1 /*
2 * Copyright (c) 1997, 2015, 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 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "asm/assembler.inline.hpp"
28 #include "compiler/disassembler.hpp"
29 #include "gc_interface/collectedHeap.inline.hpp"
30 #include "interpreter/interpreter.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "memory/universe.hpp"
34 #include "prims/methodHandles.hpp"
35 #include "runtime/biasedLocking.hpp"
36 #include "runtime/interfaceSupport.hpp"
37 #include "runtime/objectMonitor.hpp"
38 #include "runtime/os.hpp"
39 #include "runtime/sharedRuntime.hpp"
40 #include "runtime/stubRoutines.hpp"
41 #include "utilities/macros.hpp"
42 #if INCLUDE_ALL_GCS
43 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
44 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
45 #include "gc_implementation/g1/heapRegion.hpp"
46 #endif // INCLUDE_ALL_GCS
47
48 #ifdef PRODUCT
49 #define BLOCK_COMMENT(str) /* nothing */
50 #define STOP(error) stop(error)
51 #else
52 #define BLOCK_COMMENT(str) block_comment(str)
53 #define STOP(error) block_comment(error); stop(error)
54 #endif
55
56 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
57
58 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
59
60 #ifdef ASSERT
61 bool AbstractAssembler::pd_check_instruction_mark() { return true; }
62 #endif
63
64 static Assembler::Condition reverse[] = {
65 Assembler::noOverflow /* overflow = 0x0 */ ,
66 Assembler::overflow /* noOverflow = 0x1 */ ,
67 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ ,
68 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ ,
69 Assembler::notZero /* zero = 0x4, equal = 0x4 */ ,
70 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ ,
71 Assembler::above /* belowEqual = 0x6 */ ,
72 Assembler::belowEqual /* above = 0x7 */ ,
73 Assembler::positive /* negative = 0x8 */ ,
74 Assembler::negative /* positive = 0x9 */ ,
75 Assembler::noParity /* parity = 0xa */ ,
76 Assembler::parity /* noParity = 0xb */ ,
77 Assembler::greaterEqual /* less = 0xc */ ,
78 Assembler::less /* greaterEqual = 0xd */ ,
79 Assembler::greater /* lessEqual = 0xe */ ,
80 Assembler::lessEqual /* greater = 0xf, */
81
82 };
83
84
85 // Implementation of MacroAssembler
86
87 // First all the versions that have distinct versions depending on 32/64 bit
88 // Unless the difference is trivial (1 line or so).
89
90 #ifndef _LP64
91
92 // 32bit versions
93
94 Address MacroAssembler::as_Address(AddressLiteral adr) {
95 return Address(adr.target(), adr.rspec());
96 }
97
98 Address MacroAssembler::as_Address(ArrayAddress adr) {
99 return Address::make_array(adr);
100 }
101
102 void MacroAssembler::call_VM_leaf_base(address entry_point,
103 int number_of_arguments) {
104 call(RuntimeAddress(entry_point));
105 increment(rsp, number_of_arguments * wordSize);
106 }
107
108 void MacroAssembler::cmpklass(Address src1, Metadata* obj) {
109 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
110 }
111
112 void MacroAssembler::cmpklass(Register src1, Metadata* obj) {
113 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
114 }
115
116 void MacroAssembler::cmpoop(Address src1, jobject obj) {
117 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
118 }
119
120 void MacroAssembler::cmpoop(Register src1, jobject obj) {
121 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
122 }
123
124 void MacroAssembler::extend_sign(Register hi, Register lo) {
125 // According to Intel Doc. AP-526, "Integer Divide", p.18.
126 if (VM_Version::is_P6() && hi == rdx && lo == rax) {
127 cdql();
128 } else {
129 movl(hi, lo);
130 sarl(hi, 31);
131 }
132 }
133
134 void MacroAssembler::jC2(Register tmp, Label& L) {
135 // set parity bit if FPU flag C2 is set (via rax)
136 save_rax(tmp);
137 fwait(); fnstsw_ax();
138 sahf();
139 restore_rax(tmp);
140 // branch
141 jcc(Assembler::parity, L);
142 }
143
144 void MacroAssembler::jnC2(Register tmp, Label& L) {
145 // set parity bit if FPU flag C2 is set (via rax)
146 save_rax(tmp);
147 fwait(); fnstsw_ax();
148 sahf();
149 restore_rax(tmp);
150 // branch
151 jcc(Assembler::noParity, L);
152 }
153
154 // 32bit can do a case table jump in one instruction but we no longer allow the base
155 // to be installed in the Address class
156 void MacroAssembler::jump(ArrayAddress entry) {
157 jmp(as_Address(entry));
158 }
159
160 // Note: y_lo will be destroyed
161 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
162 // Long compare for Java (semantics as described in JVM spec.)
163 Label high, low, done;
164
165 cmpl(x_hi, y_hi);
166 jcc(Assembler::less, low);
167 jcc(Assembler::greater, high);
168 // x_hi is the return register
169 xorl(x_hi, x_hi);
170 cmpl(x_lo, y_lo);
171 jcc(Assembler::below, low);
172 jcc(Assembler::equal, done);
173
174 bind(high);
175 xorl(x_hi, x_hi);
176 increment(x_hi);
177 jmp(done);
178
179 bind(low);
180 xorl(x_hi, x_hi);
181 decrementl(x_hi);
182
183 bind(done);
184 }
185
186 void MacroAssembler::lea(Register dst, AddressLiteral src) {
187 mov_literal32(dst, (int32_t)src.target(), src.rspec());
188 }
189
190 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
191 // leal(dst, as_Address(adr));
192 // see note in movl as to why we must use a move
193 mov_literal32(dst, (int32_t) adr.target(), adr.rspec());
194 }
195
196 void MacroAssembler::leave() {
197 mov(rsp, rbp);
198 pop(rbp);
199 }
200
201 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) {
202 // Multiplication of two Java long values stored on the stack
203 // as illustrated below. Result is in rdx:rax.
204 //
205 // rsp ---> [ ?? ] \ \
206 // .... | y_rsp_offset |
207 // [ y_lo ] / (in bytes) | x_rsp_offset
208 // [ y_hi ] | (in bytes)
209 // .... |
210 // [ x_lo ] /
211 // [ x_hi ]
212 // ....
213 //
214 // Basic idea: lo(result) = lo(x_lo * y_lo)
215 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
216 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset);
217 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset);
218 Label quick;
219 // load x_hi, y_hi and check if quick
220 // multiplication is possible
221 movl(rbx, x_hi);
222 movl(rcx, y_hi);
223 movl(rax, rbx);
224 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0
225 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply
226 // do full multiplication
227 // 1st step
228 mull(y_lo); // x_hi * y_lo
229 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx,
230 // 2nd step
231 movl(rax, x_lo);
232 mull(rcx); // x_lo * y_hi
233 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx,
234 // 3rd step
235 bind(quick); // note: rbx, = 0 if quick multiply!
236 movl(rax, x_lo);
237 mull(y_lo); // x_lo * y_lo
238 addl(rdx, rbx); // correct hi(x_lo * y_lo)
239 }
240
241 void MacroAssembler::lneg(Register hi, Register lo) {
242 negl(lo);
243 adcl(hi, 0);
244 negl(hi);
245 }
246
247 void MacroAssembler::lshl(Register hi, Register lo) {
248 // Java shift left long support (semantics as described in JVM spec., p.305)
249 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n))
250 // shift value is in rcx !
251 assert(hi != rcx, "must not use rcx");
252 assert(lo != rcx, "must not use rcx");
253 const Register s = rcx; // shift count
254 const int n = BitsPerWord;
255 Label L;
256 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
257 cmpl(s, n); // if (s < n)
258 jcc(Assembler::less, L); // else (s >= n)
259 movl(hi, lo); // x := x << n
260 xorl(lo, lo);
261 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
262 bind(L); // s (mod n) < n
263 shldl(hi, lo); // x := x << s
264 shll(lo);
265 }
266
267
268 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) {
269 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310)
270 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n))
271 assert(hi != rcx, "must not use rcx");
272 assert(lo != rcx, "must not use rcx");
273 const Register s = rcx; // shift count
274 const int n = BitsPerWord;
275 Label L;
276 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
277 cmpl(s, n); // if (s < n)
278 jcc(Assembler::less, L); // else (s >= n)
279 movl(lo, hi); // x := x >> n
280 if (sign_extension) sarl(hi, 31);
281 else xorl(hi, hi);
282 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
283 bind(L); // s (mod n) < n
284 shrdl(lo, hi); // x := x >> s
285 if (sign_extension) sarl(hi);
286 else shrl(hi);
287 }
288
289 void MacroAssembler::movoop(Register dst, jobject obj) {
290 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
291 }
292
293 void MacroAssembler::movoop(Address dst, jobject obj) {
294 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
295 }
296
297 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
298 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
299 }
300
301 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
302 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
303 }
304
305 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
306 // scratch register is not used,
307 // it is defined to match parameters of 64-bit version of this method.
308 if (src.is_lval()) {
309 mov_literal32(dst, (intptr_t)src.target(), src.rspec());
310 } else {
311 movl(dst, as_Address(src));
312 }
313 }
314
315 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
316 movl(as_Address(dst), src);
317 }
318
319 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
320 movl(dst, as_Address(src));
321 }
322
323 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
324 void MacroAssembler::movptr(Address dst, intptr_t src) {
325 movl(dst, src);
326 }
327
328
329 void MacroAssembler::pop_callee_saved_registers() {
330 pop(rcx);
331 pop(rdx);
332 pop(rdi);
333 pop(rsi);
334 }
335
336 void MacroAssembler::pop_fTOS() {
337 fld_d(Address(rsp, 0));
338 addl(rsp, 2 * wordSize);
339 }
340
341 void MacroAssembler::push_callee_saved_registers() {
342 push(rsi);
343 push(rdi);
344 push(rdx);
345 push(rcx);
346 }
347
348 void MacroAssembler::push_fTOS() {
349 subl(rsp, 2 * wordSize);
350 fstp_d(Address(rsp, 0));
351 }
352
353
354 void MacroAssembler::pushoop(jobject obj) {
355 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate());
356 }
357
358 void MacroAssembler::pushklass(Metadata* obj) {
359 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate());
360 }
361
362 void MacroAssembler::pushptr(AddressLiteral src) {
363 if (src.is_lval()) {
364 push_literal32((int32_t)src.target(), src.rspec());
365 } else {
366 pushl(as_Address(src));
367 }
368 }
369
370 void MacroAssembler::set_word_if_not_zero(Register dst) {
371 xorl(dst, dst);
372 set_byte_if_not_zero(dst);
373 }
374
375 static void pass_arg0(MacroAssembler* masm, Register arg) {
376 masm->push(arg);
377 }
378
379 static void pass_arg1(MacroAssembler* masm, Register arg) {
380 masm->push(arg);
381 }
382
383 static void pass_arg2(MacroAssembler* masm, Register arg) {
384 masm->push(arg);
385 }
386
387 static void pass_arg3(MacroAssembler* masm, Register arg) {
388 masm->push(arg);
389 }
390
391 #ifndef PRODUCT
392 extern "C" void findpc(intptr_t x);
393 #endif
394
395 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) {
396 // In order to get locks to work, we need to fake a in_VM state
397 JavaThread* thread = JavaThread::current();
398 JavaThreadState saved_state = thread->thread_state();
399 thread->set_thread_state(_thread_in_vm);
400 if (ShowMessageBoxOnError) {
401 JavaThread* thread = JavaThread::current();
402 JavaThreadState saved_state = thread->thread_state();
403 thread->set_thread_state(_thread_in_vm);
404 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
405 ttyLocker ttyl;
406 BytecodeCounter::print();
407 }
408 // To see where a verify_oop failed, get $ebx+40/X for this frame.
409 // This is the value of eip which points to where verify_oop will return.
410 if (os::message_box(msg, "Execution stopped, print registers?")) {
411 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip);
412 BREAKPOINT;
413 }
414 } else {
415 ttyLocker ttyl;
416 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg);
417 }
418 // Don't assert holding the ttyLock
419 assert(false, err_msg("DEBUG MESSAGE: %s", msg));
420 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
421 }
422
423 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) {
424 ttyLocker ttyl;
425 FlagSetting fs(Debugging, true);
426 tty->print_cr("eip = 0x%08x", eip);
427 #ifndef PRODUCT
428 if ((WizardMode || Verbose) && PrintMiscellaneous) {
429 tty->cr();
430 findpc(eip);
431 tty->cr();
432 }
433 #endif
434 #define PRINT_REG(rax) \
435 { tty->print("%s = ", #rax); os::print_location(tty, rax); }
436 PRINT_REG(rax);
437 PRINT_REG(rbx);
438 PRINT_REG(rcx);
439 PRINT_REG(rdx);
440 PRINT_REG(rdi);
441 PRINT_REG(rsi);
442 PRINT_REG(rbp);
443 PRINT_REG(rsp);
444 #undef PRINT_REG
445 // Print some words near top of staack.
446 int* dump_sp = (int*) rsp;
447 for (int col1 = 0; col1 < 8; col1++) {
448 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
449 os::print_location(tty, *dump_sp++);
450 }
451 for (int row = 0; row < 16; row++) {
452 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
453 for (int col = 0; col < 8; col++) {
454 tty->print(" 0x%08x", *dump_sp++);
455 }
456 tty->cr();
457 }
458 // Print some instructions around pc:
459 Disassembler::decode((address)eip-64, (address)eip);
460 tty->print_cr("--------");
461 Disassembler::decode((address)eip, (address)eip+32);
462 }
463
464 void MacroAssembler::stop(const char* msg) {
465 ExternalAddress message((address)msg);
466 // push address of message
467 pushptr(message.addr());
468 { Label L; call(L, relocInfo::none); bind(L); } // push eip
469 pusha(); // push registers
470 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32)));
471 hlt();
472 }
473
474 void MacroAssembler::warn(const char* msg) {
475 push_CPU_state();
476
477 ExternalAddress message((address) msg);
478 // push address of message
479 pushptr(message.addr());
480
481 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning)));
482 addl(rsp, wordSize); // discard argument
483 pop_CPU_state();
484 }
485
486 void MacroAssembler::print_state() {
487 { Label L; call(L, relocInfo::none); bind(L); } // push eip
488 pusha(); // push registers
489
490 push_CPU_state();
491 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32)));
492 pop_CPU_state();
493
494 popa();
495 addl(rsp, wordSize);
496 }
497
498 #else // _LP64
499
500 // 64 bit versions
501
502 Address MacroAssembler::as_Address(AddressLiteral adr) {
503 // amd64 always does this as a pc-rel
504 // we can be absolute or disp based on the instruction type
505 // jmp/call are displacements others are absolute
506 assert(!adr.is_lval(), "must be rval");
507 assert(reachable(adr), "must be");
508 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc());
509
510 }
511
512 Address MacroAssembler::as_Address(ArrayAddress adr) {
513 AddressLiteral base = adr.base();
514 lea(rscratch1, base);
515 Address index = adr.index();
516 assert(index._disp == 0, "must not have disp"); // maybe it can?
517 Address array(rscratch1, index._index, index._scale, index._disp);
518 return array;
519 }
520
521 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) {
522 Label L, E;
523
524 #ifdef _WIN64
525 // Windows always allocates space for it's register args
526 assert(num_args <= 4, "only register arguments supported");
527 subq(rsp, frame::arg_reg_save_area_bytes);
528 #endif
529
530 // Align stack if necessary
531 testl(rsp, 15);
532 jcc(Assembler::zero, L);
533
534 subq(rsp, 8);
535 {
536 call(RuntimeAddress(entry_point));
537 }
538 addq(rsp, 8);
539 jmp(E);
540
541 bind(L);
542 {
543 call(RuntimeAddress(entry_point));
544 }
545
546 bind(E);
547
548 #ifdef _WIN64
549 // restore stack pointer
550 addq(rsp, frame::arg_reg_save_area_bytes);
551 #endif
552
553 }
554
555 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) {
556 assert(!src2.is_lval(), "should use cmpptr");
557
558 if (reachable(src2)) {
559 cmpq(src1, as_Address(src2));
560 } else {
561 lea(rscratch1, src2);
562 Assembler::cmpq(src1, Address(rscratch1, 0));
563 }
564 }
565
566 int MacroAssembler::corrected_idivq(Register reg) {
567 // Full implementation of Java ldiv and lrem; checks for special
568 // case as described in JVM spec., p.243 & p.271. The function
569 // returns the (pc) offset of the idivl instruction - may be needed
570 // for implicit exceptions.
571 //
572 // normal case special case
573 //
574 // input : rax: dividend min_long
575 // reg: divisor (may not be eax/edx) -1
576 //
577 // output: rax: quotient (= rax idiv reg) min_long
578 // rdx: remainder (= rax irem reg) 0
579 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register");
580 static const int64_t min_long = 0x8000000000000000;
581 Label normal_case, special_case;
582
583 // check for special case
584 cmp64(rax, ExternalAddress((address) &min_long));
585 jcc(Assembler::notEqual, normal_case);
586 xorl(rdx, rdx); // prepare rdx for possible special case (where
587 // remainder = 0)
588 cmpq(reg, -1);
589 jcc(Assembler::equal, special_case);
590
591 // handle normal case
592 bind(normal_case);
593 cdqq();
594 int idivq_offset = offset();
595 idivq(reg);
596
597 // normal and special case exit
598 bind(special_case);
599
600 return idivq_offset;
601 }
602
603 void MacroAssembler::decrementq(Register reg, int value) {
604 if (value == min_jint) { subq(reg, value); return; }
605 if (value < 0) { incrementq(reg, -value); return; }
606 if (value == 0) { ; return; }
607 if (value == 1 && UseIncDec) { decq(reg) ; return; }
608 /* else */ { subq(reg, value) ; return; }
609 }
610
611 void MacroAssembler::decrementq(Address dst, int value) {
612 if (value == min_jint) { subq(dst, value); return; }
613 if (value < 0) { incrementq(dst, -value); return; }
614 if (value == 0) { ; return; }
615 if (value == 1 && UseIncDec) { decq(dst) ; return; }
616 /* else */ { subq(dst, value) ; return; }
617 }
618
619 void MacroAssembler::incrementq(AddressLiteral dst) {
620 if (reachable(dst)) {
621 incrementq(as_Address(dst));
622 } else {
623 lea(rscratch1, dst);
624 incrementq(Address(rscratch1, 0));
625 }
626 }
627
628 void MacroAssembler::incrementq(Register reg, int value) {
629 if (value == min_jint) { addq(reg, value); return; }
630 if (value < 0) { decrementq(reg, -value); return; }
631 if (value == 0) { ; return; }
632 if (value == 1 && UseIncDec) { incq(reg) ; return; }
633 /* else */ { addq(reg, value) ; return; }
634 }
635
636 void MacroAssembler::incrementq(Address dst, int value) {
637 if (value == min_jint) { addq(dst, value); return; }
638 if (value < 0) { decrementq(dst, -value); return; }
639 if (value == 0) { ; return; }
640 if (value == 1 && UseIncDec) { incq(dst) ; return; }
641 /* else */ { addq(dst, value) ; return; }
642 }
643
644 // 32bit can do a case table jump in one instruction but we no longer allow the base
645 // to be installed in the Address class
646 void MacroAssembler::jump(ArrayAddress entry) {
647 lea(rscratch1, entry.base());
648 Address dispatch = entry.index();
649 assert(dispatch._base == noreg, "must be");
650 dispatch._base = rscratch1;
651 jmp(dispatch);
652 }
653
654 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
655 ShouldNotReachHere(); // 64bit doesn't use two regs
656 cmpq(x_lo, y_lo);
657 }
658
659 void MacroAssembler::lea(Register dst, AddressLiteral src) {
660 mov_literal64(dst, (intptr_t)src.target(), src.rspec());
661 }
662
663 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
664 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec());
665 movptr(dst, rscratch1);
666 }
667
668 void MacroAssembler::leave() {
669 // %%% is this really better? Why not on 32bit too?
670 emit_int8((unsigned char)0xC9); // LEAVE
671 }
672
673 void MacroAssembler::lneg(Register hi, Register lo) {
674 ShouldNotReachHere(); // 64bit doesn't use two regs
675 negq(lo);
676 }
677
678 void MacroAssembler::movoop(Register dst, jobject obj) {
679 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate());
680 }
681
682 void MacroAssembler::movoop(Address dst, jobject obj) {
683 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate());
684 movq(dst, rscratch1);
685 }
686
687 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
688 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
689 }
690
691 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
692 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
693 movq(dst, rscratch1);
694 }
695
696 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
697 if (src.is_lval()) {
698 mov_literal64(dst, (intptr_t)src.target(), src.rspec());
699 } else {
700 if (reachable(src)) {
701 movq(dst, as_Address(src));
702 } else {
703 lea(scratch, src);
704 movq(dst, Address(scratch, 0));
705 }
706 }
707 }
708
709 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
710 movq(as_Address(dst), src);
711 }
712
713 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
714 movq(dst, as_Address(src));
715 }
716
717 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
718 void MacroAssembler::movptr(Address dst, intptr_t src) {
719 mov64(rscratch1, src);
720 movq(dst, rscratch1);
721 }
722
723 // These are mostly for initializing NULL
724 void MacroAssembler::movptr(Address dst, int32_t src) {
725 movslq(dst, src);
726 }
727
728 void MacroAssembler::movptr(Register dst, int32_t src) {
729 mov64(dst, (intptr_t)src);
730 }
731
732 void MacroAssembler::pushoop(jobject obj) {
733 movoop(rscratch1, obj);
734 push(rscratch1);
735 }
736
737 void MacroAssembler::pushklass(Metadata* obj) {
738 mov_metadata(rscratch1, obj);
739 push(rscratch1);
740 }
741
742 void MacroAssembler::pushptr(AddressLiteral src) {
743 lea(rscratch1, src);
744 if (src.is_lval()) {
745 push(rscratch1);
746 } else {
747 pushq(Address(rscratch1, 0));
748 }
749 }
750
751 void MacroAssembler::reset_last_Java_frame(bool clear_fp,
752 bool clear_pc) {
753 // we must set sp to zero to clear frame
754 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
755 // must clear fp, so that compiled frames are not confused; it is
756 // possible that we need it only for debugging
757 if (clear_fp) {
758 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
759 }
760
761 if (clear_pc) {
762 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
763 }
764 }
765
766 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
767 Register last_java_fp,
768 address last_java_pc) {
769 // determine last_java_sp register
770 if (!last_java_sp->is_valid()) {
771 last_java_sp = rsp;
772 }
773
774 // last_java_fp is optional
775 if (last_java_fp->is_valid()) {
776 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()),
777 last_java_fp);
778 }
779
780 // last_java_pc is optional
781 if (last_java_pc != NULL) {
782 Address java_pc(r15_thread,
783 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset());
784 lea(rscratch1, InternalAddress(last_java_pc));
785 movptr(java_pc, rscratch1);
786 }
787
788 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
789 }
790
791 static void pass_arg0(MacroAssembler* masm, Register arg) {
792 if (c_rarg0 != arg ) {
793 masm->mov(c_rarg0, arg);
794 }
795 }
796
797 static void pass_arg1(MacroAssembler* masm, Register arg) {
798 if (c_rarg1 != arg ) {
799 masm->mov(c_rarg1, arg);
800 }
801 }
802
803 static void pass_arg2(MacroAssembler* masm, Register arg) {
804 if (c_rarg2 != arg ) {
805 masm->mov(c_rarg2, arg);
806 }
807 }
808
809 static void pass_arg3(MacroAssembler* masm, Register arg) {
810 if (c_rarg3 != arg ) {
811 masm->mov(c_rarg3, arg);
812 }
813 }
814
815 void MacroAssembler::stop(const char* msg) {
816 address rip = pc();
817 pusha(); // get regs on stack
818 lea(c_rarg0, ExternalAddress((address) msg));
819 lea(c_rarg1, InternalAddress(rip));
820 movq(c_rarg2, rsp); // pass pointer to regs array
821 andq(rsp, -16); // align stack as required by ABI
822 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
823 hlt();
824 }
825
826 void MacroAssembler::warn(const char* msg) {
827 push(rbp);
828 movq(rbp, rsp);
829 andq(rsp, -16); // align stack as required by push_CPU_state and call
830 push_CPU_state(); // keeps alignment at 16 bytes
831 lea(c_rarg0, ExternalAddress((address) msg));
832 call_VM_leaf(CAST_FROM_FN_PTR(address, warning), c_rarg0);
833 pop_CPU_state();
834 mov(rsp, rbp);
835 pop(rbp);
836 }
837
838 void MacroAssembler::print_state() {
839 address rip = pc();
840 pusha(); // get regs on stack
841 push(rbp);
842 movq(rbp, rsp);
843 andq(rsp, -16); // align stack as required by push_CPU_state and call
844 push_CPU_state(); // keeps alignment at 16 bytes
845
846 lea(c_rarg0, InternalAddress(rip));
847 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array
848 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1);
849
850 pop_CPU_state();
851 mov(rsp, rbp);
852 pop(rbp);
853 popa();
854 }
855
856 #ifndef PRODUCT
857 extern "C" void findpc(intptr_t x);
858 #endif
859
860 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) {
861 // In order to get locks to work, we need to fake a in_VM state
862 if (ShowMessageBoxOnError) {
863 JavaThread* thread = JavaThread::current();
864 JavaThreadState saved_state = thread->thread_state();
865 thread->set_thread_state(_thread_in_vm);
866 #ifndef PRODUCT
867 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
868 ttyLocker ttyl;
869 BytecodeCounter::print();
870 }
871 #endif
872 // To see where a verify_oop failed, get $ebx+40/X for this frame.
873 // XXX correct this offset for amd64
874 // This is the value of eip which points to where verify_oop will return.
875 if (os::message_box(msg, "Execution stopped, print registers?")) {
876 print_state64(pc, regs);
877 BREAKPOINT;
878 assert(false, "start up GDB");
879 }
880 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
881 } else {
882 ttyLocker ttyl;
883 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n",
884 msg);
885 assert(false, err_msg("DEBUG MESSAGE: %s", msg));
886 }
887 }
888
889 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) {
890 ttyLocker ttyl;
891 FlagSetting fs(Debugging, true);
892 tty->print_cr("rip = 0x%016lx", pc);
893 #ifndef PRODUCT
894 tty->cr();
895 findpc(pc);
896 tty->cr();
897 #endif
898 #define PRINT_REG(rax, value) \
899 { tty->print("%s = ", #rax); os::print_location(tty, value); }
900 PRINT_REG(rax, regs[15]);
901 PRINT_REG(rbx, regs[12]);
902 PRINT_REG(rcx, regs[14]);
903 PRINT_REG(rdx, regs[13]);
904 PRINT_REG(rdi, regs[8]);
905 PRINT_REG(rsi, regs[9]);
906 PRINT_REG(rbp, regs[10]);
907 PRINT_REG(rsp, regs[11]);
908 PRINT_REG(r8 , regs[7]);
909 PRINT_REG(r9 , regs[6]);
910 PRINT_REG(r10, regs[5]);
911 PRINT_REG(r11, regs[4]);
912 PRINT_REG(r12, regs[3]);
913 PRINT_REG(r13, regs[2]);
914 PRINT_REG(r14, regs[1]);
915 PRINT_REG(r15, regs[0]);
916 #undef PRINT_REG
917 // Print some words near top of staack.
918 int64_t* rsp = (int64_t*) regs[11];
919 int64_t* dump_sp = rsp;
920 for (int col1 = 0; col1 < 8; col1++) {
921 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp);
922 os::print_location(tty, *dump_sp++);
923 }
924 for (int row = 0; row < 25; row++) {
925 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp);
926 for (int col = 0; col < 4; col++) {
927 tty->print(" 0x%016lx", *dump_sp++);
928 }
929 tty->cr();
930 }
931 // Print some instructions around pc:
932 Disassembler::decode((address)pc-64, (address)pc);
933 tty->print_cr("--------");
934 Disassembler::decode((address)pc, (address)pc+32);
935 }
936
937 #endif // _LP64
938
939 // Now versions that are common to 32/64 bit
940
941 void MacroAssembler::addptr(Register dst, int32_t imm32) {
942 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32));
943 }
944
945 void MacroAssembler::addptr(Register dst, Register src) {
946 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
947 }
948
949 void MacroAssembler::addptr(Address dst, Register src) {
950 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
951 }
952
953 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) {
954 if (reachable(src)) {
955 Assembler::addsd(dst, as_Address(src));
956 } else {
957 lea(rscratch1, src);
958 Assembler::addsd(dst, Address(rscratch1, 0));
959 }
960 }
961
962 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) {
963 if (reachable(src)) {
964 addss(dst, as_Address(src));
965 } else {
966 lea(rscratch1, src);
967 addss(dst, Address(rscratch1, 0));
968 }
969 }
970
971 void MacroAssembler::align(int modulus) {
972 if (offset() % modulus != 0) {
973 nop(modulus - (offset() % modulus));
974 }
975 }
976
977 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) {
978 // Used in sign-masking with aligned address.
979 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
980 if (reachable(src)) {
981 Assembler::andpd(dst, as_Address(src));
982 } else {
983 lea(rscratch1, src);
984 Assembler::andpd(dst, Address(rscratch1, 0));
985 }
986 }
987
988 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) {
989 // Used in sign-masking with aligned address.
990 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
991 if (reachable(src)) {
992 Assembler::andps(dst, as_Address(src));
993 } else {
994 lea(rscratch1, src);
995 Assembler::andps(dst, Address(rscratch1, 0));
996 }
997 }
998
999 void MacroAssembler::andptr(Register dst, int32_t imm32) {
1000 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32));
1001 }
1002
1003 void MacroAssembler::atomic_incl(Address counter_addr) {
1004 if (os::is_MP())
1005 lock();
1006 incrementl(counter_addr);
1007 }
1008
1009 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) {
1010 if (reachable(counter_addr)) {
1011 atomic_incl(as_Address(counter_addr));
1012 } else {
1013 lea(scr, counter_addr);
1014 atomic_incl(Address(scr, 0));
1015 }
1016 }
1017
1018 #ifdef _LP64
1019 void MacroAssembler::atomic_incq(Address counter_addr) {
1020 if (os::is_MP())
1021 lock();
1022 incrementq(counter_addr);
1023 }
1024
1025 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) {
1026 if (reachable(counter_addr)) {
1027 atomic_incq(as_Address(counter_addr));
1028 } else {
1029 lea(scr, counter_addr);
1030 atomic_incq(Address(scr, 0));
1031 }
1032 }
1033 #endif
1034
1035 // Writes to stack successive pages until offset reached to check for
1036 // stack overflow + shadow pages. This clobbers tmp.
1037 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
1038 movptr(tmp, rsp);
1039 // Bang stack for total size given plus shadow page size.
1040 // Bang one page at a time because large size can bang beyond yellow and
1041 // red zones.
1042 Label loop;
1043 bind(loop);
1044 movl(Address(tmp, (-os::vm_page_size())), size );
1045 subptr(tmp, os::vm_page_size());
1046 subl(size, os::vm_page_size());
1047 jcc(Assembler::greater, loop);
1048
1049 // Bang down shadow pages too.
1050 // At this point, (tmp-0) is the last address touched, so don't
1051 // touch it again. (It was touched as (tmp-pagesize) but then tmp
1052 // was post-decremented.) Skip this address by starting at i=1, and
1053 // touch a few more pages below. N.B. It is important to touch all
1054 // the way down to and including i=StackShadowPages.
1055 for (int i = 1; i < StackShadowPages; i++) {
1056 // this could be any sized move but this is can be a debugging crumb
1057 // so the bigger the better.
1058 movptr(Address(tmp, (-i*os::vm_page_size())), size );
1059 }
1060 }
1061
1062 int MacroAssembler::biased_locking_enter(Register lock_reg,
1063 Register obj_reg,
1064 Register swap_reg,
1065 Register tmp_reg,
1066 bool swap_reg_contains_mark,
1067 Label& done,
1068 Label* slow_case,
1069 BiasedLockingCounters* counters) {
1070 assert(UseBiasedLocking, "why call this otherwise?");
1071 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq");
1072 LP64_ONLY( assert(tmp_reg != noreg, "tmp_reg must be supplied"); )
1073 bool need_tmp_reg = false;
1074 if (tmp_reg == noreg) {
1075 need_tmp_reg = true;
1076 tmp_reg = lock_reg;
1077 assert_different_registers(lock_reg, obj_reg, swap_reg);
1078 } else {
1079 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg);
1080 }
1081 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout");
1082 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes());
1083 Address saved_mark_addr(lock_reg, 0);
1084
1085 if (PrintBiasedLockingStatistics && counters == NULL) {
1086 counters = BiasedLocking::counters();
1087 }
1088 // Biased locking
1089 // See whether the lock is currently biased toward our thread and
1090 // whether the epoch is still valid
1091 // Note that the runtime guarantees sufficient alignment of JavaThread
1092 // pointers to allow age to be placed into low bits
1093 // First check to see whether biasing is even enabled for this object
1094 Label cas_label;
1095 int null_check_offset = -1;
1096 if (!swap_reg_contains_mark) {
1097 null_check_offset = offset();
1098 movptr(swap_reg, mark_addr);
1099 }
1100 if (need_tmp_reg) {
1101 push(tmp_reg);
1102 }
1103 movptr(tmp_reg, swap_reg);
1104 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place);
1105 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern);
1106 if (need_tmp_reg) {
1107 pop(tmp_reg);
1108 }
1109 jcc(Assembler::notEqual, cas_label);
1110 // The bias pattern is present in the object's header. Need to check
1111 // whether the bias owner and the epoch are both still current.
1112 #ifndef _LP64
1113 // Note that because there is no current thread register on x86_32 we
1114 // need to store off the mark word we read out of the object to
1115 // avoid reloading it and needing to recheck invariants below. This
1116 // store is unfortunate but it makes the overall code shorter and
1117 // simpler.
