Tue, 01 Apr 2014 09:36:49 +0200
8032410: compiler/uncommontrap/TestStackBangRbp.java times out on Solaris-Sparc V9
Summary: make compiled code bang the stack by the worst case size of the interpreter frame at deoptimization points.
Reviewed-by: twisti, kvn
2 /*
3 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
4 * Copyright 2012, 2014 SAP AG. All rights reserved.
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
27 #include "precompiled.hpp"
28 #include "asm/assembler.hpp"
29 #include "asm/macroAssembler.inline.hpp"
30 #include "interpreter/bytecodeHistogram.hpp"
31 #include "interpreter/cppInterpreter.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "interpreter/interpreterGenerator.hpp"
34 #include "interpreter/interpreterRuntime.hpp"
35 #include "oops/arrayOop.hpp"
36 #include "oops/methodData.hpp"
37 #include "oops/method.hpp"
38 #include "oops/oop.inline.hpp"
39 #include "prims/jvmtiExport.hpp"
40 #include "prims/jvmtiThreadState.hpp"
41 #include "runtime/arguments.hpp"
42 #include "runtime/deoptimization.hpp"
43 #include "runtime/frame.inline.hpp"
44 #include "runtime/interfaceSupport.hpp"
45 #include "runtime/sharedRuntime.hpp"
46 #include "runtime/stubRoutines.hpp"
47 #include "runtime/synchronizer.hpp"
48 #include "runtime/timer.hpp"
49 #include "runtime/vframeArray.hpp"
50 #include "utilities/debug.hpp"
51 #ifdef SHARK
52 #include "shark/shark_globals.hpp"
53 #endif
55 #ifdef CC_INTERP
57 #define __ _masm->
59 // Contains is used for identifying interpreter frames during a stack-walk.
60 // A frame with a PC in InterpretMethod must be identified as a normal C frame.
61 bool CppInterpreter::contains(address pc) {
62 return _code->contains(pc);
63 }
65 #ifdef PRODUCT
66 #define BLOCK_COMMENT(str) // nothing
67 #else
68 #define BLOCK_COMMENT(str) __ block_comment(str)
69 #endif
71 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
73 static address interpreter_frame_manager = NULL;
74 static address frame_manager_specialized_return = NULL;
75 static address native_entry = NULL;
77 static address interpreter_return_address = NULL;
79 static address unctrap_frame_manager_entry = NULL;
81 static address deopt_frame_manager_return_atos = NULL;
82 static address deopt_frame_manager_return_btos = NULL;
83 static address deopt_frame_manager_return_itos = NULL;
84 static address deopt_frame_manager_return_ltos = NULL;
85 static address deopt_frame_manager_return_ftos = NULL;
86 static address deopt_frame_manager_return_dtos = NULL;
87 static address deopt_frame_manager_return_vtos = NULL;
89 // A result handler converts/unboxes a native call result into
90 // a java interpreter/compiler result. The current frame is an
91 // interpreter frame.
92 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
93 return AbstractInterpreterGenerator::generate_result_handler_for(type);
94 }
96 // tosca based result to c++ interpreter stack based result.
97 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
98 //
99 // A result is in the native abi result register from a native
100 // method call. We need to return this result to the interpreter by
101 // pushing the result on the interpreter's stack.
102 //
103 // Registers alive:
104 // R3_ARG1(R3_RET)/F1_ARG1(F1_RET) - result to move
105 // R4_ARG2 - address of tos
106 // LR
107 //
108 // Registers updated:
109 // R3_RET(R3_ARG1) - address of new tos (== R17_tos for T_VOID)
110 //
112 int number_of_used_slots = 1;
114 const Register tos = R4_ARG2;
115 Label done;
116 Label is_false;
118 address entry = __ pc();
120 switch (type) {
121 case T_BOOLEAN:
122 __ cmpwi(CCR0, R3_RET, 0);
123 __ beq(CCR0, is_false);
124 __ li(R3_RET, 1);
125 __ stw(R3_RET, 0, tos);
126 __ b(done);
127 __ bind(is_false);
128 __ li(R3_RET, 0);
129 __ stw(R3_RET, 0, tos);
130 break;
131 case T_BYTE:
132 case T_CHAR:
133 case T_SHORT:
134 case T_INT:
135 __ stw(R3_RET, 0, tos);
136 break;
137 case T_LONG:
138 number_of_used_slots = 2;
139 // mark unused slot for debugging
140 // long goes to topmost slot
141 __ std(R3_RET, -BytesPerWord, tos);
142 __ li(R3_RET, 0);
143 __ std(R3_RET, 0, tos);
144 break;
145 case T_OBJECT:
146 __ verify_oop(R3_RET);
147 __ std(R3_RET, 0, tos);
148 break;
149 case T_FLOAT:
150 __ stfs(F1_RET, 0, tos);
151 break;
152 case T_DOUBLE:
153 number_of_used_slots = 2;
154 // mark unused slot for debugging
155 __ li(R3_RET, 0);
156 __ std(R3_RET, 0, tos);
157 // double goes to topmost slot
158 __ stfd(F1_RET, -BytesPerWord, tos);
159 break;
160 case T_VOID:
161 number_of_used_slots = 0;
162 break;
163 default:
164 ShouldNotReachHere();
165 }
167 __ BIND(done);
169 // new expression stack top
170 __ addi(R3_RET, tos, -BytesPerWord * number_of_used_slots);
172 __ blr();
174 return entry;
175 }
177 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
178 //
179 // Copy the result from the callee's stack to the caller's stack,
180 // caller and callee both being interpreted.
181 //
182 // Registers alive
183 // R3_ARG1 - address of callee's tos + BytesPerWord
184 // R4_ARG2 - address of caller's tos [i.e. free location]
185 // LR
186 //
187 // stack grows upwards, memory grows downwards.
188 //
189 // [ free ] <-- callee's tos
190 // [ optional result ] <-- R3_ARG1
191 // [ optional dummy ]
192 // ...
193 // [ free ] <-- caller's tos, R4_ARG2
194 // ...
195 // Registers updated
196 // R3_RET(R3_ARG1) - address of caller's new tos
197 //
198 // stack grows upwards, memory grows downwards.
199 //
200 // [ free ] <-- current tos, R3_RET
201 // [ optional result ]
202 // [ optional dummy ]
203 // ...
204 //
206 const Register from = R3_ARG1;
207 const Register ret = R3_ARG1;
208 const Register tos = R4_ARG2;
209 const Register tmp1 = R21_tmp1;
210 const Register tmp2 = R22_tmp2;
212 address entry = __ pc();
214 switch (type) {
215 case T_BOOLEAN:
216 case T_BYTE:
217 case T_CHAR:
218 case T_SHORT:
219 case T_INT:
220 case T_FLOAT:
221 __ lwz(tmp1, 0, from);
222 __ stw(tmp1, 0, tos);
223 // New expression stack top.
224 __ addi(ret, tos, - BytesPerWord);
225 break;
226 case T_LONG:
227 case T_DOUBLE:
228 // Move both entries for debug purposes even though only one is live.
229 __ ld(tmp1, BytesPerWord, from);
230 __ ld(tmp2, 0, from);
231 __ std(tmp1, 0, tos);
232 __ std(tmp2, -BytesPerWord, tos);
233 // New expression stack top.
234 __ addi(ret, tos, - 2 * BytesPerWord); // two slots
235 break;
236 case T_OBJECT:
237 __ ld(tmp1, 0, from);
238 __ verify_oop(tmp1);
239 __ std(tmp1, 0, tos);
240 // New expression stack top.
241 __ addi(ret, tos, - BytesPerWord);
242 break;
243 case T_VOID:
244 // New expression stack top.
245 __ mr(ret, tos);
246 break;
247 default:
248 ShouldNotReachHere();
249 }
251 __ blr();
253 return entry;
254 }
256 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
257 //
258 // Load a result from the callee's stack into the caller's expecting
259 // return register, callee being interpreted, caller being call stub
260 // or jit code.
261 //
262 // Registers alive
263 // R3_ARG1 - callee expression tos + BytesPerWord
264 // LR
265 //
266 // stack grows upwards, memory grows downwards.
267 //
268 // [ free ] <-- callee's tos
269 // [ optional result ] <-- R3_ARG1
270 // [ optional dummy ]
271 // ...
272 //
273 // Registers updated
274 // R3_RET(R3_ARG1)/F1_RET - result
275 //
277 const Register from = R3_ARG1;
278 const Register ret = R3_ARG1;
279 const FloatRegister fret = F1_ARG1;
281 address entry = __ pc();
283 // Implemented uniformly for both kinds of endianness. The interpreter
284 // implements boolean, byte, char, and short as jint (4 bytes).
285 switch (type) {
286 case T_BOOLEAN:
287 case T_CHAR:
288 // zero extension
289 __ lwz(ret, 0, from);
290 break;
291 case T_BYTE:
292 case T_SHORT:
293 case T_INT:
294 // sign extension
295 __ lwa(ret, 0, from);
296 break;
297 case T_LONG:
298 __ ld(ret, 0, from);
299 break;
300 case T_OBJECT:
301 __ ld(ret, 0, from);
302 __ verify_oop(ret);
303 break;
304 case T_FLOAT:
305 __ lfs(fret, 0, from);
306 break;
307 case T_DOUBLE:
308 __ lfd(fret, 0, from);
309 break;
310 case T_VOID:
311 break;
312 default:
313 ShouldNotReachHere();
314 }
316 __ blr();
318 return entry;
319 }
321 address CppInterpreter::return_entry(TosState state, int length, Bytecodes::Code code) {
322 assert(interpreter_return_address != NULL, "Not initialized");
323 return interpreter_return_address;
324 }
326 address CppInterpreter::deopt_entry(TosState state, int length) {
327 address ret = NULL;
328 if (length != 0) {
329 switch (state) {
330 case atos: ret = deopt_frame_manager_return_atos; break;
331 case btos: ret = deopt_frame_manager_return_itos; break;
332 case ctos:
333 case stos:
334 case itos: ret = deopt_frame_manager_return_itos; break;
335 case ltos: ret = deopt_frame_manager_return_ltos; break;
336 case ftos: ret = deopt_frame_manager_return_ftos; break;
337 case dtos: ret = deopt_frame_manager_return_dtos; break;
338 case vtos: ret = deopt_frame_manager_return_vtos; break;
339 default: ShouldNotReachHere();
340 }
341 } else {
342 ret = unctrap_frame_manager_entry; // re-execute the bytecode (e.g. uncommon trap, popframe)
343 }
344 assert(ret != NULL, "Not initialized");
345 return ret;
346 }
348 //
349 // Helpers for commoning out cases in the various type of method entries.
350 //
352 //
353 // Registers alive
354 // R16_thread - JavaThread*
355 // R1_SP - old stack pointer
356 // R19_method - callee's Method
357 // R17_tos - address of caller's tos (prepushed)
358 // R15_prev_state - address of caller's BytecodeInterpreter or 0
359 // return_pc in R21_tmp15 (only when called within generate_native_entry)
360 //
361 // Registers updated
362 // R14_state - address of callee's interpreter state
363 // R1_SP - new stack pointer
364 // CCR4_is_synced - current method is synchronized
365 //
366 void CppInterpreterGenerator::generate_compute_interpreter_state(Label& stack_overflow_return) {
367 //
368 // Stack layout at this point:
369 //
370 // F1 [TOP_IJAVA_FRAME_ABI] <-- R1_SP
371 // alignment (optional)
372 // [F1's outgoing Java arguments] <-- R17_tos
373 // ...
374 // F2 [PARENT_IJAVA_FRAME_ABI]
375 // ...
377 //=============================================================================
378 // Allocate space for locals other than the parameters, the
379 // interpreter state, monitors, and the expression stack.
381 const Register local_count = R21_tmp1;
382 const Register parameter_count = R22_tmp2;
383 const Register max_stack = R23_tmp3;
384 // Must not be overwritten within this method!
385 // const Register return_pc = R29_tmp9;
387 const ConditionRegister is_synced = CCR4_is_synced;
388 const ConditionRegister is_native = CCR6;
389 const ConditionRegister is_static = CCR7;
391 assert(is_synced != is_native, "condition code registers must be distinct");
392 assert(is_synced != is_static, "condition code registers must be distinct");
393 assert(is_native != is_static, "condition code registers must be distinct");
395 {
397 // Local registers
398 const Register top_frame_size = R24_tmp4;
399 const Register access_flags = R25_tmp5;
400 const Register state_offset = R26_tmp6;
401 Register mem_stack_limit = R27_tmp7;
402 const Register page_size = R28_tmp8;
404 BLOCK_COMMENT("compute_interpreter_state {");
406 // access_flags = method->access_flags();
407 // TODO: PPC port: assert(4 == sizeof(AccessFlags), "unexpected field size");
408 __ lwa(access_flags, method_(access_flags));
410 // parameter_count = method->constMethod->size_of_parameters();
411 // TODO: PPC port: assert(2 == ConstMethod::sz_size_of_parameters(), "unexpected field size");
412 __ ld(max_stack, in_bytes(Method::const_offset()), R19_method); // Max_stack holds constMethod for a while.
413 __ lhz(parameter_count, in_bytes(ConstMethod::size_of_parameters_offset()), max_stack);
415 // local_count = method->constMethod()->max_locals();
416 // TODO: PPC port: assert(2 == ConstMethod::sz_max_locals(), "unexpected field size");
417 __ lhz(local_count, in_bytes(ConstMethod::size_of_locals_offset()), max_stack);
419 // max_stack = method->constMethod()->max_stack();
420 // TODO: PPC port: assert(2 == ConstMethod::sz_max_stack(), "unexpected field size");
421 __ lhz(max_stack, in_bytes(ConstMethod::max_stack_offset()), max_stack);
423 if (EnableInvokeDynamic) {
424 // Take into account 'extra_stack_entries' needed by method handles (see method.hpp).
425 __ addi(max_stack, max_stack, Method::extra_stack_entries());
426 }
428 // mem_stack_limit = thread->stack_limit();
429 __ ld(mem_stack_limit, thread_(stack_overflow_limit));
431 // Point locals at the first argument. Method's locals are the
432 // parameters on top of caller's expression stack.
434 // tos points past last Java argument
435 __ sldi(R18_locals, parameter_count, Interpreter::logStackElementSize);
436 __ add(R18_locals, R17_tos, R18_locals);
438 // R18_locals - i*BytesPerWord points to i-th Java local (i starts at 0)
440 // Set is_native, is_synced, is_static - will be used later.
