1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "asm/assembler.inline.hpp"
28 #include "compiler/disassembler.hpp"
29 #include "gc/shared/cardTableModRefBS.hpp"
30 #include "gc/shared/collectedHeap.inline.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "memory/universe.hpp"
34 #include "oops/klass.inline.hpp"
35 #include "prims/methodHandles.hpp"
36 #include "runtime/biasedLocking.hpp"
37 #include "runtime/interfaceSupport.hpp"
38 #include "runtime/objectMonitor.hpp"
39 #include "runtime/os.hpp"
40 #include "runtime/sharedRuntime.hpp"
41 #include "runtime/stubRoutines.hpp"
42 #include "utilities/macros.hpp"
43 #if INCLUDE_ALL_GCS
44 #include "gc/g1/g1CollectedHeap.inline.hpp"
45 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
46 #include "gc/g1/heapRegion.hpp"
47 #endif // INCLUDE_ALL_GCS
48 #include "crc32c.h"
49 #ifdef COMPILER2
50 #include "opto/intrinsicnode.hpp"
51 #endif
52
53 #ifdef PRODUCT
54 #define BLOCK_COMMENT(str) /* nothing */
55 #define STOP(error) stop(error)
56 #else
57 #define BLOCK_COMMENT(str) block_comment(str)
58 #define STOP(error) block_comment(error); stop(error)
59 #endif
60
61 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
62
63 #ifdef ASSERT
64 bool AbstractAssembler::pd_check_instruction_mark() { return true; }
65 #endif
66
67 static Assembler::Condition reverse[] = {
68 Assembler::noOverflow /* overflow = 0x0 */ ,
69 Assembler::overflow /* noOverflow = 0x1 */ ,
70 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ ,
71 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ ,
72 Assembler::notZero /* zero = 0x4, equal = 0x4 */ ,
73 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ ,
74 Assembler::above /* belowEqual = 0x6 */ ,
75 Assembler::belowEqual /* above = 0x7 */ ,
76 Assembler::positive /* negative = 0x8 */ ,
77 Assembler::negative /* positive = 0x9 */ ,
78 Assembler::noParity /* parity = 0xa */ ,
79 Assembler::parity /* noParity = 0xb */ ,
80 Assembler::greaterEqual /* less = 0xc */ ,
81 Assembler::less /* greaterEqual = 0xd */ ,
82 Assembler::greater /* lessEqual = 0xe */ ,
83 Assembler::lessEqual /* greater = 0xf, */
84
85 };
86
87
88 // Implementation of MacroAssembler
89
90 // First all the versions that have distinct versions depending on 32/64 bit
91 // Unless the difference is trivial (1 line or so).
92
93 #ifndef _LP64
94
95 // 32bit versions
96
97 Address MacroAssembler::as_Address(AddressLiteral adr) {
98 return Address(adr.target(), adr.rspec());
99 }
100
101 Address MacroAssembler::as_Address(ArrayAddress adr) {
102 return Address::make_array(adr);
103 }
104
105 void MacroAssembler::call_VM_leaf_base(address entry_point,
106 int number_of_arguments) {
107 call(RuntimeAddress(entry_point));
108 increment(rsp, number_of_arguments * wordSize);
109 }
110
111 void MacroAssembler::cmpklass(Address src1, Metadata* obj) {
112 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
113 }
114
115 void MacroAssembler::cmpklass(Register src1, Metadata* obj) {
116 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate());
117 }
118
119 void MacroAssembler::cmpoop(Address src1, jobject obj) {
120 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
121 }
122
123 void MacroAssembler::cmpoop(Register src1, jobject obj) {
124 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate());
125 }
126
127 void MacroAssembler::extend_sign(Register hi, Register lo) {
128 // According to Intel Doc. AP-526, "Integer Divide", p.18.
129 if (VM_Version::is_P6() && hi == rdx && lo == rax) {
130 cdql();
131 } else {
132 movl(hi, lo);
133 sarl(hi, 31);
134 }
135 }
136
137 void MacroAssembler::jC2(Register tmp, Label& L) {
138 // set parity bit if FPU flag C2 is set (via rax)
139 save_rax(tmp);
140 fwait(); fnstsw_ax();
141 sahf();
142 restore_rax(tmp);
143 // branch
144 jcc(Assembler::parity, L);
145 }
146
147 void MacroAssembler::jnC2(Register tmp, Label& L) {
148 // set parity bit if FPU flag C2 is set (via rax)
149 save_rax(tmp);
150 fwait(); fnstsw_ax();
151 sahf();
152 restore_rax(tmp);
153 // branch
154 jcc(Assembler::noParity, L);
155 }
156
157 // 32bit can do a case table jump in one instruction but we no longer allow the base
158 // to be installed in the Address class
159 void MacroAssembler::jump(ArrayAddress entry) {
160 jmp(as_Address(entry));
161 }
162
163 // Note: y_lo will be destroyed
164 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
165 // Long compare for Java (semantics as described in JVM spec.)
166 Label high, low, done;
167
168 cmpl(x_hi, y_hi);
169 jcc(Assembler::less, low);
170 jcc(Assembler::greater, high);
171 // x_hi is the return register
172 xorl(x_hi, x_hi);
173 cmpl(x_lo, y_lo);
174 jcc(Assembler::below, low);
175 jcc(Assembler::equal, done);
176
177 bind(high);
178 xorl(x_hi, x_hi);
179 increment(x_hi);
180 jmp(done);
181
182 bind(low);
183 xorl(x_hi, x_hi);
184 decrementl(x_hi);
185
186 bind(done);
187 }
188
189 void MacroAssembler::lea(Register dst, AddressLiteral src) {
190 mov_literal32(dst, (int32_t)src.target(), src.rspec());
191 }
192
193 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
194 // leal(dst, as_Address(adr));
195 // see note in movl as to why we must use a move
196 mov_literal32(dst, (int32_t) adr.target(), adr.rspec());
197 }
198
199 void MacroAssembler::leave() {
200 mov(rsp, rbp);
201 pop(rbp);
202 }
203
204 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) {
205 // Multiplication of two Java long values stored on the stack
206 // as illustrated below. Result is in rdx:rax.
207 //
208 // rsp ---> [ ?? ] \ \
209 // .... | y_rsp_offset |
210 // [ y_lo ] / (in bytes) | x_rsp_offset
211 // [ y_hi ] | (in bytes)
212 // .... |
213 // [ x_lo ] /
214 // [ x_hi ]
215 // ....
216 //
217 // Basic idea: lo(result) = lo(x_lo * y_lo)
218 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi)
219 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset);
220 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset);
221 Label quick;
222 // load x_hi, y_hi and check if quick
223 // multiplication is possible
224 movl(rbx, x_hi);
225 movl(rcx, y_hi);
226 movl(rax, rbx);
227 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0
228 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply
229 // do full multiplication
230 // 1st step
231 mull(y_lo); // x_hi * y_lo
232 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx,
233 // 2nd step
234 movl(rax, x_lo);
235 mull(rcx); // x_lo * y_hi
236 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx,
237 // 3rd step
238 bind(quick); // note: rbx, = 0 if quick multiply!
239 movl(rax, x_lo);
240 mull(y_lo); // x_lo * y_lo
241 addl(rdx, rbx); // correct hi(x_lo * y_lo)
242 }
243
244 void MacroAssembler::lneg(Register hi, Register lo) {
245 negl(lo);
246 adcl(hi, 0);
247 negl(hi);
248 }
249
250 void MacroAssembler::lshl(Register hi, Register lo) {
251 // Java shift left long support (semantics as described in JVM spec., p.305)
252 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n))
253 // shift value is in rcx !
254 assert(hi != rcx, "must not use rcx");
255 assert(lo != rcx, "must not use rcx");
256 const Register s = rcx; // shift count
257 const int n = BitsPerWord;
258 Label L;
259 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
260 cmpl(s, n); // if (s < n)
261 jcc(Assembler::less, L); // else (s >= n)
262 movl(hi, lo); // x := x << n
263 xorl(lo, lo);
264 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
265 bind(L); // s (mod n) < n
266 shldl(hi, lo); // x := x << s
267 shll(lo);
268 }
269
270
271 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) {
272 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310)
273 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n))
274 assert(hi != rcx, "must not use rcx");
275 assert(lo != rcx, "must not use rcx");
276 const Register s = rcx; // shift count
277 const int n = BitsPerWord;
278 Label L;
279 andl(s, 0x3f); // s := s & 0x3f (s < 0x40)
280 cmpl(s, n); // if (s < n)
281 jcc(Assembler::less, L); // else (s >= n)
282 movl(lo, hi); // x := x >> n
283 if (sign_extension) sarl(hi, 31);
284 else xorl(hi, hi);
285 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n!
286 bind(L); // s (mod n) < n
287 shrdl(lo, hi); // x := x >> s
288 if (sign_extension) sarl(hi);
289 else shrl(hi);
290 }
291
292 void MacroAssembler::movoop(Register dst, jobject obj) {
293 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
294 }
295
296 void MacroAssembler::movoop(Address dst, jobject obj) {
297 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate());
298 }
299
300 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
301 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
302 }
303
304 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
305 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate());
306 }
307
308 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
309 // scratch register is not used,
310 // it is defined to match parameters of 64-bit version of this method.
311 if (src.is_lval()) {
312 mov_literal32(dst, (intptr_t)src.target(), src.rspec());
313 } else {
314 movl(dst, as_Address(src));
315 }
316 }
317
318 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
319 movl(as_Address(dst), src);
320 }
321
322 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
323 movl(dst, as_Address(src));
324 }
325
326 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
327 void MacroAssembler::movptr(Address dst, intptr_t src) {
328 movl(dst, src);
329 }
330
331
332 void MacroAssembler::pop_callee_saved_registers() {
333 pop(rcx);
334 pop(rdx);
335 pop(rdi);
336 pop(rsi);
337 }
338
339 void MacroAssembler::pop_fTOS() {
340 fld_d(Address(rsp, 0));
341 addl(rsp, 2 * wordSize);
342 }
343
344 void MacroAssembler::push_callee_saved_registers() {
345 push(rsi);
346 push(rdi);
347 push(rdx);
348 push(rcx);
349 }
350
351 void MacroAssembler::push_fTOS() {
352 subl(rsp, 2 * wordSize);
353 fstp_d(Address(rsp, 0));
354 }
355
356
357 void MacroAssembler::pushoop(jobject obj) {
358 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate());
359 }
360
361 void MacroAssembler::pushklass(Metadata* obj) {
362 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate());
363 }
364
365 void MacroAssembler::pushptr(AddressLiteral src) {
366 if (src.is_lval()) {
367 push_literal32((int32_t)src.target(), src.rspec());
368 } else {
369 pushl(as_Address(src));
370 }
371 }
372
373 void MacroAssembler::set_word_if_not_zero(Register dst) {
374 xorl(dst, dst);
375 set_byte_if_not_zero(dst);
376 }
377
378 static void pass_arg0(MacroAssembler* masm, Register arg) {
379 masm->push(arg);
380 }
381
382 static void pass_arg1(MacroAssembler* masm, Register arg) {
383 masm->push(arg);
384 }
385
386 static void pass_arg2(MacroAssembler* masm, Register arg) {
387 masm->push(arg);
388 }
389
390 static void pass_arg3(MacroAssembler* masm, Register arg) {
391 masm->push(arg);
392 }
393
394 #ifndef PRODUCT
395 extern "C" void findpc(intptr_t x);
396 #endif
397
398 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) {
399 // In order to get locks to work, we need to fake a in_VM state
400 JavaThread* thread = JavaThread::current();
401 JavaThreadState saved_state = thread->thread_state();
402 thread->set_thread_state(_thread_in_vm);
403 if (ShowMessageBoxOnError) {
404 JavaThread* thread = JavaThread::current();
405 JavaThreadState saved_state = thread->thread_state();
406 thread->set_thread_state(_thread_in_vm);
407 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
408 ttyLocker ttyl;
409 BytecodeCounter::print();
410 }
411 // To see where a verify_oop failed, get $ebx+40/X for this frame.
412 // This is the value of eip which points to where verify_oop will return.
413 if (os::message_box(msg, "Execution stopped, print registers?")) {
414 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip);
415 BREAKPOINT;
416 }
417 } else {
418 ttyLocker ttyl;
419 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg);
420 }
421 // Don't assert holding the ttyLock
422 assert(false, "DEBUG MESSAGE: %s", msg);
423 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
424 }
425
426 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) {
427 ttyLocker ttyl;
428 FlagSetting fs(Debugging, true);
429 tty->print_cr("eip = 0x%08x", eip);
430 #ifndef PRODUCT
431 if ((WizardMode || Verbose) && PrintMiscellaneous) {
432 tty->cr();
433 findpc(eip);
434 tty->cr();
435 }
436 #endif
437 #define PRINT_REG(rax) \
438 { tty->print("%s = ", #rax); os::print_location(tty, rax); }
439 PRINT_REG(rax);
440 PRINT_REG(rbx);
441 PRINT_REG(rcx);
442 PRINT_REG(rdx);
443 PRINT_REG(rdi);
444 PRINT_REG(rsi);
445 PRINT_REG(rbp);
446 PRINT_REG(rsp);
447 #undef PRINT_REG
448 // Print some words near top of staack.
449 int* dump_sp = (int*) rsp;
450 for (int col1 = 0; col1 < 8; col1++) {
451 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
452 os::print_location(tty, *dump_sp++);
453 }
454 for (int row = 0; row < 16; row++) {
455 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp);
456 for (int col = 0; col < 8; col++) {
457 tty->print(" 0x%08x", *dump_sp++);
458 }
459 tty->cr();
460 }
461 // Print some instructions around pc:
462 Disassembler::decode((address)eip-64, (address)eip);
463 tty->print_cr("--------");
464 Disassembler::decode((address)eip, (address)eip+32);
465 }
466
467 void MacroAssembler::stop(const char* msg) {
468 ExternalAddress message((address)msg);
469 // push address of message
470 pushptr(message.addr());
471 { Label L; call(L, relocInfo::none); bind(L); } // push eip
472 pusha(); // push registers
473 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32)));
474 hlt();
475 }
476
477 void MacroAssembler::warn(const char* msg) {
478 push_CPU_state();
479
480 ExternalAddress message((address) msg);
481 // push address of message
482 pushptr(message.addr());
483
484 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning)));
485 addl(rsp, wordSize); // discard argument
486 pop_CPU_state();
487 }
488
489 void MacroAssembler::print_state() {
490 { Label L; call(L, relocInfo::none); bind(L); } // push eip
491 pusha(); // push registers
492
493 push_CPU_state();
494 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32)));
495 pop_CPU_state();
496
497 popa();
498 addl(rsp, wordSize);
499 }
500
501 #else // _LP64
502
503 // 64 bit versions
504
505 Address MacroAssembler::as_Address(AddressLiteral adr) {
506 // amd64 always does this as a pc-rel
507 // we can be absolute or disp based on the instruction type
508 // jmp/call are displacements others are absolute
509 assert(!adr.is_lval(), "must be rval");
510 assert(reachable(adr), "must be");
511 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc());
512
513 }
514
515 Address MacroAssembler::as_Address(ArrayAddress adr) {
516 AddressLiteral base = adr.base();
517 lea(rscratch1, base);
518 Address index = adr.index();
519 assert(index._disp == 0, "must not have disp"); // maybe it can?
520 Address array(rscratch1, index._index, index._scale, index._disp);
521 return array;
522 }
523
524 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) {
525 Label L, E;
526
527 #ifdef _WIN64
528 // Windows always allocates space for it's register args
529 assert(num_args <= 4, "only register arguments supported");
530 subq(rsp, frame::arg_reg_save_area_bytes);
531 #endif
532
533 // Align stack if necessary
534 testl(rsp, 15);
535 jcc(Assembler::zero, L);
536
537 subq(rsp, 8);
538 {
539 call(RuntimeAddress(entry_point));
540 }
541 addq(rsp, 8);
542 jmp(E);
543
544 bind(L);
545 {
546 call(RuntimeAddress(entry_point));
547 }
548
549 bind(E);
550
551 #ifdef _WIN64
552 // restore stack pointer
553 addq(rsp, frame::arg_reg_save_area_bytes);
554 #endif
555
556 }
557
558 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) {
559 assert(!src2.is_lval(), "should use cmpptr");
560
561 if (reachable(src2)) {
562 cmpq(src1, as_Address(src2));
563 } else {
564 lea(rscratch1, src2);
565 Assembler::cmpq(src1, Address(rscratch1, 0));
566 }
567 }
568
569 int MacroAssembler::corrected_idivq(Register reg) {
570 // Full implementation of Java ldiv and lrem; checks for special
571 // case as described in JVM spec., p.243 & p.271. The function
572 // returns the (pc) offset of the idivl instruction - may be needed
573 // for implicit exceptions.
574 //
575 // normal case special case
576 //
577 // input : rax: dividend min_long
578 // reg: divisor (may not be eax/edx) -1
579 //
580 // output: rax: quotient (= rax idiv reg) min_long
581 // rdx: remainder (= rax irem reg) 0
582 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register");
583 static const int64_t min_long = 0x8000000000000000;
584 Label normal_case, special_case;
585
586 // check for special case
587 cmp64(rax, ExternalAddress((address) &min_long));
588 jcc(Assembler::notEqual, normal_case);
589 xorl(rdx, rdx); // prepare rdx for possible special case (where
590 // remainder = 0)
591 cmpq(reg, -1);
592 jcc(Assembler::equal, special_case);
593
594 // handle normal case
595 bind(normal_case);
596 cdqq();
597 int idivq_offset = offset();
598 idivq(reg);
599
600 // normal and special case exit
601 bind(special_case);
602
603 return idivq_offset;
604 }
605
606 void MacroAssembler::decrementq(Register reg, int value) {
607 if (value == min_jint) { subq(reg, value); return; }
608 if (value < 0) { incrementq(reg, -value); return; }
609 if (value == 0) { ; return; }
610 if (value == 1 && UseIncDec) { decq(reg) ; return; }
611 /* else */ { subq(reg, value) ; return; }
612 }
613
614 void MacroAssembler::decrementq(Address dst, int value) {
615 if (value == min_jint) { subq(dst, value); return; }
616 if (value < 0) { incrementq(dst, -value); return; }
617 if (value == 0) { ; return; }
618 if (value == 1 && UseIncDec) { decq(dst) ; return; }
619 /* else */ { subq(dst, value) ; return; }
620 }
621
622 void MacroAssembler::incrementq(AddressLiteral dst) {
623 if (reachable(dst)) {
624 incrementq(as_Address(dst));
625 } else {
626 lea(rscratch1, dst);
627 incrementq(Address(rscratch1, 0));
628 }
629 }
630
631 void MacroAssembler::incrementq(Register reg, int value) {
632 if (value == min_jint) { addq(reg, value); return; }
633 if (value < 0) { decrementq(reg, -value); return; }
634 if (value == 0) { ; return; }
635 if (value == 1 && UseIncDec) { incq(reg) ; return; }
636 /* else */ { addq(reg, value) ; return; }
637 }
638
639 void MacroAssembler::incrementq(Address dst, int value) {
640 if (value == min_jint) { addq(dst, value); return; }
641 if (value < 0) { decrementq(dst, -value); return; }
642 if (value == 0) { ; return; }
643 if (value == 1 && UseIncDec) { incq(dst) ; return; }
644 /* else */ { addq(dst, value) ; return; }
645 }
646
647 // 32bit can do a case table jump in one instruction but we no longer allow the base
648 // to be installed in the Address class
649 void MacroAssembler::jump(ArrayAddress entry) {
650 lea(rscratch1, entry.base());
651 Address dispatch = entry.index();
652 assert(dispatch._base == noreg, "must be");
653 dispatch._base = rscratch1;
654 jmp(dispatch);
655 }
656
657 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) {
658 ShouldNotReachHere(); // 64bit doesn't use two regs
659 cmpq(x_lo, y_lo);
660 }
661
662 void MacroAssembler::lea(Register dst, AddressLiteral src) {
663 mov_literal64(dst, (intptr_t)src.target(), src.rspec());
664 }
665
666 void MacroAssembler::lea(Address dst, AddressLiteral adr) {
667 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec());
668 movptr(dst, rscratch1);
669 }
670
671 void MacroAssembler::leave() {
672 // %%% is this really better? Why not on 32bit too?
673 emit_int8((unsigned char)0xC9); // LEAVE
674 }
675
676 void MacroAssembler::lneg(Register hi, Register lo) {
677 ShouldNotReachHere(); // 64bit doesn't use two regs
678 negq(lo);
679 }
680
681 void MacroAssembler::movoop(Register dst, jobject obj) {
682 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate());
683 }
684
685 void MacroAssembler::movoop(Address dst, jobject obj) {
686 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate());
687 movq(dst, rscratch1);
688 }
689
690 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
691 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
692 }
693
694 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) {
695 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate());
696 movq(dst, rscratch1);
697 }
698
699 void MacroAssembler::movptr(Register dst, AddressLiteral src, Register scratch) {
700 if (src.is_lval()) {
701 mov_literal64(dst, (intptr_t)src.target(), src.rspec());
702 } else {
703 if (reachable(src)) {
704 movq(dst, as_Address(src));
705 } else {
706 lea(scratch, src);
707 movq(dst, Address(scratch, 0));
708 }
709 }
710 }
711
712 void MacroAssembler::movptr(ArrayAddress dst, Register src) {
713 movq(as_Address(dst), src);
714 }
715
716 void MacroAssembler::movptr(Register dst, ArrayAddress src) {
717 movq(dst, as_Address(src));
718 }
719
720 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
721 void MacroAssembler::movptr(Address dst, intptr_t src) {
722 mov64(rscratch1, src);
723 movq(dst, rscratch1);
724 }
725
726 // These are mostly for initializing NULL
727 void MacroAssembler::movptr(Address dst, int32_t src) {
728 movslq(dst, src);
729 }
730
731 void MacroAssembler::movptr(Register dst, int32_t src) {
732 mov64(dst, (intptr_t)src);
733 }
734
735 void MacroAssembler::pushoop(jobject obj) {
736 movoop(rscratch1, obj);
737 push(rscratch1);
738 }
739
740 void MacroAssembler::pushklass(Metadata* obj) {
741 mov_metadata(rscratch1, obj);
742 push(rscratch1);
743 }
744
745 void MacroAssembler::pushptr(AddressLiteral src) {
746 lea(rscratch1, src);
747 if (src.is_lval()) {
748 push(rscratch1);
749 } else {
750 pushq(Address(rscratch1, 0));
751 }
752 }
753
754 void MacroAssembler::reset_last_Java_frame(bool clear_fp,
755 bool clear_pc) {
756 // we must set sp to zero to clear frame
757 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
758 // must clear fp, so that compiled frames are not confused; it is
759 // possible that we need it only for debugging
760 if (clear_fp) {
761 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
762 }
763
764 if (clear_pc) {
765 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
766 }
767 }
768
769 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
770 Register last_java_fp,
771 address last_java_pc) {
772 // determine last_java_sp register
773 if (!last_java_sp->is_valid()) {
774 last_java_sp = rsp;
775 }
776
777 // last_java_fp is optional
778 if (last_java_fp->is_valid()) {
779 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()),
780 last_java_fp);
781 }
782
783 // last_java_pc is optional
784 if (last_java_pc != NULL) {
785 Address java_pc(r15_thread,
786 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset());
787 lea(rscratch1, InternalAddress(last_java_pc));
788 movptr(java_pc, rscratch1);
789 }
790
791 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
792 }
793
794 static void pass_arg0(MacroAssembler* masm, Register arg) {
795 if (c_rarg0 != arg ) {
796 masm->mov(c_rarg0, arg);
797 }
798 }
799
800 static void pass_arg1(MacroAssembler* masm, Register arg) {
801 if (c_rarg1 != arg ) {
802 masm->mov(c_rarg1, arg);
803 }
804 }
805
806 static void pass_arg2(MacroAssembler* masm, Register arg) {
807 if (c_rarg2 != arg ) {
808 masm->mov(c_rarg2, arg);
809 }
810 }
811
812 static void pass_arg3(MacroAssembler* masm, Register arg) {
813 if (c_rarg3 != arg ) {
814 masm->mov(c_rarg3, arg);
815 }
816 }
817
818 void MacroAssembler::stop(const char* msg) {
819 address rip = pc();
820 pusha(); // get regs on stack
821 lea(c_rarg0, ExternalAddress((address) msg));
822 lea(c_rarg1, InternalAddress(rip));
823 movq(c_rarg2, rsp); // pass pointer to regs array
824 andq(rsp, -16); // align stack as required by ABI
825 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
826 hlt();
827 }
828
829 void MacroAssembler::warn(const char* msg) {
830 push(rbp);
831 movq(rbp, rsp);
832 andq(rsp, -16); // align stack as required by push_CPU_state and call
833 push_CPU_state(); // keeps alignment at 16 bytes
834 lea(c_rarg0, ExternalAddress((address) msg));
835 call_VM_leaf(CAST_FROM_FN_PTR(address, warning), c_rarg0);
836 pop_CPU_state();
837 mov(rsp, rbp);
838 pop(rbp);
839 }
840
841 void MacroAssembler::print_state() {
842 address rip = pc();
843 pusha(); // get regs on stack
844 push(rbp);
845 movq(rbp, rsp);
846 andq(rsp, -16); // align stack as required by push_CPU_state and call
847 push_CPU_state(); // keeps alignment at 16 bytes
848
849 lea(c_rarg0, InternalAddress(rip));
850 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array
851 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1);
852
853 pop_CPU_state();
854 mov(rsp, rbp);
855 pop(rbp);
856 popa();
857 }
858
859 #ifndef PRODUCT
860 extern "C" void findpc(intptr_t x);
861 #endif
862
863 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) {
864 // In order to get locks to work, we need to fake a in_VM state
865 if (ShowMessageBoxOnError) {
866 JavaThread* thread = JavaThread::current();
867 JavaThreadState saved_state = thread->thread_state();
868 thread->set_thread_state(_thread_in_vm);
869 #ifndef PRODUCT
870 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
871 ttyLocker ttyl;
872 BytecodeCounter::print();
873 }
874 #endif
875 // To see where a verify_oop failed, get $ebx+40/X for this frame.
876 // XXX correct this offset for amd64
877 // This is the value of eip which points to where verify_oop will return.
878 if (os::message_box(msg, "Execution stopped, print registers?")) {
879 print_state64(pc, regs);
880 BREAKPOINT;
881 assert(false, "start up GDB");
882 }
883 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state);
884 } else {
885 ttyLocker ttyl;
886 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n",
887 msg);
888 assert(false, "DEBUG MESSAGE: %s", msg);
889 }
890 }
891
892 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) {
893 ttyLocker ttyl;
894 FlagSetting fs(Debugging, true);
895 tty->print_cr("rip = 0x%016lx", pc);
896 #ifndef PRODUCT
897 tty->cr();
898 findpc(pc);
899 tty->cr();
900 #endif
901 #define PRINT_REG(rax, value) \
902 { tty->print("%s = ", #rax); os::print_location(tty, value); }
903 PRINT_REG(rax, regs[15]);
904 PRINT_REG(rbx, regs[12]);
905 PRINT_REG(rcx, regs[14]);
906 PRINT_REG(rdx, regs[13]);
907 PRINT_REG(rdi, regs[8]);
908 PRINT_REG(rsi, regs[9]);
909 PRINT_REG(rbp, regs[10]);
910 PRINT_REG(rsp, regs[11]);
911 PRINT_REG(r8 , regs[7]);
912 PRINT_REG(r9 , regs[6]);
913 PRINT_REG(r10, regs[5]);
914 PRINT_REG(r11, regs[4]);
915 PRINT_REG(r12, regs[3]);
916 PRINT_REG(r13, regs[2]);
917 PRINT_REG(r14, regs[1]);
918 PRINT_REG(r15, regs[0]);
919 #undef PRINT_REG
920 // Print some words near top of staack.
921 int64_t* rsp = (int64_t*) regs[11];
922 int64_t* dump_sp = rsp;
923 for (int col1 = 0; col1 < 8; col1++) {
924 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp);
925 os::print_location(tty, *dump_sp++);
926 }
927 for (int row = 0; row < 25; row++) {
928 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp);
929 for (int col = 0; col < 4; col++) {
930 tty->print(" 0x%016lx", *dump_sp++);
931 }
932 tty->cr();
933 }
934 // Print some instructions around pc:
935 Disassembler::decode((address)pc-64, (address)pc);
936 tty->print_cr("--------");
937 Disassembler::decode((address)pc, (address)pc+32);
938 }
939
940 #endif // _LP64
941
942 // Now versions that are common to 32/64 bit
943
944 void MacroAssembler::addptr(Register dst, int32_t imm32) {
945 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32));
946 }
947
948 void MacroAssembler::addptr(Register dst, Register src) {
949 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
950 }
951
952 void MacroAssembler::addptr(Address dst, Register src) {
953 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src));
954 }
955
956 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) {
957 if (reachable(src)) {
958 Assembler::addsd(dst, as_Address(src));
959 } else {
960 lea(rscratch1, src);
961 Assembler::addsd(dst, Address(rscratch1, 0));
962 }
963 }
964
965 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) {
966 if (reachable(src)) {
967 addss(dst, as_Address(src));
968 } else {
969 lea(rscratch1, src);
970 addss(dst, Address(rscratch1, 0));
971 }
972 }
973
974 void MacroAssembler::align(int modulus) {
975 align(modulus, offset());
976 }
977
978 void MacroAssembler::align(int modulus, int target) {
979 if (target % modulus != 0) {
980 nop(modulus - (target % modulus));
981 }
982 }
983
984 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) {
985 // Used in sign-masking with aligned address.
986 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
987 if (reachable(src)) {
988 Assembler::andpd(dst, as_Address(src));
989 } else {
990 lea(rscratch1, src);
991 Assembler::andpd(dst, Address(rscratch1, 0));
992 }
993 }
994
995 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) {
996 // Used in sign-masking with aligned address.
997 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
998 if (reachable(src)) {
999 Assembler::andps(dst, as_Address(src));
1000 } else {
1001 lea(rscratch1, src);
1002 Assembler::andps(dst, Address(rscratch1, 0));
1003 }
1004 }
1005
1006 void MacroAssembler::andptr(Register dst, int32_t imm32) {
1007 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32));
1008 }
1009
1010 void MacroAssembler::atomic_incl(Address counter_addr) {
1011 if (os::is_MP())
1012 lock();
1013 incrementl(counter_addr);
1014 }
1015
1016 void MacroAssembler::atomic_incl(AddressLiteral counter_addr, Register scr) {
1017 if (reachable(counter_addr)) {
1018 atomic_incl(as_Address(counter_addr));
1019 } else {
1020 lea(scr, counter_addr);
1021 atomic_incl(Address(scr, 0));
1022 }
1023 }
1024
1025 #ifdef _LP64
1026 void MacroAssembler::atomic_incq(Address counter_addr) {
1027 if (os::is_MP())
1028 lock();
1029 incrementq(counter_addr);
1030 }
1031
1032 void MacroAssembler::atomic_incq(AddressLiteral counter_addr, Register scr) {
1033 if (reachable(counter_addr)) {
1034 atomic_incq(as_Address(counter_addr));
1035 } else {
1036 lea(scr, counter_addr);
1037 atomic_incq(Address(scr, 0));
1038 }
1039 }
1040 #endif
1041
1042 // Writes to stack successive pages until offset reached to check for
1043 // stack overflow + shadow pages. This clobbers tmp.
1044 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
1045 movptr(tmp, rsp);
1046 // Bang stack for total size given plus shadow page size.
1047 // Bang one page at a time because large size can bang beyond yellow and
1048 // red zones.
1049 Label loop;
1050 bind(loop);
1051 movl(Address(tmp, (-os::vm_page_size())), size );
1052 subptr(tmp, os::vm_page_size());
1053 subl(size, os::vm_page_size());
1054 jcc(Assembler::greater, loop);
1055
1056 // Bang down shadow pages too.
1057 // At this point, (tmp-0) is the last address touched, so don't
1058 // touch it again. (It was touched as (tmp-pagesize) but then tmp
1059 // was post-decremented.) Skip this address by starting at i=1, and
1060 // touch a few more pages below. N.B. It is important to touch all
1061 // the way down to and including i=StackShadowPages.
1062 for (int i = 1; i < StackShadowPages; i++) {
1063 // this could be any sized move but this is can be a debugging crumb
1064 // so the bigger the better.
1065 movptr(Address(tmp, (-i*os::vm_page_size())), size );
1066 }
1067 }
1068
1069 int MacroAssembler::biased_locking_enter(Register lock_reg,
1070 Register obj_reg,
1071 Register swap_reg,
1072 Register tmp_reg,
1073 bool swap_reg_contains_mark,
1074 Label& done,
1075 Label* slow_case,
1076 BiasedLockingCounters* counters) {
1077 assert(UseBiasedLocking, "why call this otherwise?");
1078 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq");
1079 assert(tmp_reg != noreg, "tmp_reg must be supplied");
1080 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg);
1081 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout");
1082 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes());
1083 Address saved_mark_addr(lock_reg, 0);
1084
1085 if (PrintBiasedLockingStatistics && counters == NULL) {
1086 counters = BiasedLocking::counters();
1087 }
1088 // Biased locking
1089 // See whether the lock is currently biased toward our thread and
1090 // whether the epoch is still valid
1091 // Note that the runtime guarantees sufficient alignment of JavaThread
1092 // pointers to allow age to be placed into low bits
1093 // First check to see whether biasing is even enabled for this object
1094 Label cas_label;
1095 int null_check_offset = -1;
1096 if (!swap_reg_contains_mark) {
1097 null_check_offset = offset();
1098 movptr(swap_reg, mark_addr);
1099 }
1100 movptr(tmp_reg, swap_reg);
1101 andptr(tmp_reg, markOopDesc::biased_lock_mask_in_place);
1102 cmpptr(tmp_reg, markOopDesc::biased_lock_pattern);
1103 jcc(Assembler::notEqual, cas_label);
1104 // The bias pattern is present in the object's header. Need to check
1105 // whether the bias owner and the epoch are both still current.
1106 #ifndef _LP64
1107 // Note that because there is no current thread register on x86_32 we
1108 // need to store off the mark word we read out of the object to
1109 // avoid reloading it and needing to recheck invariants below. This
1110 // store is unfortunate but it makes the overall code shorter and
1111 // simpler.
1112 movptr(saved_mark_addr, swap_reg);
1113 #endif
1114 if (swap_reg_contains_mark) {
1115 null_check_offset = offset();
1116 }
1117 load_prototype_header(tmp_reg, obj_reg);
1118 #ifdef _LP64
1119 orptr(tmp_reg, r15_thread);
1120 xorptr(tmp_reg, swap_reg);
1121 Register header_reg = tmp_reg;
1122 #else
1123 xorptr(tmp_reg, swap_reg);
1124 get_thread(swap_reg);
1125 xorptr(swap_reg, tmp_reg);
1126 Register header_reg = swap_reg;
1127 #endif
1128 andptr(header_reg, ~((int) markOopDesc::age_mask_in_place));
1129 if (counters != NULL) {
1130 cond_inc32(Assembler::zero,
1131 ExternalAddress((address) counters->biased_lock_entry_count_addr()));
1132 }
1133 jcc(Assembler::equal, done);
1134
1135 Label try_revoke_bias;
1136 Label try_rebias;
1137
1138 // At this point we know that the header has the bias pattern and
1139 // that we are not the bias owner in the current epoch. We need to
1140 // figure out more details about the state of the header in order to
1141 // know what operations can be legally performed on the object's
1142 // header.
1143
1144 // If the low three bits in the xor result aren't clear, that means
1145 // the prototype header is no longer biased and we have to revoke
1146 // the bias on this object.
1147 testptr(header_reg, markOopDesc::biased_lock_mask_in_place);
1148 jccb(Assembler::notZero, try_revoke_bias);
1149
1150 // Biasing is still enabled for this data type. See whether the
1151 // epoch of the current bias is still valid, meaning that the epoch
1152 // bits of the mark word are equal to the epoch bits of the
1153 // prototype header. (Note that the prototype header's epoch bits
1154 // only change at a safepoint.) If not, attempt to rebias the object
1155 // toward the current thread. Note that we must be absolutely sure
1156 // that the current epoch is invalid in order to do this because
1157 // otherwise the manipulations it performs on the mark word are
1158 // illegal.
1159 testptr(header_reg, markOopDesc::epoch_mask_in_place);
1160 jccb(Assembler::notZero, try_rebias);
1161
1162 // The epoch of the current bias is still valid but we know nothing
1163 // about the owner; it might be set or it might be clear. Try to
1164 // acquire the bias of the object using an atomic operation. If this
1165 // fails we will go in to the runtime to revoke the object's bias.
1166 // Note that we first construct the presumed unbiased header so we
1167 // don't accidentally blow away another thread's valid bias.
1168 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1169 andptr(swap_reg,
1170 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place);
1171 #ifdef _LP64
1172 movptr(tmp_reg, swap_reg);
1173 orptr(tmp_reg, r15_thread);
1174 #else
1175 get_thread(tmp_reg);
1176 orptr(tmp_reg, swap_reg);
1177 #endif
1178 if (os::is_MP()) {
1179 lock();
1180 }
1181 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1182 // If the biasing toward our thread failed, this means that
1183 // another thread succeeded in biasing it toward itself and we
1184 // need to revoke that bias. The revocation will occur in the
1185 // interpreter runtime in the slow case.
1186 if (counters != NULL) {
1187 cond_inc32(Assembler::zero,
1188 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr()));
1189 }
1190 if (slow_case != NULL) {
1191 jcc(Assembler::notZero, *slow_case);
1192 }
1193 jmp(done);
1194
1195 bind(try_rebias);
1196 // At this point we know the epoch has expired, meaning that the
1197 // current "bias owner", if any, is actually invalid. Under these
1198 // circumstances _only_, we are allowed to use the current header's
1199 // value as the comparison value when doing the cas to acquire the
1200 // bias in the current epoch. In other words, we allow transfer of
1201 // the bias from one thread to another directly in this situation.
1202 //
1203 // FIXME: due to a lack of registers we currently blow away the age
1204 // bits in this situation. Should attempt to preserve them.
1205 load_prototype_header(tmp_reg, obj_reg);
1206 #ifdef _LP64
1207 orptr(tmp_reg, r15_thread);
1208 #else
1209 get_thread(swap_reg);
1210 orptr(tmp_reg, swap_reg);
1211 movptr(swap_reg, saved_mark_addr);
1212 #endif
1213 if (os::is_MP()) {
1214 lock();
1215 }
1216 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1217 // If the biasing toward our thread failed, then another thread
1218 // succeeded in biasing it toward itself and we need to revoke that
1219 // bias. The revocation will occur in the runtime in the slow case.
1220 if (counters != NULL) {
1221 cond_inc32(Assembler::zero,
1222 ExternalAddress((address) counters->rebiased_lock_entry_count_addr()));
1223 }
1224 if (slow_case != NULL) {
1225 jcc(Assembler::notZero, *slow_case);
1226 }
1227 jmp(done);
1228
1229 bind(try_revoke_bias);
1230 // The prototype mark in the klass doesn't have the bias bit set any
1231 // more, indicating that objects of this data type are not supposed
1232 // to be biased any more. We are going to try to reset the mark of
1233 // this object to the prototype value and fall through to the
1234 // CAS-based locking scheme. Note that if our CAS fails, it means
1235 // that another thread raced us for the privilege of revoking the
1236 // bias of this particular object, so it's okay to continue in the
1237 // normal locking code.
1238 //
1239 // FIXME: due to a lack of registers we currently blow away the age
1240 // bits in this situation. Should attempt to preserve them.
1241 NOT_LP64( movptr(swap_reg, saved_mark_addr); )
1242 load_prototype_header(tmp_reg, obj_reg);
1243 if (os::is_MP()) {
1244 lock();
1245 }
1246 cmpxchgptr(tmp_reg, mark_addr); // compare tmp_reg and swap_reg
1247 // Fall through to the normal CAS-based lock, because no matter what
1248 // the result of the above CAS, some thread must have succeeded in
1249 // removing the bias bit from the object's header.
1250 if (counters != NULL) {
1251 cond_inc32(Assembler::zero,
1252 ExternalAddress((address) counters->revoked_lock_entry_count_addr()));
1253 }
1254
1255 bind(cas_label);
1256
1257 return null_check_offset;
1258 }
1259
1260 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) {
1261 assert(UseBiasedLocking, "why call this otherwise?");
1262
1263 // Check for biased locking unlock case, which is a no-op
1264 // Note: we do not have to check the thread ID for two reasons.
1265 // First, the interpreter checks for IllegalMonitorStateException at
1266 // a higher level. Second, if the bias was revoked while we held the
1267 // lock, the object could not be rebiased toward another thread, so
1268 // the bias bit would be clear.
1269 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1270 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place);
1271 cmpptr(temp_reg, markOopDesc::biased_lock_pattern);
1272 jcc(Assembler::equal, done);
1273 }
1274
1275 #ifdef COMPILER2
1276
1277 #if INCLUDE_RTM_OPT
1278
1279 // Update rtm_counters based on abort status
1280 // input: abort_status
1281 // rtm_counters (RTMLockingCounters*)
1282 // flags are killed
1283 void MacroAssembler::rtm_counters_update(Register abort_status, Register rtm_counters) {
1284
1285 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abort_count_offset()));
1286 if (PrintPreciseRTMLockingStatistics) {
1287 for (int i = 0; i < RTMLockingCounters::ABORT_STATUS_LIMIT; i++) {
1288 Label check_abort;
1289 testl(abort_status, (1<<i));
1290 jccb(Assembler::equal, check_abort);
1291 atomic_incptr(Address(rtm_counters, RTMLockingCounters::abortX_count_offset() + (i * sizeof(uintx))));
1292 bind(check_abort);
1293 }
1294 }
1295 }
1296
1297 // Branch if (random & (count-1) != 0), count is 2^n
1298 // tmp, scr and flags are killed
1299 void MacroAssembler::branch_on_random_using_rdtsc(Register tmp, Register scr, int count, Label& brLabel) {
1300 assert(tmp == rax, "");
1301 assert(scr == rdx, "");
1302 rdtsc(); // modifies EDX:EAX
1303 andptr(tmp, count-1);
1304 jccb(Assembler::notZero, brLabel);
1305 }
1306
1307 // Perform abort ratio calculation, set no_rtm bit if high ratio
1308 // input: rtm_counters_Reg (RTMLockingCounters* address)
1309 // tmpReg, rtm_counters_Reg and flags are killed
1310 void MacroAssembler::rtm_abort_ratio_calculation(Register tmpReg,
1311 Register rtm_counters_Reg,
1312 RTMLockingCounters* rtm_counters,
1313 Metadata* method_data) {
1314 Label L_done, L_check_always_rtm1, L_check_always_rtm2;
1315
1316 if (RTMLockingCalculationDelay > 0) {
1317 // Delay calculation
1318 movptr(tmpReg, ExternalAddress((address) RTMLockingCounters::rtm_calculation_flag_addr()), tmpReg);
1319 testptr(tmpReg, tmpReg);
1320 jccb(Assembler::equal, L_done);
1321 }
1322 // Abort ratio calculation only if abort_count > RTMAbortThreshold
1323 // Aborted transactions = abort_count * 100
1324 // All transactions = total_count * RTMTotalCountIncrRate
1325 // Set no_rtm bit if (Aborted transactions >= All transactions * RTMAbortRatio)
1326
1327 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::abort_count_offset()));
1328 cmpptr(tmpReg, RTMAbortThreshold);
1329 jccb(Assembler::below, L_check_always_rtm2);
1330 imulptr(tmpReg, tmpReg, 100);
1331
1332 Register scrReg = rtm_counters_Reg;
1333 movptr(scrReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1334 imulptr(scrReg, scrReg, RTMTotalCountIncrRate);
1335 imulptr(scrReg, scrReg, RTMAbortRatio);
1336 cmpptr(tmpReg, scrReg);
1337 jccb(Assembler::below, L_check_always_rtm1);
1338 if (method_data != NULL) {
1339 // set rtm_state to "no rtm" in MDO
1340 mov_metadata(tmpReg, method_data);
1341 if (os::is_MP()) {
1342 lock();
1343 }
1344 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), NoRTM);
1345 }
1346 jmpb(L_done);
1347 bind(L_check_always_rtm1);
1348 // Reload RTMLockingCounters* address
1349 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1350 bind(L_check_always_rtm2);
1351 movptr(tmpReg, Address(rtm_counters_Reg, RTMLockingCounters::total_count_offset()));
1352 cmpptr(tmpReg, RTMLockingThreshold / RTMTotalCountIncrRate);
1353 jccb(Assembler::below, L_done);
1354 if (method_data != NULL) {
1355 // set rtm_state to "always rtm" in MDO
1356 mov_metadata(tmpReg, method_data);
1357 if (os::is_MP()) {
1358 lock();
1359 }
1360 orl(Address(tmpReg, MethodData::rtm_state_offset_in_bytes()), UseRTM);
1361 }
1362 bind(L_done);
1363 }
1364
1365 // Update counters and perform abort ratio calculation
1366 // input: abort_status_Reg
1367 // rtm_counters_Reg, flags are killed
1368 void MacroAssembler::rtm_profiling(Register abort_status_Reg,
1369 Register rtm_counters_Reg,
1370 RTMLockingCounters* rtm_counters,
1371 Metadata* method_data,
1372 bool profile_rtm) {
1373
1374 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1375 // update rtm counters based on rax value at abort
1376 // reads abort_status_Reg, updates flags
1377 lea(rtm_counters_Reg, ExternalAddress((address)rtm_counters));
1378 rtm_counters_update(abort_status_Reg, rtm_counters_Reg);
1379 if (profile_rtm) {
1380 // Save abort status because abort_status_Reg is used by following code.
