Mercurial > hg > truffle

comparison src/cpu/ppc/vm/templateInterpreter_ppc.cpp @ 17804:fd1b9f02cc91

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
8036976: PPC64: implement the template interpreter Reviewed-by: kvn, coleenp Contributed-by: axel.siebenborn@sap.com, martin.doerr@sap.com
author goetz
date Mon, 10 Mar 2014 12:58:02 +0100
parents
children 6048424d3865
comparison
equal deleted inserted replaced
17803:31e80afe3fed 17804:fd1b9f02cc91
1 /*
2 * Copyright (c) 2014, Oracle and/or its affiliates. All rights reserved.
3 * Copyright 2013, 2014 SAP AG. 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 #ifndef CC_INTERP
28 #include "asm/macroAssembler.inline.hpp"
29 #include "interpreter/bytecodeHistogram.hpp"
30 #include "interpreter/interpreter.hpp"
31 #include "interpreter/interpreterGenerator.hpp"
32 #include "interpreter/interpreterRuntime.hpp"
33 #include "interpreter/templateTable.hpp"
34 #include "oops/arrayOop.hpp"
35 #include "oops/methodData.hpp"
36 #include "oops/method.hpp"
37 #include "oops/oop.inline.hpp"
38 #include "prims/jvmtiExport.hpp"
39 #include "prims/jvmtiThreadState.hpp"
40 #include "runtime/arguments.hpp"
41 #include "runtime/deoptimization.hpp"
42 #include "runtime/frame.inline.hpp"
43 #include "runtime/sharedRuntime.hpp"
44 #include "runtime/stubRoutines.hpp"
45 #include "runtime/synchronizer.hpp"
46 #include "runtime/timer.hpp"
47 #include "runtime/vframeArray.hpp"
48 #include "utilities/debug.hpp"
49 #include "utilities/macros.hpp"
50
51 #undef __
52 #define __ _masm->
53
54 #ifdef PRODUCT
55 #define BLOCK_COMMENT(str) /* nothing */
56 #else
57 #define BLOCK_COMMENT(str) __ block_comment(str)
58 #endif
59
60 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
61
62 //-----------------------------------------------------------------------------
63
64 // Actually we should never reach here since we do stack overflow checks before pushing any frame.
65 address TemplateInterpreterGenerator::generate_StackOverflowError_handler() {
66 address entry = __ pc();
67 __ unimplemented("generate_StackOverflowError_handler");
68 return entry;
69 }
70
71 address TemplateInterpreterGenerator::generate_ArrayIndexOutOfBounds_handler(const char* name) {
72 address entry = __ pc();
73 __ empty_expression_stack();
74 __ load_const_optimized(R4_ARG2, (address) name);
75 // Index is in R17_tos.
76 __ mr(R5_ARG3, R17_tos);
77 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ArrayIndexOutOfBoundsException));
78 return entry;
79 }
80
81 #if 0
82 // Call special ClassCastException constructor taking object to cast
83 // and target class as arguments.
84 address TemplateInterpreterGenerator::generate_ClassCastException_verbose_handler(const char* name) {
85 address entry = __ pc();
86
87 // Target class oop is in register R6_ARG4 by convention!
88
89 // Expression stack must be empty before entering the VM if an
90 // exception happened.
91 __ empty_expression_stack();
92 // Setup parameters.
93 // Thread will be loaded to R3_ARG1.
94 __ load_const_optimized(R4_ARG2, (address) name);
95 __ mr(R5_ARG3, R17_tos);
96 // R6_ARG4 contains specified class.
97 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ClassCastException_verbose));
98 #ifdef ASSERT
99 // Above call must not return here since exception pending.
100 __ should_not_reach_here();
101 #endif
102 return entry;
103 }
104 #endif
105
106 address TemplateInterpreterGenerator::generate_ClassCastException_handler() {
107 address entry = __ pc();
108 // Expression stack must be empty before entering the VM if an
109 // exception happened.
110 __ empty_expression_stack();
111
112 // Load exception object.
113 // Thread will be loaded to R3_ARG1.
114 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_ClassCastException), R17_tos);
115 #ifdef ASSERT
116 // Above call must not return here since exception pending.
117 __ should_not_reach_here();
118 #endif
119 return entry;
120 }
121
122 address TemplateInterpreterGenerator::generate_exception_handler_common(const char* name, const char* message, bool pass_oop) {
123 address entry = __ pc();
124 //__ untested("generate_exception_handler_common");
125 Register Rexception = R17_tos;
126
127 // Expression stack must be empty before entering the VM if an exception happened.
128 __ empty_expression_stack();
129
130 __ load_const_optimized(R4_ARG2, (address) name, R11_scratch1);
131 if (pass_oop) {
132 __ mr(R5_ARG3, Rexception);
133 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_klass_exception), false);
134 } else {
135 __ load_const_optimized(R5_ARG3, (address) message, R11_scratch1);
136 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::create_exception), false);
137 }
138
139 // Throw exception.
140 __ mr(R3_ARG1, Rexception);
141 __ load_const_optimized(R11_scratch1, Interpreter::throw_exception_entry(), R12_scratch2);
142 __ mtctr(R11_scratch1);
143 __ bctr();
144
145 return entry;
146 }
147
148 address TemplateInterpreterGenerator::generate_continuation_for(TosState state) {
149 address entry = __ pc();
150 __ unimplemented("generate_continuation_for");
151 return entry;
152 }
153
154 // This entry is returned to when a call returns to the interpreter.
155 // When we arrive here, we expect that the callee stack frame is already popped.
156 address TemplateInterpreterGenerator::generate_return_entry_for(TosState state, int step, size_t index_size) {
157 address entry = __ pc();
158
159 // Move the value out of the return register back to the TOS cache of current frame.
160 switch (state) {
161 case ltos:
162 case btos:
163 case ctos:
164 case stos:
165 case atos:
166 case itos: __ mr(R17_tos, R3_RET); break; // RET -> TOS cache
167 case ftos:
168 case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET
169 case vtos: break; // Nothing to do, this was a void return.
170 default : ShouldNotReachHere();
171 }
172
173 __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp.
174 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
175 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
176
177 // Compiled code destroys templateTableBase, reload.
178 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R12_scratch2);
179
180 const Register cache = R11_scratch1;
181 const Register size = R12_scratch2;
182 __ get_cache_and_index_at_bcp(cache, 1, index_size);
183
184 // Big Endian (get least significant byte of 64 bit value):
185 __ lbz(size, in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::flags_offset()) + 7, cache);
186 __ sldi(size, size, Interpreter::logStackElementSize);
187 __ add(R15_esp, R15_esp, size);
188 __ dispatch_next(state, step);
189 return entry;
190 }
191
192 address TemplateInterpreterGenerator::generate_deopt_entry_for(TosState state, int step) {
193 address entry = __ pc();
194 // If state != vtos, we're returning from a native method, which put it's result
195 // into the result register. So move the value out of the return register back
196 // to the TOS cache of current frame.
197
198 switch (state) {
199 case ltos:
200 case btos:
201 case ctos:
202 case stos:
203 case atos:
204 case itos: __ mr(R17_tos, R3_RET); break; // GR_RET -> TOS cache
205 case ftos:
206 case dtos: __ fmr(F15_ftos, F1_RET); break; // TOS cache -> GR_FRET
207 case vtos: break; // Nothing to do, this was a void return.
208 default : ShouldNotReachHere();
209 }
210
211 // Load LcpoolCache @@@ should be already set!
212 __ get_constant_pool_cache(R27_constPoolCache);
213
214 // Handle a pending exception, fall through if none.
215 __ check_and_forward_exception(R11_scratch1, R12_scratch2);
216
217 // Start executing bytecodes.
218 __ dispatch_next(state, step);
219
220 return entry;
221 }
222
223 // A result handler converts the native result into java format.
224 // Use the shared code between c++ and template interpreter.
225 address TemplateInterpreterGenerator::generate_result_handler_for(BasicType type) {
226 return AbstractInterpreterGenerator::generate_result_handler_for(type);
227 }
228
229 address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state, address runtime_entry) {
230 address entry = __ pc();
231
232 __ push(state);
233 __ call_VM(noreg, runtime_entry);
234 __ dispatch_via(vtos, Interpreter::_normal_table.table_for(vtos));
235
236 return entry;
237 }
238
239 // Helpers for commoning out cases in the various type of method entries.
240
241 // Increment invocation count & check for overflow.
242 //
243 // Note: checking for negative value instead of overflow
244 // so we have a 'sticky' overflow test.
245 //
246 void TemplateInterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
247 // Note: In tiered we increment either counters in method or in MDO depending if we're profiling or not.
248 Register Rscratch1 = R11_scratch1;
249 Register Rscratch2 = R12_scratch2;
250 Register R3_counters = R3_ARG1;
251 Label done;
252
253 if (TieredCompilation) {
254 const int increment = InvocationCounter::count_increment;
255 const int mask = ((1 << Tier0InvokeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
256 Label no_mdo;
257 if (ProfileInterpreter) {
258 const Register Rmdo = Rscratch1;
259 // If no method data exists, go to profile_continue.
260 __ ld(Rmdo, in_bytes(Method::method_data_offset()), R19_method);
261 __ cmpdi(CCR0, Rmdo, 0);
262 __ beq(CCR0, no_mdo);
263
264 // Increment backedge counter in the MDO.