441 __ testbitdi(is_native, R0, access_flags, JVM_ACC_NATIVE_BIT);
442 __ testbitdi(is_synced, R0, access_flags, JVM_ACC_SYNCHRONIZED_BIT);
443 assert(is_synced->is_nonvolatile(), "is_synced must be non-volatile");
444 __ testbitdi(is_static, R0, access_flags, JVM_ACC_STATIC_BIT);
446 // PARENT_IJAVA_FRAME_ABI
447 //
448 // frame_size =
449 // round_to((local_count - parameter_count)*BytesPerWord +
450 // 2*BytesPerWord +
451 // alignment +
452 // frame::interpreter_frame_cinterpreterstate_size_in_bytes()
453 // sizeof(PARENT_IJAVA_FRAME_ABI)
454 // method->is_synchronized() ? sizeof(BasicObjectLock) : 0 +
455 // max_stack*BytesPerWord,
456 // 16)
457 //
458 // Note that this calculation is exactly mirrored by
459 // AbstractInterpreter::layout_activation_impl() [ and
460 // AbstractInterpreter::size_activation() ]. Which is used by
461 // deoptimization so that it can allocate the proper sized
462 // frame. This only happens for interpreted frames so the extra
463 // notes below about max_stack below are not important. The other
464 // thing to note is that for interpreter frames other than the
465 // current activation the size of the stack is the size of the live
466 // portion of the stack at the particular bcp and NOT the maximum
467 // stack that the method might use.
468 //
469 // If we're calling a native method, we replace max_stack (which is
470 // zero) with space for the worst-case signature handler varargs
471 // vector, which is:
472 //
473 // max_stack = max(Argument::n_register_parameters, parameter_count+2);
474 //
475 // We add two slots to the parameter_count, one for the jni
476 // environment and one for a possible native mirror. We allocate
477 // space for at least the number of ABI registers, even though
478 // InterpreterRuntime::slow_signature_handler won't write more than
479 // parameter_count+2 words when it creates the varargs vector at the
480 // top of the stack. The generated slow signature handler will just
481 // load trash into registers beyond the necessary number. We're
482 // still going to cut the stack back by the ABI register parameter
483 // count so as to get SP+16 pointing at the ABI outgoing parameter
484 // area, so we need to allocate at least that much even though we're
485 // going to throw it away.
486 //
488 // Adjust max_stack for native methods:
489 Label skip_native_calculate_max_stack;
490 __ bfalse(is_native, skip_native_calculate_max_stack);
491 // if (is_native) {
492 // max_stack = max(Argument::n_register_parameters, parameter_count+2);
493 __ addi(max_stack, parameter_count, 2*Interpreter::stackElementWords);
494 __ cmpwi(CCR0, max_stack, Argument::n_register_parameters);
495 __ bge(CCR0, skip_native_calculate_max_stack);
496 __ li(max_stack, Argument::n_register_parameters);
497 // }
498 __ bind(skip_native_calculate_max_stack);
499 // max_stack is now in bytes
500 __ slwi(max_stack, max_stack, Interpreter::logStackElementSize);
502 // Calculate number of non-parameter locals (in slots):
503 Label not_java;
504 __ btrue(is_native, not_java);
505 // if (!is_native) {
506 // local_count = non-parameter local count
507 __ sub(local_count, local_count, parameter_count);
508 // } else {
509 // // nothing to do: method->max_locals() == 0 for native methods
510 // }
511 __ bind(not_java);
514 // Calculate top_frame_size and parent_frame_resize.
515 {
516 const Register parent_frame_resize = R12_scratch2;
518 BLOCK_COMMENT("Compute top_frame_size.");
519 // top_frame_size = TOP_IJAVA_FRAME_ABI
520 // + size of interpreter state
521 __ li(top_frame_size, frame::top_ijava_frame_abi_size
522 + frame::interpreter_frame_cinterpreterstate_size_in_bytes());
523 // + max_stack
524 __ add(top_frame_size, top_frame_size, max_stack);
525 // + stack slots for a BasicObjectLock for synchronized methods
526 {
527 Label not_synced;
528 __ bfalse(is_synced, not_synced);
529 __ addi(top_frame_size, top_frame_size, frame::interpreter_frame_monitor_size_in_bytes());
530 __ bind(not_synced);
531 }
532 // align
533 __ round_to(top_frame_size, frame::alignment_in_bytes);
536 BLOCK_COMMENT("Compute parent_frame_resize.");
537 // parent_frame_resize = R1_SP - R17_tos
538 __ sub(parent_frame_resize, R1_SP, R17_tos);
539 //__ li(parent_frame_resize, 0);
540 // + PARENT_IJAVA_FRAME_ABI
541 // + extra two slots for the no-parameter/no-locals
542 // method result
543 __ addi(parent_frame_resize, parent_frame_resize,
544 frame::parent_ijava_frame_abi_size
545 + 2*Interpreter::stackElementSize);
546 // + (locals_count - params_count)
547 __ sldi(R0, local_count, Interpreter::logStackElementSize);
548 __ add(parent_frame_resize, parent_frame_resize, R0);
549 // align
550 __ round_to(parent_frame_resize, frame::alignment_in_bytes);
552 //
553 // Stack layout at this point:
554 //
555 // The new frame F0 hasn't yet been pushed, F1 is still the top frame.
556 //
557 // F0 [TOP_IJAVA_FRAME_ABI]
558 // alignment (optional)
559 // [F0's full operand stack]
560 // [F0's monitors] (optional)
561 // [F0's BytecodeInterpreter object]
562 // F1 [PARENT_IJAVA_FRAME_ABI]
563 // alignment (optional)
564 // [F0's Java result]
565 // [F0's non-arg Java locals]
566 // [F1's outgoing Java arguments] <-- R17_tos
567 // ...
568 // F2 [PARENT_IJAVA_FRAME_ABI]
569 // ...
572 // Calculate new R14_state
573 // and
574 // test that the new memory stack pointer is above the limit,
575 // throw a StackOverflowError otherwise.
576 __ sub(R11_scratch1/*F1's SP*/, R1_SP, parent_frame_resize);
577 __ addi(R14_state, R11_scratch1/*F1's SP*/,
578 -frame::interpreter_frame_cinterpreterstate_size_in_bytes());
579 __ sub(R11_scratch1/*F0's SP*/,
580 R11_scratch1/*F1's SP*/, top_frame_size);
582 BLOCK_COMMENT("Test for stack overflow:");
583 __ cmpld(CCR0/*is_stack_overflow*/, R11_scratch1, mem_stack_limit);
584 __ blt(CCR0/*is_stack_overflow*/, stack_overflow_return);
587 //=============================================================================
588 // Frame_size doesn't overflow the stack. Allocate new frame and
589 // initialize interpreter state.
591 // Register state
592 //
593 // R15 - local_count
594 // R16 - parameter_count
595 // R17 - max_stack
596 //
597 // R18 - frame_size
598 // R19 - access_flags
599 // CCR4_is_synced - is_synced
600 //
601 // GR_Lstate - pointer to the uninitialized new BytecodeInterpreter.
603 // _last_Java_pc just needs to be close enough that we can identify
604 // the frame as an interpreted frame. It does not need to be the
605 // exact return address from either calling
606 // BytecodeInterpreter::InterpretMethod or the call to a jni native method.
607 // So we can initialize it here with a value of a bundle in this
608 // code fragment. We only do this initialization for java frames
609 // where InterpretMethod needs a a way to get a good pc value to
610 // store in the thread state. For interpreter frames used to call
611 // jni native code we just zero the value in the state and move an
612 // ip as needed in the native entry code.
613 //
614 // const Register last_Java_pc_addr = GR24_SCRATCH; // QQQ 27
615 // const Register last_Java_pc = GR26_SCRATCH;
617 // Must reference stack before setting new SP since Windows
618 // will not be able to deliver the exception on a bad SP.
619 // Windows also insists that we bang each page one at a time in order
620 // for the OS to map in the reserved pages. If we bang only
621 // the final page, Windows stops delivering exceptions to our
622 // VectoredExceptionHandler and terminates our program.
623 // Linux only requires a single bang but it's rare to have
624 // to bang more than 1 page so the code is enabled for both OS's.
626 // BANG THE STACK
627 //
628 // Nothing to do for PPC, because updating the SP will automatically
629 // bang the page.
631 // Up to here we have calculated the delta for the new C-frame and
632 // checked for a stack-overflow. Now we can savely update SP and
633 // resize the C-frame.
635 // R14_state has already been calculated.
636 __ push_interpreter_frame(top_frame_size, parent_frame_resize,
637 R25_tmp5, R26_tmp6, R27_tmp7, R28_tmp8);
639 }
641 //
642 // Stack layout at this point:
643 //
644 // F0 has been been pushed!
645 //
646 // F0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP
647 // alignment (optional) (now it's here, if required)
648 // [F0's full operand stack]
649 // [F0's monitors] (optional)
650 // [F0's BytecodeInterpreter object]
651 // F1 [PARENT_IJAVA_FRAME_ABI]
652 // alignment (optional) (now it's here, if required)
653 // [F0's Java result]
654 // [F0's non-arg Java locals]
655 // [F1's outgoing Java arguments]
656 // ...
657 // F2 [PARENT_IJAVA_FRAME_ABI]
658 // ...
659 //
660 // R14_state points to F0's BytecodeInterpreter object.
661 //
663 }
665 //=============================================================================
666 // new BytecodeInterpreter-object is save, let's initialize it:
667 BLOCK_COMMENT("New BytecodeInterpreter-object is save.");
669 {
670 // Locals
671 const Register bytecode_addr = R24_tmp4;
672 const Register constants = R25_tmp5;
673 const Register tos = R26_tmp6;
674 const Register stack_base = R27_tmp7;
675 const Register local_addr = R28_tmp8;
676 {
677 Label L;
678 __ btrue(is_native, L);
679 // if (!is_native) {
680 // bytecode_addr = constMethod->codes();
681 __ ld(bytecode_addr, method_(const));
682 __ addi(bytecode_addr, bytecode_addr, in_bytes(ConstMethod::codes_offset()));
683 // }
684 __ bind(L);
685 }
687 __ ld(constants, in_bytes(Method::const_offset()), R19_method);
688 __ ld(constants, in_bytes(ConstMethod::constants_offset()), constants);
690 // state->_prev_link = prev_state;
691 __ std(R15_prev_state, state_(_prev_link));
693 // For assertions only.
694 // TODO: not needed anyway because it coincides with `_monitor_base'. remove!
695 // state->_self_link = state;
696 DEBUG_ONLY(__ std(R14_state, state_(_self_link));)
698 // state->_thread = thread;
699 __ std(R16_thread, state_(_thread));
701 // state->_method = method;
702 __ std(R19_method, state_(_method));
704 // state->_locals = locals;
705 __ std(R18_locals, state_(_locals));
707 // state->_oop_temp = NULL;
708 __ li(R0, 0);
709 __ std(R0, state_(_oop_temp));
711 // state->_last_Java_fp = *R1_SP // Use *R1_SP as fp
712 __ ld(R0, _abi(callers_sp), R1_SP);
713 __ std(R0, state_(_last_Java_fp));
715 BLOCK_COMMENT("load Stack base:");
716 {
717 // Stack_base.
718 // if (!method->synchronized()) {
719 // stack_base = state;
720 // } else {
721 // stack_base = (uintptr_t)state - sizeof(BasicObjectLock);
722 // }
723 Label L;
724 __ mr(stack_base, R14_state);
725 __ bfalse(is_synced, L);
726 __ addi(stack_base, stack_base, -frame::interpreter_frame_monitor_size_in_bytes());
727 __ bind(L);
728 }
730 // state->_mdx = NULL;
731 __ li(R0, 0);
732 __ std(R0, state_(_mdx));
734 {
735 // if (method->is_native()) state->_bcp = NULL;
736 // else state->_bcp = bytecode_addr;
737 Label label1, label2;
738 __ bfalse(is_native, label1);
739 __ std(R0, state_(_bcp));
740 __ b(label2);
741 __ bind(label1);
742 __ std(bytecode_addr, state_(_bcp));
743 __ bind(label2);
744 }
747 // state->_result._to_call._callee = NULL;
748 __ std(R0, state_(_result._to_call._callee));
750 // state->_monitor_base = state;
751 __ std(R14_state, state_(_monitor_base));
753 // state->_msg = BytecodeInterpreter::method_entry;
754 __ li(R0, BytecodeInterpreter::method_entry);
755 __ stw(R0, state_(_msg));
757 // state->_last_Java_sp = R1_SP;
758 __ std(R1_SP, state_(_last_Java_sp));
760 // state->_stack_base = stack_base;
761 __ std(stack_base, state_(_stack_base));
763 // tos = stack_base - 1 slot (prepushed);
764 // state->_stack.Tos(tos);
765 __ addi(tos, stack_base, - Interpreter::stackElementSize);
766 __ std(tos, state_(_stack));
769 {
770 BLOCK_COMMENT("get last_Java_pc:");
771 // if (!is_native) state->_last_Java_pc = <some_ip_in_this_code_buffer>;
772 // else state->_last_Java_pc = NULL; (just for neatness)
773 Label label1, label2;
774 __ btrue(is_native, label1);
775 __ get_PC_trash_LR(R0);
776 __ std(R0, state_(_last_Java_pc));
777 __ b(label2);
778 __ bind(label1);
779 __ li(R0, 0);
780 __ std(R0, state_(_last_Java_pc));
781 __ bind(label2);
782 }
785 // stack_limit = tos - max_stack;
786 __ sub(R0, tos, max_stack);
787 // state->_stack_limit = stack_limit;
788 __ std(R0, state_(_stack_limit));
791 // cache = method->constants()->cache();
792 __ ld(R0, ConstantPool::cache_offset_in_bytes(), constants);
793 // state->_constants = method->constants()->cache();
794 __ std(R0, state_(_constants));
798 //=============================================================================
799 // synchronized method, allocate and initialize method object lock.
