Mercurial > jdk8-mips64-public > hotspot / file revision
src/cpu/x86/vm/templateTable_x86_32.cpp@ea79ab313e98
src/cpu/x86/vm/templateTable_x86_32.cpp
Wed, 30 Oct 2013 15:35:25 +0100
- author
- mgerdin
- date
- Wed, 30 Oct 2013 15:35:25 +0100
- changeset 6059
- ea79ab313e98
- parent 5914
- d13d7aba8c12
- child 6064
- 42790b7e4d48
- permissions
- -rw-r--r--
8027252: Crash in interpreter because get_unsigned_2_byte_index_at_bcp reads 4 bytes
Summary: Use 2-byte loads to load indexes from the byte code stream to avoid out of bounds reads.
Reviewed-by: coleenp, sspitsyn
1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "interpreter/interpreter.hpp"
28 #include "interpreter/interpreterRuntime.hpp"
29 #include "interpreter/templateTable.hpp"
30 #include "memory/universe.inline.hpp"
31 #include "oops/methodData.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "oops/oop.inline.hpp"
34 #include "prims/methodHandles.hpp"
35 #include "runtime/sharedRuntime.hpp"
36 #include "runtime/stubRoutines.hpp"
37 #include "runtime/synchronizer.hpp"
38 #include "utilities/macros.hpp"
40 #ifndef CC_INTERP
41 #define __ _masm->
43 //----------------------------------------------------------------------------------------------------
44 // Platform-dependent initialization
46 void TemplateTable::pd_initialize() {
47 // No i486 specific initialization
48 }
50 //----------------------------------------------------------------------------------------------------
51 // Address computation
53 // local variables
54 static inline Address iaddress(int n) {
55 return Address(rdi, Interpreter::local_offset_in_bytes(n));
56 }
58 static inline Address laddress(int n) { return iaddress(n + 1); }
59 static inline Address haddress(int n) { return iaddress(n + 0); }
60 static inline Address faddress(int n) { return iaddress(n); }
61 static inline Address daddress(int n) { return laddress(n); }
62 static inline Address aaddress(int n) { return iaddress(n); }
64 static inline Address iaddress(Register r) {
65 return Address(rdi, r, Interpreter::stackElementScale());
66 }
67 static inline Address laddress(Register r) {
68 return Address(rdi, r, Interpreter::stackElementScale(), Interpreter::local_offset_in_bytes(1));
69 }
70 static inline Address haddress(Register r) {
71 return Address(rdi, r, Interpreter::stackElementScale(), Interpreter::local_offset_in_bytes(0));
72 }
74 static inline Address faddress(Register r) { return iaddress(r); }
75 static inline Address daddress(Register r) { return laddress(r); }
76 static inline Address aaddress(Register r) { return iaddress(r); }
78 // expression stack
79 // (Note: Must not use symmetric equivalents at_rsp_m1/2 since they store
80 // data beyond the rsp which is potentially unsafe in an MT environment;
81 // an interrupt may overwrite that data.)
82 static inline Address at_rsp () {
83 return Address(rsp, 0);
84 }
86 // At top of Java expression stack which may be different than rsp(). It
87 // isn't for category 1 objects.
88 static inline Address at_tos () {
89 Address tos = Address(rsp, Interpreter::expr_offset_in_bytes(0));
90 return tos;
91 }
93 static inline Address at_tos_p1() {
94 return Address(rsp, Interpreter::expr_offset_in_bytes(1));
95 }
97 static inline Address at_tos_p2() {
98 return Address(rsp, Interpreter::expr_offset_in_bytes(2));
99 }
101 // Condition conversion
102 static Assembler::Condition j_not(TemplateTable::Condition cc) {
103 switch (cc) {
104 case TemplateTable::equal : return Assembler::notEqual;
105 case TemplateTable::not_equal : return Assembler::equal;
106 case TemplateTable::less : return Assembler::greaterEqual;
107 case TemplateTable::less_equal : return Assembler::greater;
108 case TemplateTable::greater : return Assembler::lessEqual;
109 case TemplateTable::greater_equal: return Assembler::less;
110 }
111 ShouldNotReachHere();
112 return Assembler::zero;
113 }
116 //----------------------------------------------------------------------------------------------------
117 // Miscelaneous helper routines
119 // Store an oop (or NULL) at the address described by obj.
120 // If val == noreg this means store a NULL
122 static void do_oop_store(InterpreterMacroAssembler* _masm,
123 Address obj,
124 Register val,
125 BarrierSet::Name barrier,
126 bool precise) {
127 assert(val == noreg || val == rax, "parameter is just for looks");
128 switch (barrier) {
129 #if INCLUDE_ALL_GCS
130 case BarrierSet::G1SATBCT:
131 case BarrierSet::G1SATBCTLogging:
132 {
133 // flatten object address if needed
134 // We do it regardless of precise because we need the registers
135 if (obj.index() == noreg && obj.disp() == 0) {
136 if (obj.base() != rdx) {
137 __ movl(rdx, obj.base());
138 }
139 } else {
140 __ leal(rdx, obj);
141 }
142 __ get_thread(rcx);
143 __ save_bcp();
144 __ g1_write_barrier_pre(rdx /* obj */,
145 rbx /* pre_val */,
146 rcx /* thread */,
147 rsi /* tmp */,
148 val != noreg /* tosca_live */,
149 false /* expand_call */);
151 // Do the actual store
152 // noreg means NULL
153 if (val == noreg) {
154 __ movptr(Address(rdx, 0), NULL_WORD);
155 // No post barrier for NULL
156 } else {
157 __ movl(Address(rdx, 0), val);
158 __ g1_write_barrier_post(rdx /* store_adr */,
159 val /* new_val */,
160 rcx /* thread */,
161 rbx /* tmp */,
162 rsi /* tmp2 */);
163 }
164 __ restore_bcp();
166 }
167 break;
168 #endif // INCLUDE_ALL_GCS
169 case BarrierSet::CardTableModRef:
170 case BarrierSet::CardTableExtension:
171 {
172 if (val == noreg) {
173 __ movptr(obj, NULL_WORD);
174 } else {
175 __ movl(obj, val);
176 // flatten object address if needed
177 if (!precise || (obj.index() == noreg && obj.disp() == 0)) {
178 __ store_check(obj.base());
179 } else {
180 __ leal(rdx, obj);
181 __ store_check(rdx);
182 }
183 }
184 }
185 break;
186 case BarrierSet::ModRef:
187 case BarrierSet::Other:
188 if (val == noreg) {
189 __ movptr(obj, NULL_WORD);
190 } else {
191 __ movl(obj, val);
192 }
193 break;
194 default :
195 ShouldNotReachHere();
197 }
198 }
200 Address TemplateTable::at_bcp(int offset) {
201 assert(_desc->uses_bcp(), "inconsistent uses_bcp information");
202 return Address(rsi, offset);
203 }
206 void TemplateTable::patch_bytecode(Bytecodes::Code bc, Register bc_reg,
207 Register temp_reg, bool load_bc_into_bc_reg/*=true*/,
208 int byte_no) {
209 if (!RewriteBytecodes) return;
210 Label L_patch_done;
212 switch (bc) {
213 case Bytecodes::_fast_aputfield:
214 case Bytecodes::_fast_bputfield:
215 case Bytecodes::_fast_cputfield:
216 case Bytecodes::_fast_dputfield:
217 case Bytecodes::_fast_fputfield:
218 case Bytecodes::_fast_iputfield:
219 case Bytecodes::_fast_lputfield:
220 case Bytecodes::_fast_sputfield:
221 {
222 // We skip bytecode quickening for putfield instructions when
223 // the put_code written to the constant pool cache is zero.
224 // This is required so that every execution of this instruction
225 // calls out to InterpreterRuntime::resolve_get_put to do
226 // additional, required work.
227 assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range");
228 assert(load_bc_into_bc_reg, "we use bc_reg as temp");
229 __ get_cache_and_index_and_bytecode_at_bcp(bc_reg, temp_reg, temp_reg, byte_no, 1);
230 __ movl(bc_reg, bc);
231 __ cmpl(temp_reg, (int) 0);
232 __ jcc(Assembler::zero, L_patch_done); // don't patch
233 }
234 break;
235 default:
236 assert(byte_no == -1, "sanity");
237 // the pair bytecodes have already done the load.
238 if (load_bc_into_bc_reg) {
239 __ movl(bc_reg, bc);
240 }
241 }
243 if (JvmtiExport::can_post_breakpoint()) {
244 Label L_fast_patch;
245 // if a breakpoint is present we can't rewrite the stream directly
246 __ movzbl(temp_reg, at_bcp(0));
247 __ cmpl(temp_reg, Bytecodes::_breakpoint);
248 __ jcc(Assembler::notEqual, L_fast_patch);
249 __ get_method(temp_reg);
250 // Let breakpoint table handling rewrite to quicker bytecode
251 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::set_original_bytecode_at), temp_reg, rsi, bc_reg);
252 #ifndef ASSERT
253 __ jmpb(L_patch_done);
254 #else
255 __ jmp(L_patch_done);
256 #endif
257 __ bind(L_fast_patch);
258 }
260 #ifdef ASSERT
261 Label L_okay;
262 __ load_unsigned_byte(temp_reg, at_bcp(0));
263 __ cmpl(temp_reg, (int)Bytecodes::java_code(bc));
264 __ jccb(Assembler::equal, L_okay);
265 __ cmpl(temp_reg, bc_reg);
266 __ jcc(Assembler::equal, L_okay);
267 __ stop("patching the wrong bytecode");
268 __ bind(L_okay);
269 #endif
271 // patch bytecode
272 __ movb(at_bcp(0), bc_reg);
273 __ bind(L_patch_done);
274 }
276 //----------------------------------------------------------------------------------------------------
277 // Individual instructions
279 void TemplateTable::nop() {
280 transition(vtos, vtos);
281 // nothing to do
282 }
284 void TemplateTable::shouldnotreachhere() {
285 transition(vtos, vtos);
286 __ stop("shouldnotreachhere bytecode");
287 }
291 void TemplateTable::aconst_null() {
292 transition(vtos, atos);
293 __ xorptr(rax, rax);
294 }
297 void TemplateTable::iconst(int value) {
298 transition(vtos, itos);
299 if (value == 0) {
300 __ xorptr(rax, rax);
301 } else {
302 __ movptr(rax, value);
303 }
304 }
307 void TemplateTable::lconst(int value) {
308 transition(vtos, ltos);
309 if (value == 0) {
310 __ xorptr(rax, rax);
311 } else {
312 __ movptr(rax, value);
313 }
314 assert(value >= 0, "check this code");
315 __ xorptr(rdx, rdx);
316 }
319 void TemplateTable::fconst(int value) {
320 transition(vtos, ftos);
321 if (value == 0) { __ fldz();
322 } else if (value == 1) { __ fld1();
323 } else if (value == 2) { __ fld1(); __ fld1(); __ faddp(); // should do a better solution here
324 } else { ShouldNotReachHere();
325 }
326 }
329 void TemplateTable::dconst(int value) {
330 transition(vtos, dtos);
331 if (value == 0) { __ fldz();
332 } else if (value == 1) { __ fld1();
333 } else { ShouldNotReachHere();
334 }
335 }
338 void TemplateTable::bipush() {
339 transition(vtos, itos);
340 __ load_signed_byte(rax, at_bcp(1));
341 }
344 void TemplateTable::sipush() {
345 transition(vtos, itos);
346 __ load_unsigned_short(rax, at_bcp(1));
347 __ bswapl(rax);
348 __ sarl(rax, 16);
349 }
351 void TemplateTable::ldc(bool wide) {
352 transition(vtos, vtos);
353 Label call_ldc, notFloat, notClass, Done;
355 if (wide) {
356 __ get_unsigned_2_byte_index_at_bcp(rbx, 1);
357 } else {
358 __ load_unsigned_byte(rbx, at_bcp(1));
359 }
360 __ get_cpool_and_tags(rcx, rax);
361 const int base_offset = ConstantPool::header_size() * wordSize;
362 const int tags_offset = Array<u1>::base_offset_in_bytes();
364 // get type
365 __ xorptr(rdx, rdx);
366 __ movb(rdx, Address(rax, rbx, Address::times_1, tags_offset));
368 // unresolved class - get the resolved class
369 __ cmpl(rdx, JVM_CONSTANT_UnresolvedClass);
370 __ jccb(Assembler::equal, call_ldc);
372 // unresolved class in error (resolution failed) - call into runtime
373 // so that the same error from first resolution attempt is thrown.
374 __ cmpl(rdx, JVM_CONSTANT_UnresolvedClassInError);
375 __ jccb(Assembler::equal, call_ldc);
377 // resolved class - need to call vm to get java mirror of the class
378 __ cmpl(rdx, JVM_CONSTANT_Class);
379 __ jcc(Assembler::notEqual, notClass);
381 __ bind(call_ldc);
382 __ movl(rcx, wide);
383 call_VM(rax, CAST_FROM_FN_PTR(address, InterpreterRuntime::ldc), rcx);
384 __ push(atos);
385 __ jmp(Done);
387 __ bind(notClass);
388 __ cmpl(rdx, JVM_CONSTANT_Float);
389 __ jccb(Assembler::notEqual, notFloat);
390 // ftos
391 __ fld_s( Address(rcx, rbx, Address::times_ptr, base_offset));
392 __ push(ftos);
393 __ jmp(Done);
395 __ bind(notFloat);
396 #ifdef ASSERT
397 { Label L;
398 __ cmpl(rdx, JVM_CONSTANT_Integer);
399 __ jcc(Assembler::equal, L);
400 // String and Object are rewritten to fast_aldc
401 __ stop("unexpected tag type in ldc");
402 __ bind(L);
403 }
404 #endif
405 // itos JVM_CONSTANT_Integer only
406 __ movl(rax, Address(rcx, rbx, Address::times_ptr, base_offset));
407 __ push(itos);
408 __ bind(Done);
409 }
411 // Fast path for caching oop constants.
412 void TemplateTable::fast_aldc(bool wide) {
413 transition(vtos, atos);
415 Register result = rax;
416 Register tmp = rdx;
417 int index_size = wide ? sizeof(u2) : sizeof(u1);
419 Label resolved;
421 // We are resolved if the resolved reference cache entry contains a
422 // non-null object (String, MethodType, etc.)
423 assert_different_registers(result, tmp);
424 __ get_cache_index_at_bcp(tmp, 1, index_size);
425 __ load_resolved_reference_at_index(result, tmp);
426 __ testl(result, result);
427 __ jcc(Assembler::notZero, resolved);
429 address entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_ldc);
431 // first time invocation - must resolve first
432 __ movl(tmp, (int)bytecode());
433 __ call_VM(result, entry, tmp);
435 __ bind(resolved);
437 if (VerifyOops) {
438 __ verify_oop(result);
439 }
440 }
442 void TemplateTable::ldc2_w() {
443 transition(vtos, vtos);
444 Label Long, Done;
445 __ get_unsigned_2_byte_index_at_bcp(rbx, 1);
447 __ get_cpool_and_tags(rcx, rax);
448 const int base_offset = ConstantPool::header_size() * wordSize;
449 const int tags_offset = Array<u1>::base_offset_in_bytes();
451 // get type
452 __ cmpb(Address(rax, rbx, Address::times_1, tags_offset), JVM_CONSTANT_Double);
453 __ jccb(Assembler::notEqual, Long);
454 // dtos
455 __ fld_d( Address(rcx, rbx, Address::times_ptr, base_offset));
456 __ push(dtos);
457 __ jmpb(Done);
459 __ bind(Long);
460 // ltos
461 __ movptr(rax, Address(rcx, rbx, Address::times_ptr, base_offset + 0 * wordSize));
462 NOT_LP64(__ movptr(rdx, Address(rcx, rbx, Address::times_ptr, base_offset + 1 * wordSize)));
464 __ push(ltos);
466 __ bind(Done);
467 }
470 void TemplateTable::locals_index(Register reg, int offset) {
471 __ load_unsigned_byte(reg, at_bcp(offset));
472 __ negptr(reg);
473 }
476 void TemplateTable::iload() {
477 transition(vtos, itos);
478 if (RewriteFrequentPairs) {
479 Label rewrite, done;
481 // get next byte
482 __ load_unsigned_byte(rbx, at_bcp(Bytecodes::length_for(Bytecodes::_iload)));
483 // if _iload, wait to rewrite to iload2. We only want to rewrite the
484 // last two iloads in a pair. Comparing against fast_iload means that
485 // the next bytecode is neither an iload or a caload, and therefore
486 // an iload pair.
487 __ cmpl(rbx, Bytecodes::_iload);
488 __ jcc(Assembler::equal, done);
490 __ cmpl(rbx, Bytecodes::_fast_iload);
491 __ movl(rcx, Bytecodes::_fast_iload2);
492 __ jccb(Assembler::equal, rewrite);
494 // if _caload, rewrite to fast_icaload
495 __ cmpl(rbx, Bytecodes::_caload);
496 __ movl(rcx, Bytecodes::_fast_icaload);
497 __ jccb(Assembler::equal, rewrite);
499 // rewrite so iload doesn't check again.
