1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "jvm.h"
27 #include "asm/assembler.hpp"
28 #include "asm/assembler.inline.hpp"
29 #include "compiler/disassembler.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/barrierSetAssembler.hpp"
32 #include "gc/shared/collectedHeap.inline.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "memory/universe.hpp"
36 #include "oops/accessDecorators.hpp"
37 #include "oops/compressedOops.inline.hpp"
38 #include "oops/klass.inline.hpp"
39 #include "prims/methodHandles.hpp"
40 #include "runtime/biasedLocking.hpp"
41 #include "runtime/flags/flagSetting.hpp"
42 #include "runtime/interfaceSupport.inline.hpp"
43 #include "runtime/objectMonitor.hpp"
44 #include "runtime/os.hpp"
45 #include "runtime/safepoint.hpp"
46 #include "runtime/safepointMechanism.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #include "runtime/stubRoutines.hpp"
49 #include "runtime/thread.hpp"
50 #include "utilities/macros.hpp"
51 #include "crc32c.h"
52 #ifdef COMPILER2
53 #include "opto/intrinsicnode.hpp"
54 #endif
55
56 #ifdef PRODUCT
57 #define BLOCK_COMMENT(str) /* nothing */
58 #define STOP(error) stop(error)
59 #else
60 #define BLOCK_COMMENT(str) block_comment(str)
61 #define STOP(error) block_comment(error); stop(error)
62 #endif
63
64 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
65
66 #ifdef ASSERT
67 bool AbstractAssembler::pd_check_instruction_mark() { return true; }
68 #endif
69
70 static Assembler::Condition reverse[] = {
71 Assembler::noOverflow /* overflow = 0x0 */ ,
72 Assembler::overflow /* noOverflow = 0x1 */ ,
73 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ ,
74 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ ,
75 Assembler::notZero /* zero = 0x4, equal = 0x4 */ ,
76 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ ,
77 Assembler::above /* belowEqual = 0x6 */ ,
78 Assembler::belowEqual /* above = 0x7 */ ,
79 Assembler::positive /* negative = 0x8 */ ,
80 Assembler::negative /* positive = 0x9 */ ,
81 Assembler::noParity /* parity = 0xa */ ,
82 Assembler::parity /* noParity = 0xb */ ,
83 Assembler::greaterEqual /* less = 0xc */ ,
84 Assembler::less /* greaterEqual = 0xd */ ,
85 Assembler::greater /* lessEqual = 0xe */ ,
86 Assembler::lessEqual /* greater = 0xf, */
87
88 };
89
90
91 // Implementation of MacroAssembler
92
93 // First all the versions that have distinct versions depending on 32/64 bit
94 // Unless the difference is trivial (1 line or so).
95
96 #ifndef _LP64
97
98 // 32bit versions
99
100 Address MacroAssembler::as_Address(AddressLiteral adr) {
101 return Address(adr.target(), adr.rspec());
102 }
103
104 Address MacroAssembler::as_Address(ArrayAddress adr) {
105 return Address::make_array(adr);
106 }
107
108 void MacroAssembler::call_VM_leaf_base(address entry_point,
109 int number_of_arguments) {
110 call(RuntimeAddress(entry_point));
111 increment(rsp, number_of_arguments * wordSize);
112 }
113
114 void MacroAssembler::cmpklass(Address src1, Metadata* obj) {
115 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
116 }
117
118 void MacroAssembler::cmpklass(Register src1, Metadata* obj) {
119 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
120 }
121
122 void MacroAssembler::cmpoop_raw(Address src1, jobject obj) {
123 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
124 }
125
126 void MacroAssembler::cmpoop_raw(Register src1, jobject obj) {
127 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
128 }
129
130 void MacroAssembler::cmpoop(Address src1, jobject obj) {
131 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
132 bs->obj_equals(this, src1, obj);
133 }
134
135 void MacroAssembler::cmpoop(Register src1, jobject obj) {
136 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
137 bs->obj_equals(this, src1, obj);
138 }
139
140 void MacroAssembler::extend_sign(Register hi, Register lo) {
141 // According to Intel Doc. AP-526, "Integer Divide", p.18.
142 if (VM_Version::is_P6() && hi == rdx && lo == rax) {
143 cdql();
144 } else {
145 movl(hi, lo);
146 sarl(hi, 31);
147 }
148 }
149
150 void MacroAssembler::jC2(Register tmp, Label& L) {
151 // set parity bit if FPU flag C2 is set (via rax)
152 save_rax(tmp);
153 fwait(); fnstsw_ax();
154 sahf();
155 restore_rax(tmp);
156 // branch
157 jcc(Assembler::parity, L);
158 }
159
160 void MacroAssembler::jnC2(Register tmp, Label& L) {
161 // set parity bit if FPU flag C2 is set (via rax)
162 save_rax(tmp);
163 fwait(); fnstsw_ax();
164 sahf();
165 restore_rax(tmp);
166 // branch
167 jcc(Assembler::noParity, L);
168 }
169
170 // 32bit can do a case table jump in one instruction but we no longer allow the base
171 // to be installed in the Address class
172 void MacroAssembler::jump(ArrayAddress entry) {
173 jmp(as_Address(entry));
174 }
175
176 // Note: y_lo will be destroyed
177 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
178 // Long compare for Java (semantics as described in JVM spec.)
179 Label high, low, done;
180
181 cmpl(x_hi, y_hi);
182 jcc(Assembler::less, low);
183 jcc(Assembler::greater, high);
184 // x_hi is the return register
185 xorl(x_hi, x_hi);
186 cmpl(x_lo, y_lo);
187 jcc(Assembler::below, low);
188 jcc(Assembler::equal, done);
189
190 bind(high);
191 xorl(x_hi, x_hi);
192 increment(x_hi);
193 jmp(done);
194
195 bind(low);
196 xorl(x_hi, x_hi);
197 decrementl(x_hi);
198
199 bind(done);
200 }
201
202 void MacroAssembler::lea(Register dst, AddressLiteral src) {
203 mov_literal32(dst, (int32_t)src.target(), src.rspec());
204 }
205
206 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
207 // leal(dst, as_Address(adr));
208 // see note in movl as to why we must use a move
209 mov_literal32(dst, (int32_t) adr.target(), adr.rspec());
210 }
211
212 void MacroAssembler::leave() {
213 mov(rsp, rbp);
214 pop(rbp);
215 }
216
217 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) {
218 // Multiplication of two Java long values stored on the stack
219 // as illustrated below. Result is in rdx:rax.
220 //
221 // rsp ---> [ ?? ] \ \
222 // .... | y_rsp_offset |
223 // [ y_lo ] / (in bytes) | x_rsp_offset
224 // [ y_hi ] | (in bytes)
225 // .... |
226 // [ x_lo ] /
227 // [ x_hi ]
228 // ....
229 //
230 // Basic idea: lo(result) = lo(x_lo * y_lo)
231 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
232 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset);
233 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset);
234 Label quick;
235 // load x_hi, y_hi and check if quick
236 // multiplication is possible
237 movl(rbx, x_hi);
238 movl(rcx, y_hi);
239 movl(rax, rbx);
240 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0
241 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply
242 // do full multiplication
243 // 1st step
244 mull(y_lo); // x_hi * y_lo
245 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx,
246 // 2nd step
247 movl(rax, x_lo);
248 mull(rcx); // x_lo * y_hi
249 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx,
250 // 3rd step
251 bind(quick); // note: rbx, = 0 if quick multiply!
252 movl(rax, x_lo);
253 mull(y_lo); // x_lo * y_lo
254 addl(rdx, rbx); // correct hi(x_lo * y_lo)
255 }
256
257 void MacroAssembler::lneg(Register hi, Register lo) {
258 negl(lo);
259 adcl(hi, 0);
260 negl(hi);
261 }
262
263 void MacroAssembler::lshl(Register hi, Register lo) {
264 // Java shift left long support (semantics as described in JVM spec., p.305)
265 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n))
266 // shift value is in rcx !
267 assert(hi != rcx, "must not use rcx");
268 assert(lo != rcx, "must not use rcx");
269 const Register s = rcx; // shift count
270 const int n = BitsPerWord;
271 Label L;
272 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
273 cmpl(s, n); // if (s < n)
274 jcc(Assembler::less, L); // else (s >= n)
275 movl(hi, lo); // x := x << n
276 xorl(lo, lo);
277 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
278 bind(L); // s (mod n) < n
279 shldl(hi, lo); // x := x << s
280 shll(lo);
281 }
282
283
284 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) {
285 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310)
286 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n))
287 assert(hi != rcx, "must not use rcx");
288 assert(lo != rcx, "must not use rcx");
289 const Register s = rcx; // shift count
290 const int n = BitsPerWord;
291 Label L;
292 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
293 cmpl(s, n); // if (s < n)
294 jcc(Assembler::less, L); // else (s >= n)
295 movl(lo, hi); // x := x >> n
296 if (sign_extension) sarl(hi, 31);
297 else xorl(hi, hi);
298 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
299 bind(L); // s (mod n) < n
300 shrdl(lo, hi); // x := x >> s
301 if (sign_extension) sarl(hi);
302 else shrl(hi);
303 }
304
305 void MacroAssembler::movoop(Register dst, jobject obj) {
306 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
307 }
308
309 void MacroAssembler::movoop(Address dst, jobject obj) {
310 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
311 }
312
313 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
314 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
315 }
316
317 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
318 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
319 }
320
321 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
322 // scratch register is not used,
323 // it is defined to match parameters of 64-bit version of this method.
324 if (src.is_lval()) {
325 mov_literal32(dst, (intptr_t)src.target(), src.rspec());
326 } else {
327 movl(dst, as_Address(src));
328 }
329 }
330
331 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
332 movl(as_Address(dst), src);
333 }
334
335 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
336 movl(dst, as_Address(src));
337 }
338
339 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
340 void MacroAssembler::movptr(Address dst, intptr_t src) {
341 movl(dst, src);
342 }
343
344
345 void MacroAssembler::pop_callee_saved_registers() {
346 pop(rcx);
347 pop(rdx);
348 pop(rdi);
349 pop(rsi);
350 }
351
352 void MacroAssembler::pop_fTOS() {
353 fld_d(Address(rsp, 0));
354 addl(rsp, 2 * wordSize);
355 }
356
357 void MacroAssembler::push_callee_saved_registers() {
358 push(rsi);
359 push(rdi);
360 push(rdx);
361 push(rcx);
362 }
363
364 void MacroAssembler::push_fTOS() {
365 subl(rsp, 2 * wordSize);
366 fstp_d(Address(rsp, 0));
367 }
368
369
370 void MacroAssembler::pushoop(jobject obj) {
371 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate());
372 }
373
374 void MacroAssembler::pushklass(Metadata* obj) {
375 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate());
376 }
377
378 void MacroAssembler::pushptr(AddressLiteral src) {
379 if (src.is_lval()) {
380 push_literal32((int32_t)src.target(), src.rspec());
381 } else {
382 pushl(as_Address(src));
383 }
384 }
385
386 void MacroAssembler::set_word_if_not_zero(Register dst) {
387 xorl(dst, dst);
388 set_byte_if_not_zero(dst);
389 }
390
391 static void pass_arg0(MacroAssembler* masm, Register arg) {
392 masm->push(arg);
393 }
394
395 static void pass_arg1(MacroAssembler* masm, Register arg) {
396 masm->push(arg);
397 }
398
399 static void pass_arg2(MacroAssembler* masm, Register arg) {
400 masm->push(arg);
401 }
402
403 static void pass_arg3(MacroAssembler* masm, Register arg) {
404 masm->push(arg);
405 }
406
407 #ifndef PRODUCT
408 extern "C" void findpc(intptr_t x);
409 #endif
410
411 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) {
412 // In order to get locks to work, we need to fake a in_VM state
413 JavaThread* thread = JavaThread::current();
414 JavaThreadState saved_state = thread->thread_state();
415 thread->set_thread_state(_thread_in_vm);
416 if (ShowMessageBoxOnError) {
417 JavaThread* thread = JavaThread::current();
418 JavaThreadState saved_state = thread->thread_state();
419 thread->set_thread_state(_thread_in_vm);
420 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
421 ttyLocker ttyl;
422 BytecodeCounter::print();
423 }
424 // To see where a verify_oop failed, get $ebx+40/X for this frame.
425 // This is the value of eip which points to where verify_oop will return.
426 if (os::message_box(msg, "Execution stopped, print registers?")) {
427 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip);
428 BREAKPOINT;
429 }
430 } else {
431 ttyLocker ttyl;
432 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg);
433 }
434 // Don't assert holding the ttyLock
435 assert(false, "DEBUG MESSAGE: %s", msg);
436 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
437 }
438
439 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) {
440 ttyLocker ttyl;
441 FlagSetting fs(Debugging, true);
442 tty->print_cr("eip = 0x%08x", eip);
443 #ifndef PRODUCT
444 if ((WizardMode || Verbose) && PrintMiscellaneous) {
445 tty->cr();
446 findpc(eip);
447 tty->cr();
448 }
449 #endif
450 #define PRINT_REG(rax) \
451 { tty->print("%s = ", #rax); os::print_location(tty, rax); }
452 PRINT_REG(rax);
453 PRINT_REG(rbx);
454 PRINT_REG(rcx);
455 PRINT_REG(rdx);
456 PRINT_REG(rdi);
457 PRINT_REG(rsi);
458 PRINT_REG(rbp);
459 PRINT_REG(rsp);
460 #undef PRINT_REG
461 // Print some words near top of staack.
462 int* dump_sp = (int*) rsp;
463 for (int col1 = 0; col1 < 8; col1++) {
464 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
465 os::print_location(tty, *dump_sp++);
466 }
467 for (int row = 0; row < 16; row++) {
468 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
469 for (int col = 0; col < 8; col++) {
470 tty->print(" 0x%08x", *dump_sp++);
471 }
472 tty->cr();
473 }
474 // Print some instructions around pc:
475 Disassembler::decode((address)eip-64, (address)eip);
476 tty->print_cr("--------");
477 Disassembler::decode((address)eip, (address)eip+32);
478 }
479
480 void MacroAssembler::stop(const char* msg) {
481 ExternalAddress message((address)msg);
482 // push address of message
483 pushptr(message.addr());
484 { Label L; call(L, relocInfo::none); bind(L); } // push eip
485 pusha(); // push registers
486 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32)));
487 hlt();
488 }
489
490 void MacroAssembler::warn(const char* msg) {
491 push_CPU_state();
492
493 ExternalAddress message((address) msg);
494 // push address of message
495 pushptr(message.addr());
496
497 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning)));
498 addl(rsp, wordSize); // discard argument
499 pop_CPU_state();
500 }
501
502 void MacroAssembler::print_state() {
503 { Label L; call(L, relocInfo::none); bind(L); } // push eip
504 pusha(); // push registers
505
506 push_CPU_state();
507 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32)));
508 pop_CPU_state();
509
510 popa();
511 addl(rsp, wordSize);
512 }
513
514 #else // _LP64
515
516 // 64 bit versions
517
518 Address MacroAssembler::as_Address(AddressLiteral adr) {
519 // amd64 always does this as a pc-rel
520 // we can be absolute or disp based on the instruction type
521 // jmp/call are displacements others are absolute
522 assert(!adr.is_lval(), "must be rval");
523 assert(reachable(adr), "must be");
524 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc());
525
526 }
527
528 Address MacroAssembler::as_Address(ArrayAddress adr) {
529 AddressLiteral base = adr.base();
530 lea(rscratch1, base);
531 Address index = adr.index();
532 assert(index._disp == 0, "must not have disp"); // maybe it can?
533 Address array(rscratch1, index._index, index._scale, index._disp);
534 return array;
535 }
536
537 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) {
538 Label L, E;
539
540 #ifdef _WIN64
541 // Windows always allocates space for it's register args
542 assert(num_args <= 4, "only register arguments supported");
543 subq(rsp, frame::arg_reg_save_area_bytes);
544 #endif
545
546 // Align stack if necessary
547 testl(rsp, 15);
548 jcc(Assembler::zero, L);
549
550 subq(rsp, 8);
551 {
552 call(RuntimeAddress(entry_point));
553 }
554 addq(rsp, 8);
555 jmp(E);
556
557 bind(L);
558 {
559 call(RuntimeAddress(entry_point));
560 }
561
562 bind(E);
563
564 #ifdef _WIN64
565 // restore stack pointer
566 addq(rsp, frame::arg_reg_save_area_bytes);
567 #endif
568
569 }
570
571 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) {
572 assert(!src2.is_lval(), "should use cmpptr");
573
574 if (reachable(src2)) {
575 cmpq(src1, as_Address(src2));
576 } else {
577 lea(rscratch1, src2);
578 Assembler::cmpq(src1, Address(rscratch1, 0));
579 }
580 }
581
582 int MacroAssembler::corrected_idivq(Register reg) {
583 // Full implementation of Java ldiv and lrem; checks for special
584 // case as described in JVM spec., p.243 & p.271. The function
585 // returns the (pc) offset of the idivl instruction - may be needed
586 // for implicit exceptions.
587 //
588 // normal case special case
589 //
590 // input : rax: dividend min_long
591 // reg: divisor (may not be eax/edx) -1
592 //
593 // output: rax: quotient (= rax idiv reg) min_long
594 // rdx: remainder (= rax irem reg) 0
595 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register");
596 static const int64_t min_long = 0x8000000000000000;
597 Label normal_case, special_case;
598
599 // check for special case
600 cmp64(rax, ExternalAddress((address) &min_long));
601 jcc(Assembler::notEqual, normal_case);
602 xorl(rdx, rdx); // prepare rdx for possible special case (where
603 // remainder = 0)
604 cmpq(reg, -1);
605 jcc(Assembler::equal, special_case);
606
607 // handle normal case
608 bind(normal_case);
609 cdqq();
610 int idivq_offset = offset();
611 idivq(reg);
612
613 // normal and special case exit
614 bind(special_case);
615
616 return idivq_offset;
617 }
618
619 void MacroAssembler::decrementq(Register reg, int value) {
620 if (value == min_jint) { subq(reg, value); return; }
621 if (value < 0) { incrementq(reg, -value); return; }
622 if (value == 0) { ; return; }
623 if (value == 1 && UseIncDec) { decq(reg) ; return; }
624 /* else */ { subq(reg, value) ; return; }
625 }
626
627 void MacroAssembler::decrementq(Address dst, int value) {
628 if (value == min_jint) { subq(dst, value); return; }
629 if (value < 0) { incrementq(dst, -value); return; }
630 if (value == 0) { ; return; }
631 if (value == 1 && UseIncDec) { decq(dst) ; return; }
632 /* else */ { subq(dst, value) ; return; }
633 }
634
635 void MacroAssembler::incrementq(AddressLiteral dst) {
636 if (reachable(dst)) {
637 incrementq(as_Address(dst));
638 } else {
639 lea(rscratch1, dst);
640 incrementq(Address(rscratch1, 0));
641 }
642 }
643
644 void MacroAssembler::incrementq(Register reg, int value) {
645 if (value == min_jint) { addq(reg, value); return; }
646 if (value < 0) { decrementq(reg, -value); return; }
647 if (value == 0) { ; return; }
648 if (value == 1 && UseIncDec) { incq(reg) ; return; }
649 /* else */ { addq(reg, value) ; return; }
650 }
651
652 void MacroAssembler::incrementq(Address dst, int value) {
653 if (value == min_jint) { addq(dst, value); return; }
654 if (value < 0) { decrementq(dst, -value); return; }
655 if (value == 0) { ; return; }
656 if (value == 1 && UseIncDec) { incq(dst) ; return; }
657 /* else */ { addq(dst, value) ; return; }
658 }
659
660 // 32bit can do a case table jump in one instruction but we no longer allow the base
661 // to be installed in the Address class
662 void MacroAssembler::jump(ArrayAddress entry) {
663 lea(rscratch1, entry.base());
664 Address dispatch = entry.index();
665 assert(dispatch._base == noreg, "must be");
666 dispatch._base = rscratch1;
667 jmp(dispatch);
668 }
669
670 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
671 ShouldNotReachHere(); // 64bit doesn't use two regs
672 cmpq(x_lo, y_lo);
673 }
674
675 void MacroAssembler::lea(Register dst, AddressLiteral src) {
676 mov_literal64(dst, (intptr_t)src.target(), src.rspec());
677 }
678
679 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
680 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec());
681 movptr(dst, rscratch1);
682 }
683
684 void MacroAssembler::leave() {
685 // %%% is this really better? Why not on 32bit too?
686 emit_int8((unsigned char)0xC9); // LEAVE
687 }
688
689 void MacroAssembler::lneg(Register hi, Register lo) {
690 ShouldNotReachHere(); // 64bit doesn't use two regs
691 negq(lo);
692 }
693
694 void MacroAssembler::movoop(Register dst, jobject obj) {
695 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate());
696 }
697
698 void MacroAssembler::movoop(Address dst, jobject obj) {
699 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate());
700 movq(dst, rscratch1);
701 }
702
703 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
704 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
705 }
706
707 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
708 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
709 movq(dst, rscratch1);
710 }
711
712 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
713 if (src.is_lval()) {
714 mov_literal64(dst, (intptr_t)src.target(), src.rspec());
715 } else {
716 if (reachable(src)) {
717 movq(dst, as_Address(src));
718 } else {
719 lea(scratch, src);
720 movq(dst, Address(scratch, 0));
721 }
722 }
723 }
724
725 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
726 movq(as_Address(dst), src);
727 }
728
729 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
730 movq(dst, as_Address(src));
731 }
732
733 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
734 void MacroAssembler::movptr(Address dst, intptr_t src) {
735 mov64(rscratch1, src);
736 movq(dst, rscratch1);
737 }
738
739 // These are mostly for initializing NULL
740 void MacroAssembler::movptr(Address dst, int32_t src) {
741 movslq(dst, src);
742 }
743
744 void MacroAssembler::movptr(Register dst, int32_t src) {
745 mov64(dst, (intptr_t)src);
746 }
747
748 void MacroAssembler::pushoop(jobject obj) {
749 movoop(rscratch1, obj);
750 push(rscratch1);
751 }
752
753 void MacroAssembler::pushklass(Metadata* obj) {
754 mov_metadata(rscratch1, obj);
755 push(rscratch1);
756 }
757
758 void MacroAssembler::pushptr(AddressLiteral src) {
759 lea(rscratch1, src);
760 if (src.is_lval()) {
761 push(rscratch1);
762 } else {
763 pushq(Address(rscratch1, 0));
764 }
765 }
766
767 void MacroAssembler::reset_last_Java_frame(bool clear_fp) {
768 // we must set sp to zero to clear frame
769 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
770 // must clear fp, so that compiled frames are not confused; it is
771 // possible that we need it only for debugging
772 if (clear_fp) {
773 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
774 }
775
776 // Always clear the pc because it could have been set by make_walkable()
777 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
778 vzeroupper();
779 }
780
781 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
782 Register last_java_fp,
783 address last_java_pc) {
784 vzeroupper();
785 // determine last_java_sp register
786 if (!last_java_sp->is_valid()) {
787 last_java_sp = rsp;
788 }
789
790 // last_java_fp is optional
791 if (last_java_fp->is_valid()) {
792 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()),
793 last_java_fp);
794 }
795
796 // last_java_pc is optional
797 if (last_java_pc != NULL) {
798 Address java_pc(r15_thread,
799 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset());
800 lea(rscratch1, InternalAddress(last_java_pc));
801 movptr(java_pc, rscratch1);
802 }
803
804 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
805 }
806
807 static void pass_arg0(MacroAssembler* masm, Register arg) {
808 if (c_rarg0 != arg ) {
809 masm->mov(c_rarg0, arg);
810 }
811 }
812
813 static void pass_arg1(MacroAssembler* masm, Register arg) {
814 if (c_rarg1 != arg ) {
815 masm->mov(c_rarg1, arg);
816 }
817 }
818
819 static void pass_arg2(MacroAssembler* masm, Register arg) {
820 if (c_rarg2 != arg ) {
821 masm->mov(c_rarg2, arg);
822 }
823 }
824
825 static void pass_arg3(MacroAssembler* masm, Register arg) {
826 if (c_rarg3 != arg ) {
827 masm->mov(c_rarg3, arg);
828 }
829 }
830
831 void MacroAssembler::stop(const char* msg) {
832 address rip = pc();
833 pusha(); // get regs on stack
834 lea(c_rarg0, ExternalAddress((address) msg));
835 lea(c_rarg1, InternalAddress(rip));
836 movq(c_rarg2, rsp); // pass pointer to regs array
837 andq(rsp, -16); // align stack as required by ABI
838 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
839 hlt();
840 }
841
842 void MacroAssembler::warn(const char* msg) {
843 push(rbp);
844 movq(rbp, rsp);
845 andq(rsp, -16); // align stack as required by push_CPU_state and call
846 push_CPU_state(); // keeps alignment at 16 bytes
847 lea(c_rarg0, ExternalAddress((address) msg));
848 lea(rax, ExternalAddress(CAST_FROM_FN_PTR(address, warning)));
849 call(rax);
850 pop_CPU_state();
851 mov(rsp, rbp);
852 pop(rbp);
853 }
854
855 void MacroAssembler::print_state() {
856 address rip = pc();
857 pusha(); // get regs on stack
858 push(rbp);
859 movq(rbp, rsp);
860 andq(rsp, -16); // align stack as required by push_CPU_state and call
861 push_CPU_state(); // keeps alignment at 16 bytes
862
863 lea(c_rarg0, InternalAddress(rip));
864 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array
865 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1);
866
867 pop_CPU_state();
868 mov(rsp, rbp);
869 pop(rbp);
870 popa();
871 }
872
873 #ifndef PRODUCT
874 extern "C" void findpc(intptr_t x);
875 #endif
876
877 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) {
878 // In order to get locks to work, we need to fake a in_VM state
879 if (ShowMessageBoxOnError) {
880 JavaThread* thread = JavaThread::current();
881 JavaThreadState saved_state = thread->thread_state();
882 thread->set_thread_state(_thread_in_vm);
883 #ifndef PRODUCT
884 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
885 ttyLocker ttyl;
886 BytecodeCounter::print();
887 }
888 #endif
889 // To see where a verify_oop failed, get $ebx+40/X for this frame.
890 // XXX correct this offset for amd64
891 // This is the value of eip which points to where verify_oop will return.
892 if (os::message_box(msg, "Execution stopped, print registers?")) {
893 print_state64(pc, regs);
894 BREAKPOINT;
895 assert(false, "start up GDB");
896 }
897 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
898 } else {
899 ttyLocker ttyl;
900 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n",
901 msg);
902 assert(false, "DEBUG MESSAGE: %s", msg);
903 }
904 }
905
906 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) {
907 ttyLocker ttyl;
908 FlagSetting fs(Debugging, true);
909 tty->print_cr("rip = 0x%016lx", (intptr_t)pc);
910 #ifndef PRODUCT
911 tty->cr();
912 findpc(pc);
913 tty->cr();
914 #endif
915 #define PRINT_REG(rax, value) \
916 { tty->print("%s = ", #rax); os::print_location(tty, value); }
917 PRINT_REG(rax, regs[15]);
918 PRINT_REG(rbx, regs[12]);
919 PRINT_REG(rcx, regs[14]);
920 PRINT_REG(rdx, regs[13]);
921 PRINT_REG(rdi, regs[8]);
922 PRINT_REG(rsi, regs[9]);
923 PRINT_REG(rbp, regs[10]);
924 PRINT_REG(rsp, regs[11]);
925 PRINT_REG(r8 , regs[7]);
926 PRINT_REG(r9 , regs[6]);
927 PRINT_REG(r10, regs[5]);
928 PRINT_REG(r11, regs[4]);
929 PRINT_REG(r12, regs[3]);
930 PRINT_REG(r13, regs[2]);
931 PRINT_REG(r14, regs[1]);
932 PRINT_REG(r15, regs[0]);
933 #undef PRINT_REG
934 // Print some words near top of staack.
935 int64_t* rsp = (int64_t*) regs[11];
936 int64_t* dump_sp = rsp;
937 for (int col1 = 0; col1 < 8; col1++) {
938 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
939 os::print_location(tty, *dump_sp++);
940 }
941 for (int row = 0; row < 25; row++) {
942 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
943 for (int col = 0; col < 4; col++) {
944 tty->print(" 0x%016lx", (intptr_t)*dump_sp++);
945 }
946 tty->cr();
947 }
948 // Print some instructions around pc:
949 Disassembler::decode((address)pc-64, (address)pc);
950 tty->print_cr("--------");
951 Disassembler::decode((address)pc, (address)pc+32);
952 }
953
954 #endif // _LP64
955
956 // Now versions that are common to 32/64 bit
957
958 void MacroAssembler::addptr(Register dst, int32_t imm32) {
959 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32));
960 }
961
962 void MacroAssembler::addptr(Register dst, Register src) {
963 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
964 }
965
966 void MacroAssembler::addptr(Address dst, Register src) {
967 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
968 }
969
970 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) {
971 if (reachable(src)) {
972 Assembler::addsd(dst, as_Address(src));
973 } else {
974 lea(rscratch1, src);
975 Assembler::addsd(dst, Address(rscratch1, 0));
976 }
977 }
978
979 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) {
980 if (reachable(src)) {
981 addss(dst, as_Address(src));
982 } else {
983 lea(rscratch1, src);
984 addss(dst, Address(rscratch1, 0));
985 }
986 }
987
988 void MacroAssembler::addpd(XMMRegister dst, AddressLiteral src) {
989 if (reachable(src)) {
990 Assembler::addpd(dst, as_Address(src));
991 } else {
992 lea(rscratch1, src);
993 Assembler::addpd(dst, Address(rscratch1, 0));
994 }
995 }
996
997 void MacroAssembler::align(int modulus) {
998 align(modulus, offset());
999 }
1000
1001 void MacroAssembler::align(int modulus, int target) {
1002 if (target % modulus != 0) {
1003 nop(modulus - (target % modulus));
1004 }
1005 }
1006
1007 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
1008 // Used in sign-masking with aligned address.
1009 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
1010 if (reachable(src)) {
1011 Assembler::andpd(dst, as_Address(src));
1012 } else {
1013 lea(scratch_reg, src);
1014 Assembler::andpd(dst, Address(scratch_reg, 0));
1015 }
1016 }
1017
1018 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
1019 // Used in sign-masking with aligned address.
1020 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
1021 if (reachable(src)) {
1022 Assembler::andps(dst, as_Address(src));
1023 } else {
1024 lea(scratch_reg, src);
1025 Assembler::andps(dst, Address(scratch_reg, 0));
1026 }
1027 }
1028
1029 void MacroAssembler::andptr(Register dst, int32_t imm32) {
1030 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32));
1031 }
1032
1033 void MacroAssembler::atomic_incl(Address counter_addr) {
1034 lock();
1035 incrementl(counter_addr);
1036 }
1037
1038 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) {
1039 if (reachable(counter_addr)) {
1040 atomic_incl(as_Address(counter_addr));
1041 } else {
1042 lea(scr, counter_addr);
1043 atomic_incl(Address(scr, 0));
1044 }
1045 }
1046
1047 #ifdef _LP64
1048 void MacroAssembler::atomic_incq(Address counter_addr) {
1049 lock();
1050 incrementq(counter_addr);
1051 }
1052
1053 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) {
1054 if (reachable(counter_addr)) {
1055 atomic_incq(as_Address(counter_addr));
1056 } else {
1057 lea(scr, counter_addr);
1058 atomic_incq(Address(scr, 0));
1059 }
1060 }
1061 #endif
1062
1063 // Writes to stack successive pages until offset reached to check for
1064 // stack overflow + shadow pages. This clobbers tmp.
1065 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
1066 movptr(tmp, rsp);
1067 // Bang stack for total size given plus shadow page size.
1068 // Bang one page at a time because large size can bang beyond yellow and
1069 // red zones.
1070 Label loop;
1071 bind(loop);
1072 movl(Address(tmp, (-os::vm_page_size())), size );
1073 subptr(tmp, os::vm_page_size());
1074 subl(size, os::vm_page_size());
1075 jcc(Assembler::greater, loop);
1076
1077 // Bang down shadow pages too.
1078 // At this point, (tmp-0) is the last address touched, so don't
1079 // touch it again. (It was touched as (tmp-pagesize) but then tmp
1080 // was post-decremented.) Skip this address by starting at i=1, and
1081 // touch a few more pages below. N.B. It is important to touch all
1082 // the way down including all pages in the shadow zone.
1083 for (int i = 1; i < ((int)JavaThread::stack_shadow_zone_size() / os::vm_page_size()); i++) {
1084 // this could be any sized move but this is can be a debugging crumb
1085 // so the bigger the better.
1086 movptr(Address(tmp, (-i*os::vm_page_size())), size );
1087 }
1088 }
1089
1090 void MacroAssembler::reserved_stack_check() {
1091 // testing if reserved zone needs to be enabled
1092 Label no_reserved_zone_enabling;
1093 Register thread = NOT_LP64(rsi) LP64_ONLY(r15_thread);
1094 NOT_LP64(get_thread(rsi);)
1095
1096 cmpptr(rsp, Address(thread, JavaThread::reserved_stack_activation_offset()));
1097 jcc(Assembler::below, no_reserved_zone_enabling);
1098
1099 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), thread);
1100 jump(RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry()));
1101 should_not_reach_here();
1102
1103 bind(no_reserved_zone_enabling);
1104 }
1105
1106 int MacroAssembler::biased_locking_enter(Register lock_reg,
1107 Register obj_reg,
1108 Register swap_reg,
1109 Register tmp_reg,
1110 bool swap_reg_contains_mark,
1111 Label& done,
1112 Label* slow_case,
1113 BiasedLockingCounters* counters) {
1114 assert(UseBiasedLocking, "why call this otherwise?");
1115 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq");
1116 assert(tmp_reg != noreg, "tmp_reg must be supplied");
1117 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg);
1118 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout");
1119 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes());
1120 NOT_LP64( Address saved_mark_addr(lock_reg, 0); )
1121
1122 if (PrintBiasedLockingStatistics && counters == NULL) {
1123 counters = BiasedLocking::counters();
1124 }
1125 // Biased locking
1126 // See whether the lock is currently biased toward our thread and
1127 // whether the epoch is still valid
1128 // Note that the runtime guarantees sufficient alignment of JavaThread
1129 // pointers to allow age to be placed into low bits
1130 // First check to see whether biasing is even enabled for this object
1131 Label cas_label;
1132 int null_check_offset = -1;
1133 if (!swap_reg_contains_mark) {
1134 null_check_offset = offset();
1135 movptr(swap_reg, mark_addr);
1136 }
1137 movptr(tmp_reg, swap_reg);
1138 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place);
1139 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern);
1140 jcc(Assembler::notEqual, cas_label);
1141 // The bias pattern is present in the object's header. Need to check
1142 // whether the bias owner and the epoch are both still current.
1143 #ifndef _LP64
1144 // Note that because there is no current thread register on x86_32 we
1145 // need to store off the mark word we read out of the object to
1146 // avoid reloading it and needing to recheck invariants below. This
1147 // store is unfortunate but it makes the overall code shorter and
1148 // simpler.
1149 movptr(saved_mark_addr, swap_reg);
1150 #endif
1151 if (swap_reg_contains_mark) {
1152 null_check_offset = offset();
1153 }
1154 load_prototype_header(tmp_reg, obj_reg);
1155 #ifdef _LP64
1156 orptr(tmp_reg, r15_thread);
1157 xorptr(tmp_reg, swap_reg);
1158 Register header_reg = tmp_reg;
1159 #else
1160 xorptr(tmp_reg, swap_reg);
1161 get_thread(swap_reg);
1162 xorptr(swap_reg, tmp_reg);
1163 Register header_reg = swap_reg;
1164 #endif
1165 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place));
1166 if (counters != NULL) {
1167 cond_inc32(Assembler::zero,
1168 ExternalAddress((address) counters->biased_lock_entry_count_addr()));
1169 }
1170 jcc(Assembler::equal, done);
1171
1172 Label try_revoke_bias;
1173 Label try_rebias;
1174
1175 // At this point we know that the header has the bias pattern and
1176 // that we are not the bias owner in the current epoch. We need to
1177 // figure out more details about the state of the header in order to
1178 // know what operations can be legally performed on the object's
1179 // header.
1180
1181 // If the low three bits in the xor result aren't clear, that means
1182 // the prototype header is no longer biased and we have to revoke
1183 // the bias on this object.
1184 testptr(header_reg, markOopDesc::biased_lock_mask_in_place);
1185 jccb(Assembler::notZero, try_revoke_bias);
1186
1187 // Biasing is still enabled for this data type. See whether the
1188 // epoch of the current bias is still valid, meaning that the epoch
1189 // bits of the mark word are equal to the epoch bits of the
1190 // prototype header. (Note that the prototype header's epoch bits
1191 // only change at a safepoint.) If not, attempt to rebias the object
1192 // toward the current thread. Note that we must be absolutely sure
1193 // that the current epoch is invalid in order to do this because
1194 // otherwise the manipulations it performs on the mark word are
1195 // illegal.
1196 testptr(header_reg, markOopDesc::epoch_mask_in_place);
1197 jccb(Assembler::notZero, try_rebias);
1198
1199 // The epoch of the current bias is still valid but we know nothing
1200 // about the owner; it might be set or it might be clear. Try to
1201 // acquire the bias of the object using an atomic operation. If this
1202 // fails we will go in to the runtime to revoke the object's bias.
1203 // Note that we first construct the presumed unbiased header so we
1204 // don't accidentally blow away another thread's valid bias.
1205 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1206 andptr(swap_reg,
1207 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place);
1208 #ifdef _LP64
1209 movptr(tmp_reg, swap_reg);
1210 orptr(tmp_reg, r15_thread);
1211 #else
1212 get_thread(tmp_reg);
1213 orptr(tmp_reg, swap_reg);
1214 #endif
1215 lock();
1216 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1217 // If the biasing toward our thread failed, this means that
1218 // another thread succeeded in biasing it toward itself and we
1219 // need to revoke that bias. The revocation will occur in the
1220 // interpreter runtime in the slow case.
1221 if (counters != NULL) {
1222 cond_inc32(Assembler::zero,
1223 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr()));
1224 }
1225 if (slow_case != NULL) {
1226 jcc(Assembler::notZero, *slow_case);
1227 }
1228 jmp(done);
1229
1230 bind(try_rebias);
1231 // At this point we know the epoch has expired, meaning that the
1232 // current "bias owner", if any, is actually invalid. Under these
1233 // circumstances _only_, we are allowed to use the current header's
1234 // value as the comparison value when doing the cas to acquire the
1235 // bias in the current epoch. In other words, we allow transfer of
1236 // the bias from one thread to another directly in this situation.
1237 //
1238 // FIXME: due to a lack of registers we currently blow away the age
1239 // bits in this situation. Should attempt to preserve them.
1240 load_prototype_header(tmp_reg, obj_reg);
1241 #ifdef _LP64
1242 orptr(tmp_reg, r15_thread);
1243 #else
1244 get_thread(swap_reg);
1245 orptr(tmp_reg, swap_reg);
1246 movptr(swap_reg, saved_mark_addr);
1247 #endif
1248 lock();
1249 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1250 // If the biasing toward our thread failed, then another thread
1251 // succeeded in biasing it toward itself and we need to revoke that
1252 // bias. The revocation will occur in the runtime in the slow case.
1253 if (counters != NULL) {
1254 cond_inc32(Assembler::zero,
1255 ExternalAddress((address) counters->rebiased_lock_entry_count_addr()));
1256 }
1257 if (slow_case != NULL) {
1258 jcc(Assembler::notZero, *slow_case);
1259 }
1260 jmp(done);
1261
1262 bind(try_revoke_bias);
1263 // The prototype mark in the klass doesn't have the bias bit set any
1264 // more, indicating that objects of this data type are not supposed
1265 // to be biased any more. We are going to try to reset the mark of
1266 // this object to the prototype value and fall through to the
1267 // CAS-based locking scheme. Note that if our CAS fails, it means
1268 // that another thread raced us for the privilege of revoking the
1269 // bias of this particular object, so it's okay to continue in the
1270 // normal locking code.
1271 //
1272 // FIXME: due to a lack of registers we currently blow away the age
1273 // bits in this situation. Should attempt to preserve them.
1274 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1275 load_prototype_header(tmp_reg, obj_reg);
1276 lock();
1277 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1278 // Fall through to the normal CAS-based lock, because no matter what
1279 // the result of the above CAS, some thread must have succeeded in
1280 // removing the bias bit from the object's header.
1281 if (counters != NULL) {
1282 cond_inc32(Assembler::zero,
1283 ExternalAddress((address) counters->revoked_lock_entry_count_addr()));
1284 }
1285
1286 bind(cas_label);
1287
1288 return null_check_offset;
1289 }
1290
1291 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) {
1292 assert(UseBiasedLocking, "why call this otherwise?");
1293
1294 // Check for biased locking unlock case, which is a no-op
1295 // Note: we do not have to check the thread ID for two reasons.
1296 // First, the interpreter checks for IllegalMonitorStateException at
1297 // a higher level. Second, if the bias was revoked while we held the
1298 // lock, the object could not be rebiased toward another thread, so
1299 // the bias bit would be clear.
