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/klass.inline.hpp"
38 #include "prims/methodHandles.hpp"
39 #include "runtime/biasedLocking.hpp"
40 #include "runtime/flags/flagSetting.hpp"
41 #include "runtime/interfaceSupport.inline.hpp"
42 #include "runtime/objectMonitor.hpp"
43 #include "runtime/os.hpp"
44 #include "runtime/safepoint.hpp"
45 #include "runtime/safepointMechanism.hpp"
46 #include "runtime/sharedRuntime.hpp"
47 #include "runtime/stubRoutines.hpp"
48 #include "runtime/thread.hpp"
49 #include "utilities/macros.hpp"
50 #include "vmreg_x86.inline.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) {
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(rscratch1, src);
1014 Assembler::andpd(dst, Address(rscratch1, 0));
1015 }
1016 }
1017
1018 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) {
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(rscratch1, src);
1025 Assembler::andps(dst, Address(rscratch1, 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) {
2430 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2431 }
2432
2433 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2434 pass_arg0(this, arg_0);
2435 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2436 }
2437
2438 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2439
2440 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2441 pass_arg1(this, arg_1);
2442 pass_arg0(this, arg_0);
2443 MacroAssembler::call_VM_leaf_base(entry_point, 2);
2444 }
2445
2446 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2447 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2448 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2449 pass_arg2(this, arg_2);
2450 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2451 pass_arg1(this, arg_1);
2452 pass_arg0(this, arg_0);
2453 MacroAssembler::call_VM_leaf_base(entry_point, 3);
2454 }
2455
2456 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2457 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg"));
2458 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2459 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2460 pass_arg3(this, arg_3);
2461 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2462 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2463 pass_arg2(this, arg_2);
2464 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2465 pass_arg1(this, arg_1);
2466 pass_arg0(this, arg_0);
2467 MacroAssembler::call_VM_leaf_base(entry_point, 4);
2468 }
2469
2470 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
2471 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
2472 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD);
2473 verify_oop(oop_result, "broken oop in call_VM_base");
2474 }
2475
2476 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
2477 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
2478 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD);
2479 }
2480
2481 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2482 }
2483
2484 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2485 }
2486
2487 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) {
2488 if (reachable(src1)) {
2489 cmpl(as_Address(src1), imm);
2490 } else {
2491 lea(rscratch1, src1);
2492 cmpl(Address(rscratch1, 0), imm);
2493 }
2494 }
2495
2496 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) {
2497 assert(!src2.is_lval(), "use cmpptr");
2498 if (reachable(src2)) {
2499 cmpl(src1, as_Address(src2));
2500 } else {
2501 lea(rscratch1, src2);
2502 cmpl(src1, Address(rscratch1, 0));
2503 }
2504 }
2505
2506 void MacroAssembler::cmp32(Register src1, int32_t imm) {
2507 Assembler::cmpl(src1, imm);
2508 }
2509
2510 void MacroAssembler::cmp32(Register src1, Address src2) {
2511 Assembler::cmpl(src1, src2);
2512 }
2513
2514 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2515 ucomisd(opr1, opr2);
2516
2517 Label L;
2518 if (unordered_is_less) {
2519 movl(dst, -1);
2520 jcc(Assembler::parity, L);
2521 jcc(Assembler::below , L);
2522 movl(dst, 0);
2523 jcc(Assembler::equal , L);
2524 increment(dst);
2525 } else { // unordered is greater
2526 movl(dst, 1);
2527 jcc(Assembler::parity, L);
2528 jcc(Assembler::above , L);
2529 movl(dst, 0);
2530 jcc(Assembler::equal , L);
2531 decrementl(dst);
2532 }
2533 bind(L);
2534 }
2535
2536 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2537 ucomiss(opr1, opr2);
2538
2539 Label L;
2540 if (unordered_is_less) {
2541 movl(dst, -1);
2542 jcc(Assembler::parity, L);
2543 jcc(Assembler::below , L);
2544 movl(dst, 0);
2545 jcc(Assembler::equal , L);
2546 increment(dst);
2547 } else { // unordered is greater
2548 movl(dst, 1);
2549 jcc(Assembler::parity, L);
2550 jcc(Assembler::above , L);
2551 movl(dst, 0);
2552 jcc(Assembler::equal , L);
2553 decrementl(dst);
2554 }
2555 bind(L);
2556 }
2557
2558
2559 void MacroAssembler::cmp8(AddressLiteral src1, int imm) {
2560 if (reachable(src1)) {
2561 cmpb(as_Address(src1), imm);
2562 } else {
2563 lea(rscratch1, src1);
2564 cmpb(Address(rscratch1, 0), imm);
2565 }
2566 }
2567
2568 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) {
2569 #ifdef _LP64
2570 if (src2.is_lval()) {
2571 movptr(rscratch1, src2);
2572 Assembler::cmpq(src1, rscratch1);
2573 } else if (reachable(src2)) {
2574 cmpq(src1, as_Address(src2));
2575 } else {
2576 lea(rscratch1, src2);
2577 Assembler::cmpq(src1, Address(rscratch1, 0));
2578 }
2579 #else
2580 if (src2.is_lval()) {
2581 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2582 } else {
2583 cmpl(src1, as_Address(src2));
2584 }
2585 #endif // _LP64
2586 }
2587
2588 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) {
2589 assert(src2.is_lval(), "not a mem-mem compare");
2590 #ifdef _LP64
2591 // moves src2's literal address
2592 movptr(rscratch1, src2);
2593 Assembler::cmpq(src1, rscratch1);
2594 #else
2595 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2596 #endif // _LP64
2597 }
2598
2599 void MacroAssembler::cmpoop(Register src1, Register src2) {
2600 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2601 bs->obj_equals(this, src1, src2);
2602 }
2603
2604 void MacroAssembler::cmpoop(Register src1, Address src2) {
2605 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2606 bs->obj_equals(this, src1, src2);
2607 }
2608
2609 #ifdef _LP64
2610 void MacroAssembler::cmpoop(Register src1, jobject src2) {
2611 movoop(rscratch1, src2);
2612 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2613 bs->obj_equals(this, src1, rscratch1);
2614 }
2615 #endif
2616
2617 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) {
2618 if (reachable(adr)) {
2619 lock();
2620 cmpxchgptr(reg, as_Address(adr));
2621 } else {
2622 lea(rscratch1, adr);
2623 lock();
2624 cmpxchgptr(reg, Address(rscratch1, 0));
2625 }
2626 }
2627
2628 void MacroAssembler::cmpxchgptr(Register reg, Address adr) {
2629 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr));
2630 }
2631
2632 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) {
2633 if (reachable(src)) {
2634 Assembler::comisd(dst, as_Address(src));
2635 } else {
2636 lea(rscratch1, src);
2637 Assembler::comisd(dst, Address(rscratch1, 0));
2638 }
2639 }
2640
2641 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) {
2642 if (reachable(src)) {
2643 Assembler::comiss(dst, as_Address(src));
2644 } else {
2645 lea(rscratch1, src);
2646 Assembler::comiss(dst, Address(rscratch1, 0));
2647 }
2648 }
2649
2650
2651 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) {
2652 Condition negated_cond = negate_condition(cond);
2653 Label L;
2654 jcc(negated_cond, L);
2655 pushf(); // Preserve flags
2656 atomic_incl(counter_addr);
2657 popf();
2658 bind(L);
2659 }
2660
2661 int MacroAssembler::corrected_idivl(Register reg) {
2662 // Full implementation of Java idiv and irem; checks for
2663 // special case as described in JVM spec., p.243 & p.271.
2664 // The function returns the (pc) offset of the idivl
2665 // instruction - may be needed for implicit exceptions.
2666 //
2667 // normal case special case
2668 //
2669 // input : rax,: dividend min_int
2670 // reg: divisor (may not be rax,/rdx) -1
2671 //
2672 // output: rax,: quotient (= rax, idiv reg) min_int
2673 // rdx: remainder (= rax, irem reg) 0
2674 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register");
2675 const int min_int = 0x80000000;
2676 Label normal_case, special_case;
2677
2678 // check for special case
2679 cmpl(rax, min_int);
2680 jcc(Assembler::notEqual, normal_case);
2681 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0)
2682 cmpl(reg, -1);
2683 jcc(Assembler::equal, special_case);
2684
2685 // handle normal case
2686 bind(normal_case);
2687 cdql();
2688 int idivl_offset = offset();
2689 idivl(reg);
2690
2691 // normal and special case exit
2692 bind(special_case);
2693
2694 return idivl_offset;
2695 }
2696
2697
2698
2699 void MacroAssembler::decrementl(Register reg, int value) {
2700 if (value == min_jint) {subl(reg, value) ; return; }
2701 if (value < 0) { incrementl(reg, -value); return; }
2702 if (value == 0) { ; return; }
2703 if (value == 1 && UseIncDec) { decl(reg) ; return; }
2704 /* else */ { subl(reg, value) ; return; }
2705 }
2706
2707 void MacroAssembler::decrementl(Address dst, int value) {
2708 if (value == min_jint) {subl(dst, value) ; return; }
2709 if (value < 0) { incrementl(dst, -value); return; }
2710 if (value == 0) { ; return; }
2711 if (value == 1 && UseIncDec) { decl(dst) ; return; }
2712 /* else */ { subl(dst, value) ; return; }
2713 }
2714
2715 void MacroAssembler::division_with_shift (Register reg, int shift_value) {
2716 assert (shift_value > 0, "illegal shift value");
2717 Label _is_positive;
2718 testl (reg, reg);
2719 jcc (Assembler::positive, _is_positive);
2720 int offset = (1 << shift_value) - 1 ;
2721
2722 if (offset == 1) {
2723 incrementl(reg);
2724 } else {
2725 addl(reg, offset);
2726 }
2727
2728 bind (_is_positive);
2729 sarl(reg, shift_value);
2730 }
2731
2732 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) {
2733 if (reachable(src)) {
2734 Assembler::divsd(dst, as_Address(src));
2735 } else {
2736 lea(rscratch1, src);
2737 Assembler::divsd(dst, Address(rscratch1, 0));
2738 }
2739 }
2740
2741 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) {
2742 if (reachable(src)) {
2743 Assembler::divss(dst, as_Address(src));
2744 } else {
2745 lea(rscratch1, src);
2746 Assembler::divss(dst, Address(rscratch1, 0));
2747 }
2748 }
2749
2750 // !defined(COMPILER2) is because of stupid core builds
2751 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI
2752 void MacroAssembler::empty_FPU_stack() {
2753 if (VM_Version::supports_mmx()) {
2754 emms();
2755 } else {
2756 for (int i = 8; i-- > 0; ) ffree(i);
2757 }
2758 }
2759 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI
2760
2761
2762 void MacroAssembler::enter() {
2763 push(rbp);
2764 mov(rbp, rsp);
2765 }
2766
2767 // A 5 byte nop that is safe for patching (see patch_verified_entry)
2768 void MacroAssembler::fat_nop() {
2769 if (UseAddressNop) {
2770 addr_nop_5();
2771 } else {
2772 emit_int8(0x26); // es:
2773 emit_int8(0x2e); // cs:
2774 emit_int8(0x64); // fs:
2775 emit_int8(0x65); // gs:
2776 emit_int8((unsigned char)0x90);
2777 }
2778 }
2779
2780 void MacroAssembler::fcmp(Register tmp) {
2781 fcmp(tmp, 1, true, true);
2782 }
2783
2784 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) {
2785 assert(!pop_right || pop_left, "usage error");
2786 if (VM_Version::supports_cmov()) {
2787 assert(tmp == noreg, "unneeded temp");
2788 if (pop_left) {
2789 fucomip(index);
2790 } else {
2791 fucomi(index);
2792 }
2793 if (pop_right) {
2794 fpop();
2795 }
2796 } else {
2797 assert(tmp != noreg, "need temp");
2798 if (pop_left) {
2799 if (pop_right) {
2800 fcompp();
2801 } else {
2802 fcomp(index);
2803 }
2804 } else {
2805 fcom(index);
2806 }
2807 // convert FPU condition into eflags condition via rax,
2808 save_rax(tmp);
2809 fwait(); fnstsw_ax();
2810 sahf();
2811 restore_rax(tmp);
2812 }
2813 // condition codes set as follows:
2814 //
2815 // CF (corresponds to C0) if x < y
2816 // PF (corresponds to C2) if unordered
2817 // ZF (corresponds to C3) if x = y
2818 }
2819
2820 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) {
2821 fcmp2int(dst, unordered_is_less, 1, true, true);
2822 }
2823
2824 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) {
2825 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right);
2826 Label L;
2827 if (unordered_is_less) {
2828 movl(dst, -1);
2829 jcc(Assembler::parity, L);
2830 jcc(Assembler::below , L);
2831 movl(dst, 0);
2832 jcc(Assembler::equal , L);
2833 increment(dst);
2834 } else { // unordered is greater
2835 movl(dst, 1);
2836 jcc(Assembler::parity, L);
2837 jcc(Assembler::above , L);
2838 movl(dst, 0);
2839 jcc(Assembler::equal , L);
2840 decrementl(dst);
2841 }
2842 bind(L);
2843 }
2844
2845 void MacroAssembler::fld_d(AddressLiteral src) {
2846 fld_d(as_Address(src));
2847 }
2848
2849 void MacroAssembler::fld_s(AddressLiteral src) {
2850 fld_s(as_Address(src));
2851 }
2852
2853 void MacroAssembler::fld_x(AddressLiteral src) {
2854 Assembler::fld_x(as_Address(src));
2855 }
2856
2857 void MacroAssembler::fldcw(AddressLiteral src) {
2858 Assembler::fldcw(as_Address(src));
2859 }
2860
2861 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) {
2862 if (reachable(src)) {
2863 Assembler::mulpd(dst, as_Address(src));
2864 } else {
2865 lea(rscratch1, src);
2866 Assembler::mulpd(dst, Address(rscratch1, 0));
2867 }
2868 }
2869
2870 void MacroAssembler::increase_precision() {
2871 subptr(rsp, BytesPerWord);
2872 fnstcw(Address(rsp, 0));
2873 movl(rax, Address(rsp, 0));
2874 orl(rax, 0x300);
2875 push(rax);
2876 fldcw(Address(rsp, 0));
2877 pop(rax);
2878 }
2879
2880 void MacroAssembler::restore_precision() {
2881 fldcw(Address(rsp, 0));
2882 addptr(rsp, BytesPerWord);
2883 }
2884
2885 void MacroAssembler::fpop() {
2886 ffree();
2887 fincstp();
2888 }
2889
2890 void MacroAssembler::load_float(Address src) {
2891 if (UseSSE >= 1) {
2892 movflt(xmm0, src);
2893 } else {
2894 LP64_ONLY(ShouldNotReachHere());
2895 NOT_LP64(fld_s(src));
2896 }
2897 }
2898
2899 void MacroAssembler::store_float(Address dst) {
2900 if (UseSSE >= 1) {
2901 movflt(dst, xmm0);
2902 } else {
2903 LP64_ONLY(ShouldNotReachHere());
2904 NOT_LP64(fstp_s(dst));
2905 }
2906 }
2907
2908 void MacroAssembler::load_double(Address src) {
2909 if (UseSSE >= 2) {
2910 movdbl(xmm0, src);
2911 } else {
2912 LP64_ONLY(ShouldNotReachHere());
2913 NOT_LP64(fld_d(src));
2914 }
2915 }
2916
2917 void MacroAssembler::store_double(Address dst) {
2918 if (UseSSE >= 2) {
2919 movdbl(dst, xmm0);
2920 } else {
2921 LP64_ONLY(ShouldNotReachHere());
2922 NOT_LP64(fstp_d(dst));
2923 }
2924 }
2925
2926 void MacroAssembler::fremr(Register tmp) {
2927 save_rax(tmp);
2928 { Label L;
2929 bind(L);
2930 fprem();
2931 fwait(); fnstsw_ax();
2932 #ifdef _LP64
2933 testl(rax, 0x400);
2934 jcc(Assembler::notEqual, L);
2935 #else
2936 sahf();
2937 jcc(Assembler::parity, L);
2938 #endif // _LP64
2939 }
2940 restore_rax(tmp);
2941 // Result is in ST0.
2942 // Note: fxch & fpop to get rid of ST1
2943 // (otherwise FPU stack could overflow eventually)
2944 fxch(1);
2945 fpop();
2946 }
2947
2948 // dst = c = a * b + c
2949 void MacroAssembler::fmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) {
2950 Assembler::vfmadd231sd(c, a, b);
2951 if (dst != c) {
2952 movdbl(dst, c);
2953 }
2954 }
2955
2956 // dst = c = a * b + c
2957 void MacroAssembler::fmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c) {
2958 Assembler::vfmadd231ss(c, a, b);
2959 if (dst != c) {
2960 movflt(dst, c);
2961 }
2962 }
2963
2964 // dst = c = a * b + c
2965 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) {
2966 Assembler::vfmadd231pd(c, a, b, vector_len);
2967 if (dst != c) {
2968 vmovdqu(dst, c);
2969 }
2970 }
2971
2972 // dst = c = a * b + c
2973 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, XMMRegister b, XMMRegister c, int vector_len) {
2974 Assembler::vfmadd231ps(c, a, b, vector_len);
2975 if (dst != c) {
2976 vmovdqu(dst, c);
2977 }
2978 }
2979
2980 // dst = c = a * b + c
2981 void MacroAssembler::vfmad(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) {
2982 Assembler::vfmadd231pd(c, a, b, vector_len);
2983 if (dst != c) {
2984 vmovdqu(dst, c);
2985 }
2986 }
2987
2988 // dst = c = a * b + c
2989 void MacroAssembler::vfmaf(XMMRegister dst, XMMRegister a, Address b, XMMRegister c, int vector_len) {
2990 Assembler::vfmadd231ps(c, a, b, vector_len);
2991 if (dst != c) {
2992 vmovdqu(dst, c);
2993 }
2994 }
2995
2996 void MacroAssembler::incrementl(AddressLiteral dst) {
2997 if (reachable(dst)) {
2998 incrementl(as_Address(dst));
2999 } else {
3000 lea(rscratch1, dst);
3001 incrementl(Address(rscratch1, 0));
3002 }
3003 }
3004
3005 void MacroAssembler::incrementl(ArrayAddress dst) {
3006 incrementl(as_Address(dst));
3007 }
3008
3009 void MacroAssembler::incrementl(Register reg, int value) {
3010 if (value == min_jint) {addl(reg, value) ; return; }
3011 if (value < 0) { decrementl(reg, -value); return; }
3012 if (value == 0) { ; return; }
3013 if (value == 1 && UseIncDec) { incl(reg) ; return; }
3014 /* else */ { addl(reg, value) ; return; }
3015 }
3016
3017 void MacroAssembler::incrementl(Address dst, int value) {
3018 if (value == min_jint) {addl(dst, value) ; return; }
3019 if (value < 0) { decrementl(dst, -value); return; }
3020 if (value == 0) { ; return; }
3021 if (value == 1 && UseIncDec) { incl(dst) ; return; }
3022 /* else */ { addl(dst, value) ; return; }
3023 }
3024
3025 void MacroAssembler::jump(AddressLiteral dst) {
3026 if (reachable(dst)) {
3027 jmp_literal(dst.target(), dst.rspec());
3028 } else {
3029 lea(rscratch1, dst);
3030 jmp(rscratch1);
3031 }
3032 }
3033
3034 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) {
3035 if (reachable(dst)) {
3036 InstructionMark im(this);
3037 relocate(dst.reloc());
3038 const int short_size = 2;
3039 const int long_size = 6;
3040 int offs = (intptr_t)dst.target() - ((intptr_t)pc());
3041 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
3042 // 0111 tttn #8-bit disp
3043 emit_int8(0x70 | cc);
3044 emit_int8((offs - short_size) & 0xFF);
3045 } else {
3046 // 0000 1111 1000 tttn #32-bit disp
3047 emit_int8(0x0F);
3048 emit_int8((unsigned char)(0x80 | cc));
3049 emit_int32(offs - long_size);
3050 }
3051 } else {
3052 #ifdef ASSERT
3053 warning("reversing conditional branch");
3054 #endif /* ASSERT */
3055 Label skip;
3056 jccb(reverse[cc], skip);
3057 lea(rscratch1, dst);
3058 Assembler::jmp(rscratch1);
3059 bind(skip);
3060 }
3061 }
3062
3063 void MacroAssembler::ldmxcsr(AddressLiteral src) {
3064 if (reachable(src)) {
3065 Assembler::ldmxcsr(as_Address(src));
3066 } else {
3067 lea(rscratch1, src);
3068 Assembler::ldmxcsr(Address(rscratch1, 0));
3069 }
3070 }
3071
3072 int MacroAssembler::load_signed_byte(Register dst, Address src) {
3073 int off;
3074 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3075 off = offset();
3076 movsbl(dst, src); // movsxb
3077 } else {
3078 off = load_unsigned_byte(dst, src);
3079 shll(dst, 24);
3080 sarl(dst, 24);
3081 }
3082 return off;
3083 }
3084
3085 // Note: load_signed_short used to be called load_signed_word.
3086 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler
3087 // manual, which means 16 bits, that usage is found nowhere in HotSpot code.
3088 // The term "word" in HotSpot means a 32- or 64-bit machine word.
3089 int MacroAssembler::load_signed_short(Register dst, Address src) {
3090 int off;
3091 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3092 // This is dubious to me since it seems safe to do a signed 16 => 64 bit
3093 // version but this is what 64bit has always done. This seems to imply
3094 // that users are only using 32bits worth.
3095 off = offset();
3096 movswl(dst, src); // movsxw
3097 } else {
3098 off = load_unsigned_short(dst, src);
3099 shll(dst, 16);
3100 sarl(dst, 16);
3101 }
3102 return off;
3103 }
3104
3105 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
3106 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3107 // and "3.9 Partial Register Penalties", p. 22).
3108 int off;
3109 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) {
3110 off = offset();
3111 movzbl(dst, src); // movzxb
3112 } else {
3113 xorl(dst, dst);
3114 off = offset();
3115 movb(dst, src);
3116 }
3117 return off;
3118 }
3119
3120 // Note: load_unsigned_short used to be called load_unsigned_word.
3121 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
3122 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3123 // and "3.9 Partial Register Penalties", p. 22).
3124 int off;
3125 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) {
3126 off = offset();
3127 movzwl(dst, src); // movzxw
3128 } else {
3129 xorl(dst, dst);
3130 off = offset();
3131 movw(dst, src);
3132 }
3133 return off;
3134 }
3135
3136 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
3137 switch (size_in_bytes) {
3138 #ifndef _LP64
3139 case 8:
3140 assert(dst2 != noreg, "second dest register required");
3141 movl(dst, src);
3142 movl(dst2, src.plus_disp(BytesPerInt));
3143 break;
3144 #else
3145 case 8: movq(dst, src); break;
3146 #endif
3147 case 4: movl(dst, src); break;
3148 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
3149 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
3150 default: ShouldNotReachHere();
3151 }
3152 }
3153
3154 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
3155 switch (size_in_bytes) {
3156 #ifndef _LP64
3157 case 8:
3158 assert(src2 != noreg, "second source register required");
3159 movl(dst, src);
3160 movl(dst.plus_disp(BytesPerInt), src2);
3161 break;
3162 #else
3163 case 8: movq(dst, src); break;
3164 #endif
3165 case 4: movl(dst, src); break;
3166 case 2: movw(dst, src); break;
3167 case 1: movb(dst, src); break;
3168 default: ShouldNotReachHere();
3169 }
3170 }
3171
3172 void MacroAssembler::mov32(AddressLiteral dst, Register src) {
3173 if (reachable(dst)) {
3174 movl(as_Address(dst), src);
3175 } else {
3176 lea(rscratch1, dst);
3177 movl(Address(rscratch1, 0), src);
3178 }
3179 }
3180
3181 void MacroAssembler::mov32(Register dst, AddressLiteral src) {
3182 if (reachable(src)) {
3183 movl(dst, as_Address(src));
3184 } else {
3185 lea(rscratch1, src);
3186 movl(dst, Address(rscratch1, 0));
3187 }
3188 }
3189
3190 // C++ bool manipulation
3191
3192 void MacroAssembler::movbool(Register dst, Address src) {
3193 if(sizeof(bool) == 1)
3194 movb(dst, src);
3195 else if(sizeof(bool) == 2)
3196 movw(dst, src);
3197 else if(sizeof(bool) == 4)
3198 movl(dst, src);
3199 else
3200 // unsupported
3201 ShouldNotReachHere();
3202 }
3203
3204 void MacroAssembler::movbool(Address dst, bool boolconst) {
3205 if(sizeof(bool) == 1)
3206 movb(dst, (int) boolconst);
3207 else if(sizeof(bool) == 2)
3208 movw(dst, (int) boolconst);
3209 else if(sizeof(bool) == 4)
3210 movl(dst, (int) boolconst);
3211 else
3212 // unsupported
3213 ShouldNotReachHere();
3214 }
3215
3216 void MacroAssembler::movbool(Address dst, Register src) {
3217 if(sizeof(bool) == 1)
3218 movb(dst, src);
3219 else if(sizeof(bool) == 2)
3220 movw(dst, src);
3221 else if(sizeof(bool) == 4)
3222 movl(dst, src);
3223 else
3224 // unsupported
3225 ShouldNotReachHere();
3226 }
3227
3228 void MacroAssembler::movbyte(ArrayAddress dst, int src) {
3229 movb(as_Address(dst), src);
3230 }
3231
3232 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) {
3233 if (reachable(src)) {
3234 movdl(dst, as_Address(src));
3235 } else {
3236 lea(rscratch1, src);
3237 movdl(dst, Address(rscratch1, 0));
3238 }
3239 }
3240
3241 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) {
3242 if (reachable(src)) {
3243 movq(dst, as_Address(src));
3244 } else {
3245 lea(rscratch1, src);
3246 movq(dst, Address(rscratch1, 0));
3247 }
3248 }
3249
3250 #ifdef COMPILER2
3251 void MacroAssembler::setvectmask(Register dst, Register src) {
3252 guarantee(PostLoopMultiversioning, "must be");
3253 Assembler::movl(dst, 1);
3254 Assembler::shlxl(dst, dst, src);
3255 Assembler::decl(dst);
3256 Assembler::kmovdl(k1, dst);
3257 Assembler::movl(dst, src);
3258 }
3259
3260 void MacroAssembler::restorevectmask() {
3261 guarantee(PostLoopMultiversioning, "must be");
3262 Assembler::knotwl(k1, k0);
3263 }
3264 #endif // COMPILER2
3265
3266 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) {
3267 if (reachable(src)) {
3268 if (UseXmmLoadAndClearUpper) {
3269 movsd (dst, as_Address(src));
3270 } else {
3271 movlpd(dst, as_Address(src));
3272 }
3273 } else {
3274 lea(rscratch1, src);
3275 if (UseXmmLoadAndClearUpper) {
3276 movsd (dst, Address(rscratch1, 0));
3277 } else {
3278 movlpd(dst, Address(rscratch1, 0));
3279 }
3280 }
3281 }
3282
3283 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) {
3284 if (reachable(src)) {
3285 movss(dst, as_Address(src));
3286 } else {
3287 lea(rscratch1, src);
3288 movss(dst, Address(rscratch1, 0));
3289 }
3290 }
3291
3292 void MacroAssembler::movptr(Register dst, Register src) {
3293 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3294 }
3295
3296 void MacroAssembler::movptr(Register dst, Address src) {
3297 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3298 }
3299
3300 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
3301 void MacroAssembler::movptr(Register dst, intptr_t src) {
3302 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src));
3303 }
3304
3305 void MacroAssembler::movptr(Address dst, Register src) {
3306 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3307 }
3308
3309 void MacroAssembler::movdqu(Address dst, XMMRegister src) {
3310 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3311 Assembler::movdqu(dst, src);
3312 }
3313
3314 void MacroAssembler::movdqu(XMMRegister dst, Address src) {
3315 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3316 Assembler::movdqu(dst, src);
3317 }
3318
3319 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) {
3320 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3321 Assembler::movdqu(dst, src);
3322 }
3323
3324 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src, Register scratchReg) {
3325 if (reachable(src)) {
3326 movdqu(dst, as_Address(src));
3327 } else {
3328 lea(scratchReg, src);
3329 movdqu(dst, Address(scratchReg, 0));
3330 }
3331 }
3332
3333 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) {
3334 assert(((src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3335 Assembler::vmovdqu(dst, src);
3336 }
3337
3338 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) {
3339 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3340 Assembler::vmovdqu(dst, src);
3341 }
3342
3343 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) {
3344 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3345 Assembler::vmovdqu(dst, src);
3346 }
3347
3348 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) {
3349 if (reachable(src)) {
3350 vmovdqu(dst, as_Address(src));
3351 }
3352 else {
3353 lea(rscratch1, src);
3354 vmovdqu(dst, Address(rscratch1, 0));
3355 }
3356 }
3357
3358 void MacroAssembler::evmovdquq(XMMRegister dst, AddressLiteral src, int vector_len, Register rscratch) {
3359 if (reachable(src)) {
3360 Assembler::evmovdquq(dst, as_Address(src), vector_len);
3361 } else {
3362 lea(rscratch, src);
3363 Assembler::evmovdquq(dst, Address(rscratch, 0), vector_len);
3364 }
3365 }
3366
3367 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) {
3368 if (reachable(src)) {
3369 Assembler::movdqa(dst, as_Address(src));
3370 } else {
3371 lea(rscratch1, src);
3372 Assembler::movdqa(dst, Address(rscratch1, 0));
3373 }
3374 }
3375
3376 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) {
3377 if (reachable(src)) {
3378 Assembler::movsd(dst, as_Address(src));
3379 } else {
3380 lea(rscratch1, src);
3381 Assembler::movsd(dst, Address(rscratch1, 0));
3382 }
3383 }
3384
3385 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) {
3386 if (reachable(src)) {
3387 Assembler::movss(dst, as_Address(src));
3388 } else {
3389 lea(rscratch1, src);
3390 Assembler::movss(dst, Address(rscratch1, 0));
3391 }
3392 }
3393
3394 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) {
3395 if (reachable(src)) {
3396 Assembler::mulsd(dst, as_Address(src));
3397 } else {
3398 lea(rscratch1, src);
3399 Assembler::mulsd(dst, Address(rscratch1, 0));
3400 }
3401 }
3402
3403 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) {
3404 if (reachable(src)) {
3405 Assembler::mulss(dst, as_Address(src));
3406 } else {
3407 lea(rscratch1, src);
3408 Assembler::mulss(dst, Address(rscratch1, 0));
3409 }
3410 }
3411
3412 void MacroAssembler::null_check(Register reg, int offset) {
3413 if (needs_explicit_null_check(offset)) {
3414 // provoke OS NULL exception if reg = NULL by
3415 // accessing M[reg] w/o changing any (non-CC) registers
3416 // NOTE: cmpl is plenty here to provoke a segv
3417 cmpptr(rax, Address(reg, 0));
3418 // Note: should probably use testl(rax, Address(reg, 0));
3419 // may be shorter code (however, this version of
3420 // testl needs to be implemented first)
3421 } else {
3422 // nothing to do, (later) access of M[reg + offset]
3423 // will provoke OS NULL exception if reg = NULL
3424 }
3425 }
3426
3427 void MacroAssembler::test_klass_is_value(Register klass, Register temp_reg, Label& is_value) {
3428 movl(temp_reg, Address(klass, Klass::access_flags_offset()));
3429 testl(temp_reg, JVM_ACC_VALUE);
3430 jcc(Assembler::notZero, is_value);
3431 }
3432
3433 void MacroAssembler::test_field_is_flattenable(Register flags, Register temp_reg, Label& is_flattenable) {
3434 movl(temp_reg, flags);
3435 shrl(temp_reg, ConstantPoolCacheEntry::is_flattenable_field_shift);
3436 andl(temp_reg, 0x1);
3437 testl(temp_reg, temp_reg);
3438 jcc(Assembler::notZero, is_flattenable);
3439 }
3440
3441 void MacroAssembler::test_field_is_not_flattenable(Register flags, Register temp_reg, Label& notFlattenable) {
3442 movl(temp_reg, flags);
3443 shrl(temp_reg, ConstantPoolCacheEntry::is_flattenable_field_shift);
3444 andl(temp_reg, 0x1);
3445 testl(temp_reg, temp_reg);
3446 jcc(Assembler::zero, notFlattenable);
3447 }
3448
3449 void MacroAssembler::test_field_is_flattened(Register flags, Register temp_reg, Label& is_flattened) {
3450 movl(temp_reg, flags);
3451 shrl(temp_reg, ConstantPoolCacheEntry::is_flattened_field_shift);
3452 andl(temp_reg, 0x1);
3453 testl(temp_reg, temp_reg);
3454 jcc(Assembler::notZero, is_flattened);
3455 }
3456
3457 void MacroAssembler::test_flattened_array_oop(Register oop, Register temp_reg,
3458 Label&is_flattened_array) {
3459 load_storage_props(temp_reg, oop);
3460 testb(temp_reg, ArrayStorageProperties::flattened_value);
3461 jcc(Assembler::notZero, is_flattened_array);
3462 }
3463
3464 void MacroAssembler::test_null_free_array_oop(Register oop, Register temp_reg, Label&is_null_free_array) {
3465 load_storage_props(temp_reg, oop);
3466 testb(temp_reg, ArrayStorageProperties::null_free_value);
3467 jcc(Assembler::notZero, is_null_free_array);
3468 }
3469
3470 void MacroAssembler::os_breakpoint() {
3471 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability
3472 // (e.g., MSVC can't call ps() otherwise)
3473 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint)));
3474 }
3475
3476 void MacroAssembler::unimplemented(const char* what) {
3477 const char* buf = NULL;
3478 {
3479 ResourceMark rm;
3480 stringStream ss;
3481 ss.print("unimplemented: %s", what);
3482 buf = code_string(ss.as_string());
3483 }
3484 stop(buf);
3485 }
3486
3487 #ifdef _LP64
3488 #define XSTATE_BV 0x200
3489 #endif
3490
3491 void MacroAssembler::pop_CPU_state() {
3492 pop_FPU_state();
3493 pop_IU_state();
3494 }
3495
3496 void MacroAssembler::pop_FPU_state() {
3497 #ifndef _LP64
3498 frstor(Address(rsp, 0));
3499 #else
3500 fxrstor(Address(rsp, 0));
3501 #endif
3502 addptr(rsp, FPUStateSizeInWords * wordSize);
3503 }
3504
3505 void MacroAssembler::pop_IU_state() {
3506 popa();
3507 LP64_ONLY(addq(rsp, 8));
3508 popf();
3509 }
3510
3511 // Save Integer and Float state
3512 // Warning: Stack must be 16 byte aligned (64bit)
3513 void MacroAssembler::push_CPU_state() {
3514 push_IU_state();
3515 push_FPU_state();
3516 }
3517
3518 void MacroAssembler::push_FPU_state() {
3519 subptr(rsp, FPUStateSizeInWords * wordSize);
3520 #ifndef _LP64
3521 fnsave(Address(rsp, 0));
3522 fwait();
3523 #else
3524 fxsave(Address(rsp, 0));
3525 #endif // LP64
3526 }
3527
3528 void MacroAssembler::push_IU_state() {
3529 // Push flags first because pusha kills them
3530 pushf();
3531 // Make sure rsp stays 16-byte aligned
3532 LP64_ONLY(subq(rsp, 8));
3533 pusha();
3534 }
3535
3536 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp) { // determine java_thread register
3537 if (!java_thread->is_valid()) {
3538 java_thread = rdi;
3539 get_thread(java_thread);
3540 }
3541 // we must set sp to zero to clear frame
3542 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
3543 if (clear_fp) {
3544 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3545 }
3546
3547 // Always clear the pc because it could have been set by make_walkable()
3548 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
3549
3550 vzeroupper();
3551 }
3552
3553 void MacroAssembler::restore_rax(Register tmp) {
3554 if (tmp == noreg) pop(rax);
3555 else if (tmp != rax) mov(rax, tmp);
3556 }
3557
3558 void MacroAssembler::round_to(Register reg, int modulus) {
3559 addptr(reg, modulus - 1);
3560 andptr(reg, -modulus);
3561 }
3562
3563 void MacroAssembler::save_rax(Register tmp) {
3564 if (tmp == noreg) push(rax);
3565 else if (tmp != rax) mov(tmp, rax);
3566 }
3567
3568 void MacroAssembler::safepoint_poll(Label& slow_path, Register thread_reg, Register temp_reg) {
3569 if (SafepointMechanism::uses_thread_local_poll()) {
3570 #ifdef _LP64
3571 assert(thread_reg == r15_thread, "should be");
3572 #else
3573 if (thread_reg == noreg) {
3574 thread_reg = temp_reg;
3575 get_thread(thread_reg);
3576 }
3577 #endif
3578 testb(Address(thread_reg, Thread::polling_page_offset()), SafepointMechanism::poll_bit());
3579 jcc(Assembler::notZero, slow_path); // handshake bit set implies poll
3580 } else {
3581 cmp32(ExternalAddress(SafepointSynchronize::address_of_state()),
3582 SafepointSynchronize::_not_synchronized);
3583 jcc(Assembler::notEqual, slow_path);
3584 }
3585 }
3586
3587 // Calls to C land
3588 //
3589 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded
3590 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
3591 // has to be reset to 0. This is required to allow proper stack traversal.
