1 /*
2 * Copyright (c) 1997, 2018, 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/access.hpp"
37 #include "oops/klass.inline.hpp"
38 #include "prims/methodHandles.hpp"
39 #include "runtime/biasedLocking.hpp"
40 #include "runtime/interfaceSupport.inline.hpp"
41 #include "runtime/objectMonitor.hpp"
42 #include "runtime/os.hpp"
43 #include "runtime/safepoint.hpp"
44 #include "runtime/safepointMechanism.hpp"
45 #include "runtime/sharedRuntime.hpp"
46 #include "runtime/stubRoutines.hpp"
47 #include "runtime/thread.hpp"
48 #include "utilities/macros.hpp"
49 #include "crc32c.h"
50 #ifdef COMPILER2
51 #include "opto/intrinsicnode.hpp"
52 #endif
53
54 #ifdef PRODUCT
55 #define BLOCK_COMMENT(str) /* nothing */
56 #define STOP(error) stop(error)
57 #else
58 #define BLOCK_COMMENT(str) block_comment(str)
59 #define STOP(error) block_comment(error); stop(error)
60 #endif
61
62 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
63
64 #ifdef ASSERT
65 bool AbstractAssembler::pd_check_instruction_mark() { return true; }
66 #endif
67
68 static Assembler::Condition reverse[] = {
69 Assembler::noOverflow /* overflow = 0x0 */ ,
70 Assembler::overflow /* noOverflow = 0x1 */ ,
71 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ ,
72 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ ,
73 Assembler::notZero /* zero = 0x4, equal = 0x4 */ ,
74 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ ,
75 Assembler::above /* belowEqual = 0x6 */ ,
76 Assembler::belowEqual /* above = 0x7 */ ,
77 Assembler::positive /* negative = 0x8 */ ,
78 Assembler::negative /* positive = 0x9 */ ,
79 Assembler::noParity /* parity = 0xa */ ,
80 Assembler::parity /* noParity = 0xb */ ,
81 Assembler::greaterEqual /* less = 0xc */ ,
82 Assembler::less /* greaterEqual = 0xd */ ,
83 Assembler::greater /* lessEqual = 0xe */ ,
84 Assembler::lessEqual /* greater = 0xf, */
85
86 };
87
88
89 // Implementation of MacroAssembler
90
91 // First all the versions that have distinct versions depending on 32/64 bit
92 // Unless the difference is trivial (1 line or so).
93
94 #ifndef _LP64
95
96 // 32bit versions
97
98 Address MacroAssembler::as_Address(AddressLiteral adr) {
99 return Address(adr.target(), adr.rspec());
100 }
101
102 Address MacroAssembler::as_Address(ArrayAddress adr) {
103 return Address::make_array(adr);
104 }
105
106 void MacroAssembler::call_VM_leaf_base(address entry_point,
107 int number_of_arguments) {
108 call(RuntimeAddress(entry_point));
109 increment(rsp, number_of_arguments * wordSize);
110 }
111
112 void MacroAssembler::cmpklass(Address src1, Metadata* obj) {
113 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
114 }
115
116 void MacroAssembler::cmpklass(Register src1, Metadata* obj) {
117 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
118 }
119
120 void MacroAssembler::cmpoop(Address src1, jobject obj) {
121 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
122 }
123
124 void MacroAssembler::cmpoop(Register src1, jobject obj) {
125 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
126 }
127
128 void MacroAssembler::extend_sign(Register hi, Register lo) {
129 // According to Intel Doc. AP-526, "Integer Divide", p.18.
130 if (VM_Version::is_P6() && hi == rdx && lo == rax) {
131 cdql();
132 } else {
133 movl(hi, lo);
134 sarl(hi, 31);
135 }
136 }
137
138 void MacroAssembler::jC2(Register tmp, Label& L) {
139 // set parity bit if FPU flag C2 is set (via rax)
140 save_rax(tmp);
141 fwait(); fnstsw_ax();
142 sahf();
143 restore_rax(tmp);
144 // branch
145 jcc(Assembler::parity, L);
146 }
147
148 void MacroAssembler::jnC2(Register tmp, Label& L) {
149 // set parity bit if FPU flag C2 is set (via rax)
150 save_rax(tmp);
151 fwait(); fnstsw_ax();
152 sahf();
153 restore_rax(tmp);
154 // branch
155 jcc(Assembler::noParity, L);
156 }
157
158 // 32bit can do a case table jump in one instruction but we no longer allow the base
159 // to be installed in the Address class
160 void MacroAssembler::jump(ArrayAddress entry) {
161 jmp(as_Address(entry));
162 }
163
164 // Note: y_lo will be destroyed
165 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
166 // Long compare for Java (semantics as described in JVM spec.)
167 Label high, low, done;
168
169 cmpl(x_hi, y_hi);
170 jcc(Assembler::less, low);
171 jcc(Assembler::greater, high);
172 // x_hi is the return register
173 xorl(x_hi, x_hi);
174 cmpl(x_lo, y_lo);
175 jcc(Assembler::below, low);
176 jcc(Assembler::equal, done);
177
178 bind(high);
179 xorl(x_hi, x_hi);
180 increment(x_hi);
181 jmp(done);
182
183 bind(low);
184 xorl(x_hi, x_hi);
185 decrementl(x_hi);
186
187 bind(done);
188 }
189
190 void MacroAssembler::lea(Register dst, AddressLiteral src) {
191 mov_literal32(dst, (int32_t)src.target(), src.rspec());
192 }
193
194 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
195 // leal(dst, as_Address(adr));
196 // see note in movl as to why we must use a move
197 mov_literal32(dst, (int32_t) adr.target(), adr.rspec());
198 }
199
200 void MacroAssembler::leave() {
201 mov(rsp, rbp);
202 pop(rbp);
203 }
204
205 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) {
206 // Multiplication of two Java long values stored on the stack
207 // as illustrated below. Result is in rdx:rax.
208 //
209 // rsp ---> [ ?? ] \ \
210 // .... | y_rsp_offset |
211 // [ y_lo ] / (in bytes) | x_rsp_offset
212 // [ y_hi ] | (in bytes)
213 // .... |
214 // [ x_lo ] /
215 // [ x_hi ]
216 // ....
217 //
218 // Basic idea: lo(result) = lo(x_lo * y_lo)
219 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
220 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset);
221 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset);
222 Label quick;
223 // load x_hi, y_hi and check if quick
224 // multiplication is possible
225 movl(rbx, x_hi);
226 movl(rcx, y_hi);
227 movl(rax, rbx);
228 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0
229 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply
230 // do full multiplication
231 // 1st step
232 mull(y_lo); // x_hi * y_lo
233 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx,
234 // 2nd step
235 movl(rax, x_lo);
236 mull(rcx); // x_lo * y_hi
237 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx,
238 // 3rd step
239 bind(quick); // note: rbx, = 0 if quick multiply!
240 movl(rax, x_lo);
241 mull(y_lo); // x_lo * y_lo
242 addl(rdx, rbx); // correct hi(x_lo * y_lo)
243 }
244
245 void MacroAssembler::lneg(Register hi, Register lo) {
246 negl(lo);
247 adcl(hi, 0);
248 negl(hi);
249 }
250
251 void MacroAssembler::lshl(Register hi, Register lo) {
252 // Java shift left long support (semantics as described in JVM spec., p.305)
253 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n))
254 // shift value is in rcx !
255 assert(hi != rcx, "must not use rcx");
256 assert(lo != rcx, "must not use rcx");
257 const Register s = rcx; // shift count
258 const int n = BitsPerWord;
259 Label L;
260 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
261 cmpl(s, n); // if (s < n)
262 jcc(Assembler::less, L); // else (s >= n)
263 movl(hi, lo); // x := x << n
264 xorl(lo, lo);
265 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
266 bind(L); // s (mod n) < n
267 shldl(hi, lo); // x := x << s
268 shll(lo);
269 }
270
271
272 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) {
273 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310)
274 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n))
275 assert(hi != rcx, "must not use rcx");
276 assert(lo != rcx, "must not use rcx");
277 const Register s = rcx; // shift count
278 const int n = BitsPerWord;
279 Label L;
280 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
281 cmpl(s, n); // if (s < n)
282 jcc(Assembler::less, L); // else (s >= n)
283 movl(lo, hi); // x := x >> n
284 if (sign_extension) sarl(hi, 31);
285 else xorl(hi, hi);
286 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
287 bind(L); // s (mod n) < n
288 shrdl(lo, hi); // x := x >> s
289 if (sign_extension) sarl(hi);
290 else shrl(hi);
291 }
292
293 void MacroAssembler::movoop(Register dst, jobject obj) {
294 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
295 }
296
297 void MacroAssembler::movoop(Address dst, jobject obj) {
298 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
299 }
300
301 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
302 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
303 }
304
305 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
306 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
307 }
308
309 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
310 // scratch register is not used,
311 // it is defined to match parameters of 64-bit version of this method.
312 if (src.is_lval()) {
313 mov_literal32(dst, (intptr_t)src.target(), src.rspec());
314 } else {
315 movl(dst, as_Address(src));
316 }
317 }
318
319 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
320 movl(as_Address(dst), src);
321 }
322
323 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
324 movl(dst, as_Address(src));
325 }
326
327 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
328 void MacroAssembler::movptr(Address dst, intptr_t src) {
329 movl(dst, src);
330 }
331
332
333 void MacroAssembler::pop_callee_saved_registers() {
334 pop(rcx);
335 pop(rdx);
336 pop(rdi);
337 pop(rsi);
338 }
339
340 void MacroAssembler::pop_fTOS() {
341 fld_d(Address(rsp, 0));
342 addl(rsp, 2 * wordSize);
343 }
344
345 void MacroAssembler::push_callee_saved_registers() {
346 push(rsi);
347 push(rdi);
348 push(rdx);
349 push(rcx);
350 }
351
352 void MacroAssembler::push_fTOS() {
353 subl(rsp, 2 * wordSize);
354 fstp_d(Address(rsp, 0));
355 }
356
357
358 void MacroAssembler::pushoop(jobject obj) {
359 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate());
360 }
361
362 void MacroAssembler::pushklass(Metadata* obj) {
363 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate());
364 }
365
366 void MacroAssembler::pushptr(AddressLiteral src) {
367 if (src.is_lval()) {
368 push_literal32((int32_t)src.target(), src.rspec());
369 } else {
370 pushl(as_Address(src));
371 }
372 }
373
374 void MacroAssembler::set_word_if_not_zero(Register dst) {
375 xorl(dst, dst);
376 set_byte_if_not_zero(dst);
377 }
378
379 static void pass_arg0(MacroAssembler* masm, Register arg) {
380 masm->push(arg);
381 }
382
383 static void pass_arg1(MacroAssembler* masm, Register arg) {
384 masm->push(arg);
385 }
386
387 static void pass_arg2(MacroAssembler* masm, Register arg) {
388 masm->push(arg);
389 }
390
391 static void pass_arg3(MacroAssembler* masm, Register arg) {
392 masm->push(arg);
393 }
394
395 #ifndef PRODUCT
396 extern "C" void findpc(intptr_t x);
397 #endif
398
399 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) {
400 // In order to get locks to work, we need to fake a in_VM state
401 JavaThread* thread = JavaThread::current();
402 JavaThreadState saved_state = thread->thread_state();
403 thread->set_thread_state(_thread_in_vm);
404 if (ShowMessageBoxOnError) {
405 JavaThread* thread = JavaThread::current();
406 JavaThreadState saved_state = thread->thread_state();
407 thread->set_thread_state(_thread_in_vm);
408 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
409 ttyLocker ttyl;
410 BytecodeCounter::print();
411 }
412 // To see where a verify_oop failed, get $ebx+40/X for this frame.
413 // This is the value of eip which points to where verify_oop will return.
414 if (os::message_box(msg, "Execution stopped, print registers?")) {
415 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip);
416 BREAKPOINT;
417 }
418 } else {
419 ttyLocker ttyl;
420 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg);
421 }
422 // Don't assert holding the ttyLock
423 assert(false, "DEBUG MESSAGE: %s", msg);
424 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
425 }
426
427 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) {
428 ttyLocker ttyl;
429 FlagSetting fs(Debugging, true);
430 tty->print_cr("eip = 0x%08x", eip);
431 #ifndef PRODUCT
432 if ((WizardMode || Verbose) && PrintMiscellaneous) {
433 tty->cr();
434 findpc(eip);
435 tty->cr();
436 }
437 #endif
438 #define PRINT_REG(rax) \
439 { tty->print("%s = ", #rax); os::print_location(tty, rax); }
440 PRINT_REG(rax);
441 PRINT_REG(rbx);
442 PRINT_REG(rcx);
443 PRINT_REG(rdx);
444 PRINT_REG(rdi);
445 PRINT_REG(rsi);
446 PRINT_REG(rbp);
447 PRINT_REG(rsp);
448 #undef PRINT_REG
449 // Print some words near top of staack.
450 int* dump_sp = (int*) rsp;
451 for (int col1 = 0; col1 < 8; col1++) {
452 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
453 os::print_location(tty, *dump_sp++);
454 }
455 for (int row = 0; row < 16; row++) {
456 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
457 for (int col = 0; col < 8; col++) {
458 tty->print(" 0x%08x", *dump_sp++);
459 }
460 tty->cr();
461 }
462 // Print some instructions around pc:
463 Disassembler::decode((address)eip-64, (address)eip);
464 tty->print_cr("--------");
465 Disassembler::decode((address)eip, (address)eip+32);
466 }
467
468 void MacroAssembler::stop(const char* msg) {
469 ExternalAddress message((address)msg);
470 // push address of message
471 pushptr(message.addr());
472 { Label L; call(L, relocInfo::none); bind(L); } // push eip
473 pusha(); // push registers
474 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32)));
475 hlt();
476 }
477
478 void MacroAssembler::warn(const char* msg) {
479 push_CPU_state();
480
481 ExternalAddress message((address) msg);
482 // push address of message
483 pushptr(message.addr());
484
485 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning)));
486 addl(rsp, wordSize); // discard argument
487 pop_CPU_state();
488 }
489
490 void MacroAssembler::print_state() {
491 { Label L; call(L, relocInfo::none); bind(L); } // push eip
492 pusha(); // push registers
493
494 push_CPU_state();
495 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32)));
496 pop_CPU_state();
497
498 popa();
499 addl(rsp, wordSize);
500 }
501
502 #else // _LP64
503
504 // 64 bit versions
505
506 Address MacroAssembler::as_Address(AddressLiteral adr) {
507 // amd64 always does this as a pc-rel
508 // we can be absolute or disp based on the instruction type
509 // jmp/call are displacements others are absolute
510 assert(!adr.is_lval(), "must be rval");
511 assert(reachable(adr), "must be");
512 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc());
513
514 }
515
516 Address MacroAssembler::as_Address(ArrayAddress adr) {
517 AddressLiteral base = adr.base();
518 lea(rscratch1, base);
519 Address index = adr.index();
520 assert(index._disp == 0, "must not have disp"); // maybe it can?
521 Address array(rscratch1, index._index, index._scale, index._disp);
522 return array;
523 }
524
525 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) {
526 Label L, E;
527
528 #ifdef _WIN64
529 // Windows always allocates space for it's register args
530 assert(num_args <= 4, "only register arguments supported");
531 subq(rsp, frame::arg_reg_save_area_bytes);
532 #endif
533
534 // Align stack if necessary
535 testl(rsp, 15);
536 jcc(Assembler::zero, L);
537
538 subq(rsp, 8);
539 {
540 call(RuntimeAddress(entry_point));
541 }
542 addq(rsp, 8);
543 jmp(E);
544
545 bind(L);
546 {
547 call(RuntimeAddress(entry_point));
548 }
549
550 bind(E);
551
552 #ifdef _WIN64
553 // restore stack pointer
554 addq(rsp, frame::arg_reg_save_area_bytes);
555 #endif
556
557 }
558
559 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) {
560 assert(!src2.is_lval(), "should use cmpptr");
561
562 if (reachable(src2)) {
563 cmpq(src1, as_Address(src2));
564 } else {
565 lea(rscratch1, src2);
566 Assembler::cmpq(src1, Address(rscratch1, 0));
567 }
568 }
569
570 int MacroAssembler::corrected_idivq(Register reg) {
571 // Full implementation of Java ldiv and lrem; checks for special
572 // case as described in JVM spec., p.243 & p.271. The function
573 // returns the (pc) offset of the idivl instruction - may be needed
574 // for implicit exceptions.
575 //
576 // normal case special case
577 //
578 // input : rax: dividend min_long
579 // reg: divisor (may not be eax/edx) -1
580 //
581 // output: rax: quotient (= rax idiv reg) min_long
582 // rdx: remainder (= rax irem reg) 0
583 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register");
584 static const int64_t min_long = 0x8000000000000000;
585 Label normal_case, special_case;
586
587 // check for special case
588 cmp64(rax, ExternalAddress((address) &min_long));
589 jcc(Assembler::notEqual, normal_case);
590 xorl(rdx, rdx); // prepare rdx for possible special case (where
591 // remainder = 0)
592 cmpq(reg, -1);
593 jcc(Assembler::equal, special_case);
594
595 // handle normal case
596 bind(normal_case);
597 cdqq();
598 int idivq_offset = offset();
599 idivq(reg);
600
601 // normal and special case exit
602 bind(special_case);
603
604 return idivq_offset;
605 }
606
607 void MacroAssembler::decrementq(Register reg, int value) {
608 if (value == min_jint) { subq(reg, value); return; }
609 if (value < 0) { incrementq(reg, -value); return; }
610 if (value == 0) { ; return; }
611 if (value == 1 && UseIncDec) { decq(reg) ; return; }
612 /* else */ { subq(reg, value) ; return; }
613 }
614
615 void MacroAssembler::decrementq(Address dst, int value) {
616 if (value == min_jint) { subq(dst, value); return; }
617 if (value < 0) { incrementq(dst, -value); return; }
618 if (value == 0) { ; return; }
619 if (value == 1 && UseIncDec) { decq(dst) ; return; }
620 /* else */ { subq(dst, value) ; return; }
621 }
622
623 void MacroAssembler::incrementq(AddressLiteral dst) {
624 if (reachable(dst)) {
625 incrementq(as_Address(dst));
626 } else {
627 lea(rscratch1, dst);
628 incrementq(Address(rscratch1, 0));
629 }
630 }
631
632 void MacroAssembler::incrementq(Register reg, int value) {
633 if (value == min_jint) { addq(reg, value); return; }
634 if (value < 0) { decrementq(reg, -value); return; }
635 if (value == 0) { ; return; }
636 if (value == 1 && UseIncDec) { incq(reg) ; return; }
637 /* else */ { addq(reg, value) ; return; }
638 }
639
640 void MacroAssembler::incrementq(Address dst, int value) {
641 if (value == min_jint) { addq(dst, value); return; }
642 if (value < 0) { decrementq(dst, -value); return; }
643 if (value == 0) { ; return; }
644 if (value == 1 && UseIncDec) { incq(dst) ; return; }
645 /* else */ { addq(dst, value) ; return; }
646 }
647
648 // 32bit can do a case table jump in one instruction but we no longer allow the base
649 // to be installed in the Address class
650 void MacroAssembler::jump(ArrayAddress entry) {
651 lea(rscratch1, entry.base());
652 Address dispatch = entry.index();
653 assert(dispatch._base == noreg, "must be");
654 dispatch._base = rscratch1;
655 jmp(dispatch);
656 }
657
658 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
659 ShouldNotReachHere(); // 64bit doesn't use two regs
660 cmpq(x_lo, y_lo);
661 }
662
663 void MacroAssembler::lea(Register dst, AddressLiteral src) {
664 mov_literal64(dst, (intptr_t)src.target(), src.rspec());
665 }
666
667 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
668 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec());
669 movptr(dst, rscratch1);
670 }
671
672 void MacroAssembler::leave() {
673 // %%% is this really better? Why not on 32bit too?
674 emit_int8((unsigned char)0xC9); // LEAVE
675 }
676
677 void MacroAssembler::lneg(Register hi, Register lo) {
678 ShouldNotReachHere(); // 64bit doesn't use two regs
679 negq(lo);
680 }
681
682 void MacroAssembler::movoop(Register dst, jobject obj) {
683 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate());
684 }
685
686 void MacroAssembler::movoop(Address dst, jobject obj) {
687 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate());
688 movq(dst, rscratch1);
689 }
690
691 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
692 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
693 }
694
695 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
696 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
697 movq(dst, rscratch1);
698 }
699
700 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
701 if (src.is_lval()) {
702 mov_literal64(dst, (intptr_t)src.target(), src.rspec());
703 } else {
704 if (reachable(src)) {
705 movq(dst, as_Address(src));
706 } else {
707 lea(scratch, src);
708 movq(dst, Address(scratch, 0));
709 }
710 }
711 }
712
713 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
714 movq(as_Address(dst), src);
715 }
716
717 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
718 movq(dst, as_Address(src));
719 }
720
721 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
722 void MacroAssembler::movptr(Address dst, intptr_t src) {
723 mov64(rscratch1, src);
724 movq(dst, rscratch1);
725 }
726
727 // These are mostly for initializing NULL
728 void MacroAssembler::movptr(Address dst, int32_t src) {
729 movslq(dst, src);
730 }
731
732 void MacroAssembler::movptr(Register dst, int32_t src) {
733 mov64(dst, (intptr_t)src);
734 }
735
736 void MacroAssembler::pushoop(jobject obj) {
737 movoop(rscratch1, obj);
738 push(rscratch1);
739 }
740
741 void MacroAssembler::pushklass(Metadata* obj) {
742 mov_metadata(rscratch1, obj);
743 push(rscratch1);
744 }
745
746 void MacroAssembler::pushptr(AddressLiteral src) {
747 lea(rscratch1, src);
748 if (src.is_lval()) {
749 push(rscratch1);
750 } else {
751 pushq(Address(rscratch1, 0));
752 }
753 }
754
755 void MacroAssembler::reset_last_Java_frame(bool clear_fp) {
756 // we must set sp to zero to clear frame
757 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
758 // must clear fp, so that compiled frames are not confused; it is
759 // possible that we need it only for debugging
760 if (clear_fp) {
761 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
762 }
763
764 // Always clear the pc because it could have been set by make_walkable()
765 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
766 vzeroupper();
767 }
768
769 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
770 Register last_java_fp,
771 address last_java_pc) {
772 vzeroupper();
773 // determine last_java_sp register
774 if (!last_java_sp->is_valid()) {
775 last_java_sp = rsp;
776 }
777
778 // last_java_fp is optional
779 if (last_java_fp->is_valid()) {
780 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()),
781 last_java_fp);
782 }
783
784 // last_java_pc is optional
785 if (last_java_pc != NULL) {
786 Address java_pc(r15_thread,
787 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset());
788 lea(rscratch1, InternalAddress(last_java_pc));
789 movptr(java_pc, rscratch1);
790 }
791
792 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
793 }
794
795 static void pass_arg0(MacroAssembler* masm, Register arg) {
796 if (c_rarg0 != arg ) {
797 masm->mov(c_rarg0, arg);
798 }
799 }
800
801 static void pass_arg1(MacroAssembler* masm, Register arg) {
802 if (c_rarg1 != arg ) {
803 masm->mov(c_rarg1, arg);
804 }
805 }
806
807 static void pass_arg2(MacroAssembler* masm, Register arg) {
808 if (c_rarg2 != arg ) {
809 masm->mov(c_rarg2, arg);
810 }
811 }
812
813 static void pass_arg3(MacroAssembler* masm, Register arg) {
814 if (c_rarg3 != arg ) {
815 masm->mov(c_rarg3, arg);
816 }
817 }
818
819 void MacroAssembler::stop(const char* msg) {
820 address rip = pc();
821 pusha(); // get regs on stack
822 lea(c_rarg0, ExternalAddress((address) msg));
823 lea(c_rarg1, InternalAddress(rip));
824 movq(c_rarg2, rsp); // pass pointer to regs array
825 andq(rsp, -16); // align stack as required by ABI
826 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
827 hlt();
828 }
829
830 void MacroAssembler::warn(const char* msg) {
831 push(rbp);
832 movq(rbp, rsp);
833 andq(rsp, -16); // align stack as required by push_CPU_state and call
834 push_CPU_state(); // keeps alignment at 16 bytes
835 lea(c_rarg0, ExternalAddress((address) msg));
836 lea(rax, ExternalAddress(CAST_FROM_FN_PTR(address, warning)));
837 call(rax);
838 pop_CPU_state();
839 mov(rsp, rbp);
840 pop(rbp);
841 }
842
843 void MacroAssembler::print_state() {
844 address rip = pc();
845 pusha(); // get regs on stack
846 push(rbp);
847 movq(rbp, rsp);
848 andq(rsp, -16); // align stack as required by push_CPU_state and call
849 push_CPU_state(); // keeps alignment at 16 bytes
850
851 lea(c_rarg0, InternalAddress(rip));
852 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array
853 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1);
854
855 pop_CPU_state();
856 mov(rsp, rbp);
857 pop(rbp);
858 popa();
859 }
860
861 #ifndef PRODUCT
862 extern "C" void findpc(intptr_t x);
863 #endif
864
865 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) {
866 // In order to get locks to work, we need to fake a in_VM state
867 if (ShowMessageBoxOnError) {
868 JavaThread* thread = JavaThread::current();
869 JavaThreadState saved_state = thread->thread_state();
870 thread->set_thread_state(_thread_in_vm);
871 #ifndef PRODUCT
872 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
873 ttyLocker ttyl;
874 BytecodeCounter::print();
875 }
876 #endif
877 // To see where a verify_oop failed, get $ebx+40/X for this frame.
878 // XXX correct this offset for amd64
879 // This is the value of eip which points to where verify_oop will return.
880 if (os::message_box(msg, "Execution stopped, print registers?")) {
881 print_state64(pc, regs);
882 BREAKPOINT;
883 assert(false, "start up GDB");
884 }
885 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
886 } else {
887 ttyLocker ttyl;
888 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n",
889 msg);
890 assert(false, "DEBUG MESSAGE: %s", msg);
891 }
892 }
893
894 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) {
895 ttyLocker ttyl;
896 FlagSetting fs(Debugging, true);
897 tty->print_cr("rip = 0x%016lx", (intptr_t)pc);
898 #ifndef PRODUCT
899 tty->cr();
900 findpc(pc);
901 tty->cr();
902 #endif
903 #define PRINT_REG(rax, value) \
904 { tty->print("%s = ", #rax); os::print_location(tty, value); }
905 PRINT_REG(rax, regs[15]);
906 PRINT_REG(rbx, regs[12]);
907 PRINT_REG(rcx, regs[14]);
908 PRINT_REG(rdx, regs[13]);
909 PRINT_REG(rdi, regs[8]);
910 PRINT_REG(rsi, regs[9]);
911 PRINT_REG(rbp, regs[10]);
912 PRINT_REG(rsp, regs[11]);
913 PRINT_REG(r8 , regs[7]);
914 PRINT_REG(r9 , regs[6]);
915 PRINT_REG(r10, regs[5]);
916 PRINT_REG(r11, regs[4]);
917 PRINT_REG(r12, regs[3]);
918 PRINT_REG(r13, regs[2]);
919 PRINT_REG(r14, regs[1]);
920 PRINT_REG(r15, regs[0]);
921 #undef PRINT_REG
922 // Print some words near top of staack.
923 int64_t* rsp = (int64_t*) regs[11];
924 int64_t* dump_sp = rsp;
925 for (int col1 = 0; col1 < 8; col1++) {
926 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
927 os::print_location(tty, *dump_sp++);
928 }
929 for (int row = 0; row < 25; row++) {
930 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
931 for (int col = 0; col < 4; col++) {
932 tty->print(" 0x%016lx", (intptr_t)*dump_sp++);
933 }
934 tty->cr();
935 }
936 // Print some instructions around pc:
937 Disassembler::decode((address)pc-64, (address)pc);
938 tty->print_cr("--------");
939 Disassembler::decode((address)pc, (address)pc+32);
940 }
941
942 #endif // _LP64
943
944 // Now versions that are common to 32/64 bit
945
946 void MacroAssembler::addptr(Register dst, int32_t imm32) {
947 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32));
948 }
949
950 void MacroAssembler::addptr(Register dst, Register src) {
951 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
952 }
953
954 void MacroAssembler::addptr(Address dst, Register src) {
955 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
956 }
957
958 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) {
959 if (reachable(src)) {
960 Assembler::addsd(dst, as_Address(src));
961 } else {
962 lea(rscratch1, src);
963 Assembler::addsd(dst, Address(rscratch1, 0));
964 }
965 }
966
967 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) {
968 if (reachable(src)) {
969 addss(dst, as_Address(src));
970 } else {
971 lea(rscratch1, src);
972 addss(dst, Address(rscratch1, 0));
973 }
974 }
975
976 void MacroAssembler::addpd(XMMRegister dst, AddressLiteral src) {
977 if (reachable(src)) {
978 Assembler::addpd(dst, as_Address(src));
979 } else {
980 lea(rscratch1, src);
981 Assembler::addpd(dst, Address(rscratch1, 0));
982 }
983 }
984
985 void MacroAssembler::align(int modulus) {
986 align(modulus, offset());
987 }
988
989 void MacroAssembler::align(int modulus, int target) {
990 if (target % modulus != 0) {
991 nop(modulus - (target % modulus));
992 }
993 }
994
995 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) {
996 // Used in sign-masking with aligned address.
997 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
998 if (reachable(src)) {
999 Assembler::andpd(dst, as_Address(src));
1000 } else {
1001 lea(rscratch1, src);
1002 Assembler::andpd(dst, Address(rscratch1, 0));
1003 }
1004 }
1005
1006 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) {
1007 // Used in sign-masking with aligned address.
1008 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
1009 if (reachable(src)) {
1010 Assembler::andps(dst, as_Address(src));
1011 } else {
1012 lea(rscratch1, src);
1013 Assembler::andps(dst, Address(rscratch1, 0));
1014 }
1015 }
1016
1017 void MacroAssembler::andptr(Register dst, int32_t imm32) {
1018 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32));
1019 }
1020
1021 void MacroAssembler::atomic_incl(Address counter_addr) {
1022 if (os::is_MP())
1023 lock();
1024 incrementl(counter_addr);
1025 }
1026
1027 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) {
1028 if (reachable(counter_addr)) {
1029 atomic_incl(as_Address(counter_addr));
1030 } else {
1031 lea(scr, counter_addr);
1032 atomic_incl(Address(scr, 0));
1033 }
1034 }
1035
1036 #ifdef _LP64
1037 void MacroAssembler::atomic_incq(Address counter_addr) {
1038 if (os::is_MP())
1039 lock();
1040 incrementq(counter_addr);
1041 }
1042
1043 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) {
1044 if (reachable(counter_addr)) {
1045 atomic_incq(as_Address(counter_addr));
1046 } else {
1047 lea(scr, counter_addr);
1048 atomic_incq(Address(scr, 0));
1049 }
1050 }
1051 #endif
1052
1053 // Writes to stack successive pages until offset reached to check for
1054 // stack overflow + shadow pages. This clobbers tmp.
1055 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
1056 movptr(tmp, rsp);
1057 // Bang stack for total size given plus shadow page size.
1058 // Bang one page at a time because large size can bang beyond yellow and
1059 // red zones.
1060 Label loop;
1061 bind(loop);
1062 movl(Address(tmp, (-os::vm_page_size())), size );
1063 subptr(tmp, os::vm_page_size());
1064 subl(size, os::vm_page_size());
1065 jcc(Assembler::greater, loop);
1066
1067 // Bang down shadow pages too.
1068 // At this point, (tmp-0) is the last address touched, so don't
1069 // touch it again. (It was touched as (tmp-pagesize) but then tmp
1070 // was post-decremented.) Skip this address by starting at i=1, and
1071 // touch a few more pages below. N.B. It is important to touch all
1072 // the way down including all pages in the shadow zone.
1073 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) {
1074 // this could be any sized move but this is can be a debugging crumb
1075 // so the bigger the better.
1076 movptr(Address(tmp, (-i*os::vm_page_size())), size );
1077 }
1078 }
1079
1080 void MacroAssembler::reserved_stack_check() {
1081 // testing if reserved zone needs to be enabled
1082 Label no_reserved_zone_enabling;
1083 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread);
1084 NOT_LP64(get_thread(rsi);)
1085
1086 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset()));
1087 jcc(Assembler::below, no_reserved_zone_enabling);
1088
1089 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread);
1090 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry()));
1091 should_not_reach_here();
1092
1093 bind(no_reserved_zone_enabling);
1094 }
1095
1096 int MacroAssembler::biased_locking_enter(Register lock_reg,
1097 Register obj_reg,
1098 Register swap_reg,
1099 Register tmp_reg,
1100 bool swap_reg_contains_mark,
1101 Label& done,
1102 Label* slow_case,
1103 BiasedLockingCounters* counters) {
1104 assert(UseBiasedLocking, "why call this otherwise?");
1105 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq");
1106 assert(tmp_reg != noreg, "tmp_reg must be supplied");
1107 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg);
1108 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout");
1109 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes());
1110 NOT_LP64( Address saved_mark_addr(lock_reg, 0); )
1111
1112 if (PrintBiasedLockingStatistics && counters == NULL) {
1113 counters = BiasedLocking::counters();
1114 }
1115 // Biased locking
1116 // See whether the lock is currently biased toward our thread and
1117 // whether the epoch is still valid
1118 // Note that the runtime guarantees sufficient alignment of JavaThread
1119 // pointers to allow age to be placed into low bits
1120 // First check to see whether biasing is even enabled for this object
1121 Label cas_label;
1122 int null_check_offset = -1;
1123 if (!swap_reg_contains_mark) {
1124 null_check_offset = offset();
1125 movptr(swap_reg, mark_addr);
1126 }
1127 movptr(tmp_reg, swap_reg);
1128 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place);
1129 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern);
1130 jcc(Assembler::notEqual, cas_label);
1131 // The bias pattern is present in the object's header. Need to check
1132 // whether the bias owner and the epoch are both still current.
1133 #ifndef _LP64
1134 // Note that because there is no current thread register on x86_32 we
1135 // need to store off the mark word we read out of the object to
1136 // avoid reloading it and needing to recheck invariants below. This
1137 // store is unfortunate but it makes the overall code shorter and
1138 // simpler.
1139 movptr(saved_mark_addr, swap_reg);
1140 #endif
1141 if (swap_reg_contains_mark) {
1142 null_check_offset = offset();
1143 }
1144 load_prototype_header(tmp_reg, obj_reg);
1145 #ifdef _LP64
1146 orptr(tmp_reg, r15_thread);
1147 xorptr(tmp_reg, swap_reg);
1148 Register header_reg = tmp_reg;
1149 #else
1150 xorptr(tmp_reg, swap_reg);
1151 get_thread(swap_reg);
1152 xorptr(swap_reg, tmp_reg);
1153 Register header_reg = swap_reg;
1154 #endif
1155 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place));
1156 if (counters != NULL) {
1157 cond_inc32(Assembler::zero,
1158 ExternalAddress((address) counters->biased_lock_entry_count_addr()));
1159 }
1160 jcc(Assembler::equal, done);
1161
1162 Label try_revoke_bias;
1163 Label try_rebias;
1164
1165 // At this point we know that the header has the bias pattern and
1166 // that we are not the bias owner in the current epoch. We need to
1167 // figure out more details about the state of the header in order to
1168 // know what operations can be legally performed on the object's
1169 // header.
1170
1171 // If the low three bits in the xor result aren't clear, that means
1172 // the prototype header is no longer biased and we have to revoke
1173 // the bias on this object.
1174 testptr(header_reg, markOopDesc::biased_lock_mask_in_place);
1175 jccb(Assembler::notZero, try_revoke_bias);
1176
1177 // Biasing is still enabled for this data type. See whether the
1178 // epoch of the current bias is still valid, meaning that the epoch
1179 // bits of the mark word are equal to the epoch bits of the
1180 // prototype header. (Note that the prototype header's epoch bits
1181 // only change at a safepoint.) If not, attempt to rebias the object
1182 // toward the current thread. Note that we must be absolutely sure
1183 // that the current epoch is invalid in order to do this because
1184 // otherwise the manipulations it performs on the mark word are
1185 // illegal.
1186 testptr(header_reg, markOopDesc::epoch_mask_in_place);
1187 jccb(Assembler::notZero, try_rebias);
1188
1189 // The epoch of the current bias is still valid but we know nothing
1190 // about the owner; it might be set or it might be clear. Try to
1191 // acquire the bias of the object using an atomic operation. If this
1192 // fails we will go in to the runtime to revoke the object's bias.
1193 // Note that we first construct the presumed unbiased header so we
1194 // don't accidentally blow away another thread's valid bias.
1195 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1196 andptr(swap_reg,
1197 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place);
1198 #ifdef _LP64
1199 movptr(tmp_reg, swap_reg);
1200 orptr(tmp_reg, r15_thread);
1201 #else
1202 get_thread(tmp_reg);
1203 orptr(tmp_reg, swap_reg);
1204 #endif
1205 if (os::is_MP()) {
1206 lock();
1207 }
1208 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1209 // If the biasing toward our thread failed, this means that
1210 // another thread succeeded in biasing it toward itself and we
1211 // need to revoke that bias. The revocation will occur in the
1212 // interpreter runtime in the slow case.
1213 if (counters != NULL) {
1214 cond_inc32(Assembler::zero,
1215 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr()));
1216 }
1217 if (slow_case != NULL) {
1218 jcc(Assembler::notZero, *slow_case);
1219 }
1220 jmp(done);
1221
1222 bind(try_rebias);
1223 // At this point we know the epoch has expired, meaning that the
1224 // current "bias owner", if any, is actually invalid. Under these
1225 // circumstances _only_, we are allowed to use the current header's
1226 // value as the comparison value when doing the cas to acquire the
1227 // bias in the current epoch. In other words, we allow transfer of
1228 // the bias from one thread to another directly in this situation.
1229 //
1230 // FIXME: due to a lack of registers we currently blow away the age
1231 // bits in this situation. Should attempt to preserve them.
1232 load_prototype_header(tmp_reg, obj_reg);
1233 #ifdef _LP64
1234 orptr(tmp_reg, r15_thread);
1235 #else
1236 get_thread(swap_reg);
1237 orptr(tmp_reg, swap_reg);
1238 movptr(swap_reg, saved_mark_addr);
1239 #endif
1240 if (os::is_MP()) {
1241 lock();
1242 }
1243 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1244 // If the biasing toward our thread failed, then another thread
1245 // succeeded in biasing it toward itself and we need to revoke that
1246 // bias. The revocation will occur in the runtime in the slow case.
1247 if (counters != NULL) {
1248 cond_inc32(Assembler::zero,
1249 ExternalAddress((address) counters->rebiased_lock_entry_count_addr()));
1250 }
1251 if (slow_case != NULL) {
1252 jcc(Assembler::notZero, *slow_case);
1253 }
1254 jmp(done);
1255
1256 bind(try_revoke_bias);
1257 // The prototype mark in the klass doesn't have the bias bit set any
1258 // more, indicating that objects of this data type are not supposed
1259 // to be biased any more. We are going to try to reset the mark of
1260 // this object to the prototype value and fall through to the
1261 // CAS-based locking scheme. Note that if our CAS fails, it means
1262 // that another thread raced us for the privilege of revoking the
1263 // bias of this particular object, so it's okay to continue in the
1264 // normal locking code.
1265 //
1266 // FIXME: due to a lack of registers we currently blow away the age
1267 // bits in this situation. Should attempt to preserve them.
1268 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1269 load_prototype_header(tmp_reg, obj_reg);
1270 if (os::is_MP()) {
1271 lock();
1272 }
1273 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1274 // Fall through to the normal CAS-based lock, because no matter what
1275 // the result of the above CAS, some thread must have succeeded in
1276 // removing the bias bit from the object's header.
1277 if (counters != NULL) {
1278 cond_inc32(Assembler::zero,
1279 ExternalAddress((address) counters->revoked_lock_entry_count_addr()));
1280 }
1281
1282 bind(cas_label);
1283
1284 return null_check_offset;
1285 }
1286
1287 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) {
1288 assert(UseBiasedLocking, "why call this otherwise?");
1289
1290 // Check for biased locking unlock case, which is a no-op
1291 // Note: we do not have to check the thread ID for two reasons.
1292 // First, the interpreter checks for IllegalMonitorStateException at
1293 // a higher level. Second, if the bias was revoked while we held the
1294 // lock, the object could not be rebiased toward another thread, so
1295 // the bias bit would be clear.
1296 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1297 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place);
1298 cmpptr(temp_reg, markOopDesc::biased_lock_pattern);
1299 jcc(Assembler::equal, done);
1300 }
1301
1302 #ifdef COMPILER2
1303
1304 #if INCLUDE_RTM_OPT
1305
1306 // Update rtm_counters based on abort status
1307 // input: abort_status
1308 // rtm_counters (RTMLockingCounters*)
1309 // flags are killed
1310 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) {
1311
1312 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset()));
1313 if (PrintPreciseRTMLockingStatistics) {
1314 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) {
1315 Label check_abort;
1316 testl(abort_status, (1<<i));
1317 jccb(Assembler::equal, check_abort);
1318 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx))));
1319 bind(check_abort);
1320 }
1321 }
1322 }
1323
1324 // Branch if (random & (count-1) != 0), count is 2^n
1325 // tmp, scr and flags are killed
1326 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) {
1327 assert(tmp == rax, "");
1328 assert(scr == rdx, "");
1329 rdtsc(); // modifies EDX:EAX
1330 andptr(tmp, count-1);
1331 jccb(Assembler::notZero, brLabel);
1332 }
1333
1334 // Perform abort ratio calculation, set no_rtm bit if high ratio
1335 // input: rtm_counters_Reg (RTMLockingCounters* address)
1336 // tmpReg, rtm_counters_Reg and flags are killed
1337 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg,
1338 Register rtm_counters_Reg,
1339 RTMLockingCounters* rtm_counters,
1340 Metadata* method_data) {
1341 Label L_done, L_check_always_rtm1, L_check_always_rtm2;
1342
1343 if (RTMLockingCalculationDelay > 0) {
1344 // Delay calculation
1345 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg);
1346 testptr(tmpReg, tmpReg);
1347 jccb(Assembler::equal, L_done);
1348 }
1349 // Abort ratio calculation only if abort_count > RTMAbortThreshold
1350 // Aborted transactions = abort_count * 100
1351 // All transactions = total_count * RTMTotalCountIncrRate
1352 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio)
1353
1354 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset()));
1355 cmpptr(tmpReg, RTMAbortThreshold);
1356 jccb(Assembler::below, L_check_always_rtm2);
1357 imulptr(tmpReg, tmpReg, 100);
1358
1359 Register scrReg = rtm_counters_Reg;
1360 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1361 imulptr(scrReg, scrReg, RTMTotalCountIncrRate);
1362 imulptr(scrReg, scrReg, RTMAbortRatio);
1363 cmpptr(tmpReg, scrReg);
1364 jccb(Assembler::below, L_check_always_rtm1);
1365 if (method_data != NULL) {
1366 // set rtm_state to "no rtm" in MDO
1367 mov_metadata(tmpReg, method_data);
1368 if (os::is_MP()) {
1369 lock();
1370 }
1371 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM);
1372 }
1373 jmpb(L_done);
1374 bind(L_check_always_rtm1);
1375 // Reload RTMLockingCounters* address
1376 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1377 bind(L_check_always_rtm2);
1378 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1379 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate);
1380 jccb(Assembler::below, L_done);
1381 if (method_data != NULL) {
1382 // set rtm_state to "always rtm" in MDO
1383 mov_metadata(tmpReg, method_data);
1384 if (os::is_MP()) {
1385 lock();
1386 }
1387 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM);
1388 }
1389 bind(L_done);
1390 }
1391
1392 // Update counters and perform abort ratio calculation
1393 // input: abort_status_Reg
1394 // rtm_counters_Reg, flags are killed
1395 void MacroAssembler::rtm_profiling(Register abort_status_Reg,
1396 Register rtm_counters_Reg,
1397 RTMLockingCounters* rtm_counters,
1398 Metadata* method_data,
1399 bool profile_rtm) {
1400
1401 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1402 // update rtm counters based on rax value at abort
1403 // reads abort_status_Reg, updates flags
1404 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1405 rtm_counters_update(abort_status_Reg, rtm_counters_Reg);
1406 if (profile_rtm) {
1407 // Save abort status because abort_status_Reg is used by following code.