265 const int mdo_bc_offs = in_bytes(MethodData::backedge_counter_offset()) + in_bytes(InvocationCounter::counter_offset());
266 __ lwz(Rscratch2, mdo_bc_offs, Rmdo);
267 __ addi(Rscratch2, Rscratch2, increment);
268 __ stw(Rscratch2, mdo_bc_offs, Rmdo);
269 __ load_const_optimized(Rscratch1, mask, R0);
270 __ and_(Rscratch1, Rscratch2, Rscratch1);
271 __ bne(CCR0, done);
272 __ b(*overflow);
273 }
274
275 // Increment counter in MethodCounters*.
276 const int mo_bc_offs = in_bytes(MethodCounters::backedge_counter_offset()) + in_bytes(InvocationCounter::counter_offset());
277 __ bind(no_mdo);
278 __ get_method_counters(R19_method, R3_counters, done);
279 __ lwz(Rscratch2, mo_bc_offs, R3_counters);
280 __ addi(Rscratch2, Rscratch2, increment);
281 __ stw(Rscratch2, mo_bc_offs, R3_counters);
282 __ load_const_optimized(Rscratch1, mask, R0);
283 __ and_(Rscratch1, Rscratch2, Rscratch1);
284 __ beq(CCR0, *overflow);
285
286 __ bind(done);
287
288 } else {
289
290 // Update standard invocation counters.
291 Register Rsum_ivc_bec = R4_ARG2;
292 __ get_method_counters(R19_method, R3_counters, done);
293 __ increment_invocation_counter(R3_counters, Rsum_ivc_bec, R12_scratch2);
294 // Increment interpreter invocation counter.
295 if (ProfileInterpreter) { // %%% Merge this into methodDataOop.
296 __ lwz(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters);
297 __ addi(R12_scratch2, R12_scratch2, 1);
298 __ stw(R12_scratch2, in_bytes(MethodCounters::interpreter_invocation_counter_offset()), R3_counters);
299 }
300 // Check if we must create a method data obj.
301 if (ProfileInterpreter && profile_method != NULL) {
302 const Register profile_limit = Rscratch1;
303 int pl_offs = __ load_const_optimized(profile_limit, &InvocationCounter::InterpreterProfileLimit, R0, true);
304 __ lwz(profile_limit, pl_offs, profile_limit);
305 // Test to see if we should create a method data oop.
306 __ cmpw(CCR0, Rsum_ivc_bec, profile_limit);
307 __ blt(CCR0, *profile_method_continue);
308 // If no method data exists, go to profile_method.
309 __ test_method_data_pointer(*profile_method);
310 }
311 // Finally check for counter overflow.
312 if (overflow) {
313 const Register invocation_limit = Rscratch1;
314 int il_offs = __ load_const_optimized(invocation_limit, &InvocationCounter::InterpreterInvocationLimit, R0, true);
315 __ lwz(invocation_limit, il_offs, invocation_limit);
316 assert(4 == sizeof(InvocationCounter::InterpreterInvocationLimit), "unexpected field size");
317 __ cmpw(CCR0, Rsum_ivc_bec, invocation_limit);
318 __ bge(CCR0, *overflow);
319 }
320
321 __ bind(done);
322 }
323 }
324
325 // Generate code to initiate compilation on invocation counter overflow.
326 void TemplateInterpreterGenerator::generate_counter_overflow(Label& continue_entry) {
327 // Generate code to initiate compilation on the counter overflow.
328
329 // InterpreterRuntime::frequency_counter_overflow takes one arguments,
330 // which indicates if the counter overflow occurs at a backwards branch (NULL bcp)
331 // We pass zero in.
332 // The call returns the address of the verified entry point for the method or NULL
333 // if the compilation did not complete (either went background or bailed out).
334 //
335 // Unlike the C++ interpreter above: Check exceptions!
336 // Assumption: Caller must set the flag "do_not_unlock_if_sychronized" if the monitor of a sync'ed
337 // method has not yet been created. Thus, no unlocking of a non-existing monitor can occur.
338
339 __ li(R4_ARG2, 0);
340 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), R4_ARG2, true);
341
342 // Returns verified_entry_point or NULL.
343 // We ignore it in any case.
344 __ b(continue_entry);
345 }
346
347 void TemplateInterpreterGenerator::generate_stack_overflow_check(Register Rmem_frame_size, Register Rscratch1) {
348 assert_different_registers(Rmem_frame_size, Rscratch1);
349 __ generate_stack_overflow_check_with_compare_and_throw(Rmem_frame_size, Rscratch1);
350 }
351
352 void TemplateInterpreterGenerator::unlock_method(bool check_exceptions) {
353 __ unlock_object(R26_monitor, check_exceptions);
354 }
355
356 // Lock the current method, interpreter register window must be set up!
357 void TemplateInterpreterGenerator::lock_method(Register Rflags, Register Rscratch1, Register Rscratch2, bool flags_preloaded) {
358 const Register Robj_to_lock = Rscratch2;
359
360 {
361 if (!flags_preloaded) {
362 __ lwz(Rflags, method_(access_flags));
363 }
364
365 #ifdef ASSERT
366 // Check if methods needs synchronization.
367 {
368 Label Lok;
369 __ testbitdi(CCR0, R0, Rflags, JVM_ACC_SYNCHRONIZED_BIT);
370 __ btrue(CCR0,Lok);
371 __ stop("method doesn't need synchronization");
372 __ bind(Lok);
373 }
374 #endif // ASSERT
375 }
376
377 // Get synchronization object to Rscratch2.
378 {
379 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
380 Label Lstatic;
381 Label Ldone;
382
383 __ testbitdi(CCR0, R0, Rflags, JVM_ACC_STATIC_BIT);
384 __ btrue(CCR0, Lstatic);
385
386 // Non-static case: load receiver obj from stack and we're done.
387 __ ld(Robj_to_lock, R18_locals);
388 __ b(Ldone);
389
390 __ bind(Lstatic); // Static case: Lock the java mirror
391 __ ld(Robj_to_lock, in_bytes(Method::const_offset()), R19_method);
392 __ ld(Robj_to_lock, in_bytes(ConstMethod::constants_offset()), Robj_to_lock);
393 __ ld(Robj_to_lock, ConstantPool::pool_holder_offset_in_bytes(), Robj_to_lock);
394 __ ld(Robj_to_lock, mirror_offset, Robj_to_lock);
395
396 __ bind(Ldone);
397 __ verify_oop(Robj_to_lock);
398 }
399
400 // Got the oop to lock => execute!
401 __ add_monitor_to_stack(true, Rscratch1, R0);
402
403 __ std(Robj_to_lock, BasicObjectLock::obj_offset_in_bytes(), R26_monitor);
404 __ lock_object(R26_monitor, Robj_to_lock);
405 }
406
407 // Generate a fixed interpreter frame for pure interpreter
408 // and I2N native transition frames.
409 //
410 // Before (stack grows downwards):
411 //
412 // | ... |
413 // |------------- |
414 // | java arg0 |
415 // | ... |
416 // | java argn |
417 // | | <- R15_esp
418 // | |
419 // |--------------|
420 // | abi_112 |
421 // | | <- R1_SP
422 // |==============|
423 //
424 //
425 // After:
426 //
427 // | ... |
428 // | java arg0 |<- R18_locals
429 // | ... |
430 // | java argn |
431 // |--------------|
432 // | |
433 // | java locals |
434 // | |
435 // |--------------|
436 // | abi_48 |
437 // |==============|
438 // | |
439 // | istate |
440 // | |
441 // |--------------|
442 // | monitor |<- R26_monitor
443 // |--------------|
444 // | |<- R15_esp
445 // | expression |
446 // | stack |
447 // | |
448 // |--------------|
449 // | |
450 // | abi_112 |<- R1_SP
451 // |==============|
452 //
453 // The top most frame needs an abi space of 112 bytes. This space is needed,
454 // since we call to c. The c function may spill their arguments to the caller
455 // frame. When we call to java, we don't need these spill slots. In order to save
456 // space on the stack, we resize the caller. However, java local reside in
457 // the caller frame and the frame has to be increased. The frame_size for the
458 // current frame was calculated based on max_stack as size for the expression
459 // stack. At the call, just a part of the expression stack might be used.
460 // We don't want to waste this space and cut the frame back accordingly.
461 // The resulting amount for resizing is calculated as follows:
462 // resize = (number_of_locals - number_of_arguments) * slot_size
463 // + (R1_SP - R15_esp) + 48
464 //
465 // The size for the callee frame is calculated:
466 // framesize = 112 + max_stack + monitor + state_size
467 //
468 // maxstack: Max number of slots on the expression stack, loaded from the method.
469 // monitor: We statically reserve room for one monitor object.
470 // state_size: We save the current state of the interpreter to this area.
471 //
472 void TemplateInterpreterGenerator::generate_fixed_frame(bool native_call, Register Rsize_of_parameters, Register Rsize_of_locals) {
473 Register parent_frame_resize = R6_ARG4, // Frame will grow by this number of bytes.
474 top_frame_size = R7_ARG5,
475 Rconst_method = R8_ARG6;
476
477 assert_different_registers(Rsize_of_parameters, Rsize_of_locals, parent_frame_resize, top_frame_size);
478
479 __ ld(Rconst_method, method_(const));
480 __ lhz(Rsize_of_parameters /* number of params */,
481 in_bytes(ConstMethod::size_of_parameters_offset()), Rconst_method);
482 if (native_call) {
483 // If we're calling a native method, we reserve space for the worst-case signature
484 // handler varargs vector, which is max(Argument::n_register_parameters, parameter_count+2).
485 // We add two slots to the parameter_count, one for the jni
486 // environment and one for a possible native mirror.
