1 /*
2 * Copyright (c) 1997, 2019, 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 "jvm.h"
27 #include "asm/assembler.hpp"
28 #include "asm/assembler.inline.hpp"
29 #include "compiler/disassembler.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/barrierSetAssembler.hpp"
32 #include "gc/shared/collectedHeap.inline.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "memory/universe.hpp"
36 #include "oops/accessDecorators.hpp"
37 #include "oops/compressedOops.inline.hpp"
38 #include "oops/klass.inline.hpp"
39 #include "prims/methodHandles.hpp"
40 #include "runtime/biasedLocking.hpp"
41 #include "runtime/flags/flagSetting.hpp"
42 #include "runtime/interfaceSupport.inline.hpp"
43 #include "runtime/objectMonitor.hpp"
44 #include "runtime/os.hpp"
45 #include "runtime/safepoint.hpp"
46 #include "runtime/safepointMechanism.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #include "runtime/stubRoutines.hpp"
49 #include "runtime/thread.hpp"
50 #include "utilities/macros.hpp"
51 #include "crc32c.h"
52 #ifdef COMPILER2
53 #include "opto/intrinsicnode.hpp"
54 #endif
55
56 #ifdef PRODUCT
57 #define BLOCK_COMMENT(str) /* nothing */
58 #define STOP(error) stop(error)
59 #else
60 #define BLOCK_COMMENT(str) block_comment(str)
61 #define STOP(error) block_comment(error); stop(error)
62 #endif
63
64 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
65
66 #ifdef ASSERT
67 bool AbstractAssembler::pd_check_instruction_mark() { return true; }
68 #endif
69
70 static Assembler::Condition reverse[] = {
71 Assembler::noOverflow /* overflow = 0x0 */ ,
72 Assembler::overflow /* noOverflow = 0x1 */ ,
73 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ ,
74 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ ,
75 Assembler::notZero /* zero = 0x4, equal = 0x4 */ ,
76 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ ,
77 Assembler::above /* belowEqual = 0x6 */ ,
78 Assembler::belowEqual /* above = 0x7 */ ,
79 Assembler::positive /* negative = 0x8 */ ,
80 Assembler::negative /* positive = 0x9 */ ,
81 Assembler::noParity /* parity = 0xa */ ,
82 Assembler::parity /* noParity = 0xb */ ,
83 Assembler::greaterEqual /* less = 0xc */ ,
84 Assembler::less /* greaterEqual = 0xd */ ,
85 Assembler::greater /* lessEqual = 0xe */ ,
86 Assembler::lessEqual /* greater = 0xf, */
87
88 };
89
90
91 // Implementation of MacroAssembler
92
93 // First all the versions that have distinct versions depending on 32/64 bit
94 // Unless the difference is trivial (1 line or so).
95
96 #ifndef _LP64
97
98 // 32bit versions
99
100 Address MacroAssembler::as_Address(AddressLiteral adr) {
101 return Address(adr.target(), adr.rspec());
102 }
103
104 Address MacroAssembler::as_Address(ArrayAddress adr) {
105 return Address::make_array(adr);
106 }
107
108 void MacroAssembler::call_VM_leaf_base(address entry_point,
109 int number_of_arguments) {
110 call(RuntimeAddress(entry_point));
111 increment(rsp, number_of_arguments * wordSize);
112 }
113
114 void MacroAssembler::cmpklass(Address src1, Metadata* obj) {
115 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
116 }
117
118 void MacroAssembler::cmpklass(Register src1, Metadata* obj) {
119 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
120 }
121
122 void MacroAssembler::cmpoop_raw(Address src1, jobject obj) {
123 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
124 }
125
126 void MacroAssembler::cmpoop_raw(Register src1, jobject obj) {
127 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
128 }
129
130 void MacroAssembler::cmpoop(Address src1, jobject obj) {
131 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
132 bs->obj_equals(this, src1, obj);
133 }
134
135 void MacroAssembler::cmpoop(Register src1, jobject obj) {
136 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
137 bs->obj_equals(this, src1, obj);
138 }
139
140 void MacroAssembler::extend_sign(Register hi, Register lo) {
141 // According to Intel Doc. AP-526, "Integer Divide", p.18.
142 if (VM_Version::is_P6() && hi == rdx && lo == rax) {
143 cdql();
144 } else {
145 movl(hi, lo);
146 sarl(hi, 31);
147 }
148 }
149
150 void MacroAssembler::jC2(Register tmp, Label& L) {
151 // set parity bit if FPU flag C2 is set (via rax)
152 save_rax(tmp);
153 fwait(); fnstsw_ax();
154 sahf();
155 restore_rax(tmp);
156 // branch
157 jcc(Assembler::parity, L);
158 }
159
160 void MacroAssembler::jnC2(Register tmp, Label& L) {
161 // set parity bit if FPU flag C2 is set (via rax)
162 save_rax(tmp);
163 fwait(); fnstsw_ax();
164 sahf();
165 restore_rax(tmp);
166 // branch
167 jcc(Assembler::noParity, L);
168 }
169
170 // 32bit can do a case table jump in one instruction but we no longer allow the base
171 // to be installed in the Address class
172 void MacroAssembler::jump(ArrayAddress entry) {
173 jmp(as_Address(entry));
174 }
175
176 // Note: y_lo will be destroyed
177 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
178 // Long compare for Java (semantics as described in JVM spec.)
179 Label high, low, done;
180
181 cmpl(x_hi, y_hi);
182 jcc(Assembler::less, low);
183 jcc(Assembler::greater, high);
184 // x_hi is the return register
185 xorl(x_hi, x_hi);
186 cmpl(x_lo, y_lo);
187 jcc(Assembler::below, low);
188 jcc(Assembler::equal, done);
189
190 bind(high);
191 xorl(x_hi, x_hi);
192 increment(x_hi);
193 jmp(done);
194
195 bind(low);
196 xorl(x_hi, x_hi);
197 decrementl(x_hi);
198
199 bind(done);
200 }
201
202 void MacroAssembler::lea(Register dst, AddressLiteral src) {
203 mov_literal32(dst, (int32_t)src.target(), src.rspec());
204 }
205
206 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
207 // leal(dst, as_Address(adr));
208 // see note in movl as to why we must use a move
209 mov_literal32(dst, (int32_t) adr.target(), adr.rspec());
210 }
211
212 void MacroAssembler::leave() {
213 mov(rsp, rbp);
214 pop(rbp);
215 }
216
217 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) {
218 // Multiplication of two Java long values stored on the stack
219 // as illustrated below. Result is in rdx:rax.
220 //
221 // rsp ---> [ ?? ] \ \
222 // .... | y_rsp_offset |
223 // [ y_lo ] / (in bytes) | x_rsp_offset
224 // [ y_hi ] | (in bytes)
225 // .... |
226 // [ x_lo ] /
227 // [ x_hi ]
228 // ....
229 //
230 // Basic idea: lo(result) = lo(x_lo * y_lo)
231 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
232 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset);
233 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset);
234 Label quick;
235 // load x_hi, y_hi and check if quick
236 // multiplication is possible
237 movl(rbx, x_hi);
238 movl(rcx, y_hi);
239 movl(rax, rbx);
240 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0
241 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply
242 // do full multiplication
243 // 1st step
244 mull(y_lo); // x_hi * y_lo
245 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx,
246 // 2nd step
247 movl(rax, x_lo);
248 mull(rcx); // x_lo * y_hi
249 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx,
250 // 3rd step
251 bind(quick); // note: rbx, = 0 if quick multiply!
252 movl(rax, x_lo);
253 mull(y_lo); // x_lo * y_lo
254 addl(rdx, rbx); // correct hi(x_lo * y_lo)
255 }
256
257 void MacroAssembler::lneg(Register hi, Register lo) {
258 negl(lo);
259 adcl(hi, 0);
260 negl(hi);
261 }
262
263 void MacroAssembler::lshl(Register hi, Register lo) {
264 // Java shift left long support (semantics as described in JVM spec., p.305)
265 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n))
266 // shift value is in rcx !
267 assert(hi != rcx, "must not use rcx");
268 assert(lo != rcx, "must not use rcx");
269 const Register s = rcx; // shift count
270 const int n = BitsPerWord;
271 Label L;
272 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
273 cmpl(s, n); // if (s < n)
274 jcc(Assembler::less, L); // else (s >= n)
275 movl(hi, lo); // x := x << n
276 xorl(lo, lo);
277 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
278 bind(L); // s (mod n) < n
279 shldl(hi, lo); // x := x << s
280 shll(lo);
281 }
282
283
284 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) {
285 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310)
286 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n))
287 assert(hi != rcx, "must not use rcx");
288 assert(lo != rcx, "must not use rcx");
289 const Register s = rcx; // shift count
290 const int n = BitsPerWord;
291 Label L;
292 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
293 cmpl(s, n); // if (s < n)
294 jcc(Assembler::less, L); // else (s >= n)
295 movl(lo, hi); // x := x >> n
296 if (sign_extension) sarl(hi, 31);
297 else xorl(hi, hi);
298 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
299 bind(L); // s (mod n) < n
300 shrdl(lo, hi); // x := x >> s
301 if (sign_extension) sarl(hi);
302 else shrl(hi);
303 }
304
305 void MacroAssembler::movoop(Register dst, jobject obj) {
306 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
307 }
308
309 void MacroAssembler::movoop(Address dst, jobject obj) {
310 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
311 }
312
313 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
314 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
315 }
316
317 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
318 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
319 }
320
321 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
322 // scratch register is not used,
323 // it is defined to match parameters of 64-bit version of this method.
324 if (src.is_lval()) {
325 mov_literal32(dst, (intptr_t)src.target(), src.rspec());
326 } else {
327 movl(dst, as_Address(src));
328 }
329 }
330
331 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
332 movl(as_Address(dst), src);
333 }
334
335 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
336 movl(dst, as_Address(src));
337 }
338
339 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
340 void MacroAssembler::movptr(Address dst, intptr_t src) {
341 movl(dst, src);
342 }
343
344
345 void MacroAssembler::pop_callee_saved_registers() {
346 pop(rcx);
347 pop(rdx);
348 pop(rdi);
349 pop(rsi);
350 }
351
352 void MacroAssembler::push_callee_saved_registers() {
353 push(rsi);
354 push(rdi);
355 push(rdx);
356 push(rcx);
357 }
358
359 void MacroAssembler::pushoop(jobject obj) {
360 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate());
361 }
362
363 void MacroAssembler::pushklass(Metadata* obj) {
364 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate());
365 }
366
367 void MacroAssembler::pushptr(AddressLiteral src) {
368 if (src.is_lval()) {
369 push_literal32((int32_t)src.target(), src.rspec());
370 } else {
371 pushl(as_Address(src));
372 }
373 }
374
375 void MacroAssembler::set_word_if_not_zero(Register dst) {
376 xorl(dst, dst);
377 set_byte_if_not_zero(dst);
378 }
379
380 static void pass_arg0(MacroAssembler* masm, Register arg) {
381 masm->push(arg);
382 }
383
384 static void pass_arg1(MacroAssembler* masm, Register arg) {
385 masm->push(arg);
386 }
387
388 static void pass_arg2(MacroAssembler* masm, Register arg) {
389 masm->push(arg);
390 }
391
392 static void pass_arg3(MacroAssembler* masm, Register arg) {
393 masm->push(arg);
394 }
395
396 #ifndef PRODUCT
397 extern "C" void findpc(intptr_t x);
398 #endif
399
400 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) {
401 // In order to get locks to work, we need to fake a in_VM state
402 JavaThread* thread = JavaThread::current();
403 JavaThreadState saved_state = thread->thread_state();
404 thread->set_thread_state(_thread_in_vm);
405 if (ShowMessageBoxOnError) {
406 JavaThread* thread = JavaThread::current();
407 JavaThreadState saved_state = thread->thread_state();
408 thread->set_thread_state(_thread_in_vm);
409 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
410 ttyLocker ttyl;
411 BytecodeCounter::print();
412 }
413 // To see where a verify_oop failed, get $ebx+40/X for this frame.
414 // This is the value of eip which points to where verify_oop will return.
415 if (os::message_box(msg, "Execution stopped, print registers?")) {
416 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip);
417 BREAKPOINT;
418 }
419 }
420 fatal("DEBUG MESSAGE: %s", msg);
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 // we must set sp to zero to clear frame
753 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
754 // must clear fp, so that compiled frames are not confused; it is
755 // possible that we need it only for debugging
756 if (clear_fp) {
757 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
758 }
759
760 // Always clear the pc because it could have been set by make_walkable()
761 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
762 vzeroupper();
763 }
764
765 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
766 Register last_java_fp,
767 address last_java_pc) {
768 vzeroupper();
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 if (ShowMessageBoxOnError) {
817 address rip = pc();
818 pusha(); // get regs on stack
819 lea(c_rarg1, InternalAddress(rip));
820 movq(c_rarg2, rsp); // pass pointer to regs array
821 }
822 lea(c_rarg0, ExternalAddress((address) msg));
823 andq(rsp, -16); // align stack as required by ABI
824 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
825 hlt();
826 }
827
828 void MacroAssembler::warn(const char* msg) {
829 push(rbp);
830 movq(rbp, rsp);
831 andq(rsp, -16); // align stack as required by push_CPU_state and call
832 push_CPU_state(); // keeps alignment at 16 bytes
833 lea(c_rarg0, ExternalAddress((address) msg));
834 lea(rax, ExternalAddress(CAST_FROM_FN_PTR(address, warning)));
835 call(rax);
836 pop_CPU_state();
837 mov(rsp, rbp);
838 pop(rbp);
839 }
840
841 void MacroAssembler::print_state() {
842 address rip = pc();
843 pusha(); // get regs on stack
844 push(rbp);
845 movq(rbp, rsp);
846 andq(rsp, -16); // align stack as required by push_CPU_state and call
847 push_CPU_state(); // keeps alignment at 16 bytes
848
849 lea(c_rarg0, InternalAddress(rip));
850 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array
851 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1);
852
853 pop_CPU_state();
854 mov(rsp, rbp);
855 pop(rbp);
856 popa();
857 }
858
859 #ifndef PRODUCT
860 extern "C" void findpc(intptr_t x);
861 #endif
862
863 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) {
864 // In order to get locks to work, we need to fake a in_VM state
865 if (ShowMessageBoxOnError) {
866 JavaThread* thread = JavaThread::current();
867 JavaThreadState saved_state = thread->thread_state();
868 thread->set_thread_state(_thread_in_vm);
869 #ifndef PRODUCT
870 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
871 ttyLocker ttyl;
872 BytecodeCounter::print();
873 }
874 #endif
875 // To see where a verify_oop failed, get $ebx+40/X for this frame.
876 // XXX correct this offset for amd64
877 // This is the value of eip which points to where verify_oop will return.
878 if (os::message_box(msg, "Execution stopped, print registers?")) {
879 print_state64(pc, regs);
880 BREAKPOINT;
881 }
882 }
883 fatal("DEBUG MESSAGE: %s", msg);
884 }
885
886 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) {
887 ttyLocker ttyl;
888 FlagSetting fs(Debugging, true);
889 tty->print_cr("rip = 0x%016lx", (intptr_t)pc);
890 #ifndef PRODUCT
891 tty->cr();
892 findpc(pc);
893 tty->cr();
894 #endif
895 #define PRINT_REG(rax, value) \
896 { tty->print("%s = ", #rax); os::print_location(tty, value); }
897 PRINT_REG(rax, regs[15]);
898 PRINT_REG(rbx, regs[12]);
899 PRINT_REG(rcx, regs[14]);
900 PRINT_REG(rdx, regs[13]);
901 PRINT_REG(rdi, regs[8]);
902 PRINT_REG(rsi, regs[9]);
903 PRINT_REG(rbp, regs[10]);
904 PRINT_REG(rsp, regs[11]);
905 PRINT_REG(r8 , regs[7]);
906 PRINT_REG(r9 , regs[6]);
907 PRINT_REG(r10, regs[5]);
908 PRINT_REG(r11, regs[4]);
909 PRINT_REG(r12, regs[3]);
910 PRINT_REG(r13, regs[2]);
911 PRINT_REG(r14, regs[1]);
912 PRINT_REG(r15, regs[0]);
913 #undef PRINT_REG
914 // Print some words near top of staack.
915 int64_t* rsp = (int64_t*) regs[11];
916 int64_t* dump_sp = rsp;
917 for (int col1 = 0; col1 < 8; col1++) {
918 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
919 os::print_location(tty, *dump_sp++);
920 }
921 for (int row = 0; row < 25; row++) {
922 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
923 for (int col = 0; col < 4; col++) {
924 tty->print(" 0x%016lx", (intptr_t)*dump_sp++);
925 }
926 tty->cr();
927 }
928 // Print some instructions around pc:
929 Disassembler::decode((address)pc-64, (address)pc);
930 tty->print_cr("--------");
931 Disassembler::decode((address)pc, (address)pc+32);
932 }
933
934 #endif // _LP64
935
936 // Now versions that are common to 32/64 bit
937
938 void MacroAssembler::addptr(Register dst, int32_t imm32) {
939 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32));
940 }
941
942 void MacroAssembler::addptr(Register dst, Register src) {
943 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
944 }
945
946 void MacroAssembler::addptr(Address dst, Register src) {
947 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
948 }
949
950 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) {
951 if (reachable(src)) {
952 Assembler::addsd(dst, as_Address(src));
953 } else {
954 lea(rscratch1, src);
955 Assembler::addsd(dst, Address(rscratch1, 0));
956 }
957 }
958
959 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) {
960 if (reachable(src)) {
961 addss(dst, as_Address(src));
962 } else {
963 lea(rscratch1, src);
964 addss(dst, Address(rscratch1, 0));
965 }
966 }
967
968 void MacroAssembler::addpd(XMMRegister dst, AddressLiteral src) {
969 if (reachable(src)) {
970 Assembler::addpd(dst, as_Address(src));
971 } else {
972 lea(rscratch1, src);
973 Assembler::addpd(dst, Address(rscratch1, 0));
974 }
975 }
976
977 void MacroAssembler::align(int modulus) {
978 align(modulus, offset());
979 }
980
981 void MacroAssembler::align(int modulus, int target) {
982 if (target % modulus != 0) {
983 nop(modulus - (target % modulus));
984 }
985 }
986
987 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
988 // Used in sign-masking with aligned address.
989 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
990 if (reachable(src)) {
991 Assembler::andpd(dst, as_Address(src));
992 } else {
993 lea(scratch_reg, src);
994 Assembler::andpd(dst, Address(scratch_reg, 0));
995 }
996 }
997
998 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
999 // Used in sign-masking with aligned address.
1000 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
1001 if (reachable(src)) {
1002 Assembler::andps(dst, as_Address(src));
1003 } else {
1004 lea(scratch_reg, src);
1005 Assembler::andps(dst, Address(scratch_reg, 0));
1006 }
1007 }
1008
1009 void MacroAssembler::andptr(Register dst, int32_t imm32) {
1010 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32));
1011 }
1012
1013 void MacroAssembler::atomic_incl(Address counter_addr) {
1014 lock();
1015 incrementl(counter_addr);
1016 }
1017
1018 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) {
1019 if (reachable(counter_addr)) {
1020 atomic_incl(as_Address(counter_addr));
1021 } else {
1022 lea(scr, counter_addr);
1023 atomic_incl(Address(scr, 0));
1024 }
1025 }
1026
1027 #ifdef _LP64
1028 void MacroAssembler::atomic_incq(Address counter_addr) {
1029 lock();
1030 incrementq(counter_addr);
1031 }
1032
1033 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) {
1034 if (reachable(counter_addr)) {
1035 atomic_incq(as_Address(counter_addr));
1036 } else {
1037 lea(scr, counter_addr);
1038 atomic_incq(Address(scr, 0));
1039 }
1040 }
1041 #endif
1042
1043 // Writes to stack successive pages until offset reached to check for
1044 // stack overflow + shadow pages. This clobbers tmp.
1045 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
1046 movptr(tmp, rsp);
1047 // Bang stack for total size given plus shadow page size.
1048 // Bang one page at a time because large size can bang beyond yellow and
1049 // red zones.
1050 Label loop;
1051 bind(loop);
1052 movl(Address(tmp, (-os::vm_page_size())), size );
1053 subptr(tmp, os::vm_page_size());
1054 subl(size, os::vm_page_size());
1055 jcc(Assembler::greater, loop);
1056
1057 // Bang down shadow pages too.
1058 // At this point, (tmp-0) is the last address touched, so don't
1059 // touch it again. (It was touched as (tmp-pagesize) but then tmp
1060 // was post-decremented.) Skip this address by starting at i=1, and
1061 // touch a few more pages below. N.B. It is important to touch all
1062 // the way down including all pages in the shadow zone.
1063 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) {
1064 // this could be any sized move but this is can be a debugging crumb
1065 // so the bigger the better.
1066 movptr(Address(tmp, (-i*os::vm_page_size())), size );
1067 }
1068 }
1069
1070 void MacroAssembler::reserved_stack_check() {
1071 // testing if reserved zone needs to be enabled
1072 Label no_reserved_zone_enabling;
1073 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread);
1074 NOT_LP64(get_thread(rsi);)
1075
1076 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset()));
1077 jcc(Assembler::below, no_reserved_zone_enabling);
1078
1079 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread);
1080 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry()));
1081 should_not_reach_here();
1082
1083 bind(no_reserved_zone_enabling);
1084 }
1085
1086 int MacroAssembler::biased_locking_enter(Register lock_reg,
1087 Register obj_reg,
1088 Register swap_reg,
1089 Register tmp_reg,
1090 bool swap_reg_contains_mark,
1091 Label& done,
1092 Label* slow_case,
1093 BiasedLockingCounters* counters) {
1094 assert(UseBiasedLocking, "why call this otherwise?");
1095 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq");
1096 assert(tmp_reg != noreg, "tmp_reg must be supplied");
1097 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg);
1098 assert(markWord::age_shift == markWord::lock_bits + markWord::biased_lock_bits, "biased locking makes assumptions about bit layout");
1099 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes());
1100 NOT_LP64( Address saved_mark_addr(lock_reg, 0); )
1101
1102 if (PrintBiasedLockingStatistics && counters == NULL) {
1103 counters = BiasedLocking::counters();
1104 }
1105 // Biased locking
1106 // See whether the lock is currently biased toward our thread and
1107 // whether the epoch is still valid
1108 // Note that the runtime guarantees sufficient alignment of JavaThread
1109 // pointers to allow age to be placed into low bits
1110 // First check to see whether biasing is even enabled for this object
1111 Label cas_label;
1112 int null_check_offset = -1;
1113 if (!swap_reg_contains_mark) {
1114 null_check_offset = offset();
1115 movptr(swap_reg, mark_addr);
1116 }
1117 movptr(tmp_reg, swap_reg);
1118 andptr(tmp_reg, markWord::biased_lock_mask_in_place);
1119 cmpptr(tmp_reg, markWord::biased_lock_pattern);
1120 jcc(Assembler::notEqual, cas_label);
1121 // The bias pattern is present in the object's header. Need to check
1122 // whether the bias owner and the epoch are both still current.
1123 #ifndef _LP64
1124 // Note that because there is no current thread register on x86_32 we
1125 // need to store off the mark word we read out of the object to
1126 // avoid reloading it and needing to recheck invariants below. This
1127 // store is unfortunate but it makes the overall code shorter and
1128 // simpler.
1129 movptr(saved_mark_addr, swap_reg);
1130 #endif
1131 if (swap_reg_contains_mark) {
1132 null_check_offset = offset();
1133 }
1134 load_prototype_header(tmp_reg, obj_reg);
1135 #ifdef _LP64
1136 orptr(tmp_reg, r15_thread);
1137 xorptr(tmp_reg, swap_reg);
1138 Register header_reg = tmp_reg;
1139 #else
1140 xorptr(tmp_reg, swap_reg);
1141 get_thread(swap_reg);
1142 xorptr(swap_reg, tmp_reg);
1143 Register header_reg = swap_reg;
1144 #endif
1145 andptr(header_reg, ~((int) markWord::age_mask_in_place));
1146 if (counters != NULL) {
1147 cond_inc32(Assembler::zero,
1148 ExternalAddress((address) counters->biased_lock_entry_count_addr()));
1149 }
1150 jcc(Assembler::equal, done);
1151
1152 Label try_revoke_bias;
1153 Label try_rebias;
1154
1155 // At this point we know that the header has the bias pattern and
1156 // that we are not the bias owner in the current epoch. We need to
1157 // figure out more details about the state of the header in order to
1158 // know what operations can be legally performed on the object's
1159 // header.
1160
1161 // If the low three bits in the xor result aren't clear, that means
1162 // the prototype header is no longer biased and we have to revoke
1163 // the bias on this object.
1164 testptr(header_reg, markWord::biased_lock_mask_in_place);
1165 jccb(Assembler::notZero, try_revoke_bias);
1166
1167 // Biasing is still enabled for this data type. See whether the
1168 // epoch of the current bias is still valid, meaning that the epoch
1169 // bits of the mark word are equal to the epoch bits of the
1170 // prototype header. (Note that the prototype header's epoch bits
1171 // only change at a safepoint.) If not, attempt to rebias the object
1172 // toward the current thread. Note that we must be absolutely sure
1173 // that the current epoch is invalid in order to do this because
1174 // otherwise the manipulations it performs on the mark word are
1175 // illegal.
1176 testptr(header_reg, markWord::epoch_mask_in_place);
1177 jccb(Assembler::notZero, try_rebias);
1178
1179 // The epoch of the current bias is still valid but we know nothing
1180 // about the owner; it might be set or it might be clear. Try to
1181 // acquire the bias of the object using an atomic operation. If this
1182 // fails we will go in to the runtime to revoke the object's bias.
1183 // Note that we first construct the presumed unbiased header so we
1184 // don't accidentally blow away another thread's valid bias.
1185 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1186 andptr(swap_reg,
1187 markWord::biased_lock_mask_in_place | markWord::age_mask_in_place | markWord::epoch_mask_in_place);
1188 #ifdef _LP64
1189 movptr(tmp_reg, swap_reg);
1190 orptr(tmp_reg, r15_thread);
1191 #else
1192 get_thread(tmp_reg);
1193 orptr(tmp_reg, swap_reg);
1194 #endif
1195 lock();
1196 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1197 // If the biasing toward our thread failed, this means that
1198 // another thread succeeded in biasing it toward itself and we
1199 // need to revoke that bias. The revocation will occur in the
1200 // interpreter runtime in the slow case.
1201 if (counters != NULL) {
1202 cond_inc32(Assembler::zero,
1203 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr()));
1204 }
1205 if (slow_case != NULL) {
1206 jcc(Assembler::notZero, *slow_case);
1207 }
1208 jmp(done);
1209
1210 bind(try_rebias);
1211 // At this point we know the epoch has expired, meaning that the
1212 // current "bias owner", if any, is actually invalid. Under these
1213 // circumstances _only_, we are allowed to use the current header's
1214 // value as the comparison value when doing the cas to acquire the
1215 // bias in the current epoch. In other words, we allow transfer of
1216 // the bias from one thread to another directly in this situation.
1217 //
1218 // FIXME: due to a lack of registers we currently blow away the age
1219 // bits in this situation. Should attempt to preserve them.
1220 load_prototype_header(tmp_reg, obj_reg);
1221 #ifdef _LP64
1222 orptr(tmp_reg, r15_thread);
1223 #else
1224 get_thread(swap_reg);
1225 orptr(tmp_reg, swap_reg);
1226 movptr(swap_reg, saved_mark_addr);
1227 #endif
1228 lock();
1229 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1230 // If the biasing toward our thread failed, then another thread
1231 // succeeded in biasing it toward itself and we need to revoke that
1232 // bias. The revocation will occur in the runtime in the slow case.
1233 if (counters != NULL) {
1234 cond_inc32(Assembler::zero,
1235 ExternalAddress((address) counters->rebiased_lock_entry_count_addr()));
1236 }
1237 if (slow_case != NULL) {
1238 jcc(Assembler::notZero, *slow_case);
1239 }
1240 jmp(done);
1241
1242 bind(try_revoke_bias);
1243 // The prototype mark in the klass doesn't have the bias bit set any
1244 // more, indicating that objects of this data type are not supposed
1245 // to be biased any more. We are going to try to reset the mark of
1246 // this object to the prototype value and fall through to the
1247 // CAS-based locking scheme. Note that if our CAS fails, it means
1248 // that another thread raced us for the privilege of revoking the
1249 // bias of this particular object, so it's okay to continue in the
1250 // normal locking code.
1251 //
1252 // FIXME: due to a lack of registers we currently blow away the age
1253 // bits in this situation. Should attempt to preserve them.
1254 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1255 load_prototype_header(tmp_reg, obj_reg);
1256 lock();
1257 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1258 // Fall through to the normal CAS-based lock, because no matter what
1259 // the result of the above CAS, some thread must have succeeded in
1260 // removing the bias bit from the object's header.
1261 if (counters != NULL) {
1262 cond_inc32(Assembler::zero,
1263 ExternalAddress((address) counters->revoked_lock_entry_count_addr()));
1264 }
1265
1266 bind(cas_label);
1267
1268 return null_check_offset;
1269 }
1270
1271 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) {
1272 assert(UseBiasedLocking, "why call this otherwise?");
1273
1274 // Check for biased locking unlock case, which is a no-op
1275 // Note: we do not have to check the thread ID for two reasons.
1276 // First, the interpreter checks for IllegalMonitorStateException at
1277 // a higher level. Second, if the bias was revoked while we held the
1278 // lock, the object could not be rebiased toward another thread, so
1279 // the bias bit would be clear.
1280 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1281 andptr(temp_reg, markWord::biased_lock_mask_in_place);
1282 cmpptr(temp_reg, markWord::biased_lock_pattern);
1283 jcc(Assembler::equal, done);
1284 }
1285
1286 #ifdef COMPILER2
1287
1288 #if INCLUDE_RTM_OPT
1289
1290 // Update rtm_counters based on abort status
1291 // input: abort_status
1292 // rtm_counters (RTMLockingCounters*)
1293 // flags are killed
1294 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) {
1295
1296 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset()));
1297 if (PrintPreciseRTMLockingStatistics) {
1298 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) {
1299 Label check_abort;
1300 testl(abort_status, (1<<i));
1301 jccb(Assembler::equal, check_abort);
1302 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx))));
1303 bind(check_abort);
1304 }
1305 }
1306 }
1307
1308 // Branch if (random & (count-1) != 0), count is 2^n
1309 // tmp, scr and flags are killed
1310 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) {
1311 assert(tmp == rax, "");
1312 assert(scr == rdx, "");
1313 rdtsc(); // modifies EDX:EAX
1314 andptr(tmp, count-1);
1315 jccb(Assembler::notZero, brLabel);
1316 }
1317
1318 // Perform abort ratio calculation, set no_rtm bit if high ratio
1319 // input: rtm_counters_Reg (RTMLockingCounters* address)
1320 // tmpReg, rtm_counters_Reg and flags are killed
1321 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg,
1322 Register rtm_counters_Reg,
1323 RTMLockingCounters* rtm_counters,
1324 Metadata* method_data) {
1325 Label L_done, L_check_always_rtm1, L_check_always_rtm2;
1326
1327 if (RTMLockingCalculationDelay > 0) {
1328 // Delay calculation
1329 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg);
1330 testptr(tmpReg, tmpReg);
1331 jccb(Assembler::equal, L_done);
1332 }
1333 // Abort ratio calculation only if abort_count > RTMAbortThreshold
1334 // Aborted transactions = abort_count * 100
1335 // All transactions = total_count * RTMTotalCountIncrRate
1336 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio)
1337
1338 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset()));
1339 cmpptr(tmpReg, RTMAbortThreshold);
1340 jccb(Assembler::below, L_check_always_rtm2);
1341 imulptr(tmpReg, tmpReg, 100);
1342
1343 Register scrReg = rtm_counters_Reg;
1344 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1345 imulptr(scrReg, scrReg, RTMTotalCountIncrRate);
1346 imulptr(scrReg, scrReg, RTMAbortRatio);
1347 cmpptr(tmpReg, scrReg);
1348 jccb(Assembler::below, L_check_always_rtm1);
1349 if (method_data != NULL) {
1350 // set rtm_state to "no rtm" in MDO
1351 mov_metadata(tmpReg, method_data);
1352 lock();
1353 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM);
1354 }
1355 jmpb(L_done);
1356 bind(L_check_always_rtm1);
1357 // Reload RTMLockingCounters* address
1358 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1359 bind(L_check_always_rtm2);
1360 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1361 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate);
1362 jccb(Assembler::below, L_done);
1363 if (method_data != NULL) {
1364 // set rtm_state to "always rtm" in MDO
1365 mov_metadata(tmpReg, method_data);
1366 lock();
1367 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM);
1368 }
1369 bind(L_done);
1370 }
1371
1372 // Update counters and perform abort ratio calculation
1373 // input: abort_status_Reg
1374 // rtm_counters_Reg, flags are killed
1375 void MacroAssembler::rtm_profiling(Register abort_status_Reg,
1376 Register rtm_counters_Reg,
1377 RTMLockingCounters* rtm_counters,
1378 Metadata* method_data,
1379 bool profile_rtm) {
1380
1381 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1382 // update rtm counters based on rax value at abort
1383 // reads abort_status_Reg, updates flags
1384 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1385 rtm_counters_update(abort_status_Reg, rtm_counters_Reg);
1386 if (profile_rtm) {
1387 // Save abort status because abort_status_Reg is used by following code.
1388 if (RTMRetryCount > 0) {
1389 push(abort_status_Reg);
1390 }
1391 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1392 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data);
1393 // restore abort status
1394 if (RTMRetryCount > 0) {
1395 pop(abort_status_Reg);
1396 }
1397 }
1398 }
1399
1400 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4)
1401 // inputs: retry_count_Reg
1402 // : abort_status_Reg
1403 // output: retry_count_Reg decremented by 1
1404 // flags are killed
1405 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) {
1406 Label doneRetry;
1407 assert(abort_status_Reg == rax, "");
1408 // The abort reason bits are in eax (see all states in rtmLocking.hpp)
1409 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4)
1410 // if reason is in 0x6 and retry count != 0 then retry
1411 andptr(abort_status_Reg, 0x6);
1412 jccb(Assembler::zero, doneRetry);
1413 testl(retry_count_Reg, retry_count_Reg);
1414 jccb(Assembler::zero, doneRetry);
1415 pause();
1416 decrementl(retry_count_Reg);
1417 jmp(retryLabel);
1418 bind(doneRetry);
1419 }
1420
1421 // Spin and retry if lock is busy,
1422 // inputs: box_Reg (monitor address)
1423 // : retry_count_Reg
1424 // output: retry_count_Reg decremented by 1
1425 // : clear z flag if retry count exceeded
1426 // tmp_Reg, scr_Reg, flags are killed
1427 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg,
1428 Register tmp_Reg, Register scr_Reg, Label& retryLabel) {
1429 Label SpinLoop, SpinExit, doneRetry;
1430 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1431
1432 testl(retry_count_Reg, retry_count_Reg);
1433 jccb(Assembler::zero, doneRetry);
1434 decrementl(retry_count_Reg);
1435 movptr(scr_Reg, RTMSpinLoopCount);
1436
1437 bind(SpinLoop);
1438 pause();
1439 decrementl(scr_Reg);
1440 jccb(Assembler::lessEqual, SpinExit);
1441 movptr(tmp_Reg, Address(box_Reg, owner_offset));
1442 testptr(tmp_Reg, tmp_Reg);
1443 jccb(Assembler::notZero, SpinLoop);
1444
1445 bind(SpinExit);
1446 jmp(retryLabel);
1447 bind(doneRetry);
1448 incrementl(retry_count_Reg); // clear z flag
1449 }
1450
1451 // Use RTM for normal stack locks
1452 // Input: objReg (object to lock)
1453 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg,
1454 Register retry_on_abort_count_Reg,
1455 RTMLockingCounters* stack_rtm_counters,
1456 Metadata* method_data, bool profile_rtm,
1457 Label& DONE_LABEL, Label& IsInflated) {
1458 assert(UseRTMForStackLocks, "why call this otherwise?");
1459 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1460 assert(tmpReg == rax, "");
1461 assert(scrReg == rdx, "");
1462 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1463
1464 if (RTMRetryCount > 0) {
1465 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1466 bind(L_rtm_retry);
1467 }
1468 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes()));
1469 testptr(tmpReg, markWord::monitor_value); // inflated vs stack-locked|neutral|biased
1470 jcc(Assembler::notZero, IsInflated);
1471
1472 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1473 Label L_noincrement;
1474 if (RTMTotalCountIncrRate > 1) {
1475 // tmpReg, scrReg and flags are killed
1476 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement);
1477 }
1478 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM");
1479 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg);
1480 bind(L_noincrement);
1481 }
1482 xbegin(L_on_abort);
1483 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword
1484 andptr(tmpReg, markWord::biased_lock_mask_in_place); // look at 3 lock bits
1485 cmpptr(tmpReg, markWord::unlocked_value); // bits = 001 unlocked
1486 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked
1487
1488 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1489 if (UseRTMXendForLockBusy) {
1490 xend();
1491 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry)
1492 jmp(L_decrement_retry);
1493 }
1494 else {
1495 xabort(0);
1496 }
1497 bind(L_on_abort);
1498 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1499 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm);
1500 }
1501 bind(L_decrement_retry);
1502 if (RTMRetryCount > 0) {
1503 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1504 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1505 }
1506 }
1507
1508 // Use RTM for inflating locks
1509 // inputs: objReg (object to lock)
1510 // boxReg (on-stack box address (displaced header location) - KILLED)
1511 // tmpReg (ObjectMonitor address + markWord::monitor_value)
1512 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg,
1513 Register scrReg, Register retry_on_busy_count_Reg,
1514 Register retry_on_abort_count_Reg,
1515 RTMLockingCounters* rtm_counters,
1516 Metadata* method_data, bool profile_rtm,
1517 Label& DONE_LABEL) {
1518 assert(UseRTMLocking, "why call this otherwise?");
1519 assert(tmpReg == rax, "");
1520 assert(scrReg == rdx, "");
1521 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1522 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1523
1524 // Without cast to int32_t this style of movptr will destroy r10 which is typically obj.
1525 movptr(Address(boxReg, 0), (int32_t)intptr_t(markWord::unused_mark().value()));
1526 movptr(boxReg, tmpReg); // Save ObjectMonitor address
1527
1528 if (RTMRetryCount > 0) {
1529 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy
1530 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1531 bind(L_rtm_retry);
1532 }
1533 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1534 Label L_noincrement;
1535 if (RTMTotalCountIncrRate > 1) {
1536 // tmpReg, scrReg and flags are killed
1537 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement);
1538 }
1539 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1540 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg);
1541 bind(L_noincrement);
1542 }
1543 xbegin(L_on_abort);
1544 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes()));
1545 movptr(tmpReg, Address(tmpReg, owner_offset));
1546 testptr(tmpReg, tmpReg);
1547 jcc(Assembler::zero, DONE_LABEL);
1548 if (UseRTMXendForLockBusy) {
1549 xend();
1550 jmp(L_decrement_retry);
1551 }
1552 else {
1553 xabort(0);
1554 }
1555 bind(L_on_abort);
1556 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1557 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1558 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm);
1559 }
1560 if (RTMRetryCount > 0) {
1561 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1562 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1563 }
1564
1565 movptr(tmpReg, Address(boxReg, owner_offset)) ;
1566 testptr(tmpReg, tmpReg) ;
1567 jccb(Assembler::notZero, L_decrement_retry) ;
1568
1569 // Appears unlocked - try to swing _owner from null to non-null.
1570 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1571 #ifdef _LP64
1572 Register threadReg = r15_thread;
1573 #else
1574 get_thread(scrReg);
1575 Register threadReg = scrReg;
1576 #endif
1577 lock();
1578 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg
1579
1580 if (RTMRetryCount > 0) {
1581 // success done else retry
1582 jccb(Assembler::equal, DONE_LABEL) ;
1583 bind(L_decrement_retry);
1584 // Spin and retry if lock is busy.
1585 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry);
1586 }
1587 else {
1588 bind(L_decrement_retry);
1589 }
1590 }
1591
1592 #endif // INCLUDE_RTM_OPT
1593
1594 // fast_lock and fast_unlock used by C2
1595
1596 // Because the transitions from emitted code to the runtime
1597 // monitorenter/exit helper stubs are so slow it's critical that
1598 // we inline both the stack-locking fast path and the inflated fast path.
1599 //
1600 // See also: cmpFastLock and cmpFastUnlock.
