Tue, 10 Mar 2009 08:52:16 -0700
Merge
1 /*
2 * Copyright 2007-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_cppInterpreter_x86.cpp.incl"
28 #ifdef CC_INTERP
30 // Routine exists to make tracebacks look decent in debugger
31 // while we are recursed in the frame manager/c++ interpreter.
32 // We could use an address in the frame manager but having
33 // frames look natural in the debugger is a plus.
34 extern "C" void RecursiveInterpreterActivation(interpreterState istate )
35 {
36 //
37 ShouldNotReachHere();
38 }
41 #define __ _masm->
42 #define STATE(field_name) (Address(state, byte_offset_of(BytecodeInterpreter, field_name)))
44 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized
45 // c++ interpreter entry point this holds that entry point label.
47 // default registers for state and sender_sp
48 // state and sender_sp are the same on 32bit because we have no choice.
49 // state could be rsi on 64bit but it is an arg reg and not callee save
50 // so r13 is better choice.
52 const Register state = NOT_LP64(rsi) LP64_ONLY(r13);
53 const Register sender_sp_on_entry = NOT_LP64(rsi) LP64_ONLY(r13);
55 // NEEDED for JVMTI?
56 // address AbstractInterpreter::_remove_activation_preserving_args_entry;
58 static address unctrap_frame_manager_entry = NULL;
60 static address deopt_frame_manager_return_atos = NULL;
61 static address deopt_frame_manager_return_btos = NULL;
62 static address deopt_frame_manager_return_itos = NULL;
63 static address deopt_frame_manager_return_ltos = NULL;
64 static address deopt_frame_manager_return_ftos = NULL;
65 static address deopt_frame_manager_return_dtos = NULL;
66 static address deopt_frame_manager_return_vtos = NULL;
68 int AbstractInterpreter::BasicType_as_index(BasicType type) {
69 int i = 0;
70 switch (type) {
71 case T_BOOLEAN: i = 0; break;
72 case T_CHAR : i = 1; break;
73 case T_BYTE : i = 2; break;
74 case T_SHORT : i = 3; break;
75 case T_INT : i = 4; break;
76 case T_VOID : i = 5; break;
77 case T_FLOAT : i = 8; break;
78 case T_LONG : i = 9; break;
79 case T_DOUBLE : i = 6; break;
80 case T_OBJECT : // fall through
81 case T_ARRAY : i = 7; break;
82 default : ShouldNotReachHere();
83 }
84 assert(0 <= i && i < AbstractInterpreter::number_of_result_handlers, "index out of bounds");
85 return i;
86 }
88 // Is this pc anywhere within code owned by the interpreter?
89 // This only works for pc that might possibly be exposed to frame
90 // walkers. It clearly misses all of the actual c++ interpreter
91 // implementation
92 bool CppInterpreter::contains(address pc) {
93 return (_code->contains(pc) ||
94 pc == CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation));
95 }
98 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
99 address entry = __ pc();
100 switch (type) {
101 case T_BOOLEAN: __ c2bool(rax); break;
102 case T_CHAR : __ andl(rax, 0xFFFF); break;
103 case T_BYTE : __ sign_extend_byte (rax); break;
104 case T_SHORT : __ sign_extend_short(rax); break;
105 case T_VOID : // fall thru
106 case T_LONG : // fall thru
107 case T_INT : /* nothing to do */ break;
109 case T_DOUBLE :
110 case T_FLOAT :
111 {
112 const Register t = InterpreterRuntime::SignatureHandlerGenerator::temp();
113 __ pop(t); // remove return address first
114 // Must return a result for interpreter or compiler. In SSE
115 // mode, results are returned in xmm0 and the FPU stack must
116 // be empty.
117 if (type == T_FLOAT && UseSSE >= 1) {
118 #ifndef _LP64
119 // Load ST0
120 __ fld_d(Address(rsp, 0));
121 // Store as float and empty fpu stack
122 __ fstp_s(Address(rsp, 0));
123 #endif // !_LP64
124 // and reload
125 __ movflt(xmm0, Address(rsp, 0));
126 } else if (type == T_DOUBLE && UseSSE >= 2 ) {
127 __ movdbl(xmm0, Address(rsp, 0));
128 } else {
129 // restore ST0
130 __ fld_d(Address(rsp, 0));
131 }
132 // and pop the temp
133 __ addptr(rsp, 2 * wordSize);
134 __ push(t); // restore return address
135 }
136 break;
137 case T_OBJECT :
138 // retrieve result from frame
139 __ movptr(rax, STATE(_oop_temp));
140 // and verify it
141 __ verify_oop(rax);
142 break;
143 default : ShouldNotReachHere();
144 }
145 __ ret(0); // return from result handler
146 return entry;
147 }
149 // tosca based result to c++ interpreter stack based result.
150 // Result goes to top of native stack.
152 #undef EXTEND // SHOULD NOT BE NEEDED
153 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
154 // A result is in the tosca (abi result) from either a native method call or compiled
155 // code. Place this result on the java expression stack so C++ interpreter can use it.
156 address entry = __ pc();
158 const Register t = InterpreterRuntime::SignatureHandlerGenerator::temp();
159 __ pop(t); // remove return address first
160 switch (type) {
161 case T_VOID:
162 break;
163 case T_BOOLEAN:
164 #ifdef EXTEND
165 __ c2bool(rax);
166 #endif
167 __ push(rax);
168 break;
169 case T_CHAR :
170 #ifdef EXTEND
171 __ andl(rax, 0xFFFF);
172 #endif
173 __ push(rax);
174 break;
175 case T_BYTE :
176 #ifdef EXTEND
177 __ sign_extend_byte (rax);
178 #endif
179 __ push(rax);
180 break;
181 case T_SHORT :
182 #ifdef EXTEND
183 __ sign_extend_short(rax);
184 #endif
185 __ push(rax);
186 break;
187 case T_LONG :
188 __ push(rdx); // pushes useless junk on 64bit
189 __ push(rax);
190 break;
191 case T_INT :
192 __ push(rax);
193 break;
194 case T_FLOAT :
195 // Result is in ST(0)/xmm0
196 __ subptr(rsp, wordSize);
197 if ( UseSSE < 1) {
198 __ fstp_s(Address(rsp, 0));
199 } else {
200 __ movflt(Address(rsp, 0), xmm0);
201 }
202 break;
203 case T_DOUBLE :
204 __ subptr(rsp, 2*wordSize);
205 if ( UseSSE < 2 ) {
206 __ fstp_d(Address(rsp, 0));
207 } else {
208 __ movdbl(Address(rsp, 0), xmm0);
209 }
210 break;
211 case T_OBJECT :
212 __ verify_oop(rax); // verify it
213 __ push(rax);
214 break;
215 default : ShouldNotReachHere();
216 }
217 __ jmp(t); // return from result handler
218 return entry;
219 }
221 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
222 // A result is in the java expression stack of the interpreted method that has just
223 // returned. Place this result on the java expression stack of the caller.
224 //
225 // The current interpreter activation in rsi/r13 is for the method just returning its
226 // result. So we know that the result of this method is on the top of the current
227 // execution stack (which is pre-pushed) and will be return to the top of the caller
228 // stack. The top of the callers stack is the bottom of the locals of the current
229 // activation.
230 // Because of the way activation are managed by the frame manager the value of rsp is
231 // below both the stack top of the current activation and naturally the stack top
232 // of the calling activation. This enable this routine to leave the return address
233 // to the frame manager on the stack and do a vanilla return.
234 //
235 // On entry: rsi/r13 - interpreter state of activation returning a (potential) result
236 // On Return: rsi/r13 - unchanged
237 // rax - new stack top for caller activation (i.e. activation in _prev_link)
238 //
239 // Can destroy rdx, rcx.
240 //
242 address entry = __ pc();
243 const Register t = InterpreterRuntime::SignatureHandlerGenerator::temp();
244 switch (type) {
245 case T_VOID:
246 __ movptr(rax, STATE(_locals)); // pop parameters get new stack value
247 __ addptr(rax, wordSize); // account for prepush before we return
248 break;
249 case T_FLOAT :
250 case T_BOOLEAN:
251 case T_CHAR :
252 case T_BYTE :
253 case T_SHORT :
254 case T_INT :
255 // 1 word result
256 __ movptr(rdx, STATE(_stack));
257 __ movptr(rax, STATE(_locals)); // address for result
258 __ movl(rdx, Address(rdx, wordSize)); // get result
259 __ movptr(Address(rax, 0), rdx); // and store it
260 break;
261 case T_LONG :
262 case T_DOUBLE :
263 // return top two words on current expression stack to caller's expression stack
264 // The caller's expression stack is adjacent to the current frame manager's intepretState
265 // except we allocated one extra word for this intepretState so we won't overwrite it
266 // when we return a two word result.
268 __ movptr(rax, STATE(_locals)); // address for result
269 __ movptr(rcx, STATE(_stack));
270 __ subptr(rax, wordSize); // need addition word besides locals[0]
271 __ movptr(rdx, Address(rcx, 2*wordSize)); // get result word (junk in 64bit)
272 __ movptr(Address(rax, wordSize), rdx); // and store it
273 __ movptr(rdx, Address(rcx, wordSize)); // get result word
274 __ movptr(Address(rax, 0), rdx); // and store it
275 break;
276 case T_OBJECT :
277 __ movptr(rdx, STATE(_stack));
278 __ movptr(rax, STATE(_locals)); // address for result
279 __ movptr(rdx, Address(rdx, wordSize)); // get result
280 __ verify_oop(rdx); // verify it
281 __ movptr(Address(rax, 0), rdx); // and store it
282 break;
283 default : ShouldNotReachHere();
284 }
285 __ ret(0);
286 return entry;
287 }
289 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
290 // A result is in the java expression stack of the interpreted method that has just
291 // returned. Place this result in the native abi that the caller expects.
292 //
293 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
294 // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
295 // and so rather than return result onto caller's java expression stack we return the
296 // result in the expected location based on the native abi.
