Mon, 11 Oct 2010 04:18:58 -0700
6829194: JSR 292 needs to support compressed oops
Reviewed-by: kvn, jrose
1 /*
2 * Copyright (c) 2007, 2009, 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 "incls/_precompiled.incl"
26 #include "incls/_cppInterpreter_sparc.cpp.incl"
28 #ifdef CC_INTERP
30 // Routine exists to make tracebacks look decent in debugger
31 // while "shadow" interpreter frames are on stack. It is also
32 // used to distinguish interpreter frames.
34 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
35 ShouldNotReachHere();
36 }
38 bool CppInterpreter::contains(address pc) {
39 return ( _code->contains(pc) ||
40 ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
41 }
43 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
44 #define __ _masm->
46 Label frame_manager_entry;
47 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized
48 // c++ interpreter entry point this holds that entry point label.
50 static address unctrap_frame_manager_entry = NULL;
52 static address interpreter_return_address = NULL;
53 static address deopt_frame_manager_return_atos = NULL;
54 static address deopt_frame_manager_return_btos = NULL;
55 static address deopt_frame_manager_return_itos = NULL;
56 static address deopt_frame_manager_return_ltos = NULL;
57 static address deopt_frame_manager_return_ftos = NULL;
58 static address deopt_frame_manager_return_dtos = NULL;
59 static address deopt_frame_manager_return_vtos = NULL;
61 const Register prevState = G1_scratch;
63 void InterpreterGenerator::save_native_result(void) {
64 // result potentially in O0/O1: save it across calls
65 __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
66 #ifdef _LP64
67 __ stx(O0, STATE(_native_lresult));
68 #else
69 __ std(O0, STATE(_native_lresult));
70 #endif
71 }
73 void InterpreterGenerator::restore_native_result(void) {
75 // Restore any method result value
76 __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
77 #ifdef _LP64
78 __ ldx(STATE(_native_lresult), O0);
79 #else
80 __ ldd(STATE(_native_lresult), O0);
81 #endif
82 }
84 // A result handler converts/unboxes a native call result into
85 // a java interpreter/compiler result. The current frame is an
86 // interpreter frame. The activation frame unwind code must be
87 // consistent with that of TemplateTable::_return(...). In the
88 // case of native methods, the caller's SP was not modified.
89 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
90 address entry = __ pc();
91 Register Itos_i = Otos_i ->after_save();
92 Register Itos_l = Otos_l ->after_save();
93 Register Itos_l1 = Otos_l1->after_save();
94 Register Itos_l2 = Otos_l2->after_save();
95 switch (type) {
96 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
97 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
98 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
99 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
100 case T_LONG :
101 #ifndef _LP64
102 __ mov(O1, Itos_l2); // move other half of long
103 #endif // ifdef or no ifdef, fall through to the T_INT case
104 case T_INT : __ mov(O0, Itos_i); break;
105 case T_VOID : /* nothing to do */ break;
106 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
107 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
108 case T_OBJECT :
109 __ ld_ptr(STATE(_oop_temp), Itos_i);
110 __ verify_oop(Itos_i);
111 break;
112 default : ShouldNotReachHere();
113 }
114 __ ret(); // return from interpreter activation
115 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
116 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
117 return entry;
118 }
120 // tosca based result to c++ interpreter stack based result.
121 // Result goes to address in L1_scratch
123 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
124 // A result is in the native abi result register from a native method call.
125 // We need to return this result to the interpreter by pushing the result on the interpreter's
126 // stack. This is relatively simple the destination is in L1_scratch
127 // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
128 // adjust L1_scratch
129 address entry = __ pc();
130 switch (type) {
131 case T_BOOLEAN:
132 // !0 => true; 0 => false
133 __ subcc(G0, O0, G0);
134 __ addc(G0, 0, O0);
135 __ st(O0, L1_scratch, 0);
136 __ sub(L1_scratch, wordSize, L1_scratch);
137 break;
139 // cannot use and3, 0xFFFF too big as immediate value!
140 case T_CHAR :
141 __ sll(O0, 16, O0);
142 __ srl(O0, 16, O0);
143 __ st(O0, L1_scratch, 0);
144 __ sub(L1_scratch, wordSize, L1_scratch);
145 break;
147 case T_BYTE :
148 __ sll(O0, 24, O0);
149 __ sra(O0, 24, O0);
150 __ st(O0, L1_scratch, 0);
151 __ sub(L1_scratch, wordSize, L1_scratch);
152 break;
154 case T_SHORT :
155 __ sll(O0, 16, O0);
156 __ sra(O0, 16, O0);
157 __ st(O0, L1_scratch, 0);
158 __ sub(L1_scratch, wordSize, L1_scratch);
159 break;
160 case T_LONG :
161 #ifndef _LP64
162 #if defined(COMPILER2)
163 // All return values are where we want them, except for Longs. C2 returns
164 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
165 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
166 // build even if we are returning from interpreted we just do a little
167 // stupid shuffing.
168 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
169 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
170 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
171 __ stx(G1, L1_scratch, -wordSize);
172 #else
173 // native result is in O0, O1
174 __ st(O1, L1_scratch, 0); // Low order
175 __ st(O0, L1_scratch, -wordSize); // High order
176 #endif /* COMPILER2 */
177 #else
178 __ stx(O0, L1_scratch, -wordSize);
179 #endif
180 __ sub(L1_scratch, 2*wordSize, L1_scratch);
181 break;
183 case T_INT :
184 __ st(O0, L1_scratch, 0);
185 __ sub(L1_scratch, wordSize, L1_scratch);
186 break;
188 case T_VOID : /* nothing to do */
189 break;
191 case T_FLOAT :
192 __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
193 __ sub(L1_scratch, wordSize, L1_scratch);
194 break;
196 case T_DOUBLE :
197 // Every stack slot is aligned on 64 bit, However is this
198 // the correct stack slot on 64bit?? QQQ
199 __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
200 __ sub(L1_scratch, 2*wordSize, L1_scratch);
201 break;
202 case T_OBJECT :
203 __ verify_oop(O0);
204 __ st_ptr(O0, L1_scratch, 0);
205 __ sub(L1_scratch, wordSize, L1_scratch);
206 break;
207 default : ShouldNotReachHere();
208 }
209 __ retl(); // return from interpreter activation
210 __ delayed()->nop(); // schedule this better
211 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
212 return entry;
213 }
215 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
216 // A result is in the java expression stack of the interpreted method that has just
217 // returned. Place this result on the java expression stack of the caller.
218 //
219 // The current interpreter activation in Lstate is for the method just returning its
220 // result. So we know that the result of this method is on the top of the current
221 // execution stack (which is pre-pushed) and will be return to the top of the caller
222 // stack. The top of the callers stack is the bottom of the locals of the current
223 // activation.
224 // Because of the way activation are managed by the frame manager the value of esp is
225 // below both the stack top of the current activation and naturally the stack top
226 // of the calling activation. This enable this routine to leave the return address
227 // to the frame manager on the stack and do a vanilla return.
228 //
229 // On entry: O0 - points to source (callee stack top)
230 // O1 - points to destination (caller stack top [i.e. free location])
231 // destroys O2, O3
232 //
234 address entry = __ pc();
235 switch (type) {
236 case T_VOID: break;
237 break;
238 case T_FLOAT :
239 case T_BOOLEAN:
240 case T_CHAR :
241 case T_BYTE :
242 case T_SHORT :
243 case T_INT :
244 // 1 word result
245 __ ld(O0, 0, O2);
246 __ st(O2, O1, 0);
247 __ sub(O1, wordSize, O1);
248 break;
249 case T_DOUBLE :
250 case T_LONG :
251 // return top two words on current expression stack to caller's expression stack
252 // The caller's expression stack is adjacent to the current frame manager's intepretState
253 // except we allocated one extra word for this intepretState so we won't overwrite it
254 // when we return a two word result.
255 #ifdef _LP64
256 __ ld_ptr(O0, 0, O2);
257 __ st_ptr(O2, O1, -wordSize);
258 #else
259 __ ld(O0, 0, O2);
260 __ ld(O0, wordSize, O3);
261 __ st(O3, O1, 0);
262 __ st(O2, O1, -wordSize);
263 #endif
264 __ sub(O1, 2*wordSize, O1);
265 break;
266 case T_OBJECT :
267 __ ld_ptr(O0, 0, O2);
268 __ verify_oop(O2); // verify it
269 __ st_ptr(O2, O1, 0);
270 __ sub(O1, wordSize, O1);
271 break;
272 default : ShouldNotReachHere();
273 }
274 __ retl();
275 __ delayed()->nop(); // QQ schedule this better
276 return entry;
277 }
279 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
280 // A result is in the java expression stack of the interpreted method that has just
281 // returned. Place this result in the native abi that the caller expects.