800 // if (!method->is_synchronized()) goto fill_locals_with_0x0s;
801 Label fill_locals_with_0x0s;
802 __ bfalse(is_synced, fill_locals_with_0x0s);
804 // pool_holder = method->constants()->pool_holder();
805 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
806 {
807 Label label1, label2;
808 // lockee = NULL; for java methods, correct value will be inserted in BytecodeInterpretMethod.hpp
809 __ li(R0,0);
810 __ bfalse(is_native, label2);
812 __ bfalse(is_static, label1);
813 // if (method->is_static()) lockee =
814 // pool_holder->klass_part()->java_mirror();
815 __ ld(R11_scratch1/*pool_holder*/, ConstantPool::pool_holder_offset_in_bytes(), constants);
816 __ ld(R0/*lockee*/, mirror_offset, R11_scratch1/*pool_holder*/);
817 __ b(label2);
819 __ bind(label1);
820 // else lockee = *(oop*)locals;
821 __ ld(R0/*lockee*/, 0, R18_locals);
822 __ bind(label2);
824 // monitor->set_obj(lockee);
825 __ std(R0/*lockee*/, BasicObjectLock::obj_offset_in_bytes(), stack_base);
826 }
828 // See if we need to zero the locals
829 __ BIND(fill_locals_with_0x0s);
832 //=============================================================================
833 // fill locals with 0x0s
834 Label locals_zeroed;
835 __ btrue(is_native, locals_zeroed);
837 if (true /* zerolocals */ || ClearInterpreterLocals) {
838 // local_count is already num_locals_slots - num_param_slots
839 __ sldi(R0, parameter_count, Interpreter::logStackElementSize);
840 __ sub(local_addr, R18_locals, R0);
841 __ cmpdi(CCR0, local_count, 0);
842 __ ble(CCR0, locals_zeroed);
844 __ mtctr(local_count);
845 //__ ld_const_addr(R0, (address) 0xcafe0000babe);
846 __ li(R0, 0);
848 Label zero_slot;
849 __ bind(zero_slot);
851 // first local is at local_addr
852 __ std(R0, 0, local_addr);
853 __ addi(local_addr, local_addr, -BytesPerWord);
854 __ bdnz(zero_slot);
855 }
857 __ BIND(locals_zeroed);
859 }
860 BLOCK_COMMENT("} compute_interpreter_state");
861 }
863 // Generate code to initiate compilation on invocation counter overflow.
864 void CppInterpreterGenerator::generate_counter_overflow(Label& continue_entry) {
865 // Registers alive
866 // R14_state
867 // R16_thread
868 //
869 // Registers updated
870 // R14_state
871 // R3_ARG1 (=R3_RET)
872 // R4_ARG2
874 // After entering the vm we remove the activation and retry the
875 // entry point in case the compilation is complete.
877 // InterpreterRuntime::frequency_counter_overflow takes one argument
878 // that indicates if the counter overflow occurs at a backwards
879 // branch (NULL bcp). We pass zero. The call returns the address
880 // of the verified entry point for the method or NULL if the
881 // compilation did not complete (either went background or bailed
882 // out).
883 __ li(R4_ARG2, 0);
885 // Pass false to call_VM so it doesn't check for pending exceptions,
886 // since at this point in the method invocation the exception
887 // handler would try to exit the monitor of synchronized methods
888 // which haven't been entered yet.
889 //
890 // Returns verified_entry_point or NULL, we don't care which.
891 //
892 // Do not use the variant `frequency_counter_overflow' that returns
893 // a structure, because this will change the argument list by a
894 // hidden parameter (gcc 4.1).
896 __ call_VM(noreg,
897 CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow),
898 R4_ARG2,
899 false);
900 // Returns verified_entry_point or NULL, we don't care which as we ignore it
901 // and run interpreted.
903 // Reload method, it may have moved.
904 __ ld(R19_method, state_(_method));
906 // We jump now to the label "continue_after_compile".
907 __ b(continue_entry);
908 }
910 // Increment invocation count and check for overflow.
911 //
912 // R19_method must contain Method* of method to profile.
913 void CppInterpreterGenerator::generate_counter_incr(Label& overflow) {
914 Label done;
915 const Register Rcounters = R12_scratch2;
916 const Register iv_be_count = R11_scratch1;
917 const Register invocation_limit = R12_scratch2;
918 const Register invocation_limit_addr = invocation_limit;
920 // Load and ev. allocate MethodCounters object.
921 __ get_method_counters(R19_method, Rcounters, done);
923 // Update standard invocation counters.
924 __ increment_invocation_counter(Rcounters, iv_be_count, R0);
926 // Compare against limit.
927 BLOCK_COMMENT("Compare counter against limit:");
928 assert(4 == sizeof(InvocationCounter::InterpreterInvocationLimit),
929 "must be 4 bytes");
930 __ load_const(invocation_limit_addr, (address)&InvocationCounter::InterpreterInvocationLimit);
931 __ lwa(invocation_limit, 0, invocation_limit_addr);
932 __ cmpw(CCR0, iv_be_count, invocation_limit);
933 __ bge(CCR0, overflow);
934 __ bind(done);
935 }
937 //
938 // Call a JNI method.
939 //
940 // Interpreter stub for calling a native method. (C++ interpreter)
941 // This sets up a somewhat different looking stack for calling the native method
942 // than the typical interpreter frame setup.
943 //
944 address CppInterpreterGenerator::generate_native_entry(void) {
945 if (native_entry != NULL) return native_entry;
946 address entry = __ pc();
948 // Read
949 // R16_thread
950 // R15_prev_state - address of caller's BytecodeInterpreter, if this snippet
951 // gets called by the frame manager.
952 // R19_method - callee's Method
953 // R17_tos - address of caller's tos
954 // R1_SP - caller's stack pointer
955 // R21_sender_SP - initial caller sp
956 //
957 // Update
958 // R14_state - address of caller's BytecodeInterpreter
959 // R3_RET - integer result, if any.
960 // F1_RET - float result, if any.
961 //
962 //
963 // Stack layout at this point:
964 //
965 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP
966 // alignment (optional)
967 // [outgoing Java arguments] <-- R17_tos
968 // ...
969 // PARENT [PARENT_IJAVA_FRAME_ABI]
970 // ...
971 //
973 const bool inc_counter = UseCompiler || CountCompiledCalls;
975 const Register signature_handler_fd = R21_tmp1;
976 const Register pending_exception = R22_tmp2;
977 const Register result_handler_addr = R23_tmp3;
978 const Register native_method_fd = R24_tmp4;
979 const Register access_flags = R25_tmp5;
980 const Register active_handles = R26_tmp6;
981 const Register sync_state = R27_tmp7;
982 const Register sync_state_addr = sync_state; // Address is dead after use.
983 const Register suspend_flags = R24_tmp4;
985 const Register return_pc = R28_tmp8; // Register will be locked for some time.
987 const ConditionRegister is_synced = CCR4_is_synced; // Live-on-exit from compute_interpreter_state.
990 // R1_SP still points to caller's SP at this point.
992 // Save initial_caller_sp to caller's abi. The caller frame must be
993 // resized before returning to get rid of the c2i arguments (if
994 // any).
995 // Override the saved SP with the senderSP so we can pop c2i
996 // arguments (if any) off when we return
997 __ std(R21_sender_SP, _top_ijava_frame_abi(initial_caller_sp), R1_SP);
999 // Save LR to caller's frame. We don't use _abi(lr) here, because it is not safe.
1000 __ mflr(return_pc);
1001 __ std(return_pc, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1003 assert(return_pc->is_nonvolatile(), "return_pc must be a non-volatile register");
1005 __ verify_method_ptr(R19_method);
1007 //=============================================================================
1009 // If this snippet gets called by the frame manager (at label
1010 // `call_special'), then R15_prev_state is valid. If this snippet
1011 // is not called by the frame manager, but e.g. by the call stub or
1012 // by compiled code, then R15_prev_state is invalid.
1013 {
1014 // Set R15_prev_state to 0 if we don't return to the frame
1015 // manager; we will return to the call_stub or to compiled code
1016 // instead. If R15_prev_state is 0 there will be only one
1017 // interpreter frame (we will set this up later) in this C frame!
1018 // So we must take care about retrieving prev_state_(_prev_link)
1019 // and restoring R1_SP when popping that interpreter.
1020 Label prev_state_is_valid;
1022 __ load_const(R11_scratch1/*frame_manager_returnpc_addr*/, (address)&frame_manager_specialized_return);
1023 __ ld(R12_scratch2/*frame_manager_returnpc*/, 0, R11_scratch1/*frame_manager_returnpc_addr*/);
1024 __ cmpd(CCR0, return_pc, R12_scratch2/*frame_manager_returnpc*/);
1025 __ beq(CCR0, prev_state_is_valid);
1027 __ li(R15_prev_state, 0);
1029 __ BIND(prev_state_is_valid);
1030 }
1032 //=============================================================================
1033 // Allocate new frame and initialize interpreter state.
1035 Label exception_return;
1036 Label exception_return_sync_check;
1037 Label stack_overflow_return;
1039 // Generate new interpreter state and jump to stack_overflow_return in case of
1040 // a stack overflow.
1041 generate_compute_interpreter_state(stack_overflow_return);
1043 //=============================================================================
1044 // Increment invocation counter. On overflow, entry to JNI method
1045 // will be compiled.
1046 Label invocation_counter_overflow;
1047 if (inc_counter) {
1048 generate_counter_incr(invocation_counter_overflow);
1049 }
1051 Label continue_after_compile;
1052 __ BIND(continue_after_compile);
1054 // access_flags = method->access_flags();
1055 // Load access flags.
1056 assert(access_flags->is_nonvolatile(),
1057 "access_flags must be in a non-volatile register");
1058 // Type check.
1059 // TODO: PPC port: assert(4 == sizeof(AccessFlags), "unexpected field size");
1060 __ lwz(access_flags, method_(access_flags));
1062 // We don't want to reload R19_method and access_flags after calls
1063 // to some helper functions.
1064 assert(R19_method->is_nonvolatile(), "R19_method must be a non-volatile register");
1066 // Check for synchronized methods. Must happen AFTER invocation counter
1067 // check, so method is not locked if counter overflows.
1069 {
1070 Label method_is_not_synced;
1071 // Is_synced is still alive.
1072 assert(is_synced->is_nonvolatile(), "is_synced must be non-volatile");
1073 __ bfalse(is_synced, method_is_not_synced);
1075 lock_method();
1076 // Reload method, it may have moved.
1077 __ ld(R19_method, state_(_method));
1079 __ BIND(method_is_not_synced);
1080 }
1082 // jvmti/jvmpi support
1083 __ notify_method_entry();
1085 // Reload method, it may have moved.
1086 __ ld(R19_method, state_(_method));
1088 //=============================================================================
1089 // Get and call the signature handler
1091 __ ld(signature_handler_fd, method_(signature_handler));
1092 Label call_signature_handler;
1094 __ cmpdi(CCR0, signature_handler_fd, 0);
1095 __ bne(CCR0, call_signature_handler);
1097 // Method has never been called. Either generate a specialized
1098 // handler or point to the slow one.
1099 //
1100 // Pass parameter 'false' to avoid exception check in call_VM.
1101 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), R19_method, false);
1103 // Check for an exception while looking up the target method. If we
1104 // incurred one, bail.
1105 __ ld(pending_exception, thread_(pending_exception));
1106 __ cmpdi(CCR0, pending_exception, 0);
1107 __ bne(CCR0, exception_return_sync_check); // has pending exception
1109 // reload method
1110 __ ld(R19_method, state_(_method));
1112 // Reload signature handler, it may have been created/assigned in the meanwhile
1113 __ ld(signature_handler_fd, method_(signature_handler));
1115 __ BIND(call_signature_handler);
1117 // Before we call the signature handler we push a new frame to
1118 // protect the interpreter frame volatile registers when we return
1119 // from jni but before we can get back to Java.
1121 // First set the frame anchor while the SP/FP registers are
1122 // convenient and the slow signature handler can use this same frame
1123 // anchor.
1125 // We have a TOP_IJAVA_FRAME here, which belongs to us.
1126 __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/);
1128 // Now the interpreter frame (and its call chain) have been
1129 // invalidated and flushed. We are now protected against eager
1130 // being enabled in native code. Even if it goes eager the
1131 // registers will be reloaded as clean and we will invalidate after
1132 // the call so no spurious flush should be possible.
1134 // Call signature handler and pass locals address.
1135 //
1136 // Our signature handlers copy required arguments to the C stack
1137 // (outgoing C args), R3_ARG1 to R10_ARG8, and F1_ARG1 to
1138 // F13_ARG13.
1139 __ mr(R3_ARG1, R18_locals);
1140 #if !defined(ABI_ELFv2)
1141 __ ld(signature_handler_fd, 0, signature_handler_fd);
1142 #endif
1143 __ call_stub(signature_handler_fd);
1144 // reload method
1145 __ ld(R19_method, state_(_method));
1147 // Remove the register parameter varargs slots we allocated in
1148 // compute_interpreter_state. SP+16 ends up pointing to the ABI
1149 // outgoing argument area.
1150 //
1151 // Not needed on PPC64.
1152 //__ add(SP, SP, Argument::n_register_parameters*BytesPerWord);
1154 assert(result_handler_addr->is_nonvolatile(), "result_handler_addr must be in a non-volatile register");
1155 // Save across call to native method.
1156 __ mr(result_handler_addr, R3_RET);
1158 // Set up fixed parameters and call the native method.
1159 // If the method is static, get mirror into R4_ARG2.
1161 {
1162 Label method_is_not_static;
1163 // access_flags is non-volatile and still, no need to restore it
1165 // restore access flags
1166 __ testbitdi(CCR0, R0, access_flags, JVM_ACC_STATIC_BIT);
1167 __ bfalse(CCR0, method_is_not_static);
1169 // constants = method->constants();
1170 __ ld(R11_scratch1, in_bytes(Method::const_offset()), R19_method);
1171 __ ld(R11_scratch1/*constants*/, in_bytes(ConstMethod::constants_offset()), R11_scratch1);
1172 // pool_holder = method->constants()->pool_holder();
1173 __ ld(R11_scratch1/*pool_holder*/, ConstantPool::pool_holder_offset_in_bytes(),
1174 R11_scratch1/*constants*/);
1176 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
1178 // mirror = pool_holder->klass_part()->java_mirror();
1179 __ ld(R0/*mirror*/, mirror_offset, R11_scratch1/*pool_holder*/);
1180 // state->_native_mirror = mirror;
1181 __ std(R0/*mirror*/, state_(_oop_temp));
1182 // R4_ARG2 = &state->_oop_temp;
1183 __ addir(R4_ARG2, state_(_oop_temp));
1185 __ BIND(method_is_not_static);
1186 }
1188 // At this point, arguments have been copied off the stack into
1189 // their JNI positions. Oops are boxed in-place on the stack, with
1190 // handles copied to arguments. The result handler address is in a
1191 // register.