297 // On entry: rsi/r13 - interpreter state of activation returning a (potential) result
298 // On Return: rsi/r13 - unchanged
299 // Other registers changed [rax/rdx/ST(0) as needed for the result returned]
301 address entry = __ pc();
302 switch (type) {
303 case T_VOID:
304 break;
305 case T_BOOLEAN:
306 case T_CHAR :
307 case T_BYTE :
308 case T_SHORT :
309 case T_INT :
310 __ movptr(rdx, STATE(_stack)); // get top of stack
311 __ movl(rax, Address(rdx, wordSize)); // get result word 1
312 break;
313 case T_LONG :
314 __ movptr(rdx, STATE(_stack)); // get top of stack
315 __ movptr(rax, Address(rdx, wordSize)); // get result low word
316 NOT_LP64(__ movl(rdx, Address(rdx, 2*wordSize));) // get result high word
317 break;
318 case T_FLOAT :
319 __ movptr(rdx, STATE(_stack)); // get top of stack
320 if ( UseSSE >= 1) {
321 __ movflt(xmm0, Address(rdx, wordSize));
322 } else {
323 __ fld_s(Address(rdx, wordSize)); // pushd float result
324 }
325 break;
326 case T_DOUBLE :
327 __ movptr(rdx, STATE(_stack)); // get top of stack
328 if ( UseSSE > 1) {
329 __ movdbl(xmm0, Address(rdx, wordSize));
330 } else {
331 __ fld_d(Address(rdx, wordSize)); // push double result
332 }
333 break;
334 case T_OBJECT :
335 __ movptr(rdx, STATE(_stack)); // get top of stack
336 __ movptr(rax, Address(rdx, wordSize)); // get result word 1
337 __ verify_oop(rax); // verify it
338 break;
339 default : ShouldNotReachHere();
340 }
341 __ ret(0);
342 return entry;
343 }
345 address CppInterpreter::return_entry(TosState state, int length) {
346 // make it look good in the debugger
347 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation);
348 }
350 address CppInterpreter::deopt_entry(TosState state, int length) {
351 address ret = NULL;
352 if (length != 0) {
353 switch (state) {
354 case atos: ret = deopt_frame_manager_return_atos; break;
355 case btos: ret = deopt_frame_manager_return_btos; break;
356 case ctos:
357 case stos:
358 case itos: ret = deopt_frame_manager_return_itos; break;
359 case ltos: ret = deopt_frame_manager_return_ltos; break;
360 case ftos: ret = deopt_frame_manager_return_ftos; break;
361 case dtos: ret = deopt_frame_manager_return_dtos; break;
362 case vtos: ret = deopt_frame_manager_return_vtos; break;
363 }
364 } else {
365 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap)
366 }
367 assert(ret != NULL, "Not initialized");
368 return ret;
369 }
371 // C++ Interpreter
372 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
373 const Register locals,
374 const Register sender_sp,
375 bool native) {
377 // On entry the "locals" argument points to locals[0] (or where it would be in case no locals in
378 // a static method). "state" contains any previous frame manager state which we must save a link
379 // to in the newly generated state object. On return "state" is a pointer to the newly allocated
380 // state object. We must allocate and initialize a new interpretState object and the method
381 // expression stack. Because the returned result (if any) of the method will be placed on the caller's
382 // expression stack and this will overlap with locals[0] (and locals[1] if double/long) we must
383 // be sure to leave space on the caller's stack so that this result will not overwrite values when
384 // locals[0] and locals[1] do not exist (and in fact are return address and saved rbp). So when
385 // we are non-native we in essence ensure that locals[0-1] exist. We play an extra trick in
386 // non-product builds and initialize this last local with the previous interpreterState as
387 // this makes things look real nice in the debugger.
389 // State on entry
390 // Assumes locals == &locals[0]
391 // Assumes state == any previous frame manager state (assuming call path from c++ interpreter)
392 // Assumes rax = return address
393 // rcx == senders_sp
394 // rbx == method
395 // Modifies rcx, rdx, rax
396 // Returns:
397 // state == address of new interpreterState
398 // rsp == bottom of method's expression stack.
400 const Address const_offset (rbx, methodOopDesc::const_offset());
403 // On entry sp is the sender's sp. This includes the space for the arguments
404 // that the sender pushed. If the sender pushed no args (a static) and the
405 // caller returns a long then we need two words on the sender's stack which
406 // are not present (although when we return a restore full size stack the
407 // space will be present). If we didn't allocate two words here then when
408 // we "push" the result of the caller's stack we would overwrite the return
409 // address and the saved rbp. Not good. So simply allocate 2 words now
410 // just to be safe. This is the "static long no_params() method" issue.
411 // See Lo.java for a testcase.
412 // We don't need this for native calls because they return result in
413 // register and the stack is expanded in the caller before we store
414 // the results on the stack.
416 if (!native) {
417 #ifdef PRODUCT
418 __ subptr(rsp, 2*wordSize);
419 #else /* PRODUCT */
420 __ push((int32_t)NULL_WORD);
421 __ push(state); // make it look like a real argument
422 #endif /* PRODUCT */
423 }
425 // Now that we are assure of space for stack result, setup typical linkage
427 __ push(rax);
428 __ enter();
430 __ mov(rax, state); // save current state
432 __ lea(rsp, Address(rsp, -(int)sizeof(BytecodeInterpreter)));
433 __ mov(state, rsp);
435 // rsi/r13 == state/locals rax == prevstate
437 // initialize the "shadow" frame so that use since C++ interpreter not directly
438 // recursive. Simpler to recurse but we can't trim expression stack as we call
439 // new methods.
440 __ movptr(STATE(_locals), locals); // state->_locals = locals()
441 __ movptr(STATE(_self_link), state); // point to self
442 __ movptr(STATE(_prev_link), rax); // state->_link = state on entry (NULL or previous state)
443 __ movptr(STATE(_sender_sp), sender_sp); // state->_sender_sp = sender_sp
444 #ifdef _LP64
445 __ movptr(STATE(_thread), r15_thread); // state->_bcp = codes()
446 #else
447 __ get_thread(rax); // get vm's javathread*
448 __ movptr(STATE(_thread), rax); // state->_bcp = codes()
449 #endif // _LP64
450 __ movptr(rdx, Address(rbx, methodOopDesc::const_offset())); // get constantMethodOop
451 __ lea(rdx, Address(rdx, constMethodOopDesc::codes_offset())); // get code base
452 if (native) {
453 __ movptr(STATE(_bcp), (int32_t)NULL_WORD); // state->_bcp = NULL
454 } else {
455 __ movptr(STATE(_bcp), rdx); // state->_bcp = codes()
456 }
457 __ xorptr(rdx, rdx);
458 __ movptr(STATE(_oop_temp), rdx); // state->_oop_temp = NULL (only really needed for native)
459 __ movptr(STATE(_mdx), rdx); // state->_mdx = NULL
460 __ movptr(rdx, Address(rbx, methodOopDesc::constants_offset()));
461 __ movptr(rdx, Address(rdx, constantPoolOopDesc::cache_offset_in_bytes()));
462 __ movptr(STATE(_constants), rdx); // state->_constants = constants()
464 __ movptr(STATE(_method), rbx); // state->_method = method()
465 __ movl(STATE(_msg), (int32_t) BytecodeInterpreter::method_entry); // state->_msg = initial method entry
466 __ movptr(STATE(_result._to_call._callee), (int32_t) NULL_WORD); // state->_result._to_call._callee_callee = NULL
469 __ movptr(STATE(_monitor_base), rsp); // set monitor block bottom (grows down) this would point to entry [0]
470 // entries run from -1..x where &monitor[x] ==
472 {
473 // Must not attempt to lock method until we enter interpreter as gc won't be able to find the
474 // initial frame. However we allocate a free monitor so we don't have to shuffle the expression stack
475 // immediately.
477 // synchronize method
478 const Address access_flags (rbx, methodOopDesc::access_flags_offset());
479 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
480 Label not_synced;
482 __ movl(rax, access_flags);
483 __ testl(rax, JVM_ACC_SYNCHRONIZED);
484 __ jcc(Assembler::zero, not_synced);
486 // Allocate initial monitor and pre initialize it
487 // get synchronization object
489 Label done;
490 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
491 __ movl(rax, access_flags);
492 __ testl(rax, JVM_ACC_STATIC);
493 __ movptr(rax, Address(locals, 0)); // get receiver (assume this is frequent case)
494 __ jcc(Assembler::zero, done);
495 __ movptr(rax, Address(rbx, methodOopDesc::constants_offset()));
496 __ movptr(rax, Address(rax, constantPoolOopDesc::pool_holder_offset_in_bytes()));
497 __ movptr(rax, Address(rax, mirror_offset));
498 __ bind(done);
499 // add space for monitor & lock
500 __ subptr(rsp, entry_size); // add space for a monitor entry
501 __ movptr(Address(rsp, BasicObjectLock::obj_offset_in_bytes()), rax); // store object
502 __ bind(not_synced);
503 }
505 __ movptr(STATE(_stack_base), rsp); // set expression stack base ( == &monitors[-count])
506 if (native) {
507 __ movptr(STATE(_stack), rsp); // set current expression stack tos
508 __ movptr(STATE(_stack_limit), rsp);
509 } else {
510 __ subptr(rsp, wordSize); // pre-push stack
511 __ movptr(STATE(_stack), rsp); // set current expression stack tos
513 // compute full expression stack limit
515 const Address size_of_stack (rbx, methodOopDesc::max_stack_offset());
516 __ load_unsigned_short(rdx, size_of_stack); // get size of expression stack in words
517 __ negptr(rdx); // so we can subtract in next step
518 // Allocate expression stack
519 __ lea(rsp, Address(rsp, rdx, Address::times_ptr));
520 __ movptr(STATE(_stack_limit), rsp);
521 }
523 #ifdef _LP64
524 // Make sure stack is properly aligned and sized for the abi
525 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
526 __ andptr(rsp, -16); // must be 16 byte boundary (see amd64 ABI)
527 #endif // _LP64
531 }
533 // Helpers for commoning out cases in the various type of method entries.
534 //
536 // increment invocation count & check for overflow
537 //
538 // Note: checking for negative value instead of overflow
539 // so we have a 'sticky' overflow test
540 //
541 // rbx,: method
542 // rcx: invocation counter
543 //
544 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
546 const Address invocation_counter(rbx, methodOopDesc::invocation_counter_offset() + InvocationCounter::counter_offset());
547 const Address backedge_counter (rbx, methodOopDesc::backedge_counter_offset() + InvocationCounter::counter_offset());
549 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
550 __ incrementl(Address(rbx,methodOopDesc::interpreter_invocation_counter_offset()));
551 }
552 // Update standard invocation counters
553 __ movl(rax, backedge_counter); // load backedge counter
555 __ increment(rcx, InvocationCounter::count_increment);
556 __ andl(rax, InvocationCounter::count_mask_value); // mask out the status bits
558 __ movl(invocation_counter, rcx); // save invocation count
559 __ addl(rcx, rax); // add both counters
561 // profile_method is non-null only for interpreted method so
562 // profile_method != NULL == !native_call
563 // BytecodeInterpreter only calls for native so code is elided.
565 __ cmp32(rcx,
566 ExternalAddress((address)&InvocationCounter::InterpreterInvocationLimit));
567 __ jcc(Assembler::aboveEqual, *overflow);
569 }
571 void InterpreterGenerator::generate_counter_overflow(Label* do_continue) {
573 // C++ interpreter on entry
574 // rsi/r13 - new interpreter state pointer
575 // rbp - interpreter frame pointer
576 // rbx - method
578 // On return (i.e. jump to entry_point) [ back to invocation of interpreter ]
579 // rbx, - method
580 // rcx - rcvr (assuming there is one)
581 // top of stack return address of interpreter caller
582 // rsp - sender_sp
584 // C++ interpreter only
585 // rsi/r13 - previous interpreter state pointer
587 const Address size_of_parameters(rbx, methodOopDesc::size_of_parameters_offset());
589 // InterpreterRuntime::frequency_counter_overflow takes one argument
590 // indicating if the counter overflow occurs at a backwards branch (non-NULL bcp).
591 // The call returns the address of the verified entry point for the method or NULL
592 // if the compilation did not complete (either went background or bailed out).
593 __ movptr(rax, (int32_t)false);
594 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), rax);
596 // for c++ interpreter can rsi really be munged?
597 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter))); // restore state
598 __ movptr(rbx, Address(state, byte_offset_of(BytecodeInterpreter, _method))); // restore method
599 __ movptr(rdi, Address(state, byte_offset_of(BytecodeInterpreter, _locals))); // get locals pointer
601 __ jmp(*do_continue, relocInfo::none);
603 }
605 void InterpreterGenerator::generate_stack_overflow_check(void) {
606 // see if we've got enough room on the stack for locals plus overhead.
607 // the expression stack grows down incrementally, so the normal guard
608 // page mechanism will work for that.