282 // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
283 //
284 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
285 // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
286 // and so rather than return result onto caller's java expression stack we return the
287 // result in the expected location based on the native abi.
288 // On entry: O0 - source (stack top)
289 // On exit result in expected output register
290 // QQQ schedule this better
292 address entry = __ pc();
293 switch (type) {
294 case T_VOID: break;
295 break;
296 case T_FLOAT :
297 __ ldf(FloatRegisterImpl::S, O0, 0, F0);
298 break;
299 case T_BOOLEAN:
300 case T_CHAR :
301 case T_BYTE :
302 case T_SHORT :
303 case T_INT :
304 // 1 word result
305 __ ld(O0, 0, O0->after_save());
306 break;
307 case T_DOUBLE :
308 __ ldf(FloatRegisterImpl::D, O0, 0, F0);
309 break;
310 case T_LONG :
311 // return top two words on current expression stack to caller's expression stack
312 // The caller's expression stack is adjacent to the current frame manager's interpretState
313 // except we allocated one extra word for this intepretState so we won't overwrite it
314 // when we return a two word result.
315 #ifdef _LP64
316 __ ld_ptr(O0, 0, O0->after_save());
317 #else
318 __ ld(O0, wordSize, O1->after_save());
319 __ ld(O0, 0, O0->after_save());
320 #endif
321 #if defined(COMPILER2) && !defined(_LP64)
322 // C2 expects long results in G1 we can't tell if we're returning to interpreted
323 // or compiled so just be safe use G1 and O0/O1
325 // Shift bits into high (msb) of G1
326 __ sllx(Otos_l1->after_save(), 32, G1);
327 // Zero extend low bits
328 __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
329 __ or3 (Otos_l2->after_save(), G1, G1);
330 #endif /* COMPILER2 */
331 break;
332 case T_OBJECT :
333 __ ld_ptr(O0, 0, O0->after_save());
334 __ verify_oop(O0->after_save()); // verify it
335 break;
336 default : ShouldNotReachHere();
337 }
338 __ retl();
339 __ delayed()->nop();
340 return entry;
341 }
343 address CppInterpreter::return_entry(TosState state, int length) {
344 // make it look good in the debugger
345 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
346 }
348 address CppInterpreter::deopt_entry(TosState state, int length) {
349 address ret = NULL;
350 if (length != 0) {
351 switch (state) {
352 case atos: ret = deopt_frame_manager_return_atos; break;
353 case btos: ret = deopt_frame_manager_return_btos; break;
354 case ctos:
355 case stos:
356 case itos: ret = deopt_frame_manager_return_itos; break;
357 case ltos: ret = deopt_frame_manager_return_ltos; break;
358 case ftos: ret = deopt_frame_manager_return_ftos; break;
359 case dtos: ret = deopt_frame_manager_return_dtos; break;
360 case vtos: ret = deopt_frame_manager_return_vtos; break;
361 }
362 } else {
363 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap)
364 }
365 assert(ret != NULL, "Not initialized");
366 return ret;
367 }
369 //
370 // Helpers for commoning out cases in the various type of method entries.
371 //
373 // increment invocation count & check for overflow
374 //
375 // Note: checking for negative value instead of overflow
376 // so we have a 'sticky' overflow test
377 //
378 // Lmethod: method
379 // ??: invocation counter
380 //
381 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
382 // Update standard invocation counters
383 __ increment_invocation_counter(O0, G3_scratch);
384 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
385 __ ld_ptr(STATE(_method), G3_scratch);
386 Address interpreter_invocation_counter(G3_scratch, 0, in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
387 __ ld(interpreter_invocation_counter, G3_scratch);
388 __ inc(G3_scratch);
389 __ st(G3_scratch, interpreter_invocation_counter);
390 }
392 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
393 __ sethi(invocation_limit);
394 __ ld(invocation_limit, G3_scratch);
395 __ cmp(O0, G3_scratch);
396 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
397 __ delayed()->nop();
399 }
401 address InterpreterGenerator::generate_empty_entry(void) {
403 // A method that does nothing but return...
405 address entry = __ pc();
406 Label slow_path;
408 __ verify_oop(G5_method);
410 // do nothing for empty methods (do not even increment invocation counter)
411 if ( UseFastEmptyMethods) {
412 // If we need a safepoint check, generate full interpreter entry.
413 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
414 __ load_contents(sync_state, G3_scratch);
415 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
416 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
417 __ delayed()->nop();
419 // Code: _return
420 __ retl();
421 __ delayed()->mov(O5_savedSP, SP);
422 return entry;
423 }
424 return NULL;
425 }
427 // Call an accessor method (assuming it is resolved, otherwise drop into
428 // vanilla (slow path) entry
430 // Generates code to elide accessor methods
431 // Uses G3_scratch and G1_scratch as scratch
432 address InterpreterGenerator::generate_accessor_entry(void) {
434 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
435 // parameter size = 1
436 // Note: We can only use this code if the getfield has been resolved
437 // and if we don't have a null-pointer exception => check for
438 // these conditions first and use slow path if necessary.
439 address entry = __ pc();
440 Label slow_path;
442 if ( UseFastAccessorMethods) {
443 // Check if we need to reach a safepoint and generate full interpreter
444 // frame if so.
445 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
446 __ load_contents(sync_state, G3_scratch);
447 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
448 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
449 __ delayed()->nop();
451 // Check if local 0 != NULL
452 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
453 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
454 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
455 __ delayed()->nop();
458 // read first instruction word and extract bytecode @ 1 and index @ 2
459 // get first 4 bytes of the bytecodes (big endian!)
460 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), G1_scratch);
461 __ ld(Address(G1_scratch, 0, in_bytes(constMethodOopDesc::codes_offset())), G1_scratch);
463 // move index @ 2 far left then to the right most two bytes.
464 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
465 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
466 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
468 // get constant pool cache
469 __ ld_ptr(G5_method, in_bytes(methodOopDesc::constants_offset()), G3_scratch);
470 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
472 // get specific constant pool cache entry
473 __ add(G3_scratch, G1_scratch, G3_scratch);
475 // Check the constant Pool cache entry to see if it has been resolved.
476 // If not, need the slow path.
477 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
478 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
479 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
480 __ and3(G1_scratch, 0xFF, G1_scratch);
481 __ cmp(G1_scratch, Bytecodes::_getfield);
482 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
483 __ delayed()->nop();
485 // Get the type and return field offset from the constant pool cache
486 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
487 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
489 Label xreturn_path;
490 // Need to differentiate between igetfield, agetfield, bgetfield etc.
491 // because they are different sizes.
492 // Get the type from the constant pool cache
493 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
494 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
495 ConstantPoolCacheEntry::verify_tosBits();
496 __ cmp(G1_scratch, atos );
497 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
498 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
499 __ cmp(G1_scratch, itos);
500 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
501 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
502 __ cmp(G1_scratch, stos);
503 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
504 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
505 __ cmp(G1_scratch, ctos);
506 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
507 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
508 #ifdef ASSERT
509 __ cmp(G1_scratch, btos);
510 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
511 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
512 __ should_not_reach_here();
513 #endif
514 __ ldsb(Otos_i, G3_scratch, Otos_i);
515 __ bind(xreturn_path);
517 // _ireturn/_areturn
518 __ retl(); // return from leaf routine
519 __ delayed()->mov(O5_savedSP, SP);
521 // Generate regular method entry
522 __ bind(slow_path);
523 __ ba(false, fast_accessor_slow_entry_path);
524 __ delayed()->nop();
525 return entry;
526 }
527 return NULL;
528 }
530 //
531 // Interpreter stub for calling a native method. (C++ interpreter)
532 // This sets up a somewhat different looking stack for calling the native method
533 // than the typical interpreter frame setup.
534 //
536 address InterpreterGenerator::generate_native_entry(bool synchronized) {
537 address entry = __ pc();
539 // the following temporary registers are used during frame creation
540 const Register Gtmp1 = G3_scratch ;
541 const Register Gtmp2 = G1_scratch;
542 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
544 bool inc_counter = UseCompiler || CountCompiledCalls;
546 // make sure registers are different!