1300 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1301 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place);
1302 cmpptr(temp_reg, markOopDesc::biased_lock_pattern);
1303 jcc(Assembler::equal, done);
1304 }
1305
1306 #ifdef COMPILER2
1307
1308 #if INCLUDE_RTM_OPT
1309
1310 // Update rtm_counters based on abort status
1311 // input: abort_status
1312 // rtm_counters (RTMLockingCounters*)
1313 // flags are killed
1314 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) {
1315
1316 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset()));
1317 if (PrintPreciseRTMLockingStatistics) {
1318 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) {
1319 Label check_abort;
1320 testl(abort_status, (1<<i));
1321 jccb(Assembler::equal, check_abort);
1322 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx))));
1323 bind(check_abort);
1324 }
1325 }
1326 }
1327
1328 // Branch if (random & (count-1) != 0), count is 2^n
1329 // tmp, scr and flags are killed
1330 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) {
1331 assert(tmp == rax, "");
1332 assert(scr == rdx, "");
1333 rdtsc(); // modifies EDX:EAX
1334 andptr(tmp, count-1);
1335 jccb(Assembler::notZero, brLabel);
1336 }
1337
1338 // Perform abort ratio calculation, set no_rtm bit if high ratio
1339 // input: rtm_counters_Reg (RTMLockingCounters* address)
1340 // tmpReg, rtm_counters_Reg and flags are killed
1341 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg,
1342 Register rtm_counters_Reg,
1343 RTMLockingCounters* rtm_counters,
1344 Metadata* method_data) {
1345 Label L_done, L_check_always_rtm1, L_check_always_rtm2;
1346
1347 if (RTMLockingCalculationDelay > 0) {
1348 // Delay calculation
1349 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg);
1350 testptr(tmpReg, tmpReg);
1351 jccb(Assembler::equal, L_done);
1352 }
1353 // Abort ratio calculation only if abort_count > RTMAbortThreshold
1354 // Aborted transactions = abort_count * 100
1355 // All transactions = total_count * RTMTotalCountIncrRate
1356 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio)
1357
1358 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset()));
1359 cmpptr(tmpReg, RTMAbortThreshold);
1360 jccb(Assembler::below, L_check_always_rtm2);
1361 imulptr(tmpReg, tmpReg, 100);
1362
1363 Register scrReg = rtm_counters_Reg;
1364 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1365 imulptr(scrReg, scrReg, RTMTotalCountIncrRate);
1366 imulptr(scrReg, scrReg, RTMAbortRatio);
1367 cmpptr(tmpReg, scrReg);
1368 jccb(Assembler::below, L_check_always_rtm1);
1369 if (method_data != NULL) {
1370 // set rtm_state to "no rtm" in MDO
1371 mov_metadata(tmpReg, method_data);
1372 lock();
1373 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM);
1374 }
1375 jmpb(L_done);
1376 bind(L_check_always_rtm1);
1377 // Reload RTMLockingCounters* address
1378 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1379 bind(L_check_always_rtm2);
1380 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1381 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate);
1382 jccb(Assembler::below, L_done);
1383 if (method_data != NULL) {
1384 // set rtm_state to "always rtm" in MDO
1385 mov_metadata(tmpReg, method_data);
1386 lock();
1387 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM);
1388 }
1389 bind(L_done);
1390 }
1391
1392 // Update counters and perform abort ratio calculation
1393 // input: abort_status_Reg
1394 // rtm_counters_Reg, flags are killed
1395 void MacroAssembler::rtm_profiling(Register abort_status_Reg,
1396 Register rtm_counters_Reg,
1397 RTMLockingCounters* rtm_counters,
1398 Metadata* method_data,
1399 bool profile_rtm) {
1400
1401 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1402 // update rtm counters based on rax value at abort
1403 // reads abort_status_Reg, updates flags
1404 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1405 rtm_counters_update(abort_status_Reg, rtm_counters_Reg);
1406 if (profile_rtm) {
1407 // Save abort status because abort_status_Reg is used by following code.
1408 if (RTMRetryCount > 0) {
1409 push(abort_status_Reg);
1410 }
1411 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1412 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data);
1413 // restore abort status
1414 if (RTMRetryCount > 0) {
1415 pop(abort_status_Reg);
1416 }
1417 }
1418 }
1419
1420 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4)
1421 // inputs: retry_count_Reg
1422 // : abort_status_Reg
1423 // output: retry_count_Reg decremented by 1
1424 // flags are killed
1425 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) {
1426 Label doneRetry;
1427 assert(abort_status_Reg == rax, "");
1428 // The abort reason bits are in eax (see all states in rtmLocking.hpp)
1429 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4)
1430 // if reason is in 0x6 and retry count != 0 then retry
1431 andptr(abort_status_Reg, 0x6);
1432 jccb(Assembler::zero, doneRetry);
1433 testl(retry_count_Reg, retry_count_Reg);
1434 jccb(Assembler::zero, doneRetry);
1435 pause();
1436 decrementl(retry_count_Reg);
1437 jmp(retryLabel);
1438 bind(doneRetry);
1439 }
1440
1441 // Spin and retry if lock is busy,
1442 // inputs: box_Reg (monitor address)
1443 // : retry_count_Reg
1444 // output: retry_count_Reg decremented by 1
1445 // : clear z flag if retry count exceeded
1446 // tmp_Reg, scr_Reg, flags are killed
1447 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg,
1448 Register tmp_Reg, Register scr_Reg, Label& retryLabel) {
1449 Label SpinLoop, SpinExit, doneRetry;
1450 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1451
1452 testl(retry_count_Reg, retry_count_Reg);
1453 jccb(Assembler::zero, doneRetry);
1454 decrementl(retry_count_Reg);
1455 movptr(scr_Reg, RTMSpinLoopCount);
1456
1457 bind(SpinLoop);
1458 pause();
1459 decrementl(scr_Reg);
1460 jccb(Assembler::lessEqual, SpinExit);
1461 movptr(tmp_Reg, Address(box_Reg, owner_offset));
1462 testptr(tmp_Reg, tmp_Reg);
1463 jccb(Assembler::notZero, SpinLoop);
1464
1465 bind(SpinExit);
1466 jmp(retryLabel);
1467 bind(doneRetry);
1468 incrementl(retry_count_Reg); // clear z flag
1469 }
1470
1471 // Use RTM for normal stack locks
1472 // Input: objReg (object to lock)
1473 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg,
1474 Register retry_on_abort_count_Reg,
1475 RTMLockingCounters* stack_rtm_counters,
1476 Metadata* method_data, bool profile_rtm,
1477 Label& DONE_LABEL, Label& IsInflated) {
1478 assert(UseRTMForStackLocks, "why call this otherwise?");
1479 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1480 assert(tmpReg == rax, "");
1481 assert(scrReg == rdx, "");
1482 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1483
1484 if (RTMRetryCount > 0) {
1485 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1486 bind(L_rtm_retry);
1487 }
1488 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes()));
1489 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1490 jcc(Assembler::notZero, IsInflated);
1491
1492 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1493 Label L_noincrement;
1494 if (RTMTotalCountIncrRate > 1) {
1495 // tmpReg, scrReg and flags are killed
1496 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement);
1497 }
1498 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM");
1499 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg);
1500 bind(L_noincrement);
1501 }
1502 xbegin(L_on_abort);
1503 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword
1504 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1505 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1506 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked
1507
1508 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1509 if (UseRTMXendForLockBusy) {
1510 xend();
1511 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry)
1512 jmp(L_decrement_retry);
1513 }
1514 else {
1515 xabort(0);
1516 }
1517 bind(L_on_abort);
1518 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1519 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm);
1520 }
1521 bind(L_decrement_retry);
1522 if (RTMRetryCount > 0) {
1523 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1524 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1525 }
1526 }
1527
1528 // Use RTM for inflating locks
1529 // inputs: objReg (object to lock)
1530 // boxReg (on-stack box address (displaced header location) - KILLED)
1531 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value)
1532 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg,
1533 Register scrReg, Register retry_on_busy_count_Reg,
1534 Register retry_on_abort_count_Reg,
1535 RTMLockingCounters* rtm_counters,
1536 Metadata* method_data, bool profile_rtm,
1537 Label& DONE_LABEL) {
1538 assert(UseRTMLocking, "why call this otherwise?");
1539 assert(tmpReg == rax, "");
1540 assert(scrReg == rdx, "");
1541 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1542 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1543
1544 // Without cast to int32_t a movptr will destroy r10 which is typically obj
1545 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1546 movptr(boxReg, tmpReg); // Save ObjectMonitor address
1547
1548 if (RTMRetryCount > 0) {
1549 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy
1550 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1551 bind(L_rtm_retry);
1552 }
1553 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1554 Label L_noincrement;
1555 if (RTMTotalCountIncrRate > 1) {
1556 // tmpReg, scrReg and flags are killed
1557 branch_on_random_using_rdtsc(tmpReg, scrReg, RTMTotalCountIncrRate, L_noincrement);
1558 }
1559 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1560 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg);
1561 bind(L_noincrement);
1562 }
1563 xbegin(L_on_abort);
1564 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes()));
1565 movptr(tmpReg, Address(tmpReg, owner_offset));
1566 testptr(tmpReg, tmpReg);
1567 jcc(Assembler::zero, DONE_LABEL);
1568 if (UseRTMXendForLockBusy) {
1569 xend();
1570 jmp(L_decrement_retry);
1571 }
1572 else {
1573 xabort(0);
1574 }
1575 bind(L_on_abort);
1576 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1577 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1578 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm);
1579 }
1580 if (RTMRetryCount > 0) {
1581 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1582 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1583 }
1584
1585 movptr(tmpReg, Address(boxReg, owner_offset)) ;
1586 testptr(tmpReg, tmpReg) ;
1587 jccb(Assembler::notZero, L_decrement_retry) ;
1588
1589 // Appears unlocked - try to swing _owner from null to non-null.
1590 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1591 #ifdef _LP64
1592 Register threadReg = r15_thread;
1593 #else
1594 get_thread(scrReg);
1595 Register threadReg = scrReg;
1596 #endif
1597 lock();
1598 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg
1599
1600 if (RTMRetryCount > 0) {
1601 // success done else retry
1602 jccb(Assembler::equal, DONE_LABEL) ;
1603 bind(L_decrement_retry);
1604 // Spin and retry if lock is busy.
1605 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry);
1606 }
1607 else {
1608 bind(L_decrement_retry);
1609 }
1610 }
1611
1612 #endif // INCLUDE_RTM_OPT
1613
1614 // Fast_Lock and Fast_Unlock used by C2
1615
1616 // Because the transitions from emitted code to the runtime
1617 // monitorenter/exit helper stubs are so slow it's critical that
1618 // we inline both the stack-locking fast-path and the inflated fast path.
1619 //
1620 // See also: cmpFastLock and cmpFastUnlock.
1621 //
1622 // What follows is a specialized inline transliteration of the code
1623 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat
1624 // another option would be to emit TrySlowEnter and TrySlowExit methods
1625 // at startup-time. These methods would accept arguments as
1626 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
1627 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply
1628 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
1629 // In practice, however, the # of lock sites is bounded and is usually small.
1630 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
1631 // if the processor uses simple bimodal branch predictors keyed by EIP
1632 // Since the helper routines would be called from multiple synchronization
1633 // sites.
1634 //
1635 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
1636 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
1637 // to those specialized methods. That'd give us a mostly platform-independent
1638 // implementation that the JITs could optimize and inline at their pleasure.
1639 // Done correctly, the only time we'd need to cross to native could would be
1640 // to park() or unpark() threads. We'd also need a few more unsafe operators
1641 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
1642 // (b) explicit barriers or fence operations.
1643 //
1644 // TODO:
1645 //
1646 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
1647 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
1648 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
1649 // the lock operators would typically be faster than reifying Self.
1650 //
1651 // * Ideally I'd define the primitives as:
1652 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
1653 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
1654 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
1655 // Instead, we're stuck with a rather awkward and brittle register assignments below.
1656 // Furthermore the register assignments are overconstrained, possibly resulting in
1657 // sub-optimal code near the synchronization site.
1658 //
1659 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
1660 // Alternately, use a better sp-proximity test.
1661 //
1662 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
1663 // Either one is sufficient to uniquely identify a thread.
1664 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
1665 //
1666 // * Intrinsify notify() and notifyAll() for the common cases where the
1667 // object is locked by the calling thread but the waitlist is empty.
1668 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
1669 //
1670 // * use jccb and jmpb instead of jcc and jmp to improve code density.
1671 // But beware of excessive branch density on AMD Opterons.
1672 //
1673 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
1674 // or failure of the fast-path. If the fast-path fails then we pass
1675 // control to the slow-path, typically in C. In Fast_Lock and
1676 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2
1677 // will emit a conditional branch immediately after the node.
1678 // So we have branches to branches and lots of ICC.ZF games.
1679 // Instead, it might be better to have C2 pass a "FailureLabel"
1680 // into Fast_Lock and Fast_Unlock. In the case of success, control
1681 // will drop through the node. ICC.ZF is undefined at exit.
1682 // In the case of failure, the node will branch directly to the
1683 // FailureLabel
1684
1685
1686 // obj: object to lock
1687 // box: on-stack box address (displaced header location) - KILLED
1688 // rax,: tmp -- KILLED
1689 // scr: tmp -- KILLED
1690 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg,
1691 Register scrReg, Register cx1Reg, Register cx2Reg,
1692 BiasedLockingCounters* counters,
1693 RTMLockingCounters* rtm_counters,
1694 RTMLockingCounters* stack_rtm_counters,
1695 Metadata* method_data,
1696 bool use_rtm, bool profile_rtm) {
1697 // Ensure the register assignments are disjoint
1698 assert(tmpReg == rax, "");
1699
1700 if (use_rtm) {
1701 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg);
1702 } else {
1703 assert(cx1Reg == noreg, "");
1704 assert(cx2Reg == noreg, "");
1705 assert_different_registers(objReg, boxReg, tmpReg, scrReg);
1706 }
1707
1708 if (counters != NULL) {
1709 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg);
1710 }
1711
1712 // Possible cases that we'll encounter in fast_lock
1713 // ------------------------------------------------
1714 // * Inflated
1715 // -- unlocked
1716 // -- Locked
1717 // = by self
1718 // = by other
1719 // * biased
1720 // -- by Self
1721 // -- by other
1722 // * neutral
1723 // * stack-locked
1724 // -- by self
1725 // = sp-proximity test hits
1726 // = sp-proximity test generates false-negative
1727 // -- by other
1728 //
1729
1730 Label IsInflated, DONE_LABEL;
1731
1732 // it's stack-locked, biased or neutral
1733 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
1734 // order to reduce the number of conditional branches in the most common cases.
1735 // Beware -- there's a subtle invariant that fetch of the markword
1736 // at [FETCH], below, will never observe a biased encoding (*101b).
1737 // If this invariant is not held we risk exclusion (safety) failure.
1738 if (UseBiasedLocking && !UseOptoBiasInlining) {
1739 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters);
1740 }
1741
1742 #if INCLUDE_RTM_OPT
1743 if (UseRTMForStackLocks && use_rtm) {
1744 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg,
1745 stack_rtm_counters, method_data, profile_rtm,
1746 DONE_LABEL, IsInflated);
1747 }
1748 #endif // INCLUDE_RTM_OPT
1749
1750 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // [FETCH]
1751 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1752 jccb(Assembler::notZero, IsInflated);
1753
1754 // Attempt stack-locking ...
1755 orptr (tmpReg, markOopDesc::unlocked_value);
1756 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
1757 lock();
1758 cmpxchgptr(boxReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Updates tmpReg
1759 if (counters != NULL) {
1760 cond_inc32(Assembler::equal,
1761 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1762 }
1763 jcc(Assembler::equal, DONE_LABEL); // Success
1764
1765 // Recursive locking.
1766 // The object is stack-locked: markword contains stack pointer to BasicLock.
1767 // Locked by current thread if difference with current SP is less than one page.
1768 subptr(tmpReg, rsp);
1769 // Next instruction set ZFlag == 1 (Success) if difference is less then one page.
1770 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) );
1771 movptr(Address(boxReg, 0), tmpReg);
1772 if (counters != NULL) {
1773 cond_inc32(Assembler::equal,
1774 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1775 }
1776 jmp(DONE_LABEL);
1777
1778 bind(IsInflated);
1779 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value
1780
1781 #if INCLUDE_RTM_OPT
1782 // Use the same RTM locking code in 32- and 64-bit VM.
1783 if (use_rtm) {
1784 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg,
1785 rtm_counters, method_data, profile_rtm, DONE_LABEL);
1786 } else {
1787 #endif // INCLUDE_RTM_OPT
1788
1789 #ifndef _LP64
1790 // The object is inflated.
1791
1792 // boxReg refers to the on-stack BasicLock in the current frame.
1793 // We'd like to write:
1794 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices.
1795 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
1796 // additional latency as we have another ST in the store buffer that must drain.
1797
1798 // avoid ST-before-CAS
1799 // register juggle because we need tmpReg for cmpxchgptr below
1800 movptr(scrReg, boxReg);
1801 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1802
1803 // Optimistic form: consider XORL tmpReg,tmpReg
1804 movptr(tmpReg, NULL_WORD);
1805
1806 // Appears unlocked - try to swing _owner from null to non-null.
1807 // Ideally, I'd manifest "Self" with get_thread and then attempt
1808 // to CAS the register containing Self into m->Owner.
1809 // But we don't have enough registers, so instead we can either try to CAS
1810 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
1811 // we later store "Self" into m->Owner. Transiently storing a stack address
1812 // (rsp or the address of the box) into m->owner is harmless.
1813 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1814 lock();
1815 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1816 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3
1817 // If we weren't able to swing _owner from NULL to the BasicLock
1818 // then take the slow path.
1819 jccb (Assembler::notZero, DONE_LABEL);
1820 // update _owner from BasicLock to thread
1821 get_thread (scrReg); // beware: clobbers ICCs
1822 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg);
1823 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success
1824
1825 // If the CAS fails we can either retry or pass control to the slow-path.
1826 // We use the latter tactic.
1827 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1828 // If the CAS was successful ...
1829 // Self has acquired the lock
1830 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1831 // Intentional fall-through into DONE_LABEL ...
1832 #else // _LP64
1833 // It's inflated
1834 movq(scrReg, tmpReg);
1835 xorq(tmpReg, tmpReg);
1836
1837 lock();
1838 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1839 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark().
1840 // Without cast to int32_t movptr will destroy r10 which is typically obj.
1841 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1842 // Intentional fall-through into DONE_LABEL ...
1843 // Propagate ICC.ZF from CAS above into DONE_LABEL.
1844 #endif // _LP64
1845 #if INCLUDE_RTM_OPT
1846 } // use_rtm()
1847 #endif
1848 // DONE_LABEL is a hot target - we'd really like to place it at the
1849 // start of cache line by padding with NOPs.
1850 // See the AMD and Intel software optimization manuals for the
1851 // most efficient "long" NOP encodings.
1852 // Unfortunately none of our alignment mechanisms suffice.
1853 bind(DONE_LABEL);
1854
1855 // At DONE_LABEL the icc ZFlag is set as follows ...
1856 // Fast_Unlock uses the same protocol.
1857 // ZFlag == 1 -> Success
1858 // ZFlag == 0 -> Failure - force control through the slow-path
1859 }
1860
1861 // obj: object to unlock
1862 // box: box address (displaced header location), killed. Must be EAX.
1863 // tmp: killed, cannot be obj nor box.
1864 //
1865 // Some commentary on balanced locking:
1866 //
1867 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
1868 // Methods that don't have provably balanced locking are forced to run in the
1869 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
1870 // The interpreter provides two properties:
1871 // I1: At return-time the interpreter automatically and quietly unlocks any
1872 // objects acquired the current activation (frame). Recall that the
1873 // interpreter maintains an on-stack list of locks currently held by
1874 // a frame.
1875 // I2: If a method attempts to unlock an object that is not held by the
1876 // the frame the interpreter throws IMSX.
1877 //
1878 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
1879 // B() doesn't have provably balanced locking so it runs in the interpreter.
1880 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
1881 // is still locked by A().
1882 //
1883 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
1884 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
1885 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
1886 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
1887 // Arguably given that the spec legislates the JNI case as undefined our implementation
1888 // could reasonably *avoid* checking owner in Fast_Unlock().
1889 // In the interest of performance we elide m->Owner==Self check in unlock.
1890 // A perfectly viable alternative is to elide the owner check except when
1891 // Xcheck:jni is enabled.
1892
1893 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) {
1894 assert(boxReg == rax, "");
1895 assert_different_registers(objReg, boxReg, tmpReg);
1896
1897 Label DONE_LABEL, Stacked, CheckSucc;
1898
1899 // Critically, the biased locking test must have precedence over
1900 // and appear before the (box->dhw == 0) recursive stack-lock test.
1901 if (UseBiasedLocking && !UseOptoBiasInlining) {
1902 biased_locking_exit(objReg, tmpReg, DONE_LABEL);
1903 }
1904
1905 #if INCLUDE_RTM_OPT
1906 if (UseRTMForStackLocks && use_rtm) {
1907 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1908 Label L_regular_unlock;
1909 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // fetch markword
1910 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1911 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1912 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock
1913 xend(); // otherwise end...
1914 jmp(DONE_LABEL); // ... and we're done
1915 bind(L_regular_unlock);
1916 }
1917 #endif
1918
1919 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header
1920 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock
1921 movptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Examine the object's markword
1922 testptr(tmpReg, markOopDesc::monitor_value); // Inflated?
1923 jccb (Assembler::zero, Stacked);
1924
1925 // It's inflated.
1926 #if INCLUDE_RTM_OPT
1927 if (use_rtm) {
1928 Label L_regular_inflated_unlock;
1929 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1930 movptr(boxReg, Address(tmpReg, owner_offset));
1931 testptr(boxReg, boxReg);
1932 jccb(Assembler::notZero, L_regular_inflated_unlock);
1933 xend();
1934 jmpb(DONE_LABEL);
1935 bind(L_regular_inflated_unlock);
1936 }
1937 #endif
1938
1939 // Despite our balanced locking property we still check that m->_owner == Self
1940 // as java routines or native JNI code called by this thread might
1941 // have released the lock.
1942 // Refer to the comments in synchronizer.cpp for how we might encode extra
1943 // state in _succ so we can avoid fetching EntryList|cxq.
1944 //
1945 // I'd like to add more cases in fast_lock() and fast_unlock() --
1946 // such as recursive enter and exit -- but we have to be wary of
1947 // I$ bloat, T$ effects and BP$ effects.
1948 //
1949 // If there's no contention try a 1-0 exit. That is, exit without
1950 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
1951 // we detect and recover from the race that the 1-0 exit admits.
1952 //
1953 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
1954 // before it STs null into _owner, releasing the lock. Updates
1955 // to data protected by the critical section must be visible before
1956 // we drop the lock (and thus before any other thread could acquire
1957 // the lock and observe the fields protected by the lock).
1958 // IA32's memory-model is SPO, so STs are ordered with respect to
1959 // each other and there's no need for an explicit barrier (fence).
1960 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
1961 #ifndef _LP64
1962 get_thread (boxReg);
1963
1964 // Note that we could employ various encoding schemes to reduce
1965 // the number of loads below (currently 4) to just 2 or 3.
1966 // Refer to the comments in synchronizer.cpp.
1967 // In practice the chain of fetches doesn't seem to impact performance, however.
1968 xorptr(boxReg, boxReg);
1969 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
1970 jccb (Assembler::notZero, DONE_LABEL);
1971 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
1972 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
1973 jccb (Assembler::notZero, CheckSucc);
1974 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
1975 jmpb (DONE_LABEL);
1976
1977 bind (Stacked);
1978 // It's not inflated and it's not recursively stack-locked and it's not biased.
1979 // It must be stack-locked.
1980 // Try to reset the header to displaced header.
1981 // The "box" value on the stack is stable, so we can reload
1982 // and be assured we observe the same value as above.
1983 movptr(tmpReg, Address(boxReg, 0));
1984 lock();
1985 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box
1986 // Intention fall-thru into DONE_LABEL
1987
1988 // DONE_LABEL is a hot target - we'd really like to place it at the
1989 // start of cache line by padding with NOPs.
1990 // See the AMD and Intel software optimization manuals for the
1991 // most efficient "long" NOP encodings.
1992 // Unfortunately none of our alignment mechanisms suffice.
1993 bind (CheckSucc);
1994 #else // _LP64
1995 // It's inflated
1996 xorptr(boxReg, boxReg);
1997 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
1998 jccb (Assembler::notZero, DONE_LABEL);
1999 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2000 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2001 jccb (Assembler::notZero, CheckSucc);
2002 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2003 jmpb (DONE_LABEL);
2004
2005 // Try to avoid passing control into the slow_path ...
2006 Label LSuccess, LGoSlowPath ;
2007 bind (CheckSucc);
2008
2009 // The following optional optimization can be elided if necessary
2010 // Effectively: if (succ == null) goto SlowPath
2011 // The code reduces the window for a race, however,
2012 // and thus benefits performance.
2013 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2014 jccb (Assembler::zero, LGoSlowPath);
2015
2016 xorptr(boxReg, boxReg);
2017 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2018
2019 // Memory barrier/fence
2020 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ
2021 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack.
2022 // This is faster on Nehalem and AMD Shanghai/Barcelona.
2023 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2024 // We might also restructure (ST Owner=0;barrier;LD _Succ) to
2025 // (mov box,0; xchgq box, &m->Owner; LD _succ) .
2026 lock(); addl(Address(rsp, 0), 0);
2027
2028 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2029 jccb (Assembler::notZero, LSuccess);
2030
2031 // Rare inopportune interleaving - race.
2032 // The successor vanished in the small window above.
2033 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor.
2034 // We need to ensure progress and succession.
2035 // Try to reacquire the lock.
2036 // If that fails then the new owner is responsible for succession and this
2037 // thread needs to take no further action and can exit via the fast path (success).
2038 // If the re-acquire succeeds then pass control into the slow path.
2039 // As implemented, this latter mode is horrible because we generated more
2040 // coherence traffic on the lock *and* artifically extended the critical section
2041 // length while by virtue of passing control into the slow path.
2042
2043 // box is really RAX -- the following CMPXCHG depends on that binding
2044 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R)
2045 lock();
2046 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2047 // There's no successor so we tried to regrab the lock.
2048 // If that didn't work, then another thread grabbed the
2049 // lock so we're done (and exit was a success).
2050 jccb (Assembler::notEqual, LSuccess);
2051 // Intentional fall-through into slow-path
2052
2053 bind (LGoSlowPath);
2054 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure
2055 jmpb (DONE_LABEL);
2056
2057 bind (LSuccess);
2058 testl (boxReg, 0); // set ICC.ZF=1 to indicate success
2059 jmpb (DONE_LABEL);
2060
2061 bind (Stacked);
2062 movptr(tmpReg, Address (boxReg, 0)); // re-fetch
2063 lock();
2064 cmpxchgptr(tmpReg, Address(objReg, oopDesc::mark_offset_in_bytes())); // Uses RAX which is box
2065
2066 #endif
2067 bind(DONE_LABEL);
2068 }
2069 #endif // COMPILER2
2070
2071 void MacroAssembler::c2bool(Register x) {
2072 // implements x == 0 ? 0 : 1
2073 // note: must only look at least-significant byte of x
2074 // since C-style booleans are stored in one byte
2075 // only! (was bug)
2076 andl(x, 0xFF);
2077 setb(Assembler::notZero, x);
2078 }
2079
2080 // Wouldn't need if AddressLiteral version had new name
2081 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) {
2082 Assembler::call(L, rtype);
2083 }
2084
2085 void MacroAssembler::call(Register entry) {
2086 Assembler::call(entry);
2087 }
2088
2089 void MacroAssembler::call(AddressLiteral entry) {
2090 if (reachable(entry)) {
2091 Assembler::call_literal(entry.target(), entry.rspec());
2092 } else {
2093 lea(rscratch1, entry);
2094 Assembler::call(rscratch1);
2095 }
2096 }
2097
2098 void MacroAssembler::ic_call(address entry, jint method_index) {
2099 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index);
2100 movptr(rax, (intptr_t)Universe::non_oop_word());
2101 call(AddressLiteral(entry, rh));
2102 }
2103
2104 // Implementation of call_VM versions
2105
2106 void MacroAssembler::call_VM(Register oop_result,
2107 address entry_point,
2108 bool check_exceptions) {
2109 Label C, E;
2110 call(C, relocInfo::none);
2111 jmp(E);
2112
2113 bind(C);
2114 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
2115 ret(0);
2116
2117 bind(E);
2118 }
2119
2120 void MacroAssembler::call_VM(Register oop_result,
2121 address entry_point,
2122 Register arg_1,
2123 bool check_exceptions) {
2124 Label C, E;
2125 call(C, relocInfo::none);
2126 jmp(E);
2127
2128 bind(C);
2129 pass_arg1(this, arg_1);
2130 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
2131 ret(0);
2132
2133 bind(E);
2134 }
2135
2136 void MacroAssembler::call_VM(Register oop_result,
2137 address entry_point,
2138 Register arg_1,
2139 Register arg_2,
2140 bool check_exceptions) {
2141 Label C, E;
2142 call(C, relocInfo::none);
2143 jmp(E);
2144
2145 bind(C);
2146
2147 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2148
2149 pass_arg2(this, arg_2);
2150 pass_arg1(this, arg_1);
2151 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
2152 ret(0);
2153
2154 bind(E);
2155 }
2156
2157 void MacroAssembler::call_VM(Register oop_result,
2158 address entry_point,
2159 Register arg_1,
2160 Register arg_2,
2161 Register arg_3,
2162 bool check_exceptions) {
2163 Label C, E;
2164 call(C, relocInfo::none);
2165 jmp(E);
2166
2167 bind(C);
2168
2169 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2170 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2171 pass_arg3(this, arg_3);
2172
2173 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2174 pass_arg2(this, arg_2);
2175
2176 pass_arg1(this, arg_1);
2177 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
2178 ret(0);
2179
2180 bind(E);
2181 }
2182
2183 void MacroAssembler::call_VM(Register oop_result,
2184 Register last_java_sp,
2185 address entry_point,
2186 int number_of_arguments,
2187 bool check_exceptions) {
2188 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2189 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2190 }
2191
2192 void MacroAssembler::call_VM(Register oop_result,
2193 Register last_java_sp,
2194 address entry_point,
2195 Register arg_1,
2196 bool check_exceptions) {
2197 pass_arg1(this, arg_1);
2198 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2199 }
2200
2201 void MacroAssembler::call_VM(Register oop_result,
2202 Register last_java_sp,
2203 address entry_point,
2204 Register arg_1,
2205 Register arg_2,
2206 bool check_exceptions) {
2207
2208 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2209 pass_arg2(this, arg_2);
2210 pass_arg1(this, arg_1);
2211 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2212 }
2213
2214 void MacroAssembler::call_VM(Register oop_result,
2215 Register last_java_sp,
2216 address entry_point,
2217 Register arg_1,
2218 Register arg_2,
2219 Register arg_3,
2220 bool check_exceptions) {
2221 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2222 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2223 pass_arg3(this, arg_3);
2224 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2225 pass_arg2(this, arg_2);
2226 pass_arg1(this, arg_1);
2227 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2228 }
2229
2230 void MacroAssembler::super_call_VM(Register oop_result,
2231 Register last_java_sp,
2232 address entry_point,
2233 int number_of_arguments,
2234 bool check_exceptions) {
2235 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2236 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2237 }
2238
2239 void MacroAssembler::super_call_VM(Register oop_result,
2240 Register last_java_sp,
2241 address entry_point,
2242 Register arg_1,
2243 bool check_exceptions) {
2244 pass_arg1(this, arg_1);
2245 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2246 }
2247
2248 void MacroAssembler::super_call_VM(Register oop_result,
2249 Register last_java_sp,
2250 address entry_point,
2251 Register arg_1,
2252 Register arg_2,
2253 bool check_exceptions) {
2254
2255 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2256 pass_arg2(this, arg_2);
2257 pass_arg1(this, arg_1);
2258 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2259 }
2260
2261 void MacroAssembler::super_call_VM(Register oop_result,
2262 Register last_java_sp,
2263 address entry_point,
2264 Register arg_1,
2265 Register arg_2,
2266 Register arg_3,
2267 bool check_exceptions) {
2268 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2269 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2270 pass_arg3(this, arg_3);
2271 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2272 pass_arg2(this, arg_2);
2273 pass_arg1(this, arg_1);
2274 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2275 }
2276
2277 void MacroAssembler::call_VM_base(Register oop_result,
2278 Register java_thread,
2279 Register last_java_sp,
2280 address entry_point,
2281 int number_of_arguments,
2282 bool check_exceptions) {
2283 // determine java_thread register
2284 if (!java_thread->is_valid()) {
2285 #ifdef _LP64
2286 java_thread = r15_thread;
2287 #else
2288 java_thread = rdi;
2289 get_thread(java_thread);
2290 #endif // LP64
2291 }
2292 // determine last_java_sp register
2293 if (!last_java_sp->is_valid()) {
2294 last_java_sp = rsp;
2295 }
2296 // debugging support
2297 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
2298 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register"));
2299 #ifdef ASSERT
2300 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
2301 // r12 is the heapbase.
2302 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");)
2303 #endif // ASSERT
2304
2305 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
2306 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
2307
2308 // push java thread (becomes first argument of C function)
2309
2310 NOT_LP64(push(java_thread); number_of_arguments++);
2311 LP64_ONLY(mov(c_rarg0, r15_thread));
2312
2313 // set last Java frame before call
2314 assert(last_java_sp != rbp, "can't use ebp/rbp");
2315
2316 // Only interpreter should have to set fp
2317 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL);
2318
2319 // do the call, remove parameters
2320 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
2321
2322 // restore the thread (cannot use the pushed argument since arguments
2323 // may be overwritten by C code generated by an optimizing compiler);
2324 // however can use the register value directly if it is callee saved.
2325 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) {
2326 // rdi & rsi (also r15) are callee saved -> nothing to do
2327 #ifdef ASSERT
2328 guarantee(java_thread != rax, "change this code");
2329 push(rax);
2330 { Label L;
2331 get_thread(rax);
2332 cmpptr(java_thread, rax);
2333 jcc(Assembler::equal, L);
2334 STOP("MacroAssembler::call_VM_base: rdi not callee saved?");
2335 bind(L);
2336 }
2337 pop(rax);
2338 #endif
2339 } else {
2340 get_thread(java_thread);
2341 }
2342 // reset last Java frame
2343 // Only interpreter should have to clear fp
2344 reset_last_Java_frame(java_thread, true);
2345
2346 // C++ interp handles this in the interpreter
2347 check_and_handle_popframe(java_thread);
2348 check_and_handle_earlyret(java_thread);
2349
2350 if (check_exceptions) {
2351 // check for pending exceptions (java_thread is set upon return)
2352 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD);
2353 #ifndef _LP64
2354 jump_cc(Assembler::notEqual,
2355 RuntimeAddress(StubRoutines::forward_exception_entry()));
2356 #else
2357 // This used to conditionally jump to forward_exception however it is
2358 // possible if we relocate that the branch will not reach. So we must jump
2359 // around so we can always reach
2360
2361 Label ok;
2362 jcc(Assembler::equal, ok);
2363 jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2364 bind(ok);
2365 #endif // LP64
2366 }
2367
2368 // get oop result if there is one and reset the value in the thread
2369 if (oop_result->is_valid()) {
2370 get_vm_result(oop_result, java_thread);
2371 }
2372 }
2373
2374 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
2375
2376 // Calculate the value for last_Java_sp
2377 // somewhat subtle. call_VM does an intermediate call
2378 // which places a return address on the stack just under the
2379 // stack pointer as the user finsihed with it. This allows
2380 // use to retrieve last_Java_pc from last_Java_sp[-1].
2381 // On 32bit we then have to push additional args on the stack to accomplish
2382 // the actual requested call. On 64bit call_VM only can use register args
2383 // so the only extra space is the return address that call_VM created.
2384 // This hopefully explains the calculations here.
2385
2386 #ifdef _LP64
2387 // We've pushed one address, correct last_Java_sp
2388 lea(rax, Address(rsp, wordSize));
2389 #else
2390 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize));
2391 #endif // LP64
2392
2393 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions);
2394
2395 }
2396
2397 // Use this method when MacroAssembler version of call_VM_leaf_base() should be called from Interpreter.
2398 void MacroAssembler::call_VM_leaf0(address entry_point) {
2399 MacroAssembler::call_VM_leaf_base(entry_point, 0);
2400 }
2401
2402 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
2403 call_VM_leaf_base(entry_point, number_of_arguments);
2404 }
2405
2406 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
2407 pass_arg0(this, arg_0);
2408 call_VM_leaf(entry_point, 1);
2409 }
2410
2411 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2412
2413 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2414 pass_arg1(this, arg_1);
2415 pass_arg0(this, arg_0);
2416 call_VM_leaf(entry_point, 2);
2417 }
2418
2419 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2420 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2421 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2422 pass_arg2(this, arg_2);
2423 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2424 pass_arg1(this, arg_1);
2425 pass_arg0(this, arg_0);
2426 call_VM_leaf(entry_point, 3);
2427 }
2428
2429 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2430 pass_arg0(this, arg_0);
2431 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2432 }
2433
2434 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2435
2436 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2437 pass_arg1(this, arg_1);
2438 pass_arg0(this, arg_0);
2439 MacroAssembler::call_VM_leaf_base(entry_point, 2);
2440 }
2441
2442 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2443 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2444 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2445 pass_arg2(this, arg_2);
2446 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2447 pass_arg1(this, arg_1);
2448 pass_arg0(this, arg_0);
2449 MacroAssembler::call_VM_leaf_base(entry_point, 3);
2450 }
2451
2452 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2453 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg"));
2454 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2455 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2456 pass_arg3(this, arg_3);
2457 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2458 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2459 pass_arg2(this, arg_2);
2460 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2461 pass_arg1(this, arg_1);
2462 pass_arg0(this, arg_0);
2463 MacroAssembler::call_VM_leaf_base(entry_point, 4);
2464 }
2465
2466 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
2467 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
2468 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD);
2469 verify_oop(oop_result, "broken oop in call_VM_base");
2470 }
2471
2472 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
2473 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
2474 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD);
2475 }
2476
2477 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2478 }
2479
2480 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2481 }
2482
2483 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) {
2484 if (reachable(src1)) {
2485 cmpl(as_Address(src1), imm);
2486 } else {
2487 lea(rscratch1, src1);
2488 cmpl(Address(rscratch1, 0), imm);
2489 }
2490 }
2491
2492 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) {
2493 assert(!src2.is_lval(), "use cmpptr");
2494 if (reachable(src2)) {
2495 cmpl(src1, as_Address(src2));
2496 } else {
2497 lea(rscratch1, src2);
2498 cmpl(src1, Address(rscratch1, 0));
2499 }
2500 }
2501
2502 void MacroAssembler::cmp32(Register src1, int32_t imm) {
2503 Assembler::cmpl(src1, imm);
2504 }
2505
2506 void MacroAssembler::cmp32(Register src1, Address src2) {
2507 Assembler::cmpl(src1, src2);
2508 }
2509
2510 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2511 ucomisd(opr1, opr2);
2512
2513 Label L;
2514 if (unordered_is_less) {
2515 movl(dst, -1);
2516 jcc(Assembler::parity, L);
2517 jcc(Assembler::below , L);
2518 movl(dst, 0);
2519 jcc(Assembler::equal , L);
2520 increment(dst);
2521 } else { // unordered is greater
2522 movl(dst, 1);
2523 jcc(Assembler::parity, L);
2524 jcc(Assembler::above , L);
2525 movl(dst, 0);
2526 jcc(Assembler::equal , L);
2527 decrementl(dst);
2528 }
2529 bind(L);
2530 }
2531
2532 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2533 ucomiss(opr1, opr2);
2534
2535 Label L;
2536 if (unordered_is_less) {
2537 movl(dst, -1);
2538 jcc(Assembler::parity, L);
2539 jcc(Assembler::below , L);
2540 movl(dst, 0);
2541 jcc(Assembler::equal , L);
2542 increment(dst);
2543 } else { // unordered is greater
2544 movl(dst, 1);
2545 jcc(Assembler::parity, L);
2546 jcc(Assembler::above , L);
2547 movl(dst, 0);
2548 jcc(Assembler::equal , L);
2549 decrementl(dst);
2550 }
2551 bind(L);
2552 }
2553
2554
2555 void MacroAssembler::cmp8(AddressLiteral src1, int imm) {
2556 if (reachable(src1)) {
2557 cmpb(as_Address(src1), imm);
2558 } else {
2559 lea(rscratch1, src1);
2560 cmpb(Address(rscratch1, 0), imm);
2561 }
2562 }
2563
2564 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) {
2565 #ifdef _LP64
2566 if (src2.is_lval()) {
2567 movptr(rscratch1, src2);
2568 Assembler::cmpq(src1, rscratch1);
2569 } else if (reachable(src2)) {
2570 cmpq(src1, as_Address(src2));
2571 } else {
2572 lea(rscratch1, src2);
2573 Assembler::cmpq(src1, Address(rscratch1, 0));
2574 }
2575 #else
2576 if (src2.is_lval()) {
2577 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2578 } else {
2579 cmpl(src1, as_Address(src2));
2580 }
2581 #endif // _LP64
2582 }
2583
2584 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) {
2585 assert(src2.is_lval(), "not a mem-mem compare");
2586 #ifdef _LP64
2587 // moves src2's literal address
2588 movptr(rscratch1, src2);
2589 Assembler::cmpq(src1, rscratch1);
2590 #else
2591 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2592 #endif // _LP64
2593 }
2594
2595 void MacroAssembler::cmpoop(Register src1, Register src2) {
2596 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2597 bs->obj_equals(this, src1, src2);
2598 }
2599
2600 void MacroAssembler::cmpoop(Register src1, Address src2) {
2601 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2602 bs->obj_equals(this, src1, src2);
2603 }
2604
2605 #ifdef _LP64
2606 void MacroAssembler::cmpoop(Register src1, jobject src2) {
2607 movoop(rscratch1, src2);
2608 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2609 bs->obj_equals(this, src1, rscratch1);
2610 }
2611 #endif
2612
2613 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) {
2614 if (reachable(adr)) {
2615 lock();
2616 cmpxchgptr(reg, as_Address(adr));
2617 } else {
2618 lea(rscratch1, adr);
2619 lock();
2620 cmpxchgptr(reg, Address(rscratch1, 0));
2621 }
2622 }
2623
2624 void MacroAssembler::cmpxchgptr(Register reg, Address adr) {
2625 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr));
2626 }
2627
2628 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) {
2629 if (reachable(src)) {
2630 Assembler::comisd(dst, as_Address(src));
2631 } else {
2632 lea(rscratch1, src);
2633 Assembler::comisd(dst, Address(rscratch1, 0));
2634 }
2635 }
2636
2637 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) {
2638 if (reachable(src)) {
2639 Assembler::comiss(dst, as_Address(src));
2640 } else {
2641 lea(rscratch1, src);
2642 Assembler::comiss(dst, Address(rscratch1, 0));
2643 }
2644 }
2645
2646
2647 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) {
2648 Condition negated_cond = negate_condition(cond);
2649 Label L;
2650 jcc(negated_cond, L);
2651 pushf(); // Preserve flags
2652 atomic_incl(counter_addr);
2653 popf();
2654 bind(L);
2655 }
2656
2657 int MacroAssembler::corrected_idivl(Register reg) {
2658 // Full implementation of Java idiv and irem; checks for
2659 // special case as described in JVM spec., p.243 & p.271.