500 __ movl(rcx, Bytecodes::_fast_iload);
502 // rewrite
503 // rcx: fast bytecode
504 __ bind(rewrite);
505 patch_bytecode(Bytecodes::_iload, rcx, rbx, false);
506 __ bind(done);
507 }
509 // Get the local value into tos
510 locals_index(rbx);
511 __ movl(rax, iaddress(rbx));
512 }
515 void TemplateTable::fast_iload2() {
516 transition(vtos, itos);
517 locals_index(rbx);
518 __ movl(rax, iaddress(rbx));
519 __ push(itos);
520 locals_index(rbx, 3);
521 __ movl(rax, iaddress(rbx));
522 }
524 void TemplateTable::fast_iload() {
525 transition(vtos, itos);
526 locals_index(rbx);
527 __ movl(rax, iaddress(rbx));
528 }
531 void TemplateTable::lload() {
532 transition(vtos, ltos);
533 locals_index(rbx);
534 __ movptr(rax, laddress(rbx));
535 NOT_LP64(__ movl(rdx, haddress(rbx)));
536 }
539 void TemplateTable::fload() {
540 transition(vtos, ftos);
541 locals_index(rbx);
542 __ fld_s(faddress(rbx));
543 }
546 void TemplateTable::dload() {
547 transition(vtos, dtos);
548 locals_index(rbx);
549 __ fld_d(daddress(rbx));
550 }
553 void TemplateTable::aload() {
554 transition(vtos, atos);
555 locals_index(rbx);
556 __ movptr(rax, aaddress(rbx));
557 }
560 void TemplateTable::locals_index_wide(Register reg) {
561 __ load_unsigned_short(reg, at_bcp(2));
562 __ bswapl(reg);
563 __ shrl(reg, 16);
564 __ negptr(reg);
565 }
568 void TemplateTable::wide_iload() {
569 transition(vtos, itos);
570 locals_index_wide(rbx);
571 __ movl(rax, iaddress(rbx));
572 }
575 void TemplateTable::wide_lload() {
576 transition(vtos, ltos);
577 locals_index_wide(rbx);
578 __ movptr(rax, laddress(rbx));
579 NOT_LP64(__ movl(rdx, haddress(rbx)));
580 }
583 void TemplateTable::wide_fload() {
584 transition(vtos, ftos);
585 locals_index_wide(rbx);
586 __ fld_s(faddress(rbx));
587 }
590 void TemplateTable::wide_dload() {
591 transition(vtos, dtos);
592 locals_index_wide(rbx);
593 __ fld_d(daddress(rbx));
594 }
597 void TemplateTable::wide_aload() {
598 transition(vtos, atos);
599 locals_index_wide(rbx);
600 __ movptr(rax, aaddress(rbx));
601 }
603 void TemplateTable::index_check(Register array, Register index) {
604 // Pop ptr into array
605 __ pop_ptr(array);
606 index_check_without_pop(array, index);
607 }
609 void TemplateTable::index_check_without_pop(Register array, Register index) {
610 // destroys rbx,
611 // check array
612 __ null_check(array, arrayOopDesc::length_offset_in_bytes());
613 LP64_ONLY(__ movslq(index, index));
614 // check index
615 __ cmpl(index, Address(array, arrayOopDesc::length_offset_in_bytes()));
616 if (index != rbx) {
617 // ??? convention: move aberrant index into rbx, for exception message
618 assert(rbx != array, "different registers");
619 __ mov(rbx, index);
620 }
621 __ jump_cc(Assembler::aboveEqual,
622 ExternalAddress(Interpreter::_throw_ArrayIndexOutOfBoundsException_entry));
623 }
626 void TemplateTable::iaload() {
627 transition(itos, itos);
628 // rdx: array
629 index_check(rdx, rax); // kills rbx,
630 // rax,: index
631 __ movl(rax, Address(rdx, rax, Address::times_4, arrayOopDesc::base_offset_in_bytes(T_INT)));
632 }
635 void TemplateTable::laload() {
636 transition(itos, ltos);
637 // rax,: index
638 // rdx: array
639 index_check(rdx, rax);
640 __ mov(rbx, rax);
641 // rbx,: index
642 __ movptr(rax, Address(rdx, rbx, Address::times_8, arrayOopDesc::base_offset_in_bytes(T_LONG) + 0 * wordSize));
643 NOT_LP64(__ movl(rdx, Address(rdx, rbx, Address::times_8, arrayOopDesc::base_offset_in_bytes(T_LONG) + 1 * wordSize)));
644 }
647 void TemplateTable::faload() {
648 transition(itos, ftos);
649 // rdx: array
650 index_check(rdx, rax); // kills rbx,
651 // rax,: index
652 __ fld_s(Address(rdx, rax, Address::times_4, arrayOopDesc::base_offset_in_bytes(T_FLOAT)));
653 }
656 void TemplateTable::daload() {
657 transition(itos, dtos);
658 // rdx: array
659 index_check(rdx, rax); // kills rbx,
660 // rax,: index
661 __ fld_d(Address(rdx, rax, Address::times_8, arrayOopDesc::base_offset_in_bytes(T_DOUBLE)));
662 }
665 void TemplateTable::aaload() {
666 transition(itos, atos);
667 // rdx: array
668 index_check(rdx, rax); // kills rbx,
669 // rax,: index
670 __ movptr(rax, Address(rdx, rax, Address::times_ptr, arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
671 }
674 void TemplateTable::baload() {
675 transition(itos, itos);
676 // rdx: array
677 index_check(rdx, rax); // kills rbx,
678 // rax,: index
679 // can do better code for P5 - fix this at some point
680 __ load_signed_byte(rbx, Address(rdx, rax, Address::times_1, arrayOopDesc::base_offset_in_bytes(T_BYTE)));
681 __ mov(rax, rbx);
682 }
685 void TemplateTable::caload() {
686 transition(itos, itos);
687 // rdx: array
688 index_check(rdx, rax); // kills rbx,
689 // rax,: index
690 // can do better code for P5 - may want to improve this at some point
691 __ load_unsigned_short(rbx, Address(rdx, rax, Address::times_2, arrayOopDesc::base_offset_in_bytes(T_CHAR)));
692 __ mov(rax, rbx);
693 }
695 // iload followed by caload frequent pair
696 void TemplateTable::fast_icaload() {
697 transition(vtos, itos);
698 // load index out of locals
699 locals_index(rbx);
700 __ movl(rax, iaddress(rbx));
702 // rdx: array
703 index_check(rdx, rax);
704 // rax,: index
705 __ load_unsigned_short(rbx, Address(rdx, rax, Address::times_2, arrayOopDesc::base_offset_in_bytes(T_CHAR)));
706 __ mov(rax, rbx);
707 }
709 void TemplateTable::saload() {
710 transition(itos, itos);
711 // rdx: array
712 index_check(rdx, rax); // kills rbx,
713 // rax,: index
714 // can do better code for P5 - may want to improve this at some point
715 __ load_signed_short(rbx, Address(rdx, rax, Address::times_2, arrayOopDesc::base_offset_in_bytes(T_SHORT)));
716 __ mov(rax, rbx);
717 }
720 void TemplateTable::iload(int n) {
721 transition(vtos, itos);
722 __ movl(rax, iaddress(n));
723 }
726 void TemplateTable::lload(int n) {
727 transition(vtos, ltos);
728 __ movptr(rax, laddress(n));
729 NOT_LP64(__ movptr(rdx, haddress(n)));
730 }
733 void TemplateTable::fload(int n) {
734 transition(vtos, ftos);
735 __ fld_s(faddress(n));
736 }
739 void TemplateTable::dload(int n) {
740 transition(vtos, dtos);
741 __ fld_d(daddress(n));
742 }
745 void TemplateTable::aload(int n) {
746 transition(vtos, atos);
747 __ movptr(rax, aaddress(n));
748 }
751 void TemplateTable::aload_0() {
752 transition(vtos, atos);
753 // According to bytecode histograms, the pairs:
754 //
755 // _aload_0, _fast_igetfield
756 // _aload_0, _fast_agetfield
757 // _aload_0, _fast_fgetfield
758 //
759 // occur frequently. If RewriteFrequentPairs is set, the (slow) _aload_0
760 // bytecode checks if the next bytecode is either _fast_igetfield,
761 // _fast_agetfield or _fast_fgetfield and then rewrites the
762 // current bytecode into a pair bytecode; otherwise it rewrites the current
763 // bytecode into _fast_aload_0 that doesn't do the pair check anymore.
764 //
765 // Note: If the next bytecode is _getfield, the rewrite must be delayed,
766 // otherwise we may miss an opportunity for a pair.
767 //
768 // Also rewrite frequent pairs
769 // aload_0, aload_1
770 // aload_0, iload_1
771 // These bytecodes with a small amount of code are most profitable to rewrite
772 if (RewriteFrequentPairs) {
773 Label rewrite, done;
774 // get next byte
775 __ load_unsigned_byte(rbx, at_bcp(Bytecodes::length_for(Bytecodes::_aload_0)));
777 // do actual aload_0
778 aload(0);
780 // if _getfield then wait with rewrite
781 __ cmpl(rbx, Bytecodes::_getfield);
782 __ jcc(Assembler::equal, done);
784 // if _igetfield then reqrite to _fast_iaccess_0
785 assert(Bytecodes::java_code(Bytecodes::_fast_iaccess_0) == Bytecodes::_aload_0, "fix bytecode definition");
786 __ cmpl(rbx, Bytecodes::_fast_igetfield);
787 __ movl(rcx, Bytecodes::_fast_iaccess_0);
788 __ jccb(Assembler::equal, rewrite);
790 // if _agetfield then reqrite to _fast_aaccess_0
791 assert(Bytecodes::java_code(Bytecodes::_fast_aaccess_0) == Bytecodes::_aload_0, "fix bytecode definition");
792 __ cmpl(rbx, Bytecodes::_fast_agetfield);
793 __ movl(rcx, Bytecodes::_fast_aaccess_0);
794 __ jccb(Assembler::equal, rewrite);
796 // if _fgetfield then reqrite to _fast_faccess_0
797 assert(Bytecodes::java_code(Bytecodes::_fast_faccess_0) == Bytecodes::_aload_0, "fix bytecode definition");
798 __ cmpl(rbx, Bytecodes::_fast_fgetfield);
799 __ movl(rcx, Bytecodes::_fast_faccess_0);
800 __ jccb(Assembler::equal, rewrite);
802 // else rewrite to _fast_aload0
803 assert(Bytecodes::java_code(Bytecodes::_fast_aload_0) == Bytecodes::_aload_0, "fix bytecode definition");
804 __ movl(rcx, Bytecodes::_fast_aload_0);
806 // rewrite
807 // rcx: fast bytecode
808 __ bind(rewrite);
809 patch_bytecode(Bytecodes::_aload_0, rcx, rbx, false);
811 __ bind(done);
812 } else {
813 aload(0);
814 }
815 }
817 void TemplateTable::istore() {
818 transition(itos, vtos);
819 locals_index(rbx);
820 __ movl(iaddress(rbx), rax);
821 }
824 void TemplateTable::lstore() {
825 transition(ltos, vtos);
826 locals_index(rbx);
827 __ movptr(laddress(rbx), rax);
828 NOT_LP64(__ movptr(haddress(rbx), rdx));
829 }
832 void TemplateTable::fstore() {
833 transition(ftos, vtos);
834 locals_index(rbx);
835 __ fstp_s(faddress(rbx));
836 }
839 void TemplateTable::dstore() {
840 transition(dtos, vtos);
841 locals_index(rbx);
842 __ fstp_d(daddress(rbx));
843 }
846 void TemplateTable::astore() {
847 transition(vtos, vtos);
848 __ pop_ptr(rax);
849 locals_index(rbx);
850 __ movptr(aaddress(rbx), rax);
851 }
854 void TemplateTable::wide_istore() {
855 transition(vtos, vtos);
856 __ pop_i(rax);
857 locals_index_wide(rbx);
858 __ movl(iaddress(rbx), rax);
859 }
862 void TemplateTable::wide_lstore() {
863 transition(vtos, vtos);
864 __ pop_l(rax, rdx);
865 locals_index_wide(rbx);
866 __ movptr(laddress(rbx), rax);
867 NOT_LP64(__ movl(haddress(rbx), rdx));
868 }
871 void TemplateTable::wide_fstore() {
872 wide_istore();
873 }
876 void TemplateTable::wide_dstore() {
877 wide_lstore();
878 }
881 void TemplateTable::wide_astore() {
882 transition(vtos, vtos);
883 __ pop_ptr(rax);
884 locals_index_wide(rbx);
885 __ movptr(aaddress(rbx), rax);
886 }
889 void TemplateTable::iastore() {
890 transition(itos, vtos);
891 __ pop_i(rbx);
892 // rax,: value
893 // rdx: array
894 index_check(rdx, rbx); // prefer index in rbx,
895 // rbx,: index
896 __ movl(Address(rdx, rbx, Address::times_4, arrayOopDesc::base_offset_in_bytes(T_INT)), rax);
897 }
900 void TemplateTable::lastore() {
901 transition(ltos, vtos);
902 __ pop_i(rbx);
903 // rax,: low(value)
904 // rcx: array
905 // rdx: high(value)
906 index_check(rcx, rbx); // prefer index in rbx,
907 // rbx,: index
908 __ movptr(Address(rcx, rbx, Address::times_8, arrayOopDesc::base_offset_in_bytes(T_LONG) + 0 * wordSize), rax);
909 NOT_LP64(__ movl(Address(rcx, rbx, Address::times_8, arrayOopDesc::base_offset_in_bytes(T_LONG) + 1 * wordSize), rdx));
910 }
913 void TemplateTable::fastore() {
914 transition(ftos, vtos);
915 __ pop_i(rbx);
916 // rdx: array
917 // st0: value
918 index_check(rdx, rbx); // prefer index in rbx,
919 // rbx,: index
920 __ fstp_s(Address(rdx, rbx, Address::times_4, arrayOopDesc::base_offset_in_bytes(T_FLOAT)));
921 }
924 void TemplateTable::dastore() {
925 transition(dtos, vtos);
926 __ pop_i(rbx);
927 // rdx: array
928 // st0: value
929 index_check(rdx, rbx); // prefer index in rbx,
930 // rbx,: index
931 __ fstp_d(Address(rdx, rbx, Address::times_8, arrayOopDesc::base_offset_in_bytes(T_DOUBLE)));
932 }
935 void TemplateTable::aastore() {
936 Label is_null, ok_is_subtype, done;
937 transition(vtos, vtos);
938 // stack: ..., array, index, value
939 __ movptr(rax, at_tos()); // Value
940 __ movl(rcx, at_tos_p1()); // Index
941 __ movptr(rdx, at_tos_p2()); // Array
943 Address element_address(rdx, rcx, Address::times_4, arrayOopDesc::base_offset_in_bytes(T_OBJECT));
944 index_check_without_pop(rdx, rcx); // kills rbx,
945 // do array store check - check for NULL value first
946 __ testptr(rax, rax);
947 __ jcc(Assembler::zero, is_null);
949 // Move subklass into EBX
950 __ load_klass(rbx, rax);
951 // Move superklass into EAX
952 __ load_klass(rax, rdx);
953 __ movptr(rax, Address(rax, ObjArrayKlass::element_klass_offset()));
954 // Compress array+index*wordSize+12 into a single register. Frees ECX.
955 __ lea(rdx, element_address);
957 // Generate subtype check. Blows ECX. Resets EDI to locals.
958 // Superklass in EAX. Subklass in EBX.
