Fri, 25 Mar 2011 09:35:39 +0100
7029017: Additional architecture support for c2 compiler
Summary: Enables cross building of a c2 VM. Support masking of shift counts when the processor architecture mandates it.
Reviewed-by: kvn, never
1 /*
2 * Copyright (c) 2007, 2011, 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::constants_offset()), G3_scratch);
494 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
496 // get specific constant pool cache entry
497 __ add(G3_scratch, G1_scratch, G3_scratch);
499 // Check the constant Pool cache entry to see if it has been resolved.
500 // If not, need the slow path.
501 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
502 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
503 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
504 __ and3(G1_scratch, 0xFF, G1_scratch);
505 __ cmp(G1_scratch, Bytecodes::_getfield);
506 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
507 __ delayed()->nop();
509 // Get the type and return field offset from the constant pool cache
510 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
511 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
513 Label xreturn_path;
514 // Need to differentiate between igetfield, agetfield, bgetfield etc.
515 // because they are different sizes.
516 // Get the type from the constant pool cache
517 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
518 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
519 ConstantPoolCacheEntry::verify_tosBits();
520 __ cmp(G1_scratch, atos );
521 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
522 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
523 __ cmp(G1_scratch, itos);
524 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
525 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
526 __ cmp(G1_scratch, stos);
527 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
528 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
529 __ cmp(G1_scratch, ctos);
530 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
531 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
532 #ifdef ASSERT
533 __ cmp(G1_scratch, btos);
534 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
535 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
536 __ should_not_reach_here();
537 #endif
538 __ ldsb(Otos_i, G3_scratch, Otos_i);
539 __ bind(xreturn_path);
541 // _ireturn/_areturn
542 __ retl(); // return from leaf routine
543 __ delayed()->mov(O5_savedSP, SP);
545 // Generate regular method entry
546 __ bind(slow_path);
547 __ ba(false, fast_accessor_slow_entry_path);
548 __ delayed()->nop();
549 return entry;
550 }
551 return NULL;
552 }
554 //
555 // Interpreter stub for calling a native method. (C++ interpreter)
556 // This sets up a somewhat different looking stack for calling the native method
557 // than the typical interpreter frame setup.
558 //
560 address InterpreterGenerator::generate_native_entry(bool synchronized) {
561 address entry = __ pc();
563 // the following temporary registers are used during frame creation
564 const Register Gtmp1 = G3_scratch ;
565 const Register Gtmp2 = G1_scratch;
566 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
568 bool inc_counter = UseCompiler || CountCompiledCalls;
570 // make sure registers are different!
571 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
573 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
575 Label Lentry;
576 __ bind(Lentry);
578 __ verify_oop(G5_method);
580 const Register Glocals_size = G3;
581 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
583 // make sure method is native & not abstract
584 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
585 #ifdef ASSERT
586 __ ld(access_flags, Gtmp1);
587 {
588 Label L;
589 __ btst(JVM_ACC_NATIVE, Gtmp1);
590 __ br(Assembler::notZero, false, Assembler::pt, L);
591 __ delayed()->nop();
592 __ stop("tried to execute non-native method as native");
593 __ bind(L);
594 }
595 { Label L;
596 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
597 __ br(Assembler::zero, false, Assembler::pt, L);
598 __ delayed()->nop();
599 __ stop("tried to execute abstract method as non-abstract");
600 __ bind(L);
601 }
602 #endif // ASSERT
604 __ lduh(size_of_parameters, Gtmp1);
605 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes
606 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord
607 // NEW
608 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
609 // generate the code to allocate the interpreter stack frame
610 // NEW FRAME ALLOCATED HERE
611 // save callers original sp
612 // __ mov(SP, I5_savedSP->after_restore());
614 generate_compute_interpreter_state(Lstate, G0, true);
616 // At this point Lstate points to new interpreter state
617 //
619 const Address do_not_unlock_if_synchronized(G2_thread, 0,
620 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
621 // Since at this point in the method invocation the exception handler
622 // would try to exit the monitor of synchronized methods which hasn't
623 // been entered yet, we set the thread local variable
624 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
625 // runtime, exception handling i.e. unlock_if_synchronized_method will
626 // check this thread local flag.
627 // This flag has two effects, one is to force an unwind in the topmost
628 // interpreter frame and not perform an unlock while doing so.
630 __ movbool(true, G3_scratch);
631 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
634 // increment invocation counter and check for overflow
635 //
636 // Note: checking for negative value instead of overflow
637 // so we have a 'sticky' overflow test (may be of
638 // importance as soon as we have true MT/MP)
639 Label invocation_counter_overflow;
640 if (inc_counter) {
641 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
642 }
643 Label Lcontinue;
644 __ bind(Lcontinue);
646 bang_stack_shadow_pages(true);
647 // reset the _do_not_unlock_if_synchronized flag
648 __ stbool(G0, do_not_unlock_if_synchronized);
650 // check for synchronized methods
651 // Must happen AFTER invocation_counter check, so method is not locked
652 // if counter overflows.
654 if (synchronized) {
655 lock_method();
656 // Don't see how G2_thread is preserved here...
657 // __ verify_thread(); QQQ destroys L0,L1 can't use
658 } else {
659 #ifdef ASSERT
660 { Label ok;
661 __ ld_ptr(STATE(_method), G5_method);
662 __ ld(access_flags, O0);
663 __ btst(JVM_ACC_SYNCHRONIZED, O0);
664 __ br( Assembler::zero, false, Assembler::pt, ok);
665 __ delayed()->nop();
666 __ stop("method needs synchronization");
667 __ bind(ok);
668 }
669 #endif // ASSERT
670 }
672 // start execution
674 // __ verify_thread(); kills L1,L2 can't use at the moment
676 // jvmti/jvmpi support
677 __ notify_method_entry();
679 // native call
681 // (note that O0 is never an oop--at most it is a handle)
682 // It is important not to smash any handles created by this call,
683 // until any oop handle in O0 is dereferenced.
685 // (note that the space for outgoing params is preallocated)
687 // get signature handler
689 Label pending_exception_present;
691 { Label L;
692 __ ld_ptr(STATE(_method), G5_method);
693 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
694 __ tst(G3_scratch);
695 __ brx(Assembler::notZero, false, Assembler::pt, L);
696 __ delayed()->nop();
697 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
698 __ ld_ptr(STATE(_method), G5_method);
700 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
701 __ ld_ptr(exception_addr, G3_scratch);
702 __ br_notnull(G3_scratch, false, Assembler::pn, pending_exception_present);
703 __ delayed()->nop();
704 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
705 __ bind(L);
706 }
708 // Push a new frame so that the args will really be stored in
709 // Copy a few locals across so the new frame has the variables
710 // we need but these values will be dead at the jni call and
711 // therefore not gc volatile like the values in the current
712 // frame (Lstate in particular)
714 // Flush the state pointer to the register save area
715 // Which is the only register we need for a stack walk.