1601 //
1602 // What follows is a specialized inline transliteration of the code
1603 // in enter() and exit(). If we're concerned about I$ bloat another
1604 // option would be to emit TrySlowEnter and TrySlowExit methods
1605 // at startup-time. These methods would accept arguments as
1606 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
1607 // indications in the icc.ZFlag. fast_lock and fast_unlock would simply
1608 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
1609 // In practice, however, the # of lock sites is bounded and is usually small.
1610 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
1611 // if the processor uses simple bimodal branch predictors keyed by EIP
1612 // Since the helper routines would be called from multiple synchronization
1613 // sites.
1614 //
1615 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
1616 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
1617 // to those specialized methods. That'd give us a mostly platform-independent
1618 // implementation that the JITs could optimize and inline at their pleasure.
1619 // Done correctly, the only time we'd need to cross to native could would be
1620 // to park() or unpark() threads. We'd also need a few more unsafe operators
1621 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
1622 // (b) explicit barriers or fence operations.
1623 //
1624 // TODO:
1625 //
1626 // * Arrange for C2 to pass "Self" into fast_lock and fast_unlock in one of the registers (scr).
1627 // This avoids manifesting the Self pointer in the fast_lock and fast_unlock terminals.
1628 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
1629 // the lock operators would typically be faster than reifying Self.
1630 //
1631 // * Ideally I'd define the primitives as:
1632 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
1633 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
1634 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
1635 // Instead, we're stuck with a rather awkward and brittle register assignments below.
1636 // Furthermore the register assignments are overconstrained, possibly resulting in
1637 // sub-optimal code near the synchronization site.
1638 //
1639 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
1640 // Alternately, use a better sp-proximity test.
1641 //
1642 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
1643 // Either one is sufficient to uniquely identify a thread.
1644 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
1645 //
1646 // * Intrinsify notify() and notifyAll() for the common cases where the
1647 // object is locked by the calling thread but the waitlist is empty.
1648 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
1649 //
1650 // * use jccb and jmpb instead of jcc and jmp to improve code density.
1651 // But beware of excessive branch density on AMD Opterons.
1652 //
1653 // * Both fast_lock and fast_unlock set the ICC.ZF to indicate success
1654 // or failure of the fast path. If the fast path fails then we pass
1655 // control to the slow path, typically in C. In fast_lock and
1656 // fast_unlock we often branch to DONE_LABEL, just to find that C2
1657 // will emit a conditional branch immediately after the node.
1658 // So we have branches to branches and lots of ICC.ZF games.
1659 // Instead, it might be better to have C2 pass a "FailureLabel"
1660 // into fast_lock and fast_unlock. In the case of success, control
1661 // will drop through the node. ICC.ZF is undefined at exit.
1662 // In the case of failure, the node will branch directly to the
1663 // FailureLabel
1664
1665
1666 // obj: object to lock
1667 // box: on-stack box address (displaced header location) - KILLED
1668 // rax,: tmp -- KILLED
1669 // scr: tmp -- KILLED
1670 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg,
1671 Register scrReg, Register cx1Reg, Register cx2Reg,
1672 BiasedLockingCounters* counters,
1673 RTMLockingCounters* rtm_counters,
1674 RTMLockingCounters* stack_rtm_counters,
1675 Metadata* method_data,
1676 bool use_rtm, bool profile_rtm) {
1677 // Ensure the register assignments are disjoint
1678 assert(tmpReg == rax, "");
1679
1680 if (use_rtm) {
1681 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg);
1682 } else {
1683 assert(cx1Reg == noreg, "");
1684 assert(cx2Reg == noreg, "");
1685 assert_different_registers(objReg, boxReg, tmpReg, scrReg);
1686 }
1687
1688 if (counters != NULL) {
1689 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg);
1690 }
1691
1692 // Possible cases that we'll encounter in fast_lock
1693 // ------------------------------------------------
1694 // * Inflated
1695 // -- unlocked
1696 // -- Locked
1697 // = by self
1698 // = by other
1699 // * biased
1700 // -- by Self
1701 // -- by other
1702 // * neutral
1703 // * stack-locked
1704 // -- by self
1705 // = sp-proximity test hits
1706 // = sp-proximity test generates false-negative
1707 // -- by other
1708 //
1709
1710 Label IsInflated, DONE_LABEL;
1711
1712 // it's stack-locked, biased or neutral
1713 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
1714 // order to reduce the number of conditional branches in the most common cases.
1715 // Beware -- there's a subtle invariant that fetch of the markword
1716 // at [FETCH], below, will never observe a biased encoding (*101b).
1717 // If this invariant is not held we risk exclusion (safety) failure.
1718 if (UseBiasedLocking && !UseOptoBiasInlining) {
1719 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters);
1720 }
1721
1722 #if INCLUDE_RTM_OPT
1723 if (UseRTMForStackLocks && use_rtm) {
1724 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg,
1725 stack_rtm_counters, method_data, profile_rtm,
1726 DONE_LABEL, IsInflated);
1727 }
1728 #endif // INCLUDE_RTM_OPT
1729
1730 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // [FETCH]
1731 testptr(tmpReg, markWord::monitor_value); // inflated vs stack-locked|neutral|biased
1732 jccb(Assembler::notZero, IsInflated);
1733
1734 // Attempt stack-locking ...
1735 orptr (tmpReg, markWord::unlocked_value);
1736 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
1737 lock();
1738 cmpxchgptr(boxReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Updates tmpReg
1739 if (counters != NULL) {
1740 cond_inc32(Assembler::equal,
1741 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1742 }
1743 jcc(Assembler::equal, DONE_LABEL); // Success
1744
1745 // Recursive locking.
1746 // The object is stack-locked: markword contains stack pointer to BasicLock.
1747 // Locked by current thread if difference with current SP is less than one page.
1748 subptr(tmpReg, rsp);
1749 // Next instruction set ZFlag == 1 (Success) if difference is less then one page.
1750 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) );
1751 movptr(Address(boxReg, 0), tmpReg);
1752 if (counters != NULL) {
1753 cond_inc32(Assembler::equal,
1754 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1755 }
1756 jmp(DONE_LABEL);
1757
1758 bind(IsInflated);
1759 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markWord::monitor_value
1760
1761 #if INCLUDE_RTM_OPT
1762 // Use the same RTM locking code in 32- and 64-bit VM.
1763 if (use_rtm) {
1764 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg,
1765 rtm_counters, method_data, profile_rtm, DONE_LABEL);
1766 } else {
1767 #endif // INCLUDE_RTM_OPT
1768
1769 #ifndef _LP64
1770 // The object is inflated.
1771
1772 // boxReg refers to the on-stack BasicLock in the current frame.
1773 // We'd like to write:
1774 // set box->_displaced_header = markWord::unused_mark(). Any non-0 value suffices.
1775 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
1776 // additional latency as we have another ST in the store buffer that must drain.
1777
1778 // avoid ST-before-CAS
1779 // register juggle because we need tmpReg for cmpxchgptr below
1780 movptr(scrReg, boxReg);
1781 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1782
1783 // Optimistic form: consider XORL tmpReg,tmpReg
1784 movptr(tmpReg, NULL_WORD);
1785
1786 // Appears unlocked - try to swing _owner from null to non-null.
1787 // Ideally, I'd manifest "Self" with get_thread and then attempt
1788 // to CAS the register containing Self into m->Owner.
1789 // But we don't have enough registers, so instead we can either try to CAS
1790 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
1791 // we later store "Self" into m->Owner. Transiently storing a stack address
1792 // (rsp or the address of the box) into m->owner is harmless.
1793 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1794 lock();
1795 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1796 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3
1797 // If we weren't able to swing _owner from NULL to the BasicLock
1798 // then take the slow path.
1799 jccb (Assembler::notZero, DONE_LABEL);
1800 // update _owner from BasicLock to thread
1801 get_thread (scrReg); // beware: clobbers ICCs
1802 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg);
1803 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success
1804
1805 // If the CAS fails we can either retry or pass control to the slow path.
1806 // We use the latter tactic.
1807 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1808 // If the CAS was successful ...
1809 // Self has acquired the lock
1810 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1811 // Intentional fall-through into DONE_LABEL ...
1812 #else // _LP64
1813 // It's inflated and we use scrReg for ObjectMonitor* in this section.
1814 movq(scrReg, tmpReg);
1815 xorq(tmpReg, tmpReg);
1816 lock();
1817 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1818 // Unconditionally set box->_displaced_header = markWord::unused_mark().
1819 // Without cast to int32_t this style of movptr will destroy r10 which is typically obj.
1820 movptr(Address(boxReg, 0), (int32_t)intptr_t(markWord::unused_mark().value()));
1821 // Intentional fall-through into DONE_LABEL ...
1822 // Propagate ICC.ZF from CAS above into DONE_LABEL.
1823 #endif // _LP64
1824 #if INCLUDE_RTM_OPT
1825 } // use_rtm()
1826 #endif
1827 // DONE_LABEL is a hot target - we'd really like to place it at the
1828 // start of cache line by padding with NOPs.
1829 // See the AMD and Intel software optimization manuals for the
1830 // most efficient "long" NOP encodings.
1831 // Unfortunately none of our alignment mechanisms suffice.
1832 bind(DONE_LABEL);
1833
1834 // At DONE_LABEL the icc ZFlag is set as follows ...
1835 // fast_unlock uses the same protocol.
1836 // ZFlag == 1 -> Success
1837 // ZFlag == 0 -> Failure - force control through the slow path
1838 }
1839
1840 // obj: object to unlock
1841 // box: box address (displaced header location), killed. Must be EAX.
1842 // tmp: killed, cannot be obj nor box.
1843 //
1844 // Some commentary on balanced locking:
1845 //
1846 // fast_lock and fast_unlock are emitted only for provably balanced lock sites.
1847 // Methods that don't have provably balanced locking are forced to run in the
1848 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
1849 // The interpreter provides two properties:
1850 // I1: At return-time the interpreter automatically and quietly unlocks any
1851 // objects acquired the current activation (frame). Recall that the
1852 // interpreter maintains an on-stack list of locks currently held by
1853 // a frame.
1854 // I2: If a method attempts to unlock an object that is not held by the
1855 // the frame the interpreter throws IMSX.
1856 //
1857 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
1858 // B() doesn't have provably balanced locking so it runs in the interpreter.
1859 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
1860 // is still locked by A().
1861 //
1862 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
1863 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
1864 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
1865 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
1866 // Arguably given that the spec legislates the JNI case as undefined our implementation
1867 // could reasonably *avoid* checking owner in fast_unlock().
1868 // In the interest of performance we elide m->Owner==Self check in unlock.
1869 // A perfectly viable alternative is to elide the owner check except when
1870 // Xcheck:jni is enabled.
1871
1872 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) {
1873 assert(boxReg == rax, "");
1874 assert_different_registers(objReg, boxReg, tmpReg);
1875
1876 Label DONE_LABEL, Stacked, CheckSucc;
1877
1878 // Critically, the biased locking test must have precedence over
1879 // and appear before the (box->dhw == 0) recursive stack-lock test.
1880 if (UseBiasedLocking && !UseOptoBiasInlining) {
1881 biased_locking_exit(objReg, tmpReg, DONE_LABEL);
1882 }
1883
1884 #if INCLUDE_RTM_OPT
1885 if (UseRTMForStackLocks && use_rtm) {
1886 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1887 Label L_regular_unlock;
1888 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword
1889 andptr(tmpReg, markWord::biased_lock_mask_in_place); // look at 3 lock bits
1890 cmpptr(tmpReg, markWord::unlocked_value); // bits = 001 unlocked
1891 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock
1892 xend(); // otherwise end...
1893 jmp(DONE_LABEL); // ... and we're done
1894 bind(L_regular_unlock);
1895 }
1896 #endif
1897
1898 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header
1899 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock
1900 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Examine the object's markword
1901 testptr(tmpReg, markWord::monitor_value); // Inflated?
1902 jccb (Assembler::zero, Stacked);
1903
1904 // It's inflated.
1905 #if INCLUDE_RTM_OPT
1906 if (use_rtm) {
1907 Label L_regular_inflated_unlock;
1908 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1909 movptr(boxReg, Address(tmpReg, owner_offset));
1910 testptr(boxReg, boxReg);
1911 jccb(Assembler::notZero, L_regular_inflated_unlock);
1912 xend();
1913 jmpb(DONE_LABEL);
1914 bind(L_regular_inflated_unlock);
1915 }
1916 #endif
1917
1918 // Despite our balanced locking property we still check that m->_owner == Self
1919 // as java routines or native JNI code called by this thread might
1920 // have released the lock.
1921 // Refer to the comments in synchronizer.cpp for how we might encode extra
1922 // state in _succ so we can avoid fetching EntryList|cxq.
1923 //
1924 // I'd like to add more cases in fast_lock() and fast_unlock() --
1925 // such as recursive enter and exit -- but we have to be wary of
1926 // I$ bloat, T$ effects and BP$ effects.
1927 //
1928 // If there's no contention try a 1-0 exit. That is, exit without
1929 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
1930 // we detect and recover from the race that the 1-0 exit admits.
1931 //
1932 // Conceptually fast_unlock() must execute a STST|LDST "release" barrier
1933 // before it STs null into _owner, releasing the lock. Updates
1934 // to data protected by the critical section must be visible before
1935 // we drop the lock (and thus before any other thread could acquire
1936 // the lock and observe the fields protected by the lock).
1937 // IA32's memory-model is SPO, so STs are ordered with respect to
1938 // each other and there's no need for an explicit barrier (fence).
1939 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
1940 #ifndef _LP64
1941 get_thread (boxReg);
1942
1943 // Note that we could employ various encoding schemes to reduce
1944 // the number of loads below (currently 4) to just 2 or 3.
1945 // Refer to the comments in synchronizer.cpp.
1946 // In practice the chain of fetches doesn't seem to impact performance, however.
1947 xorptr(boxReg, boxReg);
1948 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
1949 jccb (Assembler::notZero, DONE_LABEL);
1950 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
1951 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
1952 jccb (Assembler::notZero, CheckSucc);
1953 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
1954 jmpb (DONE_LABEL);
1955
1956 bind (Stacked);
1957 // It's not inflated and it's not recursively stack-locked and it's not biased.
1958 // It must be stack-locked.
1959 // Try to reset the header to displaced header.
1960 // The "box" value on the stack is stable, so we can reload
1961 // and be assured we observe the same value as above.
1962 movptr(tmpReg, Address(boxReg, 0));
1963 lock();
1964 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box
1965 // Intention fall-thru into DONE_LABEL
1966
1967 // DONE_LABEL is a hot target - we'd really like to place it at the
1968 // start of cache line by padding with NOPs.
1969 // See the AMD and Intel software optimization manuals for the
1970 // most efficient "long" NOP encodings.
1971 // Unfortunately none of our alignment mechanisms suffice.
1972 bind (CheckSucc);
1973 #else // _LP64
1974 // It's inflated
1975 xorptr(boxReg, boxReg);
1976 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
1977 jccb (Assembler::notZero, DONE_LABEL);
1978 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
1979 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
1980 jccb (Assembler::notZero, CheckSucc);
1981 // Without cast to int32_t this style of movptr will destroy r10 which is typically obj.
1982 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
1983 jmpb (DONE_LABEL);
1984
1985 // Try to avoid passing control into the slow_path ...
1986 Label LSuccess, LGoSlowPath ;
1987 bind (CheckSucc);
1988
1989 // The following optional optimization can be elided if necessary
1990 // Effectively: if (succ == null) goto slow path
1991 // The code reduces the window for a race, however,
1992 // and thus benefits performance.
1993 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
1994 jccb (Assembler::zero, LGoSlowPath);
1995
1996 xorptr(boxReg, boxReg);
1997 // Without cast to int32_t this style of movptr will destroy r10 which is typically obj.
1998 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
1999
2000 // Memory barrier/fence
2001 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ
2002 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack.
2003 // This is faster on Nehalem and AMD Shanghai/Barcelona.
2004 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2005 // We might also restructure (ST Owner=0;barrier;LD _Succ) to
2006 // (mov box,0; xchgq box, &m->Owner; LD _succ) .
2007 lock(); addl(Address(rsp, 0), 0);
2008
2009 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2010 jccb (Assembler::notZero, LSuccess);
2011
2012 // Rare inopportune interleaving - race.
2013 // The successor vanished in the small window above.
2014 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor.
2015 // We need to ensure progress and succession.
2016 // Try to reacquire the lock.
2017 // If that fails then the new owner is responsible for succession and this
2018 // thread needs to take no further action and can exit via the fast path (success).
2019 // If the re-acquire succeeds then pass control into the slow path.
2020 // As implemented, this latter mode is horrible because we generated more
2021 // coherence traffic on the lock *and* artifically extended the critical section
2022 // length while by virtue of passing control into the slow path.
2023
2024 // box is really RAX -- the following CMPXCHG depends on that binding
2025 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R)
2026 lock();
2027 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2028 // There's no successor so we tried to regrab the lock.
2029 // If that didn't work, then another thread grabbed the
2030 // lock so we're done (and exit was a success).
2031 jccb (Assembler::notEqual, LSuccess);
2032 // Intentional fall-through into slow path
2033
2034 bind (LGoSlowPath);
2035 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure
2036 jmpb (DONE_LABEL);
2037
2038 bind (LSuccess);
2039 testl (boxReg, 0); // set ICC.ZF=1 to indicate success
2040 jmpb (DONE_LABEL);
2041
2042 bind (Stacked);
2043 movptr(tmpReg, Address (boxReg, 0)); // re-fetch
2044 lock();
2045 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box
2046
2047 #endif
2048 bind(DONE_LABEL);
2049 }
2050 #endif // COMPILER2
2051
2052 void MacroAssembler::c2bool(Register x) {
2053 // implements x == 0 ? 0 : 1
2054 // note: must only look at least-significant byte of x
2055 // since C-style booleans are stored in one byte
2056 // only! (was bug)
2057 andl(x, 0xFF);
2058 setb(Assembler::notZero, x);
2059 }
2060
2061 // Wouldn't need if AddressLiteral version had new name
2062 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) {
2063 Assembler::call(L, rtype);
2064 }
2065
2066 void MacroAssembler::call(Register entry) {
2067 Assembler::call(entry);
2068 }
2069
2070 void MacroAssembler::call(AddressLiteral entry) {
2071 if (reachable(entry)) {
2072 Assembler::call_literal(entry.target(), entry.rspec());
2073 } else {
2074 lea(rscratch1, entry);
2075 Assembler::call(rscratch1);
2076 }
2077 }
2078
2079 void MacroAssembler::ic_call(address entry, jint method_index) {
2080 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index);
2081 movptr(rax, (intptr_t)Universe::non_oop_word());
2082 call(AddressLiteral(entry, rh));
2083 }
2084
2085 // Implementation of call_VM versions
2086
2087 void MacroAssembler::call_VM(Register oop_result,
2088 address entry_point,
2089 bool check_exceptions) {
2090 Label C, E;
2091 call(C, relocInfo::none);
2092 jmp(E);
2093
2094 bind(C);
2095 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
2096 ret(0);
2097
2098 bind(E);
2099 }
2100
2101 void MacroAssembler::call_VM(Register oop_result,
2102 address entry_point,
2103 Register arg_1,
2104 bool check_exceptions) {
2105 Label C, E;
2106 call(C, relocInfo::none);
2107 jmp(E);
2108
2109 bind(C);
2110 pass_arg1(this, arg_1);
2111 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
2112 ret(0);
2113
2114 bind(E);
2115 }
2116
2117 void MacroAssembler::call_VM(Register oop_result,
2118 address entry_point,
2119 Register arg_1,
2120 Register arg_2,
2121 bool check_exceptions) {
2122 Label C, E;
2123 call(C, relocInfo::none);
2124 jmp(E);
2125
2126 bind(C);
2127
2128 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2129
2130 pass_arg2(this, arg_2);
2131 pass_arg1(this, arg_1);
2132 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
2133 ret(0);
2134
2135 bind(E);
2136 }
2137
2138 void MacroAssembler::call_VM(Register oop_result,
2139 address entry_point,
2140 Register arg_1,
2141 Register arg_2,
2142 Register arg_3,
2143 bool check_exceptions) {
2144 Label C, E;
2145 call(C, relocInfo::none);
2146 jmp(E);
2147
2148 bind(C);
2149
2150 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2151 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2152 pass_arg3(this, arg_3);
2153
2154 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2155 pass_arg2(this, arg_2);
2156
2157 pass_arg1(this, arg_1);
2158 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
2159 ret(0);
2160
2161 bind(E);
2162 }
2163
2164 void MacroAssembler::call_VM(Register oop_result,
2165 Register last_java_sp,
2166 address entry_point,
2167 int number_of_arguments,
2168 bool check_exceptions) {
2169 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2170 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2171 }
2172
2173 void MacroAssembler::call_VM(Register oop_result,
2174 Register last_java_sp,
2175 address entry_point,
2176 Register arg_1,
2177 bool check_exceptions) {
2178 pass_arg1(this, arg_1);
2179 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2180 }
2181
2182 void MacroAssembler::call_VM(Register oop_result,
2183 Register last_java_sp,
2184 address entry_point,
2185 Register arg_1,
2186 Register arg_2,
2187 bool check_exceptions) {
2188
2189 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2190 pass_arg2(this, arg_2);
2191 pass_arg1(this, arg_1);
2192 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2193 }
2194
2195 void MacroAssembler::call_VM(Register oop_result,
2196 Register last_java_sp,
2197 address entry_point,
2198 Register arg_1,
2199 Register arg_2,
2200 Register arg_3,
2201 bool check_exceptions) {
2202 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2203 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2204 pass_arg3(this, arg_3);
2205 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2206 pass_arg2(this, arg_2);
2207 pass_arg1(this, arg_1);
2208 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2209 }
2210
2211 void MacroAssembler::super_call_VM(Register oop_result,
2212 Register last_java_sp,
2213 address entry_point,
2214 int number_of_arguments,
2215 bool check_exceptions) {
2216 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2217 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2218 }
2219
2220 void MacroAssembler::super_call_VM(Register oop_result,
2221 Register last_java_sp,
2222 address entry_point,
2223 Register arg_1,
2224 bool check_exceptions) {
2225 pass_arg1(this, arg_1);
2226 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2227 }
2228
2229 void MacroAssembler::super_call_VM(Register oop_result,
2230 Register last_java_sp,
2231 address entry_point,
2232 Register arg_1,
2233 Register arg_2,
2234 bool check_exceptions) {
2235
2236 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2237 pass_arg2(this, arg_2);
2238 pass_arg1(this, arg_1);
2239 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2240 }
2241
2242 void MacroAssembler::super_call_VM(Register oop_result,
2243 Register last_java_sp,
2244 address entry_point,
2245 Register arg_1,
2246 Register arg_2,
2247 Register arg_3,
2248 bool check_exceptions) {
2249 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2250 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2251 pass_arg3(this, arg_3);
2252 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2253 pass_arg2(this, arg_2);
2254 pass_arg1(this, arg_1);
2255 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2256 }
2257
2258 void MacroAssembler::call_VM_base(Register oop_result,
2259 Register java_thread,
2260 Register last_java_sp,
2261 address entry_point,
2262 int number_of_arguments,
2263 bool check_exceptions) {
2264 // determine java_thread register
2265 if (!java_thread->is_valid()) {
2266 #ifdef _LP64
2267 java_thread = r15_thread;
2268 #else
2269 java_thread = rdi;
2270 get_thread(java_thread);
2271 #endif // LP64
2272 }
2273 // determine last_java_sp register
2274 if (!last_java_sp->is_valid()) {
2275 last_java_sp = rsp;
2276 }
2277 // debugging support
2278 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
2279 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register"));
2280 #ifdef ASSERT
2281 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
2282 // r12 is the heapbase.
2283 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");)
2284 #endif // ASSERT
2285
2286 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
2287 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
2288
2289 // push java thread (becomes first argument of C function)
2290
2291 NOT_LP64(push(java_thread); number_of_arguments++);
2292 LP64_ONLY(mov(c_rarg0, r15_thread));
2293
2294 // set last Java frame before call
2295 assert(last_java_sp != rbp, "can't use ebp/rbp");
2296
2297 // Only interpreter should have to set fp
2298 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL);
2299
2300 // do the call, remove parameters
2301 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
2302
2303 // restore the thread (cannot use the pushed argument since arguments
2304 // may be overwritten by C code generated by an optimizing compiler);
2305 // however can use the register value directly if it is callee saved.
2306 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) {
2307 // rdi & rsi (also r15) are callee saved -> nothing to do
2308 #ifdef ASSERT
2309 guarantee(java_thread != rax, "change this code");
2310 push(rax);
2311 { Label L;
2312 get_thread(rax);
2313 cmpptr(java_thread, rax);
2314 jcc(Assembler::equal, L);
2315 STOP("MacroAssembler::call_VM_base: rdi not callee saved?");
2316 bind(L);
2317 }
2318 pop(rax);
2319 #endif
2320 } else {
2321 get_thread(java_thread);
2322 }
2323 // reset last Java frame
2324 // Only interpreter should have to clear fp
2325 reset_last_Java_frame(java_thread, true);
2326
2327 // C++ interp handles this in the interpreter
2328 check_and_handle_popframe(java_thread);
2329 check_and_handle_earlyret(java_thread);
2330
2331 if (check_exceptions) {
2332 // check for pending exceptions (java_thread is set upon return)
2333 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD);
2334 #ifndef _LP64
2335 jump_cc(Assembler::notEqual,
2336 RuntimeAddress(StubRoutines::forward_exception_entry()));
2337 #else
2338 // This used to conditionally jump to forward_exception however it is
2339 // possible if we relocate that the branch will not reach. So we must jump
2340 // around so we can always reach
2341
2342 Label ok;
2343 jcc(Assembler::equal, ok);
2344 jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2345 bind(ok);
2346 #endif // LP64
2347 }
2348
2349 // get oop result if there is one and reset the value in the thread
2350 if (oop_result->is_valid()) {
2351 get_vm_result(oop_result, java_thread);
2352 }
2353 }
2354
2355 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
2356
2357 // Calculate the value for last_Java_sp
2358 // somewhat subtle. call_VM does an intermediate call
2359 // which places a return address on the stack just under the
2360 // stack pointer as the user finsihed with it. This allows
2361 // use to retrieve last_Java_pc from last_Java_sp[-1].
2362 // On 32bit we then have to push additional args on the stack to accomplish
2363 // the actual requested call. On 64bit call_VM only can use register args
2364 // so the only extra space is the return address that call_VM created.
2365 // This hopefully explains the calculations here.
2366
2367 #ifdef _LP64
2368 // We've pushed one address, correct last_Java_sp
2369 lea(rax, Address(rsp, wordSize));
2370 #else
2371 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize));
2372 #endif // LP64
2373
2374 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions);
2375
2376 }
2377
2378 // Use this method when MacroAssembler version of call_VM_leaf_base() should be called from Interpreter.
2379 void MacroAssembler::call_VM_leaf0(address entry_point) {
2380 MacroAssembler::call_VM_leaf_base(entry_point, 0);
2381 }
2382
2383 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
2384 call_VM_leaf_base(entry_point, number_of_arguments);
2385 }
2386
2387 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
2388 pass_arg0(this, arg_0);
2389 call_VM_leaf(entry_point, 1);
2390 }
2391
2392 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2393
2394 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2395 pass_arg1(this, arg_1);
2396 pass_arg0(this, arg_0);
2397 call_VM_leaf(entry_point, 2);
2398 }
2399
2400 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2401 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2402 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2403 pass_arg2(this, arg_2);
2404 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2405 pass_arg1(this, arg_1);
2406 pass_arg0(this, arg_0);
2407 call_VM_leaf(entry_point, 3);
2408 }
2409
2410 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2411 pass_arg0(this, arg_0);
2412 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2413 }
2414
2415 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2416
2417 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2418 pass_arg1(this, arg_1);
2419 pass_arg0(this, arg_0);
2420 MacroAssembler::call_VM_leaf_base(entry_point, 2);
2421 }
2422
2423 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2424 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2425 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2426 pass_arg2(this, arg_2);
2427 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2428 pass_arg1(this, arg_1);
2429 pass_arg0(this, arg_0);
2430 MacroAssembler::call_VM_leaf_base(entry_point, 3);
2431 }
2432
2433 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2434 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg"));
2435 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2436 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2437 pass_arg3(this, arg_3);
2438 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2439 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2440 pass_arg2(this, arg_2);
2441 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2442 pass_arg1(this, arg_1);
2443 pass_arg0(this, arg_0);
2444 MacroAssembler::call_VM_leaf_base(entry_point, 4);
2445 }
2446
2447 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
2448 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
2449 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD);
2450 verify_oop(oop_result);
2451 }
2452
2453 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
2454 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
2455 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD);
2456 }
2457
2458 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2459 }
2460
2461 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2462 }
2463
2464 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) {
2465 if (reachable(src1)) {
2466 cmpl(as_Address(src1), imm);
2467 } else {
2468 lea(rscratch1, src1);
2469 cmpl(Address(rscratch1, 0), imm);
2470 }
2471 }
2472
2473 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) {
2474 assert(!src2.is_lval(), "use cmpptr");
2475 if (reachable(src2)) {
2476 cmpl(src1, as_Address(src2));
2477 } else {
2478 lea(rscratch1, src2);
2479 cmpl(src1, Address(rscratch1, 0));
2480 }
2481 }
2482
2483 void MacroAssembler::cmp32(Register src1, int32_t imm) {
2484 Assembler::cmpl(src1, imm);
2485 }
2486
2487 void MacroAssembler::cmp32(Register src1, Address src2) {
2488 Assembler::cmpl(src1, src2);
2489 }
2490
2491 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2492 ucomisd(opr1, opr2);
2493
2494 Label L;
2495 if (unordered_is_less) {
2496 movl(dst, -1);
2497 jcc(Assembler::parity, L);
2498 jcc(Assembler::below , L);
2499 movl(dst, 0);
2500 jcc(Assembler::equal , L);
2501 increment(dst);
2502 } else { // unordered is greater
2503 movl(dst, 1);
2504 jcc(Assembler::parity, L);
2505 jcc(Assembler::above , L);
2506 movl(dst, 0);
2507 jcc(Assembler::equal , L);
2508 decrementl(dst);
2509 }
2510 bind(L);
2511 }
2512
2513 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2514 ucomiss(opr1, opr2);
2515
2516 Label L;
2517 if (unordered_is_less) {
2518 movl(dst, -1);
2519 jcc(Assembler::parity, L);
2520 jcc(Assembler::below , L);
2521 movl(dst, 0);
2522 jcc(Assembler::equal , L);
2523 increment(dst);
2524 } else { // unordered is greater
2525 movl(dst, 1);
2526 jcc(Assembler::parity, L);
2527 jcc(Assembler::above , L);
2528 movl(dst, 0);
2529 jcc(Assembler::equal , L);
2530 decrementl(dst);
2531 }
2532 bind(L);
2533 }
2534
2535
2536 void MacroAssembler::cmp8(AddressLiteral src1, int imm) {
2537 if (reachable(src1)) {
2538 cmpb(as_Address(src1), imm);
2539 } else {
2540 lea(rscratch1, src1);
2541 cmpb(Address(rscratch1, 0), imm);
2542 }
2543 }
2544
2545 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) {
2546 #ifdef _LP64
2547 if (src2.is_lval()) {
2548 movptr(rscratch1, src2);
2549 Assembler::cmpq(src1, rscratch1);
2550 } else if (reachable(src2)) {
2551 cmpq(src1, as_Address(src2));
2552 } else {
2553 lea(rscratch1, src2);
2554 Assembler::cmpq(src1, Address(rscratch1, 0));
2555 }
2556 #else
2557 if (src2.is_lval()) {
2558 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2559 } else {
2560 cmpl(src1, as_Address(src2));
2561 }
2562 #endif // _LP64
2563 }
2564
2565 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) {
2566 assert(src2.is_lval(), "not a mem-mem compare");
2567 #ifdef _LP64
2568 // moves src2's literal address
2569 movptr(rscratch1, src2);
2570 Assembler::cmpq(src1, rscratch1);
2571 #else
2572 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2573 #endif // _LP64
2574 }
2575
2576 void MacroAssembler::cmpoop(Register src1, Register src2) {
2577 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2578 bs->obj_equals(this, src1, src2);
2579 }
2580
2581 void MacroAssembler::cmpoop(Register src1, Address src2) {
2582 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2583 bs->obj_equals(this, src1, src2);
2584 }
2585
2586 #ifdef _LP64
2587 void MacroAssembler::cmpoop(Register src1, jobject src2) {
2588 movoop(rscratch1, src2);
2589 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2590 bs->obj_equals(this, src1, rscratch1);
2591 }
2592 #endif
2593
2594 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) {
2595 if (reachable(adr)) {
2596 lock();
2597 cmpxchgptr(reg, as_Address(adr));
2598 } else {
2599 lea(rscratch1, adr);
2600 lock();
2601 cmpxchgptr(reg, Address(rscratch1, 0));
2602 }
2603 }
2604
2605 void MacroAssembler::cmpxchgptr(Register reg, Address adr) {
2606 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr));
2607 }
2608
2609 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) {
2610 if (reachable(src)) {
2611 Assembler::comisd(dst, as_Address(src));
2612 } else {
2613 lea(rscratch1, src);
2614 Assembler::comisd(dst, Address(rscratch1, 0));
2615 }
2616 }
2617
2618 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) {
2619 if (reachable(src)) {
2620 Assembler::comiss(dst, as_Address(src));
2621 } else {
2622 lea(rscratch1, src);
2623 Assembler::comiss(dst, Address(rscratch1, 0));
2624 }
2625 }
2626
2627
2628 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) {
2629 Condition negated_cond = negate_condition(cond);
2630 Label L;
2631 jcc(negated_cond, L);
2632 pushf(); // Preserve flags
2633 atomic_incl(counter_addr);
2634 popf();
2635 bind(L);
2636 }
2637
2638 int MacroAssembler::corrected_idivl(Register reg) {
2639 // Full implementation of Java idiv and irem; checks for
2640 // special case as described in JVM spec., p.243 & p.271.
2641 // The function returns the (pc) offset of the idivl
2642 // instruction - may be needed for implicit exceptions.
2643 //
2644 // normal case special case
2645 //
2646 // input : rax,: dividend min_int
2647 // reg: divisor (may not be rax,/rdx) -1
2648 //
2649 // output: rax,: quotient (= rax, idiv reg) min_int
2650 // rdx: remainder (= rax, irem reg) 0
2651 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register");
2652 const int min_int = 0x80000000;
2653 Label normal_case, special_case;
2654
2655 // check for special case
2656 cmpl(rax, min_int);
2657 jcc(Assembler::notEqual, normal_case);
2658 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0)
2659 cmpl(reg, -1);
2660 jcc(Assembler::equal, special_case);
2661
2662 // handle normal case
2663 bind(normal_case);
2664 cdql();
2665 int idivl_offset = offset();
2666 idivl(reg);
2667
2668 // normal and special case exit
2669 bind(special_case);
2670
2671 return idivl_offset;
2672 }
2673
2674
2675
2676 void MacroAssembler::decrementl(Register reg, int value) {
2677 if (value == min_jint) {subl(reg, value) ; return; }
2678 if (value < 0) { incrementl(reg, -value); return; }
2679 if (value == 0) { ; return; }
2680 if (value == 1 && UseIncDec) { decl(reg) ; return; }
2681 /* else */ { subl(reg, value) ; return; }
2682 }
2683
2684 void MacroAssembler::decrementl(Address dst, int value) {
2685 if (value == min_jint) {subl(dst, value) ; return; }
2686 if (value < 0) { incrementl(dst, -value); return; }
2687 if (value == 0) { ; return; }
2688 if (value == 1 && UseIncDec) { decl(dst) ; return; }
2689 /* else */ { subl(dst, value) ; return; }
2690 }
2691
2692 void MacroAssembler::division_with_shift (Register reg, int shift_value) {
2693 assert (shift_value > 0, "illegal shift value");
2694 Label _is_positive;
2695 testl (reg, reg);
2696 jcc (Assembler::positive, _is_positive);
2697 int offset = (1 << shift_value) - 1 ;
2698
2699 if (offset == 1) {
2700 incrementl(reg);
2701 } else {
2702 addl(reg, offset);
2703 }
2704
2705 bind (_is_positive);
2706 sarl(reg, shift_value);
2707 }
2708
2709 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) {
2710 if (reachable(src)) {
2711 Assembler::divsd(dst, as_Address(src));
2712 } else {
2713 lea(rscratch1, src);
2714 Assembler::divsd(dst, Address(rscratch1, 0));
2715 }
2716 }
2717
2718 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) {
2719 if (reachable(src)) {
2720 Assembler::divss(dst, as_Address(src));
2721 } else {
2722 lea(rscratch1, src);
2723 Assembler::divss(dst, Address(rscratch1, 0));
2724 }
2725 }
2726
2727 #ifndef _LP64
2728 void MacroAssembler::empty_FPU_stack() {
2729 if (VM_Version::supports_mmx()) {
2730 emms();
2731 } else {
2732 for (int i = 8; i-- > 0; ) ffree(i);
2733 }
2734 }
2735 #endif // !LP64
2736
2737
2738 void MacroAssembler::enter() {
2739 push(rbp);
2740 mov(rbp, rsp);
2741 }
2742
2743 // A 5 byte nop that is safe for patching (see patch_verified_entry)
2744 void MacroAssembler::fat_nop() {
2745 if (UseAddressNop) {
2746 addr_nop_5();
2747 } else {
2748 emit_int8(0x26); // es:
2749 emit_int8(0x2e); // cs:
2750 emit_int8(0x64); // fs:
2751 emit_int8(0x65); // gs:
2752 emit_int8((unsigned char)0x90);
2753 }
2754 }
2755
2756 #if !defined(_LP64)
2757 void MacroAssembler::fcmp(Register tmp) {
2758 fcmp(tmp, 1, true, true);
2759 }
2760
2761 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) {
2762 assert(!pop_right || pop_left, "usage error");
2763 if (VM_Version::supports_cmov()) {
2764 assert(tmp == noreg, "unneeded temp");
2765 if (pop_left) {
2766 fucomip(index);
2767 } else {
2768 fucomi(index);
2769 }
2770 if (pop_right) {
2771 fpop();
2772 }
2773 } else {
2774 assert(tmp != noreg, "need temp");
2775 if (pop_left) {
2776 if (pop_right) {
2777 fcompp();
2778 } else {
2779 fcomp(index);
2780 }
2781 } else {
2782 fcom(index);
2783 }
2784 // convert FPU condition into eflags condition via rax,
2785 save_rax(tmp);
2786 fwait(); fnstsw_ax();
2787 sahf();
2788 restore_rax(tmp);
2789 }
2790 // condition codes set as follows:
2791 //
2792 // CF (corresponds to C0) if x < y
2793 // PF (corresponds to C2) if unordered
2794 // ZF (corresponds to C3) if x = y
2795 }
2796
2797 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) {
2798 fcmp2int(dst, unordered_is_less, 1, true, true);
2799 }
2800
2801 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) {
2802 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right);
2803 Label L;
2804 if (unordered_is_less) {
2805 movl(dst, -1);
2806 jcc(Assembler::parity, L);
2807 jcc(Assembler::below , L);
2808 movl(dst, 0);
2809 jcc(Assembler::equal , L);
2810 increment(dst);
2811 } else { // unordered is greater
2812 movl(dst, 1);
2813 jcc(Assembler::parity, L);
2814 jcc(Assembler::above , L);
2815 movl(dst, 0);
2816 jcc(Assembler::equal , L);
2817 decrementl(dst);
2818 }
2819 bind(L);
2820 }
2821
2822 void MacroAssembler::fld_d(AddressLiteral src) {
2823 fld_d(as_Address(src));
2824 }
2825
2826 void MacroAssembler::fld_s(AddressLiteral src) {
2827 fld_s(as_Address(src));
2828 }
2829
2830 void MacroAssembler::fld_x(AddressLiteral src) {
2831 Assembler::fld_x(as_Address(src));
2832 }
2833
2834 void MacroAssembler::fldcw(AddressLiteral src) {
2835 Assembler::fldcw(as_Address(src));
2836 }
2837
2838 void MacroAssembler::fpop() {
2839 ffree();
2840 fincstp();
2841 }
2842
2843 void MacroAssembler::fremr(Register tmp) {
2844 save_rax(tmp);
2845 { Label L;
2846 bind(L);
2847 fprem();
2848 fwait(); fnstsw_ax();
2849 sahf();
2850 jcc(Assembler::parity, L);
2851 }
2852 restore_rax(tmp);
2853 // Result is in ST0.