487 Label skip_native_calculate_max_stack;
488 __ addi(top_frame_size, Rsize_of_parameters, 2);
489 __ cmpwi(CCR0, top_frame_size, Argument::n_register_parameters);
490 __ bge(CCR0, skip_native_calculate_max_stack);
491 __ li(top_frame_size, Argument::n_register_parameters);
492 __ bind(skip_native_calculate_max_stack);
493 __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize);
494 __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize);
495 __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize!
496 assert(Rsize_of_locals == noreg, "Rsize_of_locals not initialized"); // Only relevant value is Rsize_of_parameters.
497 } else {
498 __ lhz(Rsize_of_locals /* number of params */, in_bytes(ConstMethod::size_of_locals_offset()), Rconst_method);
499 __ sldi(Rsize_of_parameters, Rsize_of_parameters, Interpreter::logStackElementSize);
500 __ sldi(Rsize_of_locals, Rsize_of_locals, Interpreter::logStackElementSize);
501 __ lhz(top_frame_size, in_bytes(ConstMethod::max_stack_offset()), Rconst_method);
502 __ sub(R11_scratch1, Rsize_of_locals, Rsize_of_parameters); // >=0
503 __ sub(parent_frame_resize, R1_SP, R15_esp); // <0, off by Interpreter::stackElementSize!
504 __ sldi(top_frame_size, top_frame_size, Interpreter::logStackElementSize);
505 __ add(parent_frame_resize, parent_frame_resize, R11_scratch1);
506 }
507
508 // Compute top frame size.
509 __ addi(top_frame_size, top_frame_size, frame::abi_reg_args_size + frame::ijava_state_size);
510
511 // Cut back area between esp and max_stack.
512 __ addi(parent_frame_resize, parent_frame_resize, frame::abi_minframe_size - Interpreter::stackElementSize);
513
514 __ round_to(top_frame_size, frame::alignment_in_bytes);
515 __ round_to(parent_frame_resize, frame::alignment_in_bytes);
516 // parent_frame_resize = (locals-parameters) - (ESP-SP-ABI48) Rounded to frame alignment size.
517 // Enlarge by locals-parameters (not in case of native_call), shrink by ESP-SP-ABI48.
518
519 {
520 // --------------------------------------------------------------------------
521 // Stack overflow check
522
523 Label cont;
524 __ add(R11_scratch1, parent_frame_resize, top_frame_size);
525 generate_stack_overflow_check(R11_scratch1, R12_scratch2);
526 }
527
528 // Set up interpreter state registers.
529
530 __ add(R18_locals, R15_esp, Rsize_of_parameters);
531 __ ld(R27_constPoolCache, in_bytes(ConstMethod::constants_offset()), Rconst_method);
532 __ ld(R27_constPoolCache, ConstantPool::cache_offset_in_bytes(), R27_constPoolCache);
533
534 // Set method data pointer.
535 if (ProfileInterpreter) {
536 Label zero_continue;
537 __ ld(R28_mdx, method_(method_data));
538 __ cmpdi(CCR0, R28_mdx, 0);
539 __ beq(CCR0, zero_continue);
540 __ addi(R28_mdx, R28_mdx, in_bytes(MethodData::data_offset()));
541 __ bind(zero_continue);
542 }
543
544 if (native_call) {
545 __ li(R14_bcp, 0); // Must initialize.
546 } else {
547 __ add(R14_bcp, in_bytes(ConstMethod::codes_offset()), Rconst_method);
548 }
549
550 // Resize parent frame.
551 __ mflr(R12_scratch2);
552 __ neg(parent_frame_resize, parent_frame_resize);
553 __ resize_frame(parent_frame_resize, R11_scratch1);
554 __ std(R12_scratch2, _abi(lr), R1_SP);
555
556 __ addi(R26_monitor, R1_SP, - frame::ijava_state_size);
557 __ addi(R15_esp, R26_monitor, - Interpreter::stackElementSize);
558
559 // Store values.
560 // R15_esp, R14_bcp, R26_monitor, R28_mdx are saved at java calls
561 // in InterpreterMacroAssembler::call_from_interpreter.
562 __ std(R19_method, _ijava_state_neg(method), R1_SP);
563 __ std(R21_sender_SP, _ijava_state_neg(sender_sp), R1_SP);
564 __ std(R27_constPoolCache, _ijava_state_neg(cpoolCache), R1_SP);
565 __ std(R18_locals, _ijava_state_neg(locals), R1_SP);
566
567 // Note: esp, bcp, monitor, mdx live in registers. Hence, the correct version can only
568 // be found in the frame after save_interpreter_state is done. This is always true
569 // for non-top frames. But when a signal occurs, dumping the top frame can go wrong,
570 // because e.g. frame::interpreter_frame_bcp() will not access the correct value
571 // (Enhanced Stack Trace).
572 // The signal handler does not save the interpreter state into the frame.
573 __ li(R0, 0);
574 #ifdef ASSERT
575 // Fill remaining slots with constants.
576 __ load_const_optimized(R11_scratch1, 0x5afe);
577 __ load_const_optimized(R12_scratch2, 0xdead);
578 #endif
579 // We have to initialize some frame slots for native calls (accessed by GC).
580 if (native_call) {
581 __ std(R26_monitor, _ijava_state_neg(monitors), R1_SP);
582 __ std(R14_bcp, _ijava_state_neg(bcp), R1_SP);
583 if (ProfileInterpreter) { __ std(R28_mdx, _ijava_state_neg(mdx), R1_SP); }
584 }
585 #ifdef ASSERT
586 else {
587 __ std(R12_scratch2, _ijava_state_neg(monitors), R1_SP);
588 __ std(R12_scratch2, _ijava_state_neg(bcp), R1_SP);
589 __ std(R12_scratch2, _ijava_state_neg(mdx), R1_SP);
590 }
591 __ std(R11_scratch1, _ijava_state_neg(ijava_reserved), R1_SP);
592 __ std(R12_scratch2, _ijava_state_neg(esp), R1_SP);
593 __ std(R12_scratch2, _ijava_state_neg(lresult), R1_SP);
594 __ std(R12_scratch2, _ijava_state_neg(fresult), R1_SP);
595 #endif
596 __ subf(R12_scratch2, top_frame_size, R1_SP);
597 __ std(R0, _ijava_state_neg(oop_tmp), R1_SP);
598 __ std(R12_scratch2, _ijava_state_neg(top_frame_sp), R1_SP);
599
600 // Push top frame.
601 __ push_frame(top_frame_size, R11_scratch1);
602 }
603
604 // End of helpers
605
606 // ============================================================================
607 // Various method entries
608 //
609
610 // Empty method, generate a very fast return. We must skip this entry if
611 // someone's debugging, indicated by the flag
612 // "interp_mode" in the Thread obj.
613 // Note: empty methods are generated mostly methods that do assertions, which are
614 // disabled in the "java opt build".
615 address TemplateInterpreterGenerator::generate_empty_entry(void) {
616 if (!UseFastEmptyMethods) {
617 NOT_PRODUCT(__ should_not_reach_here();)
618 return Interpreter::entry_for_kind(Interpreter::zerolocals);
619 }
620
621 Label Lslow_path;
622 const Register Rjvmti_mode = R11_scratch1;
623 address entry = __ pc();
624
625 __ lwz(Rjvmti_mode, thread_(interp_only_mode));
626 __ cmpwi(CCR0, Rjvmti_mode, 0);
627 __ bne(CCR0, Lslow_path); // jvmti_mode!=0
628
629 // Noone's debuggin: Simply return.
630 // Pop c2i arguments (if any) off when we return.
631 #ifdef ASSERT
632 __ ld(R9_ARG7, 0, R1_SP);
633 __ ld(R10_ARG8, 0, R21_sender_SP);
634 __ cmpd(CCR0, R9_ARG7, R10_ARG8);
635 __ asm_assert_eq("backlink", 0x545);
636 #endif // ASSERT
637 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
638
639 // And we're done.
640 __ blr();
641
642 __ bind(Lslow_path);
643 __ branch_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals), R11_scratch1);
644 __ flush();
645
646 return entry;
647 }
648
649 // Support abs and sqrt like in compiler.
650 // For others we can use a normal (native) entry.
651
652 inline bool math_entry_available(AbstractInterpreter::MethodKind kind) {
653 // Provide math entry with debugging on demand.
654 // Note: Debugging changes which code will get executed:
655 // Debugging or disabled InlineIntrinsics: java method will get interpreted and performs a native call.
656 // Not debugging and enabled InlineIntrinics: processor instruction will get used.
657 // Result might differ slightly due to rounding etc.
658 if (!InlineIntrinsics && (!FLAG_IS_ERGO(InlineIntrinsics))) return false; // Generate a vanilla entry.
659
660 return ((kind==Interpreter::java_lang_math_sqrt && VM_Version::has_fsqrt()) ||
661 (kind==Interpreter::java_lang_math_abs));
662 }
663
664 address TemplateInterpreterGenerator::generate_math_entry(AbstractInterpreter::MethodKind kind) {
665 if (!math_entry_available(kind)) {
666 NOT_PRODUCT(__ should_not_reach_here();)
667 return Interpreter::entry_for_kind(Interpreter::zerolocals);
668 }
669
670 Label Lslow_path;
671 const Register Rjvmti_mode = R11_scratch1;
672 address entry = __ pc();
673
674 // Provide math entry with debugging on demand.
675 __ lwz(Rjvmti_mode, thread_(interp_only_mode));
676 __ cmpwi(CCR0, Rjvmti_mode, 0);
677 __ bne(CCR0, Lslow_path); // jvmti_mode!=0
678
679 __ lfd(F1_RET, Interpreter::stackElementSize, R15_esp);
680
681 // Pop c2i arguments (if any) off when we return.
