);
2005 }
2006 #endif // ASSERT
2008 // Create a new frame where we can store values that make it look like the interpreter
2009 // really recursed.
2011 // prepare to recurse or call specialized entry
2013 // First link the registers we need
2015 // make the pc look good in debugger
2016 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
2017 // argument too
2018 __ mov(Lstate, I0);
2020 // Record our sending SP
2021 __ mov(SP, O5_savedSP);
2023 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
2024 __ set((intptr_t) entry_point, L1_scratch);
2025 __ cmp(L1_scratch, L2_scratch);
2026 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
2027 __ delayed()->mov(Lstate, prevState); // link activations
2029 // method uses specialized entry, push a return so we look like call stub setup
2030 // this path will handle fact that result is returned in registers and not
2031 // on the java stack.
2033 __ set((intptr_t)return_from_native_method - 8, O7);
2034 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
2035 __ delayed()->nop();
2037 //
2038 // Bad Message from interpreter
2039 //
2040 __ bind(bad_msg);
2041 __ stop("Bad message from interpreter");
2043 // Interpreted method "returned" with an exception pass it on...
2044 // Pass result, unwind activation and continue/return to interpreter/call_stub
2045 // We handle result (if any) differently based on return to interpreter or call_stub
2047 __ bind(throw_exception);
2048 __ ld_ptr(STATE(_prev_link), L1_scratch);
2049 __ tst(L1_scratch);
2050 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2051 __ delayed()->nop();
2053 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
2054 __ ba(unwind_recursive_activation);
2055 __ delayed()->nop();
2057 interpreter_frame_manager = entry_point;
2058 return entry_point;
2059 }
2061 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2062 : CppInterpreterGenerator(code) {
2063 generate_all(); // down here so it can be "virtual"
2064 }
2067 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2069 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2070 // expression stack, the callee will have callee_extra_locals (so we can account for
2071 // frame extension) and monitor_size for monitors. Basically we need to calculate
2072 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2073 //
2074 //
2075 // The big complicating thing here is that we must ensure that the stack stays properly
2076 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2077 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2078 // the caller so we must ensure that it is properly aligned for our callee.
2079 //
2080 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2081 // stack at all times to deal with the "stack long no_params()" method issue. This
2082 // is "slop_factor" here.
2083 const int slop_factor = 2;
2085 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2086 frame::memory_parameter_word_sp_offset; // register save area + param window
2087 const int extra_stack = 0; //6815692//Method::extra_stack_entries();
2088 return (round_to(max_stack +
2089 extra_stack +
2090 slop_factor +
2091 fixed_size +
2092 monitor_size +
2093 (callee_extra_locals * Interpreter::stackElementWords), WordsPerLong));
2095 }
2097 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
2099 // See call_stub code
2100 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2101 WordsPerLong); // 7 + register save area
2103 // Save space for one monitor to get into the interpreted method in case
2104 // the method is synchronized
2105 int monitor_size = method->is_synchronized() ?
2106 1*frame::interpreter_frame_monitor_size() : 0;
2107 return size_activation_helper(method->max_locals(), method->max_stack(),
2108 monitor_size) + call_stub_size;
2109 }
2111 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2112 frame* caller,
2113 frame* current,
2114 Method* method,
2115 intptr_t* locals,
2116 intptr_t* stack,
2117 intptr_t* stack_base,
2118 intptr_t* monitor_base,
2119 intptr_t* frame_bottom,
2120 bool is_top_frame
2121 )
2122 {
2123 // What about any vtable?
2124 //
2125 to_fill->_thread = JavaThread::current();
2126 // This gets filled in later but make it something recognizable for now
2127 to_fill->_bcp = method->code_base();
2128 to_fill->_locals = locals;
2129 to_fill->_constants = method->constants()->cache();
2130 to_fill->_method = method;
2131 to_fill->_mdx = NULL;
2132 to_fill->_stack = stack;
2133 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2134 to_fill->_msg = deopt_resume2;
2135 } else {
2136 to_fill->_msg = method_resume;
2137 }
2138 to_fill->_result._to_call._bcp_advance = 0;
2139 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2140 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2141 to_fill->_prev_link = NULL;
2143 // Fill in the registers for the frame
2145 // Need to install _sender_sp. Actually not too hard in C++!
2146 // When the skeletal frames are layed out we fill in a value
2147 // for _sender_sp. That value is only correct for the oldest
2148 // skeletal frame constructed (because there is only a single
2149 // entry for "caller_adjustment". While the skeletal frames
2150 // exist that is good enough. We correct that calculation
2151 // here and get all the frames correct.
2153 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2155 *current->register_addr(Lstate) = (intptr_t) to_fill;
2156 // skeletal already places a useful value here and this doesn't account
2157 // for alignment so don't bother.