1118 movptr(saved_mark_addr, swap_reg);
1119 #endif
1120 if (need_tmp_reg) {
1121 push(tmp_reg);
1122 }
1123 if (swap_reg_contains_mark) {
1124 null_check_offset = offset();
1125 }
1126 load_prototype_header(tmp_reg, obj_reg);
1127 #ifdef _LP64
1128 orptr(tmp_reg, r15_thread);
1129 xorptr(tmp_reg, swap_reg);
1130 Register header_reg = tmp_reg;
1131 #else
1132 xorptr(tmp_reg, swap_reg);
1133 get_thread(swap_reg);
1134 xorptr(swap_reg, tmp_reg);
1135 Register header_reg = swap_reg;
1136 #endif
1137 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place));
1138 if (need_tmp_reg) {
1139 pop(tmp_reg);
1140 }
1141 if (counters != NULL) {
1142 cond_inc32(Assembler::zero,
1143 ExternalAddress((address) counters->biased_lock_entry_count_addr()));
1144 }
1145 jcc(Assembler::equal, done);
1146
1147 Label try_revoke_bias;
1148 Label try_rebias;
1149
1150 // At this point we know that the header has the bias pattern and
1151 // that we are not the bias owner in the current epoch. We need to
1152 // figure out more details about the state of the header in order to
1153 // know what operations can be legally performed on the object's
1154 // header.
1155
1156 // If the low three bits in the xor result aren't clear, that means
1157 // the prototype header is no longer biased and we have to revoke
1158 // the bias on this object.
1159 testptr(header_reg, markOopDesc::biased_lock_mask_in_place);
1160 jccb(Assembler::notZero, try_revoke_bias);
1161
1162 // Biasing is still enabled for this data type. See whether the
1163 // epoch of the current bias is still valid, meaning that the epoch
1164 // bits of the mark word are equal to the epoch bits of the
1165 // prototype header. (Note that the prototype header's epoch bits
1166 // only change at a safepoint.) If not, attempt to rebias the object
1167 // toward the current thread. Note that we must be absolutely sure
1168 // that the current epoch is invalid in order to do this because
1169 // otherwise the manipulations it performs on the mark word are
1170 // illegal.
1171 testptr(header_reg, markOopDesc::epoch_mask_in_place);
1172 jccb(Assembler::notZero, try_rebias);
1173
1174 // The epoch of the current bias is still valid but we know nothing
1175 // about the owner; it might be set or it might be clear. Try to
1176 // acquire the bias of the object using an atomic operation. If this
1177 // fails we will go in to the runtime to revoke the object's bias.
1178 // Note that we first construct the presumed unbiased header so we
1179 // don't accidentally blow away another thread's valid bias.
1180 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1181 andptr(swap_reg,
1182 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place);
1183 if (need_tmp_reg) {
1184 push(tmp_reg);
1185 }
1186 #ifdef _LP64
1187 movptr(tmp_reg, swap_reg);
1188 orptr(tmp_reg, r15_thread);
1189 #else
1190 get_thread(tmp_reg);
1191 orptr(tmp_reg, swap_reg);
1192 #endif
1193 if (os::is_MP()) {
1194 lock();
1195 }
1196 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1197 if (need_tmp_reg) {
1198 pop(tmp_reg);
1199 }
1200 // If the biasing toward our thread failed, this means that
1201 // another thread succeeded in biasing it toward itself and we
1202 // need to revoke that bias. The revocation will occur in the
1203 // interpreter runtime in the slow case.
1204 if (counters != NULL) {
1205 cond_inc32(Assembler::zero,
1206 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr()));
1207 }
1208 if (slow_case != NULL) {
1209 jcc(Assembler::notZero, *slow_case);
1210 }
1211 jmp(done);
1212
1213 bind(try_rebias);
1214 // At this point we know the epoch has expired, meaning that the
1215 // current "bias owner", if any, is actually invalid. Under these
1216 // circumstances _only_, we are allowed to use the current header's
1217 // value as the comparison value when doing the cas to acquire the
1218 // bias in the current epoch. In other words, we allow transfer of
1219 // the bias from one thread to another directly in this situation.
1220 //
1221 // FIXME: due to a lack of registers we currently blow away the age
1222 // bits in this situation. Should attempt to preserve them.
1223 if (need_tmp_reg) {
1224 push(tmp_reg);
1225 }
1226 load_prototype_header(tmp_reg, obj_reg);
1227 #ifdef _LP64
1228 orptr(tmp_reg, r15_thread);
1229 #else
1230 get_thread(swap_reg);
1231 orptr(tmp_reg, swap_reg);
1232 movptr(swap_reg, saved_mark_addr);
1233 #endif
1234 if (os::is_MP()) {
1235 lock();
1236 }
1237 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1238 if (need_tmp_reg) {
1239 pop(tmp_reg);
1240 }
1241 // If the biasing toward our thread failed, then another thread
1242 // succeeded in biasing it toward itself and we need to revoke that
1243 // bias. The revocation will occur in the runtime in the slow case.
1244 if (counters != NULL) {
1245 cond_inc32(Assembler::zero,
1246 ExternalAddress((address) counters->rebiased_lock_entry_count_addr()));
1247 }
1248 if (slow_case != NULL) {
1249 jcc(Assembler::notZero, *slow_case);
1250 }
1251 jmp(done);
1252
1253 bind(try_revoke_bias);
1254 // The prototype mark in the klass doesn't have the bias bit set any
1255 // more, indicating that objects of this data type are not supposed
1256 // to be biased any more. We are going to try to reset the mark of
1257 // this object to the prototype value and fall through to the
1258 // CAS-based locking scheme. Note that if our CAS fails, it means
1259 // that another thread raced us for the privilege of revoking the
1260 // bias of this particular object, so it's okay to continue in the
1261 // normal locking code.
1262 //
1263 // FIXME: due to a lack of registers we currently blow away the age
1264 // bits in this situation. Should attempt to preserve them.
1265 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1266 if (need_tmp_reg) {
1267 push(tmp_reg);
1268 }
1269 load_prototype_header(tmp_reg, obj_reg);
1270 if (os::is_MP()) {
1271 lock();
1272 }
1273 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1274 if (need_tmp_reg) {
1275 pop(tmp_reg);
1276 }
1277 // Fall through to the normal CAS-based lock, because no matter what
1278 // the result of the above CAS, some thread must have succeeded in
1279 // removing the bias bit from the object's header.
1280 if (counters != NULL) {
1281 cond_inc32(Assembler::zero,
1282 ExternalAddress((address) counters->revoked_lock_entry_count_addr()));
1283 }
1284
1285 bind(cas_label);
1286
1287 return null_check_offset;
1288 }
1289
1290 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) {
1291 assert(UseBiasedLocking, "why call this otherwise?");
1292
1293 // Check for biased locking unlock case, which is a no-op
1294 // Note: we do not have to check the thread ID for two reasons.
1295 // First, the interpreter checks for IllegalMonitorStateException at
1296 // a higher level. Second, if the bias was revoked while we held the
1297 // lock, the object could not be rebiased toward another thread, so
1298 // the bias bit would be clear.
1299 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1300 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place);
1301 cmpptr(temp_reg, markOopDesc::biased_lock_pattern);
1302 jcc(Assembler::equal, done);
1303 }
1304
1305 #ifdef COMPILER2
1306
1307 #if INCLUDE_RTM_OPT
1308
1309 // Update rtm_counters based on abort status
1310 // input: abort_status
1311 // rtm_counters (RTMLockingCounters*)
1312 // flags are killed
1313 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) {
1314
1315 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset()));
1316 if (PrintPreciseRTMLockingStatistics) {
1317 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) {
1318 Label check_abort;
1319 testl(abort_status, (1<<i));
1320 jccb(Assembler::equal, check_abort);
1321 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx))));
1322 bind(check_abort);
1323 }
1324 }
1325 }
1326
1327 // Branch if (random & (count-1) != 0), count is 2^n
1328 // tmp, scr and flags are killed
1329 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) {
1330 assert(tmp == rax, "");
1331 assert(scr == rdx, "");
1332 rdtsc(); // modifies EDX:EAX
1333 andptr(tmp, count-1);
1334 jccb(Assembler::notZero, brLabel);
1335 }
1336
1337 // Perform abort ratio calculation, set no_rtm bit if high ratio
1338 // input: rtm_counters_Reg (RTMLockingCounters* address)
1339 // tmpReg, rtm_counters_Reg and flags are killed
1340 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg,
1341 Register rtm_counters_Reg,
1342 RTMLockingCounters* rtm_counters,
1343 Metadata* method_data) {
1344 Label L_done, L_check_always_rtm1, L_check_always_rtm2;
1345
1346 if (RTMLockingCalculationDelay > 0) {
1347 // Delay calculation
1348 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg);
1349 testptr(tmpReg, tmpReg);
1350 jccb(Assembler::equal, L_done);
1351 }
1352 // Abort ratio calculation only if abort_count > RTMAbortThreshold
1353 // Aborted transactions = abort_count * 100
1354 // All transactions = total_count * RTMTotalCountIncrRate
1355 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio)
1356
1357 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset()));
1358 cmpptr(tmpReg, RTMAbortThreshold);
1359 jccb(Assembler::below, L_check_always_rtm2);
1360 imulptr(tmpReg, tmpReg, 100);
1361
1362 Register scrReg = rtm_counters_Reg;
1363 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1364 imulptr(scrReg, scrReg, RTMTotalCountIncrRate);
1365 imulptr(scrReg, scrReg, RTMAbortRatio);
1366 cmpptr(tmpReg, scrReg);
1367 jccb(Assembler::below, L_check_always_rtm1);
1368 if (method_data != NULL) {
1369 // set rtm_state to "no rtm" in MDO
1370 mov_metadata(tmpReg, method_data);
1371 if (os::is_MP()) {
1372 lock();
1373 }
1374 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM);
1375 }
1376 jmpb(L_done);
1377 bind(L_check_always_rtm1);
1378 // Reload RTMLockingCounters* address
1379 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1380 bind(L_check_always_rtm2);
1381 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1382 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate);
1383 jccb(Assembler::below, L_done);
1384 if (method_data != NULL) {
1385 // set rtm_state to "always rtm" in MDO
1386 mov_metadata(tmpReg, method_data);
1387 if (os::is_MP()) {
1388 lock();
1389 }
1390 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM);
1391 }
1392 bind(L_done);
1393 }
1394
1395 // Update counters and perform abort ratio calculation
1396 // input: abort_status_Reg
1397 // rtm_counters_Reg, flags are killed
1398 void MacroAssembler::rtm_profiling(Register abort_status_Reg,
1399 Register rtm_counters_Reg,
1400 RTMLockingCounters* rtm_counters,
1401 Metadata* method_data,
1402 bool profile_rtm) {
1403
1404 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1405 // update rtm counters based on rax value at abort
1406 // reads abort_status_Reg, updates flags
1407 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1408 rtm_counters_update(abort_status_Reg, rtm_counters_Reg);
1409 if (profile_rtm) {
1410 // Save abort status because abort_status_Reg is used by following code.
1411 if (RTMRetryCount > 0) {
1412 push(abort_status_Reg);
1413 }
1414 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1415 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data);
1416 // restore abort status
1417 if (RTMRetryCount > 0) {
1418 pop(abort_status_Reg);
1419 }
1420 }
1421 }
1422
1423 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4)
1424 // inputs: retry_count_Reg
1425 // : abort_status_Reg
1426 // output: retry_count_Reg decremented by 1
1427 // flags are killed
1428 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) {
1429 Label doneRetry;
1430 assert(abort_status_Reg == rax, "");
1431 // The abort reason bits are in eax (see all states in rtmLocking.hpp)
1432 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4)
1433 // if reason is in 0x6 and retry count != 0 then retry
1434 andptr(abort_status_Reg, 0x6);
1435 jccb(Assembler::zero, doneRetry);
1436 testl(retry_count_Reg, retry_count_Reg);
1437 jccb(Assembler::zero, doneRetry);
1438 pause();
1439 decrementl(retry_count_Reg);
1440 jmp(retryLabel);
1441 bind(doneRetry);
1442 }
1443
1444 // Spin and retry if lock is busy,
1445 // inputs: box_Reg (monitor address)
1446 // : retry_count_Reg
1447 // output: retry_count_Reg decremented by 1
1448 // : clear z flag if retry count exceeded
1449 // tmp_Reg, scr_Reg, flags are killed
1450 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg,
1451 Register tmp_Reg, Register scr_Reg, Label& retryLabel) {
1452 Label SpinLoop, SpinExit, doneRetry;
1453 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1454
1455 testl(retry_count_Reg, retry_count_Reg);
1456 jccb(Assembler::zero, doneRetry);
1457 decrementl(retry_count_Reg);
1458 movptr(scr_Reg, RTMSpinLoopCount);
1459
1460 bind(SpinLoop);
1461 pause();
1462 decrementl(scr_Reg);
1463 jccb(Assembler::lessEqual, SpinExit);
1464 movptr(tmp_Reg, Address(box_Reg, owner_offset));
1465 testptr(tmp_Reg, tmp_Reg);
1466 jccb(Assembler::notZero, SpinLoop);
1467
1468 bind(SpinExit);
1469 jmp(retryLabel);
1470 bind(doneRetry);
1471 incrementl(retry_count_Reg); // clear z flag
1472 }
1473
1474 // Use RTM for normal stack locks
1475 // Input: objReg (object to lock)
1476 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg,
1477 Register retry_on_abort_count_Reg,
1478 RTMLockingCounters* stack_rtm_counters,
1479 Metadata* method_data, bool profile_rtm,
1480 Label& DONE_LABEL, Label& IsInflated) {
1481 assert(UseRTMForStackLocks, "why call this otherwise?");
1482 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1483 assert(tmpReg == rax, "");
1484 assert(scrReg == rdx, "");
1485 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1486
1487 if (RTMRetryCount > 0) {
1488 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1489 bind(L_rtm_retry);
1490 }
1491 movptr(tmpReg, Address(objReg, 0));
1492 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1493 jcc(Assembler::notZero, IsInflated);
1494
1495 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1496 Label L_noincrement;
1497 if (RTMTotalCountIncrRate > 1) {
1498 // tmpReg, scrReg and flags are killed
1499 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement);
1500 }
1501 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM");
1502 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg);
1503 bind(L_noincrement);
1504 }
1505 xbegin(L_on_abort);
1506 movptr(tmpReg, Address(objReg, 0)); // fetch markword
1507 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1508 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1509 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked
1510
1511 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1512 if (UseRTMXendForLockBusy) {
1513 xend();
1514 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry)
1515 jmp(L_decrement_retry);
1516 }
1517 else {
1518 xabort(0);
1519 }
1520 bind(L_on_abort);
1521 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1522 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm);
1523 }
1524 bind(L_decrement_retry);
1525 if (RTMRetryCount > 0) {
1526 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1527 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1528 }
1529 }
1530
1531 // Use RTM for inflating locks
1532 // inputs: objReg (object to lock)
1533 // boxReg (on-stack box address (displaced header location) - KILLED)
1534 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value)
1535 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg,
1536 Register scrReg, Register retry_on_busy_count_Reg,
1537 Register retry_on_abort_count_Reg,
1538 RTMLockingCounters* rtm_counters,
1539 Metadata* method_data, bool profile_rtm,
1540 Label& DONE_LABEL) {
1541 assert(UseRTMLocking, "why call this otherwise?");
1542 assert(tmpReg == rax, "");
1543 assert(scrReg == rdx, "");
1544 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1545 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1546
1547 // Without cast to int32_t a movptr will destroy r10 which is typically obj
1548 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1549 movptr(boxReg, tmpReg); // Save ObjectMonitor address
1550
1551 if (RTMRetryCount > 0) {
1552 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy
1553 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1554 bind(L_rtm_retry);
1555 }
1556 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1557 Label L_noincrement;
1558 if (RTMTotalCountIncrRate > 1) {
1559 // tmpReg, scrReg and flags are killed
1560 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement);
1561 }
1562 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1563 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg);
1564 bind(L_noincrement);
1565 }
1566 xbegin(L_on_abort);
1567 movptr(tmpReg, Address(objReg, 0));
1568 movptr(tmpReg, Address(tmpReg, owner_offset));
1569 testptr(tmpReg, tmpReg);
1570 jcc(Assembler::zero, DONE_LABEL);
1571 if (UseRTMXendForLockBusy) {
1572 xend();
1573 jmp(L_decrement_retry);
1574 }
1575 else {
1576 xabort(0);
1577 }
1578 bind(L_on_abort);
1579 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1580 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1581 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm);
1582 }
1583 if (RTMRetryCount > 0) {
1584 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1585 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1586 }
1587
1588 movptr(tmpReg, Address(boxReg, owner_offset)) ;
1589 testptr(tmpReg, tmpReg) ;
1590 jccb(Assembler::notZero, L_decrement_retry) ;
1591
1592 // Appears unlocked - try to swing _owner from null to non-null.
1593 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1594 #ifdef _LP64
1595 Register threadReg = r15_thread;
1596 #else
1597 get_thread(scrReg);
1598 Register threadReg = scrReg;
1599 #endif
1600 if (os::is_MP()) {
1601 lock();
1602 }
1603 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg
1604
1605 if (RTMRetryCount > 0) {
1606 // success done else retry
1607 jccb(Assembler::equal, DONE_LABEL) ;
1608 bind(L_decrement_retry);
1609 // Spin and retry if lock is busy.
1610 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry);
1611 }
1612 else {
1613 bind(L_decrement_retry);
1614 }
1615 }
1616
1617 #endif // INCLUDE_RTM_OPT
1618
1619 // Fast_Lock and Fast_Unlock used by C2
1620
1621 // Because the transitions from emitted code to the runtime
1622 // monitorenter/exit helper stubs are so slow it's critical that
1623 // we inline both the stack-locking fast-path and the inflated fast path.
1624 //
1625 // See also: cmpFastLock and cmpFastUnlock.
1626 //
1627 // What follows is a specialized inline transliteration of the code
1628 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat
1629 // another option would be to emit TrySlowEnter and TrySlowExit methods
1630 // at startup-time. These methods would accept arguments as
1631 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
1632 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply
1633 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
1634 // In practice, however, the # of lock sites is bounded and is usually small.
1635 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
1636 // if the processor uses simple bimodal branch predictors keyed by EIP
1637 // Since the helper routines would be called from multiple synchronization
1638 // sites.
1639 //
1640 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
1641 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
1642 // to those specialized methods. That'd give us a mostly platform-independent
1643 // implementation that the JITs could optimize and inline at their pleasure.
1644 // Done correctly, the only time we'd need to cross to native could would be
1645 // to park() or unpark() threads. We'd also need a few more unsafe operators
1646 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
1647 // (b) explicit barriers or fence operations.
1648 //
1649 // TODO:
1650 //
1651 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
1652 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
1653 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
1654 // the lock operators would typically be faster than reifying Self.
1655 //
1656 // * Ideally I'd define the primitives as:
1657 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
1658 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
1659 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
1660 // Instead, we're stuck with a rather awkward and brittle register assignments below.
1661 // Furthermore the register assignments are overconstrained, possibly resulting in
1662 // sub-optimal code near the synchronization site.
1663 //
1664 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
1665 // Alternately, use a better sp-proximity test.
1666 //
1667 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
1668 // Either one is sufficient to uniquely identify a thread.
1669 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
1670 //
1671 // * Intrinsify notify() and notifyAll() for the common cases where the
1672 // object is locked by the calling thread but the waitlist is empty.
1673 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
1674 //
1675 // * use jccb and jmpb instead of jcc and jmp to improve code density.
1676 // But beware of excessive branch density on AMD Opterons.
1677 //
1678 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
1679 // or failure of the fast-path. If the fast-path fails then we pass
1680 // control to the slow-path, typically in C. In Fast_Lock and
1681 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2
1682 // will emit a conditional branch immediately after the node.
1683 // So we have branches to branches and lots of ICC.ZF games.
1684 // Instead, it might be better to have C2 pass a "FailureLabel"
1685 // into Fast_Lock and Fast_Unlock. In the case of success, control
1686 // will drop through the node. ICC.ZF is undefined at exit.
1687 // In the case of failure, the node will branch directly to the
1688 // FailureLabel
1689
1690
1691 // obj: object to lock
1692 // box: on-stack box address (displaced header location) - KILLED
1693 // rax,: tmp -- KILLED
1694 // scr: tmp -- KILLED
1695 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg,
1696 Register scrReg, Register cx1Reg, Register cx2Reg,
1697 BiasedLockingCounters* counters,
1698 RTMLockingCounters* rtm_counters,
1699 RTMLockingCounters* stack_rtm_counters,
1700 Metadata* method_data,
1701 bool use_rtm, bool profile_rtm) {
1702 // Ensure the register assignents are disjoint
1703 assert(tmpReg == rax, "");
1704
1705 if (use_rtm) {
1706 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg);
1707 } else {
1708 assert(cx1Reg == noreg, "");
1709 assert(cx2Reg == noreg, "");
1710 assert_different_registers(objReg, boxReg, tmpReg, scrReg);
1711 }
1712
1713 if (counters != NULL) {
1714 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg);
1715 }
1716 if (EmitSync & 1) {
1717 // set box->dhw = markOopDesc::unused_mark()
1718 // Force all sync thru slow-path: slow_enter() and slow_exit()
1719 movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1720 cmpptr (rsp, (int32_t)NULL_WORD);
1721 } else {
1722 // Possible cases that we'll encounter in fast_lock
1723 // ------------------------------------------------
1724 // * Inflated
1725 // -- unlocked
1726 // -- Locked
1727 // = by self
1728 // = by other
1729 // * biased
1730 // -- by Self
1731 // -- by other
1732 // * neutral
1733 // * stack-locked
1734 // -- by self
1735 // = sp-proximity test hits
1736 // = sp-proximity test generates false-negative
1737 // -- by other
1738 //
1739
1740 Label IsInflated, DONE_LABEL;
1741
1742 // it's stack-locked, biased or neutral
1743 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
1744 // order to reduce the number of conditional branches in the most common cases.
1745 // Beware -- there's a subtle invariant that fetch of the markword
1746 // at [FETCH], below, will never observe a biased encoding (*101b).
1747 // If this invariant is not held we risk exclusion (safety) failure.
1748 if (UseBiasedLocking && !UseOptoBiasInlining) {
1749 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters);
1750 }
1751
1752 #if INCLUDE_RTM_OPT
1753 if (UseRTMForStackLocks && use_rtm) {
1754 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg,
1755 stack_rtm_counters, method_data, profile_rtm,
1756 DONE_LABEL, IsInflated);
1757 }
1758 #endif // INCLUDE_RTM_OPT
1759
1760 movptr(tmpReg, Address(objReg, 0)); // [FETCH]
1761 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1762 jccb(Assembler::notZero, IsInflated);
1763
1764 // Attempt stack-locking ...
1765 orptr (tmpReg, markOopDesc::unlocked_value);
1766 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
1767 if (os::is_MP()) {
1768 lock();
1769 }
1770 cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg
1771 if (counters != NULL) {
1772 cond_inc32(Assembler::equal,
1773 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1774 }
1775 jcc(Assembler::equal, DONE_LABEL); // Success
1776
1777 // Recursive locking.
1778 // The object is stack-locked: markword contains stack pointer to BasicLock.
1779 // Locked by current thread if difference with current SP is less than one page.
1780 subptr(tmpReg, rsp);
1781 // Next instruction set ZFlag == 1 (Success) if difference is less then one page.
1782 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) );
1783 movptr(Address(boxReg, 0), tmpReg);
1784 if (counters != NULL) {
1785 cond_inc32(Assembler::equal,
1786 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1787 }
1788 jmp(DONE_LABEL);
1789
1790 bind(IsInflated);
1791 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value
1792
1793 #if INCLUDE_RTM_OPT
1794 // Use the same RTM locking code in 32- and 64-bit VM.
1795 if (use_rtm) {
1796 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg,
1797 rtm_counters, method_data, profile_rtm, DONE_LABEL);
1798 } else {
1799 #endif // INCLUDE_RTM_OPT
1800
1801 #ifndef _LP64
1802 // The object is inflated.
1803
1804 // boxReg refers to the on-stack BasicLock in the current frame.
1805 // We'd like to write:
1806 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices.
1807 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
1808 // additional latency as we have another ST in the store buffer that must drain.
1809
1810 if (EmitSync & 8192) {
1811 movptr(Address(boxReg, 0), 3); // results in ST-before-CAS penalty
1812 get_thread (scrReg);
1813 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1814 movptr(tmpReg, NULL_WORD); // consider: xor vs mov
1815 if (os::is_MP()) {
1816 lock();
1817 }
1818 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1819 } else
1820 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS
1821 movptr(scrReg, boxReg);
1822 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1823
1824 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1825 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1826 // prefetchw [eax + Offset(_owner)-2]
1827 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1828 }
1829
1830 if ((EmitSync & 64) == 0) {
1831 // Optimistic form: consider XORL tmpReg,tmpReg
1832 movptr(tmpReg, NULL_WORD);
1833 } else {
1834 // Can suffer RTS->RTO upgrades on shared or cold $ lines
1835 // Test-And-CAS instead of CAS
1836 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner
1837 testptr(tmpReg, tmpReg); // Locked ?
1838 jccb (Assembler::notZero, DONE_LABEL);
1839 }
1840
1841 // Appears unlocked - try to swing _owner from null to non-null.
1842 // Ideally, I'd manifest "Self" with get_thread and then attempt
1843 // to CAS the register containing Self into m->Owner.
1844 // But we don't have enough registers, so instead we can either try to CAS
1845 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
1846 // we later store "Self" into m->Owner. Transiently storing a stack address
1847 // (rsp or the address of the box) into m->owner is harmless.
1848 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1849 if (os::is_MP()) {
1850 lock();
1851 }
1852 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1853 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3
1854 jccb (Assembler::notZero, DONE_LABEL);
1855 get_thread (scrReg); // beware: clobbers ICCs
1856 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg);
1857 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success
1858
1859 // If the CAS fails we can either retry or pass control to the slow-path.
1860 // We use the latter tactic.
1861 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1862 // If the CAS was successful ...
1863 // Self has acquired the lock
1864 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1865 // Intentional fall-through into DONE_LABEL ...
1866 } else {
1867 movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())); // results in ST-before-CAS penalty
1868 movptr(boxReg, tmpReg);
1869
1870 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1871 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1872 // prefetchw [eax + Offset(_owner)-2]
1873 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1874 }
1875
1876 if ((EmitSync & 64) == 0) {
1877 // Optimistic form
1878 xorptr (tmpReg, tmpReg);
1879 } else {
1880 // Can suffer RTS->RTO upgrades on shared or cold $ lines
1881 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner
1882 testptr(tmpReg, tmpReg); // Locked ?
1883 jccb (Assembler::notZero, DONE_LABEL);
1884 }
1885
1886 // Appears unlocked - try to swing _owner from null to non-null.
1887 // Use either "Self" (in scr) or rsp as thread identity in _owner.
1888 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1889 get_thread (scrReg);
1890 if (os::is_MP()) {
1891 lock();
1892 }
1893 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1894
1895 // If the CAS fails we can either retry or pass control to the slow-path.
1896 // We use the latter tactic.
1897 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1898 // If the CAS was successful ...
1899 // Self has acquired the lock
1900 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1901 // Intentional fall-through into DONE_LABEL ...
1902 }
1903 #else // _LP64
1904 // It's inflated
1905 movq(scrReg, tmpReg);
1906 xorq(tmpReg, tmpReg);
1907
1908 if (os::is_MP()) {
1909 lock();
1910 }
1911 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1912 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark().
1913 // Without cast to int32_t movptr will destroy r10 which is typically obj.
1914 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1915 // Intentional fall-through into DONE_LABEL ...
1916 // Propagate ICC.ZF from CAS above into DONE_LABEL.
1917 #endif // _LP64
1918 #if INCLUDE_RTM_OPT
1919 } // use_rtm()
1920 #endif
1921 // DONE_LABEL is a hot target - we'd really like to place it at the
1922 // start of cache line by padding with NOPs.
1923 // See the AMD and Intel software optimization manuals for the
1924 // most efficient "long" NOP encodings.
1925 // Unfortunately none of our alignment mechanisms suffice.
1926 bind(DONE_LABEL);
1927
1928 // At DONE_LABEL the icc ZFlag is set as follows ...
1929 // Fast_Unlock uses the same protocol.
1930 // ZFlag == 1 -> Success
1931 // ZFlag == 0 -> Failure - force control through the slow-path
1932 }
1933 }
1934
1935 // obj: object to unlock
1936 // box: box address (displaced header location), killed. Must be EAX.
1937 // tmp: killed, cannot be obj nor box.
1938 //
1939 // Some commentary on balanced locking:
1940 //
1941 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
1942 // Methods that don't have provably balanced locking are forced to run in the
1943 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
1944 // The interpreter provides two properties:
1945 // I1: At return-time the interpreter automatically and quietly unlocks any
1946 // objects acquired the current activation (frame). Recall that the
1947 // interpreter maintains an on-stack list of locks currently held by
1948 // a frame.
1949 // I2: If a method attempts to unlock an object that is not held by the
1950 // the frame the interpreter throws IMSX.
1951 //
1952 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
1953 // B() doesn't have provably balanced locking so it runs in the interpreter.
1954 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
1955 // is still locked by A().
1956 //
1957 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
1958 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
1959 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
1960 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
1961 // Arguably given that the spec legislates the JNI case as undefined our implementation
1962 // could reasonably *avoid* checking owner in Fast_Unlock().
1963 // In the interest of performance we elide m->Owner==Self check in unlock.
1964 // A perfectly viable alternative is to elide the owner check except when
1965 // Xcheck:jni is enabled.
1966
1967 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) {
1968 assert(boxReg == rax, "");
1969 assert_different_registers(objReg, boxReg, tmpReg);
1970
1971 if (EmitSync & 4) {
1972 // Disable - inhibit all inlining. Force control through the slow-path
1973 cmpptr (rsp, 0);
1974 } else {
1975 Label DONE_LABEL, Stacked, CheckSucc;
1976
1977 // Critically, the biased locking test must have precedence over
1978 // and appear before the (box->dhw == 0) recursive stack-lock test.
1979 if (UseBiasedLocking && !UseOptoBiasInlining) {
1980 biased_locking_exit(objReg, tmpReg, DONE_LABEL);
1981 }
1982
1983 #if INCLUDE_RTM_OPT
1984 if (UseRTMForStackLocks && use_rtm) {
1985 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1986 Label L_regular_unlock;
1987 movptr(tmpReg, Address(objReg, 0)); // fetch markword
1988 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1989 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1990 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock
1991 xend(); // otherwise end...
1992 jmp(DONE_LABEL); // ... and we're done
1993 bind(L_regular_unlock);
1994 }
1995 #endif
1996
1997 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header
1998 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock
1999 movptr(tmpReg, Address(objReg, 0)); // Examine the object's markword
2000 testptr(tmpReg, markOopDesc::monitor_value); // Inflated?
2001 jccb (Assembler::zero, Stacked);
2002
2003 // It's inflated.
2004 #if INCLUDE_RTM_OPT
2005 if (use_rtm) {
2006 Label L_regular_inflated_unlock;
2007 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
2008 movptr(boxReg, Address(tmpReg, owner_offset));
2009 testptr(boxReg, boxReg);
2010 jccb(Assembler::notZero, L_regular_inflated_unlock);
2011 xend();
2012 jmpb(DONE_LABEL);
2013 bind(L_regular_inflated_unlock);
2014 }
2015 #endif
2016
2017 // Despite our balanced locking property we still check that m->_owner == Self
2018 // as java routines or native JNI code called by this thread might
2019 // have released the lock.
2020 // Refer to the comments in synchronizer.cpp for how we might encode extra
2021 // state in _succ so we can avoid fetching EntryList|cxq.
2022 //
2023 // I'd like to add more cases in fast_lock() and fast_unlock() --
2024 // such as recursive enter and exit -- but we have to be wary of
2025 // I$ bloat, T$ effects and BP$ effects.
2026 //
2027 // If there's no contention try a 1-0 exit. That is, exit without
2028 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
2029 // we detect and recover from the race that the 1-0 exit admits.
2030 //
2031 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
2032 // before it STs null into _owner, releasing the lock. Updates
2033 // to data protected by the critical section must be visible before
2034 // we drop the lock (and thus before any other thread could acquire
2035 // the lock and observe the fields protected by the lock).
2036 // IA32's memory-model is SPO, so STs are ordered with respect to
2037 // each other and there's no need for an explicit barrier (fence).
2038 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
2039 #ifndef _LP64
2040 get_thread (boxReg);
2041 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
2042 // prefetchw [ebx + Offset(_owner)-2]
2043 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2044 }
2045
2046 // Note that we could employ various encoding schemes to reduce
2047 // the number of loads below (currently 4) to just 2 or 3.
2048 // Refer to the comments in synchronizer.cpp.
2049 // In practice the chain of fetches doesn't seem to impact performance, however.
2050 xorptr(boxReg, boxReg);
2051 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) {
2052 // Attempt to reduce branch density - AMD's branch predictor.
2053 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2054 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2055 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2056 jccb (Assembler::notZero, DONE_LABEL);
2057 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2058 jmpb (DONE_LABEL);
2059 } else {
2060 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2061 jccb (Assembler::notZero, DONE_LABEL);
2062 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2063 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2064 jccb (Assembler::notZero, CheckSucc);
2065 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2066 jmpb (DONE_LABEL);
2067 }
2068
2069 // The Following code fragment (EmitSync & 65536) improves the performance of
2070 // contended applications and contended synchronization microbenchmarks.
2071 // Unfortunately the emission of the code - even though not executed - causes regressions
2072 // in scimark and jetstream, evidently because of $ effects. Replacing the code
2073 // with an equal number of never-executed NOPs results in the same regression.
2074 // We leave it off by default.
