1 /*
2 * Copyright (c) 2014, 2018, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2015, 2018, SAP SE. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "gc/shared/barrierSetAssembler.hpp"
29 #include "interpreter/bytecodeHistogram.hpp"
30 #include "interpreter/interpreter.hpp"
31 #include "interpreter/interpreterRuntime.hpp"
32 #include "interpreter/interp_masm.hpp"
33 #include "interpreter/templateInterpreterGenerator.hpp"
34 #include "interpreter/templateTable.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/sharedRuntime.hpp"
45 #include "runtime/stubRoutines.hpp"
46 #include "runtime/synchronizer.hpp"
47 #include "runtime/timer.hpp"
48 #include "runtime/vframeArray.hpp"
49 #include "utilities/debug.hpp"
50 #include "utilities/macros.hpp"
51
52 #undef __
53 #define __ _masm->
54
55 // Size of interpreter code. Increase if too small. Interpreter will
56 // fail with a guarantee ("not enough space for interpreter generation");
57 // if too small.
58 // Run with +PrintInterpreter to get the VM to print out the size.
59 // Max size with JVMTI
60 int TemplateInterpreter::InterpreterCodeSize = 256*K;
61
62 #ifdef PRODUCT
63 #define BLOCK_COMMENT(str) /* nothing */
64 #else
65 #define BLOCK_COMMENT(str) __ block_comment(str)
66 #endif
67
68 #define BIND(label) __ bind(label); BLOCK_COMMENT(#label ":")
69
70 //-----------------------------------------------------------------------------
71
72 address TemplateInterpreterGenerator::generate_slow_signature_handler() {
73 // Slow_signature handler that respects the PPC C calling conventions.
74 //
75 // We get called by the native entry code with our output register
76 // area == 8. First we call InterpreterRuntime::get_result_handler
77 // to copy the pointer to the signature string temporarily to the
78 // first C-argument and to return the result_handler in
79 // R3_RET. Since native_entry will copy the jni-pointer to the
80 // first C-argument slot later on, it is OK to occupy this slot
81 // temporarilly. Then we copy the argument list on the java
82 // expression stack into native varargs format on the native stack
83 // and load arguments into argument registers. Integer arguments in
84 // the varargs vector will be sign-extended to 8 bytes.
85 //
86 // On entry:
87 // R3_ARG1 - intptr_t* Address of java argument list in memory.
88 // R15_prev_state - BytecodeInterpreter* Address of interpreter state for
89 // this method
90 // R19_method
91 //
92 // On exit (just before return instruction):
93 // R3_RET - contains the address of the result_handler.
94 // R4_ARG2 - is not updated for static methods and contains "this" otherwise.
95 // R5_ARG3-R10_ARG8: - When the (i-2)th Java argument is not of type float or double,
96 // ARGi contains this argument. Otherwise, ARGi is not updated.
97 // F1_ARG1-F13_ARG13 - contain the first 13 arguments of type float or double.
98
99 const int LogSizeOfTwoInstructions = 3;
100
101 // FIXME: use Argument:: GL: Argument names different numbers!
102 const int max_fp_register_arguments = 13;
103 const int max_int_register_arguments = 6; // first 2 are reserved
104
105 const Register arg_java = R21_tmp1;
106 const Register arg_c = R22_tmp2;
107 const Register signature = R23_tmp3; // is string
108 const Register sig_byte = R24_tmp4;
109 const Register fpcnt = R25_tmp5;
110 const Register argcnt = R26_tmp6;
111 const Register intSlot = R27_tmp7;
112 const Register target_sp = R28_tmp8;
113 const FloatRegister floatSlot = F0;
114
115 address entry = __ function_entry();
116
117 __ save_LR_CR(R0);
118 __ save_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
119 // We use target_sp for storing arguments in the C frame.
120 __ mr(target_sp, R1_SP);
121 __ push_frame_reg_args_nonvolatiles(0, R11_scratch1);
122
123 __ mr(arg_java, R3_ARG1);
124
125 __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::get_signature), R16_thread, R19_method);
126
127 // Signature is in R3_RET. Signature is callee saved.
128 __ mr(signature, R3_RET);
129
130 // Get the result handler.
131 __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::get_result_handler), R16_thread, R19_method);
132
133 {
134 Label L;
135 // test if static
136 // _access_flags._flags must be at offset 0.
137 // TODO PPC port: requires change in shared code.
138 //assert(in_bytes(AccessFlags::flags_offset()) == 0,
139 // "MethodDesc._access_flags == MethodDesc._access_flags._flags");
140 // _access_flags must be a 32 bit value.
141 assert(sizeof(AccessFlags) == 4, "wrong size");
142 __ lwa(R11_scratch1/*access_flags*/, method_(access_flags));
143 // testbit with condition register.
144 __ testbitdi(CCR0, R0, R11_scratch1/*access_flags*/, JVM_ACC_STATIC_BIT);
145 __ btrue(CCR0, L);
146 // For non-static functions, pass "this" in R4_ARG2 and copy it
147 // to 2nd C-arg slot.
148 // We need to box the Java object here, so we use arg_java
149 // (address of current Java stack slot) as argument and don't
150 // dereference it as in case of ints, floats, etc.
151 __ mr(R4_ARG2, arg_java);
152 __ addi(arg_java, arg_java, -BytesPerWord);
153 __ std(R4_ARG2, _abi(carg_2), target_sp);
154 __ bind(L);
155 }
156
157 // Will be incremented directly after loop_start. argcnt=0
158 // corresponds to 3rd C argument.
159 __ li(argcnt, -1);
160 // arg_c points to 3rd C argument
161 __ addi(arg_c, target_sp, _abi(carg_3));
162 // no floating-point args parsed so far
163 __ li(fpcnt, 0);
164
165 Label move_intSlot_to_ARG, move_floatSlot_to_FARG;
166 Label loop_start, loop_end;
167 Label do_int, do_long, do_float, do_double, do_dontreachhere, do_object, do_array, do_boxed;
168
169 // signature points to '(' at entry
170 #ifdef ASSERT
171 __ lbz(sig_byte, 0, signature);
172 __ cmplwi(CCR0, sig_byte, '(');
173 __ bne(CCR0, do_dontreachhere);
174 #endif
175
176 __ bind(loop_start);
177
178 __ addi(argcnt, argcnt, 1);
179 __ lbzu(sig_byte, 1, signature);
180
181 __ cmplwi(CCR0, sig_byte, ')'); // end of signature
182 __ beq(CCR0, loop_end);
183
184 __ cmplwi(CCR0, sig_byte, 'B'); // byte
185 __ beq(CCR0, do_int);
186
187 __ cmplwi(CCR0, sig_byte, 'C'); // char
188 __ beq(CCR0, do_int);
189
190 __ cmplwi(CCR0, sig_byte, 'D'); // double
191 __ beq(CCR0, do_double);
192
193 __ cmplwi(CCR0, sig_byte, 'F'); // float
194 __ beq(CCR0, do_float);
195
196 __ cmplwi(CCR0, sig_byte, 'I'); // int
197 __ beq(CCR0, do_int);
198
199 __ cmplwi(CCR0, sig_byte, 'J'); // long
200 __ beq(CCR0, do_long);
201
202 __ cmplwi(CCR0, sig_byte, 'S'); // short
203 __ beq(CCR0, do_int);
204
205 __ cmplwi(CCR0, sig_byte, 'Z'); // boolean
206 __ beq(CCR0, do_int);
207
208 __ cmplwi(CCR0, sig_byte, 'L'); // object
209 __ beq(CCR0, do_object);
210
211 __ cmplwi(CCR0, sig_byte, '['); // array
212 __ beq(CCR0, do_array);
213
214 // __ cmplwi(CCR0, sig_byte, 'V'); // void cannot appear since we do not parse the return type
215 // __ beq(CCR0, do_void);
216
217 __ bind(do_dontreachhere);
218
219 __ unimplemented("ShouldNotReachHere in slow_signature_handler", 120);
220
221 __ bind(do_array);
222
223 {
224 Label start_skip, end_skip;
225
226 __ bind(start_skip);
227 __ lbzu(sig_byte, 1, signature);
228 __ cmplwi(CCR0, sig_byte, '[');
229 __ beq(CCR0, start_skip); // skip further brackets
230 __ cmplwi(CCR0, sig_byte, '9');
231 __ bgt(CCR0, end_skip); // no optional size
232 __ cmplwi(CCR0, sig_byte, '0');
233 __ bge(CCR0, start_skip); // skip optional size
234 __ bind(end_skip);
235
236 __ cmplwi(CCR0, sig_byte, 'L');
237 __ beq(CCR0, do_object); // for arrays of objects, the name of the object must be skipped
238 __ b(do_boxed); // otherwise, go directly to do_boxed
239 }
240
241 __ bind(do_object);
242 {
243 Label L;
244 __ bind(L);
245 __ lbzu(sig_byte, 1, signature);
246 __ cmplwi(CCR0, sig_byte, ';');
247 __ bne(CCR0, L);
248 }
249 // Need to box the Java object here, so we use arg_java (address of
250 // current Java stack slot) as argument and don't dereference it as
251 // in case of ints, floats, etc.
252 Label do_null;
253 __ bind(do_boxed);
254 __ ld(R0,0, arg_java);
255 __ cmpdi(CCR0, R0, 0);
256 __ li(intSlot,0);
257 __ beq(CCR0, do_null);
258 __ mr(intSlot, arg_java);
259 __ bind(do_null);
260 __ std(intSlot, 0, arg_c);
261 __ addi(arg_java, arg_java, -BytesPerWord);
262 __ addi(arg_c, arg_c, BytesPerWord);
263 __ cmplwi(CCR0, argcnt, max_int_register_arguments);
264 __ blt(CCR0, move_intSlot_to_ARG);
265 __ b(loop_start);
266
267 __ bind(do_int);
268 __ lwa(intSlot, 0, arg_java);
269 __ std(intSlot, 0, arg_c);
270 __ addi(arg_java, arg_java, -BytesPerWord);
271 __ addi(arg_c, arg_c, BytesPerWord);
272 __ cmplwi(CCR0, argcnt, max_int_register_arguments);
273 __ blt(CCR0, move_intSlot_to_ARG);
274 __ b(loop_start);
275
276 __ bind(do_long);
277 __ ld(intSlot, -BytesPerWord, arg_java);
278 __ std(intSlot, 0, arg_c);
279 __ addi(arg_java, arg_java, - 2 * BytesPerWord);
280 __ addi(arg_c, arg_c, BytesPerWord);
281 __ cmplwi(CCR0, argcnt, max_int_register_arguments);
282 __ blt(CCR0, move_intSlot_to_ARG);
283 __ b(loop_start);
284
285 __ bind(do_float);
286 __ lfs(floatSlot, 0, arg_java);
287 #if defined(LINUX)
288 // Linux uses ELF ABI. Both original ELF and ELFv2 ABIs have float
289 // in the least significant word of an argument slot.
290 #if defined(VM_LITTLE_ENDIAN)
291 __ stfs(floatSlot, 0, arg_c);
292 #else
293 __ stfs(floatSlot, 4, arg_c);
294 #endif
295 #elif defined(AIX)
296 // Although AIX runs on big endian CPU, float is in most significant
297 // word of an argument slot.
