1 /*
2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "asm/assembler.inline.hpp"
28 #include "compiler/disassembler.hpp"
29 #include "gc/shared/cardTableModRefBS.hpp"
30 #include "gc/shared/collectedHeap.inline.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "memory/universe.hpp"
34 #include "oops/klass.inline.hpp"
35 #include "prims/methodHandles.hpp"
36 #include "runtime/biasedLocking.hpp"
37 #include "runtime/interfaceSupport.hpp"
38 #include "runtime/objectMonitor.hpp"
39 #include "runtime/os.hpp"
40 #include "runtime/sharedRuntime.hpp"
41 #include "runtime/stubRoutines.hpp"
42 #include "runtime/thread.hpp"
43 #include "utilities/macros.hpp"
44 #if INCLUDE_ALL_GCS
45 #include "gc/g1/g1CollectedHeap.inline.hpp"
46 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
47 #include "gc/g1/heapRegion.hpp"
48 #endif // INCLUDE_ALL_GCS
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 bool clear_pc) {
757 // we must set sp to zero to clear frame
758 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
759 // must clear fp, so that compiled frames are not confused; it is
760 // possible that we need it only for debugging
761 if (clear_fp) {
762 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
763 }
764
765 if (clear_pc) {
766 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
767 }
768 }
769
770 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
771 Register last_java_fp,
772 address last_java_pc) {
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 call_VM_leaf(CAST_FROM_FN_PTR(address, warning), c_rarg0);
837 pop_CPU_state();
838 mov(rsp, rbp);
839 pop(rbp);
840 }
841
842 void MacroAssembler::print_state() {
843 address rip = pc();
844 pusha(); // get regs on stack
845 push(rbp);
846 movq(rbp, rsp);
847 andq(rsp, -16); // align stack as required by push_CPU_state and call
848 push_CPU_state(); // keeps alignment at 16 bytes
849
850 lea(c_rarg0, InternalAddress(rip));
851 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array
852 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1);
853
854 pop_CPU_state();
855 mov(rsp, rbp);
856 pop(rbp);
857 popa();
858 }
859
860 #ifndef PRODUCT
861 extern "C" void findpc(intptr_t x);
862 #endif
863
864 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) {
865 // In order to get locks to work, we need to fake a in_VM state
866 if (ShowMessageBoxOnError) {
867 JavaThread* thread = JavaThread::current();
868 JavaThreadState saved_state = thread->thread_state();
869 thread->set_thread_state(_thread_in_vm);
870 #ifndef PRODUCT
871 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
872 ttyLocker ttyl;
873 BytecodeCounter::print();
874 }
875 #endif
876 // To see where a verify_oop failed, get $ebx+40/X for this frame.
877 // XXX correct this offset for amd64
878 // This is the value of eip which points to where verify_oop will return.
879 if (os::message_box(msg, "Execution stopped, print registers?")) {
880 print_state64(pc, regs);
881 BREAKPOINT;
882 assert(false, "start up GDB");
883 }
884 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
885 } else {
886 ttyLocker ttyl;
887 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n",
888 msg);
889 assert(false, "DEBUG MESSAGE: %s", msg);
890 }
891 }
892
893 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) {
894 ttyLocker ttyl;
895 FlagSetting fs(Debugging, true);
896 tty->print_cr("rip = 0x%016lx", pc);
897 #ifndef PRODUCT
898 tty->cr();
899 findpc(pc);
900 tty->cr();
901 #endif
902 #define PRINT_REG(rax, value) \
903 { tty->print("%s = ", #rax); os::print_location(tty, value); }
904 PRINT_REG(rax, regs[15]);
905 PRINT_REG(rbx, regs[12]);
906 PRINT_REG(rcx, regs[14]);
907 PRINT_REG(rdx, regs[13]);
908 PRINT_REG(rdi, regs[8]);
909 PRINT_REG(rsi, regs[9]);
910 PRINT_REG(rbp, regs[10]);
911 PRINT_REG(rsp, regs[11]);
912 PRINT_REG(r8 , regs[7]);
913 PRINT_REG(r9 , regs[6]);
914 PRINT_REG(r10, regs[5]);
915 PRINT_REG(r11, regs[4]);
916 PRINT_REG(r12, regs[3]);
917 PRINT_REG(r13, regs[2]);
918 PRINT_REG(r14, regs[1]);
919 PRINT_REG(r15, regs[0]);
920 #undef PRINT_REG
921 // Print some words near top of staack.
922 int64_t* rsp = (int64_t*) regs[11];
923 int64_t* dump_sp = rsp;
924 for (int col1 = 0; col1 < 8; col1++) {
925 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp);
926 os::print_location(tty, *dump_sp++);
927 }
928 for (int row = 0; row < 25; row++) {
929 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp);
930 for (int col = 0; col < 4; col++) {
931 tty->print(" 0x%016lx", *dump_sp++);
932 }
933 tty->cr();
934 }
935 // Print some instructions around pc:
936 Disassembler::decode((address)pc-64, (address)pc);
937 tty->print_cr("--------");
938 Disassembler::decode((address)pc, (address)pc+32);
939 }
940
941 #endif // _LP64
942
943 // Now versions that are common to 32/64 bit
944
945 void MacroAssembler::addptr(Register dst, int32_t imm32) {
946 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32));
947 }
948
949 void MacroAssembler::addptr(Register dst, Register src) {
950 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
951 }
952
953 void MacroAssembler::addptr(Address dst, Register src) {
954 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
955 }
956
957 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) {
958 if (reachable(src)) {
959 Assembler::addsd(dst, as_Address(src));
960 } else {
961 lea(rscratch1, src);
962 Assembler::addsd(dst, Address(rscratch1, 0));
963 }
964 }
965
966 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) {
967 if (reachable(src)) {
968 addss(dst, as_Address(src));
969 } else {
970 lea(rscratch1, src);
971 addss(dst, Address(rscratch1, 0));
972 }
973 }
974
975 void MacroAssembler::addpd(XMMRegister dst, AddressLiteral src) {
976 if (reachable(src)) {
977 Assembler::addpd(dst, as_Address(src));
978 } else {
979 lea(rscratch1, src);
980 Assembler::addpd(dst, Address(rscratch1, 0));
981 }
982 }
983
984 void MacroAssembler::align(int modulus) {
985 align(modulus, offset());
986 }
987
988 void MacroAssembler::align(int modulus, int target) {
989 if (target % modulus != 0) {
990 nop(modulus - (target % modulus));
991 }
992 }
993
994 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) {
995 // Used in sign-masking with aligned address.
996 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
997 if (reachable(src)) {
998 Assembler::andpd(dst, as_Address(src));
999 } else {
1000 lea(rscratch1, src);
1001 Assembler::andpd(dst, Address(rscratch1, 0));
1002 }
1003 }
1004
1005 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) {
1006 // Used in sign-masking with aligned address.
1007 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
1008 if (reachable(src)) {
1009 Assembler::andps(dst, as_Address(src));
1010 } else {
1011 lea(rscratch1, src);
1012 Assembler::andps(dst, Address(rscratch1, 0));
1013 }
1014 }
1015
1016 void MacroAssembler::andptr(Register dst, int32_t imm32) {
1017 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32));
1018 }
1019
1020 void MacroAssembler::atomic_incl(Address counter_addr) {
1021 if (os::is_MP())
1022 lock();
1023 incrementl(counter_addr);
1024 }
1025
1026 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) {
1027 if (reachable(counter_addr)) {
1028 atomic_incl(as_Address(counter_addr));
1029 } else {
1030 lea(scr, counter_addr);
1031 atomic_incl(Address(scr, 0));
1032 }
1033 }
1034
1035 #ifdef _LP64
1036 void MacroAssembler::atomic_incq(Address counter_addr) {
1037 if (os::is_MP())
1038 lock();
1039 incrementq(counter_addr);
1040 }
1041
1042 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) {
1043 if (reachable(counter_addr)) {
1044 atomic_incq(as_Address(counter_addr));
1045 } else {
1046 lea(scr, counter_addr);
1047 atomic_incq(Address(scr, 0));
1048 }
1049 }
1050 #endif
1051
1052 // Writes to stack successive pages until offset reached to check for
1053 // stack overflow + shadow pages. This clobbers tmp.
1054 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
1055 movptr(tmp, rsp);
1056 // Bang stack for total size given plus shadow page size.
1057 // Bang one page at a time because large size can bang beyond yellow and
1058 // red zones.
1059 Label loop;
1060 bind(loop);
1061 movl(Address(tmp, (-os::vm_page_size())), size );
1062 subptr(tmp, os::vm_page_size());
1063 subl(size, os::vm_page_size());
1064 jcc(Assembler::greater, loop);
1065
1066 // Bang down shadow pages too.
1067 // At this point, (tmp-0) is the last address touched, so don't
1068 // touch it again. (It was touched as (tmp-pagesize) but then tmp
1069 // was post-decremented.) Skip this address by starting at i=1, and
1070 // touch a few more pages below. N.B. It is important to touch all
1071 // the way down including all pages in the shadow zone.
1072 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) {
1073 // this could be any sized move but this is can be a debugging crumb
1074 // so the bigger the better.
1075 movptr(Address(tmp, (-i*os::vm_page_size())), size );
1076 }
1077 }
1078
1079 void MacroAssembler::reserved_stack_check() {
1080 // testing if reserved zone needs to be enabled
1081 Label no_reserved_zone_enabling;
1082 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread);
1083 NOT_LP64(get_thread(rsi);)
1084
1085 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset()));
1086 jcc(Assembler::below, no_reserved_zone_enabling);
1087
1088 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread);
1089 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry()));
1090 should_not_reach_here();
1091
1092 bind(no_reserved_zone_enabling);
1093 }
1094
1095 int MacroAssembler::biased_locking_enter(Register lock_reg,
1096 Register obj_reg,
1097 Register swap_reg,
1098 Register tmp_reg,
1099 bool swap_reg_contains_mark,
1100 Label& done,
1101 Label* slow_case,
1102 BiasedLockingCounters* counters) {
1103 assert(UseBiasedLocking, "why call this otherwise?");
1104 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq");
1105 assert(tmp_reg != noreg, "tmp_reg must be supplied");
1106 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg);
1107 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout");
1108 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes());
1109 NOT_LP64( Address saved_mark_addr(lock_reg, 0); )
1110
1111 if (PrintBiasedLockingStatistics && counters == NULL) {
1112 counters = BiasedLocking::counters();
1113 }
1114 // Biased locking
1115 // See whether the lock is currently biased toward our thread and
1116 // whether the epoch is still valid
1117 // Note that the runtime guarantees sufficient alignment of JavaThread
1118 // pointers to allow age to be placed into low bits
1119 // First check to see whether biasing is even enabled for this object
1120 Label cas_label;
1121 int null_check_offset = -1;
1122 if (!swap_reg_contains_mark) {
1123 null_check_offset = offset();
1124 movptr(swap_reg, mark_addr);
1125 }
1126 movptr(tmp_reg, swap_reg);
1127 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place);
1128 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern);
1129 jcc(Assembler::notEqual, cas_label);
1130 // The bias pattern is present in the object's header. Need to check
1131 // whether the bias owner and the epoch are both still current.
1132 #ifndef _LP64
1133 // Note that because there is no current thread register on x86_32 we
1134 // need to store off the mark word we read out of the object to
1135 // avoid reloading it and needing to recheck invariants below. This
1136 // store is unfortunate but it makes the overall code shorter and
1137 // simpler.
1138 movptr(saved_mark_addr, swap_reg);
1139 #endif
1140 if (swap_reg_contains_mark) {
1141 null_check_offset = offset();
1142 }
1143 load_prototype_header(tmp_reg, obj_reg);
1144 #ifdef _LP64
1145 orptr(tmp_reg, r15_thread);
1146 xorptr(tmp_reg, swap_reg);
1147 Register header_reg = tmp_reg;
1148 #else
1149 xorptr(tmp_reg, swap_reg);
1150 get_thread(swap_reg);
1151 xorptr(swap_reg, tmp_reg);
1152 Register header_reg = swap_reg;
1153 #endif
1154 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place));
1155 if (counters != NULL) {
1156 cond_inc32(Assembler::zero,
1157 ExternalAddress((address) counters->biased_lock_entry_count_addr()));
1158 }
1159 jcc(Assembler::equal, done);
1160
1161 Label try_revoke_bias;
1162 Label try_rebias;
1163
1164 // At this point we know that the header has the bias pattern and
1165 // that we are not the bias owner in the current epoch. We need to
1166 // figure out more details about the state of the header in order to
1167 // know what operations can be legally performed on the object's
1168 // header.
1169
1170 // If the low three bits in the xor result aren't clear, that means
1171 // the prototype header is no longer biased and we have to revoke
1172 // the bias on this object.
1173 testptr(header_reg, markOopDesc::biased_lock_mask_in_place);
1174 jccb(Assembler::notZero, try_revoke_bias);
1175
1176 // Biasing is still enabled for this data type. See whether the
1177 // epoch of the current bias is still valid, meaning that the epoch
1178 // bits of the mark word are equal to the epoch bits of the
1179 // prototype header. (Note that the prototype header's epoch bits
1180 // only change at a safepoint.) If not, attempt to rebias the object
1181 // toward the current thread. Note that we must be absolutely sure
1182 // that the current epoch is invalid in order to do this because
1183 // otherwise the manipulations it performs on the mark word are
1184 // illegal.
1185 testptr(header_reg, markOopDesc::epoch_mask_in_place);
1186 jccb(Assembler::notZero, try_rebias);
1187
1188 // The epoch of the current bias is still valid but we know nothing
1189 // about the owner; it might be set or it might be clear. Try to
1190 // acquire the bias of the object using an atomic operation. If this
1191 // fails we will go in to the runtime to revoke the object's bias.
1192 // Note that we first construct the presumed unbiased header so we
1193 // don't accidentally blow away another thread's valid bias.
1194 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1195 andptr(swap_reg,
1196 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place);
1197 #ifdef _LP64
1198 movptr(tmp_reg, swap_reg);
1199 orptr(tmp_reg, r15_thread);
1200 #else
1201 get_thread(tmp_reg);
1202 orptr(tmp_reg, swap_reg);
1203 #endif
1204 if (os::is_MP()) {
1205 lock();
1206 }
1207 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1208 // If the biasing toward our thread failed, this means that
1209 // another thread succeeded in biasing it toward itself and we
1210 // need to revoke that bias. The revocation will occur in the
1211 // interpreter runtime in the slow case.
1212 if (counters != NULL) {
1213 cond_inc32(Assembler::zero,
1214 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr()));
1215 }
1216 if (slow_case != NULL) {
1217 jcc(Assembler::notZero, *slow_case);
1218 }
1219 jmp(done);
1220
1221 bind(try_rebias);
1222 // At this point we know the epoch has expired, meaning that the
1223 // current "bias owner", if any, is actually invalid. Under these
1224 // circumstances _only_, we are allowed to use the current header's
1225 // value as the comparison value when doing the cas to acquire the
1226 // bias in the current epoch. In other words, we allow transfer of
1227 // the bias from one thread to another directly in this situation.
1228 //
1229 // FIXME: due to a lack of registers we currently blow away the age
1230 // bits in this situation. Should attempt to preserve them.
1231 load_prototype_header(tmp_reg, obj_reg);
1232 #ifdef _LP64
1233 orptr(tmp_reg, r15_thread);
1234 #else
1235 get_thread(swap_reg);
1236 orptr(tmp_reg, swap_reg);
1237 movptr(swap_reg, saved_mark_addr);
1238 #endif
1239 if (os::is_MP()) {
1240 lock();
1241 }
1242 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1243 // If the biasing toward our thread failed, then another thread
1244 // succeeded in biasing it toward itself and we need to revoke that
1245 // bias. The revocation will occur in the runtime in the slow case.
1246 if (counters != NULL) {
1247 cond_inc32(Assembler::zero,
1248 ExternalAddress((address) counters->rebiased_lock_entry_count_addr()));
1249 }
1250 if (slow_case != NULL) {
1251 jcc(Assembler::notZero, *slow_case);
1252 }
1253 jmp(done);
1254
1255 bind(try_revoke_bias);
1256 // The prototype mark in the klass doesn't have the bias bit set any
1257 // more, indicating that objects of this data type are not supposed
1258 // to be biased any more. We are going to try to reset the mark of
1259 // this object to the prototype value and fall through to the
1260 // CAS-based locking scheme. Note that if our CAS fails, it means
1261 // that another thread raced us for the privilege of revoking the
1262 // bias of this particular object, so it's okay to continue in the
1263 // normal locking code.
1264 //
1265 // FIXME: due to a lack of registers we currently blow away the age
1266 // bits in this situation. Should attempt to preserve them.
1267 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1268 load_prototype_header(tmp_reg, obj_reg);
1269 if (os::is_MP()) {
1270 lock();
1271 }
1272 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1273 // Fall through to the normal CAS-based lock, because no matter what
1274 // the result of the above CAS, some thread must have succeeded in
1275 // removing the bias bit from the object's header.
1276 if (counters != NULL) {
1277 cond_inc32(Assembler::zero,
1278 ExternalAddress((address) counters->revoked_lock_entry_count_addr()));
1279 }
1280
1281 bind(cas_label);
1282
1283 return null_check_offset;
1284 }
1285
1286 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) {
1287 assert(UseBiasedLocking, "why call this otherwise?");
1288
1289 // Check for biased locking unlock case, which is a no-op
1290 // Note: we do not have to check the thread ID for two reasons.
1291 // First, the interpreter checks for IllegalMonitorStateException at
1292 // a higher level. Second, if the bias was revoked while we held the
1293 // lock, the object could not be rebiased toward another thread, so
1294 // the bias bit would be clear.
1295 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1296 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place);
1297 cmpptr(temp_reg, markOopDesc::biased_lock_pattern);
1298 jcc(Assembler::equal, done);
1299 }
1300
1301 #ifdef COMPILER2
1302
1303 #if INCLUDE_RTM_OPT
1304
1305 // Update rtm_counters based on abort status
1306 // input: abort_status
1307 // rtm_counters (RTMLockingCounters*)
1308 // flags are killed
1309 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) {
1310
1311 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset()));
1312 if (PrintPreciseRTMLockingStatistics) {
1313 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) {
1314 Label check_abort;
1315 testl(abort_status, (1<<i));
1316 jccb(Assembler::equal, check_abort);
1317 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx))));
1318 bind(check_abort);
1319 }
1320 }
1321 }
1322
1323 // Branch if (random & (count-1) != 0), count is 2^n
1324 // tmp, scr and flags are killed
1325 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) {
1326 assert(tmp == rax, "");
1327 assert(scr == rdx, "");
1328 rdtsc(); // modifies EDX:EAX
1329 andptr(tmp, count-1);
1330 jccb(Assembler::notZero, brLabel);
1331 }
1332
1333 // Perform abort ratio calculation, set no_rtm bit if high ratio
1334 // input: rtm_counters_Reg (RTMLockingCounters* address)
1335 // tmpReg, rtm_counters_Reg and flags are killed
1336 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg,
1337 Register rtm_counters_Reg,
1338 RTMLockingCounters* rtm_counters,
1339 Metadata* method_data) {
1340 Label L_done, L_check_always_rtm1, L_check_always_rtm2;
1341
1342 if (RTMLockingCalculationDelay > 0) {
1343 // Delay calculation
1344 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg);
1345 testptr(tmpReg, tmpReg);
1346 jccb(Assembler::equal, L_done);
1347 }
1348 // Abort ratio calculation only if abort_count > RTMAbortThreshold
1349 // Aborted transactions = abort_count * 100
1350 // All transactions = total_count * RTMTotalCountIncrRate
1351 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio)
1352
1353 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset()));
1354 cmpptr(tmpReg, RTMAbortThreshold);
1355 jccb(Assembler::below, L_check_always_rtm2);
1356 imulptr(tmpReg, tmpReg, 100);
1357
1358 Register scrReg = rtm_counters_Reg;
1359 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1360 imulptr(scrReg, scrReg, RTMTotalCountIncrRate);
1361 imulptr(scrReg, scrReg, RTMAbortRatio);
1362 cmpptr(tmpReg, scrReg);
1363 jccb(Assembler::below, L_check_always_rtm1);
1364 if (method_data != NULL) {
1365 // set rtm_state to "no rtm" in MDO
1366 mov_metadata(tmpReg, method_data);
1367 if (os::is_MP()) {
1368 lock();
1369 }
1370 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM);
1371 }
1372 jmpb(L_done);
1373 bind(L_check_always_rtm1);
1374 // Reload RTMLockingCounters* address
1375 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1376 bind(L_check_always_rtm2);
1377 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1378 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate);
1379 jccb(Assembler::below, L_done);
1380 if (method_data != NULL) {
1381 // set rtm_state to "always rtm" in MDO
1382 mov_metadata(tmpReg, method_data);
1383 if (os::is_MP()) {
1384 lock();
1385 }
1386 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM);
1387 }
1388 bind(L_done);
1389 }
1390
1391 // Update counters and perform abort ratio calculation
1392 // input: abort_status_Reg
1393 // rtm_counters_Reg, flags are killed
1394 void MacroAssembler::rtm_profiling(Register abort_status_Reg,
1395 Register rtm_counters_Reg,
1396 RTMLockingCounters* rtm_counters,
1397 Metadata* method_data,
1398 bool profile_rtm) {
1399
1400 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1401 // update rtm counters based on rax value at abort
1402 // reads abort_status_Reg, updates flags
1403 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1404 rtm_counters_update(abort_status_Reg, rtm_counters_Reg);
1405 if (profile_rtm) {
1406 // Save abort status because abort_status_Reg is used by following code.
1407 if (RTMRetryCount > 0) {
1408 push(abort_status_Reg);
1409 }
1410 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1411 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data);
1412 // restore abort status
1413 if (RTMRetryCount > 0) {
1414 pop(abort_status_Reg);
1415 }
1416 }
1417 }
1418
1419 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4)
1420 // inputs: retry_count_Reg
1421 // : abort_status_Reg
1422 // output: retry_count_Reg decremented by 1
1423 // flags are killed
1424 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) {
1425 Label doneRetry;
1426 assert(abort_status_Reg == rax, "");
1427 // The abort reason bits are in eax (see all states in rtmLocking.hpp)
1428 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4)
1429 // if reason is in 0x6 and retry count != 0 then retry
1430 andptr(abort_status_Reg, 0x6);
1431 jccb(Assembler::zero, doneRetry);
1432 testl(retry_count_Reg, retry_count_Reg);
1433 jccb(Assembler::zero, doneRetry);
1434 pause();
1435 decrementl(retry_count_Reg);
1436 jmp(retryLabel);
1437 bind(doneRetry);
1438 }
1439
1440 // Spin and retry if lock is busy,
1441 // inputs: box_Reg (monitor address)
1442 // : retry_count_Reg
1443 // output: retry_count_Reg decremented by 1
1444 // : clear z flag if retry count exceeded
1445 // tmp_Reg, scr_Reg, flags are killed
1446 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg,
1447 Register tmp_Reg, Register scr_Reg, Label& retryLabel) {
1448 Label SpinLoop, SpinExit, doneRetry;
1449 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1450
1451 testl(retry_count_Reg, retry_count_Reg);
1452 jccb(Assembler::zero, doneRetry);
1453 decrementl(retry_count_Reg);
1454 movptr(scr_Reg, RTMSpinLoopCount);
1455
1456 bind(SpinLoop);
1457 pause();
1458 decrementl(scr_Reg);
1459 jccb(Assembler::lessEqual, SpinExit);
1460 movptr(tmp_Reg, Address(box_Reg, owner_offset));
1461 testptr(tmp_Reg, tmp_Reg);
1462 jccb(Assembler::notZero, SpinLoop);
1463
1464 bind(SpinExit);
1465 jmp(retryLabel);
1466 bind(doneRetry);
1467 incrementl(retry_count_Reg); // clear z flag
1468 }
1469
1470 // Use RTM for normal stack locks
1471 // Input: objReg (object to lock)
1472 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg,
1473 Register retry_on_abort_count_Reg,
1474 RTMLockingCounters* stack_rtm_counters,
1475 Metadata* method_data, bool profile_rtm,
1476 Label& DONE_LABEL, Label& IsInflated) {
1477 assert(UseRTMForStackLocks, "why call this otherwise?");
1478 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1479 assert(tmpReg == rax, "");
1480 assert(scrReg == rdx, "");
1481 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1482
1483 if (RTMRetryCount > 0) {
1484 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1485 bind(L_rtm_retry);
1486 }
1487 movptr(tmpReg, Address(objReg, 0));
1488 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1489 jcc(Assembler::notZero, IsInflated);
1490
1491 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1492 Label L_noincrement;
1493 if (RTMTotalCountIncrRate > 1) {
1494 // tmpReg, scrReg and flags are killed
1495 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement);
1496 }
1497 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM");
1498 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg);
1499 bind(L_noincrement);
1500 }
1501 xbegin(L_on_abort);
1502 movptr(tmpReg, Address(objReg, 0)); // fetch markword
1503 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1504 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1505 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked
1506
1507 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1508 if (UseRTMXendForLockBusy) {
1509 xend();
1510 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry)
1511 jmp(L_decrement_retry);
1512 }
1513 else {
1514 xabort(0);
1515 }
1516 bind(L_on_abort);
1517 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1518 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm);
1519 }
1520 bind(L_decrement_retry);
1521 if (RTMRetryCount > 0) {
1522 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1523 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1524 }
1525 }
1526
1527 // Use RTM for inflating locks
1528 // inputs: objReg (object to lock)
1529 // boxReg (on-stack box address (displaced header location) - KILLED)
1530 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value)
1531 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg,
1532 Register scrReg, Register retry_on_busy_count_Reg,
1533 Register retry_on_abort_count_Reg,
1534 RTMLockingCounters* rtm_counters,
1535 Metadata* method_data, bool profile_rtm,
1536 Label& DONE_LABEL) {
1537 assert(UseRTMLocking, "why call this otherwise?");
1538 assert(tmpReg == rax, "");
1539 assert(scrReg == rdx, "");
1540 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1541 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1542
1543 // Without cast to int32_t a movptr will destroy r10 which is typically obj
1544 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1545 movptr(boxReg, tmpReg); // Save ObjectMonitor address
1546
1547 if (RTMRetryCount > 0) {
1548 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy
1549 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1550 bind(L_rtm_retry);
1551 }
1552 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1553 Label L_noincrement;
1554 if (RTMTotalCountIncrRate > 1) {
1555 // tmpReg, scrReg and flags are killed
1556 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement);
1557 }
1558 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1559 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg);
1560 bind(L_noincrement);
1561 }
1562 xbegin(L_on_abort);
1563 movptr(tmpReg, Address(objReg, 0));
1564 movptr(tmpReg, Address(tmpReg, owner_offset));
1565 testptr(tmpReg, tmpReg);
1566 jcc(Assembler::zero, DONE_LABEL);
1567 if (UseRTMXendForLockBusy) {
1568 xend();
1569 jmp(L_decrement_retry);
1570 }
1571 else {
1572 xabort(0);
1573 }
1574 bind(L_on_abort);
1575 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1576 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1577 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm);
1578 }
1579 if (RTMRetryCount > 0) {
1580 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1581 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1582 }
1583
1584 movptr(tmpReg, Address(boxReg, owner_offset)) ;
1585 testptr(tmpReg, tmpReg) ;
1586 jccb(Assembler::notZero, L_decrement_retry) ;
1587
1588 // Appears unlocked - try to swing _owner from null to non-null.
1589 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1590 #ifdef _LP64
1591 Register threadReg = r15_thread;
1592 #else
1593 get_thread(scrReg);
1594 Register threadReg = scrReg;
1595 #endif
1596 if (os::is_MP()) {
1597 lock();
1598 }
1599 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg
1600
1601 if (RTMRetryCount > 0) {
1602 // success done else retry
1603 jccb(Assembler::equal, DONE_LABEL) ;
1604 bind(L_decrement_retry);
1605 // Spin and retry if lock is busy.
1606 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry);
1607 }
1608 else {
1609 bind(L_decrement_retry);
1610 }
1611 }
1612
1613 #endif // INCLUDE_RTM_OPT
1614
1615 // Fast_Lock and Fast_Unlock used by C2
1616
1617 // Because the transitions from emitted code to the runtime
1618 // monitorenter/exit helper stubs are so slow it's critical that
1619 // we inline both the stack-locking fast-path and the inflated fast path.
1620 //
1621 // See also: cmpFastLock and cmpFastUnlock.
1622 //
1623 // What follows is a specialized inline transliteration of the code
1624 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat
1625 // another option would be to emit TrySlowEnter and TrySlowExit methods
1626 // at startup-time. These methods would accept arguments as
1627 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
1628 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply
1629 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
1630 // In practice, however, the # of lock sites is bounded and is usually small.
1631 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
1632 // if the processor uses simple bimodal branch predictors keyed by EIP
1633 // Since the helper routines would be called from multiple synchronization
1634 // sites.
1635 //
1636 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
1637 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
1638 // to those specialized methods. That'd give us a mostly platform-independent
1639 // implementation that the JITs could optimize and inline at their pleasure.
1640 // Done correctly, the only time we'd need to cross to native could would be
1641 // to park() or unpark() threads. We'd also need a few more unsafe operators
1642 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
1643 // (b) explicit barriers or fence operations.
1644 //
1645 // TODO:
1646 //
1647 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
1648 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
1649 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
1650 // the lock operators would typically be faster than reifying Self.
1651 //
1652 // * Ideally I'd define the primitives as:
1653 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
1654 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
1655 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
1656 // Instead, we're stuck with a rather awkward and brittle register assignments below.
1657 // Furthermore the register assignments are overconstrained, possibly resulting in
1658 // sub-optimal code near the synchronization site.
1659 //
1660 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
1661 // Alternately, use a better sp-proximity test.
1662 //
1663 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
1664 // Either one is sufficient to uniquely identify a thread.
1665 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
1666 //
1667 // * Intrinsify notify() and notifyAll() for the common cases where the
1668 // object is locked by the calling thread but the waitlist is empty.
1669 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
1670 //
1671 // * use jccb and jmpb instead of jcc and jmp to improve code density.
1672 // But beware of excessive branch density on AMD Opterons.
1673 //
1674 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
1675 // or failure of the fast-path. If the fast-path fails then we pass
1676 // control to the slow-path, typically in C. In Fast_Lock and
1677 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2
1678 // will emit a conditional branch immediately after the node.
1679 // So we have branches to branches and lots of ICC.ZF games.
1680 // Instead, it might be better to have C2 pass a "FailureLabel"
1681 // into Fast_Lock and Fast_Unlock. In the case of success, control
1682 // will drop through the node. ICC.ZF is undefined at exit.
1683 // In the case of failure, the node will branch directly to the
1684 // FailureLabel
1685
1686
1687 // obj: object to lock
1688 // box: on-stack box address (displaced header location) - KILLED
1689 // rax,: tmp -- KILLED
1690 // scr: tmp -- KILLED
1691 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg,
1692 Register scrReg, Register cx1Reg, Register cx2Reg,
1693 BiasedLockingCounters* counters,
1694 RTMLockingCounters* rtm_counters,
1695 RTMLockingCounters* stack_rtm_counters,
1696 Metadata* method_data,
1697 bool use_rtm, bool profile_rtm) {
1698 // Ensure the register assignments are disjoint
1699 assert(tmpReg == rax, "");
1700
1701 if (use_rtm) {
1702 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg);
1703 } else {
1704 assert(cx1Reg == noreg, "");
1705 assert(cx2Reg == noreg, "");
1706 assert_different_registers(objReg, boxReg, tmpReg, scrReg);
1707 }
1708
1709 if (counters != NULL) {
1710 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg);
1711 }
1712 if (EmitSync & 1) {
1713 // set box->dhw = markOopDesc::unused_mark()
1714 // Force all sync thru slow-path: slow_enter() and slow_exit()
1715 movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1716 cmpptr (rsp, (int32_t)NULL_WORD);
1717 } else {
1718 // Possible cases that we'll encounter in fast_lock
1719 // ------------------------------------------------
1720 // * Inflated
1721 // -- unlocked
1722 // -- Locked
1723 // = by self
1724 // = by other
1725 // * biased
1726 // -- by Self
1727 // -- by other
1728 // * neutral
1729 // * stack-locked
1730 // -- by self
1731 // = sp-proximity test hits
1732 // = sp-proximity test generates false-negative
1733 // -- by other
1734 //
1735
1736 Label IsInflated, DONE_LABEL;
1737
1738 // it's stack-locked, biased or neutral
1739 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
1740 // order to reduce the number of conditional branches in the most common cases.
1741 // Beware -- there's a subtle invariant that fetch of the markword
1742 // at [FETCH], below, will never observe a biased encoding (*101b).
1743 // If this invariant is not held we risk exclusion (safety) failure.
1744 if (UseBiasedLocking && !UseOptoBiasInlining) {
1745 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters);
1746 }
1747
1748 #if INCLUDE_RTM_OPT
1749 if (UseRTMForStackLocks && use_rtm) {
1750 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg,
1751 stack_rtm_counters, method_data, profile_rtm,
1752 DONE_LABEL, IsInflated);
1753 }
1754 #endif // INCLUDE_RTM_OPT
1755
1756 movptr(tmpReg, Address(objReg, 0)); // [FETCH]
1757 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1758 jccb(Assembler::notZero, IsInflated);
1759
1760 // Attempt stack-locking ...
1761 orptr (tmpReg, markOopDesc::unlocked_value);
1762 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
1763 if (os::is_MP()) {
1764 lock();
1765 }
1766 cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg
1767 if (counters != NULL) {
1768 cond_inc32(Assembler::equal,
1769 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1770 }
1771 jcc(Assembler::equal, DONE_LABEL); // Success
1772
1773 // Recursive locking.
1774 // The object is stack-locked: markword contains stack pointer to BasicLock.
1775 // Locked by current thread if difference with current SP is less than one page.
1776 subptr(tmpReg, rsp);
1777 // Next instruction set ZFlag == 1 (Success) if difference is less then one page.
1778 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) );
1779 movptr(Address(boxReg, 0), tmpReg);
1780 if (counters != NULL) {
1781 cond_inc32(Assembler::equal,
1782 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1783 }
1784 jmp(DONE_LABEL);
1785
1786 bind(IsInflated);
1787 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value
1788
1789 #if INCLUDE_RTM_OPT
1790 // Use the same RTM locking code in 32- and 64-bit VM.
1791 if (use_rtm) {
1792 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg,
1793 rtm_counters, method_data, profile_rtm, DONE_LABEL);
1794 } else {
1795 #endif // INCLUDE_RTM_OPT
1796
1797 #ifndef _LP64
1798 // The object is inflated.
1799
1800 // boxReg refers to the on-stack BasicLock in the current frame.
1801 // We'd like to write:
1802 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices.
1803 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
1804 // additional latency as we have another ST in the store buffer that must drain.
1805
1806 if (EmitSync & 8192) {
1807 movptr(Address(boxReg, 0), 3); // results in ST-before-CAS penalty
1808 get_thread (scrReg);
1809 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1810 movptr(tmpReg, NULL_WORD); // consider: xor vs mov
1811 if (os::is_MP()) {
1812 lock();
1813 }
1814 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1815 } else
1816 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS
1817 // register juggle because we need tmpReg for cmpxchgptr below
1818 movptr(scrReg, boxReg);
1819 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1820
1821 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1822 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1823 // prefetchw [eax + Offset(_owner)-2]
1824 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1825 }
1826
1827 if ((EmitSync & 64) == 0) {
1828 // Optimistic form: consider XORL tmpReg,tmpReg
1829 movptr(tmpReg, NULL_WORD);
1830 } else {
1831 // Can suffer RTS->RTO upgrades on shared or cold $ lines
1832 // Test-And-CAS instead of CAS
1833 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner
1834 testptr(tmpReg, tmpReg); // Locked ?
1835 jccb (Assembler::notZero, DONE_LABEL);
1836 }
1837
1838 // Appears unlocked - try to swing _owner from null to non-null.
1839 // Ideally, I'd manifest "Self" with get_thread and then attempt
1840 // to CAS the register containing Self into m->Owner.
1841 // But we don't have enough registers, so instead we can either try to CAS
1842 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
1843 // we later store "Self" into m->Owner. Transiently storing a stack address
1844 // (rsp or the address of the box) into m->owner is harmless.
1845 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1846 if (os::is_MP()) {
1847 lock();
1848 }
1849 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1850 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3
1851 // If we weren't able to swing _owner from NULL to the BasicLock
1852 // then take the slow path.
1853 jccb (Assembler::notZero, DONE_LABEL);
1854 // update _owner from BasicLock to thread
1855 get_thread (scrReg); // beware: clobbers ICCs
1856 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg);
1857 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success
1858
1859 // If the CAS fails we can either retry or pass control to the slow-path.
1860 // We use the latter tactic.
1861 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1862 // If the CAS was successful ...
1863 // Self has acquired the lock
1864 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1865 // Intentional fall-through into DONE_LABEL ...
1866 } else {
1867 movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())); // results in ST-before-CAS penalty
1868 movptr(boxReg, tmpReg);
1869
1870 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1871 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1872 // prefetchw [eax + Offset(_owner)-2]
1873 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1874 }
1875
1876 if ((EmitSync & 64) == 0) {
1877 // Optimistic form
1878 xorptr (tmpReg, tmpReg);
1879 } else {
1880 // Can suffer RTS->RTO upgrades on shared or cold $ lines
1881 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner
1882 testptr(tmpReg, tmpReg); // Locked ?
1883 jccb (Assembler::notZero, DONE_LABEL);
1884 }
1885
1886 // Appears unlocked - try to swing _owner from null to non-null.
1887 // Use either "Self" (in scr) or rsp as thread identity in _owner.
1888 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1889 get_thread (scrReg);
1890 if (os::is_MP()) {
1891 lock();
1892 }
1893 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1894
1895 // If the CAS fails we can either retry or pass control to the slow-path.
1896 // We use the latter tactic.
1897 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1898 // If the CAS was successful ...
1899 // Self has acquired the lock
1900 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1901 // Intentional fall-through into DONE_LABEL ...
1902 }
1903 #else // _LP64
1904 // It's inflated
1905 movq(scrReg, tmpReg);
1906 xorq(tmpReg, tmpReg);
1907
1908 if (os::is_MP()) {
1909 lock();
1910 }
1911 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1912 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark().
1913 // Without cast to int32_t movptr will destroy r10 which is typically obj.
1914 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1915 // Intentional fall-through into DONE_LABEL ...
1916 // Propagate ICC.ZF from CAS above into DONE_LABEL.
1917 #endif // _LP64
1918 #if INCLUDE_RTM_OPT
1919 } // use_rtm()
1920 #endif
1921 // DONE_LABEL is a hot target - we'd really like to place it at the
1922 // start of cache line by padding with NOPs.
1923 // See the AMD and Intel software optimization manuals for the
1924 // most efficient "long" NOP encodings.
1925 // Unfortunately none of our alignment mechanisms suffice.
1926 bind(DONE_LABEL);
1927
1928 // At DONE_LABEL the icc ZFlag is set as follows ...
1929 // Fast_Unlock uses the same protocol.
1930 // ZFlag == 1 -> Success
1931 // ZFlag == 0 -> Failure - force control through the slow-path
1932 }
1933 }
1934
1935 // obj: object to unlock
1936 // box: box address (displaced header location), killed. Must be EAX.
1937 // tmp: killed, cannot be obj nor box.
1938 //
1939 // Some commentary on balanced locking:
1940 //
1941 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
1942 // Methods that don't have provably balanced locking are forced to run in the
1943 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
1944 // The interpreter provides two properties:
1945 // I1: At return-time the interpreter automatically and quietly unlocks any
1946 // objects acquired the current activation (frame). Recall that the
1947 // interpreter maintains an on-stack list of locks currently held by
1948 // a frame.
1949 // I2: If a method attempts to unlock an object that is not held by the
1950 // the frame the interpreter throws IMSX.
1951 //
1952 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
1953 // B() doesn't have provably balanced locking so it runs in the interpreter.
1954 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
1955 // is still locked by A().
1956 //
1957 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
1958 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
1959 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
1960 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
1961 // Arguably given that the spec legislates the JNI case as undefined our implementation
1962 // could reasonably *avoid* checking owner in Fast_Unlock().
1963 // In the interest of performance we elide m->Owner==Self check in unlock.
1964 // A perfectly viable alternative is to elide the owner check except when
1965 // Xcheck:jni is enabled.
1966
1967 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) {
1968 assert(boxReg == rax, "");
1969 assert_different_registers(objReg, boxReg, tmpReg);
1970
1971 if (EmitSync & 4) {
1972 // Disable - inhibit all inlining. Force control through the slow-path
1973 cmpptr (rsp, 0);
1974 } else {
1975 Label DONE_LABEL, Stacked, CheckSucc;
1976
1977 // Critically, the biased locking test must have precedence over
1978 // and appear before the (box->dhw == 0) recursive stack-lock test.
1979 if (UseBiasedLocking && !UseOptoBiasInlining) {
1980 biased_locking_exit(objReg, tmpReg, DONE_LABEL);
1981 }
1982
1983 #if INCLUDE_RTM_OPT
1984 if (UseRTMForStackLocks && use_rtm) {
1985 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1986 Label L_regular_unlock;
1987 movptr(tmpReg, Address(objReg, 0)); // fetch markword
1988 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1989 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1990 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock
1991 xend(); // otherwise end...
1992 jmp(DONE_LABEL); // ... and we're done
1993 bind(L_regular_unlock);
1994 }
1995 #endif
1996
1997 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header
1998 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock
1999 movptr(tmpReg, Address(objReg, 0)); // Examine the object's markword
2000 testptr(tmpReg, markOopDesc::monitor_value); // Inflated?
2001 jccb (Assembler::zero, Stacked);
2002
2003 // It's inflated.
2004 #if INCLUDE_RTM_OPT
2005 if (use_rtm) {
2006 Label L_regular_inflated_unlock;
2007 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
2008 movptr(boxReg, Address(tmpReg, owner_offset));
2009 testptr(boxReg, boxReg);
2010 jccb(Assembler::notZero, L_regular_inflated_unlock);
2011 xend();
2012 jmpb(DONE_LABEL);
2013 bind(L_regular_inflated_unlock);
2014 }
2015 #endif
2016
2017 // Despite our balanced locking property we still check that m->_owner == Self
2018 // as java routines or native JNI code called by this thread might
2019 // have released the lock.
2020 // Refer to the comments in synchronizer.cpp for how we might encode extra
2021 // state in _succ so we can avoid fetching EntryList|cxq.
2022 //
2023 // I'd like to add more cases in fast_lock() and fast_unlock() --
2024 // such as recursive enter and exit -- but we have to be wary of
2025 // I$ bloat, T$ effects and BP$ effects.
2026 //
2027 // If there's no contention try a 1-0 exit. That is, exit without
2028 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
2029 // we detect and recover from the race that the 1-0 exit admits.
2030 //
2031 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
2032 // before it STs null into _owner, releasing the lock. Updates
2033 // to data protected by the critical section must be visible before
2034 // we drop the lock (and thus before any other thread could acquire
2035 // the lock and observe the fields protected by the lock).
2036 // IA32's memory-model is SPO, so STs are ordered with respect to
2037 // each other and there's no need for an explicit barrier (fence).
2038 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
2039 #ifndef _LP64
2040 get_thread (boxReg);
2041 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
2042 // prefetchw [ebx + Offset(_owner)-2]
2043 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2044 }
2045
2046 // Note that we could employ various encoding schemes to reduce
2047 // the number of loads below (currently 4) to just 2 or 3.
2048 // Refer to the comments in synchronizer.cpp.
2049 // In practice the chain of fetches doesn't seem to impact performance, however.
2050 xorptr(boxReg, boxReg);
2051 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) {
2052 // Attempt to reduce branch density - AMD's branch predictor.
2053 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2054 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2055 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2056 jccb (Assembler::notZero, DONE_LABEL);
2057 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2058 jmpb (DONE_LABEL);
2059 } else {
2060 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2061 jccb (Assembler::notZero, DONE_LABEL);
2062 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2063 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2064 jccb (Assembler::notZero, CheckSucc);
2065 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2066 jmpb (DONE_LABEL);
2067 }
2068
2069 // The Following code fragment (EmitSync & 65536) improves the performance of
2070 // contended applications and contended synchronization microbenchmarks.