2075
2076 if ((EmitSync & 65536) != 0) {
2077 Label LSuccess, LGoSlowPath ;
2078
2079 bind (CheckSucc);
2080
2081 // Optional pre-test ... it's safe to elide this
2082 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2083 jccb(Assembler::zero, LGoSlowPath);
2084
2085 // We have a classic Dekker-style idiom:
2086 // ST m->_owner = 0 ; MEMBAR; LD m->_succ
2087 // There are a number of ways to implement the barrier:
2088 // (1) lock:andl &m->_owner, 0
2089 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form.
2090 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0
2091 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8
2092 // (2) If supported, an explicit MFENCE is appealing.
2093 // In older IA32 processors MFENCE is slower than lock:add or xchg
2094 // particularly if the write-buffer is full as might be the case if
2095 // if stores closely precede the fence or fence-equivalent instruction.
2096 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2097 // as the situation has changed with Nehalem and Shanghai.
2098 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack
2099 // The $lines underlying the top-of-stack should be in M-state.
2100 // The locked add instruction is serializing, of course.
2101 // (4) Use xchg, which is serializing
2102 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works
2103 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0.
2104 // The integer condition codes will tell us if succ was 0.
2105 // Since _succ and _owner should reside in the same $line and
2106 // we just stored into _owner, it's likely that the $line
2107 // remains in M-state for the lock:orl.
2108 //
2109 // We currently use (3), although it's likely that switching to (2)
2110 // is correct for the future.
2111
2112 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2113 if (os::is_MP()) {
2114 lock(); addptr(Address(rsp, 0), 0);
2115 }
2116 // Ratify _succ remains non-null
2117 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0);
2118 jccb (Assembler::notZero, LSuccess);
2119
2120 xorptr(boxReg, boxReg); // box is really EAX
2121 if (os::is_MP()) { lock(); }
2122 cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2123 jccb (Assembler::notEqual, LSuccess);
2124 // Since we're low on registers we installed rsp as a placeholding in _owner.
2125 // Now install Self over rsp. This is safe as we're transitioning from
2126 // non-null to non=null
2127 get_thread (boxReg);
2128 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg);
2129 // Intentional fall-through into LGoSlowPath ...
2130
2131 bind (LGoSlowPath);
2132 orptr(boxReg, 1); // set ICC.ZF=0 to indicate failure
2133 jmpb (DONE_LABEL);
2134
2135 bind (LSuccess);
2136 xorptr(boxReg, boxReg); // set ICC.ZF=1 to indicate success
2137 jmpb (DONE_LABEL);
2138 }
2139
2140 bind (Stacked);
2141 // It's not inflated and it's not recursively stack-locked and it's not biased.
2142 // It must be stack-locked.
2143 // Try to reset the header to displaced header.
2144 // The "box" value on the stack is stable, so we can reload
2145 // and be assured we observe the same value as above.
2146 movptr(tmpReg, Address(boxReg, 0));
2147 if (os::is_MP()) {
2148 lock();
2149 }
2150 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box
2151 // Intention fall-thru into DONE_LABEL
2152
2153 // DONE_LABEL is a hot target - we'd really like to place it at the
2154 // start of cache line by padding with NOPs.
2155 // See the AMD and Intel software optimization manuals for the
2156 // most efficient "long" NOP encodings.
2157 // Unfortunately none of our alignment mechanisms suffice.
2158 if ((EmitSync & 65536) == 0) {
2159 bind (CheckSucc);
2160 }
2161 #else // _LP64
2162 // It's inflated
2163 if (EmitSync & 1024) {
2164 // Emit code to check that _owner == Self
2165 // We could fold the _owner test into subsequent code more efficiently
2166 // than using a stand-alone check, but since _owner checking is off by
2167 // default we don't bother. We also might consider predicating the
2168 // _owner==Self check on Xcheck:jni or running on a debug build.
2169 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2170 xorptr(boxReg, r15_thread);
2171 } else {
2172 xorptr(boxReg, boxReg);
2173 }
2174 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2175 jccb (Assembler::notZero, DONE_LABEL);
2176 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2177 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2178 jccb (Assembler::notZero, CheckSucc);
2179 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2180 jmpb (DONE_LABEL);
2181
2182 if ((EmitSync & 65536) == 0) {
2183 // Try to avoid passing control into the slow_path ...
2184 Label LSuccess, LGoSlowPath ;
2185 bind (CheckSucc);
2186
2187 // The following optional optimization can be elided if necessary
2188 // Effectively: if (succ == null) goto SlowPath
2189 // The code reduces the window for a race, however,
2190 // and thus benefits performance.
2191 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2192 jccb (Assembler::zero, LGoSlowPath);
2193
2194 if ((EmitSync & 16) && os::is_MP()) {
2195 orptr(boxReg, boxReg);
2196 xchgptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2197 } else {
2198 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2199 if (os::is_MP()) {
2200 // Memory barrier/fence
2201 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ
2202 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack.
2203 // This is faster on Nehalem and AMD Shanghai/Barcelona.
2204 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2205 // We might also restructure (ST Owner=0;barrier;LD _Succ) to
2206 // (mov box,0; xchgq box, &m->Owner; LD _succ) .
2207 lock(); addl(Address(rsp, 0), 0);
2208 }
2209 }
2210 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2211 jccb (Assembler::notZero, LSuccess);
2212
2213 // Rare inopportune interleaving - race.
2214 // The successor vanished in the small window above.
2215 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor.
2216 // We need to ensure progress and succession.
2217 // Try to reacquire the lock.
2218 // If that fails then the new owner is responsible for succession and this
2219 // thread needs to take no further action and can exit via the fast path (success).
2220 // If the re-acquire succeeds then pass control into the slow path.
2221 // As implemented, this latter mode is horrible because we generated more
2222 // coherence traffic on the lock *and* artifically extended the critical section
2223 // length while by virtue of passing control into the slow path.
2224
2225 // box is really RAX -- the following CMPXCHG depends on that binding
2226 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R)
2227 movptr(boxReg, (int32_t)NULL_WORD);
2228 if (os::is_MP()) { lock(); }
2229 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2230 jccb (Assembler::notEqual, LSuccess);
2231 // Intentional fall-through into slow-path
2232
2233 bind (LGoSlowPath);
2234 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure
2235 jmpb (DONE_LABEL);
2236
2237 bind (LSuccess);
2238 testl (boxReg, 0); // set ICC.ZF=1 to indicate success
2239 jmpb (DONE_LABEL);
2240 }
2241
2242 bind (Stacked);
2243 movptr(tmpReg, Address (boxReg, 0)); // re-fetch
2244 if (os::is_MP()) { lock(); }
2245 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box
2246
2247 if (EmitSync & 65536) {
2248 bind (CheckSucc);
2249 }
2250 #endif
2251 bind(DONE_LABEL);
2252 }
2253 }
2254 #endif // COMPILER2
2255
2256 void MacroAssembler::c2bool(Register x) {
2257 // implements x == 0 ? 0 : 1
2258 // note: must only look at least-significant byte of x
2259 // since C-style booleans are stored in one byte
2260 // only! (was bug)
2261 andl(x, 0xFF);
2262 setb(Assembler::notZero, x);
2263 }
2264
2265 // Wouldn't need if AddressLiteral version had new name
2266 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) {
2267 Assembler::call(L, rtype);
2268 }
2269
2270 void MacroAssembler::call(Register entry) {
2271 Assembler::call(entry);
2272 }
2273
2274 void MacroAssembler::call(AddressLiteral entry) {
2275 if (reachable(entry)) {
2276 Assembler::call_literal(entry.target(), entry.rspec());
2277 } else {
2278 lea(rscratch1, entry);
2279 Assembler::call(rscratch1);
2280 }
2281 }
2282
2283 void MacroAssembler::ic_call(address entry) {
2284 RelocationHolder rh = virtual_call_Relocation::spec(pc());
2285 movptr(rax, (intptr_t)Universe::non_oop_word());
2286 call(AddressLiteral(entry, rh));
2287 }
2288
2289 // Implementation of call_VM versions
2290
2291 void MacroAssembler::call_VM(Register oop_result,
2292 address entry_point,
2293 bool check_exceptions) {
2294 Label C, E;
2295 call(C, relocInfo::none);
2296 jmp(E);
2297
2298 bind(C);
2299 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
2300 ret(0);
2301
2302 bind(E);
2303 }
2304
2305 void MacroAssembler::call_VM(Register oop_result,
2306 address entry_point,
2307 Register arg_1,
2308 bool check_exceptions) {
2309 Label C, E;
2310 call(C, relocInfo::none);
2311 jmp(E);
2312
2313 bind(C);
2314 pass_arg1(this, arg_1);
2315 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
2316 ret(0);
2317
2318 bind(E);
2319 }
2320
2321 void MacroAssembler::call_VM(Register oop_result,
2322 address entry_point,
2323 Register arg_1,
2324 Register arg_2,
2325 bool check_exceptions) {
2326 Label C, E;
2327 call(C, relocInfo::none);
2328 jmp(E);
2329
2330 bind(C);
2331
2332 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2333
2334 pass_arg2(this, arg_2);
2335 pass_arg1(this, arg_1);
2336 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
2337 ret(0);
2338
2339 bind(E);
2340 }
2341
2342 void MacroAssembler::call_VM(Register oop_result,
2343 address entry_point,
2344 Register arg_1,
2345 Register arg_2,
2346 Register arg_3,
2347 bool check_exceptions) {
2348 Label C, E;
2349 call(C, relocInfo::none);
2350 jmp(E);
2351
2352 bind(C);
2353
2354 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2355 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2356 pass_arg3(this, arg_3);
2357
2358 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2359 pass_arg2(this, arg_2);
2360
2361 pass_arg1(this, arg_1);
2362 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
2363 ret(0);
2364
2365 bind(E);
2366 }
2367
2368 void MacroAssembler::call_VM(Register oop_result,
2369 Register last_java_sp,
2370 address entry_point,
2371 int number_of_arguments,
2372 bool check_exceptions) {
2373 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2374 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2375 }
2376
2377 void MacroAssembler::call_VM(Register oop_result,
2378 Register last_java_sp,
2379 address entry_point,
2380 Register arg_1,
2381 bool check_exceptions) {
2382 pass_arg1(this, arg_1);
2383 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2384 }
2385
2386 void MacroAssembler::call_VM(Register oop_result,
2387 Register last_java_sp,
2388 address entry_point,
2389 Register arg_1,
2390 Register arg_2,
2391 bool check_exceptions) {
2392
2393 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2394 pass_arg2(this, arg_2);
2395 pass_arg1(this, arg_1);
2396 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2397 }
2398
2399 void MacroAssembler::call_VM(Register oop_result,
2400 Register last_java_sp,
2401 address entry_point,
2402 Register arg_1,
2403 Register arg_2,
2404 Register arg_3,
2405 bool check_exceptions) {
2406 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2407 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2408 pass_arg3(this, arg_3);
2409 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2410 pass_arg2(this, arg_2);
2411 pass_arg1(this, arg_1);
2412 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2413 }
2414
2415 void MacroAssembler::super_call_VM(Register oop_result,
2416 Register last_java_sp,
2417 address entry_point,
2418 int number_of_arguments,
2419 bool check_exceptions) {
2420 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2421 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2422 }
2423
2424 void MacroAssembler::super_call_VM(Register oop_result,
2425 Register last_java_sp,
2426 address entry_point,
2427 Register arg_1,
2428 bool check_exceptions) {
2429 pass_arg1(this, arg_1);
2430 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2431 }
2432
2433 void MacroAssembler::super_call_VM(Register oop_result,
2434 Register last_java_sp,
2435 address entry_point,
2436 Register arg_1,
2437 Register arg_2,
2438 bool check_exceptions) {
2439
2440 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2441 pass_arg2(this, arg_2);
2442 pass_arg1(this, arg_1);
2443 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2444 }
2445
2446 void MacroAssembler::super_call_VM(Register oop_result,
2447 Register last_java_sp,
2448 address entry_point,
2449 Register arg_1,
2450 Register arg_2,
2451 Register arg_3,
2452 bool check_exceptions) {
2453 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2454 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2455 pass_arg3(this, arg_3);
2456 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2457 pass_arg2(this, arg_2);
2458 pass_arg1(this, arg_1);
2459 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2460 }
2461
2462 void MacroAssembler::call_VM_base(Register oop_result,
2463 Register java_thread,
2464 Register last_java_sp,
2465 address entry_point,
2466 int number_of_arguments,
2467 bool check_exceptions) {
2468 // determine java_thread register
2469 if (!java_thread->is_valid()) {
2470 #ifdef _LP64
2471 java_thread = r15_thread;
2472 #else
2473 java_thread = rdi;
2474 get_thread(java_thread);
2475 #endif // LP64
2476 }
2477 // determine last_java_sp register
2478 if (!last_java_sp->is_valid()) {
2479 last_java_sp = rsp;
2480 }
2481 // debugging support
2482 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
2483 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register"));
2484 #ifdef ASSERT
2485 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
2486 // r12 is the heapbase.
2487 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");)
2488 #endif // ASSERT
2489
2490 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
2491 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
2492
2493 // push java thread (becomes first argument of C function)
2494
2495 NOT_LP64(push(java_thread); number_of_arguments++);
2496 LP64_ONLY(mov(c_rarg0, r15_thread));
2497
2498 // set last Java frame before call
2499 assert(last_java_sp != rbp, "can't use ebp/rbp");
2500
2501 // Only interpreter should have to set fp
2502 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL);
2503
2504 // do the call, remove parameters
2505 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
2506
2507 // restore the thread (cannot use the pushed argument since arguments
2508 // may be overwritten by C code generated by an optimizing compiler);
2509 // however can use the register value directly if it is callee saved.
2510 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) {
2511 // rdi & rsi (also r15) are callee saved -> nothing to do
2512 #ifdef ASSERT
2513 guarantee(java_thread != rax, "change this code");
2514 push(rax);
2515 { Label L;
2516 get_thread(rax);
2517 cmpptr(java_thread, rax);
2518 jcc(Assembler::equal, L);
2519 STOP("MacroAssembler::call_VM_base: rdi not callee saved?");
2520 bind(L);
2521 }
2522 pop(rax);
2523 #endif
2524 } else {
2525 get_thread(java_thread);
2526 }
2527 // reset last Java frame
2528 // Only interpreter should have to clear fp
2529 reset_last_Java_frame(java_thread, true, false);
2530
2531 #ifndef CC_INTERP
2532 // C++ interp handles this in the interpreter
2533 check_and_handle_popframe(java_thread);
2534 check_and_handle_earlyret(java_thread);
2535 #endif /* CC_INTERP */
2536
2537 if (check_exceptions) {
2538 // check for pending exceptions (java_thread is set upon return)
2539 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD);
2540 #ifndef _LP64
2541 jump_cc(Assembler::notEqual,
2542 RuntimeAddress(StubRoutines::forward_exception_entry()));
2543 #else
2544 // This used to conditionally jump to forward_exception however it is
2545 // possible if we relocate that the branch will not reach. So we must jump
2546 // around so we can always reach
2547
2548 Label ok;
2549 jcc(Assembler::equal, ok);
2550 jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2551 bind(ok);
2552 #endif // LP64
2553 }
2554
2555 // get oop result if there is one and reset the value in the thread
2556 if (oop_result->is_valid()) {
2557 get_vm_result(oop_result, java_thread);
2558 }
2559 }
2560
2561 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
2562
2563 // Calculate the value for last_Java_sp
2564 // somewhat subtle. call_VM does an intermediate call
2565 // which places a return address on the stack just under the
2566 // stack pointer as the user finsihed with it. This allows
2567 // use to retrieve last_Java_pc from last_Java_sp[-1].
2568 // On 32bit we then have to push additional args on the stack to accomplish
2569 // the actual requested call. On 64bit call_VM only can use register args
2570 // so the only extra space is the return address that call_VM created.
2571 // This hopefully explains the calculations here.
2572
2573 #ifdef _LP64
2574 // We've pushed one address, correct last_Java_sp
2575 lea(rax, Address(rsp, wordSize));
2576 #else
2577 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize));
2578 #endif // LP64
2579
2580 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions);
2581
2582 }
2583
2584 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
2585 call_VM_leaf_base(entry_point, number_of_arguments);
2586 }
2587
2588 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
2589 pass_arg0(this, arg_0);
2590 call_VM_leaf(entry_point, 1);
2591 }
2592
2593 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2594
2595 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2596 pass_arg1(this, arg_1);
2597 pass_arg0(this, arg_0);
2598 call_VM_leaf(entry_point, 2);
2599 }
2600
2601 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2602 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2603 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2604 pass_arg2(this, arg_2);
2605 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2606 pass_arg1(this, arg_1);
2607 pass_arg0(this, arg_0);
2608 call_VM_leaf(entry_point, 3);
2609 }
2610
2611 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2612 pass_arg0(this, arg_0);
2613 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2614 }
2615
2616 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2617
2618 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2619 pass_arg1(this, arg_1);
2620 pass_arg0(this, arg_0);
2621 MacroAssembler::call_VM_leaf_base(entry_point, 2);
2622 }
2623
2624 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2625 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2626 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2627 pass_arg2(this, arg_2);
2628 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2629 pass_arg1(this, arg_1);
2630 pass_arg0(this, arg_0);
2631 MacroAssembler::call_VM_leaf_base(entry_point, 3);
2632 }
2633
2634 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2635 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg"));
2636 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2637 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2638 pass_arg3(this, arg_3);
2639 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2640 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2641 pass_arg2(this, arg_2);
2642 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2643 pass_arg1(this, arg_1);
2644 pass_arg0(this, arg_0);
2645 MacroAssembler::call_VM_leaf_base(entry_point, 4);
2646 }
2647
2648 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
2649 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
2650 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD);
2651 verify_oop(oop_result, "broken oop in call_VM_base");
2652 }
2653
2654 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
2655 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
2656 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD);
2657 }
2658
2659 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2660 }
2661
2662 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2663 }
2664
2665 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) {
2666 if (reachable(src1)) {
2667 cmpl(as_Address(src1), imm);
2668 } else {
2669 lea(rscratch1, src1);
2670 cmpl(Address(rscratch1, 0), imm);
2671 }
2672 }
2673
2674 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) {
2675 assert(!src2.is_lval(), "use cmpptr");
2676 if (reachable(src2)) {
2677 cmpl(src1, as_Address(src2));
2678 } else {
2679 lea(rscratch1, src2);
2680 cmpl(src1, Address(rscratch1, 0));
2681 }
2682 }
2683
2684 void MacroAssembler::cmp32(Register src1, int32_t imm) {
2685 Assembler::cmpl(src1, imm);
2686 }
2687
2688 void MacroAssembler::cmp32(Register src1, Address src2) {
2689 Assembler::cmpl(src1, src2);
2690 }
2691
2692 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2693 ucomisd(opr1, opr2);
2694
2695 Label L;
2696 if (unordered_is_less) {
2697 movl(dst, -1);
2698 jcc(Assembler::parity, L);
2699 jcc(Assembler::below , L);
2700 movl(dst, 0);
2701 jcc(Assembler::equal , L);
2702 increment(dst);
2703 } else { // unordered is greater
2704 movl(dst, 1);
2705 jcc(Assembler::parity, L);
2706 jcc(Assembler::above , L);
2707 movl(dst, 0);
2708 jcc(Assembler::equal , L);
2709 decrementl(dst);
2710 }
2711 bind(L);
2712 }
2713
2714 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2715 ucomiss(opr1, opr2);
2716
2717 Label L;
2718 if (unordered_is_less) {
2719 movl(dst, -1);
2720 jcc(Assembler::parity, L);
2721 jcc(Assembler::below , L);
2722 movl(dst, 0);
2723 jcc(Assembler::equal , L);
2724 increment(dst);
2725 } else { // unordered is greater
2726 movl(dst, 1);
2727 jcc(Assembler::parity, L);
2728 jcc(Assembler::above , L);
2729 movl(dst, 0);
2730 jcc(Assembler::equal , L);
2731 decrementl(dst);
2732 }
2733 bind(L);
2734 }
2735
2736
2737 void MacroAssembler::cmp8(AddressLiteral src1, int imm) {
2738 if (reachable(src1)) {
2739 cmpb(as_Address(src1), imm);
2740 } else {
2741 lea(rscratch1, src1);
2742 cmpb(Address(rscratch1, 0), imm);
2743 }
2744 }
2745
2746 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) {
2747 #ifdef _LP64
2748 if (src2.is_lval()) {
2749 movptr(rscratch1, src2);
2750 Assembler::cmpq(src1, rscratch1);
2751 } else if (reachable(src2)) {
2752 cmpq(src1, as_Address(src2));
2753 } else {
2754 lea(rscratch1, src2);
2755 Assembler::cmpq(src1, Address(rscratch1, 0));
2756 }
2757 #else
2758 if (src2.is_lval()) {
2759 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2760 } else {
2761 cmpl(src1, as_Address(src2));
2762 }
2763 #endif // _LP64
2764 }
2765
2766 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) {
2767 assert(src2.is_lval(), "not a mem-mem compare");
2768 #ifdef _LP64
2769 // moves src2's literal address
2770 movptr(rscratch1, src2);
2771 Assembler::cmpq(src1, rscratch1);
2772 #else
2773 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2774 #endif // _LP64
2775 }
2776
2777 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) {
2778 if (reachable(adr)) {
2779 if (os::is_MP())
2780 lock();
2781 cmpxchgptr(reg, as_Address(adr));
2782 } else {
2783 lea(rscratch1, adr);
2784 if (os::is_MP())
2785 lock();
2786 cmpxchgptr(reg, Address(rscratch1, 0));
2787 }
2788 }
2789
2790 void MacroAssembler::cmpxchgptr(Register reg, Address adr) {
2791 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr));
2792 }
2793
2794 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) {
2795 if (reachable(src)) {
2796 Assembler::comisd(dst, as_Address(src));
2797 } else {
2798 lea(rscratch1, src);
2799 Assembler::comisd(dst, Address(rscratch1, 0));
2800 }
2801 }
2802
2803 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) {
2804 if (reachable(src)) {
2805 Assembler::comiss(dst, as_Address(src));
2806 } else {
2807 lea(rscratch1, src);
2808 Assembler::comiss(dst, Address(rscratch1, 0));
2809 }
2810 }
2811
2812
2813 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) {
2814 Condition negated_cond = negate_condition(cond);
2815 Label L;
2816 jcc(negated_cond, L);
2817 pushf(); // Preserve flags
2818 atomic_incl(counter_addr);
2819 popf();
2820 bind(L);
2821 }
2822
2823 int MacroAssembler::corrected_idivl(Register reg) {
2824 // Full implementation of Java idiv and irem; checks for
2825 // special case as described in JVM spec., p.243 & p.271.
2826 // The function returns the (pc) offset of the idivl
2827 // instruction - may be needed for implicit exceptions.
2828 //
2829 // normal case special case
2830 //
2831 // input : rax,: dividend min_int
2832 // reg: divisor (may not be rax,/rdx) -1
2833 //
2834 // output: rax,: quotient (= rax, idiv reg) min_int
2835 // rdx: remainder (= rax, irem reg) 0
2836 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register");
2837 const int min_int = 0x80000000;
2838 Label normal_case, special_case;
2839
2840 // check for special case
2841 cmpl(rax, min_int);
2842 jcc(Assembler::notEqual, normal_case);
2843 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0)
2844 cmpl(reg, -1);
2845 jcc(Assembler::equal, special_case);
2846
2847 // handle normal case
2848 bind(normal_case);
2849 cdql();
2850 int idivl_offset = offset();
2851 idivl(reg);
2852
2853 // normal and special case exit
2854 bind(special_case);
2855
2856 return idivl_offset;
2857 }
2858
2859
2860
2861 void MacroAssembler::decrementl(Register reg, int value) {
2862 if (value == min_jint) {subl(reg, value) ; return; }
2863 if (value < 0) { incrementl(reg, -value); return; }
2864 if (value == 0) { ; return; }
2865 if (value == 1 && UseIncDec) { decl(reg) ; return; }
2866 /* else */ { subl(reg, value) ; return; }
2867 }
2868
2869 void MacroAssembler::decrementl(Address dst, int value) {
2870 if (value == min_jint) {subl(dst, value) ; return; }
2871 if (value < 0) { incrementl(dst, -value); return; }
2872 if (value == 0) { ; return; }
2873 if (value == 1 && UseIncDec) { decl(dst) ; return; }
2874 /* else */ { subl(dst, value) ; return; }
2875 }
2876
2877 void MacroAssembler::division_with_shift (Register reg, int shift_value) {
2878 assert (shift_value > 0, "illegal shift value");
2879 Label _is_positive;
2880 testl (reg, reg);
2881 jcc (Assembler::positive, _is_positive);
2882 int offset = (1 << shift_value) - 1 ;
2883
2884 if (offset == 1) {
2885 incrementl(reg);
2886 } else {
2887 addl(reg, offset);
2888 }
2889
2890 bind (_is_positive);
2891 sarl(reg, shift_value);
2892 }
2893
2894 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) {
2895 if (reachable(src)) {
2896 Assembler::divsd(dst, as_Address(src));
2897 } else {
2898 lea(rscratch1, src);
2899 Assembler::divsd(dst, Address(rscratch1, 0));
2900 }
2901 }
2902
2903 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) {
2904 if (reachable(src)) {
2905 Assembler::divss(dst, as_Address(src));
2906 } else {
2907 lea(rscratch1, src);
2908 Assembler::divss(dst, Address(rscratch1, 0));
2909 }
2910 }
2911
2912 // !defined(COMPILER2) is because of stupid core builds
2913 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2)
2914 void MacroAssembler::empty_FPU_stack() {
2915 if (VM_Version::supports_mmx()) {
2916 emms();
2917 } else {
2918 for (int i = 8; i-- > 0; ) ffree(i);
2919 }
2920 }
2921 #endif // !LP64 || C1 || !C2
2922
2923
2924 // Defines obj, preserves var_size_in_bytes
2925 void MacroAssembler::eden_allocate(Register obj,
2926 Register var_size_in_bytes,
2927 int con_size_in_bytes,
2928 Register t1,
2929 Label& slow_case) {
2930 assert(obj == rax, "obj must be in rax, for cmpxchg");
2931 assert_different_registers(obj, var_size_in_bytes, t1);
2932 if (!Universe::heap()->supports_inline_contig_alloc()) {
2933 jmp(slow_case);
2934 } else {
2935 Register end = t1;
2936 Label retry;
2937 bind(retry);
2938 ExternalAddress heap_top((address) Universe::heap()->top_addr());
2939 movptr(obj, heap_top);
2940 if (var_size_in_bytes == noreg) {
2941 lea(end, Address(obj, con_size_in_bytes));
2942 } else {
2943 lea(end, Address(obj, var_size_in_bytes, Address::times_1));
2944 }
2945 // if end < obj then we wrapped around => object too long => slow case
2946 cmpptr(end, obj);
2947 jcc(Assembler::below, slow_case);
2948 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr()));
2949 jcc(Assembler::above, slow_case);
2950 // Compare obj with the top addr, and if still equal, store the new top addr in
2951 // end at the address of the top addr pointer. Sets ZF if was equal, and clears
2952 // it otherwise. Use lock prefix for atomicity on MPs.
2953 locked_cmpxchgptr(end, heap_top);
2954 jcc(Assembler::notEqual, retry);
2955 }
2956 }
2957
2958 void MacroAssembler::enter() {
2959 push(rbp);
2960 mov(rbp, rsp);
2961 }
2962
2963 // A 5 byte nop that is safe for patching (see patch_verified_entry)
2964 void MacroAssembler::fat_nop() {
2965 if (UseAddressNop) {
2966 addr_nop_5();
2967 } else {
2968 emit_int8(0x26); // es:
2969 emit_int8(0x2e); // cs:
2970 emit_int8(0x64); // fs:
2971 emit_int8(0x65); // gs:
2972 emit_int8((unsigned char)0x90);
2973 }
2974 }
2975
2976 void MacroAssembler::fcmp(Register tmp) {
2977 fcmp(tmp, 1, true, true);
2978 }
2979
2980 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) {
2981 assert(!pop_right || pop_left, "usage error");
2982 if (VM_Version::supports_cmov()) {
2983 assert(tmp == noreg, "unneeded temp");
2984 if (pop_left) {
2985 fucomip(index);
2986 } else {
2987 fucomi(index);
2988 }
2989 if (pop_right) {
2990 fpop();
2991 }
2992 } else {
2993 assert(tmp != noreg, "need temp");
2994 if (pop_left) {
2995 if (pop_right) {
2996 fcompp();
2997 } else {
2998 fcomp(index);
2999 }
3000 } else {
3001 fcom(index);
3002 }
3003 // convert FPU condition into eflags condition via rax,
3004 save_rax(tmp);
3005 fwait(); fnstsw_ax();
3006 sahf();
3007 restore_rax(tmp);
3008 }
3009 // condition codes set as follows:
3010 //
3011 // CF (corresponds to C0) if x < y
3012 // PF (corresponds to C2) if unordered
3013 // ZF (corresponds to C3) if x = y
3014 }
3015
3016 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) {
3017 fcmp2int(dst, unordered_is_less, 1, true, true);
3018 }
3019
3020 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) {
3021 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right);
3022 Label L;
3023 if (unordered_is_less) {
3024 movl(dst, -1);
3025 jcc(Assembler::parity, L);
3026 jcc(Assembler::below , L);
3027 movl(dst, 0);
3028 jcc(Assembler::equal , L);
3029 increment(dst);
3030 } else { // unordered is greater
3031 movl(dst, 1);
3032 jcc(Assembler::parity, L);
3033 jcc(Assembler::above , L);
3034 movl(dst, 0);
3035 jcc(Assembler::equal , L);
3036 decrementl(dst);
3037 }
3038 bind(L);
3039 }
3040
3041 void MacroAssembler::fld_d(AddressLiteral src) {
3042 fld_d(as_Address(src));
3043 }
3044
3045 void MacroAssembler::fld_s(AddressLiteral src) {
3046 fld_s(as_Address(src));
3047 }
3048
3049 void MacroAssembler::fld_x(AddressLiteral src) {
3050 Assembler::fld_x(as_Address(src));
3051 }
3052
3053 void MacroAssembler::fldcw(AddressLiteral src) {
3054 Assembler::fldcw(as_Address(src));
3055 }
3056
3057 void MacroAssembler::pow_exp_core_encoding() {
3058 // kills rax, rcx, rdx
3059 subptr(rsp,sizeof(jdouble));
3060 // computes 2^X. Stack: X ...
3061 // f2xm1 computes 2^X-1 but only operates on -1<=X<=1. Get int(X) and
3062 // keep it on the thread's stack to compute 2^int(X) later
3063 // then compute 2^(X-int(X)) as (2^(X-int(X)-1+1)
3064 // final result is obtained with: 2^X = 2^int(X) * 2^(X-int(X))
3065 fld_s(0); // Stack: X X ...
3066 frndint(); // Stack: int(X) X ...
3067 fsuba(1); // Stack: int(X) X-int(X) ...
3068 fistp_s(Address(rsp,0)); // move int(X) as integer to thread's stack. Stack: X-int(X) ...
3069 f2xm1(); // Stack: 2^(X-int(X))-1 ...
3070 fld1(); // Stack: 1 2^(X-int(X))-1 ...
3071 faddp(1); // Stack: 2^(X-int(X))
3072 // computes 2^(int(X)): add exponent bias (1023) to int(X), then
3073 // shift int(X)+1023 to exponent position.
3074 // Exponent is limited to 11 bits if int(X)+1023 does not fit in 11
3075 // bits, set result to NaN. 0x000 and 0x7FF are reserved exponent
3076 // values so detect them and set result to NaN.
3077 movl(rax,Address(rsp,0));
3078 movl(rcx, -2048); // 11 bit mask and valid NaN binary encoding
3079 addl(rax, 1023);
3080 movl(rdx,rax);
3081 shll(rax,20);
3082 // Check that 0 < int(X)+1023 < 2047. Otherwise set rax to NaN.
3083 addl(rdx,1);
3084 // Check that 1 < int(X)+1023+1 < 2048
3085 // in 3 steps:
3086 // 1- (int(X)+1023+1)&-2048 == 0 => 0 <= int(X)+1023+1 < 2048
3087 // 2- (int(X)+1023+1)&-2048 != 0
3088 // 3- (int(X)+1023+1)&-2048 != 1
3089 // Do 2- first because addl just updated the flags.
3090 cmov32(Assembler::equal,rax,rcx);
3091 cmpl(rdx,1);
3092 cmov32(Assembler::equal,rax,rcx);
3093 testl(rdx,rcx);
3094 cmov32(Assembler::notEqual,rax,rcx);
3095 movl(Address(rsp,4),rax);
3096 movl(Address(rsp,0),0);
3097 fmul_d(Address(rsp,0)); // Stack: 2^X ...
3098 addptr(rsp,sizeof(jdouble));
3099 }
3100
3101 void MacroAssembler::increase_precision() {
3102 subptr(rsp, BytesPerWord);
3103 fnstcw(Address(rsp, 0));
3104 movl(rax, Address(rsp, 0));
3105 orl(rax, 0x300);
3106 push(rax);
3107 fldcw(Address(rsp, 0));
3108 pop(rax);
3109 }
3110
3111 void MacroAssembler::restore_precision() {
3112 fldcw(Address(rsp, 0));
3113 addptr(rsp, BytesPerWord);
3114 }
3115
3116 void MacroAssembler::fast_pow() {
3117 // computes X^Y = 2^(Y * log2(X))
3118 // if fast computation is not possible, result is NaN. Requires
3119 // fallback from user of this macro.
3120 // increase precision for intermediate steps of the computation
3121 BLOCK_COMMENT("fast_pow {");
3122 increase_precision();
3123 fyl2x(); // Stack: (Y*log2(X)) ...
3124 pow_exp_core_encoding(); // Stack: exp(X) ...