609 //
610 // Registers live on entry:
611 //
612 // Asm interpreter
613 // rdx: number of additional locals this frame needs (what we must check)
614 // rbx,: methodOop
616 // C++ Interpreter
617 // rsi/r13: previous interpreter frame state object
618 // rdi: &locals[0]
619 // rcx: # of locals
620 // rdx: number of additional locals this frame needs (what we must check)
621 // rbx: methodOop
623 // destroyed on exit
624 // rax,
626 // NOTE: since the additional locals are also always pushed (wasn't obvious in
627 // generate_method_entry) so the guard should work for them too.
628 //
630 // monitor entry size: see picture of stack set (generate_method_entry) and frame_i486.hpp
631 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
633 // total overhead size: entry_size + (saved rbp, thru expr stack bottom).
634 // be sure to change this if you add/subtract anything to/from the overhead area
635 const int overhead_size = (int)sizeof(BytecodeInterpreter);
637 const int page_size = os::vm_page_size();
639 Label after_frame_check;
641 // compute rsp as if this were going to be the last frame on
642 // the stack before the red zone
644 Label after_frame_check_pop;
646 // save rsi == caller's bytecode ptr (c++ previous interp. state)
647 // QQQ problem here?? rsi overload????
648 __ push(state);
650 const Register thread = LP64_ONLY(r15_thread) NOT_LP64(rsi);
652 NOT_LP64(__ get_thread(thread));
654 const Address stack_base(thread, Thread::stack_base_offset());
655 const Address stack_size(thread, Thread::stack_size_offset());
657 // locals + overhead, in bytes
658 const Address size_of_stack (rbx, methodOopDesc::max_stack_offset());
659 // Always give one monitor to allow us to start interp if sync method.
660 // Any additional monitors need a check when moving the expression stack
661 const int one_monitor = frame::interpreter_frame_monitor_size() * wordSize;
662 __ load_unsigned_short(rax, size_of_stack); // get size of expression stack in words
663 __ lea(rax, Address(noreg, rax, Interpreter::stackElementScale(), one_monitor));
664 __ lea(rax, Address(rax, rdx, Interpreter::stackElementScale(), overhead_size));
666 #ifdef ASSERT
667 Label stack_base_okay, stack_size_okay;
668 // verify that thread stack base is non-zero
669 __ cmpptr(stack_base, (int32_t)0);
670 __ jcc(Assembler::notEqual, stack_base_okay);
671 __ stop("stack base is zero");
672 __ bind(stack_base_okay);
673 // verify that thread stack size is non-zero
674 __ cmpptr(stack_size, (int32_t)0);
675 __ jcc(Assembler::notEqual, stack_size_okay);
676 __ stop("stack size is zero");
677 __ bind(stack_size_okay);
678 #endif
680 // Add stack base to locals and subtract stack size
681 __ addptr(rax, stack_base);
682 __ subptr(rax, stack_size);
684 // We should have a magic number here for the size of the c++ interpreter frame.
685 // We can't actually tell this ahead of time. The debug version size is around 3k
686 // product is 1k and fastdebug is 4k
687 const int slop = 6 * K;
689 // Use the maximum number of pages we might bang.
690 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
691 (StackRedPages+StackYellowPages);
692 // Only need this if we are stack banging which is temporary while
693 // we're debugging.
694 __ addptr(rax, slop + 2*max_pages * page_size);
696 // check against the current stack bottom
697 __ cmpptr(rsp, rax);
698 __ jcc(Assembler::above, after_frame_check_pop);
700 __ pop(state); // get c++ prev state.
702 // throw exception return address becomes throwing pc
703 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
705 // all done with frame size check
706 __ bind(after_frame_check_pop);
707 __ pop(state);
709 __ bind(after_frame_check);
710 }
712 // Find preallocated monitor and lock method (C++ interpreter)
713 // rbx - methodOop
714 //
715 void InterpreterGenerator::lock_method(void) {
716 // assumes state == rsi/r13 == pointer to current interpreterState
717 // minimally destroys rax, rdx|c_rarg1, rdi
718 //
719 // synchronize method
720 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
721 const Address access_flags (rbx, methodOopDesc::access_flags_offset());
723 const Register monitor = NOT_LP64(rdx) LP64_ONLY(c_rarg1);
725 // find initial monitor i.e. monitors[-1]
726 __ movptr(monitor, STATE(_monitor_base)); // get monitor bottom limit
727 __ subptr(monitor, entry_size); // point to initial monitor
729 #ifdef ASSERT
730 { Label L;
731 __ movl(rax, access_flags);
732 __ testl(rax, JVM_ACC_SYNCHRONIZED);
733 __ jcc(Assembler::notZero, L);
734 __ stop("method doesn't need synchronization");
735 __ bind(L);
736 }
737 #endif // ASSERT
738 // get synchronization object
739 { Label done;
740 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
741 __ movl(rax, access_flags);
742 __ movptr(rdi, STATE(_locals)); // prepare to get receiver (assume common case)
743 __ testl(rax, JVM_ACC_STATIC);
744 __ movptr(rax, Address(rdi, 0)); // get receiver (assume this is frequent case)
745 __ jcc(Assembler::zero, done);
746 __ movptr(rax, Address(rbx, methodOopDesc::constants_offset()));
747 __ movptr(rax, Address(rax, constantPoolOopDesc::pool_holder_offset_in_bytes()));
748 __ movptr(rax, Address(rax, mirror_offset));
749 __ bind(done);
750 }
751 #ifdef ASSERT
752 { Label L;
753 __ cmpptr(rax, Address(monitor, BasicObjectLock::obj_offset_in_bytes())); // correct object?
754 __ jcc(Assembler::equal, L);
755 __ stop("wrong synchronization lobject");
756 __ bind(L);
757 }
758 #endif // ASSERT
759 // can destroy rax, rdx|c_rarg1, rcx, and (via call_VM) rdi!
760 __ lock_object(monitor);
761 }
763 // Call an accessor method (assuming it is resolved, otherwise drop into vanilla (slow path) entry
765 address InterpreterGenerator::generate_accessor_entry(void) {
767 // rbx: methodOop
769 // rsi/r13: senderSP must preserved for slow path, set SP to it on fast path
771 Label xreturn_path;
773 // do fastpath for resolved accessor methods
774 if (UseFastAccessorMethods) {
776 address entry_point = __ pc();
778 Label slow_path;
779 // If we need a safepoint check, generate full interpreter entry.
780 ExternalAddress state(SafepointSynchronize::address_of_state());
781 __ cmp32(ExternalAddress(SafepointSynchronize::address_of_state()),
782 SafepointSynchronize::_not_synchronized);
784 __ jcc(Assembler::notEqual, slow_path);
785 // ASM/C++ Interpreter
786 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof; parameter size = 1
787 // Note: We can only use this code if the getfield has been resolved
788 // and if we don't have a null-pointer exception => check for
789 // these conditions first and use slow path if necessary.
790 // rbx,: method
791 // rcx: receiver
792 __ movptr(rax, Address(rsp, wordSize));
794 // check if local 0 != NULL and read field
795 __ testptr(rax, rax);
796 __ jcc(Assembler::zero, slow_path);
798 __ movptr(rdi, Address(rbx, methodOopDesc::constants_offset()));
799 // read first instruction word and extract bytecode @ 1 and index @ 2
800 __ movptr(rdx, Address(rbx, methodOopDesc::const_offset()));
801 __ movl(rdx, Address(rdx, constMethodOopDesc::codes_offset()));
802 // Shift codes right to get the index on the right.
803 // The bytecode fetched looks like <index><0xb4><0x2a>
804 __ shrl(rdx, 2*BitsPerByte);
805 __ shll(rdx, exact_log2(in_words(ConstantPoolCacheEntry::size())));
806 __ movptr(rdi, Address(rdi, constantPoolOopDesc::cache_offset_in_bytes()));
808 // rax,: local 0
809 // rbx,: method
810 // rcx: receiver - do not destroy since it is needed for slow path!
811 // rcx: scratch
812 // rdx: constant pool cache index
813 // rdi: constant pool cache
814 // rsi/r13: sender sp
816 // check if getfield has been resolved and read constant pool cache entry
817 // check the validity of the cache entry by testing whether _indices field
818 // contains Bytecode::_getfield in b1 byte.
819 assert(in_words(ConstantPoolCacheEntry::size()) == 4, "adjust shift below");
820 __ movl(rcx,
821 Address(rdi,
822 rdx,
823 Address::times_ptr, constantPoolCacheOopDesc::base_offset() + ConstantPoolCacheEntry::indices_offset()));
824 __ shrl(rcx, 2*BitsPerByte);
825 __ andl(rcx, 0xFF);
826 __ cmpl(rcx, Bytecodes::_getfield);
827 __ jcc(Assembler::notEqual, slow_path);
829 // Note: constant pool entry is not valid before bytecode is resolved
830 __ movptr(rcx,
831 Address(rdi,
832 rdx,
833 Address::times_ptr, constantPoolCacheOopDesc::base_offset() + ConstantPoolCacheEntry::f2_offset()));
834 __ movl(rdx,
835 Address(rdi,
836 rdx,
837 Address::times_ptr, constantPoolCacheOopDesc::base_offset() + ConstantPoolCacheEntry::flags_offset()));
839 Label notByte, notShort, notChar;
840 const Address field_address (rax, rcx, Address::times_1);
842 // Need to differentiate between igetfield, agetfield, bgetfield etc.
843 // because they are different sizes.
844 // Use the type from the constant pool cache
845 __ shrl(rdx, ConstantPoolCacheEntry::tosBits);
846 // Make sure we don't need to mask rdx for tosBits after the above shift
847 ConstantPoolCacheEntry::verify_tosBits();
848 #ifdef _LP64
849 Label notObj;
850 __ cmpl(rdx, atos);
851 __ jcc(Assembler::notEqual, notObj);
852 // atos
853 __ movptr(rax, field_address);
854 __ jmp(xreturn_path);
856 __ bind(notObj);
857 #endif // _LP64
858 __ cmpl(rdx, btos);
859 __ jcc(Assembler::notEqual, notByte);
860 __ load_signed_byte(rax, field_address);
861 __ jmp(xreturn_path);
863 __ bind(notByte);
864 __ cmpl(rdx, stos);
865 __ jcc(Assembler::notEqual, notShort);
866 __ load_signed_short(rax, field_address);
867 __ jmp(xreturn_path);
869 __ bind(notShort);
870 __ cmpl(rdx, ctos);
871 __ jcc(Assembler::notEqual, notChar);
872 __ load_unsigned_short(rax, field_address);
873 __ jmp(xreturn_path);
875 __ bind(notChar);
876 #ifdef ASSERT
877 Label okay;
878 #ifndef _LP64
879 __ cmpl(rdx, atos);
880 __ jcc(Assembler::equal, okay);
881 #endif // _LP64
882 __ cmpl(rdx, itos);
883 __ jcc(Assembler::equal, okay);
884 __ stop("what type is this?");
885 __ bind(okay);
886 #endif // ASSERT
887 // All the rest are a 32 bit wordsize
888 __ movl(rax, field_address);
890 __ bind(xreturn_path);
892 // _ireturn/_areturn
893 __ pop(rdi); // get return address
894 __ mov(rsp, sender_sp_on_entry); // set sp to sender sp
895 __ jmp(rdi);
897 // generate a vanilla interpreter entry as the slow path
898 __ bind(slow_path);
899 // We will enter c++ interpreter looking like it was
900 // called by the call_stub this will cause it to return
901 // a tosca result to the invoker which might have been
902 // the c++ interpreter itself.
904 __ jmp(fast_accessor_slow_entry_path);
905 return entry_point;
907 } else {
908 return NULL;
909 }
911 }
913 //
914 // C++ Interpreter stub for calling a native method.