298 __ stfs(floatSlot, 0, arg_c);
299 #else
300 #error "unknown OS"
301 #endif
302 __ addi(arg_java, arg_java, -BytesPerWord);
303 __ addi(arg_c, arg_c, BytesPerWord);
304 __ cmplwi(CCR0, fpcnt, max_fp_register_arguments);
305 __ blt(CCR0, move_floatSlot_to_FARG);
306 __ b(loop_start);
307
308 __ bind(do_double);
309 __ lfd(floatSlot, - BytesPerWord, arg_java);
310 __ stfd(floatSlot, 0, arg_c);
311 __ addi(arg_java, arg_java, - 2 * BytesPerWord);
312 __ addi(arg_c, arg_c, BytesPerWord);
313 __ cmplwi(CCR0, fpcnt, max_fp_register_arguments);
314 __ blt(CCR0, move_floatSlot_to_FARG);
315 __ b(loop_start);
316
317 __ bind(loop_end);
318
319 __ pop_frame();
320 __ restore_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
321 __ restore_LR_CR(R0);
322
323 __ blr();
324
325 Label move_int_arg, move_float_arg;
326 __ bind(move_int_arg); // each case must consist of 2 instructions (otherwise adapt LogSizeOfTwoInstructions)
327 __ mr(R5_ARG3, intSlot); __ b(loop_start);
328 __ mr(R6_ARG4, intSlot); __ b(loop_start);
329 __ mr(R7_ARG5, intSlot); __ b(loop_start);
330 __ mr(R8_ARG6, intSlot); __ b(loop_start);
331 __ mr(R9_ARG7, intSlot); __ b(loop_start);
332 __ mr(R10_ARG8, intSlot); __ b(loop_start);
333
334 __ bind(move_float_arg); // each case must consist of 2 instructions (otherwise adapt LogSizeOfTwoInstructions)
335 __ fmr(F1_ARG1, floatSlot); __ b(loop_start);
336 __ fmr(F2_ARG2, floatSlot); __ b(loop_start);
337 __ fmr(F3_ARG3, floatSlot); __ b(loop_start);
338 __ fmr(F4_ARG4, floatSlot); __ b(loop_start);
339 __ fmr(F5_ARG5, floatSlot); __ b(loop_start);
340 __ fmr(F6_ARG6, floatSlot); __ b(loop_start);
341 __ fmr(F7_ARG7, floatSlot); __ b(loop_start);
342 __ fmr(F8_ARG8, floatSlot); __ b(loop_start);
343 __ fmr(F9_ARG9, floatSlot); __ b(loop_start);
344 __ fmr(F10_ARG10, floatSlot); __ b(loop_start);
345 __ fmr(F11_ARG11, floatSlot); __ b(loop_start);
346 __ fmr(F12_ARG12, floatSlot); __ b(loop_start);
347 __ fmr(F13_ARG13, floatSlot); __ b(loop_start);
348
349 __ bind(move_intSlot_to_ARG);
350 __ sldi(R0, argcnt, LogSizeOfTwoInstructions);
351 __ load_const(R11_scratch1, move_int_arg); // Label must be bound here.
352 __ add(R11_scratch1, R0, R11_scratch1);
353 __ mtctr(R11_scratch1/*branch_target*/);
354 __ bctr();
355 __ bind(move_floatSlot_to_FARG);
356 __ sldi(R0, fpcnt, LogSizeOfTwoInstructions);
357 __ addi(fpcnt, fpcnt, 1);
358 __ load_const(R11_scratch1, move_float_arg); // Label must be bound here.
359 __ add(R11_scratch1, R0, R11_scratch1);
360 __ mtctr(R11_scratch1/*branch_target*/);
361 __ bctr();
362
363 return entry;
364 }
365
366 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
367 //
368 // Registers alive
369 // R3_RET
370 // LR
371 //
372 // Registers updated
373 // R3_RET
374 //
375
376 Label done;
377 address entry = __ pc();
378
379 switch (type) {
380 case T_BOOLEAN:
381 // convert !=0 to 1
382 __ neg(R0, R3_RET);
383 __ orr(R0, R3_RET, R0);
384 __ srwi(R3_RET, R0, 31);
385 break;
386 case T_BYTE:
387 // sign extend 8 bits
388 __ extsb(R3_RET, R3_RET);
389 break;
390 case T_CHAR:
391 // zero extend 16 bits
392 __ clrldi(R3_RET, R3_RET, 48);
393 break;
394 case T_SHORT:
395 // sign extend 16 bits
396 __ extsh(R3_RET, R3_RET);
397 break;
398 case T_INT:
399 // sign extend 32 bits
400 __ extsw(R3_RET, R3_RET);
401 break;
402 case T_LONG:
403 break;
404 case T_OBJECT:
405 // JNIHandles::resolve result.
406 __ resolve_jobject(R3_RET, R11_scratch1, R31, /* needs_frame */ true); // kills R31
407 break;
408 case T_FLOAT:
409 break;
410 case T_DOUBLE:
411 break;
412 case T_VOID:
413 break;
414 default: ShouldNotReachHere();
415 }
416
417 BIND(done);
418 __ blr();
419
420 return entry;
421 }
422
423 // Abstract method entry.
424 //
425 address TemplateInterpreterGenerator::generate_abstract_entry(void) {
426 address entry = __ pc();
427
428 //
429 // Registers alive
430 // R16_thread - JavaThread*
431 // R19_method - callee's method (method to be invoked)
432 // R1_SP - SP prepared such that caller's outgoing args are near top
433 // LR - return address to caller
434 //
435 // Stack layout at this point:
436 //
437 // 0 [TOP_IJAVA_FRAME_ABI] <-- R1_SP
438 // alignment (optional)
439 // [outgoing Java arguments]
440 // ...
441 // PARENT [PARENT_IJAVA_FRAME_ABI]
442 // ...
443 //
444
445 // Can't use call_VM here because we have not set up a new
446 // interpreter state. Make the call to the vm and make it look like
447 // our caller set up the JavaFrameAnchor.
448 __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/);
449
450 // Push a new C frame and save LR.
451 __ save_LR_CR(R0);
452 __ push_frame_reg_args(0, R11_scratch1);
453
454 // This is not a leaf but we have a JavaFrameAnchor now and we will
455 // check (create) exceptions afterward so this is ok.
456 __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodErrorWithMethod),
457 R16_thread, R19_method);
458
459 // Pop the C frame and restore LR.
460 __ pop_frame();
461 __ restore_LR_CR(R0);
462
463 // Reset JavaFrameAnchor from call_VM_leaf above.
464 __ reset_last_Java_frame();
465
466 // We don't know our caller, so jump to the general forward exception stub,
467 // which will also pop our full frame off. Satisfy the interface of
468 // SharedRuntime::generate_forward_exception()
469 __ load_const_optimized(R11_scratch1, StubRoutines::forward_exception_entry(), R0);
470 __ mtctr(R11_scratch1);
471 __ bctr();
472
473 return entry;
474 }
475
476 // Interpreter intrinsic for WeakReference.get().
477 // 1. Don't push a full blown frame and go on dispatching, but fetch the value
478 // into R8 and return quickly
479 // 2. If G1 is active we *must* execute this intrinsic for corrrectness:
480 // It contains a GC barrier which puts the reference into the satb buffer
481 // to indicate that someone holds a strong reference to the object the
482 // weak ref points to!
483 address TemplateInterpreterGenerator::generate_Reference_get_entry(void) {
484 // Code: _aload_0, _getfield, _areturn
485 // parameter size = 1
486 //
487 // The code that gets generated by this routine is split into 2 parts:
488 // 1. the "intrinsified" code for G1 (or any SATB based GC),
489 // 2. the slow path - which is an expansion of the regular method entry.
490 //
491 // Notes:
492 // * In the G1 code we do not check whether we need to block for
493 // a safepoint. If G1 is enabled then we must execute the specialized
494 // code for Reference.get (except when the Reference object is null)
495 // so that we can log the value in the referent field with an SATB
496 // update buffer.
497 // If the code for the getfield template is modified so that the
498 // G1 pre-barrier code is executed when the current method is
499 // Reference.get() then going through the normal method entry
500 // will be fine.
501 // * The G1 code can, however, check the receiver object (the instance
502 // of java.lang.Reference) and jump to the slow path if null. If the
503 // Reference object is null then we obviously cannot fetch the referent
504 // and so we don't need to call the G1 pre-barrier. Thus we can use the
505 // regular method entry code to generate the NPE.
506 //
507
508 address entry = __ pc();
509
510 const int referent_offset = java_lang_ref_Reference::referent_offset;
511 guarantee(referent_offset > 0, "referent offset not initialized");
512
513 Label slow_path;
514
515 // Debugging not possible, so can't use __ skip_if_jvmti_mode(slow_path, GR31_SCRATCH);
516
517 // In the G1 code we don't check if we need to reach a safepoint. We
518 // continue and the thread will safepoint at the next bytecode dispatch.
519
520 // If the receiver is null then it is OK to jump to the slow path.
521 __ ld(R3_RET, Interpreter::stackElementSize, R15_esp); // get receiver
522
523 // Check if receiver == NULL and go the slow path.
524 __ cmpdi(CCR0, R3_RET, 0);
525 __ beq(CCR0, slow_path);
526
527 __ load_heap_oop(R3_RET, referent_offset, R3_RET,
528 /* non-volatile temp */ R31, R11_scratch1, true, ON_WEAK_OOP_REF);
529
530 // Generate the G1 pre-barrier code to log the value of
531 // the referent field in an SATB buffer. Note with
532 // these parameters the pre-barrier does not generate
533 // the load of the previous value.
534
535 // Restore caller sp for c2i case.
536 #ifdef ASSERT
537 __ ld(R9_ARG7, 0, R1_SP);
538 __ ld(R10_ARG8, 0, R21_sender_SP);
539 __ cmpd(CCR0, R9_ARG7, R10_ARG8);
540 __ asm_assert_eq("backlink", 0x544);
541 #endif // ASSERT
542 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
543
544 __ blr();
545
546 __ bind(slow_path);
547 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals), R11_scratch1);
548 return entry;
549 }
550
551 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
552 address entry = __ pc();
553
554 // Expression stack must be empty before entering the VM if an
555 // exception happened.
556 __ empty_expression_stack();
557 // Throw exception.
558 __ call_VM(noreg,
559 CAST_FROM_FN_PTR(address,
560 InterpreterRuntime::throw_StackOverflowError));
561 return entry;
562 }
563
564 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler() {
565 address entry = __ pc();
566 __ empty_expression_stack();
567 // R4_ARG2 already contains the array.
568 // Index is in R17_tos.
569 __ mr(R5_ARG3, R17_tos);
570 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException), R4_ARG2, R5_ARG3);
571 return entry;
572 }
573
574 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
575 address entry = __ pc();
576 // Expression stack must be empty before entering the VM if an
577 // exception happened.
578 __ empty_expression_stack();
579
580 // Load exception object.
581 // Thread will be loaded to R3_ARG1.
582 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ClassCastException), R17_tos);
583 #ifdef ASSERT
584 // Above call must not return here since exception pending.
585 __ should_not_reach_here();
586 #endif
587 return entry;
588 }
589
590 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
591 address entry = __ pc();
592 //__ untested("generate_exception_handler_common");
593 Register Rexception = R17_tos;
594
595 // Expression stack must be empty before entering the VM if an exception happened.
596 __ empty_expression_stack();
597
598 __ load_const_optimized(R4_ARG2, (address) name, R11_scratch1);
599 if (pass_oop) {
600 __ mr(R5_ARG3, Rexception);
601 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), false);
602 } else {
603 __ load_const_optimized(R5_ARG3, (address) message, R11_scratch1);
604 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), false);
605 }
606
607 // Throw exception.
608 __ mr(R3_ARG1, Rexception);
609 __ load_const_optimized(R11_scratch1, Interpreter::throw_exception_entry(), R12_scratch2);
610 __ mtctr(R11_scratch1);
611 __ bctr();
612
613 return entry;
614 }
615
616 // This entry is returned to when a call returns to the interpreter.
617 // When we arrive here, we expect that the callee stack frame is already popped.
618 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
619 address entry = __ pc();
620
621 // Move the value out of the return register back to the TOS cache of current frame.
622 switch (state) {
623 case ltos:
624 case btos:
625 case ztos:
626 case ctos:
627 case stos:
628 case atos:
629 case itos: __ mr(R17_tos, R3_RET); break; // RET -> TOS cache
630 case ftos:
631 case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET
632 case vtos: break; // Nothing to do, this was a void return.
633 default : ShouldNotReachHere();
634 }
635
636 __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp.
637 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
638 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
639
640 // Compiled code destroys templateTableBase, reload.
641 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R12_scratch2);
642
643 if (state == atos) {
644 __ profile_return_type(R3_RET, R11_scratch1, R12_scratch2);
645 }
646
647 const Register cache = R11_scratch1;
648 const Register size = R12_scratch2;
649 __ get_cache_and_index_at_bcp(cache, 1, index_size);
650
651 // Get least significant byte of 64 bit value:
652 #if defined(VM_LITTLE_ENDIAN)
653 __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()), cache);
654 #else
655 __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()) + 7, cache);
656 #endif
657 __ sldi(size, size, Interpreter::logStackElementSize);
658 __ add(R15_esp, R15_esp, size);
659
660 __ check_and_handle_popframe(R11_scratch1);
661 __ check_and_handle_earlyret(R11_scratch1);
662
663 __ dispatch_next(state, step);
664 return entry;
665 }
666
667 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step, address continuation) {
668 address entry = __ pc();
669 // If state != vtos, we're returning from a native method, which put it's result
670 // into the result register. So move the value out of the return register back
671 // to the TOS cache of current frame.