2071 // Unfortunately the emission of the code - even though not executed - causes regressions
2072 // in scimark and jetstream, evidently because of $ effects. Replacing the code
2073 // with an equal number of never-executed NOPs results in the same regression.
2074 // We leave it off by default.
2075
2076 if ((EmitSync & 65536) != 0) {
2077 Label LSuccess, LGoSlowPath ;
2078
2079 bind (CheckSucc);
2080
2081 // Optional pre-test ... it's safe to elide this
2082 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2083 jccb(Assembler::zero, LGoSlowPath);
2084
2085 // We have a classic Dekker-style idiom:
2086 // ST m->_owner = 0 ; MEMBAR; LD m->_succ
2087 // There are a number of ways to implement the barrier:
2088 // (1) lock:andl &m->_owner, 0
2089 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form.
2090 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0
2091 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8
2092 // (2) If supported, an explicit MFENCE is appealing.
2093 // In older IA32 processors MFENCE is slower than lock:add or xchg
2094 // particularly if the write-buffer is full as might be the case if
2095 // if stores closely precede the fence or fence-equivalent instruction.
2096 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2097 // as the situation has changed with Nehalem and Shanghai.
2098 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack
2099 // The $lines underlying the top-of-stack should be in M-state.
2100 // The locked add instruction is serializing, of course.
2101 // (4) Use xchg, which is serializing
2102 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works
2103 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0.
2104 // The integer condition codes will tell us if succ was 0.
2105 // Since _succ and _owner should reside in the same $line and
2106 // we just stored into _owner, it's likely that the $line
2107 // remains in M-state for the lock:orl.
2108 //
2109 // We currently use (3), although it's likely that switching to (2)
2110 // is correct for the future.
2111
2112 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2113 if (os::is_MP()) {
2114 lock(); addptr(Address(rsp, 0), 0);
2115 }
2116 // Ratify _succ remains non-null
2117 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0);
2118 jccb (Assembler::notZero, LSuccess);
2119
2120 xorptr(boxReg, boxReg); // box is really EAX
2121 if (os::is_MP()) { lock(); }
2122 cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2123 // There's no successor so we tried to regrab the lock with the
2124 // placeholder value. If that didn't work, then another thread
2125 // grabbed the lock so we're done (and exit was a success).
2126 jccb (Assembler::notEqual, LSuccess);
2127 // Since we're low on registers we installed rsp as a placeholding in _owner.
2128 // Now install Self over rsp. This is safe as we're transitioning from
2129 // non-null to non=null
2130 get_thread (boxReg);
2131 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg);
2132 // Intentional fall-through into LGoSlowPath ...
2133
2134 bind (LGoSlowPath);
2135 orptr(boxReg, 1); // set ICC.ZF=0 to indicate failure
2136 jmpb (DONE_LABEL);
2137
2138 bind (LSuccess);
2139 xorptr(boxReg, boxReg); // set ICC.ZF=1 to indicate success
2140 jmpb (DONE_LABEL);
2141 }
2142
2143 bind (Stacked);
2144 // It's not inflated and it's not recursively stack-locked and it's not biased.
2145 // It must be stack-locked.
2146 // Try to reset the header to displaced header.
2147 // The "box" value on the stack is stable, so we can reload
2148 // and be assured we observe the same value as above.
2149 movptr(tmpReg, Address(boxReg, 0));
2150 if (os::is_MP()) {
2151 lock();
2152 }
2153 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box
2154 // Intention fall-thru into DONE_LABEL
2155
2156 // DONE_LABEL is a hot target - we'd really like to place it at the
2157 // start of cache line by padding with NOPs.
2158 // See the AMD and Intel software optimization manuals for the
2159 // most efficient "long" NOP encodings.
2160 // Unfortunately none of our alignment mechanisms suffice.
2161 if ((EmitSync & 65536) == 0) {
2162 bind (CheckSucc);
2163 }
2164 #else // _LP64
2165 // It's inflated
2166 if (EmitSync & 1024) {
2167 // Emit code to check that _owner == Self
2168 // We could fold the _owner test into subsequent code more efficiently
2169 // than using a stand-alone check, but since _owner checking is off by
2170 // default we don't bother. We also might consider predicating the
2171 // _owner==Self check on Xcheck:jni or running on a debug build.
2172 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2173 xorptr(boxReg, r15_thread);
2174 } else {
2175 xorptr(boxReg, boxReg);
2176 }
2177 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2178 jccb (Assembler::notZero, DONE_LABEL);
2179 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2180 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2181 jccb (Assembler::notZero, CheckSucc);
2182 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2183 jmpb (DONE_LABEL);
2184
2185 if ((EmitSync & 65536) == 0) {
2186 // Try to avoid passing control into the slow_path ...
2187 Label LSuccess, LGoSlowPath ;
2188 bind (CheckSucc);
2189
2190 // The following optional optimization can be elided if necessary
2191 // Effectively: if (succ == null) goto SlowPath
2192 // The code reduces the window for a race, however,
2193 // and thus benefits performance.
2194 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2195 jccb (Assembler::zero, LGoSlowPath);
2196
2197 xorptr(boxReg, boxReg);
2198 if ((EmitSync & 16) && os::is_MP()) {
2199 xchgptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2200 } else {
2201 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2202 if (os::is_MP()) {
2203 // Memory barrier/fence
2204 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ
2205 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack.
2206 // This is faster on Nehalem and AMD Shanghai/Barcelona.
2207 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2208 // We might also restructure (ST Owner=0;barrier;LD _Succ) to
2209 // (mov box,0; xchgq box, &m->Owner; LD _succ) .
2210 lock(); addl(Address(rsp, 0), 0);
2211 }
2212 }
2213 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2214 jccb (Assembler::notZero, LSuccess);
2215
2216 // Rare inopportune interleaving - race.
2217 // The successor vanished in the small window above.
2218 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor.
2219 // We need to ensure progress and succession.
2220 // Try to reacquire the lock.
2221 // If that fails then the new owner is responsible for succession and this
2222 // thread needs to take no further action and can exit via the fast path (success).
2223 // If the re-acquire succeeds then pass control into the slow path.
2224 // As implemented, this latter mode is horrible because we generated more
2225 // coherence traffic on the lock *and* artifically extended the critical section
2226 // length while by virtue of passing control into the slow path.
2227
2228 // box is really RAX -- the following CMPXCHG depends on that binding
2229 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R)
2230 if (os::is_MP()) { lock(); }
2231 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2232 // There's no successor so we tried to regrab the lock.
2233 // If that didn't work, then another thread grabbed the
2234 // lock so we're done (and exit was a success).
2235 jccb (Assembler::notEqual, LSuccess);
2236 // Intentional fall-through into slow-path
2237
2238 bind (LGoSlowPath);
2239 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure
2240 jmpb (DONE_LABEL);
2241
2242 bind (LSuccess);
2243 testl (boxReg, 0); // set ICC.ZF=1 to indicate success
2244 jmpb (DONE_LABEL);
2245 }
2246
2247 bind (Stacked);
2248 movptr(tmpReg, Address (boxReg, 0)); // re-fetch
2249 if (os::is_MP()) { lock(); }
2250 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box
2251
2252 if (EmitSync & 65536) {
2253 bind (CheckSucc);
2254 }
2255 #endif
2256 bind(DONE_LABEL);
2257 }
2258 }
2259 #endif // COMPILER2
2260
2261 void MacroAssembler::c2bool(Register x) {
2262 // implements x == 0 ? 0 : 1
2263 // note: must only look at least-significant byte of x
2264 // since C-style booleans are stored in one byte
2265 // only! (was bug)
2266 andl(x, 0xFF);
2267 setb(Assembler::notZero, x);
2268 }
2269
2270 // Wouldn't need if AddressLiteral version had new name
2271 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) {
2272 Assembler::call(L, rtype);
2273 }
2274
2275 void MacroAssembler::call(Register entry) {
2276 Assembler::call(entry);
2277 }
2278
2279 void MacroAssembler::call(AddressLiteral entry) {
2280 if (reachable(entry)) {
2281 Assembler::call_literal(entry.target(), entry.rspec());
2282 } else {
2283 lea(rscratch1, entry);
2284 Assembler::call(rscratch1);
2285 }
2286 }
2287
2288 void MacroAssembler::ic_call(address entry, jint method_index) {
2289 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index);
2290 movptr(rax, (intptr_t)Universe::non_oop_word());
2291 call(AddressLiteral(entry, rh));
2292 }
2293
2294 // Implementation of call_VM versions
2295
2296 void MacroAssembler::call_VM(Register oop_result,
2297 address entry_point,
2298 bool check_exceptions) {
2299 Label C, E;
2300 call(C, relocInfo::none);
2301 jmp(E);
2302
2303 bind(C);
2304 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
2305 ret(0);
2306
2307 bind(E);
2308 }
2309
2310 void MacroAssembler::call_VM(Register oop_result,
2311 address entry_point,
2312 Register arg_1,
2313 bool check_exceptions) {
2314 Label C, E;
2315 call(C, relocInfo::none);
2316 jmp(E);
2317
2318 bind(C);
2319 pass_arg1(this, arg_1);
2320 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
2321 ret(0);
2322
2323 bind(E);
2324 }
2325
2326 void MacroAssembler::call_VM(Register oop_result,
2327 address entry_point,
2328 Register arg_1,
2329 Register arg_2,
2330 bool check_exceptions) {
2331 Label C, E;
2332 call(C, relocInfo::none);
2333 jmp(E);
2334
2335 bind(C);
2336
2337 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2338
2339 pass_arg2(this, arg_2);
2340 pass_arg1(this, arg_1);
2341 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
2342 ret(0);
2343
2344 bind(E);
2345 }
2346
2347 void MacroAssembler::call_VM(Register oop_result,
2348 address entry_point,
2349 Register arg_1,
2350 Register arg_2,
2351 Register arg_3,
2352 bool check_exceptions) {
2353 Label C, E;
2354 call(C, relocInfo::none);
2355 jmp(E);
2356
2357 bind(C);
2358
2359 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2360 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2361 pass_arg3(this, arg_3);
2362
2363 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2364 pass_arg2(this, arg_2);
2365
2366 pass_arg1(this, arg_1);
2367 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
2368 ret(0);
2369
2370 bind(E);
2371 }
2372
2373 void MacroAssembler::call_VM(Register oop_result,
2374 Register last_java_sp,
2375 address entry_point,
2376 int number_of_arguments,
2377 bool check_exceptions) {
2378 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2379 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2380 }
2381
2382 void MacroAssembler::call_VM(Register oop_result,
2383 Register last_java_sp,
2384 address entry_point,
2385 Register arg_1,
2386 bool check_exceptions) {
2387 pass_arg1(this, arg_1);
2388 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2389 }
2390
2391 void MacroAssembler::call_VM(Register oop_result,
2392 Register last_java_sp,
2393 address entry_point,
2394 Register arg_1,
2395 Register arg_2,
2396 bool check_exceptions) {
2397
2398 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2399 pass_arg2(this, arg_2);
2400 pass_arg1(this, arg_1);
2401 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2402 }
2403
2404 void MacroAssembler::call_VM(Register oop_result,
2405 Register last_java_sp,
2406 address entry_point,
2407 Register arg_1,
2408 Register arg_2,
2409 Register arg_3,
2410 bool check_exceptions) {
2411 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2412 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2413 pass_arg3(this, arg_3);
2414 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2415 pass_arg2(this, arg_2);
2416 pass_arg1(this, arg_1);
2417 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2418 }
2419
2420 void MacroAssembler::super_call_VM(Register oop_result,
2421 Register last_java_sp,
2422 address entry_point,
2423 int number_of_arguments,
2424 bool check_exceptions) {
2425 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2426 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2427 }
2428
2429 void MacroAssembler::super_call_VM(Register oop_result,
2430 Register last_java_sp,
2431 address entry_point,
2432 Register arg_1,
2433 bool check_exceptions) {
2434 pass_arg1(this, arg_1);
2435 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2436 }
2437
2438 void MacroAssembler::super_call_VM(Register oop_result,
2439 Register last_java_sp,
2440 address entry_point,
2441 Register arg_1,
2442 Register arg_2,
2443 bool check_exceptions) {
2444
2445 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2446 pass_arg2(this, arg_2);
2447 pass_arg1(this, arg_1);
2448 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2449 }
2450
2451 void MacroAssembler::super_call_VM(Register oop_result,
2452 Register last_java_sp,
2453 address entry_point,
2454 Register arg_1,
2455 Register arg_2,
2456 Register arg_3,
2457 bool check_exceptions) {
2458 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2459 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2460 pass_arg3(this, arg_3);
2461 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2462 pass_arg2(this, arg_2);
2463 pass_arg1(this, arg_1);
2464 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2465 }
2466
2467 void MacroAssembler::call_VM_base(Register oop_result,
2468 Register java_thread,
2469 Register last_java_sp,
2470 address entry_point,
2471 int number_of_arguments,
2472 bool check_exceptions) {
2473 // determine java_thread register
2474 if (!java_thread->is_valid()) {
2475 #ifdef _LP64
2476 java_thread = r15_thread;
2477 #else
2478 java_thread = rdi;
2479 get_thread(java_thread);
2480 #endif // LP64
2481 }
2482 // determine last_java_sp register
2483 if (!last_java_sp->is_valid()) {
2484 last_java_sp = rsp;
2485 }
2486 // debugging support
2487 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
2488 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register"));
2489 #ifdef ASSERT
2490 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
2491 // r12 is the heapbase.
2492 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");)
2493 #endif // ASSERT
2494
2495 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
2496 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
2497
2498 // push java thread (becomes first argument of C function)
2499
2500 NOT_LP64(push(java_thread); number_of_arguments++);
2501 LP64_ONLY(mov(c_rarg0, r15_thread));
2502
2503 // set last Java frame before call
2504 assert(last_java_sp != rbp, "can't use ebp/rbp");
2505
2506 // Only interpreter should have to set fp
2507 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL);
2508
2509 // do the call, remove parameters
2510 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
2511
2512 // restore the thread (cannot use the pushed argument since arguments
2513 // may be overwritten by C code generated by an optimizing compiler);
2514 // however can use the register value directly if it is callee saved.
2515 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) {
2516 // rdi & rsi (also r15) are callee saved -> nothing to do
2517 #ifdef ASSERT
2518 guarantee(java_thread != rax, "change this code");
2519 push(rax);
2520 { Label L;
2521 get_thread(rax);
2522 cmpptr(java_thread, rax);
2523 jcc(Assembler::equal, L);
2524 STOP("MacroAssembler::call_VM_base: rdi not callee saved?");
2525 bind(L);
2526 }
2527 pop(rax);
2528 #endif
2529 } else {
2530 get_thread(java_thread);
2531 }
2532 // reset last Java frame
2533 // Only interpreter should have to clear fp
2534 reset_last_Java_frame(java_thread, true, false);
2535
2536 // C++ interp handles this in the interpreter
2537 check_and_handle_popframe(java_thread);
2538 check_and_handle_earlyret(java_thread);
2539
2540 if (check_exceptions) {
2541 // check for pending exceptions (java_thread is set upon return)
2542 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD);
2543 #ifndef _LP64
2544 jump_cc(Assembler::notEqual,
2545 RuntimeAddress(StubRoutines::forward_exception_entry()));
2546 #else
2547 // This used to conditionally jump to forward_exception however it is
2548 // possible if we relocate that the branch will not reach. So we must jump
2549 // around so we can always reach
2550
2551 Label ok;
2552 jcc(Assembler::equal, ok);
2553 jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2554 bind(ok);
2555 #endif // LP64
2556 }
2557
2558 // get oop result if there is one and reset the value in the thread
2559 if (oop_result->is_valid()) {
2560 get_vm_result(oop_result, java_thread);
2561 }
2562 }
2563
2564 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
2565
2566 // Calculate the value for last_Java_sp
2567 // somewhat subtle. call_VM does an intermediate call
2568 // which places a return address on the stack just under the
2569 // stack pointer as the user finsihed with it. This allows
2570 // use to retrieve last_Java_pc from last_Java_sp[-1].
2571 // On 32bit we then have to push additional args on the stack to accomplish
2572 // the actual requested call. On 64bit call_VM only can use register args
2573 // so the only extra space is the return address that call_VM created.
2574 // This hopefully explains the calculations here.
2575
2576 #ifdef _LP64
2577 // We've pushed one address, correct last_Java_sp
2578 lea(rax, Address(rsp, wordSize));
2579 #else
2580 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize));
2581 #endif // LP64
2582
2583 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions);
2584
2585 }
2586
2587 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
2588 call_VM_leaf_base(entry_point, number_of_arguments);
2589 }
2590
2591 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
2592 pass_arg0(this, arg_0);
2593 call_VM_leaf(entry_point, 1);
2594 }
2595
2596 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2597
2598 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2599 pass_arg1(this, arg_1);
2600 pass_arg0(this, arg_0);
2601 call_VM_leaf(entry_point, 2);
2602 }
2603
2604 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2605 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2606 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2607 pass_arg2(this, arg_2);
2608 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2609 pass_arg1(this, arg_1);
2610 pass_arg0(this, arg_0);
2611 call_VM_leaf(entry_point, 3);
2612 }
2613
2614 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2615 pass_arg0(this, arg_0);
2616 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2617 }
2618
2619 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2620
2621 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2622 pass_arg1(this, arg_1);
2623 pass_arg0(this, arg_0);
2624 MacroAssembler::call_VM_leaf_base(entry_point, 2);
2625 }
2626
2627 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2628 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2629 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2630 pass_arg2(this, arg_2);
2631 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2632 pass_arg1(this, arg_1);
2633 pass_arg0(this, arg_0);
2634 MacroAssembler::call_VM_leaf_base(entry_point, 3);
2635 }
2636
2637 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2638 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg"));
2639 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2640 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2641 pass_arg3(this, arg_3);
2642 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2643 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2644 pass_arg2(this, arg_2);
2645 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2646 pass_arg1(this, arg_1);
2647 pass_arg0(this, arg_0);
2648 MacroAssembler::call_VM_leaf_base(entry_point, 4);
2649 }
2650
2651 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
2652 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
2653 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD);
2654 verify_oop(oop_result, "broken oop in call_VM_base");
2655 }
2656
2657 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
2658 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
2659 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD);
2660 }
2661
2662 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2663 }
2664
2665 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2666 }
2667
2668 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) {
2669 if (reachable(src1)) {
2670 cmpl(as_Address(src1), imm);
2671 } else {
2672 lea(rscratch1, src1);
2673 cmpl(Address(rscratch1, 0), imm);
2674 }
2675 }
2676
2677 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) {
2678 assert(!src2.is_lval(), "use cmpptr");
2679 if (reachable(src2)) {
2680 cmpl(src1, as_Address(src2));
2681 } else {
2682 lea(rscratch1, src2);
2683 cmpl(src1, Address(rscratch1, 0));
2684 }
2685 }
2686
2687 void MacroAssembler::cmp32(Register src1, int32_t imm) {
2688 Assembler::cmpl(src1, imm);
2689 }
2690
2691 void MacroAssembler::cmp32(Register src1, Address src2) {
2692 Assembler::cmpl(src1, src2);
2693 }
2694
2695 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2696 ucomisd(opr1, opr2);
2697
2698 Label L;
2699 if (unordered_is_less) {
2700 movl(dst, -1);
2701 jcc(Assembler::parity, L);
2702 jcc(Assembler::below , L);
2703 movl(dst, 0);
2704 jcc(Assembler::equal , L);
2705 increment(dst);
2706 } else { // unordered is greater
2707 movl(dst, 1);
2708 jcc(Assembler::parity, L);
2709 jcc(Assembler::above , L);
2710 movl(dst, 0);
2711 jcc(Assembler::equal , L);
2712 decrementl(dst);
2713 }
2714 bind(L);
2715 }
2716
2717 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2718 ucomiss(opr1, opr2);
2719
2720 Label L;
2721 if (unordered_is_less) {
2722 movl(dst, -1);
2723 jcc(Assembler::parity, L);
2724 jcc(Assembler::below , L);
2725 movl(dst, 0);
2726 jcc(Assembler::equal , L);
2727 increment(dst);
2728 } else { // unordered is greater
2729 movl(dst, 1);
2730 jcc(Assembler::parity, L);
2731 jcc(Assembler::above , L);
2732 movl(dst, 0);
2733 jcc(Assembler::equal , L);
2734 decrementl(dst);
2735 }
2736 bind(L);
2737 }
2738
2739
2740 void MacroAssembler::cmp8(AddressLiteral src1, int imm) {
2741 if (reachable(src1)) {
2742 cmpb(as_Address(src1), imm);
2743 } else {
2744 lea(rscratch1, src1);
2745 cmpb(Address(rscratch1, 0), imm);
2746 }
2747 }
2748
2749 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) {
2750 #ifdef _LP64
2751 if (src2.is_lval()) {
2752 movptr(rscratch1, src2);
2753 Assembler::cmpq(src1, rscratch1);
2754 } else if (reachable(src2)) {
2755 cmpq(src1, as_Address(src2));
2756 } else {
2757 lea(rscratch1, src2);
2758 Assembler::cmpq(src1, Address(rscratch1, 0));
2759 }
2760 #else
2761 if (src2.is_lval()) {
2762 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2763 } else {
2764 cmpl(src1, as_Address(src2));
2765 }
2766 #endif // _LP64
2767 }
2768
2769 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) {
2770 assert(src2.is_lval(), "not a mem-mem compare");
2771 #ifdef _LP64
2772 // moves src2's literal address
2773 movptr(rscratch1, src2);
2774 Assembler::cmpq(src1, rscratch1);
2775 #else
2776 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2777 #endif // _LP64
2778 }
2779
2780 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) {
2781 if (reachable(adr)) {
2782 if (os::is_MP())
2783 lock();
2784 cmpxchgptr(reg, as_Address(adr));
2785 } else {
2786 lea(rscratch1, adr);
2787 if (os::is_MP())
2788 lock();
2789 cmpxchgptr(reg, Address(rscratch1, 0));
2790 }
2791 }
2792
2793 void MacroAssembler::cmpxchgptr(Register reg, Address adr) {
2794 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr));
2795 }
2796
2797 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) {
2798 if (reachable(src)) {
2799 Assembler::comisd(dst, as_Address(src));
2800 } else {
2801 lea(rscratch1, src);
2802 Assembler::comisd(dst, Address(rscratch1, 0));
2803 }
2804 }
2805
2806 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) {
2807 if (reachable(src)) {
2808 Assembler::comiss(dst, as_Address(src));
2809 } else {
2810 lea(rscratch1, src);
2811 Assembler::comiss(dst, Address(rscratch1, 0));
2812 }
2813 }
2814
2815
2816 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) {
2817 Condition negated_cond = negate_condition(cond);
2818 Label L;
2819 jcc(negated_cond, L);
2820 pushf(); // Preserve flags
2821 atomic_incl(counter_addr);
2822 popf();
2823 bind(L);
2824 }
2825
2826 int MacroAssembler::corrected_idivl(Register reg) {
2827 // Full implementation of Java idiv and irem; checks for
2828 // special case as described in JVM spec., p.243 & p.271.
2829 // The function returns the (pc) offset of the idivl
2830 // instruction - may be needed for implicit exceptions.
2831 //
2832 // normal case special case
2833 //
2834 // input : rax,: dividend min_int
2835 // reg: divisor (may not be rax,/rdx) -1
2836 //
2837 // output: rax,: quotient (= rax, idiv reg) min_int
2838 // rdx: remainder (= rax, irem reg) 0
2839 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register");
2840 const int min_int = 0x80000000;
2841 Label normal_case, special_case;
2842
2843 // check for special case
2844 cmpl(rax, min_int);
2845 jcc(Assembler::notEqual, normal_case);
2846 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0)
2847 cmpl(reg, -1);
2848 jcc(Assembler::equal, special_case);
2849
2850 // handle normal case
2851 bind(normal_case);
2852 cdql();
2853 int idivl_offset = offset();
2854 idivl(reg);
2855
2856 // normal and special case exit
2857 bind(special_case);
2858
2859 return idivl_offset;
2860 }
2861
2862
2863
2864 void MacroAssembler::decrementl(Register reg, int value) {
2865 if (value == min_jint) {subl(reg, value) ; return; }
2866 if (value < 0) { incrementl(reg, -value); return; }
2867 if (value == 0) { ; return; }
2868 if (value == 1 && UseIncDec) { decl(reg) ; return; }
2869 /* else */ { subl(reg, value) ; return; }
2870 }
2871
2872 void MacroAssembler::decrementl(Address dst, int value) {
2873 if (value == min_jint) {subl(dst, value) ; return; }
2874 if (value < 0) { incrementl(dst, -value); return; }
2875 if (value == 0) { ; return; }
2876 if (value == 1 && UseIncDec) { decl(dst) ; return; }
2877 /* else */ { subl(dst, value) ; return; }
2878 }
2879
2880 void MacroAssembler::division_with_shift (Register reg, int shift_value) {
2881 assert (shift_value > 0, "illegal shift value");
2882 Label _is_positive;
2883 testl (reg, reg);
2884 jcc (Assembler::positive, _is_positive);
2885 int offset = (1 << shift_value) - 1 ;
2886
2887 if (offset == 1) {
2888 incrementl(reg);
2889 } else {
2890 addl(reg, offset);
2891 }
2892
2893 bind (_is_positive);
2894 sarl(reg, shift_value);
2895 }
2896
2897 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) {
2898 if (reachable(src)) {
2899 Assembler::divsd(dst, as_Address(src));
2900 } else {
2901 lea(rscratch1, src);
2902 Assembler::divsd(dst, Address(rscratch1, 0));
2903 }
2904 }
2905
2906 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) {
2907 if (reachable(src)) {
2908 Assembler::divss(dst, as_Address(src));
2909 } else {
2910 lea(rscratch1, src);
2911 Assembler::divss(dst, Address(rscratch1, 0));
2912 }
2913 }
2914
2915 // !defined(COMPILER2) is because of stupid core builds
2916 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI
2917 void MacroAssembler::empty_FPU_stack() {
2918 if (VM_Version::supports_mmx()) {
2919 emms();
2920 } else {
2921 for (int i = 8; i-- > 0; ) ffree(i);
2922 }
2923 }
2924 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI
2925
2926
2927 // Defines obj, preserves var_size_in_bytes
2928 void MacroAssembler::eden_allocate(Register obj,
2929 Register var_size_in_bytes,
2930 int con_size_in_bytes,
2931 Register t1,
2932 Label& slow_case) {
2933 assert(obj == rax, "obj must be in rax, for cmpxchg");
2934 assert_different_registers(obj, var_size_in_bytes, t1);
2935 if (!Universe::heap()->supports_inline_contig_alloc()) {
2936 jmp(slow_case);
2937 } else {
2938 Register end = t1;
2939 Label retry;
2940 bind(retry);
2941 ExternalAddress heap_top((address) Universe::heap()->top_addr());
2942 movptr(obj, heap_top);
2943 if (var_size_in_bytes == noreg) {
2944 lea(end, Address(obj, con_size_in_bytes));
2945 } else {
2946 lea(end, Address(obj, var_size_in_bytes, Address::times_1));
2947 }
2948 // if end < obj then we wrapped around => object too long => slow case
2949 cmpptr(end, obj);
2950 jcc(Assembler::below, slow_case);
2951 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr()));
2952 jcc(Assembler::above, slow_case);
2953 // Compare obj with the top addr, and if still equal, store the new top addr in
2954 // end at the address of the top addr pointer. Sets ZF if was equal, and clears
2955 // it otherwise. Use lock prefix for atomicity on MPs.
2956 locked_cmpxchgptr(end, heap_top);
2957 jcc(Assembler::notEqual, retry);
2958 }
2959 }
2960
2961 void MacroAssembler::enter() {
2962 push(rbp);
2963 mov(rbp, rsp);
2964 }
2965
2966 // A 5 byte nop that is safe for patching (see patch_verified_entry)
2967 void MacroAssembler::fat_nop() {
2968 if (UseAddressNop) {
2969 addr_nop_5();
2970 } else {
2971 emit_int8(0x26); // es:
2972 emit_int8(0x2e); // cs:
2973 emit_int8(0x64); // fs:
2974 emit_int8(0x65); // gs:
2975 emit_int8((unsigned char)0x90);
2976 }
2977 }
2978
2979 void MacroAssembler::fcmp(Register tmp) {
2980 fcmp(tmp, 1, true, true);
2981 }
2982
2983 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) {
2984 assert(!pop_right || pop_left, "usage error");
2985 if (VM_Version::supports_cmov()) {
2986 assert(tmp == noreg, "unneeded temp");
2987 if (pop_left) {
2988 fucomip(index);
2989 } else {
2990 fucomi(index);
2991 }
2992 if (pop_right) {
2993 fpop();
2994 }
2995 } else {
2996 assert(tmp != noreg, "need temp");
2997 if (pop_left) {
2998 if (pop_right) {
2999 fcompp();
3000 } else {
3001 fcomp(index);
3002 }
3003 } else {
3004 fcom(index);
3005 }
3006 // convert FPU condition into eflags condition via rax,
3007 save_rax(tmp);
3008 fwait(); fnstsw_ax();
3009 sahf();
3010 restore_rax(tmp);
3011 }
3012 // condition codes set as follows:
3013 //
3014 // CF (corresponds to C0) if x < y
3015 // PF (corresponds to C2) if unordered
3016 // ZF (corresponds to C3) if x = y
3017 }
3018
3019 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) {
3020 fcmp2int(dst, unordered_is_less, 1, true, true);
3021 }
3022
3023 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) {
3024 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right);
3025 Label L;
3026 if (unordered_is_less) {
3027 movl(dst, -1);
3028 jcc(Assembler::parity, L);
3029 jcc(Assembler::below , L);
3030 movl(dst, 0);
3031 jcc(Assembler::equal , L);
3032 increment(dst);
3033 } else { // unordered is greater
3034 movl(dst, 1);
3035 jcc(Assembler::parity, L);
3036 jcc(Assembler::above , L);
3037 movl(dst, 0);
3038 jcc(Assembler::equal , L);
3039 decrementl(dst);
3040 }
3041 bind(L);
3042 }
3043
3044 void MacroAssembler::fld_d(AddressLiteral src) {
3045 fld_d(as_Address(src));
3046 }
3047
3048 void MacroAssembler::fld_s(AddressLiteral src) {
3049 fld_s(as_Address(src));
3050 }
3051
3052 void MacroAssembler::fld_x(AddressLiteral src) {
3053 Assembler::fld_x(as_Address(src));
3054 }
3055
3056 void MacroAssembler::fldcw(AddressLiteral src) {
3057 Assembler::fldcw(as_Address(src));
3058 }
3059
3060 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) {
3061 if (reachable(src)) {
3062 Assembler::mulpd(dst, as_Address(src));
3063 } else {
3064 lea(rscratch1, src);
3065 Assembler::mulpd(dst, Address(rscratch1, 0));
3066 }
3067 }
3068
3069 void MacroAssembler::increase_precision() {
3070 subptr(rsp, BytesPerWord);
3071 fnstcw(Address(rsp, 0));
3072 movl(rax, Address(rsp, 0));
3073 orl(rax, 0x300);
3074 push(rax);
3075 fldcw(Address(rsp, 0));
3076 pop(rax);
3077 }
3078
3079 void MacroAssembler::restore_precision() {
3080 fldcw(Address(rsp, 0));
3081 addptr(rsp, BytesPerWord);
3082 }
3083
3084 void MacroAssembler::fpop() {
3085 ffree();
3086 fincstp();
3087 }
3088
3089 void MacroAssembler::load_float(Address src) {
3090 if (UseSSE >= 1) {
3091 movflt(xmm0, src);
3092 } else {
3093 LP64_ONLY(ShouldNotReachHere());
3094 NOT_LP64(fld_s(src));
3095 }
3096 }
3097
3098 void MacroAssembler::store_float(Address dst) {
3099 if (UseSSE >= 1) {
3100 movflt(dst, xmm0);
3101 } else {
3102 LP64_ONLY(ShouldNotReachHere());
3103 NOT_LP64(fstp_s(dst));
3104 }
3105 }
3106
3107 void MacroAssembler::load_double(Address src) {
3108 if (UseSSE >= 2) {
3109 movdbl(xmm0, src);
3110 } else {
3111 LP64_ONLY(ShouldNotReachHere());
3112 NOT_LP64(fld_d(src));
3113 }
3114 }
3115
3116 void MacroAssembler::store_double(Address dst) {
3117 if (UseSSE >= 2) {
3118 movdbl(dst, xmm0);
3119 } else {
3120 LP64_ONLY(ShouldNotReachHere());
3121 NOT_LP64(fstp_d(dst));
3122 }
3123 }
3124
3125 void MacroAssembler::fremr(Register tmp) {
3126 save_rax(tmp);
3127 { Label L;
3128 bind(L);
3129 fprem();
3130 fwait(); fnstsw_ax();
3131 #ifdef _LP64
3132 testl(rax, 0x400);
3133 jcc(Assembler::notEqual, L);
3134 #else
3135 sahf();
3136 jcc(Assembler::parity, L);
3137 #endif // _LP64
3138 }
3139 restore_rax(tmp);
3140 // Result is in ST0.
3141 // Note: fxch & fpop to get rid of ST1
3142 // (otherwise FPU stack could overflow eventually)
3143 fxch(1);
3144 fpop();
3145 }
3146
3147
3148 void MacroAssembler::incrementl(AddressLiteral dst) {
3149 if (reachable(dst)) {
3150 incrementl(as_Address(dst));
3151 } else {
3152 lea(rscratch1, dst);
3153 incrementl(Address(rscratch1, 0));
3154 }
3155 }
3156
3157 void MacroAssembler::incrementl(ArrayAddress dst) {
3158 incrementl(as_Address(dst));
3159 }
3160
3161 void MacroAssembler::incrementl(Register reg, int value) {
3162 if (value == min_jint) {addl(reg, value) ; return; }
3163 if (value < 0) { decrementl(reg, -value); return; }
3164 if (value == 0) { ; return; }
3165 if (value == 1 && UseIncDec) { incl(reg) ; return; }
3166 /* else */ { addl(reg, value) ; return; }
3167 }
3168
3169 void MacroAssembler::incrementl(Address dst, int value) {
3170 if (value == min_jint) {addl(dst, value) ; return; }
3171 if (value < 0) { decrementl(dst, -value); return; }
3172 if (value == 0) { ; return; }
3173 if (value == 1 && UseIncDec) { incl(dst) ; return; }
3174 /* else */ { addl(dst, value) ; return; }
3175 }
3176
3177 void MacroAssembler::jump(AddressLiteral dst) {
3178 if (reachable(dst)) {
3179 jmp_literal(dst.target(), dst.rspec());
3180 } else {
3181 lea(rscratch1, dst);
3182 jmp(rscratch1);
3183 }
3184 }
3185
3186 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) {
3187 if (reachable(dst)) {
3188 InstructionMark im(this);
3189 relocate(dst.reloc());
3190 const int short_size = 2;
3191 const int long_size = 6;
3192 int offs = (intptr_t)dst.target() - ((intptr_t)pc());
3193 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
3194 // 0111 tttn #8-bit disp
3195 emit_int8(0x70 | cc);
3196 emit_int8((offs - short_size) & 0xFF);
3197 } else {
3198 // 0000 1111 1000 tttn #32-bit disp
3199 emit_int8(0x0F);
3200 emit_int8((unsigned char)(0x80 | cc));
3201 emit_int32(offs - long_size);
3202 }
3203 } else {
3204 #ifdef ASSERT
3205 warning("reversing conditional branch");
3206 #endif /* ASSERT */
3207 Label skip;
3208 jccb(reverse[cc], skip);
3209 lea(rscratch1, dst);
3210 Assembler::jmp(rscratch1);
3211 bind(skip);
3212 }
3213 }
3214
3215 void MacroAssembler::ldmxcsr(AddressLiteral src) {
3216 if (reachable(src)) {
3217 Assembler::ldmxcsr(as_Address(src));
3218 } else {
3219 lea(rscratch1, src);
3220 Assembler::ldmxcsr(Address(rscratch1, 0));
3221 }
3222 }
3223
3224 int MacroAssembler::load_signed_byte(Register dst, Address src) {
3225 int off;
3226 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3227 off = offset();
3228 movsbl(dst, src); // movsxb
3229 } else {
3230 off = load_unsigned_byte(dst, src);
3231 shll(dst, 24);
3232 sarl(dst, 24);
3233 }
3234 return off;
3235 }
3236
3237 // Note: load_signed_short used to be called load_signed_word.
3238 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler
3239 // manual, which means 16 bits, that usage is found nowhere in HotSpot code.
3240 // The term "word" in HotSpot means a 32- or 64-bit machine word.
3241 int MacroAssembler::load_signed_short(Register dst, Address src) {
3242 int off;
3243 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3244 // This is dubious to me since it seems safe to do a signed 16 => 64 bit
3245 // version but this is what 64bit has always done. This seems to imply
3246 // that users are only using 32bits worth.
3247 off = offset();
3248 movswl(dst, src); // movsxw
3249 } else {
3250 off = load_unsigned_short(dst, src);
3251 shll(dst, 16);
3252 sarl(dst, 16);
3253 }
3254 return off;
3255 }
3256
3257 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
3258 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3259 // and "3.9 Partial Register Penalties", p. 22).
3260 int off;
3261 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) {
3262 off = offset();
3263 movzbl(dst, src); // movzxb
3264 } else {
3265 xorl(dst, dst);
3266 off = offset();
3267 movb(dst, src);
3268 }
3269 return off;
3270 }
3271
3272 // Note: load_unsigned_short used to be called load_unsigned_word.
3273 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
3274 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3275 // and "3.9 Partial Register Penalties", p. 22).
3276 int off;
3277 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) {
3278 off = offset();
3279 movzwl(dst, src); // movzxw
3280 } else {
3281 xorl(dst, dst);
3282 off = offset();
3283 movw(dst, src);
3284 }
3285 return off;
3286 }
3287
3288 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
3289 switch (size_in_bytes) {
3290 #ifndef _LP64
3291 case 8:
3292 assert(dst2 != noreg, "second dest register required");
3293 movl(dst, src);
3294 movl(dst2, src.plus_disp(BytesPerInt));
3295 break;
3296 #else
3297 case 8: movq(dst, src); break;
3298 #endif
3299 case 4: movl(dst, src); break;
3300 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
3301 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
3302 default: ShouldNotReachHere();
3303 }
3304 }
3305
3306 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
3307 switch (size_in_bytes) {
3308 #ifndef _LP64
3309 case 8:
3310 assert(src2 != noreg, "second source register required");
3311 movl(dst, src);
3312 movl(dst.plus_disp(BytesPerInt), src2);
3313 break;
3314 #else
3315 case 8: movq(dst, src); break;
3316 #endif
3317 case 4: movl(dst, src); break;
3318 case 2: movw(dst, src); break;
3319 case 1: movb(dst, src); break;
3320 default: ShouldNotReachHere();
3321 }
3322 }
3323
3324 void MacroAssembler::mov32(AddressLiteral dst, Register src) {
3325 if (reachable(dst)) {
3326 movl(as_Address(dst), src);
3327 } else {
3328 lea(rscratch1, dst);
3329 movl(Address(rscratch1, 0), src);
3330 }
3331 }
3332
3333 void MacroAssembler::mov32(Register dst, AddressLiteral src) {
3334 if (reachable(src)) {
3335 movl(dst, as_Address(src));
3336 } else {
3337 lea(rscratch1, src);
3338 movl(dst, Address(rscratch1, 0));
3339 }
3340 }
3341
3342 // C++ bool manipulation
3343
3344 void MacroAssembler::movbool(Register dst, Address src) {
3345 if(sizeof(bool) == 1)
3346 movb(dst, src);
3347 else if(sizeof(bool) == 2)
3348 movw(dst, src);
3349 else if(sizeof(bool) == 4)
3350 movl(dst, src);
3351 else
3352 // unsupported
3353 ShouldNotReachHere();
3354 }
3355
3356 void MacroAssembler::movbool(Address dst, bool boolconst) {
3357 if(sizeof(bool) == 1)
3358 movb(dst, (int) boolconst);
3359 else if(sizeof(bool) == 2)
3360 movw(dst, (int) boolconst);
3361 else if(sizeof(bool) == 4)
3362 movl(dst, (int) boolconst);
3363 else
3364 // unsupported
3365 ShouldNotReachHere();
3366 }
3367
3368 void MacroAssembler::movbool(Address dst, Register src) {
3369 if(sizeof(bool) == 1)
3370 movb(dst, src);
3371 else if(sizeof(bool) == 2)
3372 movw(dst, src);
3373 else if(sizeof(bool) == 4)
3374 movl(dst, src);
3375 else
3376 // unsupported
3377 ShouldNotReachHere();
3378 }
3379
3380 void MacroAssembler::movbyte(ArrayAddress dst, int src) {
3381 movb(as_Address(dst), src);
3382 }
3383
3384 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) {
3385 if (reachable(src)) {
3386 movdl(dst, as_Address(src));
3387 } else {
3388 lea(rscratch1, src);
3389 movdl(dst, Address(rscratch1, 0));
3390 }
3391 }
3392
3393 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) {
3394 if (reachable(src)) {
3395 movq(dst, as_Address(src));
3396 } else {
3397 lea(rscratch1, src);
3398 movq(dst, Address(rscratch1, 0));
3399 }
3400 }
3401
3402 void MacroAssembler::setvectmask(Register dst, Register src) {
3403 Assembler::movl(dst, 1);
3404 Assembler::shlxl(dst, dst, src);
3405 Assembler::decl(dst);
3406 Assembler::kmovdl(k1, dst);
3407 Assembler::movl(dst, src);
3408 }
3409
3410 void MacroAssembler::restorevectmask() {
3411 Assembler::knotwl(k1, k0);
3412 }
3413
3414 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) {
3415 if (reachable(src)) {
3416 if (UseXmmLoadAndClearUpper) {
3417 movsd (dst, as_Address(src));
3418 } else {
3419 movlpd(dst, as_Address(src));
3420 }
3421 } else {
3422 lea(rscratch1, src);
3423 if (UseXmmLoadAndClearUpper) {
3424 movsd (dst, Address(rscratch1, 0));
3425 } else {
3426 movlpd(dst, Address(rscratch1, 0));
3427 }
3428 }
3429 }
3430
3431 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) {
3432 if (reachable(src)) {
3433 movss(dst, as_Address(src));
3434 } else {
3435 lea(rscratch1, src);
3436 movss(dst, Address(rscratch1, 0));
3437 }
3438 }
3439
3440 void MacroAssembler::movptr(Register dst, Register src) {
3441 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3442 }
3443
3444 void MacroAssembler::movptr(Register dst, Address src) {
3445 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3446 }
3447
3448 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
3449 void MacroAssembler::movptr(Register dst, intptr_t src) {
3450 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src));
3451 }
3452
3453 void MacroAssembler::movptr(Address dst, Register src) {
3454 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3455 }
3456
3457 void MacroAssembler::movdqu(Address dst, XMMRegister src) {
3458 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) {
3459 Assembler::vextractf32x4(dst, src, 0);
3460 } else {
3461 Assembler::movdqu(dst, src);
3462 }
3463 }
3464
3465 void MacroAssembler::movdqu(XMMRegister dst, Address src) {
3466 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) {
3467 Assembler::vinsertf32x4(dst, dst, src, 0);
3468 } else {
3469 Assembler::movdqu(dst, src);
3470 }
3471 }
3472
3473 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) {
3474 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) {
3475 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit);
3476 } else {
3477 Assembler::movdqu(dst, src);
3478 }
3479 }
3480
3481 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src) {
3482 if (reachable(src)) {
3483 movdqu(dst, as_Address(src));
3484 } else {
3485 lea(rscratch1, src);
3486 movdqu(dst, Address(rscratch1, 0));
3487 }
3488 }
3489
3490 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) {
3491 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) {
3492 vextractf64x4_low(dst, src);
3493 } else {
3494 Assembler::vmovdqu(dst, src);
3495 }
3496 }
3497
3498 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) {
3499 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) {
3500 vinsertf64x4_low(dst, src);
3501 } else {
3502 Assembler::vmovdqu(dst, src);
3503 }
3504 }
3505
3506 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) {
3507 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) {
3508 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit);
3509 }
3510 else {
3511 Assembler::vmovdqu(dst, src);
3512 }
3513 }
3514
3515 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) {
3516 if (reachable(src)) {
3517 vmovdqu(dst, as_Address(src));
3518 }
3519 else {
3520 lea(rscratch1, src);
3521 vmovdqu(dst, Address(rscratch1, 0));
3522 }
3523 }
3524
3525 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) {
3526 if (reachable(src)) {
3527 Assembler::movdqa(dst, as_Address(src));
3528 } else {
3529 lea(rscratch1, src);
3530 Assembler::movdqa(dst, Address(rscratch1, 0));
3531 }
3532 }
3533
3534 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) {
3535 if (reachable(src)) {
3536 Assembler::movsd(dst, as_Address(src));
3537 } else {
3538 lea(rscratch1, src);
3539 Assembler::movsd(dst, Address(rscratch1, 0));
3540 }
3541 }
3542
3543 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) {
3544 if (reachable(src)) {
3545 Assembler::movss(dst, as_Address(src));
3546 } else {
3547 lea(rscratch1, src);
3548 Assembler::movss(dst, Address(rscratch1, 0));
3549 }
3550 }
3551
3552 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) {
3553 if (reachable(src)) {
3554 Assembler::mulsd(dst, as_Address(src));
3555 } else {
3556 lea(rscratch1, src);
3557 Assembler::mulsd(dst, Address(rscratch1, 0));
3558 }
3559 }
3560
3561 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) {
3562 if (reachable(src)) {
3563 Assembler::mulss(dst, as_Address(src));
3564 } else {
3565 lea(rscratch1, src);
3566 Assembler::mulss(dst, Address(rscratch1, 0));
3567 }
3568 }
3569
3570 void MacroAssembler::null_check(Register reg, int offset) {
3571 if (needs_explicit_null_check(offset)) {
3572 // provoke OS NULL exception if reg = NULL by
3573 // accessing M[reg] w/o changing any (non-CC) registers
3574 // NOTE: cmpl is plenty here to provoke a segv
3575 cmpptr(rax, Address(reg, 0));
3576 // Note: should probably use testl(rax, Address(reg, 0));
3577 // may be shorter code (however, this version of
3578 // testl needs to be implemented first)
3579 } else {
3580 // nothing to do, (later) access of M[reg + offset]
3581 // will provoke OS NULL exception if reg = NULL
3582 }
3583 }
3584
3585 void MacroAssembler::os_breakpoint() {
3586 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability
3587 // (e.g., MSVC can't call ps() otherwise)
3588 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint)));
3589 }
3590
3591 #ifdef _LP64
3592 #define XSTATE_BV 0x200
3593 #endif
3594
3595 void MacroAssembler::pop_CPU_state() {
3596 pop_FPU_state();
3597 pop_IU_state();
3598 }
3599
3600 void MacroAssembler::pop_FPU_state() {
3601 #ifndef _LP64
3602 frstor(Address(rsp, 0));
3603 #else
3604 fxrstor(Address(rsp, 0));
3605 #endif
3606 addptr(rsp, FPUStateSizeInWords * wordSize);
3607 }
3608
3609 void MacroAssembler::pop_IU_state() {
3610 popa();
3611 LP64_ONLY(addq(rsp, 8));
3612 popf();
3613 }
3614
3615 // Save Integer and Float state
3616 // Warning: Stack must be 16 byte aligned (64bit)
3617 void MacroAssembler::push_CPU_state() {
3618 push_IU_state();
3619 push_FPU_state();
3620 }
3621
3622 void MacroAssembler::push_FPU_state() {
3623 subptr(rsp, FPUStateSizeInWords * wordSize);
3624 #ifndef _LP64
3625 fnsave(Address(rsp, 0));
3626 fwait();
3627 #else
3628 fxsave(Address(rsp, 0));
3629 #endif // LP64
3630 }
3631
3632 void MacroAssembler::push_IU_state() {
3633 // Push flags first because pusha kills them
3634 pushf();
3635 // Make sure rsp stays 16-byte aligned
3636 LP64_ONLY(subq(rsp, 8));
3637 pusha();
3638 }
3639
3640 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp, bool clear_pc) {
3641 // determine java_thread register
3642 if (!java_thread->is_valid()) {
3643 java_thread = rdi;
3644 get_thread(java_thread);
3645 }
3646 // we must set sp to zero to clear frame
3647 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
3648 if (clear_fp) {
3649 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3650 }
3651
3652 if (clear_pc)
3653 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
3654
3655 }
3656
3657 void MacroAssembler::restore_rax(Register tmp) {
3658 if (tmp == noreg) pop(rax);
3659 else if (tmp != rax) mov(rax, tmp);
3660 }
3661
3662 void MacroAssembler::round_to(Register reg, int modulus) {
3663 addptr(reg, modulus - 1);
3664 andptr(reg, -modulus);
3665 }
3666
3667 void MacroAssembler::save_rax(Register tmp) {
3668 if (tmp == noreg) push(rax);
3669 else if (tmp != rax) mov(tmp, rax);
3670 }
3671
3672 // Write serialization page so VM thread can do a pseudo remote membar.