915 // This sets up a somewhat different looking stack for calling the native method
916 // than the typical interpreter frame setup but still has the pointer to
917 // an interpreter state.
918 //
920 address InterpreterGenerator::generate_native_entry(bool synchronized) {
921 // determine code generation flags
922 bool inc_counter = UseCompiler || CountCompiledCalls;
924 // rbx: methodOop
925 // rcx: receiver (unused)
926 // rsi/r13: previous interpreter state (if called from C++ interpreter) must preserve
927 // in any case. If called via c1/c2/call_stub rsi/r13 is junk (to use) but harmless
928 // to save/restore.
929 address entry_point = __ pc();
931 const Address size_of_parameters(rbx, methodOopDesc::size_of_parameters_offset());
932 const Address size_of_locals (rbx, methodOopDesc::size_of_locals_offset());
933 const Address invocation_counter(rbx, methodOopDesc::invocation_counter_offset() + InvocationCounter::counter_offset());
934 const Address access_flags (rbx, methodOopDesc::access_flags_offset());
936 // rsi/r13 == state/locals rdi == prevstate
937 const Register locals = rdi;
939 // get parameter size (always needed)
940 __ load_unsigned_short(rcx, size_of_parameters);
942 // rbx: methodOop
943 // rcx: size of parameters
944 __ pop(rax); // get return address
945 // for natives the size of locals is zero
947 // compute beginning of parameters /locals
948 __ lea(locals, Address(rsp, rcx, Address::times_ptr, -wordSize));
950 // initialize fixed part of activation frame
952 // Assumes rax = return address
954 // allocate and initialize new interpreterState and method expression stack
955 // IN(locals) -> locals
956 // IN(state) -> previous frame manager state (NULL from stub/c1/c2)
957 // destroys rax, rcx, rdx
958 // OUT (state) -> new interpreterState
959 // OUT(rsp) -> bottom of methods expression stack
961 // save sender_sp
962 __ mov(rcx, sender_sp_on_entry);
963 // start with NULL previous state
964 __ movptr(state, (int32_t)NULL_WORD);
965 generate_compute_interpreter_state(state, locals, rcx, true);
967 #ifdef ASSERT
968 { Label L;
969 __ movptr(rax, STATE(_stack_base));
970 #ifdef _LP64
971 // duplicate the alignment rsp got after setting stack_base
972 __ subptr(rax, frame::arg_reg_save_area_bytes); // windows
973 __ andptr(rax, -16); // must be 16 byte boundary (see amd64 ABI)
974 #endif // _LP64
975 __ cmpptr(rax, rsp);
976 __ jcc(Assembler::equal, L);
977 __ stop("broken stack frame setup in interpreter");
978 __ bind(L);
979 }
980 #endif
982 if (inc_counter) __ movl(rcx, invocation_counter); // (pre-)fetch invocation count
984 const Register unlock_thread = LP64_ONLY(r15_thread) NOT_LP64(rax);
985 NOT_LP64(__ movptr(unlock_thread, STATE(_thread));) // get thread
986 // Since at this point in the method invocation the exception handler
987 // would try to exit the monitor of synchronized methods which hasn't
988 // been entered yet, we set the thread local variable
989 // _do_not_unlock_if_synchronized to true. The remove_activation will
990 // check this flag.
992 const Address do_not_unlock_if_synchronized(unlock_thread,
993 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
994 __ movbool(do_not_unlock_if_synchronized, true);
996 // make sure method is native & not abstract
997 #ifdef ASSERT
998 __ movl(rax, access_flags);
999 {
1000 Label L;
1001 __ testl(rax, JVM_ACC_NATIVE);
1002 __ jcc(Assembler::notZero, L);
1003 __ stop("tried to execute non-native method as native");
1004 __ bind(L);
1005 }
1006 { Label L;
1007 __ testl(rax, JVM_ACC_ABSTRACT);
1008 __ jcc(Assembler::zero, L);
1009 __ stop("tried to execute abstract method in interpreter");
1010 __ bind(L);
1011 }
1012 #endif
1015 // increment invocation count & check for overflow
1016 Label invocation_counter_overflow;
1017 if (inc_counter) {
1018 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1019 }
1021 Label continue_after_compile;
1023 __ bind(continue_after_compile);
1025 bang_stack_shadow_pages(true);
1027 // reset the _do_not_unlock_if_synchronized flag
1028 NOT_LP64(__ movl(rax, STATE(_thread));) // get thread
1029 __ movbool(do_not_unlock_if_synchronized, false);
1032 // check for synchronized native methods
1033 //
1034 // Note: This must happen *after* invocation counter check, since
1035 // when overflow happens, the method should not be locked.
1036 if (synchronized) {
1037 // potentially kills rax, rcx, rdx, rdi
1038 lock_method();
1039 } else {
1040 // no synchronization necessary
1041 #ifdef ASSERT
1042 { Label L;
1043 __ movl(rax, access_flags);
1044 __ testl(rax, JVM_ACC_SYNCHRONIZED);
1045 __ jcc(Assembler::zero, L);
1046 __ stop("method needs synchronization");
1047 __ bind(L);
1048 }
1049 #endif
1050 }
1052 // start execution
1054 // jvmti support
1055 __ notify_method_entry();
1057 // work registers
1058 const Register method = rbx;
1059 const Register thread = LP64_ONLY(r15_thread) NOT_LP64(rdi);
1060 const Register t = InterpreterRuntime::SignatureHandlerGenerator::temp(); // rcx|rscratch1
1062 // allocate space for parameters
1063 __ movptr(method, STATE(_method));
1064 __ verify_oop(method);
1065 __ load_unsigned_short(t, Address(method, methodOopDesc::size_of_parameters_offset()));
1066 __ shll(t, 2);
1067 #ifdef _LP64
1068 __ subptr(rsp, t);
1069 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
1070 __ andptr(rsp, -16); // must be 16 byte boundary (see amd64 ABI)
1071 #else
1072 __ addptr(t, 2*wordSize); // allocate two more slots for JNIEnv and possible mirror
1073 __ subptr(rsp, t);
1074 __ andptr(rsp, -(StackAlignmentInBytes)); // gcc needs 16 byte aligned stacks to do XMM intrinsics
1075 #endif // _LP64
1077 // get signature handler
1078 Label pending_exception_present;
1080 { Label L;
1081 __ movptr(t, Address(method, methodOopDesc::signature_handler_offset()));
1082 __ testptr(t, t);
1083 __ jcc(Assembler::notZero, L);
1084 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), method, false);
1085 __ movptr(method, STATE(_method));
1086 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
1087 __ jcc(Assembler::notEqual, pending_exception_present);
1088 __ verify_oop(method);
1089 __ movptr(t, Address(method, methodOopDesc::signature_handler_offset()));
1090 __ bind(L);
1091 }
1092 #ifdef ASSERT
1093 {
1094 Label L;
1095 __ push(t);
1096 __ get_thread(t); // get vm's javathread*
1097 __ cmpptr(t, STATE(_thread));
1098 __ jcc(Assembler::equal, L);
1099 __ int3();
1100 __ bind(L);
1101 __ pop(t);
1102 }
1103 #endif //
1105 const Register from_ptr = InterpreterRuntime::SignatureHandlerGenerator::from();
1106 // call signature handler
1107 assert(InterpreterRuntime::SignatureHandlerGenerator::to () == rsp, "adjust this code");
1109 // The generated handlers do not touch RBX (the method oop).
1110 // However, large signatures cannot be cached and are generated
1111 // each time here. The slow-path generator will blow RBX
1112 // sometime, so we must reload it after the call.
1113 __ movptr(from_ptr, STATE(_locals)); // get the from pointer
1114 __ call(t);
1115 __ movptr(method, STATE(_method));
1116 __ verify_oop(method);
1118 // result handler is in rax
1119 // set result handler
1120 __ movptr(STATE(_result_handler), rax);
1123 // get native function entry point
1124 { Label L;
1125 __ movptr(rax, Address(method, methodOopDesc::native_function_offset()));
1126 __ testptr(rax, rax);
1127 __ jcc(Assembler::notZero, L);
1128 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), method);
1129 __ movptr(method, STATE(_method));
1130 __ verify_oop(method);
1131 __ movptr(rax, Address(method, methodOopDesc::native_function_offset()));
1132 __ bind(L);
1133 }
1135 // pass mirror handle if static call
1136 { Label L;
1137 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1138 __ movl(t, Address(method, methodOopDesc::access_flags_offset()));
1139 __ testl(t, JVM_ACC_STATIC);
1140 __ jcc(Assembler::zero, L);
1141 // get mirror
1142 __ movptr(t, Address(method, methodOopDesc:: constants_offset()));
1143 __ movptr(t, Address(t, constantPoolOopDesc::pool_holder_offset_in_bytes()));
1144 __ movptr(t, Address(t, mirror_offset));
1145 // copy mirror into activation object
1146 __ movptr(STATE(_oop_temp), t);
1147 // pass handle to mirror
1148 #ifdef _LP64
1149 __ lea(c_rarg1, STATE(_oop_temp));
1150 #else
1151 __ lea(t, STATE(_oop_temp));
1152 __ movptr(Address(rsp, wordSize), t);
1153 #endif // _LP64
1154 __ bind(L);
1155 }
1156 #ifdef ASSERT
1157 {
1158 Label L;
1159 __ push(t);
1160 __ get_thread(t); // get vm's javathread*
1161 __ cmpptr(t, STATE(_thread));
1162 __ jcc(Assembler::equal, L);
1163 __ int3();
1164 __ bind(L);
1165 __ pop(t);
1166 }
1167 #endif //
1169 // pass JNIEnv
1170 #ifdef _LP64
1171 __ lea(c_rarg0, Address(thread, JavaThread::jni_environment_offset()));
1172 #else
1173 __ movptr(thread, STATE(_thread)); // get thread
1174 __ lea(t, Address(thread, JavaThread::jni_environment_offset()));
1176 __ movptr(Address(rsp, 0), t);
1177 #endif // _LP64
1179 #ifdef ASSERT
1180 {
1181 Label L;
1182 __ push(t);
1183 __ get_thread(t); // get vm's javathread*
1184 __ cmpptr(t, STATE(_thread));
1185 __ jcc(Assembler::equal, L);
1186 __ int3();
1187 __ bind(L);
1188 __ pop(t);
1189 }
1190 #endif //
1192 #ifdef ASSERT
1193 { Label L;
1194 __ movl(t, Address(thread, JavaThread::thread_state_offset()));
1195 __ cmpl(t, _thread_in_Java);
1196 __ jcc(Assembler::equal, L);
1197 __ stop("Wrong thread state in native stub");
1198 __ bind(L);
1199 }
1200 #endif
1202 // Change state to native (we save the return address in the thread, since it might not
1203 // be pushed on the stack when we do a a stack traversal). It is enough that the pc()
1204 // points into the right code segment. It does not have to be the correct return pc.
1206 __ set_last_Java_frame(thread, noreg, rbp, __ pc());
1208 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native);
1210 __ call(rax);
1212 // result potentially in rdx:rax or ST0
1213 __ movptr(method, STATE(_method));
1214 NOT_LP64(__ movptr(thread, STATE(_thread));) // get thread
1216 // The potential result is in ST(0) & rdx:rax
1217 // With C++ interpreter we leave any possible result in ST(0) until we are in result handler and then
1218 // we do the appropriate stuff for returning the result. rdx:rax must always be saved because just about
1219 // anything we do here will destroy it, st(0) is only saved if we re-enter the vm where it would
1220 // be destroyed.