716 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
718 __ mov(Lstate, O1); // Need to pass the state pointer across the frame
720 // Calculate current frame size
721 __ sub(SP, FP, O3); // Calculate negative of current frame size
722 __ save(SP, O3, SP); // Allocate an identical sized frame
724 __ mov(I1, Lstate); // In the "natural" register.
726 // Note I7 has leftover trash. Slow signature handler will fill it in
727 // should we get there. Normal jni call will set reasonable last_Java_pc
728 // below (and fix I7 so the stack trace doesn't have a meaningless frame
729 // in it).
732 // call signature handler
733 __ ld_ptr(STATE(_method), Lmethod);
734 __ ld_ptr(STATE(_locals), Llocals);
736 __ callr(G3_scratch, 0);
737 __ delayed()->nop();
738 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed)
740 { Label not_static;
742 __ ld_ptr(STATE(_method), G5_method);
743 __ ld(access_flags, O0);
744 __ btst(JVM_ACC_STATIC, O0);
745 __ br( Assembler::zero, false, Assembler::pt, not_static);
746 __ delayed()->
747 // get native function entry point(O0 is a good temp until the very end)
748 ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::native_function_offset())), O0);
749 // for static methods insert the mirror argument
750 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
752 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc:: constants_offset())), O1);
753 __ ld_ptr(Address(O1, 0, constantPoolOopDesc::pool_holder_offset_in_bytes()), O1);
754 __ ld_ptr(O1, mirror_offset, O1);
755 // where the mirror handle body is allocated:
756 #ifdef ASSERT
757 if (!PrintSignatureHandlers) // do not dirty the output with this
758 { Label L;
759 __ tst(O1);
760 __ brx(Assembler::notZero, false, Assembler::pt, L);
761 __ delayed()->nop();
762 __ stop("mirror is missing");
763 __ bind(L);
764 }
765 #endif // ASSERT
766 __ st_ptr(O1, STATE(_oop_temp));
767 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add
768 __ bind(not_static);
769 }
771 // At this point, arguments have been copied off of stack into
772 // their JNI positions, which are O1..O5 and SP[68..].
773 // Oops are boxed in-place on the stack, with handles copied to arguments.
774 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
776 #ifdef ASSERT
777 { Label L;
778 __ tst(O0);
779 __ brx(Assembler::notZero, false, Assembler::pt, L);
780 __ delayed()->nop();
781 __ stop("native entry point is missing");
782 __ bind(L);
783 }
784 #endif // ASSERT
786 //
787 // setup the java frame anchor
788 //
789 // The scavenge function only needs to know that the PC of this frame is
790 // in the interpreter method entry code, it doesn't need to know the exact
791 // PC and hence we can use O7 which points to the return address from the
792 // previous call in the code stream (signature handler function)
793 //
794 // The other trick is we set last_Java_sp to FP instead of the usual SP because
795 // we have pushed the extra frame in order to protect the volatile register(s)
796 // in that frame when we return from the jni call
797 //
800 __ set_last_Java_frame(FP, O7);
801 __ mov(O7, I7); // make dummy interpreter frame look like one above,
802 // not meaningless information that'll confuse me.
804 // flush the windows now. We don't care about the current (protection) frame
805 // only the outer frames
807 __ flush_windows();
809 // mark windows as flushed
810 Address flags(G2_thread,
811 0,
812 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
813 __ set(JavaFrameAnchor::flushed, G3_scratch);
814 __ st(G3_scratch, flags);
816 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
818 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
819 #ifdef ASSERT
820 { Label L;
821 __ ld(thread_state, G3_scratch);
822 __ cmp(G3_scratch, _thread_in_Java);
823 __ br(Assembler::equal, false, Assembler::pt, L);
824 __ delayed()->nop();
825 __ stop("Wrong thread state in native stub");
826 __ bind(L);
827 }
828 #endif // ASSERT
829 __ set(_thread_in_native, G3_scratch);
830 __ st(G3_scratch, thread_state);
832 // Call the jni method, using the delay slot to set the JNIEnv* argument.
833 __ callr(O0, 0);
834 __ delayed()->
835 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
836 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
838 // must we block?
840 // Block, if necessary, before resuming in _thread_in_Java state.
841 // In order for GC to work, don't clear the last_Java_sp until after blocking.
842 { Label no_block;
843 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
845 // Switch thread to "native transition" state before reading the synchronization state.
846 // This additional state is necessary because reading and testing the synchronization
847 // state is not atomic w.r.t. GC, as this scenario demonstrates:
848 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
849 // VM thread changes sync state to synchronizing and suspends threads for GC.
850 // Thread A is resumed to finish this native method, but doesn't block here since it
851 // didn't see any synchronization is progress, and escapes.
852 __ set(_thread_in_native_trans, G3_scratch);
853 __ st(G3_scratch, thread_state);
854 if(os::is_MP()) {
855 // Write serialization page so VM thread can do a pseudo remote membar.
856 // We use the current thread pointer to calculate a thread specific
857 // offset to write to within the page. This minimizes bus traffic
858 // due to cache line collision.
859 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
860 }
861 __ load_contents(sync_state, G3_scratch);
862 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
865 Label L;
866 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
867 __ br(Assembler::notEqual, false, Assembler::pn, L);
868 __ delayed()->
869 ld(suspend_state, G3_scratch);
870 __ cmp(G3_scratch, 0);
871 __ br(Assembler::equal, false, Assembler::pt, no_block);
872 __ delayed()->nop();
873 __ bind(L);
875 // Block. Save any potential method result value before the operation and
876 // use a leaf call to leave the last_Java_frame setup undisturbed.
877 save_native_result();
878 __ call_VM_leaf(noreg,
879 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
880 G2_thread);
881 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
882 // Restore any method result value
883 restore_native_result();
884 __ bind(no_block);
885 }
887 // Clear the frame anchor now
889 __ reset_last_Java_frame();
891 // Move the result handler address
892 __ mov(Lscratch, G3_scratch);
893 // return possible result to the outer frame
894 #ifndef __LP64
895 __ mov(O0, I0);
896 __ restore(O1, G0, O1);
897 #else
898 __ restore(O0, G0, O0);
899 #endif /* __LP64 */
901 // Move result handler to expected register
902 __ mov(G3_scratch, Lscratch);
905 // thread state is thread_in_native_trans. Any safepoint blocking has
906 // happened in the trampoline we are ready to switch to thread_in_Java.
908 __ set(_thread_in_Java, G3_scratch);
909 __ st(G3_scratch, thread_state);
911 // If we have an oop result store it where it will be safe for any further gc
912 // until we return now that we've released the handle it might be protected by
914 {
915 Label no_oop, store_result;
917 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
918 __ cmp(G3_scratch, Lscratch);
919 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
920 __ delayed()->nop();
921 __ addcc(G0, O0, O0);
922 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
923 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
924 __ mov(G0, O0);
926 __ bind(store_result);
927 // Store it where gc will look for it and result handler expects it.