3592 void MacroAssembler::set_last_Java_frame(Register java_thread,
3593 Register last_java_sp,
3594 Register last_java_fp,
3595 address last_java_pc) {
3596 vzeroupper();
3597 // determine java_thread register
3598 if (!java_thread->is_valid()) {
3599 java_thread = rdi;
3600 get_thread(java_thread);
3601 }
3602 // determine last_java_sp register
3603 if (!last_java_sp->is_valid()) {
3604 last_java_sp = rsp;
3605 }
3606
3607 // last_java_fp is optional
3608
3609 if (last_java_fp->is_valid()) {
3610 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp);
3611 }
3612
3613 // last_java_pc is optional
3614
3615 if (last_java_pc != NULL) {
3616 lea(Address(java_thread,
3617 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()),
3618 InternalAddress(last_java_pc));
3619
3620 }
3621 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
3622 }
3623
3624 void MacroAssembler::shlptr(Register dst, int imm8) {
3625 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8));
3626 }
3627
3628 void MacroAssembler::shrptr(Register dst, int imm8) {
3629 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8));
3630 }
3631
3632 void MacroAssembler::sign_extend_byte(Register reg) {
3633 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) {
3634 movsbl(reg, reg); // movsxb
3635 } else {
3636 shll(reg, 24);
3637 sarl(reg, 24);
3638 }
3639 }
3640
3641 void MacroAssembler::sign_extend_short(Register reg) {
3642 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3643 movswl(reg, reg); // movsxw
3644 } else {
3645 shll(reg, 16);
3646 sarl(reg, 16);
3647 }
3648 }
3649
3650 void MacroAssembler::testl(Register dst, AddressLiteral src) {
3651 assert(reachable(src), "Address should be reachable");
3652 testl(dst, as_Address(src));
3653 }
3654
3655 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) {
3656 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3657 Assembler::pcmpeqb(dst, src);
3658 }
3659
3660 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) {
3661 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3662 Assembler::pcmpeqw(dst, src);
3663 }
3664
3665 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) {
3666 assert((dst->encoding() < 16),"XMM register should be 0-15");
3667 Assembler::pcmpestri(dst, src, imm8);
3668 }
3669
3670 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) {
3671 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3672 Assembler::pcmpestri(dst, src, imm8);
3673 }
3674
3675 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) {
3676 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3677 Assembler::pmovzxbw(dst, src);
3678 }
3679
3680 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) {
3681 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3682 Assembler::pmovzxbw(dst, src);
3683 }
3684
3685 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) {
3686 assert((src->encoding() < 16),"XMM register should be 0-15");
3687 Assembler::pmovmskb(dst, src);
3688 }
3689
3690 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) {
3691 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3692 Assembler::ptest(dst, src);
3693 }
3694
3695 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) {
3696 if (reachable(src)) {
3697 Assembler::sqrtsd(dst, as_Address(src));
3698 } else {
3699 lea(rscratch1, src);
3700 Assembler::sqrtsd(dst, Address(rscratch1, 0));
3701 }
3702 }
3703
3704 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) {
3705 if (reachable(src)) {
3706 Assembler::sqrtss(dst, as_Address(src));
3707 } else {
3708 lea(rscratch1, src);
3709 Assembler::sqrtss(dst, Address(rscratch1, 0));
3710 }
3711 }
3712
3713 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) {
3714 if (reachable(src)) {
3715 Assembler::subsd(dst, as_Address(src));
3716 } else {
3717 lea(rscratch1, src);
3718 Assembler::subsd(dst, Address(rscratch1, 0));
3719 }
3720 }
3721
3722 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
3723 if (reachable(src)) {
3724 Assembler::subss(dst, as_Address(src));
3725 } else {
3726 lea(rscratch1, src);
3727 Assembler::subss(dst, Address(rscratch1, 0));
3728 }
3729 }
3730
3731 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
3732 if (reachable(src)) {
3733 Assembler::ucomisd(dst, as_Address(src));
3734 } else {
3735 lea(rscratch1, src);
3736 Assembler::ucomisd(dst, Address(rscratch1, 0));
3737 }
3738 }
3739
3740 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
3741 if (reachable(src)) {
3742 Assembler::ucomiss(dst, as_Address(src));
3743 } else {
3744 lea(rscratch1, src);
3745 Assembler::ucomiss(dst, Address(rscratch1, 0));
3746 }
3747 }
3748
3749 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) {
3750 // Used in sign-bit flipping with aligned address.
3751 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3752 if (reachable(src)) {
3753 Assembler::xorpd(dst, as_Address(src));
3754 } else {
3755 lea(rscratch1, src);
3756 Assembler::xorpd(dst, Address(rscratch1, 0));
3757 }
3758 }
3759
3760 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) {
3761 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
3762 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
3763 }
3764 else {
3765 Assembler::xorpd(dst, src);
3766 }
3767 }
3768
3769 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) {
3770 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
3771 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
3772 } else {
3773 Assembler::xorps(dst, src);
3774 }
3775 }
3776
3777 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) {
3778 // Used in sign-bit flipping with aligned address.
3779 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
3780 if (reachable(src)) {
3781 Assembler::xorps(dst, as_Address(src));
3782 } else {
3783 lea(rscratch1, src);
3784 Assembler::xorps(dst, Address(rscratch1, 0));
3785 }
3786 }
3787
3788 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
3789 // Used in sign-bit flipping with aligned address.
3790 bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
3791 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
3792 if (reachable(src)) {
3793 Assembler::pshufb(dst, as_Address(src));
3794 } else {
3795 lea(rscratch1, src);
3796 Assembler::pshufb(dst, Address(rscratch1, 0));
3797 }
3798 }
3799
3800 // AVX 3-operands instructions
3801
3802 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3803 if (reachable(src)) {
3804 vaddsd(dst, nds, as_Address(src));
3805 } else {
3806 lea(rscratch1, src);
3807 vaddsd(dst, nds, Address(rscratch1, 0));
3808 }
3809 }
3810
3811 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3812 if (reachable(src)) {
3813 vaddss(dst, nds, as_Address(src));
3814 } else {
3815 lea(rscratch1, src);
3816 vaddss(dst, nds, Address(rscratch1, 0));
3817 }
3818 }
3819
3820 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
3821 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
3822 vandps(dst, nds, negate_field, vector_len);
3823 }
3824
3825 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
3826 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
3827 vandpd(dst, nds, negate_field, vector_len);
3828 }
3829
3830 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3831 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3832 Assembler::vpaddb(dst, nds, src, vector_len);
3833 }
3834
3835 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3836 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3837 Assembler::vpaddb(dst, nds, src, vector_len);
3838 }
3839
3840 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3841 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3842 Assembler::vpaddw(dst, nds, src, vector_len);
3843 }
3844
3845 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3846 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3847 Assembler::vpaddw(dst, nds, src, vector_len);
3848 }
3849
3850 void MacroAssembler::vpand(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
3851 if (reachable(src)) {
3852 Assembler::vpand(dst, nds, as_Address(src), vector_len);
3853 } else {
3854 lea(rscratch1, src);
3855 Assembler::vpand(dst, nds, Address(rscratch1, 0), vector_len);
3856 }
3857 }
3858
3859 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src, int vector_len) {
3860 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3861 Assembler::vpbroadcastw(dst, src, vector_len);
3862 }
3863
3864 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3865 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3866 Assembler::vpcmpeqb(dst, nds, src, vector_len);
3867 }
3868
3869 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3870 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3871 Assembler::vpcmpeqw(dst, nds, src, vector_len);
3872 }
3873
3874 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) {
3875 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3876 Assembler::vpmovzxbw(dst, src, vector_len);
3877 }
3878
3879 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) {
3880 assert((src->encoding() < 16),"XMM register should be 0-15");
3881 Assembler::vpmovmskb(dst, src);
3882 }
3883
3884 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3885 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3886 Assembler::vpmullw(dst, nds, src, vector_len);
3887 }
3888
3889 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3890 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3891 Assembler::vpmullw(dst, nds, src, vector_len);
3892 }
3893
3894 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3895 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3896 Assembler::vpsubb(dst, nds, src, vector_len);
3897 }
3898
3899 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3900 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3901 Assembler::vpsubb(dst, nds, src, vector_len);
3902 }
3903
3904 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
3905 assert(((dst->encoding() < 16 && src->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3906 Assembler::vpsubw(dst, nds, src, vector_len);
3907 }
3908
3909 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
3910 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3911 Assembler::vpsubw(dst, nds, src, vector_len);
3912 }
3913
3914 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3915 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3916 Assembler::vpsraw(dst, nds, shift, vector_len);
3917 }
3918
3919 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3920 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3921 Assembler::vpsraw(dst, nds, shift, vector_len);
3922 }
3923
3924 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3925 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3926 Assembler::vpsrlw(dst, nds, shift, vector_len);
3927 }
3928
3929 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3930 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3931 Assembler::vpsrlw(dst, nds, shift, vector_len);
3932 }
3933
3934 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
3935 assert(((dst->encoding() < 16 && shift->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3936 Assembler::vpsllw(dst, nds, shift, vector_len);
3937 }
3938
3939 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
3940 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3941 Assembler::vpsllw(dst, nds, shift, vector_len);
3942 }
3943
3944 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) {
3945 assert((dst->encoding() < 16 && src->encoding() < 16),"XMM register should be 0-15");
3946 Assembler::vptest(dst, src);
3947 }
3948
3949 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) {
3950 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3951 Assembler::punpcklbw(dst, src);
3952 }
3953
3954 void MacroAssembler::pshufd(XMMRegister dst, Address src, int mode) {
3955 assert(((dst->encoding() < 16) || VM_Version::supports_avx512vl()),"XMM register should be 0-15");
3956 Assembler::pshufd(dst, src, mode);
3957 }
3958
3959 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) {
3960 assert(((dst->encoding() < 16 && src->encoding() < 16) || VM_Version::supports_avx512vlbw()),"XMM register should be 0-15");
3961 Assembler::pshuflw(dst, src, mode);
3962 }
3963
3964 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
3965 if (reachable(src)) {
3966 vandpd(dst, nds, as_Address(src), vector_len);
3967 } else {
3968 lea(rscratch1, src);
3969 vandpd(dst, nds, Address(rscratch1, 0), vector_len);
3970 }
3971 }
3972
3973 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
3974 if (reachable(src)) {
3975 vandps(dst, nds, as_Address(src), vector_len);
3976 } else {
3977 lea(rscratch1, src);
3978 vandps(dst, nds, Address(rscratch1, 0), vector_len);
3979 }
3980 }
3981
3982 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3983 if (reachable(src)) {
3984 vdivsd(dst, nds, as_Address(src));
3985 } else {
3986 lea(rscratch1, src);
3987 vdivsd(dst, nds, Address(rscratch1, 0));
3988 }
3989 }
3990
3991 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
3992 if (reachable(src)) {
3993 vdivss(dst, nds, as_Address(src));
3994 } else {
3995 lea(rscratch1, src);
3996 vdivss(dst, nds, Address(rscratch1, 0));
3997 }
3998 }
3999
4000 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4001 if (reachable(src)) {
4002 vmulsd(dst, nds, as_Address(src));
4003 } else {
4004 lea(rscratch1, src);
4005 vmulsd(dst, nds, Address(rscratch1, 0));
4006 }
4007 }
4008
4009 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4010 if (reachable(src)) {
4011 vmulss(dst, nds, as_Address(src));
4012 } else {
4013 lea(rscratch1, src);
4014 vmulss(dst, nds, Address(rscratch1, 0));
4015 }
4016 }
4017
4018 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4019 if (reachable(src)) {
4020 vsubsd(dst, nds, as_Address(src));
4021 } else {
4022 lea(rscratch1, src);
4023 vsubsd(dst, nds, Address(rscratch1, 0));
4024 }
4025 }
4026
4027 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4028 if (reachable(src)) {
4029 vsubss(dst, nds, as_Address(src));
4030 } else {
4031 lea(rscratch1, src);
4032 vsubss(dst, nds, Address(rscratch1, 0));
4033 }
4034 }
4035
4036 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4037 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
4038 vxorps(dst, nds, src, Assembler::AVX_128bit);
4039 }
4040
4041 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4042 assert(((dst->encoding() < 16 && nds->encoding() < 16) || VM_Version::supports_avx512vldq()),"XMM register should be 0-15");
4043 vxorpd(dst, nds, src, Assembler::AVX_128bit);
4044 }
4045
4046 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
4047 if (reachable(src)) {
4048 vxorpd(dst, nds, as_Address(src), vector_len);
4049 } else {
4050 lea(rscratch1, src);
4051 vxorpd(dst, nds, Address(rscratch1, 0), vector_len);
4052 }
4053 }
4054
4055 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
4056 if (reachable(src)) {
4057 vxorps(dst, nds, as_Address(src), vector_len);
4058 } else {
4059 lea(rscratch1, src);
4060 vxorps(dst, nds, Address(rscratch1, 0), vector_len);
4061 }
4062 }
4063
4064 void MacroAssembler::clear_jweak_tag(Register possibly_jweak) {
4065 const int32_t inverted_jweak_mask = ~static_cast<int32_t>(JNIHandles::weak_tag_mask);
4066 STATIC_ASSERT(inverted_jweak_mask == -2); // otherwise check this code
4067 // The inverted mask is sign-extended
4068 andptr(possibly_jweak, inverted_jweak_mask);
4069 }
4070
4071 void MacroAssembler::resolve_jobject(Register value,
4072 Register thread,
4073 Register tmp) {
4074 assert_different_registers(value, thread, tmp);
4075 Label done, not_weak;
4076 testptr(value, value);
4077 jcc(Assembler::zero, done); // Use NULL as-is.
4078 testptr(value, JNIHandles::weak_tag_mask); // Test for jweak tag.
4079 jcc(Assembler::zero, not_weak);
4080 // Resolve jweak.
4081 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
4082 value, Address(value, -JNIHandles::weak_tag_value), tmp, thread);
4083 verify_oop(value);
4084 jmp(done);
4085 bind(not_weak);
4086 // Resolve (untagged) jobject.
4087 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, 0), tmp, thread);
4088 verify_oop(value);
4089 bind(done);
4090 }
4091
4092 void MacroAssembler::subptr(Register dst, int32_t imm32) {
4093 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32));
4094 }
4095
4096 // Force generation of a 4 byte immediate value even if it fits into 8bit
4097 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) {
4098 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32));
4099 }
4100
4101 void MacroAssembler::subptr(Register dst, Register src) {
4102 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src));
4103 }
4104
4105 // C++ bool manipulation
4106 void MacroAssembler::testbool(Register dst) {
4107 if(sizeof(bool) == 1)
4108 testb(dst, 0xff);
4109 else if(sizeof(bool) == 2) {
4110 // testw implementation needed for two byte bools
4111 ShouldNotReachHere();
4112 } else if(sizeof(bool) == 4)
4113 testl(dst, dst);
4114 else
4115 // unsupported
4116 ShouldNotReachHere();
4117 }
4118
4119 void MacroAssembler::testptr(Register dst, Register src) {
4120 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src));
4121 }
4122
4123 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
4124 void MacroAssembler::tlab_allocate(Register thread, Register obj,
4125 Register var_size_in_bytes,
4126 int con_size_in_bytes,
4127 Register t1,
4128 Register t2,
4129 Label& slow_case) {
4130 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4131 bs->tlab_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
4132 }
4133
4134 // Defines obj, preserves var_size_in_bytes
4135 void MacroAssembler::eden_allocate(Register thread, Register obj,
4136 Register var_size_in_bytes,
4137 int con_size_in_bytes,
4138 Register t1,
4139 Label& slow_case) {
4140 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4141 bs->eden_allocate(this, thread, obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case);
4142 }
4143
4144 // Preserves the contents of address, destroys the contents length_in_bytes and temp.
4145 void MacroAssembler::zero_memory(Register address, Register length_in_bytes, int offset_in_bytes, Register temp) {
4146 assert(address != length_in_bytes && address != temp && temp != length_in_bytes, "registers must be different");
4147 assert((offset_in_bytes & (BytesPerWord - 1)) == 0, "offset must be a multiple of BytesPerWord");
4148 Label done;
4149
4150 testptr(length_in_bytes, length_in_bytes);
4151 jcc(Assembler::zero, done);
4152
4153 // initialize topmost word, divide index by 2, check if odd and test if zero
4154 // note: for the remaining code to work, index must be a multiple of BytesPerWord
4155 #ifdef ASSERT
4156 {
4157 Label L;
4158 testptr(length_in_bytes, BytesPerWord - 1);
4159 jcc(Assembler::zero, L);
4160 stop("length must be a multiple of BytesPerWord");
4161 bind(L);
4162 }
4163 #endif
4164 Register index = length_in_bytes;
4165 xorptr(temp, temp); // use _zero reg to clear memory (shorter code)
4166 if (UseIncDec) {
4167 shrptr(index, 3); // divide by 8/16 and set carry flag if bit 2 was set
4168 } else {
4169 shrptr(index, 2); // use 2 instructions to avoid partial flag stall
4170 shrptr(index, 1);
4171 }
4172 #ifndef _LP64
4173 // index could have not been a multiple of 8 (i.e., bit 2 was set)
4174 {
4175 Label even;
4176 // note: if index was a multiple of 8, then it cannot
4177 // be 0 now otherwise it must have been 0 before
4178 // => if it is even, we don't need to check for 0 again
4179 jcc(Assembler::carryClear, even);
4180 // clear topmost word (no jump would be needed if conditional assignment worked here)
4181 movptr(Address(address, index, Address::times_8, offset_in_bytes - 0*BytesPerWord), temp);
4182 // index could be 0 now, must check again
4183 jcc(Assembler::zero, done);
4184 bind(even);
4185 }
4186 #endif // !_LP64
4187 // initialize remaining object fields: index is a multiple of 2 now
4188 {
4189 Label loop;
4190 bind(loop);
4191 movptr(Address(address, index, Address::times_8, offset_in_bytes - 1*BytesPerWord), temp);
4192 NOT_LP64(movptr(Address(address, index, Address::times_8, offset_in_bytes - 2*BytesPerWord), temp);)
4193 decrement(index);
4194 jcc(Assembler::notZero, loop);
4195 }
4196
4197 bind(done);
4198 }
4199
4200 // Look up the method for a megamorphic invokeinterface call.
4201 // The target method is determined by <intf_klass, itable_index>.
4202 // The receiver klass is in recv_klass.
4203 // On success, the result will be in method_result, and execution falls through.
4204 // On failure, execution transfers to the given label.
4205 void MacroAssembler::lookup_interface_method(Register recv_klass,
4206 Register intf_klass,
4207 RegisterOrConstant itable_index,
4208 Register method_result,
4209 Register scan_temp,
4210 Label& L_no_such_interface,
4211 bool return_method) {
4212 assert_different_registers(recv_klass, intf_klass, scan_temp);
4213 assert_different_registers(method_result, intf_klass, scan_temp);
4214 assert(recv_klass != method_result || !return_method,
4215 "recv_klass can be destroyed when method isn't needed");
4216
4217 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
4218 "caller must use same register for non-constant itable index as for method");
4219
4220 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
4221 int vtable_base = in_bytes(Klass::vtable_start_offset());
4222 int itentry_off = itableMethodEntry::method_offset_in_bytes();
4223 int scan_step = itableOffsetEntry::size() * wordSize;
4224 int vte_size = vtableEntry::size_in_bytes();
4225 Address::ScaleFactor times_vte_scale = Address::times_ptr;
4226 assert(vte_size == wordSize, "else adjust times_vte_scale");
4227
4228 movl(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
4229
4230 // %%% Could store the aligned, prescaled offset in the klassoop.
4231 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
4232
4233 if (return_method) {
4234 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
4235 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
4236 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
4237 }
4238
4239 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
4240 // if (scan->interface() == intf) {
4241 // result = (klass + scan->offset() + itable_index);
4242 // }
4243 // }
4244 Label search, found_method;
4245
4246 for (int peel = 1; peel >= 0; peel--) {
4247 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
4248 cmpptr(intf_klass, method_result);
4249
4250 if (peel) {
4251 jccb(Assembler::equal, found_method);
4252 } else {
4253 jccb(Assembler::notEqual, search);
4254 // (invert the test to fall through to found_method...)
4255 }
4256
4257 if (!peel) break;
4258
4259 bind(search);
4260
4261 // Check that the previous entry is non-null. A null entry means that
4262 // the receiver class doesn't implement the interface, and wasn't the
4263 // same as when the caller was compiled.
4264 testptr(method_result, method_result);
4265 jcc(Assembler::zero, L_no_such_interface);
4266 addptr(scan_temp, scan_step);
4267 }
4268
4269 bind(found_method);
4270
4271 if (return_method) {
4272 // Got a hit.
4273 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
4274 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
4275 }
4276 }
4277
4278
4279 // virtual method calling
4280 void MacroAssembler::lookup_virtual_method(Register recv_klass,
4281 RegisterOrConstant vtable_index,
4282 Register method_result) {
4283 const int base = in_bytes(Klass::vtable_start_offset());
4284 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below");
4285 Address vtable_entry_addr(recv_klass,
4286 vtable_index, Address::times_ptr,
4287 base + vtableEntry::method_offset_in_bytes());
4288 movptr(method_result, vtable_entry_addr);
4289 }
4290
4291
4292 void MacroAssembler::check_klass_subtype(Register sub_klass,
4293 Register super_klass,
4294 Register temp_reg,
4295 Label& L_success) {
4296 Label L_failure;
4297 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL);
4298 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
4299 bind(L_failure);
4300 }
4301
4302
4303 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
4304 Register super_klass,
4305 Register temp_reg,
4306 Label* L_success,
4307 Label* L_failure,
4308 Label* L_slow_path,
4309 RegisterOrConstant super_check_offset) {
4310 assert_different_registers(sub_klass, super_klass, temp_reg);
4311 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
4312 if (super_check_offset.is_register()) {
4313 assert_different_registers(sub_klass, super_klass,
4314 super_check_offset.as_register());
4315 } else if (must_load_sco) {
4316 assert(temp_reg != noreg, "supply either a temp or a register offset");
4317 }
4318
4319 Label L_fallthrough;
4320 int label_nulls = 0;
4321 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
4322 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
4323 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
4324 assert(label_nulls <= 1, "at most one NULL in the batch");
4325
4326 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
4327 int sco_offset = in_bytes(Klass::super_check_offset_offset());
4328 Address super_check_offset_addr(super_klass, sco_offset);
4329
4330 // Hacked jcc, which "knows" that L_fallthrough, at least, is in
4331 // range of a jccb. If this routine grows larger, reconsider at
4332 // least some of these.
4333 #define local_jcc(assembler_cond, label) \
4334 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \
4335 else jcc( assembler_cond, label) /*omit semi*/
4336
4337 // Hacked jmp, which may only be used just before L_fallthrough.
4338 #define final_jmp(label) \
4339 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
4340 else jmp(label) /*omit semi*/
4341
4342 // If the pointers are equal, we are done (e.g., String[] elements).
4343 // This self-check enables sharing of secondary supertype arrays among
4344 // non-primary types such as array-of-interface. Otherwise, each such
4345 // type would need its own customized SSA.
4346 // We move this check to the front of the fast path because many
4347 // type checks are in fact trivially successful in this manner,
4348 // so we get a nicely predicted branch right at the start of the check.
4349 cmpptr(sub_klass, super_klass);
4350 local_jcc(Assembler::equal, *L_success);
4351
4352 // Check the supertype display:
4353 if (must_load_sco) {
4354 // Positive movl does right thing on LP64.
4355 movl(temp_reg, super_check_offset_addr);
4356 super_check_offset = RegisterOrConstant(temp_reg);
4357 }
4358 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0);
4359 cmpptr(super_klass, super_check_addr); // load displayed supertype
4360
4361 // This check has worked decisively for primary supers.
4362 // Secondary supers are sought in the super_cache ('super_cache_addr').
4363 // (Secondary supers are interfaces and very deeply nested subtypes.)
4364 // This works in the same check above because of a tricky aliasing
4365 // between the super_cache and the primary super display elements.
4366 // (The 'super_check_addr' can address either, as the case requires.)
4367 // Note that the cache is updated below if it does not help us find
4368 // what we need immediately.
4369 // So if it was a primary super, we can just fail immediately.
4370 // Otherwise, it's the slow path for us (no success at this point).
4371
4372 if (super_check_offset.is_register()) {
4373 local_jcc(Assembler::equal, *L_success);
4374 cmpl(super_check_offset.as_register(), sc_offset);
4375 if (L_failure == &L_fallthrough) {
4376 local_jcc(Assembler::equal, *L_slow_path);
4377 } else {
4378 local_jcc(Assembler::notEqual, *L_failure);
4379 final_jmp(*L_slow_path);
4380 }
4381 } else if (super_check_offset.as_constant() == sc_offset) {
4382 // Need a slow path; fast failure is impossible.
4383 if (L_slow_path == &L_fallthrough) {
4384 local_jcc(Assembler::equal, *L_success);
4385 } else {
4386 local_jcc(Assembler::notEqual, *L_slow_path);
4387 final_jmp(*L_success);
4388 }
4389 } else {
4390 // No slow path; it's a fast decision.
4391 if (L_failure == &L_fallthrough) {
4392 local_jcc(Assembler::equal, *L_success);
4393 } else {
4394 local_jcc(Assembler::notEqual, *L_failure);
4395 final_jmp(*L_success);
4396 }
4397 }
4398
4399 bind(L_fallthrough);
4400
4401 #undef local_jcc
4402 #undef final_jmp
4403 }
4404
4405
4406 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
4407 Register super_klass,
4408 Register temp_reg,
4409 Register temp2_reg,
4410 Label* L_success,
4411 Label* L_failure,
4412 bool set_cond_codes) {
4413 assert_different_registers(sub_klass, super_klass, temp_reg);
4414 if (temp2_reg != noreg)
4415 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg);
4416 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
4417
4418 Label L_fallthrough;
4419 int label_nulls = 0;
4420 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
4421 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
4422 assert(label_nulls <= 1, "at most one NULL in the batch");
4423
4424 // a couple of useful fields in sub_klass:
4425 int ss_offset = in_bytes(Klass::secondary_supers_offset());
4426 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
4427 Address secondary_supers_addr(sub_klass, ss_offset);
4428 Address super_cache_addr( sub_klass, sc_offset);
4429
4430 // Do a linear scan of the secondary super-klass chain.
4431 // This code is rarely used, so simplicity is a virtue here.
4432 // The repne_scan instruction uses fixed registers, which we must spill.
4433 // Don't worry too much about pre-existing connections with the input regs.
4434
4435 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super)
4436 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter)
4437
4438 // Get super_klass value into rax (even if it was in rdi or rcx).
4439 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false;
4440 if (super_klass != rax || UseCompressedOops) {
4441 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; }
4442 mov(rax, super_klass);
4443 }
4444 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; }
4445 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; }
4446
4447 #ifndef PRODUCT
4448 int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
4449 ExternalAddress pst_counter_addr((address) pst_counter);
4450 NOT_LP64( incrementl(pst_counter_addr) );
4451 LP64_ONLY( lea(rcx, pst_counter_addr) );
4452 LP64_ONLY( incrementl(Address(rcx, 0)) );
4453 #endif //PRODUCT
4454
4455 // We will consult the secondary-super array.
4456 movptr(rdi, secondary_supers_addr);
4457 // Load the array length. (Positive movl does right thing on LP64.)
4458 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes()));
4459 // Skip to start of data.
4460 addptr(rdi, Array<Klass*>::base_offset_in_bytes());
4461
4462 // Scan RCX words at [RDI] for an occurrence of RAX.
4463 // Set NZ/Z based on last compare.
4464 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does
4465 // not change flags (only scas instruction which is repeated sets flags).
4466 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found.
4467
4468 testptr(rax,rax); // Set Z = 0
4469 repne_scan();
4470
4471 // Unspill the temp. registers:
4472 if (pushed_rdi) pop(rdi);
4473 if (pushed_rcx) pop(rcx);
4474 if (pushed_rax) pop(rax);
4475
4476 if (set_cond_codes) {
4477 // Special hack for the AD files: rdi is guaranteed non-zero.
4478 assert(!pushed_rdi, "rdi must be left non-NULL");
4479 // Also, the condition codes are properly set Z/NZ on succeed/failure.
4480 }
4481
4482 if (L_failure == &L_fallthrough)
4483 jccb(Assembler::notEqual, *L_failure);
4484 else jcc(Assembler::notEqual, *L_failure);
4485
4486 // Success. Cache the super we found and proceed in triumph.
4487 movptr(super_cache_addr, super_klass);
4488
4489 if (L_success != &L_fallthrough) {
4490 jmp(*L_success);
4491 }
4492
4493 #undef IS_A_TEMP
4494
4495 bind(L_fallthrough);
4496 }
4497
4498
4499 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
4500 if (VM_Version::supports_cmov()) {
4501 cmovl(cc, dst, src);
4502 } else {
4503 Label L;
4504 jccb(negate_condition(cc), L);
4505 movl(dst, src);
4506 bind(L);
4507 }
4508 }
4509
4510 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) {
4511 if (VM_Version::supports_cmov()) {
4512 cmovl(cc, dst, src);
4513 } else {
4514 Label L;
4515 jccb(negate_condition(cc), L);
4516 movl(dst, src);
4517 bind(L);
4518 }
4519 }
4520
4521 void MacroAssembler::verify_oop(Register reg, const char* s) {
4522 if (!VerifyOops || VerifyAdapterSharing) {
4523 // Below address of the code string confuses VerifyAdapterSharing
4524 // because it may differ between otherwise equivalent adapters.