959 __ gen_subtype_check( rbx, ok_is_subtype );
961 // Come here on failure
962 // object is at TOS
963 __ jump(ExternalAddress(Interpreter::_throw_ArrayStoreException_entry));
965 // Come here on success
966 __ bind(ok_is_subtype);
968 // Get the value to store
969 __ movptr(rax, at_rsp());
970 // and store it with appropriate barrier
971 do_oop_store(_masm, Address(rdx, 0), rax, _bs->kind(), true);
973 __ jmp(done);
975 // Have a NULL in EAX, EDX=array, ECX=index. Store NULL at ary[idx]
976 __ bind(is_null);
977 __ profile_null_seen(rbx);
979 // Store NULL, (noreg means NULL to do_oop_store)
980 do_oop_store(_masm, element_address, noreg, _bs->kind(), true);
982 // Pop stack arguments
983 __ bind(done);
984 __ addptr(rsp, 3 * Interpreter::stackElementSize);
985 }
988 void TemplateTable::bastore() {
989 transition(itos, vtos);
990 __ pop_i(rbx);
991 // rax,: value
992 // rdx: array
993 index_check(rdx, rbx); // prefer index in rbx,
994 // rbx,: index
995 __ movb(Address(rdx, rbx, Address::times_1, arrayOopDesc::base_offset_in_bytes(T_BYTE)), rax);
996 }
999 void TemplateTable::castore() {
1000 transition(itos, vtos);
1001 __ pop_i(rbx);
1002 // rax,: value
1003 // rdx: array
1004 index_check(rdx, rbx); // prefer index in rbx,
1005 // rbx,: index
1006 __ movw(Address(rdx, rbx, Address::times_2, arrayOopDesc::base_offset_in_bytes(T_CHAR)), rax);
1007 }
1010 void TemplateTable::sastore() {
1011 castore();
1012 }
1015 void TemplateTable::istore(int n) {
1016 transition(itos, vtos);
1017 __ movl(iaddress(n), rax);
1018 }
1021 void TemplateTable::lstore(int n) {
1022 transition(ltos, vtos);
1023 __ movptr(laddress(n), rax);
1024 NOT_LP64(__ movptr(haddress(n), rdx));
1025 }
1028 void TemplateTable::fstore(int n) {
1029 transition(ftos, vtos);
1030 __ fstp_s(faddress(n));
1031 }
1034 void TemplateTable::dstore(int n) {
1035 transition(dtos, vtos);
1036 __ fstp_d(daddress(n));
1037 }
1040 void TemplateTable::astore(int n) {
1041 transition(vtos, vtos);
1042 __ pop_ptr(rax);
1043 __ movptr(aaddress(n), rax);
1044 }
1047 void TemplateTable::pop() {
1048 transition(vtos, vtos);
1049 __ addptr(rsp, Interpreter::stackElementSize);
1050 }
1053 void TemplateTable::pop2() {
1054 transition(vtos, vtos);
1055 __ addptr(rsp, 2*Interpreter::stackElementSize);
1056 }
1059 void TemplateTable::dup() {
1060 transition(vtos, vtos);
1061 // stack: ..., a
1062 __ load_ptr(0, rax);
1063 __ push_ptr(rax);
1064 // stack: ..., a, a
1065 }
1068 void TemplateTable::dup_x1() {
1069 transition(vtos, vtos);
1070 // stack: ..., a, b
1071 __ load_ptr( 0, rax); // load b
1072 __ load_ptr( 1, rcx); // load a
1073 __ store_ptr(1, rax); // store b
1074 __ store_ptr(0, rcx); // store a
1075 __ push_ptr(rax); // push b
1076 // stack: ..., b, a, b
1077 }
1080 void TemplateTable::dup_x2() {
1081 transition(vtos, vtos);
1082 // stack: ..., a, b, c
1083 __ load_ptr( 0, rax); // load c
1084 __ load_ptr( 2, rcx); // load a
1085 __ store_ptr(2, rax); // store c in a
1086 __ push_ptr(rax); // push c
1087 // stack: ..., c, b, c, c
1088 __ load_ptr( 2, rax); // load b
1089 __ store_ptr(2, rcx); // store a in b
1090 // stack: ..., c, a, c, c
1091 __ store_ptr(1, rax); // store b in c
1092 // stack: ..., c, a, b, c
1093 }
1096 void TemplateTable::dup2() {
1097 transition(vtos, vtos);
1098 // stack: ..., a, b
1099 __ load_ptr(1, rax); // load a
1100 __ push_ptr(rax); // push a
1101 __ load_ptr(1, rax); // load b
1102 __ push_ptr(rax); // push b
1103 // stack: ..., a, b, a, b
1104 }
1107 void TemplateTable::dup2_x1() {
1108 transition(vtos, vtos);
1109 // stack: ..., a, b, c
1110 __ load_ptr( 0, rcx); // load c
1111 __ load_ptr( 1, rax); // load b
1112 __ push_ptr(rax); // push b
1113 __ push_ptr(rcx); // push c
1114 // stack: ..., a, b, c, b, c
1115 __ store_ptr(3, rcx); // store c in b
1116 // stack: ..., a, c, c, b, c
1117 __ load_ptr( 4, rcx); // load a
1118 __ store_ptr(2, rcx); // store a in 2nd c
1119 // stack: ..., a, c, a, b, c
1120 __ store_ptr(4, rax); // store b in a
1121 // stack: ..., b, c, a, b, c
1122 // stack: ..., b, c, a, b, c
1123 }
1126 void TemplateTable::dup2_x2() {
1127 transition(vtos, vtos);
1128 // stack: ..., a, b, c, d
1129 __ load_ptr( 0, rcx); // load d
1130 __ load_ptr( 1, rax); // load c
1131 __ push_ptr(rax); // push c
1132 __ push_ptr(rcx); // push d
1133 // stack: ..., a, b, c, d, c, d
1134 __ load_ptr( 4, rax); // load b
1135 __ store_ptr(2, rax); // store b in d
1136 __ store_ptr(4, rcx); // store d in b
1137 // stack: ..., a, d, c, b, c, d
1138 __ load_ptr( 5, rcx); // load a
1139 __ load_ptr( 3, rax); // load c
1140 __ store_ptr(3, rcx); // store a in c
1141 __ store_ptr(5, rax); // store c in a
1142 // stack: ..., c, d, a, b, c, d
1143 // stack: ..., c, d, a, b, c, d
1144 }
1147 void TemplateTable::swap() {
1148 transition(vtos, vtos);
1149 // stack: ..., a, b
1150 __ load_ptr( 1, rcx); // load a
1151 __ load_ptr( 0, rax); // load b
1152 __ store_ptr(0, rcx); // store a in b
1153 __ store_ptr(1, rax); // store b in a
1154 // stack: ..., b, a
1155 }
1158 void TemplateTable::iop2(Operation op) {
1159 transition(itos, itos);
1160 switch (op) {
1161 case add : __ pop_i(rdx); __ addl (rax, rdx); break;
1162 case sub : __ mov(rdx, rax); __ pop_i(rax); __ subl (rax, rdx); break;
1163 case mul : __ pop_i(rdx); __ imull(rax, rdx); break;
1164 case _and : __ pop_i(rdx); __ andl (rax, rdx); break;
1165 case _or : __ pop_i(rdx); __ orl (rax, rdx); break;
1166 case _xor : __ pop_i(rdx); __ xorl (rax, rdx); break;
1167 case shl : __ mov(rcx, rax); __ pop_i(rax); __ shll (rax); break; // implicit masking of lower 5 bits by Intel shift instr.
1168 case shr : __ mov(rcx, rax); __ pop_i(rax); __ sarl (rax); break; // implicit masking of lower 5 bits by Intel shift instr.
1169 case ushr : __ mov(rcx, rax); __ pop_i(rax); __ shrl (rax); break; // implicit masking of lower 5 bits by Intel shift instr.
1170 default : ShouldNotReachHere();
1171 }
1172 }
1175 void TemplateTable::lop2(Operation op) {
1176 transition(ltos, ltos);
1177 __ pop_l(rbx, rcx);
1178 switch (op) {
1179 case add : __ addl(rax, rbx); __ adcl(rdx, rcx); break;
1180 case sub : __ subl(rbx, rax); __ sbbl(rcx, rdx);
1181 __ mov (rax, rbx); __ mov (rdx, rcx); break;
1182 case _and : __ andl(rax, rbx); __ andl(rdx, rcx); break;
1183 case _or : __ orl (rax, rbx); __ orl (rdx, rcx); break;
1184 case _xor : __ xorl(rax, rbx); __ xorl(rdx, rcx); break;
1185 default : ShouldNotReachHere();
1186 }
1187 }
1190 void TemplateTable::idiv() {
1191 transition(itos, itos);
1192 __ mov(rcx, rax);
1193 __ pop_i(rax);
1194 // Note: could xor rax, and rcx and compare with (-1 ^ min_int). If
1195 // they are not equal, one could do a normal division (no correction
1196 // needed), which may speed up this implementation for the common case.
1197 // (see also JVM spec., p.243 & p.271)
1198 __ corrected_idivl(rcx);
1199 }
1202 void TemplateTable::irem() {
1203 transition(itos, itos);
1204 __ mov(rcx, rax);
1205 __ pop_i(rax);
1206 // Note: could xor rax, and rcx and compare with (-1 ^ min_int). If
1207 // they are not equal, one could do a normal division (no correction
1208 // needed), which may speed up this implementation for the common case.
1209 // (see also JVM spec., p.243 & p.271)
1210 __ corrected_idivl(rcx);
1211 __ mov(rax, rdx);
1212 }
1215 void TemplateTable::lmul() {
1216 transition(ltos, ltos);
1217 __ pop_l(rbx, rcx);
1218 __ push(rcx); __ push(rbx);
1219 __ push(rdx); __ push(rax);
1220 __ lmul(2 * wordSize, 0);
1221 __ addptr(rsp, 4 * wordSize); // take off temporaries
1222 }
1225 void TemplateTable::ldiv() {
1226 transition(ltos, ltos);
1227 __ pop_l(rbx, rcx);
1228 __ push(rcx); __ push(rbx);
1229 __ push(rdx); __ push(rax);
1230 // check if y = 0
1231 __ orl(rax, rdx);
1232 __ jump_cc(Assembler::zero,
1233 ExternalAddress(Interpreter::_throw_ArithmeticException_entry));
1234 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::ldiv));
1235 __ addptr(rsp, 4 * wordSize); // take off temporaries
1236 }
1239 void TemplateTable::lrem() {
1240 transition(ltos, ltos);
1241 __ pop_l(rbx, rcx);
1242 __ push(rcx); __ push(rbx);
1243 __ push(rdx); __ push(rax);
1244 // check if y = 0
1245 __ orl(rax, rdx);
1246 __ jump_cc(Assembler::zero,
1247 ExternalAddress(Interpreter::_throw_ArithmeticException_entry));
1248 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::lrem));
1249 __ addptr(rsp, 4 * wordSize);
1250 }
1253 void TemplateTable::lshl() {
1254 transition(itos, ltos);
1255 __ movl(rcx, rax); // get shift count
1256 __ pop_l(rax, rdx); // get shift value
1257 __ lshl(rdx, rax);
1258 }
1261 void TemplateTable::lshr() {
1262 transition(itos, ltos);
1263 __ mov(rcx, rax); // get shift count
1264 __ pop_l(rax, rdx); // get shift value
1265 __ lshr(rdx, rax, true);
1266 }
1269 void TemplateTable::lushr() {
1270 transition(itos, ltos);
1271 __ mov(rcx, rax); // get shift count
1272 __ pop_l(rax, rdx); // get shift value
1273 __ lshr(rdx, rax);
1274 }
1277 void TemplateTable::fop2(Operation op) {
1278 transition(ftos, ftos);
1279 switch (op) {
1280 case add: __ fadd_s (at_rsp()); break;
1281 case sub: __ fsubr_s(at_rsp()); break;
1282 case mul: __ fmul_s (at_rsp()); break;
1283 case div: __ fdivr_s(at_rsp()); break;
1284 case rem: __ fld_s (at_rsp()); __ fremr(rax); break;
1285 default : ShouldNotReachHere();
1286 }
1287 __ f2ieee();
1288 __ pop(rax); // pop float thing off
1289 }
1292 void TemplateTable::dop2(Operation op) {
1293 transition(dtos, dtos);
1295 switch (op) {
1296 case add: __ fadd_d (at_rsp()); break;
1297 case sub: __ fsubr_d(at_rsp()); break;
1298 case mul: {
1299 Label L_strict;
1300 Label L_join;
1301 const Address access_flags (rcx, Method::access_flags_offset());
1302 __ get_method(rcx);
1303 __ movl(rcx, access_flags);
1304 __ testl(rcx, JVM_ACC_STRICT);
1305 __ jccb(Assembler::notZero, L_strict);
1306 __ fmul_d (at_rsp());
1307 __ jmpb(L_join);
1308 __ bind(L_strict);
1309 __ fld_x(ExternalAddress(StubRoutines::addr_fpu_subnormal_bias1()));
1310 __ fmulp();
1311 __ fmul_d (at_rsp());
1312 __ fld_x(ExternalAddress(StubRoutines::addr_fpu_subnormal_bias2()));
1313 __ fmulp();
1314 __ bind(L_join);
1315 break;
1316 }
1317 case div: {
1318 Label L_strict;
1319 Label L_join;
1320 const Address access_flags (rcx, Method::access_flags_offset());
1321 __ get_method(rcx);
1322 __ movl(rcx, access_flags);
1323 __ testl(rcx, JVM_ACC_STRICT);
1324 __ jccb(Assembler::notZero, L_strict);
1325 __ fdivr_d(at_rsp());
1326 __ jmp(L_join);
1327 __ bind(L_strict);
1328 __ fld_x(ExternalAddress(StubRoutines::addr_fpu_subnormal_bias1()));
1329 __ fmul_d (at_rsp());
1330 __ fdivrp();
1331 __ fld_x(ExternalAddress(StubRoutines::addr_fpu_subnormal_bias2()));
1332 __ fmulp();
1333 __ bind(L_join);
1334 break;
1335 }
1336 case rem: __ fld_d (at_rsp()); __ fremr(rax); break;
1337 default : ShouldNotReachHere();
1338 }
1339 __ d2ieee();
1340 // Pop double precision number from rsp.
1341 __ pop(rax);
1342 __ pop(rdx);
1343 }
1346 void TemplateTable::ineg() {
1347 transition(itos, itos);
1348 __ negl(rax);
1349 }
1352 void TemplateTable::lneg() {
1353 transition(ltos, ltos);
1354 __ lneg(rdx, rax);
1355 }
1358 void TemplateTable::fneg() {
1359 transition(ftos, ftos);
1360 __ fchs();
1361 }
1364 void TemplateTable::dneg() {
1365 transition(dtos, dtos);
1366 __ fchs();
1367 }
1370 void TemplateTable::iinc() {
1371 transition(vtos, vtos);
1372 __ load_signed_byte(rdx, at_bcp(2)); // get constant
1373 locals_index(rbx);
1374 __ addl(iaddress(rbx), rdx);
1375 }
1378 void TemplateTable::wide_iinc() {
1379 transition(vtos, vtos);
1380 __ movl(rdx, at_bcp(4)); // get constant
1381 locals_index_wide(rbx);
1382 __ bswapl(rdx); // swap bytes & sign-extend constant
1383 __ sarl(rdx, 16);
1384 __ addl(iaddress(rbx), rdx);
1385 // Note: should probably use only one movl to get both
1386 // the index and the constant -> fix this
1387 }
1390 void TemplateTable::convert() {
1391 // Checking
1392 #ifdef ASSERT
1393 { TosState tos_in = ilgl;
1394 TosState tos_out = ilgl;
1395 switch (bytecode()) {
1396 case Bytecodes::_i2l: // fall through
1397 case Bytecodes::_i2f: // fall through
1398 case Bytecodes::_i2d: // fall through
1399 case Bytecodes::_i2b: // fall through
1400 case Bytecodes::_i2c: // fall through
1401 case Bytecodes::_i2s: tos_in = itos; break;
1402 case Bytecodes::_l2i: // fall through
1403 case Bytecodes::_l2f: // fall through
1404 case Bytecodes::_l2d: tos_in = ltos; break;
1405 case Bytecodes::_f2i: // fall through
1406 case Bytecodes::_f2l: // fall through
1407 case Bytecodes::_f2d: tos_in = ftos; break;
1408 case Bytecodes::_d2i: // fall through
1409 case Bytecodes::_d2l: // fall through
1410 case Bytecodes::_d2f: tos_in = dtos; break;
1411 default : ShouldNotReachHere();
1412 }
1413 switch (bytecode()) {
1414 case Bytecodes::_l2i: // fall through
1415 case Bytecodes::_f2i: // fall through
1416 case Bytecodes::_d2i: // fall through
1417 case Bytecodes::_i2b: // fall through
1418 case Bytecodes::_i2c: // fall through
1419 case Bytecodes::_i2s: tos_out = itos; break;
1420 case Bytecodes::_i2l: // fall through
1421 case Bytecodes::_f2l: // fall through
1422 case Bytecodes::_d2l: tos_out = ltos; break;
1423 case Bytecodes::_i2f: // fall through
1424 case Bytecodes::_l2f: // fall through
1425 case Bytecodes::_d2f: tos_out = ftos; break;
1426 case Bytecodes::_i2d: // fall through
1427 case Bytecodes::_l2d: // fall through
1428 case Bytecodes::_f2d: tos_out = dtos; break;
1429 default : ShouldNotReachHere();
1430 }
1431 transition(tos_in, tos_out);
1432 }
1433 #endif // ASSERT
1435 // Conversion
1436 // (Note: use push(rcx)/pop(rcx) for 1/2-word stack-ptr manipulation)
1437 switch (bytecode()) {
1438 case Bytecodes::_i2l:
1439 __ extend_sign(rdx, rax);
1440 break;
1441 case Bytecodes::_i2f:
1442 __ push(rax); // store int on tos
1443 __ fild_s(at_rsp()); // load int to ST0
1444 __ f2ieee(); // truncate to float size
1445 __ pop(rcx); // adjust rsp
1446 break;
1447 case Bytecodes::_i2d:
1448 __ push(rax); // add one slot for d2ieee()
1449 __ push(rax); // store int on tos
1450 __ fild_s(at_rsp()); // load int to ST0
1451 __ d2ieee(); // truncate to double size
1452 __ pop(rcx); // adjust rsp
1453 __ pop(rcx);
1454 break;
1455 case Bytecodes::_i2b:
1456 __ shll(rax, 24); // truncate upper 24 bits
1457 __ sarl(rax, 24); // and sign-extend byte
1458 LP64_ONLY(__ movsbl(rax, rax));
1459 break;
1460 case Bytecodes::_i2c:
1461 __ andl(rax, 0xFFFF); // truncate upper 16 bits
1462 LP64_ONLY(__ movzwl(rax, rax));
1463 break;
1464 case Bytecodes::_i2s:
1465 __ shll(rax, 16); // truncate upper 16 bits
1466 __ sarl(rax, 16); // and sign-extend short
1467 LP64_ONLY(__ movswl(rax, rax));
1468 break;
1469 case Bytecodes::_l2i:
1470 /* nothing to do */
1471 break;
1472 case Bytecodes::_l2f:
1473 __ push(rdx); // store long on tos
1474 __ push(rax);
1475 __ fild_d(at_rsp()); // load long to ST0
1476 __ f2ieee(); // truncate to float size
1477 __ pop(rcx); // adjust rsp
1478 __ pop(rcx);
1479 break;
1480 case Bytecodes::_l2d:
1481 __ push(rdx); // store long on tos
1482 __ push(rax);
1483 __ fild_d(at_rsp()); // load long to ST0
1484 __ d2ieee(); // truncate to double size
1485 __ pop(rcx); // adjust rsp
1486 __ pop(rcx);
1487 break;
1488 case Bytecodes::_f2i:
1489 __ push(rcx); // reserve space for argument
1490 __ fstp_s(at_rsp()); // pass float argument on stack
1491 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::f2i), 1);
1492 break;
1493 case Bytecodes::_f2l:
1494 __ push(rcx); // reserve space for argument
1495 __ fstp_s(at_rsp()); // pass float argument on stack
1496 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::f2l), 1);
1497 break;
1498 case Bytecodes::_f2d:
1499 /* nothing to do */
1500 break;
1501 case Bytecodes::_d2i:
1502 __ push(rcx); // reserve space for argument
1503 __ push(rcx);
1504 __ fstp_d(at_rsp()); // pass double argument on stack
1505 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::d2i), 2);
1506 break;
1507 case Bytecodes::_d2l:
1508 __ push(rcx); // reserve space for argument
1509 __ push(rcx);
1510 __ fstp_d(at_rsp()); // pass double argument on stack
1511 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::d2l), 2);
1512 break;
1513 case Bytecodes::_d2f:
1514 __ push(rcx); // reserve space for f2ieee()
1515 __ f2ieee(); // truncate to float size
1516 __ pop(rcx); // adjust rsp
1517 break;
1518 default :
1519 ShouldNotReachHere();
1520 }
1521 }
1524 void TemplateTable::lcmp() {
1525 transition(ltos, itos);
1526 // y = rdx:rax
1527 __ pop_l(rbx, rcx); // get x = rcx:rbx
1528 __ lcmp2int(rcx, rbx, rdx, rax);// rcx := cmp(x, y)
1529 __ mov(rax, rcx);
1530 }
1533 void TemplateTable::float_cmp(bool is_float, int unordered_result) {
1534 if (is_float) {
1535 __ fld_s(at_rsp());
1536 } else {
1537 __ fld_d(at_rsp());
1538 __ pop(rdx);
1539 }
1540 __ pop(rcx);
1541 __ fcmp2int(rax, unordered_result < 0);
1542 }
1545 void TemplateTable::branch(bool is_jsr, bool is_wide) {
1546 __ get_method(rcx); // ECX holds method
1547 __ profile_taken_branch(rax,rbx); // EAX holds updated MDP, EBX holds bumped taken count
1549 const ByteSize be_offset = MethodCounters::backedge_counter_offset() +
1550 InvocationCounter::counter_offset();
1551 const ByteSize inv_offset = MethodCounters::invocation_counter_offset() +
1552 InvocationCounter::counter_offset();
1554 // Load up EDX with the branch displacement
1555 if (is_wide) {
1556 __ movl(rdx, at_bcp(1));
1557 } else {
1558 __ load_signed_short(rdx, at_bcp(1));
1559 }
1560 __ bswapl(rdx);
1561 if (!is_wide) __ sarl(rdx, 16);
1562 LP64_ONLY(__ movslq(rdx, rdx));
1565 // Handle all the JSR stuff here, then exit.