928 __ st_ptr(O0, STATE(_oop_temp));
930 __ bind(no_oop);
932 }
934 // reset handle block
935 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
936 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
939 // handle exceptions (exception handling will handle unlocking!)
940 { Label L;
941 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
943 __ ld_ptr(exception_addr, Gtemp);
944 __ tst(Gtemp);
945 __ brx(Assembler::equal, false, Assembler::pt, L);
946 __ delayed()->nop();
947 __ bind(pending_exception_present);
948 // With c++ interpreter we just leave it pending caller will do the correct thing. However...
949 // Like x86 we ignore the result of the native call and leave the method locked. This
950 // seems wrong to leave things locked.
952 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
953 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
955 __ bind(L);
956 }
958 // jvmdi/jvmpi support (preserves thread register)
959 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
961 if (synchronized) {
962 // save and restore any potential method result value around the unlocking operation
963 save_native_result();
965 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
966 // Get the initial monitor we allocated
967 __ sub(Lstate, entry_size, O1); // initial monitor
968 __ unlock_object(O1);
969 restore_native_result();
970 }
972 #if defined(COMPILER2) && !defined(_LP64)
974 // C2 expects long results in G1 we can't tell if we're returning to interpreted
975 // or compiled so just be safe.
977 __ sllx(O0, 32, G1); // Shift bits into high G1
978 __ srl (O1, 0, O1); // Zero extend O1
979 __ or3 (O1, G1, G1); // OR 64 bits into G1
981 #endif /* COMPILER2 && !_LP64 */
983 #ifdef ASSERT
984 {
985 Label ok;
986 __ cmp(I5_savedSP, FP);
987 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
988 __ delayed()->nop();
989 __ stop("bad I5_savedSP value");
990 __ should_not_reach_here();
991 __ bind(ok);
992 }
993 #endif
994 // Calls result handler which POPS FRAME
995 if (TraceJumps) {
996 // Move target to register that is recordable
997 __ mov(Lscratch, G3_scratch);
998 __ JMP(G3_scratch, 0);
999 } else {
1000 __ jmp(Lscratch, 0);
1001 }
1002 __ delayed()->nop();
1004 if (inc_counter) {
1005 // handle invocation counter overflow
1006 __ bind(invocation_counter_overflow);
1007 generate_counter_overflow(Lcontinue);
1008 }
1011 return entry;
1012 }
1014 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1015 const Register prev_state,
1016 bool native) {
1018 // On entry
1019 // G5_method - caller's method
1020 // Gargs - points to initial parameters (i.e. locals[0])
1021 // G2_thread - valid? (C1 only??)
1022 // "prev_state" - contains any previous frame manager state which we must save a link
1023 //
1024 // On return
1025 // "state" is a pointer to the newly allocated state object. We must allocate and initialize
1026 // a new interpretState object and the method expression stack.
1028 assert_different_registers(state, prev_state);
1029 assert_different_registers(prev_state, G3_scratch);
1030 const Register Gtmp = G3_scratch;
1031 const Address constants (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1032 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1033 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1034 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1035 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1037 // slop factor is two extra slots on the expression stack so that
1038 // we always have room to store a result when returning from a call without parameters
1039 // that returns a result.
1041 const int slop_factor = 2*wordSize;
1043 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1044 //6815692//methodOopDesc::extra_stack_words() + // extra push slots for MH adapters
1045 frame::memory_parameter_word_sp_offset + // register save area + param window
1046 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1048 // XXX G5_method valid
1050 // Now compute new frame size
1052 if (native) {
1053 __ lduh( size_of_parameters, Gtmp );
1054 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words
1055 } else {
1056 __ lduh(max_stack, Gtmp); // Full size expression stack
1057 }
1058 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1060 __ neg(Gtmp); // negative space for stack/parameters in words
1061 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1062 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1064 // Need to do stack size check here before we fault on large frames
1066 Label stack_ok;
1068 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1069 (StackRedPages+StackYellowPages);
1072 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1073 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1074 // compute stack bottom
1075 __ sub(O0, O1, O0);
1077 // Avoid touching the guard pages
1078 // Also a fudge for frame size of BytecodeInterpreter::run
1079 // It varies from 1k->4k depending on build type
1080 const int fudge = 6 * K;
1082 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1084 __ add(O0, O1, O0);
1085 __ sub(O0, Gtmp, O0);
1086 __ cmp(SP, O0);
1087 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1088 __ delayed()->nop();
1090 // throw exception return address becomes throwing pc
1092 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1093 __ stop("never reached");
1095 __ bind(stack_ok);
1097 __ save(SP, Gtmp, SP); // setup new frame and register window
1099 // New window I7 call_stub or previous activation
1100 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1101 //
1102 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1103 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1105 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1107 // Initialize a new Interpreter state
1108 // orig_sp - caller's original sp
1109 // G2_thread - thread
1110 // Gargs - &locals[0] (unbiased?)
1111 // G5_method - method
1112 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1115 __ set(0xdead0004, O1);
1118 __ st_ptr(Gargs, XXX_STATE(_locals));
1119 __ st_ptr(G0, XXX_STATE(_oop_temp));
1121 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1122 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1123 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1125 if (native) {
1126 __ st_ptr(G0, XXX_STATE(_bcp));
1127 } else {
1128 __ ld_ptr(G5_method, in_bytes(methodOopDesc::const_offset()), O2); // get constMethodOop
1129 __ add(O2, in_bytes(constMethodOopDesc::codes_offset()), O2); // get bcp
1130 __ st_ptr(O2, XXX_STATE(_bcp));
1131 }
1133 __ st_ptr(G0, XXX_STATE(_mdx));
1134 __ st_ptr(G5_method, XXX_STATE(_method));
1136 __ set((int) BytecodeInterpreter::method_entry, O1);
1137 __ st(O1, XXX_STATE(_msg));
1139 __ ld_ptr(constants, O3);
1140 __ ld_ptr(O3, constantPoolOopDesc::cache_offset_in_bytes(), O2);
1141 __ st_ptr(O2, XXX_STATE(_constants));
1143 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1145 // Monitor base is just start of BytecodeInterpreter object;
1146 __ mov(state, O2);
1147 __ st_ptr(O2, XXX_STATE(_monitor_base));
1149 // Do we need a monitor for synchonized method?
