Mon, 27 Aug 2012 15:17:17 -0700
6677625: Move platform specific flags from globals.hpp to globals_<arch>.hpp
Reviewed-by: kvn, dholmes, coleenp
Contributed-by: Tao Mao <tao.mao@oracle.com>
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/methodDataOop.hpp"
34 #include "oops/methodOop.hpp"
35 #include "oops/oop.inline.hpp"
36 #include "prims/jvmtiExport.hpp"
37 #include "prims/jvmtiThreadState.hpp"
38 #include "runtime/arguments.hpp"
39 #include "runtime/deoptimization.hpp"
40 #include "runtime/frame.inline.hpp"
41 #include "runtime/interfaceSupport.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "runtime/stubRoutines.hpp"
44 #include "runtime/synchronizer.hpp"
45 #include "runtime/timer.hpp"
46 #include "runtime/vframeArray.hpp"
47 #include "utilities/debug.hpp"
48 #ifdef SHARK
49 #include "shark/shark_globals.hpp"
50 #endif
52 #ifdef CC_INTERP
54 // Routine exists to make tracebacks look decent in debugger
55 // while "shadow" interpreter frames are on stack. It is also
56 // used to distinguish interpreter frames.
58 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
59 ShouldNotReachHere();
60 }
62 bool CppInterpreter::contains(address pc) {
63 return ( _code->contains(pc) ||
64 ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
65 }
67 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
68 #define __ _masm->
70 Label frame_manager_entry;
71 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized
72 // c++ interpreter entry point this holds that entry point label.
74 static address unctrap_frame_manager_entry = NULL;
76 static address interpreter_return_address = NULL;
77 static address deopt_frame_manager_return_atos = NULL;
78 static address deopt_frame_manager_return_btos = NULL;
79 static address deopt_frame_manager_return_itos = NULL;
80 static address deopt_frame_manager_return_ltos = NULL;
81 static address deopt_frame_manager_return_ftos = NULL;
82 static address deopt_frame_manager_return_dtos = NULL;
83 static address deopt_frame_manager_return_vtos = NULL;
85 const Register prevState = G1_scratch;
87 void InterpreterGenerator::save_native_result(void) {
88 // result potentially in O0/O1: save it across calls
89 __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
90 #ifdef _LP64
91 __ stx(O0, STATE(_native_lresult));
92 #else
93 __ std(O0, STATE(_native_lresult));
94 #endif
95 }
97 void InterpreterGenerator::restore_native_result(void) {
99 // Restore any method result value
100 __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
101 #ifdef _LP64
102 __ ldx(STATE(_native_lresult), O0);
103 #else
104 __ ldd(STATE(_native_lresult), O0);
105 #endif
106 }
108 // A result handler converts/unboxes a native call result into
109 // a java interpreter/compiler result. The current frame is an
110 // interpreter frame. The activation frame unwind code must be
111 // consistent with that of TemplateTable::_return(...). In the
112 // case of native methods, the caller's SP was not modified.
113 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
114 address entry = __ pc();
115 Register Itos_i = Otos_i ->after_save();
116 Register Itos_l = Otos_l ->after_save();
117 Register Itos_l1 = Otos_l1->after_save();
118 Register Itos_l2 = Otos_l2->after_save();
119 switch (type) {
120 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
121 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
122 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
123 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
124 case T_LONG :
125 #ifndef _LP64
126 __ mov(O1, Itos_l2); // move other half of long
127 #endif // ifdef or no ifdef, fall through to the T_INT case
128 case T_INT : __ mov(O0, Itos_i); break;
129 case T_VOID : /* nothing to do */ break;
130 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
131 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
132 case T_OBJECT :
133 __ ld_ptr(STATE(_oop_temp), Itos_i);
134 __ verify_oop(Itos_i);
135 break;
136 default : ShouldNotReachHere();
137 }
138 __ ret(); // return from interpreter activation
139 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
140 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
141 return entry;
142 }
144 // tosca based result to c++ interpreter stack based result.
145 // Result goes to address in L1_scratch
147 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
148 // A result is in the native abi result register from a native method call.
149 // We need to return this result to the interpreter by pushing the result on the interpreter's
150 // stack. This is relatively simple the destination is in L1_scratch
151 // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
152 // adjust L1_scratch
153 address entry = __ pc();
154 switch (type) {
155 case T_BOOLEAN:
156 // !0 => true; 0 => false
157 __ subcc(G0, O0, G0);
158 __ addc(G0, 0, O0);
159 __ st(O0, L1_scratch, 0);
160 __ sub(L1_scratch, wordSize, L1_scratch);
161 break;
163 // cannot use and3, 0xFFFF too big as immediate value!
164 case T_CHAR :
165 __ sll(O0, 16, O0);
166 __ srl(O0, 16, O0);
167 __ st(O0, L1_scratch, 0);
168 __ sub(L1_scratch, wordSize, L1_scratch);
169 break;
171 case T_BYTE :
172 __ sll(O0, 24, O0);
173 __ sra(O0, 24, O0);
174 __ st(O0, L1_scratch, 0);
175 __ sub(L1_scratch, wordSize, L1_scratch);
176 break;
178 case T_SHORT :
179 __ sll(O0, 16, O0);
180 __ sra(O0, 16, O0);
181 __ st(O0, L1_scratch, 0);
182 __ sub(L1_scratch, wordSize, L1_scratch);
183 break;
184 case T_LONG :
185 #ifndef _LP64
186 #if defined(COMPILER2)
187 // All return values are where we want them, except for Longs. C2 returns
188 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
189 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
190 // build even if we are returning from interpreted we just do a little
191 // stupid shuffing.
192 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
193 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
194 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
195 __ stx(G1, L1_scratch, -wordSize);
196 #else
197 // native result is in O0, O1
198 __ st(O1, L1_scratch, 0); // Low order
199 __ st(O0, L1_scratch, -wordSize); // High order
200 #endif /* COMPILER2 */
201 #else
202 __ stx(O0, L1_scratch, -wordSize);
203 #endif
204 __ sub(L1_scratch, 2*wordSize, L1_scratch);
205 break;
207 case T_INT :
208 __ st(O0, L1_scratch, 0);
209 __ sub(L1_scratch, wordSize, L1_scratch);
210 break;
212 case T_VOID : /* nothing to do */
213 break;
215 case T_FLOAT :
216 __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
217 __ sub(L1_scratch, wordSize, L1_scratch);
218 break;
220 case T_DOUBLE :
221 // Every stack slot is aligned on 64 bit, However is this
222 // the correct stack slot on 64bit?? QQQ
223 __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
224 __ sub(L1_scratch, 2*wordSize, L1_scratch);
225 break;
226 case T_OBJECT :
227 __ verify_oop(O0);
228 __ st_ptr(O0, L1_scratch, 0);
229 __ sub(L1_scratch, wordSize, L1_scratch);
230 break;
231 default : ShouldNotReachHere();
232 }
233 __ retl(); // return from interpreter activation
234 __ delayed()->nop(); // schedule this better
235 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
236 return entry;
237 }
239 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
240 // A result is in the java expression stack of the interpreted method that has just
241 // returned. Place this result on the java expression stack of the caller.