1381 if (RTMRetryCount > 0) {
1382 push(abort_status_Reg);
1383 }
1384 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1385 rtm_abort_ratio_calculation(abort_status_Reg, rtm_counters_Reg, rtm_counters, method_data);
1386 // restore abort status
1387 if (RTMRetryCount > 0) {
1388 pop(abort_status_Reg);
1389 }
1390 }
1391 }
1392
1393 // Retry on abort if abort's status is 0x6: can retry (0x2) | memory conflict (0x4)
1394 // inputs: retry_count_Reg
1395 // : abort_status_Reg
1396 // output: retry_count_Reg decremented by 1
1397 // flags are killed
1398 void MacroAssembler::rtm_retry_lock_on_abort(Register retry_count_Reg, Register abort_status_Reg, Label& retryLabel) {
1399 Label doneRetry;
1400 assert(abort_status_Reg == rax, "");
1401 // The abort reason bits are in eax (see all states in rtmLocking.hpp)
1402 // 0x6 = conflict on which we can retry (0x2) | memory conflict (0x4)
1403 // if reason is in 0x6 and retry count != 0 then retry
1404 andptr(abort_status_Reg, 0x6);
1405 jccb(Assembler::zero, doneRetry);
1406 testl(retry_count_Reg, retry_count_Reg);
1407 jccb(Assembler::zero, doneRetry);
1408 pause();
1409 decrementl(retry_count_Reg);
1410 jmp(retryLabel);
1411 bind(doneRetry);
1412 }
1413
1414 // Spin and retry if lock is busy,
1415 // inputs: box_Reg (monitor address)
1416 // : retry_count_Reg
1417 // output: retry_count_Reg decremented by 1
1418 // : clear z flag if retry count exceeded
1419 // tmp_Reg, scr_Reg, flags are killed
1420 void MacroAssembler::rtm_retry_lock_on_busy(Register retry_count_Reg, Register box_Reg,
1421 Register tmp_Reg, Register scr_Reg, Label& retryLabel) {
1422 Label SpinLoop, SpinExit, doneRetry;
1423 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1424
1425 testl(retry_count_Reg, retry_count_Reg);
1426 jccb(Assembler::zero, doneRetry);
1427 decrementl(retry_count_Reg);
1428 movptr(scr_Reg, RTMSpinLoopCount);
1429
1430 bind(SpinLoop);
1431 pause();
1432 decrementl(scr_Reg);
1433 jccb(Assembler::lessEqual, SpinExit);
1434 movptr(tmp_Reg, Address(box_Reg, owner_offset));
1435 testptr(tmp_Reg, tmp_Reg);
1436 jccb(Assembler::notZero, SpinLoop);
1437
1438 bind(SpinExit);
1439 jmp(retryLabel);
1440 bind(doneRetry);
1441 incrementl(retry_count_Reg); // clear z flag
1442 }
1443
1444 // Use RTM for normal stack locks
1445 // Input: objReg (object to lock)
1446 void MacroAssembler::rtm_stack_locking(Register objReg, Register tmpReg, Register scrReg,
1447 Register retry_on_abort_count_Reg,
1448 RTMLockingCounters* stack_rtm_counters,
1449 Metadata* method_data, bool profile_rtm,
1450 Label& DONE_LABEL, Label& IsInflated) {
1451 assert(UseRTMForStackLocks, "why call this otherwise?");
1452 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1453 assert(tmpReg == rax, "");
1454 assert(scrReg == rdx, "");
1455 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1456
1457 if (RTMRetryCount > 0) {
1458 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1459 bind(L_rtm_retry);
1460 }
1461 movptr(tmpReg, Address(objReg, 0));
1462 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1463 jcc(Assembler::notZero, IsInflated);
1464
1465 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1466 Label L_noincrement;
1467 if (RTMTotalCountIncrRate > 1) {
1468 // tmpReg, scrReg and flags are killed
1469 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement);
1470 }
1471 assert(stack_rtm_counters != NULL, "should not be NULL when profiling RTM");
1472 atomic_incptr(ExternalAddress((address)stack_rtm_counters->total_count_addr()), scrReg);
1473 bind(L_noincrement);
1474 }
1475 xbegin(L_on_abort);
1476 movptr(tmpReg, Address(objReg, 0)); // fetch markword
1477 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1478 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1479 jcc(Assembler::equal, DONE_LABEL); // all done if unlocked
1480
1481 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1482 if (UseRTMXendForLockBusy) {
1483 xend();
1484 movptr(abort_status_Reg, 0x2); // Set the abort status to 2 (so we can retry)
1485 jmp(L_decrement_retry);
1486 }
1487 else {
1488 xabort(0);
1489 }
1490 bind(L_on_abort);
1491 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1492 rtm_profiling(abort_status_Reg, scrReg, stack_rtm_counters, method_data, profile_rtm);
1493 }
1494 bind(L_decrement_retry);
1495 if (RTMRetryCount > 0) {
1496 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1497 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1498 }
1499 }
1500
1501 // Use RTM for inflating locks
1502 // inputs: objReg (object to lock)
1503 // boxReg (on-stack box address (displaced header location) - KILLED)
1504 // tmpReg (ObjectMonitor address + markOopDesc::monitor_value)
1505 void MacroAssembler::rtm_inflated_locking(Register objReg, Register boxReg, Register tmpReg,
1506 Register scrReg, Register retry_on_busy_count_Reg,
1507 Register retry_on_abort_count_Reg,
1508 RTMLockingCounters* rtm_counters,
1509 Metadata* method_data, bool profile_rtm,
1510 Label& DONE_LABEL) {
1511 assert(UseRTMLocking, "why call this otherwise?");
1512 assert(tmpReg == rax, "");
1513 assert(scrReg == rdx, "");
1514 Label L_rtm_retry, L_decrement_retry, L_on_abort;
1515 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1516
1517 // Without cast to int32_t a movptr will destroy r10 which is typically obj
1518 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1519 movptr(boxReg, tmpReg); // Save ObjectMonitor address
1520
1521 if (RTMRetryCount > 0) {
1522 movl(retry_on_busy_count_Reg, RTMRetryCount); // Retry on lock busy
1523 movl(retry_on_abort_count_Reg, RTMRetryCount); // Retry on abort
1524 bind(L_rtm_retry);
1525 }
1526 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1527 Label L_noincrement;
1528 if (RTMTotalCountIncrRate > 1) {
1529 // tmpReg, scrReg and flags are killed
1530 branch_on_random_using_rdtsc(tmpReg, scrReg, (int)RTMTotalCountIncrRate, L_noincrement);
1531 }
1532 assert(rtm_counters != NULL, "should not be NULL when profiling RTM");
1533 atomic_incptr(ExternalAddress((address)rtm_counters->total_count_addr()), scrReg);
1534 bind(L_noincrement);
1535 }
1536 xbegin(L_on_abort);
1537 movptr(tmpReg, Address(objReg, 0));
1538 movptr(tmpReg, Address(tmpReg, owner_offset));
1539 testptr(tmpReg, tmpReg);
1540 jcc(Assembler::zero, DONE_LABEL);
1541 if (UseRTMXendForLockBusy) {
1542 xend();
1543 jmp(L_decrement_retry);
1544 }
1545 else {
1546 xabort(0);
1547 }
1548 bind(L_on_abort);
1549 Register abort_status_Reg = tmpReg; // status of abort is stored in RAX
1550 if (PrintPreciseRTMLockingStatistics || profile_rtm) {
1551 rtm_profiling(abort_status_Reg, scrReg, rtm_counters, method_data, profile_rtm);
1552 }
1553 if (RTMRetryCount > 0) {
1554 // retry on lock abort if abort status is 'can retry' (0x2) or 'memory conflict' (0x4)
1555 rtm_retry_lock_on_abort(retry_on_abort_count_Reg, abort_status_Reg, L_rtm_retry);
1556 }
1557
1558 movptr(tmpReg, Address(boxReg, owner_offset)) ;
1559 testptr(tmpReg, tmpReg) ;
1560 jccb(Assembler::notZero, L_decrement_retry) ;
1561
1562 // Appears unlocked - try to swing _owner from null to non-null.
1563 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1564 #ifdef _LP64
1565 Register threadReg = r15_thread;
1566 #else
1567 get_thread(scrReg);
1568 Register threadReg = scrReg;
1569 #endif
1570 if (os::is_MP()) {
1571 lock();
1572 }
1573 cmpxchgptr(threadReg, Address(boxReg, owner_offset)); // Updates tmpReg
1574
1575 if (RTMRetryCount > 0) {
1576 // success done else retry
1577 jccb(Assembler::equal, DONE_LABEL) ;
1578 bind(L_decrement_retry);
1579 // Spin and retry if lock is busy.
1580 rtm_retry_lock_on_busy(retry_on_busy_count_Reg, boxReg, tmpReg, scrReg, L_rtm_retry);
1581 }
1582 else {
1583 bind(L_decrement_retry);
1584 }
1585 }
1586
1587 #endif // INCLUDE_RTM_OPT
1588
1589 // Fast_Lock and Fast_Unlock used by C2
1590
1591 // Because the transitions from emitted code to the runtime
1592 // monitorenter/exit helper stubs are so slow it's critical that
1593 // we inline both the stack-locking fast-path and the inflated fast path.
1594 //
1595 // See also: cmpFastLock and cmpFastUnlock.
1596 //
1597 // What follows is a specialized inline transliteration of the code
1598 // in slow_enter() and slow_exit(). If we're concerned about I$ bloat
1599 // another option would be to emit TrySlowEnter and TrySlowExit methods
1600 // at startup-time. These methods would accept arguments as
1601 // (rax,=Obj, rbx=Self, rcx=box, rdx=Scratch) and return success-failure
1602 // indications in the icc.ZFlag. Fast_Lock and Fast_Unlock would simply
1603 // marshal the arguments and emit calls to TrySlowEnter and TrySlowExit.
1604 // In practice, however, the # of lock sites is bounded and is usually small.
1605 // Besides the call overhead, TrySlowEnter and TrySlowExit might suffer
1606 // if the processor uses simple bimodal branch predictors keyed by EIP
1607 // Since the helper routines would be called from multiple synchronization
1608 // sites.
1609 //
1610 // An even better approach would be write "MonitorEnter()" and "MonitorExit()"
1611 // in java - using j.u.c and unsafe - and just bind the lock and unlock sites
1612 // to those specialized methods. That'd give us a mostly platform-independent
1613 // implementation that the JITs could optimize and inline at their pleasure.
1614 // Done correctly, the only time we'd need to cross to native could would be
1615 // to park() or unpark() threads. We'd also need a few more unsafe operators
1616 // to (a) prevent compiler-JIT reordering of non-volatile accesses, and
1617 // (b) explicit barriers or fence operations.
1618 //
1619 // TODO:
1620 //
1621 // * Arrange for C2 to pass "Self" into Fast_Lock and Fast_Unlock in one of the registers (scr).
1622 // This avoids manifesting the Self pointer in the Fast_Lock and Fast_Unlock terminals.
1623 // Given TLAB allocation, Self is usually manifested in a register, so passing it into
1624 // the lock operators would typically be faster than reifying Self.
1625 //
1626 // * Ideally I'd define the primitives as:
1627 // fast_lock (nax Obj, nax box, EAX tmp, nax scr) where box, tmp and scr are KILLED.
1628 // fast_unlock (nax Obj, EAX box, nax tmp) where box and tmp are KILLED
1629 // Unfortunately ADLC bugs prevent us from expressing the ideal form.
1630 // Instead, we're stuck with a rather awkward and brittle register assignments below.
1631 // Furthermore the register assignments are overconstrained, possibly resulting in
1632 // sub-optimal code near the synchronization site.
1633 //
1634 // * Eliminate the sp-proximity tests and just use "== Self" tests instead.
1635 // Alternately, use a better sp-proximity test.
1636 //
1637 // * Currently ObjectMonitor._Owner can hold either an sp value or a (THREAD *) value.
1638 // Either one is sufficient to uniquely identify a thread.
1639 // TODO: eliminate use of sp in _owner and use get_thread(tr) instead.
1640 //
1641 // * Intrinsify notify() and notifyAll() for the common cases where the
1642 // object is locked by the calling thread but the waitlist is empty.
1643 // avoid the expensive JNI call to JVM_Notify() and JVM_NotifyAll().
1644 //
1645 // * use jccb and jmpb instead of jcc and jmp to improve code density.
1646 // But beware of excessive branch density on AMD Opterons.
1647 //
1648 // * Both Fast_Lock and Fast_Unlock set the ICC.ZF to indicate success
1649 // or failure of the fast-path. If the fast-path fails then we pass
1650 // control to the slow-path, typically in C. In Fast_Lock and
1651 // Fast_Unlock we often branch to DONE_LABEL, just to find that C2
1652 // will emit a conditional branch immediately after the node.
1653 // So we have branches to branches and lots of ICC.ZF games.
1654 // Instead, it might be better to have C2 pass a "FailureLabel"
1655 // into Fast_Lock and Fast_Unlock. In the case of success, control
1656 // will drop through the node. ICC.ZF is undefined at exit.
1657 // In the case of failure, the node will branch directly to the
1658 // FailureLabel
1659
1660
1661 // obj: object to lock
1662 // box: on-stack box address (displaced header location) - KILLED
1663 // rax,: tmp -- KILLED
1664 // scr: tmp -- KILLED
1665 void MacroAssembler::fast_lock(Register objReg, Register boxReg, Register tmpReg,
1666 Register scrReg, Register cx1Reg, Register cx2Reg,
1667 BiasedLockingCounters* counters,
1668 RTMLockingCounters* rtm_counters,
1669 RTMLockingCounters* stack_rtm_counters,
1670 Metadata* method_data,
1671 bool use_rtm, bool profile_rtm) {
1672 // Ensure the register assignents are disjoint
1673 assert(tmpReg == rax, "");
1674
1675 if (use_rtm) {
1676 assert_different_registers(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg);
1677 } else {
1678 assert(cx1Reg == noreg, "");
1679 assert(cx2Reg == noreg, "");
1680 assert_different_registers(objReg, boxReg, tmpReg, scrReg);
1681 }
1682
1683 if (counters != NULL) {
1684 atomic_incl(ExternalAddress((address)counters->total_entry_count_addr()), scrReg);
1685 }
1686 if (EmitSync & 1) {
1687 // set box->dhw = markOopDesc::unused_mark()
1688 // Force all sync thru slow-path: slow_enter() and slow_exit()
1689 movptr (Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1690 cmpptr (rsp, (int32_t)NULL_WORD);
1691 } else {
1692 // Possible cases that we'll encounter in fast_lock
1693 // ------------------------------------------------
1694 // * Inflated
1695 // -- unlocked
1696 // -- Locked
1697 // = by self
1698 // = by other
1699 // * biased
1700 // -- by Self
1701 // -- by other
1702 // * neutral
1703 // * stack-locked
1704 // -- by self
1705 // = sp-proximity test hits
1706 // = sp-proximity test generates false-negative
1707 // -- by other
1708 //
1709
1710 Label IsInflated, DONE_LABEL;
1711
1712 // it's stack-locked, biased or neutral
1713 // TODO: optimize away redundant LDs of obj->mark and improve the markword triage
1714 // order to reduce the number of conditional branches in the most common cases.
1715 // Beware -- there's a subtle invariant that fetch of the markword
1716 // at [FETCH], below, will never observe a biased encoding (*101b).
1717 // If this invariant is not held we risk exclusion (safety) failure.
1718 if (UseBiasedLocking && !UseOptoBiasInlining) {
1719 biased_locking_enter(boxReg, objReg, tmpReg, scrReg, false, DONE_LABEL, NULL, counters);
1720 }
1721
1722 #if INCLUDE_RTM_OPT
1723 if (UseRTMForStackLocks && use_rtm) {
1724 rtm_stack_locking(objReg, tmpReg, scrReg, cx2Reg,
1725 stack_rtm_counters, method_data, profile_rtm,
1726 DONE_LABEL, IsInflated);
1727 }
1728 #endif // INCLUDE_RTM_OPT
1729
1730 movptr(tmpReg, Address(objReg, 0)); // [FETCH]
1731 testptr(tmpReg, markOopDesc::monitor_value); // inflated vs stack-locked|neutral|biased
1732 jccb(Assembler::notZero, IsInflated);
1733
1734 // Attempt stack-locking ...
1735 orptr (tmpReg, markOopDesc::unlocked_value);
1736 movptr(Address(boxReg, 0), tmpReg); // Anticipate successful CAS
1737 if (os::is_MP()) {
1738 lock();
1739 }
1740 cmpxchgptr(boxReg, Address(objReg, 0)); // Updates tmpReg
1741 if (counters != NULL) {
1742 cond_inc32(Assembler::equal,
1743 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1744 }
1745 jcc(Assembler::equal, DONE_LABEL); // Success
1746
1747 // Recursive locking.
1748 // The object is stack-locked: markword contains stack pointer to BasicLock.
1749 // Locked by current thread if difference with current SP is less than one page.
1750 subptr(tmpReg, rsp);
1751 // Next instruction set ZFlag == 1 (Success) if difference is less then one page.
1752 andptr(tmpReg, (int32_t) (NOT_LP64(0xFFFFF003) LP64_ONLY(7 - os::vm_page_size())) );
1753 movptr(Address(boxReg, 0), tmpReg);
1754 if (counters != NULL) {
1755 cond_inc32(Assembler::equal,
1756 ExternalAddress((address)counters->fast_path_entry_count_addr()));
1757 }
1758 jmp(DONE_LABEL);
1759
1760 bind(IsInflated);
1761 // The object is inflated. tmpReg contains pointer to ObjectMonitor* + markOopDesc::monitor_value
1762
1763 #if INCLUDE_RTM_OPT
1764 // Use the same RTM locking code in 32- and 64-bit VM.
1765 if (use_rtm) {
1766 rtm_inflated_locking(objReg, boxReg, tmpReg, scrReg, cx1Reg, cx2Reg,
1767 rtm_counters, method_data, profile_rtm, DONE_LABEL);
1768 } else {
1769 #endif // INCLUDE_RTM_OPT
1770
1771 #ifndef _LP64
1772 // The object is inflated.
1773
1774 // boxReg refers to the on-stack BasicLock in the current frame.
1775 // We'd like to write:
1776 // set box->_displaced_header = markOopDesc::unused_mark(). Any non-0 value suffices.
1777 // This is convenient but results a ST-before-CAS penalty. The following CAS suffers
1778 // additional latency as we have another ST in the store buffer that must drain.
1779
1780 if (EmitSync & 8192) {
1781 movptr(Address(boxReg, 0), 3); // results in ST-before-CAS penalty
1782 get_thread (scrReg);
1783 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1784 movptr(tmpReg, NULL_WORD); // consider: xor vs mov
1785 if (os::is_MP()) {
1786 lock();
1787 }
1788 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1789 } else
1790 if ((EmitSync & 128) == 0) { // avoid ST-before-CAS
1791 // register juggle because we need tmpReg for cmpxchgptr below
1792 movptr(scrReg, boxReg);
1793 movptr(boxReg, tmpReg); // consider: LEA box, [tmp-2]
1794
1795 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1796 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1797 // prefetchw [eax + Offset(_owner)-2]
1798 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1799 }
1800
1801 if ((EmitSync & 64) == 0) {
1802 // Optimistic form: consider XORL tmpReg,tmpReg
1803 movptr(tmpReg, NULL_WORD);
1804 } else {
1805 // Can suffer RTS->RTO upgrades on shared or cold $ lines
1806 // Test-And-CAS instead of CAS
1807 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner
1808 testptr(tmpReg, tmpReg); // Locked ?
1809 jccb (Assembler::notZero, DONE_LABEL);
1810 }
1811
1812 // Appears unlocked - try to swing _owner from null to non-null.
1813 // Ideally, I'd manifest "Self" with get_thread and then attempt
1814 // to CAS the register containing Self into m->Owner.
1815 // But we don't have enough registers, so instead we can either try to CAS
1816 // rsp or the address of the box (in scr) into &m->owner. If the CAS succeeds
1817 // we later store "Self" into m->Owner. Transiently storing a stack address
1818 // (rsp or the address of the box) into m->owner is harmless.
1819 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1820 if (os::is_MP()) {
1821 lock();
1822 }
1823 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1824 movptr(Address(scrReg, 0), 3); // box->_displaced_header = 3
1825 // If we weren't able to swing _owner from NULL to the BasicLock
1826 // then take the slow path.
1827 jccb (Assembler::notZero, DONE_LABEL);
1828 // update _owner from BasicLock to thread
1829 get_thread (scrReg); // beware: clobbers ICCs
1830 movptr(Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), scrReg);
1831 xorptr(boxReg, boxReg); // set icc.ZFlag = 1 to indicate success
1832
1833 // If the CAS fails we can either retry or pass control to the slow-path.
1834 // We use the latter tactic.
1835 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1836 // If the CAS was successful ...
1837 // Self has acquired the lock
1838 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1839 // Intentional fall-through into DONE_LABEL ...
1840 } else {
1841 movptr(Address(boxReg, 0), intptr_t(markOopDesc::unused_mark())); // results in ST-before-CAS penalty
1842 movptr(boxReg, tmpReg);
1843
1844 // Using a prefetchw helps avoid later RTS->RTO upgrades and cache probes
1845 if ((EmitSync & 2048) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
1846 // prefetchw [eax + Offset(_owner)-2]
1847 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1848 }
1849
1850 if ((EmitSync & 64) == 0) {
1851 // Optimistic form
1852 xorptr (tmpReg, tmpReg);
1853 } else {
1854 // Can suffer RTS->RTO upgrades on shared or cold $ lines
1855 movptr(tmpReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner))); // rax, = m->_owner
1856 testptr(tmpReg, tmpReg); // Locked ?
1857 jccb (Assembler::notZero, DONE_LABEL);
1858 }
1859
1860 // Appears unlocked - try to swing _owner from null to non-null.
1861 // Use either "Self" (in scr) or rsp as thread identity in _owner.
1862 // Invariant: tmpReg == 0. tmpReg is EAX which is the implicit cmpxchg comparand.
1863 get_thread (scrReg);
1864 if (os::is_MP()) {
1865 lock();
1866 }
1867 cmpxchgptr(scrReg, Address(boxReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1868
1869 // If the CAS fails we can either retry or pass control to the slow-path.
1870 // We use the latter tactic.
1871 // Pass the CAS result in the icc.ZFlag into DONE_LABEL
1872 // If the CAS was successful ...
1873 // Self has acquired the lock
1874 // Invariant: m->_recursions should already be 0, so we don't need to explicitly set it.
1875 // Intentional fall-through into DONE_LABEL ...
1876 }
1877 #else // _LP64
1878 // It's inflated
1879 movq(scrReg, tmpReg);
1880 xorq(tmpReg, tmpReg);
1881
1882 if (os::is_MP()) {
1883 lock();
1884 }
1885 cmpxchgptr(r15_thread, Address(scrReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
1886 // Unconditionally set box->_displaced_header = markOopDesc::unused_mark().
1887 // Without cast to int32_t movptr will destroy r10 which is typically obj.
1888 movptr(Address(boxReg, 0), (int32_t)intptr_t(markOopDesc::unused_mark()));
1889 // Intentional fall-through into DONE_LABEL ...
1890 // Propagate ICC.ZF from CAS above into DONE_LABEL.
1891 #endif // _LP64
1892 #if INCLUDE_RTM_OPT
1893 } // use_rtm()
1894 #endif
1895 // DONE_LABEL is a hot target - we'd really like to place it at the
1896 // start of cache line by padding with NOPs.
1897 // See the AMD and Intel software optimization manuals for the
1898 // most efficient "long" NOP encodings.
1899 // Unfortunately none of our alignment mechanisms suffice.
1900 bind(DONE_LABEL);
1901
1902 // At DONE_LABEL the icc ZFlag is set as follows ...
1903 // Fast_Unlock uses the same protocol.
1904 // ZFlag == 1 -> Success
1905 // ZFlag == 0 -> Failure - force control through the slow-path
1906 }
1907 }
1908
1909 // obj: object to unlock
1910 // box: box address (displaced header location), killed. Must be EAX.
1911 // tmp: killed, cannot be obj nor box.
1912 //
1913 // Some commentary on balanced locking:
1914 //
1915 // Fast_Lock and Fast_Unlock are emitted only for provably balanced lock sites.
1916 // Methods that don't have provably balanced locking are forced to run in the
1917 // interpreter - such methods won't be compiled to use fast_lock and fast_unlock.
1918 // The interpreter provides two properties:
1919 // I1: At return-time the interpreter automatically and quietly unlocks any
1920 // objects acquired the current activation (frame). Recall that the
1921 // interpreter maintains an on-stack list of locks currently held by
1922 // a frame.
1923 // I2: If a method attempts to unlock an object that is not held by the
1924 // the frame the interpreter throws IMSX.
1925 //
1926 // Lets say A(), which has provably balanced locking, acquires O and then calls B().
1927 // B() doesn't have provably balanced locking so it runs in the interpreter.
1928 // Control returns to A() and A() unlocks O. By I1 and I2, above, we know that O
1929 // is still locked by A().
1930 //
1931 // The only other source of unbalanced locking would be JNI. The "Java Native Interface:
1932 // Programmer's Guide and Specification" claims that an object locked by jni_monitorenter
1933 // should not be unlocked by "normal" java-level locking and vice-versa. The specification
1934 // doesn't specify what will occur if a program engages in such mixed-mode locking, however.
1935 // Arguably given that the spec legislates the JNI case as undefined our implementation
1936 // could reasonably *avoid* checking owner in Fast_Unlock().
1937 // In the interest of performance we elide m->Owner==Self check in unlock.
1938 // A perfectly viable alternative is to elide the owner check except when
1939 // Xcheck:jni is enabled.
1940
1941 void MacroAssembler::fast_unlock(Register objReg, Register boxReg, Register tmpReg, bool use_rtm) {
1942 assert(boxReg == rax, "");
1943 assert_different_registers(objReg, boxReg, tmpReg);
1944
1945 if (EmitSync & 4) {
1946 // Disable - inhibit all inlining. Force control through the slow-path
1947 cmpptr (rsp, 0);
1948 } else {
1949 Label DONE_LABEL, Stacked, CheckSucc;
1950
1951 // Critically, the biased locking test must have precedence over
1952 // and appear before the (box->dhw == 0) recursive stack-lock test.
1953 if (UseBiasedLocking && !UseOptoBiasInlining) {
1954 biased_locking_exit(objReg, tmpReg, DONE_LABEL);
1955 }
1956
1957 #if INCLUDE_RTM_OPT
1958 if (UseRTMForStackLocks && use_rtm) {
1959 assert(!UseBiasedLocking, "Biased locking is not supported with RTM locking");
1960 Label L_regular_unlock;
1961 movptr(tmpReg, Address(objReg, 0)); // fetch markword
1962 andptr(tmpReg, markOopDesc::biased_lock_mask_in_place); // look at 3 lock bits
1963 cmpptr(tmpReg, markOopDesc::unlocked_value); // bits = 001 unlocked
1964 jccb(Assembler::notEqual, L_regular_unlock); // if !HLE RegularLock
1965 xend(); // otherwise end...
1966 jmp(DONE_LABEL); // ... and we're done
1967 bind(L_regular_unlock);
1968 }
1969 #endif
1970
1971 cmpptr(Address(boxReg, 0), (int32_t)NULL_WORD); // Examine the displaced header
1972 jcc (Assembler::zero, DONE_LABEL); // 0 indicates recursive stack-lock
1973 movptr(tmpReg, Address(objReg, 0)); // Examine the object's markword
1974 testptr(tmpReg, markOopDesc::monitor_value); // Inflated?
1975 jccb (Assembler::zero, Stacked);
1976
1977 // It's inflated.
1978 #if INCLUDE_RTM_OPT
1979 if (use_rtm) {
1980 Label L_regular_inflated_unlock;
1981 int owner_offset = OM_OFFSET_NO_MONITOR_VALUE_TAG(owner);
1982 movptr(boxReg, Address(tmpReg, owner_offset));
1983 testptr(boxReg, boxReg);
1984 jccb(Assembler::notZero, L_regular_inflated_unlock);
1985 xend();
1986 jmpb(DONE_LABEL);
1987 bind(L_regular_inflated_unlock);
1988 }
1989 #endif
1990
1991 // Despite our balanced locking property we still check that m->_owner == Self
1992 // as java routines or native JNI code called by this thread might
1993 // have released the lock.
1994 // Refer to the comments in synchronizer.cpp for how we might encode extra
1995 // state in _succ so we can avoid fetching EntryList|cxq.
1996 //
1997 // I'd like to add more cases in fast_lock() and fast_unlock() --
1998 // such as recursive enter and exit -- but we have to be wary of
1999 // I$ bloat, T$ effects and BP$ effects.
2000 //
2001 // If there's no contention try a 1-0 exit. That is, exit without
2002 // a costly MEMBAR or CAS. See synchronizer.cpp for details on how
2003 // we detect and recover from the race that the 1-0 exit admits.
2004 //
2005 // Conceptually Fast_Unlock() must execute a STST|LDST "release" barrier
2006 // before it STs null into _owner, releasing the lock. Updates
2007 // to data protected by the critical section must be visible before
2008 // we drop the lock (and thus before any other thread could acquire
2009 // the lock and observe the fields protected by the lock).
2010 // IA32's memory-model is SPO, so STs are ordered with respect to
2011 // each other and there's no need for an explicit barrier (fence).
2012 // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
2013 #ifndef _LP64
2014 get_thread (boxReg);
2015 if ((EmitSync & 4096) && VM_Version::supports_3dnow_prefetch() && os::is_MP()) {
2016 // prefetchw [ebx + Offset(_owner)-2]
2017 prefetchw(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2018 }
2019
2020 // Note that we could employ various encoding schemes to reduce
2021 // the number of loads below (currently 4) to just 2 or 3.
2022 // Refer to the comments in synchronizer.cpp.
2023 // In practice the chain of fetches doesn't seem to impact performance, however.
2024 xorptr(boxReg, boxReg);
2025 if ((EmitSync & 65536) == 0 && (EmitSync & 256)) {
2026 // Attempt to reduce branch density - AMD's branch predictor.
2027 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2028 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2029 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2030 jccb (Assembler::notZero, DONE_LABEL);
2031 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2032 jmpb (DONE_LABEL);
2033 } else {
2034 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2035 jccb (Assembler::notZero, DONE_LABEL);
2036 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2037 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2038 jccb (Assembler::notZero, CheckSucc);
2039 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2040 jmpb (DONE_LABEL);
2041 }
2042
2043 // The Following code fragment (EmitSync & 65536) improves the performance of
2044 // contended applications and contended synchronization microbenchmarks.
2045 // Unfortunately the emission of the code - even though not executed - causes regressions
2046 // in scimark and jetstream, evidently because of $ effects. Replacing the code
2047 // with an equal number of never-executed NOPs results in the same regression.
2048 // We leave it off by default.
2049
2050 if ((EmitSync & 65536) != 0) {
2051 Label LSuccess, LGoSlowPath ;
2052
2053 bind (CheckSucc);
2054
2055 // Optional pre-test ... it's safe to elide this
2056 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2057 jccb(Assembler::zero, LGoSlowPath);
2058
2059 // We have a classic Dekker-style idiom:
2060 // ST m->_owner = 0 ; MEMBAR; LD m->_succ
2061 // There are a number of ways to implement the barrier:
2062 // (1) lock:andl &m->_owner, 0
2063 // is fast, but mask doesn't currently support the "ANDL M,IMM32" form.
2064 // LOCK: ANDL [ebx+Offset(_Owner)-2], 0
2065 // Encodes as 81 31 OFF32 IMM32 or 83 63 OFF8 IMM8
2066 // (2) If supported, an explicit MFENCE is appealing.
2067 // In older IA32 processors MFENCE is slower than lock:add or xchg
2068 // particularly if the write-buffer is full as might be the case if
2069 // if stores closely precede the fence or fence-equivalent instruction.
2070 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2071 // as the situation has changed with Nehalem and Shanghai.
2072 // (3) In lieu of an explicit fence, use lock:addl to the top-of-stack
2073 // The $lines underlying the top-of-stack should be in M-state.
2074 // The locked add instruction is serializing, of course.
2075 // (4) Use xchg, which is serializing
2076 // mov boxReg, 0; xchgl boxReg, [tmpReg + Offset(_owner)-2] also works
2077 // (5) ST m->_owner = 0 and then execute lock:orl &m->_succ, 0.
2078 // The integer condition codes will tell us if succ was 0.
2079 // Since _succ and _owner should reside in the same $line and
2080 // we just stored into _owner, it's likely that the $line
2081 // remains in M-state for the lock:orl.
2082 //
2083 // We currently use (3), although it's likely that switching to (2)
2084 // is correct for the future.
2085
2086 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), NULL_WORD);
2087 if (os::is_MP()) {
2088 lock(); addptr(Address(rsp, 0), 0);
2089 }
2090 // Ratify _succ remains non-null
2091 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), 0);
2092 jccb (Assembler::notZero, LSuccess);
2093
2094 xorptr(boxReg, boxReg); // box is really EAX
2095 if (os::is_MP()) { lock(); }
2096 cmpxchgptr(rsp, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2097 // There's no successor so we tried to regrab the lock with the
2098 // placeholder value. If that didn't work, then another thread
2099 // grabbed the lock so we're done (and exit was a success).
2100 jccb (Assembler::notEqual, LSuccess);
2101 // Since we're low on registers we installed rsp as a placeholding in _owner.
2102 // Now install Self over rsp. This is safe as we're transitioning from
2103 // non-null to non=null
2104 get_thread (boxReg);
2105 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), boxReg);
2106 // Intentional fall-through into LGoSlowPath ...
2107
2108 bind (LGoSlowPath);
2109 orptr(boxReg, 1); // set ICC.ZF=0 to indicate failure
2110 jmpb (DONE_LABEL);
2111
2112 bind (LSuccess);
2113 xorptr(boxReg, boxReg); // set ICC.ZF=1 to indicate success
2114 jmpb (DONE_LABEL);
2115 }
2116
2117 bind (Stacked);
2118 // It's not inflated and it's not recursively stack-locked and it's not biased.
2119 // It must be stack-locked.
2120 // Try to reset the header to displaced header.
2121 // The "box" value on the stack is stable, so we can reload
2122 // and be assured we observe the same value as above.
2123 movptr(tmpReg, Address(boxReg, 0));
2124 if (os::is_MP()) {
2125 lock();
2126 }
2127 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box
2128 // Intention fall-thru into DONE_LABEL
2129
2130 // DONE_LABEL is a hot target - we'd really like to place it at the
2131 // start of cache line by padding with NOPs.
2132 // See the AMD and Intel software optimization manuals for the
2133 // most efficient "long" NOP encodings.
2134 // Unfortunately none of our alignment mechanisms suffice.
2135 if ((EmitSync & 65536) == 0) {
2136 bind (CheckSucc);
2137 }
2138 #else // _LP64
2139 // It's inflated
2140 if (EmitSync & 1024) {
2141 // Emit code to check that _owner == Self
2142 // We could fold the _owner test into subsequent code more efficiently
2143 // than using a stand-alone check, but since _owner checking is off by
2144 // default we don't bother. We also might consider predicating the
2145 // _owner==Self check on Xcheck:jni or running on a debug build.
2146 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2147 xorptr(boxReg, r15_thread);
2148 } else {
2149 xorptr(boxReg, boxReg);
2150 }
2151 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
2152 jccb (Assembler::notZero, DONE_LABEL);
2153 movptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
2154 orptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
2155 jccb (Assembler::notZero, CheckSucc);
2156 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2157 jmpb (DONE_LABEL);
2158
2159 if ((EmitSync & 65536) == 0) {
2160 // Try to avoid passing control into the slow_path ...
2161 Label LSuccess, LGoSlowPath ;
2162 bind (CheckSucc);
2163
2164 // The following optional optimization can be elided if necessary
2165 // Effectively: if (succ == null) goto SlowPath
2166 // The code reduces the window for a race, however,
2167 // and thus benefits performance.
2168 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2169 jccb (Assembler::zero, LGoSlowPath);
2170
2171 if ((EmitSync & 16) && os::is_MP()) {
2172 orptr(boxReg, boxReg);
2173 xchgptr(boxReg, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2174 } else {
2175 movptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), (int32_t)NULL_WORD);
2176 if (os::is_MP()) {
2177 // Memory barrier/fence
2178 // Dekker pivot point -- fulcrum : ST Owner; MEMBAR; LD Succ
2179 // Instead of MFENCE we use a dummy locked add of 0 to the top-of-stack.
2180 // This is faster on Nehalem and AMD Shanghai/Barcelona.
2181 // See https://blogs.oracle.com/dave/entry/instruction_selection_for_volatile_fences
2182 // We might also restructure (ST Owner=0;barrier;LD _Succ) to
2183 // (mov box,0; xchgq box, &m->Owner; LD _succ) .
2184 lock(); addl(Address(rsp, 0), 0);
2185 }
2186 }
2187 cmpptr(Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), (int32_t)NULL_WORD);
2188 jccb (Assembler::notZero, LSuccess);
2189
2190 // Rare inopportune interleaving - race.
2191 // The successor vanished in the small window above.
2192 // The lock is contended -- (cxq|EntryList) != null -- and there's no apparent successor.
2193 // We need to ensure progress and succession.
2194 // Try to reacquire the lock.
2195 // If that fails then the new owner is responsible for succession and this
2196 // thread needs to take no further action and can exit via the fast path (success).
2197 // If the re-acquire succeeds then pass control into the slow path.
2198 // As implemented, this latter mode is horrible because we generated more
2199 // coherence traffic on the lock *and* artifically extended the critical section
2200 // length while by virtue of passing control into the slow path.
2201
2202 // box is really RAX -- the following CMPXCHG depends on that binding
2203 // cmpxchg R,[M] is equivalent to rax = CAS(M,rax,R)
2204 movptr(boxReg, (int32_t)NULL_WORD);
2205 if (os::is_MP()) { lock(); }
2206 cmpxchgptr(r15_thread, Address(tmpReg, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2207 // There's no successor so we tried to regrab the lock.
2208 // If that didn't work, then another thread grabbed the
2209 // lock so we're done (and exit was a success).
2210 jccb (Assembler::notEqual, LSuccess);
2211 // Intentional fall-through into slow-path
2212
2213 bind (LGoSlowPath);
2214 orl (boxReg, 1); // set ICC.ZF=0 to indicate failure
2215 jmpb (DONE_LABEL);
2216
2217 bind (LSuccess);
2218 testl (boxReg, 0); // set ICC.ZF=1 to indicate success
2219 jmpb (DONE_LABEL);
2220 }
2221
2222 bind (Stacked);
2223 movptr(tmpReg, Address (boxReg, 0)); // re-fetch
2224 if (os::is_MP()) { lock(); }
2225 cmpxchgptr(tmpReg, Address(objReg, 0)); // Uses RAX which is box
2226
2227 if (EmitSync & 65536) {
2228 bind (CheckSucc);
2229 }
2230 #endif
2231 bind(DONE_LABEL);
2232 }
2233 }
2234 #endif // COMPILER2
2235
2236 void MacroAssembler::c2bool(Register x) {
2237 // implements x == 0 ? 0 : 1
2238 // note: must only look at least-significant byte of x
2239 // since C-style booleans are stored in one byte
2240 // only! (was bug)
2241 andl(x, 0xFF);
2242 setb(Assembler::notZero, x);
2243 }
2244
2245 // Wouldn't need if AddressLiteral version had new name
2246 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) {
2247 Assembler::call(L, rtype);
2248 }
2249
2250 void MacroAssembler::call(Register entry) {
2251 Assembler::call(entry);
2252 }
2253
2254 void MacroAssembler::call(AddressLiteral entry) {
2255 if (reachable(entry)) {
2256 Assembler::call_literal(entry.target(), entry.rspec());
2257 } else {
2258 lea(rscratch1, entry);
2259 Assembler::call(rscratch1);
2260 }
2261 }
2262
2263 void MacroAssembler::ic_call(address entry) {
2264 RelocationHolder rh = virtual_call_Relocation::spec(pc());
2265 movptr(rax, (intptr_t)Universe::non_oop_word());
2266 call(AddressLiteral(entry, rh));
2267 }
2268
2269 // Implementation of call_VM versions
2270
2271 void MacroAssembler::call_VM(Register oop_result,
2272 address entry_point,
2273 bool check_exceptions) {
2274 Label C, E;
2275 call(C, relocInfo::none);
2276 jmp(E);
2277
2278 bind(C);
2279 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
2280 ret(0);
2281
2282 bind(E);
2283 }
2284
2285 void MacroAssembler::call_VM(Register oop_result,
2286 address entry_point,
2287 Register arg_1,
2288 bool check_exceptions) {
2289 Label C, E;
2290 call(C, relocInfo::none);
2291 jmp(E);
2292
2293 bind(C);
2294 pass_arg1(this, arg_1);
2295 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
2296 ret(0);
2297
2298 bind(E);
2299 }
2300
2301 void MacroAssembler::call_VM(Register oop_result,
2302 address entry_point,
2303 Register arg_1,
2304 Register arg_2,
2305 bool check_exceptions) {
2306 Label C, E;
2307 call(C, relocInfo::none);
2308 jmp(E);
2309
2310 bind(C);
2311
2312 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2313
2314 pass_arg2(this, arg_2);
2315 pass_arg1(this, arg_1);
2316 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
2317 ret(0);
2318
2319 bind(E);
2320 }
2321
2322 void MacroAssembler::call_VM(Register oop_result,
2323 address entry_point,
2324 Register arg_1,
2325 Register arg_2,
2326 Register arg_3,
2327 bool check_exceptions) {
2328 Label C, E;
2329 call(C, relocInfo::none);
2330 jmp(E);
2331
2332 bind(C);
2333
2334 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2335 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2336 pass_arg3(this, arg_3);
2337
2338 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2339 pass_arg2(this, arg_2);
2340
2341 pass_arg1(this, arg_1);
2342 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
2343 ret(0);
2344
2345 bind(E);
2346 }
2347
2348 void MacroAssembler::call_VM(Register oop_result,
2349 Register last_java_sp,
2350 address entry_point,
2351 int number_of_arguments,
2352 bool check_exceptions) {
2353 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2354 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2355 }
2356
2357 void MacroAssembler::call_VM(Register oop_result,
2358 Register last_java_sp,
2359 address entry_point,
2360 Register arg_1,
2361 bool check_exceptions) {
2362 pass_arg1(this, arg_1);
2363 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2364 }
2365
2366 void MacroAssembler::call_VM(Register oop_result,
2367 Register last_java_sp,
2368 address entry_point,
2369 Register arg_1,
2370 Register arg_2,
2371 bool check_exceptions) {
2372
2373 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2374 pass_arg2(this, arg_2);
2375 pass_arg1(this, arg_1);
2376 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2377 }
2378
2379 void MacroAssembler::call_VM(Register oop_result,
2380 Register last_java_sp,
2381 address entry_point,
2382 Register arg_1,
2383 Register arg_2,
2384 Register arg_3,
2385 bool check_exceptions) {
2386 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2387 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2388 pass_arg3(this, arg_3);
2389 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2390 pass_arg2(this, arg_2);
2391 pass_arg1(this, arg_1);
2392 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2393 }
2394
2395 void MacroAssembler::super_call_VM(Register oop_result,
2396 Register last_java_sp,
2397 address entry_point,
2398 int number_of_arguments,
2399 bool check_exceptions) {
2400 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg);
2401 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
2402 }
2403
2404 void MacroAssembler::super_call_VM(Register oop_result,
2405 Register last_java_sp,
2406 address entry_point,
2407 Register arg_1,
2408 bool check_exceptions) {
2409 pass_arg1(this, arg_1);
2410 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
2411 }
2412
2413 void MacroAssembler::super_call_VM(Register oop_result,
2414 Register last_java_sp,
2415 address entry_point,
2416 Register arg_1,
2417 Register arg_2,
2418 bool check_exceptions) {
2419
2420 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2421 pass_arg2(this, arg_2);
2422 pass_arg1(this, arg_1);
2423 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
2424 }
2425
2426 void MacroAssembler::super_call_VM(Register oop_result,
2427 Register last_java_sp,
2428 address entry_point,
2429 Register arg_1,
2430 Register arg_2,
2431 Register arg_3,
2432 bool check_exceptions) {
2433 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2434 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2435 pass_arg3(this, arg_3);
2436 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2437 pass_arg2(this, arg_2);
2438 pass_arg1(this, arg_1);
2439 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
2440 }
2441
2442 void MacroAssembler::call_VM_base(Register oop_result,
2443 Register java_thread,
2444 Register last_java_sp,
2445 address entry_point,
2446 int number_of_arguments,
2447 bool check_exceptions) {
2448 // determine java_thread register
2449 if (!java_thread->is_valid()) {
2450 #ifdef _LP64
2451 java_thread = r15_thread;
2452 #else
2453 java_thread = rdi;
2454 get_thread(java_thread);
2455 #endif // LP64
2456 }
2457 // determine last_java_sp register
2458 if (!last_java_sp->is_valid()) {
2459 last_java_sp = rsp;
2460 }
2461 // debugging support
2462 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
2463 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register"));
2464 #ifdef ASSERT
2465 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
2466 // r12 is the heapbase.
2467 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");)
2468 #endif // ASSERT
2469
2470 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
2471 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
2472
2473 // push java thread (becomes first argument of C function)
2474
2475 NOT_LP64(push(java_thread); number_of_arguments++);
2476 LP64_ONLY(mov(c_rarg0, r15_thread));
2477
2478 // set last Java frame before call
2479 assert(last_java_sp != rbp, "can't use ebp/rbp");
2480
2481 // Only interpreter should have to set fp
2482 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL);
2483
2484 // do the call, remove parameters
2485 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
2486
2487 // restore the thread (cannot use the pushed argument since arguments
2488 // may be overwritten by C code generated by an optimizing compiler);
2489 // however can use the register value directly if it is callee saved.
2490 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) {
2491 // rdi & rsi (also r15) are callee saved -> nothing to do
2492 #ifdef ASSERT
2493 guarantee(java_thread != rax, "change this code");
2494 push(rax);
2495 { Label L;
2496 get_thread(rax);
2497 cmpptr(java_thread, rax);
2498 jcc(Assembler::equal, L);
2499 STOP("MacroAssembler::call_VM_base: rdi not callee saved?");
2500 bind(L);
2501 }
2502 pop(rax);
2503 #endif
2504 } else {
2505 get_thread(java_thread);
2506 }
2507 // reset last Java frame
2508 // Only interpreter should have to clear fp
2509 reset_last_Java_frame(java_thread, true, false);
2510
2511 #ifndef CC_INTERP
2512 // C++ interp handles this in the interpreter
2513 check_and_handle_popframe(java_thread);
2514 check_and_handle_earlyret(java_thread);
2515 #endif /* CC_INTERP */
2516
2517 if (check_exceptions) {
2518 // check for pending exceptions (java_thread is set upon return)
2519 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD);
2520 #ifndef _LP64
2521 jump_cc(Assembler::notEqual,
2522 RuntimeAddress(StubRoutines::forward_exception_entry()));
2523 #else
2524 // This used to conditionally jump to forward_exception however it is
2525 // possible if we relocate that the branch will not reach. So we must jump
2526 // around so we can always reach
2527
2528 Label ok;
2529 jcc(Assembler::equal, ok);
2530 jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2531 bind(ok);
2532 #endif // LP64
2533 }
2534
2535 // get oop result if there is one and reset the value in the thread
2536 if (oop_result->is_valid()) {
2537 get_vm_result(oop_result, java_thread);
2538 }
2539 }
2540
2541 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
2542
2543 // Calculate the value for last_Java_sp
2544 // somewhat subtle. call_VM does an intermediate call
2545 // which places a return address on the stack just under the
2546 // stack pointer as the user finsihed with it. This allows
2547 // use to retrieve last_Java_pc from last_Java_sp[-1].
2548 // On 32bit we then have to push additional args on the stack to accomplish
2549 // the actual requested call. On 64bit call_VM only can use register args
2550 // so the only extra space is the return address that call_VM created.
2551 // This hopefully explains the calculations here.
2552
2553 #ifdef _LP64
2554 // We've pushed one address, correct last_Java_sp
2555 lea(rax, Address(rsp, wordSize));
2556 #else
2557 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize));
2558 #endif // LP64
2559
2560 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions);
2561
2562 }
2563
2564 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
2565 call_VM_leaf_base(entry_point, number_of_arguments);
2566 }
2567
2568 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
2569 pass_arg0(this, arg_0);
2570 call_VM_leaf(entry_point, 1);
2571 }
2572
2573 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2574
2575 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2576 pass_arg1(this, arg_1);
2577 pass_arg0(this, arg_0);
2578 call_VM_leaf(entry_point, 2);
2579 }
2580
2581 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2582 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2583 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2584 pass_arg2(this, arg_2);
2585 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2586 pass_arg1(this, arg_1);
2587 pass_arg0(this, arg_0);
2588 call_VM_leaf(entry_point, 3);
2589 }
2590
2591 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2592 pass_arg0(this, arg_0);
2593 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2594 }
2595
2596 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2597
2598 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2599 pass_arg1(this, arg_1);
2600 pass_arg0(this, arg_0);
2601 MacroAssembler::call_VM_leaf_base(entry_point, 2);
2602 }
2603
2604 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2605 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2606 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2607 pass_arg2(this, arg_2);
2608 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2609 pass_arg1(this, arg_1);
2610 pass_arg0(this, arg_0);
2611 MacroAssembler::call_VM_leaf_base(entry_point, 3);
2612 }
2613
2614 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2615 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg"));
2616 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg"));
2617 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg"));
2618 pass_arg3(this, arg_3);
2619 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg"));
2620 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg"));
2621 pass_arg2(this, arg_2);
2622 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg"));
2623 pass_arg1(this, arg_1);
2624 pass_arg0(this, arg_0);
2625 MacroAssembler::call_VM_leaf_base(entry_point, 4);
2626 }
2627
2628 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
2629 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
2630 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD);
2631 verify_oop(oop_result, "broken oop in call_VM_base");
2632 }
2633
2634 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
2635 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
2636 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD);
2637 }
2638
2639 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2640 }
2641
2642 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2643 }
2644
2645 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) {
2646 if (reachable(src1)) {
2647 cmpl(as_Address(src1), imm);
2648 } else {
2649 lea(rscratch1, src1);
2650 cmpl(Address(rscratch1, 0), imm);
2651 }
2652 }
2653
2654 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) {
2655 assert(!src2.is_lval(), "use cmpptr");
2656 if (reachable(src2)) {
2657 cmpl(src1, as_Address(src2));
2658 } else {
2659 lea(rscratch1, src2);
2660 cmpl(src1, Address(rscratch1, 0));
2661 }
2662 }
2663
2664 void MacroAssembler::cmp32(Register src1, int32_t imm) {
2665 Assembler::cmpl(src1, imm);
2666 }
2667
2668 void MacroAssembler::cmp32(Register src1, Address src2) {
2669 Assembler::cmpl(src1, src2);
2670 }
2671
2672 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2673 ucomisd(opr1, opr2);
2674
2675 Label L;
2676 if (unordered_is_less) {
2677 movl(dst, -1);
2678 jcc(Assembler::parity, L);
2679 jcc(Assembler::below , L);
2680 movl(dst, 0);
2681 jcc(Assembler::equal , L);
2682 increment(dst);
2683 } else { // unordered is greater
2684 movl(dst, 1);
2685 jcc(Assembler::parity, L);
2686 jcc(Assembler::above , L);
2687 movl(dst, 0);
2688 jcc(Assembler::equal , L);
2689 decrementl(dst);
2690 }
2691 bind(L);
2692 }
2693
2694 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) {
2695 ucomiss(opr1, opr2);
2696
2697 Label L;
2698 if (unordered_is_less) {
2699 movl(dst, -1);
2700 jcc(Assembler::parity, L);
2701 jcc(Assembler::below , L);
2702 movl(dst, 0);
2703 jcc(Assembler::equal , L);
2704 increment(dst);
2705 } else { // unordered is greater
2706 movl(dst, 1);
2707 jcc(Assembler::parity, L);
2708 jcc(Assembler::above , L);
2709 movl(dst, 0);
2710 jcc(Assembler::equal , L);
2711 decrementl(dst);
2712 }
2713 bind(L);
2714 }
2715
2716
2717 void MacroAssembler::cmp8(AddressLiteral src1, int imm) {
2718 if (reachable(src1)) {
2719 cmpb(as_Address(src1), imm);
2720 } else {
2721 lea(rscratch1, src1);
2722 cmpb(Address(rscratch1, 0), imm);
2723 }
2724 }
2725
2726 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) {
2727 #ifdef _LP64
2728 if (src2.is_lval()) {
2729 movptr(rscratch1, src2);
2730 Assembler::cmpq(src1, rscratch1);
2731 } else if (reachable(src2)) {
2732 cmpq(src1, as_Address(src2));
2733 } else {
2734 lea(rscratch1, src2);
2735 Assembler::cmpq(src1, Address(rscratch1, 0));
2736 }
2737 #else
2738 if (src2.is_lval()) {
2739 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2740 } else {
2741 cmpl(src1, as_Address(src2));
2742 }
2743 #endif // _LP64
2744 }
2745
2746 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) {
2747 assert(src2.is_lval(), "not a mem-mem compare");
2748 #ifdef _LP64
2749 // moves src2's literal address
2750 movptr(rscratch1, src2);
2751 Assembler::cmpq(src1, rscratch1);
2752 #else
2753 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec());
2754 #endif // _LP64
2755 }
2756
2757 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) {
2758 if (reachable(adr)) {
2759 if (os::is_MP())
2760 lock();
2761 cmpxchgptr(reg, as_Address(adr));
2762 } else {
2763 lea(rscratch1, adr);
2764 if (os::is_MP())
2765 lock();
2766 cmpxchgptr(reg, Address(rscratch1, 0));
2767 }
2768 }
2769
2770 void MacroAssembler::cmpxchgptr(Register reg, Address adr) {
2771 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr));
2772 }
2773
2774 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) {
2775 if (reachable(src)) {
2776 Assembler::comisd(dst, as_Address(src));
2777 } else {
2778 lea(rscratch1, src);
2779 Assembler::comisd(dst, Address(rscratch1, 0));
2780 }
2781 }
2782
2783 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) {
2784 if (reachable(src)) {
2785 Assembler::comiss(dst, as_Address(src));
2786 } else {
2787 lea(rscratch1, src);
2788 Assembler::comiss(dst, Address(rscratch1, 0));
2789 }
2790 }
2791
2792
2793 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) {
2794 Condition negated_cond = negate_condition(cond);
2795 Label L;
2796 jcc(negated_cond, L);
2797 pushf(); // Preserve flags
2798 atomic_incl(counter_addr);
2799 popf();
2800 bind(L);
2801 }
2802
2803 int MacroAssembler::corrected_idivl(Register reg) {
2804 // Full implementation of Java idiv and irem; checks for
2805 // special case as described in JVM spec., p.243 & p.271.
