1 /*
2 * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "jvm.h"
27 #include "asm/macroAssembler.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.inline.hpp"
40 #include "runtime/safepoint.hpp"
41 #include "runtime/safepointMechanism.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "runtime/stubRoutines.hpp"
44 #include "utilities/align.hpp"
45 #include "utilities/macros.hpp"
46 #if INCLUDE_ALL_GCS
47 #include "gc/g1/g1CollectedHeap.inline.hpp"
48 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
49 #include "gc/g1/heapRegion.hpp"
50 #endif // INCLUDE_ALL_GCS
51 #ifdef COMPILER2
52 #include "opto/intrinsicnode.hpp"
53 #endif
54
55 #ifdef PRODUCT
56 #define BLOCK_COMMENT(str) /* nothing */
57 #define STOP(error) stop(error)
58 #else
59 #define BLOCK_COMMENT(str) block_comment(str)
60 #define STOP(error) block_comment(error); stop(error)
61 #endif
62
63 // Convert the raw encoding form into the form expected by the
64 // constructor for Address.
65 Address Address::make_raw(int base, int index, int scale, int disp, relocInfo::relocType disp_reloc) {
66 assert(scale == 0, "not supported");
67 RelocationHolder rspec;
68 if (disp_reloc != relocInfo::none) {
69 rspec = Relocation::spec_simple(disp_reloc);
70 }
71
72 Register rindex = as_Register(index);
73 if (rindex != G0) {
74 Address madr(as_Register(base), rindex);
75 madr._rspec = rspec;
76 return madr;
77 } else {
78 Address madr(as_Register(base), disp);
79 madr._rspec = rspec;
80 return madr;
81 }
82 }
83
84 Address Argument::address_in_frame() const {
85 // Warning: In LP64 mode disp will occupy more than 10 bits, but
86 // op codes such as ld or ldx, only access disp() to get
87 // their simm13 argument.
88 int disp = ((_number - Argument::n_register_parameters + frame::memory_parameter_word_sp_offset) * BytesPerWord) + STACK_BIAS;
89 if (is_in())
90 return Address(FP, disp); // In argument.
91 else
92 return Address(SP, disp); // Out argument.
93 }
94
95 static const char* argumentNames[][2] = {
96 {"A0","P0"}, {"A1","P1"}, {"A2","P2"}, {"A3","P3"}, {"A4","P4"},
97 {"A5","P5"}, {"A6","P6"}, {"A7","P7"}, {"A8","P8"}, {"A9","P9"},
98 {"A(n>9)","P(n>9)"}
99 };
100
101 const char* Argument::name() const {
102 int nofArgs = sizeof argumentNames / sizeof argumentNames[0];
103 int num = number();
104 if (num >= nofArgs) num = nofArgs - 1;
105 return argumentNames[num][is_in() ? 1 : 0];
106 }
107
108 #ifdef ASSERT
109 // On RISC, there's no benefit to verifying instruction boundaries.
110 bool AbstractAssembler::pd_check_instruction_mark() { return false; }
111 #endif
112
113 // Patch instruction inst at offset inst_pos to refer to dest_pos
114 // and return the resulting instruction.
115 // We should have pcs, not offsets, but since all is relative, it will work out
116 // OK.
117 int MacroAssembler::patched_branch(int dest_pos, int inst, int inst_pos) {
118 int m; // mask for displacement field
119 int v; // new value for displacement field
120 const int word_aligned_ones = -4;
121 switch (inv_op(inst)) {
122 default: ShouldNotReachHere();
123 case call_op: m = wdisp(word_aligned_ones, 0, 30); v = wdisp(dest_pos, inst_pos, 30); break;
124 case branch_op:
125 switch (inv_op2(inst)) {
126 case fbp_op2: m = wdisp( word_aligned_ones, 0, 19); v = wdisp( dest_pos, inst_pos, 19); break;
127 case bp_op2: m = wdisp( word_aligned_ones, 0, 19); v = wdisp( dest_pos, inst_pos, 19); break;
128 case fb_op2: m = wdisp( word_aligned_ones, 0, 22); v = wdisp( dest_pos, inst_pos, 22); break;
129 case br_op2: m = wdisp( word_aligned_ones, 0, 22); v = wdisp( dest_pos, inst_pos, 22); break;
130 case bpr_op2: {
131 if (is_cbcond(inst)) {
132 m = wdisp10(word_aligned_ones, 0);
133 v = wdisp10(dest_pos, inst_pos);
134 } else {
135 m = wdisp16(word_aligned_ones, 0);
136 v = wdisp16(dest_pos, inst_pos);
137 }
138 break;
139 }
140 default: ShouldNotReachHere();
141 }
142 }
143 return inst & ~m | v;
144 }
145
146 // Return the offset of the branch destionation of instruction inst
147 // at offset pos.
148 // Should have pcs, but since all is relative, it works out.
149 int MacroAssembler::branch_destination(int inst, int pos) {
150 int r;
151 switch (inv_op(inst)) {
152 default: ShouldNotReachHere();
153 case call_op: r = inv_wdisp(inst, pos, 30); break;
154 case branch_op:
155 switch (inv_op2(inst)) {
156 case fbp_op2: r = inv_wdisp( inst, pos, 19); break;
157 case bp_op2: r = inv_wdisp( inst, pos, 19); break;
158 case fb_op2: r = inv_wdisp( inst, pos, 22); break;
159 case br_op2: r = inv_wdisp( inst, pos, 22); break;
160 case bpr_op2: {
161 if (is_cbcond(inst)) {
162 r = inv_wdisp10(inst, pos);
163 } else {
164 r = inv_wdisp16(inst, pos);
165 }
166 break;
167 }
168 default: ShouldNotReachHere();
169 }
170 }
171 return r;
172 }
173
174 void MacroAssembler::null_check(Register reg, int offset) {
175 if (needs_explicit_null_check((intptr_t)offset)) {
176 // provoke OS NULL exception if reg = NULL by
177 // accessing M[reg] w/o changing any registers
178 ld_ptr(reg, 0, G0);
179 }
180 else {
181 // nothing to do, (later) access of M[reg + offset]
182 // will provoke OS NULL exception if reg = NULL
183 }
184 }
185
186 // Ring buffer jumps
187
188
189 void MacroAssembler::jmp2(Register r1, Register r2, const char* file, int line ) {
190 assert_not_delayed();
191 jmpl(r1, r2, G0);
192 }
193 void MacroAssembler::jmp(Register r1, int offset, const char* file, int line ) {
194 assert_not_delayed();
195 jmp(r1, offset);
196 }
197
198 // This code sequence is relocatable to any address, even on LP64.
199 void MacroAssembler::jumpl(const AddressLiteral& addrlit, Register temp, Register d, int offset, const char* file, int line) {
200 assert_not_delayed();
201 // Force fixed length sethi because NativeJump and NativeFarCall don't handle
202 // variable length instruction streams.
203 patchable_sethi(addrlit, temp);
204 Address a(temp, addrlit.low10() + offset); // Add the offset to the displacement.
205 jmpl(a.base(), a.disp(), d);
206 }
207
208 void MacroAssembler::jump(const AddressLiteral& addrlit, Register temp, int offset, const char* file, int line) {
209 jumpl(addrlit, temp, G0, offset, file, line);
210 }
211
212
213 // Conditional breakpoint (for assertion checks in assembly code)
214 void MacroAssembler::breakpoint_trap(Condition c, CC cc) {
215 trap(c, cc, G0, ST_RESERVED_FOR_USER_0);
216 }
217
218 // We want to use ST_BREAKPOINT here, but the debugger is confused by it.
219 void MacroAssembler::breakpoint_trap() {
220 trap(ST_RESERVED_FOR_USER_0);
221 }
222
223 // Write serialization page so VM thread can do a pseudo remote membar
224 // We use the current thread pointer to calculate a thread specific
225 // offset to write to within the page. This minimizes bus traffic
226 // due to cache line collision.
227 void MacroAssembler::serialize_memory(Register thread, Register tmp1, Register tmp2) {
228 srl(thread, os::get_serialize_page_shift_count(), tmp2);
229 if (Assembler::is_simm13(os::vm_page_size())) {
230 and3(tmp2, (os::vm_page_size() - sizeof(int)), tmp2);
231 }
232 else {
233 set((os::vm_page_size() - sizeof(int)), tmp1);
234 and3(tmp2, tmp1, tmp2);
235 }
236 set(os::get_memory_serialize_page(), tmp1);
237 st(G0, tmp1, tmp2);
238 }
239
240
241 void MacroAssembler::safepoint_poll(Label& slow_path, bool a, Register thread_reg, Register temp_reg) {
242 if (SafepointMechanism::uses_thread_local_poll()) {
243 ldx(Address(thread_reg, Thread::polling_page_offset()), temp_reg, 0);
244 // Armed page has poll bit set.
245 and3(temp_reg, SafepointMechanism::poll_bit(), temp_reg);
246 br_notnull(temp_reg, a, Assembler::pn, slow_path);
247 } else {
248 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
249
250 load_contents(sync_state, temp_reg);
251 cmp(temp_reg, SafepointSynchronize::_not_synchronized);
252 br(Assembler::notEqual, a, Assembler::pn, slow_path);
253 }
254 }
255
256 void MacroAssembler::enter() {
257 Unimplemented();
258 }
259
260 void MacroAssembler::leave() {
261 Unimplemented();
262 }
263
264 // Calls to C land
265
266 #ifdef ASSERT
267 // a hook for debugging
268 static Thread* reinitialize_thread() {
269 return Thread::current();
270 }
271 #else
272 #define reinitialize_thread Thread::current
273 #endif
274
275 #ifdef ASSERT
276 address last_get_thread = NULL;
277 #endif
278
279 // call this when G2_thread is not known to be valid
280 void MacroAssembler::get_thread() {
281 save_frame(0); // to avoid clobbering O0
282 mov(G1, L0); // avoid clobbering G1
283 mov(G5_method, L1); // avoid clobbering G5
284 mov(G3, L2); // avoid clobbering G3 also
285 mov(G4, L5); // avoid clobbering G4
286 #ifdef ASSERT
287 AddressLiteral last_get_thread_addrlit(&last_get_thread);
288 set(last_get_thread_addrlit, L3);
289 rdpc(L4);
290 inc(L4, 3 * BytesPerInstWord); // skip rdpc + inc + st_ptr to point L4 at call st_ptr(L4, L3, 0);
291 #endif
292 call(CAST_FROM_FN_PTR(address, reinitialize_thread), relocInfo::runtime_call_type);
293 delayed()->nop();
294 mov(L0, G1);
295 mov(L1, G5_method);
296 mov(L2, G3);
297 mov(L5, G4);
298 restore(O0, 0, G2_thread);
299 }
300
301 static Thread* verify_thread_subroutine(Thread* gthread_value) {
302 Thread* correct_value = Thread::current();
303 guarantee(gthread_value == correct_value, "G2_thread value must be the thread");
304 return correct_value;
305 }
306
307 void MacroAssembler::verify_thread() {
308 if (VerifyThread) {
309 // NOTE: this chops off the heads of the 64-bit O registers.
310 // make sure G2_thread contains the right value
311 save_frame_and_mov(0, Lmethod, Lmethod); // to avoid clobbering O0 (and propagate Lmethod)
312 mov(G1, L1); // avoid clobbering G1
313 // G2 saved below
314 mov(G3, L3); // avoid clobbering G3
315 mov(G4, L4); // avoid clobbering G4
316 mov(G5_method, L5); // avoid clobbering G5_method
317 call(CAST_FROM_FN_PTR(address,verify_thread_subroutine), relocInfo::runtime_call_type);
318 delayed()->mov(G2_thread, O0);
319
320 mov(L1, G1); // Restore G1
321 // G2 restored below
322 mov(L3, G3); // restore G3
323 mov(L4, G4); // restore G4
324 mov(L5, G5_method); // restore G5_method
325 restore(O0, 0, G2_thread);
326 }
327 }
328
329
330 void MacroAssembler::save_thread(const Register thread_cache) {
331 verify_thread();
332 if (thread_cache->is_valid()) {
333 assert(thread_cache->is_local() || thread_cache->is_in(), "bad volatile");
334 mov(G2_thread, thread_cache);
335 }
336 if (VerifyThread) {
337 // smash G2_thread, as if the VM were about to anyway
338 set(0x67676767, G2_thread);
339 }
340 }
341
342
343 void MacroAssembler::restore_thread(const Register thread_cache) {
344 if (thread_cache->is_valid()) {
345 assert(thread_cache->is_local() || thread_cache->is_in(), "bad volatile");
346 mov(thread_cache, G2_thread);
347 verify_thread();
348 } else {
349 // do it the slow way
350 get_thread();
351 }
352 }
353
354
355 // %%% maybe get rid of [re]set_last_Java_frame
356 void MacroAssembler::set_last_Java_frame(Register last_java_sp, Register last_Java_pc) {
357 assert_not_delayed();
358 Address flags(G2_thread, JavaThread::frame_anchor_offset() +
359 JavaFrameAnchor::flags_offset());
360 Address pc_addr(G2_thread, JavaThread::last_Java_pc_offset());
361
362 // Always set last_Java_pc and flags first because once last_Java_sp is visible
363 // has_last_Java_frame is true and users will look at the rest of the fields.
364 // (Note: flags should always be zero before we get here so doesn't need to be set.)
365
366 #ifdef ASSERT
367 // Verify that flags was zeroed on return to Java
368 Label PcOk;
369 save_frame(0); // to avoid clobbering O0
370 ld_ptr(pc_addr, L0);
371 br_null_short(L0, Assembler::pt, PcOk);
372 STOP("last_Java_pc not zeroed before leaving Java");
373 bind(PcOk);
374
375 // Verify that flags was zeroed on return to Java
376 Label FlagsOk;
377 ld(flags, L0);
378 tst(L0);
379 br(Assembler::zero, false, Assembler::pt, FlagsOk);
380 delayed() -> restore();
381 STOP("flags not zeroed before leaving Java");
382 bind(FlagsOk);
383 #endif /* ASSERT */
384 //
385 // When returning from calling out from Java mode the frame anchor's last_Java_pc
386 // will always be set to NULL. It is set here so that if we are doing a call to
387 // native (not VM) that we capture the known pc and don't have to rely on the
388 // native call having a standard frame linkage where we can find the pc.
389
390 if (last_Java_pc->is_valid()) {
391 st_ptr(last_Java_pc, pc_addr);
392 }
393
394 #ifdef ASSERT
395 // Make sure that we have an odd stack
396 Label StackOk;
397 andcc(last_java_sp, 0x01, G0);
398 br(Assembler::notZero, false, Assembler::pt, StackOk);
399 delayed()->nop();
400 STOP("Stack Not Biased in set_last_Java_frame");
401 bind(StackOk);
402 #endif // ASSERT
403 assert( last_java_sp != G4_scratch, "bad register usage in set_last_Java_frame");
404 add( last_java_sp, STACK_BIAS, G4_scratch );
405 st_ptr(G4_scratch, G2_thread, JavaThread::last_Java_sp_offset());
406 }
407
408 void MacroAssembler::reset_last_Java_frame(void) {
409 assert_not_delayed();
410
411 Address sp_addr(G2_thread, JavaThread::last_Java_sp_offset());
412 Address pc_addr(G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset());
413 Address flags (G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
414
415 #ifdef ASSERT
416 // check that it WAS previously set
417 save_frame_and_mov(0, Lmethod, Lmethod); // Propagate Lmethod to helper frame
418 ld_ptr(sp_addr, L0);
419 tst(L0);
420 breakpoint_trap(Assembler::zero, Assembler::ptr_cc);
421 restore();
422 #endif // ASSERT
423
424 st_ptr(G0, sp_addr);
425 // Always return last_Java_pc to zero
426 st_ptr(G0, pc_addr);
427 // Always null flags after return to Java
428 st(G0, flags);
429 }
430
431
432 void MacroAssembler::call_VM_base(
433 Register oop_result,
434 Register thread_cache,
435 Register last_java_sp,
436 address entry_point,
437 int number_of_arguments,
438 bool check_exceptions)
439 {
440 assert_not_delayed();
441
442 // determine last_java_sp register
443 if (!last_java_sp->is_valid()) {
444 last_java_sp = SP;
445 }
446 // debugging support
447 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
448
449 // 64-bit last_java_sp is biased!
450 set_last_Java_frame(last_java_sp, noreg);
451 if (VerifyThread) mov(G2_thread, O0); // about to be smashed; pass early
452 save_thread(thread_cache);
453 // do the call
454 call(entry_point, relocInfo::runtime_call_type);
455 if (!VerifyThread)
456 delayed()->mov(G2_thread, O0); // pass thread as first argument
457 else
458 delayed()->nop(); // (thread already passed)
459 restore_thread(thread_cache);
460 reset_last_Java_frame();
461
462 // check for pending exceptions. use Gtemp as scratch register.
463 if (check_exceptions) {
464 check_and_forward_exception(Gtemp);
465 }
466
467 #ifdef ASSERT
468 set(badHeapWordVal, G3);
469 set(badHeapWordVal, G4);
470 set(badHeapWordVal, G5);
471 #endif
472
473 // get oop result if there is one and reset the value in the thread
474 if (oop_result->is_valid()) {
475 get_vm_result(oop_result);
476 }
477 }
478
479 void MacroAssembler::check_and_forward_exception(Register scratch_reg)
480 {
481 Label L;
482
483 check_and_handle_popframe(scratch_reg);
484 check_and_handle_earlyret(scratch_reg);
485
486 Address exception_addr(G2_thread, Thread::pending_exception_offset());
487 ld_ptr(exception_addr, scratch_reg);
488 br_null_short(scratch_reg, pt, L);
489 // we use O7 linkage so that forward_exception_entry has the issuing PC
490 call(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
491 delayed()->nop();
492 bind(L);
493 }
494
495
496 void MacroAssembler::check_and_handle_popframe(Register scratch_reg) {
497 }
498
499
500 void MacroAssembler::check_and_handle_earlyret(Register scratch_reg) {
501 }
502
503
504 void MacroAssembler::call_VM(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
505 call_VM_base(oop_result, noreg, noreg, entry_point, number_of_arguments, check_exceptions);
506 }
507
508
509 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, bool check_exceptions) {
510 // O0 is reserved for the thread
511 mov(arg_1, O1);
512 call_VM(oop_result, entry_point, 1, check_exceptions);
513 }
514
515
516 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) {
517 // O0 is reserved for the thread
518 mov(arg_1, O1);
519 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument");
520 call_VM(oop_result, entry_point, 2, check_exceptions);
521 }
522
523
524 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, Register arg_3, bool check_exceptions) {
525 // O0 is reserved for the thread
526 mov(arg_1, O1);
527 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument");
528 mov(arg_3, O3); assert(arg_3 != O1 && arg_3 != O2, "smashed argument");
529 call_VM(oop_result, entry_point, 3, check_exceptions);
530 }
531
532
533
534 // Note: The following call_VM overloadings are useful when a "save"
535 // has already been performed by a stub, and the last Java frame is
536 // the previous one. In that case, last_java_sp must be passed as FP
537 // instead of SP.
538
539
540 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, int number_of_arguments, bool check_exceptions) {
541 call_VM_base(oop_result, noreg, last_java_sp, entry_point, number_of_arguments, check_exceptions);
542 }
543
544
545 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, bool check_exceptions) {
546 // O0 is reserved for the thread
547 mov(arg_1, O1);
548 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
549 }
550
551
552 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) {
553 // O0 is reserved for the thread
554 mov(arg_1, O1);
555 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument");
556 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
557 }
558
559
560 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, Register arg_2, Register arg_3, bool check_exceptions) {
561 // O0 is reserved for the thread
562 mov(arg_1, O1);
563 mov(arg_2, O2); assert(arg_2 != O1, "smashed argument");
564 mov(arg_3, O3); assert(arg_3 != O1 && arg_3 != O2, "smashed argument");
565 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
566 }
567
568
569
570 void MacroAssembler::call_VM_leaf_base(Register thread_cache, address entry_point, int number_of_arguments) {
571 assert_not_delayed();
572 save_thread(thread_cache);
573 // do the call
574 call(entry_point, relocInfo::runtime_call_type);
575 delayed()->nop();
576 restore_thread(thread_cache);
577 #ifdef ASSERT
578 set(badHeapWordVal, G3);
579 set(badHeapWordVal, G4);
580 set(badHeapWordVal, G5);
581 #endif
582 }
583
584
585 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, int number_of_arguments) {
586 call_VM_leaf_base(thread_cache, entry_point, number_of_arguments);
587 }
588
589
590 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1) {
591 mov(arg_1, O0);
592 call_VM_leaf(thread_cache, entry_point, 1);
593 }
594
595
596 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1, Register arg_2) {
597 mov(arg_1, O0);
598 mov(arg_2, O1); assert(arg_2 != O0, "smashed argument");
599 call_VM_leaf(thread_cache, entry_point, 2);
600 }
601
602
603 void MacroAssembler::call_VM_leaf(Register thread_cache, address entry_point, Register arg_1, Register arg_2, Register arg_3) {
604 mov(arg_1, O0);
605 mov(arg_2, O1); assert(arg_2 != O0, "smashed argument");
606 mov(arg_3, O2); assert(arg_3 != O0 && arg_3 != O1, "smashed argument");
607 call_VM_leaf(thread_cache, entry_point, 3);
608 }
609
610
611 void MacroAssembler::get_vm_result(Register oop_result) {
612 verify_thread();
613 Address vm_result_addr(G2_thread, JavaThread::vm_result_offset());
614 ld_ptr( vm_result_addr, oop_result);
615 st_ptr(G0, vm_result_addr);
616 verify_oop(oop_result);
617 }
618
619
620 void MacroAssembler::get_vm_result_2(Register metadata_result) {
621 verify_thread();
622 Address vm_result_addr_2(G2_thread, JavaThread::vm_result_2_offset());
623 ld_ptr(vm_result_addr_2, metadata_result);
624 st_ptr(G0, vm_result_addr_2);
625 }
626
627
628 // We require that C code which does not return a value in vm_result will
629 // leave it undisturbed.
630 void MacroAssembler::set_vm_result(Register oop_result) {
631 verify_thread();
632 Address vm_result_addr(G2_thread, JavaThread::vm_result_offset());
633 verify_oop(oop_result);
634
635 # ifdef ASSERT
636 // Check that we are not overwriting any other oop.
637 save_frame_and_mov(0, Lmethod, Lmethod); // Propagate Lmethod
638 ld_ptr(vm_result_addr, L0);
639 tst(L0);
640 restore();
641 breakpoint_trap(notZero, Assembler::ptr_cc);
642 // }
643 # endif
644
645 st_ptr(oop_result, vm_result_addr);
646 }
647
648
649 void MacroAssembler::ic_call(address entry, bool emit_delay, jint method_index) {
650 RelocationHolder rspec = virtual_call_Relocation::spec(pc(), method_index);
651 patchable_set((intptr_t)Universe::non_oop_word(), G5_inline_cache_reg);
652 relocate(rspec);
653 call(entry, relocInfo::none);
654 if (emit_delay) {
655 delayed()->nop();
656 }
657 }
658
659 void MacroAssembler::card_table_write(jbyte* byte_map_base,
660 Register tmp, Register obj) {
661 srlx(obj, CardTableModRefBS::card_shift, obj);
662 assert(tmp != obj, "need separate temp reg");
663 set((address) byte_map_base, tmp);
664 stb(G0, tmp, obj);
665 }
666
667
668 void MacroAssembler::internal_sethi(const AddressLiteral& addrlit, Register d, bool ForceRelocatable) {
669 address save_pc;
670 int shiftcnt;
671 #ifdef VALIDATE_PIPELINE
672 assert_no_delay("Cannot put two instructions in delay-slot.");
673 #endif
674 v9_dep();
675 save_pc = pc();
676
677 int msb32 = (int) (addrlit.value() >> 32);
678 int lsb32 = (int) (addrlit.value());
679
680 if (msb32 == 0 && lsb32 >= 0) {
681 Assembler::sethi(lsb32, d, addrlit.rspec());
682 }
683 else if (msb32 == -1) {
684 Assembler::sethi(~lsb32, d, addrlit.rspec());
685 xor3(d, ~low10(~0), d);
686 }
687 else {
688 Assembler::sethi(msb32, d, addrlit.rspec()); // msb 22-bits
689 if (msb32 & 0x3ff) // Any bits?
690 or3(d, msb32 & 0x3ff, d); // msb 32-bits are now in lsb 32
691 if (lsb32 & 0xFFFFFC00) { // done?
692 if ((lsb32 >> 20) & 0xfff) { // Any bits set?
693 sllx(d, 12, d); // Make room for next 12 bits
694 or3(d, (lsb32 >> 20) & 0xfff, d); // Or in next 12
695 shiftcnt = 0; // We already shifted
696 }
697 else
698 shiftcnt = 12;
699 if ((lsb32 >> 10) & 0x3ff) {
700 sllx(d, shiftcnt + 10, d); // Make room for last 10 bits
701 or3(d, (lsb32 >> 10) & 0x3ff, d); // Or in next 10
702 shiftcnt = 0;
703 }
704 else
705 shiftcnt = 10;
706 sllx(d, shiftcnt + 10, d); // Shift leaving disp field 0'd
707 }
708 else
709 sllx(d, 32, d);
710 }
711 // Pad out the instruction sequence so it can be patched later.