242 //
243 // The current interpreter activation in Lstate is for the method just returning its
244 // result. So we know that the result of this method is on the top of the current
245 // execution stack (which is pre-pushed) and will be return to the top of the caller
246 // stack. The top of the callers stack is the bottom of the locals of the current
247 // activation.
248 // Because of the way activation are managed by the frame manager the value of esp is
249 // below both the stack top of the current activation and naturally the stack top
250 // of the calling activation. This enable this routine to leave the return address
251 // to the frame manager on the stack and do a vanilla return.
252 //
253 // On entry: O0 - points to source (callee stack top)
254 // O1 - points to destination (caller stack top [i.e. free location])
255 // destroys O2, O3
256 //
258 address entry = __ pc();
259 switch (type) {
260 case T_VOID: break;
261 break;
262 case T_FLOAT :
263 case T_BOOLEAN:
264 case T_CHAR :
265 case T_BYTE :
266 case T_SHORT :
267 case T_INT :
268 // 1 word result
269 __ ld(O0, 0, O2);
270 __ st(O2, O1, 0);
271 __ sub(O1, wordSize, O1);
272 break;
273 case T_DOUBLE :
274 case T_LONG :
275 // return top two words on current expression stack to caller's expression stack
276 // The caller's expression stack is adjacent to the current frame manager's intepretState
277 // except we allocated one extra word for this intepretState so we won't overwrite it
278 // when we return a two word result.
279 #ifdef _LP64
280 __ ld_ptr(O0, 0, O2);
281 __ st_ptr(O2, O1, -wordSize);
282 #else
283 __ ld(O0, 0, O2);
284 __ ld(O0, wordSize, O3);
285 __ st(O3, O1, 0);
286 __ st(O2, O1, -wordSize);
287 #endif
288 __ sub(O1, 2*wordSize, O1);
289 break;
290 case T_OBJECT :
291 __ ld_ptr(O0, 0, O2);
292 __ verify_oop(O2); // verify it
293 __ st_ptr(O2, O1, 0);
294 __ sub(O1, wordSize, O1);
295 break;
296 default : ShouldNotReachHere();
297 }
298 __ retl();
299 __ delayed()->nop(); // QQ schedule this better
300 return entry;
301 }
303 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
304 // A result is in the java expression stack of the interpreted method that has just
305 // returned. Place this result in the native abi that the caller expects.
306 // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
307 //
308 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
309 // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
310 // and so rather than return result onto caller's java expression stack we return the
311 // result in the expected location based on the native abi.
312 // On entry: O0 - source (stack top)
313 // On exit result in expected output register
314 // QQQ schedule this better
316 address entry = __ pc();
317 switch (type) {
318 case T_VOID: break;
319 break;
320 case T_FLOAT :
321 __ ldf(FloatRegisterImpl::S, O0, 0, F0);
322 break;
323 case T_BOOLEAN:
324 case T_CHAR :
325 case T_BYTE :
326 case T_SHORT :
327 case T_INT :
328 // 1 word result
329 __ ld(O0, 0, O0->after_save());
330 break;
331 case T_DOUBLE :
332 __ ldf(FloatRegisterImpl::D, O0, 0, F0);
333 break;
334 case T_LONG :
335 // return top two words on current expression stack to caller's expression stack
336 // The caller's expression stack is adjacent to the current frame manager's interpretState
337 // except we allocated one extra word for this intepretState so we won't overwrite it
338 // when we return a two word result.
339 #ifdef _LP64
340 __ ld_ptr(O0, 0, O0->after_save());
341 #else
342 __ ld(O0, wordSize, O1->after_save());
343 __ ld(O0, 0, O0->after_save());
344 #endif
345 #if defined(COMPILER2) && !defined(_LP64)
346 // C2 expects long results in G1 we can't tell if we're returning to interpreted
347 // or compiled so just be safe use G1 and O0/O1
349 // Shift bits into high (msb) of G1
350 __ sllx(Otos_l1->after_save(), 32, G1);
351 // Zero extend low bits
352 __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
353 __ or3 (Otos_l2->after_save(), G1, G1);
354 #endif /* COMPILER2 */
355 break;
356 case T_OBJECT :
357 __ ld_ptr(O0, 0, O0->after_save());
358 __ verify_oop(O0->after_save()); // verify it
359 break;
360 default : ShouldNotReachHere();
361 }
362 __ retl();
363 __ delayed()->nop();
364 return entry;
365 }
367 address CppInterpreter::return_entry(TosState state, int length) {
368 // make it look good in the debugger
369 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
370 }
372 address CppInterpreter::deopt_entry(TosState state, int length) {
373 address ret = NULL;
374 if (length != 0) {
375 switch (state) {
376 case atos: ret = deopt_frame_manager_return_atos; break;
377 case btos: ret = deopt_frame_manager_return_btos; break;
378 case ctos:
379 case stos:
380 case itos: ret = deopt_frame_manager_return_itos; break;
381 case ltos: ret = deopt_frame_manager_return_ltos; break;
382 case ftos: ret = deopt_frame_manager_return_ftos; break;
383 case dtos: ret = deopt_frame_manager_return_dtos; break;
384 case vtos: ret = deopt_frame_manager_return_vtos; break;
385 }
386 } else {
387 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap)
388 }
389 assert(ret != NULL, "Not initialized");
390 return ret;
391 }
393 //
394 // Helpers for commoning out cases in the various type of method entries.
395 //
397 // increment invocation count & check for overflow
398 //
399 // Note: checking for negative value instead of overflow
400 // so we have a 'sticky' overflow test
401 //
402 // Lmethod: method
403 // ??: invocation counter
404 //
405 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
406 // Update standard invocation counters
407 __ increment_invocation_counter(O0, G3_scratch);
408 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
409 __ ld_ptr(STATE(_method), G3_scratch);
410 Address interpreter_invocation_counter(G3_scratch, 0, in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
411 __ ld(interpreter_invocation_counter, G3_scratch);
412 __ inc(G3_scratch);
413 __ st(G3_scratch, interpreter_invocation_counter);
414 }
416 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
417 __ sethi(invocation_limit);
418 __ ld(invocation_limit, G3_scratch);
419 __ cmp(O0, G3_scratch);
420 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
421 __ delayed()->nop();
423 }
425 address InterpreterGenerator::generate_empty_entry(void) {
427 // A method that does nothing but return...
429 address entry = __ pc();
430 Label slow_path;
432 __ verify_oop(G5_method);
434 // do nothing for empty methods (do not even increment invocation counter)
435 if ( UseFastEmptyMethods) {
436 // If we need a safepoint check, generate full interpreter entry.
437 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
438 __ load_contents(sync_state, G3_scratch);
439 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
440 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
441 __ delayed()->nop();
443 // Code: _return
444 __ retl();
445 __ delayed()->mov(O5_savedSP, SP);
446 return entry;
447 }
448 return NULL;
449 }
451 // Call an accessor method (assuming it is resolved, otherwise drop into
452 // vanilla (slow path) entry
454 // Generates code to elide accessor methods
455 // Uses G3_scratch and G1_scratch as scratch
456 address InterpreterGenerator::generate_accessor_entry(void) {
458 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
459 // parameter size = 1
460 // Note: We can only use this code if the getfield has been resolved
461 // and if we don't have a null-pointer exception => check for
462 // these conditions first and use slow path if necessary.