3673 // We use the current thread pointer to calculate a thread specific
3674 // offset to write to within the page. This minimizes bus traffic
3675 // due to cache line collision.
3676 void MacroAssembler::serialize_memory(Register thread, Register tmp) {
3677 movl(tmp, thread);
3678 shrl(tmp, os::get_serialize_page_shift_count());
3679 andl(tmp, (os::vm_page_size() - sizeof(int)));
3680
3681 Address index(noreg, tmp, Address::times_1);
3682 ExternalAddress page(os::get_memory_serialize_page());
3683
3684 // Size of store must match masking code above
3685 movl(as_Address(ArrayAddress(page, index)), tmp);
3686 }
3687
3688 // Calls to C land
3689 //
3690 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded
3691 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
3692 // has to be reset to 0. This is required to allow proper stack traversal.
3693 void MacroAssembler::set_last_Java_frame(Register java_thread,
3694 Register last_java_sp,
3695 Register last_java_fp,
3696 address last_java_pc) {
3697 // determine java_thread register
3698 if (!java_thread->is_valid()) {
3699 java_thread = rdi;
3700 get_thread(java_thread);
3701 }
3702 // determine last_java_sp register
3703 if (!last_java_sp->is_valid()) {
3704 last_java_sp = rsp;
3705 }
3706
3707 // last_java_fp is optional
3708
3709 if (last_java_fp->is_valid()) {
3710 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp);
3711 }
3712
3713 // last_java_pc is optional
3714
3715 if (last_java_pc != NULL) {
3716 lea(Address(java_thread,
3717 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()),
3718 InternalAddress(last_java_pc));
3719
3720 }
3721 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
3722 }
3723
3724 void MacroAssembler::shlptr(Register dst, int imm8) {
3725 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8));
3726 }
3727
3728 void MacroAssembler::shrptr(Register dst, int imm8) {
3729 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8));
3730 }
3731
3732 void MacroAssembler::sign_extend_byte(Register reg) {
3733 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) {
3734 movsbl(reg, reg); // movsxb
3735 } else {
3736 shll(reg, 24);
3737 sarl(reg, 24);
3738 }
3739 }
3740
3741 void MacroAssembler::sign_extend_short(Register reg) {
3742 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3743 movswl(reg, reg); // movsxw
3744 } else {
3745 shll(reg, 16);
3746 sarl(reg, 16);
3747 }
3748 }
3749
3750 void MacroAssembler::testl(Register dst, AddressLiteral src) {
3751 assert(reachable(src), "Address should be reachable");
3752 testl(dst, as_Address(src));
3753 }
3754
3755 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) {
3756 int dst_enc = dst->encoding();
3757 int src_enc = src->encoding();
3758 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3759 Assembler::pcmpeqb(dst, src);
3760 } else if ((dst_enc < 16) && (src_enc < 16)) {
3761 Assembler::pcmpeqb(dst, src);
3762 } else if (src_enc < 16) {
3763 subptr(rsp, 64);
3764 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3765 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3766 Assembler::pcmpeqb(xmm0, src);
3767 movdqu(dst, xmm0);
3768 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3769 addptr(rsp, 64);
3770 } else if (dst_enc < 16) {
3771 subptr(rsp, 64);
3772 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3773 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3774 Assembler::pcmpeqb(dst, xmm0);
3775 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3776 addptr(rsp, 64);
3777 } else {
3778 subptr(rsp, 64);
3779 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3780 subptr(rsp, 64);
3781 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3782 movdqu(xmm0, src);
3783 movdqu(xmm1, dst);
3784 Assembler::pcmpeqb(xmm1, xmm0);
3785 movdqu(dst, xmm1);
3786 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3787 addptr(rsp, 64);
3788 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3789 addptr(rsp, 64);
3790 }
3791 }
3792
3793 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) {
3794 int dst_enc = dst->encoding();
3795 int src_enc = src->encoding();
3796 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3797 Assembler::pcmpeqw(dst, src);
3798 } else if ((dst_enc < 16) && (src_enc < 16)) {
3799 Assembler::pcmpeqw(dst, src);
3800 } else if (src_enc < 16) {
3801 subptr(rsp, 64);
3802 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3803 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3804 Assembler::pcmpeqw(xmm0, src);
3805 movdqu(dst, xmm0);
3806 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3807 addptr(rsp, 64);
3808 } else if (dst_enc < 16) {
3809 subptr(rsp, 64);
3810 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3811 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3812 Assembler::pcmpeqw(dst, xmm0);
3813 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3814 addptr(rsp, 64);
3815 } else {
3816 subptr(rsp, 64);
3817 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3818 subptr(rsp, 64);
3819 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3820 movdqu(xmm0, src);
3821 movdqu(xmm1, dst);
3822 Assembler::pcmpeqw(xmm1, xmm0);
3823 movdqu(dst, xmm1);
3824 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3825 addptr(rsp, 64);
3826 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3827 addptr(rsp, 64);
3828 }
3829 }
3830
3831 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) {
3832 int dst_enc = dst->encoding();
3833 if (dst_enc < 16) {
3834 Assembler::pcmpestri(dst, src, imm8);
3835 } else {
3836 subptr(rsp, 64);
3837 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3838 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3839 Assembler::pcmpestri(xmm0, src, imm8);
3840 movdqu(dst, xmm0);
3841 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3842 addptr(rsp, 64);
3843 }
3844 }
3845
3846 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) {
3847 int dst_enc = dst->encoding();
3848 int src_enc = src->encoding();
3849 if ((dst_enc < 16) && (src_enc < 16)) {
3850 Assembler::pcmpestri(dst, src, imm8);
3851 } else if (src_enc < 16) {
3852 subptr(rsp, 64);
3853 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3854 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3855 Assembler::pcmpestri(xmm0, src, imm8);
3856 movdqu(dst, xmm0);
3857 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3858 addptr(rsp, 64);
3859 } else if (dst_enc < 16) {
3860 subptr(rsp, 64);
3861 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3862 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3863 Assembler::pcmpestri(dst, xmm0, imm8);
3864 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3865 addptr(rsp, 64);
3866 } else {
3867 subptr(rsp, 64);
3868 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3869 subptr(rsp, 64);
3870 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3871 movdqu(xmm0, src);
3872 movdqu(xmm1, dst);
3873 Assembler::pcmpestri(xmm1, xmm0, imm8);
3874 movdqu(dst, xmm1);
3875 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3876 addptr(rsp, 64);
3877 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3878 addptr(rsp, 64);
3879 }
3880 }
3881
3882 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) {
3883 int dst_enc = dst->encoding();
3884 int src_enc = src->encoding();
3885 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3886 Assembler::pmovzxbw(dst, src);
3887 } else if ((dst_enc < 16) && (src_enc < 16)) {
3888 Assembler::pmovzxbw(dst, src);
3889 } else if (src_enc < 16) {
3890 subptr(rsp, 64);
3891 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3892 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3893 Assembler::pmovzxbw(xmm0, src);
3894 movdqu(dst, xmm0);
3895 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3896 addptr(rsp, 64);
3897 } else if (dst_enc < 16) {
3898 subptr(rsp, 64);
3899 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3900 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3901 Assembler::pmovzxbw(dst, xmm0);
3902 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3903 addptr(rsp, 64);
3904 } else {
3905 subptr(rsp, 64);
3906 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3907 subptr(rsp, 64);
3908 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3909 movdqu(xmm0, src);
3910 movdqu(xmm1, dst);
3911 Assembler::pmovzxbw(xmm1, xmm0);
3912 movdqu(dst, xmm1);
3913 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3914 addptr(rsp, 64);
3915 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3916 addptr(rsp, 64);
3917 }
3918 }
3919
3920 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) {
3921 int dst_enc = dst->encoding();
3922 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3923 Assembler::pmovzxbw(dst, src);
3924 } else if (dst_enc < 16) {
3925 Assembler::pmovzxbw(dst, src);
3926 } else {
3927 subptr(rsp, 64);
3928 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3929 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3930 Assembler::pmovzxbw(xmm0, src);
3931 movdqu(dst, xmm0);
3932 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3933 addptr(rsp, 64);
3934 }
3935 }
3936
3937 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) {
3938 int src_enc = src->encoding();
3939 if (src_enc < 16) {
3940 Assembler::pmovmskb(dst, src);
3941 } else {
3942 subptr(rsp, 64);
3943 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3944 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3945 Assembler::pmovmskb(dst, xmm0);
3946 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3947 addptr(rsp, 64);
3948 }
3949 }
3950
3951 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) {
3952 int dst_enc = dst->encoding();
3953 int src_enc = src->encoding();
3954 if ((dst_enc < 16) && (src_enc < 16)) {
3955 Assembler::ptest(dst, src);
3956 } else if (src_enc < 16) {
3957 subptr(rsp, 64);
3958 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3959 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3960 Assembler::ptest(xmm0, src);
3961 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3962 addptr(rsp, 64);
3963 } else if (dst_enc < 16) {
3964 subptr(rsp, 64);
3965 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3966 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3967 Assembler::ptest(dst, xmm0);
3968 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3969 addptr(rsp, 64);
3970 } else {
3971 subptr(rsp, 64);
3972 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3973 subptr(rsp, 64);
3974 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3975 movdqu(xmm0, src);
3976 movdqu(xmm1, dst);
3977 Assembler::ptest(xmm1, xmm0);
3978 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3979 addptr(rsp, 64);
3980 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3981 addptr(rsp, 64);
3982 }
3983 }
3984
3985 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) {
3986 if (reachable(src)) {
3987 Assembler::sqrtsd(dst, as_Address(src));
3988 } else {
3989 lea(rscratch1, src);
3990 Assembler::sqrtsd(dst, Address(rscratch1, 0));
3991 }
3992 }
3993
3994 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) {
3995 if (reachable(src)) {
3996 Assembler::sqrtss(dst, as_Address(src));
3997 } else {
3998 lea(rscratch1, src);
3999 Assembler::sqrtss(dst, Address(rscratch1, 0));
4000 }
4001 }
4002
4003 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) {
4004 if (reachable(src)) {
4005 Assembler::subsd(dst, as_Address(src));
4006 } else {
4007 lea(rscratch1, src);
4008 Assembler::subsd(dst, Address(rscratch1, 0));
4009 }
4010 }
4011
4012 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
4013 if (reachable(src)) {
4014 Assembler::subss(dst, as_Address(src));
4015 } else {
4016 lea(rscratch1, src);
4017 Assembler::subss(dst, Address(rscratch1, 0));
4018 }
4019 }
4020
4021 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
4022 if (reachable(src)) {
4023 Assembler::ucomisd(dst, as_Address(src));
4024 } else {
4025 lea(rscratch1, src);
4026 Assembler::ucomisd(dst, Address(rscratch1, 0));
4027 }
4028 }
4029
4030 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
4031 if (reachable(src)) {
4032 Assembler::ucomiss(dst, as_Address(src));
4033 } else {
4034 lea(rscratch1, src);
4035 Assembler::ucomiss(dst, Address(rscratch1, 0));
4036 }
4037 }
4038
4039 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) {
4040 // Used in sign-bit flipping with aligned address.
4041 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
4042 if (reachable(src)) {
4043 Assembler::xorpd(dst, as_Address(src));
4044 } else {
4045 lea(rscratch1, src);
4046 Assembler::xorpd(dst, Address(rscratch1, 0));
4047 }
4048 }
4049
4050 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) {
4051 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
4052 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
4053 }
4054 else {
4055 Assembler::xorpd(dst, src);
4056 }
4057 }
4058
4059 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) {
4060 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
4061 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
4062 } else {
4063 Assembler::xorps(dst, src);
4064 }
4065 }
4066
4067 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) {
4068 // Used in sign-bit flipping with aligned address.
4069 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
4070 if (reachable(src)) {
4071 Assembler::xorps(dst, as_Address(src));
4072 } else {
4073 lea(rscratch1, src);
4074 Assembler::xorps(dst, Address(rscratch1, 0));
4075 }
4076 }
4077
4078 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
4079 // Used in sign-bit flipping with aligned address.
4080 bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
4081 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
4082 if (reachable(src)) {
4083 Assembler::pshufb(dst, as_Address(src));
4084 } else {
4085 lea(rscratch1, src);
4086 Assembler::pshufb(dst, Address(rscratch1, 0));
4087 }
4088 }
4089
4090 // AVX 3-operands instructions
4091
4092 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4093 if (reachable(src)) {
4094 vaddsd(dst, nds, as_Address(src));
4095 } else {
4096 lea(rscratch1, src);
4097 vaddsd(dst, nds, Address(rscratch1, 0));
4098 }
4099 }
4100
4101 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4102 if (reachable(src)) {
4103 vaddss(dst, nds, as_Address(src));
4104 } else {
4105 lea(rscratch1, src);
4106 vaddss(dst, nds, Address(rscratch1, 0));
4107 }
4108 }
4109
4110 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
4111 int dst_enc = dst->encoding();
4112 int nds_enc = nds->encoding();
4113 int src_enc = src->encoding();
4114 if ((dst_enc < 16) && (nds_enc < 16)) {
4115 vandps(dst, nds, negate_field, vector_len);
4116 } else if ((src_enc < 16) && (dst_enc < 16)) {
4117 movss(src, nds);
4118 vandps(dst, src, negate_field, vector_len);
4119 } else if (src_enc < 16) {
4120 movss(src, nds);
4121 vandps(src, src, negate_field, vector_len);
4122 movss(dst, src);
4123 } else if (dst_enc < 16) {
4124 movdqu(src, xmm0);
4125 movss(xmm0, nds);
4126 vandps(dst, xmm0, negate_field, vector_len);
4127 movdqu(xmm0, src);
4128 } else if (nds_enc < 16) {
4129 movdqu(src, xmm0);
4130 vandps(xmm0, nds, negate_field, vector_len);
4131 movss(dst, xmm0);
4132 movdqu(xmm0, src);
4133 } else {
4134 movdqu(src, xmm0);
4135 movss(xmm0, nds);
4136 vandps(xmm0, xmm0, negate_field, vector_len);
4137 movss(dst, xmm0);
4138 movdqu(xmm0, src);
4139 }
4140 }
4141
4142 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
4143 int dst_enc = dst->encoding();
4144 int nds_enc = nds->encoding();
4145 int src_enc = src->encoding();
4146 if ((dst_enc < 16) && (nds_enc < 16)) {
4147 vandpd(dst, nds, negate_field, vector_len);
4148 } else if ((src_enc < 16) && (dst_enc < 16)) {
4149 movsd(src, nds);
4150 vandpd(dst, src, negate_field, vector_len);
4151 } else if (src_enc < 16) {
4152 movsd(src, nds);
4153 vandpd(src, src, negate_field, vector_len);
4154 movsd(dst, src);
4155 } else if (dst_enc < 16) {
4156 movdqu(src, xmm0);
4157 movsd(xmm0, nds);
4158 vandpd(dst, xmm0, negate_field, vector_len);
4159 movdqu(xmm0, src);
4160 } else if (nds_enc < 16) {
4161 movdqu(src, xmm0);
4162 vandpd(xmm0, nds, negate_field, vector_len);
4163 movsd(dst, xmm0);
4164 movdqu(xmm0, src);
4165 } else {
4166 movdqu(src, xmm0);
4167 movsd(xmm0, nds);
4168 vandpd(xmm0, xmm0, negate_field, vector_len);
4169 movsd(dst, xmm0);
4170 movdqu(xmm0, src);
4171 }
4172 }
4173
4174 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4175 int dst_enc = dst->encoding();
4176 int nds_enc = nds->encoding();
4177 int src_enc = src->encoding();
4178 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4179 Assembler::vpaddb(dst, nds, src, vector_len);
4180 } else if ((dst_enc < 16) && (src_enc < 16)) {
4181 Assembler::vpaddb(dst, dst, src, vector_len);
4182 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4183 // use nds as scratch for src
4184 evmovdqul(nds, src, Assembler::AVX_512bit);
4185 Assembler::vpaddb(dst, dst, nds, vector_len);
4186 } else if ((src_enc < 16) && (nds_enc < 16)) {
4187 // use nds as scratch for dst
4188 evmovdqul(nds, dst, Assembler::AVX_512bit);
4189 Assembler::vpaddb(nds, nds, src, vector_len);
4190 evmovdqul(dst, nds, Assembler::AVX_512bit);
4191 } else if (dst_enc < 16) {
4192 // use nds as scatch for xmm0 to hold src
4193 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4194 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4195 Assembler::vpaddb(dst, dst, xmm0, vector_len);
4196 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4197 } else {
4198 // worse case scenario, all regs are in the upper bank
4199 subptr(rsp, 64);
4200 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4201 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4202 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4203 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4204 Assembler::vpaddb(xmm0, xmm0, xmm1, vector_len);
4205 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4206 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4207 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4208 addptr(rsp, 64);
4209 }
4210 }
4211
4212 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4213 int dst_enc = dst->encoding();
4214 int nds_enc = nds->encoding();
4215 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4216 Assembler::vpaddb(dst, nds, src, vector_len);
4217 } else if (dst_enc < 16) {
4218 Assembler::vpaddb(dst, dst, src, vector_len);
4219 } else if (nds_enc < 16) {
4220 // implies dst_enc in upper bank with src as scratch
4221 evmovdqul(nds, dst, Assembler::AVX_512bit);
4222 Assembler::vpaddb(nds, nds, src, vector_len);
4223 evmovdqul(dst, nds, Assembler::AVX_512bit);
4224 } else {
4225 // worse case scenario, all regs in upper bank
4226 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4227 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4228 Assembler::vpaddb(xmm0, xmm0, src, vector_len);
4229 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4230 }
4231 }
4232
4233 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4234 int dst_enc = dst->encoding();
4235 int nds_enc = nds->encoding();
4236 int src_enc = src->encoding();
4237 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4238 Assembler::vpaddw(dst, nds, src, vector_len);
4239 } else if ((dst_enc < 16) && (src_enc < 16)) {
4240 Assembler::vpaddw(dst, dst, src, vector_len);
4241 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4242 // use nds as scratch for src
4243 evmovdqul(nds, src, Assembler::AVX_512bit);
4244 Assembler::vpaddw(dst, dst, nds, vector_len);
4245 } else if ((src_enc < 16) && (nds_enc < 16)) {
4246 // use nds as scratch for dst
4247 evmovdqul(nds, dst, Assembler::AVX_512bit);
4248 Assembler::vpaddw(nds, nds, src, vector_len);
4249 evmovdqul(dst, nds, Assembler::AVX_512bit);
4250 } else if (dst_enc < 16) {
4251 // use nds as scatch for xmm0 to hold src
4252 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4253 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4254 Assembler::vpaddw(dst, dst, xmm0, vector_len);
4255 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4256 } else {
4257 // worse case scenario, all regs are in the upper bank
4258 subptr(rsp, 64);
4259 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4260 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4261 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4262 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4263 Assembler::vpaddw(xmm0, xmm0, xmm1, vector_len);
4264 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4265 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4266 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4267 addptr(rsp, 64);
4268 }
4269 }
4270
4271 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4272 int dst_enc = dst->encoding();
4273 int nds_enc = nds->encoding();
4274 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4275 Assembler::vpaddw(dst, nds, src, vector_len);
4276 } else if (dst_enc < 16) {
4277 Assembler::vpaddw(dst, dst, src, vector_len);
4278 } else if (nds_enc < 16) {
4279 // implies dst_enc in upper bank with src as scratch
4280 evmovdqul(nds, dst, Assembler::AVX_512bit);
4281 Assembler::vpaddw(nds, nds, src, vector_len);
4282 evmovdqul(dst, nds, Assembler::AVX_512bit);
4283 } else {
4284 // worse case scenario, all regs in upper bank
4285 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4286 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4287 Assembler::vpaddw(xmm0, xmm0, src, vector_len);
4288 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4289 }
4290 }
4291
4292 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src) {
4293 int dst_enc = dst->encoding();
4294 int src_enc = src->encoding();
4295 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4296 Assembler::vpbroadcastw(dst, src);
4297 } else if ((dst_enc < 16) && (src_enc < 16)) {
4298 Assembler::vpbroadcastw(dst, src);
4299 } else if (src_enc < 16) {
4300 subptr(rsp, 64);
4301 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4302 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4303 Assembler::vpbroadcastw(xmm0, src);
4304 movdqu(dst, xmm0);
4305 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4306 addptr(rsp, 64);
4307 } else if (dst_enc < 16) {
4308 subptr(rsp, 64);
4309 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4310 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4311 Assembler::vpbroadcastw(dst, xmm0);
4312 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4313 addptr(rsp, 64);
4314 } else {
4315 subptr(rsp, 64);
4316 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4317 subptr(rsp, 64);
4318 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4319 movdqu(xmm0, src);
4320 movdqu(xmm1, dst);
4321 Assembler::vpbroadcastw(xmm1, xmm0);
4322 movdqu(dst, xmm1);
4323 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4324 addptr(rsp, 64);
4325 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4326 addptr(rsp, 64);
4327 }
4328 }
4329
4330 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4331 int dst_enc = dst->encoding();
4332 int nds_enc = nds->encoding();
4333 int src_enc = src->encoding();
4334 assert(dst_enc == nds_enc, "");
4335 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4336 Assembler::vpcmpeqb(dst, nds, src, vector_len);
4337 } else if ((dst_enc < 16) && (src_enc < 16)) {
4338 Assembler::vpcmpeqb(dst, nds, src, vector_len);
4339 } else if (src_enc < 16) {
4340 subptr(rsp, 64);
4341 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4342 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4343 Assembler::vpcmpeqb(xmm0, xmm0, src, vector_len);
4344 movdqu(dst, xmm0);
4345 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4346 addptr(rsp, 64);
4347 } else if (dst_enc < 16) {
4348 subptr(rsp, 64);
4349 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4350 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4351 Assembler::vpcmpeqb(dst, dst, xmm0, vector_len);
4352 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4353 addptr(rsp, 64);
4354 } else {
4355 subptr(rsp, 64);
4356 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4357 subptr(rsp, 64);
4358 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4359 movdqu(xmm0, src);
4360 movdqu(xmm1, dst);
4361 Assembler::vpcmpeqb(xmm1, xmm1, xmm0, vector_len);
4362 movdqu(dst, xmm1);
4363 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4364 addptr(rsp, 64);
4365 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4366 addptr(rsp, 64);
4367 }
4368 }
4369
4370 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4371 int dst_enc = dst->encoding();
4372 int nds_enc = nds->encoding();
4373 int src_enc = src->encoding();
4374 assert(dst_enc == nds_enc, "");
4375 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4376 Assembler::vpcmpeqw(dst, nds, src, vector_len);
4377 } else if ((dst_enc < 16) && (src_enc < 16)) {
4378 Assembler::vpcmpeqw(dst, nds, src, vector_len);
4379 } else if (src_enc < 16) {
4380 subptr(rsp, 64);
4381 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4382 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4383 Assembler::vpcmpeqw(xmm0, xmm0, src, vector_len);
4384 movdqu(dst, xmm0);
4385 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4386 addptr(rsp, 64);
4387 } else if (dst_enc < 16) {
4388 subptr(rsp, 64);
4389 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4390 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4391 Assembler::vpcmpeqw(dst, dst, xmm0, vector_len);
4392 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4393 addptr(rsp, 64);
4394 } else {
4395 subptr(rsp, 64);
4396 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4397 subptr(rsp, 64);
4398 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4399 movdqu(xmm0, src);
4400 movdqu(xmm1, dst);
4401 Assembler::vpcmpeqw(xmm1, xmm1, xmm0, vector_len);
4402 movdqu(dst, xmm1);
4403 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4404 addptr(rsp, 64);
4405 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4406 addptr(rsp, 64);
4407 }
4408 }
4409
4410 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) {
4411 int dst_enc = dst->encoding();
4412 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4413 Assembler::vpmovzxbw(dst, src, vector_len);
4414 } else if (dst_enc < 16) {
4415 Assembler::vpmovzxbw(dst, src, vector_len);
4416 } else {
4417 subptr(rsp, 64);
4418 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4419 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4420 Assembler::vpmovzxbw(xmm0, src, vector_len);
4421 movdqu(dst, xmm0);
4422 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4423 addptr(rsp, 64);
4424 }
4425 }
4426
4427 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) {
4428 int src_enc = src->encoding();
4429 if (src_enc < 16) {
4430 Assembler::vpmovmskb(dst, src);
4431 } else {
4432 subptr(rsp, 64);
4433 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4434 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4435 Assembler::vpmovmskb(dst, xmm0);
4436 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4437 addptr(rsp, 64);
4438 }
4439 }
4440
4441 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4442 int dst_enc = dst->encoding();
4443 int nds_enc = nds->encoding();
4444 int src_enc = src->encoding();
4445 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4446 Assembler::vpmullw(dst, nds, src, vector_len);
4447 } else if ((dst_enc < 16) && (src_enc < 16)) {
4448 Assembler::vpmullw(dst, dst, src, vector_len);
4449 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4450 // use nds as scratch for src
4451 evmovdqul(nds, src, Assembler::AVX_512bit);
4452 Assembler::vpmullw(dst, dst, nds, vector_len);
4453 } else if ((src_enc < 16) && (nds_enc < 16)) {
4454 // use nds as scratch for dst
4455 evmovdqul(nds, dst, Assembler::AVX_512bit);
4456 Assembler::vpmullw(nds, nds, src, vector_len);
4457 evmovdqul(dst, nds, Assembler::AVX_512bit);
4458 } else if (dst_enc < 16) {
4459 // use nds as scatch for xmm0 to hold src
4460 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4461 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4462 Assembler::vpmullw(dst, dst, xmm0, vector_len);
4463 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4464 } else {
4465 // worse case scenario, all regs are in the upper bank
4466 subptr(rsp, 64);
4467 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4468 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4469 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4470 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4471 Assembler::vpmullw(xmm0, xmm0, xmm1, vector_len);
4472 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4473 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4474 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4475 addptr(rsp, 64);
4476 }
4477 }
4478
4479 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4480 int dst_enc = dst->encoding();
4481 int nds_enc = nds->encoding();
4482 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4483 Assembler::vpmullw(dst, nds, src, vector_len);
4484 } else if (dst_enc < 16) {
4485 Assembler::vpmullw(dst, dst, src, vector_len);
4486 } else if (nds_enc < 16) {
4487 // implies dst_enc in upper bank with src as scratch
4488 evmovdqul(nds, dst, Assembler::AVX_512bit);
4489 Assembler::vpmullw(nds, nds, src, vector_len);
4490 evmovdqul(dst, nds, Assembler::AVX_512bit);
4491 } else {
4492 // worse case scenario, all regs in upper bank
4493 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4494 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4495 Assembler::vpmullw(xmm0, xmm0, src, vector_len);
4496 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4497 }
4498 }
4499
4500 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4501 int dst_enc = dst->encoding();
4502 int nds_enc = nds->encoding();
4503 int src_enc = src->encoding();
4504 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4505 Assembler::vpsubb(dst, nds, src, vector_len);
4506 } else if ((dst_enc < 16) && (src_enc < 16)) {
4507 Assembler::vpsubb(dst, dst, src, vector_len);
4508 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4509 // use nds as scratch for src
4510 evmovdqul(nds, src, Assembler::AVX_512bit);
4511 Assembler::vpsubb(dst, dst, nds, vector_len);
4512 } else if ((src_enc < 16) && (nds_enc < 16)) {
4513 // use nds as scratch for dst
4514 evmovdqul(nds, dst, Assembler::AVX_512bit);
4515 Assembler::vpsubb(nds, nds, src, vector_len);
4516 evmovdqul(dst, nds, Assembler::AVX_512bit);
4517 } else if (dst_enc < 16) {
4518 // use nds as scatch for xmm0 to hold src
4519 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4520 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4521 Assembler::vpsubb(dst, dst, xmm0, vector_len);
4522 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4523 } else {
4524 // worse case scenario, all regs are in the upper bank
4525 subptr(rsp, 64);
4526 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4527 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4528 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4529 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4530 Assembler::vpsubb(xmm0, xmm0, xmm1, vector_len);
4531 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4532 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4533 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4534 addptr(rsp, 64);
4535 }
4536 }
4537
4538 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4539 int dst_enc = dst->encoding();
4540 int nds_enc = nds->encoding();
4541 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4542 Assembler::vpsubb(dst, nds, src, vector_len);
4543 } else if (dst_enc < 16) {
4544 Assembler::vpsubb(dst, dst, src, vector_len);
4545 } else if (nds_enc < 16) {
4546 // implies dst_enc in upper bank with src as scratch
4547 evmovdqul(nds, dst, Assembler::AVX_512bit);
4548 Assembler::vpsubb(nds, nds, src, vector_len);
4549 evmovdqul(dst, nds, Assembler::AVX_512bit);
4550 } else {
4551 // worse case scenario, all regs in upper bank
4552 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4553 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4554 Assembler::vpsubw(xmm0, xmm0, src, vector_len);
4555 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4556 }
4557 }
4558
4559 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4560 int dst_enc = dst->encoding();
4561 int nds_enc = nds->encoding();
4562 int src_enc = src->encoding();
4563 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4564 Assembler::vpsubw(dst, nds, src, vector_len);
4565 } else if ((dst_enc < 16) && (src_enc < 16)) {
4566 Assembler::vpsubw(dst, dst, src, vector_len);
4567 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4568 // use nds as scratch for src
4569 evmovdqul(nds, src, Assembler::AVX_512bit);
4570 Assembler::vpsubw(dst, dst, nds, vector_len);
4571 } else if ((src_enc < 16) && (nds_enc < 16)) {
4572 // use nds as scratch for dst
4573 evmovdqul(nds, dst, Assembler::AVX_512bit);
4574 Assembler::vpsubw(nds, nds, src, vector_len);
4575 evmovdqul(dst, nds, Assembler::AVX_512bit);
4576 } else if (dst_enc < 16) {
4577 // use nds as scatch for xmm0 to hold src
4578 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4579 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4580 Assembler::vpsubw(dst, dst, xmm0, vector_len);
4581 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4582 } else {
4583 // worse case scenario, all regs are in the upper bank
4584 subptr(rsp, 64);
4585 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4586 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4587 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4588 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4589 Assembler::vpsubw(xmm0, xmm0, xmm1, vector_len);
4590 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4591 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4592 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4593 addptr(rsp, 64);
4594 }
4595 }
4596
4597 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4598 int dst_enc = dst->encoding();
4599 int nds_enc = nds->encoding();
4600 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4601 Assembler::vpsubw(dst, nds, src, vector_len);
4602 } else if (dst_enc < 16) {
4603 Assembler::vpsubw(dst, dst, src, vector_len);
4604 } else if (nds_enc < 16) {
4605 // implies dst_enc in upper bank with src as scratch
4606 evmovdqul(nds, dst, Assembler::AVX_512bit);
4607 Assembler::vpsubw(nds, nds, src, vector_len);
4608 evmovdqul(dst, nds, Assembler::AVX_512bit);
4609 } else {
4610 // worse case scenario, all regs in upper bank
4611 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4612 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4613 Assembler::vpsubw(xmm0, xmm0, src, vector_len);
4614 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4615 }
4616 }
4617
4618 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4619 int dst_enc = dst->encoding();
4620 int nds_enc = nds->encoding();
4621 int shift_enc = shift->encoding();
4622 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4623 Assembler::vpsraw(dst, nds, shift, vector_len);
4624 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4625 Assembler::vpsraw(dst, dst, shift, vector_len);
4626 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4627 // use nds_enc as scratch with shift
4628 evmovdqul(nds, shift, Assembler::AVX_512bit);
4629 Assembler::vpsraw(dst, dst, nds, vector_len);
4630 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4631 // use nds as scratch with dst
4632 evmovdqul(nds, dst, Assembler::AVX_512bit);
4633 Assembler::vpsraw(nds, nds, shift, vector_len);
4634 evmovdqul(dst, nds, Assembler::AVX_512bit);
4635 } else if (dst_enc < 16) {
4636 // use nds to save a copy of xmm0 and hold shift
4637 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4638 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4639 Assembler::vpsraw(dst, dst, xmm0, vector_len);
4640 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4641 } else if (nds_enc < 16) {
4642 // use nds as dest as temps
4643 evmovdqul(nds, dst, Assembler::AVX_512bit);
4644 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4645 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4646 Assembler::vpsraw(nds, nds, xmm0, vector_len);
4647 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4648 evmovdqul(dst, nds, Assembler::AVX_512bit);
4649 } else {
4650 // worse case scenario, all regs are in the upper bank
4651 subptr(rsp, 64);
4652 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4653 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4654 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4655 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4656 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4657 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4658 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4659 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4660 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4661 addptr(rsp, 64);
4662 }
4663 }
4664
4665 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4666 int dst_enc = dst->encoding();
4667 int nds_enc = nds->encoding();
4668 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4669 Assembler::vpsraw(dst, nds, shift, vector_len);
4670 } else if (dst_enc < 16) {
4671 Assembler::vpsraw(dst, dst, shift, vector_len);
4672 } else if (nds_enc < 16) {
4673 // use nds as scratch
4674 evmovdqul(nds, dst, Assembler::AVX_512bit);
4675 Assembler::vpsraw(nds, nds, shift, vector_len);
4676 evmovdqul(dst, nds, Assembler::AVX_512bit);
4677 } else {
4678 // use nds as scratch for xmm0
4679 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4680 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4681 Assembler::vpsraw(xmm0, xmm0, shift, vector_len);
4682 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4683 }
4684 }
4685
4686 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4687 int dst_enc = dst->encoding();
4688 int nds_enc = nds->encoding();
4689 int shift_enc = shift->encoding();
4690 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4691 Assembler::vpsrlw(dst, nds, shift, vector_len);
4692 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4693 Assembler::vpsrlw(dst, dst, shift, vector_len);
4694 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4695 // use nds_enc as scratch with shift
4696 evmovdqul(nds, shift, Assembler::AVX_512bit);
4697 Assembler::vpsrlw(dst, dst, nds, vector_len);
4698 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4699 // use nds as scratch with dst
4700 evmovdqul(nds, dst, Assembler::AVX_512bit);
4701 Assembler::vpsrlw(nds, nds, shift, vector_len);
4702 evmovdqul(dst, nds, Assembler::AVX_512bit);
4703 } else if (dst_enc < 16) {
4704 // use nds to save a copy of xmm0 and hold shift
4705 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4706 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4707 Assembler::vpsrlw(dst, dst, xmm0, vector_len);
4708 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4709 } else if (nds_enc < 16) {
4710 // use nds as dest as temps
4711 evmovdqul(nds, dst, Assembler::AVX_512bit);
4712 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4713 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4714 Assembler::vpsrlw(nds, nds, xmm0, vector_len);
4715 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4716 evmovdqul(dst, nds, Assembler::AVX_512bit);
4717 } else {
4718 // worse case scenario, all regs are in the upper bank
4719 subptr(rsp, 64);
4720 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4721 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4722 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4723 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4724 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4725 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4726 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4727 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4728 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4729 addptr(rsp, 64);
4730 }
4731 }
4732
4733 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4734 int dst_enc = dst->encoding();
4735 int nds_enc = nds->encoding();
4736 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4737 Assembler::vpsrlw(dst, nds, shift, vector_len);
4738 } else if (dst_enc < 16) {
4739 Assembler::vpsrlw(dst, dst, shift, vector_len);
4740 } else if (nds_enc < 16) {
4741 // use nds as scratch
4742 evmovdqul(nds, dst, Assembler::AVX_512bit);
4743 Assembler::vpsrlw(nds, nds, shift, vector_len);
4744 evmovdqul(dst, nds, Assembler::AVX_512bit);
4745 } else {
4746 // use nds as scratch for xmm0
4747 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4748 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4749 Assembler::vpsrlw(xmm0, xmm0, shift, vector_len);
4750 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4751 }
4752 }
4753
4754 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4755 int dst_enc = dst->encoding();
4756 int nds_enc = nds->encoding();
4757 int shift_enc = shift->encoding();
4758 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4759 Assembler::vpsllw(dst, nds, shift, vector_len);
4760 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4761 Assembler::vpsllw(dst, dst, shift, vector_len);
4762 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4763 // use nds_enc as scratch with shift
4764 evmovdqul(nds, shift, Assembler::AVX_512bit);
4765 Assembler::vpsllw(dst, dst, nds, vector_len);
4766 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4767 // use nds as scratch with dst
4768 evmovdqul(nds, dst, Assembler::AVX_512bit);
4769 Assembler::vpsllw(nds, nds, shift, vector_len);
4770 evmovdqul(dst, nds, Assembler::AVX_512bit);
4771 } else if (dst_enc < 16) {
4772 // use nds to save a copy of xmm0 and hold shift
4773 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4774 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4775 Assembler::vpsllw(dst, dst, xmm0, vector_len);
4776 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4777 } else if (nds_enc < 16) {
4778 // use nds as dest as temps
4779 evmovdqul(nds, dst, Assembler::AVX_512bit);
4780 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4781 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4782 Assembler::vpsllw(nds, nds, xmm0, vector_len);
4783 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4784 evmovdqul(dst, nds, Assembler::AVX_512bit);
4785 } else {
4786 // worse case scenario, all regs are in the upper bank
4787 subptr(rsp, 64);
4788 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4789 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4790 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4791 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4792 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4793 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4794 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4795 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4796 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4797 addptr(rsp, 64);
4798 }
4799 }
4800
4801 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4802 int dst_enc = dst->encoding();
4803 int nds_enc = nds->encoding();
4804 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4805 Assembler::vpsllw(dst, nds, shift, vector_len);
4806 } else if (dst_enc < 16) {
4807 Assembler::vpsllw(dst, dst, shift, vector_len);
4808 } else if (nds_enc < 16) {
4809 // use nds as scratch
4810 evmovdqul(nds, dst, Assembler::AVX_512bit);
4811 Assembler::vpsllw(nds, nds, shift, vector_len);
4812 evmovdqul(dst, nds, Assembler::AVX_512bit);
4813 } else {
4814 // use nds as scratch for xmm0
4815 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4816 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4817 Assembler::vpsllw(xmm0, xmm0, shift, vector_len);
4818 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4819 }
4820 }
4821
4822 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) {
4823 int dst_enc = dst->encoding();
4824 int src_enc = src->encoding();
4825 if ((dst_enc < 16) && (src_enc < 16)) {
4826 Assembler::vptest(dst, src);
4827 } else if (src_enc < 16) {
4828 subptr(rsp, 64);
4829 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4830 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4831 Assembler::vptest(xmm0, src);
4832 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4833 addptr(rsp, 64);
4834 } else if (dst_enc < 16) {
4835 subptr(rsp, 64);
4836 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4837 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4838 Assembler::vptest(dst, xmm0);
4839 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4840 addptr(rsp, 64);
4841 } else {
4842 subptr(rsp, 64);
4843 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4844 subptr(rsp, 64);
4845 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4846 movdqu(xmm0, src);
4847 movdqu(xmm1, dst);
4848 Assembler::vptest(xmm1, xmm0);
4849 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4850 addptr(rsp, 64);
4851 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4852 addptr(rsp, 64);
4853 }
4854 }
4855
4856 // This instruction exists within macros, ergo we cannot control its input
4857 // when emitted through those patterns.