2806 // The function returns the (pc) offset of the idivl
2807 // instruction - may be needed for implicit exceptions.
2808 //
2809 // normal case special case
2810 //
2811 // input : rax,: dividend min_int
2812 // reg: divisor (may not be rax,/rdx) -1
2813 //
2814 // output: rax,: quotient (= rax, idiv reg) min_int
2815 // rdx: remainder (= rax, irem reg) 0
2816 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register");
2817 const int min_int = 0x80000000;
2818 Label normal_case, special_case;
2819
2820 // check for special case
2821 cmpl(rax, min_int);
2822 jcc(Assembler::notEqual, normal_case);
2823 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0)
2824 cmpl(reg, -1);
2825 jcc(Assembler::equal, special_case);
2826
2827 // handle normal case
2828 bind(normal_case);
2829 cdql();
2830 int idivl_offset = offset();
2831 idivl(reg);
2832
2833 // normal and special case exit
2834 bind(special_case);
2835
2836 return idivl_offset;
2837 }
2838
2839
2840
2841 void MacroAssembler::decrementl(Register reg, int value) {
2842 if (value == min_jint) {subl(reg, value) ; return; }
2843 if (value < 0) { incrementl(reg, -value); return; }
2844 if (value == 0) { ; return; }
2845 if (value == 1 && UseIncDec) { decl(reg) ; return; }
2846 /* else */ { subl(reg, value) ; return; }
2847 }
2848
2849 void MacroAssembler::decrementl(Address dst, int value) {
2850 if (value == min_jint) {subl(dst, value) ; return; }
2851 if (value < 0) { incrementl(dst, -value); return; }
2852 if (value == 0) { ; return; }
2853 if (value == 1 && UseIncDec) { decl(dst) ; return; }
2854 /* else */ { subl(dst, value) ; return; }
2855 }
2856
2857 void MacroAssembler::division_with_shift (Register reg, int shift_value) {
2858 assert (shift_value > 0, "illegal shift value");
2859 Label _is_positive;
2860 testl (reg, reg);
2861 jcc (Assembler::positive, _is_positive);
2862 int offset = (1 << shift_value) - 1 ;
2863
2864 if (offset == 1) {
2865 incrementl(reg);
2866 } else {
2867 addl(reg, offset);
2868 }
2869
2870 bind (_is_positive);
2871 sarl(reg, shift_value);
2872 }
2873
2874 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) {
2875 if (reachable(src)) {
2876 Assembler::divsd(dst, as_Address(src));
2877 } else {
2878 lea(rscratch1, src);
2879 Assembler::divsd(dst, Address(rscratch1, 0));
2880 }
2881 }
2882
2883 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) {
2884 if (reachable(src)) {
2885 Assembler::divss(dst, as_Address(src));
2886 } else {
2887 lea(rscratch1, src);
2888 Assembler::divss(dst, Address(rscratch1, 0));
2889 }
2890 }
2891
2892 // !defined(COMPILER2) is because of stupid core builds
2893 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || INCLUDE_JVMCI
2894 void MacroAssembler::empty_FPU_stack() {
2895 if (VM_Version::supports_mmx()) {
2896 emms();
2897 } else {
2898 for (int i = 8; i-- > 0; ) ffree(i);
2899 }
2900 }
2901 #endif // !LP64 || C1 || !C2 || INCLUDE_JVMCI
2902
2903
2904 // Defines obj, preserves var_size_in_bytes
2905 void MacroAssembler::eden_allocate(Register obj,
2906 Register var_size_in_bytes,
2907 int con_size_in_bytes,
2908 Register t1,
2909 Label& slow_case) {
2910 assert(obj == rax, "obj must be in rax, for cmpxchg");
2911 assert_different_registers(obj, var_size_in_bytes, t1);
2912 if (!Universe::heap()->supports_inline_contig_alloc()) {
2913 jmp(slow_case);
2914 } else {
2915 Register end = t1;
2916 Label retry;
2917 bind(retry);
2918 ExternalAddress heap_top((address) Universe::heap()->top_addr());
2919 movptr(obj, heap_top);
2920 if (var_size_in_bytes == noreg) {
2921 lea(end, Address(obj, con_size_in_bytes));
2922 } else {
2923 lea(end, Address(obj, var_size_in_bytes, Address::times_1));
2924 }
2925 // if end < obj then we wrapped around => object too long => slow case
2926 cmpptr(end, obj);
2927 jcc(Assembler::below, slow_case);
2928 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr()));
2929 jcc(Assembler::above, slow_case);
2930 // Compare obj with the top addr, and if still equal, store the new top addr in
2931 // end at the address of the top addr pointer. Sets ZF if was equal, and clears
2932 // it otherwise. Use lock prefix for atomicity on MPs.
2933 locked_cmpxchgptr(end, heap_top);
2934 jcc(Assembler::notEqual, retry);
2935 }
2936 }
2937
2938 void MacroAssembler::enter() {
2939 push(rbp);
2940 mov(rbp, rsp);
2941 }
2942
2943 // A 5 byte nop that is safe for patching (see patch_verified_entry)
2944 void MacroAssembler::fat_nop() {
2945 if (UseAddressNop) {
2946 addr_nop_5();
2947 } else {
2948 emit_int8(0x26); // es:
2949 emit_int8(0x2e); // cs:
2950 emit_int8(0x64); // fs:
2951 emit_int8(0x65); // gs:
2952 emit_int8((unsigned char)0x90);
2953 }
2954 }
2955
2956 void MacroAssembler::fcmp(Register tmp) {
2957 fcmp(tmp, 1, true, true);
2958 }
2959
2960 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) {
2961 assert(!pop_right || pop_left, "usage error");
2962 if (VM_Version::supports_cmov()) {
2963 assert(tmp == noreg, "unneeded temp");
2964 if (pop_left) {
2965 fucomip(index);
2966 } else {
2967 fucomi(index);
2968 }
2969 if (pop_right) {
2970 fpop();
2971 }
2972 } else {
2973 assert(tmp != noreg, "need temp");
2974 if (pop_left) {
2975 if (pop_right) {
2976 fcompp();
2977 } else {
2978 fcomp(index);
2979 }
2980 } else {
2981 fcom(index);
2982 }
2983 // convert FPU condition into eflags condition via rax,
2984 save_rax(tmp);
2985 fwait(); fnstsw_ax();
2986 sahf();
2987 restore_rax(tmp);
2988 }
2989 // condition codes set as follows:
2990 //
2991 // CF (corresponds to C0) if x < y
2992 // PF (corresponds to C2) if unordered
2993 // ZF (corresponds to C3) if x = y
2994 }
2995
2996 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) {
2997 fcmp2int(dst, unordered_is_less, 1, true, true);
2998 }
2999
3000 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) {
3001 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right);
3002 Label L;
3003 if (unordered_is_less) {
3004 movl(dst, -1);
3005 jcc(Assembler::parity, L);
3006 jcc(Assembler::below , L);
3007 movl(dst, 0);
3008 jcc(Assembler::equal , L);
3009 increment(dst);
3010 } else { // unordered is greater
3011 movl(dst, 1);
3012 jcc(Assembler::parity, L);
3013 jcc(Assembler::above , L);
3014 movl(dst, 0);
3015 jcc(Assembler::equal , L);
3016 decrementl(dst);
3017 }
3018 bind(L);
3019 }
3020
3021 void MacroAssembler::fld_d(AddressLiteral src) {
3022 fld_d(as_Address(src));
3023 }
3024
3025 void MacroAssembler::fld_s(AddressLiteral src) {
3026 fld_s(as_Address(src));
3027 }
3028
3029 void MacroAssembler::fld_x(AddressLiteral src) {
3030 Assembler::fld_x(as_Address(src));
3031 }
3032
3033 void MacroAssembler::fldcw(AddressLiteral src) {
3034 Assembler::fldcw(as_Address(src));
3035 }
3036
3037 void MacroAssembler::mulpd(XMMRegister dst, AddressLiteral src) {
3038 if (reachable(src)) {
3039 Assembler::mulpd(dst, as_Address(src));
3040 } else {
3041 lea(rscratch1, src);
3042 Assembler::mulpd(dst, Address(rscratch1, 0));
3043 }
3044 }
3045
3046 void MacroAssembler::pow_exp_core_encoding() {
3047 // kills rax, rcx, rdx
3048 subptr(rsp,sizeof(jdouble));
3049 // computes 2^X. Stack: X ...
3050 // f2xm1 computes 2^X-1 but only operates on -1<=X<=1. Get int(X) and
3051 // keep it on the thread's stack to compute 2^int(X) later
3052 // then compute 2^(X-int(X)) as (2^(X-int(X)-1+1)
3053 // final result is obtained with: 2^X = 2^int(X) * 2^(X-int(X))
3054 fld_s(0); // Stack: X X ...
3055 frndint(); // Stack: int(X) X ...
3056 fsuba(1); // Stack: int(X) X-int(X) ...
3057 fistp_s(Address(rsp,0)); // move int(X) as integer to thread's stack. Stack: X-int(X) ...
3058 f2xm1(); // Stack: 2^(X-int(X))-1 ...
3059 fld1(); // Stack: 1 2^(X-int(X))-1 ...
3060 faddp(1); // Stack: 2^(X-int(X))
3061 // computes 2^(int(X)): add exponent bias (1023) to int(X), then
3062 // shift int(X)+1023 to exponent position.
3063 // Exponent is limited to 11 bits if int(X)+1023 does not fit in 11
3064 // bits, set result to NaN. 0x000 and 0x7FF are reserved exponent
3065 // values so detect them and set result to NaN.
3066 movl(rax,Address(rsp,0));
3067 movl(rcx, -2048); // 11 bit mask and valid NaN binary encoding
3068 addl(rax, 1023);
3069 movl(rdx,rax);
3070 shll(rax,20);
3071 // Check that 0 < int(X)+1023 < 2047. Otherwise set rax to NaN.
3072 addl(rdx,1);
3073 // Check that 1 < int(X)+1023+1 < 2048
3074 // in 3 steps:
3075 // 1- (int(X)+1023+1)&-2048 == 0 => 0 <= int(X)+1023+1 < 2048
3076 // 2- (int(X)+1023+1)&-2048 != 0
3077 // 3- (int(X)+1023+1)&-2048 != 1
3078 // Do 2- first because addl just updated the flags.
3079 cmov32(Assembler::equal,rax,rcx);
3080 cmpl(rdx,1);
3081 cmov32(Assembler::equal,rax,rcx);
3082 testl(rdx,rcx);
3083 cmov32(Assembler::notEqual,rax,rcx);
3084 movl(Address(rsp,4),rax);
3085 movl(Address(rsp,0),0);
3086 fmul_d(Address(rsp,0)); // Stack: 2^X ...
3087 addptr(rsp,sizeof(jdouble));
3088 }
3089
3090 void MacroAssembler::increase_precision() {
3091 subptr(rsp, BytesPerWord);
3092 fnstcw(Address(rsp, 0));
3093 movl(rax, Address(rsp, 0));
3094 orl(rax, 0x300);
3095 push(rax);
3096 fldcw(Address(rsp, 0));
3097 pop(rax);
3098 }
3099
3100 void MacroAssembler::restore_precision() {
3101 fldcw(Address(rsp, 0));
3102 addptr(rsp, BytesPerWord);
3103 }
3104
3105 void MacroAssembler::fast_pow() {
3106 // computes X^Y = 2^(Y * log2(X))
3107 // if fast computation is not possible, result is NaN. Requires
3108 // fallback from user of this macro.
3109 // increase precision for intermediate steps of the computation
3110 BLOCK_COMMENT("fast_pow {");
3111 increase_precision();
3112 fyl2x(); // Stack: (Y*log2(X)) ...
3113 pow_exp_core_encoding(); // Stack: exp(X) ...
3114 restore_precision();
3115 BLOCK_COMMENT("} fast_pow");
3116 }
3117
3118 void MacroAssembler::pow_or_exp(int num_fpu_regs_in_use) {
3119 // kills rax, rcx, rdx
3120 // pow and exp needs 2 extra registers on the fpu stack.
3121 Label slow_case, done;
3122 Register tmp = noreg;
3123 if (!VM_Version::supports_cmov()) {
3124 // fcmp needs a temporary so preserve rdx,
3125 tmp = rdx;
3126 }
3127 Register tmp2 = rax;
3128 Register tmp3 = rcx;
3129
3130 // Stack: X Y
3131 Label x_negative, y_not_2;
3132
3133 static double two = 2.0;
3134 ExternalAddress two_addr((address)&two);
3135
3136 // constant maybe too far on 64 bit
3137 lea(tmp2, two_addr);
3138 fld_d(Address(tmp2, 0)); // Stack: 2 X Y
3139 fcmp(tmp, 2, true, false); // Stack: X Y
3140 jcc(Assembler::parity, y_not_2);
3141 jcc(Assembler::notEqual, y_not_2);
3142
3143 fxch(); fpop(); // Stack: X
3144 fmul(0); // Stack: X*X
3145
3146 jmp(done);
3147
3148 bind(y_not_2);
3149
3150 fldz(); // Stack: 0 X Y
3151 fcmp(tmp, 1, true, false); // Stack: X Y
3152 jcc(Assembler::above, x_negative);
3153
3154 // X >= 0
3155
3156 fld_s(1); // duplicate arguments for runtime call. Stack: Y X Y
3157 fld_s(1); // Stack: X Y X Y
3158 fast_pow(); // Stack: X^Y X Y
3159 fcmp(tmp, 0, false, false); // Stack: X^Y X Y
3160 // X^Y not equal to itself: X^Y is NaN go to slow case.
3161 jcc(Assembler::parity, slow_case);
3162 // get rid of duplicate arguments. Stack: X^Y
3163 if (num_fpu_regs_in_use > 0) {
3164 fxch(); fpop();
3165 fxch(); fpop();
3166 } else {
3167 ffree(2);
3168 ffree(1);
3169 }
3170 jmp(done);
3171
3172 // X <= 0
3173 bind(x_negative);
3174
3175 fld_s(1); // Stack: Y X Y
3176 frndint(); // Stack: int(Y) X Y
3177 fcmp(tmp, 2, false, false); // Stack: int(Y) X Y
3178 jcc(Assembler::notEqual, slow_case);
3179
3180 subptr(rsp, 8);
3181
3182 // For X^Y, when X < 0, Y has to be an integer and the final
3183 // result depends on whether it's odd or even. We just checked
3184 // that int(Y) == Y. We move int(Y) to gp registers as a 64 bit
3185 // integer to test its parity. If int(Y) is huge and doesn't fit
3186 // in the 64 bit integer range, the integer indefinite value will
3187 // end up in the gp registers. Huge numbers are all even, the
3188 // integer indefinite number is even so it's fine.
3189
3190 #ifdef ASSERT
3191 // Let's check we don't end up with an integer indefinite number
3192 // when not expected. First test for huge numbers: check whether
3193 // int(Y)+1 == int(Y) which is true for very large numbers and
3194 // those are all even. A 64 bit integer is guaranteed to not
3195 // overflow for numbers where y+1 != y (when precision is set to
3196 // double precision).
3197 Label y_not_huge;
3198
3199 fld1(); // Stack: 1 int(Y) X Y
3200 fadd(1); // Stack: 1+int(Y) int(Y) X Y
3201
3202 #ifdef _LP64
3203 // trip to memory to force the precision down from double extended
3204 // precision
3205 fstp_d(Address(rsp, 0));
3206 fld_d(Address(rsp, 0));
3207 #endif
3208
3209 fcmp(tmp, 1, true, false); // Stack: int(Y) X Y
3210 #endif
3211
3212 // move int(Y) as 64 bit integer to thread's stack
3213 fistp_d(Address(rsp,0)); // Stack: X Y
3214
3215 #ifdef ASSERT
3216 jcc(Assembler::notEqual, y_not_huge);
3217
3218 // Y is huge so we know it's even. It may not fit in a 64 bit
3219 // integer and we don't want the debug code below to see the
3220 // integer indefinite value so overwrite int(Y) on the thread's
3221 // stack with 0.
3222 movl(Address(rsp, 0), 0);
3223 movl(Address(rsp, 4), 0);
3224
3225 bind(y_not_huge);
3226 #endif
3227
3228 fld_s(1); // duplicate arguments for runtime call. Stack: Y X Y
3229 fld_s(1); // Stack: X Y X Y
3230 fabs(); // Stack: abs(X) Y X Y
3231 fast_pow(); // Stack: abs(X)^Y X Y
3232 fcmp(tmp, 0, false, false); // Stack: abs(X)^Y X Y
3233 // abs(X)^Y not equal to itself: abs(X)^Y is NaN go to slow case.
3234
3235 pop(tmp2);
3236 NOT_LP64(pop(tmp3));
3237 jcc(Assembler::parity, slow_case);
3238
3239 #ifdef ASSERT
3240 // Check that int(Y) is not integer indefinite value (int
3241 // overflow). Shouldn't happen because for values that would
3242 // overflow, 1+int(Y)==Y which was tested earlier.
3243 #ifndef _LP64
3244 {
3245 Label integer;
3246 testl(tmp2, tmp2);
3247 jcc(Assembler::notZero, integer);
3248 cmpl(tmp3, 0x80000000);
3249 jcc(Assembler::notZero, integer);
3250 STOP("integer indefinite value shouldn't be seen here");
3251 bind(integer);
3252 }
3253 #else
3254 {
3255 Label integer;
3256 mov(tmp3, tmp2); // preserve tmp2 for parity check below
3257 shlq(tmp3, 1);
3258 jcc(Assembler::carryClear, integer);
3259 jcc(Assembler::notZero, integer);
3260 STOP("integer indefinite value shouldn't be seen here");
3261 bind(integer);
3262 }
3263 #endif
3264 #endif
3265
3266 // get rid of duplicate arguments. Stack: X^Y
3267 if (num_fpu_regs_in_use > 0) {
3268 fxch(); fpop();
3269 fxch(); fpop();
3270 } else {
3271 ffree(2);
3272 ffree(1);
3273 }
3274
3275 testl(tmp2, 1);
3276 jcc(Assembler::zero, done); // X <= 0, Y even: X^Y = abs(X)^Y
3277 // X <= 0, Y even: X^Y = -abs(X)^Y
3278
3279 fchs(); // Stack: -abs(X)^Y Y
3280 jmp(done);
3281
3282 // slow case: runtime call
3283 bind(slow_case);
3284
3285 fpop(); // pop incorrect result or int(Y)
3286
3287 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dpow), 2, num_fpu_regs_in_use);
3288
3289 // Come here with result in F-TOS
3290 bind(done);
3291 }
3292
3293 void MacroAssembler::fpop() {
3294 ffree();
3295 fincstp();
3296 }
3297
3298 void MacroAssembler::load_float(Address src) {
3299 if (UseSSE >= 1) {
3300 movflt(xmm0, src);
3301 } else {
3302 LP64_ONLY(ShouldNotReachHere());
3303 NOT_LP64(fld_s(src));
3304 }
3305 }
3306
3307 void MacroAssembler::store_float(Address dst) {
3308 if (UseSSE >= 1) {
3309 movflt(dst, xmm0);
3310 } else {
3311 LP64_ONLY(ShouldNotReachHere());
3312 NOT_LP64(fstp_s(dst));
3313 }
3314 }
3315
3316 void MacroAssembler::load_double(Address src) {
3317 if (UseSSE >= 2) {
3318 movdbl(xmm0, src);
3319 } else {
3320 LP64_ONLY(ShouldNotReachHere());
3321 NOT_LP64(fld_d(src));
3322 }
3323 }
3324
3325 void MacroAssembler::store_double(Address dst) {
3326 if (UseSSE >= 2) {
3327 movdbl(dst, xmm0);
3328 } else {
3329 LP64_ONLY(ShouldNotReachHere());
3330 NOT_LP64(fstp_d(dst));
3331 }
3332 }
3333
3334 void MacroAssembler::fremr(Register tmp) {
3335 save_rax(tmp);
3336 { Label L;
3337 bind(L);
3338 fprem();
3339 fwait(); fnstsw_ax();
3340 #ifdef _LP64
3341 testl(rax, 0x400);
3342 jcc(Assembler::notEqual, L);
3343 #else
3344 sahf();
3345 jcc(Assembler::parity, L);
3346 #endif // _LP64
3347 }
3348 restore_rax(tmp);
3349 // Result is in ST0.
3350 // Note: fxch & fpop to get rid of ST1
3351 // (otherwise FPU stack could overflow eventually)
3352 fxch(1);
3353 fpop();
3354 }
3355
3356
3357 void MacroAssembler::incrementl(AddressLiteral dst) {
3358 if (reachable(dst)) {
3359 incrementl(as_Address(dst));
3360 } else {
3361 lea(rscratch1, dst);
3362 incrementl(Address(rscratch1, 0));
3363 }
3364 }
3365
3366 void MacroAssembler::incrementl(ArrayAddress dst) {
3367 incrementl(as_Address(dst));
3368 }
3369
3370 void MacroAssembler::incrementl(Register reg, int value) {
3371 if (value == min_jint) {addl(reg, value) ; return; }
3372 if (value < 0) { decrementl(reg, -value); return; }
3373 if (value == 0) { ; return; }
3374 if (value == 1 && UseIncDec) { incl(reg) ; return; }
3375 /* else */ { addl(reg, value) ; return; }
3376 }
3377
3378 void MacroAssembler::incrementl(Address dst, int value) {
3379 if (value == min_jint) {addl(dst, value) ; return; }
3380 if (value < 0) { decrementl(dst, -value); return; }
3381 if (value == 0) { ; return; }
3382 if (value == 1 && UseIncDec) { incl(dst) ; return; }
3383 /* else */ { addl(dst, value) ; return; }
3384 }
3385
3386 void MacroAssembler::jump(AddressLiteral dst) {
3387 if (reachable(dst)) {
3388 jmp_literal(dst.target(), dst.rspec());
3389 } else {
3390 lea(rscratch1, dst);
3391 jmp(rscratch1);
3392 }
3393 }
3394
3395 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) {
3396 if (reachable(dst)) {
3397 InstructionMark im(this);
3398 relocate(dst.reloc());
3399 const int short_size = 2;
3400 const int long_size = 6;
3401 int offs = (intptr_t)dst.target() - ((intptr_t)pc());
3402 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) {
3403 // 0111 tttn #8-bit disp
3404 emit_int8(0x70 | cc);
3405 emit_int8((offs - short_size) & 0xFF);
3406 } else {
3407 // 0000 1111 1000 tttn #32-bit disp
3408 emit_int8(0x0F);
3409 emit_int8((unsigned char)(0x80 | cc));
3410 emit_int32(offs - long_size);
3411 }
3412 } else {
3413 #ifdef ASSERT
3414 warning("reversing conditional branch");
3415 #endif /* ASSERT */
3416 Label skip;
3417 jccb(reverse[cc], skip);
3418 lea(rscratch1, dst);
3419 Assembler::jmp(rscratch1);
3420 bind(skip);
3421 }
3422 }
3423
3424 void MacroAssembler::ldmxcsr(AddressLiteral src) {
3425 if (reachable(src)) {
3426 Assembler::ldmxcsr(as_Address(src));
3427 } else {
3428 lea(rscratch1, src);
3429 Assembler::ldmxcsr(Address(rscratch1, 0));
3430 }
3431 }
3432
3433 int MacroAssembler::load_signed_byte(Register dst, Address src) {
3434 int off;
3435 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3436 off = offset();
3437 movsbl(dst, src); // movsxb
3438 } else {
3439 off = load_unsigned_byte(dst, src);
3440 shll(dst, 24);
3441 sarl(dst, 24);
3442 }
3443 return off;
3444 }
3445
3446 // Note: load_signed_short used to be called load_signed_word.
3447 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler
3448 // manual, which means 16 bits, that usage is found nowhere in HotSpot code.
3449 // The term "word" in HotSpot means a 32- or 64-bit machine word.
3450 int MacroAssembler::load_signed_short(Register dst, Address src) {
3451 int off;
3452 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3453 // This is dubious to me since it seems safe to do a signed 16 => 64 bit
3454 // version but this is what 64bit has always done. This seems to imply
3455 // that users are only using 32bits worth.
3456 off = offset();
3457 movswl(dst, src); // movsxw
3458 } else {
3459 off = load_unsigned_short(dst, src);
3460 shll(dst, 16);
3461 sarl(dst, 16);
3462 }
3463 return off;
3464 }
3465
3466 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
3467 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3468 // and "3.9 Partial Register Penalties", p. 22).
3469 int off;
3470 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) {
3471 off = offset();
3472 movzbl(dst, src); // movzxb
3473 } else {
3474 xorl(dst, dst);
3475 off = offset();
3476 movb(dst, src);
3477 }
3478 return off;
3479 }
3480
3481 // Note: load_unsigned_short used to be called load_unsigned_word.
3482 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
3483 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16,
3484 // and "3.9 Partial Register Penalties", p. 22).
3485 int off;
3486 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) {
3487 off = offset();
3488 movzwl(dst, src); // movzxw
3489 } else {
3490 xorl(dst, dst);
3491 off = offset();
3492 movw(dst, src);
3493 }
3494 return off;
3495 }
3496
3497 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
3498 switch (size_in_bytes) {
3499 #ifndef _LP64
3500 case 8:
3501 assert(dst2 != noreg, "second dest register required");
3502 movl(dst, src);
3503 movl(dst2, src.plus_disp(BytesPerInt));
3504 break;
3505 #else
3506 case 8: movq(dst, src); break;
3507 #endif
3508 case 4: movl(dst, src); break;
3509 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
3510 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
3511 default: ShouldNotReachHere();
3512 }
3513 }
3514
3515 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
3516 switch (size_in_bytes) {
3517 #ifndef _LP64
3518 case 8:
3519 assert(src2 != noreg, "second source register required");
3520 movl(dst, src);
3521 movl(dst.plus_disp(BytesPerInt), src2);
3522 break;
3523 #else
3524 case 8: movq(dst, src); break;
3525 #endif
3526 case 4: movl(dst, src); break;
3527 case 2: movw(dst, src); break;
3528 case 1: movb(dst, src); break;
3529 default: ShouldNotReachHere();
3530 }
3531 }
3532
3533 void MacroAssembler::mov32(AddressLiteral dst, Register src) {
3534 if (reachable(dst)) {
3535 movl(as_Address(dst), src);
3536 } else {
3537 lea(rscratch1, dst);
3538 movl(Address(rscratch1, 0), src);
3539 }
3540 }
3541
3542 void MacroAssembler::mov32(Register dst, AddressLiteral src) {
3543 if (reachable(src)) {
3544 movl(dst, as_Address(src));
3545 } else {
3546 lea(rscratch1, src);
3547 movl(dst, Address(rscratch1, 0));
3548 }
3549 }
3550
3551 // C++ bool manipulation
3552
3553 void MacroAssembler::movbool(Register dst, Address src) {
3554 if(sizeof(bool) == 1)
3555 movb(dst, src);
3556 else if(sizeof(bool) == 2)
3557 movw(dst, src);
3558 else if(sizeof(bool) == 4)
3559 movl(dst, src);
3560 else
3561 // unsupported
3562 ShouldNotReachHere();
3563 }
3564
3565 void MacroAssembler::movbool(Address dst, bool boolconst) {
3566 if(sizeof(bool) == 1)
3567 movb(dst, (int) boolconst);
3568 else if(sizeof(bool) == 2)
3569 movw(dst, (int) boolconst);
3570 else if(sizeof(bool) == 4)
3571 movl(dst, (int) boolconst);
3572 else
3573 // unsupported
3574 ShouldNotReachHere();
3575 }
3576
3577 void MacroAssembler::movbool(Address dst, Register src) {
3578 if(sizeof(bool) == 1)
3579 movb(dst, src);
3580 else if(sizeof(bool) == 2)
3581 movw(dst, src);
3582 else if(sizeof(bool) == 4)
3583 movl(dst, src);
3584 else
3585 // unsupported
3586 ShouldNotReachHere();
3587 }
3588
3589 void MacroAssembler::movbyte(ArrayAddress dst, int src) {
3590 movb(as_Address(dst), src);
3591 }
3592
3593 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) {
3594 if (reachable(src)) {
3595 movdl(dst, as_Address(src));
3596 } else {
3597 lea(rscratch1, src);
3598 movdl(dst, Address(rscratch1, 0));
3599 }
3600 }
3601
3602 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) {
3603 if (reachable(src)) {
3604 movq(dst, as_Address(src));
3605 } else {
3606 lea(rscratch1, src);
3607 movq(dst, Address(rscratch1, 0));
3608 }
3609 }
3610
3611 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) {
3612 if (reachable(src)) {
3613 if (UseXmmLoadAndClearUpper) {
3614 movsd (dst, as_Address(src));
3615 } else {
3616 movlpd(dst, as_Address(src));
3617 }
3618 } else {
3619 lea(rscratch1, src);
3620 if (UseXmmLoadAndClearUpper) {
3621 movsd (dst, Address(rscratch1, 0));
3622 } else {
3623 movlpd(dst, Address(rscratch1, 0));
3624 }
3625 }
3626 }
3627
3628 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) {
3629 if (reachable(src)) {
3630 movss(dst, as_Address(src));
3631 } else {
3632 lea(rscratch1, src);
3633 movss(dst, Address(rscratch1, 0));
3634 }
3635 }
3636
3637 void MacroAssembler::movptr(Register dst, Register src) {
3638 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3639 }
3640
3641 void MacroAssembler::movptr(Register dst, Address src) {
3642 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3643 }
3644
3645 // src should NEVER be a real pointer. Use AddressLiteral for true pointers
3646 void MacroAssembler::movptr(Register dst, intptr_t src) {
3647 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src));
3648 }
3649
3650 void MacroAssembler::movptr(Address dst, Register src) {
3651 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src));
3652 }
3653
3654 void MacroAssembler::movdqu(Address dst, XMMRegister src) {
3655 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) {
3656 Assembler::vextractf32x4h(dst, src, 0);
3657 } else {
3658 Assembler::movdqu(dst, src);
3659 }
3660 }
3661
3662 void MacroAssembler::movdqu(XMMRegister dst, Address src) {
3663 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) {
3664 Assembler::vinsertf32x4h(dst, src, 0);
3665 } else {
3666 Assembler::movdqu(dst, src);
3667 }
3668 }
3669
3670 void MacroAssembler::movdqu(XMMRegister dst, XMMRegister src) {
3671 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) {
3672 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit);
3673 } else {
3674 Assembler::movdqu(dst, src);
3675 }
3676 }
3677
3678 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src) {
3679 if (reachable(src)) {
3680 movdqu(dst, as_Address(src));
3681 } else {
3682 lea(rscratch1, src);
3683 movdqu(dst, Address(rscratch1, 0));
3684 }
3685 }
3686
3687 void MacroAssembler::vmovdqu(Address dst, XMMRegister src) {
3688 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (src->encoding() > 15)) {
3689 Assembler::vextractf64x4h(dst, src, 0);
3690 } else {
3691 Assembler::vmovdqu(dst, src);
3692 }
3693 }
3694
3695 void MacroAssembler::vmovdqu(XMMRegister dst, Address src) {
3696 if (UseAVX > 2 && !VM_Version::supports_avx512vl() && (dst->encoding() > 15)) {
3697 Assembler::vinsertf64x4h(dst, src, 0);
3698 } else {
3699 Assembler::vmovdqu(dst, src);
3700 }
3701 }
3702
3703 void MacroAssembler::vmovdqu(XMMRegister dst, XMMRegister src) {
3704 if (UseAVX > 2 && !VM_Version::supports_avx512vl()) {
3705 Assembler::evmovdqul(dst, src, Assembler::AVX_512bit);
3706 }
3707 else {
3708 Assembler::vmovdqu(dst, src);
3709 }
3710 }
3711
3712 void MacroAssembler::vmovdqu(XMMRegister dst, AddressLiteral src) {
3713 if (reachable(src)) {
3714 vmovdqu(dst, as_Address(src));
3715 }
3716 else {
3717 lea(rscratch1, src);
3718 vmovdqu(dst, Address(rscratch1, 0));
3719 }
3720 }
3721
3722 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) {
3723 if (reachable(src)) {
3724 Assembler::movdqa(dst, as_Address(src));
3725 } else {
3726 lea(rscratch1, src);
3727 Assembler::movdqa(dst, Address(rscratch1, 0));
3728 }
3729 }
3730
3731 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) {
3732 if (reachable(src)) {
3733 Assembler::movsd(dst, as_Address(src));
3734 } else {
3735 lea(rscratch1, src);
3736 Assembler::movsd(dst, Address(rscratch1, 0));
3737 }
3738 }
3739
3740 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) {
3741 if (reachable(src)) {
3742 Assembler::movss(dst, as_Address(src));
3743 } else {
3744 lea(rscratch1, src);
3745 Assembler::movss(dst, Address(rscratch1, 0));
3746 }
3747 }
3748
3749 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) {
3750 if (reachable(src)) {
3751 Assembler::mulsd(dst, as_Address(src));
3752 } else {
3753 lea(rscratch1, src);
3754 Assembler::mulsd(dst, Address(rscratch1, 0));
3755 }
3756 }
3757
3758 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) {
3759 if (reachable(src)) {
3760 Assembler::mulss(dst, as_Address(src));
3761 } else {
3762 lea(rscratch1, src);
3763 Assembler::mulss(dst, Address(rscratch1, 0));
3764 }
3765 }
3766
3767 void MacroAssembler::null_check(Register reg, int offset) {
3768 if (needs_explicit_null_check(offset)) {
3769 // provoke OS NULL exception if reg = NULL by
3770 // accessing M[reg] w/o changing any (non-CC) registers
3771 // NOTE: cmpl is plenty here to provoke a segv
3772 cmpptr(rax, Address(reg, 0));
3773 // Note: should probably use testl(rax, Address(reg, 0));
3774 // may be shorter code (however, this version of
3775 // testl needs to be implemented first)
3776 } else {
3777 // nothing to do, (later) access of M[reg + offset]
3778 // will provoke OS NULL exception if reg = NULL
3779 }
3780 }
3781
3782 void MacroAssembler::os_breakpoint() {
3783 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability
3784 // (e.g., MSVC can't call ps() otherwise)
3785 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint)));
3786 }
3787
3788 #ifdef _LP64
3789 #define XSTATE_BV 0x200
3790 #endif
3791
3792 void MacroAssembler::pop_CPU_state() {
3793 pop_FPU_state();
3794 pop_IU_state();
3795 }
3796
3797 void MacroAssembler::pop_FPU_state() {
3798 #ifndef _LP64
3799 frstor(Address(rsp, 0));
3800 #else
3801 fxrstor(Address(rsp, 0));
3802 #endif
3803 addptr(rsp, FPUStateSizeInWords * wordSize);
3804 }
3805
3806 void MacroAssembler::pop_IU_state() {
3807 popa();
3808 LP64_ONLY(addq(rsp, 8));
3809 popf();
3810 }
3811
3812 // Save Integer and Float state
3813 // Warning: Stack must be 16 byte aligned (64bit)
3814 void MacroAssembler::push_CPU_state() {
3815 push_IU_state();
3816 push_FPU_state();
3817 }
3818
3819 void MacroAssembler::push_FPU_state() {
3820 subptr(rsp, FPUStateSizeInWords * wordSize);
3821 #ifndef _LP64
3822 fnsave(Address(rsp, 0));
3823 fwait();
3824 #else
3825 fxsave(Address(rsp, 0));
3826 #endif // LP64
3827 }
3828
3829 void MacroAssembler::push_IU_state() {
3830 // Push flags first because pusha kills them
3831 pushf();
3832 // Make sure rsp stays 16-byte aligned
3833 LP64_ONLY(subq(rsp, 8));
3834 pusha();
3835 }
3836
3837 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp, bool clear_pc) {
3838 // determine java_thread register
3839 if (!java_thread->is_valid()) {
3840 java_thread = rdi;
3841 get_thread(java_thread);
3842 }
3843 // we must set sp to zero to clear frame
3844 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD);
3845 if (clear_fp) {
3846 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3847 }
3848
3849 if (clear_pc)
3850 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD);
3851
3852 }
3853
3854 void MacroAssembler::restore_rax(Register tmp) {
3855 if (tmp == noreg) pop(rax);
3856 else if (tmp != rax) mov(rax, tmp);
3857 }
3858
3859 void MacroAssembler::round_to(Register reg, int modulus) {
3860 addptr(reg, modulus - 1);
3861 andptr(reg, -modulus);
3862 }
3863
3864 void MacroAssembler::save_rax(Register tmp) {
3865 if (tmp == noreg) push(rax);
3866 else if (tmp != rax) mov(tmp, rax);
3867 }
3868
3869 // Write serialization page so VM thread can do a pseudo remote membar.
3870 // We use the current thread pointer to calculate a thread specific
3871 // offset to write to within the page. This minimizes bus traffic
3872 // due to cache line collision.
3873 void MacroAssembler::serialize_memory(Register thread, Register tmp) {
3874 movl(tmp, thread);
3875 shrl(tmp, os::get_serialize_page_shift_count());
3876 andl(tmp, (os::vm_page_size() - sizeof(int)));
3877
3878 Address index(noreg, tmp, Address::times_1);
3879 ExternalAddress page(os::get_memory_serialize_page());
3880
3881 // Size of store must match masking code above
3882 movl(as_Address(ArrayAddress(page, index)), tmp);
3883 }
3884
3885 // Calls to C land
3886 //
3887 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded
3888 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
3889 // has to be reset to 0. This is required to allow proper stack traversal.
3890 void MacroAssembler::set_last_Java_frame(Register java_thread,
3891 Register last_java_sp,
3892 Register last_java_fp,
3893 address last_java_pc) {
3894 // determine java_thread register
3895 if (!java_thread->is_valid()) {
3896 java_thread = rdi;
3897 get_thread(java_thread);
3898 }
3899 // determine last_java_sp register
3900 if (!last_java_sp->is_valid()) {
3901 last_java_sp = rsp;
3902 }
3903
3904 // last_java_fp is optional
3905
3906 if (last_java_fp->is_valid()) {
3907 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp);
3908 }
3909
3910 // last_java_pc is optional
3911
3912 if (last_java_pc != NULL) {
3913 lea(Address(java_thread,
3914 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()),
3915 InternalAddress(last_java_pc));
3916
3917 }
3918 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp);
3919 }
3920
3921 void MacroAssembler::shlptr(Register dst, int imm8) {
3922 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8));
3923 }
3924
3925 void MacroAssembler::shrptr(Register dst, int imm8) {
3926 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8));
3927 }
3928
3929 void MacroAssembler::sign_extend_byte(Register reg) {
3930 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) {
3931 movsbl(reg, reg); // movsxb
3932 } else {
3933 shll(reg, 24);
3934 sarl(reg, 24);
3935 }
3936 }
3937
3938 void MacroAssembler::sign_extend_short(Register reg) {
3939 if (LP64_ONLY(true ||) VM_Version::is_P6()) {
3940 movswl(reg, reg); // movsxw
3941 } else {
3942 shll(reg, 16);
3943 sarl(reg, 16);
3944 }
3945 }
3946
3947 void MacroAssembler::testl(Register dst, AddressLiteral src) {
3948 assert(reachable(src), "Address should be reachable");
3949 testl(dst, as_Address(src));
3950 }
3951
3952 void MacroAssembler::pcmpeqb(XMMRegister dst, XMMRegister src) {
3953 int dst_enc = dst->encoding();
3954 int src_enc = src->encoding();
3955 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3956 Assembler::pcmpeqb(dst, src);
3957 } else if ((dst_enc < 16) && (src_enc < 16)) {
3958 Assembler::pcmpeqb(dst, src);
3959 } else if (src_enc < 16) {
3960 subptr(rsp, 64);
3961 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3962 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
3963 Assembler::pcmpeqb(xmm0, src);
3964 movdqu(dst, xmm0);
3965 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3966 addptr(rsp, 64);
3967 } else if (dst_enc < 16) {
3968 subptr(rsp, 64);
3969 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3970 evmovdqul(xmm0, src, Assembler::AVX_512bit);
3971 Assembler::pcmpeqb(dst, xmm0);
3972 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3973 addptr(rsp, 64);
3974 } else {
3975 subptr(rsp, 64);
3976 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
3977 subptr(rsp, 64);
3978 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
3979 movdqu(xmm0, src);
3980 movdqu(xmm1, dst);
3981 Assembler::pcmpeqb(xmm1, xmm0);
3982 movdqu(dst, xmm1);
3983 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
3984 addptr(rsp, 64);
3985 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
3986 addptr(rsp, 64);
3987 }
3988 }
3989
3990 void MacroAssembler::pcmpeqw(XMMRegister dst, XMMRegister src) {
3991 int dst_enc = dst->encoding();
3992 int src_enc = src->encoding();
3993 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
3994 Assembler::pcmpeqw(dst, src);
3995 } else if ((dst_enc < 16) && (src_enc < 16)) {
3996 Assembler::pcmpeqw(dst, src);
3997 } else if (src_enc < 16) {
3998 subptr(rsp, 64);
3999 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4000 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4001 Assembler::pcmpeqw(xmm0, src);
4002 movdqu(dst, xmm0);
4003 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4004 addptr(rsp, 64);
4005 } else if (dst_enc < 16) {
4006 subptr(rsp, 64);
4007 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4008 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4009 Assembler::pcmpeqw(dst, xmm0);
4010 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4011 addptr(rsp, 64);
4012 } else {
4013 subptr(rsp, 64);
4014 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4015 subptr(rsp, 64);
4016 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4017 movdqu(xmm0, src);
4018 movdqu(xmm1, dst);
4019 Assembler::pcmpeqw(xmm1, xmm0);
4020 movdqu(dst, xmm1);
4021 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4022 addptr(rsp, 64);
4023 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4024 addptr(rsp, 64);
4025 }
4026 }
4027
4028 void MacroAssembler::pcmpestri(XMMRegister dst, Address src, int imm8) {
4029 int dst_enc = dst->encoding();
4030 if (dst_enc < 16) {
4031 Assembler::pcmpestri(dst, src, imm8);
4032 } else {
4033 subptr(rsp, 64);
4034 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4035 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4036 Assembler::pcmpestri(xmm0, src, imm8);
4037 movdqu(dst, xmm0);
4038 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4039 addptr(rsp, 64);
4040 }
4041 }
4042
4043 void MacroAssembler::pcmpestri(XMMRegister dst, XMMRegister src, int imm8) {
4044 int dst_enc = dst->encoding();
4045 int src_enc = src->encoding();
4046 if ((dst_enc < 16) && (src_enc < 16)) {
4047 Assembler::pcmpestri(dst, src, imm8);
4048 } else if (src_enc < 16) {
4049 subptr(rsp, 64);
4050 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4051 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4052 Assembler::pcmpestri(xmm0, src, imm8);
4053 movdqu(dst, xmm0);
4054 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4055 addptr(rsp, 64);
4056 } else if (dst_enc < 16) {
4057 subptr(rsp, 64);
4058 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4059 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4060 Assembler::pcmpestri(dst, xmm0, imm8);
4061 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4062 addptr(rsp, 64);
4063 } else {
4064 subptr(rsp, 64);
4065 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4066 subptr(rsp, 64);
4067 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4068 movdqu(xmm0, src);
4069 movdqu(xmm1, dst);
4070 Assembler::pcmpestri(xmm1, xmm0, imm8);
4071 movdqu(dst, xmm1);
4072 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4073 addptr(rsp, 64);
4074 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4075 addptr(rsp, 64);
4076 }
4077 }
4078
4079 void MacroAssembler::pmovzxbw(XMMRegister dst, XMMRegister src) {
4080 int dst_enc = dst->encoding();
4081 int src_enc = src->encoding();
4082 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4083 Assembler::pmovzxbw(dst, src);
4084 } else if ((dst_enc < 16) && (src_enc < 16)) {
4085 Assembler::pmovzxbw(dst, src);
4086 } else if (src_enc < 16) {
4087 subptr(rsp, 64);
4088 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4089 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4090 Assembler::pmovzxbw(xmm0, src);
4091 movdqu(dst, xmm0);
4092 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4093 addptr(rsp, 64);
4094 } else if (dst_enc < 16) {
4095 subptr(rsp, 64);
4096 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4097 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4098 Assembler::pmovzxbw(dst, xmm0);
4099 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4100 addptr(rsp, 64);
4101 } else {
4102 subptr(rsp, 64);
4103 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4104 subptr(rsp, 64);
4105 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4106 movdqu(xmm0, src);
4107 movdqu(xmm1, dst);
4108 Assembler::pmovzxbw(xmm1, xmm0);
4109 movdqu(dst, xmm1);
4110 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4111 addptr(rsp, 64);
4112 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4113 addptr(rsp, 64);
4114 }
4115 }
4116
4117 void MacroAssembler::pmovzxbw(XMMRegister dst, Address src) {
4118 int dst_enc = dst->encoding();
4119 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4120 Assembler::pmovzxbw(dst, src);
4121 } else if (dst_enc < 16) {
4122 Assembler::pmovzxbw(dst, src);
4123 } else {
4124 subptr(rsp, 64);
4125 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4126 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4127 Assembler::pmovzxbw(xmm0, src);
4128 movdqu(dst, xmm0);
4129 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4130 addptr(rsp, 64);
4131 }
4132 }
4133
4134 void MacroAssembler::pmovmskb(Register dst, XMMRegister src) {
4135 int src_enc = src->encoding();
4136 if (src_enc < 16) {
4137 Assembler::pmovmskb(dst, src);
4138 } else {
4139 subptr(rsp, 64);
4140 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4141 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4142 Assembler::pmovmskb(dst, xmm0);
4143 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4144 addptr(rsp, 64);
4145 }
4146 }
4147
4148 void MacroAssembler::ptest(XMMRegister dst, XMMRegister src) {
4149 int dst_enc = dst->encoding();
4150 int src_enc = src->encoding();
4151 if ((dst_enc < 16) && (src_enc < 16)) {
4152 Assembler::ptest(dst, src);
4153 } else if (src_enc < 16) {
4154 subptr(rsp, 64);
4155 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4156 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4157 Assembler::ptest(xmm0, src);
4158 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4159 addptr(rsp, 64);
4160 } else if (dst_enc < 16) {
4161 subptr(rsp, 64);
4162 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4163 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4164 Assembler::ptest(dst, xmm0);
4165 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4166 addptr(rsp, 64);
4167 } else {
4168 subptr(rsp, 64);
4169 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4170 subptr(rsp, 64);
4171 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4172 movdqu(xmm0, src);
4173 movdqu(xmm1, dst);
4174 Assembler::ptest(xmm1, xmm0);
4175 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4176 addptr(rsp, 64);
4177 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4178 addptr(rsp, 64);
4179 }
4180 }
4181
4182 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) {
4183 if (reachable(src)) {
4184 Assembler::sqrtsd(dst, as_Address(src));
4185 } else {
4186 lea(rscratch1, src);
4187 Assembler::sqrtsd(dst, Address(rscratch1, 0));
4188 }
4189 }
4190
4191 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) {
4192 if (reachable(src)) {
4193 Assembler::sqrtss(dst, as_Address(src));
4194 } else {
4195 lea(rscratch1, src);
4196 Assembler::sqrtss(dst, Address(rscratch1, 0));
4197 }
4198 }
4199
4200 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) {
4201 if (reachable(src)) {
4202 Assembler::subsd(dst, as_Address(src));
4203 } else {
4204 lea(rscratch1, src);
4205 Assembler::subsd(dst, Address(rscratch1, 0));
4206 }
4207 }
4208
4209 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) {
4210 if (reachable(src)) {
4211 Assembler::subss(dst, as_Address(src));
4212 } else {
4213 lea(rscratch1, src);
4214 Assembler::subss(dst, Address(rscratch1, 0));
4215 }
4216 }
4217
4218 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) {
4219 if (reachable(src)) {
4220 Assembler::ucomisd(dst, as_Address(src));
4221 } else {
4222 lea(rscratch1, src);
4223 Assembler::ucomisd(dst, Address(rscratch1, 0));
4224 }
4225 }
4226
4227 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) {
4228 if (reachable(src)) {
4229 Assembler::ucomiss(dst, as_Address(src));
4230 } else {
4231 lea(rscratch1, src);
4232 Assembler::ucomiss(dst, Address(rscratch1, 0));
4233 }
4234 }
4235
4236 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) {
4237 // Used in sign-bit flipping with aligned address.