463 address entry = __ pc();
464 Label slow_path;
466 if ( UseFastAccessorMethods) {
467 // Check if we need to reach a safepoint and generate full interpreter
468 // frame if so.
469 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
470 __ load_contents(sync_state, G3_scratch);
471 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
472 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
473 __ delayed()->nop();
475 // Check if local 0 != NULL
476 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
477 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
478 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
479 __ delayed()->nop();
482 // read first instruction word and extract bytecode @ 1 and index @ 2
483 // get first 4 bytes of the bytecodes (big endian!)
484 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), G1_scratch);
485 __ ld(Address(G1_scratch, 0, in_bytes(constMethodOopDesc::codes_offset())), G1_scratch);
487 // move index @ 2 far left then to the right most two bytes.
488 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
489 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
490 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
492 // get constant pool cache
493 __ ld_ptr(G5_method, in_bytes(methodOopDesc::const_offset()), G3_scratch);
494 __ ld_ptr(G3_scratch, in_bytes(constMethodOopDesc::constants_offset()), G3_scratch);
495 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
497 // get specific constant pool cache entry
498 __ add(G3_scratch, G1_scratch, G3_scratch);
500 // Check the constant Pool cache entry to see if it has been resolved.
501 // If not, need the slow path.
502 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
503 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
504 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
505 __ and3(G1_scratch, 0xFF, G1_scratch);
506 __ cmp(G1_scratch, Bytecodes::_getfield);
507 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
508 __ delayed()->nop();
510 // Get the type and return field offset from the constant pool cache
511 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
512 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
514 Label xreturn_path;
515 // Need to differentiate between igetfield, agetfield, bgetfield etc.
516 // because they are different sizes.
517 // Get the type from the constant pool cache
518 __ srl(G1_scratch, ConstantPoolCacheEntry::tos_state_shift, G1_scratch);
519 // Make sure we don't need to mask G1_scratch after the above shift
520 ConstantPoolCacheEntry::verify_tos_state_shift();
521 __ cmp(G1_scratch, atos );
522 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
523 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
524 __ cmp(G1_scratch, itos);
525 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
526 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
527 __ cmp(G1_scratch, stos);
528 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
529 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
530 __ cmp(G1_scratch, ctos);
531 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
532 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
533 #ifdef ASSERT
534 __ cmp(G1_scratch, btos);
535 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
536 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
537 __ should_not_reach_here();
538 #endif
539 __ ldsb(Otos_i, G3_scratch, Otos_i);
540 __ bind(xreturn_path);
542 // _ireturn/_areturn
543 __ retl(); // return from leaf routine
544 __ delayed()->mov(O5_savedSP, SP);
546 // Generate regular method entry
547 __ bind(slow_path);
548 __ ba(fast_accessor_slow_entry_path);
549 __ delayed()->nop();
550 return entry;
551 }
552 return NULL;
553 }
555 address InterpreterGenerator::generate_Reference_get_entry(void) {
556 #ifndef SERIALGC
557 if (UseG1GC) {
558 // We need to generate have a routine that generates code to:
559 // * load the value in the referent field
560 // * passes that value to the pre-barrier.
561 //
562 // In the case of G1 this will record the value of the
563 // referent in an SATB buffer if marking is active.
564 // This will cause concurrent marking to mark the referent
565 // field as live.
566 Unimplemented();
567 }
568 #endif // SERIALGC
570 // If G1 is not enabled then attempt to go through the accessor entry point
571 // Reference.get is an accessor
572 return generate_accessor_entry();
573 }
575 //
576 // Interpreter stub for calling a native method. (C++ interpreter)
577 // This sets up a somewhat different looking stack for calling the native method
578 // than the typical interpreter frame setup.
579 //
581 address InterpreterGenerator::generate_native_entry(bool synchronized) {
582 address entry = __ pc();
584 // the following temporary registers are used during frame creation
585 const Register Gtmp1 = G3_scratch ;
586 const Register Gtmp2 = G1_scratch;
587 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
589 bool inc_counter = UseCompiler || CountCompiledCalls;
591 // make sure registers are different!
592 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
594 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
596 Label Lentry;
597 __ bind(Lentry);
599 __ verify_oop(G5_method);
601 const Register Glocals_size = G3;
602 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
604 // make sure method is native & not abstract
605 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
606 #ifdef ASSERT
607 __ ld(access_flags, Gtmp1);
608 {
609 Label L;
610 __ btst(JVM_ACC_NATIVE, Gtmp1);
611 __ br(Assembler::notZero, false, Assembler::pt, L);
612 __ delayed()->nop();
613 __ stop("tried to execute non-native method as native");
614 __ bind(L);
615 }
616 { Label L;
617 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
618 __ br(Assembler::zero, false, Assembler::pt, L);
619 __ delayed()->nop();
620 __ stop("tried to execute abstract method as non-abstract");
621 __ bind(L);
622 }
623 #endif // ASSERT
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(methodOopDesc::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(methodOopDesc::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(methodOopDesc::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(methodOopDesc:: const_offset())), O1);
773 __ ld_ptr(Address(O1, 0, in_bytes(constMethodOopDesc::constants_offset())), O1);
774 __ ld_ptr(Address(O1, 0, constantPoolOopDesc::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(methodOopDesc::const_offset()));
1053 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1054 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1055 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1056 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1058 // slop factor is two extra slots on the expression stack so that
1059 // we always have room to store a result when returning from a call without parameters
1060 // that returns a result.
1062 const int slop_factor = 2*wordSize;
1064 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1065 //6815692//methodOopDesc::extra_stack_words() + // extra push slots for MH adapters
1066 frame::memory_parameter_word_sp_offset + // register save area + param window
1067 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1069 // XXX G5_method valid
1071 // Now compute new frame size
1073 if (native) {
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 __ lduh(max_stack, Gtmp); // Full size expression stack
1078 }
1079 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1081 __ neg(Gtmp); // negative space for stack/parameters in words
1082 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1083 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1085 // Need to do stack size check here before we fault on large frames
1087 Label stack_ok;
1089 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1090 (StackRedPages+StackYellowPages);
1093 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1094 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1095 // compute stack bottom
1096 __ sub(O0, O1, O0);
1098 // Avoid touching the guard pages
1099 // Also a fudge for frame size of BytecodeInterpreter::run
1100 // It varies from 1k->4k depending on build type
1101 const int fudge = 6 * K;
1103 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1105 __ add(O0, O1, O0);
1106 __ sub(O0, Gtmp, O0);
1107 __ cmp(SP, O0);
1108 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1109 __ delayed()->nop();
1111 // throw exception return address becomes throwing pc
1113 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1114 __ stop("never reached");
1116 __ bind(stack_ok);
1118 __ save(SP, Gtmp, SP); // setup new frame and register window
1120 // New window I7 call_stub or previous activation
1121 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1122 //
1123 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1124 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1126 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1128 // Initialize a new Interpreter state
1129 // orig_sp - caller's original sp
1130 // G2_thread - thread
1131 // Gargs - &locals[0] (unbiased?)