1221 // It is safe to do these pushes because state is _thread_in_native and return address will be found
1222 // via _last_native_pc and not via _last_jave_sp
1224 // Must save the value of ST(0)/xmm0 since it could be destroyed before we get to result handler
1225 { Label Lpush, Lskip;
1226 ExternalAddress float_handler(AbstractInterpreter::result_handler(T_FLOAT));
1227 ExternalAddress double_handler(AbstractInterpreter::result_handler(T_DOUBLE));
1228 __ cmpptr(STATE(_result_handler), float_handler.addr());
1229 __ jcc(Assembler::equal, Lpush);
1230 __ cmpptr(STATE(_result_handler), double_handler.addr());
1231 __ jcc(Assembler::notEqual, Lskip);
1232 __ bind(Lpush);
1233 __ subptr(rsp, 2*wordSize);
1234 if ( UseSSE < 2 ) {
1235 __ fstp_d(Address(rsp, 0));
1236 } else {
1237 __ movdbl(Address(rsp, 0), xmm0);
1238 }
1239 __ bind(Lskip);
1240 }
1242 // save rax:rdx for potential use by result handler.
1243 __ push(rax);
1244 #ifndef _LP64
1245 __ push(rdx);
1246 #endif // _LP64
1248 // Either restore the MXCSR register after returning from the JNI Call
1249 // or verify that it wasn't changed.
1250 if (VM_Version::supports_sse()) {
1251 if (RestoreMXCSROnJNICalls) {
1252 __ ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
1253 }
1254 else if (CheckJNICalls ) {
1255 __ call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
1256 }
1257 }
1259 #ifndef _LP64
1260 // Either restore the x87 floating pointer control word after returning
1261 // from the JNI call or verify that it wasn't changed.
1262 if (CheckJNICalls) {
1263 __ call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
1264 }
1265 #endif // _LP64
1268 // change thread state
1269 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native_trans);
1270 if(os::is_MP()) {
1271 // Write serialization page so VM thread can do a pseudo remote membar.
1272 // We use the current thread pointer to calculate a thread specific
1273 // offset to write to within the page. This minimizes bus traffic
1274 // due to cache line collision.
1275 __ serialize_memory(thread, rcx);
1276 }
1278 // check for safepoint operation in progress and/or pending suspend requests
1279 { Label Continue;
1281 __ cmp32(ExternalAddress(SafepointSynchronize::address_of_state()),
1282 SafepointSynchronize::_not_synchronized);
1284 // threads running native code and they are expected to self-suspend
1285 // when leaving the _thread_in_native state. We need to check for
1286 // pending suspend requests here.
1287 Label L;
1288 __ jcc(Assembler::notEqual, L);
1289 __ cmpl(Address(thread, JavaThread::suspend_flags_offset()), 0);
1290 __ jcc(Assembler::equal, Continue);
1291 __ bind(L);
1293 // Don't use call_VM as it will see a possible pending exception and forward it
1294 // and never return here preventing us from clearing _last_native_pc down below.
1295 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are
1296 // preserved and correspond to the bcp/locals pointers.
1297 //
1299 ((MacroAssembler*)_masm)->call_VM_leaf(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1300 thread);
1301 __ increment(rsp, wordSize);
1303 __ movptr(method, STATE(_method));
1304 __ verify_oop(method);
1305 __ movptr(thread, STATE(_thread)); // get thread
1307 __ bind(Continue);
1308 }
1310 // change thread state
1311 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_Java);
1313 __ reset_last_Java_frame(thread, true, true);
1315 // reset handle block
1316 __ movptr(t, Address(thread, JavaThread::active_handles_offset()));
1317 __ movptr(Address(t, JNIHandleBlock::top_offset_in_bytes()), (int32_t)NULL_WORD);
1319 // If result was an oop then unbox and save it in the frame
1320 { Label L;
1321 Label no_oop, store_result;
1322 ExternalAddress oop_handler(AbstractInterpreter::result_handler(T_OBJECT));
1323 __ cmpptr(STATE(_result_handler), oop_handler.addr());
1324 __ jcc(Assembler::notEqual, no_oop);
1325 #ifndef _LP64
1326 __ pop(rdx);
1327 #endif // _LP64
1328 __ pop(rax);
1329 __ testptr(rax, rax);
1330 __ jcc(Assembler::zero, store_result);
1331 // unbox
1332 __ movptr(rax, Address(rax, 0));
1333 __ bind(store_result);
1334 __ movptr(STATE(_oop_temp), rax);
1335 // keep stack depth as expected by pushing oop which will eventually be discarded
1336 __ push(rax);
1337 #ifndef _LP64
1338 __ push(rdx);
1339 #endif // _LP64
1340 __ bind(no_oop);
1341 }
1343 {
1344 Label no_reguard;
1345 __ cmpl(Address(thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_disabled);
1346 __ jcc(Assembler::notEqual, no_reguard);
1348 __ pusha();
1349 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)));
1350 __ popa();
1352 __ bind(no_reguard);
1353 }
1356 // QQQ Seems like for native methods we simply return and the caller will see the pending
1357 // exception and do the right thing. Certainly the interpreter will, don't know about
1358 // compiled methods.
1359 // Seems that the answer to above is no this is wrong. The old code would see the exception
1360 // and forward it before doing the unlocking and notifying jvmdi that method has exited.
1361 // This seems wrong need to investigate the spec.
1363 // handle exceptions (exception handling will handle unlocking!)
1364 { Label L;
1365 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
1366 __ jcc(Assembler::zero, L);
1367 __ bind(pending_exception_present);
1369 // There are potential results on the stack (rax/rdx, ST(0)) we ignore these and simply
1370 // return and let caller deal with exception. This skips the unlocking here which
1371 // seems wrong but seems to be what asm interpreter did. Can't find this in the spec.
1372 // Note: must preverve method in rbx
1373 //
1375 // remove activation
1377 __ movptr(t, STATE(_sender_sp));
1378 __ leave(); // remove frame anchor
1379 __ pop(rdi); // get return address
1380 __ movptr(state, STATE(_prev_link)); // get previous state for return
1381 __ mov(rsp, t); // set sp to sender sp
1382 __ push(rdi); // push throwing pc
1383 // The skips unlocking!! This seems to be what asm interpreter does but seems
1384 // very wrong. Not clear if this violates the spec.
1385 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1386 __ bind(L);
1387 }
1389 // do unlocking if necessary
1390 { Label L;
1391 __ movl(t, Address(method, methodOopDesc::access_flags_offset()));
1392 __ testl(t, JVM_ACC_SYNCHRONIZED);
1393 __ jcc(Assembler::zero, L);
1394 // the code below should be shared with interpreter macro assembler implementation
1395 { Label unlock;
1396 const Register monitor = NOT_LP64(rdx) LP64_ONLY(c_rarg1);
1397 // BasicObjectLock will be first in list, since this is a synchronized method. However, need
1398 // to check that the object has not been unlocked by an explicit monitorexit bytecode.
1399 __ movptr(monitor, STATE(_monitor_base));
1400 __ subptr(monitor, frame::interpreter_frame_monitor_size() * wordSize); // address of initial monitor
1402 __ movptr(t, Address(monitor, BasicObjectLock::obj_offset_in_bytes()));
1403 __ testptr(t, t);
1404 __ jcc(Assembler::notZero, unlock);
1406 // Entry already unlocked, need to throw exception
1407 __ MacroAssembler::call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_illegal_monitor_state_exception));
1408 __ should_not_reach_here();
1410 __ bind(unlock);
1411 __ unlock_object(monitor);
1412 // unlock can blow rbx so restore it for path that needs it below
1413 __ movptr(method, STATE(_method));
1414 }
1415 __ bind(L);
1416 }
1418 // jvmti support
1419 // Note: This must happen _after_ handling/throwing any exceptions since
1420 // the exception handler code notifies the runtime of method exits
1421 // too. If this happens before, method entry/exit notifications are
1422 // not properly paired (was bug - gri 11/22/99).
1423 __ notify_method_exit(vtos, InterpreterMacroAssembler::NotifyJVMTI);
1425 // restore potential result in rdx:rax, call result handler to restore potential result in ST0 & handle result
1426 #ifndef _LP64
1427 __ pop(rdx);
1428 #endif // _LP64
1429 __ pop(rax);
1430 __ movptr(t, STATE(_result_handler)); // get result handler
1431 __ call(t); // call result handler to convert to tosca form
1433 // remove activation
1435 __ movptr(t, STATE(_sender_sp));
1437 __ leave(); // remove frame anchor
1438 __ pop(rdi); // get return address
1439 __ movptr(state, STATE(_prev_link)); // get previous state for return (if c++ interpreter was caller)
1440 __ mov(rsp, t); // set sp to sender sp
1441 __ jmp(rdi);
1443 // invocation counter overflow
1444 if (inc_counter) {
1445 // Handle overflow of counter and compile method
1446 __ bind(invocation_counter_overflow);
1447 generate_counter_overflow(&continue_after_compile);
1448 }
1450 return entry_point;
1451 }
1453 // Generate entries that will put a result type index into rcx
1454 void CppInterpreterGenerator::generate_deopt_handling() {
1456 Label return_from_deopt_common;
1458 // Generate entries that will put a result type index into rcx
1459 // deopt needs to jump to here to enter the interpreter (return a result)
1460 deopt_frame_manager_return_atos = __ pc();
1462 // rax is live here
1463 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_OBJECT)); // Result stub address array index
1464 __ jmp(return_from_deopt_common);
1467 // deopt needs to jump to here to enter the interpreter (return a result)
1468 deopt_frame_manager_return_btos = __ pc();
1470 // rax is live here
1471 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_BOOLEAN)); // Result stub address array index
1472 __ jmp(return_from_deopt_common);
1474 // deopt needs to jump to here to enter the interpreter (return a result)
1475 deopt_frame_manager_return_itos = __ pc();
1477 // rax is live here
1478 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_INT)); // Result stub address array index
1479 __ jmp(return_from_deopt_common);
1481 // deopt needs to jump to here to enter the interpreter (return a result)
1483 deopt_frame_manager_return_ltos = __ pc();
1484 // rax,rdx are live here
1485 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_LONG)); // Result stub address array index
1486 __ jmp(return_from_deopt_common);
1488 // deopt needs to jump to here to enter the interpreter (return a result)
1490 deopt_frame_manager_return_ftos = __ pc();
1491 // st(0) is live here
1492 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_FLOAT)); // Result stub address array index
1493 __ jmp(return_from_deopt_common);
1495 // deopt needs to jump to here to enter the interpreter (return a result)
1496 deopt_frame_manager_return_dtos = __ pc();
1498 // st(0) is live here
1499 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_DOUBLE)); // Result stub address array index
1500 __ jmp(return_from_deopt_common);
1502 // deopt needs to jump to here to enter the interpreter (return a result)
1503 deopt_frame_manager_return_vtos = __ pc();
1505 __ movl(rcx, AbstractInterpreter::BasicType_as_index(T_VOID));
1507 // Deopt return common
1508 // an index is present in rcx that lets us move any possible result being
1509 // return to the interpreter's stack
1510 //
1511 // Because we have a full sized interpreter frame on the youngest
1512 // activation the stack is pushed too deep to share the tosca to
1513 // stack converters directly. We shrink the stack to the desired
1514 // amount and then push result and then re-extend the stack.
1515 // We could have the code in size_activation layout a short
1516 // frame for the top activation but that would look different
1517 // than say sparc (which needs a full size activation because
1518 // the windows are in the way. Really it could be short? QQQ
1519 //
1520 __ bind(return_from_deopt_common);
1522 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter)));
1524 // setup rsp so we can push the "result" as needed.
