Wed, 13 Mar 2013 09:44:45 +0100
8009761: Deoptimization on sparc doesn't set Llast_SP correctly in the interpreter frames it creates
Summary: deoptimization doesn't set up callee frames so that they restore caller frames correctly.
Reviewed-by: kvn
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 */
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 // Update standard invocation counters
408 __ increment_invocation_counter(O0, G3_scratch);
409 if (ProfileInterpreter) { // %%% Merge this into MethodData*
410 __ ld_ptr(STATE(_method), G3_scratch);
411 Address interpreter_invocation_counter(G3_scratch, 0, in_bytes(Method::interpreter_invocation_counter_offset()));
412 __ ld(interpreter_invocation_counter, G3_scratch);
413 __ inc(G3_scratch);
414 __ st(G3_scratch, interpreter_invocation_counter);
415 }
417 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
418 __ sethi(invocation_limit);
419 __ ld(invocation_limit, G3_scratch);
420 __ cmp(O0, G3_scratch);
421 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
422 __ delayed()->nop();
424 }
426 address InterpreterGenerator::generate_empty_entry(void) {
428 // A method that does nothing but return...
430 address entry = __ pc();
431 Label slow_path;
433 // do nothing for empty methods (do not even increment invocation counter)
434 if ( UseFastEmptyMethods) {
435 // If we need a safepoint check, generate full interpreter entry.
436 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
437 __ load_contents(sync_state, G3_scratch);
438 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
439 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
440 __ delayed()->nop();
442 // Code: _return
443 __ retl();
444 __ delayed()->mov(O5_savedSP, SP);
445 return entry;
446 }
447 return NULL;
448 }
450 // Call an accessor method (assuming it is resolved, otherwise drop into
451 // vanilla (slow path) entry
453 // Generates code to elide accessor methods
454 // Uses G3_scratch and G1_scratch as scratch
455 address InterpreterGenerator::generate_accessor_entry(void) {
457 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
458 // parameter size = 1
459 // Note: We can only use this code if the getfield has been resolved
460 // and if we don't have a null-pointer exception => check for
461 // these conditions first and use slow path if necessary.
462 address entry = __ pc();
463 Label slow_path;
465 if ( UseFastAccessorMethods) {
466 // Check if we need to reach a safepoint and generate full interpreter
467 // frame if so.
468 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
469 __ load_contents(sync_state, G3_scratch);
470 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
471 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
472 __ delayed()->nop();
474 // Check if local 0 != NULL
475 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
476 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
477 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
478 __ delayed()->nop();
481 // read first instruction word and extract bytecode @ 1 and index @ 2
482 // get first 4 bytes of the bytecodes (big endian!)
483 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::const_offset())), G1_scratch);
484 __ ld(Address(G1_scratch, 0, in_bytes(ConstMethod::codes_offset())), G1_scratch);
486 // move index @ 2 far left then to the right most two bytes.
487 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
488 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
489 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
491 // get constant pool cache
492 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), G3_scratch);
493 __ ld_ptr(G3_scratch, in_bytes(ConstMethod::constants_offset()), G3_scratch);
494 __ ld_ptr(G3_scratch, ConstantPool::cache_offset_in_bytes(), G3_scratch);
496 // get specific constant pool cache entry
497 __ add(G3_scratch, G1_scratch, G3_scratch);
499 // Check the constant Pool cache entry to see if it has been resolved.
500 // If not, need the slow path.
501 ByteSize cp_base_offset = ConstantPoolCache::base_offset();
502 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
503 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
504 __ and3(G1_scratch, 0xFF, G1_scratch);
505 __ cmp(G1_scratch, Bytecodes::_getfield);
506 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
507 __ delayed()->nop();
509 // Get the type and return field offset from the constant pool cache
510 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
511 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
513 Label xreturn_path;
514 // Need to differentiate between igetfield, agetfield, bgetfield etc.
515 // because they are different sizes.
516 // Get the type from the constant pool cache
517 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
518 // Make sure we don't need to mask G1_scratch after the above shift
519 ConstantPoolCacheEntry::verify_tos_state_shift();
520 __ cmp(G1_scratch, atos );
521 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
522 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
523 __ cmp(G1_scratch, itos);
524 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
525 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
526 __ cmp(G1_scratch, stos);
527 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
528 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
529 __ cmp(G1_scratch, ctos);
530 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
531 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
532 #ifdef ASSERT
533 __ cmp(G1_scratch, btos);
534 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
535 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
536 __ should_not_reach_here();
537 #endif
538 __ ldsb(Otos_i, G3_scratch, Otos_i);
539 __ bind(xreturn_path);
541 // _ireturn/_areturn
542 __ retl(); // return from leaf routine
543 __ delayed()->mov(O5_savedSP, SP);
545 // Generate regular method entry
546 __ bind(slow_path);
547 __ ba(fast_accessor_slow_entry_path);
548 __ delayed()->nop();
549 return entry;
550 }
551 return NULL;
552 }
554 address InterpreterGenerator::generate_Reference_get_entry(void) {
555 #if INCLUDE_ALL_GCS
556 if (UseG1GC) {
557 // We need to generate have a routine that generates code to:
558 // * load the value in the referent field
559 // * passes that value to the pre-barrier.
560 //
561 // In the case of G1 this will record the value of the
562 // referent in an SATB buffer if marking is active.
563 // This will cause concurrent marking to mark the referent
564 // field as live.
565 Unimplemented();
566 }
567 #endif // INCLUDE_ALL_GCS
569 // If G1 is not enabled then attempt to go through the accessor entry point
570 // Reference.get is an accessor
571 return generate_accessor_entry();
572 }
574 //
575 // Interpreter stub for calling a native method. (C++ interpreter)
576 // This sets up a somewhat different looking stack for calling the native method
577 // than the typical interpreter frame setup.
578 //
580 address InterpreterGenerator::generate_native_entry(bool synchronized) {
581 address entry = __ pc();
583 // the following temporary registers are used during frame creation
584 const Register Gtmp1 = G3_scratch ;
585 const Register Gtmp2 = G1_scratch;
586 const Register RconstMethod = Gtmp1;
587 const Address constMethod(G5_method, 0, in_bytes(Method::const_offset()));
588 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
590 bool inc_counter = UseCompiler || CountCompiledCalls;
592 // make sure registers are different!
593 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
595 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
597 Label Lentry;
598 __ bind(Lentry);
600 const Register Glocals_size = G3;
601 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
603 // make sure method is native & not abstract
604 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
605 #ifdef ASSERT
606 __ ld(access_flags, Gtmp1);
607 {
608 Label L;
609 __ btst(JVM_ACC_NATIVE, Gtmp1);
610 __ br(Assembler::notZero, false, Assembler::pt, L);
611 __ delayed()->nop();
612 __ stop("tried to execute non-native method as native");
613 __ bind(L);
614 }
615 { Label L;
616 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
617 __ br(Assembler::zero, false, Assembler::pt, L);
618 __ delayed()->nop();
619 __ stop("tried to execute abstract method as non-abstract");
620 __ bind(L);
621 }
622 #endif // ASSERT
624 __ ld_ptr(constMethod, RconstMethod);
625 __ lduh(size_of_parameters, Gtmp1);
626 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes
627 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord
628 // NEW
629 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
630 // generate the code to allocate the interpreter stack frame
631 // NEW FRAME ALLOCATED HERE
632 // save callers original sp
633 // __ mov(SP, I5_savedSP->after_restore());
635 generate_compute_interpreter_state(Lstate, G0, true);
637 // At this point Lstate points to new interpreter state
638 //
640 const Address do_not_unlock_if_synchronized(G2_thread, 0,
641 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
642 // Since at this point in the method invocation the exception handler
643 // would try to exit the monitor of synchronized methods which hasn't
644 // been entered yet, we set the thread local variable
645 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
646 // runtime, exception handling i.e. unlock_if_synchronized_method will
647 // check this thread local flag.