1566 // It's much shorter and cleaner than intermingling with the
1567 // non-JSR normal-branch stuff occurring below.
1568 if (is_jsr) {
1569 // Pre-load the next target bytecode into EBX
1570 __ load_unsigned_byte(rbx, Address(rsi, rdx, Address::times_1, 0));
1572 // compute return address as bci in rax,
1573 __ lea(rax, at_bcp((is_wide ? 5 : 3) - in_bytes(ConstMethod::codes_offset())));
1574 __ subptr(rax, Address(rcx, Method::const_offset()));
1575 // Adjust the bcp in RSI by the displacement in EDX
1576 __ addptr(rsi, rdx);
1577 // Push return address
1578 __ push_i(rax);
1579 // jsr returns vtos
1580 __ dispatch_only_noverify(vtos);
1581 return;
1582 }
1584 // Normal (non-jsr) branch handling
1586 // Adjust the bcp in RSI by the displacement in EDX
1587 __ addptr(rsi, rdx);
1589 assert(UseLoopCounter || !UseOnStackReplacement, "on-stack-replacement requires loop counters");
1590 Label backedge_counter_overflow;
1591 Label profile_method;
1592 Label dispatch;
1593 if (UseLoopCounter) {
1594 // increment backedge counter for backward branches
1595 // rax,: MDO
1596 // rbx,: MDO bumped taken-count
1597 // rcx: method
1598 // rdx: target offset
1599 // rsi: target bcp
1600 // rdi: locals pointer
1601 __ testl(rdx, rdx); // check if forward or backward branch
1602 __ jcc(Assembler::positive, dispatch); // count only if backward branch
1604 // check if MethodCounters exists
1605 Label has_counters;
1606 __ movptr(rax, Address(rcx, Method::method_counters_offset()));
1607 __ testptr(rax, rax);
1608 __ jcc(Assembler::notZero, has_counters);
1609 __ push(rdx);
1610 __ push(rcx);
1611 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::build_method_counters),
1612 rcx);
1613 __ pop(rcx);
1614 __ pop(rdx);
1615 __ movptr(rax, Address(rcx, Method::method_counters_offset()));
1616 __ testptr(rax, rax);
1617 __ jcc(Assembler::zero, dispatch);
1618 __ bind(has_counters);
1620 if (TieredCompilation) {
1621 Label no_mdo;
1622 int increment = InvocationCounter::count_increment;
1623 int mask = ((1 << Tier0BackedgeNotifyFreqLog) - 1) << InvocationCounter::count_shift;
1624 if (ProfileInterpreter) {
1625 // Are we profiling?
1626 __ movptr(rbx, Address(rcx, in_bytes(Method::method_data_offset())));
1627 __ testptr(rbx, rbx);
1628 __ jccb(Assembler::zero, no_mdo);
1629 // Increment the MDO backedge counter
1630 const Address mdo_backedge_counter(rbx, in_bytes(MethodData::backedge_counter_offset()) +
1631 in_bytes(InvocationCounter::counter_offset()));
1632 __ increment_mask_and_jump(mdo_backedge_counter, increment, mask,
1633 rax, false, Assembler::zero, &backedge_counter_overflow);
1634 __ jmp(dispatch);
1635 }
1636 __ bind(no_mdo);
1637 // Increment backedge counter in MethodCounters*
1638 __ movptr(rcx, Address(rcx, Method::method_counters_offset()));
1639 __ increment_mask_and_jump(Address(rcx, be_offset), increment, mask,
1640 rax, false, Assembler::zero, &backedge_counter_overflow);
1641 } else {
1642 // increment counter
1643 __ movptr(rcx, Address(rcx, Method::method_counters_offset()));
1644 __ movl(rax, Address(rcx, be_offset)); // load backedge counter
1645 __ incrementl(rax, InvocationCounter::count_increment); // increment counter
1646 __ movl(Address(rcx, be_offset), rax); // store counter
1648 __ movl(rax, Address(rcx, inv_offset)); // load invocation counter
1650 __ andl(rax, InvocationCounter::count_mask_value); // and the status bits
1651 __ addl(rax, Address(rcx, be_offset)); // add both counters
1653 if (ProfileInterpreter) {
1654 // Test to see if we should create a method data oop
1655 __ cmp32(rax,
1656 ExternalAddress((address) &InvocationCounter::InterpreterProfileLimit));
1657 __ jcc(Assembler::less, dispatch);
1659 // if no method data exists, go to profile method
1660 __ test_method_data_pointer(rax, profile_method);
1662 if (UseOnStackReplacement) {
1663 // check for overflow against rbx, which is the MDO taken count
1664 __ cmp32(rbx,
1665 ExternalAddress((address) &InvocationCounter::InterpreterBackwardBranchLimit));
1666 __ jcc(Assembler::below, dispatch);
1668 // When ProfileInterpreter is on, the backedge_count comes from the
1669 // MethodData*, which value does not get reset on the call to
1670 // frequency_counter_overflow(). To avoid excessive calls to the overflow
1671 // routine while the method is being compiled, add a second test to make
1672 // sure the overflow function is called only once every overflow_frequency.
1673 const int overflow_frequency = 1024;
1674 __ andptr(rbx, overflow_frequency-1);
1675 __ jcc(Assembler::zero, backedge_counter_overflow);
1676 }
1677 } else {
1678 if (UseOnStackReplacement) {
1679 // check for overflow against rax, which is the sum of the counters
1680 __ cmp32(rax,
1681 ExternalAddress((address) &InvocationCounter::InterpreterBackwardBranchLimit));
1682 __ jcc(Assembler::aboveEqual, backedge_counter_overflow);
1684 }
1685 }
1686 }
1687 __ bind(dispatch);
1688 }
1690 // Pre-load the next target bytecode into EBX
1691 __ load_unsigned_byte(rbx, Address(rsi, 0));
1693 // continue with the bytecode @ target
1694 // rax,: return bci for jsr's, unused otherwise
1695 // rbx,: target bytecode
1696 // rsi: target bcp
1697 __ dispatch_only(vtos);
1699 if (UseLoopCounter) {
1700 if (ProfileInterpreter) {
1701 // Out-of-line code to allocate method data oop.
1702 __ bind(profile_method);
1703 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::profile_method));
1704 __ load_unsigned_byte(rbx, Address(rsi, 0)); // restore target bytecode
1705 __ set_method_data_pointer_for_bcp();
1706 __ jmp(dispatch);
1707 }
1709 if (UseOnStackReplacement) {
1711 // invocation counter overflow
1712 __ bind(backedge_counter_overflow);
1713 __ negptr(rdx);
1714 __ addptr(rdx, rsi); // branch bcp
1715 call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::frequency_counter_overflow), rdx);
1716 __ load_unsigned_byte(rbx, Address(rsi, 0)); // restore target bytecode
1718 // rax,: osr nmethod (osr ok) or NULL (osr not possible)
1719 // rbx,: target bytecode
1720 // rdx: scratch
1721 // rdi: locals pointer
1722 // rsi: bcp
1723 __ testptr(rax, rax); // test result
1724 __ jcc(Assembler::zero, dispatch); // no osr if null
1725 // nmethod may have been invalidated (VM may block upon call_VM return)
1726 __ movl(rcx, Address(rax, nmethod::entry_bci_offset()));
1727 __ cmpl(rcx, InvalidOSREntryBci);
1728 __ jcc(Assembler::equal, dispatch);
1730 // We have the address of an on stack replacement routine in rax,
1731 // We need to prepare to execute the OSR method. First we must
1732 // migrate the locals and monitors off of the stack.
1734 __ mov(rbx, rax); // save the nmethod
1736 const Register thread = rcx;
1737 __ get_thread(thread);
1738 call_VM(noreg, CAST_FROM_FN_PTR(address, SharedRuntime::OSR_migration_begin));
1739 // rax, is OSR buffer, move it to expected parameter location
1740 __ mov(rcx, rax);
1742 // pop the interpreter frame
1743 __ movptr(rdx, Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize)); // get sender sp
1744 __ leave(); // remove frame anchor
1745 __ pop(rdi); // get return address
1746 __ mov(rsp, rdx); // set sp to sender sp
1748 // Align stack pointer for compiled code (note that caller is
1749 // responsible for undoing this fixup by remembering the old SP
1750 // in an rbp,-relative location)
1751 __ andptr(rsp, -(StackAlignmentInBytes));
1753 // push the (possibly adjusted) return address
1754 __ push(rdi);
1756 // and begin the OSR nmethod
1757 __ jmp(Address(rbx, nmethod::osr_entry_point_offset()));
1758 }
1759 }
1760 }
1763 void TemplateTable::if_0cmp(Condition cc) {
1764 transition(itos, vtos);
1765 // assume branch is more often taken than not (loops use backward branches)
1766 Label not_taken;
1767 __ testl(rax, rax);
1768 __ jcc(j_not(cc), not_taken);
1769 branch(false, false);
1770 __ bind(not_taken);
1771 __ profile_not_taken_branch(rax);
1772 }
1775 void TemplateTable::if_icmp(Condition cc) {
1776 transition(itos, vtos);
1777 // assume branch is more often taken than not (loops use backward branches)
1778 Label not_taken;
1779 __ pop_i(rdx);
1780 __ cmpl(rdx, rax);
1781 __ jcc(j_not(cc), not_taken);
1782 branch(false, false);
1783 __ bind(not_taken);
1784 __ profile_not_taken_branch(rax);
1785 }
1788 void TemplateTable::if_nullcmp(Condition cc) {
1789 transition(atos, vtos);
1790 // assume branch is more often taken than not (loops use backward branches)
1791 Label not_taken;
1792 __ testptr(rax, rax);
1793 __ jcc(j_not(cc), not_taken);
1794 branch(false, false);
1795 __ bind(not_taken);
1796 __ profile_not_taken_branch(rax);
1797 }
1800 void TemplateTable::if_acmp(Condition cc) {
1801 transition(atos, vtos);
1802 // assume branch is more often taken than not (loops use backward branches)
1803 Label not_taken;
1804 __ pop_ptr(rdx);
1805 __ cmpptr(rdx, rax);
1806 __ jcc(j_not(cc), not_taken);
1807 branch(false, false);
1808 __ bind(not_taken);
1809 __ profile_not_taken_branch(rax);
1810 }
1813 void TemplateTable::ret() {
1814 transition(vtos, vtos);
1815 locals_index(rbx);
1816 __ movptr(rbx, iaddress(rbx)); // get return bci, compute return bcp
1817 __ profile_ret(rbx, rcx);
1818 __ get_method(rax);
1819 __ movptr(rsi, Address(rax, Method::const_offset()));
1820 __ lea(rsi, Address(rsi, rbx, Address::times_1,
1821 ConstMethod::codes_offset()));
1822 __ dispatch_next(vtos);
1823 }
1826 void TemplateTable::wide_ret() {
1827 transition(vtos, vtos);
1828 locals_index_wide(rbx);
1829 __ movptr(rbx, iaddress(rbx)); // get return bci, compute return bcp
1830 __ profile_ret(rbx, rcx);
1831 __ get_method(rax);
1832 __ movptr(rsi, Address(rax, Method::const_offset()));
1833 __ lea(rsi, Address(rsi, rbx, Address::times_1, ConstMethod::codes_offset()));
1834 __ dispatch_next(vtos);
1835 }
1838 void TemplateTable::tableswitch() {
1839 Label default_case, continue_execution;
1840 transition(itos, vtos);
1841 // align rsi
1842 __ lea(rbx, at_bcp(wordSize));
1843 __ andptr(rbx, -wordSize);
1844 // load lo & hi
1845 __ movl(rcx, Address(rbx, 1 * wordSize));
1846 __ movl(rdx, Address(rbx, 2 * wordSize));
1847 __ bswapl(rcx);
1848 __ bswapl(rdx);
1849 // check against lo & hi
1850 __ cmpl(rax, rcx);
1851 __ jccb(Assembler::less, default_case);
1852 __ cmpl(rax, rdx);
1853 __ jccb(Assembler::greater, default_case);
1854 // lookup dispatch offset
1855 __ subl(rax, rcx);
1856 __ movl(rdx, Address(rbx, rax, Address::times_4, 3 * BytesPerInt));
1857 __ profile_switch_case(rax, rbx, rcx);
1858 // continue execution
1859 __ bind(continue_execution);
1860 __ bswapl(rdx);
1861 __ load_unsigned_byte(rbx, Address(rsi, rdx, Address::times_1));
1862 __ addptr(rsi, rdx);
1863 __ dispatch_only(vtos);
1864 // handle default
1865 __ bind(default_case);
1866 __ profile_switch_default(rax);
1867 __ movl(rdx, Address(rbx, 0));
1868 __ jmp(continue_execution);
1869 }
1872 void TemplateTable::lookupswitch() {
1873 transition(itos, itos);
1874 __ stop("lookupswitch bytecode should have been rewritten");
1875 }
1878 void TemplateTable::fast_linearswitch() {
1879 transition(itos, vtos);
1880 Label loop_entry, loop, found, continue_execution;
1881 // bswapl rax, so we can avoid bswapping the table entries
1882 __ bswapl(rax);
1883 // align rsi
1884 __ lea(rbx, at_bcp(wordSize)); // btw: should be able to get rid of this instruction (change offsets below)
1885 __ andptr(rbx, -wordSize);
1886 // set counter
1887 __ movl(rcx, Address(rbx, wordSize));
1888 __ bswapl(rcx);
1889 __ jmpb(loop_entry);
1890 // table search
1891 __ bind(loop);
1892 __ cmpl(rax, Address(rbx, rcx, Address::times_8, 2 * wordSize));
1893 __ jccb(Assembler::equal, found);
1894 __ bind(loop_entry);
1895 __ decrementl(rcx);
1896 __ jcc(Assembler::greaterEqual, loop);
1897 // default case
1898 __ profile_switch_default(rax);
1899 __ movl(rdx, Address(rbx, 0));
1900 __ jmpb(continue_execution);
1901 // entry found -> get offset
1902 __ bind(found);
1903 __ movl(rdx, Address(rbx, rcx, Address::times_8, 3 * wordSize));
1904 __ profile_switch_case(rcx, rax, rbx);
1905 // continue execution
1906 __ bind(continue_execution);
1907 __ bswapl(rdx);
1908 __ load_unsigned_byte(rbx, Address(rsi, rdx, Address::times_1));
1909 __ addptr(rsi, rdx);
1910 __ dispatch_only(vtos);
1911 }
1914 void TemplateTable::fast_binaryswitch() {
1915 transition(itos, vtos);
1916 // Implementation using the following core algorithm:
1917 //
1918 // int binary_search(int key, LookupswitchPair* array, int n) {
1919 // // Binary search according to "Methodik des Programmierens" by
1920 // // Edsger W. Dijkstra and W.H.J. Feijen, Addison Wesley Germany 1985.
1921 // int i = 0;
1922 // int j = n;
1923 // while (i+1 < j) {
1924 // // invariant P: 0 <= i < j <= n and (a[i] <= key < a[j] or Q)
1925 // // with Q: for all i: 0 <= i < n: key < a[i]
1926 // // where a stands for the array and assuming that the (inexisting)
1927 // // element a[n] is infinitely big.