4238 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
4239 if (reachable(src)) {
4240 Assembler::xorpd(dst, as_Address(src));
4241 } else {
4242 lea(rscratch1, src);
4243 Assembler::xorpd(dst, Address(rscratch1, 0));
4244 }
4245 }
4246
4247 void MacroAssembler::xorpd(XMMRegister dst, XMMRegister src) {
4248 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
4249 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
4250 }
4251 else {
4252 Assembler::xorpd(dst, src);
4253 }
4254 }
4255
4256 void MacroAssembler::xorps(XMMRegister dst, XMMRegister src) {
4257 if (UseAVX > 2 && !VM_Version::supports_avx512dq() && (dst->encoding() == src->encoding())) {
4258 Assembler::vpxor(dst, dst, src, Assembler::AVX_512bit);
4259 } else {
4260 Assembler::xorps(dst, src);
4261 }
4262 }
4263
4264 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) {
4265 // Used in sign-bit flipping with aligned address.
4266 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes");
4267 if (reachable(src)) {
4268 Assembler::xorps(dst, as_Address(src));
4269 } else {
4270 lea(rscratch1, src);
4271 Assembler::xorps(dst, Address(rscratch1, 0));
4272 }
4273 }
4274
4275 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) {
4276 // Used in sign-bit flipping with aligned address.
4277 bool aligned_adr = (((intptr_t)src.target() & 15) == 0);
4278 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes");
4279 if (reachable(src)) {
4280 Assembler::pshufb(dst, as_Address(src));
4281 } else {
4282 lea(rscratch1, src);
4283 Assembler::pshufb(dst, Address(rscratch1, 0));
4284 }
4285 }
4286
4287 // AVX 3-operands instructions
4288
4289 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4290 if (reachable(src)) {
4291 vaddsd(dst, nds, as_Address(src));
4292 } else {
4293 lea(rscratch1, src);
4294 vaddsd(dst, nds, Address(rscratch1, 0));
4295 }
4296 }
4297
4298 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
4299 if (reachable(src)) {
4300 vaddss(dst, nds, as_Address(src));
4301 } else {
4302 lea(rscratch1, src);
4303 vaddss(dst, nds, Address(rscratch1, 0));
4304 }
4305 }
4306
4307 void MacroAssembler::vabsss(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
4308 int dst_enc = dst->encoding();
4309 int nds_enc = nds->encoding();
4310 int src_enc = src->encoding();
4311 if ((dst_enc < 16) && (nds_enc < 16)) {
4312 vandps(dst, nds, negate_field, vector_len);
4313 } else if ((src_enc < 16) && (dst_enc < 16)) {
4314 movss(src, nds);
4315 vandps(dst, src, negate_field, vector_len);
4316 } else if (src_enc < 16) {
4317 movss(src, nds);
4318 vandps(src, src, negate_field, vector_len);
4319 movss(dst, src);
4320 } else if (dst_enc < 16) {
4321 movdqu(src, xmm0);
4322 movss(xmm0, nds);
4323 vandps(dst, xmm0, negate_field, vector_len);
4324 movdqu(xmm0, src);
4325 } else if (nds_enc < 16) {
4326 movdqu(src, xmm0);
4327 vandps(xmm0, nds, negate_field, vector_len);
4328 movss(dst, xmm0);
4329 movdqu(xmm0, src);
4330 } else {
4331 movdqu(src, xmm0);
4332 movss(xmm0, nds);
4333 vandps(xmm0, xmm0, negate_field, vector_len);
4334 movss(dst, xmm0);
4335 movdqu(xmm0, src);
4336 }
4337 }
4338
4339 void MacroAssembler::vabssd(XMMRegister dst, XMMRegister nds, XMMRegister src, AddressLiteral negate_field, int vector_len) {
4340 int dst_enc = dst->encoding();
4341 int nds_enc = nds->encoding();
4342 int src_enc = src->encoding();
4343 if ((dst_enc < 16) && (nds_enc < 16)) {
4344 vandpd(dst, nds, negate_field, vector_len);
4345 } else if ((src_enc < 16) && (dst_enc < 16)) {
4346 movsd(src, nds);
4347 vandpd(dst, src, negate_field, vector_len);
4348 } else if (src_enc < 16) {
4349 movsd(src, nds);
4350 vandpd(src, src, negate_field, vector_len);
4351 movsd(dst, src);
4352 } else if (dst_enc < 16) {
4353 movdqu(src, xmm0);
4354 movsd(xmm0, nds);
4355 vandpd(dst, xmm0, negate_field, vector_len);
4356 movdqu(xmm0, src);
4357 } else if (nds_enc < 16) {
4358 movdqu(src, xmm0);
4359 vandpd(xmm0, nds, negate_field, vector_len);
4360 movsd(dst, xmm0);
4361 movdqu(xmm0, src);
4362 } else {
4363 movdqu(src, xmm0);
4364 movsd(xmm0, nds);
4365 vandpd(xmm0, xmm0, negate_field, vector_len);
4366 movsd(dst, xmm0);
4367 movdqu(xmm0, src);
4368 }
4369 }
4370
4371 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4372 int dst_enc = dst->encoding();
4373 int nds_enc = nds->encoding();
4374 int src_enc = src->encoding();
4375 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4376 Assembler::vpaddb(dst, nds, src, vector_len);
4377 } else if ((dst_enc < 16) && (src_enc < 16)) {
4378 Assembler::vpaddb(dst, dst, src, vector_len);
4379 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4380 // use nds as scratch for src
4381 evmovdqul(nds, src, Assembler::AVX_512bit);
4382 Assembler::vpaddb(dst, dst, nds, vector_len);
4383 } else if ((src_enc < 16) && (nds_enc < 16)) {
4384 // use nds as scratch for dst
4385 evmovdqul(nds, dst, Assembler::AVX_512bit);
4386 Assembler::vpaddb(nds, nds, src, vector_len);
4387 evmovdqul(dst, nds, Assembler::AVX_512bit);
4388 } else if (dst_enc < 16) {
4389 // use nds as scatch for xmm0 to hold src
4390 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4391 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4392 Assembler::vpaddb(dst, dst, xmm0, vector_len);
4393 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4394 } else {
4395 // worse case scenario, all regs are in the upper bank
4396 subptr(rsp, 64);
4397 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4398 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4399 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4400 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4401 Assembler::vpaddb(xmm0, xmm0, xmm1, vector_len);
4402 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4403 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4404 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4405 addptr(rsp, 64);
4406 }
4407 }
4408
4409 void MacroAssembler::vpaddb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4410 int dst_enc = dst->encoding();
4411 int nds_enc = nds->encoding();
4412 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4413 Assembler::vpaddb(dst, nds, src, vector_len);
4414 } else if (dst_enc < 16) {
4415 Assembler::vpaddb(dst, dst, src, vector_len);
4416 } else if (nds_enc < 16) {
4417 // implies dst_enc in upper bank with src as scratch
4418 evmovdqul(nds, dst, Assembler::AVX_512bit);
4419 Assembler::vpaddb(nds, nds, src, vector_len);
4420 evmovdqul(dst, nds, Assembler::AVX_512bit);
4421 } else {
4422 // worse case scenario, all regs in upper bank
4423 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4424 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4425 Assembler::vpaddb(xmm0, xmm0, src, vector_len);
4426 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4427 }
4428 }
4429
4430 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4431 int dst_enc = dst->encoding();
4432 int nds_enc = nds->encoding();
4433 int src_enc = src->encoding();
4434 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4435 Assembler::vpaddw(dst, nds, src, vector_len);
4436 } else if ((dst_enc < 16) && (src_enc < 16)) {
4437 Assembler::vpaddw(dst, dst, src, vector_len);
4438 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4439 // use nds as scratch for src
4440 evmovdqul(nds, src, Assembler::AVX_512bit);
4441 Assembler::vpaddw(dst, dst, nds, vector_len);
4442 } else if ((src_enc < 16) && (nds_enc < 16)) {
4443 // use nds as scratch for dst
4444 evmovdqul(nds, dst, Assembler::AVX_512bit);
4445 Assembler::vpaddw(nds, nds, src, vector_len);
4446 evmovdqul(dst, nds, Assembler::AVX_512bit);
4447 } else if (dst_enc < 16) {
4448 // use nds as scatch for xmm0 to hold src
4449 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4450 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4451 Assembler::vpaddw(dst, dst, xmm0, vector_len);
4452 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4453 } else {
4454 // worse case scenario, all regs are in the upper bank
4455 subptr(rsp, 64);
4456 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4457 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4458 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4459 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4460 Assembler::vpaddw(xmm0, xmm0, xmm1, vector_len);
4461 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4462 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4463 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4464 addptr(rsp, 64);
4465 }
4466 }
4467
4468 void MacroAssembler::vpaddw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4469 int dst_enc = dst->encoding();
4470 int nds_enc = nds->encoding();
4471 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4472 Assembler::vpaddw(dst, nds, src, vector_len);
4473 } else if (dst_enc < 16) {
4474 Assembler::vpaddw(dst, dst, src, vector_len);
4475 } else if (nds_enc < 16) {
4476 // implies dst_enc in upper bank with src as scratch
4477 evmovdqul(nds, dst, Assembler::AVX_512bit);
4478 Assembler::vpaddw(nds, nds, src, vector_len);
4479 evmovdqul(dst, nds, Assembler::AVX_512bit);
4480 } else {
4481 // worse case scenario, all regs in upper bank
4482 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4483 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4484 Assembler::vpaddw(xmm0, xmm0, src, vector_len);
4485 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4486 }
4487 }
4488
4489 void MacroAssembler::vpbroadcastw(XMMRegister dst, XMMRegister src) {
4490 int dst_enc = dst->encoding();
4491 int src_enc = src->encoding();
4492 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4493 Assembler::vpbroadcastw(dst, src);
4494 } else if ((dst_enc < 16) && (src_enc < 16)) {
4495 Assembler::vpbroadcastw(dst, src);
4496 } else if (src_enc < 16) {
4497 subptr(rsp, 64);
4498 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4499 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4500 Assembler::vpbroadcastw(xmm0, src);
4501 movdqu(dst, xmm0);
4502 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4503 addptr(rsp, 64);
4504 } else if (dst_enc < 16) {
4505 subptr(rsp, 64);
4506 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4507 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4508 Assembler::vpbroadcastw(dst, xmm0);
4509 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4510 addptr(rsp, 64);
4511 } else {
4512 subptr(rsp, 64);
4513 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4514 subptr(rsp, 64);
4515 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4516 movdqu(xmm0, src);
4517 movdqu(xmm1, dst);
4518 Assembler::vpbroadcastw(xmm1, xmm0);
4519 movdqu(dst, xmm1);
4520 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4521 addptr(rsp, 64);
4522 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4523 addptr(rsp, 64);
4524 }
4525 }
4526
4527 void MacroAssembler::vpcmpeqb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4528 int dst_enc = dst->encoding();
4529 int nds_enc = nds->encoding();
4530 int src_enc = src->encoding();
4531 assert(dst_enc == nds_enc, "");
4532 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4533 Assembler::vpcmpeqb(dst, nds, src, vector_len);
4534 } else if ((dst_enc < 16) && (src_enc < 16)) {
4535 Assembler::vpcmpeqb(dst, nds, src, vector_len);
4536 } else if (src_enc < 16) {
4537 subptr(rsp, 64);
4538 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4539 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4540 Assembler::vpcmpeqb(xmm0, xmm0, src, vector_len);
4541 movdqu(dst, xmm0);
4542 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4543 addptr(rsp, 64);
4544 } else if (dst_enc < 16) {
4545 subptr(rsp, 64);
4546 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4547 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4548 Assembler::vpcmpeqb(dst, dst, xmm0, vector_len);
4549 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4550 addptr(rsp, 64);
4551 } else {
4552 subptr(rsp, 64);
4553 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4554 subptr(rsp, 64);
4555 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4556 movdqu(xmm0, src);
4557 movdqu(xmm1, dst);
4558 Assembler::vpcmpeqb(xmm1, xmm1, xmm0, vector_len);
4559 movdqu(dst, xmm1);
4560 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4561 addptr(rsp, 64);
4562 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4563 addptr(rsp, 64);
4564 }
4565 }
4566
4567 void MacroAssembler::vpcmpeqw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4568 int dst_enc = dst->encoding();
4569 int nds_enc = nds->encoding();
4570 int src_enc = src->encoding();
4571 assert(dst_enc == nds_enc, "");
4572 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4573 Assembler::vpcmpeqw(dst, nds, src, vector_len);
4574 } else if ((dst_enc < 16) && (src_enc < 16)) {
4575 Assembler::vpcmpeqw(dst, nds, src, vector_len);
4576 } else if (src_enc < 16) {
4577 subptr(rsp, 64);
4578 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4579 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4580 Assembler::vpcmpeqw(xmm0, xmm0, src, vector_len);
4581 movdqu(dst, xmm0);
4582 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4583 addptr(rsp, 64);
4584 } else if (dst_enc < 16) {
4585 subptr(rsp, 64);
4586 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4587 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4588 Assembler::vpcmpeqw(dst, dst, xmm0, vector_len);
4589 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4590 addptr(rsp, 64);
4591 } else {
4592 subptr(rsp, 64);
4593 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4594 subptr(rsp, 64);
4595 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4596 movdqu(xmm0, src);
4597 movdqu(xmm1, dst);
4598 Assembler::vpcmpeqw(xmm1, xmm1, xmm0, vector_len);
4599 movdqu(dst, xmm1);
4600 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4601 addptr(rsp, 64);
4602 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4603 addptr(rsp, 64);
4604 }
4605 }
4606
4607 void MacroAssembler::vpmovzxbw(XMMRegister dst, Address src, int vector_len) {
4608 int dst_enc = dst->encoding();
4609 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4610 Assembler::vpmovzxbw(dst, src, vector_len);
4611 } else if (dst_enc < 16) {
4612 Assembler::vpmovzxbw(dst, src, vector_len);
4613 } else {
4614 subptr(rsp, 64);
4615 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4616 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4617 Assembler::vpmovzxbw(xmm0, src, vector_len);
4618 movdqu(dst, xmm0);
4619 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4620 addptr(rsp, 64);
4621 }
4622 }
4623
4624 void MacroAssembler::vpmovmskb(Register dst, XMMRegister src) {
4625 int src_enc = src->encoding();
4626 if (src_enc < 16) {
4627 Assembler::vpmovmskb(dst, src);
4628 } else {
4629 subptr(rsp, 64);
4630 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
4631 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4632 Assembler::vpmovmskb(dst, xmm0);
4633 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
4634 addptr(rsp, 64);
4635 }
4636 }
4637
4638 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4639 int dst_enc = dst->encoding();
4640 int nds_enc = nds->encoding();
4641 int src_enc = src->encoding();
4642 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4643 Assembler::vpmullw(dst, nds, src, vector_len);
4644 } else if ((dst_enc < 16) && (src_enc < 16)) {
4645 Assembler::vpmullw(dst, dst, src, vector_len);
4646 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4647 // use nds as scratch for src
4648 evmovdqul(nds, src, Assembler::AVX_512bit);
4649 Assembler::vpmullw(dst, dst, nds, vector_len);
4650 } else if ((src_enc < 16) && (nds_enc < 16)) {
4651 // use nds as scratch for dst
4652 evmovdqul(nds, dst, Assembler::AVX_512bit);
4653 Assembler::vpmullw(nds, nds, src, vector_len);
4654 evmovdqul(dst, nds, Assembler::AVX_512bit);
4655 } else if (dst_enc < 16) {
4656 // use nds as scatch for xmm0 to hold src
4657 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4658 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4659 Assembler::vpmullw(dst, dst, xmm0, vector_len);
4660 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4661 } else {
4662 // worse case scenario, all regs are in the upper bank
4663 subptr(rsp, 64);
4664 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4665 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4666 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4667 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4668 Assembler::vpmullw(xmm0, xmm0, xmm1, vector_len);
4669 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4670 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4671 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4672 addptr(rsp, 64);
4673 }
4674 }
4675
4676 void MacroAssembler::vpmullw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4677 int dst_enc = dst->encoding();
4678 int nds_enc = nds->encoding();
4679 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4680 Assembler::vpmullw(dst, nds, src, vector_len);
4681 } else if (dst_enc < 16) {
4682 Assembler::vpmullw(dst, dst, src, vector_len);
4683 } else if (nds_enc < 16) {
4684 // implies dst_enc in upper bank with src as scratch
4685 evmovdqul(nds, dst, Assembler::AVX_512bit);
4686 Assembler::vpmullw(nds, nds, src, vector_len);
4687 evmovdqul(dst, nds, Assembler::AVX_512bit);
4688 } else {
4689 // worse case scenario, all regs in upper bank
4690 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4691 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4692 Assembler::vpmullw(xmm0, xmm0, src, vector_len);
4693 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4694 }
4695 }
4696
4697 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4698 int dst_enc = dst->encoding();
4699 int nds_enc = nds->encoding();
4700 int src_enc = src->encoding();
4701 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4702 Assembler::vpsubb(dst, nds, src, vector_len);
4703 } else if ((dst_enc < 16) && (src_enc < 16)) {
4704 Assembler::vpsubb(dst, dst, src, vector_len);
4705 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4706 // use nds as scratch for src
4707 evmovdqul(nds, src, Assembler::AVX_512bit);
4708 Assembler::vpsubb(dst, dst, nds, vector_len);
4709 } else if ((src_enc < 16) && (nds_enc < 16)) {
4710 // use nds as scratch for dst
4711 evmovdqul(nds, dst, Assembler::AVX_512bit);
4712 Assembler::vpsubb(nds, nds, src, vector_len);
4713 evmovdqul(dst, nds, Assembler::AVX_512bit);
4714 } else if (dst_enc < 16) {
4715 // use nds as scatch for xmm0 to hold src
4716 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4717 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4718 Assembler::vpsubb(dst, dst, xmm0, vector_len);
4719 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4720 } else {
4721 // worse case scenario, all regs are in the upper bank
4722 subptr(rsp, 64);
4723 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4724 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4725 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4726 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4727 Assembler::vpsubb(xmm0, xmm0, xmm1, vector_len);
4728 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4729 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4730 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4731 addptr(rsp, 64);
4732 }
4733 }
4734
4735 void MacroAssembler::vpsubb(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4736 int dst_enc = dst->encoding();
4737 int nds_enc = nds->encoding();
4738 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4739 Assembler::vpsubb(dst, nds, src, vector_len);
4740 } else if (dst_enc < 16) {
4741 Assembler::vpsubb(dst, dst, src, vector_len);
4742 } else if (nds_enc < 16) {
4743 // implies dst_enc in upper bank with src as scratch
4744 evmovdqul(nds, dst, Assembler::AVX_512bit);
4745 Assembler::vpsubb(nds, nds, src, vector_len);
4746 evmovdqul(dst, nds, Assembler::AVX_512bit);
4747 } else {
4748 // worse case scenario, all regs in upper bank
4749 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4750 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4751 Assembler::vpsubw(xmm0, xmm0, src, vector_len);
4752 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4753 }
4754 }
4755
4756 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, XMMRegister src, int vector_len) {
4757 int dst_enc = dst->encoding();
4758 int nds_enc = nds->encoding();
4759 int src_enc = src->encoding();
4760 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4761 Assembler::vpsubw(dst, nds, src, vector_len);
4762 } else if ((dst_enc < 16) && (src_enc < 16)) {
4763 Assembler::vpsubw(dst, dst, src, vector_len);
4764 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4765 // use nds as scratch for src
4766 evmovdqul(nds, src, Assembler::AVX_512bit);
4767 Assembler::vpsubw(dst, dst, nds, vector_len);
4768 } else if ((src_enc < 16) && (nds_enc < 16)) {
4769 // use nds as scratch for dst
4770 evmovdqul(nds, dst, Assembler::AVX_512bit);
4771 Assembler::vpsubw(nds, nds, src, vector_len);
4772 evmovdqul(dst, nds, Assembler::AVX_512bit);
4773 } else if (dst_enc < 16) {
4774 // use nds as scatch for xmm0 to hold src
4775 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4776 evmovdqul(xmm0, src, Assembler::AVX_512bit);
4777 Assembler::vpsubw(dst, dst, xmm0, vector_len);
4778 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4779 } else {
4780 // worse case scenario, all regs are in the upper bank
4781 subptr(rsp, 64);
4782 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4783 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4784 evmovdqul(xmm1, src, Assembler::AVX_512bit);
4785 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4786 Assembler::vpsubw(xmm0, xmm0, xmm1, vector_len);
4787 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4788 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4789 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4790 addptr(rsp, 64);
4791 }
4792 }
4793
4794 void MacroAssembler::vpsubw(XMMRegister dst, XMMRegister nds, Address src, int vector_len) {
4795 int dst_enc = dst->encoding();
4796 int nds_enc = nds->encoding();
4797 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4798 Assembler::vpsubw(dst, nds, src, vector_len);
4799 } else if (dst_enc < 16) {
4800 Assembler::vpsubw(dst, dst, src, vector_len);
4801 } else if (nds_enc < 16) {
4802 // implies dst_enc in upper bank with src as scratch
4803 evmovdqul(nds, dst, Assembler::AVX_512bit);
4804 Assembler::vpsubw(nds, nds, src, vector_len);
4805 evmovdqul(dst, nds, Assembler::AVX_512bit);
4806 } else {
4807 // worse case scenario, all regs in upper bank
4808 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4809 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4810 Assembler::vpsubw(xmm0, xmm0, src, vector_len);
4811 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4812 }
4813 }
4814
4815 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4816 int dst_enc = dst->encoding();
4817 int nds_enc = nds->encoding();
4818 int shift_enc = shift->encoding();
4819 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4820 Assembler::vpsraw(dst, nds, shift, vector_len);
4821 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4822 Assembler::vpsraw(dst, dst, shift, vector_len);
4823 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4824 // use nds_enc as scratch with shift
4825 evmovdqul(nds, shift, Assembler::AVX_512bit);
4826 Assembler::vpsraw(dst, dst, nds, vector_len);
4827 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4828 // use nds as scratch with dst
4829 evmovdqul(nds, dst, Assembler::AVX_512bit);
4830 Assembler::vpsraw(nds, nds, shift, vector_len);
4831 evmovdqul(dst, nds, Assembler::AVX_512bit);
4832 } else if (dst_enc < 16) {
4833 // use nds to save a copy of xmm0 and hold shift
4834 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4835 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4836 Assembler::vpsraw(dst, dst, xmm0, vector_len);
4837 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4838 } else if (nds_enc < 16) {
4839 // use nds as dest as temps
4840 evmovdqul(nds, dst, Assembler::AVX_512bit);
4841 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4842 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4843 Assembler::vpsraw(nds, nds, xmm0, vector_len);
4844 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4845 evmovdqul(dst, nds, Assembler::AVX_512bit);
4846 } else {
4847 // worse case scenario, all regs are in the upper bank
4848 subptr(rsp, 64);
4849 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4850 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4851 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4852 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4853 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4854 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4855 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4856 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4857 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4858 addptr(rsp, 64);
4859 }
4860 }
4861
4862 void MacroAssembler::vpsraw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4863 int dst_enc = dst->encoding();
4864 int nds_enc = nds->encoding();
4865 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4866 Assembler::vpsraw(dst, nds, shift, vector_len);
4867 } else if (dst_enc < 16) {
4868 Assembler::vpsraw(dst, dst, shift, vector_len);
4869 } else if (nds_enc < 16) {
4870 // use nds as scratch
4871 evmovdqul(nds, dst, Assembler::AVX_512bit);
4872 Assembler::vpsraw(nds, nds, shift, vector_len);
4873 evmovdqul(dst, nds, Assembler::AVX_512bit);
4874 } else {
4875 // use nds as scratch for xmm0
4876 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4877 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4878 Assembler::vpsraw(xmm0, xmm0, shift, vector_len);
4879 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4880 }
4881 }
4882
4883 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4884 int dst_enc = dst->encoding();
4885 int nds_enc = nds->encoding();
4886 int shift_enc = shift->encoding();
4887 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4888 Assembler::vpsrlw(dst, nds, shift, vector_len);
4889 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4890 Assembler::vpsrlw(dst, dst, shift, vector_len);
4891 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4892 // use nds_enc as scratch with shift
4893 evmovdqul(nds, shift, Assembler::AVX_512bit);
4894 Assembler::vpsrlw(dst, dst, nds, vector_len);
4895 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4896 // use nds as scratch with dst
4897 evmovdqul(nds, dst, Assembler::AVX_512bit);
4898 Assembler::vpsrlw(nds, nds, shift, vector_len);
4899 evmovdqul(dst, nds, Assembler::AVX_512bit);
4900 } else if (dst_enc < 16) {
4901 // use nds to save a copy of xmm0 and hold shift
4902 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4903 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4904 Assembler::vpsrlw(dst, dst, xmm0, vector_len);
4905 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4906 } else if (nds_enc < 16) {
4907 // use nds as dest as temps
4908 evmovdqul(nds, dst, Assembler::AVX_512bit);
4909 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4910 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4911 Assembler::vpsrlw(nds, nds, xmm0, vector_len);
4912 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4913 evmovdqul(dst, nds, Assembler::AVX_512bit);
4914 } else {
4915 // worse case scenario, all regs are in the upper bank
4916 subptr(rsp, 64);
4917 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4918 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4919 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4920 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4921 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4922 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4923 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4924 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4925 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4926 addptr(rsp, 64);
4927 }
4928 }
4929
4930 void MacroAssembler::vpsrlw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4931 int dst_enc = dst->encoding();
4932 int nds_enc = nds->encoding();
4933 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4934 Assembler::vpsrlw(dst, nds, shift, vector_len);
4935 } else if (dst_enc < 16) {
4936 Assembler::vpsrlw(dst, dst, shift, vector_len);
4937 } else if (nds_enc < 16) {
4938 // use nds as scratch
4939 evmovdqul(nds, dst, Assembler::AVX_512bit);
4940 Assembler::vpsrlw(nds, nds, shift, vector_len);
4941 evmovdqul(dst, nds, Assembler::AVX_512bit);
4942 } else {
4943 // use nds as scratch for xmm0
4944 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4945 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4946 Assembler::vpsrlw(xmm0, xmm0, shift, vector_len);
4947 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4948 }
4949 }
4950
4951 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, XMMRegister shift, int vector_len) {
4952 int dst_enc = dst->encoding();
4953 int nds_enc = nds->encoding();
4954 int shift_enc = shift->encoding();
4955 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
4956 Assembler::vpsllw(dst, nds, shift, vector_len);
4957 } else if ((dst_enc < 16) && (shift_enc < 16)) {
4958 Assembler::vpsllw(dst, dst, shift, vector_len);
4959 } else if ((dst_enc < 16) && (nds_enc < 16)) {
4960 // use nds_enc as scratch with shift
4961 evmovdqul(nds, shift, Assembler::AVX_512bit);
4962 Assembler::vpsllw(dst, dst, nds, vector_len);
4963 } else if ((shift_enc < 16) && (nds_enc < 16)) {
4964 // use nds as scratch with dst
4965 evmovdqul(nds, dst, Assembler::AVX_512bit);
4966 Assembler::vpsllw(nds, nds, shift, vector_len);
4967 evmovdqul(dst, nds, Assembler::AVX_512bit);
4968 } else if (dst_enc < 16) {
4969 // use nds to save a copy of xmm0 and hold shift
4970 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4971 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4972 Assembler::vpsllw(dst, dst, xmm0, vector_len);
4973 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4974 } else if (nds_enc < 16) {
4975 // use nds as dest as temps
4976 evmovdqul(nds, dst, Assembler::AVX_512bit);
4977 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4978 evmovdqul(xmm0, shift, Assembler::AVX_512bit);
4979 Assembler::vpsllw(nds, nds, xmm0, vector_len);
4980 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4981 evmovdqul(dst, nds, Assembler::AVX_512bit);
4982 } else {
4983 // worse case scenario, all regs are in the upper bank
4984 subptr(rsp, 64);
4985 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
4986 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
4987 evmovdqul(xmm1, shift, Assembler::AVX_512bit);
4988 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
4989 Assembler::vpsllw(xmm0, xmm0, xmm1, vector_len);
4990 evmovdqul(xmm1, dst, Assembler::AVX_512bit);
4991 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
4992 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
4993 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
4994 addptr(rsp, 64);
4995 }
4996 }
4997
4998 void MacroAssembler::vpsllw(XMMRegister dst, XMMRegister nds, int shift, int vector_len) {
4999 int dst_enc = dst->encoding();
5000 int nds_enc = nds->encoding();
5001 if (VM_Version::supports_avxonly() || VM_Version::supports_avx512bw()) {
5002 Assembler::vpsllw(dst, nds, shift, vector_len);
5003 } else if (dst_enc < 16) {
5004 Assembler::vpsllw(dst, dst, shift, vector_len);
5005 } else if (nds_enc < 16) {
5006 // use nds as scratch
5007 evmovdqul(nds, dst, Assembler::AVX_512bit);
5008 Assembler::vpsllw(nds, nds, shift, vector_len);
5009 evmovdqul(dst, nds, Assembler::AVX_512bit);
5010 } else {
5011 // use nds as scratch for xmm0
5012 evmovdqul(nds, xmm0, Assembler::AVX_512bit);
5013 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5014 Assembler::vpsllw(xmm0, xmm0, shift, vector_len);
5015 evmovdqul(xmm0, nds, Assembler::AVX_512bit);
5016 }
5017 }
5018
5019 void MacroAssembler::vptest(XMMRegister dst, XMMRegister src) {
5020 int dst_enc = dst->encoding();
5021 int src_enc = src->encoding();
5022 if ((dst_enc < 16) && (src_enc < 16)) {
5023 Assembler::vptest(dst, src);
5024 } else if (src_enc < 16) {
5025 subptr(rsp, 64);
5026 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5027 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5028 Assembler::vptest(xmm0, src);
5029 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5030 addptr(rsp, 64);
5031 } else if (dst_enc < 16) {
5032 subptr(rsp, 64);
5033 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5034 evmovdqul(xmm0, src, Assembler::AVX_512bit);
5035 Assembler::vptest(dst, xmm0);
5036 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5037 addptr(rsp, 64);
5038 } else {
5039 subptr(rsp, 64);
5040 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5041 subptr(rsp, 64);
5042 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
5043 movdqu(xmm0, src);
5044 movdqu(xmm1, dst);
5045 Assembler::vptest(xmm1, xmm0);
5046 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
5047 addptr(rsp, 64);
5048 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5049 addptr(rsp, 64);
5050 }
5051 }
5052
5053 // This instruction exists within macros, ergo we cannot control its input
5054 // when emitted through those patterns.
5055 void MacroAssembler::punpcklbw(XMMRegister dst, XMMRegister src) {
5056 if (VM_Version::supports_avx512nobw()) {
5057 int dst_enc = dst->encoding();
5058 int src_enc = src->encoding();
5059 if (dst_enc == src_enc) {
5060 if (dst_enc < 16) {
5061 Assembler::punpcklbw(dst, src);
5062 } else {
5063 subptr(rsp, 64);
5064 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5065 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5066 Assembler::punpcklbw(xmm0, xmm0);
5067 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5068 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5069 addptr(rsp, 64);
5070 }
5071 } else {
5072 if ((src_enc < 16) && (dst_enc < 16)) {
5073 Assembler::punpcklbw(dst, src);
5074 } else if (src_enc < 16) {
5075 subptr(rsp, 64);
5076 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5077 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5078 Assembler::punpcklbw(xmm0, src);
5079 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5080 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5081 addptr(rsp, 64);
5082 } else if (dst_enc < 16) {
5083 subptr(rsp, 64);
5084 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5085 evmovdqul(xmm0, src, Assembler::AVX_512bit);
5086 Assembler::punpcklbw(dst, xmm0);
5087 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5088 addptr(rsp, 64);
5089 } else {
5090 subptr(rsp, 64);
5091 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5092 subptr(rsp, 64);
5093 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
5094 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5095 evmovdqul(xmm1, src, Assembler::AVX_512bit);
5096 Assembler::punpcklbw(xmm0, xmm1);
5097 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5098 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
5099 addptr(rsp, 64);
5100 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5101 addptr(rsp, 64);
5102 }
5103 }
5104 } else {
5105 Assembler::punpcklbw(dst, src);
5106 }
5107 }
5108
5109 // This instruction exists within macros, ergo we cannot control its input
5110 // when emitted through those patterns.
5111 void MacroAssembler::pshuflw(XMMRegister dst, XMMRegister src, int mode) {
5112 if (VM_Version::supports_avx512nobw()) {
5113 int dst_enc = dst->encoding();
5114 int src_enc = src->encoding();
5115 if (dst_enc == src_enc) {
5116 if (dst_enc < 16) {
5117 Assembler::pshuflw(dst, src, mode);
5118 } else {
5119 subptr(rsp, 64);
5120 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5121 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5122 Assembler::pshuflw(xmm0, xmm0, mode);
5123 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5124 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5125 addptr(rsp, 64);
5126 }
5127 } else {
5128 if ((src_enc < 16) && (dst_enc < 16)) {
5129 Assembler::pshuflw(dst, src, mode);
5130 } else if (src_enc < 16) {
5131 subptr(rsp, 64);
5132 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5133 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5134 Assembler::pshuflw(xmm0, src, mode);
5135 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5136 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5137 addptr(rsp, 64);
5138 } else if (dst_enc < 16) {
5139 subptr(rsp, 64);
5140 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5141 evmovdqul(xmm0, src, Assembler::AVX_512bit);
5142 Assembler::pshuflw(dst, xmm0, mode);
5143 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5144 addptr(rsp, 64);
5145 } else {
5146 subptr(rsp, 64);
5147 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5148 subptr(rsp, 64);
5149 evmovdqul(Address(rsp, 0), xmm1, Assembler::AVX_512bit);
5150 evmovdqul(xmm0, dst, Assembler::AVX_512bit);
5151 evmovdqul(xmm1, src, Assembler::AVX_512bit);
5152 Assembler::pshuflw(xmm0, xmm1, mode);
5153 evmovdqul(dst, xmm0, Assembler::AVX_512bit);
5154 evmovdqul(xmm1, Address(rsp, 0), Assembler::AVX_512bit);
5155 addptr(rsp, 64);
5156 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5157 addptr(rsp, 64);
5158 }
5159 }
5160 } else {
5161 Assembler::pshuflw(dst, src, mode);
5162 }
5163 }
5164
5165 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5166 if (reachable(src)) {
5167 vandpd(dst, nds, as_Address(src), vector_len);
5168 } else {
5169 lea(rscratch1, src);
5170 vandpd(dst, nds, Address(rscratch1, 0), vector_len);
5171 }
5172 }
5173
5174 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5175 if (reachable(src)) {
5176 vandps(dst, nds, as_Address(src), vector_len);
5177 } else {
5178 lea(rscratch1, src);
5179 vandps(dst, nds, Address(rscratch1, 0), vector_len);
5180 }
5181 }
5182
5183 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5184 if (reachable(src)) {
5185 vdivsd(dst, nds, as_Address(src));
5186 } else {
5187 lea(rscratch1, src);
5188 vdivsd(dst, nds, Address(rscratch1, 0));
5189 }
5190 }
5191
5192 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5193 if (reachable(src)) {
5194 vdivss(dst, nds, as_Address(src));
5195 } else {
5196 lea(rscratch1, src);
5197 vdivss(dst, nds, Address(rscratch1, 0));
5198 }
5199 }
5200
5201 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5202 if (reachable(src)) {
5203 vmulsd(dst, nds, as_Address(src));
5204 } else {
5205 lea(rscratch1, src);
5206 vmulsd(dst, nds, Address(rscratch1, 0));
5207 }
5208 }
5209
5210 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5211 if (reachable(src)) {
5212 vmulss(dst, nds, as_Address(src));
5213 } else {
5214 lea(rscratch1, src);
5215 vmulss(dst, nds, Address(rscratch1, 0));
5216 }
5217 }
5218
5219 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5220 if (reachable(src)) {
5221 vsubsd(dst, nds, as_Address(src));
5222 } else {
5223 lea(rscratch1, src);
5224 vsubsd(dst, nds, Address(rscratch1, 0));
5225 }
5226 }
5227
5228 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5229 if (reachable(src)) {
5230 vsubss(dst, nds, as_Address(src));
5231 } else {
5232 lea(rscratch1, src);
5233 vsubss(dst, nds, Address(rscratch1, 0));
5234 }
5235 }
5236
5237 void MacroAssembler::vnegatess(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5238 int nds_enc = nds->encoding();
5239 int dst_enc = dst->encoding();
5240 bool dst_upper_bank = (dst_enc > 15);
5241 bool nds_upper_bank = (nds_enc > 15);
5242 if (VM_Version::supports_avx512novl() &&
5243 (nds_upper_bank || dst_upper_bank)) {
5244 if (dst_upper_bank) {
5245 subptr(rsp, 64);
5246 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5247 movflt(xmm0, nds);
5248 vxorps(xmm0, xmm0, src, Assembler::AVX_128bit);
5249 movflt(dst, xmm0);
5250 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5251 addptr(rsp, 64);
5252 } else {
5253 movflt(dst, nds);
5254 vxorps(dst, dst, src, Assembler::AVX_128bit);
5255 }
5256 } else {
5257 vxorps(dst, nds, src, Assembler::AVX_128bit);
5258 }
5259 }
5260
5261 void MacroAssembler::vnegatesd(XMMRegister dst, XMMRegister nds, AddressLiteral src) {
5262 int nds_enc = nds->encoding();
5263 int dst_enc = dst->encoding();
5264 bool dst_upper_bank = (dst_enc > 15);
5265 bool nds_upper_bank = (nds_enc > 15);
5266 if (VM_Version::supports_avx512novl() &&
5267 (nds_upper_bank || dst_upper_bank)) {
5268 if (dst_upper_bank) {
5269 subptr(rsp, 64);
5270 evmovdqul(Address(rsp, 0), xmm0, Assembler::AVX_512bit);
5271 movdbl(xmm0, nds);
5272 vxorpd(xmm0, xmm0, src, Assembler::AVX_128bit);
5273 movdbl(dst, xmm0);
5274 evmovdqul(xmm0, Address(rsp, 0), Assembler::AVX_512bit);
5275 addptr(rsp, 64);
5276 } else {
5277 movdbl(dst, nds);
5278 vxorpd(dst, dst, src, Assembler::AVX_128bit);
5279 }
5280 } else {
5281 vxorpd(dst, nds, src, Assembler::AVX_128bit);
5282 }
5283 }
5284
5285 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5286 if (reachable(src)) {
5287 vxorpd(dst, nds, as_Address(src), vector_len);
5288 } else {
5289 lea(rscratch1, src);
5290 vxorpd(dst, nds, Address(rscratch1, 0), vector_len);
5291 }
5292 }
5293
5294 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, int vector_len) {
5295 if (reachable(src)) {
5296 vxorps(dst, nds, as_Address(src), vector_len);
5297 } else {
5298 lea(rscratch1, src);
5299 vxorps(dst, nds, Address(rscratch1, 0), vector_len);
5300 }
5301 }
5302
5303
5304 //////////////////////////////////////////////////////////////////////////////////
5305 #if INCLUDE_ALL_GCS
5306
5307 void MacroAssembler::g1_write_barrier_pre(Register obj,
5308 Register pre_val,
5309 Register thread,
5310 Register tmp,
5311 bool tosca_live,
5312 bool expand_call) {
5313
5314 // If expand_call is true then we expand the call_VM_leaf macro
5315 // directly to skip generating the check by
5316 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
5317
5318 #ifdef _LP64
5319 assert(thread == r15_thread, "must be");
5320 #endif // _LP64
5321
5322 Label done;
5323 Label runtime;
5324
5325 assert(pre_val != noreg, "check this code");
5326
5327 if (obj != noreg) {
5328 assert_different_registers(obj, pre_val, tmp);
5329 assert(pre_val != rax, "check this code");
5330 }
5331
5332 Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
5333 SATBMarkQueue::byte_offset_of_active()));
5334 Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
5335 SATBMarkQueue::byte_offset_of_index()));
5336 Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() +
5337 SATBMarkQueue::byte_offset_of_buf()));
5338
5339
5340 // Is marking active?
5341 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
5342 cmpl(in_progress, 0);
5343 } else {
5344 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
5345 cmpb(in_progress, 0);
5346 }
5347 jcc(Assembler::equal, done);
5348
5349 // Do we need to load the previous value?
5350 if (obj != noreg) {
5351 load_heap_oop(pre_val, Address(obj, 0));
5352 }
5353
5354 // Is the previous value null?
5355 cmpptr(pre_val, (int32_t) NULL_WORD);
5356 jcc(Assembler::equal, done);
5357
5358 // Can we store original value in the thread's buffer?
5359 // Is index == 0?
5360 // (The index field is typed as size_t.)
5361
5362 movptr(tmp, index); // tmp := *index_adr
5363 cmpptr(tmp, 0); // tmp == 0?
5364 jcc(Assembler::equal, runtime); // If yes, goto runtime
5365
5366 subptr(tmp, wordSize); // tmp := tmp - wordSize
5367 movptr(index, tmp); // *index_adr := tmp
5368 addptr(tmp, buffer); // tmp := tmp + *buffer_adr
5369
5370 // Record the previous value
5371 movptr(Address(tmp, 0), pre_val);
5372 jmp(done);
5373
5374 bind(runtime);
5375 // save the live input values
5376 if(tosca_live) push(rax);
5377
5378 if (obj != noreg && obj != rax)
5379 push(obj);
5380
5381 if (pre_val != rax)
5382 push(pre_val);
5383
5384 // Calling the runtime using the regular call_VM_leaf mechanism generates
5385 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
5386 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
5387 //
5388 // If we care generating the pre-barrier without a frame (e.g. in the
5389 // intrinsified Reference.get() routine) then ebp might be pointing to
5390 // the caller frame and so this check will most likely fail at runtime.
5391 //
5392 // Expanding the call directly bypasses the generation of the check.
5393 // So when we do not have have a full interpreter frame on the stack
5394 // expand_call should be passed true.
5395
5396 NOT_LP64( push(thread); )
5397
5398 if (expand_call) {
5399 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
5400 pass_arg1(this, thread);
5401 pass_arg0(this, pre_val);
5402 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2);
5403 } else {
5404 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread);
5405 }
5406
5407 NOT_LP64( pop(thread); )
5408
5409 // save the live input values
5410 if (pre_val != rax)
5411 pop(pre_val);
5412
5413 if (obj != noreg && obj != rax)
5414 pop(obj);
5415
5416 if(tosca_live) pop(rax);
5417
5418 bind(done);
5419 }
5420
5421 void MacroAssembler::g1_write_barrier_post(Register store_addr,
5422 Register new_val,
5423 Register thread,
5424 Register tmp,
5425 Register tmp2) {
5426 #ifdef _LP64
5427 assert(thread == r15_thread, "must be");
5428 #endif // _LP64
5429
5430 Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() +
5431 DirtyCardQueue::byte_offset_of_index()));
5432 Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() +
5433 DirtyCardQueue::byte_offset_of_buf()));
5434
5435 CardTableModRefBS* ct =
5436 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
5437 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
5438
5439 Label done;
5440 Label runtime;
5441
5442 // Does store cross heap regions?
5443
5444 movptr(tmp, store_addr);
5445 xorptr(tmp, new_val);
5446 shrptr(tmp, HeapRegion::LogOfHRGrainBytes);
5447 jcc(Assembler::equal, done);
5448
5449 // crosses regions, storing NULL?
5450
5451 cmpptr(new_val, (int32_t) NULL_WORD);
5452 jcc(Assembler::equal, done);
5453
5454 // storing region crossing non-NULL, is card already dirty?
5455
5456 const Register card_addr = tmp;
5457 const Register cardtable = tmp2;
5458
5459 movptr(card_addr, store_addr);
5460 shrptr(card_addr, CardTableModRefBS::card_shift);
5461 // Do not use ExternalAddress to load 'byte_map_base', since 'byte_map_base' is NOT
5462 // a valid address and therefore is not properly handled by the relocation code.
5463 movptr(cardtable, (intptr_t)ct->byte_map_base);
5464 addptr(card_addr, cardtable);
5465
5466 cmpb(Address(card_addr, 0), (int)G1SATBCardTableModRefBS::g1_young_card_val());
5467 jcc(Assembler::equal, done);
5468
5469 membar(Assembler::Membar_mask_bits(Assembler::StoreLoad));
5470 cmpb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val());
5471 jcc(Assembler::equal, done);
5472
5473
5474 // storing a region crossing, non-NULL oop, card is clean.
5475 // dirty card and log.
5476
5477 movb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val());
5478
5479 cmpl(queue_index, 0);
5480 jcc(Assembler::equal, runtime);
5481 subl(queue_index, wordSize);
5482 movptr(tmp2, buffer);
5483 #ifdef _LP64
5484 movslq(rscratch1, queue_index);
5485 addq(tmp2, rscratch1);
5486 movq(Address(tmp2, 0), card_addr);
5487 #else
5488 addl(tmp2, queue_index);
5489 movl(Address(tmp2, 0), card_addr);
5490 #endif
5491 jmp(done);
5492
5493 bind(runtime);
5494 // save the live input values
5495 push(store_addr);
5496 push(new_val);
5497 #ifdef _LP64
5498 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread);
5499 #else
5500 push(thread);
5501 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread);
5502 pop(thread);
5503 #endif
5504 pop(new_val);
5505 pop(store_addr);
5506
5507 bind(done);
5508 }
5509
5510 #endif // INCLUDE_ALL_GCS
5511 //////////////////////////////////////////////////////////////////////////////////
5512
5513
5514 void MacroAssembler::store_check(Register obj, Address dst) {
5515 store_check(obj);
5516 }
5517
5518 void MacroAssembler::store_check(Register obj) {
5519 // Does a store check for the oop in register obj. The content of
5520 // register obj is destroyed afterwards.