1193 // pass JNIEnv address as first parameter
1194 __ addir(R3_ARG1, thread_(jni_environment));
1196 // Load the native_method entry before we change the thread state.
1197 __ ld(native_method_fd, method_(native_function));
1199 //=============================================================================
1200 // Transition from _thread_in_Java to _thread_in_native. As soon as
1201 // we make this change the safepoint code needs to be certain that
1202 // the last Java frame we established is good. The pc in that frame
1203 // just needs to be near here not an actual return address.
1205 // We use release_store_fence to update values like the thread state, where
1206 // we don't want the current thread to continue until all our prior memory
1207 // accesses (including the new thread state) are visible to other threads.
1208 __ li(R0, _thread_in_native);
1209 __ release();
1211 // TODO: PPC port: assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
1212 __ stw(R0, thread_(thread_state));
1214 if (UseMembar) {
1215 __ fence();
1216 }
1218 //=============================================================================
1219 // Call the native method. Argument registers must not have been
1220 // overwritten since "__ call_stub(signature_handler);" (except for
1221 // ARG1 and ARG2 for static methods)
1222 __ call_c(native_method_fd);
1224 __ std(R3_RET, state_(_native_lresult));
1225 __ stfd(F1_RET, state_(_native_fresult));
1227 // The frame_manager_lr field, which we use for setting the last
1228 // java frame, gets overwritten by the signature handler. Restore
1229 // it now.
1230 __ get_PC_trash_LR(R11_scratch1);
1231 __ std(R11_scratch1, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1233 // Because of GC R19_method may no longer be valid.
1235 // Block, if necessary, before resuming in _thread_in_Java state.
1236 // In order for GC to work, don't clear the last_Java_sp until after
1237 // blocking.
1241 //=============================================================================
1242 // Switch thread to "native transition" state before reading the
1243 // synchronization state. This additional state is necessary
1244 // because reading and testing the synchronization state is not
1245 // atomic w.r.t. GC, as this scenario demonstrates: Java thread A,
1246 // in _thread_in_native state, loads _not_synchronized and is
1247 // preempted. VM thread changes sync state to synchronizing and
1248 // suspends threads for GC. Thread A is resumed to finish this
1249 // native method, but doesn't block here since it didn't see any
1250 // synchronization in progress, and escapes.
1252 // We use release_store_fence to update values like the thread state, where
1253 // we don't want the current thread to continue until all our prior memory
1254 // accesses (including the new thread state) are visible to other threads.
1255 __ li(R0/*thread_state*/, _thread_in_native_trans);
1256 __ release();
1257 __ stw(R0/*thread_state*/, thread_(thread_state));
1258 if (UseMembar) {
1259 __ fence();
1260 }
1261 // Write serialization page so that the VM thread can do a pseudo remote
1262 // membar. We use the current thread pointer to calculate a thread
1263 // specific offset to write to within the page. This minimizes bus
1264 // traffic due to cache line collision.
1265 else {
1266 __ serialize_memory(R16_thread, R11_scratch1, R12_scratch2);
1267 }
1269 // Now before we return to java we must look for a current safepoint
1270 // (a new safepoint can not start since we entered native_trans).
1271 // We must check here because a current safepoint could be modifying
1272 // the callers registers right this moment.
1274 // Acquire isn't strictly necessary here because of the fence, but
1275 // sync_state is declared to be volatile, so we do it anyway.
1276 __ load_const(sync_state_addr, SafepointSynchronize::address_of_state());
1278 // TODO: PPC port: assert(4 == SafepointSynchronize::sz_state(), "unexpected field size");
1279 __ lwz(sync_state, 0, sync_state_addr);
1281 // TODO: PPC port: assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
1282 __ lwz(suspend_flags, thread_(suspend_flags));
1284 __ acquire();
1286 Label sync_check_done;
1287 Label do_safepoint;
1288 // No synchronization in progress nor yet synchronized
1289 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized);
1290 // not suspended
1291 __ cmpwi(CCR1, suspend_flags, 0);
1293 __ bne(CCR0, do_safepoint);
1294 __ beq(CCR1, sync_check_done);
1295 __ bind(do_safepoint);
1296 // Block. We do the call directly and leave the current
1297 // last_Java_frame setup undisturbed. We must save any possible
1298 // native result acrosss the call. No oop is present
1300 __ mr(R3_ARG1, R16_thread);
1301 #if defined(ABI_ELFv2)
1302 __ call_c(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1303 relocInfo::none);
1304 #else
1305 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, JavaThread::check_special_condition_for_native_trans),
1306 relocInfo::none);
1307 #endif
1308 __ bind(sync_check_done);
1310 //=============================================================================
1311 // <<<<<< Back in Interpreter Frame >>>>>
1313 // We are in thread_in_native_trans here and back in the normal
1314 // interpreter frame. We don't have to do anything special about
1315 // safepoints and we can switch to Java mode anytime we are ready.
1317 // Note: frame::interpreter_frame_result has a dependency on how the
1318 // method result is saved across the call to post_method_exit. For
1319 // native methods it assumes that the non-FPU/non-void result is
1320 // saved in _native_lresult and a FPU result in _native_fresult. If
1321 // this changes then the interpreter_frame_result implementation
1322 // will need to be updated too.
1324 // On PPC64, we have stored the result directly after the native call.
1326 //=============================================================================
1327 // back in Java
1329 // We use release_store_fence to update values like the thread state, where
1330 // we don't want the current thread to continue until all our prior memory
1331 // accesses (including the new thread state) are visible to other threads.
1332 __ li(R0/*thread_state*/, _thread_in_Java);
1333 __ release();
1334 __ stw(R0/*thread_state*/, thread_(thread_state));
1335 if (UseMembar) {
1336 __ fence();
1337 }
1339 __ reset_last_Java_frame();
1341 // Reload GR27_method, call killed it. We can't look at
1342 // state->_method until we're back in java state because in java
1343 // state gc can't happen until we get to a safepoint.
1344 //
1345 // We've set thread_state to _thread_in_Java already, so restoring
1346 // R19_method from R14_state works; R19_method is invalid, because
1347 // GC may have happened.
1348 __ ld(R19_method, state_(_method)); // reload method, may have moved
1350 // jvmdi/jvmpi support. Whether we've got an exception pending or
1351 // not, and whether unlocking throws an exception or not, we notify
1352 // on native method exit. If we do have an exception, we'll end up
1353 // in the caller's context to handle it, so if we don't do the
1354 // notify here, we'll drop it on the floor.
1356 __ notify_method_exit(true/*native method*/,
1357 ilgl /*illegal state (not used for native methods)*/,
1358 InterpreterMacroAssembler::NotifyJVMTI,
1359 false /*check_exceptions*/);
1361 //=============================================================================
1362 // Handle exceptions
1364 // See if we must unlock.
1365 //
1366 {
1367 Label method_is_not_synced;
1368 // is_synced is still alive
1369 assert(is_synced->is_nonvolatile(), "is_synced must be non-volatile");
1370 __ bfalse(is_synced, method_is_not_synced);
1372 unlock_method();
1374 __ bind(method_is_not_synced);
1375 }
1377 // Reset active handles after returning from native.
1378 // thread->active_handles()->clear();
1379 __ ld(active_handles, thread_(active_handles));
1380 // JNIHandleBlock::_top is an int.
1381 // TODO: PPC port: assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
1382 __ li(R0, 0);
1383 __ stw(R0, JNIHandleBlock::top_offset_in_bytes(), active_handles);
1385 Label no_pending_exception_from_native_method;
1386 __ ld(R0/*pending_exception*/, thread_(pending_exception));
1387 __ cmpdi(CCR0, R0/*pending_exception*/, 0);
1388 __ beq(CCR0, no_pending_exception_from_native_method);
1391 //-----------------------------------------------------------------------------
1392 // An exception is pending. We call into the runtime only if the
1393 // caller was not interpreted. If it was interpreted the
1394 // interpreter will do the correct thing. If it isn't interpreted
1395 // (call stub/compiled code) we will change our return and continue.
1396 __ BIND(exception_return);
1398 Label return_to_initial_caller_with_pending_exception;
1399 __ cmpdi(CCR0, R15_prev_state, 0);
1400 __ beq(CCR0, return_to_initial_caller_with_pending_exception);
1402 // We are returning to an interpreter activation, just pop the state,
1403 // pop our frame, leave the exception pending, and return.
1404 __ pop_interpreter_state(/*prev_state_may_be_0=*/false);
1405 __ pop_interpreter_frame(R11_scratch1, R12_scratch2, R21_tmp1 /* set to return pc */, R22_tmp2);
1406 __ mtlr(R21_tmp1);
1407 __ blr();
1409 __ BIND(exception_return_sync_check);
1411 assert(is_synced->is_nonvolatile(), "is_synced must be non-volatile");
1412 __ bfalse(is_synced, exception_return);
1413 unlock_method();
1414 __ b(exception_return);
1417 __ BIND(return_to_initial_caller_with_pending_exception);
1418 // We are returning to a c2i-adapter / call-stub, get the address of the
1419 // exception handler, pop the frame and return to the handler.
1421 // First, pop to caller's frame.
1422 __ pop_interpreter_frame(R11_scratch1, R12_scratch2, R21_tmp1 /* set to return pc */, R22_tmp2);
1424 __ push_frame_reg_args(0, R11_scratch1);
1425 // Get the address of the exception handler.
1426 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1427 R16_thread,
1428 R21_tmp1 /* return pc */);
1429 __ pop_frame();
1431 // Load the PC of the the exception handler into LR.
1432 __ mtlr(R3_RET);
1434 // Load exception into R3_ARG1 and clear pending exception in thread.
1435 __ ld(R3_ARG1/*exception*/, thread_(pending_exception));
1436 __ li(R4_ARG2, 0);
1437 __ std(R4_ARG2, thread_(pending_exception));
1439 // Load the original return pc into R4_ARG2.
1440 __ mr(R4_ARG2/*issuing_pc*/, R21_tmp1);
1442 // Resize frame to get rid of a potential extension.
1443 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2);
1445 // Return to exception handler.
1446 __ blr();
1449 //-----------------------------------------------------------------------------
1450 // No exception pending.
1451 __ BIND(no_pending_exception_from_native_method);
1453 // Move native method result back into proper registers and return.
1454 // Invoke result handler (may unbox/promote).
1455 __ ld(R3_RET, state_(_native_lresult));
1456 __ lfd(F1_RET, state_(_native_fresult));
1457 __ call_stub(result_handler_addr);
1459 // We have created a new BytecodeInterpreter object, now we must destroy it.
1460 //
1461 // Restore previous R14_state and caller's SP. R15_prev_state may
1462 // be 0 here, because our caller may be the call_stub or compiled
1463 // code.
1464 __ pop_interpreter_state(/*prev_state_may_be_0=*/true);
1465 __ pop_interpreter_frame(R11_scratch1, R12_scratch2, R21_tmp1 /* set to return pc */, R22_tmp2);
1466 // Resize frame to get rid of a potential extension.
1467 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2);
1469 // Must use the return pc which was loaded from the caller's frame
1470 // as the VM uses return-pc-patching for deoptimization.
1471 __ mtlr(R21_tmp1);
1472 __ blr();
1476 //=============================================================================
1477 // We encountered an exception while computing the interpreter
1478 // state, so R14_state isn't valid. Act as if we just returned from
1479 // the callee method with a pending exception.
1480 __ BIND(stack_overflow_return);
1482 //
1483 // Register state:
1484 // R14_state invalid; trashed by compute_interpreter_state
1485 // R15_prev_state valid, but may be 0
1486 //
1487 // R1_SP valid, points to caller's SP; wasn't yet updated by
1488 // compute_interpreter_state
1489 //
1491 // Create exception oop and make it pending.
1493 // Throw the exception via RuntimeStub "throw_StackOverflowError_entry".
1494 //
1495 // Previously, we called C-Code directly. As a consequence, a
1496 // possible GC tried to process the argument oops of the top frame
1497 // (see RegisterMap::clear, which sets the corresponding flag to
1498 // true). This lead to crashes because:
1499 // 1. The top register map did not contain locations for the argument registers
1500 // 2. The arguments are dead anyway, could be already overwritten in the worst case
1501 // Solution: Call via special runtime stub that pushes it's own
1502 // frame. This runtime stub has the flag "CodeBlob::caller_must_gc_arguments()"
1503 // set to "false", what prevents the dead arguments getting GC'd.
1504 //
1505 // 2 cases exist:
1506 // 1. We were called by the c2i adapter / call stub
1507 // 2. We were called by the frame manager
1508 //
1509 // Both cases are handled by this code:
1510 // 1. - initial_caller_sp was saved in both cases on entry, so it's safe to load it back even if it was not changed.
1511 // - control flow will be:
1512 // throw_stackoverflow_stub->VM->throw_stackoverflow_stub->forward_excep->excp_blob of caller method
1513 // 2. - control flow will be:
1514 // throw_stackoverflow_stub->VM->throw_stackoverflow_stub->forward_excep->rethrow_excp_entry of frame manager->resume_method
1515 // Since we restored the caller SP above, the rethrow_excp_entry can restore the original interpreter state
1516 // registers using the stack and resume the calling method with a pending excp.
1518 // Pop any c2i extension from the stack, restore LR just to be sure
1519 __ ld(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1520 __ mtlr(R0);
1521 // Resize frame to get rid of a potential extension.
1522 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2);
1524 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "generated in wrong order");
1525 // Load target address of the runtime stub.
1526 __ load_const(R12_scratch2, (StubRoutines::throw_StackOverflowError_entry()));
1527 __ mtctr(R12_scratch2);
1528 __ bctr();
1531 //=============================================================================
1532 // Counter overflow.
1534 if (inc_counter) {
1535 // Handle invocation counter overflow
1536 __ bind(invocation_counter_overflow);
1538 generate_counter_overflow(continue_after_compile);
1539 }
1541 native_entry = entry;
1542 return entry;
1543 }
1545 bool AbstractInterpreter::can_be_compiled(methodHandle m) {
1546 // No special entry points that preclude compilation.
1547 return true;
1548 }
1550 // Unlock the current method.
1551 //
1552 void CppInterpreterGenerator::unlock_method(void) {
1553 // Find preallocated monitor and unlock method. Method monitor is
1554 // the first one.