1928 // int h = (i + j) >> 1;
1929 // // i < h < j
1930 // if (key < array[h].fast_match()) {
1931 // j = h;
1932 // } else {
1933 // i = h;
1934 // }
1935 // }
1936 // // R: a[i] <= key < a[i+1] or Q
1937 // // (i.e., if key is within array, i is the correct index)
1938 // return i;
1939 // }
1941 // register allocation
1942 const Register key = rax; // already set (tosca)
1943 const Register array = rbx;
1944 const Register i = rcx;
1945 const Register j = rdx;
1946 const Register h = rdi; // needs to be restored
1947 const Register temp = rsi;
1948 // setup array
1949 __ save_bcp();
1951 __ lea(array, at_bcp(3*wordSize)); // btw: should be able to get rid of this instruction (change offsets below)
1952 __ andptr(array, -wordSize);
1953 // initialize i & j
1954 __ xorl(i, i); // i = 0;
1955 __ movl(j, Address(array, -wordSize)); // j = length(array);
1956 // Convert j into native byteordering
1957 __ bswapl(j);
1958 // and start
1959 Label entry;
1960 __ jmp(entry);
1962 // binary search loop
1963 { Label loop;
1964 __ bind(loop);
1965 // int h = (i + j) >> 1;
1966 __ leal(h, Address(i, j, Address::times_1)); // h = i + j;
1967 __ sarl(h, 1); // h = (i + j) >> 1;
1968 // if (key < array[h].fast_match()) {
1969 // j = h;
1970 // } else {
1971 // i = h;
1972 // }
1973 // Convert array[h].match to native byte-ordering before compare
1974 __ movl(temp, Address(array, h, Address::times_8, 0*wordSize));
1975 __ bswapl(temp);
1976 __ cmpl(key, temp);
1977 // j = h if (key < array[h].fast_match())
1978 __ cmov32(Assembler::less , j, h);
1979 // i = h if (key >= array[h].fast_match())
1980 __ cmov32(Assembler::greaterEqual, i, h);
1981 // while (i+1 < j)
1982 __ bind(entry);
1983 __ leal(h, Address(i, 1)); // i+1
1984 __ cmpl(h, j); // i+1 < j
1985 __ jcc(Assembler::less, loop);
1986 }
1988 // end of binary search, result index is i (must check again!)
1989 Label default_case;
1990 // Convert array[i].match to native byte-ordering before compare
1991 __ movl(temp, Address(array, i, Address::times_8, 0*wordSize));
1992 __ bswapl(temp);
1993 __ cmpl(key, temp);
1994 __ jcc(Assembler::notEqual, default_case);
1996 // entry found -> j = offset
1997 __ movl(j , Address(array, i, Address::times_8, 1*wordSize));
1998 __ profile_switch_case(i, key, array);
1999 __ bswapl(j);
2000 LP64_ONLY(__ movslq(j, j));
2001 __ restore_bcp();
2002 __ restore_locals(); // restore rdi
2003 __ load_unsigned_byte(rbx, Address(rsi, j, Address::times_1));
2005 __ addptr(rsi, j);
2006 __ dispatch_only(vtos);
2008 // default case -> j = default offset
2009 __ bind(default_case);
2010 __ profile_switch_default(i);
2011 __ movl(j, Address(array, -2*wordSize));
2012 __ bswapl(j);
2013 LP64_ONLY(__ movslq(j, j));
2014 __ restore_bcp();
2015 __ restore_locals(); // restore rdi
2016 __ load_unsigned_byte(rbx, Address(rsi, j, Address::times_1));
2017 __ addptr(rsi, j);
2018 __ dispatch_only(vtos);
2019 }
2022 void TemplateTable::_return(TosState state) {
2023 transition(state, state);
2024 assert(_desc->calls_vm(), "inconsistent calls_vm information"); // call in remove_activation
2026 if (_desc->bytecode() == Bytecodes::_return_register_finalizer) {
2027 assert(state == vtos, "only valid state");
2028 __ movptr(rax, aaddress(0));
2029 __ load_klass(rdi, rax);
2030 __ movl(rdi, Address(rdi, Klass::access_flags_offset()));
2031 __ testl(rdi, JVM_ACC_HAS_FINALIZER);
2032 Label skip_register_finalizer;
2033 __ jcc(Assembler::zero, skip_register_finalizer);
2035 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::register_finalizer), rax);
2037 __ bind(skip_register_finalizer);
2038 }
2040 __ remove_activation(state, rsi);
2041 __ jmp(rsi);
2042 }
2045 // ----------------------------------------------------------------------------
2046 // Volatile variables demand their effects be made known to all CPU's in
2047 // order. Store buffers on most chips allow reads & writes to reorder; the
2048 // JMM's ReadAfterWrite.java test fails in -Xint mode without some kind of
2049 // memory barrier (i.e., it's not sufficient that the interpreter does not
2050 // reorder volatile references, the hardware also must not reorder them).
2051 //
2052 // According to the new Java Memory Model (JMM):
2053 // (1) All volatiles are serialized wrt to each other.
2054 // ALSO reads & writes act as aquire & release, so:
2055 // (2) A read cannot let unrelated NON-volatile memory refs that happen after
2056 // the read float up to before the read. It's OK for non-volatile memory refs
2057 // that happen before the volatile read to float down below it.
2058 // (3) Similar a volatile write cannot let unrelated NON-volatile memory refs
2059 // that happen BEFORE the write float down to after the write. It's OK for
2060 // non-volatile memory refs that happen after the volatile write to float up
2061 // before it.
2062 //
2063 // We only put in barriers around volatile refs (they are expensive), not
2064 // _between_ memory refs (that would require us to track the flavor of the
2065 // previous memory refs). Requirements (2) and (3) require some barriers
2066 // before volatile stores and after volatile loads. These nearly cover
2067 // requirement (1) but miss the volatile-store-volatile-load case. This final
2068 // case is placed after volatile-stores although it could just as well go
2069 // before volatile-loads.
2070 void TemplateTable::volatile_barrier(Assembler::Membar_mask_bits order_constraint ) {
2071 // Helper function to insert a is-volatile test and memory barrier
2072 if( !os::is_MP() ) return; // Not needed on single CPU
2073 __ membar(order_constraint);
2074 }
2076 void TemplateTable::resolve_cache_and_index(int byte_no,
2077 Register Rcache,
2078 Register index,
2079 size_t index_size) {
2080 const Register temp = rbx;
2081 assert_different_registers(Rcache, index, temp);
2083 Label resolved;
2084 assert(byte_no == f1_byte || byte_no == f2_byte, "byte_no out of range");
2085 __ get_cache_and_index_and_bytecode_at_bcp(Rcache, index, temp, byte_no, 1, index_size);
2086 __ cmpl(temp, (int) bytecode()); // have we resolved this bytecode?
2087 __ jcc(Assembler::equal, resolved);
2089 // resolve first time through
2090 address entry;
2091 switch (bytecode()) {
2092 case Bytecodes::_getstatic : // fall through
2093 case Bytecodes::_putstatic : // fall through
2094 case Bytecodes::_getfield : // fall through
2095 case Bytecodes::_putfield : entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_get_put); break;
2096 case Bytecodes::_invokevirtual : // fall through
2097 case Bytecodes::_invokespecial : // fall through
2098 case Bytecodes::_invokestatic : // fall through
2099 case Bytecodes::_invokeinterface: entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_invoke); break;
2100 case Bytecodes::_invokehandle : entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_invokehandle); break;
2101 case Bytecodes::_invokedynamic : entry = CAST_FROM_FN_PTR(address, InterpreterRuntime::resolve_invokedynamic); break;
2102 default:
2103 fatal(err_msg("unexpected bytecode: %s", Bytecodes::name(bytecode())));
2104 break;
2105 }
2106 __ movl(temp, (int)bytecode());
2107 __ call_VM(noreg, entry, temp);
2108 // Update registers with resolved info
2109 __ get_cache_and_index_at_bcp(Rcache, index, 1, index_size);
2110 __ bind(resolved);
2111 }
2114 // The cache and index registers must be set before call
2115 void TemplateTable::load_field_cp_cache_entry(Register obj,
2116 Register cache,
2117 Register index,
2118 Register off,
2119 Register flags,
2120 bool is_static = false) {
2121 assert_different_registers(cache, index, flags, off);
2123 ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2124 // Field offset
2125 __ movptr(off, Address(cache, index, Address::times_ptr,
2126 in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset())));
2127 // Flags
2128 __ movl(flags, Address(cache, index, Address::times_ptr,
2129 in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset())));
2131 // klass overwrite register
2132 if (is_static) {
2133 __ movptr(obj, Address(cache, index, Address::times_ptr,
2134 in_bytes(cp_base_offset + ConstantPoolCacheEntry::f1_offset())));
2135 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
2136 __ movptr(obj, Address(obj, mirror_offset));
2137 }
2138 }
2140 void TemplateTable::load_invoke_cp_cache_entry(int byte_no,
2141 Register method,
2142 Register itable_index,
2143 Register flags,
2144 bool is_invokevirtual,
2145 bool is_invokevfinal, /*unused*/
2146 bool is_invokedynamic) {
2147 // setup registers
2148 const Register cache = rcx;
2149 const Register index = rdx;
2150 assert_different_registers(method, flags);
2151 assert_different_registers(method, cache, index);
2152 assert_different_registers(itable_index, flags);
2153 assert_different_registers(itable_index, cache, index);
2154 // determine constant pool cache field offsets
2155 assert(is_invokevirtual == (byte_no == f2_byte), "is_invokevirtual flag redundant");
2156 const int method_offset = in_bytes(
2157 ConstantPoolCache::base_offset() +
2158 ((byte_no == f2_byte)
2159 ? ConstantPoolCacheEntry::f2_offset()
2160 : ConstantPoolCacheEntry::f1_offset()));
2161 const int flags_offset = in_bytes(ConstantPoolCache::base_offset() +
2162 ConstantPoolCacheEntry::flags_offset());
2163 // access constant pool cache fields
2164 const int index_offset = in_bytes(ConstantPoolCache::base_offset() +
2165 ConstantPoolCacheEntry::f2_offset());
2167 size_t index_size = (is_invokedynamic ? sizeof(u4) : sizeof(u2));
2168 resolve_cache_and_index(byte_no, cache, index, index_size);
2169 __ movptr(method, Address(cache, index, Address::times_ptr, method_offset));
2171 if (itable_index != noreg) {
2172 __ movptr(itable_index, Address(cache, index, Address::times_ptr, index_offset));
2173 }
2174 __ movl(flags, Address(cache, index, Address::times_ptr, flags_offset));
2175 }
2178 // The registers cache and index expected to be set before call.
2179 // Correct values of the cache and index registers are preserved.
2180 void TemplateTable::jvmti_post_field_access(Register cache,
2181 Register index,
2182 bool is_static,
2183 bool has_tos) {
2184 if (JvmtiExport::can_post_field_access()) {
2185 // Check to see if a field access watch has been set before we take
2186 // the time to call into the VM.
2187 Label L1;
2188 assert_different_registers(cache, index, rax);
2189 __ mov32(rax, ExternalAddress((address) JvmtiExport::get_field_access_count_addr()));
2190 __ testl(rax,rax);
2191 __ jcc(Assembler::zero, L1);
2193 // cache entry pointer
2194 __ addptr(cache, in_bytes(ConstantPoolCache::base_offset()));
2195 __ shll(index, LogBytesPerWord);
2196 __ addptr(cache, index);
2197 if (is_static) {
2198 __ xorptr(rax, rax); // NULL object reference
2199 } else {
2200 __ pop(atos); // Get the object
2201 __ verify_oop(rax);
2202 __ push(atos); // Restore stack state
2203 }
2204 // rax,: object pointer or NULL
2205 // cache: cache entry pointer
2206 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_field_access),
2207 rax, cache);
2208 __ get_cache_and_index_at_bcp(cache, index, 1);
2209 __ bind(L1);
2210 }
2211 }
2213 void TemplateTable::pop_and_check_object(Register r) {
2214 __ pop_ptr(r);
2215 __ null_check(r); // for field access must check obj.
2216 __ verify_oop(r);
2217 }
2219 void TemplateTable::getfield_or_static(int byte_no, bool is_static) {
2220 transition(vtos, vtos);
2222 const Register cache = rcx;
2223 const Register index = rdx;
2224 const Register obj = rcx;
2225 const Register off = rbx;
2226 const Register flags = rax;
2228 resolve_cache_and_index(byte_no, cache, index, sizeof(u2));
2229 jvmti_post_field_access(cache, index, is_static, false);
2230 load_field_cp_cache_entry(obj, cache, index, off, flags, is_static);
2232 if (!is_static) pop_and_check_object(obj);
2234 const Address lo(obj, off, Address::times_1, 0*wordSize);
2235 const Address hi(obj, off, Address::times_1, 1*wordSize);
2237 Label Done, notByte, notInt, notShort, notChar, notLong, notFloat, notObj, notDouble;
2239 __ shrl(flags, ConstantPoolCacheEntry::tos_state_shift);
2240 assert(btos == 0, "change code, btos != 0");
2241 // btos
2242 __ andptr(flags, ConstantPoolCacheEntry::tos_state_mask);
2243 __ jcc(Assembler::notZero, notByte);
2245 __ load_signed_byte(rax, lo );
2246 __ push(btos);
2247 // Rewrite bytecode to be faster
2248 if (!is_static) {
2249 patch_bytecode(Bytecodes::_fast_bgetfield, rcx, rbx);
2250 }
2251 __ jmp(Done);
2253 __ bind(notByte);
2254 // itos
2255 __ cmpl(flags, itos );
2256 __ jcc(Assembler::notEqual, notInt);
2258 __ movl(rax, lo );
2259 __ push(itos);
2260 // Rewrite bytecode to be faster
2261 if (!is_static) {
2262 patch_bytecode(Bytecodes::_fast_igetfield, rcx, rbx);
2263 }
2264 __ jmp(Done);
2266 __ bind(notInt);
2267 // atos
2268 __ cmpl(flags, atos );
2269 __ jcc(Assembler::notEqual, notObj);
2271 __ movl(rax, lo );
2272 __ push(atos);
2273 if (!is_static) {
2274 patch_bytecode(Bytecodes::_fast_agetfield, rcx, rbx);
2275 }
2276 __ jmp(Done);
2278 __ bind(notObj);
2279 // ctos
2280 __ cmpl(flags, ctos );
2281 __ jcc(Assembler::notEqual, notChar);
2283 __ load_unsigned_short(rax, lo );
2284 __ push(ctos);
2285 if (!is_static) {
2286 patch_bytecode(Bytecodes::_fast_cgetfield, rcx, rbx);
2287 }
2288 __ jmp(Done);
2290 __ bind(notChar);
2291 // stos
2292 __ cmpl(flags, stos );
2293 __ jcc(Assembler::notEqual, notShort);
2295 __ load_signed_short(rax, lo );
2296 __ push(stos);
2297 if (!is_static) {
2298 patch_bytecode(Bytecodes::_fast_sgetfield, rcx, rbx);
2299 }
2300 __ jmp(Done);
2302 __ bind(notShort);
2303 // ltos
2304 __ cmpl(flags, ltos );
2305 __ jcc(Assembler::notEqual, notLong);
2307 // Generate code as if volatile. There just aren't enough registers to
2308 // save that information and this code is faster than the test.
2309 __ fild_d(lo); // Must load atomically
2310 __ subptr(rsp,2*wordSize); // Make space for store
2311 __ fistp_d(Address(rsp,0));
2312 __ pop(rax);
2313 __ pop(rdx);
2315 __ push(ltos);
2316 // Don't rewrite to _fast_lgetfield for potential volatile case.
2317 __ jmp(Done);
2319 __ bind(notLong);
2320 // ftos
2321 __ cmpl(flags, ftos );
2322 __ jcc(Assembler::notEqual, notFloat);
2324 __ fld_s(lo);
2325 __ push(ftos);
2326 if (!is_static) {
2327 patch_bytecode(Bytecodes::_fast_fgetfield, rcx, rbx);
2328 }
2329 __ jmp(Done);
2331 __ bind(notFloat);
2332 // dtos
2333 __ cmpl(flags, dtos );
2334 __ jcc(Assembler::notEqual, notDouble);
2336 __ fld_d(lo);
2337 __ push(dtos);
2338 if (!is_static) {
2339 patch_bytecode(Bytecodes::_fast_dgetfield, rcx, rbx);
2340 }
2341 __ jmpb(Done);
2343 __ bind(notDouble);
2345 __ stop("Bad state");
2347 __ bind(Done);
2348 // Doug Lea believes this is not needed with current Sparcs (TSO) and Intel (PSO).
2349 // volatile_barrier( );
2350 }
2353 void TemplateTable::getfield(int byte_no) {
2354 getfield_or_static(byte_no, false);
2355 }
2358 void TemplateTable::getstatic(int byte_no) {
2359 getfield_or_static(byte_no, true);
2360 }
2362 // The registers cache and index expected to be set before call.
2363 // The function may destroy various registers, just not the cache and index registers.
2364 void TemplateTable::jvmti_post_field_mod(Register cache, Register index, bool is_static) {
2366 ByteSize cp_base_offset = ConstantPoolCache::base_offset();
2368 if (JvmtiExport::can_post_field_modification()) {
2369 // Check to see if a field modification watch has been set before we take
2370 // the time to call into the VM.
2371 Label L1;
2372 assert_different_registers(cache, index, rax);
2373 __ mov32(rax, ExternalAddress((address)JvmtiExport::get_field_modification_count_addr()));
2374 __ testl(rax, rax);
2375 __ jcc(Assembler::zero, L1);
2377 // The cache and index registers have been already set.
2378 // This allows to eliminate this call but the cache and index
2379 // registers have to be correspondingly used after this line.