1408 if (RTMRetryCount > 0) {
1409 push(abort_status_Reg);
1410 }
1411 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1412 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data);
1413 // restore abort status
1414 if (RTMRetryCount > 0) {
1415 pop(abort_status_Reg);
1416 }
1417 }
1418 }
1419
1420 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4)
1421 // inputs: retry_count_Reg
1422 // : abort_status_Reg
1423 // output: retry_count_Reg decremented by 1
1424 // flags are killed
1425 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) {
1426 Label doneRetry;
1427 assert(abort_status_Reg == rax, "");
1428 // The abort reason bits are in eax (see all states in rtmLocking.hpp)
1429 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4)
1430 // if reason is in 0x6 and retry count != 0 then retry
1431 andptr(abort_status_Reg, 0x6);
1432 jccb(Assembler::zero, doneRetry);
1433 testl(retry_count_Reg, retry_count_Reg);
1434 jccb(Assembler::zero, doneRetry);
1435 pause();
1436 decrementl(retry_count_Reg);
1437 jmp(retryLabel);
1438 bind(doneRetry);
1439 }
1440
1441 // Spin and retry if lock is busy,
1442 // inputs: box_Reg (monitor address)
1443 // : retry_count_Reg
1444 // output: retry_count_Reg decremented by 1
1445 // : clear z flag if retry count exceeded
1446 // tmp_Reg, scr_Reg, flags are killed
1447 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg,
1448 Register tmp_Reg, Register scr_Reg, Label& retryLabel) {
1449 Label SpinLoop, SpinExit, doneRetry;
1450 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1451
1452 testl(retry_count_Reg, retry_count_Reg);
1453 jccb(Assembler::zero, doneRetry);
1454 decrementl(retry_count_Reg);
1455 movptr(scr_Reg, RTMSpinLoopCount);
1456
1457 bind(SpinLoop);
1458 pause();
1459 decrementl(scr_Reg);
1460 jccb(Assembler::lessEqual, SpinExit);
1461 movptr(tmp_Reg, Address(box_Reg, owner_offset));
1462 testptr(tmp_Reg, tmp_Reg);
1463 jccb(Assembler::notZero, SpinLoop);
1464
1465 bind(SpinExit);
1466 jmp(retryLabel);
1467 bind(doneRetry);
1468 incrementl(retry_count_Reg); // clear z flag
1469 }
1470
1471 // Use RTM for normal stack locks
1472 // Input: objReg (object to lock)
1473 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg,
1474 Register retry_on_abort_count_Reg,
1475 RTMLockingCounters* stack_rtm_counters,
1476 Metadata* method_data, bool profile_rtm,
1477 Label& DONE_LABEL, Label& IsInflated) {
1478 assert(UseRTMForStackLocks, "why call this otherwise?");
1479 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1480 assert(tmpReg == rax, "");
1481 assert(scrReg == rdx, "");
1482 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1483
1484 if (RTMRetryCount > 0) {
1485 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1486 bind(L_rtm_retry);
1487 }
1488 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes()));
1489 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1490 jcc(Assembler::notZero, IsInflated);
1491
1492 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1493 Label L_noincrement;
1494 if (RTMTotalCountIncrRate > 1) {
1495 // tmpReg, scrReg and flags are killed
1496 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement);
1497 }
1498 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM");
1499 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg);
1500 bind(L_noincrement);
1501 }
1502 xbegin(L_on_abort);
1503 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword
1504 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1505 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1506 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked
1507
1508 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1509 if (UseRTMXendForLockBusy) {
1510 xend();
1511 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry)
1512 jmp(L_decrement_retry);
1513 }
1514 else {
1515 xabort(0);
1516 }
1517 bind(L_on_abort);
1518 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1519 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm);
1520 }
1521 bind(L_decrement_retry);
1522 if (RTMRetryCount > 0) {
1523 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1524 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1525 }
1526 }
1527
1528 // Use RTM for inflating locks
1529 // inputs: objReg (object to lock)
1530 // boxReg (on-stack box address (displaced header location) - KILLED)
1531 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value)
1532 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg,
1533 Register scrReg, Register retry_on_busy_count_Reg,
1534 Register retry_on_abort_count_Reg,
1535 RTMLockingCounters* rtm_counters,
1536 Metadata* method_data, bool profile_rtm,
1537 Label& DONE_LABEL) {
1538 assert(UseRTMLocking, "why call this otherwise?");
1539 assert(tmpReg == rax, "");
1540 assert(scrReg == rdx, "");
1541 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1542 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1543
1544 // Without cast to int32_t a movptr will destroy r10 which is typically obj
1545 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1546 movptr(boxReg, tmpReg); // Save ObjectMonitor address
1547
1548 if (RTMRetryCount > 0) {
1549 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy
1550 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1551 bind(L_rtm_retry);
1552 }
1553 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1554 Label L_noincrement;
1555 if (RTMTotalCountIncrRate > 1) {
1556 // tmpReg, scrReg and flags are killed
1557 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement);
1558 }
1559 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1560 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg);
1561 bind(L_noincrement);
1562 }
1563 xbegin(L_on_abort);
1564 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes()));
1565 movptr(tmpReg, Address(tmpReg, owner_offset));
1566 testptr(tmpReg, tmpReg);
1567 jcc(Assembler::zero, DONE_LABEL);
1568 if (UseRTMXendForLockBusy) {
1569 xend();
1570 jmp(L_decrement_retry);
1571 }
1572 else {
1573 xabort(0);
1574 }
1575 bind(L_on_abort);
1576 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1577 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1578 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm);
1579 }
1580 if (RTMRetryCount > 0) {
1581 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1582 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1583 }
1584
1585 movptr(tmpReg, Address(boxReg, owner_offset)) ;
1586 testptr(tmpReg, tmpReg) ;
1587 jccb(Assembler::notZero, L_decrement_retry) ;
1588
1589 // Appears unlocked - try to swing _owner from null to non-null.
1590 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1591 #ifdef _LP64
1592 Register threadReg = r15_thread;
1593 #else
1594 get_thread(scrReg);
1595 Register threadReg = scrReg;
1596 #endif
1597 if (os::is_MP()) {
1598 lock();
1599 }
1600 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg
1601
1602 if (RTMRetryCount > 0) {
1603 // success done else retry
1604 jccb(Assembler::equal, DONE_LABEL) ;
1605 bind(L_decrement_retry);
1606 // Spin and retry if lock is busy.
1607 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry);
1608 }
1609 else {
1610 bind(L_decrement_retry);
1611 }
1612 }
1613
1614 #endif // INCLUDE_RTM_OPT
1615
1616 // Fast_Lock and Fast_Unlock used by C2
1617
1618 // Because the transitions from emitted code to the runtime
1619 // monitorenter/exit helper stubs are so slow it's critical that
1620 // we inline both the stack-locking fast-path and the inflated fast path.
1621 //
1622 // See also: cmpFastLock and cmpFastUnlock.
1623 //
1624 // What follows is a specialized inline transliteration of the code
1625 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat
1626 // another option would be to emit TrySlowEnter and TrySlowExit methods
1627 // at startup-time. These methods would accept arguments as
1628 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
1629 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply
1630 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
1631 // In practice, however, the # of lock sites is bounded and is usually small.
1632 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
1633 // if the processor uses simple bimodal branch predictors keyed by EIP
1634 // Since the helper routines would be called from multiple synchronization
1635 // sites.
1636 //
1637 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
1638 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
1639 // to those specialized methods. That'd give us a mostly platform-independent
1640 // implementation that the JITs could optimize and inline at their pleasure.
1641 // Done correctly, the only time we'd need to cross to native could would be
1642 // to park() or unpark() threads. We'd also need a few more unsafe operators
1643 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
1644 // (b) explicit barriers or fence operations.
1645 //
1646 // TODO:
1647 //
1648 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
1649 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
1650 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
1651 // the lock operators would typically be faster than reifying Self.
1652 //
1653 // * Ideally I'd define the primitives as:
1654 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
1655 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
1656 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
1657 // Instead, we're stuck with a rather awkward and brittle register assignments below.
1658 // Furthermore the register assignments are overconstrained, possibly resulting in
1659 // sub-optimal code near the synchronization site.
1660 //
1661 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
1662 // Alternately, use a better sp-proximity test.
1663 //
1664 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
1665 // Either one is sufficient to uniquely identify a thread.
1666 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
1667 //
1668 // * Intrinsify notify() and notifyAll() for the common cases where the
1669 // object is locked by the calling thread but the waitlist is empty.
1670 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
1671 //
1672 // * use jccb and jmpb instead of jcc and jmp to improve code density.
1673 // But beware of excessive branch density on AMD Opterons.
1674 //
1675 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
1676 // or failure of the fast-path. If the fast-path fails then we pass
1677 // control to the slow-path, typically in C. In Fast_Lock and
1678 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2
1679 // will emit a conditional branch immediately after the node.
1680 // So we have branches to branches and lots of ICC.ZF games.
1681 // Instead, it might be better to have C2 pass a "FailureLabel"
1682 // into Fast_Lock and Fast_Unlock. In the case of success, control
1683 // will drop through the node. ICC.ZF is undefined at exit.
1684 // In the case of failure, the node will branch directly to the
1685 // FailureLabel
1686
1687
1688 // obj: object to lock
1689 // box: on-stack box address (displaced header location) - KILLED
1690 // rax,: tmp -- KILLED
1691 // scr: tmp -- KILLED
1692 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg,
1693 Register scrReg, Register cx1Reg, Register cx2Reg,
1694 BiasedLockingCounters* counters,
1695 RTMLockingCounters* rtm_counters,
1696 RTMLockingCounters* stack_rtm_counters,
1697 Metadata* method_data,
1698 bool use_rtm, bool profile_rtm) {
1699 // Ensure the register assignments are disjoint
1700 assert(tmpReg == rax, "");
1701
1702 if (use_rtm) {
1703 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg);
1704 } else {
1705 assert(cx1Reg == noreg, "");
1706 assert(cx2Reg == noreg, "");
1707 assert_different_registers(objReg, boxReg, tmpReg, scrReg);
1708 }
1709
1710 if (counters != NULL) {
1711 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg);
1712 }
1713 if (EmitSync & 1) {
1714 // set box->dhw = markOopDesc::unused_mark()
1715 // Force all sync thru slow-path: slow_enter() and slow_exit()
1716 movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1717 cmpptr (rsp, (int32_t)NULL_WORD);
1718 } else {
1719 // Possible cases that we'll encounter in fast_lock
1720 // ------------------------------------------------
1721 // * Inflated
1722 // -- unlocked
1723 // -- Locked
1724 // = by self
1725 // = by other
1726 // * biased
1727 // -- by Self
1728 // -- by other
1729 // * neutral
1730 // * stack-locked
1731 // -- by self
1732 // = sp-proximity test hits
1733 // = sp-proximity test generates false-negative
1734 // -- by other
1735 //
1736
1737 Label IsInflated, DONE_LABEL;
1738
1739 // it's stack-locked, biased or neutral
1740 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
1741 // order to reduce the number of conditional branches in the most common cases.
1742 // Beware -- there's a subtle invariant that fetch of the markword
1743 // at [FETCH], below, will never observe a biased encoding (*101b).
1744 // If this invariant is not held we risk exclusion (safety) failure.
1745 if (UseBiasedLocking && !UseOptoBiasInlining) {
1746 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters);
1747 }
1748
1749 #if INCLUDE_RTM_OPT
1750 if (UseRTMForStackLocks && use_rtm) {
1751 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg,
1752 stack_rtm_counters, method_data, profile_rtm,
1753 DONE_LABEL, IsInflated);
1754 }
1755 #endif // INCLUDE_RTM_OPT
1756
1757 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // [FETCH]
1758 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1759 jccb(Assembler::notZero, IsInflated);
1760
1761 // Attempt stack-locking ...
1762 orptr (tmpReg, markOopDesc::unlocked_value);
1763 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
1764 if (os::is_MP()) {
1765 lock();
1766 }
1767 cmpxchgptr(boxReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Updates tmpReg
1768 if (counters != NULL) {
1769 cond_inc32(Assembler::equal,
1770 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1771 }
1772 jcc(Assembler::equal, DONE_LABEL); // Success
1773
1774 // Recursive locking.
1775 // The object is stack-locked: markword contains stack pointer to BasicLock.
1776 // Locked by current thread if difference with current SP is less than one page.
1777 subptr(tmpReg, rsp);
1778 // Next instruction set ZFlag == 1 (Success) if difference is less then one page.
1779 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) );
1780 movptr(Address(boxReg, 0), tmpReg);
1781 if (counters != NULL) {
1782 cond_inc32(Assembler::equal,
1783 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1784 }
1785 jmp(DONE_LABEL);
1786
1787 bind(IsInflated);
1788 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value
1789
1790 #if INCLUDE_RTM_OPT
1791 // Use the same RTM locking code in 32- and 64-bit VM.
1792 if (use_rtm) {
1793 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg,
1794 rtm_counters, method_data, profile_rtm, DONE_LABEL);
1795 } else {
1796 #endif // INCLUDE_RTM_OPT
1797
1798 #ifndef _LP64
1799 // The object is inflated.
1800
1801 // boxReg refers to the on-stack BasicLock in the current frame.
1802 // We'd like to write:
1803 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices.
1804 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
1805 // additional latency as we have another ST in the store buffer that must drain.
1806
1807 if (EmitSync & 8192) {
1808 movptr(Address(boxReg, 0), 3); // results in ST-before-CAS penalty
1809 get_thread (scrReg);
1810 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1811 movptr(tmpReg, NULL_WORD); // consider: xor vs mov
1812 if (os::is_MP()) {
1813 lock();
1814 }
1815 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1816 } else
1817 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS
1818 // register juggle because we need tmpReg for cmpxchgptr below
1819 movptr(scrReg, boxReg);
1820 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1821
1822 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1823 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1824 // prefetchw [eax + Offset(_owner)-2]
1825 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1826 }
1827
1828 if ((EmitSync & 64) == 0) {
1829 // Optimistic form: consider XORL tmpReg,tmpReg
1830 movptr(tmpReg, NULL_WORD);
1831 } else {
1832 // Can suffer RTS->RTO upgrades on shared or cold $ lines
1833 // Test-And-CAS instead of CAS
1834 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner
1835 testptr(tmpReg, tmpReg); // Locked ?
1836 jccb (Assembler::notZero, DONE_LABEL);
1837 }
1838
1839 // Appears unlocked - try to swing _owner from null to non-null.
1840 // Ideally, I'd manifest "Self" with get_thread and then attempt
1841 // to CAS the register containing Self into m->Owner.
1842 // But we don't have enough registers, so instead we can either try to CAS
1843 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
1844 // we later store "Self" into m->Owner. Transiently storing a stack address
1845 // (rsp or the address of the box) into m->owner is harmless.
1846 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1847 if (os::is_MP()) {
1848 lock();
1849 }
1850 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1851 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3
1852 // If we weren't able to swing _owner from NULL to the BasicLock
1853 // then take the slow path.
1854 jccb (Assembler::notZero, DONE_LABEL);
1855 // update _owner from BasicLock to thread
1856 get_thread (scrReg); // beware: clobbers ICCs
1857 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg);
1858 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success
1859
1860 // If the CAS fails we can either retry or pass control to the slow-path.
1861 // We use the latter tactic.
1862 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1863 // If the CAS was successful ...
1864 // Self has acquired the lock
1865 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1866 // Intentional fall-through into DONE_LABEL ...
1867 } else {
1868 movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())); // results in ST-before-CAS penalty
1869 movptr(boxReg, tmpReg);
1870
1871 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1872 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1873 // prefetchw [eax + Offset(_owner)-2]
1874 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1875 }
1876
1877 if ((EmitSync & 64) == 0) {
1878 // Optimistic form
1879 xorptr (tmpReg, tmpReg);
1880 } else {
1881 // Can suffer RTS->RTO upgrades on shared or cold $ lines
1882 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner
1883 testptr(tmpReg, tmpReg); // Locked ?
1884 jccb (Assembler::notZero, DONE_LABEL);
1885 }
1886
1887 // Appears unlocked - try to swing _owner from null to non-null.
1888 // Use either "Self" (in scr) or rsp as thread identity in _owner.
1889 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1890 get_thread (scrReg);
1891 if (os::is_MP()) {
1892 lock();
1893 }
1894 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1895
1896 // If the CAS fails we can either retry or pass control to the slow-path.
1897 // We use the latter tactic.
1898 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1899 // If the CAS was successful ...
1900 // Self has acquired the lock
1901 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1902 // Intentional fall-through into DONE_LABEL ...
1903 }
1904 #else // _LP64
1905 // It's inflated
1906 movq(scrReg, tmpReg);
1907 xorq(tmpReg, tmpReg);
1908
1909 if (os::is_MP()) {
1910 lock();
1911 }
1912 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1913 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark().
1914 // Without cast to int32_t movptr will destroy r10 which is typically obj.
1915 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1916 // Intentional fall-through into DONE_LABEL ...
1917 // Propagate ICC.ZF from CAS above into DONE_LABEL.
1918 #endif // _LP64
1919 #if INCLUDE_RTM_OPT
1920 } // use_rtm()
1921 #endif
1922 // DONE_LABEL is a hot target - we'd really like to place it at the
1923 // start of cache line by padding with NOPs.
1924 // See the AMD and Intel software optimization manuals for the
1925 // most efficient "long" NOP encodings.
1926 // Unfortunately none of our alignment mechanisms suffice.
1927 bind(DONE_LABEL);
1928
1929 // At DONE_LABEL the icc ZFlag is set as follows ...
1930 // Fast_Unlock uses the same protocol.
1931 // ZFlag == 1 -> Success
1932 // ZFlag == 0 -> Failure - force control through the slow-path
1933 }
1934 }
1935
1936 // obj: object to unlock
1937 // box: box address (displaced header location), killed. Must be EAX.
1938 // tmp: killed, cannot be obj nor box.
1939 //
1940 // Some commentary on balanced locking:
1941 //
1942 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
1943 // Methods that don't have provably balanced locking are forced to run in the
1944 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
1945 // The interpreter provides two properties:
1946 // I1: At return-time the interpreter automatically and quietly unlocks any
1947 // objects acquired the current activation (frame). Recall that the
1948 // interpreter maintains an on-stack list of locks currently held by
1949 // a frame.
1950 // I2: If a method attempts to unlock an object that is not held by the
1951 // the frame the interpreter throws IMSX.
1952 //
1953 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
1954 // B() doesn't have provably balanced locking so it runs in the interpreter.
1955 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
1956 // is still locked by A().
1957 //
1958 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
1959 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
1960 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
1961 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
1962 // Arguably given that the spec legislates the JNI case as undefined our implementation
1963 // could reasonably *avoid* checking owner in Fast_Unlock().
1964 // In the interest of performance we elide m->Owner==Self check in unlock.
1965 // A perfectly viable alternative is to elide the owner check except when
1966 // Xcheck:jni is enabled.
1967
1968 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) {
1969 assert(boxReg == rax, "");
1970 assert_different_registers(objReg, boxReg, tmpReg);
1971
1972 if (EmitSync & 4) {
1973 // Disable - inhibit all inlining. Force control through the slow-path
1974 cmpptr (rsp, 0);
1975 } else {
1976 Label DONE_LABEL, Stacked, CheckSucc;
1977
1978 // Critically, the biased locking test must have precedence over
1979 // and appear before the (box->dhw == 0) recursive stack-lock test.
1980 if (UseBiasedLocking && !UseOptoBiasInlining) {
1981 biased_locking_exit(objReg, tmpReg, DONE_LABEL);
1982 }
1983
1984 #if INCLUDE_RTM_OPT
1985 if (UseRTMForStackLocks && use_rtm) {
1986 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1987 Label L_regular_unlock;
1988 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword
1989 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1990 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1991 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock
1992 xend(); // otherwise end...
1993 jmp(DONE_LABEL); // ... and we're done
1994 bind(L_regular_unlock);
1995 }
1996 #endif
1997
1998 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header
1999 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock
2000 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Examine the object's markword
2001 testptr(tmpReg, markOopDesc::monitor_value); // Inflated?
2002 jccb (Assembler::zero, Stacked);
2003
2004 // It's inflated.
2005 #if INCLUDE_RTM_OPT
2006 if (use_rtm) {
2007 Label L_regular_inflated_unlock;
2008 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
2009 movptr(boxReg, Address(tmpReg, owner_offset));
2010 testptr(boxReg, boxReg);
2011 jccb(Assembler::notZero, L_regular_inflated_unlock);
2012 xend();
2013 jmpb(DONE_LABEL);
2014 bind(L_regular_inflated_unlock);
2015 }
2016 #endif
2017
2018 // Despite our balanced locking property we still check that m->_owner == Self
2019 // as java routines or native JNI code called by this thread might
2020 // have released the lock.
2021 // Refer to the comments in synchronizer.cpp for how we might encode extra
2022 // state in _succ so we can avoid fetching EntryList|cxq.
2023 //
2024 // I'd like to add more cases in fast_lock() and fast_unlock() --
2025 // such as recursive enter and exit -- but we have to be wary of
2026 // I$ bloat, T$ effects and BP$ effects.
2027 //
2028 // If there's no contention try a 1-0 exit. That is, exit without
2029 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
2030 // we detect and recover from the race that the 1-0 exit admits.
2031 //
2032 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
2033 // before it STs null into _owner, releasing the lock. Updates
2034 // to data protected by the critical section must be visible before
2035 // we drop the lock (and thus before any other thread could acquire
2036 // the lock and observe the fields protected by the lock).
2037 // IA32's memory-model is SPO, so STs are ordered with respect to
2038 // each other and there's no need for an explicit barrier (fence).
2039 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
2040 #ifndef _LP64
2041 get_thread (boxReg);
2042 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
2043 // prefetchw [ebx + Offset(_owner)-2]
2044 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2045 }
2046
2047 // Note that we could employ various encoding schemes to reduce
2048 // the number of loads below (currently 4) to just 2 or 3.
2049 // Refer to the comments in synchronizer.cpp.
2050 // In practice the chain of fetches doesn't seem to impact performance, however.
2051 xorptr(boxReg, boxReg);
2052 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) {
2053 // Attempt to reduce branch density - AMD's branch predictor.
2054 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2055 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2056 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2057 jccb (Assembler::notZero, DONE_LABEL);
2058 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2059 jmpb (DONE_LABEL);
2060 } else {
2061 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2062 jccb (Assembler::notZero, DONE_LABEL);
2063 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2064 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2065 jccb (Assembler::notZero, CheckSucc);
2066 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2067 jmpb (DONE_LABEL);
2068 }
2069
2070 // The Following code fragment (EmitSync & 65536) improves the performance of
2071 // contended applications and contended synchronization microbenchmarks.
2072 // Unfortunately the emission of the code - even though not executed - causes regressions
2073 // in scimark and jetstream, evidently because of $ effects. Replacing the code
2074 // with an equal number of never-executed NOPs results in the same regression.
2075 // We leave it off by default.
2076
2077 if ((EmitSync & 65536) != 0) {
2078 Label LSuccess, LGoSlowPath ;
2079
2080 bind (CheckSucc);
2081
2082 // Optional pre-test ... it's safe to elide this
2083 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2084 jccb(Assembler::zero, LGoSlowPath);
2085
2086 // We have a classic Dekker-style idiom:
2087 // ST m->_owner = 0 ; MEMBAR; LD m->_succ
2088 // There are a number of ways to implement the barrier:
2089 // (1) lock:andl &m->_owner, 0
2090 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form.
2091 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0
2092 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8
2093 // (2) If supported, an explicit MFENCE is appealing.
2094 // In older IA32 processors MFENCE is slower than lock:add or xchg
2095 // particularly if the write-buffer is full as might be the case if
2096 // if stores closely precede the fence or fence-equivalent instruction.
2097 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2098 // as the situation has changed with Nehalem and Shanghai.
2099 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack
2100 // The $lines underlying the top-of-stack should be in M-state.
2101 // The locked add instruction is serializing, of course.
2102 // (4) Use xchg, which is serializing
2103 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works
2104 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0.
2105 // The integer condition codes will tell us if succ was 0.
2106 // Since _succ and _owner should reside in the same $line and
2107 // we just stored into _owner, it's likely that the $line
2108 // remains in M-state for the lock:orl.
2109 //
2110 // We currently use (3), although it's likely that switching to (2)
2111 // is correct for the future.
2112
2113 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2114 if (os::is_MP()) {
2115 lock(); addptr(Address(rsp, 0), 0);
2116 }
2117 // Ratify _succ remains non-null
2118 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0);
2119 jccb (Assembler::notZero, LSuccess);
2120
2121 xorptr(boxReg, boxReg); // box is really EAX
2122 if (os::is_MP()) { lock(); }
2123 cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2124 // There's no successor so we tried to regrab the lock with the
2125 // placeholder value. If that didn't work, then another thread
2126 // grabbed the lock so we're done (and exit was a success).
2127 jccb (Assembler::notEqual, LSuccess);
2128 // Since we're low on registers we installed rsp as a placeholding in _owner.
2129 // Now install Self over rsp. This is safe as we're transitioning from
2130 // non-null to non=null
2131 get_thread (boxReg);
2132 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg);
2133 // Intentional fall-through into LGoSlowPath ...
2134
2135 bind (LGoSlowPath);
2136 orptr(boxReg, 1); // set ICC.ZF=0 to indicate failure
2137 jmpb (DONE_LABEL);
2138
2139 bind (LSuccess);
2140 xorptr(boxReg, boxReg); // set ICC.ZF=1 to indicate success
2141 jmpb (DONE_LABEL);
2142 }
2143
2144 bind (Stacked);
2145 // It's not inflated and it's not recursively stack-locked and it's not biased.
2146 // It must be stack-locked.
2147 // Try to reset the header to displaced header.
2148 // The "box" value on the stack is stable, so we can reload
2149 // and be assured we observe the same value as above.
2150 movptr(tmpReg, Address(boxReg, 0));
2151 if (os::is_MP()) {
2152 lock();
2153 }
2154 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box
2155 // Intention fall-thru into DONE_LABEL
2156
2157 // DONE_LABEL is a hot target - we'd really like to place it at the
2158 // start of cache line by padding with NOPs.
2159 // See the AMD and Intel software optimization manuals for the
2160 // most efficient "long" NOP encodings.
2161 // Unfortunately none of our alignment mechanisms suffice.
2162 if ((EmitSync & 65536) == 0) {
2163 bind (CheckSucc);
2164 }
2165 #else // _LP64
2166 // It's inflated
2167 if (EmitSync & 1024) {
2168 // Emit code to check that _owner == Self
2169 // We could fold the _owner test into subsequent code more efficiently
2170 // than using a stand-alone check, but since _owner checking is off by
2171 // default we don't bother. We also might consider predicating the
2172 // _owner==Self check on Xcheck:jni or running on a debug build.
2173 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2174 xorptr(boxReg, r15_thread);
2175 } else {
2176 xorptr(boxReg, boxReg);
2177 }
2178 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2179 jccb (Assembler::notZero, DONE_LABEL);
2180 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2181 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2182 jccb (Assembler::notZero, CheckSucc);
2183 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2184 jmpb (DONE_LABEL);
2185
2186 if ((EmitSync & 65536) == 0) {
2187 // Try to avoid passing control into the slow_path ...
2188 Label LSuccess, LGoSlowPath ;
2189 bind (CheckSucc);
2190
2191 // The following optional optimization can be elided if necessary
2192 // Effectively: if (succ == null) goto SlowPath
2193 // The code reduces the window for a race, however,
2194 // and thus benefits performance.
2195 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2196 jccb (Assembler::zero, LGoSlowPath);
2197
2198 xorptr(boxReg, boxReg);
2199 if ((EmitSync & 16) && os::is_MP()) {
2200 xchgptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2201 } else {
2202 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2203 if (os::is_MP()) {
2204 // Memory barrier/fence
2205 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ
2206 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack.
2207 // This is faster on Nehalem and AMD Shanghai/Barcelona.
2208 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2209 // We might also restructure (ST Owner=0;barrier;LD _Succ) to
2210 // (mov box,0; xchgq box, &m->Owner; LD _succ) .
2211 lock(); addl(Address(rsp, 0), 0);
2212 }
2213 }
2214 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2215 jccb (Assembler::notZero, LSuccess);
2216
2217 // Rare inopportune interleaving - race.
2218 // The successor vanished in the small window above.
2219 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor.
2220 // We need to ensure progress and succession.
2221 // Try to reacquire the lock.
2222 // If that fails then the new owner is responsible for succession and this
2223 // thread needs to take no further action and can exit via the fast path (success).
2224 // If the re-acquire succeeds then pass control into the slow path.
2225 // As implemented, this latter mode is horrible because we generated more
2226 // coherence traffic on the lock *and* artifically extended the critical section
2227 // length while by virtue of passing control into the slow path.
2228
2229 // box is really RAX -- the following CMPXCHG depends on that binding
2230 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R)
2231 if (os::is_MP()) { lock(); }
2232 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2233 // There's no successor so we tried to regrab the lock.
2234 // If that didn't work, then another thread grabbed the
2235 // lock so we're done (and exit was a success).
2236 jccb (Assembler::notEqual, LSuccess);
2237 // Intentional fall-through into slow-path
2238
2239 bind (LGoSlowPath);
2240 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure
2241 jmpb (DONE_LABEL);
2242
2243 bind (LSuccess);
2244 testl (boxReg, 0); // set ICC.ZF=1 to indicate success
2245 jmpb (DONE_LABEL);
2246 }
2247
2248 bind (Stacked);
2249 movptr(tmpReg, Address (boxReg, 0)); // re-fetch
2250 if (os::is_MP()) { lock(); }
2251 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box
2252
2253 if (EmitSync & 65536) {
2254 bind (CheckSucc);
2255 }
2256 #endif
2257 bind(DONE_LABEL);
2258 }
2259 }
2260 #endif // COMPILER2
2261
2262 void MacroAssembler::c2bool(Register x) {
2263 // implements x == 0 ? 0 : 1
2264 // note: must only look at least-significant byte of x
2265 // since C-style booleans are stored in one byte
2266 // only! (was bug)
2267 andl(x, 0xFF);
2268 setb(Assembler::notZero, x);
2269 }
2270
2271 // Wouldn't need if AddressLiteral version had new name
2272 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) {
2273 Assembler::call(L, rtype);
2274 }
2275
2276 void MacroAssembler::call(Register entry) {
2277 Assembler::call(entry);
2278 }
2279
2280 void MacroAssembler::call(AddressLiteral entry) {
2281 if (reachable(entry)) {
2282 Assembler::call_literal(entry.target(), entry.rspec());
2283 } else {
2284 lea(rscratch1, entry);
2285 Assembler::call(rscratch1);
2286 }
2287 }
2288
2289 void MacroAssembler::ic_call(address entry, jint method_index) {
2290 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index);
2291 movptr(rax, (intptr_t)Universe::non_oop_word());
2292 call(AddressLiteral(entry, rh));
2293 }
2294
2295 // Implementation of call_VM versions
2296
2297 void MacroAssembler::call_VM(Register oop_result,
2298 address entry_point,
2299 bool check_exceptions) {
2300 Label C, E;
2301 call(C, relocInfo::none);
2302 jmp(E);
2303
2304 bind(C);
2305 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
2306 ret(0);
2307
2308 bind(E);
2309 }
2310
2311 void MacroAssembler::call_VM(Register oop_result,
2312 address entry_point,
2313 Register arg_1,
2314 bool check_exceptions) {
2315 Label C, E;
2316 call(C, relocInfo::none);
2317 jmp(E);
2318
2319 bind(C);
2320 pass_arg1(this, arg_1);
2321 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
2322 ret(0);
2323
2324 bind(E);
2325 }
2326
2327 void MacroAssembler::call_VM(Register oop_result,
2328 address entry_point,
2329 Register arg_1,
2330 Register arg_2,
2331 bool check_exceptions) {
2332 Label C, E;
2333 call(C, relocInfo::none);
2334 jmp(E);
2335
2336 bind(C);
2337
2338 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2339
2340 pass_arg2(this, arg_2);
2341 pass_arg1(this, arg_1);
2342 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
2343 ret(0);
2344
2345 bind(E);
2346 }
2347
2348 void MacroAssembler::call_VM(Register oop_result,
2349 address entry_point,
2350 Register arg_1,
2351 Register arg_2,
2352 Register arg_3,
2353 bool check_exceptions) {
2354 Label C, E;
2355 call(C, relocInfo::none);
2356 jmp(E);
2357
2358 bind(C);
2359
2360 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2361 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2362 pass_arg3(this, arg_3);
2363
2364 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2365 pass_arg2(this, arg_2);
2366
2367 pass_arg1(this, arg_1);
2368 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
2369 ret(0);
2370
2371 bind(E);
2372 }
2373
2374 void MacroAssembler::call_VM(Register oop_result,
2375 Register last_java_sp,
2376 address entry_point,
2377 int number_of_arguments,
2378 bool check_exceptions) {
2379 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2380 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2381 }
2382
2383 void MacroAssembler::call_VM(Register oop_result,
2384 Register last_java_sp,
2385 address entry_point,
2386 Register arg_1,
2387 bool check_exceptions) {
2388 pass_arg1(this, arg_1);
2389 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2390 }
2391
2392 void MacroAssembler::call_VM(Register oop_result,
2393 Register last_java_sp,
2394 address entry_point,
2395 Register arg_1,
2396 Register arg_2,
2397 bool check_exceptions) {
2398
2399 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2400 pass_arg2(this, arg_2);
2401 pass_arg1(this, arg_1);
2402 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2403 }
2404
2405 void MacroAssembler::call_VM(Register oop_result,
2406 Register last_java_sp,
2407 address entry_point,
2408 Register arg_1,
2409 Register arg_2,
2410 Register arg_3,
2411 bool check_exceptions) {
2412 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2413 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2414 pass_arg3(this, arg_3);
2415 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2416 pass_arg2(this, arg_2);
2417 pass_arg1(this, arg_1);
2418 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2419 }
2420
2421 void MacroAssembler::super_call_VM(Register oop_result,
2422 Register last_java_sp,
2423 address entry_point,
2424 int number_of_arguments,
2425 bool check_exceptions) {
2426 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2427 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2428 }
2429
2430 void MacroAssembler::super_call_VM(Register oop_result,
2431 Register last_java_sp,
2432 address entry_point,
2433 Register arg_1,
2434 bool check_exceptions) {
2435 pass_arg1(this, arg_1);
2436 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2437 }
2438
2439 void MacroAssembler::super_call_VM(Register oop_result,
2440 Register last_java_sp,
2441 address entry_point,
2442 Register arg_1,
2443 Register arg_2,
2444 bool check_exceptions) {
2445
2446 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2447 pass_arg2(this, arg_2);
2448 pass_arg1(this, arg_1);
2449 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2450 }
2451
2452 void MacroAssembler::super_call_VM(Register oop_result,
2453 Register last_java_sp,
2454 address entry_point,
2455 Register arg_1,
2456 Register arg_2,
2457 Register arg_3,
2458 bool check_exceptions) {
2459 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2460 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2461 pass_arg3(this, arg_3);
2462 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2463 pass_arg2(this, arg_2);
2464 pass_arg1(this, arg_1);
2465 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2466 }
2467
2468 void MacroAssembler::call_VM_base(Register oop_result,
2469 Register java_thread,
2470 Register last_java_sp,
2471 address entry_point,
2472 int number_of_arguments,
2473 bool check_exceptions) {
2474 // determine java_thread register
2475 if (!java_thread->is_valid()) {
2476 #ifdef _LP64
2477 java_thread = r15_thread;
2478 #else
2479 java_thread = rdi;
2480 get_thread(java_thread);
2481 #endif // LP64
2482 }
2483 // determine last_java_sp register
2484 if (!last_java_sp->is_valid()) {
2485 last_java_sp = rsp;
2486 }
2487 // debugging support
2488 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
2489 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register"));
2490 #ifdef ASSERT
2491 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
2492 // r12 is the heapbase.
2493 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");)
2494 #endif // ASSERT
2495
2496 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
2497 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
2498
2499 // push java thread (becomes first argument of C function)
2500
2501 NOT_LP64(push(java_thread); number_of_arguments++);
2502 LP64_ONLY(mov(c_rarg0, r15_thread));
2503
2504 // set last Java frame before call
2505 assert(last_java_sp != rbp, "can't use ebp/rbp");
2506
2507 // Only interpreter should have to set fp
2508 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL);
2509
2510 // do the call, remove parameters
2511 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
2512
2513 // restore the thread (cannot use the pushed argument since arguments
2514 // may be overwritten by C code generated by an optimizing compiler);
2515 // however can use the register value directly if it is callee saved.
2516 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) {
2517 // rdi & rsi (also r15) are callee saved -> nothing to do
2518 #ifdef ASSERT
2519 guarantee(java_thread != rax, "change this code");
2520 push(rax);
2521 { Label L;
2522 get_thread(rax);
2523 cmpptr(java_thread, rax);
2524 jcc(Assembler::equal, L);
2525 STOP("MacroAssembler::call_VM_base: rdi not callee saved?");
2526 bind(L);
2527 }
2528 pop(rax);
2529 #endif
2530 } else {
2531 get_thread(java_thread);
2532 }
2533 // reset last Java frame
2534 // Only interpreter should have to clear fp
2535 reset_last_Java_frame(java_thread, true);
2536
2537 // C++ interp handles this in the interpreter
2538 check_and_handle_popframe(java_thread);
2539 check_and_handle_earlyret(java_thread);
2540
2541 if (check_exceptions) {
2542 // check for pending exceptions (java_thread is set upon return)
2543 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD);
2544 #ifndef _LP64
2545 jump_cc(Assembler::notEqual,
2546 RuntimeAddress(StubRoutines::forward_exception_entry()));
2547 #else
2548 // This used to conditionally jump to forward_exception however it is
2549 // possible if we relocate that the branch will not reach. So we must jump
2550 // around so we can always reach
2551
2552 Label ok;
2553 jcc(Assembler::equal, ok);
2554 jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2555 bind(ok);
2556 #endif // LP64
2557 }
2558
2559 // get oop result if there is one and reset the value in the thread
2560 if (oop_result->is_valid()) {
2561 get_vm_result(oop_result, java_thread);
2562 }
2563 }
2564
2565 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
2566
2567 // Calculate the value for last_Java_sp
2568 // somewhat subtle. call_VM does an intermediate call
2569 // which places a return address on the stack just under the
2570 // stack pointer as the user finsihed with it. This allows
2571 // use to retrieve last_Java_pc from last_Java_sp[-1].
2572 // On 32bit we then have to push additional args on the stack to accomplish
2573 // the actual requested call. On 64bit call_VM only can use register args
2574 // so the only extra space is the return address that call_VM created.
2575 // This hopefully explains the calculations here.
2576
2577 #ifdef _LP64
2578 // We've pushed one address, correct last_Java_sp
2579 lea(rax, Address(rsp, wordSize));
2580 #else
2581 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize));
2582 #endif // LP64
2583
2584 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions);
2585
2586 }
2587
2588 // Use this method when MacroAssembler version of call_VM_leaf_base() should be called from Interpreter.
2589 void MacroAssembler::call_VM_leaf0(address entry_point) {
2590 MacroAssembler::call_VM_leaf_base(entry_point, 0);
2591 }
2592
2593 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
2594 call_VM_leaf_base(entry_point, number_of_arguments);
2595 }
2596
2597 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
2598 pass_arg0(this, arg_0);
2599 call_VM_leaf(entry_point, 1);
2600 }
2601
2602 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2603
2604 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2605 pass_arg1(this, arg_1);
2606 pass_arg0(this, arg_0);
2607 call_VM_leaf(entry_point, 2);
2608 }
2609
2610 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2611 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2612 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2613 pass_arg2(this, arg_2);
2614 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2615 pass_arg1(this, arg_1);
2616 pass_arg0(this, arg_0);
2617 call_VM_leaf(entry_point, 3);
2618 }
2619
2620 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2621 pass_arg0(this, arg_0);
2622 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2623 }
2624
2625 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2626
2627 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2628 pass_arg1(this, arg_1);
2629 pass_arg0(this, arg_0);
2630 MacroAssembler::call_VM_leaf_base(entry_point, 2);
2631 }
2632
2633 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2634 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2635 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2636 pass_arg2(this, arg_2);
2637 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2638 pass_arg1(this, arg_1);
2639 pass_arg0(this, arg_0);
2640 MacroAssembler::call_VM_leaf_base(entry_point, 3);
2641 }
2642
2643 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2644 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg"));
2645 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2646 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2647 pass_arg3(this, arg_3);
2648 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2649 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2650 pass_arg2(this, arg_2);
2651 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2652 pass_arg1(this, arg_1);
2653 pass_arg0(this, arg_0);
2654 MacroAssembler::call_VM_leaf_base(entry_point, 4);
2655 }
2656
2657 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
2658 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
2659 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD);
2660 verify_oop(oop_result, "broken oop in call_VM_base");
2661 }
2662
2663 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
2664 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
2665 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD);
2666 }
2667
2668 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2669 }
2670
2671 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2672 }
2673
2674 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) {
2675 if (reachable(src1)) {
2676 cmpl(as_Address(src1), imm);
2677 } else {
2678 lea(rscratch1, src1);
2679 cmpl(Address(rscratch1, 0), imm);
2680 }
2681 }
2682
2683 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) {
2684 assert(!src2.is_lval(), "use cmpptr");
2685 if (reachable(src2)) {
2686 cmpl(src1, as_Address(src2));
2687 } else {
2688 lea(rscratch1, src2);
2689 cmpl(src1, Address(rscratch1, 0));
2690 }
2691 }
2692
2693 void MacroAssembler::cmp32(Register src1, int32_t imm) {
2694 Assembler::cmpl(src1, imm);
2695 }
2696
2697 void MacroAssembler::cmp32(Register src1, Address src2) {
2698 Assembler::cmpl(src1, src2);
2699 }
2700
2701 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2702 ucomisd(opr1, opr2);
2703
2704 Label L;
2705 if (unordered_is_less) {
2706 movl(dst, -1);
2707 jcc(Assembler::parity, L);
2708 jcc(Assembler::below , L);
2709 movl(dst, 0);
2710 jcc(Assembler::equal , L);
2711 increment(dst);
2712 } else { // unordered is greater
2713 movl(dst, 1);
2714 jcc(Assembler::parity, L);
2715 jcc(Assembler::above , L);
2716 movl(dst, 0);
2717 jcc(Assembler::equal , L);
2718 decrementl(dst);
2719 }
2720 bind(L);
2721 }
2722
2723 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2724 ucomiss(opr1, opr2);
2725
2726 Label L;
2727 if (unordered_is_less) {
2728 movl(dst, -1);
2729 jcc(Assembler::parity, L);
2730 jcc(Assembler::below , L);
2731 movl(dst, 0);
2732 jcc(Assembler::equal , L);
2733 increment(dst);
2734 } else { // unordered is greater
2735 movl(dst, 1);
2736 jcc(Assembler::parity, L);
2737 jcc(Assembler::above , L);
2738 movl(dst, 0);
2739 jcc(Assembler::equal , L);
2740 decrementl(dst);
2741 }
2742 bind(L);
2743 }
2744
2745
2746 void MacroAssembler::cmp8(AddressLiteral src1, int imm) {
2747 if (reachable(src1)) {
2748 cmpb(as_Address(src1), imm);
2749 } else {
2750 lea(rscratch1, src1);
2751 cmpb(Address(rscratch1, 0), imm);
2752 }
2753 }
2754
2755 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) {
2756 #ifdef _LP64
2757 if (src2.is_lval()) {
2758 movptr(rscratch1, src2);
2759 Assembler::cmpq(src1, rscratch1);
2760 } else if (reachable(src2)) {
2761 cmpq(src1, as_Address(src2));
2762 } else {
2763 lea(rscratch1, src2);
2764 Assembler::cmpq(src1, Address(rscratch1, 0));
2765 }
2766 #else
2767 if (src2.is_lval()) {
2768 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2769 } else {
2770 cmpl(src1, as_Address(src2));
2771 }
2772 #endif // _LP64
2773 }
2774
2775 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) {
2776 assert(src2.is_lval(), "not a mem-mem compare");
2777 #ifdef _LP64
2778 // moves src2's literal address
2779 movptr(rscratch1, src2);
2780 Assembler::cmpq(src1, rscratch1);
2781 #else
2782 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2783 #endif // _LP64
2784 }
2785
2786 void MacroAssembler::cmpoop(Register src1, Register src2) {
2787 cmpptr(src1, src2);
2788 }
2789
2790 void MacroAssembler::cmpoop(Register src1, Address src2) {
2791 cmpptr(src1, src2);
2792 }
2793
2794 #ifdef _LP64
2795 void MacroAssembler::cmpoop(Register src1, jobject src2) {
2796 movoop(rscratch1, src2);
2797 cmpptr(src1, rscratch1);
2798 }
2799 #endif
2800
2801 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) {
2802 if (reachable(adr)) {
2803 if (os::is_MP())
2804 lock();
2805 cmpxchgptr(reg, as_Address(adr));
2806 } else {
2807 lea(rscratch1, adr);
2808 if (os::is_MP())
2809 lock();
2810 cmpxchgptr(reg, Address(rscratch1, 0));
2811 }
2812 }
2813
2814 void MacroAssembler::cmpxchgptr(Register reg, Address adr) {
2815 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr));
2816 }
2817
2818 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) {
2819 if (reachable(src)) {
2820 Assembler::comisd(dst, as_Address(src));
2821 } else {
2822 lea(rscratch1, src);
2823 Assembler::comisd(dst, Address(rscratch1, 0));
2824 }
2825 }
2826
2827 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) {
2828 if (reachable(src)) {
2829 Assembler::comiss(dst, as_Address(src));
2830 } else {
2831 lea(rscratch1, src);
2832 Assembler::comiss(dst, Address(rscratch1, 0));
2833 }
2834 }
2835
2836
2837 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) {
2838 Condition negated_cond = negate_condition(cond);
2839 Label L;
2840 jcc(negated_cond, L);
2841 pushf(); // Preserve flags
2842 atomic_incl(counter_addr);
2843 popf();
2844 bind(L);
2845 }
2846
2847 int MacroAssembler::corrected_idivl(Register reg) {
2848 // Full implementation of Java idiv and irem; checks for
2849 // special case as described in JVM spec., p.243 & p.271.
2850 // The function returns the (pc) offset of the idivl
2851 // instruction - may be needed for implicit exceptions.