682 #ifdef ASSERT
683 __ ld(R9_ARG7, 0, R1_SP);
684 __ ld(R10_ARG8, 0, R21_sender_SP);
685 __ cmpd(CCR0, R9_ARG7, R10_ARG8);
686 __ asm_assert_eq("backlink", 0x545);
687 #endif // ASSERT
688 __ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
689
690 if (kind == Interpreter::java_lang_math_sqrt) {
691 __ fsqrt(F1_RET, F1_RET);
692 } else if (kind == Interpreter::java_lang_math_abs) {
693 __ fabs(F1_RET, F1_RET);
694 } else {
695 ShouldNotReachHere();
696 }
697
698 // And we're done.
699 __ blr();
700
701 // Provide slow path for JVMTI case.
702 __ bind(Lslow_path);
703 __ branch_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals), R12_scratch2);
704 __ flush();
705
706 return entry;
707 }
708
709 // Interpreter stub for calling a native method. (asm interpreter)
710 // This sets up a somewhat different looking stack for calling the
711 // native method than the typical interpreter frame setup.
712 //
713 // On entry:
714 // R19_method - method
715 // R16_thread - JavaThread*
716 // R15_esp - intptr_t* sender tos
717 //
718 // abstract stack (grows up)
719 // [ IJava (caller of JNI callee) ] <-- ASP
720 // ...
721 address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
722
723 address entry = __ pc();
724
725 const bool inc_counter = UseCompiler || CountCompiledCalls;
726
727 // -----------------------------------------------------------------------------
728 // Allocate a new frame that represents the native callee (i2n frame).
729 // This is not a full-blown interpreter frame, but in particular, the
730 // following registers are valid after this:
731 // - R19_method
732 // - R18_local (points to start of argumuments to native function)
733 //
734 // abstract stack (grows up)
735 // [ IJava (caller of JNI callee) ] <-- ASP
736 // ...
737
738 const Register signature_handler_fd = R11_scratch1;
739 const Register pending_exception = R0;
740 const Register result_handler_addr = R31;
741 const Register native_method_fd = R11_scratch1;
742 const Register access_flags = R22_tmp2;
743 const Register active_handles = R11_scratch1; // R26_monitor saved to state.
744 const Register sync_state = R12_scratch2;
745 const Register sync_state_addr = sync_state; // Address is dead after use.
746 const Register suspend_flags = R11_scratch1;
747
748 //=============================================================================
749 // Allocate new frame and initialize interpreter state.
750
751 Label exception_return;
752 Label exception_return_sync_check;
753 Label stack_overflow_return;
754
755 // Generate new interpreter state and jump to stack_overflow_return in case of
756 // a stack overflow.
757 //generate_compute_interpreter_state(stack_overflow_return);
758
759 Register size_of_parameters = R22_tmp2;
760
761 generate_fixed_frame(true, size_of_parameters, noreg /* unused */);
762
763 //=============================================================================
764 // Increment invocation counter. On overflow, entry to JNI method
765 // will be compiled.
766 Label invocation_counter_overflow, continue_after_compile;
767 if (inc_counter) {
768 if (synchronized) {
769 // Since at this point in the method invocation the exception handler
770 // would try to exit the monitor of synchronized methods which hasn't
771 // been entered yet, we set the thread local variable
772 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
773 // runtime, exception handling i.e. unlock_if_synchronized_method will
774 // check this thread local flag.
775 // This flag has two effects, one is to force an unwind in the topmost
776 // interpreter frame and not perform an unlock while doing so.
777 __ li(R0, 1);
778 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
779 }
780 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
781
782 __ BIND(continue_after_compile);
783 // Reset the _do_not_unlock_if_synchronized flag.
784 if (synchronized) {
785 __ li(R0, 0);
786 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
787 }
788 }
789
790 // access_flags = method->access_flags();
791 // Load access flags.
792 assert(access_flags->is_nonvolatile(),
793 "access_flags must be in a non-volatile register");
794 // Type check.
795 assert(4 == sizeof(AccessFlags), "unexpected field size");
796 __ lwz(access_flags, method_(access_flags));
797
798 // We don't want to reload R19_method and access_flags after calls
799 // to some helper functions.
800 assert(R19_method->is_nonvolatile(),
801 "R19_method must be a non-volatile register");
802
803 // Check for synchronized methods. Must happen AFTER invocation counter
804 // check, so method is not locked if counter overflows.
805
806 if (synchronized) {
807 lock_method(access_flags, R11_scratch1, R12_scratch2, true);
808
809 // Update monitor in state.
810 __ ld(R11_scratch1, 0, R1_SP);
811 __ std(R26_monitor, _ijava_state_neg(monitors), R11_scratch1);
812 }
813
814 // jvmti/jvmpi support
815 __ notify_method_entry();
816
817 //=============================================================================
818 // Get and call the signature handler.
819
820 __ ld(signature_handler_fd, method_(signature_handler));
821 Label call_signature_handler;
822
823 __ cmpdi(CCR0, signature_handler_fd, 0);
824 __ bne(CCR0, call_signature_handler);
825
826 // Method has never been called. Either generate a specialized
827 // handler or point to the slow one.
828 //
829 // Pass parameter 'false' to avoid exception check in call_VM.
830 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), R19_method, false);
831
832 // Check for an exception while looking up the target method. If we
833 // incurred one, bail.
834 __ ld(pending_exception, thread_(pending_exception));
835 __ cmpdi(CCR0, pending_exception, 0);
836 __ bne(CCR0, exception_return_sync_check); // Has pending exception.
837
838 // Reload signature handler, it may have been created/assigned in the meanwhile.
839 __ ld(signature_handler_fd, method_(signature_handler));
840 __ twi_0(signature_handler_fd); // Order wrt. load of klass mirror and entry point (isync is below).
841
842 __ BIND(call_signature_handler);
843
844 // Before we call the signature handler we push a new frame to
845 // protect the interpreter frame volatile registers when we return
846 // from jni but before we can get back to Java.
847
848 // First set the frame anchor while the SP/FP registers are
849 // convenient and the slow signature handler can use this same frame
850 // anchor.
851
852 // We have a TOP_IJAVA_FRAME here, which belongs to us.
853 __ set_top_ijava_frame_at_SP_as_last_Java_frame(R1_SP, R12_scratch2/*tmp*/);
854
855 // Now the interpreter frame (and its call chain) have been
856 // invalidated and flushed. We are now protected against eager
857 // being enabled in native code. Even if it goes eager the
858 // registers will be reloaded as clean and we will invalidate after
859 // the call so no spurious flush should be possible.
860
861 // Call signature handler and pass locals address.
862 //
863 // Our signature handlers copy required arguments to the C stack
864 // (outgoing C args), R3_ARG1 to R10_ARG8, and FARG1 to FARG13.
865 __ mr(R3_ARG1, R18_locals);
866 __ ld(signature_handler_fd, 0, signature_handler_fd);
867
868 __ call_stub(signature_handler_fd);
869
870 // Remove the register parameter varargs slots we allocated in
871 // compute_interpreter_state. SP+16 ends up pointing to the ABI
872 // outgoing argument area.
873 //
874 // Not needed on PPC64.
875 //__ add(SP, SP, Argument::n_register_parameters*BytesPerWord);
876
877 assert(result_handler_addr->is_nonvolatile(), "result_handler_addr must be in a non-volatile register");
878 // Save across call to native method.
879 __ mr(result_handler_addr, R3_RET);
880
881 __ isync(); // Acquire signature handler before trying to fetch the native entry point and klass mirror.
882
883 // Set up fixed parameters and call the native method.
884 // If the method is static, get mirror into R4_ARG2.
885 {
886 Label method_is_not_static;
887 // Access_flags is non-volatile and still, no need to restore it.
888
889 // Restore access flags.
890 __ testbitdi(CCR0, R0, access_flags, JVM_ACC_STATIC_BIT);
891 __ bfalse(CCR0, method_is_not_static);
892
893 // constants = method->constants();
894 __ ld(R11_scratch1, in_bytes(Method::const_offset()), R19_method);
895 __ ld(R11_scratch1, in_bytes(ConstMethod::constants_offset()), R11_scratch1);
896 // pool_holder = method->constants()->pool_holder();
897 __ ld(R11_scratch1/*pool_holder*/, ConstantPool::pool_holder_offset_in_bytes(),
898 R11_scratch1/*constants*/);
899
900 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
901
902 // mirror = pool_holder->klass_part()->java_mirror();
903 __ ld(R0/*mirror*/, mirror_offset, R11_scratch1/*pool_holder*/);
904 // state->_native_mirror = mirror;
905
906 __ ld(R11_scratch1, 0, R1_SP);
907 __ std(R0/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1);
908 // R4_ARG2 = &state->_oop_temp;
909 __ addi(R4_ARG2, R11_scratch1, _ijava_state_neg(oop_tmp));
910 __ BIND(method_is_not_static);
911 }
912
913 // At this point, arguments have been copied off the stack into
914 // their JNI positions. Oops are boxed in-place on the stack, with
915 // handles copied to arguments. The result handler address is in a
916 // register.
917
918 // Pass JNIEnv address as first parameter.
919 __ addir(R3_ARG1, thread_(jni_environment));
920
921 // Load the native_method entry before we change the thread state.
922 __ ld(native_method_fd, method_(native_function));
923
924 //=============================================================================
925 // Transition from _thread_in_Java to _thread_in_native. As soon as
926 // we make this change the safepoint code needs to be certain that
927 // the last Java frame we established is good. The pc in that frame
928 // just needs to be near here not an actual return address.
929
930 // We use release_store_fence to update values like the thread state, where
931 // we don't want the current thread to continue until all our prior memory
932 // accesses (including the new thread state) are visible to other threads.