3125 restore_precision();
3126 BLOCK_COMMENT("} fast_pow");
3127 }
3128
3129 void MacroAssembler::fast_exp() {
3130 // computes exp(X) = 2^(X * log2(e))
3131 // if fast computation is not possible, result is NaN. Requires
3132 // fallback from user of this macro.
3133 // increase precision for intermediate steps of the computation
3134 increase_precision();
3135 fldl2e(); // Stack: log2(e) X ...
3136 fmulp(1); // Stack: (X*log2(e)) ...
3137 pow_exp_core_encoding(); // Stack: exp(X) ...
3138 restore_precision();
3139 }
3140
3141 void MacroAssembler::pow_or_exp(bool is_exp, int num_fpu_regs_in_use) {
3142 // kills rax, rcx, rdx
3143 // pow and exp needs 2 extra registers on the fpu stack.
3144 Label slow_case, done;
3145 Register tmp = noreg;
3146 if (!VM_Version::supports_cmov()) {
3147 // fcmp needs a temporary so preserve rdx,
3148 tmp = rdx;
3149 }
3150 Register tmp2 = rax;
3151 Register tmp3 = rcx;
3152
3153 if (is_exp) {
3154 // Stack: X
3155 fld_s(0); // duplicate argument for runtime call. Stack: X X
3156 fast_exp(); // Stack: exp(X) X
3157 fcmp(tmp, 0, false, false); // Stack: exp(X) X
3158 // exp(X) not equal to itself: exp(X) is NaN go to slow case.
3159 jcc(Assembler::parity, slow_case);
3160 // get rid of duplicate argument. Stack: exp(X)
3161 if (num_fpu_regs_in_use > 0) {
3162 fxch();
3163 fpop();
3164 } else {
3165 ffree(1);
3166 }
3167 jmp(done);
3168 } else {
3169 // Stack: X Y
3170 Label x_negative, y_not_2;
3171
3172 static double two = 2.0;
3173 ExternalAddress two_addr((address)&two);
3174
3175 // constant maybe too far on 64 bit
3176 lea(tmp2, two_addr);
3177 fld_d(Address(tmp2, 0)); // Stack: 2 X Y
3178 fcmp(tmp, 2, true, false); // Stack: X Y
3179 jcc(Assembler::parity, y_not_2);
3180 jcc(Assembler::notEqual, y_not_2);
3181
3182 fxch(); fpop(); // Stack: X
3183 fmul(0); // Stack: X*X
3184
3185 jmp(done);
3186
3187 bind(y_not_2);
3188
3189 fldz(); // Stack: 0 X Y
3190 fcmp(tmp, 1, true, false); // Stack: X Y
3191 jcc(Assembler::above, x_negative);
3192
3193 // X >= 0
3194
3195 fld_s(1); // duplicate arguments for runtime call. Stack: Y X Y
3196 fld_s(1); // Stack: X Y X Y
3197 fast_pow(); // Stack: X^Y X Y
3198 fcmp(tmp, 0, false, false); // Stack: X^Y X Y
3199 // X^Y not equal to itself: X^Y is NaN go to slow case.
3200 jcc(Assembler::parity, slow_case);
3201 // get rid of duplicate arguments. Stack: X^Y
3202 if (num_fpu_regs_in_use > 0) {
3203 fxch(); fpop();
3204 fxch(); fpop();
3205 } else {
3206 ffree(2);
3207 ffree(1);
3208 }
3209 jmp(done);
3210
3211 // X <= 0
3212 bind(x_negative);
3213
3214 fld_s(1); // Stack: Y X Y
3215 frndint(); // Stack: int(Y) X Y
3216 fcmp(tmp, 2, false, false); // Stack: int(Y) X Y
3217 jcc(Assembler::notEqual, slow_case);
3218
3219 subptr(rsp, 8);
3220
3221 // For X^Y, when X < 0, Y has to be an integer and the final
3222 // result depends on whether it's odd or even. We just checked
3223 // that int(Y) == Y. We move int(Y) to gp registers as a 64 bit
3224 // integer to test its parity. If int(Y) is huge and doesn't fit
3225 // in the 64 bit integer range, the integer indefinite value will
3226 // end up in the gp registers. Huge numbers are all even, the
3227 // integer indefinite number is even so it's fine.
3228
3229 #ifdef ASSERT
3230 // Let's check we don't end up with an integer indefinite number
3231 // when not expected. First test for huge numbers: check whether
3232 // int(Y)+1 == int(Y) which is true for very large numbers and
3233 // those are all even. A 64 bit integer is guaranteed to not
3234 // overflow for numbers where y+1 != y (when precision is set to
3235 // double precision).
3236 Label y_not_huge;
3237
3238 fld1(); // Stack: 1 int(Y) X Y
3239 fadd(1); // Stack: 1+int(Y) int(Y) X Y
3240
3241 #ifdef _LP64
3242 // trip to memory to force the precision down from double extended
3243 // precision
3244 fstp_d(Address(rsp, 0));
3245 fld_d(Address(rsp, 0));
3246 #endif
3247
3248 fcmp(tmp, 1, true, false); // Stack: int(Y) X Y
3249 #endif
3250
3251 // move int(Y) as 64 bit integer to thread's stack
3252 fistp_d(Address(rsp,0)); // Stack: X Y
3253
3254 #ifdef ASSERT
3255 jcc(Assembler::notEqual, y_not_huge);
3256
3257 // Y is huge so we know it's even. It may not fit in a 64 bit
3258 // integer and we don't want the debug code below to see the
3259 // integer indefinite value so overwrite int(Y) on the thread's
3260 // stack with 0.
3261 movl(Address(rsp, 0), 0);
3262 movl(Address(rsp, 4), 0);
3263
3264 bind(y_not_huge);
3265 #endif
3266
3267 fld_s(1); // duplicate arguments for runtime call. Stack: Y X Y
3268 fld_s(1); // Stack: X Y X Y
3269 fabs(); // Stack: abs(X) Y X Y
3270 fast_pow(); // Stack: abs(X)^Y X Y
3271 fcmp(tmp, 0, false, false); // Stack: abs(X)^Y X Y
3272 // abs(X)^Y not equal to itself: abs(X)^Y is NaN go to slow case.
3273
3274 pop(tmp2);
3275 NOT_LP64(pop(tmp3));
3276 jcc(Assembler::parity, slow_case);
3277
3278 #ifdef ASSERT
3279 // Check that int(Y) is not integer indefinite value (int
3280 // overflow). Shouldn't happen because for values that would
3281 // overflow, 1+int(Y)==Y which was tested earlier.
3282 #ifndef _LP64
3283 {
3284 Label integer;
3285 testl(tmp2, tmp2);
3286 jcc(Assembler::notZero, integer);
3287 cmpl(tmp3, 0x80000000);
3288 jcc(Assembler::notZero, integer);
3289 STOP("integer indefinite value shouldn't be seen here");
3290 bind(integer);
3291 }
3292 #else
3293 {
3294 Label integer;
3295 mov(tmp3, tmp2); // preserve tmp2 for parity check below
3296 shlq(tmp3, 1);
3297 jcc(Assembler::carryClear, integer);
3298 jcc(Assembler::notZero, integer);
3299 STOP("integer indefinite value shouldn't be seen here");
3300 bind(integer);
3301 }
3302 #endif
3303 #endif
3304
3305 // get rid of duplicate arguments. Stack: X^Y
3306 if (num_fpu_regs_in_use > 0) {
3307 fxch(); fpop();
3308 fxch(); fpop();
3309 } else {
3310 ffree(2);
3311 ffree(1);
3312 }
3313
3314 testl(tmp2, 1);
3315 jcc(Assembler::zero, done); // X <= 0, Y even: X^Y = abs(X)^Y
3316 // X <= 0, Y even: X^Y = -abs(X)^Y
3317
3318 fchs(); // Stack: -abs(X)^Y Y
3319 jmp(done);
3320 }
3321
3322 // slow case: runtime call
3323 bind(slow_case);
3324
3325 fpop(); // pop incorrect result or int(Y)
3326
3327 fp_runtime_fallback(is_exp ? CAST_FROM_FN_PTR(address, SharedRuntime::dexp) : CAST_FROM_FN_PTR(address, SharedRuntime::dpow),
3328 is_exp ? 1 : 2, num_fpu_regs_in_use);
3329
3330 // Come here with result in F-TOS
3331 bind(done);
3332 }
3333
3334 void MacroAssembler::fpop() {
3335 ffree();
3336 fincstp();
3337 }
3338
3339 void MacroAssembler::fremr(Register tmp) {
3340 save_rax(tmp);
3341 { Label L;
3342 bind(L);
3343 fprem();
3344 fwait(); fnstsw_ax();
3345 #ifdef _LP64
3346 testl(rax, 0x400);
3347 jcc(Assembler::notEqual, L);
3348 #else
3349 sahf();
3350 jcc(Assembler::parity, L);
3351 #endif // _LP64
3352 }
3353 restore_rax(tmp);
3354 // Result is in ST0.
3355 // Note: fxch & fpop to get rid of ST1
3356 // (otherwise FPU stack could overflow eventually)
3357 fxch(1);
3358 fpop();
3359 }
3360
3361
3362 void MacroAssembler::incrementl(AddressLiteral dst) {
3363 if (reachable(dst)) {
3364 incrementl(as_Address(dst));
3365 } else {
3366 lea(rscratch1, dst);
3367 incrementl(Address(rscratch1, 0));
3368 }
3369 }
3370
3371 void MacroAssembler::incrementl(ArrayAddress dst) {
3372 incrementl(as_Address(dst));
3373 }
3374
3375 void MacroAssembler::incrementl(Register reg, int value) {
3376 if (value == min_jint) {addl(reg, value) ; return; }
3377 if (value < 0) { decrementl(reg, -value); return; }
3378 if (value == 0) { ; return; }
3379 if (value == 1 && UseIncDec) { incl(reg) ; return; }
3380 /* else */ { addl(reg, value) ; return; }
3381 }
3382
3383 void MacroAssembler::incrementl(Address dst, int value) {
3384 if (value == min_jint) {addl(dst, value) ; return; }
3385 if (value < 0) { decrementl(dst, -value); return; }
3386 if (value == 0) { ; return; }
3387 if (value == 1 && UseIncDec) { incl(dst) ; return; }
3388 /* else */ { addl(dst, value) ; return; }
3389 }
3390
3391 void MacroAssembler::jump(AddressLiteral dst) {
3392 if (reachable(dst)) {
3393 jmp_literal(dst.target(), dst.rspec());
3394 } else {
3395 lea(rscratch1, dst);
3396 jmp(rscratch1);
3397 }
3398 }
3399
3400 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) {
3401 if (reachable(dst)) {
3402 InstructionMark im(this);
3403 relocate(dst.reloc());
3404 const int short_size = 2;
3405 const int long_size = 6;
3406 int offs = (intptr_t)dst.target() - ((intptr_t)pc());
3407 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
3408 // 0111 tttn #8-bit disp
3409 emit_int8(0x70 | cc);
3410 emit_int8((offs - short_size) & 0xFF);
3411 } else {
3412 // 0000 1111 1000 tttn #32-bit disp
3413 emit_int8(0x0F);
3414 emit_int8((unsigned char)(0x80 | cc));
3415 emit_int32(offs - long_size);
3416 }
3417 } else {
3418 #ifdef ASSERT
3419 warning("reversing conditional branch");
3420 #endif /* ASSERT */
3421 Label skip;
3422 jccb(reverse[cc], skip);
3423 lea(rscratch1, dst);
3424 Assembler::jmp(rscratch1);
3425 bind(skip);
3426 }
3427 }
3428
3429 void MacroAssembler::ldmxcsr(AddressLiteral src) {
3430 if (reachable(src)) {
3431 Assembler::ldmxcsr(as_Address(src));
3432 } else {
3433 lea(rscratch1, src);
3434 Assembler::ldmxcsr(Address(rscratch1, 0));
3435 }
3436 }
3437
3438 int MacroAssembler::load_signed_byte(Register dst, Address src) {
3439 int off;
3440 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3441 off = offset();
3442 movsbl(dst, src); // movsxb
3443 } else {
3444 off = load_unsigned_byte(dst, src);
3445 shll(dst, 24);
3446 sarl(dst, 24);
3447 }
3448 return off;
3449 }
3450
3451 // Note: load_signed_short used to be called load_signed_word.
3452 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler
3453 // manual, which means 16 bits, that usage is found nowhere in HotSpot code.
3454 // The term "word" in HotSpot means a 32- or 64-bit machine word.
3455 int MacroAssembler::load_signed_short(Register dst, Address src) {
3456 int off;
3457 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3458 // This is dubious to me since it seems safe to do a signed 16 => 64 bit
3459 // version but this is what 64bit has always done. This seems to imply
3460 // that users are only using 32bits worth.
3461 off = offset();
3462 movswl(dst, src); // movsxw
3463 } else {
3464 off = load_unsigned_short(dst, src);
3465 shll(dst, 16);
3466 sarl(dst, 16);
3467 }
3468 return off;
3469 }
3470
3471 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
3472 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3473 // and "3.9 Partial Register Penalties", p. 22).
3474 int off;
3475 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) {
3476 off = offset();
3477 movzbl(dst, src); // movzxb
3478 } else {
3479 xorl(dst, dst);
3480 off = offset();
3481 movb(dst, src);
3482 }
3483 return off;
3484 }
3485
3486 // Note: load_unsigned_short used to be called load_unsigned_word.
3487 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
3488 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3489 // and "3.9 Partial Register Penalties", p. 22).
3490 int off;
3491 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) {
3492 off = offset();
3493 movzwl(dst, src); // movzxw
3494 } else {
3495 xorl(dst, dst);
3496 off = offset();
3497 movw(dst, src);
3498 }
3499 return off;
3500 }
3501
3502 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
3503 switch (size_in_bytes) {
3504 #ifndef _LP64
3505 case 8:
3506 assert(dst2 != noreg, "second dest register required");
3507 movl(dst, src);
3508 movl(dst2, src.plus_disp(BytesPerInt));
3509 break;
3510 #else
3511 case 8: movq(dst, src); break;
3512 #endif
3513 case 4: movl(dst, src); break;
3514 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
3515 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
3516 default: ShouldNotReachHere();
3517 }
3518 }
3519
3520 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
3521 switch (size_in_bytes) {
3522 #ifndef _LP64
3523 case 8:
3524 assert(src2 != noreg, "second source register required");
3525 movl(dst, src);
3526 movl(dst.plus_disp(BytesPerInt), src2);
3527 break;
3528 #else
3529 case 8: movq(dst, src); break;
3530 #endif
3531 case 4: movl(dst, src); break;
3532 case 2: movw(dst, src); break;
3533 case 1: movb(dst, src); break;
3534 default: ShouldNotReachHere();
3535 }
3536 }
3537
3538 void MacroAssembler::mov32(AddressLiteral dst, Register src) {
3539 if (reachable(dst)) {
3540 movl(as_Address(dst), src);
3541 } else {
3542 lea(rscratch1, dst);
3543 movl(Address(rscratch1, 0), src);
3544 }
3545 }
3546
3547 void MacroAssembler::mov32(Register dst, AddressLiteral src) {
3548 if (reachable(src)) {
3549 movl(dst, as_Address(src));
3550 } else {
3551 lea(rscratch1, src);
3552 movl(dst, Address(rscratch1, 0));
3553 }
3554 }
3555
3556 // C++ bool manipulation
3557
3558 void MacroAssembler::movbool(Register dst, Address src) {
3559 if(sizeof(bool) == 1)
3560 movb(dst, src);
3561 else if(sizeof(bool) == 2)
3562 movw(dst, src);
3563 else if(sizeof(bool) == 4)
3564 movl(dst, src);
3565 else
3566 // unsupported
3567 ShouldNotReachHere();
3568 }
3569
3570 void MacroAssembler::movbool(Address dst, bool boolconst) {
3571 if(sizeof(bool) == 1)
3572 movb(dst, (int) boolconst);
3573 else if(sizeof(bool) == 2)
3574 movw(dst, (int) boolconst);
3575 else if(sizeof(bool) == 4)
3576 movl(dst, (int) boolconst);
3577 else
3578 // unsupported
3579 ShouldNotReachHere();
3580 }
3581
3582 void MacroAssembler::movbool(Address dst, Register src) {
3583 if(sizeof(bool) == 1)
3584 movb(dst, src);
3585 else if(sizeof(bool) == 2)
3586 movw(dst, src);
3587 else if(sizeof(bool) == 4)
3588 movl(dst, src);
3589 else
3590 // unsupported
3591 ShouldNotReachHere();
3592 }
3593
3594 void MacroAssembler::movbyte(ArrayAddress dst, int src) {
3595 movb(as_Address(dst), src);
3596 }
3597
3598 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) {
3599 if (reachable(src)) {
3600 movdl(dst, as_Address(src));
3601 } else {
3602 lea(rscratch1, src);
3603 movdl(dst, Address(rscratch1, 0));
3604 }
3605 }
3606
3607 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) {
3608 if (reachable(src)) {
3609 movq(dst, as_Address(src));
3610 } else {
3611 lea(rscratch1, src);
3612 movq(dst, Address(rscratch1, 0));
3613 }
3614 }
3615
3616 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) {
3617 if (reachable(src)) {
3618 if (UseXmmLoadAndClearUpper) {
3619 movsd (dst, as_Address(src));
3620 } else {
3621 movlpd(dst, as_Address(src));
3622 }
3623 } else {
3624 lea(rscratch1, src);
3625 if (UseXmmLoadAndClearUpper) {
3626 movsd (dst, Address(rscratch1, 0));
3627 } else {
3628 movlpd(dst, Address(rscratch1, 0));
3629 }
3630 }
3631 }
3632
3633 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) {
3634 if (reachable(src)) {
3635 movss(dst, as_Address(src));
3636 } else {
3637 lea(rscratch1, src);
3638 movss(dst, Address(rscratch1, 0));
3639 }
3640 }
3641
3642 void MacroAssembler::movptr(Register dst, Register src) {
3643 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3644 }
3645
3646 void MacroAssembler::movptr(Register dst, Address src) {
3647 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3648 }
3649
3650 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
3651 void MacroAssembler::movptr(Register dst, intptr_t src) {
3652 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src));
3653 }
3654
3655 void MacroAssembler::movptr(Address dst, Register src) {
3656 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3657 }
3658
3659 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src) {
3660 if (reachable(src)) {
3661 Assembler::movdqu(dst, as_Address(src));
3662 } else {
3663 lea(rscratch1, src);
3664 Assembler::movdqu(dst, Address(rscratch1, 0));
3665 }
3666 }
3667
3668 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) {
3669 if (reachable(src)) {
3670 Assembler::movdqa(dst, as_Address(src));
3671 } else {
3672 lea(rscratch1, src);
3673 Assembler::movdqa(dst, Address(rscratch1, 0));
3674 }
3675 }
3676
3677 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) {
3678 if (reachable(src)) {
3679 Assembler::movsd(dst, as_Address(src));
3680 } else {
3681 lea(rscratch1, src);
3682 Assembler::movsd(dst, Address(rscratch1, 0));
3683 }
3684 }
3685
3686 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) {
3687 if (reachable(src)) {
3688 Assembler::movss(dst, as_Address(src));
3689 } else {
3690 lea(rscratch1, src);
3691 Assembler::movss(dst, Address(rscratch1, 0));
3692 }
3693 }
3694
3695 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) {
3696 if (reachable(src)) {
3697 Assembler::mulsd(dst, as_Address(src));
3698 } else {
3699 lea(rscratch1, src);
3700 Assembler::mulsd(dst, Address(rscratch1, 0));
3701 }
3702 }
3703
3704 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) {
3705 if (reachable(src)) {
3706 Assembler::mulss(dst, as_Address(src));
3707 } else {
3708 lea(rscratch1, src);
3709 Assembler::mulss(dst, Address(rscratch1, 0));
3710 }
3711 }
3712
3713 void MacroAssembler::null_check(Register reg, int offset) {
3714 if (needs_explicit_null_check(offset)) {
3715 // provoke OS NULL exception if reg = NULL by
3716 // accessing M[reg] w/o changing any (non-CC) registers
3717 // NOTE: cmpl is plenty here to provoke a segv
3718 cmpptr(rax, Address(reg, 0));
3719 // Note: should probably use testl(rax, Address(reg, 0));
3720 // may be shorter code (however, this version of
3721 // testl needs to be implemented first)
3722 } else {
3723 // nothing to do, (later) access of M[reg + offset]
3724 // will provoke OS NULL exception if reg = NULL
3725 }
3726 }
3727
3728 void MacroAssembler::os_breakpoint() {
3729 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability
3730 // (e.g., MSVC can't call ps() otherwise)
3731 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint)));
3732 }
3733
3734 void MacroAssembler::pop_CPU_state() {
3735 pop_FPU_state();
3736 pop_IU_state();
3737 }
3738
3739 void MacroAssembler::pop_FPU_state() {
3740 NOT_LP64(frstor(Address(rsp, 0));)
3741 LP64_ONLY(fxrstor(Address(rsp, 0));)
3742 addptr(rsp, FPUStateSizeInWords * wordSize);
3743 }
3744
3745 void MacroAssembler::pop_IU_state() {
3746 popa();
3747 LP64_ONLY(addq(rsp, 8));
3748 popf();
3749 }
3750
3751 // Save Integer and Float state
3752 // Warning: Stack must be 16 byte aligned (64bit)
3753 void MacroAssembler::push_CPU_state() {
3754 push_IU_state();
3755 push_FPU_state();
3756 }
3757
3758 void MacroAssembler::push_FPU_state() {
3759 subptr(rsp, FPUStateSizeInWords * wordSize);
3760 #ifndef _LP64
3761 fnsave(Address(rsp, 0));
3762 fwait();
3763 #else
3764 fxsave(Address(rsp, 0));
3765 #endif // LP64
3766 }
3767
3768 void MacroAssembler::push_IU_state() {
3769 // Push flags first because pusha kills them
3770 pushf();
3771 // Make sure rsp stays 16-byte aligned
3772 LP64_ONLY(subq(rsp, 8));
3773 pusha();
3774 }
3775
3776 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp, bool clear_pc) {
3777 // determine java_thread register
3778 if (!java_thread->is_valid()) {
3779 java_thread = rdi;
3780 get_thread(java_thread);
3781 }
3782 // we must set sp to zero to clear frame
3783 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
3784 if (clear_fp) {
3785 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3786 }
3787
3788 if (clear_pc)
3789 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
3790
3791 }
3792
3793 void MacroAssembler::restore_rax(Register tmp) {
3794 if (tmp == noreg) pop(rax);
3795 else if (tmp != rax) mov(rax, tmp);
3796 }
3797
3798 void MacroAssembler::round_to(Register reg, int modulus) {
3799 addptr(reg, modulus - 1);
3800 andptr(reg, -modulus);
3801 }
3802
3803 void MacroAssembler::save_rax(Register tmp) {
3804 if (tmp == noreg) push(rax);
3805 else if (tmp != rax) mov(tmp, rax);
3806 }
3807
3808 // Write serialization page so VM thread can do a pseudo remote membar.
3809 // We use the current thread pointer to calculate a thread specific
3810 // offset to write to within the page. This minimizes bus traffic
3811 // due to cache line collision.
3812 void MacroAssembler::serialize_memory(Register thread, Register tmp) {
3813 movl(tmp, thread);
3814 shrl(tmp, os::get_serialize_page_shift_count());
3815 andl(tmp, (os::vm_page_size() - sizeof(int)));
3816
3817 Address index(noreg, tmp, Address::times_1);
3818 ExternalAddress page(os::get_memory_serialize_page());
3819
3820 // Size of store must match masking code above
3821 movl(as_Address(ArrayAddress(page, index)), tmp);
3822 }
3823
3824 // Calls to C land
3825 //
3826 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded
3827 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
3828 // has to be reset to 0. This is required to allow proper stack traversal.
3829 void MacroAssembler::set_last_Java_frame(Register java_thread,
3830 Register last_java_sp,
3831 Register last_java_fp,
3832 address last_java_pc) {
3833 // determine java_thread register
3834 if (!java_thread->is_valid()) {
3835 java_thread = rdi;
3836 get_thread(java_thread);
3837 }
3838 // determine last_java_sp register
3839 if (!last_java_sp->is_valid()) {
3840 last_java_sp = rsp;
3841 }
3842
3843 // last_java_fp is optional
3844
3845 if (last_java_fp->is_valid()) {
3846 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp);
3847 }
3848
3849 // last_java_pc is optional
3850
3851 if (last_java_pc != NULL) {
3852 lea(Address(java_thread,
3853 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()),
3854 InternalAddress(last_java_pc));
3855
3856 }
3857 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
3858 }
3859
3860 void MacroAssembler::shlptr(Register dst, int imm8) {
3861 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8));
3862 }
3863
3864 void MacroAssembler::shrptr(Register dst, int imm8) {
3865 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8));
3866 }
3867
3868 void MacroAssembler::sign_extend_byte(Register reg) {
3869 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) {
3870 movsbl(reg, reg); // movsxb
3871 } else {
3872 shll(reg, 24);
3873 sarl(reg, 24);
3874 }
3875 }
3876
3877 void MacroAssembler::sign_extend_short(Register reg) {
3878 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3879 movswl(reg, reg); // movsxw
3880 } else {
3881 shll(reg, 16);
3882 sarl(reg, 16);
3883 }
3884 }
3885
3886 void MacroAssembler::testl(Register dst, AddressLiteral src) {
3887 assert(reachable(src), "Address should be reachable");
3888 testl(dst, as_Address(src));
3889 }
3890
3891 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) {
3892 if (reachable(src)) {
3893 Assembler::sqrtsd(dst, as_Address(src));
3894 } else {
3895 lea(rscratch1, src);
3896 Assembler::sqrtsd(dst, Address(rscratch1, 0));
3897 }
3898 }
3899
3900 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) {
3901 if (reachable(src)) {
3902 Assembler::sqrtss(dst, as_Address(src));
3903 } else {
3904 lea(rscratch1, src);
3905 Assembler::sqrtss(dst, Address(rscratch1, 0));
3906 }
3907 }
3908
3909 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) {
3910 if (reachable(src)) {
3911 Assembler::subsd(dst, as_Address(src));
3912 } else {
3913 lea(rscratch1, src);
3914 Assembler::subsd(dst, Address(rscratch1, 0));
3915 }
3916 }
3917
3918 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
3919 if (reachable(src)) {
3920 Assembler::subss(dst, as_Address(src));
3921 } else {
3922 lea(rscratch1, src);
3923 Assembler::subss(dst, Address(rscratch1, 0));
3924 }
3925 }
3926
3927 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
3928 if (reachable(src)) {
3929 Assembler::ucomisd(dst, as_Address(src));
3930 } else {
3931 lea(rscratch1, src);
3932 Assembler::ucomisd(dst, Address(rscratch1, 0));
3933 }
3934 }
3935
3936 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
3937 if (reachable(src)) {
3938 Assembler::ucomiss(dst, as_Address(src));
3939 } else {
3940 lea(rscratch1, src);
3941 Assembler::ucomiss(dst, Address(rscratch1, 0));
3942 }
3943 }
3944
3945 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) {
3946 // Used in sign-bit flipping with aligned address.
3947 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3948 if (reachable(src)) {
3949 Assembler::xorpd(dst, as_Address(src));
3950 } else {
3951 lea(rscratch1, src);
3952 Assembler::xorpd(dst, Address(rscratch1, 0));
3953 }
3954 }
3955
3956 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) {
3957 // Used in sign-bit flipping with aligned address.
3958 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3959 if (reachable(src)) {
3960 Assembler::xorps(dst, as_Address(src));
3961 } else {
3962 lea(rscratch1, src);
3963 Assembler::xorps(dst, Address(rscratch1, 0));
3964 }
3965 }
3966
3967 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
3968 // Used in sign-bit flipping with aligned address.
3969 bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
3970 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
3971 if (reachable(src)) {
3972 Assembler::pshufb(dst, as_Address(src));
3973 } else {
3974 lea(rscratch1, src);
3975 Assembler::pshufb(dst, Address(rscratch1, 0));
3976 }
3977 }
3978
3979 // AVX 3-operands instructions
3980
3981 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3982 if (reachable(src)) {
3983 vaddsd(dst, nds, as_Address(src));
3984 } else {
3985 lea(rscratch1, src);
3986 vaddsd(dst, nds, Address(rscratch1, 0));
3987 }
3988 }
3989
3990 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3991 if (reachable(src)) {
3992 vaddss(dst, nds, as_Address(src));
3993 } else {
3994 lea(rscratch1, src);
3995 vaddss(dst, nds, Address(rscratch1, 0));
3996 }
3997 }
3998
3999 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) {
4000 if (reachable(src)) {
4001 vandpd(dst, nds, as_Address(src), vector256);
4002 } else {
4003 lea(rscratch1, src);
4004 vandpd(dst, nds, Address(rscratch1, 0), vector256);
4005 }
4006 }
4007
4008 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) {
4009 if (reachable(src)) {
4010 vandps(dst, nds, as_Address(src), vector256);
4011 } else {
4012 lea(rscratch1, src);
4013 vandps(dst, nds, Address(rscratch1, 0), vector256);
4014 }
4015 }
4016
4017 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4018 if (reachable(src)) {
4019 vdivsd(dst, nds, as_Address(src));
4020 } else {
4021 lea(rscratch1, src);
4022 vdivsd(dst, nds, Address(rscratch1, 0));
4023 }
4024 }
4025
4026 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4027 if (reachable(src)) {
4028 vdivss(dst, nds, as_Address(src));
4029 } else {
4030 lea(rscratch1, src);
4031 vdivss(dst, nds, Address(rscratch1, 0));
4032 }
4033 }
4034
4035 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4036 if (reachable(src)) {
4037 vmulsd(dst, nds, as_Address(src));
4038 } else {
4039 lea(rscratch1, src);
4040 vmulsd(dst, nds, Address(rscratch1, 0));
4041 }
4042 }
4043
4044 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4045 if (reachable(src)) {
4046 vmulss(dst, nds, as_Address(src));
4047 } else {
4048 lea(rscratch1, src);
4049 vmulss(dst, nds, Address(rscratch1, 0));
4050 }
4051 }
4052
4053 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4054 if (reachable(src)) {
4055 vsubsd(dst, nds, as_Address(src));
4056 } else {
4057 lea(rscratch1, src);
4058 vsubsd(dst, nds, Address(rscratch1, 0));
4059 }
4060 }
4061
4062 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4063 if (reachable(src)) {
4064 vsubss(dst, nds, as_Address(src));
4065 } else {
4066 lea(rscratch1, src);
4067 vsubss(dst, nds, Address(rscratch1, 0));
4068 }
4069 }
4070
4071 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) {
4072 if (reachable(src)) {
4073 vxorpd(dst, nds, as_Address(src), vector256);
4074 } else {
4075 lea(rscratch1, src);
4076 vxorpd(dst, nds, Address(rscratch1, 0), vector256);
4077 }
4078 }
4079
4080 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) {
4081 if (reachable(src)) {
4082 vxorps(dst, nds, as_Address(src), vector256);
4083 } else {
4084 lea(rscratch1, src);
4085 vxorps(dst, nds, Address(rscratch1, 0), vector256);
4086 }
4087 }
4088
4089
4090 //////////////////////////////////////////////////////////////////////////////////
4091 #if INCLUDE_ALL_GCS
4092
4093 void MacroAssembler::g1_write_barrier_pre(Register obj,
4094 Register pre_val,
4095 Register thread,
4096 Register tmp,
4097 bool tosca_live,
4098 bool expand_call) {
4099
4100 // If expand_call is true then we expand the call_VM_leaf macro
4101 // directly to skip generating the check by
4102 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
4103
4104 #ifdef _LP64
4105 assert(thread == r15_thread, "must be");
4106 #endif // _LP64
4107
4108 Label done;
4109 Label runtime;
4110
4111 assert(pre_val != noreg, "check this code");
4112
4113 if (obj != noreg) {
4114 assert_different_registers(obj, pre_val, tmp);
4115 assert(pre_val != rax, "check this code");
4116 }
4117
4118 Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
4119 PtrQueue::byte_offset_of_active()));
4120 Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
4121 PtrQueue::byte_offset_of_index()));
4122 Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
4123 PtrQueue::byte_offset_of_buf()));
4124
4125
4126 // Is marking active?
4127 if (in_bytes(PtrQueue::byte_width_of_active()) == 4) {
4128 cmpl(in_progress, 0);
4129 } else {
4130 assert(in_bytes(PtrQueue::byte_width_of_active()) == 1, "Assumption");
4131 cmpb(in_progress, 0);
4132 }
4133 jcc(Assembler::equal, done);
4134
4135 // Do we need to load the previous value?
4136 if (obj != noreg) {
4137 load_heap_oop(pre_val, Address(obj, 0));
4138 }
4139
4140 // Is the previous value null?
4141 cmpptr(pre_val, (int32_t) NULL_WORD);
4142 jcc(Assembler::equal, done);
4143
4144 // Can we store original value in the thread's buffer?
4145 // Is index == 0?
4146 // (The index field is typed as size_t.)
4147
4148 movptr(tmp, index); // tmp := *index_adr
4149 cmpptr(tmp, 0); // tmp == 0?
4150 jcc(Assembler::equal, runtime); // If yes, goto runtime
4151
4152 subptr(tmp, wordSize); // tmp := tmp - wordSize
4153 movptr(index, tmp); // *index_adr := tmp
4154 addptr(tmp, buffer); // tmp := tmp + *buffer_adr
4155
4156 // Record the previous value
4157 movptr(Address(tmp, 0), pre_val);
4158 jmp(done);
4159
4160 bind(runtime);
4161 // save the live input values
4162 if(tosca_live) push(rax);
4163
4164 if (obj != noreg && obj != rax)
4165 push(obj);
4166
4167 if (pre_val != rax)
4168 push(pre_val);
4169
4170 // Calling the runtime using the regular call_VM_leaf mechanism generates
4171 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
4172 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
4173 //
4174 // If we care generating the pre-barrier without a frame (e.g. in the
4175 // intrinsified Reference.get() routine) then ebp might be pointing to
4176 // the caller frame and so this check will most likely fail at runtime.