1525 __ movptr(rsp, STATE(_stack)); // trim stack (is prepushed)
1526 __ addptr(rsp, wordSize); // undo prepush
1528 ExternalAddress tosca_to_stack((address)CppInterpreter::_tosca_to_stack);
1529 // Address index(noreg, rcx, Address::times_ptr);
1530 __ movptr(rcx, ArrayAddress(tosca_to_stack, Address(noreg, rcx, Address::times_ptr)));
1531 // __ movl(rcx, Address(noreg, rcx, Address::times_ptr, int(AbstractInterpreter::_tosca_to_stack)));
1532 __ call(rcx); // call result converter
1534 __ movl(STATE(_msg), (int)BytecodeInterpreter::deopt_resume);
1535 __ lea(rsp, Address(rsp, -wordSize)); // prepush stack (result if any already present)
1536 __ movptr(STATE(_stack), rsp); // inform interpreter of new stack depth (parameters removed,
1537 // result if any on stack already )
1538 __ movptr(rsp, STATE(_stack_limit)); // restore expression stack to full depth
1539 }
1541 // Generate the code to handle a more_monitors message from the c++ interpreter
1542 void CppInterpreterGenerator::generate_more_monitors() {
1545 Label entry, loop;
1546 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1547 // 1. compute new pointers // rsp: old expression stack top
1548 __ movptr(rdx, STATE(_stack_base)); // rdx: old expression stack bottom
1549 __ subptr(rsp, entry_size); // move expression stack top limit
1550 __ subptr(STATE(_stack), entry_size); // update interpreter stack top
1551 __ subptr(STATE(_stack_limit), entry_size); // inform interpreter
1552 __ subptr(rdx, entry_size); // move expression stack bottom
1553 __ movptr(STATE(_stack_base), rdx); // inform interpreter
1554 __ movptr(rcx, STATE(_stack)); // set start value for copy loop
1555 __ jmp(entry);
1556 // 2. move expression stack contents
1557 __ bind(loop);
1558 __ movptr(rbx, Address(rcx, entry_size)); // load expression stack word from old location
1559 __ movptr(Address(rcx, 0), rbx); // and store it at new location
1560 __ addptr(rcx, wordSize); // advance to next word
1561 __ bind(entry);
1562 __ cmpptr(rcx, rdx); // check if bottom reached
1563 __ jcc(Assembler::notEqual, loop); // if not at bottom then copy next word
1564 // now zero the slot so we can find it.
1565 __ movptr(Address(rdx, BasicObjectLock::obj_offset_in_bytes()), (int32_t) NULL_WORD);
1566 __ movl(STATE(_msg), (int)BytecodeInterpreter::got_monitors);
1567 }
1570 // Initial entry to C++ interpreter from the call_stub.
1571 // This entry point is called the frame manager since it handles the generation
1572 // of interpreter activation frames via requests directly from the vm (via call_stub)
1573 // and via requests from the interpreter. The requests from the call_stub happen
1574 // directly thru the entry point. Requests from the interpreter happen via returning
1575 // from the interpreter and examining the message the interpreter has returned to
1576 // the frame manager. The frame manager can take the following requests:
1578 // NO_REQUEST - error, should never happen.
1579 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1580 // allocate a new monitor.
1581 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1582 // happens during entry during the entry via the call stub.
1583 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1584 //
1585 // Arguments:
1586 //
1587 // rbx: methodOop
1588 // rcx: receiver - unused (retrieved from stack as needed)
1589 // rsi/r13: previous frame manager state (NULL from the call_stub/c1/c2)
1590 //
1591 //
1592 // Stack layout at entry
1593 //
1594 // [ return address ] <--- rsp
1595 // [ parameter n ]
1596 // ...
1597 // [ parameter 1 ]
1598 // [ expression stack ]
1599 //
1600 //
1601 // We are free to blow any registers we like because the call_stub which brought us here
1602 // initially has preserved the callee save registers already.
1603 //
1604 //
1606 static address interpreter_frame_manager = NULL;
1608 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1610 // rbx: methodOop
1611 // rsi/r13: sender sp
1613 // Because we redispatch "recursive" interpreter entries thru this same entry point
1614 // the "input" register usage is a little strange and not what you expect coming
1615 // from the call_stub. From the call stub rsi/rdi (current/previous) interpreter
1616 // state are NULL but on "recursive" dispatches they are what you'd expect.
1617 // rsi: current interpreter state (C++ interpreter) must preserve (null from call_stub/c1/c2)
1620 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1621 // the code is pretty much ready. Would need to change the test below and for good measure
1622 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1623 // routines. Not clear this is worth it yet.
1625 if (interpreter_frame_manager) return interpreter_frame_manager;
1627 address entry_point = __ pc();
1629 // Fast accessor methods share this entry point.
1630 // This works because frame manager is in the same codelet
1631 if (UseFastAccessorMethods && !synchronized) __ bind(fast_accessor_slow_entry_path);
1633 Label dispatch_entry_2;
1634 __ movptr(rcx, sender_sp_on_entry);
1635 __ movptr(state, (int32_t)NULL_WORD); // no current activation
1637 __ jmp(dispatch_entry_2);
1639 const Register locals = rdi;
1641 Label re_dispatch;
1643 __ bind(re_dispatch);
1645 // save sender sp (doesn't include return address
1646 __ lea(rcx, Address(rsp, wordSize));
1648 __ bind(dispatch_entry_2);
1650 // save sender sp
1651 __ push(rcx);
1653 const Address size_of_parameters(rbx, methodOopDesc::size_of_parameters_offset());
1654 const Address size_of_locals (rbx, methodOopDesc::size_of_locals_offset());
1655 const Address access_flags (rbx, methodOopDesc::access_flags_offset());
1657 // const Address monitor_block_top (rbp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
1658 // const Address monitor_block_bot (rbp, frame::interpreter_frame_initial_sp_offset * wordSize);
1659 // const Address monitor(rbp, frame::interpreter_frame_initial_sp_offset * wordSize - (int)sizeof(BasicObjectLock));
1661 // get parameter size (always needed)
1662 __ load_unsigned_short(rcx, size_of_parameters);
1664 // rbx: methodOop
1665 // rcx: size of parameters
1666 __ load_unsigned_short(rdx, size_of_locals); // get size of locals in words
1668 __ subptr(rdx, rcx); // rdx = no. of additional locals
1670 // see if we've got enough room on the stack for locals plus overhead.
1671 generate_stack_overflow_check(); // C++
1673 // c++ interpreter does not use stack banging or any implicit exceptions
1674 // leave for now to verify that check is proper.
1675 bang_stack_shadow_pages(false);
1679 // compute beginning of parameters (rdi)
1680 __ lea(locals, Address(rsp, rcx, Address::times_ptr, wordSize));
1682 // save sender's sp
1683 // __ movl(rcx, rsp);
1685 // get sender's sp
1686 __ pop(rcx);
1688 // get return address
1689 __ pop(rax);
1691 // rdx - # of additional locals
1692 // allocate space for locals
1693 // explicitly initialize locals
1694 {
1695 Label exit, loop;
1696 __ testl(rdx, rdx); // (32bit ok)
1697 __ jcc(Assembler::lessEqual, exit); // do nothing if rdx <= 0
1698 __ bind(loop);
1699 __ push((int32_t)NULL_WORD); // initialize local variables
1700 __ decrement(rdx); // until everything initialized
1701 __ jcc(Assembler::greater, loop);
1702 __ bind(exit);
1703 }
1706 // Assumes rax = return address
1708 // allocate and initialize new interpreterState and method expression stack
1709 // IN(locals) -> locals
1710 // IN(state) -> any current interpreter activation
1711 // destroys rax, rcx, rdx, rdi
1712 // OUT (state) -> new interpreterState
1713 // OUT(rsp) -> bottom of methods expression stack
1715 generate_compute_interpreter_state(state, locals, rcx, false);
1717 // Call interpreter
1719 Label call_interpreter;
1720 __ bind(call_interpreter);
1722 // c++ interpreter does not use stack banging or any implicit exceptions
1723 // leave for now to verify that check is proper.
1724 bang_stack_shadow_pages(false);
1727 // Call interpreter enter here if message is
1728 // set and we know stack size is valid
1730 Label call_interpreter_2;
1732 __ bind(call_interpreter_2);
1734 {
1735 const Register thread = NOT_LP64(rcx) LP64_ONLY(r15_thread);
1737 #ifdef _LP64
1738 __ mov(c_rarg0, state);
1739 #else
1740 __ push(state); // push arg to interpreter
1741 __ movptr(thread, STATE(_thread));
1742 #endif // _LP64
1744 // We can setup the frame anchor with everything we want at this point
1745 // as we are thread_in_Java and no safepoints can occur until we go to
1746 // vm mode. We do have to clear flags on return from vm but that is it
1747 //
1748 __ movptr(Address(thread, JavaThread::last_Java_fp_offset()), rbp);
1749 __ movptr(Address(thread, JavaThread::last_Java_sp_offset()), rsp);
1751 // Call the interpreter
1753 RuntimeAddress normal(CAST_FROM_FN_PTR(address, BytecodeInterpreter::run));
1754 RuntimeAddress checking(CAST_FROM_FN_PTR(address, BytecodeInterpreter::runWithChecks));
1756 __ call(JvmtiExport::can_post_interpreter_events() ? checking : normal);
1757 NOT_LP64(__ pop(rax);) // discard parameter to run
1758 //
1759 // state is preserved since it is callee saved
1760 //
1762 // reset_last_Java_frame
1764 NOT_LP64(__ movl(thread, STATE(_thread));)
1765 __ reset_last_Java_frame(thread, true, true);
1766 }
1768 // examine msg from interpreter to determine next action
1770 __ movl(rdx, STATE(_msg)); // Get new message
1772 Label call_method;
1773 Label return_from_interpreted_method;
1774 Label throw_exception;
1775 Label bad_msg;
1776 Label do_OSR;
1778 __ cmpl(rdx, (int32_t)BytecodeInterpreter::call_method);
1779 __ jcc(Assembler::equal, call_method);
1780 __ cmpl(rdx, (int32_t)BytecodeInterpreter::return_from_method);
1781 __ jcc(Assembler::equal, return_from_interpreted_method);
1782 __ cmpl(rdx, (int32_t)BytecodeInterpreter::do_osr);
1783 __ jcc(Assembler::equal, do_OSR);
1784 __ cmpl(rdx, (int32_t)BytecodeInterpreter::throwing_exception);
1785 __ jcc(Assembler::equal, throw_exception);
1786 __ cmpl(rdx, (int32_t)BytecodeInterpreter::more_monitors);
1787 __ jcc(Assembler::notEqual, bad_msg);
1789 // Allocate more monitor space, shuffle expression stack....
1791 generate_more_monitors();
1793 __ jmp(call_interpreter);
1795 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1796 unctrap_frame_manager_entry = __ pc();
1797 //
1798 // Load the registers we need.
1799 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter)));
1800 __ movptr(rsp, STATE(_stack_limit)); // restore expression stack to full depth
1801 __ jmp(call_interpreter_2);
1805 //=============================================================================
1806 // Returning from a compiled method into a deopted method. The bytecode at the
1807 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1808 // for the template based interpreter). Any stack space that was used by the
1809 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1810 // so all that we have to do is place any pending result on the expression stack
1811 // and resume execution on the next bytecode.
1814 generate_deopt_handling();
1815 __ jmp(call_interpreter);
1818 // Current frame has caught an exception we need to dispatch to the
1819 // handler. We can get here because a native interpreter frame caught
1820 // an exception in which case there is no handler and we must rethrow
1821 // If it is a vanilla interpreted frame the we simply drop into the
1822 // interpreter and let it do the lookup.