1150 {
1151 __ ld(access_flags, O1);
1152 Label done;
1153 Label got_obj;
1154 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1155 __ br( Assembler::zero, false, Assembler::pt, done);
1157 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1158 __ delayed()->btst(JVM_ACC_STATIC, O1);
1159 __ ld_ptr(XXX_STATE(_locals), O1);
1160 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1161 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1162 __ ld_ptr(constants, O1);
1163 __ ld_ptr( O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
1164 // lock the mirror, not the klassOop
1165 __ ld_ptr( O1, mirror_offset, O1);
1167 __ bind(got_obj);
1169 #ifdef ASSERT
1170 __ tst(O1);
1171 __ breakpoint_trap(Assembler::zero);
1172 #endif // ASSERT
1174 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1175 __ sub(SP, entry_size, SP); // account for initial monitor
1176 __ sub(O2, entry_size, O2); // initial monitor
1177 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1178 __ bind(done);
1179 }
1181 // Remember initial frame bottom
1183 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1185 __ st_ptr(O2, XXX_STATE(_stack_base));
1187 __ sub(O2, wordSize, O2); // prepush
1188 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1190 __ lduh(max_stack, O3); // Full size expression stack
1191 guarantee(!EnableMethodHandles, "no support yet for java.lang.invoke.MethodHandle"); //6815692
1192 //6815692//if (EnableMethodHandles)
1193 //6815692// __ inc(O3, methodOopDesc::extra_stack_entries());
1194 __ sll(O3, LogBytesPerWord, O3);
1195 __ sub(O2, O3, O3);
1196 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1197 __ st_ptr(O3, XXX_STATE(_stack_limit));
1199 if (!native) {
1200 //
1201 // Code to initialize locals
1202 //
1203 Register init_value = noreg; // will be G0 if we must clear locals
1204 // Now zero locals
1205 if (true /* zerolocals */ || ClearInterpreterLocals) {
1206 // explicitly initialize locals
1207 init_value = G0;
1208 } else {
1209 #ifdef ASSERT
1210 // initialize locals to a garbage pattern for better debugging
1211 init_value = O3;
1212 __ set( 0x0F0F0F0F, init_value );
1213 #endif // ASSERT
1214 }
1215 if (init_value != noreg) {
1216 Label clear_loop;
1218 // NOTE: If you change the frame layout, this code will need to
1219 // be updated!
1220 __ lduh( size_of_locals, O2 );
1221 __ lduh( size_of_parameters, O1 );
1222 __ sll( O2, LogBytesPerWord, O2);
1223 __ sll( O1, LogBytesPerWord, O1 );
1224 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1225 __ sub( L2_scratch, O2, O2 );
1226 __ sub( L2_scratch, O1, O1 );
1228 __ bind( clear_loop );
1229 __ inc( O2, wordSize );
1231 __ cmp( O2, O1 );
1232 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1233 __ delayed()->st_ptr( init_value, O2, 0 );
1234 }
1235 }
1236 }
1237 // Find preallocated monitor and lock method (C++ interpreter)
1238 //
1239 void InterpreterGenerator::lock_method(void) {
1240 // Lock the current method.
1241 // Destroys registers L2_scratch, L3_scratch, O0
1242 //
1243 // Find everything relative to Lstate
1245 #ifdef ASSERT
1246 __ ld_ptr(STATE(_method), L2_scratch);
1247 __ ld(L2_scratch, in_bytes(methodOopDesc::access_flags_offset()), O0);
1249 { Label ok;
1250 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1251 __ br( Assembler::notZero, false, Assembler::pt, ok);
1252 __ delayed()->nop();
1253 __ stop("method doesn't need synchronization");
1254 __ bind(ok);
1255 }
1256 #endif // ASSERT
1258 // monitor is already allocated at stack base
1259 // and the lockee is already present
1260 __ ld_ptr(STATE(_stack_base), L2_scratch);
1261 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1262 __ lock_object(L2_scratch, O0);
1264 }
1266 // Generate code for handling resuming a deopted method
1267 void CppInterpreterGenerator::generate_deopt_handling() {
1269 Label return_from_deopt_common;
1271 // deopt needs to jump to here to enter the interpreter (return a result)
1272 deopt_frame_manager_return_atos = __ pc();
1274 // O0/O1 live
1275 __ ba(false, return_from_deopt_common);
1276 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1279 // deopt needs to jump to here to enter the interpreter (return a result)
1280 deopt_frame_manager_return_btos = __ pc();
1282 // O0/O1 live
1283 __ ba(false, return_from_deopt_common);
1284 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1286 // deopt needs to jump to here to enter the interpreter (return a result)
1287 deopt_frame_manager_return_itos = __ pc();
1289 // O0/O1 live
1290 __ ba(false, return_from_deopt_common);
1291 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1293 // deopt needs to jump to here to enter the interpreter (return a result)
1295 deopt_frame_manager_return_ltos = __ pc();
1296 #if !defined(_LP64) && defined(COMPILER2)
1297 // All return values are where we want them, except for Longs. C2 returns
1298 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1299 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1300 // build even if we are returning from interpreted we just do a little
1301 // stupid shuffing.
1302 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1303 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1304 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1306 __ srl (G1, 0,O1);
1307 __ srlx(G1,32,O0);
1308 #endif /* !_LP64 && COMPILER2 */
1309 // O0/O1 live
1310 __ ba(false, return_from_deopt_common);
1311 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1313 // deopt needs to jump to here to enter the interpreter (return a result)
1315 deopt_frame_manager_return_ftos = __ pc();
1316 // O0/O1 live
1317 __ ba(false, return_from_deopt_common);
1318 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1320 // deopt needs to jump to here to enter the interpreter (return a result)
1321 deopt_frame_manager_return_dtos = __ pc();
1323 // O0/O1 live
1324 __ ba(false, return_from_deopt_common);
1325 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1327 // deopt needs to jump to here to enter the interpreter (return a result)
1328 deopt_frame_manager_return_vtos = __ pc();
1330 // O0/O1 live
1331 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1333 // Deopt return common
1334 // an index is present that lets us move any possible result being
1335 // return to the interpreter's stack
1336 //
1337 __ bind(return_from_deopt_common);
1339 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1340 // stack is in the state that the calling convention left it.
1341 // Copy the result from native abi result and place it on java expression stack.
1343 // Current interpreter state is present in Lstate
1345 // Get current pre-pushed top of interpreter stack
1346 // Any result (if any) is in native abi
1347 // result type index is in L3_scratch
1349 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1351 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1352 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1353 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1354 __ jmpl(Lscratch, G0, O7); // and convert it
1355 __ delayed()->nop();
1357 // L1_scratch points to top of stack (prepushed)
1358 __ st_ptr(L1_scratch, STATE(_stack));
1359 }
1361 // Generate the code to handle a more_monitors message from the c++ interpreter
1362 void CppInterpreterGenerator::generate_more_monitors() {
1364 Label entry, loop;
1365 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1366 // 1. compute new pointers // esp: old expression stack top
1367 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1368 __ sub(L4_scratch, entry_size, L4_scratch);
1369 __ st_ptr(L4_scratch, STATE(_stack_base));
1371 __ sub(SP, entry_size, SP); // Grow stack
1372 __ st_ptr(SP, STATE(_frame_bottom));
1374 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1375 __ sub(L2_scratch, entry_size, L2_scratch);
1376 __ st_ptr(L2_scratch, STATE(_stack_limit));
1378 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1379 __ sub(L1_scratch, entry_size, L1_scratch);
1380 __ st_ptr(L1_scratch, STATE(_stack));
1381 __ ba(false, entry);
1382 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1384 // 2. move expression stack
1386 __ bind(loop);
1387 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1388 __ add(L1_scratch, wordSize, L1_scratch);
1389 __ bind(entry);
1390 __ cmp(L1_scratch, L4_scratch);
1391 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1392 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1394 // now zero the slot so we can find it.