547 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
549 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
551 Label Lentry;
552 __ bind(Lentry);
554 __ verify_oop(G5_method);
556 const Register Glocals_size = G3;
557 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
559 // make sure method is native & not abstract
560 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
561 #ifdef ASSERT
562 __ ld(access_flags, Gtmp1);
563 {
564 Label L;
565 __ btst(JVM_ACC_NATIVE, Gtmp1);
566 __ br(Assembler::notZero, false, Assembler::pt, L);
567 __ delayed()->nop();
568 __ stop("tried to execute non-native method as native");
569 __ bind(L);
570 }
571 { Label L;
572 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
573 __ br(Assembler::zero, false, Assembler::pt, L);
574 __ delayed()->nop();
575 __ stop("tried to execute abstract method as non-abstract");
576 __ bind(L);
577 }
578 #endif // ASSERT
580 __ lduh(size_of_parameters, Gtmp1);
581 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes
582 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord
583 // NEW
584 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
585 // generate the code to allocate the interpreter stack frame
586 // NEW FRAME ALLOCATED HERE
587 // save callers original sp
588 // __ mov(SP, I5_savedSP->after_restore());
590 generate_compute_interpreter_state(Lstate, G0, true);
592 // At this point Lstate points to new interpreter state
593 //
595 const Address do_not_unlock_if_synchronized(G2_thread, 0,
596 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
597 // Since at this point in the method invocation the exception handler
598 // would try to exit the monitor of synchronized methods which hasn't
599 // been entered yet, we set the thread local variable
600 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
601 // runtime, exception handling i.e. unlock_if_synchronized_method will
602 // check this thread local flag.
603 // This flag has two effects, one is to force an unwind in the topmost
604 // interpreter frame and not perform an unlock while doing so.
606 __ movbool(true, G3_scratch);
607 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
610 // increment invocation counter and check for overflow
611 //
612 // Note: checking for negative value instead of overflow
613 // so we have a 'sticky' overflow test (may be of
614 // importance as soon as we have true MT/MP)
615 Label invocation_counter_overflow;
616 if (inc_counter) {
617 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
618 }
619 Label Lcontinue;
620 __ bind(Lcontinue);
622 bang_stack_shadow_pages(true);
623 // reset the _do_not_unlock_if_synchronized flag
624 __ stbool(G0, do_not_unlock_if_synchronized);
626 // check for synchronized methods
627 // Must happen AFTER invocation_counter check, so method is not locked
628 // if counter overflows.
630 if (synchronized) {
631 lock_method();
632 // Don't see how G2_thread is preserved here...
633 // __ verify_thread(); QQQ destroys L0,L1 can't use
634 } else {
635 #ifdef ASSERT
636 { Label ok;
637 __ ld_ptr(STATE(_method), G5_method);
638 __ ld(access_flags, O0);
639 __ btst(JVM_ACC_SYNCHRONIZED, O0);
640 __ br( Assembler::zero, false, Assembler::pt, ok);
641 __ delayed()->nop();
642 __ stop("method needs synchronization");
643 __ bind(ok);
644 }
645 #endif // ASSERT
646 }
648 // start execution
650 // __ verify_thread(); kills L1,L2 can't use at the moment
652 // jvmti/jvmpi support
653 __ notify_method_entry();
655 // native call
657 // (note that O0 is never an oop--at most it is a handle)
658 // It is important not to smash any handles created by this call,
659 // until any oop handle in O0 is dereferenced.
661 // (note that the space for outgoing params is preallocated)
663 // get signature handler
665 Label pending_exception_present;
667 { Label L;
668 __ ld_ptr(STATE(_method), G5_method);
669 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
670 __ tst(G3_scratch);
671 __ brx(Assembler::notZero, false, Assembler::pt, L);
672 __ delayed()->nop();
673 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
674 __ ld_ptr(STATE(_method), G5_method);
676 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
677 __ ld_ptr(exception_addr, G3_scratch);
678 __ br_notnull(G3_scratch, false, Assembler::pn, pending_exception_present);
679 __ delayed()->nop();
680 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
681 __ bind(L);
682 }
684 // Push a new frame so that the args will really be stored in
685 // Copy a few locals across so the new frame has the variables
686 // we need but these values will be dead at the jni call and
687 // therefore not gc volatile like the values in the current
688 // frame (Lstate in particular)
690 // Flush the state pointer to the register save area
691 // Which is the only register we need for a stack walk.
692 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
694 __ mov(Lstate, O1); // Need to pass the state pointer across the frame
696 // Calculate current frame size
697 __ sub(SP, FP, O3); // Calculate negative of current frame size
698 __ save(SP, O3, SP); // Allocate an identical sized frame
700 __ mov(I1, Lstate); // In the "natural" register.
702 // Note I7 has leftover trash. Slow signature handler will fill it in
703 // should we get there. Normal jni call will set reasonable last_Java_pc
704 // below (and fix I7 so the stack trace doesn't have a meaningless frame
705 // in it).
708 // call signature handler
709 __ ld_ptr(STATE(_method), Lmethod);
710 __ ld_ptr(STATE(_locals), Llocals);
712 __ callr(G3_scratch, 0);
713 __ delayed()->nop();
714 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed)
716 { Label not_static;
718 __ ld_ptr(STATE(_method), G5_method);
719 __ ld(access_flags, O0);
720 __ btst(JVM_ACC_STATIC, O0);
721 __ br( Assembler::zero, false, Assembler::pt, not_static);
722 __ delayed()->
723 // get native function entry point(O0 is a good temp until the very end)
724 ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::native_function_offset())), O0);
725 // for static methods insert the mirror argument
726 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
728 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc:: constants_offset())), O1);
729 __ ld_ptr(Address(O1, 0, constantPoolOopDesc::pool_holder_offset_in_bytes()), O1);
730 __ ld_ptr(O1, mirror_offset, O1);
731 // where the mirror handle body is allocated:
732 #ifdef ASSERT
733 if (!PrintSignatureHandlers) // do not dirty the output with this
734 { Label L;
735 __ tst(O1);
736 __ brx(Assembler::notZero, false, Assembler::pt, L);
737 __ delayed()->nop();
738 __ stop("mirror is missing");
739 __ bind(L);
740 }
741 #endif // ASSERT
742 __ st_ptr(O1, STATE(_oop_temp));
743 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add
744 __ bind(not_static);
745 }
747 // At this point, arguments have been copied off of stack into
748 // their JNI positions, which are O1..O5 and SP[68..].
749 // Oops are boxed in-place on the stack, with handles copied to arguments.
750 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
752 #ifdef ASSERT
753 { Label L;
754 __ tst(O0);
755 __ brx(Assembler::notZero, false, Assembler::pt, L);
756 __ delayed()->nop();
757 __ stop("native entry point is missing");
758 __ bind(L);
759 }
760 #endif // ASSERT
762 //
763 // setup the java frame anchor
764 //
765 // The scavenge function only needs to know that the PC of this frame is
766 // in the interpreter method entry code, it doesn't need to know the exact
767 // PC and hence we can use O7 which points to the return address from the
768 // previous call in the code stream (signature handler function)
769 //
770 // The other trick is we set last_Java_sp to FP instead of the usual SP because
771 // we have pushed the extra frame in order to protect the volatile register(s)
772 // in that frame when we return from the jni call
773 //
776 __ set_last_Java_frame(FP, O7);
777 __ mov(O7, I7); // make dummy interpreter frame look like one above,
778 // not meaningless information that'll confuse me.
780 // flush the windows now. We don't care about the current (protection) frame
781 // only the outer frames
783 __ flush_windows();
785 // mark windows as flushed
786 Address flags(G2_thread,
787 0,
788 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
789 __ set(JavaFrameAnchor::flushed, G3_scratch);
790 __ st(G3_scratch, flags);
792 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
794 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
795 #ifdef ASSERT
796 { Label L;
797 __ ld(thread_state, G3_scratch);
798 __ cmp(G3_scratch, _thread_in_Java);
799 __ br(Assembler::equal, false, Assembler::pt, L);
800 __ delayed()->nop();
801 __ stop("Wrong thread state in native stub");
802 __ bind(L);
803 }
804 #endif // ASSERT
805 __ set(_thread_in_native, G3_scratch);
806 __ st(G3_scratch, thread_state);
808 // Call the jni method, using the delay slot to set the JNIEnv* argument.
809 __ callr(O0, 0);
810 __ delayed()->
811 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
812 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
814 // must we block?
816 // Block, if necessary, before resuming in _thread_in_Java state.
817 // In order for GC to work, don't clear the last_Java_sp until after blocking.
818 { Label no_block;
819 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
821 // Switch thread to "native transition" state before reading the synchronization state.
822 // This additional state is necessary because reading and testing the synchronization
823 // state is not atomic w.r.t. GC, as this scenario demonstrates:
824 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
825 // VM thread changes sync state to synchronizing and suspends threads for GC.