2660 // The function returns the (pc) offset of the idivl
2661 // instruction - may be needed for implicit exceptions.
2662 //
2663 // normal case special case
2664 //
2665 // input : rax,: dividend min_int
2666 // reg: divisor (may not be rax,/rdx) -1
2667 //
2668 // output: rax,: quotient (= rax, idiv reg) min_int
2669 // rdx: remainder (= rax, irem reg) 0
2670 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register");
2671 const int min_int = 0x80000000;
2672 Label normal_case, special_case;
2673
2674 // check for special case
2675 cmpl(rax, min_int);
2676 jcc(Assembler::notEqual, normal_case);
2677 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0)
2678 cmpl(reg, -1);
2679 jcc(Assembler::equal, special_case);
2680
2681 // handle normal case
2682 bind(normal_case);
2683 cdql();
2684 int idivl_offset = offset();
2685 idivl(reg);
2686
2687 // normal and special case exit
2688 bind(special_case);
2689
2690 return idivl_offset;
2691 }
2692
2693
2694
2695 void MacroAssembler::decrementl(Register reg, int value) {
2696 if (value == min_jint) {subl(reg, value) ; return; }
2697 if (value < 0) { incrementl(reg, -value); return; }
2698 if (value == 0) { ; return; }
2699 if (value == 1 && UseIncDec) { decl(reg) ; return; }
2700 /* else */ { subl(reg, value) ; return; }
2701 }
2702
2703 void MacroAssembler::decrementl(Address dst, int value) {
2704 if (value == min_jint) {subl(dst, value) ; return; }
2705 if (value < 0) { incrementl(dst, -value); return; }
2706 if (value == 0) { ; return; }
2707 if (value == 1 && UseIncDec) { decl(dst) ; return; }
2708 /* else */ { subl(dst, value) ; return; }
2709 }
2710
2711 void MacroAssembler::division_with_shift (Register reg, int shift_value) {
2712 assert (shift_value > 0, "illegal shift value");
2713 Label _is_positive;
2714 testl (reg, reg);
2715 jcc (Assembler::positive, _is_positive);
2716 int offset = (1 << shift_value) - 1 ;
2717
2718 if (offset == 1) {
2719 incrementl(reg);
2720 } else {
2721 addl(reg, offset);
2722 }
2723
2724 bind (_is_positive);
2725 sarl(reg, shift_value);
2726 }
2727
2728 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) {
2729 if (reachable(src)) {
2730 Assembler::divsd(dst, as_Address(src));
2731 } else {
2732 lea(rscratch1, src);
2733 Assembler::divsd(dst, Address(rscratch1, 0));
2734 }
2735 }
2736
2737 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) {
2738 if (reachable(src)) {
2739 Assembler::divss(dst, as_Address(src));
2740 } else {
2741 lea(rscratch1, src);
2742 Assembler::divss(dst, Address(rscratch1, 0));
2743 }
2744 }
2745
2746 // !defined(COMPILER2) is because of stupid core builds
2747 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI
2748 void MacroAssembler::empty_FPU_stack() {
2749 if (VM_Version::supports_mmx()) {
2750 emms();
2751 } else {
2752 for (int i = 8; i-- > 0; ) ffree(i);
2753 }
2754 }
2755 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI
2756
2757
2758 void MacroAssembler::enter() {
2759 push(rbp);
2760 mov(rbp, rsp);
2761 }
2762
2763 // A 5 byte nop that is safe for patching (see patch_verified_entry)
2764 void MacroAssembler::fat_nop() {
2765 if (UseAddressNop) {
2766 addr_nop_5();
2767 } else {
2768 emit_int8(0x26); // es:
2769 emit_int8(0x2e); // cs:
2770 emit_int8(0x64); // fs:
2771 emit_int8(0x65); // gs:
2772 emit_int8((unsigned char)0x90);
2773 }
2774 }
2775
2776 void MacroAssembler::fcmp(Register tmp) {
2777 fcmp(tmp, 1, true, true);
2778 }
2779
2780 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) {
2781 assert(!pop_right || pop_left, "usage error");
2782 if (VM_Version::supports_cmov()) {
2783 assert(tmp == noreg, "unneeded temp");
2784 if (pop_left) {
2785 fucomip(index);
2786 } else {
2787 fucomi(index);
2788 }
2789 if (pop_right) {
2790 fpop();
2791 }
2792 } else {
2793 assert(tmp != noreg, "need temp");
2794 if (pop_left) {
2795 if (pop_right) {
2796 fcompp();
2797 } else {
2798 fcomp(index);
2799 }
2800 } else {
2801 fcom(index);
2802 }
2803 // convert FPU condition into eflags condition via rax,
2804 save_rax(tmp);
2805 fwait(); fnstsw_ax();
2806 sahf();
2807 restore_rax(tmp);
2808 }
2809 // condition codes set as follows:
2810 //
2811 // CF (corresponds to C0) if x < y
2812 // PF (corresponds to C2) if unordered
2813 // ZF (corresponds to C3) if x = y
2814 }
2815
2816 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) {
2817 fcmp2int(dst, unordered_is_less, 1, true, true);
2818 }
2819
2820 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) {
2821 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right);
2822 Label L;
2823 if (unordered_is_less) {
2824 movl(dst, -1);
2825 jcc(Assembler::parity, L);
2826 jcc(Assembler::below , L);
2827 movl(dst, 0);
2828 jcc(Assembler::equal , L);
2829 increment(dst);
2830 } else { // unordered is greater
2831 movl(dst, 1);
2832 jcc(Assembler::parity, L);
2833 jcc(Assembler::above , L);
2834 movl(dst, 0);
2835 jcc(Assembler::equal , L);
2836 decrementl(dst);
2837 }
2838 bind(L);
2839 }
2840
2841 void MacroAssembler::fld_d(AddressLiteral src) {
2842 fld_d(as_Address(src));
2843 }
2844
2845 void MacroAssembler::fld_s(AddressLiteral src) {
2846 fld_s(as_Address(src));
2847 }
2848
2849 void MacroAssembler::fld_x(AddressLiteral src) {
2850 Assembler::fld_x(as_Address(src));
2851 }
2852
2853 void MacroAssembler::fldcw(AddressLiteral src) {
2854 Assembler::fldcw(as_Address(src));
2855 }
2856
2857 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) {
2858 if (reachable(src)) {
2859 Assembler::mulpd(dst, as_Address(src));
2860 } else {
2861 lea(rscratch1, src);
2862 Assembler::mulpd(dst, Address(rscratch1, 0));
2863 }
2864 }
2865
2866 void MacroAssembler::increase_precision() {
2867 subptr(rsp, BytesPerWord);
2868 fnstcw(Address(rsp, 0));
2869 movl(rax, Address(rsp, 0));
2870 orl(rax, 0x300);
2871 push(rax);
2872 fldcw(Address(rsp, 0));
2873 pop(rax);
2874 }
2875
2876 void MacroAssembler::restore_precision() {
2877 fldcw(Address(rsp, 0));
2878 addptr(rsp, BytesPerWord);
2879 }
2880
2881 void MacroAssembler::fpop() {
2882 ffree();
2883 fincstp();
2884 }
2885
2886 void MacroAssembler::load_float(Address src) {
2887 if (UseSSE >= 1) {
2888 movflt(xmm0, src);
2889 } else {
2890 LP64_ONLY(ShouldNotReachHere());
2891 NOT_LP64(fld_s(src));
2892 }
2893 }
2894
2895 void MacroAssembler::store_float(Address dst) {
2896 if (UseSSE >= 1) {
2897 movflt(dst, xmm0);
2898 } else {
2899 LP64_ONLY(ShouldNotReachHere());
2900 NOT_LP64(fstp_s(dst));
2901 }
2902 }
2903
2904 void MacroAssembler::load_double(Address src) {
2905 if (UseSSE >= 2) {
2906 movdbl(xmm0, src);
2907 } else {
2908 LP64_ONLY(ShouldNotReachHere());
2909 NOT_LP64(fld_d(src));
2910 }
2911 }
2912
2913 void MacroAssembler::store_double(Address dst) {
2914 if (UseSSE >= 2) {
2915 movdbl(dst, xmm0);
2916 } else {
2917 LP64_ONLY(ShouldNotReachHere());
2918 NOT_LP64(fstp_d(dst));
2919 }
2920 }
2921
2922 void MacroAssembler::fremr(Register tmp) {
2923 save_rax(tmp);
2924 { Label L;
2925 bind(L);
2926 fprem();
2927 fwait(); fnstsw_ax();
2928 #ifdef _LP64
2929 testl(rax, 0x400);
2930 jcc(Assembler::notEqual, L);
2931 #else
2932 sahf();
2933 jcc(Assembler::parity, L);
2934 #endif // _LP64
2935 }
2936 restore_rax(tmp);
2937 // Result is in ST0.
2938 // Note: fxch & fpop to get rid of ST1
2939 // (otherwise FPU stack could overflow eventually)
2940 fxch(1);
2941 fpop();
2942 }
2943
2944 // dst = c = a * b + c
2945 void MacroAssembler::fmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) {
2946 Assembler::vfmadd231sd(c, a, b);
2947 if (dst != c) {
2948 movdbl(dst, c);
2949 }
2950 }
2951
2952 // dst = c = a * b + c
2953 void MacroAssembler::fmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) {
2954 Assembler::vfmadd231ss(c, a, b);
2955 if (dst != c) {
2956 movflt(dst, c);
2957 }
2958 }
2959
2960 // dst = c = a * b + c
2961 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) {
2962 Assembler::vfmadd231pd(c, a, b, vector_len);
2963 if (dst != c) {
2964 vmovdqu(dst, c);
2965 }
2966 }
2967
2968 // dst = c = a * b + c
2969 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) {
2970 Assembler::vfmadd231ps(c, a, b, vector_len);
2971 if (dst != c) {
2972 vmovdqu(dst, c);
2973 }
2974 }
2975
2976 // dst = c = a * b + c
2977 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) {
2978 Assembler::vfmadd231pd(c, a, b, vector_len);
2979 if (dst != c) {
2980 vmovdqu(dst, c);
2981 }
2982 }
2983
2984 // dst = c = a * b + c
2985 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) {
2986 Assembler::vfmadd231ps(c, a, b, vector_len);
2987 if (dst != c) {
2988 vmovdqu(dst, c);
2989 }
2990 }
2991
2992 void MacroAssembler::incrementl(AddressLiteral dst) {
2993 if (reachable(dst)) {
2994 incrementl(as_Address(dst));
2995 } else {
2996 lea(rscratch1, dst);
2997 incrementl(Address(rscratch1, 0));
2998 }
2999 }
3000
3001 void MacroAssembler::incrementl(ArrayAddress dst) {
3002 incrementl(as_Address(dst));
3003 }
3004
3005 void MacroAssembler::incrementl(Register reg, int value) {
3006 if (value == min_jint) {addl(reg, value) ; return; }
3007 if (value < 0) { decrementl(reg, -value); return; }
3008 if (value == 0) { ; return; }
3009 if (value == 1 && UseIncDec) { incl(reg) ; return; }
3010 /* else */ { addl(reg, value) ; return; }
3011 }
3012
3013 void MacroAssembler::incrementl(Address dst, int value) {
3014 if (value == min_jint) {addl(dst, value) ; return; }
3015 if (value < 0) { decrementl(dst, -value); return; }
3016 if (value == 0) { ; return; }
3017 if (value == 1 && UseIncDec) { incl(dst) ; return; }
3018 /* else */ { addl(dst, value) ; return; }
3019 }
3020
3021 void MacroAssembler::jump(AddressLiteral dst) {
3022 if (reachable(dst)) {
3023 jmp_literal(dst.target(), dst.rspec());
3024 } else {
3025 lea(rscratch1, dst);
3026 jmp(rscratch1);
3027 }
3028 }
3029
3030 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) {
3031 if (reachable(dst)) {
3032 InstructionMark im(this);
3033 relocate(dst.reloc());
3034 const int short_size = 2;
3035 const int long_size = 6;
3036 int offs = (intptr_t)dst.target() - ((intptr_t)pc());
3037 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
3038 // 0111 tttn #8-bit disp
3039 emit_int8(0x70 | cc);
3040 emit_int8((offs - short_size) & 0xFF);
3041 } else {
3042 // 0000 1111 1000 tttn #32-bit disp
3043 emit_int8(0x0F);
3044 emit_int8((unsigned char)(0x80 | cc));
3045 emit_int32(offs - long_size);
3046 }
3047 } else {
3048 #ifdef ASSERT
3049 warning("reversing conditional branch");
3050 #endif /* ASSERT */
3051 Label skip;
3052 jccb(reverse[cc], skip);
3053 lea(rscratch1, dst);
3054 Assembler::jmp(rscratch1);
3055 bind(skip);
3056 }
3057 }
3058
3059 void MacroAssembler::ldmxcsr(AddressLiteral src) {
3060 if (reachable(src)) {
3061 Assembler::ldmxcsr(as_Address(src));
3062 } else {
3063 lea(rscratch1, src);
3064 Assembler::ldmxcsr(Address(rscratch1, 0));
3065 }
3066 }
3067
3068 int MacroAssembler::load_signed_byte(Register dst, Address src) {
3069 int off;
3070 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3071 off = offset();
3072 movsbl(dst, src); // movsxb
3073 } else {
3074 off = load_unsigned_byte(dst, src);
3075 shll(dst, 24);
3076 sarl(dst, 24);
3077 }
3078 return off;
3079 }
3080
3081 // Note: load_signed_short used to be called load_signed_word.
3082 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler
3083 // manual, which means 16 bits, that usage is found nowhere in HotSpot code.
3084 // The term "word" in HotSpot means a 32- or 64-bit machine word.
3085 int MacroAssembler::load_signed_short(Register dst, Address src) {
3086 int off;
3087 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3088 // This is dubious to me since it seems safe to do a signed 16 => 64 bit
3089 // version but this is what 64bit has always done. This seems to imply
3090 // that users are only using 32bits worth.
3091 off = offset();
3092 movswl(dst, src); // movsxw
3093 } else {
3094 off = load_unsigned_short(dst, src);
3095 shll(dst, 16);
3096 sarl(dst, 16);
3097 }
3098 return off;
3099 }
3100
3101 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
3102 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3103 // and "3.9 Partial Register Penalties", p. 22).
3104 int off;
3105 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) {
3106 off = offset();
3107 movzbl(dst, src); // movzxb
3108 } else {
3109 xorl(dst, dst);
3110 off = offset();
3111 movb(dst, src);
3112 }
3113 return off;
3114 }
3115
3116 // Note: load_unsigned_short used to be called load_unsigned_word.
3117 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
3118 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3119 // and "3.9 Partial Register Penalties", p. 22).
3120 int off;
3121 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) {
3122 off = offset();
3123 movzwl(dst, src); // movzxw
3124 } else {
3125 xorl(dst, dst);
3126 off = offset();
3127 movw(dst, src);
3128 }
3129 return off;
3130 }
3131
3132 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
3133 switch (size_in_bytes) {
3134 #ifndef _LP64
3135 case 8:
3136 assert(dst2 != noreg, "second dest register required");
3137 movl(dst, src);
3138 movl(dst2, src.plus_disp(BytesPerInt));
3139 break;
3140 #else
3141 case 8: movq(dst, src); break;
3142 #endif
3143 case 4: movl(dst, src); break;
3144 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
3145 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
3146 default: ShouldNotReachHere();
3147 }
3148 }
3149
3150 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
3151 switch (size_in_bytes) {
3152 #ifndef _LP64
3153 case 8:
3154 assert(src2 != noreg, "second source register required");
3155 movl(dst, src);
3156 movl(dst.plus_disp(BytesPerInt), src2);
3157 break;
3158 #else
3159 case 8: movq(dst, src); break;
3160 #endif
3161 case 4: movl(dst, src); break;
3162 case 2: movw(dst, src); break;
3163 case 1: movb(dst, src); break;
3164 default: ShouldNotReachHere();
3165 }
3166 }
3167
3168 void MacroAssembler::mov32(AddressLiteral dst, Register src) {
3169 if (reachable(dst)) {
3170 movl(as_Address(dst), src);
3171 } else {
3172 lea(rscratch1, dst);
3173 movl(Address(rscratch1, 0), src);
3174 }
3175 }
3176
3177 void MacroAssembler::mov32(Register dst, AddressLiteral src) {
3178 if (reachable(src)) {
3179 movl(dst, as_Address(src));
3180 } else {
3181 lea(rscratch1, src);
3182 movl(dst, Address(rscratch1, 0));
3183 }
3184 }
3185
3186 // C++ bool manipulation
3187
3188 void MacroAssembler::movbool(Register dst, Address src) {
3189 if(sizeof(bool) == 1)
3190 movb(dst, src);
3191 else if(sizeof(bool) == 2)
3192 movw(dst, src);
3193 else if(sizeof(bool) == 4)
3194 movl(dst, src);
3195 else
3196 // unsupported
3197 ShouldNotReachHere();
3198 }
3199
3200 void MacroAssembler::movbool(Address dst, bool boolconst) {
3201 if(sizeof(bool) == 1)
3202 movb(dst, (int) boolconst);
3203 else if(sizeof(bool) == 2)
3204 movw(dst, (int) boolconst);
3205 else if(sizeof(bool) == 4)
3206 movl(dst, (int) boolconst);
3207 else
3208 // unsupported
3209 ShouldNotReachHere();
3210 }
3211
3212 void MacroAssembler::movbool(Address dst, Register src) {
3213 if(sizeof(bool) == 1)
3214 movb(dst, src);
3215 else if(sizeof(bool) == 2)
3216 movw(dst, src);
3217 else if(sizeof(bool) == 4)
3218 movl(dst, src);
3219 else
3220 // unsupported
3221 ShouldNotReachHere();
3222 }
3223
3224 void MacroAssembler::movbyte(ArrayAddress dst, int src) {
3225 movb(as_Address(dst), src);
3226 }
3227
3228 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) {
3229 if (reachable(src)) {
3230 movdl(dst, as_Address(src));
3231 } else {
3232 lea(rscratch1, src);
3233 movdl(dst, Address(rscratch1, 0));
3234 }
3235 }
3236
3237 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) {
3238 if (reachable(src)) {
3239 movq(dst, as_Address(src));
3240 } else {
3241 lea(rscratch1, src);
3242 movq(dst, Address(rscratch1, 0));
3243 }
3244 }
3245
3246 #ifdef COMPILER2
3247 void MacroAssembler::setvectmask(Register dst, Register src) {
3248 guarantee(PostLoopMultiversioning, "must be");
3249 Assembler::movl(dst, 1);
3250 Assembler::shlxl(dst, dst, src);
3251 Assembler::decl(dst);
3252 Assembler::kmovdl(k1, dst);
3253 Assembler::movl(dst, src);
3254 }
3255
3256 void MacroAssembler::restorevectmask() {
3257 guarantee(PostLoopMultiversioning, "must be");
3258 Assembler::knotwl(k1, k0);
3259 }
3260 #endif // COMPILER2
3261
3262 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) {
3263 if (reachable(src)) {
3264 if (UseXmmLoadAndClearUpper) {
3265 movsd (dst, as_Address(src));
3266 } else {
3267 movlpd(dst, as_Address(src));
3268 }
3269 } else {
3270 lea(rscratch1, src);
3271 if (UseXmmLoadAndClearUpper) {
3272 movsd (dst, Address(rscratch1, 0));
3273 } else {
3274 movlpd(dst, Address(rscratch1, 0));
3275 }
3276 }
3277 }
3278
3279 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) {
3280 if (reachable(src)) {
3281 movss(dst, as_Address(src));
3282 } else {
3283 lea(rscratch1, src);
3284 movss(dst, Address(rscratch1, 0));
3285 }
3286 }
3287
3288 void MacroAssembler::movptr(Register dst, Register src) {
3289 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3290 }
3291
3292 void MacroAssembler::movptr(Register dst, Address src) {
3293 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3294 }
3295
3296 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
3297 void MacroAssembler::movptr(Register dst, intptr_t src) {
3298 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src));
3299 }
3300
3301 void MacroAssembler::movptr(Address dst, Register src) {
3302 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3303 }
3304
3305 void MacroAssembler::movdqu(Address dst, XMMRegister src) {
3306 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3307 Assembler::movdqu(dst, src);
3308 }
3309
3310 void MacroAssembler::movdqu(XMMRegister dst, Address src) {
3311 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3312 Assembler::movdqu(dst, src);
3313 }
3314
3315 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) {
3316 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3317 Assembler::movdqu(dst, src);
3318 }
3319
3320 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src, Register scratchReg) {
3321 if (reachable(src)) {
3322 movdqu(dst, as_Address(src));
3323 } else {
3324 lea(scratchReg, src);
3325 movdqu(dst, Address(scratchReg, 0));
3326 }
3327 }
3328
3329 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) {
3330 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3331 Assembler::vmovdqu(dst, src);
3332 }
3333
3334 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) {
3335 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3336 Assembler::vmovdqu(dst, src);
3337 }
3338
3339 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) {
3340 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3341 Assembler::vmovdqu(dst, src);
3342 }
3343
3344 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
3345 if (reachable(src)) {
3346 vmovdqu(dst, as_Address(src));
3347 }
3348 else {
3349 lea(scratch_reg, src);
3350 vmovdqu(dst, Address(scratch_reg, 0));
3351 }
3352 }
3353
3354 void MacroAssembler::evmovdquq(XMMRegister dst, AddressLiteral src, int vector_len, Register rscratch) {
3355 if (reachable(src)) {
3356 Assembler::evmovdquq(dst, as_Address(src), vector_len);
3357 } else {
3358 lea(rscratch, src);
3359 Assembler::evmovdquq(dst, Address(rscratch, 0), vector_len);
3360 }
3361 }
3362
3363 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) {
3364 if (reachable(src)) {
3365 Assembler::movdqa(dst, as_Address(src));
3366 } else {
3367 lea(rscratch1, src);
3368 Assembler::movdqa(dst, Address(rscratch1, 0));
3369 }
3370 }
3371
3372 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) {
3373 if (reachable(src)) {
3374 Assembler::movsd(dst, as_Address(src));
3375 } else {
3376 lea(rscratch1, src);
3377 Assembler::movsd(dst, Address(rscratch1, 0));
3378 }
3379 }
3380
3381 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) {
3382 if (reachable(src)) {
3383 Assembler::movss(dst, as_Address(src));
3384 } else {
3385 lea(rscratch1, src);
3386 Assembler::movss(dst, Address(rscratch1, 0));
3387 }
3388 }
3389
3390 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) {
3391 if (reachable(src)) {
3392 Assembler::mulsd(dst, as_Address(src));
3393 } else {
3394 lea(rscratch1, src);
3395 Assembler::mulsd(dst, Address(rscratch1, 0));
3396 }
3397 }
3398
3399 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) {
3400 if (reachable(src)) {
3401 Assembler::mulss(dst, as_Address(src));
3402 } else {
3403 lea(rscratch1, src);
3404 Assembler::mulss(dst, Address(rscratch1, 0));
3405 }
3406 }
3407
3408 void MacroAssembler::null_check(Register reg, int offset) {
3409 if (needs_explicit_null_check(offset)) {
3410 // provoke OS NULL exception if reg = NULL by
3411 // accessing M[reg] w/o changing any (non-CC) registers
3412 // NOTE: cmpl is plenty here to provoke a segv
3413 cmpptr(rax, Address(reg, 0));
3414 // Note: should probably use testl(rax, Address(reg, 0));
3415 // may be shorter code (however, this version of
3416 // testl needs to be implemented first)
3417 } else {
3418 // nothing to do, (later) access of M[reg + offset]
3419 // will provoke OS NULL exception if reg = NULL
3420 }
3421 }
3422
3423 void MacroAssembler::os_breakpoint() {
3424 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability
3425 // (e.g., MSVC can't call ps() otherwise)
3426 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint)));
3427 }
3428
3429 void MacroAssembler::unimplemented(const char* what) {
3430 const char* buf = NULL;
3431 {
3432 ResourceMark rm;
3433 stringStream ss;
3434 ss.print("unimplemented: %s", what);
3435 buf = code_string(ss.as_string());
3436 }
3437 stop(buf);
3438 }
3439
3440 #ifdef _LP64
3441 #define XSTATE_BV 0x200
3442 #endif
3443
3444 void MacroAssembler::pop_CPU_state() {
3445 pop_FPU_state();
3446 pop_IU_state();
3447 }
3448
3449 void MacroAssembler::pop_FPU_state() {
3450 #ifndef _LP64
3451 frstor(Address(rsp, 0));
3452 #else
3453 fxrstor(Address(rsp, 0));
3454 #endif
3455 addptr(rsp, FPUStateSizeInWords * wordSize);
3456 }
3457
3458 void MacroAssembler::pop_IU_state() {
3459 popa();
3460 LP64_ONLY(addq(rsp, 8));
3461 popf();
3462 }
3463
3464 // Save Integer and Float state
3465 // Warning: Stack must be 16 byte aligned (64bit)
3466 void MacroAssembler::push_CPU_state() {
3467 push_IU_state();
3468 push_FPU_state();
3469 }
3470
3471 void MacroAssembler::push_FPU_state() {
3472 subptr(rsp, FPUStateSizeInWords * wordSize);
3473 #ifndef _LP64
3474 fnsave(Address(rsp, 0));
3475 fwait();
3476 #else
3477 fxsave(Address(rsp, 0));
3478 #endif // LP64
3479 }
3480
3481 void MacroAssembler::push_IU_state() {
3482 // Push flags first because pusha kills them
3483 pushf();
3484 // Make sure rsp stays 16-byte aligned
3485 LP64_ONLY(subq(rsp, 8));
3486 pusha();
3487 }
3488
3489 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp) { // determine java_thread register
3490 if (!java_thread->is_valid()) {
3491 java_thread = rdi;
3492 get_thread(java_thread);
3493 }
3494 // we must set sp to zero to clear frame
3495 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
3496 if (clear_fp) {
3497 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3498 }
3499
3500 // Always clear the pc because it could have been set by make_walkable()
3501 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
3502
3503 vzeroupper();
3504 }
3505
3506 void MacroAssembler::restore_rax(Register tmp) {
3507 if (tmp == noreg) pop(rax);
3508 else if (tmp != rax) mov(rax, tmp);
3509 }
3510
3511 void MacroAssembler::round_to(Register reg, int modulus) {
3512 addptr(reg, modulus - 1);
3513 andptr(reg, -modulus);
3514 }
3515
3516 void MacroAssembler::save_rax(Register tmp) {
3517 if (tmp == noreg) push(rax);
3518 else if (tmp != rax) mov(tmp, rax);
3519 }
3520
3521 void MacroAssembler::safepoint_poll(Label& slow_path, Register thread_reg, Register temp_reg) {
3522 if (SafepointMechanism::uses_thread_local_poll()) {
3523 #ifdef _LP64
3524 assert(thread_reg == r15_thread, "should be");
3525 #else
3526 if (thread_reg == noreg) {
3527 thread_reg = temp_reg;
3528 get_thread(thread_reg);
3529 }
3530 #endif
3531 testb(Address(thread_reg, Thread::polling_page_offset()), SafepointMechanism::poll_bit());
3532 jcc(Assembler::notZero, slow_path); // handshake bit set implies poll
3533 } else {
3534 cmp32(ExternalAddress(SafepointSynchronize::address_of_state()),
3535 SafepointSynchronize::_not_synchronized);
3536 jcc(Assembler::notEqual, slow_path);
3537 }
3538 }
3539
3540 // Calls to C land
3541 //
3542 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded
3543 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
3544 // has to be reset to 0. This is required to allow proper stack traversal.
3545 void MacroAssembler::set_last_Java_frame(Register java_thread,
3546 Register last_java_sp,
3547 Register last_java_fp,
3548 address last_java_pc) {
3549 vzeroupper();
3550 // determine java_thread register
3551 if (!java_thread->is_valid()) {
3552 java_thread = rdi;
3553 get_thread(java_thread);
3554 }
3555 // determine last_java_sp register
3556 if (!last_java_sp->is_valid()) {
3557 last_java_sp = rsp;
3558 }
3559
3560 // last_java_fp is optional
3561
3562 if (last_java_fp->is_valid()) {
3563 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp);
3564 }
3565
3566 // last_java_pc is optional
3567
3568 if (last_java_pc != NULL) {
3569 lea(Address(java_thread,
3570 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()),
3571 InternalAddress(last_java_pc));
3572
3573 }
3574 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
3575 }
3576
3577 void MacroAssembler::shlptr(Register dst, int imm8) {
3578 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8));
3579 }
3580
3581 void MacroAssembler::shrptr(Register dst, int imm8) {
3582 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8));
3583 }
3584
3585 void MacroAssembler::sign_extend_byte(Register reg) {
3586 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) {
3587 movsbl(reg, reg); // movsxb
3588 } else {
3589 shll(reg, 24);
3590 sarl(reg, 24);
3591 }
3592 }
3593
3594 void MacroAssembler::sign_extend_short(Register reg) {
3595 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3596 movswl(reg, reg); // movsxw
3597 } else {
3598 shll(reg, 16);
3599 sarl(reg, 16);
3600 }
3601 }
3602
3603 void MacroAssembler::testl(Register dst, AddressLiteral src) {
3604 assert(reachable(src), "Address should be reachable");
3605 testl(dst, as_Address(src));
3606 }
3607
3608 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) {
3609 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3610 Assembler::pcmpeqb(dst, src);
3611 }
3612
3613 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) {
3614 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3615 Assembler::pcmpeqw(dst, src);
3616 }
3617
3618 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) {
3619 assert((dst->encoding() < 16),"XMM register should be 0-15");
3620 Assembler::pcmpestri(dst, src, imm8);
3621 }
3622
3623 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) {
3624 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3625 Assembler::pcmpestri(dst, src, imm8);
3626 }
3627
3628 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) {
3629 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3630 Assembler::pmovzxbw(dst, src);
3631 }
3632
3633 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) {
3634 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3635 Assembler::pmovzxbw(dst, src);
3636 }
3637
3638 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) {
3639 assert((src->encoding() < 16),"XMM register should be 0-15");
3640 Assembler::pmovmskb(dst, src);
3641 }
3642
3643 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) {
3644 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3645 Assembler::ptest(dst, src);
3646 }
3647
3648 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) {
3649 if (reachable(src)) {
3650 Assembler::sqrtsd(dst, as_Address(src));
3651 } else {
3652 lea(rscratch1, src);
3653 Assembler::sqrtsd(dst, Address(rscratch1, 0));
3654 }
3655 }
3656
3657 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) {
3658 if (reachable(src)) {
3659 Assembler::sqrtss(dst, as_Address(src));
3660 } else {
3661 lea(rscratch1, src);
3662 Assembler::sqrtss(dst, Address(rscratch1, 0));
3663 }
3664 }
3665
3666 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) {
3667 if (reachable(src)) {
3668 Assembler::subsd(dst, as_Address(src));
3669 } else {
3670 lea(rscratch1, src);
3671 Assembler::subsd(dst, Address(rscratch1, 0));
3672 }
3673 }
3674
3675 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
3676 if (reachable(src)) {
3677 Assembler::subss(dst, as_Address(src));
3678 } else {
3679 lea(rscratch1, src);
3680 Assembler::subss(dst, Address(rscratch1, 0));
3681 }
3682 }
3683
3684 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
3685 if (reachable(src)) {
3686 Assembler::ucomisd(dst, as_Address(src));
3687 } else {
3688 lea(rscratch1, src);
3689 Assembler::ucomisd(dst, Address(rscratch1, 0));
3690 }
3691 }
3692
3693 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
3694 if (reachable(src)) {
3695 Assembler::ucomiss(dst, as_Address(src));
3696 } else {
3697 lea(rscratch1, src);
3698 Assembler::ucomiss(dst, Address(rscratch1, 0));
3699 }
3700 }
3701
3702 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
3703 // Used in sign-bit flipping with aligned address.
3704 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3705 if (reachable(src)) {
3706 Assembler::xorpd(dst, as_Address(src));
3707 } else {
3708 lea(scratch_reg, src);
3709 Assembler::xorpd(dst, Address(scratch_reg, 0));
3710 }
3711 }
3712
3713 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) {
3714 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
3715 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
3716 }
3717 else {
3718 Assembler::xorpd(dst, src);
3719 }
3720 }
3721
3722 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) {
3723 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
3724 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
3725 } else {
3726 Assembler::xorps(dst, src);
3727 }
3728 }
3729
3730 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src, Register scratch_reg) {
3731 // Used in sign-bit flipping with aligned address.
3732 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3733 if (reachable(src)) {
3734 Assembler::xorps(dst, as_Address(src));
3735 } else {
3736 lea(scratch_reg, src);
3737 Assembler::xorps(dst, Address(scratch_reg, 0));
3738 }
3739 }
3740
3741 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
3742 // Used in sign-bit flipping with aligned address.