2854 // Note: fxch & fpop to get rid of ST1
2855 // (otherwise FPU stack could overflow eventually)
2856 fxch(1);
2857 fpop();
2858 }
2859 #endif // !LP64
2860
2861 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) {
2862 if (reachable(src)) {
2863 Assembler::mulpd(dst, as_Address(src));
2864 } else {
2865 lea(rscratch1, src);
2866 Assembler::mulpd(dst, Address(rscratch1, 0));
2867 }
2868 }
2869
2870 void MacroAssembler::load_float(Address src) {
2871 if (UseSSE >= 1) {
2872 movflt(xmm0, src);
2873 } else {
2874 LP64_ONLY(ShouldNotReachHere());
2875 NOT_LP64(fld_s(src));
2876 }
2877 }
2878
2879 void MacroAssembler::store_float(Address dst) {
2880 if (UseSSE >= 1) {
2881 movflt(dst, xmm0);
2882 } else {
2883 LP64_ONLY(ShouldNotReachHere());
2884 NOT_LP64(fstp_s(dst));
2885 }
2886 }
2887
2888 void MacroAssembler::load_double(Address src) {
2889 if (UseSSE >= 2) {
2890 movdbl(xmm0, src);
2891 } else {
2892 LP64_ONLY(ShouldNotReachHere());
2893 NOT_LP64(fld_d(src));
2894 }
2895 }
2896
2897 void MacroAssembler::store_double(Address dst) {
2898 if (UseSSE >= 2) {
2899 movdbl(dst, xmm0);
2900 } else {
2901 LP64_ONLY(ShouldNotReachHere());
2902 NOT_LP64(fstp_d(dst));
2903 }
2904 }
2905
2906 // dst = c = a * b + c
2907 void MacroAssembler::fmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) {
2908 Assembler::vfmadd231sd(c, a, b);
2909 if (dst != c) {
2910 movdbl(dst, c);
2911 }
2912 }
2913
2914 // dst = c = a * b + c
2915 void MacroAssembler::fmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) {
2916 Assembler::vfmadd231ss(c, a, b);
2917 if (dst != c) {
2918 movflt(dst, c);
2919 }
2920 }
2921
2922 // dst = c = a * b + c
2923 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) {
2924 Assembler::vfmadd231pd(c, a, b, vector_len);
2925 if (dst != c) {
2926 vmovdqu(dst, c);
2927 }
2928 }
2929
2930 // dst = c = a * b + c
2931 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) {
2932 Assembler::vfmadd231ps(c, a, b, vector_len);
2933 if (dst != c) {
2934 vmovdqu(dst, c);
2935 }
2936 }
2937
2938 // dst = c = a * b + c
2939 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) {
2940 Assembler::vfmadd231pd(c, a, b, vector_len);
2941 if (dst != c) {
2942 vmovdqu(dst, c);
2943 }
2944 }
2945
2946 // dst = c = a * b + c
2947 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) {
2948 Assembler::vfmadd231ps(c, a, b, vector_len);
2949 if (dst != c) {
2950 vmovdqu(dst, c);
2951 }
2952 }
2953
2954 void MacroAssembler::incrementl(AddressLiteral dst) {
2955 if (reachable(dst)) {
2956 incrementl(as_Address(dst));
2957 } else {
2958 lea(rscratch1, dst);
2959 incrementl(Address(rscratch1, 0));
2960 }
2961 }
2962
2963 void MacroAssembler::incrementl(ArrayAddress dst) {
2964 incrementl(as_Address(dst));
2965 }
2966
2967 void MacroAssembler::incrementl(Register reg, int value) {
2968 if (value == min_jint) {addl(reg, value) ; return; }
2969 if (value < 0) { decrementl(reg, -value); return; }
2970 if (value == 0) { ; return; }
2971 if (value == 1 && UseIncDec) { incl(reg) ; return; }
2972 /* else */ { addl(reg, value) ; return; }
2973 }
2974
2975 void MacroAssembler::incrementl(Address dst, int value) {
2976 if (value == min_jint) {addl(dst, value) ; return; }
2977 if (value < 0) { decrementl(dst, -value); return; }
2978 if (value == 0) { ; return; }
2979 if (value == 1 && UseIncDec) { incl(dst) ; return; }
2980 /* else */ { addl(dst, value) ; return; }
2981 }
2982
2983 void MacroAssembler::jump(AddressLiteral dst) {
2984 if (reachable(dst)) {
2985 jmp_literal(dst.target(), dst.rspec());
2986 } else {
2987 lea(rscratch1, dst);
2988 jmp(rscratch1);
2989 }
2990 }
2991
2992 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) {
2993 if (reachable(dst)) {
2994 InstructionMark im(this);
2995 relocate(dst.reloc());
2996 const int short_size = 2;
2997 const int long_size = 6;
2998 int offs = (intptr_t)dst.target() - ((intptr_t)pc());
2999 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
3000 // 0111 tttn #8-bit disp
3001 emit_int8(0x70 | cc);
3002 emit_int8((offs - short_size) & 0xFF);
3003 } else {
3004 // 0000 1111 1000 tttn #32-bit disp
3005 emit_int8(0x0F);
3006 emit_int8((unsigned char)(0x80 | cc));
3007 emit_int32(offs - long_size);
3008 }
3009 } else {
3010 #ifdef ASSERT
3011 warning("reversing conditional branch");
3012 #endif /* ASSERT */
3013 Label skip;
3014 jccb(reverse[cc], skip);
3015 lea(rscratch1, dst);
3016 Assembler::jmp(rscratch1);
3017 bind(skip);
3018 }
3019 }
3020
3021 void MacroAssembler::ldmxcsr(AddressLiteral src) {
3022 if (reachable(src)) {
3023 Assembler::ldmxcsr(as_Address(src));
3024 } else {
3025 lea(rscratch1, src);
3026 Assembler::ldmxcsr(Address(rscratch1, 0));
3027 }
3028 }
3029
3030 int MacroAssembler::load_signed_byte(Register dst, Address src) {
3031 int off;
3032 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3033 off = offset();
3034 movsbl(dst, src); // movsxb
3035 } else {
3036 off = load_unsigned_byte(dst, src);
3037 shll(dst, 24);
3038 sarl(dst, 24);
3039 }
3040 return off;
3041 }
3042
3043 // Note: load_signed_short used to be called load_signed_word.
3044 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler
3045 // manual, which means 16 bits, that usage is found nowhere in HotSpot code.
3046 // The term "word" in HotSpot means a 32- or 64-bit machine word.
3047 int MacroAssembler::load_signed_short(Register dst, Address src) {
3048 int off;
3049 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3050 // This is dubious to me since it seems safe to do a signed 16 => 64 bit
3051 // version but this is what 64bit has always done. This seems to imply
3052 // that users are only using 32bits worth.
3053 off = offset();
3054 movswl(dst, src); // movsxw
3055 } else {
3056 off = load_unsigned_short(dst, src);
3057 shll(dst, 16);
3058 sarl(dst, 16);
3059 }
3060 return off;
3061 }
3062
3063 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
3064 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3065 // and "3.9 Partial Register Penalties", p. 22).
3066 int off;
3067 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) {
3068 off = offset();
3069 movzbl(dst, src); // movzxb
3070 } else {
3071 xorl(dst, dst);
3072 off = offset();
3073 movb(dst, src);
3074 }
3075 return off;
3076 }
3077
3078 // Note: load_unsigned_short used to be called load_unsigned_word.
3079 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
3080 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3081 // and "3.9 Partial Register Penalties", p. 22).
3082 int off;
3083 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) {
3084 off = offset();
3085 movzwl(dst, src); // movzxw
3086 } else {
3087 xorl(dst, dst);
3088 off = offset();
3089 movw(dst, src);
3090 }
3091 return off;
3092 }
3093
3094 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
3095 switch (size_in_bytes) {
3096 #ifndef _LP64
3097 case 8:
3098 assert(dst2 != noreg, "second dest register required");
3099 movl(dst, src);
3100 movl(dst2, src.plus_disp(BytesPerInt));
3101 break;
3102 #else
3103 case 8: movq(dst, src); break;
3104 #endif
3105 case 4: movl(dst, src); break;
3106 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
3107 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
3108 default: ShouldNotReachHere();
3109 }
3110 }
3111
3112 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
3113 switch (size_in_bytes) {
3114 #ifndef _LP64
3115 case 8:
3116 assert(src2 != noreg, "second source register required");
3117 movl(dst, src);
3118 movl(dst.plus_disp(BytesPerInt), src2);
3119 break;
3120 #else
3121 case 8: movq(dst, src); break;
3122 #endif
3123 case 4: movl(dst, src); break;
3124 case 2: movw(dst, src); break;
3125 case 1: movb(dst, src); break;
3126 default: ShouldNotReachHere();
3127 }
3128 }
3129
3130 void MacroAssembler::mov32(AddressLiteral dst, Register src) {
3131 if (reachable(dst)) {
3132 movl(as_Address(dst), src);
3133 } else {
3134 lea(rscratch1, dst);
3135 movl(Address(rscratch1, 0), src);
3136 }
3137 }
3138
3139 void MacroAssembler::mov32(Register dst, AddressLiteral src) {
3140 if (reachable(src)) {
3141 movl(dst, as_Address(src));
3142 } else {
3143 lea(rscratch1, src);
3144 movl(dst, Address(rscratch1, 0));
3145 }
3146 }
3147
3148 // C++ bool manipulation
3149
3150 void MacroAssembler::movbool(Register dst, Address src) {
3151 if(sizeof(bool) == 1)
3152 movb(dst, src);
3153 else if(sizeof(bool) == 2)
3154 movw(dst, src);
3155 else if(sizeof(bool) == 4)
3156 movl(dst, src);
3157 else
3158 // unsupported
3159 ShouldNotReachHere();
3160 }
3161
3162 void MacroAssembler::movbool(Address dst, bool boolconst) {
3163 if(sizeof(bool) == 1)
3164 movb(dst, (int) boolconst);
3165 else if(sizeof(bool) == 2)
3166 movw(dst, (int) boolconst);
3167 else if(sizeof(bool) == 4)
3168 movl(dst, (int) boolconst);
3169 else
3170 // unsupported
3171 ShouldNotReachHere();
3172 }
3173
3174 void MacroAssembler::movbool(Address dst, Register src) {
3175 if(sizeof(bool) == 1)
3176 movb(dst, src);
3177 else if(sizeof(bool) == 2)
3178 movw(dst, src);
3179 else if(sizeof(bool) == 4)
3180 movl(dst, src);
3181 else
3182 // unsupported
3183 ShouldNotReachHere();
3184 }
3185
3186 void MacroAssembler::movbyte(ArrayAddress dst, int src) {
3187 movb(as_Address(dst), src);
3188 }
3189
3190 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) {
3191 if (reachable(src)) {
3192 movdl(dst, as_Address(src));
3193 } else {
3194 lea(rscratch1, src);
3195 movdl(dst, Address(rscratch1, 0));
3196 }
3197 }
3198
3199 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) {
3200 if (reachable(src)) {
3201 movq(dst, as_Address(src));
3202 } else {
3203 lea(rscratch1, src);
3204 movq(dst, Address(rscratch1, 0));
3205 }
3206 }
3207
3208 #ifdef COMPILER2
3209 void MacroAssembler::setvectmask(Register dst, Register src) {
3210 guarantee(PostLoopMultiversioning, "must be");
3211 Assembler::movl(dst, 1);
3212 Assembler::shlxl(dst, dst, src);
3213 Assembler::decl(dst);
3214 Assembler::kmovdl(k1, dst);
3215 Assembler::movl(dst, src);
3216 }
3217
3218 void MacroAssembler::restorevectmask() {
3219 guarantee(PostLoopMultiversioning, "must be");
3220 Assembler::knotwl(k1, k0);
3221 }
3222 #endif // COMPILER2
3223
3224 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) {
3225 if (reachable(src)) {
3226 if (UseXmmLoadAndClearUpper) {
3227 movsd (dst, as_Address(src));
3228 } else {
3229 movlpd(dst, as_Address(src));
3230 }
3231 } else {
3232 lea(rscratch1, src);
3233 if (UseXmmLoadAndClearUpper) {
3234 movsd (dst, Address(rscratch1, 0));
3235 } else {
3236 movlpd(dst, Address(rscratch1, 0));
3237 }
3238 }
3239 }
3240
3241 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) {
3242 if (reachable(src)) {
3243 movss(dst, as_Address(src));
3244 } else {
3245 lea(rscratch1, src);
3246 movss(dst, Address(rscratch1, 0));
3247 }
3248 }
3249
3250 void MacroAssembler::movptr(Register dst, Register src) {
3251 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3252 }
3253
3254 void MacroAssembler::movptr(Register dst, Address src) {
3255 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3256 }
3257
3258 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
3259 void MacroAssembler::movptr(Register dst, intptr_t src) {
3260 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src));
3261 }
3262
3263 void MacroAssembler::movptr(Address dst, Register src) {
3264 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3265 }
3266
3267 void MacroAssembler::movdqu(Address dst, XMMRegister src) {
3268 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3269 Assembler::movdqu(dst, src);
3270 }
3271
3272 void MacroAssembler::movdqu(XMMRegister dst, Address src) {
3273 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3274 Assembler::movdqu(dst, src);
3275 }
3276
3277 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) {
3278 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3279 Assembler::movdqu(dst, src);
3280 }
3281
3282 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src, Register scratchReg) {
3283 if (reachable(src)) {
3284 movdqu(dst, as_Address(src));
3285 } else {
3286 lea(scratchReg, src);
3287 movdqu(dst, Address(scratchReg, 0));
3288 }
3289 }
3290
3291 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) {
3292 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3293 Assembler::vmovdqu(dst, src);
3294 }
3295
3296 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) {
3297 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3298 Assembler::vmovdqu(dst, src);
3299 }
3300
3301 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) {
3302 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3303 Assembler::vmovdqu(dst, src);
3304 }
3305
3306 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
3307 if (reachable(src)) {
3308 vmovdqu(dst, as_Address(src));
3309 }
3310 else {
3311 lea(scratch_reg, src);
3312 vmovdqu(dst, Address(scratch_reg, 0));
3313 }
3314 }
3315
3316 void MacroAssembler::evmovdquq(XMMRegister dst, AddressLiteral src, int vector_len, Register rscratch) {
3317 if (reachable(src)) {
3318 Assembler::evmovdquq(dst, as_Address(src), vector_len);
3319 } else {
3320 lea(rscratch, src);
3321 Assembler::evmovdquq(dst, Address(rscratch, 0), vector_len);
3322 }
3323 }
3324
3325 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) {
3326 if (reachable(src)) {
3327 Assembler::movdqa(dst, as_Address(src));
3328 } else {
3329 lea(rscratch1, src);
3330 Assembler::movdqa(dst, Address(rscratch1, 0));
3331 }
3332 }
3333
3334 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) {
3335 if (reachable(src)) {
3336 Assembler::movsd(dst, as_Address(src));
3337 } else {
3338 lea(rscratch1, src);
3339 Assembler::movsd(dst, Address(rscratch1, 0));
3340 }
3341 }
3342
3343 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) {
3344 if (reachable(src)) {
3345 Assembler::movss(dst, as_Address(src));
3346 } else {
3347 lea(rscratch1, src);
3348 Assembler::movss(dst, Address(rscratch1, 0));
3349 }
3350 }
3351
3352 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) {
3353 if (reachable(src)) {
3354 Assembler::mulsd(dst, as_Address(src));
3355 } else {
3356 lea(rscratch1, src);
3357 Assembler::mulsd(dst, Address(rscratch1, 0));
3358 }
3359 }
3360
3361 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) {
3362 if (reachable(src)) {
3363 Assembler::mulss(dst, as_Address(src));
3364 } else {
3365 lea(rscratch1, src);
3366 Assembler::mulss(dst, Address(rscratch1, 0));
3367 }
3368 }
3369
3370 void MacroAssembler::null_check(Register reg, int offset) {
3371 if (needs_explicit_null_check(offset)) {
3372 // provoke OS NULL exception if reg = NULL by
3373 // accessing M[reg] w/o changing any (non-CC) registers
3374 // NOTE: cmpl is plenty here to provoke a segv
3375 cmpptr(rax, Address(reg, 0));
3376 // Note: should probably use testl(rax, Address(reg, 0));
3377 // may be shorter code (however, this version of
3378 // testl needs to be implemented first)
3379 } else {
3380 // nothing to do, (later) access of M[reg + offset]
3381 // will provoke OS NULL exception if reg = NULL
3382 }
3383 }
3384
3385 void MacroAssembler::os_breakpoint() {
3386 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability
3387 // (e.g., MSVC can't call ps() otherwise)
3388 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint)));
3389 }
3390
3391 void MacroAssembler::unimplemented(const char* what) {
3392 const char* buf = NULL;
3393 {
3394 ResourceMark rm;
3395 stringStream ss;
3396 ss.print("unimplemented: %s", what);
3397 buf = code_string(ss.as_string());
3398 }
3399 stop(buf);
3400 }
3401
3402 #ifdef _LP64
3403 #define XSTATE_BV 0x200
3404 #endif
3405
3406 void MacroAssembler::pop_CPU_state() {
3407 pop_FPU_state();
3408 pop_IU_state();
3409 }
3410
3411 void MacroAssembler::pop_FPU_state() {
3412 #ifndef _LP64
3413 frstor(Address(rsp, 0));
3414 #else
3415 fxrstor(Address(rsp, 0));
3416 #endif
3417 addptr(rsp, FPUStateSizeInWords * wordSize);
3418 }
3419
3420 void MacroAssembler::pop_IU_state() {
3421 popa();
3422 LP64_ONLY(addq(rsp, 8));
3423 popf();
3424 }
3425
3426 // Save Integer and Float state
3427 // Warning: Stack must be 16 byte aligned (64bit)
3428 void MacroAssembler::push_CPU_state() {
3429 push_IU_state();
3430 push_FPU_state();
3431 }
3432
3433 void MacroAssembler::push_FPU_state() {
3434 subptr(rsp, FPUStateSizeInWords * wordSize);
3435 #ifndef _LP64
3436 fnsave(Address(rsp, 0));
3437 fwait();
3438 #else
3439 fxsave(Address(rsp, 0));
3440 #endif // LP64
3441 }
3442
3443 void MacroAssembler::push_IU_state() {
3444 // Push flags first because pusha kills them
3445 pushf();
3446 // Make sure rsp stays 16-byte aligned
3447 LP64_ONLY(subq(rsp, 8));
3448 pusha();
3449 }
3450
3451 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp) { // determine java_thread register
3452 if (!java_thread->is_valid()) {
3453 java_thread = rdi;
3454 get_thread(java_thread);
3455 }
3456 // we must set sp to zero to clear frame
3457 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
3458 if (clear_fp) {
3459 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3460 }
3461
3462 // Always clear the pc because it could have been set by make_walkable()
3463 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
3464
3465 vzeroupper();
3466 }
3467
3468 void MacroAssembler::restore_rax(Register tmp) {
3469 if (tmp == noreg) pop(rax);
3470 else if (tmp != rax) mov(rax, tmp);
3471 }
3472
3473 void MacroAssembler::round_to(Register reg, int modulus) {
3474 addptr(reg, modulus - 1);
3475 andptr(reg, -modulus);
3476 }
3477
3478 void MacroAssembler::save_rax(Register tmp) {
3479 if (tmp == noreg) push(rax);
3480 else if (tmp != rax) mov(tmp, rax);
3481 }
3482
3483 void MacroAssembler::safepoint_poll(Label& slow_path, Register thread_reg, Register temp_reg) {
3484 if (SafepointMechanism::uses_thread_local_poll()) {
3485 #ifdef _LP64
3486 assert(thread_reg == r15_thread, "should be");
3487 #else
3488 if (thread_reg == noreg) {
3489 thread_reg = temp_reg;
3490 get_thread(thread_reg);
3491 }
3492 #endif
3493 testb(Address(thread_reg, Thread::polling_page_offset()), SafepointMechanism::poll_bit());
3494 jcc(Assembler::notZero, slow_path); // handshake bit set implies poll
3495 } else {
3496 cmp32(ExternalAddress(SafepointSynchronize::address_of_state()),
3497 SafepointSynchronize::_not_synchronized);
3498 jcc(Assembler::notEqual, slow_path);
3499 }
3500 }
3501
3502 // Calls to C land
3503 //
3504 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded
3505 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
3506 // has to be reset to 0. This is required to allow proper stack traversal.
3507 void MacroAssembler::set_last_Java_frame(Register java_thread,
3508 Register last_java_sp,
3509 Register last_java_fp,
3510 address last_java_pc) {
3511 vzeroupper();
3512 // determine java_thread register
3513 if (!java_thread->is_valid()) {
3514 java_thread = rdi;
3515 get_thread(java_thread);
3516 }
3517 // determine last_java_sp register
3518 if (!last_java_sp->is_valid()) {
3519 last_java_sp = rsp;
3520 }
3521
3522 // last_java_fp is optional
3523
3524 if (last_java_fp->is_valid()) {
3525 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp);
3526 }
3527
3528 // last_java_pc is optional
3529
3530 if (last_java_pc != NULL) {
3531 lea(Address(java_thread,
3532 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()),
3533 InternalAddress(last_java_pc));
3534
3535 }
3536 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
3537 }
3538
3539 void MacroAssembler::shlptr(Register dst, int imm8) {
3540 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8));
3541 }
3542
3543 void MacroAssembler::shrptr(Register dst, int imm8) {
3544 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8));
3545 }
3546
3547 void MacroAssembler::sign_extend_byte(Register reg) {
3548 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) {
3549 movsbl(reg, reg); // movsxb
3550 } else {
3551 shll(reg, 24);
3552 sarl(reg, 24);
3553 }
3554 }
3555
3556 void MacroAssembler::sign_extend_short(Register reg) {
3557 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3558 movswl(reg, reg); // movsxw
3559 } else {
3560 shll(reg, 16);
3561 sarl(reg, 16);
3562 }
3563 }
3564
3565 void MacroAssembler::testl(Register dst, AddressLiteral src) {
3566 assert(reachable(src), "Address should be reachable");
3567 testl(dst, as_Address(src));
3568 }
3569
3570 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) {
3571 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3572 Assembler::pcmpeqb(dst, src);
3573 }
3574
3575 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) {
3576 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3577 Assembler::pcmpeqw(dst, src);
3578 }
3579
3580 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) {
3581 assert((dst->encoding() < 16),"XMM register should be 0-15");
3582 Assembler::pcmpestri(dst, src, imm8);
3583 }
3584
3585 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) {
3586 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3587 Assembler::pcmpestri(dst, src, imm8);
3588 }
3589
3590 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) {
3591 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3592 Assembler::pmovzxbw(dst, src);
3593 }
3594
3595 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) {
3596 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3597 Assembler::pmovzxbw(dst, src);
3598 }
3599
3600 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) {
3601 assert((src->encoding() < 16),"XMM register should be 0-15");
3602 Assembler::pmovmskb(dst, src);
3603 }
3604
3605 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) {
3606 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3607 Assembler::ptest(dst, src);
3608 }
3609
3610 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) {
3611 if (reachable(src)) {
3612 Assembler::sqrtsd(dst, as_Address(src));
3613 } else {
3614 lea(rscratch1, src);
3615 Assembler::sqrtsd(dst, Address(rscratch1, 0));
3616 }
3617 }
3618
3619 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) {
3620 if (reachable(src)) {
3621 Assembler::sqrtss(dst, as_Address(src));
3622 } else {
3623 lea(rscratch1, src);
3624 Assembler::sqrtss(dst, Address(rscratch1, 0));
3625 }
3626 }
3627
3628 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) {
3629 if (reachable(src)) {
3630 Assembler::subsd(dst, as_Address(src));
3631 } else {
3632 lea(rscratch1, src);
3633 Assembler::subsd(dst, Address(rscratch1, 0));
3634 }
3635 }
3636
3637 void MacroAssembler::roundsd(XMMRegister dst, AddressLiteral src, int32_t rmode, Register scratch_reg) {
3638 if (reachable(src)) {
3639 Assembler::roundsd(dst, as_Address(src), rmode);
3640 } else {
3641 lea(scratch_reg, src);
3642 Assembler::roundsd(dst, Address(scratch_reg, 0), rmode);
3643 }
3644 }
3645
3646 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
3647 if (reachable(src)) {
3648 Assembler::subss(dst, as_Address(src));
3649 } else {
3650 lea(rscratch1, src);
3651 Assembler::subss(dst, Address(rscratch1, 0));
3652 }
3653 }
3654
3655 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
3656 if (reachable(src)) {
3657 Assembler::ucomisd(dst, as_Address(src));
3658 } else {
3659 lea(rscratch1, src);
3660 Assembler::ucomisd(dst, Address(rscratch1, 0));
3661 }
3662 }
3663
3664 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
3665 if (reachable(src)) {
3666 Assembler::ucomiss(dst, as_Address(src));
3667 } else {
3668 lea(rscratch1, src);
3669 Assembler::ucomiss(dst, Address(rscratch1, 0));
3670 }
3671 }
3672
3673 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
3674 // Used in sign-bit flipping with aligned address.
3675 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3676 if (reachable(src)) {
3677 Assembler::xorpd(dst, as_Address(src));
3678 } else {
3679 lea(scratch_reg, src);
3680 Assembler::xorpd(dst, Address(scratch_reg, 0));
3681 }
3682 }
3683
3684 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) {
3685 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
3686 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
3687 }
3688 else {
3689 Assembler::xorpd(dst, src);
3690 }
3691 }
3692
3693 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) {
3694 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
3695 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
3696 } else {
3697 Assembler::xorps(dst, src);
3698 }
3699 }
3700
3701 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
3702 // Used in sign-bit flipping with aligned address.
3703 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3704 if (reachable(src)) {
3705 Assembler::xorps(dst, as_Address(src));
3706 } else {
3707 lea(scratch_reg, src);
3708 Assembler::xorps(dst, Address(scratch_reg, 0));
3709 }
3710 }
3711
3712 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
3713 // Used in sign-bit flipping with aligned address.
3714 bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
3715 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
3716 if (reachable(src)) {
3717 Assembler::pshufb(dst, as_Address(src));
3718 } else {
3719 lea(rscratch1, src);
3720 Assembler::pshufb(dst, Address(rscratch1, 0));
3721 }
3722 }
3723
3724 // AVX 3-operands instructions
3725
3726 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3727 if (reachable(src)) {
3728 vaddsd(dst, nds, as_Address(src));
3729 } else {
3730 lea(rscratch1, src);
3731 vaddsd(dst, nds, Address(rscratch1, 0));
3732 }
3733 }
3734
3735 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3736 if (reachable(src)) {
3737 vaddss(dst, nds, as_Address(src));
3738 } else {
3739 lea(rscratch1, src);
3740 vaddss(dst, nds, Address(rscratch1, 0));
3741 }
3742 }
3743
3744 void MacroAssembler::vpaddd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register rscratch) {
3745 assert(UseAVX > 0, "requires some form of AVX");
3746 if (reachable(src)) {
3747 Assembler::vpaddd(dst, nds, as_Address(src), vector_len);
3748 } else {
3749 lea(rscratch, src);
3750 Assembler::vpaddd(dst, nds, Address(rscratch, 0), vector_len);
3751 }
3752 }
3753
3754 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
3755 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
3756 vandps(dst, nds, negate_field, vector_len);
3757 }
3758
3759 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
3760 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
3761 vandpd(dst, nds, negate_field, vector_len);
3762 }
3763
3764 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3765 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3766 Assembler::vpaddb(dst, nds, src, vector_len);
3767 }
3768
3769 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3770 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3771 Assembler::vpaddb(dst, nds, src, vector_len);
3772 }
3773
3774 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3775 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3776 Assembler::vpaddw(dst, nds, src, vector_len);
3777 }
3778
3779 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3780 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3781 Assembler::vpaddw(dst, nds, src, vector_len);
3782 }
3783
3784 void MacroAssembler::vpand(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
3785 if (reachable(src)) {
3786 Assembler::vpand(dst, nds, as_Address(src), vector_len);
3787 } else {
3788 lea(scratch_reg, src);
3789 Assembler::vpand(dst, nds, Address(scratch_reg, 0), vector_len);
3790 }
3791 }
3792
3793 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src, int vector_len) {
3794 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3795 Assembler::vpbroadcastw(dst, src, vector_len);
3796 }
3797
3798 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3799 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3800 Assembler::vpcmpeqb(dst, nds, src, vector_len);
3801 }
3802
3803 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3804 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3805 Assembler::vpcmpeqw(dst, nds, src, vector_len);
3806 }
3807
3808 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) {
3809 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3810 Assembler::vpmovzxbw(dst, src, vector_len);
3811 }
3812
3813 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) {
3814 assert((src->encoding() < 16),"XMM register should be 0-15");
3815 Assembler::vpmovmskb(dst, src);
3816 }
3817
3818 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3819 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3820 Assembler::vpmullw(dst, nds, src, vector_len);
3821 }
3822
3823 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3824 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3825 Assembler::vpmullw(dst, nds, src, vector_len);
3826 }
3827
3828 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3829 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3830 Assembler::vpsubb(dst, nds, src, vector_len);
3831 }
3832
3833 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3834 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3835 Assembler::vpsubb(dst, nds, src, vector_len);
3836 }
3837
3838 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3839 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3840 Assembler::vpsubw(dst, nds, src, vector_len);
3841 }
3842
3843 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3844 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3845 Assembler::vpsubw(dst, nds, src, vector_len);
3846 }
3847
3848 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3849 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3850 Assembler::vpsraw(dst, nds, shift, vector_len);
3851 }
3852
3853 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3854 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3855 Assembler::vpsraw(dst, nds, shift, vector_len);
3856 }
3857
3858 void MacroAssembler::evpsraq(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3859 assert(UseAVX > 2,"");
3860 if (!VM_Version::supports_avx512vl() && vector_len < 2) {
3861 vector_len = 2;
3862 }
3863 Assembler::evpsraq(dst, nds, shift, vector_len);
3864 }
3865
3866 void MacroAssembler::evpsraq(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3867 assert(UseAVX > 2,"");
3868 if (!VM_Version::supports_avx512vl() && vector_len < 2) {
3869 vector_len = 2;
3870 }
3871 Assembler::evpsraq(dst, nds, shift, vector_len);
3872 }
3873
3874 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3875 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3876 Assembler::vpsrlw(dst, nds, shift, vector_len);
3877 }
3878
3879 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3880 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3881 Assembler::vpsrlw(dst, nds, shift, vector_len);
3882 }
3883
3884 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3885 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3886 Assembler::vpsllw(dst, nds, shift, vector_len);
3887 }
3888
3889 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3890 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3891 Assembler::vpsllw(dst, nds, shift, vector_len);
3892 }
3893
3894 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) {
3895 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3896 Assembler::vptest(dst, src);
3897 }
3898
3899 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) {
3900 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3901 Assembler::punpcklbw(dst, src);
3902 }
3903
3904 void MacroAssembler::pshufd(XMMRegister dst, Address src, int mode) {
3905 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3906 Assembler::pshufd(dst, src, mode);
3907 }
3908
3909 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) {
3910 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3911 Assembler::pshuflw(dst, src, mode);
3912 }
3913
3914 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
3915 if (reachable(src)) {
3916 vandpd(dst, nds, as_Address(src), vector_len);
3917 } else {
3918 lea(scratch_reg, src);
3919 vandpd(dst, nds, Address(scratch_reg, 0), vector_len);
3920 }
3921 }
3922
3923 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
3924 if (reachable(src)) {
3925 vandps(dst, nds, as_Address(src), vector_len);
3926 } else {
3927 lea(scratch_reg, src);
3928 vandps(dst, nds, Address(scratch_reg, 0), vector_len);
3929 }
3930 }
3931
3932 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3933 if (reachable(src)) {
3934 vdivsd(dst, nds, as_Address(src));
3935 } else {
3936 lea(rscratch1, src);
3937 vdivsd(dst, nds, Address(rscratch1, 0));
3938 }
3939 }
3940
3941 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3942 if (reachable(src)) {
3943 vdivss(dst, nds, as_Address(src));
3944 } else {
3945 lea(rscratch1, src);
3946 vdivss(dst, nds, Address(rscratch1, 0));
3947 }
3948 }
3949
3950 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3951 if (reachable(src)) {
3952 vmulsd(dst, nds, as_Address(src));
3953 } else {
3954 lea(rscratch1, src);
3955 vmulsd(dst, nds, Address(rscratch1, 0));
3956 }
3957 }
3958
3959 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3960 if (reachable(src)) {
3961 vmulss(dst, nds, as_Address(src));
3962 } else {
3963 lea(rscratch1, src);
3964 vmulss(dst, nds, Address(rscratch1, 0));
3965 }
3966 }
3967
3968 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3969 if (reachable(src)) {
3970 vsubsd(dst, nds, as_Address(src));
3971 } else {
3972 lea(rscratch1, src);
3973 vsubsd(dst, nds, Address(rscratch1, 0));
3974 }
3975 }
3976
3977 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3978 if (reachable(src)) {
3979 vsubss(dst, nds, as_Address(src));
3980 } else {
3981 lea(rscratch1, src);
3982 vsubss(dst, nds, Address(rscratch1, 0));
3983 }
3984 }
3985
3986 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3987 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
3988 vxorps(dst, nds, src, Assembler::AVX_128bit);
3989 }
3990
3991 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3992 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
3993 vxorpd(dst, nds, src, Assembler::AVX_128bit);
3994 }
3995
3996 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
3997 if (reachable(src)) {
3998 vxorpd(dst, nds, as_Address(src), vector_len);
3999 } else {
4000 lea(scratch_reg, src);
4001 vxorpd(dst, nds, Address(scratch_reg, 0), vector_len);
4002 }
4003 }
4004
4005 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
4006 if (reachable(src)) {
4007 vxorps(dst, nds, as_Address(src), vector_len);
4008 } else {
4009 lea(scratch_reg, src);
4010 vxorps(dst, nds, Address(scratch_reg, 0), vector_len);
4011 }
4012 }
4013
4014 void MacroAssembler::vpxor(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
4015 if (UseAVX > 1 || (vector_len < 1)) {
4016 if (reachable(src)) {
4017 Assembler::vpxor(dst, nds, as_Address(src), vector_len);
4018 } else {
4019 lea(scratch_reg, src);
4020 Assembler::vpxor(dst, nds, Address(scratch_reg, 0), vector_len);
4021 }
4022 }
4023 else {
4024 MacroAssembler::vxorpd(dst, nds, src, vector_len, scratch_reg);
4025 }
4026 }
4027
4028 //-------------------------------------------------------------------------------------------
4029 #ifdef COMPILER2
4030 // Generic instructions support for use in .ad files C2 code generation
4031
4032 void MacroAssembler::vabsnegd(int opcode, XMMRegister dst, XMMRegister src, Register scr) {
4033 if (dst != src) {
4034 movdqu(dst, src);
4035 }
4036 if (opcode == Op_AbsVD) {
4037 andpd(dst, ExternalAddress(StubRoutines::x86::vector_double_sign_mask()), scr);
4038 } else {
4039 assert((opcode == Op_NegVD),"opcode should be Op_NegD");
4040 xorpd(dst, ExternalAddress(StubRoutines::x86::vector_double_sign_flip()), scr);
4041 }
4042 }
4043
4044 void MacroAssembler::vabsnegd(int opcode, XMMRegister dst, XMMRegister src, int vector_len, Register scr) {
4045 if (opcode == Op_AbsVD) {
4046 vandpd(dst, src, ExternalAddress(StubRoutines::x86::vector_double_sign_mask()), vector_len, scr);
4047 } else {
4048 assert((opcode == Op_NegVD),"opcode should be Op_NegD");
4049 vxorpd(dst, src, ExternalAddress(StubRoutines::x86::vector_double_sign_flip()), vector_len, scr);
4050 }
4051 }
4052
4053 void MacroAssembler::vabsnegf(int opcode, XMMRegister dst, XMMRegister src, Register scr) {
4054 if (dst != src) {
4055 movdqu(dst, src);
4056 }
4057 if (opcode == Op_AbsVF) {
4058 andps(dst, ExternalAddress(StubRoutines::x86::vector_float_sign_mask()), scr);
4059 } else {
4060 assert((opcode == Op_NegVF),"opcode should be Op_NegF");
4061 xorps(dst, ExternalAddress(StubRoutines::x86::vector_float_sign_flip()), scr);
4062 }
4063 }
4064
4065 void MacroAssembler::vabsnegf(int opcode, XMMRegister dst, XMMRegister src, int vector_len, Register scr) {
4066 if (opcode == Op_AbsVF) {
4067 vandps(dst, src, ExternalAddress(StubRoutines::x86::vector_float_sign_mask()), vector_len, scr);
4068 } else {
4069 assert((opcode == Op_NegVF),"opcode should be Op_NegF");
4070 vxorps(dst, src, ExternalAddress(StubRoutines::x86::vector_float_sign_flip()), vector_len, scr);
4071 }
4072 }
4073
4074 void MacroAssembler::vextendbw(bool sign, XMMRegister dst, XMMRegister src) {
4075 if (sign) {
4076 pmovsxbw(dst, src);
4077 } else {
4078 pmovzxbw(dst, src);
4079 }
4080 }
4081
4082 void MacroAssembler::vextendbw(bool sign, XMMRegister dst, XMMRegister src, int vector_len) {
4083 if (sign) {
4084 vpmovsxbw(dst, src, vector_len);
4085 } else {
4086 vpmovzxbw(dst, src, vector_len);
4087 }
4088 }
4089
4090 void MacroAssembler::vshiftd(int opcode, XMMRegister dst, XMMRegister src) {
4091 if (opcode == Op_RShiftVI) {
4092 psrad(dst, src);
4093 } else if (opcode == Op_LShiftVI) {
4094 pslld(dst, src);
4095 } else {
4096 assert((opcode == Op_URShiftVI),"opcode should be Op_URShiftVI");
4097 psrld(dst, src);
4098 }
4099 }
4100
4101 void MacroAssembler::vshiftd(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4102 if (opcode == Op_RShiftVI) {
4103 vpsrad(dst, nds, src, vector_len);
4104 } else if (opcode == Op_LShiftVI) {
4105 vpslld(dst, nds, src, vector_len);
4106 } else {
4107 assert((opcode == Op_URShiftVI),"opcode should be Op_URShiftVI");
4108 vpsrld(dst, nds, src, vector_len);
4109 }
4110 }
4111
4112 void MacroAssembler::vshiftw(int opcode, XMMRegister dst, XMMRegister src) {
4113 if ((opcode == Op_RShiftVS) || (opcode == Op_RShiftVB)) {
4114 psraw(dst, src);
4115 } else if ((opcode == Op_LShiftVS) || (opcode == Op_LShiftVB)) {
4116 psllw(dst, src);
4117 } else {
4118 assert(((opcode == Op_URShiftVS) || (opcode == Op_URShiftVB)),"opcode should be one of Op_URShiftVS or Op_URShiftVB");
4119 psrlw(dst, src);
4120 }
4121 }
4122
4123 void MacroAssembler::vshiftw(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4124 if ((opcode == Op_RShiftVS) || (opcode == Op_RShiftVB)) {
4125 vpsraw(dst, nds, src, vector_len);
4126 } else if ((opcode == Op_LShiftVS) || (opcode == Op_LShiftVB)) {
4127 vpsllw(dst, nds, src, vector_len);
4128 } else {
4129 assert(((opcode == Op_URShiftVS) || (opcode == Op_URShiftVB)),"opcode should be one of Op_URShiftVS or Op_URShiftVB");
4130 vpsrlw(dst, nds, src, vector_len);
4131 }
4132 }
4133
4134 void MacroAssembler::vshiftq(int opcode, XMMRegister dst, XMMRegister src) {
4135 if (opcode == Op_RShiftVL) {
4136 psrlq(dst, src); // using srl to implement sra on pre-avs512 systems
4137 } else if (opcode == Op_LShiftVL) {
4138 psllq(dst, src);
4139 } else {
4140 assert((opcode == Op_URShiftVL),"opcode should be Op_URShiftVL");
4141 psrlq(dst, src);
4142 }
4143 }
4144
4145 void MacroAssembler::vshiftq(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4146 if (opcode == Op_RShiftVL) {
4147 evpsraq(dst, nds, src, vector_len);
4148 } else if (opcode == Op_LShiftVL) {
4149 vpsllq(dst, nds, src, vector_len);
4150 } else {
4151 assert((opcode == Op_URShiftVL),"opcode should be Op_URShiftVL");
4152 vpsrlq(dst, nds, src, vector_len);
4153 }
4154 }
4155 #endif
4156 //-------------------------------------------------------------------------------------------
4157
4158 void MacroAssembler::clear_jweak_tag(Register possibly_jweak) {
4159 const int32_t inverted_jweak_mask = ~static_cast<int32_t>(JNIHandles::weak_tag_mask);
4160 STATIC_ASSERT(inverted_jweak_mask == -2); // otherwise check this code
4161 // The inverted mask is sign-extended
4162 andptr(possibly_jweak, inverted_jweak_mask);
4163 }
4164
4165 void MacroAssembler::resolve_jobject(Register value,
4166 Register thread,
4167 Register tmp) {
4168 assert_different_registers(value, thread, tmp);
4169 Label done, not_weak;
4170 testptr(value, value);
4171 jcc(Assembler::zero, done); // Use NULL as-is.