2852 //
2853 // normal case special case
2854 //
2855 // input : rax,: dividend min_int
2856 // reg: divisor (may not be rax,/rdx) -1
2857 //
2858 // output: rax,: quotient (= rax, idiv reg) min_int
2859 // rdx: remainder (= rax, irem reg) 0
2860 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register");
2861 const int min_int = 0x80000000;
2862 Label normal_case, special_case;
2863
2864 // check for special case
2865 cmpl(rax, min_int);
2866 jcc(Assembler::notEqual, normal_case);
2867 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0)
2868 cmpl(reg, -1);
2869 jcc(Assembler::equal, special_case);
2870
2871 // handle normal case
2872 bind(normal_case);
2873 cdql();
2874 int idivl_offset = offset();
2875 idivl(reg);
2876
2877 // normal and special case exit
2878 bind(special_case);
2879
2880 return idivl_offset;
2881 }
2882
2883
2884
2885 void MacroAssembler::decrementl(Register reg, int value) {
2886 if (value == min_jint) {subl(reg, value) ; return; }
2887 if (value < 0) { incrementl(reg, -value); return; }
2888 if (value == 0) { ; return; }
2889 if (value == 1 && UseIncDec) { decl(reg) ; return; }
2890 /* else */ { subl(reg, value) ; return; }
2891 }
2892
2893 void MacroAssembler::decrementl(Address dst, int value) {
2894 if (value == min_jint) {subl(dst, value) ; return; }
2895 if (value < 0) { incrementl(dst, -value); return; }
2896 if (value == 0) { ; return; }
2897 if (value == 1 && UseIncDec) { decl(dst) ; return; }
2898 /* else */ { subl(dst, value) ; return; }
2899 }
2900
2901 void MacroAssembler::division_with_shift (Register reg, int shift_value) {
2902 assert (shift_value > 0, "illegal shift value");
2903 Label _is_positive;
2904 testl (reg, reg);
2905 jcc (Assembler::positive, _is_positive);
2906 int offset = (1 << shift_value) - 1 ;
2907
2908 if (offset == 1) {
2909 incrementl(reg);
2910 } else {
2911 addl(reg, offset);
2912 }
2913
2914 bind (_is_positive);
2915 sarl(reg, shift_value);
2916 }
2917
2918 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) {
2919 if (reachable(src)) {
2920 Assembler::divsd(dst, as_Address(src));
2921 } else {
2922 lea(rscratch1, src);
2923 Assembler::divsd(dst, Address(rscratch1, 0));
2924 }
2925 }
2926
2927 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) {
2928 if (reachable(src)) {
2929 Assembler::divss(dst, as_Address(src));
2930 } else {
2931 lea(rscratch1, src);
2932 Assembler::divss(dst, Address(rscratch1, 0));
2933 }
2934 }
2935
2936 // !defined(COMPILER2) is because of stupid core builds
2937 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI
2938 void MacroAssembler::empty_FPU_stack() {
2939 if (VM_Version::supports_mmx()) {
2940 emms();
2941 } else {
2942 for (int i = 8; i-- > 0; ) ffree(i);
2943 }
2944 }
2945 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI
2946
2947
2948 // Defines obj, preserves var_size_in_bytes
2949 void MacroAssembler::eden_allocate(Register obj,
2950 Register var_size_in_bytes,
2951 int con_size_in_bytes,
2952 Register t1,
2953 Label& slow_case) {
2954 assert(obj == rax, "obj must be in rax, for cmpxchg");
2955 assert_different_registers(obj, var_size_in_bytes, t1);
2956 if (!Universe::heap()->supports_inline_contig_alloc()) {
2957 jmp(slow_case);
2958 } else {
2959 Register end = t1;
2960 Label retry;
2961 bind(retry);
2962 ExternalAddress heap_top((address) Universe::heap()->top_addr());
2963 movptr(obj, heap_top);
2964 if (var_size_in_bytes == noreg) {
2965 lea(end, Address(obj, con_size_in_bytes));
2966 } else {
2967 lea(end, Address(obj, var_size_in_bytes, Address::times_1));
2968 }
2969 // if end < obj then we wrapped around => object too long => slow case
2970 cmpptr(end, obj);
2971 jcc(Assembler::below, slow_case);
2972 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr()));
2973 jcc(Assembler::above, slow_case);
2974 // Compare obj with the top addr, and if still equal, store the new top addr in
2975 // end at the address of the top addr pointer. Sets ZF if was equal, and clears
2976 // it otherwise. Use lock prefix for atomicity on MPs.
2977 locked_cmpxchgptr(end, heap_top);
2978 jcc(Assembler::notEqual, retry);
2979 }
2980 }
2981
2982 void MacroAssembler::enter() {
2983 push(rbp);
2984 mov(rbp, rsp);
2985 }
2986
2987 // A 5 byte nop that is safe for patching (see patch_verified_entry)
2988 void MacroAssembler::fat_nop() {
2989 if (UseAddressNop) {
2990 addr_nop_5();
2991 } else {
2992 emit_int8(0x26); // es:
2993 emit_int8(0x2e); // cs:
2994 emit_int8(0x64); // fs:
2995 emit_int8(0x65); // gs:
2996 emit_int8((unsigned char)0x90);
2997 }
2998 }
2999
3000 void MacroAssembler::fcmp(Register tmp) {
3001 fcmp(tmp, 1, true, true);
3002 }
3003
3004 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) {
3005 assert(!pop_right || pop_left, "usage error");
3006 if (VM_Version::supports_cmov()) {
3007 assert(tmp == noreg, "unneeded temp");
3008 if (pop_left) {
3009 fucomip(index);
3010 } else {
3011 fucomi(index);
3012 }
3013 if (pop_right) {
3014 fpop();
3015 }
3016 } else {
3017 assert(tmp != noreg, "need temp");
3018 if (pop_left) {
3019 if (pop_right) {
3020 fcompp();
3021 } else {
3022 fcomp(index);
3023 }
3024 } else {
3025 fcom(index);
3026 }
3027 // convert FPU condition into eflags condition via rax,
3028 save_rax(tmp);
3029 fwait(); fnstsw_ax();
3030 sahf();
3031 restore_rax(tmp);
3032 }
3033 // condition codes set as follows:
3034 //
3035 // CF (corresponds to C0) if x < y
3036 // PF (corresponds to C2) if unordered
3037 // ZF (corresponds to C3) if x = y
3038 }
3039
3040 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) {
3041 fcmp2int(dst, unordered_is_less, 1, true, true);
3042 }
3043
3044 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) {
3045 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right);
3046 Label L;
3047 if (unordered_is_less) {
3048 movl(dst, -1);
3049 jcc(Assembler::parity, L);
3050 jcc(Assembler::below , L);
3051 movl(dst, 0);
3052 jcc(Assembler::equal , L);
3053 increment(dst);
3054 } else { // unordered is greater
3055 movl(dst, 1);
3056 jcc(Assembler::parity, L);
3057 jcc(Assembler::above , L);
3058 movl(dst, 0);
3059 jcc(Assembler::equal , L);
3060 decrementl(dst);
3061 }
3062 bind(L);
3063 }
3064
3065 void MacroAssembler::fld_d(AddressLiteral src) {
3066 fld_d(as_Address(src));
3067 }
3068
3069 void MacroAssembler::fld_s(AddressLiteral src) {
3070 fld_s(as_Address(src));
3071 }
3072
3073 void MacroAssembler::fld_x(AddressLiteral src) {
3074 Assembler::fld_x(as_Address(src));
3075 }
3076
3077 void MacroAssembler::fldcw(AddressLiteral src) {
3078 Assembler::fldcw(as_Address(src));
3079 }
3080
3081 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) {
3082 if (reachable(src)) {
3083 Assembler::mulpd(dst, as_Address(src));
3084 } else {
3085 lea(rscratch1, src);
3086 Assembler::mulpd(dst, Address(rscratch1, 0));
3087 }
3088 }
3089
3090 void MacroAssembler::increase_precision() {
3091 subptr(rsp, BytesPerWord);
3092 fnstcw(Address(rsp, 0));
3093 movl(rax, Address(rsp, 0));
3094 orl(rax, 0x300);
3095 push(rax);
3096 fldcw(Address(rsp, 0));
3097 pop(rax);
3098 }
3099
3100 void MacroAssembler::restore_precision() {
3101 fldcw(Address(rsp, 0));
3102 addptr(rsp, BytesPerWord);
3103 }
3104
3105 void MacroAssembler::fpop() {
3106 ffree();
3107 fincstp();
3108 }
3109
3110 void MacroAssembler::load_float(Address src) {
3111 if (UseSSE >= 1) {
3112 movflt(xmm0, src);
3113 } else {
3114 LP64_ONLY(ShouldNotReachHere());
3115 NOT_LP64(fld_s(src));
3116 }
3117 }
3118
3119 void MacroAssembler::store_float(Address dst) {
3120 if (UseSSE >= 1) {
3121 movflt(dst, xmm0);
3122 } else {
3123 LP64_ONLY(ShouldNotReachHere());
3124 NOT_LP64(fstp_s(dst));
3125 }
3126 }
3127
3128 void MacroAssembler::load_double(Address src) {
3129 if (UseSSE >= 2) {
3130 movdbl(xmm0, src);
3131 } else {
3132 LP64_ONLY(ShouldNotReachHere());
3133 NOT_LP64(fld_d(src));
3134 }
3135 }
3136
3137 void MacroAssembler::store_double(Address dst) {
3138 if (UseSSE >= 2) {
3139 movdbl(dst, xmm0);
3140 } else {
3141 LP64_ONLY(ShouldNotReachHere());
3142 NOT_LP64(fstp_d(dst));
3143 }
3144 }
3145
3146 void MacroAssembler::fremr(Register tmp) {
3147 save_rax(tmp);
3148 { Label L;
3149 bind(L);
3150 fprem();
3151 fwait(); fnstsw_ax();
3152 #ifdef _LP64
3153 testl(rax, 0x400);
3154 jcc(Assembler::notEqual, L);
3155 #else
3156 sahf();
3157 jcc(Assembler::parity, L);
3158 #endif // _LP64
3159 }
3160 restore_rax(tmp);
3161 // Result is in ST0.
3162 // Note: fxch & fpop to get rid of ST1
3163 // (otherwise FPU stack could overflow eventually)
3164 fxch(1);
3165 fpop();
3166 }
3167
3168 // dst = c = a * b + c
3169 void MacroAssembler::fmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) {
3170 Assembler::vfmadd231sd(c, a, b);
3171 if (dst != c) {
3172 movdbl(dst, c);
3173 }
3174 }
3175
3176 // dst = c = a * b + c
3177 void MacroAssembler::fmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) {
3178 Assembler::vfmadd231ss(c, a, b);
3179 if (dst != c) {
3180 movflt(dst, c);
3181 }
3182 }
3183
3184 // dst = c = a * b + c
3185 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) {
3186 Assembler::vfmadd231pd(c, a, b, vector_len);
3187 if (dst != c) {
3188 vmovdqu(dst, c);
3189 }
3190 }
3191
3192 // dst = c = a * b + c
3193 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) {
3194 Assembler::vfmadd231ps(c, a, b, vector_len);
3195 if (dst != c) {
3196 vmovdqu(dst, c);
3197 }
3198 }
3199
3200 // dst = c = a * b + c
3201 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) {
3202 Assembler::vfmadd231pd(c, a, b, vector_len);
3203 if (dst != c) {
3204 vmovdqu(dst, c);
3205 }
3206 }
3207
3208 // dst = c = a * b + c
3209 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) {
3210 Assembler::vfmadd231ps(c, a, b, vector_len);
3211 if (dst != c) {
3212 vmovdqu(dst, c);
3213 }
3214 }
3215
3216 void MacroAssembler::incrementl(AddressLiteral dst) {
3217 if (reachable(dst)) {
3218 incrementl(as_Address(dst));
3219 } else {
3220 lea(rscratch1, dst);
3221 incrementl(Address(rscratch1, 0));
3222 }
3223 }
3224
3225 void MacroAssembler::incrementl(ArrayAddress dst) {
3226 incrementl(as_Address(dst));
3227 }
3228
3229 void MacroAssembler::incrementl(Register reg, int value) {
3230 if (value == min_jint) {addl(reg, value) ; return; }
3231 if (value < 0) { decrementl(reg, -value); return; }
3232 if (value == 0) { ; return; }
3233 if (value == 1 && UseIncDec) { incl(reg) ; return; }
3234 /* else */ { addl(reg, value) ; return; }
3235 }
3236
3237 void MacroAssembler::incrementl(Address dst, int value) {
3238 if (value == min_jint) {addl(dst, value) ; return; }
3239 if (value < 0) { decrementl(dst, -value); return; }
3240 if (value == 0) { ; return; }
3241 if (value == 1 && UseIncDec) { incl(dst) ; return; }
3242 /* else */ { addl(dst, value) ; return; }
3243 }
3244
3245 void MacroAssembler::jump(AddressLiteral dst) {
3246 if (reachable(dst)) {
3247 jmp_literal(dst.target(), dst.rspec());
3248 } else {
3249 lea(rscratch1, dst);
3250 jmp(rscratch1);
3251 }
3252 }
3253
3254 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) {
3255 if (reachable(dst)) {
3256 InstructionMark im(this);
3257 relocate(dst.reloc());
3258 const int short_size = 2;
3259 const int long_size = 6;
3260 int offs = (intptr_t)dst.target() - ((intptr_t)pc());
3261 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
3262 // 0111 tttn #8-bit disp
3263 emit_int8(0x70 | cc);
3264 emit_int8((offs - short_size) & 0xFF);
3265 } else {
3266 // 0000 1111 1000 tttn #32-bit disp
3267 emit_int8(0x0F);
3268 emit_int8((unsigned char)(0x80 | cc));
3269 emit_int32(offs - long_size);
3270 }
3271 } else {
3272 #ifdef ASSERT
3273 warning("reversing conditional branch");
3274 #endif /* ASSERT */
3275 Label skip;
3276 jccb(reverse[cc], skip);
3277 lea(rscratch1, dst);
3278 Assembler::jmp(rscratch1);
3279 bind(skip);
3280 }
3281 }
3282
3283 void MacroAssembler::ldmxcsr(AddressLiteral src) {
3284 if (reachable(src)) {
3285 Assembler::ldmxcsr(as_Address(src));
3286 } else {
3287 lea(rscratch1, src);
3288 Assembler::ldmxcsr(Address(rscratch1, 0));
3289 }
3290 }
3291
3292 int MacroAssembler::load_signed_byte(Register dst, Address src) {
3293 int off;
3294 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3295 off = offset();
3296 movsbl(dst, src); // movsxb
3297 } else {
3298 off = load_unsigned_byte(dst, src);
3299 shll(dst, 24);
3300 sarl(dst, 24);
3301 }
3302 return off;
3303 }
3304
3305 // Note: load_signed_short used to be called load_signed_word.
3306 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler
3307 // manual, which means 16 bits, that usage is found nowhere in HotSpot code.
3308 // The term "word" in HotSpot means a 32- or 64-bit machine word.
3309 int MacroAssembler::load_signed_short(Register dst, Address src) {
3310 int off;
3311 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3312 // This is dubious to me since it seems safe to do a signed 16 => 64 bit
3313 // version but this is what 64bit has always done. This seems to imply
3314 // that users are only using 32bits worth.
3315 off = offset();
3316 movswl(dst, src); // movsxw
3317 } else {
3318 off = load_unsigned_short(dst, src);
3319 shll(dst, 16);
3320 sarl(dst, 16);
3321 }
3322 return off;
3323 }
3324
3325 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
3326 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3327 // and "3.9 Partial Register Penalties", p. 22).
3328 int off;
3329 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) {
3330 off = offset();
3331 movzbl(dst, src); // movzxb
3332 } else {
3333 xorl(dst, dst);
3334 off = offset();
3335 movb(dst, src);
3336 }
3337 return off;
3338 }
3339
3340 // Note: load_unsigned_short used to be called load_unsigned_word.
3341 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
3342 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3343 // and "3.9 Partial Register Penalties", p. 22).
3344 int off;
3345 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) {
3346 off = offset();
3347 movzwl(dst, src); // movzxw
3348 } else {
3349 xorl(dst, dst);
3350 off = offset();
3351 movw(dst, src);
3352 }
3353 return off;
3354 }
3355
3356 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
3357 switch (size_in_bytes) {
3358 #ifndef _LP64
3359 case 8:
3360 assert(dst2 != noreg, "second dest register required");
3361 movl(dst, src);
3362 movl(dst2, src.plus_disp(BytesPerInt));
3363 break;
3364 #else
3365 case 8: movq(dst, src); break;
3366 #endif
3367 case 4: movl(dst, src); break;
3368 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
3369 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
3370 default: ShouldNotReachHere();
3371 }
3372 }
3373
3374 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
3375 switch (size_in_bytes) {
3376 #ifndef _LP64
3377 case 8:
3378 assert(src2 != noreg, "second source register required");
3379 movl(dst, src);
3380 movl(dst.plus_disp(BytesPerInt), src2);
3381 break;
3382 #else
3383 case 8: movq(dst, src); break;
3384 #endif
3385 case 4: movl(dst, src); break;
3386 case 2: movw(dst, src); break;
3387 case 1: movb(dst, src); break;
3388 default: ShouldNotReachHere();
3389 }
3390 }
3391
3392 void MacroAssembler::mov32(AddressLiteral dst, Register src) {
3393 if (reachable(dst)) {
3394 movl(as_Address(dst), src);
3395 } else {
3396 lea(rscratch1, dst);
3397 movl(Address(rscratch1, 0), src);
3398 }
3399 }
3400
3401 void MacroAssembler::mov32(Register dst, AddressLiteral src) {
3402 if (reachable(src)) {
3403 movl(dst, as_Address(src));
3404 } else {
3405 lea(rscratch1, src);
3406 movl(dst, Address(rscratch1, 0));
3407 }
3408 }
3409
3410 // C++ bool manipulation
3411
3412 void MacroAssembler::movbool(Register dst, Address src) {
3413 if(sizeof(bool) == 1)
3414 movb(dst, src);
3415 else if(sizeof(bool) == 2)
3416 movw(dst, src);
3417 else if(sizeof(bool) == 4)
3418 movl(dst, src);
3419 else
3420 // unsupported
3421 ShouldNotReachHere();
3422 }
3423
3424 void MacroAssembler::movbool(Address dst, bool boolconst) {
3425 if(sizeof(bool) == 1)
3426 movb(dst, (int) boolconst);
3427 else if(sizeof(bool) == 2)
3428 movw(dst, (int) boolconst);
3429 else if(sizeof(bool) == 4)
3430 movl(dst, (int) boolconst);
3431 else
3432 // unsupported
3433 ShouldNotReachHere();
3434 }
3435
3436 void MacroAssembler::movbool(Address dst, Register src) {
3437 if(sizeof(bool) == 1)
3438 movb(dst, src);
3439 else if(sizeof(bool) == 2)
3440 movw(dst, src);
3441 else if(sizeof(bool) == 4)
3442 movl(dst, src);
3443 else
3444 // unsupported
3445 ShouldNotReachHere();
3446 }
3447
3448 void MacroAssembler::movbyte(ArrayAddress dst, int src) {
3449 movb(as_Address(dst), src);
3450 }
3451
3452 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) {
3453 if (reachable(src)) {
3454 movdl(dst, as_Address(src));
3455 } else {
3456 lea(rscratch1, src);
3457 movdl(dst, Address(rscratch1, 0));
3458 }
3459 }
3460
3461 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) {
3462 if (reachable(src)) {
3463 movq(dst, as_Address(src));
3464 } else {
3465 lea(rscratch1, src);
3466 movq(dst, Address(rscratch1, 0));
3467 }
3468 }
3469
3470 void MacroAssembler::setvectmask(Register dst, Register src) {
3471 Assembler::movl(dst, 1);
3472 Assembler::shlxl(dst, dst, src);
3473 Assembler::decl(dst);
3474 Assembler::kmovdl(k1, dst);
3475 Assembler::movl(dst, src);
3476 }
3477
3478 void MacroAssembler::restorevectmask() {
3479 Assembler::knotwl(k1, k0);
3480 }
3481
3482 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) {
3483 if (reachable(src)) {
3484 if (UseXmmLoadAndClearUpper) {
3485 movsd (dst, as_Address(src));
3486 } else {
3487 movlpd(dst, as_Address(src));
3488 }
3489 } else {
3490 lea(rscratch1, src);
3491 if (UseXmmLoadAndClearUpper) {
3492 movsd (dst, Address(rscratch1, 0));
3493 } else {
3494 movlpd(dst, Address(rscratch1, 0));
3495 }
3496 }
3497 }
3498
3499 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) {
3500 if (reachable(src)) {
3501 movss(dst, as_Address(src));
3502 } else {
3503 lea(rscratch1, src);
3504 movss(dst, Address(rscratch1, 0));
3505 }
3506 }
3507
3508 void MacroAssembler::movptr(Register dst, Register src) {
3509 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3510 }
3511
3512 void MacroAssembler::movptr(Register dst, Address src) {
3513 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3514 }
3515
3516 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
3517 void MacroAssembler::movptr(Register dst, intptr_t src) {
3518 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src));
3519 }
3520
3521 void MacroAssembler::movptr(Address dst, Register src) {
3522 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3523 }
3524
3525 void MacroAssembler::movdqu(Address dst, XMMRegister src) {
3526 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) {
3527 Assembler::vextractf32x4(dst, src, 0);
3528 } else {
3529 Assembler::movdqu(dst, src);
3530 }
3531 }
3532
3533 void MacroAssembler::movdqu(XMMRegister dst, Address src) {
3534 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) {
3535 Assembler::vinsertf32x4(dst, dst, src, 0);
3536 } else {
3537 Assembler::movdqu(dst, src);
3538 }
3539 }
3540
3541 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) {
3542 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) {
3543 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit);
3544 } else {
3545 Assembler::movdqu(dst, src);
3546 }
3547 }
3548
3549 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src, Register scratchReg) {
3550 if (reachable(src)) {
3551 movdqu(dst, as_Address(src));
3552 } else {
3553 lea(scratchReg, src);
3554 movdqu(dst, Address(scratchReg, 0));
3555 }
3556 }
3557
3558 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) {
3559 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) {
3560 vextractf64x4_low(dst, src);
3561 } else {
3562 Assembler::vmovdqu(dst, src);
3563 }
3564 }
3565
3566 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) {
3567 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) {
3568 vinsertf64x4_low(dst, src);
3569 } else {
3570 Assembler::vmovdqu(dst, src);
3571 }
3572 }
3573
3574 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) {
3575 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) {
3576 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit);
3577 }
3578 else {
3579 Assembler::vmovdqu(dst, src);
3580 }
3581 }
3582
3583 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) {
3584 if (reachable(src)) {
3585 vmovdqu(dst, as_Address(src));
3586 }
3587 else {
3588 lea(rscratch1, src);
3589 vmovdqu(dst, Address(rscratch1, 0));
3590 }
3591 }
3592
3593 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) {
3594 if (reachable(src)) {
3595 Assembler::movdqa(dst, as_Address(src));
3596 } else {
3597 lea(rscratch1, src);
3598 Assembler::movdqa(dst, Address(rscratch1, 0));
3599 }
3600 }
3601
3602 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) {
3603 if (reachable(src)) {
3604 Assembler::movsd(dst, as_Address(src));
3605 } else {
3606 lea(rscratch1, src);
3607 Assembler::movsd(dst, Address(rscratch1, 0));
3608 }
3609 }
3610
3611 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) {
3612 if (reachable(src)) {
3613 Assembler::movss(dst, as_Address(src));
3614 } else {
3615 lea(rscratch1, src);
3616 Assembler::movss(dst, Address(rscratch1, 0));
3617 }
3618 }
3619
3620 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) {
3621 if (reachable(src)) {
3622 Assembler::mulsd(dst, as_Address(src));
3623 } else {
3624 lea(rscratch1, src);
3625 Assembler::mulsd(dst, Address(rscratch1, 0));
3626 }
3627 }
3628
3629 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) {
3630 if (reachable(src)) {
3631 Assembler::mulss(dst, as_Address(src));
3632 } else {
3633 lea(rscratch1, src);
3634 Assembler::mulss(dst, Address(rscratch1, 0));
3635 }
3636 }
3637
3638 void MacroAssembler::null_check(Register reg, int offset) {
3639 if (needs_explicit_null_check(offset)) {
3640 // provoke OS NULL exception if reg = NULL by
3641 // accessing M[reg] w/o changing any (non-CC) registers
3642 // NOTE: cmpl is plenty here to provoke a segv
3643 cmpptr(rax, Address(reg, 0));
3644 // Note: should probably use testl(rax, Address(reg, 0));
3645 // may be shorter code (however, this version of
3646 // testl needs to be implemented first)
3647 } else {
3648 // nothing to do, (later) access of M[reg + offset]
3649 // will provoke OS NULL exception if reg = NULL
3650 }
3651 }
3652
3653 void MacroAssembler::os_breakpoint() {
3654 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability
3655 // (e.g., MSVC can't call ps() otherwise)
3656 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint)));
3657 }
3658
3659 void MacroAssembler::unimplemented(const char* what) {
3660 const char* buf = NULL;
3661 {
3662 ResourceMark rm;
3663 stringStream ss;
3664 ss.print("unimplemented: %s", what);
3665 buf = code_string(ss.as_string());
3666 }
3667 stop(buf);
3668 }
3669
3670 #ifdef _LP64
3671 #define XSTATE_BV 0x200
3672 #endif
3673
3674 void MacroAssembler::pop_CPU_state() {
3675 pop_FPU_state();
3676 pop_IU_state();
3677 }
3678
3679 void MacroAssembler::pop_FPU_state() {
3680 #ifndef _LP64
3681 frstor(Address(rsp, 0));
3682 #else
3683 fxrstor(Address(rsp, 0));
3684 #endif
3685 addptr(rsp, FPUStateSizeInWords * wordSize);
3686 }
3687
3688 void MacroAssembler::pop_IU_state() {
3689 popa();
3690 LP64_ONLY(addq(rsp, 8));
3691 popf();
3692 }
3693
3694 // Save Integer and Float state
3695 // Warning: Stack must be 16 byte aligned (64bit)
3696 void MacroAssembler::push_CPU_state() {
3697 push_IU_state();
3698 push_FPU_state();
3699 }
3700
3701 void MacroAssembler::push_FPU_state() {
3702 subptr(rsp, FPUStateSizeInWords * wordSize);
3703 #ifndef _LP64
3704 fnsave(Address(rsp, 0));
3705 fwait();
3706 #else
3707 fxsave(Address(rsp, 0));
3708 #endif // LP64
3709 }
3710
3711 void MacroAssembler::push_IU_state() {
3712 // Push flags first because pusha kills them
3713 pushf();
3714 // Make sure rsp stays 16-byte aligned
3715 LP64_ONLY(subq(rsp, 8));
3716 pusha();
3717 }
3718
3719 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp) { // determine java_thread register
3720 if (!java_thread->is_valid()) {
3721 java_thread = rdi;
3722 get_thread(java_thread);
3723 }
3724 // we must set sp to zero to clear frame
3725 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
3726 if (clear_fp) {
3727 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3728 }
3729
3730 // Always clear the pc because it could have been set by make_walkable()
3731 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
3732
3733 vzeroupper();
3734 }
3735
3736 void MacroAssembler::restore_rax(Register tmp) {
3737 if (tmp == noreg) pop(rax);
3738 else if (tmp != rax) mov(rax, tmp);
3739 }
3740
3741 void MacroAssembler::round_to(Register reg, int modulus) {
3742 addptr(reg, modulus - 1);
3743 andptr(reg, -modulus);
3744 }
3745
3746 void MacroAssembler::save_rax(Register tmp) {
3747 if (tmp == noreg) push(rax);
3748 else if (tmp != rax) mov(tmp, rax);
3749 }
3750
3751 // Write serialization page so VM thread can do a pseudo remote membar.
3752 // We use the current thread pointer to calculate a thread specific
3753 // offset to write to within the page. This minimizes bus traffic
3754 // due to cache line collision.
3755 void MacroAssembler::serialize_memory(Register thread, Register tmp) {
3756 movl(tmp, thread);
3757 shrl(tmp, os::get_serialize_page_shift_count());
3758 andl(tmp, (os::vm_page_size() - sizeof(int)));
3759
3760 Address index(noreg, tmp, Address::times_1);
3761 ExternalAddress page(os::get_memory_serialize_page());
3762
3763 // Size of store must match masking code above
3764 movl(as_Address(ArrayAddress(page, index)), tmp);
3765 }
3766
3767 void MacroAssembler::safepoint_poll(Label& slow_path, Register thread_reg, Register temp_reg) {
3768 if (SafepointMechanism::uses_thread_local_poll()) {
3769 #ifdef _LP64
3770 assert(thread_reg == r15_thread, "should be");
3771 #else
3772 if (thread_reg == noreg) {
3773 thread_reg = temp_reg;
3774 get_thread(thread_reg);
3775 }
3776 #endif
3777 testb(Address(thread_reg, Thread::polling_page_offset()), SafepointMechanism::poll_bit());
3778 jcc(Assembler::notZero, slow_path); // handshake bit set implies poll
3779 } else {
3780 cmp32(ExternalAddress(SafepointSynchronize::address_of_state()),
3781 SafepointSynchronize::_not_synchronized);
3782 jcc(Assembler::notEqual, slow_path);
3783 }
3784 }
3785
3786 // Calls to C land
3787 //
3788 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded
3789 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
3790 // has to be reset to 0. This is required to allow proper stack traversal.
3791 void MacroAssembler::set_last_Java_frame(Register java_thread,
3792 Register last_java_sp,
3793 Register last_java_fp,
3794 address last_java_pc) {
3795 vzeroupper();
3796 // determine java_thread register
3797 if (!java_thread->is_valid()) {
3798 java_thread = rdi;
3799 get_thread(java_thread);
3800 }
3801 // determine last_java_sp register
3802 if (!last_java_sp->is_valid()) {
3803 last_java_sp = rsp;
3804 }
3805
3806 // last_java_fp is optional
3807
3808 if (last_java_fp->is_valid()) {
3809 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp);
3810 }
3811
3812 // last_java_pc is optional
3813
3814 if (last_java_pc != NULL) {
3815 lea(Address(java_thread,
3816 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()),
3817 InternalAddress(last_java_pc));
3818
3819 }
3820 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
3821 }
3822
3823 void MacroAssembler::shlptr(Register dst, int imm8) {
3824 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8));
3825 }
3826
3827 void MacroAssembler::shrptr(Register dst, int imm8) {
3828 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8));
3829 }
3830
3831 void MacroAssembler::sign_extend_byte(Register reg) {
3832 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) {
3833 movsbl(reg, reg); // movsxb
3834 } else {
3835 shll(reg, 24);
3836 sarl(reg, 24);
3837 }
3838 }
3839
3840 void MacroAssembler::sign_extend_short(Register reg) {
3841 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3842 movswl(reg, reg); // movsxw
3843 } else {
3844 shll(reg, 16);
3845 sarl(reg, 16);
3846 }
3847 }
3848
3849 void MacroAssembler::testl(Register dst, AddressLiteral src) {
3850 assert(reachable(src), "Address should be reachable");
3851 testl(dst, as_Address(src));
3852 }
3853
3854 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) {
3855 int dst_enc = dst->encoding();
3856 int src_enc = src->encoding();
3857 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3858 Assembler::pcmpeqb(dst, src);
3859 } else if ((dst_enc < 16) && (src_enc < 16)) {
3860 Assembler::pcmpeqb(dst, src);
3861 } else if (src_enc < 16) {
3862 subptr(rsp, 64);
3863 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3864 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3865 Assembler::pcmpeqb(xmm0, src);
3866 movdqu(dst, xmm0);
3867 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3868 addptr(rsp, 64);
3869 } else if (dst_enc < 16) {
3870 subptr(rsp, 64);
3871 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3872 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3873 Assembler::pcmpeqb(dst, xmm0);
3874 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3875 addptr(rsp, 64);
3876 } else {
3877 subptr(rsp, 64);
3878 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3879 subptr(rsp, 64);
3880 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3881 movdqu(xmm0, src);
3882 movdqu(xmm1, dst);
3883 Assembler::pcmpeqb(xmm1, xmm0);
3884 movdqu(dst, xmm1);
3885 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3886 addptr(rsp, 64);
3887 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3888 addptr(rsp, 64);
3889 }
3890 }
3891
3892 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) {
3893 int dst_enc = dst->encoding();
3894 int src_enc = src->encoding();
3895 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3896 Assembler::pcmpeqw(dst, src);
3897 } else if ((dst_enc < 16) && (src_enc < 16)) {
3898 Assembler::pcmpeqw(dst, src);
3899 } else if (src_enc < 16) {
3900 subptr(rsp, 64);
3901 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3902 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3903 Assembler::pcmpeqw(xmm0, src);
3904 movdqu(dst, xmm0);
3905 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3906 addptr(rsp, 64);
3907 } else if (dst_enc < 16) {
3908 subptr(rsp, 64);
3909 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3910 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3911 Assembler::pcmpeqw(dst, xmm0);
3912 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3913 addptr(rsp, 64);
3914 } else {
3915 subptr(rsp, 64);
3916 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3917 subptr(rsp, 64);
3918 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3919 movdqu(xmm0, src);
3920 movdqu(xmm1, dst);
3921 Assembler::pcmpeqw(xmm1, xmm0);
3922 movdqu(dst, xmm1);
3923 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3924 addptr(rsp, 64);
3925 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3926 addptr(rsp, 64);
3927 }
3928 }
3929
3930 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) {
3931 int dst_enc = dst->encoding();
3932 if (dst_enc < 16) {
3933 Assembler::pcmpestri(dst, src, imm8);
3934 } else {
3935 subptr(rsp, 64);
3936 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3937 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3938 Assembler::pcmpestri(xmm0, src, imm8);
3939 movdqu(dst, xmm0);
3940 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3941 addptr(rsp, 64);
3942 }
3943 }
3944
3945 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) {
3946 int dst_enc = dst->encoding();
3947 int src_enc = src->encoding();
3948 if ((dst_enc < 16) && (src_enc < 16)) {
3949 Assembler::pcmpestri(dst, src, imm8);
3950 } else if (src_enc < 16) {
3951 subptr(rsp, 64);
3952 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3953 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3954 Assembler::pcmpestri(xmm0, src, imm8);
3955 movdqu(dst, xmm0);
3956 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3957 addptr(rsp, 64);
3958 } else if (dst_enc < 16) {
3959 subptr(rsp, 64);
3960 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3961 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3962 Assembler::pcmpestri(dst, xmm0, imm8);
3963 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3964 addptr(rsp, 64);
3965 } else {
3966 subptr(rsp, 64);
3967 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3968 subptr(rsp, 64);
3969 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3970 movdqu(xmm0, src);
3971 movdqu(xmm1, dst);
3972 Assembler::pcmpestri(xmm1, xmm0, imm8);
3973 movdqu(dst, xmm1);
3974 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3975 addptr(rsp, 64);
3976 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3977 addptr(rsp, 64);
3978 }
3979 }
3980
3981 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) {
3982 int dst_enc = dst->encoding();
3983 int src_enc = src->encoding();
3984 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3985 Assembler::pmovzxbw(dst, src);
3986 } else if ((dst_enc < 16) && (src_enc < 16)) {
3987 Assembler::pmovzxbw(dst, src);
3988 } else if (src_enc < 16) {
3989 subptr(rsp, 64);
3990 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3991 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3992 Assembler::pmovzxbw(xmm0, src);
3993 movdqu(dst, xmm0);
3994 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3995 addptr(rsp, 64);
3996 } else if (dst_enc < 16) {
3997 subptr(rsp, 64);
3998 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3999 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4000 Assembler::pmovzxbw(dst, xmm0);
4001 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4002 addptr(rsp, 64);
4003 } else {
4004 subptr(rsp, 64);
4005 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4006 subptr(rsp, 64);
4007 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4008 movdqu(xmm0, src);
4009 movdqu(xmm1, dst);
4010 Assembler::pmovzxbw(xmm1, xmm0);
4011 movdqu(dst, xmm1);
4012 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4013 addptr(rsp, 64);
4014 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4015 addptr(rsp, 64);
4016 }
4017 }
4018
4019 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) {
4020 int dst_enc = dst->encoding();
4021 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4022 Assembler::pmovzxbw(dst, src);
4023 } else if (dst_enc < 16) {
4024 Assembler::pmovzxbw(dst, src);
4025 } else {
4026 subptr(rsp, 64);
4027 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4028 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4029 Assembler::pmovzxbw(xmm0, src);
4030 movdqu(dst, xmm0);
4031 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4032 addptr(rsp, 64);
4033 }
4034 }
4035
4036 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) {
4037 int src_enc = src->encoding();
4038 if (src_enc < 16) {
4039 Assembler::pmovmskb(dst, src);
4040 } else {
4041 subptr(rsp, 64);
4042 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4043 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4044 Assembler::pmovmskb(dst, xmm0);
4045 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4046 addptr(rsp, 64);
4047 }
4048 }
4049
4050 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) {
4051 int dst_enc = dst->encoding();
4052 int src_enc = src->encoding();
4053 if ((dst_enc < 16) && (src_enc < 16)) {
4054 Assembler::ptest(dst, src);
4055 } else if (src_enc < 16) {
4056 subptr(rsp, 64);
4057 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4058 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4059 Assembler::ptest(xmm0, src);
4060 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4061 addptr(rsp, 64);
4062 } else if (dst_enc < 16) {
4063 subptr(rsp, 64);
4064 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4065 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4066 Assembler::ptest(dst, xmm0);
4067 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4068 addptr(rsp, 64);
4069 } else {
4070 subptr(rsp, 64);
4071 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4072 subptr(rsp, 64);
4073 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4074 movdqu(xmm0, src);
4075 movdqu(xmm1, dst);
4076 Assembler::ptest(xmm1, xmm0);
4077 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4078 addptr(rsp, 64);
4079 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4080 addptr(rsp, 64);
4081 }
4082 }
4083
4084 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) {
4085 if (reachable(src)) {
4086 Assembler::sqrtsd(dst, as_Address(src));
4087 } else {
4088 lea(rscratch1, src);
4089 Assembler::sqrtsd(dst, Address(rscratch1, 0));
4090 }
4091 }
4092
4093 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) {
4094 if (reachable(src)) {
4095 Assembler::sqrtss(dst, as_Address(src));
4096 } else {
4097 lea(rscratch1, src);
4098 Assembler::sqrtss(dst, Address(rscratch1, 0));
4099 }
4100 }
4101
4102 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) {
4103 if (reachable(src)) {
4104 Assembler::subsd(dst, as_Address(src));
4105 } else {
4106 lea(rscratch1, src);
4107 Assembler::subsd(dst, Address(rscratch1, 0));
4108 }
4109 }
4110
4111 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
4112 if (reachable(src)) {
4113 Assembler::subss(dst, as_Address(src));
4114 } else {
4115 lea(rscratch1, src);
4116 Assembler::subss(dst, Address(rscratch1, 0));
4117 }
4118 }
4119
4120 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
4121 if (reachable(src)) {
4122 Assembler::ucomisd(dst, as_Address(src));
4123 } else {
4124 lea(rscratch1, src);
4125 Assembler::ucomisd(dst, Address(rscratch1, 0));
4126 }
4127 }
4128
4129 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
4130 if (reachable(src)) {
4131 Assembler::ucomiss(dst, as_Address(src));
4132 } else {
4133 lea(rscratch1, src);
4134 Assembler::ucomiss(dst, Address(rscratch1, 0));
4135 }
4136 }
4137
4138 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) {
4139 // Used in sign-bit flipping with aligned address.
4140 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
4141 if (reachable(src)) {
4142 Assembler::xorpd(dst, as_Address(src));
4143 } else {
4144 lea(rscratch1, src);
4145 Assembler::xorpd(dst, Address(rscratch1, 0));
4146 }
4147 }
4148
4149 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) {
4150 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
4151 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
4152 }
4153 else {
4154 Assembler::xorpd(dst, src);
4155 }
4156 }
4157
4158 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) {
4159 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
4160 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
4161 } else {
4162 Assembler::xorps(dst, src);
4163 }
4164 }
4165
4166 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) {
4167 // Used in sign-bit flipping with aligned address.
4168 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
4169 if (reachable(src)) {
4170 Assembler::xorps(dst, as_Address(src));
4171 } else {
4172 lea(rscratch1, src);
4173 Assembler::xorps(dst, Address(rscratch1, 0));
4174 }
4175 }
4176
4177 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
4178 // Used in sign-bit flipping with aligned address.