712 if (ForceRelocatable || (addrlit.rtype() != relocInfo::none &&
713 addrlit.rtype() != relocInfo::runtime_call_type)) {
714 while (pc() < (save_pc + (7 * BytesPerInstWord)))
715 nop();
716 }
717 }
718
719
720 void MacroAssembler::sethi(const AddressLiteral& addrlit, Register d) {
721 internal_sethi(addrlit, d, false);
722 }
723
724
725 void MacroAssembler::patchable_sethi(const AddressLiteral& addrlit, Register d) {
726 internal_sethi(addrlit, d, true);
727 }
728
729
730 int MacroAssembler::insts_for_sethi(address a, bool worst_case) {
731 if (worst_case) return 7;
732 intptr_t iaddr = (intptr_t) a;
733 int msb32 = (int) (iaddr >> 32);
734 int lsb32 = (int) (iaddr);
735 int count;
736 if (msb32 == 0 && lsb32 >= 0)
737 count = 1;
738 else if (msb32 == -1)
739 count = 2;
740 else {
741 count = 2;
742 if (msb32 & 0x3ff)
743 count++;
744 if (lsb32 & 0xFFFFFC00 ) {
745 if ((lsb32 >> 20) & 0xfff) count += 2;
746 if ((lsb32 >> 10) & 0x3ff) count += 2;
747 }
748 }
749 return count;
750 }
751
752 int MacroAssembler::worst_case_insts_for_set() {
753 return insts_for_sethi(NULL, true) + 1;
754 }
755
756
757 // Keep in sync with MacroAssembler::insts_for_internal_set
758 void MacroAssembler::internal_set(const AddressLiteral& addrlit, Register d, bool ForceRelocatable) {
759 intptr_t value = addrlit.value();
760
761 if (!ForceRelocatable && addrlit.rspec().type() == relocInfo::none) {
762 // can optimize
763 if (-4096 <= value && value <= 4095) {
764 or3(G0, value, d); // setsw (this leaves upper 32 bits sign-extended)
765 return;
766 }
767 if (inv_hi22(hi22(value)) == value) {
768 sethi(addrlit, d);
769 return;
770 }
771 }
772 assert_no_delay("Cannot put two instructions in delay-slot.");
773 internal_sethi(addrlit, d, ForceRelocatable);
774 if (ForceRelocatable || addrlit.rspec().type() != relocInfo::none || addrlit.low10() != 0) {
775 add(d, addrlit.low10(), d, addrlit.rspec());
776 }
777 }
778
779 // Keep in sync with MacroAssembler::internal_set
780 int MacroAssembler::insts_for_internal_set(intptr_t value) {
781 // can optimize
782 if (-4096 <= value && value <= 4095) {
783 return 1;
784 }
785 if (inv_hi22(hi22(value)) == value) {
786 return insts_for_sethi((address) value);
787 }
788 int count = insts_for_sethi((address) value);
789 AddressLiteral al(value);
790 if (al.low10() != 0) {
791 count++;
792 }
793 return count;
794 }
795
796 void MacroAssembler::set(const AddressLiteral& al, Register d) {
797 internal_set(al, d, false);
798 }
799
800 void MacroAssembler::set(intptr_t value, Register d) {
801 AddressLiteral al(value);
802 internal_set(al, d, false);
803 }
804
805 void MacroAssembler::set(address addr, Register d, RelocationHolder const& rspec) {
806 AddressLiteral al(addr, rspec);
807 internal_set(al, d, false);
808 }
809
810 void MacroAssembler::patchable_set(const AddressLiteral& al, Register d) {
811 internal_set(al, d, true);
812 }
813
814 void MacroAssembler::patchable_set(intptr_t value, Register d) {
815 AddressLiteral al(value);
816 internal_set(al, d, true);
817 }
818
819
820 void MacroAssembler::set64(jlong value, Register d, Register tmp) {
821 assert_not_delayed();
822 v9_dep();
823
824 int hi = (int)(value >> 32);
825 int lo = (int)(value & ~0);
826 int bits_33to2 = (int)((value >> 2) & ~0);
827 // (Matcher::isSimpleConstant64 knows about the following optimizations.)
828 if (Assembler::is_simm13(lo) && value == lo) {
829 or3(G0, lo, d);
830 } else if (hi == 0) {
831 Assembler::sethi(lo, d); // hardware version zero-extends to upper 32
832 if (low10(lo) != 0)
833 or3(d, low10(lo), d);
834 }
835 else if ((hi >> 2) == 0) {
836 Assembler::sethi(bits_33to2, d); // hardware version zero-extends to upper 32
837 sllx(d, 2, d);
838 if (low12(lo) != 0)
839 or3(d, low12(lo), d);
840 }
841 else if (hi == -1) {
842 Assembler::sethi(~lo, d); // hardware version zero-extends to upper 32
843 xor3(d, low10(lo) ^ ~low10(~0), d);
844 }
845 else if (lo == 0) {
846 if (Assembler::is_simm13(hi)) {
847 or3(G0, hi, d);
848 } else {
849 Assembler::sethi(hi, d); // hardware version zero-extends to upper 32
850 if (low10(hi) != 0)
851 or3(d, low10(hi), d);
852 }
853 sllx(d, 32, d);
854 }
855 else {
856 Assembler::sethi(hi, tmp);
857 Assembler::sethi(lo, d); // macro assembler version sign-extends
858 if (low10(hi) != 0)
859 or3 (tmp, low10(hi), tmp);
860 if (low10(lo) != 0)
861 or3 ( d, low10(lo), d);
862 sllx(tmp, 32, tmp);
863 or3 (d, tmp, d);
864 }
865 }
866
867 int MacroAssembler::insts_for_set64(jlong value) {
868 v9_dep();
869
870 int hi = (int) (value >> 32);
871 int lo = (int) (value & ~0);
872 int count = 0;
873
874 // (Matcher::isSimpleConstant64 knows about the following optimizations.)
875 if (Assembler::is_simm13(lo) && value == lo) {
876 count++;
877 } else if (hi == 0) {
878 count++;
879 if (low10(lo) != 0)
880 count++;
881 }
882 else if (hi == -1) {
883 count += 2;
884 }
885 else if (lo == 0) {
886 if (Assembler::is_simm13(hi)) {
887 count++;
888 } else {
889 count++;
890 if (low10(hi) != 0)
891 count++;
892 }
893 count++;
894 }
895 else {
896 count += 2;
897 if (low10(hi) != 0)
898 count++;
899 if (low10(lo) != 0)
900 count++;
901 count += 2;
902 }
903 return count;
904 }
905
906 // compute size in bytes of sparc frame, given
907 // number of extraWords
908 int MacroAssembler::total_frame_size_in_bytes(int extraWords) {
909
910 int nWords = frame::memory_parameter_word_sp_offset;
911
912 nWords += extraWords;
913
914 if (nWords & 1) ++nWords; // round up to double-word
915
916 return nWords * BytesPerWord;
917 }
918
919
920 // save_frame: given number of "extra" words in frame,
921 // issue approp. save instruction (p 200, v8 manual)
922
923 void MacroAssembler::save_frame(int extraWords) {
924 int delta = -total_frame_size_in_bytes(extraWords);
925 if (is_simm13(delta)) {
926 save(SP, delta, SP);
927 } else {
928 set(delta, G3_scratch);
929 save(SP, G3_scratch, SP);
930 }
931 }
932
933
934 void MacroAssembler::save_frame_c1(int size_in_bytes) {
935 if (is_simm13(-size_in_bytes)) {
936 save(SP, -size_in_bytes, SP);
937 } else {
938 set(-size_in_bytes, G3_scratch);
939 save(SP, G3_scratch, SP);
940 }
941 }
942
943
944 void MacroAssembler::save_frame_and_mov(int extraWords,
945 Register s1, Register d1,
946 Register s2, Register d2) {
947 assert_not_delayed();
948
949 // The trick here is to use precisely the same memory word
950 // that trap handlers also use to save the register.
951 // This word cannot be used for any other purpose, but
952 // it works fine to save the register's value, whether or not
953 // an interrupt flushes register windows at any given moment!
954 Address s1_addr;
955 if (s1->is_valid() && (s1->is_in() || s1->is_local())) {
956 s1_addr = s1->address_in_saved_window();
957 st_ptr(s1, s1_addr);
958 }
959
960 Address s2_addr;
961 if (s2->is_valid() && (s2->is_in() || s2->is_local())) {
962 s2_addr = s2->address_in_saved_window();
963 st_ptr(s2, s2_addr);
964 }
965
966 save_frame(extraWords);
967
968 if (s1_addr.base() == SP) {
969 ld_ptr(s1_addr.after_save(), d1);
970 } else if (s1->is_valid()) {
971 mov(s1->after_save(), d1);
972 }
973
974 if (s2_addr.base() == SP) {
975 ld_ptr(s2_addr.after_save(), d2);
976 } else if (s2->is_valid()) {
977 mov(s2->after_save(), d2);
978 }
979 }
980
981
982 AddressLiteral MacroAssembler::allocate_metadata_address(Metadata* obj) {
983 assert(oop_recorder() != NULL, "this assembler needs a Recorder");
984 int index = oop_recorder()->allocate_metadata_index(obj);
985 RelocationHolder rspec = metadata_Relocation::spec(index);
986 return AddressLiteral((address)obj, rspec);
987 }
988
989 AddressLiteral MacroAssembler::constant_metadata_address(Metadata* obj) {
990 assert(oop_recorder() != NULL, "this assembler needs a Recorder");
991 int index = oop_recorder()->find_index(obj);
992 RelocationHolder rspec = metadata_Relocation::spec(index);
993 return AddressLiteral((address)obj, rspec);
994 }
995
996
997 AddressLiteral MacroAssembler::constant_oop_address(jobject obj) {
998 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
999 assert(Universe::heap()->is_in_reserved(JNIHandles::resolve(obj)), "not an oop");
1000 int oop_index = oop_recorder()->find_index(obj);
1001 return AddressLiteral(obj, oop_Relocation::spec(oop_index));
1002 }
1003
1004 void MacroAssembler::set_narrow_oop(jobject obj, Register d) {
1005 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1006 int oop_index = oop_recorder()->find_index(obj);
1007 RelocationHolder rspec = oop_Relocation::spec(oop_index);
1008
1009 assert_not_delayed();
1010 // Relocation with special format (see relocInfo_sparc.hpp).
1011 relocate(rspec, 1);
1012 // Assembler::sethi(0x3fffff, d);
1013 emit_int32( op(branch_op) | rd(d) | op2(sethi_op2) | hi22(0x3fffff) );
1014 // Don't add relocation for 'add'. Do patching during 'sethi' processing.
1015 add(d, 0x3ff, d);
1016
1017 }
1018
1019 void MacroAssembler::set_narrow_klass(Klass* k, Register d) {
1020 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1021 int klass_index = oop_recorder()->find_index(k);
1022 RelocationHolder rspec = metadata_Relocation::spec(klass_index);
1023 narrowOop encoded_k = Klass::encode_klass(k);
1024
1025 assert_not_delayed();
1026 // Relocation with special format (see relocInfo_sparc.hpp).
1027 relocate(rspec, 1);
1028 // Assembler::sethi(encoded_k, d);
1029 emit_int32( op(branch_op) | rd(d) | op2(sethi_op2) | hi22(encoded_k) );
1030 // Don't add relocation for 'add'. Do patching during 'sethi' processing.
1031 add(d, low10(encoded_k), d);
1032
1033 }
1034
1035 void MacroAssembler::align(int modulus) {
1036 while (offset() % modulus != 0) nop();
1037 }
1038
1039 void RegistersForDebugging::print(outputStream* s) {
1040 FlagSetting fs(Debugging, true);
1041 int j;
1042 for (j = 0; j < 8; ++j) {
1043 if (j != 6) { s->print("i%d = ", j); os::print_location(s, i[j]); }
1044 else { s->print( "fp = " ); os::print_location(s, i[j]); }
1045 }
1046 s->cr();
1047
1048 for (j = 0; j < 8; ++j) {
1049 s->print("l%d = ", j); os::print_location(s, l[j]);
1050 }
1051 s->cr();
1052
1053 for (j = 0; j < 8; ++j) {
1054 if (j != 6) { s->print("o%d = ", j); os::print_location(s, o[j]); }
1055 else { s->print( "sp = " ); os::print_location(s, o[j]); }
1056 }
1057 s->cr();
1058
1059 for (j = 0; j < 8; ++j) {
1060 s->print("g%d = ", j); os::print_location(s, g[j]);
1061 }
1062 s->cr();
1063
1064 // print out floats with compression
1065 for (j = 0; j < 32; ) {
1066 jfloat val = f[j];
1067 int last = j;
1068 for ( ; last+1 < 32; ++last ) {
1069 char b1[1024], b2[1024];
1070 sprintf(b1, "%f", val);
1071 sprintf(b2, "%f", f[last+1]);
1072 if (strcmp(b1, b2))
1073 break;
1074 }
1075 s->print("f%d", j);
1076 if ( j != last ) s->print(" - f%d", last);
1077 s->print(" = %f", val);
1078 s->fill_to(25);
1079 s->print_cr(" (0x%x)", *(int*)&val);
1080 j = last + 1;
1081 }
1082 s->cr();
1083
1084 // and doubles (evens only)
1085 for (j = 0; j < 32; ) {
1086 jdouble val = d[j];
1087 int last = j;
1088 for ( ; last+1 < 32; ++last ) {
1089 char b1[1024], b2[1024];
1090 sprintf(b1, "%f", val);
1091 sprintf(b2, "%f", d[last+1]);
1092 if (strcmp(b1, b2))
1093 break;
1094 }
1095 s->print("d%d", 2 * j);
1096 if ( j != last ) s->print(" - d%d", last);
1097 s->print(" = %f", val);
1098 s->fill_to(30);
1099 s->print("(0x%x)", *(int*)&val);
1100 s->fill_to(42);
1101 s->print_cr("(0x%x)", *(1 + (int*)&val));
1102 j = last + 1;
1103 }
1104 s->cr();
1105 }
1106
1107 void RegistersForDebugging::save_registers(MacroAssembler* a) {
1108 a->sub(FP, align_up(sizeof(RegistersForDebugging), sizeof(jdouble)) - STACK_BIAS, O0);
1109 a->flushw();
1110 int i;
1111 for (i = 0; i < 8; ++i) {
1112 a->ld_ptr(as_iRegister(i)->address_in_saved_window().after_save(), L1); a->st_ptr( L1, O0, i_offset(i));
1113 a->ld_ptr(as_lRegister(i)->address_in_saved_window().after_save(), L1); a->st_ptr( L1, O0, l_offset(i));
1114 a->st_ptr(as_oRegister(i)->after_save(), O0, o_offset(i));
1115 a->st_ptr(as_gRegister(i)->after_save(), O0, g_offset(i));
1116 }
1117 for (i = 0; i < 32; ++i) {
1118 a->stf(FloatRegisterImpl::S, as_FloatRegister(i), O0, f_offset(i));
1119 }
1120 for (i = 0; i < 64; i += 2) {
1121 a->stf(FloatRegisterImpl::D, as_FloatRegister(i), O0, d_offset(i));
1122 }
1123 }
1124
1125 void RegistersForDebugging::restore_registers(MacroAssembler* a, Register r) {
1126 for (int i = 1; i < 8; ++i) {
1127 a->ld_ptr(r, g_offset(i), as_gRegister(i));
1128 }
1129 for (int j = 0; j < 32; ++j) {
1130 a->ldf(FloatRegisterImpl::S, O0, f_offset(j), as_FloatRegister(j));
1131 }
1132 for (int k = 0; k < 64; k += 2) {
1133 a->ldf(FloatRegisterImpl::D, O0, d_offset(k), as_FloatRegister(k));
1134 }
1135 }
1136
1137
1138 // pushes double TOS element of FPU stack on CPU stack; pops from FPU stack
1139 void MacroAssembler::push_fTOS() {
1140 // %%%%%% need to implement this
1141 }
1142
1143 // pops double TOS element from CPU stack and pushes on FPU stack
1144 void MacroAssembler::pop_fTOS() {
1145 // %%%%%% need to implement this
1146 }
1147
1148 void MacroAssembler::empty_FPU_stack() {
1149 // %%%%%% need to implement this
1150 }
1151
1152 void MacroAssembler::_verify_oop(Register reg, const char* msg, const char * file, int line) {
1153 // plausibility check for oops
1154 if (!VerifyOops) return;
1155
1156 if (reg == G0) return; // always NULL, which is always an oop
1157
1158 BLOCK_COMMENT("verify_oop {");
1159 char buffer[64];
1160 #ifdef COMPILER1
1161 if (CommentedAssembly) {
1162 snprintf(buffer, sizeof(buffer), "verify_oop at %d", offset());
1163 block_comment(buffer);
1164 }
1165 #endif
1166
1167 const char* real_msg = NULL;
1168 {
1169 ResourceMark rm;
1170 stringStream ss;
1171 ss.print("%s at offset %d (%s:%d)", msg, offset(), file, line);
1172 real_msg = code_string(ss.as_string());
1173 }
1174
1175 // Call indirectly to solve generation ordering problem
1176 AddressLiteral a(StubRoutines::verify_oop_subroutine_entry_address());
1177
1178 // Make some space on stack above the current register window.
1179 // Enough to hold 8 64-bit registers.
1180 add(SP,-8*8,SP);
1181
1182 // Save some 64-bit registers; a normal 'save' chops the heads off
1183 // of 64-bit longs in the 32-bit build.
1184 stx(O0,SP,frame::register_save_words*wordSize+STACK_BIAS+0*8);
1185 stx(O1,SP,frame::register_save_words*wordSize+STACK_BIAS+1*8);
1186 mov(reg,O0); // Move arg into O0; arg might be in O7 which is about to be crushed
1187 stx(O7,SP,frame::register_save_words*wordSize+STACK_BIAS+7*8);
1188
1189 // Size of set() should stay the same
1190 patchable_set((intptr_t)real_msg, O1);
1191 // Load address to call to into O7
1192 load_ptr_contents(a, O7);
1193 // Register call to verify_oop_subroutine
1194 callr(O7, G0);
1195 delayed()->nop();
1196 // recover frame size
1197 add(SP, 8*8,SP);
1198 BLOCK_COMMENT("} verify_oop");
1199 }
1200
1201 void MacroAssembler::_verify_oop_addr(Address addr, const char* msg, const char * file, int line) {
1202 // plausibility check for oops
1203 if (!VerifyOops) return;
1204
1205 const char* real_msg = NULL;
1206 {
1207 ResourceMark rm;
1208 stringStream ss;
1209 ss.print("%s at SP+%d (%s:%d)", msg, addr.disp(), file, line);
1210 real_msg = code_string(ss.as_string());
1211 }
1212
1213 // Call indirectly to solve generation ordering problem
1214 AddressLiteral a(StubRoutines::verify_oop_subroutine_entry_address());
1215
1216 // Make some space on stack above the current register window.
1217 // Enough to hold 8 64-bit registers.
1218 add(SP,-8*8,SP);
1219
1220 // Save some 64-bit registers; a normal 'save' chops the heads off
1221 // of 64-bit longs in the 32-bit build.
1222 stx(O0,SP,frame::register_save_words*wordSize+STACK_BIAS+0*8);
1223 stx(O1,SP,frame::register_save_words*wordSize+STACK_BIAS+1*8);
1224 ld_ptr(addr.base(), addr.disp() + 8*8, O0); // Load arg into O0; arg might be in O7 which is about to be crushed
1225 stx(O7,SP,frame::register_save_words*wordSize+STACK_BIAS+7*8);
1226
1227 // Size of set() should stay the same
1228 patchable_set((intptr_t)real_msg, O1);
1229 // Load address to call to into O7
1230 load_ptr_contents(a, O7);
1231 // Register call to verify_oop_subroutine
1232 callr(O7, G0);
1233 delayed()->nop();
1234 // recover frame size
1235 add(SP, 8*8,SP);
1236 }
1237
1238 // side-door communication with signalHandler in os_solaris.cpp
1239 address MacroAssembler::_verify_oop_implicit_branch[3] = { NULL };
1240
1241 // This macro is expanded just once; it creates shared code. Contract:
1242 // receives an oop in O0. Must restore O0 & O7 from TLS. Must not smash ANY
1243 // registers, including flags. May not use a register 'save', as this blows
1244 // the high bits of the O-regs if they contain Long values. Acts as a 'leaf'
1245 // call.
1246 void MacroAssembler::verify_oop_subroutine() {
1247 // Leaf call; no frame.
1248 Label succeed, fail, null_or_fail;
1249
1250 // O0 and O7 were saved already (O0 in O0's TLS home, O7 in O5's TLS home).
1251 // O0 is now the oop to be checked. O7 is the return address.
1252 Register O0_obj = O0;
1253
1254 // Save some more registers for temps.
1255 stx(O2,SP,frame::register_save_words*wordSize+STACK_BIAS+2*8);
1256 stx(O3,SP,frame::register_save_words*wordSize+STACK_BIAS+3*8);
1257 stx(O4,SP,frame::register_save_words*wordSize+STACK_BIAS+4*8);
1258 stx(O5,SP,frame::register_save_words*wordSize+STACK_BIAS+5*8);
1259
1260 // Save flags
1261 Register O5_save_flags = O5;
1262 rdccr( O5_save_flags );
1263
1264 { // count number of verifies
1265 Register O2_adr = O2;
1266 Register O3_accum = O3;
1267 inc_counter(StubRoutines::verify_oop_count_addr(), O2_adr, O3_accum);
1268 }
1269
1270 Register O2_mask = O2;
1271 Register O3_bits = O3;
1272 Register O4_temp = O4;
1273
1274 // mark lower end of faulting range
1275 assert(_verify_oop_implicit_branch[0] == NULL, "set once");
1276 _verify_oop_implicit_branch[0] = pc();
1277
1278 // We can't check the mark oop because it could be in the process of
1279 // locking or unlocking while this is running.
1280 set(Universe::verify_oop_mask (), O2_mask);
1281 set(Universe::verify_oop_bits (), O3_bits);
1282
1283 // assert((obj & oop_mask) == oop_bits);
1284 and3(O0_obj, O2_mask, O4_temp);
1285 cmp_and_brx_short(O4_temp, O3_bits, notEqual, pn, null_or_fail);
1286
1287 if ((NULL_WORD & Universe::verify_oop_mask()) == Universe::verify_oop_bits()) {
1288 // the null_or_fail case is useless; must test for null separately
1289 br_null_short(O0_obj, pn, succeed);
1290 }
1291
1292 // Check the Klass* of this object for being in the right area of memory.
1293 // Cannot do the load in the delay above slot in case O0 is null
1294 load_klass(O0_obj, O0_obj);
1295 // assert((klass != NULL)
1296 br_null_short(O0_obj, pn, fail);
1297
1298 wrccr( O5_save_flags ); // Restore CCR's
1299
1300 // mark upper end of faulting range
1301 _verify_oop_implicit_branch[1] = pc();
1302
1303 //-----------------------
1304 // all tests pass
1305 bind(succeed);
1306
1307 // Restore prior 64-bit registers
1308 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+0*8,O0);
1309 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+1*8,O1);
1310 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+2*8,O2);
1311 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+3*8,O3);
1312 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+4*8,O4);
1313 ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+5*8,O5);
1314
1315 retl(); // Leaf return; restore prior O7 in delay slot
1316 delayed()->ldx(SP,frame::register_save_words*wordSize+STACK_BIAS+7*8,O7);
1317
1318 //-----------------------
1319 bind(null_or_fail); // nulls are less common but OK
1320 br_null(O0_obj, false, pt, succeed);
1321 delayed()->wrccr( O5_save_flags ); // Restore CCR's
1322
1323 //-----------------------
1324 // report failure:
1325 bind(fail);
1326 _verify_oop_implicit_branch[2] = pc();
1327
1328 wrccr( O5_save_flags ); // Restore CCR's
1329
1330 save_frame(align_up(sizeof(RegistersForDebugging) / BytesPerWord, 2));
1331
1332 // stop_subroutine expects message pointer in I1.
1333 mov(I1, O1);
1334
1335 // Restore prior 64-bit registers
1336 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+0*8,I0);
1337 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+1*8,I1);
1338 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+2*8,I2);
1339 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+3*8,I3);
1340 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+4*8,I4);
1341 ldx(FP,frame::register_save_words*wordSize+STACK_BIAS+5*8,I5);
1342
1343 // factor long stop-sequence into subroutine to save space
1344 assert(StubRoutines::Sparc::stop_subroutine_entry_address(), "hasn't been generated yet");
1345
1346 // call indirectly to solve generation ordering problem
1347 AddressLiteral al(StubRoutines::Sparc::stop_subroutine_entry_address());
1348 load_ptr_contents(al, O5);
1349 jmpl(O5, 0, O7);
1350 delayed()->nop();
1351 }
1352
1353
1354 void MacroAssembler::stop(const char* msg) {
1355 // save frame first to get O7 for return address
1356 // add one word to size in case struct is odd number of words long
1357 // It must be doubleword-aligned for storing doubles into it.
1358
1359 save_frame(align_up(sizeof(RegistersForDebugging) / BytesPerWord, 2));
1360
1361 // stop_subroutine expects message pointer in I1.
1362 // Size of set() should stay the same
1363 patchable_set((intptr_t)msg, O1);
1364
1365 // factor long stop-sequence into subroutine to save space
1366 assert(StubRoutines::Sparc::stop_subroutine_entry_address(), "hasn't been generated yet");
1367
1368 // call indirectly to solve generation ordering problem
1369 AddressLiteral a(StubRoutines::Sparc::stop_subroutine_entry_address());
1370 load_ptr_contents(a, O5);
1371 jmpl(O5, 0, O7);
1372 delayed()->nop();
1373
1374 breakpoint_trap(); // make stop actually stop rather than writing
1375 // unnoticeable results in the output files.
1376
1377 // restore(); done in callee to save space!
1378 }
1379
1380
1381 void MacroAssembler::warn(const char* msg) {
1382 save_frame(align_up(sizeof(RegistersForDebugging) / BytesPerWord, 2));
1383 RegistersForDebugging::save_registers(this);
1384 mov(O0, L0);
1385 // Size of set() should stay the same
1386 patchable_set((intptr_t)msg, O0);
1387 call( CAST_FROM_FN_PTR(address, warning) );
1388 delayed()->nop();
1389 // ret();
1390 // delayed()->restore();
1391 RegistersForDebugging::restore_registers(this, L0);
1392 restore();
1393 }
1394
1395
1396 void MacroAssembler::untested(const char* what) {
1397 // We must be able to turn interactive prompting off
1398 // in order to run automated test scripts on the VM
1399 // Use the flag ShowMessageBoxOnError
1400
1401 const char* b = NULL;
1402 {
1403 ResourceMark rm;
1404 stringStream ss;
1405 ss.print("untested: %s", what);
1406 b = code_string(ss.as_string());
1407 }
1408 if (ShowMessageBoxOnError) { STOP(b); }
1409 else { warn(b); }
1410 }
1411
1412
1413 void MacroAssembler::unimplemented(const char* what) {
1414 char* b = new char[1024];
1415 jio_snprintf(b, 1024, "unimplemented: %s", what);
1416 stop(b);
1417 }
1418
1419
1420 void MacroAssembler::stop_subroutine() {
1421 RegistersForDebugging::save_registers(this);
1422
1423 // for the sake of the debugger, stick a PC on the current frame
1424 // (this assumes that the caller has performed an extra "save")
1425 mov(I7, L7);
1426 add(O7, -7 * BytesPerInt, I7);
1427
1428 save_frame(); // one more save to free up another O7 register
1429 mov(I0, O1); // addr of reg save area
1430
1431 // We expect pointer to message in I1. Caller must set it up in O1
1432 mov(I1, O0); // get msg
1433 call (CAST_FROM_FN_PTR(address, MacroAssembler::debug), relocInfo::runtime_call_type);
1434 delayed()->nop();
1435
1436 restore();
1437
1438 RegistersForDebugging::restore_registers(this, O0);
1439
1440 save_frame(0);
1441 call(CAST_FROM_FN_PTR(address,breakpoint));
1442 delayed()->nop();
1443 restore();
1444
1445 mov(L7, I7);
1446 retl();
1447 delayed()->restore(); // see stop above
1448 }
1449
1450
1451 void MacroAssembler::debug(char* msg, RegistersForDebugging* regs) {
1452 if ( ShowMessageBoxOnError ) {
1453 JavaThread* thread = JavaThread::current();
1454 JavaThreadState saved_state = thread->thread_state();
1455 thread->set_thread_state(_thread_in_vm);
1456 {
1457 // In order to get locks work, we need to fake a in_VM state
1458 ttyLocker ttyl;
1459 ::tty->print_cr("EXECUTION STOPPED: %s\n", msg);
1460 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
1461 BytecodeCounter::print();
1462 }
1463 if (os::message_box(msg, "Execution stopped, print registers?"))