826 // Thread A is resumed to finish this native method, but doesn't block here since it
827 // didn't see any synchronization is progress, and escapes.
828 __ set(_thread_in_native_trans, G3_scratch);
829 __ st(G3_scratch, thread_state);
830 if(os::is_MP()) {
831 // Write serialization page so VM thread can do a pseudo remote membar.
832 // We use the current thread pointer to calculate a thread specific
833 // offset to write to within the page. This minimizes bus traffic
834 // due to cache line collision.
835 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
836 }
837 __ load_contents(sync_state, G3_scratch);
838 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
841 Label L;
842 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
843 __ br(Assembler::notEqual, false, Assembler::pn, L);
844 __ delayed()->
845 ld(suspend_state, G3_scratch);
846 __ cmp(G3_scratch, 0);
847 __ br(Assembler::equal, false, Assembler::pt, no_block);
848 __ delayed()->nop();
849 __ bind(L);
851 // Block. Save any potential method result value before the operation and
852 // use a leaf call to leave the last_Java_frame setup undisturbed.
853 save_native_result();
854 __ call_VM_leaf(noreg,
855 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
856 G2_thread);
857 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
858 // Restore any method result value
859 restore_native_result();
860 __ bind(no_block);
861 }
863 // Clear the frame anchor now
865 __ reset_last_Java_frame();
867 // Move the result handler address
868 __ mov(Lscratch, G3_scratch);
869 // return possible result to the outer frame
870 #ifndef __LP64
871 __ mov(O0, I0);
872 __ restore(O1, G0, O1);
873 #else
874 __ restore(O0, G0, O0);
875 #endif /* __LP64 */
877 // Move result handler to expected register
878 __ mov(G3_scratch, Lscratch);
881 // thread state is thread_in_native_trans. Any safepoint blocking has
882 // happened in the trampoline we are ready to switch to thread_in_Java.
884 __ set(_thread_in_Java, G3_scratch);
885 __ st(G3_scratch, thread_state);
887 // If we have an oop result store it where it will be safe for any further gc
888 // until we return now that we've released the handle it might be protected by
890 {
891 Label no_oop, store_result;
893 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
894 __ cmp(G3_scratch, Lscratch);
895 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
896 __ delayed()->nop();
897 __ addcc(G0, O0, O0);
898 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
899 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
900 __ mov(G0, O0);
902 __ bind(store_result);
903 // Store it where gc will look for it and result handler expects it.
904 __ st_ptr(O0, STATE(_oop_temp));
906 __ bind(no_oop);
908 }
910 // reset handle block
911 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
912 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
915 // handle exceptions (exception handling will handle unlocking!)
916 { Label L;
917 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
919 __ ld_ptr(exception_addr, Gtemp);
920 __ tst(Gtemp);
921 __ brx(Assembler::equal, false, Assembler::pt, L);
922 __ delayed()->nop();
923 __ bind(pending_exception_present);
924 // With c++ interpreter we just leave it pending caller will do the correct thing. However...
925 // Like x86 we ignore the result of the native call and leave the method locked. This
926 // seems wrong to leave things locked.
928 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
929 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
931 __ bind(L);
932 }
934 // jvmdi/jvmpi support (preserves thread register)
935 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
937 if (synchronized) {
938 // save and restore any potential method result value around the unlocking operation
939 save_native_result();
941 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
942 // Get the initial monitor we allocated
943 __ sub(Lstate, entry_size, O1); // initial monitor
944 __ unlock_object(O1);
945 restore_native_result();
946 }
948 #if defined(COMPILER2) && !defined(_LP64)
950 // C2 expects long results in G1 we can't tell if we're returning to interpreted
951 // or compiled so just be safe.
953 __ sllx(O0, 32, G1); // Shift bits into high G1
954 __ srl (O1, 0, O1); // Zero extend O1
955 __ or3 (O1, G1, G1); // OR 64 bits into G1
957 #endif /* COMPILER2 && !_LP64 */
959 #ifdef ASSERT
960 {
961 Label ok;
962 __ cmp(I5_savedSP, FP);
963 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
964 __ delayed()->nop();
965 __ stop("bad I5_savedSP value");
966 __ should_not_reach_here();
967 __ bind(ok);
968 }
969 #endif
970 // Calls result handler which POPS FRAME
971 if (TraceJumps) {
972 // Move target to register that is recordable
973 __ mov(Lscratch, G3_scratch);
974 __ JMP(G3_scratch, 0);
975 } else {
976 __ jmp(Lscratch, 0);
977 }
978 __ delayed()->nop();
980 if (inc_counter) {
981 // handle invocation counter overflow
982 __ bind(invocation_counter_overflow);
983 generate_counter_overflow(Lcontinue);
984 }
987 return entry;
988 }
990 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
991 const Register prev_state,
992 bool native) {
994 // On entry
995 // G5_method - caller's method
996 // Gargs - points to initial parameters (i.e. locals[0])
997 // G2_thread - valid? (C1 only??)
998 // "prev_state" - contains any previous frame manager state which we must save a link
999 //
1000 // On return
1001 // "state" is a pointer to the newly allocated state object. We must allocate and initialize
1002 // a new interpretState object and the method expression stack.
1004 assert_different_registers(state, prev_state);
1005 assert_different_registers(prev_state, G3_scratch);
1006 const Register Gtmp = G3_scratch;
1007 const Address constants (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1008 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1009 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1010 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1011 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1013 // slop factor is two extra slots on the expression stack so that
1014 // we always have room to store a result when returning from a call without parameters
1015 // that returns a result.
1017 const int slop_factor = 2*wordSize;
1019 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1020 //6815692//methodOopDesc::extra_stack_words() + // extra push slots for MH adapters
1021 frame::memory_parameter_word_sp_offset + // register save area + param window
1022 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1024 // XXX G5_method valid
1026 // Now compute new frame size
1028 if (native) {
1029 __ lduh( size_of_parameters, Gtmp );
1030 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words
1031 } else {
1032 __ lduh(max_stack, Gtmp); // Full size expression stack
1033 }
1034 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1036 __ neg(Gtmp); // negative space for stack/parameters in words
1037 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1038 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1040 // Need to do stack size check here before we fault on large frames
1042 Label stack_ok;
1044 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1045 (StackRedPages+StackYellowPages);
1048 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1049 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1050 // compute stack bottom
1051 __ sub(O0, O1, O0);
1053 // Avoid touching the guard pages
1054 // Also a fudge for frame size of BytecodeInterpreter::run
1055 // It varies from 1k->4k depending on build type
1056 const int fudge = 6 * K;
1058 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1060 __ add(O0, O1, O0);
1061 __ sub(O0, Gtmp, O0);
1062 __ cmp(SP, O0);
1063 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1064 __ delayed()->nop();
1066 // throw exception return address becomes throwing pc
1068 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1069 __ stop("never reached");
1071 __ bind(stack_ok);
1073 __ save(SP, Gtmp, SP); // setup new frame and register window
1075 // New window I7 call_stub or previous activation
1076 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1077 //
1078 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1079 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1081 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1083 // Initialize a new Interpreter state
1084 // orig_sp - caller's original sp
1085 // G2_thread - thread
1086 // Gargs - &locals[0] (unbiased?)
1087 // G5_method - method
1088 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1091 __ set(0xdead0004, O1);
1094 __ st_ptr(Gargs, XXX_STATE(_locals));
1095 __ st_ptr(G0, XXX_STATE(_oop_temp));
1097 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1098 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1099 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1101 if (native) {
1102 __ st_ptr(G0, XXX_STATE(_bcp));
1103 } else {
1104 __ ld_ptr(G5_method, in_bytes(methodOopDesc::const_offset()), O2); // get constMethodOop
1105 __ add(O2, in_bytes(constMethodOopDesc::codes_offset()), O2); // get bcp
1106 __ st_ptr(O2, XXX_STATE(_bcp));
1107 }
1109 __ st_ptr(G0, XXX_STATE(_mdx));
1110 __ st_ptr(G5_method, XXX_STATE(_method));
1112 __ set((int) BytecodeInterpreter::method_entry, O1);
1113 __ st(O1, XXX_STATE(_msg));
1115 __ ld_ptr(constants, O3);
1116 __ ld_ptr(O3, constantPoolOopDesc::cache_offset_in_bytes(), O2);
1117 __ st_ptr(O2, XXX_STATE(_constants));
1119 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1121 // Monitor base is just start of BytecodeInterpreter object;
1122 __ mov(state, O2);
1123 __ st_ptr(O2, XXX_STATE(_monitor_base));
1125 // Do we need a monitor for synchonized method?