3743 bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
3744 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
3745 if (reachable(src)) {
3746 Assembler::pshufb(dst, as_Address(src));
3747 } else {
3748 lea(rscratch1, src);
3749 Assembler::pshufb(dst, Address(rscratch1, 0));
3750 }
3751 }
3752
3753 // AVX 3-operands instructions
3754
3755 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3756 if (reachable(src)) {
3757 vaddsd(dst, nds, as_Address(src));
3758 } else {
3759 lea(rscratch1, src);
3760 vaddsd(dst, nds, Address(rscratch1, 0));
3761 }
3762 }
3763
3764 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3765 if (reachable(src)) {
3766 vaddss(dst, nds, as_Address(src));
3767 } else {
3768 lea(rscratch1, src);
3769 vaddss(dst, nds, Address(rscratch1, 0));
3770 }
3771 }
3772
3773 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
3774 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
3775 vandps(dst, nds, negate_field, vector_len);
3776 }
3777
3778 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
3779 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
3780 vandpd(dst, nds, negate_field, vector_len);
3781 }
3782
3783 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3784 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3785 Assembler::vpaddb(dst, nds, src, vector_len);
3786 }
3787
3788 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3789 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3790 Assembler::vpaddb(dst, nds, src, vector_len);
3791 }
3792
3793 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3794 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3795 Assembler::vpaddw(dst, nds, src, vector_len);
3796 }
3797
3798 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3799 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3800 Assembler::vpaddw(dst, nds, src, vector_len);
3801 }
3802
3803 void MacroAssembler::vpand(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
3804 if (reachable(src)) {
3805 Assembler::vpand(dst, nds, as_Address(src), vector_len);
3806 } else {
3807 lea(scratch_reg, src);
3808 Assembler::vpand(dst, nds, Address(scratch_reg, 0), vector_len);
3809 }
3810 }
3811
3812 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src, int vector_len) {
3813 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3814 Assembler::vpbroadcastw(dst, src, vector_len);
3815 }
3816
3817 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3818 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3819 Assembler::vpcmpeqb(dst, nds, src, vector_len);
3820 }
3821
3822 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3823 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3824 Assembler::vpcmpeqw(dst, nds, src, vector_len);
3825 }
3826
3827 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) {
3828 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3829 Assembler::vpmovzxbw(dst, src, vector_len);
3830 }
3831
3832 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) {
3833 assert((src->encoding() < 16),"XMM register should be 0-15");
3834 Assembler::vpmovmskb(dst, src);
3835 }
3836
3837 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3838 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3839 Assembler::vpmullw(dst, nds, src, vector_len);
3840 }
3841
3842 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3843 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3844 Assembler::vpmullw(dst, nds, src, vector_len);
3845 }
3846
3847 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3848 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3849 Assembler::vpsubb(dst, nds, src, vector_len);
3850 }
3851
3852 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3853 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3854 Assembler::vpsubb(dst, nds, src, vector_len);
3855 }
3856
3857 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3858 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3859 Assembler::vpsubw(dst, nds, src, vector_len);
3860 }
3861
3862 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3863 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3864 Assembler::vpsubw(dst, nds, src, vector_len);
3865 }
3866
3867 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3868 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3869 Assembler::vpsraw(dst, nds, shift, vector_len);
3870 }
3871
3872 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3873 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3874 Assembler::vpsraw(dst, nds, shift, vector_len);
3875 }
3876
3877 void MacroAssembler::evpsraq(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3878 assert(UseAVX > 2,"");
3879 if (!VM_Version::supports_avx512vl() && vector_len < 2) {
3880 vector_len = 2;
3881 }
3882 Assembler::evpsraq(dst, nds, shift, vector_len);
3883 }
3884
3885 void MacroAssembler::evpsraq(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3886 assert(UseAVX > 2,"");
3887 if (!VM_Version::supports_avx512vl() && vector_len < 2) {
3888 vector_len = 2;
3889 }
3890 Assembler::evpsraq(dst, nds, shift, vector_len);
3891 }
3892
3893 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3894 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3895 Assembler::vpsrlw(dst, nds, shift, vector_len);
3896 }
3897
3898 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3899 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3900 Assembler::vpsrlw(dst, nds, shift, vector_len);
3901 }
3902
3903 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3904 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3905 Assembler::vpsllw(dst, nds, shift, vector_len);
3906 }
3907
3908 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3909 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3910 Assembler::vpsllw(dst, nds, shift, vector_len);
3911 }
3912
3913 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) {
3914 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3915 Assembler::vptest(dst, src);
3916 }
3917
3918 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) {
3919 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3920 Assembler::punpcklbw(dst, src);
3921 }
3922
3923 void MacroAssembler::pshufd(XMMRegister dst, Address src, int mode) {
3924 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3925 Assembler::pshufd(dst, src, mode);
3926 }
3927
3928 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) {
3929 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3930 Assembler::pshuflw(dst, src, mode);
3931 }
3932
3933 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
3934 if (reachable(src)) {
3935 vandpd(dst, nds, as_Address(src), vector_len);
3936 } else {
3937 lea(scratch_reg, src);
3938 vandpd(dst, nds, Address(scratch_reg, 0), vector_len);
3939 }
3940 }
3941
3942 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
3943 if (reachable(src)) {
3944 vandps(dst, nds, as_Address(src), vector_len);
3945 } else {
3946 lea(scratch_reg, src);
3947 vandps(dst, nds, Address(scratch_reg, 0), vector_len);
3948 }
3949 }
3950
3951 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3952 if (reachable(src)) {
3953 vdivsd(dst, nds, as_Address(src));
3954 } else {
3955 lea(rscratch1, src);
3956 vdivsd(dst, nds, Address(rscratch1, 0));
3957 }
3958 }
3959
3960 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3961 if (reachable(src)) {
3962 vdivss(dst, nds, as_Address(src));
3963 } else {
3964 lea(rscratch1, src);
3965 vdivss(dst, nds, Address(rscratch1, 0));
3966 }
3967 }
3968
3969 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3970 if (reachable(src)) {
3971 vmulsd(dst, nds, as_Address(src));
3972 } else {
3973 lea(rscratch1, src);
3974 vmulsd(dst, nds, Address(rscratch1, 0));
3975 }
3976 }
3977
3978 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3979 if (reachable(src)) {
3980 vmulss(dst, nds, as_Address(src));
3981 } else {
3982 lea(rscratch1, src);
3983 vmulss(dst, nds, Address(rscratch1, 0));
3984 }
3985 }
3986
3987 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3988 if (reachable(src)) {
3989 vsubsd(dst, nds, as_Address(src));
3990 } else {
3991 lea(rscratch1, src);
3992 vsubsd(dst, nds, Address(rscratch1, 0));
3993 }
3994 }
3995
3996 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3997 if (reachable(src)) {
3998 vsubss(dst, nds, as_Address(src));
3999 } else {
4000 lea(rscratch1, src);
4001 vsubss(dst, nds, Address(rscratch1, 0));
4002 }
4003 }
4004
4005 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4006 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
4007 vxorps(dst, nds, src, Assembler::AVX_128bit);
4008 }
4009
4010 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4011 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
4012 vxorpd(dst, nds, src, Assembler::AVX_128bit);
4013 }
4014
4015 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
4016 if (reachable(src)) {
4017 vxorpd(dst, nds, as_Address(src), vector_len);
4018 } else {
4019 lea(scratch_reg, src);
4020 vxorpd(dst, nds, Address(scratch_reg, 0), vector_len);
4021 }
4022 }
4023
4024 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
4025 if (reachable(src)) {
4026 vxorps(dst, nds, as_Address(src), vector_len);
4027 } else {
4028 lea(scratch_reg, src);
4029 vxorps(dst, nds, Address(scratch_reg, 0), vector_len);
4030 }
4031 }
4032
4033 void MacroAssembler::vpxor(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len, Register scratch_reg) {
4034 if (UseAVX > 1 || (vector_len < 1)) {
4035 if (reachable(src)) {
4036 Assembler::vpxor(dst, nds, as_Address(src), vector_len);
4037 } else {
4038 lea(scratch_reg, src);
4039 Assembler::vpxor(dst, nds, Address(scratch_reg, 0), vector_len);
4040 }
4041 }
4042 else {
4043 MacroAssembler::vxorpd(dst, nds, src, vector_len, scratch_reg);
4044 }
4045 }
4046
4047 //-------------------------------------------------------------------------------------------
4048 #ifdef COMPILER2
4049 // Generic instructions support for use in .ad files C2 code generation
4050
4051 void MacroAssembler::vabsnegd(int opcode, XMMRegister dst, Register scr) {
4052 if (opcode == Op_AbsVD) {
4053 andpd(dst, ExternalAddress(StubRoutines::x86::vector_double_sign_mask()), scr);
4054 } else {
4055 assert((opcode == Op_NegVD),"opcode should be Op_NegD");
4056 xorpd(dst, ExternalAddress(StubRoutines::x86::vector_double_sign_flip()), scr);
4057 }
4058 }
4059
4060 void MacroAssembler::vabsnegd(int opcode, XMMRegister dst, XMMRegister src, int vector_len, Register scr) {
4061 if (opcode == Op_AbsVD) {
4062 vandpd(dst, src, ExternalAddress(StubRoutines::x86::vector_double_sign_mask()), vector_len, scr);
4063 } else {
4064 assert((opcode == Op_NegVD),"opcode should be Op_NegD");
4065 vxorpd(dst, src, ExternalAddress(StubRoutines::x86::vector_double_sign_flip()), vector_len, scr);
4066 }
4067 }
4068
4069 void MacroAssembler::vabsnegf(int opcode, XMMRegister dst, Register scr) {
4070 if (opcode == Op_AbsVF) {
4071 andps(dst, ExternalAddress(StubRoutines::x86::vector_float_sign_mask()), scr);
4072 } else {
4073 assert((opcode == Op_NegVF),"opcode should be Op_NegF");
4074 xorps(dst, ExternalAddress(StubRoutines::x86::vector_float_sign_flip()), scr);
4075 }
4076 }
4077
4078 void MacroAssembler::vabsnegf(int opcode, XMMRegister dst, XMMRegister src, int vector_len, Register scr) {
4079 if (opcode == Op_AbsVF) {
4080 vandps(dst, src, ExternalAddress(StubRoutines::x86::vector_float_sign_mask()), vector_len, scr);
4081 } else {
4082 assert((opcode == Op_NegVF),"opcode should be Op_NegF");
4083 vxorps(dst, src, ExternalAddress(StubRoutines::x86::vector_float_sign_flip()), vector_len, scr);
4084 }
4085 }
4086
4087 void MacroAssembler::vextendbw(bool sign, XMMRegister dst, XMMRegister src) {
4088 if (sign) {
4089 pmovsxbw(dst, src);
4090 } else {
4091 pmovzxbw(dst, src);
4092 }
4093 }
4094
4095 void MacroAssembler::vextendbw(bool sign, XMMRegister dst, XMMRegister src, int vector_len) {
4096 if (sign) {
4097 vpmovsxbw(dst, src, vector_len);
4098 } else {
4099 vpmovzxbw(dst, src, vector_len);
4100 }
4101 }
4102
4103 void MacroAssembler::vshiftd(int opcode, XMMRegister dst, XMMRegister src) {
4104 if (opcode == Op_RShiftVI) {
4105 psrad(dst, src);
4106 } else if (opcode == Op_LShiftVI) {
4107 pslld(dst, src);
4108 } else {
4109 assert((opcode == Op_URShiftVI),"opcode should be Op_URShiftVI");
4110 psrld(dst, src);
4111 }
4112 }
4113
4114 void MacroAssembler::vshiftd(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4115 if (opcode == Op_RShiftVI) {
4116 vpsrad(dst, nds, src, vector_len);
4117 } else if (opcode == Op_LShiftVI) {
4118 vpslld(dst, nds, src, vector_len);
4119 } else {
4120 assert((opcode == Op_URShiftVI),"opcode should be Op_URShiftVI");
4121 vpsrld(dst, nds, src, vector_len);
4122 }
4123 }
4124
4125 void MacroAssembler::vshiftw(int opcode, XMMRegister dst, XMMRegister src) {
4126 if ((opcode == Op_RShiftVS) || (opcode == Op_RShiftVB)) {
4127 psraw(dst, src);
4128 } else if ((opcode == Op_LShiftVS) || (opcode == Op_LShiftVB)) {
4129 psllw(dst, src);
4130 } else {
4131 assert(((opcode == Op_URShiftVS) || (opcode == Op_URShiftVB)),"opcode should be one of Op_URShiftVS or Op_URShiftVB");
4132 psrlw(dst, src);
4133 }
4134 }
4135
4136 void MacroAssembler::vshiftw(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4137 if ((opcode == Op_RShiftVS) || (opcode == Op_RShiftVB)) {
4138 vpsraw(dst, nds, src, vector_len);
4139 } else if ((opcode == Op_LShiftVS) || (opcode == Op_LShiftVB)) {
4140 vpsllw(dst, nds, src, vector_len);
4141 } else {
4142 assert(((opcode == Op_URShiftVS) || (opcode == Op_URShiftVB)),"opcode should be one of Op_URShiftVS or Op_URShiftVB");
4143 vpsrlw(dst, nds, src, vector_len);
4144 }
4145 }
4146
4147 void MacroAssembler::vshiftq(int opcode, XMMRegister dst, XMMRegister src) {
4148 if (opcode == Op_RShiftVL) {
4149 psrlq(dst, src); // using srl to implement sra on pre-avs512 systems
4150 } else if (opcode == Op_LShiftVL) {
4151 psllq(dst, src);
4152 } else {
4153 assert((opcode == Op_URShiftVL),"opcode should be Op_URShiftVL");
4154 psrlq(dst, src);
4155 }
4156 }
4157
4158 void MacroAssembler::vshiftq(int opcode, XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4159 if (opcode == Op_RShiftVL) {
4160 evpsraq(dst, nds, src, vector_len);
4161 } else if (opcode == Op_LShiftVL) {
4162 vpsllq(dst, nds, src, vector_len);
4163 } else {
4164 assert((opcode == Op_URShiftVL),"opcode should be Op_URShiftVL");
4165 vpsrlq(dst, nds, src, vector_len);
4166 }
4167 }
4168 #endif
4169 //-------------------------------------------------------------------------------------------
4170
4171 void MacroAssembler::clear_jweak_tag(Register possibly_jweak) {
4172 const int32_t inverted_jweak_mask = ~static_cast<int32_t>(JNIHandles::weak_tag_mask);
4173 STATIC_ASSERT(inverted_jweak_mask == -2); // otherwise check this code
4174 // The inverted mask is sign-extended
4175 andptr(possibly_jweak, inverted_jweak_mask);
4176 }
4177
4178 void MacroAssembler::resolve_jobject(Register value,
4179 Register thread,
4180 Register tmp) {
4181 assert_different_registers(value, thread, tmp);
4182 Label done, not_weak;
4183 testptr(value, value);
4184 jcc(Assembler::zero, done); // Use NULL as-is.
4185 testptr(value, JNIHandles::weak_tag_mask); // Test for jweak tag.
4186 jcc(Assembler::zero, not_weak);
4187 // Resolve jweak.
4188 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
4189 value, Address(value, -JNIHandles::weak_tag_value), tmp, thread);
4190 verify_oop(value);
4191 jmp(done);
4192 bind(not_weak);
4193 // Resolve (untagged) jobject.
4194 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, 0), tmp, thread);
4195 verify_oop(value);
4196 bind(done);
4197 }
4198
4199 void MacroAssembler::subptr(Register dst, int32_t imm32) {
4200 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32));
4201 }
4202
4203 // Force generation of a 4 byte immediate value even if it fits into 8bit
4204 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) {
4205 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32));
4206 }
4207
4208 void MacroAssembler::subptr(Register dst, Register src) {
4209 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src));
4210 }
4211
4212 // C++ bool manipulation
4213 void MacroAssembler::testbool(Register dst) {
4214 if(sizeof(bool) == 1)
4215 testb(dst, 0xff);
4216 else if(sizeof(bool) == 2) {
4217 // testw implementation needed for two byte bools
4218 ShouldNotReachHere();
4219 } else if(sizeof(bool) == 4)
4220 testl(dst, dst);
4221 else
4222 // unsupported
4223 ShouldNotReachHere();
4224 }
4225
4226 void MacroAssembler::testptr(Register dst, Register src) {
4227 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src));
4228 }
4229
4230 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
4231 void MacroAssembler::tlab_allocate(Register thread, Register obj,
4232 Register var_size_in_bytes,
4233 int con_size_in_bytes,
4234 Register t1,
4235 Register t2,
4236 Label& slow_case) {
4237 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4238 bs->tlab_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
4239 }
4240
4241 // Defines obj, preserves var_size_in_bytes
4242 void MacroAssembler::eden_allocate(Register thread, Register obj,
4243 Register var_size_in_bytes,
4244 int con_size_in_bytes,
4245 Register t1,
4246 Label& slow_case) {
4247 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4248 bs->eden_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case);
4249 }
4250
4251 // Preserves the contents of address, destroys the contents length_in_bytes and temp.
4252 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) {
4253 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different");
4254 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord");
4255 Label done;
4256
4257 testptr(length_in_bytes, length_in_bytes);
4258 jcc(Assembler::zero, done);
4259
4260 // initialize topmost word, divide index by 2, check if odd and test if zero
4261 // note: for the remaining code to work, index must be a multiple of BytesPerWord
4262 #ifdef ASSERT
4263 {
4264 Label L;
4265 testptr(length_in_bytes, BytesPerWord - 1);
4266 jcc(Assembler::zero, L);
4267 stop("length must be a multiple of BytesPerWord");
4268 bind(L);
4269 }
4270 #endif
4271 Register index = length_in_bytes;
4272 xorptr(temp, temp); // use _zero reg to clear memory (shorter code)
4273 if (UseIncDec) {
4274 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set
4275 } else {
4276 shrptr(index, 2); // use 2 instructions to avoid partial flag stall
4277 shrptr(index, 1);
4278 }
4279 #ifndef _LP64
4280 // index could have not been a multiple of 8 (i.e., bit 2 was set)
4281 {
4282 Label even;
4283 // note: if index was a multiple of 8, then it cannot
4284 // be 0 now otherwise it must have been 0 before
4285 // => if it is even, we don't need to check for 0 again
4286 jcc(Assembler::carryClear, even);
4287 // clear topmost word (no jump would be needed if conditional assignment worked here)
4288 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp);
4289 // index could be 0 now, must check again
4290 jcc(Assembler::zero, done);
4291 bind(even);
4292 }
4293 #endif // !_LP64
4294 // initialize remaining object fields: index is a multiple of 2 now
4295 {
4296 Label loop;
4297 bind(loop);
4298 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp);
4299 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);)
4300 decrement(index);
4301 jcc(Assembler::notZero, loop);
4302 }
4303
4304 bind(done);
4305 }
4306
4307 // Look up the method for a megamorphic invokeinterface call.
4308 // The target method is determined by <intf_klass, itable_index>.
4309 // The receiver klass is in recv_klass.
4310 // On success, the result will be in method_result, and execution falls through.
4311 // On failure, execution transfers to the given label.
4312 void MacroAssembler::lookup_interface_method(Register recv_klass,
4313 Register intf_klass,
4314 RegisterOrConstant itable_index,
4315 Register method_result,
4316 Register scan_temp,
4317 Label& L_no_such_interface,
4318 bool return_method) {
4319 assert_different_registers(recv_klass, intf_klass, scan_temp);
4320 assert_different_registers(method_result, intf_klass, scan_temp);
4321 assert(recv_klass != method_result || !return_method,
4322 "recv_klass can be destroyed when method isn't needed");
4323
4324 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
4325 "caller must use same register for non-constant itable index as for method");
4326
4327 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
4328 int vtable_base = in_bytes(Klass::vtable_start_offset());
4329 int itentry_off = itableMethodEntry::method_offset_in_bytes();
4330 int scan_step = itableOffsetEntry::size() * wordSize;
4331 int vte_size = vtableEntry::size_in_bytes();
4332 Address::ScaleFactor times_vte_scale = Address::times_ptr;
4333 assert(vte_size == wordSize, "else adjust times_vte_scale");
4334
4335 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
4336
4337 // %%% Could store the aligned, prescaled offset in the klassoop.
4338 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
4339
4340 if (return_method) {
4341 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
4342 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
4343 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
4344 }
4345
4346 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
4347 // if (scan->interface() == intf) {
4348 // result = (klass + scan->offset() + itable_index);
4349 // }
4350 // }
4351 Label search, found_method;
4352
4353 for (int peel = 1; peel >= 0; peel--) {
4354 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
4355 cmpptr(intf_klass, method_result);
4356
4357 if (peel) {
4358 jccb(Assembler::equal, found_method);
4359 } else {
4360 jccb(Assembler::notEqual, search);
4361 // (invert the test to fall through to found_method...)
4362 }
4363
4364 if (!peel) break;
4365
4366 bind(search);
4367
4368 // Check that the previous entry is non-null. A null entry means that
4369 // the receiver class doesn't implement the interface, and wasn't the
4370 // same as when the caller was compiled.
4371 testptr(method_result, method_result);
4372 jcc(Assembler::zero, L_no_such_interface);
4373 addptr(scan_temp, scan_step);
4374 }
4375
4376 bind(found_method);
4377
4378 if (return_method) {
4379 // Got a hit.
4380 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
4381 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
4382 }
4383 }
4384
4385
4386 // virtual method calling
4387 void MacroAssembler::lookup_virtual_method(Register recv_klass,
4388 RegisterOrConstant vtable_index,
4389 Register method_result) {
4390 const int base = in_bytes(Klass::vtable_start_offset());
4391 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below");
4392 Address vtable_entry_addr(recv_klass,
4393 vtable_index, Address::times_ptr,
4394 base + vtableEntry::method_offset_in_bytes());
4395 movptr(method_result, vtable_entry_addr);
4396 }
4397
4398
4399 void MacroAssembler::check_klass_subtype(Register sub_klass,
4400 Register super_klass,
4401 Register temp_reg,
4402 Label& L_success) {
4403 Label L_failure;
4404 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL);
4405 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
4406 bind(L_failure);
4407 }
4408
4409
4410 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
4411 Register super_klass,
4412 Register temp_reg,
4413 Label* L_success,
4414 Label* L_failure,
4415 Label* L_slow_path,
4416 RegisterOrConstant super_check_offset) {
4417 assert_different_registers(sub_klass, super_klass, temp_reg);
4418 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
4419 if (super_check_offset.is_register()) {
4420 assert_different_registers(sub_klass, super_klass,
4421 super_check_offset.as_register());
4422 } else if (must_load_sco) {
4423 assert(temp_reg != noreg, "supply either a temp or a register offset");
4424 }
4425
4426 Label L_fallthrough;
4427 int label_nulls = 0;
4428 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
4429 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
4430 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
4431 assert(label_nulls <= 1, "at most one NULL in the batch");
4432
4433 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
4434 int sco_offset = in_bytes(Klass::super_check_offset_offset());
4435 Address super_check_offset_addr(super_klass, sco_offset);
4436
4437 // Hacked jcc, which "knows" that L_fallthrough, at least, is in
4438 // range of a jccb. If this routine grows larger, reconsider at
4439 // least some of these.
4440 #define local_jcc(assembler_cond, label) \
4441 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \
4442 else jcc( assembler_cond, label) /*omit semi*/
4443
4444 // Hacked jmp, which may only be used just before L_fallthrough.
4445 #define final_jmp(label) \
4446 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
4447 else jmp(label) /*omit semi*/
4448
4449 // If the pointers are equal, we are done (e.g., String[] elements).
4450 // This self-check enables sharing of secondary supertype arrays among
4451 // non-primary types such as array-of-interface. Otherwise, each such
4452 // type would need its own customized SSA.
4453 // We move this check to the front of the fast path because many
4454 // type checks are in fact trivially successful in this manner,
4455 // so we get a nicely predicted branch right at the start of the check.
4456 cmpptr(sub_klass, super_klass);
4457 local_jcc(Assembler::equal, *L_success);
4458
4459 // Check the supertype display:
4460 if (must_load_sco) {
4461 // Positive movl does right thing on LP64.
4462 movl(temp_reg, super_check_offset_addr);
4463 super_check_offset = RegisterOrConstant(temp_reg);
4464 }
4465 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0);
4466 cmpptr(super_klass, super_check_addr); // load displayed supertype
4467
4468 // This check has worked decisively for primary supers.
4469 // Secondary supers are sought in the super_cache ('super_cache_addr').
4470 // (Secondary supers are interfaces and very deeply nested subtypes.)
4471 // This works in the same check above because of a tricky aliasing
4472 // between the super_cache and the primary super display elements.
4473 // (The 'super_check_addr' can address either, as the case requires.)
4474 // Note that the cache is updated below if it does not help us find
4475 // what we need immediately.
4476 // So if it was a primary super, we can just fail immediately.
4477 // Otherwise, it's the slow path for us (no success at this point).
4478
4479 if (super_check_offset.is_register()) {
4480 local_jcc(Assembler::equal, *L_success);
4481 cmpl(super_check_offset.as_register(), sc_offset);
4482 if (L_failure == &L_fallthrough) {
4483 local_jcc(Assembler::equal, *L_slow_path);
4484 } else {
4485 local_jcc(Assembler::notEqual, *L_failure);
4486 final_jmp(*L_slow_path);
4487 }
4488 } else if (super_check_offset.as_constant() == sc_offset) {
4489 // Need a slow path; fast failure is impossible.
4490 if (L_slow_path == &L_fallthrough) {
4491 local_jcc(Assembler::equal, *L_success);
4492 } else {
4493 local_jcc(Assembler::notEqual, *L_slow_path);
4494 final_jmp(*L_success);
4495 }
4496 } else {
4497 // No slow path; it's a fast decision.
4498 if (L_failure == &L_fallthrough) {
4499 local_jcc(Assembler::equal, *L_success);
4500 } else {
4501 local_jcc(Assembler::notEqual, *L_failure);
4502 final_jmp(*L_success);
4503 }
4504 }
4505
4506 bind(L_fallthrough);
4507
4508 #undef local_jcc
4509 #undef final_jmp
4510 }
4511
4512
4513 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
4514 Register super_klass,
4515 Register temp_reg,
4516 Register temp2_reg,
4517 Label* L_success,
4518 Label* L_failure,
4519 bool set_cond_codes) {
4520 assert_different_registers(sub_klass, super_klass, temp_reg);
4521 if (temp2_reg != noreg)
4522 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg);
4523 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
4524
4525 Label L_fallthrough;
4526 int label_nulls = 0;
4527 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
4528 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
4529 assert(label_nulls <= 1, "at most one NULL in the batch");
4530
4531 // a couple of useful fields in sub_klass:
4532 int ss_offset = in_bytes(Klass::secondary_supers_offset());
4533 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
4534 Address secondary_supers_addr(sub_klass, ss_offset);
4535 Address super_cache_addr( sub_klass, sc_offset);
4536
4537 // Do a linear scan of the secondary super-klass chain.
4538 // This code is rarely used, so simplicity is a virtue here.
4539 // The repne_scan instruction uses fixed registers, which we must spill.
4540 // Don't worry too much about pre-existing connections with the input regs.
4541
4542 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super)
4543 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter)
4544
4545 // Get super_klass value into rax (even if it was in rdi or rcx).
4546 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false;
4547 if (super_klass != rax || UseCompressedOops) {
4548 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; }
4549 mov(rax, super_klass);
4550 }
4551 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; }
4552 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; }
4553
4554 #ifndef PRODUCT
4555 int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
4556 ExternalAddress pst_counter_addr((address) pst_counter);
4557 NOT_LP64( incrementl(pst_counter_addr) );
4558 LP64_ONLY( lea(rcx, pst_counter_addr) );
4559 LP64_ONLY( incrementl(Address(rcx, 0)) );
4560 #endif //PRODUCT
4561
4562 // We will consult the secondary-super array.
4563 movptr(rdi, secondary_supers_addr);
4564 // Load the array length. (Positive movl does right thing on LP64.)
4565 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes()));
4566 // Skip to start of data.
4567 addptr(rdi, Array<Klass*>::base_offset_in_bytes());
4568
4569 // Scan RCX words at [RDI] for an occurrence of RAX.
4570 // Set NZ/Z based on last compare.
4571 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does
4572 // not change flags (only scas instruction which is repeated sets flags).
4573 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found.
4574
4575 testptr(rax,rax); // Set Z = 0
4576 repne_scan();
4577
4578 // Unspill the temp. registers:
4579 if (pushed_rdi) pop(rdi);
4580 if (pushed_rcx) pop(rcx);
4581 if (pushed_rax) pop(rax);
4582
4583 if (set_cond_codes) {
4584 // Special hack for the AD files: rdi is guaranteed non-zero.
4585 assert(!pushed_rdi, "rdi must be left non-NULL");
4586 // Also, the condition codes are properly set Z/NZ on succeed/failure.
4587 }
4588
4589 if (L_failure == &L_fallthrough)
4590 jccb(Assembler::notEqual, *L_failure);
4591 else jcc(Assembler::notEqual, *L_failure);
4592
4593 // Success. Cache the super we found and proceed in triumph.
4594 movptr(super_cache_addr, super_klass);
4595
4596 if (L_success != &L_fallthrough) {
4597 jmp(*L_success);
4598 }
4599
4600 #undef IS_A_TEMP
4601
4602 bind(L_fallthrough);
4603 }
4604
4605
4606 void MacroAssembler::clinit_barrier(Register klass, Register thread, Label* L_fast_path, Label* L_slow_path) {
4607 assert(L_fast_path != NULL || L_slow_path != NULL, "at least one is required");
4608
4609 Label L_fallthrough;
4610 if (L_fast_path == NULL) {
4611 L_fast_path = &L_fallthrough;
4612 }
4613
4614 // Fast path check: class is fully initialized
4615 cmpb(Address(klass, InstanceKlass::init_state_offset()), InstanceKlass::fully_initialized);
4616 jcc(Assembler::equal, *L_fast_path);
4617
4618 // Fast path check: current thread is initializer thread
4619 cmpptr(thread, Address(klass, InstanceKlass::init_thread_offset()));
4620 if (L_slow_path != NULL) {
4621 jcc(Assembler::notEqual, *L_slow_path);
4622 } else {
4623 jcc(Assembler::equal, *L_fast_path);
4624 }
4625
4626 bind(L_fallthrough);
4627 }
4628
4629 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
4630 if (VM_Version::supports_cmov()) {
4631 cmovl(cc, dst, src);
4632 } else {
4633 Label L;
4634 jccb(negate_condition(cc), L);
4635 movl(dst, src);
4636 bind(L);
4637 }
4638 }
4639
4640 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) {
4641 if (VM_Version::supports_cmov()) {
4642 cmovl(cc, dst, src);
4643 } else {
4644 Label L;
4645 jccb(negate_condition(cc), L);
4646 movl(dst, src);
4647 bind(L);
4648 }
4649 }
4650
4651 void MacroAssembler::verify_oop(Register reg, const char* s) {
4652 if (!VerifyOops) return;
4653
4654 // Pass register number to verify_oop_subroutine
4655 const char* b = NULL;
4656 {
4657 ResourceMark rm;
4658 stringStream ss;
4659 ss.print("verify_oop: %s: %s", reg->name(), s);
4660 b = code_string(ss.as_string());
4661 }
4662 BLOCK_COMMENT("verify_oop {");
4663 #ifdef _LP64
4664 push(rscratch1); // save r10, trashed by movptr()
4665 #endif
4666 push(rax); // save rax,
4667 push(reg); // pass register argument
4668 ExternalAddress buffer((address) b);
4669 // avoid using pushptr, as it modifies scratch registers
4670 // and our contract is not to modify anything
4671 movptr(rax, buffer.addr());
4672 push(rax);
4673 // call indirectly to solve generation ordering problem
4674 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
4675 call(rax);
4676 // Caller pops the arguments (oop, message) and restores rax, r10
4677 BLOCK_COMMENT("} verify_oop");
4678 }
4679
4680
4681 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
4682 Register tmp,
4683 int offset) {
4684 intptr_t value = *delayed_value_addr;
4685 if (value != 0)
4686 return RegisterOrConstant(value + offset);
4687
4688 // load indirectly to solve generation ordering problem
4689 movptr(tmp, ExternalAddress((address) delayed_value_addr));
4690
4691 #ifdef ASSERT
4692 { Label L;
4693 testptr(tmp, tmp);
4694 if (WizardMode) {
4695 const char* buf = NULL;
4696 {
4697 ResourceMark rm;
4698 stringStream ss;
4699 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]);
4700 buf = code_string(ss.as_string());
4701 }
4702 jcc(Assembler::notZero, L);
4703 STOP(buf);
4704 } else {
4705 jccb(Assembler::notZero, L);
4706 hlt();
4707 }
4708 bind(L);
4709 }
4710 #endif
4711
4712 if (offset != 0)
4713 addptr(tmp, offset);
4714
4715 return RegisterOrConstant(tmp);
4716 }
4717
4718
4719 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
4720 int extra_slot_offset) {
4721 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
4722 int stackElementSize = Interpreter::stackElementSize;
4723 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
4724 #ifdef ASSERT
4725 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
4726 assert(offset1 - offset == stackElementSize, "correct arithmetic");
4727 #endif
4728 Register scale_reg = noreg;
4729 Address::ScaleFactor scale_factor = Address::no_scale;
4730 if (arg_slot.is_constant()) {
4731 offset += arg_slot.as_constant() * stackElementSize;
4732 } else {
4733 scale_reg = arg_slot.as_register();
4734 scale_factor = Address::times(stackElementSize);
4735 }
4736 offset += wordSize; // return PC is on stack
4737 return Address(rsp, scale_reg, scale_factor, offset);
4738 }
4739
4740
4741 void MacroAssembler::verify_oop_addr(Address addr, const char* s) {
4742 if (!VerifyOops) return;
4743
4744 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord);
4745 // Pass register number to verify_oop_subroutine
4746 const char* b = NULL;
4747 {
4748 ResourceMark rm;
4749 stringStream ss;
4750 ss.print("verify_oop_addr: %s", s);
4751 b = code_string(ss.as_string());
4752 }
4753 #ifdef _LP64
4754 push(rscratch1); // save r10, trashed by movptr()
4755 #endif
4756 push(rax); // save rax,
4757 // addr may contain rsp so we will have to adjust it based on the push
4758 // we just did (and on 64 bit we do two pushes)
4759 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which
4760 // stores rax into addr which is backwards of what was intended.
4761 if (addr.uses(rsp)) {
4762 lea(rax, addr);
4763 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord));
4764 } else {
4765 pushptr(addr);
4766 }
4767
4768 ExternalAddress buffer((address) b);
4769 // pass msg argument
4770 // avoid using pushptr, as it modifies scratch registers
4771 // and our contract is not to modify anything
4772 movptr(rax, buffer.addr());
4773 push(rax);
4774
4775 // call indirectly to solve generation ordering problem
4776 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
4777 call(rax);
4778 // Caller pops the arguments (addr, message) and restores rax, r10.
4779 }
4780
4781 void MacroAssembler::verify_tlab() {
4782 #ifdef ASSERT
4783 if (UseTLAB && VerifyOops) {
4784 Label next, ok;
4785 Register t1 = rsi;
4786 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread);
4787
4788 push(t1);
4789 NOT_LP64(push(thread_reg));
4790 NOT_LP64(get_thread(thread_reg));
4791
4792 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
4793 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
4794 jcc(Assembler::aboveEqual, next);
4795 STOP("assert(top >= start)");
4796 should_not_reach_here();
4797
4798 bind(next);
4799 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
4800 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
4801 jcc(Assembler::aboveEqual, ok);
4802 STOP("assert(top <= end)");
4803 should_not_reach_here();
4804
4805 bind(ok);
4806 NOT_LP64(pop(thread_reg));
4807 pop(t1);
4808 }
4809 #endif
4810 }
4811
4812 class ControlWord {
4813 public:
4814 int32_t _value;
4815
4816 int rounding_control() const { return (_value >> 10) & 3 ; }
4817 int precision_control() const { return (_value >> 8) & 3 ; }
4818 bool precision() const { return ((_value >> 5) & 1) != 0; }
4819 bool underflow() const { return ((_value >> 4) & 1) != 0; }
4820 bool overflow() const { return ((_value >> 3) & 1) != 0; }
4821 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
4822 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
4823 bool invalid() const { return ((_value >> 0) & 1) != 0; }
4824
4825 void print() const {
4826 // rounding control
4827 const char* rc;
4828 switch (rounding_control()) {
4829 case 0: rc = "round near"; break;
4830 case 1: rc = "round down"; break;
4831 case 2: rc = "round up "; break;
4832 case 3: rc = "chop "; break;
4833 };
4834 // precision control
4835 const char* pc;
4836 switch (precision_control()) {
4837 case 0: pc = "24 bits "; break;
4838 case 1: pc = "reserved"; break;
4839 case 2: pc = "53 bits "; break;
4840 case 3: pc = "64 bits "; break;
4841 };
4842 // flags
4843 char f[9];
4844 f[0] = ' ';
4845 f[1] = ' ';
4846 f[2] = (precision ()) ? 'P' : 'p';
4847 f[3] = (underflow ()) ? 'U' : 'u';
4848 f[4] = (overflow ()) ? 'O' : 'o';
4849 f[5] = (zero_divide ()) ? 'Z' : 'z';
4850 f[6] = (denormalized()) ? 'D' : 'd';
4851 f[7] = (invalid ()) ? 'I' : 'i';
4852 f[8] = '\x0';
4853 // output
4854 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc);
4855 }
4856
4857 };
4858
4859 class StatusWord {
4860 public:
4861 int32_t _value;
4862
4863 bool busy() const { return ((_value >> 15) & 1) != 0; }
4864 bool C3() const { return ((_value >> 14) & 1) != 0; }
4865 bool C2() const { return ((_value >> 10) & 1) != 0; }
4866 bool C1() const { return ((_value >> 9) & 1) != 0; }
4867 bool C0() const { return ((_value >> 8) & 1) != 0; }
4868 int top() const { return (_value >> 11) & 7 ; }
4869 bool error_status() const { return ((_value >> 7) & 1) != 0; }
4870 bool stack_fault() const { return ((_value >> 6) & 1) != 0; }
4871 bool precision() const { return ((_value >> 5) & 1) != 0; }
4872 bool underflow() const { return ((_value >> 4) & 1) != 0; }
4873 bool overflow() const { return ((_value >> 3) & 1) != 0; }
4874 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
4875 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
4876 bool invalid() const { return ((_value >> 0) & 1) != 0; }
4877
4878 void print() const {
4879 // condition codes
4880 char c[5];
4881 c[0] = (C3()) ? '3' : '-';
4882 c[1] = (C2()) ? '2' : '-';
4883 c[2] = (C1()) ? '1' : '-';
4884 c[3] = (C0()) ? '0' : '-';
4885 c[4] = '\x0';
4886 // flags
4887 char f[9];
4888 f[0] = (error_status()) ? 'E' : '-';
4889 f[1] = (stack_fault ()) ? 'S' : '-';
4890 f[2] = (precision ()) ? 'P' : '-';
4891 f[3] = (underflow ()) ? 'U' : '-';
4892 f[4] = (overflow ()) ? 'O' : '-';
4893 f[5] = (zero_divide ()) ? 'Z' : '-';
4894 f[6] = (denormalized()) ? 'D' : '-';
4895 f[7] = (invalid ()) ? 'I' : '-';
4896 f[8] = '\x0';
4897 // output
4898 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top());
4899 }
4900
4901 };
4902
4903 class TagWord {
4904 public:
4905 int32_t _value;
4906
4907 int tag_at(int i) const { return (_value >> (i*2)) & 3; }
4908
4909 void print() const {
4910 printf("%04x", _value & 0xFFFF);
4911 }
4912
4913 };
4914
4915 class FPU_Register {
4916 public:
4917 int32_t _m0;
4918 int32_t _m1;
4919 int16_t _ex;
4920
4921 bool is_indefinite() const {
4922 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0;
4923 }
4924
4925 void print() const {
4926 char sign = (_ex < 0) ? '-' : '+';
4927 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " ";
4928 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind);
4929 };
4930
4931 };
4932
4933 class FPU_State {
4934 public:
4935 enum {
4936 register_size = 10,
4937 number_of_registers = 8,
4938 register_mask = 7
4939 };
4940
4941 ControlWord _control_word;
4942 StatusWord _status_word;
4943 TagWord _tag_word;
4944 int32_t _error_offset;
4945 int32_t _error_selector;
4946 int32_t _data_offset;
4947 int32_t _data_selector;
4948 int8_t _register[register_size * number_of_registers];
4949
4950 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); }
4951 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; }
4952
4953 const char* tag_as_string(int tag) const {
4954 switch (tag) {
4955 case 0: return "valid";
4956 case 1: return "zero";
4957 case 2: return "special";
4958 case 3: return "empty";
4959 }
4960 ShouldNotReachHere();
4961 return NULL;
4962 }
4963
4964 void print() const {
4965 // print computation registers
4966 { int t = _status_word.top();
4967 for (int i = 0; i < number_of_registers; i++) {
4968 int j = (i - t) & register_mask;
4969 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j);
4970 st(j)->print();
4971 printf(" %s\n", tag_as_string(_tag_word.tag_at(i)));
4972 }
4973 }
4974 printf("\n");
4975 // print control registers
4976 printf("ctrl = "); _control_word.print(); printf("\n");
4977 printf("stat = "); _status_word .print(); printf("\n");
4978 printf("tags = "); _tag_word .print(); printf("\n");
4979 }
4980
4981 };
4982
4983 class Flag_Register {
4984 public:
4985 int32_t _value;
4986
4987 bool overflow() const { return ((_value >> 11) & 1) != 0; }
4988 bool direction() const { return ((_value >> 10) & 1) != 0; }
4989 bool sign() const { return ((_value >> 7) & 1) != 0; }
4990 bool zero() const { return ((_value >> 6) & 1) != 0; }
4991 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; }
4992 bool parity() const { return ((_value >> 2) & 1) != 0; }
4993 bool carry() const { return ((_value >> 0) & 1) != 0; }
4994
4995 void print() const {
4996 // flags
4997 char f[8];
4998 f[0] = (overflow ()) ? 'O' : '-';
4999 f[1] = (direction ()) ? 'D' : '-';
5000 f[2] = (sign ()) ? 'S' : '-';
5001 f[3] = (zero ()) ? 'Z' : '-';
5002 f[4] = (auxiliary_carry()) ? 'A' : '-';
5003 f[5] = (parity ()) ? 'P' : '-';
5004 f[6] = (carry ()) ? 'C' : '-';
5005 f[7] = '\x0';
5006 // output
5007 printf("%08x flags = %s", _value, f);
5008 }
5009
5010 };
5011
5012 class IU_Register {
5013 public:
5014 int32_t _value;
5015
5016 void print() const {
5017 printf("%08x %11d", _value, _value);
5018 }
5019
5020 };
5021
5022 class IU_State {
5023 public:
5024 Flag_Register _eflags;
5025 IU_Register _rdi;
5026 IU_Register _rsi;
5027 IU_Register _rbp;
5028 IU_Register _rsp;
5029 IU_Register _rbx;
5030 IU_Register _rdx;
5031 IU_Register _rcx;
5032 IU_Register _rax;
5033
5034 void print() const {
5035 // computation registers
5036 printf("rax, = "); _rax.print(); printf("\n");
5037 printf("rbx, = "); _rbx.print(); printf("\n");
5038 printf("rcx = "); _rcx.print(); printf("\n");
5039 printf("rdx = "); _rdx.print(); printf("\n");
5040 printf("rdi = "); _rdi.print(); printf("\n");
5041 printf("rsi = "); _rsi.print(); printf("\n");
5042 printf("rbp, = "); _rbp.print(); printf("\n");
5043 printf("rsp = "); _rsp.print(); printf("\n");
5044 printf("\n");
5045 // control registers
5046 printf("flgs = "); _eflags.print(); printf("\n");
5047 }
5048 };
5049
5050
5051 class CPU_State {
5052 public:
5053 FPU_State _fpu_state;
5054 IU_State _iu_state;
5055
5056 void print() const {
5057 printf("--------------------------------------------------\n");
5058 _iu_state .print();
5059 printf("\n");
5060 _fpu_state.print();
5061 printf("--------------------------------------------------\n");
5062 }
5063
5064 };
5065
5066
5067 static void _print_CPU_state(CPU_State* state) {
5068 state->print();
5069 };
5070
5071
5072 void MacroAssembler::print_CPU_state() {
5073 push_CPU_state();
5074 push(rsp); // pass CPU state
5075 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state)));
5076 addptr(rsp, wordSize); // discard argument
5077 pop_CPU_state();
5078 }
5079
5080
5081 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) {
5082 static int counter = 0;
5083 FPU_State* fs = &state->_fpu_state;
5084 counter++;
5085 // For leaf calls, only verify that the top few elements remain empty.
5086 // We only need 1 empty at the top for C2 code.
5087 if( stack_depth < 0 ) {
5088 if( fs->tag_for_st(7) != 3 ) {
5089 printf("FPR7 not empty\n");
5090 state->print();
5091 assert(false, "error");
5092 return false;
5093 }
5094 return true; // All other stack states do not matter
5095 }
5096
5097 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std,
5098 "bad FPU control word");
5099
5100 // compute stack depth
5101 int i = 0;
5102 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++;
5103 int d = i;
5104 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++;
5105 // verify findings
5106 if (i != FPU_State::number_of_registers) {
5107 // stack not contiguous
5108 printf("%s: stack not contiguous at ST%d\n", s, i);
5109 state->print();
5110 assert(false, "error");
5111 return false;
5112 }
5113 // check if computed stack depth corresponds to expected stack depth
5114 if (stack_depth < 0) {
5115 // expected stack depth is -stack_depth or less
5116 if (d > -stack_depth) {
5117 // too many elements on the stack
5118 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d);
5119 state->print();
5120 assert(false, "error");
5121 return false;
5122 }
5123 } else {
5124 // expected stack depth is stack_depth
5125 if (d != stack_depth) {
5126 // wrong stack depth
5127 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d);
5128 state->print();
5129 assert(false, "error");
5130 return false;
5131 }
5132 }
5133 // everything is cool
5134 return true;
5135 }
5136
5137
5138 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
5139 if (!VerifyFPU) return;
5140 push_CPU_state();
5141 push(rsp); // pass CPU state
5142 ExternalAddress msg((address) s);
5143 // pass message string s
5144 pushptr(msg.addr());
5145 push(stack_depth); // pass stack depth
5146 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU)));
5147 addptr(rsp, 3 * wordSize); // discard arguments
5148 // check for error
5149 { Label L;
5150 testl(rax, rax);
5151 jcc(Assembler::notZero, L);
5152 int3(); // break if error condition
5153 bind(L);
5154 }
5155 pop_CPU_state();
5156 }
5157
5158 void MacroAssembler::restore_cpu_control_state_after_jni() {
5159 // Either restore the MXCSR register after returning from the JNI Call
5160 // or verify that it wasn't changed (with -Xcheck:jni flag).
5161 if (VM_Version::supports_sse()) {
5162 if (RestoreMXCSROnJNICalls) {
5163 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
5164 } else if (CheckJNICalls) {
5165 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
5166 }
5167 }
5168 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty.
5169 vzeroupper();
5170 // Reset k1 to 0xffff.
5171
5172 #ifdef COMPILER2
5173 if (PostLoopMultiversioning && VM_Version::supports_evex()) {
5174 push(rcx);
5175 movl(rcx, 0xffff);
5176 kmovwl(k1, rcx);
5177 pop(rcx);
5178 }
5179 #endif // COMPILER2
5180
5181 #ifndef _LP64
5182 // Either restore the x87 floating pointer control word after returning
5183 // from the JNI call or verify that it wasn't changed.
5184 if (CheckJNICalls) {
5185 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
5186 }
5187 #endif // _LP64
5188 }
5189
5190 // ((OopHandle)result).resolve();
5191 void MacroAssembler::resolve_oop_handle(Register result, Register tmp) {
5192 assert_different_registers(result, tmp);
5193
5194 // Only 64 bit platforms support GCs that require a tmp register
5195 // Only IN_HEAP loads require a thread_tmp register
5196 // OopHandle::resolve is an indirection like jobject.
5197 access_load_at(T_OBJECT, IN_NATIVE,
5198 result, Address(result, 0), tmp, /*tmp_thread*/noreg);
5199 }
5200
5201 // ((WeakHandle)result).resolve();
5202 void MacroAssembler::resolve_weak_handle(Register rresult, Register rtmp) {
5203 assert_different_registers(rresult, rtmp);
5204 Label resolved;
5205
5206 // A null weak handle resolves to null.