1464 regs->print(::tty);
1465 }
1466 BREAKPOINT;
1467 ThreadStateTransition::transition(JavaThread::current(), _thread_in_vm, saved_state);
1468 }
1469 else {
1470 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg);
1471 }
1472 assert(false, "DEBUG MESSAGE: %s", msg);
1473 }
1474
1475
1476 void MacroAssembler::calc_mem_param_words(Register Rparam_words, Register Rresult) {
1477 subcc( Rparam_words, Argument::n_register_parameters, Rresult); // how many mem words?
1478 Label no_extras;
1479 br( negative, true, pt, no_extras ); // if neg, clear reg
1480 delayed()->set(0, Rresult); // annuled, so only if taken
1481 bind( no_extras );
1482 }
1483
1484
1485 void MacroAssembler::calc_frame_size(Register Rextra_words, Register Rresult) {
1486 add(Rextra_words, frame::memory_parameter_word_sp_offset, Rresult);
1487 bclr(1, Rresult);
1488 sll(Rresult, LogBytesPerWord, Rresult); // Rresult has total frame bytes
1489 }
1490
1491
1492 void MacroAssembler::calc_frame_size_and_save(Register Rextra_words, Register Rresult) {
1493 calc_frame_size(Rextra_words, Rresult);
1494 neg(Rresult);
1495 save(SP, Rresult, SP);
1496 }
1497
1498
1499 // ---------------------------------------------------------
1500 Assembler::RCondition cond2rcond(Assembler::Condition c) {
1501 switch (c) {
1502 /*case zero: */
1503 case Assembler::equal: return Assembler::rc_z;
1504 case Assembler::lessEqual: return Assembler::rc_lez;
1505 case Assembler::less: return Assembler::rc_lz;
1506 /*case notZero:*/
1507 case Assembler::notEqual: return Assembler::rc_nz;
1508 case Assembler::greater: return Assembler::rc_gz;
1509 case Assembler::greaterEqual: return Assembler::rc_gez;
1510 }
1511 ShouldNotReachHere();
1512 return Assembler::rc_z;
1513 }
1514
1515 // compares (32 bit) register with zero and branches. NOT FOR USE WITH 64-bit POINTERS
1516 void MacroAssembler::cmp_zero_and_br(Condition c, Register s1, Label& L, bool a, Predict p) {
1517 tst(s1);
1518 br (c, a, p, L);
1519 }
1520
1521 // Compares a pointer register with zero and branches on null.
1522 // Does a test & branch on 32-bit systems and a register-branch on 64-bit.
1523 void MacroAssembler::br_null( Register s1, bool a, Predict p, Label& L ) {
1524 assert_not_delayed();
1525 bpr( rc_z, a, p, s1, L );
1526 }
1527
1528 void MacroAssembler::br_notnull( Register s1, bool a, Predict p, Label& L ) {
1529 assert_not_delayed();
1530 bpr( rc_nz, a, p, s1, L );
1531 }
1532
1533 // Compare registers and branch with nop in delay slot or cbcond without delay slot.
1534
1535 // Compare integer (32 bit) values (icc only).
1536 void MacroAssembler::cmp_and_br_short(Register s1, Register s2, Condition c,
1537 Predict p, Label& L) {
1538 assert_not_delayed();
1539 if (use_cbcond(L)) {
1540 Assembler::cbcond(c, icc, s1, s2, L);
1541 } else {
1542 cmp(s1, s2);
1543 br(c, false, p, L);
1544 delayed()->nop();
1545 }
1546 }
1547
1548 // Compare integer (32 bit) values (icc only).
1549 void MacroAssembler::cmp_and_br_short(Register s1, int simm13a, Condition c,
1550 Predict p, Label& L) {
1551 assert_not_delayed();
1552 if (is_simm(simm13a,5) && use_cbcond(L)) {
1553 Assembler::cbcond(c, icc, s1, simm13a, L);
1554 } else {
1555 cmp(s1, simm13a);
1556 br(c, false, p, L);
1557 delayed()->nop();
1558 }
1559 }
1560
1561 // Branch that tests xcc in LP64 and icc in !LP64
1562 void MacroAssembler::cmp_and_brx_short(Register s1, Register s2, Condition c,
1563 Predict p, Label& L) {
1564 assert_not_delayed();
1565 if (use_cbcond(L)) {
1566 Assembler::cbcond(c, ptr_cc, s1, s2, L);
1567 } else {
1568 cmp(s1, s2);
1569 brx(c, false, p, L);
1570 delayed()->nop();
1571 }
1572 }
1573
1574 // Branch that tests xcc in LP64 and icc in !LP64
1575 void MacroAssembler::cmp_and_brx_short(Register s1, int simm13a, Condition c,
1576 Predict p, Label& L) {
1577 assert_not_delayed();
1578 if (is_simm(simm13a,5) && use_cbcond(L)) {
1579 Assembler::cbcond(c, ptr_cc, s1, simm13a, L);
1580 } else {
1581 cmp(s1, simm13a);
1582 brx(c, false, p, L);
1583 delayed()->nop();
1584 }
1585 }
1586
1587 // Short branch version for compares a pointer with zero.
1588
1589 void MacroAssembler::br_null_short(Register s1, Predict p, Label& L) {
1590 assert_not_delayed();
1591 if (use_cbcond(L)) {
1592 Assembler::cbcond(zero, ptr_cc, s1, 0, L);
1593 } else {
1594 br_null(s1, false, p, L);
1595 delayed()->nop();
1596 }
1597 }
1598
1599 void MacroAssembler::br_notnull_short(Register s1, Predict p, Label& L) {
1600 assert_not_delayed();
1601 if (use_cbcond(L)) {
1602 Assembler::cbcond(notZero, ptr_cc, s1, 0, L);
1603 } else {
1604 br_notnull(s1, false, p, L);
1605 delayed()->nop();
1606 }
1607 }
1608
1609 // Unconditional short branch
1610 void MacroAssembler::ba_short(Label& L) {
1611 assert_not_delayed();
1612 if (use_cbcond(L)) {
1613 Assembler::cbcond(equal, icc, G0, G0, L);
1614 } else {
1615 br(always, false, pt, L);
1616 delayed()->nop();
1617 }
1618 }
1619
1620 // Branch if 'icc' says zero or not (i.e. icc.z == 1|0).
1621
1622 void MacroAssembler::br_icc_zero(bool iszero, Predict p, Label &L) {
1623 assert_not_delayed();
1624 Condition cf = (iszero ? Assembler::zero : Assembler::notZero);
1625 br(cf, false, p, L);
1626 delayed()->nop();
1627 }
1628
1629 // instruction sequences factored across compiler & interpreter
1630
1631
1632 void MacroAssembler::lcmp( Register Ra_hi, Register Ra_low,
1633 Register Rb_hi, Register Rb_low,
1634 Register Rresult) {
1635
1636 Label check_low_parts, done;
1637
1638 cmp(Ra_hi, Rb_hi ); // compare hi parts
1639 br(equal, true, pt, check_low_parts);
1640 delayed()->cmp(Ra_low, Rb_low); // test low parts
1641
1642 // And, with an unsigned comparison, it does not matter if the numbers
1643 // are negative or not.
1644 // E.g., -2 cmp -1: the low parts are 0xfffffffe and 0xffffffff.
1645 // The second one is bigger (unsignedly).
1646
1647 // Other notes: The first move in each triplet can be unconditional
1648 // (and therefore probably prefetchable).
1649 // And the equals case for the high part does not need testing,
1650 // since that triplet is reached only after finding the high halves differ.
1651
1652 mov(-1, Rresult);
1653 ba(done);
1654 delayed()->movcc(greater, false, icc, 1, Rresult);
1655
1656 bind(check_low_parts);
1657
1658 mov( -1, Rresult);
1659 movcc(equal, false, icc, 0, Rresult);
1660 movcc(greaterUnsigned, false, icc, 1, Rresult);
1661
1662 bind(done);
1663 }
1664
1665 void MacroAssembler::lneg( Register Rhi, Register Rlow ) {
1666 subcc( G0, Rlow, Rlow );
1667 subc( G0, Rhi, Rhi );
1668 }
1669
1670 void MacroAssembler::lshl( Register Rin_high, Register Rin_low,
1671 Register Rcount,
1672 Register Rout_high, Register Rout_low,
1673 Register Rtemp ) {
1674
1675
1676 Register Ralt_count = Rtemp;
1677 Register Rxfer_bits = Rtemp;
1678
1679 assert( Ralt_count != Rin_high
1680 && Ralt_count != Rin_low
1681 && Ralt_count != Rcount
1682 && Rxfer_bits != Rin_low
1683 && Rxfer_bits != Rin_high
1684 && Rxfer_bits != Rcount
1685 && Rxfer_bits != Rout_low
1686 && Rout_low != Rin_high,
1687 "register alias checks");
1688
1689 Label big_shift, done;
1690
1691 // This code can be optimized to use the 64 bit shifts in V9.
1692 // Here we use the 32 bit shifts.
1693
1694 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits
1695 subcc(Rcount, 31, Ralt_count);
1696 br(greater, true, pn, big_shift);
1697 delayed()->dec(Ralt_count);
1698
1699 // shift < 32 bits, Ralt_count = Rcount-31
1700
1701 // We get the transfer bits by shifting right by 32-count the low
1702 // register. This is done by shifting right by 31-count and then by one
1703 // more to take care of the special (rare) case where count is zero
1704 // (shifting by 32 would not work).
1705
1706 neg(Ralt_count);
1707
1708 // The order of the next two instructions is critical in the case where
1709 // Rin and Rout are the same and should not be reversed.
1710
1711 srl(Rin_low, Ralt_count, Rxfer_bits); // shift right by 31-count
1712 if (Rcount != Rout_low) {
1713 sll(Rin_low, Rcount, Rout_low); // low half
1714 }
1715 sll(Rin_high, Rcount, Rout_high);
1716 if (Rcount == Rout_low) {
1717 sll(Rin_low, Rcount, Rout_low); // low half
1718 }
1719 srl(Rxfer_bits, 1, Rxfer_bits ); // shift right by one more
1720 ba(done);
1721 delayed()->or3(Rout_high, Rxfer_bits, Rout_high); // new hi value: or in shifted old hi part and xfer from low
1722
1723 // shift >= 32 bits, Ralt_count = Rcount-32
1724 bind(big_shift);
1725 sll(Rin_low, Ralt_count, Rout_high );
1726 clr(Rout_low);
1727
1728 bind(done);
1729 }
1730
1731
1732 void MacroAssembler::lshr( Register Rin_high, Register Rin_low,
1733 Register Rcount,
1734 Register Rout_high, Register Rout_low,
1735 Register Rtemp ) {
1736
1737 Register Ralt_count = Rtemp;
1738 Register Rxfer_bits = Rtemp;
1739
1740 assert( Ralt_count != Rin_high
1741 && Ralt_count != Rin_low
1742 && Ralt_count != Rcount
1743 && Rxfer_bits != Rin_low
1744 && Rxfer_bits != Rin_high
1745 && Rxfer_bits != Rcount
1746 && Rxfer_bits != Rout_high
1747 && Rout_high != Rin_low,
1748 "register alias checks");
1749
1750 Label big_shift, done;
1751
1752 // This code can be optimized to use the 64 bit shifts in V9.
1753 // Here we use the 32 bit shifts.
1754
1755 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits
1756 subcc(Rcount, 31, Ralt_count);
1757 br(greater, true, pn, big_shift);
1758 delayed()->dec(Ralt_count);
1759
1760 // shift < 32 bits, Ralt_count = Rcount-31
1761
1762 // We get the transfer bits by shifting left by 32-count the high
1763 // register. This is done by shifting left by 31-count and then by one
1764 // more to take care of the special (rare) case where count is zero
1765 // (shifting by 32 would not work).
1766
1767 neg(Ralt_count);
1768 if (Rcount != Rout_low) {
1769 srl(Rin_low, Rcount, Rout_low);
1770 }
1771
1772 // The order of the next two instructions is critical in the case where
1773 // Rin and Rout are the same and should not be reversed.
1774
1775 sll(Rin_high, Ralt_count, Rxfer_bits); // shift left by 31-count
1776 sra(Rin_high, Rcount, Rout_high ); // high half
1777 sll(Rxfer_bits, 1, Rxfer_bits); // shift left by one more
1778 if (Rcount == Rout_low) {
1779 srl(Rin_low, Rcount, Rout_low);
1780 }
1781 ba(done);
1782 delayed()->or3(Rout_low, Rxfer_bits, Rout_low); // new low value: or shifted old low part and xfer from high
1783
1784 // shift >= 32 bits, Ralt_count = Rcount-32
1785 bind(big_shift);
1786
1787 sra(Rin_high, Ralt_count, Rout_low);
1788 sra(Rin_high, 31, Rout_high); // sign into hi
1789
1790 bind( done );
1791 }
1792
1793
1794
1795 void MacroAssembler::lushr( Register Rin_high, Register Rin_low,
1796 Register Rcount,
1797 Register Rout_high, Register Rout_low,
1798 Register Rtemp ) {
1799
1800 Register Ralt_count = Rtemp;
1801 Register Rxfer_bits = Rtemp;
1802
1803 assert( Ralt_count != Rin_high
1804 && Ralt_count != Rin_low
1805 && Ralt_count != Rcount
1806 && Rxfer_bits != Rin_low
1807 && Rxfer_bits != Rin_high
1808 && Rxfer_bits != Rcount
1809 && Rxfer_bits != Rout_high
1810 && Rout_high != Rin_low,
1811 "register alias checks");
1812
1813 Label big_shift, done;
1814
1815 // This code can be optimized to use the 64 bit shifts in V9.
1816 // Here we use the 32 bit shifts.
1817
1818 and3( Rcount, 0x3f, Rcount); // take least significant 6 bits
1819 subcc(Rcount, 31, Ralt_count);
1820 br(greater, true, pn, big_shift);
1821 delayed()->dec(Ralt_count);
1822
1823 // shift < 32 bits, Ralt_count = Rcount-31
1824
1825 // We get the transfer bits by shifting left by 32-count the high
1826 // register. This is done by shifting left by 31-count and then by one
1827 // more to take care of the special (rare) case where count is zero
1828 // (shifting by 32 would not work).
1829
1830 neg(Ralt_count);
1831 if (Rcount != Rout_low) {
1832 srl(Rin_low, Rcount, Rout_low);
1833 }
1834
1835 // The order of the next two instructions is critical in the case where
1836 // Rin and Rout are the same and should not be reversed.
1837
1838 sll(Rin_high, Ralt_count, Rxfer_bits); // shift left by 31-count
1839 srl(Rin_high, Rcount, Rout_high ); // high half
1840 sll(Rxfer_bits, 1, Rxfer_bits); // shift left by one more
1841 if (Rcount == Rout_low) {
1842 srl(Rin_low, Rcount, Rout_low);
1843 }
1844 ba(done);
1845 delayed()->or3(Rout_low, Rxfer_bits, Rout_low); // new low value: or shifted old low part and xfer from high
1846
1847 // shift >= 32 bits, Ralt_count = Rcount-32
1848 bind(big_shift);
1849
1850 srl(Rin_high, Ralt_count, Rout_low);
1851 clr(Rout_high);
1852
1853 bind( done );
1854 }
1855
1856 void MacroAssembler::lcmp( Register Ra, Register Rb, Register Rresult) {
1857 cmp(Ra, Rb);
1858 mov(-1, Rresult);
1859 movcc(equal, false, xcc, 0, Rresult);
1860 movcc(greater, false, xcc, 1, Rresult);
1861 }
1862
1863
1864 void MacroAssembler::load_sized_value(Address src, Register dst, size_t size_in_bytes, bool is_signed) {
1865 switch (size_in_bytes) {
1866 case 8: ld_long(src, dst); break;
1867 case 4: ld( src, dst); break;
1868 case 2: is_signed ? ldsh(src, dst) : lduh(src, dst); break;
1869 case 1: is_signed ? ldsb(src, dst) : ldub(src, dst); break;
1870 default: ShouldNotReachHere();
1871 }
1872 }
1873
1874 void MacroAssembler::store_sized_value(Register src, Address dst, size_t size_in_bytes) {
1875 switch (size_in_bytes) {
1876 case 8: st_long(src, dst); break;
1877 case 4: st( src, dst); break;
1878 case 2: sth( src, dst); break;
1879 case 1: stb( src, dst); break;
1880 default: ShouldNotReachHere();
1881 }
1882 }
1883
1884
1885 void MacroAssembler::float_cmp( bool is_float, int unordered_result,
1886 FloatRegister Fa, FloatRegister Fb,
1887 Register Rresult) {
1888 if (is_float) {
1889 fcmp(FloatRegisterImpl::S, fcc0, Fa, Fb);
1890 } else {
1891 fcmp(FloatRegisterImpl::D, fcc0, Fa, Fb);
1892 }
1893
1894 if (unordered_result == 1) {
1895 mov( -1, Rresult);
1896 movcc(f_equal, true, fcc0, 0, Rresult);
1897 movcc(f_unorderedOrGreater, true, fcc0, 1, Rresult);
1898 } else {
1899 mov( -1, Rresult);
1900 movcc(f_equal, true, fcc0, 0, Rresult);
1901 movcc(f_greater, true, fcc0, 1, Rresult);
1902 }
1903 }
1904
1905
1906 void MacroAssembler::save_all_globals_into_locals() {
1907 mov(G1,L1);
1908 mov(G2,L2);
1909 mov(G3,L3);
1910 mov(G4,L4);
1911 mov(G5,L5);
1912 mov(G6,L6);
1913 mov(G7,L7);
1914 }
1915
1916 void MacroAssembler::restore_globals_from_locals() {
1917 mov(L1,G1);
1918 mov(L2,G2);
1919 mov(L3,G3);
1920 mov(L4,G4);
1921 mov(L5,G5);
1922 mov(L6,G6);
1923 mov(L7,G7);
1924 }
1925
1926 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
1927 Register tmp,
1928 int offset) {
1929 intptr_t value = *delayed_value_addr;
1930 if (value != 0)
1931 return RegisterOrConstant(value + offset);
1932
1933 // load indirectly to solve generation ordering problem
1934 AddressLiteral a(delayed_value_addr);
1935 load_ptr_contents(a, tmp);
1936
1937 #ifdef ASSERT
1938 tst(tmp);
1939 breakpoint_trap(zero, xcc);
1940 #endif
1941
1942 if (offset != 0)
1943 add(tmp, offset, tmp);
1944
1945 return RegisterOrConstant(tmp);
1946 }
1947
1948
1949 RegisterOrConstant MacroAssembler::regcon_andn_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) {
1950 assert(d.register_or_noreg() != G0, "lost side effect");
1951 if ((s2.is_constant() && s2.as_constant() == 0) ||
1952 (s2.is_register() && s2.as_register() == G0)) {
1953 // Do nothing, just move value.
1954 if (s1.is_register()) {
1955 if (d.is_constant()) d = temp;
1956 mov(s1.as_register(), d.as_register());
1957 return d;
1958 } else {
1959 return s1;
1960 }
1961 }
1962
1963 if (s1.is_register()) {
1964 assert_different_registers(s1.as_register(), temp);
1965 if (d.is_constant()) d = temp;
1966 andn(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register());
1967 return d;
1968 } else {
1969 if (s2.is_register()) {
1970 assert_different_registers(s2.as_register(), temp);
1971 if (d.is_constant()) d = temp;
1972 set(s1.as_constant(), temp);
1973 andn(temp, s2.as_register(), d.as_register());
1974 return d;
1975 } else {
1976 intptr_t res = s1.as_constant() & ~s2.as_constant();
1977 return res;
1978 }
1979 }
1980 }
1981
1982 RegisterOrConstant MacroAssembler::regcon_inc_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) {
1983 assert(d.register_or_noreg() != G0, "lost side effect");
1984 if ((s2.is_constant() && s2.as_constant() == 0) ||
1985 (s2.is_register() && s2.as_register() == G0)) {
1986 // Do nothing, just move value.
1987 if (s1.is_register()) {
1988 if (d.is_constant()) d = temp;
1989 mov(s1.as_register(), d.as_register());
1990 return d;
1991 } else {
1992 return s1;
1993 }
1994 }
1995
1996 if (s1.is_register()) {
1997 assert_different_registers(s1.as_register(), temp);
1998 if (d.is_constant()) d = temp;
1999 add(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register());
2000 return d;
2001 } else {
2002 if (s2.is_register()) {
2003 assert_different_registers(s2.as_register(), temp);
2004 if (d.is_constant()) d = temp;
2005 add(s2.as_register(), ensure_simm13_or_reg(s1, temp), d.as_register());
2006 return d;
2007 } else {
2008 intptr_t res = s1.as_constant() + s2.as_constant();
2009 return res;
2010 }
2011 }
2012 }
2013
2014 RegisterOrConstant MacroAssembler::regcon_sll_ptr(RegisterOrConstant s1, RegisterOrConstant s2, RegisterOrConstant d, Register temp) {
2015 assert(d.register_or_noreg() != G0, "lost side effect");
2016 if (!is_simm13(s2.constant_or_zero()))
2017 s2 = (s2.as_constant() & 0xFF);
2018 if ((s2.is_constant() && s2.as_constant() == 0) ||
2019 (s2.is_register() && s2.as_register() == G0)) {
2020 // Do nothing, just move value.
2021 if (s1.is_register()) {
2022 if (d.is_constant()) d = temp;
2023 mov(s1.as_register(), d.as_register());
2024 return d;
2025 } else {
2026 return s1;
2027 }
2028 }
2029
2030 if (s1.is_register()) {
2031 assert_different_registers(s1.as_register(), temp);
2032 if (d.is_constant()) d = temp;
2033 sll_ptr(s1.as_register(), ensure_simm13_or_reg(s2, temp), d.as_register());
2034 return d;
2035 } else {
2036 if (s2.is_register()) {
2037 assert_different_registers(s2.as_register(), temp);
2038 if (d.is_constant()) d = temp;
2039 set(s1.as_constant(), temp);
2040 sll_ptr(temp, s2.as_register(), d.as_register());
2041 return d;
2042 } else {
2043 intptr_t res = s1.as_constant() << s2.as_constant();
2044 return res;
2045 }
2046 }
2047 }
2048
2049
2050 // Look up the method for a megamorphic invokeinterface call.
2051 // The target method is determined by <intf_klass, itable_index>.
2052 // The receiver klass is in recv_klass.
2053 // On success, the result will be in method_result, and execution falls through.
2054 // On failure, execution transfers to the given label.
2055 void MacroAssembler::lookup_interface_method(Register recv_klass,
2056 Register intf_klass,
2057 RegisterOrConstant itable_index,
2058 Register method_result,
2059 Register scan_temp,
2060 Register sethi_temp,
2061 Label& L_no_such_interface,
2062 bool return_method) {
2063 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
2064 assert(!return_method || itable_index.is_constant() || itable_index.as_register() == method_result,
2065 "caller must use same register for non-constant itable index as for method");
2066
2067 Label L_no_such_interface_restore;
2068 bool did_save = false;
2069 if (scan_temp == noreg || sethi_temp == noreg) {
2070 Register recv_2 = recv_klass->is_global() ? recv_klass : L0;
2071 Register intf_2 = intf_klass->is_global() ? intf_klass : L1;
2072 assert(method_result->is_global(), "must be able to return value");
2073 scan_temp = L2;
2074 sethi_temp = L3;
2075 save_frame_and_mov(0, recv_klass, recv_2, intf_klass, intf_2);
2076 recv_klass = recv_2;
2077 intf_klass = intf_2;
2078 did_save = true;
2079 }
2080
2081 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
2082 int vtable_base = in_bytes(Klass::vtable_start_offset());
2083 int scan_step = itableOffsetEntry::size() * wordSize;
2084 int vte_size = vtableEntry::size_in_bytes();
2085
2086 lduw(recv_klass, in_bytes(Klass::vtable_length_offset()), scan_temp);
2087 // %%% We should store the aligned, prescaled offset in the klassoop.
2088 // Then the next several instructions would fold away.
2089
2090 int itb_offset = vtable_base;
2091 int itb_scale = exact_log2(vtableEntry::size_in_bytes());
2092 sll(scan_temp, itb_scale, scan_temp);
2093 add(scan_temp, itb_offset, scan_temp);
2094 add(recv_klass, scan_temp, scan_temp);
2095
2096 if (return_method) {
2097 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
2098 RegisterOrConstant itable_offset = itable_index;
2099 itable_offset = regcon_sll_ptr(itable_index, exact_log2(itableMethodEntry::size() * wordSize), itable_offset);
2100 itable_offset = regcon_inc_ptr(itable_offset, itableMethodEntry::method_offset_in_bytes(), itable_offset);
2101 add(recv_klass, ensure_simm13_or_reg(itable_offset, sethi_temp), recv_klass);
2102 }
2103
2104 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
2105 // if (scan->interface() == intf) {
2106 // result = (klass + scan->offset() + itable_index);
2107 // }
2108 // }
2109 Label L_search, L_found_method;
2110
2111 for (int peel = 1; peel >= 0; peel--) {
2112 // %%%% Could load both offset and interface in one ldx, if they were
2113 // in the opposite order. This would save a load.
2114 ld_ptr(scan_temp, itableOffsetEntry::interface_offset_in_bytes(), method_result);
2115
2116 // Check that this entry is non-null. A null entry means that
2117 // the receiver class doesn't implement the interface, and wasn't the
2118 // same as when the caller was compiled.
2119 bpr(Assembler::rc_z, false, Assembler::pn, method_result, did_save ? L_no_such_interface_restore : L_no_such_interface);
2120 delayed()->cmp(method_result, intf_klass);
2121
2122 if (peel) {
2123 brx(Assembler::equal, false, Assembler::pt, L_found_method);
2124 } else {
2125 brx(Assembler::notEqual, false, Assembler::pn, L_search);
2126 // (invert the test to fall through to found_method...)