1556 // Registers alive
1557 // R14_state
1558 //
1559 // Registers updated
1560 // volatiles
1561 //
1562 const Register monitor = R4_ARG2;
1564 // Pass address of initial monitor we allocated.
1565 //
1566 // First monitor.
1567 __ addi(monitor, R14_state, -frame::interpreter_frame_monitor_size_in_bytes());
1569 // Unlock method
1570 __ unlock_object(monitor);
1571 }
1573 // Lock the current method.
1574 //
1575 void CppInterpreterGenerator::lock_method(void) {
1576 // Find preallocated monitor and lock method. Method monitor is the
1577 // first one.
1579 //
1580 // Registers alive
1581 // R14_state
1582 //
1583 // Registers updated
1584 // volatiles
1585 //
1587 const Register monitor = R4_ARG2;
1588 const Register object = R5_ARG3;
1590 // Pass address of initial monitor we allocated.
1591 __ addi(monitor, R14_state, -frame::interpreter_frame_monitor_size_in_bytes());
1593 // Pass object address.
1594 __ ld(object, BasicObjectLock::obj_offset_in_bytes(), monitor);
1596 // Lock method.
1597 __ lock_object(monitor, object);
1598 }
1600 // Generate code for handling resuming a deopted method.
1601 void CppInterpreterGenerator::generate_deopt_handling(Register result_index) {
1603 //=============================================================================
1604 // Returning from a compiled method into a deopted method. The
1605 // bytecode at the bcp has completed. The result of the bytecode is
1606 // in the native abi (the tosca for the template based
1607 // interpreter). Any stack space that was used by the bytecode that
1608 // has completed has been removed (e.g. parameters for an invoke) so
1609 // all that we have to do is place any pending result on the
1610 // expression stack and resume execution on the next bytecode.
1612 Label return_from_deopt_common;
1614 // R3_RET and F1_RET are live here! Load the array index of the
1615 // required result stub address and continue at return_from_deopt_common.
1617 // Deopt needs to jump to here to enter the interpreter (return a result).
1618 deopt_frame_manager_return_atos = __ pc();
1619 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_OBJECT));
1620 __ b(return_from_deopt_common);
1622 deopt_frame_manager_return_btos = __ pc();
1623 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_BOOLEAN));
1624 __ b(return_from_deopt_common);
1626 deopt_frame_manager_return_itos = __ pc();
1627 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_INT));
1628 __ b(return_from_deopt_common);
1630 deopt_frame_manager_return_ltos = __ pc();
1631 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_LONG));
1632 __ b(return_from_deopt_common);
1634 deopt_frame_manager_return_ftos = __ pc();
1635 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_FLOAT));
1636 __ b(return_from_deopt_common);
1638 deopt_frame_manager_return_dtos = __ pc();
1639 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_DOUBLE));
1640 __ b(return_from_deopt_common);
1642 deopt_frame_manager_return_vtos = __ pc();
1643 __ li(result_index, AbstractInterpreter::BasicType_as_index(T_VOID));
1644 // Last one, fall-through to return_from_deopt_common.
1646 // Deopt return common. An index is present that lets us move any
1647 // possible result being return to the interpreter's stack.
1648 //
1649 __ BIND(return_from_deopt_common);
1651 }
1653 // Generate the code to handle a more_monitors message from the c++ interpreter.
1654 void CppInterpreterGenerator::generate_more_monitors() {
1656 //
1657 // Registers alive
1658 // R16_thread - JavaThread*
1659 // R15_prev_state - previous BytecodeInterpreter or 0
1660 // R14_state - BytecodeInterpreter* address of receiver's interpreter state
1661 // R1_SP - old stack pointer
1662 //
1663 // Registers updated
1664 // R1_SP - new stack pointer
1665 //
1667 // Very-local scratch registers.
1668 const Register old_tos = R21_tmp1;
1669 const Register new_tos = R22_tmp2;
1670 const Register stack_base = R23_tmp3;
1671 const Register stack_limit = R24_tmp4;
1672 const Register slot = R25_tmp5;
1673 const Register n_slots = R25_tmp5;
1675 // Interpreter state fields.
1676 const Register msg = R24_tmp4;
1678 // Load up relevant interpreter state.
1680 __ ld(stack_base, state_(_stack_base)); // Old stack_base
1681 __ ld(old_tos, state_(_stack)); // Old tos
1682 __ ld(stack_limit, state_(_stack_limit)); // Old stack_limit
1684 // extracted monitor_size
1685 int monitor_size = frame::interpreter_frame_monitor_size_in_bytes();
1686 assert(Assembler::is_aligned((unsigned int)monitor_size,
1687 (unsigned int)frame::alignment_in_bytes),
1688 "size of a monitor must respect alignment of SP");
1690 // Save and restore top LR
1691 __ ld(R12_scratch2, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1692 __ resize_frame(-monitor_size, R11_scratch1);// Allocate space for new monitor
1693 __ std(R12_scratch2, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1694 // Initial_caller_sp is used as unextended_sp for non initial callers.
1695 __ std(R1_SP, _top_ijava_frame_abi(initial_caller_sp), R1_SP);
1696 __ addi(stack_base, stack_base, -monitor_size); // New stack_base
1697 __ addi(new_tos, old_tos, -monitor_size); // New tos
1698 __ addi(stack_limit, stack_limit, -monitor_size); // New stack_limit
1700 __ std(R1_SP, state_(_last_Java_sp)); // Update frame_bottom
1702 __ std(stack_base, state_(_stack_base)); // Update stack_base
1703 __ std(new_tos, state_(_stack)); // Update tos
1704 __ std(stack_limit, state_(_stack_limit)); // Update stack_limit
1706 __ li(msg, BytecodeInterpreter::got_monitors); // Tell interpreter we allocated the lock
1707 __ stw(msg, state_(_msg));
1709 // Shuffle expression stack down. Recall that stack_base points
1710 // just above the new expression stack bottom. Old_tos and new_tos
1711 // are used to scan thru the old and new expression stacks.
1713 Label copy_slot, copy_slot_finished;
1714 __ sub(n_slots, stack_base, new_tos);
1715 __ srdi_(n_slots, n_slots, LogBytesPerWord); // compute number of slots to copy
1716 assert(LogBytesPerWord == 3, "conflicts assembler instructions");
1717 __ beq(CCR0, copy_slot_finished); // nothing to copy
1719 __ mtctr(n_slots);
1721 // loop
1722 __ bind(copy_slot);
1723 __ ldu(slot, BytesPerWord, old_tos); // slot = *++old_tos;
1724 __ stdu(slot, BytesPerWord, new_tos); // *++new_tos = slot;
1725 __ bdnz(copy_slot);
1727 __ bind(copy_slot_finished);
1729 // Restart interpreter
1730 __ li(R0, 0);
1731 __ std(R0, BasicObjectLock::obj_offset_in_bytes(), stack_base); // Mark lock as unused
1732 }
1734 address CppInterpreterGenerator::generate_normal_entry(void) {
1735 if (interpreter_frame_manager != NULL) return interpreter_frame_manager;
1737 address entry = __ pc();
1739 address return_from_native_pc = (address) NULL;
1741 // Initial entry to frame manager (from call_stub or c2i_adapter)
1743 //
1744 // Registers alive
1745 // R16_thread - JavaThread*
1746 // R19_method - callee's Method (method to be invoked)
1747 // R17_tos - address of sender tos (prepushed)
1748 // R1_SP - SP prepared by call stub such that caller's outgoing args are near top
1749 // LR - return address to caller (call_stub or c2i_adapter)
1750 // R21_sender_SP - initial caller sp
1751 //
1752 // Registers updated
1753 // R15_prev_state - 0
1754 //
1755 // Stack layout at this point:
1756 //
1757 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP
1758 // alignment (optional)
1759 // [outgoing Java arguments] <-- R17_tos
1760 // ...
1761 // PARENT [PARENT_IJAVA_FRAME_ABI]
1762 // ...
1763 //
1765 // Save initial_caller_sp to caller's abi.
1766 // The caller frame must be resized before returning to get rid of
1767 // the c2i part on top of the calling compiled frame (if any).
1768 // R21_tmp1 must match sender_sp in gen_c2i_adapter.
1769 // Now override the saved SP with the senderSP so we can pop c2i
1770 // arguments (if any) off when we return.
1771 __ std(R21_sender_SP, _top_ijava_frame_abi(initial_caller_sp), R1_SP);
1773 // Save LR to caller's frame. We don't use _abi(lr) here,
1774 // because it is not safe.
1775 __ mflr(R0);
1776 __ std(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1778 // If we come here, it is the first invocation of the frame manager.
1779 // So there is no previous interpreter state.
1780 __ li(R15_prev_state, 0);
1783 // Fall through to where "recursive" invocations go.
1785 //=============================================================================
1786 // Dispatch an instance of the interpreter. Recursive activations
1787 // come here.
1789 Label re_dispatch;
1790 __ BIND(re_dispatch);
1792 //
1793 // Registers alive
1794 // R16_thread - JavaThread*
1795 // R19_method - callee's Method
1796 // R17_tos - address of caller's tos (prepushed)
1797 // R15_prev_state - address of caller's BytecodeInterpreter or 0
1798 // R1_SP - caller's SP trimmed such that caller's outgoing args are near top.
1799 //
1800 // Stack layout at this point:
1801 //
1802 // 0 [TOP_IJAVA_FRAME_ABI]
1803 // alignment (optional)
1804 // [outgoing Java arguments]
1805 // ...
1806 // PARENT [PARENT_IJAVA_FRAME_ABI]
1807 // ...
1809 // fall through to interpreted execution
1811 //=============================================================================
1812 // Allocate a new Java frame and initialize the new interpreter state.
1814 Label stack_overflow_return;
1816 // Create a suitable new Java frame plus a new BytecodeInterpreter instance
1817 // in the current (frame manager's) C frame.
1818 generate_compute_interpreter_state(stack_overflow_return);
1820 // fall through
1822 //=============================================================================
1823 // Interpreter dispatch.
1825 Label call_interpreter;
1826 __ BIND(call_interpreter);
1828 //
1829 // Registers alive
1830 // R16_thread - JavaThread*
1831 // R15_prev_state - previous BytecodeInterpreter or 0
1832 // R14_state - address of receiver's BytecodeInterpreter
1833 // R1_SP - receiver's stack pointer
1834 //
1836 // Thread fields.
1837 const Register pending_exception = R21_tmp1;
1839 // Interpreter state fields.
1840 const Register msg = R24_tmp4;
1842 // Method fields.
1843 const Register parameter_count = R25_tmp5;
1844 const Register result_index = R26_tmp6;
1846 const Register dummy = R28_tmp8;
1848 // Address of various interpreter stubs.
1849 // R29_tmp9 is reserved.
1850 const Register stub_addr = R27_tmp7;
1852 // Uncommon trap needs to jump to here to enter the interpreter
1853 // (re-execute current bytecode).
1854 unctrap_frame_manager_entry = __ pc();
1856 // If we are profiling, store our fp (BSP) in the thread so we can
1857 // find it during a tick.
1858 if (Arguments::has_profile()) {
1859 // On PPC64 we store the pointer to the current BytecodeInterpreter,
1860 // instead of the bsp of ia64. This should suffice to be able to
1861 // find all interesting information.
1862 __ std(R14_state, thread_(last_interpreter_fp));
1863 }
1865 // R16_thread, R14_state and R15_prev_state are nonvolatile
1866 // registers. There is no need to save these. If we needed to save
1867 // some state in the current Java frame, this could be a place to do
1868 // so.
1870 // Call Java bytecode dispatcher passing "BytecodeInterpreter* istate".
1871 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
1872 JvmtiExport::can_post_interpreter_events()
1873 ? BytecodeInterpreter::runWithChecks
1874 : BytecodeInterpreter::run),
1875 R14_state);
1877 interpreter_return_address = __ last_calls_return_pc();
1879 // R16_thread, R14_state and R15_prev_state have their values preserved.
1881 // If we are profiling, clear the fp in the thread to tell
1882 // the profiler that we are no longer in the interpreter.
1883 if (Arguments::has_profile()) {
1884 __ li(R11_scratch1, 0);
1885 __ std(R11_scratch1, thread_(last_interpreter_fp));
1886 }
1888 // Load message from bytecode dispatcher.
1889 // TODO: PPC port: guarantee(4 == BytecodeInterpreter::sz_msg(), "unexpected field size");
1890 __ lwz(msg, state_(_msg));
1893 Label more_monitors;
1894 Label return_from_native;
1895 Label return_from_native_common;
1896 Label return_from_native_no_exception;
1897 Label return_from_interpreted_method;
1898 Label return_from_recursive_activation;
1899 Label unwind_recursive_activation;
1900 Label resume_interpreter;
1901 Label return_to_initial_caller;
1902 Label unwind_initial_activation;
1903 Label unwind_initial_activation_pending_exception;
1904 Label call_method;
1905 Label call_special;
1906 Label retry_method;
1907 Label retry_method_osr;
1908 Label popping_frame;
1909 Label throwing_exception;
1911 // Branch according to the received message
1913 __ cmpwi(CCR1, msg, BytecodeInterpreter::call_method);
1914 __ cmpwi(CCR2, msg, BytecodeInterpreter::return_from_method);
1916 __ beq(CCR1, call_method);
1917 __ beq(CCR2, return_from_interpreted_method);
1919 __ cmpwi(CCR3, msg, BytecodeInterpreter::more_monitors);
1920 __ cmpwi(CCR4, msg, BytecodeInterpreter::throwing_exception);
1922 __ beq(CCR3, more_monitors);
1923 __ beq(CCR4, throwing_exception);
1925 __ cmpwi(CCR5, msg, BytecodeInterpreter::popping_frame);
1926 __ cmpwi(CCR6, msg, BytecodeInterpreter::do_osr);
1928 __ beq(CCR5, popping_frame);
1929 __ beq(CCR6, retry_method_osr);
1931 __ stop("bad message from interpreter");
1934 //=============================================================================
1935 // Add a monitor just below the existing one(s). State->_stack_base
1936 // points to the lowest existing one, so we insert the new one just
1937 // below it and shuffle the expression stack down. Ref. the above
1938 // stack layout picture, we must update _stack_base, _stack, _stack_limit
1939 // and _last_Java_sp in the interpreter state.
1941 __ BIND(more_monitors);
1943 generate_more_monitors();
1944 __ b(call_interpreter);
1946 generate_deopt_handling(result_index);
1948 // Restoring the R14_state is already done by the deopt_blob.