4179 bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
4180 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
4181 if (reachable(src)) {
4182 Assembler::pshufb(dst, as_Address(src));
4183 } else {
4184 lea(rscratch1, src);
4185 Assembler::pshufb(dst, Address(rscratch1, 0));
4186 }
4187 }
4188
4189 // AVX 3-operands instructions
4190
4191 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4192 if (reachable(src)) {
4193 vaddsd(dst, nds, as_Address(src));
4194 } else {
4195 lea(rscratch1, src);
4196 vaddsd(dst, nds, Address(rscratch1, 0));
4197 }
4198 }
4199
4200 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4201 if (reachable(src)) {
4202 vaddss(dst, nds, as_Address(src));
4203 } else {
4204 lea(rscratch1, src);
4205 vaddss(dst, nds, Address(rscratch1, 0));
4206 }
4207 }
4208
4209 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
4210 int dst_enc = dst->encoding();
4211 int nds_enc = nds->encoding();
4212 int src_enc = src->encoding();
4213 if ((dst_enc < 16) && (nds_enc < 16)) {
4214 vandps(dst, nds, negate_field, vector_len);
4215 } else if ((src_enc < 16) && (dst_enc < 16)) {
4216 evmovdqul(src, nds, Assembler::AVX_512bit);
4217 vandps(dst, src, negate_field, vector_len);
4218 } else if (src_enc < 16) {
4219 evmovdqul(src, nds, Assembler::AVX_512bit);
4220 vandps(src, src, negate_field, vector_len);
4221 evmovdqul(dst, src, Assembler::AVX_512bit);
4222 } else if (dst_enc < 16) {
4223 evmovdqul(src, xmm0, Assembler::AVX_512bit);
4224 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4225 vandps(dst, xmm0, negate_field, vector_len);
4226 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4227 } else {
4228 if (src_enc != dst_enc) {
4229 evmovdqul(src, xmm0, Assembler::AVX_512bit);
4230 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4231 vandps(xmm0, xmm0, negate_field, vector_len);
4232 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4233 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4234 } else {
4235 subptr(rsp, 64);
4236 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4237 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4238 vandps(xmm0, xmm0, negate_field, vector_len);
4239 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4240 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4241 addptr(rsp, 64);
4242 }
4243 }
4244 }
4245
4246 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
4247 int dst_enc = dst->encoding();
4248 int nds_enc = nds->encoding();
4249 int src_enc = src->encoding();
4250 if ((dst_enc < 16) && (nds_enc < 16)) {
4251 vandpd(dst, nds, negate_field, vector_len);
4252 } else if ((src_enc < 16) && (dst_enc < 16)) {
4253 evmovdqul(src, nds, Assembler::AVX_512bit);
4254 vandpd(dst, src, negate_field, vector_len);
4255 } else if (src_enc < 16) {
4256 evmovdqul(src, nds, Assembler::AVX_512bit);
4257 vandpd(src, src, negate_field, vector_len);
4258 evmovdqul(dst, src, Assembler::AVX_512bit);
4259 } else if (dst_enc < 16) {
4260 evmovdqul(src, xmm0, Assembler::AVX_512bit);
4261 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4262 vandpd(dst, xmm0, negate_field, vector_len);
4263 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4264 } else {
4265 if (src_enc != dst_enc) {
4266 evmovdqul(src, xmm0, Assembler::AVX_512bit);
4267 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4268 vandpd(xmm0, xmm0, negate_field, vector_len);
4269 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4270 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4271 } else {
4272 subptr(rsp, 64);
4273 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4274 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4275 vandpd(xmm0, xmm0, negate_field, vector_len);
4276 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4277 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4278 addptr(rsp, 64);
4279 }
4280 }
4281 }
4282
4283 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4284 int dst_enc = dst->encoding();
4285 int nds_enc = nds->encoding();
4286 int src_enc = src->encoding();
4287 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4288 Assembler::vpaddb(dst, nds, src, vector_len);
4289 } else if ((dst_enc < 16) && (src_enc < 16)) {
4290 Assembler::vpaddb(dst, dst, src, vector_len);
4291 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4292 // use nds as scratch for src
4293 evmovdqul(nds, src, Assembler::AVX_512bit);
4294 Assembler::vpaddb(dst, dst, nds, vector_len);
4295 } else if ((src_enc < 16) && (nds_enc < 16)) {
4296 // use nds as scratch for dst
4297 evmovdqul(nds, dst, Assembler::AVX_512bit);
4298 Assembler::vpaddb(nds, nds, src, vector_len);
4299 evmovdqul(dst, nds, Assembler::AVX_512bit);
4300 } else if (dst_enc < 16) {
4301 // use nds as scatch for xmm0 to hold src
4302 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4303 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4304 Assembler::vpaddb(dst, dst, xmm0, vector_len);
4305 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4306 } else {
4307 // worse case scenario, all regs are in the upper bank
4308 subptr(rsp, 64);
4309 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4310 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4311 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4312 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4313 Assembler::vpaddb(xmm0, xmm0, xmm1, vector_len);
4314 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4315 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4316 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4317 addptr(rsp, 64);
4318 }
4319 }
4320
4321 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4322 int dst_enc = dst->encoding();
4323 int nds_enc = nds->encoding();
4324 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4325 Assembler::vpaddb(dst, nds, src, vector_len);
4326 } else if (dst_enc < 16) {
4327 Assembler::vpaddb(dst, dst, src, vector_len);
4328 } else if (nds_enc < 16) {
4329 // implies dst_enc in upper bank with src as scratch
4330 evmovdqul(nds, dst, Assembler::AVX_512bit);
4331 Assembler::vpaddb(nds, nds, src, vector_len);
4332 evmovdqul(dst, nds, Assembler::AVX_512bit);
4333 } else {
4334 // worse case scenario, all regs in upper bank
4335 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4336 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4337 Assembler::vpaddb(xmm0, xmm0, src, vector_len);
4338 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4339 }
4340 }
4341
4342 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4343 int dst_enc = dst->encoding();
4344 int nds_enc = nds->encoding();
4345 int src_enc = src->encoding();
4346 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4347 Assembler::vpaddw(dst, nds, src, vector_len);
4348 } else if ((dst_enc < 16) && (src_enc < 16)) {
4349 Assembler::vpaddw(dst, dst, src, vector_len);
4350 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4351 // use nds as scratch for src
4352 evmovdqul(nds, src, Assembler::AVX_512bit);
4353 Assembler::vpaddw(dst, dst, nds, vector_len);
4354 } else if ((src_enc < 16) && (nds_enc < 16)) {
4355 // use nds as scratch for dst
4356 evmovdqul(nds, dst, Assembler::AVX_512bit);
4357 Assembler::vpaddw(nds, nds, src, vector_len);
4358 evmovdqul(dst, nds, Assembler::AVX_512bit);
4359 } else if (dst_enc < 16) {
4360 // use nds as scatch for xmm0 to hold src
4361 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4362 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4363 Assembler::vpaddw(dst, dst, xmm0, vector_len);
4364 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4365 } else {
4366 // worse case scenario, all regs are in the upper bank
4367 subptr(rsp, 64);
4368 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4369 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4370 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4371 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4372 Assembler::vpaddw(xmm0, xmm0, xmm1, vector_len);
4373 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4374 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4375 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4376 addptr(rsp, 64);
4377 }
4378 }
4379
4380 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4381 int dst_enc = dst->encoding();
4382 int nds_enc = nds->encoding();
4383 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4384 Assembler::vpaddw(dst, nds, src, vector_len);
4385 } else if (dst_enc < 16) {
4386 Assembler::vpaddw(dst, dst, src, vector_len);
4387 } else if (nds_enc < 16) {
4388 // implies dst_enc in upper bank with src as scratch
4389 evmovdqul(nds, dst, Assembler::AVX_512bit);
4390 Assembler::vpaddw(nds, nds, src, vector_len);
4391 evmovdqul(dst, nds, Assembler::AVX_512bit);
4392 } else {
4393 // worse case scenario, all regs in upper bank
4394 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4395 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4396 Assembler::vpaddw(xmm0, xmm0, src, vector_len);
4397 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4398 }
4399 }
4400
4401 void MacroAssembler::vpand(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
4402 if (reachable(src)) {
4403 Assembler::vpand(dst, nds, as_Address(src), vector_len);
4404 } else {
4405 lea(rscratch1, src);
4406 Assembler::vpand(dst, nds, Address(rscratch1, 0), vector_len);
4407 }
4408 }
4409
4410 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src) {
4411 int dst_enc = dst->encoding();
4412 int src_enc = src->encoding();
4413 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4414 Assembler::vpbroadcastw(dst, src);
4415 } else if ((dst_enc < 16) && (src_enc < 16)) {
4416 Assembler::vpbroadcastw(dst, src);
4417 } else if (src_enc < 16) {
4418 subptr(rsp, 64);
4419 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4420 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4421 Assembler::vpbroadcastw(xmm0, src);
4422 movdqu(dst, xmm0);
4423 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4424 addptr(rsp, 64);
4425 } else if (dst_enc < 16) {
4426 subptr(rsp, 64);
4427 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4428 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4429 Assembler::vpbroadcastw(dst, xmm0);
4430 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4431 addptr(rsp, 64);
4432 } else {
4433 subptr(rsp, 64);
4434 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4435 subptr(rsp, 64);
4436 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4437 movdqu(xmm0, src);
4438 movdqu(xmm1, dst);
4439 Assembler::vpbroadcastw(xmm1, xmm0);
4440 movdqu(dst, xmm1);
4441 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4442 addptr(rsp, 64);
4443 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4444 addptr(rsp, 64);
4445 }
4446 }
4447
4448 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4449 int dst_enc = dst->encoding();
4450 int nds_enc = nds->encoding();
4451 int src_enc = src->encoding();
4452 assert(dst_enc == nds_enc, "");
4453 if ((dst_enc < 16) && (src_enc < 16)) {
4454 Assembler::vpcmpeqb(dst, nds, src, vector_len);
4455 } else if (src_enc < 16) {
4456 subptr(rsp, 64);
4457 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4458 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4459 Assembler::vpcmpeqb(xmm0, xmm0, src, vector_len);
4460 movdqu(dst, xmm0);
4461 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4462 addptr(rsp, 64);
4463 } else if (dst_enc < 16) {
4464 subptr(rsp, 64);
4465 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4466 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4467 Assembler::vpcmpeqb(dst, dst, xmm0, vector_len);
4468 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4469 addptr(rsp, 64);
4470 } else {
4471 subptr(rsp, 64);
4472 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4473 subptr(rsp, 64);
4474 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4475 movdqu(xmm0, src);
4476 movdqu(xmm1, dst);
4477 Assembler::vpcmpeqb(xmm1, xmm1, xmm0, vector_len);
4478 movdqu(dst, xmm1);
4479 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4480 addptr(rsp, 64);
4481 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4482 addptr(rsp, 64);
4483 }
4484 }
4485
4486 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4487 int dst_enc = dst->encoding();
4488 int nds_enc = nds->encoding();
4489 int src_enc = src->encoding();
4490 assert(dst_enc == nds_enc, "");
4491 if ((dst_enc < 16) && (src_enc < 16)) {
4492 Assembler::vpcmpeqw(dst, nds, src, vector_len);
4493 } else if (src_enc < 16) {
4494 subptr(rsp, 64);
4495 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4496 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4497 Assembler::vpcmpeqw(xmm0, xmm0, src, vector_len);
4498 movdqu(dst, xmm0);
4499 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4500 addptr(rsp, 64);
4501 } else if (dst_enc < 16) {
4502 subptr(rsp, 64);
4503 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4504 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4505 Assembler::vpcmpeqw(dst, dst, xmm0, vector_len);
4506 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4507 addptr(rsp, 64);
4508 } else {
4509 subptr(rsp, 64);
4510 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4511 subptr(rsp, 64);
4512 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4513 movdqu(xmm0, src);
4514 movdqu(xmm1, dst);
4515 Assembler::vpcmpeqw(xmm1, xmm1, xmm0, vector_len);
4516 movdqu(dst, xmm1);
4517 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4518 addptr(rsp, 64);
4519 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4520 addptr(rsp, 64);
4521 }
4522 }
4523
4524 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) {
4525 int dst_enc = dst->encoding();
4526 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4527 Assembler::vpmovzxbw(dst, src, vector_len);
4528 } else if (dst_enc < 16) {
4529 Assembler::vpmovzxbw(dst, src, vector_len);
4530 } else {
4531 subptr(rsp, 64);
4532 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4533 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4534 Assembler::vpmovzxbw(xmm0, src, vector_len);
4535 movdqu(dst, xmm0);
4536 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4537 addptr(rsp, 64);
4538 }
4539 }
4540
4541 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) {
4542 int src_enc = src->encoding();
4543 if (src_enc < 16) {
4544 Assembler::vpmovmskb(dst, src);
4545 } else {
4546 subptr(rsp, 64);
4547 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4548 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4549 Assembler::vpmovmskb(dst, xmm0);
4550 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4551 addptr(rsp, 64);
4552 }
4553 }
4554
4555 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4556 int dst_enc = dst->encoding();
4557 int nds_enc = nds->encoding();
4558 int src_enc = src->encoding();
4559 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4560 Assembler::vpmullw(dst, nds, src, vector_len);
4561 } else if ((dst_enc < 16) && (src_enc < 16)) {
4562 Assembler::vpmullw(dst, dst, src, vector_len);
4563 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4564 // use nds as scratch for src
4565 evmovdqul(nds, src, Assembler::AVX_512bit);
4566 Assembler::vpmullw(dst, dst, nds, vector_len);
4567 } else if ((src_enc < 16) && (nds_enc < 16)) {
4568 // use nds as scratch for dst
4569 evmovdqul(nds, dst, Assembler::AVX_512bit);
4570 Assembler::vpmullw(nds, nds, src, vector_len);
4571 evmovdqul(dst, nds, Assembler::AVX_512bit);
4572 } else if (dst_enc < 16) {
4573 // use nds as scatch for xmm0 to hold src
4574 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4575 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4576 Assembler::vpmullw(dst, dst, xmm0, vector_len);
4577 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4578 } else {
4579 // worse case scenario, all regs are in the upper bank
4580 subptr(rsp, 64);
4581 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4582 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4583 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4584 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4585 Assembler::vpmullw(xmm0, xmm0, xmm1, vector_len);
4586 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4587 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4588 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4589 addptr(rsp, 64);
4590 }
4591 }
4592
4593 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4594 int dst_enc = dst->encoding();
4595 int nds_enc = nds->encoding();
4596 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4597 Assembler::vpmullw(dst, nds, src, vector_len);
4598 } else if (dst_enc < 16) {
4599 Assembler::vpmullw(dst, dst, src, vector_len);
4600 } else if (nds_enc < 16) {
4601 // implies dst_enc in upper bank with src as scratch
4602 evmovdqul(nds, dst, Assembler::AVX_512bit);
4603 Assembler::vpmullw(nds, nds, src, vector_len);
4604 evmovdqul(dst, nds, Assembler::AVX_512bit);
4605 } else {
4606 // worse case scenario, all regs in upper bank
4607 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4608 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4609 Assembler::vpmullw(xmm0, xmm0, src, vector_len);
4610 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4611 }
4612 }
4613
4614 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4615 int dst_enc = dst->encoding();
4616 int nds_enc = nds->encoding();
4617 int src_enc = src->encoding();
4618 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4619 Assembler::vpsubb(dst, nds, src, vector_len);
4620 } else if ((dst_enc < 16) && (src_enc < 16)) {
4621 Assembler::vpsubb(dst, dst, src, vector_len);
4622 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4623 // use nds as scratch for src
4624 evmovdqul(nds, src, Assembler::AVX_512bit);
4625 Assembler::vpsubb(dst, dst, nds, vector_len);
4626 } else if ((src_enc < 16) && (nds_enc < 16)) {
4627 // use nds as scratch for dst
4628 evmovdqul(nds, dst, Assembler::AVX_512bit);
4629 Assembler::vpsubb(nds, nds, src, vector_len);
4630 evmovdqul(dst, nds, Assembler::AVX_512bit);
4631 } else if (dst_enc < 16) {
4632 // use nds as scatch for xmm0 to hold src
4633 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4634 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4635 Assembler::vpsubb(dst, dst, xmm0, vector_len);
4636 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4637 } else {
4638 // worse case scenario, all regs are in the upper bank
4639 subptr(rsp, 64);
4640 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4641 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4642 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4643 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4644 Assembler::vpsubb(xmm0, xmm0, xmm1, vector_len);
4645 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4646 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4647 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4648 addptr(rsp, 64);
4649 }
4650 }
4651
4652 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4653 int dst_enc = dst->encoding();
4654 int nds_enc = nds->encoding();
4655 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4656 Assembler::vpsubb(dst, nds, src, vector_len);
4657 } else if (dst_enc < 16) {
4658 Assembler::vpsubb(dst, dst, src, vector_len);
4659 } else if (nds_enc < 16) {
4660 // implies dst_enc in upper bank with src as scratch
4661 evmovdqul(nds, dst, Assembler::AVX_512bit);
4662 Assembler::vpsubb(nds, nds, src, vector_len);
4663 evmovdqul(dst, nds, Assembler::AVX_512bit);
4664 } else {
4665 // worse case scenario, all regs in upper bank
4666 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4667 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4668 Assembler::vpsubw(xmm0, xmm0, src, vector_len);
4669 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4670 }
4671 }
4672
4673 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4674 int dst_enc = dst->encoding();
4675 int nds_enc = nds->encoding();
4676 int src_enc = src->encoding();
4677 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4678 Assembler::vpsubw(dst, nds, src, vector_len);
4679 } else if ((dst_enc < 16) && (src_enc < 16)) {
4680 Assembler::vpsubw(dst, dst, src, vector_len);
4681 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4682 // use nds as scratch for src
4683 evmovdqul(nds, src, Assembler::AVX_512bit);
4684 Assembler::vpsubw(dst, dst, nds, vector_len);
4685 } else if ((src_enc < 16) && (nds_enc < 16)) {
4686 // use nds as scratch for dst
4687 evmovdqul(nds, dst, Assembler::AVX_512bit);
4688 Assembler::vpsubw(nds, nds, src, vector_len);
4689 evmovdqul(dst, nds, Assembler::AVX_512bit);
4690 } else if (dst_enc < 16) {
4691 // use nds as scatch for xmm0 to hold src
4692 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4693 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4694 Assembler::vpsubw(dst, dst, xmm0, vector_len);
4695 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4696 } else {
4697 // worse case scenario, all regs are in the upper bank
4698 subptr(rsp, 64);
4699 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4700 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4701 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4702 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4703 Assembler::vpsubw(xmm0, xmm0, xmm1, vector_len);
4704 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4705 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4706 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4707 addptr(rsp, 64);
4708 }
4709 }
4710
4711 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4712 int dst_enc = dst->encoding();
4713 int nds_enc = nds->encoding();
4714 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4715 Assembler::vpsubw(dst, nds, src, vector_len);
4716 } else if (dst_enc < 16) {
4717 Assembler::vpsubw(dst, dst, src, vector_len);
4718 } else if (nds_enc < 16) {
4719 // implies dst_enc in upper bank with src as scratch
4720 evmovdqul(nds, dst, Assembler::AVX_512bit);
4721 Assembler::vpsubw(nds, nds, src, vector_len);
4722 evmovdqul(dst, nds, Assembler::AVX_512bit);
4723 } else {
4724 // worse case scenario, all regs in upper bank
4725 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4726 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4727 Assembler::vpsubw(xmm0, xmm0, src, vector_len);
4728 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4729 }
4730 }
4731
4732 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4733 int dst_enc = dst->encoding();
4734 int nds_enc = nds->encoding();
4735 int shift_enc = shift->encoding();
4736 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4737 Assembler::vpsraw(dst, nds, shift, vector_len);
4738 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4739 Assembler::vpsraw(dst, dst, shift, vector_len);
4740 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4741 // use nds_enc as scratch with shift
4742 evmovdqul(nds, shift, Assembler::AVX_512bit);
4743 Assembler::vpsraw(dst, dst, nds, vector_len);
4744 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4745 // use nds as scratch with dst
4746 evmovdqul(nds, dst, Assembler::AVX_512bit);
4747 Assembler::vpsraw(nds, nds, shift, vector_len);
4748 evmovdqul(dst, nds, Assembler::AVX_512bit);
4749 } else if (dst_enc < 16) {
4750 // use nds to save a copy of xmm0 and hold shift
4751 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4752 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4753 Assembler::vpsraw(dst, dst, xmm0, vector_len);
4754 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4755 } else if (nds_enc < 16) {
4756 // use nds as dest as temps
4757 evmovdqul(nds, dst, Assembler::AVX_512bit);
4758 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4759 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4760 Assembler::vpsraw(nds, nds, xmm0, vector_len);
4761 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4762 evmovdqul(dst, nds, Assembler::AVX_512bit);
4763 } else {
4764 // worse case scenario, all regs are in the upper bank
4765 subptr(rsp, 64);
4766 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4767 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4768 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4769 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4770 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4771 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4772 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4773 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4774 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4775 addptr(rsp, 64);
4776 }
4777 }
4778
4779 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4780 int dst_enc = dst->encoding();
4781 int nds_enc = nds->encoding();
4782 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4783 Assembler::vpsraw(dst, nds, shift, vector_len);
4784 } else if (dst_enc < 16) {
4785 Assembler::vpsraw(dst, dst, shift, vector_len);
4786 } else if (nds_enc < 16) {
4787 // use nds as scratch
4788 evmovdqul(nds, dst, Assembler::AVX_512bit);
4789 Assembler::vpsraw(nds, nds, shift, vector_len);
4790 evmovdqul(dst, nds, Assembler::AVX_512bit);
4791 } else {
4792 // use nds as scratch for xmm0
4793 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4794 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4795 Assembler::vpsraw(xmm0, xmm0, shift, vector_len);
4796 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4797 }
4798 }
4799
4800 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4801 int dst_enc = dst->encoding();
4802 int nds_enc = nds->encoding();
4803 int shift_enc = shift->encoding();
4804 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4805 Assembler::vpsrlw(dst, nds, shift, vector_len);
4806 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4807 Assembler::vpsrlw(dst, dst, shift, vector_len);
4808 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4809 // use nds_enc as scratch with shift
4810 evmovdqul(nds, shift, Assembler::AVX_512bit);
4811 Assembler::vpsrlw(dst, dst, nds, vector_len);
4812 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4813 // use nds as scratch with dst
4814 evmovdqul(nds, dst, Assembler::AVX_512bit);
4815 Assembler::vpsrlw(nds, nds, shift, vector_len);
4816 evmovdqul(dst, nds, Assembler::AVX_512bit);
4817 } else if (dst_enc < 16) {
4818 // use nds to save a copy of xmm0 and hold shift
4819 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4820 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4821 Assembler::vpsrlw(dst, dst, xmm0, vector_len);
4822 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4823 } else if (nds_enc < 16) {
4824 // use nds as dest as temps
4825 evmovdqul(nds, dst, Assembler::AVX_512bit);
4826 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4827 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4828 Assembler::vpsrlw(nds, nds, xmm0, vector_len);
4829 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4830 evmovdqul(dst, nds, Assembler::AVX_512bit);
4831 } else {
4832 // worse case scenario, all regs are in the upper bank
4833 subptr(rsp, 64);
4834 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4835 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4836 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4837 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4838 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4839 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4840 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4841 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4842 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4843 addptr(rsp, 64);
4844 }
4845 }
4846
4847 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4848 int dst_enc = dst->encoding();
4849 int nds_enc = nds->encoding();
4850 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4851 Assembler::vpsrlw(dst, nds, shift, vector_len);
4852 } else if (dst_enc < 16) {
4853 Assembler::vpsrlw(dst, dst, shift, vector_len);
4854 } else if (nds_enc < 16) {
4855 // use nds as scratch
4856 evmovdqul(nds, dst, Assembler::AVX_512bit);
4857 Assembler::vpsrlw(nds, nds, shift, vector_len);
4858 evmovdqul(dst, nds, Assembler::AVX_512bit);
4859 } else {
4860 // use nds as scratch for xmm0
4861 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4862 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4863 Assembler::vpsrlw(xmm0, xmm0, shift, vector_len);
4864 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4865 }
4866 }
4867
4868 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4869 int dst_enc = dst->encoding();
4870 int nds_enc = nds->encoding();
4871 int shift_enc = shift->encoding();
4872 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4873 Assembler::vpsllw(dst, nds, shift, vector_len);
4874 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4875 Assembler::vpsllw(dst, dst, shift, vector_len);
4876 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4877 // use nds_enc as scratch with shift
4878 evmovdqul(nds, shift, Assembler::AVX_512bit);
4879 Assembler::vpsllw(dst, dst, nds, vector_len);
4880 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4881 // use nds as scratch with dst
4882 evmovdqul(nds, dst, Assembler::AVX_512bit);
4883 Assembler::vpsllw(nds, nds, shift, vector_len);
4884 evmovdqul(dst, nds, Assembler::AVX_512bit);
4885 } else if (dst_enc < 16) {
4886 // use nds to save a copy of xmm0 and hold shift
4887 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4888 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4889 Assembler::vpsllw(dst, dst, xmm0, vector_len);
4890 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4891 } else if (nds_enc < 16) {
4892 // use nds as dest as temps
4893 evmovdqul(nds, dst, Assembler::AVX_512bit);
4894 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4895 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4896 Assembler::vpsllw(nds, nds, xmm0, vector_len);
4897 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4898 evmovdqul(dst, nds, Assembler::AVX_512bit);
4899 } else {
4900 // worse case scenario, all regs are in the upper bank
4901 subptr(rsp, 64);
4902 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4903 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4904 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4905 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4906 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4907 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4908 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4909 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4910 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4911 addptr(rsp, 64);
4912 }
4913 }
4914
4915 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4916 int dst_enc = dst->encoding();
4917 int nds_enc = nds->encoding();
4918 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4919 Assembler::vpsllw(dst, nds, shift, vector_len);
4920 } else if (dst_enc < 16) {
4921 Assembler::vpsllw(dst, dst, shift, vector_len);
4922 } else if (nds_enc < 16) {
4923 // use nds as scratch
4924 evmovdqul(nds, dst, Assembler::AVX_512bit);
4925 Assembler::vpsllw(nds, nds, shift, vector_len);
4926 evmovdqul(dst, nds, Assembler::AVX_512bit);
4927 } else {
4928 // use nds as scratch for xmm0
4929 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4930 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4931 Assembler::vpsllw(xmm0, xmm0, shift, vector_len);
4932 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4933 }
4934 }
4935
4936 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) {
4937 int dst_enc = dst->encoding();
4938 int src_enc = src->encoding();
4939 if ((dst_enc < 16) && (src_enc < 16)) {
4940 Assembler::vptest(dst, src);
4941 } else if (src_enc < 16) {
4942 subptr(rsp, 64);
4943 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4944 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4945 Assembler::vptest(xmm0, src);
4946 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4947 addptr(rsp, 64);
4948 } else if (dst_enc < 16) {
4949 subptr(rsp, 64);
4950 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4951 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4952 Assembler::vptest(dst, xmm0);
4953 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4954 addptr(rsp, 64);
4955 } else {
4956 subptr(rsp, 64);
4957 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4958 subptr(rsp, 64);
4959 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4960 movdqu(xmm0, src);
4961 movdqu(xmm1, dst);
4962 Assembler::vptest(xmm1, xmm0);
4963 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4964 addptr(rsp, 64);
4965 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4966 addptr(rsp, 64);
4967 }
4968 }
4969
4970 // This instruction exists within macros, ergo we cannot control its input
4971 // when emitted through those patterns.
4972 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) {
4973 if (VM_Version::supports_avx512nobw()) {
4974 int dst_enc = dst->encoding();
4975 int src_enc = src->encoding();
4976 if (dst_enc == src_enc) {
4977 if (dst_enc < 16) {
4978 Assembler::punpcklbw(dst, src);
4979 } else {
4980 subptr(rsp, 64);
4981 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4982 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4983 Assembler::punpcklbw(xmm0, xmm0);
4984 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4985 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4986 addptr(rsp, 64);
4987 }
4988 } else {
4989 if ((src_enc < 16) && (dst_enc < 16)) {
4990 Assembler::punpcklbw(dst, src);
4991 } else if (src_enc < 16) {
4992 subptr(rsp, 64);
4993 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4994 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4995 Assembler::punpcklbw(xmm0, src);
4996 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4997 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4998 addptr(rsp, 64);
4999 } else if (dst_enc < 16) {
5000 subptr(rsp, 64);
5001 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5002 evmovdqul(xmm0, src, Assembler::AVX_512bit);
5003 Assembler::punpcklbw(dst, xmm0);
5004 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5005 addptr(rsp, 64);
5006 } else {
5007 subptr(rsp, 64);
5008 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5009 subptr(rsp, 64);
5010 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
5011 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5012 evmovdqul(xmm1, src, Assembler::AVX_512bit);
5013 Assembler::punpcklbw(xmm0, xmm1);
5014 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5015 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
5016 addptr(rsp, 64);
5017 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5018 addptr(rsp, 64);
5019 }
5020 }
5021 } else {
5022 Assembler::punpcklbw(dst, src);
5023 }
5024 }
5025
5026 void MacroAssembler::pshufd(XMMRegister dst, Address src, int mode) {
5027 if (VM_Version::supports_avx512vl()) {
5028 Assembler::pshufd(dst, src, mode);
5029 } else {
5030 int dst_enc = dst->encoding();
5031 if (dst_enc < 16) {
5032 Assembler::pshufd(dst, src, mode);
5033 } else {
5034 subptr(rsp, 64);
5035 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5036 Assembler::pshufd(xmm0, src, mode);
5037 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5038 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5039 addptr(rsp, 64);
5040 }
5041 }
5042 }
5043
5044 // This instruction exists within macros, ergo we cannot control its input
5045 // when emitted through those patterns.
5046 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) {
5047 if (VM_Version::supports_avx512nobw()) {
5048 int dst_enc = dst->encoding();
5049 int src_enc = src->encoding();
5050 if (dst_enc == src_enc) {
5051 if (dst_enc < 16) {
5052 Assembler::pshuflw(dst, src, mode);
5053 } else {
5054 subptr(rsp, 64);
5055 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5056 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5057 Assembler::pshuflw(xmm0, xmm0, mode);
5058 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5059 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5060 addptr(rsp, 64);
5061 }
5062 } else {
5063 if ((src_enc < 16) && (dst_enc < 16)) {
5064 Assembler::pshuflw(dst, src, mode);
5065 } else if (src_enc < 16) {
5066 subptr(rsp, 64);
5067 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5068 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5069 Assembler::pshuflw(xmm0, src, mode);
5070 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5071 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5072 addptr(rsp, 64);
5073 } else if (dst_enc < 16) {
5074 subptr(rsp, 64);
5075 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5076 evmovdqul(xmm0, src, Assembler::AVX_512bit);
5077 Assembler::pshuflw(dst, xmm0, mode);
5078 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5079 addptr(rsp, 64);
5080 } else {
5081 subptr(rsp, 64);
5082 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5083 subptr(rsp, 64);
5084 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
5085 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5086 evmovdqul(xmm1, src, Assembler::AVX_512bit);
5087 Assembler::pshuflw(xmm0, xmm1, mode);
5088 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5089 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
5090 addptr(rsp, 64);
5091 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5092 addptr(rsp, 64);
5093 }
5094 }
5095 } else {
5096 Assembler::pshuflw(dst, src, mode);
5097 }
5098 }
5099
5100 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5101 if (reachable(src)) {
5102 vandpd(dst, nds, as_Address(src), vector_len);
5103 } else {
5104 lea(rscratch1, src);
5105 vandpd(dst, nds, Address(rscratch1, 0), vector_len);
5106 }
5107 }
5108
5109 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5110 if (reachable(src)) {
5111 vandps(dst, nds, as_Address(src), vector_len);
5112 } else {
5113 lea(rscratch1, src);
5114 vandps(dst, nds, Address(rscratch1, 0), vector_len);
5115 }
5116 }
5117
5118 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5119 if (reachable(src)) {
5120 vdivsd(dst, nds, as_Address(src));
5121 } else {
5122 lea(rscratch1, src);
5123 vdivsd(dst, nds, Address(rscratch1, 0));
5124 }
5125 }
5126
5127 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5128 if (reachable(src)) {
5129 vdivss(dst, nds, as_Address(src));
5130 } else {
5131 lea(rscratch1, src);
5132 vdivss(dst, nds, Address(rscratch1, 0));
5133 }
5134 }
5135
5136 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5137 if (reachable(src)) {
5138 vmulsd(dst, nds, as_Address(src));
5139 } else {
5140 lea(rscratch1, src);
5141 vmulsd(dst, nds, Address(rscratch1, 0));
5142 }
5143 }
5144
5145 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5146 if (reachable(src)) {
5147 vmulss(dst, nds, as_Address(src));
5148 } else {
5149 lea(rscratch1, src);
5150 vmulss(dst, nds, Address(rscratch1, 0));
5151 }
5152 }
5153
5154 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5155 if (reachable(src)) {
5156 vsubsd(dst, nds, as_Address(src));
5157 } else {
5158 lea(rscratch1, src);
5159 vsubsd(dst, nds, Address(rscratch1, 0));
5160 }
5161 }
5162
5163 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5164 if (reachable(src)) {
5165 vsubss(dst, nds, as_Address(src));
5166 } else {
5167 lea(rscratch1, src);
5168 vsubss(dst, nds, Address(rscratch1, 0));
5169 }
5170 }
5171
5172 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5173 int nds_enc = nds->encoding();
5174 int dst_enc = dst->encoding();
5175 bool dst_upper_bank = (dst_enc > 15);
5176 bool nds_upper_bank = (nds_enc > 15);
5177 if (VM_Version::supports_avx512novl() &&
5178 (nds_upper_bank || dst_upper_bank)) {
5179 if (dst_upper_bank) {
5180 subptr(rsp, 64);
5181 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5182 movflt(xmm0, nds);
5183 vxorps(xmm0, xmm0, src, Assembler::AVX_128bit);
5184 movflt(dst, xmm0);
5185 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5186 addptr(rsp, 64);
5187 } else {
5188 movflt(dst, nds);
5189 vxorps(dst, dst, src, Assembler::AVX_128bit);
5190 }
5191 } else {
5192 vxorps(dst, nds, src, Assembler::AVX_128bit);
5193 }
5194 }
5195
5196 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5197 int nds_enc = nds->encoding();
5198 int dst_enc = dst->encoding();
5199 bool dst_upper_bank = (dst_enc > 15);
5200 bool nds_upper_bank = (nds_enc > 15);
5201 if (VM_Version::supports_avx512novl() &&
5202 (nds_upper_bank || dst_upper_bank)) {
5203 if (dst_upper_bank) {
5204 subptr(rsp, 64);
5205 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5206 movdbl(xmm0, nds);
5207 vxorpd(xmm0, xmm0, src, Assembler::AVX_128bit);
5208 movdbl(dst, xmm0);
5209 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5210 addptr(rsp, 64);
5211 } else {
5212 movdbl(dst, nds);
5213 vxorpd(dst, dst, src, Assembler::AVX_128bit);
5214 }
5215 } else {
5216 vxorpd(dst, nds, src, Assembler::AVX_128bit);
5217 }
5218 }
5219
5220 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5221 if (reachable(src)) {
5222 vxorpd(dst, nds, as_Address(src), vector_len);
5223 } else {
5224 lea(rscratch1, src);
5225 vxorpd(dst, nds, Address(rscratch1, 0), vector_len);
5226 }
5227 }
5228
5229 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5230 if (reachable(src)) {
5231 vxorps(dst, nds, as_Address(src), vector_len);
5232 } else {
5233 lea(rscratch1, src);
5234 vxorps(dst, nds, Address(rscratch1, 0), vector_len);
5235 }
5236 }
5237
5238 void MacroAssembler::clear_jweak_tag(Register possibly_jweak) {
5239 const int32_t inverted_jweak_mask = ~static_cast<int32_t>(JNIHandles::weak_tag_mask);
5240 STATIC_ASSERT(inverted_jweak_mask == -2); // otherwise check this code
5241 // The inverted mask is sign-extended
5242 andptr(possibly_jweak, inverted_jweak_mask);
5243 }
5244
5245 void MacroAssembler::resolve_jobject(Register value,
5246 Register thread,
5247 Register tmp) {
5248 assert_different_registers(value, thread, tmp);
5249 Label done, not_weak;
5250 testptr(value, value);
5251 jcc(Assembler::zero, done); // Use NULL as-is.
5252 testptr(value, JNIHandles::weak_tag_mask); // Test for jweak tag.
5253 jcc(Assembler::zero, not_weak);
5254 // Resolve jweak.
5255 access_load_at(T_OBJECT, IN_ROOT | ON_PHANTOM_OOP_REF,
5256 value, Address(value, -JNIHandles::weak_tag_value), tmp, thread);
5257 verify_oop(value);
5258 jmp(done);
5259 bind(not_weak);
5260 // Resolve (untagged) jobject.
5261 access_load_at(T_OBJECT, IN_ROOT | IN_CONCURRENT_ROOT,
5262 value, Address(value, 0), tmp, thread);
5263 verify_oop(value);
5264 bind(done);
5265 }
5266
5267 void MacroAssembler::subptr(Register dst, int32_t imm32) {
5268 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32));
5269 }
5270
5271 // Force generation of a 4 byte immediate value even if it fits into 8bit
5272 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) {
5273 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32));
5274 }
5275
5276 void MacroAssembler::subptr(Register dst, Register src) {
5277 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src));
5278 }
5279
5280 // C++ bool manipulation
5281 void MacroAssembler::testbool(Register dst) {
5282 if(sizeof(bool) == 1)
5283 testb(dst, 0xff);
5284 else if(sizeof(bool) == 2) {
5285 // testw implementation needed for two byte bools
5286 ShouldNotReachHere();
5287 } else if(sizeof(bool) == 4)
5288 testl(dst, dst);
5289 else
5290 // unsupported
5291 ShouldNotReachHere();
5292 }
5293
5294 void MacroAssembler::testptr(Register dst, Register src) {
5295 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src));
5296 }
5297
5298 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
5299 void MacroAssembler::tlab_allocate(Register obj,
5300 Register var_size_in_bytes,
5301 int con_size_in_bytes,
5302 Register t1,
5303 Register t2,
5304 Label& slow_case) {
5305 assert_different_registers(obj, t1, t2);
5306 assert_different_registers(obj, var_size_in_bytes, t1);
5307 Register end = t2;
5308 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread);
5309
5310 verify_tlab();
5311
5312 NOT_LP64(get_thread(thread));
5313
5314 movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
5315 if (var_size_in_bytes == noreg) {
5316 lea(end, Address(obj, con_size_in_bytes));
5317 } else {
5318 lea(end, Address(obj, var_size_in_bytes, Address::times_1));
5319 }
5320 cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
5321 jcc(Assembler::above, slow_case);
5322
5323 // update the tlab top pointer
5324 movptr(Address(thread, JavaThread::tlab_top_offset()), end);
5325
5326 // recover var_size_in_bytes if necessary
5327 if (var_size_in_bytes == end) {
5328 subptr(var_size_in_bytes, obj);
5329 }
5330 verify_tlab();
5331 }
5332
5333 // Preserves the contents of address, destroys the contents length_in_bytes and temp.
5334 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) {
5335 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different");
5336 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord");
5337 Label done;
5338
5339 testptr(length_in_bytes, length_in_bytes);
5340 jcc(Assembler::zero, done);
5341
5342 // initialize topmost word, divide index by 2, check if odd and test if zero
5343 // note: for the remaining code to work, index must be a multiple of BytesPerWord
5344 #ifdef ASSERT
5345 {
5346 Label L;
5347 testptr(length_in_bytes, BytesPerWord - 1);
5348 jcc(Assembler::zero, L);
5349 stop("length must be a multiple of BytesPerWord");
5350 bind(L);
5351 }
5352 #endif
5353 Register index = length_in_bytes;
5354 xorptr(temp, temp); // use _zero reg to clear memory (shorter code)
5355 if (UseIncDec) {
5356 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set
5357 } else {
5358 shrptr(index, 2); // use 2 instructions to avoid partial flag stall
5359 shrptr(index, 1);
5360 }
5361 #ifndef _LP64
5362 // index could have not been a multiple of 8 (i.e., bit 2 was set)
5363 {
5364 Label even;
5365 // note: if index was a multiple of 8, then it cannot
5366 // be 0 now otherwise it must have been 0 before
5367 // => if it is even, we don't need to check for 0 again
5368 jcc(Assembler::carryClear, even);
5369 // clear topmost word (no jump would be needed if conditional assignment worked here)
5370 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp);
5371 // index could be 0 now, must check again
5372 jcc(Assembler::zero, done);
5373 bind(even);
5374 }
5375 #endif // !_LP64
5376 // initialize remaining object fields: index is a multiple of 2 now
5377 {
5378 Label loop;
5379 bind(loop);
5380 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp);
5381 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);)
5382 decrement(index);
5383 jcc(Assembler::notZero, loop);
5384 }
5385
5386 bind(done);
5387 }
5388
5389 void MacroAssembler::incr_allocated_bytes(Register thread,
5390 Register var_size_in_bytes,
5391 int con_size_in_bytes,
5392 Register t1) {
5393 if (!thread->is_valid()) {
5394 #ifdef _LP64
5395 thread = r15_thread;
5396 #else
5397 assert(t1->is_valid(), "need temp reg");
5398 thread = t1;
5399 get_thread(thread);
5400 #endif
5401 }
5402
5403 #ifdef _LP64
5404 if (var_size_in_bytes->is_valid()) {
5405 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
5406 } else {
5407 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
5408 }
5409 #else
5410 if (var_size_in_bytes->is_valid()) {
5411 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
5412 } else {
5413 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
5414 }
5415 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0);
5416 #endif
5417 }
5418
5419 // Look up the method for a megamorphic invokeinterface call.
5420 // The target method is determined by <intf_klass, itable_index>.
5421 // The receiver klass is in recv_klass.
5422 // On success, the result will be in method_result, and execution falls through.
5423 // On failure, execution transfers to the given label.
5424 void MacroAssembler::lookup_interface_method(Register recv_klass,
5425 Register intf_klass,
5426 RegisterOrConstant itable_index,
5427 Register method_result,
5428 Register scan_temp,
5429 Label& L_no_such_interface,
5430 bool return_method) {
5431 assert_different_registers(recv_klass, intf_klass, scan_temp);
5432 assert_different_registers(method_result, intf_klass, scan_temp);
5433 assert(recv_klass != method_result || !return_method,
5434 "recv_klass can be destroyed when method isn't needed");
5435
5436 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
5437 "caller must use same register for non-constant itable index as for method");
5438
5439 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
5440 int vtable_base = in_bytes(Klass::vtable_start_offset());
5441 int itentry_off = itableMethodEntry::method_offset_in_bytes();
5442 int scan_step = itableOffsetEntry::size() * wordSize;
5443 int vte_size = vtableEntry::size_in_bytes();
5444 Address::ScaleFactor times_vte_scale = Address::times_ptr;
5445 assert(vte_size == wordSize, "else adjust times_vte_scale");
5446
5447 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
5448
5449 // %%% Could store the aligned, prescaled offset in the klassoop.
5450 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
5451
5452 if (return_method) {
5453 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
5454 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
5455 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
5456 }
5457
5458 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
5459 // if (scan->interface() == intf) {
5460 // result = (klass + scan->offset() + itable_index);
5461 // }
5462 // }
5463 Label search, found_method;
5464
5465 for (int peel = 1; peel >= 0; peel--) {
5466 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
5467 cmpptr(intf_klass, method_result);
5468
5469 if (peel) {
5470 jccb(Assembler::equal, found_method);
5471 } else {
5472 jccb(Assembler::notEqual, search);
5473 // (invert the test to fall through to found_method...)
5474 }
5475
5476 if (!peel) break;
5477
5478 bind(search);
5479
5480 // Check that the previous entry is non-null. A null entry means that
5481 // the receiver class doesn't implement the interface, and wasn't the
5482 // same as when the caller was compiled.
5483 testptr(method_result, method_result);
5484 jcc(Assembler::zero, L_no_such_interface);
5485 addptr(scan_temp, scan_step);
5486 }
5487
5488 bind(found_method);
5489
5490 if (return_method) {
5491 // Got a hit.
5492 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
5493 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
5494 }
5495 }
5496
5497
5498 // virtual method calling
5499 void MacroAssembler::lookup_virtual_method(Register recv_klass,
5500 RegisterOrConstant vtable_index,
5501 Register method_result) {
5502 const int base = in_bytes(Klass::vtable_start_offset());
5503 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below");
5504 Address vtable_entry_addr(recv_klass,
5505 vtable_index, Address::times_ptr,
5506 base + vtableEntry::method_offset_in_bytes());
5507 movptr(method_result, vtable_entry_addr);
5508 }
5509
5510
5511 void MacroAssembler::check_klass_subtype(Register sub_klass,
5512 Register super_klass,
5513 Register temp_reg,
5514 Label& L_success) {
5515 Label L_failure;
5516 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL);
5517 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
5518 bind(L_failure);
5519 }
5520
5521
5522 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
5523 Register super_klass,
5524 Register temp_reg,
5525 Label* L_success,
5526 Label* L_failure,
5527 Label* L_slow_path,
5528 RegisterOrConstant super_check_offset) {
5529 assert_different_registers(sub_klass, super_klass, temp_reg);
5530 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
5531 if (super_check_offset.is_register()) {
5532 assert_different_registers(sub_klass, super_klass,
5533 super_check_offset.as_register());
5534 } else if (must_load_sco) {
5535 assert(temp_reg != noreg, "supply either a temp or a register offset");
5536 }
5537
5538 Label L_fallthrough;
5539 int label_nulls = 0;
5540 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
5541 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
5542 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
5543 assert(label_nulls <= 1, "at most one NULL in the batch");
5544
5545 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
5546 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5547 Address super_check_offset_addr(super_klass, sco_offset);
5548
5549 // Hacked jcc, which "knows" that L_fallthrough, at least, is in
5550 // range of a jccb. If this routine grows larger, reconsider at
5551 // least some of these.
5552 #define local_jcc(assembler_cond, label) \
5553 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \
5554 else jcc( assembler_cond, label) /*omit semi*/
5555
5556 // Hacked jmp, which may only be used just before L_fallthrough.
5557 #define final_jmp(label) \
5558 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
5559 else jmp(label) /*omit semi*/
5560
5561 // If the pointers are equal, we are done (e.g., String[] elements).
5562 // This self-check enables sharing of secondary supertype arrays among
5563 // non-primary types such as array-of-interface. Otherwise, each such
5564 // type would need its own customized SSA.
5565 // We move this check to the front of the fast path because many
5566 // type checks are in fact trivially successful in this manner,
5567 // so we get a nicely predicted branch right at the start of the check.
5568 cmpptr(sub_klass, super_klass);
5569 local_jcc(Assembler::equal, *L_success);
5570
5571 // Check the supertype display:
5572 if (must_load_sco) {
5573 // Positive movl does right thing on LP64.
5574 movl(temp_reg, super_check_offset_addr);
5575 super_check_offset = RegisterOrConstant(temp_reg);
5576 }
5577 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0);
5578 cmpptr(super_klass, super_check_addr); // load displayed supertype
5579
5580 // This check has worked decisively for primary supers.
5581 // Secondary supers are sought in the super_cache ('super_cache_addr').
5582 // (Secondary supers are interfaces and very deeply nested subtypes.)
5583 // This works in the same check above because of a tricky aliasing
5584 // between the super_cache and the primary super display elements.
5585 // (The 'super_check_addr' can address either, as the case requires.)
5586 // Note that the cache is updated below if it does not help us find
5587 // what we need immediately.
5588 // So if it was a primary super, we can just fail immediately.
5589 // Otherwise, it's the slow path for us (no success at this point).
5590
5591 if (super_check_offset.is_register()) {
5592 local_jcc(Assembler::equal, *L_success);
5593 cmpl(super_check_offset.as_register(), sc_offset);
5594 if (L_failure == &L_fallthrough) {
5595 local_jcc(Assembler::equal, *L_slow_path);
5596 } else {
5597 local_jcc(Assembler::notEqual, *L_failure);
5598 final_jmp(*L_slow_path);
5599 }
5600 } else if (super_check_offset.as_constant() == sc_offset) {
5601 // Need a slow path; fast failure is impossible.
5602 if (L_slow_path == &L_fallthrough) {
5603 local_jcc(Assembler::equal, *L_success);
5604 } else {
5605 local_jcc(Assembler::notEqual, *L_slow_path);
5606 final_jmp(*L_success);
5607 }
5608 } else {
5609 // No slow path; it's a fast decision.
5610 if (L_failure == &L_fallthrough) {
5611 local_jcc(Assembler::equal, *L_success);
5612 } else {
5613 local_jcc(Assembler::notEqual, *L_failure);
5614 final_jmp(*L_success);
5615 }
5616 }
5617
5618 bind(L_fallthrough);
5619
5620 #undef local_jcc
5621 #undef final_jmp
5622 }
5623
5624
5625 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
5626 Register super_klass,
5627 Register temp_reg,
5628 Register temp2_reg,
5629 Label* L_success,
5630 Label* L_failure,
5631 bool set_cond_codes) {
5632 assert_different_registers(sub_klass, super_klass, temp_reg);
5633 if (temp2_reg != noreg)
5634 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg);
5635 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
5636
5637 Label L_fallthrough;
5638 int label_nulls = 0;
5639 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
5640 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
5641 assert(label_nulls <= 1, "at most one NULL in the batch");
5642
5643 // a couple of useful fields in sub_klass:
5644 int ss_offset = in_bytes(Klass::secondary_supers_offset());
5645 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
5646 Address secondary_supers_addr(sub_klass, ss_offset);
5647 Address super_cache_addr( sub_klass, sc_offset);
5648
5649 // Do a linear scan of the secondary super-klass chain.
5650 // This code is rarely used, so simplicity is a virtue here.
5651 // The repne_scan instruction uses fixed registers, which we must spill.
5652 // Don't worry too much about pre-existing connections with the input regs.
5653
5654 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super)
5655 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter)
5656
5657 // Get super_klass value into rax (even if it was in rdi or rcx).
5658 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false;
5659 if (super_klass != rax || UseCompressedOops) {
5660 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; }
5661 mov(rax, super_klass);
5662 }
5663 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; }
5664 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; }
5665
5666 #ifndef PRODUCT
5667 int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
5668 ExternalAddress pst_counter_addr((address) pst_counter);
5669 NOT_LP64( incrementl(pst_counter_addr) );
5670 LP64_ONLY( lea(rcx, pst_counter_addr) );
5671 LP64_ONLY( incrementl(Address(rcx, 0)) );
5672 #endif //PRODUCT
5673
5674 // We will consult the secondary-super array.
5675 movptr(rdi, secondary_supers_addr);
5676 // Load the array length. (Positive movl does right thing on LP64.)
5677 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes()));
5678 // Skip to start of data.
5679 addptr(rdi, Array<Klass*>::base_offset_in_bytes());
5680
5681 // Scan RCX words at [RDI] for an occurrence of RAX.
5682 // Set NZ/Z based on last compare.