2127 }
2128 delayed()->add(scan_temp, scan_step, scan_temp);
2129
2130 if (!peel) break;
2131
2132 bind(L_search);
2133 }
2134
2135 bind(L_found_method);
2136
2137 if (return_method) {
2138 // Got a hit.
2139 int ito_offset = itableOffsetEntry::offset_offset_in_bytes();
2140 // scan_temp[-scan_step] points to the vtable offset we need
2141 ito_offset -= scan_step;
2142 lduw(scan_temp, ito_offset, scan_temp);
2143 ld_ptr(recv_klass, scan_temp, method_result);
2144 }
2145
2146 if (did_save) {
2147 Label L_done;
2148 ba(L_done);
2149 delayed()->restore();
2150
2151 bind(L_no_such_interface_restore);
2152 ba(L_no_such_interface);
2153 delayed()->restore();
2154
2155 bind(L_done);
2156 }
2157 }
2158
2159
2160 // virtual method calling
2161 void MacroAssembler::lookup_virtual_method(Register recv_klass,
2162 RegisterOrConstant vtable_index,
2163 Register method_result) {
2164 assert_different_registers(recv_klass, method_result, vtable_index.register_or_noreg());
2165 Register sethi_temp = method_result;
2166 const int base = in_bytes(Klass::vtable_start_offset()) +
2167 // method pointer offset within the vtable entry:
2168 vtableEntry::method_offset_in_bytes();
2169 RegisterOrConstant vtable_offset = vtable_index;
2170 // Each of the following three lines potentially generates an instruction.
2171 // But the total number of address formation instructions will always be
2172 // at most two, and will often be zero. In any case, it will be optimal.
2173 // If vtable_index is a register, we will have (sll_ptr N,x; inc_ptr B,x; ld_ptr k,x).
2174 // If vtable_index is a constant, we will have at most (set B+X<<N,t; ld_ptr k,t).
2175 vtable_offset = regcon_sll_ptr(vtable_index, exact_log2(vtableEntry::size_in_bytes()), vtable_offset);
2176 vtable_offset = regcon_inc_ptr(vtable_offset, base, vtable_offset, sethi_temp);
2177 Address vtable_entry_addr(recv_klass, ensure_simm13_or_reg(vtable_offset, sethi_temp));
2178 ld_ptr(vtable_entry_addr, method_result);
2179 }
2180
2181
2182 void MacroAssembler::check_klass_subtype(Register sub_klass,
2183 Register super_klass,
2184 Register temp_reg,
2185 Register temp2_reg,
2186 Label& L_success) {
2187 Register sub_2 = sub_klass;
2188 Register sup_2 = super_klass;
2189 if (!sub_2->is_global()) sub_2 = L0;
2190 if (!sup_2->is_global()) sup_2 = L1;
2191 bool did_save = false;
2192 if (temp_reg == noreg || temp2_reg == noreg) {
2193 temp_reg = L2;
2194 temp2_reg = L3;
2195 save_frame_and_mov(0, sub_klass, sub_2, super_klass, sup_2);
2196 sub_klass = sub_2;
2197 super_klass = sup_2;
2198 did_save = true;
2199 }
2200 Label L_failure, L_pop_to_failure, L_pop_to_success;
2201 check_klass_subtype_fast_path(sub_klass, super_klass,
2202 temp_reg, temp2_reg,
2203 (did_save ? &L_pop_to_success : &L_success),
2204 (did_save ? &L_pop_to_failure : &L_failure), NULL);
2205
2206 if (!did_save)
2207 save_frame_and_mov(0, sub_klass, sub_2, super_klass, sup_2);
2208 check_klass_subtype_slow_path(sub_2, sup_2,
2209 L2, L3, L4, L5,
2210 NULL, &L_pop_to_failure);
2211
2212 // on success:
2213 bind(L_pop_to_success);
2214 restore();
2215 ba_short(L_success);
2216
2217 // on failure:
2218 bind(L_pop_to_failure);
2219 restore();
2220 bind(L_failure);
2221 }
2222
2223
2224 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
2225 Register super_klass,
2226 Register temp_reg,
2227 Register temp2_reg,
2228 Label* L_success,
2229 Label* L_failure,
2230 Label* L_slow_path,
2231 RegisterOrConstant super_check_offset) {
2232 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
2233 int sco_offset = in_bytes(Klass::super_check_offset_offset());
2234
2235 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
2236 bool need_slow_path = (must_load_sco ||
2237 super_check_offset.constant_or_zero() == sco_offset);
2238
2239 assert_different_registers(sub_klass, super_klass, temp_reg);
2240 if (super_check_offset.is_register()) {
2241 assert_different_registers(sub_klass, super_klass, temp_reg,
2242 super_check_offset.as_register());
2243 } else if (must_load_sco) {
2244 assert(temp2_reg != noreg, "supply either a temp or a register offset");
2245 }
2246
2247 Label L_fallthrough;
2248 int label_nulls = 0;
2249 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
2250 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
2251 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
2252 assert(label_nulls <= 1 ||
2253 (L_slow_path == &L_fallthrough && label_nulls <= 2 && !need_slow_path),
2254 "at most one NULL in the batch, usually");
2255
2256 // If the pointers are equal, we are done (e.g., String[] elements).
2257 // This self-check enables sharing of secondary supertype arrays among
2258 // non-primary types such as array-of-interface. Otherwise, each such
2259 // type would need its own customized SSA.
2260 // We move this check to the front of the fast path because many
2261 // type checks are in fact trivially successful in this manner,
2262 // so we get a nicely predicted branch right at the start of the check.
2263 cmp(super_klass, sub_klass);
2264 brx(Assembler::equal, false, Assembler::pn, *L_success);
2265 delayed()->nop();
2266
2267 // Check the supertype display:
2268 if (must_load_sco) {
2269 // The super check offset is always positive...
2270 lduw(super_klass, sco_offset, temp2_reg);
2271 super_check_offset = RegisterOrConstant(temp2_reg);
2272 // super_check_offset is register.
2273 assert_different_registers(sub_klass, super_klass, temp_reg, super_check_offset.as_register());
2274 }
2275 ld_ptr(sub_klass, super_check_offset, temp_reg);
2276 cmp(super_klass, temp_reg);
2277
2278 // This check has worked decisively for primary supers.
2279 // Secondary supers are sought in the super_cache ('super_cache_addr').
2280 // (Secondary supers are interfaces and very deeply nested subtypes.)
2281 // This works in the same check above because of a tricky aliasing
2282 // between the super_cache and the primary super display elements.
2283 // (The 'super_check_addr' can address either, as the case requires.)
2284 // Note that the cache is updated below if it does not help us find
2285 // what we need immediately.
2286 // So if it was a primary super, we can just fail immediately.
2287 // Otherwise, it's the slow path for us (no success at this point).
2288
2289 // Hacked ba(), which may only be used just before L_fallthrough.
2290 #define FINAL_JUMP(label) \
2291 if (&(label) != &L_fallthrough) { \
2292 ba(label); delayed()->nop(); \
2293 }
2294
2295 if (super_check_offset.is_register()) {
2296 brx(Assembler::equal, false, Assembler::pn, *L_success);
2297 delayed()->cmp(super_check_offset.as_register(), sc_offset);
2298
2299 if (L_failure == &L_fallthrough) {
2300 brx(Assembler::equal, false, Assembler::pt, *L_slow_path);
2301 delayed()->nop();
2302 } else {
2303 brx(Assembler::notEqual, false, Assembler::pn, *L_failure);
2304 delayed()->nop();
2305 FINAL_JUMP(*L_slow_path);
2306 }
2307 } else if (super_check_offset.as_constant() == sc_offset) {
2308 // Need a slow path; fast failure is impossible.
2309 if (L_slow_path == &L_fallthrough) {
2310 brx(Assembler::equal, false, Assembler::pt, *L_success);
2311 delayed()->nop();
2312 } else {
2313 brx(Assembler::notEqual, false, Assembler::pn, *L_slow_path);
2314 delayed()->nop();
2315 FINAL_JUMP(*L_success);
2316 }
2317 } else {
2318 // No slow path; it's a fast decision.
2319 if (L_failure == &L_fallthrough) {
2320 brx(Assembler::equal, false, Assembler::pt, *L_success);
2321 delayed()->nop();
2322 } else {
2323 brx(Assembler::notEqual, false, Assembler::pn, *L_failure);
2324 delayed()->nop();
2325 FINAL_JUMP(*L_success);
2326 }
2327 }
2328
2329 bind(L_fallthrough);
2330
2331 #undef FINAL_JUMP
2332 }
2333
2334
2335 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
2336 Register super_klass,
2337 Register count_temp,
2338 Register scan_temp,
2339 Register scratch_reg,
2340 Register coop_reg,
2341 Label* L_success,
2342 Label* L_failure) {
2343 assert_different_registers(sub_klass, super_klass,
2344 count_temp, scan_temp, scratch_reg, coop_reg);
2345
2346 Label L_fallthrough, L_loop;
2347 int label_nulls = 0;
2348 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
2349 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
2350 assert(label_nulls <= 1, "at most one NULL in the batch");
2351
2352 // a couple of useful fields in sub_klass:
2353 int ss_offset = in_bytes(Klass::secondary_supers_offset());
2354 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
2355
2356 // Do a linear scan of the secondary super-klass chain.
2357 // This code is rarely used, so simplicity is a virtue here.
2358
2359 #ifndef PRODUCT
2360 int* pst_counter = &SharedRuntime::_partial_subtype_ctr;
2361 inc_counter((address) pst_counter, count_temp, scan_temp);
2362 #endif
2363
2364 // We will consult the secondary-super array.
2365 ld_ptr(sub_klass, ss_offset, scan_temp);
2366
2367 Register search_key = super_klass;
2368
2369 // Load the array length. (Positive movl does right thing on LP64.)
2370 lduw(scan_temp, Array<Klass*>::length_offset_in_bytes(), count_temp);
2371
2372 // Check for empty secondary super list
2373 tst(count_temp);
2374
2375 // In the array of super classes elements are pointer sized.
2376 int element_size = wordSize;
2377
2378 // Top of search loop
2379 bind(L_loop);
2380 br(Assembler::equal, false, Assembler::pn, *L_failure);
2381 delayed()->add(scan_temp, element_size, scan_temp);
2382
2383 // Skip the array header in all array accesses.
2384 int elem_offset = Array<Klass*>::base_offset_in_bytes();
2385 elem_offset -= element_size; // the scan pointer was pre-incremented also
2386
2387 // Load next super to check
2388 ld_ptr( scan_temp, elem_offset, scratch_reg );
2389
2390 // Look for Rsuper_klass on Rsub_klass's secondary super-class-overflow list
2391 cmp(scratch_reg, search_key);
2392
2393 // A miss means we are NOT a subtype and need to keep looping
2394 brx(Assembler::notEqual, false, Assembler::pn, L_loop);
2395 delayed()->deccc(count_temp); // decrement trip counter in delay slot
2396
2397 // Success. Cache the super we found and proceed in triumph.
2398 st_ptr(super_klass, sub_klass, sc_offset);
2399
2400 if (L_success != &L_fallthrough) {
2401 ba(*L_success);
2402 delayed()->nop();
2403 }
2404
2405 bind(L_fallthrough);
2406 }
2407
2408
2409 RegisterOrConstant MacroAssembler::argument_offset(RegisterOrConstant arg_slot,
2410 Register temp_reg,
2411 int extra_slot_offset) {
2412 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
2413 int stackElementSize = Interpreter::stackElementSize;
2414 int offset = extra_slot_offset * stackElementSize;
2415 if (arg_slot.is_constant()) {
2416 offset += arg_slot.as_constant() * stackElementSize;
2417 return offset;
2418 } else {
2419 assert(temp_reg != noreg, "must specify");
2420 sll_ptr(arg_slot.as_register(), exact_log2(stackElementSize), temp_reg);
2421 if (offset != 0)
2422 add(temp_reg, offset, temp_reg);
2423 return temp_reg;
2424 }
2425 }
2426
2427
2428 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
2429 Register temp_reg,
2430 int extra_slot_offset) {
2431 return Address(Gargs, argument_offset(arg_slot, temp_reg, extra_slot_offset));
2432 }
2433
2434
2435 void MacroAssembler::biased_locking_enter(Register obj_reg, Register mark_reg,
2436 Register temp_reg,
2437 Label& done, Label* slow_case,
2438 BiasedLockingCounters* counters) {
2439 assert(UseBiasedLocking, "why call this otherwise?");
2440
2441 if (PrintBiasedLockingStatistics) {
2442 assert_different_registers(obj_reg, mark_reg, temp_reg, O7);
2443 if (counters == NULL)
2444 counters = BiasedLocking::counters();
2445 }
2446
2447 Label cas_label;
2448
2449 // Biased locking
2450 // See whether the lock is currently biased toward our thread and
2451 // whether the epoch is still valid
2452 // Note that the runtime guarantees sufficient alignment of JavaThread
2453 // pointers to allow age to be placed into low bits
2454 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout");
2455 and3(mark_reg, markOopDesc::biased_lock_mask_in_place, temp_reg);
2456 cmp_and_brx_short(temp_reg, markOopDesc::biased_lock_pattern, Assembler::notEqual, Assembler::pn, cas_label);
2457
2458 load_klass(obj_reg, temp_reg);
2459 ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg);
2460 or3(G2_thread, temp_reg, temp_reg);
2461 xor3(mark_reg, temp_reg, temp_reg);
2462 andcc(temp_reg, ~((int) markOopDesc::age_mask_in_place), temp_reg);
2463 if (counters != NULL) {
2464 cond_inc(Assembler::equal, (address) counters->biased_lock_entry_count_addr(), mark_reg, temp_reg);
2465 // Reload mark_reg as we may need it later
2466 ld_ptr(Address(obj_reg, oopDesc::mark_offset_in_bytes()), mark_reg);
2467 }
2468 brx(Assembler::equal, true, Assembler::pt, done);
2469 delayed()->nop();
2470
2471 Label try_revoke_bias;
2472 Label try_rebias;
2473 Address mark_addr = Address(obj_reg, oopDesc::mark_offset_in_bytes());
2474 assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2475
2476 // At this point we know that the header has the bias pattern and
2477 // that we are not the bias owner in the current epoch. We need to
2478 // figure out more details about the state of the header in order to
2479 // know what operations can be legally performed on the object's
2480 // header.
2481
2482 // If the low three bits in the xor result aren't clear, that means
2483 // the prototype header is no longer biased and we have to revoke
2484 // the bias on this object.
2485 btst(markOopDesc::biased_lock_mask_in_place, temp_reg);
2486 brx(Assembler::notZero, false, Assembler::pn, try_revoke_bias);
2487
2488 // Biasing is still enabled for this data type. See whether the
2489 // epoch of the current bias is still valid, meaning that the epoch
2490 // bits of the mark word are equal to the epoch bits of the
2491 // prototype header. (Note that the prototype header's epoch bits
2492 // only change at a safepoint.) If not, attempt to rebias the object
2493 // toward the current thread. Note that we must be absolutely sure
2494 // that the current epoch is invalid in order to do this because
2495 // otherwise the manipulations it performs on the mark word are
2496 // illegal.
2497 delayed()->btst(markOopDesc::epoch_mask_in_place, temp_reg);
2498 brx(Assembler::notZero, false, Assembler::pn, try_rebias);
2499
2500 // The epoch of the current bias is still valid but we know nothing
2501 // about the owner; it might be set or it might be clear. Try to
2502 // acquire the bias of the object using an atomic operation. If this
2503 // fails we will go in to the runtime to revoke the object's bias.
2504 // Note that we first construct the presumed unbiased header so we
2505 // don't accidentally blow away another thread's valid bias.
2506 delayed()->and3(mark_reg,
2507 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place,
2508 mark_reg);
2509 or3(G2_thread, mark_reg, temp_reg);
2510 cas_ptr(mark_addr.base(), mark_reg, temp_reg);
2511 // If the biasing toward our thread failed, this means that
2512 // another thread succeeded in biasing it toward itself and we
2513 // need to revoke that bias. The revocation will occur in the
2514 // interpreter runtime in the slow case.
2515 cmp(mark_reg, temp_reg);
2516 if (counters != NULL) {
2517 cond_inc(Assembler::zero, (address) counters->anonymously_biased_lock_entry_count_addr(), mark_reg, temp_reg);
2518 }
2519 if (slow_case != NULL) {
2520 brx(Assembler::notEqual, true, Assembler::pn, *slow_case);
2521 delayed()->nop();
2522 }
2523 ba_short(done);
2524
2525 bind(try_rebias);
2526 // At this point we know the epoch has expired, meaning that the
2527 // current "bias owner", if any, is actually invalid. Under these
2528 // circumstances _only_, we are allowed to use the current header's
2529 // value as the comparison value when doing the cas to acquire the
2530 // bias in the current epoch. In other words, we allow transfer of
2531 // the bias from one thread to another directly in this situation.
2532 //
2533 // FIXME: due to a lack of registers we currently blow away the age
2534 // bits in this situation. Should attempt to preserve them.
2535 load_klass(obj_reg, temp_reg);
2536 ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg);
2537 or3(G2_thread, temp_reg, temp_reg);
2538 cas_ptr(mark_addr.base(), mark_reg, temp_reg);
2539 // If the biasing toward our thread failed, this means that
2540 // another thread succeeded in biasing it toward itself and we
2541 // need to revoke that bias. The revocation will occur in the
2542 // interpreter runtime in the slow case.
2543 cmp(mark_reg, temp_reg);
2544 if (counters != NULL) {
2545 cond_inc(Assembler::zero, (address) counters->rebiased_lock_entry_count_addr(), mark_reg, temp_reg);
2546 }
2547 if (slow_case != NULL) {
2548 brx(Assembler::notEqual, true, Assembler::pn, *slow_case);
2549 delayed()->nop();
2550 }
2551 ba_short(done);
2552
2553 bind(try_revoke_bias);
2554 // The prototype mark in the klass doesn't have the bias bit set any
2555 // more, indicating that objects of this data type are not supposed
2556 // to be biased any more. We are going to try to reset the mark of
2557 // this object to the prototype value and fall through to the
2558 // CAS-based locking scheme. Note that if our CAS fails, it means
2559 // that another thread raced us for the privilege of revoking the
2560 // bias of this particular object, so it's okay to continue in the
2561 // normal locking code.
2562 //
2563 // FIXME: due to a lack of registers we currently blow away the age
2564 // bits in this situation. Should attempt to preserve them.
2565 load_klass(obj_reg, temp_reg);
2566 ld_ptr(Address(temp_reg, Klass::prototype_header_offset()), temp_reg);
2567 cas_ptr(mark_addr.base(), mark_reg, temp_reg);
2568 // Fall through to the normal CAS-based lock, because no matter what
2569 // the result of the above CAS, some thread must have succeeded in
2570 // removing the bias bit from the object's header.
2571 if (counters != NULL) {
2572 cmp(mark_reg, temp_reg);
2573 cond_inc(Assembler::zero, (address) counters->revoked_lock_entry_count_addr(), mark_reg, temp_reg);
2574 }
2575
2576 bind(cas_label);
2577 }
2578
2579 void MacroAssembler::biased_locking_exit (Address mark_addr, Register temp_reg, Label& done,
2580 bool allow_delay_slot_filling) {
2581 // Check for biased locking unlock case, which is a no-op
2582 // Note: we do not have to check the thread ID for two reasons.
2583 // First, the interpreter checks for IllegalMonitorStateException at
2584 // a higher level. Second, if the bias was revoked while we held the
2585 // lock, the object could not be rebiased toward another thread, so
2586 // the bias bit would be clear.
2587 ld_ptr(mark_addr, temp_reg);
2588 and3(temp_reg, markOopDesc::biased_lock_mask_in_place, temp_reg);
2589 cmp(temp_reg, markOopDesc::biased_lock_pattern);
2590 brx(Assembler::equal, allow_delay_slot_filling, Assembler::pt, done);
2591 delayed();
2592 if (!allow_delay_slot_filling) {
2593 nop();
2594 }
2595 }
2596
2597
2598 // compiler_lock_object() and compiler_unlock_object() are direct transliterations
2599 // of i486.ad fast_lock() and fast_unlock(). See those methods for detailed comments.
2600 // The code could be tightened up considerably.
2601 //
2602 // box->dhw disposition - post-conditions at DONE_LABEL.
2603 // - Successful inflated lock: box->dhw != 0.
2604 // Any non-zero value suffices.
2605 // Consider G2_thread, rsp, boxReg, or markOopDesc::unused_mark()
2606 // - Successful Stack-lock: box->dhw == mark.
2607 // box->dhw must contain the displaced mark word value
2608 // - Failure -- icc.ZFlag == 0 and box->dhw is undefined.
2609 // The slow-path fast_enter() and slow_enter() operators
2610 // are responsible for setting box->dhw = NonZero (typically markOopDesc::unused_mark()).
2611 // - Biased: box->dhw is undefined
2612 //
2613 // SPARC refworkload performance - specifically jetstream and scimark - are
2614 // extremely sensitive to the size of the code emitted by compiler_lock_object
2615 // and compiler_unlock_object. Critically, the key factor is code size, not path
2616 // length. (Simply experiments to pad CLO with unexecuted NOPs demonstrte the
2617 // effect).
2618
2619
2620 void MacroAssembler::compiler_lock_object(Register Roop, Register Rmark,
2621 Register Rbox, Register Rscratch,
2622 BiasedLockingCounters* counters,
2623 bool try_bias) {
2624 Address mark_addr(Roop, oopDesc::mark_offset_in_bytes());
2625
2626 verify_oop(Roop);
2627 Label done ;
2628
2629 if (counters != NULL) {
2630 inc_counter((address) counters->total_entry_count_addr(), Rmark, Rscratch);
2631 }
2632
2633 if (EmitSync & 1) {
2634 mov(3, Rscratch);
2635 st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2636 cmp(SP, G0);
2637 return ;
2638 }
2639
2640 if (EmitSync & 2) {
2641
2642 // Fetch object's markword
2643 ld_ptr(mark_addr, Rmark);
2644
2645 if (try_bias) {
2646 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters);
2647 }
2648
2649 // Save Rbox in Rscratch to be used for the cas operation
2650 mov(Rbox, Rscratch);
2651
2652 // set Rmark to markOop | markOopDesc::unlocked_value
2653 or3(Rmark, markOopDesc::unlocked_value, Rmark);
2654
2655 // Initialize the box. (Must happen before we update the object mark!)
2656 st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes());
2657
2658 // compare object markOop with Rmark and if equal exchange Rscratch with object markOop
2659 assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2660 cas_ptr(mark_addr.base(), Rmark, Rscratch);
2661
2662 // if compare/exchange succeeded we found an unlocked object and we now have locked it
2663 // hence we are done
2664 cmp(Rmark, Rscratch);
2665 sub(Rscratch, STACK_BIAS, Rscratch);
2666 brx(Assembler::equal, false, Assembler::pt, done);
2667 delayed()->sub(Rscratch, SP, Rscratch); //pull next instruction into delay slot
2668
2669 // we did not find an unlocked object so see if this is a recursive case
2670 // sub(Rscratch, SP, Rscratch);
2671 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant");
2672 andcc(Rscratch, 0xfffff003, Rscratch);
2673 st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2674 bind (done);
2675 return ;
2676 }
2677
2678 Label Egress ;
2679
2680 if (EmitSync & 256) {
2681 Label IsInflated ;
2682
2683 ld_ptr(mark_addr, Rmark); // fetch obj->mark
2684 // Triage: biased, stack-locked, neutral, inflated
2685 if (try_bias) {
2686 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters);
2687 // Invariant: if control reaches this point in the emitted stream
2688 // then Rmark has not been modified.
2689 }
2690
2691 // Store mark into displaced mark field in the on-stack basic-lock "box"
2692 // Critically, this must happen before the CAS
2693 // Maximize the ST-CAS distance to minimize the ST-before-CAS penalty.
2694 st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes());
2695 andcc(Rmark, 2, G0);
2696 brx(Assembler::notZero, false, Assembler::pn, IsInflated);
2697 delayed()->
2698
2699 // Try stack-lock acquisition.
2700 // Beware: the 1st instruction is in a delay slot
2701 mov(Rbox, Rscratch);
2702 or3(Rmark, markOopDesc::unlocked_value, Rmark);
2703 assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2704 cas_ptr(mark_addr.base(), Rmark, Rscratch);
2705 cmp(Rmark, Rscratch);
2706 brx(Assembler::equal, false, Assembler::pt, done);
2707 delayed()->sub(Rscratch, SP, Rscratch);
2708
2709 // Stack-lock attempt failed - check for recursive stack-lock.
2710 // See the comments below about how we might remove this case.
2711 sub(Rscratch, STACK_BIAS, Rscratch);
2712 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant");
2713 andcc(Rscratch, 0xfffff003, Rscratch);
2714 br(Assembler::always, false, Assembler::pt, done);
2715 delayed()-> st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2716
2717 bind(IsInflated);
2718 if (EmitSync & 64) {
2719 // If m->owner != null goto IsLocked
2720 // Pessimistic form: Test-and-CAS vs CAS
2721 // The optimistic form avoids RTS->RTO cache line upgrades.
2722 ld_ptr(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rscratch);
2723 andcc(Rscratch, Rscratch, G0);
2724 brx(Assembler::notZero, false, Assembler::pn, done);
2725 delayed()->nop();
2726 // m->owner == null : it's unlocked.
2727 }
2728
2729 // Try to CAS m->owner from null to Self
2730 // Invariant: if we acquire the lock then _recursions should be 0.
2731 add(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rmark);
2732 mov(G2_thread, Rscratch);
2733 cas_ptr(Rmark, G0, Rscratch);
2734 cmp(Rscratch, G0);
2735 // Intentional fall-through into done
2736 } else {
2737 // Aggressively avoid the Store-before-CAS penalty
2738 // Defer the store into box->dhw until after the CAS
2739 Label IsInflated, Recursive ;
2740
2741 // Anticipate CAS -- Avoid RTS->RTO upgrade
2742 // prefetch (mark_addr, Assembler::severalWritesAndPossiblyReads);
2743
2744 ld_ptr(mark_addr, Rmark); // fetch obj->mark
2745 // Triage: biased, stack-locked, neutral, inflated
2746
2747 if (try_bias) {
2748 biased_locking_enter(Roop, Rmark, Rscratch, done, NULL, counters);
2749 // Invariant: if control reaches this point in the emitted stream
2750 // then Rmark has not been modified.
2751 }
2752 andcc(Rmark, 2, G0);
2753 brx(Assembler::notZero, false, Assembler::pn, IsInflated);
2754 delayed()-> // Beware - dangling delay-slot
2755
2756 // Try stack-lock acquisition.