648 // This flag has two effects, one is to force an unwind in the topmost
649 // interpreter frame and not perform an unlock while doing so.
651 __ movbool(true, G3_scratch);
652 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
655 // increment invocation counter and check for overflow
656 //
657 // Note: checking for negative value instead of overflow
658 // so we have a 'sticky' overflow test (may be of
659 // importance as soon as we have true MT/MP)
660 Label invocation_counter_overflow;
661 if (inc_counter) {
662 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
663 }
664 Label Lcontinue;
665 __ bind(Lcontinue);
667 bang_stack_shadow_pages(true);
668 // reset the _do_not_unlock_if_synchronized flag
669 __ stbool(G0, do_not_unlock_if_synchronized);
671 // check for synchronized methods
672 // Must happen AFTER invocation_counter check, so method is not locked
673 // if counter overflows.
675 if (synchronized) {
676 lock_method();
677 // Don't see how G2_thread is preserved here...
678 // __ verify_thread(); QQQ destroys L0,L1 can't use
679 } else {
680 #ifdef ASSERT
681 { Label ok;
682 __ ld_ptr(STATE(_method), G5_method);
683 __ ld(access_flags, O0);
684 __ btst(JVM_ACC_SYNCHRONIZED, O0);
685 __ br( Assembler::zero, false, Assembler::pt, ok);
686 __ delayed()->nop();
687 __ stop("method needs synchronization");
688 __ bind(ok);
689 }
690 #endif // ASSERT
691 }
693 // start execution
695 // __ verify_thread(); kills L1,L2 can't use at the moment
697 // jvmti/jvmpi support
698 __ notify_method_entry();
700 // native call
702 // (note that O0 is never an oop--at most it is a handle)
703 // It is important not to smash any handles created by this call,
704 // until any oop handle in O0 is dereferenced.
706 // (note that the space for outgoing params is preallocated)
708 // get signature handler
710 Label pending_exception_present;
712 { Label L;
713 __ ld_ptr(STATE(_method), G5_method);
714 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
715 __ tst(G3_scratch);
716 __ brx(Assembler::notZero, false, Assembler::pt, L);
717 __ delayed()->nop();
718 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
719 __ ld_ptr(STATE(_method), G5_method);
721 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
722 __ ld_ptr(exception_addr, G3_scratch);
723 __ br_notnull_short(G3_scratch, Assembler::pn, pending_exception_present);
724 __ ld_ptr(Address(G5_method, 0, in_bytes(Method::signature_handler_offset())), G3_scratch);
725 __ bind(L);
726 }
728 // Push a new frame so that the args will really be stored in
729 // Copy a few locals across so the new frame has the variables
730 // we need but these values will be dead at the jni call and
731 // therefore not gc volatile like the values in the current
732 // frame (Lstate in particular)
734 // Flush the state pointer to the register save area
735 // Which is the only register we need for a stack walk.
736 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
738 __ mov(Lstate, O1); // Need to pass the state pointer across the frame
740 // Calculate current frame size
741 __ sub(SP, FP, O3); // Calculate negative of current frame size
742 __ save(SP, O3, SP); // Allocate an identical sized frame
744 __ mov(I1, Lstate); // In the "natural" register.
746 // Note I7 has leftover trash. Slow signature handler will fill it in
747 // should we get there. Normal jni call will set reasonable last_Java_pc
748 // below (and fix I7 so the stack trace doesn't have a meaningless frame
749 // in it).
752 // call signature handler
753 __ ld_ptr(STATE(_method), Lmethod);
754 __ ld_ptr(STATE(_locals), Llocals);
756 __ callr(G3_scratch, 0);
757 __ delayed()->nop();
758 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed)
760 { Label not_static;
762 __ ld_ptr(STATE(_method), G5_method);
763 __ ld(access_flags, O0);
764 __ btst(JVM_ACC_STATIC, O0);
765 __ br( Assembler::zero, false, Assembler::pt, not_static);
766 __ delayed()->
767 // get native function entry point(O0 is a good temp until the very end)
768 ld_ptr(Address(G5_method, 0, in_bytes(Method::native_function_offset())), O0);
769 // for static methods insert the mirror argument
770 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
772 __ ld_ptr(Address(G5_method, 0, in_bytes(Method:: const_offset())), O1);
773 __ ld_ptr(Address(O1, 0, in_bytes(ConstMethod::constants_offset())), O1);
774 __ ld_ptr(Address(O1, 0, ConstantPool::pool_holder_offset_in_bytes()), O1);
775 __ ld_ptr(O1, mirror_offset, O1);
776 // where the mirror handle body is allocated:
777 #ifdef ASSERT
778 if (!PrintSignatureHandlers) // do not dirty the output with this
779 { Label L;
780 __ tst(O1);
781 __ brx(Assembler::notZero, false, Assembler::pt, L);
782 __ delayed()->nop();
783 __ stop("mirror is missing");
784 __ bind(L);
785 }
786 #endif // ASSERT
787 __ st_ptr(O1, STATE(_oop_temp));
788 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add
789 __ bind(not_static);
790 }
792 // At this point, arguments have been copied off of stack into
793 // their JNI positions, which are O1..O5 and SP[68..].
794 // Oops are boxed in-place on the stack, with handles copied to arguments.
795 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
797 #ifdef ASSERT
798 { Label L;
799 __ tst(O0);
800 __ brx(Assembler::notZero, false, Assembler::pt, L);
801 __ delayed()->nop();
802 __ stop("native entry point is missing");
803 __ bind(L);
804 }
805 #endif // ASSERT
807 //
808 // setup the java frame anchor
809 //
810 // The scavenge function only needs to know that the PC of this frame is
811 // in the interpreter method entry code, it doesn't need to know the exact
812 // PC and hence we can use O7 which points to the return address from the
813 // previous call in the code stream (signature handler function)
814 //
815 // The other trick is we set last_Java_sp to FP instead of the usual SP because
816 // we have pushed the extra frame in order to protect the volatile register(s)
817 // in that frame when we return from the jni call
818 //
821 __ set_last_Java_frame(FP, O7);
822 __ mov(O7, I7); // make dummy interpreter frame look like one above,
823 // not meaningless information that'll confuse me.
825 // flush the windows now. We don't care about the current (protection) frame
826 // only the outer frames
828 __ flush_windows();
830 // mark windows as flushed
831 Address flags(G2_thread,
832 0,
833 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
834 __ set(JavaFrameAnchor::flushed, G3_scratch);
835 __ st(G3_scratch, flags);
837 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
839 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
840 #ifdef ASSERT
841 { Label L;
842 __ ld(thread_state, G3_scratch);
843 __ cmp(G3_scratch, _thread_in_Java);
844 __ br(Assembler::equal, false, Assembler::pt, L);
845 __ delayed()->nop();
846 __ stop("Wrong thread state in native stub");
847 __ bind(L);
848 }
849 #endif // ASSERT
850 __ set(_thread_in_native, G3_scratch);
851 __ st(G3_scratch, thread_state);
853 // Call the jni method, using the delay slot to set the JNIEnv* argument.