1824 Interpreter::_rethrow_exception_entry = __ pc();
1825 // rax: exception
1826 // rdx: return address/pc that threw exception
1828 Label return_with_exception;
1829 Label unwind_and_forward;
1831 // restore state pointer.
1832 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter)));
1834 __ movptr(rbx, STATE(_method)); // get method
1835 #ifdef _LP64
1836 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
1837 #else
1838 __ movl(rcx, STATE(_thread)); // get thread
1840 // Store exception with interpreter will expect it
1841 __ movptr(Address(rcx, Thread::pending_exception_offset()), rax);
1842 #endif // _LP64
1844 // is current frame vanilla or native?
1846 __ movl(rdx, access_flags);
1847 __ testl(rdx, JVM_ACC_NATIVE);
1848 __ jcc(Assembler::zero, return_with_exception); // vanilla interpreted frame, handle directly
1850 // We drop thru to unwind a native interpreted frame with a pending exception
1851 // We jump here for the initial interpreter frame with exception pending
1852 // We unwind the current acivation and forward it to our caller.
1854 __ bind(unwind_and_forward);
1856 // unwind rbp, return stack to unextended value and re-push return address
1858 __ movptr(rcx, STATE(_sender_sp));
1859 __ leave();
1860 __ pop(rdx);
1861 __ mov(rsp, rcx);
1862 __ push(rdx);
1863 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1865 // Return point from a call which returns a result in the native abi
1866 // (c1/c2/jni-native). This result must be processed onto the java
1867 // expression stack.
1868 //
1869 // A pending exception may be present in which case there is no result present
1871 Label resume_interpreter;
1872 Label do_float;
1873 Label do_double;
1874 Label done_conv;
1876 address compiled_entry = __ pc();
1878 // The FPU stack is clean if UseSSE >= 2 but must be cleaned in other cases
1879 if (UseSSE < 2) {
1880 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter)));
1881 __ movptr(rbx, STATE(_result._to_call._callee)); // get method just executed
1882 __ movl(rcx, Address(rbx, methodOopDesc::result_index_offset()));
1883 __ cmpl(rcx, AbstractInterpreter::BasicType_as_index(T_FLOAT)); // Result stub address array index
1884 __ jcc(Assembler::equal, do_float);
1885 __ cmpl(rcx, AbstractInterpreter::BasicType_as_index(T_DOUBLE)); // Result stub address array index
1886 __ jcc(Assembler::equal, do_double);
1887 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2)
1888 __ empty_FPU_stack();
1889 #endif // COMPILER2
1890 __ jmp(done_conv);
1892 __ bind(do_float);
1893 #ifdef COMPILER2
1894 for (int i = 1; i < 8; i++) {
1895 __ ffree(i);
1896 }
1897 #endif // COMPILER2
1898 __ jmp(done_conv);
1899 __ bind(do_double);
1900 #ifdef COMPILER2
1901 for (int i = 1; i < 8; i++) {
1902 __ ffree(i);
1903 }
1904 #endif // COMPILER2
1905 __ jmp(done_conv);
1906 } else {
1907 __ MacroAssembler::verify_FPU(0, "generate_return_entry_for compiled");
1908 __ jmp(done_conv);
1909 }
1911 #if 0
1912 // emit a sentinel we can test for when converting an interpreter
1913 // entry point to a compiled entry point.
1914 __ a_long(Interpreter::return_sentinel);
1915 __ a_long((int)compiled_entry);
1916 #endif
1918 // Return point to interpreter from compiled/native method
1920 InternalAddress return_from_native_method(__ pc());
1922 __ bind(done_conv);
1925 // Result if any is in tosca. The java expression stack is in the state that the
1926 // calling convention left it (i.e. params may or may not be present)
1927 // Copy the result from tosca and place it on java expression stack.
1929 // Restore rsi/r13 as compiled code may not preserve it
1931 __ lea(state, Address(rbp, -(int)sizeof(BytecodeInterpreter)));
1933 // restore stack to what we had when we left (in case i2c extended it)
1935 __ movptr(rsp, STATE(_stack));
1936 __ lea(rsp, Address(rsp, wordSize));
1938 // If there is a pending exception then we don't really have a result to process
1940 #ifdef _LP64
1941 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
1942 #else
1943 __ movptr(rcx, STATE(_thread)); // get thread
1944 __ cmpptr(Address(rcx, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
1945 #endif // _LP64
1946 __ jcc(Assembler::notZero, return_with_exception);
1948 // get method just executed
1949 __ movptr(rbx, STATE(_result._to_call._callee));
1951 // callee left args on top of expression stack, remove them
1952 __ load_unsigned_short(rcx, Address(rbx, methodOopDesc::size_of_parameters_offset()));
1953 __ lea(rsp, Address(rsp, rcx, Address::times_ptr));
1955 __ movl(rcx, Address(rbx, methodOopDesc::result_index_offset()));
1956 ExternalAddress tosca_to_stack((address)CppInterpreter::_tosca_to_stack);
1957 // Address index(noreg, rax, Address::times_ptr);
1958 __ movptr(rcx, ArrayAddress(tosca_to_stack, Address(noreg, rcx, Address::times_ptr)));
1959 // __ movl(rcx, Address(noreg, rcx, Address::times_ptr, int(AbstractInterpreter::_tosca_to_stack)));
1960 __ call(rcx); // call result converter
1961 __ jmp(resume_interpreter);
1963 // An exception is being caught on return to a vanilla interpreter frame.
1964 // Empty the stack and resume interpreter
1966 __ bind(return_with_exception);
1968 // Exception present, empty stack
1969 __ movptr(rsp, STATE(_stack_base));
1970 __ jmp(resume_interpreter);
1972 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1973 // interpreter call, or native) and unwind this interpreter activation.
1974 // All monitors should be unlocked.
1976 __ bind(return_from_interpreted_method);
1978 Label return_to_initial_caller;
1980 __ movptr(rbx, STATE(_method)); // get method just executed
1981 __ cmpptr(STATE(_prev_link), (int32_t)NULL_WORD); // returning from "recursive" interpreter call?
1982 __ movl(rax, Address(rbx, methodOopDesc::result_index_offset())); // get result type index
1983 __ jcc(Assembler::equal, return_to_initial_caller); // back to native code (call_stub/c1/c2)
1985 // Copy result to callers java stack
1986 ExternalAddress stack_to_stack((address)CppInterpreter::_stack_to_stack);
1987 // Address index(noreg, rax, Address::times_ptr);
1989 __ movptr(rax, ArrayAddress(stack_to_stack, Address(noreg, rax, Address::times_ptr)));
1990 // __ movl(rax, Address(noreg, rax, Address::times_ptr, int(AbstractInterpreter::_stack_to_stack)));
1991 __ call(rax); // call result converter
1993 Label unwind_recursive_activation;
1994 __ bind(unwind_recursive_activation);
1996 // returning to interpreter method from "recursive" interpreter call
1997 // result converter left rax pointing to top of the java stack for method we are returning
1998 // to. Now all we must do is unwind the state from the completed call
2000 __ movptr(state, STATE(_prev_link)); // unwind state
2001 __ leave(); // pop the frame
2002 __ mov(rsp, rax); // unwind stack to remove args
2004 // Resume the interpreter. The current frame contains the current interpreter
2005 // state object.
2006 //
2008 __ bind(resume_interpreter);
2010 // state == interpreterState object for method we are resuming
2012 __ movl(STATE(_msg), (int)BytecodeInterpreter::method_resume);
2013 __ lea(rsp, Address(rsp, -wordSize)); // prepush stack (result if any already present)
2014 __ movptr(STATE(_stack), rsp); // inform interpreter of new stack depth (parameters removed,
2015 // result if any on stack already )
2016 __ movptr(rsp, STATE(_stack_limit)); // restore expression stack to full depth
2017 __ jmp(call_interpreter_2); // No need to bang
2019 // interpreter returning to native code (call_stub/c1/c2)
2020 // convert result and unwind initial activation
2021 // rax - result index
2023 __ bind(return_to_initial_caller);
2024 ExternalAddress stack_to_native((address)CppInterpreter::_stack_to_native_abi);
2025 // Address index(noreg, rax, Address::times_ptr);
2027 __ movptr(rax, ArrayAddress(stack_to_native, Address(noreg, rax, Address::times_ptr)));
2028 __ call(rax); // call result converter
2030 Label unwind_initial_activation;
2031 __ bind(unwind_initial_activation);
2033 // RETURN TO CALL_STUB/C1/C2 code (result if any in rax/rdx ST(0))
2035 /* Current stack picture
2037 [ incoming parameters ]
2038 [ extra locals ]
2039 [ return address to CALL_STUB/C1/C2]
2040 fp -> [ CALL_STUB/C1/C2 fp ]
2041 BytecodeInterpreter object
2042 expression stack
2043 sp ->
2045 */
2047 // return restoring the stack to the original sender_sp value
2049 __ movptr(rcx, STATE(_sender_sp));
2050 __ leave();
2051 __ pop(rdi); // get return address
2052 // set stack to sender's sp
2053 __ mov(rsp, rcx);
2054 __ jmp(rdi); // return to call_stub
2056 // OSR request, adjust return address to make current frame into adapter frame
2057 // and enter OSR nmethod
2059 __ bind(do_OSR);
2061 Label remove_initial_frame;
2063 // We are going to pop this frame. Is there another interpreter frame underneath
2064 // it or is it callstub/compiled?
2066 // Move buffer to the expected parameter location
2067 __ movptr(rcx, STATE(_result._osr._osr_buf));
2069 __ movptr(rax, STATE(_result._osr._osr_entry));
2071 __ cmpptr(STATE(_prev_link), (int32_t)NULL_WORD); // returning from "recursive" interpreter call?
2072 __ jcc(Assembler::equal, remove_initial_frame); // back to native code (call_stub/c1/c2)
2074 __ movptr(sender_sp_on_entry, STATE(_sender_sp)); // get sender's sp in expected register
2075 __ leave(); // pop the frame
2076 __ mov(rsp, sender_sp_on_entry); // trim any stack expansion
2079 // We know we are calling compiled so push specialized return
2080 // method uses specialized entry, push a return so we look like call stub setup
2081 // this path will handle fact that result is returned in registers and not
2082 // on the java stack.
2084 __ pushptr(return_from_native_method.addr());
2086 __ jmp(rax);
2088 __ bind(remove_initial_frame);
2090 __ movptr(rdx, STATE(_sender_sp));
2091 __ leave();
2092 // get real return
2093 __ pop(rsi);
2094 // set stack to sender's sp
2095 __ mov(rsp, rdx);
2096 // repush real return
2097 __ push(rsi);
2098 // Enter OSR nmethod
2099 __ jmp(rax);
2104 // Call a new method. All we do is (temporarily) trim the expression stack
2105 // push a return address to bring us back to here and leap to the new entry.
2107 __ bind(call_method);
2109 // stack points to next free location and not top element on expression stack
2110 // method expects sp to be pointing to topmost element
2112 __ movptr(rsp, STATE(_stack)); // pop args to c++ interpreter, set sp to java stack top
2113 __ lea(rsp, Address(rsp, wordSize));
2115 __ movptr(rbx, STATE(_result._to_call._callee)); // get method to execute
2117 // don't need a return address if reinvoking interpreter
2119 // Make it look like call_stub calling conventions
2121 // Get (potential) receiver
2122 __ load_unsigned_short(rcx, size_of_parameters); // get size of parameters in words
2124 ExternalAddress recursive(CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation));
2125 __ pushptr(recursive.addr()); // make it look good in the debugger
2127 InternalAddress entry(entry_point);
2128 __ cmpptr(STATE(_result._to_call._callee_entry_point), entry.addr()); // returning to interpreter?