1395 __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1397 }
1399 // Initial entry to C++ interpreter from the call_stub.
1400 // This entry point is called the frame manager since it handles the generation
1401 // of interpreter activation frames via requests directly from the vm (via call_stub)
1402 // and via requests from the interpreter. The requests from the call_stub happen
1403 // directly thru the entry point. Requests from the interpreter happen via returning
1404 // from the interpreter and examining the message the interpreter has returned to
1405 // the frame manager. The frame manager can take the following requests:
1407 // NO_REQUEST - error, should never happen.
1408 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1409 // allocate a new monitor.
1410 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1411 // happens during entry during the entry via the call stub.
1412 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1413 //
1414 // Arguments:
1415 //
1416 // ebx: methodOop
1417 // ecx: receiver - unused (retrieved from stack as needed)
1418 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1419 //
1420 //
1421 // Stack layout at entry
1422 //
1423 // [ return address ] <--- esp
1424 // [ parameter n ]
1425 // ...
1426 // [ parameter 1 ]
1427 // [ expression stack ]
1428 //
1429 //
1430 // We are free to blow any registers we like because the call_stub which brought us here
1431 // initially has preserved the callee save registers already.
1432 //
1433 //
1435 static address interpreter_frame_manager = NULL;
1437 #ifdef ASSERT
1438 #define VALIDATE_STATE(scratch, marker) \
1439 { \
1440 Label skip; \
1441 __ ld_ptr(STATE(_self_link), scratch); \
1442 __ cmp(Lstate, scratch); \
1443 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1444 __ delayed()->nop(); \
1445 __ breakpoint_trap(); \
1446 __ emit_long(marker); \
1447 __ bind(skip); \
1448 }
1449 #else
1450 #define VALIDATE_STATE(scratch, marker)
1451 #endif /* ASSERT */
1453 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1454 //
1455 // Adjust caller's stack so that all the locals can be contiguous with
1456 // the parameters.
1457 // Worries about stack overflow make this a pain.
1458 //
1459 // Destroys args, G3_scratch, G3_scratch
1460 // In/Out O5_savedSP (sender's original SP)
1461 //
1462 // assert_different_registers(state, prev_state);
1463 const Register Gtmp = G3_scratch;
1464 const Register tmp = O2;
1465 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1466 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1468 __ lduh(size_of_parameters, tmp);
1469 __ sll(tmp, LogBytesPerWord, Gtmp); // parameter size in bytes
1470 __ add(args, Gtmp, Gargs); // points to first local + BytesPerWord
1471 // NEW
1472 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1473 // determine extra space for non-argument locals & adjust caller's SP
1474 // Gtmp1: parameter size in words
1475 __ lduh(size_of_locals, Gtmp);
1476 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1478 #if 1
1479 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1480 // the call_stub will place the final interpreter argument at
1481 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1482 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1483 // and try to make it look good in the debugger we will store the argument to
1484 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1485 // extra space for the compiler this will overwrite locals in the local array of the
1486 // interpreter.
1487 // QQQ still needed with frameless adapters???
1489 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1491 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1492 #endif // 1
1495 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1496 }
1498 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1500 // G5_method: methodOop
1501 // G2_thread: thread (unused)
1502 // Gargs: bottom of args (sender_sp)
1503 // O5: sender's sp
1505 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1506 // the code is pretty much ready. Would need to change the test below and for good measure
1507 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1508 // routines. Not clear this is worth it yet.
1510 if (interpreter_frame_manager) {
1511 return interpreter_frame_manager;
1512 }
1514 __ bind(frame_manager_entry);
1516 // the following temporary registers are used during frame creation
1517 const Register Gtmp1 = G3_scratch;
1518 // const Register Lmirror = L1; // native mirror (native calls only)
1520 const Address constants (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1521 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1522 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1523 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1524 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1526 address entry_point = __ pc();
1527 __ mov(G0, prevState); // no current activation
1530 Label re_dispatch;
1532 __ bind(re_dispatch);
1534 // Interpreter needs to have locals completely contiguous. In order to do that
1535 // We must adjust the caller's stack pointer for any locals beyond just the
1536 // parameters
1537 adjust_callers_stack(Gargs);
1539 // O5_savedSP still contains sender's sp
1541 // NEW FRAME
1543 generate_compute_interpreter_state(Lstate, prevState, false);
1545 // At this point a new interpreter frame and state object are created and initialized
1546 // Lstate has the pointer to the new activation
1547 // Any stack banging or limit check should already be done.
1549 Label call_interpreter;
1551 __ bind(call_interpreter);
1554 #if 1
1555 __ set(0xdead002, Lmirror);
1556 __ set(0xdead002, L2_scratch);
1557 __ set(0xdead003, L3_scratch);
1558 __ set(0xdead004, L4_scratch);
1559 __ set(0xdead005, Lscratch);
1560 __ set(0xdead006, Lscratch2);
1561 __ set(0xdead007, L7_scratch);
1563 __ set(0xdeaf002, O2);
1564 __ set(0xdeaf003, O3);
1565 __ set(0xdeaf004, O4);
1566 __ set(0xdeaf005, O5);
1567 #endif
1569 // Call interpreter (stack bang complete) enter here if message is
1570 // set and we know stack size is valid
1572 Label call_interpreter_2;
1574 __ bind(call_interpreter_2);
1576 #ifdef ASSERT
1577 {
1578 Label skip;
1579 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1580 __ cmp(G3_scratch, SP);
1581 __ brx(Assembler::equal, false, Assembler::pt, skip);
1582 __ delayed()->nop();
1583 __ stop("SP not restored to frame bottom");
1584 __ bind(skip);
1585 }
1586 #endif
1588 VALIDATE_STATE(G3_scratch, 4);
1589 __ set_last_Java_frame(SP, noreg);
1590 __ mov(Lstate, O0); // (arg) pointer to current state
1592 __ call(CAST_FROM_FN_PTR(address,
1593 JvmtiExport::can_post_interpreter_events() ?
1594 BytecodeInterpreter::runWithChecks
1595 : BytecodeInterpreter::run),
1596 relocInfo::runtime_call_type);
1598 __ delayed()->nop();
1600 __ ld_ptr(STATE(_thread), G2_thread);
1601 __ reset_last_Java_frame();
1603 // examine msg from interpreter to determine next action
1604 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1606 __ ld(STATE(_msg), L1_scratch); // Get new message
1608 Label call_method;
1609 Label return_from_interpreted_method;
1610 Label throw_exception;
1611 Label do_OSR;
1612 Label bad_msg;
1613 Label resume_interpreter;
1615 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1616 __ br(Assembler::equal, false, Assembler::pt, call_method);
1617 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1618 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1619 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1620 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1621 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1622 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1623 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1624 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1626 // Allocate more monitor space, shuffle expression stack....