854 __ callr(O0, 0);
855 __ delayed()->
856 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
857 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
859 // must we block?
861 // Block, if necessary, before resuming in _thread_in_Java state.
862 // In order for GC to work, don't clear the last_Java_sp until after blocking.
863 { Label no_block;
864 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
866 // Switch thread to "native transition" state before reading the synchronization state.
867 // This additional state is necessary because reading and testing the synchronization
868 // state is not atomic w.r.t. GC, as this scenario demonstrates:
869 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
870 // VM thread changes sync state to synchronizing and suspends threads for GC.
871 // Thread A is resumed to finish this native method, but doesn't block here since it
872 // didn't see any synchronization is progress, and escapes.
873 __ set(_thread_in_native_trans, G3_scratch);
874 __ st(G3_scratch, thread_state);
875 if(os::is_MP()) {
876 // Write serialization page so VM thread can do a pseudo remote membar.
877 // We use the current thread pointer to calculate a thread specific
878 // offset to write to within the page. This minimizes bus traffic
879 // due to cache line collision.
880 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
881 }
882 __ load_contents(sync_state, G3_scratch);
883 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
886 Label L;
887 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
888 __ br(Assembler::notEqual, false, Assembler::pn, L);
889 __ delayed()->
890 ld(suspend_state, G3_scratch);
891 __ cmp(G3_scratch, 0);
892 __ br(Assembler::equal, false, Assembler::pt, no_block);
893 __ delayed()->nop();
894 __ bind(L);
896 // Block. Save any potential method result value before the operation and
897 // use a leaf call to leave the last_Java_frame setup undisturbed.
898 save_native_result();
899 __ call_VM_leaf(noreg,
900 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
901 G2_thread);
902 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
903 // Restore any method result value
904 restore_native_result();
905 __ bind(no_block);
906 }
908 // Clear the frame anchor now
910 __ reset_last_Java_frame();
912 // Move the result handler address
913 __ mov(Lscratch, G3_scratch);
914 // return possible result to the outer frame
915 #ifndef __LP64
916 __ mov(O0, I0);
917 __ restore(O1, G0, O1);
918 #else
919 __ restore(O0, G0, O0);
920 #endif /* __LP64 */
922 // Move result handler to expected register
923 __ mov(G3_scratch, Lscratch);
926 // thread state is thread_in_native_trans. Any safepoint blocking has
927 // happened in the trampoline we are ready to switch to thread_in_Java.
929 __ set(_thread_in_Java, G3_scratch);
930 __ st(G3_scratch, thread_state);
932 // If we have an oop result store it where it will be safe for any further gc
933 // until we return now that we've released the handle it might be protected by
935 {
936 Label no_oop, store_result;
938 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
939 __ cmp(G3_scratch, Lscratch);
940 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
941 __ delayed()->nop();
942 __ addcc(G0, O0, O0);
943 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
944 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
945 __ mov(G0, O0);
947 __ bind(store_result);
948 // Store it where gc will look for it and result handler expects it.
949 __ st_ptr(O0, STATE(_oop_temp));
951 __ bind(no_oop);
953 }
955 // reset handle block
956 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
957 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
960 // handle exceptions (exception handling will handle unlocking!)
961 { Label L;
962 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
964 __ ld_ptr(exception_addr, Gtemp);
965 __ tst(Gtemp);
966 __ brx(Assembler::equal, false, Assembler::pt, L);
967 __ delayed()->nop();
968 __ bind(pending_exception_present);
969 // With c++ interpreter we just leave it pending caller will do the correct thing. However...
970 // Like x86 we ignore the result of the native call and leave the method locked. This
971 // seems wrong to leave things locked.
973 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
974 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
976 __ bind(L);
977 }
979 // jvmdi/jvmpi support (preserves thread register)
980 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
982 if (synchronized) {
983 // save and restore any potential method result value around the unlocking operation
984 save_native_result();
986 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
987 // Get the initial monitor we allocated
988 __ sub(Lstate, entry_size, O1); // initial monitor
989 __ unlock_object(O1);
990 restore_native_result();
991 }
993 #if defined(COMPILER2) && !defined(_LP64)
995 // C2 expects long results in G1 we can't tell if we're returning to interpreted
996 // or compiled so just be safe.
998 __ sllx(O0, 32, G1); // Shift bits into high G1
999 __ srl (O1, 0, O1); // Zero extend O1
1000 __ or3 (O1, G1, G1); // OR 64 bits into G1
1002 #endif /* COMPILER2 && !_LP64 */
1004 #ifdef ASSERT
1005 {
1006 Label ok;
1007 __ cmp(I5_savedSP, FP);
1008 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
1009 __ delayed()->nop();
1010 __ stop("bad I5_savedSP value");
1011 __ should_not_reach_here();
1012 __ bind(ok);
1013 }
1014 #endif
1015 // Calls result handler which POPS FRAME
1016 if (TraceJumps) {
1017 // Move target to register that is recordable
1018 __ mov(Lscratch, G3_scratch);
1019 __ JMP(G3_scratch, 0);
1020 } else {
1021 __ jmp(Lscratch, 0);
1022 }
1023 __ delayed()->nop();
1025 if (inc_counter) {
1026 // handle invocation counter overflow
1027 __ bind(invocation_counter_overflow);
1028 generate_counter_overflow(Lcontinue);
1029 }
1032 return entry;
1033 }
1035 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1036 const Register prev_state,
1037 bool native) {
1039 // On entry
1040 // G5_method - caller's method
1041 // Gargs - points to initial parameters (i.e. locals[0])
1042 // G2_thread - valid? (C1 only??)
1043 // "prev_state" - contains any previous frame manager state which we must save a link
1044 //
1045 // On return
1046 // "state" is a pointer to the newly allocated state object. We must allocate and initialize
1047 // a new interpretState object and the method expression stack.
1049 assert_different_registers(state, prev_state);
1050 assert_different_registers(prev_state, G3_scratch);
1051 const Register Gtmp = G3_scratch;
1052 const Address constMethod (G5_method, 0, in_bytes(Method::const_offset()));
1053 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
1055 // slop factor is two extra slots on the expression stack so that
1056 // we always have room to store a result when returning from a call without parameters
1057 // that returns a result.
1059 const int slop_factor = 2*wordSize;
1061 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1062 //6815692//Method::extra_stack_words() + // extra push slots for MH adapters
1063 frame::memory_parameter_word_sp_offset + // register save area + param window
1064 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1066 // XXX G5_method valid
1068 // Now compute new frame size
1070 if (native) {
1071 const Register RconstMethod = Gtmp;
1072 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1073 __ ld_ptr(constMethod, RconstMethod);
1074 __ lduh( size_of_parameters, Gtmp );
1075 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words
1076 } else {
1077 // Full size expression stack
1078 __ ld_ptr(constMethod, Gtmp);
1079 __ lduh(Gtmp, in_bytes(ConstMethod::max_stack_offset()), Gtmp);
1080 }
1081 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1083 __ neg(Gtmp); // negative space for stack/parameters in words
1084 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1085 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1087 // Need to do stack size check here before we fault on large frames
1089 Label stack_ok;
1091 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1092 (StackRedPages+StackYellowPages);
1095 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1096 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1097 // compute stack bottom
1098 __ sub(O0, O1, O0);
1100 // Avoid touching the guard pages
1101 // Also a fudge for frame size of BytecodeInterpreter::run
1102 // It varies from 1k->4k depending on build type
1103 const int fudge = 6 * K;
1105 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1107 __ add(O0, O1, O0);
1108 __ sub(O0, Gtmp, O0);
1109 __ cmp(SP, O0);
1110 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1111 __ delayed()->nop();
1113 // throw exception return address becomes throwing pc
1115 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1116 __ stop("never reached");
1118 __ bind(stack_ok);
1120 __ save(SP, Gtmp, SP); // setup new frame and register window
1122 // New window I7 call_stub or previous activation
1123 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1124 //
1125 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1126 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1128 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1130 // Initialize a new Interpreter state
1131 // orig_sp - caller's original sp
1132 // G2_thread - thread
1133 // Gargs - &locals[0] (unbiased?)