5521 BarrierSet* bs = Universe::heap()->barrier_set();
5522 assert(bs->kind() == BarrierSet::CardTableForRS ||
5523 bs->kind() == BarrierSet::CardTableExtension,
5524 "Wrong barrier set kind");
5525
5526 CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
5527 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
5528
5529 shrptr(obj, CardTableModRefBS::card_shift);
5530
5531 Address card_addr;
5532
5533 // The calculation for byte_map_base is as follows:
5534 // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift);
5535 // So this essentially converts an address to a displacement and it will
5536 // never need to be relocated. On 64bit however the value may be too
5537 // large for a 32bit displacement.
5538 intptr_t disp = (intptr_t) ct->byte_map_base;
5539 if (is_simm32(disp)) {
5540 card_addr = Address(noreg, obj, Address::times_1, disp);
5541 } else {
5542 // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative
5543 // displacement and done in a single instruction given favorable mapping and a
5544 // smarter version of as_Address. However, 'ExternalAddress' generates a relocation
5545 // entry and that entry is not properly handled by the relocation code.
5546 AddressLiteral cardtable((address)ct->byte_map_base, relocInfo::none);
5547 Address index(noreg, obj, Address::times_1);
5548 card_addr = as_Address(ArrayAddress(cardtable, index));
5549 }
5550
5551 int dirty = CardTableModRefBS::dirty_card_val();
5552 if (UseCondCardMark) {
5553 Label L_already_dirty;
5554 if (UseConcMarkSweepGC) {
5555 membar(Assembler::StoreLoad);
5556 }
5557 cmpb(card_addr, dirty);
5558 jcc(Assembler::equal, L_already_dirty);
5559 movb(card_addr, dirty);
5560 bind(L_already_dirty);
5561 } else {
5562 movb(card_addr, dirty);
5563 }
5564 }
5565
5566 void MacroAssembler::subptr(Register dst, int32_t imm32) {
5567 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32));
5568 }
5569
5570 // Force generation of a 4 byte immediate value even if it fits into 8bit
5571 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) {
5572 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32));
5573 }
5574
5575 void MacroAssembler::subptr(Register dst, Register src) {
5576 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src));
5577 }
5578
5579 // C++ bool manipulation
5580 void MacroAssembler::testbool(Register dst) {
5581 if(sizeof(bool) == 1)
5582 testb(dst, 0xff);
5583 else if(sizeof(bool) == 2) {
5584 // testw implementation needed for two byte bools
5585 ShouldNotReachHere();
5586 } else if(sizeof(bool) == 4)
5587 testl(dst, dst);
5588 else
5589 // unsupported
5590 ShouldNotReachHere();
5591 }
5592
5593 void MacroAssembler::testptr(Register dst, Register src) {
5594 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src));
5595 }
5596
5597 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
5598 void MacroAssembler::tlab_allocate(Register obj,
5599 Register var_size_in_bytes,
5600 int con_size_in_bytes,
5601 Register t1,
5602 Register t2,
5603 Label& slow_case) {
5604 assert_different_registers(obj, t1, t2);
5605 assert_different_registers(obj, var_size_in_bytes, t1);
5606 Register end = t2;
5607 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread);
5608
5609 verify_tlab();
5610
5611 NOT_LP64(get_thread(thread));
5612
5613 movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
5614 if (var_size_in_bytes == noreg) {
5615 lea(end, Address(obj, con_size_in_bytes));
5616 } else {
5617 lea(end, Address(obj, var_size_in_bytes, Address::times_1));
5618 }
5619 cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
5620 jcc(Assembler::above, slow_case);
5621
5622 // update the tlab top pointer
5623 movptr(Address(thread, JavaThread::tlab_top_offset()), end);
5624
5625 // recover var_size_in_bytes if necessary
5626 if (var_size_in_bytes == end) {
5627 subptr(var_size_in_bytes, obj);
5628 }
5629 verify_tlab();
5630 }
5631
5632 // Preserves rbx, and rdx.
5633 Register MacroAssembler::tlab_refill(Label& retry,
5634 Label& try_eden,
5635 Label& slow_case) {
5636 Register top = rax;
5637 Register t1 = rcx;
5638 Register t2 = rsi;
5639 Register thread_reg = NOT_LP64(rdi) LP64_ONLY(r15_thread);
5640 assert_different_registers(top, thread_reg, t1, t2, /* preserve: */ rbx, rdx);
5641 Label do_refill, discard_tlab;
5642
5643 if (!Universe::heap()->supports_inline_contig_alloc()) {
5644 // No allocation in the shared eden.
5645 jmp(slow_case);
5646 }
5647
5648 NOT_LP64(get_thread(thread_reg));
5649
5650 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
5651 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
5652
5653 // calculate amount of free space
5654 subptr(t1, top);
5655 shrptr(t1, LogHeapWordSize);
5656
5657 // Retain tlab and allocate object in shared space if
5658 // the amount free in the tlab is too large to discard.
5659 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())));
5660 jcc(Assembler::lessEqual, discard_tlab);
5661
5662 // Retain
5663 // %%% yuck as movptr...
5664 movptr(t2, (int32_t) ThreadLocalAllocBuffer::refill_waste_limit_increment());
5665 addptr(Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())), t2);
5666 if (TLABStats) {
5667 // increment number of slow_allocations
5668 addl(Address(thread_reg, in_bytes(JavaThread::tlab_slow_allocations_offset())), 1);
5669 }
5670 jmp(try_eden);
5671
5672 bind(discard_tlab);
5673 if (TLABStats) {
5674 // increment number of refills
5675 addl(Address(thread_reg, in_bytes(JavaThread::tlab_number_of_refills_offset())), 1);
5676 // accumulate wastage -- t1 is amount free in tlab
5677 addl(Address(thread_reg, in_bytes(JavaThread::tlab_fast_refill_waste_offset())), t1);
5678 }
5679
5680 // if tlab is currently allocated (top or end != null) then
5681 // fill [top, end + alignment_reserve) with array object
5682 testptr(top, top);
5683 jcc(Assembler::zero, do_refill);
5684
5685 // set up the mark word
5686 movptr(Address(top, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::prototype()->copy_set_hash(0x2));
5687 // set the length to the remaining space
5688 subptr(t1, typeArrayOopDesc::header_size(T_INT));
5689 addptr(t1, (int32_t)ThreadLocalAllocBuffer::alignment_reserve());
5690 shlptr(t1, log2_intptr(HeapWordSize/sizeof(jint)));
5691 movl(Address(top, arrayOopDesc::length_offset_in_bytes()), t1);
5692 // set klass to intArrayKlass
5693 // dubious reloc why not an oop reloc?
5694 movptr(t1, ExternalAddress((address)Universe::intArrayKlassObj_addr()));
5695 // store klass last. concurrent gcs assumes klass length is valid if
5696 // klass field is not null.
5697 store_klass(top, t1);
5698
5699 movptr(t1, top);
5700 subptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
5701 incr_allocated_bytes(thread_reg, t1, 0);
5702
5703 // refill the tlab with an eden allocation
5704 bind(do_refill);
5705 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset())));
5706 shlptr(t1, LogHeapWordSize);
5707 // allocate new tlab, address returned in top
5708 eden_allocate(top, t1, 0, t2, slow_case);
5709
5710 // Check that t1 was preserved in eden_allocate.
5711 #ifdef ASSERT
5712 if (UseTLAB) {
5713 Label ok;
5714 Register tsize = rsi;
5715 assert_different_registers(tsize, thread_reg, t1);
5716 push(tsize);
5717 movptr(tsize, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset())));
5718 shlptr(tsize, LogHeapWordSize);
5719 cmpptr(t1, tsize);
5720 jcc(Assembler::equal, ok);
5721 STOP("assert(t1 != tlab size)");
5722 should_not_reach_here();
5723
5724 bind(ok);
5725 pop(tsize);
5726 }
5727 #endif
5728 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())), top);
5729 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())), top);
5730 addptr(top, t1);
5731 subptr(top, (int32_t)ThreadLocalAllocBuffer::alignment_reserve_in_bytes());
5732 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())), top);
5733 verify_tlab();
5734 jmp(retry);
5735
5736 return thread_reg; // for use by caller
5737 }
5738
5739 void MacroAssembler::incr_allocated_bytes(Register thread,
5740 Register var_size_in_bytes,
5741 int con_size_in_bytes,
5742 Register t1) {
5743 if (!thread->is_valid()) {
5744 #ifdef _LP64
5745 thread = r15_thread;
5746 #else
5747 assert(t1->is_valid(), "need temp reg");
5748 thread = t1;
5749 get_thread(thread);
5750 #endif
5751 }
5752
5753 #ifdef _LP64
5754 if (var_size_in_bytes->is_valid()) {
5755 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
5756 } else {
5757 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
5758 }
5759 #else
5760 if (var_size_in_bytes->is_valid()) {
5761 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes);
5762 } else {
5763 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes);
5764 }
5765 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0);
5766 #endif
5767 }
5768
5769 void MacroAssembler::fp_runtime_fallback(address runtime_entry, int nb_args, int num_fpu_regs_in_use) {
5770 pusha();
5771
5772 // if we are coming from c1, xmm registers may be live
5773 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
5774 if (UseAVX > 2) {
5775 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
5776 }
5777
5778 if (UseSSE == 1) {
5779 subptr(rsp, sizeof(jdouble)*8);
5780 for (int n = 0; n < 8; n++) {
5781 movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n));
5782 }
5783 } else if (UseSSE >= 2) {
5784 if (UseAVX > 2) {
5785 push(rbx);
5786 movl(rbx, 0xffff);
5787 kmovwl(k1, rbx);
5788 pop(rbx);
5789 }
5790 #ifdef COMPILER2
5791 if (MaxVectorSize > 16) {
5792 if(UseAVX > 2) {
5793 // Save upper half of ZMM registers
5794 subptr(rsp, 32*num_xmm_regs);
5795 for (int n = 0; n < num_xmm_regs; n++) {
5796 vextractf64x4h(Address(rsp, n*32), as_XMMRegister(n), 1);
5797 }
5798 }
5799 assert(UseAVX > 0, "256 bit vectors are supported only with AVX");
5800 // Save upper half of YMM registers
5801 subptr(rsp, 16*num_xmm_regs);
5802 for (int n = 0; n < num_xmm_regs; n++) {
5803 vextractf128h(Address(rsp, n*16), as_XMMRegister(n));
5804 }
5805 }
5806 #endif
5807 // Save whole 128bit (16 bytes) XMM registers
5808 subptr(rsp, 16*num_xmm_regs);
5809 #ifdef _LP64
5810 if (VM_Version::supports_evex()) {
5811 for (int n = 0; n < num_xmm_regs; n++) {
5812 vextractf32x4h(Address(rsp, n*16), as_XMMRegister(n), 0);
5813 }
5814 } else {
5815 for (int n = 0; n < num_xmm_regs; n++) {
5816 movdqu(Address(rsp, n*16), as_XMMRegister(n));
5817 }
5818 }
5819 #else
5820 for (int n = 0; n < num_xmm_regs; n++) {
5821 movdqu(Address(rsp, n*16), as_XMMRegister(n));
5822 }
5823 #endif
5824 }
5825
5826 // Preserve registers across runtime call
5827 int incoming_argument_and_return_value_offset = -1;
5828 if (num_fpu_regs_in_use > 1) {
5829 // Must preserve all other FPU regs (could alternatively convert
5830 // SharedRuntime::dsin, dcos etc. into assembly routines known not to trash
5831 // FPU state, but can not trust C compiler)
5832 NEEDS_CLEANUP;
5833 // NOTE that in this case we also push the incoming argument(s) to
5834 // the stack and restore it later; we also use this stack slot to
5835 // hold the return value from dsin, dcos etc.
5836 for (int i = 0; i < num_fpu_regs_in_use; i++) {
5837 subptr(rsp, sizeof(jdouble));
5838 fstp_d(Address(rsp, 0));
5839 }
5840 incoming_argument_and_return_value_offset = sizeof(jdouble)*(num_fpu_regs_in_use-1);
5841 for (int i = nb_args-1; i >= 0; i--) {
5842 fld_d(Address(rsp, incoming_argument_and_return_value_offset-i*sizeof(jdouble)));
5843 }
5844 }
5845
5846 subptr(rsp, nb_args*sizeof(jdouble));
5847 for (int i = 0; i < nb_args; i++) {
5848 fstp_d(Address(rsp, i*sizeof(jdouble)));
5849 }
5850
5851 #ifdef _LP64
5852 if (nb_args > 0) {
5853 movdbl(xmm0, Address(rsp, 0));
5854 }
5855 if (nb_args > 1) {
5856 movdbl(xmm1, Address(rsp, sizeof(jdouble)));
5857 }
5858 assert(nb_args <= 2, "unsupported number of args");
5859 #endif // _LP64
5860
5861 // NOTE: we must not use call_VM_leaf here because that requires a
5862 // complete interpreter frame in debug mode -- same bug as 4387334
5863 // MacroAssembler::call_VM_leaf_base is perfectly safe and will
5864 // do proper 64bit abi
5865
5866 NEEDS_CLEANUP;
5867 // Need to add stack banging before this runtime call if it needs to
5868 // be taken; however, there is no generic stack banging routine at
5869 // the MacroAssembler level
5870
5871 MacroAssembler::call_VM_leaf_base(runtime_entry, 0);
5872
5873 #ifdef _LP64
5874 movsd(Address(rsp, 0), xmm0);
5875 fld_d(Address(rsp, 0));
5876 #endif // _LP64
5877 addptr(rsp, sizeof(jdouble)*nb_args);
5878 if (num_fpu_regs_in_use > 1) {
5879 // Must save return value to stack and then restore entire FPU
5880 // stack except incoming arguments
5881 fstp_d(Address(rsp, incoming_argument_and_return_value_offset));
5882 for (int i = 0; i < num_fpu_regs_in_use - nb_args; i++) {
5883 fld_d(Address(rsp, 0));
5884 addptr(rsp, sizeof(jdouble));
5885 }
5886 fld_d(Address(rsp, (nb_args-1)*sizeof(jdouble)));
5887 addptr(rsp, sizeof(jdouble)*nb_args);
5888 }
5889
5890 if (UseSSE == 1) {
5891 for (int n = 0; n < 8; n++) {
5892 movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble)));
5893 }
5894 addptr(rsp, sizeof(jdouble)*8);
5895 } else if (UseSSE >= 2) {
5896 // Restore whole 128bit (16 bytes) XMM registers
5897 #ifdef _LP64
5898 if (VM_Version::supports_evex()) {
5899 for (int n = 0; n < num_xmm_regs; n++) {
5900 vinsertf32x4h(as_XMMRegister(n), Address(rsp, n*16), 0);
5901 }
5902 } else {
5903 for (int n = 0; n < num_xmm_regs; n++) {
5904 movdqu(as_XMMRegister(n), Address(rsp, n*16));
5905 }
5906 }
5907 #else
5908 for (int n = 0; n < num_xmm_regs; n++) {
5909 movdqu(as_XMMRegister(n), Address(rsp, n*16));
5910 }
5911 #endif
5912 addptr(rsp, 16*num_xmm_regs);
5913
5914 #ifdef COMPILER2
5915 if (MaxVectorSize > 16) {
5916 // Restore upper half of YMM registers.
5917 for (int n = 0; n < num_xmm_regs; n++) {
5918 vinsertf128h(as_XMMRegister(n), Address(rsp, n*16));
5919 }
5920 addptr(rsp, 16*num_xmm_regs);
5921 if(UseAVX > 2) {
5922 for (int n = 0; n < num_xmm_regs; n++) {
5923 vinsertf64x4h(as_XMMRegister(n), Address(rsp, n*32), 1);
5924 }
5925 addptr(rsp, 32*num_xmm_regs);
5926 }
5927 }
5928 #endif
5929 }
5930 popa();
5931 }
5932
5933 static const double pi_4 = 0.7853981633974483;
5934
5935 void MacroAssembler::trigfunc(char trig, int num_fpu_regs_in_use) {
5936 // A hand-coded argument reduction for values in fabs(pi/4, pi/2)
5937 // was attempted in this code; unfortunately it appears that the
5938 // switch to 80-bit precision and back causes this to be
5939 // unprofitable compared with simply performing a runtime call if
5940 // the argument is out of the (-pi/4, pi/4) range.
5941
5942 Register tmp = noreg;
5943 if (!VM_Version::supports_cmov()) {
5944 // fcmp needs a temporary so preserve rbx,
5945 tmp = rbx;
5946 push(tmp);
5947 }
5948
5949 Label slow_case, done;
5950
5951 ExternalAddress pi4_adr = (address)&pi_4;
5952 if (reachable(pi4_adr)) {
5953 // x ?<= pi/4
5954 fld_d(pi4_adr);
5955 fld_s(1); // Stack: X PI/4 X
5956 fabs(); // Stack: |X| PI/4 X
5957 fcmp(tmp);
5958 jcc(Assembler::above, slow_case);
5959
5960 // fastest case: -pi/4 <= x <= pi/4
5961 switch(trig) {
5962 case 's':
5963 fsin();
5964 break;
5965 case 'c':
5966 fcos();
5967 break;
5968 case 't':
5969 ftan();
5970 break;
5971 default:
5972 assert(false, "bad intrinsic");
5973 break;
5974 }
5975 jmp(done);
5976 }
5977
5978 // slow case: runtime call
5979 bind(slow_case);
5980
5981 switch(trig) {
5982 case 's':
5983 {
5984 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1, num_fpu_regs_in_use);
5985 }
5986 break;
5987 case 'c':
5988 {
5989 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1, num_fpu_regs_in_use);
5990 }
5991 break;
5992 case 't':
5993 {
5994 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1, num_fpu_regs_in_use);
5995 }
5996 break;
5997 default:
5998 assert(false, "bad intrinsic");
5999 break;
6000 }
6001
6002 // Come here with result in F-TOS
6003 bind(done);
6004
6005 if (tmp != noreg) {
6006 pop(tmp);
6007 }
6008 }
6009
6010
6011 // Look up the method for a megamorphic invokeinterface call.
6012 // The target method is determined by <intf_klass, itable_index>.
6013 // The receiver klass is in recv_klass.
6014 // On success, the result will be in method_result, and execution falls through.
6015 // On failure, execution transfers to the given label.
6016 void MacroAssembler::lookup_interface_method(Register recv_klass,
6017 Register intf_klass,
6018 RegisterOrConstant itable_index,
6019 Register method_result,
6020 Register scan_temp,
6021 Label& L_no_such_interface) {
6022 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
6023 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
6024 "caller must use same register for non-constant itable index as for method");
6025
6026 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
6027 int vtable_base = InstanceKlass::vtable_start_offset() * wordSize;
6028 int itentry_off = itableMethodEntry::method_offset_in_bytes();
6029 int scan_step = itableOffsetEntry::size() * wordSize;
6030 int vte_size = vtableEntry::size() * wordSize;
6031 Address::ScaleFactor times_vte_scale = Address::times_ptr;
6032 assert(vte_size == wordSize, "else adjust times_vte_scale");
6033
6034 movl(scan_temp, Address(recv_klass, InstanceKlass::vtable_length_offset() * wordSize));
6035
6036 // %%% Could store the aligned, prescaled offset in the klassoop.
6037 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
6038 if (HeapWordsPerLong > 1) {
6039 // Round up to align_object_offset boundary
6040 // see code for InstanceKlass::start_of_itable!
6041 round_to(scan_temp, BytesPerLong);
6042 }
6043
6044 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
6045 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
6046 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
6047
6048 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
6049 // if (scan->interface() == intf) {
6050 // result = (klass + scan->offset() + itable_index);
6051 // }
6052 // }
6053 Label search, found_method;
6054
6055 for (int peel = 1; peel >= 0; peel--) {
6056 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
6057 cmpptr(intf_klass, method_result);
6058
6059 if (peel) {
6060 jccb(Assembler::equal, found_method);
6061 } else {
6062 jccb(Assembler::notEqual, search);
6063 // (invert the test to fall through to found_method...)
6064 }
6065
6066 if (!peel) break;
6067
6068 bind(search);
6069
6070 // Check that the previous entry is non-null. A null entry means that
6071 // the receiver class doesn't implement the interface, and wasn't the
6072 // same as when the caller was compiled.
6073 testptr(method_result, method_result);
6074 jcc(Assembler::zero, L_no_such_interface);
6075 addptr(scan_temp, scan_step);
6076 }
6077
6078 bind(found_method);
6079
6080 // Got a hit.
6081 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
6082 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1));
6083 }
6084
6085
6086 // virtual method calling
6087 void MacroAssembler::lookup_virtual_method(Register recv_klass,
6088 RegisterOrConstant vtable_index,
6089 Register method_result) {
6090 const int base = InstanceKlass::vtable_start_offset() * wordSize;
6091 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below");
6092 Address vtable_entry_addr(recv_klass,
6093 vtable_index, Address::times_ptr,
6094 base + vtableEntry::method_offset_in_bytes());
6095 movptr(method_result, vtable_entry_addr);
6096 }
6097
6098
6099 void MacroAssembler::check_klass_subtype(Register sub_klass,
6100 Register super_klass,
6101 Register temp_reg,
6102 Label& L_success) {
6103 Label L_failure;
6104 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL);
6105 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
6106 bind(L_failure);
6107 }
6108
6109
6110 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
6111 Register super_klass,
6112 Register temp_reg,
6113 Label* L_success,
6114 Label* L_failure,
6115 Label* L_slow_path,
6116 RegisterOrConstant super_check_offset) {
6117 assert_different_registers(sub_klass, super_klass, temp_reg);
6118 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
6119 if (super_check_offset.is_register()) {
6120 assert_different_registers(sub_klass, super_klass,
6121 super_check_offset.as_register());
6122 } else if (must_load_sco) {
6123 assert(temp_reg != noreg, "supply either a temp or a register offset");
6124 }
6125
6126 Label L_fallthrough;
6127 int label_nulls = 0;
6128 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
6129 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
6130 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
6131 assert(label_nulls <= 1, "at most one NULL in the batch");
6132
6133 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
6134 int sco_offset = in_bytes(Klass::super_check_offset_offset());
6135 Address super_check_offset_addr(super_klass, sco_offset);
6136
6137 // Hacked jcc, which "knows" that L_fallthrough, at least, is in
6138 // range of a jccb. If this routine grows larger, reconsider at
6139 // least some of these.
6140 #define local_jcc(assembler_cond, label) \
6141 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \
6142 else jcc( assembler_cond, label) /*omit semi*/
6143
6144 // Hacked jmp, which may only be used just before L_fallthrough.
6145 #define final_jmp(label) \
6146 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
6147 else jmp(label) /*omit semi*/
6148
6149 // If the pointers are equal, we are done (e.g., String[] elements).
6150 // This self-check enables sharing of secondary supertype arrays among
6151 // non-primary types such as array-of-interface. Otherwise, each such
6152 // type would need its own customized SSA.
6153 // We move this check to the front of the fast path because many
6154 // type checks are in fact trivially successful in this manner,
6155 // so we get a nicely predicted branch right at the start of the check.
6156 cmpptr(sub_klass, super_klass);
6157 local_jcc(Assembler::equal, *L_success);
6158
6159 // Check the supertype display:
6160 if (must_load_sco) {
6161 // Positive movl does right thing on LP64.
6162 movl(temp_reg, super_check_offset_addr);
6163 super_check_offset = RegisterOrConstant(temp_reg);
6164 }
6165 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0);
6166 cmpptr(super_klass, super_check_addr); // load displayed supertype
6167
6168 // This check has worked decisively for primary supers.
6169 // Secondary supers are sought in the super_cache ('super_cache_addr').
6170 // (Secondary supers are interfaces and very deeply nested subtypes.)
6171 // This works in the same check above because of a tricky aliasing
6172 // between the super_cache and the primary super display elements.
6173 // (The 'super_check_addr' can address either, as the case requires.)
6174 // Note that the cache is updated below if it does not help us find
6175 // what we need immediately.
6176 // So if it was a primary super, we can just fail immediately.
6177 // Otherwise, it's the slow path for us (no success at this point).
6178
6179 if (super_check_offset.is_register()) {
6180 local_jcc(Assembler::equal, *L_success);
6181 cmpl(super_check_offset.as_register(), sc_offset);
6182 if (L_failure == &L_fallthrough) {
6183 local_jcc(Assembler::equal, *L_slow_path);
6184 } else {
6185 local_jcc(Assembler::notEqual, *L_failure);
6186 final_jmp(*L_slow_path);
6187 }
6188 } else if (super_check_offset.as_constant() == sc_offset) {
6189 // Need a slow path; fast failure is impossible.
6190 if (L_slow_path == &L_fallthrough) {
6191 local_jcc(Assembler::equal, *L_success);
6192 } else {
6193 local_jcc(Assembler::notEqual, *L_slow_path);
6194 final_jmp(*L_success);
6195 }
6196 } else {
6197 // No slow path; it's a fast decision.
6198 if (L_failure == &L_fallthrough) {
6199 local_jcc(Assembler::equal, *L_success);
6200 } else {
6201 local_jcc(Assembler::notEqual, *L_failure);
6202 final_jmp(*L_success);
6203 }
6204 }
6205
6206 bind(L_fallthrough);
6207
6208 #undef local_jcc
6209 #undef final_jmp
6210 }
6211
6212
6213 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
6214 Register super_klass,
6215 Register temp_reg,
6216 Register temp2_reg,
6217 Label* L_success,
6218 Label* L_failure,
6219 bool set_cond_codes) {
6220 assert_different_registers(sub_klass, super_klass, temp_reg);
6221 if (temp2_reg != noreg)
6222 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg);
6223 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
6224
6225 Label L_fallthrough;
6226 int label_nulls = 0;
6227 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
6228 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
6229 assert(label_nulls <= 1, "at most one NULL in the batch");
6230
6231 // a couple of useful fields in sub_klass:
6232 int ss_offset = in_bytes(Klass::secondary_supers_offset());
6233 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
6234 Address secondary_supers_addr(sub_klass, ss_offset);
6235 Address super_cache_addr( sub_klass, sc_offset);
6236
6237 // Do a linear scan of the secondary super-klass chain.
6238 // This code is rarely used, so simplicity is a virtue here.
6239 // The repne_scan instruction uses fixed registers, which we must spill.
6240 // Don't worry too much about pre-existing connections with the input regs.
6241
6242 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super)
6243 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter)
6244
6245 // Get super_klass value into rax (even if it was in rdi or rcx).
6246 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false;
6247 if (super_klass != rax || UseCompressedOops) {
6248 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; }
6249 mov(rax, super_klass);
6250 }
6251 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; }
6252 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; }
6253
6254 #ifndef PRODUCT
6255 int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
6256 ExternalAddress pst_counter_addr((address) pst_counter);
6257 NOT_LP64( incrementl(pst_counter_addr) );
6258 LP64_ONLY( lea(rcx, pst_counter_addr) );
6259 LP64_ONLY( incrementl(Address(rcx, 0)) );
6260 #endif //PRODUCT
6261
6262 // We will consult the secondary-super array.
6263 movptr(rdi, secondary_supers_addr);
6264 // Load the array length. (Positive movl does right thing on LP64.)
6265 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes()));
6266 // Skip to start of data.
6267 addptr(rdi, Array<Klass*>::base_offset_in_bytes());
6268
6269 // Scan RCX words at [RDI] for an occurrence of RAX.
6270 // Set NZ/Z based on last compare.
6271 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does
6272 // not change flags (only scas instruction which is repeated sets flags).
6273 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found.
6274
6275 testptr(rax,rax); // Set Z = 0
6276 repne_scan();
6277
6278 // Unspill the temp. registers:
6279 if (pushed_rdi) pop(rdi);
6280 if (pushed_rcx) pop(rcx);
6281 if (pushed_rax) pop(rax);
6282
6283 if (set_cond_codes) {
6284 // Special hack for the AD files: rdi is guaranteed non-zero.
6285 assert(!pushed_rdi, "rdi must be left non-NULL");
6286 // Also, the condition codes are properly set Z/NZ on succeed/failure.
6287 }
6288
6289 if (L_failure == &L_fallthrough)
6290 jccb(Assembler::notEqual, *L_failure);
6291 else jcc(Assembler::notEqual, *L_failure);
6292
6293 // Success. Cache the super we found and proceed in triumph.
6294 movptr(super_cache_addr, super_klass);
6295
6296 if (L_success != &L_fallthrough) {
6297 jmp(*L_success);
6298 }
6299
6300 #undef IS_A_TEMP
6301
6302 bind(L_fallthrough);
6303 }
6304
6305
6306 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) {
6307 if (VM_Version::supports_cmov()) {
6308 cmovl(cc, dst, src);
6309 } else {
6310 Label L;
6311 jccb(negate_condition(cc), L);
6312 movl(dst, src);
6313 bind(L);
6314 }
6315 }
6316
6317 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) {
6318 if (VM_Version::supports_cmov()) {
6319 cmovl(cc, dst, src);
6320 } else {
6321 Label L;
6322 jccb(negate_condition(cc), L);
6323 movl(dst, src);
6324 bind(L);
6325 }
6326 }
6327
6328 void MacroAssembler::verify_oop(Register reg, const char* s) {
6329 if (!VerifyOops) return;
6330
6331 // Pass register number to verify_oop_subroutine
6332 const char* b = NULL;
6333 {
6334 ResourceMark rm;
6335 stringStream ss;
6336 ss.print("verify_oop: %s: %s", reg->name(), s);
6337 b = code_string(ss.as_string());
6338 }
6339 BLOCK_COMMENT("verify_oop {");
6340 #ifdef _LP64
6341 push(rscratch1); // save r10, trashed by movptr()
6342 #endif
6343 push(rax); // save rax,
6344 push(reg); // pass register argument
6345 ExternalAddress buffer((address) b);
6346 // avoid using pushptr, as it modifies scratch registers
6347 // and our contract is not to modify anything
6348 movptr(rax, buffer.addr());
6349 push(rax);
6350 // call indirectly to solve generation ordering problem
6351 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
6352 call(rax);
6353 // Caller pops the arguments (oop, message) and restores rax, r10
6354 BLOCK_COMMENT("} verify_oop");
6355 }
6356
6357
6358 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
6359 Register tmp,
6360 int offset) {
6361 intptr_t value = *delayed_value_addr;
6362 if (value != 0)
6363 return RegisterOrConstant(value + offset);
6364
6365 // load indirectly to solve generation ordering problem
6366 movptr(tmp, ExternalAddress((address) delayed_value_addr));
6367
6368 #ifdef ASSERT
6369 { Label L;
6370 testptr(tmp, tmp);
6371 if (WizardMode) {
6372 const char* buf = NULL;
6373 {
6374 ResourceMark rm;
6375 stringStream ss;
6376 ss.print("DelayedValue=" INTPTR_FORMAT, delayed_value_addr[1]);
6377 buf = code_string(ss.as_string());
6378 }
6379 jcc(Assembler::notZero, L);
6380 STOP(buf);
6381 } else {
6382 jccb(Assembler::notZero, L);
6383 hlt();
6384 }
6385 bind(L);
6386 }
6387 #endif
6388
6389 if (offset != 0)
6390 addptr(tmp, offset);
6391
6392 return RegisterOrConstant(tmp);
6393 }
6394
6395
6396 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
6397 int extra_slot_offset) {
6398 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
6399 int stackElementSize = Interpreter::stackElementSize;
6400 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
6401 #ifdef ASSERT
6402 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
6403 assert(offset1 - offset == stackElementSize, "correct arithmetic");
6404 #endif
6405 Register scale_reg = noreg;
6406 Address::ScaleFactor scale_factor = Address::no_scale;
6407 if (arg_slot.is_constant()) {
6408 offset += arg_slot.as_constant() * stackElementSize;
6409 } else {
6410 scale_reg = arg_slot.as_register();
6411 scale_factor = Address::times(stackElementSize);
6412 }
6413 offset += wordSize; // return PC is on stack
6414 return Address(rsp, scale_reg, scale_factor, offset);
6415 }
6416
6417
6418 void MacroAssembler::verify_oop_addr(Address addr, const char* s) {
6419 if (!VerifyOops) return;
6420
6421 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord);
6422 // Pass register number to verify_oop_subroutine
6423 const char* b = NULL;
6424 {
6425 ResourceMark rm;
6426 stringStream ss;
6427 ss.print("verify_oop_addr: %s", s);
6428 b = code_string(ss.as_string());
6429 }
6430 #ifdef _LP64
6431 push(rscratch1); // save r10, trashed by movptr()
6432 #endif
6433 push(rax); // save rax,
6434 // addr may contain rsp so we will have to adjust it based on the push
6435 // we just did (and on 64 bit we do two pushes)
6436 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which
6437 // stores rax into addr which is backwards of what was intended.
6438 if (addr.uses(rsp)) {
6439 lea(rax, addr);
6440 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord));
6441 } else {
6442 pushptr(addr);
6443 }
6444
6445 ExternalAddress buffer((address) b);
6446 // pass msg argument
6447 // avoid using pushptr, as it modifies scratch registers
6448 // and our contract is not to modify anything
6449 movptr(rax, buffer.addr());
6450 push(rax);
6451
6452 // call indirectly to solve generation ordering problem
6453 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
6454 call(rax);
6455 // Caller pops the arguments (addr, message) and restores rax, r10.
6456 }
6457
6458 void MacroAssembler::verify_tlab() {
6459 #ifdef ASSERT
6460 if (UseTLAB && VerifyOops) {
6461 Label next, ok;
6462 Register t1 = rsi;
6463 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread);
6464
6465 push(t1);
6466 NOT_LP64(push(thread_reg));
6467 NOT_LP64(get_thread(thread_reg));
6468
6469 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
6470 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())));
6471 jcc(Assembler::aboveEqual, next);
6472 STOP("assert(top >= start)");
6473 should_not_reach_here();
6474
6475 bind(next);
6476 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())));
6477 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())));
6478 jcc(Assembler::aboveEqual, ok);
6479 STOP("assert(top <= end)");
6480 should_not_reach_here();
6481
6482 bind(ok);
6483 NOT_LP64(pop(thread_reg));
6484 pop(t1);
6485 }
6486 #endif
6487 }
6488
6489 class ControlWord {
6490 public:
6491 int32_t _value;
6492
6493 int rounding_control() const { return (_value >> 10) & 3 ; }
6494 int precision_control() const { return (_value >> 8) & 3 ; }
6495 bool precision() const { return ((_value >> 5) & 1) != 0; }
6496 bool underflow() const { return ((_value >> 4) & 1) != 0; }
6497 bool overflow() const { return ((_value >> 3) & 1) != 0; }
6498 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
6499 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
6500 bool invalid() const { return ((_value >> 0) & 1) != 0; }
6501
6502 void print() const {
6503 // rounding control
6504 const char* rc;
6505 switch (rounding_control()) {
6506 case 0: rc = "round near"; break;
6507 case 1: rc = "round down"; break;
6508 case 2: rc = "round up "; break;
6509 case 3: rc = "chop "; break;
6510 };
6511 // precision control
6512 const char* pc;
6513 switch (precision_control()) {
6514 case 0: pc = "24 bits "; break;
6515 case 1: pc = "reserved"; break;
6516 case 2: pc = "53 bits "; break;
6517 case 3: pc = "64 bits "; break;
6518 };
6519 // flags
6520 char f[9];
6521 f[0] = ' ';
6522 f[1] = ' ';
6523 f[2] = (precision ()) ? 'P' : 'p';
6524 f[3] = (underflow ()) ? 'U' : 'u';
6525 f[4] = (overflow ()) ? 'O' : 'o';
6526 f[5] = (zero_divide ()) ? 'Z' : 'z';
6527 f[6] = (denormalized()) ? 'D' : 'd';
6528 f[7] = (invalid ()) ? 'I' : 'i';
6529 f[8] = '\x0';
6530 // output
6531 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc);
6532 }
6533
6534 };
6535
6536 class StatusWord {
6537 public:
6538 int32_t _value;
6539
6540 bool busy() const { return ((_value >> 15) & 1) != 0; }
6541 bool C3() const { return ((_value >> 14) & 1) != 0; }
6542 bool C2() const { return ((_value >> 10) & 1) != 0; }
6543 bool C1() const { return ((_value >> 9) & 1) != 0; }
6544 bool C0() const { return ((_value >> 8) & 1) != 0; }
6545 int top() const { return (_value >> 11) & 7 ; }
6546 bool error_status() const { return ((_value >> 7) & 1) != 0; }
6547 bool stack_fault() const { return ((_value >> 6) & 1) != 0; }
6548 bool precision() const { return ((_value >> 5) & 1) != 0; }
6549 bool underflow() const { return ((_value >> 4) & 1) != 0; }
6550 bool overflow() const { return ((_value >> 3) & 1) != 0; }
6551 bool zero_divide() const { return ((_value >> 2) & 1) != 0; }
6552 bool denormalized() const { return ((_value >> 1) & 1) != 0; }
6553 bool invalid() const { return ((_value >> 0) & 1) != 0; }
6554
6555 void print() const {
6556 // condition codes
6557 char c[5];
6558 c[0] = (C3()) ? '3' : '-';
6559 c[1] = (C2()) ? '2' : '-';
6560 c[2] = (C1()) ? '1' : '-';
6561 c[3] = (C0()) ? '0' : '-';
6562 c[4] = '\x0';
6563 // flags
6564 char f[9];
6565 f[0] = (error_status()) ? 'E' : '-';
6566 f[1] = (stack_fault ()) ? 'S' : '-';
6567 f[2] = (precision ()) ? 'P' : '-';
6568 f[3] = (underflow ()) ? 'U' : '-';
6569 f[4] = (overflow ()) ? 'O' : '-';
6570 f[5] = (zero_divide ()) ? 'Z' : '-';
6571 f[6] = (denormalized()) ? 'D' : '-';
6572 f[7] = (invalid ()) ? 'I' : '-';
6573 f[8] = '\x0';
6574 // output
6575 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top());
6576 }
6577
6578 };
6579
6580 class TagWord {
6581 public:
6582 int32_t _value;
6583
6584 int tag_at(int i) const { return (_value >> (i*2)) & 3; }
6585
6586 void print() const {
6587 printf("%04x", _value & 0xFFFF);
6588 }
6589
6590 };
6591
6592 class FPU_Register {
6593 public:
6594 int32_t _m0;
6595 int32_t _m1;
6596 int16_t _ex;
6597
6598 bool is_indefinite() const {
6599 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0;
6600 }
6601
6602 void print() const {
6603 char sign = (_ex < 0) ? '-' : '+';
6604 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " ";
6605 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind);
6606 };
6607
6608 };
6609
6610 class FPU_State {
6611 public:
6612 enum {
6613 register_size = 10,
6614 number_of_registers = 8,
6615 register_mask = 7
6616 };
6617
6618 ControlWord _control_word;
6619 StatusWord _status_word;
6620 TagWord _tag_word;
6621 int32_t _error_offset;
6622 int32_t _error_selector;
6623 int32_t _data_offset;
6624 int32_t _data_selector;
6625 int8_t _register[register_size * number_of_registers];
6626
6627 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); }
6628 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; }
6629
6630 const char* tag_as_string(int tag) const {
6631 switch (tag) {
6632 case 0: return "valid";
6633 case 1: return "zero";
6634 case 2: return "special";
6635 case 3: return "empty";
6636 }
6637 ShouldNotReachHere();
6638 return NULL;
6639 }
6640
6641 void print() const {
6642 // print computation registers
6643 { int t = _status_word.top();
6644 for (int i = 0; i < number_of_registers; i++) {
6645 int j = (i - t) & register_mask;
6646 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j);
6647 st(j)->print();
6648 printf(" %s\n", tag_as_string(_tag_word.tag_at(i)));
6649 }
6650 }
6651 printf("\n");
6652 // print control registers
6653 printf("ctrl = "); _control_word.print(); printf("\n");
6654 printf("stat = "); _status_word .print(); printf("\n");
6655 printf("tags = "); _tag_word .print(); printf("\n");
6656 }
6657
6658 };
6659
6660 class Flag_Register {
6661 public:
6662 int32_t _value;
6663
6664 bool overflow() const { return ((_value >> 11) & 1) != 0; }
6665 bool direction() const { return ((_value >> 10) & 1) != 0; }
6666 bool sign() const { return ((_value >> 7) & 1) != 0; }
6667 bool zero() const { return ((_value >> 6) & 1) != 0; }
6668 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; }
6669 bool parity() const { return ((_value >> 2) & 1) != 0; }
6670 bool carry() const { return ((_value >> 0) & 1) != 0; }
6671
6672 void print() const {
6673 // flags
6674 char f[8];
6675 f[0] = (overflow ()) ? 'O' : '-';
6676 f[1] = (direction ()) ? 'D' : '-';
6677 f[2] = (sign ()) ? 'S' : '-';
6678 f[3] = (zero ()) ? 'Z' : '-';
6679 f[4] = (auxiliary_carry()) ? 'A' : '-';
6680 f[5] = (parity ()) ? 'P' : '-';
6681 f[6] = (carry ()) ? 'C' : '-';
6682 f[7] = '\x0';
6683 // output
6684 printf("%08x flags = %s", _value, f);
6685 }
6686
6687 };
6688
6689 class IU_Register {
6690 public:
6691 int32_t _value;
6692
6693 void print() const {
6694 printf("%08x %11d", _value, _value);
6695 }
6696
6697 };
6698
6699 class IU_State {
6700 public:
6701 Flag_Register _eflags;
6702 IU_Register _rdi;
6703 IU_Register _rsi;
6704 IU_Register _rbp;
6705 IU_Register _rsp;
6706 IU_Register _rbx;
6707 IU_Register _rdx;
6708 IU_Register _rcx;
6709 IU_Register _rax;
6710
6711 void print() const {
6712 // computation registers
6713 printf("rax, = "); _rax.print(); printf("\n");
6714 printf("rbx, = "); _rbx.print(); printf("\n");
6715 printf("rcx = "); _rcx.print(); printf("\n");
6716 printf("rdx = "); _rdx.print(); printf("\n");
6717 printf("rdi = "); _rdi.print(); printf("\n");
6718 printf("rsi = "); _rsi.print(); printf("\n");
6719 printf("rbp, = "); _rbp.print(); printf("\n");
6720 printf("rsp = "); _rsp.print(); printf("\n");
6721 printf("\n");
6722 // control registers
6723 printf("flgs = "); _eflags.print(); printf("\n");
6724 }
6725 };
6726
6727
6728 class CPU_State {
6729 public:
6730 FPU_State _fpu_state;
6731 IU_State _iu_state;
6732
6733 void print() const {
6734 printf("--------------------------------------------------\n");
6735 _iu_state .print();
6736 printf("\n");
6737 _fpu_state.print();
6738 printf("--------------------------------------------------\n");
6739 }
6740
6741 };
6742
6743
6744 static void _print_CPU_state(CPU_State* state) {
6745 state->print();
6746 };
6747
6748
6749 void MacroAssembler::print_CPU_state() {
6750 push_CPU_state();
6751 push(rsp); // pass CPU state
6752 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state)));
6753 addptr(rsp, wordSize); // discard argument
6754 pop_CPU_state();
6755 }
6756
6757
6758 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) {
6759 static int counter = 0;
6760 FPU_State* fs = &state->_fpu_state;
6761 counter++;
6762 // For leaf calls, only verify that the top few elements remain empty.
6763 // We only need 1 empty at the top for C2 code.
6764 if( stack_depth < 0 ) {
6765 if( fs->tag_for_st(7) != 3 ) {
6766 printf("FPR7 not empty\n");
6767 state->print();
6768 assert(false, "error");
6769 return false;
6770 }
6771 return true; // All other stack states do not matter
6772 }
6773
6774 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std,
6775 "bad FPU control word");
6776
6777 // compute stack depth
6778 int i = 0;
6779 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++;
6780 int d = i;
6781 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++;
6782 // verify findings
6783 if (i != FPU_State::number_of_registers) {
6784 // stack not contiguous
6785 printf("%s: stack not contiguous at ST%d\n", s, i);
6786 state->print();
6787 assert(false, "error");
6788 return false;
6789 }
6790 // check if computed stack depth corresponds to expected stack depth
6791 if (stack_depth < 0) {
6792 // expected stack depth is -stack_depth or less
6793 if (d > -stack_depth) {
6794 // too many elements on the stack
6795 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d);
6796 state->print();
6797 assert(false, "error");
6798 return false;
6799 }
6800 } else {
6801 // expected stack depth is stack_depth
6802 if (d != stack_depth) {
6803 // wrong stack depth
6804 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d);
6805 state->print();
6806 assert(false, "error");
6807 return false;
6808 }
6809 }
6810 // everything is cool
6811 return true;
6812 }
6813
6814
6815 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
6816 if (!VerifyFPU) return;
6817 push_CPU_state();
6818 push(rsp); // pass CPU state
6819 ExternalAddress msg((address) s);
6820 // pass message string s
6821 pushptr(msg.addr());
6822 push(stack_depth); // pass stack depth
6823 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU)));
6824 addptr(rsp, 3 * wordSize); // discard arguments
6825 // check for error
6826 { Label L;
6827 testl(rax, rax);
6828 jcc(Assembler::notZero, L);
6829 int3(); // break if error condition
6830 bind(L);
6831 }
6832 pop_CPU_state();
6833 }
6834
6835 void MacroAssembler::restore_cpu_control_state_after_jni() {
6836 // Either restore the MXCSR register after returning from the JNI Call
6837 // or verify that it wasn't changed (with -Xcheck:jni flag).
6838 if (VM_Version::supports_sse()) {
6839 if (RestoreMXCSROnJNICalls) {
6840 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std()));
6841 } else if (CheckJNICalls) {
6842 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry()));
6843 }
6844 }
6845 if (VM_Version::supports_avx()) {
6846 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty.
6847 vzeroupper();
6848 }
6849
6850 #ifndef _LP64
6851 // Either restore the x87 floating pointer control word after returning
6852 // from the JNI call or verify that it wasn't changed.
6853 if (CheckJNICalls) {
6854 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry()));
6855 }
6856 #endif // _LP64
6857 }
6858
6859
6860 void MacroAssembler::load_klass(Register dst, Register src) {
6861 #ifdef _LP64
6862 if (UseCompressedClassPointers) {
6863 movl(dst, Address(src, oopDesc::klass_offset_in_bytes()));
6864 decode_klass_not_null(dst);
6865 } else
6866 #endif
6867 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
6868 }
6869
6870 void MacroAssembler::load_prototype_header(Register dst, Register src) {
6871 load_klass(dst, src);
6872 movptr(dst, Address(dst, Klass::prototype_header_offset()));
6873 }
6874
6875 void MacroAssembler::store_klass(Register dst, Register src) {
6876 #ifdef _LP64
6877 if (UseCompressedClassPointers) {
6878 encode_klass_not_null(src);
6879 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src);
6880 } else
6881 #endif
6882 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src);
6883 }
6884
6885 void MacroAssembler::load_heap_oop(Register dst, Address src) {
6886 #ifdef _LP64
6887 // FIXME: Must change all places where we try to load the klass.
6888 if (UseCompressedOops) {
6889 movl(dst, src);
6890 decode_heap_oop(dst);
6891 } else
6892 #endif
6893 movptr(dst, src);
6894 }
6895
6896 // Doesn't do verfication, generates fixed size code
6897 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) {
6898 #ifdef _LP64
6899 if (UseCompressedOops) {
6900 movl(dst, src);
6901 decode_heap_oop_not_null(dst);
6902 } else
6903 #endif
6904 movptr(dst, src);
6905 }
6906
6907 void MacroAssembler::store_heap_oop(Address dst, Register src) {
6908 #ifdef _LP64
6909 if (UseCompressedOops) {
6910 assert(!dst.uses(src), "not enough registers");
6911 encode_heap_oop(src);
6912 movl(dst, src);
6913 } else
6914 #endif
6915 movptr(dst, src);
6916 }
6917
6918 void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) {
6919 assert_different_registers(src1, tmp);
6920 #ifdef _LP64
6921 if (UseCompressedOops) {
6922 bool did_push = false;
6923 if (tmp == noreg) {
6924 tmp = rax;
6925 push(tmp);
6926 did_push = true;
6927 assert(!src2.uses(rsp), "can't push");
6928 }
6929 load_heap_oop(tmp, src2);
6930 cmpptr(src1, tmp);
6931 if (did_push) pop(tmp);
6932 } else
6933 #endif
6934 cmpptr(src1, src2);
6935 }
6936
6937 // Used for storing NULLs.