1950 // Current tos includes no parameter slots.
1951 __ ld(R17_tos, state_(_stack));
1952 __ li(msg, BytecodeInterpreter::deopt_resume);
1953 __ b(return_from_native_common);
1955 // We are sent here when we are unwinding from a native method or
1956 // adapter with an exception pending. We need to notify the interpreter
1957 // that there is an exception to process.
1958 // We arrive here also if the frame manager called an (interpreted) target
1959 // which returns with a StackOverflow exception.
1960 // The control flow is in this case is:
1961 // frame_manager->throw_excp_stub->forward_excp->rethrow_excp_entry
1963 AbstractInterpreter::_rethrow_exception_entry = __ pc();
1965 // Restore R14_state.
1966 __ ld(R14_state, 0, R1_SP);
1967 __ addi(R14_state, R14_state,
1968 -frame::interpreter_frame_cinterpreterstate_size_in_bytes());
1970 // Store exception oop into thread object.
1971 __ std(R3_RET, thread_(pending_exception));
1972 __ li(msg, BytecodeInterpreter::method_resume /*rethrow_exception*/);
1973 //
1974 // NOTE: the interpreter frame as setup be deopt does NOT include
1975 // any parameter slots (good thing since we have no callee here
1976 // and couldn't remove them) so we don't have to do any calculations
1977 // here to figure it out.
1978 //
1979 __ ld(R17_tos, state_(_stack));
1980 __ b(return_from_native_common);
1983 //=============================================================================
1984 // Returning from a native method. Result is in the native abi
1985 // location so we must move it to the java expression stack.
1987 __ BIND(return_from_native);
1988 guarantee(return_from_native_pc == (address) NULL, "precondition");
1989 return_from_native_pc = __ pc();
1991 // Restore R14_state.
1992 __ ld(R14_state, 0, R1_SP);
1993 __ addi(R14_state, R14_state, -frame::interpreter_frame_cinterpreterstate_size_in_bytes());
1995 //
1996 // Registers alive
1997 // R16_thread
1998 // R14_state - address of caller's BytecodeInterpreter.
1999 // R3_RET - integer result, if any.
2000 // F1_RET - float result, if any.
2001 //
2002 // Registers updated
2003 // R19_method - callee's Method
2004 // R17_tos - caller's tos, with outgoing args popped
2005 // result_index - index of result handler.
2006 // msg - message for resuming interpreter.
2007 //
2009 // Very-local scratch registers.
2011 const ConditionRegister have_pending_exception = CCR0;
2013 // Load callee Method, gc may have moved it.
2014 __ ld(R19_method, state_(_result._to_call._callee));
2016 // Load address of caller's tos. includes parameter slots.
2017 __ ld(R17_tos, state_(_stack));
2019 // Pop callee's parameters.
2021 __ ld(parameter_count, in_bytes(Method::const_offset()), R19_method);
2022 __ lhz(parameter_count, in_bytes(ConstMethod::size_of_parameters_offset()), parameter_count);
2023 __ sldi(parameter_count, parameter_count, Interpreter::logStackElementSize);
2024 __ add(R17_tos, R17_tos, parameter_count);
2026 // Result stub address array index
2027 // TODO: PPC port: assert(4 == sizeof(AccessFlags), "unexpected field size");
2028 __ lwa(result_index, method_(result_index));
2030 __ li(msg, BytecodeInterpreter::method_resume);
2032 //
2033 // Registers alive
2034 // R16_thread
2035 // R14_state - address of caller's BytecodeInterpreter.
2036 // R17_tos - address of caller's tos with outgoing args already popped
2037 // R3_RET - integer return value, if any.
2038 // F1_RET - float return value, if any.
2039 // result_index - index of result handler.
2040 // msg - message for resuming interpreter.
2041 //
2042 // Registers updated
2043 // R3_RET - new address of caller's tos, including result, if any
2044 //
2046 __ BIND(return_from_native_common);
2048 // Check for pending exception
2049 __ ld(pending_exception, thread_(pending_exception));
2050 __ cmpdi(CCR0, pending_exception, 0);
2051 __ beq(CCR0, return_from_native_no_exception);
2053 // If there's a pending exception, we really have no result, so
2054 // R3_RET is dead. Resume_interpreter assumes the new tos is in
2055 // R3_RET.
2056 __ mr(R3_RET, R17_tos);
2057 // `resume_interpreter' expects R15_prev_state to be alive.
2058 __ ld(R15_prev_state, state_(_prev_link));
2059 __ b(resume_interpreter);
2061 __ BIND(return_from_native_no_exception);
2063 // No pending exception, copy method result from native ABI register
2064 // to tos.
2066 // Address of stub descriptor address array.
2067 __ load_const(stub_addr, CppInterpreter::tosca_result_to_stack());
2069 // Pass address of tos to stub.
2070 __ mr(R4_ARG2, R17_tos);
2072 // Address of stub descriptor address.
2073 __ sldi(result_index, result_index, LogBytesPerWord);
2074 __ add(stub_addr, stub_addr, result_index);
2076 // Stub descriptor address.
2077 __ ld(stub_addr, 0, stub_addr);
2079 // TODO: don't do this via a call, do it in place!
2080 //
2081 // call stub via descriptor
2082 // in R3_ARG1/F1_ARG1: result value (R3_RET or F1_RET)
2083 __ call_stub(stub_addr);
2085 // new tos = result of call in R3_RET
2087 // `resume_interpreter' expects R15_prev_state to be alive.
2088 __ ld(R15_prev_state, state_(_prev_link));
2089 __ b(resume_interpreter);
2091 //=============================================================================
2092 // We encountered an exception while computing the interpreter
2093 // state, so R14_state isn't valid. Act as if we just returned from
2094 // the callee method with a pending exception.
2095 __ BIND(stack_overflow_return);
2097 //
2098 // Registers alive
2099 // R16_thread - JavaThread*
2100 // R1_SP - old stack pointer
2101 // R19_method - callee's Method
2102 // R17_tos - address of caller's tos (prepushed)
2103 // R15_prev_state - address of caller's BytecodeInterpreter or 0
2104 // R18_locals - address of callee's locals array
2105 //
2106 // Registers updated
2107 // R3_RET - address of resuming tos, if recursive unwind
2109 Label Lskip_unextend_SP;
2111 {
2112 const ConditionRegister is_initial_call = CCR0;
2113 const Register tos_save = R21_tmp1;
2114 const Register tmp = R22_tmp2;
2116 assert(tos_save->is_nonvolatile(), "need a nonvolatile");
2118 // Is the exception thrown in the initial Java frame of this frame
2119 // manager frame?
2120 __ cmpdi(is_initial_call, R15_prev_state, 0);
2121 __ bne(is_initial_call, Lskip_unextend_SP);
2123 // Pop any c2i extension from the stack. This is necessary in the
2124 // non-recursive case (that is we were called by the c2i adapter,
2125 // meaning we have to prev state). In this case we entered the frame
2126 // manager through a special entry which pushes the orignal
2127 // unextended SP to the stack. Here we load it back.
2128 __ ld(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
2129 __ mtlr(R0);
2130 // Resize frame to get rid of a potential extension.
2131 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2);
2133 // Fall through
2135 __ bind(Lskip_unextend_SP);
2137 // Throw the exception via RuntimeStub "throw_StackOverflowError_entry".
2138 //
2139 // Previously, we called C-Code directly. As a consequence, a
2140 // possible GC tried to process the argument oops of the top frame
2141 // (see RegisterMap::clear, which sets the corresponding flag to
2142 // true). This lead to crashes because:
2143 // 1. The top register map did not contain locations for the argument registers
2144 // 2. The arguments are dead anyway, could be already overwritten in the worst case
2145 // Solution: Call via special runtime stub that pushes it's own frame. This runtime stub has the flag
2146 // "CodeBlob::caller_must_gc_arguments()" set to "false", what prevents the dead arguments getting GC'd.
2147 //
2148 // 2 cases exist:
2149 // 1. We were called by the c2i adapter / call stub
2150 // 2. We were called by the frame manager
2151 //
2152 // Both cases are handled by this code:
2153 // 1. - initial_caller_sp was saved on stack => Load it back and we're ok
2154 // - control flow will be:
2155 // throw_stackoverflow_stub->VM->throw_stackoverflow_stub->forward_excep->excp_blob of calling method
2156 // 2. - control flow will be:
2157 // throw_stackoverflow_stub->VM->throw_stackoverflow_stub->forward_excep->
2158 // ->rethrow_excp_entry of frame manager->resume_method
2159 // Since we restored the caller SP above, the rethrow_excp_entry can restore the original interpreter state
2160 // registers using the stack and resume the calling method with a pending excp.
2162 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "generated in wrong order");
2163 __ load_const(R3_ARG1, (StubRoutines::throw_StackOverflowError_entry()));
2164 __ mtctr(R3_ARG1);
2165 __ bctr();
2166 }
2167 //=============================================================================
2168 // We have popped a frame from an interpreted call. We are assured
2169 // of returning to an interpreted call by the popframe abi. We have
2170 // no return value all we have to do is pop the current frame and
2171 // then make sure that the top of stack (of the caller) gets set to
2172 // where it was when we entered the callee (i.e. the args are still
2173 // in place). Or we are returning to the interpreter. In the first
2174 // case we must extract result (if any) from the java expression
2175 // stack and store it in the location the native abi would expect
2176 // for a call returning this type. In the second case we must simply
2177 // do a stack to stack move as we unwind.
2179 __ BIND(popping_frame);
2181 // Registers alive
2182 // R14_state
2183 // R15_prev_state
2184 // R17_tos
2185 //
2186 // Registers updated
2187 // R19_method
2188 // R3_RET
2189 // msg
2190 {
2191 Label L;
2193 // Reload callee method, gc may have moved it.
2194 __ ld(R19_method, state_(_method));
2196 // We may be returning to a deoptimized frame in which case the
2197 // usual assumption of a recursive return is not true.
2199 // not equal = is recursive call
2200 __ cmpdi(CCR0, R15_prev_state, 0);
2202 __ bne(CCR0, L);
2204 // Pop_frame capability.
2205 // The pop_frame api says that the underlying frame is a Java frame, in this case
2206 // (prev_state==null) it must be a compiled frame:
2207 //
2208 // Stack at this point: I, C2I + C, ...
2209 //
2210 // The outgoing arguments of the call have just been copied (popframe_preserve_args).
2211 // By the pop_frame api, we must end up in an interpreted frame. So the compiled frame
2212 // will be deoptimized. Deoptimization will restore the outgoing arguments from
2213 // popframe_preserve_args, adjust the tos such that it includes the popframe_preserve_args,
2214 // and adjust the bci such that the call will be executed again.
2215 // We have no results, just pop the interpreter frame, resize the compiled frame to get rid
2216 // of the c2i extension and return to the deopt_handler.
2217 __ b(unwind_initial_activation);
2219 // is recursive call
2220 __ bind(L);
2222 // Resume_interpreter expects the original tos in R3_RET.
2223 __ ld(R3_RET, prev_state_(_stack));
2225 // We're done.
2226 __ li(msg, BytecodeInterpreter::popping_frame);
2228 __ b(unwind_recursive_activation);
2229 }
2232 //=============================================================================
2234 // We have finished an interpreted call. We are either returning to
2235 // native (call_stub/c2) or we are returning to the interpreter.
2236 // When returning to native, we must extract the result (if any)
2237 // from the java expression stack and store it in the location the
2238 // native abi expects. When returning to the interpreter we must
2239 // simply do a stack to stack move as we unwind.
2241 __ BIND(return_from_interpreted_method);
2243 //
2244 // Registers alive
2245 // R16_thread - JavaThread*
2246 // R15_prev_state - address of caller's BytecodeInterpreter or 0
2247 // R14_state - address of callee's interpreter state
2248 // R1_SP - callee's stack pointer
2249 //
2250 // Registers updated
2251 // R19_method - callee's method
2252 // R3_RET - address of result (new caller's tos),
2253 //
2254 // if returning to interpreted
2255 // msg - message for interpreter,
2256 // if returning to interpreted
2257 //
2259 // Check if this is the initial invocation of the frame manager.
2260 // If so, R15_prev_state will be null.
2261 __ cmpdi(CCR0, R15_prev_state, 0);
2263 // Reload callee method, gc may have moved it.
2264 __ ld(R19_method, state_(_method));
2266 // Load the method's result type.
2267 __ lwz(result_index, method_(result_index));
2269 // Go to return_to_initial_caller if R15_prev_state is null.
2270 __ beq(CCR0, return_to_initial_caller);
2272 // Copy callee's result to caller's expression stack via inline stack-to-stack
2273 // converters.
2274 {
2275 Register new_tos = R3_RET;
2276 Register from_temp = R4_ARG2;
2277 Register from = R5_ARG3;
2278 Register tos = R6_ARG4;
2279 Register tmp1 = R7_ARG5;
2280 Register tmp2 = R8_ARG6;
2282 ConditionRegister result_type_is_void = CCR1;
2283 ConditionRegister result_type_is_long = CCR2;
2284 ConditionRegister result_type_is_double = CCR3;
2286 Label stack_to_stack_void;
2287 Label stack_to_stack_double_slot; // T_LONG, T_DOUBLE
2288 Label stack_to_stack_single_slot; // T_BOOLEAN, T_BYTE, T_CHAR, T_SHORT, T_INT, T_FLOAT, T_OBJECT
2289 Label stack_to_stack_done;
2291 // Pass callee's address of tos + BytesPerWord
2292 __ ld(from_temp, state_(_stack));
2294 // result type: void
2295 __ cmpwi(result_type_is_void, result_index, AbstractInterpreter::BasicType_as_index(T_VOID));
2297 // Pass caller's tos == callee's locals address
2298 __ ld(tos, state_(_locals));
2300 // result type: long
2301 __ cmpwi(result_type_is_long, result_index, AbstractInterpreter::BasicType_as_index(T_LONG));
2303 __ addi(from, from_temp, Interpreter::stackElementSize);
2305 // !! don't branch above this line !!