672
673 switch (state) {
674 case ltos:
675 case btos:
676 case ztos:
677 case ctos:
678 case stos:
679 case atos:
680 case itos: __ mr(R17_tos, R3_RET); break; // GR_RET -> TOS cache
681 case ftos:
682 case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET
683 case vtos: break; // Nothing to do, this was a void return.
684 default : ShouldNotReachHere();
685 }
686
687 // Load LcpoolCache @@@ should be already set!
688 __ get_constant_pool_cache(R27_constPoolCache);
689
690 // Handle a pending exception, fall through if none.
691 __ check_and_forward_exception(R11_scratch1, R12_scratch2);
692
693 // Start executing bytecodes.
694 if (continuation == NULL) {
695 __ dispatch_next(state, step);
696 } else {
697 __ jump_to_entry(continuation, R11_scratch1);
698 }
699
700 return entry;
701 }
702
703 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
704 address entry = __ pc();
705
706 __ push(state);
707 __ call_VM(noreg, runtime_entry);
708 __ dispatch_via(vtos, Interpreter::_normal_table.table_for(vtos));
709
710 return entry;
711 }
712
713 // Helpers for commoning out cases in the various type of method entries.
714
715 // Increment invocation count & check for overflow.
716 //
717 // Note: checking for negative value instead of overflow
718 // so we have a 'sticky' overflow test.
719 //
720 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
721 // Note: In tiered we increment either counters in method or in MDO depending if we're profiling or not.
722 Register Rscratch1 = R11_scratch1;
723 Register Rscratch2 = R12_scratch2;
724 Register R3_counters = R3_ARG1;
725 Label done;
726
727 if (TieredCompilation) {
728 const int increment = InvocationCounter::count_increment;
729 Label no_mdo;
730 if (ProfileInterpreter) {
731 const Register Rmdo = R3_counters;
732 // If no method data exists, go to profile_continue.
733 __ ld(Rmdo, in_bytes(Method::method_data_offset()), R19_method);
734 __ cmpdi(CCR0, Rmdo, 0);
735 __ beq(CCR0, no_mdo);
736
737 // Increment invocation counter in the MDO.
738 const int mdo_ic_offs = in_bytes(MethodData::invocation_counter_offset()) + in_bytes(InvocationCounter::counter_offset());
739 __ lwz(Rscratch2, mdo_ic_offs, Rmdo);
740 __ lwz(Rscratch1, in_bytes(MethodData::invoke_mask_offset()), Rmdo);
741 __ addi(Rscratch2, Rscratch2, increment);
742 __ stw(Rscratch2, mdo_ic_offs, Rmdo);
743 __ and_(Rscratch1, Rscratch2, Rscratch1);
744 __ bne(CCR0, done);
745 __ b(*overflow);
746 }
747
748 // Increment counter in MethodCounters*.
749 const int mo_ic_offs = in_bytes(MethodCounters::invocation_counter_offset()) + in_bytes(InvocationCounter::counter_offset());
750 __ bind(no_mdo);
751 __ get_method_counters(R19_method, R3_counters, done);
752 __ lwz(Rscratch2, mo_ic_offs, R3_counters);
753 __ lwz(Rscratch1, in_bytes(MethodCounters::invoke_mask_offset()), R3_counters);
754 __ addi(Rscratch2, Rscratch2, increment);
755 __ stw(Rscratch2, mo_ic_offs, R3_counters);
756 __ and_(Rscratch1, Rscratch2, Rscratch1);
757 __ beq(CCR0, *overflow);
758
759 __ bind(done);
760
761 } else {
762
763 // Update standard invocation counters.
764 Register Rsum_ivc_bec = R4_ARG2;
765 __ get_method_counters(R19_method, R3_counters, done);
766 __ increment_invocation_counter(R3_counters, Rsum_ivc_bec, R12_scratch2);
767 // Increment interpreter invocation counter.
768 if (ProfileInterpreter) { // %%% Merge this into methodDataOop.
769 __ lwz(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters);
770 __ addi(R12_scratch2, R12_scratch2, 1);
771 __ stw(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters);
772 }
773 // Check if we must create a method data obj.
774 if (ProfileInterpreter && profile_method != NULL) {
775 const Register profile_limit = Rscratch1;
776 __ lwz(profile_limit, in_bytes(MethodCounters::interpreter_profile_limit_offset()), R3_counters);
777 // Test to see if we should create a method data oop.
778 __ cmpw(CCR0, Rsum_ivc_bec, profile_limit);
779 __ blt(CCR0, *profile_method_continue);
780 // If no method data exists, go to profile_method.
781 __ test_method_data_pointer(*profile_method);
782 }
783 // Finally check for counter overflow.
784 if (overflow) {
785 const Register invocation_limit = Rscratch1;
786 __ lwz(invocation_limit, in_bytes(MethodCounters::interpreter_invocation_limit_offset()), R3_counters);
787 __ cmpw(CCR0, Rsum_ivc_bec, invocation_limit);
788 __ bge(CCR0, *overflow);
789 }
790
791 __ bind(done);
792 }
793 }
794
795 // Generate code to initiate compilation on invocation counter overflow.
796 void TemplateInterpreterGenerator::generate_counter_overflow(Label& continue_entry) {
797 // Generate code to initiate compilation on the counter overflow.
798
799 // InterpreterRuntime::frequency_counter_overflow takes one arguments,
800 // which indicates if the counter overflow occurs at a backwards branch (NULL bcp)
801 // We pass zero in.
802 // The call returns the address of the verified entry point for the method or NULL
803 // if the compilation did not complete (either went background or bailed out).
804 //
805 // Unlike the C++ interpreter above: Check exceptions!
806 // Assumption: Caller must set the flag "do_not_unlock_if_sychronized" if the monitor of a sync'ed
807 // method has not yet been created. Thus, no unlocking of a non-existing monitor can occur.
808
809 __ li(R4_ARG2, 0);
810 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), R4_ARG2, true);
811
812 // Returns verified_entry_point or NULL.
813 // We ignore it in any case.
814 __ b(continue_entry);
815 }
816
817 // See if we've got enough room on the stack for locals plus overhead below
818 // JavaThread::stack_overflow_limit(). If not, throw a StackOverflowError
819 // without going through the signal handler, i.e., reserved and yellow zones
820 // will not be made usable. The shadow zone must suffice to handle the
821 // overflow.
822 //
823 // Kills Rmem_frame_size, Rscratch1.
824 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rmem_frame_size, Register Rscratch1) {
825 Label done;
826 assert_different_registers(Rmem_frame_size, Rscratch1);
827
828 BLOCK_COMMENT("stack_overflow_check_with_compare {");
829 __ sub(Rmem_frame_size, R1_SP, Rmem_frame_size);
830 __ ld(Rscratch1, thread_(stack_overflow_limit));
831 __ cmpld(CCR0/*is_stack_overflow*/, Rmem_frame_size, Rscratch1);
832 __ bgt(CCR0/*is_stack_overflow*/, done);
833
834 // The stack overflows. Load target address of the runtime stub and call it.
835 assert(StubRoutines::throw_StackOverflowError_entry() != NULL, "generated in wrong order");
836 __ load_const_optimized(Rscratch1, (StubRoutines::throw_StackOverflowError_entry()), R0);
837 __ mtctr(Rscratch1);
838 // Restore caller_sp.
839 #ifdef ASSERT
840 __ ld(Rscratch1, 0, R1_SP);
841 __ ld(R0, 0, R21_sender_SP);
842 __ cmpd(CCR0, R0, Rscratch1);
843 __ asm_assert_eq("backlink", 0x547);
844 #endif // ASSERT
845 __ mr(R1_SP, R21_sender_SP);
846 __ bctr();
847
848 __ align(32, 12);
849 __ bind(done);
850 BLOCK_COMMENT("} stack_overflow_check_with_compare");
851 }
852
853 // Lock the current method, interpreter register window must be set up!
854 void TemplateInterpreterGenerator::lock_method(Register Rflags, Register Rscratch1, Register Rscratch2, bool flags_preloaded) {
855 const Register Robj_to_lock = Rscratch2;
856
857 {
858 if (!flags_preloaded) {
859 __ lwz(Rflags, method_(access_flags));
860 }
861
862 #ifdef ASSERT
863 // Check if methods needs synchronization.
864 {
865 Label Lok;
866 __ testbitdi(CCR0, R0, Rflags, JVM_ACC_SYNCHRONIZED_BIT);
867 __ btrue(CCR0,Lok);
868 __ stop("method doesn't need synchronization");
869 __ bind(Lok);
870 }
871 #endif // ASSERT
872 }
873
874 // Get synchronization object to Rscratch2.
875 {
876 Label Lstatic;
877 Label Ldone;
878
879 __ testbitdi(CCR0, R0, Rflags, JVM_ACC_STATIC_BIT);
880 __ btrue(CCR0, Lstatic);
881
882 // Non-static case: load receiver obj from stack and we're done.
883 __ ld(Robj_to_lock, R18_locals);
884 __ b(Ldone);
885
886 __ bind(Lstatic); // Static case: Lock the java mirror
887 // Load mirror from interpreter frame.
888 __ ld(Robj_to_lock, _abi(callers_sp), R1_SP);
889 __ ld(Robj_to_lock, _ijava_state_neg(mirror), Robj_to_lock);
890
891 __ bind(Ldone);
892 __ verify_oop(Robj_to_lock);
893 }
894
895 // Got the oop to lock => execute!
896 __ add_monitor_to_stack(true, Rscratch1, R0);
897
898 __ std(Robj_to_lock, BasicObjectLock::obj_offset_in_bytes(), R26_monitor);
899 __ lock_object(R26_monitor, Robj_to_lock);
900 }
901
902 // Generate a fixed interpreter frame for pure interpreter
903 // and I2N native transition frames.
904 //
905 // Before (stack grows downwards):
906 //
907 // | ... |
908 // |------------- |
909 // | java arg0 |
910 // | ... |
911 // | java argn |
912 // | | <- R15_esp
913 // | |
914 // |--------------|
915 // | abi_112 |
916 // | | <- R1_SP
917 // |==============|
918 //
919 //
920 // After:
921 //
922 // | ... |
923 // | java arg0 |<- R18_locals
924 // | ... |
925 // | java argn |
926 // |--------------|
927 // | |
928 // | java locals |
929 // | |
930 // |--------------|
931 // | abi_48 |
932 // |==============|
933 // | |
934 // | istate |
935 // | |
936 // |--------------|
937 // | monitor |<- R26_monitor
938 // |--------------|
939 // | |<- R15_esp
940 // | expression |
941 // | stack |
942 // | |
943 // |--------------|
944 // | |
945 // | abi_112 |<- R1_SP
946 // |==============|
947 //
948 // The top most frame needs an abi space of 112 bytes. This space is needed,
949 // since we call to c. The c function may spill their arguments to the caller
950 // frame. When we call to java, we don't need these spill slots. In order to save
951 // space on the stack, we resize the caller. However, java locals reside in
952 // the caller frame and the frame has to be increased. The frame_size for the
953 // current frame was calculated based on max_stack as size for the expression
954 // stack. At the call, just a part of the expression stack might be used.
955 // We don't want to waste this space and cut the frame back accordingly.
956 // The resulting amount for resizing is calculated as follows:
957 // resize = (number_of_locals - number_of_arguments) * slot_size
958 // + (R1_SP - R15_esp) + 48
959 //
960 // The size for the callee frame is calculated:
961 // framesize = 112 + max_stack + monitor + state_size
962 //
963 // maxstack: Max number of slots on the expression stack, loaded from the method.
964 // monitor: We statically reserve room for one monitor object.
965 // state_size: We save the current state of the interpreter to this area.
966 //
967 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call, Register Rsize_of_parameters, Register Rsize_of_locals) {
968 Register parent_frame_resize = R6_ARG4, // Frame will grow by this number of bytes.
969 top_frame_size = R7_ARG5,
970 Rconst_method = R8_ARG6;
971
972 assert_different_registers(Rsize_of_parameters, Rsize_of_locals, parent_frame_resize, top_frame_size);
973
974 __ ld(Rconst_method, method_(const));
975 __ lhz(Rsize_of_parameters /* number of params */,
976 in_bytes(ConstMethod::size_of_parameters_offset()), Rconst_method);
977 if (native_call) {
978 // If we're calling a native method, we reserve space for the worst-case signature
979 // handler varargs vector, which is max(Argument::n_register_parameters, parameter_count+2).
980 // We add two slots to the parameter_count, one for the jni
981 // environment and one for a possible native mirror.