1126 {
1127 __ ld(access_flags, O1);
1128 Label done;
1129 Label got_obj;
1130 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1131 __ br( Assembler::zero, false, Assembler::pt, done);
1133 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1134 __ delayed()->btst(JVM_ACC_STATIC, O1);
1135 __ ld_ptr(XXX_STATE(_locals), O1);
1136 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1137 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1138 __ ld_ptr(constants, O1);
1139 __ ld_ptr( O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
1140 // lock the mirror, not the klassOop
1141 __ ld_ptr( O1, mirror_offset, O1);
1143 __ bind(got_obj);
1145 #ifdef ASSERT
1146 __ tst(O1);
1147 __ breakpoint_trap(Assembler::zero);
1148 #endif // ASSERT
1150 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1151 __ sub(SP, entry_size, SP); // account for initial monitor
1152 __ sub(O2, entry_size, O2); // initial monitor
1153 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1154 __ bind(done);
1155 }
1157 // Remember initial frame bottom
1159 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1161 __ st_ptr(O2, XXX_STATE(_stack_base));
1163 __ sub(O2, wordSize, O2); // prepush
1164 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1166 __ lduh(max_stack, O3); // Full size expression stack
1167 guarantee(!EnableMethodHandles, "no support yet for java.dyn.MethodHandle"); //6815692
1168 //6815692//if (EnableMethodHandles)
1169 //6815692// __ inc(O3, methodOopDesc::extra_stack_entries());
1170 __ sll(O3, LogBytesPerWord, O3);
1171 __ sub(O2, O3, O3);
1172 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1173 __ st_ptr(O3, XXX_STATE(_stack_limit));
1175 if (!native) {
1176 //
1177 // Code to initialize locals
1178 //
1179 Register init_value = noreg; // will be G0 if we must clear locals
1180 // Now zero locals
1181 if (true /* zerolocals */ || ClearInterpreterLocals) {
1182 // explicitly initialize locals
1183 init_value = G0;
1184 } else {
1185 #ifdef ASSERT
1186 // initialize locals to a garbage pattern for better debugging
1187 init_value = O3;
1188 __ set( 0x0F0F0F0F, init_value );
1189 #endif // ASSERT
1190 }
1191 if (init_value != noreg) {
1192 Label clear_loop;
1194 // NOTE: If you change the frame layout, this code will need to
1195 // be updated!
1196 __ lduh( size_of_locals, O2 );
1197 __ lduh( size_of_parameters, O1 );
1198 __ sll( O2, LogBytesPerWord, O2);
1199 __ sll( O1, LogBytesPerWord, O1 );
1200 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1201 __ sub( L2_scratch, O2, O2 );
1202 __ sub( L2_scratch, O1, O1 );
1204 __ bind( clear_loop );
1205 __ inc( O2, wordSize );
1207 __ cmp( O2, O1 );
1208 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1209 __ delayed()->st_ptr( init_value, O2, 0 );
1210 }
1211 }
1212 }
1213 // Find preallocated monitor and lock method (C++ interpreter)
1214 //
1215 void InterpreterGenerator::lock_method(void) {
1216 // Lock the current method.
1217 // Destroys registers L2_scratch, L3_scratch, O0
1218 //
1219 // Find everything relative to Lstate
1221 #ifdef ASSERT
1222 __ ld_ptr(STATE(_method), L2_scratch);
1223 __ ld(L2_scratch, in_bytes(methodOopDesc::access_flags_offset()), O0);
1225 { Label ok;
1226 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1227 __ br( Assembler::notZero, false, Assembler::pt, ok);
1228 __ delayed()->nop();
1229 __ stop("method doesn't need synchronization");
1230 __ bind(ok);
1231 }
1232 #endif // ASSERT
1234 // monitor is already allocated at stack base
1235 // and the lockee is already present
1236 __ ld_ptr(STATE(_stack_base), L2_scratch);
1237 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1238 __ lock_object(L2_scratch, O0);
1240 }
1242 // Generate code for handling resuming a deopted method
1243 void CppInterpreterGenerator::generate_deopt_handling() {
1245 Label return_from_deopt_common;
1247 // deopt needs to jump to here to enter the interpreter (return a result)
1248 deopt_frame_manager_return_atos = __ pc();
1250 // O0/O1 live
1251 __ ba(false, return_from_deopt_common);
1252 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1255 // deopt needs to jump to here to enter the interpreter (return a result)
1256 deopt_frame_manager_return_btos = __ pc();
1258 // O0/O1 live
1259 __ ba(false, return_from_deopt_common);
1260 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1262 // deopt needs to jump to here to enter the interpreter (return a result)
1263 deopt_frame_manager_return_itos = __ pc();
1265 // O0/O1 live
1266 __ ba(false, return_from_deopt_common);
1267 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1269 // deopt needs to jump to here to enter the interpreter (return a result)
1271 deopt_frame_manager_return_ltos = __ pc();
1272 #if !defined(_LP64) && defined(COMPILER2)
1273 // All return values are where we want them, except for Longs. C2 returns
1274 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1275 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1276 // build even if we are returning from interpreted we just do a little
1277 // stupid shuffing.
1278 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1279 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1280 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1282 __ srl (G1, 0,O1);
1283 __ srlx(G1,32,O0);
1284 #endif /* !_LP64 && COMPILER2 */
1285 // O0/O1 live
1286 __ ba(false, return_from_deopt_common);
1287 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1289 // deopt needs to jump to here to enter the interpreter (return a result)
1291 deopt_frame_manager_return_ftos = __ pc();
1292 // O0/O1 live
1293 __ ba(false, return_from_deopt_common);
1294 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1296 // deopt needs to jump to here to enter the interpreter (return a result)
1297 deopt_frame_manager_return_dtos = __ pc();
1299 // O0/O1 live
1300 __ ba(false, return_from_deopt_common);
1301 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1303 // deopt needs to jump to here to enter the interpreter (return a result)
1304 deopt_frame_manager_return_vtos = __ pc();
1306 // O0/O1 live
1307 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1309 // Deopt return common
1310 // an index is present that lets us move any possible result being
1311 // return to the interpreter's stack
1312 //
1313 __ bind(return_from_deopt_common);
1315 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1316 // stack is in the state that the calling convention left it.
1317 // Copy the result from native abi result and place it on java expression stack.
1319 // Current interpreter state is present in Lstate
1321 // Get current pre-pushed top of interpreter stack
1322 // Any result (if any) is in native abi
1323 // result type index is in L3_scratch
1325 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1327 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1328 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1329 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1330 __ jmpl(Lscratch, G0, O7); // and convert it
1331 __ delayed()->nop();
1333 // L1_scratch points to top of stack (prepushed)
1334 __ st_ptr(L1_scratch, STATE(_stack));
1335 }
1337 // Generate the code to handle a more_monitors message from the c++ interpreter
1338 void CppInterpreterGenerator::generate_more_monitors() {
1340 Label entry, loop;
1341 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1342 // 1. compute new pointers // esp: old expression stack top
1343 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1344 __ sub(L4_scratch, entry_size, L4_scratch);
1345 __ st_ptr(L4_scratch, STATE(_stack_base));
1347 __ sub(SP, entry_size, SP); // Grow stack
1348 __ st_ptr(SP, STATE(_frame_bottom));
1350 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1351 __ sub(L2_scratch, entry_size, L2_scratch);
1352 __ st_ptr(L2_scratch, STATE(_stack_limit));
1354 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1355 __ sub(L1_scratch, entry_size, L1_scratch);
1356 __ st_ptr(L1_scratch, STATE(_stack));
1357 __ ba(false, entry);
1358 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1360 // 2. move expression stack
1362 __ bind(loop);
1363 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1364 __ add(L1_scratch, wordSize, L1_scratch);
1365 __ bind(entry);
1366 __ cmp(L1_scratch, L4_scratch);
1367 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1368 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1370 // now zero the slot so we can find it.
1371 __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1373 }
1375 // Initial entry to C++ interpreter from the call_stub.
1376 // This entry point is called the frame manager since it handles the generation
1377 // of interpreter activation frames via requests directly from the vm (via call_stub)
1378 // and via requests from the interpreter. The requests from the call_stub happen
1379 // directly thru the entry point. Requests from the interpreter happen via returning
1380 // from the interpreter and examining the message the interpreter has returned to
1381 // the frame manager. The frame manager can take the following requests:
1383 // NO_REQUEST - error, should never happen.
1384 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1385 // allocate a new monitor.
1386 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1387 // happens during entry during the entry via the call stub.