1132 // G5_method - method
1133 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1136 __ set(0xdead0004, O1);
1139 __ st_ptr(Gargs, XXX_STATE(_locals));
1140 __ st_ptr(G0, XXX_STATE(_oop_temp));
1142 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1143 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1144 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1146 if (native) {
1147 __ st_ptr(G0, XXX_STATE(_bcp));
1148 } else {
1149 __ ld_ptr(G5_method, in_bytes(methodOopDesc::const_offset()), O2); // get constMethodOop
1150 __ add(O2, in_bytes(constMethodOopDesc::codes_offset()), O2); // get bcp
1151 __ st_ptr(O2, XXX_STATE(_bcp));
1152 }
1154 __ st_ptr(G0, XXX_STATE(_mdx));
1155 __ st_ptr(G5_method, XXX_STATE(_method));
1157 __ set((int) BytecodeInterpreter::method_entry, O1);
1158 __ st(O1, XXX_STATE(_msg));
1160 __ ld_ptr(constMethod, O3);
1161 __ ld_ptr(O3, in_bytes(constMethodOopDesc::constants_offset()), O3);
1162 __ ld_ptr(O3, constantPoolOopDesc::cache_offset_in_bytes(), O2);
1163 __ st_ptr(O2, XXX_STATE(_constants));
1165 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1167 // Monitor base is just start of BytecodeInterpreter object;
1168 __ mov(state, O2);
1169 __ st_ptr(O2, XXX_STATE(_monitor_base));
1171 // Do we need a monitor for synchonized method?
1172 {
1173 __ ld(access_flags, O1);
1174 Label done;
1175 Label got_obj;
1176 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1177 __ br( Assembler::zero, false, Assembler::pt, done);
1179 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1180 __ delayed()->btst(JVM_ACC_STATIC, O1);
1181 __ ld_ptr(XXX_STATE(_locals), O1);
1182 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1183 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1184 __ ld_ptr(constMethod, O1);
1185 __ ld_ptr( O1, in_bytes(constMethodOopDesc::constants_offset()), O1);
1186 __ ld_ptr( O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
1187 // lock the mirror, not the klassOop
1188 __ ld_ptr( O1, mirror_offset, O1);
1190 __ bind(got_obj);
1192 #ifdef ASSERT
1193 __ tst(O1);
1194 __ breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
1195 #endif // ASSERT
1197 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1198 __ sub(SP, entry_size, SP); // account for initial monitor
1199 __ sub(O2, entry_size, O2); // initial monitor
1200 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1201 __ bind(done);
1202 }
1204 // Remember initial frame bottom
1206 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1208 __ st_ptr(O2, XXX_STATE(_stack_base));
1210 __ sub(O2, wordSize, O2); // prepush
1211 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1213 __ lduh(max_stack, O3); // Full size expression stack
1214 guarantee(!EnableInvokeDynamic, "no support yet for java.lang.invoke.MethodHandle"); //6815692
1215 //6815692//if (EnableInvokeDynamic)
1216 //6815692// __ inc(O3, methodOopDesc::extra_stack_entries());
1217 __ sll(O3, LogBytesPerWord, O3);
1218 __ sub(O2, O3, O3);
1219 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1220 __ st_ptr(O3, XXX_STATE(_stack_limit));
1222 if (!native) {
1223 //
1224 // Code to initialize locals
1225 //
1226 Register init_value = noreg; // will be G0 if we must clear locals
1227 // Now zero locals
1228 if (true /* zerolocals */ || ClearInterpreterLocals) {
1229 // explicitly initialize locals
1230 init_value = G0;
1231 } else {
1232 #ifdef ASSERT
1233 // initialize locals to a garbage pattern for better debugging
1234 init_value = O3;
1235 __ set( 0x0F0F0F0F, init_value );
1236 #endif // ASSERT
1237 }
1238 if (init_value != noreg) {
1239 Label clear_loop;
1241 // NOTE: If you change the frame layout, this code will need to
1242 // be updated!
1243 __ lduh( size_of_locals, O2 );
1244 __ lduh( size_of_parameters, O1 );
1245 __ sll( O2, LogBytesPerWord, O2);
1246 __ sll( O1, LogBytesPerWord, O1 );
1247 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1248 __ sub( L2_scratch, O2, O2 );
1249 __ sub( L2_scratch, O1, O1 );
1251 __ bind( clear_loop );
1252 __ inc( O2, wordSize );
1254 __ cmp( O2, O1 );
1255 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1256 __ delayed()->st_ptr( init_value, O2, 0 );
1257 }
1258 }
1259 }
1260 // Find preallocated monitor and lock method (C++ interpreter)
1261 //
1262 void InterpreterGenerator::lock_method(void) {
1263 // Lock the current method.
1264 // Destroys registers L2_scratch, L3_scratch, O0
1265 //
1266 // Find everything relative to Lstate
1268 #ifdef ASSERT
1269 __ ld_ptr(STATE(_method), L2_scratch);
1270 __ ld(L2_scratch, in_bytes(methodOopDesc::access_flags_offset()), O0);
1272 { Label ok;
1273 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1274 __ br( Assembler::notZero, false, Assembler::pt, ok);
1275 __ delayed()->nop();
1276 __ stop("method doesn't need synchronization");
1277 __ bind(ok);
1278 }
1279 #endif // ASSERT
1281 // monitor is already allocated at stack base
1282 // and the lockee is already present
1283 __ ld_ptr(STATE(_stack_base), L2_scratch);
1284 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1285 __ lock_object(L2_scratch, O0);
1287 }
1289 // Generate code for handling resuming a deopted method
1290 void CppInterpreterGenerator::generate_deopt_handling() {
1292 Label return_from_deopt_common;
1294 // deopt needs to jump to here to enter the interpreter (return a result)
1295 deopt_frame_manager_return_atos = __ pc();
1297 // O0/O1 live
1298 __ ba(return_from_deopt_common);
1299 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1302 // deopt needs to jump to here to enter the interpreter (return a result)
1303 deopt_frame_manager_return_btos = __ pc();
1305 // O0/O1 live
1306 __ ba(return_from_deopt_common);
1307 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1309 // deopt needs to jump to here to enter the interpreter (return a result)
1310 deopt_frame_manager_return_itos = __ pc();
1312 // O0/O1 live
1313 __ ba(return_from_deopt_common);
1314 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1316 // deopt needs to jump to here to enter the interpreter (return a result)
1318 deopt_frame_manager_return_ltos = __ pc();
1319 #if !defined(_LP64) && defined(COMPILER2)
1320 // All return values are where we want them, except for Longs. C2 returns
1321 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1322 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1323 // build even if we are returning from interpreted we just do a little
1324 // stupid shuffing.