1134 // G5_method - method
1135 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1138 __ set(0xdead0004, O1);
1141 __ st_ptr(Gargs, XXX_STATE(_locals));
1142 __ st_ptr(G0, XXX_STATE(_oop_temp));
1144 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1145 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1146 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1148 if (native) {
1149 __ st_ptr(G0, XXX_STATE(_bcp));
1150 } else {
1151 __ ld_ptr(G5_method, in_bytes(Method::const_offset()), O2); // get ConstMethod*
1152 __ add(O2, in_bytes(ConstMethod::codes_offset()), O2); // get bcp
1153 __ st_ptr(O2, XXX_STATE(_bcp));
1154 }
1156 __ st_ptr(G0, XXX_STATE(_mdx));
1157 __ st_ptr(G5_method, XXX_STATE(_method));
1159 __ set((int) BytecodeInterpreter::method_entry, O1);
1160 __ st(O1, XXX_STATE(_msg));
1162 __ ld_ptr(constMethod, O3);
1163 __ ld_ptr(O3, in_bytes(ConstMethod::constants_offset()), O3);
1164 __ ld_ptr(O3, ConstantPool::cache_offset_in_bytes(), O2);
1165 __ st_ptr(O2, XXX_STATE(_constants));
1167 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1169 // Monitor base is just start of BytecodeInterpreter object;
1170 __ mov(state, O2);
1171 __ st_ptr(O2, XXX_STATE(_monitor_base));
1173 // Do we need a monitor for synchonized method?
1174 {
1175 __ ld(access_flags, O1);
1176 Label done;
1177 Label got_obj;
1178 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1179 __ br( Assembler::zero, false, Assembler::pt, done);
1181 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1182 __ delayed()->btst(JVM_ACC_STATIC, O1);
1183 __ ld_ptr(XXX_STATE(_locals), O1);
1184 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1185 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1186 __ ld_ptr(constMethod, O1);
1187 __ ld_ptr( O1, in_bytes(ConstMethod::constants_offset()), O1);
1188 __ ld_ptr( O1, ConstantPool::pool_holder_offset_in_bytes(), O1);
1189 // lock the mirror, not the Klass*
1190 __ ld_ptr( O1, mirror_offset, O1);
1192 __ bind(got_obj);
1194 #ifdef ASSERT
1195 __ tst(O1);
1196 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
1197 #endif // ASSERT
1199 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1200 __ sub(SP, entry_size, SP); // account for initial monitor
1201 __ sub(O2, entry_size, O2); // initial monitor
1202 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1203 __ bind(done);
1204 }
1206 // Remember initial frame bottom
1208 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1210 __ st_ptr(O2, XXX_STATE(_stack_base));
1212 __ sub(O2, wordSize, O2); // prepush
1213 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1215 // Full size expression stack
1216 __ ld_ptr(constMethod, O3);
1217 __ lduh(O3, in_bytes(ConstMethod::max_stack_offset()), O3);
1218 guarantee(!EnableInvokeDynamic, "no support yet for java.lang.invoke.MethodHandle"); //6815692
1219 //6815692//if (EnableInvokeDynamic)
1220 //6815692// __ inc(O3, Method::extra_stack_entries());
1221 __ sll(O3, LogBytesPerWord, O3);
1222 __ sub(O2, O3, O3);
1223 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1224 __ st_ptr(O3, XXX_STATE(_stack_limit));
1226 if (!native) {
1227 //
1228 // Code to initialize locals
1229 //
1230 Register init_value = noreg; // will be G0 if we must clear locals
1231 // Now zero locals
1232 if (true /* zerolocals */ || ClearInterpreterLocals) {
1233 // explicitly initialize locals
1234 init_value = G0;
1235 } else {
1236 #ifdef ASSERT
1237 // initialize locals to a garbage pattern for better debugging
1238 init_value = O3;
1239 __ set( 0x0F0F0F0F, init_value );
1240 #endif // ASSERT
1241 }
1242 if (init_value != noreg) {
1243 Label clear_loop;
1244 const Register RconstMethod = O1;
1245 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1246 const Address size_of_locals (RconstMethod, 0, in_bytes(ConstMethod::size_of_locals_offset()));
1248 // NOTE: If you change the frame layout, this code will need to
1249 // be updated!
1250 __ ld_ptr( constMethod, RconstMethod );
1251 __ lduh( size_of_locals, O2 );
1252 __ lduh( size_of_parameters, O1 );
1253 __ sll( O2, LogBytesPerWord, O2);
1254 __ sll( O1, LogBytesPerWord, O1 );
1255 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1256 __ sub( L2_scratch, O2, O2 );
1257 __ sub( L2_scratch, O1, O1 );
1259 __ bind( clear_loop );
1260 __ inc( O2, wordSize );
1262 __ cmp( O2, O1 );
1263 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1264 __ delayed()->st_ptr( init_value, O2, 0 );
1265 }
1266 }
1267 }
1268 // Find preallocated monitor and lock method (C++ interpreter)
1269 //
1270 void InterpreterGenerator::lock_method(void) {
1271 // Lock the current method.
1272 // Destroys registers L2_scratch, L3_scratch, O0
1273 //
1274 // Find everything relative to Lstate
1276 #ifdef ASSERT
1277 __ ld_ptr(STATE(_method), L2_scratch);
1278 __ ld(L2_scratch, in_bytes(Method::access_flags_offset()), O0);
1280 { Label ok;
1281 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1282 __ br( Assembler::notZero, false, Assembler::pt, ok);
1283 __ delayed()->nop();
1284 __ stop("method doesn't need synchronization");
1285 __ bind(ok);
1286 }
1287 #endif // ASSERT
1289 // monitor is already allocated at stack base
1290 // and the lockee is already present
1291 __ ld_ptr(STATE(_stack_base), L2_scratch);
1292 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1293 __ lock_object(L2_scratch, O0);
1295 }
1297 // Generate code for handling resuming a deopted method
1298 void CppInterpreterGenerator::generate_deopt_handling() {
1300 Label return_from_deopt_common;
1302 // deopt needs to jump to here to enter the interpreter (return a result)
1303 deopt_frame_manager_return_atos = __ pc();
1305 // O0/O1 live
1306 __ ba(return_from_deopt_common);
1307 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1310 // deopt needs to jump to here to enter the interpreter (return a result)
1311 deopt_frame_manager_return_btos = __ pc();
1313 // O0/O1 live
1314 __ ba(return_from_deopt_common);
1315 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1317 // deopt needs to jump to here to enter the interpreter (return a result)
1318 deopt_frame_manager_return_itos = __ pc();
1320 // O0/O1 live
1321 __ ba(return_from_deopt_common);
1322 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1324 // deopt needs to jump to here to enter the interpreter (return a result)
1326 deopt_frame_manager_return_ltos = __ pc();
1327 #if !defined(_LP64) && defined(COMPILER2)
1328 // All return values are where we want them, except for Longs. C2 returns
1329 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1330 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1331 // build even if we are returning from interpreted we just do a little
1332 // stupid shuffing.