2129 __ jcc(Assembler::equal, re_dispatch); // yes
2131 __ pop(rax); // pop dummy address
2134 // get specialized entry
2135 __ movptr(rax, STATE(_result._to_call._callee_entry_point));
2136 // set sender SP
2137 __ mov(sender_sp_on_entry, rsp);
2139 // method uses specialized entry, push a return so we look like call stub setup
2140 // this path will handle fact that result is returned in registers and not
2141 // on the java stack.
2143 __ pushptr(return_from_native_method.addr());
2145 __ jmp(rax);
2147 __ bind(bad_msg);
2148 __ stop("Bad message from interpreter");
2150 // Interpreted method "returned" with an exception pass it on...
2151 // Pass result, unwind activation and continue/return to interpreter/call_stub
2152 // We handle result (if any) differently based on return to interpreter or call_stub
2154 Label unwind_initial_with_pending_exception;
2156 __ bind(throw_exception);
2157 __ cmpptr(STATE(_prev_link), (int32_t)NULL_WORD); // returning from recursive interpreter call?
2158 __ jcc(Assembler::equal, unwind_initial_with_pending_exception); // no, back to native code (call_stub/c1/c2)
2159 __ movptr(rax, STATE(_locals)); // pop parameters get new stack value
2160 __ addptr(rax, wordSize); // account for prepush before we return
2161 __ jmp(unwind_recursive_activation);
2163 __ bind(unwind_initial_with_pending_exception);
2165 // We will unwind the current (initial) interpreter frame and forward
2166 // the exception to the caller. We must put the exception in the
2167 // expected register and clear pending exception and then forward.
2169 __ jmp(unwind_and_forward);
2171 interpreter_frame_manager = entry_point;
2172 return entry_point;
2173 }
2175 address AbstractInterpreterGenerator::generate_method_entry(AbstractInterpreter::MethodKind kind) {
2176 // determine code generation flags
2177 bool synchronized = false;
2178 address entry_point = NULL;
2180 switch (kind) {
2181 case Interpreter::zerolocals : break;
2182 case Interpreter::zerolocals_synchronized: synchronized = true; break;
2183 case Interpreter::native : entry_point = ((InterpreterGenerator*)this)->generate_native_entry(false); break;
2184 case Interpreter::native_synchronized : entry_point = ((InterpreterGenerator*)this)->generate_native_entry(true); break;
2185 case Interpreter::empty : entry_point = ((InterpreterGenerator*)this)->generate_empty_entry(); break;
2186 case Interpreter::accessor : entry_point = ((InterpreterGenerator*)this)->generate_accessor_entry(); break;
2187 case Interpreter::abstract : entry_point = ((InterpreterGenerator*)this)->generate_abstract_entry(); break;
2189 case Interpreter::java_lang_math_sin : // fall thru
2190 case Interpreter::java_lang_math_cos : // fall thru
2191 case Interpreter::java_lang_math_tan : // fall thru
2192 case Interpreter::java_lang_math_abs : // fall thru
2193 case Interpreter::java_lang_math_log : // fall thru
2194 case Interpreter::java_lang_math_log10 : // fall thru
2195 case Interpreter::java_lang_math_sqrt : entry_point = ((InterpreterGenerator*)this)->generate_math_entry(kind); break;
2196 default : ShouldNotReachHere(); break;
2197 }
2199 if (entry_point) return entry_point;
2201 return ((InterpreterGenerator*)this)->generate_normal_entry(synchronized);
2203 }
2205 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2206 : CppInterpreterGenerator(code) {
2207 generate_all(); // down here so it can be "virtual"
2208 }
2210 // Deoptimization helpers for C++ interpreter
2212 // How much stack a method activation needs in words.
2213 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
2215 const int stub_code = 4; // see generate_call_stub
2216 // Save space for one monitor to get into the interpreted method in case
2217 // the method is synchronized
2218 int monitor_size = method->is_synchronized() ?
2219 1*frame::interpreter_frame_monitor_size() : 0;
2221 // total static overhead size. Account for interpreter state object, return
2222 // address, saved rbp and 2 words for a "static long no_params() method" issue.
2224 const int overhead_size = sizeof(BytecodeInterpreter)/wordSize +
2225 ( frame::sender_sp_offset - frame::link_offset) + 2;
2227 const int method_stack = (method->max_locals() + method->max_stack()) *
2228 Interpreter::stackElementWords();
2229 return overhead_size + method_stack + stub_code;
2230 }
2232 // returns the activation size.
2233 static int size_activation_helper(int extra_locals_size, int monitor_size) {
2234 return (extra_locals_size + // the addition space for locals
2235 2*BytesPerWord + // return address and saved rbp
2236 2*BytesPerWord + // "static long no_params() method" issue
2237 sizeof(BytecodeInterpreter) + // interpreterState
2238 monitor_size); // monitors
2239 }
2241 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2242 frame* caller,
2243 frame* current,
2244 methodOop method,
2245 intptr_t* locals,
2246 intptr_t* stack,
2247 intptr_t* stack_base,
2248 intptr_t* monitor_base,
2249 intptr_t* frame_bottom,
2250 bool is_top_frame
2251 )
2252 {
2253 // What about any vtable?
2254 //
2255 to_fill->_thread = JavaThread::current();
2256 // This gets filled in later but make it something recognizable for now
2257 to_fill->_bcp = method->code_base();
2258 to_fill->_locals = locals;
2259 to_fill->_constants = method->constants()->cache();
2260 to_fill->_method = method;
2261 to_fill->_mdx = NULL;
2262 to_fill->_stack = stack;
2263 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2264 to_fill->_msg = deopt_resume2;
2265 } else {
2266 to_fill->_msg = method_resume;
2267 }
2268 to_fill->_result._to_call._bcp_advance = 0;
2269 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2270 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2271 to_fill->_prev_link = NULL;
2273 to_fill->_sender_sp = caller->unextended_sp();
2275 if (caller->is_interpreted_frame()) {
2276 interpreterState prev = caller->get_interpreterState();
2277 to_fill->_prev_link = prev;
2278 // *current->register_addr(GR_Iprev_state) = (intptr_t) prev;
2279 // Make the prev callee look proper
2280 prev->_result._to_call._callee = method;
2281 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2282 prev->_result._to_call._bcp_advance = 5;
2283 } else {
2284 prev->_result._to_call._bcp_advance = 3;
2285 }
2286 }
2287 to_fill->_oop_temp = NULL;
2288 to_fill->_stack_base = stack_base;
2289 // Need +1 here because stack_base points to the word just above the first expr stack entry
2290 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2291 // See generate_compute_interpreter_state.
2292 to_fill->_stack_limit = stack_base - (method->max_stack() + 1);
2293 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2295 to_fill->_self_link = to_fill;
2296 assert(stack >= to_fill->_stack_limit && stack < to_fill->_stack_base,
2297 "Stack top out of range");
2298 }
2300 int AbstractInterpreter::layout_activation(methodOop method,
2301 int tempcount, //
2302 int popframe_extra_args,
2303 int moncount,
2304 int callee_param_count,
2305 int callee_locals,
2306 frame* caller,
2307 frame* interpreter_frame,
2308 bool is_top_frame) {
2310 assert(popframe_extra_args == 0, "FIX ME");
2311 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2312 // does as far as allocating an interpreter frame.
2313 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2314 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2315 // as determined by a previous call to this method.
2316 // It is also guaranteed to be walkable even though it is in a skeletal state
2317 // NOTE: return size is in words not bytes
2318 // NOTE: tempcount is the current size of the java expression stack. For top most
2319 // frames we will allocate a full sized expression stack and not the curback
2320 // version that non-top frames have.
2322 // Calculate the amount our frame will be adjust by the callee. For top frame
2323 // this is zero.
2325 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2326 // calculates the extra locals based on itself. Not what the callee does
2327 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2328 // as getting sender_sp correct.
2330 int extra_locals_size = (callee_locals - callee_param_count) * BytesPerWord;
2331 int monitor_size = sizeof(BasicObjectLock) * moncount;
2333 // First calculate the frame size without any java expression stack
2334 int short_frame_size = size_activation_helper(extra_locals_size,
2335 monitor_size);
2337 // Now with full size expression stack
2338 int full_frame_size = short_frame_size + method->max_stack() * BytesPerWord;
2340 // and now with only live portion of the expression stack
2341 short_frame_size = short_frame_size + tempcount * BytesPerWord;
2343 // the size the activation is right now. Only top frame is full size
2344 int frame_size = (is_top_frame ? full_frame_size : short_frame_size);
2346 if (interpreter_frame != NULL) {
2347 #ifdef ASSERT
2348 assert(caller->unextended_sp() == interpreter_frame->interpreter_frame_sender_sp(), "Frame not properly walkable");
2349 #endif
2351 // MUCHO HACK
2353 intptr_t* frame_bottom = (intptr_t*) ((intptr_t)interpreter_frame->sp() - (full_frame_size - frame_size));
2355 /* Now fillin the interpreterState object */
2357 // The state object is the first thing on the frame and easily located
2359 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2362 // Find the locals pointer. This is rather simple on x86 because there is no
2363 // confusing rounding at the callee to account for. We can trivially locate
2364 // our locals based on the current fp().
2365 // Note: the + 2 is for handling the "static long no_params() method" issue.
2366 // (too bad I don't really remember that issue well...)
2368 intptr_t* locals;
2369 // If the caller is interpreted we need to make sure that locals points to the first
2370 // argument that the caller passed and not in an area where the stack might have been extended.
2371 // because the stack to stack to converter needs a proper locals value in order to remove the
2372 // arguments from the caller and place the result in the proper location. Hmm maybe it'd be
2373 // simpler if we simply stored the result in the BytecodeInterpreter object and let the c++ code
2374 // adjust the stack?? HMMM QQQ
2375 //
2376 if (caller->is_interpreted_frame()) {
2377 // locals must agree with the caller because it will be used to set the
2378 // caller's tos when we return.
2379 interpreterState prev = caller->get_interpreterState();
2380 // stack() is prepushed.
2381 locals = prev->stack() + method->size_of_parameters();
2382 // locals = caller->unextended_sp() + (method->size_of_parameters() - 1);
2383 if (locals != interpreter_frame->fp() + frame::sender_sp_offset + (method->max_locals() - 1) + 2) {
2384 // os::breakpoint();
2385 }
2386 } else {
2387 // this is where a c2i would have placed locals (except for the +2)
2388 locals = interpreter_frame->fp() + frame::sender_sp_offset + (method->max_locals() - 1) + 2;
2389 }
2391 intptr_t* monitor_base = (intptr_t*) cur_state;
2392 intptr_t* stack_base = (intptr_t*) ((intptr_t) monitor_base - monitor_size);
2393 /* +1 because stack is always prepushed */
2394 intptr_t* stack = (intptr_t*) ((intptr_t) stack_base - (tempcount + 1) * BytesPerWord);
2397 BytecodeInterpreter::layout_interpreterState(cur_state,
2398 caller,
2399 interpreter_frame,
2400 method,
2401 locals,
2402 stack,
2403 stack_base,
2404 monitor_base,
2405 frame_bottom,
2406 is_top_frame);
2408 // BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2409 }
2410 return frame_size/BytesPerWord;
2411 }
2413 #endif // CC_INTERP (all)