1325 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1326 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1327 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1329 __ srl (G1, 0,O1);
1330 __ srlx(G1,32,O0);
1331 #endif /* !_LP64 && COMPILER2 */
1332 // O0/O1 live
1333 __ ba(return_from_deopt_common);
1334 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1336 // deopt needs to jump to here to enter the interpreter (return a result)
1338 deopt_frame_manager_return_ftos = __ pc();
1339 // O0/O1 live
1340 __ ba(return_from_deopt_common);
1341 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1343 // deopt needs to jump to here to enter the interpreter (return a result)
1344 deopt_frame_manager_return_dtos = __ pc();
1346 // O0/O1 live
1347 __ ba(return_from_deopt_common);
1348 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1350 // deopt needs to jump to here to enter the interpreter (return a result)
1351 deopt_frame_manager_return_vtos = __ pc();
1353 // O0/O1 live
1354 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1356 // Deopt return common
1357 // an index is present that lets us move any possible result being
1358 // return to the interpreter's stack
1359 //
1360 __ bind(return_from_deopt_common);
1362 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1363 // stack is in the state that the calling convention left it.
1364 // Copy the result from native abi result and place it on java expression stack.
1366 // Current interpreter state is present in Lstate
1368 // Get current pre-pushed top of interpreter stack
1369 // Any result (if any) is in native abi
1370 // result type index is in L3_scratch
1372 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1374 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1375 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1376 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1377 __ jmpl(Lscratch, G0, O7); // and convert it
1378 __ delayed()->nop();
1380 // L1_scratch points to top of stack (prepushed)
1381 __ st_ptr(L1_scratch, STATE(_stack));
1382 }
1384 // Generate the code to handle a more_monitors message from the c++ interpreter
1385 void CppInterpreterGenerator::generate_more_monitors() {
1387 Label entry, loop;
1388 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1389 // 1. compute new pointers // esp: old expression stack top
1390 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1391 __ sub(L4_scratch, entry_size, L4_scratch);
1392 __ st_ptr(L4_scratch, STATE(_stack_base));
1394 __ sub(SP, entry_size, SP); // Grow stack
1395 __ st_ptr(SP, STATE(_frame_bottom));
1397 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1398 __ sub(L2_scratch, entry_size, L2_scratch);
1399 __ st_ptr(L2_scratch, STATE(_stack_limit));
1401 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1402 __ sub(L1_scratch, entry_size, L1_scratch);
1403 __ st_ptr(L1_scratch, STATE(_stack));
1404 __ ba(entry);
1405 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1407 // 2. move expression stack
1409 __ bind(loop);
1410 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1411 __ add(L1_scratch, wordSize, L1_scratch);
1412 __ bind(entry);
1413 __ cmp(L1_scratch, L4_scratch);
1414 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1415 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1417 // now zero the slot so we can find it.
1418 __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1420 }
1422 // Initial entry to C++ interpreter from the call_stub.
1423 // This entry point is called the frame manager since it handles the generation
1424 // of interpreter activation frames via requests directly from the vm (via call_stub)
1425 // and via requests from the interpreter. The requests from the call_stub happen
1426 // directly thru the entry point. Requests from the interpreter happen via returning
1427 // from the interpreter and examining the message the interpreter has returned to
1428 // the frame manager. The frame manager can take the following requests:
1430 // NO_REQUEST - error, should never happen.
1431 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1432 // allocate a new monitor.
1433 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1434 // happens during entry during the entry via the call stub.
1435 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1436 //
1437 // Arguments:
1438 //
1439 // ebx: methodOop
1440 // ecx: receiver - unused (retrieved from stack as needed)
1441 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1442 //
1443 //
1444 // Stack layout at entry
1445 //
1446 // [ return address ] <--- esp
1447 // [ parameter n ]
1448 // ...
1449 // [ parameter 1 ]
1450 // [ expression stack ]
1451 //
1452 //
1453 // We are free to blow any registers we like because the call_stub which brought us here
1454 // initially has preserved the callee save registers already.
1455 //
1456 //
1458 static address interpreter_frame_manager = NULL;
1460 #ifdef ASSERT
1461 #define VALIDATE_STATE(scratch, marker) \
1462 { \
1463 Label skip; \
1464 __ ld_ptr(STATE(_self_link), scratch); \
1465 __ cmp(Lstate, scratch); \
1466 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1467 __ delayed()->nop(); \
1468 __ breakpoint_trap(); \
1469 __ emit_long(marker); \
1470 __ bind(skip); \
1471 }
1472 #else
1473 #define VALIDATE_STATE(scratch, marker)
1474 #endif /* ASSERT */
1476 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1477 //
1478 // Adjust caller's stack so that all the locals can be contiguous with
1479 // the parameters.
1480 // Worries about stack overflow make this a pain.
1481 //
1482 // Destroys args, G3_scratch, G3_scratch
1483 // In/Out O5_savedSP (sender's original SP)
1484 //
1485 // assert_different_registers(state, prev_state);
1486 const Register Gtmp = G3_scratch;
1487 const Register tmp = O2;
1488 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1489 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1491 __ lduh(size_of_parameters, tmp);
1492 __ sll(tmp, LogBytesPerWord, Gtmp); // parameter size in bytes
1493 __ add(args, Gtmp, Gargs); // points to first local + BytesPerWord
1494 // NEW
1495 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1496 // determine extra space for non-argument locals & adjust caller's SP
1497 // Gtmp1: parameter size in words
1498 __ lduh(size_of_locals, Gtmp);
1499 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1501 #if 1
1502 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1503 // the call_stub will place the final interpreter argument at
1504 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1505 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1506 // and try to make it look good in the debugger we will store the argument to
1507 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1508 // extra space for the compiler this will overwrite locals in the local array of the
1509 // interpreter.
1510 // QQQ still needed with frameless adapters???
1512 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1514 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1515 #endif // 1
1518 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1519 }
1521 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1523 // G5_method: methodOop
1524 // G2_thread: thread (unused)
1525 // Gargs: bottom of args (sender_sp)
1526 // O5: sender's sp
1528 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1529 // the code is pretty much ready. Would need to change the test below and for good measure
1530 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1531 // routines. Not clear this is worth it yet.
1533 if (interpreter_frame_manager) {
1534 return interpreter_frame_manager;
1535 }
1537 __ bind(frame_manager_entry);
1539 // the following temporary registers are used during frame creation
1540 const Register Gtmp1 = G3_scratch;
1541 // const Register Lmirror = L1; // native mirror (native calls only)
1543 const Address constMethod (G5_method, 0, in_bytes(methodOopDesc::const_offset()));
1544 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1545 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1546 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1547 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1549 address entry_point = __ pc();
1550 __ mov(G0, prevState); // no current activation
1553 Label re_dispatch;
1555 __ bind(re_dispatch);
1557 // Interpreter needs to have locals completely contiguous. In order to do that
1558 // We must adjust the caller's stack pointer for any locals beyond just the
1559 // parameters
1560 adjust_callers_stack(Gargs);
1562 // O5_savedSP still contains sender's sp
1564 // NEW FRAME
1566 generate_compute_interpreter_state(Lstate, prevState, false);
1568 // At this point a new interpreter frame and state object are created and initialized
1569 // Lstate has the pointer to the new activation
1570 // Any stack banging or limit check should already be done.