1333 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1334 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1335 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1337 __ srl (G1, 0,O1);
1338 __ srlx(G1,32,O0);
1339 #endif /* !_LP64 && COMPILER2 */
1340 // O0/O1 live
1341 __ ba(return_from_deopt_common);
1342 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1344 // deopt needs to jump to here to enter the interpreter (return a result)
1346 deopt_frame_manager_return_ftos = __ pc();
1347 // O0/O1 live
1348 __ ba(return_from_deopt_common);
1349 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1351 // deopt needs to jump to here to enter the interpreter (return a result)
1352 deopt_frame_manager_return_dtos = __ pc();
1354 // O0/O1 live
1355 __ ba(return_from_deopt_common);
1356 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1358 // deopt needs to jump to here to enter the interpreter (return a result)
1359 deopt_frame_manager_return_vtos = __ pc();
1361 // O0/O1 live
1362 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1364 // Deopt return common
1365 // an index is present that lets us move any possible result being
1366 // return to the interpreter's stack
1367 //
1368 __ bind(return_from_deopt_common);
1370 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1371 // stack is in the state that the calling convention left it.
1372 // Copy the result from native abi result and place it on java expression stack.
1374 // Current interpreter state is present in Lstate
1376 // Get current pre-pushed top of interpreter stack
1377 // Any result (if any) is in native abi
1378 // result type index is in L3_scratch
1380 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1382 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1383 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1384 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1385 __ jmpl(Lscratch, G0, O7); // and convert it
1386 __ delayed()->nop();
1388 // L1_scratch points to top of stack (prepushed)
1389 __ st_ptr(L1_scratch, STATE(_stack));
1390 }
1392 // Generate the code to handle a more_monitors message from the c++ interpreter
1393 void CppInterpreterGenerator::generate_more_monitors() {
1395 Label entry, loop;
1396 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1397 // 1. compute new pointers // esp: old expression stack top
1398 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1399 __ sub(L4_scratch, entry_size, L4_scratch);
1400 __ st_ptr(L4_scratch, STATE(_stack_base));
1402 __ sub(SP, entry_size, SP); // Grow stack
1403 __ st_ptr(SP, STATE(_frame_bottom));
1405 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1406 __ sub(L2_scratch, entry_size, L2_scratch);
1407 __ st_ptr(L2_scratch, STATE(_stack_limit));
1409 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1410 __ sub(L1_scratch, entry_size, L1_scratch);
1411 __ st_ptr(L1_scratch, STATE(_stack));
1412 __ ba(entry);
1413 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1415 // 2. move expression stack
1417 __ bind(loop);
1418 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1419 __ add(L1_scratch, wordSize, L1_scratch);
1420 __ bind(entry);
1421 __ cmp(L1_scratch, L4_scratch);
1422 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1423 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1425 // now zero the slot so we can find it.
1426 __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1428 }
1430 // Initial entry to C++ interpreter from the call_stub.
1431 // This entry point is called the frame manager since it handles the generation
1432 // of interpreter activation frames via requests directly from the vm (via call_stub)
1433 // and via requests from the interpreter. The requests from the call_stub happen
1434 // directly thru the entry point. Requests from the interpreter happen via returning
1435 // from the interpreter and examining the message the interpreter has returned to
1436 // the frame manager. The frame manager can take the following requests:
1438 // NO_REQUEST - error, should never happen.
1439 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1440 // allocate a new monitor.
1441 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1442 // happens during entry during the entry via the call stub.
1443 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1444 //
1445 // Arguments:
1446 //
1447 // ebx: Method*
1448 // ecx: receiver - unused (retrieved from stack as needed)
1449 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1450 //
1451 //
1452 // Stack layout at entry
1453 //
1454 // [ return address ] <--- esp
1455 // [ parameter n ]
1456 // ...
1457 // [ parameter 1 ]
1458 // [ expression stack ]
1459 //
1460 //
1461 // We are free to blow any registers we like because the call_stub which brought us here
1462 // initially has preserved the callee save registers already.
1463 //
1464 //
1466 static address interpreter_frame_manager = NULL;
1468 #ifdef ASSERT
1469 #define VALIDATE_STATE(scratch, marker) \
1470 { \
1471 Label skip; \
1472 __ ld_ptr(STATE(_self_link), scratch); \
1473 __ cmp(Lstate, scratch); \
1474 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1475 __ delayed()->nop(); \
1476 __ breakpoint_trap(); \
1477 __ emit_int32(marker); \
1478 __ bind(skip); \
1479 }
1480 #else
1481 #define VALIDATE_STATE(scratch, marker)
1482 #endif /* ASSERT */
1484 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1485 //
1486 // Adjust caller's stack so that all the locals can be contiguous with
1487 // the parameters.
1488 // Worries about stack overflow make this a pain.
1489 //
1490 // Destroys args, G3_scratch, G3_scratch
1491 // In/Out O5_savedSP (sender's original SP)
1492 //
1493 // assert_different_registers(state, prev_state);
1494 const Register Gtmp = G3_scratch;
1495 const RconstMethod = G3_scratch;
1496 const Register tmp = O2;
1497 const Address constMethod(G5_method, 0, in_bytes(Method::const_offset()));
1498 const Address size_of_parameters(RconstMethod, 0, in_bytes(ConstMethod::size_of_parameters_offset()));
1499 const Address size_of_locals (RconstMethod, 0, in_bytes(ConstMethod::size_of_locals_offset()));
1501 __ ld_ptr(constMethod, RconstMethod);
1502 __ lduh(size_of_parameters, tmp);
1503 __ sll(tmp, LogBytesPerWord, Gargs); // parameter size in bytes
1504 __ add(args, Gargs, Gargs); // points to first local + BytesPerWord
1505 // NEW
1506 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1507 // determine extra space for non-argument locals & adjust caller's SP
1508 // Gtmp1: parameter size in words
1509 __ lduh(size_of_locals, Gtmp);
1510 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1512 #if 1
1513 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1514 // the call_stub will place the final interpreter argument at
1515 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1516 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1517 // and try to make it look good in the debugger we will store the argument to
1518 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1519 // extra space for the compiler this will overwrite locals in the local array of the
1520 // interpreter.
1521 // QQQ still needed with frameless adapters???
1523 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1525 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1526 #endif // 1
1529 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1530 }
1532 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1534 // G5_method: Method*
1535 // G2_thread: thread (unused)
1536 // Gargs: bottom of args (sender_sp)
1537 // O5: sender's sp
1539 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1540 // the code is pretty much ready. Would need to change the test below and for good measure
1541 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1542 // routines. Not clear this is worth it yet.
1544 if (interpreter_frame_manager) {
1545 return interpreter_frame_manager;
1546 }
1548 __ bind(frame_manager_entry);
1550 // the following temporary registers are used during frame creation
1551 const Register Gtmp1 = G3_scratch;
1552 // const Register Lmirror = L1; // native mirror (native calls only)
1554 const Address constMethod (G5_method, 0, in_bytes(Method::const_offset()));
1555 const Address access_flags (G5_method, 0, in_bytes(Method::access_flags_offset()));
1557 address entry_point = __ pc();
1558 __ mov(G0, prevState); // no current activation
1561 Label re_dispatch;
1563 __ bind(re_dispatch);
1565 // Interpreter needs to have locals completely contiguous. In order to do that
1566 // We must adjust the caller's stack pointer for any locals beyond just the
1567 // parameters
1568 adjust_callers_stack(Gargs);
1570 // O5_savedSP still contains sender's sp
1572 // NEW FRAME
1574 generate_compute_interpreter_state(Lstate, prevState, false);
1576 // At this point a new interpreter frame and state object are created and initialized
1577 // Lstate has the pointer to the new activation
1578 // Any stack banging or limit check should already be done.