4172 testptr(value, JNIHandles::weak_tag_mask); // Test for jweak tag.
4173 jcc(Assembler::zero, not_weak);
4174 // Resolve jweak.
4175 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
4176 value, Address(value, -JNIHandles::weak_tag_value), tmp, thread);
4177 verify_oop(value);
4178 jmp(done);
4179 bind(not_weak);
4180 // Resolve (untagged) jobject.
4181 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, 0), tmp, thread);
4182 verify_oop(value);
4183 bind(done);
4184 }
4185
4186 void MacroAssembler::subptr(Register dst, int32_t imm32) {
4187 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32));
4188 }
4189
4190 // Force generation of a 4 byte immediate value even if it fits into 8bit
4191 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) {
4192 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32));
4193 }
4194
4195 void MacroAssembler::subptr(Register dst, Register src) {
4196 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src));
4197 }
4198
4199 // C++ bool manipulation
4200 void MacroAssembler::testbool(Register dst) {
4201 if(sizeof(bool) == 1)
4202 testb(dst, 0xff);
4203 else if(sizeof(bool) == 2) {
4204 // testw implementation needed for two byte bools
4205 ShouldNotReachHere();
4206 } else if(sizeof(bool) == 4)
4207 testl(dst, dst);
4208 else
4209 // unsupported
4210 ShouldNotReachHere();
4211 }
4212
4213 void MacroAssembler::testptr(Register dst, Register src) {
4214 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src));
4215 }
4216
4217 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
4218 void MacroAssembler::tlab_allocate(Register thread, Register obj,
4219 Register var_size_in_bytes,
4220 int con_size_in_bytes,
4221 Register t1,
4222 Register t2,
4223 Label& slow_case) {
4224 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4225 bs->tlab_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
4226 }
4227
4228 // Defines obj, preserves var_size_in_bytes
4229 void MacroAssembler::eden_allocate(Register thread, Register obj,
4230 Register var_size_in_bytes,
4231 int con_size_in_bytes,
4232 Register t1,
4233 Label& slow_case) {
4234 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4235 bs->eden_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case);
4236 }
4237
4238 // Preserves the contents of address, destroys the contents length_in_bytes and temp.
4239 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) {
4240 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different");
4241 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord");
4242 Label done;
4243
4244 testptr(length_in_bytes, length_in_bytes);
4245 jcc(Assembler::zero, done);
4246
4247 // initialize topmost word, divide index by 2, check if odd and test if zero
4248 // note: for the remaining code to work, index must be a multiple of BytesPerWord
4249 #ifdef ASSERT
4250 {
4251 Label L;
4252 testptr(length_in_bytes, BytesPerWord - 1);
4253 jcc(Assembler::zero, L);
4254 stop("length must be a multiple of BytesPerWord");
4255 bind(L);
4256 }
4257 #endif
4258 Register index = length_in_bytes;
4259 xorptr(temp, temp); // use _zero reg to clear memory (shorter code)
4260 if (UseIncDec) {
4261 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set
4262 } else {
4263 shrptr(index, 2); // use 2 instructions to avoid partial flag stall
4264 shrptr(index, 1);
4265 }
4266 #ifndef _LP64
4267 // index could have not been a multiple of 8 (i.e., bit 2 was set)
4268 {
4269 Label even;
4270 // note: if index was a multiple of 8, then it cannot
4271 // be 0 now otherwise it must have been 0 before
4272 // => if it is even, we don't need to check for 0 again
4273 jcc(Assembler::carryClear, even);
4274 // clear topmost word (no jump would be needed if conditional assignment worked here)
4275 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp);
4276 // index could be 0 now, must check again
4277 jcc(Assembler::zero, done);
4278 bind(even);
4279 }
4280 #endif // !_LP64
4281 // initialize remaining object fields: index is a multiple of 2 now
4282 {
4283 Label loop;
4284 bind(loop);
4285 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp);
4286 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);)
4287 decrement(index);
4288 jcc(Assembler::notZero, loop);
4289 }
4290
4291 bind(done);
4292 }
4293
4294 // Look up the method for a megamorphic invokeinterface call.
4295 // The target method is determined by <intf_klass, itable_index>.
4296 // The receiver klass is in recv_klass.
4297 // On success, the result will be in method_result, and execution falls through.
4298 // On failure, execution transfers to the given label.
4299 void MacroAssembler::lookup_interface_method(Register recv_klass,
4300 Register intf_klass,
4301 RegisterOrConstant itable_index,
4302 Register method_result,
4303 Register scan_temp,
4304 Label& L_no_such_interface,
4305 bool return_method) {
4306 assert_different_registers(recv_klass, intf_klass, scan_temp);
4307 assert_different_registers(method_result, intf_klass, scan_temp);
4308 assert(recv_klass != method_result || !return_method,
4309 "recv_klass can be destroyed when method isn't needed");
4310
4311 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
4312 "caller must use same register for non-constant itable index as for method");
4313
4314 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
4315 int vtable_base = in_bytes(Klass::vtable_start_offset());
4316 int itentry_off = itableMethodEntry::method_offset_in_bytes();
4317 int scan_step = itableOffsetEntry::size() * wordSize;
4318 int vte_size = vtableEntry::size_in_bytes();
4319 Address::ScaleFactor times_vte_scale = Address::times_ptr;
4320 assert(vte_size == wordSize, "else adjust times_vte_scale");
4321
4322 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
4323
4324 // %%% Could store the aligned, prescaled offset in the klassoop.
4325 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
4326
4327 if (return_method) {
4328 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
4329 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
4330 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
4331 }
4332
4333 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
4334 // if (scan->interface() == intf) {
4335 // result = (klass + scan->offset() + itable_index);
4336 // }
4337 // }
4338 Label search, found_method;
4339
4340 for (int peel = 1; peel >= 0; peel--) {
4341 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
4342 cmpptr(intf_klass, method_result);
4343
4344 if (peel) {
4345 jccb(Assembler::equal, found_method);
4346 } else {
4347 jccb(Assembler::notEqual, search);
4348 // (invert the test to fall through to found_method...)
4349 }
4350
4351 if (!peel) break;
4352
4353 bind(search);
4354
4355 // Check that the previous entry is non-null. A null entry means that
4356 // the receiver class doesn't implement the interface, and wasn't the
4357 // same as when the caller was compiled.
4358 testptr(method_result, method_result);
4359 jcc(Assembler::zero, L_no_such_interface);
4360 addptr(scan_temp, scan_step);
4361 }
4362
4363 bind(found_method);
4364
4365 if (return_method) {
4366 // Got a hit.
4367 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
4368 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
4369 }
4370 }
4371
4372
4373 // virtual method calling
4374 void MacroAssembler::lookup_virtual_method(Register recv_klass,
4375 RegisterOrConstant vtable_index,
4376 Register method_result) {
4377 const int base = in_bytes(Klass::vtable_start_offset());
4378 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below");
4379 Address vtable_entry_addr(recv_klass,
4380 vtable_index, Address::times_ptr,
4381 base + vtableEntry::method_offset_in_bytes());
4382 movptr(method_result, vtable_entry_addr);
4383 }
4384
4385
4386 void MacroAssembler::check_klass_subtype(Register sub_klass,
4387 Register super_klass,
4388 Register temp_reg,
4389 Label& L_success) {
4390 Label L_failure;
4391 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL);
4392 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
4393 bind(L_failure);
4394 }
4395
4396
4397 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
4398 Register super_klass,
4399 Register temp_reg,
4400 Label* L_success,
4401 Label* L_failure,
4402 Label* L_slow_path,
4403 RegisterOrConstant super_check_offset) {
4404 assert_different_registers(sub_klass, super_klass, temp_reg);
4405 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
4406 if (super_check_offset.is_register()) {
4407 assert_different_registers(sub_klass, super_klass,
4408 super_check_offset.as_register());
4409 } else if (must_load_sco) {
4410 assert(temp_reg != noreg, "supply either a temp or a register offset");
4411 }
4412
4413 Label L_fallthrough;
4414 int label_nulls = 0;
4415 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
4416 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
4417 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
4418 assert(label_nulls <= 1, "at most one NULL in the batch");
4419
4420 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
4421 int sco_offset = in_bytes(Klass::super_check_offset_offset());
4422 Address super_check_offset_addr(super_klass, sco_offset);
4423
4424 // Hacked jcc, which "knows" that L_fallthrough, at least, is in
4425 // range of a jccb. If this routine grows larger, reconsider at
4426 // least some of these.
4427 #define local_jcc(assembler_cond, label) \
4428 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \
4429 else jcc( assembler_cond, label) /*omit semi*/
4430
4431 // Hacked jmp, which may only be used just before L_fallthrough.
4432 #define final_jmp(label) \
4433 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
4434 else jmp(label) /*omit semi*/
4435
4436 // If the pointers are equal, we are done (e.g., String[] elements).
4437 // This self-check enables sharing of secondary supertype arrays among
4438 // non-primary types such as array-of-interface. Otherwise, each such
4439 // type would need its own customized SSA.
4440 // We move this check to the front of the fast path because many
4441 // type checks are in fact trivially successful in this manner,
4442 // so we get a nicely predicted branch right at the start of the check.
4443 cmpptr(sub_klass, super_klass);
4444 local_jcc(Assembler::equal, *L_success);
4445
4446 // Check the supertype display:
4447 if (must_load_sco) {
4448 // Positive movl does right thing on LP64.
4449 movl(temp_reg, super_check_offset_addr);
4450 super_check_offset = RegisterOrConstant(temp_reg);
4451 }
4452 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0);
4453 cmpptr(super_klass, super_check_addr); // load displayed supertype
4454
4455 // This check has worked decisively for primary supers.
4456 // Secondary supers are sought in the super_cache ('super_cache_addr').
4457 // (Secondary supers are interfaces and very deeply nested subtypes.)
4458 // This works in the same check above because of a tricky aliasing
4459 // between the super_cache and the primary super display elements.
4460 // (The 'super_check_addr' can address either, as the case requires.)
4461 // Note that the cache is updated below if it does not help us find
4462 // what we need immediately.
4463 // So if it was a primary super, we can just fail immediately.
4464 // Otherwise, it's the slow path for us (no success at this point).
4465
4466 if (super_check_offset.is_register()) {
4467 local_jcc(Assembler::equal, *L_success);
4468 cmpl(super_check_offset.as_register(), sc_offset);
4469 if (L_failure == &L_fallthrough) {
4470 local_jcc(Assembler::equal, *L_slow_path);
4471 } else {
4472 local_jcc(Assembler::notEqual, *L_failure);
4473 final_jmp(*L_slow_path);
4474 }
4475 } else if (super_check_offset.as_constant() == sc_offset) {
4476 // Need a slow path; fast failure is impossible.
4477 if (L_slow_path == &L_fallthrough) {
4478 local_jcc(Assembler::equal, *L_success);
4479 } else {
4480 local_jcc(Assembler::notEqual, *L_slow_path);
4481 final_jmp(*L_success);
4482 }
4483 } else {
4484 // No slow path; it's a fast decision.
4485 if (L_failure == &L_fallthrough) {
4486 local_jcc(Assembler::equal, *L_success);
4487 } else {
4488 local_jcc(Assembler::notEqual, *L_failure);
4489 final_jmp(*L_success);
4490 }
4491 }
4492
4493 bind(L_fallthrough);
4494
4495 #undef local_jcc
4496 #undef final_jmp
4497 }
4498
4499
4500 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
4501 Register super_klass,
4502 Register temp_reg,
4503 Register temp2_reg,
4504 Label* L_success,
4505 Label* L_failure,
4506 bool set_cond_codes) {
4507 assert_different_registers(sub_klass, super_klass, temp_reg);
4508 if (temp2_reg != noreg)
4509 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg);
4510 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
4511
4512 Label L_fallthrough;
4513 int label_nulls = 0;
4514 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
4515 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
4516 assert(label_nulls <= 1, "at most one NULL in the batch");
4517
4518 // a couple of useful fields in sub_klass:
4519 int ss_offset = in_bytes(Klass::secondary_supers_offset());
4520 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
4521 Address secondary_supers_addr(sub_klass, ss_offset);
4522 Address super_cache_addr( sub_klass, sc_offset);
4523
4524 // Do a linear scan of the secondary super-klass chain.
4525 // This code is rarely used, so simplicity is a virtue here.
4526 // The repne_scan instruction uses fixed registers, which we must spill.
4527 // Don't worry too much about pre-existing connections with the input regs.
4528
4529 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super)
4530 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter)
4531
4532 // Get super_klass value into rax (even if it was in rdi or rcx).
4533 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false;
4534 if (super_klass != rax || UseCompressedOops) {
4535 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; }
4536 mov(rax, super_klass);
4537 }
4538 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; }
4539 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; }
4540
4541 #ifndef PRODUCT
4542 int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
4543 ExternalAddress pst_counter_addr((address) pst_counter);
4544 NOT_LP64( incrementl(pst_counter_addr) );
4545 LP64_ONLY( lea(rcx, pst_counter_addr) );
4546 LP64_ONLY( incrementl(Address(rcx, 0)) );
4547 #endif //PRODUCT
4548
4549 // We will consult the secondary-super array.
4550 movptr(rdi, secondary_supers_addr);
4551 // Load the array length. (Positive movl does right thing on LP64.)
4552 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes()));
4553 // Skip to start of data.
4554 addptr(rdi, Array<Klass*>::base_offset_in_bytes());
4555
4556 // Scan RCX words at [RDI] for an occurrence of RAX.
4557 // Set NZ/Z based on last compare.
4558 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does
4559 // not change flags (only scas instruction which is repeated sets flags).
4560 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found.
4561
4562 testptr(rax,rax); // Set Z = 0
4563 repne_scan();
4564
4565 // Unspill the temp. registers:
4566 if (pushed_rdi) pop(rdi);
4567 if (pushed_rcx) pop(rcx);
4568 if (pushed_rax) pop(rax);
4569
4570 if (set_cond_codes) {
4571 // Special hack for the AD files: rdi is guaranteed non-zero.
4572 assert(!pushed_rdi, "rdi must be left non-NULL");
4573 // Also, the condition codes are properly set Z/NZ on succeed/failure.
4574 }
4575
4576 if (L_failure == &L_fallthrough)
4577 jccb(Assembler::notEqual, *L_failure);
4578 else jcc(Assembler::notEqual, *L_failure);
4579
4580 // Success. Cache the super we found and proceed in triumph.
4581 movptr(super_cache_addr, super_klass);
4582
4583 if (L_success != &L_fallthrough) {
4584 jmp(*L_success);
4585 }
4586
4587 #undef IS_A_TEMP
4588
4589 bind(L_fallthrough);
4590 }
4591
4592 void MacroAssembler::clinit_barrier(Register klass, Register thread, Label* L_fast_path, Label* L_slow_path) {
4593 assert(L_fast_path != NULL || L_slow_path != NULL, "at least one is required");
4594
4595 Label L_fallthrough;
4596 if (L_fast_path == NULL) {
4597 L_fast_path = &L_fallthrough;
4598 } else if (L_slow_path == NULL) {
4599 L_slow_path = &L_fallthrough;
4600 }
4601
4602 // Fast path check: class is fully initialized
4603 cmpb(Address(klass, InstanceKlass::init_state_offset()), InstanceKlass::fully_initialized);
4604 jcc(Assembler::equal, *L_fast_path);
4605
4606 // Fast path check: current thread is initializer thread
4607 cmpptr(thread, Address(klass, InstanceKlass::init_thread_offset()));
4608 if (L_slow_path == &L_fallthrough) {
4609 jcc(Assembler::equal, *L_fast_path);
4610 bind(*L_slow_path);
4611 } else if (L_fast_path == &L_fallthrough) {
4612 jcc(Assembler::notEqual, *L_slow_path);
4613 bind(*L_fast_path);
4614 } else {
4615 Unimplemented();
4616 }
4617 }
4618
4619 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
4620 if (VM_Version::supports_cmov()) {
4621 cmovl(cc, dst, src);
4622 } else {
4623 Label L;
4624 jccb(negate_condition(cc), L);
4625 movl(dst, src);
4626 bind(L);
4627 }
4628 }
4629
4630 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) {
4631 if (VM_Version::supports_cmov()) {
4632 cmovl(cc, dst, src);
4633 } else {
4634 Label L;
4635 jccb(negate_condition(cc), L);
4636 movl(dst, src);
4637 bind(L);
4638 }
4639 }
4640
4641 void MacroAssembler::_verify_oop(Register reg, const char* s, const char* file, int line) {
4642 if (!VerifyOops) return;
4643
4644 // Pass register number to verify_oop_subroutine
4645 const char* b = NULL;
4646 {
4647 ResourceMark rm;
4648 stringStream ss;
4649 ss.print("verify_oop: %s: %s (%s:%d)", reg->name(), s, file, line);
4650 b = code_string(ss.as_string());
4651 }
4652 BLOCK_COMMENT("verify_oop {");
4653 #ifdef _LP64
4654 push(rscratch1); // save r10, trashed by movptr()
4655 #endif
4656 push(rax); // save rax,
4657 push(reg); // pass register argument
4658 ExternalAddress buffer((address) b);
4659 // avoid using pushptr, as it modifies scratch registers
4660 // and our contract is not to modify anything
4661 movptr(rax, buffer.addr());
4662 push(rax);
4663 // call indirectly to solve generation ordering problem
4664 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
4665 call(rax);
4666 // Caller pops the arguments (oop, message) and restores rax, r10
4667 BLOCK_COMMENT("} verify_oop");
4668 }
4669
4670
4671 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
4672 Register tmp,
4673 int offset) {
4674 intptr_t value = *delayed_value_addr;
4675 if (value != 0)
4676 return RegisterOrConstant(value + offset);
4677
4678 // load indirectly to solve generation ordering problem
4679 movptr(tmp, ExternalAddress((address) delayed_value_addr));
4680
4681 #ifdef ASSERT
4682 { Label L;
4683 testptr(tmp, tmp);
4684 if (WizardMode) {
4685 const char* buf = NULL;
4686 {
4687 ResourceMark rm;
4688 stringStream ss;
4689 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]);
4690 buf = code_string(ss.as_string());
4691 }
4692 jcc(Assembler::notZero, L);
4693 STOP(buf);
4694 } else {
4695 jccb(Assembler::notZero, L);
4696 hlt();
4697 }
4698 bind(L);
4699 }
4700 #endif
4701
4702 if (offset != 0)
4703 addptr(tmp, offset);
4704
4705 return RegisterOrConstant(tmp);
4706 }
4707
4708
4709 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
4710 int extra_slot_offset) {
4711 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
4712 int stackElementSize = Interpreter::stackElementSize;
4713 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
4714 #ifdef ASSERT
4715 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
4716 assert(offset1 - offset == stackElementSize, "correct arithmetic");
4717 #endif
4718 Register scale_reg = noreg;
4719 Address::ScaleFactor scale_factor = Address::no_scale;
4720 if (arg_slot.is_constant()) {
4721 offset += arg_slot.as_constant() * stackElementSize;
4722 } else {
4723 scale_reg = arg_slot.as_register();
4724 scale_factor = Address::times(stackElementSize);
4725 }
4726 offset += wordSize; // return PC is on stack
4727 return Address(rsp, scale_reg, scale_factor, offset);
4728 }
4729
4730
4731 void MacroAssembler::_verify_oop_addr(Address addr, const char* s, const char* file, int line) {
4732 if (!VerifyOops) return;
4733
4734 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord);
4735 // Pass register number to verify_oop_subroutine
4736 const char* b = NULL;
4737 {
4738 ResourceMark rm;
4739 stringStream ss;
4740 ss.print("verify_oop_addr: %s (%s:%d)", s, file, line);
4741 b = code_string(ss.as_string());
4742 }
4743 #ifdef _LP64
4744 push(rscratch1); // save r10, trashed by movptr()
4745 #endif
4746 push(rax); // save rax,
4747 // addr may contain rsp so we will have to adjust it based on the push
4748 // we just did (and on 64 bit we do two pushes)
4749 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which
4750 // stores rax into addr which is backwards of what was intended.
4751 if (addr.uses(rsp)) {
4752 lea(rax, addr);
4753 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord));
4754 } else {
4755 pushptr(addr);
4756 }
4757
4758 ExternalAddress buffer((address) b);
4759 // pass msg argument
4760 // avoid using pushptr, as it modifies scratch registers
4761 // and our contract is not to modify anything
4762 movptr(rax, buffer.addr());
4763 push(rax);
4764
4765 // call indirectly to solve generation ordering problem
4766 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
4767 call(rax);
4768 // Caller pops the arguments (addr, message) and restores rax, r10.
4769 }
4770
4771 void MacroAssembler::verify_tlab() {
4772 #ifdef ASSERT
4773 if (UseTLAB && VerifyOops) {
4774 Label next, ok;
4775 Register t1 = rsi;
4776 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread);
4777
4778 push(t1);
4779 NOT_LP64(push(thread_reg));
4780 NOT_LP64(get_thread(thread_reg));
4781
4782 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
4783 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
4784 jcc(Assembler::aboveEqual, next);
4785 STOP("assert(top >= start)");
4786 should_not_reach_here();
4787
4788 bind(next);
4789 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
4790 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
4791 jcc(Assembler::aboveEqual, ok);
4792 STOP("assert(top <= end)");
4793 should_not_reach_here();
4794
4795 bind(ok);
4796 NOT_LP64(pop(thread_reg));
4797 pop(t1);
4798 }
4799 #endif
4800 }
4801
4802 class ControlWord {
4803 public:
4804 int32_t _value;
4805
4806 int rounding_control() const { return (_value >> 10) & 3 ; }
4807 int precision_control() const { return (_value >> 8) & 3 ; }
4808 bool precision() const { return ((_value >> 5) & 1) != 0; }
4809 bool underflow() const { return ((_value >> 4) & 1) != 0; }
4810 bool overflow() const { return ((_value >> 3) & 1) != 0; }
4811 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
4812 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
4813 bool invalid() const { return ((_value >> 0) & 1) != 0; }
4814
4815 void print() const {
4816 // rounding control
4817 const char* rc;
4818 switch (rounding_control()) {
4819 case 0: rc = "round near"; break;
4820 case 1: rc = "round down"; break;
4821 case 2: rc = "round up "; break;
4822 case 3: rc = "chop "; break;
4823 };
4824 // precision control
4825 const char* pc;
4826 switch (precision_control()) {
4827 case 0: pc = "24 bits "; break;
4828 case 1: pc = "reserved"; break;
4829 case 2: pc = "53 bits "; break;
4830 case 3: pc = "64 bits "; break;
4831 };
4832 // flags
4833 char f[9];
4834 f[0] = ' ';
4835 f[1] = ' ';
4836 f[2] = (precision ()) ? 'P' : 'p';
4837 f[3] = (underflow ()) ? 'U' : 'u';
4838 f[4] = (overflow ()) ? 'O' : 'o';
4839 f[5] = (zero_divide ()) ? 'Z' : 'z';
4840 f[6] = (denormalized()) ? 'D' : 'd';
4841 f[7] = (invalid ()) ? 'I' : 'i';
4842 f[8] = '\x0';
4843 // output
4844 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc);
4845 }
4846
4847 };
4848
4849 class StatusWord {
4850 public:
4851 int32_t _value;
4852
4853 bool busy() const { return ((_value >> 15) & 1) != 0; }
4854 bool C3() const { return ((_value >> 14) & 1) != 0; }
4855 bool C2() const { return ((_value >> 10) & 1) != 0; }
4856 bool C1() const { return ((_value >> 9) & 1) != 0; }
4857 bool C0() const { return ((_value >> 8) & 1) != 0; }
4858 int top() const { return (_value >> 11) & 7 ; }
4859 bool error_status() const { return ((_value >> 7) & 1) != 0; }
4860 bool stack_fault() const { return ((_value >> 6) & 1) != 0; }
4861 bool precision() const { return ((_value >> 5) & 1) != 0; }
4862 bool underflow() const { return ((_value >> 4) & 1) != 0; }
4863 bool overflow() const { return ((_value >> 3) & 1) != 0; }
4864 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
4865 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
4866 bool invalid() const { return ((_value >> 0) & 1) != 0; }
4867
4868 void print() const {
4869 // condition codes
4870 char c[5];
4871 c[0] = (C3()) ? '3' : '-';
4872 c[1] = (C2()) ? '2' : '-';
4873 c[2] = (C1()) ? '1' : '-';
4874 c[3] = (C0()) ? '0' : '-';
4875 c[4] = '\x0';
4876 // flags
4877 char f[9];
4878 f[0] = (error_status()) ? 'E' : '-';
4879 f[1] = (stack_fault ()) ? 'S' : '-';
4880 f[2] = (precision ()) ? 'P' : '-';
4881 f[3] = (underflow ()) ? 'U' : '-';
4882 f[4] = (overflow ()) ? 'O' : '-';
4883 f[5] = (zero_divide ()) ? 'Z' : '-';
4884 f[6] = (denormalized()) ? 'D' : '-';
4885 f[7] = (invalid ()) ? 'I' : '-';
4886 f[8] = '\x0';
4887 // output
4888 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top());
4889 }
4890
4891 };
4892
4893 class TagWord {
4894 public:
4895 int32_t _value;
4896
4897 int tag_at(int i) const { return (_value >> (i*2)) & 3; }
4898
4899 void print() const {
4900 printf("%04x", _value & 0xFFFF);
4901 }
4902
4903 };
4904
4905 class FPU_Register {
4906 public:
4907 int32_t _m0;
4908 int32_t _m1;
4909 int16_t _ex;
4910
4911 bool is_indefinite() const {
4912 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0;
4913 }
4914
4915 void print() const {
4916 char sign = (_ex < 0) ? '-' : '+';
4917 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " ";
4918 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind);
4919 };
4920
4921 };
4922
4923 class FPU_State {
4924 public:
4925 enum {
4926 register_size = 10,
4927 number_of_registers = 8,
4928 register_mask = 7
4929 };
4930
4931 ControlWord _control_word;
4932 StatusWord _status_word;
4933 TagWord _tag_word;
4934 int32_t _error_offset;
4935 int32_t _error_selector;
4936 int32_t _data_offset;
4937 int32_t _data_selector;
4938 int8_t _register[register_size * number_of_registers];
4939
4940 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); }
4941 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; }
4942
4943 const char* tag_as_string(int tag) const {
4944 switch (tag) {
4945 case 0: return "valid";
4946 case 1: return "zero";
4947 case 2: return "special";
4948 case 3: return "empty";
4949 }
4950 ShouldNotReachHere();
4951 return NULL;
4952 }
4953
4954 void print() const {
4955 // print computation registers
4956 { int t = _status_word.top();
4957 for (int i = 0; i < number_of_registers; i++) {
4958 int j = (i - t) & register_mask;
4959 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j);
4960 st(j)->print();
4961 printf(" %s\n", tag_as_string(_tag_word.tag_at(i)));
4962 }
4963 }
4964 printf("\n");
4965 // print control registers
4966 printf("ctrl = "); _control_word.print(); printf("\n");
4967 printf("stat = "); _status_word .print(); printf("\n");
4968 printf("tags = "); _tag_word .print(); printf("\n");
4969 }
4970
4971 };
4972
4973 class Flag_Register {
4974 public:
4975 int32_t _value;
4976
4977 bool overflow() const { return ((_value >> 11) & 1) != 0; }
4978 bool direction() const { return ((_value >> 10) & 1) != 0; }
4979 bool sign() const { return ((_value >> 7) & 1) != 0; }
4980 bool zero() const { return ((_value >> 6) & 1) != 0; }
4981 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; }
4982 bool parity() const { return ((_value >> 2) & 1) != 0; }
4983 bool carry() const { return ((_value >> 0) & 1) != 0; }
4984
4985 void print() const {
4986 // flags
4987 char f[8];
4988 f[0] = (overflow ()) ? 'O' : '-';
4989 f[1] = (direction ()) ? 'D' : '-';
4990 f[2] = (sign ()) ? 'S' : '-';
4991 f[3] = (zero ()) ? 'Z' : '-';
4992 f[4] = (auxiliary_carry()) ? 'A' : '-';
4993 f[5] = (parity ()) ? 'P' : '-';
4994 f[6] = (carry ()) ? 'C' : '-';
4995 f[7] = '\x0';
4996 // output
4997 printf("%08x flags = %s", _value, f);
4998 }
4999
5000 };
5001
5002 class IU_Register {
5003 public:
5004 int32_t _value;
5005
5006 void print() const {
5007 printf("%08x %11d", _value, _value);
5008 }
5009
5010 };
5011
5012 class IU_State {
5013 public:
5014 Flag_Register _eflags;
5015 IU_Register _rdi;
5016 IU_Register _rsi;
5017 IU_Register _rbp;
5018 IU_Register _rsp;
5019 IU_Register _rbx;
5020 IU_Register _rdx;
5021 IU_Register _rcx;
5022 IU_Register _rax;
5023
5024 void print() const {
5025 // computation registers
5026 printf("rax, = "); _rax.print(); printf("\n");
5027 printf("rbx, = "); _rbx.print(); printf("\n");
5028 printf("rcx = "); _rcx.print(); printf("\n");
5029 printf("rdx = "); _rdx.print(); printf("\n");
5030 printf("rdi = "); _rdi.print(); printf("\n");
5031 printf("rsi = "); _rsi.print(); printf("\n");
5032 printf("rbp, = "); _rbp.print(); printf("\n");
5033 printf("rsp = "); _rsp.print(); printf("\n");
5034 printf("\n");
5035 // control registers
5036 printf("flgs = "); _eflags.print(); printf("\n");
5037 }
5038 };
5039
5040
5041 class CPU_State {
5042 public:
5043 FPU_State _fpu_state;
5044 IU_State _iu_state;
5045
5046 void print() const {
5047 printf("--------------------------------------------------\n");
5048 _iu_state .print();
5049 printf("\n");
5050 _fpu_state.print();
5051 printf("--------------------------------------------------\n");
5052 }
5053
5054 };
5055
5056
5057 static void _print_CPU_state(CPU_State* state) {
5058 state->print();
5059 };
5060
5061
5062 void MacroAssembler::print_CPU_state() {
5063 push_CPU_state();
5064 push(rsp); // pass CPU state
5065 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state)));
5066 addptr(rsp, wordSize); // discard argument
5067 pop_CPU_state();
5068 }
5069
5070
5071 #ifndef _LP64
5072 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) {
5073 static int counter = 0;
5074 FPU_State* fs = &state->_fpu_state;
5075 counter++;
5076 // For leaf calls, only verify that the top few elements remain empty.
5077 // We only need 1 empty at the top for C2 code.
5078 if( stack_depth < 0 ) {
5079 if( fs->tag_for_st(7) != 3 ) {
5080 printf("FPR7 not empty\n");
5081 state->print();
5082 assert(false, "error");
5083 return false;
5084 }
5085 return true; // All other stack states do not matter
5086 }
5087
5088 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std,
5089 "bad FPU control word");
5090
5091 // compute stack depth
5092 int i = 0;
5093 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++;
5094 int d = i;
5095 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++;
5096 // verify findings
5097 if (i != FPU_State::number_of_registers) {
5098 // stack not contiguous
5099 printf("%s: stack not contiguous at ST%d\n", s, i);
5100 state->print();
5101 assert(false, "error");
5102 return false;
5103 }
5104 // check if computed stack depth corresponds to expected stack depth
5105 if (stack_depth < 0) {
5106 // expected stack depth is -stack_depth or less
5107 if (d > -stack_depth) {
5108 // too many elements on the stack
5109 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d);
5110 state->print();
5111 assert(false, "error");
5112 return false;
5113 }
5114 } else {
5115 // expected stack depth is stack_depth
5116 if (d != stack_depth) {
5117 // wrong stack depth
5118 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d);
5119 state->print();
5120 assert(false, "error");
5121 return false;
5122 }
5123 }
5124 // everything is cool
5125 return true;
5126 }
5127
5128 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
5129 if (!VerifyFPU) return;
5130 push_CPU_state();
5131 push(rsp); // pass CPU state
5132 ExternalAddress msg((address) s);
5133 // pass message string s
5134 pushptr(msg.addr());
5135 push(stack_depth); // pass stack depth
5136 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU)));
5137 addptr(rsp, 3 * wordSize); // discard arguments
5138 // check for error
5139 { Label L;
5140 testl(rax, rax);
5141 jcc(Assembler::notZero, L);
5142 int3(); // break if error condition
5143 bind(L);
5144 }
5145 pop_CPU_state();
5146 }
5147 #endif // _LP64
5148
5149 void MacroAssembler::restore_cpu_control_state_after_jni() {
5150 // Either restore the MXCSR register after returning from the JNI Call
5151 // or verify that it wasn't changed (with -Xcheck:jni flag).
5152 if (VM_Version::supports_sse()) {
5153 if (RestoreMXCSROnJNICalls) {
5154 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
5155 } else if (CheckJNICalls) {
5156 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
5157 }
5158 }
5159 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty.
5160 vzeroupper();
5161 // Reset k1 to 0xffff.
5162
5163 #ifdef COMPILER2
5164 if (PostLoopMultiversioning && VM_Version::supports_evex()) {
5165 push(rcx);
5166 movl(rcx, 0xffff);
5167 kmovwl(k1, rcx);
5168 pop(rcx);
5169 }
5170 #endif // COMPILER2
5171
5172 #ifndef _LP64
5173 // Either restore the x87 floating pointer control word after returning
5174 // from the JNI call or verify that it wasn't changed.
5175 if (CheckJNICalls) {
5176 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
5177 }
5178 #endif // _LP64
5179 }
5180
5181 // ((OopHandle)result).resolve();
5182 void MacroAssembler::resolve_oop_handle(Register result, Register tmp) {
5183 assert_different_registers(result, tmp);
5184
5185 // Only 64 bit platforms support GCs that require a tmp register
5186 // Only IN_HEAP loads require a thread_tmp register
5187 // OopHandle::resolve is an indirection like jobject.
5188 access_load_at(T_OBJECT, IN_NATIVE,
5189 result, Address(result, 0), tmp, /*tmp_thread*/noreg);
5190 }
5191
5192 // ((WeakHandle)result).resolve();
5193 void MacroAssembler::resolve_weak_handle(Register rresult, Register rtmp) {
5194 assert_different_registers(rresult, rtmp);
5195 Label resolved;
5196
5197 // A null weak handle resolves to null.
5198 cmpptr(rresult, 0);
5199 jcc(Assembler::equal, resolved);
5200
5201 // Only 64 bit platforms support GCs that require a tmp register
5202 // Only IN_HEAP loads require a thread_tmp register
5203 // WeakHandle::resolve is an indirection like jweak.
5204 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
5205 rresult, Address(rresult, 0), rtmp, /*tmp_thread*/noreg);
5206 bind(resolved);
5207 }
5208
5209 void MacroAssembler::load_mirror(Register mirror, Register method, Register tmp) {
5210 // get mirror
5211 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
5212 load_method_holder(mirror, method);
5213 movptr(mirror, Address(mirror, mirror_offset));
5214 resolve_oop_handle(mirror, tmp);
5215 }
5216
5217 void MacroAssembler::load_method_holder_cld(Register rresult, Register rmethod) {
5218 load_method_holder(rresult, rmethod);
5219 movptr(rresult, Address(rresult, InstanceKlass::class_loader_data_offset()));
5220 }
5221
5222 void MacroAssembler::load_method_holder(Register holder, Register method) {
5223 movptr(holder, Address(method, Method::const_offset())); // ConstMethod*
5224 movptr(holder, Address(holder, ConstMethod::constants_offset())); // ConstantPool*
5225 movptr(holder, Address(holder, ConstantPool::pool_holder_offset_in_bytes())); // InstanceKlass*
5226 }
5227
5228 void MacroAssembler::load_klass(Register dst, Register src) {
5229 #ifdef _LP64
5230 if (UseCompressedClassPointers) {
5231 movl(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5232 decode_klass_not_null(dst);
5233 } else
5234 #endif
5235 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5236 }
5237
5238 void MacroAssembler::load_prototype_header(Register dst, Register src) {
5239 load_klass(dst, src);
5240 movptr(dst, Address(dst, Klass::prototype_header_offset()));
5241 }
5242
5243 void MacroAssembler::store_klass(Register dst, Register src) {
5244 #ifdef _LP64
5245 if (UseCompressedClassPointers) {
5246 encode_klass_not_null(src);
5247 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5248 } else
5249 #endif
5250 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5251 }
5252
5253 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators, Register dst, Address src,
5254 Register tmp1, Register thread_tmp) {
5255 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5256 decorators = AccessInternal::decorator_fixup(decorators);
5257 bool as_raw = (decorators & AS_RAW) != 0;
5258 if (as_raw) {
5259 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
5260 } else {
5261 bs->load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
5262 }
5263 }
5264
5265 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators, Address dst, Register src,
5266 Register tmp1, Register tmp2) {
5267 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5268 decorators = AccessInternal::decorator_fixup(decorators);
5269 bool as_raw = (decorators & AS_RAW) != 0;
5270 if (as_raw) {
5271 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, src, tmp1, tmp2);
5272 } else {
5273 bs->store_at(this, decorators, type, dst, src, tmp1, tmp2);
5274 }
5275 }
5276
5277 void MacroAssembler::resolve(DecoratorSet decorators, Register obj) {
5278 // Use stronger ACCESS_WRITE|ACCESS_READ by default.
5279 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
5280 decorators |= ACCESS_READ | ACCESS_WRITE;
5281 }
5282 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5283 return bs->resolve(this, decorators, obj);
5284 }
5285
5286 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
5287 Register thread_tmp, DecoratorSet decorators) {
5288 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp);
5289 }
5290
5291 // Doesn't do verfication, generates fixed size code
5292 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1,
5293 Register thread_tmp, DecoratorSet decorators) {
5294 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, thread_tmp);
5295 }
5296
5297 void MacroAssembler::store_heap_oop(Address dst, Register src, Register tmp1,
5298 Register tmp2, DecoratorSet decorators) {
5299 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2);
5300 }
5301
5302 // Used for storing NULLs.
5303 void MacroAssembler::store_heap_oop_null(Address dst) {
5304 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg);
5305 }
5306
5307 #ifdef _LP64
5308 void MacroAssembler::store_klass_gap(Register dst, Register src) {
5309 if (UseCompressedClassPointers) {
5310 // Store to klass gap in destination
5311 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src);
5312 }
5313 }
5314
5315 #ifdef ASSERT
5316 void MacroAssembler::verify_heapbase(const char* msg) {
5317 assert (UseCompressedOops, "should be compressed");
5318 assert (Universe::heap() != NULL, "java heap should be initialized");
5319 if (CheckCompressedOops) {
5320 Label ok;
5321 push(rscratch1); // cmpptr trashes rscratch1
5322 cmpptr(r12_heapbase, ExternalAddress((address)CompressedOops::ptrs_base_addr()));
5323 jcc(Assembler::equal, ok);
5324 STOP(msg);
5325 bind(ok);
5326 pop(rscratch1);
5327 }
5328 }
5329 #endif
5330
5331 // Algorithm must match oop.inline.hpp encode_heap_oop.