5683 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does
5684 // not change flags (only scas instruction which is repeated sets flags).
5685 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found.
5686
5687 testptr(rax,rax); // Set Z = 0
5688 repne_scan();
5689
5690 // Unspill the temp. registers:
5691 if (pushed_rdi) pop(rdi);
5692 if (pushed_rcx) pop(rcx);
5693 if (pushed_rax) pop(rax);
5694
5695 if (set_cond_codes) {
5696 // Special hack for the AD files: rdi is guaranteed non-zero.
5697 assert(!pushed_rdi, "rdi must be left non-NULL");
5698 // Also, the condition codes are properly set Z/NZ on succeed/failure.
5699 }
5700
5701 if (L_failure == &L_fallthrough)
5702 jccb(Assembler::notEqual, *L_failure);
5703 else jcc(Assembler::notEqual, *L_failure);
5704
5705 // Success. Cache the super we found and proceed in triumph.
5706 movptr(super_cache_addr, super_klass);
5707
5708 if (L_success != &L_fallthrough) {
5709 jmp(*L_success);
5710 }
5711
5712 #undef IS_A_TEMP
5713
5714 bind(L_fallthrough);
5715 }
5716
5717
5718 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
5719 if (VM_Version::supports_cmov()) {
5720 cmovl(cc, dst, src);
5721 } else {
5722 Label L;
5723 jccb(negate_condition(cc), L);
5724 movl(dst, src);
5725 bind(L);
5726 }
5727 }
5728
5729 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) {
5730 if (VM_Version::supports_cmov()) {
5731 cmovl(cc, dst, src);
5732 } else {
5733 Label L;
5734 jccb(negate_condition(cc), L);
5735 movl(dst, src);
5736 bind(L);
5737 }
5738 }
5739
5740 void MacroAssembler::verify_oop(Register reg, const char* s) {
5741 if (!VerifyOops) return;
5742
5743 // Pass register number to verify_oop_subroutine
5744 const char* b = NULL;
5745 {
5746 ResourceMark rm;
5747 stringStream ss;
5748 ss.print("verify_oop: %s: %s", reg->name(), s);
5749 b = code_string(ss.as_string());
5750 }
5751 BLOCK_COMMENT("verify_oop {");
5752 #ifdef _LP64
5753 push(rscratch1); // save r10, trashed by movptr()
5754 #endif
5755 push(rax); // save rax,
5756 push(reg); // pass register argument
5757 ExternalAddress buffer((address) b);
5758 // avoid using pushptr, as it modifies scratch registers
5759 // and our contract is not to modify anything
5760 movptr(rax, buffer.addr());
5761 push(rax);
5762 // call indirectly to solve generation ordering problem
5763 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
5764 call(rax);
5765 // Caller pops the arguments (oop, message) and restores rax, r10
5766 BLOCK_COMMENT("} verify_oop");
5767 }
5768
5769
5770 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
5771 Register tmp,
5772 int offset) {
5773 intptr_t value = *delayed_value_addr;
5774 if (value != 0)
5775 return RegisterOrConstant(value + offset);
5776
5777 // load indirectly to solve generation ordering problem
5778 movptr(tmp, ExternalAddress((address) delayed_value_addr));
5779
5780 #ifdef ASSERT
5781 { Label L;
5782 testptr(tmp, tmp);
5783 if (WizardMode) {
5784 const char* buf = NULL;
5785 {
5786 ResourceMark rm;
5787 stringStream ss;
5788 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]);
5789 buf = code_string(ss.as_string());
5790 }
5791 jcc(Assembler::notZero, L);
5792 STOP(buf);
5793 } else {
5794 jccb(Assembler::notZero, L);
5795 hlt();
5796 }
5797 bind(L);
5798 }
5799 #endif
5800
5801 if (offset != 0)
5802 addptr(tmp, offset);
5803
5804 return RegisterOrConstant(tmp);
5805 }
5806
5807
5808 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
5809 int extra_slot_offset) {
5810 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
5811 int stackElementSize = Interpreter::stackElementSize;
5812 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
5813 #ifdef ASSERT
5814 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
5815 assert(offset1 - offset == stackElementSize, "correct arithmetic");
5816 #endif
5817 Register scale_reg = noreg;
5818 Address::ScaleFactor scale_factor = Address::no_scale;
5819 if (arg_slot.is_constant()) {
5820 offset += arg_slot.as_constant() * stackElementSize;
5821 } else {
5822 scale_reg = arg_slot.as_register();
5823 scale_factor = Address::times(stackElementSize);
5824 }
5825 offset += wordSize; // return PC is on stack
5826 return Address(rsp, scale_reg, scale_factor, offset);
5827 }
5828
5829
5830 void MacroAssembler::verify_oop_addr(Address addr, const char* s) {
5831 if (!VerifyOops) return;
5832
5833 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord);
5834 // Pass register number to verify_oop_subroutine
5835 const char* b = NULL;
5836 {
5837 ResourceMark rm;
5838 stringStream ss;
5839 ss.print("verify_oop_addr: %s", s);
5840 b = code_string(ss.as_string());
5841 }
5842 #ifdef _LP64
5843 push(rscratch1); // save r10, trashed by movptr()
5844 #endif
5845 push(rax); // save rax,
5846 // addr may contain rsp so we will have to adjust it based on the push
5847 // we just did (and on 64 bit we do two pushes)
5848 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which
5849 // stores rax into addr which is backwards of what was intended.
5850 if (addr.uses(rsp)) {
5851 lea(rax, addr);
5852 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord));
5853 } else {
5854 pushptr(addr);
5855 }
5856
5857 ExternalAddress buffer((address) b);
5858 // pass msg argument
5859 // avoid using pushptr, as it modifies scratch registers
5860 // and our contract is not to modify anything
5861 movptr(rax, buffer.addr());
5862 push(rax);
5863
5864 // call indirectly to solve generation ordering problem
5865 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
5866 call(rax);
5867 // Caller pops the arguments (addr, message) and restores rax, r10.
5868 }
5869
5870 void MacroAssembler::verify_tlab() {
5871 #ifdef ASSERT
5872 if (UseTLAB && VerifyOops) {
5873 Label next, ok;
5874 Register t1 = rsi;
5875 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread);
5876
5877 push(t1);
5878 NOT_LP64(push(thread_reg));
5879 NOT_LP64(get_thread(thread_reg));
5880
5881 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
5882 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
5883 jcc(Assembler::aboveEqual, next);
5884 STOP("assert(top >= start)");
5885 should_not_reach_here();
5886
5887 bind(next);
5888 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
5889 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
5890 jcc(Assembler::aboveEqual, ok);
5891 STOP("assert(top <= end)");
5892 should_not_reach_here();
5893
5894 bind(ok);
5895 NOT_LP64(pop(thread_reg));
5896 pop(t1);
5897 }
5898 #endif
5899 }
5900
5901 class ControlWord {
5902 public:
5903 int32_t _value;
5904
5905 int rounding_control() const { return (_value >> 10) & 3 ; }
5906 int precision_control() const { return (_value >> 8) & 3 ; }
5907 bool precision() const { return ((_value >> 5) & 1) != 0; }
5908 bool underflow() const { return ((_value >> 4) & 1) != 0; }
5909 bool overflow() const { return ((_value >> 3) & 1) != 0; }
5910 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
5911 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
5912 bool invalid() const { return ((_value >> 0) & 1) != 0; }
5913
5914 void print() const {
5915 // rounding control
5916 const char* rc;
5917 switch (rounding_control()) {
5918 case 0: rc = "round near"; break;
5919 case 1: rc = "round down"; break;
5920 case 2: rc = "round up "; break;
5921 case 3: rc = "chop "; break;
5922 };
5923 // precision control
5924 const char* pc;
5925 switch (precision_control()) {
5926 case 0: pc = "24 bits "; break;
5927 case 1: pc = "reserved"; break;
5928 case 2: pc = "53 bits "; break;
5929 case 3: pc = "64 bits "; break;
5930 };
5931 // flags
5932 char f[9];
5933 f[0] = ' ';
5934 f[1] = ' ';
5935 f[2] = (precision ()) ? 'P' : 'p';
5936 f[3] = (underflow ()) ? 'U' : 'u';
5937 f[4] = (overflow ()) ? 'O' : 'o';
5938 f[5] = (zero_divide ()) ? 'Z' : 'z';
5939 f[6] = (denormalized()) ? 'D' : 'd';
5940 f[7] = (invalid ()) ? 'I' : 'i';
5941 f[8] = '\x0';
5942 // output
5943 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc);
5944 }
5945
5946 };
5947
5948 class StatusWord {
5949 public:
5950 int32_t _value;
5951
5952 bool busy() const { return ((_value >> 15) & 1) != 0; }
5953 bool C3() const { return ((_value >> 14) & 1) != 0; }
5954 bool C2() const { return ((_value >> 10) & 1) != 0; }
5955 bool C1() const { return ((_value >> 9) & 1) != 0; }
5956 bool C0() const { return ((_value >> 8) & 1) != 0; }
5957 int top() const { return (_value >> 11) & 7 ; }
5958 bool error_status() const { return ((_value >> 7) & 1) != 0; }
5959 bool stack_fault() const { return ((_value >> 6) & 1) != 0; }
5960 bool precision() const { return ((_value >> 5) & 1) != 0; }
5961 bool underflow() const { return ((_value >> 4) & 1) != 0; }
5962 bool overflow() const { return ((_value >> 3) & 1) != 0; }
5963 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
5964 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
5965 bool invalid() const { return ((_value >> 0) & 1) != 0; }
5966
5967 void print() const {
5968 // condition codes
5969 char c[5];
5970 c[0] = (C3()) ? '3' : '-';
5971 c[1] = (C2()) ? '2' : '-';
5972 c[2] = (C1()) ? '1' : '-';
5973 c[3] = (C0()) ? '0' : '-';
5974 c[4] = '\x0';
5975 // flags
5976 char f[9];
5977 f[0] = (error_status()) ? 'E' : '-';
5978 f[1] = (stack_fault ()) ? 'S' : '-';
5979 f[2] = (precision ()) ? 'P' : '-';
5980 f[3] = (underflow ()) ? 'U' : '-';
5981 f[4] = (overflow ()) ? 'O' : '-';
5982 f[5] = (zero_divide ()) ? 'Z' : '-';
5983 f[6] = (denormalized()) ? 'D' : '-';
5984 f[7] = (invalid ()) ? 'I' : '-';
5985 f[8] = '\x0';
5986 // output
5987 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top());
5988 }
5989
5990 };
5991
5992 class TagWord {
5993 public:
5994 int32_t _value;
5995
5996 int tag_at(int i) const { return (_value >> (i*2)) & 3; }
5997
5998 void print() const {
5999 printf("%04x", _value & 0xFFFF);
6000 }
6001
6002 };
6003
6004 class FPU_Register {
6005 public:
6006 int32_t _m0;
6007 int32_t _m1;
6008 int16_t _ex;
6009
6010 bool is_indefinite() const {
6011 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0;
6012 }
6013
6014 void print() const {
6015 char sign = (_ex < 0) ? '-' : '+';
6016 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " ";
6017 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind);
6018 };
6019
6020 };
6021
6022 class FPU_State {
6023 public:
6024 enum {
6025 register_size = 10,
6026 number_of_registers = 8,
6027 register_mask = 7
6028 };
6029
6030 ControlWord _control_word;
6031 StatusWord _status_word;
6032 TagWord _tag_word;
6033 int32_t _error_offset;
6034 int32_t _error_selector;
6035 int32_t _data_offset;
6036 int32_t _data_selector;
6037 int8_t _register[register_size * number_of_registers];
6038
6039 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); }
6040 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; }
6041
6042 const char* tag_as_string(int tag) const {
6043 switch (tag) {
6044 case 0: return "valid";
6045 case 1: return "zero";
6046 case 2: return "special";
6047 case 3: return "empty";
6048 }
6049 ShouldNotReachHere();
6050 return NULL;
6051 }
6052
6053 void print() const {
6054 // print computation registers
6055 { int t = _status_word.top();
6056 for (int i = 0; i < number_of_registers; i++) {
6057 int j = (i - t) & register_mask;
6058 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j);
6059 st(j)->print();
6060 printf(" %s\n", tag_as_string(_tag_word.tag_at(i)));
6061 }
6062 }
6063 printf("\n");
6064 // print control registers
6065 printf("ctrl = "); _control_word.print(); printf("\n");
6066 printf("stat = "); _status_word .print(); printf("\n");
6067 printf("tags = "); _tag_word .print(); printf("\n");
6068 }
6069
6070 };
6071
6072 class Flag_Register {
6073 public:
6074 int32_t _value;
6075
6076 bool overflow() const { return ((_value >> 11) & 1) != 0; }
6077 bool direction() const { return ((_value >> 10) & 1) != 0; }
6078 bool sign() const { return ((_value >> 7) & 1) != 0; }
6079 bool zero() const { return ((_value >> 6) & 1) != 0; }
6080 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; }
6081 bool parity() const { return ((_value >> 2) & 1) != 0; }
6082 bool carry() const { return ((_value >> 0) & 1) != 0; }
6083
6084 void print() const {
6085 // flags
6086 char f[8];
6087 f[0] = (overflow ()) ? 'O' : '-';
6088 f[1] = (direction ()) ? 'D' : '-';
6089 f[2] = (sign ()) ? 'S' : '-';
6090 f[3] = (zero ()) ? 'Z' : '-';
6091 f[4] = (auxiliary_carry()) ? 'A' : '-';
6092 f[5] = (parity ()) ? 'P' : '-';
6093 f[6] = (carry ()) ? 'C' : '-';
6094 f[7] = '\x0';
6095 // output
6096 printf("%08x flags = %s", _value, f);
6097 }
6098
6099 };
6100
6101 class IU_Register {
6102 public:
6103 int32_t _value;
6104
6105 void print() const {
6106 printf("%08x %11d", _value, _value);
6107 }
6108
6109 };
6110
6111 class IU_State {
6112 public:
6113 Flag_Register _eflags;
6114 IU_Register _rdi;
6115 IU_Register _rsi;
6116 IU_Register _rbp;
6117 IU_Register _rsp;
6118 IU_Register _rbx;
6119 IU_Register _rdx;
6120 IU_Register _rcx;
6121 IU_Register _rax;
6122
6123 void print() const {
6124 // computation registers
6125 printf("rax, = "); _rax.print(); printf("\n");
6126 printf("rbx, = "); _rbx.print(); printf("\n");
6127 printf("rcx = "); _rcx.print(); printf("\n");
6128 printf("rdx = "); _rdx.print(); printf("\n");
6129 printf("rdi = "); _rdi.print(); printf("\n");
6130 printf("rsi = "); _rsi.print(); printf("\n");
6131 printf("rbp, = "); _rbp.print(); printf("\n");
6132 printf("rsp = "); _rsp.print(); printf("\n");
6133 printf("\n");
6134 // control registers
6135 printf("flgs = "); _eflags.print(); printf("\n");
6136 }
6137 };
6138
6139
6140 class CPU_State {
6141 public:
6142 FPU_State _fpu_state;
6143 IU_State _iu_state;
6144
6145 void print() const {
6146 printf("--------------------------------------------------\n");
6147 _iu_state .print();
6148 printf("\n");
6149 _fpu_state.print();
6150 printf("--------------------------------------------------\n");
6151 }
6152
6153 };
6154
6155
6156 static void _print_CPU_state(CPU_State* state) {
6157 state->print();
6158 };
6159
6160
6161 void MacroAssembler::print_CPU_state() {
6162 push_CPU_state();
6163 push(rsp); // pass CPU state
6164 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state)));
6165 addptr(rsp, wordSize); // discard argument
6166 pop_CPU_state();
6167 }
6168
6169
6170 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) {
6171 static int counter = 0;
6172 FPU_State* fs = &state->_fpu_state;
6173 counter++;
6174 // For leaf calls, only verify that the top few elements remain empty.
6175 // We only need 1 empty at the top for C2 code.
6176 if( stack_depth < 0 ) {
6177 if( fs->tag_for_st(7) != 3 ) {
6178 printf("FPR7 not empty\n");
6179 state->print();
6180 assert(false, "error");
6181 return false;
6182 }
6183 return true; // All other stack states do not matter
6184 }
6185
6186 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std,
6187 "bad FPU control word");
6188
6189 // compute stack depth
6190 int i = 0;
6191 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++;
6192 int d = i;
6193 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++;
6194 // verify findings
6195 if (i != FPU_State::number_of_registers) {
6196 // stack not contiguous
6197 printf("%s: stack not contiguous at ST%d\n", s, i);
6198 state->print();
6199 assert(false, "error");
6200 return false;
6201 }
6202 // check if computed stack depth corresponds to expected stack depth
6203 if (stack_depth < 0) {
6204 // expected stack depth is -stack_depth or less
6205 if (d > -stack_depth) {
6206 // too many elements on the stack
6207 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d);
6208 state->print();
6209 assert(false, "error");
6210 return false;
6211 }
6212 } else {
6213 // expected stack depth is stack_depth
6214 if (d != stack_depth) {
6215 // wrong stack depth
6216 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d);
6217 state->print();
6218 assert(false, "error");
6219 return false;
6220 }
6221 }
6222 // everything is cool
6223 return true;
6224 }
6225
6226
6227 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
6228 if (!VerifyFPU) return;
6229 push_CPU_state();
6230 push(rsp); // pass CPU state
6231 ExternalAddress msg((address) s);
6232 // pass message string s
6233 pushptr(msg.addr());
6234 push(stack_depth); // pass stack depth
6235 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU)));
6236 addptr(rsp, 3 * wordSize); // discard arguments
6237 // check for error
6238 { Label L;
6239 testl(rax, rax);
6240 jcc(Assembler::notZero, L);
6241 int3(); // break if error condition
6242 bind(L);
6243 }
6244 pop_CPU_state();
6245 }
6246
6247 void MacroAssembler::restore_cpu_control_state_after_jni() {
6248 // Either restore the MXCSR register after returning from the JNI Call
6249 // or verify that it wasn't changed (with -Xcheck:jni flag).
6250 if (VM_Version::supports_sse()) {
6251 if (RestoreMXCSROnJNICalls) {
6252 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
6253 } else if (CheckJNICalls) {
6254 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
6255 }
6256 }
6257 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty.
6258 vzeroupper();
6259 // Reset k1 to 0xffff.
6260 if (VM_Version::supports_evex()) {
6261 push(rcx);
6262 movl(rcx, 0xffff);
6263 kmovwl(k1, rcx);
6264 pop(rcx);
6265 }
6266
6267 #ifndef _LP64
6268 // Either restore the x87 floating pointer control word after returning
6269 // from the JNI call or verify that it wasn't changed.
6270 if (CheckJNICalls) {
6271 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
6272 }
6273 #endif // _LP64
6274 }
6275
6276 // ((OopHandle)result).resolve();
6277 void MacroAssembler::resolve_oop_handle(Register result) {
6278 // OopHandle::resolve is an indirection.
6279 movptr(result, Address(result, 0));
6280 }
6281
6282 void MacroAssembler::load_mirror(Register mirror, Register method) {
6283 // get mirror
6284 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
6285 movptr(mirror, Address(method, Method::const_offset()));
6286 movptr(mirror, Address(mirror, ConstMethod::constants_offset()));
6287 movptr(mirror, Address(mirror, ConstantPool::pool_holder_offset_in_bytes()));
6288 movptr(mirror, Address(mirror, mirror_offset));
6289 resolve_oop_handle(mirror);
6290 }
6291
6292 void MacroAssembler::load_klass(Register dst, Register src) {
6293 #ifdef _LP64
6294 if (UseCompressedClassPointers) {
6295 movl(dst, Address(src, oopDesc::klass_offset_in_bytes()));
6296 decode_klass_not_null(dst);
6297 } else
6298 #endif
6299 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
6300 }
6301
6302 void MacroAssembler::load_prototype_header(Register dst, Register src) {
6303 load_klass(dst, src);
6304 movptr(dst, Address(dst, Klass::prototype_header_offset()));
6305 }
6306
6307 void MacroAssembler::store_klass(Register dst, Register src) {
6308 #ifdef _LP64
6309 if (UseCompressedClassPointers) {
6310 encode_klass_not_null(src);
6311 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src);
6312 } else
6313 #endif
6314 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src);
6315 }
6316
6317 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators, Register dst, Address src,
6318 Register tmp1, Register thread_tmp) {
6319 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
6320 bool as_raw = (decorators & AS_RAW) != 0;
6321 if (as_raw) {
6322 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
6323 } else {
6324 bs->load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
6325 }
6326 }
6327
6328 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators, Address dst, Register src,
6329 Register tmp1, Register tmp2) {
6330 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
6331 bool as_raw = (decorators & AS_RAW) != 0;
6332 if (as_raw) {
6333 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, src, tmp1, tmp2);
6334 } else {
6335 bs->store_at(this, decorators, type, dst, src, tmp1, tmp2);
6336 }
6337 }
6338
6339 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
6340 Register thread_tmp, DecoratorSet decorators) {
6341 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp);
6342 }
6343
6344 // Doesn't do verfication, generates fixed size code
6345 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1,
6346 Register thread_tmp, DecoratorSet decorators) {
6347 access_load_at(T_OBJECT, IN_HEAP | OOP_NOT_NULL | decorators, dst, src, tmp1, thread_tmp);
6348 }
6349
6350 void MacroAssembler::store_heap_oop(Address dst, Register src, Register tmp1,
6351 Register tmp2, DecoratorSet decorators) {
6352 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2);
6353 }
6354
6355 // Used for storing NULLs.
6356 void MacroAssembler::store_heap_oop_null(Address dst) {
6357 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg);
6358 }
6359
6360 #ifdef _LP64
6361 void MacroAssembler::store_klass_gap(Register dst, Register src) {
6362 if (UseCompressedClassPointers) {
6363 // Store to klass gap in destination
6364 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src);
6365 }
6366 }
6367
6368 #ifdef ASSERT
6369 void MacroAssembler::verify_heapbase(const char* msg) {
6370 assert (UseCompressedOops, "should be compressed");
6371 assert (Universe::heap() != NULL, "java heap should be initialized");
6372 if (CheckCompressedOops) {
6373 Label ok;
6374 push(rscratch1); // cmpptr trashes rscratch1
6375 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
6376 jcc(Assembler::equal, ok);
6377 STOP(msg);
6378 bind(ok);
6379 pop(rscratch1);
6380 }
6381 }
6382 #endif
6383
6384 // Algorithm must match oop.inline.hpp encode_heap_oop.
6385 void MacroAssembler::encode_heap_oop(Register r) {
6386 #ifdef ASSERT
6387 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
6388 #endif
6389 verify_oop(r, "broken oop in encode_heap_oop");
6390 if (Universe::narrow_oop_base() == NULL) {
6391 if (Universe::narrow_oop_shift() != 0) {
6392 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6393 shrq(r, LogMinObjAlignmentInBytes);
6394 }
6395 return;
6396 }
6397 testq(r, r);
6398 cmovq(Assembler::equal, r, r12_heapbase);
6399 subq(r, r12_heapbase);
6400 shrq(r, LogMinObjAlignmentInBytes);
6401 }
6402
6403 void MacroAssembler::encode_heap_oop_not_null(Register r) {
6404 #ifdef ASSERT
6405 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
6406 if (CheckCompressedOops) {
6407 Label ok;
6408 testq(r, r);
6409 jcc(Assembler::notEqual, ok);
6410 STOP("null oop passed to encode_heap_oop_not_null");
6411 bind(ok);
6412 }
6413 #endif
6414 verify_oop(r, "broken oop in encode_heap_oop_not_null");
6415 if (Universe::narrow_oop_base() != NULL) {
6416 subq(r, r12_heapbase);
6417 }
6418 if (Universe::narrow_oop_shift() != 0) {
6419 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6420 shrq(r, LogMinObjAlignmentInBytes);
6421 }
6422 }
6423
6424 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
6425 #ifdef ASSERT
6426 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
6427 if (CheckCompressedOops) {
6428 Label ok;
6429 testq(src, src);
6430 jcc(Assembler::notEqual, ok);
6431 STOP("null oop passed to encode_heap_oop_not_null2");
6432 bind(ok);
6433 }
6434 #endif
6435 verify_oop(src, "broken oop in encode_heap_oop_not_null2");
6436 if (dst != src) {
6437 movq(dst, src);
6438 }
6439 if (Universe::narrow_oop_base() != NULL) {
6440 subq(dst, r12_heapbase);
6441 }
6442 if (Universe::narrow_oop_shift() != 0) {
6443 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6444 shrq(dst, LogMinObjAlignmentInBytes);
6445 }
6446 }
6447
6448 void MacroAssembler::decode_heap_oop(Register r) {
6449 #ifdef ASSERT
6450 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
6451 #endif
6452 if (Universe::narrow_oop_base() == NULL) {
6453 if (Universe::narrow_oop_shift() != 0) {
6454 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6455 shlq(r, LogMinObjAlignmentInBytes);
6456 }
6457 } else {
6458 Label done;
6459 shlq(r, LogMinObjAlignmentInBytes);
6460 jccb(Assembler::equal, done);
6461 addq(r, r12_heapbase);
6462 bind(done);
6463 }
6464 verify_oop(r, "broken oop in decode_heap_oop");
6465 }
6466
6467 void MacroAssembler::decode_heap_oop_not_null(Register r) {
6468 // Note: it will change flags
6469 assert (UseCompressedOops, "should only be used for compressed headers");
6470 assert (Universe::heap() != NULL, "java heap should be initialized");
6471 // Cannot assert, unverified entry point counts instructions (see .ad file)
6472 // vtableStubs also counts instructions in pd_code_size_limit.
6473 // Also do not verify_oop as this is called by verify_oop.
6474 if (Universe::narrow_oop_shift() != 0) {
6475 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6476 shlq(r, LogMinObjAlignmentInBytes);
6477 if (Universe::narrow_oop_base() != NULL) {
6478 addq(r, r12_heapbase);
6479 }
6480 } else {
6481 assert (Universe::narrow_oop_base() == NULL, "sanity");
6482 }
6483 }
6484
6485 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
6486 // Note: it will change flags
6487 assert (UseCompressedOops, "should only be used for compressed headers");
6488 assert (Universe::heap() != NULL, "java heap should be initialized");
6489 // Cannot assert, unverified entry point counts instructions (see .ad file)
6490 // vtableStubs also counts instructions in pd_code_size_limit.
6491 // Also do not verify_oop as this is called by verify_oop.
6492 if (Universe::narrow_oop_shift() != 0) {
6493 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6494 if (LogMinObjAlignmentInBytes == Address::times_8) {
6495 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0));
6496 } else {
6497 if (dst != src) {
6498 movq(dst, src);
6499 }
6500 shlq(dst, LogMinObjAlignmentInBytes);
6501 if (Universe::narrow_oop_base() != NULL) {
6502 addq(dst, r12_heapbase);
6503 }
6504 }
6505 } else {
6506 assert (Universe::narrow_oop_base() == NULL, "sanity");
6507 if (dst != src) {
6508 movq(dst, src);
6509 }
6510 }
6511 }
6512
6513 void MacroAssembler::encode_klass_not_null(Register r) {
6514 if (Universe::narrow_klass_base() != NULL) {
6515 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
6516 assert(r != r12_heapbase, "Encoding a klass in r12");
6517 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
6518 subq(r, r12_heapbase);
6519 }
6520 if (Universe::narrow_klass_shift() != 0) {
6521 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
6522 shrq(r, LogKlassAlignmentInBytes);
6523 }
6524 if (Universe::narrow_klass_base() != NULL) {
6525 reinit_heapbase();
6526 }
6527 }
6528
6529 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
6530 if (dst == src) {
6531 encode_klass_not_null(src);
6532 } else {
6533 if (Universe::narrow_klass_base() != NULL) {
6534 mov64(dst, (int64_t)Universe::narrow_klass_base());
6535 negq(dst);
6536 addq(dst, src);
6537 } else {
6538 movptr(dst, src);
6539 }
6540 if (Universe::narrow_klass_shift() != 0) {
6541 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
6542 shrq(dst, LogKlassAlignmentInBytes);
6543 }
6544 }
6545 }
6546
6547 // Function instr_size_for_decode_klass_not_null() counts the instructions
6548 // generated by decode_klass_not_null(register r) and reinit_heapbase(),
6549 // when (Universe::heap() != NULL). Hence, if the instructions they
6550 // generate change, then this method needs to be updated.
6551 int MacroAssembler::instr_size_for_decode_klass_not_null() {
6552 assert (UseCompressedClassPointers, "only for compressed klass ptrs");
6553 if (Universe::narrow_klass_base() != NULL) {
6554 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()).
6555 return (Universe::narrow_klass_shift() == 0 ? 20 : 24);
6556 } else {
6557 // longest load decode klass function, mov64, leaq
6558 return 16;
6559 }
6560 }
6561
6562 // !!! If the instructions that get generated here change then function
6563 // instr_size_for_decode_klass_not_null() needs to get updated.
6564 void MacroAssembler::decode_klass_not_null(Register r) {
6565 // Note: it will change flags
6566 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6567 assert(r != r12_heapbase, "Decoding a klass in r12");
6568 // Cannot assert, unverified entry point counts instructions (see .ad file)
6569 // vtableStubs also counts instructions in pd_code_size_limit.
6570 // Also do not verify_oop as this is called by verify_oop.
6571 if (Universe::narrow_klass_shift() != 0) {
6572 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
6573 shlq(r, LogKlassAlignmentInBytes);
6574 }
6575 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
6576 if (Universe::narrow_klass_base() != NULL) {
6577 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
6578 addq(r, r12_heapbase);
6579 reinit_heapbase();
6580 }
6581 }
6582
6583 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
6584 // Note: it will change flags
6585 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6586 if (dst == src) {
6587 decode_klass_not_null(dst);
6588 } else {
6589 // Cannot assert, unverified entry point counts instructions (see .ad file)
6590 // vtableStubs also counts instructions in pd_code_size_limit.
6591 // Also do not verify_oop as this is called by verify_oop.
6592 mov64(dst, (int64_t)Universe::narrow_klass_base());
6593 if (Universe::narrow_klass_shift() != 0) {
6594 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
6595 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?");
6596 leaq(dst, Address(dst, src, Address::times_8, 0));
6597 } else {
6598 addq(dst, src);
6599 }
6600 }
6601 }
6602
6603 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
6604 assert (UseCompressedOops, "should only be used for compressed headers");
6605 assert (Universe::heap() != NULL, "java heap should be initialized");
6606 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6607 int oop_index = oop_recorder()->find_index(obj);
6608 RelocationHolder rspec = oop_Relocation::spec(oop_index);
6609 mov_narrow_oop(dst, oop_index, rspec);
6610 }
6611
6612 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) {
6613 assert (UseCompressedOops, "should only be used for compressed headers");
6614 assert (Universe::heap() != NULL, "java heap should be initialized");
6615 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6616 int oop_index = oop_recorder()->find_index(obj);
6617 RelocationHolder rspec = oop_Relocation::spec(oop_index);
6618 mov_narrow_oop(dst, oop_index, rspec);
6619 }
6620
6621 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
6622 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6623 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6624 int klass_index = oop_recorder()->find_index(k);
6625 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6626 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
6627 }
6628
6629 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) {
6630 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6631 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6632 int klass_index = oop_recorder()->find_index(k);
6633 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6634 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
6635 }
6636
6637 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) {
6638 assert (UseCompressedOops, "should only be used for compressed headers");
6639 assert (Universe::heap() != NULL, "java heap should be initialized");
6640 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6641 int oop_index = oop_recorder()->find_index(obj);
6642 RelocationHolder rspec = oop_Relocation::spec(oop_index);
6643 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
6644 }
6645
6646 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) {
6647 assert (UseCompressedOops, "should only be used for compressed headers");
6648 assert (Universe::heap() != NULL, "java heap should be initialized");
6649 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6650 int oop_index = oop_recorder()->find_index(obj);
6651 RelocationHolder rspec = oop_Relocation::spec(oop_index);
6652 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
6653 }
6654
6655 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) {
6656 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6657 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6658 int klass_index = oop_recorder()->find_index(k);
6659 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6660 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
6661 }
6662
6663 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) {
6664 assert (UseCompressedClassPointers, "should only be used for compressed headers");
6665 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
6666 int klass_index = oop_recorder()->find_index(k);
6667 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
6668 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
6669 }
6670
6671 void MacroAssembler::reinit_heapbase() {
6672 if (UseCompressedOops || UseCompressedClassPointers) {
6673 if (Universe::heap() != NULL) {
6674 if (Universe::narrow_oop_base() == NULL) {
6675 MacroAssembler::xorptr(r12_heapbase, r12_heapbase);
6676 } else {
6677 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base());
6678 }
6679 } else {
6680 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
6681 }
6682 }
6683 }
6684
6685 #endif // _LP64
6686
6687 // C2 compiled method's prolog code.
6688 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) {
6689
6690 // WARNING: Initial instruction MUST be 5 bytes or longer so that
6691 // NativeJump::patch_verified_entry will be able to patch out the entry
6692 // code safely. The push to verify stack depth is ok at 5 bytes,
6693 // the frame allocation can be either 3 or 6 bytes. So if we don't do
6694 // stack bang then we must use the 6 byte frame allocation even if
6695 // we have no frame. :-(
6696 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect");
6697
6698 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
6699 // Remove word for return addr
6700 framesize -= wordSize;
6701 stack_bang_size -= wordSize;
6702
6703 // Calls to C2R adapters often do not accept exceptional returns.
6704 // We require that their callers must bang for them. But be careful, because
6705 // some VM calls (such as call site linkage) can use several kilobytes of
6706 // stack. But the stack safety zone should account for that.
6707 // See bugs 4446381, 4468289, 4497237.
6708 if (stack_bang_size > 0) {
6709 generate_stack_overflow_check(stack_bang_size);
6710
6711 // We always push rbp, so that on return to interpreter rbp, will be
6712 // restored correctly and we can correct the stack.
6713 push(rbp);
6714 // Save caller's stack pointer into RBP if the frame pointer is preserved.
6715 if (PreserveFramePointer) {
6716 mov(rbp, rsp);
6717 }
6718 // Remove word for ebp
6719 framesize -= wordSize;
6720
6721 // Create frame
6722 if (framesize) {
6723 subptr(rsp, framesize);
6724 }
6725 } else {
6726 // Create frame (force generation of a 4 byte immediate value)
6727 subptr_imm32(rsp, framesize);
6728
6729 // Save RBP register now.
6730 framesize -= wordSize;
6731 movptr(Address(rsp, framesize), rbp);
6732 // Save caller's stack pointer into RBP if the frame pointer is preserved.
6733 if (PreserveFramePointer) {
6734 movptr(rbp, rsp);
6735 if (framesize > 0) {
6736 addptr(rbp, framesize);
6737 }
6738 }
6739 }
6740
6741 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
6742 framesize -= wordSize;
6743 movptr(Address(rsp, framesize), (int32_t)0xbadb100d);
6744 }
6745
6746 #ifndef _LP64
6747 // If method sets FPU control word do it now
6748 if (fp_mode_24b) {
6749 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
6750 }
6751 if (UseSSE >= 2 && VerifyFPU) {
6752 verify_FPU(0, "FPU stack must be clean on entry");
6753 }
6754 #endif
6755
6756 #ifdef ASSERT
6757 if (VerifyStackAtCalls) {
6758 Label L;
6759 push(rax);
6760 mov(rax, rsp);
6761 andptr(rax, StackAlignmentInBytes-1);
6762 cmpptr(rax, StackAlignmentInBytes-wordSize);
6763 pop(rax);
6764 jcc(Assembler::equal, L);
6765 STOP("Stack is not properly aligned!");
6766 bind(L);
6767 }
6768 #endif
6769
6770 }
6771
6772 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp, bool is_large) {
6773 // cnt - number of qwords (8-byte words).
6774 // base - start address, qword aligned.
6775 // is_large - if optimizers know cnt is larger than InitArrayShortSize
6776 assert(base==rdi, "base register must be edi for rep stos");
6777 assert(tmp==rax, "tmp register must be eax for rep stos");
6778 assert(cnt==rcx, "cnt register must be ecx for rep stos");
6779 assert(InitArrayShortSize % BytesPerLong == 0,
6780 "InitArrayShortSize should be the multiple of BytesPerLong");
6781
6782 Label DONE;
6783
6784 xorptr(tmp, tmp);
6785
6786 if (!is_large) {
6787 Label LOOP, LONG;
6788 cmpptr(cnt, InitArrayShortSize/BytesPerLong);
6789 jccb(Assembler::greater, LONG);
6790
6791 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
6792
6793 decrement(cnt);
6794 jccb(Assembler::negative, DONE); // Zero length
6795
6796 // Use individual pointer-sized stores for small counts:
6797 BIND(LOOP);
6798 movptr(Address(base, cnt, Address::times_ptr), tmp);
6799 decrement(cnt);
6800 jccb(Assembler::greaterEqual, LOOP);
6801 jmpb(DONE);
6802
6803 BIND(LONG);
6804 }
6805
6806 // Use longer rep-prefixed ops for non-small counts:
6807 if (UseFastStosb) {
6808 shlptr(cnt, 3); // convert to number of bytes
6809 rep_stosb();
6810 } else {
6811 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
6812 rep_stos();
6813 }
6814
6815 BIND(DONE);
6816 }
6817
6818 #ifdef COMPILER2
6819
6820 // IndexOf for constant substrings with size >= 8 chars
6821 // which don't need to be loaded through stack.
6822 void MacroAssembler::string_indexofC8(Register str1, Register str2,
6823 Register cnt1, Register cnt2,
6824 int int_cnt2, Register result,
6825 XMMRegister vec, Register tmp,
6826 int ae) {
6827 ShortBranchVerifier sbv(this);
6828 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
6829 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
6830
6831 // This method uses the pcmpestri instruction with bound registers
6832 // inputs:
6833 // xmm - substring
6834 // rax - substring length (elements count)
6835 // mem - scanned string
6836 // rdx - string length (elements count)
6837 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6838 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
6839 // outputs:
6840 // rcx - matched index in string
6841 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6842 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
6843 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
6844 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
6845 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
6846
6847 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR,
6848 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR,
6849 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE;
6850
6851 // Note, inline_string_indexOf() generates checks:
6852 // if (substr.count > string.count) return -1;
6853 // if (substr.count == 0) return 0;
6854 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars");
6855
6856 // Load substring.
6857 if (ae == StrIntrinsicNode::UL) {
6858 pmovzxbw(vec, Address(str2, 0));
6859 } else {
6860 movdqu(vec, Address(str2, 0));
6861 }
6862 movl(cnt2, int_cnt2);
6863 movptr(result, str1); // string addr
6864
6865 if (int_cnt2 > stride) {
6866 jmpb(SCAN_TO_SUBSTR);
6867
6868 // Reload substr for rescan, this code
6869 // is executed only for large substrings (> 8 chars)
6870 bind(RELOAD_SUBSTR);
6871 if (ae == StrIntrinsicNode::UL) {
6872 pmovzxbw(vec, Address(str2, 0));
6873 } else {
6874 movdqu(vec, Address(str2, 0));
6875 }
6876 negptr(cnt2); // Jumped here with negative cnt2, convert to positive
6877
6878 bind(RELOAD_STR);
6879 // We came here after the beginning of the substring was
6880 // matched but the rest of it was not so we need to search
6881 // again. Start from the next element after the previous match.
6882
6883 // cnt2 is number of substring reminding elements and
6884 // cnt1 is number of string reminding elements when cmp failed.
6885 // Restored cnt1 = cnt1 - cnt2 + int_cnt2
6886 subl(cnt1, cnt2);
6887 addl(cnt1, int_cnt2);
6888 movl(cnt2, int_cnt2); // Now restore cnt2
6889
6890 decrementl(cnt1); // Shift to next element
6891 cmpl(cnt1, cnt2);
6892 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6893
6894 addptr(result, (1<<scale1));
6895
6896 } // (int_cnt2 > 8)
6897
6898 // Scan string for start of substr in 16-byte vectors
6899 bind(SCAN_TO_SUBSTR);
6900 pcmpestri(vec, Address(result, 0), mode);
6901 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
6902 subl(cnt1, stride);
6903 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6904 cmpl(cnt1, cnt2);
6905 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6906 addptr(result, 16);
6907 jmpb(SCAN_TO_SUBSTR);
6908
6909 // Found a potential substr
6910 bind(FOUND_CANDIDATE);
6911 // Matched whole vector if first element matched (tmp(rcx) == 0).
6912 if (int_cnt2 == stride) {
6913 jccb(Assembler::overflow, RET_FOUND); // OF == 1
6914 } else { // int_cnt2 > 8
6915 jccb(Assembler::overflow, FOUND_SUBSTR);
6916 }
6917 // After pcmpestri tmp(rcx) contains matched element index
6918 // Compute start addr of substr
6919 lea(result, Address(result, tmp, scale1));
6920
6921 // Make sure string is still long enough
6922 subl(cnt1, tmp);
6923 cmpl(cnt1, cnt2);
6924 if (int_cnt2 == stride) {
6925 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6926 } else { // int_cnt2 > 8
6927 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD);
6928 }
6929 // Left less then substring.
6930
6931 bind(RET_NOT_FOUND);
6932 movl(result, -1);
6933 jmp(EXIT);
6934
6935 if (int_cnt2 > stride) {
6936 // This code is optimized for the case when whole substring
6937 // is matched if its head is matched.
6938 bind(MATCH_SUBSTR_HEAD);
6939 pcmpestri(vec, Address(result, 0), mode);
6940 // Reload only string if does not match
6941 jcc(Assembler::noOverflow, RELOAD_STR); // OF == 0
6942
6943 Label CONT_SCAN_SUBSTR;
6944 // Compare the rest of substring (> 8 chars).
6945 bind(FOUND_SUBSTR);
6946 // First 8 chars are already matched.
6947 negptr(cnt2);
6948 addptr(cnt2, stride);
6949
6950 bind(SCAN_SUBSTR);
6951 subl(cnt1, stride);
6952 cmpl(cnt2, -stride); // Do not read beyond substring
6953 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR);
6954 // Back-up strings to avoid reading beyond substring:
6955 // cnt1 = cnt1 - cnt2 + 8
6956 addl(cnt1, cnt2); // cnt2 is negative
6957 addl(cnt1, stride);
6958 movl(cnt2, stride); negptr(cnt2);
6959 bind(CONT_SCAN_SUBSTR);
6960 if (int_cnt2 < (int)G) {
6961 int tail_off1 = int_cnt2<<scale1;
6962 int tail_off2 = int_cnt2<<scale2;
6963 if (ae == StrIntrinsicNode::UL) {
6964 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2));
6965 } else {
6966 movdqu(vec, Address(str2, cnt2, scale2, tail_off2));
6967 }
6968 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode);
6969 } else {
6970 // calculate index in register to avoid integer overflow (int_cnt2*2)
6971 movl(tmp, int_cnt2);
6972 addptr(tmp, cnt2);
6973 if (ae == StrIntrinsicNode::UL) {
6974 pmovzxbw(vec, Address(str2, tmp, scale2, 0));
6975 } else {
6976 movdqu(vec, Address(str2, tmp, scale2, 0));
6977 }
6978 pcmpestri(vec, Address(result, tmp, scale1, 0), mode);
6979 }
6980 // Need to reload strings pointers if not matched whole vector
6981 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6982 addptr(cnt2, stride);
6983 jcc(Assembler::negative, SCAN_SUBSTR);
6984 // Fall through if found full substring
6985
6986 } // (int_cnt2 > 8)
6987
6988 bind(RET_FOUND);
6989 // Found result if we matched full small substring.
6990 // Compute substr offset
6991 subptr(result, str1);
6992 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6993 shrl(result, 1); // index
6994 }
6995 bind(EXIT);
6996
6997 } // string_indexofC8
6998
6999 // Small strings are loaded through stack if they cross page boundary.
7000 void MacroAssembler::string_indexof(Register str1, Register str2,
7001 Register cnt1, Register cnt2,
7002 int int_cnt2, Register result,
7003 XMMRegister vec, Register tmp,
7004 int ae) {
7005 ShortBranchVerifier sbv(this);
7006 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
7007 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
7008
7009 //
7010 // int_cnt2 is length of small (< 8 chars) constant substring
7011 // or (-1) for non constant substring in which case its length
7012 // is in cnt2 register.
7013 //
7014 // Note, inline_string_indexOf() generates checks:
7015 // if (substr.count > string.count) return -1;
7016 // if (substr.count == 0) return 0;
7017 //
7018 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
7019 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0");
7020 // This method uses the pcmpestri instruction with bound registers
7021 // inputs:
7022 // xmm - substring
7023 // rax - substring length (elements count)
7024 // mem - scanned string
7025 // rdx - string length (elements count)
7026 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
7027 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
7028 // outputs:
7029 // rcx - matched index in string
7030 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
7031 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
7032 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
7033 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
7034
7035 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR,
7036 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR,
7037 FOUND_CANDIDATE;
7038
7039 { //========================================================
7040 // We don't know where these strings are located
7041 // and we can't read beyond them. Load them through stack.
7042 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR;
7043
7044 movptr(tmp, rsp); // save old SP
7045
7046 if (int_cnt2 > 0) { // small (< 8 chars) constant substring
7047 if (int_cnt2 == (1>>scale2)) { // One byte
7048 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding");
7049 load_unsigned_byte(result, Address(str2, 0));
7050 movdl(vec, result); // move 32 bits
7051 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes
7052 // Not enough header space in 32-bit VM: 12+3 = 15.