4525 return;
4526 }
4527
4528 // Pass register number to verify_oop_subroutine
4529 const char* b = NULL;
4530 {
4531 ResourceMark rm;
4532 stringStream ss;
4533 ss.print("verify_oop: %s: %s", reg->name(), s);
4534 b = code_string(ss.as_string());
4535 }
4536 BLOCK_COMMENT("verify_oop {");
4537 #ifdef _LP64
4538 push(rscratch1); // save r10, trashed by movptr()
4539 #endif
4540 push(rax); // save rax,
4541 push(reg); // pass register argument
4542 ExternalAddress buffer((address) b);
4543 // avoid using pushptr, as it modifies scratch registers
4544 // and our contract is not to modify anything
4545 movptr(rax, buffer.addr());
4546 push(rax);
4547 // call indirectly to solve generation ordering problem
4548 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
4549 call(rax);
4550 // Caller pops the arguments (oop, message) and restores rax, r10
4551 BLOCK_COMMENT("} verify_oop");
4552 }
4553
4554
4555 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
4556 Register tmp,
4557 int offset) {
4558 intptr_t value = *delayed_value_addr;
4559 if (value != 0)
4560 return RegisterOrConstant(value + offset);
4561
4562 // load indirectly to solve generation ordering problem
4563 movptr(tmp, ExternalAddress((address) delayed_value_addr));
4564
4565 #ifdef ASSERT
4566 { Label L;
4567 testptr(tmp, tmp);
4568 if (WizardMode) {
4569 const char* buf = NULL;
4570 {
4571 ResourceMark rm;
4572 stringStream ss;
4573 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]);
4574 buf = code_string(ss.as_string());
4575 }
4576 jcc(Assembler::notZero, L);
4577 STOP(buf);
4578 } else {
4579 jccb(Assembler::notZero, L);
4580 hlt();
4581 }
4582 bind(L);
4583 }
4584 #endif
4585
4586 if (offset != 0)
4587 addptr(tmp, offset);
4588
4589 return RegisterOrConstant(tmp);
4590 }
4591
4592
4593 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
4594 int extra_slot_offset) {
4595 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
4596 int stackElementSize = Interpreter::stackElementSize;
4597 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
4598 #ifdef ASSERT
4599 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
4600 assert(offset1 - offset == stackElementSize, "correct arithmetic");
4601 #endif
4602 Register scale_reg = noreg;
4603 Address::ScaleFactor scale_factor = Address::no_scale;
4604 if (arg_slot.is_constant()) {
4605 offset += arg_slot.as_constant() * stackElementSize;
4606 } else {
4607 scale_reg = arg_slot.as_register();
4608 scale_factor = Address::times(stackElementSize);
4609 }
4610 offset += wordSize; // return PC is on stack
4611 return Address(rsp, scale_reg, scale_factor, offset);
4612 }
4613
4614
4615 void MacroAssembler::verify_oop_addr(Address addr, const char* s) {
4616 if (!VerifyOops || VerifyAdapterSharing) {
4617 // Below address of the code string confuses VerifyAdapterSharing
4618 // because it may differ between otherwise equivalent adapters.
4619 return;
4620 }
4621
4622 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord);
4623 // Pass register number to verify_oop_subroutine
4624 const char* b = NULL;
4625 {
4626 ResourceMark rm;
4627 stringStream ss;
4628 ss.print("verify_oop_addr: %s", s);
4629 b = code_string(ss.as_string());
4630 }
4631 #ifdef _LP64
4632 push(rscratch1); // save r10, trashed by movptr()
4633 #endif
4634 push(rax); // save rax,
4635 // addr may contain rsp so we will have to adjust it based on the push
4636 // we just did (and on 64 bit we do two pushes)
4637 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which
4638 // stores rax into addr which is backwards of what was intended.
4639 if (addr.uses(rsp)) {
4640 lea(rax, addr);
4641 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord));
4642 } else {
4643 pushptr(addr);
4644 }
4645
4646 ExternalAddress buffer((address) b);
4647 // pass msg argument
4648 // avoid using pushptr, as it modifies scratch registers
4649 // and our contract is not to modify anything
4650 movptr(rax, buffer.addr());
4651 push(rax);
4652
4653 // call indirectly to solve generation ordering problem
4654 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
4655 call(rax);
4656 // Caller pops the arguments (addr, message) and restores rax, r10.
4657 }
4658
4659 void MacroAssembler::verify_tlab() {
4660 #ifdef ASSERT
4661 if (UseTLAB && VerifyOops) {
4662 Label next, ok;
4663 Register t1 = rsi;
4664 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread);
4665
4666 push(t1);
4667 NOT_LP64(push(thread_reg));
4668 NOT_LP64(get_thread(thread_reg));
4669
4670 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
4671 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
4672 jcc(Assembler::aboveEqual, next);
4673 STOP("assert(top >= start)");
4674 should_not_reach_here();
4675
4676 bind(next);
4677 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
4678 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
4679 jcc(Assembler::aboveEqual, ok);
4680 STOP("assert(top <= end)");
4681 should_not_reach_here();
4682
4683 bind(ok);
4684 NOT_LP64(pop(thread_reg));
4685 pop(t1);
4686 }
4687 #endif
4688 }
4689
4690 class ControlWord {
4691 public:
4692 int32_t _value;
4693
4694 int rounding_control() const { return (_value >> 10) & 3 ; }
4695 int precision_control() const { return (_value >> 8) & 3 ; }
4696 bool precision() const { return ((_value >> 5) & 1) != 0; }
4697 bool underflow() const { return ((_value >> 4) & 1) != 0; }
4698 bool overflow() const { return ((_value >> 3) & 1) != 0; }
4699 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
4700 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
4701 bool invalid() const { return ((_value >> 0) & 1) != 0; }
4702
4703 void print() const {
4704 // rounding control
4705 const char* rc;
4706 switch (rounding_control()) {
4707 case 0: rc = "round near"; break;
4708 case 1: rc = "round down"; break;
4709 case 2: rc = "round up "; break;
4710 case 3: rc = "chop "; break;
4711 };
4712 // precision control
4713 const char* pc;
4714 switch (precision_control()) {
4715 case 0: pc = "24 bits "; break;
4716 case 1: pc = "reserved"; break;
4717 case 2: pc = "53 bits "; break;
4718 case 3: pc = "64 bits "; break;
4719 };
4720 // flags
4721 char f[9];
4722 f[0] = ' ';
4723 f[1] = ' ';
4724 f[2] = (precision ()) ? 'P' : 'p';
4725 f[3] = (underflow ()) ? 'U' : 'u';
4726 f[4] = (overflow ()) ? 'O' : 'o';
4727 f[5] = (zero_divide ()) ? 'Z' : 'z';
4728 f[6] = (denormalized()) ? 'D' : 'd';
4729 f[7] = (invalid ()) ? 'I' : 'i';
4730 f[8] = '\x0';
4731 // output
4732 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc);
4733 }
4734
4735 };
4736
4737 class StatusWord {
4738 public:
4739 int32_t _value;
4740
4741 bool busy() const { return ((_value >> 15) & 1) != 0; }
4742 bool C3() const { return ((_value >> 14) & 1) != 0; }
4743 bool C2() const { return ((_value >> 10) & 1) != 0; }
4744 bool C1() const { return ((_value >> 9) & 1) != 0; }
4745 bool C0() const { return ((_value >> 8) & 1) != 0; }
4746 int top() const { return (_value >> 11) & 7 ; }
4747 bool error_status() const { return ((_value >> 7) & 1) != 0; }
4748 bool stack_fault() const { return ((_value >> 6) & 1) != 0; }
4749 bool precision() const { return ((_value >> 5) & 1) != 0; }
4750 bool underflow() const { return ((_value >> 4) & 1) != 0; }
4751 bool overflow() const { return ((_value >> 3) & 1) != 0; }
4752 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
4753 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
4754 bool invalid() const { return ((_value >> 0) & 1) != 0; }
4755
4756 void print() const {
4757 // condition codes
4758 char c[5];
4759 c[0] = (C3()) ? '3' : '-';
4760 c[1] = (C2()) ? '2' : '-';
4761 c[2] = (C1()) ? '1' : '-';
4762 c[3] = (C0()) ? '0' : '-';
4763 c[4] = '\x0';
4764 // flags
4765 char f[9];
4766 f[0] = (error_status()) ? 'E' : '-';
4767 f[1] = (stack_fault ()) ? 'S' : '-';
4768 f[2] = (precision ()) ? 'P' : '-';
4769 f[3] = (underflow ()) ? 'U' : '-';
4770 f[4] = (overflow ()) ? 'O' : '-';
4771 f[5] = (zero_divide ()) ? 'Z' : '-';
4772 f[6] = (denormalized()) ? 'D' : '-';
4773 f[7] = (invalid ()) ? 'I' : '-';
4774 f[8] = '\x0';
4775 // output
4776 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top());
4777 }
4778
4779 };
4780
4781 class TagWord {
4782 public:
4783 int32_t _value;
4784
4785 int tag_at(int i) const { return (_value >> (i*2)) & 3; }
4786
4787 void print() const {
4788 printf("%04x", _value & 0xFFFF);
4789 }
4790
4791 };
4792
4793 class FPU_Register {
4794 public:
4795 int32_t _m0;
4796 int32_t _m1;
4797 int16_t _ex;
4798
4799 bool is_indefinite() const {
4800 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0;
4801 }
4802
4803 void print() const {
4804 char sign = (_ex < 0) ? '-' : '+';
4805 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " ";
4806 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind);
4807 };
4808
4809 };
4810
4811 class FPU_State {
4812 public:
4813 enum {
4814 register_size = 10,
4815 number_of_registers = 8,
4816 register_mask = 7
4817 };
4818
4819 ControlWord _control_word;
4820 StatusWord _status_word;
4821 TagWord _tag_word;
4822 int32_t _error_offset;
4823 int32_t _error_selector;
4824 int32_t _data_offset;
4825 int32_t _data_selector;
4826 int8_t _register[register_size * number_of_registers];
4827
4828 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); }
4829 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; }
4830
4831 const char* tag_as_string(int tag) const {
4832 switch (tag) {
4833 case 0: return "valid";
4834 case 1: return "zero";
4835 case 2: return "special";
4836 case 3: return "empty";
4837 }
4838 ShouldNotReachHere();
4839 return NULL;
4840 }
4841
4842 void print() const {
4843 // print computation registers
4844 { int t = _status_word.top();
4845 for (int i = 0; i < number_of_registers; i++) {
4846 int j = (i - t) & register_mask;
4847 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j);
4848 st(j)->print();
4849 printf(" %s\n", tag_as_string(_tag_word.tag_at(i)));
4850 }
4851 }
4852 printf("\n");
4853 // print control registers
4854 printf("ctrl = "); _control_word.print(); printf("\n");
4855 printf("stat = "); _status_word .print(); printf("\n");
4856 printf("tags = "); _tag_word .print(); printf("\n");
4857 }
4858
4859 };
4860
4861 class Flag_Register {
4862 public:
4863 int32_t _value;
4864
4865 bool overflow() const { return ((_value >> 11) & 1) != 0; }
4866 bool direction() const { return ((_value >> 10) & 1) != 0; }
4867 bool sign() const { return ((_value >> 7) & 1) != 0; }
4868 bool zero() const { return ((_value >> 6) & 1) != 0; }
4869 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; }
4870 bool parity() const { return ((_value >> 2) & 1) != 0; }
4871 bool carry() const { return ((_value >> 0) & 1) != 0; }
4872
4873 void print() const {
4874 // flags
4875 char f[8];
4876 f[0] = (overflow ()) ? 'O' : '-';
4877 f[1] = (direction ()) ? 'D' : '-';
4878 f[2] = (sign ()) ? 'S' : '-';
4879 f[3] = (zero ()) ? 'Z' : '-';
4880 f[4] = (auxiliary_carry()) ? 'A' : '-';
4881 f[5] = (parity ()) ? 'P' : '-';
4882 f[6] = (carry ()) ? 'C' : '-';
4883 f[7] = '\x0';
4884 // output
4885 printf("%08x flags = %s", _value, f);
4886 }
4887
4888 };
4889
4890 class IU_Register {
4891 public:
4892 int32_t _value;
4893
4894 void print() const {
4895 printf("%08x %11d", _value, _value);
4896 }
4897
4898 };
4899
4900 class IU_State {
4901 public:
4902 Flag_Register _eflags;
4903 IU_Register _rdi;
4904 IU_Register _rsi;
4905 IU_Register _rbp;
4906 IU_Register _rsp;
4907 IU_Register _rbx;
4908 IU_Register _rdx;
4909 IU_Register _rcx;
4910 IU_Register _rax;
4911
4912 void print() const {
4913 // computation registers
4914 printf("rax, = "); _rax.print(); printf("\n");
4915 printf("rbx, = "); _rbx.print(); printf("\n");
4916 printf("rcx = "); _rcx.print(); printf("\n");
4917 printf("rdx = "); _rdx.print(); printf("\n");
4918 printf("rdi = "); _rdi.print(); printf("\n");
4919 printf("rsi = "); _rsi.print(); printf("\n");
4920 printf("rbp, = "); _rbp.print(); printf("\n");
4921 printf("rsp = "); _rsp.print(); printf("\n");
4922 printf("\n");
4923 // control registers
4924 printf("flgs = "); _eflags.print(); printf("\n");
4925 }
4926 };
4927
4928
4929 class CPU_State {
4930 public:
4931 FPU_State _fpu_state;
4932 IU_State _iu_state;
4933
4934 void print() const {
4935 printf("--------------------------------------------------\n");
4936 _iu_state .print();
4937 printf("\n");
4938 _fpu_state.print();
4939 printf("--------------------------------------------------\n");
4940 }
4941
4942 };
4943
4944
4945 static void _print_CPU_state(CPU_State* state) {
4946 state->print();
4947 };
4948
4949
4950 void MacroAssembler::print_CPU_state() {
4951 push_CPU_state();
4952 push(rsp); // pass CPU state
4953 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state)));
4954 addptr(rsp, wordSize); // discard argument
4955 pop_CPU_state();
4956 }
4957
4958
4959 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) {
4960 static int counter = 0;
4961 FPU_State* fs = &state->_fpu_state;
4962 counter++;
4963 // For leaf calls, only verify that the top few elements remain empty.
4964 // We only need 1 empty at the top for C2 code.
4965 if( stack_depth < 0 ) {
4966 if( fs->tag_for_st(7) != 3 ) {
4967 printf("FPR7 not empty\n");
4968 state->print();
4969 assert(false, "error");
4970 return false;
4971 }
4972 return true; // All other stack states do not matter
4973 }
4974
4975 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std,
4976 "bad FPU control word");
4977
4978 // compute stack depth
4979 int i = 0;
4980 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++;
4981 int d = i;
4982 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++;
4983 // verify findings
4984 if (i != FPU_State::number_of_registers) {
4985 // stack not contiguous
4986 printf("%s: stack not contiguous at ST%d\n", s, i);
4987 state->print();
4988 assert(false, "error");
4989 return false;
4990 }
4991 // check if computed stack depth corresponds to expected stack depth
4992 if (stack_depth < 0) {
4993 // expected stack depth is -stack_depth or less
4994 if (d > -stack_depth) {
4995 // too many elements on the stack
4996 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d);
4997 state->print();
4998 assert(false, "error");
4999 return false;
5000 }
5001 } else {
5002 // expected stack depth is stack_depth
5003 if (d != stack_depth) {
5004 // wrong stack depth
5005 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d);
5006 state->print();
5007 assert(false, "error");
5008 return false;
5009 }
5010 }
5011 // everything is cool
5012 return true;
5013 }
5014
5015
5016 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
5017 if (!VerifyFPU) return;
5018 push_CPU_state();
5019 push(rsp); // pass CPU state
5020 ExternalAddress msg((address) s);
5021 // pass message string s
5022 pushptr(msg.addr());
5023 push(stack_depth); // pass stack depth
5024 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU)));
5025 addptr(rsp, 3 * wordSize); // discard arguments
5026 // check for error
5027 { Label L;
5028 testl(rax, rax);
5029 jcc(Assembler::notZero, L);
5030 int3(); // break if error condition
5031 bind(L);
5032 }
5033 pop_CPU_state();
5034 }
5035
5036 void MacroAssembler::restore_cpu_control_state_after_jni() {
5037 // Either restore the MXCSR register after returning from the JNI Call
5038 // or verify that it wasn't changed (with -Xcheck:jni flag).
5039 if (VM_Version::supports_sse()) {
5040 if (RestoreMXCSROnJNICalls) {
5041 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
5042 } else if (CheckJNICalls) {
5043 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
5044 }
5045 }
5046 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty.
5047 vzeroupper();
5048 // Reset k1 to 0xffff.
5049
5050 #ifdef COMPILER2
5051 if (PostLoopMultiversioning && VM_Version::supports_evex()) {
5052 push(rcx);
5053 movl(rcx, 0xffff);
5054 kmovwl(k1, rcx);
5055 pop(rcx);
5056 }
5057 #endif // COMPILER2
5058
5059 #ifndef _LP64
5060 // Either restore the x87 floating pointer control word after returning
5061 // from the JNI call or verify that it wasn't changed.
5062 if (CheckJNICalls) {
5063 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
5064 }
5065 #endif // _LP64
5066 }
5067
5068 // ((OopHandle)result).resolve();
5069 void MacroAssembler::resolve_oop_handle(Register result, Register tmp) {
5070 assert_different_registers(result, tmp);
5071
5072 // Only 64 bit platforms support GCs that require a tmp register
5073 // Only IN_HEAP loads require a thread_tmp register
5074 // OopHandle::resolve is an indirection like jobject.
5075 access_load_at(T_OBJECT, IN_NATIVE,
5076 result, Address(result, 0), tmp, /*tmp_thread*/noreg);
5077 }
5078
5079 void MacroAssembler::load_mirror(Register mirror, Register method, Register tmp) {
5080 // get mirror
5081 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
5082 movptr(mirror, Address(method, Method::const_offset()));
5083 movptr(mirror, Address(mirror, ConstMethod::constants_offset()));
5084 movptr(mirror, Address(mirror, ConstantPool::pool_holder_offset_in_bytes()));
5085 movptr(mirror, Address(mirror, mirror_offset));
5086 resolve_oop_handle(mirror, tmp);
5087 }
5088
5089 void MacroAssembler::load_metadata(Register dst, Register src) {
5090 if (UseCompressedClassPointers) {
5091 movl(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5092 } else {
5093 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5094 }
5095 }
5096
5097 void MacroAssembler::load_storage_props(Register dst, Register src) {
5098 load_metadata(dst, src);
5099 if (UseCompressedClassPointers) {
5100 shrl(dst, oopDesc::narrow_storage_props_shift);
5101 } else {
5102 shrq(dst, oopDesc::wide_storage_props_shift);
5103 }
5104 }
5105
5106 void MacroAssembler::load_klass(Register dst, Register src) {
5107 load_metadata(dst, src);
5108 #ifdef _LP64
5109 if (UseCompressedClassPointers) {
5110 andl(dst, oopDesc::compressed_klass_mask());
5111 decode_klass_not_null(dst);
5112 } else
5113 #endif
5114 {
5115 #ifdef _LP64
5116 shlq(dst, oopDesc::storage_props_nof_bits);
5117 shrq(dst, oopDesc::storage_props_nof_bits);
5118 #else
5119 andl(dst, oopDesc::wide_klass_mask());
5120 #endif
5121 }
5122 }
5123
5124 void MacroAssembler::load_prototype_header(Register dst, Register src) {
5125 load_klass(dst, src);
5126 movptr(dst, Address(dst, Klass::prototype_header_offset()));
5127 }
5128
5129 void MacroAssembler::store_klass(Register dst, Register src) {
5130 #ifdef _LP64
5131 if (UseCompressedClassPointers) {
5132 encode_klass_not_null(src);
5133 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5134 } else
5135 #endif
5136 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src);
5137 }
5138
5139 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators, Register dst, Address src,
5140 Register tmp1, Register thread_tmp) {
5141 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5142 decorators = AccessInternal::decorator_fixup(decorators);
5143 bool as_raw = (decorators & AS_RAW) != 0;
5144 if (as_raw) {
5145 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
5146 } else {
5147 bs->load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
5148 }
5149 }
5150
5151 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators, Address dst, Register src,
5152 Register tmp1, Register tmp2) {
5153 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5154 decorators = AccessInternal::decorator_fixup(decorators);
5155 bool as_raw = (decorators & AS_RAW) != 0;
5156 if (as_raw) {
5157 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, src, tmp1, tmp2);
5158 } else {
5159 bs->store_at(this, decorators, type, dst, src, tmp1, tmp2);
5160 }
5161 }
5162
5163 void MacroAssembler::resolve(DecoratorSet decorators, Register obj) {
5164 // Use stronger ACCESS_WRITE|ACCESS_READ by default.
5165 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
5166 decorators |= ACCESS_READ | ACCESS_WRITE;
5167 }
5168 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5169 return bs->resolve(this, decorators, obj);
5170 }
5171
5172 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
5173 Register thread_tmp, DecoratorSet decorators) {
5174 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp);
5175 }
5176
5177 // Doesn't do verfication, generates fixed size code
5178 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1,
5179 Register thread_tmp, DecoratorSet decorators) {
5180 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, thread_tmp);
5181 }
5182
5183 void MacroAssembler::store_heap_oop(Address dst, Register src, Register tmp1,
5184 Register tmp2, DecoratorSet decorators) {
5185 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2);
5186 }
5187
5188 // Used for storing NULLs.
5189 void MacroAssembler::store_heap_oop_null(Address dst) {
5190 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg);
5191 }
5192
5193 #ifdef _LP64
5194 void MacroAssembler::store_klass_gap(Register dst, Register src) {
5195 if (UseCompressedClassPointers) {
5196 // Store to klass gap in destination
5197 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src);
5198 }
5199 }
5200
5201 #ifdef ASSERT
5202 void MacroAssembler::verify_heapbase(const char* msg) {
5203 assert (UseCompressedOops, "should be compressed");
5204 assert (Universe::heap() != NULL, "java heap should be initialized");
5205 if (CheckCompressedOops) {
5206 Label ok;
5207 push(rscratch1); // cmpptr trashes rscratch1
5208 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
5209 jcc(Assembler::equal, ok);
5210 STOP(msg);
5211 bind(ok);
5212 pop(rscratch1);
5213 }
5214 }
5215 #endif
5216
5217 // Algorithm must match oop.inline.hpp encode_heap_oop.
5218 void MacroAssembler::encode_heap_oop(Register r) {
5219 #ifdef ASSERT
5220 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
5221 #endif
5222 verify_oop(r, "broken oop in encode_heap_oop");
5223 if (Universe::narrow_oop_base() == NULL) {
5224 if (Universe::narrow_oop_shift() != 0) {
5225 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5226 shrq(r, LogMinObjAlignmentInBytes);
5227 }
5228 return;
5229 }
5230 testq(r, r);
5231 cmovq(Assembler::equal, r, r12_heapbase);
5232 subq(r, r12_heapbase);
5233 shrq(r, LogMinObjAlignmentInBytes);
5234 }
5235
5236 void MacroAssembler::encode_heap_oop_not_null(Register r) {
5237 #ifdef ASSERT
5238 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
5239 if (CheckCompressedOops) {
5240 Label ok;
5241 testq(r, r);
5242 jcc(Assembler::notEqual, ok);
5243 STOP("null oop passed to encode_heap_oop_not_null");
5244 bind(ok);
5245 }
5246 #endif
5247 verify_oop(r, "broken oop in encode_heap_oop_not_null");
5248 if (Universe::narrow_oop_base() != NULL) {
5249 subq(r, r12_heapbase);
5250 }
5251 if (Universe::narrow_oop_shift() != 0) {
5252 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5253 shrq(r, LogMinObjAlignmentInBytes);
5254 }
5255 }
5256
5257 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
5258 #ifdef ASSERT
5259 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
5260 if (CheckCompressedOops) {
5261 Label ok;
5262 testq(src, src);
5263 jcc(Assembler::notEqual, ok);
5264 STOP("null oop passed to encode_heap_oop_not_null2");
5265 bind(ok);
5266 }
5267 #endif
5268 verify_oop(src, "broken oop in encode_heap_oop_not_null2");
5269 if (dst != src) {
5270 movq(dst, src);
5271 }
5272 if (Universe::narrow_oop_base() != NULL) {
5273 subq(dst, r12_heapbase);
5274 }
5275 if (Universe::narrow_oop_shift() != 0) {
5276 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5277 shrq(dst, LogMinObjAlignmentInBytes);
5278 }
5279 }
5280
5281 void MacroAssembler::decode_heap_oop(Register r) {
5282 #ifdef ASSERT
5283 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
5284 #endif
5285 if (Universe::narrow_oop_base() == NULL) {
5286 if (Universe::narrow_oop_shift() != 0) {
5287 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5288 shlq(r, LogMinObjAlignmentInBytes);
5289 }
5290 } else {
5291 Label done;
5292 shlq(r, LogMinObjAlignmentInBytes);
5293 jccb(Assembler::equal, done);
5294 addq(r, r12_heapbase);
5295 bind(done);
5296 }
5297 verify_oop(r, "broken oop in decode_heap_oop");
5298 }
5299
5300 void MacroAssembler::decode_heap_oop_not_null(Register r) {
5301 // Note: it will change flags
5302 assert (UseCompressedOops, "should only be used for compressed headers");
5303 assert (Universe::heap() != NULL, "java heap should be initialized");
5304 // Cannot assert, unverified entry point counts instructions (see .ad file)
5305 // vtableStubs also counts instructions in pd_code_size_limit.
5306 // Also do not verify_oop as this is called by verify_oop.
5307 if (Universe::narrow_oop_shift() != 0) {
5308 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5309 shlq(r, LogMinObjAlignmentInBytes);
5310 if (Universe::narrow_oop_base() != NULL) {
5311 addq(r, r12_heapbase);
5312 }
5313 } else {
5314 assert (Universe::narrow_oop_base() == NULL, "sanity");
5315 }
5316 }
5317
5318 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
5319 // Note: it will change flags
5320 assert (UseCompressedOops, "should only be used for compressed headers");
5321 assert (Universe::heap() != NULL, "java heap should be initialized");
5322 // Cannot assert, unverified entry point counts instructions (see .ad file)
5323 // vtableStubs also counts instructions in pd_code_size_limit.
5324 // Also do not verify_oop as this is called by verify_oop.
5325 if (Universe::narrow_oop_shift() != 0) {
5326 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
5327 if (LogMinObjAlignmentInBytes == Address::times_8) {
5328 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0));
5329 } else {
5330 if (dst != src) {
5331 movq(dst, src);
5332 }
5333 shlq(dst, LogMinObjAlignmentInBytes);
5334 if (Universe::narrow_oop_base() != NULL) {
5335 addq(dst, r12_heapbase);
5336 }
5337 }
5338 } else {
5339 assert (Universe::narrow_oop_base() == NULL, "sanity");
5340 if (dst != src) {
5341 movq(dst, src);
5342 }
5343 }
5344 }
5345
5346 void MacroAssembler::encode_klass_not_null(Register r) {
5347 if (Universe::narrow_klass_base() != NULL) {
5348 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5349 assert(r != r12_heapbase, "Encoding a klass in r12");
5350 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
5351 subq(r, r12_heapbase);
5352 }
5353 if (Universe::narrow_klass_shift() != 0) {
5354 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5355 shrq(r, LogKlassAlignmentInBytes);
5356 }
5357 if (Universe::narrow_klass_base() != NULL) {
5358 reinit_heapbase();
5359 }
5360 }
5361
5362 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
5363 if (dst == src) {
5364 encode_klass_not_null(src);
5365 } else {
5366 if (Universe::narrow_klass_base() != NULL) {
5367 mov64(dst, (int64_t)Universe::narrow_klass_base());
5368 negq(dst);
5369 addq(dst, src);
5370 } else {
5371 movptr(dst, src);
5372 }
5373 if (Universe::narrow_klass_shift() != 0) {
5374 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5375 shrq(dst, LogKlassAlignmentInBytes);
5376 }
5377 }
5378 }
5379
5380 // Function instr_size_for_decode_klass_not_null() counts the instructions
5381 // generated by decode_klass_not_null(register r) and reinit_heapbase(),
5382 // when (Universe::heap() != NULL). Hence, if the instructions they
5383 // generate change, then this method needs to be updated.
5384 int MacroAssembler::instr_size_for_decode_klass_not_null() {
5385 assert (UseCompressedClassPointers, "only for compressed klass ptrs");
5386 if (Universe::narrow_klass_base() != NULL) {
5387 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()).
5388 return (Universe::narrow_klass_shift() == 0 ? 20 : 24);
5389 } else {
5390 // longest load decode klass function, mov64, leaq
5391 return 16;
5392 }
5393 }
5394
5395 // !!! If the instructions that get generated here change then function
5396 // instr_size_for_decode_klass_not_null() needs to get updated.
5397 void MacroAssembler::decode_klass_not_null(Register r) {
5398 // Note: it will change flags
5399 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5400 assert(r != r12_heapbase, "Decoding a klass in r12");
5401 // Cannot assert, unverified entry point counts instructions (see .ad file)
5402 // vtableStubs also counts instructions in pd_code_size_limit.
5403 // Also do not verify_oop as this is called by verify_oop.
5404 if (Universe::narrow_klass_shift() != 0) {
5405 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5406 shlq(r, LogKlassAlignmentInBytes);
5407 }
5408 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
5409 if (Universe::narrow_klass_base() != NULL) {
5410 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
5411 addq(r, r12_heapbase);
5412 reinit_heapbase();
5413 }
5414 }
5415
5416 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
5417 // Note: it will change flags
5418 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5419 if (dst == src) {
5420 decode_klass_not_null(dst);
5421 } else {
5422 // Cannot assert, unverified entry point counts instructions (see .ad file)
5423 // vtableStubs also counts instructions in pd_code_size_limit.
5424 // Also do not verify_oop as this is called by verify_oop.
5425 mov64(dst, (int64_t)Universe::narrow_klass_base());
5426 if (Universe::narrow_klass_shift() != 0) {
5427 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
5428 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?");
5429 leaq(dst, Address(dst, src, Address::times_8, 0));
5430 } else {
5431 addq(dst, src);
5432 }
5433 }
5434 }
5435
5436 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
5437 assert (UseCompressedOops, "should only be used for compressed headers");
5438 assert (Universe::heap() != NULL, "java heap should be initialized");
5439 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5440 int oop_index = oop_recorder()->find_index(obj);
5441 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5442 mov_narrow_oop(dst, oop_index, rspec);
5443 }
5444
5445 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) {
5446 assert (UseCompressedOops, "should only be used for compressed headers");
5447 assert (Universe::heap() != NULL, "java heap should be initialized");
5448 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5449 int oop_index = oop_recorder()->find_index(obj);
5450 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5451 mov_narrow_oop(dst, oop_index, rspec);
5452 }
5453
5454 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
5455 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5456 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5457 int klass_index = oop_recorder()->find_index(k);
5458 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5459 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
5460 }
5461
5462 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) {
5463 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5464 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5465 int klass_index = oop_recorder()->find_index(k);
5466 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5467 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
5468 }
5469
5470 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) {
5471 assert (UseCompressedOops, "should only be used for compressed headers");
5472 assert (Universe::heap() != NULL, "java heap should be initialized");
5473 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5474 int oop_index = oop_recorder()->find_index(obj);
5475 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5476 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
5477 }
5478
5479 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) {
5480 assert (UseCompressedOops, "should only be used for compressed headers");
5481 assert (Universe::heap() != NULL, "java heap should be initialized");
5482 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5483 int oop_index = oop_recorder()->find_index(obj);
5484 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5485 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
5486 }
5487
5488 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) {
5489 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5490 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5491 int klass_index = oop_recorder()->find_index(k);
5492 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5493 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
5494 }
5495
5496 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) {
5497 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5498 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
5499 int klass_index = oop_recorder()->find_index(k);
5500 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
5501 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
5502 }
5503
5504 void MacroAssembler::reinit_heapbase() {
5505 if (UseCompressedOops || UseCompressedClassPointers) {
5506 if (Universe::heap() != NULL) {
5507 if (Universe::narrow_oop_base() == NULL) {
5508 MacroAssembler::xorptr(r12_heapbase, r12_heapbase);
5509 } else {
5510 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base());
5511 }
5512 } else {
5513 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
5514 }
5515 }
5516 }
5517
5518 #endif // _LP64
5519
5520 // C2 compiled method's prolog code.
5521 void MacroAssembler::verified_entry(Compile* C, int sp_inc) {
5522 int framesize = C->frame_size_in_bytes();
5523 int bangsize = C->bang_size_in_bytes();
5524 bool fp_mode_24b = C->in_24_bit_fp_mode();
5525 int stack_bang_size = C->need_stack_bang(bangsize) ? bangsize : 0;
5526 bool is_stub = C->stub_function() != NULL;
5527
5528 // WARNING: Initial instruction MUST be 5 bytes or longer so that
5529 // NativeJump::patch_verified_entry will be able to patch out the entry
5530 // code safely. The push to verify stack depth is ok at 5 bytes,
5531 // the frame allocation can be either 3 or 6 bytes. So if we don't do
5532 // stack bang then we must use the 6 byte frame allocation even if
5533 // we have no frame. :-(
5534 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect");
5535
5536 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
5537 // Remove word for return addr
5538 framesize -= wordSize;
5539 stack_bang_size -= wordSize;
5540
5541 // Calls to C2R adapters often do not accept exceptional returns.
5542 // We require that their callers must bang for them. But be careful, because
5543 // some VM calls (such as call site linkage) can use several kilobytes of
5544 // stack. But the stack safety zone should account for that.
5545 // See bugs 4446381, 4468289, 4497237.
5546 if (stack_bang_size > 0) {
5547 generate_stack_overflow_check(stack_bang_size);
5548
5549 // We always push rbp, so that on return to interpreter rbp, will be
5550 // restored correctly and we can correct the stack.
5551 push(rbp);
5552 // Save caller's stack pointer into RBP if the frame pointer is preserved.