6938 void MacroAssembler::store_heap_oop_null(Address dst) {
6939 #ifdef _LP64
6940 if (UseCompressedOops) {
6941 movl(dst, (int32_t)NULL_WORD);
6942 } else {
6943 movslq(dst, (int32_t)NULL_WORD);
6944 }
6945 #else
6946 movl(dst, (int32_t)NULL_WORD);
6947 #endif
6948 }
6949
6950 #ifdef _LP64
6951 void MacroAssembler::store_klass_gap(Register dst, Register src) {
6952 if (UseCompressedClassPointers) {
6953 // Store to klass gap in destination
6954 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src);
6955 }
6956 }
6957
6958 #ifdef ASSERT
6959 void MacroAssembler::verify_heapbase(const char* msg) {
6960 assert (UseCompressedOops, "should be compressed");
6961 assert (Universe::heap() != NULL, "java heap should be initialized");
6962 if (CheckCompressedOops) {
6963 Label ok;
6964 push(rscratch1); // cmpptr trashes rscratch1
6965 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
6966 jcc(Assembler::equal, ok);
6967 STOP(msg);
6968 bind(ok);
6969 pop(rscratch1);
6970 }
6971 }
6972 #endif
6973
6974 // Algorithm must match oop.inline.hpp encode_heap_oop.
6975 void MacroAssembler::encode_heap_oop(Register r) {
6976 #ifdef ASSERT
6977 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
6978 #endif
6979 verify_oop(r, "broken oop in encode_heap_oop");
6980 if (Universe::narrow_oop_base() == NULL) {
6981 if (Universe::narrow_oop_shift() != 0) {
6982 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
6983 shrq(r, LogMinObjAlignmentInBytes);
6984 }
6985 return;
6986 }
6987 testq(r, r);
6988 cmovq(Assembler::equal, r, r12_heapbase);
6989 subq(r, r12_heapbase);
6990 shrq(r, LogMinObjAlignmentInBytes);
6991 }
6992
6993 void MacroAssembler::encode_heap_oop_not_null(Register r) {
6994 #ifdef ASSERT
6995 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
6996 if (CheckCompressedOops) {
6997 Label ok;
6998 testq(r, r);
6999 jcc(Assembler::notEqual, ok);
7000 STOP("null oop passed to encode_heap_oop_not_null");
7001 bind(ok);
7002 }
7003 #endif
7004 verify_oop(r, "broken oop in encode_heap_oop_not_null");
7005 if (Universe::narrow_oop_base() != NULL) {
7006 subq(r, r12_heapbase);
7007 }
7008 if (Universe::narrow_oop_shift() != 0) {
7009 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
7010 shrq(r, LogMinObjAlignmentInBytes);
7011 }
7012 }
7013
7014 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
7015 #ifdef ASSERT
7016 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
7017 if (CheckCompressedOops) {
7018 Label ok;
7019 testq(src, src);
7020 jcc(Assembler::notEqual, ok);
7021 STOP("null oop passed to encode_heap_oop_not_null2");
7022 bind(ok);
7023 }
7024 #endif
7025 verify_oop(src, "broken oop in encode_heap_oop_not_null2");
7026 if (dst != src) {
7027 movq(dst, src);
7028 }
7029 if (Universe::narrow_oop_base() != NULL) {
7030 subq(dst, r12_heapbase);
7031 }
7032 if (Universe::narrow_oop_shift() != 0) {
7033 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
7034 shrq(dst, LogMinObjAlignmentInBytes);
7035 }
7036 }
7037
7038 void MacroAssembler::decode_heap_oop(Register r) {
7039 #ifdef ASSERT
7040 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
7041 #endif
7042 if (Universe::narrow_oop_base() == NULL) {
7043 if (Universe::narrow_oop_shift() != 0) {
7044 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
7045 shlq(r, LogMinObjAlignmentInBytes);
7046 }
7047 } else {
7048 Label done;
7049 shlq(r, LogMinObjAlignmentInBytes);
7050 jccb(Assembler::equal, done);
7051 addq(r, r12_heapbase);
7052 bind(done);
7053 }
7054 verify_oop(r, "broken oop in decode_heap_oop");
7055 }
7056
7057 void MacroAssembler::decode_heap_oop_not_null(Register r) {
7058 // Note: it will change flags
7059 assert (UseCompressedOops, "should only be used for compressed headers");
7060 assert (Universe::heap() != NULL, "java heap should be initialized");
7061 // Cannot assert, unverified entry point counts instructions (see .ad file)
7062 // vtableStubs also counts instructions in pd_code_size_limit.
7063 // Also do not verify_oop as this is called by verify_oop.
7064 if (Universe::narrow_oop_shift() != 0) {
7065 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
7066 shlq(r, LogMinObjAlignmentInBytes);
7067 if (Universe::narrow_oop_base() != NULL) {
7068 addq(r, r12_heapbase);
7069 }
7070 } else {
7071 assert (Universe::narrow_oop_base() == NULL, "sanity");
7072 }
7073 }
7074
7075 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
7076 // Note: it will change flags
7077 assert (UseCompressedOops, "should only be used for compressed headers");
7078 assert (Universe::heap() != NULL, "java heap should be initialized");
7079 // Cannot assert, unverified entry point counts instructions (see .ad file)
7080 // vtableStubs also counts instructions in pd_code_size_limit.
7081 // Also do not verify_oop as this is called by verify_oop.
7082 if (Universe::narrow_oop_shift() != 0) {
7083 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
7084 if (LogMinObjAlignmentInBytes == Address::times_8) {
7085 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0));
7086 } else {
7087 if (dst != src) {
7088 movq(dst, src);
7089 }
7090 shlq(dst, LogMinObjAlignmentInBytes);
7091 if (Universe::narrow_oop_base() != NULL) {
7092 addq(dst, r12_heapbase);
7093 }
7094 }
7095 } else {
7096 assert (Universe::narrow_oop_base() == NULL, "sanity");
7097 if (dst != src) {
7098 movq(dst, src);
7099 }
7100 }
7101 }
7102
7103 void MacroAssembler::encode_klass_not_null(Register r) {
7104 if (Universe::narrow_klass_base() != NULL) {
7105 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
7106 assert(r != r12_heapbase, "Encoding a klass in r12");
7107 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
7108 subq(r, r12_heapbase);
7109 }
7110 if (Universe::narrow_klass_shift() != 0) {
7111 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
7112 shrq(r, LogKlassAlignmentInBytes);
7113 }
7114 if (Universe::narrow_klass_base() != NULL) {
7115 reinit_heapbase();
7116 }
7117 }
7118
7119 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
7120 if (dst == src) {
7121 encode_klass_not_null(src);
7122 } else {
7123 if (Universe::narrow_klass_base() != NULL) {
7124 mov64(dst, (int64_t)Universe::narrow_klass_base());
7125 negq(dst);
7126 addq(dst, src);
7127 } else {
7128 movptr(dst, src);
7129 }
7130 if (Universe::narrow_klass_shift() != 0) {
7131 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
7132 shrq(dst, LogKlassAlignmentInBytes);
7133 }
7134 }
7135 }
7136
7137 // Function instr_size_for_decode_klass_not_null() counts the instructions
7138 // generated by decode_klass_not_null(register r) and reinit_heapbase(),
7139 // when (Universe::heap() != NULL). Hence, if the instructions they
7140 // generate change, then this method needs to be updated.
7141 int MacroAssembler::instr_size_for_decode_klass_not_null() {
7142 assert (UseCompressedClassPointers, "only for compressed klass ptrs");
7143 if (Universe::narrow_klass_base() != NULL) {
7144 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()).
7145 return (Universe::narrow_klass_shift() == 0 ? 20 : 24);
7146 } else {
7147 // longest load decode klass function, mov64, leaq
7148 return 16;
7149 }
7150 }
7151
7152 // !!! If the instructions that get generated here change then function
7153 // instr_size_for_decode_klass_not_null() needs to get updated.
7154 void MacroAssembler::decode_klass_not_null(Register r) {
7155 // Note: it will change flags
7156 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7157 assert(r != r12_heapbase, "Decoding a klass in r12");
7158 // Cannot assert, unverified entry point counts instructions (see .ad file)
7159 // vtableStubs also counts instructions in pd_code_size_limit.
7160 // Also do not verify_oop as this is called by verify_oop.
7161 if (Universe::narrow_klass_shift() != 0) {
7162 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
7163 shlq(r, LogKlassAlignmentInBytes);
7164 }
7165 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base.
7166 if (Universe::narrow_klass_base() != NULL) {
7167 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base());
7168 addq(r, r12_heapbase);
7169 reinit_heapbase();
7170 }
7171 }
7172
7173 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
7174 // Note: it will change flags
7175 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7176 if (dst == src) {
7177 decode_klass_not_null(dst);
7178 } else {
7179 // Cannot assert, unverified entry point counts instructions (see .ad file)
7180 // vtableStubs also counts instructions in pd_code_size_limit.
7181 // Also do not verify_oop as this is called by verify_oop.
7182 mov64(dst, (int64_t)Universe::narrow_klass_base());
7183 if (Universe::narrow_klass_shift() != 0) {
7184 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
7185 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?");
7186 leaq(dst, Address(dst, src, Address::times_8, 0));
7187 } else {
7188 addq(dst, src);
7189 }
7190 }
7191 }
7192
7193 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
7194 assert (UseCompressedOops, "should only be used for compressed headers");
7195 assert (Universe::heap() != NULL, "java heap should be initialized");
7196 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7197 int oop_index = oop_recorder()->find_index(obj);
7198 RelocationHolder rspec = oop_Relocation::spec(oop_index);
7199 mov_narrow_oop(dst, oop_index, rspec);
7200 }
7201
7202 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) {
7203 assert (UseCompressedOops, "should only be used for compressed headers");
7204 assert (Universe::heap() != NULL, "java heap should be initialized");
7205 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7206 int oop_index = oop_recorder()->find_index(obj);
7207 RelocationHolder rspec = oop_Relocation::spec(oop_index);
7208 mov_narrow_oop(dst, oop_index, rspec);
7209 }
7210
7211 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
7212 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7213 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7214 int klass_index = oop_recorder()->find_index(k);
7215 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
7216 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
7217 }
7218
7219 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) {
7220 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7221 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7222 int klass_index = oop_recorder()->find_index(k);
7223 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
7224 mov_narrow_oop(dst, Klass::encode_klass(k), rspec);
7225 }
7226
7227 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) {
7228 assert (UseCompressedOops, "should only be used for compressed headers");
7229 assert (Universe::heap() != NULL, "java heap should be initialized");
7230 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7231 int oop_index = oop_recorder()->find_index(obj);
7232 RelocationHolder rspec = oop_Relocation::spec(oop_index);
7233 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
7234 }
7235
7236 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) {
7237 assert (UseCompressedOops, "should only be used for compressed headers");
7238 assert (Universe::heap() != NULL, "java heap should be initialized");
7239 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7240 int oop_index = oop_recorder()->find_index(obj);
7241 RelocationHolder rspec = oop_Relocation::spec(oop_index);
7242 Assembler::cmp_narrow_oop(dst, oop_index, rspec);
7243 }
7244
7245 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) {
7246 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7247 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7248 int klass_index = oop_recorder()->find_index(k);
7249 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
7250 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
7251 }
7252
7253 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) {
7254 assert (UseCompressedClassPointers, "should only be used for compressed headers");
7255 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
7256 int klass_index = oop_recorder()->find_index(k);
7257 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
7258 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec);
7259 }
7260
7261 void MacroAssembler::reinit_heapbase() {
7262 if (UseCompressedOops || UseCompressedClassPointers) {
7263 if (Universe::heap() != NULL) {
7264 if (Universe::narrow_oop_base() == NULL) {
7265 MacroAssembler::xorptr(r12_heapbase, r12_heapbase);
7266 } else {
7267 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base());
7268 }
7269 } else {
7270 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr()));
7271 }
7272 }
7273 }
7274
7275 #endif // _LP64
7276
7277
7278 // C2 compiled method's prolog code.
7279 void MacroAssembler::verified_entry(int framesize, int stack_bang_size, bool fp_mode_24b) {
7280
7281 // WARNING: Initial instruction MUST be 5 bytes or longer so that
7282 // NativeJump::patch_verified_entry will be able to patch out the entry
7283 // code safely. The push to verify stack depth is ok at 5 bytes,
7284 // the frame allocation can be either 3 or 6 bytes. So if we don't do
7285 // stack bang then we must use the 6 byte frame allocation even if
7286 // we have no frame. :-(
7287 assert(stack_bang_size >= framesize || stack_bang_size <= 0, "stack bang size incorrect");
7288
7289 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned");
7290 // Remove word for return addr
7291 framesize -= wordSize;
7292 stack_bang_size -= wordSize;
7293
7294 // Calls to C2R adapters often do not accept exceptional returns.
7295 // We require that their callers must bang for them. But be careful, because
7296 // some VM calls (such as call site linkage) can use several kilobytes of
7297 // stack. But the stack safety zone should account for that.
7298 // See bugs 4446381, 4468289, 4497237.
7299 if (stack_bang_size > 0) {
7300 generate_stack_overflow_check(stack_bang_size);
7301
7302 // We always push rbp, so that on return to interpreter rbp, will be
7303 // restored correctly and we can correct the stack.
7304 push(rbp);
7305 // Save caller's stack pointer into RBP if the frame pointer is preserved.
7306 if (PreserveFramePointer) {
7307 mov(rbp, rsp);
7308 }
7309 // Remove word for ebp
7310 framesize -= wordSize;
7311
7312 // Create frame
7313 if (framesize) {
7314 subptr(rsp, framesize);
7315 }
7316 } else {
7317 // Create frame (force generation of a 4 byte immediate value)
7318 subptr_imm32(rsp, framesize);
7319
7320 // Save RBP register now.
7321 framesize -= wordSize;
7322 movptr(Address(rsp, framesize), rbp);
7323 // Save caller's stack pointer into RBP if the frame pointer is preserved.
7324 if (PreserveFramePointer) {
7325 movptr(rbp, rsp);
7326 if (framesize > 0) {
7327 addptr(rbp, framesize);
7328 }
7329 }
7330 }
7331
7332 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth
7333 framesize -= wordSize;
7334 movptr(Address(rsp, framesize), (int32_t)0xbadb100d);
7335 }
7336
7337 #ifndef _LP64
7338 // If method sets FPU control word do it now
7339 if (fp_mode_24b) {
7340 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24()));
7341 }
7342 if (UseSSE >= 2 && VerifyFPU) {
7343 verify_FPU(0, "FPU stack must be clean on entry");
7344 }
7345 #endif
7346
7347 #ifdef ASSERT
7348 if (VerifyStackAtCalls) {
7349 Label L;
7350 push(rax);
7351 mov(rax, rsp);
7352 andptr(rax, StackAlignmentInBytes-1);
7353 cmpptr(rax, StackAlignmentInBytes-wordSize);
7354 pop(rax);
7355 jcc(Assembler::equal, L);
7356 STOP("Stack is not properly aligned!");
7357 bind(L);
7358 }
7359 #endif
7360
7361 }
7362
7363 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp) {
7364 // cnt - number of qwords (8-byte words).
7365 // base - start address, qword aligned.
7366 assert(base==rdi, "base register must be edi for rep stos");
7367 assert(tmp==rax, "tmp register must be eax for rep stos");
7368 assert(cnt==rcx, "cnt register must be ecx for rep stos");
7369
7370 xorptr(tmp, tmp);
7371 if (UseFastStosb) {
7372 shlptr(cnt,3); // convert to number of bytes
7373 rep_stosb();
7374 } else {
7375 NOT_LP64(shlptr(cnt,1);) // convert to number of dwords for 32-bit VM
7376 rep_stos();
7377 }
7378 }
7379
7380 #ifdef COMPILER2
7381
7382 // IndexOf for constant substrings with size >= 8 chars
7383 // which don't need to be loaded through stack.
7384 void MacroAssembler::string_indexofC8(Register str1, Register str2,
7385 Register cnt1, Register cnt2,
7386 int int_cnt2, Register result,
7387 XMMRegister vec, Register tmp,
7388 int ae) {
7389 ShortBranchVerifier sbv(this);
7390 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
7391 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
7392 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
7393
7394 // This method uses the pcmpestri instruction with bound registers
7395 // inputs:
7396 // xmm - substring
7397 // rax - substring length (elements count)
7398 // mem - scanned string
7399 // rdx - string length (elements count)
7400 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
7401 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
7402 // outputs:
7403 // rcx - matched index in string
7404 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
7405 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
7406 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
7407 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
7408 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
7409
7410 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR,
7411 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR,
7412 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE;
7413
7414 // Note, inline_string_indexOf() generates checks:
7415 // if (substr.count > string.count) return -1;
7416 // if (substr.count == 0) return 0;
7417 assert(int_cnt2 >= stride, "this code is used only for cnt2 >= 8 chars");
7418
7419 // Load substring.
7420 if (ae == StrIntrinsicNode::UL) {
7421 pmovzxbw(vec, Address(str2, 0));
7422 } else {
7423 movdqu(vec, Address(str2, 0));
7424 }
7425 movl(cnt2, int_cnt2);
7426 movptr(result, str1); // string addr
7427
7428 if (int_cnt2 > stride) {
7429 jmpb(SCAN_TO_SUBSTR);
7430
7431 // Reload substr for rescan, this code
7432 // is executed only for large substrings (> 8 chars)
7433 bind(RELOAD_SUBSTR);
7434 if (ae == StrIntrinsicNode::UL) {
7435 pmovzxbw(vec, Address(str2, 0));
7436 } else {
7437 movdqu(vec, Address(str2, 0));
7438 }
7439 negptr(cnt2); // Jumped here with negative cnt2, convert to positive
7440
7441 bind(RELOAD_STR);
7442 // We came here after the beginning of the substring was
7443 // matched but the rest of it was not so we need to search
7444 // again. Start from the next element after the previous match.
7445
7446 // cnt2 is number of substring reminding elements and
7447 // cnt1 is number of string reminding elements when cmp failed.
7448 // Restored cnt1 = cnt1 - cnt2 + int_cnt2
7449 subl(cnt1, cnt2);
7450 addl(cnt1, int_cnt2);
7451 movl(cnt2, int_cnt2); // Now restore cnt2
7452
7453 decrementl(cnt1); // Shift to next element
7454 cmpl(cnt1, cnt2);
7455 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7456
7457 addptr(result, (1<<scale1));
7458
7459 } // (int_cnt2 > 8)
7460
7461 // Scan string for start of substr in 16-byte vectors
7462 bind(SCAN_TO_SUBSTR);
7463 pcmpestri(vec, Address(result, 0), mode);
7464 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
7465 subl(cnt1, stride);
7466 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
7467 cmpl(cnt1, cnt2);
7468 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7469 addptr(result, 16);
7470 jmpb(SCAN_TO_SUBSTR);
7471
7472 // Found a potential substr
7473 bind(FOUND_CANDIDATE);
7474 // Matched whole vector if first element matched (tmp(rcx) == 0).
7475 if (int_cnt2 == stride) {
7476 jccb(Assembler::overflow, RET_FOUND); // OF == 1
7477 } else { // int_cnt2 > 8
7478 jccb(Assembler::overflow, FOUND_SUBSTR);
7479 }
7480 // After pcmpestri tmp(rcx) contains matched element index
7481 // Compute start addr of substr
7482 lea(result, Address(result, tmp, scale1));
7483
7484 // Make sure string is still long enough
7485 subl(cnt1, tmp);
7486 cmpl(cnt1, cnt2);
7487 if (int_cnt2 == stride) {
7488 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
7489 } else { // int_cnt2 > 8
7490 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD);
7491 }
7492 // Left less then substring.
7493
7494 bind(RET_NOT_FOUND);
7495 movl(result, -1);
7496 jmpb(EXIT);
7497
7498 if (int_cnt2 > stride) {
7499 // This code is optimized for the case when whole substring
7500 // is matched if its head is matched.
7501 bind(MATCH_SUBSTR_HEAD);
7502 pcmpestri(vec, Address(result, 0), mode);
7503 // Reload only string if does not match
7504 jccb(Assembler::noOverflow, RELOAD_STR); // OF == 0
7505
7506 Label CONT_SCAN_SUBSTR;
7507 // Compare the rest of substring (> 8 chars).
7508 bind(FOUND_SUBSTR);
7509 // First 8 chars are already matched.
7510 negptr(cnt2);
7511 addptr(cnt2, stride);
7512
7513 bind(SCAN_SUBSTR);
7514 subl(cnt1, stride);
7515 cmpl(cnt2, -stride); // Do not read beyond substring
7516 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR);
7517 // Back-up strings to avoid reading beyond substring:
7518 // cnt1 = cnt1 - cnt2 + 8
7519 addl(cnt1, cnt2); // cnt2 is negative
7520 addl(cnt1, stride);
7521 movl(cnt2, stride); negptr(cnt2);
7522 bind(CONT_SCAN_SUBSTR);
7523 if (int_cnt2 < (int)G) {
7524 int tail_off1 = int_cnt2<<scale1;
7525 int tail_off2 = int_cnt2<<scale2;
7526 if (ae == StrIntrinsicNode::UL) {
7527 pmovzxbw(vec, Address(str2, cnt2, scale2, tail_off2));
7528 } else {
7529 movdqu(vec, Address(str2, cnt2, scale2, tail_off2));
7530 }
7531 pcmpestri(vec, Address(result, cnt2, scale1, tail_off1), mode);
7532 } else {
7533 // calculate index in register to avoid integer overflow (int_cnt2*2)
7534 movl(tmp, int_cnt2);
7535 addptr(tmp, cnt2);
7536 if (ae == StrIntrinsicNode::UL) {
7537 pmovzxbw(vec, Address(str2, tmp, scale2, 0));
7538 } else {
7539 movdqu(vec, Address(str2, tmp, scale2, 0));
7540 }
7541 pcmpestri(vec, Address(result, tmp, scale1, 0), mode);
7542 }
7543 // Need to reload strings pointers if not matched whole vector
7544 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
7545 addptr(cnt2, stride);
7546 jcc(Assembler::negative, SCAN_SUBSTR);
7547 // Fall through if found full substring
7548
7549 } // (int_cnt2 > 8)
7550
7551 bind(RET_FOUND);
7552 // Found result if we matched full small substring.
7553 // Compute substr offset
7554 subptr(result, str1);
7555 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
7556 shrl(result, 1); // index
7557 }
7558 bind(EXIT);
7559
7560 } // string_indexofC8
7561
7562 // Small strings are loaded through stack if they cross page boundary.
7563 void MacroAssembler::string_indexof(Register str1, Register str2,
7564 Register cnt1, Register cnt2,
7565 int int_cnt2, Register result,
7566 XMMRegister vec, Register tmp,
7567 int ae) {
7568 ShortBranchVerifier sbv(this);
7569 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
7570 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
7571 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
7572
7573 //
7574 // int_cnt2 is length of small (< 8 chars) constant substring
7575 // or (-1) for non constant substring in which case its length
7576 // is in cnt2 register.
7577 //
7578 // Note, inline_string_indexOf() generates checks:
7579 // if (substr.count > string.count) return -1;
7580 // if (substr.count == 0) return 0;
7581 //
7582 int stride = (ae == StrIntrinsicNode::LL) ? 16 : 8; //UU, UL -> 8
7583 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < stride), "should be != 0");
7584 // This method uses the pcmpestri instruction with bound registers
7585 // inputs:
7586 // xmm - substring
7587 // rax - substring length (elements count)
7588 // mem - scanned string
7589 // rdx - string length (elements count)
7590 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts)
7591 // 0xc - mode: 1100 (substring search) + 00 (unsigned bytes)
7592 // outputs:
7593 // rcx - matched index in string
7594 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
7595 int mode = (ae == StrIntrinsicNode::LL) ? 0x0c : 0x0d; // bytes or shorts
7596 Address::ScaleFactor scale1 = (ae == StrIntrinsicNode::LL) ? Address::times_1 : Address::times_2;
7597 Address::ScaleFactor scale2 = (ae == StrIntrinsicNode::UL) ? Address::times_1 : scale1;
7598
7599 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR,
7600 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR,
7601 FOUND_CANDIDATE;
7602
7603 { //========================================================
7604 // We don't know where these strings are located
7605 // and we can't read beyond them. Load them through stack.
7606 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR;
7607
7608 movptr(tmp, rsp); // save old SP
7609
7610 if (int_cnt2 > 0) { // small (< 8 chars) constant substring
7611 if (int_cnt2 == (1>>scale2)) { // One byte
7612 assert((ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL), "Only possible for latin1 encoding");
7613 load_unsigned_byte(result, Address(str2, 0));
7614 movdl(vec, result); // move 32 bits
7615 } else if (ae == StrIntrinsicNode::LL && int_cnt2 == 3) { // Three bytes
7616 // Not enough header space in 32-bit VM: 12+3 = 15.
7617 movl(result, Address(str2, -1));
7618 shrl(result, 8);
7619 movdl(vec, result); // move 32 bits
7620 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (2>>scale2)) { // One char
7621 load_unsigned_short(result, Address(str2, 0));
7622 movdl(vec, result); // move 32 bits
7623 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (4>>scale2)) { // Two chars
7624 movdl(vec, Address(str2, 0)); // move 32 bits
7625 } else if (ae != StrIntrinsicNode::UL && int_cnt2 == (8>>scale2)) { // Four chars
7626 movq(vec, Address(str2, 0)); // move 64 bits
7627 } else { // cnt2 = { 3, 5, 6, 7 } || (ae == StrIntrinsicNode::UL && cnt2 ={2, ..., 7})
7628 // Array header size is 12 bytes in 32-bit VM
7629 // + 6 bytes for 3 chars == 18 bytes,
7630 // enough space to load vec and shift.
7631 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity");
7632 if (ae == StrIntrinsicNode::UL) {
7633 int tail_off = int_cnt2-8;
7634 pmovzxbw(vec, Address(str2, tail_off));
7635 psrldq(vec, -2*tail_off);
7636 }
7637 else {
7638 int tail_off = int_cnt2*(1<<scale2);
7639 movdqu(vec, Address(str2, tail_off-16));
7640 psrldq(vec, 16-tail_off);
7641 }
7642 }
7643 } else { // not constant substring
7644 cmpl(cnt2, stride);
7645 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough
7646
7647 // We can read beyond string if srt+16 does not cross page boundary
7648 // since heaps are aligned and mapped by pages.
7649 assert(os::vm_page_size() < (int)G, "default page should be small");
7650 movl(result, str2); // We need only low 32 bits
7651 andl(result, (os::vm_page_size()-1));
7652 cmpl(result, (os::vm_page_size()-16));
7653 jccb(Assembler::belowEqual, CHECK_STR);
7654
7655 // Move small strings to stack to allow load 16 bytes into vec.
7656 subptr(rsp, 16);
7657 int stk_offset = wordSize-(1<<scale2);
7658 push(cnt2);
7659
7660 bind(COPY_SUBSTR);
7661 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UL) {
7662 load_unsigned_byte(result, Address(str2, cnt2, scale2, -1));
7663 movb(Address(rsp, cnt2, scale2, stk_offset), result);
7664 } else if (ae == StrIntrinsicNode::UU) {
7665 load_unsigned_short(result, Address(str2, cnt2, scale2, -2));
7666 movw(Address(rsp, cnt2, scale2, stk_offset), result);
7667 }
7668 decrement(cnt2);
7669 jccb(Assembler::notZero, COPY_SUBSTR);
7670
7671 pop(cnt2);
7672 movptr(str2, rsp); // New substring address
7673 } // non constant
7674
7675 bind(CHECK_STR);
7676 cmpl(cnt1, stride);
7677 jccb(Assembler::aboveEqual, BIG_STRINGS);
7678
7679 // Check cross page boundary.
7680 movl(result, str1); // We need only low 32 bits
7681 andl(result, (os::vm_page_size()-1));
7682 cmpl(result, (os::vm_page_size()-16));
7683 jccb(Assembler::belowEqual, BIG_STRINGS);
7684
7685 subptr(rsp, 16);
7686 int stk_offset = -(1<<scale1);
7687 if (int_cnt2 < 0) { // not constant
7688 push(cnt2);
7689 stk_offset += wordSize;
7690 }
7691 movl(cnt2, cnt1);
7692
7693 bind(COPY_STR);
7694 if (ae == StrIntrinsicNode::LL) {
7695 load_unsigned_byte(result, Address(str1, cnt2, scale1, -1));
7696 movb(Address(rsp, cnt2, scale1, stk_offset), result);
7697 } else {
7698 load_unsigned_short(result, Address(str1, cnt2, scale1, -2));
7699 movw(Address(rsp, cnt2, scale1, stk_offset), result);
7700 }
7701 decrement(cnt2);
7702 jccb(Assembler::notZero, COPY_STR);
7703
7704 if (int_cnt2 < 0) { // not constant
7705 pop(cnt2);
7706 }
7707 movptr(str1, rsp); // New string address
7708
7709 bind(BIG_STRINGS);
7710 // Load substring.
7711 if (int_cnt2 < 0) { // -1
7712 if (ae == StrIntrinsicNode::UL) {
7713 pmovzxbw(vec, Address(str2, 0));
7714 } else {
7715 movdqu(vec, Address(str2, 0));
7716 }
7717 push(cnt2); // substr count
7718 push(str2); // substr addr
7719 push(str1); // string addr
7720 } else {
7721 // Small (< 8 chars) constant substrings are loaded already.
7722 movl(cnt2, int_cnt2);
7723 }
7724 push(tmp); // original SP
7725
7726 } // Finished loading
7727
7728 //========================================================
7729 // Start search
7730 //
7731
7732 movptr(result, str1); // string addr
7733
7734 if (int_cnt2 < 0) { // Only for non constant substring
7735 jmpb(SCAN_TO_SUBSTR);
7736
7737 // SP saved at sp+0
7738 // String saved at sp+1*wordSize
7739 // Substr saved at sp+2*wordSize
7740 // Substr count saved at sp+3*wordSize
7741
7742 // Reload substr for rescan, this code
7743 // is executed only for large substrings (> 8 chars)
7744 bind(RELOAD_SUBSTR);
7745 movptr(str2, Address(rsp, 2*wordSize));
7746 movl(cnt2, Address(rsp, 3*wordSize));
7747 if (ae == StrIntrinsicNode::UL) {
7748 pmovzxbw(vec, Address(str2, 0));
7749 } else {
7750 movdqu(vec, Address(str2, 0));
7751 }
7752 // We came here after the beginning of the substring was
7753 // matched but the rest of it was not so we need to search
7754 // again. Start from the next element after the previous match.
7755 subptr(str1, result); // Restore counter
7756 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
7757 shrl(str1, 1);
7758 }
7759 addl(cnt1, str1);
7760 decrementl(cnt1); // Shift to next element
7761 cmpl(cnt1, cnt2);
7762 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7763
7764 addptr(result, (1<<scale1));
7765 } // non constant
7766
7767 // Scan string for start of substr in 16-byte vectors
7768 bind(SCAN_TO_SUBSTR);
7769 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri");
7770 pcmpestri(vec, Address(result, 0), mode);
7771 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1
7772 subl(cnt1, stride);
7773 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string
7774 cmpl(cnt1, cnt2);
7775 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring
7776 addptr(result, 16);
7777
7778 bind(ADJUST_STR);
7779 cmpl(cnt1, stride); // Do not read beyond string
7780 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR);
7781 // Back-up string to avoid reading beyond string.
7782 lea(result, Address(result, cnt1, scale1, -16));
7783 movl(cnt1, stride);
7784 jmpb(SCAN_TO_SUBSTR);
7785
7786 // Found a potential substr
7787 bind(FOUND_CANDIDATE);
7788 // After pcmpestri tmp(rcx) contains matched element index
7789
7790 // Make sure string is still long enough
7791 subl(cnt1, tmp);
7792 cmpl(cnt1, cnt2);
7793 jccb(Assembler::greaterEqual, FOUND_SUBSTR);
7794 // Left less then substring.
7795
7796 bind(RET_NOT_FOUND);
7797 movl(result, -1);
7798 jmpb(CLEANUP);
7799
7800 bind(FOUND_SUBSTR);
7801 // Compute start addr of substr
7802 lea(result, Address(result, tmp, scale1));
7803 if (int_cnt2 > 0) { // Constant substring
7804 // Repeat search for small substring (< 8 chars)
7805 // from new point without reloading substring.
7806 // Have to check that we don't read beyond string.
7807 cmpl(tmp, stride-int_cnt2);
7808 jccb(Assembler::greater, ADJUST_STR);
7809 // Fall through if matched whole substring.
7810 } else { // non constant
7811 assert(int_cnt2 == -1, "should be != 0");
7812
7813 addl(tmp, cnt2);
7814 // Found result if we matched whole substring.
7815 cmpl(tmp, stride);
7816 jccb(Assembler::lessEqual, RET_FOUND);
7817
7818 // Repeat search for small substring (<= 8 chars)
7819 // from new point 'str1' without reloading substring.
7820 cmpl(cnt2, stride);
7821 // Have to check that we don't read beyond string.
7822 jccb(Assembler::lessEqual, ADJUST_STR);
7823
7824 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG;
7825 // Compare the rest of substring (> 8 chars).
7826 movptr(str1, result);
7827
7828 cmpl(tmp, cnt2);
7829 // First 8 chars are already matched.
7830 jccb(Assembler::equal, CHECK_NEXT);
7831
7832 bind(SCAN_SUBSTR);
7833 pcmpestri(vec, Address(str1, 0), mode);
7834 // Need to reload strings pointers if not matched whole vector
7835 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0
7836
7837 bind(CHECK_NEXT);
7838 subl(cnt2, stride);
7839 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring
7840 addptr(str1, 16);
7841 if (ae == StrIntrinsicNode::UL) {
7842 addptr(str2, 8);
7843 } else {
7844 addptr(str2, 16);
7845 }
7846 subl(cnt1, stride);
7847 cmpl(cnt2, stride); // Do not read beyond substring
7848 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR);
7849 // Back-up strings to avoid reading beyond substring.
7850
7851 if (ae == StrIntrinsicNode::UL) {
7852 lea(str2, Address(str2, cnt2, scale2, -8));
7853 lea(str1, Address(str1, cnt2, scale1, -16));
7854 } else {
7855 lea(str2, Address(str2, cnt2, scale2, -16));
7856 lea(str1, Address(str1, cnt2, scale1, -16));
7857 }
7858 subl(cnt1, cnt2);
7859 movl(cnt2, stride);
7860 addl(cnt1, stride);
7861 bind(CONT_SCAN_SUBSTR);
7862 if (ae == StrIntrinsicNode::UL) {
7863 pmovzxbw(vec, Address(str2, 0));
7864 } else {
7865 movdqu(vec, Address(str2, 0));
7866 }
7867 jmpb(SCAN_SUBSTR);
7868
7869 bind(RET_FOUND_LONG);
7870 movptr(str1, Address(rsp, wordSize));
7871 } // non constant
7872
7873 bind(RET_FOUND);
7874 // Compute substr offset
7875 subptr(result, str1);
7876 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
7877 shrl(result, 1); // index
7878 }
7879 bind(CLEANUP);
7880 pop(rsp); // restore SP
7881
7882 } // string_indexof
7883
7884 void MacroAssembler::string_indexof_char(Register str1, Register cnt1, Register ch, Register result,
7885 XMMRegister vec1, XMMRegister vec2, XMMRegister vec3, Register tmp) {
7886 ShortBranchVerifier sbv(this);
7887 assert(UseSSE42Intrinsics, "SSE4.2 intrinsics are required");
7888 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
7889
7890 int stride = 8;
7891
7892 Label FOUND_CHAR, SCAN_TO_CHAR, SCAN_TO_CHAR_LOOP,
7893 SCAN_TO_8_CHAR, SCAN_TO_8_CHAR_LOOP, SCAN_TO_16_CHAR_LOOP,
7894 RET_NOT_FOUND, SCAN_TO_8_CHAR_INIT,
7895 FOUND_SEQ_CHAR, DONE_LABEL;
7896
7897 movptr(result, str1);
7898 if (UseAVX >= 2) {
7899 cmpl(cnt1, stride);
7900 jccb(Assembler::less, SCAN_TO_CHAR_LOOP);
7901 cmpl(cnt1, 2*stride);
7902 jccb(Assembler::less, SCAN_TO_8_CHAR_INIT);
7903 movdl(vec1, ch);
7904 vpbroadcastw(vec1, vec1);
7905 vpxor(vec2, vec2);
7906 movl(tmp, cnt1);
7907 andl(tmp, 0xFFFFFFF0); //vector count (in chars)
7908 andl(cnt1,0x0000000F); //tail count (in chars)
7909
7910 bind(SCAN_TO_16_CHAR_LOOP);
7911 vmovdqu(vec3, Address(result, 0));
7912 vpcmpeqw(vec3, vec3, vec1, 1);
7913 vptest(vec2, vec3);
7914 jcc(Assembler::carryClear, FOUND_CHAR);
7915 addptr(result, 32);
7916 subl(tmp, 2*stride);
7917 jccb(Assembler::notZero, SCAN_TO_16_CHAR_LOOP);
7918 jmp(SCAN_TO_8_CHAR);
7919 bind(SCAN_TO_8_CHAR_INIT);
7920 movdl(vec1, ch);
7921 pshuflw(vec1, vec1, 0x00);
7922 pshufd(vec1, vec1, 0);
7923 pxor(vec2, vec2);
7924 }
7925 bind(SCAN_TO_8_CHAR);
7926 cmpl(cnt1, stride);
7927 if (UseAVX >= 2) {
7928 jccb(Assembler::less, SCAN_TO_CHAR);
7929 } else {
7930 jccb(Assembler::less, SCAN_TO_CHAR_LOOP);
7931 movdl(vec1, ch);
7932 pshuflw(vec1, vec1, 0x00);
7933 pshufd(vec1, vec1, 0);
7934 pxor(vec2, vec2);
7935 }
7936 movl(tmp, cnt1);
7937 andl(tmp, 0xFFFFFFF8); //vector count (in chars)
7938 andl(cnt1,0x00000007); //tail count (in chars)
7939
7940 bind(SCAN_TO_8_CHAR_LOOP);
7941 movdqu(vec3, Address(result, 0));
7942 pcmpeqw(vec3, vec1);
7943 ptest(vec2, vec3);
7944 jcc(Assembler::carryClear, FOUND_CHAR);
7945 addptr(result, 16);
7946 subl(tmp, stride);
7947 jccb(Assembler::notZero, SCAN_TO_8_CHAR_LOOP);
7948 bind(SCAN_TO_CHAR);
7949 testl(cnt1, cnt1);
7950 jcc(Assembler::zero, RET_NOT_FOUND);
7951 bind(SCAN_TO_CHAR_LOOP);
7952 load_unsigned_short(tmp, Address(result, 0));
7953 cmpl(ch, tmp);
7954 jccb(Assembler::equal, FOUND_SEQ_CHAR);
7955 addptr(result, 2);
7956 subl(cnt1, 1);
7957 jccb(Assembler::zero, RET_NOT_FOUND);
7958 jmp(SCAN_TO_CHAR_LOOP);
7959
7960 bind(RET_NOT_FOUND);
7961 movl(result, -1);
7962 jmpb(DONE_LABEL);
7963
7964 bind(FOUND_CHAR);
7965 if (UseAVX >= 2) {
7966 vpmovmskb(tmp, vec3);
7967 } else {
7968 pmovmskb(tmp, vec3);
7969 }
7970 bsfl(ch, tmp);
7971 addl(result, ch);
7972
7973 bind(FOUND_SEQ_CHAR);
7974 subptr(result, str1);
7975 shrl(result, 1);
7976
7977 bind(DONE_LABEL);
7978 } // string_indexof_char
7979
7980 // helper function for string_compare
7981 void MacroAssembler::load_next_elements(Register elem1, Register elem2, Register str1, Register str2,
7982 Address::ScaleFactor scale, Address::ScaleFactor scale1,
7983 Address::ScaleFactor scale2, Register index, int ae) {
7984 if (ae == StrIntrinsicNode::LL) {
7985 load_unsigned_byte(elem1, Address(str1, index, scale, 0));
7986 load_unsigned_byte(elem2, Address(str2, index, scale, 0));
7987 } else if (ae == StrIntrinsicNode::UU) {
7988 load_unsigned_short(elem1, Address(str1, index, scale, 0));
7989 load_unsigned_short(elem2, Address(str2, index, scale, 0));
7990 } else {
7991 load_unsigned_byte(elem1, Address(str1, index, scale1, 0));
7992 load_unsigned_short(elem2, Address(str2, index, scale2, 0));
7993 }
7994 }
7995
7996 // Compare strings, used for char[] and byte[].
7997 void MacroAssembler::string_compare(Register str1, Register str2,
7998 Register cnt1, Register cnt2, Register result,
7999 XMMRegister vec1, int ae) {
8000 ShortBranchVerifier sbv(this);
8001 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL;
8002 int stride, stride2, adr_stride, adr_stride1, adr_stride2;
8003 Address::ScaleFactor scale, scale1, scale2;
8004
8005 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
8006 shrl(cnt2, 1);
8007 }
8008 // Compute the minimum of the string lengths and the
8009 // difference of the string lengths (stack).
8010 // Do the conditional move stuff
8011 movl(result, cnt1);
8012 subl(cnt1, cnt2);
8013 push(cnt1);
8014 cmov32(Assembler::lessEqual, cnt2, result);
8015
8016 // Is the minimum length zero?
8017 testl(cnt2, cnt2);
8018 jcc(Assembler::zero, LENGTH_DIFF_LABEL);
8019 if (ae == StrIntrinsicNode::LL) {
8020 // Load first bytes
8021 load_unsigned_byte(result, Address(str1, 0));
8022 load_unsigned_byte(cnt1, Address(str2, 0));
8023 } else if (ae == StrIntrinsicNode::UU) {
8024 // Load first characters
8025 load_unsigned_short(result, Address(str1, 0));
8026 load_unsigned_short(cnt1, Address(str2, 0));
8027 } else {
8028 load_unsigned_byte(result, Address(str1, 0));
8029 load_unsigned_short(cnt1, Address(str2, 0));
8030 }
8031 subl(result, cnt1);
8032 jcc(Assembler::notZero, POP_LABEL);
8033
8034 if (ae == StrIntrinsicNode::UU) {
8035 // Divide length by 2 to get number of chars
8036 shrl(cnt2, 1);
8037 }
8038 cmpl(cnt2, 1);
8039 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
8040
8041 // Check if the strings start at the same location and setup scale and stride
8042 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8043 cmpptr(str1, str2);
8044 jcc(Assembler::equal, LENGTH_DIFF_LABEL);
8045 if (ae == StrIntrinsicNode::LL) {
8046 scale = Address::times_1;
8047 stride = 16;
8048 } else {
8049 scale = Address::times_2;
8050 stride = 8;
8051 }
8052 } else {
8053 scale = Address::no_scale; // not used
8054 scale1 = Address::times_1;
8055 scale2 = Address::times_2;
8056 stride = 8;
8057 }
8058
8059 if (UseAVX >= 2 && UseSSE42Intrinsics) {
8060 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
8061 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR;
8062 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR;
8063 Label COMPARE_TAIL_LONG;
8064 int pcmpmask = 0x19;
8065 if (ae == StrIntrinsicNode::LL) {
8066 pcmpmask &= ~0x01;
8067 }
8068
8069 // Setup to compare 16-chars (32-bytes) vectors,
8070 // start from first character again because it has aligned address.
8071 if (ae == StrIntrinsicNode::LL) {
8072 stride2 = 32;
8073 } else {
8074 stride2 = 16;
8075 }
8076 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8077 adr_stride = stride << scale;
8078 } else {
8079 adr_stride1 = 8; //stride << scale1;
8080 adr_stride2 = 16; //stride << scale2;
8081 }
8082
8083 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
8084 // rax and rdx are used by pcmpestri as elements counters
8085 movl(result, cnt2);
8086 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count
8087 jcc(Assembler::zero, COMPARE_TAIL_LONG);
8088
8089 // fast path : compare first 2 8-char vectors.
8090 bind(COMPARE_16_CHARS);
8091 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8092 movdqu(vec1, Address(str1, 0));
8093 } else {
8094 pmovzxbw(vec1, Address(str1, 0));
8095 }
8096 pcmpestri(vec1, Address(str2, 0), pcmpmask);
8097 jccb(Assembler::below, COMPARE_INDEX_CHAR);
8098
8099 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8100 movdqu(vec1, Address(str1, adr_stride));
8101 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask);
8102 } else {
8103 pmovzxbw(vec1, Address(str1, adr_stride1));
8104 pcmpestri(vec1, Address(str2, adr_stride2), pcmpmask);
8105 }
8106 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS);
8107 addl(cnt1, stride);
8108
8109 // Compare the characters at index in cnt1
8110 bind(COMPARE_INDEX_CHAR); // cnt1 has the offset of the mismatching character
8111 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
8112 subl(result, cnt2);
8113 jmp(POP_LABEL);
8114
8115 // Setup the registers to start vector comparison loop
8116 bind(COMPARE_WIDE_VECTORS);
8117 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8118 lea(str1, Address(str1, result, scale));
8119 lea(str2, Address(str2, result, scale));
8120 } else {
8121 lea(str1, Address(str1, result, scale1));
8122 lea(str2, Address(str2, result, scale2));
8123 }
8124 subl(result, stride2);
8125 subl(cnt2, stride2);
8126 jccb(Assembler::zero, COMPARE_WIDE_TAIL);
8127 negptr(result);
8128
8129 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest)
8130 bind(COMPARE_WIDE_VECTORS_LOOP);
8131 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8132 vmovdqu(vec1, Address(str1, result, scale));
8133 vpxor(vec1, Address(str2, result, scale));
8134 } else {
8135 vpmovzxbw(vec1, Address(str1, result, scale1), Assembler::AVX_256bit);
8136 vpxor(vec1, Address(str2, result, scale2));
8137 }
8138 vptest(vec1, vec1);
8139 jccb(Assembler::notZero, VECTOR_NOT_EQUAL);
8140 addptr(result, stride2);
8141 subl(cnt2, stride2);
8142 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP);
8143 // clean upper bits of YMM registers
8144 vpxor(vec1, vec1);
8145
8146 // compare wide vectors tail
8147 bind(COMPARE_WIDE_TAIL);
8148 testptr(result, result);
8149 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
8150
8151 movl(result, stride2);
8152 movl(cnt2, result);
8153 negptr(result);
8154 jmpb(COMPARE_WIDE_VECTORS_LOOP);
8155
8156 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors.
8157 bind(VECTOR_NOT_EQUAL);
8158 // clean upper bits of YMM registers
8159 vpxor(vec1, vec1);
8160 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8161 lea(str1, Address(str1, result, scale));
8162 lea(str2, Address(str2, result, scale));
8163 } else {
8164 lea(str1, Address(str1, result, scale1));
8165 lea(str2, Address(str2, result, scale2));
8166 }
8167 jmp(COMPARE_16_CHARS);
8168
8169 // Compare tail chars, length between 1 to 15 chars
8170 bind(COMPARE_TAIL_LONG);
8171 movl(cnt2, result);
8172 cmpl(cnt2, stride);
8173 jccb(Assembler::less, COMPARE_SMALL_STR);
8174
8175 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8176 movdqu(vec1, Address(str1, 0));
8177 } else {
8178 pmovzxbw(vec1, Address(str1, 0));
8179 }
8180 pcmpestri(vec1, Address(str2, 0), pcmpmask);
8181 jcc(Assembler::below, COMPARE_INDEX_CHAR);
8182 subptr(cnt2, stride);
8183 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
8184 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8185 lea(str1, Address(str1, result, scale));
8186 lea(str2, Address(str2, result, scale));
8187 } else {
8188 lea(str1, Address(str1, result, scale1));
8189 lea(str2, Address(str2, result, scale2));
8190 }
8191 negptr(cnt2);
8192 jmpb(WHILE_HEAD_LABEL);
8193
8194 bind(COMPARE_SMALL_STR);
8195 } else if (UseSSE42Intrinsics) {
8196 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
8197 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL;
8198 int pcmpmask = 0x19;
8199 // Setup to compare 8-char (16-byte) vectors,
8200 // start from first character again because it has aligned address.