4858 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) {
4859 if (VM_Version::supports_avx512nobw()) {
4860 int dst_enc = dst->encoding();
4861 int src_enc = src->encoding();
4862 if (dst_enc == src_enc) {
4863 if (dst_enc < 16) {
4864 Assembler::punpcklbw(dst, src);
4865 } else {
4866 subptr(rsp, 64);
4867 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4868 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4869 Assembler::punpcklbw(xmm0, xmm0);
4870 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4871 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4872 addptr(rsp, 64);
4873 }
4874 } else {
4875 if ((src_enc < 16) && (dst_enc < 16)) {
4876 Assembler::punpcklbw(dst, src);
4877 } else if (src_enc < 16) {
4878 subptr(rsp, 64);
4879 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4880 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4881 Assembler::punpcklbw(xmm0, src);
4882 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4883 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4884 addptr(rsp, 64);
4885 } else if (dst_enc < 16) {
4886 subptr(rsp, 64);
4887 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4888 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4889 Assembler::punpcklbw(dst, xmm0);
4890 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4891 addptr(rsp, 64);
4892 } else {
4893 subptr(rsp, 64);
4894 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4895 subptr(rsp, 64);
4896 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4897 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4898 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4899 Assembler::punpcklbw(xmm0, xmm1);
4900 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4901 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4902 addptr(rsp, 64);
4903 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4904 addptr(rsp, 64);
4905 }
4906 }
4907 } else {
4908 Assembler::punpcklbw(dst, src);
4909 }
4910 }
4911
4912 // This instruction exists within macros, ergo we cannot control its input
4913 // when emitted through those patterns.
4914 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) {
4915 if (VM_Version::supports_avx512nobw()) {
4916 int dst_enc = dst->encoding();
4917 int src_enc = src->encoding();
4918 if (dst_enc == src_enc) {
4919 if (dst_enc < 16) {
4920 Assembler::pshuflw(dst, src, mode);
4921 } else {
4922 subptr(rsp, 64);
4923 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4924 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4925 Assembler::pshuflw(xmm0, xmm0, mode);
4926 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4927 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4928 addptr(rsp, 64);
4929 }
4930 } else {
4931 if ((src_enc < 16) && (dst_enc < 16)) {
4932 Assembler::pshuflw(dst, src, mode);
4933 } else if (src_enc < 16) {
4934 subptr(rsp, 64);
4935 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4936 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4937 Assembler::pshuflw(xmm0, src, mode);
4938 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4939 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4940 addptr(rsp, 64);
4941 } else if (dst_enc < 16) {
4942 subptr(rsp, 64);
4943 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4944 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4945 Assembler::pshuflw(dst, xmm0, mode);
4946 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4947 addptr(rsp, 64);
4948 } else {
4949 subptr(rsp, 64);
4950 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4951 subptr(rsp, 64);
4952 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4953 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4954 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4955 Assembler::pshuflw(xmm0, xmm1, mode);
4956 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4957 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4958 addptr(rsp, 64);
4959 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4960 addptr(rsp, 64);
4961 }
4962 }
4963 } else {
4964 Assembler::pshuflw(dst, src, mode);
4965 }
4966 }
4967
4968 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
4969 if (reachable(src)) {
4970 vandpd(dst, nds, as_Address(src), vector_len);
4971 } else {
4972 lea(rscratch1, src);
4973 vandpd(dst, nds, Address(rscratch1, 0), vector_len);
4974 }
4975 }
4976
4977 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
4978 if (reachable(src)) {
4979 vandps(dst, nds, as_Address(src), vector_len);
4980 } else {
4981 lea(rscratch1, src);
4982 vandps(dst, nds, Address(rscratch1, 0), vector_len);
4983 }
4984 }
4985
4986 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4987 if (reachable(src)) {
4988 vdivsd(dst, nds, as_Address(src));
4989 } else {
4990 lea(rscratch1, src);
4991 vdivsd(dst, nds, Address(rscratch1, 0));
4992 }
4993 }
4994
4995 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4996 if (reachable(src)) {
4997 vdivss(dst, nds, as_Address(src));
4998 } else {
4999 lea(rscratch1, src);
5000 vdivss(dst, nds, Address(rscratch1, 0));
5001 }
5002 }
5003
5004 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5005 if (reachable(src)) {
5006 vmulsd(dst, nds, as_Address(src));
5007 } else {
5008 lea(rscratch1, src);
5009 vmulsd(dst, nds, Address(rscratch1, 0));
5010 }
5011 }
5012
5013 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5014 if (reachable(src)) {
5015 vmulss(dst, nds, as_Address(src));
5016 } else {
5017 lea(rscratch1, src);
5018 vmulss(dst, nds, Address(rscratch1, 0));
5019 }
5020 }
5021
5022 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5023 if (reachable(src)) {
5024 vsubsd(dst, nds, as_Address(src));
5025 } else {
5026 lea(rscratch1, src);
5027 vsubsd(dst, nds, Address(rscratch1, 0));
5028 }
5029 }
5030
5031 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5032 if (reachable(src)) {
5033 vsubss(dst, nds, as_Address(src));
5034 } else {
5035 lea(rscratch1, src);
5036 vsubss(dst, nds, Address(rscratch1, 0));
5037 }
5038 }
5039
5040 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5041 int nds_enc = nds->encoding();
5042 int dst_enc = dst->encoding();
5043 bool dst_upper_bank = (dst_enc > 15);
5044 bool nds_upper_bank = (nds_enc > 15);
5045 if (VM_Version::supports_avx512novl() &&
5046 (nds_upper_bank || dst_upper_bank)) {
5047 if (dst_upper_bank) {
5048 subptr(rsp, 64);
5049 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5050 movflt(xmm0, nds);
5051 vxorps(xmm0, xmm0, src, Assembler::AVX_128bit);
5052 movflt(dst, xmm0);
5053 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5054 addptr(rsp, 64);
5055 } else {
5056 movflt(dst, nds);
5057 vxorps(dst, dst, src, Assembler::AVX_128bit);
5058 }
5059 } else {
5060 vxorps(dst, nds, src, Assembler::AVX_128bit);
5061 }
5062 }
5063
5064 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5065 int nds_enc = nds->encoding();
5066 int dst_enc = dst->encoding();
5067 bool dst_upper_bank = (dst_enc > 15);
5068 bool nds_upper_bank = (nds_enc > 15);
5069 if (VM_Version::supports_avx512novl() &&
5070 (nds_upper_bank || dst_upper_bank)) {
5071 if (dst_upper_bank) {
5072 subptr(rsp, 64);
5073 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5074 movdbl(xmm0, nds);
5075 vxorpd(xmm0, xmm0, src, Assembler::AVX_128bit);
5076 movdbl(dst, xmm0);
5077 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5078 addptr(rsp, 64);
5079 } else {
5080 movdbl(dst, nds);
5081 vxorpd(dst, dst, src, Assembler::AVX_128bit);
5082 }
5083 } else {
5084 vxorpd(dst, nds, src, Assembler::AVX_128bit);
5085 }
5086 }
5087
5088 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5089 if (reachable(src)) {
5090 vxorpd(dst, nds, as_Address(src), vector_len);
5091 } else {
5092 lea(rscratch1, src);
5093 vxorpd(dst, nds, Address(rscratch1, 0), vector_len);
5094 }
5095 }
5096
5097 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5098 if (reachable(src)) {
5099 vxorps(dst, nds, as_Address(src), vector_len);
5100 } else {
5101 lea(rscratch1, src);
5102 vxorps(dst, nds, Address(rscratch1, 0), vector_len);
5103 }
5104 }
5105
5106
5107 //////////////////////////////////////////////////////////////////////////////////
5108 #if INCLUDE_ALL_GCS
5109
5110 void MacroAssembler::g1_write_barrier_pre(Register obj,
5111 Register pre_val,
5112 Register thread,
5113 Register tmp,
5114 bool tosca_live,
5115 bool expand_call) {
5116
5117 // If expand_call is true then we expand the call_VM_leaf macro
5118 // directly to skip generating the check by
5119 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
5120
5121 #ifdef _LP64
5122 assert(thread == r15_thread, "must be");
5123 #endif // _LP64
5124
5125 Label done;
5126 Label runtime;
5127
5128 assert(pre_val != noreg, "check this code");
5129
5130 if (obj != noreg) {
5131 assert_different_registers(obj, pre_val, tmp);
5132 assert(pre_val != rax, "check this code");
5133 }
5134
5135 Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
5136 SATBMarkQueue::byte_offset_of_active()));
5137 Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
5138 SATBMarkQueue::byte_offset_of_index()));
5139 Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
5140 SATBMarkQueue::byte_offset_of_buf()));
5141
5142
5143 // Is marking active?
5144 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
5145 cmpl(in_progress, 0);
5146 } else {
5147 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
5148 cmpb(in_progress, 0);
5149 }
5150 jcc(Assembler::equal, done);
5151
5152 // Do we need to load the previous value?
5153 if (obj != noreg) {
5154 load_heap_oop(pre_val, Address(obj, 0));
5155 }
5156
5157 // Is the previous value null?
5158 cmpptr(pre_val, (int32_t) NULL_WORD);
5159 jcc(Assembler::equal, done);
5160
5161 // Can we store original value in the thread's buffer?
5162 // Is index == 0?
5163 // (The index field is typed as size_t.)
5164
5165 movptr(tmp, index); // tmp := *index_adr
5166 cmpptr(tmp, 0); // tmp == 0?
5167 jcc(Assembler::equal, runtime); // If yes, goto runtime
5168
5169 subptr(tmp, wordSize); // tmp := tmp - wordSize
5170 movptr(index, tmp); // *index_adr := tmp
5171 addptr(tmp, buffer); // tmp := tmp + *buffer_adr
5172
5173 // Record the previous value
5174 movptr(Address(tmp, 0), pre_val);
5175 jmp(done);
5176
5177 bind(runtime);
5178 // save the live input values
5179 if(tosca_live) push(rax);
5180
5181 if (obj != noreg && obj != rax)
5182 push(obj);
5183
5184 if (pre_val != rax)
5185 push(pre_val);
5186
5187 // Calling the runtime using the regular call_VM_leaf mechanism generates
5188 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
5189 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
5190 //
5191 // If we care generating the pre-barrier without a frame (e.g. in the
5192 // intrinsified Reference.get() routine) then ebp might be pointing to
5193 // the caller frame and so this check will most likely fail at runtime.
5194 //
5195 // Expanding the call directly bypasses the generation of the check.
5196 // So when we do not have have a full interpreter frame on the stack
5197 // expand_call should be passed true.
5198
5199 NOT_LP64( push(thread); )
5200
5201 if (expand_call) {
5202 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
5203 pass_arg1(this, thread);
5204 pass_arg0(this, pre_val);
5205 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2);
5206 } else {
5207 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread);
5208 }
5209
5210 NOT_LP64( pop(thread); )
5211
5212 // save the live input values
5213 if (pre_val != rax)
5214 pop(pre_val);
5215
5216 if (obj != noreg && obj != rax)
5217 pop(obj);
5218
5219 if(tosca_live) pop(rax);
5220
5221 bind(done);
5222 }
5223
5224 void MacroAssembler::g1_write_barrier_post(Register store_addr,
5225 Register new_val,
5226 Register thread,
5227 Register tmp,
5228 Register tmp2) {
5229 #ifdef _LP64
5230 assert(thread == r15_thread, "must be");
5231 #endif // _LP64
5232
5233 Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() +
5234 DirtyCardQueue::byte_offset_of_index()));
5235 Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() +
5236 DirtyCardQueue::byte_offset_of_buf()));
5237
5238 CardTableModRefBS* ct =
5239 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
5240 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
5241
5242 Label done;
5243 Label runtime;
5244
5245 // Does store cross heap regions?
5246
5247 movptr(tmp, store_addr);
5248 xorptr(tmp, new_val);
5249 shrptr(tmp, HeapRegion::LogOfHRGrainBytes);
5250 jcc(Assembler::equal, done);
5251
5252 // crosses regions, storing NULL?
5253
5254 cmpptr(new_val, (int32_t) NULL_WORD);
5255 jcc(Assembler::equal, done);
5256
5257 // storing region crossing non-NULL, is card already dirty?
5258
5259 const Register card_addr = tmp;
5260 const Register cardtable = tmp2;
5261
5262 movptr(card_addr, store_addr);
5263 shrptr(card_addr, CardTableModRefBS::card_shift);
5264 // Do not use ExternalAddress to load 'byte_map_base', since 'byte_map_base' is NOT
5265 // a valid address and therefore is not properly handled by the relocation code.
5266 movptr(cardtable, (intptr_t)ct->byte_map_base);
5267 addptr(card_addr, cardtable);
5268
5269 cmpb(Address(card_addr, 0), (int)G1SATBCardTableModRefBS::g1_young_card_val());
5270 jcc(Assembler::equal, done);
5271
5272 membar(Assembler::Membar_mask_bits(Assembler::StoreLoad));
5273 cmpb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val());
5274 jcc(Assembler::equal, done);
5275
5276
5277 // storing a region crossing, non-NULL oop, card is clean.
5278 // dirty card and log.
5279
5280 movb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val());
5281
5282 cmpl(queue_index, 0);
5283 jcc(Assembler::equal, runtime);
5284 subl(queue_index, wordSize);
5285 movptr(tmp2, buffer);
5286 #ifdef _LP64
5287 movslq(rscratch1, queue_index);
5288 addq(tmp2, rscratch1);
5289 movq(Address(tmp2, 0), card_addr);
5290 #else
5291 addl(tmp2, queue_index);
5292 movl(Address(tmp2, 0), card_addr);
5293 #endif
5294 jmp(done);
5295
5296 bind(runtime);
5297 // save the live input values
5298 push(store_addr);
5299 push(new_val);
5300 #ifdef _LP64
5301 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread);
5302 #else
5303 push(thread);
5304 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread);
5305 pop(thread);
5306 #endif
5307 pop(new_val);
5308 pop(store_addr);
5309
5310 bind(done);
5311 }
5312
5313 #endif // INCLUDE_ALL_GCS
5314 //////////////////////////////////////////////////////////////////////////////////
5315
5316
5317 void MacroAssembler::store_check(Register obj, Address dst) {
5318 store_check(obj);
5319 }
5320
5321 void MacroAssembler::store_check(Register obj) {
5322 // Does a store check for the oop in register obj. The content of
5323 // register obj is destroyed afterwards.
5324 BarrierSet* bs = Universe::heap()->barrier_set();
5325 assert(bs->kind() == BarrierSet::CardTableForRS ||
5326 bs->kind() == BarrierSet::CardTableExtension,
5327 "Wrong barrier set kind");
5328
5329 CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
5330 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
5331
5332 shrptr(obj, CardTableModRefBS::card_shift);
5333
5334 Address card_addr;
5335
5336 // The calculation for byte_map_base is as follows:
5337 // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift);
5338 // So this essentially converts an address to a displacement and it will
5339 // never need to be relocated. On 64bit however the value may be too
5340 // large for a 32bit displacement.
5341 intptr_t disp = (intptr_t) ct->byte_map_base;
5342 if (is_simm32(disp)) {
5343 card_addr = Address(noreg, obj, Address::times_1, disp);
5344 } else {
5345 // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative
5346 // displacement and done in a single instruction given favorable mapping and a
5347 // smarter version of as_Address. However, 'ExternalAddress' generates a relocation
5348 // entry and that entry is not properly handled by the relocation code.
5349 AddressLiteral cardtable((address)ct->byte_map_base, relocInfo::none);
5350 Address index(noreg, obj, Address::times_1);
5351 card_addr = as_Address(ArrayAddress(cardtable, index));
5352 }
5353
5354 int dirty = CardTableModRefBS::dirty_card_val();
5355 if (UseCondCardMark) {
5356 Label L_already_dirty;
5357 if (UseConcMarkSweepGC) {
5358 membar(Assembler::StoreLoad);
5359 }
5360 cmpb(card_addr, dirty);
5361 jcc(Assembler::equal, L_already_dirty);
5362 movb(card_addr, dirty);
5363 bind(L_already_dirty);
5364 } else {
5365 movb(card_addr, dirty);
5366 }
5367 }
5368
5369 void MacroAssembler::subptr(Register dst, int32_t imm32) {
5370 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32));
5371 }
5372
5373 // Force generation of a 4 byte immediate value even if it fits into 8bit
5374 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) {
5375 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32));
5376 }
5377
5378 void MacroAssembler::subptr(Register dst, Register src) {
5379 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src));
5380 }
5381
5382 // C++ bool manipulation
5383 void MacroAssembler::testbool(Register dst) {
5384 if(sizeof(bool) == 1)
5385 testb(dst, 0xff);
5386 else if(sizeof(bool) == 2) {
5387 // testw implementation needed for two byte bools
5388 ShouldNotReachHere();
5389 } else if(sizeof(bool) == 4)
5390 testl(dst, dst);
5391 else
5392 // unsupported
5393 ShouldNotReachHere();
5394 }
5395
5396 void MacroAssembler::testptr(Register dst, Register src) {
5397 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src));
5398 }
5399
5400 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
5401 void MacroAssembler::tlab_allocate(Register obj,
5402 Register var_size_in_bytes,
5403 int con_size_in_bytes,
5404 Register t1,
5405 Register t2,
5406 Label& slow_case) {
5407 assert_different_registers(obj, t1, t2);
5408 assert_different_registers(obj, var_size_in_bytes, t1);
5409 Register end = t2;
5410 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread);
5411
5412 verify_tlab();
5413
5414 NOT_LP64(get_thread(thread));
5415
5416 movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
5417 if (var_size_in_bytes == noreg) {
5418 lea(end, Address(obj, con_size_in_bytes));
5419 } else {
5420 lea(end, Address(obj, var_size_in_bytes, Address::times_1));
5421 }
5422 cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
5423 jcc(Assembler::above, slow_case);
5424
5425 // update the tlab top pointer
5426 movptr(Address(thread, JavaThread::tlab_top_offset()), end);
5427
5428 // recover var_size_in_bytes if necessary
5429 if (var_size_in_bytes == end) {
5430 subptr(var_size_in_bytes, obj);
5431 }
5432 verify_tlab();
5433 }
5434
5435 // Preserves rbx, and rdx.
5436 Register MacroAssembler::tlab_refill(Label& retry,
5437 Label& try_eden,
5438 Label& slow_case) {
5439 Register top = rax;
5440 Register t1 = rcx; // object size
5441 Register t2 = rsi;
5442 Register thread_reg = NOT_LP64(rdi) LP64_ONLY(r15_thread);
5443 assert_different_registers(top, thread_reg, t1, t2, /* preserve: */ rbx, rdx);
5444 Label do_refill, discard_tlab;
5445
5446 if (!Universe::heap()->supports_inline_contig_alloc()) {
5447 // No allocation in the shared eden.
5448 jmp(slow_case);
5449 }
5450
5451 NOT_LP64(get_thread(thread_reg));
5452
5453 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
5454 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
5455
5456 // calculate amount of free space
5457 subptr(t1, top);
5458 shrptr(t1, LogHeapWordSize);
5459
5460 // Retain tlab and allocate object in shared space if
5461 // the amount free in the tlab is too large to discard.
5462 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())));
5463 jcc(Assembler::lessEqual, discard_tlab);
5464
5465 // Retain
5466 // %%% yuck as movptr...
5467 movptr(t2, (int32_t) ThreadLocalAllocBuffer::refill_waste_limit_increment());
5468 addptr(Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())), t2);
5469 if (TLABStats) {
5470 // increment number of slow_allocations
5471 addl(Address(thread_reg, in_bytes(JavaThread::tlab_slow_allocations_offset())), 1);
5472 }
5473 jmp(try_eden);
5474
5475 bind(discard_tlab);
5476 if (TLABStats) {
5477 // increment number of refills
5478 addl(Address(thread_reg, in_bytes(JavaThread::tlab_number_of_refills_offset())), 1);
5479 // accumulate wastage -- t1 is amount free in tlab
5480 addl(Address(thread_reg, in_bytes(JavaThread::tlab_fast_refill_waste_offset())), t1);
5481 }
5482
5483 // if tlab is currently allocated (top or end != null) then
5484 // fill [top, end + alignment_reserve) with array object
5485 testptr(top, top);
5486 jcc(Assembler::zero, do_refill);
5487
5488 // set up the mark word
5489 movptr(Address(top, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::prototype()->copy_set_hash(0x2));
5490 // set the length to the remaining space
5491 subptr(t1, typeArrayOopDesc::header_size(T_INT));
5492 addptr(t1, (int32_t)ThreadLocalAllocBuffer::alignment_reserve());
5493 shlptr(t1, log2_intptr(HeapWordSize/sizeof(jint)));
5494 movl(Address(top, arrayOopDesc::length_offset_in_bytes()), t1);
5495 // set klass to intArrayKlass
5496 // dubious reloc why not an oop reloc?
5497 movptr(t1, ExternalAddress((address)Universe::intArrayKlassObj_addr()));
5498 // store klass last. concurrent gcs assumes klass length is valid if
5499 // klass field is not null.
5500 store_klass(top, t1);
5501
5502 movptr(t1, top);
5503 subptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
5504 incr_allocated_bytes(thread_reg, t1, 0);
5505
5506 // refill the tlab with an eden allocation
5507 bind(do_refill);
5508 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset())));
5509 shlptr(t1, LogHeapWordSize);
5510 // allocate new tlab, address returned in top
5511 eden_allocate(top, t1, 0, t2, slow_case);
5512
5513 // Check that t1 was preserved in eden_allocate.
5514 #ifdef ASSERT
5515 if (UseTLAB) {
5516 Label ok;
5517 Register tsize = rsi;
5518 assert_different_registers(tsize, thread_reg, t1);
5519 push(tsize);
5520 movptr(tsize, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset())));
5521 shlptr(tsize, LogHeapWordSize);
5522 cmpptr(t1, tsize);
5523 jcc(Assembler::equal, ok);
5524 STOP("assert(t1 != tlab size)");
5525 should_not_reach_here();
5526
5527 bind(ok);
5528 pop(tsize);
5529 }
5530 #endif
5531 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())), top);
5532 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())), top);
5533 addptr(top, t1);
5534 subptr(top, (int32_t)ThreadLocalAllocBuffer::alignment_reserve_in_bytes());
5535 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())), top);
5536
5537 if (ZeroTLAB) {
5538 // This is a fast TLAB refill, therefore the GC is not notified of it.
5539 // So compiled code must fill the new TLAB with zeroes.
5540 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
5541 zero_memory(top, t1, 0, t2);
5542 }
5543
5544 verify_tlab();
5545 jmp(retry);
5546
5547 return thread_reg; // for use by caller
5548 }
5549
5550 // Preserves the contents of address, destroys the contents length_in_bytes and temp.
5551 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) {
5552 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different");
5553 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord");
5554 Label done;
5555
5556 testptr(length_in_bytes, length_in_bytes);
5557 jcc(Assembler::zero, done);
5558
5559 // initialize topmost word, divide index by 2, check if odd and test if zero
5560 // note: for the remaining code to work, index must be a multiple of BytesPerWord
5561 #ifdef ASSERT
5562 {
5563 Label L;
5564 testptr(length_in_bytes, BytesPerWord - 1);
5565 jcc(Assembler::zero, L);
5566 stop("length must be a multiple of BytesPerWord");
5567 bind(L);
5568 }
5569 #endif
5570 Register index = length_in_bytes;
5571 xorptr(temp, temp); // use _zero reg to clear memory (shorter code)
5572 if (UseIncDec) {
5573 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set
5574 } else {
5575 shrptr(index, 2); // use 2 instructions to avoid partial flag stall
5576 shrptr(index, 1);
5577 }
5578 #ifndef _LP64
5579 // index could have not been a multiple of 8 (i.e., bit 2 was set)
5580 {
5581 Label even;
5582 // note: if index was a multiple of 8, then it cannot
5583 // be 0 now otherwise it must have been 0 before
5584 // => if it is even, we don't need to check for 0 again
5585 jcc(Assembler::carryClear, even);
5586 // clear topmost word (no jump would be needed if conditional assignment worked here)
5587 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp);
5588 // index could be 0 now, must check again
5589 jcc(Assembler::zero, done);
5590 bind(even);
5591 }
5592 #endif // !_LP64
5593 // initialize remaining object fields: index is a multiple of 2 now
5594 {
5595 Label loop;
5596 bind(loop);
5597 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp);
5598 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);)
5599 decrement(index);
5600 jcc(Assembler::notZero, loop);
5601 }
5602
5603 bind(done);
5604 }
5605
5606 void MacroAssembler::incr_allocated_bytes(Register thread,
5607 Register var_size_in_bytes,
5608 int con_size_in_bytes,
5609 Register t1) {
5610 if (!thread->is_valid()) {
5611 #ifdef _LP64
5612 thread = r15_thread;
5613 #else
5614 assert(t1->is_valid(), "need temp reg");
5615 thread = t1;
5616 get_thread(thread);
5617 #endif
5618 }
5619
5620 #ifdef _LP64
5621 if (var_size_in_bytes->is_valid()) {
5622 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
5623 } else {
5624 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
5625 }
5626 #else
5627 if (var_size_in_bytes->is_valid()) {
5628 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
5629 } else {
5630 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
5631 }
5632 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0);
5633 #endif
5634 }
5635
5636 void MacroAssembler::fp_runtime_fallback(address runtime_entry, int nb_args, int num_fpu_regs_in_use) {
5637 pusha();
5638
5639 // if we are coming from c1, xmm registers may be live
5640 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
5641 if (UseAVX > 2) {
5642 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
5643 }
5644
5645 if (UseSSE == 1) {
5646 subptr(rsp, sizeof(jdouble)*8);
5647 for (int n = 0; n < 8; n++) {
5648 movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n));
5649 }
5650 } else if (UseSSE >= 2) {
5651 if (UseAVX > 2) {
5652 push(rbx);
5653 movl(rbx, 0xffff);
5654 kmovwl(k1, rbx);
5655 pop(rbx);
5656 }
5657 #ifdef COMPILER2
5658 if (MaxVectorSize > 16) {
5659 if(UseAVX > 2) {
5660 // Save upper half of ZMM registers
5661 subptr(rsp, 32*num_xmm_regs);
5662 for (int n = 0; n < num_xmm_regs; n++) {
5663 vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n));
5664 }
5665 }
5666 assert(UseAVX > 0, "256 bit vectors are supported only with AVX");
5667 // Save upper half of YMM registers
5668 subptr(rsp, 16*num_xmm_regs);
5669 for (int n = 0; n < num_xmm_regs; n++) {
5670 vextractf128_high(Address(rsp, n*16), as_XMMRegister(n));
5671 }
5672 }
5673 #endif
5674 // Save whole 128bit (16 bytes) XMM registers
5675 subptr(rsp, 16*num_xmm_regs);
5676 #ifdef _LP64
5677 if (VM_Version::supports_evex()) {
5678 for (int n = 0; n < num_xmm_regs; n++) {
5679 vextractf32x4(Address(rsp, n*16), as_XMMRegister(n), 0);
5680 }
5681 } else {
5682 for (int n = 0; n < num_xmm_regs; n++) {
5683 movdqu(Address(rsp, n*16), as_XMMRegister(n));
5684 }
5685 }
5686 #else
5687 for (int n = 0; n < num_xmm_regs; n++) {
5688 movdqu(Address(rsp, n*16), as_XMMRegister(n));
5689 }
5690 #endif
5691 }
5692
5693 // Preserve registers across runtime call
5694 int incoming_argument_and_return_value_offset = -1;
5695 if (num_fpu_regs_in_use > 1) {
5696 // Must preserve all other FPU regs (could alternatively convert
5697 // SharedRuntime::dsin, dcos etc. into assembly routines known not to trash
5698 // FPU state, but can not trust C compiler)
5699 NEEDS_CLEANUP;
5700 // NOTE that in this case we also push the incoming argument(s) to
5701 // the stack and restore it later; we also use this stack slot to
5702 // hold the return value from dsin, dcos etc.
5703 for (int i = 0; i < num_fpu_regs_in_use; i++) {
5704 subptr(rsp, sizeof(jdouble));
5705 fstp_d(Address(rsp, 0));
5706 }
5707 incoming_argument_and_return_value_offset = sizeof(jdouble)*(num_fpu_regs_in_use-1);
5708 for (int i = nb_args-1; i >= 0; i--) {
5709 fld_d(Address(rsp, incoming_argument_and_return_value_offset-i*sizeof(jdouble)));
5710 }
5711 }
5712
5713 subptr(rsp, nb_args*sizeof(jdouble));
5714 for (int i = 0; i < nb_args; i++) {
5715 fstp_d(Address(rsp, i*sizeof(jdouble)));
5716 }
5717
5718 #ifdef _LP64
5719 if (nb_args > 0) {
5720 movdbl(xmm0, Address(rsp, 0));
5721 }
5722 if (nb_args > 1) {
5723 movdbl(xmm1, Address(rsp, sizeof(jdouble)));
5724 }
5725 assert(nb_args <= 2, "unsupported number of args");
5726 #endif // _LP64
5727
5728 // NOTE: we must not use call_VM_leaf here because that requires a
5729 // complete interpreter frame in debug mode -- same bug as 4387334
5730 // MacroAssembler::call_VM_leaf_base is perfectly safe and will
5731 // do proper 64bit abi
5732
5733 NEEDS_CLEANUP;
5734 // Need to add stack banging before this runtime call if it needs to
5735 // be taken; however, there is no generic stack banging routine at
5736 // the MacroAssembler level
5737
5738 MacroAssembler::call_VM_leaf_base(runtime_entry, 0);
5739
5740 #ifdef _LP64
5741 movsd(Address(rsp, 0), xmm0);
5742 fld_d(Address(rsp, 0));
5743 #endif // _LP64
5744 addptr(rsp, sizeof(jdouble)*nb_args);
5745 if (num_fpu_regs_in_use > 1) {
5746 // Must save return value to stack and then restore entire FPU
5747 // stack except incoming arguments
5748 fstp_d(Address(rsp, incoming_argument_and_return_value_offset));
5749 for (int i = 0; i < num_fpu_regs_in_use - nb_args; i++) {
5750 fld_d(Address(rsp, 0));
5751 addptr(rsp, sizeof(jdouble));
5752 }
5753 fld_d(Address(rsp, (nb_args-1)*sizeof(jdouble)));
5754 addptr(rsp, sizeof(jdouble)*nb_args);
5755 }
5756
5757 if (UseSSE == 1) {
5758 for (int n = 0; n < 8; n++) {
5759 movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble)));
5760 }
5761 addptr(rsp, sizeof(jdouble)*8);
5762 } else if (UseSSE >= 2) {
5763 // Restore whole 128bit (16 bytes) XMM registers
5764 #ifdef _LP64
5765 if (VM_Version::supports_evex()) {
5766 for (int n = 0; n < num_xmm_regs; n++) {
5767 vinsertf32x4(as_XMMRegister(n), as_XMMRegister(n), Address(rsp, n*16), 0);
5768 }
5769 } else {
5770 for (int n = 0; n < num_xmm_regs; n++) {
5771 movdqu(as_XMMRegister(n), Address(rsp, n*16));
5772 }
5773 }
5774 #else
5775 for (int n = 0; n < num_xmm_regs; n++) {
5776 movdqu(as_XMMRegister(n), Address(rsp, n*16));
5777 }
5778 #endif
5779 addptr(rsp, 16*num_xmm_regs);
5780
5781 #ifdef COMPILER2
5782 if (MaxVectorSize > 16) {
5783 // Restore upper half of YMM registers.
5784 for (int n = 0; n < num_xmm_regs; n++) {
5785 vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16));
5786 }
5787 addptr(rsp, 16*num_xmm_regs);
5788 if(UseAVX > 2) {
5789 for (int n = 0; n < num_xmm_regs; n++) {
5790 vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32));
5791 }
5792 addptr(rsp, 32*num_xmm_regs);
5793 }
5794 }
5795 #endif
5796 }
5797 popa();
5798 }
5799
5800 static const double pi_4 = 0.7853981633974483;
5801
5802 void MacroAssembler::trigfunc(char trig, int num_fpu_regs_in_use) {
5803 // A hand-coded argument reduction for values in fabs(pi/4, pi/2)
5804 // was attempted in this code; unfortunately it appears that the
5805 // switch to 80-bit precision and back causes this to be
5806 // unprofitable compared with simply performing a runtime call if
5807 // the argument is out of the (-pi/4, pi/4) range.
5808
5809 Register tmp = noreg;
5810 if (!VM_Version::supports_cmov()) {
5811 // fcmp needs a temporary so preserve rbx,
5812 tmp = rbx;
5813 push(tmp);
5814 }
5815
5816 Label slow_case, done;
5817 if (trig == 't') {
5818 ExternalAddress pi4_adr = (address)&pi_4;
5819 if (reachable(pi4_adr)) {
5820 // x ?<= pi/4
5821 fld_d(pi4_adr);
5822 fld_s(1); // Stack: X PI/4 X
5823 fabs(); // Stack: |X| PI/4 X
5824 fcmp(tmp);
5825 jcc(Assembler::above, slow_case);
5826
5827 // fastest case: -pi/4 <= x <= pi/4
5828 ftan();
5829
5830 jmp(done);
5831 }
5832 }
5833 // slow case: runtime call
5834 bind(slow_case);
5835
5836 switch(trig) {
5837 case 's':
5838 {
5839 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1, num_fpu_regs_in_use);
5840 }
5841 break;
5842 case 'c':
5843 {
5844 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1, num_fpu_regs_in_use);
5845 }
5846 break;
5847 case 't':
5848 {
5849 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1, num_fpu_regs_in_use);
5850 }
5851 break;
5852 default:
5853 assert(false, "bad intrinsic");
5854 break;
5855 }
5856
5857 // Come here with result in F-TOS
5858 bind(done);
5859
5860 if (tmp != noreg) {
5861 pop(tmp);
5862 }
5863 }
5864
5865 // Look up the method for a megamorphic invokeinterface call.
5866 // The target method is determined by <intf_klass, itable_index>.
5867 // The receiver klass is in recv_klass.
5868 // On success, the result will be in method_result, and execution falls through.
5869 // On failure, execution transfers to the given label.
5870 void MacroAssembler::lookup_interface_method(Register recv_klass,
5871 Register intf_klass,
5872 RegisterOrConstant itable_index,
5873 Register method_result,
5874 Register scan_temp,
5875 Label& L_no_such_interface) {
5876 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
5877 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
5878 "caller must use same register for non-constant itable index as for method");
5879
5880 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
5881 int vtable_base = in_bytes(Klass::vtable_start_offset());
5882 int itentry_off = itableMethodEntry::method_offset_in_bytes();
5883 int scan_step = itableOffsetEntry::size() * wordSize;
5884 int vte_size = vtableEntry::size_in_bytes();
5885 Address::ScaleFactor times_vte_scale = Address::times_ptr;
5886 assert(vte_size == wordSize, "else adjust times_vte_scale");
5887
5888 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
5889
5890 // %%% Could store the aligned, prescaled offset in the klassoop.
5891 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
5892
5893 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
5894 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
5895 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
5896
5897 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
5898 // if (scan->interface() == intf) {
5899 // result = (klass + scan->offset() + itable_index);
5900 // }
5901 // }
5902 Label search, found_method;
5903
5904 for (int peel = 1; peel >= 0; peel--) {
5905 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
5906 cmpptr(intf_klass, method_result);
5907
5908 if (peel) {
5909 jccb(Assembler::equal, found_method);
5910 } else {
5911 jccb(Assembler::notEqual, search);
5912 // (invert the test to fall through to found_method...)
5913 }
5914
5915 if (!peel) break;
5916
5917 bind(search);
5918
5919 // Check that the previous entry is non-null. A null entry means that
5920 // the receiver class doesn't implement the interface, and wasn't the
5921 // same as when the caller was compiled.
5922 testptr(method_result, method_result);
5923 jcc(Assembler::zero, L_no_such_interface);
5924 addptr(scan_temp, scan_step);
5925 }
5926
5927 bind(found_method);
5928
5929 // Got a hit.
5930 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
5931 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
5932 }
5933
5934
5935 // virtual method calling
5936 void MacroAssembler::lookup_virtual_method(Register recv_klass,
5937 RegisterOrConstant vtable_index,
5938 Register method_result) {
5939 const int base = in_bytes(Klass::vtable_start_offset());
5940 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below");
5941 Address vtable_entry_addr(recv_klass,
5942 vtable_index, Address::times_ptr,
5943 base + vtableEntry::method_offset_in_bytes());
5944 movptr(method_result, vtable_entry_addr);
5945 }
5946
5947
5948 void MacroAssembler::check_klass_subtype(Register sub_klass,
5949 Register super_klass,
5950 Register temp_reg,
5951 Label& L_success) {
5952 Label L_failure;
5953 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL);
5954 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
5955 bind(L_failure);
5956 }
5957
5958
5959 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
5960 Register super_klass,
5961 Register temp_reg,
5962 Label* L_success,
5963 Label* L_failure,
5964 Label* L_slow_path,
5965 RegisterOrConstant super_check_offset) {
5966 assert_different_registers(sub_klass, super_klass, temp_reg);
5967 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
5968 if (super_check_offset.is_register()) {
5969 assert_different_registers(sub_klass, super_klass,
5970 super_check_offset.as_register());
5971 } else if (must_load_sco) {
5972 assert(temp_reg != noreg, "supply either a temp or a register offset");
5973 }
5974
5975 Label L_fallthrough;
5976 int label_nulls = 0;
5977 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
5978 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
5979 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
5980 assert(label_nulls <= 1, "at most one NULL in the batch");
5981
5982 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
5983 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5984 Address super_check_offset_addr(super_klass, sco_offset);
5985
5986 // Hacked jcc, which "knows" that L_fallthrough, at least, is in
5987 // range of a jccb. If this routine grows larger, reconsider at
5988 // least some of these.
5989 #define local_jcc(assembler_cond, label) \
5990 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \
5991 else jcc( assembler_cond, label) /*omit semi*/
5992
5993 // Hacked jmp, which may only be used just before L_fallthrough.
5994 #define final_jmp(label) \
5995 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
5996 else jmp(label) /*omit semi*/
5997
5998 // If the pointers are equal, we are done (e.g., String[] elements).
5999 // This self-check enables sharing of secondary supertype arrays among
6000 // non-primary types such as array-of-interface. Otherwise, each such
6001 // type would need its own customized SSA.
6002 // We move this check to the front of the fast path because many
6003 // type checks are in fact trivially successful in this manner,
6004 // so we get a nicely predicted branch right at the start of the check.
6005 cmpptr(sub_klass, super_klass);
6006 local_jcc(Assembler::equal, *L_success);
6007
6008 // Check the supertype display:
6009 if (must_load_sco) {
6010 // Positive movl does right thing on LP64.
6011 movl(temp_reg, super_check_offset_addr);
6012 super_check_offset = RegisterOrConstant(temp_reg);
6013 }
6014 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0);
6015 cmpptr(super_klass, super_check_addr); // load displayed supertype
6016
6017 // This check has worked decisively for primary supers.
6018 // Secondary supers are sought in the super_cache ('super_cache_addr').
6019 // (Secondary supers are interfaces and very deeply nested subtypes.)
6020 // This works in the same check above because of a tricky aliasing
6021 // between the super_cache and the primary super display elements.
6022 // (The 'super_check_addr' can address either, as the case requires.)
6023 // Note that the cache is updated below if it does not help us find
6024 // what we need immediately.
6025 // So if it was a primary super, we can just fail immediately.
6026 // Otherwise, it's the slow path for us (no success at this point).
6027
6028 if (super_check_offset.is_register()) {
6029 local_jcc(Assembler::equal, *L_success);
6030 cmpl(super_check_offset.as_register(), sc_offset);
6031 if (L_failure == &L_fallthrough) {
6032 local_jcc(Assembler::equal, *L_slow_path);
6033 } else {
6034 local_jcc(Assembler::notEqual, *L_failure);
6035 final_jmp(*L_slow_path);
6036 }
6037 } else if (super_check_offset.as_constant() == sc_offset) {
6038 // Need a slow path; fast failure is impossible.
6039 if (L_slow_path == &L_fallthrough) {
6040 local_jcc(Assembler::equal, *L_success);
6041 } else {
6042 local_jcc(Assembler::notEqual, *L_slow_path);
6043 final_jmp(*L_success);
6044 }
6045 } else {
6046 // No slow path; it's a fast decision.
6047 if (L_failure == &L_fallthrough) {
6048 local_jcc(Assembler::equal, *L_success);
6049 } else {
6050 local_jcc(Assembler::notEqual, *L_failure);
6051 final_jmp(*L_success);
6052 }
6053 }
6054
6055 bind(L_fallthrough);
6056
6057 #undef local_jcc
6058 #undef final_jmp
6059 }
6060
6061
6062 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
6063 Register super_klass,
6064 Register temp_reg,
6065 Register temp2_reg,
6066 Label* L_success,
6067 Label* L_failure,
6068 bool set_cond_codes) {
6069 assert_different_registers(sub_klass, super_klass, temp_reg);
6070 if (temp2_reg != noreg)
6071 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg);
6072 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
6073
6074 Label L_fallthrough;
6075 int label_nulls = 0;
6076 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
6077 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
6078 assert(label_nulls <= 1, "at most one NULL in the batch");
6079
6080 // a couple of useful fields in sub_klass:
6081 int ss_offset = in_bytes(Klass::secondary_supers_offset());
6082 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
6083 Address secondary_supers_addr(sub_klass, ss_offset);
6084 Address super_cache_addr( sub_klass, sc_offset);
6085
6086 // Do a linear scan of the secondary super-klass chain.
6087 // This code is rarely used, so simplicity is a virtue here.
6088 // The repne_scan instruction uses fixed registers, which we must spill.
6089 // Don't worry too much about pre-existing connections with the input regs.
6090
6091 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super)
6092 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter)
6093
6094 // Get super_klass value into rax (even if it was in rdi or rcx).
6095 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false;
6096 if (super_klass != rax || UseCompressedOops) {
6097 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; }
6098 mov(rax, super_klass);
6099 }
6100 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; }
6101 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; }
6102
6103 #ifndef PRODUCT
6104 int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
6105 ExternalAddress pst_counter_addr((address) pst_counter);
6106 NOT_LP64( incrementl(pst_counter_addr) );
6107 LP64_ONLY( lea(rcx, pst_counter_addr) );
6108 LP64_ONLY( incrementl(Address(rcx, 0)) );
6109 #endif //PRODUCT
6110
6111 // We will consult the secondary-super array.
6112 movptr(rdi, secondary_supers_addr);
6113 // Load the array length. (Positive movl does right thing on LP64.)
6114 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes()));
6115 // Skip to start of data.
6116 addptr(rdi, Array<Klass*>::base_offset_in_bytes());
6117
6118 // Scan RCX words at [RDI] for an occurrence of RAX.
6119 // Set NZ/Z based on last compare.
6120 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does
6121 // not change flags (only scas instruction which is repeated sets flags).
6122 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found.
6123
6124 testptr(rax,rax); // Set Z = 0
6125 repne_scan();
6126
6127 // Unspill the temp. registers:
6128 if (pushed_rdi) pop(rdi);
6129 if (pushed_rcx) pop(rcx);
6130 if (pushed_rax) pop(rax);
6131
6132 if (set_cond_codes) {
6133 // Special hack for the AD files: rdi is guaranteed non-zero.
6134 assert(!pushed_rdi, "rdi must be left non-NULL");
6135 // Also, the condition codes are properly set Z/NZ on succeed/failure.
6136 }
6137
6138 if (L_failure == &L_fallthrough)
6139 jccb(Assembler::notEqual, *L_failure);
6140 else jcc(Assembler::notEqual, *L_failure);
6141
6142 // Success. Cache the super we found and proceed in triumph.