5553 if (PreserveFramePointer) {
5554 mov(rbp, rsp);
5555 }
5556 // Remove word for ebp
5557 framesize -= wordSize;
5558
5559 // Create frame
5560 if (framesize) {
5561 subptr(rsp, framesize);
5562 }
5563 } else {
5564 // Create frame (force generation of a 4 byte immediate value)
5565 subptr_imm32(rsp, framesize);
5566
5567 // Save RBP register now.
5568 framesize -= wordSize;
5569 movptr(Address(rsp, framesize), rbp);
5570 // Save caller's stack pointer into RBP if the frame pointer is preserved.
5571 if (PreserveFramePointer) {
5572 movptr(rbp, rsp);
5573 if (framesize > 0) {
5574 addptr(rbp, framesize);
5575 }
5576 }
5577 }
5578
5579 if (C->needs_stack_repair()) {
5580 // Save stack increment (also account for fixed framesize and rbp)
5581 assert((sp_inc & (StackAlignmentInBytes-1)) == 0, "stack increment not aligned");
5582 movptr(Address(rsp, C->sp_inc_offset()), sp_inc + framesize + wordSize);
5583 }
5584
5585 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
5586 framesize -= wordSize;
5587 movptr(Address(rsp, framesize), (int32_t)0xbadb100d);
5588 }
5589
5590 #ifndef _LP64
5591 // If method sets FPU control word do it now
5592 if (fp_mode_24b) {
5593 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
5594 }
5595 if (UseSSE >= 2 && VerifyFPU) {
5596 verify_FPU(0, "FPU stack must be clean on entry");
5597 }
5598 #endif
5599
5600 #ifdef ASSERT
5601 if (VerifyStackAtCalls) {
5602 Label L;
5603 push(rax);
5604 mov(rax, rsp);
5605 andptr(rax, StackAlignmentInBytes-1);
5606 cmpptr(rax, StackAlignmentInBytes-wordSize);
5607 pop(rax);
5608 jcc(Assembler::equal, L);
5609 STOP("Stack is not properly aligned!");
5610 bind(L);
5611 }
5612 #endif
5613
5614 if (!is_stub) {
5615 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
5616 bs->nmethod_entry_barrier(this);
5617 }
5618 }
5619
5620 // clear memory of size 'cnt' qwords, starting at 'base' using XMM/YMM registers
5621 void MacroAssembler::xmm_clear_mem(Register base, Register cnt, Register val, XMMRegister xtmp) {
5622 // cnt - number of qwords (8-byte words).
5623 // base - start address, qword aligned.
5624 Label L_zero_64_bytes, L_loop, L_sloop, L_tail, L_end;
5625 movdq(xtmp, val);
5626 if (UseAVX >= 2) {
5627 punpcklqdq(xtmp, xtmp);
5628 vinserti128_high(xtmp, xtmp);
5629 } else {
5630 punpcklqdq(xtmp, xtmp);
5631 }
5632 jmp(L_zero_64_bytes);
5633
5634 BIND(L_loop);
5635 if (UseAVX >= 2) {
5636 vmovdqu(Address(base, 0), xtmp);
5637 vmovdqu(Address(base, 32), xtmp);
5638 } else {
5639 movdqu(Address(base, 0), xtmp);
5640 movdqu(Address(base, 16), xtmp);
5641 movdqu(Address(base, 32), xtmp);
5642 movdqu(Address(base, 48), xtmp);
5643 }
5644 addptr(base, 64);
5645
5646 BIND(L_zero_64_bytes);
5647 subptr(cnt, 8);
5648 jccb(Assembler::greaterEqual, L_loop);
5649 addptr(cnt, 4);
5650 jccb(Assembler::less, L_tail);
5651 // Copy trailing 32 bytes
5652 if (UseAVX >= 2) {
5653 vmovdqu(Address(base, 0), xtmp);
5654 } else {
5655 movdqu(Address(base, 0), xtmp);
5656 movdqu(Address(base, 16), xtmp);
5657 }
5658 addptr(base, 32);
5659 subptr(cnt, 4);
5660
5661 BIND(L_tail);
5662 addptr(cnt, 4);
5663 jccb(Assembler::lessEqual, L_end);
5664 decrement(cnt);
5665
5666 BIND(L_sloop);
5667 movq(Address(base, 0), xtmp);
5668 addptr(base, 8);
5669 decrement(cnt);
5670 jccb(Assembler::greaterEqual, L_sloop);
5671 BIND(L_end);
5672 }
5673
5674 // Move a value between registers/stack slots and update the reg_state
5675 bool MacroAssembler::move_helper(VMReg from, VMReg to, BasicType bt, RegState reg_state[], int ret_off) {
5676 if (reg_state[to->value()] == reg_written) {
5677 return true; // Already written
5678 }
5679 if (from != to && bt != T_VOID) {
5680 if (reg_state[to->value()] == reg_readonly) {
5681 return false; // Not yet writable
5682 }
5683 if (from->is_reg()) {
5684 if (to->is_reg()) {
5685 if (from->is_XMMRegister()) {
5686 if (bt == T_DOUBLE) {
5687 movdbl(to->as_XMMRegister(), from->as_XMMRegister());
5688 } else {
5689 assert(bt == T_FLOAT, "must be float");
5690 movflt(to->as_XMMRegister(), from->as_XMMRegister());
5691 }
5692 } else {
5693 movq(to->as_Register(), from->as_Register());
5694 }
5695 } else {
5696 int st_off = to->reg2stack() * VMRegImpl::stack_slot_size + wordSize;
5697 assert(st_off != ret_off, "overwriting return address at %d", st_off);
5698 Address to_addr = Address(rsp, st_off);
5699 if (from->is_XMMRegister()) {
5700 if (bt == T_DOUBLE) {
5701 movdbl(to_addr, from->as_XMMRegister());
5702 } else {
5703 assert(bt == T_FLOAT, "must be float");
5704 movflt(to_addr, from->as_XMMRegister());
5705 }
5706 } else {
5707 movq(to_addr, from->as_Register());
5708 }
5709 }
5710 } else {
5711 Address from_addr = Address(rsp, from->reg2stack() * VMRegImpl::stack_slot_size + wordSize);
5712 if (to->is_reg()) {
5713 if (to->is_XMMRegister()) {
5714 if (bt == T_DOUBLE) {
5715 movdbl(to->as_XMMRegister(), from_addr);
5716 } else {
5717 assert(bt == T_FLOAT, "must be float");
5718 movflt(to->as_XMMRegister(), from_addr);
5719 }
5720 } else {
5721 movq(to->as_Register(), from_addr);
5722 }
5723 } else {
5724 int st_off = to->reg2stack() * VMRegImpl::stack_slot_size + wordSize;
5725 assert(st_off != ret_off, "overwriting return address at %d", st_off);
5726 movq(r13, from_addr);
5727 movq(Address(rsp, st_off), r13);
5728 }
5729 }
5730 }
5731 // Update register states
5732 reg_state[from->value()] = reg_writable;
5733 reg_state[to->value()] = reg_written;
5734 return true;
5735 }
5736
5737 // Read all fields from a value type oop and store the values in registers/stack slots
5738 bool MacroAssembler::unpack_value_helper(const GrowableArray<SigEntry>* sig, int& sig_index, VMReg from, VMRegPair* regs_to, int& to_index, RegState reg_state[], int ret_off) {
5739 Register fromReg = from->is_reg() ? from->as_Register() : noreg;
5740 assert(sig->at(sig_index)._bt == T_VOID, "should be at end delimiter");
5741
5742 int vt = 1;
5743 bool done = true;
5744 bool mark_done = true;
5745 do {
5746 sig_index--;
5747 BasicType bt = sig->at(sig_index)._bt;
5748 if (bt == T_VALUETYPE) {
5749 vt--;
5750 } else if (bt == T_VOID &&
5751 sig->at(sig_index-1)._bt != T_LONG &&
5752 sig->at(sig_index-1)._bt != T_DOUBLE) {
5753 vt++;
5754 } else if (SigEntry::is_reserved_entry(sig, sig_index)) {
5755 to_index--; // Ignore this
5756 } else {
5757 assert(to_index >= 0, "invalid to_index");
5758 VMRegPair pair_to = regs_to[to_index--];
5759 VMReg to = pair_to.first();
5760
5761 if (bt == T_VOID) continue;
5762
5763 int idx = (int)to->value();
5764 if (reg_state[idx] == reg_readonly) {
5765 if (idx != from->value()) {
5766 mark_done = false;
5767 }
5768 done = false;
5769 continue;
5770 } else if (reg_state[idx] == reg_written) {
5771 continue;
5772 } else {
5773 assert(reg_state[idx] == reg_writable, "must be writable");
5774 reg_state[idx] = reg_written;
5775 }
5776
5777 if (fromReg == noreg) {
5778 int st_off = from->reg2stack() * VMRegImpl::stack_slot_size + wordSize;
5779 movq(r10, Address(rsp, st_off));
5780 fromReg = r10;
5781 }
5782
5783 int off = sig->at(sig_index)._offset;
5784 assert(off > 0, "offset in object should be positive");
5785 bool is_oop = (bt == T_OBJECT || bt == T_ARRAY);
5786
5787 Address fromAddr = Address(fromReg, off);
5788 bool is_signed = (bt != T_CHAR) && (bt != T_BOOLEAN);
5789 if (!to->is_XMMRegister()) {
5790 Register dst = to->is_stack() ? r13 : to->as_Register();
5791 if (is_oop) {
5792 load_heap_oop(dst, fromAddr);
5793 } else {
5794 load_sized_value(dst, fromAddr, type2aelembytes(bt), is_signed);
5795 }
5796 if (to->is_stack()) {
5797 int st_off = to->reg2stack() * VMRegImpl::stack_slot_size + wordSize;
5798 assert(st_off != ret_off, "overwriting return address at %d", st_off);
5799 movq(Address(rsp, st_off), dst);
5800 }
5801 } else {
5802 if (bt == T_DOUBLE) {
5803 movdbl(to->as_XMMRegister(), fromAddr);
5804 } else {
5805 assert(bt == T_FLOAT, "must be float");
5806 movflt(to->as_XMMRegister(), fromAddr);
5807 }
5808 }
5809 }
5810 } while (vt != 0);
5811 if (mark_done && reg_state[from->value()] != reg_written) {
5812 // This is okay because no one else will write to that slot
5813 reg_state[from->value()] = reg_writable;
5814 }
5815 return done;
5816 }
5817
5818 // Unpack all value type arguments passed as oops
5819 void MacroAssembler::unpack_value_args(Compile* C, bool receiver_only) {
5820 assert(C->has_scalarized_args(), "value type argument scalarization is disabled");
5821 Method* method = C->method()->get_Method();
5822 const GrowableArray<SigEntry>* sig_cc = method->adapter()->get_sig_cc();
5823 assert(sig_cc != NULL, "must have scalarized signature");
5824
5825 // Get unscalarized calling convention
5826 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, sig_cc->length());
5827 int args_passed = 0;
5828 if (!method->is_static()) {
5829 sig_bt[args_passed++] = T_OBJECT;
5830 }
5831 if (!receiver_only) {
5832 for (SignatureStream ss(method->signature()); !ss.at_return_type(); ss.next()) {
5833 BasicType bt = ss.type();
5834 sig_bt[args_passed++] = bt;
5835 if (type2size[bt] == 2) {
5836 sig_bt[args_passed++] = T_VOID;
5837 }
5838 }
5839 } else {
5840 // Only unpack the receiver, all other arguments are already scalarized
5841 InstanceKlass* holder = method->method_holder();
5842 int rec_len = holder->is_value() ? ValueKlass::cast(holder)->extended_sig()->length() : 1;
5843 // Copy scalarized signature but skip receiver, value type delimiters and reserved entries
5844 for (int i = 0; i < sig_cc->length(); i++) {
5845 if (!SigEntry::is_reserved_entry(sig_cc, i)) {
5846 if (SigEntry::skip_value_delimiters(sig_cc, i) && rec_len <= 0) {
5847 sig_bt[args_passed++] = sig_cc->at(i)._bt;
5848 }
5849 rec_len--;
5850 }
5851 }
5852 }
5853 VMRegPair* regs = NEW_RESOURCE_ARRAY(VMRegPair, args_passed);
5854 int args_on_stack = SharedRuntime::java_calling_convention(sig_bt, regs, args_passed, false);
5855
5856 // Get scalarized calling convention
5857 int args_passed_cc = SigEntry::fill_sig_bt(sig_cc, sig_bt);
5858 VMRegPair* regs_cc = NEW_RESOURCE_ARRAY(VMRegPair, sig_cc->length());
5859 int args_on_stack_cc = SharedRuntime::java_calling_convention(sig_bt, regs_cc, args_passed_cc, false);
5860
5861 // Check if we need to extend the stack for unpacking
5862 int sp_inc = (args_on_stack_cc - args_on_stack) * VMRegImpl::stack_slot_size;
5863 if (sp_inc > 0) {
5864 // Save the return address, adjust the stack (make sure it is properly
5865 // 16-byte aligned) and copy the return address to the new top of the stack.
5866 pop(r13);
5867 sp_inc = align_up(sp_inc, StackAlignmentInBytes);
5868 subptr(rsp, sp_inc);
5869 push(r13);
5870 } else {
5871 // The scalarized calling convention needs less stack space than the unscalarized one.
5872 // No need to extend the stack, the caller will take care of these adjustments.
5873 sp_inc = 0;
5874 }
5875
5876 // Initialize register/stack slot states (make all writable)
5877 int max_stack = MAX2(args_on_stack + sp_inc/VMRegImpl::stack_slot_size, args_on_stack_cc);
5878 int max_reg = VMRegImpl::stack2reg(max_stack)->value();
5879 RegState* reg_state = NEW_RESOURCE_ARRAY(RegState, max_reg);
5880 for (int i = 0; i < max_reg; ++i) {
5881 reg_state[i] = reg_writable;
5882 }
5883 // Set all source registers/stack slots to readonly to prevent accidental overwriting
5884 for (int i = 0; i < args_passed; ++i) {
5885 VMReg reg = regs[i].first();
5886 if (!reg->is_valid()) continue;
5887 if (reg->is_stack()) {
5888 // Update source stack location by adding stack increment
5889 reg = VMRegImpl::stack2reg(reg->reg2stack() + sp_inc/VMRegImpl::stack_slot_size);
5890 regs[i] = reg;
5891 }
5892 assert(reg->value() >= 0 && reg->value() < max_reg, "reg value out of bounds");
5893 reg_state[reg->value()] = reg_readonly;
5894 }
5895
5896 // Emit code for unpacking value type arguments
5897 // We try multiple times and eventually start spilling to resolve (circular) dependencies
5898 bool done = false;
5899 for (int i = 0; i < 2*args_passed_cc && !done; ++i) {
5900 done = true;
5901 bool spill = (i > args_passed_cc); // Start spilling?
5902 // Iterate over all arguments (in reverse)
5903 for (int from_index = args_passed-1, to_index = args_passed_cc-1, sig_index = sig_cc->length()-1; sig_index >= 0; sig_index--) {
5904 if (SigEntry::is_reserved_entry(sig_cc, sig_index)) {
5905 to_index--; // Skip reserved entry
5906 } else {
5907 assert(from_index >= 0, "index out of bounds");
5908 VMReg reg = regs[from_index].first();
5909 if (spill && reg->is_valid() && reg_state[reg->value()] == reg_readonly) {
5910 // Spill argument to be able to write the source and resolve circular dependencies
5911 VMReg spill_reg = reg->is_XMMRegister() ? xmm8->as_VMReg() : r14->as_VMReg();
5912 bool res = move_helper(reg, spill_reg, T_DOUBLE, reg_state, sp_inc);
5913 assert(res, "Spilling should not fail");
5914 // Set spill_reg as new source and update state
5915 reg = spill_reg;
5916 regs[from_index].set1(reg);
5917 reg_state[reg->value()] = reg_readonly;
5918 spill = false; // Do not spill again in this round
5919 }
5920 BasicType bt = sig_cc->at(sig_index)._bt;
5921 if (SigEntry::skip_value_delimiters(sig_cc, sig_index)) {
5922 assert(to_index >= 0, "index out of bounds");
5923 done &= move_helper(reg, regs_cc[to_index].first(), bt, reg_state, sp_inc);
5924 to_index--;
5925 } else if (!receiver_only || (from_index == 0 && bt == T_VOID)) {
5926 done &= unpack_value_helper(sig_cc, sig_index, reg, regs_cc, to_index, reg_state, sp_inc);
5927 } else {
5928 continue;
5929 }
5930 from_index--;
5931 }
5932 }
5933 }
5934 guarantee(done, "Could not resolve circular dependency when unpacking value type arguments");
5935
5936 // Emit code for verified entry and save increment for stack repair on return
5937 verified_entry(C, sp_inc);
5938 }
5939
5940 // Restores the stack on return
5941 void MacroAssembler::restore_stack(Compile* C) {
5942 int framesize = C->frame_size_in_bytes();
5943 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
5944 // Remove word for return addr already pushed and RBP
5945 framesize -= 2*wordSize;
5946
5947 if (C->needs_stack_repair()) {
5948 // Restore rbp and repair rsp by adding the stack increment
5949 movq(rbp, Address(rsp, framesize));
5950 addq(rsp, Address(rsp, C->sp_inc_offset()));
5951 } else {
5952 if (framesize > 0) {
5953 addq(rsp, framesize);
5954 }
5955 pop(rbp);
5956 }
5957 }
5958
5959 void MacroAssembler::clear_mem(Register base, Register cnt, Register val, XMMRegister xtmp, bool is_large, bool word_copy_only) {
5960 // cnt - number of qwords (8-byte words).
5961 // base - start address, qword aligned.
5962 // is_large - if optimizers know cnt is larger than InitArrayShortSize
5963 assert(base==rdi, "base register must be edi for rep stos");
5964 assert(val==rax, "tmp register must be eax for rep stos");
5965 assert(cnt==rcx, "cnt register must be ecx for rep stos");
5966 assert(InitArrayShortSize % BytesPerLong == 0,
5967 "InitArrayShortSize should be the multiple of BytesPerLong");
5968
5969 Label DONE;
5970
5971 if (!is_large) {
5972 Label LOOP, LONG;
5973 cmpptr(cnt, InitArrayShortSize/BytesPerLong);
5974 jccb(Assembler::greater, LONG);
5975
5976 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
5977
5978 decrement(cnt);
5979 jccb(Assembler::negative, DONE); // Zero length
5980
5981 // Use individual pointer-sized stores for small counts:
5982 BIND(LOOP);
5983 movptr(Address(base, cnt, Address::times_ptr), val);
5984 decrement(cnt);
5985 jccb(Assembler::greaterEqual, LOOP);
5986 jmpb(DONE);
5987
5988 BIND(LONG);
5989 }
5990
5991 // Use longer rep-prefixed ops for non-small counts:
5992 if (UseFastStosb && !word_copy_only) {
5993 shlptr(cnt, 3); // convert to number of bytes
5994 rep_stosb();
5995 } else if (UseXMMForObjInit) {
5996 xmm_clear_mem(base, cnt, val, xtmp);
5997 } else {
5998 NOT_LP64(shlptr(cnt, 1);) // convert to number of 32-bit words for 32-bit VM
5999 rep_stos();
6000 }
6001
6002 BIND(DONE);
6003 }
6004
6005 #ifdef COMPILER2
6006
6007 // IndexOf for constant substrings with size >= 8 chars
6008 // which don't need to be loaded through stack.
6009 void MacroAssembler::string_indexofC8(Register str1, Register str2,
6010 Register cnt1, Register cnt2,
6011 int int_cnt2, Register result,
6012 XMMRegister vec, Register tmp,
6013 int ae) {
6014 ShortBranchVerifier sbv(this);
6015 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
6016 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
6017
6018 // This method uses the pcmpestri instruction with bound registers
6019 // inputs:
6020 // xmm - substring
6021 // rax - substring length (elements count)
6022 // mem - scanned string
6023 // rdx - string length (elements count)
6024 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6025 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
6026 // outputs:
6027 // rcx - matched index in string
6028 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6029 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
6030 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
6031 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
6032 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
6033
6034 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR,
6035 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR,
6036 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE;
6037
6038 // Note, inline_string_indexOf() generates checks:
6039 // if (substr.count > string.count) return -1;
6040 // if (substr.count == 0) return 0;
6041 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars");
6042
6043 // Load substring.
6044 if (ae == StrIntrinsicNode::UL) {
6045 pmovzxbw(vec, Address(str2, 0));
6046 } else {
6047 movdqu(vec, Address(str2, 0));
6048 }
6049 movl(cnt2, int_cnt2);
6050 movptr(result, str1); // string addr
6051
6052 if (int_cnt2 > stride) {
6053 jmpb(SCAN_TO_SUBSTR);
6054
6055 // Reload substr for rescan, this code
6056 // is executed only for large substrings (> 8 chars)
6057 bind(RELOAD_SUBSTR);
6058 if (ae == StrIntrinsicNode::UL) {
6059 pmovzxbw(vec, Address(str2, 0));
6060 } else {
6061 movdqu(vec, Address(str2, 0));
6062 }
6063 negptr(cnt2); // Jumped here with negative cnt2, convert to positive
6064
6065 bind(RELOAD_STR);
6066 // We came here after the beginning of the substring was
6067 // matched but the rest of it was not so we need to search
6068 // again. Start from the next element after the previous match.
6069
6070 // cnt2 is number of substring reminding elements and
6071 // cnt1 is number of string reminding elements when cmp failed.
6072 // Restored cnt1 = cnt1 - cnt2 + int_cnt2
6073 subl(cnt1, cnt2);
6074 addl(cnt1, int_cnt2);
6075 movl(cnt2, int_cnt2); // Now restore cnt2
6076
6077 decrementl(cnt1); // Shift to next element
6078 cmpl(cnt1, cnt2);
6079 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6080
6081 addptr(result, (1<<scale1));
6082
6083 } // (int_cnt2 > 8)
6084
6085 // Scan string for start of substr in 16-byte vectors
6086 bind(SCAN_TO_SUBSTR);
6087 pcmpestri(vec, Address(result, 0), mode);
6088 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
6089 subl(cnt1, stride);
6090 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6091 cmpl(cnt1, cnt2);
6092 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6093 addptr(result, 16);
6094 jmpb(SCAN_TO_SUBSTR);
6095
6096 // Found a potential substr
6097 bind(FOUND_CANDIDATE);
6098 // Matched whole vector if first element matched (tmp(rcx) == 0).
6099 if (int_cnt2 == stride) {
6100 jccb(Assembler::overflow, RET_FOUND); // OF == 1
6101 } else { // int_cnt2 > 8
6102 jccb(Assembler::overflow, FOUND_SUBSTR);
6103 }
6104 // After pcmpestri tmp(rcx) contains matched element index
6105 // Compute start addr of substr
6106 lea(result, Address(result, tmp, scale1));
6107
6108 // Make sure string is still long enough
6109 subl(cnt1, tmp);
6110 cmpl(cnt1, cnt2);
6111 if (int_cnt2 == stride) {
6112 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6113 } else { // int_cnt2 > 8
6114 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD);
6115 }
6116 // Left less then substring.
6117
6118 bind(RET_NOT_FOUND);
6119 movl(result, -1);
6120 jmp(EXIT);
6121
6122 if (int_cnt2 > stride) {
6123 // This code is optimized for the case when whole substring
6124 // is matched if its head is matched.
6125 bind(MATCH_SUBSTR_HEAD);
6126 pcmpestri(vec, Address(result, 0), mode);
6127 // Reload only string if does not match
6128 jcc(Assembler::noOverflow, RELOAD_STR); // OF == 0
6129
6130 Label CONT_SCAN_SUBSTR;
6131 // Compare the rest of substring (> 8 chars).
6132 bind(FOUND_SUBSTR);
6133 // First 8 chars are already matched.
6134 negptr(cnt2);
6135 addptr(cnt2, stride);
6136
6137 bind(SCAN_SUBSTR);
6138 subl(cnt1, stride);
6139 cmpl(cnt2, -stride); // Do not read beyond substring
6140 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR);
6141 // Back-up strings to avoid reading beyond substring:
6142 // cnt1 = cnt1 - cnt2 + 8
6143 addl(cnt1, cnt2); // cnt2 is negative
6144 addl(cnt1, stride);
6145 movl(cnt2, stride); negptr(cnt2);
6146 bind(CONT_SCAN_SUBSTR);
6147 if (int_cnt2 < (int)G) {
6148 int tail_off1 = int_cnt2<<scale1;
6149 int tail_off2 = int_cnt2<<scale2;
6150 if (ae == StrIntrinsicNode::UL) {
6151 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2));
6152 } else {
6153 movdqu(vec, Address(str2, cnt2, scale2, tail_off2));
6154 }
6155 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode);
6156 } else {
6157 // calculate index in register to avoid integer overflow (int_cnt2*2)
6158 movl(tmp, int_cnt2);
6159 addptr(tmp, cnt2);
6160 if (ae == StrIntrinsicNode::UL) {
6161 pmovzxbw(vec, Address(str2, tmp, scale2, 0));
6162 } else {
6163 movdqu(vec, Address(str2, tmp, scale2, 0));
6164 }
6165 pcmpestri(vec, Address(result, tmp, scale1, 0), mode);
6166 }
6167 // Need to reload strings pointers if not matched whole vector
6168 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6169 addptr(cnt2, stride);
6170 jcc(Assembler::negative, SCAN_SUBSTR);
6171 // Fall through if found full substring
6172
6173 } // (int_cnt2 > 8)
6174
6175 bind(RET_FOUND);
6176 // Found result if we matched full small substring.
6177 // Compute substr offset
6178 subptr(result, str1);
6179 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6180 shrl(result, 1); // index
6181 }
6182 bind(EXIT);
6183
6184 } // string_indexofC8
6185
6186 // Small strings are loaded through stack if they cross page boundary.
6187 void MacroAssembler::string_indexof(Register str1, Register str2,
6188 Register cnt1, Register cnt2,
6189 int int_cnt2, Register result,
6190 XMMRegister vec, Register tmp,
6191 int ae) {
6192 ShortBranchVerifier sbv(this);
6193 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
6194 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
6195
6196 //
6197 // int_cnt2 is length of small (< 8 chars) constant substring
6198 // or (-1) for non constant substring in which case its length
6199 // is in cnt2 register.
6200 //
6201 // Note, inline_string_indexOf() generates checks:
6202 // if (substr.count > string.count) return -1;
6203 // if (substr.count == 0) return 0;
6204 //
6205 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
6206 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0");
6207 // This method uses the pcmpestri instruction with bound registers
6208 // inputs:
6209 // xmm - substring
6210 // rax - substring length (elements count)
6211 // mem - scanned string
6212 // rdx - string length (elements count)
6213 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
6214 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
6215 // outputs:
6216 // rcx - matched index in string
6217 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6218 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
6219 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
6220 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
6221
6222 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR,
6223 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR,
6224 FOUND_CANDIDATE;
6225
6226 { //========================================================
6227 // We don't know where these strings are located
6228 // and we can't read beyond them. Load them through stack.
6229 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR;
6230
6231 movptr(tmp, rsp); // save old SP
6232
6233 if (int_cnt2 > 0) { // small (< 8 chars) constant substring
6234 if (int_cnt2 == (1>>scale2)) { // One byte
6235 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding");
6236 load_unsigned_byte(result, Address(str2, 0));
6237 movdl(vec, result); // move 32 bits
6238 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes
6239 // Not enough header space in 32-bit VM: 12+3 = 15.
6240 movl(result, Address(str2, -1));
6241 shrl(result, 8);
6242 movdl(vec, result); // move 32 bits
6243 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char
6244 load_unsigned_short(result, Address(str2, 0));
6245 movdl(vec, result); // move 32 bits
6246 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars
6247 movdl(vec, Address(str2, 0)); // move 32 bits
6248 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars
6249 movq(vec, Address(str2, 0)); // move 64 bits
6250 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7})
6251 // Array header size is 12 bytes in 32-bit VM
6252 // + 6 bytes for 3 chars == 18 bytes,
6253 // enough space to load vec and shift.
6254 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity");
6255 if (ae == StrIntrinsicNode::UL) {
6256 int tail_off = int_cnt2-8;
6257 pmovzxbw(vec, Address(str2, tail_off));
6258 psrldq(vec, -2*tail_off);
6259 }
6260 else {
6261 int tail_off = int_cnt2*(1<<scale2);
6262 movdqu(vec, Address(str2, tail_off-16));
6263 psrldq(vec, 16-tail_off);
6264 }
6265 }
6266 } else { // not constant substring
6267 cmpl(cnt2, stride);
6268 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough
6269
6270 // We can read beyond string if srt+16 does not cross page boundary
6271 // since heaps are aligned and mapped by pages.
6272 assert(os::vm_page_size() < (int)G, "default page should be small");
6273 movl(result, str2); // We need only low 32 bits
6274 andl(result, (os::vm_page_size()-1));
6275 cmpl(result, (os::vm_page_size()-16));
6276 jccb(Assembler::belowEqual, CHECK_STR);
6277
6278 // Move small strings to stack to allow load 16 bytes into vec.
6279 subptr(rsp, 16);
6280 int stk_offset = wordSize-(1<<scale2);
6281 push(cnt2);
6282
6283 bind(COPY_SUBSTR);
6284 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) {
6285 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1));
6286 movb(Address(rsp, cnt2, scale2, stk_offset), result);
6287 } else if (ae == StrIntrinsicNode::UU) {
6288 load_unsigned_short(result, Address(str2, cnt2, scale2, -2));
6289 movw(Address(rsp, cnt2, scale2, stk_offset), result);
6290 }
6291 decrement(cnt2);
6292 jccb(Assembler::notZero, COPY_SUBSTR);
6293
6294 pop(cnt2);
6295 movptr(str2, rsp); // New substring address
6296 } // non constant
6297
6298 bind(CHECK_STR);
6299 cmpl(cnt1, stride);
6300 jccb(Assembler::aboveEqual, BIG_STRINGS);
6301
6302 // Check cross page boundary.
6303 movl(result, str1); // We need only low 32 bits
6304 andl(result, (os::vm_page_size()-1));
6305 cmpl(result, (os::vm_page_size()-16));
6306 jccb(Assembler::belowEqual, BIG_STRINGS);
6307
6308 subptr(rsp, 16);
6309 int stk_offset = -(1<<scale1);
6310 if (int_cnt2 < 0) { // not constant
6311 push(cnt2);
6312 stk_offset += wordSize;
6313 }
6314 movl(cnt2, cnt1);
6315
6316 bind(COPY_STR);
6317 if (ae == StrIntrinsicNode::LL) {
6318 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1));
6319 movb(Address(rsp, cnt2, scale1, stk_offset), result);
6320 } else {
6321 load_unsigned_short(result, Address(str1, cnt2, scale1, -2));
6322 movw(Address(rsp, cnt2, scale1, stk_offset), result);
6323 }
6324 decrement(cnt2);
6325 jccb(Assembler::notZero, COPY_STR);
6326
6327 if (int_cnt2 < 0) { // not constant
6328 pop(cnt2);
6329 }
6330 movptr(str1, rsp); // New string address
6331
6332 bind(BIG_STRINGS);
6333 // Load substring.
6334 if (int_cnt2 < 0) { // -1
6335 if (ae == StrIntrinsicNode::UL) {
6336 pmovzxbw(vec, Address(str2, 0));
6337 } else {
6338 movdqu(vec, Address(str2, 0));
6339 }
6340 push(cnt2); // substr count
6341 push(str2); // substr addr
6342 push(str1); // string addr
6343 } else {
6344 // Small (< 8 chars) constant substrings are loaded already.
6345 movl(cnt2, int_cnt2);
6346 }
6347 push(tmp); // original SP
6348
6349 } // Finished loading
6350
6351 //========================================================
6352 // Start search
6353 //
6354
6355 movptr(result, str1); // string addr
6356
6357 if (int_cnt2 < 0) { // Only for non constant substring
6358 jmpb(SCAN_TO_SUBSTR);
6359
6360 // SP saved at sp+0
6361 // String saved at sp+1*wordSize
6362 // Substr saved at sp+2*wordSize
6363 // Substr count saved at sp+3*wordSize
6364
6365 // Reload substr for rescan, this code
6366 // is executed only for large substrings (> 8 chars)
6367 bind(RELOAD_SUBSTR);
6368 movptr(str2, Address(rsp, 2*wordSize));
6369 movl(cnt2, Address(rsp, 3*wordSize));
6370 if (ae == StrIntrinsicNode::UL) {
6371 pmovzxbw(vec, Address(str2, 0));
6372 } else {
6373 movdqu(vec, Address(str2, 0));
6374 }
6375 // We came here after the beginning of the substring was
6376 // matched but the rest of it was not so we need to search
6377 // again. Start from the next element after the previous match.
6378 subptr(str1, result); // Restore counter
6379 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6380 shrl(str1, 1);
6381 }
6382 addl(cnt1, str1);
6383 decrementl(cnt1); // Shift to next element
6384 cmpl(cnt1, cnt2);
6385 jcc(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6386
6387 addptr(result, (1<<scale1));
6388 } // non constant
6389
6390 // Scan string for start of substr in 16-byte vectors
6391 bind(SCAN_TO_SUBSTR);
6392 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
6393 pcmpestri(vec, Address(result, 0), mode);
6394 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
6395 subl(cnt1, stride);
6396 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
6397 cmpl(cnt1, cnt2);
6398 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
6399 addptr(result, 16);
6400
6401 bind(ADJUST_STR);
6402 cmpl(cnt1, stride); // Do not read beyond string
6403 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
6404 // Back-up string to avoid reading beyond string.