982 Label skip_native_calculate_max_stack;
983 __ addi(top_frame_size, Rsize_of_parameters, 2);
984 __ cmpwi(CCR0, top_frame_size, Argument::n_register_parameters);
985 __ bge(CCR0, skip_native_calculate_max_stack);
986 __ li(top_frame_size, Argument::n_register_parameters);
987 __ bind(skip_native_calculate_max_stack);
988 __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize);
989 __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize);
990 __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize!
991 assert(Rsize_of_locals == noreg, "Rsize_of_locals not initialized"); // Only relevant value is Rsize_of_parameters.
992 } else {
993 __ lhz(Rsize_of_locals /* number of params */, in_bytes(ConstMethod::size_of_locals_offset()), Rconst_method);
994 __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize);
995 __ sldi(Rsize_of_locals, Rsize_of_locals, Interpreter::logStackElementSize);
996 __ lhz(top_frame_size, in_bytes(ConstMethod::max_stack_offset()), Rconst_method);
997 __ sub(R11_scratch1, Rsize_of_locals, Rsize_of_parameters); // >=0
998 __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize!
999 __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize);
1000 __ add(parent_frame_resize, parent_frame_resize, R11_scratch1);
1001 }
1002
1003 // Compute top frame size.
1004 __ addi(top_frame_size, top_frame_size, frame::abi_reg_args_size + frame::ijava_state_size);
1005
1006 // Cut back area between esp and max_stack.
1007 __ addi(parent_frame_resize, parent_frame_resize, frame::abi_minframe_size - Interpreter::stackElementSize);
1008
1009 __ round_to(top_frame_size, frame::alignment_in_bytes);
1010 __ round_to(parent_frame_resize, frame::alignment_in_bytes);
1011 // parent_frame_resize = (locals-parameters) - (ESP-SP-ABI48) Rounded to frame alignment size.
1012 // Enlarge by locals-parameters (not in case of native_call), shrink by ESP-SP-ABI48.
1013
1014 if (!native_call) {
1015 // Stack overflow check.
1016 // Native calls don't need the stack size check since they have no
1017 // expression stack and the arguments are already on the stack and
1018 // we only add a handful of words to the stack.
1019 __ add(R11_scratch1, parent_frame_resize, top_frame_size);
1020 generate_stack_overflow_check(R11_scratch1, R12_scratch2);
1021 }
1022
1023 // Set up interpreter state registers.
1024
1025 __ add(R18_locals, R15_esp, Rsize_of_parameters);
1026 __ ld(R27_constPoolCache, in_bytes(ConstMethod::constants_offset()), Rconst_method);
1027 __ ld(R27_constPoolCache, ConstantPool::cache_offset_in_bytes(), R27_constPoolCache);
1028
1029 // Set method data pointer.
1030 if (ProfileInterpreter) {
1031 Label zero_continue;
1032 __ ld(R28_mdx, method_(method_data));
1033 __ cmpdi(CCR0, R28_mdx, 0);
1034 __ beq(CCR0, zero_continue);
1035 __ addi(R28_mdx, R28_mdx, in_bytes(MethodData::data_offset()));
1036 __ bind(zero_continue);
1037 }
1038
1039 if (native_call) {
1040 __ li(R14_bcp, 0); // Must initialize.
1041 } else {
1042 __ add(R14_bcp, in_bytes(ConstMethod::codes_offset()), Rconst_method);
1043 }
1044
1045 // Resize parent frame.
1046 __ mflr(R12_scratch2);
1047 __ neg(parent_frame_resize, parent_frame_resize);
1048 __ resize_frame(parent_frame_resize, R11_scratch1);
1049 __ std(R12_scratch2, _abi(lr), R1_SP);
1050
1051 // Get mirror and store it in the frame as GC root for this Method*.
1052 __ load_mirror_from_const_method(R12_scratch2, Rconst_method);
1053
1054 __ addi(R26_monitor, R1_SP, - frame::ijava_state_size);
1055 __ addi(R15_esp, R26_monitor, - Interpreter::stackElementSize);
1056
1057 // Store values.
1058 // R15_esp, R14_bcp, R26_monitor, R28_mdx are saved at java calls
1059 // in InterpreterMacroAssembler::call_from_interpreter.
1060 __ std(R19_method, _ijava_state_neg(method), R1_SP);
1061 __ std(R12_scratch2, _ijava_state_neg(mirror), R1_SP);
1062 __ std(R21_sender_SP, _ijava_state_neg(sender_sp), R1_SP);
1063 __ std(R27_constPoolCache, _ijava_state_neg(cpoolCache), R1_SP);
1064 __ std(R18_locals, _ijava_state_neg(locals), R1_SP);
1065
1066 // Note: esp, bcp, monitor, mdx live in registers. Hence, the correct version can only
1067 // be found in the frame after save_interpreter_state is done. This is always true
1068 // for non-top frames. But when a signal occurs, dumping the top frame can go wrong,
1069 // because e.g. frame::interpreter_frame_bcp() will not access the correct value
1070 // (Enhanced Stack Trace).
1071 // The signal handler does not save the interpreter state into the frame.
1072 __ li(R0, 0);
1073 #ifdef ASSERT
1074 // Fill remaining slots with constants.
1075 __ load_const_optimized(R11_scratch1, 0x5afe);
1076 __ load_const_optimized(R12_scratch2, 0xdead);
1077 #endif
1078 // We have to initialize some frame slots for native calls (accessed by GC).
1079 if (native_call) {
1080 __ std(R26_monitor, _ijava_state_neg(monitors), R1_SP);
1081 __ std(R14_bcp, _ijava_state_neg(bcp), R1_SP);
1082 if (ProfileInterpreter) { __ std(R28_mdx, _ijava_state_neg(mdx), R1_SP); }
1083 }
1084 #ifdef ASSERT
1085 else {
1086 __ std(R12_scratch2, _ijava_state_neg(monitors), R1_SP);
1087 __ std(R12_scratch2, _ijava_state_neg(bcp), R1_SP);
1088 __ std(R12_scratch2, _ijava_state_neg(mdx), R1_SP);
1089 }
1090 __ std(R11_scratch1, _ijava_state_neg(ijava_reserved), R1_SP);
1091 __ std(R12_scratch2, _ijava_state_neg(esp), R1_SP);
1092 __ std(R12_scratch2, _ijava_state_neg(lresult), R1_SP);
1093 __ std(R12_scratch2, _ijava_state_neg(fresult), R1_SP);
1094 #endif
1095 __ subf(R12_scratch2, top_frame_size, R1_SP);
1096 __ std(R0, _ijava_state_neg(oop_tmp), R1_SP);
1097 __ std(R12_scratch2, _ijava_state_neg(top_frame_sp), R1_SP);
1098
1099 // Push top frame.
1100 __ push_frame(top_frame_size, R11_scratch1);
1101 }
1102
1103 // End of helpers
1104
1105 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
1106
1107 // Decide what to do: Use same platform specific instructions and runtime calls as compilers.
1108 bool use_instruction = false;
1109 address runtime_entry = NULL;
1110 int num_args = 1;
1111 bool double_precision = true;
1112
1113 // PPC64 specific:
1114 switch (kind) {
1115 case Interpreter::java_lang_math_sqrt: use_instruction = VM_Version::has_fsqrt(); break;
1116 case Interpreter::java_lang_math_abs: use_instruction = true; break;
1117 case Interpreter::java_lang_math_fmaF:
1118 case Interpreter::java_lang_math_fmaD: use_instruction = UseFMA; break;
1119 default: break; // Fall back to runtime call.
1120 }
1121
1122 switch (kind) {
1123 case Interpreter::java_lang_math_sin : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dsin); break;
1124 case Interpreter::java_lang_math_cos : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dcos); break;
1125 case Interpreter::java_lang_math_tan : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dtan); break;
1126 case Interpreter::java_lang_math_abs : /* run interpreted */ break;
1127 case Interpreter::java_lang_math_sqrt : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dsqrt); break;
1128 case Interpreter::java_lang_math_log : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dlog); break;
1129 case Interpreter::java_lang_math_log10: runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dlog10); break;
1130 case Interpreter::java_lang_math_pow : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dpow); num_args = 2; break;
1131 case Interpreter::java_lang_math_exp : runtime_entry = CAST_FROM_FN_PTR(address, SharedRuntime::dexp); break;
1132 case Interpreter::java_lang_math_fmaF : /* run interpreted */ num_args = 3; double_precision = false; break;
1133 case Interpreter::java_lang_math_fmaD : /* run interpreted */ num_args = 3; break;
1134 default: ShouldNotReachHere();
1135 }
1136
1137 // Use normal entry if neither instruction nor runtime call is used.
1138 if (!use_instruction && runtime_entry == NULL) return NULL;
1139
1140 address entry = __ pc();
1141
1142 // Load arguments
1143 assert(num_args <= 13, "passed in registers");
1144 if (double_precision) {
1145 int offset = (2 * num_args - 1) * Interpreter::stackElementSize;
1146 for (int i = 0; i < num_args; ++i) {
1147 __ lfd(as_FloatRegister(F1_ARG1->encoding() + i), offset, R15_esp);
1148 offset -= 2 * Interpreter::stackElementSize;
1149 }
1150 } else {
1151 int offset = num_args * Interpreter::stackElementSize;
1152 for (int i = 0; i < num_args; ++i) {
1153 __ lfs(as_FloatRegister(F1_ARG1->encoding() + i), offset, R15_esp);
1154 offset -= Interpreter::stackElementSize;
1155 }
1156 }
1157
1158 // Pop c2i arguments (if any) off when we return.
1159 #ifdef ASSERT
1160 __ ld(R9_ARG7, 0, R1_SP);
1161 __ ld(R10_ARG8, 0, R21_sender_SP);
1162 __ cmpd(CCR0, R9_ARG7, R10_ARG8);
1163 __ asm_assert_eq("backlink", 0x545);
1164 #endif // ASSERT
1165 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1166
1167 if (use_instruction) {
1168 switch (kind) {
1169 case Interpreter::java_lang_math_sqrt: __ fsqrt(F1_RET, F1); break;
1170 case Interpreter::java_lang_math_abs: __ fabs(F1_RET, F1); break;
1171 case Interpreter::java_lang_math_fmaF: __ fmadds(F1_RET, F1, F2, F3); break;
1172 case Interpreter::java_lang_math_fmaD: __ fmadd(F1_RET, F1, F2, F3); break;
1173 default: ShouldNotReachHere();
1174 }
1175 } else {
1176 // Comment: Can use tail call if the unextended frame is always C ABI compliant:
1177 //__ load_const_optimized(R12_scratch2, runtime_entry, R0);
1178 //__ call_c_and_return_to_caller(R12_scratch2);
1179
1180 // Push a new C frame and save LR.
1181 __ save_LR_CR(R0);
1182 __ push_frame_reg_args(0, R11_scratch1);
1183
1184 __ call_VM_leaf(runtime_entry);
1185
1186 // Pop the C frame and restore LR.
1187 __ pop_frame();
1188 __ restore_LR_CR(R0);
1189 }
1190
1191 __ blr();
1192
1193 __ flush();
1194
1195 return entry;
1196 }
1197
1198 void TemplateInterpreterGenerator::bang_stack_shadow_pages(bool native_call) {
1199 // Quick & dirty stack overflow checking: bang the stack & handle trap.
1200 // Note that we do the banging after the frame is setup, since the exception
1201 // handling code expects to find a valid interpreter frame on the stack.
1202 // Doing the banging earlier fails if the caller frame is not an interpreter
1203 // frame.
1204 // (Also, the exception throwing code expects to unlock any synchronized
1205 // method receiever, so do the banging after locking the receiver.)
1206
1207 // Bang each page in the shadow zone. We can't assume it's been done for
1208 // an interpreter frame with greater than a page of locals, so each page
1209 // needs to be checked. Only true for non-native.
1210 if (UseStackBanging) {
1211 const int page_size = os::vm_page_size();
1212 const int n_shadow_pages = ((int)JavaThread::stack_shadow_zone_size()) / page_size;
1213 const int start_page = native_call ? n_shadow_pages : 1;
1214 BLOCK_COMMENT("bang_stack_shadow_pages:");
1215 for (int pages = start_page; pages <= n_shadow_pages; pages++) {
1216 __ bang_stack_with_offset(pages*page_size);
1217 }
1218 }
1219 }
1220
1221 // Interpreter stub for calling a native method. (asm interpreter)
1222 // This sets up a somewhat different looking stack for calling the
1223 // native method than the typical interpreter frame setup.