5332 void MacroAssembler::encode_heap_oop(Register r) {
5333 #ifdef ASSERT
5334 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
5335 #endif
5336 verify_oop(r);
5337 if (CompressedOops::base() == NULL) {
5338 if (CompressedOops::shift() != 0) {
5339 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5340 shrq(r, LogMinObjAlignmentInBytes);
5341 }
5342 return;
5343 }
5344 testq(r, r);
5345 cmovq(Assembler::equal, r, r12_heapbase);
5346 subq(r, r12_heapbase);
5347 shrq(r, LogMinObjAlignmentInBytes);
5348 }
5349
5350 void MacroAssembler::encode_heap_oop_not_null(Register r) {
5351 #ifdef ASSERT
5352 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
5353 if (CheckCompressedOops) {
5354 Label ok;
5355 testq(r, r);
5356 jcc(Assembler::notEqual, ok);
5357 STOP("null oop passed to encode_heap_oop_not_null");
5358 bind(ok);
5359 }
5360 #endif
5361 verify_oop(r);
5362 if (CompressedOops::base() != NULL) {
5363 subq(r, r12_heapbase);
5364 }
5365 if (CompressedOops::shift() != 0) {
5366 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5367 shrq(r, LogMinObjAlignmentInBytes);
5368 }
5369 }
5370
5371 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
5372 #ifdef ASSERT
5373 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
5374 if (CheckCompressedOops) {
5375 Label ok;
5376 testq(src, src);
5377 jcc(Assembler::notEqual, ok);
5378 STOP("null oop passed to encode_heap_oop_not_null2");
5379 bind(ok);
5380 }
5381 #endif
5382 verify_oop(src);
5383 if (dst != src) {
5384 movq(dst, src);
5385 }
5386 if (CompressedOops::base() != NULL) {
5387 subq(dst, r12_heapbase);
5388 }
5389 if (CompressedOops::shift() != 0) {
5390 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5391 shrq(dst, LogMinObjAlignmentInBytes);
5392 }
5393 }
5394
5395 void MacroAssembler::decode_heap_oop(Register r) {
5396 #ifdef ASSERT
5397 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
5398 #endif
5399 if (CompressedOops::base() == NULL) {
5400 if (CompressedOops::shift() != 0) {
5401 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5402 shlq(r, LogMinObjAlignmentInBytes);
5403 }
5404 } else {
5405 Label done;
5406 shlq(r, LogMinObjAlignmentInBytes);
5407 jccb(Assembler::equal, done);
5408 addq(r, r12_heapbase);
5409 bind(done);
5410 }
5411 verify_oop(r);
5412 }
5413
5414 void MacroAssembler::decode_heap_oop_not_null(Register r) {
5415 // Note: it will change flags
5416 assert (UseCompressedOops, "should only be used for compressed headers");
5417 assert (Universe::heap() != NULL, "java heap should be initialized");
5418 // Cannot assert, unverified entry point counts instructions (see .ad file)
5419 // vtableStubs also counts instructions in pd_code_size_limit.
5420 // Also do not verify_oop as this is called by verify_oop.
5421 if (CompressedOops::shift() != 0) {
5422 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5423 shlq(r, LogMinObjAlignmentInBytes);
5424 if (CompressedOops::base() != NULL) {
5425 addq(r, r12_heapbase);
5426 }
5427 } else {
5428 assert (CompressedOops::base() == NULL, "sanity");
5429 }
5430 }
5431
5432 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
5433 // Note: it will change flags
5434 assert (UseCompressedOops, "should only be used for compressed headers");
5435 assert (Universe::heap() != NULL, "java heap should be initialized");
5436 // Cannot assert, unverified entry point counts instructions (see .ad file)
5437 // vtableStubs also counts instructions in pd_code_size_limit.
5438 // Also do not verify_oop as this is called by verify_oop.
5439 if (CompressedOops::shift() != 0) {
5440 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5441 if (LogMinObjAlignmentInBytes == Address::times_8) {
5442 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0));
5443 } else {
5444 if (dst != src) {
5445 movq(dst, src);
5446 }
5447 shlq(dst, LogMinObjAlignmentInBytes);
5448 if (CompressedOops::base() != NULL) {
5449 addq(dst, r12_heapbase);
5450 }
5451 }
5452 } else {
5453 assert (CompressedOops::base() == NULL, "sanity");
5454 if (dst != src) {
5455 movq(dst, src);
5456 }
5457 }
5458 }
5459
5460 void MacroAssembler::encode_klass_not_null(Register r) {
5461 if (CompressedKlassPointers::base() != NULL) {
5462 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5463 assert(r != r12_heapbase, "Encoding a klass in r12");
5464 mov64(r12_heapbase, (int64_t)CompressedKlassPointers::base());
5465 subq(r, r12_heapbase);
5466 }
5467 if (CompressedKlassPointers::shift() != 0) {
5468 assert (LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong");
5469 shrq(r, LogKlassAlignmentInBytes);
5470 }
5471 if (CompressedKlassPointers::base() != NULL) {
5472 reinit_heapbase();
5473 }
5474 }
5475
5476 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
5477 if (dst == src) {
5478 encode_klass_not_null(src);
5479 } else {
5480 if (CompressedKlassPointers::base() != NULL) {
5481 mov64(dst, (int64_t)CompressedKlassPointers::base());
5482 negq(dst);
5483 addq(dst, src);
5484 } else {
5485 movptr(dst, src);
5486 }
5487 if (CompressedKlassPointers::shift() != 0) {
5488 assert (LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong");
5489 shrq(dst, LogKlassAlignmentInBytes);
5490 }
5491 }
5492 }
5493
5494 // Function instr_size_for_decode_klass_not_null() counts the instructions
5495 // generated by decode_klass_not_null(register r) and reinit_heapbase(),
5496 // when (Universe::heap() != NULL). Hence, if the instructions they
5497 // generate change, then this method needs to be updated.
5498 int MacroAssembler::instr_size_for_decode_klass_not_null() {
5499 assert (UseCompressedClassPointers, "only for compressed klass ptrs");
5500 if (CompressedKlassPointers::base() != NULL) {
5501 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()).
5502 return (CompressedKlassPointers::shift() == 0 ? 20 : 24);
5503 } else {
5504 // longest load decode klass function, mov64, leaq
5505 return 16;
5506 }
5507 }
5508
5509 // !!! If the instructions that get generated here change then function
5510 // instr_size_for_decode_klass_not_null() needs to get updated.
5511 void MacroAssembler::decode_klass_not_null(Register r) {
5512 // Note: it will change flags
5513 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5514 assert(r != r12_heapbase, "Decoding a klass in r12");
5515 // Cannot assert, unverified entry point counts instructions (see .ad file)
5516 // vtableStubs also counts instructions in pd_code_size_limit.
5517 // Also do not verify_oop as this is called by verify_oop.
5518 if (CompressedKlassPointers::shift() != 0) {
5519 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong");
5520 shlq(r, LogKlassAlignmentInBytes);
5521 }
5522 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5523 if (CompressedKlassPointers::base() != NULL) {
5524 mov64(r12_heapbase, (int64_t)CompressedKlassPointers::base());
5525 addq(r, r12_heapbase);
5526 reinit_heapbase();
5527 }
5528 }
5529
5530 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
5531 // Note: it will change flags
5532 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5533 if (dst == src) {
5534 decode_klass_not_null(dst);
5535 } else {
5536 // Cannot assert, unverified entry point counts instructions (see .ad file)
5537 // vtableStubs also counts instructions in pd_code_size_limit.
5538 // Also do not verify_oop as this is called by verify_oop.
5539 mov64(dst, (int64_t)CompressedKlassPointers::base());
5540 if (CompressedKlassPointers::shift() != 0) {
5541 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong");
5542 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?");
5543 leaq(dst, Address(dst, src, Address::times_8, 0));
5544 } else {
5545 addq(dst, src);
5546 }
5547 }
5548 }
5549
5550 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
5551 assert (UseCompressedOops, "should only be used for compressed headers");
5552 assert (Universe::heap() != NULL, "java heap should be initialized");
5553 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5554 int oop_index = oop_recorder()->find_index(obj);
5555 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5556 mov_narrow_oop(dst, oop_index, rspec);
5557 }
5558
5559 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) {
5560 assert (UseCompressedOops, "should only be used for compressed headers");
5561 assert (Universe::heap() != NULL, "java heap should be initialized");
5562 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5563 int oop_index = oop_recorder()->find_index(obj);
5564 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5565 mov_narrow_oop(dst, oop_index, rspec);
5566 }
5567
5568 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
5569 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5570 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5571 int klass_index = oop_recorder()->find_index(k);
5572 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5573 mov_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec);
5574 }
5575
5576 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) {
5577 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5578 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5579 int klass_index = oop_recorder()->find_index(k);
5580 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5581 mov_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec);
5582 }
5583
5584 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) {
5585 assert (UseCompressedOops, "should only be used for compressed headers");
5586 assert (Universe::heap() != NULL, "java heap should be initialized");
5587 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5588 int oop_index = oop_recorder()->find_index(obj);
5589 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5590 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
5591 }
5592
5593 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) {
5594 assert (UseCompressedOops, "should only be used for compressed headers");
5595 assert (Universe::heap() != NULL, "java heap should be initialized");
5596 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5597 int oop_index = oop_recorder()->find_index(obj);
5598 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5599 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
5600 }
5601
5602 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) {
5603 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5604 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5605 int klass_index = oop_recorder()->find_index(k);
5606 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5607 Assembler::cmp_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec);
5608 }
5609
5610 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) {
5611 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5612 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5613 int klass_index = oop_recorder()->find_index(k);
5614 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5615 Assembler::cmp_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec);
5616 }
5617
5618 void MacroAssembler::reinit_heapbase() {
5619 if (UseCompressedOops || UseCompressedClassPointers) {
5620 if (Universe::heap() != NULL) {
5621 if (CompressedOops::base() == NULL) {
5622 MacroAssembler::xorptr(r12_heapbase, r12_heapbase);
5623 } else {
5624 mov64(r12_heapbase, (int64_t)CompressedOops::ptrs_base());
5625 }
5626 } else {
5627 movptr(r12_heapbase, ExternalAddress((address)CompressedOops::ptrs_base_addr()));
5628 }
5629 }
5630 }
5631
5632 #endif // _LP64
5633
5634 // C2 compiled method's prolog code.
5635 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b, bool is_stub) {
5636
5637 // WARNING: Initial instruction MUST be 5 bytes or longer so that
5638 // NativeJump::patch_verified_entry will be able to patch out the entry
5639 // code safely. The push to verify stack depth is ok at 5 bytes,
5640 // the frame allocation can be either 3 or 6 bytes. So if we don't do
5641 // stack bang then we must use the 6 byte frame allocation even if
5642 // we have no frame. :-(
5643 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect");
5644
5645 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
5646 // Remove word for return addr
5647 framesize -= wordSize;
5648 stack_bang_size -= wordSize;
5649
5650 // Calls to C2R adapters often do not accept exceptional returns.
5651 // We require that their callers must bang for them. But be careful, because
5652 // some VM calls (such as call site linkage) can use several kilobytes of
5653 // stack. But the stack safety zone should account for that.
5654 // See bugs 4446381, 4468289, 4497237.
5655 if (stack_bang_size > 0) {
5656 generate_stack_overflow_check(stack_bang_size);
5657
5658 // We always push rbp, so that on return to interpreter rbp, will be
5659 // restored correctly and we can correct the stack.
5660 push(rbp);
5661 // Save caller's stack pointer into RBP if the frame pointer is preserved.
5662 if (PreserveFramePointer) {
5663 mov(rbp, rsp);
5664 }
5665 // Remove word for ebp
5666 framesize -= wordSize;
5667
5668 // Create frame
5669 if (framesize) {
5670 subptr(rsp, framesize);
5671 }
5672 } else {
5673 // Create frame (force generation of a 4 byte immediate value)
5674 subptr_imm32(rsp, framesize);
5675
5676 // Save RBP register now.
5677 framesize -= wordSize;
5678 movptr(Address(rsp, framesize), rbp);
5679 // Save caller's stack pointer into RBP if the frame pointer is preserved.
5680 if (PreserveFramePointer) {
5681 movptr(rbp, rsp);
5682 if (framesize > 0) {
5683 addptr(rbp, framesize);
5684 }
5685 }
5686 }
5687
5688 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
5689 framesize -= wordSize;
5690 movptr(Address(rsp, framesize), (int32_t)0xbadb100d);
5691 }
5692
5693 #ifndef _LP64
5694 // If method sets FPU control word do it now
5695 if (fp_mode_24b) {
5696 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
5697 }
5698 if (UseSSE >= 2 && VerifyFPU) {
5699 verify_FPU(0, "FPU stack must be clean on entry");
5700 }
5701 #endif
5702
5703 #ifdef ASSERT
5704 if (VerifyStackAtCalls) {
5705 Label L;
5706 push(rax);
5707 mov(rax, rsp);
5708 andptr(rax, StackAlignmentInBytes-1);
5709 cmpptr(rax, StackAlignmentInBytes-wordSize);
5710 pop(rax);
5711 jcc(Assembler::equal, L);
5712 STOP("Stack is not properly aligned!");
5713 bind(L);
5714 }
5715 #endif
5716
5717 if (!is_stub) {
5718 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5719 bs->nmethod_entry_barrier(this);
5720 }
5721 }
5722
5723 // clear memory of size 'cnt' qwords, starting at 'base' using XMM/YMM registers
5724 void MacroAssembler::xmm_clear_mem(Register base, Register cnt, XMMRegister xtmp) {
5725 // cnt - number of qwords (8-byte words).
5726 // base - start address, qword aligned.
5727 Label L_zero_64_bytes, L_loop, L_sloop, L_tail, L_end;
5728 if (UseAVX >= 2) {
5729 vpxor(xtmp, xtmp, xtmp, AVX_256bit);
5730 } else {
5731 pxor(xtmp, xtmp);
5732 }
5733 jmp(L_zero_64_bytes);
5734
5735 BIND(L_loop);
5736 if (UseAVX >= 2) {
5737 vmovdqu(Address(base, 0), xtmp);
5738 vmovdqu(Address(base, 32), xtmp);
5739 } else {
5740 movdqu(Address(base, 0), xtmp);
5741 movdqu(Address(base, 16), xtmp);
5742 movdqu(Address(base, 32), xtmp);
5743 movdqu(Address(base, 48), xtmp);
5744 }
5745 addptr(base, 64);
5746
5747 BIND(L_zero_64_bytes);
5748 subptr(cnt, 8);
5749 jccb(Assembler::greaterEqual, L_loop);
5750 addptr(cnt, 4);
5751 jccb(Assembler::less, L_tail);
5752 // Copy trailing 32 bytes
5753 if (UseAVX >= 2) {
5754 vmovdqu(Address(base, 0), xtmp);
5755 } else {
5756 movdqu(Address(base, 0), xtmp);
5757 movdqu(Address(base, 16), xtmp);
5758 }
5759 addptr(base, 32);
5760 subptr(cnt, 4);
5761
5762 BIND(L_tail);
5763 addptr(cnt, 4);
5764 jccb(Assembler::lessEqual, L_end);
5765 decrement(cnt);
5766
5767 BIND(L_sloop);
5768 movq(Address(base, 0), xtmp);
5769 addptr(base, 8);
5770 decrement(cnt);
5771 jccb(Assembler::greaterEqual, L_sloop);
5772 BIND(L_end);
5773 }
5774
5775 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp, XMMRegister xtmp, bool is_large) {
5776 // cnt - number of qwords (8-byte words).
5777 // base - start address, qword aligned.
5778 // is_large - if optimizers know cnt is larger than InitArrayShortSize
5779 assert(base==rdi, "base register must be edi for rep stos");
5780 assert(tmp==rax, "tmp register must be eax for rep stos");
5781 assert(cnt==rcx, "cnt register must be ecx for rep stos");
5782 assert(InitArrayShortSize % BytesPerLong == 0,
5783 "InitArrayShortSize should be the multiple of BytesPerLong");
5784
5785 Label DONE;
5786
5787 if (!is_large || !UseXMMForObjInit) {
5788 xorptr(tmp, tmp);
5789 }
5790
5791 if (!is_large) {
5792 Label LOOP, LONG;
5793 cmpptr(cnt, InitArrayShortSize/BytesPerLong);
5794 jccb(Assembler::greater, LONG);
5795
5796 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
5797
5798 decrement(cnt);
5799 jccb(Assembler::negative, DONE); // Zero length
5800
5801 // Use individual pointer-sized stores for small counts:
5802 BIND(LOOP);
5803 movptr(Address(base, cnt, Address::times_ptr), tmp);
5804 decrement(cnt);
5805 jccb(Assembler::greaterEqual, LOOP);
5806 jmpb(DONE);
5807
5808 BIND(LONG);
5809 }
5810
5811 // Use longer rep-prefixed ops for non-small counts:
5812 if (UseFastStosb) {
5813 shlptr(cnt, 3); // convert to number of bytes
5814 rep_stosb();
5815 } else if (UseXMMForObjInit) {
5816 movptr(tmp, base);
5817 xmm_clear_mem(tmp, cnt, xtmp);
5818 } else {
5819 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
5820 rep_stos();
5821 }
5822
5823 BIND(DONE);
5824 }
5825
5826 #ifdef COMPILER2
5827
5828 // IndexOf for constant substrings with size >= 8 chars
5829 // which don't need to be loaded through stack.
5830 void MacroAssembler::string_indexofC8(Register str1, Register str2,
5831 Register cnt1, Register cnt2,
5832 int int_cnt2, Register result,
5833 XMMRegister vec, Register tmp,
5834 int ae) {
5835 ShortBranchVerifier sbv(this);
5836 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
5837 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
5838
5839 // This method uses the pcmpestri instruction with bound registers
5840 // inputs:
5841 // xmm - substring
5842 // rax - substring length (elements count)
5843 // mem - scanned string
5844 // rdx - string length (elements count)
5845 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
5846 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
5847 // outputs:
5848 // rcx - matched index in string
5849 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
5850 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
5851 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
5852 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
5853 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
5854
5855 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR,
5856 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR,
5857 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE;
5858
5859 // Note, inline_string_indexOf() generates checks:
5860 // if (substr.count > string.count) return -1;
5861 // if (substr.count == 0) return 0;
5862 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars");
5863
5864 // Load substring.
5865 if (ae == StrIntrinsicNode::UL) {
5866 pmovzxbw(vec, Address(str2, 0));
5867 } else {
5868 movdqu(vec, Address(str2, 0));
5869 }
5870 movl(cnt2, int_cnt2);
5871 movptr(result, str1); // string addr
5872
5873 if (int_cnt2 > stride) {
5874 jmpb(SCAN_TO_SUBSTR);
5875
5876 // Reload substr for rescan, this code
5877 // is executed only for large substrings (> 8 chars)
5878 bind(RELOAD_SUBSTR);
5879 if (ae == StrIntrinsicNode::UL) {
5880 pmovzxbw(vec, Address(str2, 0));
5881 } else {
5882 movdqu(vec, Address(str2, 0));
5883 }
5884 negptr(cnt2); // Jumped here with negative cnt2, convert to positive
5885
5886 bind(RELOAD_STR);
5887 // We came here after the beginning of the substring was
5888 // matched but the rest of it was not so we need to search
5889 // again. Start from the next element after the previous match.
5890
5891 // cnt2 is number of substring reminding elements and
5892 // cnt1 is number of string reminding elements when cmp failed.
5893 // Restored cnt1 = cnt1 - cnt2 + int_cnt2
5894 subl(cnt1, cnt2);
5895 addl(cnt1, int_cnt2);
5896 movl(cnt2, int_cnt2); // Now restore cnt2
5897
5898 decrementl(cnt1); // Shift to next element
5899 cmpl(cnt1, cnt2);
5900 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring
5901
5902 addptr(result, (1<<scale1));
5903
5904 } // (int_cnt2 > 8)
5905
5906 // Scan string for start of substr in 16-byte vectors
5907 bind(SCAN_TO_SUBSTR);
5908 pcmpestri(vec, Address(result, 0), mode);
5909 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
5910 subl(cnt1, stride);
5911 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
5912 cmpl(cnt1, cnt2);
5913 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
5914 addptr(result, 16);
5915 jmpb(SCAN_TO_SUBSTR);
5916
5917 // Found a potential substr
5918 bind(FOUND_CANDIDATE);
5919 // Matched whole vector if first element matched (tmp(rcx) == 0).
5920 if (int_cnt2 == stride) {
5921 jccb(Assembler::overflow, RET_FOUND); // OF == 1
5922 } else { // int_cnt2 > 8
5923 jccb(Assembler::overflow, FOUND_SUBSTR);
5924 }
5925 // After pcmpestri tmp(rcx) contains matched element index
5926 // Compute start addr of substr
5927 lea(result, Address(result, tmp, scale1));
5928
5929 // Make sure string is still long enough
5930 subl(cnt1, tmp);
5931 cmpl(cnt1, cnt2);
5932 if (int_cnt2 == stride) {
5933 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
5934 } else { // int_cnt2 > 8
5935 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD);
5936 }
5937 // Left less then substring.
5938
5939 bind(RET_NOT_FOUND);
5940 movl(result, -1);
5941 jmp(EXIT);
5942
5943 if (int_cnt2 > stride) {
5944 // This code is optimized for the case when whole substring
5945 // is matched if its head is matched.
5946 bind(MATCH_SUBSTR_HEAD);
5947 pcmpestri(vec, Address(result, 0), mode);
5948 // Reload only string if does not match
5949 jcc(Assembler::noOverflow, RELOAD_STR); // OF == 0
5950
5951 Label CONT_SCAN_SUBSTR;
5952 // Compare the rest of substring (> 8 chars).
5953 bind(FOUND_SUBSTR);
5954 // First 8 chars are already matched.
5955 negptr(cnt2);
5956 addptr(cnt2, stride);
5957
5958 bind(SCAN_SUBSTR);
5959 subl(cnt1, stride);
5960 cmpl(cnt2, -stride); // Do not read beyond substring
5961 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR);
5962 // Back-up strings to avoid reading beyond substring:
5963 // cnt1 = cnt1 - cnt2 + 8
5964 addl(cnt1, cnt2); // cnt2 is negative
5965 addl(cnt1, stride);
5966 movl(cnt2, stride); negptr(cnt2);
5967 bind(CONT_SCAN_SUBSTR);
5968 if (int_cnt2 < (int)G) {
5969 int tail_off1 = int_cnt2<<scale1;
5970 int tail_off2 = int_cnt2<<scale2;
5971 if (ae == StrIntrinsicNode::UL) {
5972 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2));
5973 } else {
5974 movdqu(vec, Address(str2, cnt2, scale2, tail_off2));
5975 }
5976 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode);
5977 } else {
5978 // calculate index in register to avoid integer overflow (int_cnt2*2)
5979 movl(tmp, int_cnt2);
5980 addptr(tmp, cnt2);
5981 if (ae == StrIntrinsicNode::UL) {
5982 pmovzxbw(vec, Address(str2, tmp, scale2, 0));
5983 } else {
5984 movdqu(vec, Address(str2, tmp, scale2, 0));
5985 }
5986 pcmpestri(vec, Address(result, tmp, scale1, 0), mode);
5987 }
5988 // Need to reload strings pointers if not matched whole vector
5989 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
5990 addptr(cnt2, stride);
5991 jcc(Assembler::negative, SCAN_SUBSTR);
5992 // Fall through if found full substring
5993
5994 } // (int_cnt2 > 8)
5995
5996 bind(RET_FOUND);
5997 // Found result if we matched full small substring.
5998 // Compute substr offset
5999 subptr(result, str1);
6000 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6001 shrl(result, 1); // index
6002 }
6003 bind(EXIT);
6004
6005 } // string_indexofC8
6006
6007 // Small strings are loaded through stack if they cross page boundary.
6008 void MacroAssembler::string_indexof(Register str1, Register str2,
6009 Register cnt1, Register cnt2,
6010 int int_cnt2, Register result,
6011 XMMRegister vec, Register tmp,
6012 int ae) {
6013 ShortBranchVerifier sbv(this);
6014 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
6015 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
6016
6017 //
6018 // int_cnt2 is length of small (< 8 chars) constant substring
6019 // or (-1) for non constant substring in which case its length
6020 // is in cnt2 register.
6021 //
6022 // Note, inline_string_indexOf() generates checks:
6023 // if (substr.count > string.count) return -1;
6024 // if (substr.count == 0) return 0;
6025 //
6026 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
6027 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0");
6028 // This method uses the pcmpestri instruction with bound registers
6029 // inputs:
6030 // xmm - substring
6031 // rax - substring length (elements count)
6032 // mem - scanned string
6033 // rdx - string length (elements count)
6034 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6035 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
6036 // outputs:
6037 // rcx - matched index in string
6038 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6039 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
6040 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
6041 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
6042
6043 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR,
6044 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR,
6045 FOUND_CANDIDATE;
6046
6047 { //========================================================
6048 // We don't know where these strings are located
6049 // and we can't read beyond them. Load them through stack.
6050 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR;
6051
6052 movptr(tmp, rsp); // save old SP
6053
6054 if (int_cnt2 > 0) { // small (< 8 chars) constant substring
6055 if (int_cnt2 == (1>>scale2)) { // One byte
6056 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding");
6057 load_unsigned_byte(result, Address(str2, 0));
6058 movdl(vec, result); // move 32 bits
6059 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes
6060 // Not enough header space in 32-bit VM: 12+3 = 15.
6061 movl(result, Address(str2, -1));
6062 shrl(result, 8);
6063 movdl(vec, result); // move 32 bits
6064 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char
6065 load_unsigned_short(result, Address(str2, 0));
6066 movdl(vec, result); // move 32 bits
6067 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars
6068 movdl(vec, Address(str2, 0)); // move 32 bits
6069 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars
6070 movq(vec, Address(str2, 0)); // move 64 bits
6071 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7})
6072 // Array header size is 12 bytes in 32-bit VM
6073 // + 6 bytes for 3 chars == 18 bytes,
6074 // enough space to load vec and shift.
6075 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity");
6076 if (ae == StrIntrinsicNode::UL) {
6077 int tail_off = int_cnt2-8;
6078 pmovzxbw(vec, Address(str2, tail_off));
6079 psrldq(vec, -2*tail_off);
6080 }
6081 else {
6082 int tail_off = int_cnt2*(1<<scale2);
6083 movdqu(vec, Address(str2, tail_off-16));
6084 psrldq(vec, 16-tail_off);
6085 }
6086 }
6087 } else { // not constant substring
6088 cmpl(cnt2, stride);
6089 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough
6090
6091 // We can read beyond string if srt+16 does not cross page boundary
6092 // since heaps are aligned and mapped by pages.
6093 assert(os::vm_page_size() < (int)G, "default page should be small");
6094 movl(result, str2); // We need only low 32 bits
6095 andl(result, (os::vm_page_size()-1));
6096 cmpl(result, (os::vm_page_size()-16));
6097 jccb(Assembler::belowEqual, CHECK_STR);
6098
6099 // Move small strings to stack to allow load 16 bytes into vec.
6100 subptr(rsp, 16);
6101 int stk_offset = wordSize-(1<<scale2);
6102 push(cnt2);
6103
6104 bind(COPY_SUBSTR);
6105 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) {
6106 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1));
6107 movb(Address(rsp, cnt2, scale2, stk_offset), result);
6108 } else if (ae == StrIntrinsicNode::UU) {
6109 load_unsigned_short(result, Address(str2, cnt2, scale2, -2));
6110 movw(Address(rsp, cnt2, scale2, stk_offset), result);
6111 }
6112 decrement(cnt2);
6113 jccb(Assembler::notZero, COPY_SUBSTR);
6114
6115 pop(cnt2);
6116 movptr(str2, rsp); // New substring address
6117 } // non constant
6118
6119 bind(CHECK_STR);
6120 cmpl(cnt1, stride);
6121 jccb(Assembler::aboveEqual, BIG_STRINGS);
6122
6123 // Check cross page boundary.
6124 movl(result, str1); // We need only low 32 bits
6125 andl(result, (os::vm_page_size()-1));
6126 cmpl(result, (os::vm_page_size()-16));
6127 jccb(Assembler::belowEqual, BIG_STRINGS);
6128
6129 subptr(rsp, 16);
6130 int stk_offset = -(1<<scale1);
6131 if (int_cnt2 < 0) { // not constant
6132 push(cnt2);
6133 stk_offset += wordSize;
6134 }
6135 movl(cnt2, cnt1);
6136
6137 bind(COPY_STR);
6138 if (ae == StrIntrinsicNode::LL) {
6139 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1));
6140 movb(Address(rsp, cnt2, scale1, stk_offset), result);
6141 } else {
6142 load_unsigned_short(result, Address(str1, cnt2, scale1, -2));
6143 movw(Address(rsp, cnt2, scale1, stk_offset), result);
6144 }
6145 decrement(cnt2);
6146 jccb(Assembler::notZero, COPY_STR);
6147
6148 if (int_cnt2 < 0) { // not constant
6149 pop(cnt2);
6150 }
6151 movptr(str1, rsp); // New string address
6152
6153 bind(BIG_STRINGS);
6154 // Load substring.
6155 if (int_cnt2 < 0) { // -1
6156 if (ae == StrIntrinsicNode::UL) {
6157 pmovzxbw(vec, Address(str2, 0));
6158 } else {
6159 movdqu(vec, Address(str2, 0));
6160 }
6161 push(cnt2); // substr count
6162 push(str2); // substr addr
6163 push(str1); // string addr
6164 } else {
6165 // Small (< 8 chars) constant substrings are loaded already.
6166 movl(cnt2, int_cnt2);
6167 }
6168 push(tmp); // original SP
6169
6170 } // Finished loading
6171
6172 //========================================================
6173 // Start search
6174 //
6175
6176 movptr(result, str1); // string addr
6177
6178 if (int_cnt2 < 0) { // Only for non constant substring
6179 jmpb(SCAN_TO_SUBSTR);
6180
6181 // SP saved at sp+0
6182 // String saved at sp+1*wordSize
6183 // Substr saved at sp+2*wordSize
6184 // Substr count saved at sp+3*wordSize
6185
6186 // Reload substr for rescan, this code
6187 // is executed only for large substrings (> 8 chars)
6188 bind(RELOAD_SUBSTR);
6189 movptr(str2, Address(rsp, 2*wordSize));
6190 movl(cnt2, Address(rsp, 3*wordSize));
6191 if (ae == StrIntrinsicNode::UL) {
6192 pmovzxbw(vec, Address(str2, 0));
6193 } else {
6194 movdqu(vec, Address(str2, 0));
6195 }
6196 // We came here after the beginning of the substring was
6197 // matched but the rest of it was not so we need to search
6198 // again. Start from the next element after the previous match.
6199 subptr(str1, result); // Restore counter
6200 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6201 shrl(str1, 1);
6202 }
6203 addl(cnt1, str1);
6204 decrementl(cnt1); // Shift to next element
6205 cmpl(cnt1, cnt2);
6206 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6207
6208 addptr(result, (1<<scale1));
6209 } // non constant
6210
6211 // Scan string for start of substr in 16-byte vectors
6212 bind(SCAN_TO_SUBSTR);
6213 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6214 pcmpestri(vec, Address(result, 0), mode);
6215 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
6216 subl(cnt1, stride);
6217 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6218 cmpl(cnt1, cnt2);
6219 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6220 addptr(result, 16);
6221
6222 bind(ADJUST_STR);
6223 cmpl(cnt1, stride); // Do not read beyond string
6224 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6225 // Back-up string to avoid reading beyond string.
6226 lea(result, Address(result, cnt1, scale1, -16));
6227 movl(cnt1, stride);
6228 jmpb(SCAN_TO_SUBSTR);
6229
6230 // Found a potential substr
6231 bind(FOUND_CANDIDATE);
6232 // After pcmpestri tmp(rcx) contains matched element index
6233
6234 // Make sure string is still long enough
6235 subl(cnt1, tmp);
6236 cmpl(cnt1, cnt2);
6237 jccb(Assembler::greaterEqual, FOUND_SUBSTR);
6238 // Left less then substring.
6239
6240 bind(RET_NOT_FOUND);
6241 movl(result, -1);
6242 jmp(CLEANUP);
6243
6244 bind(FOUND_SUBSTR);
6245 // Compute start addr of substr
6246 lea(result, Address(result, tmp, scale1));
6247 if (int_cnt2 > 0) { // Constant substring
6248 // Repeat search for small substring (< 8 chars)
6249 // from new point without reloading substring.
6250 // Have to check that we don't read beyond string.
6251 cmpl(tmp, stride-int_cnt2);
6252 jccb(Assembler::greater, ADJUST_STR);
6253 // Fall through if matched whole substring.
6254 } else { // non constant
6255 assert(int_cnt2 == -1, "should be != 0");
6256
6257 addl(tmp, cnt2);
6258 // Found result if we matched whole substring.
6259 cmpl(tmp, stride);
6260 jcc(Assembler::lessEqual, RET_FOUND);
6261
6262 // Repeat search for small substring (<= 8 chars)
6263 // from new point 'str1' without reloading substring.
6264 cmpl(cnt2, stride);
6265 // Have to check that we don't read beyond string.
6266 jccb(Assembler::lessEqual, ADJUST_STR);
6267
6268 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG;
6269 // Compare the rest of substring (> 8 chars).
6270 movptr(str1, result);
6271
6272 cmpl(tmp, cnt2);
6273 // First 8 chars are already matched.
6274 jccb(Assembler::equal, CHECK_NEXT);
6275
6276 bind(SCAN_SUBSTR);
6277 pcmpestri(vec, Address(str1, 0), mode);
6278 // Need to reload strings pointers if not matched whole vector
6279 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6280
6281 bind(CHECK_NEXT);
6282 subl(cnt2, stride);
6283 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring
6284 addptr(str1, 16);
6285 if (ae == StrIntrinsicNode::UL) {
6286 addptr(str2, 8);
6287 } else {
6288 addptr(str2, 16);
6289 }
6290 subl(cnt1, stride);
6291 cmpl(cnt2, stride); // Do not read beyond substring
6292 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR);
6293 // Back-up strings to avoid reading beyond substring.
6294
6295 if (ae == StrIntrinsicNode::UL) {
6296 lea(str2, Address(str2, cnt2, scale2, -8));
6297 lea(str1, Address(str1, cnt2, scale1, -16));
6298 } else {
6299 lea(str2, Address(str2, cnt2, scale2, -16));
6300 lea(str1, Address(str1, cnt2, scale1, -16));
6301 }
6302 subl(cnt1, cnt2);
6303 movl(cnt2, stride);
6304 addl(cnt1, stride);
6305 bind(CONT_SCAN_SUBSTR);
6306 if (ae == StrIntrinsicNode::UL) {
6307 pmovzxbw(vec, Address(str2, 0));
6308 } else {
6309 movdqu(vec, Address(str2, 0));
6310 }
6311 jmp(SCAN_SUBSTR);
6312
6313 bind(RET_FOUND_LONG);
6314 movptr(str1, Address(rsp, wordSize));
6315 } // non constant
6316
6317 bind(RET_FOUND);
6318 // Compute substr offset
6319 subptr(result, str1);
6320 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6321 shrl(result, 1); // index
6322 }
6323 bind(CLEANUP);
6324 pop(rsp); // restore SP
6325
6326 } // string_indexof
6327
6328 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result,
6329 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) {
6330 ShortBranchVerifier sbv(this);
6331 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
6332
6333 int stride = 8;
6334
6335 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP,
6336 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP,
6337 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT,
6338 FOUND_SEQ_CHAR, DONE_LABEL;
6339
6340 movptr(result, str1);
6341 if (UseAVX >= 2) {
6342 cmpl(cnt1, stride);
6343 jcc(Assembler::less, SCAN_TO_CHAR);
6344 cmpl(cnt1, 2*stride);
6345 jcc(Assembler::less, SCAN_TO_8_CHAR_INIT);
6346 movdl(vec1, ch);
6347 vpbroadcastw(vec1, vec1, Assembler::AVX_256bit);
6348 vpxor(vec2, vec2);
6349 movl(tmp, cnt1);
6350 andl(tmp, 0xFFFFFFF0); //vector count (in chars)
6351 andl(cnt1,0x0000000F); //tail count (in chars)
6352
6353 bind(SCAN_TO_16_CHAR_LOOP);
6354 vmovdqu(vec3, Address(result, 0));
6355 vpcmpeqw(vec3, vec3, vec1, 1);
6356 vptest(vec2, vec3);
6357 jcc(Assembler::carryClear, FOUND_CHAR);
6358 addptr(result, 32);
6359 subl(tmp, 2*stride);
6360 jcc(Assembler::notZero, SCAN_TO_16_CHAR_LOOP);
6361 jmp(SCAN_TO_8_CHAR);
6362 bind(SCAN_TO_8_CHAR_INIT);
6363 movdl(vec1, ch);
6364 pshuflw(vec1, vec1, 0x00);
6365 pshufd(vec1, vec1, 0);
6366 pxor(vec2, vec2);
6367 }
6368 bind(SCAN_TO_8_CHAR);
6369 cmpl(cnt1, stride);
6370 jcc(Assembler::less, SCAN_TO_CHAR);
6371 if (UseAVX < 2) {
6372 movdl(vec1, ch);
6373 pshuflw(vec1, vec1, 0x00);
6374 pshufd(vec1, vec1, 0);
6375 pxor(vec2, vec2);
6376 }
6377 movl(tmp, cnt1);
6378 andl(tmp, 0xFFFFFFF8); //vector count (in chars)
6379 andl(cnt1,0x00000007); //tail count (in chars)
6380
6381 bind(SCAN_TO_8_CHAR_LOOP);
6382 movdqu(vec3, Address(result, 0));
6383 pcmpeqw(vec3, vec1);
6384 ptest(vec2, vec3);
6385 jcc(Assembler::carryClear, FOUND_CHAR);
6386 addptr(result, 16);
6387 subl(tmp, stride);
6388 jcc(Assembler::notZero, SCAN_TO_8_CHAR_LOOP);
6389 bind(SCAN_TO_CHAR);
6390 testl(cnt1, cnt1);
6391 jcc(Assembler::zero, RET_NOT_FOUND);
6392 bind(SCAN_TO_CHAR_LOOP);
6393 load_unsigned_short(tmp, Address(result, 0));
6394 cmpl(ch, tmp);
6395 jccb(Assembler::equal, FOUND_SEQ_CHAR);
6396 addptr(result, 2);
6397 subl(cnt1, 1);
6398 jccb(Assembler::zero, RET_NOT_FOUND);
6399 jmp(SCAN_TO_CHAR_LOOP);
6400
6401 bind(RET_NOT_FOUND);
6402 movl(result, -1);
6403 jmpb(DONE_LABEL);
6404
6405 bind(FOUND_CHAR);
6406 if (UseAVX >= 2) {
6407 vpmovmskb(tmp, vec3);
6408 } else {
6409 pmovmskb(tmp, vec3);
6410 }
6411 bsfl(ch, tmp);
6412 addl(result, ch);
6413
6414 bind(FOUND_SEQ_CHAR);
6415 subptr(result, str1);
6416 shrl(result, 1);
6417
6418 bind(DONE_LABEL);
6419 } // string_indexof_char
6420
6421 // helper function for string_compare
6422 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2,
6423 Address::ScaleFactor scale, Address::ScaleFactor scale1,
6424 Address::ScaleFactor scale2, Register index, int ae) {
6425 if (ae == StrIntrinsicNode::LL) {
6426 load_unsigned_byte(elem1, Address(str1, index, scale, 0));
6427 load_unsigned_byte(elem2, Address(str2, index, scale, 0));
6428 } else if (ae == StrIntrinsicNode::UU) {
6429 load_unsigned_short(elem1, Address(str1, index, scale, 0));
6430 load_unsigned_short(elem2, Address(str2, index, scale, 0));
6431 } else {
6432 load_unsigned_byte(elem1, Address(str1, index, scale1, 0));
6433 load_unsigned_short(elem2, Address(str2, index, scale2, 0));
6434 }
6435 }
6436
6437 // Compare strings, used for char[] and byte[].
6438 void MacroAssembler::string_compare(Register str1, Register str2,
6439 Register cnt1, Register cnt2, Register result,
6440 XMMRegister vec1, int ae) {
6441 ShortBranchVerifier sbv(this);
6442 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
6443 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3
6444 int stride, stride2, adr_stride, adr_stride1, adr_stride2;
6445 int stride2x2 = 0x40;
6446 Address::ScaleFactor scale = Address::no_scale;
6447 Address::ScaleFactor scale1 = Address::no_scale;
6448 Address::ScaleFactor scale2 = Address::no_scale;
6449
6450 if (ae != StrIntrinsicNode::LL) {
6451 stride2x2 = 0x20;
6452 }
6453
6454 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
6455 shrl(cnt2, 1);
6456 }
6457 // Compute the minimum of the string lengths and the
6458 // difference of the string lengths (stack).
6459 // Do the conditional move stuff
6460 movl(result, cnt1);
6461 subl(cnt1, cnt2);
6462 push(cnt1);
6463 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2)
6464
6465 // Is the minimum length zero?