2158 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2160 if (caller->is_interpreted_frame()) {
2161 interpreterState prev = caller->get_interpreterState();
2162 to_fill->_prev_link = prev;
2163 // Make the prev callee look proper
2164 prev->_result._to_call._callee = method;
2165 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2166 prev->_result._to_call._bcp_advance = 5;
2167 } else {
2168 prev->_result._to_call._bcp_advance = 3;
2169 }
2170 }
2171 to_fill->_oop_temp = NULL;
2172 to_fill->_stack_base = stack_base;
2173 // Need +1 here because stack_base points to the word just above the first expr stack entry
2174 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2175 // See generate_compute_interpreter_state.
2176 int extra_stack = 0; //6815692//Method::extra_stack_entries();
2177 to_fill->_stack_limit = stack_base - (method->max_stack() + 1 + extra_stack);
2178 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2180 // sparc specific
2181 to_fill->_frame_bottom = frame_bottom;
2182 to_fill->_self_link = to_fill;
2183 #ifdef ASSERT
2184 to_fill->_native_fresult = 123456.789;
2185 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2186 #endif
2187 }
2189 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2190 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2191 }
2194 int AbstractInterpreter::layout_activation(Method* method,
2195 int tempcount, // Number of slots on java expression stack in use
2196 int popframe_extra_args,
2197 int moncount, // Number of active monitors
2198 int caller_actual_parameters,
2199 int callee_param_size,
2200 int callee_locals_size,
2201 frame* caller,
2202 frame* interpreter_frame,
2203 bool is_top_frame,
2204 bool is_bottom_frame) {
2206 assert(popframe_extra_args == 0, "NEED TO FIX");
2207 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2208 // does as far as allocating an interpreter frame.
2209 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2210 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2211 // as determined by a previous call to this method.
2212 // It is also guaranteed to be walkable even though it is in a skeletal state
2213 // NOTE: return size is in words not bytes
2214 // NOTE: tempcount is the current size of the java expression stack. For top most
2215 // frames we will allocate a full sized expression stack and not the curback
2216 // version that non-top frames have.
2218 // Calculate the amount our frame will be adjust by the callee. For top frame
2219 // this is zero.
2221 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2222 // calculates the extra locals based on itself. Not what the callee does
2223 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2224 // as getting sender_sp correct.
2226 int extra_locals_size = callee_locals_size - callee_param_size;
2227 int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2228 int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2229 int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2230 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2233 /*
2234 if we actually have a frame to layout we must now fill in all the pieces. This means both
2235 the interpreterState and the registers.
2236 */
2237 if (interpreter_frame != NULL) {
2239 // MUCHO HACK
2241 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2242 // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2243 assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2244 frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2246 /* Now fillin the interpreterState object */
2248 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2251 intptr_t* locals;
2253 // Calculate the postion of locals[0]. This is painful because of
2254 // stack alignment (same as ia64). The problem is that we can
2255 // not compute the location of locals from fp(). fp() will account
2256 // for the extra locals but it also accounts for aligning the stack
2257 // and we can't determine if the locals[0] was misaligned but max_locals
2258 // was enough to have the
2259 // calculate postion of locals. fp already accounts for extra locals.
2260 // +2 for the static long no_params() issue.
2262 if (caller->is_interpreted_frame()) {
2263 // locals must agree with the caller because it will be used to set the
2264 // caller's tos when we return.
2265 interpreterState prev = caller->get_interpreterState();
2266 // stack() is prepushed.
2267 locals = prev->stack() + method->size_of_parameters();
2268 } else {
2269 // Lay out locals block in the caller adjacent to the register window save area.
2270 //
2271 // Compiled frames do not allocate a varargs area which is why this if
2272 // statement is needed.
2273 //
2274 intptr_t* fp = interpreter_frame->fp();
2275 int local_words = method->max_locals() * Interpreter::stackElementWords;
2277 if (caller->is_compiled_frame()) {
2278 locals = fp + frame::register_save_words + local_words - 1;
2279 } else {
2280 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2281 }
2283 }
2284 // END MUCHO HACK
2286 intptr_t* monitor_base = (intptr_t*) cur_state;
2287 intptr_t* stack_base = monitor_base - monitor_size;
2288 /* +1 because stack is always prepushed */
2289 intptr_t* stack = stack_base - (tempcount + 1);
2292 BytecodeInterpreter::layout_interpreterState(cur_state,
2293 caller,
2294 interpreter_frame,
2295 method,
2296 locals,
2297 stack,
2298 stack_base,
2299 monitor_base,
2300 frame_bottom,
2301 is_top_frame);
2303 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2305 }
2306 return frame_words;
2307 }
2309 #endif // CC_INTERP