2380 __ get_cache_and_index_at_bcp(rax, rdx, 1);
2382 if (is_static) {
2383 // Life is simple. Null out the object pointer.
2384 __ xorptr(rbx, rbx);
2385 } else {
2386 // Life is harder. The stack holds the value on top, followed by the object.
2387 // We don't know the size of the value, though; it could be one or two words
2388 // depending on its type. As a result, we must find the type to determine where
2389 // the object is.
2390 Label two_word, valsize_known;
2391 __ movl(rcx, Address(rax, rdx, Address::times_ptr, in_bytes(cp_base_offset +
2392 ConstantPoolCacheEntry::flags_offset())));
2393 __ mov(rbx, rsp);
2394 __ shrl(rcx, ConstantPoolCacheEntry::tos_state_shift);
2395 // Make sure we don't need to mask rcx after the above shift
2396 ConstantPoolCacheEntry::verify_tos_state_shift();
2397 __ cmpl(rcx, ltos);
2398 __ jccb(Assembler::equal, two_word);
2399 __ cmpl(rcx, dtos);
2400 __ jccb(Assembler::equal, two_word);
2401 __ addptr(rbx, Interpreter::expr_offset_in_bytes(1)); // one word jvalue (not ltos, dtos)
2402 __ jmpb(valsize_known);
2404 __ bind(two_word);
2405 __ addptr(rbx, Interpreter::expr_offset_in_bytes(2)); // two words jvalue
2407 __ bind(valsize_known);
2408 // setup object pointer
2409 __ movptr(rbx, Address(rbx, 0));
2410 }
2411 // cache entry pointer
2412 __ addptr(rax, in_bytes(cp_base_offset));
2413 __ shll(rdx, LogBytesPerWord);
2414 __ addptr(rax, rdx);
2415 // object (tos)
2416 __ mov(rcx, rsp);
2417 // rbx,: object pointer set up above (NULL if static)
2418 // rax,: cache entry pointer
2419 // rcx: jvalue object on the stack
2420 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_field_modification),
2421 rbx, rax, rcx);
2422 __ get_cache_and_index_at_bcp(cache, index, 1);
2423 __ bind(L1);
2424 }
2425 }
2428 void TemplateTable::putfield_or_static(int byte_no, bool is_static) {
2429 transition(vtos, vtos);
2431 const Register cache = rcx;
2432 const Register index = rdx;
2433 const Register obj = rcx;
2434 const Register off = rbx;
2435 const Register flags = rax;
2437 resolve_cache_and_index(byte_no, cache, index, sizeof(u2));
2438 jvmti_post_field_mod(cache, index, is_static);
2439 load_field_cp_cache_entry(obj, cache, index, off, flags, is_static);
2441 // Doug Lea believes this is not needed with current Sparcs (TSO) and Intel (PSO).
2442 // volatile_barrier( );
2444 Label notVolatile, Done;
2445 __ movl(rdx, flags);
2446 __ shrl(rdx, ConstantPoolCacheEntry::is_volatile_shift);
2447 __ andl(rdx, 0x1);
2449 // field addresses
2450 const Address lo(obj, off, Address::times_1, 0*wordSize);
2451 const Address hi(obj, off, Address::times_1, 1*wordSize);
2453 Label notByte, notInt, notShort, notChar, notLong, notFloat, notObj, notDouble;
2455 __ shrl(flags, ConstantPoolCacheEntry::tos_state_shift);
2456 assert(btos == 0, "change code, btos != 0");
2457 __ andl(flags, ConstantPoolCacheEntry::tos_state_mask);
2458 __ jcc(Assembler::notZero, notByte);
2460 // btos
2461 {
2462 __ pop(btos);
2463 if (!is_static) pop_and_check_object(obj);
2464 __ movb(lo, rax);
2465 if (!is_static) {
2466 patch_bytecode(Bytecodes::_fast_bputfield, rcx, rbx, true, byte_no);
2467 }
2468 __ jmp(Done);
2469 }
2471 __ bind(notByte);
2472 __ cmpl(flags, itos);
2473 __ jcc(Assembler::notEqual, notInt);
2475 // itos
2476 {
2477 __ pop(itos);
2478 if (!is_static) pop_and_check_object(obj);
2479 __ movl(lo, rax);
2480 if (!is_static) {
2481 patch_bytecode(Bytecodes::_fast_iputfield, rcx, rbx, true, byte_no);
2482 }
2483 __ jmp(Done);
2484 }
2486 __ bind(notInt);
2487 __ cmpl(flags, atos);
2488 __ jcc(Assembler::notEqual, notObj);
2490 // atos
2491 {
2492 __ pop(atos);
2493 if (!is_static) pop_and_check_object(obj);
2494 do_oop_store(_masm, lo, rax, _bs->kind(), false);
2495 if (!is_static) {
2496 patch_bytecode(Bytecodes::_fast_aputfield, rcx, rbx, true, byte_no);
2497 }
2498 __ jmp(Done);
2499 }
2501 __ bind(notObj);
2502 __ cmpl(flags, ctos);
2503 __ jcc(Assembler::notEqual, notChar);
2505 // ctos
2506 {
2507 __ pop(ctos);
2508 if (!is_static) pop_and_check_object(obj);
2509 __ movw(lo, rax);
2510 if (!is_static) {
2511 patch_bytecode(Bytecodes::_fast_cputfield, rcx, rbx, true, byte_no);
2512 }
2513 __ jmp(Done);
2514 }
2516 __ bind(notChar);
2517 __ cmpl(flags, stos);
2518 __ jcc(Assembler::notEqual, notShort);
2520 // stos
2521 {
2522 __ pop(stos);
2523 if (!is_static) pop_and_check_object(obj);
2524 __ movw(lo, rax);
2525 if (!is_static) {
2526 patch_bytecode(Bytecodes::_fast_sputfield, rcx, rbx, true, byte_no);
2527 }
2528 __ jmp(Done);
2529 }
2531 __ bind(notShort);
2532 __ cmpl(flags, ltos);
2533 __ jcc(Assembler::notEqual, notLong);
2535 // ltos
2536 {
2537 Label notVolatileLong;
2538 __ testl(rdx, rdx);
2539 __ jcc(Assembler::zero, notVolatileLong);
2541 __ pop(ltos); // overwrites rdx, do this after testing volatile.
2542 if (!is_static) pop_and_check_object(obj);
2544 // Replace with real volatile test
2545 __ push(rdx);
2546 __ push(rax); // Must update atomically with FIST
2547 __ fild_d(Address(rsp,0)); // So load into FPU register
2548 __ fistp_d(lo); // and put into memory atomically
2549 __ addptr(rsp, 2*wordSize);
2550 // volatile_barrier();
2551 volatile_barrier(Assembler::Membar_mask_bits(Assembler::StoreLoad |
2552 Assembler::StoreStore));
2553 // Don't rewrite volatile version
2554 __ jmp(notVolatile);
2556 __ bind(notVolatileLong);
2558 __ pop(ltos); // overwrites rdx
2559 if (!is_static) pop_and_check_object(obj);
2560 NOT_LP64(__ movptr(hi, rdx));
2561 __ movptr(lo, rax);
2562 if (!is_static) {
2563 patch_bytecode(Bytecodes::_fast_lputfield, rcx, rbx, true, byte_no);
2564 }
2565 __ jmp(notVolatile);
2566 }
2568 __ bind(notLong);
2569 __ cmpl(flags, ftos);
2570 __ jcc(Assembler::notEqual, notFloat);
2572 // ftos
2573 {
2574 __ pop(ftos);
2575 if (!is_static) pop_and_check_object(obj);
2576 __ fstp_s(lo);
2577 if (!is_static) {
2578 patch_bytecode(Bytecodes::_fast_fputfield, rcx, rbx, true, byte_no);
2579 }
2580 __ jmp(Done);
2581 }
2583 __ bind(notFloat);
2584 #ifdef ASSERT
2585 __ cmpl(flags, dtos);
2586 __ jcc(Assembler::notEqual, notDouble);
2587 #endif
2589 // dtos
2590 {
2591 __ pop(dtos);
2592 if (!is_static) pop_and_check_object(obj);
2593 __ fstp_d(lo);
2594 if (!is_static) {
2595 patch_bytecode(Bytecodes::_fast_dputfield, rcx, rbx, true, byte_no);
2596 }
2597 __ jmp(Done);
2598 }
2600 #ifdef ASSERT
2601 __ bind(notDouble);
2602 __ stop("Bad state");
2603 #endif
2605 __ bind(Done);
2607 // Check for volatile store
2608 __ testl(rdx, rdx);
2609 __ jcc(Assembler::zero, notVolatile);
2610 volatile_barrier(Assembler::Membar_mask_bits(Assembler::StoreLoad |
2611 Assembler::StoreStore));
2612 __ bind(notVolatile);
2613 }
2616 void TemplateTable::putfield(int byte_no) {
2617 putfield_or_static(byte_no, false);
2618 }
2621 void TemplateTable::putstatic(int byte_no) {
2622 putfield_or_static(byte_no, true);
2623 }
2625 void TemplateTable::jvmti_post_fast_field_mod() {
2626 if (JvmtiExport::can_post_field_modification()) {
2627 // Check to see if a field modification watch has been set before we take
2628 // the time to call into the VM.
2629 Label L2;
2630 __ mov32(rcx, ExternalAddress((address)JvmtiExport::get_field_modification_count_addr()));
2631 __ testl(rcx,rcx);
2632 __ jcc(Assembler::zero, L2);
2633 __ pop_ptr(rbx); // copy the object pointer from tos
2634 __ verify_oop(rbx);
2635 __ push_ptr(rbx); // put the object pointer back on tos
2637 // Save tos values before call_VM() clobbers them. Since we have
2638 // to do it for every data type, we use the saved values as the
2639 // jvalue object.
2640 switch (bytecode()) { // load values into the jvalue object
2641 case Bytecodes::_fast_aputfield: __ push_ptr(rax); break;
2642 case Bytecodes::_fast_bputfield: // fall through
2643 case Bytecodes::_fast_sputfield: // fall through
2644 case Bytecodes::_fast_cputfield: // fall through
2645 case Bytecodes::_fast_iputfield: __ push_i(rax); break;
2646 case Bytecodes::_fast_dputfield: __ push_d(); break;
2647 case Bytecodes::_fast_fputfield: __ push_f(); break;
2648 case Bytecodes::_fast_lputfield: __ push_l(rax); break;
2650 default:
2651 ShouldNotReachHere();
2652 }
2653 __ mov(rcx, rsp); // points to jvalue on the stack
2654 // access constant pool cache entry
2655 __ get_cache_entry_pointer_at_bcp(rax, rdx, 1);
2656 __ verify_oop(rbx);
2657 // rbx,: object pointer copied above
2658 // rax,: cache entry pointer
2659 // rcx: jvalue object on the stack
2660 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_field_modification), rbx, rax, rcx);
2662 switch (bytecode()) { // restore tos values
2663 case Bytecodes::_fast_aputfield: __ pop_ptr(rax); break;
2664 case Bytecodes::_fast_bputfield: // fall through
2665 case Bytecodes::_fast_sputfield: // fall through
2666 case Bytecodes::_fast_cputfield: // fall through
2667 case Bytecodes::_fast_iputfield: __ pop_i(rax); break;
2668 case Bytecodes::_fast_dputfield: __ pop_d(); break;
2669 case Bytecodes::_fast_fputfield: __ pop_f(); break;
2670 case Bytecodes::_fast_lputfield: __ pop_l(rax); break;
2671 }
2672 __ bind(L2);
2673 }
2674 }
2676 void TemplateTable::fast_storefield(TosState state) {
2677 transition(state, vtos);
2679 ByteSize base = ConstantPoolCache::base_offset();
2681 jvmti_post_fast_field_mod();
2683 // access constant pool cache
2684 __ get_cache_and_index_at_bcp(rcx, rbx, 1);
2686 // test for volatile with rdx but rdx is tos register for lputfield.
2687 if (bytecode() == Bytecodes::_fast_lputfield) __ push(rdx);
2688 __ movl(rdx, Address(rcx, rbx, Address::times_ptr, in_bytes(base +
2689 ConstantPoolCacheEntry::flags_offset())));
2691 // replace index with field offset from cache entry
2692 __ movptr(rbx, Address(rcx, rbx, Address::times_ptr, in_bytes(base + ConstantPoolCacheEntry::f2_offset())));
2694 // Doug Lea believes this is not needed with current Sparcs (TSO) and Intel (PSO).
2695 // volatile_barrier( );
2697 Label notVolatile, Done;
2698 __ shrl(rdx, ConstantPoolCacheEntry::is_volatile_shift);
2699 __ andl(rdx, 0x1);
2700 // Check for volatile store
2701 __ testl(rdx, rdx);
2702 __ jcc(Assembler::zero, notVolatile);
2704 if (bytecode() == Bytecodes::_fast_lputfield) __ pop(rdx);
2706 // Get object from stack
2707 pop_and_check_object(rcx);
2709 // field addresses
2710 const Address lo(rcx, rbx, Address::times_1, 0*wordSize);
2711 const Address hi(rcx, rbx, Address::times_1, 1*wordSize);
2713 // access field
2714 switch (bytecode()) {
2715 case Bytecodes::_fast_bputfield: __ movb(lo, rax); break;
2716 case Bytecodes::_fast_sputfield: // fall through
2717 case Bytecodes::_fast_cputfield: __ movw(lo, rax); break;
2718 case Bytecodes::_fast_iputfield: __ movl(lo, rax); break;
2719 case Bytecodes::_fast_lputfield:
2720 NOT_LP64(__ movptr(hi, rdx));
2721 __ movptr(lo, rax);
2722 break;
2723 case Bytecodes::_fast_fputfield: __ fstp_s(lo); break;
2724 case Bytecodes::_fast_dputfield: __ fstp_d(lo); break;
2725 case Bytecodes::_fast_aputfield: {
2726 do_oop_store(_masm, lo, rax, _bs->kind(), false);
2727 break;
2728 }
2729 default:
2730 ShouldNotReachHere();
2731 }
2733 Label done;
2734 volatile_barrier(Assembler::Membar_mask_bits(Assembler::StoreLoad |
2735 Assembler::StoreStore));
2736 // Barriers are so large that short branch doesn't reach!
2737 __ jmp(done);
2739 // Same code as above, but don't need rdx to test for volatile.
2740 __ bind(notVolatile);
2742 if (bytecode() == Bytecodes::_fast_lputfield) __ pop(rdx);
2744 // Get object from stack
2745 pop_and_check_object(rcx);
2747 // access field
2748 switch (bytecode()) {
2749 case Bytecodes::_fast_bputfield: __ movb(lo, rax); break;
2750 case Bytecodes::_fast_sputfield: // fall through
2751 case Bytecodes::_fast_cputfield: __ movw(lo, rax); break;
2752 case Bytecodes::_fast_iputfield: __ movl(lo, rax); break;
2753 case Bytecodes::_fast_lputfield:
2754 NOT_LP64(__ movptr(hi, rdx));
2755 __ movptr(lo, rax);
2756 break;
2757 case Bytecodes::_fast_fputfield: __ fstp_s(lo); break;
2758 case Bytecodes::_fast_dputfield: __ fstp_d(lo); break;
2759 case Bytecodes::_fast_aputfield: {
2760 do_oop_store(_masm, lo, rax, _bs->kind(), false);
2761 break;
2762 }
2763 default:
2764 ShouldNotReachHere();
2765 }
2766 __ bind(done);
2767 }
2770 void TemplateTable::fast_accessfield(TosState state) {
2771 transition(atos, state);
2773 // do the JVMTI work here to avoid disturbing the register state below
2774 if (JvmtiExport::can_post_field_access()) {
2775 // Check to see if a field access watch has been set before we take
2776 // the time to call into the VM.