2757 // Transiently install BUSY (0) encoding in the mark word.
2758 // if the CAS of 0 into the mark was successful then we execute:
2759 // ST box->dhw = mark -- save fetched mark in on-stack basiclock box
2760 // ST obj->mark = box -- overwrite transient 0 value
2761 // This presumes TSO, of course.
2762
2763 mov(0, Rscratch);
2764 or3(Rmark, markOopDesc::unlocked_value, Rmark);
2765 assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2766 cas_ptr(mark_addr.base(), Rmark, Rscratch);
2767 // prefetch (mark_addr, Assembler::severalWritesAndPossiblyReads);
2768 cmp(Rscratch, Rmark);
2769 brx(Assembler::notZero, false, Assembler::pn, Recursive);
2770 delayed()->st_ptr(Rmark, Rbox, BasicLock::displaced_header_offset_in_bytes());
2771 if (counters != NULL) {
2772 cond_inc(Assembler::equal, (address) counters->fast_path_entry_count_addr(), Rmark, Rscratch);
2773 }
2774 ba(done);
2775 delayed()->st_ptr(Rbox, mark_addr);
2776
2777 bind(Recursive);
2778 // Stack-lock attempt failed - check for recursive stack-lock.
2779 // Tests show that we can remove the recursive case with no impact
2780 // on refworkload 0.83. If we need to reduce the size of the code
2781 // emitted by compiler_lock_object() the recursive case is perfect
2782 // candidate.
2783 //
2784 // A more extreme idea is to always inflate on stack-lock recursion.
2785 // This lets us eliminate the recursive checks in compiler_lock_object
2786 // and compiler_unlock_object and the (box->dhw == 0) encoding.
2787 // A brief experiment - requiring changes to synchronizer.cpp, interpreter,
2788 // and showed a performance *increase*. In the same experiment I eliminated
2789 // the fast-path stack-lock code from the interpreter and always passed
2790 // control to the "slow" operators in synchronizer.cpp.
2791
2792 // RScratch contains the fetched obj->mark value from the failed CAS.
2793 sub(Rscratch, STACK_BIAS, Rscratch);
2794 sub(Rscratch, SP, Rscratch);
2795 assert(os::vm_page_size() > 0xfff, "page size too small - change the constant");
2796 andcc(Rscratch, 0xfffff003, Rscratch);
2797 if (counters != NULL) {
2798 // Accounting needs the Rscratch register
2799 st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2800 cond_inc(Assembler::equal, (address) counters->fast_path_entry_count_addr(), Rmark, Rscratch);
2801 ba_short(done);
2802 } else {
2803 ba(done);
2804 delayed()->st_ptr(Rscratch, Rbox, BasicLock::displaced_header_offset_in_bytes());
2805 }
2806
2807 bind (IsInflated);
2808
2809 // Try to CAS m->owner from null to Self
2810 // Invariant: if we acquire the lock then _recursions should be 0.
2811 add(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rmark);
2812 mov(G2_thread, Rscratch);
2813 cas_ptr(Rmark, G0, Rscratch);
2814 andcc(Rscratch, Rscratch, G0); // set ICCs for done: icc.zf iff success
2815 // set icc.zf : 1=success 0=failure
2816 // ST box->displaced_header = NonZero.
2817 // Any non-zero value suffices:
2818 // markOopDesc::unused_mark(), G2_thread, RBox, RScratch, rsp, etc.
2819 st_ptr(Rbox, Rbox, BasicLock::displaced_header_offset_in_bytes());
2820 // Intentional fall-through into done
2821 }
2822
2823 bind (done);
2824 }
2825
2826 void MacroAssembler::compiler_unlock_object(Register Roop, Register Rmark,
2827 Register Rbox, Register Rscratch,
2828 bool try_bias) {
2829 Address mark_addr(Roop, oopDesc::mark_offset_in_bytes());
2830
2831 Label done ;
2832
2833 if (EmitSync & 4) {
2834 cmp(SP, G0);
2835 return ;
2836 }
2837
2838 if (EmitSync & 8) {
2839 if (try_bias) {
2840 biased_locking_exit(mark_addr, Rscratch, done);
2841 }
2842
2843 // Test first if it is a fast recursive unlock
2844 ld_ptr(Rbox, BasicLock::displaced_header_offset_in_bytes(), Rmark);
2845 br_null_short(Rmark, Assembler::pt, done);
2846
2847 // Check if it is still a light weight lock, this is is true if we see
2848 // the stack address of the basicLock in the markOop of the object
2849 assert(mark_addr.disp() == 0, "cas must take a zero displacement");
2850 cas_ptr(mark_addr.base(), Rbox, Rmark);
2851 ba(done);
2852 delayed()->cmp(Rbox, Rmark);
2853 bind(done);
2854 return ;
2855 }
2856
2857 // Beware ... If the aggregate size of the code emitted by CLO and CUO is
2858 // is too large performance rolls abruptly off a cliff.
2859 // This could be related to inlining policies, code cache management, or
2860 // I$ effects.
2861 Label LStacked ;
2862
2863 if (try_bias) {
2864 // TODO: eliminate redundant LDs of obj->mark
2865 biased_locking_exit(mark_addr, Rscratch, done);
2866 }
2867
2868 ld_ptr(Roop, oopDesc::mark_offset_in_bytes(), Rmark);
2869 ld_ptr(Rbox, BasicLock::displaced_header_offset_in_bytes(), Rscratch);
2870 andcc(Rscratch, Rscratch, G0);
2871 brx(Assembler::zero, false, Assembler::pn, done);
2872 delayed()->nop(); // consider: relocate fetch of mark, above, into this DS
2873 andcc(Rmark, 2, G0);
2874 brx(Assembler::zero, false, Assembler::pt, LStacked);
2875 delayed()->nop();
2876
2877 // It's inflated
2878 // Conceptually we need a #loadstore|#storestore "release" MEMBAR before
2879 // the ST of 0 into _owner which releases the lock. This prevents loads
2880 // and stores within the critical section from reordering (floating)
2881 // past the store that releases the lock. But TSO is a strong memory model
2882 // and that particular flavor of barrier is a noop, so we can safely elide it.
2883 // Note that we use 1-0 locking by default for the inflated case. We
2884 // close the resultant (and rare) race by having contended threads in
2885 // monitorenter periodically poll _owner.
2886
2887 if (EmitSync & 1024) {
2888 // Emit code to check that _owner == Self
2889 // We could fold the _owner test into subsequent code more efficiently
2890 // than using a stand-alone check, but since _owner checking is off by
2891 // default we don't bother. We also might consider predicating the
2892 // _owner==Self check on Xcheck:jni or running on a debug build.
2893 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)), Rscratch);
2894 orcc(Rscratch, G0, G0);
2895 brx(Assembler::notZero, false, Assembler::pn, done);
2896 delayed()->nop();
2897 }
2898
2899 if (EmitSync & 512) {
2900 // classic lock release code absent 1-0 locking
2901 // m->Owner = null;
2902 // membar #storeload
2903 // if (m->cxq|m->EntryList) == null goto Success
2904 // if (m->succ != null) goto Success
2905 // if CAS (&m->Owner,0,Self) != 0 goto Success
2906 // goto SlowPath
2907 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)), Rbox);
2908 orcc(Rbox, G0, G0);
2909 brx(Assembler::notZero, false, Assembler::pn, done);
2910 delayed()->nop();
2911 st_ptr(G0, Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2912 if (os::is_MP()) { membar(StoreLoad); }
2913 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)), Rscratch);
2914 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)), Rbox);
2915 orcc(Rbox, Rscratch, G0);
2916 brx(Assembler::zero, false, Assembler::pt, done);
2917 delayed()->
2918 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), Rscratch);
2919 andcc(Rscratch, Rscratch, G0);
2920 brx(Assembler::notZero, false, Assembler::pt, done);
2921 delayed()->andcc(G0, G0, G0);
2922 add(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rmark);
2923 mov(G2_thread, Rscratch);
2924 cas_ptr(Rmark, G0, Rscratch);
2925 cmp(Rscratch, G0);
2926 // invert icc.zf and goto done
2927 brx(Assembler::notZero, false, Assembler::pt, done);
2928 delayed()->cmp(G0, G0);
2929 br(Assembler::always, false, Assembler::pt, done);
2930 delayed()->cmp(G0, 1);
2931 } else {
2932 // 1-0 form : avoids CAS and MEMBAR in the common case
2933 // Do not bother to ratify that m->Owner == Self.
2934 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)), Rbox);
2935 orcc(Rbox, G0, G0);
2936 brx(Assembler::notZero, false, Assembler::pn, done);
2937 delayed()->
2938 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)), Rscratch);
2939 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)), Rbox);
2940 orcc(Rbox, Rscratch, G0);
2941 if (EmitSync & 16384) {
2942 // As an optional optimization, if (EntryList|cxq) != null and _succ is null then
2943 // we should transfer control directly to the slow-path.
2944 // This test makes the reacquire operation below very infrequent.
2945 // The logic is equivalent to :
2946 // if (cxq|EntryList) == null : Owner=null; goto Success
2947 // if succ == null : goto SlowPath
2948 // Owner=null; membar #storeload
2949 // if succ != null : goto Success
2950 // if CAS(&Owner,null,Self) != null goto Success
2951 // goto SlowPath
2952 brx(Assembler::zero, true, Assembler::pt, done);
2953 delayed()->
2954 st_ptr(G0, Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2955 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), Rscratch);
2956 andcc(Rscratch, Rscratch, G0) ;
2957 brx(Assembler::zero, false, Assembler::pt, done);
2958 delayed()->orcc(G0, 1, G0);
2959 st_ptr(G0, Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2960 } else {
2961 brx(Assembler::zero, false, Assembler::pt, done);
2962 delayed()->
2963 st_ptr(G0, Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
2964 }
2965 if (os::is_MP()) { membar(StoreLoad); }
2966 // Check that _succ is (or remains) non-zero
2967 ld_ptr(Address(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(succ)), Rscratch);
2968 andcc(Rscratch, Rscratch, G0);
2969 brx(Assembler::notZero, false, Assembler::pt, done);
2970 delayed()->andcc(G0, G0, G0);
2971 add(Rmark, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), Rmark);
2972 mov(G2_thread, Rscratch);
2973 cas_ptr(Rmark, G0, Rscratch);
2974 cmp(Rscratch, G0);
2975 // invert icc.zf and goto done
2976 // A slightly better v8+/v9 idiom would be the following:
2977 // movrnz Rscratch,1,Rscratch
2978 // ba done
2979 // xorcc Rscratch,1,G0
2980 // In v8+ mode the idiom would be valid IFF Rscratch was a G or O register
2981 brx(Assembler::notZero, false, Assembler::pt, done);
2982 delayed()->cmp(G0, G0);
2983 br(Assembler::always, false, Assembler::pt, done);
2984 delayed()->cmp(G0, 1);
2985 }
2986
2987 bind (LStacked);
2988 // Consider: we could replace the expensive CAS in the exit
2989 // path with a simple ST of the displaced mark value fetched from
2990 // the on-stack basiclock box. That admits a race where a thread T2
2991 // in the slow lock path -- inflating with monitor M -- could race a
2992 // thread T1 in the fast unlock path, resulting in a missed wakeup for T2.
2993 // More precisely T1 in the stack-lock unlock path could "stomp" the
2994 // inflated mark value M installed by T2, resulting in an orphan
2995 // object monitor M and T2 becoming stranded. We can remedy that situation
2996 // by having T2 periodically poll the object's mark word using timed wait
2997 // operations. If T2 discovers that a stomp has occurred it vacates
2998 // the monitor M and wakes any other threads stranded on the now-orphan M.
2999 // In addition the monitor scavenger, which performs deflation,
3000 // would also need to check for orpan monitors and stranded threads.
3001 //
3002 // Finally, inflation is also used when T2 needs to assign a hashCode
3003 // to O and O is stack-locked by T1. The "stomp" race could cause
3004 // an assigned hashCode value to be lost. We can avoid that condition
3005 // and provide the necessary hashCode stability invariants by ensuring
3006 // that hashCode generation is idempotent between copying GCs.
3007 // For example we could compute the hashCode of an object O as
3008 // O's heap address XOR some high quality RNG value that is refreshed
3009 // at GC-time. The monitor scavenger would install the hashCode
3010 // found in any orphan monitors. Again, the mechanism admits a
3011 // lost-update "stomp" WAW race but detects and recovers as needed.
3012 //
3013 // A prototype implementation showed excellent results, although
3014 // the scavenger and timeout code was rather involved.
3015
3016 cas_ptr(mark_addr.base(), Rbox, Rscratch);
3017 cmp(Rbox, Rscratch);
3018 // Intentional fall through into done ...
3019
3020 bind(done);
3021 }
3022
3023
3024
3025 void MacroAssembler::print_CPU_state() {
3026 // %%%%% need to implement this
3027 }
3028
3029 void MacroAssembler::verify_FPU(int stack_depth, const char* s) {
3030 // %%%%% need to implement this
3031 }
3032
3033 void MacroAssembler::push_IU_state() {
3034 // %%%%% need to implement this
3035 }
3036
3037
3038 void MacroAssembler::pop_IU_state() {
3039 // %%%%% need to implement this
3040 }
3041
3042
3043 void MacroAssembler::push_FPU_state() {
3044 // %%%%% need to implement this
3045 }
3046
3047
3048 void MacroAssembler::pop_FPU_state() {
3049 // %%%%% need to implement this
3050 }
3051
3052
3053 void MacroAssembler::push_CPU_state() {
3054 // %%%%% need to implement this
3055 }
3056
3057
3058 void MacroAssembler::pop_CPU_state() {
3059 // %%%%% need to implement this
3060 }
3061
3062
3063
3064 void MacroAssembler::verify_tlab() {
3065 #ifdef ASSERT
3066 if (UseTLAB && VerifyOops) {
3067 Label next, next2, ok;
3068 Register t1 = L0;
3069 Register t2 = L1;
3070 Register t3 = L2;
3071
3072 save_frame(0);
3073 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), t1);
3074 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t2);
3075 or3(t1, t2, t3);
3076 cmp_and_br_short(t1, t2, Assembler::greaterEqual, Assembler::pn, next);
3077 STOP("assert(top >= start)");
3078 should_not_reach_here();
3079
3080 bind(next);
3081 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), t1);
3082 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), t2);
3083 or3(t3, t2, t3);
3084 cmp_and_br_short(t1, t2, Assembler::lessEqual, Assembler::pn, next2);
3085 STOP("assert(top <= end)");
3086 should_not_reach_here();
3087
3088 bind(next2);
3089 and3(t3, MinObjAlignmentInBytesMask, t3);
3090 cmp_and_br_short(t3, 0, Assembler::lessEqual, Assembler::pn, ok);
3091 STOP("assert(aligned)");
3092 should_not_reach_here();
3093
3094 bind(ok);
3095 restore();
3096 }
3097 #endif
3098 }
3099
3100
3101 void MacroAssembler::eden_allocate(
3102 Register obj, // result: pointer to object after successful allocation
3103 Register var_size_in_bytes, // object size in bytes if unknown at compile time; invalid otherwise
3104 int con_size_in_bytes, // object size in bytes if known at compile time
3105 Register t1, // temp register
3106 Register t2, // temp register
3107 Label& slow_case // continuation point if fast allocation fails
3108 ){
3109 // make sure arguments make sense
3110 assert_different_registers(obj, var_size_in_bytes, t1, t2);
3111 assert(0 <= con_size_in_bytes && Assembler::is_simm13(con_size_in_bytes), "illegal object size");
3112 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment");
3113
3114 if (!Universe::heap()->supports_inline_contig_alloc()) {
3115 // No allocation in the shared eden.
3116 ba(slow_case);
3117 delayed()->nop();
3118 } else {
3119 // get eden boundaries
3120 // note: we need both top & top_addr!
3121 const Register top_addr = t1;
3122 const Register end = t2;
3123
3124 CollectedHeap* ch = Universe::heap();
3125 set((intx)ch->top_addr(), top_addr);
3126 intx delta = (intx)ch->end_addr() - (intx)ch->top_addr();
3127 ld_ptr(top_addr, delta, end);
3128 ld_ptr(top_addr, 0, obj);
3129
3130 // try to allocate
3131 Label retry;
3132 bind(retry);
3133 #ifdef ASSERT
3134 // make sure eden top is properly aligned
3135 {
3136 Label L;
3137 btst(MinObjAlignmentInBytesMask, obj);
3138 br(Assembler::zero, false, Assembler::pt, L);
3139 delayed()->nop();
3140 STOP("eden top is not properly aligned");
3141 bind(L);
3142 }
3143 #endif // ASSERT
3144 const Register free = end;
3145 sub(end, obj, free); // compute amount of free space
3146 if (var_size_in_bytes->is_valid()) {
3147 // size is unknown at compile time
3148 cmp(free, var_size_in_bytes);
3149 brx(Assembler::lessUnsigned, false, Assembler::pn, slow_case); // if there is not enough space go the slow case
3150 delayed()->add(obj, var_size_in_bytes, end);
3151 } else {
3152 // size is known at compile time
3153 cmp(free, con_size_in_bytes);
3154 brx(Assembler::lessUnsigned, false, Assembler::pn, slow_case); // if there is not enough space go the slow case
3155 delayed()->add(obj, con_size_in_bytes, end);
3156 }
3157 // Compare obj with the value at top_addr; if still equal, swap the value of
3158 // end with the value at top_addr. If not equal, read the value at top_addr
3159 // into end.
3160 cas_ptr(top_addr, obj, end);
3161 // if someone beat us on the allocation, try again, otherwise continue
3162 cmp(obj, end);
3163 brx(Assembler::notEqual, false, Assembler::pn, retry);
3164 delayed()->mov(end, obj); // nop if successfull since obj == end
3165
3166 #ifdef ASSERT
3167 // make sure eden top is properly aligned
3168 {
3169 Label L;
3170 const Register top_addr = t1;
3171
3172 set((intx)ch->top_addr(), top_addr);
3173 ld_ptr(top_addr, 0, top_addr);
3174 btst(MinObjAlignmentInBytesMask, top_addr);
3175 br(Assembler::zero, false, Assembler::pt, L);
3176 delayed()->nop();
3177 STOP("eden top is not properly aligned");
3178 bind(L);
3179 }
3180 #endif // ASSERT
3181 }
3182 }
3183
3184
3185 void MacroAssembler::tlab_allocate(
3186 Register obj, // result: pointer to object after successful allocation
3187 Register var_size_in_bytes, // object size in bytes if unknown at compile time; invalid otherwise
3188 int con_size_in_bytes, // object size in bytes if known at compile time
3189 Register t1, // temp register
3190 Label& slow_case // continuation point if fast allocation fails
3191 ){
3192 // make sure arguments make sense
3193 assert_different_registers(obj, var_size_in_bytes, t1);
3194 assert(0 <= con_size_in_bytes && is_simm13(con_size_in_bytes), "illegal object size");
3195 assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0, "object size is not multiple of alignment");
3196
3197 const Register free = t1;
3198
3199 verify_tlab();
3200
3201 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), obj);
3202
3203 // calculate amount of free space
3204 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), free);
3205 sub(free, obj, free);
3206
3207 Label done;
3208 if (var_size_in_bytes == noreg) {
3209 cmp(free, con_size_in_bytes);
3210 } else {
3211 cmp(free, var_size_in_bytes);
3212 }
3213 br(Assembler::less, false, Assembler::pn, slow_case);
3214 // calculate the new top pointer
3215 if (var_size_in_bytes == noreg) {
3216 delayed()->add(obj, con_size_in_bytes, free);
3217 } else {
3218 delayed()->add(obj, var_size_in_bytes, free);
3219 }
3220
3221 bind(done);
3222
3223 #ifdef ASSERT
3224 // make sure new free pointer is properly aligned
3225 {
3226 Label L;
3227 btst(MinObjAlignmentInBytesMask, free);
3228 br(Assembler::zero, false, Assembler::pt, L);
3229 delayed()->nop();
3230 STOP("updated TLAB free is not properly aligned");
3231 bind(L);
3232 }
3233 #endif // ASSERT
3234
3235 // update the tlab top pointer
3236 st_ptr(free, G2_thread, in_bytes(JavaThread::tlab_top_offset()));
3237 verify_tlab();
3238 }
3239
3240
3241 void MacroAssembler::tlab_refill(Label& retry, Label& try_eden, Label& slow_case) {
3242 Register top = O0;
3243 Register t1 = G1;
3244 Register t2 = G3;
3245 Register t3 = O1;
3246 assert_different_registers(top, t1, t2, t3, G4, G5 /* preserve G4 and G5 */);
3247 Label do_refill, discard_tlab;
3248
3249 if (!Universe::heap()->supports_inline_contig_alloc()) {
3250 // No allocation in the shared eden.
3251 ba(slow_case);
3252 delayed()->nop();
3253 }
3254
3255 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_top_offset()), top);
3256 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_end_offset()), t1);
3257 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_refill_waste_limit_offset()), t2);
3258
3259 // calculate amount of free space
3260 sub(t1, top, t1);
3261 srl_ptr(t1, LogHeapWordSize, t1);
3262
3263 // Retain tlab and allocate object in shared space if
3264 // the amount free in the tlab is too large to discard.
3265 cmp(t1, t2);
3266
3267 brx(Assembler::lessEqual, false, Assembler::pt, discard_tlab);
3268 // increment waste limit to prevent getting stuck on this slow path
3269 if (Assembler::is_simm13(ThreadLocalAllocBuffer::refill_waste_limit_increment())) {
3270 delayed()->add(t2, ThreadLocalAllocBuffer::refill_waste_limit_increment(), t2);
3271 } else {
3272 delayed()->nop();
3273 // set64 does not use the temp register if the given constant is 32 bit. So
3274 // we can just use any register; using G0 results in ignoring of the upper 32 bit
3275 // of that value.
3276 set64(ThreadLocalAllocBuffer::refill_waste_limit_increment(), t3, G0);
3277 add(t2, t3, t2);
3278 }
3279
3280 st_ptr(t2, G2_thread, in_bytes(JavaThread::tlab_refill_waste_limit_offset()));
3281 if (TLABStats) {
3282 // increment number of slow_allocations
3283 ld(G2_thread, in_bytes(JavaThread::tlab_slow_allocations_offset()), t2);
3284 add(t2, 1, t2);
3285 stw(t2, G2_thread, in_bytes(JavaThread::tlab_slow_allocations_offset()));
3286 }
3287 ba(try_eden);
3288 delayed()->nop();
3289
3290 bind(discard_tlab);
3291 if (TLABStats) {
3292 // increment number of refills
3293 ld(G2_thread, in_bytes(JavaThread::tlab_number_of_refills_offset()), t2);
3294 add(t2, 1, t2);
3295 stw(t2, G2_thread, in_bytes(JavaThread::tlab_number_of_refills_offset()));
3296 // accumulate wastage
3297 ld(G2_thread, in_bytes(JavaThread::tlab_fast_refill_waste_offset()), t2);
3298 add(t2, t1, t2);
3299 stw(t2, G2_thread, in_bytes(JavaThread::tlab_fast_refill_waste_offset()));
3300 }
3301
3302 // if tlab is currently allocated (top or end != null) then
3303 // fill [top, end + alignment_reserve) with array object
3304 br_null_short(top, Assembler::pn, do_refill);
3305
3306 set((intptr_t)markOopDesc::prototype()->copy_set_hash(0x2), t2);
3307 st_ptr(t2, top, oopDesc::mark_offset_in_bytes()); // set up the mark word
3308 // set klass to intArrayKlass
3309 sub(t1, typeArrayOopDesc::header_size(T_INT), t1);
3310 add(t1, ThreadLocalAllocBuffer::alignment_reserve(), t1);
3311 sll_ptr(t1, log2_intptr(HeapWordSize/sizeof(jint)), t1);
3312 st(t1, top, arrayOopDesc::length_offset_in_bytes());
3313 set((intptr_t)Universe::intArrayKlassObj_addr(), t2);
3314 ld_ptr(t2, 0, t2);
3315 // store klass last. concurrent gcs assumes klass length is valid if
3316 // klass field is not null.
3317 store_klass(t2, top);
3318 verify_oop(top);
3319
3320 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t1);
3321 sub(top, t1, t1); // size of tlab's allocated portion
3322 incr_allocated_bytes(t1, t2, t3);
3323
3324 // refill the tlab with an eden allocation
3325 bind(do_refill);
3326 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_size_offset()), t1);
3327 sll_ptr(t1, LogHeapWordSize, t1);
3328 // allocate new tlab, address returned in top
3329 eden_allocate(top, t1, 0, t2, t3, slow_case);
3330
3331 st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_start_offset()));
3332 st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_top_offset()));
3333 #ifdef ASSERT
3334 // check that tlab_size (t1) is still valid
3335 {
3336 Label ok;
3337 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_size_offset()), t2);
3338 sll_ptr(t2, LogHeapWordSize, t2);
3339 cmp_and_br_short(t1, t2, Assembler::equal, Assembler::pt, ok);
3340 STOP("assert(t1 == tlab_size)");
3341 should_not_reach_here();
3342
3343 bind(ok);
3344 }
3345 #endif // ASSERT
3346 add(top, t1, top); // t1 is tlab_size
3347 sub(top, ThreadLocalAllocBuffer::alignment_reserve_in_bytes(), top);
3348 st_ptr(top, G2_thread, in_bytes(JavaThread::tlab_end_offset()));
3349
3350 if (ZeroTLAB) {
3351 // This is a fast TLAB refill, therefore the GC is not notified of it.
3352 // So compiled code must fill the new TLAB with zeroes.