4177 //
4178 // Expanding the call directly bypasses the generation of the check.
4179 // So when we do not have have a full interpreter frame on the stack
4180 // expand_call should be passed true.
4181
4182 NOT_LP64( push(thread); )
4183
4184 if (expand_call) {
4185 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
4186 pass_arg1(this, thread);
4187 pass_arg0(this, pre_val);
4188 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2);
4189 } else {
4190 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread);
4191 }
4192
4193 NOT_LP64( pop(thread); )
4194
4195 // save the live input values
4196 if (pre_val != rax)
4197 pop(pre_val);
4198
4199 if (obj != noreg && obj != rax)
4200 pop(obj);
4201
4202 if(tosca_live) pop(rax);
4203
4204 bind(done);
4205 }
4206
4207 void MacroAssembler::g1_write_barrier_post(Register store_addr,
4208 Register new_val,
4209 Register thread,
4210 Register tmp,
4211 Register tmp2) {
4212 #ifdef _LP64
4213 assert(thread == r15_thread, "must be");
4214 #endif // _LP64
4215
4216 Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() +
4217 PtrQueue::byte_offset_of_index()));
4218 Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() +
4219 PtrQueue::byte_offset_of_buf()));
4220
4221 CardTableModRefBS* ct =
4222 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
4223 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
4224
4225 Label done;
4226 Label runtime;
4227
4228 // Does store cross heap regions?
4229
4230 movptr(tmp, store_addr);
4231 xorptr(tmp, new_val);
4232 shrptr(tmp, HeapRegion::LogOfHRGrainBytes);
4233 jcc(Assembler::equal, done);
4234
4235 // crosses regions, storing NULL?
4236
4237 cmpptr(new_val, (int32_t) NULL_WORD);
4238 jcc(Assembler::equal, done);
4239
4240 // storing region crossing non-NULL, is card already dirty?
4241
4242 const Register card_addr = tmp;
4243 const Register cardtable = tmp2;
4244
4245 movptr(card_addr, store_addr);
4246 shrptr(card_addr, CardTableModRefBS::card_shift);
4247 // Do not use ExternalAddress to load 'byte_map_base', since 'byte_map_base' is NOT
4248 // a valid address and therefore is not properly handled by the relocation code.
4249 movptr(cardtable, (intptr_t)ct->byte_map_base);
4250 addptr(card_addr, cardtable);
4251
4252 cmpb(Address(card_addr, 0), (int)G1SATBCardTableModRefBS::g1_young_card_val());
4253 jcc(Assembler::equal, done);
4254
4255 membar(Assembler::Membar_mask_bits(Assembler::StoreLoad));
4256 cmpb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val());
4257 jcc(Assembler::equal, done);
4258
4259
4260 // storing a region crossing, non-NULL oop, card is clean.
4261 // dirty card and log.
4262
4263 movb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val());
4264
4265 cmpl(queue_index, 0);
4266 jcc(Assembler::equal, runtime);
4267 subl(queue_index, wordSize);
4268 movptr(tmp2, buffer);
4269 #ifdef _LP64
4270 movslq(rscratch1, queue_index);
4271 addq(tmp2, rscratch1);
4272 movq(Address(tmp2, 0), card_addr);
4273 #else
4274 addl(tmp2, queue_index);
4275 movl(Address(tmp2, 0), card_addr);
4276 #endif
4277 jmp(done);
4278
4279 bind(runtime);
4280 // save the live input values
4281 push(store_addr);
4282 push(new_val);
4283 #ifdef _LP64
4284 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread);
4285 #else
4286 push(thread);
4287 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread);
4288 pop(thread);
4289 #endif
4290 pop(new_val);
4291 pop(store_addr);
4292
4293 bind(done);
4294 }
4295
4296 #endif // INCLUDE_ALL_GCS
4297 //////////////////////////////////////////////////////////////////////////////////
4298
4299
4300 void MacroAssembler::store_check(Register obj, Address dst) {
4301 store_check(obj);
4302 }
4303
4304 void MacroAssembler::store_check(Register obj) {
4305 // Does a store check for the oop in register obj. The content of
4306 // register obj is destroyed afterwards.
4307
4308 BarrierSet* bs = Universe::heap()->barrier_set();
4309 assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind");
4310
4311 CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
4312 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
4313
4314 shrptr(obj, CardTableModRefBS::card_shift);
4315
4316 Address card_addr;
4317
4318 // The calculation for byte_map_base is as follows:
4319 // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift);
4320 // So this essentially converts an address to a displacement and it will
4321 // never need to be relocated. On 64bit however the value may be too
4322 // large for a 32bit displacement.
4323 intptr_t disp = (intptr_t) ct->byte_map_base;
4324 if (is_simm32(disp)) {
4325 card_addr = Address(noreg, obj, Address::times_1, disp);
4326 } else {
4327 // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative
4328 // displacement and done in a single instruction given favorable mapping and a
4329 // smarter version of as_Address. However, 'ExternalAddress' generates a relocation
4330 // entry and that entry is not properly handled by the relocation code.
4331 AddressLiteral cardtable((address)ct->byte_map_base, relocInfo::none);
4332 Address index(noreg, obj, Address::times_1);
4333 card_addr = as_Address(ArrayAddress(cardtable, index));
4334 }
4335
4336 assert(card_addr != null, "sanity");
4337
4338 int dirty = CardTableModRefBS::dirty_card_val();
4339 if (UseCondCardMark) {
4340 Label L_already_dirty;
4341 cmpb(card_addr, dirty);
4342 jcc(Assembler::equal, L_already_dirty);
4343 movb(card_addr, dirty);
4344 bind(L_already_dirty);
4345 } else {
4346 movb(card_addr, dirty);
4347 }
4348 }
4349
4350 void MacroAssembler::subptr(Register dst, int32_t imm32) {
4351 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32));
4352 }
4353
4354 // Force generation of a 4 byte immediate value even if it fits into 8bit
4355 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) {
4356 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32));
4357 }
4358
4359 void MacroAssembler::subptr(Register dst, Register src) {
4360 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src));
4361 }
4362
4363 // C++ bool manipulation
4364 void MacroAssembler::testbool(Register dst) {
4365 if(sizeof(bool) == 1)
4366 testb(dst, 0xff);
4367 else if(sizeof(bool) == 2) {
4368 // testw implementation needed for two byte bools
4369 ShouldNotReachHere();
4370 } else if(sizeof(bool) == 4)
4371 testl(dst, dst);
4372 else
4373 // unsupported
4374 ShouldNotReachHere();
4375 }
4376
4377 void MacroAssembler::testptr(Register dst, Register src) {
4378 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src));
4379 }
4380
4381 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
4382 void MacroAssembler::tlab_allocate(Register obj,
4383 Register var_size_in_bytes,
4384 int con_size_in_bytes,
4385 Register t1,
4386 Register t2,
4387 Label& slow_case) {
4388 assert_different_registers(obj, t1, t2);
4389 assert_different_registers(obj, var_size_in_bytes, t1);
4390 Register end = t2;
4391 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread);
4392
4393 verify_tlab();
4394
4395 NOT_LP64(get_thread(thread));
4396
4397 movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
4398 if (var_size_in_bytes == noreg) {
4399 lea(end, Address(obj, con_size_in_bytes));
4400 } else {
4401 lea(end, Address(obj, var_size_in_bytes, Address::times_1));
4402 }
4403 cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
4404 jcc(Assembler::above, slow_case);
4405
4406 // update the tlab top pointer
4407 movptr(Address(thread, JavaThread::tlab_top_offset()), end);
4408
4409 // recover var_size_in_bytes if necessary
4410 if (var_size_in_bytes == end) {
4411 subptr(var_size_in_bytes, obj);
4412 }
4413 verify_tlab();
4414 }
4415
4416 // Preserves rbx, and rdx.
4417 Register MacroAssembler::tlab_refill(Label& retry,
4418 Label& try_eden,
4419 Label& slow_case) {
4420 Register top = rax;
4421 Register t1 = rcx;
4422 Register t2 = rsi;
4423 Register thread_reg = NOT_LP64(rdi) LP64_ONLY(r15_thread);
4424 assert_different_registers(top, thread_reg, t1, t2, /* preserve: */ rbx, rdx);
4425 Label do_refill, discard_tlab;
4426
4427 if (!Universe::heap()->supports_inline_contig_alloc()) {
4428 // No allocation in the shared eden.
4429 jmp(slow_case);
4430 }
4431
4432 NOT_LP64(get_thread(thread_reg));
4433
4434 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
4435 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
4436
4437 // calculate amount of free space
4438 subptr(t1, top);
4439 shrptr(t1, LogHeapWordSize);
4440
4441 // Retain tlab and allocate object in shared space if
4442 // the amount free in the tlab is too large to discard.
4443 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())));
4444 jcc(Assembler::lessEqual, discard_tlab);
4445
4446 // Retain
4447 // %%% yuck as movptr...
4448 movptr(t2, (int32_t) ThreadLocalAllocBuffer::refill_waste_limit_increment());
4449 addptr(Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())), t2);
4450 if (TLABStats) {
4451 // increment number of slow_allocations
4452 addl(Address(thread_reg, in_bytes(JavaThread::tlab_slow_allocations_offset())), 1);
4453 }
4454 jmp(try_eden);
4455
4456 bind(discard_tlab);
4457 if (TLABStats) {
4458 // increment number of refills
4459 addl(Address(thread_reg, in_bytes(JavaThread::tlab_number_of_refills_offset())), 1);
4460 // accumulate wastage -- t1 is amount free in tlab
4461 addl(Address(thread_reg, in_bytes(JavaThread::tlab_fast_refill_waste_offset())), t1);
4462 }
4463
4464 // if tlab is currently allocated (top or end != null) then
4465 // fill [top, end + alignment_reserve) with array object
4466 testptr(top, top);
4467 jcc(Assembler::zero, do_refill);
4468
4469 // set up the mark word
4470 movptr(Address(top, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::prototype()->copy_set_hash(0x2));
4471 // set the length to the remaining space
4472 subptr(t1, typeArrayOopDesc::header_size(T_INT));
4473 addptr(t1, (int32_t)ThreadLocalAllocBuffer::alignment_reserve());
4474 shlptr(t1, log2_intptr(HeapWordSize/sizeof(jint)));
4475 movl(Address(top, arrayOopDesc::length_offset_in_bytes()), t1);
4476 // set klass to intArrayKlass
4477 // dubious reloc why not an oop reloc?
4478 movptr(t1, ExternalAddress((address)Universe::intArrayKlassObj_addr()));
4479 // store klass last. concurrent gcs assumes klass length is valid if
4480 // klass field is not null.
4481 store_klass(top, t1);
4482
4483 movptr(t1, top);
4484 subptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
4485 incr_allocated_bytes(thread_reg, t1, 0);
4486
4487 // refill the tlab with an eden allocation
4488 bind(do_refill);
4489 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset())));
4490 shlptr(t1, LogHeapWordSize);
4491 // allocate new tlab, address returned in top
4492 eden_allocate(top, t1, 0, t2, slow_case);
4493
4494 // Check that t1 was preserved in eden_allocate.
4495 #ifdef ASSERT
4496 if (UseTLAB) {
4497 Label ok;
4498 Register tsize = rsi;
4499 assert_different_registers(tsize, thread_reg, t1);
4500 push(tsize);
4501 movptr(tsize, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset())));
4502 shlptr(tsize, LogHeapWordSize);
4503 cmpptr(t1, tsize);
4504 jcc(Assembler::equal, ok);
4505 STOP("assert(t1 != tlab size)");
4506 should_not_reach_here();
4507
4508 bind(ok);
4509 pop(tsize);
4510 }
4511 #endif
4512 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())), top);
4513 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())), top);
4514 addptr(top, t1);
4515 subptr(top, (int32_t)ThreadLocalAllocBuffer::alignment_reserve_in_bytes());
4516 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())), top);
4517 verify_tlab();
4518 jmp(retry);
4519
4520 return thread_reg; // for use by caller
4521 }
4522
4523 void MacroAssembler::incr_allocated_bytes(Register thread,
4524 Register var_size_in_bytes,
4525 int con_size_in_bytes,
4526 Register t1) {
4527 if (!thread->is_valid()) {
4528 #ifdef _LP64
4529 thread = r15_thread;
4530 #else
4531 assert(t1->is_valid(), "need temp reg");
4532 thread = t1;
4533 get_thread(thread);
4534 #endif
4535 }
4536
4537 #ifdef _LP64
4538 if (var_size_in_bytes->is_valid()) {
4539 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
4540 } else {
4541 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
4542 }
4543 #else
4544 if (var_size_in_bytes->is_valid()) {
4545 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
4546 } else {
4547 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
4548 }
4549 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0);
4550 #endif
4551 }
4552
4553 void MacroAssembler::fp_runtime_fallback(address runtime_entry, int nb_args, int num_fpu_regs_in_use) {
4554 pusha();
4555
4556 // if we are coming from c1, xmm registers may be live
4557 int off = 0;
4558 if (UseSSE == 1) {
4559 subptr(rsp, sizeof(jdouble)*8);
4560 movflt(Address(rsp,off++*sizeof(jdouble)),xmm0);
4561 movflt(Address(rsp,off++*sizeof(jdouble)),xmm1);
4562 movflt(Address(rsp,off++*sizeof(jdouble)),xmm2);
4563 movflt(Address(rsp,off++*sizeof(jdouble)),xmm3);
4564 movflt(Address(rsp,off++*sizeof(jdouble)),xmm4);
4565 movflt(Address(rsp,off++*sizeof(jdouble)),xmm5);
4566 movflt(Address(rsp,off++*sizeof(jdouble)),xmm6);
4567 movflt(Address(rsp,off++*sizeof(jdouble)),xmm7);
4568 } else if (UseSSE >= 2) {
4569 #ifdef COMPILER2
4570 if (MaxVectorSize > 16) {
4571 assert(UseAVX > 0, "256bit vectors are supported only with AVX");
4572 // Save upper half of YMM registes
4573 subptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
4574 vextractf128h(Address(rsp, 0),xmm0);
4575 vextractf128h(Address(rsp, 16),xmm1);
4576 vextractf128h(Address(rsp, 32),xmm2);
4577 vextractf128h(Address(rsp, 48),xmm3);
4578 vextractf128h(Address(rsp, 64),xmm4);
4579 vextractf128h(Address(rsp, 80),xmm5);
4580 vextractf128h(Address(rsp, 96),xmm6);
4581 vextractf128h(Address(rsp,112),xmm7);
4582 #ifdef _LP64
4583 vextractf128h(Address(rsp,128),xmm8);
4584 vextractf128h(Address(rsp,144),xmm9);
4585 vextractf128h(Address(rsp,160),xmm10);
4586 vextractf128h(Address(rsp,176),xmm11);
4587 vextractf128h(Address(rsp,192),xmm12);
4588 vextractf128h(Address(rsp,208),xmm13);
4589 vextractf128h(Address(rsp,224),xmm14);
4590 vextractf128h(Address(rsp,240),xmm15);
4591 #endif
4592 }
4593 #endif
4594 // Save whole 128bit (16 bytes) XMM regiters
4595 subptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
4596 movdqu(Address(rsp,off++*16),xmm0);
4597 movdqu(Address(rsp,off++*16),xmm1);
4598 movdqu(Address(rsp,off++*16),xmm2);
4599 movdqu(Address(rsp,off++*16),xmm3);
4600 movdqu(Address(rsp,off++*16),xmm4);
4601 movdqu(Address(rsp,off++*16),xmm5);
4602 movdqu(Address(rsp,off++*16),xmm6);
4603 movdqu(Address(rsp,off++*16),xmm7);
4604 #ifdef _LP64
4605 movdqu(Address(rsp,off++*16),xmm8);
4606 movdqu(Address(rsp,off++*16),xmm9);
4607 movdqu(Address(rsp,off++*16),xmm10);
4608 movdqu(Address(rsp,off++*16),xmm11);
4609 movdqu(Address(rsp,off++*16),xmm12);
4610 movdqu(Address(rsp,off++*16),xmm13);
4611 movdqu(Address(rsp,off++*16),xmm14);
4612 movdqu(Address(rsp,off++*16),xmm15);
4613 #endif
4614 }
4615
4616 // Preserve registers across runtime call
4617 int incoming_argument_and_return_value_offset = -1;
4618 if (num_fpu_regs_in_use > 1) {
4619 // Must preserve all other FPU regs (could alternatively convert
4620 // SharedRuntime::dsin, dcos etc. into assembly routines known not to trash
4621 // FPU state, but can not trust C compiler)
4622 NEEDS_CLEANUP;
4623 // NOTE that in this case we also push the incoming argument(s) to
4624 // the stack and restore it later; we also use this stack slot to
4625 // hold the return value from dsin, dcos etc.
4626 for (int i = 0; i < num_fpu_regs_in_use; i++) {
4627 subptr(rsp, sizeof(jdouble));
4628 fstp_d(Address(rsp, 0));
4629 }
4630 incoming_argument_and_return_value_offset = sizeof(jdouble)*(num_fpu_regs_in_use-1);
4631 for (int i = nb_args-1; i >= 0; i--) {
4632 fld_d(Address(rsp, incoming_argument_and_return_value_offset-i*sizeof(jdouble)));
4633 }
4634 }
4635
4636 subptr(rsp, nb_args*sizeof(jdouble));
4637 for (int i = 0; i < nb_args; i++) {
4638 fstp_d(Address(rsp, i*sizeof(jdouble)));
4639 }
4640
4641 #ifdef _LP64
4642 if (nb_args > 0) {
4643 movdbl(xmm0, Address(rsp, 0));
4644 }
4645 if (nb_args > 1) {
4646 movdbl(xmm1, Address(rsp, sizeof(jdouble)));
4647 }
4648 assert(nb_args <= 2, "unsupported number of args");
4649 #endif // _LP64
4650
4651 // NOTE: we must not use call_VM_leaf here because that requires a
4652 // complete interpreter frame in debug mode -- same bug as 4387334
4653 // MacroAssembler::call_VM_leaf_base is perfectly safe and will
4654 // do proper 64bit abi
4655
4656 NEEDS_CLEANUP;
4657 // Need to add stack banging before this runtime call if it needs to
4658 // be taken; however, there is no generic stack banging routine at
4659 // the MacroAssembler level
4660
4661 MacroAssembler::call_VM_leaf_base(runtime_entry, 0);
4662
4663 #ifdef _LP64
4664 movsd(Address(rsp, 0), xmm0);
4665 fld_d(Address(rsp, 0));
4666 #endif // _LP64
4667 addptr(rsp, sizeof(jdouble) * nb_args);
4668 if (num_fpu_regs_in_use > 1) {
4669 // Must save return value to stack and then restore entire FPU
4670 // stack except incoming arguments
4671 fstp_d(Address(rsp, incoming_argument_and_return_value_offset));
4672 for (int i = 0; i < num_fpu_regs_in_use - nb_args; i++) {
4673 fld_d(Address(rsp, 0));
4674 addptr(rsp, sizeof(jdouble));
4675 }
4676 fld_d(Address(rsp, (nb_args-1)*sizeof(jdouble)));
4677 addptr(rsp, sizeof(jdouble) * nb_args);
4678 }
4679
4680 off = 0;
4681 if (UseSSE == 1) {
4682 movflt(xmm0, Address(rsp,off++*sizeof(jdouble)));
4683 movflt(xmm1, Address(rsp,off++*sizeof(jdouble)));
4684 movflt(xmm2, Address(rsp,off++*sizeof(jdouble)));
4685 movflt(xmm3, Address(rsp,off++*sizeof(jdouble)));
4686 movflt(xmm4, Address(rsp,off++*sizeof(jdouble)));
4687 movflt(xmm5, Address(rsp,off++*sizeof(jdouble)));
4688 movflt(xmm6, Address(rsp,off++*sizeof(jdouble)));
4689 movflt(xmm7, Address(rsp,off++*sizeof(jdouble)));
4690 addptr(rsp, sizeof(jdouble)*8);
4691 } else if (UseSSE >= 2) {
4692 // Restore whole 128bit (16 bytes) XMM regiters
4693 movdqu(xmm0, Address(rsp,off++*16));
4694 movdqu(xmm1, Address(rsp,off++*16));
4695 movdqu(xmm2, Address(rsp,off++*16));
4696 movdqu(xmm3, Address(rsp,off++*16));
4697 movdqu(xmm4, Address(rsp,off++*16));
4698 movdqu(xmm5, Address(rsp,off++*16));
4699 movdqu(xmm6, Address(rsp,off++*16));
4700 movdqu(xmm7, Address(rsp,off++*16));
4701 #ifdef _LP64
4702 movdqu(xmm8, Address(rsp,off++*16));
4703 movdqu(xmm9, Address(rsp,off++*16));
4704 movdqu(xmm10, Address(rsp,off++*16));
4705 movdqu(xmm11, Address(rsp,off++*16));
4706 movdqu(xmm12, Address(rsp,off++*16));
4707 movdqu(xmm13, Address(rsp,off++*16));
4708 movdqu(xmm14, Address(rsp,off++*16));
4709 movdqu(xmm15, Address(rsp,off++*16));
4710 #endif
4711 addptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
4712 #ifdef COMPILER2
4713 if (MaxVectorSize > 16) {
4714 // Restore upper half of YMM registes.
4715 vinsertf128h(xmm0, Address(rsp, 0));
4716 vinsertf128h(xmm1, Address(rsp, 16));
4717 vinsertf128h(xmm2, Address(rsp, 32));
4718 vinsertf128h(xmm3, Address(rsp, 48));
4719 vinsertf128h(xmm4, Address(rsp, 64));
4720 vinsertf128h(xmm5, Address(rsp, 80));
4721 vinsertf128h(xmm6, Address(rsp, 96));
4722 vinsertf128h(xmm7, Address(rsp,112));
4723 #ifdef _LP64
4724 vinsertf128h(xmm8, Address(rsp,128));
4725 vinsertf128h(xmm9, Address(rsp,144));
4726 vinsertf128h(xmm10, Address(rsp,160));
4727 vinsertf128h(xmm11, Address(rsp,176));
4728 vinsertf128h(xmm12, Address(rsp,192));
4729 vinsertf128h(xmm13, Address(rsp,208));
4730 vinsertf128h(xmm14, Address(rsp,224));
4731 vinsertf128h(xmm15, Address(rsp,240));
4732 #endif
4733 addptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8));
4734 }
4735 #endif
4736 }
4737 popa();
4738 }
4739
4740 static const double pi_4 = 0.7853981633974483;
4741
4742 void MacroAssembler::trigfunc(char trig, int num_fpu_regs_in_use) {
4743 // A hand-coded argument reduction for values in fabs(pi/4, pi/2)
4744 // was attempted in this code; unfortunately it appears that the
4745 // switch to 80-bit precision and back causes this to be
4746 // unprofitable compared with simply performing a runtime call if
4747 // the argument is out of the (-pi/4, pi/4) range.
4748
4749 Register tmp = noreg;
4750 if (!VM_Version::supports_cmov()) {
4751 // fcmp needs a temporary so preserve rbx,
4752 tmp = rbx;
4753 push(tmp);
4754 }
4755
4756 Label slow_case, done;
4757
4758 ExternalAddress pi4_adr = (address)&pi_4;
4759 if (reachable(pi4_adr)) {
4760 // x ?<= pi/4
4761 fld_d(pi4_adr);
4762 fld_s(1); // Stack: X PI/4 X
4763 fabs(); // Stack: |X| PI/4 X
4764 fcmp(tmp);
4765 jcc(Assembler::above, slow_case);
4766
4767 // fastest case: -pi/4 <= x <= pi/4
4768 switch(trig) {
4769 case 's':
4770 fsin();
4771 break;
4772 case 'c':
4773 fcos();
4774 break;
4775 case 't':
4776 ftan();
4777 break;
4778 default:
4779 assert(false, "bad intrinsic");
4780 break;
4781 }
4782 jmp(done);
4783 }
4784
4785 // slow case: runtime call
4786 bind(slow_case);
4787
4788 switch(trig) {
4789 case 's':
4790 {
4791 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1, num_fpu_regs_in_use);
4792 }
4793 break;
4794 case 'c':
4795 {
4796 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1, num_fpu_regs_in_use);
4797 }
4798 break;
4799 case 't':
4800 {
4801 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1, num_fpu_regs_in_use);
4802 }
4803 break;
4804 default:
4805 assert(false, "bad intrinsic");
4806 break;
4807 }
4808
4809 // Come here with result in F-TOS
4810 bind(done);
4811
4812 if (tmp != noreg) {
4813 pop(tmp);
4814 }
4815 }
4816
4817
4818 // Look up the method for a megamorphic invokeinterface call.
4819 // The target method is determined by <intf_klass, itable_index>.
4820 // The receiver klass is in recv_klass.
4821 // On success, the result will be in method_result, and execution falls through.
4822 // On failure, execution transfers to the given label.
4823 void MacroAssembler::lookup_interface_method(Register recv_klass,
4824 Register intf_klass,
4825 RegisterOrConstant itable_index,
4826 Register method_result,
4827 Register scan_temp,
4828 Label& L_no_such_interface) {
4829 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
4830 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
4831 "caller must use same register for non-constant itable index as for method");
4832
4833 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
4834 int vtable_base = InstanceKlass::vtable_start_offset() * wordSize;
4835 int itentry_off = itableMethodEntry::method_offset_in_bytes();
4836 int scan_step = itableOffsetEntry::size() * wordSize;
4837 int vte_size = vtableEntry::size() * wordSize;
4838 Address::ScaleFactor times_vte_scale = Address::times_ptr;
4839 assert(vte_size == wordSize, "else adjust times_vte_scale");
4840
4841 movl(scan_temp, Address(recv_klass, InstanceKlass::vtable_length_offset() * wordSize));
4842
4843 // %%% Could store the aligned, prescaled offset in the klassoop.
4844 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
4845 if (HeapWordsPerLong > 1) {
4846 // Round up to align_object_offset boundary
4847 // see code for InstanceKlass::start_of_itable!
4848 round_to(scan_temp, BytesPerLong);
4849 }
4850
4851 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
4852 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
4853 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
4854
4855 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
4856 // if (scan->interface() == intf) {
4857 // result = (klass + scan->offset() + itable_index);
4858 // }
4859 // }
4860 Label search, found_method;
4861
4862 for (int peel = 1; peel >= 0; peel--) {
4863 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
4864 cmpptr(intf_klass, method_result);
4865
4866 if (peel) {
4867 jccb(Assembler::equal, found_method);
4868 } else {
4869 jccb(Assembler::notEqual, search);
4870 // (invert the test to fall through to found_method...)
4871 }
4872
4873 if (!peel) break;
4874
4875 bind(search);
4876
4877 // Check that the previous entry is non-null. A null entry means that
4878 // the receiver class doesn't implement the interface, and wasn't the
4879 // same as when the caller was compiled.
4880 testptr(method_result, method_result);
4881 jcc(Assembler::zero, L_no_such_interface);
4882 addptr(scan_temp, scan_step);
4883 }
4884
4885 bind(found_method);
4886
4887 // Got a hit.
4888 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
4889 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
4890 }
4891
4892
4893 // virtual method calling
4894 void MacroAssembler::lookup_virtual_method(Register recv_klass,
4895 RegisterOrConstant vtable_index,
4896 Register method_result) {
4897 const int base = InstanceKlass::vtable_start_offset() * wordSize;
4898 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below");
4899 Address vtable_entry_addr(recv_klass,
4900 vtable_index, Address::times_ptr,
4901 base + vtableEntry::method_offset_in_bytes());
4902 movptr(method_result, vtable_entry_addr);
4903 }
4904
4905
4906 void MacroAssembler::check_klass_subtype(Register sub_klass,
4907 Register super_klass,
4908 Register temp_reg,
4909 Label& L_success) {
4910 Label L_failure;
4911 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL);
4912 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
4913 bind(L_failure);
4914 }
4915
4916
4917 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
4918 Register super_klass,
4919 Register temp_reg,
4920 Label* L_success,
4921 Label* L_failure,
4922 Label* L_slow_path,
4923 RegisterOrConstant super_check_offset) {
4924 assert_different_registers(sub_klass, super_klass, temp_reg);
4925 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
4926 if (super_check_offset.is_register()) {
4927 assert_different_registers(sub_klass, super_klass,
4928 super_check_offset.as_register());
4929 } else if (must_load_sco) {
4930 assert(temp_reg != noreg, "supply either a temp or a register offset");
4931 }
4932
4933 Label L_fallthrough;
4934 int label_nulls = 0;
4935 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
4936 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
4937 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
4938 assert(label_nulls <= 1, "at most one NULL in the batch");
4939
4940 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
4941 int sco_offset = in_bytes(Klass::super_check_offset_offset());
4942 Address super_check_offset_addr(super_klass, sco_offset);
4943
4944 // Hacked jcc, which "knows" that L_fallthrough, at least, is in
4945 // range of a jccb. If this routine grows larger, reconsider at
4946 // least some of these.
4947 #define local_jcc(assembler_cond, label) \
4948 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \
4949 else jcc( assembler_cond, label) /*omit semi*/
4950
4951 // Hacked jmp, which may only be used just before L_fallthrough.
4952 #define final_jmp(label) \
4953 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
4954 else jmp(label) /*omit semi*/
4955
4956 // If the pointers are equal, we are done (e.g., String[] elements).
4957 // This self-check enables sharing of secondary supertype arrays among
4958 // non-primary types such as array-of-interface. Otherwise, each such
4959 // type would need its own customized SSA.
4960 // We move this check to the front of the fast path because many
4961 // type checks are in fact trivially successful in this manner,
4962 // so we get a nicely predicted branch right at the start of the check.
4963 cmpptr(sub_klass, super_klass);
4964 local_jcc(Assembler::equal, *L_success);
4965
4966 // Check the supertype display:
4967 if (must_load_sco) {
4968 // Positive movl does right thing on LP64.
4969 movl(temp_reg, super_check_offset_addr);
4970 super_check_offset = RegisterOrConstant(temp_reg);
4971 }
4972 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0);
4973 cmpptr(super_klass, super_check_addr); // load displayed supertype
4974
4975 // This check has worked decisively for primary supers.
4976 // Secondary supers are sought in the super_cache ('super_cache_addr').
4977 // (Secondary supers are interfaces and very deeply nested subtypes.)
4978 // This works in the same check above because of a tricky aliasing
4979 // between the super_cache and the primary super display elements.
4980 // (The 'super_check_addr' can address either, as the case requires.)
4981 // Note that the cache is updated below if it does not help us find
4982 // what we need immediately.
4983 // So if it was a primary super, we can just fail immediately.
4984 // Otherwise, it's the slow path for us (no success at this point).
4985
4986 if (super_check_offset.is_register()) {
4987 local_jcc(Assembler::equal, *L_success);
4988 cmpl(super_check_offset.as_register(), sc_offset);
4989 if (L_failure == &L_fallthrough) {
4990 local_jcc(Assembler::equal, *L_slow_path);
4991 } else {
4992 local_jcc(Assembler::notEqual, *L_failure);
4993 final_jmp(*L_slow_path);
4994 }
4995 } else if (super_check_offset.as_constant() == sc_offset) {
4996 // Need a slow path; fast failure is impossible.
4997 if (L_slow_path == &L_fallthrough) {
4998 local_jcc(Assembler::equal, *L_success);
4999 } else {
5000 local_jcc(Assembler::notEqual, *L_slow_path);
5001 final_jmp(*L_success);
5002 }
5003 } else {
5004 // No slow path; it's a fast decision.
5005 if (L_failure == &L_fallthrough) {
5006 local_jcc(Assembler::equal, *L_success);
5007 } else {
5008 local_jcc(Assembler::notEqual, *L_failure);
5009 final_jmp(*L_success);
5010 }
5011 }
5012
5013 bind(L_fallthrough);
5014
5015 #undef local_jcc
5016 #undef final_jmp
5017 }
5018
5019
5020 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
5021 Register super_klass,
5022 Register temp_reg,
5023 Register temp2_reg,
5024 Label* L_success,
5025 Label* L_failure,
5026 bool set_cond_codes) {
5027 assert_different_registers(sub_klass, super_klass, temp_reg);
5028 if (temp2_reg != noreg)
5029 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg);
5030 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
5031
5032 Label L_fallthrough;
5033 int label_nulls = 0;
5034 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
5035 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
5036 assert(label_nulls <= 1, "at most one NULL in the batch");
5037
5038 // a couple of useful fields in sub_klass:
5039 int ss_offset = in_bytes(Klass::secondary_supers_offset());
5040 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
5041 Address secondary_supers_addr(sub_klass, ss_offset);
5042 Address super_cache_addr( sub_klass, sc_offset);
5043
5044 // Do a linear scan of the secondary super-klass chain.
5045 // This code is rarely used, so simplicity is a virtue here.
5046 // The repne_scan instruction uses fixed registers, which we must spill.
5047 // Don't worry too much about pre-existing connections with the input regs.
5048
5049 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super)
5050 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter)
5051
5052 // Get super_klass value into rax (even if it was in rdi or rcx).
5053 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false;
5054 if (super_klass != rax || UseCompressedOops) {
5055 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; }
5056 mov(rax, super_klass);
5057 }
5058 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; }
5059 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; }
5060
5061 #ifndef PRODUCT
5062 int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
5063 ExternalAddress pst_counter_addr((address) pst_counter);
5064 NOT_LP64( incrementl(pst_counter_addr) );
5065 LP64_ONLY( lea(rcx, pst_counter_addr) );
5066 LP64_ONLY( incrementl(Address(rcx, 0)) );
5067 #endif //PRODUCT
5068
5069 // We will consult the secondary-super array.