1580 Label call_interpreter;
1582 __ bind(call_interpreter);
1585 #if 1
1586 __ set(0xdead002, Lmirror);
1587 __ set(0xdead002, L2_scratch);
1588 __ set(0xdead003, L3_scratch);
1589 __ set(0xdead004, L4_scratch);
1590 __ set(0xdead005, Lscratch);
1591 __ set(0xdead006, Lscratch2);
1592 __ set(0xdead007, L7_scratch);
1594 __ set(0xdeaf002, O2);
1595 __ set(0xdeaf003, O3);
1596 __ set(0xdeaf004, O4);
1597 __ set(0xdeaf005, O5);
1598 #endif
1600 // Call interpreter (stack bang complete) enter here if message is
1601 // set and we know stack size is valid
1603 Label call_interpreter_2;
1605 __ bind(call_interpreter_2);
1607 #ifdef ASSERT
1608 {
1609 Label skip;
1610 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1611 __ cmp(G3_scratch, SP);
1612 __ brx(Assembler::equal, false, Assembler::pt, skip);
1613 __ delayed()->nop();
1614 __ stop("SP not restored to frame bottom");
1615 __ bind(skip);
1616 }
1617 #endif
1619 VALIDATE_STATE(G3_scratch, 4);
1620 __ set_last_Java_frame(SP, noreg);
1621 __ mov(Lstate, O0); // (arg) pointer to current state
1623 __ call(CAST_FROM_FN_PTR(address,
1624 JvmtiExport::can_post_interpreter_events() ?
1625 BytecodeInterpreter::runWithChecks
1626 : BytecodeInterpreter::run),
1627 relocInfo::runtime_call_type);
1629 __ delayed()->nop();
1631 __ ld_ptr(STATE(_thread), G2_thread);
1632 __ reset_last_Java_frame();
1634 // examine msg from interpreter to determine next action
1635 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1637 __ ld(STATE(_msg), L1_scratch); // Get new message
1639 Label call_method;
1640 Label return_from_interpreted_method;
1641 Label throw_exception;
1642 Label do_OSR;
1643 Label bad_msg;
1644 Label resume_interpreter;
1646 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1647 __ br(Assembler::equal, false, Assembler::pt, call_method);
1648 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1649 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1650 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1651 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1652 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1653 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1654 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1655 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1657 // Allocate more monitor space, shuffle expression stack....
1659 generate_more_monitors();
1661 // new monitor slot allocated, resume the interpreter.
1663 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1664 VALIDATE_STATE(G3_scratch, 5);
1665 __ ba(call_interpreter);
1666 __ delayed()->st(L1_scratch, STATE(_msg));
1668 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1669 unctrap_frame_manager_entry = __ pc();
1671 // QQQ what message do we send
1673 __ ba(call_interpreter);
1674 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1676 //=============================================================================
1677 // Returning from a compiled method into a deopted method. The bytecode at the
1678 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1679 // for the template based interpreter). Any stack space that was used by the
1680 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1681 // so all that we have to do is place any pending result on the expression stack
1682 // and resume execution on the next bytecode.
1684 generate_deopt_handling();
1686 // ready to resume the interpreter
1688 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1689 __ ba(call_interpreter);
1690 __ delayed()->st(L1_scratch, STATE(_msg));
1692 // Current frame has caught an exception we need to dispatch to the
1693 // handler. We can get here because a native interpreter frame caught
1694 // an exception in which case there is no handler and we must rethrow
1695 // If it is a vanilla interpreted frame the we simply drop into the
1696 // interpreter and let it do the lookup.
1698 Interpreter::_rethrow_exception_entry = __ pc();
1700 Label return_with_exception;
1701 Label unwind_and_forward;
1703 // O0: exception
1704 // O7: throwing pc
1706 // We want exception in the thread no matter what we ultimately decide about frame type.
1708 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1709 __ verify_thread();
1710 __ st_ptr(O0, exception_addr);
1712 // get the Method*
1713 __ ld_ptr(STATE(_method), G5_method);
1715 // if this current frame vanilla or native?
1717 __ ld(access_flags, Gtmp1);
1718 __ btst(JVM_ACC_NATIVE, Gtmp1);
1719 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1720 __ delayed()->nop();
1722 // We drop thru to unwind a native interpreted frame with a pending exception
1723 // We jump here for the initial interpreter frame with exception pending
1724 // We unwind the current acivation and forward it to our caller.
1726 __ bind(unwind_and_forward);
1728 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1729 // as expected by forward_exception.
1731 __ restore(FP, G0, SP); // unwind interpreter state frame
1732 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1733 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1735 // Return point from a call which returns a result in the native abi
1736 // (c1/c2/jni-native). This result must be processed onto the java
1737 // expression stack.
1738 //
1739 // A pending exception may be present in which case there is no result present
1741 address return_from_native_method = __ pc();
1743 VALIDATE_STATE(G3_scratch, 6);
1745 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1746 // stack is in the state that the calling convention left it.
1747 // Copy the result from native abi result and place it on java expression stack.
1749 // Current interpreter state is present in Lstate
1751 // Exception pending?
1753 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1754 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1755 __ tst(Lscratch); // exception pending?
1756 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1757 __ delayed()->nop();
1759 // Process the native abi result to java expression stack
1761 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1762 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1763 // get parameter size
1764 __ ld_ptr(L4_scratch, in_bytes(Method::const_offset()), L2_scratch);
1765 __ lduh(L2_scratch, in_bytes(ConstMethod::size_of_parameters_offset()), L2_scratch);
1766 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1767 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1768 __ ld(L4_scratch, in_bytes(Method::result_index_offset()), L3_scratch); // called method result type index
1770 // tosca is really just native abi
1771 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1772 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1773 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1774 __ jmpl(Lscratch, G0, O7); // and convert it
1775 __ delayed()->nop();
1777 // L1_scratch points to top of stack (prepushed)
1779 __ ba(resume_interpreter);
1780 __ delayed()->mov(L1_scratch, O1);
1782 // An exception is being caught on return to a vanilla interpreter frame.
1783 // Empty the stack and resume interpreter
1785 __ bind(return_with_exception);
1787 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1788 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1789 __ ba(resume_interpreter);
1790 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1792 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1793 // interpreter call, or native) and unwind this interpreter activation.
1794 // All monitors should be unlocked.
1796 __ bind(return_from_interpreted_method);
1798 VALIDATE_STATE(G3_scratch, 7);
1800 Label return_to_initial_caller;
1802 // Interpreted result is on the top of the completed activation expression stack.