7053 movl(result, Address(str2, -1));
7054 shrl(result, 8);
7055 movdl(vec, result); // move 32 bits
7056 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char
7057 load_unsigned_short(result, Address(str2, 0));
7058 movdl(vec, result); // move 32 bits
7059 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars
7060 movdl(vec, Address(str2, 0)); // move 32 bits
7061 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars
7062 movq(vec, Address(str2, 0)); // move 64 bits
7063 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7})
7064 // Array header size is 12 bytes in 32-bit VM
7065 // + 6 bytes for 3 chars == 18 bytes,
7066 // enough space to load vec and shift.
7067 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity");
7068 if (ae == StrIntrinsicNode::UL) {
7069 int tail_off = int_cnt2-8;
7070 pmovzxbw(vec, Address(str2, tail_off));
7071 psrldq(vec, -2*tail_off);
7072 }
7073 else {
7074 int tail_off = int_cnt2*(1<<scale2);
7075 movdqu(vec, Address(str2, tail_off-16));
7076 psrldq(vec, 16-tail_off);
7077 }
7078 }
7079 } else { // not constant substring
7080 cmpl(cnt2, stride);
7081 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough
7082
7083 // We can read beyond string if srt+16 does not cross page boundary
7084 // since heaps are aligned and mapped by pages.
7085 assert(os::vm_page_size() < (int)G, "default page should be small");
7086 movl(result, str2); // We need only low 32 bits
7087 andl(result, (os::vm_page_size()-1));
7088 cmpl(result, (os::vm_page_size()-16));
7089 jccb(Assembler::belowEqual, CHECK_STR);
7090
7091 // Move small strings to stack to allow load 16 bytes into vec.
7092 subptr(rsp, 16);
7093 int stk_offset = wordSize-(1<<scale2);
7094 push(cnt2);
7095
7096 bind(COPY_SUBSTR);
7097 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) {
7098 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1));
7099 movb(Address(rsp, cnt2, scale2, stk_offset), result);
7100 } else if (ae == StrIntrinsicNode::UU) {
7101 load_unsigned_short(result, Address(str2, cnt2, scale2, -2));
7102 movw(Address(rsp, cnt2, scale2, stk_offset), result);
7103 }
7104 decrement(cnt2);
7105 jccb(Assembler::notZero, COPY_SUBSTR);
7106
7107 pop(cnt2);
7108 movptr(str2, rsp); // New substring address
7109 } // non constant
7110
7111 bind(CHECK_STR);
7112 cmpl(cnt1, stride);
7113 jccb(Assembler::aboveEqual, BIG_STRINGS);
7114
7115 // Check cross page boundary.
7116 movl(result, str1); // We need only low 32 bits
7117 andl(result, (os::vm_page_size()-1));
7118 cmpl(result, (os::vm_page_size()-16));
7119 jccb(Assembler::belowEqual, BIG_STRINGS);
7120
7121 subptr(rsp, 16);
7122 int stk_offset = -(1<<scale1);
7123 if (int_cnt2 < 0) { // not constant
7124 push(cnt2);
7125 stk_offset += wordSize;
7126 }
7127 movl(cnt2, cnt1);
7128
7129 bind(COPY_STR);
7130 if (ae == StrIntrinsicNode::LL) {
7131 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1));
7132 movb(Address(rsp, cnt2, scale1, stk_offset), result);
7133 } else {
7134 load_unsigned_short(result, Address(str1, cnt2, scale1, -2));
7135 movw(Address(rsp, cnt2, scale1, stk_offset), result);
7136 }
7137 decrement(cnt2);
7138 jccb(Assembler::notZero, COPY_STR);
7139
7140 if (int_cnt2 < 0) { // not constant
7141 pop(cnt2);
7142 }
7143 movptr(str1, rsp); // New string address
7144
7145 bind(BIG_STRINGS);
7146 // Load substring.
7147 if (int_cnt2 < 0) { // -1
7148 if (ae == StrIntrinsicNode::UL) {
7149 pmovzxbw(vec, Address(str2, 0));
7150 } else {
7151 movdqu(vec, Address(str2, 0));
7152 }
7153 push(cnt2); // substr count
7154 push(str2); // substr addr
7155 push(str1); // string addr
7156 } else {
7157 // Small (< 8 chars) constant substrings are loaded already.
7158 movl(cnt2, int_cnt2);
7159 }
7160 push(tmp); // original SP
7161
7162 } // Finished loading
7163
7164 //========================================================
7165 // Start search
7166 //
7167
7168 movptr(result, str1); // string addr
7169
7170 if (int_cnt2 < 0) { // Only for non constant substring
7171 jmpb(SCAN_TO_SUBSTR);
7172
7173 // SP saved at sp+0
7174 // String saved at sp+1*wordSize
7175 // Substr saved at sp+2*wordSize
7176 // Substr count saved at sp+3*wordSize
7177
7178 // Reload substr for rescan, this code
7179 // is executed only for large substrings (> 8 chars)
7180 bind(RELOAD_SUBSTR);
7181 movptr(str2, Address(rsp, 2*wordSize));
7182 movl(cnt2, Address(rsp, 3*wordSize));
7183 if (ae == StrIntrinsicNode::UL) {
7184 pmovzxbw(vec, Address(str2, 0));
7185 } else {
7186 movdqu(vec, Address(str2, 0));
7187 }
7188 // We came here after the beginning of the substring was
7189 // matched but the rest of it was not so we need to search
7190 // again. Start from the next element after the previous match.
7191 subptr(str1, result); // Restore counter
7192 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
7193 shrl(str1, 1);
7194 }
7195 addl(cnt1, str1);
7196 decrementl(cnt1); // Shift to next element
7197 cmpl(cnt1, cnt2);
7198 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7199
7200 addptr(result, (1<<scale1));
7201 } // non constant
7202
7203 // Scan string for start of substr in 16-byte vectors
7204 bind(SCAN_TO_SUBSTR);
7205 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
7206 pcmpestri(vec, Address(result, 0), mode);
7207 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
7208 subl(cnt1, stride);
7209 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
7210 cmpl(cnt1, cnt2);
7211 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7212 addptr(result, 16);
7213
7214 bind(ADJUST_STR);
7215 cmpl(cnt1, stride); // Do not read beyond string
7216 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
7217 // Back-up string to avoid reading beyond string.
7218 lea(result, Address(result, cnt1, scale1, -16));
7219 movl(cnt1, stride);
7220 jmpb(SCAN_TO_SUBSTR);
7221
7222 // Found a potential substr
7223 bind(FOUND_CANDIDATE);
7224 // After pcmpestri tmp(rcx) contains matched element index
7225
7226 // Make sure string is still long enough
7227 subl(cnt1, tmp);
7228 cmpl(cnt1, cnt2);
7229 jccb(Assembler::greaterEqual, FOUND_SUBSTR);
7230 // Left less then substring.
7231
7232 bind(RET_NOT_FOUND);
7233 movl(result, -1);
7234 jmpb(CLEANUP);
7235
7236 bind(FOUND_SUBSTR);
7237 // Compute start addr of substr
7238 lea(result, Address(result, tmp, scale1));
7239 if (int_cnt2 > 0) { // Constant substring
7240 // Repeat search for small substring (< 8 chars)
7241 // from new point without reloading substring.
7242 // Have to check that we don't read beyond string.
7243 cmpl(tmp, stride-int_cnt2);
7244 jccb(Assembler::greater, ADJUST_STR);
7245 // Fall through if matched whole substring.
7246 } else { // non constant
7247 assert(int_cnt2 == -1, "should be != 0");
7248
7249 addl(tmp, cnt2);
7250 // Found result if we matched whole substring.
7251 cmpl(tmp, stride);
7252 jccb(Assembler::lessEqual, RET_FOUND);
7253
7254 // Repeat search for small substring (<= 8 chars)
7255 // from new point 'str1' without reloading substring.
7256 cmpl(cnt2, stride);
7257 // Have to check that we don't read beyond string.
7258 jccb(Assembler::lessEqual, ADJUST_STR);
7259
7260 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG;
7261 // Compare the rest of substring (> 8 chars).
7262 movptr(str1, result);
7263
7264 cmpl(tmp, cnt2);
7265 // First 8 chars are already matched.
7266 jccb(Assembler::equal, CHECK_NEXT);
7267
7268 bind(SCAN_SUBSTR);
7269 pcmpestri(vec, Address(str1, 0), mode);
7270 // Need to reload strings pointers if not matched whole vector
7271 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
7272
7273 bind(CHECK_NEXT);
7274 subl(cnt2, stride);
7275 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring
7276 addptr(str1, 16);
7277 if (ae == StrIntrinsicNode::UL) {
7278 addptr(str2, 8);
7279 } else {
7280 addptr(str2, 16);
7281 }
7282 subl(cnt1, stride);
7283 cmpl(cnt2, stride); // Do not read beyond substring
7284 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR);
7285 // Back-up strings to avoid reading beyond substring.
7286
7287 if (ae == StrIntrinsicNode::UL) {
7288 lea(str2, Address(str2, cnt2, scale2, -8));
7289 lea(str1, Address(str1, cnt2, scale1, -16));
7290 } else {
7291 lea(str2, Address(str2, cnt2, scale2, -16));
7292 lea(str1, Address(str1, cnt2, scale1, -16));
7293 }
7294 subl(cnt1, cnt2);
7295 movl(cnt2, stride);
7296 addl(cnt1, stride);
7297 bind(CONT_SCAN_SUBSTR);
7298 if (ae == StrIntrinsicNode::UL) {
7299 pmovzxbw(vec, Address(str2, 0));
7300 } else {
7301 movdqu(vec, Address(str2, 0));
7302 }
7303 jmp(SCAN_SUBSTR);
7304
7305 bind(RET_FOUND_LONG);
7306 movptr(str1, Address(rsp, wordSize));
7307 } // non constant
7308
7309 bind(RET_FOUND);
7310 // Compute substr offset
7311 subptr(result, str1);
7312 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
7313 shrl(result, 1); // index
7314 }
7315 bind(CLEANUP);
7316 pop(rsp); // restore SP
7317
7318 } // string_indexof
7319
7320 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result,
7321 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) {
7322 ShortBranchVerifier sbv(this);
7323 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
7324
7325 int stride = 8;
7326
7327 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP,
7328 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP,
7329 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT,
7330 FOUND_SEQ_CHAR, DONE_LABEL;
7331
7332 movptr(result, str1);
7333 if (UseAVX >= 2) {
7334 cmpl(cnt1, stride);
7335 jcc(Assembler::less, SCAN_TO_CHAR_LOOP);
7336 cmpl(cnt1, 2*stride);
7337 jcc(Assembler::less, SCAN_TO_8_CHAR_INIT);
7338 movdl(vec1, ch);
7339 vpbroadcastw(vec1, vec1);
7340 vpxor(vec2, vec2);
7341 movl(tmp, cnt1);
7342 andl(tmp, 0xFFFFFFF0); //vector count (in chars)
7343 andl(cnt1,0x0000000F); //tail count (in chars)
7344
7345 bind(SCAN_TO_16_CHAR_LOOP);
7346 vmovdqu(vec3, Address(result, 0));
7347 vpcmpeqw(vec3, vec3, vec1, 1);
7348 vptest(vec2, vec3);
7349 jcc(Assembler::carryClear, FOUND_CHAR);
7350 addptr(result, 32);
7351 subl(tmp, 2*stride);
7352 jccb(Assembler::notZero, SCAN_TO_16_CHAR_LOOP);
7353 jmp(SCAN_TO_8_CHAR);
7354 bind(SCAN_TO_8_CHAR_INIT);
7355 movdl(vec1, ch);
7356 pshuflw(vec1, vec1, 0x00);
7357 pshufd(vec1, vec1, 0);
7358 pxor(vec2, vec2);
7359 }
7360 bind(SCAN_TO_8_CHAR);
7361 cmpl(cnt1, stride);
7362 if (UseAVX >= 2) {
7363 jcc(Assembler::less, SCAN_TO_CHAR);
7364 } else {
7365 jcc(Assembler::less, SCAN_TO_CHAR_LOOP);
7366 movdl(vec1, ch);
7367 pshuflw(vec1, vec1, 0x00);
7368 pshufd(vec1, vec1, 0);
7369 pxor(vec2, vec2);
7370 }
7371 movl(tmp, cnt1);
7372 andl(tmp, 0xFFFFFFF8); //vector count (in chars)
7373 andl(cnt1,0x00000007); //tail count (in chars)
7374
7375 bind(SCAN_TO_8_CHAR_LOOP);
7376 movdqu(vec3, Address(result, 0));
7377 pcmpeqw(vec3, vec1);
7378 ptest(vec2, vec3);
7379 jcc(Assembler::carryClear, FOUND_CHAR);
7380 addptr(result, 16);
7381 subl(tmp, stride);
7382 jccb(Assembler::notZero, SCAN_TO_8_CHAR_LOOP);
7383 bind(SCAN_TO_CHAR);
7384 testl(cnt1, cnt1);
7385 jcc(Assembler::zero, RET_NOT_FOUND);
7386 bind(SCAN_TO_CHAR_LOOP);
7387 load_unsigned_short(tmp, Address(result, 0));
7388 cmpl(ch, tmp);
7389 jccb(Assembler::equal, FOUND_SEQ_CHAR);
7390 addptr(result, 2);
7391 subl(cnt1, 1);
7392 jccb(Assembler::zero, RET_NOT_FOUND);
7393 jmp(SCAN_TO_CHAR_LOOP);
7394
7395 bind(RET_NOT_FOUND);
7396 movl(result, -1);
7397 jmpb(DONE_LABEL);
7398
7399 bind(FOUND_CHAR);
7400 if (UseAVX >= 2) {
7401 vpmovmskb(tmp, vec3);
7402 } else {
7403 pmovmskb(tmp, vec3);
7404 }
7405 bsfl(ch, tmp);
7406 addl(result, ch);
7407
7408 bind(FOUND_SEQ_CHAR);
7409 subptr(result, str1);
7410 shrl(result, 1);
7411
7412 bind(DONE_LABEL);
7413 } // string_indexof_char
7414
7415 // helper function for string_compare
7416 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2,
7417 Address::ScaleFactor scale, Address::ScaleFactor scale1,
7418 Address::ScaleFactor scale2, Register index, int ae) {
7419 if (ae == StrIntrinsicNode::LL) {
7420 load_unsigned_byte(elem1, Address(str1, index, scale, 0));
7421 load_unsigned_byte(elem2, Address(str2, index, scale, 0));
7422 } else if (ae == StrIntrinsicNode::UU) {
7423 load_unsigned_short(elem1, Address(str1, index, scale, 0));
7424 load_unsigned_short(elem2, Address(str2, index, scale, 0));
7425 } else {
7426 load_unsigned_byte(elem1, Address(str1, index, scale1, 0));
7427 load_unsigned_short(elem2, Address(str2, index, scale2, 0));
7428 }
7429 }
7430
7431 // Compare strings, used for char[] and byte[].
7432 void MacroAssembler::string_compare(Register str1, Register str2,
7433 Register cnt1, Register cnt2, Register result,
7434 XMMRegister vec1, int ae) {
7435 ShortBranchVerifier sbv(this);
7436 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
7437 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3
7438 int stride, stride2, adr_stride, adr_stride1, adr_stride2;
7439 int stride2x2 = 0x40;
7440 Address::ScaleFactor scale = Address::no_scale;
7441 Address::ScaleFactor scale1 = Address::no_scale;
7442 Address::ScaleFactor scale2 = Address::no_scale;
7443
7444 if (ae != StrIntrinsicNode::LL) {
7445 stride2x2 = 0x20;
7446 }
7447
7448 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
7449 shrl(cnt2, 1);
7450 }
7451 // Compute the minimum of the string lengths and the
7452 // difference of the string lengths (stack).
7453 // Do the conditional move stuff
7454 movl(result, cnt1);
7455 subl(cnt1, cnt2);
7456 push(cnt1);
7457 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2)
7458
7459 // Is the minimum length zero?
7460 testl(cnt2, cnt2);
7461 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
7462 if (ae == StrIntrinsicNode::LL) {
7463 // Load first bytes
7464 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0]
7465 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0]
7466 } else if (ae == StrIntrinsicNode::UU) {
7467 // Load first characters
7468 load_unsigned_short(result, Address(str1, 0));
7469 load_unsigned_short(cnt1, Address(str2, 0));
7470 } else {
7471 load_unsigned_byte(result, Address(str1, 0));
7472 load_unsigned_short(cnt1, Address(str2, 0));
7473 }
7474 subl(result, cnt1);
7475 jcc(Assembler::notZero, POP_LABEL);
7476
7477 if (ae == StrIntrinsicNode::UU) {
7478 // Divide length by 2 to get number of chars
7479 shrl(cnt2, 1);
7480 }
7481 cmpl(cnt2, 1);
7482 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
7483
7484 // Check if the strings start at the same location and setup scale and stride
7485 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7486 cmpptr(str1, str2);
7487 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
7488 if (ae == StrIntrinsicNode::LL) {
7489 scale = Address::times_1;
7490 stride = 16;
7491 } else {
7492 scale = Address::times_2;
7493 stride = 8;
7494 }
7495 } else {
7496 scale1 = Address::times_1;
7497 scale2 = Address::times_2;
7498 // scale not used
7499 stride = 8;
7500 }
7501
7502 if (UseAVX >= 2 && UseSSE42Intrinsics) {
7503 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
7504 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
7505 Label COMPARE_WIDE_VECTORS_LOOP_AVX2;
7506 Label COMPARE_TAIL_LONG;
7507 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3
7508
7509 int pcmpmask = 0x19;
7510 if (ae == StrIntrinsicNode::LL) {
7511 pcmpmask &= ~0x01;
7512 }
7513
7514 // Setup to compare 16-chars (32-bytes) vectors,
7515 // start from first character again because it has aligned address.
7516 if (ae == StrIntrinsicNode::LL) {
7517 stride2 = 32;
7518 } else {
7519 stride2 = 16;
7520 }
7521 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7522 adr_stride = stride << scale;
7523 } else {
7524 adr_stride1 = 8; //stride << scale1;
7525 adr_stride2 = 16; //stride << scale2;
7526 }
7527
7528 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
7529 // rax and rdx are used by pcmpestri as elements counters
7530 movl(result, cnt2);
7531 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count
7532 jcc(Assembler::zero, COMPARE_TAIL_LONG);
7533
7534 // fast path : compare first 2 8-char vectors.
7535 bind(COMPARE_16_CHARS);
7536 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7537 movdqu(vec1, Address(str1, 0));
7538 } else {
7539 pmovzxbw(vec1, Address(str1, 0));
7540 }
7541 pcmpestri(vec1, Address(str2, 0), pcmpmask);
7542 jccb(Assembler::below, COMPARE_INDEX_CHAR);
7543
7544 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7545 movdqu(vec1, Address(str1, adr_stride));
7546 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
7547 } else {
7548 pmovzxbw(vec1, Address(str1, adr_stride1));
7549 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask);
7550 }
7551 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
7552 addl(cnt1, stride);
7553
7554 // Compare the characters at index in cnt1
7555 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character
7556 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
7557 subl(result, cnt2);
7558 jmp(POP_LABEL);
7559
7560 // Setup the registers to start vector comparison loop
7561 bind(COMPARE_WIDE_VECTORS);
7562 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7563 lea(str1, Address(str1, result, scale));
7564 lea(str2, Address(str2, result, scale));
7565 } else {
7566 lea(str1, Address(str1, result, scale1));
7567 lea(str2, Address(str2, result, scale2));
7568 }
7569 subl(result, stride2);
7570 subl(cnt2, stride2);
7571 jcc(Assembler::zero, COMPARE_WIDE_TAIL);
7572 negptr(result);
7573
7574 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
7575 bind(COMPARE_WIDE_VECTORS_LOOP);
7576
7577 #ifdef _LP64
7578 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
7579 cmpl(cnt2, stride2x2);
7580 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2);
7581 testl(cnt2, stride2x2-1); // cnt2 holds the vector count
7582 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40
7583
7584 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
7585 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7586 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit);
7587 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
7588 } else {
7589 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit);
7590 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
7591 }
7592 kortestql(k7, k7);
7593 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare
7594 addptr(result, stride2x2); // update since we already compared at this addr
7595 subl(cnt2, stride2x2); // and sub the size too
7596 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3);
7597
7598 vpxor(vec1, vec1);
7599 jmpb(COMPARE_WIDE_TAIL);
7600 }//if (VM_Version::supports_avx512vlbw())
7601 #endif // _LP64
7602
7603
7604 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
7605 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7606 vmovdqu(vec1, Address(str1, result, scale));
7607 vpxor(vec1, Address(str2, result, scale));
7608 } else {
7609 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit);
7610 vpxor(vec1, Address(str2, result, scale2));
7611 }
7612 vptest(vec1, vec1);
7613 jcc(Assembler::notZero, VECTOR_NOT_EQUAL);
7614 addptr(result, stride2);
7615 subl(cnt2, stride2);
7616 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
7617 // clean upper bits of YMM registers
7618 vpxor(vec1, vec1);
7619
7620 // compare wide vectors tail
7621 bind(COMPARE_WIDE_TAIL);
7622 testptr(result, result);
7623 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
7624
7625 movl(result, stride2);
7626 movl(cnt2, result);
7627 negptr(result);
7628 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2);
7629
7630 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
7631 bind(VECTOR_NOT_EQUAL);
7632 // clean upper bits of YMM registers
7633 vpxor(vec1, vec1);
7634 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7635 lea(str1, Address(str1, result, scale));
7636 lea(str2, Address(str2, result, scale));
7637 } else {
7638 lea(str1, Address(str1, result, scale1));
7639 lea(str2, Address(str2, result, scale2));
7640 }
7641 jmp(COMPARE_16_CHARS);
7642
7643 // Compare tail chars, length between 1 to 15 chars
7644 bind(COMPARE_TAIL_LONG);
7645 movl(cnt2, result);
7646 cmpl(cnt2, stride);
7647 jcc(Assembler::less, COMPARE_SMALL_STR);
7648
7649 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7650 movdqu(vec1, Address(str1, 0));
7651 } else {
7652 pmovzxbw(vec1, Address(str1, 0));
7653 }
7654 pcmpestri(vec1, Address(str2, 0), pcmpmask);
7655 jcc(Assembler::below, COMPARE_INDEX_CHAR);
7656 subptr(cnt2, stride);
7657 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
7658 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7659 lea(str1, Address(str1, result, scale));
7660 lea(str2, Address(str2, result, scale));
7661 } else {
7662 lea(str1, Address(str1, result, scale1));
7663 lea(str2, Address(str2, result, scale2));
7664 }
7665 negptr(cnt2);
7666 jmpb(WHILE_HEAD_LABEL);
7667
7668 bind(COMPARE_SMALL_STR);
7669 } else if (UseSSE42Intrinsics) {
7670 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
7671 int pcmpmask = 0x19;
7672 // Setup to compare 8-char (16-byte) vectors,
7673 // start from first character again because it has aligned address.
7674 movl(result, cnt2);
7675 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count
7676 if (ae == StrIntrinsicNode::LL) {
7677 pcmpmask &= ~0x01;
7678 }
7679 jcc(Assembler::zero, COMPARE_TAIL);
7680 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7681 lea(str1, Address(str1, result, scale));
7682 lea(str2, Address(str2, result, scale));
7683 } else {
7684 lea(str1, Address(str1, result, scale1));
7685 lea(str2, Address(str2, result, scale2));
7686 }
7687 negptr(result);
7688
7689 // pcmpestri
7690 // inputs:
7691 // vec1- substring
7692 // rax - negative string length (elements count)
7693 // mem - scanned string
7694 // rdx - string length (elements count)
7695 // pcmpmask - cmp mode: 11000 (string compare with negated result)
7696 // + 00 (unsigned bytes) or + 01 (unsigned shorts)
7697 // outputs:
7698 // rcx - first mismatched element index
7699 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
7700
7701 bind(COMPARE_WIDE_VECTORS);
7702 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7703 movdqu(vec1, Address(str1, result, scale));
7704 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
7705 } else {
7706 pmovzxbw(vec1, Address(str1, result, scale1));
7707 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
7708 }
7709 // After pcmpestri cnt1(rcx) contains mismatched element index
7710
7711 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1
7712 addptr(result, stride);
7713 subptr(cnt2, stride);
7714 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
7715
7716 // compare wide vectors tail
7717 testptr(result, result);
7718 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
7719
7720 movl(cnt2, stride);
7721 movl(result, stride);
7722 negptr(result);
7723 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7724 movdqu(vec1, Address(str1, result, scale));
7725 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
7726 } else {
7727 pmovzxbw(vec1, Address(str1, result, scale1));
7728 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
7729 }
7730 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
7731
7732 // Mismatched characters in the vectors
7733 bind(VECTOR_NOT_EQUAL);
7734 addptr(cnt1, result);
7735 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
7736 subl(result, cnt2);
7737 jmpb(POP_LABEL);
7738
7739 bind(COMPARE_TAIL); // limit is zero
7740 movl(cnt2, result);
7741 // Fallthru to tail compare
7742 }
7743 // Shift str2 and str1 to the end of the arrays, negate min
7744 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7745 lea(str1, Address(str1, cnt2, scale));
7746 lea(str2, Address(str2, cnt2, scale));
7747 } else {
7748 lea(str1, Address(str1, cnt2, scale1));
7749 lea(str2, Address(str2, cnt2, scale2));
7750 }
7751 decrementl(cnt2); // first character was compared already
7752 negptr(cnt2);
7753
7754 // Compare the rest of the elements
7755 bind(WHILE_HEAD_LABEL);
7756 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae);
7757 subl(result, cnt1);
7758 jccb(Assembler::notZero, POP_LABEL);
7759 increment(cnt2);
7760 jccb(Assembler::notZero, WHILE_HEAD_LABEL);
7761
7762 // Strings are equal up to min length. Return the length difference.
7763 bind(LENGTH_DIFF_LABEL);
7764 pop(result);
7765 if (ae == StrIntrinsicNode::UU) {
7766 // Divide diff by 2 to get number of chars
7767 sarl(result, 1);
7768 }
7769 jmpb(DONE_LABEL);
7770
7771 #ifdef _LP64
7772 if (VM_Version::supports_avx512vlbw()) {
7773
7774 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED);
7775
7776 kmovql(cnt1, k7);
7777 notq(cnt1);
7778 bsfq(cnt2, cnt1);
7779 if (ae != StrIntrinsicNode::LL) {
7780 // Divide diff by 2 to get number of chars
7781 sarl(cnt2, 1);
7782 }
7783 addq(result, cnt2);
7784 if (ae == StrIntrinsicNode::LL) {
7785 load_unsigned_byte(cnt1, Address(str2, result));
7786 load_unsigned_byte(result, Address(str1, result));
7787 } else if (ae == StrIntrinsicNode::UU) {
7788 load_unsigned_short(cnt1, Address(str2, result, scale));
7789 load_unsigned_short(result, Address(str1, result, scale));
7790 } else {
7791 load_unsigned_short(cnt1, Address(str2, result, scale2));
7792 load_unsigned_byte(result, Address(str1, result, scale1));
7793 }
7794 subl(result, cnt1);
7795 jmpb(POP_LABEL);
7796 }//if (VM_Version::supports_avx512vlbw())
7797 #endif // _LP64
7798
7799 // Discard the stored length difference
7800 bind(POP_LABEL);
7801 pop(cnt1);
7802
7803 // That's it
7804 bind(DONE_LABEL);
7805 if(ae == StrIntrinsicNode::UL) {
7806 negl(result);
7807 }
7808
7809 }
7810
7811 // Search for Non-ASCII character (Negative byte value) in a byte array,
7812 // return true if it has any and false otherwise.
7813 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java
7814 // @HotSpotIntrinsicCandidate
7815 // private static boolean hasNegatives(byte[] ba, int off, int len) {
7816 // for (int i = off; i < off + len; i++) {
7817 // if (ba[i] < 0) {
7818 // return true;
7819 // }
7820 // }
7821 // return false;
7822 // }
7823 void MacroAssembler::has_negatives(Register ary1, Register len,
7824 Register result, Register tmp1,
7825 XMMRegister vec1, XMMRegister vec2) {
7826 // rsi: byte array
7827 // rcx: len
7828 // rax: result
7829 ShortBranchVerifier sbv(this);
7830 assert_different_registers(ary1, len, result, tmp1);
7831 assert_different_registers(vec1, vec2);
7832 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE;
7833
7834 // len == 0
7835 testl(len, len);
7836 jcc(Assembler::zero, FALSE_LABEL);
7837
7838 if ((UseAVX > 2) && // AVX512
7839 VM_Version::supports_avx512vlbw() &&
7840 VM_Version::supports_bmi2()) {
7841
7842 set_vector_masking(); // opening of the stub context for programming mask registers
7843
7844 Label test_64_loop, test_tail;
7845 Register tmp3_aliased = len;
7846
7847 movl(tmp1, len);
7848 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit);
7849
7850 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F
7851 andl(len, ~(64 - 1)); // vector count (in chars)
7852 jccb(Assembler::zero, test_tail);
7853
7854 lea(ary1, Address(ary1, len, Address::times_1));
7855 negptr(len);
7856
7857 bind(test_64_loop);
7858 // Check whether our 64 elements of size byte contain negatives
7859 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit);
7860 kortestql(k2, k2);
7861 jcc(Assembler::notZero, TRUE_LABEL);
7862
7863 addptr(len, 64);
7864 jccb(Assembler::notZero, test_64_loop);
7865
7866
7867 bind(test_tail);
7868 // bail out when there is nothing to be done
7869 testl(tmp1, -1);
7870 jcc(Assembler::zero, FALSE_LABEL);
7871
7872 // Save k1
7873 kmovql(k3, k1);
7874
7875 // ~(~0 << len) applied up to two times (for 32-bit scenario)
7876 #ifdef _LP64
7877 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF);
7878 shlxq(tmp3_aliased, tmp3_aliased, tmp1);
7879 notq(tmp3_aliased);
7880 kmovql(k1, tmp3_aliased);
7881 #else
7882 Label k_init;
7883 jmp(k_init);
7884
7885 // We could not read 64-bits from a general purpose register thus we move
7886 // data required to compose 64 1's to the instruction stream
7887 // We emit 64 byte wide series of elements from 0..63 which later on would
7888 // be used as a compare targets with tail count contained in tmp1 register.
7889 // Result would be a k1 register having tmp1 consecutive number or 1
7890 // counting from least significant bit.
7891 address tmp = pc();
7892 emit_int64(0x0706050403020100);
7893 emit_int64(0x0F0E0D0C0B0A0908);
7894 emit_int64(0x1716151413121110);
7895 emit_int64(0x1F1E1D1C1B1A1918);
7896 emit_int64(0x2726252423222120);
7897 emit_int64(0x2F2E2D2C2B2A2928);
7898 emit_int64(0x3736353433323130);
7899 emit_int64(0x3F3E3D3C3B3A3938);
7900
7901 bind(k_init);
7902 lea(len, InternalAddress(tmp));
7903 // create mask to test for negative byte inside a vector
7904 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit);
7905 evpcmpgtb(k1, vec1, Address(len, 0), Assembler::AVX_512bit);
7906
7907 #endif
7908 evpcmpgtb(k2, k1, vec2, Address(ary1, 0), Assembler::AVX_512bit);
7909 ktestq(k2, k1);
7910 // Restore k1
7911 kmovql(k1, k3);
7912 jcc(Assembler::notZero, TRUE_LABEL);
7913
7914 jmp(FALSE_LABEL);
7915
7916 clear_vector_masking(); // closing of the stub context for programming mask registers
7917 } else {
7918 movl(result, len); // copy
7919
7920 if (UseAVX == 2 && UseSSE >= 2) {
7921 // With AVX2, use 32-byte vector compare
7922 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7923
7924 // Compare 32-byte vectors
7925 andl(result, 0x0000001f); // tail count (in bytes)
7926 andl(len, 0xffffffe0); // vector count (in bytes)
7927 jccb(Assembler::zero, COMPARE_TAIL);
7928
7929 lea(ary1, Address(ary1, len, Address::times_1));
7930 negptr(len);
7931
7932 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector
7933 movdl(vec2, tmp1);
7934 vpbroadcastd(vec2, vec2);
7935
7936 bind(COMPARE_WIDE_VECTORS);
7937 vmovdqu(vec1, Address(ary1, len, Address::times_1));
7938 vptest(vec1, vec2);
7939 jccb(Assembler::notZero, TRUE_LABEL);
7940 addptr(len, 32);
7941 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7942
7943 testl(result, result);
7944 jccb(Assembler::zero, FALSE_LABEL);
7945
7946 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
7947 vptest(vec1, vec2);
7948 jccb(Assembler::notZero, TRUE_LABEL);
7949 jmpb(FALSE_LABEL);
7950
7951 bind(COMPARE_TAIL); // len is zero
7952 movl(len, result);
7953 // Fallthru to tail compare
7954 } else if (UseSSE42Intrinsics) {
7955 // With SSE4.2, use double quad vector compare
7956 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7957
7958 // Compare 16-byte vectors
7959 andl(result, 0x0000000f); // tail count (in bytes)
7960 andl(len, 0xfffffff0); // vector count (in bytes)
7961 jccb(Assembler::zero, COMPARE_TAIL);
7962
7963 lea(ary1, Address(ary1, len, Address::times_1));
7964 negptr(len);
7965
7966 movl(tmp1, 0x80808080);
7967 movdl(vec2, tmp1);
7968 pshufd(vec2, vec2, 0);
7969
7970 bind(COMPARE_WIDE_VECTORS);
7971 movdqu(vec1, Address(ary1, len, Address::times_1));
7972 ptest(vec1, vec2);
7973 jccb(Assembler::notZero, TRUE_LABEL);
7974 addptr(len, 16);
7975 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7976
7977 testl(result, result);
7978 jccb(Assembler::zero, FALSE_LABEL);
7979
7980 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
7981 ptest(vec1, vec2);
7982 jccb(Assembler::notZero, TRUE_LABEL);
7983 jmpb(FALSE_LABEL);
7984
7985 bind(COMPARE_TAIL); // len is zero
7986 movl(len, result);
7987 // Fallthru to tail compare
7988 }
7989 }
7990 // Compare 4-byte vectors
7991 andl(len, 0xfffffffc); // vector count (in bytes)
7992 jccb(Assembler::zero, COMPARE_CHAR);
7993
7994 lea(ary1, Address(ary1, len, Address::times_1));
7995 negptr(len);
7996
7997 bind(COMPARE_VECTORS);
7998 movl(tmp1, Address(ary1, len, Address::times_1));
7999 andl(tmp1, 0x80808080);
8000 jccb(Assembler::notZero, TRUE_LABEL);
8001 addptr(len, 4);
8002 jcc(Assembler::notZero, COMPARE_VECTORS);
8003
8004 // Compare trailing char (final 2 bytes), if any
8005 bind(COMPARE_CHAR);
8006 testl(result, 0x2); // tail char
8007 jccb(Assembler::zero, COMPARE_BYTE);
8008 load_unsigned_short(tmp1, Address(ary1, 0));
8009 andl(tmp1, 0x00008080);
8010 jccb(Assembler::notZero, TRUE_LABEL);
8011 subptr(result, 2);
8012 lea(ary1, Address(ary1, 2));
8013
8014 bind(COMPARE_BYTE);
8015 testl(result, 0x1); // tail byte
8016 jccb(Assembler::zero, FALSE_LABEL);
8017 load_unsigned_byte(tmp1, Address(ary1, 0));
8018 andl(tmp1, 0x00000080);
8019 jccb(Assembler::notEqual, TRUE_LABEL);
8020 jmpb(FALSE_LABEL);
8021
8022 bind(TRUE_LABEL);
8023 movl(result, 1); // return true
8024 jmpb(DONE);
8025
8026 bind(FALSE_LABEL);
8027 xorl(result, result); // return false
8028
8029 // That's it
8030 bind(DONE);
8031 if (UseAVX >= 2 && UseSSE >= 2) {
8032 // clean upper bits of YMM registers
8033 vpxor(vec1, vec1);
8034 vpxor(vec2, vec2);
8035 }
8036 }
8037 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings.
8038 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2,
8039 Register limit, Register result, Register chr,
8040 XMMRegister vec1, XMMRegister vec2, bool is_char) {
8041 ShortBranchVerifier sbv(this);
8042 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE;
8043
8044 int length_offset = arrayOopDesc::length_offset_in_bytes();
8045 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE);
8046
8047 if (is_array_equ) {
8048 // Check the input args
8049 cmpoop(ary1, ary2);
8050 jcc(Assembler::equal, TRUE_LABEL);
8051
8052 // Need additional checks for arrays_equals.
8053 testptr(ary1, ary1);
8054 jcc(Assembler::zero, FALSE_LABEL);
8055 testptr(ary2, ary2);
8056 jcc(Assembler::zero, FALSE_LABEL);
8057
8058 // Check the lengths
8059 movl(limit, Address(ary1, length_offset));
8060 cmpl(limit, Address(ary2, length_offset));
8061 jcc(Assembler::notEqual, FALSE_LABEL);
8062 }
8063
8064 // count == 0
8065 testl(limit, limit);
8066 jcc(Assembler::zero, TRUE_LABEL);
8067
8068 if (is_array_equ) {
8069 // Load array address
8070 lea(ary1, Address(ary1, base_offset));
8071 lea(ary2, Address(ary2, base_offset));
8072 }
8073
8074 if (is_array_equ && is_char) {
8075 // arrays_equals when used for char[].
8076 shll(limit, 1); // byte count != 0
8077 }
8078 movl(result, limit); // copy
8079
8080 if (UseAVX >= 2) {
8081 // With AVX2, use 32-byte vector compare
8082 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8083
8084 // Compare 32-byte vectors
8085 andl(result, 0x0000001f); // tail count (in bytes)
8086 andl(limit, 0xffffffe0); // vector count (in bytes)
8087 jcc(Assembler::zero, COMPARE_TAIL);
8088
8089 lea(ary1, Address(ary1, limit, Address::times_1));
8090 lea(ary2, Address(ary2, limit, Address::times_1));
8091 negptr(limit);
8092
8093 bind(COMPARE_WIDE_VECTORS);
8094
8095 #ifdef _LP64
8096 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
8097 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3;
8098
8099 cmpl(limit, -64);
8100 jccb(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2);
8101
8102 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
8103
8104 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit);
8105 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit);
8106 kortestql(k7, k7);
8107 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
8108 addptr(limit, 64); // update since we already compared at this addr
8109 cmpl(limit, -64);
8110 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3);
8111
8112 // At this point we may still need to compare -limit+result bytes.
8113 // We could execute the next two instruction and just continue via non-wide path:
8114 // cmpl(limit, 0);
8115 // jcc(Assembler::equal, COMPARE_TAIL); // true
8116 // But since we stopped at the points ary{1,2}+limit which are
8117 // not farther than 64 bytes from the ends of arrays ary{1,2}+result
8118 // (|limit| <= 32 and result < 32),
8119 // we may just compare the last 64 bytes.