1388 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1389 //
1390 // Arguments:
1391 //
1392 // ebx: methodOop
1393 // ecx: receiver - unused (retrieved from stack as needed)
1394 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1395 //
1396 //
1397 // Stack layout at entry
1398 //
1399 // [ return address ] <--- esp
1400 // [ parameter n ]
1401 // ...
1402 // [ parameter 1 ]
1403 // [ expression stack ]
1404 //
1405 //
1406 // We are free to blow any registers we like because the call_stub which brought us here
1407 // initially has preserved the callee save registers already.
1408 //
1409 //
1411 static address interpreter_frame_manager = NULL;
1413 #ifdef ASSERT
1414 #define VALIDATE_STATE(scratch, marker) \
1415 { \
1416 Label skip; \
1417 __ ld_ptr(STATE(_self_link), scratch); \
1418 __ cmp(Lstate, scratch); \
1419 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1420 __ delayed()->nop(); \
1421 __ breakpoint_trap(); \
1422 __ emit_long(marker); \
1423 __ bind(skip); \
1424 }
1425 #else
1426 #define VALIDATE_STATE(scratch, marker)
1427 #endif /* ASSERT */
1429 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1430 //
1431 // Adjust caller's stack so that all the locals can be contiguous with
1432 // the parameters.
1433 // Worries about stack overflow make this a pain.
1434 //
1435 // Destroys args, G3_scratch, G3_scratch
1436 // In/Out O5_savedSP (sender's original SP)
1437 //
1438 // assert_different_registers(state, prev_state);
1439 const Register Gtmp = G3_scratch;
1440 const Register tmp = O2;
1441 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1442 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1444 __ lduh(size_of_parameters, tmp);
1445 __ sll(tmp, LogBytesPerWord, Gtmp); // parameter size in bytes
1446 __ add(args, Gtmp, Gargs); // points to first local + BytesPerWord
1447 // NEW
1448 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1449 // determine extra space for non-argument locals & adjust caller's SP
1450 // Gtmp1: parameter size in words
1451 __ lduh(size_of_locals, Gtmp);
1452 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1454 #if 1
1455 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1456 // the call_stub will place the final interpreter argument at
1457 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1458 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1459 // and try to make it look good in the debugger we will store the argument to
1460 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1461 // extra space for the compiler this will overwrite locals in the local array of the
1462 // interpreter.
1463 // QQQ still needed with frameless adapters???
1465 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1467 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1468 #endif // 1
1471 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1472 }
1474 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1476 // G5_method: methodOop
1477 // G2_thread: thread (unused)
1478 // Gargs: bottom of args (sender_sp)
1479 // O5: sender's sp
1481 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1482 // the code is pretty much ready. Would need to change the test below and for good measure
1483 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1484 // routines. Not clear this is worth it yet.
1486 if (interpreter_frame_manager) {
1487 return interpreter_frame_manager;
1488 }
1490 __ bind(frame_manager_entry);
1492 // the following temporary registers are used during frame creation
1493 const Register Gtmp1 = G3_scratch;
1494 // const Register Lmirror = L1; // native mirror (native calls only)
1496 const Address constants (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1497 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1498 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1499 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1500 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1502 address entry_point = __ pc();
1503 __ mov(G0, prevState); // no current activation
1506 Label re_dispatch;
1508 __ bind(re_dispatch);
1510 // Interpreter needs to have locals completely contiguous. In order to do that
1511 // We must adjust the caller's stack pointer for any locals beyond just the
1512 // parameters
1513 adjust_callers_stack(Gargs);
1515 // O5_savedSP still contains sender's sp
1517 // NEW FRAME
1519 generate_compute_interpreter_state(Lstate, prevState, false);
1521 // At this point a new interpreter frame and state object are created and initialized
1522 // Lstate has the pointer to the new activation
1523 // Any stack banging or limit check should already be done.
1525 Label call_interpreter;
1527 __ bind(call_interpreter);
1530 #if 1
1531 __ set(0xdead002, Lmirror);
1532 __ set(0xdead002, L2_scratch);
1533 __ set(0xdead003, L3_scratch);
1534 __ set(0xdead004, L4_scratch);
1535 __ set(0xdead005, Lscratch);
1536 __ set(0xdead006, Lscratch2);
1537 __ set(0xdead007, L7_scratch);
1539 __ set(0xdeaf002, O2);
1540 __ set(0xdeaf003, O3);
1541 __ set(0xdeaf004, O4);
1542 __ set(0xdeaf005, O5);
1543 #endif
1545 // Call interpreter (stack bang complete) enter here if message is
1546 // set and we know stack size is valid
1548 Label call_interpreter_2;
1550 __ bind(call_interpreter_2);
1552 #ifdef ASSERT
1553 {
1554 Label skip;
1555 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1556 __ cmp(G3_scratch, SP);
1557 __ brx(Assembler::equal, false, Assembler::pt, skip);
1558 __ delayed()->nop();
1559 __ stop("SP not restored to frame bottom");
1560 __ bind(skip);
1561 }
1562 #endif
1564 VALIDATE_STATE(G3_scratch, 4);
1565 __ set_last_Java_frame(SP, noreg);
1566 __ mov(Lstate, O0); // (arg) pointer to current state
1568 __ call(CAST_FROM_FN_PTR(address,
1569 JvmtiExport::can_post_interpreter_events() ?
1570 BytecodeInterpreter::runWithChecks
1571 : BytecodeInterpreter::run),
1572 relocInfo::runtime_call_type);
1574 __ delayed()->nop();
1576 __ ld_ptr(STATE(_thread), G2_thread);
1577 __ reset_last_Java_frame();
1579 // examine msg from interpreter to determine next action
1580 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1582 __ ld(STATE(_msg), L1_scratch); // Get new message
1584 Label call_method;
1585 Label return_from_interpreted_method;
1586 Label throw_exception;
1587 Label do_OSR;
1588 Label bad_msg;
1589 Label resume_interpreter;
1591 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1592 __ br(Assembler::equal, false, Assembler::pt, call_method);
1593 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1594 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1595 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1596 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1597 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1598 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1599 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1600 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1602 // Allocate more monitor space, shuffle expression stack....
1604 generate_more_monitors();
1606 // new monitor slot allocated, resume the interpreter.
1608 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1609 VALIDATE_STATE(G3_scratch, 5);
1610 __ ba(false, call_interpreter);
1611 __ delayed()->st(L1_scratch, STATE(_msg));
1613 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1614 unctrap_frame_manager_entry = __ pc();
1616 // QQQ what message do we send
1618 __ ba(false, call_interpreter);
1619 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1621 //=============================================================================
1622 // Returning from a compiled method into a deopted method. The bytecode at the
1623 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1624 // for the template based interpreter). Any stack space that was used by the
1625 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1626 // so all that we have to do is place any pending result on the expression stack
1627 // and resume execution on the next bytecode.
1629 generate_deopt_handling();
1631 // ready to resume the interpreter
1633 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1634 __ ba(false, call_interpreter);
1635 __ delayed()->st(L1_scratch, STATE(_msg));
1637 // Current frame has caught an exception we need to dispatch to the
1638 // handler. We can get here because a native interpreter frame caught
1639 // an exception in which case there is no handler and we must rethrow
1640 // If it is a vanilla interpreted frame the we simply drop into the
1641 // interpreter and let it do the lookup.
1643 Interpreter::_rethrow_exception_entry = __ pc();
1645 Label return_with_exception;
1646 Label unwind_and_forward;
1648 // O0: exception
1649 // O7: throwing pc
1651 // We want exception in the thread no matter what we ultimately decide about frame type.
1653 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1654 __ verify_thread();
1655 __ st_ptr(O0, exception_addr);
1657 // get the methodOop
1658 __ ld_ptr(STATE(_method), G5_method);
1660 // if this current frame vanilla or native?
1662 __ ld(access_flags, Gtmp1);
1663 __ btst(JVM_ACC_NATIVE, Gtmp1);
1664 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1665 __ delayed()->nop();
1667 // We drop thru to unwind a native interpreted frame with a pending exception
1668 // We jump here for the initial interpreter frame with exception pending
1669 // We unwind the current acivation and forward it to our caller.
1671 __ bind(unwind_and_forward);
1673 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1674 // as expected by forward_exception.
1676 __ restore(FP, G0, SP); // unwind interpreter state frame
1677 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1678 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1680 // Return point from a call which returns a result in the native abi
1681 // (c1/c2/jni-native). This result must be processed onto the java
1682 // expression stack.