5207 cmpptr(rresult, 0);
5208 jcc(Assembler::equal, resolved);
5209
5210 // Only 64 bit platforms support GCs that require a tmp register
5211 // Only IN_HEAP loads require a thread_tmp register
5212 // WeakHandle::resolve is an indirection like jweak.
5213 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
5214 rresult, Address(rresult, 0), rtmp, /*tmp_thread*/noreg);
5215 bind(resolved);
5216 }
5217
5218 void MacroAssembler::load_mirror(Register mirror, Register method, Register tmp) {
5219 // get mirror
5220 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
5221 load_method_holder(mirror, method);
5222 movptr(mirror, Address(mirror, mirror_offset));
5223 resolve_oop_handle(mirror, tmp);
5224 }
5225
5226 void MacroAssembler::load_method_holder_cld(Register rresult, Register rmethod) {
5227 load_method_holder(rresult, rmethod);
5228 movptr(rresult, Address(rresult, InstanceKlass::class_loader_data_offset()));
5229 }
5230
5231 void MacroAssembler::load_method_holder(Register holder, Register method) {
5232 movptr(holder, Address(method, Method::const_offset())); // ConstMethod*
5233 movptr(holder, Address(holder, ConstMethod::constants_offset())); // ConstantPool*
5234 movptr(holder, Address(holder, ConstantPool::pool_holder_offset_in_bytes())); // InstanceKlass*
5235 }
5236
5237 void MacroAssembler::load_klass(Register dst, Register src) {
5238 #ifdef _LP64
5239 if (UseCompressedClassPointers) {
5240 movl(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5241 decode_klass_not_null(dst);
5242 } else
5243 #endif
5244 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5245 }
5246
5247 void MacroAssembler::load_prototype_header(Register dst, Register src) {
5248 load_klass(dst, src);
5249 movptr(dst, Address(dst, Klass::prototype_header_offset()));
5250 }
5251
5252 void MacroAssembler::store_klass(Register dst, Register src) {
5253 #ifdef _LP64
5254 if (UseCompressedClassPointers) {
5255 encode_klass_not_null(src);
5256 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5257 } else
5258 #endif
5259 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5260 }
5261
5262 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators, Register dst, Address src,
5263 Register tmp1, Register thread_tmp) {
5264 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5265 decorators = AccessInternal::decorator_fixup(decorators);
5266 bool as_raw = (decorators & AS_RAW) != 0;
5267 if (as_raw) {
5268 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
5269 } else {
5270 bs->load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
5271 }
5272 }
5273
5274 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators, Address dst, Register src,
5275 Register tmp1, Register tmp2) {
5276 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5277 decorators = AccessInternal::decorator_fixup(decorators);
5278 bool as_raw = (decorators & AS_RAW) != 0;
5279 if (as_raw) {
5280 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, src, tmp1, tmp2);
5281 } else {
5282 bs->store_at(this, decorators, type, dst, src, tmp1, tmp2);
5283 }
5284 }
5285
5286 void MacroAssembler::resolve(DecoratorSet decorators, Register obj) {
5287 // Use stronger ACCESS_WRITE|ACCESS_READ by default.
5288 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
5289 decorators |= ACCESS_READ | ACCESS_WRITE;
5290 }
5291 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5292 return bs->resolve(this, decorators, obj);
5293 }
5294
5295 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
5296 Register thread_tmp, DecoratorSet decorators) {
5297 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp);
5298 }
5299
5300 // Doesn't do verfication, generates fixed size code
5301 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1,
5302 Register thread_tmp, DecoratorSet decorators) {
5303 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, thread_tmp);
5304 }
5305
5306 void MacroAssembler::store_heap_oop(Address dst, Register src, Register tmp1,
5307 Register tmp2, DecoratorSet decorators) {
5308 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2);
5309 }
5310
5311 // Used for storing NULLs.
5312 void MacroAssembler::store_heap_oop_null(Address dst) {
5313 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg);
5314 }
5315
5316 #ifdef _LP64
5317 void MacroAssembler::store_klass_gap(Register dst, Register src) {
5318 if (UseCompressedClassPointers) {
5319 // Store to klass gap in destination
5320 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src);
5321 }
5322 }
5323
5324 #ifdef ASSERT
5325 void MacroAssembler::verify_heapbase(const char* msg) {
5326 assert (UseCompressedOops, "should be compressed");
5327 assert (Universe::heap() != NULL, "java heap should be initialized");
5328 if (CheckCompressedOops) {
5329 Label ok;
5330 push(rscratch1); // cmpptr trashes rscratch1
5331 cmpptr(r12_heapbase, ExternalAddress((address)CompressedOops::ptrs_base_addr()));
5332 jcc(Assembler::equal, ok);
5333 STOP(msg);
5334 bind(ok);
5335 pop(rscratch1);
5336 }
5337 }
5338 #endif
5339
5340 // Algorithm must match oop.inline.hpp encode_heap_oop.
5341 void MacroAssembler::encode_heap_oop(Register r) {
5342 #ifdef ASSERT
5343 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
5344 #endif
5345 verify_oop(r, "broken oop in encode_heap_oop");
5346 if (CompressedOops::base() == NULL) {
5347 if (CompressedOops::shift() != 0) {
5348 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5349 shrq(r, LogMinObjAlignmentInBytes);
5350 }
5351 return;
5352 }
5353 testq(r, r);
5354 cmovq(Assembler::equal, r, r12_heapbase);
5355 subq(r, r12_heapbase);
5356 shrq(r, LogMinObjAlignmentInBytes);
5357 }
5358
5359 void MacroAssembler::encode_heap_oop_not_null(Register r) {
5360 #ifdef ASSERT
5361 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
5362 if (CheckCompressedOops) {
5363 Label ok;
5364 testq(r, r);
5365 jcc(Assembler::notEqual, ok);
5366 STOP("null oop passed to encode_heap_oop_not_null");
5367 bind(ok);
5368 }
5369 #endif
5370 verify_oop(r, "broken oop in encode_heap_oop_not_null");
5371 if (CompressedOops::base() != NULL) {
5372 subq(r, r12_heapbase);
5373 }
5374 if (CompressedOops::shift() != 0) {
5375 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5376 shrq(r, LogMinObjAlignmentInBytes);
5377 }
5378 }
5379
5380 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
5381 #ifdef ASSERT
5382 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
5383 if (CheckCompressedOops) {
5384 Label ok;
5385 testq(src, src);
5386 jcc(Assembler::notEqual, ok);
5387 STOP("null oop passed to encode_heap_oop_not_null2");
5388 bind(ok);
5389 }
5390 #endif
5391 verify_oop(src, "broken oop in encode_heap_oop_not_null2");
5392 if (dst != src) {
5393 movq(dst, src);
5394 }
5395 if (CompressedOops::base() != NULL) {
5396 subq(dst, r12_heapbase);
5397 }
5398 if (CompressedOops::shift() != 0) {
5399 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5400 shrq(dst, LogMinObjAlignmentInBytes);
5401 }
5402 }
5403
5404 void MacroAssembler::decode_heap_oop(Register r) {
5405 #ifdef ASSERT
5406 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
5407 #endif
5408 if (CompressedOops::base() == NULL) {
5409 if (CompressedOops::shift() != 0) {
5410 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5411 shlq(r, LogMinObjAlignmentInBytes);
5412 }
5413 } else {
5414 Label done;
5415 shlq(r, LogMinObjAlignmentInBytes);
5416 jccb(Assembler::equal, done);
5417 addq(r, r12_heapbase);
5418 bind(done);
5419 }
5420 verify_oop(r, "broken oop in decode_heap_oop");
5421 }
5422
5423 void MacroAssembler::decode_heap_oop_not_null(Register r) {
5424 // Note: it will change flags
5425 assert (UseCompressedOops, "should only be used for compressed headers");
5426 assert (Universe::heap() != NULL, "java heap should be initialized");
5427 // Cannot assert, unverified entry point counts instructions (see .ad file)
5428 // vtableStubs also counts instructions in pd_code_size_limit.
5429 // Also do not verify_oop as this is called by verify_oop.
5430 if (CompressedOops::shift() != 0) {
5431 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5432 shlq(r, LogMinObjAlignmentInBytes);
5433 if (CompressedOops::base() != NULL) {
5434 addq(r, r12_heapbase);
5435 }
5436 } else {
5437 assert (CompressedOops::base() == NULL, "sanity");
5438 }
5439 }
5440
5441 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
5442 // Note: it will change flags
5443 assert (UseCompressedOops, "should only be used for compressed headers");
5444 assert (Universe::heap() != NULL, "java heap should be initialized");
5445 // Cannot assert, unverified entry point counts instructions (see .ad file)
5446 // vtableStubs also counts instructions in pd_code_size_limit.
5447 // Also do not verify_oop as this is called by verify_oop.
5448 if (CompressedOops::shift() != 0) {
5449 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5450 if (LogMinObjAlignmentInBytes == Address::times_8) {
5451 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0));
5452 } else {
5453 if (dst != src) {
5454 movq(dst, src);
5455 }
5456 shlq(dst, LogMinObjAlignmentInBytes);
5457 if (CompressedOops::base() != NULL) {
5458 addq(dst, r12_heapbase);
5459 }
5460 }
5461 } else {
5462 assert (CompressedOops::base() == NULL, "sanity");
5463 if (dst != src) {
5464 movq(dst, src);
5465 }
5466 }
5467 }
5468
5469 void MacroAssembler::encode_klass_not_null(Register r) {
5470 if (CompressedKlassPointers::base() != NULL) {
5471 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5472 assert(r != r12_heapbase, "Encoding a klass in r12");
5473 mov64(r12_heapbase, (int64_t)CompressedKlassPointers::base());
5474 subq(r, r12_heapbase);
5475 }
5476 if (CompressedKlassPointers::shift() != 0) {
5477 assert (LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong");
5478 shrq(r, LogKlassAlignmentInBytes);
5479 }
5480 if (CompressedKlassPointers::base() != NULL) {
5481 reinit_heapbase();
5482 }
5483 }
5484
5485 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
5486 if (dst == src) {
5487 encode_klass_not_null(src);
5488 } else {
5489 if (CompressedKlassPointers::base() != NULL) {
5490 mov64(dst, (int64_t)CompressedKlassPointers::base());
5491 negq(dst);
5492 addq(dst, src);
5493 } else {
5494 movptr(dst, src);
5495 }
5496 if (CompressedKlassPointers::shift() != 0) {
5497 assert (LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong");
5498 shrq(dst, LogKlassAlignmentInBytes);
5499 }
5500 }
5501 }
5502
5503 // Function instr_size_for_decode_klass_not_null() counts the instructions
5504 // generated by decode_klass_not_null(register r) and reinit_heapbase(),
5505 // when (Universe::heap() != NULL). Hence, if the instructions they
5506 // generate change, then this method needs to be updated.
5507 int MacroAssembler::instr_size_for_decode_klass_not_null() {
5508 assert (UseCompressedClassPointers, "only for compressed klass ptrs");
5509 if (CompressedKlassPointers::base() != NULL) {
5510 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()).
5511 return (CompressedKlassPointers::shift() == 0 ? 20 : 24);
5512 } else {
5513 // longest load decode klass function, mov64, leaq
5514 return 16;
5515 }
5516 }
5517
5518 // !!! If the instructions that get generated here change then function
5519 // instr_size_for_decode_klass_not_null() needs to get updated.
5520 void MacroAssembler::decode_klass_not_null(Register r) {
5521 // Note: it will change flags
5522 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5523 assert(r != r12_heapbase, "Decoding a klass in r12");
5524 // Cannot assert, unverified entry point counts instructions (see .ad file)
5525 // vtableStubs also counts instructions in pd_code_size_limit.
5526 // Also do not verify_oop as this is called by verify_oop.
5527 if (CompressedKlassPointers::shift() != 0) {
5528 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong");
5529 shlq(r, LogKlassAlignmentInBytes);
5530 }
5531 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5532 if (CompressedKlassPointers::base() != NULL) {
5533 mov64(r12_heapbase, (int64_t)CompressedKlassPointers::base());
5534 addq(r, r12_heapbase);
5535 reinit_heapbase();
5536 }
5537 }
5538
5539 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
5540 // Note: it will change flags
5541 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5542 if (dst == src) {
5543 decode_klass_not_null(dst);
5544 } else {
5545 // Cannot assert, unverified entry point counts instructions (see .ad file)
5546 // vtableStubs also counts instructions in pd_code_size_limit.
5547 // Also do not verify_oop as this is called by verify_oop.
5548 mov64(dst, (int64_t)CompressedKlassPointers::base());
5549 if (CompressedKlassPointers::shift() != 0) {
5550 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift(), "decode alg wrong");
5551 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?");
5552 leaq(dst, Address(dst, src, Address::times_8, 0));
5553 } else {
5554 addq(dst, src);
5555 }
5556 }
5557 }
5558
5559 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
5560 assert (UseCompressedOops, "should only be used for compressed headers");
5561 assert (Universe::heap() != NULL, "java heap should be initialized");
5562 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5563 int oop_index = oop_recorder()->find_index(obj);
5564 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5565 mov_narrow_oop(dst, oop_index, rspec);
5566 }
5567
5568 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) {
5569 assert (UseCompressedOops, "should only be used for compressed headers");
5570 assert (Universe::heap() != NULL, "java heap should be initialized");
5571 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5572 int oop_index = oop_recorder()->find_index(obj);
5573 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5574 mov_narrow_oop(dst, oop_index, rspec);
5575 }
5576
5577 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
5578 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5579 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5580 int klass_index = oop_recorder()->find_index(k);
5581 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5582 mov_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec);
5583 }
5584
5585 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) {
5586 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5587 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5588 int klass_index = oop_recorder()->find_index(k);
5589 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5590 mov_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec);
5591 }
5592
5593 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) {
5594 assert (UseCompressedOops, "should only be used for compressed headers");
5595 assert (Universe::heap() != NULL, "java heap should be initialized");
5596 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5597 int oop_index = oop_recorder()->find_index(obj);
5598 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5599 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
5600 }
5601
5602 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) {
5603 assert (UseCompressedOops, "should only be used for compressed headers");
5604 assert (Universe::heap() != NULL, "java heap should be initialized");
5605 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5606 int oop_index = oop_recorder()->find_index(obj);
5607 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5608 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
5609 }
5610
5611 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) {
5612 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5613 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5614 int klass_index = oop_recorder()->find_index(k);
5615 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5616 Assembler::cmp_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec);
5617 }
5618
5619 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) {
5620 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5621 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5622 int klass_index = oop_recorder()->find_index(k);
5623 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5624 Assembler::cmp_narrow_oop(dst, CompressedKlassPointers::encode(k), rspec);
5625 }
5626
5627 void MacroAssembler::reinit_heapbase() {
5628 if (UseCompressedOops || UseCompressedClassPointers) {
5629 if (Universe::heap() != NULL) {
5630 if (CompressedOops::base() == NULL) {
5631 MacroAssembler::xorptr(r12_heapbase, r12_heapbase);
5632 } else {
5633 mov64(r12_heapbase, (int64_t)CompressedOops::ptrs_base());
5634 }
5635 } else {
5636 movptr(r12_heapbase, ExternalAddress((address)CompressedOops::ptrs_base_addr()));
5637 }
5638 }
5639 }
5640
5641 #endif // _LP64
5642
5643 // C2 compiled method's prolog code.
5644 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b, bool is_stub) {
5645
5646 // WARNING: Initial instruction MUST be 5 bytes or longer so that
5647 // NativeJump::patch_verified_entry will be able to patch out the entry
5648 // code safely. The push to verify stack depth is ok at 5 bytes,
5649 // the frame allocation can be either 3 or 6 bytes. So if we don't do
5650 // stack bang then we must use the 6 byte frame allocation even if
5651 // we have no frame. :-(
5652 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect");
5653
5654 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
5655 // Remove word for return addr
5656 framesize -= wordSize;
5657 stack_bang_size -= wordSize;
5658
5659 // Calls to C2R adapters often do not accept exceptional returns.
5660 // We require that their callers must bang for them. But be careful, because
5661 // some VM calls (such as call site linkage) can use several kilobytes of
5662 // stack. But the stack safety zone should account for that.
5663 // See bugs 4446381, 4468289, 4497237.
5664 if (stack_bang_size > 0) {
5665 generate_stack_overflow_check(stack_bang_size);
5666
5667 // We always push rbp, so that on return to interpreter rbp, will be
5668 // restored correctly and we can correct the stack.
5669 push(rbp);
5670 // Save caller's stack pointer into RBP if the frame pointer is preserved.
5671 if (PreserveFramePointer) {
5672 mov(rbp, rsp);
5673 }
5674 // Remove word for ebp
5675 framesize -= wordSize;
5676
5677 // Create frame
5678 if (framesize) {
5679 subptr(rsp, framesize);
5680 }
5681 } else {
5682 // Create frame (force generation of a 4 byte immediate value)
5683 subptr_imm32(rsp, framesize);
5684
5685 // Save RBP register now.
5686 framesize -= wordSize;
5687 movptr(Address(rsp, framesize), rbp);
5688 // Save caller's stack pointer into RBP if the frame pointer is preserved.
5689 if (PreserveFramePointer) {
5690 movptr(rbp, rsp);
5691 if (framesize > 0) {
5692 addptr(rbp, framesize);
5693 }
5694 }
5695 }
5696
5697 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
5698 framesize -= wordSize;
5699 movptr(Address(rsp, framesize), (int32_t)0xbadb100d);
5700 }
5701
5702 #ifndef _LP64
5703 // If method sets FPU control word do it now
5704 if (fp_mode_24b) {
5705 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
5706 }
5707 if (UseSSE >= 2 && VerifyFPU) {
5708 verify_FPU(0, "FPU stack must be clean on entry");
5709 }
5710 #endif
5711
5712 #ifdef ASSERT
5713 if (VerifyStackAtCalls) {
5714 Label L;
5715 push(rax);
5716 mov(rax, rsp);
5717 andptr(rax, StackAlignmentInBytes-1);
5718 cmpptr(rax, StackAlignmentInBytes-wordSize);
5719 pop(rax);
5720 jcc(Assembler::equal, L);
5721 STOP("Stack is not properly aligned!");
5722 bind(L);
5723 }
5724 #endif
5725
5726 if (!is_stub) {
5727 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5728 bs->nmethod_entry_barrier(this);
5729 }
5730 }
5731
5732 // clear memory of size 'cnt' qwords, starting at 'base' using XMM/YMM registers
5733 void MacroAssembler::xmm_clear_mem(Register base, Register cnt, XMMRegister xtmp) {
5734 // cnt - number of qwords (8-byte words).
5735 // base - start address, qword aligned.
5736 Label L_zero_64_bytes, L_loop, L_sloop, L_tail, L_end;
5737 if (UseAVX >= 2) {
5738 vpxor(xtmp, xtmp, xtmp, AVX_256bit);
5739 } else {
5740 pxor(xtmp, xtmp);
5741 }
5742 jmp(L_zero_64_bytes);
5743
5744 BIND(L_loop);
5745 if (UseAVX >= 2) {
5746 vmovdqu(Address(base, 0), xtmp);
5747 vmovdqu(Address(base, 32), xtmp);
5748 } else {
5749 movdqu(Address(base, 0), xtmp);
5750 movdqu(Address(base, 16), xtmp);
5751 movdqu(Address(base, 32), xtmp);
5752 movdqu(Address(base, 48), xtmp);
5753 }
5754 addptr(base, 64);
5755
5756 BIND(L_zero_64_bytes);
5757 subptr(cnt, 8);
5758 jccb(Assembler::greaterEqual, L_loop);
5759 addptr(cnt, 4);
5760 jccb(Assembler::less, L_tail);
5761 // Copy trailing 32 bytes
5762 if (UseAVX >= 2) {
5763 vmovdqu(Address(base, 0), xtmp);
5764 } else {
5765 movdqu(Address(base, 0), xtmp);
5766 movdqu(Address(base, 16), xtmp);
5767 }
5768 addptr(base, 32);
5769 subptr(cnt, 4);
5770
5771 BIND(L_tail);
5772 addptr(cnt, 4);
5773 jccb(Assembler::lessEqual, L_end);
5774 decrement(cnt);
5775
5776 BIND(L_sloop);
5777 movq(Address(base, 0), xtmp);
5778 addptr(base, 8);
5779 decrement(cnt);
5780 jccb(Assembler::greaterEqual, L_sloop);
5781 BIND(L_end);
5782 }
5783
5784 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp, XMMRegister xtmp, bool is_large) {
5785 // cnt - number of qwords (8-byte words).
5786 // base - start address, qword aligned.
5787 // is_large - if optimizers know cnt is larger than InitArrayShortSize
5788 assert(base==rdi, "base register must be edi for rep stos");
5789 assert(tmp==rax, "tmp register must be eax for rep stos");
5790 assert(cnt==rcx, "cnt register must be ecx for rep stos");
5791 assert(InitArrayShortSize % BytesPerLong == 0,
5792 "InitArrayShortSize should be the multiple of BytesPerLong");
5793
5794 Label DONE;
5795
5796 if (!is_large || !UseXMMForObjInit) {
5797 xorptr(tmp, tmp);
5798 }
5799
5800 if (!is_large) {
5801 Label LOOP, LONG;
5802 cmpptr(cnt, InitArrayShortSize/BytesPerLong);
5803 jccb(Assembler::greater, LONG);
5804
5805 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
5806
5807 decrement(cnt);
5808 jccb(Assembler::negative, DONE); // Zero length
5809
5810 // Use individual pointer-sized stores for small counts:
5811 BIND(LOOP);
5812 movptr(Address(base, cnt, Address::times_ptr), tmp);
5813 decrement(cnt);
5814 jccb(Assembler::greaterEqual, LOOP);
5815 jmpb(DONE);
5816
5817 BIND(LONG);
5818 }
5819
5820 // Use longer rep-prefixed ops for non-small counts:
5821 if (UseFastStosb) {
5822 shlptr(cnt, 3); // convert to number of bytes
5823 rep_stosb();
5824 } else if (UseXMMForObjInit) {
5825 movptr(tmp, base);
5826 xmm_clear_mem(tmp, cnt, xtmp);
5827 } else {
5828 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
5829 rep_stos();
5830 }
5831
5832 BIND(DONE);
5833 }
5834
5835 #ifdef COMPILER2
5836
5837 // IndexOf for constant substrings with size >= 8 chars
5838 // which don't need to be loaded through stack.
5839 void MacroAssembler::string_indexofC8(Register str1, Register str2,
5840 Register cnt1, Register cnt2,
5841 int int_cnt2, Register result,
5842 XMMRegister vec, Register tmp,
5843 int ae) {
5844 ShortBranchVerifier sbv(this);
5845 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
5846 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
5847
5848 // This method uses the pcmpestri instruction with bound registers
5849 // inputs:
5850 // xmm - substring
5851 // rax - substring length (elements count)
5852 // mem - scanned string
5853 // rdx - string length (elements count)
5854 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
5855 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
5856 // outputs:
5857 // rcx - matched index in string
5858 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
5859 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
5860 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
5861 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
5862 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
5863
5864 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR,
5865 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR,
5866 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE;
5867
5868 // Note, inline_string_indexOf() generates checks:
5869 // if (substr.count > string.count) return -1;
5870 // if (substr.count == 0) return 0;
5871 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars");
5872
5873 // Load substring.
5874 if (ae == StrIntrinsicNode::UL) {
5875 pmovzxbw(vec, Address(str2, 0));
5876 } else {
5877 movdqu(vec, Address(str2, 0));
5878 }
5879 movl(cnt2, int_cnt2);
5880 movptr(result, str1); // string addr
5881
5882 if (int_cnt2 > stride) {
5883 jmpb(SCAN_TO_SUBSTR);
5884
5885 // Reload substr for rescan, this code
5886 // is executed only for large substrings (> 8 chars)
5887 bind(RELOAD_SUBSTR);
5888 if (ae == StrIntrinsicNode::UL) {
5889 pmovzxbw(vec, Address(str2, 0));
5890 } else {
5891 movdqu(vec, Address(str2, 0));
5892 }
5893 negptr(cnt2); // Jumped here with negative cnt2, convert to positive
5894
5895 bind(RELOAD_STR);
5896 // We came here after the beginning of the substring was
5897 // matched but the rest of it was not so we need to search
5898 // again. Start from the next element after the previous match.
5899
5900 // cnt2 is number of substring reminding elements and
5901 // cnt1 is number of string reminding elements when cmp failed.
5902 // Restored cnt1 = cnt1 - cnt2 + int_cnt2
5903 subl(cnt1, cnt2);
5904 addl(cnt1, int_cnt2);
5905 movl(cnt2, int_cnt2); // Now restore cnt2
5906
5907 decrementl(cnt1); // Shift to next element
5908 cmpl(cnt1, cnt2);
5909 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring
5910
5911 addptr(result, (1<<scale1));
5912
5913 } // (int_cnt2 > 8)
5914
5915 // Scan string for start of substr in 16-byte vectors
5916 bind(SCAN_TO_SUBSTR);
5917 pcmpestri(vec, Address(result, 0), mode);
5918 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
5919 subl(cnt1, stride);
5920 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
5921 cmpl(cnt1, cnt2);
5922 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
5923 addptr(result, 16);
5924 jmpb(SCAN_TO_SUBSTR);
5925
5926 // Found a potential substr
5927 bind(FOUND_CANDIDATE);
5928 // Matched whole vector if first element matched (tmp(rcx) == 0).
5929 if (int_cnt2 == stride) {
5930 jccb(Assembler::overflow, RET_FOUND); // OF == 1
5931 } else { // int_cnt2 > 8
5932 jccb(Assembler::overflow, FOUND_SUBSTR);
5933 }
5934 // After pcmpestri tmp(rcx) contains matched element index
5935 // Compute start addr of substr
5936 lea(result, Address(result, tmp, scale1));
5937
5938 // Make sure string is still long enough
5939 subl(cnt1, tmp);
5940 cmpl(cnt1, cnt2);
5941 if (int_cnt2 == stride) {
5942 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
5943 } else { // int_cnt2 > 8
5944 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD);
5945 }
5946 // Left less then substring.
5947
5948 bind(RET_NOT_FOUND);
5949 movl(result, -1);
5950 jmp(EXIT);
5951
5952 if (int_cnt2 > stride) {
5953 // This code is optimized for the case when whole substring
5954 // is matched if its head is matched.
5955 bind(MATCH_SUBSTR_HEAD);
5956 pcmpestri(vec, Address(result, 0), mode);
5957 // Reload only string if does not match
5958 jcc(Assembler::noOverflow, RELOAD_STR); // OF == 0
5959
5960 Label CONT_SCAN_SUBSTR;
5961 // Compare the rest of substring (> 8 chars).
5962 bind(FOUND_SUBSTR);
5963 // First 8 chars are already matched.
5964 negptr(cnt2);
5965 addptr(cnt2, stride);
5966
5967 bind(SCAN_SUBSTR);
5968 subl(cnt1, stride);
5969 cmpl(cnt2, -stride); // Do not read beyond substring
5970 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR);
5971 // Back-up strings to avoid reading beyond substring:
5972 // cnt1 = cnt1 - cnt2 + 8
5973 addl(cnt1, cnt2); // cnt2 is negative
5974 addl(cnt1, stride);
5975 movl(cnt2, stride); negptr(cnt2);
5976 bind(CONT_SCAN_SUBSTR);
5977 if (int_cnt2 < (int)G) {
5978 int tail_off1 = int_cnt2<<scale1;
5979 int tail_off2 = int_cnt2<<scale2;
5980 if (ae == StrIntrinsicNode::UL) {
5981 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2));
5982 } else {
5983 movdqu(vec, Address(str2, cnt2, scale2, tail_off2));
5984 }
5985 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode);
5986 } else {
5987 // calculate index in register to avoid integer overflow (int_cnt2*2)
5988 movl(tmp, int_cnt2);
5989 addptr(tmp, cnt2);
5990 if (ae == StrIntrinsicNode::UL) {
5991 pmovzxbw(vec, Address(str2, tmp, scale2, 0));
5992 } else {
5993 movdqu(vec, Address(str2, tmp, scale2, 0));
5994 }
5995 pcmpestri(vec, Address(result, tmp, scale1, 0), mode);
5996 }
5997 // Need to reload strings pointers if not matched whole vector
5998 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
5999 addptr(cnt2, stride);
6000 jcc(Assembler::negative, SCAN_SUBSTR);
6001 // Fall through if found full substring
6002
6003 } // (int_cnt2 > 8)
6004
6005 bind(RET_FOUND);
6006 // Found result if we matched full small substring.
6007 // Compute substr offset
6008 subptr(result, str1);
6009 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6010 shrl(result, 1); // index
6011 }
6012 bind(EXIT);
6013
6014 } // string_indexofC8
6015
6016 // Small strings are loaded through stack if they cross page boundary.
6017 void MacroAssembler::string_indexof(Register str1, Register str2,
6018 Register cnt1, Register cnt2,
6019 int int_cnt2, Register result,
6020 XMMRegister vec, Register tmp,
6021 int ae) {
6022 ShortBranchVerifier sbv(this);
6023 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
6024 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
6025
6026 //
6027 // int_cnt2 is length of small (< 8 chars) constant substring
6028 // or (-1) for non constant substring in which case its length
6029 // is in cnt2 register.
6030 //
6031 // Note, inline_string_indexOf() generates checks:
6032 // if (substr.count > string.count) return -1;
6033 // if (substr.count == 0) return 0;
6034 //
6035 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
6036 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0");
6037 // This method uses the pcmpestri instruction with bound registers
6038 // inputs:
6039 // xmm - substring
6040 // rax - substring length (elements count)
6041 // mem - scanned string
6042 // rdx - string length (elements count)
6043 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6044 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
6045 // outputs:
6046 // rcx - matched index in string
6047 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6048 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
6049 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
6050 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
6051
6052 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR,
6053 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR,
6054 FOUND_CANDIDATE;
6055
6056 { //========================================================
6057 // We don't know where these strings are located
6058 // and we can't read beyond them. Load them through stack.
6059 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR;
6060
6061 movptr(tmp, rsp); // save old SP
6062
6063 if (int_cnt2 > 0) { // small (< 8 chars) constant substring
6064 if (int_cnt2 == (1>>scale2)) { // One byte
6065 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding");
6066 load_unsigned_byte(result, Address(str2, 0));
6067 movdl(vec, result); // move 32 bits
6068 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes
6069 // Not enough header space in 32-bit VM: 12+3 = 15.
6070 movl(result, Address(str2, -1));
6071 shrl(result, 8);
6072 movdl(vec, result); // move 32 bits
6073 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char
6074 load_unsigned_short(result, Address(str2, 0));
6075 movdl(vec, result); // move 32 bits
6076 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars
6077 movdl(vec, Address(str2, 0)); // move 32 bits
6078 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars
6079 movq(vec, Address(str2, 0)); // move 64 bits
6080 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7})
6081 // Array header size is 12 bytes in 32-bit VM
6082 // + 6 bytes for 3 chars == 18 bytes,
6083 // enough space to load vec and shift.
6084 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity");
6085 if (ae == StrIntrinsicNode::UL) {
6086 int tail_off = int_cnt2-8;
6087 pmovzxbw(vec, Address(str2, tail_off));
6088 psrldq(vec, -2*tail_off);
6089 }
6090 else {
6091 int tail_off = int_cnt2*(1<<scale2);
6092 movdqu(vec, Address(str2, tail_off-16));
6093 psrldq(vec, 16-tail_off);
6094 }
6095 }
6096 } else { // not constant substring
6097 cmpl(cnt2, stride);
6098 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough
6099
6100 // We can read beyond string if srt+16 does not cross page boundary
6101 // since heaps are aligned and mapped by pages.
6102 assert(os::vm_page_size() < (int)G, "default page should be small");
6103 movl(result, str2); // We need only low 32 bits
6104 andl(result, (os::vm_page_size()-1));
6105 cmpl(result, (os::vm_page_size()-16));
6106 jccb(Assembler::belowEqual, CHECK_STR);
6107
6108 // Move small strings to stack to allow load 16 bytes into vec.
6109 subptr(rsp, 16);
6110 int stk_offset = wordSize-(1<<scale2);
6111 push(cnt2);
6112
6113 bind(COPY_SUBSTR);
6114 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) {
6115 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1));
6116 movb(Address(rsp, cnt2, scale2, stk_offset), result);
6117 } else if (ae == StrIntrinsicNode::UU) {
6118 load_unsigned_short(result, Address(str2, cnt2, scale2, -2));
6119 movw(Address(rsp, cnt2, scale2, stk_offset), result);
6120 }
6121 decrement(cnt2);
6122 jccb(Assembler::notZero, COPY_SUBSTR);
6123
6124 pop(cnt2);
6125 movptr(str2, rsp); // New substring address
6126 } // non constant
6127
6128 bind(CHECK_STR);
6129 cmpl(cnt1, stride);
6130 jccb(Assembler::aboveEqual, BIG_STRINGS);
6131
6132 // Check cross page boundary.
6133 movl(result, str1); // We need only low 32 bits
6134 andl(result, (os::vm_page_size()-1));
6135 cmpl(result, (os::vm_page_size()-16));
6136 jccb(Assembler::belowEqual, BIG_STRINGS);
6137
6138 subptr(rsp, 16);
6139 int stk_offset = -(1<<scale1);
6140 if (int_cnt2 < 0) { // not constant
6141 push(cnt2);
6142 stk_offset += wordSize;
6143 }
6144 movl(cnt2, cnt1);
6145
6146 bind(COPY_STR);
6147 if (ae == StrIntrinsicNode::LL) {
6148 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1));
6149 movb(Address(rsp, cnt2, scale1, stk_offset), result);
6150 } else {
6151 load_unsigned_short(result, Address(str1, cnt2, scale1, -2));
6152 movw(Address(rsp, cnt2, scale1, stk_offset), result);
6153 }
6154 decrement(cnt2);
6155 jccb(Assembler::notZero, COPY_STR);
6156
6157 if (int_cnt2 < 0) { // not constant
6158 pop(cnt2);
6159 }
6160 movptr(str1, rsp); // New string address
6161
6162 bind(BIG_STRINGS);
6163 // Load substring.
6164 if (int_cnt2 < 0) { // -1
6165 if (ae == StrIntrinsicNode::UL) {
6166 pmovzxbw(vec, Address(str2, 0));
6167 } else {
6168 movdqu(vec, Address(str2, 0));
6169 }
6170 push(cnt2); // substr count
6171 push(str2); // substr addr
6172 push(str1); // string addr
6173 } else {
6174 // Small (< 8 chars) constant substrings are loaded already.
6175 movl(cnt2, int_cnt2);
6176 }
6177 push(tmp); // original SP
6178
6179 } // Finished loading
6180
6181 //========================================================
6182 // Start search
6183 //
6184
6185 movptr(result, str1); // string addr
6186
6187 if (int_cnt2 < 0) { // Only for non constant substring
6188 jmpb(SCAN_TO_SUBSTR);
6189
6190 // SP saved at sp+0
6191 // String saved at sp+1*wordSize
6192 // Substr saved at sp+2*wordSize
6193 // Substr count saved at sp+3*wordSize
6194
6195 // Reload substr for rescan, this code
6196 // is executed only for large substrings (> 8 chars)
6197 bind(RELOAD_SUBSTR);
6198 movptr(str2, Address(rsp, 2*wordSize));
6199 movl(cnt2, Address(rsp, 3*wordSize));
6200 if (ae == StrIntrinsicNode::UL) {
6201 pmovzxbw(vec, Address(str2, 0));
6202 } else {
6203 movdqu(vec, Address(str2, 0));
6204 }
6205 // We came here after the beginning of the substring was
6206 // matched but the rest of it was not so we need to search
6207 // again. Start from the next element after the previous match.
6208 subptr(str1, result); // Restore counter
6209 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6210 shrl(str1, 1);
6211 }
6212 addl(cnt1, str1);
6213 decrementl(cnt1); // Shift to next element
6214 cmpl(cnt1, cnt2);
6215 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6216
6217 addptr(result, (1<<scale1));
6218 } // non constant
6219
6220 // Scan string for start of substr in 16-byte vectors
6221 bind(SCAN_TO_SUBSTR);
6222 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6223 pcmpestri(vec, Address(result, 0), mode);
6224 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
6225 subl(cnt1, stride);
6226 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6227 cmpl(cnt1, cnt2);
6228 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6229 addptr(result, 16);
6230
6231 bind(ADJUST_STR);
6232 cmpl(cnt1, stride); // Do not read beyond string
6233 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6234 // Back-up string to avoid reading beyond string.
6235 lea(result, Address(result, cnt1, scale1, -16));
6236 movl(cnt1, stride);
6237 jmpb(SCAN_TO_SUBSTR);
6238
6239 // Found a potential substr
6240 bind(FOUND_CANDIDATE);
6241 // After pcmpestri tmp(rcx) contains matched element index
6242
6243 // Make sure string is still long enough
6244 subl(cnt1, tmp);
6245 cmpl(cnt1, cnt2);
6246 jccb(Assembler::greaterEqual, FOUND_SUBSTR);
6247 // Left less then substring.
6248
6249 bind(RET_NOT_FOUND);
6250 movl(result, -1);
6251 jmp(CLEANUP);
6252
6253 bind(FOUND_SUBSTR);
6254 // Compute start addr of substr
6255 lea(result, Address(result, tmp, scale1));
6256 if (int_cnt2 > 0) { // Constant substring
6257 // Repeat search for small substring (< 8 chars)
6258 // from new point without reloading substring.
6259 // Have to check that we don't read beyond string.
6260 cmpl(tmp, stride-int_cnt2);
6261 jccb(Assembler::greater, ADJUST_STR);
6262 // Fall through if matched whole substring.
6263 } else { // non constant
6264 assert(int_cnt2 == -1, "should be != 0");
6265
6266 addl(tmp, cnt2);
6267 // Found result if we matched whole substring.
6268 cmpl(tmp, stride);
6269 jcc(Assembler::lessEqual, RET_FOUND);
6270
6271 // Repeat search for small substring (<= 8 chars)
6272 // from new point 'str1' without reloading substring.
6273 cmpl(cnt2, stride);
6274 // Have to check that we don't read beyond string.
6275 jccb(Assembler::lessEqual, ADJUST_STR);
6276
6277 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG;
6278 // Compare the rest of substring (> 8 chars).
6279 movptr(str1, result);
6280
6281 cmpl(tmp, cnt2);
6282 // First 8 chars are already matched.
6283 jccb(Assembler::equal, CHECK_NEXT);
6284
6285 bind(SCAN_SUBSTR);
6286 pcmpestri(vec, Address(str1, 0), mode);
6287 // Need to reload strings pointers if not matched whole vector
6288 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6289
6290 bind(CHECK_NEXT);
6291 subl(cnt2, stride);
6292 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring
6293 addptr(str1, 16);
6294 if (ae == StrIntrinsicNode::UL) {
6295 addptr(str2, 8);
6296 } else {
6297 addptr(str2, 16);
6298 }
6299 subl(cnt1, stride);
6300 cmpl(cnt2, stride); // Do not read beyond substring
6301 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR);
6302 // Back-up strings to avoid reading beyond substring.