5070 movptr(rdi, secondary_supers_addr);
5071 // Load the array length. (Positive movl does right thing on LP64.)
5072 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes()));
5073 // Skip to start of data.
5074 addptr(rdi, Array<Klass*>::base_offset_in_bytes());
5075
5076 // Scan RCX words at [RDI] for an occurrence of RAX.
5077 // Set NZ/Z based on last compare.
5078 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does
5079 // not change flags (only scas instruction which is repeated sets flags).
5080 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found.
5081
5082 testptr(rax,rax); // Set Z = 0
5083 repne_scan();
5084
5085 // Unspill the temp. registers:
5086 if (pushed_rdi) pop(rdi);
5087 if (pushed_rcx) pop(rcx);
5088 if (pushed_rax) pop(rax);
5089
5090 if (set_cond_codes) {
5091 // Special hack for the AD files: rdi is guaranteed non-zero.
5092 assert(!pushed_rdi, "rdi must be left non-NULL");
5093 // Also, the condition codes are properly set Z/NZ on succeed/failure.
5094 }
5095
5096 if (L_failure == &L_fallthrough)
5097 jccb(Assembler::notEqual, *L_failure);
5098 else jcc(Assembler::notEqual, *L_failure);
5099
5100 // Success. Cache the super we found and proceed in triumph.
5101 movptr(super_cache_addr, super_klass);
5102
5103 if (L_success != &L_fallthrough) {
5104 jmp(*L_success);
5105 }
5106
5107 #undef IS_A_TEMP
5108
5109 bind(L_fallthrough);
5110 }
5111
5112
5113 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
5114 if (VM_Version::supports_cmov()) {
5115 cmovl(cc, dst, src);
5116 } else {
5117 Label L;
5118 jccb(negate_condition(cc), L);
5119 movl(dst, src);
5120 bind(L);
5121 }
5122 }
5123
5124 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) {
5125 if (VM_Version::supports_cmov()) {
5126 cmovl(cc, dst, src);
5127 } else {
5128 Label L;
5129 jccb(negate_condition(cc), L);
5130 movl(dst, src);
5131 bind(L);
5132 }
5133 }
5134
5135 void MacroAssembler::verify_oop(Register reg, const char* s) {
5136 if (!VerifyOops) return;
5137
5138 // Pass register number to verify_oop_subroutine
5139 const char* b = NULL;
5140 {
5141 ResourceMark rm;
5142 stringStream ss;
5143 ss.print("verify_oop: %s: %s", reg->name(), s);
5144 b = code_string(ss.as_string());
5145 }
5146 BLOCK_COMMENT("verify_oop {");
5147 #ifdef _LP64
5148 push(rscratch1); // save r10, trashed by movptr()
5149 #endif
5150 push(rax); // save rax,
5151 push(reg); // pass register argument
5152 ExternalAddress buffer((address) b);
5153 // avoid using pushptr, as it modifies scratch registers
5154 // and our contract is not to modify anything
5155 movptr(rax, buffer.addr());
5156 push(rax);
5157 // call indirectly to solve generation ordering problem
5158 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
5159 call(rax);
5160 // Caller pops the arguments (oop, message) and restores rax, r10
5161 BLOCK_COMMENT("} verify_oop");
5162 }
5163
5164
5165 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
5166 Register tmp,
5167 int offset) {
5168 intptr_t value = *delayed_value_addr;
5169 if (value != 0)
5170 return RegisterOrConstant(value + offset);
5171
5172 // load indirectly to solve generation ordering problem
5173 movptr(tmp, ExternalAddress((address) delayed_value_addr));
5174
5175 #ifdef ASSERT
5176 { Label L;
5177 testptr(tmp, tmp);
5178 if (WizardMode) {
5179 const char* buf = NULL;
5180 {
5181 ResourceMark rm;
5182 stringStream ss;
5183 ss.print("DelayedValue="INTPTR_FORMAT, delayed_value_addr[1]);
5184 buf = code_string(ss.as_string());
5185 }
5186 jcc(Assembler::notZero, L);
5187 STOP(buf);
5188 } else {
5189 jccb(Assembler::notZero, L);
5190 hlt();
5191 }
5192 bind(L);
5193 }
5194 #endif
5195
5196 if (offset != 0)
5197 addptr(tmp, offset);
5198
5199 return RegisterOrConstant(tmp);
5200 }
5201
5202
5203 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
5204 int extra_slot_offset) {
5205 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
5206 int stackElementSize = Interpreter::stackElementSize;
5207 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
5208 #ifdef ASSERT
5209 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
5210 assert(offset1 - offset == stackElementSize, "correct arithmetic");
5211 #endif
5212 Register scale_reg = noreg;
5213 Address::ScaleFactor scale_factor = Address::no_scale;
5214 if (arg_slot.is_constant()) {
5215 offset += arg_slot.as_constant() * stackElementSize;
5216 } else {
5217 scale_reg = arg_slot.as_register();
5218 scale_factor = Address::times(stackElementSize);
5219 }
5220 offset += wordSize; // return PC is on stack
5221 return Address(rsp, scale_reg, scale_factor, offset);
5222 }
5223
5224
5225 void MacroAssembler::verify_oop_addr(Address addr, const char* s) {
5226 if (!VerifyOops) return;
5227
5228 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord);
5229 // Pass register number to verify_oop_subroutine
5230 const char* b = NULL;
5231 {
5232 ResourceMark rm;
5233 stringStream ss;
5234 ss.print("verify_oop_addr: %s", s);
5235 b = code_string(ss.as_string());
5236 }
5237 #ifdef _LP64
5238 push(rscratch1); // save r10, trashed by movptr()
5239 #endif
5240 push(rax); // save rax,
5241 // addr may contain rsp so we will have to adjust it based on the push
5242 // we just did (and on 64 bit we do two pushes)
5243 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which
5244 // stores rax into addr which is backwards of what was intended.
5245 if (addr.uses(rsp)) {
5246 lea(rax, addr);
5247 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord));
5248 } else {
5249 pushptr(addr);
5250 }
5251
5252 ExternalAddress buffer((address) b);
5253 // pass msg argument
5254 // avoid using pushptr, as it modifies scratch registers
5255 // and our contract is not to modify anything
5256 movptr(rax, buffer.addr());
5257 push(rax);
5258
5259 // call indirectly to solve generation ordering problem
5260 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
5261 call(rax);
5262 // Caller pops the arguments (addr, message) and restores rax, r10.
5263 }
5264
5265 void MacroAssembler::verify_tlab() {
5266 #ifdef ASSERT
5267 if (UseTLAB && VerifyOops) {
5268 Label next, ok;
5269 Register t1 = rsi;
5270 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread);
5271
5272 push(t1);
5273 NOT_LP64(push(thread_reg));
5274 NOT_LP64(get_thread(thread_reg));
5275
5276 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
5277 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
5278 jcc(Assembler::aboveEqual, next);
5279 STOP("assert(top >= start)");
5280 should_not_reach_here();
5281
5282 bind(next);
5283 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
5284 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
5285 jcc(Assembler::aboveEqual, ok);
5286 STOP("assert(top <= end)");
5287 should_not_reach_here();
5288
5289 bind(ok);
5290 NOT_LP64(pop(thread_reg));
5291 pop(t1);
5292 }
5293 #endif
5294 }
5295
5296 class ControlWord {
5297 public:
5298 int32_t _value;
5299
5300 int rounding_control() const { return (_value >> 10) & 3 ; }
5301 int precision_control() const { return (_value >> 8) & 3 ; }
5302 bool precision() const { return ((_value >> 5) & 1) != 0; }
5303 bool underflow() const { return ((_value >> 4) & 1) != 0; }
5304 bool overflow() const { return ((_value >> 3) & 1) != 0; }
5305 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
5306 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
5307 bool invalid() const { return ((_value >> 0) & 1) != 0; }
5308
5309 void print() const {
5310 // rounding control
5311 const char* rc;
5312 switch (rounding_control()) {
5313 case 0: rc = "round near"; break;
5314 case 1: rc = "round down"; break;
5315 case 2: rc = "round up "; break;
5316 case 3: rc = "chop "; break;
5317 };
5318 // precision control
5319 const char* pc;
5320 switch (precision_control()) {
5321 case 0: pc = "24 bits "; break;
5322 case 1: pc = "reserved"; break;
5323 case 2: pc = "53 bits "; break;
5324 case 3: pc = "64 bits "; break;
5325 };
5326 // flags
5327 char f[9];
5328 f[0] = ' ';
5329 f[1] = ' ';
5330 f[2] = (precision ()) ? 'P' : 'p';
5331 f[3] = (underflow ()) ? 'U' : 'u';
5332 f[4] = (overflow ()) ? 'O' : 'o';
5333 f[5] = (zero_divide ()) ? 'Z' : 'z';
5334 f[6] = (denormalized()) ? 'D' : 'd';
5335 f[7] = (invalid ()) ? 'I' : 'i';
5336 f[8] = '\x0';
5337 // output
5338 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc);
5339 }
5340
5341 };
5342
5343 class StatusWord {
5344 public:
5345 int32_t _value;
5346
5347 bool busy() const { return ((_value >> 15) & 1) != 0; }
5348 bool C3() const { return ((_value >> 14) & 1) != 0; }
5349 bool C2() const { return ((_value >> 10) & 1) != 0; }
5350 bool C1() const { return ((_value >> 9) & 1) != 0; }
5351 bool C0() const { return ((_value >> 8) & 1) != 0; }
5352 int top() const { return (_value >> 11) & 7 ; }
5353 bool error_status() const { return ((_value >> 7) & 1) != 0; }
5354 bool stack_fault() const { return ((_value >> 6) & 1) != 0; }
5355 bool precision() const { return ((_value >> 5) & 1) != 0; }
5356 bool underflow() const { return ((_value >> 4) & 1) != 0; }
5357 bool overflow() const { return ((_value >> 3) & 1) != 0; }
5358 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
5359 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
5360 bool invalid() const { return ((_value >> 0) & 1) != 0; }
5361
5362 void print() const {
5363 // condition codes
5364 char c[5];
5365 c[0] = (C3()) ? '3' : '-';
5366 c[1] = (C2()) ? '2' : '-';
5367 c[2] = (C1()) ? '1' : '-';
5368 c[3] = (C0()) ? '0' : '-';
5369 c[4] = '\x0';
5370 // flags
5371 char f[9];
5372 f[0] = (error_status()) ? 'E' : '-';
5373 f[1] = (stack_fault ()) ? 'S' : '-';
5374 f[2] = (precision ()) ? 'P' : '-';
5375 f[3] = (underflow ()) ? 'U' : '-';
5376 f[4] = (overflow ()) ? 'O' : '-';
5377 f[5] = (zero_divide ()) ? 'Z' : '-';
5378 f[6] = (denormalized()) ? 'D' : '-';
5379 f[7] = (invalid ()) ? 'I' : '-';
5380 f[8] = '\x0';
5381 // output
5382 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top());
5383 }
5384
5385 };
5386
5387 class TagWord {
5388 public:
5389 int32_t _value;
5390
5391 int tag_at(int i) const { return (_value >> (i*2)) & 3; }
5392
5393 void print() const {
5394 printf("%04x", _value & 0xFFFF);
5395 }
5396
5397 };
5398
5399 class FPU_Register {
5400 public:
5401 int32_t _m0;
5402 int32_t _m1;
5403 int16_t _ex;
5404
5405 bool is_indefinite() const {
5406 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0;
5407 }
5408
5409 void print() const {
5410 char sign = (_ex < 0) ? '-' : '+';
5411 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " ";
5412 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind);
5413 };
5414
5415 };
5416
5417 class FPU_State {
5418 public:
5419 enum {
5420 register_size = 10,
5421 number_of_registers = 8,
5422 register_mask = 7
5423 };
5424
5425 ControlWord _control_word;
5426 StatusWord _status_word;
5427 TagWord _tag_word;
5428 int32_t _error_offset;
5429 int32_t _error_selector;
5430 int32_t _data_offset;
5431 int32_t _data_selector;
5432 int8_t _register[register_size * number_of_registers];
5433
5434 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); }
5435 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; }
5436
5437 const char* tag_as_string(int tag) const {
5438 switch (tag) {
5439 case 0: return "valid";
5440 case 1: return "zero";
5441 case 2: return "special";
5442 case 3: return "empty";
5443 }
5444 ShouldNotReachHere();
5445 return NULL;
5446 }
5447
5448 void print() const {
5449 // print computation registers
5450 { int t = _status_word.top();
5451 for (int i = 0; i < number_of_registers; i++) {
5452 int j = (i - t) & register_mask;
5453 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j);
5454 st(j)->print();
5455 printf(" %s\n", tag_as_string(_tag_word.tag_at(i)));
5456 }
5457 }
5458 printf("\n");
5459 // print control registers
5460 printf("ctrl = "); _control_word.print(); printf("\n");
5461 printf("stat = "); _status_word .print(); printf("\n");
5462 printf("tags = "); _tag_word .print(); printf("\n");
5463 }
5464
5465 };
5466
5467 class Flag_Register {
5468 public:
5469 int32_t _value;
5470
5471 bool overflow() const { return ((_value >> 11) & 1) != 0; }
5472 bool direction() const { return ((_value >> 10) & 1) != 0; }
5473 bool sign() const { return ((_value >> 7) & 1) != 0; }
5474 bool zero() const { return ((_value >> 6) & 1) != 0; }
5475 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; }
5476 bool parity() const { return ((_value >> 2) & 1) != 0; }
5477 bool carry() const { return ((_value >> 0) & 1) != 0; }
5478
5479 void print() const {
5480 // flags
5481 char f[8];
5482 f[0] = (overflow ()) ? 'O' : '-';
5483 f[1] = (direction ()) ? 'D' : '-';
5484 f[2] = (sign ()) ? 'S' : '-';
5485 f[3] = (zero ()) ? 'Z' : '-';
5486 f[4] = (auxiliary_carry()) ? 'A' : '-';
5487 f[5] = (parity ()) ? 'P' : '-';
5488 f[6] = (carry ()) ? 'C' : '-';
5489 f[7] = '\x0';
5490 // output
5491 printf("%08x flags = %s", _value, f);
5492 }
5493
5494 };
5495
5496 class IU_Register {
5497 public:
5498 int32_t _value;
5499
5500 void print() const {
5501 printf("%08x %11d", _value, _value);
5502 }
5503
5504 };
5505
5506 class IU_State {
5507 public:
5508 Flag_Register _eflags;
5509 IU_Register _rdi;
5510 IU_Register _rsi;
5511 IU_Register _rbp;
5512 IU_Register _rsp;
5513 IU_Register _rbx;
5514 IU_Register _rdx;
5515 IU_Register _rcx;
5516 IU_Register _rax;
5517
5518 void print() const {
5519 // computation registers
5520 printf("rax, = "); _rax.print(); printf("\n");
5521 printf("rbx, = "); _rbx.print(); printf("\n");
5522 printf("rcx = "); _rcx.print(); printf("\n");
5523 printf("rdx = "); _rdx.print(); printf("\n");
5524 printf("rdi = "); _rdi.print(); printf("\n");
5525 printf("rsi = "); _rsi.print(); printf("\n");
5526 printf("rbp, = "); _rbp.print(); printf("\n");
5527 printf("rsp = "); _rsp.print(); printf("\n");
5528 printf("\n");
5529 // control registers
5530 printf("flgs = "); _eflags.print(); printf("\n");
5531 }
5532 };
5533
5534
5535 class CPU_State {
5536 public:
5537 FPU_State _fpu_state;
5538 IU_State _iu_state;
5539
5540 void print() const {
5541 printf("--------------------------------------------------\n");
5542 _iu_state .print();
5543 printf("\n");
5544 _fpu_state.print();
5545 printf("--------------------------------------------------\n");
5546 }
5547
5548 };
5549
5550
5551 static void _print_CPU_state(CPU_State* state) {
5552 state->print();
5553 };
5554
5555
5556 void MacroAssembler::print_CPU_state() {
5557 push_CPU_state();
5558 push(rsp); // pass CPU state
5559 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state)));
5560 addptr(rsp, wordSize); // discard argument
5561 pop_CPU_state();
5562 }
5563
5564
5565 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) {
5566 static int counter = 0;
5567 FPU_State* fs = &state->_fpu_state;
5568 counter++;
5569 // For leaf calls, only verify that the top few elements remain empty.
5570 // We only need 1 empty at the top for C2 code.
5571 if( stack_depth < 0 ) {
5572 if( fs->tag_for_st(7) != 3 ) {
5573 printf("FPR7 not empty\n");
5574 state->print();
5575 assert(false, "error");
5576 return false;
5577 }
5578 return true; // All other stack states do not matter
5579 }
5580
5581 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std,
5582 "bad FPU control word");
5583
5584 // compute stack depth
5585 int i = 0;
5586 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++;
5587 int d = i;
5588 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++;
5589 // verify findings
5590 if (i != FPU_State::number_of_registers) {
5591 // stack not contiguous
5592 printf("%s: stack not contiguous at ST%d\n", s, i);
5593 state->print();
5594 assert(false, "error");
5595 return false;
5596 }
5597 // check if computed stack depth corresponds to expected stack depth
5598 if (stack_depth < 0) {
5599 // expected stack depth is -stack_depth or less
5600 if (d > -stack_depth) {
5601 // too many elements on the stack
5602 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d);
5603 state->print();
5604 assert(false, "error");
5605 return false;
5606 }
5607 } else {
5608 // expected stack depth is stack_depth
5609 if (d != stack_depth) {
5610 // wrong stack depth
5611 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d);
5612 state->print();
5613 assert(false, "error");
5614 return false;
5615 }
5616 }
5617 // everything is cool
5618 return true;
5619 }
5620
5621
5622 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
5623 if (!VerifyFPU) return;
5624 push_CPU_state();
5625 push(rsp); // pass CPU state
5626 ExternalAddress msg((address) s);
5627 // pass message string s
5628 pushptr(msg.addr());
5629 push(stack_depth); // pass stack depth
5630 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU)));
5631 addptr(rsp, 3 * wordSize); // discard arguments
5632 // check for error
5633 { Label L;
5634 testl(rax, rax);
5635 jcc(Assembler::notZero, L);
5636 int3(); // break if error condition
5637 bind(L);
5638 }
5639 pop_CPU_state();
5640 }
5641
5642 void MacroAssembler::restore_cpu_control_state_after_jni() {
5643 // Either restore the MXCSR register after returning from the JNI Call
5644 // or verify that it wasn't changed (with -Xcheck:jni flag).
5645 if (VM_Version::supports_sse()) {
5646 if (RestoreMXCSROnJNICalls) {
5647 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
5648 } else if (CheckJNICalls) {
5649 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
5650 }
5651 }
5652 if (VM_Version::supports_avx()) {
5653 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty.
5654 vzeroupper();
5655 }
5656
5657 #ifndef _LP64
5658 // Either restore the x87 floating pointer control word after returning
5659 // from the JNI call or verify that it wasn't changed.
5660 if (CheckJNICalls) {
5661 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
5662 }
5663 #endif // _LP64
5664 }
5665
5666
5667 void MacroAssembler::load_klass(Register dst, Register src) {
5668 #ifdef _LP64
5669 if (UseCompressedClassPointers) {
5670 movl(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5671 decode_klass_not_null(dst);
5672 } else
5673 #endif
5674 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5675 }
5676
5677 void MacroAssembler::load_prototype_header(Register dst, Register src) {
5678 load_klass(dst, src);
5679 movptr(dst, Address(dst, Klass::prototype_header_offset()));
5680 }
5681
5682 void MacroAssembler::store_klass(Register dst, Register src) {
5683 #ifdef _LP64
5684 if (UseCompressedClassPointers) {
5685 encode_klass_not_null(src);
5686 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5687 } else
5688 #endif
5689 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5690 }
5691
5692 void MacroAssembler::load_heap_oop(Register dst, Address src) {
5693 #ifdef _LP64
5694 // FIXME: Must change all places where we try to load the klass.
5695 if (UseCompressedOops) {
5696 movl(dst, src);
5697 decode_heap_oop(dst);
5698 } else
5699 #endif
5700 movptr(dst, src);
5701 }
5702
5703 // Doesn't do verfication, generates fixed size code
5704 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) {
5705 #ifdef _LP64
5706 if (UseCompressedOops) {
5707 movl(dst, src);
5708 decode_heap_oop_not_null(dst);
5709 } else
5710 #endif
5711 movptr(dst, src);
5712 }
5713
5714 void MacroAssembler::store_heap_oop(Address dst, Register src) {
5715 #ifdef _LP64
5716 if (UseCompressedOops) {
5717 assert(!dst.uses(src), "not enough registers");
5718 encode_heap_oop(src);
5719 movl(dst, src);
5720 } else
5721 #endif
5722 movptr(dst, src);
5723 }
5724
5725 void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) {
5726 assert_different_registers(src1, tmp);
5727 #ifdef _LP64
5728 if (UseCompressedOops) {
5729 bool did_push = false;
5730 if (tmp == noreg) {
5731 tmp = rax;
5732 push(tmp);
5733 did_push = true;
5734 assert(!src2.uses(rsp), "can't push");
5735 }
5736 load_heap_oop(tmp, src2);
5737 cmpptr(src1, tmp);
5738 if (did_push) pop(tmp);
5739 } else
5740 #endif
5741 cmpptr(src1, src2);
5742 }
5743
5744 // Used for storing NULLs.
5745 void MacroAssembler::store_heap_oop_null(Address dst) {
5746 #ifdef _LP64
5747 if (UseCompressedOops) {
5748 movl(dst, (int32_t)NULL_WORD);
5749 } else {
5750 movslq(dst, (int32_t)NULL_WORD);
5751 }
5752 #else
5753 movl(dst, (int32_t)NULL_WORD);
5754 #endif
5755 }
5756
5757 #ifdef _LP64
5758 void MacroAssembler::store_klass_gap(Register dst, Register src) {
5759 if (UseCompressedClassPointers) {
5760 // Store to klass gap in destination
5761 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src);
5762 }
5763 }
5764
5765 #ifdef ASSERT
5766 void MacroAssembler::verify_heapbase(const char* msg) {
5767 assert (UseCompressedOops, "should be compressed");
5768 assert (Universe::heap() != NULL, "java heap should be initialized");
5769 if (CheckCompressedOops) {
5770 Label ok;
5771 push(rscratch1); // cmpptr trashes rscratch1
5772 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
5773 jcc(Assembler::equal, ok);
5774 STOP(msg);
5775 bind(ok);
5776 pop(rscratch1);
5777 }
5778 }
5779 #endif
5780
5781 // Algorithm must match oop.inline.hpp encode_heap_oop.
5782 void MacroAssembler::encode_heap_oop(Register r) {
5783 #ifdef ASSERT
5784 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
5785 #endif
5786 verify_oop(r, "broken oop in encode_heap_oop");
5787 if (Universe::narrow_oop_base() == NULL) {
5788 if (Universe::narrow_oop_shift() != 0) {
5789 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5790 shrq(r, LogMinObjAlignmentInBytes);
5791 }
5792 return;
5793 }
5794 testq(r, r);
5795 cmovq(Assembler::equal, r, r12_heapbase);
5796 subq(r, r12_heapbase);
5797 shrq(r, LogMinObjAlignmentInBytes);
5798 }
5799
5800 void MacroAssembler::encode_heap_oop_not_null(Register r) {
5801 #ifdef ASSERT
5802 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
5803 if (CheckCompressedOops) {
5804 Label ok;
5805 testq(r, r);
5806 jcc(Assembler::notEqual, ok);
5807 STOP("null oop passed to encode_heap_oop_not_null");
5808 bind(ok);
5809 }
5810 #endif
5811 verify_oop(r, "broken oop in encode_heap_oop_not_null");
5812 if (Universe::narrow_oop_base() != NULL) {
5813 subq(r, r12_heapbase);
5814 }
5815 if (Universe::narrow_oop_shift() != 0) {
5816 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5817 shrq(r, LogMinObjAlignmentInBytes);
5818 }
5819 }
5820
5821 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
5822 #ifdef ASSERT
5823 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
5824 if (CheckCompressedOops) {
5825 Label ok;
5826 testq(src, src);
5827 jcc(Assembler::notEqual, ok);
5828 STOP("null oop passed to encode_heap_oop_not_null2");
5829 bind(ok);
5830 }
5831 #endif
5832 verify_oop(src, "broken oop in encode_heap_oop_not_null2");
5833 if (dst != src) {
5834 movq(dst, src);
5835 }
5836 if (Universe::narrow_oop_base() != NULL) {
5837 subq(dst, r12_heapbase);
5838 }
5839 if (Universe::narrow_oop_shift() != 0) {
5840 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5841 shrq(dst, LogMinObjAlignmentInBytes);
5842 }
5843 }
5844
5845 void MacroAssembler::decode_heap_oop(Register r) {
5846 #ifdef ASSERT
5847 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
5848 #endif
5849 if (Universe::narrow_oop_base() == NULL) {
5850 if (Universe::narrow_oop_shift() != 0) {
5851 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5852 shlq(r, LogMinObjAlignmentInBytes);
5853 }
5854 } else {
5855 Label done;
5856 shlq(r, LogMinObjAlignmentInBytes);
5857 jccb(Assembler::equal, done);
5858 addq(r, r12_heapbase);
5859 bind(done);
5860 }
5861 verify_oop(r, "broken oop in decode_heap_oop");
5862 }
5863
5864 void MacroAssembler::decode_heap_oop_not_null(Register r) {
5865 // Note: it will change flags
5866 assert (UseCompressedOops, "should only be used for compressed headers");
5867 assert (Universe::heap() != NULL, "java heap should be initialized");
5868 // Cannot assert, unverified entry point counts instructions (see .ad file)
5869 // vtableStubs also counts instructions in pd_code_size_limit.
5870 // Also do not verify_oop as this is called by verify_oop.
5871 if (Universe::narrow_oop_shift() != 0) {
5872 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5873 shlq(r, LogMinObjAlignmentInBytes);
5874 if (Universe::narrow_oop_base() != NULL) {
5875 addq(r, r12_heapbase);
5876 }
5877 } else {
5878 assert (Universe::narrow_oop_base() == NULL, "sanity");
5879 }
5880 }
5881
5882 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
5883 // Note: it will change flags
5884 assert (UseCompressedOops, "should only be used for compressed headers");
5885 assert (Universe::heap() != NULL, "java heap should be initialized");
5886 // Cannot assert, unverified entry point counts instructions (see .ad file)
5887 // vtableStubs also counts instructions in pd_code_size_limit.
5888 // Also do not verify_oop as this is called by verify_oop.
5889 if (Universe::narrow_oop_shift() != 0) {
5890 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5891 if (LogMinObjAlignmentInBytes == Address::times_8) {
5892 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0));
5893 } else {
5894 if (dst != src) {
5895 movq(dst, src);
5896 }
5897 shlq(dst, LogMinObjAlignmentInBytes);
5898 if (Universe::narrow_oop_base() != NULL) {
5899 addq(dst, r12_heapbase);
5900 }
5901 }
5902 } else {
5903 assert (Universe::narrow_oop_base() == NULL, "sanity");
5904 if (dst != src) {
5905 movq(dst, src);
5906 }
5907 }
5908 }
5909
5910 void MacroAssembler::encode_klass_not_null(Register r) {
5911 if (Universe::narrow_klass_base() != NULL) {
5912 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5913 assert(r != r12_heapbase, "Encoding a klass in r12");
5914 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
5915 subq(r, r12_heapbase);
5916 }
5917 if (Universe::narrow_klass_shift() != 0) {
5918 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5919 shrq(r, LogKlassAlignmentInBytes);
5920 }
5921 if (Universe::narrow_klass_base() != NULL) {
5922 reinit_heapbase();
5923 }
5924 }
5925
5926 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
5927 if (dst == src) {
5928 encode_klass_not_null(src);
5929 } else {
5930 if (Universe::narrow_klass_base() != NULL) {
5931 mov64(dst, (int64_t)Universe::narrow_klass_base());
5932 negq(dst);
5933 addq(dst, src);
5934 } else {
5935 movptr(dst, src);
5936 }
5937 if (Universe::narrow_klass_shift() != 0) {
5938 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5939 shrq(dst, LogKlassAlignmentInBytes);
5940 }
5941 }
5942 }
5943
5944 // Function instr_size_for_decode_klass_not_null() counts the instructions
5945 // generated by decode_klass_not_null(register r) and reinit_heapbase(),
5946 // when (Universe::heap() != NULL). Hence, if the instructions they
5947 // generate change, then this method needs to be updated.
5948 int MacroAssembler::instr_size_for_decode_klass_not_null() {
5949 assert (UseCompressedClassPointers, "only for compressed klass ptrs");
5950 if (Universe::narrow_klass_base() != NULL) {
5951 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()).
5952 return (Universe::narrow_klass_shift() == 0 ? 20 : 24);
5953 } else {
5954 // longest load decode klass function, mov64, leaq
5955 return 16;
5956 }
5957 }
5958
5959 // !!! If the instructions that get generated here change then function
5960 // instr_size_for_decode_klass_not_null() needs to get updated.
5961 void MacroAssembler::decode_klass_not_null(Register r) {
5962 // Note: it will change flags
5963 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5964 assert(r != r12_heapbase, "Decoding a klass in r12");
5965 // Cannot assert, unverified entry point counts instructions (see .ad file)
5966 // vtableStubs also counts instructions in pd_code_size_limit.
5967 // Also do not verify_oop as this is called by verify_oop.
5968 if (Universe::narrow_klass_shift() != 0) {
5969 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5970 shlq(r, LogKlassAlignmentInBytes);
5971 }
5972 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5973 if (Universe::narrow_klass_base() != NULL) {
5974 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
5975 addq(r, r12_heapbase);
5976 reinit_heapbase();
5977 }
5978 }
5979
5980 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
5981 // Note: it will change flags
5982 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5983 if (dst == src) {
5984 decode_klass_not_null(dst);
5985 } else {
5986 // Cannot assert, unverified entry point counts instructions (see .ad file)
5987 // vtableStubs also counts instructions in pd_code_size_limit.
5988 // Also do not verify_oop as this is called by verify_oop.
5989 mov64(dst, (int64_t)Universe::narrow_klass_base());
5990 if (Universe::narrow_klass_shift() != 0) {
5991 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5992 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?");
5993 leaq(dst, Address(dst, src, Address::times_8, 0));
5994 } else {
5995 addq(dst, src);
5996 }
5997 }
5998 }
5999
6000 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
6001 assert (UseCompressedOops, "should only be used for compressed headers");
6002 assert (Universe::heap() != NULL, "java heap should be initialized");
6003 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6004 int oop_index = oop_recorder()->find_index(obj);
6005 RelocationHolder rspec = oop_Relocation::spec(oop_index);
6006 mov_narrow_oop(dst, oop_index, rspec);
6007 }
6008
6009 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) {
6010 assert (UseCompressedOops, "should only be used for compressed headers");
6011 assert (Universe::heap() != NULL, "java heap should be initialized");
6012 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6013 int oop_index = oop_recorder()->find_index(obj);
6014 RelocationHolder rspec = oop_Relocation::spec(oop_index);
6015 mov_narrow_oop(dst, oop_index, rspec);
6016 }
6017
6018 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
6019 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6020 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6021 int klass_index = oop_recorder()->find_index(k);
6022 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6023 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
6024 }
6025
6026 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) {
6027 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6028 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6029 int klass_index = oop_recorder()->find_index(k);
6030 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6031 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
6032 }
6033
6034 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) {
6035 assert (UseCompressedOops, "should only be used for compressed headers");
6036 assert (Universe::heap() != NULL, "java heap should be initialized");
6037 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6038 int oop_index = oop_recorder()->find_index(obj);
6039 RelocationHolder rspec = oop_Relocation::spec(oop_index);
6040 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
6041 }
6042
6043 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) {
6044 assert (UseCompressedOops, "should only be used for compressed headers");
6045 assert (Universe::heap() != NULL, "java heap should be initialized");
6046 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6047 int oop_index = oop_recorder()->find_index(obj);
6048 RelocationHolder rspec = oop_Relocation::spec(oop_index);
6049 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
6050 }
6051
6052 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) {
6053 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6054 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6055 int klass_index = oop_recorder()->find_index(k);
6056 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6057 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
6058 }
6059
6060 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) {
6061 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6062 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6063 int klass_index = oop_recorder()->find_index(k);
6064 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6065 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
6066 }
6067
6068 void MacroAssembler::reinit_heapbase() {
6069 if (UseCompressedOops || UseCompressedClassPointers) {
6070 if (Universe::heap() != NULL) {
6071 if (Universe::narrow_oop_base() == NULL) {
6072 MacroAssembler::xorptr(r12_heapbase, r12_heapbase);
6073 } else {
6074 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base());
6075 }
6076 } else {
6077 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
6078 }
6079 }
6080 }
6081
6082 #endif // _LP64
6083
6084
6085 // C2 compiled method's prolog code.
6086 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) {
6087
6088 // WARNING: Initial instruction MUST be 5 bytes or longer so that
6089 // NativeJump::patch_verified_entry will be able to patch out the entry
6090 // code safely. The push to verify stack depth is ok at 5 bytes,
6091 // the frame allocation can be either 3 or 6 bytes. So if we don't do
6092 // stack bang then we must use the 6 byte frame allocation even if
6093 // we have no frame. :-(
6094 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect");
6095
6096 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
6097 // Remove word for return addr
6098 framesize -= wordSize;
6099 stack_bang_size -= wordSize;
6100
6101 // Calls to C2R adapters often do not accept exceptional returns.
6102 // We require that their callers must bang for them. But be careful, because
6103 // some VM calls (such as call site linkage) can use several kilobytes of
6104 // stack. But the stack safety zone should account for that.
6105 // See bugs 4446381, 4468289, 4497237.