1683 //
1684 // A pending exception may be present in which case there is no result present
1686 address return_from_native_method = __ pc();
1688 VALIDATE_STATE(G3_scratch, 6);
1690 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1691 // stack is in the state that the calling convention left it.
1692 // Copy the result from native abi result and place it on java expression stack.
1694 // Current interpreter state is present in Lstate
1696 // Exception pending?
1698 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1699 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1700 __ tst(Lscratch); // exception pending?
1701 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1702 __ delayed()->nop();
1704 // Process the native abi result to java expression stack
1706 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1707 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1708 __ lduh(L4_scratch, in_bytes(methodOopDesc::size_of_parameters_offset()), L2_scratch); // get parameter size
1709 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1710 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1711 __ ld(L4_scratch, in_bytes(methodOopDesc::result_index_offset()), L3_scratch); // called method result type index
1713 // tosca is really just native abi
1714 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1715 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1716 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1717 __ jmpl(Lscratch, G0, O7); // and convert it
1718 __ delayed()->nop();
1720 // L1_scratch points to top of stack (prepushed)
1722 __ ba(false, resume_interpreter);
1723 __ delayed()->mov(L1_scratch, O1);
1725 // An exception is being caught on return to a vanilla interpreter frame.
1726 // Empty the stack and resume interpreter
1728 __ bind(return_with_exception);
1730 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1731 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1732 __ ba(false, resume_interpreter);
1733 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1735 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1736 // interpreter call, or native) and unwind this interpreter activation.
1737 // All monitors should be unlocked.
1739 __ bind(return_from_interpreted_method);
1741 VALIDATE_STATE(G3_scratch, 7);
1743 Label return_to_initial_caller;
1745 // Interpreted result is on the top of the completed activation expression stack.
1746 // We must return it to the top of the callers stack if caller was interpreted
1747 // otherwise we convert to native abi result and return to call_stub/c1/c2
1748 // The caller's expression stack was truncated by the call however the current activation
1749 // has enough stuff on the stack that we have usable space there no matter what. The
1750 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1751 // for the current activation
1753 __ ld_ptr(STATE(_prev_link), L1_scratch);
1754 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1755 __ ld(L2_scratch, in_bytes(methodOopDesc::result_index_offset()), L2_scratch);
1756 __ tst(L1_scratch);
1757 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1758 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1760 // Copy result to callers java stack
1762 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1763 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1764 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1765 __ ld_ptr(STATE(_locals), O1); // stack destination
1767 // O0 - will be source, O1 - will be destination (preserved)
1768 __ jmpl(Lscratch, G0, O7); // and convert it
1769 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1771 // O1 == &locals[0]
1773 // Result is now on caller's stack. Just unwind current activation and resume
1775 Label unwind_recursive_activation;
1778 __ bind(unwind_recursive_activation);
1780 // O1 == &locals[0] (really callers stacktop) for activation now returning
1781 // returning to interpreter method from "recursive" interpreter call
1782 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1783 // to. Now all we must do is unwind the state from the completed call
1785 // Must restore stack
1786 VALIDATE_STATE(G3_scratch, 8);
1788 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1789 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1790 // frame and tell interpreter to resume.
1793 __ mov(O1, I1); // pass back new stack top across activation
1794 // POP FRAME HERE ==================================
1795 __ restore(FP, G0, SP); // unwind interpreter state frame
1796 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1799 // Resume the interpreter. The current frame contains the current interpreter
1800 // state object.
1801 //
1802 // O1 == new java stack pointer
1804 __ bind(resume_interpreter);
1805 VALIDATE_STATE(G3_scratch, 10);
1807 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1809 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1810 __ st(L1_scratch, STATE(_msg));
1811 __ ba(false, call_interpreter_2);
1812 __ delayed()->st_ptr(O1, STATE(_stack));
1815 // Fast accessor methods share this entry point.
1816 // This works because frame manager is in the same codelet
1817 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1818 // we need to do a little register fixup here once we distinguish the two of them
1819 if (UseFastAccessorMethods && !synchronized) {
1820 // Call stub_return address still in O7
1821 __ bind(fast_accessor_slow_entry_path);
1822 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1823 __ cmp(Gtmp1, O7); // returning to interpreter?
1824 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1825 __ delayed()->nop();
1826 __ ba(false, re_dispatch);
1827 __ delayed()->mov(G0, prevState); // initial entry
1829 }
1831 // interpreter returning to native code (call_stub/c1/c2)
1832 // convert result and unwind initial activation
1833 // L2_scratch - scaled result type index
1835 __ bind(return_to_initial_caller);
1837 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1838 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1839 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1840 __ jmpl(Lscratch, G0, O7); // and convert it
1841 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1843 Label unwind_initial_activation;
1844 __ bind(unwind_initial_activation);
1846 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1847 // we can return here with an exception that wasn't handled by interpreted code
1848 // how does c1/c2 see it on return?
1850 // compute resulting sp before/after args popped depending upon calling convention
1851 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1852 //
1853 // POP FRAME HERE ==================================
1854 __ restore(FP, G0, SP);
1855 __ retl();
1856 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1858 // OSR request, unwind the current frame and transfer to the OSR entry
1859 // and enter OSR nmethod
1861 __ bind(do_OSR);
1862 Label remove_initial_frame;
1863 __ ld_ptr(STATE(_prev_link), L1_scratch);
1864 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1866 // We are going to pop this frame. Is there another interpreter frame underneath
1867 // it or is it callstub/compiled?
1869 __ tst(L1_scratch);
1870 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1871 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1873 // Frame underneath is an interpreter frame simply unwind
1874 // POP FRAME HERE ==================================
1875 __ restore(FP, G0, SP); // unwind interpreter state frame
1876 __ mov(I5_savedSP->after_restore(), SP);
1878 // Since we are now calling native need to change our "return address" from the
1879 // dummy RecursiveInterpreterActivation to a return from native
1881 __ set((intptr_t)return_from_native_method - 8, O7);
1883 __ jmpl(G3_scratch, G0, G0);
1884 __ delayed()->mov(G1_scratch, O0);
1886 __ bind(remove_initial_frame);
1888 // POP FRAME HERE ==================================
1889 __ restore(FP, G0, SP);
1890 __ mov(I5_savedSP->after_restore(), SP);
1891 __ jmpl(G3_scratch, G0, G0);
1892 __ delayed()->mov(G1_scratch, O0);
1894 // Call a new method. All we do is (temporarily) trim the expression stack
1895 // push a return address to bring us back to here and leap to the new entry.
1896 // At this point we have a topmost frame that was allocated by the frame manager
1897 // which contains the current method interpreted state. We trim this frame
1898 // of excess java expression stack entries and then recurse.
1900 __ bind(call_method);
1902 // stack points to next free location and not top element on expression stack
1903 // method expects sp to be pointing to topmost element
1905 __ ld_ptr(STATE(_thread), G2_thread);
1906 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1909 // SP already takes in to account the 2 extra words we use for slop
1910 // when we call a "static long no_params()" method. So if
1911 // we trim back sp by the amount of unused java expression stack
1912 // there will be automagically the 2 extra words we need.
1913 // We also have to worry about keeping SP aligned.
1915 __ ld_ptr(STATE(_stack), Gargs);
1916 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1918 // compute the unused java stack size
1919 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1921 // Round down the unused space to that stack is always 16-byte aligned
1922 // by making the unused space a multiple of the size of two longs.
1924 __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1926 // Now trim the stack
1927 __ add(SP, L2_scratch, SP);
1930 // Now point to the final argument (account for prepush)
1931 __ add(Gargs, wordSize, Gargs);
1932 #ifdef ASSERT
1933 // Make sure we have space for the window
1934 __ sub(Gargs, SP, L1_scratch);
1935 __ cmp(L1_scratch, 16*wordSize);
1936 {
1937 Label skip;
1938 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1939 __ delayed()->nop();
1940 __ stop("killed stack");
1941 __ bind(skip);
1942 }
1943 #endif // ASSERT
1945 // Create a new frame where we can store values that make it look like the interpreter
1946 // really recursed.
1948 // prepare to recurse or call specialized entry
1950 // First link the registers we need
1952 // make the pc look good in debugger
1953 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
1954 // argument too
1955 __ mov(Lstate, I0);
1957 // Record our sending SP
1958 __ mov(SP, O5_savedSP);
1960 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
1961 __ set((intptr_t) entry_point, L1_scratch);
1962 __ cmp(L1_scratch, L2_scratch);
1963 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
1964 __ delayed()->mov(Lstate, prevState); // link activations
1966 // method uses specialized entry, push a return so we look like call stub setup
1967 // this path will handle fact that result is returned in registers and not
1968 // on the java stack.