3353 ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t2);
3354 zero_memory(t2, t1);
3355 }
3356 verify_tlab();
3357 ba(retry);
3358 delayed()->nop();
3359 }
3360
3361 void MacroAssembler::zero_memory(Register base, Register index) {
3362 assert_different_registers(base, index);
3363 Label loop;
3364 bind(loop);
3365 subcc(index, HeapWordSize, index);
3366 brx(Assembler::greaterEqual, true, Assembler::pt, loop);
3367 delayed()->st_ptr(G0, base, index);
3368 }
3369
3370 void MacroAssembler::incr_allocated_bytes(RegisterOrConstant size_in_bytes,
3371 Register t1, Register t2) {
3372 // Bump total bytes allocated by this thread
3373 assert(t1->is_global(), "must be global reg"); // so all 64 bits are saved on a context switch
3374 assert_different_registers(size_in_bytes.register_or_noreg(), t1, t2);
3375 // v8 support has gone the way of the dodo
3376 ldx(G2_thread, in_bytes(JavaThread::allocated_bytes_offset()), t1);
3377 add(t1, ensure_simm13_or_reg(size_in_bytes, t2), t1);
3378 stx(t1, G2_thread, in_bytes(JavaThread::allocated_bytes_offset()));
3379 }
3380
3381 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) {
3382 switch (cond) {
3383 // Note some conditions are synonyms for others
3384 case Assembler::never: return Assembler::always;
3385 case Assembler::zero: return Assembler::notZero;
3386 case Assembler::lessEqual: return Assembler::greater;
3387 case Assembler::less: return Assembler::greaterEqual;
3388 case Assembler::lessEqualUnsigned: return Assembler::greaterUnsigned;
3389 case Assembler::lessUnsigned: return Assembler::greaterEqualUnsigned;
3390 case Assembler::negative: return Assembler::positive;
3391 case Assembler::overflowSet: return Assembler::overflowClear;
3392 case Assembler::always: return Assembler::never;
3393 case Assembler::notZero: return Assembler::zero;
3394 case Assembler::greater: return Assembler::lessEqual;
3395 case Assembler::greaterEqual: return Assembler::less;
3396 case Assembler::greaterUnsigned: return Assembler::lessEqualUnsigned;
3397 case Assembler::greaterEqualUnsigned: return Assembler::lessUnsigned;
3398 case Assembler::positive: return Assembler::negative;
3399 case Assembler::overflowClear: return Assembler::overflowSet;
3400 }
3401
3402 ShouldNotReachHere(); return Assembler::overflowClear;
3403 }
3404
3405 void MacroAssembler::cond_inc(Assembler::Condition cond, address counter_ptr,
3406 Register Rtmp1, Register Rtmp2 /*, Register Rtmp3, Register Rtmp4 */) {
3407 Condition negated_cond = negate_condition(cond);
3408 Label L;
3409 brx(negated_cond, false, Assembler::pt, L);
3410 delayed()->nop();
3411 inc_counter(counter_ptr, Rtmp1, Rtmp2);
3412 bind(L);
3413 }
3414
3415 void MacroAssembler::inc_counter(address counter_addr, Register Rtmp1, Register Rtmp2) {
3416 AddressLiteral addrlit(counter_addr);
3417 sethi(addrlit, Rtmp1); // Move hi22 bits into temporary register.
3418 Address addr(Rtmp1, addrlit.low10()); // Build an address with low10 bits.
3419 ld(addr, Rtmp2);
3420 inc(Rtmp2);
3421 st(Rtmp2, addr);
3422 }
3423
3424 void MacroAssembler::inc_counter(int* counter_addr, Register Rtmp1, Register Rtmp2) {
3425 inc_counter((address) counter_addr, Rtmp1, Rtmp2);
3426 }
3427
3428 SkipIfEqual::SkipIfEqual(
3429 MacroAssembler* masm, Register temp, const bool* flag_addr,
3430 Assembler::Condition condition) {
3431 _masm = masm;
3432 AddressLiteral flag(flag_addr);
3433 _masm->sethi(flag, temp);
3434 _masm->ldub(temp, flag.low10(), temp);
3435 _masm->tst(temp);
3436 _masm->br(condition, false, Assembler::pt, _label);
3437 _masm->delayed()->nop();
3438 }
3439
3440 SkipIfEqual::~SkipIfEqual() {
3441 _masm->bind(_label);
3442 }
3443
3444
3445 // Writes to stack successive pages until offset reached to check for
3446 // stack overflow + shadow pages. This clobbers tsp and scratch.
3447 void MacroAssembler::bang_stack_size(Register Rsize, Register Rtsp,
3448 Register Rscratch) {
3449 // Use stack pointer in temp stack pointer
3450 mov(SP, Rtsp);
3451
3452 // Bang stack for total size given plus stack shadow page size.
3453 // Bang one page at a time because a large size can overflow yellow and
3454 // red zones (the bang will fail but stack overflow handling can't tell that
3455 // it was a stack overflow bang vs a regular segv).
3456 int offset = os::vm_page_size();
3457 Register Roffset = Rscratch;
3458
3459 Label loop;
3460 bind(loop);
3461 set((-offset)+STACK_BIAS, Rscratch);
3462 st(G0, Rtsp, Rscratch);
3463 set(offset, Roffset);
3464 sub(Rsize, Roffset, Rsize);
3465 cmp(Rsize, G0);
3466 br(Assembler::greater, false, Assembler::pn, loop);
3467 delayed()->sub(Rtsp, Roffset, Rtsp);
3468
3469 // Bang down shadow pages too.
3470 // At this point, (tmp-0) is the last address touched, so don't
3471 // touch it again. (It was touched as (tmp-pagesize) but then tmp
3472 // was post-decremented.) Skip this address by starting at i=1, and
3473 // touch a few more pages below. N.B. It is important to touch all
3474 // the way down to and including i=StackShadowPages.
3475 for (int i = 1; i < JavaThread::stack_shadow_zone_size() / os::vm_page_size(); i++) {
3476 set((-i*offset)+STACK_BIAS, Rscratch);
3477 st(G0, Rtsp, Rscratch);
3478 }
3479 }
3480
3481 void MacroAssembler::reserved_stack_check() {
3482 // testing if reserved zone needs to be enabled
3483 Label no_reserved_zone_enabling;
3484
3485 ld_ptr(G2_thread, JavaThread::reserved_stack_activation_offset(), G4_scratch);
3486 cmp_and_brx_short(SP, G4_scratch, Assembler::lessUnsigned, Assembler::pt, no_reserved_zone_enabling);
3487
3488 call_VM_leaf(L0, CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), G2_thread);
3489
3490 AddressLiteral stub(StubRoutines::throw_delayed_StackOverflowError_entry());
3491 jump_to(stub, G4_scratch);
3492 delayed()->restore();
3493
3494 should_not_reach_here();
3495
3496 bind(no_reserved_zone_enabling);
3497 }
3498
3499 ///////////////////////////////////////////////////////////////////////////////////
3500 #if INCLUDE_ALL_GCS
3501
3502 static address satb_log_enqueue_with_frame = NULL;
3503 static u_char* satb_log_enqueue_with_frame_end = NULL;
3504
3505 static address satb_log_enqueue_frameless = NULL;
3506 static u_char* satb_log_enqueue_frameless_end = NULL;
3507
3508 static int EnqueueCodeSize = 128 DEBUG_ONLY( + 256); // Instructions?
3509
3510 static void generate_satb_log_enqueue(bool with_frame) {
3511 BufferBlob* bb = BufferBlob::create("enqueue_with_frame", EnqueueCodeSize);
3512 CodeBuffer buf(bb);
3513 MacroAssembler masm(&buf);
3514
3515 #define __ masm.
3516
3517 address start = __ pc();
3518 Register pre_val;
3519
3520 Label refill, restart;
3521 if (with_frame) {
3522 __ save_frame(0);
3523 pre_val = I0; // Was O0 before the save.
3524 } else {
3525 pre_val = O0;
3526 }
3527
3528 int satb_q_index_byte_offset =
3529 in_bytes(JavaThread::satb_mark_queue_offset() +
3530 SATBMarkQueue::byte_offset_of_index());
3531
3532 int satb_q_buf_byte_offset =
3533 in_bytes(JavaThread::satb_mark_queue_offset() +
3534 SATBMarkQueue::byte_offset_of_buf());
3535
3536 assert(in_bytes(SATBMarkQueue::byte_width_of_index()) == sizeof(intptr_t) &&
3537 in_bytes(SATBMarkQueue::byte_width_of_buf()) == sizeof(intptr_t),
3538 "check sizes in assembly below");
3539
3540 __ bind(restart);
3541
3542 // Load the index into the SATB buffer. SATBMarkQueue::_index is a size_t
3543 // so ld_ptr is appropriate.
3544 __ ld_ptr(G2_thread, satb_q_index_byte_offset, L0);
3545
3546 // index == 0?
3547 __ cmp_and_brx_short(L0, G0, Assembler::equal, Assembler::pn, refill);
3548
3549 __ ld_ptr(G2_thread, satb_q_buf_byte_offset, L1);
3550 __ sub(L0, oopSize, L0);
3551
3552 __ st_ptr(pre_val, L1, L0); // [_buf + index] := I0
3553 if (!with_frame) {
3554 // Use return-from-leaf
3555 __ retl();
3556 __ delayed()->st_ptr(L0, G2_thread, satb_q_index_byte_offset);
3557 } else {
3558 // Not delayed.
3559 __ st_ptr(L0, G2_thread, satb_q_index_byte_offset);
3560 }
3561 if (with_frame) {
3562 __ ret();
3563 __ delayed()->restore();
3564 }
3565 __ bind(refill);
3566
3567 address handle_zero =
3568 CAST_FROM_FN_PTR(address,
3569 &SATBMarkQueueSet::handle_zero_index_for_thread);
3570 // This should be rare enough that we can afford to save all the
3571 // scratch registers that the calling context might be using.
3572 __ mov(G1_scratch, L0);
3573 __ mov(G3_scratch, L1);
3574 __ mov(G4, L2);
3575 // We need the value of O0 above (for the write into the buffer), so we
3576 // save and restore it.
3577 __ mov(O0, L3);
3578 // Since the call will overwrite O7, we save and restore that, as well.
3579 __ mov(O7, L4);
3580 __ call_VM_leaf(L5, handle_zero, G2_thread);
3581 __ mov(L0, G1_scratch);
3582 __ mov(L1, G3_scratch);
3583 __ mov(L2, G4);
3584 __ mov(L3, O0);
3585 __ br(Assembler::always, /*annul*/false, Assembler::pt, restart);
3586 __ delayed()->mov(L4, O7);
3587
3588 if (with_frame) {
3589 satb_log_enqueue_with_frame = start;
3590 satb_log_enqueue_with_frame_end = __ pc();
3591 } else {
3592 satb_log_enqueue_frameless = start;
3593 satb_log_enqueue_frameless_end = __ pc();
3594 }
3595
3596 #undef __
3597 }
3598
3599 void MacroAssembler::g1_write_barrier_pre(Register obj,
3600 Register index,
3601 int offset,
3602 Register pre_val,
3603 Register tmp,
3604 bool preserve_o_regs) {
3605 Label filtered;
3606
3607 if (obj == noreg) {
3608 // We are not loading the previous value so make
3609 // sure that we don't trash the value in pre_val
3610 // with the code below.
3611 assert_different_registers(pre_val, tmp);
3612 } else {
3613 // We will be loading the previous value
3614 // in this code so...
3615 assert(offset == 0 || index == noreg, "choose one");
3616 assert(pre_val == noreg, "check this code");
3617 }
3618
3619 // Is marking active?
3620 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
3621 ld(G2,
3622 in_bytes(JavaThread::satb_mark_queue_offset() +
3623 SATBMarkQueue::byte_offset_of_active()),
3624 tmp);
3625 } else {
3626 guarantee(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1,
3627 "Assumption");
3628 ldsb(G2,
3629 in_bytes(JavaThread::satb_mark_queue_offset() +
3630 SATBMarkQueue::byte_offset_of_active()),
3631 tmp);
3632 }
3633
3634 // Is marking active?
3635 cmp_and_br_short(tmp, G0, Assembler::equal, Assembler::pt, filtered);
3636
3637 // Do we need to load the previous value?
3638 if (obj != noreg) {
3639 // Load the previous value...
3640 if (index == noreg) {
3641 if (Assembler::is_simm13(offset)) {
3642 load_heap_oop(obj, offset, tmp);
3643 } else {
3644 set(offset, tmp);
3645 load_heap_oop(obj, tmp, tmp);
3646 }
3647 } else {
3648 load_heap_oop(obj, index, tmp);
3649 }
3650 // Previous value has been loaded into tmp
3651 pre_val = tmp;
3652 }
3653
3654 assert(pre_val != noreg, "must have a real register");
3655
3656 // Is the previous value null?
3657 cmp_and_brx_short(pre_val, G0, Assembler::equal, Assembler::pt, filtered);
3658
3659 // OK, it's not filtered, so we'll need to call enqueue. In the normal
3660 // case, pre_val will be a scratch G-reg, but there are some cases in
3661 // which it's an O-reg. In the first case, do a normal call. In the
3662 // latter, do a save here and call the frameless version.
3663
3664 guarantee(pre_val->is_global() || pre_val->is_out(),
3665 "Or we need to think harder.");
3666
3667 if (pre_val->is_global() && !preserve_o_regs) {
3668 call(satb_log_enqueue_with_frame);
3669 delayed()->mov(pre_val, O0);
3670 } else {
3671 save_frame(0);
3672 call(satb_log_enqueue_frameless);
3673 delayed()->mov(pre_val->after_save(), O0);
3674 restore();
3675 }
3676
3677 bind(filtered);
3678 }
3679
3680 static address dirty_card_log_enqueue = 0;
3681 static u_char* dirty_card_log_enqueue_end = 0;
3682
3683 // This gets to assume that o0 contains the object address.
3684 static void generate_dirty_card_log_enqueue(jbyte* byte_map_base) {
3685 BufferBlob* bb = BufferBlob::create("dirty_card_enqueue", EnqueueCodeSize*2);
3686 CodeBuffer buf(bb);
3687 MacroAssembler masm(&buf);
3688 #define __ masm.
3689 address start = __ pc();
3690
3691 Label not_already_dirty, restart, refill, young_card;
3692
3693 __ srlx(O0, CardTableModRefBS::card_shift, O0);
3694 AddressLiteral addrlit(byte_map_base);
3695 __ set(addrlit, O1); // O1 := <card table base>
3696 __ ldub(O0, O1, O2); // O2 := [O0 + O1]
3697
3698 __ cmp_and_br_short(O2, G1SATBCardTableModRefBS::g1_young_card_val(), Assembler::equal, Assembler::pt, young_card);
3699
3700 __ membar(Assembler::Membar_mask_bits(Assembler::StoreLoad));
3701 __ ldub(O0, O1, O2); // O2 := [O0 + O1]
3702
3703 assert(CardTableModRefBS::dirty_card_val() == 0, "otherwise check this code");
3704 __ cmp_and_br_short(O2, G0, Assembler::notEqual, Assembler::pt, not_already_dirty);
3705
3706 __ bind(young_card);
3707 // We didn't take the branch, so we're already dirty: return.
3708 // Use return-from-leaf
3709 __ retl();
3710 __ delayed()->nop();
3711
3712 // Not dirty.
3713 __ bind(not_already_dirty);
3714
3715 // Get O0 + O1 into a reg by itself
3716 __ add(O0, O1, O3);
3717
3718 // First, dirty it.
3719 __ stb(G0, O3, G0); // [cardPtr] := 0 (i.e., dirty).
3720
3721 int dirty_card_q_index_byte_offset =
3722 in_bytes(JavaThread::dirty_card_queue_offset() +
3723 DirtyCardQueue::byte_offset_of_index());
3724 int dirty_card_q_buf_byte_offset =
3725 in_bytes(JavaThread::dirty_card_queue_offset() +
3726 DirtyCardQueue::byte_offset_of_buf());
3727 __ bind(restart);
3728
3729 // Load the index into the update buffer. DirtyCardQueue::_index is
3730 // a size_t so ld_ptr is appropriate here.
3731 __ ld_ptr(G2_thread, dirty_card_q_index_byte_offset, L0);
3732
3733 // index == 0?
3734 __ cmp_and_brx_short(L0, G0, Assembler::equal, Assembler::pn, refill);
3735
3736 __ ld_ptr(G2_thread, dirty_card_q_buf_byte_offset, L1);
3737 __ sub(L0, oopSize, L0);
3738
3739 __ st_ptr(O3, L1, L0); // [_buf + index] := I0
3740 // Use return-from-leaf
3741 __ retl();
3742 __ delayed()->st_ptr(L0, G2_thread, dirty_card_q_index_byte_offset);
3743
3744 __ bind(refill);
3745 address handle_zero =
3746 CAST_FROM_FN_PTR(address,
3747 &DirtyCardQueueSet::handle_zero_index_for_thread);
3748 // This should be rare enough that we can afford to save all the
3749 // scratch registers that the calling context might be using.
3750 __ mov(G1_scratch, L3);
3751 __ mov(G3_scratch, L5);
3752 // We need the value of O3 above (for the write into the buffer), so we
3753 // save and restore it.
3754 __ mov(O3, L6);
3755 // Since the call will overwrite O7, we save and restore that, as well.
3756 __ mov(O7, L4);
3757
3758 __ call_VM_leaf(L7_thread_cache, handle_zero, G2_thread);
3759 __ mov(L3, G1_scratch);
3760 __ mov(L5, G3_scratch);
3761 __ mov(L6, O3);
3762 __ br(Assembler::always, /*annul*/false, Assembler::pt, restart);
3763 __ delayed()->mov(L4, O7);
3764
3765 dirty_card_log_enqueue = start;
3766 dirty_card_log_enqueue_end = __ pc();
3767 // XXX Should have a guarantee here about not going off the end!
3768 // Does it already do so? Do an experiment...
3769
3770 #undef __
3771
3772 }
3773
3774 void MacroAssembler::g1_write_barrier_post(Register store_addr, Register new_val, Register tmp) {
3775
3776 Label filtered;
3777 MacroAssembler* post_filter_masm = this;
3778
3779 if (new_val == G0) return;
3780
3781 G1SATBCardTableLoggingModRefBS* bs =
3782 barrier_set_cast<G1SATBCardTableLoggingModRefBS>(Universe::heap()->barrier_set());
3783
3784 if (G1RSBarrierRegionFilter) {
3785 xor3(store_addr, new_val, tmp);
3786 srlx(tmp, HeapRegion::LogOfHRGrainBytes, tmp);
3787
3788 // XXX Should I predict this taken or not? Does it matter?
3789 cmp_and_brx_short(tmp, G0, Assembler::equal, Assembler::pt, filtered);
3790 }
3791
3792 // If the "store_addr" register is an "in" or "local" register, move it to
3793 // a scratch reg so we can pass it as an argument.
3794 bool use_scr = !(store_addr->is_global() || store_addr->is_out());
3795 // Pick a scratch register different from "tmp".
3796 Register scr = (tmp == G1_scratch ? G3_scratch : G1_scratch);
3797 // Make sure we use up the delay slot!
3798 if (use_scr) {
3799 post_filter_masm->mov(store_addr, scr);
3800 } else {
3801 post_filter_masm->nop();
3802 }
3803 save_frame(0);
3804 call(dirty_card_log_enqueue);
3805 if (use_scr) {
3806 delayed()->mov(scr, O0);
3807 } else {
3808 delayed()->mov(store_addr->after_save(), O0);
3809 }
3810 restore();
3811
3812 bind(filtered);
3813 }
3814
3815 // Called from init_globals() after universe_init() and before interpreter_init()
3816 void g1_barrier_stubs_init() {
3817 CollectedHeap* heap = Universe::heap();
3818 if (heap->kind() == CollectedHeap::G1CollectedHeap) {
3819 // Only needed for G1
3820 if (dirty_card_log_enqueue == 0) {
3821 G1SATBCardTableLoggingModRefBS* bs =
3822 barrier_set_cast<G1SATBCardTableLoggingModRefBS>(heap->barrier_set());
3823 generate_dirty_card_log_enqueue(bs->byte_map_base);
3824 assert(dirty_card_log_enqueue != 0, "postcondition.");
3825 }
3826 if (satb_log_enqueue_with_frame == 0) {
3827 generate_satb_log_enqueue(true);
3828 assert(satb_log_enqueue_with_frame != 0, "postcondition.");
3829 }
3830 if (satb_log_enqueue_frameless == 0) {
3831 generate_satb_log_enqueue(false);
3832 assert(satb_log_enqueue_frameless != 0, "postcondition.");
3833 }
3834 }
3835 }
3836
3837 #endif // INCLUDE_ALL_GCS
3838 ///////////////////////////////////////////////////////////////////////////////////
3839
3840 void MacroAssembler::card_write_barrier_post(Register store_addr, Register new_val, Register tmp) {
3841 // If we're writing constant NULL, we can skip the write barrier.
3842 if (new_val == G0) return;
3843 CardTableModRefBS* bs =
3844 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3845 assert(bs->kind() == BarrierSet::CardTableForRS ||
3846 bs->kind() == BarrierSet::CardTableExtension, "wrong barrier");
3847 card_table_write(bs->byte_map_base, tmp, store_addr);
3848 }
3849
3850 // ((OopHandle)result).resolve();
3851 void MacroAssembler::resolve_oop_handle(Register result) {
3852 // OopHandle::resolve is an indirection.
3853 ld_ptr(result, 0, result);
3854 }
3855
3856 void MacroAssembler::load_mirror(Register mirror, Register method) {
3857 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
3858 ld_ptr(method, in_bytes(Method::const_offset()), mirror);
3859 ld_ptr(mirror, in_bytes(ConstMethod::constants_offset()), mirror);
3860 ld_ptr(mirror, ConstantPool::pool_holder_offset_in_bytes(), mirror);
3861 ld_ptr(mirror, mirror_offset, mirror);
3862 resolve_oop_handle(mirror);
3863 }
3864
3865 void MacroAssembler::load_klass(Register src_oop, Register klass) {
3866 // The number of bytes in this code is used by
3867 // MachCallDynamicJavaNode::ret_addr_offset()
3868 // if this changes, change that.
3869 if (UseCompressedClassPointers) {
3870 lduw(src_oop, oopDesc::klass_offset_in_bytes(), klass);
3871 decode_klass_not_null(klass);
3872 } else {
3873 ld_ptr(src_oop, oopDesc::klass_offset_in_bytes(), klass);
3874 }
3875 }
3876
3877 void MacroAssembler::store_klass(Register klass, Register dst_oop) {
3878 if (UseCompressedClassPointers) {
3879 assert(dst_oop != klass, "not enough registers");
3880 encode_klass_not_null(klass);
3881 st(klass, dst_oop, oopDesc::klass_offset_in_bytes());
3882 } else {
3883 st_ptr(klass, dst_oop, oopDesc::klass_offset_in_bytes());
3884 }
3885 }
3886
3887 void MacroAssembler::store_klass_gap(Register s, Register d) {
3888 if (UseCompressedClassPointers) {
3889 assert(s != d, "not enough registers");
3890 st(s, d, oopDesc::klass_gap_offset_in_bytes());
3891 }
3892 }
3893
3894 void MacroAssembler::load_heap_oop(const Address& s, Register d) {
3895 if (UseCompressedOops) {
3896 lduw(s, d);
3897 decode_heap_oop(d);
3898 } else {
3899 ld_ptr(s, d);
3900 }
3901 }
3902
3903 void MacroAssembler::load_heap_oop(Register s1, Register s2, Register d) {
3904 if (UseCompressedOops) {
3905 lduw(s1, s2, d);
3906 decode_heap_oop(d, d);
3907 } else {
3908 ld_ptr(s1, s2, d);
3909 }
3910 }
3911
3912 void MacroAssembler::load_heap_oop(Register s1, int simm13a, Register d) {
3913 if (UseCompressedOops) {
3914 lduw(s1, simm13a, d);
3915 decode_heap_oop(d, d);
3916 } else {
3917 ld_ptr(s1, simm13a, d);
3918 }
3919 }
3920
3921 void MacroAssembler::load_heap_oop(Register s1, RegisterOrConstant s2, Register d) {
3922 if (s2.is_constant()) load_heap_oop(s1, s2.as_constant(), d);
3923 else load_heap_oop(s1, s2.as_register(), d);
3924 }
3925
3926 void MacroAssembler::store_heap_oop(Register d, Register s1, Register s2) {
3927 if (UseCompressedOops) {
3928 assert(s1 != d && s2 != d, "not enough registers");
3929 encode_heap_oop(d);
3930 st(d, s1, s2);
3931 } else {
3932 st_ptr(d, s1, s2);
3933 }
3934 }
3935
3936 void MacroAssembler::store_heap_oop(Register d, Register s1, int simm13a) {
3937 if (UseCompressedOops) {
3938 assert(s1 != d, "not enough registers");
3939 encode_heap_oop(d);
3940 st(d, s1, simm13a);
3941 } else {
3942 st_ptr(d, s1, simm13a);
3943 }
3944 }
3945
3946 void MacroAssembler::store_heap_oop(Register d, const Address& a, int offset) {
3947 if (UseCompressedOops) {
3948 assert(a.base() != d, "not enough registers");
3949 encode_heap_oop(d);
3950 st(d, a, offset);
3951 } else {
3952 st_ptr(d, a, offset);
3953 }
3954 }
3955
3956
3957 void MacroAssembler::encode_heap_oop(Register src, Register dst) {
3958 assert (UseCompressedOops, "must be compressed");
3959 assert (Universe::heap() != NULL, "java heap should be initialized");
3960 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
3961 verify_oop(src);
3962 if (Universe::narrow_oop_base() == NULL) {
3963 srlx(src, LogMinObjAlignmentInBytes, dst);
3964 return;
3965 }
3966 Label done;
3967 if (src == dst) {
3968 // optimize for frequent case src == dst
3969 bpr(rc_nz, true, Assembler::pt, src, done);
3970 delayed() -> sub(src, G6_heapbase, dst); // annuled if not taken
3971 bind(done);
3972 srlx(src, LogMinObjAlignmentInBytes, dst);
3973 } else {
3974 bpr(rc_z, false, Assembler::pn, src, done);
3975 delayed() -> mov(G0, dst);
3976 // could be moved before branch, and annulate delay,
3977 // but may add some unneeded work decoding null
3978 sub(src, G6_heapbase, dst);
3979 srlx(dst, LogMinObjAlignmentInBytes, dst);
3980 bind(done);
3981 }
3982 }
3983
3984
3985 void MacroAssembler::encode_heap_oop_not_null(Register r) {
3986 assert (UseCompressedOops, "must be compressed");
3987 assert (Universe::heap() != NULL, "java heap should be initialized");
3988 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
3989 verify_oop(r);
3990 if (Universe::narrow_oop_base() != NULL)
3991 sub(r, G6_heapbase, r);
3992 srlx(r, LogMinObjAlignmentInBytes, r);
3993 }
3994
3995 void MacroAssembler::encode_heap_oop_not_null(Register src, Register dst) {
3996 assert (UseCompressedOops, "must be compressed");
3997 assert (Universe::heap() != NULL, "java heap should be initialized");
3998 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
3999 verify_oop(src);
4000 if (Universe::narrow_oop_base() == NULL) {
4001 srlx(src, LogMinObjAlignmentInBytes, dst);
4002 } else {
4003 sub(src, G6_heapbase, dst);
4004 srlx(dst, LogMinObjAlignmentInBytes, dst);
4005 }
4006 }
4007
4008 // Same algorithm as oops.inline.hpp decode_heap_oop.