6106 if (stack_bang_size > 0) {
6107 generate_stack_overflow_check(stack_bang_size);
6108
6109 // We always push rbp, so that on return to interpreter rbp, will be
6110 // restored correctly and we can correct the stack.
6111 push(rbp);
6112 // Save caller's stack pointer into RBP if the frame pointer is preserved.
6113 if (PreserveFramePointer) {
6114 mov(rbp, rsp);
6115 }
6116 // Remove word for ebp
6117 framesize -= wordSize;
6118
6119 // Create frame
6120 if (framesize) {
6121 subptr(rsp, framesize);
6122 }
6123 } else {
6124 // Create frame (force generation of a 4 byte immediate value)
6125 subptr_imm32(rsp, framesize);
6126
6127 // Save RBP register now.
6128 framesize -= wordSize;
6129 movptr(Address(rsp, framesize), rbp);
6130 // Save caller's stack pointer into RBP if the frame pointer is preserved.
6131 if (PreserveFramePointer) {
6132 movptr(rbp, rsp);
6133 addptr(rbp, framesize + wordSize);
6134 }
6135 }
6136
6137 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
6138 framesize -= wordSize;
6139 movptr(Address(rsp, framesize), (int32_t)0xbadb100d);
6140 }
6141
6142 #ifndef _LP64
6143 // If method sets FPU control word do it now
6144 if (fp_mode_24b) {
6145 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
6146 }
6147 if (UseSSE >= 2 && VerifyFPU) {
6148 verify_FPU(0, "FPU stack must be clean on entry");
6149 }
6150 #endif
6151
6152 #ifdef ASSERT
6153 if (VerifyStackAtCalls) {
6154 Label L;
6155 push(rax);
6156 mov(rax, rsp);
6157 andptr(rax, StackAlignmentInBytes-1);
6158 cmpptr(rax, StackAlignmentInBytes-wordSize);
6159 pop(rax);
6160 jcc(Assembler::equal, L);
6161 STOP("Stack is not properly aligned!");
6162 bind(L);
6163 }
6164 #endif
6165
6166 }
6167
6168 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp) {
6169 // cnt - number of qwords (8-byte words).
6170 // base - start address, qword aligned.
6171 assert(base==rdi, "base register must be edi for rep stos");
6172 assert(tmp==rax, "tmp register must be eax for rep stos");
6173 assert(cnt==rcx, "cnt register must be ecx for rep stos");
6174
6175 xorptr(tmp, tmp);
6176 if (UseFastStosb) {
6177 shlptr(cnt,3); // convert to number of bytes
6178 rep_stosb();
6179 } else {
6180 NOT_LP64(shlptr(cnt,1);) // convert to number of dwords for 32-bit VM
6181 rep_stos();
6182 }
6183 }
6184
6185 // IndexOf for constant substrings with size >= 8 chars
6186 // which don't need to be loaded through stack.
6187 void MacroAssembler::string_indexofC8(Register str1, Register str2,
6188 Register cnt1, Register cnt2,
6189 int int_cnt2, Register result,
6190 XMMRegister vec, Register tmp) {
6191 ShortBranchVerifier sbv(this);
6192 assert(UseSSE42Intrinsics, "SSE4.2 is required");
6193
6194 // This method uses pcmpestri instruction with bound registers
6195 // inputs:
6196 // xmm - substring
6197 // rax - substring length (elements count)
6198 // mem - scanned string
6199 // rdx - string length (elements count)
6200 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6201 // outputs:
6202 // rcx - matched index in string
6203 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6204
6205 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR,
6206 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR,
6207 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE;
6208
6209 // Note, inline_string_indexOf() generates checks:
6210 // if (substr.count > string.count) return -1;
6211 // if (substr.count == 0) return 0;
6212 assert(int_cnt2 >= 8, "this code isused only for cnt2 >= 8 chars");
6213
6214 // Load substring.
6215 movdqu(vec, Address(str2, 0));
6216 movl(cnt2, int_cnt2);
6217 movptr(result, str1); // string addr
6218
6219 if (int_cnt2 > 8) {
6220 jmpb(SCAN_TO_SUBSTR);
6221
6222 // Reload substr for rescan, this code
6223 // is executed only for large substrings (> 8 chars)
6224 bind(RELOAD_SUBSTR);
6225 movdqu(vec, Address(str2, 0));
6226 negptr(cnt2); // Jumped here with negative cnt2, convert to positive
6227
6228 bind(RELOAD_STR);
6229 // We came here after the beginning of the substring was
6230 // matched but the rest of it was not so we need to search
6231 // again. Start from the next element after the previous match.
6232
6233 // cnt2 is number of substring reminding elements and
6234 // cnt1 is number of string reminding elements when cmp failed.
6235 // Restored cnt1 = cnt1 - cnt2 + int_cnt2
6236 subl(cnt1, cnt2);
6237 addl(cnt1, int_cnt2);
6238 movl(cnt2, int_cnt2); // Now restore cnt2
6239
6240 decrementl(cnt1); // Shift to next element
6241 cmpl(cnt1, cnt2);
6242 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6243
6244 addptr(result, 2);
6245
6246 } // (int_cnt2 > 8)
6247
6248 // Scan string for start of substr in 16-byte vectors
6249 bind(SCAN_TO_SUBSTR);
6250 pcmpestri(vec, Address(result, 0), 0x0d);
6251 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
6252 subl(cnt1, 8);
6253 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6254 cmpl(cnt1, cnt2);
6255 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6256 addptr(result, 16);
6257 jmpb(SCAN_TO_SUBSTR);
6258
6259 // Found a potential substr
6260 bind(FOUND_CANDIDATE);
6261 // Matched whole vector if first element matched (tmp(rcx) == 0).
6262 if (int_cnt2 == 8) {
6263 jccb(Assembler::overflow, RET_FOUND); // OF == 1
6264 } else { // int_cnt2 > 8
6265 jccb(Assembler::overflow, FOUND_SUBSTR);
6266 }
6267 // After pcmpestri tmp(rcx) contains matched element index
6268 // Compute start addr of substr
6269 lea(result, Address(result, tmp, Address::times_2));
6270
6271 // Make sure string is still long enough
6272 subl(cnt1, tmp);
6273 cmpl(cnt1, cnt2);
6274 if (int_cnt2 == 8) {
6275 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6276 } else { // int_cnt2 > 8
6277 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD);
6278 }
6279 // Left less then substring.
6280
6281 bind(RET_NOT_FOUND);
6282 movl(result, -1);
6283 jmpb(EXIT);
6284
6285 if (int_cnt2 > 8) {
6286 // This code is optimized for the case when whole substring
6287 // is matched if its head is matched.
6288 bind(MATCH_SUBSTR_HEAD);
6289 pcmpestri(vec, Address(result, 0), 0x0d);
6290 // Reload only string if does not match
6291 jccb(Assembler::noOverflow, RELOAD_STR); // OF == 0
6292
6293 Label CONT_SCAN_SUBSTR;
6294 // Compare the rest of substring (> 8 chars).
6295 bind(FOUND_SUBSTR);
6296 // First 8 chars are already matched.
6297 negptr(cnt2);
6298 addptr(cnt2, 8);
6299
6300 bind(SCAN_SUBSTR);
6301 subl(cnt1, 8);
6302 cmpl(cnt2, -8); // Do not read beyond substring
6303 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR);
6304 // Back-up strings to avoid reading beyond substring:
6305 // cnt1 = cnt1 - cnt2 + 8
6306 addl(cnt1, cnt2); // cnt2 is negative
6307 addl(cnt1, 8);
6308 movl(cnt2, 8); negptr(cnt2);
6309 bind(CONT_SCAN_SUBSTR);
6310 if (int_cnt2 < (int)G) {
6311 movdqu(vec, Address(str2, cnt2, Address::times_2, int_cnt2*2));
6312 pcmpestri(vec, Address(result, cnt2, Address::times_2, int_cnt2*2), 0x0d);
6313 } else {
6314 // calculate index in register to avoid integer overflow (int_cnt2*2)
6315 movl(tmp, int_cnt2);
6316 addptr(tmp, cnt2);
6317 movdqu(vec, Address(str2, tmp, Address::times_2, 0));
6318 pcmpestri(vec, Address(result, tmp, Address::times_2, 0), 0x0d);
6319 }
6320 // Need to reload strings pointers if not matched whole vector
6321 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6322 addptr(cnt2, 8);
6323 jcc(Assembler::negative, SCAN_SUBSTR);
6324 // Fall through if found full substring
6325
6326 } // (int_cnt2 > 8)
6327
6328 bind(RET_FOUND);
6329 // Found result if we matched full small substring.
6330 // Compute substr offset
6331 subptr(result, str1);
6332 shrl(result, 1); // index
6333 bind(EXIT);
6334
6335 } // string_indexofC8
6336
6337 // Small strings are loaded through stack if they cross page boundary.
6338 void MacroAssembler::string_indexof(Register str1, Register str2,
6339 Register cnt1, Register cnt2,
6340 int int_cnt2, Register result,
6341 XMMRegister vec, Register tmp) {
6342 ShortBranchVerifier sbv(this);
6343 assert(UseSSE42Intrinsics, "SSE4.2 is required");
6344 //
6345 // int_cnt2 is length of small (< 8 chars) constant substring
6346 // or (-1) for non constant substring in which case its length
6347 // is in cnt2 register.
6348 //
6349 // Note, inline_string_indexOf() generates checks:
6350 // if (substr.count > string.count) return -1;
6351 // if (substr.count == 0) return 0;
6352 //
6353 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < 8), "should be != 0");
6354
6355 // This method uses pcmpestri instruction with bound registers
6356 // inputs:
6357 // xmm - substring
6358 // rax - substring length (elements count)
6359 // mem - scanned string
6360 // rdx - string length (elements count)
6361 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6362 // outputs:
6363 // rcx - matched index in string
6364 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6365
6366 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR,
6367 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR,
6368 FOUND_CANDIDATE;
6369
6370 { //========================================================
6371 // We don't know where these strings are located
6372 // and we can't read beyond them. Load them through stack.
6373 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR;
6374
6375 movptr(tmp, rsp); // save old SP
6376
6377 if (int_cnt2 > 0) { // small (< 8 chars) constant substring
6378 if (int_cnt2 == 1) { // One char
6379 load_unsigned_short(result, Address(str2, 0));
6380 movdl(vec, result); // move 32 bits
6381 } else if (int_cnt2 == 2) { // Two chars
6382 movdl(vec, Address(str2, 0)); // move 32 bits
6383 } else if (int_cnt2 == 4) { // Four chars
6384 movq(vec, Address(str2, 0)); // move 64 bits
6385 } else { // cnt2 = { 3, 5, 6, 7 }
6386 // Array header size is 12 bytes in 32-bit VM
6387 // + 6 bytes for 3 chars == 18 bytes,
6388 // enough space to load vec and shift.
6389 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity");
6390 movdqu(vec, Address(str2, (int_cnt2*2)-16));
6391 psrldq(vec, 16-(int_cnt2*2));
6392 }
6393 } else { // not constant substring
6394 cmpl(cnt2, 8);
6395 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough
6396
6397 // We can read beyond string if srt+16 does not cross page boundary
6398 // since heaps are aligned and mapped by pages.
6399 assert(os::vm_page_size() < (int)G, "default page should be small");
6400 movl(result, str2); // We need only low 32 bits
6401 andl(result, (os::vm_page_size()-1));
6402 cmpl(result, (os::vm_page_size()-16));
6403 jccb(Assembler::belowEqual, CHECK_STR);
6404
6405 // Move small strings to stack to allow load 16 bytes into vec.
6406 subptr(rsp, 16);
6407 int stk_offset = wordSize-2;
6408 push(cnt2);
6409
6410 bind(COPY_SUBSTR);
6411 load_unsigned_short(result, Address(str2, cnt2, Address::times_2, -2));
6412 movw(Address(rsp, cnt2, Address::times_2, stk_offset), result);
6413 decrement(cnt2);
6414 jccb(Assembler::notZero, COPY_SUBSTR);
6415
6416 pop(cnt2);
6417 movptr(str2, rsp); // New substring address
6418 } // non constant
6419
6420 bind(CHECK_STR);
6421 cmpl(cnt1, 8);
6422 jccb(Assembler::aboveEqual, BIG_STRINGS);
6423
6424 // Check cross page boundary.
6425 movl(result, str1); // We need only low 32 bits
6426 andl(result, (os::vm_page_size()-1));
6427 cmpl(result, (os::vm_page_size()-16));
6428 jccb(Assembler::belowEqual, BIG_STRINGS);
6429
6430 subptr(rsp, 16);
6431 int stk_offset = -2;
6432 if (int_cnt2 < 0) { // not constant
6433 push(cnt2);
6434 stk_offset += wordSize;
6435 }
6436 movl(cnt2, cnt1);
6437
6438 bind(COPY_STR);
6439 load_unsigned_short(result, Address(str1, cnt2, Address::times_2, -2));
6440 movw(Address(rsp, cnt2, Address::times_2, stk_offset), result);
6441 decrement(cnt2);
6442 jccb(Assembler::notZero, COPY_STR);
6443
6444 if (int_cnt2 < 0) { // not constant
6445 pop(cnt2);
6446 }
6447 movptr(str1, rsp); // New string address
6448
6449 bind(BIG_STRINGS);
6450 // Load substring.
6451 if (int_cnt2 < 0) { // -1
6452 movdqu(vec, Address(str2, 0));
6453 push(cnt2); // substr count
6454 push(str2); // substr addr
6455 push(str1); // string addr
6456 } else {
6457 // Small (< 8 chars) constant substrings are loaded already.
6458 movl(cnt2, int_cnt2);
6459 }
6460 push(tmp); // original SP
6461
6462 } // Finished loading
6463
6464 //========================================================
6465 // Start search
6466 //
6467
6468 movptr(result, str1); // string addr
6469
6470 if (int_cnt2 < 0) { // Only for non constant substring
6471 jmpb(SCAN_TO_SUBSTR);
6472
6473 // SP saved at sp+0
6474 // String saved at sp+1*wordSize
6475 // Substr saved at sp+2*wordSize
6476 // Substr count saved at sp+3*wordSize
6477
6478 // Reload substr for rescan, this code
6479 // is executed only for large substrings (> 8 chars)
6480 bind(RELOAD_SUBSTR);
6481 movptr(str2, Address(rsp, 2*wordSize));
6482 movl(cnt2, Address(rsp, 3*wordSize));
6483 movdqu(vec, Address(str2, 0));
6484 // We came here after the beginning of the substring was
6485 // matched but the rest of it was not so we need to search
6486 // again. Start from the next element after the previous match.
6487 subptr(str1, result); // Restore counter
6488 shrl(str1, 1);
6489 addl(cnt1, str1);
6490 decrementl(cnt1); // Shift to next element
6491 cmpl(cnt1, cnt2);
6492 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6493
6494 addptr(result, 2);
6495 } // non constant
6496
6497 // Scan string for start of substr in 16-byte vectors
6498 bind(SCAN_TO_SUBSTR);
6499 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6500 pcmpestri(vec, Address(result, 0), 0x0d);
6501 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
6502 subl(cnt1, 8);
6503 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6504 cmpl(cnt1, cnt2);
6505 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6506 addptr(result, 16);
6507
6508 bind(ADJUST_STR);
6509 cmpl(cnt1, 8); // Do not read beyond string
6510 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6511 // Back-up string to avoid reading beyond string.
6512 lea(result, Address(result, cnt1, Address::times_2, -16));
6513 movl(cnt1, 8);
6514 jmpb(SCAN_TO_SUBSTR);
6515
6516 // Found a potential substr
6517 bind(FOUND_CANDIDATE);
6518 // After pcmpestri tmp(rcx) contains matched element index
6519
6520 // Make sure string is still long enough
6521 subl(cnt1, tmp);
6522 cmpl(cnt1, cnt2);
6523 jccb(Assembler::greaterEqual, FOUND_SUBSTR);
6524 // Left less then substring.
6525
6526 bind(RET_NOT_FOUND);
6527 movl(result, -1);
6528 jmpb(CLEANUP);
6529
6530 bind(FOUND_SUBSTR);
6531 // Compute start addr of substr
6532 lea(result, Address(result, tmp, Address::times_2));
6533
6534 if (int_cnt2 > 0) { // Constant substring
6535 // Repeat search for small substring (< 8 chars)
6536 // from new point without reloading substring.
6537 // Have to check that we don't read beyond string.
6538 cmpl(tmp, 8-int_cnt2);
6539 jccb(Assembler::greater, ADJUST_STR);
6540 // Fall through if matched whole substring.
6541 } else { // non constant
6542 assert(int_cnt2 == -1, "should be != 0");
6543
6544 addl(tmp, cnt2);
6545 // Found result if we matched whole substring.
6546 cmpl(tmp, 8);
6547 jccb(Assembler::lessEqual, RET_FOUND);
6548
6549 // Repeat search for small substring (<= 8 chars)
6550 // from new point 'str1' without reloading substring.
6551 cmpl(cnt2, 8);
6552 // Have to check that we don't read beyond string.
6553 jccb(Assembler::lessEqual, ADJUST_STR);
6554
6555 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG;
6556 // Compare the rest of substring (> 8 chars).
6557 movptr(str1, result);
6558
6559 cmpl(tmp, cnt2);
6560 // First 8 chars are already matched.
6561 jccb(Assembler::equal, CHECK_NEXT);
6562
6563 bind(SCAN_SUBSTR);
6564 pcmpestri(vec, Address(str1, 0), 0x0d);
6565 // Need to reload strings pointers if not matched whole vector
6566 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6567
6568 bind(CHECK_NEXT);
6569 subl(cnt2, 8);
6570 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring
6571 addptr(str1, 16);
6572 addptr(str2, 16);
6573 subl(cnt1, 8);
6574 cmpl(cnt2, 8); // Do not read beyond substring
6575 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR);
6576 // Back-up strings to avoid reading beyond substring.
6577 lea(str2, Address(str2, cnt2, Address::times_2, -16));
6578 lea(str1, Address(str1, cnt2, Address::times_2, -16));
6579 subl(cnt1, cnt2);
6580 movl(cnt2, 8);
6581 addl(cnt1, 8);
6582 bind(CONT_SCAN_SUBSTR);
6583 movdqu(vec, Address(str2, 0));
6584 jmpb(SCAN_SUBSTR);
6585
6586 bind(RET_FOUND_LONG);
6587 movptr(str1, Address(rsp, wordSize));
6588 } // non constant
6589
6590 bind(RET_FOUND);
6591 // Compute substr offset
6592 subptr(result, str1);
6593 shrl(result, 1); // index
6594
6595 bind(CLEANUP);
6596 pop(rsp); // restore SP
6597
6598 } // string_indexof
6599
6600 // Compare strings.
6601 void MacroAssembler::string_compare(Register str1, Register str2,
6602 Register cnt1, Register cnt2, Register result,
6603 XMMRegister vec1) {
6604 ShortBranchVerifier sbv(this);
6605 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
6606
6607 // Compute the minimum of the string lengths and the
6608 // difference of the string lengths (stack).
6609 // Do the conditional move stuff
6610 movl(result, cnt1);
6611 subl(cnt1, cnt2);
6612 push(cnt1);
6613 cmov32(Assembler::lessEqual, cnt2, result);
6614
6615 // Is the minimum length zero?
6616 testl(cnt2, cnt2);
6617 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6618
6619 // Compare first characters
6620 load_unsigned_short(result, Address(str1, 0));
6621 load_unsigned_short(cnt1, Address(str2, 0));
6622 subl(result, cnt1);
6623 jcc(Assembler::notZero, POP_LABEL);
6624 cmpl(cnt2, 1);
6625 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6626
6627 // Check if the strings start at the same location.
6628 cmpptr(str1, str2);
6629 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6630
6631 Address::ScaleFactor scale = Address::times_2;
6632 int stride = 8;
6633
6634 if (UseAVX >= 2 && UseSSE42Intrinsics) {
6635 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
6636 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
6637 Label COMPARE_TAIL_LONG;
6638 int pcmpmask = 0x19;
6639
6640 // Setup to compare 16-chars (32-bytes) vectors,
6641 // start from first character again because it has aligned address.
6642 int stride2 = 16;
6643 int adr_stride = stride << scale;
6644
6645 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6646 // rax and rdx are used by pcmpestri as elements counters
6647 movl(result, cnt2);
6648 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count
6649 jcc(Assembler::zero, COMPARE_TAIL_LONG);
6650
6651 // fast path : compare first 2 8-char vectors.
6652 bind(COMPARE_16_CHARS);
6653 movdqu(vec1, Address(str1, 0));
6654 pcmpestri(vec1, Address(str2, 0), pcmpmask);
6655 jccb(Assembler::below, COMPARE_INDEX_CHAR);
6656
6657 movdqu(vec1, Address(str1, adr_stride));
6658 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
6659 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
6660 addl(cnt1, stride);
6661
6662 // Compare the characters at index in cnt1
6663 bind(COMPARE_INDEX_CHAR); //cnt1 has the offset of the mismatching character
6664 load_unsigned_short(result, Address(str1, cnt1, scale));
6665 load_unsigned_short(cnt2, Address(str2, cnt1, scale));
6666 subl(result, cnt2);
6667 jmp(POP_LABEL);
6668
6669 // Setup the registers to start vector comparison loop
6670 bind(COMPARE_WIDE_VECTORS);
6671 lea(str1, Address(str1, result, scale));
6672 lea(str2, Address(str2, result, scale));
6673 subl(result, stride2);
6674 subl(cnt2, stride2);
6675 jccb(Assembler::zero, COMPARE_WIDE_TAIL);
6676 negptr(result);
6677
6678 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
6679 bind(COMPARE_WIDE_VECTORS_LOOP);
6680 vmovdqu(vec1, Address(str1, result, scale));
6681 vpxor(vec1, Address(str2, result, scale));
6682 vptest(vec1, vec1);
6683 jccb(Assembler::notZero, VECTOR_NOT_EQUAL);
6684 addptr(result, stride2);
6685 subl(cnt2, stride2);
6686 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
6687 // clean upper bits of YMM registers
6688 vpxor(vec1, vec1);
6689
6690 // compare wide vectors tail
6691 bind(COMPARE_WIDE_TAIL);
6692 testptr(result, result);
6693 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
6694
6695 movl(result, stride2);
6696 movl(cnt2, result);
6697 negptr(result);
6698 jmpb(COMPARE_WIDE_VECTORS_LOOP);
6699
6700 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
6701 bind(VECTOR_NOT_EQUAL);
6702 // clean upper bits of YMM registers
6703 vpxor(vec1, vec1);
6704 lea(str1, Address(str1, result, scale));
6705 lea(str2, Address(str2, result, scale));
6706 jmp(COMPARE_16_CHARS);
6707
6708 // Compare tail chars, length between 1 to 15 chars
6709 bind(COMPARE_TAIL_LONG);
6710 movl(cnt2, result);
6711 cmpl(cnt2, stride);
6712 jccb(Assembler::less, COMPARE_SMALL_STR);
6713
6714 movdqu(vec1, Address(str1, 0));
6715 pcmpestri(vec1, Address(str2, 0), pcmpmask);
6716 jcc(Assembler::below, COMPARE_INDEX_CHAR);
6717 subptr(cnt2, stride);
6718 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
6719 lea(str1, Address(str1, result, scale));
6720 lea(str2, Address(str2, result, scale));
6721 negptr(cnt2);
6722 jmpb(WHILE_HEAD_LABEL);
6723
6724 bind(COMPARE_SMALL_STR);
6725 } else if (UseSSE42Intrinsics) {
6726 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
6727 int pcmpmask = 0x19;
6728 // Setup to compare 8-char (16-byte) vectors,
6729 // start from first character again because it has aligned address.
6730 movl(result, cnt2);
6731 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count
6732 jccb(Assembler::zero, COMPARE_TAIL);
6733
6734 lea(str1, Address(str1, result, scale));
6735 lea(str2, Address(str2, result, scale));
6736 negptr(result);
6737
6738 // pcmpestri
6739 // inputs:
6740 // vec1- substring
6741 // rax - negative string length (elements count)
6742 // mem - scanned string
6743 // rdx - string length (elements count)
6744 // pcmpmask - cmp mode: 11000 (string compare with negated result)
6745 // + 00 (unsigned bytes) or + 01 (unsigned shorts)
6746 // outputs:
6747 // rcx - first mismatched element index
6748 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6749
6750 bind(COMPARE_WIDE_VECTORS);
6751 movdqu(vec1, Address(str1, result, scale));
6752 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6753 // After pcmpestri cnt1(rcx) contains mismatched element index
6754
6755 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1
6756 addptr(result, stride);
6757 subptr(cnt2, stride);
6758 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
6759
6760 // compare wide vectors tail
6761 testptr(result, result);
6762 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
6763
6764 movl(cnt2, stride);
6765 movl(result, stride);
6766 negptr(result);
6767 movdqu(vec1, Address(str1, result, scale));
6768 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6769 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
6770
6771 // Mismatched characters in the vectors
6772 bind(VECTOR_NOT_EQUAL);
6773 addptr(cnt1, result);
6774 load_unsigned_short(result, Address(str1, cnt1, scale));
6775 load_unsigned_short(cnt2, Address(str2, cnt1, scale));
6776 subl(result, cnt2);
6777 jmpb(POP_LABEL);
6778
6779 bind(COMPARE_TAIL); // limit is zero
6780 movl(cnt2, result);
6781 // Fallthru to tail compare
6782 }
6783 // Shift str2 and str1 to the end of the arrays, negate min
6784 lea(str1, Address(str1, cnt2, scale));
6785 lea(str2, Address(str2, cnt2, scale));
6786 decrementl(cnt2); // first character was compared already
6787 negptr(cnt2);
6788
6789 // Compare the rest of the elements
6790 bind(WHILE_HEAD_LABEL);
6791 load_unsigned_short(result, Address(str1, cnt2, scale, 0));
6792 load_unsigned_short(cnt1, Address(str2, cnt2, scale, 0));
6793 subl(result, cnt1);
6794 jccb(Assembler::notZero, POP_LABEL);
6795 increment(cnt2);
6796 jccb(Assembler::notZero, WHILE_HEAD_LABEL);
6797
6798 // Strings are equal up to min length. Return the length difference.
6799 bind(LENGTH_DIFF_LABEL);
6800 pop(result);
6801 jmpb(DONE_LABEL);
6802
6803 // Discard the stored length difference
6804 bind(POP_LABEL);
6805 pop(cnt1);
6806
6807 // That's it
6808 bind(DONE_LABEL);
6809 }
6810
6811 // Compare char[] arrays aligned to 4 bytes or substrings.
6812 void MacroAssembler::char_arrays_equals(bool is_array_equ, Register ary1, Register ary2,
6813 Register limit, Register result, Register chr,
6814 XMMRegister vec1, XMMRegister vec2) {
6815 ShortBranchVerifier sbv(this);
6816 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR;
6817
6818 int length_offset = arrayOopDesc::length_offset_in_bytes();
6819 int base_offset = arrayOopDesc::base_offset_in_bytes(T_CHAR);
6820
6821 // Check the input args
6822 cmpptr(ary1, ary2);
6823 jcc(Assembler::equal, TRUE_LABEL);
6824
6825 if (is_array_equ) {
6826 // Need additional checks for arrays_equals.