933 __ li(R0, _thread_in_native);
934 __ release();
935
936 // TODO PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
937 __ stw(R0, thread_(thread_state));
938
939 if (UseMembar) {
940 __ fence();
941 }
942
943 //=============================================================================
944 // Call the native method. Argument registers must not have been
945 // overwritten since "__ call_stub(signature_handler);" (except for
946 // ARG1 and ARG2 for static methods).
947 __ call_c(native_method_fd);
948
949 __ li(R0, 0);
950 __ ld(R11_scratch1, 0, R1_SP);
951 __ std(R3_RET, _ijava_state_neg(lresult), R11_scratch1);
952 __ stfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1);
953 __ std(R0/*mirror*/, _ijava_state_neg(oop_tmp), R11_scratch1); // reset
954
955 // Note: C++ interpreter needs the following here:
956 // The frame_manager_lr field, which we use for setting the last
957 // java frame, gets overwritten by the signature handler. Restore
958 // it now.
959 //__ get_PC_trash_LR(R11_scratch1);
960 //__ std(R11_scratch1, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
961
962 // Because of GC R19_method may no longer be valid.
963
964 // Block, if necessary, before resuming in _thread_in_Java state.
965 // In order for GC to work, don't clear the last_Java_sp until after
966 // blocking.
967
968 //=============================================================================
969 // Switch thread to "native transition" state before reading the
970 // synchronization state. This additional state is necessary
971 // because reading and testing the synchronization state is not
972 // atomic w.r.t. GC, as this scenario demonstrates: Java thread A,
973 // in _thread_in_native state, loads _not_synchronized and is
974 // preempted. VM thread changes sync state to synchronizing and
975 // suspends threads for GC. Thread A is resumed to finish this
976 // native method, but doesn't block here since it didn't see any
977 // synchronization in progress, and escapes.
978
979 // We use release_store_fence to update values like the thread state, where
980 // we don't want the current thread to continue until all our prior memory
981 // accesses (including the new thread state) are visible to other threads.
982 __ li(R0/*thread_state*/, _thread_in_native_trans);
983 __ release();
984 __ stw(R0/*thread_state*/, thread_(thread_state));
985 if (UseMembar) {
986 __ fence();
987 }
988 // Write serialization page so that the VM thread can do a pseudo remote
989 // membar. We use the current thread pointer to calculate a thread
990 // specific offset to write to within the page. This minimizes bus
991 // traffic due to cache line collision.
992 else {
993 __ serialize_memory(R16_thread, R11_scratch1, R12_scratch2);
994 }
995
996 // Now before we return to java we must look for a current safepoint
997 // (a new safepoint can not start since we entered native_trans).
998 // We must check here because a current safepoint could be modifying
999 // the callers registers right this moment.
1000
1001 // Acquire isn't strictly necessary here because of the fence, but
1002 // sync_state is declared to be volatile, so we do it anyway
1003 // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
1004 int sync_state_offs = __ load_const_optimized(sync_state_addr, SafepointSynchronize::address_of_state(), /*temp*/R0, true);
1005
1006 // TODO PPC port assert(4 == SafepointSynchronize::sz_state(), "unexpected field size");
1007 __ lwz(sync_state, sync_state_offs, sync_state_addr);
1008
1009 // TODO PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
1010 __ lwz(suspend_flags, thread_(suspend_flags));
1011
1012 Label sync_check_done;
1013 Label do_safepoint;
1014 // No synchronization in progress nor yet synchronized.
1015 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized);
1016 // Not suspended.
1017 __ cmpwi(CCR1, suspend_flags, 0);
1018
1019 __ bne(CCR0, do_safepoint);
1020 __ beq(CCR1, sync_check_done);
1021 __ bind(do_safepoint);
1022 __ isync();
1023 // Block. We do the call directly and leave the current
1024 // last_Java_frame setup undisturbed. We must save any possible
1025 // native result across the call. No oop is present.
1026
1027 __ mr(R3_ARG1, R16_thread);
1028 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, JavaThread::check_special_condition_for_native_trans),
1029 relocInfo::none);
1030
1031 __ bind(sync_check_done);
1032
1033 //=============================================================================
1034 // <<<<<< Back in Interpreter Frame >>>>>
1035
1036 // We are in thread_in_native_trans here and back in the normal
1037 // interpreter frame. We don't have to do anything special about
1038 // safepoints and we can switch to Java mode anytime we are ready.
1039
1040 // Note: frame::interpreter_frame_result has a dependency on how the
1041 // method result is saved across the call to post_method_exit. For
1042 // native methods it assumes that the non-FPU/non-void result is
1043 // saved in _native_lresult and a FPU result in _native_fresult. If
1044 // this changes then the interpreter_frame_result implementation
1045 // will need to be updated too.
1046
1047 // On PPC64, we have stored the result directly after the native call.
1048
1049 //=============================================================================
1050 // Back in Java
1051
1052 // We use release_store_fence to update values like the thread state, where
1053 // we don't want the current thread to continue until all our prior memory
1054 // accesses (including the new thread state) are visible to other threads.
1055 __ li(R0/*thread_state*/, _thread_in_Java);
1056 __ release();
1057 __ stw(R0/*thread_state*/, thread_(thread_state));
1058 if (UseMembar) {
1059 __ fence();
1060 }
1061
1062 __ reset_last_Java_frame();
1063
1064 // Jvmdi/jvmpi support. Whether we've got an exception pending or
1065 // not, and whether unlocking throws an exception or not, we notify
1066 // on native method exit. If we do have an exception, we'll end up
1067 // in the caller's context to handle it, so if we don't do the
1068 // notify here, we'll drop it on the floor.
1069 __ notify_method_exit(true/*native method*/,
1070 ilgl /*illegal state (not used for native methods)*/,
1071 InterpreterMacroAssembler::NotifyJVMTI,
1072 false /*check_exceptions*/);
1073
1074 //=============================================================================
1075 // Handle exceptions
1076
1077 if (synchronized) {
1078 // Don't check for exceptions since we're still in the i2n frame. Do that
1079 // manually afterwards.
1080 unlock_method(false);
1081 }
1082
1083 // Reset active handles after returning from native.
1084 // thread->active_handles()->clear();
1085 __ ld(active_handles, thread_(active_handles));
1086 // TODO PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
1087 __ li(R0, 0);
1088 __ stw(R0, JNIHandleBlock::top_offset_in_bytes(), active_handles);
1089
1090 Label exception_return_sync_check_already_unlocked;
1091 __ ld(R0/*pending_exception*/, thread_(pending_exception));
1092 __ cmpdi(CCR0, R0/*pending_exception*/, 0);
1093 __ bne(CCR0, exception_return_sync_check_already_unlocked);
1094
1095 //-----------------------------------------------------------------------------
1096 // No exception pending.
1097
1098 // Move native method result back into proper registers and return.
1099 // Invoke result handler (may unbox/promote).
1100 __ ld(R11_scratch1, 0, R1_SP);
1101 __ ld(R3_RET, _ijava_state_neg(lresult), R11_scratch1);
1102 __ lfd(F1_RET, _ijava_state_neg(fresult), R11_scratch1);
1103 __ call_stub(result_handler_addr);
1104
1105 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ R0, R11_scratch1, R12_scratch2);
1106
1107 // Must use the return pc which was loaded from the caller's frame
1108 // as the VM uses return-pc-patching for deoptimization.
1109 __ mtlr(R0);
1110 __ blr();
1111
1112 //-----------------------------------------------------------------------------
1113 // An exception is pending. We call into the runtime only if the
1114 // caller was not interpreted. If it was interpreted the
1115 // interpreter will do the correct thing. If it isn't interpreted
1116 // (call stub/compiled code) we will change our return and continue.
1117
1118 __ BIND(exception_return_sync_check);
1119
1120 if (synchronized) {
1121 // Don't check for exceptions since we're still in the i2n frame. Do that
1122 // manually afterwards.
1123 unlock_method(false);
1124 }
1125 __ BIND(exception_return_sync_check_already_unlocked);
1126
1127 const Register return_pc = R31;
1128
1129 __ ld(return_pc, 0, R1_SP);
1130 __ ld(return_pc, _abi(lr), return_pc);
1131
1132 // Get the address of the exception handler.
1133 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
1134 R16_thread,
1135 return_pc /* return pc */);
1136 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, noreg, R11_scratch1, R12_scratch2);
1137
1138 // Load the PC of the the exception handler into LR.
1139 __ mtlr(R3_RET);
1140
1141 // Load exception into R3_ARG1 and clear pending exception in thread.
1142 __ ld(R3_ARG1/*exception*/, thread_(pending_exception));
1143 __ li(R4_ARG2, 0);
1144 __ std(R4_ARG2, thread_(pending_exception));
1145
1146 // Load the original return pc into R4_ARG2.
1147 __ mr(R4_ARG2/*issuing_pc*/, return_pc);
1148
1149 // Return to exception handler.
1150 __ blr();
1151
1152 //=============================================================================
1153 // Counter overflow.
1154
1155 if (inc_counter) {
1156 // Handle invocation counter overflow.
1157 __ bind(invocation_counter_overflow);
1158
1159 generate_counter_overflow(continue_after_compile);
1160 }
1161
1162 return entry;
1163 }
1164
1165 // Generic interpreted method entry to (asm) interpreter.
1166 //
1167 address TemplateInterpreterGenerator::generate_normal_entry(bool synchronized) {
1168 bool inc_counter = UseCompiler || CountCompiledCalls;
1169 address entry = __ pc();
1170 // Generate the code to allocate the interpreter stack frame.