4009 void MacroAssembler::decode_heap_oop(Register src, Register dst) {
4010 assert (UseCompressedOops, "must be compressed");
4011 assert (Universe::heap() != NULL, "java heap should be initialized");
4012 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4013 sllx(src, LogMinObjAlignmentInBytes, dst);
4014 if (Universe::narrow_oop_base() != NULL) {
4015 Label done;
4016 bpr(rc_nz, true, Assembler::pt, dst, done);
4017 delayed() -> add(dst, G6_heapbase, dst); // annuled if not taken
4018 bind(done);
4019 }
4020 verify_oop(dst);
4021 }
4022
4023 void MacroAssembler::decode_heap_oop_not_null(Register r) {
4024 // Do not add assert code to this unless you change vtableStubs_sparc.cpp
4025 // pd_code_size_limit.
4026 // Also do not verify_oop as this is called by verify_oop.
4027 assert (UseCompressedOops, "must be compressed");
4028 assert (Universe::heap() != NULL, "java heap should be initialized");
4029 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4030 sllx(r, LogMinObjAlignmentInBytes, r);
4031 if (Universe::narrow_oop_base() != NULL)
4032 add(r, G6_heapbase, r);
4033 }
4034
4035 void MacroAssembler::decode_heap_oop_not_null(Register src, Register dst) {
4036 // Do not add assert code to this unless you change vtableStubs_sparc.cpp
4037 // pd_code_size_limit.
4038 // Also do not verify_oop as this is called by verify_oop.
4039 assert (UseCompressedOops, "must be compressed");
4040 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
4041 sllx(src, LogMinObjAlignmentInBytes, dst);
4042 if (Universe::narrow_oop_base() != NULL)
4043 add(dst, G6_heapbase, dst);
4044 }
4045
4046 void MacroAssembler::encode_klass_not_null(Register r) {
4047 assert (UseCompressedClassPointers, "must be compressed");
4048 if (Universe::narrow_klass_base() != NULL) {
4049 assert(r != G6_heapbase, "bad register choice");
4050 set((intptr_t)Universe::narrow_klass_base(), G6_heapbase);
4051 sub(r, G6_heapbase, r);
4052 if (Universe::narrow_klass_shift() != 0) {
4053 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong");
4054 srlx(r, LogKlassAlignmentInBytes, r);
4055 }
4056 reinit_heapbase();
4057 } else {
4058 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift() || Universe::narrow_klass_shift() == 0, "decode alg wrong");
4059 srlx(r, Universe::narrow_klass_shift(), r);
4060 }
4061 }
4062
4063 void MacroAssembler::encode_klass_not_null(Register src, Register dst) {
4064 if (src == dst) {
4065 encode_klass_not_null(src);
4066 } else {
4067 assert (UseCompressedClassPointers, "must be compressed");
4068 if (Universe::narrow_klass_base() != NULL) {
4069 set((intptr_t)Universe::narrow_klass_base(), dst);
4070 sub(src, dst, dst);
4071 if (Universe::narrow_klass_shift() != 0) {
4072 srlx(dst, LogKlassAlignmentInBytes, dst);
4073 }
4074 } else {
4075 // shift src into dst
4076 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift() || Universe::narrow_klass_shift() == 0, "decode alg wrong");
4077 srlx(src, Universe::narrow_klass_shift(), dst);
4078 }
4079 }
4080 }
4081
4082 // Function instr_size_for_decode_klass_not_null() counts the instructions
4083 // generated by decode_klass_not_null() and reinit_heapbase(). Hence, if
4084 // the instructions they generate change, then this method needs to be updated.
4085 int MacroAssembler::instr_size_for_decode_klass_not_null() {
4086 assert (UseCompressedClassPointers, "only for compressed klass ptrs");
4087 int num_instrs = 1; // shift src,dst or add
4088 if (Universe::narrow_klass_base() != NULL) {
4089 // set + add + set
4090 num_instrs += insts_for_internal_set((intptr_t)Universe::narrow_klass_base()) +
4091 insts_for_internal_set((intptr_t)Universe::narrow_ptrs_base());
4092 if (Universe::narrow_klass_shift() != 0) {
4093 num_instrs += 1; // sllx
4094 }
4095 }
4096 return num_instrs * BytesPerInstWord;
4097 }
4098
4099 // !!! If the instructions that get generated here change then function
4100 // instr_size_for_decode_klass_not_null() needs to get updated.
4101 void MacroAssembler::decode_klass_not_null(Register r) {
4102 // Do not add assert code to this unless you change vtableStubs_sparc.cpp
4103 // pd_code_size_limit.
4104 assert (UseCompressedClassPointers, "must be compressed");
4105 if (Universe::narrow_klass_base() != NULL) {
4106 assert(r != G6_heapbase, "bad register choice");
4107 set((intptr_t)Universe::narrow_klass_base(), G6_heapbase);
4108 if (Universe::narrow_klass_shift() != 0)
4109 sllx(r, LogKlassAlignmentInBytes, r);
4110 add(r, G6_heapbase, r);
4111 reinit_heapbase();
4112 } else {
4113 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift() || Universe::narrow_klass_shift() == 0, "decode alg wrong");
4114 sllx(r, Universe::narrow_klass_shift(), r);
4115 }
4116 }
4117
4118 void MacroAssembler::decode_klass_not_null(Register src, Register dst) {
4119 if (src == dst) {
4120 decode_klass_not_null(src);
4121 } else {
4122 // Do not add assert code to this unless you change vtableStubs_sparc.cpp
4123 // pd_code_size_limit.
4124 assert (UseCompressedClassPointers, "must be compressed");
4125 if (Universe::narrow_klass_base() != NULL) {
4126 if (Universe::narrow_klass_shift() != 0) {
4127 assert((src != G6_heapbase) && (dst != G6_heapbase), "bad register choice");
4128 set((intptr_t)Universe::narrow_klass_base(), G6_heapbase);
4129 sllx(src, LogKlassAlignmentInBytes, dst);
4130 add(dst, G6_heapbase, dst);
4131 reinit_heapbase();
4132 } else {
4133 set((intptr_t)Universe::narrow_klass_base(), dst);
4134 add(src, dst, dst);
4135 }
4136 } else {
4137 // shift/mov src into dst.
4138 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift() || Universe::narrow_klass_shift() == 0, "decode alg wrong");
4139 sllx(src, Universe::narrow_klass_shift(), dst);
4140 }
4141 }
4142 }
4143
4144 void MacroAssembler::reinit_heapbase() {
4145 if (UseCompressedOops || UseCompressedClassPointers) {
4146 if (Universe::heap() != NULL) {
4147 set((intptr_t)Universe::narrow_ptrs_base(), G6_heapbase);
4148 } else {
4149 AddressLiteral base(Universe::narrow_ptrs_base_addr());
4150 load_ptr_contents(base, G6_heapbase);
4151 }
4152 }
4153 }
4154
4155 #ifdef COMPILER2
4156
4157 // Compress char[] to byte[] by compressing 16 bytes at once. Return 0 on failure.
4158 void MacroAssembler::string_compress_16(Register src, Register dst, Register cnt, Register result,
4159 Register tmp1, Register tmp2, Register tmp3, Register tmp4,
4160 FloatRegister ftmp1, FloatRegister ftmp2, FloatRegister ftmp3, Label& Ldone) {
4161 Label Lloop, Lslow;
4162 assert(UseVIS >= 3, "VIS3 is required");
4163 assert_different_registers(src, dst, cnt, tmp1, tmp2, tmp3, tmp4, result);
4164 assert_different_registers(ftmp1, ftmp2, ftmp3);
4165
4166 // Check if cnt >= 8 (= 16 bytes)
4167 cmp(cnt, 8);
4168 br(Assembler::less, false, Assembler::pn, Lslow);
4169 delayed()->mov(cnt, result); // copy count
4170
4171 // Check for 8-byte alignment of src and dst
4172 or3(src, dst, tmp1);
4173 andcc(tmp1, 7, G0);
4174 br(Assembler::notZero, false, Assembler::pn, Lslow);
4175 delayed()->nop();
4176
4177 // Set mask for bshuffle instruction
4178 Register mask = tmp4;
4179 set(0x13579bdf, mask);
4180 bmask(mask, G0, G0);
4181
4182 // Set mask to 0xff00 ff00 ff00 ff00 to check for non-latin1 characters
4183 Assembler::sethi(0xff00fc00, mask); // mask = 0x0000 0000 ff00 fc00
4184 add(mask, 0x300, mask); // mask = 0x0000 0000 ff00 ff00
4185 sllx(mask, 32, tmp1); // tmp1 = 0xff00 ff00 0000 0000
4186 or3(mask, tmp1, mask); // mask = 0xff00 ff00 ff00 ff00
4187
4188 // Load first 8 bytes
4189 ldx(src, 0, tmp1);
4190
4191 bind(Lloop);
4192 // Load next 8 bytes
4193 ldx(src, 8, tmp2);
4194
4195 // Check for non-latin1 character by testing if the most significant byte of a char is set.
4196 // Although we have to move the data between integer and floating point registers, this is
4197 // still faster than the corresponding VIS instructions (ford/fand/fcmpd).
4198 or3(tmp1, tmp2, tmp3);
4199 btst(tmp3, mask);
4200 // annul zeroing if branch is not taken to preserve original count
4201 brx(Assembler::notZero, true, Assembler::pn, Ldone);
4202 delayed()->mov(G0, result); // 0 - failed
4203
4204 // Move bytes into float register
4205 movxtod(tmp1, ftmp1);
4206 movxtod(tmp2, ftmp2);
4207
4208 // Compress by copying one byte per char from ftmp1 and ftmp2 to ftmp3
4209 bshuffle(ftmp1, ftmp2, ftmp3);
4210 stf(FloatRegisterImpl::D, ftmp3, dst, 0);
4211
4212 // Increment addresses and decrement count
4213 inc(src, 16);
4214 inc(dst, 8);
4215 dec(cnt, 8);
4216
4217 cmp(cnt, 8);
4218 // annul LDX if branch is not taken to prevent access past end of string
4219 br(Assembler::greaterEqual, true, Assembler::pt, Lloop);
4220 delayed()->ldx(src, 0, tmp1);
4221
4222 // Fallback to slow version
4223 bind(Lslow);
4224 }
4225
4226 // Compress char[] to byte[]. Return 0 on failure.
4227 void MacroAssembler::string_compress(Register src, Register dst, Register cnt, Register result, Register tmp, Label& Ldone) {
4228 Label Lloop;
4229 assert_different_registers(src, dst, cnt, tmp, result);
4230
4231 lduh(src, 0, tmp);
4232
4233 bind(Lloop);
4234 inc(src, sizeof(jchar));
4235 cmp(tmp, 0xff);
4236 // annul zeroing if branch is not taken to preserve original count
4237 br(Assembler::greater, true, Assembler::pn, Ldone); // don't check xcc
4238 delayed()->mov(G0, result); // 0 - failed
4239 deccc(cnt);
4240 stb(tmp, dst, 0);
4241 inc(dst);
4242 // annul LDUH if branch is not taken to prevent access past end of string
4243 br(Assembler::notZero, true, Assembler::pt, Lloop);
4244 delayed()->lduh(src, 0, tmp); // hoisted
4245 }
4246
4247 // Inflate byte[] to char[] by inflating 16 bytes at once.
4248 void MacroAssembler::string_inflate_16(Register src, Register dst, Register cnt, Register tmp,
4249 FloatRegister ftmp1, FloatRegister ftmp2, FloatRegister ftmp3, FloatRegister ftmp4, Label& Ldone) {
4250 Label Lloop, Lslow;
4251 assert(UseVIS >= 3, "VIS3 is required");
4252 assert_different_registers(src, dst, cnt, tmp);
4253 assert_different_registers(ftmp1, ftmp2, ftmp3, ftmp4);
4254
4255 // Check if cnt >= 8 (= 16 bytes)
4256 cmp(cnt, 8);
4257 br(Assembler::less, false, Assembler::pn, Lslow);
4258 delayed()->nop();
4259
4260 // Check for 8-byte alignment of src and dst
4261 or3(src, dst, tmp);
4262 andcc(tmp, 7, G0);
4263 br(Assembler::notZero, false, Assembler::pn, Lslow);
4264 // Initialize float register to zero
4265 FloatRegister zerof = ftmp4;
4266 delayed()->fzero(FloatRegisterImpl::D, zerof);
4267
4268 // Load first 8 bytes
4269 ldf(FloatRegisterImpl::D, src, 0, ftmp1);
4270
4271 bind(Lloop);
4272 inc(src, 8);
4273 dec(cnt, 8);
4274
4275 // Inflate the string by interleaving each byte from the source array
4276 // with a zero byte and storing the result in the destination array.
4277 fpmerge(zerof, ftmp1->successor(), ftmp2);
4278 stf(FloatRegisterImpl::D, ftmp2, dst, 8);
4279 fpmerge(zerof, ftmp1, ftmp3);
4280 stf(FloatRegisterImpl::D, ftmp3, dst, 0);
4281
4282 inc(dst, 16);
4283
4284 cmp(cnt, 8);
4285 // annul LDX if branch is not taken to prevent access past end of string
4286 br(Assembler::greaterEqual, true, Assembler::pt, Lloop);
4287 delayed()->ldf(FloatRegisterImpl::D, src, 0, ftmp1);
4288
4289 // Fallback to slow version
4290 bind(Lslow);
4291 }
4292
4293 // Inflate byte[] to char[].
4294 void MacroAssembler::string_inflate(Register src, Register dst, Register cnt, Register tmp, Label& Ldone) {
4295 Label Loop;
4296 assert_different_registers(src, dst, cnt, tmp);
4297
4298 ldub(src, 0, tmp);
4299 bind(Loop);
4300 inc(src);
4301 deccc(cnt);
4302 sth(tmp, dst, 0);
4303 inc(dst, sizeof(jchar));
4304 // annul LDUB if branch is not taken to prevent access past end of string
4305 br(Assembler::notZero, true, Assembler::pt, Loop);
4306 delayed()->ldub(src, 0, tmp); // hoisted
4307 }
4308
4309 void MacroAssembler::string_compare(Register str1, Register str2,
4310 Register cnt1, Register cnt2,
4311 Register tmp1, Register tmp2,
4312 Register result, int ae) {
4313 Label Ldone, Lloop;
4314 assert_different_registers(str1, str2, cnt1, cnt2, tmp1, result);
4315 int stride1, stride2;
4316
4317 // Note: Making use of the fact that compareTo(a, b) == -compareTo(b, a)
4318 // we interchange str1 and str2 in the UL case and negate the result.
4319 // Like this, str1 is always latin1 encoded, expect for the UU case.
4320
4321 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
4322 srl(cnt2, 1, cnt2);
4323 }
4324
4325 // See if the lengths are different, and calculate min in cnt1.
4326 // Save diff in case we need it for a tie-breaker.
4327 Label Lskip;
4328 Register diff = tmp1;
4329 subcc(cnt1, cnt2, diff);
4330 br(Assembler::greater, true, Assembler::pt, Lskip);
4331 // cnt2 is shorter, so use its count:
4332 delayed()->mov(cnt2, cnt1);
4333 bind(Lskip);
4334
4335 // Rename registers
4336 Register limit1 = cnt1;
4337 Register limit2 = limit1;
4338 Register chr1 = result;
4339 Register chr2 = cnt2;
4340 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
4341 // We need an additional register to keep track of two limits
4342 assert_different_registers(str1, str2, cnt1, cnt2, tmp1, tmp2, result);
4343 limit2 = tmp2;
4344 }
4345
4346 // Is the minimum length zero?
4347 cmp(limit1, (int)0); // use cast to resolve overloading ambiguity
4348 br(Assembler::equal, true, Assembler::pn, Ldone);
4349 // result is difference in lengths
4350 if (ae == StrIntrinsicNode::UU) {
4351 delayed()->sra(diff, 1, result); // Divide by 2 to get number of chars
4352 } else {
4353 delayed()->mov(diff, result);
4354 }
4355
4356 // Load first characters
4357 if (ae == StrIntrinsicNode::LL) {
4358 stride1 = stride2 = sizeof(jbyte);
4359 ldub(str1, 0, chr1);
4360 ldub(str2, 0, chr2);
4361 } else if (ae == StrIntrinsicNode::UU) {
4362 stride1 = stride2 = sizeof(jchar);
4363 lduh(str1, 0, chr1);
4364 lduh(str2, 0, chr2);
4365 } else {
4366 stride1 = sizeof(jbyte);
4367 stride2 = sizeof(jchar);
4368 ldub(str1, 0, chr1);
4369 lduh(str2, 0, chr2);
4370 }
4371
4372 // Compare first characters
4373 subcc(chr1, chr2, chr1);
4374 br(Assembler::notZero, false, Assembler::pt, Ldone);
4375 assert(chr1 == result, "result must be pre-placed");
4376 delayed()->nop();
4377
4378 // Check if the strings start at same location
4379 cmp(str1, str2);
4380 brx(Assembler::equal, true, Assembler::pn, Ldone);
4381 delayed()->mov(G0, result); // result is zero
4382
4383 // We have no guarantee that on 64 bit the higher half of limit is 0
4384 signx(limit1);
4385
4386 // Get limit
4387 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
4388 sll(limit1, 1, limit2);
4389 subcc(limit2, stride2, chr2);
4390 }
4391 subcc(limit1, stride1, chr1);
4392 br(Assembler::zero, true, Assembler::pn, Ldone);
4393 // result is difference in lengths
4394 if (ae == StrIntrinsicNode::UU) {
4395 delayed()->sra(diff, 1, result); // Divide by 2 to get number of chars
4396 } else {
4397 delayed()->mov(diff, result);
4398 }
4399
4400 // Shift str1 and str2 to the end of the arrays, negate limit
4401 add(str1, limit1, str1);
4402 add(str2, limit2, str2);
4403 neg(chr1, limit1); // limit1 = -(limit1-stride1)
4404 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
4405 neg(chr2, limit2); // limit2 = -(limit2-stride2)
4406 }
4407
4408 // Compare the rest of the characters
4409 load_sized_value(Address(str1, limit1), chr1, (ae == StrIntrinsicNode::UU) ? 2 : 1, false);
4410
4411 bind(Lloop);
4412 load_sized_value(Address(str2, limit2), chr2, (ae == StrIntrinsicNode::LL) ? 1 : 2, false);
4413
4414 subcc(chr1, chr2, chr1);
4415 br(Assembler::notZero, false, Assembler::pt, Ldone);
4416 assert(chr1 == result, "result must be pre-placed");
4417 delayed()->inccc(limit1, stride1);
4418 if (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL) {
4419 inccc(limit2, stride2);
4420 }
4421
4422 // annul LDUB if branch is not taken to prevent access past end of string
4423 br(Assembler::notZero, true, Assembler::pt, Lloop);
4424 delayed()->load_sized_value(Address(str1, limit1), chr1, (ae == StrIntrinsicNode::UU) ? 2 : 1, false);
4425
4426 // If strings are equal up to min length, return the length difference.
4427 if (ae == StrIntrinsicNode::UU) {
4428 // Divide by 2 to get number of chars
4429 sra(diff, 1, result);
4430 } else {
4431 mov(diff, result);
4432 }
4433
4434 // Otherwise, return the difference between the first mismatched chars.
4435 bind(Ldone);
4436 if(ae == StrIntrinsicNode::UL) {
4437 // Negate result (see note above)
4438 neg(result);
4439 }
4440 }
4441
4442 void MacroAssembler::array_equals(bool is_array_equ, Register ary1, Register ary2,
4443 Register limit, Register tmp, Register result, bool is_byte) {
4444 Label Ldone, Lloop, Lremaining;
4445 assert_different_registers(ary1, ary2, limit, tmp, result);
4446
4447 int length_offset = arrayOopDesc::length_offset_in_bytes();
4448 int base_offset = arrayOopDesc::base_offset_in_bytes(is_byte ? T_BYTE : T_CHAR);
4449 assert(base_offset % 8 == 0, "Base offset must be 8-byte aligned");
4450
4451 if (is_array_equ) {
4452 // return true if the same array
4453 cmp(ary1, ary2);
4454 brx(Assembler::equal, true, Assembler::pn, Ldone);
4455 delayed()->mov(1, result); // equal
4456
4457 br_null(ary1, true, Assembler::pn, Ldone);
4458 delayed()->clr(result); // not equal
4459
4460 br_null(ary2, true, Assembler::pn, Ldone);
4461 delayed()->clr(result); // not equal
4462
4463 // load the lengths of arrays
4464 ld(Address(ary1, length_offset), limit);
4465 ld(Address(ary2, length_offset), tmp);
4466
4467 // return false if the two arrays are not equal length
4468 cmp(limit, tmp);
4469 br(Assembler::notEqual, true, Assembler::pn, Ldone);
4470 delayed()->clr(result); // not equal
4471 }
4472
4473 cmp_zero_and_br(Assembler::zero, limit, Ldone, true, Assembler::pn);
4474 delayed()->mov(1, result); // zero-length arrays are equal
4475
4476 if (is_array_equ) {
4477 // load array addresses
4478 add(ary1, base_offset, ary1);
4479 add(ary2, base_offset, ary2);
4480 // set byte count
4481 if (!is_byte) {
4482 sll(limit, exact_log2(sizeof(jchar)), limit);
4483 }
4484 } else {
4485 // We have no guarantee that on 64 bit the higher half of limit is 0
4486 signx(limit);
4487 }
4488
4489 #ifdef ASSERT
4490 // Sanity check for doubleword (8-byte) alignment of ary1 and ary2.
4491 // Guaranteed on 64-bit systems (see arrayOopDesc::header_size_in_bytes()).
4492 Label Laligned;
4493 or3(ary1, ary2, tmp);
4494 andcc(tmp, 7, tmp);
4495 br_null_short(tmp, Assembler::pn, Laligned);
4496 STOP("First array element is not 8-byte aligned.");
4497 should_not_reach_here();
4498 bind(Laligned);
4499 #endif
4500
4501 // Shift ary1 and ary2 to the end of the arrays, negate limit
4502 add(ary1, limit, ary1);
4503 add(ary2, limit, ary2);
4504 neg(limit, limit);
4505
4506 // MAIN LOOP
4507 // Load and compare array elements of size 'byte_width' until the elements are not
4508 // equal or we reached the end of the arrays. If the size of the arrays is not a
4509 // multiple of 'byte_width', we simply read over the end of the array, bail out and
4510 // compare the remaining bytes below by skipping the garbage bytes.
4511 ldx(ary1, limit, result);
4512 bind(Lloop);
4513 ldx(ary2, limit, tmp);
4514 inccc(limit, 8);
4515 // Bail out if we reached the end (but still do the comparison)
4516 br(Assembler::positive, false, Assembler::pn, Lremaining);
4517 delayed()->cmp(result, tmp);
4518 // Check equality of elements
4519 brx(Assembler::equal, false, Assembler::pt, target(Lloop));
4520 delayed()->ldx(ary1, limit, result);
4521
4522 ba(Ldone);
4523 delayed()->clr(result); // not equal
4524
4525 // TAIL COMPARISON
4526 // We got here because we reached the end of the arrays. 'limit' is the number of
4527 // garbage bytes we may have compared by reading over the end of the arrays. Shift
4528 // out the garbage and compare the remaining elements.
4529 bind(Lremaining);
4530 // Optimistic shortcut: elements potentially including garbage are equal
4531 brx(Assembler::equal, true, Assembler::pt, target(Ldone));
4532 delayed()->mov(1, result); // equal
4533 // Shift 'limit' bytes to the right and compare
4534 sll(limit, 3, limit); // bytes to bits
4535 srlx(result, limit, result);
4536 srlx(tmp, limit, tmp);
4537 cmp(result, tmp);
4538 clr(result);
4539 movcc(Assembler::equal, false, xcc, 1, result);
4540
4541 bind(Ldone);
4542 }
4543
4544 void MacroAssembler::has_negatives(Register inp, Register size, Register result, Register t2, Register t3, Register t4, Register t5) {
4545
4546 // test for negative bytes in input string of a given size
4547 // result 1 if found, 0 otherwise.
4548
4549 Label Lcore, Ltail, Lreturn, Lcore_rpt;
4550
4551 assert_different_registers(inp, size, t2, t3, t4, t5, result);
4552
4553 Register i = result; // result used as integer index i until very end
4554 Register lmask = t2; // t2 is aliased to lmask
4555
4556 // INITIALIZATION
4557 // ===========================================================
4558 // initialize highbits mask -> lmask = 0x8080808080808080 (8B/64b)
4559 // compute unaligned offset -> i
4560 // compute core end index -> t5
4561 Assembler::sethi(0x80808000, t2); //! sethi macro fails to emit optimal
4562 add(t2, 0x80, t2);
4563 sllx(t2, 32, t3);
4564 or3(t3, t2, lmask); // 0x8080808080808080 -> lmask
4565 sra(size,0,size);
4566 andcc(inp, 0x7, i); // unaligned offset -> i
4567 br(Assembler::zero, true, Assembler::pn, Lcore); // starts 8B aligned?
4568 delayed()->add(size, -8, t5); // (annuled) core end index -> t5
4569
4570 // ===========================================================
4571
4572 // UNALIGNED HEAD
4573 // ===========================================================
4574 // * unaligned head handling: grab aligned 8B containing unaligned inp(ut)
4575 // * obliterate (ignore) bytes outside string by shifting off reg ends
4576 // * compare with bitmask, short circuit return true if one or more high
4577 // bits set.
4578 cmp(size, 0);
4579 br(Assembler::zero, true, Assembler::pn, Lreturn); // short-circuit?
4580 delayed()->mov(0,result); // annuled so i not clobbered for following
4581 neg(i, t4);
4582 add(i, size, t5);
4583 ldx(inp, t4, t3); // raw aligned 8B containing unaligned head -> t3
4584 mov(8, t4);
4585 sub(t4, t5, t4);
4586 sra(t4, 31, t5);
4587 andn(t4, t5, t5);
4588 add(i, t5, t4);
4589 sll(t5, 3, t5);
4590 sll(t4, 3, t4); // # bits to shift right, left -> t5,t4
4591 srlx(t3, t5, t3);
4592 sllx(t3, t4, t3); // bytes outside string in 8B header obliterated -> t3
4593 andcc(lmask, t3, G0);
4594 brx(Assembler::notZero, true, Assembler::pn, Lreturn); // short circuit?
4595 delayed()->mov(1,result); // annuled so i not clobbered for following
4596 add(size, -8, t5); // core end index -> t5
4597 mov(8, t4);
4598 sub(t4, i, i); // # bytes examined in unalgn head (<8) -> i
4599 // ===========================================================
4600
4601 // ALIGNED CORE
4602 // ===========================================================
4603 // * iterate index i over aligned 8B sections of core, comparing with
4604 // bitmask, short circuit return true if one or more high bits set
4605 // t5 contains core end index/loop limit which is the index
4606 // of the MSB of last (unaligned) 8B fully contained in the string.