1224 //
1225 // On entry:
1226 // R19_method - method
1227 // R16_thread - JavaThread*
1228 // R15_esp - intptr_t* sender tos
1229 //
1230 // abstract stack (grows up)
1231 // [ IJava (caller of JNI callee) ] <-- ASP
1232 // ...
1233 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
1234
1235 address entry = __ pc();
1236
1237 const bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1238
1239 // -----------------------------------------------------------------------------
1240 // Allocate a new frame that represents the native callee (i2n frame).
1241 // This is not a full-blown interpreter frame, but in particular, the
1242 // following registers are valid after this:
1243 // - R19_method
1244 // - R18_local (points to start of arguments to native function)
1245 //
1246 // abstract stack (grows up)
1247 // [ IJava (caller of JNI callee) ] <-- ASP
1248 // ...
1249
1250 const Register signature_handler_fd = R11_scratch1;
1251 const Register pending_exception = R0;
1252 const Register result_handler_addr = R31;
1253 const Register native_method_fd = R11_scratch1;
1254 const Register access_flags = R22_tmp2;
1255 const Register active_handles = R11_scratch1; // R26_monitor saved to state.
1256 const Register sync_state = R12_scratch2;
1257 const Register sync_state_addr = sync_state; // Address is dead after use.
1258 const Register suspend_flags = R11_scratch1;
1259
1260 //=============================================================================
1261 // Allocate new frame and initialize interpreter state.
1262
1263 Label exception_return;
1264 Label exception_return_sync_check;
1265 Label stack_overflow_return;
1266
1267 // Generate new interpreter state and jump to stack_overflow_return in case of
1268 // a stack overflow.
1269 //generate_compute_interpreter_state(stack_overflow_return);
1270
1271 Register size_of_parameters = R22_tmp2;
1272
1273 generate_fixed_frame(true, size_of_parameters, noreg /* unused */);
1274
1275 //=============================================================================
1276 // Increment invocation counter. On overflow, entry to JNI method
1277 // will be compiled.
1278 Label invocation_counter_overflow, continue_after_compile;
1279 if (inc_counter) {
1280 if (synchronized) {
1281 // Since at this point in the method invocation the exception handler
1282 // would try to exit the monitor of synchronized methods which hasn't
1283 // been entered yet, we set the thread local variable
1284 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1285 // runtime, exception handling i.e. unlock_if_synchronized_method will
1286 // check this thread local flag.
1287 // This flag has two effects, one is to force an unwind in the topmost
1288 // interpreter frame and not perform an unlock while doing so.
1289 __ li(R0, 1);
1290 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1291 }
1292 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
1293
1294 BIND(continue_after_compile);
1295 }
1296
1297 bang_stack_shadow_pages(true);
1298
1299 if (inc_counter) {
1300 // Reset the _do_not_unlock_if_synchronized flag.
1301 if (synchronized) {
1302 __ li(R0, 0);
1303 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1304 }
1305 }
1306
1307 // access_flags = method->access_flags();
1308 // Load access flags.
1309 assert(access_flags->is_nonvolatile(),
1310 "access_flags must be in a non-volatile register");
1311 // Type check.
1312 assert(4 == sizeof(AccessFlags), "unexpected field size");
1313 __ lwz(access_flags, method_(access_flags));
1314
1315 // We don't want to reload R19_method and access_flags after calls
1316 // to some helper functions.
1317 assert(R19_method->is_nonvolatile(),
1318 "R19_method must be a non-volatile register");
1319
1320 // Check for synchronized methods. Must happen AFTER invocation counter
1321 // check, so method is not locked if counter overflows.
1322
1323 if (synchronized) {
1324 lock_method(access_flags, R11_scratch1, R12_scratch2, true);
1325
1326 // Update monitor in state.
1327 __ ld(R11_scratch1, 0, R1_SP);
1328 __ std(R26_monitor, _ijava_state_neg(monitors), R11_scratch1);
1329 }
1330
1331 // jvmti/jvmpi support
1332 __ notify_method_entry();
1333
1334 //=============================================================================
1335 // Get and call the signature handler.
1336
1337 __ ld(signature_handler_fd, method_(signature_handler));
1338 Label call_signature_handler;
1339
1340 __ cmpdi(CCR0, signature_handler_fd, 0);
1341 __ bne(CCR0, call_signature_handler);
1342
1343 // Method has never been called. Either generate a specialized
1344 // handler or point to the slow one.
1345 //
1346 // Pass parameter 'false' to avoid exception check in call_VM.
1347 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), R19_method, false);
1348
1349 // Check for an exception while looking up the target method. If we
1350 // incurred one, bail.
1351 __ ld(pending_exception, thread_(pending_exception));
1352 __ cmpdi(CCR0, pending_exception, 0);
1353 __ bne(CCR0, exception_return_sync_check); // Has pending exception.
1354
1355 // Reload signature handler, it may have been created/assigned in the meanwhile.
1356 __ ld(signature_handler_fd, method_(signature_handler));
1357 __ twi_0(signature_handler_fd); // Order wrt. load of klass mirror and entry point (isync is below).
1358
1359 BIND(call_signature_handler);
1360
1361 // Before we call the signature handler we push a new frame to
1362 // protect the interpreter frame volatile registers when we return
1363 // from jni but before we can get back to Java.
1364
1365 // First set the frame anchor while the SP/FP registers are
1366 // convenient and the slow signature handler can use this same frame
1367 // anchor.
1368
1369 // We have a TOP_IJAVA_FRAME here, which belongs to us.
1370 __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/);
1371
1372 // Now the interpreter frame (and its call chain) have been
1373 // invalidated and flushed. We are now protected against eager
1374 // being enabled in native code. Even if it goes eager the
1375 // registers will be reloaded as clean and we will invalidate after
1376 // the call so no spurious flush should be possible.
1377
1378 // Call signature handler and pass locals address.
1379 //
1380 // Our signature handlers copy required arguments to the C stack
1381 // (outgoing C args), R3_ARG1 to R10_ARG8, and FARG1 to FARG13.
1382 __ mr(R3_ARG1, R18_locals);
1383 #if !defined(ABI_ELFv2)
1384 __ ld(signature_handler_fd, 0, signature_handler_fd);
1385 #endif
1386
1387 __ call_stub(signature_handler_fd);
1388
1389 // Remove the register parameter varargs slots we allocated in
1390 // compute_interpreter_state. SP+16 ends up pointing to the ABI
1391 // outgoing argument area.
1392 //
1393 // Not needed on PPC64.
1394 //__ add(SP, SP, Argument::n_register_parameters*BytesPerWord);
1395
1396 assert(result_handler_addr->is_nonvolatile(), "result_handler_addr must be in a non-volatile register");
1397 // Save across call to native method.
1398 __ mr(result_handler_addr, R3_RET);
1399
1400 __ isync(); // Acquire signature handler before trying to fetch the native entry point and klass mirror.
1401
1402 // Set up fixed parameters and call the native method.
1403 // If the method is static, get mirror into R4_ARG2.
1404 {
1405 Label method_is_not_static;
1406 // Access_flags is non-volatile and still, no need to restore it.
1407
1408 // Restore access flags.
1409 __ testbitdi(CCR0, R0, access_flags, JVM_ACC_STATIC_BIT);
1410 __ bfalse(CCR0, method_is_not_static);
1411
1412 __ ld(R11_scratch1, _abi(callers_sp), R1_SP);
1413 // Load mirror from interpreter frame.
1414 __ ld(R12_scratch2, _ijava_state_neg(mirror), R11_scratch1);
1415 // R4_ARG2 = &state->_oop_temp;
1416 __ addi(R4_ARG2, R11_scratch1, _ijava_state_neg(oop_tmp));
1417 __ std(R12_scratch2/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1);
1418 BIND(method_is_not_static);
1419 }
1420
1421 // At this point, arguments have been copied off the stack into
1422 // their JNI positions. Oops are boxed in-place on the stack, with
1423 // handles copied to arguments. The result handler address is in a
1424 // register.
1425
1426 // Pass JNIEnv address as first parameter.
1427 __ addir(R3_ARG1, thread_(jni_environment));
1428
1429 // Load the native_method entry before we change the thread state.
1430 __ ld(native_method_fd, method_(native_function));
1431
1432 //=============================================================================
1433 // Transition from _thread_in_Java to _thread_in_native. As soon as
1434 // we make this change the safepoint code needs to be certain that
1435 // the last Java frame we established is good. The pc in that frame
1436 // just needs to be near here not an actual return address.
1437
1438 // We use release_store_fence to update values like the thread state, where
1439 // we don't want the current thread to continue until all our prior memory
1440 // accesses (including the new thread state) are visible to other threads.
1441 __ li(R0, _thread_in_native);
1442 __ release();
1443
1444 // TODO PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
1445 __ stw(R0, thread_(thread_state));
1446
1447 //=============================================================================
1448 // Call the native method. Argument registers must not have been
1449 // overwritten since "__ call_stub(signature_handler);" (except for
1450 // ARG1 and ARG2 for static methods).
1451 __ call_c(native_method_fd);
1452
1453 __ li(R0, 0);
1454 __ ld(R11_scratch1, 0, R1_SP);
1455 __ std(R3_RET, _ijava_state_neg(lresult), R11_scratch1);
1456 __ stfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1);
1457 __ std(R0/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1); // reset
1458
1459 // Note: C++ interpreter needs the following here:
1460 // The frame_manager_lr field, which we use for setting the last
1461 // java frame, gets overwritten by the signature handler. Restore
1462 // it now.
1463 //__ get_PC_trash_LR(R11_scratch1);
1464 //__ std(R11_scratch1, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
1465
1466 // Because of GC R19_method may no longer be valid.
1467
1468 // Block, if necessary, before resuming in _thread_in_Java state.
1469 // In order for GC to work, don't clear the last_Java_sp until after
1470 // blocking.
1471
1472 //=============================================================================
1473 // Switch thread to "native transition" state before reading the
1474 // synchronization state. This additional state is necessary
1475 // because reading and testing the synchronization state is not
1476 // atomic w.r.t. GC, as this scenario demonstrates: Java thread A,
1477 // in _thread_in_native state, loads _not_synchronized and is
1478 // preempted. VM thread changes sync state to synchronizing and
1479 // suspends threads for GC. Thread A is resumed to finish this
1480 // native method, but doesn't block here since it didn't see any
1481 // synchronization in progress, and escapes.
1482
1483 // We use release_store_fence to update values like the thread state, where
1484 // we don't want the current thread to continue until all our prior memory
1485 // accesses (including the new thread state) are visible to other threads.
1486 __ li(R0/*thread_state*/, _thread_in_native_trans);
1487 __ release();
1488 __ stw(R0/*thread_state*/, thread_(thread_state));
1489 __ fence();
1490
1491 // Now before we return to java we must look for a current safepoint
1492 // (a new safepoint can not start since we entered native_trans).
1493 // We must check here because a current safepoint could be modifying
1494 // the callers registers right this moment.
1495
1496 // Acquire isn't strictly necessary here because of the fence, but
1497 // sync_state is declared to be volatile, so we do it anyway
1498 // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
1499
1500 Label do_safepoint, sync_check_done;
1501 // No synchronization in progress nor yet synchronized.
1502 __ safepoint_poll(do_safepoint, sync_state);
1503
1504 // Not suspended.
1505 // TODO PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
1506 __ lwz(suspend_flags, thread_(suspend_flags));
1507 __ cmpwi(CCR1, suspend_flags, 0);
1508 __ beq(CCR1, sync_check_done);
1509
1510 __ bind(do_safepoint);
1511 __ isync();
1512 // Block. We do the call directly and leave the current
1513 // last_Java_frame setup undisturbed. We must save any possible
1514 // native result across the call. No oop is present.
1515
1516 __ mr(R3_ARG1, R16_thread);
1517 #if defined(ABI_ELFv2)
1518 __ call_c(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
1519 relocInfo::none);
1520 #else
1521 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, JavaThread::check_special_condition_for_native_trans),
1522 relocInfo::none);
1523 #endif
1524
1525 __ bind(sync_check_done);
1526
1527 //=============================================================================
1528 // <<<<<< Back in Interpreter Frame >>>>>
1529
1530 // We are in thread_in_native_trans here and back in the normal
1531 // interpreter frame. We don't have to do anything special about
1532 // safepoints and we can switch to Java mode anytime we are ready.