6405 lea(result, Address(result, cnt1, scale1, -16));
6406 movl(cnt1, stride);
6407 jmpb(SCAN_TO_SUBSTR);
6408
6409 // Found a potential substr
6410 bind(FOUND_CANDIDATE);
6411 // After pcmpestri tmp(rcx) contains matched element index
6412
6413 // Make sure string is still long enough
6414 subl(cnt1, tmp);
6415 cmpl(cnt1, cnt2);
6416 jccb(Assembler::greaterEqual, FOUND_SUBSTR);
6417 // Left less then substring.
6418
6419 bind(RET_NOT_FOUND);
6420 movl(result, -1);
6421 jmp(CLEANUP);
6422
6423 bind(FOUND_SUBSTR);
6424 // Compute start addr of substr
6425 lea(result, Address(result, tmp, scale1));
6426 if (int_cnt2 > 0) { // Constant substring
6427 // Repeat search for small substring (< 8 chars)
6428 // from new point without reloading substring.
6429 // Have to check that we don't read beyond string.
6430 cmpl(tmp, stride-int_cnt2);
6431 jccb(Assembler::greater, ADJUST_STR);
6432 // Fall through if matched whole substring.
6433 } else { // non constant
6434 assert(int_cnt2 == -1, "should be != 0");
6435
6436 addl(tmp, cnt2);
6437 // Found result if we matched whole substring.
6438 cmpl(tmp, stride);
6439 jcc(Assembler::lessEqual, RET_FOUND);
6440
6441 // Repeat search for small substring (<= 8 chars)
6442 // from new point 'str1' without reloading substring.
6443 cmpl(cnt2, stride);
6444 // Have to check that we don't read beyond string.
6445 jccb(Assembler::lessEqual, ADJUST_STR);
6446
6447 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG;
6448 // Compare the rest of substring (> 8 chars).
6449 movptr(str1, result);
6450
6451 cmpl(tmp, cnt2);
6452 // First 8 chars are already matched.
6453 jccb(Assembler::equal, CHECK_NEXT);
6454
6455 bind(SCAN_SUBSTR);
6456 pcmpestri(vec, Address(str1, 0), mode);
6457 // Need to reload strings pointers if not matched whole vector
6458 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
6459
6460 bind(CHECK_NEXT);
6461 subl(cnt2, stride);
6462 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring
6463 addptr(str1, 16);
6464 if (ae == StrIntrinsicNode::UL) {
6465 addptr(str2, 8);
6466 } else {
6467 addptr(str2, 16);
6468 }
6469 subl(cnt1, stride);
6470 cmpl(cnt2, stride); // Do not read beyond substring
6471 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR);
6472 // Back-up strings to avoid reading beyond substring.
6473
6474 if (ae == StrIntrinsicNode::UL) {
6475 lea(str2, Address(str2, cnt2, scale2, -8));
6476 lea(str1, Address(str1, cnt2, scale1, -16));
6477 } else {
6478 lea(str2, Address(str2, cnt2, scale2, -16));
6479 lea(str1, Address(str1, cnt2, scale1, -16));
6480 }
6481 subl(cnt1, cnt2);
6482 movl(cnt2, stride);
6483 addl(cnt1, stride);
6484 bind(CONT_SCAN_SUBSTR);
6485 if (ae == StrIntrinsicNode::UL) {
6486 pmovzxbw(vec, Address(str2, 0));
6487 } else {
6488 movdqu(vec, Address(str2, 0));
6489 }
6490 jmp(SCAN_SUBSTR);
6491
6492 bind(RET_FOUND_LONG);
6493 movptr(str1, Address(rsp, wordSize));
6494 } // non constant
6495
6496 bind(RET_FOUND);
6497 // Compute substr offset
6498 subptr(result, str1);
6499 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
6500 shrl(result, 1); // index
6501 }
6502 bind(CLEANUP);
6503 pop(rsp); // restore SP
6504
6505 } // string_indexof
6506
6507 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result,
6508 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) {
6509 ShortBranchVerifier sbv(this);
6510 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
6511
6512 int stride = 8;
6513
6514 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP,
6515 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP,
6516 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT,
6517 FOUND_SEQ_CHAR, DONE_LABEL;
6518
6519 movptr(result, str1);
6520 if (UseAVX >= 2) {
6521 cmpl(cnt1, stride);
6522 jcc(Assembler::less, SCAN_TO_CHAR_LOOP);
6523 cmpl(cnt1, 2*stride);
6524 jcc(Assembler::less, SCAN_TO_8_CHAR_INIT);
6525 movdl(vec1, ch);
6526 vpbroadcastw(vec1, vec1, Assembler::AVX_256bit);
6527 vpxor(vec2, vec2);
6528 movl(tmp, cnt1);
6529 andl(tmp, 0xFFFFFFF0); //vector count (in chars)
6530 andl(cnt1,0x0000000F); //tail count (in chars)
6531
6532 bind(SCAN_TO_16_CHAR_LOOP);
6533 vmovdqu(vec3, Address(result, 0));
6534 vpcmpeqw(vec3, vec3, vec1, 1);
6535 vptest(vec2, vec3);
6536 jcc(Assembler::carryClear, FOUND_CHAR);
6537 addptr(result, 32);
6538 subl(tmp, 2*stride);
6539 jcc(Assembler::notZero, SCAN_TO_16_CHAR_LOOP);
6540 jmp(SCAN_TO_8_CHAR);
6541 bind(SCAN_TO_8_CHAR_INIT);
6542 movdl(vec1, ch);
6543 pshuflw(vec1, vec1, 0x00);
6544 pshufd(vec1, vec1, 0);
6545 pxor(vec2, vec2);
6546 }
6547 bind(SCAN_TO_8_CHAR);
6548 cmpl(cnt1, stride);
6549 if (UseAVX >= 2) {
6550 jcc(Assembler::less, SCAN_TO_CHAR);
6551 } else {
6552 jcc(Assembler::less, SCAN_TO_CHAR_LOOP);
6553 movdl(vec1, ch);
6554 pshuflw(vec1, vec1, 0x00);
6555 pshufd(vec1, vec1, 0);
6556 pxor(vec2, vec2);
6557 }
6558 movl(tmp, cnt1);
6559 andl(tmp, 0xFFFFFFF8); //vector count (in chars)
6560 andl(cnt1,0x00000007); //tail count (in chars)
6561
6562 bind(SCAN_TO_8_CHAR_LOOP);
6563 movdqu(vec3, Address(result, 0));
6564 pcmpeqw(vec3, vec1);
6565 ptest(vec2, vec3);
6566 jcc(Assembler::carryClear, FOUND_CHAR);
6567 addptr(result, 16);
6568 subl(tmp, stride);
6569 jcc(Assembler::notZero, SCAN_TO_8_CHAR_LOOP);
6570 bind(SCAN_TO_CHAR);
6571 testl(cnt1, cnt1);
6572 jcc(Assembler::zero, RET_NOT_FOUND);
6573 bind(SCAN_TO_CHAR_LOOP);
6574 load_unsigned_short(tmp, Address(result, 0));
6575 cmpl(ch, tmp);
6576 jccb(Assembler::equal, FOUND_SEQ_CHAR);
6577 addptr(result, 2);
6578 subl(cnt1, 1);
6579 jccb(Assembler::zero, RET_NOT_FOUND);
6580 jmp(SCAN_TO_CHAR_LOOP);
6581
6582 bind(RET_NOT_FOUND);
6583 movl(result, -1);
6584 jmpb(DONE_LABEL);
6585
6586 bind(FOUND_CHAR);
6587 if (UseAVX >= 2) {
6588 vpmovmskb(tmp, vec3);
6589 } else {
6590 pmovmskb(tmp, vec3);
6591 }
6592 bsfl(ch, tmp);
6593 addl(result, ch);
6594
6595 bind(FOUND_SEQ_CHAR);
6596 subptr(result, str1);
6597 shrl(result, 1);
6598
6599 bind(DONE_LABEL);
6600 } // string_indexof_char
6601
6602 // helper function for string_compare
6603 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2,
6604 Address::ScaleFactor scale, Address::ScaleFactor scale1,
6605 Address::ScaleFactor scale2, Register index, int ae) {
6606 if (ae == StrIntrinsicNode::LL) {
6607 load_unsigned_byte(elem1, Address(str1, index, scale, 0));
6608 load_unsigned_byte(elem2, Address(str2, index, scale, 0));
6609 } else if (ae == StrIntrinsicNode::UU) {
6610 load_unsigned_short(elem1, Address(str1, index, scale, 0));
6611 load_unsigned_short(elem2, Address(str2, index, scale, 0));
6612 } else {
6613 load_unsigned_byte(elem1, Address(str1, index, scale1, 0));
6614 load_unsigned_short(elem2, Address(str2, index, scale2, 0));
6615 }
6616 }
6617
6618 // Compare strings, used for char[] and byte[].
6619 void MacroAssembler::string_compare(Register str1, Register str2,
6620 Register cnt1, Register cnt2, Register result,
6621 XMMRegister vec1, int ae) {
6622 ShortBranchVerifier sbv(this);
6623 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
6624 Label COMPARE_WIDE_VECTORS_LOOP_FAILED; // used only _LP64 && AVX3
6625 int stride, stride2, adr_stride, adr_stride1, adr_stride2;
6626 int stride2x2 = 0x40;
6627 Address::ScaleFactor scale = Address::no_scale;
6628 Address::ScaleFactor scale1 = Address::no_scale;
6629 Address::ScaleFactor scale2 = Address::no_scale;
6630
6631 if (ae != StrIntrinsicNode::LL) {
6632 stride2x2 = 0x20;
6633 }
6634
6635 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
6636 shrl(cnt2, 1);
6637 }
6638 // Compute the minimum of the string lengths and the
6639 // difference of the string lengths (stack).
6640 // Do the conditional move stuff
6641 movl(result, cnt1);
6642 subl(cnt1, cnt2);
6643 push(cnt1);
6644 cmov32(Assembler::lessEqual, cnt2, result); // cnt2 = min(cnt1, cnt2)
6645
6646 // Is the minimum length zero?
6647 testl(cnt2, cnt2);
6648 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6649 if (ae == StrIntrinsicNode::LL) {
6650 // Load first bytes
6651 load_unsigned_byte(result, Address(str1, 0)); // result = str1[0]
6652 load_unsigned_byte(cnt1, Address(str2, 0)); // cnt1 = str2[0]
6653 } else if (ae == StrIntrinsicNode::UU) {
6654 // Load first characters
6655 load_unsigned_short(result, Address(str1, 0));
6656 load_unsigned_short(cnt1, Address(str2, 0));
6657 } else {
6658 load_unsigned_byte(result, Address(str1, 0));
6659 load_unsigned_short(cnt1, Address(str2, 0));
6660 }
6661 subl(result, cnt1);
6662 jcc(Assembler::notZero, POP_LABEL);
6663
6664 if (ae == StrIntrinsicNode::UU) {
6665 // Divide length by 2 to get number of chars
6666 shrl(cnt2, 1);
6667 }
6668 cmpl(cnt2, 1);
6669 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6670
6671 // Check if the strings start at the same location and setup scale and stride
6672 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6673 cmpptr(str1, str2);
6674 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
6675 if (ae == StrIntrinsicNode::LL) {
6676 scale = Address::times_1;
6677 stride = 16;
6678 } else {
6679 scale = Address::times_2;
6680 stride = 8;
6681 }
6682 } else {
6683 scale1 = Address::times_1;
6684 scale2 = Address::times_2;
6685 // scale not used
6686 stride = 8;
6687 }
6688
6689 if (UseAVX >= 2 && UseSSE42Intrinsics) {
6690 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
6691 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
6692 Label COMPARE_WIDE_VECTORS_LOOP_AVX2;
6693 Label COMPARE_TAIL_LONG;
6694 Label COMPARE_WIDE_VECTORS_LOOP_AVX3; // used only _LP64 && AVX3
6695
6696 int pcmpmask = 0x19;
6697 if (ae == StrIntrinsicNode::LL) {
6698 pcmpmask &= ~0x01;
6699 }
6700
6701 // Setup to compare 16-chars (32-bytes) vectors,
6702 // start from first character again because it has aligned address.
6703 if (ae == StrIntrinsicNode::LL) {
6704 stride2 = 32;
6705 } else {
6706 stride2 = 16;
6707 }
6708 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6709 adr_stride = stride << scale;
6710 } else {
6711 adr_stride1 = 8; //stride << scale1;
6712 adr_stride2 = 16; //stride << scale2;
6713 }
6714
6715 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6716 // rax and rdx are used by pcmpestri as elements counters
6717 movl(result, cnt2);
6718 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count
6719 jcc(Assembler::zero, COMPARE_TAIL_LONG);
6720
6721 // fast path : compare first 2 8-char vectors.
6722 bind(COMPARE_16_CHARS);
6723 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6724 movdqu(vec1, Address(str1, 0));
6725 } else {
6726 pmovzxbw(vec1, Address(str1, 0));
6727 }
6728 pcmpestri(vec1, Address(str2, 0), pcmpmask);
6729 jccb(Assembler::below, COMPARE_INDEX_CHAR);
6730
6731 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6732 movdqu(vec1, Address(str1, adr_stride));
6733 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
6734 } else {
6735 pmovzxbw(vec1, Address(str1, adr_stride1));
6736 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask);
6737 }
6738 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
6739 addl(cnt1, stride);
6740
6741 // Compare the characters at index in cnt1
6742 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character
6743 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
6744 subl(result, cnt2);
6745 jmp(POP_LABEL);
6746
6747 // Setup the registers to start vector comparison loop
6748 bind(COMPARE_WIDE_VECTORS);
6749 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6750 lea(str1, Address(str1, result, scale));
6751 lea(str2, Address(str2, result, scale));
6752 } else {
6753 lea(str1, Address(str1, result, scale1));
6754 lea(str2, Address(str2, result, scale2));
6755 }
6756 subl(result, stride2);
6757 subl(cnt2, stride2);
6758 jcc(Assembler::zero, COMPARE_WIDE_TAIL);
6759 negptr(result);
6760
6761 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
6762 bind(COMPARE_WIDE_VECTORS_LOOP);
6763
6764 #ifdef _LP64
6765 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
6766 cmpl(cnt2, stride2x2);
6767 jccb(Assembler::below, COMPARE_WIDE_VECTORS_LOOP_AVX2);
6768 testl(cnt2, stride2x2-1); // cnt2 holds the vector count
6769 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX2); // means we cannot subtract by 0x40
6770
6771 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
6772 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6773 evmovdquq(vec1, Address(str1, result, scale), Assembler::AVX_512bit);
6774 evpcmpeqb(k7, vec1, Address(str2, result, scale), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
6775 } else {
6776 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_512bit);
6777 evpcmpeqb(k7, vec1, Address(str2, result, scale2), Assembler::AVX_512bit); // k7 == 11..11, if operands equal, otherwise k7 has some 0
6778 }
6779 kortestql(k7, k7);
6780 jcc(Assembler::aboveEqual, COMPARE_WIDE_VECTORS_LOOP_FAILED); // miscompare
6781 addptr(result, stride2x2); // update since we already compared at this addr
6782 subl(cnt2, stride2x2); // and sub the size too
6783 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP_AVX3);
6784
6785 vpxor(vec1, vec1);
6786 jmpb(COMPARE_WIDE_TAIL);
6787 }//if (VM_Version::supports_avx512vlbw())
6788 #endif // _LP64
6789
6790
6791 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
6792 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6793 vmovdqu(vec1, Address(str1, result, scale));
6794 vpxor(vec1, Address(str2, result, scale));
6795 } else {
6796 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit);
6797 vpxor(vec1, Address(str2, result, scale2));
6798 }
6799 vptest(vec1, vec1);
6800 jcc(Assembler::notZero, VECTOR_NOT_EQUAL);
6801 addptr(result, stride2);
6802 subl(cnt2, stride2);
6803 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
6804 // clean upper bits of YMM registers
6805 vpxor(vec1, vec1);
6806
6807 // compare wide vectors tail
6808 bind(COMPARE_WIDE_TAIL);
6809 testptr(result, result);
6810 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6811
6812 movl(result, stride2);
6813 movl(cnt2, result);
6814 negptr(result);
6815 jmp(COMPARE_WIDE_VECTORS_LOOP_AVX2);
6816
6817 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
6818 bind(VECTOR_NOT_EQUAL);
6819 // clean upper bits of YMM registers
6820 vpxor(vec1, vec1);
6821 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6822 lea(str1, Address(str1, result, scale));
6823 lea(str2, Address(str2, result, scale));
6824 } else {
6825 lea(str1, Address(str1, result, scale1));
6826 lea(str2, Address(str2, result, scale2));
6827 }
6828 jmp(COMPARE_16_CHARS);
6829
6830 // Compare tail chars, length between 1 to 15 chars
6831 bind(COMPARE_TAIL_LONG);
6832 movl(cnt2, result);
6833 cmpl(cnt2, stride);
6834 jcc(Assembler::less, COMPARE_SMALL_STR);
6835
6836 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6837 movdqu(vec1, Address(str1, 0));
6838 } else {
6839 pmovzxbw(vec1, Address(str1, 0));
6840 }
6841 pcmpestri(vec1, Address(str2, 0), pcmpmask);
6842 jcc(Assembler::below, COMPARE_INDEX_CHAR);
6843 subptr(cnt2, stride);
6844 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6845 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6846 lea(str1, Address(str1, result, scale));
6847 lea(str2, Address(str2, result, scale));
6848 } else {
6849 lea(str1, Address(str1, result, scale1));
6850 lea(str2, Address(str2, result, scale2));
6851 }
6852 negptr(cnt2);
6853 jmpb(WHILE_HEAD_LABEL);
6854
6855 bind(COMPARE_SMALL_STR);
6856 } else if (UseSSE42Intrinsics) {
6857 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
6858 int pcmpmask = 0x19;
6859 // Setup to compare 8-char (16-byte) vectors,
6860 // start from first character again because it has aligned address.
6861 movl(result, cnt2);
6862 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count
6863 if (ae == StrIntrinsicNode::LL) {
6864 pcmpmask &= ~0x01;
6865 }
6866 jcc(Assembler::zero, COMPARE_TAIL);
6867 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6868 lea(str1, Address(str1, result, scale));
6869 lea(str2, Address(str2, result, scale));
6870 } else {
6871 lea(str1, Address(str1, result, scale1));
6872 lea(str2, Address(str2, result, scale2));
6873 }
6874 negptr(result);
6875
6876 // pcmpestri
6877 // inputs:
6878 // vec1- substring
6879 // rax - negative string length (elements count)
6880 // mem - scanned string
6881 // rdx - string length (elements count)
6882 // pcmpmask - cmp mode: 11000 (string compare with negated result)
6883 // + 00 (unsigned bytes) or + 01 (unsigned shorts)
6884 // outputs:
6885 // rcx - first mismatched element index
6886 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
6887
6888 bind(COMPARE_WIDE_VECTORS);
6889 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6890 movdqu(vec1, Address(str1, result, scale));
6891 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6892 } else {
6893 pmovzxbw(vec1, Address(str1, result, scale1));
6894 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
6895 }
6896 // After pcmpestri cnt1(rcx) contains mismatched element index
6897
6898 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1
6899 addptr(result, stride);
6900 subptr(cnt2, stride);
6901 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
6902
6903 // compare wide vectors tail
6904 testptr(result, result);
6905 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
6906
6907 movl(cnt2, stride);
6908 movl(result, stride);
6909 negptr(result);
6910 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6911 movdqu(vec1, Address(str1, result, scale));
6912 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
6913 } else {
6914 pmovzxbw(vec1, Address(str1, result, scale1));
6915 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
6916 }
6917 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
6918
6919 // Mismatched characters in the vectors
6920 bind(VECTOR_NOT_EQUAL);
6921 addptr(cnt1, result);
6922 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
6923 subl(result, cnt2);
6924 jmpb(POP_LABEL);
6925
6926 bind(COMPARE_TAIL); // limit is zero
6927 movl(cnt2, result);
6928 // Fallthru to tail compare
6929 }
6930 // Shift str2 and str1 to the end of the arrays, negate min
6931 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
6932 lea(str1, Address(str1, cnt2, scale));
6933 lea(str2, Address(str2, cnt2, scale));
6934 } else {
6935 lea(str1, Address(str1, cnt2, scale1));
6936 lea(str2, Address(str2, cnt2, scale2));
6937 }
6938 decrementl(cnt2); // first character was compared already
6939 negptr(cnt2);
6940
6941 // Compare the rest of the elements
6942 bind(WHILE_HEAD_LABEL);
6943 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae);
6944 subl(result, cnt1);
6945 jccb(Assembler::notZero, POP_LABEL);
6946 increment(cnt2);
6947 jccb(Assembler::notZero, WHILE_HEAD_LABEL);
6948
6949 // Strings are equal up to min length. Return the length difference.
6950 bind(LENGTH_DIFF_LABEL);
6951 pop(result);
6952 if (ae == StrIntrinsicNode::UU) {
6953 // Divide diff by 2 to get number of chars
6954 sarl(result, 1);
6955 }
6956 jmpb(DONE_LABEL);
6957
6958 #ifdef _LP64
6959 if (VM_Version::supports_avx512vlbw()) {
6960
6961 bind(COMPARE_WIDE_VECTORS_LOOP_FAILED);
6962
6963 kmovql(cnt1, k7);
6964 notq(cnt1);
6965 bsfq(cnt2, cnt1);
6966 if (ae != StrIntrinsicNode::LL) {
6967 // Divide diff by 2 to get number of chars
6968 sarl(cnt2, 1);
6969 }
6970 addq(result, cnt2);
6971 if (ae == StrIntrinsicNode::LL) {
6972 load_unsigned_byte(cnt1, Address(str2, result));
6973 load_unsigned_byte(result, Address(str1, result));
6974 } else if (ae == StrIntrinsicNode::UU) {
6975 load_unsigned_short(cnt1, Address(str2, result, scale));
6976 load_unsigned_short(result, Address(str1, result, scale));
6977 } else {
6978 load_unsigned_short(cnt1, Address(str2, result, scale2));
6979 load_unsigned_byte(result, Address(str1, result, scale1));
6980 }
6981 subl(result, cnt1);
6982 jmpb(POP_LABEL);
6983 }//if (VM_Version::supports_avx512vlbw())
6984 #endif // _LP64
6985
6986 // Discard the stored length difference
6987 bind(POP_LABEL);
6988 pop(cnt1);
6989
6990 // That's it
6991 bind(DONE_LABEL);
6992 if(ae == StrIntrinsicNode::UL) {
6993 negl(result);
6994 }
6995
6996 }
6997
6998 // Search for Non-ASCII character (Negative byte value) in a byte array,
6999 // return true if it has any and false otherwise.
7000 // ..\jdk\src\java.base\share\classes\java\lang\StringCoding.java
7001 // @HotSpotIntrinsicCandidate
7002 // private static boolean hasNegatives(byte[] ba, int off, int len) {
7003 // for (int i = off; i < off + len; i++) {
7004 // if (ba[i] < 0) {
7005 // return true;
7006 // }
7007 // }
7008 // return false;
7009 // }
7010 void MacroAssembler::has_negatives(Register ary1, Register len,
7011 Register result, Register tmp1,
7012 XMMRegister vec1, XMMRegister vec2) {
7013 // rsi: byte array
7014 // rcx: len
7015 // rax: result
7016 ShortBranchVerifier sbv(this);
7017 assert_different_registers(ary1, len, result, tmp1);
7018 assert_different_registers(vec1, vec2);
7019 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE;
7020
7021 // len == 0
7022 testl(len, len);
7023 jcc(Assembler::zero, FALSE_LABEL);
7024
7025 if ((UseAVX > 2) && // AVX512
7026 VM_Version::supports_avx512vlbw() &&
7027 VM_Version::supports_bmi2()) {
7028
7029 Label test_64_loop, test_tail;
7030 Register tmp3_aliased = len;
7031
7032 movl(tmp1, len);
7033 vpxor(vec2, vec2, vec2, Assembler::AVX_512bit);
7034
7035 andl(tmp1, 64 - 1); // tail count (in chars) 0x3F
7036 andl(len, ~(64 - 1)); // vector count (in chars)
7037 jccb(Assembler::zero, test_tail);
7038
7039 lea(ary1, Address(ary1, len, Address::times_1));
7040 negptr(len);
7041
7042 bind(test_64_loop);
7043 // Check whether our 64 elements of size byte contain negatives
7044 evpcmpgtb(k2, vec2, Address(ary1, len, Address::times_1), Assembler::AVX_512bit);
7045 kortestql(k2, k2);
7046 jcc(Assembler::notZero, TRUE_LABEL);
7047
7048 addptr(len, 64);
7049 jccb(Assembler::notZero, test_64_loop);
7050
7051
7052 bind(test_tail);
7053 // bail out when there is nothing to be done
7054 testl(tmp1, -1);
7055 jcc(Assembler::zero, FALSE_LABEL);
7056
7057 // ~(~0 << len) applied up to two times (for 32-bit scenario)
7058 #ifdef _LP64
7059 mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF);
7060 shlxq(tmp3_aliased, tmp3_aliased, tmp1);
7061 notq(tmp3_aliased);
7062 kmovql(k3, tmp3_aliased);
7063 #else
7064 Label k_init;
7065 jmp(k_init);
7066
7067 // We could not read 64-bits from a general purpose register thus we move
7068 // data required to compose 64 1's to the instruction stream
7069 // We emit 64 byte wide series of elements from 0..63 which later on would
7070 // be used as a compare targets with tail count contained in tmp1 register.
7071 // Result would be a k register having tmp1 consecutive number or 1
7072 // counting from least significant bit.
7073 address tmp = pc();
7074 emit_int64(0x0706050403020100);
7075 emit_int64(0x0F0E0D0C0B0A0908);
7076 emit_int64(0x1716151413121110);
7077 emit_int64(0x1F1E1D1C1B1A1918);
7078 emit_int64(0x2726252423222120);
7079 emit_int64(0x2F2E2D2C2B2A2928);
7080 emit_int64(0x3736353433323130);
7081 emit_int64(0x3F3E3D3C3B3A3938);
7082
7083 bind(k_init);
7084 lea(len, InternalAddress(tmp));
7085 // create mask to test for negative byte inside a vector
7086 evpbroadcastb(vec1, tmp1, Assembler::AVX_512bit);
7087 evpcmpgtb(k3, vec1, Address(len, 0), Assembler::AVX_512bit);
7088
7089 #endif
7090 evpcmpgtb(k2, k3, vec2, Address(ary1, 0), Assembler::AVX_512bit);
7091 ktestq(k2, k3);
7092 jcc(Assembler::notZero, TRUE_LABEL);
7093
7094 jmp(FALSE_LABEL);
7095 } else {
7096 movl(result, len); // copy
7097
7098 if (UseAVX == 2 && UseSSE >= 2) {
7099 // With AVX2, use 32-byte vector compare
7100 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7101
7102 // Compare 32-byte vectors
7103 andl(result, 0x0000001f); // tail count (in bytes)
7104 andl(len, 0xffffffe0); // vector count (in bytes)
7105 jccb(Assembler::zero, COMPARE_TAIL);
7106
7107 lea(ary1, Address(ary1, len, Address::times_1));
7108 negptr(len);
7109
7110 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector
7111 movdl(vec2, tmp1);
7112 vpbroadcastd(vec2, vec2, Assembler::AVX_256bit);
7113
7114 bind(COMPARE_WIDE_VECTORS);
7115 vmovdqu(vec1, Address(ary1, len, Address::times_1));
7116 vptest(vec1, vec2);
7117 jccb(Assembler::notZero, TRUE_LABEL);
7118 addptr(len, 32);
7119 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7120
7121 testl(result, result);
7122 jccb(Assembler::zero, FALSE_LABEL);
7123
7124 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
7125 vptest(vec1, vec2);
7126 jccb(Assembler::notZero, TRUE_LABEL);
7127 jmpb(FALSE_LABEL);
7128
7129 bind(COMPARE_TAIL); // len is zero
7130 movl(len, result);
7131 // Fallthru to tail compare
7132 } else if (UseSSE42Intrinsics) {
7133 // With SSE4.2, use double quad vector compare
7134 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7135
7136 // Compare 16-byte vectors
7137 andl(result, 0x0000000f); // tail count (in bytes)
7138 andl(len, 0xfffffff0); // vector count (in bytes)
7139 jcc(Assembler::zero, COMPARE_TAIL);
7140
7141 lea(ary1, Address(ary1, len, Address::times_1));
7142 negptr(len);
7143
7144 movl(tmp1, 0x80808080);
7145 movdl(vec2, tmp1);
7146 pshufd(vec2, vec2, 0);
7147
7148 bind(COMPARE_WIDE_VECTORS);
7149 movdqu(vec1, Address(ary1, len, Address::times_1));
7150 ptest(vec1, vec2);
7151 jcc(Assembler::notZero, TRUE_LABEL);
7152 addptr(len, 16);
7153 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7154
7155 testl(result, result);
7156 jcc(Assembler::zero, FALSE_LABEL);
7157
7158 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
7159 ptest(vec1, vec2);
7160 jccb(Assembler::notZero, TRUE_LABEL);
7161 jmpb(FALSE_LABEL);
7162
7163 bind(COMPARE_TAIL); // len is zero
7164 movl(len, result);
7165 // Fallthru to tail compare
7166 }
7167 }
7168 // Compare 4-byte vectors
7169 andl(len, 0xfffffffc); // vector count (in bytes)
7170 jccb(Assembler::zero, COMPARE_CHAR);
7171
7172 lea(ary1, Address(ary1, len, Address::times_1));
7173 negptr(len);
7174
7175 bind(COMPARE_VECTORS);
7176 movl(tmp1, Address(ary1, len, Address::times_1));
7177 andl(tmp1, 0x80808080);
7178 jccb(Assembler::notZero, TRUE_LABEL);
7179 addptr(len, 4);
7180 jcc(Assembler::notZero, COMPARE_VECTORS);
7181
7182 // Compare trailing char (final 2 bytes), if any
7183 bind(COMPARE_CHAR);
7184 testl(result, 0x2); // tail char
7185 jccb(Assembler::zero, COMPARE_BYTE);
7186 load_unsigned_short(tmp1, Address(ary1, 0));
7187 andl(tmp1, 0x00008080);
7188 jccb(Assembler::notZero, TRUE_LABEL);
7189 subptr(result, 2);
7190 lea(ary1, Address(ary1, 2));
7191
7192 bind(COMPARE_BYTE);
7193 testl(result, 0x1); // tail byte
7194 jccb(Assembler::zero, FALSE_LABEL);
7195 load_unsigned_byte(tmp1, Address(ary1, 0));
7196 andl(tmp1, 0x00000080);
7197 jccb(Assembler::notEqual, TRUE_LABEL);
7198 jmpb(FALSE_LABEL);
7199
7200 bind(TRUE_LABEL);
7201 movl(result, 1); // return true
7202 jmpb(DONE);
7203
7204 bind(FALSE_LABEL);
7205 xorl(result, result); // return false
7206
7207 // That's it
7208 bind(DONE);
7209 if (UseAVX >= 2 && UseSSE >= 2) {
7210 // clean upper bits of YMM registers
7211 vpxor(vec1, vec1);
7212 vpxor(vec2, vec2);
7213 }
7214 }
7215 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings.
7216 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2,
7217 Register limit, Register result, Register chr,
7218 XMMRegister vec1, XMMRegister vec2, bool is_char) {
7219 ShortBranchVerifier sbv(this);
7220 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE;
7221
7222 int length_offset = arrayOopDesc::length_offset_in_bytes();
7223 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE);
7224
7225 if (is_array_equ) {
7226 // Check the input args
7227 cmpoop(ary1, ary2);
7228 jcc(Assembler::equal, TRUE_LABEL);
7229
7230 // Need additional checks for arrays_equals.
7231 testptr(ary1, ary1);
7232 jcc(Assembler::zero, FALSE_LABEL);
7233 testptr(ary2, ary2);
7234 jcc(Assembler::zero, FALSE_LABEL);
7235
7236 // Check the lengths
7237 movl(limit, Address(ary1, length_offset));
7238 cmpl(limit, Address(ary2, length_offset));
7239 jcc(Assembler::notEqual, FALSE_LABEL);
7240 }
7241
7242 // count == 0
7243 testl(limit, limit);
7244 jcc(Assembler::zero, TRUE_LABEL);
7245
7246 if (is_array_equ) {
7247 // Load array address
7248 lea(ary1, Address(ary1, base_offset));
7249 lea(ary2, Address(ary2, base_offset));
7250 }
7251
7252 if (is_array_equ && is_char) {
7253 // arrays_equals when used for char[].
7254 shll(limit, 1); // byte count != 0
7255 }
7256 movl(result, limit); // copy
7257
7258 if (UseAVX >= 2) {
7259 // With AVX2, use 32-byte vector compare
7260 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7261
7262 // Compare 32-byte vectors
7263 andl(result, 0x0000001f); // tail count (in bytes)
7264 andl(limit, 0xffffffe0); // vector count (in bytes)
7265 jcc(Assembler::zero, COMPARE_TAIL);
7266
7267 lea(ary1, Address(ary1, limit, Address::times_1));
7268 lea(ary2, Address(ary2, limit, Address::times_1));
7269 negptr(limit);
7270
7271 bind(COMPARE_WIDE_VECTORS);
7272
7273 #ifdef _LP64
7274 if (VM_Version::supports_avx512vlbw()) { // trying 64 bytes fast loop
7275 Label COMPARE_WIDE_VECTORS_LOOP_AVX2, COMPARE_WIDE_VECTORS_LOOP_AVX3;
7276
7277 cmpl(limit, -64);
7278 jccb(Assembler::greater, COMPARE_WIDE_VECTORS_LOOP_AVX2);
7279
7280 bind(COMPARE_WIDE_VECTORS_LOOP_AVX3); // the hottest loop
7281
7282 evmovdquq(vec1, Address(ary1, limit, Address::times_1), Assembler::AVX_512bit);
7283 evpcmpeqb(k7, vec1, Address(ary2, limit, Address::times_1), Assembler::AVX_512bit);
7284 kortestql(k7, k7);
7285 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
7286 addptr(limit, 64); // update since we already compared at this addr
7287 cmpl(limit, -64);
7288 jccb(Assembler::lessEqual, COMPARE_WIDE_VECTORS_LOOP_AVX3);
7289
7290 // At this point we may still need to compare -limit+result bytes.