1628 generate_more_monitors();
1630 // new monitor slot allocated, resume the interpreter.
1632 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1633 VALIDATE_STATE(G3_scratch, 5);
1634 __ ba(false, call_interpreter);
1635 __ delayed()->st(L1_scratch, STATE(_msg));
1637 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1638 unctrap_frame_manager_entry = __ pc();
1640 // QQQ what message do we send
1642 __ ba(false, call_interpreter);
1643 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1645 //=============================================================================
1646 // Returning from a compiled method into a deopted method. The bytecode at the
1647 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1648 // for the template based interpreter). Any stack space that was used by the
1649 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1650 // so all that we have to do is place any pending result on the expression stack
1651 // and resume execution on the next bytecode.
1653 generate_deopt_handling();
1655 // ready to resume the interpreter
1657 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1658 __ ba(false, call_interpreter);
1659 __ delayed()->st(L1_scratch, STATE(_msg));
1661 // Current frame has caught an exception we need to dispatch to the
1662 // handler. We can get here because a native interpreter frame caught
1663 // an exception in which case there is no handler and we must rethrow
1664 // If it is a vanilla interpreted frame the we simply drop into the
1665 // interpreter and let it do the lookup.
1667 Interpreter::_rethrow_exception_entry = __ pc();
1669 Label return_with_exception;
1670 Label unwind_and_forward;
1672 // O0: exception
1673 // O7: throwing pc
1675 // We want exception in the thread no matter what we ultimately decide about frame type.
1677 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1678 __ verify_thread();
1679 __ st_ptr(O0, exception_addr);
1681 // get the methodOop
1682 __ ld_ptr(STATE(_method), G5_method);
1684 // if this current frame vanilla or native?
1686 __ ld(access_flags, Gtmp1);
1687 __ btst(JVM_ACC_NATIVE, Gtmp1);
1688 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1689 __ delayed()->nop();
1691 // We drop thru to unwind a native interpreted frame with a pending exception
1692 // We jump here for the initial interpreter frame with exception pending
1693 // We unwind the current acivation and forward it to our caller.
1695 __ bind(unwind_and_forward);
1697 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1698 // as expected by forward_exception.
1700 __ restore(FP, G0, SP); // unwind interpreter state frame
1701 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1702 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1704 // Return point from a call which returns a result in the native abi
1705 // (c1/c2/jni-native). This result must be processed onto the java
1706 // expression stack.
1707 //
1708 // A pending exception may be present in which case there is no result present
1710 address return_from_native_method = __ pc();
1712 VALIDATE_STATE(G3_scratch, 6);
1714 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1715 // stack is in the state that the calling convention left it.
1716 // Copy the result from native abi result and place it on java expression stack.
1718 // Current interpreter state is present in Lstate
1720 // Exception pending?
1722 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1723 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1724 __ tst(Lscratch); // exception pending?
1725 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1726 __ delayed()->nop();
1728 // Process the native abi result to java expression stack
1730 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1731 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1732 __ lduh(L4_scratch, in_bytes(methodOopDesc::size_of_parameters_offset()), L2_scratch); // get parameter size
1733 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1734 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1735 __ ld(L4_scratch, in_bytes(methodOopDesc::result_index_offset()), L3_scratch); // called method result type index
1737 // tosca is really just native abi
1738 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1739 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1740 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1741 __ jmpl(Lscratch, G0, O7); // and convert it
1742 __ delayed()->nop();
1744 // L1_scratch points to top of stack (prepushed)
1746 __ ba(false, resume_interpreter);
1747 __ delayed()->mov(L1_scratch, O1);
1749 // An exception is being caught on return to a vanilla interpreter frame.
1750 // Empty the stack and resume interpreter
1752 __ bind(return_with_exception);
1754 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1755 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1756 __ ba(false, resume_interpreter);
1757 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1759 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1760 // interpreter call, or native) and unwind this interpreter activation.
1761 // All monitors should be unlocked.
1763 __ bind(return_from_interpreted_method);
1765 VALIDATE_STATE(G3_scratch, 7);
1767 Label return_to_initial_caller;
1769 // Interpreted result is on the top of the completed activation expression stack.
1770 // We must return it to the top of the callers stack if caller was interpreted
1771 // otherwise we convert to native abi result and return to call_stub/c1/c2
1772 // The caller's expression stack was truncated by the call however the current activation
1773 // has enough stuff on the stack that we have usable space there no matter what. The
1774 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1775 // for the current activation
1777 __ ld_ptr(STATE(_prev_link), L1_scratch);
1778 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1779 __ ld(L2_scratch, in_bytes(methodOopDesc::result_index_offset()), L2_scratch);
1780 __ tst(L1_scratch);
1781 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1782 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1784 // Copy result to callers java stack
1786 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1787 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1788 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1789 __ ld_ptr(STATE(_locals), O1); // stack destination
1791 // O0 - will be source, O1 - will be destination (preserved)
1792 __ jmpl(Lscratch, G0, O7); // and convert it
1793 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1795 // O1 == &locals[0]
1797 // Result is now on caller's stack. Just unwind current activation and resume
1799 Label unwind_recursive_activation;
1802 __ bind(unwind_recursive_activation);
1804 // O1 == &locals[0] (really callers stacktop) for activation now returning
1805 // returning to interpreter method from "recursive" interpreter call
1806 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1807 // to. Now all we must do is unwind the state from the completed call
1809 // Must restore stack
1810 VALIDATE_STATE(G3_scratch, 8);
1812 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1813 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1814 // frame and tell interpreter to resume.
1817 __ mov(O1, I1); // pass back new stack top across activation
1818 // POP FRAME HERE ==================================
1819 __ restore(FP, G0, SP); // unwind interpreter state frame
1820 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1823 // Resume the interpreter. The current frame contains the current interpreter
1824 // state object.
1825 //
1826 // O1 == new java stack pointer
1828 __ bind(resume_interpreter);
1829 VALIDATE_STATE(G3_scratch, 10);
1831 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1833 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1834 __ st(L1_scratch, STATE(_msg));
1835 __ ba(false, call_interpreter_2);
1836 __ delayed()->st_ptr(O1, STATE(_stack));
1839 // Fast accessor methods share this entry point.