1533
1534 // Note: frame::interpreter_frame_result has a dependency on how the
1535 // method result is saved across the call to post_method_exit. For
1536 // native methods it assumes that the non-FPU/non-void result is
1537 // saved in _native_lresult and a FPU result in _native_fresult. If
1538 // this changes then the interpreter_frame_result implementation
1539 // will need to be updated too.
1540
1541 // On PPC64, we have stored the result directly after the native call.
1542
1543 //=============================================================================
1544 // Back in Java
1545
1546 // We use release_store_fence to update values like the thread state, where
1547 // we don't want the current thread to continue until all our prior memory
1548 // accesses (including the new thread state) are visible to other threads.
1549 __ li(R0/*thread_state*/, _thread_in_Java);
1550 __ lwsync(); // Acquire safepoint and suspend state, release thread state.
1551 __ stw(R0/*thread_state*/, thread_(thread_state));
1552
1553 if (CheckJNICalls) {
1554 // clear_pending_jni_exception_check
1555 __ load_const_optimized(R0, 0L);
1556 __ st_ptr(R0, JavaThread::pending_jni_exception_check_fn_offset(), R16_thread);
1557 }
1558
1559 __ reset_last_Java_frame();
1560
1561 // Jvmdi/jvmpi support. Whether we've got an exception pending or
1562 // not, and whether unlocking throws an exception or not, we notify
1563 // on native method exit. If we do have an exception, we'll end up
1564 // in the caller's context to handle it, so if we don't do the
1565 // notify here, we'll drop it on the floor.
1566 __ notify_method_exit(true/*native method*/,
1567 ilgl /*illegal state (not used for native methods)*/,
1568 InterpreterMacroAssembler::NotifyJVMTI,
1569 false /*check_exceptions*/);
1570
1571 //=============================================================================
1572 // Handle exceptions
1573
1574 if (synchronized) {
1575 // Don't check for exceptions since we're still in the i2n frame. Do that
1576 // manually afterwards.
1577 __ unlock_object(R26_monitor, false); // Can also unlock methods.
1578 }
1579
1580 // Reset active handles after returning from native.
1581 // thread->active_handles()->clear();
1582 __ ld(active_handles, thread_(active_handles));
1583 // TODO PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
1584 __ li(R0, 0);
1585 __ stw(R0, JNIHandleBlock::top_offset_in_bytes(), active_handles);
1586
1587 Label exception_return_sync_check_already_unlocked;
1588 __ ld(R0/*pending_exception*/, thread_(pending_exception));
1589 __ cmpdi(CCR0, R0/*pending_exception*/, 0);
1590 __ bne(CCR0, exception_return_sync_check_already_unlocked);
1591
1592 //-----------------------------------------------------------------------------
1593 // No exception pending.
1594
1595 // Move native method result back into proper registers and return.
1596 // Invoke result handler (may unbox/promote).
1597 __ ld(R11_scratch1, 0, R1_SP);
1598 __ ld(R3_RET, _ijava_state_neg(lresult), R11_scratch1);
1599 __ lfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1);
1600 __ call_stub(result_handler_addr);
1601
1602 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ R0, R11_scratch1, R12_scratch2);
1603
1604 // Must use the return pc which was loaded from the caller's frame
1605 // as the VM uses return-pc-patching for deoptimization.
1606 __ mtlr(R0);
1607 __ blr();
1608
1609 //-----------------------------------------------------------------------------
1610 // An exception is pending. We call into the runtime only if the
1611 // caller was not interpreted. If it was interpreted the
1612 // interpreter will do the correct thing. If it isn't interpreted
1613 // (call stub/compiled code) we will change our return and continue.
1614
1615 BIND(exception_return_sync_check);
1616
1617 if (synchronized) {
1618 // Don't check for exceptions since we're still in the i2n frame. Do that
1619 // manually afterwards.
1620 __ unlock_object(R26_monitor, false); // Can also unlock methods.
1621 }
1622 BIND(exception_return_sync_check_already_unlocked);
1623
1624 const Register return_pc = R31;
1625
1626 __ ld(return_pc, 0, R1_SP);
1627 __ ld(return_pc, _abi(lr), return_pc);
1628
1629 // Get the address of the exception handler.
1630 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1631 R16_thread,
1632 return_pc /* return pc */);
1633 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, noreg, R11_scratch1, R12_scratch2);
1634
1635 // Load the PC of the the exception handler into LR.
1636 __ mtlr(R3_RET);
1637
1638 // Load exception into R3_ARG1 and clear pending exception in thread.
1639 __ ld(R3_ARG1/*exception*/, thread_(pending_exception));
1640 __ li(R4_ARG2, 0);
1641 __ std(R4_ARG2, thread_(pending_exception));
1642
1643 // Load the original return pc into R4_ARG2.
1644 __ mr(R4_ARG2/*issuing_pc*/, return_pc);
1645
1646 // Return to exception handler.
1647 __ blr();
1648
1649 //=============================================================================
1650 // Counter overflow.
1651
1652 if (inc_counter) {
1653 // Handle invocation counter overflow.
1654 __ bind(invocation_counter_overflow);
1655
1656 generate_counter_overflow(continue_after_compile);
1657 }
1658
1659 return entry;
1660 }
1661
1662 // Generic interpreted method entry to (asm) interpreter.
1663 //
1664 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1665 bool inc_counter = UseCompiler || CountCompiledCalls || LogTouchedMethods;
1666 address entry = __ pc();
1667 // Generate the code to allocate the interpreter stack frame.
1668 Register Rsize_of_parameters = R4_ARG2, // Written by generate_fixed_frame.
1669 Rsize_of_locals = R5_ARG3; // Written by generate_fixed_frame.
1670
1671 // Does also a stack check to assure this frame fits on the stack.
1672 generate_fixed_frame(false, Rsize_of_parameters, Rsize_of_locals);
1673
1674 // --------------------------------------------------------------------------
1675 // Zero out non-parameter locals.
1676 // Note: *Always* zero out non-parameter locals as Sparc does. It's not
1677 // worth to ask the flag, just do it.
1678 Register Rslot_addr = R6_ARG4,
1679 Rnum = R7_ARG5;
1680 Label Lno_locals, Lzero_loop;
1681
1682 // Set up the zeroing loop.
1683 __ subf(Rnum, Rsize_of_parameters, Rsize_of_locals);
1684 __ subf(Rslot_addr, Rsize_of_parameters, R18_locals);
1685 __ srdi_(Rnum, Rnum, Interpreter::logStackElementSize);
1686 __ beq(CCR0, Lno_locals);
1687 __ li(R0, 0);
1688 __ mtctr(Rnum);
1689
1690 // The zero locals loop.
1691 __ bind(Lzero_loop);
1692 __ std(R0, 0, Rslot_addr);
1693 __ addi(Rslot_addr, Rslot_addr, -Interpreter::stackElementSize);
1694 __ bdnz(Lzero_loop);
1695
1696 __ bind(Lno_locals);
1697
1698 // --------------------------------------------------------------------------
1699 // Counter increment and overflow check.
1700 Label invocation_counter_overflow,
1701 profile_method,
1702 profile_method_continue;
1703 if (inc_counter || ProfileInterpreter) {
1704
1705 Register Rdo_not_unlock_if_synchronized_addr = R11_scratch1;
1706 if (synchronized) {
1707 // Since at this point in the method invocation the exception handler
1708 // would try to exit the monitor of synchronized methods which hasn't
1709 // been entered yet, we set the thread local variable
1710 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1711 // runtime, exception handling i.e. unlock_if_synchronized_method will
1712 // check this thread local flag.
1713 // This flag has two effects, one is to force an unwind in the topmost
1714 // interpreter frame and not perform an unlock while doing so.
1715 __ li(R0, 1);
1716 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1717 }
1718
1719 // Argument and return type profiling.
1720 __ profile_parameters_type(R3_ARG1, R4_ARG2, R5_ARG3, R6_ARG4);
1721
1722 // Increment invocation counter and check for overflow.
1723 if (inc_counter) {
1724 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1725 }
1726
1727 __ bind(profile_method_continue);
1728 }
1729
1730 bang_stack_shadow_pages(false);
1731
1732 if (inc_counter || ProfileInterpreter) {
1733 // Reset the _do_not_unlock_if_synchronized flag.
1734 if (synchronized) {
1735 __ li(R0, 0);
1736 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1737 }
1738 }
1739
1740 // --------------------------------------------------------------------------
1741 // Locking of synchronized methods. Must happen AFTER invocation_counter
1742 // check and stack overflow check, so method is not locked if overflows.
1743 if (synchronized) {
1744 lock_method(R3_ARG1, R4_ARG2, R5_ARG3);
1745 }
1746 #ifdef ASSERT
1747 else {
1748 Label Lok;
1749 __ lwz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1750 __ andi_(R0, R0, JVM_ACC_SYNCHRONIZED);
1751 __ asm_assert_eq("method needs synchronization", 0x8521);
1752 __ bind(Lok);
1753 }
1754 #endif // ASSERT
1755
1756 __ verify_thread();
1757
1758 // --------------------------------------------------------------------------
1759 // JVMTI support
1760 __ notify_method_entry();
1761
1762 // --------------------------------------------------------------------------
1763 // Start executing instructions.
1764 __ dispatch_next(vtos);
1765
1766 // --------------------------------------------------------------------------
1767 // Out of line counter overflow and MDO creation code.
1768 if (ProfileInterpreter) {
1769 // We have decided to profile this method in the interpreter.
1770 __ bind(profile_method);
1771 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1772 __ set_method_data_pointer_for_bcp();
1773 __ b(profile_method_continue);
1774 }
1775
1776 if (inc_counter) {
1777 // Handle invocation counter overflow.
1778 __ bind(invocation_counter_overflow);
1779 generate_counter_overflow(profile_method_continue);
1780 }
1781 return entry;
1782 }
1783
1784 // CRC32 Intrinsics.
1785 //
1786 // Contract on scratch and work registers.
1787 // =======================================
1788 //
1789 // On ppc, the register set {R2..R12} is available in the interpreter as scratch/work registers.
1790 // You should, however, keep in mind that {R3_ARG1..R10_ARG8} is the C-ABI argument register set.
1791 // You can't rely on these registers across calls.
1792 //
1793 // The generators for CRC32_update and for CRC32_updateBytes use the
1794 // scratch/work register set internally, passing the work registers
1795 // as arguments to the MacroAssembler emitters as required.
1796 //
1797 // R3_ARG1..R6_ARG4 are preset to hold the incoming java arguments.
1798 // Their contents is not constant but may change according to the requirements
1799 // of the emitted code.
1800 //
1801 // All other registers from the scratch/work register set are used "internally"
1802 // and contain garbage (i.e. unpredictable values) once blr() is reached.
1803 // Basically, only R3_RET contains a defined value which is the function result.
1804 //
1805 /**
1806 * Method entry for static native methods:
1807 * int java.util.zip.CRC32.update(int crc, int b)
1808 */
1809 address TemplateInterpreterGenerator::generate_CRC32_update_entry() {
1810 if (UseCRC32Intrinsics) {
1811 address start = __ pc(); // Remember stub start address (is rtn value).
1812 Label slow_path;
1813
1814 // Safepoint check
1815 const Register sync_state = R11_scratch1;
1816 __ safepoint_poll(slow_path, sync_state);
1817
1818 // We don't generate local frame and don't align stack because
1819 // we not even call stub code (we generate the code inline)
1820 // and there is no safepoint on this path.
1821
1822 // Load java parameters.
1823 // R15_esp is callers operand stack pointer, i.e. it points to the parameters.
1824 const Register argP = R15_esp;
1825 const Register crc = R3_ARG1; // crc value
1826 const Register data = R4_ARG2; // address of java byte value (kernel_crc32 needs address)
1827 const Register dataLen = R5_ARG3; // source data len (1 byte). Not used because calling the single-byte emitter.
1828 const Register table = R6_ARG4; // address of crc32 table
1829 const Register tmp = dataLen; // Reuse unused len register to show we don't actually need a separate tmp here.
1830
1831 BLOCK_COMMENT("CRC32_update {");
1832
1833 // Arguments are reversed on java expression stack
1834 #ifdef VM_LITTLE_ENDIAN
1835 __ addi(data, argP, 0+1*wordSize); // (stack) address of byte value. Emitter expects address, not value.
1836 // Being passed as an int, the single byte is at offset +0.
1837 #else
1838 __ addi(data, argP, 3+1*wordSize); // (stack) address of byte value. Emitter expects address, not value.
1839 // Being passed from java as an int, the single byte is at offset +3.
1840 #endif
1841 __ lwz(crc, 2*wordSize, argP); // Current crc state, zero extend to 64 bit to have a clean register.