2777 Label L1;
2778 __ mov32(rcx, ExternalAddress((address) JvmtiExport::get_field_access_count_addr()));
2779 __ testl(rcx,rcx);
2780 __ jcc(Assembler::zero, L1);
2781 // access constant pool cache entry
2782 __ get_cache_entry_pointer_at_bcp(rcx, rdx, 1);
2783 __ push_ptr(rax); // save object pointer before call_VM() clobbers it
2784 __ verify_oop(rax);
2785 // rax,: object pointer copied above
2786 // rcx: cache entry pointer
2787 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_field_access), rax, rcx);
2788 __ pop_ptr(rax); // restore object pointer
2789 __ bind(L1);
2790 }
2792 // access constant pool cache
2793 __ get_cache_and_index_at_bcp(rcx, rbx, 1);
2794 // replace index with field offset from cache entry
2795 __ movptr(rbx, Address(rcx,
2796 rbx,
2797 Address::times_ptr,
2798 in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::f2_offset())));
2801 // rax,: object
2802 __ verify_oop(rax);
2803 __ null_check(rax);
2804 // field addresses
2805 const Address lo = Address(rax, rbx, Address::times_1, 0*wordSize);
2806 const Address hi = Address(rax, rbx, Address::times_1, 1*wordSize);
2808 // access field
2809 switch (bytecode()) {
2810 case Bytecodes::_fast_bgetfield: __ movsbl(rax, lo ); break;
2811 case Bytecodes::_fast_sgetfield: __ load_signed_short(rax, lo ); break;
2812 case Bytecodes::_fast_cgetfield: __ load_unsigned_short(rax, lo ); break;
2813 case Bytecodes::_fast_igetfield: __ movl(rax, lo); break;
2814 case Bytecodes::_fast_lgetfield: __ stop("should not be rewritten"); break;
2815 case Bytecodes::_fast_fgetfield: __ fld_s(lo); break;
2816 case Bytecodes::_fast_dgetfield: __ fld_d(lo); break;
2817 case Bytecodes::_fast_agetfield: __ movptr(rax, lo); __ verify_oop(rax); break;
2818 default:
2819 ShouldNotReachHere();
2820 }
2822 // Doug Lea believes this is not needed with current Sparcs(TSO) and Intel(PSO)
2823 // volatile_barrier( );
2824 }
2826 void TemplateTable::fast_xaccess(TosState state) {
2827 transition(vtos, state);
2828 // get receiver
2829 __ movptr(rax, aaddress(0));
2830 // access constant pool cache
2831 __ get_cache_and_index_at_bcp(rcx, rdx, 2);
2832 __ movptr(rbx, Address(rcx,
2833 rdx,
2834 Address::times_ptr,
2835 in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::f2_offset())));
2836 // make sure exception is reported in correct bcp range (getfield is next instruction)
2837 __ increment(rsi);
2838 __ null_check(rax);
2839 const Address lo = Address(rax, rbx, Address::times_1, 0*wordSize);
2840 if (state == itos) {
2841 __ movl(rax, lo);
2842 } else if (state == atos) {
2843 __ movptr(rax, lo);
2844 __ verify_oop(rax);
2845 } else if (state == ftos) {
2846 __ fld_s(lo);
2847 } else {
2848 ShouldNotReachHere();
2849 }
2850 __ decrement(rsi);
2851 }
2855 //----------------------------------------------------------------------------------------------------
2856 // Calls
2858 void TemplateTable::count_calls(Register method, Register temp) {
2859 // implemented elsewhere
2860 ShouldNotReachHere();
2861 }
2864 void TemplateTable::prepare_invoke(int byte_no,
2865 Register method, // linked method (or i-klass)
2866 Register index, // itable index, MethodType, etc.
2867 Register recv, // if caller wants to see it
2868 Register flags // if caller wants to test it
2869 ) {
2870 // determine flags
2871 const Bytecodes::Code code = bytecode();
2872 const bool is_invokeinterface = code == Bytecodes::_invokeinterface;
2873 const bool is_invokedynamic = code == Bytecodes::_invokedynamic;
2874 const bool is_invokehandle = code == Bytecodes::_invokehandle;
2875 const bool is_invokevirtual = code == Bytecodes::_invokevirtual;
2876 const bool is_invokespecial = code == Bytecodes::_invokespecial;
2877 const bool load_receiver = (recv != noreg);
2878 const bool save_flags = (flags != noreg);
2879 assert(load_receiver == (code != Bytecodes::_invokestatic && code != Bytecodes::_invokedynamic), "");
2880 assert(save_flags == (is_invokeinterface || is_invokevirtual), "need flags for vfinal");
2881 assert(flags == noreg || flags == rdx, "");
2882 assert(recv == noreg || recv == rcx, "");
2884 // setup registers & access constant pool cache
2885 if (recv == noreg) recv = rcx;
2886 if (flags == noreg) flags = rdx;
2887 assert_different_registers(method, index, recv, flags);
2889 // save 'interpreter return address'
2890 __ save_bcp();
2892 load_invoke_cp_cache_entry(byte_no, method, index, flags, is_invokevirtual, false, is_invokedynamic);
2894 // maybe push appendix to arguments (just before return address)
2895 if (is_invokedynamic || is_invokehandle) {
2896 Label L_no_push;
2897 __ testl(flags, (1 << ConstantPoolCacheEntry::has_appendix_shift));
2898 __ jccb(Assembler::zero, L_no_push);
2899 // Push the appendix as a trailing parameter.
2900 // This must be done before we get the receiver,
2901 // since the parameter_size includes it.
2902 __ push(rbx);
2903 __ mov(rbx, index);
2904 assert(ConstantPoolCacheEntry::_indy_resolved_references_appendix_offset == 0, "appendix expected at index+0");
2905 __ load_resolved_reference_at_index(index, rbx);
2906 __ pop(rbx);
2907 __ push(index); // push appendix (MethodType, CallSite, etc.)
2908 __ bind(L_no_push);
2909 }
2911 // load receiver if needed (note: no return address pushed yet)
2912 if (load_receiver) {
2913 __ movl(recv, flags);
2914 __ andl(recv, ConstantPoolCacheEntry::parameter_size_mask);
2915 const int no_return_pc_pushed_yet = -1; // argument slot correction before we push return address
2916 const int receiver_is_at_end = -1; // back off one slot to get receiver
2917 Address recv_addr = __ argument_address(recv, no_return_pc_pushed_yet + receiver_is_at_end);
2918 __ movptr(recv, recv_addr);
2919 __ verify_oop(recv);
2920 }
2922 if (save_flags) {
2923 __ mov(rsi, flags);
2924 }
2926 // compute return type
2927 __ shrl(flags, ConstantPoolCacheEntry::tos_state_shift);
2928 // Make sure we don't need to mask flags after the above shift
2929 ConstantPoolCacheEntry::verify_tos_state_shift();
2930 // load return address
2931 {
2932 const address table_addr = (is_invokeinterface || is_invokedynamic) ?
2933 (address)Interpreter::return_5_addrs_by_index_table() :
2934 (address)Interpreter::return_3_addrs_by_index_table();
2935 ExternalAddress table(table_addr);
2936 __ movptr(flags, ArrayAddress(table, Address(noreg, flags, Address::times_ptr)));
2937 }
2939 // push return address
2940 __ push(flags);
2942 // Restore flags value from the constant pool cache, and restore rsi
2943 // for later null checks. rsi is the bytecode pointer
2944 if (save_flags) {
2945 __ mov(flags, rsi);
2946 __ restore_bcp();
2947 }
2948 }
2951 void TemplateTable::invokevirtual_helper(Register index,
2952 Register recv,
2953 Register flags) {
2954 // Uses temporary registers rax, rdx
2955 assert_different_registers(index, recv, rax, rdx);
2956 assert(index == rbx, "");
2957 assert(recv == rcx, "");
2959 // Test for an invoke of a final method
2960 Label notFinal;
2961 __ movl(rax, flags);
2962 __ andl(rax, (1 << ConstantPoolCacheEntry::is_vfinal_shift));
2963 __ jcc(Assembler::zero, notFinal);
2965 const Register method = index; // method must be rbx
2966 assert(method == rbx,
2967 "Method* must be rbx for interpreter calling convention");
2969 // do the call - the index is actually the method to call
2970 // that is, f2 is a vtable index if !is_vfinal, else f2 is a Method*
2972 // It's final, need a null check here!
2973 __ null_check(recv);
2975 // profile this call
2976 __ profile_final_call(rax);
2977 __ profile_arguments_type(rax, method, rsi, true);
2979 __ jump_from_interpreted(method, rax);
2981 __ bind(notFinal);
2983 // get receiver klass
2984 __ null_check(recv, oopDesc::klass_offset_in_bytes());
2985 __ load_klass(rax, recv);
2987 // profile this call
2988 __ profile_virtual_call(rax, rdi, rdx);
2990 // get target Method* & entry point
2991 __ lookup_virtual_method(rax, index, method);
2992 __ profile_arguments_type(rdx, method, rsi, true);
2993 __ jump_from_interpreted(method, rdx);
2994 }
2997 void TemplateTable::invokevirtual(int byte_no) {
2998 transition(vtos, vtos);
2999 assert(byte_no == f2_byte, "use this argument");
3000 prepare_invoke(byte_no,
3001 rbx, // method or vtable index
3002 noreg, // unused itable index
3003 rcx, rdx); // recv, flags
3005 // rbx: index
3006 // rcx: receiver
3007 // rdx: flags
3009 invokevirtual_helper(rbx, rcx, rdx);
3010 }
3013 void TemplateTable::invokespecial(int byte_no) {
3014 transition(vtos, vtos);
3015 assert(byte_no == f1_byte, "use this argument");
3016 prepare_invoke(byte_no, rbx, noreg, // get f1 Method*
3017 rcx); // get receiver also for null check
3018 __ verify_oop(rcx);
3019 __ null_check(rcx);
3020 // do the call
3021 __ profile_call(rax);
3022 __ profile_arguments_type(rax, rbx, rsi, false);
3023 __ jump_from_interpreted(rbx, rax);
3024 }
3027 void TemplateTable::invokestatic(int byte_no) {
3028 transition(vtos, vtos);
3029 assert(byte_no == f1_byte, "use this argument");
3030 prepare_invoke(byte_no, rbx); // get f1 Method*
3031 // do the call
3032 __ profile_call(rax);
3033 __ profile_arguments_type(rax, rbx, rsi, false);
3034 __ jump_from_interpreted(rbx, rax);
3035 }
3038 void TemplateTable::fast_invokevfinal(int byte_no) {
3039 transition(vtos, vtos);
3040 assert(byte_no == f2_byte, "use this argument");
3041 __ stop("fast_invokevfinal not used on x86");
3042 }
3045 void TemplateTable::invokeinterface(int byte_no) {
3046 transition(vtos, vtos);
3047 assert(byte_no == f1_byte, "use this argument");
3048 prepare_invoke(byte_no, rax, rbx, // get f1 Klass*, f2 itable index
3049 rcx, rdx); // recv, flags
3051 // rax: interface klass (from f1)
3052 // rbx: itable index (from f2)
3053 // rcx: receiver
3054 // rdx: flags
3056 // Special case of invokeinterface called for virtual method of
3057 // java.lang.Object. See cpCacheOop.cpp for details.
3058 // This code isn't produced by javac, but could be produced by
3059 // another compliant java compiler.
3060 Label notMethod;
3061 __ movl(rdi, rdx);
3062 __ andl(rdi, (1 << ConstantPoolCacheEntry::is_forced_virtual_shift));
3063 __ jcc(Assembler::zero, notMethod);
3065 invokevirtual_helper(rbx, rcx, rdx);
3066 __ bind(notMethod);
3068 // Get receiver klass into rdx - also a null check
3069 __ restore_locals(); // restore rdi
3070 __ null_check(rcx, oopDesc::klass_offset_in_bytes());
3071 __ load_klass(rdx, rcx);
3073 // profile this call
3074 __ profile_virtual_call(rdx, rsi, rdi);
3076 Label no_such_interface, no_such_method;
3078 __ lookup_interface_method(// inputs: rec. class, interface, itable index
3079 rdx, rax, rbx,
3080 // outputs: method, scan temp. reg
3081 rbx, rsi,
3082 no_such_interface);
3084 // rbx: Method* to call
3085 // rcx: receiver
3086 // Check for abstract method error
3087 // Note: This should be done more efficiently via a throw_abstract_method_error
3088 // interpreter entry point and a conditional jump to it in case of a null
3089 // method.
3090 __ testptr(rbx, rbx);
3091 __ jcc(Assembler::zero, no_such_method);
3093 __ profile_arguments_type(rdx, rbx, rsi, true);
3095 // do the call
3096 // rcx: receiver
3097 // rbx,: Method*
3098 __ jump_from_interpreted(rbx, rdx);
3099 __ should_not_reach_here();
3101 // exception handling code follows...
3102 // note: must restore interpreter registers to canonical
3103 // state for exception handling to work correctly!
3105 __ bind(no_such_method);
3106 // throw exception
3107 __ pop(rbx); // pop return address (pushed by prepare_invoke)
3108 __ restore_bcp(); // rsi must be correct for exception handler (was destroyed)
3109 __ restore_locals(); // make sure locals pointer is correct as well (was destroyed)
3110 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_AbstractMethodError));
3111 // the call_VM checks for exception, so we should never return here.
3112 __ should_not_reach_here();
3114 __ bind(no_such_interface);
3115 // throw exception
3116 __ pop(rbx); // pop return address (pushed by prepare_invoke)
3117 __ restore_bcp(); // rsi must be correct for exception handler (was destroyed)
3118 __ restore_locals(); // make sure locals pointer is correct as well (was destroyed)
3119 __ call_VM(noreg, CAST_FROM_FN_PTR(address,
3120 InterpreterRuntime::throw_IncompatibleClassChangeError));
3121 // the call_VM checks for exception, so we should never return here.
3122 __ should_not_reach_here();
3123 }
3125 void TemplateTable::invokehandle(int byte_no) {
3126 transition(vtos, vtos);
3127 assert(byte_no == f1_byte, "use this argument");
3128 const Register rbx_method = rbx;
3129 const Register rax_mtype = rax;
3130 const Register rcx_recv = rcx;
3131 const Register rdx_flags = rdx;
3133 if (!EnableInvokeDynamic) {
3134 // rewriter does not generate this bytecode
3135 __ should_not_reach_here();
3136 return;
3137 }
3139 prepare_invoke(byte_no, rbx_method, rax_mtype, rcx_recv);
3140 __ verify_method_ptr(rbx_method);
3141 __ verify_oop(rcx_recv);
3142 __ null_check(rcx_recv);
3144 // rax: MethodType object (from cpool->resolved_references[f1], if necessary)
3145 // rbx: MH.invokeExact_MT method (from f2)
3147 // Note: rax_mtype is already pushed (if necessary) by prepare_invoke
3149 // FIXME: profile the LambdaForm also
3150 __ profile_final_call(rax);
3151 __ profile_arguments_type(rdx, rbx_method, rsi, true);
3153 __ jump_from_interpreted(rbx_method, rdx);
3154 }
3157 void TemplateTable::invokedynamic(int byte_no) {
3158 transition(vtos, vtos);
3159 assert(byte_no == f1_byte, "use this argument");
3161 if (!EnableInvokeDynamic) {
3162 // We should not encounter this bytecode if !EnableInvokeDynamic.
3163 // The verifier will stop it. However, if we get past the verifier,
3164 // this will stop the thread in a reasonable way, without crashing the JVM.
3165 __ call_VM(noreg, CAST_FROM_FN_PTR(address,
3166 InterpreterRuntime::throw_IncompatibleClassChangeError));
3167 // the call_VM checks for exception, so we should never return here.
3168 __ should_not_reach_here();
3169 return;
3170 }
3172 const Register rbx_method = rbx;
3173 const Register rax_callsite = rax;
3175 prepare_invoke(byte_no, rbx_method, rax_callsite);
3177 // rax: CallSite object (from cpool->resolved_references[f1])
3178 // rbx: MH.linkToCallSite method (from f2)
3180 // Note: rax_callsite is already pushed by prepare_invoke
3182 // %%% should make a type profile for any invokedynamic that takes a ref argument
3183 // profile this call
3184 __ profile_call(rsi);
3185 __ profile_arguments_type(rdx, rbx, rsi, false);
3187 __ verify_oop(rax_callsite);
3189 __ jump_from_interpreted(rbx_method, rdx);
3190 }
3192 //----------------------------------------------------------------------------------------------------
3193 // Allocation
3195 void TemplateTable::_new() {
3196 transition(vtos, atos);
3197 __ get_unsigned_2_byte_index_at_bcp(rdx, 1);
3198 Label slow_case;
3199 Label slow_case_no_pop;
3200 Label done;
3201 Label initialize_header;
3202 Label initialize_object; // including clearing the fields
3203 Label allocate_shared;
3205 __ get_cpool_and_tags(rcx, rax);
3207 // Make sure the class we're about to instantiate has been resolved.
3208 // This is done before loading InstanceKlass to be consistent with the order
3209 // how Constant Pool is updated (see ConstantPool::klass_at_put)
3210 const int tags_offset = Array<u1>::base_offset_in_bytes();
3211 __ cmpb(Address(rax, rdx, Address::times_1, tags_offset), JVM_CONSTANT_Class);
3212 __ jcc(Assembler::notEqual, slow_case_no_pop);
3214 // get InstanceKlass
3215 __ movptr(rcx, Address(rcx, rdx, Address::times_ptr, sizeof(ConstantPool)));
3216 __ push(rcx); // save the contexts of klass for initializing the header
3218 // make sure klass is initialized & doesn't have finalizer
3219 // make sure klass is fully initialized
3220 __ cmpb(Address(rcx, InstanceKlass::init_state_offset()), InstanceKlass::fully_initialized);
3221 __ jcc(Assembler::notEqual, slow_case);
3223 // get instance_size in InstanceKlass (scaled to a count of bytes)
3224 __ movl(rdx, Address(rcx, Klass::layout_helper_offset()));
3225 // test to see if it has a finalizer or is malformed in some way
3226 __ testl(rdx, Klass::_lh_instance_slow_path_bit);
3227 __ jcc(Assembler::notZero, slow_case);
3229 //
3230 // Allocate the instance
3231 // 1) Try to allocate in the TLAB
3232 // 2) if fail and the object is large allocate in the shared Eden
3233 // 3) if the above fails (or is not applicable), go to a slow case
3234 // (creates a new TLAB, etc.)
3236 const bool allow_shared_alloc =
3237 Universe::heap()->supports_inline_contig_alloc() && !CMSIncrementalMode;
3239 const Register thread = rcx;
3240 if (UseTLAB || allow_shared_alloc) {
3241 __ get_thread(thread);
3242 }
3244 if (UseTLAB) {
3245 __ movptr(rax, Address(thread, in_bytes(JavaThread::tlab_top_offset())));
3246 __ lea(rbx, Address(rax, rdx, Address::times_1));
3247 __ cmpptr(rbx, Address(thread, in_bytes(JavaThread::tlab_end_offset())));
3248 __ jcc(Assembler::above, allow_shared_alloc ? allocate_shared : slow_case);
3249 __ movptr(Address(thread, in_bytes(JavaThread::tlab_top_offset())), rbx);
3250 if (ZeroTLAB) {
3251 // the fields have been already cleared
3252 __ jmp(initialize_header);
3253 } else {
3254 // initialize both the header and fields
3255 __ jmp(initialize_object);
3256 }
3257 }
3259 // Allocation in the shared Eden, if allowed.