1840 // This works because frame manager is in the same codelet
1841 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1842 // we need to do a little register fixup here once we distinguish the two of them
1843 if (UseFastAccessorMethods && !synchronized) {
1844 // Call stub_return address still in O7
1845 __ bind(fast_accessor_slow_entry_path);
1846 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1847 __ cmp(Gtmp1, O7); // returning to interpreter?
1848 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1849 __ delayed()->nop();
1850 __ ba(false, re_dispatch);
1851 __ delayed()->mov(G0, prevState); // initial entry
1853 }
1855 // interpreter returning to native code (call_stub/c1/c2)
1856 // convert result and unwind initial activation
1857 // L2_scratch - scaled result type index
1859 __ bind(return_to_initial_caller);
1861 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1862 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1863 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1864 __ jmpl(Lscratch, G0, O7); // and convert it
1865 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1867 Label unwind_initial_activation;
1868 __ bind(unwind_initial_activation);
1870 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1871 // we can return here with an exception that wasn't handled by interpreted code
1872 // how does c1/c2 see it on return?
1874 // compute resulting sp before/after args popped depending upon calling convention
1875 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1876 //
1877 // POP FRAME HERE ==================================
1878 __ restore(FP, G0, SP);
1879 __ retl();
1880 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1882 // OSR request, unwind the current frame and transfer to the OSR entry
1883 // and enter OSR nmethod
1885 __ bind(do_OSR);
1886 Label remove_initial_frame;
1887 __ ld_ptr(STATE(_prev_link), L1_scratch);
1888 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1890 // We are going to pop this frame. Is there another interpreter frame underneath
1891 // it or is it callstub/compiled?
1893 __ tst(L1_scratch);
1894 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1895 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1897 // Frame underneath is an interpreter frame simply unwind
1898 // POP FRAME HERE ==================================
1899 __ restore(FP, G0, SP); // unwind interpreter state frame
1900 __ mov(I5_savedSP->after_restore(), SP);
1902 // Since we are now calling native need to change our "return address" from the
1903 // dummy RecursiveInterpreterActivation to a return from native
1905 __ set((intptr_t)return_from_native_method - 8, O7);
1907 __ jmpl(G3_scratch, G0, G0);
1908 __ delayed()->mov(G1_scratch, O0);
1910 __ bind(remove_initial_frame);
1912 // POP FRAME HERE ==================================
1913 __ restore(FP, G0, SP);
1914 __ mov(I5_savedSP->after_restore(), SP);
1915 __ jmpl(G3_scratch, G0, G0);
1916 __ delayed()->mov(G1_scratch, O0);
1918 // Call a new method. All we do is (temporarily) trim the expression stack
1919 // push a return address to bring us back to here and leap to the new entry.
1920 // At this point we have a topmost frame that was allocated by the frame manager
1921 // which contains the current method interpreted state. We trim this frame
1922 // of excess java expression stack entries and then recurse.
1924 __ bind(call_method);
1926 // stack points to next free location and not top element on expression stack
1927 // method expects sp to be pointing to topmost element
1929 __ ld_ptr(STATE(_thread), G2_thread);
1930 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1933 // SP already takes in to account the 2 extra words we use for slop
1934 // when we call a "static long no_params()" method. So if
1935 // we trim back sp by the amount of unused java expression stack
1936 // there will be automagically the 2 extra words we need.
1937 // We also have to worry about keeping SP aligned.
1939 __ ld_ptr(STATE(_stack), Gargs);
1940 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1942 // compute the unused java stack size
1943 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1945 // Round down the unused space to that stack is always 16-byte aligned
1946 // by making the unused space a multiple of the size of two longs.
1948 __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1950 // Now trim the stack
1951 __ add(SP, L2_scratch, SP);
1954 // Now point to the final argument (account for prepush)
1955 __ add(Gargs, wordSize, Gargs);
1956 #ifdef ASSERT
1957 // Make sure we have space for the window
1958 __ sub(Gargs, SP, L1_scratch);
1959 __ cmp(L1_scratch, 16*wordSize);
1960 {
1961 Label skip;
1962 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1963 __ delayed()->nop();
1964 __ stop("killed stack");
1965 __ bind(skip);
1966 }
1967 #endif // ASSERT
1969 // Create a new frame where we can store values that make it look like the interpreter
1970 // really recursed.
1972 // prepare to recurse or call specialized entry
1974 // First link the registers we need
1976 // make the pc look good in debugger
1977 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
1978 // argument too
1979 __ mov(Lstate, I0);
1981 // Record our sending SP
1982 __ mov(SP, O5_savedSP);
1984 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
1985 __ set((intptr_t) entry_point, L1_scratch);
1986 __ cmp(L1_scratch, L2_scratch);
1987 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
1988 __ delayed()->mov(Lstate, prevState); // link activations
1990 // method uses specialized entry, push a return so we look like call stub setup
1991 // this path will handle fact that result is returned in registers and not
1992 // on the java stack.
1994 __ set((intptr_t)return_from_native_method - 8, O7);
1995 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
1996 __ delayed()->nop();
1998 //
1999 // Bad Message from interpreter
2000 //
2001 __ bind(bad_msg);
2002 __ stop("Bad message from interpreter");
2004 // Interpreted method "returned" with an exception pass it on...
2005 // Pass result, unwind activation and continue/return to interpreter/call_stub
2006 // We handle result (if any) differently based on return to interpreter or call_stub
2008 __ bind(throw_exception);
2009 __ ld_ptr(STATE(_prev_link), L1_scratch);
2010 __ tst(L1_scratch);
2011 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
2012 __ delayed()->nop();
2014 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
2015 __ ba(false, unwind_recursive_activation);
2016 __ delayed()->nop();
2018 interpreter_frame_manager = entry_point;
2019 return entry_point;
2020 }
2022 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2023 : CppInterpreterGenerator(code) {
2024 generate_all(); // down here so it can be "virtual"
2025 }
2028 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2030 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2031 // expression stack, the callee will have callee_extra_locals (so we can account for
2032 // frame extension) and monitor_size for monitors. Basically we need to calculate
2033 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2034 //
2035 //
2036 // The big complicating thing here is that we must ensure that the stack stays properly
2037 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2038 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2039 // the caller so we must ensure that it is properly aligned for our callee.
2040 //
2041 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2042 // stack at all times to deal with the "stack long no_params()" method issue. This
2043 // is "slop_factor" here.