1842
1843 StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table);
1844 __ kernel_crc32_singleByte(crc, data, dataLen, table, tmp, true);
1845
1846 // Restore caller sp for c2i case and return.
1847 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1848 __ blr();
1849
1850 // Generate a vanilla native entry as the slow path.
1851 BLOCK_COMMENT("} CRC32_update");
1852 BIND(slow_path);
1853 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native), R11_scratch1);
1854 return start;
1855 }
1856
1857 return NULL;
1858 }
1859
1860 /**
1861 * Method entry for static native methods:
1862 * int java.util.zip.CRC32.updateBytes( int crc, byte[] b, int off, int len)
1863 * int java.util.zip.CRC32.updateByteBuffer(int crc, long* buf, int off, int len)
1864 */
1865 address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1866 if (UseCRC32Intrinsics) {
1867 address start = __ pc(); // Remember stub start address (is rtn value).
1868 Label slow_path;
1869
1870 // Safepoint check
1871 const Register sync_state = R11_scratch1;
1872 __ safepoint_poll(slow_path, sync_state);
1873
1874 // We don't generate local frame and don't align stack because
1875 // we not even call stub code (we generate the code inline)
1876 // and there is no safepoint on this path.
1877
1878 // Load parameters.
1879 // Z_esp is callers operand stack pointer, i.e. it points to the parameters.
1880 const Register argP = R15_esp;
1881 const Register crc = R3_ARG1; // crc value
1882 const Register data = R4_ARG2; // address of java byte array
1883 const Register dataLen = R5_ARG3; // source data len
1884 const Register table = R6_ARG4; // address of crc32 table
1885
1886 const Register t0 = R9; // scratch registers for crc calculation
1887 const Register t1 = R10;
1888 const Register t2 = R11;
1889 const Register t3 = R12;
1890
1891 const Register tc0 = R2; // registers to hold pre-calculated column addresses
1892 const Register tc1 = R7;
1893 const Register tc2 = R8;
1894 const Register tc3 = table; // table address is reconstructed at the end of kernel_crc32_* emitters
1895
1896 const Register tmp = t0; // Only used very locally to calculate byte buffer address.
1897
1898 // Arguments are reversed on java expression stack.
1899 // Calculate address of start element.
1900 if (kind == Interpreter::java_util_zip_CRC32_updateByteBuffer) { // Used for "updateByteBuffer direct".
1901 BLOCK_COMMENT("CRC32_updateByteBuffer {");
1902 // crc @ (SP + 5W) (32bit)
1903 // buf @ (SP + 3W) (64bit ptr to long array)
1904 // off @ (SP + 2W) (32bit)
1905 // dataLen @ (SP + 1W) (32bit)
1906 // data = buf + off
1907 __ ld( data, 3*wordSize, argP); // start of byte buffer
1908 __ lwa( tmp, 2*wordSize, argP); // byte buffer offset
1909 __ lwa( dataLen, 1*wordSize, argP); // #bytes to process
1910 __ lwz( crc, 5*wordSize, argP); // current crc state
1911 __ add( data, data, tmp); // Add byte buffer offset.
1912 } else { // Used for "updateBytes update".
1913 BLOCK_COMMENT("CRC32_updateBytes {");
1914 // crc @ (SP + 4W) (32bit)
1915 // buf @ (SP + 3W) (64bit ptr to byte array)
1916 // off @ (SP + 2W) (32bit)
1917 // dataLen @ (SP + 1W) (32bit)
1918 // data = buf + off + base_offset
1919 __ ld( data, 3*wordSize, argP); // start of byte buffer
1920 __ lwa( tmp, 2*wordSize, argP); // byte buffer offset
1921 __ lwa( dataLen, 1*wordSize, argP); // #bytes to process
1922 __ add( data, data, tmp); // add byte buffer offset
1923 __ lwz( crc, 4*wordSize, argP); // current crc state
1924 __ addi(data, data, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1925 }
1926
1927 StubRoutines::ppc64::generate_load_crc_table_addr(_masm, table);
1928
1929 // Performance measurements show the 1word and 2word variants to be almost equivalent,
1930 // with very light advantages for the 1word variant. We chose the 1word variant for
1931 // code compactness.
1932 __ kernel_crc32_1word(crc, data, dataLen, table, t0, t1, t2, t3, tc0, tc1, tc2, tc3, true);
1933
1934 // Restore caller sp for c2i case and return.
1935 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
1936 __ blr();
1937
1938 // Generate a vanilla native entry as the slow path.
1939 BLOCK_COMMENT("} CRC32_updateBytes(Buffer)");
1940 BIND(slow_path);
1941 __ jump_to_entry(Interpreter::entry_for_kind(Interpreter::native), R11_scratch1);
1942 return start;
1943 }
1944
1945 return NULL;
1946 }
1947
1948
1949 /**
1950 * Method entry for intrinsic-candidate (non-native) methods:
1951 * int java.util.zip.CRC32C.updateBytes( int crc, byte[] b, int off, int end)
1952 * int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long* buf, int off, int end)
1953 * Unlike CRC32, CRC32C does not have any methods marked as native
1954 * CRC32C also uses an "end" variable instead of the length variable CRC32 uses
1955 **/
1956 address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(AbstractInterpreter::MethodKind kind) {
1957 if (UseCRC32CIntrinsics) {
1958 address start = __ pc(); // Remember stub start address (is rtn value).
1959
1960 // We don't generate local frame and don't align stack because
1961 // we not even call stub code (we generate the code inline)
1962 // and there is no safepoint on this path.
1963
1964 // Load parameters.
1965 // Z_esp is callers operand stack pointer, i.e. it points to the parameters.
1966 const Register argP = R15_esp;
1967 const Register crc = R3_ARG1; // crc value
1968 const Register data = R4_ARG2; // address of java byte array
1969 const Register dataLen = R5_ARG3; // source data len
1970 const Register table = R6_ARG4; // address of crc32c table
1971
1972 const Register t0 = R9; // scratch registers for crc calculation
1973 const Register t1 = R10;
1974 const Register t2 = R11;
1975 const Register t3 = R12;
1976
1977 const Register tc0 = R2; // registers to hold pre-calculated column addresses
1978 const Register tc1 = R7;
1979 const Register tc2 = R8;
1980 const Register tc3 = table; // table address is reconstructed at the end of kernel_crc32_* emitters
1981
1982 const Register tmp = t0; // Only used very locally to calculate byte buffer address.
1983
1984 // Arguments are reversed on java expression stack.
1985 // Calculate address of start element.
1986 if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) { // Used for "updateDirectByteBuffer".
1987 BLOCK_COMMENT("CRC32C_updateDirectByteBuffer {");
1988 // crc @ (SP + 5W) (32bit)
1989 // buf @ (SP + 3W) (64bit ptr to long array)
1990 // off @ (SP + 2W) (32bit)
1991 // dataLen @ (SP + 1W) (32bit)
1992 // data = buf + off
1993 __ ld( data, 3*wordSize, argP); // start of byte buffer
1994 __ lwa( tmp, 2*wordSize, argP); // byte buffer offset
1995 __ lwa( dataLen, 1*wordSize, argP); // #bytes to process
1996 __ lwz( crc, 5*wordSize, argP); // current crc state
1997 __ add( data, data, tmp); // Add byte buffer offset.
1998 __ sub( dataLen, dataLen, tmp); // (end_index - offset)
1999 } else { // Used for "updateBytes update".
2000 BLOCK_COMMENT("CRC32C_updateBytes {");
2001 // crc @ (SP + 4W) (32bit)
2002 // buf @ (SP + 3W) (64bit ptr to byte array)
2003 // off @ (SP + 2W) (32bit)
2004 // dataLen @ (SP + 1W) (32bit)
2005 // data = buf + off + base_offset
2006 __ ld( data, 3*wordSize, argP); // start of byte buffer
2007 __ lwa( tmp, 2*wordSize, argP); // byte buffer offset
2008 __ lwa( dataLen, 1*wordSize, argP); // #bytes to process
2009 __ add( data, data, tmp); // add byte buffer offset
2010 __ sub( dataLen, dataLen, tmp); // (end_index - offset)
2011 __ lwz( crc, 4*wordSize, argP); // current crc state
2012 __ addi(data, data, arrayOopDesc::base_offset_in_bytes(T_BYTE));
2013 }
2014
2015 StubRoutines::ppc64::generate_load_crc32c_table_addr(_masm, table);
2016
2017 // Performance measurements show the 1word and 2word variants to be almost equivalent,
2018 // with very light advantages for the 1word variant. We chose the 1word variant for
2019 // code compactness.
2020 __ kernel_crc32_1word(crc, data, dataLen, table, t0, t1, t2, t3, tc0, tc1, tc2, tc3, false);
2021
2022 // Restore caller sp for c2i case and return.
2023 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
2024 __ blr();
2025
2026 BLOCK_COMMENT("} CRC32C_update{Bytes|DirectByteBuffer}");
2027 return start;
2028 }
2029
2030 return NULL;
2031 }
2032
2033 // =============================================================================
2034 // Exceptions
2035
2036 void TemplateInterpreterGenerator::generate_throw_exception() {
2037 Register Rexception = R17_tos,
2038 Rcontinuation = R3_RET;
2039
2040 // --------------------------------------------------------------------------
2041 // Entry point if an method returns with a pending exception (rethrow).
2042 Interpreter::_rethrow_exception_entry = __ pc();
2043 {
2044 __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp.
2045 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
2046 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
2047
2048 // Compiled code destroys templateTableBase, reload.
2049 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
2050 }
2051
2052 // Entry point if a interpreted method throws an exception (throw).
2053 Interpreter::_throw_exception_entry = __ pc();
2054 {
2055 __ mr(Rexception, R3_RET);
2056
2057 __ verify_thread();
2058 __ verify_oop(Rexception);
2059
2060 // Expression stack must be empty before entering the VM in case of an exception.
2061 __ empty_expression_stack();
2062 // Find exception handler address and preserve exception oop.
2063 // Call C routine to find handler and jump to it.
2064 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Rexception);
2065 __ mtctr(Rcontinuation);
2066 // Push exception for exception handler bytecodes.
2067 __ push_ptr(Rexception);
2068
2069 // Jump to exception handler (may be remove activation entry!).
2070 __ bctr();
2071 }
2072
2073 // If the exception is not handled in the current frame the frame is
2074 // removed and the exception is rethrown (i.e. exception
2075 // continuation is _rethrow_exception).
2076 //
2077 // Note: At this point the bci is still the bxi for the instruction
2078 // which caused the exception and the expression stack is
2079 // empty. Thus, for any VM calls at this point, GC will find a legal
2080 // oop map (with empty expression stack).
2081
2082 // In current activation
2083 // tos: exception
2084 // bcp: exception bcp
2085
2086 // --------------------------------------------------------------------------
2087 // JVMTI PopFrame support
2088
2089 Interpreter::_remove_activation_preserving_args_entry = __ pc();
2090 {
2091 // Set the popframe_processing bit in popframe_condition indicating that we are
2092 // currently handling popframe, so that call_VMs that may happen later do not
2093 // trigger new popframe handling cycles.
2094 __ lwz(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2095 __ ori(R11_scratch1, R11_scratch1, JavaThread::popframe_processing_bit);
2096 __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2097
2098 // Empty the expression stack, as in normal exception handling.
2099 __ empty_expression_stack();
2100 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
2101
2102 // Check to see whether we are returning to a deoptimized frame.
2103 // (The PopFrame call ensures that the caller of the popped frame is
2104 // either interpreted or compiled and deoptimizes it if compiled.)
2105 // Note that we don't compare the return PC against the
2106 // deoptimization blob's unpack entry because of the presence of
2107 // adapter frames in C2.
2108 Label Lcaller_not_deoptimized;
2109 Register return_pc = R3_ARG1;
2110 __ ld(return_pc, 0, R1_SP);
2111 __ ld(return_pc, _abi(lr), return_pc);
2112 __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), return_pc);
2113 __ cmpdi(CCR0, R3_RET, 0);
2114 __ bne(CCR0, Lcaller_not_deoptimized);
2115
2116 // The deoptimized case.
2117 // In this case, we can't call dispatch_next() after the frame is
2118 // popped, but instead must save the incoming arguments and restore
2119 // them after deoptimization has occurred.