1572 Label call_interpreter;
1574 __ bind(call_interpreter);
1577 #if 1
1578 __ set(0xdead002, Lmirror);
1579 __ set(0xdead002, L2_scratch);
1580 __ set(0xdead003, L3_scratch);
1581 __ set(0xdead004, L4_scratch);
1582 __ set(0xdead005, Lscratch);
1583 __ set(0xdead006, Lscratch2);
1584 __ set(0xdead007, L7_scratch);
1586 __ set(0xdeaf002, O2);
1587 __ set(0xdeaf003, O3);
1588 __ set(0xdeaf004, O4);
1589 __ set(0xdeaf005, O5);
1590 #endif
1592 // Call interpreter (stack bang complete) enter here if message is
1593 // set and we know stack size is valid
1595 Label call_interpreter_2;
1597 __ bind(call_interpreter_2);
1599 #ifdef ASSERT
1600 {
1601 Label skip;
1602 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1603 __ cmp(G3_scratch, SP);
1604 __ brx(Assembler::equal, false, Assembler::pt, skip);
1605 __ delayed()->nop();
1606 __ stop("SP not restored to frame bottom");
1607 __ bind(skip);
1608 }
1609 #endif
1611 VALIDATE_STATE(G3_scratch, 4);
1612 __ set_last_Java_frame(SP, noreg);
1613 __ mov(Lstate, O0); // (arg) pointer to current state
1615 __ call(CAST_FROM_FN_PTR(address,
1616 JvmtiExport::can_post_interpreter_events() ?
1617 BytecodeInterpreter::runWithChecks
1618 : BytecodeInterpreter::run),
1619 relocInfo::runtime_call_type);
1621 __ delayed()->nop();
1623 __ ld_ptr(STATE(_thread), G2_thread);
1624 __ reset_last_Java_frame();
1626 // examine msg from interpreter to determine next action
1627 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1629 __ ld(STATE(_msg), L1_scratch); // Get new message
1631 Label call_method;
1632 Label return_from_interpreted_method;
1633 Label throw_exception;
1634 Label do_OSR;
1635 Label bad_msg;
1636 Label resume_interpreter;
1638 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1639 __ br(Assembler::equal, false, Assembler::pt, call_method);
1640 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1641 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1642 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1643 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1644 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1645 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1646 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1647 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1649 // Allocate more monitor space, shuffle expression stack....
1651 generate_more_monitors();
1653 // new monitor slot allocated, resume the interpreter.
1655 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1656 VALIDATE_STATE(G3_scratch, 5);
1657 __ ba(call_interpreter);
1658 __ delayed()->st(L1_scratch, STATE(_msg));
1660 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1661 unctrap_frame_manager_entry = __ pc();
1663 // QQQ what message do we send
1665 __ ba(call_interpreter);
1666 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1668 //=============================================================================
1669 // Returning from a compiled method into a deopted method. The bytecode at the
1670 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1671 // for the template based interpreter). Any stack space that was used by the
1672 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1673 // so all that we have to do is place any pending result on the expression stack
1674 // and resume execution on the next bytecode.
1676 generate_deopt_handling();
1678 // ready to resume the interpreter
1680 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1681 __ ba(call_interpreter);
1682 __ delayed()->st(L1_scratch, STATE(_msg));
1684 // Current frame has caught an exception we need to dispatch to the
1685 // handler. We can get here because a native interpreter frame caught
1686 // an exception in which case there is no handler and we must rethrow
1687 // If it is a vanilla interpreted frame the we simply drop into the
1688 // interpreter and let it do the lookup.
1690 Interpreter::_rethrow_exception_entry = __ pc();
1692 Label return_with_exception;
1693 Label unwind_and_forward;
1695 // O0: exception
1696 // O7: throwing pc
1698 // We want exception in the thread no matter what we ultimately decide about frame type.
1700 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1701 __ verify_thread();
1702 __ st_ptr(O0, exception_addr);
1704 // get the methodOop
1705 __ ld_ptr(STATE(_method), G5_method);
1707 // if this current frame vanilla or native?
1709 __ ld(access_flags, Gtmp1);
1710 __ btst(JVM_ACC_NATIVE, Gtmp1);
1711 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1712 __ delayed()->nop();
1714 // We drop thru to unwind a native interpreted frame with a pending exception
1715 // We jump here for the initial interpreter frame with exception pending
1716 // We unwind the current acivation and forward it to our caller.
1718 __ bind(unwind_and_forward);
1720 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1721 // as expected by forward_exception.
1723 __ restore(FP, G0, SP); // unwind interpreter state frame
1724 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1725 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1727 // Return point from a call which returns a result in the native abi
1728 // (c1/c2/jni-native). This result must be processed onto the java
1729 // expression stack.
1730 //
1731 // A pending exception may be present in which case there is no result present
1733 address return_from_native_method = __ pc();
1735 VALIDATE_STATE(G3_scratch, 6);
1737 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1738 // stack is in the state that the calling convention left it.
1739 // Copy the result from native abi result and place it on java expression stack.
1741 // Current interpreter state is present in Lstate
1743 // Exception pending?
1745 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1746 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1747 __ tst(Lscratch); // exception pending?
1748 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1749 __ delayed()->nop();
1751 // Process the native abi result to java expression stack
1753 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1754 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1755 __ lduh(L4_scratch, in_bytes(methodOopDesc::size_of_parameters_offset()), L2_scratch); // get parameter size
1756 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1757 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1758 __ ld(L4_scratch, in_bytes(methodOopDesc::result_index_offset()), L3_scratch); // called method result type index
1760 // tosca is really just native abi
1761 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1762 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1763 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1764 __ jmpl(Lscratch, G0, O7); // and convert it
1765 __ delayed()->nop();
1767 // L1_scratch points to top of stack (prepushed)
1769 __ ba(resume_interpreter);
1770 __ delayed()->mov(L1_scratch, O1);
1772 // An exception is being caught on return to a vanilla interpreter frame.
1773 // Empty the stack and resume interpreter
1775 __ bind(return_with_exception);
1777 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1778 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1779 __ ba(resume_interpreter);
1780 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1782 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1783 // interpreter call, or native) and unwind this interpreter activation.
1784 // All monitors should be unlocked.
1786 __ bind(return_from_interpreted_method);
1788 VALIDATE_STATE(G3_scratch, 7);
1790 Label return_to_initial_caller;
1792 // Interpreted result is on the top of the completed activation expression stack.
1793 // We must return it to the top of the callers stack if caller was interpreted
1794 // otherwise we convert to native abi result and return to call_stub/c1/c2
1795 // The caller's expression stack was truncated by the call however the current activation
1796 // has enough stuff on the stack that we have usable space there no matter what. The
1797 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1798 // for the current activation
1800 __ ld_ptr(STATE(_prev_link), L1_scratch);
1801 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1802 __ ld(L2_scratch, in_bytes(methodOopDesc::result_index_offset()), L2_scratch);
1803 __ tst(L1_scratch);
1804 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1805 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1807 // Copy result to callers java stack
1809 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1810 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1811 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1812 __ ld_ptr(STATE(_locals), O1); // stack destination
1814 // O0 - will be source, O1 - will be destination (preserved)
1815 __ jmpl(Lscratch, G0, O7); // and convert it
1816 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1818 // O1 == &locals[0]
1820 // Result is now on caller's stack. Just unwind current activation and resume
1822 Label unwind_recursive_activation;
1825 __ bind(unwind_recursive_activation);
1827 // O1 == &locals[0] (really callers stacktop) for activation now returning
1828 // returning to interpreter method from "recursive" interpreter call
1829 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1830 // to. Now all we must do is unwind the state from the completed call
1832 // Must restore stack
1833 VALIDATE_STATE(G3_scratch, 8);
1835 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1836 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1837 // frame and tell interpreter to resume.