1803 // We must return it to the top of the callers stack if caller was interpreted
1804 // otherwise we convert to native abi result and return to call_stub/c1/c2
1805 // The caller's expression stack was truncated by the call however the current activation
1806 // has enough stuff on the stack that we have usable space there no matter what. The
1807 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1808 // for the current activation
1810 __ ld_ptr(STATE(_prev_link), L1_scratch);
1811 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1812 __ ld(L2_scratch, in_bytes(Method::result_index_offset()), L2_scratch);
1813 __ tst(L1_scratch);
1814 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1815 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1817 // Copy result to callers java stack
1819 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1820 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1821 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1822 __ ld_ptr(STATE(_locals), O1); // stack destination
1824 // O0 - will be source, O1 - will be destination (preserved)
1825 __ jmpl(Lscratch, G0, O7); // and convert it
1826 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1828 // O1 == &locals[0]
1830 // Result is now on caller's stack. Just unwind current activation and resume
1832 Label unwind_recursive_activation;
1835 __ bind(unwind_recursive_activation);
1837 // O1 == &locals[0] (really callers stacktop) for activation now returning
1838 // returning to interpreter method from "recursive" interpreter call
1839 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1840 // to. Now all we must do is unwind the state from the completed call
1842 // Must restore stack
1843 VALIDATE_STATE(G3_scratch, 8);
1845 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1846 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1847 // frame and tell interpreter to resume.
1850 __ mov(O1, I1); // pass back new stack top across activation
1851 // POP FRAME HERE ==================================
1852 __ restore(FP, G0, SP); // unwind interpreter state frame
1853 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1856 // Resume the interpreter. The current frame contains the current interpreter
1857 // state object.
1858 //
1859 // O1 == new java stack pointer
1861 __ bind(resume_interpreter);
1862 VALIDATE_STATE(G3_scratch, 10);
1864 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1866 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1867 __ st(L1_scratch, STATE(_msg));
1868 __ ba(call_interpreter_2);
1869 __ delayed()->st_ptr(O1, STATE(_stack));
1872 // Fast accessor methods share this entry point.
1873 // This works because frame manager is in the same codelet
1874 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1875 // we need to do a little register fixup here once we distinguish the two of them
1876 if (UseFastAccessorMethods && !synchronized) {
1877 // Call stub_return address still in O7
1878 __ bind(fast_accessor_slow_entry_path);
1879 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1880 __ cmp(Gtmp1, O7); // returning to interpreter?
1881 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1882 __ delayed()->nop();
1883 __ ba(re_dispatch);
1884 __ delayed()->mov(G0, prevState); // initial entry
1886 }
1888 // interpreter returning to native code (call_stub/c1/c2)
1889 // convert result and unwind initial activation
1890 // L2_scratch - scaled result type index
1892 __ bind(return_to_initial_caller);
1894 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1895 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1896 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1897 __ jmpl(Lscratch, G0, O7); // and convert it
1898 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1900 Label unwind_initial_activation;
1901 __ bind(unwind_initial_activation);
1903 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1904 // we can return here with an exception that wasn't handled by interpreted code
1905 // how does c1/c2 see it on return?
1907 // compute resulting sp before/after args popped depending upon calling convention
1908 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1909 //
1910 // POP FRAME HERE ==================================
1911 __ restore(FP, G0, SP);
1912 __ retl();
1913 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1915 // OSR request, unwind the current frame and transfer to the OSR entry
1916 // and enter OSR nmethod
1918 __ bind(do_OSR);
1919 Label remove_initial_frame;
1920 __ ld_ptr(STATE(_prev_link), L1_scratch);
1921 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1923 // We are going to pop this frame. Is there another interpreter frame underneath
1924 // it or is it callstub/compiled?
1926 __ tst(L1_scratch);
1927 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1928 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1930 // Frame underneath is an interpreter frame simply unwind
1931 // POP FRAME HERE ==================================
1932 __ restore(FP, G0, SP); // unwind interpreter state frame
1933 __ mov(I5_savedSP->after_restore(), SP);
1935 // Since we are now calling native need to change our "return address" from the
1936 // dummy RecursiveInterpreterActivation to a return from native
1938 __ set((intptr_t)return_from_native_method - 8, O7);
1940 __ jmpl(G3_scratch, G0, G0);
1941 __ delayed()->mov(G1_scratch, O0);
1943 __ bind(remove_initial_frame);
1945 // POP FRAME HERE ==================================
1946 __ restore(FP, G0, SP);
1947 __ mov(I5_savedSP->after_restore(), SP);
1948 __ jmpl(G3_scratch, G0, G0);
1949 __ delayed()->mov(G1_scratch, O0);
1951 // Call a new method. All we do is (temporarily) trim the expression stack
1952 // push a return address to bring us back to here and leap to the new entry.
1953 // At this point we have a topmost frame that was allocated by the frame manager
1954 // which contains the current method interpreted state. We trim this frame
1955 // of excess java expression stack entries and then recurse.
1957 __ bind(call_method);
1959 // stack points to next free location and not top element on expression stack
1960 // method expects sp to be pointing to topmost element
1962 __ ld_ptr(STATE(_thread), G2_thread);
1963 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1966 // SP already takes in to account the 2 extra words we use for slop
1967 // when we call a "static long no_params()" method. So if
1968 // we trim back sp by the amount of unused java expression stack
1969 // there will be automagically the 2 extra words we need.
1970 // We also have to worry about keeping SP aligned.
1972 __ ld_ptr(STATE(_stack), Gargs);
1973 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1975 // compute the unused java stack size
1976 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1978 // Round down the unused space to that stack is always 16-byte aligned
1979 // by making the unused space a multiple of the size of two longs.
1981 __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1983 // Now trim the stack
1984 __ add(SP, L2_scratch, SP);
1987 // Now point to the final argument (account for prepush)
1988 __ add(Gargs, wordSize, Gargs);
1989 #ifdef ASSERT
1990 // Make sure we have space for the window
1991 __ sub(Gargs, SP, L1_scratch);
1992 __ cmp(L1_scratch, 16*wordSize);
1993 {
1994 Label skip;
1995 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1996 __ delayed()->nop();
1997 __ stop("killed stack");
1998 __ bind(skip);
1999 }
2000 #endif // ASSERT
2002 // Create a new frame where we can store values that make it look like the interpreter
2003 // really recursed.
2005 // prepare to recurse or call specialized entry
2007 // First link the registers we need
2009 // make the pc look good in debugger
2010 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
2011 // argument too
2012 __ mov(Lstate, I0);
2014 // Record our sending SP
2015 __ mov(SP, O5_savedSP);
2017 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
2018 __ set((intptr_t) entry_point, L1_scratch);
2019 __ cmp(L1_scratch, L2_scratch);
2020 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
2021 __ delayed()->mov(Lstate, prevState); // link activations
2023 // method uses specialized entry, push a return so we look like call stub setup
2024 // this path will handle fact that result is returned in registers and not
2025 // on the java stack.
2027 __ set((intptr_t)return_from_native_method - 8, O7);
2028 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
2029 __ delayed()->nop();
2031 //
2032 // Bad Message from interpreter
2033 //
2034 __ bind(bad_msg);
2035 __ stop("Bad message from interpreter");
2037 // Interpreted method "returned" with an exception pass it on...
2038 // Pass result, unwind activation and continue/return to interpreter/call_stub
2039 // We handle result (if any) differently based on return to interpreter or call_stub
2041 __ bind(throw_exception);
2042 __ ld_ptr(STATE(_prev_link), L1_scratch);
2043 __ tst(L1_scratch);
2044 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2045 __ delayed()->nop();
2047 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
2048 __ ba(unwind_recursive_activation);
2049 __ delayed()->nop();
2051 interpreter_frame_manager = entry_point;
2052 return entry_point;
2053 }
2055 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2056 : CppInterpreterGenerator(code) {
2057 generate_all(); // down here so it can be "virtual"
2058 }
2061 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2063 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2064 // expression stack, the callee will have callee_extra_locals (so we can account for
2065 // frame extension) and monitor_size for monitors. Basically we need to calculate
2066 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2067 //
2068 //
2069 // The big complicating thing here is that we must ensure that the stack stays properly
2070 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2071 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2072 // the caller so we must ensure that it is properly aligned for our callee.