6143 movptr(super_cache_addr, super_klass);
6144
6145 if (L_success != &L_fallthrough) {
6146 jmp(*L_success);
6147 }
6148
6149 #undef IS_A_TEMP
6150
6151 bind(L_fallthrough);
6152 }
6153
6154
6155 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
6156 if (VM_Version::supports_cmov()) {
6157 cmovl(cc, dst, src);
6158 } else {
6159 Label L;
6160 jccb(negate_condition(cc), L);
6161 movl(dst, src);
6162 bind(L);
6163 }
6164 }
6165
6166 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) {
6167 if (VM_Version::supports_cmov()) {
6168 cmovl(cc, dst, src);
6169 } else {
6170 Label L;
6171 jccb(negate_condition(cc), L);
6172 movl(dst, src);
6173 bind(L);
6174 }
6175 }
6176
6177 void MacroAssembler::verify_oop(Register reg, const char* s) {
6178 if (!VerifyOops) return;
6179
6180 // Pass register number to verify_oop_subroutine
6181 const char* b = NULL;
6182 {
6183 ResourceMark rm;
6184 stringStream ss;
6185 ss.print("verify_oop: %s: %s", reg->name(), s);
6186 b = code_string(ss.as_string());
6187 }
6188 BLOCK_COMMENT("verify_oop {");
6189 #ifdef _LP64
6190 push(rscratch1); // save r10, trashed by movptr()
6191 #endif
6192 push(rax); // save rax,
6193 push(reg); // pass register argument
6194 ExternalAddress buffer((address) b);
6195 // avoid using pushptr, as it modifies scratch registers
6196 // and our contract is not to modify anything
6197 movptr(rax, buffer.addr());
6198 push(rax);
6199 // call indirectly to solve generation ordering problem
6200 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
6201 call(rax);
6202 // Caller pops the arguments (oop, message) and restores rax, r10
6203 BLOCK_COMMENT("} verify_oop");
6204 }
6205
6206
6207 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
6208 Register tmp,
6209 int offset) {
6210 intptr_t value = *delayed_value_addr;
6211 if (value != 0)
6212 return RegisterOrConstant(value + offset);
6213
6214 // load indirectly to solve generation ordering problem
6215 movptr(tmp, ExternalAddress((address) delayed_value_addr));
6216
6217 #ifdef ASSERT
6218 { Label L;
6219 testptr(tmp, tmp);
6220 if (WizardMode) {
6221 const char* buf = NULL;
6222 {
6223 ResourceMark rm;
6224 stringStream ss;
6225 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]);
6226 buf = code_string(ss.as_string());
6227 }
6228 jcc(Assembler::notZero, L);
6229 STOP(buf);
6230 } else {
6231 jccb(Assembler::notZero, L);
6232 hlt();
6233 }
6234 bind(L);
6235 }
6236 #endif
6237
6238 if (offset != 0)
6239 addptr(tmp, offset);
6240
6241 return RegisterOrConstant(tmp);
6242 }
6243
6244
6245 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
6246 int extra_slot_offset) {
6247 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
6248 int stackElementSize = Interpreter::stackElementSize;
6249 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
6250 #ifdef ASSERT
6251 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
6252 assert(offset1 - offset == stackElementSize, "correct arithmetic");
6253 #endif
6254 Register scale_reg = noreg;
6255 Address::ScaleFactor scale_factor = Address::no_scale;
6256 if (arg_slot.is_constant()) {
6257 offset += arg_slot.as_constant() * stackElementSize;
6258 } else {
6259 scale_reg = arg_slot.as_register();
6260 scale_factor = Address::times(stackElementSize);
6261 }
6262 offset += wordSize; // return PC is on stack
6263 return Address(rsp, scale_reg, scale_factor, offset);
6264 }
6265
6266
6267 void MacroAssembler::verify_oop_addr(Address addr, const char* s) {
6268 if (!VerifyOops) return;
6269
6270 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord);
6271 // Pass register number to verify_oop_subroutine
6272 const char* b = NULL;
6273 {
6274 ResourceMark rm;
6275 stringStream ss;
6276 ss.print("verify_oop_addr: %s", s);
6277 b = code_string(ss.as_string());
6278 }
6279 #ifdef _LP64
6280 push(rscratch1); // save r10, trashed by movptr()
6281 #endif
6282 push(rax); // save rax,
6283 // addr may contain rsp so we will have to adjust it based on the push
6284 // we just did (and on 64 bit we do two pushes)
6285 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which
6286 // stores rax into addr which is backwards of what was intended.
6287 if (addr.uses(rsp)) {
6288 lea(rax, addr);
6289 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord));
6290 } else {
6291 pushptr(addr);
6292 }
6293
6294 ExternalAddress buffer((address) b);
6295 // pass msg argument
6296 // avoid using pushptr, as it modifies scratch registers
6297 // and our contract is not to modify anything
6298 movptr(rax, buffer.addr());
6299 push(rax);
6300
6301 // call indirectly to solve generation ordering problem
6302 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
6303 call(rax);
6304 // Caller pops the arguments (addr, message) and restores rax, r10.
6305 }
6306
6307 void MacroAssembler::verify_tlab() {
6308 #ifdef ASSERT
6309 if (UseTLAB && VerifyOops) {
6310 Label next, ok;
6311 Register t1 = rsi;
6312 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread);
6313
6314 push(t1);
6315 NOT_LP64(push(thread_reg));
6316 NOT_LP64(get_thread(thread_reg));
6317
6318 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
6319 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
6320 jcc(Assembler::aboveEqual, next);
6321 STOP("assert(top >= start)");
6322 should_not_reach_here();
6323
6324 bind(next);
6325 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
6326 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
6327 jcc(Assembler::aboveEqual, ok);
6328 STOP("assert(top <= end)");
6329 should_not_reach_here();
6330
6331 bind(ok);
6332 NOT_LP64(pop(thread_reg));
6333 pop(t1);
6334 }
6335 #endif
6336 }
6337
6338 class ControlWord {
6339 public:
6340 int32_t _value;
6341
6342 int rounding_control() const { return (_value >> 10) & 3 ; }
6343 int precision_control() const { return (_value >> 8) & 3 ; }
6344 bool precision() const { return ((_value >> 5) & 1) != 0; }
6345 bool underflow() const { return ((_value >> 4) & 1) != 0; }
6346 bool overflow() const { return ((_value >> 3) & 1) != 0; }
6347 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
6348 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
6349 bool invalid() const { return ((_value >> 0) & 1) != 0; }
6350
6351 void print() const {
6352 // rounding control
6353 const char* rc;
6354 switch (rounding_control()) {
6355 case 0: rc = "round near"; break;
6356 case 1: rc = "round down"; break;
6357 case 2: rc = "round up "; break;
6358 case 3: rc = "chop "; break;
6359 };
6360 // precision control
6361 const char* pc;
6362 switch (precision_control()) {
6363 case 0: pc = "24 bits "; break;
6364 case 1: pc = "reserved"; break;
6365 case 2: pc = "53 bits "; break;
6366 case 3: pc = "64 bits "; break;
6367 };
6368 // flags
6369 char f[9];
6370 f[0] = ' ';
6371 f[1] = ' ';
6372 f[2] = (precision ()) ? 'P' : 'p';
6373 f[3] = (underflow ()) ? 'U' : 'u';
6374 f[4] = (overflow ()) ? 'O' : 'o';
6375 f[5] = (zero_divide ()) ? 'Z' : 'z';
6376 f[6] = (denormalized()) ? 'D' : 'd';
6377 f[7] = (invalid ()) ? 'I' : 'i';
6378 f[8] = '\x0';
6379 // output
6380 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc);
6381 }
6382
6383 };
6384
6385 class StatusWord {
6386 public:
6387 int32_t _value;
6388
6389 bool busy() const { return ((_value >> 15) & 1) != 0; }
6390 bool C3() const { return ((_value >> 14) & 1) != 0; }
6391 bool C2() const { return ((_value >> 10) & 1) != 0; }
6392 bool C1() const { return ((_value >> 9) & 1) != 0; }
6393 bool C0() const { return ((_value >> 8) & 1) != 0; }
6394 int top() const { return (_value >> 11) & 7 ; }
6395 bool error_status() const { return ((_value >> 7) & 1) != 0; }
6396 bool stack_fault() const { return ((_value >> 6) & 1) != 0; }
6397 bool precision() const { return ((_value >> 5) & 1) != 0; }
6398 bool underflow() const { return ((_value >> 4) & 1) != 0; }
6399 bool overflow() const { return ((_value >> 3) & 1) != 0; }
6400 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
6401 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
6402 bool invalid() const { return ((_value >> 0) & 1) != 0; }
6403
6404 void print() const {
6405 // condition codes
6406 char c[5];
6407 c[0] = (C3()) ? '3' : '-';
6408 c[1] = (C2()) ? '2' : '-';
6409 c[2] = (C1()) ? '1' : '-';
6410 c[3] = (C0()) ? '0' : '-';
6411 c[4] = '\x0';
6412 // flags
6413 char f[9];
6414 f[0] = (error_status()) ? 'E' : '-';
6415 f[1] = (stack_fault ()) ? 'S' : '-';
6416 f[2] = (precision ()) ? 'P' : '-';
6417 f[3] = (underflow ()) ? 'U' : '-';
6418 f[4] = (overflow ()) ? 'O' : '-';
6419 f[5] = (zero_divide ()) ? 'Z' : '-';
6420 f[6] = (denormalized()) ? 'D' : '-';
6421 f[7] = (invalid ()) ? 'I' : '-';
6422 f[8] = '\x0';
6423 // output
6424 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top());
6425 }
6426
6427 };
6428
6429 class TagWord {
6430 public:
6431 int32_t _value;
6432
6433 int tag_at(int i) const { return (_value >> (i*2)) & 3; }
6434
6435 void print() const {
6436 printf("%04x", _value & 0xFFFF);
6437 }
6438
6439 };
6440
6441 class FPU_Register {
6442 public:
6443 int32_t _m0;
6444 int32_t _m1;
6445 int16_t _ex;
6446
6447 bool is_indefinite() const {
6448 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0;
6449 }
6450
6451 void print() const {
6452 char sign = (_ex < 0) ? '-' : '+';
6453 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " ";
6454 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind);
6455 };
6456
6457 };
6458
6459 class FPU_State {
6460 public:
6461 enum {
6462 register_size = 10,
6463 number_of_registers = 8,
6464 register_mask = 7
6465 };
6466
6467 ControlWord _control_word;
6468 StatusWord _status_word;
6469 TagWord _tag_word;
6470 int32_t _error_offset;
6471 int32_t _error_selector;
6472 int32_t _data_offset;
6473 int32_t _data_selector;
6474 int8_t _register[register_size * number_of_registers];
6475
6476 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); }
6477 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; }
6478
6479 const char* tag_as_string(int tag) const {
6480 switch (tag) {
6481 case 0: return "valid";
6482 case 1: return "zero";
6483 case 2: return "special";
6484 case 3: return "empty";
6485 }
6486 ShouldNotReachHere();
6487 return NULL;
6488 }
6489
6490 void print() const {
6491 // print computation registers
6492 { int t = _status_word.top();
6493 for (int i = 0; i < number_of_registers; i++) {
6494 int j = (i - t) & register_mask;
6495 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j);
6496 st(j)->print();
6497 printf(" %s\n", tag_as_string(_tag_word.tag_at(i)));
6498 }
6499 }
6500 printf("\n");
6501 // print control registers
6502 printf("ctrl = "); _control_word.print(); printf("\n");
6503 printf("stat = "); _status_word .print(); printf("\n");
6504 printf("tags = "); _tag_word .print(); printf("\n");
6505 }
6506
6507 };
6508
6509 class Flag_Register {
6510 public:
6511 int32_t _value;
6512
6513 bool overflow() const { return ((_value >> 11) & 1) != 0; }
6514 bool direction() const { return ((_value >> 10) & 1) != 0; }
6515 bool sign() const { return ((_value >> 7) & 1) != 0; }
6516 bool zero() const { return ((_value >> 6) & 1) != 0; }
6517 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; }
6518 bool parity() const { return ((_value >> 2) & 1) != 0; }
6519 bool carry() const { return ((_value >> 0) & 1) != 0; }
6520
6521 void print() const {
6522 // flags
6523 char f[8];
6524 f[0] = (overflow ()) ? 'O' : '-';
6525 f[1] = (direction ()) ? 'D' : '-';
6526 f[2] = (sign ()) ? 'S' : '-';
6527 f[3] = (zero ()) ? 'Z' : '-';
6528 f[4] = (auxiliary_carry()) ? 'A' : '-';
6529 f[5] = (parity ()) ? 'P' : '-';
6530 f[6] = (carry ()) ? 'C' : '-';
6531 f[7] = '\x0';
6532 // output
6533 printf("%08x flags = %s", _value, f);
6534 }
6535
6536 };
6537
6538 class IU_Register {
6539 public:
6540 int32_t _value;
6541
6542 void print() const {
6543 printf("%08x %11d", _value, _value);
6544 }
6545
6546 };
6547
6548 class IU_State {
6549 public:
6550 Flag_Register _eflags;
6551 IU_Register _rdi;
6552 IU_Register _rsi;
6553 IU_Register _rbp;
6554 IU_Register _rsp;
6555 IU_Register _rbx;
6556 IU_Register _rdx;
6557 IU_Register _rcx;
6558 IU_Register _rax;
6559
6560 void print() const {
6561 // computation registers
6562 printf("rax, = "); _rax.print(); printf("\n");
6563 printf("rbx, = "); _rbx.print(); printf("\n");
6564 printf("rcx = "); _rcx.print(); printf("\n");
6565 printf("rdx = "); _rdx.print(); printf("\n");
6566 printf("rdi = "); _rdi.print(); printf("\n");
6567 printf("rsi = "); _rsi.print(); printf("\n");
6568 printf("rbp, = "); _rbp.print(); printf("\n");
6569 printf("rsp = "); _rsp.print(); printf("\n");
6570 printf("\n");
6571 // control registers
6572 printf("flgs = "); _eflags.print(); printf("\n");
6573 }
6574 };
6575
6576
6577 class CPU_State {
6578 public:
6579 FPU_State _fpu_state;
6580 IU_State _iu_state;
6581
6582 void print() const {
6583 printf("--------------------------------------------------\n");
6584 _iu_state .print();
6585 printf("\n");
6586 _fpu_state.print();
6587 printf("--------------------------------------------------\n");
6588 }
6589
6590 };
6591
6592
6593 static void _print_CPU_state(CPU_State* state) {
6594 state->print();
6595 };
6596
6597
6598 void MacroAssembler::print_CPU_state() {
6599 push_CPU_state();
6600 push(rsp); // pass CPU state
6601 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state)));
6602 addptr(rsp, wordSize); // discard argument
6603 pop_CPU_state();
6604 }
6605
6606
6607 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) {
6608 static int counter = 0;
6609 FPU_State* fs = &state->_fpu_state;
6610 counter++;
6611 // For leaf calls, only verify that the top few elements remain empty.
6612 // We only need 1 empty at the top for C2 code.
6613 if( stack_depth < 0 ) {
6614 if( fs->tag_for_st(7) != 3 ) {
6615 printf("FPR7 not empty\n");
6616 state->print();
6617 assert(false, "error");
6618 return false;
6619 }
6620 return true; // All other stack states do not matter
6621 }
6622
6623 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std,
6624 "bad FPU control word");
6625
6626 // compute stack depth
6627 int i = 0;
6628 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++;
6629 int d = i;
6630 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++;
6631 // verify findings
6632 if (i != FPU_State::number_of_registers) {
6633 // stack not contiguous
6634 printf("%s: stack not contiguous at ST%d\n", s, i);
6635 state->print();
6636 assert(false, "error");
6637 return false;
6638 }
6639 // check if computed stack depth corresponds to expected stack depth
6640 if (stack_depth < 0) {
6641 // expected stack depth is -stack_depth or less
6642 if (d > -stack_depth) {
6643 // too many elements on the stack
6644 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d);
6645 state->print();
6646 assert(false, "error");
6647 return false;
6648 }
6649 } else {
6650 // expected stack depth is stack_depth
6651 if (d != stack_depth) {
6652 // wrong stack depth
6653 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d);
6654 state->print();
6655 assert(false, "error");
6656 return false;
6657 }
6658 }
6659 // everything is cool
6660 return true;
6661 }
6662
6663
6664 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
6665 if (!VerifyFPU) return;
6666 push_CPU_state();
6667 push(rsp); // pass CPU state
6668 ExternalAddress msg((address) s);
6669 // pass message string s
6670 pushptr(msg.addr());
6671 push(stack_depth); // pass stack depth
6672 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU)));
6673 addptr(rsp, 3 * wordSize); // discard arguments
6674 // check for error
6675 { Label L;
6676 testl(rax, rax);
6677 jcc(Assembler::notZero, L);
6678 int3(); // break if error condition
6679 bind(L);
6680 }
6681 pop_CPU_state();
6682 }
6683
6684 void MacroAssembler::restore_cpu_control_state_after_jni() {
6685 // Either restore the MXCSR register after returning from the JNI Call
6686 // or verify that it wasn't changed (with -Xcheck:jni flag).
6687 if (VM_Version::supports_sse()) {
6688 if (RestoreMXCSROnJNICalls) {
6689 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
6690 } else if (CheckJNICalls) {
6691 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
6692 }
6693 }
6694 if (VM_Version::supports_avx()) {
6695 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty.
6696 vzeroupper();
6697 }
6698
6699 #ifndef _LP64
6700 // Either restore the x87 floating pointer control word after returning
6701 // from the JNI call or verify that it wasn't changed.
6702 if (CheckJNICalls) {
6703 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
6704 }
6705 #endif // _LP64
6706 }
6707
6708
6709 void MacroAssembler::load_klass(Register dst, Register src) {
6710 #ifdef _LP64
6711 if (UseCompressedClassPointers) {
6712 movl(dst, Address(src, oopDesc::klass_offset_in_bytes()));
6713 decode_klass_not_null(dst);
6714 } else
6715 #endif
6716 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
6717 }
6718
6719 void MacroAssembler::load_prototype_header(Register dst, Register src) {
6720 load_klass(dst, src);
6721 movptr(dst, Address(dst, Klass::prototype_header_offset()));
6722 }
6723
6724 void MacroAssembler::store_klass(Register dst, Register src) {
6725 #ifdef _LP64
6726 if (UseCompressedClassPointers) {
6727 encode_klass_not_null(src);
6728 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src);
6729 } else
6730 #endif
6731 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src);
6732 }
6733
6734 void MacroAssembler::load_heap_oop(Register dst, Address src) {
6735 #ifdef _LP64
6736 // FIXME: Must change all places where we try to load the klass.
6737 if (UseCompressedOops) {
6738 movl(dst, src);
6739 decode_heap_oop(dst);
6740 } else
6741 #endif
6742 movptr(dst, src);
6743 }
6744
6745 // Doesn't do verfication, generates fixed size code
6746 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) {
6747 #ifdef _LP64
6748 if (UseCompressedOops) {
6749 movl(dst, src);
6750 decode_heap_oop_not_null(dst);
6751 } else
6752 #endif
6753 movptr(dst, src);
6754 }
6755
6756 void MacroAssembler::store_heap_oop(Address dst, Register src) {
6757 #ifdef _LP64
6758 if (UseCompressedOops) {
6759 assert(!dst.uses(src), "not enough registers");
6760 encode_heap_oop(src);
6761 movl(dst, src);
6762 } else
6763 #endif
6764 movptr(dst, src);
6765 }
6766
6767 void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) {
6768 assert_different_registers(src1, tmp);
6769 #ifdef _LP64
6770 if (UseCompressedOops) {
6771 bool did_push = false;
6772 if (tmp == noreg) {
6773 tmp = rax;
6774 push(tmp);
6775 did_push = true;
6776 assert(!src2.uses(rsp), "can't push");
6777 }
6778 load_heap_oop(tmp, src2);
6779 cmpptr(src1, tmp);
6780 if (did_push) pop(tmp);
6781 } else
6782 #endif
6783 cmpptr(src1, src2);
6784 }
6785
6786 // Used for storing NULLs.
6787 void MacroAssembler::store_heap_oop_null(Address dst) {
6788 #ifdef _LP64
6789 if (UseCompressedOops) {
6790 movl(dst, (int32_t)NULL_WORD);
6791 } else {
6792 movslq(dst, (int32_t)NULL_WORD);
6793 }
6794 #else
6795 movl(dst, (int32_t)NULL_WORD);
6796 #endif
6797 }
6798
6799 #ifdef _LP64
6800 void MacroAssembler::store_klass_gap(Register dst, Register src) {
6801 if (UseCompressedClassPointers) {
6802 // Store to klass gap in destination
6803 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src);
6804 }
6805 }
6806
6807 #ifdef ASSERT
6808 void MacroAssembler::verify_heapbase(const char* msg) {
6809 assert (UseCompressedOops, "should be compressed");
6810 assert (Universe::heap() != NULL, "java heap should be initialized");
6811 if (CheckCompressedOops) {
6812 Label ok;
6813 push(rscratch1); // cmpptr trashes rscratch1
6814 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
6815 jcc(Assembler::equal, ok);
6816 STOP(msg);
6817 bind(ok);
6818 pop(rscratch1);
6819 }
6820 }
6821 #endif
6822
6823 // Algorithm must match oop.inline.hpp encode_heap_oop.
6824 void MacroAssembler::encode_heap_oop(Register r) {
6825 #ifdef ASSERT
6826 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
6827 #endif
6828 verify_oop(r, "broken oop in encode_heap_oop");
6829 if (Universe::narrow_oop_base() == NULL) {
6830 if (Universe::narrow_oop_shift() != 0) {
6831 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6832 shrq(r, LogMinObjAlignmentInBytes);
6833 }
6834 return;
6835 }
6836 testq(r, r);
6837 cmovq(Assembler::equal, r, r12_heapbase);
6838 subq(r, r12_heapbase);
6839 shrq(r, LogMinObjAlignmentInBytes);
6840 }
6841
6842 void MacroAssembler::encode_heap_oop_not_null(Register r) {
6843 #ifdef ASSERT
6844 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
6845 if (CheckCompressedOops) {
6846 Label ok;
6847 testq(r, r);
6848 jcc(Assembler::notEqual, ok);
6849 STOP("null oop passed to encode_heap_oop_not_null");
6850 bind(ok);
6851 }
6852 #endif
6853 verify_oop(r, "broken oop in encode_heap_oop_not_null");
6854 if (Universe::narrow_oop_base() != NULL) {
6855 subq(r, r12_heapbase);
6856 }
6857 if (Universe::narrow_oop_shift() != 0) {
6858 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6859 shrq(r, LogMinObjAlignmentInBytes);
6860 }
6861 }
6862
6863 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
6864 #ifdef ASSERT
6865 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
6866 if (CheckCompressedOops) {
6867 Label ok;
6868 testq(src, src);
6869 jcc(Assembler::notEqual, ok);
6870 STOP("null oop passed to encode_heap_oop_not_null2");
6871 bind(ok);
6872 }
6873 #endif
6874 verify_oop(src, "broken oop in encode_heap_oop_not_null2");
6875 if (dst != src) {
6876 movq(dst, src);
6877 }
6878 if (Universe::narrow_oop_base() != NULL) {
6879 subq(dst, r12_heapbase);
6880 }
6881 if (Universe::narrow_oop_shift() != 0) {
6882 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6883 shrq(dst, LogMinObjAlignmentInBytes);
6884 }
6885 }
6886
6887 void MacroAssembler::decode_heap_oop(Register r) {
6888 #ifdef ASSERT
6889 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
6890 #endif
6891 if (Universe::narrow_oop_base() == NULL) {
6892 if (Universe::narrow_oop_shift() != 0) {
6893 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6894 shlq(r, LogMinObjAlignmentInBytes);
6895 }
6896 } else {
6897 Label done;
6898 shlq(r, LogMinObjAlignmentInBytes);
6899 jccb(Assembler::equal, done);
6900 addq(r, r12_heapbase);
6901 bind(done);
6902 }
6903 verify_oop(r, "broken oop in decode_heap_oop");
6904 }
6905
6906 void MacroAssembler::decode_heap_oop_not_null(Register r) {
6907 // Note: it will change flags
6908 assert (UseCompressedOops, "should only be used for compressed headers");
6909 assert (Universe::heap() != NULL, "java heap should be initialized");
6910 // Cannot assert, unverified entry point counts instructions (see .ad file)
6911 // vtableStubs also counts instructions in pd_code_size_limit.
6912 // Also do not verify_oop as this is called by verify_oop.
6913 if (Universe::narrow_oop_shift() != 0) {
6914 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6915 shlq(r, LogMinObjAlignmentInBytes);
6916 if (Universe::narrow_oop_base() != NULL) {
6917 addq(r, r12_heapbase);
6918 }
6919 } else {
6920 assert (Universe::narrow_oop_base() == NULL, "sanity");
6921 }
6922 }
6923
6924 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
6925 // Note: it will change flags
6926 assert (UseCompressedOops, "should only be used for compressed headers");
6927 assert (Universe::heap() != NULL, "java heap should be initialized");
6928 // Cannot assert, unverified entry point counts instructions (see .ad file)
6929 // vtableStubs also counts instructions in pd_code_size_limit.
6930 // Also do not verify_oop as this is called by verify_oop.
6931 if (Universe::narrow_oop_shift() != 0) {
6932 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6933 if (LogMinObjAlignmentInBytes == Address::times_8) {
6934 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0));
6935 } else {
6936 if (dst != src) {
6937 movq(dst, src);
6938 }
6939 shlq(dst, LogMinObjAlignmentInBytes);
6940 if (Universe::narrow_oop_base() != NULL) {
6941 addq(dst, r12_heapbase);
6942 }
6943 }
6944 } else {
6945 assert (Universe::narrow_oop_base() == NULL, "sanity");
6946 if (dst != src) {
6947 movq(dst, src);
6948 }
6949 }
6950 }
6951
6952 void MacroAssembler::encode_klass_not_null(Register r) {
6953 if (Universe::narrow_klass_base() != NULL) {
6954 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
6955 assert(r != r12_heapbase, "Encoding a klass in r12");
6956 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
6957 subq(r, r12_heapbase);
6958 }
6959 if (Universe::narrow_klass_shift() != 0) {
6960 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
6961 shrq(r, LogKlassAlignmentInBytes);
6962 }
6963 if (Universe::narrow_klass_base() != NULL) {
6964 reinit_heapbase();
6965 }
6966 }
6967
6968 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
6969 if (dst == src) {
6970 encode_klass_not_null(src);
6971 } else {
6972 if (Universe::narrow_klass_base() != NULL) {
6973 mov64(dst, (int64_t)Universe::narrow_klass_base());
6974 negq(dst);
6975 addq(dst, src);
6976 } else {
6977 movptr(dst, src);
6978 }
6979 if (Universe::narrow_klass_shift() != 0) {
6980 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
6981 shrq(dst, LogKlassAlignmentInBytes);
6982 }
6983 }
6984 }
6985
6986 // Function instr_size_for_decode_klass_not_null() counts the instructions
6987 // generated by decode_klass_not_null(register r) and reinit_heapbase(),
6988 // when (Universe::heap() != NULL). Hence, if the instructions they
6989 // generate change, then this method needs to be updated.
6990 int MacroAssembler::instr_size_for_decode_klass_not_null() {
6991 assert (UseCompressedClassPointers, "only for compressed klass ptrs");
6992 if (Universe::narrow_klass_base() != NULL) {
6993 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()).
6994 return (Universe::narrow_klass_shift() == 0 ? 20 : 24);
6995 } else {
6996 // longest load decode klass function, mov64, leaq
6997 return 16;
6998 }
6999 }
7000
7001 // !!! If the instructions that get generated here change then function
7002 // instr_size_for_decode_klass_not_null() needs to get updated.
7003 void MacroAssembler::decode_klass_not_null(Register r) {
7004 // Note: it will change flags
7005 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7006 assert(r != r12_heapbase, "Decoding a klass in r12");
7007 // Cannot assert, unverified entry point counts instructions (see .ad file)
7008 // vtableStubs also counts instructions in pd_code_size_limit.
7009 // Also do not verify_oop as this is called by verify_oop.
7010 if (Universe::narrow_klass_shift() != 0) {
7011 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
7012 shlq(r, LogKlassAlignmentInBytes);
7013 }
7014 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
7015 if (Universe::narrow_klass_base() != NULL) {
7016 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
7017 addq(r, r12_heapbase);
7018 reinit_heapbase();
7019 }
7020 }
7021
7022 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
7023 // Note: it will change flags
7024 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7025 if (dst == src) {
7026 decode_klass_not_null(dst);
7027 } else {
7028 // Cannot assert, unverified entry point counts instructions (see .ad file)
7029 // vtableStubs also counts instructions in pd_code_size_limit.
7030 // Also do not verify_oop as this is called by verify_oop.
7031 mov64(dst, (int64_t)Universe::narrow_klass_base());
7032 if (Universe::narrow_klass_shift() != 0) {
7033 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
7034 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?");
7035 leaq(dst, Address(dst, src, Address::times_8, 0));
7036 } else {
7037 addq(dst, src);
7038 }
7039 }
7040 }
7041
7042 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
7043 assert (UseCompressedOops, "should only be used for compressed headers");
7044 assert (Universe::heap() != NULL, "java heap should be initialized");
7045 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7046 int oop_index = oop_recorder()->find_index(obj);
7047 RelocationHolder rspec = oop_Relocation::spec(oop_index);
7048 mov_narrow_oop(dst, oop_index, rspec);
7049 }
7050
7051 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) {
7052 assert (UseCompressedOops, "should only be used for compressed headers");
7053 assert (Universe::heap() != NULL, "java heap should be initialized");
7054 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7055 int oop_index = oop_recorder()->find_index(obj);
7056 RelocationHolder rspec = oop_Relocation::spec(oop_index);
7057 mov_narrow_oop(dst, oop_index, rspec);
7058 }
7059
7060 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
7061 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7062 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7063 int klass_index = oop_recorder()->find_index(k);
7064 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
7065 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
7066 }
7067
7068 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) {
7069 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7070 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7071 int klass_index = oop_recorder()->find_index(k);
7072 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
7073 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
7074 }
7075
7076 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) {
7077 assert (UseCompressedOops, "should only be used for compressed headers");
7078 assert (Universe::heap() != NULL, "java heap should be initialized");
7079 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7080 int oop_index = oop_recorder()->find_index(obj);
7081 RelocationHolder rspec = oop_Relocation::spec(oop_index);
7082 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
7083 }
7084
7085 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) {
7086 assert (UseCompressedOops, "should only be used for compressed headers");
7087 assert (Universe::heap() != NULL, "java heap should be initialized");
7088 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7089 int oop_index = oop_recorder()->find_index(obj);
7090 RelocationHolder rspec = oop_Relocation::spec(oop_index);
7091 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
7092 }
7093
7094 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) {
7095 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7096 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7097 int klass_index = oop_recorder()->find_index(k);
7098 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
7099 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
7100 }
7101
7102 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) {
7103 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7104 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7105 int klass_index = oop_recorder()->find_index(k);
7106 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
7107 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
7108 }
7109
7110 void MacroAssembler::reinit_heapbase() {
7111 if (UseCompressedOops || UseCompressedClassPointers) {
7112 if (Universe::heap() != NULL) {
7113 if (Universe::narrow_oop_base() == NULL) {
7114 MacroAssembler::xorptr(r12_heapbase, r12_heapbase);
7115 } else {
7116 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base());
7117 }
7118 } else {
7119 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
7120 }
7121 }
7122 }
7123
7124 #endif // _LP64
7125
7126
7127 // C2 compiled method's prolog code.
7128 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) {
7129
7130 // WARNING: Initial instruction MUST be 5 bytes or longer so that
7131 // NativeJump::patch_verified_entry will be able to patch out the entry
7132 // code safely. The push to verify stack depth is ok at 5 bytes,
7133 // the frame allocation can be either 3 or 6 bytes. So if we don't do
7134 // stack bang then we must use the 6 byte frame allocation even if
7135 // we have no frame. :-(
7136 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect");
7137
7138 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
7139 // Remove word for return addr
7140 framesize -= wordSize;
7141 stack_bang_size -= wordSize;
7142
7143 // Calls to C2R adapters often do not accept exceptional returns.
7144 // We require that their callers must bang for them. But be careful, because
7145 // some VM calls (such as call site linkage) can use several kilobytes of
7146 // stack. But the stack safety zone should account for that.
7147 // See bugs 4446381, 4468289, 4497237.
7148 if (stack_bang_size > 0) {
7149 generate_stack_overflow_check(stack_bang_size);
7150
7151 // We always push rbp, so that on return to interpreter rbp, will be
7152 // restored correctly and we can correct the stack.
7153 push(rbp);
7154 // Save caller's stack pointer into RBP if the frame pointer is preserved.
7155 if (PreserveFramePointer) {
7156 mov(rbp, rsp);
7157 }
7158 // Remove word for ebp
7159 framesize -= wordSize;
7160
7161 // Create frame
7162 if (framesize) {
7163 subptr(rsp, framesize);
7164 }
7165 } else {
7166 // Create frame (force generation of a 4 byte immediate value)
7167 subptr_imm32(rsp, framesize);
7168
7169 // Save RBP register now.
7170 framesize -= wordSize;
7171 movptr(Address(rsp, framesize), rbp);
7172 // Save caller's stack pointer into RBP if the frame pointer is preserved.
7173 if (PreserveFramePointer) {
7174 movptr(rbp, rsp);
7175 if (framesize > 0) {
7176 addptr(rbp, framesize);
7177 }
7178 }
7179 }
7180
7181 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
7182 framesize -= wordSize;
7183 movptr(Address(rsp, framesize), (int32_t)0xbadb100d);
7184 }
7185
7186 #ifndef _LP64
7187 // If method sets FPU control word do it now
7188 if (fp_mode_24b) {
7189 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
7190 }
7191 if (UseSSE >= 2 && VerifyFPU) {
7192 verify_FPU(0, "FPU stack must be clean on entry");
7193 }
7194 #endif
7195
7196 #ifdef ASSERT
7197 if (VerifyStackAtCalls) {
7198 Label L;
7199 push(rax);
7200 mov(rax, rsp);
7201 andptr(rax, StackAlignmentInBytes-1);
7202 cmpptr(rax, StackAlignmentInBytes-wordSize);
7203 pop(rax);
7204 jcc(Assembler::equal, L);
7205 STOP("Stack is not properly aligned!");
7206 bind(L);
7207 }
7208 #endif
7209
7210 }
7211
7212 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp, bool is_large) {
7213 // cnt - number of qwords (8-byte words).
7214 // base - start address, qword aligned.
7215 // is_large - if optimizers know cnt is larger than InitArrayShortSize
7216 assert(base==rdi, "base register must be edi for rep stos");
7217 assert(tmp==rax, "tmp register must be eax for rep stos");
7218 assert(cnt==rcx, "cnt register must be ecx for rep stos");
7219 assert(InitArrayShortSize % BytesPerLong == 0,
7220 "InitArrayShortSize should be the multiple of BytesPerLong");
7221
7222 Label DONE;
7223
7224 xorptr(tmp, tmp);
7225
7226 if (!is_large) {
7227 Label LOOP, LONG;
7228 cmpptr(cnt, InitArrayShortSize/BytesPerLong);
7229 jccb(Assembler::greater, LONG);
7230
7231 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
7232
7233 decrement(cnt);
7234 jccb(Assembler::negative, DONE); // Zero length
7235
7236 // Use individual pointer-sized stores for small counts:
7237 BIND(LOOP);
7238 movptr(Address(base, cnt, Address::times_ptr), tmp);
7239 decrement(cnt);
7240 jccb(Assembler::greaterEqual, LOOP);
7241 jmpb(DONE);
7242
7243 BIND(LONG);
7244 }
7245
7246 // Use longer rep-prefixed ops for non-small counts:
7247 if (UseFastStosb) {
7248 shlptr(cnt, 3); // convert to number of bytes
7249 rep_stosb();
7250 } else {
7251 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
7252 rep_stos();
7253 }
7254
7255 BIND(DONE);
7256 }
7257
7258 #ifdef COMPILER2
7259
7260 // IndexOf for constant substrings with size >= 8 chars
7261 // which don't need to be loaded through stack.
7262 void MacroAssembler::string_indexofC8(Register str1, Register str2,
7263 Register cnt1, Register cnt2,
7264 int int_cnt2, Register result,
7265 XMMRegister vec, Register tmp,
7266 int ae) {
7267 ShortBranchVerifier sbv(this);
7268 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
7269 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
7270 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
7271
7272 // This method uses the pcmpestri instruction with bound registers
7273 // inputs:
7274 // xmm - substring
7275 // rax - substring length (elements count)
7276 // mem - scanned string
7277 // rdx - string length (elements count)
7278 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
7279 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
7280 // outputs:
7281 // rcx - matched index in string
7282 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
7283 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
7284 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
7285 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
7286 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
7287
7288 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR,
7289 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR,
7290 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE;
7291
7292 // Note, inline_string_indexOf() generates checks:
7293 // if (substr.count > string.count) return -1;
7294 // if (substr.count == 0) return 0;
7295 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars");
7296
7297 // Load substring.
7298 if (ae == StrIntrinsicNode::UL) {
7299 pmovzxbw(vec, Address(str2, 0));
7300 } else {
7301 movdqu(vec, Address(str2, 0));
7302 }
7303 movl(cnt2, int_cnt2);
7304 movptr(result, str1); // string addr
7305
7306 if (int_cnt2 > stride) {
7307 jmpb(SCAN_TO_SUBSTR);
7308
7309 // Reload substr for rescan, this code
7310 // is executed only for large substrings (> 8 chars)
7311 bind(RELOAD_SUBSTR);
7312 if (ae == StrIntrinsicNode::UL) {
7313 pmovzxbw(vec, Address(str2, 0));
7314 } else {
7315 movdqu(vec, Address(str2, 0));
7316 }
7317 negptr(cnt2); // Jumped here with negative cnt2, convert to positive
7318
7319 bind(RELOAD_STR);
7320 // We came here after the beginning of the substring was
7321 // matched but the rest of it was not so we need to search
7322 // again. Start from the next element after the previous match.
7323
7324 // cnt2 is number of substring reminding elements and
7325 // cnt1 is number of string reminding elements when cmp failed.
7326 // Restored cnt1 = cnt1 - cnt2 + int_cnt2
7327 subl(cnt1, cnt2);
7328 addl(cnt1, int_cnt2);
7329 movl(cnt2, int_cnt2); // Now restore cnt2
7330
7331 decrementl(cnt1); // Shift to next element
7332 cmpl(cnt1, cnt2);
7333 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7334
7335 addptr(result, (1<<scale1));
7336
7337 } // (int_cnt2 > 8)
7338
7339 // Scan string for start of substr in 16-byte vectors
7340 bind(SCAN_TO_SUBSTR);
7341 pcmpestri(vec, Address(result, 0), mode);
7342 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
7343 subl(cnt1, stride);
7344 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
7345 cmpl(cnt1, cnt2);
7346 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7347 addptr(result, 16);
7348 jmpb(SCAN_TO_SUBSTR);
7349
7350 // Found a potential substr
7351 bind(FOUND_CANDIDATE);
7352 // Matched whole vector if first element matched (tmp(rcx) == 0).
7353 if (int_cnt2 == stride) {
7354 jccb(Assembler::overflow, RET_FOUND); // OF == 1
7355 } else { // int_cnt2 > 8
7356 jccb(Assembler::overflow, FOUND_SUBSTR);
7357 }
7358 // After pcmpestri tmp(rcx) contains matched element index
7359 // Compute start addr of substr
7360 lea(result, Address(result, tmp, scale1));
7361
7362 // Make sure string is still long enough
7363 subl(cnt1, tmp);
7364 cmpl(cnt1, cnt2);
7365 if (int_cnt2 == stride) {
7366 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
7367 } else { // int_cnt2 > 8
7368 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD);
7369 }
7370 // Left less then substring.
7371
7372 bind(RET_NOT_FOUND);
7373 movl(result, -1);
7374 jmpb(EXIT);
7375
7376 if (int_cnt2 > stride) {
7377 // This code is optimized for the case when whole substring
7378 // is matched if its head is matched.
7379 bind(MATCH_SUBSTR_HEAD);
7380 pcmpestri(vec, Address(result, 0), mode);
7381 // Reload only string if does not match
7382 jccb(Assembler::noOverflow, RELOAD_STR); // OF == 0
7383
7384 Label CONT_SCAN_SUBSTR;
7385 // Compare the rest of substring (> 8 chars).
7386 bind(FOUND_SUBSTR);
7387 // First 8 chars are already matched.
7388 negptr(cnt2);
7389 addptr(cnt2, stride);
7390
7391 bind(SCAN_SUBSTR);
7392 subl(cnt1, stride);
7393 cmpl(cnt2, -stride); // Do not read beyond substring
7394 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR);
7395 // Back-up strings to avoid reading beyond substring:
7396 // cnt1 = cnt1 - cnt2 + 8
7397 addl(cnt1, cnt2); // cnt2 is negative
7398 addl(cnt1, stride);
7399 movl(cnt2, stride); negptr(cnt2);
7400 bind(CONT_SCAN_SUBSTR);
7401 if (int_cnt2 < (int)G) {
7402 int tail_off1 = int_cnt2<<scale1;
7403 int tail_off2 = int_cnt2<<scale2;
7404 if (ae == StrIntrinsicNode::UL) {
7405 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2));
7406 } else {
7407 movdqu(vec, Address(str2, cnt2, scale2, tail_off2));
7408 }
7409 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode);
7410 } else {
7411 // calculate index in register to avoid integer overflow (int_cnt2*2)
7412 movl(tmp, int_cnt2);
7413 addptr(tmp, cnt2);
7414 if (ae == StrIntrinsicNode::UL) {
7415 pmovzxbw(vec, Address(str2, tmp, scale2, 0));
7416 } else {
7417 movdqu(vec, Address(str2, tmp, scale2, 0));
7418 }
7419 pcmpestri(vec, Address(result, tmp, scale1, 0), mode);
7420 }
7421 // Need to reload strings pointers if not matched whole vector
7422 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
7423 addptr(cnt2, stride);
7424 jcc(Assembler::negative, SCAN_SUBSTR);
7425 // Fall through if found full substring
7426
7427 } // (int_cnt2 > 8)
7428
7429 bind(RET_FOUND);
7430 // Found result if we matched full small substring.
7431 // Compute substr offset
7432 subptr(result, str1);
7433 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
7434 shrl(result, 1); // index
7435 }
7436 bind(EXIT);
7437
7438 } // string_indexofC8
7439
7440 // Small strings are loaded through stack if they cross page boundary.
7441 void MacroAssembler::string_indexof(Register str1, Register str2,
7442 Register cnt1, Register cnt2,
7443 int int_cnt2, Register result,
7444 XMMRegister vec, Register tmp,
7445 int ae) {
7446 ShortBranchVerifier sbv(this);
7447 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
7448 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
7449 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
7450
7451 //
7452 // int_cnt2 is length of small (< 8 chars) constant substring
7453 // or (-1) for non constant substring in which case its length
7454 // is in cnt2 register.
7455 //
7456 // Note, inline_string_indexOf() generates checks:
7457 // if (substr.count > string.count) return -1;
7458 // if (substr.count == 0) return 0;
7459 //
7460 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
7461 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0");
7462 // This method uses the pcmpestri instruction with bound registers
7463 // inputs:
7464 // xmm - substring
7465 // rax - substring length (elements count)
7466 // mem - scanned string
7467 // rdx - string length (elements count)
7468 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
7469 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
7470 // outputs:
7471 // rcx - matched index in string
7472 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
7473 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
7474 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
7475 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
7476
7477 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR,
7478 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR,
7479 FOUND_CANDIDATE;
7480
7481 { //========================================================
7482 // We don't know where these strings are located
7483 // and we can't read beyond them. Load them through stack.
7484 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR;
7485
7486 movptr(tmp, rsp); // save old SP
7487
7488 if (int_cnt2 > 0) { // small (< 8 chars) constant substring
7489 if (int_cnt2 == (1>>scale2)) { // One byte
7490 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding");
7491 load_unsigned_byte(result, Address(str2, 0));
7492 movdl(vec, result); // move 32 bits
7493 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes
7494 // Not enough header space in 32-bit VM: 12+3 = 15.
7495 movl(result, Address(str2, -1));
7496 shrl(result, 8);
7497 movdl(vec, result); // move 32 bits
7498 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char
7499 load_unsigned_short(result, Address(str2, 0));
7500 movdl(vec, result); // move 32 bits
7501 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars
7502 movdl(vec, Address(str2, 0)); // move 32 bits
7503 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars
7504 movq(vec, Address(str2, 0)); // move 64 bits
7505 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7})
7506 // Array header size is 12 bytes in 32-bit VM
7507 // + 6 bytes for 3 chars == 18 bytes,
7508 // enough space to load vec and shift.