1840 __ mov(O1, I1); // pass back new stack top across activation
1841 // POP FRAME HERE ==================================
1842 __ restore(FP, G0, SP); // unwind interpreter state frame
1843 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1846 // Resume the interpreter. The current frame contains the current interpreter
1847 // state object.
1848 //
1849 // O1 == new java stack pointer
1851 __ bind(resume_interpreter);
1852 VALIDATE_STATE(G3_scratch, 10);
1854 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1856 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1857 __ st(L1_scratch, STATE(_msg));
1858 __ ba(call_interpreter_2);
1859 __ delayed()->st_ptr(O1, STATE(_stack));
1862 // Fast accessor methods share this entry point.
1863 // This works because frame manager is in the same codelet
1864 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1865 // we need to do a little register fixup here once we distinguish the two of them
1866 if (UseFastAccessorMethods && !synchronized) {
1867 // Call stub_return address still in O7
1868 __ bind(fast_accessor_slow_entry_path);
1869 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1870 __ cmp(Gtmp1, O7); // returning to interpreter?
1871 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1872 __ delayed()->nop();
1873 __ ba(re_dispatch);
1874 __ delayed()->mov(G0, prevState); // initial entry
1876 }
1878 // interpreter returning to native code (call_stub/c1/c2)
1879 // convert result and unwind initial activation
1880 // L2_scratch - scaled result type index
1882 __ bind(return_to_initial_caller);
1884 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1885 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1886 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1887 __ jmpl(Lscratch, G0, O7); // and convert it
1888 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1890 Label unwind_initial_activation;
1891 __ bind(unwind_initial_activation);
1893 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1894 // we can return here with an exception that wasn't handled by interpreted code
1895 // how does c1/c2 see it on return?
1897 // compute resulting sp before/after args popped depending upon calling convention
1898 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1899 //
1900 // POP FRAME HERE ==================================
1901 __ restore(FP, G0, SP);
1902 __ retl();
1903 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1905 // OSR request, unwind the current frame and transfer to the OSR entry
1906 // and enter OSR nmethod
1908 __ bind(do_OSR);
1909 Label remove_initial_frame;
1910 __ ld_ptr(STATE(_prev_link), L1_scratch);
1911 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1913 // We are going to pop this frame. Is there another interpreter frame underneath
1914 // it or is it callstub/compiled?
1916 __ tst(L1_scratch);
1917 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1918 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1920 // Frame underneath is an interpreter frame simply unwind
1921 // POP FRAME HERE ==================================
1922 __ restore(FP, G0, SP); // unwind interpreter state frame
1923 __ mov(I5_savedSP->after_restore(), SP);
1925 // Since we are now calling native need to change our "return address" from the
1926 // dummy RecursiveInterpreterActivation to a return from native
1928 __ set((intptr_t)return_from_native_method - 8, O7);
1930 __ jmpl(G3_scratch, G0, G0);
1931 __ delayed()->mov(G1_scratch, O0);
1933 __ bind(remove_initial_frame);
1935 // POP FRAME HERE ==================================
1936 __ restore(FP, G0, SP);
1937 __ mov(I5_savedSP->after_restore(), SP);
1938 __ jmpl(G3_scratch, G0, G0);
1939 __ delayed()->mov(G1_scratch, O0);
1941 // Call a new method. All we do is (temporarily) trim the expression stack
1942 // push a return address to bring us back to here and leap to the new entry.
1943 // At this point we have a topmost frame that was allocated by the frame manager
1944 // which contains the current method interpreted state. We trim this frame
1945 // of excess java expression stack entries and then recurse.
1947 __ bind(call_method);
1949 // stack points to next free location and not top element on expression stack
1950 // method expects sp to be pointing to topmost element
1952 __ ld_ptr(STATE(_thread), G2_thread);
1953 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1956 // SP already takes in to account the 2 extra words we use for slop
1957 // when we call a "static long no_params()" method. So if
1958 // we trim back sp by the amount of unused java expression stack
1959 // there will be automagically the 2 extra words we need.
1960 // We also have to worry about keeping SP aligned.
1962 __ ld_ptr(STATE(_stack), Gargs);
1963 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1965 // compute the unused java stack size
1966 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1968 // Round down the unused space to that stack is always 16-byte aligned
1969 // by making the unused space a multiple of the size of two longs.
1971 __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1973 // Now trim the stack
1974 __ add(SP, L2_scratch, SP);
1977 // Now point to the final argument (account for prepush)
1978 __ add(Gargs, wordSize, Gargs);
1979 #ifdef ASSERT
1980 // Make sure we have space for the window
1981 __ sub(Gargs, SP, L1_scratch);
1982 __ cmp(L1_scratch, 16*wordSize);
1983 {
1984 Label skip;
1985 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1986 __ delayed()->nop();
1987 __ stop("killed stack");
1988 __ bind(skip);
1989 }
1990 #endif // ASSERT
1992 // Create a new frame where we can store values that make it look like the interpreter
1993 // really recursed.
1995 // prepare to recurse or call specialized entry
1997 // First link the registers we need
1999 // make the pc look good in debugger
2000 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
2001 // argument too
2002 __ mov(Lstate, I0);
2004 // Record our sending SP
2005 __ mov(SP, O5_savedSP);
2007 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
2008 __ set((intptr_t) entry_point, L1_scratch);
2009 __ cmp(L1_scratch, L2_scratch);
2010 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
2011 __ delayed()->mov(Lstate, prevState); // link activations
2013 // method uses specialized entry, push a return so we look like call stub setup
2014 // this path will handle fact that result is returned in registers and not
2015 // on the java stack.
2017 __ set((intptr_t)return_from_native_method - 8, O7);
2018 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
2019 __ delayed()->nop();
2021 //
2022 // Bad Message from interpreter
2023 //
2024 __ bind(bad_msg);
2025 __ stop("Bad message from interpreter");
2027 // Interpreted method "returned" with an exception pass it on...
2028 // Pass result, unwind activation and continue/return to interpreter/call_stub
2029 // We handle result (if any) differently based on return to interpreter or call_stub
2031 __ bind(throw_exception);
2032 __ ld_ptr(STATE(_prev_link), L1_scratch);
2033 __ tst(L1_scratch);
2034 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2035 __ delayed()->nop();
2037 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
2038 __ ba(unwind_recursive_activation);
2039 __ delayed()->nop();
2041 interpreter_frame_manager = entry_point;
2042 return entry_point;
2043 }
2045 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2046 : CppInterpreterGenerator(code) {
2047 generate_all(); // down here so it can be "virtual"
2048 }
2051 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2053 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2054 // expression stack, the callee will have callee_extra_locals (so we can account for
2055 // frame extension) and monitor_size for monitors. Basically we need to calculate
2056 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2057 //
2058 //
2059 // The big complicating thing here is that we must ensure that the stack stays properly
2060 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2061 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2062 // the caller so we must ensure that it is properly aligned for our callee.