6466 testl(cnt2, cnt2);
6467 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6468 if (ae == StrIntrinsicNode::LL) {
6469 // Load first bytes
6470 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0]
6471 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0]
6472 } else if (ae == StrIntrinsicNode::UU) {
6473 // Load first characters
6474 load_unsigned_short(result, Address(str1, 0));
6475 load_unsigned_short(cnt1, Address(str2, 0));
6476 } else {
6477 load_unsigned_byte(result, Address(str1, 0));
6478 load_unsigned_short(cnt1, Address(str2, 0));
6479 }
6480 subl(result, cnt1);
6481 jcc(Assembler::notZero, POP_LABEL);
6482
6483 if (ae == StrIntrinsicNode::UU) {
6484 // Divide length by 2 to get number of chars
6485 shrl(cnt2, 1);
6486 }
6487 cmpl(cnt2, 1);
6488 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6489
6490 // Check if the strings start at the same location and setup scale and stride
6491 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6492 cmpptr(str1, str2);
6493 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6494 if (ae == StrIntrinsicNode::LL) {
6495 scale = Address::times_1;
6496 stride = 16;
6497 } else {
6498 scale = Address::times_2;
6499 stride = 8;
6500 }
6501 } else {
6502 scale1 = Address::times_1;
6503 scale2 = Address::times_2;
6504 // scale not used
6505 stride = 8;
6506 }
6507
6508 if (UseAVX >= 2 && UseSSE42Intrinsics) {
6509 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
6510 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
6511 Label COMPARE_WIDE_VECTORS_LOOP_AVX2;
6512 Label COMPARE_TAIL_LONG;
6513 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3
6514
6515 int pcmpmask = 0x19;
6516 if (ae == StrIntrinsicNode::LL) {
6517 pcmpmask &= ~0x01;
6518 }
6519
6520 // Setup to compare 16-chars (32-bytes) vectors,
6521 // start from first character again because it has aligned address.
6522 if (ae == StrIntrinsicNode::LL) {
6523 stride2 = 32;
6524 } else {
6525 stride2 = 16;
6526 }
6527 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6528 adr_stride = stride << scale;
6529 } else {
6530 adr_stride1 = 8; //stride << scale1;
6531 adr_stride2 = 16; //stride << scale2;
6532 }
6533
6534 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6535 // rax and rdx are used by pcmpestri as elements counters
6536 movl(result, cnt2);
6537 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count
6538 jcc(Assembler::zero, COMPARE_TAIL_LONG);
6539
6540 // fast path : compare first 2 8-char vectors.
6541 bind(COMPARE_16_CHARS);
6542 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6543 movdqu(vec1, Address(str1, 0));
6544 } else {
6545 pmovzxbw(vec1, Address(str1, 0));
6546 }
6547 pcmpestri(vec1, Address(str2, 0), pcmpmask);
6548 jccb(Assembler::below, COMPARE_INDEX_CHAR);
6549
6550 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6551 movdqu(vec1, Address(str1, adr_stride));
6552 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
6553 } else {
6554 pmovzxbw(vec1, Address(str1, adr_stride1));
6555 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask);
6556 }
6557 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
6558 addl(cnt1, stride);
6559
6560 // Compare the characters at index in cnt1
6561 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character
6562 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
6563 subl(result, cnt2);
6564 jmp(POP_LABEL);
6565
6566 // Setup the registers to start vector comparison loop
6567 bind(COMPARE_WIDE_VECTORS);
6568 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6569 lea(str1, Address(str1, result, scale));
6570 lea(str2, Address(str2, result, scale));
6571 } else {
6572 lea(str1, Address(str1, result, scale1));
6573 lea(str2, Address(str2, result, scale2));
6574 }
6575 subl(result, stride2);
6576 subl(cnt2, stride2);
6577 jcc(Assembler::zero, COMPARE_WIDE_TAIL);
6578 negptr(result);
6579
6580 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
6581 bind(COMPARE_WIDE_VECTORS_LOOP);
6582
6583 #ifdef _LP64
6584 if ((AVX3Threshold == 0) && VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
6585 cmpl(cnt2, stride2x2);
6586 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2);
6587 testl(cnt2, stride2x2-1); // cnt2 holds the vector count
6588 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40
6589
6590 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
6591 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6592 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit);
6593 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
6594 } else {
6595 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit);
6596 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
6597 }
6598 kortestql(k7, k7);
6599 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare
6600 addptr(result, stride2x2); // update since we already compared at this addr
6601 subl(cnt2, stride2x2); // and sub the size too
6602 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3);
6603
6604 vpxor(vec1, vec1);
6605 jmpb(COMPARE_WIDE_TAIL);
6606 }//if (VM_Version::supports_avx512vlbw())
6607 #endif // _LP64
6608
6609
6610 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
6611 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6612 vmovdqu(vec1, Address(str1, result, scale));
6613 vpxor(vec1, Address(str2, result, scale));
6614 } else {
6615 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit);
6616 vpxor(vec1, Address(str2, result, scale2));
6617 }
6618 vptest(vec1, vec1);
6619 jcc(Assembler::notZero, VECTOR_NOT_EQUAL);
6620 addptr(result, stride2);
6621 subl(cnt2, stride2);
6622 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
6623 // clean upper bits of YMM registers
6624 vpxor(vec1, vec1);
6625
6626 // compare wide vectors tail
6627 bind(COMPARE_WIDE_TAIL);
6628 testptr(result, result);
6629 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6630
6631 movl(result, stride2);
6632 movl(cnt2, result);
6633 negptr(result);
6634 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2);
6635
6636 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
6637 bind(VECTOR_NOT_EQUAL);
6638 // clean upper bits of YMM registers
6639 vpxor(vec1, vec1);
6640 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6641 lea(str1, Address(str1, result, scale));
6642 lea(str2, Address(str2, result, scale));
6643 } else {
6644 lea(str1, Address(str1, result, scale1));
6645 lea(str2, Address(str2, result, scale2));
6646 }
6647 jmp(COMPARE_16_CHARS);
6648
6649 // Compare tail chars, length between 1 to 15 chars
6650 bind(COMPARE_TAIL_LONG);
6651 movl(cnt2, result);
6652 cmpl(cnt2, stride);
6653 jcc(Assembler::less, COMPARE_SMALL_STR);
6654
6655 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6656 movdqu(vec1, Address(str1, 0));
6657 } else {
6658 pmovzxbw(vec1, Address(str1, 0));
6659 }
6660 pcmpestri(vec1, Address(str2, 0), pcmpmask);
6661 jcc(Assembler::below, COMPARE_INDEX_CHAR);
6662 subptr(cnt2, stride);
6663 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6664 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6665 lea(str1, Address(str1, result, scale));
6666 lea(str2, Address(str2, result, scale));
6667 } else {
6668 lea(str1, Address(str1, result, scale1));
6669 lea(str2, Address(str2, result, scale2));
6670 }
6671 negptr(cnt2);
6672 jmpb(WHILE_HEAD_LABEL);
6673
6674 bind(COMPARE_SMALL_STR);
6675 } else if (UseSSE42Intrinsics) {
6676 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
6677 int pcmpmask = 0x19;
6678 // Setup to compare 8-char (16-byte) vectors,
6679 // start from first character again because it has aligned address.
6680 movl(result, cnt2);
6681 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count
6682 if (ae == StrIntrinsicNode::LL) {
6683 pcmpmask &= ~0x01;
6684 }
6685 jcc(Assembler::zero, COMPARE_TAIL);
6686 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6687 lea(str1, Address(str1, result, scale));
6688 lea(str2, Address(str2, result, scale));
6689 } else {
6690 lea(str1, Address(str1, result, scale1));
6691 lea(str2, Address(str2, result, scale2));
6692 }
6693 negptr(result);
6694
6695 // pcmpestri
6696 // inputs:
6697 // vec1- substring
6698 // rax - negative string length (elements count)
6699 // mem - scanned string
6700 // rdx - string length (elements count)
6701 // pcmpmask - cmp mode: 11000 (string compare with negated result)
6702 // + 00 (unsigned bytes) or + 01 (unsigned shorts)
6703 // outputs:
6704 // rcx - first mismatched element index
6705 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6706
6707 bind(COMPARE_WIDE_VECTORS);
6708 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6709 movdqu(vec1, Address(str1, result, scale));
6710 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6711 } else {
6712 pmovzxbw(vec1, Address(str1, result, scale1));
6713 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
6714 }
6715 // After pcmpestri cnt1(rcx) contains mismatched element index
6716
6717 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1
6718 addptr(result, stride);
6719 subptr(cnt2, stride);
6720 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
6721
6722 // compare wide vectors tail
6723 testptr(result, result);
6724 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6725
6726 movl(cnt2, stride);
6727 movl(result, stride);
6728 negptr(result);
6729 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6730 movdqu(vec1, Address(str1, result, scale));
6731 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6732 } else {
6733 pmovzxbw(vec1, Address(str1, result, scale1));
6734 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
6735 }
6736 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
6737
6738 // Mismatched characters in the vectors
6739 bind(VECTOR_NOT_EQUAL);
6740 addptr(cnt1, result);
6741 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
6742 subl(result, cnt2);
6743 jmpb(POP_LABEL);
6744
6745 bind(COMPARE_TAIL); // limit is zero
6746 movl(cnt2, result);
6747 // Fallthru to tail compare
6748 }
6749 // Shift str2 and str1 to the end of the arrays, negate min
6750 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6751 lea(str1, Address(str1, cnt2, scale));
6752 lea(str2, Address(str2, cnt2, scale));
6753 } else {
6754 lea(str1, Address(str1, cnt2, scale1));
6755 lea(str2, Address(str2, cnt2, scale2));
6756 }
6757 decrementl(cnt2); // first character was compared already
6758 negptr(cnt2);
6759
6760 // Compare the rest of the elements
6761 bind(WHILE_HEAD_LABEL);
6762 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae);
6763 subl(result, cnt1);
6764 jccb(Assembler::notZero, POP_LABEL);
6765 increment(cnt2);
6766 jccb(Assembler::notZero, WHILE_HEAD_LABEL);
6767
6768 // Strings are equal up to min length. Return the length difference.
6769 bind(LENGTH_DIFF_LABEL);
6770 pop(result);
6771 if (ae == StrIntrinsicNode::UU) {
6772 // Divide diff by 2 to get number of chars
6773 sarl(result, 1);
6774 }
6775 jmpb(DONE_LABEL);
6776
6777 #ifdef _LP64
6778 if (VM_Version::supports_avx512vlbw()) {
6779
6780 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED);
6781
6782 kmovql(cnt1, k7);
6783 notq(cnt1);
6784 bsfq(cnt2, cnt1);
6785 if (ae != StrIntrinsicNode::LL) {
6786 // Divide diff by 2 to get number of chars
6787 sarl(cnt2, 1);
6788 }
6789 addq(result, cnt2);
6790 if (ae == StrIntrinsicNode::LL) {
6791 load_unsigned_byte(cnt1, Address(str2, result));
6792 load_unsigned_byte(result, Address(str1, result));
6793 } else if (ae == StrIntrinsicNode::UU) {
6794 load_unsigned_short(cnt1, Address(str2, result, scale));
6795 load_unsigned_short(result, Address(str1, result, scale));
6796 } else {
6797 load_unsigned_short(cnt1, Address(str2, result, scale2));
6798 load_unsigned_byte(result, Address(str1, result, scale1));
6799 }
6800 subl(result, cnt1);
6801 jmpb(POP_LABEL);
6802 }//if (VM_Version::supports_avx512vlbw())
6803 #endif // _LP64
6804
6805 // Discard the stored length difference
6806 bind(POP_LABEL);
6807 pop(cnt1);
6808
6809 // That's it
6810 bind(DONE_LABEL);
6811 if(ae == StrIntrinsicNode::UL) {
6812 negl(result);
6813 }
6814
6815 }
6816
6817 // Search for Non-ASCII character (Negative byte value) in a byte array,
6818 // return true if it has any and false otherwise.
6819 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java
6820 // @HotSpotIntrinsicCandidate
6821 // private static boolean hasNegatives(byte[] ba, int off, int len) {
6822 // for (int i = off; i < off + len; i++) {
6823 // if (ba[i] < 0) {
6824 // return true;
6825 // }
6826 // }
6827 // return false;
6828 // }
6829 void MacroAssembler::has_negatives(Register ary1, Register len,
6830 Register result, Register tmp1,
6831 XMMRegister vec1, XMMRegister vec2) {
6832 // rsi: byte array
6833 // rcx: len
6834 // rax: result
6835 ShortBranchVerifier sbv(this);
6836 assert_different_registers(ary1, len, result, tmp1);
6837 assert_different_registers(vec1, vec2);
6838 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE;
6839
6840 // len == 0
6841 testl(len, len);
6842 jcc(Assembler::zero, FALSE_LABEL);
6843
6844 if ((AVX3Threshold == 0) && (UseAVX > 2) && // AVX512
6845 VM_Version::supports_avx512vlbw() &&
6846 VM_Version::supports_bmi2()) {
6847
6848 Label test_64_loop, test_tail;
6849 Register tmp3_aliased = len;
6850
6851 movl(tmp1, len);
6852 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit);
6853
6854 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F
6855 andl(len, ~(64 - 1)); // vector count (in chars)
6856 jccb(Assembler::zero, test_tail);
6857
6858 lea(ary1, Address(ary1, len, Address::times_1));
6859 negptr(len);
6860
6861 bind(test_64_loop);
6862 // Check whether our 64 elements of size byte contain negatives
6863 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit);
6864 kortestql(k2, k2);
6865 jcc(Assembler::notZero, TRUE_LABEL);
6866
6867 addptr(len, 64);
6868 jccb(Assembler::notZero, test_64_loop);
6869
6870
6871 bind(test_tail);
6872 // bail out when there is nothing to be done
6873 testl(tmp1, -1);
6874 jcc(Assembler::zero, FALSE_LABEL);
6875
6876 // ~(~0 << len) applied up to two times (for 32-bit scenario)
6877 #ifdef _LP64
6878 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF);
6879 shlxq(tmp3_aliased, tmp3_aliased, tmp1);
6880 notq(tmp3_aliased);
6881 kmovql(k3, tmp3_aliased);
6882 #else
6883 Label k_init;
6884 jmp(k_init);
6885
6886 // We could not read 64-bits from a general purpose register thus we move
6887 // data required to compose 64 1's to the instruction stream
6888 // We emit 64 byte wide series of elements from 0..63 which later on would
6889 // be used as a compare targets with tail count contained in tmp1 register.
6890 // Result would be a k register having tmp1 consecutive number or 1
6891 // counting from least significant bit.
6892 address tmp = pc();
6893 emit_int64(0x0706050403020100);
6894 emit_int64(0x0F0E0D0C0B0A0908);
6895 emit_int64(0x1716151413121110);
6896 emit_int64(0x1F1E1D1C1B1A1918);
6897 emit_int64(0x2726252423222120);
6898 emit_int64(0x2F2E2D2C2B2A2928);
6899 emit_int64(0x3736353433323130);
6900 emit_int64(0x3F3E3D3C3B3A3938);
6901
6902 bind(k_init);
6903 lea(len, InternalAddress(tmp));
6904 // create mask to test for negative byte inside a vector
6905 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit);
6906 evpcmpgtb(k3, vec1, Address(len, 0), Assembler::AVX_512bit);
6907
6908 #endif
6909 evpcmpgtb(k2, k3, vec2, Address(ary1, 0), Assembler::AVX_512bit);
6910 ktestq(k2, k3);
6911 jcc(Assembler::notZero, TRUE_LABEL);
6912
6913 jmp(FALSE_LABEL);
6914 } else {
6915 movl(result, len); // copy
6916
6917 if (UseAVX >= 2 && UseSSE >= 2) {
6918 // With AVX2, use 32-byte vector compare
6919 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
6920
6921 // Compare 32-byte vectors
6922 andl(result, 0x0000001f); // tail count (in bytes)
6923 andl(len, 0xffffffe0); // vector count (in bytes)
6924 jccb(Assembler::zero, COMPARE_TAIL);
6925
6926 lea(ary1, Address(ary1, len, Address::times_1));
6927 negptr(len);
6928
6929 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector
6930 movdl(vec2, tmp1);
6931 vpbroadcastd(vec2, vec2, Assembler::AVX_256bit);
6932
6933 bind(COMPARE_WIDE_VECTORS);
6934 vmovdqu(vec1, Address(ary1, len, Address::times_1));
6935 vptest(vec1, vec2);
6936 jccb(Assembler::notZero, TRUE_LABEL);
6937 addptr(len, 32);
6938 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
6939
6940 testl(result, result);
6941 jccb(Assembler::zero, FALSE_LABEL);
6942
6943 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
6944 vptest(vec1, vec2);
6945 jccb(Assembler::notZero, TRUE_LABEL);
6946 jmpb(FALSE_LABEL);
6947
6948 bind(COMPARE_TAIL); // len is zero
6949 movl(len, result);
6950 // Fallthru to tail compare
6951 } else if (UseSSE42Intrinsics) {
6952 // With SSE4.2, use double quad vector compare
6953 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
6954
6955 // Compare 16-byte vectors
6956 andl(result, 0x0000000f); // tail count (in bytes)
6957 andl(len, 0xfffffff0); // vector count (in bytes)
6958 jcc(Assembler::zero, COMPARE_TAIL);
6959
6960 lea(ary1, Address(ary1, len, Address::times_1));
6961 negptr(len);
6962
6963 movl(tmp1, 0x80808080);
6964 movdl(vec2, tmp1);
6965 pshufd(vec2, vec2, 0);
6966
6967 bind(COMPARE_WIDE_VECTORS);
6968 movdqu(vec1, Address(ary1, len, Address::times_1));
6969 ptest(vec1, vec2);
6970 jcc(Assembler::notZero, TRUE_LABEL);
6971 addptr(len, 16);
6972 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
6973
6974 testl(result, result);
6975 jcc(Assembler::zero, FALSE_LABEL);
6976
6977 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
6978 ptest(vec1, vec2);
6979 jccb(Assembler::notZero, TRUE_LABEL);
6980 jmpb(FALSE_LABEL);
6981
6982 bind(COMPARE_TAIL); // len is zero
6983 movl(len, result);
6984 // Fallthru to tail compare
6985 }
6986 }
6987 // Compare 4-byte vectors
6988 andl(len, 0xfffffffc); // vector count (in bytes)
6989 jccb(Assembler::zero, COMPARE_CHAR);
6990
6991 lea(ary1, Address(ary1, len, Address::times_1));
6992 negptr(len);
6993
6994 bind(COMPARE_VECTORS);
6995 movl(tmp1, Address(ary1, len, Address::times_1));
6996 andl(tmp1, 0x80808080);
6997 jccb(Assembler::notZero, TRUE_LABEL);
6998 addptr(len, 4);
6999 jcc(Assembler::notZero, COMPARE_VECTORS);
7000
7001 // Compare trailing char (final 2 bytes), if any
7002 bind(COMPARE_CHAR);
7003 testl(result, 0x2); // tail char
7004 jccb(Assembler::zero, COMPARE_BYTE);
7005 load_unsigned_short(tmp1, Address(ary1, 0));
7006 andl(tmp1, 0x00008080);
7007 jccb(Assembler::notZero, TRUE_LABEL);
7008 subptr(result, 2);
7009 lea(ary1, Address(ary1, 2));
7010
7011 bind(COMPARE_BYTE);
7012 testl(result, 0x1); // tail byte
7013 jccb(Assembler::zero, FALSE_LABEL);
7014 load_unsigned_byte(tmp1, Address(ary1, 0));
7015 andl(tmp1, 0x00000080);
7016 jccb(Assembler::notEqual, TRUE_LABEL);
7017 jmpb(FALSE_LABEL);
7018
7019 bind(TRUE_LABEL);
7020 movl(result, 1); // return true
7021 jmpb(DONE);
7022
7023 bind(FALSE_LABEL);
7024 xorl(result, result); // return false
7025
7026 // That's it
7027 bind(DONE);
7028 if (UseAVX >= 2 && UseSSE >= 2) {
7029 // clean upper bits of YMM registers
7030 vpxor(vec1, vec1);
7031 vpxor(vec2, vec2);
7032 }
7033 }
7034 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings.
7035 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2,
7036 Register limit, Register result, Register chr,
7037 XMMRegister vec1, XMMRegister vec2, bool is_char) {
7038 ShortBranchVerifier sbv(this);
7039 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE;
7040
7041 int length_offset = arrayOopDesc::length_offset_in_bytes();
7042 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE);
7043
7044 if (is_array_equ) {
7045 // Check the input args
7046 cmpoop(ary1, ary2);
7047 jcc(Assembler::equal, TRUE_LABEL);
7048
7049 // Need additional checks for arrays_equals.
7050 testptr(ary1, ary1);
7051 jcc(Assembler::zero, FALSE_LABEL);
7052 testptr(ary2, ary2);
7053 jcc(Assembler::zero, FALSE_LABEL);
7054
7055 // Check the lengths
7056 movl(limit, Address(ary1, length_offset));
7057 cmpl(limit, Address(ary2, length_offset));
7058 jcc(Assembler::notEqual, FALSE_LABEL);
7059 }
7060
7061 // count == 0
7062 testl(limit, limit);
7063 jcc(Assembler::zero, TRUE_LABEL);
7064
7065 if (is_array_equ) {
7066 // Load array address
7067 lea(ary1, Address(ary1, base_offset));
7068 lea(ary2, Address(ary2, base_offset));
7069 }
7070
7071 if (is_array_equ && is_char) {
7072 // arrays_equals when used for char[].
7073 shll(limit, 1); // byte count != 0
7074 }
7075 movl(result, limit); // copy
7076
7077 if (UseAVX >= 2) {
7078 // With AVX2, use 32-byte vector compare
7079 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7080
7081 // Compare 32-byte vectors
7082 andl(result, 0x0000001f); // tail count (in bytes)
7083 andl(limit, 0xffffffe0); // vector count (in bytes)
7084 jcc(Assembler::zero, COMPARE_TAIL);
7085
7086 lea(ary1, Address(ary1, limit, Address::times_1));
7087 lea(ary2, Address(ary2, limit, Address::times_1));
7088 negptr(limit);
7089
7090 #ifdef _LP64
7091 if ((AVX3Threshold == 0) && VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
7092 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3;
7093
7094 cmpl(limit, -64);
7095 jcc(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2);
7096
7097 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
7098
7099 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit);
7100 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit);
7101 kortestql(k7, k7);
7102 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
7103 addptr(limit, 64); // update since we already compared at this addr
7104 cmpl(limit, -64);
7105 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3);
7106
7107 // At this point we may still need to compare -limit+result bytes.
7108 // We could execute the next two instruction and just continue via non-wide path:
7109 // cmpl(limit, 0);
7110 // jcc(Assembler::equal, COMPARE_TAIL); // true
7111 // But since we stopped at the points ary{1,2}+limit which are
7112 // not farther than 64 bytes from the ends of arrays ary{1,2}+result
7113 // (|limit| <= 32 and result < 32),
7114 // we may just compare the last 64 bytes.
7115 //
7116 addptr(result, -64); // it is safe, bc we just came from this area
7117 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit);
7118 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit);
7119 kortestql(k7, k7);
7120 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
7121
7122 jmp(TRUE_LABEL);
7123
7124 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
7125
7126 }//if (VM_Version::supports_avx512vlbw())
7127 #endif //_LP64
7128 bind(COMPARE_WIDE_VECTORS);
7129 vmovdqu(vec1, Address(ary1, limit, Address::times_1));
7130 vmovdqu(vec2, Address(ary2, limit, Address::times_1));
7131 vpxor(vec1, vec2);
7132
7133 vptest(vec1, vec1);
7134 jcc(Assembler::notZero, FALSE_LABEL);
7135 addptr(limit, 32);
7136 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7137
7138 testl(result, result);
7139 jcc(Assembler::zero, TRUE_LABEL);
7140
7141 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
7142 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
7143 vpxor(vec1, vec2);
7144
7145 vptest(vec1, vec1);
7146 jccb(Assembler::notZero, FALSE_LABEL);
7147 jmpb(TRUE_LABEL);
7148
7149 bind(COMPARE_TAIL); // limit is zero
7150 movl(limit, result);
7151 // Fallthru to tail compare
7152 } else if (UseSSE42Intrinsics) {
7153 // With SSE4.2, use double quad vector compare
7154 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7155
7156 // Compare 16-byte vectors
7157 andl(result, 0x0000000f); // tail count (in bytes)
7158 andl(limit, 0xfffffff0); // vector count (in bytes)
7159 jcc(Assembler::zero, COMPARE_TAIL);
7160
7161 lea(ary1, Address(ary1, limit, Address::times_1));
7162 lea(ary2, Address(ary2, limit, Address::times_1));
7163 negptr(limit);
7164
7165 bind(COMPARE_WIDE_VECTORS);
7166 movdqu(vec1, Address(ary1, limit, Address::times_1));
7167 movdqu(vec2, Address(ary2, limit, Address::times_1));
7168 pxor(vec1, vec2);
7169
7170 ptest(vec1, vec1);
7171 jcc(Assembler::notZero, FALSE_LABEL);
7172 addptr(limit, 16);
7173 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7174
7175 testl(result, result);
7176 jcc(Assembler::zero, TRUE_LABEL);
7177
7178 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
7179 movdqu(vec2, Address(ary2, result, Address::times_1, -16));
7180 pxor(vec1, vec2);
7181
7182 ptest(vec1, vec1);
7183 jccb(Assembler::notZero, FALSE_LABEL);
7184 jmpb(TRUE_LABEL);
7185
7186 bind(COMPARE_TAIL); // limit is zero
7187 movl(limit, result);
7188 // Fallthru to tail compare
7189 }
7190
7191 // Compare 4-byte vectors
7192 andl(limit, 0xfffffffc); // vector count (in bytes)
7193 jccb(Assembler::zero, COMPARE_CHAR);
7194
7195 lea(ary1, Address(ary1, limit, Address::times_1));
7196 lea(ary2, Address(ary2, limit, Address::times_1));
7197 negptr(limit);
7198
7199 bind(COMPARE_VECTORS);
7200 movl(chr, Address(ary1, limit, Address::times_1));
7201 cmpl(chr, Address(ary2, limit, Address::times_1));
7202 jccb(Assembler::notEqual, FALSE_LABEL);
7203 addptr(limit, 4);
7204 jcc(Assembler::notZero, COMPARE_VECTORS);
7205
7206 // Compare trailing char (final 2 bytes), if any
7207 bind(COMPARE_CHAR);
7208 testl(result, 0x2); // tail char
7209 jccb(Assembler::zero, COMPARE_BYTE);
7210 load_unsigned_short(chr, Address(ary1, 0));
7211 load_unsigned_short(limit, Address(ary2, 0));
7212 cmpl(chr, limit);
7213 jccb(Assembler::notEqual, FALSE_LABEL);
7214
7215 if (is_array_equ && is_char) {
7216 bind(COMPARE_BYTE);
7217 } else {
7218 lea(ary1, Address(ary1, 2));
7219 lea(ary2, Address(ary2, 2));
7220
7221 bind(COMPARE_BYTE);
7222 testl(result, 0x1); // tail byte
7223 jccb(Assembler::zero, TRUE_LABEL);
7224 load_unsigned_byte(chr, Address(ary1, 0));
7225 load_unsigned_byte(limit, Address(ary2, 0));
7226 cmpl(chr, limit);
7227 jccb(Assembler::notEqual, FALSE_LABEL);
7228 }
7229 bind(TRUE_LABEL);
7230 movl(result, 1); // return true
7231 jmpb(DONE);
7232
7233 bind(FALSE_LABEL);
7234 xorl(result, result); // return false
7235
7236 // That's it
7237 bind(DONE);
7238 if (UseAVX >= 2) {
7239 // clean upper bits of YMM registers
7240 vpxor(vec1, vec1);
7241 vpxor(vec2, vec2);
7242 }
7243 }
7244
7245 #endif
7246
7247 void MacroAssembler::generate_fill(BasicType t, bool aligned,
7248 Register to, Register value, Register count,
7249 Register rtmp, XMMRegister xtmp) {
7250 ShortBranchVerifier sbv(this);
7251 assert_different_registers(to, value, count, rtmp);
7252 Label L_exit;
7253 Label L_fill_2_bytes, L_fill_4_bytes;
7254
7255 int shift = -1;
7256 switch (t) {
7257 case T_BYTE:
7258 shift = 2;
7259 break;
7260 case T_SHORT:
7261 shift = 1;
7262 break;
7263 case T_INT:
7264 shift = 0;
7265 break;
7266 default: ShouldNotReachHere();
7267 }
7268
7269 if (t == T_BYTE) {
7270 andl(value, 0xff);
7271 movl(rtmp, value);
7272 shll(rtmp, 8);
7273 orl(value, rtmp);
7274 }
7275 if (t == T_SHORT) {
7276 andl(value, 0xffff);
7277 }
7278 if (t == T_BYTE || t == T_SHORT) {
7279 movl(rtmp, value);
7280 shll(rtmp, 16);
7281 orl(value, rtmp);
7282 }
7283
7284 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
7285 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp
7286 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) {
7287 Label L_skip_align2;
7288 // align source address at 4 bytes address boundary
7289 if (t == T_BYTE) {
7290 Label L_skip_align1;
7291 // One byte misalignment happens only for byte arrays
7292 testptr(to, 1);
7293 jccb(Assembler::zero, L_skip_align1);
7294 movb(Address(to, 0), value);
7295 increment(to);
7296 decrement(count);
7297 BIND(L_skip_align1);
7298 }
7299 // Two bytes misalignment happens only for byte and short (char) arrays
7300 testptr(to, 2);
7301 jccb(Assembler::zero, L_skip_align2);
7302 movw(Address(to, 0), value);
7303 addptr(to, 2);
7304 subl(count, 1<<(shift-1));
7305 BIND(L_skip_align2);
7306 }
7307 if (UseSSE < 2) {
7308 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7309 // Fill 32-byte chunks
7310 subl(count, 8 << shift);
7311 jcc(Assembler::less, L_check_fill_8_bytes);
7312 align(16);
7313
7314 BIND(L_fill_32_bytes_loop);
7315
7316 for (int i = 0; i < 32; i += 4) {
7317 movl(Address(to, i), value);
7318 }
7319
7320 addptr(to, 32);
7321 subl(count, 8 << shift);
7322 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7323 BIND(L_check_fill_8_bytes);
7324 addl(count, 8 << shift);
7325 jccb(Assembler::zero, L_exit);
7326 jmpb(L_fill_8_bytes);
7327
7328 //
7329 // length is too short, just fill qwords
7330 //
7331 BIND(L_fill_8_bytes_loop);
7332 movl(Address(to, 0), value);
7333 movl(Address(to, 4), value);
7334 addptr(to, 8);
7335 BIND(L_fill_8_bytes);
7336 subl(count, 1 << (shift + 1));
7337 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7338 // fall through to fill 4 bytes
7339 } else {
7340 Label L_fill_32_bytes;
7341 if (!UseUnalignedLoadStores) {
7342 // align to 8 bytes, we know we are 4 byte aligned to start
7343 testptr(to, 4);
7344 jccb(Assembler::zero, L_fill_32_bytes);
7345 movl(Address(to, 0), value);
7346 addptr(to, 4);
7347 subl(count, 1<<shift);
7348 }
7349 BIND(L_fill_32_bytes);
7350 {
7351 assert( UseSSE >= 2, "supported cpu only" );
7352 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7353 movdl(xtmp, value);
7354 if (UseAVX >= 2 && UseUnalignedLoadStores) {
7355 Label L_check_fill_32_bytes;
7356 if (UseAVX > 2) {
7357 // Fill 64-byte chunks
7358 Label L_fill_64_bytes_loop_avx3, L_check_fill_64_bytes_avx2;
7359
7360 // If number of bytes to fill < AVX3Threshold, perform fill using AVX2
7361 cmpl(count, AVX3Threshold);
7362 jccb(Assembler::below, L_check_fill_64_bytes_avx2);
7363
7364 vpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit);
7365
7366 subl(count, 16 << shift);
7367 jccb(Assembler::less, L_check_fill_32_bytes);
7368 align(16);
7369
7370 BIND(L_fill_64_bytes_loop_avx3);
7371 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit);
7372 addptr(to, 64);
7373 subl(count, 16 << shift);
7374 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop_avx3);
7375 jmpb(L_check_fill_32_bytes);
7376
7377 BIND(L_check_fill_64_bytes_avx2);
7378 }
7379 // Fill 64-byte chunks
7380 Label L_fill_64_bytes_loop;
7381 vpbroadcastd(xtmp, xtmp, Assembler::AVX_256bit);
7382
7383 subl(count, 16 << shift);
7384 jcc(Assembler::less, L_check_fill_32_bytes);
7385 align(16);
7386
7387 BIND(L_fill_64_bytes_loop);
7388 vmovdqu(Address(to, 0), xtmp);
7389 vmovdqu(Address(to, 32), xtmp);
7390 addptr(to, 64);
7391 subl(count, 16 << shift);
7392 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
7393
7394 BIND(L_check_fill_32_bytes);
7395 addl(count, 8 << shift);
7396 jccb(Assembler::less, L_check_fill_8_bytes);
7397 vmovdqu(Address(to, 0), xtmp);
7398 addptr(to, 32);
7399 subl(count, 8 << shift);
7400
7401 BIND(L_check_fill_8_bytes);
7402 // clean upper bits of YMM registers
7403 movdl(xtmp, value);
7404 pshufd(xtmp, xtmp, 0);
7405 } else {
7406 // Fill 32-byte chunks
7407 pshufd(xtmp, xtmp, 0);
7408
7409 subl(count, 8 << shift);
7410 jcc(Assembler::less, L_check_fill_8_bytes);
7411 align(16);
7412
7413 BIND(L_fill_32_bytes_loop);
7414
7415 if (UseUnalignedLoadStores) {
7416 movdqu(Address(to, 0), xtmp);
7417 movdqu(Address(to, 16), xtmp);
7418 } else {
7419 movq(Address(to, 0), xtmp);
7420 movq(Address(to, 8), xtmp);
7421 movq(Address(to, 16), xtmp);
7422 movq(Address(to, 24), xtmp);
7423 }
7424
7425 addptr(to, 32);
7426 subl(count, 8 << shift);
7427 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7428
7429 BIND(L_check_fill_8_bytes);
7430 }
7431 addl(count, 8 << shift);
7432 jccb(Assembler::zero, L_exit);
7433 jmpb(L_fill_8_bytes);
7434
7435 //
7436 // length is too short, just fill qwords
7437 //
7438 BIND(L_fill_8_bytes_loop);
7439 movq(Address(to, 0), xtmp);
7440 addptr(to, 8);
7441 BIND(L_fill_8_bytes);
7442 subl(count, 1 << (shift + 1));
7443 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7444 }
7445 }
7446 // fill trailing 4 bytes
7447 BIND(L_fill_4_bytes);
7448 testl(count, 1<<shift);
7449 jccb(Assembler::zero, L_fill_2_bytes);
7450 movl(Address(to, 0), value);
7451 if (t == T_BYTE || t == T_SHORT) {
7452 Label L_fill_byte;
7453 addptr(to, 4);
7454 BIND(L_fill_2_bytes);
7455 // fill trailing 2 bytes
7456 testl(count, 1<<(shift-1));
7457 jccb(Assembler::zero, L_fill_byte);
7458 movw(Address(to, 0), value);
7459 if (t == T_BYTE) {
7460 addptr(to, 2);
7461 BIND(L_fill_byte);
7462 // fill trailing byte
7463 testl(count, 1);
7464 jccb(Assembler::zero, L_exit);
7465 movb(Address(to, 0), value);
7466 } else {
7467 BIND(L_fill_byte);
7468 }
7469 } else {
7470 BIND(L_fill_2_bytes);
7471 }
7472 BIND(L_exit);
7473 }
7474
7475 // encode char[] to byte[] in ISO_8859_1
7476 //@HotSpotIntrinsicCandidate
7477 //private static int implEncodeISOArray(byte[] sa, int sp,
7478 //byte[] da, int dp, int len) {
7479 // int i = 0;
7480 // for (; i < len; i++) {
7481 // char c = StringUTF16.getChar(sa, sp++);
7482 // if (c > '\u00FF')
7483 // break;
7484 // da[dp++] = (byte)c;
7485 // }
7486 // return i;
7487 //}
7488 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len,
7489 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
7490 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
7491 Register tmp5, Register result) {
7492
7493 // rsi: src
7494 // rdi: dst
7495 // rdx: len
7496 // rcx: tmp5
7497 // rax: result
7498 ShortBranchVerifier sbv(this);
7499 assert_different_registers(src, dst, len, tmp5, result);
7500 Label L_done, L_copy_1_char, L_copy_1_char_exit;
7501
7502 // set result
7503 xorl(result, result);
7504 // check for zero length
7505 testl(len, len);
7506 jcc(Assembler::zero, L_done);
7507
7508 movl(result, len);
7509
7510 // Setup pointers
7511 lea(src, Address(src, len, Address::times_2)); // char[]
7512 lea(dst, Address(dst, len, Address::times_1)); // byte[]
7513 negptr(len);
7514
7515 if (UseSSE42Intrinsics || UseAVX >= 2) {
7516 Label L_copy_8_chars, L_copy_8_chars_exit;
7517 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit;
7518
7519 if (UseAVX >= 2) {
7520 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit;
7521 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
7522 movdl(tmp1Reg, tmp5);
7523 vpbroadcastd(tmp1Reg, tmp1Reg, Assembler::AVX_256bit);
7524 jmp(L_chars_32_check);
7525
7526 bind(L_copy_32_chars);
7527 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64));
7528 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32));
7529 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
7530 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
7531 jccb(Assembler::notZero, L_copy_32_chars_exit);
7532 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
7533 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1);
7534 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg);
7535
7536 bind(L_chars_32_check);
7537 addptr(len, 32);
7538 jcc(Assembler::lessEqual, L_copy_32_chars);
7539
7540 bind(L_copy_32_chars_exit);
7541 subptr(len, 16);
7542 jccb(Assembler::greater, L_copy_16_chars_exit);
7543
7544 } else if (UseSSE42Intrinsics) {
7545 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
7546 movdl(tmp1Reg, tmp5);
7547 pshufd(tmp1Reg, tmp1Reg, 0);
7548 jmpb(L_chars_16_check);
7549 }
7550
7551 bind(L_copy_16_chars);
7552 if (UseAVX >= 2) {
7553 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32));
7554 vptest(tmp2Reg, tmp1Reg);
7555 jcc(Assembler::notZero, L_copy_16_chars_exit);
7556 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1);
7557 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1);
7558 } else {
7559 if (UseAVX > 0) {
7560 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7561 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7562 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0);
7563 } else {
7564 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7565 por(tmp2Reg, tmp3Reg);
7566 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7567 por(tmp2Reg, tmp4Reg);
7568 }
7569 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
7570 jccb(Assembler::notZero, L_copy_16_chars_exit);
7571 packuswb(tmp3Reg, tmp4Reg);
7572 }
7573 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg);
7574
7575 bind(L_chars_16_check);
7576 addptr(len, 16);
7577 jcc(Assembler::lessEqual, L_copy_16_chars);
7578
7579 bind(L_copy_16_chars_exit);
7580 if (UseAVX >= 2) {
7581 // clean upper bits of YMM registers
7582 vpxor(tmp2Reg, tmp2Reg);
7583 vpxor(tmp3Reg, tmp3Reg);
7584 vpxor(tmp4Reg, tmp4Reg);
7585 movdl(tmp1Reg, tmp5);
7586 pshufd(tmp1Reg, tmp1Reg, 0);
7587 }
7588 subptr(len, 8);
7589 jccb(Assembler::greater, L_copy_8_chars_exit);
7590
7591 bind(L_copy_8_chars);
7592 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16));
7593 ptest(tmp3Reg, tmp1Reg);
7594 jccb(Assembler::notZero, L_copy_8_chars_exit);
7595 packuswb(tmp3Reg, tmp1Reg);
7596 movq(Address(dst, len, Address::times_1, -8), tmp3Reg);
7597 addptr(len, 8);
7598 jccb(Assembler::lessEqual, L_copy_8_chars);
7599
7600 bind(L_copy_8_chars_exit);
7601 subptr(len, 8);
7602 jccb(Assembler::zero, L_done);
7603 }
7604
7605 bind(L_copy_1_char);
7606 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0));
7607 testl(tmp5, 0xff00); // check if Unicode char
7608 jccb(Assembler::notZero, L_copy_1_char_exit);
7609 movb(Address(dst, len, Address::times_1, 0), tmp5);
7610 addptr(len, 1);
7611 jccb(Assembler::less, L_copy_1_char);
7612
7613 bind(L_copy_1_char_exit);
7614 addptr(result, len); // len is negative count of not processed elements
7615
7616 bind(L_done);
7617 }
7618
7619 #ifdef _LP64
7620 /**
7621 * Helper for multiply_to_len().
7622 */
7623 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) {
7624 addq(dest_lo, src1);
7625 adcq(dest_hi, 0);
7626 addq(dest_lo, src2);
7627 adcq(dest_hi, 0);
7628 }
7629
7630 /**
7631 * Multiply 64 bit by 64 bit first loop.