7509 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity");
7510 if (ae == StrIntrinsicNode::UL) {
7511 int tail_off = int_cnt2-8;
7512 pmovzxbw(vec, Address(str2, tail_off));
7513 psrldq(vec, -2*tail_off);
7514 }
7515 else {
7516 int tail_off = int_cnt2*(1<<scale2);
7517 movdqu(vec, Address(str2, tail_off-16));
7518 psrldq(vec, 16-tail_off);
7519 }
7520 }
7521 } else { // not constant substring
7522 cmpl(cnt2, stride);
7523 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough
7524
7525 // We can read beyond string if srt+16 does not cross page boundary
7526 // since heaps are aligned and mapped by pages.
7527 assert(os::vm_page_size() < (int)G, "default page should be small");
7528 movl(result, str2); // We need only low 32 bits
7529 andl(result, (os::vm_page_size()-1));
7530 cmpl(result, (os::vm_page_size()-16));
7531 jccb(Assembler::belowEqual, CHECK_STR);
7532
7533 // Move small strings to stack to allow load 16 bytes into vec.
7534 subptr(rsp, 16);
7535 int stk_offset = wordSize-(1<<scale2);
7536 push(cnt2);
7537
7538 bind(COPY_SUBSTR);
7539 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) {
7540 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1));
7541 movb(Address(rsp, cnt2, scale2, stk_offset), result);
7542 } else if (ae == StrIntrinsicNode::UU) {
7543 load_unsigned_short(result, Address(str2, cnt2, scale2, -2));
7544 movw(Address(rsp, cnt2, scale2, stk_offset), result);
7545 }
7546 decrement(cnt2);
7547 jccb(Assembler::notZero, COPY_SUBSTR);
7548
7549 pop(cnt2);
7550 movptr(str2, rsp); // New substring address
7551 } // non constant
7552
7553 bind(CHECK_STR);
7554 cmpl(cnt1, stride);
7555 jccb(Assembler::aboveEqual, BIG_STRINGS);
7556
7557 // Check cross page boundary.
7558 movl(result, str1); // We need only low 32 bits
7559 andl(result, (os::vm_page_size()-1));
7560 cmpl(result, (os::vm_page_size()-16));
7561 jccb(Assembler::belowEqual, BIG_STRINGS);
7562
7563 subptr(rsp, 16);
7564 int stk_offset = -(1<<scale1);
7565 if (int_cnt2 < 0) { // not constant
7566 push(cnt2);
7567 stk_offset += wordSize;
7568 }
7569 movl(cnt2, cnt1);
7570
7571 bind(COPY_STR);
7572 if (ae == StrIntrinsicNode::LL) {
7573 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1));
7574 movb(Address(rsp, cnt2, scale1, stk_offset), result);
7575 } else {
7576 load_unsigned_short(result, Address(str1, cnt2, scale1, -2));
7577 movw(Address(rsp, cnt2, scale1, stk_offset), result);
7578 }
7579 decrement(cnt2);
7580 jccb(Assembler::notZero, COPY_STR);
7581
7582 if (int_cnt2 < 0) { // not constant
7583 pop(cnt2);
7584 }
7585 movptr(str1, rsp); // New string address
7586
7587 bind(BIG_STRINGS);
7588 // Load substring.
7589 if (int_cnt2 < 0) { // -1
7590 if (ae == StrIntrinsicNode::UL) {
7591 pmovzxbw(vec, Address(str2, 0));
7592 } else {
7593 movdqu(vec, Address(str2, 0));
7594 }
7595 push(cnt2); // substr count
7596 push(str2); // substr addr
7597 push(str1); // string addr
7598 } else {
7599 // Small (< 8 chars) constant substrings are loaded already.
7600 movl(cnt2, int_cnt2);
7601 }
7602 push(tmp); // original SP
7603
7604 } // Finished loading
7605
7606 //========================================================
7607 // Start search
7608 //
7609
7610 movptr(result, str1); // string addr
7611
7612 if (int_cnt2 < 0) { // Only for non constant substring
7613 jmpb(SCAN_TO_SUBSTR);
7614
7615 // SP saved at sp+0
7616 // String saved at sp+1*wordSize
7617 // Substr saved at sp+2*wordSize
7618 // Substr count saved at sp+3*wordSize
7619
7620 // Reload substr for rescan, this code
7621 // is executed only for large substrings (> 8 chars)
7622 bind(RELOAD_SUBSTR);
7623 movptr(str2, Address(rsp, 2*wordSize));
7624 movl(cnt2, Address(rsp, 3*wordSize));
7625 if (ae == StrIntrinsicNode::UL) {
7626 pmovzxbw(vec, Address(str2, 0));
7627 } else {
7628 movdqu(vec, Address(str2, 0));
7629 }
7630 // We came here after the beginning of the substring was
7631 // matched but the rest of it was not so we need to search
7632 // again. Start from the next element after the previous match.
7633 subptr(str1, result); // Restore counter
7634 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
7635 shrl(str1, 1);
7636 }
7637 addl(cnt1, str1);
7638 decrementl(cnt1); // Shift to next element
7639 cmpl(cnt1, cnt2);
7640 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7641
7642 addptr(result, (1<<scale1));
7643 } // non constant
7644
7645 // Scan string for start of substr in 16-byte vectors
7646 bind(SCAN_TO_SUBSTR);
7647 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
7648 pcmpestri(vec, Address(result, 0), mode);
7649 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
7650 subl(cnt1, stride);
7651 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
7652 cmpl(cnt1, cnt2);
7653 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7654 addptr(result, 16);
7655
7656 bind(ADJUST_STR);
7657 cmpl(cnt1, stride); // Do not read beyond string
7658 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
7659 // Back-up string to avoid reading beyond string.
7660 lea(result, Address(result, cnt1, scale1, -16));
7661 movl(cnt1, stride);
7662 jmpb(SCAN_TO_SUBSTR);
7663
7664 // Found a potential substr
7665 bind(FOUND_CANDIDATE);
7666 // After pcmpestri tmp(rcx) contains matched element index
7667
7668 // Make sure string is still long enough
7669 subl(cnt1, tmp);
7670 cmpl(cnt1, cnt2);
7671 jccb(Assembler::greaterEqual, FOUND_SUBSTR);
7672 // Left less then substring.
7673
7674 bind(RET_NOT_FOUND);
7675 movl(result, -1);
7676 jmpb(CLEANUP);
7677
7678 bind(FOUND_SUBSTR);
7679 // Compute start addr of substr
7680 lea(result, Address(result, tmp, scale1));
7681 if (int_cnt2 > 0) { // Constant substring
7682 // Repeat search for small substring (< 8 chars)
7683 // from new point without reloading substring.
7684 // Have to check that we don't read beyond string.
7685 cmpl(tmp, stride-int_cnt2);
7686 jccb(Assembler::greater, ADJUST_STR);
7687 // Fall through if matched whole substring.
7688 } else { // non constant
7689 assert(int_cnt2 == -1, "should be != 0");
7690
7691 addl(tmp, cnt2);
7692 // Found result if we matched whole substring.
7693 cmpl(tmp, stride);
7694 jccb(Assembler::lessEqual, RET_FOUND);
7695
7696 // Repeat search for small substring (<= 8 chars)
7697 // from new point 'str1' without reloading substring.
7698 cmpl(cnt2, stride);
7699 // Have to check that we don't read beyond string.
7700 jccb(Assembler::lessEqual, ADJUST_STR);
7701
7702 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG;
7703 // Compare the rest of substring (> 8 chars).
7704 movptr(str1, result);
7705
7706 cmpl(tmp, cnt2);
7707 // First 8 chars are already matched.
7708 jccb(Assembler::equal, CHECK_NEXT);
7709
7710 bind(SCAN_SUBSTR);
7711 pcmpestri(vec, Address(str1, 0), mode);
7712 // Need to reload strings pointers if not matched whole vector
7713 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
7714
7715 bind(CHECK_NEXT);
7716 subl(cnt2, stride);
7717 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring
7718 addptr(str1, 16);
7719 if (ae == StrIntrinsicNode::UL) {
7720 addptr(str2, 8);
7721 } else {
7722 addptr(str2, 16);
7723 }
7724 subl(cnt1, stride);
7725 cmpl(cnt2, stride); // Do not read beyond substring
7726 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR);
7727 // Back-up strings to avoid reading beyond substring.
7728
7729 if (ae == StrIntrinsicNode::UL) {
7730 lea(str2, Address(str2, cnt2, scale2, -8));
7731 lea(str1, Address(str1, cnt2, scale1, -16));
7732 } else {
7733 lea(str2, Address(str2, cnt2, scale2, -16));
7734 lea(str1, Address(str1, cnt2, scale1, -16));
7735 }
7736 subl(cnt1, cnt2);
7737 movl(cnt2, stride);
7738 addl(cnt1, stride);
7739 bind(CONT_SCAN_SUBSTR);
7740 if (ae == StrIntrinsicNode::UL) {
7741 pmovzxbw(vec, Address(str2, 0));
7742 } else {
7743 movdqu(vec, Address(str2, 0));
7744 }
7745 jmpb(SCAN_SUBSTR);
7746
7747 bind(RET_FOUND_LONG);
7748 movptr(str1, Address(rsp, wordSize));
7749 } // non constant
7750
7751 bind(RET_FOUND);
7752 // Compute substr offset
7753 subptr(result, str1);
7754 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
7755 shrl(result, 1); // index
7756 }
7757 bind(CLEANUP);
7758 pop(rsp); // restore SP
7759
7760 } // string_indexof
7761
7762 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result,
7763 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) {
7764 ShortBranchVerifier sbv(this);
7765 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
7766 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
7767
7768 int stride = 8;
7769
7770 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP,
7771 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP,
7772 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT,
7773 FOUND_SEQ_CHAR, DONE_LABEL;
7774
7775 movptr(result, str1);
7776 if (UseAVX >= 2) {
7777 cmpl(cnt1, stride);
7778 jccb(Assembler::less, SCAN_TO_CHAR_LOOP);
7779 cmpl(cnt1, 2*stride);
7780 jccb(Assembler::less, SCAN_TO_8_CHAR_INIT);
7781 movdl(vec1, ch);
7782 vpbroadcastw(vec1, vec1);
7783 vpxor(vec2, vec2);
7784 movl(tmp, cnt1);
7785 andl(tmp, 0xFFFFFFF0); //vector count (in chars)
7786 andl(cnt1,0x0000000F); //tail count (in chars)
7787
7788 bind(SCAN_TO_16_CHAR_LOOP);
7789 vmovdqu(vec3, Address(result, 0));
7790 vpcmpeqw(vec3, vec3, vec1, 1);
7791 vptest(vec2, vec3);
7792 jcc(Assembler::carryClear, FOUND_CHAR);
7793 addptr(result, 32);
7794 subl(tmp, 2*stride);
7795 jccb(Assembler::notZero, SCAN_TO_16_CHAR_LOOP);
7796 jmp(SCAN_TO_8_CHAR);
7797 bind(SCAN_TO_8_CHAR_INIT);
7798 movdl(vec1, ch);
7799 pshuflw(vec1, vec1, 0x00);
7800 pshufd(vec1, vec1, 0);
7801 pxor(vec2, vec2);
7802 }
7803 bind(SCAN_TO_8_CHAR);
7804 cmpl(cnt1, stride);
7805 if (UseAVX >= 2) {
7806 jccb(Assembler::less, SCAN_TO_CHAR);
7807 } else {
7808 jccb(Assembler::less, SCAN_TO_CHAR_LOOP);
7809 movdl(vec1, ch);
7810 pshuflw(vec1, vec1, 0x00);
7811 pshufd(vec1, vec1, 0);
7812 pxor(vec2, vec2);
7813 }
7814 movl(tmp, cnt1);
7815 andl(tmp, 0xFFFFFFF8); //vector count (in chars)
7816 andl(cnt1,0x00000007); //tail count (in chars)
7817
7818 bind(SCAN_TO_8_CHAR_LOOP);
7819 movdqu(vec3, Address(result, 0));
7820 pcmpeqw(vec3, vec1);
7821 ptest(vec2, vec3);
7822 jcc(Assembler::carryClear, FOUND_CHAR);
7823 addptr(result, 16);
7824 subl(tmp, stride);
7825 jccb(Assembler::notZero, SCAN_TO_8_CHAR_LOOP);
7826 bind(SCAN_TO_CHAR);
7827 testl(cnt1, cnt1);
7828 jcc(Assembler::zero, RET_NOT_FOUND);
7829 bind(SCAN_TO_CHAR_LOOP);
7830 load_unsigned_short(tmp, Address(result, 0));
7831 cmpl(ch, tmp);
7832 jccb(Assembler::equal, FOUND_SEQ_CHAR);
7833 addptr(result, 2);
7834 subl(cnt1, 1);
7835 jccb(Assembler::zero, RET_NOT_FOUND);
7836 jmp(SCAN_TO_CHAR_LOOP);
7837
7838 bind(RET_NOT_FOUND);
7839 movl(result, -1);
7840 jmpb(DONE_LABEL);
7841
7842 bind(FOUND_CHAR);
7843 if (UseAVX >= 2) {
7844 vpmovmskb(tmp, vec3);
7845 } else {
7846 pmovmskb(tmp, vec3);
7847 }
7848 bsfl(ch, tmp);
7849 addl(result, ch);
7850
7851 bind(FOUND_SEQ_CHAR);
7852 subptr(result, str1);
7853 shrl(result, 1);
7854
7855 bind(DONE_LABEL);
7856 } // string_indexof_char
7857
7858 // helper function for string_compare
7859 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2,
7860 Address::ScaleFactor scale, Address::ScaleFactor scale1,
7861 Address::ScaleFactor scale2, Register index, int ae) {
7862 if (ae == StrIntrinsicNode::LL) {
7863 load_unsigned_byte(elem1, Address(str1, index, scale, 0));
7864 load_unsigned_byte(elem2, Address(str2, index, scale, 0));
7865 } else if (ae == StrIntrinsicNode::UU) {
7866 load_unsigned_short(elem1, Address(str1, index, scale, 0));
7867 load_unsigned_short(elem2, Address(str2, index, scale, 0));
7868 } else {
7869 load_unsigned_byte(elem1, Address(str1, index, scale1, 0));
7870 load_unsigned_short(elem2, Address(str2, index, scale2, 0));
7871 }
7872 }
7873
7874 // Compare strings, used for char[] and byte[].
7875 void MacroAssembler::string_compare(Register str1, Register str2,
7876 Register cnt1, Register cnt2, Register result,
7877 XMMRegister vec1, int ae) {
7878 ShortBranchVerifier sbv(this);
7879 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
7880 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3
7881 int stride, stride2, adr_stride, adr_stride1, adr_stride2;
7882 int stride2x2 = 0x40;
7883 Address::ScaleFactor scale = Address::no_scale;
7884 Address::ScaleFactor scale1 = Address::no_scale;
7885 Address::ScaleFactor scale2 = Address::no_scale;
7886
7887 if (ae != StrIntrinsicNode::LL) {
7888 stride2x2 = 0x20;
7889 }
7890
7891 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
7892 shrl(cnt2, 1);
7893 }
7894 // Compute the minimum of the string lengths and the
7895 // difference of the string lengths (stack).
7896 // Do the conditional move stuff
7897 movl(result, cnt1);
7898 subl(cnt1, cnt2);
7899 push(cnt1);
7900 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2)
7901
7902 // Is the minimum length zero?
7903 testl(cnt2, cnt2);
7904 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
7905 if (ae == StrIntrinsicNode::LL) {
7906 // Load first bytes
7907 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0]
7908 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0]
7909 } else if (ae == StrIntrinsicNode::UU) {
7910 // Load first characters
7911 load_unsigned_short(result, Address(str1, 0));
7912 load_unsigned_short(cnt1, Address(str2, 0));
7913 } else {
7914 load_unsigned_byte(result, Address(str1, 0));
7915 load_unsigned_short(cnt1, Address(str2, 0));
7916 }
7917 subl(result, cnt1);
7918 jcc(Assembler::notZero, POP_LABEL);
7919
7920 if (ae == StrIntrinsicNode::UU) {
7921 // Divide length by 2 to get number of chars
7922 shrl(cnt2, 1);
7923 }
7924 cmpl(cnt2, 1);
7925 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
7926
7927 // Check if the strings start at the same location and setup scale and stride
7928 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7929 cmpptr(str1, str2);
7930 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
7931 if (ae == StrIntrinsicNode::LL) {
7932 scale = Address::times_1;
7933 stride = 16;
7934 } else {
7935 scale = Address::times_2;
7936 stride = 8;
7937 }
7938 } else {
7939 scale1 = Address::times_1;
7940 scale2 = Address::times_2;
7941 // scale not used
7942 stride = 8;
7943 }
7944
7945 if (UseAVX >= 2 && UseSSE42Intrinsics) {
7946 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
7947 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
7948 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
7949 Label COMPARE_WIDE_VECTORS_LOOP_AVX2;
7950 Label COMPARE_TAIL_LONG;
7951 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3
7952
7953 int pcmpmask = 0x19;
7954 if (ae == StrIntrinsicNode::LL) {
7955 pcmpmask &= ~0x01;
7956 }
7957
7958 // Setup to compare 16-chars (32-bytes) vectors,
7959 // start from first character again because it has aligned address.
7960 if (ae == StrIntrinsicNode::LL) {
7961 stride2 = 32;
7962 } else {
7963 stride2 = 16;
7964 }
7965 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7966 adr_stride = stride << scale;
7967 } else {
7968 adr_stride1 = 8; //stride << scale1;
7969 adr_stride2 = 16; //stride << scale2;
7970 }
7971
7972 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
7973 // rax and rdx are used by pcmpestri as elements counters
7974 movl(result, cnt2);
7975 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count
7976 jcc(Assembler::zero, COMPARE_TAIL_LONG);
7977
7978 // fast path : compare first 2 8-char vectors.
7979 bind(COMPARE_16_CHARS);
7980 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7981 movdqu(vec1, Address(str1, 0));
7982 } else {
7983 pmovzxbw(vec1, Address(str1, 0));
7984 }
7985 pcmpestri(vec1, Address(str2, 0), pcmpmask);
7986 jccb(Assembler::below, COMPARE_INDEX_CHAR);
7987
7988 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
7989 movdqu(vec1, Address(str1, adr_stride));
7990 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
7991 } else {
7992 pmovzxbw(vec1, Address(str1, adr_stride1));
7993 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask);
7994 }
7995 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
7996 addl(cnt1, stride);
7997
7998 // Compare the characters at index in cnt1
7999 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character
8000 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
8001 subl(result, cnt2);
8002 jmp(POP_LABEL);
8003
8004 // Setup the registers to start vector comparison loop
8005 bind(COMPARE_WIDE_VECTORS);
8006 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8007 lea(str1, Address(str1, result, scale));
8008 lea(str2, Address(str2, result, scale));
8009 } else {
8010 lea(str1, Address(str1, result, scale1));
8011 lea(str2, Address(str2, result, scale2));
8012 }
8013 subl(result, stride2);
8014 subl(cnt2, stride2);
8015 jcc(Assembler::zero, COMPARE_WIDE_TAIL);
8016 negptr(result);
8017
8018 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
8019 bind(COMPARE_WIDE_VECTORS_LOOP);
8020
8021 #ifdef _LP64
8022 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
8023 cmpl(cnt2, stride2x2);
8024 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2);
8025 testl(cnt2, stride2x2-1); // cnt2 holds the vector count
8026 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40
8027
8028 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
8029 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8030 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit);
8031 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
8032 } else {
8033 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit);
8034 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
8035 }
8036 kortestql(k7, k7);
8037 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare
8038 addptr(result, stride2x2); // update since we already compared at this addr
8039 subl(cnt2, stride2x2); // and sub the size too
8040 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3);
8041
8042 vpxor(vec1, vec1);
8043 jmpb(COMPARE_WIDE_TAIL);
8044 }//if (VM_Version::supports_avx512vlbw())
8045 #endif // _LP64
8046
8047
8048 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
8049 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8050 vmovdqu(vec1, Address(str1, result, scale));
8051 vpxor(vec1, Address(str2, result, scale));
8052 } else {
8053 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit);
8054 vpxor(vec1, Address(str2, result, scale2));
8055 }
8056 vptest(vec1, vec1);
8057 jcc(Assembler::notZero, VECTOR_NOT_EQUAL);
8058 addptr(result, stride2);
8059 subl(cnt2, stride2);
8060 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
8061 // clean upper bits of YMM registers
8062 vpxor(vec1, vec1);
8063
8064 // compare wide vectors tail
8065 bind(COMPARE_WIDE_TAIL);
8066 testptr(result, result);
8067 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
8068
8069 movl(result, stride2);
8070 movl(cnt2, result);
8071 negptr(result);
8072 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2);
8073
8074 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
8075 bind(VECTOR_NOT_EQUAL);
8076 // clean upper bits of YMM registers
8077 vpxor(vec1, vec1);
8078 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8079 lea(str1, Address(str1, result, scale));
8080 lea(str2, Address(str2, result, scale));
8081 } else {
8082 lea(str1, Address(str1, result, scale1));
8083 lea(str2, Address(str2, result, scale2));
8084 }
8085 jmp(COMPARE_16_CHARS);
8086
8087 // Compare tail chars, length between 1 to 15 chars
8088 bind(COMPARE_TAIL_LONG);
8089 movl(cnt2, result);
8090 cmpl(cnt2, stride);
8091 jccb(Assembler::less, COMPARE_SMALL_STR);
8092
8093 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8094 movdqu(vec1, Address(str1, 0));
8095 } else {
8096 pmovzxbw(vec1, Address(str1, 0));
8097 }
8098 pcmpestri(vec1, Address(str2, 0), pcmpmask);
8099 jcc(Assembler::below, COMPARE_INDEX_CHAR);
8100 subptr(cnt2, stride);
8101 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
8102 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8103 lea(str1, Address(str1, result, scale));
8104 lea(str2, Address(str2, result, scale));
8105 } else {
8106 lea(str1, Address(str1, result, scale1));
8107 lea(str2, Address(str2, result, scale2));
8108 }
8109 negptr(cnt2);
8110 jmpb(WHILE_HEAD_LABEL);
8111
8112 bind(COMPARE_SMALL_STR);
8113 } else if (UseSSE42Intrinsics) {
8114 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
8115 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
8116 int pcmpmask = 0x19;
8117 // Setup to compare 8-char (16-byte) vectors,
8118 // start from first character again because it has aligned address.
8119 movl(result, cnt2);
8120 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count
8121 if (ae == StrIntrinsicNode::LL) {
8122 pcmpmask &= ~0x01;
8123 }
8124 jccb(Assembler::zero, COMPARE_TAIL);
8125 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8126 lea(str1, Address(str1, result, scale));
8127 lea(str2, Address(str2, result, scale));
8128 } else {
8129 lea(str1, Address(str1, result, scale1));
8130 lea(str2, Address(str2, result, scale2));
8131 }
8132 negptr(result);
8133
8134 // pcmpestri
8135 // inputs:
8136 // vec1- substring
8137 // rax - negative string length (elements count)
8138 // mem - scanned string
8139 // rdx - string length (elements count)
8140 // pcmpmask - cmp mode: 11000 (string compare with negated result)
8141 // + 00 (unsigned bytes) or + 01 (unsigned shorts)
8142 // outputs:
8143 // rcx - first mismatched element index
8144 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
8145
8146 bind(COMPARE_WIDE_VECTORS);
8147 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8148 movdqu(vec1, Address(str1, result, scale));
8149 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
8150 } else {
8151 pmovzxbw(vec1, Address(str1, result, scale1));
8152 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
8153 }
8154 // After pcmpestri cnt1(rcx) contains mismatched element index
8155
8156 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1
8157 addptr(result, stride);
8158 subptr(cnt2, stride);
8159 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
8160
8161 // compare wide vectors tail
8162 testptr(result, result);
8163 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
8164
8165 movl(cnt2, stride);
8166 movl(result, stride);
8167 negptr(result);
8168 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8169 movdqu(vec1, Address(str1, result, scale));
8170 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
8171 } else {
8172 pmovzxbw(vec1, Address(str1, result, scale1));
8173 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
8174 }
8175 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
8176
8177 // Mismatched characters in the vectors
8178 bind(VECTOR_NOT_EQUAL);
8179 addptr(cnt1, result);
8180 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
8181 subl(result, cnt2);
8182 jmpb(POP_LABEL);
8183
8184 bind(COMPARE_TAIL); // limit is zero
8185 movl(cnt2, result);
8186 // Fallthru to tail compare
8187 }
8188 // Shift str2 and str1 to the end of the arrays, negate min
8189 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8190 lea(str1, Address(str1, cnt2, scale));
8191 lea(str2, Address(str2, cnt2, scale));
8192 } else {
8193 lea(str1, Address(str1, cnt2, scale1));
8194 lea(str2, Address(str2, cnt2, scale2));
8195 }
8196 decrementl(cnt2); // first character was compared already
8197 negptr(cnt2);
8198
8199 // Compare the rest of the elements
8200 bind(WHILE_HEAD_LABEL);
8201 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae);
8202 subl(result, cnt1);
8203 jccb(Assembler::notZero, POP_LABEL);
8204 increment(cnt2);
8205 jccb(Assembler::notZero, WHILE_HEAD_LABEL);
8206
8207 // Strings are equal up to min length. Return the length difference.
8208 bind(LENGTH_DIFF_LABEL);
8209 pop(result);
8210 if (ae == StrIntrinsicNode::UU) {
8211 // Divide diff by 2 to get number of chars
8212 sarl(result, 1);
8213 }
8214 jmpb(DONE_LABEL);
8215
8216 #ifdef _LP64
8217 if (VM_Version::supports_avx512vlbw()) {
8218
8219 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED);
8220
8221 kmovql(cnt1, k7);
8222 notq(cnt1);
8223 bsfq(cnt2, cnt1);
8224 if (ae != StrIntrinsicNode::LL) {
8225 // Divide diff by 2 to get number of chars
8226 sarl(cnt2, 1);
8227 }
8228 addq(result, cnt2);
8229 if (ae == StrIntrinsicNode::LL) {
8230 load_unsigned_byte(cnt1, Address(str2, result));
8231 load_unsigned_byte(result, Address(str1, result));
8232 } else if (ae == StrIntrinsicNode::UU) {
8233 load_unsigned_short(cnt1, Address(str2, result, scale));
8234 load_unsigned_short(result, Address(str1, result, scale));
8235 } else {
8236 load_unsigned_short(cnt1, Address(str2, result, scale2));
8237 load_unsigned_byte(result, Address(str1, result, scale1));
8238 }
8239 subl(result, cnt1);
8240 jmpb(POP_LABEL);
8241 }//if (VM_Version::supports_avx512vlbw())
8242 #endif // _LP64
8243
8244 // Discard the stored length difference
8245 bind(POP_LABEL);
8246 pop(cnt1);
8247
8248 // That's it
8249 bind(DONE_LABEL);
8250 if(ae == StrIntrinsicNode::UL) {
8251 negl(result);
8252 }
8253
8254 }
8255
8256 // Search for Non-ASCII character (Negative byte value) in a byte array,
8257 // return true if it has any and false otherwise.
8258 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java
8259 // @HotSpotIntrinsicCandidate
8260 // private static boolean hasNegatives(byte[] ba, int off, int len) {
8261 // for (int i = off; i < off + len; i++) {
8262 // if (ba[i] < 0) {
8263 // return true;
8264 // }
8265 // }
8266 // return false;
8267 // }
8268 void MacroAssembler::has_negatives(Register ary1, Register len,
8269 Register result, Register tmp1,
8270 XMMRegister vec1, XMMRegister vec2) {
8271 // rsi: byte array
8272 // rcx: len
8273 // rax: result
8274 ShortBranchVerifier sbv(this);
8275 assert_different_registers(ary1, len, result, tmp1);
8276 assert_different_registers(vec1, vec2);
8277 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE;
8278
8279 // len == 0
8280 testl(len, len);
8281 jcc(Assembler::zero, FALSE_LABEL);
8282
8283 if ((UseAVX > 2) && // AVX512
8284 VM_Version::supports_avx512vlbw() &&
8285 VM_Version::supports_bmi2()) {
8286
8287 set_vector_masking(); // opening of the stub context for programming mask registers
8288
8289 Label test_64_loop, test_tail;
8290 Register tmp3_aliased = len;
8291
8292 movl(tmp1, len);
8293 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit);
8294
8295 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F
8296 andl(len, ~(64 - 1)); // vector count (in chars)
8297 jccb(Assembler::zero, test_tail);
8298
8299 lea(ary1, Address(ary1, len, Address::times_1));
8300 negptr(len);
8301
8302 bind(test_64_loop);
8303 // Check whether our 64 elements of size byte contain negatives
8304 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit);
8305 kortestql(k2, k2);
8306 jcc(Assembler::notZero, TRUE_LABEL);
8307
8308 addptr(len, 64);
8309 jccb(Assembler::notZero, test_64_loop);
8310
8311
8312 bind(test_tail);
8313 // bail out when there is nothing to be done
8314 testl(tmp1, -1);
8315 jcc(Assembler::zero, FALSE_LABEL);
8316
8317 // Save k1
8318 kmovql(k3, k1);
8319
8320 // ~(~0 << len) applied up to two times (for 32-bit scenario)
8321 #ifdef _LP64
8322 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF);
8323 shlxq(tmp3_aliased, tmp3_aliased, tmp1);
8324 notq(tmp3_aliased);
8325 kmovql(k1, tmp3_aliased);
8326 #else
8327 Label k_init;
8328 jmp(k_init);
8329
8330 // We could not read 64-bits from a general purpose register thus we move
8331 // data required to compose 64 1's to the instruction stream
8332 // We emit 64 byte wide series of elements from 0..63 which later on would
8333 // be used as a compare targets with tail count contained in tmp1 register.
8334 // Result would be a k1 register having tmp1 consecutive number or 1
8335 // counting from least significant bit.
8336 address tmp = pc();
8337 emit_int64(0x0706050403020100);
8338 emit_int64(0x0F0E0D0C0B0A0908);
8339 emit_int64(0x1716151413121110);
8340 emit_int64(0x1F1E1D1C1B1A1918);
8341 emit_int64(0x2726252423222120);
8342 emit_int64(0x2F2E2D2C2B2A2928);
8343 emit_int64(0x3736353433323130);
8344 emit_int64(0x3F3E3D3C3B3A3938);
8345
8346 bind(k_init);
8347 lea(len, InternalAddress(tmp));
8348 // create mask to test for negative byte inside a vector
8349 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit);
8350 evpcmpgtb(k1, vec1, Address(len, 0), Assembler::AVX_512bit);
8351
8352 #endif
8353 evpcmpgtb(k2, k1, vec2, Address(ary1, 0), Assembler::AVX_512bit);
8354 ktestq(k2, k1);
8355 // Restore k1
8356 kmovql(k1, k3);
8357 jcc(Assembler::notZero, TRUE_LABEL);
8358
8359 jmp(FALSE_LABEL);
8360
8361 clear_vector_masking(); // closing of the stub context for programming mask registers
8362 } else {
8363 movl(result, len); // copy
8364
8365 if (UseAVX == 2 && UseSSE >= 2) {
8366 // With AVX2, use 32-byte vector compare
8367 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8368
8369 // Compare 32-byte vectors
8370 andl(result, 0x0000001f); // tail count (in bytes)
8371 andl(len, 0xffffffe0); // vector count (in bytes)
8372 jccb(Assembler::zero, COMPARE_TAIL);
8373
8374 lea(ary1, Address(ary1, len, Address::times_1));
8375 negptr(len);
8376
8377 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector
8378 movdl(vec2, tmp1);
8379 vpbroadcastd(vec2, vec2);
8380
8381 bind(COMPARE_WIDE_VECTORS);
8382 vmovdqu(vec1, Address(ary1, len, Address::times_1));
8383 vptest(vec1, vec2);
8384 jccb(Assembler::notZero, TRUE_LABEL);
8385 addptr(len, 32);
8386 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8387
8388 testl(result, result);
8389 jccb(Assembler::zero, FALSE_LABEL);
8390
8391 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
8392 vptest(vec1, vec2);
8393 jccb(Assembler::notZero, TRUE_LABEL);
8394 jmpb(FALSE_LABEL);
8395
8396 bind(COMPARE_TAIL); // len is zero
8397 movl(len, result);
8398 // Fallthru to tail compare
8399 } else if (UseSSE42Intrinsics) {
8400 assert(UseSSE >= 4, "SSE4 must be for SSE4.2 intrinsics to be available");
8401 // With SSE4.2, use double quad vector compare
8402 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8403
8404 // Compare 16-byte vectors
8405 andl(result, 0x0000000f); // tail count (in bytes)
8406 andl(len, 0xfffffff0); // vector count (in bytes)
8407 jccb(Assembler::zero, COMPARE_TAIL);
8408
8409 lea(ary1, Address(ary1, len, Address::times_1));
8410 negptr(len);
8411
8412 movl(tmp1, 0x80808080);
8413 movdl(vec2, tmp1);
8414 pshufd(vec2, vec2, 0);
8415
8416 bind(COMPARE_WIDE_VECTORS);
8417 movdqu(vec1, Address(ary1, len, Address::times_1));
8418 ptest(vec1, vec2);
8419 jccb(Assembler::notZero, TRUE_LABEL);
8420 addptr(len, 16);
8421 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8422
8423 testl(result, result);
8424 jccb(Assembler::zero, FALSE_LABEL);
8425
8426 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
8427 ptest(vec1, vec2);
8428 jccb(Assembler::notZero, TRUE_LABEL);
8429 jmpb(FALSE_LABEL);
8430
8431 bind(COMPARE_TAIL); // len is zero
8432 movl(len, result);
8433 // Fallthru to tail compare
8434 }
8435 }
8436 // Compare 4-byte vectors
8437 andl(len, 0xfffffffc); // vector count (in bytes)
8438 jccb(Assembler::zero, COMPARE_CHAR);
8439
8440 lea(ary1, Address(ary1, len, Address::times_1));
8441 negptr(len);
8442
8443 bind(COMPARE_VECTORS);
8444 movl(tmp1, Address(ary1, len, Address::times_1));
8445 andl(tmp1, 0x80808080);
8446 jccb(Assembler::notZero, TRUE_LABEL);
8447 addptr(len, 4);
8448 jcc(Assembler::notZero, COMPARE_VECTORS);
8449
8450 // Compare trailing char (final 2 bytes), if any
8451 bind(COMPARE_CHAR);
8452 testl(result, 0x2); // tail char
8453 jccb(Assembler::zero, COMPARE_BYTE);
8454 load_unsigned_short(tmp1, Address(ary1, 0));
8455 andl(tmp1, 0x00008080);
8456 jccb(Assembler::notZero, TRUE_LABEL);
8457 subptr(result, 2);
8458 lea(ary1, Address(ary1, 2));
8459
8460 bind(COMPARE_BYTE);
8461 testl(result, 0x1); // tail byte
8462 jccb(Assembler::zero, FALSE_LABEL);
8463 load_unsigned_byte(tmp1, Address(ary1, 0));
8464 andl(tmp1, 0x00000080);
8465 jccb(Assembler::notEqual, TRUE_LABEL);
8466 jmpb(FALSE_LABEL);
8467
8468 bind(TRUE_LABEL);
8469 movl(result, 1); // return true
8470 jmpb(DONE);
8471
8472 bind(FALSE_LABEL);
8473 xorl(result, result); // return false
8474
8475 // That's it
8476 bind(DONE);
8477 if (UseAVX >= 2 && UseSSE >= 2) {
8478 // clean upper bits of YMM registers
8479 vpxor(vec1, vec1);
8480 vpxor(vec2, vec2);
8481 }
8482 }
8483
8484 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings.
8485 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2,
8486 Register limit, Register result, Register chr,
8487 XMMRegister vec1, XMMRegister vec2, bool is_char) {
8488 ShortBranchVerifier sbv(this);
8489 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE;
8490
8491 int length_offset = arrayOopDesc::length_offset_in_bytes();
8492 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE);
8493
8494 if (is_array_equ) {
8495 // Check the input args
8496 cmpptr(ary1, ary2);
8497 jcc(Assembler::equal, TRUE_LABEL);
8498
8499 // Need additional checks for arrays_equals.
8500 testptr(ary1, ary1);
8501 jcc(Assembler::zero, FALSE_LABEL);
8502 testptr(ary2, ary2);
8503 jcc(Assembler::zero, FALSE_LABEL);
8504
8505 // Check the lengths
8506 movl(limit, Address(ary1, length_offset));
8507 cmpl(limit, Address(ary2, length_offset));
8508 jcc(Assembler::notEqual, FALSE_LABEL);
8509 }
8510
8511 // count == 0
8512 testl(limit, limit);
8513 jcc(Assembler::zero, TRUE_LABEL);
8514
8515 if (is_array_equ) {
8516 // Load array address
8517 lea(ary1, Address(ary1, base_offset));
8518 lea(ary2, Address(ary2, base_offset));
8519 }
8520
8521 if (is_array_equ && is_char) {
8522 // arrays_equals when used for char[].
8523 shll(limit, 1); // byte count != 0
8524 }
8525 movl(result, limit); // copy
8526
8527 if (UseAVX >= 2) {
8528 // With AVX2, use 32-byte vector compare
8529 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8530
8531 // Compare 32-byte vectors
8532 andl(result, 0x0000001f); // tail count (in bytes)
8533 andl(limit, 0xffffffe0); // vector count (in bytes)
8534 jcc(Assembler::zero, COMPARE_TAIL);
8535
8536 lea(ary1, Address(ary1, limit, Address::times_1));
8537 lea(ary2, Address(ary2, limit, Address::times_1));
8538 negptr(limit);
8539
8540 bind(COMPARE_WIDE_VECTORS);
8541
8542 #ifdef _LP64
8543 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
8544 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3;
8545
8546 cmpl(limit, -64);
8547 jccb(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2);
8548
8549 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
8550
8551 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit);
8552 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit);
8553 kortestql(k7, k7);
8554 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
8555 addptr(limit, 64); // update since we already compared at this addr
8556 cmpl(limit, -64);
8557 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3);
8558
8559 // At this point we may still need to compare -limit+result bytes.
8560 // We could execute the next two instruction and just continue via non-wide path:
8561 // cmpl(limit, 0);
8562 // jcc(Assembler::equal, COMPARE_TAIL); // true
8563 // But since we stopped at the points ary{1,2}+limit which are
8564 // not farther than 64 bytes from the ends of arrays ary{1,2}+result
8565 // (|limit| <= 32 and result < 32),
8566 // we may just compare the last 64 bytes.