7291 // We could execute the next two instruction and just continue via non-wide path:
7292 // cmpl(limit, 0);
7293 // jcc(Assembler::equal, COMPARE_TAIL); // true
7294 // But since we stopped at the points ary{1,2}+limit which are
7295 // not farther than 64 bytes from the ends of arrays ary{1,2}+result
7296 // (|limit| <= 32 and result < 32),
7297 // we may just compare the last 64 bytes.
7298 //
7299 addptr(result, -64); // it is safe, bc we just came from this area
7300 evmovdquq(vec1, Address(ary1, result, Address::times_1), Assembler::AVX_512bit);
7301 evpcmpeqb(k7, vec1, Address(ary2, result, Address::times_1), Assembler::AVX_512bit);
7302 kortestql(k7, k7);
7303 jcc(Assembler::aboveEqual, FALSE_LABEL); // miscompare
7304
7305 jmp(TRUE_LABEL);
7306
7307 bind(COMPARE_WIDE_VECTORS_LOOP_AVX2);
7308
7309 }//if (VM_Version::supports_avx512vlbw())
7310 #endif //_LP64
7311
7312 vmovdqu(vec1, Address(ary1, limit, Address::times_1));
7313 vmovdqu(vec2, Address(ary2, limit, Address::times_1));
7314 vpxor(vec1, vec2);
7315
7316 vptest(vec1, vec1);
7317 jcc(Assembler::notZero, FALSE_LABEL);
7318 addptr(limit, 32);
7319 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7320
7321 testl(result, result);
7322 jcc(Assembler::zero, TRUE_LABEL);
7323
7324 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
7325 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
7326 vpxor(vec1, vec2);
7327
7328 vptest(vec1, vec1);
7329 jccb(Assembler::notZero, FALSE_LABEL);
7330 jmpb(TRUE_LABEL);
7331
7332 bind(COMPARE_TAIL); // limit is zero
7333 movl(limit, result);
7334 // Fallthru to tail compare
7335 } else if (UseSSE42Intrinsics) {
7336 // With SSE4.2, use double quad vector compare
7337 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
7338
7339 // Compare 16-byte vectors
7340 andl(result, 0x0000000f); // tail count (in bytes)
7341 andl(limit, 0xfffffff0); // vector count (in bytes)
7342 jcc(Assembler::zero, COMPARE_TAIL);
7343
7344 lea(ary1, Address(ary1, limit, Address::times_1));
7345 lea(ary2, Address(ary2, limit, Address::times_1));
7346 negptr(limit);
7347
7348 bind(COMPARE_WIDE_VECTORS);
7349 movdqu(vec1, Address(ary1, limit, Address::times_1));
7350 movdqu(vec2, Address(ary2, limit, Address::times_1));
7351 pxor(vec1, vec2);
7352
7353 ptest(vec1, vec1);
7354 jcc(Assembler::notZero, FALSE_LABEL);
7355 addptr(limit, 16);
7356 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
7357
7358 testl(result, result);
7359 jcc(Assembler::zero, TRUE_LABEL);
7360
7361 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
7362 movdqu(vec2, Address(ary2, result, Address::times_1, -16));
7363 pxor(vec1, vec2);
7364
7365 ptest(vec1, vec1);
7366 jccb(Assembler::notZero, FALSE_LABEL);
7367 jmpb(TRUE_LABEL);
7368
7369 bind(COMPARE_TAIL); // limit is zero
7370 movl(limit, result);
7371 // Fallthru to tail compare
7372 }
7373
7374 // Compare 4-byte vectors
7375 andl(limit, 0xfffffffc); // vector count (in bytes)
7376 jccb(Assembler::zero, COMPARE_CHAR);
7377
7378 lea(ary1, Address(ary1, limit, Address::times_1));
7379 lea(ary2, Address(ary2, limit, Address::times_1));
7380 negptr(limit);
7381
7382 bind(COMPARE_VECTORS);
7383 movl(chr, Address(ary1, limit, Address::times_1));
7384 cmpl(chr, Address(ary2, limit, Address::times_1));
7385 jccb(Assembler::notEqual, FALSE_LABEL);
7386 addptr(limit, 4);
7387 jcc(Assembler::notZero, COMPARE_VECTORS);
7388
7389 // Compare trailing char (final 2 bytes), if any
7390 bind(COMPARE_CHAR);
7391 testl(result, 0x2); // tail char
7392 jccb(Assembler::zero, COMPARE_BYTE);
7393 load_unsigned_short(chr, Address(ary1, 0));
7394 load_unsigned_short(limit, Address(ary2, 0));
7395 cmpl(chr, limit);
7396 jccb(Assembler::notEqual, FALSE_LABEL);
7397
7398 if (is_array_equ && is_char) {
7399 bind(COMPARE_BYTE);
7400 } else {
7401 lea(ary1, Address(ary1, 2));
7402 lea(ary2, Address(ary2, 2));
7403
7404 bind(COMPARE_BYTE);
7405 testl(result, 0x1); // tail byte
7406 jccb(Assembler::zero, TRUE_LABEL);
7407 load_unsigned_byte(chr, Address(ary1, 0));
7408 load_unsigned_byte(limit, Address(ary2, 0));
7409 cmpl(chr, limit);
7410 jccb(Assembler::notEqual, FALSE_LABEL);
7411 }
7412 bind(TRUE_LABEL);
7413 movl(result, 1); // return true
7414 jmpb(DONE);
7415
7416 bind(FALSE_LABEL);
7417 xorl(result, result); // return false
7418
7419 // That's it
7420 bind(DONE);
7421 if (UseAVX >= 2) {
7422 // clean upper bits of YMM registers
7423 vpxor(vec1, vec1);
7424 vpxor(vec2, vec2);
7425 }
7426 }
7427
7428 #endif
7429
7430 void MacroAssembler::generate_fill(BasicType t, bool aligned,
7431 Register to, Register value, Register count,
7432 Register rtmp, XMMRegister xtmp) {
7433 ShortBranchVerifier sbv(this);
7434 assert_different_registers(to, value, count, rtmp);
7435 Label L_exit;
7436 Label L_fill_2_bytes, L_fill_4_bytes;
7437
7438 int shift = -1;
7439 switch (t) {
7440 case T_BYTE:
7441 shift = 2;
7442 break;
7443 case T_SHORT:
7444 shift = 1;
7445 break;
7446 case T_INT:
7447 shift = 0;
7448 break;
7449 default: ShouldNotReachHere();
7450 }
7451
7452 if (t == T_BYTE) {
7453 andl(value, 0xff);
7454 movl(rtmp, value);
7455 shll(rtmp, 8);
7456 orl(value, rtmp);
7457 }
7458 if (t == T_SHORT) {
7459 andl(value, 0xffff);
7460 }
7461 if (t == T_BYTE || t == T_SHORT) {
7462 movl(rtmp, value);
7463 shll(rtmp, 16);
7464 orl(value, rtmp);
7465 }
7466
7467 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
7468 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp
7469 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) {
7470 Label L_skip_align2;
7471 // align source address at 4 bytes address boundary
7472 if (t == T_BYTE) {
7473 Label L_skip_align1;
7474 // One byte misalignment happens only for byte arrays
7475 testptr(to, 1);
7476 jccb(Assembler::zero, L_skip_align1);
7477 movb(Address(to, 0), value);
7478 increment(to);
7479 decrement(count);
7480 BIND(L_skip_align1);
7481 }
7482 // Two bytes misalignment happens only for byte and short (char) arrays
7483 testptr(to, 2);
7484 jccb(Assembler::zero, L_skip_align2);
7485 movw(Address(to, 0), value);
7486 addptr(to, 2);
7487 subl(count, 1<<(shift-1));
7488 BIND(L_skip_align2);
7489 }
7490 if (UseSSE < 2) {
7491 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7492 // Fill 32-byte chunks
7493 subl(count, 8 << shift);
7494 jcc(Assembler::less, L_check_fill_8_bytes);
7495 align(16);
7496
7497 BIND(L_fill_32_bytes_loop);
7498
7499 for (int i = 0; i < 32; i += 4) {
7500 movl(Address(to, i), value);
7501 }
7502
7503 addptr(to, 32);
7504 subl(count, 8 << shift);
7505 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7506 BIND(L_check_fill_8_bytes);
7507 addl(count, 8 << shift);
7508 jccb(Assembler::zero, L_exit);
7509 jmpb(L_fill_8_bytes);
7510
7511 //
7512 // length is too short, just fill qwords
7513 //
7514 BIND(L_fill_8_bytes_loop);
7515 movl(Address(to, 0), value);
7516 movl(Address(to, 4), value);
7517 addptr(to, 8);
7518 BIND(L_fill_8_bytes);
7519 subl(count, 1 << (shift + 1));
7520 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7521 // fall through to fill 4 bytes
7522 } else {
7523 Label L_fill_32_bytes;
7524 if (!UseUnalignedLoadStores) {
7525 // align to 8 bytes, we know we are 4 byte aligned to start
7526 testptr(to, 4);
7527 jccb(Assembler::zero, L_fill_32_bytes);
7528 movl(Address(to, 0), value);
7529 addptr(to, 4);
7530 subl(count, 1<<shift);
7531 }
7532 BIND(L_fill_32_bytes);
7533 {
7534 assert( UseSSE >= 2, "supported cpu only" );
7535 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
7536 movdl(xtmp, value);
7537 if (UseAVX > 2 && UseUnalignedLoadStores) {
7538 // Fill 64-byte chunks
7539 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
7540 vpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit);
7541
7542 subl(count, 16 << shift);
7543 jcc(Assembler::less, L_check_fill_32_bytes);
7544 align(16);
7545
7546 BIND(L_fill_64_bytes_loop);
7547 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit);
7548 addptr(to, 64);
7549 subl(count, 16 << shift);
7550 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
7551
7552 BIND(L_check_fill_32_bytes);
7553 addl(count, 8 << shift);
7554 jccb(Assembler::less, L_check_fill_8_bytes);
7555 vmovdqu(Address(to, 0), xtmp);
7556 addptr(to, 32);
7557 subl(count, 8 << shift);
7558
7559 BIND(L_check_fill_8_bytes);
7560 } else if (UseAVX == 2 && UseUnalignedLoadStores) {
7561 // Fill 64-byte chunks
7562 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
7563 vpbroadcastd(xtmp, xtmp, Assembler::AVX_256bit);
7564
7565 subl(count, 16 << shift);
7566 jcc(Assembler::less, L_check_fill_32_bytes);
7567 align(16);
7568
7569 BIND(L_fill_64_bytes_loop);
7570 vmovdqu(Address(to, 0), xtmp);
7571 vmovdqu(Address(to, 32), xtmp);
7572 addptr(to, 64);
7573 subl(count, 16 << shift);
7574 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
7575
7576 BIND(L_check_fill_32_bytes);
7577 addl(count, 8 << shift);
7578 jccb(Assembler::less, L_check_fill_8_bytes);
7579 vmovdqu(Address(to, 0), xtmp);
7580 addptr(to, 32);
7581 subl(count, 8 << shift);
7582
7583 BIND(L_check_fill_8_bytes);
7584 // clean upper bits of YMM registers
7585 movdl(xtmp, value);
7586 pshufd(xtmp, xtmp, 0);
7587 } else {
7588 // Fill 32-byte chunks
7589 pshufd(xtmp, xtmp, 0);
7590
7591 subl(count, 8 << shift);
7592 jcc(Assembler::less, L_check_fill_8_bytes);
7593 align(16);
7594
7595 BIND(L_fill_32_bytes_loop);
7596
7597 if (UseUnalignedLoadStores) {
7598 movdqu(Address(to, 0), xtmp);
7599 movdqu(Address(to, 16), xtmp);
7600 } else {
7601 movq(Address(to, 0), xtmp);
7602 movq(Address(to, 8), xtmp);
7603 movq(Address(to, 16), xtmp);
7604 movq(Address(to, 24), xtmp);
7605 }
7606
7607 addptr(to, 32);
7608 subl(count, 8 << shift);
7609 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
7610
7611 BIND(L_check_fill_8_bytes);
7612 }
7613 addl(count, 8 << shift);
7614 jccb(Assembler::zero, L_exit);
7615 jmpb(L_fill_8_bytes);
7616
7617 //
7618 // length is too short, just fill qwords
7619 //
7620 BIND(L_fill_8_bytes_loop);
7621 movq(Address(to, 0), xtmp);
7622 addptr(to, 8);
7623 BIND(L_fill_8_bytes);
7624 subl(count, 1 << (shift + 1));
7625 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
7626 }
7627 }
7628 // fill trailing 4 bytes
7629 BIND(L_fill_4_bytes);
7630 testl(count, 1<<shift);
7631 jccb(Assembler::zero, L_fill_2_bytes);
7632 movl(Address(to, 0), value);
7633 if (t == T_BYTE || t == T_SHORT) {
7634 Label L_fill_byte;
7635 addptr(to, 4);
7636 BIND(L_fill_2_bytes);
7637 // fill trailing 2 bytes
7638 testl(count, 1<<(shift-1));
7639 jccb(Assembler::zero, L_fill_byte);
7640 movw(Address(to, 0), value);
7641 if (t == T_BYTE) {
7642 addptr(to, 2);
7643 BIND(L_fill_byte);
7644 // fill trailing byte
7645 testl(count, 1);
7646 jccb(Assembler::zero, L_exit);
7647 movb(Address(to, 0), value);
7648 } else {
7649 BIND(L_fill_byte);
7650 }
7651 } else {
7652 BIND(L_fill_2_bytes);
7653 }
7654 BIND(L_exit);
7655 }
7656
7657 // encode char[] to byte[] in ISO_8859_1
7658 //@HotSpotIntrinsicCandidate
7659 //private static int implEncodeISOArray(byte[] sa, int sp,
7660 //byte[] da, int dp, int len) {
7661 // int i = 0;
7662 // for (; i < len; i++) {
7663 // char c = StringUTF16.getChar(sa, sp++);
7664 // if (c > '\u00FF')
7665 // break;
7666 // da[dp++] = (byte)c;
7667 // }
7668 // return i;
7669 //}
7670 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len,
7671 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
7672 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
7673 Register tmp5, Register result) {
7674
7675 // rsi: src
7676 // rdi: dst
7677 // rdx: len
7678 // rcx: tmp5
7679 // rax: result
7680 ShortBranchVerifier sbv(this);
7681 assert_different_registers(src, dst, len, tmp5, result);
7682 Label L_done, L_copy_1_char, L_copy_1_char_exit;
7683
7684 // set result
7685 xorl(result, result);
7686 // check for zero length
7687 testl(len, len);
7688 jcc(Assembler::zero, L_done);
7689
7690 movl(result, len);
7691
7692 // Setup pointers
7693 lea(src, Address(src, len, Address::times_2)); // char[]
7694 lea(dst, Address(dst, len, Address::times_1)); // byte[]
7695 negptr(len);
7696
7697 if (UseSSE42Intrinsics || UseAVX >= 2) {
7698 Label L_copy_8_chars, L_copy_8_chars_exit;
7699 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit;
7700
7701 if (UseAVX >= 2) {
7702 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit;
7703 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
7704 movdl(tmp1Reg, tmp5);
7705 vpbroadcastd(tmp1Reg, tmp1Reg, Assembler::AVX_256bit);
7706 jmp(L_chars_32_check);
7707
7708 bind(L_copy_32_chars);
7709 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64));
7710 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32));
7711 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
7712 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
7713 jccb(Assembler::notZero, L_copy_32_chars_exit);
7714 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
7715 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1);
7716 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg);
7717
7718 bind(L_chars_32_check);
7719 addptr(len, 32);
7720 jcc(Assembler::lessEqual, L_copy_32_chars);
7721
7722 bind(L_copy_32_chars_exit);
7723 subptr(len, 16);
7724 jccb(Assembler::greater, L_copy_16_chars_exit);
7725
7726 } else if (UseSSE42Intrinsics) {
7727 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
7728 movdl(tmp1Reg, tmp5);
7729 pshufd(tmp1Reg, tmp1Reg, 0);
7730 jmpb(L_chars_16_check);
7731 }
7732
7733 bind(L_copy_16_chars);
7734 if (UseAVX >= 2) {
7735 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32));
7736 vptest(tmp2Reg, tmp1Reg);
7737 jcc(Assembler::notZero, L_copy_16_chars_exit);
7738 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1);
7739 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1);
7740 } else {
7741 if (UseAVX > 0) {
7742 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7743 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7744 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0);
7745 } else {
7746 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
7747 por(tmp2Reg, tmp3Reg);
7748 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
7749 por(tmp2Reg, tmp4Reg);
7750 }
7751 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
7752 jccb(Assembler::notZero, L_copy_16_chars_exit);
7753 packuswb(tmp3Reg, tmp4Reg);
7754 }
7755 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg);
7756
7757 bind(L_chars_16_check);
7758 addptr(len, 16);
7759 jcc(Assembler::lessEqual, L_copy_16_chars);
7760
7761 bind(L_copy_16_chars_exit);
7762 if (UseAVX >= 2) {
7763 // clean upper bits of YMM registers
7764 vpxor(tmp2Reg, tmp2Reg);
7765 vpxor(tmp3Reg, tmp3Reg);
7766 vpxor(tmp4Reg, tmp4Reg);
7767 movdl(tmp1Reg, tmp5);
7768 pshufd(tmp1Reg, tmp1Reg, 0);
7769 }
7770 subptr(len, 8);
7771 jccb(Assembler::greater, L_copy_8_chars_exit);
7772
7773 bind(L_copy_8_chars);
7774 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16));
7775 ptest(tmp3Reg, tmp1Reg);
7776 jccb(Assembler::notZero, L_copy_8_chars_exit);
7777 packuswb(tmp3Reg, tmp1Reg);
7778 movq(Address(dst, len, Address::times_1, -8), tmp3Reg);
7779 addptr(len, 8);
7780 jccb(Assembler::lessEqual, L_copy_8_chars);
7781
7782 bind(L_copy_8_chars_exit);
7783 subptr(len, 8);
7784 jccb(Assembler::zero, L_done);
7785 }
7786
7787 bind(L_copy_1_char);
7788 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0));
7789 testl(tmp5, 0xff00); // check if Unicode char
7790 jccb(Assembler::notZero, L_copy_1_char_exit);
7791 movb(Address(dst, len, Address::times_1, 0), tmp5);
7792 addptr(len, 1);
7793 jccb(Assembler::less, L_copy_1_char);
7794
7795 bind(L_copy_1_char_exit);
7796 addptr(result, len); // len is negative count of not processed elements
7797
7798 bind(L_done);
7799 }
7800
7801 #ifdef _LP64
7802 /**
7803 * Helper for multiply_to_len().
7804 */
7805 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) {
7806 addq(dest_lo, src1);
7807 adcq(dest_hi, 0);
7808 addq(dest_lo, src2);
7809 adcq(dest_hi, 0);
7810 }
7811
7812 /**
7813 * Multiply 64 bit by 64 bit first loop.
7814 */
7815 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
7816 Register y, Register y_idx, Register z,
7817 Register carry, Register product,
7818 Register idx, Register kdx) {
7819 //
7820 // jlong carry, x[], y[], z[];
7821 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
7822 // huge_128 product = y[idx] * x[xstart] + carry;
7823 // z[kdx] = (jlong)product;
7824 // carry = (jlong)(product >>> 64);
7825 // }
7826 // z[xstart] = carry;
7827 //
7828
7829 Label L_first_loop, L_first_loop_exit;
7830 Label L_one_x, L_one_y, L_multiply;
7831
7832 decrementl(xstart);
7833 jcc(Assembler::negative, L_one_x);
7834
7835 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
7836 rorq(x_xstart, 32); // convert big-endian to little-endian
7837
7838 bind(L_first_loop);
7839 decrementl(idx);
7840 jcc(Assembler::negative, L_first_loop_exit);
7841 decrementl(idx);
7842 jcc(Assembler::negative, L_one_y);
7843 movq(y_idx, Address(y, idx, Address::times_4, 0));
7844 rorq(y_idx, 32); // convert big-endian to little-endian
7845 bind(L_multiply);
7846 movq(product, x_xstart);
7847 mulq(y_idx); // product(rax) * y_idx -> rdx:rax
7848 addq(product, carry);
7849 adcq(rdx, 0);
7850 subl(kdx, 2);
7851 movl(Address(z, kdx, Address::times_4, 4), product);
7852 shrq(product, 32);
7853 movl(Address(z, kdx, Address::times_4, 0), product);
7854 movq(carry, rdx);
7855 jmp(L_first_loop);
7856
7857 bind(L_one_y);
7858 movl(y_idx, Address(y, 0));
7859 jmp(L_multiply);
7860
7861 bind(L_one_x);
7862 movl(x_xstart, Address(x, 0));
7863 jmp(L_first_loop);
7864
7865 bind(L_first_loop_exit);
7866 }
7867
7868 /**
7869 * Multiply 64 bit by 64 bit and add 128 bit.
7870 */
7871 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z,
7872 Register yz_idx, Register idx,
7873 Register carry, Register product, int offset) {
7874 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
7875 // z[kdx] = (jlong)product;
7876
7877 movq(yz_idx, Address(y, idx, Address::times_4, offset));
7878 rorq(yz_idx, 32); // convert big-endian to little-endian
7879 movq(product, x_xstart);
7880 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7881 movq(yz_idx, Address(z, idx, Address::times_4, offset));
7882 rorq(yz_idx, 32); // convert big-endian to little-endian
7883
7884 add2_with_carry(rdx, product, carry, yz_idx);
7885
7886 movl(Address(z, idx, Address::times_4, offset+4), product);
7887 shrq(product, 32);
7888 movl(Address(z, idx, Address::times_4, offset), product);
7889
7890 }
7891
7892 /**
7893 * Multiply 128 bit by 128 bit. Unrolled inner loop.
7894 */
7895 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z,
7896 Register yz_idx, Register idx, Register jdx,
7897 Register carry, Register product,
7898 Register carry2) {
7899 // jlong carry, x[], y[], z[];
7900 // int kdx = ystart+1;
7901 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7902 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
7903 // z[kdx+idx+1] = (jlong)product;
7904 // jlong carry2 = (jlong)(product >>> 64);
7905 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
7906 // z[kdx+idx] = (jlong)product;
7907 // carry = (jlong)(product >>> 64);
7908 // }
7909 // idx += 2;
7910 // if (idx > 0) {
7911 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
7912 // z[kdx+idx] = (jlong)product;
7913 // carry = (jlong)(product >>> 64);
7914 // }
7915 //
7916
7917 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
7918
7919 movl(jdx, idx);
7920 andl(jdx, 0xFFFFFFFC);
7921 shrl(jdx, 2);
7922
7923 bind(L_third_loop);
7924 subl(jdx, 1);
7925 jcc(Assembler::negative, L_third_loop_exit);
7926 subl(idx, 4);
7927
7928 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
7929 movq(carry2, rdx);
7930
7931 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
7932 movq(carry, rdx);
7933 jmp(L_third_loop);
7934
7935 bind (L_third_loop_exit);
7936
7937 andl (idx, 0x3);
7938 jcc(Assembler::zero, L_post_third_loop_done);
7939
7940 Label L_check_1;
7941 subl(idx, 2);
7942 jcc(Assembler::negative, L_check_1);
7943
7944 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
7945 movq(carry, rdx);
7946
7947 bind (L_check_1);
7948 addl (idx, 0x2);
7949 andl (idx, 0x1);
7950 subl(idx, 1);
7951 jcc(Assembler::negative, L_post_third_loop_done);
7952
7953 movl(yz_idx, Address(y, idx, Address::times_4, 0));
7954 movq(product, x_xstart);
7955 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
7956 movl(yz_idx, Address(z, idx, Address::times_4, 0));
7957
7958 add2_with_carry(rdx, product, yz_idx, carry);
7959
7960 movl(Address(z, idx, Address::times_4, 0), product);
7961 shrq(product, 32);
7962
7963 shlq(rdx, 32);
7964 orq(product, rdx);
7965 movq(carry, product);
7966
7967 bind(L_post_third_loop_done);
7968 }
7969
7970 /**
7971 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop.
7972 *
7973 */
7974 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z,
7975 Register carry, Register carry2,
7976 Register idx, Register jdx,
7977 Register yz_idx1, Register yz_idx2,
7978 Register tmp, Register tmp3, Register tmp4) {
7979 assert(UseBMI2Instructions, "should be used only when BMI2 is available");
7980
7981 // jlong carry, x[], y[], z[];
7982 // int kdx = ystart+1;
7983 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
7984 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry;
7985 // jlong carry2 = (jlong)(tmp3 >>> 64);
7986 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2;
7987 // carry = (jlong)(tmp4 >>> 64);
7988 // z[kdx+idx+1] = (jlong)tmp3;
7989 // z[kdx+idx] = (jlong)tmp4;
7990 // }
7991 // idx += 2;
7992 // if (idx > 0) {
7993 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry;
7994 // z[kdx+idx] = (jlong)yz_idx1;
7995 // carry = (jlong)(yz_idx1 >>> 64);
7996 // }
7997 //
7998
7999 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
8000
8001 movl(jdx, idx);
8002 andl(jdx, 0xFFFFFFFC);
8003 shrl(jdx, 2);
8004
8005 bind(L_third_loop);
8006 subl(jdx, 1);
8007 jcc(Assembler::negative, L_third_loop_exit);
8008 subl(idx, 4);
8009
8010 movq(yz_idx1, Address(y, idx, Address::times_4, 8));
8011 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
8012 movq(yz_idx2, Address(y, idx, Address::times_4, 0));
8013 rorxq(yz_idx2, yz_idx2, 32);
8014
8015 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
8016 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp
8017
8018 movq(yz_idx1, Address(z, idx, Address::times_4, 8));
8019 rorxq(yz_idx1, yz_idx1, 32);
8020 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
8021 rorxq(yz_idx2, yz_idx2, 32);
8022
8023 if (VM_Version::supports_adx()) {
8024 adcxq(tmp3, carry);
8025 adoxq(tmp3, yz_idx1);
8026
8027 adcxq(tmp4, tmp);
8028 adoxq(tmp4, yz_idx2);
8029
8030 movl(carry, 0); // does not affect flags
8031 adcxq(carry2, carry);
8032 adoxq(carry2, carry);
8033 } else {
8034 add2_with_carry(tmp4, tmp3, carry, yz_idx1);
8035 add2_with_carry(carry2, tmp4, tmp, yz_idx2);
8036 }
8037 movq(carry, carry2);
8038
8039 movl(Address(z, idx, Address::times_4, 12), tmp3);
8040 shrq(tmp3, 32);
8041 movl(Address(z, idx, Address::times_4, 8), tmp3);
8042
8043 movl(Address(z, idx, Address::times_4, 4), tmp4);
8044 shrq(tmp4, 32);
8045 movl(Address(z, idx, Address::times_4, 0), tmp4);
8046
8047 jmp(L_third_loop);
8048
8049 bind (L_third_loop_exit);
8050
8051 andl (idx, 0x3);
8052 jcc(Assembler::zero, L_post_third_loop_done);
8053
8054 Label L_check_1;
8055 subl(idx, 2);
8056 jcc(Assembler::negative, L_check_1);
8057
8058 movq(yz_idx1, Address(y, idx, Address::times_4, 0));
8059 rorxq(yz_idx1, yz_idx1, 32);
8060 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
8061 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
8062 rorxq(yz_idx2, yz_idx2, 32);
8063
8064 add2_with_carry(tmp4, tmp3, carry, yz_idx2);
8065
8066 movl(Address(z, idx, Address::times_4, 4), tmp3);
8067 shrq(tmp3, 32);
8068 movl(Address(z, idx, Address::times_4, 0), tmp3);
8069 movq(carry, tmp4);
8070
8071 bind (L_check_1);
8072 addl (idx, 0x2);
8073 andl (idx, 0x1);
8074 subl(idx, 1);
8075 jcc(Assembler::negative, L_post_third_loop_done);
8076 movl(tmp4, Address(y, idx, Address::times_4, 0));
8077 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3
8078 movl(tmp4, Address(z, idx, Address::times_4, 0));
8079
8080 add2_with_carry(carry2, tmp3, tmp4, carry);
8081
8082 movl(Address(z, idx, Address::times_4, 0), tmp3);
8083 shrq(tmp3, 32);
8084
8085 shlq(carry2, 32);
8086 orq(tmp3, carry2);
8087 movq(carry, tmp3);
8088
8089 bind(L_post_third_loop_done);
8090 }
8091
8092 /**
8093 * Code for BigInteger::multiplyToLen() instrinsic.
8094 *
8095 * rdi: x
8096 * rax: xlen
8097 * rsi: y
8098 * rcx: ylen
8099 * r8: z
8100 * r11: zlen
8101 * r12: tmp1
8102 * r13: tmp2
8103 * r14: tmp3
8104 * r15: tmp4
8105 * rbx: tmp5
8106 *
8107 */
8108 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen,
8109 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) {
8110 ShortBranchVerifier sbv(this);
8111 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx);
8112
8113 push(tmp1);
8114 push(tmp2);
8115 push(tmp3);
8116 push(tmp4);
8117 push(tmp5);
8118
8119 push(xlen);
8120 push(zlen);
8121
8122 const Register idx = tmp1;
8123 const Register kdx = tmp2;
8124 const Register xstart = tmp3;
8125
8126 const Register y_idx = tmp4;
8127 const Register carry = tmp5;
8128 const Register product = xlen;
8129 const Register x_xstart = zlen; // reuse register
8130
8131 // First Loop.
8132 //
8133 // final static long LONG_MASK = 0xffffffffL;
8134 // int xstart = xlen - 1;
8135 // int ystart = ylen - 1;
8136 // long carry = 0;
8137 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
8138 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
8139 // z[kdx] = (int)product;
8140 // carry = product >>> 32;
8141 // }
8142 // z[xstart] = (int)carry;
8143 //
8144
8145 movl(idx, ylen); // idx = ylen;
8146 movl(kdx, zlen); // kdx = xlen+ylen;
8147 xorq(carry, carry); // carry = 0;
8148
8149 Label L_done;
8150
8151 movl(xstart, xlen);
8152 decrementl(xstart);
8153 jcc(Assembler::negative, L_done);
8154
8155 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
8156
8157 Label L_second_loop;
8158 testl(kdx, kdx);
8159 jcc(Assembler::zero, L_second_loop);
8160
8161 Label L_carry;
8162 subl(kdx, 1);
8163 jcc(Assembler::zero, L_carry);
8164
8165 movl(Address(z, kdx, Address::times_4, 0), carry);
8166 shrq(carry, 32);
8167 subl(kdx, 1);
8168
8169 bind(L_carry);
8170 movl(Address(z, kdx, Address::times_4, 0), carry);
8171
8172 // Second and third (nested) loops.
8173 //
8174 // for (int i = xstart-1; i >= 0; i--) { // Second loop
8175 // carry = 0;
8176 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
8177 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
8178 // (z[k] & LONG_MASK) + carry;
8179 // z[k] = (int)product;
8180 // carry = product >>> 32;
8181 // }
8182 // z[i] = (int)carry;
8183 // }
8184 //
8185 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
8186
8187 const Register jdx = tmp1;
8188
8189 bind(L_second_loop);
8190 xorl(carry, carry); // carry = 0;
8191 movl(jdx, ylen); // j = ystart+1
8192
8193 subl(xstart, 1); // i = xstart-1;
8194 jcc(Assembler::negative, L_done);
8195
8196 push (z);
8197
8198 Label L_last_x;
8199 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j
8200 subl(xstart, 1); // i = xstart-1;
8201 jcc(Assembler::negative, L_last_x);
8202
8203 if (UseBMI2Instructions) {
8204 movq(rdx, Address(x, xstart, Address::times_4, 0));
8205 rorxq(rdx, rdx, 32); // convert big-endian to little-endian
8206 } else {
8207 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
8208 rorq(x_xstart, 32); // convert big-endian to little-endian
8209 }
8210
8211 Label L_third_loop_prologue;
8212 bind(L_third_loop_prologue);
8213
8214 push (x);
8215 push (xstart);
8216 push (ylen);
8217
8218
8219 if (UseBMI2Instructions) {
8220 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4);
8221 } else { // !UseBMI2Instructions
8222 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
8223 }
8224
8225 pop(ylen);
8226 pop(xlen);
8227 pop(x);
8228 pop(z);
8229
8230 movl(tmp3, xlen);
8231 addl(tmp3, 1);
8232 movl(Address(z, tmp3, Address::times_4, 0), carry);
8233 subl(tmp3, 1);
8234 jccb(Assembler::negative, L_done);
8235
8236 shrq(carry, 32);
8237 movl(Address(z, tmp3, Address::times_4, 0), carry);
8238 jmp(L_second_loop);
8239
8240 // Next infrequent code is moved outside loops.