7632 */
7633 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
7634 Register y, Register y_idx, Register z,
7635 Register carry, Register product,
7636 Register idx, Register kdx) {
7637 //
7638 // jlong carry, x[], y[], z[];
7639 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7640 // huge_128 product = y[idx] * x[xstart] + carry;
7641 // z[kdx] = (jlong)product;
7642 // carry = (jlong)(product >>> 64);
7643 // }
7644 // z[xstart] = carry;
7645 //
7646
7647 Label L_first_loop, L_first_loop_exit;
7648 Label L_one_x, L_one_y, L_multiply;
7649
7650 decrementl(xstart);
7651 jcc(Assembler::negative, L_one_x);
7652
7653 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
7654 rorq(x_xstart, 32); // convert big-endian to little-endian
7655
7656 bind(L_first_loop);
7657 decrementl(idx);
7658 jcc(Assembler::negative, L_first_loop_exit);
7659 decrementl(idx);
7660 jcc(Assembler::negative, L_one_y);
7661 movq(y_idx, Address(y, idx, Address::times_4, 0));
7662 rorq(y_idx, 32); // convert big-endian to little-endian
7663 bind(L_multiply);
7664 movq(product, x_xstart);
7665 mulq(y_idx); // product(rax) * y_idx -> rdx:rax
7666 addq(product, carry);
7667 adcq(rdx, 0);
7668 subl(kdx, 2);
7669 movl(Address(z, kdx, Address::times_4, 4), product);
7670 shrq(product, 32);
7671 movl(Address(z, kdx, Address::times_4, 0), product);
7672 movq(carry, rdx);
7673 jmp(L_first_loop);
7674
7675 bind(L_one_y);
7676 movl(y_idx, Address(y, 0));
7677 jmp(L_multiply);
7678
7679 bind(L_one_x);
7680 movl(x_xstart, Address(x, 0));
7681 jmp(L_first_loop);
7682
7683 bind(L_first_loop_exit);
7684 }
7685
7686 /**
7687 * Multiply 64 bit by 64 bit and add 128 bit.
7688 */
7689 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z,
7690 Register yz_idx, Register idx,
7691 Register carry, Register product, int offset) {
7692 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
7693 // z[kdx] = (jlong)product;
7694
7695 movq(yz_idx, Address(y, idx, Address::times_4, offset));
7696 rorq(yz_idx, 32); // convert big-endian to little-endian
7697 movq(product, x_xstart);
7698 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7699 movq(yz_idx, Address(z, idx, Address::times_4, offset));
7700 rorq(yz_idx, 32); // convert big-endian to little-endian
7701
7702 add2_with_carry(rdx, product, carry, yz_idx);
7703
7704 movl(Address(z, idx, Address::times_4, offset+4), product);
7705 shrq(product, 32);
7706 movl(Address(z, idx, Address::times_4, offset), product);
7707
7708 }
7709
7710 /**
7711 * Multiply 128 bit by 128 bit. Unrolled inner loop.
7712 */
7713 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z,
7714 Register yz_idx, Register idx, Register jdx,
7715 Register carry, Register product,
7716 Register carry2) {
7717 // jlong carry, x[], y[], z[];
7718 // int kdx = ystart+1;
7719 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7720 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
7721 // z[kdx+idx+1] = (jlong)product;
7722 // jlong carry2 = (jlong)(product >>> 64);
7723 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
7724 // z[kdx+idx] = (jlong)product;
7725 // carry = (jlong)(product >>> 64);
7726 // }
7727 // idx += 2;
7728 // if (idx > 0) {
7729 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
7730 // z[kdx+idx] = (jlong)product;
7731 // carry = (jlong)(product >>> 64);
7732 // }
7733 //
7734
7735 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7736
7737 movl(jdx, idx);
7738 andl(jdx, 0xFFFFFFFC);
7739 shrl(jdx, 2);
7740
7741 bind(L_third_loop);
7742 subl(jdx, 1);
7743 jcc(Assembler::negative, L_third_loop_exit);
7744 subl(idx, 4);
7745
7746 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
7747 movq(carry2, rdx);
7748
7749 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
7750 movq(carry, rdx);
7751 jmp(L_third_loop);
7752
7753 bind (L_third_loop_exit);
7754
7755 andl (idx, 0x3);
7756 jcc(Assembler::zero, L_post_third_loop_done);
7757
7758 Label L_check_1;
7759 subl(idx, 2);
7760 jcc(Assembler::negative, L_check_1);
7761
7762 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
7763 movq(carry, rdx);
7764
7765 bind (L_check_1);
7766 addl (idx, 0x2);
7767 andl (idx, 0x1);
7768 subl(idx, 1);
7769 jcc(Assembler::negative, L_post_third_loop_done);
7770
7771 movl(yz_idx, Address(y, idx, Address::times_4, 0));
7772 movq(product, x_xstart);
7773 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7774 movl(yz_idx, Address(z, idx, Address::times_4, 0));
7775
7776 add2_with_carry(rdx, product, yz_idx, carry);
7777
7778 movl(Address(z, idx, Address::times_4, 0), product);
7779 shrq(product, 32);
7780
7781 shlq(rdx, 32);
7782 orq(product, rdx);
7783 movq(carry, product);
7784
7785 bind(L_post_third_loop_done);
7786 }
7787
7788 /**
7789 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop.
7790 *
7791 */
7792 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z,
7793 Register carry, Register carry2,
7794 Register idx, Register jdx,
7795 Register yz_idx1, Register yz_idx2,
7796 Register tmp, Register tmp3, Register tmp4) {
7797 assert(UseBMI2Instructions, "should be used only when BMI2 is available");
7798
7799 // jlong carry, x[], y[], z[];
7800 // int kdx = ystart+1;
7801 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7802 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry;
7803 // jlong carry2 = (jlong)(tmp3 >>> 64);
7804 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2;
7805 // carry = (jlong)(tmp4 >>> 64);
7806 // z[kdx+idx+1] = (jlong)tmp3;
7807 // z[kdx+idx] = (jlong)tmp4;
7808 // }
7809 // idx += 2;
7810 // if (idx > 0) {
7811 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry;
7812 // z[kdx+idx] = (jlong)yz_idx1;
7813 // carry = (jlong)(yz_idx1 >>> 64);
7814 // }
7815 //
7816
7817 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7818
7819 movl(jdx, idx);
7820 andl(jdx, 0xFFFFFFFC);
7821 shrl(jdx, 2);
7822
7823 bind(L_third_loop);
7824 subl(jdx, 1);
7825 jcc(Assembler::negative, L_third_loop_exit);
7826 subl(idx, 4);
7827
7828 movq(yz_idx1, Address(y, idx, Address::times_4, 8));
7829 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
7830 movq(yz_idx2, Address(y, idx, Address::times_4, 0));
7831 rorxq(yz_idx2, yz_idx2, 32);
7832
7833 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
7834 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp
7835
7836 movq(yz_idx1, Address(z, idx, Address::times_4, 8));
7837 rorxq(yz_idx1, yz_idx1, 32);
7838 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
7839 rorxq(yz_idx2, yz_idx2, 32);
7840
7841 if (VM_Version::supports_adx()) {
7842 adcxq(tmp3, carry);
7843 adoxq(tmp3, yz_idx1);
7844
7845 adcxq(tmp4, tmp);
7846 adoxq(tmp4, yz_idx2);
7847
7848 movl(carry, 0); // does not affect flags
7849 adcxq(carry2, carry);
7850 adoxq(carry2, carry);
7851 } else {
7852 add2_with_carry(tmp4, tmp3, carry, yz_idx1);
7853 add2_with_carry(carry2, tmp4, tmp, yz_idx2);
7854 }
7855 movq(carry, carry2);
7856
7857 movl(Address(z, idx, Address::times_4, 12), tmp3);
7858 shrq(tmp3, 32);
7859 movl(Address(z, idx, Address::times_4, 8), tmp3);
7860
7861 movl(Address(z, idx, Address::times_4, 4), tmp4);
7862 shrq(tmp4, 32);
7863 movl(Address(z, idx, Address::times_4, 0), tmp4);
7864
7865 jmp(L_third_loop);
7866
7867 bind (L_third_loop_exit);
7868
7869 andl (idx, 0x3);
7870 jcc(Assembler::zero, L_post_third_loop_done);
7871
7872 Label L_check_1;
7873 subl(idx, 2);
7874 jcc(Assembler::negative, L_check_1);
7875
7876 movq(yz_idx1, Address(y, idx, Address::times_4, 0));
7877 rorxq(yz_idx1, yz_idx1, 32);
7878 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
7879 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
7880 rorxq(yz_idx2, yz_idx2, 32);
7881
7882 add2_with_carry(tmp4, tmp3, carry, yz_idx2);
7883
7884 movl(Address(z, idx, Address::times_4, 4), tmp3);
7885 shrq(tmp3, 32);
7886 movl(Address(z, idx, Address::times_4, 0), tmp3);
7887 movq(carry, tmp4);
7888
7889 bind (L_check_1);
7890 addl (idx, 0x2);
7891 andl (idx, 0x1);
7892 subl(idx, 1);
7893 jcc(Assembler::negative, L_post_third_loop_done);
7894 movl(tmp4, Address(y, idx, Address::times_4, 0));
7895 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3
7896 movl(tmp4, Address(z, idx, Address::times_4, 0));
7897
7898 add2_with_carry(carry2, tmp3, tmp4, carry);
7899
7900 movl(Address(z, idx, Address::times_4, 0), tmp3);
7901 shrq(tmp3, 32);
7902
7903 shlq(carry2, 32);
7904 orq(tmp3, carry2);
7905 movq(carry, tmp3);
7906
7907 bind(L_post_third_loop_done);
7908 }
7909
7910 /**
7911 * Code for BigInteger::multiplyToLen() instrinsic.
7912 *
7913 * rdi: x
7914 * rax: xlen
7915 * rsi: y
7916 * rcx: ylen
7917 * r8: z
7918 * r11: zlen
7919 * r12: tmp1
7920 * r13: tmp2
7921 * r14: tmp3
7922 * r15: tmp4
7923 * rbx: tmp5
7924 *
7925 */
7926 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen,
7927 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) {
7928 ShortBranchVerifier sbv(this);
7929 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx);
7930
7931 push(tmp1);
7932 push(tmp2);
7933 push(tmp3);
7934 push(tmp4);
7935 push(tmp5);
7936
7937 push(xlen);
7938 push(zlen);
7939
7940 const Register idx = tmp1;
7941 const Register kdx = tmp2;
7942 const Register xstart = tmp3;
7943
7944 const Register y_idx = tmp4;
7945 const Register carry = tmp5;
7946 const Register product = xlen;
7947 const Register x_xstart = zlen; // reuse register
7948
7949 // First Loop.
7950 //
7951 // final static long LONG_MASK = 0xffffffffL;
7952 // int xstart = xlen - 1;
7953 // int ystart = ylen - 1;
7954 // long carry = 0;
7955 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7956 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
7957 // z[kdx] = (int)product;
7958 // carry = product >>> 32;
7959 // }
7960 // z[xstart] = (int)carry;
7961 //
7962
7963 movl(idx, ylen); // idx = ylen;
7964 movl(kdx, zlen); // kdx = xlen+ylen;
7965 xorq(carry, carry); // carry = 0;
7966
7967 Label L_done;
7968
7969 movl(xstart, xlen);
7970 decrementl(xstart);
7971 jcc(Assembler::negative, L_done);
7972
7973 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
7974
7975 Label L_second_loop;
7976 testl(kdx, kdx);
7977 jcc(Assembler::zero, L_second_loop);
7978
7979 Label L_carry;
7980 subl(kdx, 1);
7981 jcc(Assembler::zero, L_carry);
7982
7983 movl(Address(z, kdx, Address::times_4, 0), carry);
7984 shrq(carry, 32);
7985 subl(kdx, 1);
7986
7987 bind(L_carry);
7988 movl(Address(z, kdx, Address::times_4, 0), carry);
7989
7990 // Second and third (nested) loops.
7991 //
7992 // for (int i = xstart-1; i >= 0; i--) { // Second loop
7993 // carry = 0;
7994 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
7995 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
7996 // (z[k] & LONG_MASK) + carry;
7997 // z[k] = (int)product;
7998 // carry = product >>> 32;
7999 // }
8000 // z[i] = (int)carry;
8001 // }
8002 //
8003 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
8004
8005 const Register jdx = tmp1;
8006
8007 bind(L_second_loop);
8008 xorl(carry, carry); // carry = 0;
8009 movl(jdx, ylen); // j = ystart+1
8010
8011 subl(xstart, 1); // i = xstart-1;
8012 jcc(Assembler::negative, L_done);
8013
8014 push (z);
8015
8016 Label L_last_x;
8017 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j
8018 subl(xstart, 1); // i = xstart-1;
8019 jcc(Assembler::negative, L_last_x);
8020
8021 if (UseBMI2Instructions) {
8022 movq(rdx, Address(x, xstart, Address::times_4, 0));
8023 rorxq(rdx, rdx, 32); // convert big-endian to little-endian
8024 } else {
8025 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
8026 rorq(x_xstart, 32); // convert big-endian to little-endian
8027 }
8028
8029 Label L_third_loop_prologue;
8030 bind(L_third_loop_prologue);
8031
8032 push (x);
8033 push (xstart);
8034 push (ylen);
8035
8036
8037 if (UseBMI2Instructions) {
8038 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4);
8039 } else { // !UseBMI2Instructions
8040 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
8041 }
8042
8043 pop(ylen);
8044 pop(xlen);
8045 pop(x);
8046 pop(z);
8047
8048 movl(tmp3, xlen);
8049 addl(tmp3, 1);
8050 movl(Address(z, tmp3, Address::times_4, 0), carry);
8051 subl(tmp3, 1);
8052 jccb(Assembler::negative, L_done);
8053
8054 shrq(carry, 32);
8055 movl(Address(z, tmp3, Address::times_4, 0), carry);
8056 jmp(L_second_loop);
8057
8058 // Next infrequent code is moved outside loops.
8059 bind(L_last_x);
8060 if (UseBMI2Instructions) {
8061 movl(rdx, Address(x, 0));
8062 } else {
8063 movl(x_xstart, Address(x, 0));
8064 }
8065 jmp(L_third_loop_prologue);
8066
8067 bind(L_done);
8068
8069 pop(zlen);
8070 pop(xlen);
8071
8072 pop(tmp5);
8073 pop(tmp4);
8074 pop(tmp3);
8075 pop(tmp2);
8076 pop(tmp1);
8077 }
8078
8079 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale,
8080 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){
8081 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled.");
8082 Label VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP;
8083 Label VECTOR8_TAIL, VECTOR4_TAIL;
8084 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL;
8085 Label SAME_TILL_END, DONE;
8086 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL;
8087
8088 //scale is in rcx in both Win64 and Unix
8089 ShortBranchVerifier sbv(this);
8090
8091 shlq(length);
8092 xorq(result, result);
8093
8094 if ((AVX3Threshold == 0) && (UseAVX > 2) &&
8095 VM_Version::supports_avx512vlbw()) {
8096 Label VECTOR64_LOOP, VECTOR64_NOT_EQUAL, VECTOR32_TAIL;
8097
8098 cmpq(length, 64);
8099 jcc(Assembler::less, VECTOR32_TAIL);
8100
8101 movq(tmp1, length);
8102 andq(tmp1, 0x3F); // tail count
8103 andq(length, ~(0x3F)); //vector count
8104
8105 bind(VECTOR64_LOOP);
8106 // AVX512 code to compare 64 byte vectors.
8107 evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit);
8108 evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit);
8109 kortestql(k7, k7);
8110 jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL); // mismatch
8111 addq(result, 64);
8112 subq(length, 64);
8113 jccb(Assembler::notZero, VECTOR64_LOOP);
8114
8115 //bind(VECTOR64_TAIL);
8116 testq(tmp1, tmp1);
8117 jcc(Assembler::zero, SAME_TILL_END);
8118
8119 //bind(VECTOR64_TAIL);
8120 // AVX512 code to compare upto 63 byte vectors.
8121 mov64(tmp2, 0xFFFFFFFFFFFFFFFF);
8122 shlxq(tmp2, tmp2, tmp1);
8123 notq(tmp2);
8124 kmovql(k3, tmp2);
8125
8126 evmovdqub(rymm0, k3, Address(obja, result), Assembler::AVX_512bit);
8127 evpcmpeqb(k7, k3, rymm0, Address(objb, result), Assembler::AVX_512bit);
8128
8129 ktestql(k7, k3);
8130 jcc(Assembler::below, SAME_TILL_END); // not mismatch
8131
8132 bind(VECTOR64_NOT_EQUAL);
8133 kmovql(tmp1, k7);
8134 notq(tmp1);
8135 tzcntq(tmp1, tmp1);
8136 addq(result, tmp1);
8137 shrq(result);
8138 jmp(DONE);
8139 bind(VECTOR32_TAIL);
8140 }
8141
8142 cmpq(length, 8);
8143 jcc(Assembler::equal, VECTOR8_LOOP);
8144 jcc(Assembler::less, VECTOR4_TAIL);
8145
8146 if (UseAVX >= 2) {
8147 Label VECTOR16_TAIL, VECTOR32_LOOP;
8148
8149 cmpq(length, 16);
8150 jcc(Assembler::equal, VECTOR16_LOOP);
8151 jcc(Assembler::less, VECTOR8_LOOP);
8152
8153 cmpq(length, 32);
8154 jccb(Assembler::less, VECTOR16_TAIL);
8155
8156 subq(length, 32);
8157 bind(VECTOR32_LOOP);
8158 vmovdqu(rymm0, Address(obja, result));
8159 vmovdqu(rymm1, Address(objb, result));
8160 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit);
8161 vptest(rymm2, rymm2);
8162 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found
8163 addq(result, 32);
8164 subq(length, 32);
8165 jcc(Assembler::greaterEqual, VECTOR32_LOOP);
8166 addq(length, 32);
8167 jcc(Assembler::equal, SAME_TILL_END);
8168 //falling through if less than 32 bytes left //close the branch here.
8169
8170 bind(VECTOR16_TAIL);
8171 cmpq(length, 16);
8172 jccb(Assembler::less, VECTOR8_TAIL);
8173 bind(VECTOR16_LOOP);
8174 movdqu(rymm0, Address(obja, result));
8175 movdqu(rymm1, Address(objb, result));
8176 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit);
8177 ptest(rymm2, rymm2);
8178 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
8179 addq(result, 16);
8180 subq(length, 16);
8181 jcc(Assembler::equal, SAME_TILL_END);
8182 //falling through if less than 16 bytes left
8183 } else {//regular intrinsics
8184
8185 cmpq(length, 16);
8186 jccb(Assembler::less, VECTOR8_TAIL);
8187
8188 subq(length, 16);
8189 bind(VECTOR16_LOOP);
8190 movdqu(rymm0, Address(obja, result));
8191 movdqu(rymm1, Address(objb, result));
8192 pxor(rymm0, rymm1);
8193 ptest(rymm0, rymm0);
8194 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
8195 addq(result, 16);
8196 subq(length, 16);
8197 jccb(Assembler::greaterEqual, VECTOR16_LOOP);
8198 addq(length, 16);
8199 jcc(Assembler::equal, SAME_TILL_END);
8200 //falling through if less than 16 bytes left
8201 }
8202
8203 bind(VECTOR8_TAIL);
8204 cmpq(length, 8);
8205 jccb(Assembler::less, VECTOR4_TAIL);
8206 bind(VECTOR8_LOOP);
8207 movq(tmp1, Address(obja, result));
8208 movq(tmp2, Address(objb, result));
8209 xorq(tmp1, tmp2);
8210 testq(tmp1, tmp1);
8211 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found
8212 addq(result, 8);
8213 subq(length, 8);
8214 jcc(Assembler::equal, SAME_TILL_END);
8215 //falling through if less than 8 bytes left
8216
8217 bind(VECTOR4_TAIL);
8218 cmpq(length, 4);
8219 jccb(Assembler::less, BYTES_TAIL);
8220 bind(VECTOR4_LOOP);
8221 movl(tmp1, Address(obja, result));
8222 xorl(tmp1, Address(objb, result));
8223 testl(tmp1, tmp1);
8224 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found
8225 addq(result, 4);
8226 subq(length, 4);
8227 jcc(Assembler::equal, SAME_TILL_END);
8228 //falling through if less than 4 bytes left
8229
8230 bind(BYTES_TAIL);
8231 bind(BYTES_LOOP);
8232 load_unsigned_byte(tmp1, Address(obja, result));
8233 load_unsigned_byte(tmp2, Address(objb, result));
8234 xorl(tmp1, tmp2);
8235 testl(tmp1, tmp1);
8236 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8237 decq(length);
8238 jcc(Assembler::zero, SAME_TILL_END);
8239 incq(result);
8240 load_unsigned_byte(tmp1, Address(obja, result));
8241 load_unsigned_byte(tmp2, Address(objb, result));
8242 xorl(tmp1, tmp2);
8243 testl(tmp1, tmp1);
8244 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8245 decq(length);
8246 jcc(Assembler::zero, SAME_TILL_END);
8247 incq(result);
8248 load_unsigned_byte(tmp1, Address(obja, result));
8249 load_unsigned_byte(tmp2, Address(objb, result));
8250 xorl(tmp1, tmp2);
8251 testl(tmp1, tmp1);
8252 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8253 jmp(SAME_TILL_END);
8254
8255 if (UseAVX >= 2) {
8256 bind(VECTOR32_NOT_EQUAL);
8257 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit);
8258 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit);
8259 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit);
8260 vpmovmskb(tmp1, rymm0);
8261 bsfq(tmp1, tmp1);
8262 addq(result, tmp1);
8263 shrq(result);
8264 jmp(DONE);
8265 }
8266
8267 bind(VECTOR16_NOT_EQUAL);
8268 if (UseAVX >= 2) {
8269 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit);
8270 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit);
8271 pxor(rymm0, rymm2);
8272 } else {
8273 pcmpeqb(rymm2, rymm2);
8274 pxor(rymm0, rymm1);
8275 pcmpeqb(rymm0, rymm1);
8276 pxor(rymm0, rymm2);
8277 }
8278 pmovmskb(tmp1, rymm0);
8279 bsfq(tmp1, tmp1);
8280 addq(result, tmp1);
8281 shrq(result);
8282 jmpb(DONE);
8283
8284 bind(VECTOR8_NOT_EQUAL);
8285 bind(VECTOR4_NOT_EQUAL);
8286 bsfq(tmp1, tmp1);
8287 shrq(tmp1, 3);
8288 addq(result, tmp1);
8289 bind(BYTES_NOT_EQUAL);
8290 shrq(result);
8291 jmpb(DONE);
8292
8293 bind(SAME_TILL_END);
8294 mov64(result, -1);
8295
8296 bind(DONE);
8297 }
8298
8299 //Helper functions for square_to_len()
8300
8301 /**
8302 * Store the squares of x[], right shifted one bit (divided by 2) into z[]
8303 * Preserves x and z and modifies rest of the registers.
8304 */
8305 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8306 // Perform square and right shift by 1
8307 // Handle odd xlen case first, then for even xlen do the following
8308 // jlong carry = 0;
8309 // for (int j=0, i=0; j < xlen; j+=2, i+=4) {
8310 // huge_128 product = x[j:j+1] * x[j:j+1];
8311 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65);
8312 // z[i+2:i+3] = (jlong)(product >>> 1);
8313 // carry = (jlong)product;
8314 // }
8315
8316 xorq(tmp5, tmp5); // carry
8317 xorq(rdxReg, rdxReg);
8318 xorl(tmp1, tmp1); // index for x
8319 xorl(tmp4, tmp4); // index for z
8320
8321 Label L_first_loop, L_first_loop_exit;
8322
8323 testl(xlen, 1);
8324 jccb(Assembler::zero, L_first_loop); //jump if xlen is even
8325
8326 // Square and right shift by 1 the odd element using 32 bit multiply
8327 movl(raxReg, Address(x, tmp1, Address::times_4, 0));
8328 imulq(raxReg, raxReg);
8329 shrq(raxReg, 1);
8330 adcq(tmp5, 0);
8331 movq(Address(z, tmp4, Address::times_4, 0), raxReg);
8332 incrementl(tmp1);
8333 addl(tmp4, 2);
8334
8335 // Square and right shift by 1 the rest using 64 bit multiply
8336 bind(L_first_loop);
8337 cmpptr(tmp1, xlen);
8338 jccb(Assembler::equal, L_first_loop_exit);
8339
8340 // Square
8341 movq(raxReg, Address(x, tmp1, Address::times_4, 0));
8342 rorq(raxReg, 32); // convert big-endian to little-endian
8343 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax
8344
8345 // Right shift by 1 and save carry
8346 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1
8347 rcrq(rdxReg, 1);
8348 rcrq(raxReg, 1);
8349 adcq(tmp5, 0);
8350
8351 // Store result in z
8352 movq(Address(z, tmp4, Address::times_4, 0), rdxReg);
8353 movq(Address(z, tmp4, Address::times_4, 8), raxReg);
8354
8355 // Update indices for x and z
8356 addl(tmp1, 2);
8357 addl(tmp4, 4);
8358 jmp(L_first_loop);
8359
8360 bind(L_first_loop_exit);
8361 }
8362
8363
8364 /**
8365 * Perform the following multiply add operation using BMI2 instructions
8366 * carry:sum = sum + op1*op2 + carry
8367 * op2 should be in rdx
8368 * op2 is preserved, all other registers are modified
8369 */
8370 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) {
8371 // assert op2 is rdx
8372 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1
8373 addq(sum, carry);
8374 adcq(tmp2, 0);
8375 addq(sum, op1);
8376 adcq(tmp2, 0);
8377 movq(carry, tmp2);
8378 }
8379
8380 /**
8381 * Perform the following multiply add operation:
8382 * carry:sum = sum + op1*op2 + carry
8383 * Preserves op1, op2 and modifies rest of registers
8384 */
8385 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) {
8386 // rdx:rax = op1 * op2
8387 movq(raxReg, op2);
8388 mulq(op1);
8389
8390 // rdx:rax = sum + carry + rdx:rax
8391 addq(sum, carry);
8392 adcq(rdxReg, 0);
8393 addq(sum, raxReg);
8394 adcq(rdxReg, 0);
8395
8396 // carry:sum = rdx:sum
8397 movq(carry, rdxReg);
8398 }
8399
8400 /**
8401 * Add 64 bit long carry into z[] with carry propogation.
8402 * Preserves z and carry register values and modifies rest of registers.
8403 *
8404 */
8405 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) {
8406 Label L_fourth_loop, L_fourth_loop_exit;
8407
8408 movl(tmp1, 1);
8409 subl(zlen, 2);
8410 addq(Address(z, zlen, Address::times_4, 0), carry);
8411
8412 bind(L_fourth_loop);
8413 jccb(Assembler::carryClear, L_fourth_loop_exit);
8414 subl(zlen, 2);
8415 jccb(Assembler::negative, L_fourth_loop_exit);
8416 addq(Address(z, zlen, Address::times_4, 0), tmp1);
8417 jmp(L_fourth_loop);
8418 bind(L_fourth_loop_exit);
8419 }
8420
8421 /**
8422 * Shift z[] left by 1 bit.
8423 * Preserves x, len, z and zlen registers and modifies rest of the registers.
8424 *
8425 */
8426 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) {
8427
8428 Label L_fifth_loop, L_fifth_loop_exit;
8429
8430 // Fifth loop
8431 // Perform primitiveLeftShift(z, zlen, 1)
8432
8433 const Register prev_carry = tmp1;
8434 const Register new_carry = tmp4;
8435 const Register value = tmp2;
8436 const Register zidx = tmp3;
8437
8438 // int zidx, carry;
8439 // long value;
8440 // carry = 0;
8441 // for (zidx = zlen-2; zidx >=0; zidx -= 2) {
8442 // (carry:value) = (z[i] << 1) | carry ;
8443 // z[i] = value;
8444 // }
8445
8446 movl(zidx, zlen);
8447 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register
8448
8449 bind(L_fifth_loop);
8450 decl(zidx); // Use decl to preserve carry flag
8451 decl(zidx);
8452 jccb(Assembler::negative, L_fifth_loop_exit);
8453
8454 if (UseBMI2Instructions) {
8455 movq(value, Address(z, zidx, Address::times_4, 0));
8456 rclq(value, 1);
8457 rorxq(value, value, 32);
8458 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
8459 }
8460 else {
8461 // clear new_carry
8462 xorl(new_carry, new_carry);
8463
8464 // Shift z[i] by 1, or in previous carry and save new carry
8465 movq(value, Address(z, zidx, Address::times_4, 0));
8466 shlq(value, 1);
8467 adcl(new_carry, 0);
8468
8469 orq(value, prev_carry);
8470 rorq(value, 0x20);
8471 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
8472
8473 // Set previous carry = new carry
8474 movl(prev_carry, new_carry);
8475 }
8476 jmp(L_fifth_loop);
8477
8478 bind(L_fifth_loop_exit);
8479 }
8480
8481
8482 /**
8483 * Code for BigInteger::squareToLen() intrinsic
8484 *
8485 * rdi: x
8486 * rsi: len
8487 * r8: z
8488 * rcx: zlen
8489 * r12: tmp1
8490 * r13: tmp2
8491 * r14: tmp3
8492 * r15: tmp4
8493 * rbx: tmp5
8494 *
8495 */
8496 void MacroAssembler::square_to_len(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8497
8498 Label L_second_loop, L_second_loop_exit, L_third_loop, L_third_loop_exit, L_last_x, L_multiply;
8499 push(tmp1);
8500 push(tmp2);
8501 push(tmp3);
8502 push(tmp4);
8503 push(tmp5);
8504
8505 // First loop
8506 // Store the squares, right shifted one bit (i.e., divided by 2).
8507 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg);
8508
8509 // Add in off-diagonal sums.
8510 //
8511 // Second, third (nested) and fourth loops.
8512 // zlen +=2;
8513 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) {
8514 // carry = 0;
8515 // long op2 = x[xidx:xidx+1];
8516 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) {
8517 // k -= 2;
8518 // long op1 = x[j:j+1];
8519 // long sum = z[k:k+1];
8520 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs);
8521 // z[k:k+1] = sum;
8522 // }
8523 // add_one_64(z, k, carry, tmp_regs);
8524 // }
8525
8526 const Register carry = tmp5;
8527 const Register sum = tmp3;
8528 const Register op1 = tmp4;
8529 Register op2 = tmp2;
8530
8531 push(zlen);
8532 push(len);
8533 addl(zlen,2);
8534 bind(L_second_loop);
8535 xorq(carry, carry);
8536 subl(zlen, 4);
8537 subl(len, 2);
8538 push(zlen);
8539 push(len);
8540 cmpl(len, 0);
8541 jccb(Assembler::lessEqual, L_second_loop_exit);
8542
8543 // Multiply an array by one 64 bit long.
8544 if (UseBMI2Instructions) {
8545 op2 = rdxReg;
8546 movq(op2, Address(x, len, Address::times_4, 0));
8547 rorxq(op2, op2, 32);
8548 }
8549 else {
8550 movq(op2, Address(x, len, Address::times_4, 0));
8551 rorq(op2, 32);
8552 }
8553
8554 bind(L_third_loop);
8555 decrementl(len);
8556 jccb(Assembler::negative, L_third_loop_exit);
8557 decrementl(len);
8558 jccb(Assembler::negative, L_last_x);
8559
8560 movq(op1, Address(x, len, Address::times_4, 0));
8561 rorq(op1, 32);
8562
8563 bind(L_multiply);
8564 subl(zlen, 2);
8565 movq(sum, Address(z, zlen, Address::times_4, 0));
8566
8567 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry.
8568 if (UseBMI2Instructions) {
8569 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2);
8570 }
8571 else {
8572 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8573 }
8574
8575 movq(Address(z, zlen, Address::times_4, 0), sum);
8576
8577 jmp(L_third_loop);
8578 bind(L_third_loop_exit);
8579
8580 // Fourth loop
8581 // Add 64 bit long carry into z with carry propogation.
8582 // Uses offsetted zlen.
8583 add_one_64(z, zlen, carry, tmp1);
8584
8585 pop(len);
8586 pop(zlen);
8587 jmp(L_second_loop);
8588
8589 // Next infrequent code is moved outside loops.
8590 bind(L_last_x);
8591 movl(op1, Address(x, 0));
8592 jmp(L_multiply);
8593
8594 bind(L_second_loop_exit);
8595 pop(len);
8596 pop(zlen);
8597 pop(len);
8598 pop(zlen);
8599
8600 // Fifth loop
8601 // Shift z left 1 bit.
8602 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4);
8603
8604 // z[zlen-1] |= x[len-1] & 1;
8605 movl(tmp3, Address(x, len, Address::times_4, -4));
8606 andl(tmp3, 1);
8607 orl(Address(z, zlen, Address::times_4, -4), tmp3);
8608
8609 pop(tmp5);
8610 pop(tmp4);
8611 pop(tmp3);
8612 pop(tmp2);
8613 pop(tmp1);
8614 }
8615
8616 /**
8617 * Helper function for mul_add()
8618 * Multiply the in[] by int k and add to out[] starting at offset offs using
8619 * 128 bit by 32 bit multiply and return the carry in tmp5.
8620 * Only quad int aligned length of in[] is operated on in this function.
8621 * k is in rdxReg for BMI2Instructions, for others it is in tmp2.
8622 * This function preserves out, in and k registers.
8623 * len and offset point to the appropriate index in "in" & "out" correspondingly
8624 * tmp5 has the carry.
8625 * other registers are temporary and are modified.
8626 *
8627 */
8628 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in,
8629 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3,
8630 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8631
8632 Label L_first_loop, L_first_loop_exit;
8633
8634 movl(tmp1, len);
8635 shrl(tmp1, 2);
8636
8637 bind(L_first_loop);
8638 subl(tmp1, 1);
8639 jccb(Assembler::negative, L_first_loop_exit);
8640
8641 subl(len, 4);
8642 subl(offset, 4);
8643
8644 Register op2 = tmp2;
8645 const Register sum = tmp3;
8646 const Register op1 = tmp4;
8647 const Register carry = tmp5;
8648
8649 if (UseBMI2Instructions) {
8650 op2 = rdxReg;
8651 }
8652
8653 movq(op1, Address(in, len, Address::times_4, 8));
8654 rorq(op1, 32);
8655 movq(sum, Address(out, offset, Address::times_4, 8));
8656 rorq(sum, 32);
8657 if (UseBMI2Instructions) {
8658 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8659 }
8660 else {
8661 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8662 }
8663 // Store back in big endian from little endian
8664 rorq(sum, 0x20);
8665 movq(Address(out, offset, Address::times_4, 8), sum);
8666
8667 movq(op1, Address(in, len, Address::times_4, 0));
8668 rorq(op1, 32);
8669 movq(sum, Address(out, offset, Address::times_4, 0));
8670 rorq(sum, 32);
8671 if (UseBMI2Instructions) {
8672 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8673 }
8674 else {
8675 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8676 }
8677 // Store back in big endian from little endian
8678 rorq(sum, 0x20);
8679 movq(Address(out, offset, Address::times_4, 0), sum);
8680
8681 jmp(L_first_loop);
8682 bind(L_first_loop_exit);
8683 }
8684
8685 /**
8686 * Code for BigInteger::mulAdd() intrinsic
8687 *
8688 * rdi: out
8689 * rsi: in
8690 * r11: offs (out.length - offset)
8691 * rcx: len
8692 * r8: k
8693 * r12: tmp1
8694 * r13: tmp2
8695 * r14: tmp3
8696 * r15: tmp4
8697 * rbx: tmp5
8698 * Multiply the in[] by word k and add to out[], return the carry in rax
8699 */
8700 void MacroAssembler::mul_add(Register out, Register in, Register offs,
8701 Register len, Register k, Register tmp1, Register tmp2, Register tmp3,
8702 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8703
8704 Label L_carry, L_last_in, L_done;
8705
8706 // carry = 0;
8707 // for (int j=len-1; j >= 0; j--) {
8708 // long product = (in[j] & LONG_MASK) * kLong +
8709 // (out[offs] & LONG_MASK) + carry;
8710 // out[offs--] = (int)product;
8711 // carry = product >>> 32;
8712 // }
8713 //
8714 push(tmp1);
8715 push(tmp2);
8716 push(tmp3);
8717 push(tmp4);
8718 push(tmp5);
8719
8720 Register op2 = tmp2;
8721 const Register sum = tmp3;
8722 const Register op1 = tmp4;
8723 const Register carry = tmp5;
8724
8725 if (UseBMI2Instructions) {
8726 op2 = rdxReg;
8727 movl(op2, k);
8728 }
8729 else {
8730 movl(op2, k);
8731 }
8732
8733 xorq(carry, carry);
8734
8735 //First loop
8736
8737 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply
8738 //The carry is in tmp5
8739 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg);
8740
8741 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any
8742 decrementl(len);
8743 jccb(Assembler::negative, L_carry);
8744 decrementl(len);
8745 jccb(Assembler::negative, L_last_in);
8746
8747 movq(op1, Address(in, len, Address::times_4, 0));
8748 rorq(op1, 32);
8749
8750 subl(offs, 2);
8751 movq(sum, Address(out, offs, Address::times_4, 0));
8752 rorq(sum, 32);
8753
8754 if (UseBMI2Instructions) {
8755 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8756 }
8757 else {
8758 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8759 }
8760
8761 // Store back in big endian from little endian
8762 rorq(sum, 0x20);
8763 movq(Address(out, offs, Address::times_4, 0), sum);
8764
8765 testl(len, len);
8766 jccb(Assembler::zero, L_carry);
8767
8768 //Multiply the last in[] entry, if any
8769 bind(L_last_in);
8770 movl(op1, Address(in, 0));
8771 movl(sum, Address(out, offs, Address::times_4, -4));
8772
8773 movl(raxReg, k);
8774 mull(op1); //tmp4 * eax -> edx:eax
8775 addl(sum, carry);
8776 adcl(rdxReg, 0);
8777 addl(sum, raxReg);
8778 adcl(rdxReg, 0);
8779 movl(carry, rdxReg);
8780
8781 movl(Address(out, offs, Address::times_4, -4), sum);
8782
8783 bind(L_carry);
8784 //return tmp5/carry as carry in rax
8785 movl(rax, carry);
8786
8787 bind(L_done);
8788 pop(tmp5);
8789 pop(tmp4);
8790 pop(tmp3);
8791 pop(tmp2);
8792 pop(tmp1);
8793 }
8794 #endif
8795
8796 /**
8797 * Emits code to update CRC-32 with a byte value according to constants in table
8798 *
8799 * @param [in,out]crc Register containing the crc.
8800 * @param [in]val Register containing the byte to fold into the CRC.
8801 * @param [in]table Register containing the table of crc constants.
8802 *
8803 * uint32_t crc;
8804 * val = crc_table[(val ^ crc) & 0xFF];
8805 * crc = val ^ (crc >> 8);
8806 *
8807 */
8808 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
8809 xorl(val, crc);
8810 andl(val, 0xFF);
8811 shrl(crc, 8); // unsigned shift
8812 xorl(crc, Address(table, val, Address::times_4, 0));
8813 }
8814
8815 /**
8816 * Fold four 128-bit data chunks
8817 */
8818 void MacroAssembler::fold_128bit_crc32_avx512(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
8819 evpclmulhdq(xtmp, xK, xcrc, Assembler::AVX_512bit); // [123:64]
8820 evpclmulldq(xcrc, xK, xcrc, Assembler::AVX_512bit); // [63:0]
8821 evpxorq(xcrc, xcrc, Address(buf, offset), Assembler::AVX_512bit /* vector_len */);
8822 evpxorq(xcrc, xcrc, xtmp, Assembler::AVX_512bit /* vector_len */);
8823 }
8824
8825 /**
8826 * Fold 128-bit data chunk
8827 */
8828 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
8829 if (UseAVX > 0) {
8830 vpclmulhdq(xtmp, xK, xcrc); // [123:64]
8831 vpclmulldq(xcrc, xK, xcrc); // [63:0]
8832 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */);
8833 pxor(xcrc, xtmp);
8834 } else {
8835 movdqa(xtmp, xcrc);
8836 pclmulhdq(xtmp, xK); // [123:64]
8837 pclmulldq(xcrc, xK); // [63:0]
8838 pxor(xcrc, xtmp);
8839 movdqu(xtmp, Address(buf, offset));
8840 pxor(xcrc, xtmp);
8841 }
8842 }
8843
8844 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
8845 if (UseAVX > 0) {
8846 vpclmulhdq(xtmp, xK, xcrc);
8847 vpclmulldq(xcrc, xK, xcrc);
8848 pxor(xcrc, xbuf);
8849 pxor(xcrc, xtmp);
8850 } else {
8851 movdqa(xtmp, xcrc);
8852 pclmulhdq(xtmp, xK);
8853 pclmulldq(xcrc, xK);
8854 pxor(xcrc, xbuf);
8855 pxor(xcrc, xtmp);
8856 }
8857 }
8858
8859 /**
8860 * 8-bit folds to compute 32-bit CRC
8861 *
8862 * uint64_t xcrc;
8863 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8);
8864 */
8865 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) {
8866 movdl(tmp, xcrc);
8867 andl(tmp, 0xFF);
8868 movdl(xtmp, Address(table, tmp, Address::times_4, 0));
8869 psrldq(xcrc, 1); // unsigned shift one byte
8870 pxor(xcrc, xtmp);
8871 }
8872
8873 /**
8874 * uint32_t crc;
8875 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
8876 */
8877 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
8878 movl(tmp, crc);
8879 andl(tmp, 0xFF);
8880 shrl(crc, 8);
8881 xorl(crc, Address(table, tmp, Address::times_4, 0));
8882 }
8883
8884 /**
8885 * @param crc register containing existing CRC (32-bit)
8886 * @param buf register pointing to input byte buffer (byte*)
8887 * @param len register containing number of bytes
8888 * @param table register that will contain address of CRC table
8889 * @param tmp scratch register
8890 */
8891 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) {
8892 assert_different_registers(crc, buf, len, table, tmp, rax);
8893
8894 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned;
8895 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop;
8896
8897 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge
8898 // context for the registers used, where all instructions below are using 128-bit mode
8899 // On EVEX without VL and BW, these instructions will all be AVX.