8567 //
8568 addptr(result, -64); // it is safe, bc we just came from this area
8569 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit);
8570 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit);
8571 kortestql(k7, k7);
8572 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
8573
8574 jmp(TRUE_LABEL);
8575
8576 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
8577
8578 }//if (VM_Version::supports_avx512vlbw())
8579 #endif //_LP64
8580
8581 vmovdqu(vec1, Address(ary1, limit, Address::times_1));
8582 vmovdqu(vec2, Address(ary2, limit, Address::times_1));
8583 vpxor(vec1, vec2);
8584
8585 vptest(vec1, vec1);
8586 jccb(Assembler::notZero, FALSE_LABEL);
8587 addptr(limit, 32);
8588 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8589
8590 testl(result, result);
8591 jccb(Assembler::zero, TRUE_LABEL);
8592
8593 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
8594 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
8595 vpxor(vec1, vec2);
8596
8597 vptest(vec1, vec1);
8598 jccb(Assembler::notZero, FALSE_LABEL);
8599 jmpb(TRUE_LABEL);
8600
8601 bind(COMPARE_TAIL); // limit is zero
8602 movl(limit, result);
8603 // Fallthru to tail compare
8604 } else if (UseSSE42Intrinsics) {
8605 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
8606 // With SSE4.2, use double quad vector compare
8607 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8608
8609 // Compare 16-byte vectors
8610 andl(result, 0x0000000f); // tail count (in bytes)
8611 andl(limit, 0xfffffff0); // vector count (in bytes)
8612 jccb(Assembler::zero, COMPARE_TAIL);
8613
8614 lea(ary1, Address(ary1, limit, Address::times_1));
8615 lea(ary2, Address(ary2, limit, Address::times_1));
8616 negptr(limit);
8617
8618 bind(COMPARE_WIDE_VECTORS);
8619 movdqu(vec1, Address(ary1, limit, Address::times_1));
8620 movdqu(vec2, Address(ary2, limit, Address::times_1));
8621 pxor(vec1, vec2);
8622
8623 ptest(vec1, vec1);
8624 jccb(Assembler::notZero, FALSE_LABEL);
8625 addptr(limit, 16);
8626 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8627
8628 testl(result, result);
8629 jccb(Assembler::zero, TRUE_LABEL);
8630
8631 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
8632 movdqu(vec2, Address(ary2, result, Address::times_1, -16));
8633 pxor(vec1, vec2);
8634
8635 ptest(vec1, vec1);
8636 jccb(Assembler::notZero, FALSE_LABEL);
8637 jmpb(TRUE_LABEL);
8638
8639 bind(COMPARE_TAIL); // limit is zero
8640 movl(limit, result);
8641 // Fallthru to tail compare
8642 }
8643
8644 // Compare 4-byte vectors
8645 andl(limit, 0xfffffffc); // vector count (in bytes)
8646 jccb(Assembler::zero, COMPARE_CHAR);
8647
8648 lea(ary1, Address(ary1, limit, Address::times_1));
8649 lea(ary2, Address(ary2, limit, Address::times_1));
8650 negptr(limit);
8651
8652 bind(COMPARE_VECTORS);
8653 movl(chr, Address(ary1, limit, Address::times_1));
8654 cmpl(chr, Address(ary2, limit, Address::times_1));
8655 jccb(Assembler::notEqual, FALSE_LABEL);
8656 addptr(limit, 4);
8657 jcc(Assembler::notZero, COMPARE_VECTORS);
8658
8659 // Compare trailing char (final 2 bytes), if any
8660 bind(COMPARE_CHAR);
8661 testl(result, 0x2); // tail char
8662 jccb(Assembler::zero, COMPARE_BYTE);
8663 load_unsigned_short(chr, Address(ary1, 0));
8664 load_unsigned_short(limit, Address(ary2, 0));
8665 cmpl(chr, limit);
8666 jccb(Assembler::notEqual, FALSE_LABEL);
8667
8668 if (is_array_equ && is_char) {
8669 bind(COMPARE_BYTE);
8670 } else {
8671 lea(ary1, Address(ary1, 2));
8672 lea(ary2, Address(ary2, 2));
8673
8674 bind(COMPARE_BYTE);
8675 testl(result, 0x1); // tail byte
8676 jccb(Assembler::zero, TRUE_LABEL);
8677 load_unsigned_byte(chr, Address(ary1, 0));
8678 load_unsigned_byte(limit, Address(ary2, 0));
8679 cmpl(chr, limit);
8680 jccb(Assembler::notEqual, FALSE_LABEL);
8681 }
8682 bind(TRUE_LABEL);
8683 movl(result, 1); // return true
8684 jmpb(DONE);
8685
8686 bind(FALSE_LABEL);
8687 xorl(result, result); // return false
8688
8689 // That's it
8690 bind(DONE);
8691 if (UseAVX >= 2) {
8692 // clean upper bits of YMM registers
8693 vpxor(vec1, vec1);
8694 vpxor(vec2, vec2);
8695 }
8696 }
8697
8698 #endif
8699
8700 void MacroAssembler::generate_fill(BasicType t, bool aligned,
8701 Register to, Register value, Register count,
8702 Register rtmp, XMMRegister xtmp) {
8703 ShortBranchVerifier sbv(this);
8704 assert_different_registers(to, value, count, rtmp);
8705 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
8706 Label L_fill_2_bytes, L_fill_4_bytes;
8707
8708 int shift = -1;
8709 switch (t) {
8710 case T_BYTE:
8711 shift = 2;
8712 break;
8713 case T_SHORT:
8714 shift = 1;
8715 break;
8716 case T_INT:
8717 shift = 0;
8718 break;
8719 default: ShouldNotReachHere();
8720 }
8721
8722 if (t == T_BYTE) {
8723 andl(value, 0xff);
8724 movl(rtmp, value);
8725 shll(rtmp, 8);
8726 orl(value, rtmp);
8727 }
8728 if (t == T_SHORT) {
8729 andl(value, 0xffff);
8730 }
8731 if (t == T_BYTE || t == T_SHORT) {
8732 movl(rtmp, value);
8733 shll(rtmp, 16);
8734 orl(value, rtmp);
8735 }
8736
8737 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
8738 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp
8739 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) {
8740 // align source address at 4 bytes address boundary
8741 if (t == T_BYTE) {
8742 // One byte misalignment happens only for byte arrays
8743 testptr(to, 1);
8744 jccb(Assembler::zero, L_skip_align1);
8745 movb(Address(to, 0), value);
8746 increment(to);
8747 decrement(count);
8748 BIND(L_skip_align1);
8749 }
8750 // Two bytes misalignment happens only for byte and short (char) arrays
8751 testptr(to, 2);
8752 jccb(Assembler::zero, L_skip_align2);
8753 movw(Address(to, 0), value);
8754 addptr(to, 2);
8755 subl(count, 1<<(shift-1));
8756 BIND(L_skip_align2);
8757 }
8758 if (UseSSE < 2) {
8759 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
8760 // Fill 32-byte chunks
8761 subl(count, 8 << shift);
8762 jcc(Assembler::less, L_check_fill_8_bytes);
8763 align(16);
8764
8765 BIND(L_fill_32_bytes_loop);
8766
8767 for (int i = 0; i < 32; i += 4) {
8768 movl(Address(to, i), value);
8769 }
8770
8771 addptr(to, 32);
8772 subl(count, 8 << shift);
8773 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
8774 BIND(L_check_fill_8_bytes);
8775 addl(count, 8 << shift);
8776 jccb(Assembler::zero, L_exit);
8777 jmpb(L_fill_8_bytes);
8778
8779 //
8780 // length is too short, just fill qwords
8781 //
8782 BIND(L_fill_8_bytes_loop);
8783 movl(Address(to, 0), value);
8784 movl(Address(to, 4), value);
8785 addptr(to, 8);
8786 BIND(L_fill_8_bytes);
8787 subl(count, 1 << (shift + 1));
8788 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
8789 // fall through to fill 4 bytes
8790 } else {
8791 Label L_fill_32_bytes;
8792 if (!UseUnalignedLoadStores) {
8793 // align to 8 bytes, we know we are 4 byte aligned to start
8794 testptr(to, 4);
8795 jccb(Assembler::zero, L_fill_32_bytes);
8796 movl(Address(to, 0), value);
8797 addptr(to, 4);
8798 subl(count, 1<<shift);
8799 }
8800 BIND(L_fill_32_bytes);
8801 {
8802 assert( UseSSE >= 2, "supported cpu only" );
8803 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
8804 if (UseAVX > 2) {
8805 movl(rtmp, 0xffff);
8806 kmovwl(k1, rtmp);
8807 }
8808 movdl(xtmp, value);
8809 if (UseAVX > 2 && UseUnalignedLoadStores) {
8810 // Fill 64-byte chunks
8811 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
8812 evpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit);
8813
8814 subl(count, 16 << shift);
8815 jcc(Assembler::less, L_check_fill_32_bytes);
8816 align(16);
8817
8818 BIND(L_fill_64_bytes_loop);
8819 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit);
8820 addptr(to, 64);
8821 subl(count, 16 << shift);
8822 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
8823
8824 BIND(L_check_fill_32_bytes);
8825 addl(count, 8 << shift);
8826 jccb(Assembler::less, L_check_fill_8_bytes);
8827 vmovdqu(Address(to, 0), xtmp);
8828 addptr(to, 32);
8829 subl(count, 8 << shift);
8830
8831 BIND(L_check_fill_8_bytes);
8832 } else if (UseAVX == 2 && UseUnalignedLoadStores) {
8833 // Fill 64-byte chunks
8834 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
8835 vpbroadcastd(xtmp, xtmp);
8836
8837 subl(count, 16 << shift);
8838 jcc(Assembler::less, L_check_fill_32_bytes);
8839 align(16);
8840
8841 BIND(L_fill_64_bytes_loop);
8842 vmovdqu(Address(to, 0), xtmp);
8843 vmovdqu(Address(to, 32), xtmp);
8844 addptr(to, 64);
8845 subl(count, 16 << shift);
8846 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
8847
8848 BIND(L_check_fill_32_bytes);
8849 addl(count, 8 << shift);
8850 jccb(Assembler::less, L_check_fill_8_bytes);
8851 vmovdqu(Address(to, 0), xtmp);
8852 addptr(to, 32);
8853 subl(count, 8 << shift);
8854
8855 BIND(L_check_fill_8_bytes);
8856 // clean upper bits of YMM registers
8857 movdl(xtmp, value);
8858 pshufd(xtmp, xtmp, 0);
8859 } else {
8860 // Fill 32-byte chunks
8861 pshufd(xtmp, xtmp, 0);
8862
8863 subl(count, 8 << shift);
8864 jcc(Assembler::less, L_check_fill_8_bytes);
8865 align(16);
8866
8867 BIND(L_fill_32_bytes_loop);
8868
8869 if (UseUnalignedLoadStores) {
8870 movdqu(Address(to, 0), xtmp);
8871 movdqu(Address(to, 16), xtmp);
8872 } else {
8873 movq(Address(to, 0), xtmp);
8874 movq(Address(to, 8), xtmp);
8875 movq(Address(to, 16), xtmp);
8876 movq(Address(to, 24), xtmp);
8877 }
8878
8879 addptr(to, 32);
8880 subl(count, 8 << shift);
8881 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
8882
8883 BIND(L_check_fill_8_bytes);
8884 }
8885 addl(count, 8 << shift);
8886 jccb(Assembler::zero, L_exit);
8887 jmpb(L_fill_8_bytes);
8888
8889 //
8890 // length is too short, just fill qwords
8891 //
8892 BIND(L_fill_8_bytes_loop);
8893 movq(Address(to, 0), xtmp);
8894 addptr(to, 8);
8895 BIND(L_fill_8_bytes);
8896 subl(count, 1 << (shift + 1));
8897 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
8898 }
8899 }
8900 // fill trailing 4 bytes
8901 BIND(L_fill_4_bytes);
8902 testl(count, 1<<shift);
8903 jccb(Assembler::zero, L_fill_2_bytes);
8904 movl(Address(to, 0), value);
8905 if (t == T_BYTE || t == T_SHORT) {
8906 addptr(to, 4);
8907 BIND(L_fill_2_bytes);
8908 // fill trailing 2 bytes
8909 testl(count, 1<<(shift-1));
8910 jccb(Assembler::zero, L_fill_byte);
8911 movw(Address(to, 0), value);
8912 if (t == T_BYTE) {
8913 addptr(to, 2);
8914 BIND(L_fill_byte);
8915 // fill trailing byte
8916 testl(count, 1);
8917 jccb(Assembler::zero, L_exit);
8918 movb(Address(to, 0), value);
8919 } else {
8920 BIND(L_fill_byte);
8921 }
8922 } else {
8923 BIND(L_fill_2_bytes);
8924 }
8925 BIND(L_exit);
8926 }
8927
8928 // encode char[] to byte[] in ISO_8859_1
8929 //@HotSpotIntrinsicCandidate
8930 //private static int implEncodeISOArray(byte[] sa, int sp,
8931 //byte[] da, int dp, int len) {
8932 // int i = 0;
8933 // for (; i < len; i++) {
8934 // char c = StringUTF16.getChar(sa, sp++);
8935 // if (c > '\u00FF')
8936 // break;
8937 // da[dp++] = (byte)c;
8938 // }
8939 // return i;
8940 //}
8941 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len,
8942 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
8943 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
8944 Register tmp5, Register result) {
8945
8946 // rsi: src
8947 // rdi: dst
8948 // rdx: len
8949 // rcx: tmp5
8950 // rax: result
8951 ShortBranchVerifier sbv(this);
8952 assert_different_registers(src, dst, len, tmp5, result);
8953 Label L_done, L_copy_1_char, L_copy_1_char_exit;
8954
8955 // set result
8956 xorl(result, result);
8957 // check for zero length
8958 testl(len, len);
8959 jcc(Assembler::zero, L_done);
8960
8961 movl(result, len);
8962
8963 // Setup pointers
8964 lea(src, Address(src, len, Address::times_2)); // char[]
8965 lea(dst, Address(dst, len, Address::times_1)); // byte[]
8966 negptr(len);
8967
8968 if (UseSSE42Intrinsics || UseAVX >= 2) {
8969 assert(UseSSE42Intrinsics ? UseSSE >= 4 : true, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
8970 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit;
8971 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit;
8972
8973 if (UseAVX >= 2) {
8974 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit;
8975 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
8976 movdl(tmp1Reg, tmp5);
8977 vpbroadcastd(tmp1Reg, tmp1Reg);
8978 jmpb(L_chars_32_check);
8979
8980 bind(L_copy_32_chars);
8981 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64));
8982 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32));
8983 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
8984 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
8985 jccb(Assembler::notZero, L_copy_32_chars_exit);
8986 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
8987 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1);
8988 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg);
8989
8990 bind(L_chars_32_check);
8991 addptr(len, 32);
8992 jccb(Assembler::lessEqual, L_copy_32_chars);
8993
8994 bind(L_copy_32_chars_exit);
8995 subptr(len, 16);
8996 jccb(Assembler::greater, L_copy_16_chars_exit);
8997
8998 } else if (UseSSE42Intrinsics) {
8999 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
9000 movdl(tmp1Reg, tmp5);
9001 pshufd(tmp1Reg, tmp1Reg, 0);
9002 jmpb(L_chars_16_check);
9003 }
9004
9005 bind(L_copy_16_chars);
9006 if (UseAVX >= 2) {
9007 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32));
9008 vptest(tmp2Reg, tmp1Reg);
9009 jccb(Assembler::notZero, L_copy_16_chars_exit);
9010 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1);
9011 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1);
9012 } else {
9013 if (UseAVX > 0) {
9014 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
9015 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
9016 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0);
9017 } else {
9018 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
9019 por(tmp2Reg, tmp3Reg);
9020 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
9021 por(tmp2Reg, tmp4Reg);
9022 }
9023 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
9024 jccb(Assembler::notZero, L_copy_16_chars_exit);
9025 packuswb(tmp3Reg, tmp4Reg);
9026 }
9027 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg);
9028
9029 bind(L_chars_16_check);
9030 addptr(len, 16);
9031 jccb(Assembler::lessEqual, L_copy_16_chars);
9032
9033 bind(L_copy_16_chars_exit);
9034 if (UseAVX >= 2) {
9035 // clean upper bits of YMM registers
9036 vpxor(tmp2Reg, tmp2Reg);
9037 vpxor(tmp3Reg, tmp3Reg);
9038 vpxor(tmp4Reg, tmp4Reg);
9039 movdl(tmp1Reg, tmp5);
9040 pshufd(tmp1Reg, tmp1Reg, 0);
9041 }
9042 subptr(len, 8);
9043 jccb(Assembler::greater, L_copy_8_chars_exit);
9044
9045 bind(L_copy_8_chars);
9046 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16));
9047 ptest(tmp3Reg, tmp1Reg);
9048 jccb(Assembler::notZero, L_copy_8_chars_exit);
9049 packuswb(tmp3Reg, tmp1Reg);
9050 movq(Address(dst, len, Address::times_1, -8), tmp3Reg);
9051 addptr(len, 8);
9052 jccb(Assembler::lessEqual, L_copy_8_chars);
9053
9054 bind(L_copy_8_chars_exit);
9055 subptr(len, 8);
9056 jccb(Assembler::zero, L_done);
9057 }
9058
9059 bind(L_copy_1_char);
9060 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0));
9061 testl(tmp5, 0xff00); // check if Unicode char
9062 jccb(Assembler::notZero, L_copy_1_char_exit);
9063 movb(Address(dst, len, Address::times_1, 0), tmp5);
9064 addptr(len, 1);
9065 jccb(Assembler::less, L_copy_1_char);
9066
9067 bind(L_copy_1_char_exit);
9068 addptr(result, len); // len is negative count of not processed elements
9069
9070 bind(L_done);
9071 }
9072
9073 #ifdef _LP64
9074 /**
9075 * Helper for multiply_to_len().
9076 */
9077 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) {
9078 addq(dest_lo, src1);
9079 adcq(dest_hi, 0);
9080 addq(dest_lo, src2);
9081 adcq(dest_hi, 0);
9082 }
9083
9084 /**
9085 * Multiply 64 bit by 64 bit first loop.
9086 */
9087 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
9088 Register y, Register y_idx, Register z,
9089 Register carry, Register product,
9090 Register idx, Register kdx) {
9091 //
9092 // jlong carry, x[], y[], z[];
9093 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
9094 // huge_128 product = y[idx] * x[xstart] + carry;
9095 // z[kdx] = (jlong)product;
9096 // carry = (jlong)(product >>> 64);
9097 // }
9098 // z[xstart] = carry;
9099 //
9100
9101 Label L_first_loop, L_first_loop_exit;
9102 Label L_one_x, L_one_y, L_multiply;
9103
9104 decrementl(xstart);
9105 jcc(Assembler::negative, L_one_x);
9106
9107 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
9108 rorq(x_xstart, 32); // convert big-endian to little-endian
9109
9110 bind(L_first_loop);
9111 decrementl(idx);
9112 jcc(Assembler::negative, L_first_loop_exit);
9113 decrementl(idx);
9114 jcc(Assembler::negative, L_one_y);
9115 movq(y_idx, Address(y, idx, Address::times_4, 0));
9116 rorq(y_idx, 32); // convert big-endian to little-endian
9117 bind(L_multiply);
9118 movq(product, x_xstart);
9119 mulq(y_idx); // product(rax) * y_idx -> rdx:rax
9120 addq(product, carry);
9121 adcq(rdx, 0);
9122 subl(kdx, 2);
9123 movl(Address(z, kdx, Address::times_4, 4), product);
9124 shrq(product, 32);
9125 movl(Address(z, kdx, Address::times_4, 0), product);
9126 movq(carry, rdx);
9127 jmp(L_first_loop);
9128
9129 bind(L_one_y);
9130 movl(y_idx, Address(y, 0));
9131 jmp(L_multiply);
9132
9133 bind(L_one_x);
9134 movl(x_xstart, Address(x, 0));
9135 jmp(L_first_loop);
9136
9137 bind(L_first_loop_exit);
9138 }
9139
9140 /**
9141 * Multiply 64 bit by 64 bit and add 128 bit.
9142 */
9143 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z,
9144 Register yz_idx, Register idx,
9145 Register carry, Register product, int offset) {
9146 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
9147 // z[kdx] = (jlong)product;
9148
9149 movq(yz_idx, Address(y, idx, Address::times_4, offset));
9150 rorq(yz_idx, 32); // convert big-endian to little-endian
9151 movq(product, x_xstart);
9152 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
9153 movq(yz_idx, Address(z, idx, Address::times_4, offset));
9154 rorq(yz_idx, 32); // convert big-endian to little-endian
9155
9156 add2_with_carry(rdx, product, carry, yz_idx);
9157
9158 movl(Address(z, idx, Address::times_4, offset+4), product);
9159 shrq(product, 32);
9160 movl(Address(z, idx, Address::times_4, offset), product);
9161
9162 }
9163
9164 /**
9165 * Multiply 128 bit by 128 bit. Unrolled inner loop.
9166 */
9167 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z,
9168 Register yz_idx, Register idx, Register jdx,
9169 Register carry, Register product,
9170 Register carry2) {
9171 // jlong carry, x[], y[], z[];
9172 // int kdx = ystart+1;
9173 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
9174 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
9175 // z[kdx+idx+1] = (jlong)product;
9176 // jlong carry2 = (jlong)(product >>> 64);
9177 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
9178 // z[kdx+idx] = (jlong)product;
9179 // carry = (jlong)(product >>> 64);
9180 // }
9181 // idx += 2;
9182 // if (idx > 0) {
9183 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
9184 // z[kdx+idx] = (jlong)product;
9185 // carry = (jlong)(product >>> 64);
9186 // }
9187 //
9188
9189 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
9190
9191 movl(jdx, idx);
9192 andl(jdx, 0xFFFFFFFC);
9193 shrl(jdx, 2);
9194
9195 bind(L_third_loop);
9196 subl(jdx, 1);
9197 jcc(Assembler::negative, L_third_loop_exit);
9198 subl(idx, 4);
9199
9200 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
9201 movq(carry2, rdx);
9202
9203 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
9204 movq(carry, rdx);
9205 jmp(L_third_loop);
9206
9207 bind (L_third_loop_exit);
9208
9209 andl (idx, 0x3);
9210 jcc(Assembler::zero, L_post_third_loop_done);
9211
9212 Label L_check_1;
9213 subl(idx, 2);
9214 jcc(Assembler::negative, L_check_1);
9215
9216 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
9217 movq(carry, rdx);
9218
9219 bind (L_check_1);
9220 addl (idx, 0x2);
9221 andl (idx, 0x1);
9222 subl(idx, 1);
9223 jcc(Assembler::negative, L_post_third_loop_done);
9224
9225 movl(yz_idx, Address(y, idx, Address::times_4, 0));
9226 movq(product, x_xstart);
9227 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
9228 movl(yz_idx, Address(z, idx, Address::times_4, 0));
9229
9230 add2_with_carry(rdx, product, yz_idx, carry);
9231
9232 movl(Address(z, idx, Address::times_4, 0), product);
9233 shrq(product, 32);
9234
9235 shlq(rdx, 32);
9236 orq(product, rdx);
9237 movq(carry, product);
9238
9239 bind(L_post_third_loop_done);
9240 }
9241
9242 /**
9243 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop.
9244 *
9245 */
9246 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z,
9247 Register carry, Register carry2,
9248 Register idx, Register jdx,
9249 Register yz_idx1, Register yz_idx2,
9250 Register tmp, Register tmp3, Register tmp4) {
9251 assert(UseBMI2Instructions, "should be used only when BMI2 is available");
9252
9253 // jlong carry, x[], y[], z[];
9254 // int kdx = ystart+1;
9255 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
9256 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry;
9257 // jlong carry2 = (jlong)(tmp3 >>> 64);
9258 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2;
9259 // carry = (jlong)(tmp4 >>> 64);
9260 // z[kdx+idx+1] = (jlong)tmp3;
9261 // z[kdx+idx] = (jlong)tmp4;
9262 // }
9263 // idx += 2;
9264 // if (idx > 0) {
9265 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry;
9266 // z[kdx+idx] = (jlong)yz_idx1;
9267 // carry = (jlong)(yz_idx1 >>> 64);
9268 // }
9269 //
9270
9271 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
9272
9273 movl(jdx, idx);
9274 andl(jdx, 0xFFFFFFFC);
9275 shrl(jdx, 2);
9276
9277 bind(L_third_loop);
9278 subl(jdx, 1);
9279 jcc(Assembler::negative, L_third_loop_exit);
9280 subl(idx, 4);
9281
9282 movq(yz_idx1, Address(y, idx, Address::times_4, 8));
9283 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
9284 movq(yz_idx2, Address(y, idx, Address::times_4, 0));
9285 rorxq(yz_idx2, yz_idx2, 32);
9286
9287 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
9288 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp
9289
9290 movq(yz_idx1, Address(z, idx, Address::times_4, 8));
9291 rorxq(yz_idx1, yz_idx1, 32);
9292 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
9293 rorxq(yz_idx2, yz_idx2, 32);
9294
9295 if (VM_Version::supports_adx()) {
9296 adcxq(tmp3, carry);
9297 adoxq(tmp3, yz_idx1);
9298
9299 adcxq(tmp4, tmp);
9300 adoxq(tmp4, yz_idx2);
9301
9302 movl(carry, 0); // does not affect flags
9303 adcxq(carry2, carry);
9304 adoxq(carry2, carry);
9305 } else {
9306 add2_with_carry(tmp4, tmp3, carry, yz_idx1);
9307 add2_with_carry(carry2, tmp4, tmp, yz_idx2);
9308 }
9309 movq(carry, carry2);
9310
9311 movl(Address(z, idx, Address::times_4, 12), tmp3);
9312 shrq(tmp3, 32);
9313 movl(Address(z, idx, Address::times_4, 8), tmp3);
9314
9315 movl(Address(z, idx, Address::times_4, 4), tmp4);
9316 shrq(tmp4, 32);
9317 movl(Address(z, idx, Address::times_4, 0), tmp4);
9318
9319 jmp(L_third_loop);
9320
9321 bind (L_third_loop_exit);
9322
9323 andl (idx, 0x3);
9324 jcc(Assembler::zero, L_post_third_loop_done);
9325
9326 Label L_check_1;
9327 subl(idx, 2);
9328 jcc(Assembler::negative, L_check_1);
9329
9330 movq(yz_idx1, Address(y, idx, Address::times_4, 0));
9331 rorxq(yz_idx1, yz_idx1, 32);
9332 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
9333 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
9334 rorxq(yz_idx2, yz_idx2, 32);
9335
9336 add2_with_carry(tmp4, tmp3, carry, yz_idx2);
9337
9338 movl(Address(z, idx, Address::times_4, 4), tmp3);
9339 shrq(tmp3, 32);
9340 movl(Address(z, idx, Address::times_4, 0), tmp3);
9341 movq(carry, tmp4);
9342
9343 bind (L_check_1);
9344 addl (idx, 0x2);
9345 andl (idx, 0x1);
9346 subl(idx, 1);
9347 jcc(Assembler::negative, L_post_third_loop_done);
9348 movl(tmp4, Address(y, idx, Address::times_4, 0));
9349 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3
9350 movl(tmp4, Address(z, idx, Address::times_4, 0));
9351
9352 add2_with_carry(carry2, tmp3, tmp4, carry);
9353
9354 movl(Address(z, idx, Address::times_4, 0), tmp3);
9355 shrq(tmp3, 32);
9356
9357 shlq(carry2, 32);
9358 orq(tmp3, carry2);
9359 movq(carry, tmp3);
9360
9361 bind(L_post_third_loop_done);
9362 }
9363
9364 /**
9365 * Code for BigInteger::multiplyToLen() instrinsic.
9366 *
9367 * rdi: x
9368 * rax: xlen
9369 * rsi: y
9370 * rcx: ylen
9371 * r8: z
9372 * r11: zlen
9373 * r12: tmp1
9374 * r13: tmp2
9375 * r14: tmp3
9376 * r15: tmp4
9377 * rbx: tmp5
9378 *
9379 */
9380 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen,
9381 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) {
9382 ShortBranchVerifier sbv(this);
9383 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx);
9384
9385 push(tmp1);
9386 push(tmp2);
9387 push(tmp3);
9388 push(tmp4);
9389 push(tmp5);
9390
9391 push(xlen);
9392 push(zlen);
9393
9394 const Register idx = tmp1;
9395 const Register kdx = tmp2;
9396 const Register xstart = tmp3;
9397
9398 const Register y_idx = tmp4;
9399 const Register carry = tmp5;
9400 const Register product = xlen;
9401 const Register x_xstart = zlen; // reuse register
9402
9403 // First Loop.
9404 //
9405 // final static long LONG_MASK = 0xffffffffL;
9406 // int xstart = xlen - 1;
9407 // int ystart = ylen - 1;
9408 // long carry = 0;
9409 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
9410 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
9411 // z[kdx] = (int)product;
9412 // carry = product >>> 32;
9413 // }
9414 // z[xstart] = (int)carry;
9415 //
9416
9417 movl(idx, ylen); // idx = ylen;
9418 movl(kdx, zlen); // kdx = xlen+ylen;
9419 xorq(carry, carry); // carry = 0;
9420
9421 Label L_done;
9422
9423 movl(xstart, xlen);
9424 decrementl(xstart);
9425 jcc(Assembler::negative, L_done);
9426
9427 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
9428
9429 Label L_second_loop;
9430 testl(kdx, kdx);
9431 jcc(Assembler::zero, L_second_loop);
9432
9433 Label L_carry;
9434 subl(kdx, 1);
9435 jcc(Assembler::zero, L_carry);
9436
9437 movl(Address(z, kdx, Address::times_4, 0), carry);
9438 shrq(carry, 32);
9439 subl(kdx, 1);
9440
9441 bind(L_carry);
9442 movl(Address(z, kdx, Address::times_4, 0), carry);
9443
9444 // Second and third (nested) loops.
9445 //
9446 // for (int i = xstart-1; i >= 0; i--) { // Second loop
9447 // carry = 0;
9448 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
9449 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
9450 // (z[k] & LONG_MASK) + carry;
9451 // z[k] = (int)product;
9452 // carry = product >>> 32;
9453 // }
9454 // z[i] = (int)carry;
9455 // }
9456 //
9457 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
9458
9459 const Register jdx = tmp1;
9460
9461 bind(L_second_loop);
9462 xorl(carry, carry); // carry = 0;
9463 movl(jdx, ylen); // j = ystart+1
9464
9465 subl(xstart, 1); // i = xstart-1;
9466 jcc(Assembler::negative, L_done);
9467
9468 push (z);
9469
9470 Label L_last_x;
9471 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j
9472 subl(xstart, 1); // i = xstart-1;
9473 jcc(Assembler::negative, L_last_x);
9474
9475 if (UseBMI2Instructions) {
9476 movq(rdx, Address(x, xstart, Address::times_4, 0));
9477 rorxq(rdx, rdx, 32); // convert big-endian to little-endian
9478 } else {
9479 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
9480 rorq(x_xstart, 32); // convert big-endian to little-endian
9481 }
9482
9483 Label L_third_loop_prologue;
9484 bind(L_third_loop_prologue);
9485
9486 push (x);
9487 push (xstart);
9488 push (ylen);
9489
9490
9491 if (UseBMI2Instructions) {
9492 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4);
9493 } else { // !UseBMI2Instructions
9494 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
9495 }
9496
9497 pop(ylen);
9498 pop(xlen);
9499 pop(x);
9500 pop(z);
9501
9502 movl(tmp3, xlen);
9503 addl(tmp3, 1);
9504 movl(Address(z, tmp3, Address::times_4, 0), carry);
9505 subl(tmp3, 1);
9506 jccb(Assembler::negative, L_done);
9507
9508 shrq(carry, 32);
9509 movl(Address(z, tmp3, Address::times_4, 0), carry);
9510 jmp(L_second_loop);
9511
9512 // Next infrequent code is moved outside loops.
9513 bind(L_last_x);
9514 if (UseBMI2Instructions) {
9515 movl(rdx, Address(x, 0));
9516 } else {
9517 movl(x_xstart, Address(x, 0));
9518 }
9519 jmp(L_third_loop_prologue);
9520
9521 bind(L_done);
9522
9523 pop(zlen);
9524 pop(xlen);
9525
9526 pop(tmp5);
9527 pop(tmp4);
9528 pop(tmp3);
9529 pop(tmp2);
9530 pop(tmp1);
9531 }
9532
9533 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale,
9534 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){
9535 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled.");
9536 Label VECTOR32_LOOP, VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP;
9537 Label VECTOR16_TAIL, VECTOR8_TAIL, VECTOR4_TAIL;
9538 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL;
9539 Label SAME_TILL_END, DONE;
9540 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL;
9541
9542 //scale is in rcx in both Win64 and Unix
9543 ShortBranchVerifier sbv(this);
9544
9545 shlq(length);
9546 xorq(result, result);
9547
9548 cmpq(length, 8);
9549 jcc(Assembler::equal, VECTOR8_LOOP);
9550 jcc(Assembler::less, VECTOR4_TAIL);
9551
9552 if (UseAVX >= 2){
9553
9554 cmpq(length, 16);
9555 jcc(Assembler::equal, VECTOR16_LOOP);
9556 jcc(Assembler::less, VECTOR8_LOOP);
9557
9558 cmpq(length, 32);
9559 jccb(Assembler::less, VECTOR16_TAIL);
9560
9561 subq(length, 32);
9562 bind(VECTOR32_LOOP);
9563 vmovdqu(rymm0, Address(obja, result));
9564 vmovdqu(rymm1, Address(objb, result));
9565 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit);
9566 vptest(rymm2, rymm2);
9567 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found
9568 addq(result, 32);
9569 subq(length, 32);
9570 jccb(Assembler::greaterEqual, VECTOR32_LOOP);
9571 addq(length, 32);
9572 jcc(Assembler::equal, SAME_TILL_END);
9573 //falling through if less than 32 bytes left //close the branch here.
9574
9575 bind(VECTOR16_TAIL);
9576 cmpq(length, 16);
9577 jccb(Assembler::less, VECTOR8_TAIL);
9578 bind(VECTOR16_LOOP);
9579 movdqu(rymm0, Address(obja, result));
9580 movdqu(rymm1, Address(objb, result));
9581 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit);
9582 ptest(rymm2, rymm2);
9583 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
9584 addq(result, 16);
9585 subq(length, 16);
9586 jcc(Assembler::equal, SAME_TILL_END);
9587 //falling through if less than 16 bytes left
9588 } else {//regular intrinsics
9589
9590 cmpq(length, 16);
9591 jccb(Assembler::less, VECTOR8_TAIL);
9592
9593 subq(length, 16);
9594 bind(VECTOR16_LOOP);
9595 movdqu(rymm0, Address(obja, result));
9596 movdqu(rymm1, Address(objb, result));
9597 pxor(rymm0, rymm1);
9598 ptest(rymm0, rymm0);
9599 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
9600 addq(result, 16);
9601 subq(length, 16);
9602 jccb(Assembler::greaterEqual, VECTOR16_LOOP);
9603 addq(length, 16);
9604 jcc(Assembler::equal, SAME_TILL_END);
9605 //falling through if less than 16 bytes left
9606 }
9607
9608 bind(VECTOR8_TAIL);
9609 cmpq(length, 8);
9610 jccb(Assembler::less, VECTOR4_TAIL);
9611 bind(VECTOR8_LOOP);
9612 movq(tmp1, Address(obja, result));
9613 movq(tmp2, Address(objb, result));
9614 xorq(tmp1, tmp2);
9615 testq(tmp1, tmp1);
9616 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found
9617 addq(result, 8);
9618 subq(length, 8);
9619 jcc(Assembler::equal, SAME_TILL_END);
9620 //falling through if less than 8 bytes left
9621
9622 bind(VECTOR4_TAIL);
9623 cmpq(length, 4);
9624 jccb(Assembler::less, BYTES_TAIL);
9625 bind(VECTOR4_LOOP);
9626 movl(tmp1, Address(obja, result));
9627 xorl(tmp1, Address(objb, result));
9628 testl(tmp1, tmp1);
9629 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found
9630 addq(result, 4);
9631 subq(length, 4);
9632 jcc(Assembler::equal, SAME_TILL_END);
9633 //falling through if less than 4 bytes left
9634
9635 bind(BYTES_TAIL);
9636 bind(BYTES_LOOP);
9637 load_unsigned_byte(tmp1, Address(obja, result));
9638 load_unsigned_byte(tmp2, Address(objb, result));
9639 xorl(tmp1, tmp2);
9640 testl(tmp1, tmp1);
9641 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
9642 decq(length);
9643 jccb(Assembler::zero, SAME_TILL_END);
9644 incq(result);
9645 load_unsigned_byte(tmp1, Address(obja, result));
9646 load_unsigned_byte(tmp2, Address(objb, result));
9647 xorl(tmp1, tmp2);
9648 testl(tmp1, tmp1);
9649 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
9650 decq(length);
9651 jccb(Assembler::zero, SAME_TILL_END);
9652 incq(result);
9653 load_unsigned_byte(tmp1, Address(obja, result));
9654 load_unsigned_byte(tmp2, Address(objb, result));
9655 xorl(tmp1, tmp2);
9656 testl(tmp1, tmp1);
9657 jccb(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
9658 jmpb(SAME_TILL_END);
9659
9660 if (UseAVX >= 2){
9661 bind(VECTOR32_NOT_EQUAL);
9662 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit);
9663 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit);
9664 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit);
9665 vpmovmskb(tmp1, rymm0);
9666 bsfq(tmp1, tmp1);
9667 addq(result, tmp1);
9668 shrq(result);
9669 jmpb(DONE);
9670 }
9671
9672 bind(VECTOR16_NOT_EQUAL);
9673 if (UseAVX >= 2){
9674 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit);
9675 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit);
9676 pxor(rymm0, rymm2);
9677 } else {
9678 pcmpeqb(rymm2, rymm2);
9679 pxor(rymm0, rymm1);
9680 pcmpeqb(rymm0, rymm1);
9681 pxor(rymm0, rymm2);
9682 }
9683 pmovmskb(tmp1, rymm0);
9684 bsfq(tmp1, tmp1);
9685 addq(result, tmp1);
9686 shrq(result);
9687 jmpb(DONE);
9688
9689 bind(VECTOR8_NOT_EQUAL);
9690 bind(VECTOR4_NOT_EQUAL);
9691 bsfq(tmp1, tmp1);
9692 shrq(tmp1, 3);
9693 addq(result, tmp1);
9694 bind(BYTES_NOT_EQUAL);
9695 shrq(result);
9696 jmpb(DONE);
9697
9698 bind(SAME_TILL_END);
9699 mov64(result, -1);
9700
9701 bind(DONE);
9702 }
9703
9704
9705 //Helper functions for square_to_len()
9706
9707 /**
9708 * Store the squares of x[], right shifted one bit (divided by 2) into z[]
9709 * Preserves x and z and modifies rest of the registers.
9710 */
9711 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
9712 // Perform square and right shift by 1
9713 // Handle odd xlen case first, then for even xlen do the following
9714 // jlong carry = 0;
9715 // for (int j=0, i=0; j < xlen; j+=2, i+=4) {
9716 // huge_128 product = x[j:j+1] * x[j:j+1];
9717 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65);
9718 // z[i+2:i+3] = (jlong)(product >>> 1);
9719 // carry = (jlong)product;
9720 // }
9721
9722 xorq(tmp5, tmp5); // carry
9723 xorq(rdxReg, rdxReg);
9724 xorl(tmp1, tmp1); // index for x
9725 xorl(tmp4, tmp4); // index for z
9726
9727 Label L_first_loop, L_first_loop_exit;
9728
9729 testl(xlen, 1);
9730 jccb(Assembler::zero, L_first_loop); //jump if xlen is even
9731
9732 // Square and right shift by 1 the odd element using 32 bit multiply
9733 movl(raxReg, Address(x, tmp1, Address::times_4, 0));
9734 imulq(raxReg, raxReg);
9735 shrq(raxReg, 1);
9736 adcq(tmp5, 0);
9737 movq(Address(z, tmp4, Address::times_4, 0), raxReg);
9738 incrementl(tmp1);
9739 addl(tmp4, 2);
9740
9741 // Square and right shift by 1 the rest using 64 bit multiply
9742 bind(L_first_loop);
9743 cmpptr(tmp1, xlen);
9744 jccb(Assembler::equal, L_first_loop_exit);
9745
9746 // Square
9747 movq(raxReg, Address(x, tmp1, Address::times_4, 0));
9748 rorq(raxReg, 32); // convert big-endian to little-endian
9749 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax
9750
9751 // Right shift by 1 and save carry
9752 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1
9753 rcrq(rdxReg, 1);
9754 rcrq(raxReg, 1);
9755 adcq(tmp5, 0);
9756
9757 // Store result in z
9758 movq(Address(z, tmp4, Address::times_4, 0), rdxReg);
9759 movq(Address(z, tmp4, Address::times_4, 8), raxReg);
9760
9761 // Update indices for x and z
9762 addl(tmp1, 2);
9763 addl(tmp4, 4);
9764 jmp(L_first_loop);
9765
9766 bind(L_first_loop_exit);
9767 }
9768
9769
9770 /**
9771 * Perform the following multiply add operation using BMI2 instructions
9772 * carry:sum = sum + op1*op2 + carry
9773 * op2 should be in rdx
9774 * op2 is preserved, all other registers are modified
9775 */
9776 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) {
9777 // assert op2 is rdx
9778 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1
9779 addq(sum, carry);
9780 adcq(tmp2, 0);
9781 addq(sum, op1);
9782 adcq(tmp2, 0);
9783 movq(carry, tmp2);
9784 }
9785
9786 /**
9787 * Perform the following multiply add operation:
9788 * carry:sum = sum + op1*op2 + carry
9789 * Preserves op1, op2 and modifies rest of registers
9790 */
9791 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) {
9792 // rdx:rax = op1 * op2
9793 movq(raxReg, op2);
9794 mulq(op1);
9795
9796 // rdx:rax = sum + carry + rdx:rax
9797 addq(sum, carry);
9798 adcq(rdxReg, 0);
9799 addq(sum, raxReg);
9800 adcq(rdxReg, 0);
9801
9802 // carry:sum = rdx:sum
9803 movq(carry, rdxReg);
9804 }
9805
9806 /**
9807 * Add 64 bit long carry into z[] with carry propogation.
9808 * Preserves z and carry register values and modifies rest of registers.
9809 *
9810 */
9811 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) {
9812 Label L_fourth_loop, L_fourth_loop_exit;
9813
9814 movl(tmp1, 1);
9815 subl(zlen, 2);
9816 addq(Address(z, zlen, Address::times_4, 0), carry);
9817
9818 bind(L_fourth_loop);
9819 jccb(Assembler::carryClear, L_fourth_loop_exit);
9820 subl(zlen, 2);
9821 jccb(Assembler::negative, L_fourth_loop_exit);
9822 addq(Address(z, zlen, Address::times_4, 0), tmp1);
9823 jmp(L_fourth_loop);
9824 bind(L_fourth_loop_exit);
9825 }
9826
9827 /**
9828 * Shift z[] left by 1 bit.
9829 * Preserves x, len, z and zlen registers and modifies rest of the registers.
9830 *
9831 */
9832 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) {
9833
9834 Label L_fifth_loop, L_fifth_loop_exit;
9835
9836 // Fifth loop
9837 // Perform primitiveLeftShift(z, zlen, 1)
9838
9839 const Register prev_carry = tmp1;
9840 const Register new_carry = tmp4;
9841 const Register value = tmp2;
9842 const Register zidx = tmp3;
9843
9844 // int zidx, carry;
9845 // long value;
9846 // carry = 0;
9847 // for (zidx = zlen-2; zidx >=0; zidx -= 2) {
9848 // (carry:value) = (z[i] << 1) | carry ;
9849 // z[i] = value;
9850 // }
9851
9852 movl(zidx, zlen);
9853 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register
9854
9855 bind(L_fifth_loop);
9856 decl(zidx); // Use decl to preserve carry flag
9857 decl(zidx);
9858 jccb(Assembler::negative, L_fifth_loop_exit);
9859
9860 if (UseBMI2Instructions) {
9861 movq(value, Address(z, zidx, Address::times_4, 0));
9862 rclq(value, 1);
9863 rorxq(value, value, 32);
9864 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
9865 }
9866 else {
9867 // clear new_carry
9868 xorl(new_carry, new_carry);
9869
9870 // Shift z[i] by 1, or in previous carry and save new carry
9871 movq(value, Address(z, zidx, Address::times_4, 0));
9872 shlq(value, 1);
9873 adcl(new_carry, 0);
9874
9875 orq(value, prev_carry);
9876 rorq(value, 0x20);
9877 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
9878
9879 // Set previous carry = new carry
9880 movl(prev_carry, new_carry);
9881 }
9882 jmp(L_fifth_loop);
9883
9884 bind(L_fifth_loop_exit);
9885 }
9886
9887
9888 /**
9889 * Code for BigInteger::squareToLen() intrinsic
9890 *
9891 * rdi: x
9892 * rsi: len
9893 * r8: z
9894 * rcx: zlen
9895 * r12: tmp1
9896 * r13: tmp2
9897 * r14: tmp3
9898 * r15: tmp4
9899 * rbx: tmp5
9900 *
9901 */
9902 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) {
9903
9904 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;
9905 push(tmp1);
9906 push(tmp2);
9907 push(tmp3);
9908 push(tmp4);
9909 push(tmp5);
9910
9911 // First loop
9912 // Store the squares, right shifted one bit (i.e., divided by 2).
9913 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg);
9914
9915 // Add in off-diagonal sums.
9916 //
9917 // Second, third (nested) and fourth loops.
9918 // zlen +=2;
9919 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) {
9920 // carry = 0;
9921 // long op2 = x[xidx:xidx+1];
9922 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) {
9923 // k -= 2;
9924 // long op1 = x[j:j+1];
9925 // long sum = z[k:k+1];
9926 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs);
9927 // z[k:k+1] = sum;
9928 // }
9929 // add_one_64(z, k, carry, tmp_regs);
9930 // }
9931
9932 const Register carry = tmp5;
9933 const Register sum = tmp3;
9934 const Register op1 = tmp4;
9935 Register op2 = tmp2;
9936
9937 push(zlen);
9938 push(len);
9939 addl(zlen,2);
9940 bind(L_second_loop);
9941 xorq(carry, carry);
9942 subl(zlen, 4);
9943 subl(len, 2);
9944 push(zlen);
9945 push(len);
9946 cmpl(len, 0);
9947 jccb(Assembler::lessEqual, L_second_loop_exit);
9948
9949 // Multiply an array by one 64 bit long.
9950 if (UseBMI2Instructions) {
9951 op2 = rdxReg;
9952 movq(op2, Address(x, len, Address::times_4, 0));
9953 rorxq(op2, op2, 32);
9954 }
9955 else {
9956 movq(op2, Address(x, len, Address::times_4, 0));
9957 rorq(op2, 32);
9958 }
9959
9960 bind(L_third_loop);
9961 decrementl(len);
9962 jccb(Assembler::negative, L_third_loop_exit);
9963 decrementl(len);
9964 jccb(Assembler::negative, L_last_x);
9965
9966 movq(op1, Address(x, len, Address::times_4, 0));
9967 rorq(op1, 32);
9968
9969 bind(L_multiply);
9970 subl(zlen, 2);
9971 movq(sum, Address(z, zlen, Address::times_4, 0));
9972
9973 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry.
9974 if (UseBMI2Instructions) {
9975 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2);
9976 }
9977 else {
9978 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9979 }
9980
9981 movq(Address(z, zlen, Address::times_4, 0), sum);
9982
9983 jmp(L_third_loop);
9984 bind(L_third_loop_exit);
9985
9986 // Fourth loop
9987 // Add 64 bit long carry into z with carry propogation.
9988 // Uses offsetted zlen.
9989 add_one_64(z, zlen, carry, tmp1);
9990
9991 pop(len);
9992 pop(zlen);
9993 jmp(L_second_loop);
9994
9995 // Next infrequent code is moved outside loops.
9996 bind(L_last_x);
9997 movl(op1, Address(x, 0));
9998 jmp(L_multiply);
9999
10000 bind(L_second_loop_exit);
10001 pop(len);
10002 pop(zlen);
10003 pop(len);
10004 pop(zlen);
10005
10006 // Fifth loop
10007 // Shift z left 1 bit.
10008 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4);
10009
10010 // z[zlen-1] |= x[len-1] & 1;
10011 movl(tmp3, Address(x, len, Address::times_4, -4));
10012 andl(tmp3, 1);
10013 orl(Address(z, zlen, Address::times_4, -4), tmp3);
10014
10015 pop(tmp5);
10016 pop(tmp4);
10017 pop(tmp3);
10018 pop(tmp2);
10019 pop(tmp1);
10020 }
10021
10022 /**
10023 * Helper function for mul_add()
10024 * Multiply the in[] by int k and add to out[] starting at offset offs using
10025 * 128 bit by 32 bit multiply and return the carry in tmp5.
10026 * Only quad int aligned length of in[] is operated on in this function.
10027 * k is in rdxReg for BMI2Instructions, for others it is in tmp2.
10028 * This function preserves out, in and k registers.
10029 * len and offset point to the appropriate index in "in" & "out" correspondingly
10030 * tmp5 has the carry.
10031 * other registers are temporary and are modified.