8201 movl(result, cnt2);
8202 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count
8203 if (ae == StrIntrinsicNode::LL) {
8204 pcmpmask &= ~0x01;
8205 }
8206 jccb(Assembler::zero, COMPARE_TAIL);
8207 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8208 lea(str1, Address(str1, result, scale));
8209 lea(str2, Address(str2, result, scale));
8210 } else {
8211 lea(str1, Address(str1, result, scale1));
8212 lea(str2, Address(str2, result, scale2));
8213 }
8214 negptr(result);
8215
8216 // pcmpestri
8217 // inputs:
8218 // vec1- substring
8219 // rax - negative string length (elements count)
8220 // mem - scanned string
8221 // rdx - string length (elements count)
8222 // pcmpmask - cmp mode: 11000 (string compare with negated result)
8223 // + 00 (unsigned bytes) or + 01 (unsigned shorts)
8224 // outputs:
8225 // rcx - first mismatched element index
8226 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri");
8227
8228 bind(COMPARE_WIDE_VECTORS);
8229 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8230 movdqu(vec1, Address(str1, result, scale));
8231 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
8232 } else {
8233 pmovzxbw(vec1, Address(str1, result, scale1));
8234 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
8235 }
8236 // After pcmpestri cnt1(rcx) contains mismatched element index
8237
8238 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1
8239 addptr(result, stride);
8240 subptr(cnt2, stride);
8241 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS);
8242
8243 // compare wide vectors tail
8244 testptr(result, result);
8245 jccb(Assembler::zero, LENGTH_DIFF_LABEL);
8246
8247 movl(cnt2, stride);
8248 movl(result, stride);
8249 negptr(result);
8250 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8251 movdqu(vec1, Address(str1, result, scale));
8252 pcmpestri(vec1, Address(str2, result, scale), pcmpmask);
8253 } else {
8254 pmovzxbw(vec1, Address(str1, result, scale1));
8255 pcmpestri(vec1, Address(str2, result, scale2), pcmpmask);
8256 }
8257 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL);
8258
8259 // Mismatched characters in the vectors
8260 bind(VECTOR_NOT_EQUAL);
8261 addptr(cnt1, result);
8262 load_next_elements(result, cnt2, str1, str2, scale, scale1, scale2, cnt1, ae);
8263 subl(result, cnt2);
8264 jmpb(POP_LABEL);
8265
8266 bind(COMPARE_TAIL); // limit is zero
8267 movl(cnt2, result);
8268 // Fallthru to tail compare
8269 }
8270 // Shift str2 and str1 to the end of the arrays, negate min
8271 if (ae == StrIntrinsicNode::LL || ae == StrIntrinsicNode::UU) {
8272 lea(str1, Address(str1, cnt2, scale));
8273 lea(str2, Address(str2, cnt2, scale));
8274 } else {
8275 lea(str1, Address(str1, cnt2, scale1));
8276 lea(str2, Address(str2, cnt2, scale2));
8277 }
8278 decrementl(cnt2); // first character was compared already
8279 negptr(cnt2);
8280
8281 // Compare the rest of the elements
8282 bind(WHILE_HEAD_LABEL);
8283 load_next_elements(result, cnt1, str1, str2, scale, scale1, scale2, cnt2, ae);
8284 subl(result, cnt1);
8285 jccb(Assembler::notZero, POP_LABEL);
8286 increment(cnt2);
8287 jccb(Assembler::notZero, WHILE_HEAD_LABEL);
8288
8289 // Strings are equal up to min length. Return the length difference.
8290 bind(LENGTH_DIFF_LABEL);
8291 pop(result);
8292 if (ae == StrIntrinsicNode::UU) {
8293 // Divide diff by 2 to get number of chars
8294 sarl(result, 1);
8295 }
8296 jmpb(DONE_LABEL);
8297
8298 // Discard the stored length difference
8299 bind(POP_LABEL);
8300 pop(cnt1);
8301
8302 // That's it
8303 bind(DONE_LABEL);
8304 if(ae == StrIntrinsicNode::UL) {
8305 negl(result);
8306 }
8307 }
8308
8309 // Search for Non-ASCII character (Negative byte value) in a byte array,
8310 // return true if it has any and false otherwise.
8311 void MacroAssembler::has_negatives(Register ary1, Register len,
8312 Register result, Register tmp1,
8313 XMMRegister vec1, XMMRegister vec2) {
8314
8315 // rsi: byte array
8316 // rcx: len
8317 // rax: result
8318 ShortBranchVerifier sbv(this);
8319 assert_different_registers(ary1, len, result, tmp1);
8320 assert_different_registers(vec1, vec2);
8321 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_CHAR, COMPARE_VECTORS, COMPARE_BYTE;
8322
8323 // len == 0
8324 testl(len, len);
8325 jcc(Assembler::zero, FALSE_LABEL);
8326
8327 movl(result, len); // copy
8328
8329 if (UseAVX >= 2 && UseSSE >= 2) {
8330 // With AVX2, use 32-byte vector compare
8331 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8332
8333 // Compare 32-byte vectors
8334 andl(result, 0x0000001f); // tail count (in bytes)
8335 andl(len, 0xffffffe0); // vector count (in bytes)
8336 jccb(Assembler::zero, COMPARE_TAIL);
8337
8338 lea(ary1, Address(ary1, len, Address::times_1));
8339 negptr(len);
8340
8341 movl(tmp1, 0x80808080); // create mask to test for Unicode chars in vector
8342 movdl(vec2, tmp1);
8343 vpbroadcastd(vec2, vec2);
8344
8345 bind(COMPARE_WIDE_VECTORS);
8346 vmovdqu(vec1, Address(ary1, len, Address::times_1));
8347 vptest(vec1, vec2);
8348 jccb(Assembler::notZero, TRUE_LABEL);
8349 addptr(len, 32);
8350 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8351
8352 testl(result, result);
8353 jccb(Assembler::zero, FALSE_LABEL);
8354
8355 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
8356 vptest(vec1, vec2);
8357 jccb(Assembler::notZero, TRUE_LABEL);
8358 jmpb(FALSE_LABEL);
8359
8360 bind(COMPARE_TAIL); // len is zero
8361 movl(len, result);
8362 // Fallthru to tail compare
8363 } else if (UseSSE42Intrinsics) {
8364 assert(UseSSE >= 4, "SSE4 must be for SSE4.2 intrinsics to be available");
8365 // With SSE4.2, use double quad vector compare
8366 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8367
8368 // Compare 16-byte vectors
8369 andl(result, 0x0000000f); // tail count (in bytes)
8370 andl(len, 0xfffffff0); // vector count (in bytes)
8371 jccb(Assembler::zero, COMPARE_TAIL);
8372
8373 lea(ary1, Address(ary1, len, Address::times_1));
8374 negptr(len);
8375
8376 movl(tmp1, 0x80808080);
8377 movdl(vec2, tmp1);
8378 pshufd(vec2, vec2, 0);
8379
8380 bind(COMPARE_WIDE_VECTORS);
8381 movdqu(vec1, Address(ary1, len, Address::times_1));
8382 ptest(vec1, vec2);
8383 jccb(Assembler::notZero, TRUE_LABEL);
8384 addptr(len, 16);
8385 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8386
8387 testl(result, result);
8388 jccb(Assembler::zero, FALSE_LABEL);
8389
8390 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
8391 ptest(vec1, vec2);
8392 jccb(Assembler::notZero, TRUE_LABEL);
8393 jmpb(FALSE_LABEL);
8394
8395 bind(COMPARE_TAIL); // len is zero
8396 movl(len, result);
8397 // Fallthru to tail compare
8398 }
8399
8400 // Compare 4-byte vectors
8401 andl(len, 0xfffffffc); // vector count (in bytes)
8402 jccb(Assembler::zero, COMPARE_CHAR);
8403
8404 lea(ary1, Address(ary1, len, Address::times_1));
8405 negptr(len);
8406
8407 bind(COMPARE_VECTORS);
8408 movl(tmp1, Address(ary1, len, Address::times_1));
8409 andl(tmp1, 0x80808080);
8410 jccb(Assembler::notZero, TRUE_LABEL);
8411 addptr(len, 4);
8412 jcc(Assembler::notZero, COMPARE_VECTORS);
8413
8414 // Compare trailing char (final 2 bytes), if any
8415 bind(COMPARE_CHAR);
8416 testl(result, 0x2); // tail char
8417 jccb(Assembler::zero, COMPARE_BYTE);
8418 load_unsigned_short(tmp1, Address(ary1, 0));
8419 andl(tmp1, 0x00008080);
8420 jccb(Assembler::notZero, TRUE_LABEL);
8421 subptr(result, 2);
8422 lea(ary1, Address(ary1, 2));
8423
8424 bind(COMPARE_BYTE);
8425 testl(result, 0x1); // tail byte
8426 jccb(Assembler::zero, FALSE_LABEL);
8427 load_unsigned_byte(tmp1, Address(ary1, 0));
8428 andl(tmp1, 0x00000080);
8429 jccb(Assembler::notEqual, TRUE_LABEL);
8430 jmpb(FALSE_LABEL);
8431
8432 bind(TRUE_LABEL);
8433 movl(result, 1); // return true
8434 jmpb(DONE);
8435
8436 bind(FALSE_LABEL);
8437 xorl(result, result); // return false
8438
8439 // That's it
8440 bind(DONE);
8441 if (UseAVX >= 2 && UseSSE >= 2) {
8442 // clean upper bits of YMM registers
8443 vpxor(vec1, vec1);
8444 vpxor(vec2, vec2);
8445 }
8446 }
8447
8448 // Compare char[] or byte[] arrays aligned to 4 bytes or substrings.
8449 void MacroAssembler::arrays_equals(bool is_array_equ, Register ary1, Register ary2,
8450 Register limit, Register result, Register chr,
8451 XMMRegister vec1, XMMRegister vec2, bool is_char) {
8452 ShortBranchVerifier sbv(this);
8453 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR, COMPARE_BYTE;
8454
8455 int length_offset = arrayOopDesc::length_offset_in_bytes();
8456 int base_offset = arrayOopDesc::base_offset_in_bytes(is_char ? T_CHAR : T_BYTE);
8457
8458 if (is_array_equ) {
8459 // Check the input args
8460 cmpptr(ary1, ary2);
8461 jcc(Assembler::equal, TRUE_LABEL);
8462
8463 // Need additional checks for arrays_equals.
8464 testptr(ary1, ary1);
8465 jcc(Assembler::zero, FALSE_LABEL);
8466 testptr(ary2, ary2);
8467 jcc(Assembler::zero, FALSE_LABEL);
8468
8469 // Check the lengths
8470 movl(limit, Address(ary1, length_offset));
8471 cmpl(limit, Address(ary2, length_offset));
8472 jcc(Assembler::notEqual, FALSE_LABEL);
8473 }
8474
8475 // count == 0
8476 testl(limit, limit);
8477 jcc(Assembler::zero, TRUE_LABEL);
8478
8479 if (is_array_equ) {
8480 // Load array address
8481 lea(ary1, Address(ary1, base_offset));
8482 lea(ary2, Address(ary2, base_offset));
8483 }
8484
8485 if (is_array_equ && is_char) {
8486 // arrays_equals when used for char[].
8487 shll(limit, 1); // byte count != 0
8488 }
8489 movl(result, limit); // copy
8490
8491 if (UseAVX >= 2) {
8492 // With AVX2, use 32-byte vector compare
8493 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8494
8495 // Compare 32-byte vectors
8496 andl(result, 0x0000001f); // tail count (in bytes)
8497 andl(limit, 0xffffffe0); // vector count (in bytes)
8498 jccb(Assembler::zero, COMPARE_TAIL);
8499
8500 lea(ary1, Address(ary1, limit, Address::times_1));
8501 lea(ary2, Address(ary2, limit, Address::times_1));
8502 negptr(limit);
8503
8504 bind(COMPARE_WIDE_VECTORS);
8505 vmovdqu(vec1, Address(ary1, limit, Address::times_1));
8506 vmovdqu(vec2, Address(ary2, limit, Address::times_1));
8507 vpxor(vec1, vec2);
8508
8509 vptest(vec1, vec1);
8510 jccb(Assembler::notZero, FALSE_LABEL);
8511 addptr(limit, 32);
8512 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8513
8514 testl(result, result);
8515 jccb(Assembler::zero, TRUE_LABEL);
8516
8517 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32));
8518 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32));
8519 vpxor(vec1, vec2);
8520
8521 vptest(vec1, vec1);
8522 jccb(Assembler::notZero, FALSE_LABEL);
8523 jmpb(TRUE_LABEL);
8524
8525 bind(COMPARE_TAIL); // limit is zero
8526 movl(limit, result);
8527 // Fallthru to tail compare
8528 } else if (UseSSE42Intrinsics) {
8529 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
8530 // With SSE4.2, use double quad vector compare
8531 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL;
8532
8533 // Compare 16-byte vectors
8534 andl(result, 0x0000000f); // tail count (in bytes)
8535 andl(limit, 0xfffffff0); // vector count (in bytes)
8536 jccb(Assembler::zero, COMPARE_TAIL);
8537
8538 lea(ary1, Address(ary1, limit, Address::times_1));
8539 lea(ary2, Address(ary2, limit, Address::times_1));
8540 negptr(limit);
8541
8542 bind(COMPARE_WIDE_VECTORS);
8543 movdqu(vec1, Address(ary1, limit, Address::times_1));
8544 movdqu(vec2, Address(ary2, limit, Address::times_1));
8545 pxor(vec1, vec2);
8546
8547 ptest(vec1, vec1);
8548 jccb(Assembler::notZero, FALSE_LABEL);
8549 addptr(limit, 16);
8550 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS);
8551
8552 testl(result, result);
8553 jccb(Assembler::zero, TRUE_LABEL);
8554
8555 movdqu(vec1, Address(ary1, result, Address::times_1, -16));
8556 movdqu(vec2, Address(ary2, result, Address::times_1, -16));
8557 pxor(vec1, vec2);
8558
8559 ptest(vec1, vec1);
8560 jccb(Assembler::notZero, FALSE_LABEL);
8561 jmpb(TRUE_LABEL);
8562
8563 bind(COMPARE_TAIL); // limit is zero
8564 movl(limit, result);
8565 // Fallthru to tail compare
8566 }
8567
8568 // Compare 4-byte vectors
8569 andl(limit, 0xfffffffc); // vector count (in bytes)
8570 jccb(Assembler::zero, COMPARE_CHAR);
8571
8572 lea(ary1, Address(ary1, limit, Address::times_1));
8573 lea(ary2, Address(ary2, limit, Address::times_1));
8574 negptr(limit);
8575
8576 bind(COMPARE_VECTORS);
8577 movl(chr, Address(ary1, limit, Address::times_1));
8578 cmpl(chr, Address(ary2, limit, Address::times_1));
8579 jccb(Assembler::notEqual, FALSE_LABEL);
8580 addptr(limit, 4);
8581 jcc(Assembler::notZero, COMPARE_VECTORS);
8582
8583 // Compare trailing char (final 2 bytes), if any
8584 bind(COMPARE_CHAR);
8585 testl(result, 0x2); // tail char
8586 jccb(Assembler::zero, COMPARE_BYTE);
8587 load_unsigned_short(chr, Address(ary1, 0));
8588 load_unsigned_short(limit, Address(ary2, 0));
8589 cmpl(chr, limit);
8590 jccb(Assembler::notEqual, FALSE_LABEL);
8591
8592 if (is_array_equ && is_char) {
8593 bind(COMPARE_BYTE);
8594 } else {
8595 lea(ary1, Address(ary1, 2));
8596 lea(ary2, Address(ary2, 2));
8597
8598 bind(COMPARE_BYTE);
8599 testl(result, 0x1); // tail byte
8600 jccb(Assembler::zero, TRUE_LABEL);
8601 load_unsigned_byte(chr, Address(ary1, 0));
8602 load_unsigned_byte(limit, Address(ary2, 0));
8603 cmpl(chr, limit);
8604 jccb(Assembler::notEqual, FALSE_LABEL);
8605 }
8606 bind(TRUE_LABEL);
8607 movl(result, 1); // return true
8608 jmpb(DONE);
8609
8610 bind(FALSE_LABEL);
8611 xorl(result, result); // return false
8612
8613 // That's it
8614 bind(DONE);
8615 if (UseAVX >= 2) {
8616 // clean upper bits of YMM registers
8617 vpxor(vec1, vec1);
8618 vpxor(vec2, vec2);
8619 }
8620 }
8621
8622 #endif
8623
8624 void MacroAssembler::generate_fill(BasicType t, bool aligned,
8625 Register to, Register value, Register count,
8626 Register rtmp, XMMRegister xtmp) {
8627 ShortBranchVerifier sbv(this);
8628 assert_different_registers(to, value, count, rtmp);
8629 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
8630 Label L_fill_2_bytes, L_fill_4_bytes;
8631
8632 int shift = -1;
8633 switch (t) {
8634 case T_BYTE:
8635 shift = 2;
8636 break;
8637 case T_SHORT:
8638 shift = 1;
8639 break;
8640 case T_INT:
8641 shift = 0;
8642 break;
8643 default: ShouldNotReachHere();
8644 }
8645
8646 if (t == T_BYTE) {
8647 andl(value, 0xff);
8648 movl(rtmp, value);
8649 shll(rtmp, 8);
8650 orl(value, rtmp);
8651 }
8652 if (t == T_SHORT) {
8653 andl(value, 0xffff);
8654 }
8655 if (t == T_BYTE || t == T_SHORT) {
8656 movl(rtmp, value);
8657 shll(rtmp, 16);
8658 orl(value, rtmp);
8659 }
8660
8661 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
8662 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp
8663 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) {
8664 // align source address at 4 bytes address boundary
8665 if (t == T_BYTE) {
8666 // One byte misalignment happens only for byte arrays
8667 testptr(to, 1);
8668 jccb(Assembler::zero, L_skip_align1);
8669 movb(Address(to, 0), value);
8670 increment(to);
8671 decrement(count);
8672 BIND(L_skip_align1);
8673 }
8674 // Two bytes misalignment happens only for byte and short (char) arrays
8675 testptr(to, 2);
8676 jccb(Assembler::zero, L_skip_align2);
8677 movw(Address(to, 0), value);
8678 addptr(to, 2);
8679 subl(count, 1<<(shift-1));
8680 BIND(L_skip_align2);
8681 }
8682 if (UseSSE < 2) {
8683 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
8684 // Fill 32-byte chunks
8685 subl(count, 8 << shift);
8686 jcc(Assembler::less, L_check_fill_8_bytes);
8687 align(16);
8688
8689 BIND(L_fill_32_bytes_loop);
8690
8691 for (int i = 0; i < 32; i += 4) {
8692 movl(Address(to, i), value);
8693 }
8694
8695 addptr(to, 32);
8696 subl(count, 8 << shift);
8697 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
8698 BIND(L_check_fill_8_bytes);
8699 addl(count, 8 << shift);
8700 jccb(Assembler::zero, L_exit);
8701 jmpb(L_fill_8_bytes);
8702
8703 //
8704 // length is too short, just fill qwords
8705 //
8706 BIND(L_fill_8_bytes_loop);
8707 movl(Address(to, 0), value);
8708 movl(Address(to, 4), value);
8709 addptr(to, 8);
8710 BIND(L_fill_8_bytes);
8711 subl(count, 1 << (shift + 1));
8712 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
8713 // fall through to fill 4 bytes
8714 } else {
8715 Label L_fill_32_bytes;
8716 if (!UseUnalignedLoadStores) {
8717 // align to 8 bytes, we know we are 4 byte aligned to start
8718 testptr(to, 4);
8719 jccb(Assembler::zero, L_fill_32_bytes);
8720 movl(Address(to, 0), value);
8721 addptr(to, 4);
8722 subl(count, 1<<shift);
8723 }
8724 BIND(L_fill_32_bytes);
8725 {
8726 assert( UseSSE >= 2, "supported cpu only" );
8727 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes;
8728 if (UseAVX > 2) {
8729 movl(rtmp, 0xffff);
8730 kmovwl(k1, rtmp);
8731 }
8732 movdl(xtmp, value);
8733 if (UseAVX > 2 && UseUnalignedLoadStores) {
8734 // Fill 64-byte chunks
8735 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
8736 evpbroadcastd(xtmp, xtmp, Assembler::AVX_512bit);
8737
8738 subl(count, 16 << shift);
8739 jcc(Assembler::less, L_check_fill_32_bytes);
8740 align(16);
8741
8742 BIND(L_fill_64_bytes_loop);
8743 evmovdqul(Address(to, 0), xtmp, Assembler::AVX_512bit);
8744 addptr(to, 64);
8745 subl(count, 16 << shift);
8746 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
8747
8748 BIND(L_check_fill_32_bytes);
8749 addl(count, 8 << shift);
8750 jccb(Assembler::less, L_check_fill_8_bytes);
8751 vmovdqu(Address(to, 0), xtmp);
8752 addptr(to, 32);
8753 subl(count, 8 << shift);
8754
8755 BIND(L_check_fill_8_bytes);
8756 } else if (UseAVX == 2 && UseUnalignedLoadStores) {
8757 // Fill 64-byte chunks
8758 Label L_fill_64_bytes_loop, L_check_fill_32_bytes;
8759 vpbroadcastd(xtmp, xtmp);
8760
8761 subl(count, 16 << shift);
8762 jcc(Assembler::less, L_check_fill_32_bytes);
8763 align(16);
8764
8765 BIND(L_fill_64_bytes_loop);
8766 vmovdqu(Address(to, 0), xtmp);
8767 vmovdqu(Address(to, 32), xtmp);
8768 addptr(to, 64);
8769 subl(count, 16 << shift);
8770 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop);
8771
8772 BIND(L_check_fill_32_bytes);
8773 addl(count, 8 << shift);
8774 jccb(Assembler::less, L_check_fill_8_bytes);
8775 vmovdqu(Address(to, 0), xtmp);
8776 addptr(to, 32);
8777 subl(count, 8 << shift);
8778
8779 BIND(L_check_fill_8_bytes);
8780 // clean upper bits of YMM registers
8781 movdl(xtmp, value);
8782 pshufd(xtmp, xtmp, 0);
8783 } else {
8784 // Fill 32-byte chunks
8785 pshufd(xtmp, xtmp, 0);
8786
8787 subl(count, 8 << shift);
8788 jcc(Assembler::less, L_check_fill_8_bytes);
8789 align(16);
8790
8791 BIND(L_fill_32_bytes_loop);
8792
8793 if (UseUnalignedLoadStores) {
8794 movdqu(Address(to, 0), xtmp);
8795 movdqu(Address(to, 16), xtmp);
8796 } else {
8797 movq(Address(to, 0), xtmp);
8798 movq(Address(to, 8), xtmp);
8799 movq(Address(to, 16), xtmp);
8800 movq(Address(to, 24), xtmp);
8801 }
8802
8803 addptr(to, 32);
8804 subl(count, 8 << shift);
8805 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop);
8806
8807 BIND(L_check_fill_8_bytes);
8808 }
8809 addl(count, 8 << shift);
8810 jccb(Assembler::zero, L_exit);
8811 jmpb(L_fill_8_bytes);
8812
8813 //
8814 // length is too short, just fill qwords
8815 //
8816 BIND(L_fill_8_bytes_loop);
8817 movq(Address(to, 0), xtmp);
8818 addptr(to, 8);
8819 BIND(L_fill_8_bytes);
8820 subl(count, 1 << (shift + 1));
8821 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop);
8822 }
8823 }
8824 // fill trailing 4 bytes
8825 BIND(L_fill_4_bytes);
8826 testl(count, 1<<shift);
8827 jccb(Assembler::zero, L_fill_2_bytes);
8828 movl(Address(to, 0), value);
8829 if (t == T_BYTE || t == T_SHORT) {
8830 addptr(to, 4);
8831 BIND(L_fill_2_bytes);
8832 // fill trailing 2 bytes
8833 testl(count, 1<<(shift-1));
8834 jccb(Assembler::zero, L_fill_byte);
8835 movw(Address(to, 0), value);
8836 if (t == T_BYTE) {
8837 addptr(to, 2);
8838 BIND(L_fill_byte);
8839 // fill trailing byte
8840 testl(count, 1);
8841 jccb(Assembler::zero, L_exit);
8842 movb(Address(to, 0), value);
8843 } else {
8844 BIND(L_fill_byte);
8845 }
8846 } else {
8847 BIND(L_fill_2_bytes);
8848 }
8849 BIND(L_exit);
8850 }
8851
8852 // encode char[] to byte[] in ISO_8859_1
8853 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len,
8854 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
8855 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
8856 Register tmp5, Register result) {
8857 // rsi: src
8858 // rdi: dst
8859 // rdx: len
8860 // rcx: tmp5
8861 // rax: result
8862 ShortBranchVerifier sbv(this);
8863 assert_different_registers(src, dst, len, tmp5, result);
8864 Label L_done, L_copy_1_char, L_copy_1_char_exit;
8865
8866 // set result
8867 xorl(result, result);
8868 // check for zero length
8869 testl(len, len);
8870 jcc(Assembler::zero, L_done);
8871 movl(result, len);
8872
8873 // Setup pointers
8874 lea(src, Address(src, len, Address::times_2)); // char[]
8875 lea(dst, Address(dst, len, Address::times_1)); // byte[]
8876 negptr(len);
8877
8878 if (UseSSE42Intrinsics || UseAVX >= 2) {
8879 assert(UseSSE42Intrinsics ? UseSSE >= 4 : true, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
8880 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit;
8881 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit;
8882
8883 if (UseAVX >= 2) {
8884 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit;
8885 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
8886 movdl(tmp1Reg, tmp5);
8887 vpbroadcastd(tmp1Reg, tmp1Reg);
8888 jmpb(L_chars_32_check);
8889
8890 bind(L_copy_32_chars);
8891 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64));
8892 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32));
8893 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
8894 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
8895 jccb(Assembler::notZero, L_copy_32_chars_exit);
8896 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector_len */ 1);
8897 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector_len */ 1);
8898 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg);
8899
8900 bind(L_chars_32_check);
8901 addptr(len, 32);
8902 jccb(Assembler::lessEqual, L_copy_32_chars);
8903
8904 bind(L_copy_32_chars_exit);
8905 subptr(len, 16);
8906 jccb(Assembler::greater, L_copy_16_chars_exit);
8907
8908 } else if (UseSSE42Intrinsics) {
8909 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector
8910 movdl(tmp1Reg, tmp5);
8911 pshufd(tmp1Reg, tmp1Reg, 0);
8912 jmpb(L_chars_16_check);
8913 }
8914
8915 bind(L_copy_16_chars);
8916 if (UseAVX >= 2) {
8917 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32));
8918 vptest(tmp2Reg, tmp1Reg);
8919 jccb(Assembler::notZero, L_copy_16_chars_exit);
8920 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector_len */ 1);
8921 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector_len */ 1);
8922 } else {
8923 if (UseAVX > 0) {
8924 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
8925 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
8926 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector_len */ 0);
8927 } else {
8928 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32));
8929 por(tmp2Reg, tmp3Reg);
8930 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16));
8931 por(tmp2Reg, tmp4Reg);
8932 }
8933 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
8934 jccb(Assembler::notZero, L_copy_16_chars_exit);
8935 packuswb(tmp3Reg, tmp4Reg);
8936 }
8937 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg);
8938
8939 bind(L_chars_16_check);
8940 addptr(len, 16);
8941 jccb(Assembler::lessEqual, L_copy_16_chars);
8942
8943 bind(L_copy_16_chars_exit);
8944 if (UseAVX >= 2) {
8945 // clean upper bits of YMM registers
8946 vpxor(tmp2Reg, tmp2Reg);
8947 vpxor(tmp3Reg, tmp3Reg);
8948 vpxor(tmp4Reg, tmp4Reg);
8949 movdl(tmp1Reg, tmp5);
8950 pshufd(tmp1Reg, tmp1Reg, 0);
8951 }
8952 subptr(len, 8);
8953 jccb(Assembler::greater, L_copy_8_chars_exit);
8954
8955 bind(L_copy_8_chars);
8956 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16));
8957 ptest(tmp3Reg, tmp1Reg);
8958 jccb(Assembler::notZero, L_copy_8_chars_exit);
8959 packuswb(tmp3Reg, tmp1Reg);
8960 movq(Address(dst, len, Address::times_1, -8), tmp3Reg);
8961 addptr(len, 8);
8962 jccb(Assembler::lessEqual, L_copy_8_chars);
8963
8964 bind(L_copy_8_chars_exit);
8965 subptr(len, 8);
8966 jccb(Assembler::zero, L_done);
8967 }
8968
8969 bind(L_copy_1_char);
8970 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0));
8971 testl(tmp5, 0xff00); // check if Unicode char
8972 jccb(Assembler::notZero, L_copy_1_char_exit);
8973 movb(Address(dst, len, Address::times_1, 0), tmp5);
8974 addptr(len, 1);
8975 jccb(Assembler::less, L_copy_1_char);
8976
8977 bind(L_copy_1_char_exit);
8978 addptr(result, len); // len is negative count of not processed elements
8979 bind(L_done);
8980 }
8981
8982 #ifdef _LP64
8983 /**
8984 * Helper for multiply_to_len().
8985 */
8986 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo, Register src1, Register src2) {
8987 addq(dest_lo, src1);
8988 adcq(dest_hi, 0);
8989 addq(dest_lo, src2);
8990 adcq(dest_hi, 0);
8991 }
8992
8993 /**
8994 * Multiply 64 bit by 64 bit first loop.
8995 */
8996 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
8997 Register y, Register y_idx, Register z,
8998 Register carry, Register product,
8999 Register idx, Register kdx) {
9000 //
9001 // jlong carry, x[], y[], z[];
9002 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
9003 // huge_128 product = y[idx] * x[xstart] + carry;
9004 // z[kdx] = (jlong)product;
9005 // carry = (jlong)(product >>> 64);
9006 // }
9007 // z[xstart] = carry;
9008 //
9009
9010 Label L_first_loop, L_first_loop_exit;
9011 Label L_one_x, L_one_y, L_multiply;
9012
9013 decrementl(xstart);
9014 jcc(Assembler::negative, L_one_x);
9015
9016 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
9017 rorq(x_xstart, 32); // convert big-endian to little-endian
9018
9019 bind(L_first_loop);
9020 decrementl(idx);
9021 jcc(Assembler::negative, L_first_loop_exit);
9022 decrementl(idx);
9023 jcc(Assembler::negative, L_one_y);
9024 movq(y_idx, Address(y, idx, Address::times_4, 0));
9025 rorq(y_idx, 32); // convert big-endian to little-endian
9026 bind(L_multiply);
9027 movq(product, x_xstart);
9028 mulq(y_idx); // product(rax) * y_idx -> rdx:rax
9029 addq(product, carry);
9030 adcq(rdx, 0);
9031 subl(kdx, 2);
9032 movl(Address(z, kdx, Address::times_4, 4), product);
9033 shrq(product, 32);
9034 movl(Address(z, kdx, Address::times_4, 0), product);
9035 movq(carry, rdx);
9036 jmp(L_first_loop);
9037
9038 bind(L_one_y);
9039 movl(y_idx, Address(y, 0));
9040 jmp(L_multiply);
9041
9042 bind(L_one_x);
9043 movl(x_xstart, Address(x, 0));
9044 jmp(L_first_loop);
9045
9046 bind(L_first_loop_exit);
9047 }
9048
9049 /**
9050 * Multiply 64 bit by 64 bit and add 128 bit.
9051 */
9052 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y, Register z,
9053 Register yz_idx, Register idx,
9054 Register carry, Register product, int offset) {
9055 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
9056 // z[kdx] = (jlong)product;
9057
9058 movq(yz_idx, Address(y, idx, Address::times_4, offset));
9059 rorq(yz_idx, 32); // convert big-endian to little-endian
9060 movq(product, x_xstart);
9061 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
9062 movq(yz_idx, Address(z, idx, Address::times_4, offset));
9063 rorq(yz_idx, 32); // convert big-endian to little-endian
9064
9065 add2_with_carry(rdx, product, carry, yz_idx);
9066
9067 movl(Address(z, idx, Address::times_4, offset+4), product);
9068 shrq(product, 32);
9069 movl(Address(z, idx, Address::times_4, offset), product);
9070
9071 }
9072
9073 /**
9074 * Multiply 128 bit by 128 bit. Unrolled inner loop.
9075 */
9076 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart, Register y, Register z,
9077 Register yz_idx, Register idx, Register jdx,
9078 Register carry, Register product,
9079 Register carry2) {
9080 // jlong carry, x[], y[], z[];
9081 // int kdx = ystart+1;
9082 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
9083 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
9084 // z[kdx+idx+1] = (jlong)product;
9085 // jlong carry2 = (jlong)(product >>> 64);
9086 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
9087 // z[kdx+idx] = (jlong)product;
9088 // carry = (jlong)(product >>> 64);
9089 // }
9090 // idx += 2;
9091 // if (idx > 0) {
9092 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
9093 // z[kdx+idx] = (jlong)product;
9094 // carry = (jlong)(product >>> 64);
9095 // }
9096 //
9097
9098 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
9099
9100 movl(jdx, idx);
9101 andl(jdx, 0xFFFFFFFC);
9102 shrl(jdx, 2);
9103
9104 bind(L_third_loop);
9105 subl(jdx, 1);
9106 jcc(Assembler::negative, L_third_loop_exit);
9107 subl(idx, 4);
9108
9109 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
9110 movq(carry2, rdx);
9111
9112 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
9113 movq(carry, rdx);
9114 jmp(L_third_loop);
9115
9116 bind (L_third_loop_exit);
9117
9118 andl (idx, 0x3);
9119 jcc(Assembler::zero, L_post_third_loop_done);
9120
9121 Label L_check_1;
9122 subl(idx, 2);
9123 jcc(Assembler::negative, L_check_1);
9124
9125 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
9126 movq(carry, rdx);
9127
9128 bind (L_check_1);
9129 addl (idx, 0x2);
9130 andl (idx, 0x1);
9131 subl(idx, 1);
9132 jcc(Assembler::negative, L_post_third_loop_done);
9133
9134 movl(yz_idx, Address(y, idx, Address::times_4, 0));
9135 movq(product, x_xstart);
9136 mulq(yz_idx); // product(rax) * yz_idx -> rdx:product(rax)
9137 movl(yz_idx, Address(z, idx, Address::times_4, 0));
9138
9139 add2_with_carry(rdx, product, yz_idx, carry);
9140
9141 movl(Address(z, idx, Address::times_4, 0), product);
9142 shrq(product, 32);
9143
9144 shlq(rdx, 32);
9145 orq(product, rdx);
9146 movq(carry, product);
9147
9148 bind(L_post_third_loop_done);
9149 }
9150
9151 /**
9152 * Multiply 128 bit by 128 bit using BMI2. Unrolled inner loop.
9153 *
9154 */
9155 void MacroAssembler::multiply_128_x_128_bmi2_loop(Register y, Register z,
9156 Register carry, Register carry2,
9157 Register idx, Register jdx,
9158 Register yz_idx1, Register yz_idx2,
9159 Register tmp, Register tmp3, Register tmp4) {
9160 assert(UseBMI2Instructions, "should be used only when BMI2 is available");
9161
9162 // jlong carry, x[], y[], z[];
9163 // int kdx = ystart+1;
9164 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
9165 // huge_128 tmp3 = (y[idx+1] * rdx) + z[kdx+idx+1] + carry;
9166 // jlong carry2 = (jlong)(tmp3 >>> 64);
9167 // huge_128 tmp4 = (y[idx] * rdx) + z[kdx+idx] + carry2;
9168 // carry = (jlong)(tmp4 >>> 64);
9169 // z[kdx+idx+1] = (jlong)tmp3;
9170 // z[kdx+idx] = (jlong)tmp4;
9171 // }
9172 // idx += 2;
9173 // if (idx > 0) {
9174 // yz_idx1 = (y[idx] * rdx) + z[kdx+idx] + carry;
9175 // z[kdx+idx] = (jlong)yz_idx1;
9176 // carry = (jlong)(yz_idx1 >>> 64);
9177 // }
9178 //
9179
9180 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
9181
9182 movl(jdx, idx);
9183 andl(jdx, 0xFFFFFFFC);
9184 shrl(jdx, 2);
9185
9186 bind(L_third_loop);
9187 subl(jdx, 1);
9188 jcc(Assembler::negative, L_third_loop_exit);
9189 subl(idx, 4);
9190
9191 movq(yz_idx1, Address(y, idx, Address::times_4, 8));
9192 rorxq(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
9193 movq(yz_idx2, Address(y, idx, Address::times_4, 0));
9194 rorxq(yz_idx2, yz_idx2, 32);
9195
9196 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
9197 mulxq(carry2, tmp, yz_idx2); // yz_idx2 * rdx -> carry2:tmp
9198
9199 movq(yz_idx1, Address(z, idx, Address::times_4, 8));
9200 rorxq(yz_idx1, yz_idx1, 32);
9201 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
9202 rorxq(yz_idx2, yz_idx2, 32);
9203
9204 if (VM_Version::supports_adx()) {
9205 adcxq(tmp3, carry);
9206 adoxq(tmp3, yz_idx1);
9207
9208 adcxq(tmp4, tmp);
9209 adoxq(tmp4, yz_idx2);
9210
9211 movl(carry, 0); // does not affect flags
9212 adcxq(carry2, carry);
9213 adoxq(carry2, carry);
9214 } else {
9215 add2_with_carry(tmp4, tmp3, carry, yz_idx1);
9216 add2_with_carry(carry2, tmp4, tmp, yz_idx2);
9217 }
9218 movq(carry, carry2);
9219
9220 movl(Address(z, idx, Address::times_4, 12), tmp3);
9221 shrq(tmp3, 32);
9222 movl(Address(z, idx, Address::times_4, 8), tmp3);
9223
9224 movl(Address(z, idx, Address::times_4, 4), tmp4);
9225 shrq(tmp4, 32);
9226 movl(Address(z, idx, Address::times_4, 0), tmp4);
9227
9228 jmp(L_third_loop);
9229
9230 bind (L_third_loop_exit);
9231
9232 andl (idx, 0x3);
9233 jcc(Assembler::zero, L_post_third_loop_done);
9234
9235 Label L_check_1;
9236 subl(idx, 2);
9237 jcc(Assembler::negative, L_check_1);
9238
9239 movq(yz_idx1, Address(y, idx, Address::times_4, 0));
9240 rorxq(yz_idx1, yz_idx1, 32);
9241 mulxq(tmp4, tmp3, yz_idx1); // yz_idx1 * rdx -> tmp4:tmp3
9242 movq(yz_idx2, Address(z, idx, Address::times_4, 0));
9243 rorxq(yz_idx2, yz_idx2, 32);
9244
9245 add2_with_carry(tmp4, tmp3, carry, yz_idx2);
9246
9247 movl(Address(z, idx, Address::times_4, 4), tmp3);
9248 shrq(tmp3, 32);
9249 movl(Address(z, idx, Address::times_4, 0), tmp3);
9250 movq(carry, tmp4);
9251
9252 bind (L_check_1);
9253 addl (idx, 0x2);
9254 andl (idx, 0x1);
9255 subl(idx, 1);
9256 jcc(Assembler::negative, L_post_third_loop_done);
9257 movl(tmp4, Address(y, idx, Address::times_4, 0));
9258 mulxq(carry2, tmp3, tmp4); // tmp4 * rdx -> carry2:tmp3
9259 movl(tmp4, Address(z, idx, Address::times_4, 0));
9260
9261 add2_with_carry(carry2, tmp3, tmp4, carry);
9262
9263 movl(Address(z, idx, Address::times_4, 0), tmp3);
9264 shrq(tmp3, 32);
9265
9266 shlq(carry2, 32);
9267 orq(tmp3, carry2);
9268 movq(carry, tmp3);
9269
9270 bind(L_post_third_loop_done);
9271 }
9272
9273 /**
9274 * Code for BigInteger::multiplyToLen() instrinsic.
9275 *
9276 * rdi: x
9277 * rax: xlen
9278 * rsi: y
9279 * rcx: ylen
9280 * r8: z
9281 * r11: zlen
9282 * r12: tmp1
9283 * r13: tmp2
9284 * r14: tmp3
9285 * r15: tmp4
9286 * rbx: tmp5
9287 *
9288 */
9289 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen, Register z, Register zlen,
9290 Register tmp1, Register tmp2, Register tmp3, Register tmp4, Register tmp5) {
9291 ShortBranchVerifier sbv(this);
9292 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx);
9293
9294 push(tmp1);
9295 push(tmp2);
9296 push(tmp3);
9297 push(tmp4);
9298 push(tmp5);
9299
9300 push(xlen);
9301 push(zlen);
9302
9303 const Register idx = tmp1;
9304 const Register kdx = tmp2;
9305 const Register xstart = tmp3;
9306
9307 const Register y_idx = tmp4;
9308 const Register carry = tmp5;
9309 const Register product = xlen;
9310 const Register x_xstart = zlen; // reuse register
9311
9312 // First Loop.
9313 //
9314 // final static long LONG_MASK = 0xffffffffL;
9315 // int xstart = xlen - 1;
9316 // int ystart = ylen - 1;
9317 // long carry = 0;
9318 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
9319 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
9320 // z[kdx] = (int)product;
9321 // carry = product >>> 32;
9322 // }
9323 // z[xstart] = (int)carry;
9324 //
9325
9326 movl(idx, ylen); // idx = ylen;
9327 movl(kdx, zlen); // kdx = xlen+ylen;
9328 xorq(carry, carry); // carry = 0;
9329
9330 Label L_done;
9331
9332 movl(xstart, xlen);
9333 decrementl(xstart);
9334 jcc(Assembler::negative, L_done);
9335
9336 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
9337
9338 Label L_second_loop;
9339 testl(kdx, kdx);
9340 jcc(Assembler::zero, L_second_loop);
9341
9342 Label L_carry;
9343 subl(kdx, 1);
9344 jcc(Assembler::zero, L_carry);
9345
9346 movl(Address(z, kdx, Address::times_4, 0), carry);
9347 shrq(carry, 32);
9348 subl(kdx, 1);
9349
9350 bind(L_carry);
9351 movl(Address(z, kdx, Address::times_4, 0), carry);
9352
9353 // Second and third (nested) loops.
9354 //
9355 // for (int i = xstart-1; i >= 0; i--) { // Second loop
9356 // carry = 0;
9357 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
9358 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
9359 // (z[k] & LONG_MASK) + carry;
9360 // z[k] = (int)product;
9361 // carry = product >>> 32;
9362 // }
9363 // z[i] = (int)carry;
9364 // }
9365 //
9366 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
9367
9368 const Register jdx = tmp1;
9369
9370 bind(L_second_loop);
9371 xorl(carry, carry); // carry = 0;
9372 movl(jdx, ylen); // j = ystart+1
9373
9374 subl(xstart, 1); // i = xstart-1;
9375 jcc(Assembler::negative, L_done);
9376
9377 push (z);
9378
9379 Label L_last_x;
9380 lea(z, Address(z, xstart, Address::times_4, 4)); // z = z + k - j
9381 subl(xstart, 1); // i = xstart-1;
9382 jcc(Assembler::negative, L_last_x);
9383
9384 if (UseBMI2Instructions) {
9385 movq(rdx, Address(x, xstart, Address::times_4, 0));
9386 rorxq(rdx, rdx, 32); // convert big-endian to little-endian
9387 } else {
9388 movq(x_xstart, Address(x, xstart, Address::times_4, 0));
9389 rorq(x_xstart, 32); // convert big-endian to little-endian
9390 }
9391
9392 Label L_third_loop_prologue;
9393 bind(L_third_loop_prologue);
9394
9395 push (x);
9396 push (xstart);
9397 push (ylen);
9398
9399
9400 if (UseBMI2Instructions) {
9401 multiply_128_x_128_bmi2_loop(y, z, carry, x, jdx, ylen, product, tmp2, x_xstart, tmp3, tmp4);
9402 } else { // !UseBMI2Instructions
9403 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
9404 }
9405
9406 pop(ylen);
9407 pop(xlen);
9408 pop(x);
9409 pop(z);
9410
9411 movl(tmp3, xlen);
9412 addl(tmp3, 1);
9413 movl(Address(z, tmp3, Address::times_4, 0), carry);
9414 subl(tmp3, 1);
9415 jccb(Assembler::negative, L_done);
9416
9417 shrq(carry, 32);
9418 movl(Address(z, tmp3, Address::times_4, 0), carry);
9419 jmp(L_second_loop);
9420
9421 // Next infrequent code is moved outside loops.
9422 bind(L_last_x);
9423 if (UseBMI2Instructions) {
9424 movl(rdx, Address(x, 0));
9425 } else {
9426 movl(x_xstart, Address(x, 0));
9427 }
9428 jmp(L_third_loop_prologue);
9429
9430 bind(L_done);
9431
9432 pop(zlen);
9433 pop(xlen);
9434
9435 pop(tmp5);
9436 pop(tmp4);
9437 pop(tmp3);
9438 pop(tmp2);
9439 pop(tmp1);
9440 }
9441
9442 //Helper functions for square_to_len()
9443
9444 /**
9445 * Store the squares of x[], right shifted one bit (divided by 2) into z[]
9446 * Preserves x and z and modifies rest of the registers.
9447 */
9448
9449 void MacroAssembler::square_rshift(Register x, Register xlen, Register z, Register tmp1, Register tmp3, Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
9450 // Perform square and right shift by 1
9451 // Handle odd xlen case first, then for even xlen do the following
9452 // jlong carry = 0;
9453 // for (int j=0, i=0; j < xlen; j+=2, i+=4) {
9454 // huge_128 product = x[j:j+1] * x[j:j+1];
9455 // z[i:i+1] = (carry << 63) | (jlong)(product >>> 65);
9456 // z[i+2:i+3] = (jlong)(product >>> 1);
9457 // carry = (jlong)product;
9458 // }
9459
9460 xorq(tmp5, tmp5); // carry
9461 xorq(rdxReg, rdxReg);
9462 xorl(tmp1, tmp1); // index for x
9463 xorl(tmp4, tmp4); // index for z
9464
9465 Label L_first_loop, L_first_loop_exit;
9466
9467 testl(xlen, 1);
9468 jccb(Assembler::zero, L_first_loop); //jump if xlen is even
9469
9470 // Square and right shift by 1 the odd element using 32 bit multiply
9471 movl(raxReg, Address(x, tmp1, Address::times_4, 0));
9472 imulq(raxReg, raxReg);
9473 shrq(raxReg, 1);
9474 adcq(tmp5, 0);
9475 movq(Address(z, tmp4, Address::times_4, 0), raxReg);
9476 incrementl(tmp1);
9477 addl(tmp4, 2);
9478
9479 // Square and right shift by 1 the rest using 64 bit multiply
9480 bind(L_first_loop);
9481 cmpptr(tmp1, xlen);
9482 jccb(Assembler::equal, L_first_loop_exit);
9483
9484 // Square
9485 movq(raxReg, Address(x, tmp1, Address::times_4, 0));
9486 rorq(raxReg, 32); // convert big-endian to little-endian
9487 mulq(raxReg); // 64-bit multiply rax * rax -> rdx:rax
9488
9489 // Right shift by 1 and save carry
9490 shrq(tmp5, 1); // rdx:rax:tmp5 = (tmp5:rdx:rax) >>> 1
9491 rcrq(rdxReg, 1);
9492 rcrq(raxReg, 1);
9493 adcq(tmp5, 0);
9494
9495 // Store result in z
9496 movq(Address(z, tmp4, Address::times_4, 0), rdxReg);
9497 movq(Address(z, tmp4, Address::times_4, 8), raxReg);
9498
9499 // Update indices for x and z
9500 addl(tmp1, 2);
9501 addl(tmp4, 4);
9502 jmp(L_first_loop);
9503
9504 bind(L_first_loop_exit);
9505 }
9506
9507
9508 /**
9509 * Perform the following multiply add operation using BMI2 instructions
9510 * carry:sum = sum + op1*op2 + carry
9511 * op2 should be in rdx
9512 * op2 is preserved, all other registers are modified
9513 */
9514 void MacroAssembler::multiply_add_64_bmi2(Register sum, Register op1, Register op2, Register carry, Register tmp2) {
9515 // assert op2 is rdx
9516 mulxq(tmp2, op1, op1); // op1 * op2 -> tmp2:op1
9517 addq(sum, carry);
9518 adcq(tmp2, 0);
9519 addq(sum, op1);
9520 adcq(tmp2, 0);
9521 movq(carry, tmp2);
9522 }
9523
9524 /**
9525 * Perform the following multiply add operation:
9526 * carry:sum = sum + op1*op2 + carry
9527 * Preserves op1, op2 and modifies rest of registers
9528 */
9529 void MacroAssembler::multiply_add_64(Register sum, Register op1, Register op2, Register carry, Register rdxReg, Register raxReg) {
9530 // rdx:rax = op1 * op2
9531 movq(raxReg, op2);
9532 mulq(op1);
9533
9534 // rdx:rax = sum + carry + rdx:rax
9535 addq(sum, carry);
9536 adcq(rdxReg, 0);
9537 addq(sum, raxReg);
9538 adcq(rdxReg, 0);
9539
9540 // carry:sum = rdx:sum
9541 movq(carry, rdxReg);
9542 }
9543
9544 /**
9545 * Add 64 bit long carry into z[] with carry propogation.
9546 * Preserves z and carry register values and modifies rest of registers.