1970 __ set((intptr_t)return_from_native_method - 8, O7);
1971 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
1972 __ delayed()->nop();
1974 //
1975 // Bad Message from interpreter
1976 //
1977 __ bind(bad_msg);
1978 __ stop("Bad message from interpreter");
1980 // Interpreted method "returned" with an exception pass it on...
1981 // Pass result, unwind activation and continue/return to interpreter/call_stub
1982 // We handle result (if any) differently based on return to interpreter or call_stub
1984 __ bind(throw_exception);
1985 __ ld_ptr(STATE(_prev_link), L1_scratch);
1986 __ tst(L1_scratch);
1987 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
1988 __ delayed()->nop();
1990 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
1991 __ ba(false, unwind_recursive_activation);
1992 __ delayed()->nop();
1994 interpreter_frame_manager = entry_point;
1995 return entry_point;
1996 }
1998 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1999 : CppInterpreterGenerator(code) {
2000 generate_all(); // down here so it can be "virtual"
2001 }
2004 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2006 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2007 // expression stack, the callee will have callee_extra_locals (so we can account for
2008 // frame extension) and monitor_size for monitors. Basically we need to calculate
2009 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2010 //
2011 //
2012 // The big complicating thing here is that we must ensure that the stack stays properly
2013 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2014 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2015 // the caller so we must ensure that it is properly aligned for our callee.
2016 //
2017 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2018 // stack at all times to deal with the "stack long no_params()" method issue. This
2019 // is "slop_factor" here.
2020 const int slop_factor = 2;
2022 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2023 frame::memory_parameter_word_sp_offset; // register save area + param window
2024 const int extra_stack = 0; //6815692//methodOopDesc::extra_stack_entries();
2025 return (round_to(max_stack +
2026 extra_stack +
2027 slop_factor +
2028 fixed_size +
2029 monitor_size +
2030 (callee_extra_locals * Interpreter::stackElementWords()), WordsPerLong));
2032 }
2034 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
2036 // See call_stub code
2037 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2038 WordsPerLong); // 7 + register save area
2040 // Save space for one monitor to get into the interpreted method in case
2041 // the method is synchronized
2042 int monitor_size = method->is_synchronized() ?
2043 1*frame::interpreter_frame_monitor_size() : 0;
2044 return size_activation_helper(method->max_locals(), method->max_stack(),
2045 monitor_size) + call_stub_size;
2046 }
2048 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2049 frame* caller,
2050 frame* current,
2051 methodOop method,
2052 intptr_t* locals,
2053 intptr_t* stack,
2054 intptr_t* stack_base,
2055 intptr_t* monitor_base,
2056 intptr_t* frame_bottom,
2057 bool is_top_frame
2058 )
2059 {
2060 // What about any vtable?
2061 //
2062 to_fill->_thread = JavaThread::current();
2063 // This gets filled in later but make it something recognizable for now
2064 to_fill->_bcp = method->code_base();
2065 to_fill->_locals = locals;
2066 to_fill->_constants = method->constants()->cache();
2067 to_fill->_method = method;
2068 to_fill->_mdx = NULL;
2069 to_fill->_stack = stack;
2070 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2071 to_fill->_msg = deopt_resume2;
2072 } else {
2073 to_fill->_msg = method_resume;
2074 }
2075 to_fill->_result._to_call._bcp_advance = 0;
2076 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2077 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2078 to_fill->_prev_link = NULL;
2080 // Fill in the registers for the frame
2082 // Need to install _sender_sp. Actually not too hard in C++!
2083 // When the skeletal frames are layed out we fill in a value
2084 // for _sender_sp. That value is only correct for the oldest
2085 // skeletal frame constructed (because there is only a single
2086 // entry for "caller_adjustment". While the skeletal frames
2087 // exist that is good enough. We correct that calculation
2088 // here and get all the frames correct.
2090 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2092 *current->register_addr(Lstate) = (intptr_t) to_fill;
2093 // skeletal already places a useful value here and this doesn't account
2094 // for alignment so don't bother.
2095 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2097 if (caller->is_interpreted_frame()) {
2098 interpreterState prev = caller->get_interpreterState();
2099 to_fill->_prev_link = prev;
2100 // Make the prev callee look proper
2101 prev->_result._to_call._callee = method;
2102 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2103 prev->_result._to_call._bcp_advance = 5;
2104 } else {
2105 prev->_result._to_call._bcp_advance = 3;
2106 }
2107 }
2108 to_fill->_oop_temp = NULL;
2109 to_fill->_stack_base = stack_base;
2110 // Need +1 here because stack_base points to the word just above the first expr stack entry
2111 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2112 // See generate_compute_interpreter_state.
2113 int extra_stack = 0; //6815692//methodOopDesc::extra_stack_entries();
2114 to_fill->_stack_limit = stack_base - (method->max_stack() + 1 + extra_stack);
2115 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2117 // sparc specific
2118 to_fill->_frame_bottom = frame_bottom;
2119 to_fill->_self_link = to_fill;
2120 #ifdef ASSERT
2121 to_fill->_native_fresult = 123456.789;
2122 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2123 #endif
2124 }
2126 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2127 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2128 }
2131 int AbstractInterpreter::layout_activation(methodOop method,
2132 int tempcount, // Number of slots on java expression stack in use
2133 int popframe_extra_args,
2134 int moncount, // Number of active monitors
2135 int callee_param_size,
2136 int callee_locals_size,
2137 frame* caller,
2138 frame* interpreter_frame,
2139 bool is_top_frame) {
2141 assert(popframe_extra_args == 0, "NEED TO FIX");
2142 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2143 // does as far as allocating an interpreter frame.
2144 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2145 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2146 // as determined by a previous call to this method.
2147 // It is also guaranteed to be walkable even though it is in a skeletal state
2148 // NOTE: return size is in words not bytes
2149 // NOTE: tempcount is the current size of the java expression stack. For top most
2150 // frames we will allocate a full sized expression stack and not the curback
2151 // version that non-top frames have.
2153 // Calculate the amount our frame will be adjust by the callee. For top frame
2154 // this is zero.
2156 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2157 // calculates the extra locals based on itself. Not what the callee does
2158 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2159 // as getting sender_sp correct.
2161 int extra_locals_size = callee_locals_size - callee_param_size;
2162 int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2163 int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2164 int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2165 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2168 /*
2169 if we actually have a frame to layout we must now fill in all the pieces. This means both
2170 the interpreterState and the registers.
2171 */
2172 if (interpreter_frame != NULL) {
2174 // MUCHO HACK
2176 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2177 // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2178 assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2179 frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2181 /* Now fillin the interpreterState object */
2183 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2186 intptr_t* locals;
2188 // Calculate the postion of locals[0]. This is painful because of
2189 // stack alignment (same as ia64). The problem is that we can
2190 // not compute the location of locals from fp(). fp() will account
2191 // for the extra locals but it also accounts for aligning the stack
2192 // and we can't determine if the locals[0] was misaligned but max_locals
2193 // was enough to have the
2194 // calculate postion of locals. fp already accounts for extra locals.
2195 // +2 for the static long no_params() issue.
2197 if (caller->is_interpreted_frame()) {
2198 // locals must agree with the caller because it will be used to set the
2199 // caller's tos when we return.
2200 interpreterState prev = caller->get_interpreterState();
2201 // stack() is prepushed.
2202 locals = prev->stack() + method->size_of_parameters();
2203 } else {
2204 // Lay out locals block in the caller adjacent to the register window save area.
2205 //
2206 // Compiled frames do not allocate a varargs area which is why this if
2207 // statement is needed.
2208 //
2209 intptr_t* fp = interpreter_frame->fp();
2210 int local_words = method->max_locals() * Interpreter::stackElementWords();
2212 if (caller->is_compiled_frame()) {
2213 locals = fp + frame::register_save_words + local_words - 1;
2214 } else {
2215 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2216 }
2218 }
2219 // END MUCHO HACK
2221 intptr_t* monitor_base = (intptr_t*) cur_state;
2222 intptr_t* stack_base = monitor_base - monitor_size;
2223 /* +1 because stack is always prepushed */
2224 intptr_t* stack = stack_base - (tempcount + 1);
2227 BytecodeInterpreter::layout_interpreterState(cur_state,
2228 caller,
2229 interpreter_frame,
2230 method,
2231 locals,
2232 stack,
2233 stack_base,
2234 monitor_base,
2235 frame_bottom,
2236 is_top_frame);
2238 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2240 }
2241 return frame_words;
2242 }
2244 #endif // CC_INTERP