10032 *
10033 */
10034 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in,
10035 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3,
10036 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
10037
10038 Label L_first_loop, L_first_loop_exit;
10039
10040 movl(tmp1, len);
10041 shrl(tmp1, 2);
10042
10043 bind(L_first_loop);
10044 subl(tmp1, 1);
10045 jccb(Assembler::negative, L_first_loop_exit);
10046
10047 subl(len, 4);
10048 subl(offset, 4);
10049
10050 Register op2 = tmp2;
10051 const Register sum = tmp3;
10052 const Register op1 = tmp4;
10053 const Register carry = tmp5;
10054
10055 if (UseBMI2Instructions) {
10056 op2 = rdxReg;
10057 }
10058
10059 movq(op1, Address(in, len, Address::times_4, 8));
10060 rorq(op1, 32);
10061 movq(sum, Address(out, offset, Address::times_4, 8));
10062 rorq(sum, 32);
10063 if (UseBMI2Instructions) {
10064 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
10065 }
10066 else {
10067 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
10068 }
10069 // Store back in big endian from little endian
10070 rorq(sum, 0x20);
10071 movq(Address(out, offset, Address::times_4, 8), sum);
10072
10073 movq(op1, Address(in, len, Address::times_4, 0));
10074 rorq(op1, 32);
10075 movq(sum, Address(out, offset, Address::times_4, 0));
10076 rorq(sum, 32);
10077 if (UseBMI2Instructions) {
10078 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
10079 }
10080 else {
10081 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
10082 }
10083 // Store back in big endian from little endian
10084 rorq(sum, 0x20);
10085 movq(Address(out, offset, Address::times_4, 0), sum);
10086
10087 jmp(L_first_loop);
10088 bind(L_first_loop_exit);
10089 }
10090
10091 /**
10092 * Code for BigInteger::mulAdd() intrinsic
10093 *
10094 * rdi: out
10095 * rsi: in
10096 * r11: offs (out.length - offset)
10097 * rcx: len
10098 * r8: k
10099 * r12: tmp1
10100 * r13: tmp2
10101 * r14: tmp3
10102 * r15: tmp4
10103 * rbx: tmp5
10104 * Multiply the in[] by word k and add to out[], return the carry in rax
10105 */
10106 void MacroAssembler::mul_add(Register out, Register in, Register offs,
10107 Register len, Register k, Register tmp1, Register tmp2, Register tmp3,
10108 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
10109
10110 Label L_carry, L_last_in, L_done;
10111
10112 // carry = 0;
10113 // for (int j=len-1; j >= 0; j--) {
10114 // long product = (in[j] & LONG_MASK) * kLong +
10115 // (out[offs] & LONG_MASK) + carry;
10116 // out[offs--] = (int)product;
10117 // carry = product >>> 32;
10118 // }
10119 //
10120 push(tmp1);
10121 push(tmp2);
10122 push(tmp3);
10123 push(tmp4);
10124 push(tmp5);
10125
10126 Register op2 = tmp2;
10127 const Register sum = tmp3;
10128 const Register op1 = tmp4;
10129 const Register carry = tmp5;
10130
10131 if (UseBMI2Instructions) {
10132 op2 = rdxReg;
10133 movl(op2, k);
10134 }
10135 else {
10136 movl(op2, k);
10137 }
10138
10139 xorq(carry, carry);
10140
10141 //First loop
10142
10143 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply
10144 //The carry is in tmp5
10145 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg);
10146
10147 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any
10148 decrementl(len);
10149 jccb(Assembler::negative, L_carry);
10150 decrementl(len);
10151 jccb(Assembler::negative, L_last_in);
10152
10153 movq(op1, Address(in, len, Address::times_4, 0));
10154 rorq(op1, 32);
10155
10156 subl(offs, 2);
10157 movq(sum, Address(out, offs, Address::times_4, 0));
10158 rorq(sum, 32);
10159
10160 if (UseBMI2Instructions) {
10161 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
10162 }
10163 else {
10164 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
10165 }
10166
10167 // Store back in big endian from little endian
10168 rorq(sum, 0x20);
10169 movq(Address(out, offs, Address::times_4, 0), sum);
10170
10171 testl(len, len);
10172 jccb(Assembler::zero, L_carry);
10173
10174 //Multiply the last in[] entry, if any
10175 bind(L_last_in);
10176 movl(op1, Address(in, 0));
10177 movl(sum, Address(out, offs, Address::times_4, -4));
10178
10179 movl(raxReg, k);
10180 mull(op1); //tmp4 * eax -> edx:eax
10181 addl(sum, carry);
10182 adcl(rdxReg, 0);
10183 addl(sum, raxReg);
10184 adcl(rdxReg, 0);
10185 movl(carry, rdxReg);
10186
10187 movl(Address(out, offs, Address::times_4, -4), sum);
10188
10189 bind(L_carry);
10190 //return tmp5/carry as carry in rax
10191 movl(rax, carry);
10192
10193 bind(L_done);
10194 pop(tmp5);
10195 pop(tmp4);
10196 pop(tmp3);
10197 pop(tmp2);
10198 pop(tmp1);
10199 }
10200 #endif
10201
10202 /**
10203 * Emits code to update CRC-32 with a byte value according to constants in table
10204 *
10205 * @param [in,out]crc Register containing the crc.
10206 * @param [in]val Register containing the byte to fold into the CRC.
10207 * @param [in]table Register containing the table of crc constants.
10208 *
10209 * uint32_t crc;
10210 * val = crc_table[(val ^ crc) & 0xFF];
10211 * crc = val ^ (crc >> 8);
10212 *
10213 */
10214 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
10215 xorl(val, crc);
10216 andl(val, 0xFF);
10217 shrl(crc, 8); // unsigned shift
10218 xorl(crc, Address(table, val, Address::times_4, 0));
10219 }
10220
10221 /**
10222 * Fold 128-bit data chunk
10223 */
10224 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
10225 if (UseAVX > 0) {
10226 vpclmulhdq(xtmp, xK, xcrc); // [123:64]
10227 vpclmulldq(xcrc, xK, xcrc); // [63:0]
10228 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */);
10229 pxor(xcrc, xtmp);
10230 } else {
10231 movdqa(xtmp, xcrc);
10232 pclmulhdq(xtmp, xK); // [123:64]
10233 pclmulldq(xcrc, xK); // [63:0]
10234 pxor(xcrc, xtmp);
10235 movdqu(xtmp, Address(buf, offset));
10236 pxor(xcrc, xtmp);
10237 }
10238 }
10239
10240 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
10241 if (UseAVX > 0) {
10242 vpclmulhdq(xtmp, xK, xcrc);
10243 vpclmulldq(xcrc, xK, xcrc);
10244 pxor(xcrc, xbuf);
10245 pxor(xcrc, xtmp);
10246 } else {
10247 movdqa(xtmp, xcrc);
10248 pclmulhdq(xtmp, xK);
10249 pclmulldq(xcrc, xK);
10250 pxor(xcrc, xbuf);
10251 pxor(xcrc, xtmp);
10252 }
10253 }
10254
10255 /**
10256 * 8-bit folds to compute 32-bit CRC
10257 *
10258 * uint64_t xcrc;
10259 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8);
10260 */
10261 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) {
10262 movdl(tmp, xcrc);
10263 andl(tmp, 0xFF);
10264 movdl(xtmp, Address(table, tmp, Address::times_4, 0));
10265 psrldq(xcrc, 1); // unsigned shift one byte
10266 pxor(xcrc, xtmp);
10267 }
10268
10269 /**
10270 * uint32_t crc;
10271 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
10272 */
10273 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
10274 movl(tmp, crc);
10275 andl(tmp, 0xFF);
10276 shrl(crc, 8);
10277 xorl(crc, Address(table, tmp, Address::times_4, 0));
10278 }
10279
10280 /**
10281 * @param crc register containing existing CRC (32-bit)
10282 * @param buf register pointing to input byte buffer (byte*)
10283 * @param len register containing number of bytes
10284 * @param table register that will contain address of CRC table
10285 * @param tmp scratch register
10286 */
10287 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) {
10288 assert_different_registers(crc, buf, len, table, tmp, rax);
10289
10290 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned;
10291 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop;
10292
10293 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge
10294 // context for the registers used, where all instructions below are using 128-bit mode
10295 // On EVEX without VL and BW, these instructions will all be AVX.
10296 if (VM_Version::supports_avx512vlbw()) {
10297 movl(tmp, 0xffff);
10298 kmovwl(k1, tmp);
10299 }
10300
10301 lea(table, ExternalAddress(StubRoutines::crc_table_addr()));
10302 notl(crc); // ~crc
10303 cmpl(len, 16);
10304 jcc(Assembler::less, L_tail);
10305
10306 // Align buffer to 16 bytes
10307 movl(tmp, buf);
10308 andl(tmp, 0xF);
10309 jccb(Assembler::zero, L_aligned);
10310 subl(tmp, 16);
10311 addl(len, tmp);
10312
10313 align(4);
10314 BIND(L_align_loop);
10315 movsbl(rax, Address(buf, 0)); // load byte with sign extension
10316 update_byte_crc32(crc, rax, table);
10317 increment(buf);
10318 incrementl(tmp);
10319 jccb(Assembler::less, L_align_loop);
10320
10321 BIND(L_aligned);
10322 movl(tmp, len); // save
10323 shrl(len, 4);
10324 jcc(Assembler::zero, L_tail_restore);
10325
10326 // Fold crc into first bytes of vector
10327 movdqa(xmm1, Address(buf, 0));
10328 movdl(rax, xmm1);
10329 xorl(crc, rax);
10330 pinsrd(xmm1, crc, 0);
10331 addptr(buf, 16);
10332 subl(len, 4); // len > 0
10333 jcc(Assembler::less, L_fold_tail);
10334
10335 movdqa(xmm2, Address(buf, 0));
10336 movdqa(xmm3, Address(buf, 16));
10337 movdqa(xmm4, Address(buf, 32));
10338 addptr(buf, 48);
10339 subl(len, 3);
10340 jcc(Assembler::lessEqual, L_fold_512b);
10341
10342 // Fold total 512 bits of polynomial on each iteration,
10343 // 128 bits per each of 4 parallel streams.
10344 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
10345
10346 align(32);
10347 BIND(L_fold_512b_loop);
10348 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
10349 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16);
10350 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32);
10351 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48);
10352 addptr(buf, 64);
10353 subl(len, 4);
10354 jcc(Assembler::greater, L_fold_512b_loop);
10355
10356 // Fold 512 bits to 128 bits.
10357 BIND(L_fold_512b);
10358 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
10359 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2);
10360 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3);
10361 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4);
10362
10363 // Fold the rest of 128 bits data chunks
10364 BIND(L_fold_tail);
10365 addl(len, 3);
10366 jccb(Assembler::lessEqual, L_fold_128b);
10367 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
10368
10369 BIND(L_fold_tail_loop);
10370 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
10371 addptr(buf, 16);
10372 decrementl(len);
10373 jccb(Assembler::greater, L_fold_tail_loop);
10374
10375 // Fold 128 bits in xmm1 down into 32 bits in crc register.
10376 BIND(L_fold_128b);
10377 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
10378 if (UseAVX > 0) {
10379 vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
10380 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */);
10381 vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
10382 } else {
10383 movdqa(xmm2, xmm0);
10384 pclmulqdq(xmm2, xmm1, 0x1);
10385 movdqa(xmm3, xmm0);
10386 pand(xmm3, xmm2);
10387 pclmulqdq(xmm0, xmm3, 0x1);
10388 }
10389 psrldq(xmm1, 8);
10390 psrldq(xmm2, 4);
10391 pxor(xmm0, xmm1);
10392 pxor(xmm0, xmm2);
10393
10394 // 8 8-bit folds to compute 32-bit CRC.
10395 for (int j = 0; j < 4; j++) {
10396 fold_8bit_crc32(xmm0, table, xmm1, rax);
10397 }
10398 movdl(crc, xmm0); // mov 32 bits to general register
10399 for (int j = 0; j < 4; j++) {
10400 fold_8bit_crc32(crc, table, rax);
10401 }
10402
10403 BIND(L_tail_restore);
10404 movl(len, tmp); // restore
10405 BIND(L_tail);
10406 andl(len, 0xf);
10407 jccb(Assembler::zero, L_exit);
10408
10409 // Fold the rest of bytes
10410 align(4);
10411 BIND(L_tail_loop);
10412 movsbl(rax, Address(buf, 0)); // load byte with sign extension
10413 update_byte_crc32(crc, rax, table);
10414 increment(buf);
10415 decrementl(len);
10416 jccb(Assembler::greater, L_tail_loop);
10417
10418 BIND(L_exit);
10419 notl(crc); // ~c
10420 }
10421
10422 #ifdef _LP64
10423 // S. Gueron / Information Processing Letters 112 (2012) 184
10424 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table.
10425 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0].
10426 // Output: the 64-bit carry-less product of B * CONST
10427 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n,
10428 Register tmp1, Register tmp2, Register tmp3) {
10429 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
10430 if (n > 0) {
10431 addq(tmp3, n * 256 * 8);
10432 }
10433 // Q1 = TABLEExt[n][B & 0xFF];
10434 movl(tmp1, in);
10435 andl(tmp1, 0x000000FF);
10436 shll(tmp1, 3);
10437 addq(tmp1, tmp3);
10438 movq(tmp1, Address(tmp1, 0));
10439
10440 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
10441 movl(tmp2, in);
10442 shrl(tmp2, 8);
10443 andl(tmp2, 0x000000FF);
10444 shll(tmp2, 3);
10445 addq(tmp2, tmp3);
10446 movq(tmp2, Address(tmp2, 0));
10447
10448 shlq(tmp2, 8);
10449 xorq(tmp1, tmp2);
10450
10451 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
10452 movl(tmp2, in);
10453 shrl(tmp2, 16);
10454 andl(tmp2, 0x000000FF);
10455 shll(tmp2, 3);
10456 addq(tmp2, tmp3);
10457 movq(tmp2, Address(tmp2, 0));
10458
10459 shlq(tmp2, 16);
10460 xorq(tmp1, tmp2);
10461
10462 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
10463 shrl(in, 24);
10464 andl(in, 0x000000FF);
10465 shll(in, 3);
10466 addq(in, tmp3);
10467 movq(in, Address(in, 0));
10468
10469 shlq(in, 24);
10470 xorq(in, tmp1);
10471 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
10472 }
10473
10474 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
10475 Register in_out,
10476 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
10477 XMMRegister w_xtmp2,
10478 Register tmp1,
10479 Register n_tmp2, Register n_tmp3) {
10480 if (is_pclmulqdq_supported) {
10481 movdl(w_xtmp1, in_out); // modified blindly
10482
10483 movl(tmp1, const_or_pre_comp_const_index);
10484 movdl(w_xtmp2, tmp1);
10485 pclmulqdq(w_xtmp1, w_xtmp2, 0);
10486
10487 movdq(in_out, w_xtmp1);
10488 } else {
10489 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3);
10490 }
10491 }
10492
10493 // Recombination Alternative 2: No bit-reflections
10494 // T1 = (CRC_A * U1) << 1
10495 // T2 = (CRC_B * U2) << 1
10496 // C1 = T1 >> 32
10497 // C2 = T2 >> 32
10498 // T1 = T1 & 0xFFFFFFFF
10499 // T2 = T2 & 0xFFFFFFFF
10500 // T1 = CRC32(0, T1)
10501 // T2 = CRC32(0, T2)
10502 // C1 = C1 ^ T1
10503 // C2 = C2 ^ T2
10504 // CRC = C1 ^ C2 ^ CRC_C
10505 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,
10506 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10507 Register tmp1, Register tmp2,
10508 Register n_tmp3) {
10509 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10510 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10511 shlq(in_out, 1);
10512 movl(tmp1, in_out);
10513 shrq(in_out, 32);
10514 xorl(tmp2, tmp2);
10515 crc32(tmp2, tmp1, 4);
10516 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here
10517 shlq(in1, 1);
10518 movl(tmp1, in1);
10519 shrq(in1, 32);
10520 xorl(tmp2, tmp2);
10521 crc32(tmp2, tmp1, 4);
10522 xorl(in1, tmp2);
10523 xorl(in_out, in1);
10524 xorl(in_out, in2);
10525 }
10526
10527 // Set N to predefined value
10528 // Subtract from a lenght of a buffer
10529 // execute in a loop:
10530 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0
10531 // for i = 1 to N do
10532 // CRC_A = CRC32(CRC_A, A[i])
10533 // CRC_B = CRC32(CRC_B, B[i])
10534 // CRC_C = CRC32(CRC_C, C[i])
10535 // end for
10536 // Recombine
10537 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,
10538 Register in_out1, Register in_out2, Register in_out3,
10539 Register tmp1, Register tmp2, Register tmp3,
10540 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10541 Register tmp4, Register tmp5,
10542 Register n_tmp6) {
10543 Label L_processPartitions;
10544 Label L_processPartition;
10545 Label L_exit;
10546
10547 bind(L_processPartitions);
10548 cmpl(in_out1, 3 * size);
10549 jcc(Assembler::less, L_exit);
10550 xorl(tmp1, tmp1);
10551 xorl(tmp2, tmp2);
10552 movq(tmp3, in_out2);
10553 addq(tmp3, size);
10554
10555 bind(L_processPartition);
10556 crc32(in_out3, Address(in_out2, 0), 8);
10557 crc32(tmp1, Address(in_out2, size), 8);
10558 crc32(tmp2, Address(in_out2, size * 2), 8);
10559 addq(in_out2, 8);
10560 cmpq(in_out2, tmp3);
10561 jcc(Assembler::less, L_processPartition);
10562 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
10563 w_xtmp1, w_xtmp2, w_xtmp3,
10564 tmp4, tmp5,
10565 n_tmp6);
10566 addq(in_out2, 2 * size);
10567 subl(in_out1, 3 * size);
10568 jmp(L_processPartitions);
10569
10570 bind(L_exit);
10571 }
10572 #else
10573 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n,
10574 Register tmp1, Register tmp2, Register tmp3,
10575 XMMRegister xtmp1, XMMRegister xtmp2) {
10576 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
10577 if (n > 0) {
10578 addl(tmp3, n * 256 * 8);
10579 }
10580 // Q1 = TABLEExt[n][B & 0xFF];
10581 movl(tmp1, in_out);
10582 andl(tmp1, 0x000000FF);
10583 shll(tmp1, 3);
10584 addl(tmp1, tmp3);
10585 movq(xtmp1, Address(tmp1, 0));
10586
10587 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
10588 movl(tmp2, in_out);
10589 shrl(tmp2, 8);
10590 andl(tmp2, 0x000000FF);
10591 shll(tmp2, 3);
10592 addl(tmp2, tmp3);
10593 movq(xtmp2, Address(tmp2, 0));
10594
10595 psllq(xtmp2, 8);
10596 pxor(xtmp1, xtmp2);
10597
10598 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
10599 movl(tmp2, in_out);
10600 shrl(tmp2, 16);
10601 andl(tmp2, 0x000000FF);
10602 shll(tmp2, 3);
10603 addl(tmp2, tmp3);
10604 movq(xtmp2, Address(tmp2, 0));
10605
10606 psllq(xtmp2, 16);
10607 pxor(xtmp1, xtmp2);
10608
10609 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
10610 shrl(in_out, 24);
10611 andl(in_out, 0x000000FF);
10612 shll(in_out, 3);
10613 addl(in_out, tmp3);
10614 movq(xtmp2, Address(in_out, 0));
10615
10616 psllq(xtmp2, 24);
10617 pxor(xtmp1, xtmp2); // Result in CXMM
10618 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
10619 }
10620
10621 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
10622 Register in_out,
10623 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
10624 XMMRegister w_xtmp2,
10625 Register tmp1,
10626 Register n_tmp2, Register n_tmp3) {
10627 if (is_pclmulqdq_supported) {
10628 movdl(w_xtmp1, in_out);
10629
10630 movl(tmp1, const_or_pre_comp_const_index);
10631 movdl(w_xtmp2, tmp1);
10632 pclmulqdq(w_xtmp1, w_xtmp2, 0);
10633 // Keep result in XMM since GPR is 32 bit in length
10634 } else {
10635 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2);
10636 }
10637 }
10638
10639 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,
10640 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10641 Register tmp1, Register tmp2,
10642 Register n_tmp3) {
10643 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10644 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10645
10646 psllq(w_xtmp1, 1);
10647 movdl(tmp1, w_xtmp1);
10648 psrlq(w_xtmp1, 32);
10649 movdl(in_out, w_xtmp1);
10650
10651 xorl(tmp2, tmp2);
10652 crc32(tmp2, tmp1, 4);
10653 xorl(in_out, tmp2);
10654
10655 psllq(w_xtmp2, 1);
10656 movdl(tmp1, w_xtmp2);
10657 psrlq(w_xtmp2, 32);
10658 movdl(in1, w_xtmp2);
10659
10660 xorl(tmp2, tmp2);
10661 crc32(tmp2, tmp1, 4);
10662 xorl(in1, tmp2);
10663 xorl(in_out, in1);
10664 xorl(in_out, in2);
10665 }
10666
10667 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,
10668 Register in_out1, Register in_out2, Register in_out3,
10669 Register tmp1, Register tmp2, Register tmp3,
10670 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10671 Register tmp4, Register tmp5,
10672 Register n_tmp6) {
10673 Label L_processPartitions;
10674 Label L_processPartition;
10675 Label L_exit;
10676
10677 bind(L_processPartitions);
10678 cmpl(in_out1, 3 * size);
10679 jcc(Assembler::less, L_exit);
10680 xorl(tmp1, tmp1);
10681 xorl(tmp2, tmp2);
10682 movl(tmp3, in_out2);
10683 addl(tmp3, size);
10684
10685 bind(L_processPartition);
10686 crc32(in_out3, Address(in_out2, 0), 4);
10687 crc32(tmp1, Address(in_out2, size), 4);
10688 crc32(tmp2, Address(in_out2, size*2), 4);
10689 crc32(in_out3, Address(in_out2, 0+4), 4);
10690 crc32(tmp1, Address(in_out2, size+4), 4);
10691 crc32(tmp2, Address(in_out2, size*2+4), 4);
10692 addl(in_out2, 8);
10693 cmpl(in_out2, tmp3);
10694 jcc(Assembler::less, L_processPartition);
10695
10696 push(tmp3);
10697 push(in_out1);
10698 push(in_out2);
10699 tmp4 = tmp3;
10700 tmp5 = in_out1;
10701 n_tmp6 = in_out2;
10702
10703 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
10704 w_xtmp1, w_xtmp2, w_xtmp3,
10705 tmp4, tmp5,
10706 n_tmp6);
10707
10708 pop(in_out2);
10709 pop(in_out1);
10710 pop(tmp3);
10711
10712 addl(in_out2, 2 * size);
10713 subl(in_out1, 3 * size);
10714 jmp(L_processPartitions);
10715
10716 bind(L_exit);
10717 }
10718 #endif //LP64
10719
10720 #ifdef _LP64
10721 // Algorithm 2: Pipelined usage of the CRC32 instruction.
10722 // Input: A buffer I of L bytes.
10723 // Output: the CRC32C value of the buffer.
10724 // Notations:
10725 // Write L = 24N + r, with N = floor (L/24).
10726 // r = L mod 24 (0 <= r < 24).
10727 // Consider I as the concatenation of A|B|C|R, where A, B, C, each,
10728 // N quadwords, and R consists of r bytes.
10729 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1
10730 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1
10731 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1
10732 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1
10733 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
10734 Register tmp1, Register tmp2, Register tmp3,
10735 Register tmp4, Register tmp5, Register tmp6,
10736 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10737 bool is_pclmulqdq_supported) {
10738 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
10739 Label L_wordByWord;
10740 Label L_byteByByteProlog;
10741 Label L_byteByByte;
10742 Label L_exit;
10743
10744 if (is_pclmulqdq_supported ) {
10745 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
10746 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1);
10747
10748 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
10749 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
10750
10751 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
10752 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
10753 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\"");
10754 } else {
10755 const_or_pre_comp_const_index[0] = 1;
10756 const_or_pre_comp_const_index[1] = 0;
10757
10758 const_or_pre_comp_const_index[2] = 3;
10759 const_or_pre_comp_const_index[3] = 2;
10760
10761 const_or_pre_comp_const_index[4] = 5;
10762 const_or_pre_comp_const_index[5] = 4;
10763 }
10764 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
10765 in2, in1, in_out,
10766 tmp1, tmp2, tmp3,
10767 w_xtmp1, w_xtmp2, w_xtmp3,
10768 tmp4, tmp5,
10769 tmp6);
10770 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
10771 in2, in1, in_out,
10772 tmp1, tmp2, tmp3,
10773 w_xtmp1, w_xtmp2, w_xtmp3,
10774 tmp4, tmp5,
10775 tmp6);
10776 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
10777 in2, in1, in_out,
10778 tmp1, tmp2, tmp3,
10779 w_xtmp1, w_xtmp2, w_xtmp3,
10780 tmp4, tmp5,
10781 tmp6);
10782 movl(tmp1, in2);
10783 andl(tmp1, 0x00000007);
10784 negl(tmp1);
10785 addl(tmp1, in2);
10786 addq(tmp1, in1);
10787
10788 BIND(L_wordByWord);
10789 cmpq(in1, tmp1);
10790 jcc(Assembler::greaterEqual, L_byteByByteProlog);
10791 crc32(in_out, Address(in1, 0), 4);
10792 addq(in1, 4);
10793 jmp(L_wordByWord);
10794
10795 BIND(L_byteByByteProlog);
10796 andl(in2, 0x00000007);
10797 movl(tmp2, 1);
10798
10799 BIND(L_byteByByte);
10800 cmpl(tmp2, in2);
10801 jccb(Assembler::greater, L_exit);
10802 crc32(in_out, Address(in1, 0), 1);
10803 incq(in1);
10804 incl(tmp2);
10805 jmp(L_byteByByte);
10806
10807 BIND(L_exit);
10808 }
10809 #else
10810 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
10811 Register tmp1, Register tmp2, Register tmp3,
10812 Register tmp4, Register tmp5, Register tmp6,
10813 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10814 bool is_pclmulqdq_supported) {
10815 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
10816 Label L_wordByWord;
10817 Label L_byteByByteProlog;
10818 Label L_byteByByte;
10819 Label L_exit;
10820
10821 if (is_pclmulqdq_supported) {
10822 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
10823 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1);
10824
10825 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
10826 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
10827
10828 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
10829 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
10830 } else {
10831 const_or_pre_comp_const_index[0] = 1;
10832 const_or_pre_comp_const_index[1] = 0;
10833
10834 const_or_pre_comp_const_index[2] = 3;
10835 const_or_pre_comp_const_index[3] = 2;
10836
10837 const_or_pre_comp_const_index[4] = 5;
10838 const_or_pre_comp_const_index[5] = 4;
10839 }
10840 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
10841 in2, in1, in_out,
10842 tmp1, tmp2, tmp3,
10843 w_xtmp1, w_xtmp2, w_xtmp3,
10844 tmp4, tmp5,
10845 tmp6);
10846 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
10847 in2, in1, in_out,
10848 tmp1, tmp2, tmp3,
10849 w_xtmp1, w_xtmp2, w_xtmp3,
10850 tmp4, tmp5,
10851 tmp6);
10852 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
10853 in2, in1, in_out,
10854 tmp1, tmp2, tmp3,
10855 w_xtmp1, w_xtmp2, w_xtmp3,
10856 tmp4, tmp5,
10857 tmp6);
10858 movl(tmp1, in2);
10859 andl(tmp1, 0x00000007);
10860 negl(tmp1);
10861 addl(tmp1, in2);
10862 addl(tmp1, in1);
10863
10864 BIND(L_wordByWord);
10865 cmpl(in1, tmp1);
10866 jcc(Assembler::greaterEqual, L_byteByByteProlog);
10867 crc32(in_out, Address(in1,0), 4);
10868 addl(in1, 4);
10869 jmp(L_wordByWord);
10870
10871 BIND(L_byteByByteProlog);
10872 andl(in2, 0x00000007);
10873 movl(tmp2, 1);
10874
10875 BIND(L_byteByByte);
10876 cmpl(tmp2, in2);
10877 jccb(Assembler::greater, L_exit);
10878 movb(tmp1, Address(in1, 0));
10879 crc32(in_out, tmp1, 1);
10880 incl(in1);
10881 incl(tmp2);
10882 jmp(L_byteByByte);
10883
10884 BIND(L_exit);
10885 }
10886 #endif // LP64
10887 #undef BIND
10888 #undef BLOCK_COMMENT
10889
10890 // Compress char[] array to byte[].
10891 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java
10892 // @HotSpotIntrinsicCandidate
10893 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) {
10894 // for (int i = 0; i < len; i++) {
10895 // int c = src[srcOff++];
10896 // if (c >>> 8 != 0) {
10897 // return 0;
10898 // }
10899 // dst[dstOff++] = (byte)c;
10900 // }
10901 // return len;
10902 // }
10903 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
10904 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
10905 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
10906 Register tmp5, Register result) {
10907 Label copy_chars_loop, return_length, return_zero, done, below_threshold;
10908
10909 // rsi: src
10910 // rdi: dst
10911 // rdx: len
10912 // rcx: tmp5
10913 // rax: result
10914
10915 // rsi holds start addr of source char[] to be compressed
10916 // rdi holds start addr of destination byte[]
10917 // rdx holds length
10918
10919 assert(len != result, "");
10920
10921 // save length for return
10922 push(len);
10923
10924 if ((UseAVX > 2) && // AVX512
10925 VM_Version::supports_avx512vlbw() &&
10926 VM_Version::supports_bmi2()) {
10927
10928 set_vector_masking(); // opening of the stub context for programming mask registers
10929
10930 Label copy_32_loop, copy_loop_tail, copy_just_portion_of_candidates;
10931
10932 // alignement
10933 Label post_alignement;
10934
10935 // if length of the string is less than 16, handle it in an old fashioned
10936 // way
10937 testl(len, -32);
10938 jcc(Assembler::zero, below_threshold);
10939
10940 // First check whether a character is compressable ( <= 0xFF).
10941 // Create mask to test for Unicode chars inside zmm vector
10942 movl(result, 0x00FF);
10943 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit);
10944
10945 testl(len, -64);
10946 jcc(Assembler::zero, post_alignement);
10947
10948 // Save k1
10949 kmovql(k3, k1);
10950
10951 movl(tmp5, dst);
10952 andl(tmp5, (64 - 1));
10953 negl(tmp5);
10954 andl(tmp5, (64 - 1));
10955
10956 // bail out when there is nothing to be done
10957 testl(tmp5, 0xFFFFFFFF);
10958 jcc(Assembler::zero, post_alignement);
10959
10960 // ~(~0 << len), where len is the # of remaining elements to process
10961 movl(result, 0xFFFFFFFF);
10962 shlxl(result, result, tmp5);
10963 notl(result);
10964
10965 kmovdl(k1, result);
10966
10967 evmovdquw(tmp1Reg, k1, Address(src, 0), Assembler::AVX_512bit);
10968 evpcmpuw(k2, k1, tmp1Reg, tmp2Reg, ComparisonPredicate::le, Assembler::AVX_512bit);
10969 ktestd(k2, k1);
10970 jcc(Assembler::carryClear, copy_just_portion_of_candidates);
10971
10972 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit);
10973
10974 addptr(src, tmp5);
10975 addptr(src, tmp5);
10976 addptr(dst, tmp5);
10977 subl(len, tmp5);
10978
10979 bind(post_alignement);
10980 // end of alignement
10981
10982 movl(tmp5, len);
10983 andl(tmp5, (32 - 1)); // tail count (in chars)
10984 andl(len, ~(32 - 1)); // vector count (in chars)
10985 jcc(Assembler::zero, copy_loop_tail);
10986
10987 lea(src, Address(src, len, Address::times_2));
10988 lea(dst, Address(dst, len, Address::times_1));
10989 negptr(len);
10990
10991 bind(copy_32_loop);
10992 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit);
10993 evpcmpuw(k2, tmp1Reg, tmp2Reg, ComparisonPredicate::le, Assembler::AVX_512bit);
10994 kortestdl(k2, k2);
10995 jcc(Assembler::carryClear, copy_just_portion_of_candidates);
10996
10997 // All elements in current processed chunk are valid candidates for
10998 // compression. Write a truncated byte elements to the memory.
10999 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit);
11000 addptr(len, 32);
11001 jcc(Assembler::notZero, copy_32_loop);
11002
11003 bind(copy_loop_tail);
11004 // bail out when there is nothing to be done
11005 testl(tmp5, 0xFFFFFFFF);
11006 jcc(Assembler::zero, return_length);
11007
11008 // Save k1
11009 kmovql(k3, k1);
11010
11011 movl(len, tmp5);
11012
11013 // ~(~0 << len), where len is the # of remaining elements to process
11014 movl(result, 0xFFFFFFFF);
11015 shlxl(result, result, len);
11016 notl(result);
11017
11018 kmovdl(k1, result);
11019
11020 evmovdquw(tmp1Reg, k1, Address(src, 0), Assembler::AVX_512bit);
11021 evpcmpuw(k2, k1, tmp1Reg, tmp2Reg, ComparisonPredicate::le, Assembler::AVX_512bit);
11022 ktestd(k2, k1);
11023 jcc(Assembler::carryClear, copy_just_portion_of_candidates);
11024
11025 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit);
11026 // Restore k1
11027 kmovql(k1, k3);
11028
11029 jmp(return_length);
11030
11031 bind(copy_just_portion_of_candidates);
11032 kmovdl(tmp5, k2);
11033 tzcntl(tmp5, tmp5);
11034
11035 // ~(~0 << tmp5), where tmp5 is a number of elements in an array from the
11036 // result to the first element larger than 0xFF
11037 movl(result, 0xFFFFFFFF);
11038 shlxl(result, result, tmp5);
11039 notl(result);
11040
11041 kmovdl(k1, result);
11042
11043 evpmovwb(Address(dst, 0), k1, tmp1Reg, Assembler::AVX_512bit);
11044 // Restore k1
11045 kmovql(k1, k3);
11046
11047 jmp(return_zero);
11048
11049 clear_vector_masking(); // closing of the stub context for programming mask registers
11050 }
11051 if (UseSSE42Intrinsics) {
11052 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
11053 Label copy_32_loop, copy_16, copy_tail;
11054
11055 bind(below_threshold);
11056
11057 movl(result, len);
11058
11059 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors
11060
11061 // vectored compression
11062 andl(len, 0xfffffff0); // vector count (in chars)
11063 andl(result, 0x0000000f); // tail count (in chars)
11064 testl(len, len);
11065 jccb(Assembler::zero, copy_16);
11066
11067 // compress 16 chars per iter
11068 movdl(tmp1Reg, tmp5);
11069 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
11070 pxor(tmp4Reg, tmp4Reg);
11071
11072 lea(src, Address(src, len, Address::times_2));
11073 lea(dst, Address(dst, len, Address::times_1));
11074 negptr(len);
11075
11076 bind(copy_32_loop);
11077 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters
11078 por(tmp4Reg, tmp2Reg);
11079 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters
11080 por(tmp4Reg, tmp3Reg);
11081 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector
11082 jcc(Assembler::notZero, return_zero);
11083 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte
11084 movdqu(Address(dst, len, Address::times_1), tmp2Reg);
11085 addptr(len, 16);
11086 jcc(Assembler::notZero, copy_32_loop);
11087
11088 // compress next vector of 8 chars (if any)
11089 bind(copy_16);
11090 movl(len, result);
11091 andl(len, 0xfffffff8); // vector count (in chars)
11092 andl(result, 0x00000007); // tail count (in chars)
11093 testl(len, len);
11094 jccb(Assembler::zero, copy_tail);
11095
11096 movdl(tmp1Reg, tmp5);
11097 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
11098 pxor(tmp3Reg, tmp3Reg);
11099
11100 movdqu(tmp2Reg, Address(src, 0));
11101 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
11102 jccb(Assembler::notZero, return_zero);
11103 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte
11104 movq(Address(dst, 0), tmp2Reg);
11105 addptr(src, 16);
11106 addptr(dst, 8);
11107
11108 bind(copy_tail);
11109 movl(len, result);
11110 }
11111 // compress 1 char per iter
11112 testl(len, len);
11113 jccb(Assembler::zero, return_length);
11114 lea(src, Address(src, len, Address::times_2));
11115 lea(dst, Address(dst, len, Address::times_1));
11116 negptr(len);
11117
11118 bind(copy_chars_loop);
11119 load_unsigned_short(result, Address(src, len, Address::times_2));
11120 testl(result, 0xff00); // check if Unicode char
11121 jccb(Assembler::notZero, return_zero);
11122 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte
11123 increment(len);
11124 jcc(Assembler::notZero, copy_chars_loop);
11125
11126 // if compression succeeded, return length
11127 bind(return_length);
11128 pop(result);
11129 jmpb(done);
11130
11131 // if compression failed, return 0
11132 bind(return_zero);
11133 xorl(result, result);
11134 addptr(rsp, wordSize);
11135
11136 bind(done);
11137 }
11138
11139 // Inflate byte[] array to char[].
11140 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java
11141 // @HotSpotIntrinsicCandidate
11142 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) {
11143 // for (int i = 0; i < len; i++) {
11144 // dst[dstOff++] = (char)(src[srcOff++] & 0xff);
11145 // }
11146 // }
11147 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
11148 XMMRegister tmp1, Register tmp2) {
11149 Label copy_chars_loop, done, below_threshold;
11150 // rsi: src
11151 // rdi: dst
11152 // rdx: len
11153 // rcx: tmp2
11154
11155 // rsi holds start addr of source byte[] to be inflated
11156 // rdi holds start addr of destination char[]
11157 // rdx holds length
11158 assert_different_registers(src, dst, len, tmp2);
11159
11160 if ((UseAVX > 2) && // AVX512
11161 VM_Version::supports_avx512vlbw() &&
11162 VM_Version::supports_bmi2()) {
11163
11164 set_vector_masking(); // opening of the stub context for programming mask registers
11165
11166 Label copy_32_loop, copy_tail;
11167 Register tmp3_aliased = len;
11168
11169 // if length of the string is less than 16, handle it in an old fashioned
11170 // way
11171 testl(len, -16);
11172 jcc(Assembler::zero, below_threshold);
11173
11174 // In order to use only one arithmetic operation for the main loop we use
11175 // this pre-calculation
11176 movl(tmp2, len);
11177 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop
11178 andl(len, -32); // vector count
11179 jccb(Assembler::zero, copy_tail);
11180
11181 lea(src, Address(src, len, Address::times_1));
11182 lea(dst, Address(dst, len, Address::times_2));
11183 negptr(len);
11184
11185
11186 // inflate 32 chars per iter
11187 bind(copy_32_loop);
11188 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit);
11189 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit);
11190 addptr(len, 32);
11191 jcc(Assembler::notZero, copy_32_loop);
11192
11193 bind(copy_tail);
11194 // bail out when there is nothing to be done
11195 testl(tmp2, -1); // we don't destroy the contents of tmp2 here
11196 jcc(Assembler::zero, done);
11197
11198 // Save k1
11199 kmovql(k2, k1);
11200
11201 // ~(~0 << length), where length is the # of remaining elements to process
11202 movl(tmp3_aliased, -1);
11203 shlxl(tmp3_aliased, tmp3_aliased, tmp2);
11204 notl(tmp3_aliased);
11205 kmovdl(k1, tmp3_aliased);
11206 evpmovzxbw(tmp1, k1, Address(src, 0), Assembler::AVX_512bit);
11207 evmovdquw(Address(dst, 0), k1, tmp1, Assembler::AVX_512bit);
11208
11209 // Restore k1
11210 kmovql(k1, k2);
11211 jmp(done);
11212
11213 clear_vector_masking(); // closing of the stub context for programming mask registers
11214 }
11215 if (UseSSE42Intrinsics) {
11216 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
11217 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail;
11218
11219 movl(tmp2, len);
11220
11221 if (UseAVX > 1) {
11222 andl(tmp2, (16 - 1));
11223 andl(len, -16);
11224 jccb(Assembler::zero, copy_new_tail);
11225 } else {
11226 andl(tmp2, 0x00000007); // tail count (in chars)
11227 andl(len, 0xfffffff8); // vector count (in chars)
11228 jccb(Assembler::zero, copy_tail);
11229 }
11230
11231 // vectored inflation
11232 lea(src, Address(src, len, Address::times_1));
11233 lea(dst, Address(dst, len, Address::times_2));
11234 negptr(len);
11235
11236 if (UseAVX > 1) {
11237 bind(copy_16_loop);
11238 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit);
11239 vmovdqu(Address(dst, len, Address::times_2), tmp1);
11240 addptr(len, 16);
11241 jcc(Assembler::notZero, copy_16_loop);
11242
11243 bind(below_threshold);
11244 bind(copy_new_tail);
11245 if (UseAVX > 2) {
11246 movl(tmp2, len);
11247 }
11248 else {
11249 movl(len, tmp2);
11250 }
11251 andl(tmp2, 0x00000007);
11252 andl(len, 0xFFFFFFF8);
11253 jccb(Assembler::zero, copy_tail);
11254
11255 pmovzxbw(tmp1, Address(src, 0));
11256 movdqu(Address(dst, 0), tmp1);
11257 addptr(src, 8);
11258 addptr(dst, 2 * 8);
11259
11260 jmp(copy_tail, true);
11261 }
11262
11263 // inflate 8 chars per iter
11264 bind(copy_8_loop);
11265 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words
11266 movdqu(Address(dst, len, Address::times_2), tmp1);
11267 addptr(len, 8);
11268 jcc(Assembler::notZero, copy_8_loop);
11269
11270 bind(copy_tail);
11271 movl(len, tmp2);
11272
11273 cmpl(len, 4);
11274 jccb(Assembler::less, copy_bytes);
11275
11276 movdl(tmp1, Address(src, 0)); // load 4 byte chars
11277 pmovzxbw(tmp1, tmp1);
11278 movq(Address(dst, 0), tmp1);
11279 subptr(len, 4);
11280 addptr(src, 4);
11281 addptr(dst, 8);
11282
11283 bind(copy_bytes);
11284 }
11285 testl(len, len);
11286 jccb(Assembler::zero, done);
11287 lea(src, Address(src, len, Address::times_1));
11288 lea(dst, Address(dst, len, Address::times_2));
11289 negptr(len);
11290
11291 // inflate 1 char per iter
11292 bind(copy_chars_loop);
11293 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char
11294 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word
11295 increment(len);
11296 jcc(Assembler::notZero, copy_chars_loop);
11297
11298 bind(done);
11299 }
11300
11301 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
11302 switch (cond) {
11303 // Note some conditions are synonyms for others
11304 case Assembler::zero: return Assembler::notZero;
11305 case Assembler::notZero: return Assembler::zero;
11306 case Assembler::less: return Assembler::greaterEqual;
11307 case Assembler::lessEqual: return Assembler::greater;
11308 case Assembler::greater: return Assembler::lessEqual;
11309 case Assembler::greaterEqual: return Assembler::less;
11310 case Assembler::below: return Assembler::aboveEqual;
11311 case Assembler::belowEqual: return Assembler::above;
11312 case Assembler::above: return Assembler::belowEqual;
11313 case Assembler::aboveEqual: return Assembler::below;
11314 case Assembler::overflow: return Assembler::noOverflow;
11315 case Assembler::noOverflow: return Assembler::overflow;
11316 case Assembler::negative: return Assembler::positive;
11317 case Assembler::positive: return Assembler::negative;
11318 case Assembler::parity: return Assembler::noParity;
11319 case Assembler::noParity: return Assembler::parity;
11320 }
11321 ShouldNotReachHere(); return Assembler::overflow;
11322 }
11323
11324 SkipIfEqual::SkipIfEqual(
11325 MacroAssembler* masm, const bool* flag_addr, bool value) {
11326 _masm = masm;
11327 _masm->cmp8(ExternalAddress((address)flag_addr), value);
11328 _masm->jcc(Assembler::equal, _label);
11329 }
11330
11331 SkipIfEqual::~SkipIfEqual() {
11332 _masm->bind(_label);
11333 }
11334
11335 // 32-bit Windows has its own fast-path implementation
11336 // of get_thread
11337 #if !defined(WIN32) || defined(_LP64)
11338
11339 // This is simply a call to Thread::current()
11340 void MacroAssembler::get_thread(Register thread) {
11341 if (thread != rax) {
11342 push(rax);
11343 }
11344 LP64_ONLY(push(rdi);)
11345 LP64_ONLY(push(rsi);)
11346 push(rdx);
11347 push(rcx);
11348 #ifdef _LP64
11349 push(r8);
11350 push(r9);
11351 push(r10);
11352 push(r11);
11353 #endif
11354
11355 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0);
11356
11357 #ifdef _LP64
11358 pop(r11);
11359 pop(r10);
11360 pop(r9);
11361 pop(r8);
11362 #endif
11363 pop(rcx);
11364 pop(rdx);
11365 LP64_ONLY(pop(rsi);)
11366 LP64_ONLY(pop(rdi);)
11367 if (thread != rax) {
11368 mov(thread, rax);
11369 pop(rax);
11370 }
11371 }
11372
11373 #endif