2120 __ ld(R4_ARG2, in_bytes(Method::const_offset()), R19_method);
2121 __ lhz(R4_ARG2 /* number of params */, in_bytes(ConstMethod::size_of_parameters_offset()), R4_ARG2);
2122 __ slwi(R4_ARG2, R4_ARG2, Interpreter::logStackElementSize);
2123 __ addi(R5_ARG3, R18_locals, Interpreter::stackElementSize);
2124 __ subf(R5_ARG3, R4_ARG2, R5_ARG3);
2125 // Save these arguments.
2126 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), R16_thread, R4_ARG2, R5_ARG3);
2127
2128 // Inform deoptimization that it is responsible for restoring these arguments.
2129 __ load_const_optimized(R11_scratch1, JavaThread::popframe_force_deopt_reexecution_bit);
2130 __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2131
2132 // Return from the current method into the deoptimization blob. Will eventually
2133 // end up in the deopt interpeter entry, deoptimization prepared everything that
2134 // we will reexecute the call that called us.
2135 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*reload return_pc*/ return_pc, R11_scratch1, R12_scratch2);
2136 __ mtlr(return_pc);
2137 __ blr();
2138
2139 // The non-deoptimized case.
2140 __ bind(Lcaller_not_deoptimized);
2141
2142 // Clear the popframe condition flag.
2143 __ li(R0, 0);
2144 __ stw(R0, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
2145
2146 // Get out of the current method and re-execute the call that called us.
2147 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2);
2148 __ restore_interpreter_state(R11_scratch1);
2149 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
2150 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
2151 if (ProfileInterpreter) {
2152 __ set_method_data_pointer_for_bcp();
2153 __ ld(R11_scratch1, 0, R1_SP);
2154 __ std(R28_mdx, _ijava_state_neg(mdx), R11_scratch1);
2155 }
2156 #if INCLUDE_JVMTI
2157 Label L_done;
2158
2159 __ lbz(R11_scratch1, 0, R14_bcp);
2160 __ cmpwi(CCR0, R11_scratch1, Bytecodes::_invokestatic);
2161 __ bne(CCR0, L_done);
2162
2163 // The member name argument must be restored if _invokestatic is re-executed after a PopFrame call.
2164 // Detect such a case in the InterpreterRuntime function and return the member name argument, or NULL.
2165 __ ld(R4_ARG2, 0, R18_locals);
2166 __ MacroAssembler::call_VM(R4_ARG2, CAST_FROM_FN_PTR(address, InterpreterRuntime::member_name_arg_or_null), R4_ARG2, R19_method, R14_bcp, false);
2167 __ restore_interpreter_state(R11_scratch1, /*bcp_and_mdx_only*/ true);
2168 __ cmpdi(CCR0, R4_ARG2, 0);
2169 __ beq(CCR0, L_done);
2170 __ std(R4_ARG2, wordSize, R15_esp);
2171 __ bind(L_done);
2172 #endif // INCLUDE_JVMTI
2173 __ dispatch_next(vtos);
2174 }
2175 // end of JVMTI PopFrame support
2176
2177 // --------------------------------------------------------------------------
2178 // Remove activation exception entry.
2179 // This is jumped to if an interpreted method can't handle an exception itself
2180 // (we come from the throw/rethrow exception entry above). We're going to call
2181 // into the VM to find the exception handler in the caller, pop the current
2182 // frame and return the handler we calculated.
2183 Interpreter::_remove_activation_entry = __ pc();
2184 {
2185 __ pop_ptr(Rexception);
2186 __ verify_thread();
2187 __ verify_oop(Rexception);
2188 __ std(Rexception, in_bytes(JavaThread::vm_result_offset()), R16_thread);
2189
2190 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, true);
2191 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI, false);
2192
2193 __ get_vm_result(Rexception);
2194
2195 // We are done with this activation frame; find out where to go next.
2196 // The continuation point will be an exception handler, which expects
2197 // the following registers set up:
2198 //
2199 // RET: exception oop
2200 // ARG2: Issuing PC (see generate_exception_blob()), only used if the caller is compiled.
2201
2202 Register return_pc = R31; // Needs to survive the runtime call.
2203 __ ld(return_pc, 0, R1_SP);
2204 __ ld(return_pc, _abi(lr), return_pc);
2205 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, return_pc);
2206
2207 // Remove the current activation.
2208 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2);
2209
2210 __ mr(R4_ARG2, return_pc);
2211 __ mtlr(R3_RET);
2212 __ mr(R3_RET, Rexception);
2213 __ blr();
2214 }
2215 }
2216
2217 // JVMTI ForceEarlyReturn support.
2218 // Returns "in the middle" of a method with a "fake" return value.
2219 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
2220
2221 Register Rscratch1 = R11_scratch1,
2222 Rscratch2 = R12_scratch2;
2223
2224 address entry = __ pc();
2225 __ empty_expression_stack();
2226
2227 __ load_earlyret_value(state, Rscratch1);
2228
2229 __ ld(Rscratch1, in_bytes(JavaThread::jvmti_thread_state_offset()), R16_thread);
2230 // Clear the earlyret state.
2231 __ li(R0, 0);
2232 __ stw(R0, in_bytes(JvmtiThreadState::earlyret_state_offset()), Rscratch1);
2233
2234 __ remove_activation(state, false, false);
2235 // Copied from TemplateTable::_return.
2236 // Restoration of lr done by remove_activation.
2237 switch (state) {
2238 // Narrow result if state is itos but result type is smaller.
2239 case btos:
2240 case ztos:
2241 case ctos:
2242 case stos:
2243 case itos: __ narrow(R17_tos); /* fall through */
2244 case ltos:
2245 case atos: __ mr(R3_RET, R17_tos); break;
2246 case ftos:
2247 case dtos: __ fmr(F1_RET, F15_ftos); break;
2248 case vtos: // This might be a constructor. Final fields (and volatile fields on PPC64) need
2249 // to get visible before the reference to the object gets stored anywhere.
2250 __ membar(Assembler::StoreStore); break;
2251 default : ShouldNotReachHere();
2252 }
2253 __ blr();
2254
2255 return entry;
2256 } // end of ForceEarlyReturn support
2257
2258 //-----------------------------------------------------------------------------
2259 // Helper for vtos entry point generation
2260
2261 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t,
2262 address& bep,
2263 address& cep,
2264 address& sep,
2265 address& aep,
2266 address& iep,
2267 address& lep,
2268 address& fep,
2269 address& dep,
2270 address& vep) {
2271 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
2272 Label L;
2273
2274 aep = __ pc(); __ push_ptr(); __ b(L);
2275 fep = __ pc(); __ push_f(); __ b(L);
2276 dep = __ pc(); __ push_d(); __ b(L);
2277 lep = __ pc(); __ push_l(); __ b(L);
2278 __ align(32, 12, 24); // align L
2279 bep = cep = sep =
2280 iep = __ pc(); __ push_i();
2281 vep = __ pc();
2282 __ bind(L);
2283 generate_and_dispatch(t);
2284 }
2285
2286 //-----------------------------------------------------------------------------
2287
2288 // Non-product code
2289 #ifndef PRODUCT
2290 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
2291 //__ flush_bundle();
2292 address entry = __ pc();
2293
2294 const char *bname = NULL;
2295 uint tsize = 0;
2296 switch(state) {
2297 case ftos:
2298 bname = "trace_code_ftos {";
2299 tsize = 2;
2300 break;
2301 case btos:
2302 bname = "trace_code_btos {";
2303 tsize = 2;
2304 break;
2305 case ztos:
2306 bname = "trace_code_ztos {";
2307 tsize = 2;
2308 break;
2309 case ctos:
2310 bname = "trace_code_ctos {";
2311 tsize = 2;
2312 break;
2313 case stos:
2314 bname = "trace_code_stos {";
2315 tsize = 2;
2316 break;
2317 case itos:
2318 bname = "trace_code_itos {";
2319 tsize = 2;
2320 break;
2321 case ltos:
2322 bname = "trace_code_ltos {";
2323 tsize = 3;
2324 break;
2325 case atos:
2326 bname = "trace_code_atos {";
2327 tsize = 2;
2328 break;
2329 case vtos:
2330 // Note: In case of vtos, the topmost of stack value could be a int or doubl
2331 // In case of a double (2 slots) we won't see the 2nd stack value.
2332 // Maybe we simply should print the topmost 3 stack slots to cope with the problem.
2333 bname = "trace_code_vtos {";
2334 tsize = 2;
2335
2336 break;
2337 case dtos:
2338 bname = "trace_code_dtos {";
2339 tsize = 3;
2340 break;
2341 default:
2342 ShouldNotReachHere();
2343 }
2344 BLOCK_COMMENT(bname);
2345
2346 // Support short-cut for TraceBytecodesAt.
2347 // Don't call into the VM if we don't want to trace to speed up things.
2348 Label Lskip_vm_call;
2349 if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) {
2350 int offs1 = __ load_const_optimized(R11_scratch1, (address) &TraceBytecodesAt, R0, true);
2351 int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true);
2352 __ ld(R11_scratch1, offs1, R11_scratch1);
2353 __ lwa(R12_scratch2, offs2, R12_scratch2);
2354 __ cmpd(CCR0, R12_scratch2, R11_scratch1);
2355 __ blt(CCR0, Lskip_vm_call);
2356 }
2357
2358 __ push(state);
2359 // Load 2 topmost expression stack values.
2360 __ ld(R6_ARG4, tsize*Interpreter::stackElementSize, R15_esp);
2361 __ ld(R5_ARG3, Interpreter::stackElementSize, R15_esp);
2362 __ mflr(R31);
2363 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::trace_bytecode), /* unused */ R4_ARG2, R5_ARG3, R6_ARG4, false);
2364 __ mtlr(R31);
2365 __ pop(state);
2366
2367 if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) {
2368 __ bind(Lskip_vm_call);
2369 }
2370 __ blr();
2371 BLOCK_COMMENT("} trace_code");
2372 return entry;
2373 }
2374
2375 void TemplateInterpreterGenerator::count_bytecode() {
2376 int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeCounter::_counter_value, R12_scratch2, true);
2377 __ lwz(R12_scratch2, offs, R11_scratch1);
2378 __ addi(R12_scratch2, R12_scratch2, 1);
2379 __ stw(R12_scratch2, offs, R11_scratch1);
2380 }
2381
2382 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
2383 int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeHistogram::_counters[t->bytecode()], R12_scratch2, true);
2384 __ lwz(R12_scratch2, offs, R11_scratch1);
2385 __ addi(R12_scratch2, R12_scratch2, 1);
2386 __ stw(R12_scratch2, offs, R11_scratch1);
2387 }
2388
2389 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
2390 const Register addr = R11_scratch1,
2391 tmp = R12_scratch2;
2392 // Get index, shift out old bytecode, bring in new bytecode, and store it.
2393 // _index = (_index >> log2_number_of_codes) |
2394 // (bytecode << log2_number_of_codes);
2395 int offs1 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_index, tmp, true);
2396 __ lwz(tmp, offs1, addr);
2397 __ srwi(tmp, tmp, BytecodePairHistogram::log2_number_of_codes);
2398 __ ori(tmp, tmp, ((int) t->bytecode()) << BytecodePairHistogram::log2_number_of_codes);
2399 __ stw(tmp, offs1, addr);
2400
2401 // Bump bucket contents.
2402 // _counters[_index] ++;
2403 int offs2 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_counters, R0, true);
2404 __ sldi(tmp, tmp, LogBytesPerInt);
2405 __ add(addr, tmp, addr);
2406 __ lwz(tmp, offs2, addr);
2407 __ addi(tmp, tmp, 1);
2408 __ stw(tmp, offs2, addr);
2409 }
2410
2411 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
2412 // Call a little run-time stub to avoid blow-up for each bytecode.
2413 // The run-time runtime saves the right registers, depending on
2414 // the tosca in-state for the given template.
2415
2416 assert(Interpreter::trace_code(t->tos_in()) != NULL,
2417 "entry must have been generated");
2418
2419 // Note: we destroy LR here.
2420 __ bl(Interpreter::trace_code(t->tos_in()));
2421 }
2422
2423 void TemplateInterpreterGenerator::stop_interpreter_at() {
2424 Label L;
2425 int offs1 = __ load_const_optimized(R11_scratch1, (address) &StopInterpreterAt, R0, true);
2426 int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true);
2427 __ ld(R11_scratch1, offs1, R11_scratch1);
2428 __ lwa(R12_scratch2, offs2, R12_scratch2);
2429 __ cmpd(CCR0, R12_scratch2, R11_scratch1);
2430 __ bne(CCR0, L);
2431 __ illtrap();
2432 __ bind(L);
2433 }
2434
2435 #endif // !PRODUCT