8241 bind(L_last_x);
8242 if (UseBMI2Instructions) {
8243 movl(rdx, Address(x, 0));
8244 } else {
8245 movl(x_xstart, Address(x, 0));
8246 }
8247 jmp(L_third_loop_prologue);
8248
8249 bind(L_done);
8250
8251 pop(zlen);
8252 pop(xlen);
8253
8254 pop(tmp5);
8255 pop(tmp4);
8256 pop(tmp3);
8257 pop(tmp2);
8258 pop(tmp1);
8259 }
8260
8261 void MacroAssembler::vectorized_mismatch(Register obja, Register objb, Register length, Register log2_array_indxscale,
8262 Register result, Register tmp1, Register tmp2, XMMRegister rymm0, XMMRegister rymm1, XMMRegister rymm2){
8263 assert(UseSSE42Intrinsics, "SSE4.2 must be enabled.");
8264 Label VECTOR16_LOOP, VECTOR8_LOOP, VECTOR4_LOOP;
8265 Label VECTOR8_TAIL, VECTOR4_TAIL;
8266 Label VECTOR32_NOT_EQUAL, VECTOR16_NOT_EQUAL, VECTOR8_NOT_EQUAL, VECTOR4_NOT_EQUAL;
8267 Label SAME_TILL_END, DONE;
8268 Label BYTES_LOOP, BYTES_TAIL, BYTES_NOT_EQUAL;
8269
8270 //scale is in rcx in both Win64 and Unix
8271 ShortBranchVerifier sbv(this);
8272
8273 shlq(length);
8274 xorq(result, result);
8275
8276 if ((UseAVX > 2) &&
8277 VM_Version::supports_avx512vlbw()) {
8278 Label VECTOR64_LOOP, VECTOR64_NOT_EQUAL, VECTOR32_TAIL;
8279
8280 cmpq(length, 64);
8281 jcc(Assembler::less, VECTOR32_TAIL);
8282 movq(tmp1, length);
8283 andq(tmp1, 0x3F); // tail count
8284 andq(length, ~(0x3F)); //vector count
8285
8286 bind(VECTOR64_LOOP);
8287 // AVX512 code to compare 64 byte vectors.
8288 evmovdqub(rymm0, Address(obja, result), Assembler::AVX_512bit);
8289 evpcmpeqb(k7, rymm0, Address(objb, result), Assembler::AVX_512bit);
8290 kortestql(k7, k7);
8291 jcc(Assembler::aboveEqual, VECTOR64_NOT_EQUAL); // mismatch
8292 addq(result, 64);
8293 subq(length, 64);
8294 jccb(Assembler::notZero, VECTOR64_LOOP);
8295
8296 //bind(VECTOR64_TAIL);
8297 testq(tmp1, tmp1);
8298 jcc(Assembler::zero, SAME_TILL_END);
8299
8300 //bind(VECTOR64_TAIL);
8301 // AVX512 code to compare upto 63 byte vectors.
8302 mov64(tmp2, 0xFFFFFFFFFFFFFFFF);
8303 shlxq(tmp2, tmp2, tmp1);
8304 notq(tmp2);
8305 kmovql(k3, tmp2);
8306
8307 evmovdqub(rymm0, k3, Address(obja, result), Assembler::AVX_512bit);
8308 evpcmpeqb(k7, k3, rymm0, Address(objb, result), Assembler::AVX_512bit);
8309
8310 ktestql(k7, k3);
8311 jcc(Assembler::below, SAME_TILL_END); // not mismatch
8312
8313 bind(VECTOR64_NOT_EQUAL);
8314 kmovql(tmp1, k7);
8315 notq(tmp1);
8316 tzcntq(tmp1, tmp1);
8317 addq(result, tmp1);
8318 shrq(result);
8319 jmp(DONE);
8320 bind(VECTOR32_TAIL);
8321 }
8322
8323 cmpq(length, 8);
8324 jcc(Assembler::equal, VECTOR8_LOOP);
8325 jcc(Assembler::less, VECTOR4_TAIL);
8326
8327 if (UseAVX >= 2) {
8328 Label VECTOR16_TAIL, VECTOR32_LOOP;
8329
8330 cmpq(length, 16);
8331 jcc(Assembler::equal, VECTOR16_LOOP);
8332 jcc(Assembler::less, VECTOR8_LOOP);
8333
8334 cmpq(length, 32);
8335 jccb(Assembler::less, VECTOR16_TAIL);
8336
8337 subq(length, 32);
8338 bind(VECTOR32_LOOP);
8339 vmovdqu(rymm0, Address(obja, result));
8340 vmovdqu(rymm1, Address(objb, result));
8341 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_256bit);
8342 vptest(rymm2, rymm2);
8343 jcc(Assembler::notZero, VECTOR32_NOT_EQUAL);//mismatch found
8344 addq(result, 32);
8345 subq(length, 32);
8346 jcc(Assembler::greaterEqual, VECTOR32_LOOP);
8347 addq(length, 32);
8348 jcc(Assembler::equal, SAME_TILL_END);
8349 //falling through if less than 32 bytes left //close the branch here.
8350
8351 bind(VECTOR16_TAIL);
8352 cmpq(length, 16);
8353 jccb(Assembler::less, VECTOR8_TAIL);
8354 bind(VECTOR16_LOOP);
8355 movdqu(rymm0, Address(obja, result));
8356 movdqu(rymm1, Address(objb, result));
8357 vpxor(rymm2, rymm0, rymm1, Assembler::AVX_128bit);
8358 ptest(rymm2, rymm2);
8359 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
8360 addq(result, 16);
8361 subq(length, 16);
8362 jcc(Assembler::equal, SAME_TILL_END);
8363 //falling through if less than 16 bytes left
8364 } else {//regular intrinsics
8365
8366 cmpq(length, 16);
8367 jccb(Assembler::less, VECTOR8_TAIL);
8368
8369 subq(length, 16);
8370 bind(VECTOR16_LOOP);
8371 movdqu(rymm0, Address(obja, result));
8372 movdqu(rymm1, Address(objb, result));
8373 pxor(rymm0, rymm1);
8374 ptest(rymm0, rymm0);
8375 jcc(Assembler::notZero, VECTOR16_NOT_EQUAL);//mismatch found
8376 addq(result, 16);
8377 subq(length, 16);
8378 jccb(Assembler::greaterEqual, VECTOR16_LOOP);
8379 addq(length, 16);
8380 jcc(Assembler::equal, SAME_TILL_END);
8381 //falling through if less than 16 bytes left
8382 }
8383
8384 bind(VECTOR8_TAIL);
8385 cmpq(length, 8);
8386 jccb(Assembler::less, VECTOR4_TAIL);
8387 bind(VECTOR8_LOOP);
8388 movq(tmp1, Address(obja, result));
8389 movq(tmp2, Address(objb, result));
8390 xorq(tmp1, tmp2);
8391 testq(tmp1, tmp1);
8392 jcc(Assembler::notZero, VECTOR8_NOT_EQUAL);//mismatch found
8393 addq(result, 8);
8394 subq(length, 8);
8395 jcc(Assembler::equal, SAME_TILL_END);
8396 //falling through if less than 8 bytes left
8397
8398 bind(VECTOR4_TAIL);
8399 cmpq(length, 4);
8400 jccb(Assembler::less, BYTES_TAIL);
8401 bind(VECTOR4_LOOP);
8402 movl(tmp1, Address(obja, result));
8403 xorl(tmp1, Address(objb, result));
8404 testl(tmp1, tmp1);
8405 jcc(Assembler::notZero, VECTOR4_NOT_EQUAL);//mismatch found
8406 addq(result, 4);
8407 subq(length, 4);
8408 jcc(Assembler::equal, SAME_TILL_END);
8409 //falling through if less than 4 bytes left
8410
8411 bind(BYTES_TAIL);
8412 bind(BYTES_LOOP);
8413 load_unsigned_byte(tmp1, Address(obja, result));
8414 load_unsigned_byte(tmp2, Address(objb, result));
8415 xorl(tmp1, tmp2);
8416 testl(tmp1, tmp1);
8417 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8418 decq(length);
8419 jcc(Assembler::zero, SAME_TILL_END);
8420 incq(result);
8421 load_unsigned_byte(tmp1, Address(obja, result));
8422 load_unsigned_byte(tmp2, Address(objb, result));
8423 xorl(tmp1, tmp2);
8424 testl(tmp1, tmp1);
8425 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8426 decq(length);
8427 jcc(Assembler::zero, SAME_TILL_END);
8428 incq(result);
8429 load_unsigned_byte(tmp1, Address(obja, result));
8430 load_unsigned_byte(tmp2, Address(objb, result));
8431 xorl(tmp1, tmp2);
8432 testl(tmp1, tmp1);
8433 jcc(Assembler::notZero, BYTES_NOT_EQUAL);//mismatch found
8434 jmp(SAME_TILL_END);
8435
8436 if (UseAVX >= 2) {
8437 bind(VECTOR32_NOT_EQUAL);
8438 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_256bit);
8439 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_256bit);
8440 vpxor(rymm0, rymm0, rymm2, Assembler::AVX_256bit);
8441 vpmovmskb(tmp1, rymm0);
8442 bsfq(tmp1, tmp1);
8443 addq(result, tmp1);
8444 shrq(result);
8445 jmp(DONE);
8446 }
8447
8448 bind(VECTOR16_NOT_EQUAL);
8449 if (UseAVX >= 2) {
8450 vpcmpeqb(rymm2, rymm2, rymm2, Assembler::AVX_128bit);
8451 vpcmpeqb(rymm0, rymm0, rymm1, Assembler::AVX_128bit);
8452 pxor(rymm0, rymm2);
8453 } else {
8454 pcmpeqb(rymm2, rymm2);
8455 pxor(rymm0, rymm1);
8456 pcmpeqb(rymm0, rymm1);
8457 pxor(rymm0, rymm2);
8458 }
8459 pmovmskb(tmp1, rymm0);
8460 bsfq(tmp1, tmp1);
8461 addq(result, tmp1);
8462 shrq(result);
8463 jmpb(DONE);
8464
8465 bind(VECTOR8_NOT_EQUAL);
8466 bind(VECTOR4_NOT_EQUAL);
8467 bsfq(tmp1, tmp1);
8468 shrq(tmp1, 3);
8469 addq(result, tmp1);
8470 bind(BYTES_NOT_EQUAL);
8471 shrq(result);
8472 jmpb(DONE);
8473
8474 bind(SAME_TILL_END);
8475 mov64(result, -1);
8476
8477 bind(DONE);
8478 }
8479
8480 //Helper functions for square_to_len()
8481
8482 /**
8483 * Store the squares of x[], right shifted one bit (divided by 2) into z[]
8484 * Preserves x and z and modifies rest of the registers.
8485 */
8486 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8487 // Perform square and right shift by 1
8488 // Handle odd xlen case first, then for even xlen do the following
8489 // jlong carry = 0;
8490 // for (int j=0, i=0; j < xlen; j+=2, i+=4) {
8491 // huge_128 product = x[j:j+1] * x[j:j+1];
8492 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65);
8493 // z[i+2:i+3] = (jlong)(product >>> 1);
8494 // carry = (jlong)product;
8495 // }
8496
8497 xorq(tmp5, tmp5); // carry
8498 xorq(rdxReg, rdxReg);
8499 xorl(tmp1, tmp1); // index for x
8500 xorl(tmp4, tmp4); // index for z
8501
8502 Label L_first_loop, L_first_loop_exit;
8503
8504 testl(xlen, 1);
8505 jccb(Assembler::zero, L_first_loop); //jump if xlen is even
8506
8507 // Square and right shift by 1 the odd element using 32 bit multiply
8508 movl(raxReg, Address(x, tmp1, Address::times_4, 0));
8509 imulq(raxReg, raxReg);
8510 shrq(raxReg, 1);
8511 adcq(tmp5, 0);
8512 movq(Address(z, tmp4, Address::times_4, 0), raxReg);
8513 incrementl(tmp1);
8514 addl(tmp4, 2);
8515
8516 // Square and right shift by 1 the rest using 64 bit multiply
8517 bind(L_first_loop);
8518 cmpptr(tmp1, xlen);
8519 jccb(Assembler::equal, L_first_loop_exit);
8520
8521 // Square
8522 movq(raxReg, Address(x, tmp1, Address::times_4, 0));
8523 rorq(raxReg, 32); // convert big-endian to little-endian
8524 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax
8525
8526 // Right shift by 1 and save carry
8527 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1
8528 rcrq(rdxReg, 1);
8529 rcrq(raxReg, 1);
8530 adcq(tmp5, 0);
8531
8532 // Store result in z
8533 movq(Address(z, tmp4, Address::times_4, 0), rdxReg);
8534 movq(Address(z, tmp4, Address::times_4, 8), raxReg);
8535
8536 // Update indices for x and z
8537 addl(tmp1, 2);
8538 addl(tmp4, 4);
8539 jmp(L_first_loop);
8540
8541 bind(L_first_loop_exit);
8542 }
8543
8544
8545 /**
8546 * Perform the following multiply add operation using BMI2 instructions
8547 * carry:sum = sum + op1*op2 + carry
8548 * op2 should be in rdx
8549 * op2 is preserved, all other registers are modified
8550 */
8551 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) {
8552 // assert op2 is rdx
8553 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1
8554 addq(sum, carry);
8555 adcq(tmp2, 0);
8556 addq(sum, op1);
8557 adcq(tmp2, 0);
8558 movq(carry, tmp2);
8559 }
8560
8561 /**
8562 * Perform the following multiply add operation:
8563 * carry:sum = sum + op1*op2 + carry
8564 * Preserves op1, op2 and modifies rest of registers
8565 */
8566 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) {
8567 // rdx:rax = op1 * op2
8568 movq(raxReg, op2);
8569 mulq(op1);
8570
8571 // rdx:rax = sum + carry + rdx:rax
8572 addq(sum, carry);
8573 adcq(rdxReg, 0);
8574 addq(sum, raxReg);
8575 adcq(rdxReg, 0);
8576
8577 // carry:sum = rdx:sum
8578 movq(carry, rdxReg);
8579 }
8580
8581 /**
8582 * Add 64 bit long carry into z[] with carry propogation.
8583 * Preserves z and carry register values and modifies rest of registers.
8584 *
8585 */
8586 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) {
8587 Label L_fourth_loop, L_fourth_loop_exit;
8588
8589 movl(tmp1, 1);
8590 subl(zlen, 2);
8591 addq(Address(z, zlen, Address::times_4, 0), carry);
8592
8593 bind(L_fourth_loop);
8594 jccb(Assembler::carryClear, L_fourth_loop_exit);
8595 subl(zlen, 2);
8596 jccb(Assembler::negative, L_fourth_loop_exit);
8597 addq(Address(z, zlen, Address::times_4, 0), tmp1);
8598 jmp(L_fourth_loop);
8599 bind(L_fourth_loop_exit);
8600 }
8601
8602 /**
8603 * Shift z[] left by 1 bit.
8604 * Preserves x, len, z and zlen registers and modifies rest of the registers.
8605 *
8606 */
8607 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) {
8608
8609 Label L_fifth_loop, L_fifth_loop_exit;
8610
8611 // Fifth loop
8612 // Perform primitiveLeftShift(z, zlen, 1)
8613
8614 const Register prev_carry = tmp1;
8615 const Register new_carry = tmp4;
8616 const Register value = tmp2;
8617 const Register zidx = tmp3;
8618
8619 // int zidx, carry;
8620 // long value;
8621 // carry = 0;
8622 // for (zidx = zlen-2; zidx >=0; zidx -= 2) {
8623 // (carry:value) = (z[i] << 1) | carry ;
8624 // z[i] = value;
8625 // }
8626
8627 movl(zidx, zlen);
8628 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register
8629
8630 bind(L_fifth_loop);
8631 decl(zidx); // Use decl to preserve carry flag
8632 decl(zidx);
8633 jccb(Assembler::negative, L_fifth_loop_exit);
8634
8635 if (UseBMI2Instructions) {
8636 movq(value, Address(z, zidx, Address::times_4, 0));
8637 rclq(value, 1);
8638 rorxq(value, value, 32);
8639 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
8640 }
8641 else {
8642 // clear new_carry
8643 xorl(new_carry, new_carry);
8644
8645 // Shift z[i] by 1, or in previous carry and save new carry
8646 movq(value, Address(z, zidx, Address::times_4, 0));
8647 shlq(value, 1);
8648 adcl(new_carry, 0);
8649
8650 orq(value, prev_carry);
8651 rorq(value, 0x20);
8652 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
8653
8654 // Set previous carry = new carry
8655 movl(prev_carry, new_carry);
8656 }
8657 jmp(L_fifth_loop);
8658
8659 bind(L_fifth_loop_exit);
8660 }
8661
8662
8663 /**
8664 * Code for BigInteger::squareToLen() intrinsic
8665 *
8666 * rdi: x
8667 * rsi: len
8668 * r8: z
8669 * rcx: zlen
8670 * r12: tmp1
8671 * r13: tmp2
8672 * r14: tmp3
8673 * r15: tmp4
8674 * rbx: tmp5
8675 *
8676 */
8677 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) {
8678
8679 Label L_second_loop, L_second_loop_exit, L_third_loop, L_third_loop_exit, L_last_x, L_multiply;
8680 push(tmp1);
8681 push(tmp2);
8682 push(tmp3);
8683 push(tmp4);
8684 push(tmp5);
8685
8686 // First loop
8687 // Store the squares, right shifted one bit (i.e., divided by 2).
8688 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg);
8689
8690 // Add in off-diagonal sums.
8691 //
8692 // Second, third (nested) and fourth loops.
8693 // zlen +=2;
8694 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) {
8695 // carry = 0;
8696 // long op2 = x[xidx:xidx+1];
8697 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) {
8698 // k -= 2;
8699 // long op1 = x[j:j+1];
8700 // long sum = z[k:k+1];
8701 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs);
8702 // z[k:k+1] = sum;
8703 // }
8704 // add_one_64(z, k, carry, tmp_regs);
8705 // }
8706
8707 const Register carry = tmp5;
8708 const Register sum = tmp3;
8709 const Register op1 = tmp4;
8710 Register op2 = tmp2;
8711
8712 push(zlen);
8713 push(len);
8714 addl(zlen,2);
8715 bind(L_second_loop);
8716 xorq(carry, carry);
8717 subl(zlen, 4);
8718 subl(len, 2);
8719 push(zlen);
8720 push(len);
8721 cmpl(len, 0);
8722 jccb(Assembler::lessEqual, L_second_loop_exit);
8723
8724 // Multiply an array by one 64 bit long.
8725 if (UseBMI2Instructions) {
8726 op2 = rdxReg;
8727 movq(op2, Address(x, len, Address::times_4, 0));
8728 rorxq(op2, op2, 32);
8729 }
8730 else {
8731 movq(op2, Address(x, len, Address::times_4, 0));
8732 rorq(op2, 32);
8733 }
8734
8735 bind(L_third_loop);
8736 decrementl(len);
8737 jccb(Assembler::negative, L_third_loop_exit);
8738 decrementl(len);
8739 jccb(Assembler::negative, L_last_x);
8740
8741 movq(op1, Address(x, len, Address::times_4, 0));
8742 rorq(op1, 32);
8743
8744 bind(L_multiply);
8745 subl(zlen, 2);
8746 movq(sum, Address(z, zlen, Address::times_4, 0));
8747
8748 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry.
8749 if (UseBMI2Instructions) {
8750 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2);
8751 }
8752 else {
8753 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8754 }
8755
8756 movq(Address(z, zlen, Address::times_4, 0), sum);
8757
8758 jmp(L_third_loop);
8759 bind(L_third_loop_exit);
8760
8761 // Fourth loop
8762 // Add 64 bit long carry into z with carry propogation.
8763 // Uses offsetted zlen.
8764 add_one_64(z, zlen, carry, tmp1);
8765
8766 pop(len);
8767 pop(zlen);
8768 jmp(L_second_loop);
8769
8770 // Next infrequent code is moved outside loops.
8771 bind(L_last_x);
8772 movl(op1, Address(x, 0));
8773 jmp(L_multiply);
8774
8775 bind(L_second_loop_exit);
8776 pop(len);
8777 pop(zlen);
8778 pop(len);
8779 pop(zlen);
8780
8781 // Fifth loop
8782 // Shift z left 1 bit.
8783 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4);
8784
8785 // z[zlen-1] |= x[len-1] & 1;
8786 movl(tmp3, Address(x, len, Address::times_4, -4));
8787 andl(tmp3, 1);
8788 orl(Address(z, zlen, Address::times_4, -4), tmp3);
8789
8790 pop(tmp5);
8791 pop(tmp4);
8792 pop(tmp3);
8793 pop(tmp2);
8794 pop(tmp1);
8795 }
8796
8797 /**
8798 * Helper function for mul_add()
8799 * Multiply the in[] by int k and add to out[] starting at offset offs using
8800 * 128 bit by 32 bit multiply and return the carry in tmp5.
8801 * Only quad int aligned length of in[] is operated on in this function.
8802 * k is in rdxReg for BMI2Instructions, for others it is in tmp2.
8803 * This function preserves out, in and k registers.
8804 * len and offset point to the appropriate index in "in" & "out" correspondingly
8805 * tmp5 has the carry.
8806 * other registers are temporary and are modified.
8807 *
8808 */
8809 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in,
8810 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3,
8811 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8812
8813 Label L_first_loop, L_first_loop_exit;
8814
8815 movl(tmp1, len);
8816 shrl(tmp1, 2);
8817
8818 bind(L_first_loop);
8819 subl(tmp1, 1);
8820 jccb(Assembler::negative, L_first_loop_exit);
8821
8822 subl(len, 4);
8823 subl(offset, 4);
8824
8825 Register op2 = tmp2;
8826 const Register sum = tmp3;
8827 const Register op1 = tmp4;
8828 const Register carry = tmp5;
8829
8830 if (UseBMI2Instructions) {
8831 op2 = rdxReg;
8832 }
8833
8834 movq(op1, Address(in, len, Address::times_4, 8));
8835 rorq(op1, 32);
8836 movq(sum, Address(out, offset, Address::times_4, 8));
8837 rorq(sum, 32);
8838 if (UseBMI2Instructions) {
8839 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8840 }
8841 else {
8842 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8843 }
8844 // Store back in big endian from little endian
8845 rorq(sum, 0x20);
8846 movq(Address(out, offset, Address::times_4, 8), sum);
8847
8848 movq(op1, Address(in, len, Address::times_4, 0));
8849 rorq(op1, 32);
8850 movq(sum, Address(out, offset, Address::times_4, 0));
8851 rorq(sum, 32);
8852 if (UseBMI2Instructions) {
8853 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8854 }
8855 else {
8856 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8857 }
8858 // Store back in big endian from little endian
8859 rorq(sum, 0x20);
8860 movq(Address(out, offset, Address::times_4, 0), sum);
8861
8862 jmp(L_first_loop);
8863 bind(L_first_loop_exit);
8864 }
8865
8866 /**
8867 * Code for BigInteger::mulAdd() intrinsic
8868 *
8869 * rdi: out
8870 * rsi: in
8871 * r11: offs (out.length - offset)
8872 * rcx: len
8873 * r8: k
8874 * r12: tmp1
8875 * r13: tmp2
8876 * r14: tmp3
8877 * r15: tmp4
8878 * rbx: tmp5
8879 * Multiply the in[] by word k and add to out[], return the carry in rax
8880 */
8881 void MacroAssembler::mul_add(Register out, Register in, Register offs,
8882 Register len, Register k, Register tmp1, Register tmp2, Register tmp3,
8883 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
8884
8885 Label L_carry, L_last_in, L_done;
8886
8887 // carry = 0;
8888 // for (int j=len-1; j >= 0; j--) {
8889 // long product = (in[j] & LONG_MASK) * kLong +
8890 // (out[offs] & LONG_MASK) + carry;
8891 // out[offs--] = (int)product;
8892 // carry = product >>> 32;
8893 // }
8894 //
8895 push(tmp1);
8896 push(tmp2);
8897 push(tmp3);
8898 push(tmp4);
8899 push(tmp5);
8900
8901 Register op2 = tmp2;
8902 const Register sum = tmp3;
8903 const Register op1 = tmp4;
8904 const Register carry = tmp5;
8905
8906 if (UseBMI2Instructions) {
8907 op2 = rdxReg;
8908 movl(op2, k);
8909 }
8910 else {
8911 movl(op2, k);
8912 }
8913
8914 xorq(carry, carry);
8915
8916 //First loop
8917
8918 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply
8919 //The carry is in tmp5
8920 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg);
8921
8922 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any
8923 decrementl(len);
8924 jccb(Assembler::negative, L_carry);
8925 decrementl(len);
8926 jccb(Assembler::negative, L_last_in);
8927
8928 movq(op1, Address(in, len, Address::times_4, 0));
8929 rorq(op1, 32);
8930
8931 subl(offs, 2);
8932 movq(sum, Address(out, offs, Address::times_4, 0));
8933 rorq(sum, 32);
8934
8935 if (UseBMI2Instructions) {
8936 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
8937 }
8938 else {
8939 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
8940 }
8941
8942 // Store back in big endian from little endian
8943 rorq(sum, 0x20);
8944 movq(Address(out, offs, Address::times_4, 0), sum);
8945
8946 testl(len, len);
8947 jccb(Assembler::zero, L_carry);
8948
8949 //Multiply the last in[] entry, if any
8950 bind(L_last_in);
8951 movl(op1, Address(in, 0));
8952 movl(sum, Address(out, offs, Address::times_4, -4));
8953
8954 movl(raxReg, k);
8955 mull(op1); //tmp4 * eax -> edx:eax
8956 addl(sum, carry);
8957 adcl(rdxReg, 0);
8958 addl(sum, raxReg);
8959 adcl(rdxReg, 0);
8960 movl(carry, rdxReg);
8961
8962 movl(Address(out, offs, Address::times_4, -4), sum);
8963
8964 bind(L_carry);
8965 //return tmp5/carry as carry in rax
8966 movl(rax, carry);
8967
8968 bind(L_done);
8969 pop(tmp5);
8970 pop(tmp4);
8971 pop(tmp3);
8972 pop(tmp2);
8973 pop(tmp1);
8974 }
8975 #endif
8976
8977 /**
8978 * Emits code to update CRC-32 with a byte value according to constants in table
8979 *
8980 * @param [in,out]crc Register containing the crc.
8981 * @param [in]val Register containing the byte to fold into the CRC.
8982 * @param [in]table Register containing the table of crc constants.
8983 *
8984 * uint32_t crc;
8985 * val = crc_table[(val ^ crc) & 0xFF];
8986 * crc = val ^ (crc >> 8);
8987 *
8988 */
8989 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
8990 xorl(val, crc);
8991 andl(val, 0xFF);
8992 shrl(crc, 8); // unsigned shift
8993 xorl(crc, Address(table, val, Address::times_4, 0));
8994 }
8995
8996 /**
8997 * Fold four 128-bit data chunks
8998 */
8999 void MacroAssembler::fold_128bit_crc32_avx512(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
9000 evpclmulhdq(xtmp, xK, xcrc, Assembler::AVX_512bit); // [123:64]
9001 evpclmulldq(xcrc, xK, xcrc, Assembler::AVX_512bit); // [63:0]
9002 evpxorq(xcrc, xcrc, Address(buf, offset), Assembler::AVX_512bit /* vector_len */);
9003 evpxorq(xcrc, xcrc, xtmp, Assembler::AVX_512bit /* vector_len */);
9004 }
9005
9006 /**
9007 * Fold 128-bit data chunk
9008 */
9009 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
9010 if (UseAVX > 0) {
9011 vpclmulhdq(xtmp, xK, xcrc); // [123:64]
9012 vpclmulldq(xcrc, xK, xcrc); // [63:0]
9013 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */);
9014 pxor(xcrc, xtmp);
9015 } else {
9016 movdqa(xtmp, xcrc);
9017 pclmulhdq(xtmp, xK); // [123:64]
9018 pclmulldq(xcrc, xK); // [63:0]
9019 pxor(xcrc, xtmp);
9020 movdqu(xtmp, Address(buf, offset));
9021 pxor(xcrc, xtmp);
9022 }
9023 }
9024
9025 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
9026 if (UseAVX > 0) {
9027 vpclmulhdq(xtmp, xK, xcrc);
9028 vpclmulldq(xcrc, xK, xcrc);
9029 pxor(xcrc, xbuf);
9030 pxor(xcrc, xtmp);
9031 } else {
9032 movdqa(xtmp, xcrc);
9033 pclmulhdq(xtmp, xK);
9034 pclmulldq(xcrc, xK);
9035 pxor(xcrc, xbuf);
9036 pxor(xcrc, xtmp);
9037 }
9038 }
9039
9040 /**
9041 * 8-bit folds to compute 32-bit CRC
9042 *
9043 * uint64_t xcrc;
9044 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8);
9045 */
9046 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) {
9047 movdl(tmp, xcrc);
9048 andl(tmp, 0xFF);
9049 movdl(xtmp, Address(table, tmp, Address::times_4, 0));
9050 psrldq(xcrc, 1); // unsigned shift one byte
9051 pxor(xcrc, xtmp);
9052 }
9053
9054 /**
9055 * uint32_t crc;
9056 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
9057 */
9058 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
9059 movl(tmp, crc);
9060 andl(tmp, 0xFF);
9061 shrl(crc, 8);
9062 xorl(crc, Address(table, tmp, Address::times_4, 0));
9063 }
9064
9065 /**
9066 * @param crc register containing existing CRC (32-bit)
9067 * @param buf register pointing to input byte buffer (byte*)
9068 * @param len register containing number of bytes
9069 * @param table register that will contain address of CRC table
9070 * @param tmp scratch register
9071 */
9072 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) {
9073 assert_different_registers(crc, buf, len, table, tmp, rax);
9074
9075 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned;
9076 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop;
9077
9078 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge
9079 // context for the registers used, where all instructions below are using 128-bit mode
9080 // On EVEX without VL and BW, these instructions will all be AVX.
9081 lea(table, ExternalAddress(StubRoutines::crc_table_addr()));
9082 notl(crc); // ~crc
9083 cmpl(len, 16);
9084 jcc(Assembler::less, L_tail);
9085
9086 // Align buffer to 16 bytes
9087 movl(tmp, buf);
9088 andl(tmp, 0xF);
9089 jccb(Assembler::zero, L_aligned);
9090 subl(tmp, 16);
9091 addl(len, tmp);
9092
9093 align(4);
9094 BIND(L_align_loop);
9095 movsbl(rax, Address(buf, 0)); // load byte with sign extension
9096 update_byte_crc32(crc, rax, table);
9097 increment(buf);
9098 incrementl(tmp);
9099 jccb(Assembler::less, L_align_loop);
9100
9101 BIND(L_aligned);
9102 movl(tmp, len); // save
9103 shrl(len, 4);
9104 jcc(Assembler::zero, L_tail_restore);
9105
9106 // Fold total 512 bits of polynomial on each iteration
9107 if (VM_Version::supports_vpclmulqdq()) {
9108 Label Parallel_loop, L_No_Parallel;
9109
9110 cmpl(len, 8);
9111 jccb(Assembler::less, L_No_Parallel);
9112
9113 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
9114 evmovdquq(xmm1, Address(buf, 0), Assembler::AVX_512bit);
9115 movdl(xmm5, crc);
9116 evpxorq(xmm1, xmm1, xmm5, Assembler::AVX_512bit);
9117 addptr(buf, 64);
9118 subl(len, 7);
9119 evshufi64x2(xmm0, xmm0, xmm0, 0x00, Assembler::AVX_512bit); //propagate the mask from 128 bits to 512 bits
9120
9121 BIND(Parallel_loop);
9122 fold_128bit_crc32_avx512(xmm1, xmm0, xmm5, buf, 0);
9123 addptr(buf, 64);
9124 subl(len, 4);
9125 jcc(Assembler::greater, Parallel_loop);
9126
9127 vextracti64x2(xmm2, xmm1, 0x01);
9128 vextracti64x2(xmm3, xmm1, 0x02);
9129 vextracti64x2(xmm4, xmm1, 0x03);
9130 jmp(L_fold_512b);
9131
9132 BIND(L_No_Parallel);
9133 }
9134 // Fold crc into first bytes of vector
9135 movdqa(xmm1, Address(buf, 0));
9136 movdl(rax, xmm1);
9137 xorl(crc, rax);
9138 if (VM_Version::supports_sse4_1()) {
9139 pinsrd(xmm1, crc, 0);
9140 } else {
9141 pinsrw(xmm1, crc, 0);
9142 shrl(crc, 16);
9143 pinsrw(xmm1, crc, 1);
9144 }
9145 addptr(buf, 16);
9146 subl(len, 4); // len > 0
9147 jcc(Assembler::less, L_fold_tail);
9148
9149 movdqa(xmm2, Address(buf, 0));
9150 movdqa(xmm3, Address(buf, 16));
9151 movdqa(xmm4, Address(buf, 32));
9152 addptr(buf, 48);
9153 subl(len, 3);
9154 jcc(Assembler::lessEqual, L_fold_512b);
9155
9156 // Fold total 512 bits of polynomial on each iteration,
9157 // 128 bits per each of 4 parallel streams.
9158 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
9159
9160 align(32);
9161 BIND(L_fold_512b_loop);
9162 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
9163 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16);
9164 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32);
9165 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48);
9166 addptr(buf, 64);
9167 subl(len, 4);
9168 jcc(Assembler::greater, L_fold_512b_loop);
9169
9170 // Fold 512 bits to 128 bits.
9171 BIND(L_fold_512b);
9172 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
9173 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2);
9174 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3);
9175 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4);
9176
9177 // Fold the rest of 128 bits data chunks
9178 BIND(L_fold_tail);
9179 addl(len, 3);
9180 jccb(Assembler::lessEqual, L_fold_128b);
9181 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
9182
9183 BIND(L_fold_tail_loop);
9184 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
9185 addptr(buf, 16);
9186 decrementl(len);
9187 jccb(Assembler::greater, L_fold_tail_loop);
9188
9189 // Fold 128 bits in xmm1 down into 32 bits in crc register.