3260 //
3261 // rdx: instance size in bytes
3262 if (allow_shared_alloc) {
3263 __ bind(allocate_shared);
3265 ExternalAddress heap_top((address)Universe::heap()->top_addr());
3267 Label retry;
3268 __ bind(retry);
3269 __ movptr(rax, heap_top);
3270 __ lea(rbx, Address(rax, rdx, Address::times_1));
3271 __ cmpptr(rbx, ExternalAddress((address)Universe::heap()->end_addr()));
3272 __ jcc(Assembler::above, slow_case);
3274 // Compare rax, with the top addr, and if still equal, store the new
3275 // top addr in rbx, at the address of the top addr pointer. Sets ZF if was
3276 // equal, and clears it otherwise. Use lock prefix for atomicity on MPs.
3277 //
3278 // rax,: object begin
3279 // rbx,: object end
3280 // rdx: instance size in bytes
3281 __ locked_cmpxchgptr(rbx, heap_top);
3283 // if someone beat us on the allocation, try again, otherwise continue
3284 __ jcc(Assembler::notEqual, retry);
3286 __ incr_allocated_bytes(thread, rdx, 0);
3287 }
3289 if (UseTLAB || Universe::heap()->supports_inline_contig_alloc()) {
3290 // The object is initialized before the header. If the object size is
3291 // zero, go directly to the header initialization.
3292 __ bind(initialize_object);
3293 __ decrement(rdx, sizeof(oopDesc));
3294 __ jcc(Assembler::zero, initialize_header);
3296 // Initialize topmost object field, divide rdx by 8, check if odd and
3297 // test if zero.
3298 __ xorl(rcx, rcx); // use zero reg to clear memory (shorter code)
3299 __ shrl(rdx, LogBytesPerLong); // divide by 2*oopSize and set carry flag if odd
3301 // rdx must have been multiple of 8
3302 #ifdef ASSERT
3303 // make sure rdx was multiple of 8
3304 Label L;
3305 // Ignore partial flag stall after shrl() since it is debug VM
3306 __ jccb(Assembler::carryClear, L);
3307 __ stop("object size is not multiple of 2 - adjust this code");
3308 __ bind(L);
3309 // rdx must be > 0, no extra check needed here
3310 #endif
3312 // initialize remaining object fields: rdx was a multiple of 8
3313 { Label loop;
3314 __ bind(loop);
3315 __ movptr(Address(rax, rdx, Address::times_8, sizeof(oopDesc) - 1*oopSize), rcx);
3316 NOT_LP64(__ movptr(Address(rax, rdx, Address::times_8, sizeof(oopDesc) - 2*oopSize), rcx));
3317 __ decrement(rdx);
3318 __ jcc(Assembler::notZero, loop);
3319 }
3321 // initialize object header only.
3322 __ bind(initialize_header);
3323 if (UseBiasedLocking) {
3324 __ pop(rcx); // get saved klass back in the register.
3325 __ movptr(rbx, Address(rcx, Klass::prototype_header_offset()));
3326 __ movptr(Address(rax, oopDesc::mark_offset_in_bytes ()), rbx);
3327 } else {
3328 __ movptr(Address(rax, oopDesc::mark_offset_in_bytes ()),
3329 (int32_t)markOopDesc::prototype()); // header
3330 __ pop(rcx); // get saved klass back in the register.
3331 }
3332 __ store_klass(rax, rcx); // klass
3334 {
3335 SkipIfEqual skip_if(_masm, &DTraceAllocProbes, 0);
3336 // Trigger dtrace event for fastpath
3337 __ push(atos);
3338 __ call_VM_leaf(
3339 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc), rax);
3340 __ pop(atos);
3341 }
3343 __ jmp(done);
3344 }
3346 // slow case
3347 __ bind(slow_case);
3348 __ pop(rcx); // restore stack pointer to what it was when we came in.
3349 __ bind(slow_case_no_pop);
3350 __ get_constant_pool(rax);
3351 __ get_unsigned_2_byte_index_at_bcp(rdx, 1);
3352 call_VM(rax, CAST_FROM_FN_PTR(address, InterpreterRuntime::_new), rax, rdx);
3354 // continue
3355 __ bind(done);
3356 }
3359 void TemplateTable::newarray() {
3360 transition(itos, atos);
3361 __ push_i(rax); // make sure everything is on the stack
3362 __ load_unsigned_byte(rdx, at_bcp(1));
3363 call_VM(rax, CAST_FROM_FN_PTR(address, InterpreterRuntime::newarray), rdx, rax);
3364 __ pop_i(rdx); // discard size
3365 }
3368 void TemplateTable::anewarray() {
3369 transition(itos, atos);
3370 __ get_unsigned_2_byte_index_at_bcp(rdx, 1);
3371 __ get_constant_pool(rcx);
3372 call_VM(rax, CAST_FROM_FN_PTR(address, InterpreterRuntime::anewarray), rcx, rdx, rax);
3373 }
3376 void TemplateTable::arraylength() {
3377 transition(atos, itos);
3378 __ null_check(rax, arrayOopDesc::length_offset_in_bytes());
3379 __ movl(rax, Address(rax, arrayOopDesc::length_offset_in_bytes()));
3380 }
3383 void TemplateTable::checkcast() {
3384 transition(atos, atos);
3385 Label done, is_null, ok_is_subtype, quicked, resolved;
3386 __ testptr(rax, rax); // Object is in EAX
3387 __ jcc(Assembler::zero, is_null);
3389 // Get cpool & tags index
3390 __ get_cpool_and_tags(rcx, rdx); // ECX=cpool, EDX=tags array
3391 __ get_unsigned_2_byte_index_at_bcp(rbx, 1); // EBX=index
3392 // See if bytecode has already been quicked
3393 __ cmpb(Address(rdx, rbx, Address::times_1, Array<u1>::base_offset_in_bytes()), JVM_CONSTANT_Class);
3394 __ jcc(Assembler::equal, quicked);
3396 __ push(atos);
3397 call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::quicken_io_cc) );
3398 // vm_result_2 has metadata result
3399 // borrow rdi from locals
3400 __ get_thread(rdi);
3401 __ get_vm_result_2(rax, rdi);
3402 __ restore_locals();
3403 __ pop_ptr(rdx);
3404 __ jmpb(resolved);
3406 // Get superklass in EAX and subklass in EBX
3407 __ bind(quicked);
3408 __ mov(rdx, rax); // Save object in EDX; EAX needed for subtype check
3409 __ movptr(rax, Address(rcx, rbx, Address::times_ptr, sizeof(ConstantPool)));
3411 __ bind(resolved);
3412 __ load_klass(rbx, rdx);
3414 // Generate subtype check. Blows ECX. Resets EDI. Object in EDX.
3415 // Superklass in EAX. Subklass in EBX.
3416 __ gen_subtype_check( rbx, ok_is_subtype );
3418 // Come here on failure
3419 __ push(rdx);
3420 // object is at TOS
3421 __ jump(ExternalAddress(Interpreter::_throw_ClassCastException_entry));
3423 // Come here on success
3424 __ bind(ok_is_subtype);
3425 __ mov(rax,rdx); // Restore object in EDX
3427 // Collect counts on whether this check-cast sees NULLs a lot or not.
3428 if (ProfileInterpreter) {
3429 __ jmp(done);
3430 __ bind(is_null);
3431 __ profile_null_seen(rcx);
3432 } else {
3433 __ bind(is_null); // same as 'done'
3434 }
3435 __ bind(done);
3436 }
3439 void TemplateTable::instanceof() {
3440 transition(atos, itos);
3441 Label done, is_null, ok_is_subtype, quicked, resolved;
3442 __ testptr(rax, rax);
3443 __ jcc(Assembler::zero, is_null);
3445 // Get cpool & tags index
3446 __ get_cpool_and_tags(rcx, rdx); // ECX=cpool, EDX=tags array
3447 __ get_unsigned_2_byte_index_at_bcp(rbx, 1); // EBX=index
3448 // See if bytecode has already been quicked
3449 __ cmpb(Address(rdx, rbx, Address::times_1, Array<u1>::base_offset_in_bytes()), JVM_CONSTANT_Class);
3450 __ jcc(Assembler::equal, quicked);
3452 __ push(atos);
3453 call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::quicken_io_cc) );
3454 // vm_result_2 has metadata result
3455 // borrow rdi from locals
3456 __ get_thread(rdi);
3457 __ get_vm_result_2(rax, rdi);
3458 __ restore_locals();
3459 __ pop_ptr(rdx);
3460 __ load_klass(rdx, rdx);
3461 __ jmp(resolved);
3463 // Get superklass in EAX and subklass in EDX
3464 __ bind(quicked);
3465 __ load_klass(rdx, rax);
3466 __ movptr(rax, Address(rcx, rbx, Address::times_ptr, sizeof(ConstantPool)));
3468 __ bind(resolved);
3470 // Generate subtype check. Blows ECX. Resets EDI.
3471 // Superklass in EAX. Subklass in EDX.
3472 __ gen_subtype_check( rdx, ok_is_subtype );
3474 // Come here on failure
3475 __ xorl(rax,rax);
3476 __ jmpb(done);
3477 // Come here on success
3478 __ bind(ok_is_subtype);
3479 __ movl(rax, 1);
3481 // Collect counts on whether this test sees NULLs a lot or not.
3482 if (ProfileInterpreter) {
3483 __ jmp(done);
3484 __ bind(is_null);
3485 __ profile_null_seen(rcx);
3486 } else {
3487 __ bind(is_null); // same as 'done'
3488 }
3489 __ bind(done);
3490 // rax, = 0: obj == NULL or obj is not an instanceof the specified klass
3491 // rax, = 1: obj != NULL and obj is an instanceof the specified klass
3492 }
3495 //----------------------------------------------------------------------------------------------------
3496 // Breakpoints
3497 void TemplateTable::_breakpoint() {
3499 // Note: We get here even if we are single stepping..
3500 // jbug inists on setting breakpoints at every bytecode
3501 // even if we are in single step mode.
3503 transition(vtos, vtos);
3505 // get the unpatched byte code
3506 __ get_method(rcx);
3507 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::get_original_bytecode_at), rcx, rsi);
3508 __ mov(rbx, rax);
3510 // post the breakpoint event
3511 __ get_method(rcx);
3512 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::_breakpoint), rcx, rsi);
3514 // complete the execution of original bytecode
3515 __ dispatch_only_normal(vtos);
3516 }
3519 //----------------------------------------------------------------------------------------------------
3520 // Exceptions
3522 void TemplateTable::athrow() {
3523 transition(atos, vtos);
3524 __ null_check(rax);
3525 __ jump(ExternalAddress(Interpreter::throw_exception_entry()));
3526 }
3529 //----------------------------------------------------------------------------------------------------
3530 // Synchronization
3531 //
3532 // Note: monitorenter & exit are symmetric routines; which is reflected
3533 // in the assembly code structure as well
3534 //
3535 // Stack layout:
3536 //
3537 // [expressions ] <--- rsp = expression stack top
3538 // ..
3539 // [expressions ]
3540 // [monitor entry] <--- monitor block top = expression stack bot
3541 // ..
3542 // [monitor entry]
3543 // [frame data ] <--- monitor block bot
3544 // ...
3545 // [saved rbp, ] <--- rbp,
3548 void TemplateTable::monitorenter() {
3549 transition(atos, vtos);
3551 // check for NULL object
3552 __ null_check(rax);
3554 const Address monitor_block_top(rbp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
3555 const Address monitor_block_bot(rbp, frame::interpreter_frame_initial_sp_offset * wordSize);
3556 const int entry_size = ( frame::interpreter_frame_monitor_size() * wordSize);
3557 Label allocated;
3559 // initialize entry pointer
3560 __ xorl(rdx, rdx); // points to free slot or NULL
3562 // find a free slot in the monitor block (result in rdx)
3563 { Label entry, loop, exit;
3564 __ movptr(rcx, monitor_block_top); // points to current entry, starting with top-most entry
3566 __ lea(rbx, monitor_block_bot); // points to word before bottom of monitor block
3567 __ jmpb(entry);
3569 __ bind(loop);
3570 __ cmpptr(Address(rcx, BasicObjectLock::obj_offset_in_bytes()), (int32_t)NULL_WORD); // check if current entry is used
3571 __ cmovptr(Assembler::equal, rdx, rcx); // if not used then remember entry in rdx
3572 __ cmpptr(rax, Address(rcx, BasicObjectLock::obj_offset_in_bytes())); // check if current entry is for same object
3573 __ jccb(Assembler::equal, exit); // if same object then stop searching
3574 __ addptr(rcx, entry_size); // otherwise advance to next entry
3575 __ bind(entry);
3576 __ cmpptr(rcx, rbx); // check if bottom reached
3577 __ jcc(Assembler::notEqual, loop); // if not at bottom then check this entry
3578 __ bind(exit);
3579 }
3581 __ testptr(rdx, rdx); // check if a slot has been found
3582 __ jccb(Assembler::notZero, allocated); // if found, continue with that one
3584 // allocate one if there's no free slot
3585 { Label entry, loop;
3586 // 1. compute new pointers // rsp: old expression stack top
3587 __ movptr(rdx, monitor_block_bot); // rdx: old expression stack bottom
3588 __ subptr(rsp, entry_size); // move expression stack top
3589 __ subptr(rdx, entry_size); // move expression stack bottom
3590 __ mov(rcx, rsp); // set start value for copy loop
3591 __ movptr(monitor_block_bot, rdx); // set new monitor block top
3592 __ jmp(entry);
3593 // 2. move expression stack contents
3594 __ bind(loop);
3595 __ movptr(rbx, Address(rcx, entry_size)); // load expression stack word from old location
3596 __ movptr(Address(rcx, 0), rbx); // and store it at new location
3597 __ addptr(rcx, wordSize); // advance to next word
3598 __ bind(entry);
3599 __ cmpptr(rcx, rdx); // check if bottom reached
3600 __ jcc(Assembler::notEqual, loop); // if not at bottom then copy next word
3601 }
3603 // call run-time routine
3604 // rdx: points to monitor entry
3605 __ bind(allocated);
3607 // Increment bcp to point to the next bytecode, so exception handling for async. exceptions work correctly.
3608 // The object has already been poped from the stack, so the expression stack looks correct.
3609 __ increment(rsi);
3611 __ movptr(Address(rdx, BasicObjectLock::obj_offset_in_bytes()), rax); // store object
3612 __ lock_object(rdx);
3614 // check to make sure this monitor doesn't cause stack overflow after locking
3615 __ save_bcp(); // in case of exception
3616 __ generate_stack_overflow_check(0);
3618 // The bcp has already been incremented. Just need to dispatch to next instruction.
3619 __ dispatch_next(vtos);
3620 }
3623 void TemplateTable::monitorexit() {
3624 transition(atos, vtos);
3626 // check for NULL object
3627 __ null_check(rax);
3629 const Address monitor_block_top(rbp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
3630 const Address monitor_block_bot(rbp, frame::interpreter_frame_initial_sp_offset * wordSize);
3631 const int entry_size = ( frame::interpreter_frame_monitor_size() * wordSize);
3632 Label found;
3634 // find matching slot
3635 { Label entry, loop;
3636 __ movptr(rdx, monitor_block_top); // points to current entry, starting with top-most entry
3637 __ lea(rbx, monitor_block_bot); // points to word before bottom of monitor block
3638 __ jmpb(entry);
3640 __ bind(loop);
3641 __ cmpptr(rax, Address(rdx, BasicObjectLock::obj_offset_in_bytes())); // check if current entry is for same object
3642 __ jcc(Assembler::equal, found); // if same object then stop searching
3643 __ addptr(rdx, entry_size); // otherwise advance to next entry
3644 __ bind(entry);
3645 __ cmpptr(rdx, rbx); // check if bottom reached
3646 __ jcc(Assembler::notEqual, loop); // if not at bottom then check this entry
3647 }
3649 // error handling. Unlocking was not block-structured
3650 Label end;
3651 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_illegal_monitor_state_exception));
3652 __ should_not_reach_here();
3654 // call run-time routine
3655 // rcx: points to monitor entry
3656 __ bind(found);
3657 __ push_ptr(rax); // make sure object is on stack (contract with oopMaps)
3658 __ unlock_object(rdx);
3659 __ pop_ptr(rax); // discard object
3660 __ bind(end);
3661 }
3664 //----------------------------------------------------------------------------------------------------
3665 // Wide instructions
3667 void TemplateTable::wide() {
3668 transition(vtos, vtos);
3669 __ load_unsigned_byte(rbx, at_bcp(1));
3670 ExternalAddress wtable((address)Interpreter::_wentry_point);
3671 __ jump(ArrayAddress(wtable, Address(noreg, rbx, Address::times_ptr)));
3672 // Note: the rsi increment step is part of the individual wide bytecode implementations
3673 }
3676 //----------------------------------------------------------------------------------------------------
3677 // Multi arrays
3679 void TemplateTable::multianewarray() {
3680 transition(vtos, atos);
3681 __ load_unsigned_byte(rax, at_bcp(3)); // get number of dimensions
3682 // last dim is on top of stack; we want address of first one:
3683 // first_addr = last_addr + (ndims - 1) * stackElementSize - 1*wordsize
3684 // the latter wordSize to point to the beginning of the array.
3685 __ lea( rax, Address(rsp, rax, Interpreter::stackElementScale(), -wordSize));
3686 call_VM(rax, CAST_FROM_FN_PTR(address, InterpreterRuntime::multianewarray), rax); // pass in rax,
3687 __ load_unsigned_byte(rbx, at_bcp(3));
3688 __ lea(rsp, Address(rsp, rbx, Interpreter::stackElementScale())); // get rid of counts
3689 }
3691 #endif /* !CC_INTERP */