2073 //
2074 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2075 // stack at all times to deal with the "stack long no_params()" method issue. This
2076 // is "slop_factor" here.
2077 const int slop_factor = 2;
2079 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2080 frame::memory_parameter_word_sp_offset; // register save area + param window
2081 const int extra_stack = 0; //6815692//Method::extra_stack_entries();
2082 return (round_to(max_stack +
2083 extra_stack +
2084 slop_factor +
2085 fixed_size +
2086 monitor_size +
2087 (callee_extra_locals * Interpreter::stackElementWords), WordsPerLong));
2089 }
2091 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
2093 // See call_stub code
2094 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2095 WordsPerLong); // 7 + register save area
2097 // Save space for one monitor to get into the interpreted method in case
2098 // the method is synchronized
2099 int monitor_size = method->is_synchronized() ?
2100 1*frame::interpreter_frame_monitor_size() : 0;
2101 return size_activation_helper(method->max_locals(), method->max_stack(),
2102 monitor_size) + call_stub_size;
2103 }
2105 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2106 frame* caller,
2107 frame* current,
2108 Method* method,
2109 intptr_t* locals,
2110 intptr_t* stack,
2111 intptr_t* stack_base,
2112 intptr_t* monitor_base,
2113 intptr_t* frame_bottom,
2114 bool is_top_frame
2115 )
2116 {
2117 // What about any vtable?
2118 //
2119 to_fill->_thread = JavaThread::current();
2120 // This gets filled in later but make it something recognizable for now
2121 to_fill->_bcp = method->code_base();
2122 to_fill->_locals = locals;
2123 to_fill->_constants = method->constants()->cache();
2124 to_fill->_method = method;
2125 to_fill->_mdx = NULL;
2126 to_fill->_stack = stack;
2127 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2128 to_fill->_msg = deopt_resume2;
2129 } else {
2130 to_fill->_msg = method_resume;
2131 }
2132 to_fill->_result._to_call._bcp_advance = 0;
2133 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2134 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2135 to_fill->_prev_link = NULL;
2137 // Fill in the registers for the frame
2139 // Need to install _sender_sp. Actually not too hard in C++!
2140 // When the skeletal frames are layed out we fill in a value
2141 // for _sender_sp. That value is only correct for the oldest
2142 // skeletal frame constructed (because there is only a single
2143 // entry for "caller_adjustment". While the skeletal frames
2144 // exist that is good enough. We correct that calculation
2145 // here and get all the frames correct.
2147 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2149 *current->register_addr(Lstate) = (intptr_t) to_fill;
2150 // skeletal already places a useful value here and this doesn't account
2151 // for alignment so don't bother.
2152 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2154 if (caller->is_interpreted_frame()) {
2155 interpreterState prev = caller->get_interpreterState();
2156 to_fill->_prev_link = prev;
2157 // Make the prev callee look proper
2158 prev->_result._to_call._callee = method;
2159 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2160 prev->_result._to_call._bcp_advance = 5;
2161 } else {
2162 prev->_result._to_call._bcp_advance = 3;
2163 }
2164 }
2165 to_fill->_oop_temp = NULL;
2166 to_fill->_stack_base = stack_base;
2167 // Need +1 here because stack_base points to the word just above the first expr stack entry
2168 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2169 // See generate_compute_interpreter_state.
2170 int extra_stack = 0; //6815692//Method::extra_stack_entries();
2171 to_fill->_stack_limit = stack_base - (method->max_stack() + 1 + extra_stack);
2172 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2174 // sparc specific
2175 to_fill->_frame_bottom = frame_bottom;
2176 to_fill->_self_link = to_fill;
2177 #ifdef ASSERT
2178 to_fill->_native_fresult = 123456.789;
2179 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2180 #endif
2181 }
2183 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2184 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2185 }
2188 int AbstractInterpreter::layout_activation(Method* method,
2189 int tempcount, // Number of slots on java expression stack in use
2190 int popframe_extra_args,
2191 int moncount, // Number of active monitors
2192 int caller_actual_parameters,
2193 int callee_param_size,
2194 int callee_locals_size,
2195 frame* caller,
2196 frame* interpreter_frame,
2197 bool is_top_frame,
2198 bool is_bottom_frame) {
2200 assert(popframe_extra_args == 0, "NEED TO FIX");
2201 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2202 // does as far as allocating an interpreter frame.
2203 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2204 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2205 // as determined by a previous call to this method.
2206 // It is also guaranteed to be walkable even though it is in a skeletal state
2207 // NOTE: return size is in words not bytes
2208 // NOTE: tempcount is the current size of the java expression stack. For top most
2209 // frames we will allocate a full sized expression stack and not the curback
2210 // version that non-top frames have.
2212 // Calculate the amount our frame will be adjust by the callee. For top frame
2213 // this is zero.
2215 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2216 // calculates the extra locals based on itself. Not what the callee does
2217 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2218 // as getting sender_sp correct.
2220 int extra_locals_size = callee_locals_size - callee_param_size;
2221 int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2222 int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2223 int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2224 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2227 /*
2228 if we actually have a frame to layout we must now fill in all the pieces. This means both
2229 the interpreterState and the registers.
2230 */
2231 if (interpreter_frame != NULL) {
2233 // MUCHO HACK
2235 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2236 // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2237 assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2238 frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2240 /* Now fillin the interpreterState object */
2242 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2245 intptr_t* locals;
2247 // Calculate the postion of locals[0]. This is painful because of
2248 // stack alignment (same as ia64). The problem is that we can
2249 // not compute the location of locals from fp(). fp() will account
2250 // for the extra locals but it also accounts for aligning the stack
2251 // and we can't determine if the locals[0] was misaligned but max_locals
2252 // was enough to have the
2253 // calculate postion of locals. fp already accounts for extra locals.
2254 // +2 for the static long no_params() issue.
2256 if (caller->is_interpreted_frame()) {
2257 // locals must agree with the caller because it will be used to set the
2258 // caller's tos when we return.
2259 interpreterState prev = caller->get_interpreterState();
2260 // stack() is prepushed.
2261 locals = prev->stack() + method->size_of_parameters();
2262 } else {
2263 // Lay out locals block in the caller adjacent to the register window save area.
2264 //
2265 // Compiled frames do not allocate a varargs area which is why this if
2266 // statement is needed.
2267 //
2268 intptr_t* fp = interpreter_frame->fp();
2269 int local_words = method->max_locals() * Interpreter::stackElementWords;
2271 if (caller->is_compiled_frame()) {
2272 locals = fp + frame::register_save_words + local_words - 1;
2273 } else {
2274 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2275 }
2277 }
2278 // END MUCHO HACK
2280 intptr_t* monitor_base = (intptr_t*) cur_state;
2281 intptr_t* stack_base = monitor_base - monitor_size;
2282 /* +1 because stack is always prepushed */
2283 intptr_t* stack = stack_base - (tempcount + 1);
2286 BytecodeInterpreter::layout_interpreterState(cur_state,
2287 caller,
2288 interpreter_frame,
2289 method,
2290 locals,
2291 stack,
2292 stack_base,
2293 monitor_base,
2294 frame_bottom,
2295 is_top_frame);
2297 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2299 }
2300 return frame_words;
2301 }
2303 #endif // CC_INTERP