6303
6304 if (ae == StrIntrinsicNode::UL) {
6305 lea(str2, Address(str2, cnt2, scale2, -8));
6306 lea(str1, Address(str1, cnt2, scale1, -16));
6307 } else {
6308 lea(str2, Address(str2, cnt2, scale2, -16));
6309 lea(str1, Address(str1, cnt2, scale1, -16));
6310 }
6311 subl(cnt1, cnt2);
6312 movl(cnt2, stride);
6313 addl(cnt1, stride);
6314 bind(CONT_SCAN_SUBSTR);
6315 if (ae == StrIntrinsicNode::UL) {
6316 pmovzxbw(vec, Address(str2, 0));
6317 } else {
6318 movdqu(vec, Address(str2, 0));
6319 }
6320 jmp(SCAN_SUBSTR);
6321
6322 bind(RET_FOUND_LONG);
6323 movptr(str1, Address(rsp, wordSize));
6324 } // non constant
6325
6326 bind(RET_FOUND);
6327 // Compute substr offset
6328 subptr(result, str1);
6329 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6330 shrl(result, 1); // index
6331 }
6332 bind(CLEANUP);
6333 pop(rsp); // restore SP
6334
6335 } // string_indexof
6336
6337 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result,
6338 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) {
6339 ShortBranchVerifier sbv(this);
6340 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
6341
6342 int stride = 8;
6343
6344 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP,
6345 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP,
6346 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT,
6347 FOUND_SEQ_CHAR, DONE_LABEL;
6348
6349 movptr(result, str1);
6350 if (UseAVX >= 2) {
6351 cmpl(cnt1, stride);
6352 jcc(Assembler::less, SCAN_TO_CHAR_LOOP);
6353 cmpl(cnt1, 2*stride);
6354 jcc(Assembler::less, SCAN_TO_8_CHAR_INIT);
6355 movdl(vec1, ch);
6356 vpbroadcastw(vec1, vec1, Assembler::AVX_256bit);
6357 vpxor(vec2, vec2);
6358 movl(tmp, cnt1);
6359 andl(tmp, 0xFFFFFFF0); //vector count (in chars)
6360 andl(cnt1,0x0000000F); //tail count (in chars)
6361
6362 bind(SCAN_TO_16_CHAR_LOOP);
6363 vmovdqu(vec3, Address(result, 0));
6364 vpcmpeqw(vec3, vec3, vec1, 1);
6365 vptest(vec2, vec3);
6366 jcc(Assembler::carryClear, FOUND_CHAR);
6367 addptr(result, 32);
6368 subl(tmp, 2*stride);
6369 jcc(Assembler::notZero, SCAN_TO_16_CHAR_LOOP);
6370 jmp(SCAN_TO_8_CHAR);
6371 bind(SCAN_TO_8_CHAR_INIT);
6372 movdl(vec1, ch);
6373 pshuflw(vec1, vec1, 0x00);
6374 pshufd(vec1, vec1, 0);
6375 pxor(vec2, vec2);
6376 }
6377 bind(SCAN_TO_8_CHAR);
6378 cmpl(cnt1, stride);
6379 if (UseAVX >= 2) {
6380 jcc(Assembler::less, SCAN_TO_CHAR);
6381 } else {
6382 jcc(Assembler::less, SCAN_TO_CHAR_LOOP);
6383 movdl(vec1, ch);
6384 pshuflw(vec1, vec1, 0x00);
6385 pshufd(vec1, vec1, 0);
6386 pxor(vec2, vec2);
6387 }
6388 movl(tmp, cnt1);
6389 andl(tmp, 0xFFFFFFF8); //vector count (in chars)
6390 andl(cnt1,0x00000007); //tail count (in chars)
6391
6392 bind(SCAN_TO_8_CHAR_LOOP);
6393 movdqu(vec3, Address(result, 0));
6394 pcmpeqw(vec3, vec1);
6395 ptest(vec2, vec3);
6396 jcc(Assembler::carryClear, FOUND_CHAR);
6397 addptr(result, 16);
6398 subl(tmp, stride);
6399 jcc(Assembler::notZero, SCAN_TO_8_CHAR_LOOP);
6400 bind(SCAN_TO_CHAR);
6401 testl(cnt1, cnt1);
6402 jcc(Assembler::zero, RET_NOT_FOUND);
6403 bind(SCAN_TO_CHAR_LOOP);
6404 load_unsigned_short(tmp, Address(result, 0));
6405 cmpl(ch, tmp);
6406 jccb(Assembler::equal, FOUND_SEQ_CHAR);
6407 addptr(result, 2);
6408 subl(cnt1, 1);
6409 jccb(Assembler::zero, RET_NOT_FOUND);
6410 jmp(SCAN_TO_CHAR_LOOP);
6411
6412 bind(RET_NOT_FOUND);
6413 movl(result, -1);
6414 jmpb(DONE_LABEL);
6415
6416 bind(FOUND_CHAR);
6417 if (UseAVX >= 2) {
6418 vpmovmskb(tmp, vec3);
6419 } else {
6420 pmovmskb(tmp, vec3);
6421 }
6422 bsfl(ch, tmp);
6423 addl(result, ch);
6424
6425 bind(FOUND_SEQ_CHAR);
6426 subptr(result, str1);
6427 shrl(result, 1);
6428
6429 bind(DONE_LABEL);
6430 } // string_indexof_char
6431
6432 // helper function for string_compare
6433 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2,
6434 Address::ScaleFactor scale, Address::ScaleFactor scale1,
6435 Address::ScaleFactor scale2, Register index, int ae) {
6436 if (ae == StrIntrinsicNode::LL) {
6437 load_unsigned_byte(elem1, Address(str1, index, scale, 0));
6438 load_unsigned_byte(elem2, Address(str2, index, scale, 0));
6439 } else if (ae == StrIntrinsicNode::UU) {
6440 load_unsigned_short(elem1, Address(str1, index, scale, 0));
6441 load_unsigned_short(elem2, Address(str2, index, scale, 0));
6442 } else {
6443 load_unsigned_byte(elem1, Address(str1, index, scale1, 0));
6444 load_unsigned_short(elem2, Address(str2, index, scale2, 0));
6445 }
6446 }
6447
6448 // Compare strings, used for char[] and byte[].
6449 void MacroAssembler::string_compare(Register str1, Register str2,
6450 Register cnt1, Register cnt2, Register result,
6451 XMMRegister vec1, int ae) {
6452 ShortBranchVerifier sbv(this);
6453 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
6454 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3
6455 int stride, stride2, adr_stride, adr_stride1, adr_stride2;
6456 int stride2x2 = 0x40;
6457 Address::ScaleFactor scale = Address::no_scale;
6458 Address::ScaleFactor scale1 = Address::no_scale;
6459 Address::ScaleFactor scale2 = Address::no_scale;
6460
6461 if (ae != StrIntrinsicNode::LL) {
6462 stride2x2 = 0x20;
6463 }
6464
6465 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
6466 shrl(cnt2, 1);
6467 }
6468 // Compute the minimum of the string lengths and the
6469 // difference of the string lengths (stack).
6470 // Do the conditional move stuff
6471 movl(result, cnt1);
6472 subl(cnt1, cnt2);
6473 push(cnt1);
6474 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2)
6475
6476 // Is the minimum length zero?
6477 testl(cnt2, cnt2);
6478 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6479 if (ae == StrIntrinsicNode::LL) {
6480 // Load first bytes
6481 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0]
6482 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0]
6483 } else if (ae == StrIntrinsicNode::UU) {
6484 // Load first characters
6485 load_unsigned_short(result, Address(str1, 0));
6486 load_unsigned_short(cnt1, Address(str2, 0));
6487 } else {
6488 load_unsigned_byte(result, Address(str1, 0));
6489 load_unsigned_short(cnt1, Address(str2, 0));
6490 }
6491 subl(result, cnt1);
6492 jcc(Assembler::notZero, POP_LABEL);
6493
6494 if (ae == StrIntrinsicNode::UU) {
6495 // Divide length by 2 to get number of chars
6496 shrl(cnt2, 1);
6497 }
6498 cmpl(cnt2, 1);
6499 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6500
6501 // Check if the strings start at the same location and setup scale and stride
6502 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6503 cmpptr(str1, str2);
6504 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6505 if (ae == StrIntrinsicNode::LL) {
6506 scale = Address::times_1;
6507 stride = 16;
6508 } else {
6509 scale = Address::times_2;
6510 stride = 8;
6511 }
6512 } else {
6513 scale1 = Address::times_1;
6514 scale2 = Address::times_2;
6515 // scale not used
6516 stride = 8;
6517 }
6518
6519 if (UseAVX >= 2 && UseSSE42Intrinsics) {
6520 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
6521 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
6522 Label COMPARE_WIDE_VECTORS_LOOP_AVX2;
6523 Label COMPARE_TAIL_LONG;
6524 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3
6525
6526 int pcmpmask = 0x19;
6527 if (ae == StrIntrinsicNode::LL) {
6528 pcmpmask &= ~0x01;
6529 }
6530
6531 // Setup to compare 16-chars (32-bytes) vectors,
6532 // start from first character again because it has aligned address.
6533 if (ae == StrIntrinsicNode::LL) {
6534 stride2 = 32;
6535 } else {
6536 stride2 = 16;
6537 }
6538 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6539 adr_stride = stride << scale;
6540 } else {
6541 adr_stride1 = 8; //stride << scale1;
6542 adr_stride2 = 16; //stride << scale2;
6543 }
6544
6545 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6546 // rax and rdx are used by pcmpestri as elements counters
6547 movl(result, cnt2);
6548 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count
6549 jcc(Assembler::zero, COMPARE_TAIL_LONG);
6550
6551 // fast path : compare first 2 8-char vectors.
6552 bind(COMPARE_16_CHARS);
6553 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6554 movdqu(vec1, Address(str1, 0));
6555 } else {
6556 pmovzxbw(vec1, Address(str1, 0));
6557 }
6558 pcmpestri(vec1, Address(str2, 0), pcmpmask);
6559 jccb(Assembler::below, COMPARE_INDEX_CHAR);
6560
6561 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6562 movdqu(vec1, Address(str1, adr_stride));
6563 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
6564 } else {
6565 pmovzxbw(vec1, Address(str1, adr_stride1));
6566 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask);
6567 }
6568 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
6569 addl(cnt1, stride);
6570
6571 // Compare the characters at index in cnt1
6572 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character
6573 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
6574 subl(result, cnt2);
6575 jmp(POP_LABEL);
6576
6577 // Setup the registers to start vector comparison loop
6578 bind(COMPARE_WIDE_VECTORS);
6579 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6580 lea(str1, Address(str1, result, scale));
6581 lea(str2, Address(str2, result, scale));
6582 } else {
6583 lea(str1, Address(str1, result, scale1));
6584 lea(str2, Address(str2, result, scale2));
6585 }
6586 subl(result, stride2);
6587 subl(cnt2, stride2);
6588 jcc(Assembler::zero, COMPARE_WIDE_TAIL);
6589 negptr(result);
6590
6591 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
6592 bind(COMPARE_WIDE_VECTORS_LOOP);
6593
6594 #ifdef _LP64
6595 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
6596 cmpl(cnt2, stride2x2);
6597 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2);
6598 testl(cnt2, stride2x2-1); // cnt2 holds the vector count
6599 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40
6600
6601 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
6602 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6603 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit);
6604 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
6605 } else {
6606 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit);
6607 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
6608 }
6609 kortestql(k7, k7);
6610 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare
6611 addptr(result, stride2x2); // update since we already compared at this addr
6612 subl(cnt2, stride2x2); // and sub the size too
6613 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3);
6614
6615 vpxor(vec1, vec1);
6616 jmpb(COMPARE_WIDE_TAIL);
6617 }//if (VM_Version::supports_avx512vlbw())
6618 #endif // _LP64
6619
6620
6621 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
6622 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6623 vmovdqu(vec1, Address(str1, result, scale));
6624 vpxor(vec1, Address(str2, result, scale));
6625 } else {
6626 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit);
6627 vpxor(vec1, Address(str2, result, scale2));
6628 }
6629 vptest(vec1, vec1);
6630 jcc(Assembler::notZero, VECTOR_NOT_EQUAL);
6631 addptr(result, stride2);
6632 subl(cnt2, stride2);
6633 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
6634 // clean upper bits of YMM registers
6635 vpxor(vec1, vec1);
6636
6637 // compare wide vectors tail
6638 bind(COMPARE_WIDE_TAIL);
6639 testptr(result, result);
6640 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6641
6642 movl(result, stride2);
6643 movl(cnt2, result);
6644 negptr(result);
6645 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2);
6646
6647 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
6648 bind(VECTOR_NOT_EQUAL);
6649 // clean upper bits of YMM registers
6650 vpxor(vec1, vec1);
6651 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6652 lea(str1, Address(str1, result, scale));
6653 lea(str2, Address(str2, result, scale));
6654 } else {
6655 lea(str1, Address(str1, result, scale1));
6656 lea(str2, Address(str2, result, scale2));
6657 }
6658 jmp(COMPARE_16_CHARS);
6659
6660 // Compare tail chars, length between 1 to 15 chars
6661 bind(COMPARE_TAIL_LONG);
6662 movl(cnt2, result);
6663 cmpl(cnt2, stride);
6664 jcc(Assembler::less, COMPARE_SMALL_STR);
6665
6666 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6667 movdqu(vec1, Address(str1, 0));
6668 } else {
6669 pmovzxbw(vec1, Address(str1, 0));
6670 }
6671 pcmpestri(vec1, Address(str2, 0), pcmpmask);
6672 jcc(Assembler::below, COMPARE_INDEX_CHAR);
6673 subptr(cnt2, stride);
6674 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6675 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6676 lea(str1, Address(str1, result, scale));
6677 lea(str2, Address(str2, result, scale));
6678 } else {
6679 lea(str1, Address(str1, result, scale1));
6680 lea(str2, Address(str2, result, scale2));
6681 }
6682 negptr(cnt2);
6683 jmpb(WHILE_HEAD_LABEL);
6684
6685 bind(COMPARE_SMALL_STR);
6686 } else if (UseSSE42Intrinsics) {
6687 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
6688 int pcmpmask = 0x19;
6689 // Setup to compare 8-char (16-byte) vectors,
6690 // start from first character again because it has aligned address.
6691 movl(result, cnt2);
6692 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count
6693 if (ae == StrIntrinsicNode::LL) {
6694 pcmpmask &= ~0x01;
6695 }
6696 jcc(Assembler::zero, COMPARE_TAIL);
6697 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6698 lea(str1, Address(str1, result, scale));
6699 lea(str2, Address(str2, result, scale));
6700 } else {
6701 lea(str1, Address(str1, result, scale1));
6702 lea(str2, Address(str2, result, scale2));
6703 }
6704 negptr(result);
6705
6706 // pcmpestri
6707 // inputs:
6708 // vec1- substring
6709 // rax - negative string length (elements count)
6710 // mem - scanned string
6711 // rdx - string length (elements count)
6712 // pcmpmask - cmp mode: 11000 (string compare with negated result)
6713 // + 00 (unsigned bytes) or + 01 (unsigned shorts)
6714 // outputs:
6715 // rcx - first mismatched element index
6716 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6717
6718 bind(COMPARE_WIDE_VECTORS);
6719 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6720 movdqu(vec1, Address(str1, result, scale));
6721 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6722 } else {
6723 pmovzxbw(vec1, Address(str1, result, scale1));
6724 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
6725 }
6726 // After pcmpestri cnt1(rcx) contains mismatched element index
6727
6728 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1
6729 addptr(result, stride);
6730 subptr(cnt2, stride);
6731 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
6732
6733 // compare wide vectors tail
6734 testptr(result, result);
6735 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6736
6737 movl(cnt2, stride);
6738 movl(result, stride);
6739 negptr(result);
6740 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6741 movdqu(vec1, Address(str1, result, scale));
6742 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6743 } else {
6744 pmovzxbw(vec1, Address(str1, result, scale1));
6745 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
6746 }
6747 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
6748
6749 // Mismatched characters in the vectors
6750 bind(VECTOR_NOT_EQUAL);
6751 addptr(cnt1, result);
6752 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
6753 subl(result, cnt2);
6754 jmpb(POP_LABEL);
6755
6756 bind(COMPARE_TAIL); // limit is zero
6757 movl(cnt2, result);
6758 // Fallthru to tail compare
6759 }
6760 // Shift str2 and str1 to the end of the arrays, negate min
6761 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6762 lea(str1, Address(str1, cnt2, scale));
6763 lea(str2, Address(str2, cnt2, scale));
6764 } else {
6765 lea(str1, Address(str1, cnt2, scale1));
6766 lea(str2, Address(str2, cnt2, scale2));
6767 }
6768 decrementl(cnt2); // first character was compared already
6769 negptr(cnt2);
6770
6771 // Compare the rest of the elements
6772 bind(WHILE_HEAD_LABEL);
6773 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae);
6774 subl(result, cnt1);
6775 jccb(Assembler::notZero, POP_LABEL);
6776 increment(cnt2);
6777 jccb(Assembler::notZero, WHILE_HEAD_LABEL);
6778
6779 // Strings are equal up to min length. Return the length difference.
6780 bind(LENGTH_DIFF_LABEL);
6781 pop(result);
6782 if (ae == StrIntrinsicNode::UU) {
6783 // Divide diff by 2 to get number of chars
6784 sarl(result, 1);
6785 }
6786 jmpb(DONE_LABEL);
6787
6788 #ifdef _LP64
6789 if (VM_Version::supports_avx512vlbw()) {
6790
6791 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED);
6792
6793 kmovql(cnt1, k7);
6794 notq(cnt1);
6795 bsfq(cnt2, cnt1);
6796 if (ae != StrIntrinsicNode::LL) {
6797 // Divide diff by 2 to get number of chars
6798 sarl(cnt2, 1);
6799 }
6800 addq(result, cnt2);
6801 if (ae == StrIntrinsicNode::LL) {
6802 load_unsigned_byte(cnt1, Address(str2, result));
6803 load_unsigned_byte(result, Address(str1, result));
6804 } else if (ae == StrIntrinsicNode::UU) {
6805 load_unsigned_short(cnt1, Address(str2, result, scale));
6806 load_unsigned_short(result, Address(str1, result, scale));
6807 } else {
6808 load_unsigned_short(cnt1, Address(str2, result, scale2));
6809 load_unsigned_byte(result, Address(str1, result, scale1));
6810 }
6811 subl(result, cnt1);
6812 jmpb(POP_LABEL);
6813 }//if (VM_Version::supports_avx512vlbw())
6814 #endif // _LP64
6815
6816 // Discard the stored length difference
6817 bind(POP_LABEL);
6818 pop(cnt1);
6819
6820 // That's it
6821 bind(DONE_LABEL);
6822 if(ae == StrIntrinsicNode::UL) {
6823 negl(result);
6824 }
6825
6826 }
6827
6828 // Search for Non-ASCII character (Negative byte value) in a byte array,
6829 // return true if it has any and false otherwise.
6830 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java
6831 // @HotSpotIntrinsicCandidate
6832 // private static boolean hasNegatives(byte[] ba, int off, int len) {
6833 // for (int i = off; i < off + len; i++) {
6834 // if (ba[i] < 0) {
6835 // return true;
6836 // }
6837 // }
6838 // return false;
6839 // }
6840 void MacroAssembler::has_negatives(Register ary1, Register len,
6841 Register result, Register tmp1,
6842 XMMRegister vec1, XMMRegister vec2) {
6843 // rsi: byte array
6844 // rcx: len
6845 // rax: result
6846 ShortBranchVerifier sbv(this);
6847 assert_different_registers(ary1, len, result, tmp1);
6848 assert_different_registers(vec1, vec2);
6849 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE;
6850
6851 // len == 0
6852 testl(len, len);
6853 jcc(Assembler::zero, FALSE_LABEL);
6854
6855 if ((UseAVX > 2) && // AVX512
6856 VM_Version::supports_avx512vlbw() &&
6857 VM_Version::supports_bmi2()) {
6858
6859 Label test_64_loop, test_tail;
6860 Register tmp3_aliased = len;
6861
6862 movl(tmp1, len);
6863 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit);
6864
6865 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F
6866 andl(len, ~(64 - 1)); // vector count (in chars)
6867 jccb(Assembler::zero, test_tail);
6868
6869 lea(ary1, Address(ary1, len, Address::times_1));
6870 negptr(len);
6871
6872 bind(test_64_loop);
6873 // Check whether our 64 elements of size byte contain negatives
6874 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit);
6875 kortestql(k2, k2);
6876 jcc(Assembler::notZero, TRUE_LABEL);
6877
6878 addptr(len, 64);
6879 jccb(Assembler::notZero, test_64_loop);
6880
6881
6882 bind(test_tail);
6883 // bail out when there is nothing to be done
6884 testl(tmp1, -1);
6885 jcc(Assembler::zero, FALSE_LABEL);
6886
6887 // ~(~0 << len) applied up to two times (for 32-bit scenario)
6888 #ifdef _LP64
6889 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF);
6890 shlxq(tmp3_aliased, tmp3_aliased, tmp1);
6891 notq(tmp3_aliased);
6892 kmovql(k3, tmp3_aliased);
6893 #else
6894 Label k_init;
6895 jmp(k_init);
6896
6897 // We could not read 64-bits from a general purpose register thus we move
6898 // data required to compose 64 1's to the instruction stream
6899 // We emit 64 byte wide series of elements from 0..63 which later on would
6900 // be used as a compare targets with tail count contained in tmp1 register.
6901 // Result would be a k register having tmp1 consecutive number or 1
6902 // counting from least significant bit.
6903 address tmp = pc();
6904 emit_int64(0x0706050403020100);
6905 emit_int64(0x0F0E0D0C0B0A0908);
6906 emit_int64(0x1716151413121110);
6907 emit_int64(0x1F1E1D1C1B1A1918);
6908 emit_int64(0x2726252423222120);
6909 emit_int64(0x2F2E2D2C2B2A2928);
6910 emit_int64(0x3736353433323130);
6911 emit_int64(0x3F3E3D3C3B3A3938);
6912
6913 bind(k_init);
6914 lea(len, InternalAddress(tmp));
6915 // create mask to test for negative byte inside a vector
6916 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit);
6917 evpcmpgtb(k3, vec1, Address(len, 0), Assembler::AVX_512bit);
6918
6919 #endif
6920 evpcmpgtb(k2, k3, vec2, Address(ary1, 0), Assembler::AVX_512bit);
6921 ktestq(k2, k3);
6922 jcc(Assembler::notZero, TRUE_LABEL);
6923
6924 jmp(FALSE_LABEL);
6925 } else {
6926 movl(result, len); // copy
6927
6928 if (UseAVX == 2 && UseSSE >= 2) {
6929 // With AVX2, use 32-byte vector compare
6930 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
6931
6932 // Compare 32-byte vectors
6933 andl(result, 0x0000001f); // tail count (in bytes)
6934 andl(len, 0xffffffe0); // vector count (in bytes)
6935 jccb(Assembler::zero, COMPARE_TAIL);
6936
6937 lea(ary1, Address(ary1, len, Address::times_1));
6938 negptr(len);
6939
6940 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector
6941 movdl(vec2, tmp1);
6942 vpbroadcastd(vec2, vec2, Assembler::AVX_256bit);
6943
6944 bind(COMPARE_WIDE_VECTORS);
6945 vmovdqu(vec1, Address(ary1, len, Address::times_1));
6946 vptest(vec1, vec2);
6947 jccb(Assembler::notZero, TRUE_LABEL);
6948 addptr(len, 32);
6949 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
6950
6951 testl(result, result);
6952 jccb(Assembler::zero, FALSE_LABEL);
6953
6954 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
6955 vptest(vec1, vec2);
6956 jccb(Assembler::notZero, TRUE_LABEL);
6957 jmpb(FALSE_LABEL);
6958
6959 bind(COMPARE_TAIL); // len is zero
6960 movl(len, result);
6961 // Fallthru to tail compare
6962 } else if (UseSSE42Intrinsics) {
6963 // With SSE4.2, use double quad vector compare
6964 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
6965
6966 // Compare 16-byte vectors
6967 andl(result, 0x0000000f); // tail count (in bytes)
6968 andl(len, 0xfffffff0); // vector count (in bytes)
6969 jcc(Assembler::zero, COMPARE_TAIL);
6970
6971 lea(ary1, Address(ary1, len, Address::times_1));
6972 negptr(len);
6973
6974 movl(tmp1, 0x80808080);
6975 movdl(vec2, tmp1);
6976 pshufd(vec2, vec2, 0);
6977
6978 bind(COMPARE_WIDE_VECTORS);
6979 movdqu(vec1, Address(ary1, len, Address::times_1));
6980 ptest(vec1, vec2);
6981 jcc(Assembler::notZero, TRUE_LABEL);
6982 addptr(len, 16);
6983 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
6984
6985 testl(result, result);
6986 jcc(Assembler::zero, FALSE_LABEL);
6987
6988 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
6989 ptest(vec1, vec2);
6990 jccb(Assembler::notZero, TRUE_LABEL);
6991 jmpb(FALSE_LABEL);
6992
6993 bind(COMPARE_TAIL); // len is zero
6994 movl(len, result);
6995 // Fallthru to tail compare
6996 }
6997 }
6998 // Compare 4-byte vectors
6999 andl(len, 0xfffffffc); // vector count (in bytes)
7000 jccb(Assembler::zero, COMPARE_CHAR);
7001
7002 lea(ary1, Address(ary1, len, Address::times_1));
7003 negptr(len);
7004
7005 bind(COMPARE_VECTORS);
7006 movl(tmp1, Address(ary1, len, Address::times_1));
7007 andl(tmp1, 0x80808080);
7008 jccb(Assembler::notZero, TRUE_LABEL);
7009 addptr(len, 4);
7010 jcc(Assembler::notZero, COMPARE_VECTORS);
7011
7012 // Compare trailing char (final 2 bytes), if any
7013 bind(COMPARE_CHAR);
7014 testl(result, 0x2); // tail char
7015 jccb(Assembler::zero, COMPARE_BYTE);
7016 load_unsigned_short(tmp1, Address(ary1, 0));
7017 andl(tmp1, 0x00008080);
7018 jccb(Assembler::notZero, TRUE_LABEL);
7019 subptr(result, 2);
7020 lea(ary1, Address(ary1, 2));
7021
7022 bind(COMPARE_BYTE);
7023 testl(result, 0x1); // tail byte
7024 jccb(Assembler::zero, FALSE_LABEL);
7025 load_unsigned_byte(tmp1, Address(ary1, 0));
7026 andl(tmp1, 0x00000080);
7027 jccb(Assembler::notEqual, TRUE_LABEL);
7028 jmpb(FALSE_LABEL);
7029
7030 bind(TRUE_LABEL);
7031 movl(result, 1); // return true
7032 jmpb(DONE);
7033
7034 bind(FALSE_LABEL);
7035 xorl(result, result); // return false
7036
7037 // That's it
7038 bind(DONE);
7039 if (UseAVX >= 2 && UseSSE >= 2) {
7040 // clean upper bits of YMM registers
7041 vpxor(vec1, vec1);
7042 vpxor(vec2, vec2);
7043 }
7044 }
7045 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings.
7046 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2,
7047 Register limit, Register result, Register chr,
7048 XMMRegister vec1, XMMRegister vec2, bool is_char) {
7049 ShortBranchVerifier sbv(this);
7050 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE;
7051
7052 int length_offset = arrayOopDesc::length_offset_in_bytes();
7053 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE);
7054
7055 if (is_array_equ) {
7056 // Check the input args
7057 cmpoop(ary1, ary2);
7058 jcc(Assembler::equal, TRUE_LABEL);
7059
7060 // Need additional checks for arrays_equals.
7061 testptr(ary1, ary1);
7062 jcc(Assembler::zero, FALSE_LABEL);
7063 testptr(ary2, ary2);
7064 jcc(Assembler::zero, FALSE_LABEL);
7065
7066 // Check the lengths
7067 movl(limit, Address(ary1, length_offset));
7068 cmpl(limit, Address(ary2, length_offset));
7069 jcc(Assembler::notEqual, FALSE_LABEL);
7070 }
7071
7072 // count == 0
7073 testl(limit, limit);
7074 jcc(Assembler::zero, TRUE_LABEL);
7075
7076 if (is_array_equ) {
7077 // Load array address
7078 lea(ary1, Address(ary1, base_offset));
7079 lea(ary2, Address(ary2, base_offset));
7080 }
7081
7082 if (is_array_equ && is_char) {
7083 // arrays_equals when used for char[].
7084 shll(limit, 1); // byte count != 0
7085 }
7086 movl(result, limit); // copy
7087
7088 if (UseAVX >= 2) {
7089 // With AVX2, use 32-byte vector compare
7090 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7091
7092 // Compare 32-byte vectors
7093 andl(result, 0x0000001f); // tail count (in bytes)
7094 andl(limit, 0xffffffe0); // vector count (in bytes)
7095 jcc(Assembler::zero, COMPARE_TAIL);
7096
7097 lea(ary1, Address(ary1, limit, Address::times_1));
7098 lea(ary2, Address(ary2, limit, Address::times_1));
7099 negptr(limit);
7100
7101 bind(COMPARE_WIDE_VECTORS);
7102
7103 #ifdef _LP64
7104 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
7105 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3;
7106
7107 cmpl(limit, -64);
7108 jccb(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2);
7109
7110 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
7111
7112 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit);
7113 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit);
7114 kortestql(k7, k7);
7115 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
7116 addptr(limit, 64); // update since we already compared at this addr
7117 cmpl(limit, -64);
7118 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3);
7119
7120 // At this point we may still need to compare -limit+result bytes.
7121 // We could execute the next two instruction and just continue via non-wide path:
7122 // cmpl(limit, 0);
7123 // jcc(Assembler::equal, COMPARE_TAIL); // true
7124 // But since we stopped at the points ary{1,2}+limit which are
7125 // not farther than 64 bytes from the ends of arrays ary{1,2}+result
7126 // (|limit| <= 32 and result < 32),
7127 // we may just compare the last 64 bytes.
7128 //
7129 addptr(result, -64); // it is safe, bc we just came from this area
7130 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit);
7131 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit);
7132 kortestql(k7, k7);
7133 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
7134
7135 jmp(TRUE_LABEL);
7136
7137 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
7138
7139 }//if (VM_Version::supports_avx512vlbw())
7140 #endif //_LP64
7141
7142 vmovdqu(vec1, Address(ary1, limit, Address::times_1));
7143 vmovdqu(vec2, Address(ary2, limit, Address::times_1));
7144 vpxor(vec1, vec2);
7145
7146 vptest(vec1, vec1);
7147 jcc(Assembler::notZero, FALSE_LABEL);
7148 addptr(limit, 32);
7149 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7150
7151 testl(result, result);
7152 jcc(Assembler::zero, TRUE_LABEL);
7153
7154 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
7155 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
7156 vpxor(vec1, vec2);
7157
7158 vptest(vec1, vec1);
7159 jccb(Assembler::notZero, FALSE_LABEL);
7160 jmpb(TRUE_LABEL);
7161
7162 bind(COMPARE_TAIL); // limit is zero
7163 movl(limit, result);
7164 // Fallthru to tail compare
7165 } else if (UseSSE42Intrinsics) {
7166 // With SSE4.2, use double quad vector compare
7167 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7168
7169 // Compare 16-byte vectors
7170 andl(result, 0x0000000f); // tail count (in bytes)
7171 andl(limit, 0xfffffff0); // vector count (in bytes)
7172 jcc(Assembler::zero, COMPARE_TAIL);
7173
7174 lea(ary1, Address(ary1, limit, Address::times_1));
7175 lea(ary2, Address(ary2, limit, Address::times_1));
7176 negptr(limit);
7177
7178 bind(COMPARE_WIDE_VECTORS);
7179 movdqu(vec1, Address(ary1, limit, Address::times_1));
7180 movdqu(vec2, Address(ary2, limit, Address::times_1));
7181 pxor(vec1, vec2);
7182
7183 ptest(vec1, vec1);
7184 jcc(Assembler::notZero, FALSE_LABEL);
7185 addptr(limit, 16);
7186 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7187
7188 testl(result, result);
7189 jcc(Assembler::zero, TRUE_LABEL);
7190
7191 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
7192 movdqu(vec2, Address(ary2, result, Address::times_1, -16));
7193 pxor(vec1, vec2);
7194
7195 ptest(vec1, vec1);
7196 jccb(Assembler::notZero, FALSE_LABEL);
7197 jmpb(TRUE_LABEL);
7198
7199 bind(COMPARE_TAIL); // limit is zero
7200 movl(limit, result);
7201 // Fallthru to tail compare
7202 }
7203
7204 // Compare 4-byte vectors
7205 andl(limit, 0xfffffffc); // vector count (in bytes)
7206 jccb(Assembler::zero, COMPARE_CHAR);
7207
7208 lea(ary1, Address(ary1, limit, Address::times_1));
7209 lea(ary2, Address(ary2, limit, Address::times_1));
7210 negptr(limit);
7211
7212 bind(COMPARE_VECTORS);
7213 movl(chr, Address(ary1, limit, Address::times_1));
7214 cmpl(chr, Address(ary2, limit, Address::times_1));
7215 jccb(Assembler::notEqual, FALSE_LABEL);
7216 addptr(limit, 4);
7217 jcc(Assembler::notZero, COMPARE_VECTORS);
7218
7219 // Compare trailing char (final 2 bytes), if any
7220 bind(COMPARE_CHAR);
7221 testl(result, 0x2); // tail char
7222 jccb(Assembler::zero, COMPARE_BYTE);
7223 load_unsigned_short(chr, Address(ary1, 0));
7224 load_unsigned_short(limit, Address(ary2, 0));
7225 cmpl(chr, limit);
7226 jccb(Assembler::notEqual, FALSE_LABEL);
7227
7228 if (is_array_equ && is_char) {
7229 bind(COMPARE_BYTE);
7230 } else {
7231 lea(ary1, Address(ary1, 2));
7232 lea(ary2, Address(ary2, 2));
7233
7234 bind(COMPARE_BYTE);
7235 testl(result, 0x1); // tail byte
7236 jccb(Assembler::zero, TRUE_LABEL);
7237 load_unsigned_byte(chr, Address(ary1, 0));
7238 load_unsigned_byte(limit, Address(ary2, 0));
7239 cmpl(chr, limit);
7240 jccb(Assembler::notEqual, FALSE_LABEL);
7241 }
7242 bind(TRUE_LABEL);
7243 movl(result, 1); // return true
7244 jmpb(DONE);
7245
7246 bind(FALSE_LABEL);
7247 xorl(result, result); // return false
7248
7249 // That's it
7250 bind(DONE);
7251 if (UseAVX >= 2) {
7252 // clean upper bits of YMM registers
7253 vpxor(vec1, vec1);
7254 vpxor(vec2, vec2);
7255 }
7256 }
7257
7258 #endif
7259
7260 void MacroAssembler::generate_fill(BasicType t, bool aligned,
7261 Register to, Register value, Register count,
7262 Register rtmp, XMMRegister xtmp) {
7263 ShortBranchVerifier sbv(this);
7264 assert_different_registers(to, value, count, rtmp);
7265 Label L_exit;
7266 Label L_fill_2_bytes, L_fill_4_bytes;
7267
7268 int shift = -1;
7269 switch (t) {
7270 case T_BYTE:
7271 shift = 2;
7272 break;
7273 case T_SHORT:
7274 shift = 1;
7275 break;
7276 case T_INT:
7277 shift = 0;
7278 break;
7279 default: ShouldNotReachHere();
7280 }
7281
7282 if (t == T_BYTE) {
7283 andl(value, 0xff);
7284 movl(rtmp, value);
7285 shll(rtmp, 8);
7286 orl(value, rtmp);
7287 }
7288 if (t == T_SHORT) {
7289 andl(value, 0xffff);
7290 }
7291 if (t == T_BYTE || t == T_SHORT) {
7292 movl(rtmp, value);
7293 shll(rtmp, 16);
7294 orl(value, rtmp);
7295 }
7296
7297 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
7298 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp
7299 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) {
7300 Label L_skip_align2;
7301 // align source address at 4 bytes address boundary
7302 if (t == T_BYTE) {
7303 Label L_skip_align1;
7304 // One byte misalignment happens only for byte arrays
7305 testptr(to, 1);
7306 jccb(Assembler::zero, L_skip_align1);
7307 movb(Address(to, 0), value);
7308 increment(to);
7309 decrement(count);
7310 BIND(L_skip_align1);
7311 }
7312 // Two bytes misalignment happens only for byte and short (char) arrays
7313 testptr(to, 2);
7314 jccb(Assembler::zero, L_skip_align2);
7315 movw(Address(to, 0), value);
7316 addptr(to, 2);
7317 subl(count, 1<<(shift-1));
7318 BIND(L_skip_align2);
7319 }
7320 if (UseSSE < 2) {
7321 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7322 // Fill 32-byte chunks
7323 subl(count, 8 << shift);
7324 jcc(Assembler::less, L_check_fill_8_bytes);
7325 align(16);
7326
7327 BIND(L_fill_32_bytes_loop);
7328
7329 for (int i = 0; i < 32; i += 4) {
7330 movl(Address(to, i), value);
7331 }
7332
7333 addptr(to, 32);
7334 subl(count, 8 << shift);
7335 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7336 BIND(L_check_fill_8_bytes);
7337 addl(count, 8 << shift);
7338 jccb(Assembler::zero, L_exit);
7339 jmpb(L_fill_8_bytes);
7340
7341 //
7342 // length is too short, just fill qwords
7343 //
7344 BIND(L_fill_8_bytes_loop);
7345 movl(Address(to, 0), value);
7346 movl(Address(to, 4), value);
7347 addptr(to, 8);
7348 BIND(L_fill_8_bytes);
7349 subl(count, 1 << (shift + 1));
7350 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7351 // fall through to fill 4 bytes
7352 } else {
7353 Label L_fill_32_bytes;
7354 if (!UseUnalignedLoadStores) {
7355 // align to 8 bytes, we know we are 4 byte aligned to start
7356 testptr(to, 4);
7357 jccb(Assembler::zero, L_fill_32_bytes);
7358 movl(Address(to, 0), value);
7359 addptr(to, 4);
7360 subl(count, 1<<shift);
7361 }
7362 BIND(L_fill_32_bytes);
7363 {
7364 assert( UseSSE >= 2, "supported cpu only" );
7365 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7366 movdl(xtmp, value);
7367 if (UseAVX > 2 && UseUnalignedLoadStores) {
7368 // Fill 64-byte chunks
7369 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
7370 vpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit);
7371
7372 subl(count, 16 << shift);
7373 jcc(Assembler::less, L_check_fill_32_bytes);
7374 align(16);
7375
7376 BIND(L_fill_64_bytes_loop);
7377 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit);
7378 addptr(to, 64);
7379 subl(count, 16 << shift);
7380 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
7381
7382 BIND(L_check_fill_32_bytes);
7383 addl(count, 8 << shift);
7384 jccb(Assembler::less, L_check_fill_8_bytes);
7385 vmovdqu(Address(to, 0), xtmp);
7386 addptr(to, 32);
7387 subl(count, 8 << shift);
7388
7389 BIND(L_check_fill_8_bytes);
7390 } else if (UseAVX == 2 && UseUnalignedLoadStores) {
7391 // Fill 64-byte chunks
7392 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
7393 vpbroadcastd(xtmp, xtmp, Assembler::AVX_256bit);
7394
7395 subl(count, 16 << shift);
7396 jcc(Assembler::less, L_check_fill_32_bytes);
7397 align(16);
7398
7399 BIND(L_fill_64_bytes_loop);
7400 vmovdqu(Address(to, 0), xtmp);
7401 vmovdqu(Address(to, 32), xtmp);
7402 addptr(to, 64);
7403 subl(count, 16 << shift);
7404 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
7405
7406 BIND(L_check_fill_32_bytes);
7407 addl(count, 8 << shift);
7408 jccb(Assembler::less, L_check_fill_8_bytes);
7409 vmovdqu(Address(to, 0), xtmp);
7410 addptr(to, 32);
7411 subl(count, 8 << shift);
7412
7413 BIND(L_check_fill_8_bytes);
7414 // clean upper bits of YMM registers
7415 movdl(xtmp, value);
7416 pshufd(xtmp, xtmp, 0);
7417 } else {
7418 // Fill 32-byte chunks
7419 pshufd(xtmp, xtmp, 0);
7420
7421 subl(count, 8 << shift);
7422 jcc(Assembler::less, L_check_fill_8_bytes);
7423 align(16);
7424
7425 BIND(L_fill_32_bytes_loop);
7426
7427 if (UseUnalignedLoadStores) {
7428 movdqu(Address(to, 0), xtmp);
7429 movdqu(Address(to, 16), xtmp);
7430 } else {
7431 movq(Address(to, 0), xtmp);
7432 movq(Address(to, 8), xtmp);
7433 movq(Address(to, 16), xtmp);
7434 movq(Address(to, 24), xtmp);
7435 }
7436
7437 addptr(to, 32);
7438 subl(count, 8 << shift);
7439 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7440
7441 BIND(L_check_fill_8_bytes);
7442 }
7443 addl(count, 8 << shift);
7444 jccb(Assembler::zero, L_exit);
7445 jmpb(L_fill_8_bytes);
7446
7447 //
7448 // length is too short, just fill qwords
7449 //
7450 BIND(L_fill_8_bytes_loop);
7451 movq(Address(to, 0), xtmp);
7452 addptr(to, 8);
7453 BIND(L_fill_8_bytes);
7454 subl(count, 1 << (shift + 1));
7455 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7456 }
7457 }
7458 // fill trailing 4 bytes
7459 BIND(L_fill_4_bytes);
7460 testl(count, 1<<shift);
7461 jccb(Assembler::zero, L_fill_2_bytes);
7462 movl(Address(to, 0), value);
7463 if (t == T_BYTE || t == T_SHORT) {
7464 Label L_fill_byte;
7465 addptr(to, 4);
7466 BIND(L_fill_2_bytes);
7467 // fill trailing 2 bytes
7468 testl(count, 1<<(shift-1));
7469 jccb(Assembler::zero, L_fill_byte);
7470 movw(Address(to, 0), value);
7471 if (t == T_BYTE) {
7472 addptr(to, 2);
7473 BIND(L_fill_byte);
7474 // fill trailing byte
7475 testl(count, 1);
7476 jccb(Assembler::zero, L_exit);
7477 movb(Address(to, 0), value);
7478 } else {
7479 BIND(L_fill_byte);
7480 }
7481 } else {
7482 BIND(L_fill_2_bytes);
7483 }
7484 BIND(L_exit);
7485 }
7486
7487 // encode char[] to byte[] in ISO_8859_1
7488 //@HotSpotIntrinsicCandidate
7489 //private static int implEncodeISOArray(byte[] sa, int sp,
7490 //byte[] da, int dp, int len) {
7491 // int i = 0;
7492 // for (; i < len; i++) {
7493 // char c = StringUTF16.getChar(sa, sp++);
7494 // if (c > '\u00FF')
7495 // break;
7496 // da[dp++] = (byte)c;
7497 // }
7498 // return i;
7499 //}
7500 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len,
7501 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
7502 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
7503 Register tmp5, Register result) {
7504
7505 // rsi: src
7506 // rdi: dst
7507 // rdx: len
7508 // rcx: tmp5
7509 // rax: result
7510 ShortBranchVerifier sbv(this);
7511 assert_different_registers(src, dst, len, tmp5, result);
7512 Label L_done, L_copy_1_char, L_copy_1_char_exit;
7513
7514 // set result
7515 xorl(result, result);
7516 // check for zero length
7517 testl(len, len);
7518 jcc(Assembler::zero, L_done);
7519
7520 movl(result, len);
7521
7522 // Setup pointers
7523 lea(src, Address(src, len, Address::times_2)); // char[]
7524 lea(dst, Address(dst, len, Address::times_1)); // byte[]
7525 negptr(len);
7526
7527 if (UseSSE42Intrinsics || UseAVX >= 2) {
7528 Label L_copy_8_chars, L_copy_8_chars_exit;
7529 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit;
7530
7531 if (UseAVX >= 2) {
7532 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit;
7533 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
7534 movdl(tmp1Reg, tmp5);
7535 vpbroadcastd(tmp1Reg, tmp1Reg, Assembler::AVX_256bit);
7536 jmp(L_chars_32_check);
7537
7538 bind(L_copy_32_chars);
7539 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64));
7540 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32));
7541 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
7542 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
7543 jccb(Assembler::notZero, L_copy_32_chars_exit);
7544 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
7545 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1);
7546 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg);
7547
7548 bind(L_chars_32_check);
7549 addptr(len, 32);
7550 jcc(Assembler::lessEqual, L_copy_32_chars);
7551
7552 bind(L_copy_32_chars_exit);
7553 subptr(len, 16);
7554 jccb(Assembler::greater, L_copy_16_chars_exit);
7555
7556 } else if (UseSSE42Intrinsics) {
7557 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
7558 movdl(tmp1Reg, tmp5);
7559 pshufd(tmp1Reg, tmp1Reg, 0);
7560 jmpb(L_chars_16_check);
7561 }
7562
7563 bind(L_copy_16_chars);
7564 if (UseAVX >= 2) {
7565 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32));
7566 vptest(tmp2Reg, tmp1Reg);
7567 jcc(Assembler::notZero, L_copy_16_chars_exit);
7568 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1);
7569 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1);
7570 } else {
7571 if (UseAVX > 0) {
7572 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7573 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7574 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0);
7575 } else {
7576 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7577 por(tmp2Reg, tmp3Reg);
7578 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7579 por(tmp2Reg, tmp4Reg);
7580 }
7581 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
7582 jccb(Assembler::notZero, L_copy_16_chars_exit);
7583 packuswb(tmp3Reg, tmp4Reg);
7584 }
7585 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg);
7586
7587 bind(L_chars_16_check);
7588 addptr(len, 16);
7589 jcc(Assembler::lessEqual, L_copy_16_chars);
7590
7591 bind(L_copy_16_chars_exit);
7592 if (UseAVX >= 2) {
7593 // clean upper bits of YMM registers
7594 vpxor(tmp2Reg, tmp2Reg);
7595 vpxor(tmp3Reg, tmp3Reg);
7596 vpxor(tmp4Reg, tmp4Reg);
7597 movdl(tmp1Reg, tmp5);
7598 pshufd(tmp1Reg, tmp1Reg, 0);
7599 }
7600 subptr(len, 8);
7601 jccb(Assembler::greater, L_copy_8_chars_exit);
7602
7603 bind(L_copy_8_chars);
7604 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16));
7605 ptest(tmp3Reg, tmp1Reg);
7606 jccb(Assembler::notZero, L_copy_8_chars_exit);
7607 packuswb(tmp3Reg, tmp1Reg);
7608 movq(Address(dst, len, Address::times_1, -8), tmp3Reg);
7609 addptr(len, 8);
7610 jccb(Assembler::lessEqual, L_copy_8_chars);
7611
7612 bind(L_copy_8_chars_exit);
7613 subptr(len, 8);
7614 jccb(Assembler::zero, L_done);
7615 }
7616
7617 bind(L_copy_1_char);
7618 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0));
7619 testl(tmp5, 0xff00); // check if Unicode char
7620 jccb(Assembler::notZero, L_copy_1_char_exit);
7621 movb(Address(dst, len, Address::times_1, 0), tmp5);
7622 addptr(len, 1);
7623 jccb(Assembler::less, L_copy_1_char);
7624
7625 bind(L_copy_1_char_exit);
7626 addptr(result, len); // len is negative count of not processed elements
7627
7628 bind(L_done);
7629 }
7630
7631 #ifdef _LP64
7632 /**
7633 * Helper for multiply_to_len().