6827 testptr(ary1, ary1);
6828 jcc(Assembler::zero, FALSE_LABEL);
6829 testptr(ary2, ary2);
6830 jcc(Assembler::zero, FALSE_LABEL);
6831
6832 // Check the lengths
6833 movl(limit, Address(ary1, length_offset));
6834 cmpl(limit, Address(ary2, length_offset));
6835 jcc(Assembler::notEqual, FALSE_LABEL);
6836 }
6837
6838 // count == 0
6839 testl(limit, limit);
6840 jcc(Assembler::zero, TRUE_LABEL);
6841
6842 if (is_array_equ) {
6843 // Load array address
6844 lea(ary1, Address(ary1, base_offset));
6845 lea(ary2, Address(ary2, base_offset));
6846 }
6847
6848 shll(limit, 1); // byte count != 0
6849 movl(result, limit); // copy
6850
6851 if (UseAVX >= 2) {
6852 // With AVX2, use 32-byte vector compare
6853 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
6854
6855 // Compare 32-byte vectors
6856 andl(result, 0x0000001e); // tail count (in bytes)
6857 andl(limit, 0xffffffe0); // vector count (in bytes)
6858 jccb(Assembler::zero, COMPARE_TAIL);
6859
6860 lea(ary1, Address(ary1, limit, Address::times_1));
6861 lea(ary2, Address(ary2, limit, Address::times_1));
6862 negptr(limit);
6863
6864 bind(COMPARE_WIDE_VECTORS);
6865 vmovdqu(vec1, Address(ary1, limit, Address::times_1));
6866 vmovdqu(vec2, Address(ary2, limit, Address::times_1));
6867 vpxor(vec1, vec2);
6868
6869 vptest(vec1, vec1);
6870 jccb(Assembler::notZero, FALSE_LABEL);
6871 addptr(limit, 32);
6872 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
6873
6874 testl(result, result);
6875 jccb(Assembler::zero, TRUE_LABEL);
6876
6877 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
6878 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
6879 vpxor(vec1, vec2);
6880
6881 vptest(vec1, vec1);
6882 jccb(Assembler::notZero, FALSE_LABEL);
6883 jmpb(TRUE_LABEL);
6884
6885 bind(COMPARE_TAIL); // limit is zero
6886 movl(limit, result);
6887 // Fallthru to tail compare
6888 } else if (UseSSE42Intrinsics) {
6889 // With SSE4.2, use double quad vector compare
6890 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
6891
6892 // Compare 16-byte vectors
6893 andl(result, 0x0000000e); // tail count (in bytes)
6894 andl(limit, 0xfffffff0); // vector count (in bytes)
6895 jccb(Assembler::zero, COMPARE_TAIL);
6896
6897 lea(ary1, Address(ary1, limit, Address::times_1));
6898 lea(ary2, Address(ary2, limit, Address::times_1));
6899 negptr(limit);
6900
6901 bind(COMPARE_WIDE_VECTORS);
6902 movdqu(vec1, Address(ary1, limit, Address::times_1));
6903 movdqu(vec2, Address(ary2, limit, Address::times_1));
6904 pxor(vec1, vec2);
6905
6906 ptest(vec1, vec1);
6907 jccb(Assembler::notZero, FALSE_LABEL);
6908 addptr(limit, 16);
6909 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
6910
6911 testl(result, result);
6912 jccb(Assembler::zero, TRUE_LABEL);
6913
6914 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
6915 movdqu(vec2, Address(ary2, result, Address::times_1, -16));
6916 pxor(vec1, vec2);
6917
6918 ptest(vec1, vec1);
6919 jccb(Assembler::notZero, FALSE_LABEL);
6920 jmpb(TRUE_LABEL);
6921
6922 bind(COMPARE_TAIL); // limit is zero
6923 movl(limit, result);
6924 // Fallthru to tail compare
6925 }
6926
6927 // Compare 4-byte vectors
6928 andl(limit, 0xfffffffc); // vector count (in bytes)
6929 jccb(Assembler::zero, COMPARE_CHAR);
6930
6931 lea(ary1, Address(ary1, limit, Address::times_1));
6932 lea(ary2, Address(ary2, limit, Address::times_1));
6933 negptr(limit);
6934
6935 bind(COMPARE_VECTORS);
6936 movl(chr, Address(ary1, limit, Address::times_1));
6937 cmpl(chr, Address(ary2, limit, Address::times_1));
6938 jccb(Assembler::notEqual, FALSE_LABEL);
6939 addptr(limit, 4);
6940 jcc(Assembler::notZero, COMPARE_VECTORS);
6941
6942 // Compare trailing char (final 2 bytes), if any
6943 bind(COMPARE_CHAR);
6944 testl(result, 0x2); // tail char
6945 jccb(Assembler::zero, TRUE_LABEL);
6946 load_unsigned_short(chr, Address(ary1, 0));
6947 load_unsigned_short(limit, Address(ary2, 0));
6948 cmpl(chr, limit);
6949 jccb(Assembler::notEqual, FALSE_LABEL);
6950
6951 bind(TRUE_LABEL);
6952 movl(result, 1); // return true
6953 jmpb(DONE);
6954
6955 bind(FALSE_LABEL);
6956 xorl(result, result); // return false
6957
6958 // That's it
6959 bind(DONE);
6960 if (UseAVX >= 2) {
6961 // clean upper bits of YMM registers
6962 vpxor(vec1, vec1);
6963 vpxor(vec2, vec2);
6964 }
6965 }
6966
6967 void MacroAssembler::generate_fill(BasicType t, bool aligned,
6968 Register to, Register value, Register count,
6969 Register rtmp, XMMRegister xtmp) {
6970 ShortBranchVerifier sbv(this);
6971 assert_different_registers(to, value, count, rtmp);
6972 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
6973 Label L_fill_2_bytes, L_fill_4_bytes;
6974
6975 int shift = -1;
6976 switch (t) {
6977 case T_BYTE:
6978 shift = 2;
6979 break;
6980 case T_SHORT:
6981 shift = 1;
6982 break;
6983 case T_INT:
6984 shift = 0;
6985 break;
6986 default: ShouldNotReachHere();
6987 }
6988
6989 if (t == T_BYTE) {
6990 andl(value, 0xff);
6991 movl(rtmp, value);
6992 shll(rtmp, 8);
6993 orl(value, rtmp);
6994 }
6995 if (t == T_SHORT) {
6996 andl(value, 0xffff);
6997 }
6998 if (t == T_BYTE || t == T_SHORT) {
6999 movl(rtmp, value);
7000 shll(rtmp, 16);
7001 orl(value, rtmp);
7002 }
7003
7004 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
7005 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp
7006 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) {
7007 // align source address at 4 bytes address boundary
7008 if (t == T_BYTE) {
7009 // One byte misalignment happens only for byte arrays
7010 testptr(to, 1);
7011 jccb(Assembler::zero, L_skip_align1);
7012 movb(Address(to, 0), value);
7013 increment(to);
7014 decrement(count);
7015 BIND(L_skip_align1);
7016 }
7017 // Two bytes misalignment happens only for byte and short (char) arrays
7018 testptr(to, 2);
7019 jccb(Assembler::zero, L_skip_align2);
7020 movw(Address(to, 0), value);
7021 addptr(to, 2);
7022 subl(count, 1<<(shift-1));
7023 BIND(L_skip_align2);
7024 }
7025 if (UseSSE < 2) {
7026 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7027 // Fill 32-byte chunks
7028 subl(count, 8 << shift);
7029 jcc(Assembler::less, L_check_fill_8_bytes);
7030 align(16);
7031
7032 BIND(L_fill_32_bytes_loop);
7033
7034 for (int i = 0; i < 32; i += 4) {
7035 movl(Address(to, i), value);
7036 }
7037
7038 addptr(to, 32);
7039 subl(count, 8 << shift);
7040 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7041 BIND(L_check_fill_8_bytes);
7042 addl(count, 8 << shift);
7043 jccb(Assembler::zero, L_exit);
7044 jmpb(L_fill_8_bytes);
7045
7046 //
7047 // length is too short, just fill qwords
7048 //
7049 BIND(L_fill_8_bytes_loop);
7050 movl(Address(to, 0), value);
7051 movl(Address(to, 4), value);
7052 addptr(to, 8);
7053 BIND(L_fill_8_bytes);
7054 subl(count, 1 << (shift + 1));
7055 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7056 // fall through to fill 4 bytes
7057 } else {
7058 Label L_fill_32_bytes;
7059 if (!UseUnalignedLoadStores) {
7060 // align to 8 bytes, we know we are 4 byte aligned to start
7061 testptr(to, 4);
7062 jccb(Assembler::zero, L_fill_32_bytes);
7063 movl(Address(to, 0), value);
7064 addptr(to, 4);
7065 subl(count, 1<<shift);
7066 }
7067 BIND(L_fill_32_bytes);
7068 {
7069 assert( UseSSE >= 2, "supported cpu only" );
7070 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7071 movdl(xtmp, value);
7072 if (UseAVX >= 2 && UseUnalignedLoadStores) {
7073 // Fill 64-byte chunks
7074 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
7075 vpbroadcastd(xtmp, xtmp);
7076
7077 subl(count, 16 << shift);
7078 jcc(Assembler::less, L_check_fill_32_bytes);
7079 align(16);
7080
7081 BIND(L_fill_64_bytes_loop);
7082 vmovdqu(Address(to, 0), xtmp);
7083 vmovdqu(Address(to, 32), xtmp);
7084 addptr(to, 64);
7085 subl(count, 16 << shift);
7086 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
7087
7088 BIND(L_check_fill_32_bytes);
7089 addl(count, 8 << shift);
7090 jccb(Assembler::less, L_check_fill_8_bytes);
7091 vmovdqu(Address(to, 0), xtmp);
7092 addptr(to, 32);
7093 subl(count, 8 << shift);
7094
7095 BIND(L_check_fill_8_bytes);
7096 // clean upper bits of YMM registers
7097 movdl(xtmp, value);
7098 pshufd(xtmp, xtmp, 0);
7099 } else {
7100 // Fill 32-byte chunks
7101 pshufd(xtmp, xtmp, 0);
7102
7103 subl(count, 8 << shift);
7104 jcc(Assembler::less, L_check_fill_8_bytes);
7105 align(16);
7106
7107 BIND(L_fill_32_bytes_loop);
7108
7109 if (UseUnalignedLoadStores) {
7110 movdqu(Address(to, 0), xtmp);
7111 movdqu(Address(to, 16), xtmp);
7112 } else {
7113 movq(Address(to, 0), xtmp);
7114 movq(Address(to, 8), xtmp);
7115 movq(Address(to, 16), xtmp);
7116 movq(Address(to, 24), xtmp);
7117 }
7118
7119 addptr(to, 32);
7120 subl(count, 8 << shift);
7121 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7122
7123 BIND(L_check_fill_8_bytes);
7124 }
7125 addl(count, 8 << shift);
7126 jccb(Assembler::zero, L_exit);
7127 jmpb(L_fill_8_bytes);
7128
7129 //
7130 // length is too short, just fill qwords
7131 //
7132 BIND(L_fill_8_bytes_loop);
7133 movq(Address(to, 0), xtmp);
7134 addptr(to, 8);
7135 BIND(L_fill_8_bytes);
7136 subl(count, 1 << (shift + 1));
7137 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7138 }
7139 }
7140 // fill trailing 4 bytes
7141 BIND(L_fill_4_bytes);
7142 testl(count, 1<<shift);
7143 jccb(Assembler::zero, L_fill_2_bytes);
7144 movl(Address(to, 0), value);
7145 if (t == T_BYTE || t == T_SHORT) {
7146 addptr(to, 4);
7147 BIND(L_fill_2_bytes);
7148 // fill trailing 2 bytes
7149 testl(count, 1<<(shift-1));
7150 jccb(Assembler::zero, L_fill_byte);
7151 movw(Address(to, 0), value);
7152 if (t == T_BYTE) {
7153 addptr(to, 2);
7154 BIND(L_fill_byte);
7155 // fill trailing byte
7156 testl(count, 1);
7157 jccb(Assembler::zero, L_exit);
7158 movb(Address(to, 0), value);
7159 } else {
7160 BIND(L_fill_byte);
7161 }
7162 } else {
7163 BIND(L_fill_2_bytes);
7164 }
7165 BIND(L_exit);
7166 }
7167
7168 // encode char[] to byte[] in ISO_8859_1
7169 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len,
7170 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
7171 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
7172 Register tmp5, Register result) {
7173 // rsi: src
7174 // rdi: dst
7175 // rdx: len
7176 // rcx: tmp5
7177 // rax: result
7178 ShortBranchVerifier sbv(this);
7179 assert_different_registers(src, dst, len, tmp5, result);
7180 Label L_done, L_copy_1_char, L_copy_1_char_exit;
7181
7182 // set result
7183 xorl(result, result);
7184 // check for zero length
7185 testl(len, len);
7186 jcc(Assembler::zero, L_done);
7187 movl(result, len);
7188
7189 // Setup pointers
7190 lea(src, Address(src, len, Address::times_2)); // char[]
7191 lea(dst, Address(dst, len, Address::times_1)); // byte[]
7192 negptr(len);
7193
7194 if (UseSSE42Intrinsics || UseAVX >= 2) {
7195 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit;
7196 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit;
7197
7198 if (UseAVX >= 2) {
7199 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit;
7200 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
7201 movdl(tmp1Reg, tmp5);
7202 vpbroadcastd(tmp1Reg, tmp1Reg);
7203 jmpb(L_chars_32_check);
7204
7205 bind(L_copy_32_chars);
7206 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64));
7207 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32));
7208 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector256 */ true);
7209 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
7210 jccb(Assembler::notZero, L_copy_32_chars_exit);
7211 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector256 */ true);
7212 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector256 */ true);
7213 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg);
7214
7215 bind(L_chars_32_check);
7216 addptr(len, 32);
7217 jccb(Assembler::lessEqual, L_copy_32_chars);
7218
7219 bind(L_copy_32_chars_exit);
7220 subptr(len, 16);
7221 jccb(Assembler::greater, L_copy_16_chars_exit);
7222
7223 } else if (UseSSE42Intrinsics) {
7224 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
7225 movdl(tmp1Reg, tmp5);
7226 pshufd(tmp1Reg, tmp1Reg, 0);
7227 jmpb(L_chars_16_check);
7228 }
7229
7230 bind(L_copy_16_chars);
7231 if (UseAVX >= 2) {
7232 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32));
7233 vptest(tmp2Reg, tmp1Reg);
7234 jccb(Assembler::notZero, L_copy_16_chars_exit);
7235 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector256 */ true);
7236 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector256 */ true);
7237 } else {
7238 if (UseAVX > 0) {
7239 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7240 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7241 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector256 */ false);
7242 } else {
7243 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7244 por(tmp2Reg, tmp3Reg);
7245 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7246 por(tmp2Reg, tmp4Reg);
7247 }
7248 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
7249 jccb(Assembler::notZero, L_copy_16_chars_exit);
7250 packuswb(tmp3Reg, tmp4Reg);
7251 }
7252 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg);
7253
7254 bind(L_chars_16_check);
7255 addptr(len, 16);
7256 jccb(Assembler::lessEqual, L_copy_16_chars);
7257
7258 bind(L_copy_16_chars_exit);
7259 if (UseAVX >= 2) {
7260 // clean upper bits of YMM registers
7261 vpxor(tmp2Reg, tmp2Reg);
7262 vpxor(tmp3Reg, tmp3Reg);
7263 vpxor(tmp4Reg, tmp4Reg);
7264 movdl(tmp1Reg, tmp5);
7265 pshufd(tmp1Reg, tmp1Reg, 0);
7266 }
7267 subptr(len, 8);
7268 jccb(Assembler::greater, L_copy_8_chars_exit);
7269
7270 bind(L_copy_8_chars);
7271 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16));
7272 ptest(tmp3Reg, tmp1Reg);
7273 jccb(Assembler::notZero, L_copy_8_chars_exit);
7274 packuswb(tmp3Reg, tmp1Reg);
7275 movq(Address(dst, len, Address::times_1, -8), tmp3Reg);
7276 addptr(len, 8);
7277 jccb(Assembler::lessEqual, L_copy_8_chars);
7278
7279 bind(L_copy_8_chars_exit);
7280 subptr(len, 8);
7281 jccb(Assembler::zero, L_done);
7282 }
7283
7284 bind(L_copy_1_char);
7285 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0));
7286 testl(tmp5, 0xff00); // check if Unicode char
7287 jccb(Assembler::notZero, L_copy_1_char_exit);
7288 movb(Address(dst, len, Address::times_1, 0), tmp5);
7289 addptr(len, 1);
7290 jccb(Assembler::less, L_copy_1_char);
7291
7292 bind(L_copy_1_char_exit);
7293 addptr(result, len); // len is negative count of not processed elements
7294 bind(L_done);
7295 }
7296
7297 #ifdef _LP64
7298 /**
7299 * Helper for multiply_to_len().
7300 */
7301 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) {
7302 addq(dest_lo, src1);
7303 adcq(dest_hi, 0);
7304 addq(dest_lo, src2);
7305 adcq(dest_hi, 0);
7306 }
7307
7308 /**
7309 * Multiply 64 bit by 64 bit first loop.
7310 */
7311 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
7312 Register y, Register y_idx, Register z,
7313 Register carry, Register product,
7314 Register idx, Register kdx) {
7315 //
7316 // jlong carry, x[], y[], z[];
7317 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7318 // huge_128 product = y[idx] * x[xstart] + carry;
7319 // z[kdx] = (jlong)product;
7320 // carry = (jlong)(product >>> 64);
7321 // }
7322 // z[xstart] = carry;
7323 //
7324
7325 Label L_first_loop, L_first_loop_exit;
7326 Label L_one_x, L_one_y, L_multiply;
7327
7328 decrementl(xstart);
7329 jcc(Assembler::negative, L_one_x);
7330
7331 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
7332 rorq(x_xstart, 32); // convert big-endian to little-endian
7333
7334 bind(L_first_loop);
7335 decrementl(idx);
7336 jcc(Assembler::negative, L_first_loop_exit);
7337 decrementl(idx);
7338 jcc(Assembler::negative, L_one_y);
7339 movq(y_idx, Address(y, idx, Address::times_4, 0));
7340 rorq(y_idx, 32); // convert big-endian to little-endian
7341 bind(L_multiply);
7342 movq(product, x_xstart);
7343 mulq(y_idx); // product(rax) * y_idx -> rdx:rax
7344 addq(product, carry);
7345 adcq(rdx, 0);
7346 subl(kdx, 2);
7347 movl(Address(z, kdx, Address::times_4, 4), product);
7348 shrq(product, 32);
7349 movl(Address(z, kdx, Address::times_4, 0), product);
7350 movq(carry, rdx);
7351 jmp(L_first_loop);
7352
7353 bind(L_one_y);
7354 movl(y_idx, Address(y, 0));
7355 jmp(L_multiply);
7356
7357 bind(L_one_x);
7358 movl(x_xstart, Address(x, 0));
7359 jmp(L_first_loop);
7360
7361 bind(L_first_loop_exit);
7362 }
7363
7364 /**
7365 * Multiply 64 bit by 64 bit and add 128 bit.
7366 */
7367 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z,
7368 Register yz_idx, Register idx,
7369 Register carry, Register product, int offset) {
7370 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
7371 // z[kdx] = (jlong)product;
7372
7373 movq(yz_idx, Address(y, idx, Address::times_4, offset));
7374 rorq(yz_idx, 32); // convert big-endian to little-endian
7375 movq(product, x_xstart);
7376 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7377 movq(yz_idx, Address(z, idx, Address::times_4, offset));
7378 rorq(yz_idx, 32); // convert big-endian to little-endian
7379
7380 add2_with_carry(rdx, product, carry, yz_idx);
7381
7382 movl(Address(z, idx, Address::times_4, offset+4), product);
7383 shrq(product, 32);
7384 movl(Address(z, idx, Address::times_4, offset), product);
7385
7386 }
7387
7388 /**
7389 * Multiply 128 bit by 128 bit. Unrolled inner loop.
7390 */
7391 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z,
7392 Register yz_idx, Register idx, Register jdx,
7393 Register carry, Register product,
7394 Register carry2) {
7395 // jlong carry, x[], y[], z[];
7396 // int kdx = ystart+1;
7397 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7398 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
7399 // z[kdx+idx+1] = (jlong)product;
7400 // jlong carry2 = (jlong)(product >>> 64);
7401 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
7402 // z[kdx+idx] = (jlong)product;
7403 // carry = (jlong)(product >>> 64);
7404 // }
7405 // idx += 2;
7406 // if (idx > 0) {
7407 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
7408 // z[kdx+idx] = (jlong)product;
7409 // carry = (jlong)(product >>> 64);
7410 // }
7411 //
7412
7413 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7414
7415 movl(jdx, idx);
7416 andl(jdx, 0xFFFFFFFC);
7417 shrl(jdx, 2);
7418
7419 bind(L_third_loop);
7420 subl(jdx, 1);
7421 jcc(Assembler::negative, L_third_loop_exit);
7422 subl(idx, 4);
7423
7424 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
7425 movq(carry2, rdx);
7426
7427 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
7428 movq(carry, rdx);
7429 jmp(L_third_loop);
7430
7431 bind (L_third_loop_exit);
7432
7433 andl (idx, 0x3);
7434 jcc(Assembler::zero, L_post_third_loop_done);
7435
7436 Label L_check_1;
7437 subl(idx, 2);
7438 jcc(Assembler::negative, L_check_1);
7439
7440 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
7441 movq(carry, rdx);
7442
7443 bind (L_check_1);
7444 addl (idx, 0x2);
7445 andl (idx, 0x1);
7446 subl(idx, 1);
7447 jcc(Assembler::negative, L_post_third_loop_done);
7448
7449 movl(yz_idx, Address(y, idx, Address::times_4, 0));
7450 movq(product, x_xstart);
7451 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7452 movl(yz_idx, Address(z, idx, Address::times_4, 0));
7453
7454 add2_with_carry(rdx, product, yz_idx, carry);
7455
7456 movl(Address(z, idx, Address::times_4, 0), product);
7457 shrq(product, 32);
7458
7459 shlq(rdx, 32);
7460 orq(product, rdx);
7461 movq(carry, product);
7462
7463 bind(L_post_third_loop_done);
7464 }
7465
7466 /**
7467 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop.
7468 *
7469 */
7470 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z,
7471 Register carry, Register carry2,
7472 Register idx, Register jdx,
7473 Register yz_idx1, Register yz_idx2,
7474 Register tmp, Register tmp3, Register tmp4) {
7475 assert(UseBMI2Instructions, "should be used only when BMI2 is available");
7476
7477 // jlong carry, x[], y[], z[];
7478 // int kdx = ystart+1;
7479 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7480 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry;
7481 // jlong carry2 = (jlong)(tmp3 >>> 64);
7482 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2;
7483 // carry = (jlong)(tmp4 >>> 64);
7484 // z[kdx+idx+1] = (jlong)tmp3;
7485 // z[kdx+idx] = (jlong)tmp4;
7486 // }
7487 // idx += 2;
7488 // if (idx > 0) {
7489 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry;
7490 // z[kdx+idx] = (jlong)yz_idx1;
7491 // carry = (jlong)(yz_idx1 >>> 64);
7492 // }
7493 //
7494
7495 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7496
7497 movl(jdx, idx);
7498 andl(jdx, 0xFFFFFFFC);
7499 shrl(jdx, 2);
7500
7501 bind(L_third_loop);
7502 subl(jdx, 1);
7503 jcc(Assembler::negative, L_third_loop_exit);
7504 subl(idx, 4);
7505
7506 movq(yz_idx1, Address(y, idx, Address::times_4, 8));
7507 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
7508 movq(yz_idx2, Address(y, idx, Address::times_4, 0));
7509 rorxq(yz_idx2, yz_idx2, 32);
7510
7511 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
7512 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp
7513
7514 movq(yz_idx1, Address(z, idx, Address::times_4, 8));
7515 rorxq(yz_idx1, yz_idx1, 32);
7516 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
7517 rorxq(yz_idx2, yz_idx2, 32);
7518
7519 if (VM_Version::supports_adx()) {
7520 adcxq(tmp3, carry);
7521 adoxq(tmp3, yz_idx1);
7522
7523 adcxq(tmp4, tmp);
7524 adoxq(tmp4, yz_idx2);
7525
7526 movl(carry, 0); // does not affect flags
7527 adcxq(carry2, carry);
7528 adoxq(carry2, carry);
7529 } else {
7530 add2_with_carry(tmp4, tmp3, carry, yz_idx1);
7531 add2_with_carry(carry2, tmp4, tmp, yz_idx2);
7532 }
7533 movq(carry, carry2);
7534
7535 movl(Address(z, idx, Address::times_4, 12), tmp3);
7536 shrq(tmp3, 32);
7537 movl(Address(z, idx, Address::times_4, 8), tmp3);
7538
7539 movl(Address(z, idx, Address::times_4, 4), tmp4);
7540 shrq(tmp4, 32);
7541 movl(Address(z, idx, Address::times_4, 0), tmp4);
7542
7543 jmp(L_third_loop);
7544
7545 bind (L_third_loop_exit);
7546
7547 andl (idx, 0x3);
7548 jcc(Assembler::zero, L_post_third_loop_done);
7549
7550 Label L_check_1;
7551 subl(idx, 2);
7552 jcc(Assembler::negative, L_check_1);
7553
7554 movq(yz_idx1, Address(y, idx, Address::times_4, 0));
7555 rorxq(yz_idx1, yz_idx1, 32);
7556 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
7557 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
7558 rorxq(yz_idx2, yz_idx2, 32);
7559
7560 add2_with_carry(tmp4, tmp3, carry, yz_idx2);
7561
7562 movl(Address(z, idx, Address::times_4, 4), tmp3);
7563 shrq(tmp3, 32);
7564 movl(Address(z, idx, Address::times_4, 0), tmp3);
7565 movq(carry, tmp4);
7566
7567 bind (L_check_1);
7568 addl (idx, 0x2);
7569 andl (idx, 0x1);
7570 subl(idx, 1);
7571 jcc(Assembler::negative, L_post_third_loop_done);
7572 movl(tmp4, Address(y, idx, Address::times_4, 0));
7573 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3
7574 movl(tmp4, Address(z, idx, Address::times_4, 0));
7575
7576 add2_with_carry(carry2, tmp3, tmp4, carry);
7577
7578 movl(Address(z, idx, Address::times_4, 0), tmp3);
7579 shrq(tmp3, 32);
7580
7581 shlq(carry2, 32);
7582 orq(tmp3, carry2);
7583 movq(carry, tmp3);
7584
7585 bind(L_post_third_loop_done);
7586 }
7587
7588 /**
7589 * Code for BigInteger::multiplyToLen() instrinsic.
7590 *
7591 * rdi: x
7592 * rax: xlen
7593 * rsi: y
7594 * rcx: ylen
7595 * r8: z
7596 * r11: zlen
7597 * r12: tmp1
7598 * r13: tmp2
7599 * r14: tmp3
7600 * r15: tmp4
7601 * rbx: tmp5
7602 *
7603 */
7604 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen,
7605 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) {
7606 ShortBranchVerifier sbv(this);
7607 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx);
7608
7609 push(tmp1);
7610 push(tmp2);
7611 push(tmp3);
7612 push(tmp4);
7613 push(tmp5);
7614
7615 push(xlen);
7616 push(zlen);
7617
7618 const Register idx = tmp1;
7619 const Register kdx = tmp2;
7620 const Register xstart = tmp3;
7621
7622 const Register y_idx = tmp4;
7623 const Register carry = tmp5;
7624 const Register product = xlen;
7625 const Register x_xstart = zlen; // reuse register
7626
7627 // First Loop.
7628 //
7629 // final static long LONG_MASK = 0xffffffffL;
7630 // int xstart = xlen - 1;
7631 // int ystart = ylen - 1;
7632 // long carry = 0;
7633 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7634 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
7635 // z[kdx] = (int)product;
7636 // carry = product >>> 32;
7637 // }
7638 // z[xstart] = (int)carry;
7639 //
7640
7641 movl(idx, ylen); // idx = ylen;
7642 movl(kdx, zlen); // kdx = xlen+ylen;
7643 xorq(carry, carry); // carry = 0;
7644
7645 Label L_done;
7646
7647 movl(xstart, xlen);
7648 decrementl(xstart);
7649 jcc(Assembler::negative, L_done);
7650
7651 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
7652
7653 Label L_second_loop;
7654 testl(kdx, kdx);
7655 jcc(Assembler::zero, L_second_loop);
7656
7657 Label L_carry;
7658 subl(kdx, 1);
7659 jcc(Assembler::zero, L_carry);
7660
7661 movl(Address(z, kdx, Address::times_4, 0), carry);
7662 shrq(carry, 32);
7663 subl(kdx, 1);
7664
7665 bind(L_carry);
7666 movl(Address(z, kdx, Address::times_4, 0), carry);
7667
7668 // Second and third (nested) loops.
7669 //
7670 // for (int i = xstart-1; i >= 0; i--) { // Second loop
7671 // carry = 0;
7672 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
7673 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
7674 // (z[k] & LONG_MASK) + carry;
7675 // z[k] = (int)product;
7676 // carry = product >>> 32;
7677 // }
7678 // z[i] = (int)carry;
7679 // }
7680 //
7681 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
7682
7683 const Register jdx = tmp1;
7684
7685 bind(L_second_loop);
7686 xorl(carry, carry); // carry = 0;
7687 movl(jdx, ylen); // j = ystart+1
7688
7689 subl(xstart, 1); // i = xstart-1;
7690 jcc(Assembler::negative, L_done);
7691
7692 push (z);
7693
7694 Label L_last_x;
7695 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j
7696 subl(xstart, 1); // i = xstart-1;
7697 jcc(Assembler::negative, L_last_x);
7698
7699 if (UseBMI2Instructions) {
7700 movq(rdx, Address(x, xstart, Address::times_4, 0));
7701 rorxq(rdx, rdx, 32); // convert big-endian to little-endian
7702 } else {
7703 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
7704 rorq(x_xstart, 32); // convert big-endian to little-endian
7705 }
7706
7707 Label L_third_loop_prologue;
7708 bind(L_third_loop_prologue);
7709
7710 push (x);
7711 push (xstart);
7712 push (ylen);
7713
7714
7715 if (UseBMI2Instructions) {
7716 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4);
7717 } else { // !UseBMI2Instructions
7718 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
7719 }
7720
7721 pop(ylen);
7722 pop(xlen);
7723 pop(x);
7724 pop(z);
7725
7726 movl(tmp3, xlen);
7727 addl(tmp3, 1);
7728 movl(Address(z, tmp3, Address::times_4, 0), carry);
7729 subl(tmp3, 1);
7730 jccb(Assembler::negative, L_done);
7731
7732 shrq(carry, 32);
7733 movl(Address(z, tmp3, Address::times_4, 0), carry);
7734 jmp(L_second_loop);
7735
7736 // Next infrequent code is moved outside loops.
7737 bind(L_last_x);
7738 if (UseBMI2Instructions) {
7739 movl(rdx, Address(x, 0));
7740 } else {
7741 movl(x_xstart, Address(x, 0));
7742 }
7743 jmp(L_third_loop_prologue);
7744
7745 bind(L_done);
7746
7747 pop(zlen);
7748 pop(xlen);
7749
7750 pop(tmp5);
7751 pop(tmp4);
7752 pop(tmp3);
7753 pop(tmp2);
7754 pop(tmp1);
7755 }
7756 #endif
7757
7758 /**
7759 * Emits code to update CRC-32 with a byte value according to constants in table
7760 *
7761 * @param [in,out]crc Register containing the crc.
7762 * @param [in]val Register containing the byte to fold into the CRC.
7763 * @param [in]table Register containing the table of crc constants.
7764 *
7765 * uint32_t crc;
7766 * val = crc_table[(val ^ crc) & 0xFF];
7767 * crc = val ^ (crc >> 8);
7768 *
7769 */
7770 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
7771 xorl(val, crc);
7772 andl(val, 0xFF);
7773 shrl(crc, 8); // unsigned shift
7774 xorl(crc, Address(table, val, Address::times_4, 0));
7775 }
7776
7777 /**
7778 * Fold 128-bit data chunk
7779 */
7780 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
7781 if (UseAVX > 0) {
7782 vpclmulhdq(xtmp, xK, xcrc); // [123:64]
7783 vpclmulldq(xcrc, xK, xcrc); // [63:0]
7784 vpxor(xcrc, xcrc, Address(buf, offset), false /* vector256 */);
7785 pxor(xcrc, xtmp);
7786 } else {
7787 movdqa(xtmp, xcrc);
7788 pclmulhdq(xtmp, xK); // [123:64]
7789 pclmulldq(xcrc, xK); // [63:0]
7790 pxor(xcrc, xtmp);
7791 movdqu(xtmp, Address(buf, offset));
7792 pxor(xcrc, xtmp);
7793 }
7794 }
7795
7796 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
7797 if (UseAVX > 0) {
7798 vpclmulhdq(xtmp, xK, xcrc);
7799 vpclmulldq(xcrc, xK, xcrc);
7800 pxor(xcrc, xbuf);
7801 pxor(xcrc, xtmp);
7802 } else {
7803 movdqa(xtmp, xcrc);
7804 pclmulhdq(xtmp, xK);
7805 pclmulldq(xcrc, xK);
7806 pxor(xcrc, xbuf);
7807 pxor(xcrc, xtmp);
7808 }
7809 }
7810
7811 /**
7812 * 8-bit folds to compute 32-bit CRC
7813 *
7814 * uint64_t xcrc;
7815 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8);
7816 */
7817 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) {
7818 movdl(tmp, xcrc);
7819 andl(tmp, 0xFF);
7820 movdl(xtmp, Address(table, tmp, Address::times_4, 0));
7821 psrldq(xcrc, 1); // unsigned shift one byte
7822 pxor(xcrc, xtmp);
7823 }
7824
7825 /**
7826 * uint32_t crc;
7827 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
7828 */
7829 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
7830 movl(tmp, crc);
7831 andl(tmp, 0xFF);
7832 shrl(crc, 8);
7833 xorl(crc, Address(table, tmp, Address::times_4, 0));
7834 }
7835
7836 /**
7837 * @param crc register containing existing CRC (32-bit)
7838 * @param buf register pointing to input byte buffer (byte*)
7839 * @param len register containing number of bytes
7840 * @param table register that will contain address of CRC table
7841 * @param tmp scratch register
7842 */
7843 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) {
7844 assert_different_registers(crc, buf, len, table, tmp, rax);
7845
7846 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned;
7847 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop;
7848
7849 lea(table, ExternalAddress(StubRoutines::crc_table_addr()));
7850 notl(crc); // ~crc
7851 cmpl(len, 16);
7852 jcc(Assembler::less, L_tail);
7853
7854 // Align buffer to 16 bytes
7855 movl(tmp, buf);
7856 andl(tmp, 0xF);
7857 jccb(Assembler::zero, L_aligned);
7858 subl(tmp, 16);
7859 addl(len, tmp);
7860
7861 align(4);
7862 BIND(L_align_loop);
7863 movsbl(rax, Address(buf, 0)); // load byte with sign extension
7864 update_byte_crc32(crc, rax, table);
7865 increment(buf);
7866 incrementl(tmp);
7867 jccb(Assembler::less, L_align_loop);
7868
7869 BIND(L_aligned);
7870 movl(tmp, len); // save
7871 shrl(len, 4);
7872 jcc(Assembler::zero, L_tail_restore);
7873
7874 // Fold crc into first bytes of vector
7875 movdqa(xmm1, Address(buf, 0));
7876 movdl(rax, xmm1);
7877 xorl(crc, rax);
7878 pinsrd(xmm1, crc, 0);
7879 addptr(buf, 16);
7880 subl(len, 4); // len > 0
7881 jcc(Assembler::less, L_fold_tail);
7882
7883 movdqa(xmm2, Address(buf, 0));
7884 movdqa(xmm3, Address(buf, 16));
7885 movdqa(xmm4, Address(buf, 32));
7886 addptr(buf, 48);
7887 subl(len, 3);
7888 jcc(Assembler::lessEqual, L_fold_512b);
7889
7890 // Fold total 512 bits of polynomial on each iteration,
7891 // 128 bits per each of 4 parallel streams.
7892 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
7893
7894 align(32);
7895 BIND(L_fold_512b_loop);
7896 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
7897 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16);
7898 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32);
7899 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48);
7900 addptr(buf, 64);
7901 subl(len, 4);
7902 jcc(Assembler::greater, L_fold_512b_loop);
7903
7904 // Fold 512 bits to 128 bits.
7905 BIND(L_fold_512b);
7906 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
7907 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2);
7908 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3);
7909 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4);
7910
7911 // Fold the rest of 128 bits data chunks
7912 BIND(L_fold_tail);
7913 addl(len, 3);
7914 jccb(Assembler::lessEqual, L_fold_128b);
7915 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
7916
7917 BIND(L_fold_tail_loop);
7918 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
7919 addptr(buf, 16);
7920 decrementl(len);
7921 jccb(Assembler::greater, L_fold_tail_loop);
7922
7923 // Fold 128 bits in xmm1 down into 32 bits in crc register.
7924 BIND(L_fold_128b);
7925 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
7926 if (UseAVX > 0) {
7927 vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
7928 vpand(xmm3, xmm0, xmm2, false /* vector256 */);
7929 vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
7930 } else {
7931 movdqa(xmm2, xmm0);
7932 pclmulqdq(xmm2, xmm1, 0x1);
7933 movdqa(xmm3, xmm0);
7934 pand(xmm3, xmm2);
7935 pclmulqdq(xmm0, xmm3, 0x1);
7936 }
7937 psrldq(xmm1, 8);
7938 psrldq(xmm2, 4);
7939 pxor(xmm0, xmm1);
7940 pxor(xmm0, xmm2);
7941
7942 // 8 8-bit folds to compute 32-bit CRC.
7943 for (int j = 0; j < 4; j++) {
7944 fold_8bit_crc32(xmm0, table, xmm1, rax);
7945 }
7946 movdl(crc, xmm0); // mov 32 bits to general register
7947 for (int j = 0; j < 4; j++) {
7948 fold_8bit_crc32(crc, table, rax);
7949 }
7950
7951 BIND(L_tail_restore);
7952 movl(len, tmp); // restore
7953 BIND(L_tail);
7954 andl(len, 0xf);
7955 jccb(Assembler::zero, L_exit);
7956
7957 // Fold the rest of bytes
7958 align(4);
7959 BIND(L_tail_loop);
7960 movsbl(rax, Address(buf, 0)); // load byte with sign extension
7961 update_byte_crc32(crc, rax, table);
7962 increment(buf);
7963 decrementl(len);
7964 jccb(Assembler::greater, L_tail_loop);
7965
7966 BIND(L_exit);
7967 notl(crc); // ~c
7968 }
7969
7970 #undef BIND
7971 #undef BLOCK_COMMENT
7972
7973
7974 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
7975 switch (cond) {
7976 // Note some conditions are synonyms for others
7977 case Assembler::zero: return Assembler::notZero;
7978 case Assembler::notZero: return Assembler::zero;
7979 case Assembler::less: return Assembler::greaterEqual;
7980 case Assembler::lessEqual: return Assembler::greater;
7981 case Assembler::greater: return Assembler::lessEqual;
7982 case Assembler::greaterEqual: return Assembler::less;
7983 case Assembler::below: return Assembler::aboveEqual;
7984 case Assembler::belowEqual: return Assembler::above;
7985 case Assembler::above: return Assembler::belowEqual;
7986 case Assembler::aboveEqual: return Assembler::below;
7987 case Assembler::overflow: return Assembler::noOverflow;
7988 case Assembler::noOverflow: return Assembler::overflow;
7989 case Assembler::negative: return Assembler::positive;
7990 case Assembler::positive: return Assembler::negative;
7991 case Assembler::parity: return Assembler::noParity;
7992 case Assembler::noParity: return Assembler::parity;
7993 }
7994 ShouldNotReachHere(); return Assembler::overflow;
7995 }
7996
7997 SkipIfEqual::SkipIfEqual(
7998 MacroAssembler* masm, const bool* flag_addr, bool value) {
7999 _masm = masm;
8000 _masm->cmp8(ExternalAddress((address)flag_addr), value);
8001 _masm->jcc(Assembler::equal, _label);
8002 }
8003
8004 SkipIfEqual::~SkipIfEqual() {
8005 _masm->bind(_label);
8006 }