9190 BIND(L_fold_128b);
9191 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
9192 if (UseAVX > 0) {
9193 vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
9194 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */);
9195 vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
9196 } else {
9197 movdqa(xmm2, xmm0);
9198 pclmulqdq(xmm2, xmm1, 0x1);
9199 movdqa(xmm3, xmm0);
9200 pand(xmm3, xmm2);
9201 pclmulqdq(xmm0, xmm3, 0x1);
9202 }
9203 psrldq(xmm1, 8);
9204 psrldq(xmm2, 4);
9205 pxor(xmm0, xmm1);
9206 pxor(xmm0, xmm2);
9207
9208 // 8 8-bit folds to compute 32-bit CRC.
9209 for (int j = 0; j < 4; j++) {
9210 fold_8bit_crc32(xmm0, table, xmm1, rax);
9211 }
9212 movdl(crc, xmm0); // mov 32 bits to general register
9213 for (int j = 0; j < 4; j++) {
9214 fold_8bit_crc32(crc, table, rax);
9215 }
9216
9217 BIND(L_tail_restore);
9218 movl(len, tmp); // restore
9219 BIND(L_tail);
9220 andl(len, 0xf);
9221 jccb(Assembler::zero, L_exit);
9222
9223 // Fold the rest of bytes
9224 align(4);
9225 BIND(L_tail_loop);
9226 movsbl(rax, Address(buf, 0)); // load byte with sign extension
9227 update_byte_crc32(crc, rax, table);
9228 increment(buf);
9229 decrementl(len);
9230 jccb(Assembler::greater, L_tail_loop);
9231
9232 BIND(L_exit);
9233 notl(crc); // ~c
9234 }
9235
9236 #ifdef _LP64
9237 // S. Gueron / Information Processing Letters 112 (2012) 184
9238 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table.
9239 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0].
9240 // Output: the 64-bit carry-less product of B * CONST
9241 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n,
9242 Register tmp1, Register tmp2, Register tmp3) {
9243 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
9244 if (n > 0) {
9245 addq(tmp3, n * 256 * 8);
9246 }
9247 // Q1 = TABLEExt[n][B & 0xFF];
9248 movl(tmp1, in);
9249 andl(tmp1, 0x000000FF);
9250 shll(tmp1, 3);
9251 addq(tmp1, tmp3);
9252 movq(tmp1, Address(tmp1, 0));
9253
9254 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
9255 movl(tmp2, in);
9256 shrl(tmp2, 8);
9257 andl(tmp2, 0x000000FF);
9258 shll(tmp2, 3);
9259 addq(tmp2, tmp3);
9260 movq(tmp2, Address(tmp2, 0));
9261
9262 shlq(tmp2, 8);
9263 xorq(tmp1, tmp2);
9264
9265 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
9266 movl(tmp2, in);
9267 shrl(tmp2, 16);
9268 andl(tmp2, 0x000000FF);
9269 shll(tmp2, 3);
9270 addq(tmp2, tmp3);
9271 movq(tmp2, Address(tmp2, 0));
9272
9273 shlq(tmp2, 16);
9274 xorq(tmp1, tmp2);
9275
9276 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
9277 shrl(in, 24);
9278 andl(in, 0x000000FF);
9279 shll(in, 3);
9280 addq(in, tmp3);
9281 movq(in, Address(in, 0));
9282
9283 shlq(in, 24);
9284 xorq(in, tmp1);
9285 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
9286 }
9287
9288 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
9289 Register in_out,
9290 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
9291 XMMRegister w_xtmp2,
9292 Register tmp1,
9293 Register n_tmp2, Register n_tmp3) {
9294 if (is_pclmulqdq_supported) {
9295 movdl(w_xtmp1, in_out); // modified blindly
9296
9297 movl(tmp1, const_or_pre_comp_const_index);
9298 movdl(w_xtmp2, tmp1);
9299 pclmulqdq(w_xtmp1, w_xtmp2, 0);
9300
9301 movdq(in_out, w_xtmp1);
9302 } else {
9303 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3);
9304 }
9305 }
9306
9307 // Recombination Alternative 2: No bit-reflections
9308 // T1 = (CRC_A * U1) << 1
9309 // T2 = (CRC_B * U2) << 1
9310 // C1 = T1 >> 32
9311 // C2 = T2 >> 32
9312 // T1 = T1 & 0xFFFFFFFF
9313 // T2 = T2 & 0xFFFFFFFF
9314 // T1 = CRC32(0, T1)
9315 // T2 = CRC32(0, T2)
9316 // C1 = C1 ^ T1
9317 // C2 = C2 ^ T2
9318 // CRC = C1 ^ C2 ^ CRC_C
9319 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,
9320 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9321 Register tmp1, Register tmp2,
9322 Register n_tmp3) {
9323 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9324 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9325 shlq(in_out, 1);
9326 movl(tmp1, in_out);
9327 shrq(in_out, 32);
9328 xorl(tmp2, tmp2);
9329 crc32(tmp2, tmp1, 4);
9330 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here
9331 shlq(in1, 1);
9332 movl(tmp1, in1);
9333 shrq(in1, 32);
9334 xorl(tmp2, tmp2);
9335 crc32(tmp2, tmp1, 4);
9336 xorl(in1, tmp2);
9337 xorl(in_out, in1);
9338 xorl(in_out, in2);
9339 }
9340
9341 // Set N to predefined value
9342 // Subtract from a lenght of a buffer
9343 // execute in a loop:
9344 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0
9345 // for i = 1 to N do
9346 // CRC_A = CRC32(CRC_A, A[i])
9347 // CRC_B = CRC32(CRC_B, B[i])
9348 // CRC_C = CRC32(CRC_C, C[i])
9349 // end for
9350 // Recombine
9351 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,
9352 Register in_out1, Register in_out2, Register in_out3,
9353 Register tmp1, Register tmp2, Register tmp3,
9354 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9355 Register tmp4, Register tmp5,
9356 Register n_tmp6) {
9357 Label L_processPartitions;
9358 Label L_processPartition;
9359 Label L_exit;
9360
9361 bind(L_processPartitions);
9362 cmpl(in_out1, 3 * size);
9363 jcc(Assembler::less, L_exit);
9364 xorl(tmp1, tmp1);
9365 xorl(tmp2, tmp2);
9366 movq(tmp3, in_out2);
9367 addq(tmp3, size);
9368
9369 bind(L_processPartition);
9370 crc32(in_out3, Address(in_out2, 0), 8);
9371 crc32(tmp1, Address(in_out2, size), 8);
9372 crc32(tmp2, Address(in_out2, size * 2), 8);
9373 addq(in_out2, 8);
9374 cmpq(in_out2, tmp3);
9375 jcc(Assembler::less, L_processPartition);
9376 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
9377 w_xtmp1, w_xtmp2, w_xtmp3,
9378 tmp4, tmp5,
9379 n_tmp6);
9380 addq(in_out2, 2 * size);
9381 subl(in_out1, 3 * size);
9382 jmp(L_processPartitions);
9383
9384 bind(L_exit);
9385 }
9386 #else
9387 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n,
9388 Register tmp1, Register tmp2, Register tmp3,
9389 XMMRegister xtmp1, XMMRegister xtmp2) {
9390 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
9391 if (n > 0) {
9392 addl(tmp3, n * 256 * 8);
9393 }
9394 // Q1 = TABLEExt[n][B & 0xFF];
9395 movl(tmp1, in_out);
9396 andl(tmp1, 0x000000FF);
9397 shll(tmp1, 3);
9398 addl(tmp1, tmp3);
9399 movq(xtmp1, Address(tmp1, 0));
9400
9401 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
9402 movl(tmp2, in_out);
9403 shrl(tmp2, 8);
9404 andl(tmp2, 0x000000FF);
9405 shll(tmp2, 3);
9406 addl(tmp2, tmp3);
9407 movq(xtmp2, Address(tmp2, 0));
9408
9409 psllq(xtmp2, 8);
9410 pxor(xtmp1, xtmp2);
9411
9412 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
9413 movl(tmp2, in_out);
9414 shrl(tmp2, 16);
9415 andl(tmp2, 0x000000FF);
9416 shll(tmp2, 3);
9417 addl(tmp2, tmp3);
9418 movq(xtmp2, Address(tmp2, 0));
9419
9420 psllq(xtmp2, 16);
9421 pxor(xtmp1, xtmp2);
9422
9423 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
9424 shrl(in_out, 24);
9425 andl(in_out, 0x000000FF);
9426 shll(in_out, 3);
9427 addl(in_out, tmp3);
9428 movq(xtmp2, Address(in_out, 0));
9429
9430 psllq(xtmp2, 24);
9431 pxor(xtmp1, xtmp2); // Result in CXMM
9432 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
9433 }
9434
9435 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
9436 Register in_out,
9437 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
9438 XMMRegister w_xtmp2,
9439 Register tmp1,
9440 Register n_tmp2, Register n_tmp3) {
9441 if (is_pclmulqdq_supported) {
9442 movdl(w_xtmp1, in_out);
9443
9444 movl(tmp1, const_or_pre_comp_const_index);
9445 movdl(w_xtmp2, tmp1);
9446 pclmulqdq(w_xtmp1, w_xtmp2, 0);
9447 // Keep result in XMM since GPR is 32 bit in length
9448 } else {
9449 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2);
9450 }
9451 }
9452
9453 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,
9454 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9455 Register tmp1, Register tmp2,
9456 Register n_tmp3) {
9457 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9458 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
9459
9460 psllq(w_xtmp1, 1);
9461 movdl(tmp1, w_xtmp1);
9462 psrlq(w_xtmp1, 32);
9463 movdl(in_out, w_xtmp1);
9464
9465 xorl(tmp2, tmp2);
9466 crc32(tmp2, tmp1, 4);
9467 xorl(in_out, tmp2);
9468
9469 psllq(w_xtmp2, 1);
9470 movdl(tmp1, w_xtmp2);
9471 psrlq(w_xtmp2, 32);
9472 movdl(in1, w_xtmp2);
9473
9474 xorl(tmp2, tmp2);
9475 crc32(tmp2, tmp1, 4);
9476 xorl(in1, tmp2);
9477 xorl(in_out, in1);
9478 xorl(in_out, in2);
9479 }
9480
9481 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,
9482 Register in_out1, Register in_out2, Register in_out3,
9483 Register tmp1, Register tmp2, Register tmp3,
9484 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9485 Register tmp4, Register tmp5,
9486 Register n_tmp6) {
9487 Label L_processPartitions;
9488 Label L_processPartition;
9489 Label L_exit;
9490
9491 bind(L_processPartitions);
9492 cmpl(in_out1, 3 * size);
9493 jcc(Assembler::less, L_exit);
9494 xorl(tmp1, tmp1);
9495 xorl(tmp2, tmp2);
9496 movl(tmp3, in_out2);
9497 addl(tmp3, size);
9498
9499 bind(L_processPartition);
9500 crc32(in_out3, Address(in_out2, 0), 4);
9501 crc32(tmp1, Address(in_out2, size), 4);
9502 crc32(tmp2, Address(in_out2, size*2), 4);
9503 crc32(in_out3, Address(in_out2, 0+4), 4);
9504 crc32(tmp1, Address(in_out2, size+4), 4);
9505 crc32(tmp2, Address(in_out2, size*2+4), 4);
9506 addl(in_out2, 8);
9507 cmpl(in_out2, tmp3);
9508 jcc(Assembler::less, L_processPartition);
9509
9510 push(tmp3);
9511 push(in_out1);
9512 push(in_out2);
9513 tmp4 = tmp3;
9514 tmp5 = in_out1;
9515 n_tmp6 = in_out2;
9516
9517 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
9518 w_xtmp1, w_xtmp2, w_xtmp3,
9519 tmp4, tmp5,
9520 n_tmp6);
9521
9522 pop(in_out2);
9523 pop(in_out1);
9524 pop(tmp3);
9525
9526 addl(in_out2, 2 * size);
9527 subl(in_out1, 3 * size);
9528 jmp(L_processPartitions);
9529
9530 bind(L_exit);
9531 }
9532 #endif //LP64
9533
9534 #ifdef _LP64
9535 // Algorithm 2: Pipelined usage of the CRC32 instruction.
9536 // Input: A buffer I of L bytes.
9537 // Output: the CRC32C value of the buffer.
9538 // Notations:
9539 // Write L = 24N + r, with N = floor (L/24).
9540 // r = L mod 24 (0 <= r < 24).
9541 // Consider I as the concatenation of A|B|C|R, where A, B, C, each,
9542 // N quadwords, and R consists of r bytes.
9543 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1
9544 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1
9545 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1
9546 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1
9547 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
9548 Register tmp1, Register tmp2, Register tmp3,
9549 Register tmp4, Register tmp5, Register tmp6,
9550 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9551 bool is_pclmulqdq_supported) {
9552 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
9553 Label L_wordByWord;
9554 Label L_byteByByteProlog;
9555 Label L_byteByByte;
9556 Label L_exit;
9557
9558 if (is_pclmulqdq_supported ) {
9559 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
9560 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1);
9561
9562 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
9563 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
9564
9565 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
9566 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
9567 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\"");
9568 } else {
9569 const_or_pre_comp_const_index[0] = 1;
9570 const_or_pre_comp_const_index[1] = 0;
9571
9572 const_or_pre_comp_const_index[2] = 3;
9573 const_or_pre_comp_const_index[3] = 2;
9574
9575 const_or_pre_comp_const_index[4] = 5;
9576 const_or_pre_comp_const_index[5] = 4;
9577 }
9578 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
9579 in2, in1, in_out,
9580 tmp1, tmp2, tmp3,
9581 w_xtmp1, w_xtmp2, w_xtmp3,
9582 tmp4, tmp5,
9583 tmp6);
9584 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
9585 in2, in1, in_out,
9586 tmp1, tmp2, tmp3,
9587 w_xtmp1, w_xtmp2, w_xtmp3,
9588 tmp4, tmp5,
9589 tmp6);
9590 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
9591 in2, in1, in_out,
9592 tmp1, tmp2, tmp3,
9593 w_xtmp1, w_xtmp2, w_xtmp3,
9594 tmp4, tmp5,
9595 tmp6);
9596 movl(tmp1, in2);
9597 andl(tmp1, 0x00000007);
9598 negl(tmp1);
9599 addl(tmp1, in2);
9600 addq(tmp1, in1);
9601
9602 BIND(L_wordByWord);
9603 cmpq(in1, tmp1);
9604 jcc(Assembler::greaterEqual, L_byteByByteProlog);
9605 crc32(in_out, Address(in1, 0), 4);
9606 addq(in1, 4);
9607 jmp(L_wordByWord);
9608
9609 BIND(L_byteByByteProlog);
9610 andl(in2, 0x00000007);
9611 movl(tmp2, 1);
9612
9613 BIND(L_byteByByte);
9614 cmpl(tmp2, in2);
9615 jccb(Assembler::greater, L_exit);
9616 crc32(in_out, Address(in1, 0), 1);
9617 incq(in1);
9618 incl(tmp2);
9619 jmp(L_byteByByte);
9620
9621 BIND(L_exit);
9622 }
9623 #else
9624 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
9625 Register tmp1, Register tmp2, Register tmp3,
9626 Register tmp4, Register tmp5, Register tmp6,
9627 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
9628 bool is_pclmulqdq_supported) {
9629 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
9630 Label L_wordByWord;
9631 Label L_byteByByteProlog;
9632 Label L_byteByByte;
9633 Label L_exit;
9634
9635 if (is_pclmulqdq_supported) {
9636 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
9637 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1);
9638
9639 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
9640 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
9641
9642 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
9643 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
9644 } else {
9645 const_or_pre_comp_const_index[0] = 1;
9646 const_or_pre_comp_const_index[1] = 0;
9647
9648 const_or_pre_comp_const_index[2] = 3;
9649 const_or_pre_comp_const_index[3] = 2;
9650
9651 const_or_pre_comp_const_index[4] = 5;
9652 const_or_pre_comp_const_index[5] = 4;
9653 }
9654 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
9655 in2, in1, in_out,
9656 tmp1, tmp2, tmp3,
9657 w_xtmp1, w_xtmp2, w_xtmp3,
9658 tmp4, tmp5,
9659 tmp6);
9660 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
9661 in2, in1, in_out,
9662 tmp1, tmp2, tmp3,
9663 w_xtmp1, w_xtmp2, w_xtmp3,
9664 tmp4, tmp5,
9665 tmp6);
9666 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
9667 in2, in1, in_out,
9668 tmp1, tmp2, tmp3,
9669 w_xtmp1, w_xtmp2, w_xtmp3,
9670 tmp4, tmp5,
9671 tmp6);
9672 movl(tmp1, in2);
9673 andl(tmp1, 0x00000007);
9674 negl(tmp1);
9675 addl(tmp1, in2);
9676 addl(tmp1, in1);
9677
9678 BIND(L_wordByWord);
9679 cmpl(in1, tmp1);
9680 jcc(Assembler::greaterEqual, L_byteByByteProlog);
9681 crc32(in_out, Address(in1,0), 4);
9682 addl(in1, 4);
9683 jmp(L_wordByWord);
9684
9685 BIND(L_byteByByteProlog);
9686 andl(in2, 0x00000007);
9687 movl(tmp2, 1);
9688
9689 BIND(L_byteByByte);
9690 cmpl(tmp2, in2);
9691 jccb(Assembler::greater, L_exit);
9692 movb(tmp1, Address(in1, 0));
9693 crc32(in_out, tmp1, 1);
9694 incl(in1);
9695 incl(tmp2);
9696 jmp(L_byteByByte);
9697
9698 BIND(L_exit);
9699 }
9700 #endif // LP64
9701 #undef BIND
9702 #undef BLOCK_COMMENT
9703
9704 // Compress char[] array to byte[].
9705 // ..\jdk\src\java.base\share\classes\java\lang\StringUTF16.java
9706 // @HotSpotIntrinsicCandidate
9707 // private static int compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) {
9708 // for (int i = 0; i < len; i++) {
9709 // int c = src[srcOff++];
9710 // if (c >>> 8 != 0) {
9711 // return 0;
9712 // }
9713 // dst[dstOff++] = (byte)c;
9714 // }
9715 // return len;
9716 // }
9717 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
9718 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
9719 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
9720 Register tmp5, Register result) {
9721 Label copy_chars_loop, return_length, return_zero, done;
9722
9723 // rsi: src
9724 // rdi: dst
9725 // rdx: len
9726 // rcx: tmp5
9727 // rax: result
9728
9729 // rsi holds start addr of source char[] to be compressed
9730 // rdi holds start addr of destination byte[]
9731 // rdx holds length
9732
9733 assert(len != result, "");
9734
9735 // save length for return
9736 push(len);
9737
9738 if ((UseAVX > 2) && // AVX512
9739 VM_Version::supports_avx512vlbw() &&
9740 VM_Version::supports_bmi2()) {
9741
9742 Label copy_32_loop, copy_loop_tail, below_threshold;
9743
9744 // alignment
9745 Label post_alignment;
9746
9747 // if length of the string is less than 16, handle it in an old fashioned way
9748 testl(len, -32);
9749 jcc(Assembler::zero, below_threshold);
9750
9751 // First check whether a character is compressable ( <= 0xFF).
9752 // Create mask to test for Unicode chars inside zmm vector
9753 movl(result, 0x00FF);
9754 evpbroadcastw(tmp2Reg, result, Assembler::AVX_512bit);
9755
9756 testl(len, -64);
9757 jcc(Assembler::zero, post_alignment);
9758
9759 movl(tmp5, dst);
9760 andl(tmp5, (32 - 1));
9761 negl(tmp5);
9762 andl(tmp5, (32 - 1));
9763
9764 // bail out when there is nothing to be done
9765 testl(tmp5, 0xFFFFFFFF);
9766 jcc(Assembler::zero, post_alignment);
9767
9768 // ~(~0 << len), where len is the # of remaining elements to process
9769 movl(result, 0xFFFFFFFF);
9770 shlxl(result, result, tmp5);
9771 notl(result);
9772 kmovdl(k3, result);
9773
9774 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit);
9775 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9776 ktestd(k2, k3);
9777 jcc(Assembler::carryClear, return_zero);
9778
9779 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit);
9780
9781 addptr(src, tmp5);
9782 addptr(src, tmp5);
9783 addptr(dst, tmp5);
9784 subl(len, tmp5);
9785
9786 bind(post_alignment);
9787 // end of alignment
9788
9789 movl(tmp5, len);
9790 andl(tmp5, (32 - 1)); // tail count (in chars)
9791 andl(len, ~(32 - 1)); // vector count (in chars)
9792 jcc(Assembler::zero, copy_loop_tail);
9793
9794 lea(src, Address(src, len, Address::times_2));
9795 lea(dst, Address(dst, len, Address::times_1));
9796 negptr(len);
9797
9798 bind(copy_32_loop);
9799 evmovdquw(tmp1Reg, Address(src, len, Address::times_2), Assembler::AVX_512bit);
9800 evpcmpuw(k2, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9801 kortestdl(k2, k2);
9802 jcc(Assembler::carryClear, return_zero);
9803
9804 // All elements in current processed chunk are valid candidates for
9805 // compression. Write a truncated byte elements to the memory.
9806 evpmovwb(Address(dst, len, Address::times_1), tmp1Reg, Assembler::AVX_512bit);
9807 addptr(len, 32);
9808 jcc(Assembler::notZero, copy_32_loop);
9809
9810 bind(copy_loop_tail);
9811 // bail out when there is nothing to be done
9812 testl(tmp5, 0xFFFFFFFF);
9813 jcc(Assembler::zero, return_length);
9814
9815 movl(len, tmp5);
9816
9817 // ~(~0 << len), where len is the # of remaining elements to process
9818 movl(result, 0xFFFFFFFF);
9819 shlxl(result, result, len);
9820 notl(result);
9821
9822 kmovdl(k3, result);
9823
9824 evmovdquw(tmp1Reg, k3, Address(src, 0), Assembler::AVX_512bit);
9825 evpcmpuw(k2, k3, tmp1Reg, tmp2Reg, Assembler::le, Assembler::AVX_512bit);
9826 ktestd(k2, k3);
9827 jcc(Assembler::carryClear, return_zero);
9828
9829 evpmovwb(Address(dst, 0), k3, tmp1Reg, Assembler::AVX_512bit);
9830 jmp(return_length);
9831
9832 bind(below_threshold);
9833 }
9834
9835 if (UseSSE42Intrinsics) {
9836 Label copy_32_loop, copy_16, copy_tail;
9837
9838 movl(result, len);
9839
9840 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors
9841
9842 // vectored compression
9843 andl(len, 0xfffffff0); // vector count (in chars)
9844 andl(result, 0x0000000f); // tail count (in chars)
9845 testl(len, len);
9846 jcc(Assembler::zero, copy_16);
9847
9848 // compress 16 chars per iter
9849 movdl(tmp1Reg, tmp5);
9850 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
9851 pxor(tmp4Reg, tmp4Reg);
9852
9853 lea(src, Address(src, len, Address::times_2));
9854 lea(dst, Address(dst, len, Address::times_1));
9855 negptr(len);
9856
9857 bind(copy_32_loop);
9858 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters
9859 por(tmp4Reg, tmp2Reg);
9860 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters
9861 por(tmp4Reg, tmp3Reg);
9862 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector
9863 jcc(Assembler::notZero, return_zero);
9864 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte
9865 movdqu(Address(dst, len, Address::times_1), tmp2Reg);
9866 addptr(len, 16);
9867 jcc(Assembler::notZero, copy_32_loop);
9868
9869 // compress next vector of 8 chars (if any)
9870 bind(copy_16);
9871 movl(len, result);
9872 andl(len, 0xfffffff8); // vector count (in chars)
9873 andl(result, 0x00000007); // tail count (in chars)
9874 testl(len, len);
9875 jccb(Assembler::zero, copy_tail);
9876
9877 movdl(tmp1Reg, tmp5);
9878 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
9879 pxor(tmp3Reg, tmp3Reg);
9880
9881 movdqu(tmp2Reg, Address(src, 0));
9882 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
9883 jccb(Assembler::notZero, return_zero);
9884 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte
9885 movq(Address(dst, 0), tmp2Reg);
9886 addptr(src, 16);
9887 addptr(dst, 8);
9888
9889 bind(copy_tail);
9890 movl(len, result);
9891 }
9892 // compress 1 char per iter
9893 testl(len, len);
9894 jccb(Assembler::zero, return_length);
9895 lea(src, Address(src, len, Address::times_2));
9896 lea(dst, Address(dst, len, Address::times_1));
9897 negptr(len);
9898
9899 bind(copy_chars_loop);
9900 load_unsigned_short(result, Address(src, len, Address::times_2));
9901 testl(result, 0xff00); // check if Unicode char
9902 jccb(Assembler::notZero, return_zero);
9903 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte
9904 increment(len);
9905 jcc(Assembler::notZero, copy_chars_loop);
9906
9907 // if compression succeeded, return length
9908 bind(return_length);
9909 pop(result);
9910 jmpb(done);
9911
9912 // if compression failed, return 0
9913 bind(return_zero);
9914 xorl(result, result);
9915 addptr(rsp, wordSize);
9916
9917 bind(done);
9918 }
9919
9920 // Inflate byte[] array to char[].
9921 // ..\jdk\src\java.base\share\classes\java\lang\StringLatin1.java
9922 // @HotSpotIntrinsicCandidate
9923 // private static void inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) {
9924 // for (int i = 0; i < len; i++) {
9925 // dst[dstOff++] = (char)(src[srcOff++] & 0xff);
9926 // }
9927 // }
9928 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
9929 XMMRegister tmp1, Register tmp2) {
9930 Label copy_chars_loop, done, below_threshold;
9931 // rsi: src
9932 // rdi: dst
9933 // rdx: len
9934 // rcx: tmp2
9935
9936 // rsi holds start addr of source byte[] to be inflated
9937 // rdi holds start addr of destination char[]
9938 // rdx holds length
9939 assert_different_registers(src, dst, len, tmp2);
9940
9941 if ((UseAVX > 2) && // AVX512
9942 VM_Version::supports_avx512vlbw() &&
9943 VM_Version::supports_bmi2()) {
9944
9945 Label copy_32_loop, copy_tail;
9946 Register tmp3_aliased = len;
9947
9948 // if length of the string is less than 16, handle it in an old fashioned way
9949 testl(len, -16);
9950 jcc(Assembler::zero, below_threshold);
9951
9952 // In order to use only one arithmetic operation for the main loop we use
9953 // this pre-calculation
9954 movl(tmp2, len);
9955 andl(tmp2, (32 - 1)); // tail count (in chars), 32 element wide loop
9956 andl(len, -32); // vector count
9957 jccb(Assembler::zero, copy_tail);
9958
9959 lea(src, Address(src, len, Address::times_1));
9960 lea(dst, Address(dst, len, Address::times_2));
9961 negptr(len);
9962
9963
9964 // inflate 32 chars per iter
9965 bind(copy_32_loop);
9966 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_512bit);
9967 evmovdquw(Address(dst, len, Address::times_2), tmp1, Assembler::AVX_512bit);
9968 addptr(len, 32);
9969 jcc(Assembler::notZero, copy_32_loop);
9970
9971 bind(copy_tail);
9972 // bail out when there is nothing to be done
9973 testl(tmp2, -1); // we don't destroy the contents of tmp2 here
9974 jcc(Assembler::zero, done);
9975
9976 // ~(~0 << length), where length is the # of remaining elements to process
9977 movl(tmp3_aliased, -1);
9978 shlxl(tmp3_aliased, tmp3_aliased, tmp2);
9979 notl(tmp3_aliased);
9980 kmovdl(k2, tmp3_aliased);
9981 evpmovzxbw(tmp1, k2, Address(src, 0), Assembler::AVX_512bit);
9982 evmovdquw(Address(dst, 0), k2, tmp1, Assembler::AVX_512bit);
9983
9984 jmp(done);
9985 }
9986 if (UseSSE42Intrinsics) {
9987 Label copy_16_loop, copy_8_loop, copy_bytes, copy_new_tail, copy_tail;
9988
9989 movl(tmp2, len);
9990
9991 if (UseAVX > 1) {
9992 andl(tmp2, (16 - 1));
9993 andl(len, -16);
9994 jccb(Assembler::zero, copy_new_tail);
9995 } else {
9996 andl(tmp2, 0x00000007); // tail count (in chars)
9997 andl(len, 0xfffffff8); // vector count (in chars)
9998 jccb(Assembler::zero, copy_tail);
9999 }
10000
10001 // vectored inflation
10002 lea(src, Address(src, len, Address::times_1));
10003 lea(dst, Address(dst, len, Address::times_2));
10004 negptr(len);
10005
10006 if (UseAVX > 1) {
10007 bind(copy_16_loop);
10008 vpmovzxbw(tmp1, Address(src, len, Address::times_1), Assembler::AVX_256bit);
10009 vmovdqu(Address(dst, len, Address::times_2), tmp1);
10010 addptr(len, 16);
10011 jcc(Assembler::notZero, copy_16_loop);
10012
10013 bind(below_threshold);
10014 bind(copy_new_tail);
10015 if ((UseAVX > 2) &&
10016 VM_Version::supports_avx512vlbw() &&
10017 VM_Version::supports_bmi2()) {
10018 movl(tmp2, len);
10019 } else {
10020 movl(len, tmp2);
10021 }
10022 andl(tmp2, 0x00000007);
10023 andl(len, 0xFFFFFFF8);
10024 jccb(Assembler::zero, copy_tail);
10025
10026 pmovzxbw(tmp1, Address(src, 0));
10027 movdqu(Address(dst, 0), tmp1);
10028 addptr(src, 8);
10029 addptr(dst, 2 * 8);
10030
10031 jmp(copy_tail, true);
10032 }
10033
10034 // inflate 8 chars per iter
10035 bind(copy_8_loop);
10036 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words
10037 movdqu(Address(dst, len, Address::times_2), tmp1);
10038 addptr(len, 8);
10039 jcc(Assembler::notZero, copy_8_loop);
10040
10041 bind(copy_tail);
10042 movl(len, tmp2);
10043
10044 cmpl(len, 4);
10045 jccb(Assembler::less, copy_bytes);
10046
10047 movdl(tmp1, Address(src, 0)); // load 4 byte chars
10048 pmovzxbw(tmp1, tmp1);
10049 movq(Address(dst, 0), tmp1);
10050 subptr(len, 4);
10051 addptr(src, 4);
10052 addptr(dst, 8);
10053
10054 bind(copy_bytes);
10055 } else {
10056 bind(below_threshold);
10057 }
10058
10059 testl(len, len);
10060 jccb(Assembler::zero, done);
10061 lea(src, Address(src, len, Address::times_1));
10062 lea(dst, Address(dst, len, Address::times_2));
10063 negptr(len);
10064
10065 // inflate 1 char per iter
10066 bind(copy_chars_loop);
10067 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char
10068 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word
10069 increment(len);
10070 jcc(Assembler::notZero, copy_chars_loop);
10071
10072 bind(done);
10073 }
10074
10075 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
10076 switch (cond) {
10077 // Note some conditions are synonyms for others
10078 case Assembler::zero: return Assembler::notZero;
10079 case Assembler::notZero: return Assembler::zero;
10080 case Assembler::less: return Assembler::greaterEqual;
10081 case Assembler::lessEqual: return Assembler::greater;
10082 case Assembler::greater: return Assembler::lessEqual;
10083 case Assembler::greaterEqual: return Assembler::less;
10084 case Assembler::below: return Assembler::aboveEqual;
10085 case Assembler::belowEqual: return Assembler::above;
10086 case Assembler::above: return Assembler::belowEqual;
10087 case Assembler::aboveEqual: return Assembler::below;
10088 case Assembler::overflow: return Assembler::noOverflow;
10089 case Assembler::noOverflow: return Assembler::overflow;
10090 case Assembler::negative: return Assembler::positive;
10091 case Assembler::positive: return Assembler::negative;
10092 case Assembler::parity: return Assembler::noParity;
10093 case Assembler::noParity: return Assembler::parity;
10094 }
10095 ShouldNotReachHere(); return Assembler::overflow;
10096 }
10097
10098 SkipIfEqual::SkipIfEqual(
10099 MacroAssembler* masm, const bool* flag_addr, bool value) {
10100 _masm = masm;
10101 _masm->cmp8(ExternalAddress((address)flag_addr), value);
10102 _masm->jcc(Assembler::equal, _label);
10103 }
10104
10105 SkipIfEqual::~SkipIfEqual() {
10106 _masm->bind(_label);
10107 }
10108
10109 // 32-bit Windows has its own fast-path implementation
10110 // of get_thread
10111 #if !defined(WIN32) || defined(_LP64)
10112
10113 // This is simply a call to Thread::current()
10114 void MacroAssembler::get_thread(Register thread) {
10115 if (thread != rax) {
10116 push(rax);
10117 }
10118 LP64_ONLY(push(rdi);)
10119 LP64_ONLY(push(rsi);)
10120 push(rdx);
10121 push(rcx);
10122 #ifdef _LP64
10123 push(r8);
10124 push(r9);
10125 push(r10);
10126 push(r11);
10127 #endif
10128
10129 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, Thread::current), 0);
10130
10131 #ifdef _LP64
10132 pop(r11);
10133 pop(r10);
10134 pop(r9);
10135 pop(r8);
10136 #endif
10137 pop(rcx);
10138 pop(rdx);
10139 LP64_ONLY(pop(rsi);)
10140 LP64_ONLY(pop(rdi);)
10141 if (thread != rax) {
10142 mov(thread, rax);
10143 pop(rax);
10144 }
10145 }
10146
10147 #endif