7634 */
7635 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) {
7636 addq(dest_lo, src1);
7637 adcq(dest_hi, 0);
7638 addq(dest_lo, src2);
7639 adcq(dest_hi, 0);
7640 }
7641
7642 /**
7643 * Multiply 64 bit by 64 bit first loop.
7644 */
7645 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
7646 Register y, Register y_idx, Register z,
7647 Register carry, Register product,
7648 Register idx, Register kdx) {
7649 //
7650 // jlong carry, x[], y[], z[];
7651 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7652 // huge_128 product = y[idx] * x[xstart] + carry;
7653 // z[kdx] = (jlong)product;
7654 // carry = (jlong)(product >>> 64);
7655 // }
7656 // z[xstart] = carry;
7657 //
7658
7659 Label L_first_loop, L_first_loop_exit;
7660 Label L_one_x, L_one_y, L_multiply;
7661
7662 decrementl(xstart);
7663 jcc(Assembler::negative, L_one_x);
7664
7665 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
7666 rorq(x_xstart, 32); // convert big-endian to little-endian
7667
7668 bind(L_first_loop);
7669 decrementl(idx);
7670 jcc(Assembler::negative, L_first_loop_exit);
7671 decrementl(idx);
7672 jcc(Assembler::negative, L_one_y);
7673 movq(y_idx, Address(y, idx, Address::times_4, 0));
7674 rorq(y_idx, 32); // convert big-endian to little-endian
7675 bind(L_multiply);
7676 movq(product, x_xstart);
7677 mulq(y_idx); // product(rax) * y_idx -> rdx:rax
7678 addq(product, carry);
7679 adcq(rdx, 0);
7680 subl(kdx, 2);
7681 movl(Address(z, kdx, Address::times_4, 4), product);
7682 shrq(product, 32);
7683 movl(Address(z, kdx, Address::times_4, 0), product);
7684 movq(carry, rdx);
7685 jmp(L_first_loop);
7686
7687 bind(L_one_y);
7688 movl(y_idx, Address(y, 0));
7689 jmp(L_multiply);
7690
7691 bind(L_one_x);
7692 movl(x_xstart, Address(x, 0));
7693 jmp(L_first_loop);
7694
7695 bind(L_first_loop_exit);
7696 }
7697
7698 /**
7699 * Multiply 64 bit by 64 bit and add 128 bit.
7700 */
7701 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z,
7702 Register yz_idx, Register idx,
7703 Register carry, Register product, int offset) {
7704 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
7705 // z[kdx] = (jlong)product;
7706
7707 movq(yz_idx, Address(y, idx, Address::times_4, offset));
7708 rorq(yz_idx, 32); // convert big-endian to little-endian
7709 movq(product, x_xstart);
7710 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7711 movq(yz_idx, Address(z, idx, Address::times_4, offset));
7712 rorq(yz_idx, 32); // convert big-endian to little-endian
7713
7714 add2_with_carry(rdx, product, carry, yz_idx);
7715
7716 movl(Address(z, idx, Address::times_4, offset+4), product);
7717 shrq(product, 32);
7718 movl(Address(z, idx, Address::times_4, offset), product);
7719
7720 }
7721
7722 /**
7723 * Multiply 128 bit by 128 bit. Unrolled inner loop.
7724 */
7725 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z,
7726 Register yz_idx, Register idx, Register jdx,
7727 Register carry, Register product,
7728 Register carry2) {
7729 // jlong carry, x[], y[], z[];
7730 // int kdx = ystart+1;
7731 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7732 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
7733 // z[kdx+idx+1] = (jlong)product;
7734 // jlong carry2 = (jlong)(product >>> 64);
7735 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
7736 // z[kdx+idx] = (jlong)product;
7737 // carry = (jlong)(product >>> 64);
7738 // }
7739 // idx += 2;
7740 // if (idx > 0) {
7741 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
7742 // z[kdx+idx] = (jlong)product;
7743 // carry = (jlong)(product >>> 64);
7744 // }
7745 //
7746
7747 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7748
7749 movl(jdx, idx);
7750 andl(jdx, 0xFFFFFFFC);
7751 shrl(jdx, 2);
7752
7753 bind(L_third_loop);
7754 subl(jdx, 1);
7755 jcc(Assembler::negative, L_third_loop_exit);
7756 subl(idx, 4);
7757
7758 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
7759 movq(carry2, rdx);
7760
7761 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
7762 movq(carry, rdx);
7763 jmp(L_third_loop);
7764
7765 bind (L_third_loop_exit);
7766
7767 andl (idx, 0x3);
7768 jcc(Assembler::zero, L_post_third_loop_done);
7769
7770 Label L_check_1;
7771 subl(idx, 2);
7772 jcc(Assembler::negative, L_check_1);
7773
7774 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
7775 movq(carry, rdx);
7776
7777 bind (L_check_1);
7778 addl (idx, 0x2);
7779 andl (idx, 0x1);
7780 subl(idx, 1);
7781 jcc(Assembler::negative, L_post_third_loop_done);
7782
7783 movl(yz_idx, Address(y, idx, Address::times_4, 0));
7784 movq(product, x_xstart);
7785 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7786 movl(yz_idx, Address(z, idx, Address::times_4, 0));
7787
7788 add2_with_carry(rdx, product, yz_idx, carry);
7789
7790 movl(Address(z, idx, Address::times_4, 0), product);
7791 shrq(product, 32);
7792
7793 shlq(rdx, 32);
7794 orq(product, rdx);
7795 movq(carry, product);
7796
7797 bind(L_post_third_loop_done);
7798 }
7799
7800 /**
7801 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop.
7802 *
7803 */
7804 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z,
7805 Register carry, Register carry2,
7806 Register idx, Register jdx,
7807 Register yz_idx1, Register yz_idx2,
7808 Register tmp, Register tmp3, Register tmp4) {
7809 assert(UseBMI2Instructions, "should be used only when BMI2 is available");
7810
7811 // jlong carry, x[], y[], z[];
7812 // int kdx = ystart+1;
7813 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7814 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry;
7815 // jlong carry2 = (jlong)(tmp3 >>> 64);
7816 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2;
7817 // carry = (jlong)(tmp4 >>> 64);
7818 // z[kdx+idx+1] = (jlong)tmp3;
7819 // z[kdx+idx] = (jlong)tmp4;
7820 // }
7821 // idx += 2;
7822 // if (idx > 0) {
7823 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry;
7824 // z[kdx+idx] = (jlong)yz_idx1;
7825 // carry = (jlong)(yz_idx1 >>> 64);
7826 // }
7827 //
7828
7829 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7830
7831 movl(jdx, idx);
7832 andl(jdx, 0xFFFFFFFC);
7833 shrl(jdx, 2);
7834
7835 bind(L_third_loop);
7836 subl(jdx, 1);
7837 jcc(Assembler::negative, L_third_loop_exit);
7838 subl(idx, 4);
7839
7840 movq(yz_idx1, Address(y, idx, Address::times_4, 8));
7841 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
7842 movq(yz_idx2, Address(y, idx, Address::times_4, 0));
7843 rorxq(yz_idx2, yz_idx2, 32);
7844
7845 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
7846 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp
7847
7848 movq(yz_idx1, Address(z, idx, Address::times_4, 8));
7849 rorxq(yz_idx1, yz_idx1, 32);
7850 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
7851 rorxq(yz_idx2, yz_idx2, 32);
7852
7853 if (VM_Version::supports_adx()) {
7854 adcxq(tmp3, carry);
7855 adoxq(tmp3, yz_idx1);
7856
7857 adcxq(tmp4, tmp);
7858 adoxq(tmp4, yz_idx2);
7859
7860 movl(carry, 0); // does not affect flags
7861 adcxq(carry2, carry);
7862 adoxq(carry2, carry);
7863 } else {
7864 add2_with_carry(tmp4, tmp3, carry, yz_idx1);
7865 add2_with_carry(carry2, tmp4, tmp, yz_idx2);
7866 }
7867 movq(carry, carry2);
7868
7869 movl(Address(z, idx, Address::times_4, 12), tmp3);
7870 shrq(tmp3, 32);
7871 movl(Address(z, idx, Address::times_4, 8), tmp3);
7872
7873 movl(Address(z, idx, Address::times_4, 4), tmp4);
7874 shrq(tmp4, 32);
7875 movl(Address(z, idx, Address::times_4, 0), tmp4);
7876
7877 jmp(L_third_loop);
7878
7879 bind (L_third_loop_exit);
7880
7881 andl (idx, 0x3);
7882 jcc(Assembler::zero, L_post_third_loop_done);
7883
7884 Label L_check_1;
7885 subl(idx, 2);
7886 jcc(Assembler::negative, L_check_1);
7887
7888 movq(yz_idx1, Address(y, idx, Address::times_4, 0));
7889 rorxq(yz_idx1, yz_idx1, 32);
7890 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
7891 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
7892 rorxq(yz_idx2, yz_idx2, 32);
7893
7894 add2_with_carry(tmp4, tmp3, carry, yz_idx2);
7895
7896 movl(Address(z, idx, Address::times_4, 4), tmp3);
7897 shrq(tmp3, 32);
7898 movl(Address(z, idx, Address::times_4, 0), tmp3);
7899 movq(carry, tmp4);
7900
7901 bind (L_check_1);
7902 addl (idx, 0x2);
7903 andl (idx, 0x1);
7904 subl(idx, 1);
7905 jcc(Assembler::negative, L_post_third_loop_done);
7906 movl(tmp4, Address(y, idx, Address::times_4, 0));
7907 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3
7908 movl(tmp4, Address(z, idx, Address::times_4, 0));
7909
7910 add2_with_carry(carry2, tmp3, tmp4, carry);
7911
7912 movl(Address(z, idx, Address::times_4, 0), tmp3);
7913 shrq(tmp3, 32);
7914
7915 shlq(carry2, 32);
7916 orq(tmp3, carry2);
7917 movq(carry, tmp3);
7918
7919 bind(L_post_third_loop_done);
7920 }
7921
7922 /**
7923 * Code for BigInteger::multiplyToLen() instrinsic.
7924 *
7925 * rdi: x
7926 * rax: xlen
7927 * rsi: y
7928 * rcx: ylen
7929 * r8: z
7930 * r11: zlen
7931 * r12: tmp1
7932 * r13: tmp2
7933 * r14: tmp3
7934 * r15: tmp4
7935 * rbx: tmp5
7936 *
7937 */
7938 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen,
7939 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) {
7940 ShortBranchVerifier sbv(this);
7941 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx);
7942
7943 push(tmp1);
7944 push(tmp2);
7945 push(tmp3);
7946 push(tmp4);
7947 push(tmp5);
7948
7949 push(xlen);
7950 push(zlen);
7951
7952 const Register idx = tmp1;
7953 const Register kdx = tmp2;
7954 const Register xstart = tmp3;
7955
7956 const Register y_idx = tmp4;
7957 const Register carry = tmp5;
7958 const Register product = xlen;
7959 const Register x_xstart = zlen; // reuse register
7960
7961 // First Loop.
7962 //
7963 // final static long LONG_MASK = 0xffffffffL;
7964 // int xstart = xlen - 1;
7965 // int ystart = ylen - 1;
7966 // long carry = 0;
7967 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7968 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
7969 // z[kdx] = (int)product;
7970 // carry = product >>> 32;
7971 // }
7972 // z[xstart] = (int)carry;
7973 //
7974
7975 movl(idx, ylen); // idx = ylen;
7976 movl(kdx, zlen); // kdx = xlen+ylen;
7977 xorq(carry, carry); // carry = 0;
7978
7979 Label L_done;
7980
7981 movl(xstart, xlen);
7982 decrementl(xstart);
7983 jcc(Assembler::negative, L_done);
7984
7985 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
7986
7987 Label L_second_loop;
7988 testl(kdx, kdx);
7989 jcc(Assembler::zero, L_second_loop);
7990
7991 Label L_carry;
7992 subl(kdx, 1);
7993 jcc(Assembler::zero, L_carry);
7994
7995 movl(Address(z, kdx, Address::times_4, 0), carry);
7996 shrq(carry, 32);
7997 subl(kdx, 1);
7998
7999 bind(L_carry);
8000 movl(Address(z, kdx, Address::times_4, 0), carry);
8001
8002 // Second and third (nested) loops.
8003 //
8004 // for (int i = xstart-1; i >= 0; i--) { // Second loop
8005 // carry = 0;
8006 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
8007 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
8008 // (z[k] & LONG_MASK) + carry;
8009 // z[k] = (int)product;
8010 // carry = product >>> 32;
8011 // }
8012 // z[i] = (int)carry;
8013 // }
8014 //
8015 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
8016
8017 const Register jdx = tmp1;
8018
8019 bind(L_second_loop);
8020 xorl(carry, carry); // carry = 0;
8021 movl(jdx, ylen); // j = ystart+1
8022
8023 subl(xstart, 1); // i = xstart-1;
8024 jcc(Assembler::negative, L_done);
8025
8026 push (z);
8027
8028 Label L_last_x;
8029 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j
8030 subl(xstart, 1); // i = xstart-1;
8031 jcc(Assembler::negative, L_last_x);
8032
8033 if (UseBMI2Instructions) {
8034 movq(rdx, Address(x, xstart, Address::times_4, 0));
8035 rorxq(rdx, rdx, 32); // convert big-endian to little-endian
8036 } else {
8037 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
8038 rorq(x_xstart, 32); // convert big-endian to little-endian
8039 }
8040
8041 Label L_third_loop_prologue;
8042 bind(L_third_loop_prologue);
8043
8044 push (x);
8045 push (xstart);
8046 push (ylen);
8047
8048
8049 if (UseBMI2Instructions) {
8050 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4);
8051 } else { // !UseBMI2Instructions
8052 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
8053 }
8054
8055 pop(ylen);
8056 pop(xlen);
8057 pop(x);
8058 pop(z);
8059
8060 movl(tmp3, xlen);
8061 addl(tmp3, 1);
8062 movl(Address(z, tmp3, Address::times_4, 0), carry);
8063 subl(tmp3, 1);
8064 jccb(Assembler::negative, L_done);
8065
8066 shrq(carry, 32);
8067 movl(Address(z, tmp3, Address::times_4, 0), carry);
8068 jmp(L_second_loop);
8069
8070 // Next infrequent code is moved outside loops.
8071 bind(L_last_x);
8072 if (UseBMI2Instructions) {
8073 movl(rdx, Address(x, 0));
8074 } else {
8075 movl(x_xstart, Address(x, 0));
8076 }
8077 jmp(L_third_loop_prologue);
8078
8079 bind(L_done);
8080
8081 pop(zlen);
8082 pop(xlen);
8083
8084 pop(tmp5);
8085 pop(tmp4);
8086 pop(tmp3);
8087 pop(tmp2);
8088 pop(tmp1);
8089 }
8090
8091 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale,
8092 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){
8093 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled.");
8094 Label VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP;
8095 Label VECTOR8_TAIL, VECTOR4_TAIL;
8096 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL;
8097 Label SAME_TILL_END, DONE;
8098 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL;
8099
8100 //scale is in rcx in both Win64 and Unix
8101 ShortBranchVerifier sbv(this);
8102
8103 shlq(length);
8104 xorq(result, result);
8105
8106 if ((UseAVX > 2) &&
8107 VM_Version::supports_avx512vlbw()) {
8108 Label VECTOR64_LOOP, VECTOR64_NOT_EQUAL, VECTOR32_TAIL;
8109
8110 cmpq(length, 64);
8111 jcc(Assembler::less, VECTOR32_TAIL);
8112 movq(tmp1, length);
8113 andq(tmp1, 0x3F); // tail count
8114 andq(length, ~(0x3F)); //vector count
8115
8116 bind(VECTOR64_LOOP);
8117 // AVX512 code to compare 64 byte vectors.
8118 evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit);
8119 evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit);
8120 kortestql(k7, k7);
8121 jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL); // mismatch
8122 addq(result, 64);
8123 subq(length, 64);
8124 jccb(Assembler::notZero, VECTOR64_LOOP);
8125
8126 //bind(VECTOR64_TAIL);
8127 testq(tmp1, tmp1);
8128 jcc(Assembler::zero, SAME_TILL_END);
8129
8130 //bind(VECTOR64_TAIL);
8131 // AVX512 code to compare upto 63 byte vectors.
8132 mov64(tmp2, 0xFFFFFFFFFFFFFFFF);
8133 shlxq(tmp2, tmp2, tmp1);
8134 notq(tmp2);
8135 kmovql(k3, tmp2);
8136
8137 evmovdqub(rymm0, k3, Address(obja, result), Assembler::AVX_512bit);
8138 evpcmpeqb(k7, k3, rymm0, Address(objb, result), Assembler::AVX_512bit);
8139
8140 ktestql(k7, k3);
8141 jcc(Assembler::below, SAME_TILL_END); // not mismatch
8142
8143 bind(VECTOR64_NOT_EQUAL);
8144 kmovql(tmp1, k7);
8145 notq(tmp1);
8146 tzcntq(tmp1, tmp1);
8147 addq(result, tmp1);
8148 shrq(result);
8149 jmp(DONE);
8150 bind(VECTOR32_TAIL);
8151 }
8152
8153 cmpq(length, 8);
8154 jcc(Assembler::equal, VECTOR8_LOOP);
8155 jcc(Assembler::less, VECTOR4_TAIL);
8156
8157 if (UseAVX >= 2) {
8158 Label VECTOR16_TAIL, VECTOR32_LOOP;
8159
8160 cmpq(length, 16);
8161 jcc(Assembler::equal, VECTOR16_LOOP);
8162 jcc(Assembler::less, VECTOR8_LOOP);
8163
8164 cmpq(length, 32);
8165 jccb(Assembler::less, VECTOR16_TAIL);
8166
8167 subq(length, 32);
8168 bind(VECTOR32_LOOP);
8169 vmovdqu(rymm0, Address(obja, result));
8170 vmovdqu(rymm1, Address(objb, result));
8171 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit);
8172 vptest(rymm2, rymm2);
8173 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found
8174 addq(result, 32);
8175 subq(length, 32);
8176 jcc(Assembler::greaterEqual, VECTOR32_LOOP);
8177 addq(length, 32);
8178 jcc(Assembler::equal, SAME_TILL_END);
8179 //falling through if less than 32 bytes left //close the branch here.
8180
8181 bind(VECTOR16_TAIL);
8182 cmpq(length, 16);
8183 jccb(Assembler::less, VECTOR8_TAIL);
8184 bind(VECTOR16_LOOP);
8185 movdqu(rymm0, Address(obja, result));
8186 movdqu(rymm1, Address(objb, result));
8187 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit);
8188 ptest(rymm2, rymm2);
8189 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
8190 addq(result, 16);
8191 subq(length, 16);
8192 jcc(Assembler::equal, SAME_TILL_END);
8193 //falling through if less than 16 bytes left
8194 } else {//regular intrinsics
8195
8196 cmpq(length, 16);
8197 jccb(Assembler::less, VECTOR8_TAIL);
8198
8199 subq(length, 16);
8200 bind(VECTOR16_LOOP);
8201 movdqu(rymm0, Address(obja, result));
8202 movdqu(rymm1, Address(objb, result));
8203 pxor(rymm0, rymm1);
8204 ptest(rymm0, rymm0);
8205 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
8206 addq(result, 16);
8207 subq(length, 16);
8208 jccb(Assembler::greaterEqual, VECTOR16_LOOP);
8209 addq(length, 16);
8210 jcc(Assembler::equal, SAME_TILL_END);
8211 //falling through if less than 16 bytes left
8212 }
8213
8214 bind(VECTOR8_TAIL);
8215 cmpq(length, 8);
8216 jccb(Assembler::less, VECTOR4_TAIL);
8217 bind(VECTOR8_LOOP);
8218 movq(tmp1, Address(obja, result));
8219 movq(tmp2, Address(objb, result));
8220 xorq(tmp1, tmp2);
8221 testq(tmp1, tmp1);
8222 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found
8223 addq(result, 8);
8224 subq(length, 8);
8225 jcc(Assembler::equal, SAME_TILL_END);
8226 //falling through if less than 8 bytes left
8227
8228 bind(VECTOR4_TAIL);
8229 cmpq(length, 4);
8230 jccb(Assembler::less, BYTES_TAIL);
8231 bind(VECTOR4_LOOP);
8232 movl(tmp1, Address(obja, result));
8233 xorl(tmp1, Address(objb, result));
8234 testl(tmp1, tmp1);
8235 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found
8236 addq(result, 4);
8237 subq(length, 4);
8238 jcc(Assembler::equal, SAME_TILL_END);
8239 //falling through if less than 4 bytes left
8240
8241 bind(BYTES_TAIL);
8242 bind(BYTES_LOOP);
8243 load_unsigned_byte(tmp1, Address(obja, result));
8244 load_unsigned_byte(tmp2, Address(objb, result));
8245 xorl(tmp1, tmp2);
8246 testl(tmp1, tmp1);
8247 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8248 decq(length);
8249 jcc(Assembler::zero, SAME_TILL_END);
8250 incq(result);
8251 load_unsigned_byte(tmp1, Address(obja, result));
8252 load_unsigned_byte(tmp2, Address(objb, result));
8253 xorl(tmp1, tmp2);
8254 testl(tmp1, tmp1);
8255 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8256 decq(length);
8257 jcc(Assembler::zero, SAME_TILL_END);
8258 incq(result);
8259 load_unsigned_byte(tmp1, Address(obja, result));
8260 load_unsigned_byte(tmp2, Address(objb, result));
8261 xorl(tmp1, tmp2);
8262 testl(tmp1, tmp1);
8263 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8264 jmp(SAME_TILL_END);
8265
8266 if (UseAVX >= 2) {
8267 bind(VECTOR32_NOT_EQUAL);
8268 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit);
8269 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit);
8270 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit);
8271 vpmovmskb(tmp1, rymm0);
8272 bsfq(tmp1, tmp1);
8273 addq(result, tmp1);
8274 shrq(result);
8275 jmp(DONE);
8276 }
8277
8278 bind(VECTOR16_NOT_EQUAL);
8279 if (UseAVX >= 2) {
8280 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit);
8281 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit);
8282 pxor(rymm0, rymm2);
8283 } else {
8284 pcmpeqb(rymm2, rymm2);
8285 pxor(rymm0, rymm1);
8286 pcmpeqb(rymm0, rymm1);
8287 pxor(rymm0, rymm2);
8288 }
8289 pmovmskb(tmp1, rymm0);
8290 bsfq(tmp1, tmp1);
8291 addq(result, tmp1);
8292 shrq(result);
8293 jmpb(DONE);
8294
8295 bind(VECTOR8_NOT_EQUAL);
8296 bind(VECTOR4_NOT_EQUAL);
8297 bsfq(tmp1, tmp1);
8298 shrq(tmp1, 3);
8299 addq(result, tmp1);
8300 bind(BYTES_NOT_EQUAL);
8301 shrq(result);
8302 jmpb(DONE);
8303
8304 bind(SAME_TILL_END);
8305 mov64(result, -1);
8306
8307 bind(DONE);
8308 }
8309
8310 //Helper functions for square_to_len()
8311
8312 /**
8313 * Store the squares of x[], right shifted one bit (divided by 2) into z[]
8314 * Preserves x and z and modifies rest of the registers.
8315 */
8316 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8317 // Perform square and right shift by 1
8318 // Handle odd xlen case first, then for even xlen do the following
8319 // jlong carry = 0;
8320 // for (int j=0, i=0; j < xlen; j+=2, i+=4) {
8321 // huge_128 product = x[j:j+1] * x[j:j+1];
8322 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65);
8323 // z[i+2:i+3] = (jlong)(product >>> 1);
8324 // carry = (jlong)product;
8325 // }
8326
8327 xorq(tmp5, tmp5); // carry
8328 xorq(rdxReg, rdxReg);
8329 xorl(tmp1, tmp1); // index for x
8330 xorl(tmp4, tmp4); // index for z
8331
8332 Label L_first_loop, L_first_loop_exit;
8333
8334 testl(xlen, 1);
8335 jccb(Assembler::zero, L_first_loop); //jump if xlen is even
8336
8337 // Square and right shift by 1 the odd element using 32 bit multiply
8338 movl(raxReg, Address(x, tmp1, Address::times_4, 0));
8339 imulq(raxReg, raxReg);
8340 shrq(raxReg, 1);
8341 adcq(tmp5, 0);
8342 movq(Address(z, tmp4, Address::times_4, 0), raxReg);
8343 incrementl(tmp1);
8344 addl(tmp4, 2);
8345
8346 // Square and right shift by 1 the rest using 64 bit multiply
8347 bind(L_first_loop);
8348 cmpptr(tmp1, xlen);
8349 jccb(Assembler::equal, L_first_loop_exit);
8350
8351 // Square
8352 movq(raxReg, Address(x, tmp1, Address::times_4, 0));
8353 rorq(raxReg, 32); // convert big-endian to little-endian
8354 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax
8355
8356 // Right shift by 1 and save carry
8357 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1
8358 rcrq(rdxReg, 1);
8359 rcrq(raxReg, 1);
8360 adcq(tmp5, 0);
8361
8362 // Store result in z
8363 movq(Address(z, tmp4, Address::times_4, 0), rdxReg);
8364 movq(Address(z, tmp4, Address::times_4, 8), raxReg);
8365
8366 // Update indices for x and z
8367 addl(tmp1, 2);
8368 addl(tmp4, 4);
8369 jmp(L_first_loop);
8370
8371 bind(L_first_loop_exit);
8372 }
8373
8374
8375 /**
8376 * Perform the following multiply add operation using BMI2 instructions
8377 * carry:sum = sum + op1*op2 + carry
8378 * op2 should be in rdx
8379 * op2 is preserved, all other registers are modified
8380 */
8381 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) {
8382 // assert op2 is rdx
8383 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1
8384 addq(sum, carry);
8385 adcq(tmp2, 0);
8386 addq(sum, op1);
8387 adcq(tmp2, 0);
8388 movq(carry, tmp2);
8389 }
8390
8391 /**
8392 * Perform the following multiply add operation:
8393 * carry:sum = sum + op1*op2 + carry
8394 * Preserves op1, op2 and modifies rest of registers
8395 */
8396 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) {
8397 // rdx:rax = op1 * op2
8398 movq(raxReg, op2);
8399 mulq(op1);
8400
8401 // rdx:rax = sum + carry + rdx:rax
8402 addq(sum, carry);
8403 adcq(rdxReg, 0);
8404 addq(sum, raxReg);
8405 adcq(rdxReg, 0);
8406
8407 // carry:sum = rdx:sum
8408 movq(carry, rdxReg);
8409 }
8410
8411 /**
8412 * Add 64 bit long carry into z[] with carry propogation.
8413 * Preserves z and carry register values and modifies rest of registers.
8414 *
8415 */
8416 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) {
8417 Label L_fourth_loop, L_fourth_loop_exit;
8418
8419 movl(tmp1, 1);
8420 subl(zlen, 2);
8421 addq(Address(z, zlen, Address::times_4, 0), carry);
8422
8423 bind(L_fourth_loop);
8424 jccb(Assembler::carryClear, L_fourth_loop_exit);
8425 subl(zlen, 2);
8426 jccb(Assembler::negative, L_fourth_loop_exit);
8427 addq(Address(z, zlen, Address::times_4, 0), tmp1);
8428 jmp(L_fourth_loop);
8429 bind(L_fourth_loop_exit);
8430 }
8431
8432 /**
8433 * Shift z[] left by 1 bit.
8434 * Preserves x, len, z and zlen registers and modifies rest of the registers.
8435 *
8436 */
8437 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) {
8438
8439 Label L_fifth_loop, L_fifth_loop_exit;
8440
8441 // Fifth loop
8442 // Perform primitiveLeftShift(z, zlen, 1)
8443
8444 const Register prev_carry = tmp1;
8445 const Register new_carry = tmp4;
8446 const Register value = tmp2;
8447 const Register zidx = tmp3;
8448
8449 // int zidx, carry;
8450 // long value;
8451 // carry = 0;
8452 // for (zidx = zlen-2; zidx >=0; zidx -= 2) {
8453 // (carry:value) = (z[i] << 1) | carry ;
8454 // z[i] = value;
8455 // }
8456
8457 movl(zidx, zlen);
8458 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register
8459
8460 bind(L_fifth_loop);
8461 decl(zidx); // Use decl to preserve carry flag
8462 decl(zidx);
8463 jccb(Assembler::negative, L_fifth_loop_exit);
8464
8465 if (UseBMI2Instructions) {
8466 movq(value, Address(z, zidx, Address::times_4, 0));
8467 rclq(value, 1);
8468 rorxq(value, value, 32);
8469 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
8470 }
8471 else {
8472 // clear new_carry
8473 xorl(new_carry, new_carry);
8474
8475 // Shift z[i] by 1, or in previous carry and save new carry
8476 movq(value, Address(z, zidx, Address::times_4, 0));
8477 shlq(value, 1);
8478 adcl(new_carry, 0);
8479
8480 orq(value, prev_carry);
8481 rorq(value, 0x20);
8482 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
8483
8484 // Set previous carry = new carry
8485 movl(prev_carry, new_carry);
8486 }
8487 jmp(L_fifth_loop);
8488
8489 bind(L_fifth_loop_exit);
8490 }
8491
8492
8493 /**
8494 * Code for BigInteger::squareToLen() intrinsic
8495 *
8496 * rdi: x
8497 * rsi: len
8498 * r8: z
8499 * rcx: zlen
8500 * r12: tmp1
8501 * r13: tmp2
8502 * r14: tmp3
8503 * r15: tmp4
8504 * rbx: tmp5
8505 *
8506 */
8507 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) {
8508
8509 Label L_second_loop, L_second_loop_exit, L_third_loop, L_third_loop_exit, L_last_x, L_multiply;
8510 push(tmp1);
8511 push(tmp2);
8512 push(tmp3);
8513 push(tmp4);
8514 push(tmp5);
8515
8516 // First loop
8517 // Store the squares, right shifted one bit (i.e., divided by 2).
8518 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg);
8519
8520 // Add in off-diagonal sums.
8521 //
8522 // Second, third (nested) and fourth loops.
8523 // zlen +=2;
8524 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) {
8525 // carry = 0;
8526 // long op2 = x[xidx:xidx+1];
8527 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) {
8528 // k -= 2;
8529 // long op1 = x[j:j+1];
8530 // long sum = z[k:k+1];
8531 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs);
8532 // z[k:k+1] = sum;
8533 // }
8534 // add_one_64(z, k, carry, tmp_regs);
8535 // }
8536
8537 const Register carry = tmp5;
8538 const Register sum = tmp3;
8539 const Register op1 = tmp4;
8540 Register op2 = tmp2;
8541
8542 push(zlen);
8543 push(len);
8544 addl(zlen,2);
8545 bind(L_second_loop);
8546 xorq(carry, carry);
8547 subl(zlen, 4);
8548 subl(len, 2);
8549 push(zlen);
8550 push(len);
8551 cmpl(len, 0);
8552 jccb(Assembler::lessEqual, L_second_loop_exit);
8553
8554 // Multiply an array by one 64 bit long.
8555 if (UseBMI2Instructions) {
8556 op2 = rdxReg;
8557 movq(op2, Address(x, len, Address::times_4, 0));
8558 rorxq(op2, op2, 32);
8559 }
8560 else {
8561 movq(op2, Address(x, len, Address::times_4, 0));
8562 rorq(op2, 32);
8563 }
8564
8565 bind(L_third_loop);
8566 decrementl(len);
8567 jccb(Assembler::negative, L_third_loop_exit);
8568 decrementl(len);
8569 jccb(Assembler::negative, L_last_x);
8570
8571 movq(op1, Address(x, len, Address::times_4, 0));
8572 rorq(op1, 32);
8573
8574 bind(L_multiply);
8575 subl(zlen, 2);
8576 movq(sum, Address(z, zlen, Address::times_4, 0));
8577
8578 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry.
8579 if (UseBMI2Instructions) {
8580 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2);
8581 }
8582 else {
8583 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8584 }
8585
8586 movq(Address(z, zlen, Address::times_4, 0), sum);
8587
8588 jmp(L_third_loop);
8589 bind(L_third_loop_exit);
8590
8591 // Fourth loop
8592 // Add 64 bit long carry into z with carry propogation.
8593 // Uses offsetted zlen.
8594 add_one_64(z, zlen, carry, tmp1);
8595
8596 pop(len);
8597 pop(zlen);
8598 jmp(L_second_loop);
8599
8600 // Next infrequent code is moved outside loops.
8601 bind(L_last_x);
8602 movl(op1, Address(x, 0));
8603 jmp(L_multiply);
8604
8605 bind(L_second_loop_exit);
8606 pop(len);
8607 pop(zlen);
8608 pop(len);
8609 pop(zlen);
8610
8611 // Fifth loop
8612 // Shift z left 1 bit.
8613 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4);
8614
8615 // z[zlen-1] |= x[len-1] & 1;
8616 movl(tmp3, Address(x, len, Address::times_4, -4));
8617 andl(tmp3, 1);
8618 orl(Address(z, zlen, Address::times_4, -4), tmp3);
8619
8620 pop(tmp5);
8621 pop(tmp4);
8622 pop(tmp3);
8623 pop(tmp2);
8624 pop(tmp1);
8625 }
8626
8627 /**
8628 * Helper function for mul_add()
8629 * Multiply the in[] by int k and add to out[] starting at offset offs using
8630 * 128 bit by 32 bit multiply and return the carry in tmp5.
8631 * Only quad int aligned length of in[] is operated on in this function.
8632 * k is in rdxReg for BMI2Instructions, for others it is in tmp2.
8633 * This function preserves out, in and k registers.
8634 * len and offset point to the appropriate index in "in" & "out" correspondingly
8635 * tmp5 has the carry.
8636 * other registers are temporary and are modified.
8637 *
8638 */
8639 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in,
8640 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3,
8641 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8642
8643 Label L_first_loop, L_first_loop_exit;
8644
8645 movl(tmp1, len);
8646 shrl(tmp1, 2);
8647
8648 bind(L_first_loop);
8649 subl(tmp1, 1);
8650 jccb(Assembler::negative, L_first_loop_exit);
8651
8652 subl(len, 4);
8653 subl(offset, 4);
8654
8655 Register op2 = tmp2;
8656 const Register sum = tmp3;
8657 const Register op1 = tmp4;
8658 const Register carry = tmp5;
8659
8660 if (UseBMI2Instructions) {
8661 op2 = rdxReg;
8662 }
8663
8664 movq(op1, Address(in, len, Address::times_4, 8));
8665 rorq(op1, 32);
8666 movq(sum, Address(out, offset, Address::times_4, 8));
8667 rorq(sum, 32);
8668 if (UseBMI2Instructions) {
8669 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8670 }
8671 else {
8672 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8673 }
8674 // Store back in big endian from little endian
8675 rorq(sum, 0x20);
8676 movq(Address(out, offset, Address::times_4, 8), sum);
8677
8678 movq(op1, Address(in, len, Address::times_4, 0));
8679 rorq(op1, 32);
8680 movq(sum, Address(out, offset, Address::times_4, 0));
8681 rorq(sum, 32);
8682 if (UseBMI2Instructions) {
8683 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8684 }
8685 else {
8686 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8687 }
8688 // Store back in big endian from little endian
8689 rorq(sum, 0x20);
8690 movq(Address(out, offset, Address::times_4, 0), sum);
8691
8692 jmp(L_first_loop);
8693 bind(L_first_loop_exit);
8694 }
8695
8696 /**
8697 * Code for BigInteger::mulAdd() intrinsic
8698 *
8699 * rdi: out
8700 * rsi: in
8701 * r11: offs (out.length - offset)
8702 * rcx: len
8703 * r8: k
8704 * r12: tmp1
8705 * r13: tmp2
8706 * r14: tmp3
8707 * r15: tmp4
8708 * rbx: tmp5
8709 * Multiply the in[] by word k and add to out[], return the carry in rax
8710 */
8711 void MacroAssembler::mul_add(Register out, Register in, Register offs,
8712 Register len, Register k, Register tmp1, Register tmp2, Register tmp3,
8713 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8714
8715 Label L_carry, L_last_in, L_done;
8716
8717 // carry = 0;
8718 // for (int j=len-1; j >= 0; j--) {
8719 // long product = (in[j] & LONG_MASK) * kLong +
8720 // (out[offs] & LONG_MASK) + carry;
8721 // out[offs--] = (int)product;
8722 // carry = product >>> 32;
8723 // }
8724 //
8725 push(tmp1);
8726 push(tmp2);
8727 push(tmp3);
8728 push(tmp4);
8729 push(tmp5);
8730
8731 Register op2 = tmp2;
8732 const Register sum = tmp3;
8733 const Register op1 = tmp4;
8734 const Register carry = tmp5;
8735
8736 if (UseBMI2Instructions) {
8737 op2 = rdxReg;
8738 movl(op2, k);
8739 }
8740 else {
8741 movl(op2, k);
8742 }
8743
8744 xorq(carry, carry);
8745
8746 //First loop
8747
8748 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply
8749 //The carry is in tmp5
8750 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg);
8751
8752 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any
8753 decrementl(len);
8754 jccb(Assembler::negative, L_carry);
8755 decrementl(len);
8756 jccb(Assembler::negative, L_last_in);
8757
8758 movq(op1, Address(in, len, Address::times_4, 0));
8759 rorq(op1, 32);
8760
8761 subl(offs, 2);
8762 movq(sum, Address(out, offs, Address::times_4, 0));
8763 rorq(sum, 32);
8764
8765 if (UseBMI2Instructions) {
8766 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8767 }
8768 else {
8769 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8770 }
8771
8772 // Store back in big endian from little endian
8773 rorq(sum, 0x20);
8774 movq(Address(out, offs, Address::times_4, 0), sum);
8775
8776 testl(len, len);
8777 jccb(Assembler::zero, L_carry);
8778
8779 //Multiply the last in[] entry, if any
8780 bind(L_last_in);
8781 movl(op1, Address(in, 0));
8782 movl(sum, Address(out, offs, Address::times_4, -4));
8783
8784 movl(raxReg, k);
8785 mull(op1); //tmp4 * eax -> edx:eax
8786 addl(sum, carry);
8787 adcl(rdxReg, 0);
8788 addl(sum, raxReg);
8789 adcl(rdxReg, 0);
8790 movl(carry, rdxReg);
8791
8792 movl(Address(out, offs, Address::times_4, -4), sum);
8793
8794 bind(L_carry);
8795 //return tmp5/carry as carry in rax
8796 movl(rax, carry);
8797
8798 bind(L_done);
8799 pop(tmp5);
8800 pop(tmp4);
8801 pop(tmp3);
8802 pop(tmp2);
8803 pop(tmp1);
8804 }
8805 #endif
8806
8807 /**
8808 * Emits code to update CRC-32 with a byte value according to constants in table
8809 *
8810 * @param [in,out]crc Register containing the crc.