8900 lea(table, ExternalAddress(StubRoutines::crc_table_addr()));
8901 notl(crc); // ~crc
8902 cmpl(len, 16);
8903 jcc(Assembler::less, L_tail);
8904
8905 // Align buffer to 16 bytes
8906 movl(tmp, buf);
8907 andl(tmp, 0xF);
8908 jccb(Assembler::zero, L_aligned);
8909 subl(tmp, 16);
8910 addl(len, tmp);
8911
8912 align(4);
8913 BIND(L_align_loop);
8914 movsbl(rax, Address(buf, 0)); // load byte with sign extension
8915 update_byte_crc32(crc, rax, table);
8916 increment(buf);
8917 incrementl(tmp);
8918 jccb(Assembler::less, L_align_loop);
8919
8920 BIND(L_aligned);
8921 movl(tmp, len); // save
8922 shrl(len, 4);
8923 jcc(Assembler::zero, L_tail_restore);
8924
8925 // Fold crc into first bytes of vector
8926 movdqa(xmm1, Address(buf, 0));
8927 movdl(rax, xmm1);
8928 xorl(crc, rax);
8929 if (VM_Version::supports_sse4_1()) {
8930 pinsrd(xmm1, crc, 0);
8931 } else {
8932 pinsrw(xmm1, crc, 0);
8933 shrl(crc, 16);
8934 pinsrw(xmm1, crc, 1);
8935 }
8936 addptr(buf, 16);
8937 subl(len, 4); // len > 0
8938 jcc(Assembler::less, L_fold_tail);
8939
8940 movdqa(xmm2, Address(buf, 0));
8941 movdqa(xmm3, Address(buf, 16));
8942 movdqa(xmm4, Address(buf, 32));
8943 addptr(buf, 48);
8944 subl(len, 3);
8945 jcc(Assembler::lessEqual, L_fold_512b);
8946
8947 // Fold total 512 bits of polynomial on each iteration,
8948 // 128 bits per each of 4 parallel streams.
8949 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
8950
8951 align(32);
8952 BIND(L_fold_512b_loop);
8953 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
8954 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16);
8955 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32);
8956 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48);
8957 addptr(buf, 64);
8958 subl(len, 4);
8959 jcc(Assembler::greater, L_fold_512b_loop);
8960
8961 // Fold 512 bits to 128 bits.
8962 BIND(L_fold_512b);
8963 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
8964 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2);
8965 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3);
8966 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4);
8967
8968 // Fold the rest of 128 bits data chunks
8969 BIND(L_fold_tail);
8970 addl(len, 3);
8971 jccb(Assembler::lessEqual, L_fold_128b);
8972 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
8973
8974 BIND(L_fold_tail_loop);
8975 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
8976 addptr(buf, 16);
8977 decrementl(len);
8978 jccb(Assembler::greater, L_fold_tail_loop);
8979
8980 // Fold 128 bits in xmm1 down into 32 bits in crc register.
8981 BIND(L_fold_128b);
8982 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
8983 if (UseAVX > 0) {
8984 vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
8985 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */);
8986 vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
8987 } else {
8988 movdqa(xmm2, xmm0);
8989 pclmulqdq(xmm2, xmm1, 0x1);
8990 movdqa(xmm3, xmm0);
8991 pand(xmm3, xmm2);
8992 pclmulqdq(xmm0, xmm3, 0x1);
8993 }
8994 psrldq(xmm1, 8);
8995 psrldq(xmm2, 4);
8996 pxor(xmm0, xmm1);
8997 pxor(xmm0, xmm2);
8998
8999 // 8 8-bit folds to compute 32-bit CRC.
9000 for (int j = 0; j < 4; j++) {
9001 fold_8bit_crc32(xmm0, table, xmm1, rax);
9002 }
9003 movdl(crc, xmm0); // mov 32 bits to general register
9004 for (int j = 0; j < 4; j++) {
9005 fold_8bit_crc32(crc, table, rax);
9006 }
9007
9008 BIND(L_tail_restore);
9009 movl(len, tmp); // restore
9010 BIND(L_tail);
9011 andl(len, 0xf);
9012 jccb(Assembler::zero, L_exit);
9013
9014 // Fold the rest of bytes
9015 align(4);
9016 BIND(L_tail_loop);
9017 movsbl(rax, Address(buf, 0)); // load byte with sign extension
9018 update_byte_crc32(crc, rax, table);
9019 increment(buf);
9020 decrementl(len);
9021 jccb(Assembler::greater, L_tail_loop);
9022
9023 BIND(L_exit);
9024 notl(crc); // ~c
9025 }
9026
9027 #ifdef _LP64
9028 // S. Gueron / Information Processing Letters 112 (2012) 184
9029 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table.
9030 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0].
9031 // Output: the 64-bit carry-less product of B * CONST
9032 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n,
9033 Register tmp1, Register tmp2, Register tmp3) {
9034 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
9035 if (n > 0) {
9036 addq(tmp3, n * 256 * 8);
9037 }
9038 // Q1 = TABLEExt[n][B & 0xFF];
9039 movl(tmp1, in);
9040 andl(tmp1, 0x000000FF);
9041 shll(tmp1, 3);
9042 addq(tmp1, tmp3);
9043 movq(tmp1, Address(tmp1, 0));
9044
9045 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
9046 movl(tmp2, in);
9047 shrl(tmp2, 8);
9048 andl(tmp2, 0x000000FF);
9049 shll(tmp2, 3);
9050 addq(tmp2, tmp3);
9051 movq(tmp2, Address(tmp2, 0));
9052
9053 shlq(tmp2, 8);
9054 xorq(tmp1, tmp2);
9055
9056 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
9057 movl(tmp2, in);
9058 shrl(tmp2, 16);
9059 andl(tmp2, 0x000000FF);
9060 shll(tmp2, 3);
9061 addq(tmp2, tmp3);
9062 movq(tmp2, Address(tmp2, 0));
9063
9064 shlq(tmp2, 16);
9065 xorq(tmp1, tmp2);
9066
9067 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
9068 shrl(in, 24);
9069 andl(in, 0x000000FF);
9070 shll(in, 3);
9071 addq(in, tmp3);
9072 movq(in, Address(in, 0));
9073
9074 shlq(in, 24);
9075 xorq(in, tmp1);
9076 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
9077 }
9078
9079 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
9080 Register in_out,
9081 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
9082 XMMRegister w_xtmp2,
9083 Register tmp1,
9084 Register n_tmp2, Register n_tmp3) {
9085 if (is_pclmulqdq_supported) {
9086 movdl(w_xtmp1, in_out); // modified blindly
9087
9088 movl(tmp1, const_or_pre_comp_const_index);
9089 movdl(w_xtmp2, tmp1);
9090 pclmulqdq(w_xtmp1, w_xtmp2, 0);
9091
9092 movdq(in_out, w_xtmp1);
9093 } else {
9094 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3);
9095 }
9096 }
9097
9098 // Recombination Alternative 2: No bit-reflections
9099 // T1 = (CRC_A * U1) << 1
9100 // T2 = (CRC_B * U2) << 1
9101 // C1 = T1 >> 32
9102 // C2 = T2 >> 32
9103 // T1 = T1 & 0xFFFFFFFF
9104 // T2 = T2 & 0xFFFFFFFF
9105 // T1 = CRC32(0, T1)
9106 // T2 = CRC32(0, T2)
9107 // C1 = C1 ^ T1
9108 // C2 = C2 ^ T2
9109 // CRC = C1 ^ C2 ^ CRC_C
9110 void MacroAssembler::crc32c_rec_alt2(uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, Register in_out, Register in1, Register in2,
9111 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9112 Register tmp1, Register tmp2,
9113 Register n_tmp3) {
9114 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9115 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9116 shlq(in_out, 1);
9117 movl(tmp1, in_out);
9118 shrq(in_out, 32);
9119 xorl(tmp2, tmp2);
9120 crc32(tmp2, tmp1, 4);
9121 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here
9122 shlq(in1, 1);
9123 movl(tmp1, in1);
9124 shrq(in1, 32);
9125 xorl(tmp2, tmp2);
9126 crc32(tmp2, tmp1, 4);
9127 xorl(in1, tmp2);
9128 xorl(in_out, in1);
9129 xorl(in_out, in2);
9130 }
9131
9132 // Set N to predefined value
9133 // Subtract from a lenght of a buffer
9134 // execute in a loop:
9135 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0
9136 // for i = 1 to N do
9137 // CRC_A = CRC32(CRC_A, A[i])
9138 // CRC_B = CRC32(CRC_B, B[i])
9139 // CRC_C = CRC32(CRC_C, C[i])
9140 // end for
9141 // Recombine
9142 void MacroAssembler::crc32c_proc_chunk(uint32_t size, uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported,
9143 Register in_out1, Register in_out2, Register in_out3,
9144 Register tmp1, Register tmp2, Register tmp3,
9145 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9146 Register tmp4, Register tmp5,
9147 Register n_tmp6) {
9148 Label L_processPartitions;
9149 Label L_processPartition;
9150 Label L_exit;
9151
9152 bind(L_processPartitions);
9153 cmpl(in_out1, 3 * size);
9154 jcc(Assembler::less, L_exit);
9155 xorl(tmp1, tmp1);
9156 xorl(tmp2, tmp2);
9157 movq(tmp3, in_out2);
9158 addq(tmp3, size);
9159
9160 bind(L_processPartition);
9161 crc32(in_out3, Address(in_out2, 0), 8);
9162 crc32(tmp1, Address(in_out2, size), 8);
9163 crc32(tmp2, Address(in_out2, size * 2), 8);
9164 addq(in_out2, 8);
9165 cmpq(in_out2, tmp3);
9166 jcc(Assembler::less, L_processPartition);
9167 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
9168 w_xtmp1, w_xtmp2, w_xtmp3,
9169 tmp4, tmp5,
9170 n_tmp6);
9171 addq(in_out2, 2 * size);
9172 subl(in_out1, 3 * size);
9173 jmp(L_processPartitions);
9174
9175 bind(L_exit);
9176 }
9177 #else
9178 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n,
9179 Register tmp1, Register tmp2, Register tmp3,
9180 XMMRegister xtmp1, XMMRegister xtmp2) {
9181 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
9182 if (n > 0) {
9183 addl(tmp3, n * 256 * 8);
9184 }
9185 // Q1 = TABLEExt[n][B & 0xFF];
9186 movl(tmp1, in_out);
9187 andl(tmp1, 0x000000FF);
9188 shll(tmp1, 3);
9189 addl(tmp1, tmp3);
9190 movq(xtmp1, Address(tmp1, 0));
9191
9192 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
9193 movl(tmp2, in_out);
9194 shrl(tmp2, 8);
9195 andl(tmp2, 0x000000FF);
9196 shll(tmp2, 3);
9197 addl(tmp2, tmp3);
9198 movq(xtmp2, Address(tmp2, 0));
9199
9200 psllq(xtmp2, 8);
9201 pxor(xtmp1, xtmp2);
9202
9203 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
9204 movl(tmp2, in_out);
9205 shrl(tmp2, 16);
9206 andl(tmp2, 0x000000FF);
9207 shll(tmp2, 3);
9208 addl(tmp2, tmp3);
9209 movq(xtmp2, Address(tmp2, 0));
9210
9211 psllq(xtmp2, 16);
9212 pxor(xtmp1, xtmp2);
9213
9214 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
9215 shrl(in_out, 24);
9216 andl(in_out, 0x000000FF);
9217 shll(in_out, 3);
9218 addl(in_out, tmp3);
9219 movq(xtmp2, Address(in_out, 0));
9220
9221 psllq(xtmp2, 24);
9222 pxor(xtmp1, xtmp2); // Result in CXMM
9223 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
9224 }
9225
9226 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
9227 Register in_out,
9228 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
9229 XMMRegister w_xtmp2,
9230 Register tmp1,
9231 Register n_tmp2, Register n_tmp3) {
9232 if (is_pclmulqdq_supported) {
9233 movdl(w_xtmp1, in_out);
9234
9235 movl(tmp1, const_or_pre_comp_const_index);
9236 movdl(w_xtmp2, tmp1);
9237 pclmulqdq(w_xtmp1, w_xtmp2, 0);
9238 // Keep result in XMM since GPR is 32 bit in length
9239 } else {
9240 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2);
9241 }
9242 }
9243
9244 void MacroAssembler::crc32c_rec_alt2(uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported, Register in_out, Register in1, Register in2,
9245 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9246 Register tmp1, Register tmp2,
9247 Register n_tmp3) {
9248 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9249 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9250
9251 psllq(w_xtmp1, 1);
9252 movdl(tmp1, w_xtmp1);
9253 psrlq(w_xtmp1, 32);
9254 movdl(in_out, w_xtmp1);
9255
9256 xorl(tmp2, tmp2);
9257 crc32(tmp2, tmp1, 4);
9258 xorl(in_out, tmp2);
9259
9260 psllq(w_xtmp2, 1);
9261 movdl(tmp1, w_xtmp2);
9262 psrlq(w_xtmp2, 32);
9263 movdl(in1, w_xtmp2);
9264
9265 xorl(tmp2, tmp2);
9266 crc32(tmp2, tmp1, 4);
9267 xorl(in1, tmp2);
9268 xorl(in_out, in1);
9269 xorl(in_out, in2);
9270 }
9271
9272 void MacroAssembler::crc32c_proc_chunk(uint32_t size, uint32_t const_or_pre_comp_const_index_u1, uint32_t const_or_pre_comp_const_index_u2, bool is_pclmulqdq_supported,
9273 Register in_out1, Register in_out2, Register in_out3,
9274 Register tmp1, Register tmp2, Register tmp3,
9275 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9276 Register tmp4, Register tmp5,
9277 Register n_tmp6) {
9278 Label L_processPartitions;
9279 Label L_processPartition;
9280 Label L_exit;
9281
9282 bind(L_processPartitions);
9283 cmpl(in_out1, 3 * size);
9284 jcc(Assembler::less, L_exit);
9285 xorl(tmp1, tmp1);
9286 xorl(tmp2, tmp2);
9287 movl(tmp3, in_out2);
9288 addl(tmp3, size);
9289
9290 bind(L_processPartition);
9291 crc32(in_out3, Address(in_out2, 0), 4);
9292 crc32(tmp1, Address(in_out2, size), 4);
9293 crc32(tmp2, Address(in_out2, size*2), 4);
9294 crc32(in_out3, Address(in_out2, 0+4), 4);
9295 crc32(tmp1, Address(in_out2, size+4), 4);
9296 crc32(tmp2, Address(in_out2, size*2+4), 4);
9297 addl(in_out2, 8);
9298 cmpl(in_out2, tmp3);
9299 jcc(Assembler::less, L_processPartition);
9300
9301 push(tmp3);
9302 push(in_out1);
9303 push(in_out2);
9304 tmp4 = tmp3;
9305 tmp5 = in_out1;
9306 n_tmp6 = in_out2;
9307
9308 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
9309 w_xtmp1, w_xtmp2, w_xtmp3,
9310 tmp4, tmp5,
9311 n_tmp6);
9312
9313 pop(in_out2);
9314 pop(in_out1);
9315 pop(tmp3);
9316
9317 addl(in_out2, 2 * size);
9318 subl(in_out1, 3 * size);
9319 jmp(L_processPartitions);
9320
9321 bind(L_exit);
9322 }
9323 #endif //LP64
9324
9325 #ifdef _LP64
9326 // Algorithm 2: Pipelined usage of the CRC32 instruction.
9327 // Input: A buffer I of L bytes.
9328 // Output: the CRC32C value of the buffer.
9329 // Notations:
9330 // Write L = 24N + r, with N = floor (L/24).
9331 // r = L mod 24 (0 <= r < 24).
9332 // Consider I as the concatenation of A|B|C|R, where A, B, C, each,
9333 // N quadwords, and R consists of r bytes.
9334 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1
9335 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1
9336 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1
9337 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1
9338 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
9339 Register tmp1, Register tmp2, Register tmp3,
9340 Register tmp4, Register tmp5, Register tmp6,
9341 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9342 bool is_pclmulqdq_supported) {
9343 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
9344 Label L_wordByWord;
9345 Label L_byteByByteProlog;
9346 Label L_byteByByte;
9347 Label L_exit;
9348
9349 if (is_pclmulqdq_supported ) {
9350 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
9351 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1);
9352
9353 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
9354 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
9355
9356 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
9357 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
9358 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\"");
9359 } else {
9360 const_or_pre_comp_const_index[0] = 1;
9361 const_or_pre_comp_const_index[1] = 0;
9362
9363 const_or_pre_comp_const_index[2] = 3;
9364 const_or_pre_comp_const_index[3] = 2;
9365
9366 const_or_pre_comp_const_index[4] = 5;
9367 const_or_pre_comp_const_index[5] = 4;
9368 }
9369 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
9370 in2, in1, in_out,
9371 tmp1, tmp2, tmp3,
9372 w_xtmp1, w_xtmp2, w_xtmp3,
9373 tmp4, tmp5,
9374 tmp6);
9375 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
9376 in2, in1, in_out,
9377 tmp1, tmp2, tmp3,
9378 w_xtmp1, w_xtmp2, w_xtmp3,
9379 tmp4, tmp5,
9380 tmp6);
9381 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
9382 in2, in1, in_out,
9383 tmp1, tmp2, tmp3,
9384 w_xtmp1, w_xtmp2, w_xtmp3,
9385 tmp4, tmp5,
9386 tmp6);
9387 movl(tmp1, in2);
9388 andl(tmp1, 0x00000007);
9389 negl(tmp1);
9390 addl(tmp1, in2);
9391 addq(tmp1, in1);
9392
9393 BIND(L_wordByWord);
9394 cmpq(in1, tmp1);
9395 jcc(Assembler::greaterEqual, L_byteByByteProlog);
9396 crc32(in_out, Address(in1, 0), 4);
9397 addq(in1, 4);
9398 jmp(L_wordByWord);
9399
9400 BIND(L_byteByByteProlog);
9401 andl(in2, 0x00000007);
9402 movl(tmp2, 1);
9403
9404 BIND(L_byteByByte);
9405 cmpl(tmp2, in2);
9406 jccb(Assembler::greater, L_exit);
9407 crc32(in_out, Address(in1, 0), 1);
9408 incq(in1);
9409 incl(tmp2);
9410 jmp(L_byteByByte);
9411
9412 BIND(L_exit);
9413 }
9414 #else
9415 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
9416 Register tmp1, Register tmp2, Register tmp3,
9417 Register tmp4, Register tmp5, Register tmp6,
9418 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9419 bool is_pclmulqdq_supported) {
9420 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
9421 Label L_wordByWord;
9422 Label L_byteByByteProlog;
9423 Label L_byteByByte;
9424 Label L_exit;
9425
9426 if (is_pclmulqdq_supported) {
9427 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
9428 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1);
9429
9430 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
9431 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
9432
9433 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
9434 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
9435 } else {
9436 const_or_pre_comp_const_index[0] = 1;
9437 const_or_pre_comp_const_index[1] = 0;
9438
9439 const_or_pre_comp_const_index[2] = 3;
9440 const_or_pre_comp_const_index[3] = 2;
9441
9442 const_or_pre_comp_const_index[4] = 5;
9443 const_or_pre_comp_const_index[5] = 4;
9444 }
9445 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
9446 in2, in1, in_out,
9447 tmp1, tmp2, tmp3,
9448 w_xtmp1, w_xtmp2, w_xtmp3,
9449 tmp4, tmp5,
9450 tmp6);
9451 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
9452 in2, in1, in_out,
9453 tmp1, tmp2, tmp3,
9454 w_xtmp1, w_xtmp2, w_xtmp3,
9455 tmp4, tmp5,
9456 tmp6);
9457 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
9458 in2, in1, in_out,
9459 tmp1, tmp2, tmp3,
9460 w_xtmp1, w_xtmp2, w_xtmp3,
9461 tmp4, tmp5,
9462 tmp6);
9463 movl(tmp1, in2);
9464 andl(tmp1, 0x00000007);
9465 negl(tmp1);
9466 addl(tmp1, in2);
9467 addl(tmp1, in1);
9468
9469 BIND(L_wordByWord);
9470 cmpl(in1, tmp1);
9471 jcc(Assembler::greaterEqual, L_byteByByteProlog);
9472 crc32(in_out, Address(in1,0), 4);
9473 addl(in1, 4);
9474 jmp(L_wordByWord);
9475
9476 BIND(L_byteByByteProlog);
9477 andl(in2, 0x00000007);
9478 movl(tmp2, 1);
9479
9480 BIND(L_byteByByte);
9481 cmpl(tmp2, in2);
9482 jccb(Assembler::greater, L_exit);
9483 movb(tmp1, Address(in1, 0));
9484 crc32(in_out, tmp1, 1);
9485 incl(in1);
9486 incl(tmp2);
9487 jmp(L_byteByByte);
9488
9489 BIND(L_exit);
9490 }
9491 #endif // LP64
9492 #undef BIND
9493 #undef BLOCK_COMMENT
9494
9495 // Compress char[] array to byte[].
9496 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java
9497 // @HotSpotIntrinsicCandidate
9498 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) {
9499 // for (int i = 0; i < len; i++) {
9500 // int c = src[srcOff++];
9501 // if (c >>> 8 != 0) {
9502 // return 0;
9503 // }
9504 // dst[dstOff++] = (byte)c;
9505 // }
9506 // return len;
9507 // }
9508 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
9509 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
9510 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
9511 Register tmp5, Register result) {
9512 Label copy_chars_loop, return_length, return_zero, done;
9513
9514 // rsi: src
9515 // rdi: dst
9516 // rdx: len
9517 // rcx: tmp5
9518 // rax: result
9519
9520 // rsi holds start addr of source char[] to be compressed
9521 // rdi holds start addr of destination byte[]
9522 // rdx holds length
9523
9524 assert(len != result, "");
9525
9526 // save length for return
9527 push(len);
9528
9529 if ((AVX3Threshold == 0) && (UseAVX > 2) && // AVX512
9530 VM_Version::supports_avx512vlbw() &&
9531 VM_Version::supports_bmi2()) {
9532
9533 Label copy_32_loop, copy_loop_tail, below_threshold;
9534
9535 // alignment
9536 Label post_alignment;
9537
9538 // if length of the string is less than 16, handle it in an old fashioned way
9539 testl(len, -32);
9540 jcc(Assembler::zero, below_threshold);
9541
9542 // First check whether a character is compressable ( <= 0xFF).
9543 // Create mask to test for Unicode chars inside zmm vector
9544 movl(result, 0x00FF);
9545 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit);
9546
9547 testl(len, -64);
9548 jcc(Assembler::zero, post_alignment);
9549
9550 movl(tmp5, dst);
9551 andl(tmp5, (32 - 1));
9552 negl(tmp5);
9553 andl(tmp5, (32 - 1));
9554
9555 // bail out when there is nothing to be done
9556 testl(tmp5, 0xFFFFFFFF);
9557 jcc(Assembler::zero, post_alignment);
9558
9559 // ~(~0 << len), where len is the # of remaining elements to process
9560 movl(result, 0xFFFFFFFF);
9561 shlxl(result, result, tmp5);
9562 notl(result);
9563 kmovdl(k3, result);
9564
9565 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit);
9566 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9567 ktestd(k2, k3);
9568 jcc(Assembler::carryClear, return_zero);
9569
9570 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit);
9571
9572 addptr(src, tmp5);
9573 addptr(src, tmp5);
9574 addptr(dst, tmp5);
9575 subl(len, tmp5);
9576
9577 bind(post_alignment);
9578 // end of alignment
9579
9580 movl(tmp5, len);
9581 andl(tmp5, (32 - 1)); // tail count (in chars)
9582 andl(len, ~(32 - 1)); // vector count (in chars)
9583 jcc(Assembler::zero, copy_loop_tail);
9584
9585 lea(src, Address(src, len, Address::times_2));
9586 lea(dst, Address(dst, len, Address::times_1));
9587 negptr(len);
9588
9589 bind(copy_32_loop);
9590 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit);
9591 evpcmpuw(k2, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9592 kortestdl(k2, k2);
9593 jcc(Assembler::carryClear, return_zero);
9594
9595 // All elements in current processed chunk are valid candidates for
9596 // compression. Write a truncated byte elements to the memory.
9597 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit);
9598 addptr(len, 32);
9599 jcc(Assembler::notZero, copy_32_loop);
9600
9601 bind(copy_loop_tail);
9602 // bail out when there is nothing to be done
9603 testl(tmp5, 0xFFFFFFFF);
9604 jcc(Assembler::zero, return_length);
9605
9606 movl(len, tmp5);
9607
9608 // ~(~0 << len), where len is the # of remaining elements to process
9609 movl(result, 0xFFFFFFFF);
9610 shlxl(result, result, len);
9611 notl(result);
9612
9613 kmovdl(k3, result);
9614
9615 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit);
9616 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9617 ktestd(k2, k3);
9618 jcc(Assembler::carryClear, return_zero);
9619
9620 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit);
9621 jmp(return_length);
9622
9623 bind(below_threshold);
9624 }
9625
9626 if (UseSSE42Intrinsics) {
9627 Label copy_32_loop, copy_16, copy_tail;
9628
9629 movl(result, len);
9630
9631 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors
9632
9633 // vectored compression
9634 andl(len, 0xfffffff0); // vector count (in chars)
9635 andl(result, 0x0000000f); // tail count (in chars)
9636 testl(len, len);
9637 jcc(Assembler::zero, copy_16);
9638
9639 // compress 16 chars per iter
9640 movdl(tmp1Reg, tmp5);
9641 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
9642 pxor(tmp4Reg, tmp4Reg);
9643
9644 lea(src, Address(src, len, Address::times_2));
9645 lea(dst, Address(dst, len, Address::times_1));
9646 negptr(len);
9647
9648 bind(copy_32_loop);
9649 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters
9650 por(tmp4Reg, tmp2Reg);
9651 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters
9652 por(tmp4Reg, tmp3Reg);
9653 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector
9654 jcc(Assembler::notZero, return_zero);
9655 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte
9656 movdqu(Address(dst, len, Address::times_1), tmp2Reg);
9657 addptr(len, 16);
9658 jcc(Assembler::notZero, copy_32_loop);
9659
9660 // compress next vector of 8 chars (if any)
9661 bind(copy_16);
9662 movl(len, result);
9663 andl(len, 0xfffffff8); // vector count (in chars)
9664 andl(result, 0x00000007); // tail count (in chars)
9665 testl(len, len);
9666 jccb(Assembler::zero, copy_tail);
9667
9668 movdl(tmp1Reg, tmp5);
9669 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
9670 pxor(tmp3Reg, tmp3Reg);
9671
9672 movdqu(tmp2Reg, Address(src, 0));
9673 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
9674 jccb(Assembler::notZero, return_zero);
9675 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte
9676 movq(Address(dst, 0), tmp2Reg);
9677 addptr(src, 16);
9678 addptr(dst, 8);
9679
9680 bind(copy_tail);
9681 movl(len, result);
9682 }
9683 // compress 1 char per iter
9684 testl(len, len);
9685 jccb(Assembler::zero, return_length);
9686 lea(src, Address(src, len, Address::times_2));
9687 lea(dst, Address(dst, len, Address::times_1));
9688 negptr(len);
9689
9690 bind(copy_chars_loop);
9691 load_unsigned_short(result, Address(src, len, Address::times_2));
9692 testl(result, 0xff00); // check if Unicode char
9693 jccb(Assembler::notZero, return_zero);
9694 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte
9695 increment(len);
9696 jcc(Assembler::notZero, copy_chars_loop);
9697
9698 // if compression succeeded, return length
9699 bind(return_length);
9700 pop(result);
9701 jmpb(done);
9702
9703 // if compression failed, return 0
9704 bind(return_zero);
9705 xorl(result, result);
9706 addptr(rsp, wordSize);
9707
9708 bind(done);
9709 }
9710
9711 // Inflate byte[] array to char[].
9712 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java
9713 // @HotSpotIntrinsicCandidate
9714 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) {
9715 // for (int i = 0; i < len; i++) {
9716 // dst[dstOff++] = (char)(src[srcOff++] & 0xff);
9717 // }
9718 // }
9719 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
9720 XMMRegister tmp1, Register tmp2) {
9721 Label copy_chars_loop, done, below_threshold, avx3_threshold;
9722 // rsi: src
9723 // rdi: dst
9724 // rdx: len
9725 // rcx: tmp2
9726
9727 // rsi holds start addr of source byte[] to be inflated
9728 // rdi holds start addr of destination char[]
9729 // rdx holds length
9730 assert_different_registers(src, dst, len, tmp2);
9731 movl(tmp2, len);
9732 if ((UseAVX > 2) && // AVX512
9733 VM_Version::supports_avx512vlbw() &&
9734 VM_Version::supports_bmi2()) {
9735
9736 Label copy_32_loop, copy_tail;
9737 Register tmp3_aliased = len;
9738
9739 // if length of the string is less than 16, handle it in an old fashioned way
9740 testl(len, -16);
9741 jcc(Assembler::zero, below_threshold);
9742
9743 testl(len, -1 * AVX3Threshold);
9744 jcc(Assembler::zero, avx3_threshold);
9745
9746 // In order to use only one arithmetic operation for the main loop we use
9747 // this pre-calculation
9748 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop
9749 andl(len, -32); // vector count
9750 jccb(Assembler::zero, copy_tail);
9751
9752 lea(src, Address(src, len, Address::times_1));
9753 lea(dst, Address(dst, len, Address::times_2));
9754 negptr(len);
9755
9756
9757 // inflate 32 chars per iter
9758 bind(copy_32_loop);
9759 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit);
9760 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit);
9761 addptr(len, 32);
9762 jcc(Assembler::notZero, copy_32_loop);
9763
9764 bind(copy_tail);
9765 // bail out when there is nothing to be done
9766 testl(tmp2, -1); // we don't destroy the contents of tmp2 here
9767 jcc(Assembler::zero, done);
9768
9769 // ~(~0 << length), where length is the # of remaining elements to process
9770 movl(tmp3_aliased, -1);
9771 shlxl(tmp3_aliased, tmp3_aliased, tmp2);
9772 notl(tmp3_aliased);
9773 kmovdl(k2, tmp3_aliased);
9774 evpmovzxbw(tmp1, k2, Address(src, 0), Assembler::AVX_512bit);
9775 evmovdquw(Address(dst, 0), k2, tmp1, Assembler::AVX_512bit);
9776
9777 jmp(done);
9778 bind(avx3_threshold);
9779 }
9780 if (UseSSE42Intrinsics) {
9781 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail;
9782
9783 if (UseAVX > 1) {
9784 andl(tmp2, (16 - 1));
9785 andl(len, -16);
9786 jccb(Assembler::zero, copy_new_tail);
9787 } else {
9788 andl(tmp2, 0x00000007); // tail count (in chars)
9789 andl(len, 0xfffffff8); // vector count (in chars)
9790 jccb(Assembler::zero, copy_tail);
9791 }
9792
9793 // vectored inflation
9794 lea(src, Address(src, len, Address::times_1));
9795 lea(dst, Address(dst, len, Address::times_2));
9796 negptr(len);
9797
9798 if (UseAVX > 1) {
9799 bind(copy_16_loop);
9800 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit);
9801 vmovdqu(Address(dst, len, Address::times_2), tmp1);
9802 addptr(len, 16);
9803 jcc(Assembler::notZero, copy_16_loop);
9804
9805 bind(below_threshold);
9806 bind(copy_new_tail);
9807 movl(len, tmp2);
9808 andl(tmp2, 0x00000007);
9809 andl(len, 0xFFFFFFF8);
9810 jccb(Assembler::zero, copy_tail);
9811
9812 pmovzxbw(tmp1, Address(src, 0));
9813 movdqu(Address(dst, 0), tmp1);
9814 addptr(src, 8);
9815 addptr(dst, 2 * 8);
9816
9817 jmp(copy_tail, true);
9818 }
9819
9820 // inflate 8 chars per iter
9821 bind(copy_8_loop);
9822 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words
9823 movdqu(Address(dst, len, Address::times_2), tmp1);
9824 addptr(len, 8);
9825 jcc(Assembler::notZero, copy_8_loop);
9826
9827 bind(copy_tail);
9828 movl(len, tmp2);
9829
9830 cmpl(len, 4);
9831 jccb(Assembler::less, copy_bytes);
9832
9833 movdl(tmp1, Address(src, 0)); // load 4 byte chars
9834 pmovzxbw(tmp1, tmp1);
9835 movq(Address(dst, 0), tmp1);
9836 subptr(len, 4);
9837 addptr(src, 4);
9838 addptr(dst, 8);
9839
9840 bind(copy_bytes);
9841 } else {
9842 bind(below_threshold);
9843 }
9844
9845 testl(len, len);
9846 jccb(Assembler::zero, done);
9847 lea(src, Address(src, len, Address::times_1));
9848 lea(dst, Address(dst, len, Address::times_2));
9849 negptr(len);
9850
9851 // inflate 1 char per iter
9852 bind(copy_chars_loop);
9853 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char
9854 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word
9855 increment(len);
9856 jcc(Assembler::notZero, copy_chars_loop);
9857
9858 bind(done);
9859 }
9860
9861 #ifdef _LP64
9862 void MacroAssembler::convert_f2i(Register dst, XMMRegister src) {
9863 Label done;
9864 cvttss2sil(dst, src);
9865 // Conversion instructions do not match JLS for overflow, underflow and NaN -> fixup in stub
9866 cmpl(dst, 0x80000000); // float_sign_flip
9867 jccb(Assembler::notEqual, done);
9868 subptr(rsp, 8);
9869 movflt(Address(rsp, 0), src);
9870 call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::x86::f2i_fixup())));
9871 pop(dst);
9872 bind(done);
9873 }
9874
9875 void MacroAssembler::convert_d2i(Register dst, XMMRegister src) {
9876 Label done;
9877 cvttsd2sil(dst, src);
9878 // Conversion instructions do not match JLS for overflow, underflow and NaN -> fixup in stub
9879 cmpl(dst, 0x80000000); // float_sign_flip
9880 jccb(Assembler::notEqual, done);
9881 subptr(rsp, 8);
9882 movdbl(Address(rsp, 0), src);
9883 call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::x86::d2i_fixup())));
9884 pop(dst);
9885 bind(done);
9886 }
9887
9888 void MacroAssembler::convert_f2l(Register dst, XMMRegister src) {
9889 Label done;
9890 cvttss2siq(dst, src);
9891 cmp64(dst, ExternalAddress((address) StubRoutines::x86::double_sign_flip()));
9892 jccb(Assembler::notEqual, done);
9893 subptr(rsp, 8);
9894 movflt(Address(rsp, 0), src);
9895 call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::x86::f2l_fixup())));
9896 pop(dst);
9897 bind(done);
9898 }
9899
9900 void MacroAssembler::convert_d2l(Register dst, XMMRegister src) {
9901 Label done;
9902 cvttsd2siq(dst, src);
9903 cmp64(dst, ExternalAddress((address) StubRoutines::x86::double_sign_flip()));
9904 jccb(Assembler::notEqual, done);
9905 subptr(rsp, 8);
9906 movdbl(Address(rsp, 0), src);
9907 call(RuntimeAddress(CAST_FROM_FN_PTR(address, StubRoutines::x86::d2l_fixup())));
9908 pop(dst);
9909 bind(done);
9910 }
9911
9912 void MacroAssembler::cache_wb(Address line)
9913 {
9914 // 64 bit cpus always support clflush
9915 assert(VM_Version::supports_clflush(), "clflush should be available");
9916 bool optimized = VM_Version::supports_clflushopt();
9917 bool no_evict = VM_Version::supports_clwb();
9918
9919 // prefer clwb (writeback without evict) otherwise
9920 // prefer clflushopt (potentially parallel writeback with evict)
9921 // otherwise fallback on clflush (serial writeback with evict)
9922
9923 if (optimized) {
9924 if (no_evict) {
9925 clwb(line);
9926 } else {
9927 clflushopt(line);
9928 }
9929 } else {
9930 // no need for fence when using CLFLUSH
9931 clflush(line);
9932 }
9933 }
9934
9935 void MacroAssembler::cache_wbsync(bool is_pre)
9936 {
9937 assert(VM_Version::supports_clflush(), "clflush should be available");
9938 bool optimized = VM_Version::supports_clflushopt();
9939 bool no_evict = VM_Version::supports_clwb();
9940
9941 // pick the correct implementation
9942
9943 if (!is_pre && (optimized || no_evict)) {
9944 // need an sfence for post flush when using clflushopt or clwb
9945 // otherwise no no need for any synchroniaztion
9946
9947 sfence();
9948 }
9949 }
9950 #endif // _LP64
9951
9952 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
9953 switch (cond) {
9954 // Note some conditions are synonyms for others
9955 case Assembler::zero: return Assembler::notZero;
9956 case Assembler::notZero: return Assembler::zero;
9957 case Assembler::less: return Assembler::greaterEqual;
9958 case Assembler::lessEqual: return Assembler::greater;
9959 case Assembler::greater: return Assembler::lessEqual;
9960 case Assembler::greaterEqual: return Assembler::less;
9961 case Assembler::below: return Assembler::aboveEqual;
9962 case Assembler::belowEqual: return Assembler::above;
9963 case Assembler::above: return Assembler::belowEqual;
9964 case Assembler::aboveEqual: return Assembler::below;
9965 case Assembler::overflow: return Assembler::noOverflow;
9966 case Assembler::noOverflow: return Assembler::overflow;
9967 case Assembler::negative: return Assembler::positive;
9968 case Assembler::positive: return Assembler::negative;
9969 case Assembler::parity: return Assembler::noParity;
9970 case Assembler::noParity: return Assembler::parity;
9971 }
9972 ShouldNotReachHere(); return Assembler::overflow;
9973 }
9974
9975 SkipIfEqual::SkipIfEqual(
9976 MacroAssembler* masm, const bool* flag_addr, bool value) {
9977 _masm = masm;
9978 _masm->cmp8(ExternalAddress((address)flag_addr), value);
9979 _masm->jcc(Assembler::equal, _label);
9980 }
9981
9982 SkipIfEqual::~SkipIfEqual() {
9983 _masm->bind(_label);
9984 }
9985
9986 // 32-bit Windows has its own fast-path implementation
9987 // of get_thread
9988 #if !defined(WIN32) || defined(_LP64)
9989
9990 // This is simply a call to Thread::current()
9991 void MacroAssembler::get_thread(Register thread) {
9992 if (thread != rax) {
9993 push(rax);
9994 }
9995 LP64_ONLY(push(rdi);)
9996 LP64_ONLY(push(rsi);)
9997 push(rdx);
9998 push(rcx);
9999 #ifdef _LP64
10000 push(r8);
10001 push(r9);
10002 push(r10);
10003 push(r11);
10004 #endif
10005
10006 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0);
10007
10008 #ifdef _LP64
10009 pop(r11);
10010 pop(r10);
10011 pop(r9);
10012 pop(r8);
10013 #endif
10014 pop(rcx);
10015 pop(rdx);
10016 LP64_ONLY(pop(rsi);)
10017 LP64_ONLY(pop(rdi);)
10018 if (thread != rax) {
10019 mov(thread, rax);
10020 pop(rax);
10021 }
10022 }
10023
10024 #endif // !WIN32 || _LP64