1171 Register Rsize_of_parameters = R4_ARG2, // Written by generate_fixed_frame.
1172 Rsize_of_locals = R5_ARG3; // Written by generate_fixed_frame.
1173
1174 generate_fixed_frame(false, Rsize_of_parameters, Rsize_of_locals);
1175
1176 #ifdef FAST_DISPATCH
1177 __ unimplemented("Fast dispatch in generate_normal_entry");
1178 #if 0
1179 __ set((intptr_t)Interpreter::dispatch_table(), IdispatchTables);
1180 // Set bytecode dispatch table base.
1181 #endif
1182 #endif
1183
1184 // --------------------------------------------------------------------------
1185 // Zero out non-parameter locals.
1186 // Note: *Always* zero out non-parameter locals as Sparc does. It's not
1187 // worth to ask the flag, just do it.
1188 Register Rslot_addr = R6_ARG4,
1189 Rnum = R7_ARG5;
1190 Label Lno_locals, Lzero_loop;
1191
1192 // Set up the zeroing loop.
1193 __ subf(Rnum, Rsize_of_parameters, Rsize_of_locals);
1194 __ subf(Rslot_addr, Rsize_of_parameters, R18_locals);
1195 __ srdi_(Rnum, Rnum, Interpreter::logStackElementSize);
1196 __ beq(CCR0, Lno_locals);
1197 __ li(R0, 0);
1198 __ mtctr(Rnum);
1199
1200 // The zero locals loop.
1201 __ bind(Lzero_loop);
1202 __ std(R0, 0, Rslot_addr);
1203 __ addi(Rslot_addr, Rslot_addr, -Interpreter::stackElementSize);
1204 __ bdnz(Lzero_loop);
1205
1206 __ bind(Lno_locals);
1207
1208 // --------------------------------------------------------------------------
1209 // Counter increment and overflow check.
1210 Label invocation_counter_overflow,
1211 profile_method,
1212 profile_method_continue;
1213 if (inc_counter || ProfileInterpreter) {
1214
1215 Register Rdo_not_unlock_if_synchronized_addr = R11_scratch1;
1216 if (synchronized) {
1217 // Since at this point in the method invocation the exception handler
1218 // would try to exit the monitor of synchronized methods which hasn't
1219 // been entered yet, we set the thread local variable
1220 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
1221 // runtime, exception handling i.e. unlock_if_synchronized_method will
1222 // check this thread local flag.
1223 // This flag has two effects, one is to force an unwind in the topmost
1224 // interpreter frame and not perform an unlock while doing so.
1225 __ li(R0, 1);
1226 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1227 }
1228 // Increment invocation counter and check for overflow.
1229 if (inc_counter) {
1230 generate_counter_incr(&invocation_counter_overflow, &profile_method, &profile_method_continue);
1231 }
1232
1233 __ bind(profile_method_continue);
1234
1235 // Reset the _do_not_unlock_if_synchronized flag.
1236 if (synchronized) {
1237 __ li(R0, 0);
1238 __ stb(R0, in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()), R16_thread);
1239 }
1240 }
1241
1242 // --------------------------------------------------------------------------
1243 // Locking of synchronized methods. Must happen AFTER invocation_counter
1244 // check and stack overflow check, so method is not locked if overflows.
1245 if (synchronized) {
1246 lock_method(R3_ARG1, R4_ARG2, R5_ARG3);
1247 }
1248 #ifdef ASSERT
1249 else {
1250 Label Lok;
1251 __ lwz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1252 __ andi_(R0, R0, JVM_ACC_SYNCHRONIZED);
1253 __ asm_assert_eq("method needs synchronization", 0x8521);
1254 __ bind(Lok);
1255 }
1256 #endif // ASSERT
1257
1258 __ verify_thread();
1259
1260 // --------------------------------------------------------------------------
1261 // JVMTI support
1262 __ notify_method_entry();
1263
1264 // --------------------------------------------------------------------------
1265 // Start executing instructions.
1266 __ dispatch_next(vtos);
1267
1268 // --------------------------------------------------------------------------
1269 // Out of line counter overflow and MDO creation code.
1270 if (ProfileInterpreter) {
1271 // We have decided to profile this method in the interpreter.
1272 __ bind(profile_method);
1273 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1274 __ set_method_data_pointer_for_bcp();
1275 __ b(profile_method_continue);
1276 }
1277
1278 if (inc_counter) {
1279 // Handle invocation counter overflow.
1280 __ bind(invocation_counter_overflow);
1281 generate_counter_overflow(profile_method_continue);
1282 }
1283 return entry;
1284 }
1285
1286 // =============================================================================
1287 // Entry points
1288
1289 address AbstractInterpreterGenerator::generate_method_entry(
1290 AbstractInterpreter::MethodKind kind) {
1291 // Determine code generation flags.
1292 bool synchronized = false;
1293 address entry_point = NULL;
1294
1295 switch (kind) {
1296 case Interpreter::zerolocals : break;
1297 case Interpreter::zerolocals_synchronized: synchronized = true; break;
1298 case Interpreter::native : entry_point = ((InterpreterGenerator*) this)->generate_native_entry(false); break;
1299 case Interpreter::native_synchronized : entry_point = ((InterpreterGenerator*) this)->generate_native_entry(true); break;
1300 case Interpreter::empty : entry_point = ((InterpreterGenerator*) this)->generate_empty_entry(); break;
1301 case Interpreter::accessor : entry_point = ((InterpreterGenerator*) this)->generate_accessor_entry(); break;
1302 case Interpreter::abstract : entry_point = ((InterpreterGenerator*) this)->generate_abstract_entry(); break;
1303
1304 case Interpreter::java_lang_math_sin : // fall thru
1305 case Interpreter::java_lang_math_cos : // fall thru
1306 case Interpreter::java_lang_math_tan : // fall thru
1307 case Interpreter::java_lang_math_abs : // fall thru
1308 case Interpreter::java_lang_math_log : // fall thru
1309 case Interpreter::java_lang_math_log10 : // fall thru
1310 case Interpreter::java_lang_math_sqrt : // fall thru
1311 case Interpreter::java_lang_math_pow : // fall thru
1312 case Interpreter::java_lang_math_exp : entry_point = ((InterpreterGenerator*) this)->generate_math_entry(kind); break;
1313 case Interpreter::java_lang_ref_reference_get
1314 : entry_point = ((InterpreterGenerator*)this)->generate_Reference_get_entry(); break;
1315 default : ShouldNotReachHere(); break;
1316 }
1317
1318 if (entry_point) {
1319 return entry_point;
1320 }
1321
1322 return ((InterpreterGenerator*) this)->generate_normal_entry(synchronized);
1323 }
1324
1325 // These should never be compiled since the interpreter will prefer
1326 // the compiled version to the intrinsic version.
1327 bool AbstractInterpreter::can_be_compiled(methodHandle m) {
1328 return !math_entry_available(method_kind(m));
1329 }
1330
1331 // How much stack a method activation needs in stack slots.
1332 // We must calc this exactly like in generate_fixed_frame.
1333 // Note: This returns the conservative size assuming maximum alignment.
1334 int AbstractInterpreter::size_top_interpreter_activation(Method* method) {
1335 const int max_alignment_size = 2;
1336 const int abi_scratch = frame::abi_reg_args_size;
1337 return method->max_locals() + method->max_stack() + frame::interpreter_frame_monitor_size() + max_alignment_size + abi_scratch;
1338 }
1339
1340 // Fills a sceletal interpreter frame generated during deoptimizations
1341 // and returns the frame size in slots.
1342 //
1343 // Parameters:
1344 //
1345 // interpreter_frame == NULL:
1346 // Only calculate the size of an interpreter activation, no actual layout.
1347 // Note: This calculation must exactly parallel the frame setup
1348 // in TemplateInterpreter::generate_normal_entry. But it does not
1349 // account for the SP alignment, that might further enhance the
1350 // frame size, depending on FP.
1351 //
1352 // interpreter_frame != NULL:
1353 // set up the method, locals, and monitors.
1354 // The frame interpreter_frame, if not NULL, is guaranteed to be the
1355 // right size, as determined by a previous call to this method.
1356 // It is also guaranteed to be walkable even though it is in a skeletal state
1357 //
1358 // is_top_frame == true:
1359 // We're processing the *oldest* interpreter frame!
1360 //
1361 // pop_frame_extra_args:
1362 // If this is != 0 we are returning to a deoptimized frame by popping
1363 // off the callee frame. We want to re-execute the call that called the
1364 // callee interpreted, but since the return to the interpreter would pop
1365 // the arguments off advance the esp by dummy popframe_extra_args slots.
1366 // Popping off those will establish the stack layout as it was before the call.
1367 //
1368 int AbstractInterpreter::layout_activation(Method* method,
1369 int tempcount,
1370 int popframe_extra_args,
1371 int moncount,
1372 int caller_actual_parameters,
1373 int callee_param_count,
1374 int callee_locals,
1375 frame* caller,
1376 frame* interpreter_frame,
1377 bool is_top_frame,
1378 bool is_bottom_frame) {
1379
1380 const int max_alignment_space = 2;
1381 const int abi_scratch = is_top_frame ? (frame::abi_reg_args_size / Interpreter::stackElementSize) :
1382 (frame::abi_minframe_size / Interpreter::stackElementSize) ;
1383 const int conservative_framesize_in_slots =
1384 method->max_stack() + callee_locals - callee_param_count +
1385 (moncount * frame::interpreter_frame_monitor_size()) + max_alignment_space +
1386 abi_scratch + frame::ijava_state_size / Interpreter::stackElementSize;
1387
1388 assert(!is_top_frame || conservative_framesize_in_slots * 8 > frame::abi_reg_args_size + frame::ijava_state_size, "frame too small");
1389
1390 if (interpreter_frame == NULL) {
1391 // Since we don't know the exact alignment, we return the conservative size.