8120 //
8121 addptr(result, -64); // it is safe, bc we just came from this area
8122 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit);
8123 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit);
8124 kortestql(k7, k7);
8125 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
8126
8127 jmp(TRUE_LABEL);
8128
8129 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
8130
8131 }//if (VM_Version::supports_avx512vlbw())
8132 #endif //_LP64
8133
8134 vmovdqu(vec1, Address(ary1, limit, Address::times_1));
8135 vmovdqu(vec2, Address(ary2, limit, Address::times_1));
8136 vpxor(vec1, vec2);
8137
8138 vptest(vec1, vec1);
8139 jcc(Assembler::notZero, FALSE_LABEL);
8140 addptr(limit, 32);
8141 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8142
8143 testl(result, result);
8144 jcc(Assembler::zero, TRUE_LABEL);
8145
8146 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
8147 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
8148 vpxor(vec1, vec2);
8149
8150 vptest(vec1, vec1);
8151 jccb(Assembler::notZero, FALSE_LABEL);
8152 jmpb(TRUE_LABEL);
8153
8154 bind(COMPARE_TAIL); // limit is zero
8155 movl(limit, result);
8156 // Fallthru to tail compare
8157 } else if (UseSSE42Intrinsics) {
8158 // With SSE4.2, use double quad vector compare
8159 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8160
8161 // Compare 16-byte vectors
8162 andl(result, 0x0000000f); // tail count (in bytes)
8163 andl(limit, 0xfffffff0); // vector count (in bytes)
8164 jcc(Assembler::zero, COMPARE_TAIL);
8165
8166 lea(ary1, Address(ary1, limit, Address::times_1));
8167 lea(ary2, Address(ary2, limit, Address::times_1));
8168 negptr(limit);
8169
8170 bind(COMPARE_WIDE_VECTORS);
8171 movdqu(vec1, Address(ary1, limit, Address::times_1));
8172 movdqu(vec2, Address(ary2, limit, Address::times_1));
8173 pxor(vec1, vec2);
8174
8175 ptest(vec1, vec1);
8176 jcc(Assembler::notZero, FALSE_LABEL);
8177 addptr(limit, 16);
8178 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8179
8180 testl(result, result);
8181 jcc(Assembler::zero, TRUE_LABEL);
8182
8183 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
8184 movdqu(vec2, Address(ary2, result, Address::times_1, -16));
8185 pxor(vec1, vec2);
8186
8187 ptest(vec1, vec1);
8188 jccb(Assembler::notZero, FALSE_LABEL);
8189 jmpb(TRUE_LABEL);
8190
8191 bind(COMPARE_TAIL); // limit is zero
8192 movl(limit, result);
8193 // Fallthru to tail compare
8194 }
8195
8196 // Compare 4-byte vectors
8197 andl(limit, 0xfffffffc); // vector count (in bytes)
8198 jccb(Assembler::zero, COMPARE_CHAR);
8199
8200 lea(ary1, Address(ary1, limit, Address::times_1));
8201 lea(ary2, Address(ary2, limit, Address::times_1));
8202 negptr(limit);
8203
8204 bind(COMPARE_VECTORS);
8205 movl(chr, Address(ary1, limit, Address::times_1));
8206 cmpl(chr, Address(ary2, limit, Address::times_1));
8207 jccb(Assembler::notEqual, FALSE_LABEL);
8208 addptr(limit, 4);
8209 jcc(Assembler::notZero, COMPARE_VECTORS);
8210
8211 // Compare trailing char (final 2 bytes), if any
8212 bind(COMPARE_CHAR);
8213 testl(result, 0x2); // tail char
8214 jccb(Assembler::zero, COMPARE_BYTE);
8215 load_unsigned_short(chr, Address(ary1, 0));
8216 load_unsigned_short(limit, Address(ary2, 0));
8217 cmpl(chr, limit);
8218 jccb(Assembler::notEqual, FALSE_LABEL);
8219
8220 if (is_array_equ && is_char) {
8221 bind(COMPARE_BYTE);
8222 } else {
8223 lea(ary1, Address(ary1, 2));
8224 lea(ary2, Address(ary2, 2));
8225
8226 bind(COMPARE_BYTE);
8227 testl(result, 0x1); // tail byte
8228 jccb(Assembler::zero, TRUE_LABEL);
8229 load_unsigned_byte(chr, Address(ary1, 0));
8230 load_unsigned_byte(limit, Address(ary2, 0));
8231 cmpl(chr, limit);
8232 jccb(Assembler::notEqual, FALSE_LABEL);
8233 }
8234 bind(TRUE_LABEL);
8235 movl(result, 1); // return true
8236 jmpb(DONE);
8237
8238 bind(FALSE_LABEL);
8239 xorl(result, result); // return false
8240
8241 // That's it
8242 bind(DONE);
8243 if (UseAVX >= 2) {
8244 // clean upper bits of YMM registers
8245 vpxor(vec1, vec1);
8246 vpxor(vec2, vec2);
8247 }
8248 }
8249
8250 #endif
8251
8252 void MacroAssembler::generate_fill(BasicType t, bool aligned,
8253 Register to, Register value, Register count,
8254 Register rtmp, XMMRegister xtmp) {
8255 ShortBranchVerifier sbv(this);
8256 assert_different_registers(to, value, count, rtmp);
8257 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
8258 Label L_fill_2_bytes, L_fill_4_bytes;
8259
8260 int shift = -1;
8261 switch (t) {
8262 case T_BYTE:
8263 shift = 2;
8264 break;
8265 case T_SHORT:
8266 shift = 1;
8267 break;
8268 case T_INT:
8269 shift = 0;
8270 break;
8271 default: ShouldNotReachHere();
8272 }
8273
8274 if (t == T_BYTE) {
8275 andl(value, 0xff);
8276 movl(rtmp, value);
8277 shll(rtmp, 8);
8278 orl(value, rtmp);
8279 }
8280 if (t == T_SHORT) {
8281 andl(value, 0xffff);
8282 }
8283 if (t == T_BYTE || t == T_SHORT) {
8284 movl(rtmp, value);
8285 shll(rtmp, 16);
8286 orl(value, rtmp);
8287 }
8288
8289 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
8290 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp
8291 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) {
8292 // align source address at 4 bytes address boundary
8293 if (t == T_BYTE) {
8294 // One byte misalignment happens only for byte arrays
8295 testptr(to, 1);
8296 jccb(Assembler::zero, L_skip_align1);
8297 movb(Address(to, 0), value);
8298 increment(to);
8299 decrement(count);
8300 BIND(L_skip_align1);
8301 }
8302 // Two bytes misalignment happens only for byte and short (char) arrays
8303 testptr(to, 2);
8304 jccb(Assembler::zero, L_skip_align2);
8305 movw(Address(to, 0), value);
8306 addptr(to, 2);
8307 subl(count, 1<<(shift-1));
8308 BIND(L_skip_align2);
8309 }
8310 if (UseSSE < 2) {
8311 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
8312 // Fill 32-byte chunks
8313 subl(count, 8 << shift);
8314 jcc(Assembler::less, L_check_fill_8_bytes);
8315 align(16);
8316
8317 BIND(L_fill_32_bytes_loop);
8318
8319 for (int i = 0; i < 32; i += 4) {
8320 movl(Address(to, i), value);
8321 }
8322
8323 addptr(to, 32);
8324 subl(count, 8 << shift);
8325 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
8326 BIND(L_check_fill_8_bytes);
8327 addl(count, 8 << shift);
8328 jccb(Assembler::zero, L_exit);
8329 jmpb(L_fill_8_bytes);
8330
8331 //
8332 // length is too short, just fill qwords
8333 //
8334 BIND(L_fill_8_bytes_loop);
8335 movl(Address(to, 0), value);
8336 movl(Address(to, 4), value);
8337 addptr(to, 8);
8338 BIND(L_fill_8_bytes);
8339 subl(count, 1 << (shift + 1));
8340 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
8341 // fall through to fill 4 bytes
8342 } else {
8343 Label L_fill_32_bytes;
8344 if (!UseUnalignedLoadStores) {
8345 // align to 8 bytes, we know we are 4 byte aligned to start
8346 testptr(to, 4);
8347 jccb(Assembler::zero, L_fill_32_bytes);
8348 movl(Address(to, 0), value);
8349 addptr(to, 4);
8350 subl(count, 1<<shift);
8351 }
8352 BIND(L_fill_32_bytes);
8353 {
8354 assert( UseSSE >= 2, "supported cpu only" );
8355 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
8356 if (UseAVX > 2) {
8357 movl(rtmp, 0xffff);
8358 kmovwl(k1, rtmp);
8359 }
8360 movdl(xtmp, value);
8361 if (UseAVX > 2 && UseUnalignedLoadStores) {
8362 // Fill 64-byte chunks
8363 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
8364 evpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit);
8365
8366 subl(count, 16 << shift);
8367 jcc(Assembler::less, L_check_fill_32_bytes);
8368 align(16);
8369
8370 BIND(L_fill_64_bytes_loop);
8371 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit);
8372 addptr(to, 64);
8373 subl(count, 16 << shift);
8374 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
8375
8376 BIND(L_check_fill_32_bytes);
8377 addl(count, 8 << shift);
8378 jccb(Assembler::less, L_check_fill_8_bytes);
8379 vmovdqu(Address(to, 0), xtmp);
8380 addptr(to, 32);
8381 subl(count, 8 << shift);
8382
8383 BIND(L_check_fill_8_bytes);
8384 } else if (UseAVX == 2 && UseUnalignedLoadStores) {
8385 // Fill 64-byte chunks
8386 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
8387 vpbroadcastd(xtmp, xtmp);
8388
8389 subl(count, 16 << shift);
8390 jcc(Assembler::less, L_check_fill_32_bytes);
8391 align(16);
8392
8393 BIND(L_fill_64_bytes_loop);
8394 vmovdqu(Address(to, 0), xtmp);
8395 vmovdqu(Address(to, 32), xtmp);
8396 addptr(to, 64);
8397 subl(count, 16 << shift);
8398 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
8399
8400 BIND(L_check_fill_32_bytes);
8401 addl(count, 8 << shift);
8402 jccb(Assembler::less, L_check_fill_8_bytes);
8403 vmovdqu(Address(to, 0), xtmp);
8404 addptr(to, 32);
8405 subl(count, 8 << shift);
8406
8407 BIND(L_check_fill_8_bytes);
8408 // clean upper bits of YMM registers
8409 movdl(xtmp, value);
8410 pshufd(xtmp, xtmp, 0);
8411 } else {
8412 // Fill 32-byte chunks
8413 pshufd(xtmp, xtmp, 0);
8414
8415 subl(count, 8 << shift);
8416 jcc(Assembler::less, L_check_fill_8_bytes);
8417 align(16);
8418
8419 BIND(L_fill_32_bytes_loop);
8420
8421 if (UseUnalignedLoadStores) {
8422 movdqu(Address(to, 0), xtmp);
8423 movdqu(Address(to, 16), xtmp);
8424 } else {
8425 movq(Address(to, 0), xtmp);
8426 movq(Address(to, 8), xtmp);
8427 movq(Address(to, 16), xtmp);
8428 movq(Address(to, 24), xtmp);
8429 }
8430
8431 addptr(to, 32);
8432 subl(count, 8 << shift);
8433 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
8434
8435 BIND(L_check_fill_8_bytes);
8436 }
8437 addl(count, 8 << shift);
8438 jccb(Assembler::zero, L_exit);
8439 jmpb(L_fill_8_bytes);
8440
8441 //
8442 // length is too short, just fill qwords
8443 //
8444 BIND(L_fill_8_bytes_loop);
8445 movq(Address(to, 0), xtmp);
8446 addptr(to, 8);
8447 BIND(L_fill_8_bytes);
8448 subl(count, 1 << (shift + 1));
8449 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
8450 }
8451 }
8452 // fill trailing 4 bytes
8453 BIND(L_fill_4_bytes);
8454 testl(count, 1<<shift);
8455 jccb(Assembler::zero, L_fill_2_bytes);
8456 movl(Address(to, 0), value);
8457 if (t == T_BYTE || t == T_SHORT) {
8458 addptr(to, 4);
8459 BIND(L_fill_2_bytes);
8460 // fill trailing 2 bytes
8461 testl(count, 1<<(shift-1));
8462 jccb(Assembler::zero, L_fill_byte);
8463 movw(Address(to, 0), value);
8464 if (t == T_BYTE) {
8465 addptr(to, 2);
8466 BIND(L_fill_byte);
8467 // fill trailing byte
8468 testl(count, 1);
8469 jccb(Assembler::zero, L_exit);
8470 movb(Address(to, 0), value);
8471 } else {
8472 BIND(L_fill_byte);
8473 }
8474 } else {
8475 BIND(L_fill_2_bytes);
8476 }
8477 BIND(L_exit);
8478 }
8479
8480 // encode char[] to byte[] in ISO_8859_1
8481 //@HotSpotIntrinsicCandidate
8482 //private static int implEncodeISOArray(byte[] sa, int sp,
8483 //byte[] da, int dp, int len) {
8484 // int i = 0;
8485 // for (; i < len; i++) {
8486 // char c = StringUTF16.getChar(sa, sp++);
8487 // if (c > '\u00FF')
8488 // break;
8489 // da[dp++] = (byte)c;
8490 // }
8491 // return i;
8492 //}
8493 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len,
8494 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
8495 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
8496 Register tmp5, Register result) {
8497
8498 // rsi: src
8499 // rdi: dst
8500 // rdx: len
8501 // rcx: tmp5
8502 // rax: result
8503 ShortBranchVerifier sbv(this);
8504 assert_different_registers(src, dst, len, tmp5, result);
8505 Label L_done, L_copy_1_char, L_copy_1_char_exit;
8506
8507 // set result
8508 xorl(result, result);
8509 // check for zero length
8510 testl(len, len);
8511 jcc(Assembler::zero, L_done);
8512
8513 movl(result, len);
8514
8515 // Setup pointers
8516 lea(src, Address(src, len, Address::times_2)); // char[]
8517 lea(dst, Address(dst, len, Address::times_1)); // byte[]
8518 negptr(len);
8519
8520 if (UseSSE42Intrinsics || UseAVX >= 2) {
8521 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit;
8522 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit;
8523
8524 if (UseAVX >= 2) {
8525 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit;
8526 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
8527 movdl(tmp1Reg, tmp5);
8528 vpbroadcastd(tmp1Reg, tmp1Reg);
8529 jmp(L_chars_32_check);
8530
8531 bind(L_copy_32_chars);
8532 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64));
8533 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32));
8534 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
8535 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
8536 jccb(Assembler::notZero, L_copy_32_chars_exit);
8537 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
8538 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1);
8539 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg);
8540
8541 bind(L_chars_32_check);
8542 addptr(len, 32);
8543 jcc(Assembler::lessEqual, L_copy_32_chars);
8544
8545 bind(L_copy_32_chars_exit);
8546 subptr(len, 16);
8547 jccb(Assembler::greater, L_copy_16_chars_exit);
8548
8549 } else if (UseSSE42Intrinsics) {
8550 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
8551 movdl(tmp1Reg, tmp5);
8552 pshufd(tmp1Reg, tmp1Reg, 0);
8553 jmpb(L_chars_16_check);
8554 }
8555
8556 bind(L_copy_16_chars);
8557 if (UseAVX >= 2) {
8558 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32));
8559 vptest(tmp2Reg, tmp1Reg);
8560 jcc(Assembler::notZero, L_copy_16_chars_exit);
8561 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1);
8562 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1);
8563 } else {
8564 if (UseAVX > 0) {
8565 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
8566 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
8567 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0);
8568 } else {
8569 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
8570 por(tmp2Reg, tmp3Reg);
8571 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
8572 por(tmp2Reg, tmp4Reg);
8573 }
8574 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
8575 jccb(Assembler::notZero, L_copy_16_chars_exit);
8576 packuswb(tmp3Reg, tmp4Reg);
8577 }
8578 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg);
8579
8580 bind(L_chars_16_check);
8581 addptr(len, 16);
8582 jcc(Assembler::lessEqual, L_copy_16_chars);
8583
8584 bind(L_copy_16_chars_exit);
8585 if (UseAVX >= 2) {
8586 // clean upper bits of YMM registers
8587 vpxor(tmp2Reg, tmp2Reg);
8588 vpxor(tmp3Reg, tmp3Reg);
8589 vpxor(tmp4Reg, tmp4Reg);
8590 movdl(tmp1Reg, tmp5);
8591 pshufd(tmp1Reg, tmp1Reg, 0);
8592 }
8593 subptr(len, 8);
8594 jccb(Assembler::greater, L_copy_8_chars_exit);
8595
8596 bind(L_copy_8_chars);
8597 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16));
8598 ptest(tmp3Reg, tmp1Reg);
8599 jccb(Assembler::notZero, L_copy_8_chars_exit);
8600 packuswb(tmp3Reg, tmp1Reg);
8601 movq(Address(dst, len, Address::times_1, -8), tmp3Reg);
8602 addptr(len, 8);
8603 jccb(Assembler::lessEqual, L_copy_8_chars);
8604
8605 bind(L_copy_8_chars_exit);
8606 subptr(len, 8);
8607 jccb(Assembler::zero, L_done);
8608 }
8609
8610 bind(L_copy_1_char);
8611 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0));
8612 testl(tmp5, 0xff00); // check if Unicode char
8613 jccb(Assembler::notZero, L_copy_1_char_exit);
8614 movb(Address(dst, len, Address::times_1, 0), tmp5);
8615 addptr(len, 1);
8616 jccb(Assembler::less, L_copy_1_char);
8617
8618 bind(L_copy_1_char_exit);
8619 addptr(result, len); // len is negative count of not processed elements
8620
8621 bind(L_done);
8622 }
8623
8624 #ifdef _LP64
8625 /**
8626 * Helper for multiply_to_len().
8627 */
8628 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) {
8629 addq(dest_lo, src1);
8630 adcq(dest_hi, 0);
8631 addq(dest_lo, src2);
8632 adcq(dest_hi, 0);
8633 }
8634
8635 /**
8636 * Multiply 64 bit by 64 bit first loop.
8637 */
8638 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
8639 Register y, Register y_idx, Register z,
8640 Register carry, Register product,
8641 Register idx, Register kdx) {
8642 //
8643 // jlong carry, x[], y[], z[];
8644 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
8645 // huge_128 product = y[idx] * x[xstart] + carry;
8646 // z[kdx] = (jlong)product;
8647 // carry = (jlong)(product >>> 64);
8648 // }
8649 // z[xstart] = carry;
8650 //
8651
8652 Label L_first_loop, L_first_loop_exit;
8653 Label L_one_x, L_one_y, L_multiply;
8654
8655 decrementl(xstart);
8656 jcc(Assembler::negative, L_one_x);
8657
8658 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
8659 rorq(x_xstart, 32); // convert big-endian to little-endian
8660
8661 bind(L_first_loop);
8662 decrementl(idx);
8663 jcc(Assembler::negative, L_first_loop_exit);
8664 decrementl(idx);
8665 jcc(Assembler::negative, L_one_y);
8666 movq(y_idx, Address(y, idx, Address::times_4, 0));
8667 rorq(y_idx, 32); // convert big-endian to little-endian
8668 bind(L_multiply);
8669 movq(product, x_xstart);
8670 mulq(y_idx); // product(rax) * y_idx -> rdx:rax
8671 addq(product, carry);
8672 adcq(rdx, 0);
8673 subl(kdx, 2);
8674 movl(Address(z, kdx, Address::times_4, 4), product);
8675 shrq(product, 32);
8676 movl(Address(z, kdx, Address::times_4, 0), product);
8677 movq(carry, rdx);
8678 jmp(L_first_loop);
8679
8680 bind(L_one_y);
8681 movl(y_idx, Address(y, 0));
8682 jmp(L_multiply);
8683
8684 bind(L_one_x);
8685 movl(x_xstart, Address(x, 0));
8686 jmp(L_first_loop);
8687
8688 bind(L_first_loop_exit);
8689 }
8690
8691 /**
8692 * Multiply 64 bit by 64 bit and add 128 bit.
8693 */
8694 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z,
8695 Register yz_idx, Register idx,
8696 Register carry, Register product, int offset) {
8697 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
8698 // z[kdx] = (jlong)product;
8699
8700 movq(yz_idx, Address(y, idx, Address::times_4, offset));
8701 rorq(yz_idx, 32); // convert big-endian to little-endian
8702 movq(product, x_xstart);
8703 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
8704 movq(yz_idx, Address(z, idx, Address::times_4, offset));
8705 rorq(yz_idx, 32); // convert big-endian to little-endian
8706
8707 add2_with_carry(rdx, product, carry, yz_idx);
8708
8709 movl(Address(z, idx, Address::times_4, offset+4), product);
8710 shrq(product, 32);
8711 movl(Address(z, idx, Address::times_4, offset), product);
8712
8713 }
8714
8715 /**
8716 * Multiply 128 bit by 128 bit. Unrolled inner loop.
8717 */
8718 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z,
8719 Register yz_idx, Register idx, Register jdx,
8720 Register carry, Register product,
8721 Register carry2) {
8722 // jlong carry, x[], y[], z[];
8723 // int kdx = ystart+1;
8724 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
8725 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
8726 // z[kdx+idx+1] = (jlong)product;
8727 // jlong carry2 = (jlong)(product >>> 64);
8728 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
8729 // z[kdx+idx] = (jlong)product;
8730 // carry = (jlong)(product >>> 64);
8731 // }
8732 // idx += 2;
8733 // if (idx > 0) {
8734 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
8735 // z[kdx+idx] = (jlong)product;
8736 // carry = (jlong)(product >>> 64);
8737 // }
8738 //
8739
8740 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
8741
8742 movl(jdx, idx);
8743 andl(jdx, 0xFFFFFFFC);
8744 shrl(jdx, 2);
8745
8746 bind(L_third_loop);
8747 subl(jdx, 1);
8748 jcc(Assembler::negative, L_third_loop_exit);
8749 subl(idx, 4);
8750
8751 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
8752 movq(carry2, rdx);
8753
8754 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
8755 movq(carry, rdx);
8756 jmp(L_third_loop);
8757
8758 bind (L_third_loop_exit);
8759
8760 andl (idx, 0x3);
8761 jcc(Assembler::zero, L_post_third_loop_done);
8762
8763 Label L_check_1;
8764 subl(idx, 2);
8765 jcc(Assembler::negative, L_check_1);
8766
8767 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
8768 movq(carry, rdx);
8769
8770 bind (L_check_1);
8771 addl (idx, 0x2);
8772 andl (idx, 0x1);
8773 subl(idx, 1);
8774 jcc(Assembler::negative, L_post_third_loop_done);
8775
8776 movl(yz_idx, Address(y, idx, Address::times_4, 0));
8777 movq(product, x_xstart);
8778 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
8779 movl(yz_idx, Address(z, idx, Address::times_4, 0));
8780
8781 add2_with_carry(rdx, product, yz_idx, carry);
8782
8783 movl(Address(z, idx, Address::times_4, 0), product);
8784 shrq(product, 32);
8785
8786 shlq(rdx, 32);
8787 orq(product, rdx);
8788 movq(carry, product);
8789
8790 bind(L_post_third_loop_done);
8791 }
8792
8793 /**
8794 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop.
8795 *
8796 */
8797 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z,
8798 Register carry, Register carry2,
8799 Register idx, Register jdx,
8800 Register yz_idx1, Register yz_idx2,
8801 Register tmp, Register tmp3, Register tmp4) {
8802 assert(UseBMI2Instructions, "should be used only when BMI2 is available");
8803
8804 // jlong carry, x[], y[], z[];
8805 // int kdx = ystart+1;
8806 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
8807 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry;
8808 // jlong carry2 = (jlong)(tmp3 >>> 64);
8809 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2;
8810 // carry = (jlong)(tmp4 >>> 64);
8811 // z[kdx+idx+1] = (jlong)tmp3;
8812 // z[kdx+idx] = (jlong)tmp4;
8813 // }
8814 // idx += 2;
8815 // if (idx > 0) {
8816 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry;
8817 // z[kdx+idx] = (jlong)yz_idx1;
8818 // carry = (jlong)(yz_idx1 >>> 64);
8819 // }
8820 //
8821
8822 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
8823
8824 movl(jdx, idx);
8825 andl(jdx, 0xFFFFFFFC);
8826 shrl(jdx, 2);
8827
8828 bind(L_third_loop);
8829 subl(jdx, 1);
8830 jcc(Assembler::negative, L_third_loop_exit);
8831 subl(idx, 4);
8832
8833 movq(yz_idx1, Address(y, idx, Address::times_4, 8));
8834 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
8835 movq(yz_idx2, Address(y, idx, Address::times_4, 0));
8836 rorxq(yz_idx2, yz_idx2, 32);
8837
8838 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
8839 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp
8840
8841 movq(yz_idx1, Address(z, idx, Address::times_4, 8));
8842 rorxq(yz_idx1, yz_idx1, 32);
8843 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
8844 rorxq(yz_idx2, yz_idx2, 32);
8845
8846 if (VM_Version::supports_adx()) {
8847 adcxq(tmp3, carry);
8848 adoxq(tmp3, yz_idx1);
8849
8850 adcxq(tmp4, tmp);
8851 adoxq(tmp4, yz_idx2);
8852
8853 movl(carry, 0); // does not affect flags
8854 adcxq(carry2, carry);
8855 adoxq(carry2, carry);
8856 } else {
8857 add2_with_carry(tmp4, tmp3, carry, yz_idx1);
8858 add2_with_carry(carry2, tmp4, tmp, yz_idx2);
8859 }
8860 movq(carry, carry2);
8861
8862 movl(Address(z, idx, Address::times_4, 12), tmp3);
8863 shrq(tmp3, 32);
8864 movl(Address(z, idx, Address::times_4, 8), tmp3);
8865
8866 movl(Address(z, idx, Address::times_4, 4), tmp4);
8867 shrq(tmp4, 32);
8868 movl(Address(z, idx, Address::times_4, 0), tmp4);
8869
8870 jmp(L_third_loop);
8871
8872 bind (L_third_loop_exit);
8873
8874 andl (idx, 0x3);
8875 jcc(Assembler::zero, L_post_third_loop_done);
8876
8877 Label L_check_1;
8878 subl(idx, 2);
8879 jcc(Assembler::negative, L_check_1);
8880
8881 movq(yz_idx1, Address(y, idx, Address::times_4, 0));
8882 rorxq(yz_idx1, yz_idx1, 32);
8883 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
8884 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
8885 rorxq(yz_idx2, yz_idx2, 32);
8886
8887 add2_with_carry(tmp4, tmp3, carry, yz_idx2);
8888
8889 movl(Address(z, idx, Address::times_4, 4), tmp3);
8890 shrq(tmp3, 32);
8891 movl(Address(z, idx, Address::times_4, 0), tmp3);
8892 movq(carry, tmp4);
8893
8894 bind (L_check_1);
8895 addl (idx, 0x2);
8896 andl (idx, 0x1);
8897 subl(idx, 1);
8898 jcc(Assembler::negative, L_post_third_loop_done);
8899 movl(tmp4, Address(y, idx, Address::times_4, 0));
8900 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3
8901 movl(tmp4, Address(z, idx, Address::times_4, 0));
8902
8903 add2_with_carry(carry2, tmp3, tmp4, carry);
8904
8905 movl(Address(z, idx, Address::times_4, 0), tmp3);
8906 shrq(tmp3, 32);
8907
8908 shlq(carry2, 32);
8909 orq(tmp3, carry2);
8910 movq(carry, tmp3);
8911
8912 bind(L_post_third_loop_done);
8913 }
8914
8915 /**
8916 * Code for BigInteger::multiplyToLen() instrinsic.
8917 *
8918 * rdi: x
8919 * rax: xlen
8920 * rsi: y
8921 * rcx: ylen
8922 * r8: z
8923 * r11: zlen
8924 * r12: tmp1
8925 * r13: tmp2
8926 * r14: tmp3
8927 * r15: tmp4
8928 * rbx: tmp5
8929 *
8930 */
8931 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen,
8932 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) {
8933 ShortBranchVerifier sbv(this);
8934 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx);
8935
8936 push(tmp1);
8937 push(tmp2);
8938 push(tmp3);
8939 push(tmp4);
8940 push(tmp5);
8941
8942 push(xlen);
8943 push(zlen);
8944
8945 const Register idx = tmp1;
8946 const Register kdx = tmp2;
8947 const Register xstart = tmp3;
8948
8949 const Register y_idx = tmp4;
8950 const Register carry = tmp5;
8951 const Register product = xlen;
8952 const Register x_xstart = zlen; // reuse register
8953
8954 // First Loop.
8955 //
8956 // final static long LONG_MASK = 0xffffffffL;
8957 // int xstart = xlen - 1;
8958 // int ystart = ylen - 1;
8959 // long carry = 0;
8960 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
8961 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
8962 // z[kdx] = (int)product;
8963 // carry = product >>> 32;
8964 // }
8965 // z[xstart] = (int)carry;
8966 //
8967
8968 movl(idx, ylen); // idx = ylen;
8969 movl(kdx, zlen); // kdx = xlen+ylen;
8970 xorq(carry, carry); // carry = 0;
8971
8972 Label L_done;
8973
8974 movl(xstart, xlen);
8975 decrementl(xstart);
8976 jcc(Assembler::negative, L_done);
8977
8978 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
8979
8980 Label L_second_loop;
8981 testl(kdx, kdx);
8982 jcc(Assembler::zero, L_second_loop);
8983
8984 Label L_carry;
8985 subl(kdx, 1);
8986 jcc(Assembler::zero, L_carry);
8987
8988 movl(Address(z, kdx, Address::times_4, 0), carry);
8989 shrq(carry, 32);
8990 subl(kdx, 1);
8991
8992 bind(L_carry);
8993 movl(Address(z, kdx, Address::times_4, 0), carry);
8994
8995 // Second and third (nested) loops.
8996 //
8997 // for (int i = xstart-1; i >= 0; i--) { // Second loop
8998 // carry = 0;
8999 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
9000 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
9001 // (z[k] & LONG_MASK) + carry;
9002 // z[k] = (int)product;
9003 // carry = product >>> 32;
9004 // }
9005 // z[i] = (int)carry;
9006 // }
9007 //
9008 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
9009
9010 const Register jdx = tmp1;
9011
9012 bind(L_second_loop);
9013 xorl(carry, carry); // carry = 0;
9014 movl(jdx, ylen); // j = ystart+1
9015
9016 subl(xstart, 1); // i = xstart-1;
9017 jcc(Assembler::negative, L_done);
9018
9019 push (z);
9020
9021 Label L_last_x;
9022 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j
9023 subl(xstart, 1); // i = xstart-1;
9024 jcc(Assembler::negative, L_last_x);
9025
9026 if (UseBMI2Instructions) {
9027 movq(rdx, Address(x, xstart, Address::times_4, 0));
9028 rorxq(rdx, rdx, 32); // convert big-endian to little-endian
9029 } else {
9030 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
9031 rorq(x_xstart, 32); // convert big-endian to little-endian
9032 }
9033
9034 Label L_third_loop_prologue;
9035 bind(L_third_loop_prologue);
9036
9037 push (x);
9038 push (xstart);
9039 push (ylen);
9040
9041
9042 if (UseBMI2Instructions) {
9043 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4);
9044 } else { // !UseBMI2Instructions
9045 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
9046 }
9047
9048 pop(ylen);
9049 pop(xlen);
9050 pop(x);
9051 pop(z);
9052
9053 movl(tmp3, xlen);
9054 addl(tmp3, 1);
9055 movl(Address(z, tmp3, Address::times_4, 0), carry);
9056 subl(tmp3, 1);
9057 jccb(Assembler::negative, L_done);
9058
9059 shrq(carry, 32);
9060 movl(Address(z, tmp3, Address::times_4, 0), carry);
9061 jmp(L_second_loop);
9062
9063 // Next infrequent code is moved outside loops.
9064 bind(L_last_x);
9065 if (UseBMI2Instructions) {
9066 movl(rdx, Address(x, 0));
9067 } else {
9068 movl(x_xstart, Address(x, 0));
9069 }
9070 jmp(L_third_loop_prologue);
9071
9072 bind(L_done);
9073
9074 pop(zlen);
9075 pop(xlen);
9076
9077 pop(tmp5);
9078 pop(tmp4);
9079 pop(tmp3);
9080 pop(tmp2);
9081 pop(tmp1);
9082 }
9083
9084 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale,
9085 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){
9086 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled.");
9087 Label VECTOR64_LOOP, VECTOR64_TAIL, VECTOR64_NOT_EQUAL, VECTOR32_TAIL;
9088 Label VECTOR32_LOOP, VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP;
9089 Label VECTOR16_TAIL, VECTOR8_TAIL, VECTOR4_TAIL;
9090 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL;
9091 Label SAME_TILL_END, DONE;
9092 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL;
9093
9094 //scale is in rcx in both Win64 and Unix
9095 ShortBranchVerifier sbv(this);
9096
9097 shlq(length);
9098 xorq(result, result);
9099
9100 if ((UseAVX > 2) &&
9101 VM_Version::supports_avx512vlbw()) {
9102 set_vector_masking(); // opening of the stub context for programming mask registers
9103 cmpq(length, 64);
9104 jcc(Assembler::less, VECTOR32_TAIL);
9105 movq(tmp1, length);
9106 andq(tmp1, 0x3F); // tail count
9107 andq(length, ~(0x3F)); //vector count
9108
9109 bind(VECTOR64_LOOP);
9110 // AVX512 code to compare 64 byte vectors.
9111 evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit);
9112 evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit);
9113 kortestql(k7, k7);
9114 jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL); // mismatch
9115 addq(result, 64);
9116 subq(length, 64);
9117 jccb(Assembler::notZero, VECTOR64_LOOP);
9118
9119 //bind(VECTOR64_TAIL);
9120 testq(tmp1, tmp1);
9121 jcc(Assembler::zero, SAME_TILL_END);
9122
9123 bind(VECTOR64_TAIL);
9124 // AVX512 code to compare upto 63 byte vectors.
9125 // Save k1
9126 kmovql(k3, k1);
9127 mov64(tmp2, 0xFFFFFFFFFFFFFFFF);
9128 shlxq(tmp2, tmp2, tmp1);
9129 notq(tmp2);
9130 kmovql(k1, tmp2);
9131
9132 evmovdqub(rymm0, k1, Address(obja, result), Assembler::AVX_512bit);
9133 evpcmpeqb(k7, k1, rymm0, Address(objb, result), Assembler::AVX_512bit);
9134
9135 ktestql(k7, k1);
9136 // Restore k1
9137 kmovql(k1, k3);
9138 jcc(Assembler::below, SAME_TILL_END); // not mismatch
9139
9140 bind(VECTOR64_NOT_EQUAL);
9141 kmovql(tmp1, k7);
9142 notq(tmp1);
9143 tzcntq(tmp1, tmp1);
9144 addq(result, tmp1);
9145 shrq(result);
9146 jmp(DONE);
9147 bind(VECTOR32_TAIL);
9148 clear_vector_masking(); // closing of the stub context for programming mask registers
9149 }
9150
9151 cmpq(length, 8);
9152 jcc(Assembler::equal, VECTOR8_LOOP);
9153 jcc(Assembler::less, VECTOR4_TAIL);
9154
9155 if (UseAVX >= 2) {
9156
9157 cmpq(length, 16);
9158 jcc(Assembler::equal, VECTOR16_LOOP);
9159 jcc(Assembler::less, VECTOR8_LOOP);
9160
9161 cmpq(length, 32);
9162 jccb(Assembler::less, VECTOR16_TAIL);
9163
9164 subq(length, 32);
9165 bind(VECTOR32_LOOP);
9166 vmovdqu(rymm0, Address(obja, result));
9167 vmovdqu(rymm1, Address(objb, result));
9168 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit);
9169 vptest(rymm2, rymm2);
9170 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found
9171 addq(result, 32);
9172 subq(length, 32);
9173 jccb(Assembler::greaterEqual, VECTOR32_LOOP);
9174 addq(length, 32);
9175 jcc(Assembler::equal, SAME_TILL_END);
9176 //falling through if less than 32 bytes left //close the branch here.
9177
9178 bind(VECTOR16_TAIL);
9179 cmpq(length, 16);
9180 jccb(Assembler::less, VECTOR8_TAIL);
9181 bind(VECTOR16_LOOP);
9182 movdqu(rymm0, Address(obja, result));
9183 movdqu(rymm1, Address(objb, result));
9184 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit);
9185 ptest(rymm2, rymm2);
9186 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
9187 addq(result, 16);
9188 subq(length, 16);
9189 jcc(Assembler::equal, SAME_TILL_END);
9190 //falling through if less than 16 bytes left
9191 } else {//regular intrinsics
9192
9193 cmpq(length, 16);
9194 jccb(Assembler::less, VECTOR8_TAIL);
9195
9196 subq(length, 16);
9197 bind(VECTOR16_LOOP);
9198 movdqu(rymm0, Address(obja, result));
9199 movdqu(rymm1, Address(objb, result));
9200 pxor(rymm0, rymm1);
9201 ptest(rymm0, rymm0);
9202 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
9203 addq(result, 16);
9204 subq(length, 16);
9205 jccb(Assembler::greaterEqual, VECTOR16_LOOP);
9206 addq(length, 16);
9207 jcc(Assembler::equal, SAME_TILL_END);
9208 //falling through if less than 16 bytes left
9209 }
9210
9211 bind(VECTOR8_TAIL);
9212 cmpq(length, 8);
9213 jccb(Assembler::less, VECTOR4_TAIL);
9214 bind(VECTOR8_LOOP);
9215 movq(tmp1, Address(obja, result));
9216 movq(tmp2, Address(objb, result));
9217 xorq(tmp1, tmp2);
9218 testq(tmp1, tmp1);
9219 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found
9220 addq(result, 8);
9221 subq(length, 8);
9222 jcc(Assembler::equal, SAME_TILL_END);
9223 //falling through if less than 8 bytes left
9224
9225 bind(VECTOR4_TAIL);
9226 cmpq(length, 4);
9227 jccb(Assembler::less, BYTES_TAIL);
9228 bind(VECTOR4_LOOP);
9229 movl(tmp1, Address(obja, result));
9230 xorl(tmp1, Address(objb, result));
9231 testl(tmp1, tmp1);
9232 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found
9233 addq(result, 4);
9234 subq(length, 4);
9235 jcc(Assembler::equal, SAME_TILL_END);
9236 //falling through if less than 4 bytes left
9237
9238 bind(BYTES_TAIL);
9239 bind(BYTES_LOOP);
9240 load_unsigned_byte(tmp1, Address(obja, result));
9241 load_unsigned_byte(tmp2, Address(objb, result));
9242 xorl(tmp1, tmp2);
9243 testl(tmp1, tmp1);
9244 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
9245 decq(length);
9246 jccb(Assembler::zero, SAME_TILL_END);
9247 incq(result);
9248 load_unsigned_byte(tmp1, Address(obja, result));
9249 load_unsigned_byte(tmp2, Address(objb, result));
9250 xorl(tmp1, tmp2);
9251 testl(tmp1, tmp1);
9252 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
9253 decq(length);
9254 jccb(Assembler::zero, SAME_TILL_END);
9255 incq(result);
9256 load_unsigned_byte(tmp1, Address(obja, result));
9257 load_unsigned_byte(tmp2, Address(objb, result));
9258 xorl(tmp1, tmp2);
9259 testl(tmp1, tmp1);
9260 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
9261 jmpb(SAME_TILL_END);
9262
9263 if (UseAVX >= 2) {
9264 bind(VECTOR32_NOT_EQUAL);
9265 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit);
9266 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit);
9267 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit);
9268 vpmovmskb(tmp1, rymm0);
9269 bsfq(tmp1, tmp1);
9270 addq(result, tmp1);
9271 shrq(result);
9272 jmpb(DONE);
9273 }
9274
9275 bind(VECTOR16_NOT_EQUAL);
9276 if (UseAVX >= 2) {
9277 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit);
9278 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit);
9279 pxor(rymm0, rymm2);
9280 } else {
9281 pcmpeqb(rymm2, rymm2);
9282 pxor(rymm0, rymm1);
9283 pcmpeqb(rymm0, rymm1);
9284 pxor(rymm0, rymm2);
9285 }
9286 pmovmskb(tmp1, rymm0);
9287 bsfq(tmp1, tmp1);
9288 addq(result, tmp1);
9289 shrq(result);
9290 jmpb(DONE);
9291
9292 bind(VECTOR8_NOT_EQUAL);
9293 bind(VECTOR4_NOT_EQUAL);
9294 bsfq(tmp1, tmp1);
9295 shrq(tmp1, 3);
9296 addq(result, tmp1);
9297 bind(BYTES_NOT_EQUAL);
9298 shrq(result);
9299 jmpb(DONE);
9300
9301 bind(SAME_TILL_END);
9302 mov64(result, -1);
9303
9304 bind(DONE);
9305 }
9306
9307 //Helper functions for square_to_len()
9308
9309 /**
9310 * Store the squares of x[], right shifted one bit (divided by 2) into z[]
9311 * Preserves x and z and modifies rest of the registers.
9312 */
9313 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
9314 // Perform square and right shift by 1
9315 // Handle odd xlen case first, then for even xlen do the following
9316 // jlong carry = 0;
9317 // for (int j=0, i=0; j < xlen; j+=2, i+=4) {
9318 // huge_128 product = x[j:j+1] * x[j:j+1];
9319 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65);
9320 // z[i+2:i+3] = (jlong)(product >>> 1);
9321 // carry = (jlong)product;
9322 // }
9323
9324 xorq(tmp5, tmp5); // carry
9325 xorq(rdxReg, rdxReg);
9326 xorl(tmp1, tmp1); // index for x
9327 xorl(tmp4, tmp4); // index for z
9328
9329 Label L_first_loop, L_first_loop_exit;
9330
9331 testl(xlen, 1);
9332 jccb(Assembler::zero, L_first_loop); //jump if xlen is even
9333
9334 // Square and right shift by 1 the odd element using 32 bit multiply
9335 movl(raxReg, Address(x, tmp1, Address::times_4, 0));
9336 imulq(raxReg, raxReg);
9337 shrq(raxReg, 1);
9338 adcq(tmp5, 0);
9339 movq(Address(z, tmp4, Address::times_4, 0), raxReg);
9340 incrementl(tmp1);
9341 addl(tmp4, 2);
9342
9343 // Square and right shift by 1 the rest using 64 bit multiply
9344 bind(L_first_loop);
9345 cmpptr(tmp1, xlen);
9346 jccb(Assembler::equal, L_first_loop_exit);
9347
9348 // Square
9349 movq(raxReg, Address(x, tmp1, Address::times_4, 0));
9350 rorq(raxReg, 32); // convert big-endian to little-endian
9351 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax
9352
9353 // Right shift by 1 and save carry
9354 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1
9355 rcrq(rdxReg, 1);
9356 rcrq(raxReg, 1);
9357 adcq(tmp5, 0);
9358
9359 // Store result in z
9360 movq(Address(z, tmp4, Address::times_4, 0), rdxReg);
9361 movq(Address(z, tmp4, Address::times_4, 8), raxReg);
9362
9363 // Update indices for x and z
9364 addl(tmp1, 2);
9365 addl(tmp4, 4);
9366 jmp(L_first_loop);
9367
9368 bind(L_first_loop_exit);
9369 }
9370
9371
9372 /**
9373 * Perform the following multiply add operation using BMI2 instructions
9374 * carry:sum = sum + op1*op2 + carry
9375 * op2 should be in rdx
9376 * op2 is preserved, all other registers are modified
9377 */
9378 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) {
9379 // assert op2 is rdx
9380 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1
9381 addq(sum, carry);
9382 adcq(tmp2, 0);
9383 addq(sum, op1);
9384 adcq(tmp2, 0);
9385 movq(carry, tmp2);
9386 }
9387
9388 /**
9389 * Perform the following multiply add operation:
9390 * carry:sum = sum + op1*op2 + carry
9391 * Preserves op1, op2 and modifies rest of registers
9392 */
9393 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) {
9394 // rdx:rax = op1 * op2
9395 movq(raxReg, op2);
9396 mulq(op1);
9397
9398 // rdx:rax = sum + carry + rdx:rax
9399 addq(sum, carry);
9400 adcq(rdxReg, 0);
9401 addq(sum, raxReg);
9402 adcq(rdxReg, 0);
9403
9404 // carry:sum = rdx:sum
9405 movq(carry, rdxReg);
9406 }
9407
9408 /**
9409 * Add 64 bit long carry into z[] with carry propogation.
9410 * Preserves z and carry register values and modifies rest of registers.
9411 *
9412 */
9413 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) {
9414 Label L_fourth_loop, L_fourth_loop_exit;
9415
9416 movl(tmp1, 1);
9417 subl(zlen, 2);
9418 addq(Address(z, zlen, Address::times_4, 0), carry);
9419
9420 bind(L_fourth_loop);
9421 jccb(Assembler::carryClear, L_fourth_loop_exit);
9422 subl(zlen, 2);
9423 jccb(Assembler::negative, L_fourth_loop_exit);
9424 addq(Address(z, zlen, Address::times_4, 0), tmp1);
9425 jmp(L_fourth_loop);
9426 bind(L_fourth_loop_exit);
9427 }
9428
9429 /**
9430 * Shift z[] left by 1 bit.
9431 * Preserves x, len, z and zlen registers and modifies rest of the registers.
9432 *
9433 */
9434 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) {
9435
9436 Label L_fifth_loop, L_fifth_loop_exit;
9437
9438 // Fifth loop
9439 // Perform primitiveLeftShift(z, zlen, 1)
9440
9441 const Register prev_carry = tmp1;
9442 const Register new_carry = tmp4;
9443 const Register value = tmp2;
9444 const Register zidx = tmp3;
9445
9446 // int zidx, carry;
9447 // long value;
9448 // carry = 0;
9449 // for (zidx = zlen-2; zidx >=0; zidx -= 2) {
9450 // (carry:value) = (z[i] << 1) | carry ;
9451 // z[i] = value;
9452 // }
9453
9454 movl(zidx, zlen);
9455 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register
9456
9457 bind(L_fifth_loop);
9458 decl(zidx); // Use decl to preserve carry flag
9459 decl(zidx);
9460 jccb(Assembler::negative, L_fifth_loop_exit);
9461
9462 if (UseBMI2Instructions) {
9463 movq(value, Address(z, zidx, Address::times_4, 0));
9464 rclq(value, 1);
9465 rorxq(value, value, 32);
9466 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
9467 }
9468 else {
9469 // clear new_carry
9470 xorl(new_carry, new_carry);
9471
9472 // Shift z[i] by 1, or in previous carry and save new carry
9473 movq(value, Address(z, zidx, Address::times_4, 0));
9474 shlq(value, 1);
9475 adcl(new_carry, 0);
9476
9477 orq(value, prev_carry);
9478 rorq(value, 0x20);
9479 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
9480
9481 // Set previous carry = new carry
9482 movl(prev_carry, new_carry);
9483 }
9484 jmp(L_fifth_loop);
9485
9486 bind(L_fifth_loop_exit);
9487 }
9488
9489
9490 /**
9491 * Code for BigInteger::squareToLen() intrinsic
9492 *
9493 * rdi: x
9494 * rsi: len
9495 * r8: z
9496 * rcx: zlen
9497 * r12: tmp1
9498 * r13: tmp2
9499 * r14: tmp3
9500 * r15: tmp4
9501 * rbx: tmp5
9502 *
9503 */
9504 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) {
9505
9506 Label L_second_loop, L_second_loop_exit, L_third_loop, L_third_loop_exit, fifth_loop, fifth_loop_exit, L_last_x, L_multiply;
9507 push(tmp1);
9508 push(tmp2);
9509 push(tmp3);
9510 push(tmp4);
9511 push(tmp5);
9512
9513 // First loop
9514 // Store the squares, right shifted one bit (i.e., divided by 2).
9515 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg);
9516
9517 // Add in off-diagonal sums.
9518 //
9519 // Second, third (nested) and fourth loops.
9520 // zlen +=2;
9521 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) {
9522 // carry = 0;
9523 // long op2 = x[xidx:xidx+1];
9524 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) {
9525 // k -= 2;
9526 // long op1 = x[j:j+1];
9527 // long sum = z[k:k+1];
9528 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs);
9529 // z[k:k+1] = sum;
9530 // }
9531 // add_one_64(z, k, carry, tmp_regs);
9532 // }
9533
9534 const Register carry = tmp5;
9535 const Register sum = tmp3;
9536 const Register op1 = tmp4;
9537 Register op2 = tmp2;
9538
9539 push(zlen);
9540 push(len);
9541 addl(zlen,2);
9542 bind(L_second_loop);
9543 xorq(carry, carry);
9544 subl(zlen, 4);
9545 subl(len, 2);
9546 push(zlen);
9547 push(len);
9548 cmpl(len, 0);
9549 jccb(Assembler::lessEqual, L_second_loop_exit);
9550
9551 // Multiply an array by one 64 bit long.
9552 if (UseBMI2Instructions) {
9553 op2 = rdxReg;
9554 movq(op2, Address(x, len, Address::times_4, 0));
9555 rorxq(op2, op2, 32);
9556 }
9557 else {
9558 movq(op2, Address(x, len, Address::times_4, 0));
9559 rorq(op2, 32);
9560 }
9561
9562 bind(L_third_loop);
9563 decrementl(len);
9564 jccb(Assembler::negative, L_third_loop_exit);
9565 decrementl(len);
9566 jccb(Assembler::negative, L_last_x);
9567
9568 movq(op1, Address(x, len, Address::times_4, 0));
9569 rorq(op1, 32);
9570
9571 bind(L_multiply);
9572 subl(zlen, 2);
9573 movq(sum, Address(z, zlen, Address::times_4, 0));
9574
9575 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry.
9576 if (UseBMI2Instructions) {
9577 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2);
9578 }
9579 else {
9580 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9581 }
9582
9583 movq(Address(z, zlen, Address::times_4, 0), sum);
9584
9585 jmp(L_third_loop);
9586 bind(L_third_loop_exit);
9587
9588 // Fourth loop
9589 // Add 64 bit long carry into z with carry propogation.
9590 // Uses offsetted zlen.
9591 add_one_64(z, zlen, carry, tmp1);
9592
9593 pop(len);
9594 pop(zlen);
9595 jmp(L_second_loop);
9596
9597 // Next infrequent code is moved outside loops.
9598 bind(L_last_x);
9599 movl(op1, Address(x, 0));
9600 jmp(L_multiply);
9601
9602 bind(L_second_loop_exit);
9603 pop(len);
9604 pop(zlen);
9605 pop(len);
9606 pop(zlen);
9607
9608 // Fifth loop
9609 // Shift z left 1 bit.
9610 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4);
9611
9612 // z[zlen-1] |= x[len-1] & 1;
9613 movl(tmp3, Address(x, len, Address::times_4, -4));
9614 andl(tmp3, 1);
9615 orl(Address(z, zlen, Address::times_4, -4), tmp3);
9616
9617 pop(tmp5);
9618 pop(tmp4);
9619 pop(tmp3);
9620 pop(tmp2);
9621 pop(tmp1);
9622 }
9623
9624 /**
9625 * Helper function for mul_add()
9626 * Multiply the in[] by int k and add to out[] starting at offset offs using
9627 * 128 bit by 32 bit multiply and return the carry in tmp5.
9628 * Only quad int aligned length of in[] is operated on in this function.
9629 * k is in rdxReg for BMI2Instructions, for others it is in tmp2.
9630 * This function preserves out, in and k registers.
9631 * len and offset point to the appropriate index in "in" & "out" correspondingly
9632 * tmp5 has the carry.
9633 * other registers are temporary and are modified.