2044 const int slop_factor = 2;
2046 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2047 frame::memory_parameter_word_sp_offset; // register save area + param window
2048 const int extra_stack = 0; //6815692//methodOopDesc::extra_stack_entries();
2049 return (round_to(max_stack +
2050 extra_stack +
2051 slop_factor +
2052 fixed_size +
2053 monitor_size +
2054 (callee_extra_locals * Interpreter::stackElementWords()), WordsPerLong));
2056 }
2058 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
2060 // See call_stub code
2061 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2062 WordsPerLong); // 7 + register save area
2064 // Save space for one monitor to get into the interpreted method in case
2065 // the method is synchronized
2066 int monitor_size = method->is_synchronized() ?
2067 1*frame::interpreter_frame_monitor_size() : 0;
2068 return size_activation_helper(method->max_locals(), method->max_stack(),
2069 monitor_size) + call_stub_size;
2070 }
2072 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2073 frame* caller,
2074 frame* current,
2075 methodOop method,
2076 intptr_t* locals,
2077 intptr_t* stack,
2078 intptr_t* stack_base,
2079 intptr_t* monitor_base,
2080 intptr_t* frame_bottom,
2081 bool is_top_frame
2082 )
2083 {
2084 // What about any vtable?
2085 //
2086 to_fill->_thread = JavaThread::current();
2087 // This gets filled in later but make it something recognizable for now
2088 to_fill->_bcp = method->code_base();
2089 to_fill->_locals = locals;
2090 to_fill->_constants = method->constants()->cache();
2091 to_fill->_method = method;
2092 to_fill->_mdx = NULL;
2093 to_fill->_stack = stack;
2094 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2095 to_fill->_msg = deopt_resume2;
2096 } else {
2097 to_fill->_msg = method_resume;
2098 }
2099 to_fill->_result._to_call._bcp_advance = 0;
2100 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2101 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2102 to_fill->_prev_link = NULL;
2104 // Fill in the registers for the frame
2106 // Need to install _sender_sp. Actually not too hard in C++!
2107 // When the skeletal frames are layed out we fill in a value
2108 // for _sender_sp. That value is only correct for the oldest
2109 // skeletal frame constructed (because there is only a single
2110 // entry for "caller_adjustment". While the skeletal frames
2111 // exist that is good enough. We correct that calculation
2112 // here and get all the frames correct.
2114 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2116 *current->register_addr(Lstate) = (intptr_t) to_fill;
2117 // skeletal already places a useful value here and this doesn't account
2118 // for alignment so don't bother.
2119 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2121 if (caller->is_interpreted_frame()) {
2122 interpreterState prev = caller->get_interpreterState();
2123 to_fill->_prev_link = prev;
2124 // Make the prev callee look proper
2125 prev->_result._to_call._callee = method;
2126 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2127 prev->_result._to_call._bcp_advance = 5;
2128 } else {
2129 prev->_result._to_call._bcp_advance = 3;
2130 }
2131 }
2132 to_fill->_oop_temp = NULL;
2133 to_fill->_stack_base = stack_base;
2134 // Need +1 here because stack_base points to the word just above the first expr stack entry
2135 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2136 // See generate_compute_interpreter_state.
2137 int extra_stack = 0; //6815692//methodOopDesc::extra_stack_entries();
2138 to_fill->_stack_limit = stack_base - (method->max_stack() + 1 + extra_stack);
2139 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2141 // sparc specific
2142 to_fill->_frame_bottom = frame_bottom;
2143 to_fill->_self_link = to_fill;
2144 #ifdef ASSERT
2145 to_fill->_native_fresult = 123456.789;
2146 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2147 #endif
2148 }
2150 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2151 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2152 }
2155 int AbstractInterpreter::layout_activation(methodOop method,
2156 int tempcount, // Number of slots on java expression stack in use
2157 int popframe_extra_args,
2158 int moncount, // Number of active monitors
2159 int callee_param_size,
2160 int callee_locals_size,
2161 frame* caller,
2162 frame* interpreter_frame,
2163 bool is_top_frame) {
2165 assert(popframe_extra_args == 0, "NEED TO FIX");
2166 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2167 // does as far as allocating an interpreter frame.
2168 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2169 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2170 // as determined by a previous call to this method.
2171 // It is also guaranteed to be walkable even though it is in a skeletal state
2172 // NOTE: return size is in words not bytes
2173 // NOTE: tempcount is the current size of the java expression stack. For top most
2174 // frames we will allocate a full sized expression stack and not the curback
2175 // version that non-top frames have.
2177 // Calculate the amount our frame will be adjust by the callee. For top frame
2178 // this is zero.
2180 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2181 // calculates the extra locals based on itself. Not what the callee does
2182 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2183 // as getting sender_sp correct.
2185 int extra_locals_size = callee_locals_size - callee_param_size;
2186 int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2187 int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2188 int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2189 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2192 /*
2193 if we actually have a frame to layout we must now fill in all the pieces. This means both
2194 the interpreterState and the registers.
2195 */
2196 if (interpreter_frame != NULL) {
2198 // MUCHO HACK
2200 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2201 // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode.
2202 assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2203 frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2205 /* Now fillin the interpreterState object */
2207 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2210 intptr_t* locals;
2212 // Calculate the postion of locals[0]. This is painful because of
2213 // stack alignment (same as ia64). The problem is that we can
2214 // not compute the location of locals from fp(). fp() will account
2215 // for the extra locals but it also accounts for aligning the stack
2216 // and we can't determine if the locals[0] was misaligned but max_locals
2217 // was enough to have the
2218 // calculate postion of locals. fp already accounts for extra locals.
2219 // +2 for the static long no_params() issue.
2221 if (caller->is_interpreted_frame()) {
2222 // locals must agree with the caller because it will be used to set the
2223 // caller's tos when we return.
2224 interpreterState prev = caller->get_interpreterState();
2225 // stack() is prepushed.
2226 locals = prev->stack() + method->size_of_parameters();
2227 } else {
2228 // Lay out locals block in the caller adjacent to the register window save area.
2229 //
2230 // Compiled frames do not allocate a varargs area which is why this if
2231 // statement is needed.
2232 //
2233 intptr_t* fp = interpreter_frame->fp();
2234 int local_words = method->max_locals() * Interpreter::stackElementWords();
2236 if (caller->is_compiled_frame()) {
2237 locals = fp + frame::register_save_words + local_words - 1;
2238 } else {
2239 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2240 }
2242 }
2243 // END MUCHO HACK
2245 intptr_t* monitor_base = (intptr_t*) cur_state;
2246 intptr_t* stack_base = monitor_base - monitor_size;
2247 /* +1 because stack is always prepushed */
2248 intptr_t* stack = stack_base - (tempcount + 1);
2251 BytecodeInterpreter::layout_interpreterState(cur_state,
2252 caller,
2253 interpreter_frame,
2254 method,
2255 locals,
2256 stack,
2257 stack_base,
2258 monitor_base,
2259 frame_bottom,
2260 is_top_frame);
2262 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2264 }
2265 return frame_words;
2266 }
2268 #endif // CC_INTERP