2307 // handle void
2308 __ beq(result_type_is_void, stack_to_stack_void);
2310 // result type: double
2311 __ cmpwi(result_type_is_double, result_index, AbstractInterpreter::BasicType_as_index(T_DOUBLE));
2313 // handle long or double
2314 __ beq(result_type_is_long, stack_to_stack_double_slot);
2315 __ beq(result_type_is_double, stack_to_stack_double_slot);
2317 // fall through to single slot types (incl. object)
2319 {
2320 __ BIND(stack_to_stack_single_slot);
2321 // T_BOOLEAN, T_BYTE, T_CHAR, T_SHORT, T_INT, T_FLOAT, T_OBJECT
2323 __ ld(tmp1, 0, from);
2324 __ std(tmp1, 0, tos);
2325 // New expression stack top
2326 __ addi(new_tos, tos, - BytesPerWord);
2328 __ b(stack_to_stack_done);
2329 }
2331 {
2332 __ BIND(stack_to_stack_double_slot);
2333 // T_LONG, T_DOUBLE
2335 // Move both entries for debug purposes even though only one is live
2336 __ ld(tmp1, BytesPerWord, from);
2337 __ ld(tmp2, 0, from);
2338 __ std(tmp1, 0, tos);
2339 __ std(tmp2, -BytesPerWord, tos);
2341 // new expression stack top
2342 __ addi(new_tos, tos, - 2 * BytesPerWord); // two slots
2343 __ b(stack_to_stack_done);
2344 }
2346 {
2347 __ BIND(stack_to_stack_void);
2348 // T_VOID
2350 // new expression stack top
2351 __ mr(new_tos, tos);
2352 // fall through to stack_to_stack_done
2353 }
2355 __ BIND(stack_to_stack_done);
2356 }
2358 // new tos = R3_RET
2360 // Get the message for the interpreter
2361 __ li(msg, BytecodeInterpreter::method_resume);
2363 // And fall thru
2366 //=============================================================================
2367 // Restore caller's interpreter state and pass pointer to caller's
2368 // new tos to caller.
2370 __ BIND(unwind_recursive_activation);
2372 //
2373 // Registers alive
2374 // R15_prev_state - address of caller's BytecodeInterpreter
2375 // R3_RET - address of caller's tos
2376 // msg - message for caller's BytecodeInterpreter
2377 // R1_SP - callee's stack pointer
2378 //
2379 // Registers updated
2380 // R14_state - address of caller's BytecodeInterpreter
2381 // R15_prev_state - address of its parent or 0
2382 //
2384 // Pop callee's interpreter and set R14_state to caller's interpreter.
2385 __ pop_interpreter_state(/*prev_state_may_be_0=*/false);
2387 // And fall thru
2390 //=============================================================================
2391 // Resume the (calling) interpreter after a call.
2393 __ BIND(resume_interpreter);
2395 //
2396 // Registers alive
2397 // R14_state - address of resuming BytecodeInterpreter
2398 // R15_prev_state - address of its parent or 0
2399 // R3_RET - address of resuming tos
2400 // msg - message for resuming interpreter
2401 // R1_SP - callee's stack pointer
2402 //
2403 // Registers updated
2404 // R1_SP - caller's stack pointer
2405 //
2407 // Restore C stack pointer of caller (resuming interpreter),
2408 // R14_state already points to the resuming BytecodeInterpreter.
2409 __ pop_interpreter_frame_to_state(R14_state, R21_tmp1, R11_scratch1, R12_scratch2);
2411 // Store new address of tos (holding return value) in interpreter state.
2412 __ std(R3_RET, state_(_stack));
2414 // Store message for interpreter.
2415 __ stw(msg, state_(_msg));
2417 __ b(call_interpreter);
2419 //=============================================================================
2420 // Interpreter returning to native code (call_stub/c1/c2) from
2421 // initial activation. Convert stack result and unwind activation.
2423 __ BIND(return_to_initial_caller);
2425 //
2426 // Registers alive
2427 // R19_method - callee's Method
2428 // R14_state - address of callee's interpreter state
2429 // R16_thread - JavaThread
2430 // R1_SP - callee's stack pointer
2431 //
2432 // Registers updated
2433 // R3_RET/F1_RET - result in expected output register
2434 //
2436 // If we have an exception pending we have no result and we
2437 // must figure out where to really return to.
2438 //
2439 __ ld(pending_exception, thread_(pending_exception));
2440 __ cmpdi(CCR0, pending_exception, 0);
2441 __ bne(CCR0, unwind_initial_activation_pending_exception);
2443 __ lwa(result_index, method_(result_index));
2445 // Address of stub descriptor address array.
2446 __ load_const(stub_addr, CppInterpreter::stack_result_to_native());
2448 // Pass address of callee's tos + BytesPerWord.
2449 // Will then point directly to result.
2450 __ ld(R3_ARG1, state_(_stack));
2451 __ addi(R3_ARG1, R3_ARG1, Interpreter::stackElementSize);
2453 // Address of stub descriptor address
2454 __ sldi(result_index, result_index, LogBytesPerWord);
2455 __ add(stub_addr, stub_addr, result_index);
2457 // Stub descriptor address
2458 __ ld(stub_addr, 0, stub_addr);
2460 // TODO: don't do this via a call, do it in place!
2461 //
2462 // call stub via descriptor
2463 __ call_stub(stub_addr);
2465 __ BIND(unwind_initial_activation);
2467 // Unwind from initial activation. No exception is pending.
2469 //
2470 // Stack layout at this point:
2471 //
2472 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP
2473 // ...
2474 // CALLER [PARENT_IJAVA_FRAME_ABI]
2475 // ...
2476 // CALLER [unextended ABI]
2477 // ...
2478 //
2479 // The CALLER frame has a C2I adapter or is an entry-frame.
2480 //
2482 // An interpreter frame exists, we may pop the TOP_IJAVA_FRAME and
2483 // turn the caller's PARENT_IJAVA_FRAME back into a TOP_IJAVA_FRAME.
2484 // But, we simply restore the return pc from the caller's frame and
2485 // use the caller's initial_caller_sp as the new SP which pops the
2486 // interpreter frame and "resizes" the caller's frame to its "unextended"
2487 // size.
2489 // get rid of top frame
2490 __ pop_frame();
2492 // Load return PC from parent frame.
2493 __ ld(R21_tmp1, _parent_ijava_frame_abi(lr), R1_SP);
2495 // Resize frame to get rid of a potential extension.
2496 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2);
2498 // update LR
2499 __ mtlr(R21_tmp1);
2501 // return
2502 __ blr();
2504 //=============================================================================
2505 // Unwind from initial activation. An exception is pending
2507 __ BIND(unwind_initial_activation_pending_exception);
2509 //
2510 // Stack layout at this point:
2511 //
2512 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP
2513 // ...
2514 // CALLER [PARENT_IJAVA_FRAME_ABI]
2515 // ...
2516 // CALLER [unextended ABI]
2517 // ...
2518 //
2519 // The CALLER frame has a C2I adapter or is an entry-frame.
2520 //
2522 // An interpreter frame exists, we may pop the TOP_IJAVA_FRAME and
2523 // turn the caller's PARENT_IJAVA_FRAME back into a TOP_IJAVA_FRAME.
2524 // But, we just pop the current TOP_IJAVA_FRAME and fall through
2526 __ pop_frame();
2527 __ ld(R3_ARG1, _top_ijava_frame_abi(lr), R1_SP);
2529 //
2530 // Stack layout at this point:
2531 //
2532 // CALLER [PARENT_IJAVA_FRAME_ABI] <-- R1_SP
2533 // ...
2534 // CALLER [unextended ABI]
2535 // ...
2536 //
2537 // The CALLER frame has a C2I adapter or is an entry-frame.
2538 //
2539 // Registers alive
2540 // R16_thread
2541 // R3_ARG1 - return address to caller
2542 //
2543 // Registers updated
2544 // R3_ARG1 - address of pending exception
2545 // R4_ARG2 - issuing pc = return address to caller
2546 // LR - address of exception handler stub
2547 //
2549 // Resize frame to get rid of a potential extension.
2550 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2);
2552 __ mr(R14, R3_ARG1); // R14 := ARG1
2553 __ mr(R4_ARG2, R3_ARG1); // ARG2 := ARG1
2555 // Find the address of the "catch_exception" stub.
2556 __ push_frame_reg_args(0, R11_scratch1);
2557 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
2558 R16_thread,
2559 R4_ARG2);
2560 __ pop_frame();
2562 // Load continuation address into LR.
2563 __ mtlr(R3_RET);
2565 // Load address of pending exception and clear it in thread object.
2566 __ ld(R3_ARG1/*R3_RET*/, thread_(pending_exception));
2567 __ li(R4_ARG2, 0);
2568 __ std(R4_ARG2, thread_(pending_exception));
2570 // re-load issuing pc
2571 __ mr(R4_ARG2, R14);
2573 // Branch to found exception handler.
2574 __ blr();
2576 //=============================================================================
2577 // Call a new method. Compute new args and trim the expression stack
2578 // to only what we are currently using and then recurse.
2580 __ BIND(call_method);
2582 //
2583 // Registers alive
2584 // R16_thread
2585 // R14_state - address of caller's BytecodeInterpreter
2586 // R1_SP - caller's stack pointer
2587 //
2588 // Registers updated
2589 // R15_prev_state - address of caller's BytecodeInterpreter
2590 // R17_tos - address of caller's tos
2591 // R19_method - callee's Method
2592 // R1_SP - trimmed back
2593 //
2595 // Very-local scratch registers.
2597 const Register offset = R21_tmp1;
2598 const Register tmp = R22_tmp2;
2599 const Register self_entry = R23_tmp3;
2600 const Register stub_entry = R24_tmp4;
2602 const ConditionRegister cr = CCR0;
2604 // Load the address of the frame manager.
2605 __ load_const(self_entry, &interpreter_frame_manager);
2606 __ ld(self_entry, 0, self_entry);
2608 // Load BytecodeInterpreter._result._to_call._callee (callee's Method).
2609 __ ld(R19_method, state_(_result._to_call._callee));
2610 // Load BytecodeInterpreter._stack (outgoing tos).
2611 __ ld(R17_tos, state_(_stack));
2613 // Save address of caller's BytecodeInterpreter.
2614 __ mr(R15_prev_state, R14_state);
2616 // Load the callee's entry point.
2617 // Load BytecodeInterpreter._result._to_call._callee_entry_point.
2618 __ ld(stub_entry, state_(_result._to_call._callee_entry_point));
2620 // Check whether stub_entry is equal to self_entry.
2621 __ cmpd(cr, self_entry, stub_entry);
2622 // if (self_entry == stub_entry)
2623 // do a re-dispatch
2624 __ beq(cr, re_dispatch);
2625 // else
2626 // call the specialized entry (adapter for jni or compiled code)
2627 __ BIND(call_special);
2629 //
2630 // Call the entry generated by `InterpreterGenerator::generate_native_entry'.
2631 //
2632 // Registers alive
2633 // R16_thread
2634 // R15_prev_state - address of caller's BytecodeInterpreter
2635 // R19_method - callee's Method
2636 // R17_tos - address of caller's tos
2637 // R1_SP - caller's stack pointer
2638 //
2640 // Mark return from specialized entry for generate_native_entry.
2641 guarantee(return_from_native_pc != (address) NULL, "precondition");
2642 frame_manager_specialized_return = return_from_native_pc;
2644 // Set sender_SP in case we call interpreter native wrapper which
2645 // will expect it. Compiled code should not care.
2646 __ mr(R21_sender_SP, R1_SP);
2648 // Do a tail call here, and let the link register point to
2649 // frame_manager_specialized_return which is return_from_native_pc.
2650 __ load_const(tmp, frame_manager_specialized_return);
2651 __ call_stub_and_return_to(stub_entry, tmp /* return_pc=tmp */);
2654 //=============================================================================
2655 //
2656 // InterpretMethod triggered OSR compilation of some Java method M
2657 // and now asks to run the compiled code. We call this code the
2658 // `callee'.
2659 //
2660 // This is our current idea on how OSR should look like on PPC64:
2661 //
2662 // While interpreting a Java method M the stack is:
2663 //
2664 // (InterpretMethod (M), IJAVA_FRAME (M), ANY_FRAME, ...).
2665 //
2666 // After having OSR compiled M, `InterpretMethod' returns to the
2667 // frame manager, sending the message `retry_method_osr'. The stack
2668 // is:
2669 //
2670 // (IJAVA_FRAME (M), ANY_FRAME, ...).
2671 //
2672 // The compiler will have generated an `nmethod' suitable for
2673 // continuing execution of M at the bytecode index at which OSR took
2674 // place. So now the frame manager calls the OSR entry. The OSR
2675 // entry sets up a JIT_FRAME for M and continues execution of M with
2676 // initial state determined by the IJAVA_FRAME.
2677 //
2678 // (JIT_FRAME (M), IJAVA_FRAME (M), ANY_FRAME, ...).
2679 //
2681 __ BIND(retry_method_osr);
2682 {
2683 //
2684 // Registers alive
2685 // R16_thread
2686 // R15_prev_state - address of caller's BytecodeInterpreter
2687 // R14_state - address of callee's BytecodeInterpreter
2688 // R1_SP - callee's SP before call to InterpretMethod
2689 //
2690 // Registers updated
2691 // R17 - pointer to callee's locals array
2692 // (declared via `interpreter_arg_ptr_reg' in the AD file)
2693 // R19_method - callee's Method
2694 // R1_SP - callee's SP (will become SP of OSR adapter frame)
2695 //
2697 // Provide a debugger breakpoint in the frame manager if breakpoints
2698 // in osr'd methods are requested.
2699 #ifdef COMPILER2
2700 NOT_PRODUCT( if (OptoBreakpointOSR) { __ illtrap(); } )
2701 #endif
2703 // Load callee's pointer to locals array from callee's state.
2704 // __ ld(R17, state_(_locals));
2706 // Load osr entry.
2707 __ ld(R12_scratch2, state_(_result._osr._osr_entry));
2709 // Load address of temporary osr buffer to arg1.
2710 __ ld(R3_ARG1, state_(_result._osr._osr_buf));
2711 __ mtctr(R12_scratch2);
2713 // Load method, gc may move it during execution of osr'd method.
2714 __ ld(R22_tmp2, state_(_method));
2715 // Load message 'call_method'.
2716 __ li(R23_tmp3, BytecodeInterpreter::call_method);
2718 {
2719 // Pop the IJAVA frame of the method which we are going to call osr'd.
2720 Label no_state, skip_no_state;
2721 __ pop_interpreter_state(/*prev_state_may_be_0=*/true);
2722 __ cmpdi(CCR0, R14_state,0);
2723 __ beq(CCR0, no_state);
2724 // return to interpreter
2725 __ pop_interpreter_frame_to_state(R14_state, R11_scratch1, R12_scratch2, R21_tmp1);
2727 // Init _result._to_call._callee and tell gc that it contains a valid oop
2728 // by setting _msg to 'call_method'.