2063 //
2064 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2065 // stack at all times to deal with the "stack long no_params()" method issue. This
2066 // is "slop_factor" here.
2067 const int slop_factor = 2;
2069 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2070 frame::memory_parameter_word_sp_offset; // register save area + param window
2071 const int extra_stack = 0; //6815692//methodOopDesc::extra_stack_entries();
2072 return (round_to(max_stack +
2073 extra_stack +
2074 slop_factor +
2075 fixed_size +
2076 monitor_size +
2077 (callee_extra_locals * Interpreter::stackElementWords()), WordsPerLong));
2079 }
2081 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
2083 // See call_stub code
2084 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2085 WordsPerLong); // 7 + register save area
2087 // Save space for one monitor to get into the interpreted method in case
2088 // the method is synchronized
2089 int monitor_size = method->is_synchronized() ?
2090 1*frame::interpreter_frame_monitor_size() : 0;
2091 return size_activation_helper(method->max_locals(), method->max_stack(),
2092 monitor_size) + call_stub_size;
2093 }
2095 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2096 frame* caller,
2097 frame* current,
2098 methodOop method,
2099 intptr_t* locals,
2100 intptr_t* stack,
2101 intptr_t* stack_base,
2102 intptr_t* monitor_base,
2103 intptr_t* frame_bottom,
2104 bool is_top_frame
2105 )
2106 {
2107 // What about any vtable?
2108 //
2109 to_fill->_thread = JavaThread::current();
2110 // This gets filled in later but make it something recognizable for now
2111 to_fill->_bcp = method->code_base();
2112 to_fill->_locals = locals;
2113 to_fill->_constants = method->constants()->cache();
2114 to_fill->_method = method;
2115 to_fill->_mdx = NULL;
2116 to_fill->_stack = stack;
2117 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2118 to_fill->_msg = deopt_resume2;
2119 } else {
2120 to_fill->_msg = method_resume;
2121 }
2122 to_fill->_result._to_call._bcp_advance = 0;
2123 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2124 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2125 to_fill->_prev_link = NULL;
2127 // Fill in the registers for the frame
2129 // Need to install _sender_sp. Actually not too hard in C++!
2130 // When the skeletal frames are layed out we fill in a value
2131 // for _sender_sp. That value is only correct for the oldest
2132 // skeletal frame constructed (because there is only a single
2133 // entry for "caller_adjustment". While the skeletal frames
2134 // exist that is good enough. We correct that calculation
2135 // here and get all the frames correct.
2137 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2139 *current->register_addr(Lstate) = (intptr_t) to_fill;
2140 // skeletal already places a useful value here and this doesn't account
2141 // for alignment so don't bother.
2142 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2144 if (caller->is_interpreted_frame()) {
2145 interpreterState prev = caller->get_interpreterState();
2146 to_fill->_prev_link = prev;
2147 // Make the prev callee look proper
2148 prev->_result._to_call._callee = method;
2149 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2150 prev->_result._to_call._bcp_advance = 5;
2151 } else {
2152 prev->_result._to_call._bcp_advance = 3;
2153 }
2154 }
2155 to_fill->_oop_temp = NULL;
2156 to_fill->_stack_base = stack_base;
2157 // Need +1 here because stack_base points to the word just above the first expr stack entry
2158 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2159 // See generate_compute_interpreter_state.
2160 int extra_stack = 0; //6815692//methodOopDesc::extra_stack_entries();
2161 to_fill->_stack_limit = stack_base - (method->max_stack() + 1 + extra_stack);
2162 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2164 // sparc specific
2165 to_fill->_frame_bottom = frame_bottom;
2166 to_fill->_self_link = to_fill;
2167 #ifdef ASSERT
2168 to_fill->_native_fresult = 123456.789;
2169 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2170 #endif
2171 }
2173 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2174 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2175 }
2178 int AbstractInterpreter::layout_activation(methodOop method,
2179 int tempcount, // Number of slots on java expression stack in use
2180 int popframe_extra_args,
2181 int moncount, // Number of active monitors
2182 int caller_actual_parameters,
2183 int callee_param_size,
2184 int callee_locals_size,
2185 frame* caller,
2186 frame* interpreter_frame,
2187 bool is_top_frame) {
2189 assert(popframe_extra_args == 0, "NEED TO FIX");
2190 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2191 // does as far as allocating an interpreter frame.
2192 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2193 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2194 // as determined by a previous call to this method.
2195 // It is also guaranteed to be walkable even though it is in a skeletal state
2196 // NOTE: return size is in words not bytes
2197 // NOTE: tempcount is the current size of the java expression stack. For top most
2198 // frames we will allocate a full sized expression stack and not the curback
2199 // version that non-top frames have.
2201 // Calculate the amount our frame will be adjust by the callee. For top frame
2202 // this is zero.
2204 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2205 // calculates the extra locals based on itself. Not what the callee does
2206 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2207 // as getting sender_sp correct.
2209 int extra_locals_size = callee_locals_size - callee_param_size;
2210 int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2211 int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2212 int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2213 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2216 /*
2217 if we actually have a frame to layout we must now fill in all the pieces. This means both
2218 the interpreterState and the registers.
2219 */
2220 if (interpreter_frame != NULL) {
2222 // MUCHO HACK
2224 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2225 // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2226 assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2227 frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2229 /* Now fillin the interpreterState object */
2231 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2234 intptr_t* locals;
2236 // Calculate the postion of locals[0]. This is painful because of
2237 // stack alignment (same as ia64). The problem is that we can
2238 // not compute the location of locals from fp(). fp() will account
2239 // for the extra locals but it also accounts for aligning the stack
2240 // and we can't determine if the locals[0] was misaligned but max_locals
2241 // was enough to have the
2242 // calculate postion of locals. fp already accounts for extra locals.
2243 // +2 for the static long no_params() issue.
2245 if (caller->is_interpreted_frame()) {
2246 // locals must agree with the caller because it will be used to set the
2247 // caller's tos when we return.
2248 interpreterState prev = caller->get_interpreterState();
2249 // stack() is prepushed.
2250 locals = prev->stack() + method->size_of_parameters();
2251 } else {
2252 // Lay out locals block in the caller adjacent to the register window save area.
2253 //
2254 // Compiled frames do not allocate a varargs area which is why this if
2255 // statement is needed.
2256 //
2257 intptr_t* fp = interpreter_frame->fp();
2258 int local_words = method->max_locals() * Interpreter::stackElementWords();
2260 if (caller->is_compiled_frame()) {
2261 locals = fp + frame::register_save_words + local_words - 1;
2262 } else {
2263 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2264 }
2266 }
2267 // END MUCHO HACK
2269 intptr_t* monitor_base = (intptr_t*) cur_state;
2270 intptr_t* stack_base = monitor_base - monitor_size;
2271 /* +1 because stack is always prepushed */
2272 intptr_t* stack = stack_base - (tempcount + 1);
2275 BytecodeInterpreter::layout_interpreterState(cur_state,
2276 caller,
2277 interpreter_frame,
2278 method,
2279 locals,
2280 stack,
2281 stack_base,
2282 monitor_base,
2283 frame_bottom,
2284 is_top_frame);
2286 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2288 }
2289 return frame_words;
2290 }
2292 #endif // CC_INTERP