9634 *
9635 */
9636 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in,
9637 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3,
9638 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
9639
9640 Label L_first_loop, L_first_loop_exit;
9641
9642 movl(tmp1, len);
9643 shrl(tmp1, 2);
9644
9645 bind(L_first_loop);
9646 subl(tmp1, 1);
9647 jccb(Assembler::negative, L_first_loop_exit);
9648
9649 subl(len, 4);
9650 subl(offset, 4);
9651
9652 Register op2 = tmp2;
9653 const Register sum = tmp3;
9654 const Register op1 = tmp4;
9655 const Register carry = tmp5;
9656
9657 if (UseBMI2Instructions) {
9658 op2 = rdxReg;
9659 }
9660
9661 movq(op1, Address(in, len, Address::times_4, 8));
9662 rorq(op1, 32);
9663 movq(sum, Address(out, offset, Address::times_4, 8));
9664 rorq(sum, 32);
9665 if (UseBMI2Instructions) {
9666 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
9667 }
9668 else {
9669 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9670 }
9671 // Store back in big endian from little endian
9672 rorq(sum, 0x20);
9673 movq(Address(out, offset, Address::times_4, 8), sum);
9674
9675 movq(op1, Address(in, len, Address::times_4, 0));
9676 rorq(op1, 32);
9677 movq(sum, Address(out, offset, Address::times_4, 0));
9678 rorq(sum, 32);
9679 if (UseBMI2Instructions) {
9680 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
9681 }
9682 else {
9683 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9684 }
9685 // Store back in big endian from little endian
9686 rorq(sum, 0x20);
9687 movq(Address(out, offset, Address::times_4, 0), sum);
9688
9689 jmp(L_first_loop);
9690 bind(L_first_loop_exit);
9691 }
9692
9693 /**
9694 * Code for BigInteger::mulAdd() intrinsic
9695 *
9696 * rdi: out
9697 * rsi: in
9698 * r11: offs (out.length - offset)
9699 * rcx: len
9700 * r8: k
9701 * r12: tmp1
9702 * r13: tmp2
9703 * r14: tmp3
9704 * r15: tmp4
9705 * rbx: tmp5
9706 * Multiply the in[] by word k and add to out[], return the carry in rax
9707 */
9708 void MacroAssembler::mul_add(Register out, Register in, Register offs,
9709 Register len, Register k, Register tmp1, Register tmp2, Register tmp3,
9710 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
9711
9712 Label L_carry, L_last_in, L_done;
9713
9714 // carry = 0;
9715 // for (int j=len-1; j >= 0; j--) {
9716 // long product = (in[j] & LONG_MASK) * kLong +
9717 // (out[offs] & LONG_MASK) + carry;
9718 // out[offs--] = (int)product;
9719 // carry = product >>> 32;
9720 // }
9721 //
9722 push(tmp1);
9723 push(tmp2);
9724 push(tmp3);
9725 push(tmp4);
9726 push(tmp5);
9727
9728 Register op2 = tmp2;
9729 const Register sum = tmp3;
9730 const Register op1 = tmp4;
9731 const Register carry = tmp5;
9732
9733 if (UseBMI2Instructions) {
9734 op2 = rdxReg;
9735 movl(op2, k);
9736 }
9737 else {
9738 movl(op2, k);
9739 }
9740
9741 xorq(carry, carry);
9742
9743 //First loop
9744
9745 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply
9746 //The carry is in tmp5
9747 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg);
9748
9749 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any
9750 decrementl(len);
9751 jccb(Assembler::negative, L_carry);
9752 decrementl(len);
9753 jccb(Assembler::negative, L_last_in);
9754
9755 movq(op1, Address(in, len, Address::times_4, 0));
9756 rorq(op1, 32);
9757
9758 subl(offs, 2);
9759 movq(sum, Address(out, offs, Address::times_4, 0));
9760 rorq(sum, 32);
9761
9762 if (UseBMI2Instructions) {
9763 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
9764 }
9765 else {
9766 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9767 }
9768
9769 // Store back in big endian from little endian
9770 rorq(sum, 0x20);
9771 movq(Address(out, offs, Address::times_4, 0), sum);
9772
9773 testl(len, len);
9774 jccb(Assembler::zero, L_carry);
9775
9776 //Multiply the last in[] entry, if any
9777 bind(L_last_in);
9778 movl(op1, Address(in, 0));
9779 movl(sum, Address(out, offs, Address::times_4, -4));
9780
9781 movl(raxReg, k);
9782 mull(op1); //tmp4 * eax -> edx:eax
9783 addl(sum, carry);
9784 adcl(rdxReg, 0);
9785 addl(sum, raxReg);
9786 adcl(rdxReg, 0);
9787 movl(carry, rdxReg);
9788
9789 movl(Address(out, offs, Address::times_4, -4), sum);
9790
9791 bind(L_carry);
9792 //return tmp5/carry as carry in rax
9793 movl(rax, carry);
9794
9795 bind(L_done);
9796 pop(tmp5);
9797 pop(tmp4);
9798 pop(tmp3);
9799 pop(tmp2);
9800 pop(tmp1);
9801 }
9802 #endif
9803
9804 /**
9805 * Emits code to update CRC-32 with a byte value according to constants in table
9806 *
9807 * @param [in,out]crc Register containing the crc.
9808 * @param [in]val Register containing the byte to fold into the CRC.
9809 * @param [in]table Register containing the table of crc constants.
9810 *
9811 * uint32_t crc;
9812 * val = crc_table[(val ^ crc) & 0xFF];
9813 * crc = val ^ (crc >> 8);
9814 *
9815 */
9816 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
9817 xorl(val, crc);
9818 andl(val, 0xFF);
9819 shrl(crc, 8); // unsigned shift
9820 xorl(crc, Address(table, val, Address::times_4, 0));
9821 }
9822
9823 /**
9824 * Fold four 128-bit data chunks
9825 */
9826 void MacroAssembler::fold_128bit_crc32_avx512(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
9827 evpclmulhdq(xtmp, xK, xcrc, Assembler::AVX_512bit); // [123:64]
9828 evpclmulldq(xcrc, xK, xcrc, Assembler::AVX_512bit); // [63:0]
9829 evpxorq(xcrc, xcrc, Address(buf, offset), Assembler::AVX_512bit /* vector_len */);
9830 evpxorq(xcrc, xcrc, xtmp, Assembler::AVX_512bit /* vector_len */);
9831 }
9832
9833 /**
9834 * Fold 128-bit data chunk
9835 */
9836 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
9837 if (UseAVX > 0) {
9838 vpclmulhdq(xtmp, xK, xcrc); // [123:64]
9839 vpclmulldq(xcrc, xK, xcrc); // [63:0]
9840 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */);
9841 pxor(xcrc, xtmp);
9842 } else {
9843 movdqa(xtmp, xcrc);
9844 pclmulhdq(xtmp, xK); // [123:64]
9845 pclmulldq(xcrc, xK); // [63:0]
9846 pxor(xcrc, xtmp);
9847 movdqu(xtmp, Address(buf, offset));
9848 pxor(xcrc, xtmp);
9849 }
9850 }
9851
9852 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
9853 if (UseAVX > 0) {
9854 vpclmulhdq(xtmp, xK, xcrc);
9855 vpclmulldq(xcrc, xK, xcrc);
9856 pxor(xcrc, xbuf);
9857 pxor(xcrc, xtmp);
9858 } else {
9859 movdqa(xtmp, xcrc);
9860 pclmulhdq(xtmp, xK);
9861 pclmulldq(xcrc, xK);
9862 pxor(xcrc, xbuf);
9863 pxor(xcrc, xtmp);
9864 }
9865 }
9866
9867 /**
9868 * 8-bit folds to compute 32-bit CRC
9869 *
9870 * uint64_t xcrc;
9871 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8);
9872 */
9873 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) {
9874 movdl(tmp, xcrc);
9875 andl(tmp, 0xFF);
9876 movdl(xtmp, Address(table, tmp, Address::times_4, 0));
9877 psrldq(xcrc, 1); // unsigned shift one byte
9878 pxor(xcrc, xtmp);
9879 }
9880
9881 /**
9882 * uint32_t crc;
9883 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
9884 */
9885 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
9886 movl(tmp, crc);
9887 andl(tmp, 0xFF);
9888 shrl(crc, 8);
9889 xorl(crc, Address(table, tmp, Address::times_4, 0));
9890 }
9891
9892 /**
9893 * @param crc register containing existing CRC (32-bit)
9894 * @param buf register pointing to input byte buffer (byte*)
9895 * @param len register containing number of bytes
9896 * @param table register that will contain address of CRC table
9897 * @param tmp scratch register
9898 */
9899 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) {
9900 assert_different_registers(crc, buf, len, table, tmp, rax);
9901
9902 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned;
9903 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop;
9904
9905 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge
9906 // context for the registers used, where all instructions below are using 128-bit mode
9907 // On EVEX without VL and BW, these instructions will all be AVX.
9908 if (VM_Version::supports_avx512vlbw()) {
9909 movl(tmp, 0xffff);
9910 kmovwl(k1, tmp);
9911 }
9912
9913 lea(table, ExternalAddress(StubRoutines::crc_table_addr()));
9914 notl(crc); // ~crc
9915 cmpl(len, 16);
9916 jcc(Assembler::less, L_tail);
9917
9918 // Align buffer to 16 bytes
9919 movl(tmp, buf);
9920 andl(tmp, 0xF);
9921 jccb(Assembler::zero, L_aligned);
9922 subl(tmp, 16);
9923 addl(len, tmp);
9924
9925 align(4);
9926 BIND(L_align_loop);
9927 movsbl(rax, Address(buf, 0)); // load byte with sign extension
9928 update_byte_crc32(crc, rax, table);
9929 increment(buf);
9930 incrementl(tmp);
9931 jccb(Assembler::less, L_align_loop);
9932
9933 BIND(L_aligned);
9934 movl(tmp, len); // save
9935 shrl(len, 4);
9936 jcc(Assembler::zero, L_tail_restore);
9937
9938 // Fold total 512 bits of polynomial on each iteration
9939 if (VM_Version::supports_vpclmulqdq()) {
9940 Label Parallel_loop, L_No_Parallel;
9941
9942 cmpl(len, 8);
9943 jccb(Assembler::less, L_No_Parallel);
9944
9945 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
9946 evmovdquq(xmm1, Address(buf, 0), Assembler::AVX_512bit);
9947 movdl(xmm5, crc);
9948 evpxorq(xmm1, xmm1, xmm5, Assembler::AVX_512bit);
9949 addptr(buf, 64);
9950 subl(len, 7);
9951 evshufi64x2(xmm0, xmm0, xmm0, 0x00, Assembler::AVX_512bit); //propagate the mask from 128 bits to 512 bits
9952
9953 BIND(Parallel_loop);
9954 fold_128bit_crc32_avx512(xmm1, xmm0, xmm5, buf, 0);
9955 addptr(buf, 64);
9956 subl(len, 4);
9957 jcc(Assembler::greater, Parallel_loop);
9958
9959 vextracti64x2(xmm2, xmm1, 0x01);
9960 vextracti64x2(xmm3, xmm1, 0x02);
9961 vextracti64x2(xmm4, xmm1, 0x03);
9962 jmp(L_fold_512b);
9963
9964 BIND(L_No_Parallel);
9965 }
9966 // Fold crc into first bytes of vector
9967 movdqa(xmm1, Address(buf, 0));
9968 movdl(rax, xmm1);
9969 xorl(crc, rax);
9970 if (VM_Version::supports_sse4_1()) {
9971 pinsrd(xmm1, crc, 0);
9972 } else {
9973 pinsrw(xmm1, crc, 0);
9974 shrl(crc, 16);
9975 pinsrw(xmm1, crc, 1);
9976 }
9977 addptr(buf, 16);
9978 subl(len, 4); // len > 0
9979 jcc(Assembler::less, L_fold_tail);
9980
9981 movdqa(xmm2, Address(buf, 0));
9982 movdqa(xmm3, Address(buf, 16));
9983 movdqa(xmm4, Address(buf, 32));
9984 addptr(buf, 48);
9985 subl(len, 3);
9986 jcc(Assembler::lessEqual, L_fold_512b);
9987
9988 // Fold total 512 bits of polynomial on each iteration,
9989 // 128 bits per each of 4 parallel streams.
9990 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
9991
9992 align(32);
9993 BIND(L_fold_512b_loop);
9994 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
9995 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16);
9996 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32);
9997 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48);
9998 addptr(buf, 64);
9999 subl(len, 4);
10000 jcc(Assembler::greater, L_fold_512b_loop);
10001
10002 // Fold 512 bits to 128 bits.
10003 BIND(L_fold_512b);
10004 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
10005 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2);
10006 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3);
10007 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4);
10008
10009 // Fold the rest of 128 bits data chunks
10010 BIND(L_fold_tail);
10011 addl(len, 3);
10012 jccb(Assembler::lessEqual, L_fold_128b);
10013 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
10014
10015 BIND(L_fold_tail_loop);
10016 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
10017 addptr(buf, 16);
10018 decrementl(len);
10019 jccb(Assembler::greater, L_fold_tail_loop);
10020
10021 // Fold 128 bits in xmm1 down into 32 bits in crc register.
10022 BIND(L_fold_128b);
10023 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
10024 if (UseAVX > 0) {
10025 vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
10026 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */);
10027 vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
10028 } else {
10029 movdqa(xmm2, xmm0);
10030 pclmulqdq(xmm2, xmm1, 0x1);
10031 movdqa(xmm3, xmm0);
10032 pand(xmm3, xmm2);
10033 pclmulqdq(xmm0, xmm3, 0x1);
10034 }
10035 psrldq(xmm1, 8);
10036 psrldq(xmm2, 4);
10037 pxor(xmm0, xmm1);
10038 pxor(xmm0, xmm2);
10039
10040 // 8 8-bit folds to compute 32-bit CRC.
10041 for (int j = 0; j < 4; j++) {
10042 fold_8bit_crc32(xmm0, table, xmm1, rax);
10043 }
10044 movdl(crc, xmm0); // mov 32 bits to general register
10045 for (int j = 0; j < 4; j++) {
10046 fold_8bit_crc32(crc, table, rax);
10047 }
10048
10049 BIND(L_tail_restore);
10050 movl(len, tmp); // restore
10051 BIND(L_tail);
10052 andl(len, 0xf);
10053 jccb(Assembler::zero, L_exit);
10054
10055 // Fold the rest of bytes
10056 align(4);
10057 BIND(L_tail_loop);
10058 movsbl(rax, Address(buf, 0)); // load byte with sign extension
10059 update_byte_crc32(crc, rax, table);
10060 increment(buf);
10061 decrementl(len);
10062 jccb(Assembler::greater, L_tail_loop);
10063
10064 BIND(L_exit);
10065 notl(crc); // ~c
10066 }
10067
10068 #ifdef _LP64
10069 // S. Gueron / Information Processing Letters 112 (2012) 184
10070 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table.
10071 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0].
10072 // Output: the 64-bit carry-less product of B * CONST
10073 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n,
10074 Register tmp1, Register tmp2, Register tmp3) {
10075 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
10076 if (n > 0) {
10077 addq(tmp3, n * 256 * 8);
10078 }
10079 // Q1 = TABLEExt[n][B & 0xFF];
10080 movl(tmp1, in);
10081 andl(tmp1, 0x000000FF);
10082 shll(tmp1, 3);
10083 addq(tmp1, tmp3);
10084 movq(tmp1, Address(tmp1, 0));
10085
10086 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
10087 movl(tmp2, in);
10088 shrl(tmp2, 8);
10089 andl(tmp2, 0x000000FF);
10090 shll(tmp2, 3);
10091 addq(tmp2, tmp3);
10092 movq(tmp2, Address(tmp2, 0));
10093
10094 shlq(tmp2, 8);
10095 xorq(tmp1, tmp2);
10096
10097 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
10098 movl(tmp2, in);
10099 shrl(tmp2, 16);
10100 andl(tmp2, 0x000000FF);
10101 shll(tmp2, 3);
10102 addq(tmp2, tmp3);
10103 movq(tmp2, Address(tmp2, 0));
10104
10105 shlq(tmp2, 16);
10106 xorq(tmp1, tmp2);
10107
10108 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
10109 shrl(in, 24);
10110 andl(in, 0x000000FF);
10111 shll(in, 3);
10112 addq(in, tmp3);
10113 movq(in, Address(in, 0));
10114
10115 shlq(in, 24);
10116 xorq(in, tmp1);
10117 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
10118 }
10119
10120 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
10121 Register in_out,
10122 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
10123 XMMRegister w_xtmp2,
10124 Register tmp1,
10125 Register n_tmp2, Register n_tmp3) {
10126 if (is_pclmulqdq_supported) {
10127 movdl(w_xtmp1, in_out); // modified blindly
10128
10129 movl(tmp1, const_or_pre_comp_const_index);
10130 movdl(w_xtmp2, tmp1);
10131 pclmulqdq(w_xtmp1, w_xtmp2, 0);
10132
10133 movdq(in_out, w_xtmp1);
10134 } else {
10135 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3);
10136 }
10137 }
10138
10139 // Recombination Alternative 2: No bit-reflections
10140 // T1 = (CRC_A * U1) << 1
10141 // T2 = (CRC_B * U2) << 1
10142 // C1 = T1 >> 32
10143 // C2 = T2 >> 32
10144 // T1 = T1 & 0xFFFFFFFF
10145 // T2 = T2 & 0xFFFFFFFF
10146 // T1 = CRC32(0, T1)
10147 // T2 = CRC32(0, T2)
10148 // C1 = C1 ^ T1
10149 // C2 = C2 ^ T2
10150 // CRC = C1 ^ C2 ^ CRC_C
10151 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,
10152 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10153 Register tmp1, Register tmp2,
10154 Register n_tmp3) {
10155 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10156 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10157 shlq(in_out, 1);
10158 movl(tmp1, in_out);
10159 shrq(in_out, 32);
10160 xorl(tmp2, tmp2);
10161 crc32(tmp2, tmp1, 4);
10162 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here
10163 shlq(in1, 1);
10164 movl(tmp1, in1);
10165 shrq(in1, 32);
10166 xorl(tmp2, tmp2);
10167 crc32(tmp2, tmp1, 4);
10168 xorl(in1, tmp2);
10169 xorl(in_out, in1);
10170 xorl(in_out, in2);
10171 }
10172
10173 // Set N to predefined value
10174 // Subtract from a lenght of a buffer
10175 // execute in a loop:
10176 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0
10177 // for i = 1 to N do
10178 // CRC_A = CRC32(CRC_A, A[i])
10179 // CRC_B = CRC32(CRC_B, B[i])
10180 // CRC_C = CRC32(CRC_C, C[i])
10181 // end for
10182 // Recombine
10183 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,
10184 Register in_out1, Register in_out2, Register in_out3,
10185 Register tmp1, Register tmp2, Register tmp3,
10186 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10187 Register tmp4, Register tmp5,
10188 Register n_tmp6) {
10189 Label L_processPartitions;
10190 Label L_processPartition;
10191 Label L_exit;
10192
10193 bind(L_processPartitions);
10194 cmpl(in_out1, 3 * size);
10195 jcc(Assembler::less, L_exit);
10196 xorl(tmp1, tmp1);
10197 xorl(tmp2, tmp2);
10198 movq(tmp3, in_out2);
10199 addq(tmp3, size);
10200
10201 bind(L_processPartition);
10202 crc32(in_out3, Address(in_out2, 0), 8);
10203 crc32(tmp1, Address(in_out2, size), 8);
10204 crc32(tmp2, Address(in_out2, size * 2), 8);
10205 addq(in_out2, 8);
10206 cmpq(in_out2, tmp3);
10207 jcc(Assembler::less, L_processPartition);
10208 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
10209 w_xtmp1, w_xtmp2, w_xtmp3,
10210 tmp4, tmp5,
10211 n_tmp6);
10212 addq(in_out2, 2 * size);
10213 subl(in_out1, 3 * size);
10214 jmp(L_processPartitions);
10215
10216 bind(L_exit);
10217 }
10218 #else
10219 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n,
10220 Register tmp1, Register tmp2, Register tmp3,
10221 XMMRegister xtmp1, XMMRegister xtmp2) {
10222 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
10223 if (n > 0) {
10224 addl(tmp3, n * 256 * 8);
10225 }
10226 // Q1 = TABLEExt[n][B & 0xFF];
10227 movl(tmp1, in_out);
10228 andl(tmp1, 0x000000FF);
10229 shll(tmp1, 3);
10230 addl(tmp1, tmp3);
10231 movq(xtmp1, Address(tmp1, 0));
10232
10233 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
10234 movl(tmp2, in_out);
10235 shrl(tmp2, 8);
10236 andl(tmp2, 0x000000FF);
10237 shll(tmp2, 3);
10238 addl(tmp2, tmp3);
10239 movq(xtmp2, Address(tmp2, 0));
10240
10241 psllq(xtmp2, 8);
10242 pxor(xtmp1, xtmp2);
10243
10244 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
10245 movl(tmp2, in_out);
10246 shrl(tmp2, 16);
10247 andl(tmp2, 0x000000FF);
10248 shll(tmp2, 3);
10249 addl(tmp2, tmp3);
10250 movq(xtmp2, Address(tmp2, 0));
10251
10252 psllq(xtmp2, 16);
10253 pxor(xtmp1, xtmp2);
10254
10255 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
10256 shrl(in_out, 24);
10257 andl(in_out, 0x000000FF);
10258 shll(in_out, 3);
10259 addl(in_out, tmp3);
10260 movq(xtmp2, Address(in_out, 0));
10261
10262 psllq(xtmp2, 24);
10263 pxor(xtmp1, xtmp2); // Result in CXMM
10264 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
10265 }
10266
10267 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
10268 Register in_out,
10269 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
10270 XMMRegister w_xtmp2,
10271 Register tmp1,
10272 Register n_tmp2, Register n_tmp3) {
10273 if (is_pclmulqdq_supported) {
10274 movdl(w_xtmp1, in_out);
10275
10276 movl(tmp1, const_or_pre_comp_const_index);
10277 movdl(w_xtmp2, tmp1);
10278 pclmulqdq(w_xtmp1, w_xtmp2, 0);
10279 // Keep result in XMM since GPR is 32 bit in length
10280 } else {
10281 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2);
10282 }
10283 }
10284
10285 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,
10286 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10287 Register tmp1, Register tmp2,
10288 Register n_tmp3) {
10289 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10290 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10291
10292 psllq(w_xtmp1, 1);
10293 movdl(tmp1, w_xtmp1);
10294 psrlq(w_xtmp1, 32);
10295 movdl(in_out, w_xtmp1);
10296
10297 xorl(tmp2, tmp2);
10298 crc32(tmp2, tmp1, 4);
10299 xorl(in_out, tmp2);
10300
10301 psllq(w_xtmp2, 1);
10302 movdl(tmp1, w_xtmp2);
10303 psrlq(w_xtmp2, 32);
10304 movdl(in1, w_xtmp2);
10305
10306 xorl(tmp2, tmp2);
10307 crc32(tmp2, tmp1, 4);
10308 xorl(in1, tmp2);
10309 xorl(in_out, in1);
10310 xorl(in_out, in2);
10311 }
10312
10313 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,
10314 Register in_out1, Register in_out2, Register in_out3,
10315 Register tmp1, Register tmp2, Register tmp3,
10316 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10317 Register tmp4, Register tmp5,
10318 Register n_tmp6) {
10319 Label L_processPartitions;
10320 Label L_processPartition;
10321 Label L_exit;
10322
10323 bind(L_processPartitions);
10324 cmpl(in_out1, 3 * size);
10325 jcc(Assembler::less, L_exit);
10326 xorl(tmp1, tmp1);
10327 xorl(tmp2, tmp2);
10328 movl(tmp3, in_out2);
10329 addl(tmp3, size);
10330
10331 bind(L_processPartition);
10332 crc32(in_out3, Address(in_out2, 0), 4);
10333 crc32(tmp1, Address(in_out2, size), 4);
10334 crc32(tmp2, Address(in_out2, size*2), 4);
10335 crc32(in_out3, Address(in_out2, 0+4), 4);
10336 crc32(tmp1, Address(in_out2, size+4), 4);
10337 crc32(tmp2, Address(in_out2, size*2+4), 4);
10338 addl(in_out2, 8);
10339 cmpl(in_out2, tmp3);
10340 jcc(Assembler::less, L_processPartition);
10341
10342 push(tmp3);
10343 push(in_out1);
10344 push(in_out2);
10345 tmp4 = tmp3;
10346 tmp5 = in_out1;
10347 n_tmp6 = in_out2;
10348
10349 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
10350 w_xtmp1, w_xtmp2, w_xtmp3,
10351 tmp4, tmp5,
10352 n_tmp6);
10353
10354 pop(in_out2);
10355 pop(in_out1);
10356 pop(tmp3);
10357
10358 addl(in_out2, 2 * size);
10359 subl(in_out1, 3 * size);
10360 jmp(L_processPartitions);
10361
10362 bind(L_exit);
10363 }
10364 #endif //LP64
10365
10366 #ifdef _LP64
10367 // Algorithm 2: Pipelined usage of the CRC32 instruction.
10368 // Input: A buffer I of L bytes.
10369 // Output: the CRC32C value of the buffer.
10370 // Notations:
10371 // Write L = 24N + r, with N = floor (L/24).
10372 // r = L mod 24 (0 <= r < 24).
10373 // Consider I as the concatenation of A|B|C|R, where A, B, C, each,
10374 // N quadwords, and R consists of r bytes.
10375 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1
10376 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1
10377 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1
10378 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1
10379 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
10380 Register tmp1, Register tmp2, Register tmp3,
10381 Register tmp4, Register tmp5, Register tmp6,
10382 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10383 bool is_pclmulqdq_supported) {
10384 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
10385 Label L_wordByWord;
10386 Label L_byteByByteProlog;
10387 Label L_byteByByte;
10388 Label L_exit;
10389
10390 if (is_pclmulqdq_supported ) {
10391 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
10392 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1);
10393
10394 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
10395 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
10396
10397 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
10398 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
10399 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\"");
10400 } else {
10401 const_or_pre_comp_const_index[0] = 1;
10402 const_or_pre_comp_const_index[1] = 0;
10403
10404 const_or_pre_comp_const_index[2] = 3;
10405 const_or_pre_comp_const_index[3] = 2;
10406
10407 const_or_pre_comp_const_index[4] = 5;
10408 const_or_pre_comp_const_index[5] = 4;
10409 }
10410 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
10411 in2, in1, in_out,
10412 tmp1, tmp2, tmp3,
10413 w_xtmp1, w_xtmp2, w_xtmp3,
10414 tmp4, tmp5,
10415 tmp6);
10416 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
10417 in2, in1, in_out,
10418 tmp1, tmp2, tmp3,
10419 w_xtmp1, w_xtmp2, w_xtmp3,
10420 tmp4, tmp5,
10421 tmp6);
10422 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
10423 in2, in1, in_out,
10424 tmp1, tmp2, tmp3,
10425 w_xtmp1, w_xtmp2, w_xtmp3,
10426 tmp4, tmp5,
10427 tmp6);
10428 movl(tmp1, in2);
10429 andl(tmp1, 0x00000007);
10430 negl(tmp1);
10431 addl(tmp1, in2);
10432 addq(tmp1, in1);
10433
10434 BIND(L_wordByWord);
10435 cmpq(in1, tmp1);
10436 jcc(Assembler::greaterEqual, L_byteByByteProlog);
10437 crc32(in_out, Address(in1, 0), 4);
10438 addq(in1, 4);
10439 jmp(L_wordByWord);
10440
10441 BIND(L_byteByByteProlog);
10442 andl(in2, 0x00000007);
10443 movl(tmp2, 1);
10444
10445 BIND(L_byteByByte);
10446 cmpl(tmp2, in2);
10447 jccb(Assembler::greater, L_exit);
10448 crc32(in_out, Address(in1, 0), 1);
10449 incq(in1);
10450 incl(tmp2);
10451 jmp(L_byteByByte);
10452
10453 BIND(L_exit);
10454 }
10455 #else
10456 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
10457 Register tmp1, Register tmp2, Register tmp3,
10458 Register tmp4, Register tmp5, Register tmp6,
10459 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10460 bool is_pclmulqdq_supported) {
10461 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
10462 Label L_wordByWord;
10463 Label L_byteByByteProlog;
10464 Label L_byteByByte;
10465 Label L_exit;
10466
10467 if (is_pclmulqdq_supported) {
10468 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
10469 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1);
10470
10471 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
10472 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
10473
10474 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
10475 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
10476 } else {
10477 const_or_pre_comp_const_index[0] = 1;
10478 const_or_pre_comp_const_index[1] = 0;
10479
10480 const_or_pre_comp_const_index[2] = 3;
10481 const_or_pre_comp_const_index[3] = 2;
10482
10483 const_or_pre_comp_const_index[4] = 5;
10484 const_or_pre_comp_const_index[5] = 4;
10485 }
10486 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
10487 in2, in1, in_out,
10488 tmp1, tmp2, tmp3,
10489 w_xtmp1, w_xtmp2, w_xtmp3,
10490 tmp4, tmp5,
10491 tmp6);
10492 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
10493 in2, in1, in_out,
10494 tmp1, tmp2, tmp3,
10495 w_xtmp1, w_xtmp2, w_xtmp3,
10496 tmp4, tmp5,
10497 tmp6);
10498 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
10499 in2, in1, in_out,
10500 tmp1, tmp2, tmp3,
10501 w_xtmp1, w_xtmp2, w_xtmp3,
10502 tmp4, tmp5,
10503 tmp6);
10504 movl(tmp1, in2);
10505 andl(tmp1, 0x00000007);
10506 negl(tmp1);
10507 addl(tmp1, in2);
10508 addl(tmp1, in1);
10509
10510 BIND(L_wordByWord);
10511 cmpl(in1, tmp1);
10512 jcc(Assembler::greaterEqual, L_byteByByteProlog);
10513 crc32(in_out, Address(in1,0), 4);
10514 addl(in1, 4);
10515 jmp(L_wordByWord);
10516
10517 BIND(L_byteByByteProlog);
10518 andl(in2, 0x00000007);
10519 movl(tmp2, 1);
10520
10521 BIND(L_byteByByte);
10522 cmpl(tmp2, in2);
10523 jccb(Assembler::greater, L_exit);
10524 movb(tmp1, Address(in1, 0));
10525 crc32(in_out, tmp1, 1);
10526 incl(in1);
10527 incl(tmp2);
10528 jmp(L_byteByByte);
10529
10530 BIND(L_exit);
10531 }
10532 #endif // LP64
10533 #undef BIND
10534 #undef BLOCK_COMMENT
10535
10536 // Compress char[] array to byte[].
10537 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java
10538 // @HotSpotIntrinsicCandidate
10539 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) {
10540 // for (int i = 0; i < len; i++) {
10541 // int c = src[srcOff++];
10542 // if (c >>> 8 != 0) {
10543 // return 0;
10544 // }
10545 // dst[dstOff++] = (byte)c;
10546 // }
10547 // return len;
10548 // }
10549 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
10550 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
10551 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
10552 Register tmp5, Register result) {
10553 Label copy_chars_loop, return_length, return_zero, done, below_threshold;
10554
10555 // rsi: src
10556 // rdi: dst
10557 // rdx: len
10558 // rcx: tmp5
10559 // rax: result
10560
10561 // rsi holds start addr of source char[] to be compressed
10562 // rdi holds start addr of destination byte[]
10563 // rdx holds length
10564
10565 assert(len != result, "");
10566
10567 // save length for return
10568 push(len);
10569
10570 if ((UseAVX > 2) && // AVX512
10571 VM_Version::supports_avx512vlbw() &&
10572 VM_Version::supports_bmi2()) {
10573
10574 set_vector_masking(); // opening of the stub context for programming mask registers
10575
10576 Label copy_32_loop, copy_loop_tail, restore_k1_return_zero;
10577
10578 // alignement
10579 Label post_alignement;
10580
10581 // if length of the string is less than 16, handle it in an old fashioned
10582 // way
10583 testl(len, -32);
10584 jcc(Assembler::zero, below_threshold);
10585
10586 // First check whether a character is compressable ( <= 0xFF).
10587 // Create mask to test for Unicode chars inside zmm vector
10588 movl(result, 0x00FF);
10589 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit);
10590
10591 // Save k1
10592 kmovql(k3, k1);
10593
10594 testl(len, -64);
10595 jcc(Assembler::zero, post_alignement);
10596
10597 movl(tmp5, dst);
10598 andl(tmp5, (32 - 1));
10599 negl(tmp5);
10600 andl(tmp5, (32 - 1));
10601
10602 // bail out when there is nothing to be done
10603 testl(tmp5, 0xFFFFFFFF);
10604 jcc(Assembler::zero, post_alignement);
10605
10606 // ~(~0 << len), where len is the # of remaining elements to process
10607 movl(result, 0xFFFFFFFF);
10608 shlxl(result, result, tmp5);
10609 notl(result);
10610 kmovdl(k1, result);
10611
10612 evmovdquw(tmp1Reg, k1, Address(src, 0), Assembler::AVX_512bit);
10613 evpcmpuw(k2, k1, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
10614 ktestd(k2, k1);
10615 jcc(Assembler::carryClear, restore_k1_return_zero);
10616
10617 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit);
10618
10619 addptr(src, tmp5);
10620 addptr(src, tmp5);
10621 addptr(dst, tmp5);
10622 subl(len, tmp5);
10623
10624 bind(post_alignement);
10625 // end of alignement
10626
10627 movl(tmp5, len);
10628 andl(tmp5, (32 - 1)); // tail count (in chars)
10629 andl(len, ~(32 - 1)); // vector count (in chars)
10630 jcc(Assembler::zero, copy_loop_tail);
10631
10632 lea(src, Address(src, len, Address::times_2));
10633 lea(dst, Address(dst, len, Address::times_1));
10634 negptr(len);
10635
10636 bind(copy_32_loop);
10637 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit);
10638 evpcmpuw(k2, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
10639 kortestdl(k2, k2);
10640 jcc(Assembler::carryClear, restore_k1_return_zero);
10641
10642 // All elements in current processed chunk are valid candidates for
10643 // compression. Write a truncated byte elements to the memory.
10644 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit);
10645 addptr(len, 32);
10646 jcc(Assembler::notZero, copy_32_loop);
10647
10648 bind(copy_loop_tail);
10649 // bail out when there is nothing to be done
10650 testl(tmp5, 0xFFFFFFFF);
10651 // Restore k1
10652 kmovql(k1, k3);
10653 jcc(Assembler::zero, return_length);
10654
10655 movl(len, tmp5);
10656
10657 // ~(~0 << len), where len is the # of remaining elements to process
10658 movl(result, 0xFFFFFFFF);
10659 shlxl(result, result, len);
10660 notl(result);
10661
10662 kmovdl(k1, result);
10663
10664 evmovdquw(tmp1Reg, k1, Address(src, 0), Assembler::AVX_512bit);
10665 evpcmpuw(k2, k1, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
10666 ktestd(k2, k1);
10667 jcc(Assembler::carryClear, restore_k1_return_zero);
10668
10669 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit);
10670 // Restore k1
10671 kmovql(k1, k3);
10672 jmp(return_length);
10673
10674 bind(restore_k1_return_zero);
10675 // Restore k1
10676 kmovql(k1, k3);
10677 jmp(return_zero);
10678
10679 clear_vector_masking(); // closing of the stub context for programming mask registers
10680 }
10681 if (UseSSE42Intrinsics) {
10682 Label copy_32_loop, copy_16, copy_tail;
10683
10684 bind(below_threshold);
10685
10686 movl(result, len);
10687
10688 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors
10689
10690 // vectored compression
10691 andl(len, 0xfffffff0); // vector count (in chars)
10692 andl(result, 0x0000000f); // tail count (in chars)
10693 testl(len, len);
10694 jccb(Assembler::zero, copy_16);
10695
10696 // compress 16 chars per iter
10697 movdl(tmp1Reg, tmp5);
10698 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
10699 pxor(tmp4Reg, tmp4Reg);
10700
10701 lea(src, Address(src, len, Address::times_2));
10702 lea(dst, Address(dst, len, Address::times_1));
10703 negptr(len);
10704
10705 bind(copy_32_loop);
10706 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters
10707 por(tmp4Reg, tmp2Reg);
10708 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters
10709 por(tmp4Reg, tmp3Reg);
10710 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector
10711 jcc(Assembler::notZero, return_zero);
10712 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte
10713 movdqu(Address(dst, len, Address::times_1), tmp2Reg);
10714 addptr(len, 16);
10715 jcc(Assembler::notZero, copy_32_loop);
10716
10717 // compress next vector of 8 chars (if any)
10718 bind(copy_16);
10719 movl(len, result);
10720 andl(len, 0xfffffff8); // vector count (in chars)
10721 andl(result, 0x00000007); // tail count (in chars)
10722 testl(len, len);
10723 jccb(Assembler::zero, copy_tail);
10724
10725 movdl(tmp1Reg, tmp5);
10726 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
10727 pxor(tmp3Reg, tmp3Reg);
10728
10729 movdqu(tmp2Reg, Address(src, 0));
10730 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
10731 jccb(Assembler::notZero, return_zero);
10732 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte
10733 movq(Address(dst, 0), tmp2Reg);
10734 addptr(src, 16);
10735 addptr(dst, 8);
10736
10737 bind(copy_tail);
10738 movl(len, result);
10739 }
10740 // compress 1 char per iter
10741 testl(len, len);
10742 jccb(Assembler::zero, return_length);
10743 lea(src, Address(src, len, Address::times_2));
10744 lea(dst, Address(dst, len, Address::times_1));
10745 negptr(len);
10746
10747 bind(copy_chars_loop);
10748 load_unsigned_short(result, Address(src, len, Address::times_2));
10749 testl(result, 0xff00); // check if Unicode char
10750 jccb(Assembler::notZero, return_zero);
10751 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte
10752 increment(len);
10753 jcc(Assembler::notZero, copy_chars_loop);
10754
10755 // if compression succeeded, return length
10756 bind(return_length);
10757 pop(result);
10758 jmpb(done);
10759
10760 // if compression failed, return 0
10761 bind(return_zero);
10762 xorl(result, result);
10763 addptr(rsp, wordSize);
10764
10765 bind(done);
10766 }
10767
10768 // Inflate byte[] array to char[].
10769 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java
10770 // @HotSpotIntrinsicCandidate
10771 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) {
10772 // for (int i = 0; i < len; i++) {
10773 // dst[dstOff++] = (char)(src[srcOff++] & 0xff);
10774 // }
10775 // }
10776 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
10777 XMMRegister tmp1, Register tmp2) {
10778 Label copy_chars_loop, done, below_threshold;
10779 // rsi: src
10780 // rdi: dst
10781 // rdx: len
10782 // rcx: tmp2
10783
10784 // rsi holds start addr of source byte[] to be inflated
10785 // rdi holds start addr of destination char[]
10786 // rdx holds length
10787 assert_different_registers(src, dst, len, tmp2);
10788
10789 if ((UseAVX > 2) && // AVX512
10790 VM_Version::supports_avx512vlbw() &&
10791 VM_Version::supports_bmi2()) {
10792
10793 set_vector_masking(); // opening of the stub context for programming mask registers
10794
10795 Label copy_32_loop, copy_tail;
10796 Register tmp3_aliased = len;
10797
10798 // if length of the string is less than 16, handle it in an old fashioned
10799 // way
10800 testl(len, -16);
10801 jcc(Assembler::zero, below_threshold);
10802
10803 // In order to use only one arithmetic operation for the main loop we use
10804 // this pre-calculation
10805 movl(tmp2, len);
10806 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop
10807 andl(len, -32); // vector count
10808 jccb(Assembler::zero, copy_tail);
10809
10810 lea(src, Address(src, len, Address::times_1));
10811 lea(dst, Address(dst, len, Address::times_2));
10812 negptr(len);
10813
10814
10815 // inflate 32 chars per iter
10816 bind(copy_32_loop);
10817 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit);
10818 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit);
10819 addptr(len, 32);
10820 jcc(Assembler::notZero, copy_32_loop);
10821
10822 bind(copy_tail);
10823 // bail out when there is nothing to be done
10824 testl(tmp2, -1); // we don't destroy the contents of tmp2 here
10825 jcc(Assembler::zero, done);
10826
10827 // Save k1
10828 kmovql(k2, k1);
10829
10830 // ~(~0 << length), where length is the # of remaining elements to process
10831 movl(tmp3_aliased, -1);
10832 shlxl(tmp3_aliased, tmp3_aliased, tmp2);
10833 notl(tmp3_aliased);
10834 kmovdl(k1, tmp3_aliased);
10835 evpmovzxbw(tmp1, k1, Address(src, 0), Assembler::AVX_512bit);
10836 evmovdquw(Address(dst, 0), k1, tmp1, Assembler::AVX_512bit);
10837
10838 // Restore k1
10839 kmovql(k1, k2);
10840 jmp(done);
10841
10842 clear_vector_masking(); // closing of the stub context for programming mask registers
10843 }
10844 if (UseSSE42Intrinsics) {
10845 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail;
10846
10847 movl(tmp2, len);
10848
10849 if (UseAVX > 1) {
10850 andl(tmp2, (16 - 1));
10851 andl(len, -16);
10852 jccb(Assembler::zero, copy_new_tail);
10853 } else {
10854 andl(tmp2, 0x00000007); // tail count (in chars)
10855 andl(len, 0xfffffff8); // vector count (in chars)
10856 jccb(Assembler::zero, copy_tail);
10857 }
10858
10859 // vectored inflation
10860 lea(src, Address(src, len, Address::times_1));
10861 lea(dst, Address(dst, len, Address::times_2));
10862 negptr(len);
10863
10864 if (UseAVX > 1) {
10865 bind(copy_16_loop);
10866 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit);
10867 vmovdqu(Address(dst, len, Address::times_2), tmp1);
10868 addptr(len, 16);
10869 jcc(Assembler::notZero, copy_16_loop);
10870
10871 bind(below_threshold);
10872 bind(copy_new_tail);
10873 if ((UseAVX > 2) &&
10874 VM_Version::supports_avx512vlbw() &&
10875 VM_Version::supports_bmi2()) {
10876 movl(tmp2, len);
10877 } else {
10878 movl(len, tmp2);
10879 }
10880 andl(tmp2, 0x00000007);
10881 andl(len, 0xFFFFFFF8);
10882 jccb(Assembler::zero, copy_tail);
10883
10884 pmovzxbw(tmp1, Address(src, 0));
10885 movdqu(Address(dst, 0), tmp1);
10886 addptr(src, 8);
10887 addptr(dst, 2 * 8);
10888
10889 jmp(copy_tail, true);
10890 }
10891
10892 // inflate 8 chars per iter
10893 bind(copy_8_loop);
10894 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words
10895 movdqu(Address(dst, len, Address::times_2), tmp1);
10896 addptr(len, 8);
10897 jcc(Assembler::notZero, copy_8_loop);
10898
10899 bind(copy_tail);
10900 movl(len, tmp2);
10901
10902 cmpl(len, 4);
10903 jccb(Assembler::less, copy_bytes);
10904
10905 movdl(tmp1, Address(src, 0)); // load 4 byte chars
10906 pmovzxbw(tmp1, tmp1);
10907 movq(Address(dst, 0), tmp1);
10908 subptr(len, 4);
10909 addptr(src, 4);
10910 addptr(dst, 8);
10911
10912 bind(copy_bytes);
10913 }
10914 testl(len, len);
10915 jccb(Assembler::zero, done);
10916 lea(src, Address(src, len, Address::times_1));
10917 lea(dst, Address(dst, len, Address::times_2));
10918 negptr(len);
10919
10920 // inflate 1 char per iter
10921 bind(copy_chars_loop);
10922 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char
10923 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word
10924 increment(len);
10925 jcc(Assembler::notZero, copy_chars_loop);
10926
10927 bind(done);
10928 }
10929
10930 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
10931 switch (cond) {
10932 // Note some conditions are synonyms for others
10933 case Assembler::zero: return Assembler::notZero;
10934 case Assembler::notZero: return Assembler::zero;
10935 case Assembler::less: return Assembler::greaterEqual;
10936 case Assembler::lessEqual: return Assembler::greater;
10937 case Assembler::greater: return Assembler::lessEqual;
10938 case Assembler::greaterEqual: return Assembler::less;
10939 case Assembler::below: return Assembler::aboveEqual;
10940 case Assembler::belowEqual: return Assembler::above;
10941 case Assembler::above: return Assembler::belowEqual;
10942 case Assembler::aboveEqual: return Assembler::below;
10943 case Assembler::overflow: return Assembler::noOverflow;
10944 case Assembler::noOverflow: return Assembler::overflow;
10945 case Assembler::negative: return Assembler::positive;
10946 case Assembler::positive: return Assembler::negative;
10947 case Assembler::parity: return Assembler::noParity;
10948 case Assembler::noParity: return Assembler::parity;
10949 }
10950 ShouldNotReachHere(); return Assembler::overflow;
10951 }
10952
10953 SkipIfEqual::SkipIfEqual(
10954 MacroAssembler* masm, const bool* flag_addr, bool value) {
10955 _masm = masm;
10956 _masm->cmp8(ExternalAddress((address)flag_addr), value);
10957 _masm->jcc(Assembler::equal, _label);
10958 }
10959
10960 SkipIfEqual::~SkipIfEqual() {
10961 _masm->bind(_label);
10962 }
10963
10964 // 32-bit Windows has its own fast-path implementation
10965 // of get_thread
10966 #if !defined(WIN32) || defined(_LP64)
10967
10968 // This is simply a call to Thread::current()
10969 void MacroAssembler::get_thread(Register thread) {
10970 if (thread != rax) {
10971 push(rax);
10972 }
10973 LP64_ONLY(push(rdi);)
10974 LP64_ONLY(push(rsi);)
10975 push(rdx);
10976 push(rcx);
10977 #ifdef _LP64
10978 push(r8);
10979 push(r9);
10980 push(r10);
10981 push(r11);
10982 #endif
10983
10984 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0);
10985
10986 #ifdef _LP64
10987 pop(r11);
10988 pop(r10);
10989 pop(r9);
10990 pop(r8);
10991 #endif
10992 pop(rcx);
10993 pop(rdx);
10994 LP64_ONLY(pop(rsi);)
10995 LP64_ONLY(pop(rdi);)
10996 if (thread != rax) {
10997 mov(thread, rax);
10998 pop(rax);
10999 }
11000 }
11001
11002 #endif