1392 return (conservative_framesize_in_slots & -2);
1393 } else {
1394 // Now we know our caller, calc the exact frame layout and size.
1395 intptr_t* locals_base = (caller->is_interpreted_frame()) ?
1396 caller->interpreter_frame_esp() + caller_actual_parameters :
1397 caller->sp() + method->max_locals() - 1 + (frame::abi_minframe_size / Interpreter::stackElementSize) ;
1398
1399 intptr_t* monitor_base = caller->sp() - frame::ijava_state_size / Interpreter::stackElementSize ;
1400 intptr_t* monitor = monitor_base - (moncount * frame::interpreter_frame_monitor_size());
1401 intptr_t* esp_base = monitor - 1;
1402 intptr_t* esp = esp_base - tempcount - popframe_extra_args;
1403 intptr_t* sp = (intptr_t *) (((intptr_t) (esp_base- callee_locals + callee_param_count - method->max_stack()- abi_scratch)) & -StackAlignmentInBytes);
1404 intptr_t* sender_sp = caller->sp() + (frame::abi_minframe_size - frame::abi_reg_args_size) / Interpreter::stackElementSize;
1405 intptr_t* top_frame_sp = is_top_frame ? sp : sp + (frame::abi_minframe_size - frame::abi_reg_args_size) / Interpreter::stackElementSize;
1406
1407 interpreter_frame->interpreter_frame_set_method(method);
1408 interpreter_frame->interpreter_frame_set_locals(locals_base);
1409 interpreter_frame->interpreter_frame_set_cpcache(method->constants()->cache());
1410 interpreter_frame->interpreter_frame_set_esp(esp);
1411 interpreter_frame->interpreter_frame_set_monitor_end((BasicObjectLock *)monitor);
1412 interpreter_frame->interpreter_frame_set_top_frame_sp(top_frame_sp);
1413 if (!is_bottom_frame) {
1414 interpreter_frame->interpreter_frame_set_sender_sp(sender_sp);
1415 }
1416
1417 int framesize_in_slots = caller->sp() - sp;
1418 assert(!is_top_frame ||framesize_in_slots >= (frame::abi_reg_args_size / Interpreter::stackElementSize) + frame::ijava_state_size / Interpreter::stackElementSize, "frame too small");
1419 assert(framesize_in_slots <= conservative_framesize_in_slots, "exact frame size must be smaller than the convervative size!");
1420 return framesize_in_slots;
1421 }
1422 }
1423
1424 // =============================================================================
1425 // Exceptions
1426
1427 void TemplateInterpreterGenerator::generate_throw_exception() {
1428 Register Rexception = R17_tos,
1429 Rcontinuation = R3_RET;
1430
1431 // --------------------------------------------------------------------------
1432 // Entry point if an method returns with a pending exception (rethrow).
1433 Interpreter::_rethrow_exception_entry = __ pc();
1434 {
1435 __ restore_interpreter_state(R11_scratch1); // Sets R11_scratch1 = fp.
1436 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
1437 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
1438
1439 // Compiled code destroys templateTableBase, reload.
1440 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
1441 }
1442
1443 // Entry point if a interpreted method throws an exception (throw).
1444 Interpreter::_throw_exception_entry = __ pc();
1445 {
1446 __ mr(Rexception, R3_RET);
1447
1448 __ verify_thread();
1449 __ verify_oop(Rexception);
1450
1451 // Expression stack must be empty before entering the VM in case of an exception.
1452 __ empty_expression_stack();
1453 // Find exception handler address and preserve exception oop.
1454 // Call C routine to find handler and jump to it.
1455 __ call_VM(Rexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::exception_handler_for_exception), Rexception);
1456 __ mtctr(Rcontinuation);
1457 // Push exception for exception handler bytecodes.
1458 __ push_ptr(Rexception);
1459
1460 // Jump to exception handler (may be remove activation entry!).
1461 __ bctr();
1462 }
1463
1464 // If the exception is not handled in the current frame the frame is
1465 // removed and the exception is rethrown (i.e. exception
1466 // continuation is _rethrow_exception).
1467 //
1468 // Note: At this point the bci is still the bxi for the instruction
1469 // which caused the exception and the expression stack is
1470 // empty. Thus, for any VM calls at this point, GC will find a legal
1471 // oop map (with empty expression stack).
1472
1473 // In current activation
1474 // tos: exception
1475 // bcp: exception bcp
1476
1477 // --------------------------------------------------------------------------
1478 // JVMTI PopFrame support
1479
1480 Interpreter::_remove_activation_preserving_args_entry = __ pc();
1481 {
1482 // Set the popframe_processing bit in popframe_condition indicating that we are
1483 // currently handling popframe, so that call_VMs that may happen later do not
1484 // trigger new popframe handling cycles.
1485 __ lwz(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
1486 __ ori(R11_scratch1, R11_scratch1, JavaThread::popframe_processing_bit);
1487 __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
1488
1489 // Empty the expression stack, as in normal exception handling.
1490 __ empty_expression_stack();
1491 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, /* install_monitor_exception */ false);
1492
1493 // Check to see whether we are returning to a deoptimized frame.
1494 // (The PopFrame call ensures that the caller of the popped frame is
1495 // either interpreted or compiled and deoptimizes it if compiled.)
1496 // Note that we don't compare the return PC against the
1497 // deoptimization blob's unpack entry because of the presence of
1498 // adapter frames in C2.
1499 Label Lcaller_not_deoptimized;
1500 Register return_pc = R3_ARG1;
1501 __ ld(return_pc, 0, R1_SP);
1502 __ ld(return_pc, _abi(lr), return_pc);
1503 __ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::interpreter_contains), return_pc);
1504 __ cmpdi(CCR0, R3_RET, 0);
1505 __ bne(CCR0, Lcaller_not_deoptimized);
1506
1507 // The deoptimized case.
1508 // In this case, we can't call dispatch_next() after the frame is
1509 // popped, but instead must save the incoming arguments and restore
1510 // them after deoptimization has occurred.
1511 __ ld(R4_ARG2, in_bytes(Method::const_offset()), R19_method);
1512 __ lhz(R4_ARG2 /* number of params */, in_bytes(ConstMethod::size_of_parameters_offset()), R4_ARG2);
1513 __ slwi(R4_ARG2, R4_ARG2, Interpreter::logStackElementSize);
1514 __ addi(R5_ARG3, R18_locals, Interpreter::stackElementSize);
1515 __ subf(R5_ARG3, R4_ARG2, R5_ARG3);
1516 // Save these arguments.
1517 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::popframe_preserve_args), R16_thread, R4_ARG2, R5_ARG3);
1518
1519 // Inform deoptimization that it is responsible for restoring these arguments.
1520 __ load_const_optimized(R11_scratch1, JavaThread::popframe_force_deopt_reexecution_bit);
1521 __ stw(R11_scratch1, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
1522
1523 // Return from the current method into the deoptimization blob. Will eventually
1524 // end up in the deopt interpeter entry, deoptimization prepared everything that
1525 // we will reexecute the call that called us.
1526 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*reload return_pc*/ return_pc, R11_scratch1, R12_scratch2);
1527 __ mtlr(return_pc);
1528 __ blr();
1529
1530 // The non-deoptimized case.
1531 __ bind(Lcaller_not_deoptimized);
1532
1533 // Clear the popframe condition flag.
1534 __ li(R0, 0);
1535 __ stw(R0, in_bytes(JavaThread::popframe_condition_offset()), R16_thread);
1536
1537 // Get out of the current method and re-execute the call that called us.
1538 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ return_pc, R11_scratch1, R12_scratch2);
1539 __ restore_interpreter_state(R11_scratch1);
1540 __ ld(R12_scratch2, _ijava_state_neg(top_frame_sp), R11_scratch1);
1541 __ resize_frame_absolute(R12_scratch2, R11_scratch1, R0);
1542 __ mtlr(return_pc);
1543 if (ProfileInterpreter) {
1544 __ set_method_data_pointer_for_bcp();
1545 }
1546 __ dispatch_next(vtos);
1547 }
1548 // end of JVMTI PopFrame support
1549
1550 // --------------------------------------------------------------------------
1551 // Remove activation exception entry.
1552 // This is jumped to if an interpreted method can't handle an exception itself
1553 // (we come from the throw/rethrow exception entry above). We're going to call
1554 // into the VM to find the exception handler in the caller, pop the current
1555 // frame and return the handler we calculated.
1556 Interpreter::_remove_activation_entry = __ pc();
1557 {
1558 __ pop_ptr(Rexception);
1559 __ verify_thread();
1560 __ verify_oop(Rexception);
1561 __ std(Rexception, in_bytes(JavaThread::vm_result_offset()), R16_thread);
1562
1563 __ unlock_if_synchronized_method(vtos, /* throw_monitor_exception */ false, true);
1564 __ notify_method_exit(false, vtos, InterpreterMacroAssembler::SkipNotifyJVMTI, false);
1565
1566 __ get_vm_result(Rexception);
1567
1568 // We are done with this activation frame; find out where to go next.
1569 // The continuation point will be an exception handler, which expects
1570 // the following registers set up:
1571 //
1572 // RET: exception oop
1573 // ARG2: Issuing PC (see generate_exception_blob()), only used if the caller is compiled.
1574
1575 Register return_pc = R31; // Needs to survive the runtime call.
1576 __ ld(return_pc, 0, R1_SP);
1577 __ ld(return_pc, _abi(lr), return_pc);
1578 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, return_pc);
1579
1580 // Remove the current activation.