8811 * @param [in]val Register containing the byte to fold into the CRC.
8812 * @param [in]table Register containing the table of crc constants.
8813 *
8814 * uint32_t crc;
8815 * val = crc_table[(val ^ crc) & 0xFF];
8816 * crc = val ^ (crc >> 8);
8817 *
8818 */
8819 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
8820 xorl(val, crc);
8821 andl(val, 0xFF);
8822 shrl(crc, 8); // unsigned shift
8823 xorl(crc, Address(table, val, Address::times_4, 0));
8824 }
8825
8826 /**
8827 * Fold four 128-bit data chunks
8828 */
8829 void MacroAssembler::fold_128bit_crc32_avx512(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
8830 evpclmulhdq(xtmp, xK, xcrc, Assembler::AVX_512bit); // [123:64]
8831 evpclmulldq(xcrc, xK, xcrc, Assembler::AVX_512bit); // [63:0]
8832 evpxorq(xcrc, xcrc, Address(buf, offset), Assembler::AVX_512bit /* vector_len */);
8833 evpxorq(xcrc, xcrc, xtmp, Assembler::AVX_512bit /* vector_len */);
8834 }
8835
8836 /**
8837 * Fold 128-bit data chunk
8838 */
8839 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
8840 if (UseAVX > 0) {
8841 vpclmulhdq(xtmp, xK, xcrc); // [123:64]
8842 vpclmulldq(xcrc, xK, xcrc); // [63:0]
8843 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */);
8844 pxor(xcrc, xtmp);
8845 } else {
8846 movdqa(xtmp, xcrc);
8847 pclmulhdq(xtmp, xK); // [123:64]
8848 pclmulldq(xcrc, xK); // [63:0]
8849 pxor(xcrc, xtmp);
8850 movdqu(xtmp, Address(buf, offset));
8851 pxor(xcrc, xtmp);
8852 }
8853 }
8854
8855 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
8856 if (UseAVX > 0) {
8857 vpclmulhdq(xtmp, xK, xcrc);
8858 vpclmulldq(xcrc, xK, xcrc);
8859 pxor(xcrc, xbuf);
8860 pxor(xcrc, xtmp);
8861 } else {
8862 movdqa(xtmp, xcrc);
8863 pclmulhdq(xtmp, xK);
8864 pclmulldq(xcrc, xK);
8865 pxor(xcrc, xbuf);
8866 pxor(xcrc, xtmp);
8867 }
8868 }
8869
8870 /**
8871 * 8-bit folds to compute 32-bit CRC
8872 *
8873 * uint64_t xcrc;
8874 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8);
8875 */
8876 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) {
8877 movdl(tmp, xcrc);
8878 andl(tmp, 0xFF);
8879 movdl(xtmp, Address(table, tmp, Address::times_4, 0));
8880 psrldq(xcrc, 1); // unsigned shift one byte
8881 pxor(xcrc, xtmp);
8882 }
8883
8884 /**
8885 * uint32_t crc;
8886 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
8887 */
8888 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
8889 movl(tmp, crc);
8890 andl(tmp, 0xFF);
8891 shrl(crc, 8);
8892 xorl(crc, Address(table, tmp, Address::times_4, 0));
8893 }
8894
8895 /**
8896 * @param crc register containing existing CRC (32-bit)
8897 * @param buf register pointing to input byte buffer (byte*)
8898 * @param len register containing number of bytes
8899 * @param table register that will contain address of CRC table
8900 * @param tmp scratch register
8901 */
8902 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) {
8903 assert_different_registers(crc, buf, len, table, tmp, rax);
8904
8905 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned;
8906 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop;
8907
8908 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge
8909 // context for the registers used, where all instructions below are using 128-bit mode
8910 // On EVEX without VL and BW, these instructions will all be AVX.
8911 lea(table, ExternalAddress(StubRoutines::crc_table_addr()));
8912 notl(crc); // ~crc
8913 cmpl(len, 16);
8914 jcc(Assembler::less, L_tail);
8915
8916 // Align buffer to 16 bytes
8917 movl(tmp, buf);
8918 andl(tmp, 0xF);
8919 jccb(Assembler::zero, L_aligned);
8920 subl(tmp, 16);
8921 addl(len, tmp);
8922
8923 align(4);
8924 BIND(L_align_loop);
8925 movsbl(rax, Address(buf, 0)); // load byte with sign extension
8926 update_byte_crc32(crc, rax, table);
8927 increment(buf);
8928 incrementl(tmp);
8929 jccb(Assembler::less, L_align_loop);
8930
8931 BIND(L_aligned);
8932 movl(tmp, len); // save
8933 shrl(len, 4);
8934 jcc(Assembler::zero, L_tail_restore);
8935
8936 // Fold total 512 bits of polynomial on each iteration
8937 if (VM_Version::supports_vpclmulqdq()) {
8938 Label Parallel_loop, L_No_Parallel;
8939
8940 cmpl(len, 8);
8941 jccb(Assembler::less, L_No_Parallel);
8942
8943 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
8944 evmovdquq(xmm1, Address(buf, 0), Assembler::AVX_512bit);
8945 movdl(xmm5, crc);
8946 evpxorq(xmm1, xmm1, xmm5, Assembler::AVX_512bit);
8947 addptr(buf, 64);
8948 subl(len, 7);
8949 evshufi64x2(xmm0, xmm0, xmm0, 0x00, Assembler::AVX_512bit); //propagate the mask from 128 bits to 512 bits
8950
8951 BIND(Parallel_loop);
8952 fold_128bit_crc32_avx512(xmm1, xmm0, xmm5, buf, 0);
8953 addptr(buf, 64);
8954 subl(len, 4);
8955 jcc(Assembler::greater, Parallel_loop);
8956
8957 vextracti64x2(xmm2, xmm1, 0x01);
8958 vextracti64x2(xmm3, xmm1, 0x02);
8959 vextracti64x2(xmm4, xmm1, 0x03);
8960 jmp(L_fold_512b);
8961
8962 BIND(L_No_Parallel);
8963 }
8964 // Fold crc into first bytes of vector
8965 movdqa(xmm1, Address(buf, 0));
8966 movdl(rax, xmm1);
8967 xorl(crc, rax);
8968 if (VM_Version::supports_sse4_1()) {
8969 pinsrd(xmm1, crc, 0);
8970 } else {
8971 pinsrw(xmm1, crc, 0);
8972 shrl(crc, 16);
8973 pinsrw(xmm1, crc, 1);
8974 }
8975 addptr(buf, 16);
8976 subl(len, 4); // len > 0
8977 jcc(Assembler::less, L_fold_tail);
8978
8979 movdqa(xmm2, Address(buf, 0));
8980 movdqa(xmm3, Address(buf, 16));
8981 movdqa(xmm4, Address(buf, 32));
8982 addptr(buf, 48);
8983 subl(len, 3);
8984 jcc(Assembler::lessEqual, L_fold_512b);
8985
8986 // Fold total 512 bits of polynomial on each iteration,
8987 // 128 bits per each of 4 parallel streams.
8988 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
8989
8990 align(32);
8991 BIND(L_fold_512b_loop);
8992 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
8993 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16);
8994 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32);
8995 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48);
8996 addptr(buf, 64);
8997 subl(len, 4);
8998 jcc(Assembler::greater, L_fold_512b_loop);
8999
9000 // Fold 512 bits to 128 bits.
9001 BIND(L_fold_512b);
9002 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
9003 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2);
9004 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3);
9005 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4);
9006
9007 // Fold the rest of 128 bits data chunks
9008 BIND(L_fold_tail);
9009 addl(len, 3);
9010 jccb(Assembler::lessEqual, L_fold_128b);
9011 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
9012
9013 BIND(L_fold_tail_loop);
9014 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
9015 addptr(buf, 16);
9016 decrementl(len);
9017 jccb(Assembler::greater, L_fold_tail_loop);
9018
9019 // Fold 128 bits in xmm1 down into 32 bits in crc register.
9020 BIND(L_fold_128b);
9021 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
9022 if (UseAVX > 0) {
9023 vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
9024 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */);
9025 vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
9026 } else {
9027 movdqa(xmm2, xmm0);
9028 pclmulqdq(xmm2, xmm1, 0x1);
9029 movdqa(xmm3, xmm0);
9030 pand(xmm3, xmm2);
9031 pclmulqdq(xmm0, xmm3, 0x1);
9032 }
9033 psrldq(xmm1, 8);
9034 psrldq(xmm2, 4);
9035 pxor(xmm0, xmm1);
9036 pxor(xmm0, xmm2);
9037
9038 // 8 8-bit folds to compute 32-bit CRC.
9039 for (int j = 0; j < 4; j++) {
9040 fold_8bit_crc32(xmm0, table, xmm1, rax);
9041 }
9042 movdl(crc, xmm0); // mov 32 bits to general register
9043 for (int j = 0; j < 4; j++) {
9044 fold_8bit_crc32(crc, table, rax);
9045 }
9046
9047 BIND(L_tail_restore);
9048 movl(len, tmp); // restore
9049 BIND(L_tail);
9050 andl(len, 0xf);
9051 jccb(Assembler::zero, L_exit);
9052
9053 // Fold the rest of bytes
9054 align(4);
9055 BIND(L_tail_loop);
9056 movsbl(rax, Address(buf, 0)); // load byte with sign extension
9057 update_byte_crc32(crc, rax, table);
9058 increment(buf);
9059 decrementl(len);
9060 jccb(Assembler::greater, L_tail_loop);
9061
9062 BIND(L_exit);
9063 notl(crc); // ~c
9064 }
9065
9066 #ifdef _LP64
9067 // S. Gueron / Information Processing Letters 112 (2012) 184
9068 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table.
9069 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0].
9070 // Output: the 64-bit carry-less product of B * CONST
9071 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n,
9072 Register tmp1, Register tmp2, Register tmp3) {
9073 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
9074 if (n > 0) {
9075 addq(tmp3, n * 256 * 8);
9076 }
9077 // Q1 = TABLEExt[n][B & 0xFF];
9078 movl(tmp1, in);
9079 andl(tmp1, 0x000000FF);
9080 shll(tmp1, 3);
9081 addq(tmp1, tmp3);
9082 movq(tmp1, Address(tmp1, 0));
9083
9084 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
9085 movl(tmp2, in);
9086 shrl(tmp2, 8);
9087 andl(tmp2, 0x000000FF);
9088 shll(tmp2, 3);
9089 addq(tmp2, tmp3);
9090 movq(tmp2, Address(tmp2, 0));
9091
9092 shlq(tmp2, 8);
9093 xorq(tmp1, tmp2);
9094
9095 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
9096 movl(tmp2, in);
9097 shrl(tmp2, 16);
9098 andl(tmp2, 0x000000FF);
9099 shll(tmp2, 3);
9100 addq(tmp2, tmp3);
9101 movq(tmp2, Address(tmp2, 0));
9102
9103 shlq(tmp2, 16);
9104 xorq(tmp1, tmp2);
9105
9106 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
9107 shrl(in, 24);
9108 andl(in, 0x000000FF);
9109 shll(in, 3);
9110 addq(in, tmp3);
9111 movq(in, Address(in, 0));
9112
9113 shlq(in, 24);
9114 xorq(in, tmp1);
9115 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
9116 }
9117
9118 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
9119 Register in_out,
9120 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
9121 XMMRegister w_xtmp2,
9122 Register tmp1,
9123 Register n_tmp2, Register n_tmp3) {
9124 if (is_pclmulqdq_supported) {
9125 movdl(w_xtmp1, in_out); // modified blindly
9126
9127 movl(tmp1, const_or_pre_comp_const_index);
9128 movdl(w_xtmp2, tmp1);
9129 pclmulqdq(w_xtmp1, w_xtmp2, 0);
9130
9131 movdq(in_out, w_xtmp1);
9132 } else {
9133 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3);
9134 }
9135 }
9136
9137 // Recombination Alternative 2: No bit-reflections
9138 // T1 = (CRC_A * U1) << 1
9139 // T2 = (CRC_B * U2) << 1
9140 // C1 = T1 >> 32
9141 // C2 = T2 >> 32
9142 // T1 = T1 & 0xFFFFFFFF
9143 // T2 = T2 & 0xFFFFFFFF
9144 // T1 = CRC32(0, T1)
9145 // T2 = CRC32(0, T2)
9146 // C1 = C1 ^ T1
9147 // C2 = C2 ^ T2
9148 // CRC = C1 ^ C2 ^ CRC_C
9149 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,
9150 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9151 Register tmp1, Register tmp2,
9152 Register n_tmp3) {
9153 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9154 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9155 shlq(in_out, 1);
9156 movl(tmp1, in_out);
9157 shrq(in_out, 32);
9158 xorl(tmp2, tmp2);
9159 crc32(tmp2, tmp1, 4);
9160 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here
9161 shlq(in1, 1);
9162 movl(tmp1, in1);
9163 shrq(in1, 32);
9164 xorl(tmp2, tmp2);
9165 crc32(tmp2, tmp1, 4);
9166 xorl(in1, tmp2);
9167 xorl(in_out, in1);
9168 xorl(in_out, in2);
9169 }
9170
9171 // Set N to predefined value
9172 // Subtract from a lenght of a buffer
9173 // execute in a loop:
9174 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0
9175 // for i = 1 to N do
9176 // CRC_A = CRC32(CRC_A, A[i])
9177 // CRC_B = CRC32(CRC_B, B[i])
9178 // CRC_C = CRC32(CRC_C, C[i])
9179 // end for
9180 // Recombine
9181 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,
9182 Register in_out1, Register in_out2, Register in_out3,
9183 Register tmp1, Register tmp2, Register tmp3,
9184 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9185 Register tmp4, Register tmp5,
9186 Register n_tmp6) {
9187 Label L_processPartitions;
9188 Label L_processPartition;
9189 Label L_exit;
9190
9191 bind(L_processPartitions);
9192 cmpl(in_out1, 3 * size);
9193 jcc(Assembler::less, L_exit);
9194 xorl(tmp1, tmp1);
9195 xorl(tmp2, tmp2);
9196 movq(tmp3, in_out2);
9197 addq(tmp3, size);
9198
9199 bind(L_processPartition);
9200 crc32(in_out3, Address(in_out2, 0), 8);
9201 crc32(tmp1, Address(in_out2, size), 8);
9202 crc32(tmp2, Address(in_out2, size * 2), 8);
9203 addq(in_out2, 8);
9204 cmpq(in_out2, tmp3);
9205 jcc(Assembler::less, L_processPartition);
9206 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
9207 w_xtmp1, w_xtmp2, w_xtmp3,
9208 tmp4, tmp5,
9209 n_tmp6);
9210 addq(in_out2, 2 * size);
9211 subl(in_out1, 3 * size);
9212 jmp(L_processPartitions);
9213
9214 bind(L_exit);
9215 }
9216 #else
9217 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n,
9218 Register tmp1, Register tmp2, Register tmp3,
9219 XMMRegister xtmp1, XMMRegister xtmp2) {
9220 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
9221 if (n > 0) {
9222 addl(tmp3, n * 256 * 8);
9223 }
9224 // Q1 = TABLEExt[n][B & 0xFF];
9225 movl(tmp1, in_out);
9226 andl(tmp1, 0x000000FF);
9227 shll(tmp1, 3);
9228 addl(tmp1, tmp3);
9229 movq(xtmp1, Address(tmp1, 0));
9230
9231 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
9232 movl(tmp2, in_out);
9233 shrl(tmp2, 8);
9234 andl(tmp2, 0x000000FF);
9235 shll(tmp2, 3);
9236 addl(tmp2, tmp3);
9237 movq(xtmp2, Address(tmp2, 0));
9238
9239 psllq(xtmp2, 8);
9240 pxor(xtmp1, xtmp2);
9241
9242 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
9243 movl(tmp2, in_out);
9244 shrl(tmp2, 16);
9245 andl(tmp2, 0x000000FF);
9246 shll(tmp2, 3);
9247 addl(tmp2, tmp3);
9248 movq(xtmp2, Address(tmp2, 0));
9249
9250 psllq(xtmp2, 16);
9251 pxor(xtmp1, xtmp2);
9252
9253 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
9254 shrl(in_out, 24);
9255 andl(in_out, 0x000000FF);
9256 shll(in_out, 3);
9257 addl(in_out, tmp3);
9258 movq(xtmp2, Address(in_out, 0));
9259
9260 psllq(xtmp2, 24);
9261 pxor(xtmp1, xtmp2); // Result in CXMM
9262 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
9263 }
9264
9265 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
9266 Register in_out,
9267 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
9268 XMMRegister w_xtmp2,
9269 Register tmp1,
9270 Register n_tmp2, Register n_tmp3) {
9271 if (is_pclmulqdq_supported) {
9272 movdl(w_xtmp1, in_out);
9273
9274 movl(tmp1, const_or_pre_comp_const_index);
9275 movdl(w_xtmp2, tmp1);
9276 pclmulqdq(w_xtmp1, w_xtmp2, 0);
9277 // Keep result in XMM since GPR is 32 bit in length
9278 } else {
9279 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2);
9280 }
9281 }
9282
9283 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,
9284 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9285 Register tmp1, Register tmp2,
9286 Register n_tmp3) {
9287 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9288 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9289
9290 psllq(w_xtmp1, 1);
9291 movdl(tmp1, w_xtmp1);
9292 psrlq(w_xtmp1, 32);
9293 movdl(in_out, w_xtmp1);
9294
9295 xorl(tmp2, tmp2);
9296 crc32(tmp2, tmp1, 4);
9297 xorl(in_out, tmp2);
9298
9299 psllq(w_xtmp2, 1);
9300 movdl(tmp1, w_xtmp2);
9301 psrlq(w_xtmp2, 32);
9302 movdl(in1, w_xtmp2);
9303
9304 xorl(tmp2, tmp2);
9305 crc32(tmp2, tmp1, 4);
9306 xorl(in1, tmp2);
9307 xorl(in_out, in1);
9308 xorl(in_out, in2);
9309 }
9310
9311 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,
9312 Register in_out1, Register in_out2, Register in_out3,
9313 Register tmp1, Register tmp2, Register tmp3,
9314 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9315 Register tmp4, Register tmp5,
9316 Register n_tmp6) {
9317 Label L_processPartitions;
9318 Label L_processPartition;
9319 Label L_exit;
9320
9321 bind(L_processPartitions);
9322 cmpl(in_out1, 3 * size);
9323 jcc(Assembler::less, L_exit);
9324 xorl(tmp1, tmp1);
9325 xorl(tmp2, tmp2);
9326 movl(tmp3, in_out2);
9327 addl(tmp3, size);
9328
9329 bind(L_processPartition);
9330 crc32(in_out3, Address(in_out2, 0), 4);
9331 crc32(tmp1, Address(in_out2, size), 4);
9332 crc32(tmp2, Address(in_out2, size*2), 4);
9333 crc32(in_out3, Address(in_out2, 0+4), 4);
9334 crc32(tmp1, Address(in_out2, size+4), 4);
9335 crc32(tmp2, Address(in_out2, size*2+4), 4);
9336 addl(in_out2, 8);
9337 cmpl(in_out2, tmp3);
9338 jcc(Assembler::less, L_processPartition);
9339
9340 push(tmp3);
9341 push(in_out1);
9342 push(in_out2);
9343 tmp4 = tmp3;
9344 tmp5 = in_out1;
9345 n_tmp6 = in_out2;
9346
9347 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
9348 w_xtmp1, w_xtmp2, w_xtmp3,
9349 tmp4, tmp5,
9350 n_tmp6);
9351
9352 pop(in_out2);
9353 pop(in_out1);
9354 pop(tmp3);
9355
9356 addl(in_out2, 2 * size);
9357 subl(in_out1, 3 * size);
9358 jmp(L_processPartitions);
9359
9360 bind(L_exit);
9361 }
9362 #endif //LP64
9363
9364 #ifdef _LP64
9365 // Algorithm 2: Pipelined usage of the CRC32 instruction.
9366 // Input: A buffer I of L bytes.
9367 // Output: the CRC32C value of the buffer.
9368 // Notations:
9369 // Write L = 24N + r, with N = floor (L/24).
9370 // r = L mod 24 (0 <= r < 24).
9371 // Consider I as the concatenation of A|B|C|R, where A, B, C, each,
9372 // N quadwords, and R consists of r bytes.
9373 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1
9374 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1
9375 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1
9376 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1
9377 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
9378 Register tmp1, Register tmp2, Register tmp3,
9379 Register tmp4, Register tmp5, Register tmp6,
9380 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9381 bool is_pclmulqdq_supported) {
9382 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
9383 Label L_wordByWord;
9384 Label L_byteByByteProlog;
9385 Label L_byteByByte;
9386 Label L_exit;
9387
9388 if (is_pclmulqdq_supported ) {
9389 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
9390 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1);
9391
9392 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
9393 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
9394
9395 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
9396 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
9397 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\"");
9398 } else {
9399 const_or_pre_comp_const_index[0] = 1;
9400 const_or_pre_comp_const_index[1] = 0;
9401
9402 const_or_pre_comp_const_index[2] = 3;
9403 const_or_pre_comp_const_index[3] = 2;
9404
9405 const_or_pre_comp_const_index[4] = 5;
9406 const_or_pre_comp_const_index[5] = 4;
9407 }
9408 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
9409 in2, in1, in_out,
9410 tmp1, tmp2, tmp3,
9411 w_xtmp1, w_xtmp2, w_xtmp3,
9412 tmp4, tmp5,
9413 tmp6);
9414 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
9415 in2, in1, in_out,
9416 tmp1, tmp2, tmp3,
9417 w_xtmp1, w_xtmp2, w_xtmp3,
9418 tmp4, tmp5,
9419 tmp6);
9420 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
9421 in2, in1, in_out,
9422 tmp1, tmp2, tmp3,
9423 w_xtmp1, w_xtmp2, w_xtmp3,
9424 tmp4, tmp5,
9425 tmp6);
9426 movl(tmp1, in2);
9427 andl(tmp1, 0x00000007);
9428 negl(tmp1);
9429 addl(tmp1, in2);
9430 addq(tmp1, in1);
9431
9432 BIND(L_wordByWord);
9433 cmpq(in1, tmp1);
9434 jcc(Assembler::greaterEqual, L_byteByByteProlog);
9435 crc32(in_out, Address(in1, 0), 4);
9436 addq(in1, 4);
9437 jmp(L_wordByWord);
9438
9439 BIND(L_byteByByteProlog);
9440 andl(in2, 0x00000007);
9441 movl(tmp2, 1);
9442
9443 BIND(L_byteByByte);
9444 cmpl(tmp2, in2);
9445 jccb(Assembler::greater, L_exit);
9446 crc32(in_out, Address(in1, 0), 1);
9447 incq(in1);
9448 incl(tmp2);
9449 jmp(L_byteByByte);
9450
9451 BIND(L_exit);
9452 }
9453 #else
9454 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
9455 Register tmp1, Register tmp2, Register tmp3,
9456 Register tmp4, Register tmp5, Register tmp6,
9457 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9458 bool is_pclmulqdq_supported) {
9459 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
9460 Label L_wordByWord;
9461 Label L_byteByByteProlog;
9462 Label L_byteByByte;
9463 Label L_exit;
9464
9465 if (is_pclmulqdq_supported) {
9466 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
9467 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1);
9468
9469 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
9470 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
9471
9472 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
9473 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
9474 } else {
9475 const_or_pre_comp_const_index[0] = 1;
9476 const_or_pre_comp_const_index[1] = 0;
9477
9478 const_or_pre_comp_const_index[2] = 3;
9479 const_or_pre_comp_const_index[3] = 2;
9480
9481 const_or_pre_comp_const_index[4] = 5;
9482 const_or_pre_comp_const_index[5] = 4;
9483 }
9484 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
9485 in2, in1, in_out,
9486 tmp1, tmp2, tmp3,
9487 w_xtmp1, w_xtmp2, w_xtmp3,
9488 tmp4, tmp5,
9489 tmp6);
9490 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
9491 in2, in1, in_out,
9492 tmp1, tmp2, tmp3,
9493 w_xtmp1, w_xtmp2, w_xtmp3,
9494 tmp4, tmp5,
9495 tmp6);
9496 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
9497 in2, in1, in_out,
9498 tmp1, tmp2, tmp3,
9499 w_xtmp1, w_xtmp2, w_xtmp3,
9500 tmp4, tmp5,
9501 tmp6);
9502 movl(tmp1, in2);
9503 andl(tmp1, 0x00000007);
9504 negl(tmp1);
9505 addl(tmp1, in2);
9506 addl(tmp1, in1);
9507
9508 BIND(L_wordByWord);
9509 cmpl(in1, tmp1);
9510 jcc(Assembler::greaterEqual, L_byteByByteProlog);
9511 crc32(in_out, Address(in1,0), 4);
9512 addl(in1, 4);
9513 jmp(L_wordByWord);
9514
9515 BIND(L_byteByByteProlog);
9516 andl(in2, 0x00000007);
9517 movl(tmp2, 1);
9518
9519 BIND(L_byteByByte);
9520 cmpl(tmp2, in2);
9521 jccb(Assembler::greater, L_exit);
9522 movb(tmp1, Address(in1, 0));
9523 crc32(in_out, tmp1, 1);
9524 incl(in1);
9525 incl(tmp2);
9526 jmp(L_byteByByte);
9527
9528 BIND(L_exit);
9529 }
9530 #endif // LP64
9531 #undef BIND
9532 #undef BLOCK_COMMENT
9533
9534 // Compress char[] array to byte[].
9535 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java
9536 // @HotSpotIntrinsicCandidate
9537 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) {
9538 // for (int i = 0; i < len; i++) {
9539 // int c = src[srcOff++];
9540 // if (c >>> 8 != 0) {
9541 // return 0;
9542 // }
9543 // dst[dstOff++] = (byte)c;
9544 // }
9545 // return len;
9546 // }
9547 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
9548 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
9549 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
9550 Register tmp5, Register result) {
9551 Label copy_chars_loop, return_length, return_zero, done;
9552
9553 // rsi: src
9554 // rdi: dst
9555 // rdx: len
9556 // rcx: tmp5
9557 // rax: result
9558
9559 // rsi holds start addr of source char[] to be compressed
9560 // rdi holds start addr of destination byte[]
9561 // rdx holds length
9562
9563 assert(len != result, "");
9564
9565 // save length for return
9566 push(len);
9567
9568 if ((UseAVX > 2) && // AVX512
9569 VM_Version::supports_avx512vlbw() &&
9570 VM_Version::supports_bmi2()) {
9571
9572 Label copy_32_loop, copy_loop_tail, below_threshold;
9573
9574 // alignment
9575 Label post_alignment;
9576
9577 // if length of the string is less than 16, handle it in an old fashioned way
9578 testl(len, -32);
9579 jcc(Assembler::zero, below_threshold);
9580
9581 // First check whether a character is compressable ( <= 0xFF).
9582 // Create mask to test for Unicode chars inside zmm vector
9583 movl(result, 0x00FF);
9584 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit);
9585
9586 testl(len, -64);
9587 jcc(Assembler::zero, post_alignment);
9588
9589 movl(tmp5, dst);
9590 andl(tmp5, (32 - 1));
9591 negl(tmp5);
9592 andl(tmp5, (32 - 1));
9593
9594 // bail out when there is nothing to be done
9595 testl(tmp5, 0xFFFFFFFF);
9596 jcc(Assembler::zero, post_alignment);
9597
9598 // ~(~0 << len), where len is the # of remaining elements to process
9599 movl(result, 0xFFFFFFFF);
9600 shlxl(result, result, tmp5);
9601 notl(result);
9602 kmovdl(k3, result);
9603
9604 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit);
9605 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9606 ktestd(k2, k3);
9607 jcc(Assembler::carryClear, return_zero);
9608
9609 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit);
9610
9611 addptr(src, tmp5);
9612 addptr(src, tmp5);
9613 addptr(dst, tmp5);
9614 subl(len, tmp5);
9615
9616 bind(post_alignment);
9617 // end of alignment
9618
9619 movl(tmp5, len);
9620 andl(tmp5, (32 - 1)); // tail count (in chars)
9621 andl(len, ~(32 - 1)); // vector count (in chars)
9622 jcc(Assembler::zero, copy_loop_tail);
9623
9624 lea(src, Address(src, len, Address::times_2));
9625 lea(dst, Address(dst, len, Address::times_1));
9626 negptr(len);
9627
9628 bind(copy_32_loop);
9629 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit);
9630 evpcmpuw(k2, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9631 kortestdl(k2, k2);
9632 jcc(Assembler::carryClear, return_zero);
9633
9634 // All elements in current processed chunk are valid candidates for
9635 // compression. Write a truncated byte elements to the memory.
9636 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit);
9637 addptr(len, 32);
9638 jcc(Assembler::notZero, copy_32_loop);
9639
9640 bind(copy_loop_tail);
9641 // bail out when there is nothing to be done
9642 testl(tmp5, 0xFFFFFFFF);
9643 jcc(Assembler::zero, return_length);
9644
9645 movl(len, tmp5);
9646
9647 // ~(~0 << len), where len is the # of remaining elements to process
9648 movl(result, 0xFFFFFFFF);
9649 shlxl(result, result, len);
9650 notl(result);
9651
9652 kmovdl(k3, result);
9653
9654 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit);
9655 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9656 ktestd(k2, k3);
9657 jcc(Assembler::carryClear, return_zero);
9658
9659 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit);
9660 jmp(return_length);
9661
9662 bind(below_threshold);
9663 }
9664
9665 if (UseSSE42Intrinsics) {
9666 Label copy_32_loop, copy_16, copy_tail;
9667
9668 movl(result, len);
9669
9670 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors
9671
9672 // vectored compression
9673 andl(len, 0xfffffff0); // vector count (in chars)
9674 andl(result, 0x0000000f); // tail count (in chars)
9675 testl(len, len);
9676 jcc(Assembler::zero, copy_16);
9677
9678 // compress 16 chars per iter
9679 movdl(tmp1Reg, tmp5);
9680 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
9681 pxor(tmp4Reg, tmp4Reg);
9682
9683 lea(src, Address(src, len, Address::times_2));
9684 lea(dst, Address(dst, len, Address::times_1));
9685 negptr(len);
9686
9687 bind(copy_32_loop);
9688 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters
9689 por(tmp4Reg, tmp2Reg);
9690 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters
9691 por(tmp4Reg, tmp3Reg);
9692 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector
9693 jcc(Assembler::notZero, return_zero);
9694 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte
9695 movdqu(Address(dst, len, Address::times_1), tmp2Reg);
9696 addptr(len, 16);
9697 jcc(Assembler::notZero, copy_32_loop);
9698
9699 // compress next vector of 8 chars (if any)
9700 bind(copy_16);
9701 movl(len, result);
9702 andl(len, 0xfffffff8); // vector count (in chars)
9703 andl(result, 0x00000007); // tail count (in chars)
9704 testl(len, len);
9705 jccb(Assembler::zero, copy_tail);
9706
9707 movdl(tmp1Reg, tmp5);
9708 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
9709 pxor(tmp3Reg, tmp3Reg);
9710
9711 movdqu(tmp2Reg, Address(src, 0));
9712 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
9713 jccb(Assembler::notZero, return_zero);
9714 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte
9715 movq(Address(dst, 0), tmp2Reg);
9716 addptr(src, 16);
9717 addptr(dst, 8);
9718
9719 bind(copy_tail);
9720 movl(len, result);
9721 }
9722 // compress 1 char per iter
9723 testl(len, len);
9724 jccb(Assembler::zero, return_length);
9725 lea(src, Address(src, len, Address::times_2));
9726 lea(dst, Address(dst, len, Address::times_1));
9727 negptr(len);
9728
9729 bind(copy_chars_loop);
9730 load_unsigned_short(result, Address(src, len, Address::times_2));
9731 testl(result, 0xff00); // check if Unicode char
9732 jccb(Assembler::notZero, return_zero);
9733 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte
9734 increment(len);
9735 jcc(Assembler::notZero, copy_chars_loop);
9736
9737 // if compression succeeded, return length
9738 bind(return_length);
9739 pop(result);
9740 jmpb(done);
9741
9742 // if compression failed, return 0
9743 bind(return_zero);
9744 xorl(result, result);
9745 addptr(rsp, wordSize);
9746
9747 bind(done);
9748 }
9749
9750 // Inflate byte[] array to char[].
9751 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java
9752 // @HotSpotIntrinsicCandidate
9753 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) {
9754 // for (int i = 0; i < len; i++) {
9755 // dst[dstOff++] = (char)(src[srcOff++] & 0xff);
9756 // }
9757 // }
9758 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
9759 XMMRegister tmp1, Register tmp2) {
9760 Label copy_chars_loop, done, below_threshold;
9761 // rsi: src
9762 // rdi: dst
9763 // rdx: len
9764 // rcx: tmp2
9765
9766 // rsi holds start addr of source byte[] to be inflated
9767 // rdi holds start addr of destination char[]
9768 // rdx holds length
9769 assert_different_registers(src, dst, len, tmp2);
9770
9771 if ((UseAVX > 2) && // AVX512
9772 VM_Version::supports_avx512vlbw() &&
9773 VM_Version::supports_bmi2()) {
9774
9775 Label copy_32_loop, copy_tail;
9776 Register tmp3_aliased = len;
9777
9778 // if length of the string is less than 16, handle it in an old fashioned way
9779 testl(len, -16);
9780 jcc(Assembler::zero, below_threshold);
9781
9782 // In order to use only one arithmetic operation for the main loop we use
9783 // this pre-calculation
9784 movl(tmp2, len);
9785 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop
9786 andl(len, -32); // vector count
9787 jccb(Assembler::zero, copy_tail);
9788
9789 lea(src, Address(src, len, Address::times_1));
9790 lea(dst, Address(dst, len, Address::times_2));
9791 negptr(len);
9792
9793
9794 // inflate 32 chars per iter
9795 bind(copy_32_loop);
9796 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit);
9797 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit);
9798 addptr(len, 32);
9799 jcc(Assembler::notZero, copy_32_loop);
9800
9801 bind(copy_tail);
9802 // bail out when there is nothing to be done
9803 testl(tmp2, -1); // we don't destroy the contents of tmp2 here
9804 jcc(Assembler::zero, done);
9805
9806 // ~(~0 << length), where length is the # of remaining elements to process
9807 movl(tmp3_aliased, -1);
9808 shlxl(tmp3_aliased, tmp3_aliased, tmp2);
9809 notl(tmp3_aliased);
9810 kmovdl(k2, tmp3_aliased);
9811 evpmovzxbw(tmp1, k2, Address(src, 0), Assembler::AVX_512bit);
9812 evmovdquw(Address(dst, 0), k2, tmp1, Assembler::AVX_512bit);
9813
9814 jmp(done);
9815 }
9816 if (UseSSE42Intrinsics) {
9817 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail;
9818
9819 movl(tmp2, len);
9820
9821 if (UseAVX > 1) {
9822 andl(tmp2, (16 - 1));
9823 andl(len, -16);
9824 jccb(Assembler::zero, copy_new_tail);
9825 } else {
9826 andl(tmp2, 0x00000007); // tail count (in chars)
9827 andl(len, 0xfffffff8); // vector count (in chars)
9828 jccb(Assembler::zero, copy_tail);
9829 }
9830
9831 // vectored inflation
9832 lea(src, Address(src, len, Address::times_1));
9833 lea(dst, Address(dst, len, Address::times_2));
9834 negptr(len);
9835
9836 if (UseAVX > 1) {
9837 bind(copy_16_loop);
9838 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit);
9839 vmovdqu(Address(dst, len, Address::times_2), tmp1);
9840 addptr(len, 16);
9841 jcc(Assembler::notZero, copy_16_loop);
9842
9843 bind(below_threshold);
9844 bind(copy_new_tail);
9845 if ((UseAVX > 2) &&
9846 VM_Version::supports_avx512vlbw() &&
9847 VM_Version::supports_bmi2()) {
9848 movl(tmp2, len);
9849 } else {
9850 movl(len, tmp2);
9851 }
9852 andl(tmp2, 0x00000007);
9853 andl(len, 0xFFFFFFF8);
9854 jccb(Assembler::zero, copy_tail);
9855
9856 pmovzxbw(tmp1, Address(src, 0));
9857 movdqu(Address(dst, 0), tmp1);
9858 addptr(src, 8);
9859 addptr(dst, 2 * 8);
9860
9861 jmp(copy_tail, true);
9862 }
9863
9864 // inflate 8 chars per iter
9865 bind(copy_8_loop);
9866 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words
9867 movdqu(Address(dst, len, Address::times_2), tmp1);
9868 addptr(len, 8);
9869 jcc(Assembler::notZero, copy_8_loop);
9870
9871 bind(copy_tail);
9872 movl(len, tmp2);
9873
9874 cmpl(len, 4);
9875 jccb(Assembler::less, copy_bytes);
9876
9877 movdl(tmp1, Address(src, 0)); // load 4 byte chars
9878 pmovzxbw(tmp1, tmp1);
9879 movq(Address(dst, 0), tmp1);
9880 subptr(len, 4);
9881 addptr(src, 4);
9882 addptr(dst, 8);
9883
9884 bind(copy_bytes);
9885 } else {
9886 bind(below_threshold);
9887 }
9888
9889 testl(len, len);
9890 jccb(Assembler::zero, done);
9891 lea(src, Address(src, len, Address::times_1));
9892 lea(dst, Address(dst, len, Address::times_2));
9893 negptr(len);
9894
9895 // inflate 1 char per iter
9896 bind(copy_chars_loop);
9897 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char
9898 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word
9899 increment(len);
9900 jcc(Assembler::notZero, copy_chars_loop);
9901
9902 bind(done);
9903 }
9904
9905 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
9906 switch (cond) {
9907 // Note some conditions are synonyms for others
9908 case Assembler::zero: return Assembler::notZero;
9909 case Assembler::notZero: return Assembler::zero;
9910 case Assembler::less: return Assembler::greaterEqual;
9911 case Assembler::lessEqual: return Assembler::greater;
9912 case Assembler::greater: return Assembler::lessEqual;
9913 case Assembler::greaterEqual: return Assembler::less;
9914 case Assembler::below: return Assembler::aboveEqual;
9915 case Assembler::belowEqual: return Assembler::above;
9916 case Assembler::above: return Assembler::belowEqual;
9917 case Assembler::aboveEqual: return Assembler::below;
9918 case Assembler::overflow: return Assembler::noOverflow;
9919 case Assembler::noOverflow: return Assembler::overflow;
9920 case Assembler::negative: return Assembler::positive;
9921 case Assembler::positive: return Assembler::negative;
9922 case Assembler::parity: return Assembler::noParity;
9923 case Assembler::noParity: return Assembler::parity;
9924 }
9925 ShouldNotReachHere(); return Assembler::overflow;
9926 }
9927
9928 SkipIfEqual::SkipIfEqual(
9929 MacroAssembler* masm, const bool* flag_addr, bool value) {
9930 _masm = masm;
9931 _masm->cmp8(ExternalAddress((address)flag_addr), value);
9932 _masm->jcc(Assembler::equal, _label);
9933 }
9934
9935 SkipIfEqual::~SkipIfEqual() {
9936 _masm->bind(_label);
9937 }
9938
9939 // 32-bit Windows has its own fast-path implementation
9940 // of get_thread
9941 #if !defined(WIN32) || defined(_LP64)
9942
9943 // This is simply a call to Thread::current()
9944 void MacroAssembler::get_thread(Register thread) {
9945 if (thread != rax) {
9946 push(rax);
9947 }
9948 LP64_ONLY(push(rdi);)
9949 LP64_ONLY(push(rsi);)
9950 push(rdx);
9951 push(rcx);
9952 #ifdef _LP64
9953 push(r8);
9954 push(r9);
9955 push(r10);
9956 push(r11);
9957 #endif
9958
9959 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0);
9960
9961 #ifdef _LP64
9962 pop(r11);
9963 pop(r10);
9964 pop(r9);
9965 pop(r8);
9966 #endif
9967 pop(rcx);
9968 pop(rdx);
9969 LP64_ONLY(pop(rsi);)
9970 LP64_ONLY(pop(rdi);)
9971 if (thread != rax) {
9972 mov(thread, rax);
9973 pop(rax);
9974 }
9975 }
9976
9977 #endif