2729 __ std(R22_tmp2, state_(_result._to_call._callee));
2730 // TODO: PPC port: assert(4 == BytecodeInterpreter::sz_msg(), "unexpected field size");
2731 __ stw(R23_tmp3, state_(_msg));
2733 __ load_const(R21_tmp1, frame_manager_specialized_return);
2734 __ b(skip_no_state);
2735 __ bind(no_state);
2737 // Return to initial caller.
2739 // Get rid of top frame.
2740 __ pop_frame();
2742 // Load return PC from parent frame.
2743 __ ld(R21_tmp1, _parent_ijava_frame_abi(lr), R1_SP);
2745 // Resize frame to get rid of a potential extension.
2746 __ resize_frame_to_initial_caller(R11_scratch1, R12_scratch2);
2748 __ bind(skip_no_state);
2750 // Update LR with return pc.
2751 __ mtlr(R21_tmp1);
2752 }
2753 // Jump to the osr entry point.
2754 __ bctr();
2756 }
2758 //=============================================================================
2759 // Interpreted method "returned" with an exception, pass it on.
2760 // Pass no result, unwind activation and continue/return to
2761 // interpreter/call_stub/c2.
2763 __ BIND(throwing_exception);
2765 // Check if this is the initial invocation of the frame manager. If
2766 // so, previous interpreter state in R15_prev_state will be null.
2768 // New tos of caller is callee's first parameter address, that is
2769 // callee's incoming arguments are popped.
2770 __ ld(R3_RET, state_(_locals));
2772 // Check whether this is an initial call.
2773 __ cmpdi(CCR0, R15_prev_state, 0);
2774 // Yes, called from the call stub or from generated code via a c2i frame.
2775 __ beq(CCR0, unwind_initial_activation_pending_exception);
2777 // Send resume message, interpreter will see the exception first.
2779 __ li(msg, BytecodeInterpreter::method_resume);
2780 __ b(unwind_recursive_activation);
2783 //=============================================================================
2784 // Push the last instruction out to the code buffer.
2786 {
2787 __ unimplemented("end of InterpreterGenerator::generate_normal_entry", 128);
2788 }
2790 interpreter_frame_manager = entry;
2791 return interpreter_frame_manager;
2792 }
2794 // Generate code for various sorts of method entries
2795 //
2796 address AbstractInterpreterGenerator::generate_method_entry(AbstractInterpreter::MethodKind kind) {
2797 address entry_point = NULL;
2799 switch (kind) {
2800 case Interpreter::zerolocals : break;
2801 case Interpreter::zerolocals_synchronized : break;
2802 case Interpreter::native : // Fall thru
2803 case Interpreter::native_synchronized : entry_point = ((CppInterpreterGenerator*)this)->generate_native_entry(); break;
2804 case Interpreter::empty : break;
2805 case Interpreter::accessor : entry_point = ((InterpreterGenerator*)this)->generate_accessor_entry(); break;
2806 case Interpreter::abstract : entry_point = ((InterpreterGenerator*)this)->generate_abstract_entry(); break;
2807 // These are special interpreter intrinsics which we don't support so far.
2808 case Interpreter::java_lang_math_sin : break;
2809 case Interpreter::java_lang_math_cos : break;
2810 case Interpreter::java_lang_math_tan : break;
2811 case Interpreter::java_lang_math_abs : break;
2812 case Interpreter::java_lang_math_log : break;
2813 case Interpreter::java_lang_math_log10 : break;
2814 case Interpreter::java_lang_math_sqrt : break;
2815 case Interpreter::java_lang_math_pow : break;
2816 case Interpreter::java_lang_math_exp : break;
2817 case Interpreter::java_lang_ref_reference_get: entry_point = ((InterpreterGenerator*)this)->generate_Reference_get_entry(); break;
2818 default : ShouldNotReachHere(); break;
2819 }
2821 if (entry_point) {
2822 return entry_point;
2823 }
2824 return ((InterpreterGenerator*)this)->generate_normal_entry();
2825 }
2827 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2828 : CppInterpreterGenerator(code) {
2829 generate_all(); // down here so it can be "virtual"
2830 }
2832 // How much stack a topmost interpreter method activation needs in words.
2833 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
2834 // Computation is in bytes not words to match layout_activation_impl
2835 // below, but the return is in words.
2837 //
2838 // 0 [TOP_IJAVA_FRAME_ABI] \
2839 // alignment (optional) \ |
2840 // [operand stack / Java parameters] > stack | |
2841 // [monitors] (optional) > monitors | |
2842 // [PARENT_IJAVA_FRAME_ABI] \ | |
2843 // [BytecodeInterpreter object] > interpreter \ | | |
2844 // alignment (optional) | round | parent | round | top
2845 // [Java result] (2 slots) > result | | | |
2846 // [Java non-arg locals] \ locals | | | |
2847 // [arg locals] / / / / /
2848 //
2850 int locals = method->max_locals() * BytesPerWord;
2851 int interpreter = frame::interpreter_frame_cinterpreterstate_size_in_bytes();
2852 int result = 2 * BytesPerWord;
2854 int parent = round_to(interpreter + result + locals, 16) + frame::parent_ijava_frame_abi_size;
2856 int stack = method->max_stack() * BytesPerWord;
2857 int monitors = method->is_synchronized() ? frame::interpreter_frame_monitor_size_in_bytes() : 0;
2858 int top = round_to(parent + monitors + stack, 16) + frame::top_ijava_frame_abi_size;
2860 return (top / BytesPerWord);
2861 }
2863 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2864 frame* caller,
2865 frame* current,
2866 Method* method,
2867 intptr_t* locals,
2868 intptr_t* stack,
2869 intptr_t* stack_base,
2870 intptr_t* monitor_base,
2871 intptr_t* frame_sp,
2872 bool is_top_frame) {
2873 // What about any vtable?
2874 //
2875 to_fill->_thread = JavaThread::current();
2876 // This gets filled in later but make it something recognizable for now.
2877 to_fill->_bcp = method->code_base();
2878 to_fill->_locals = locals;
2879 to_fill->_constants = method->constants()->cache();
2880 to_fill->_method = method;
2881 to_fill->_mdx = NULL;
2882 to_fill->_stack = stack;
2884 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution()) {
2885 to_fill->_msg = deopt_resume2;
2886 } else {
2887 to_fill->_msg = method_resume;
2888 }
2889 to_fill->_result._to_call._bcp_advance = 0;
2890 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2891 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2892 to_fill->_prev_link = NULL;
2894 if (caller->is_interpreted_frame()) {
2895 interpreterState prev = caller->get_interpreterState();
2897 // Support MH calls. Make sure the interpreter will return the right address:
2898 // 1. Caller did ordinary interpreted->compiled call call: Set a prev_state
2899 // which makes the CPP interpreter return to frame manager "return_from_interpreted_method"
2900 // entry after finishing execution.
2901 // 2. Caller did a MH call: If the caller has a MethodHandleInvoke in it's
2902 // state (invariant: must be the caller of the bottom vframe) we used the
2903 // "call_special" entry to do the call, meaning the arguments have not been
2904 // popped from the stack. Therefore, don't enter a prev state in this case
2905 // in order to return to "return_from_native" frame manager entry which takes
2906 // care of popping arguments. Also, don't overwrite the MH.invoke Method in
2907 // the prev_state in order to be able to figure out the number of arguments to
2908 // pop.
2909 // The parameter method can represent MethodHandle.invokeExact(...).
2910 // The MethodHandleCompiler generates these synthetic Methods,
2911 // including bytecodes, if an invokedynamic call gets inlined. In
2912 // this case we want to return like from any other interpreted
2913 // Java call, so we set _prev_link.
2914 to_fill->_prev_link = prev;
2916 if (*prev->_bcp == Bytecodes::_invokeinterface || *prev->_bcp == Bytecodes::_invokedynamic) {
2917 prev->_result._to_call._bcp_advance = 5;
2918 } else {
2919 prev->_result._to_call._bcp_advance = 3;
2920 }
2921 }
2922 to_fill->_oop_temp = NULL;
2923 to_fill->_stack_base = stack_base;
2924 // Need +1 here because stack_base points to the word just above the
2925 // first expr stack entry and stack_limit is supposed to point to
2926 // the word just below the last expr stack entry. See
2927 // generate_compute_interpreter_state.
2928 to_fill->_stack_limit = stack_base - (method->max_stack() + 1);
2929 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2931 to_fill->_frame_bottom = frame_sp;
2933 // PPC64 specific
2934 to_fill->_last_Java_pc = NULL;
2935 to_fill->_last_Java_fp = NULL;
2936 to_fill->_last_Java_sp = frame_sp;
2937 #ifdef ASSERT
2938 to_fill->_self_link = to_fill;
2939 to_fill->_native_fresult = 123456.789;
2940 to_fill->_native_lresult = CONST64(0xdeafcafedeadc0de);
2941 #endif
2942 }
2944 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate,
2945 address last_Java_pc,
2946 intptr_t* last_Java_fp) {
2947 istate->_last_Java_pc = last_Java_pc;
2948 istate->_last_Java_fp = last_Java_fp;
2949 }
2951 // Computes monitor_size and top_frame_size in bytes.
2952 static void frame_size_helper(int max_stack,
2953 int monitors,
2954 int& monitor_size,
2955 int& top_frame_size) {
2956 monitor_size = frame::interpreter_frame_monitor_size_in_bytes() * monitors;
2957 top_frame_size = round_to(frame::interpreter_frame_cinterpreterstate_size_in_bytes()
2958 + monitor_size
2959 + max_stack * Interpreter::stackElementSize
2960 + 2 * Interpreter::stackElementSize,
2961 frame::alignment_in_bytes)
2962 + frame::top_ijava_frame_abi_size;
2963 }
2965 // Returns number of stackElementWords needed for the interpreter frame with the
2966 // given sections.
2967 int AbstractInterpreter::size_activation(int max_stack,
2968 int temps,
2969 int extra_args,
2970 int monitors,
2971 int callee_params,
2972 int callee_locals,
2973 bool is_top_frame) {
2974 int monitor_size = 0;
2975 int top_frame_size = 0;
2976 frame_size_helper(max_stack, monitors, monitor_size, top_frame_size);
2978 int frame_size;
2979 if (is_top_frame) {
2980 frame_size = top_frame_size;
2981 } else {
2982 frame_size = round_to(frame::interpreter_frame_cinterpreterstate_size_in_bytes()
2983 + monitor_size
2984 + (temps - callee_params + callee_locals) * Interpreter::stackElementSize
2985 + 2 * Interpreter::stackElementSize,
2986 frame::alignment_in_bytes)
2987 + frame::parent_ijava_frame_abi_size;
2988 assert(extra_args == 0, "non-zero for top_frame only");
2989 }
2991 return frame_size / Interpreter::stackElementSize;
2992 }
2994 void AbstractInterpreter::layout_activation(Method* method,
2995 int temps, // Number of slots on java expression stack in use.
2996 int popframe_args,
2997 int monitors, // Number of active monitors.
2998 int caller_actual_parameters,
2999 int callee_params,// Number of slots for callee parameters.
3000 int callee_locals,// Number of slots for locals.
3001 frame* caller,
3002 frame* interpreter_frame,
3003 bool is_top_frame,
3004 bool is_bottom_frame) {
3006 // NOTE this code must exactly mimic what
3007 // InterpreterGenerator::generate_compute_interpreter_state() does
3008 // as far as allocating an interpreter frame. However there is an
3009 // exception. With the C++ based interpreter only the top most frame
3010 // has a full sized expression stack. The 16 byte slop factor is
3011 // both the abi scratch area and a place to hold a result from a
3012 // callee on its way to the callers stack.
3014 int monitor_size = 0;
3015 int top_frame_size = 0;
3016 frame_size_helper(method->max_stack(), monitors, monitor_size, top_frame_size);
3018 intptr_t sp = (intptr_t)interpreter_frame->sp();
3019 intptr_t fp = *(intptr_t *)sp;
3020 assert(fp == (intptr_t)caller->sp(), "fp must match");
3021 interpreterState cur_state =
3022 (interpreterState)(fp - frame::interpreter_frame_cinterpreterstate_size_in_bytes());
3024 // Now fill in the interpreterState object.
3026 intptr_t* locals;
3027 if (caller->is_interpreted_frame()) {
3028 // Locals must agree with the caller because it will be used to set the
3029 // caller's tos when we return.
3030 interpreterState prev = caller->get_interpreterState();
3031 // Calculate start of "locals" for MH calls. For MH calls, the
3032 // current method() (= MH target) and prev->callee() (=
3033 // MH.invoke*()) are different and especially have different
3034 // signatures. To pop the argumentsof the caller, we must use
3035 // the prev->callee()->size_of_arguments() because that's what
3036 // the caller actually pushed. Currently, for synthetic MH
3037 // calls (deoptimized from inlined MH calls), detected by
3038 // is_method_handle_invoke(), we use the callee's arguments
3039 // because here, the caller's and callee's signature match.
3040 if (true /*!caller->is_at_mh_callsite()*/) {
3041 locals = prev->stack() + method->size_of_parameters();
3042 } else {
3043 // Normal MH call.
3044 locals = prev->stack() + prev->callee()->size_of_parameters();
3045 }
3046 } else {
3047 bool is_deopted;
3048 locals = (intptr_t*) (fp + ((method->max_locals() - 1) * BytesPerWord) +
3049 frame::parent_ijava_frame_abi_size);
3050 }
3052 intptr_t* monitor_base = (intptr_t*) cur_state;
3053 intptr_t* stack_base = (intptr_t*) ((intptr_t) monitor_base - monitor_size);
3055 // Provide pop_frame capability on PPC64, add popframe_args.
3056 // +1 because stack is always prepushed.
3057 intptr_t* stack = (intptr_t*) ((intptr_t) stack_base - (temps + popframe_args + 1) * BytesPerWord);
3059 BytecodeInterpreter::layout_interpreterState(cur_state,
3060 caller,
3061 interpreter_frame,
3062 method,
3063 locals,
3064 stack,
3065 stack_base,
3066 monitor_base,
3067 (intptr_t*)(((intptr_t)fp) - top_frame_size),
3068 is_top_frame);
3070 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address,
3071 interpreter_frame->fp());
3072 }
3074 #endif // CC_INTERP