1581 __ merge_frames(/*top_frame_sp*/ R21_sender_SP, /*return_pc*/ noreg, R11_scratch1, R12_scratch2);
1582
1583 __ mr(R4_ARG2, return_pc);
1584 __ mtlr(R3_RET);
1585 __ mr(R3_RET, Rexception);
1586 __ blr();
1587 }
1588 }
1589
1590 // JVMTI ForceEarlyReturn support.
1591 // Returns "in the middle" of a method with a "fake" return value.
1592 address TemplateInterpreterGenerator::generate_earlyret_entry_for(TosState state) {
1593
1594 Register Rscratch1 = R11_scratch1,
1595 Rscratch2 = R12_scratch2;
1596
1597 address entry = __ pc();
1598 __ empty_expression_stack();
1599
1600 __ load_earlyret_value(state, Rscratch1);
1601
1602 __ ld(Rscratch1, in_bytes(JavaThread::jvmti_thread_state_offset()), R16_thread);
1603 // Clear the earlyret state.
1604 __ li(R0, 0);
1605 __ stw(R0, in_bytes(JvmtiThreadState::earlyret_state_offset()), Rscratch1);
1606
1607 __ remove_activation(state, false, false);
1608 // Copied from TemplateTable::_return.
1609 // Restoration of lr done by remove_activation.
1610 switch (state) {
1611 case ltos:
1612 case btos:
1613 case ctos:
1614 case stos:
1615 case atos:
1616 case itos: __ mr(R3_RET, R17_tos); break;
1617 case ftos:
1618 case dtos: __ fmr(F1_RET, F15_ftos); break;
1619 case vtos: // This might be a constructor. Final fields (and volatile fields on PPC64) need
1620 // to get visible before the reference to the object gets stored anywhere.
1621 __ membar(Assembler::StoreStore); break;
1622 default : ShouldNotReachHere();
1623 }
1624 __ blr();
1625
1626 return entry;
1627 } // end of ForceEarlyReturn support
1628
1629 //-----------------------------------------------------------------------------
1630 // Helper for vtos entry point generation
1631
1632 void TemplateInterpreterGenerator::set_vtos_entry_points(Template* t,
1633 address& bep,
1634 address& cep,
1635 address& sep,
1636 address& aep,
1637 address& iep,
1638 address& lep,
1639 address& fep,
1640 address& dep,
1641 address& vep) {
1642 assert(t->is_valid() && t->tos_in() == vtos, "illegal template");
1643 Label L;
1644
1645 aep = __ pc(); __ push_ptr(); __ b(L);
1646 fep = __ pc(); __ push_f(); __ b(L);
1647 dep = __ pc(); __ push_d(); __ b(L);
1648 lep = __ pc(); __ push_l(); __ b(L);
1649 __ align(32, 12, 24); // align L
1650 bep = cep = sep =
1651 iep = __ pc(); __ push_i();
1652 vep = __ pc();
1653 __ bind(L);
1654 generate_and_dispatch(t);
1655 }
1656
1657 //-----------------------------------------------------------------------------
1658 // Generation of individual instructions
1659
1660 // helpers for generate_and_dispatch
1661
1662 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
1663 : TemplateInterpreterGenerator(code) {
1664 generate_all(); // Down here so it can be "virtual".
1665 }
1666
1667 //-----------------------------------------------------------------------------
1668
1669 // Non-product code
1670 #ifndef PRODUCT
1671 address TemplateInterpreterGenerator::generate_trace_code(TosState state) {
1672 //__ flush_bundle();
1673 address entry = __ pc();
1674
1675 char *bname = NULL;
1676 uint tsize = 0;
1677 switch(state) {
1678 case ftos:
1679 bname = "trace_code_ftos {";
1680 tsize = 2;
1681 break;
1682 case btos:
1683 bname = "trace_code_btos {";
1684 tsize = 2;
1685 break;
1686 case ctos:
1687 bname = "trace_code_ctos {";
1688 tsize = 2;
1689 break;
1690 case stos:
1691 bname = "trace_code_stos {";
1692 tsize = 2;
1693 break;
1694 case itos:
1695 bname = "trace_code_itos {";
1696 tsize = 2;
1697 break;
1698 case ltos:
1699 bname = "trace_code_ltos {";
1700 tsize = 3;
1701 break;
1702 case atos:
1703 bname = "trace_code_atos {";
1704 tsize = 2;
1705 break;
1706 case vtos:
1707 // Note: In case of vtos, the topmost of stack value could be a int or doubl
1708 // In case of a double (2 slots) we won't see the 2nd stack value.
1709 // Maybe we simply should print the topmost 3 stack slots to cope with the problem.
1710 bname = "trace_code_vtos {";
1711 tsize = 2;
1712
1713 break;
1714 case dtos:
1715 bname = "trace_code_dtos {";
1716 tsize = 3;
1717 break;
1718 default:
1719 ShouldNotReachHere();
1720 }
1721 BLOCK_COMMENT(bname);
1722
1723 // Support short-cut for TraceBytecodesAt.
1724 // Don't call into the VM if we don't want to trace to speed up things.
1725 Label Lskip_vm_call;
1726 if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) {
1727 int offs1 = __ load_const_optimized(R11_scratch1, (address) &TraceBytecodesAt, R0, true);
1728 int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true);
1729 __ ld(R11_scratch1, offs1, R11_scratch1);
1730 __ lwa(R12_scratch2, offs2, R12_scratch2);
1731 __ cmpd(CCR0, R12_scratch2, R11_scratch1);
1732 __ blt(CCR0, Lskip_vm_call);
1733 }
1734
1735 __ push(state);
1736 // Load 2 topmost expression stack values.
1737 __ ld(R6_ARG4, tsize*Interpreter::stackElementSize, R15_esp);
1738 __ ld(R5_ARG3, Interpreter::stackElementSize, R15_esp);
1739 __ mflr(R31);
1740 __ call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::trace_bytecode), /* unused */ R4_ARG2, R5_ARG3, R6_ARG4, false);
1741 __ mtlr(R31);
1742 __ pop(state);
1743
1744 if (TraceBytecodesAt > 0 && TraceBytecodesAt < max_intx) {
1745 __ bind(Lskip_vm_call);
1746 }
1747 __ blr();
1748 BLOCK_COMMENT("} trace_code");
1749 return entry;
1750 }
1751
1752 void TemplateInterpreterGenerator::count_bytecode() {
1753 int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeCounter::_counter_value, R12_scratch2, true);
1754 __ lwz(R12_scratch2, offs, R11_scratch1);
1755 __ addi(R12_scratch2, R12_scratch2, 1);
1756 __ stw(R12_scratch2, offs, R11_scratch1);
1757 }
1758
1759 void TemplateInterpreterGenerator::histogram_bytecode(Template* t) {
1760 int offs = __ load_const_optimized(R11_scratch1, (address) &BytecodeHistogram::_counters[t->bytecode()], R12_scratch2, true);
1761 __ lwz(R12_scratch2, offs, R11_scratch1);
1762 __ addi(R12_scratch2, R12_scratch2, 1);
1763 __ stw(R12_scratch2, offs, R11_scratch1);
1764 }
1765
1766 void TemplateInterpreterGenerator::histogram_bytecode_pair(Template* t) {
1767 const Register addr = R11_scratch1,
1768 tmp = R12_scratch2;
1769 // Get index, shift out old bytecode, bring in new bytecode, and store it.
1770 // _index = (_index >> log2_number_of_codes) |
1771 // (bytecode << log2_number_of_codes);
1772 int offs1 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_index, tmp, true);
1773 __ lwz(tmp, offs1, addr);
1774 __ srwi(tmp, tmp, BytecodePairHistogram::log2_number_of_codes);
1775 __ ori(tmp, tmp, ((int) t->bytecode()) << BytecodePairHistogram::log2_number_of_codes);
1776 __ stw(tmp, offs1, addr);
1777
1778 // Bump bucket contents.
1779 // _counters[_index] ++;
1780 int offs2 = __ load_const_optimized(addr, (address)&BytecodePairHistogram::_counters, R0, true);
1781 __ sldi(tmp, tmp, LogBytesPerInt);
1782 __ add(addr, tmp, addr);
1783 __ lwz(tmp, offs2, addr);
1784 __ addi(tmp, tmp, 1);
1785 __ stw(tmp, offs2, addr);
1786 }
1787
1788 void TemplateInterpreterGenerator::trace_bytecode(Template* t) {
1789 // Call a little run-time stub to avoid blow-up for each bytecode.
1790 // The run-time runtime saves the right registers, depending on
1791 // the tosca in-state for the given template.
1792
1793 assert(Interpreter::trace_code(t->tos_in()) != NULL,
1794 "entry must have been generated");
1795
1796 // Note: we destroy LR here.
1797 __ bl(Interpreter::trace_code(t->tos_in()));
1798 }
1799
1800 void TemplateInterpreterGenerator::stop_interpreter_at() {
1801 Label L;
1802 int offs1 = __ load_const_optimized(R11_scratch1, (address) &StopInterpreterAt, R0, true);
1803 int offs2 = __ load_const_optimized(R12_scratch2, (address) &BytecodeCounter::_counter_value, R0, true);
1804 __ ld(R11_scratch1, offs1, R11_scratch1);
1805 __ lwa(R12_scratch2, offs2, R12_scratch2);
1806 __ cmpd(CCR0, R12_scratch2, R11_scratch1);
1807 __ bne(CCR0, L);
1808 __ illtrap();
1809 __ bind(L);
1810 }
1811
1812 #endif // !PRODUCT
1813 #endif // !CC_INTERP