9547 *
9548 */
9549 void MacroAssembler::add_one_64(Register z, Register zlen, Register carry, Register tmp1) {
9550 Label L_fourth_loop, L_fourth_loop_exit;
9551
9552 movl(tmp1, 1);
9553 subl(zlen, 2);
9554 addq(Address(z, zlen, Address::times_4, 0), carry);
9555
9556 bind(L_fourth_loop);
9557 jccb(Assembler::carryClear, L_fourth_loop_exit);
9558 subl(zlen, 2);
9559 jccb(Assembler::negative, L_fourth_loop_exit);
9560 addq(Address(z, zlen, Address::times_4, 0), tmp1);
9561 jmp(L_fourth_loop);
9562 bind(L_fourth_loop_exit);
9563 }
9564
9565 /**
9566 * Shift z[] left by 1 bit.
9567 * Preserves x, len, z and zlen registers and modifies rest of the registers.
9568 *
9569 */
9570 void MacroAssembler::lshift_by_1(Register x, Register len, Register z, Register zlen, Register tmp1, Register tmp2, Register tmp3, Register tmp4) {
9571
9572 Label L_fifth_loop, L_fifth_loop_exit;
9573
9574 // Fifth loop
9575 // Perform primitiveLeftShift(z, zlen, 1)
9576
9577 const Register prev_carry = tmp1;
9578 const Register new_carry = tmp4;
9579 const Register value = tmp2;
9580 const Register zidx = tmp3;
9581
9582 // int zidx, carry;
9583 // long value;
9584 // carry = 0;
9585 // for (zidx = zlen-2; zidx >=0; zidx -= 2) {
9586 // (carry:value) = (z[i] << 1) | carry ;
9587 // z[i] = value;
9588 // }
9589
9590 movl(zidx, zlen);
9591 xorl(prev_carry, prev_carry); // clear carry flag and prev_carry register
9592
9593 bind(L_fifth_loop);
9594 decl(zidx); // Use decl to preserve carry flag
9595 decl(zidx);
9596 jccb(Assembler::negative, L_fifth_loop_exit);
9597
9598 if (UseBMI2Instructions) {
9599 movq(value, Address(z, zidx, Address::times_4, 0));
9600 rclq(value, 1);
9601 rorxq(value, value, 32);
9602 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
9603 }
9604 else {
9605 // clear new_carry
9606 xorl(new_carry, new_carry);
9607
9608 // Shift z[i] by 1, or in previous carry and save new carry
9609 movq(value, Address(z, zidx, Address::times_4, 0));
9610 shlq(value, 1);
9611 adcl(new_carry, 0);
9612
9613 orq(value, prev_carry);
9614 rorq(value, 0x20);
9615 movq(Address(z, zidx, Address::times_4, 0), value); // Store back in big endian form
9616
9617 // Set previous carry = new carry
9618 movl(prev_carry, new_carry);
9619 }
9620 jmp(L_fifth_loop);
9621
9622 bind(L_fifth_loop_exit);
9623 }
9624
9625
9626 /**
9627 * Code for BigInteger::squareToLen() intrinsic
9628 *
9629 * rdi: x
9630 * rsi: len
9631 * r8: z
9632 * rcx: zlen
9633 * r12: tmp1
9634 * r13: tmp2
9635 * r14: tmp3
9636 * r15: tmp4
9637 * rbx: tmp5
9638 *
9639 */
9640 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) {
9641
9642 Label L_second_loop, L_second_loop_exit, L_third_loop, L_third_loop_exit, fifth_loop, fifth_loop_exit, L_last_x, L_multiply;
9643 push(tmp1);
9644 push(tmp2);
9645 push(tmp3);
9646 push(tmp4);
9647 push(tmp5);
9648
9649 // First loop
9650 // Store the squares, right shifted one bit (i.e., divided by 2).
9651 square_rshift(x, len, z, tmp1, tmp3, tmp4, tmp5, rdxReg, raxReg);
9652
9653 // Add in off-diagonal sums.
9654 //
9655 // Second, third (nested) and fourth loops.
9656 // zlen +=2;
9657 // for (int xidx=len-2,zidx=zlen-4; xidx > 0; xidx-=2,zidx-=4) {
9658 // carry = 0;
9659 // long op2 = x[xidx:xidx+1];
9660 // for (int j=xidx-2,k=zidx; j >= 0; j-=2) {
9661 // k -= 2;
9662 // long op1 = x[j:j+1];
9663 // long sum = z[k:k+1];
9664 // carry:sum = multiply_add_64(sum, op1, op2, carry, tmp_regs);
9665 // z[k:k+1] = sum;
9666 // }
9667 // add_one_64(z, k, carry, tmp_regs);
9668 // }
9669
9670 const Register carry = tmp5;
9671 const Register sum = tmp3;
9672 const Register op1 = tmp4;
9673 Register op2 = tmp2;
9674
9675 push(zlen);
9676 push(len);
9677 addl(zlen,2);
9678 bind(L_second_loop);
9679 xorq(carry, carry);
9680 subl(zlen, 4);
9681 subl(len, 2);
9682 push(zlen);
9683 push(len);
9684 cmpl(len, 0);
9685 jccb(Assembler::lessEqual, L_second_loop_exit);
9686
9687 // Multiply an array by one 64 bit long.
9688 if (UseBMI2Instructions) {
9689 op2 = rdxReg;
9690 movq(op2, Address(x, len, Address::times_4, 0));
9691 rorxq(op2, op2, 32);
9692 }
9693 else {
9694 movq(op2, Address(x, len, Address::times_4, 0));
9695 rorq(op2, 32);
9696 }
9697
9698 bind(L_third_loop);
9699 decrementl(len);
9700 jccb(Assembler::negative, L_third_loop_exit);
9701 decrementl(len);
9702 jccb(Assembler::negative, L_last_x);
9703
9704 movq(op1, Address(x, len, Address::times_4, 0));
9705 rorq(op1, 32);
9706
9707 bind(L_multiply);
9708 subl(zlen, 2);
9709 movq(sum, Address(z, zlen, Address::times_4, 0));
9710
9711 // Multiply 64 bit by 64 bit and add 64 bits lower half and upper 64 bits as carry.
9712 if (UseBMI2Instructions) {
9713 multiply_add_64_bmi2(sum, op1, op2, carry, tmp2);
9714 }
9715 else {
9716 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9717 }
9718
9719 movq(Address(z, zlen, Address::times_4, 0), sum);
9720
9721 jmp(L_third_loop);
9722 bind(L_third_loop_exit);
9723
9724 // Fourth loop
9725 // Add 64 bit long carry into z with carry propogation.
9726 // Uses offsetted zlen.
9727 add_one_64(z, zlen, carry, tmp1);
9728
9729 pop(len);
9730 pop(zlen);
9731 jmp(L_second_loop);
9732
9733 // Next infrequent code is moved outside loops.
9734 bind(L_last_x);
9735 movl(op1, Address(x, 0));
9736 jmp(L_multiply);
9737
9738 bind(L_second_loop_exit);
9739 pop(len);
9740 pop(zlen);
9741 pop(len);
9742 pop(zlen);
9743
9744 // Fifth loop
9745 // Shift z left 1 bit.
9746 lshift_by_1(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4);
9747
9748 // z[zlen-1] |= x[len-1] & 1;
9749 movl(tmp3, Address(x, len, Address::times_4, -4));
9750 andl(tmp3, 1);
9751 orl(Address(z, zlen, Address::times_4, -4), tmp3);
9752
9753 pop(tmp5);
9754 pop(tmp4);
9755 pop(tmp3);
9756 pop(tmp2);
9757 pop(tmp1);
9758 }
9759
9760 /**
9761 * Helper function for mul_add()
9762 * Multiply the in[] by int k and add to out[] starting at offset offs using
9763 * 128 bit by 32 bit multiply and return the carry in tmp5.
9764 * Only quad int aligned length of in[] is operated on in this function.
9765 * k is in rdxReg for BMI2Instructions, for others it is in tmp2.
9766 * This function preserves out, in and k registers.
9767 * len and offset point to the appropriate index in "in" & "out" correspondingly
9768 * tmp5 has the carry.
9769 * other registers are temporary and are modified.
9770 *
9771 */
9772 void MacroAssembler::mul_add_128_x_32_loop(Register out, Register in,
9773 Register offset, Register len, Register tmp1, Register tmp2, Register tmp3,
9774 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
9775
9776 Label L_first_loop, L_first_loop_exit;
9777
9778 movl(tmp1, len);
9779 shrl(tmp1, 2);
9780
9781 bind(L_first_loop);
9782 subl(tmp1, 1);
9783 jccb(Assembler::negative, L_first_loop_exit);
9784
9785 subl(len, 4);
9786 subl(offset, 4);
9787
9788 Register op2 = tmp2;
9789 const Register sum = tmp3;
9790 const Register op1 = tmp4;
9791 const Register carry = tmp5;
9792
9793 if (UseBMI2Instructions) {
9794 op2 = rdxReg;
9795 }
9796
9797 movq(op1, Address(in, len, Address::times_4, 8));
9798 rorq(op1, 32);
9799 movq(sum, Address(out, offset, Address::times_4, 8));
9800 rorq(sum, 32);
9801 if (UseBMI2Instructions) {
9802 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
9803 }
9804 else {
9805 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9806 }
9807 // Store back in big endian from little endian
9808 rorq(sum, 0x20);
9809 movq(Address(out, offset, Address::times_4, 8), sum);
9810
9811 movq(op1, Address(in, len, Address::times_4, 0));
9812 rorq(op1, 32);
9813 movq(sum, Address(out, offset, Address::times_4, 0));
9814 rorq(sum, 32);
9815 if (UseBMI2Instructions) {
9816 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
9817 }
9818 else {
9819 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9820 }
9821 // Store back in big endian from little endian
9822 rorq(sum, 0x20);
9823 movq(Address(out, offset, Address::times_4, 0), sum);
9824
9825 jmp(L_first_loop);
9826 bind(L_first_loop_exit);
9827 }
9828
9829 /**
9830 * Code for BigInteger::mulAdd() intrinsic
9831 *
9832 * rdi: out
9833 * rsi: in
9834 * r11: offs (out.length - offset)
9835 * rcx: len
9836 * r8: k
9837 * r12: tmp1
9838 * r13: tmp2
9839 * r14: tmp3
9840 * r15: tmp4
9841 * rbx: tmp5
9842 * Multiply the in[] by word k and add to out[], return the carry in rax
9843 */
9844 void MacroAssembler::mul_add(Register out, Register in, Register offs,
9845 Register len, Register k, Register tmp1, Register tmp2, Register tmp3,
9846 Register tmp4, Register tmp5, Register rdxReg, Register raxReg) {
9847
9848 Label L_carry, L_last_in, L_done;
9849
9850 // carry = 0;
9851 // for (int j=len-1; j >= 0; j--) {
9852 // long product = (in[j] & LONG_MASK) * kLong +
9853 // (out[offs] & LONG_MASK) + carry;
9854 // out[offs--] = (int)product;
9855 // carry = product >>> 32;
9856 // }
9857 //
9858 push(tmp1);
9859 push(tmp2);
9860 push(tmp3);
9861 push(tmp4);
9862 push(tmp5);
9863
9864 Register op2 = tmp2;
9865 const Register sum = tmp3;
9866 const Register op1 = tmp4;
9867 const Register carry = tmp5;
9868
9869 if (UseBMI2Instructions) {
9870 op2 = rdxReg;
9871 movl(op2, k);
9872 }
9873 else {
9874 movl(op2, k);
9875 }
9876
9877 xorq(carry, carry);
9878
9879 //First loop
9880
9881 //Multiply in[] by k in a 4 way unrolled loop using 128 bit by 32 bit multiply
9882 //The carry is in tmp5
9883 mul_add_128_x_32_loop(out, in, offs, len, tmp1, tmp2, tmp3, tmp4, tmp5, rdxReg, raxReg);
9884
9885 //Multiply the trailing in[] entry using 64 bit by 32 bit, if any
9886 decrementl(len);
9887 jccb(Assembler::negative, L_carry);
9888 decrementl(len);
9889 jccb(Assembler::negative, L_last_in);
9890
9891 movq(op1, Address(in, len, Address::times_4, 0));
9892 rorq(op1, 32);
9893
9894 subl(offs, 2);
9895 movq(sum, Address(out, offs, Address::times_4, 0));
9896 rorq(sum, 32);
9897
9898 if (UseBMI2Instructions) {
9899 multiply_add_64_bmi2(sum, op1, op2, carry, raxReg);
9900 }
9901 else {
9902 multiply_add_64(sum, op1, op2, carry, rdxReg, raxReg);
9903 }
9904
9905 // Store back in big endian from little endian
9906 rorq(sum, 0x20);
9907 movq(Address(out, offs, Address::times_4, 0), sum);
9908
9909 testl(len, len);
9910 jccb(Assembler::zero, L_carry);
9911
9912 //Multiply the last in[] entry, if any
9913 bind(L_last_in);
9914 movl(op1, Address(in, 0));
9915 movl(sum, Address(out, offs, Address::times_4, -4));
9916
9917 movl(raxReg, k);
9918 mull(op1); //tmp4 * eax -> edx:eax
9919 addl(sum, carry);
9920 adcl(rdxReg, 0);
9921 addl(sum, raxReg);
9922 adcl(rdxReg, 0);
9923 movl(carry, rdxReg);
9924
9925 movl(Address(out, offs, Address::times_4, -4), sum);
9926
9927 bind(L_carry);
9928 //return tmp5/carry as carry in rax
9929 movl(rax, carry);
9930
9931 bind(L_done);
9932 pop(tmp5);
9933 pop(tmp4);
9934 pop(tmp3);
9935 pop(tmp2);
9936 pop(tmp1);
9937 }
9938 #endif
9939
9940 /**
9941 * Emits code to update CRC-32 with a byte value according to constants in table
9942 *
9943 * @param [in,out]crc Register containing the crc.
9944 * @param [in]val Register containing the byte to fold into the CRC.
9945 * @param [in]table Register containing the table of crc constants.
9946 *
9947 * uint32_t crc;
9948 * val = crc_table[(val ^ crc) & 0xFF];
9949 * crc = val ^ (crc >> 8);
9950 *
9951 */
9952 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
9953 xorl(val, crc);
9954 andl(val, 0xFF);
9955 shrl(crc, 8); // unsigned shift
9956 xorl(crc, Address(table, val, Address::times_4, 0));
9957 }
9958
9959 /**
9960 * Fold 128-bit data chunk
9961 */
9962 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) {
9963 if (UseAVX > 0) {
9964 vpclmulhdq(xtmp, xK, xcrc); // [123:64]
9965 vpclmulldq(xcrc, xK, xcrc); // [63:0]
9966 vpxor(xcrc, xcrc, Address(buf, offset), 0 /* vector_len */);
9967 pxor(xcrc, xtmp);
9968 } else {
9969 movdqa(xtmp, xcrc);
9970 pclmulhdq(xtmp, xK); // [123:64]
9971 pclmulldq(xcrc, xK); // [63:0]
9972 pxor(xcrc, xtmp);
9973 movdqu(xtmp, Address(buf, offset));
9974 pxor(xcrc, xtmp);
9975 }
9976 }
9977
9978 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) {
9979 if (UseAVX > 0) {
9980 vpclmulhdq(xtmp, xK, xcrc);
9981 vpclmulldq(xcrc, xK, xcrc);
9982 pxor(xcrc, xbuf);
9983 pxor(xcrc, xtmp);
9984 } else {
9985 movdqa(xtmp, xcrc);
9986 pclmulhdq(xtmp, xK);
9987 pclmulldq(xcrc, xK);
9988 pxor(xcrc, xbuf);
9989 pxor(xcrc, xtmp);
9990 }
9991 }
9992
9993 /**
9994 * 8-bit folds to compute 32-bit CRC
9995 *
9996 * uint64_t xcrc;
9997 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8);
9998 */
9999 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) {
10000 movdl(tmp, xcrc);
10001 andl(tmp, 0xFF);
10002 movdl(xtmp, Address(table, tmp, Address::times_4, 0));
10003 psrldq(xcrc, 1); // unsigned shift one byte
10004 pxor(xcrc, xtmp);
10005 }
10006
10007 /**
10008 * uint32_t crc;
10009 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
10010 */
10011 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
10012 movl(tmp, crc);
10013 andl(tmp, 0xFF);
10014 shrl(crc, 8);
10015 xorl(crc, Address(table, tmp, Address::times_4, 0));
10016 }
10017
10018 /**
10019 * @param crc register containing existing CRC (32-bit)
10020 * @param buf register pointing to input byte buffer (byte*)
10021 * @param len register containing number of bytes
10022 * @param table register that will contain address of CRC table
10023 * @param tmp scratch register
10024 */
10025 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) {
10026 assert_different_registers(crc, buf, len, table, tmp, rax);
10027
10028 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned;
10029 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop;
10030
10031 // For EVEX with VL and BW, provide a standard mask, VL = 128 will guide the merge
10032 // context for the registers used, where all instructions below are using 128-bit mode
10033 // On EVEX without VL and BW, these instructions will all be AVX.
10034 if (VM_Version::supports_avx512vlbw()) {
10035 movl(tmp, 0xffff);
10036 kmovwl(k1, tmp);
10037 }
10038
10039 lea(table, ExternalAddress(StubRoutines::crc_table_addr()));
10040 notl(crc); // ~crc
10041 cmpl(len, 16);
10042 jcc(Assembler::less, L_tail);
10043
10044 // Align buffer to 16 bytes
10045 movl(tmp, buf);
10046 andl(tmp, 0xF);
10047 jccb(Assembler::zero, L_aligned);
10048 subl(tmp, 16);
10049 addl(len, tmp);
10050
10051 align(4);
10052 BIND(L_align_loop);
10053 movsbl(rax, Address(buf, 0)); // load byte with sign extension
10054 update_byte_crc32(crc, rax, table);
10055 increment(buf);
10056 incrementl(tmp);
10057 jccb(Assembler::less, L_align_loop);
10058
10059 BIND(L_aligned);
10060 movl(tmp, len); // save
10061 shrl(len, 4);
10062 jcc(Assembler::zero, L_tail_restore);
10063
10064 // Fold crc into first bytes of vector
10065 movdqa(xmm1, Address(buf, 0));
10066 movdl(rax, xmm1);
10067 xorl(crc, rax);
10068 pinsrd(xmm1, crc, 0);
10069 addptr(buf, 16);
10070 subl(len, 4); // len > 0
10071 jcc(Assembler::less, L_fold_tail);
10072
10073 movdqa(xmm2, Address(buf, 0));
10074 movdqa(xmm3, Address(buf, 16));
10075 movdqa(xmm4, Address(buf, 32));
10076 addptr(buf, 48);
10077 subl(len, 3);
10078 jcc(Assembler::lessEqual, L_fold_512b);
10079
10080 // Fold total 512 bits of polynomial on each iteration,
10081 // 128 bits per each of 4 parallel streams.
10082 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32));
10083
10084 align(32);
10085 BIND(L_fold_512b_loop);
10086 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
10087 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16);
10088 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32);
10089 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48);
10090 addptr(buf, 64);
10091 subl(len, 4);
10092 jcc(Assembler::greater, L_fold_512b_loop);
10093
10094 // Fold 512 bits to 128 bits.
10095 BIND(L_fold_512b);
10096 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
10097 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2);
10098 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3);
10099 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4);
10100
10101 // Fold the rest of 128 bits data chunks
10102 BIND(L_fold_tail);
10103 addl(len, 3);
10104 jccb(Assembler::lessEqual, L_fold_128b);
10105 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16));
10106
10107 BIND(L_fold_tail_loop);
10108 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0);
10109 addptr(buf, 16);
10110 decrementl(len);
10111 jccb(Assembler::greater, L_fold_tail_loop);
10112
10113 // Fold 128 bits in xmm1 down into 32 bits in crc register.
10114 BIND(L_fold_128b);
10115 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr()));
10116 if (UseAVX > 0) {
10117 vpclmulqdq(xmm2, xmm0, xmm1, 0x1);
10118 vpand(xmm3, xmm0, xmm2, 0 /* vector_len */);
10119 vpclmulqdq(xmm0, xmm0, xmm3, 0x1);
10120 } else {
10121 movdqa(xmm2, xmm0);
10122 pclmulqdq(xmm2, xmm1, 0x1);
10123 movdqa(xmm3, xmm0);
10124 pand(xmm3, xmm2);
10125 pclmulqdq(xmm0, xmm3, 0x1);
10126 }
10127 psrldq(xmm1, 8);
10128 psrldq(xmm2, 4);
10129 pxor(xmm0, xmm1);
10130 pxor(xmm0, xmm2);
10131
10132 // 8 8-bit folds to compute 32-bit CRC.
10133 for (int j = 0; j < 4; j++) {
10134 fold_8bit_crc32(xmm0, table, xmm1, rax);
10135 }
10136 movdl(crc, xmm0); // mov 32 bits to general register
10137 for (int j = 0; j < 4; j++) {
10138 fold_8bit_crc32(crc, table, rax);
10139 }
10140
10141 BIND(L_tail_restore);
10142 movl(len, tmp); // restore
10143 BIND(L_tail);
10144 andl(len, 0xf);
10145 jccb(Assembler::zero, L_exit);
10146
10147 // Fold the rest of bytes
10148 align(4);
10149 BIND(L_tail_loop);
10150 movsbl(rax, Address(buf, 0)); // load byte with sign extension
10151 update_byte_crc32(crc, rax, table);
10152 increment(buf);
10153 decrementl(len);
10154 jccb(Assembler::greater, L_tail_loop);
10155
10156 BIND(L_exit);
10157 notl(crc); // ~c
10158 }
10159
10160 #ifdef _LP64
10161 // S. Gueron / Information Processing Letters 112 (2012) 184
10162 // Algorithm 4: Computing carry-less multiplication using a precomputed lookup table.
10163 // Input: A 32 bit value B = [byte3, byte2, byte1, byte0].
10164 // Output: the 64-bit carry-less product of B * CONST
10165 void MacroAssembler::crc32c_ipl_alg4(Register in, uint32_t n,
10166 Register tmp1, Register tmp2, Register tmp3) {
10167 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
10168 if (n > 0) {
10169 addq(tmp3, n * 256 * 8);
10170 }
10171 // Q1 = TABLEExt[n][B & 0xFF];
10172 movl(tmp1, in);
10173 andl(tmp1, 0x000000FF);
10174 shll(tmp1, 3);
10175 addq(tmp1, tmp3);
10176 movq(tmp1, Address(tmp1, 0));
10177
10178 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
10179 movl(tmp2, in);
10180 shrl(tmp2, 8);
10181 andl(tmp2, 0x000000FF);
10182 shll(tmp2, 3);
10183 addq(tmp2, tmp3);
10184 movq(tmp2, Address(tmp2, 0));
10185
10186 shlq(tmp2, 8);
10187 xorq(tmp1, tmp2);
10188
10189 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
10190 movl(tmp2, in);
10191 shrl(tmp2, 16);
10192 andl(tmp2, 0x000000FF);
10193 shll(tmp2, 3);
10194 addq(tmp2, tmp3);
10195 movq(tmp2, Address(tmp2, 0));
10196
10197 shlq(tmp2, 16);
10198 xorq(tmp1, tmp2);
10199
10200 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
10201 shrl(in, 24);
10202 andl(in, 0x000000FF);
10203 shll(in, 3);
10204 addq(in, tmp3);
10205 movq(in, Address(in, 0));
10206
10207 shlq(in, 24);
10208 xorq(in, tmp1);
10209 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
10210 }
10211
10212 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
10213 Register in_out,
10214 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
10215 XMMRegister w_xtmp2,
10216 Register tmp1,
10217 Register n_tmp2, Register n_tmp3) {
10218 if (is_pclmulqdq_supported) {
10219 movdl(w_xtmp1, in_out); // modified blindly
10220
10221 movl(tmp1, const_or_pre_comp_const_index);
10222 movdl(w_xtmp2, tmp1);
10223 pclmulqdq(w_xtmp1, w_xtmp2, 0);
10224
10225 movdq(in_out, w_xtmp1);
10226 } else {
10227 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3);
10228 }
10229 }
10230
10231 // Recombination Alternative 2: No bit-reflections
10232 // T1 = (CRC_A * U1) << 1
10233 // T2 = (CRC_B * U2) << 1
10234 // C1 = T1 >> 32
10235 // C2 = T2 >> 32
10236 // T1 = T1 & 0xFFFFFFFF
10237 // T2 = T2 & 0xFFFFFFFF
10238 // T1 = CRC32(0, T1)
10239 // T2 = CRC32(0, T2)
10240 // C1 = C1 ^ T1
10241 // C2 = C2 ^ T2
10242 // CRC = C1 ^ C2 ^ CRC_C
10243 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,
10244 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10245 Register tmp1, Register tmp2,
10246 Register n_tmp3) {
10247 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10248 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10249 shlq(in_out, 1);
10250 movl(tmp1, in_out);
10251 shrq(in_out, 32);
10252 xorl(tmp2, tmp2);
10253 crc32(tmp2, tmp1, 4);
10254 xorl(in_out, tmp2); // we don't care about upper 32 bit contents here
10255 shlq(in1, 1);
10256 movl(tmp1, in1);
10257 shrq(in1, 32);
10258 xorl(tmp2, tmp2);
10259 crc32(tmp2, tmp1, 4);
10260 xorl(in1, tmp2);
10261 xorl(in_out, in1);
10262 xorl(in_out, in2);
10263 }
10264
10265 // Set N to predefined value
10266 // Subtract from a lenght of a buffer
10267 // execute in a loop:
10268 // CRC_A = 0xFFFFFFFF, CRC_B = 0, CRC_C = 0
10269 // for i = 1 to N do
10270 // CRC_A = CRC32(CRC_A, A[i])
10271 // CRC_B = CRC32(CRC_B, B[i])
10272 // CRC_C = CRC32(CRC_C, C[i])
10273 // end for
10274 // Recombine
10275 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,
10276 Register in_out1, Register in_out2, Register in_out3,
10277 Register tmp1, Register tmp2, Register tmp3,
10278 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10279 Register tmp4, Register tmp5,
10280 Register n_tmp6) {
10281 Label L_processPartitions;
10282 Label L_processPartition;
10283 Label L_exit;
10284
10285 bind(L_processPartitions);
10286 cmpl(in_out1, 3 * size);
10287 jcc(Assembler::less, L_exit);
10288 xorl(tmp1, tmp1);
10289 xorl(tmp2, tmp2);
10290 movq(tmp3, in_out2);
10291 addq(tmp3, size);
10292
10293 bind(L_processPartition);
10294 crc32(in_out3, Address(in_out2, 0), 8);
10295 crc32(tmp1, Address(in_out2, size), 8);
10296 crc32(tmp2, Address(in_out2, size * 2), 8);
10297 addq(in_out2, 8);
10298 cmpq(in_out2, tmp3);
10299 jcc(Assembler::less, L_processPartition);
10300 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
10301 w_xtmp1, w_xtmp2, w_xtmp3,
10302 tmp4, tmp5,
10303 n_tmp6);
10304 addq(in_out2, 2 * size);
10305 subl(in_out1, 3 * size);
10306 jmp(L_processPartitions);
10307
10308 bind(L_exit);
10309 }
10310 #else
10311 void MacroAssembler::crc32c_ipl_alg4(Register in_out, uint32_t n,
10312 Register tmp1, Register tmp2, Register tmp3,
10313 XMMRegister xtmp1, XMMRegister xtmp2) {
10314 lea(tmp3, ExternalAddress(StubRoutines::crc32c_table_addr()));
10315 if (n > 0) {
10316 addl(tmp3, n * 256 * 8);
10317 }
10318 // Q1 = TABLEExt[n][B & 0xFF];
10319 movl(tmp1, in_out);
10320 andl(tmp1, 0x000000FF);
10321 shll(tmp1, 3);
10322 addl(tmp1, tmp3);
10323 movq(xtmp1, Address(tmp1, 0));
10324
10325 // Q2 = TABLEExt[n][B >> 8 & 0xFF];
10326 movl(tmp2, in_out);
10327 shrl(tmp2, 8);
10328 andl(tmp2, 0x000000FF);
10329 shll(tmp2, 3);
10330 addl(tmp2, tmp3);
10331 movq(xtmp2, Address(tmp2, 0));
10332
10333 psllq(xtmp2, 8);
10334 pxor(xtmp1, xtmp2);
10335
10336 // Q3 = TABLEExt[n][B >> 16 & 0xFF];
10337 movl(tmp2, in_out);
10338 shrl(tmp2, 16);
10339 andl(tmp2, 0x000000FF);
10340 shll(tmp2, 3);
10341 addl(tmp2, tmp3);
10342 movq(xtmp2, Address(tmp2, 0));
10343
10344 psllq(xtmp2, 16);
10345 pxor(xtmp1, xtmp2);
10346
10347 // Q4 = TABLEExt[n][B >> 24 & 0xFF];
10348 shrl(in_out, 24);
10349 andl(in_out, 0x000000FF);
10350 shll(in_out, 3);
10351 addl(in_out, tmp3);
10352 movq(xtmp2, Address(in_out, 0));
10353
10354 psllq(xtmp2, 24);
10355 pxor(xtmp1, xtmp2); // Result in CXMM
10356 // return Q1 ^ Q2 << 8 ^ Q3 << 16 ^ Q4 << 24;
10357 }
10358
10359 void MacroAssembler::crc32c_pclmulqdq(XMMRegister w_xtmp1,
10360 Register in_out,
10361 uint32_t const_or_pre_comp_const_index, bool is_pclmulqdq_supported,
10362 XMMRegister w_xtmp2,
10363 Register tmp1,
10364 Register n_tmp2, Register n_tmp3) {
10365 if (is_pclmulqdq_supported) {
10366 movdl(w_xtmp1, in_out);
10367
10368 movl(tmp1, const_or_pre_comp_const_index);
10369 movdl(w_xtmp2, tmp1);
10370 pclmulqdq(w_xtmp1, w_xtmp2, 0);
10371 // Keep result in XMM since GPR is 32 bit in length
10372 } else {
10373 crc32c_ipl_alg4(in_out, const_or_pre_comp_const_index, tmp1, n_tmp2, n_tmp3, w_xtmp1, w_xtmp2);
10374 }
10375 }
10376
10377 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,
10378 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10379 Register tmp1, Register tmp2,
10380 Register n_tmp3) {
10381 crc32c_pclmulqdq(w_xtmp1, in_out, const_or_pre_comp_const_index_u1, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10382 crc32c_pclmulqdq(w_xtmp2, in1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, w_xtmp3, tmp1, tmp2, n_tmp3);
10383
10384 psllq(w_xtmp1, 1);
10385 movdl(tmp1, w_xtmp1);
10386 psrlq(w_xtmp1, 32);
10387 movdl(in_out, w_xtmp1);
10388
10389 xorl(tmp2, tmp2);
10390 crc32(tmp2, tmp1, 4);
10391 xorl(in_out, tmp2);
10392
10393 psllq(w_xtmp2, 1);
10394 movdl(tmp1, w_xtmp2);
10395 psrlq(w_xtmp2, 32);
10396 movdl(in1, w_xtmp2);
10397
10398 xorl(tmp2, tmp2);
10399 crc32(tmp2, tmp1, 4);
10400 xorl(in1, tmp2);
10401 xorl(in_out, in1);
10402 xorl(in_out, in2);
10403 }
10404
10405 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,
10406 Register in_out1, Register in_out2, Register in_out3,
10407 Register tmp1, Register tmp2, Register tmp3,
10408 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10409 Register tmp4, Register tmp5,
10410 Register n_tmp6) {
10411 Label L_processPartitions;
10412 Label L_processPartition;
10413 Label L_exit;
10414
10415 bind(L_processPartitions);
10416 cmpl(in_out1, 3 * size);
10417 jcc(Assembler::less, L_exit);
10418 xorl(tmp1, tmp1);
10419 xorl(tmp2, tmp2);
10420 movl(tmp3, in_out2);
10421 addl(tmp3, size);
10422
10423 bind(L_processPartition);
10424 crc32(in_out3, Address(in_out2, 0), 4);
10425 crc32(tmp1, Address(in_out2, size), 4);
10426 crc32(tmp2, Address(in_out2, size*2), 4);
10427 crc32(in_out3, Address(in_out2, 0+4), 4);
10428 crc32(tmp1, Address(in_out2, size+4), 4);
10429 crc32(tmp2, Address(in_out2, size*2+4), 4);
10430 addl(in_out2, 8);
10431 cmpl(in_out2, tmp3);
10432 jcc(Assembler::less, L_processPartition);
10433
10434 push(tmp3);
10435 push(in_out1);
10436 push(in_out2);
10437 tmp4 = tmp3;
10438 tmp5 = in_out1;
10439 n_tmp6 = in_out2;
10440
10441 crc32c_rec_alt2(const_or_pre_comp_const_index_u1, const_or_pre_comp_const_index_u2, is_pclmulqdq_supported, in_out3, tmp1, tmp2,
10442 w_xtmp1, w_xtmp2, w_xtmp3,
10443 tmp4, tmp5,
10444 n_tmp6);
10445
10446 pop(in_out2);
10447 pop(in_out1);
10448 pop(tmp3);
10449
10450 addl(in_out2, 2 * size);
10451 subl(in_out1, 3 * size);
10452 jmp(L_processPartitions);
10453
10454 bind(L_exit);
10455 }
10456 #endif //LP64
10457
10458 #ifdef _LP64
10459 // Algorithm 2: Pipelined usage of the CRC32 instruction.
10460 // Input: A buffer I of L bytes.
10461 // Output: the CRC32C value of the buffer.
10462 // Notations:
10463 // Write L = 24N + r, with N = floor (L/24).
10464 // r = L mod 24 (0 <= r < 24).
10465 // Consider I as the concatenation of A|B|C|R, where A, B, C, each,
10466 // N quadwords, and R consists of r bytes.
10467 // A[j] = I [8j+7:8j], j= 0, 1, ..., N-1
10468 // B[j] = I [N + 8j+7:N + 8j], j= 0, 1, ..., N-1
10469 // C[j] = I [2N + 8j+7:2N + 8j], j= 0, 1, ..., N-1
10470 // if r > 0 R[j] = I [3N +j], j= 0, 1, ...,r-1
10471 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
10472 Register tmp1, Register tmp2, Register tmp3,
10473 Register tmp4, Register tmp5, Register tmp6,
10474 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10475 bool is_pclmulqdq_supported) {
10476 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
10477 Label L_wordByWord;
10478 Label L_byteByByteProlog;
10479 Label L_byteByByte;
10480 Label L_exit;
10481
10482 if (is_pclmulqdq_supported ) {
10483 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
10484 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr+1);
10485
10486 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
10487 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
10488
10489 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
10490 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
10491 assert((CRC32C_NUM_PRECOMPUTED_CONSTANTS - 1 ) == 5, "Checking whether you declared all of the constants based on the number of \"chunks\"");
10492 } else {
10493 const_or_pre_comp_const_index[0] = 1;
10494 const_or_pre_comp_const_index[1] = 0;
10495
10496 const_or_pre_comp_const_index[2] = 3;
10497 const_or_pre_comp_const_index[3] = 2;
10498
10499 const_or_pre_comp_const_index[4] = 5;
10500 const_or_pre_comp_const_index[5] = 4;
10501 }
10502 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
10503 in2, in1, in_out,
10504 tmp1, tmp2, tmp3,
10505 w_xtmp1, w_xtmp2, w_xtmp3,
10506 tmp4, tmp5,
10507 tmp6);
10508 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
10509 in2, in1, in_out,
10510 tmp1, tmp2, tmp3,
10511 w_xtmp1, w_xtmp2, w_xtmp3,
10512 tmp4, tmp5,
10513 tmp6);
10514 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
10515 in2, in1, in_out,
10516 tmp1, tmp2, tmp3,
10517 w_xtmp1, w_xtmp2, w_xtmp3,
10518 tmp4, tmp5,
10519 tmp6);
10520 movl(tmp1, in2);
10521 andl(tmp1, 0x00000007);
10522 negl(tmp1);
10523 addl(tmp1, in2);
10524 addq(tmp1, in1);
10525
10526 BIND(L_wordByWord);
10527 cmpq(in1, tmp1);
10528 jcc(Assembler::greaterEqual, L_byteByByteProlog);
10529 crc32(in_out, Address(in1, 0), 4);
10530 addq(in1, 4);
10531 jmp(L_wordByWord);
10532
10533 BIND(L_byteByByteProlog);
10534 andl(in2, 0x00000007);
10535 movl(tmp2, 1);
10536
10537 BIND(L_byteByByte);
10538 cmpl(tmp2, in2);
10539 jccb(Assembler::greater, L_exit);
10540 crc32(in_out, Address(in1, 0), 1);
10541 incq(in1);
10542 incl(tmp2);
10543 jmp(L_byteByByte);
10544
10545 BIND(L_exit);
10546 }
10547 #else
10548 void MacroAssembler::crc32c_ipl_alg2_alt2(Register in_out, Register in1, Register in2,
10549 Register tmp1, Register tmp2, Register tmp3,
10550 Register tmp4, Register tmp5, Register tmp6,
10551 XMMRegister w_xtmp1, XMMRegister w_xtmp2, XMMRegister w_xtmp3,
10552 bool is_pclmulqdq_supported) {
10553 uint32_t const_or_pre_comp_const_index[CRC32C_NUM_PRECOMPUTED_CONSTANTS];
10554 Label L_wordByWord;
10555 Label L_byteByByteProlog;
10556 Label L_byteByByte;
10557 Label L_exit;
10558
10559 if (is_pclmulqdq_supported) {
10560 const_or_pre_comp_const_index[1] = *(uint32_t *)StubRoutines::_crc32c_table_addr;
10561 const_or_pre_comp_const_index[0] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 1);
10562
10563 const_or_pre_comp_const_index[3] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 2);
10564 const_or_pre_comp_const_index[2] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 3);
10565
10566 const_or_pre_comp_const_index[5] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 4);
10567 const_or_pre_comp_const_index[4] = *((uint32_t *)StubRoutines::_crc32c_table_addr + 5);
10568 } else {
10569 const_or_pre_comp_const_index[0] = 1;
10570 const_or_pre_comp_const_index[1] = 0;
10571
10572 const_or_pre_comp_const_index[2] = 3;
10573 const_or_pre_comp_const_index[3] = 2;
10574
10575 const_or_pre_comp_const_index[4] = 5;
10576 const_or_pre_comp_const_index[5] = 4;
10577 }
10578 crc32c_proc_chunk(CRC32C_HIGH, const_or_pre_comp_const_index[0], const_or_pre_comp_const_index[1], is_pclmulqdq_supported,
10579 in2, in1, in_out,
10580 tmp1, tmp2, tmp3,
10581 w_xtmp1, w_xtmp2, w_xtmp3,
10582 tmp4, tmp5,
10583 tmp6);
10584 crc32c_proc_chunk(CRC32C_MIDDLE, const_or_pre_comp_const_index[2], const_or_pre_comp_const_index[3], is_pclmulqdq_supported,
10585 in2, in1, in_out,
10586 tmp1, tmp2, tmp3,
10587 w_xtmp1, w_xtmp2, w_xtmp3,
10588 tmp4, tmp5,
10589 tmp6);
10590 crc32c_proc_chunk(CRC32C_LOW, const_or_pre_comp_const_index[4], const_or_pre_comp_const_index[5], is_pclmulqdq_supported,
10591 in2, in1, in_out,
10592 tmp1, tmp2, tmp3,
10593 w_xtmp1, w_xtmp2, w_xtmp3,
10594 tmp4, tmp5,
10595 tmp6);
10596 movl(tmp1, in2);
10597 andl(tmp1, 0x00000007);
10598 negl(tmp1);
10599 addl(tmp1, in2);
10600 addl(tmp1, in1);
10601
10602 BIND(L_wordByWord);
10603 cmpl(in1, tmp1);
10604 jcc(Assembler::greaterEqual, L_byteByByteProlog);
10605 crc32(in_out, Address(in1,0), 4);
10606 addl(in1, 4);
10607 jmp(L_wordByWord);
10608
10609 BIND(L_byteByByteProlog);
10610 andl(in2, 0x00000007);
10611 movl(tmp2, 1);
10612
10613 BIND(L_byteByByte);
10614 cmpl(tmp2, in2);
10615 jccb(Assembler::greater, L_exit);
10616 movb(tmp1, Address(in1, 0));
10617 crc32(in_out, tmp1, 1);
10618 incl(in1);
10619 incl(tmp2);
10620 jmp(L_byteByByte);
10621
10622 BIND(L_exit);
10623 }
10624 #endif // LP64
10625 #undef BIND
10626 #undef BLOCK_COMMENT
10627
10628
10629 // Compress char[] array to byte[].
10630 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
10631 XMMRegister tmp1Reg, XMMRegister tmp2Reg,
10632 XMMRegister tmp3Reg, XMMRegister tmp4Reg,
10633 Register tmp5, Register result) {
10634 Label copy_chars_loop, return_length, return_zero, done;
10635
10636 // rsi: src
10637 // rdi: dst
10638 // rdx: len
10639 // rcx: tmp5
10640 // rax: result
10641
10642 // rsi holds start addr of source char[] to be compressed
10643 // rdi holds start addr of destination byte[]
10644 // rdx holds length
10645
10646 assert(len != result, "");
10647
10648 // save length for return
10649 push(len);
10650
10651 if (UseSSE42Intrinsics) {
10652 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
10653 Label copy_32_loop, copy_16, copy_tail;
10654
10655 movl(result, len);
10656 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vectors
10657
10658 // vectored compression
10659 andl(len, 0xfffffff0); // vector count (in chars)
10660 andl(result, 0x0000000f); // tail count (in chars)
10661 testl(len, len);
10662 jccb(Assembler::zero, copy_16);
10663
10664 // compress 16 chars per iter
10665 movdl(tmp1Reg, tmp5);
10666 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
10667 pxor(tmp4Reg, tmp4Reg);
10668
10669 lea(src, Address(src, len, Address::times_2));
10670 lea(dst, Address(dst, len, Address::times_1));
10671 negptr(len);
10672
10673 bind(copy_32_loop);
10674 movdqu(tmp2Reg, Address(src, len, Address::times_2)); // load 1st 8 characters
10675 por(tmp4Reg, tmp2Reg);
10676 movdqu(tmp3Reg, Address(src, len, Address::times_2, 16)); // load next 8 characters
10677 por(tmp4Reg, tmp3Reg);
10678 ptest(tmp4Reg, tmp1Reg); // check for Unicode chars in next vector
10679 jcc(Assembler::notZero, return_zero);
10680 packuswb(tmp2Reg, tmp3Reg); // only ASCII chars; compress each to 1 byte
10681 movdqu(Address(dst, len, Address::times_1), tmp2Reg);
10682 addptr(len, 16);
10683 jcc(Assembler::notZero, copy_32_loop);
10684
10685 // compress next vector of 8 chars (if any)
10686 bind(copy_16);
10687 movl(len, result);
10688 andl(len, 0xfffffff8); // vector count (in chars)
10689 andl(result, 0x00000007); // tail count (in chars)
10690 testl(len, len);
10691 jccb(Assembler::zero, copy_tail);
10692
10693 movdl(tmp1Reg, tmp5);
10694 pshufd(tmp1Reg, tmp1Reg, 0); // store Unicode mask in tmp1Reg
10695 pxor(tmp3Reg, tmp3Reg);
10696
10697 movdqu(tmp2Reg, Address(src, 0));
10698 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector
10699 jccb(Assembler::notZero, return_zero);
10700 packuswb(tmp2Reg, tmp3Reg); // only LATIN1 chars; compress each to 1 byte
10701 movq(Address(dst, 0), tmp2Reg);
10702 addptr(src, 16);
10703 addptr(dst, 8);
10704
10705 bind(copy_tail);
10706 movl(len, result);
10707 }
10708 // compress 1 char per iter
10709 testl(len, len);
10710 jccb(Assembler::zero, return_length);
10711 lea(src, Address(src, len, Address::times_2));
10712 lea(dst, Address(dst, len, Address::times_1));
10713 negptr(len);
10714
10715 bind(copy_chars_loop);
10716 load_unsigned_short(result, Address(src, len, Address::times_2));
10717 testl(result, 0xff00); // check if Unicode char
10718 jccb(Assembler::notZero, return_zero);
10719 movb(Address(dst, len, Address::times_1), result); // ASCII char; compress to 1 byte
10720 increment(len);
10721 jcc(Assembler::notZero, copy_chars_loop);
10722
10723 // if compression succeeded, return length
10724 bind(return_length);
10725 pop(result);
10726 jmpb(done);
10727
10728 // if compression failed, return 0
10729 bind(return_zero);
10730 xorl(result, result);
10731 addptr(rsp, wordSize);
10732
10733 bind(done);
10734 }
10735
10736 // Inflate byte[] array to char[].
10737 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
10738 XMMRegister tmp1, Register tmp2) {
10739 Label copy_chars_loop, done;
10740
10741 // rsi: src
10742 // rdi: dst
10743 // rdx: len
10744 // rcx: tmp2
10745
10746 // rsi holds start addr of source byte[] to be inflated
10747 // rdi holds start addr of destination char[]
10748 // rdx holds length
10749 assert_different_registers(src, dst, len, tmp2);
10750
10751 if (UseSSE42Intrinsics) {
10752 assert(UseSSE >= 4, "SSE4 must be enabled for SSE4.2 intrinsics to be available");
10753 Label copy_8_loop, copy_bytes, copy_tail;
10754
10755 movl(tmp2, len);
10756 andl(tmp2, 0x00000007); // tail count (in chars)
10757 andl(len, 0xfffffff8); // vector count (in chars)
10758 jccb(Assembler::zero, copy_tail);
10759
10760 // vectored inflation
10761 lea(src, Address(src, len, Address::times_1));
10762 lea(dst, Address(dst, len, Address::times_2));
10763 negptr(len);
10764
10765 // inflate 8 chars per iter
10766 bind(copy_8_loop);
10767 pmovzxbw(tmp1, Address(src, len, Address::times_1)); // unpack to 8 words
10768 movdqu(Address(dst, len, Address::times_2), tmp1);
10769 addptr(len, 8);
10770 jcc(Assembler::notZero, copy_8_loop);
10771
10772 bind(copy_tail);
10773 movl(len, tmp2);
10774
10775 cmpl(len, 4);
10776 jccb(Assembler::less, copy_bytes);
10777
10778 movdl(tmp1, Address(src, 0)); // load 4 byte chars
10779 pmovzxbw(tmp1, tmp1);
10780 movq(Address(dst, 0), tmp1);
10781 subptr(len, 4);
10782 addptr(src, 4);
10783 addptr(dst, 8);
10784
10785 bind(copy_bytes);
10786 }
10787 testl(len, len);
10788 jccb(Assembler::zero, done);
10789 lea(src, Address(src, len, Address::times_1));
10790 lea(dst, Address(dst, len, Address::times_2));
10791 negptr(len);
10792
10793 // inflate 1 char per iter
10794 bind(copy_chars_loop);
10795 load_unsigned_byte(tmp2, Address(src, len, Address::times_1)); // load byte char
10796 movw(Address(dst, len, Address::times_2), tmp2); // inflate byte char to word
10797 increment(len);
10798 jcc(Assembler::notZero, copy_chars_loop);
10799
10800 bind(done);
10801 }
10802
10803
10804 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
10805 switch (cond) {
10806 // Note some conditions are synonyms for others
10807 case Assembler::zero: return Assembler::notZero;
10808 case Assembler::notZero: return Assembler::zero;
10809 case Assembler::less: return Assembler::greaterEqual;
10810 case Assembler::lessEqual: return Assembler::greater;
10811 case Assembler::greater: return Assembler::lessEqual;
10812 case Assembler::greaterEqual: return Assembler::less;
10813 case Assembler::below: return Assembler::aboveEqual;
10814 case Assembler::belowEqual: return Assembler::above;
10815 case Assembler::above: return Assembler::belowEqual;
10816 case Assembler::aboveEqual: return Assembler::below;
10817 case Assembler::overflow: return Assembler::noOverflow;
10818 case Assembler::noOverflow: return Assembler::overflow;
10819 case Assembler::negative: return Assembler::positive;
10820 case Assembler::positive: return Assembler::negative;
10821 case Assembler::parity: return Assembler::noParity;
10822 case Assembler::noParity: return Assembler::parity;
10823 }
10824 ShouldNotReachHere(); return Assembler::overflow;
10825 }
10826
10827 SkipIfEqual::SkipIfEqual(
10828 MacroAssembler* masm, const bool* flag_addr, bool value) {
10829 _masm = masm;
10830 _masm->cmp8(ExternalAddress((address)flag_addr), value);
10831 _masm->jcc(Assembler::equal, _label);
10832 }
10833
10834 SkipIfEqual::~SkipIfEqual() {
10835 _masm->bind(_label);
10836 }