4607 // inp contains address of first byte in string/array
4608 // lmask contains 8B high bit mask for comparison
4609 // i contains next index to be processed (adr. inp+i is on 8B boundary)
4610 bind(Lcore);
4611 cmp_and_br_short(i, t5, Assembler::greater, Assembler::pn, Ltail);
4612 bind(Lcore_rpt);
4613 ldx(inp, i, t3);
4614 andcc(t3, lmask, G0);
4615 brx(Assembler::notZero, true, Assembler::pn, Lreturn);
4616 delayed()->mov(1, result); // annuled so i not clobbered for following
4617 add(i, 8, i);
4618 cmp_and_br_short(i, t5, Assembler::lessEqual, Assembler::pn, Lcore_rpt);
4619 // ===========================================================
4620
4621 // ALIGNED TAIL (<8B)
4622 // ===========================================================
4623 // handle aligned tail of 7B or less as complete 8B, obliterating end of
4624 // string bytes by shifting them off end, compare what's left with bitmask
4625 // inp contains address of first byte in string/array
4626 // lmask contains 8B high bit mask for comparison
4627 // i contains next index to be processed (adr. inp+i is on 8B boundary)
4628 bind(Ltail);
4629 subcc(size, i, t4); // # of remaining bytes in string -> t4
4630 // return 0 if no more remaining bytes
4631 br(Assembler::lessEqual, true, Assembler::pn, Lreturn);
4632 delayed()->mov(0, result); // annuled so i not clobbered for following
4633 ldx(inp, i, t3); // load final 8B (aligned) containing tail -> t3
4634 mov(8, t5);
4635 sub(t5, t4, t4);
4636 mov(0, result); // ** i clobbered at this point
4637 sll(t4, 3, t4); // bits beyond end of string -> t4
4638 srlx(t3, t4, t3); // bytes beyond end now obliterated -> t3
4639 andcc(lmask, t3, G0);
4640 movcc(Assembler::notZero, false, xcc, 1, result);
4641 bind(Lreturn);
4642 }
4643
4644 #endif
4645
4646
4647 // Use BIS for zeroing (count is in bytes).
4648 void MacroAssembler::bis_zeroing(Register to, Register count, Register temp, Label& Ldone) {
4649 assert(UseBlockZeroing && VM_Version::has_blk_zeroing(), "only works with BIS zeroing");
4650 Register end = count;
4651 int cache_line_size = VM_Version::prefetch_data_size();
4652 assert(cache_line_size > 0, "cache line size should be known for this code");
4653 // Minimum count when BIS zeroing can be used since
4654 // it needs membar which is expensive.
4655 int block_zero_size = MAX2(cache_line_size*3, (int)BlockZeroingLowLimit);
4656
4657 Label small_loop;
4658 // Check if count is negative (dead code) or zero.
4659 // Note, count uses 64bit in 64 bit VM.
4660 cmp_and_brx_short(count, 0, Assembler::lessEqual, Assembler::pn, Ldone);
4661
4662 // Use BIS zeroing only for big arrays since it requires membar.
4663 if (Assembler::is_simm13(block_zero_size)) { // < 4096
4664 cmp(count, block_zero_size);
4665 } else {
4666 set(block_zero_size, temp);
4667 cmp(count, temp);
4668 }
4669 br(Assembler::lessUnsigned, false, Assembler::pt, small_loop);
4670 delayed()->add(to, count, end);
4671
4672 // Note: size is >= three (32 bytes) cache lines.
4673
4674 // Clean the beginning of space up to next cache line.
4675 for (int offs = 0; offs < cache_line_size; offs += 8) {
4676 stx(G0, to, offs);
4677 }
4678
4679 // align to next cache line
4680 add(to, cache_line_size, to);
4681 and3(to, -cache_line_size, to);
4682
4683 // Note: size left >= two (32 bytes) cache lines.
4684
4685 // BIS should not be used to zero tail (64 bytes)
4686 // to avoid zeroing a header of the following object.
4687 sub(end, (cache_line_size*2)-8, end);
4688
4689 Label bis_loop;
4690 bind(bis_loop);
4691 stxa(G0, to, G0, Assembler::ASI_ST_BLKINIT_PRIMARY);
4692 add(to, cache_line_size, to);
4693 cmp_and_brx_short(to, end, Assembler::lessUnsigned, Assembler::pt, bis_loop);
4694
4695 // BIS needs membar.
4696 membar(Assembler::StoreLoad);
4697
4698 add(end, (cache_line_size*2)-8, end); // restore end
4699 cmp_and_brx_short(to, end, Assembler::greaterEqualUnsigned, Assembler::pn, Ldone);
4700
4701 // Clean the tail.
4702 bind(small_loop);
4703 stx(G0, to, 0);
4704 add(to, 8, to);
4705 cmp_and_brx_short(to, end, Assembler::lessUnsigned, Assembler::pt, small_loop);
4706 nop(); // Separate short branches
4707 }
4708
4709 /**
4710 * Update CRC-32[C] with a byte value according to constants in table
4711 *
4712 * @param [in,out]crc Register containing the crc.
4713 * @param [in]val Register containing the byte to fold into the CRC.
4714 * @param [in]table Register containing the table of crc constants.
4715 *
4716 * uint32_t crc;
4717 * val = crc_table[(val ^ crc) & 0xFF];
4718 * crc = val ^ (crc >> 8);
4719 */
4720 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
4721 xor3(val, crc, val);
4722 and3(val, 0xFF, val);
4723 sllx(val, 2, val);
4724 lduw(table, val, val);
4725 srlx(crc, 8, crc);
4726 xor3(val, crc, crc);
4727 }
4728
4729 // Reverse byte order of lower 32 bits, assuming upper 32 bits all zeros
4730 void MacroAssembler::reverse_bytes_32(Register src, Register dst, Register tmp) {
4731 srlx(src, 24, dst);
4732
4733 sllx(src, 32+8, tmp);
4734 srlx(tmp, 32+24, tmp);
4735 sllx(tmp, 8, tmp);
4736 or3(dst, tmp, dst);
4737
4738 sllx(src, 32+16, tmp);
4739 srlx(tmp, 32+24, tmp);
4740 sllx(tmp, 16, tmp);
4741 or3(dst, tmp, dst);
4742
4743 sllx(src, 32+24, tmp);
4744 srlx(tmp, 32, tmp);
4745 or3(dst, tmp, dst);
4746 }
4747
4748 void MacroAssembler::movitof_revbytes(Register src, FloatRegister dst, Register tmp1, Register tmp2) {
4749 reverse_bytes_32(src, tmp1, tmp2);
4750 movxtod(tmp1, dst);
4751 }
4752
4753 void MacroAssembler::movftoi_revbytes(FloatRegister src, Register dst, Register tmp1, Register tmp2) {
4754 movdtox(src, tmp1);
4755 reverse_bytes_32(tmp1, dst, tmp2);
4756 }
4757
4758 void MacroAssembler::fold_128bit_crc32(Register xcrc_hi, Register xcrc_lo, Register xK_hi, Register xK_lo, Register xtmp_hi, Register xtmp_lo, Register buf, int offset) {
4759 xmulx(xcrc_hi, xK_hi, xtmp_lo);
4760 xmulxhi(xcrc_hi, xK_hi, xtmp_hi);
4761 xmulxhi(xcrc_lo, xK_lo, xcrc_hi);
4762 xmulx(xcrc_lo, xK_lo, xcrc_lo);
4763 xor3(xcrc_lo, xtmp_lo, xcrc_lo);
4764 xor3(xcrc_hi, xtmp_hi, xcrc_hi);
4765 ldxl(buf, G0, xtmp_lo);
4766 inc(buf, 8);
4767 ldxl(buf, G0, xtmp_hi);
4768 inc(buf, 8);
4769 xor3(xcrc_lo, xtmp_lo, xcrc_lo);
4770 xor3(xcrc_hi, xtmp_hi, xcrc_hi);
4771 }
4772
4773 void MacroAssembler::fold_128bit_crc32(Register xcrc_hi, Register xcrc_lo, Register xK_hi, Register xK_lo, Register xtmp_hi, Register xtmp_lo, Register xbuf_hi, Register xbuf_lo) {
4774 mov(xcrc_lo, xtmp_lo);
4775 mov(xcrc_hi, xtmp_hi);
4776 xmulx(xtmp_hi, xK_hi, xtmp_lo);
4777 xmulxhi(xtmp_hi, xK_hi, xtmp_hi);
4778 xmulxhi(xcrc_lo, xK_lo, xcrc_hi);
4779 xmulx(xcrc_lo, xK_lo, xcrc_lo);
4780 xor3(xcrc_lo, xbuf_lo, xcrc_lo);
4781 xor3(xcrc_hi, xbuf_hi, xcrc_hi);
4782 xor3(xcrc_lo, xtmp_lo, xcrc_lo);
4783 xor3(xcrc_hi, xtmp_hi, xcrc_hi);
4784 }
4785
4786 void MacroAssembler::fold_8bit_crc32(Register xcrc, Register table, Register xtmp, Register tmp) {
4787 and3(xcrc, 0xFF, tmp);
4788 sllx(tmp, 2, tmp);
4789 lduw(table, tmp, xtmp);
4790 srlx(xcrc, 8, xcrc);
4791 xor3(xtmp, xcrc, xcrc);
4792 }
4793
4794 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
4795 and3(crc, 0xFF, tmp);
4796 srlx(crc, 8, crc);
4797 sllx(tmp, 2, tmp);
4798 lduw(table, tmp, tmp);
4799 xor3(tmp, crc, crc);
4800 }
4801
4802 #define CRC32_TMP_REG_NUM 18
4803
4804 #define CRC32_CONST_64 0x163cd6124
4805 #define CRC32_CONST_96 0x0ccaa009e
4806 #define CRC32_CONST_160 0x1751997d0
4807 #define CRC32_CONST_480 0x1c6e41596
4808 #define CRC32_CONST_544 0x154442bd4
4809
4810 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table) {
4811
4812 Label L_cleanup_loop, L_cleanup_check, L_align_loop, L_align_check;
4813 Label L_main_loop_prologue;
4814 Label L_fold_512b, L_fold_512b_loop, L_fold_128b;
4815 Label L_fold_tail, L_fold_tail_loop;
4816 Label L_8byte_fold_loop, L_8byte_fold_check;
4817
4818 const Register tmp[CRC32_TMP_REG_NUM] = {L0, L1, L2, L3, L4, L5, L6, G1, I0, I1, I2, I3, I4, I5, I7, O4, O5, G3};
4819
4820 Register const_64 = tmp[CRC32_TMP_REG_NUM-1];
4821 Register const_96 = tmp[CRC32_TMP_REG_NUM-1];
4822 Register const_160 = tmp[CRC32_TMP_REG_NUM-2];
4823 Register const_480 = tmp[CRC32_TMP_REG_NUM-1];
4824 Register const_544 = tmp[CRC32_TMP_REG_NUM-2];
4825
4826 set(ExternalAddress(StubRoutines::crc_table_addr()), table);
4827
4828 not1(crc); // ~c
4829 clruwu(crc); // clear upper 32 bits of crc
4830
4831 // Check if below cutoff, proceed directly to cleanup code
4832 mov(31, G4);
4833 cmp_and_br_short(len, G4, Assembler::lessEqualUnsigned, Assembler::pt, L_cleanup_check);
4834
4835 // Align buffer to 8 byte boundry
4836 mov(8, O5);
4837 and3(buf, 0x7, O4);
4838 sub(O5, O4, O5);
4839 and3(O5, 0x7, O5);
4840 sub(len, O5, len);
4841 ba(L_align_check);
4842 delayed()->nop();
4843
4844 // Alignment loop, table look up method for up to 7 bytes
4845 bind(L_align_loop);
4846 ldub(buf, 0, O4);
4847 inc(buf);
4848 dec(O5);
4849 xor3(O4, crc, O4);
4850 and3(O4, 0xFF, O4);
4851 sllx(O4, 2, O4);
4852 lduw(table, O4, O4);
4853 srlx(crc, 8, crc);
4854 xor3(O4, crc, crc);
4855 bind(L_align_check);
4856 nop();
4857 cmp_and_br_short(O5, 0, Assembler::notEqual, Assembler::pt, L_align_loop);
4858
4859 // Aligned on 64-bit (8-byte) boundry at this point
4860 // Check if still above cutoff (31-bytes)
4861 mov(31, G4);
4862 cmp_and_br_short(len, G4, Assembler::lessEqualUnsigned, Assembler::pt, L_cleanup_check);
4863 // At least 32 bytes left to process
4864
4865 // Free up registers by storing them to FP registers
4866 for (int i = 0; i < CRC32_TMP_REG_NUM; i++) {
4867 movxtod(tmp[i], as_FloatRegister(2*i));
4868 }
4869
4870 // Determine which loop to enter
4871 // Shared prologue
4872 ldxl(buf, G0, tmp[0]);
4873 inc(buf, 8);
4874 ldxl(buf, G0, tmp[1]);
4875 inc(buf, 8);
4876 xor3(tmp[0], crc, tmp[0]); // Fold CRC into first few bytes
4877 and3(crc, 0, crc); // Clear out the crc register
4878 // Main loop needs 128-bytes at least
4879 mov(128, G4);
4880 mov(64, tmp[2]);
4881 cmp_and_br_short(len, G4, Assembler::greaterEqualUnsigned, Assembler::pt, L_main_loop_prologue);
4882 // Less than 64 bytes
4883 nop();
4884 cmp_and_br_short(len, tmp[2], Assembler::lessUnsigned, Assembler::pt, L_fold_tail);
4885 // Between 64 and 127 bytes
4886 set64(CRC32_CONST_96, const_96, tmp[8]);
4887 set64(CRC32_CONST_160, const_160, tmp[9]);
4888 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[2], tmp[3], buf, 0);
4889 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[4], tmp[5], buf, 16);
4890 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[6], tmp[7], buf, 32);
4891 dec(len, 48);
4892 ba(L_fold_tail);
4893 delayed()->nop();
4894
4895 bind(L_main_loop_prologue);
4896 for (int i = 2; i < 8; i++) {
4897 ldxl(buf, G0, tmp[i]);
4898 inc(buf, 8);
4899 }
4900
4901 // Fold total 512 bits of polynomial on each iteration,
4902 // 128 bits per each of 4 parallel streams
4903 set64(CRC32_CONST_480, const_480, tmp[8]);
4904 set64(CRC32_CONST_544, const_544, tmp[9]);
4905
4906 mov(128, G4);
4907 bind(L_fold_512b_loop);
4908 fold_128bit_crc32(tmp[1], tmp[0], const_480, const_544, tmp[9], tmp[8], buf, 0);
4909 fold_128bit_crc32(tmp[3], tmp[2], const_480, const_544, tmp[11], tmp[10], buf, 16);
4910 fold_128bit_crc32(tmp[5], tmp[4], const_480, const_544, tmp[13], tmp[12], buf, 32);
4911 fold_128bit_crc32(tmp[7], tmp[6], const_480, const_544, tmp[15], tmp[14], buf, 64);
4912 dec(len, 64);
4913 cmp_and_br_short(len, G4, Assembler::greaterEqualUnsigned, Assembler::pt, L_fold_512b_loop);
4914
4915 // Fold 512 bits to 128 bits
4916 bind(L_fold_512b);
4917 set64(CRC32_CONST_96, const_96, tmp[8]);
4918 set64(CRC32_CONST_160, const_160, tmp[9]);
4919
4920 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[8], tmp[9], tmp[3], tmp[2]);
4921 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[8], tmp[9], tmp[5], tmp[4]);
4922 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[8], tmp[9], tmp[7], tmp[6]);
4923 dec(len, 48);
4924
4925 // Fold the rest of 128 bits data chunks
4926 bind(L_fold_tail);
4927 mov(32, G4);
4928 cmp_and_br_short(len, G4, Assembler::lessEqualUnsigned, Assembler::pt, L_fold_128b);
4929
4930 set64(CRC32_CONST_96, const_96, tmp[8]);
4931 set64(CRC32_CONST_160, const_160, tmp[9]);
4932
4933 bind(L_fold_tail_loop);
4934 fold_128bit_crc32(tmp[1], tmp[0], const_96, const_160, tmp[2], tmp[3], buf, 0);
4935 sub(len, 16, len);
4936 cmp_and_br_short(len, G4, Assembler::greaterEqualUnsigned, Assembler::pt, L_fold_tail_loop);
4937
4938 // Fold the 128 bits in tmps 0 - 1 into tmp 1
4939 bind(L_fold_128b);
4940
4941 set64(CRC32_CONST_64, const_64, tmp[4]);
4942
4943 xmulx(const_64, tmp[0], tmp[2]);
4944 xmulxhi(const_64, tmp[0], tmp[3]);
4945
4946 srl(tmp[2], G0, tmp[4]);
4947 xmulx(const_64, tmp[4], tmp[4]);
4948
4949 srlx(tmp[2], 32, tmp[2]);
4950 sllx(tmp[3], 32, tmp[3]);
4951 or3(tmp[2], tmp[3], tmp[2]);
4952
4953 xor3(tmp[4], tmp[1], tmp[4]);
4954 xor3(tmp[4], tmp[2], tmp[1]);
4955 dec(len, 8);
4956
4957 // Use table lookup for the 8 bytes left in tmp[1]
4958 dec(len, 8);
4959
4960 // 8 8-bit folds to compute 32-bit CRC.
4961 for (int j = 0; j < 4; j++) {
4962 fold_8bit_crc32(tmp[1], table, tmp[2], tmp[3]);
4963 }
4964 srl(tmp[1], G0, crc); // move 32 bits to general register
4965 for (int j = 0; j < 4; j++) {
4966 fold_8bit_crc32(crc, table, tmp[3]);
4967 }
4968
4969 bind(L_8byte_fold_check);
4970
4971 // Restore int registers saved in FP registers
4972 for (int i = 0; i < CRC32_TMP_REG_NUM; i++) {
4973 movdtox(as_FloatRegister(2*i), tmp[i]);
4974 }
4975
4976 ba(L_cleanup_check);
4977 delayed()->nop();
4978
4979 // Table look-up method for the remaining few bytes
4980 bind(L_cleanup_loop);
4981 ldub(buf, 0, O4);
4982 inc(buf);
4983 dec(len);
4984 xor3(O4, crc, O4);
4985 and3(O4, 0xFF, O4);
4986 sllx(O4, 2, O4);
4987 lduw(table, O4, O4);
4988 srlx(crc, 8, crc);
4989 xor3(O4, crc, crc);
4990 bind(L_cleanup_check);
4991 nop();
4992 cmp_and_br_short(len, 0, Assembler::greaterUnsigned, Assembler::pt, L_cleanup_loop);
4993
4994 not1(crc);
4995 }
4996
4997 #define CHUNK_LEN 128 /* 128 x 8B = 1KB */
4998 #define CHUNK_K1 0x1307a0206 /* reverseBits(pow(x, CHUNK_LEN*8*8*3 - 32) mod P(x)) << 1 */
4999 #define CHUNK_K2 0x1a0f717c4 /* reverseBits(pow(x, CHUNK_LEN*8*8*2 - 32) mod P(x)) << 1 */
5000 #define CHUNK_K3 0x0170076fa /* reverseBits(pow(x, CHUNK_LEN*8*8*1 - 32) mod P(x)) << 1 */
5001
5002 void MacroAssembler::kernel_crc32c(Register crc, Register buf, Register len, Register table) {
5003
5004 Label L_crc32c_head, L_crc32c_aligned;
5005 Label L_crc32c_parallel, L_crc32c_parallel_loop;
5006 Label L_crc32c_serial, L_crc32c_x32_loop, L_crc32c_x8, L_crc32c_x8_loop;
5007 Label L_crc32c_done, L_crc32c_tail, L_crc32c_return;
5008
5009 set(ExternalAddress(StubRoutines::crc32c_table_addr()), table);
5010
5011 cmp_and_br_short(len, 0, Assembler::lessEqual, Assembler::pn, L_crc32c_return);
5012
5013 // clear upper 32 bits of crc
5014 clruwu(crc);
5015
5016 and3(buf, 7, G4);
5017 cmp_and_brx_short(G4, 0, Assembler::equal, Assembler::pt, L_crc32c_aligned);
5018
5019 mov(8, G1);
5020 sub(G1, G4, G4);
5021
5022 // ------ process the misaligned head (7 bytes or less) ------
5023 bind(L_crc32c_head);
5024
5025 // crc = (crc >>> 8) ^ byteTable[(crc ^ b) & 0xFF];
5026 ldub(buf, 0, G1);
5027 update_byte_crc32(crc, G1, table);
5028
5029 inc(buf);
5030 dec(len);
5031 cmp_and_br_short(len, 0, Assembler::equal, Assembler::pn, L_crc32c_return);
5032 dec(G4);
5033 cmp_and_br_short(G4, 0, Assembler::greater, Assembler::pt, L_crc32c_head);
5034
5035 // ------ process the 8-byte-aligned body ------
5036 bind(L_crc32c_aligned);
5037 nop();
5038 cmp_and_br_short(len, 8, Assembler::less, Assembler::pn, L_crc32c_tail);
5039
5040 // reverse the byte order of lower 32 bits to big endian, and move to FP side
5041 movitof_revbytes(crc, F0, G1, G3);
5042
5043 set(CHUNK_LEN*8*4, G4);
5044 cmp_and_br_short(len, G4, Assembler::less, Assembler::pt, L_crc32c_serial);
5045
5046 // ------ process four 1KB chunks in parallel ------
5047 bind(L_crc32c_parallel);
5048
5049 fzero(FloatRegisterImpl::D, F2);
5050 fzero(FloatRegisterImpl::D, F4);
5051 fzero(FloatRegisterImpl::D, F6);
5052
5053 mov(CHUNK_LEN - 1, G4);
5054 bind(L_crc32c_parallel_loop);
5055 // schedule ldf's ahead of crc32c's to hide the load-use latency
5056 ldf(FloatRegisterImpl::D, buf, 0, F8);
5057 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*8, F10);
5058 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*16, F12);
5059 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*24, F14);
5060 crc32c(F0, F8, F0);
5061 crc32c(F2, F10, F2);
5062 crc32c(F4, F12, F4);
5063 crc32c(F6, F14, F6);
5064 inc(buf, 8);
5065 dec(G4);
5066 cmp_and_br_short(G4, 0, Assembler::greater, Assembler::pt, L_crc32c_parallel_loop);
5067
5068 ldf(FloatRegisterImpl::D, buf, 0, F8);
5069 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*8, F10);
5070 ldf(FloatRegisterImpl::D, buf, CHUNK_LEN*16, F12);
5071 crc32c(F0, F8, F0);
5072 crc32c(F2, F10, F2);
5073 crc32c(F4, F12, F4);
5074
5075 inc(buf, CHUNK_LEN*24);
5076 ldfl(FloatRegisterImpl::D, buf, G0, F14); // load in little endian
5077 inc(buf, 8);
5078
5079 prefetch(buf, 0, Assembler::severalReads);
5080 prefetch(buf, CHUNK_LEN*8, Assembler::severalReads);
5081 prefetch(buf, CHUNK_LEN*16, Assembler::severalReads);
5082 prefetch(buf, CHUNK_LEN*24, Assembler::severalReads);
5083
5084 // move to INT side, and reverse the byte order of lower 32 bits to little endian
5085 movftoi_revbytes(F0, O4, G1, G4);
5086 movftoi_revbytes(F2, O5, G1, G4);
5087 movftoi_revbytes(F4, G5, G1, G4);
5088
5089 // combine the results of 4 chunks
5090 set64(CHUNK_K1, G3, G1);
5091 xmulx(O4, G3, O4);
5092 set64(CHUNK_K2, G3, G1);
5093 xmulx(O5, G3, O5);
5094 set64(CHUNK_K3, G3, G1);
5095 xmulx(G5, G3, G5);
5096
5097 movdtox(F14, G4);
5098 xor3(O4, O5, O5);
5099 xor3(G5, O5, O5);
5100 xor3(G4, O5, O5);
5101
5102 // reverse the byte order to big endian, via stack, and move to FP side
5103 // TODO: use new revb instruction
5104 add(SP, -8, G1);
5105 srlx(G1, 3, G1);
5106 sllx(G1, 3, G1);
5107 stx(O5, G1, G0);
5108 ldfl(FloatRegisterImpl::D, G1, G0, F2); // load in little endian
5109
5110 crc32c(F6, F2, F0);
5111
5112 set(CHUNK_LEN*8*4, G4);
5113 sub(len, G4, len);
5114 cmp_and_br_short(len, G4, Assembler::greaterEqual, Assembler::pt, L_crc32c_parallel);
5115 nop();
5116 cmp_and_br_short(len, 0, Assembler::equal, Assembler::pt, L_crc32c_done);
5117
5118 bind(L_crc32c_serial);
5119
5120 mov(32, G4);
5121 cmp_and_br_short(len, G4, Assembler::less, Assembler::pn, L_crc32c_x8);
5122
5123 // ------ process 32B chunks ------
5124 bind(L_crc32c_x32_loop);
5125 ldf(FloatRegisterImpl::D, buf, 0, F2);
5126 crc32c(F0, F2, F0);
5127 ldf(FloatRegisterImpl::D, buf, 8, F2);
5128 crc32c(F0, F2, F0);
5129 ldf(FloatRegisterImpl::D, buf, 16, F2);
5130 crc32c(F0, F2, F0);
5131 ldf(FloatRegisterImpl::D, buf, 24, F2);
5132 inc(buf, 32);
5133 crc32c(F0, F2, F0);
5134 dec(len, 32);
5135 cmp_and_br_short(len, G4, Assembler::greaterEqual, Assembler::pt, L_crc32c_x32_loop);
5136
5137 bind(L_crc32c_x8);
5138 nop();
5139 cmp_and_br_short(len, 8, Assembler::less, Assembler::pt, L_crc32c_done);
5140
5141 // ------ process 8B chunks ------
5142 bind(L_crc32c_x8_loop);
5143 ldf(FloatRegisterImpl::D, buf, 0, F2);
5144 inc(buf, 8);
5145 crc32c(F0, F2, F0);
5146 dec(len, 8);
5147 cmp_and_br_short(len, 8, Assembler::greaterEqual, Assembler::pt, L_crc32c_x8_loop);
5148
5149 bind(L_crc32c_done);
5150
5151 // move to INT side, and reverse the byte order of lower 32 bits to little endian
5152 movftoi_revbytes(F0, crc, G1, G3);
5153
5154 cmp_and_br_short(len, 0, Assembler::equal, Assembler::pt, L_crc32c_return);
5155
5156 // ------ process the misaligned tail (7 bytes or less) ------
5157 bind(L_crc32c_tail);
5158
5159 // crc = (crc >>> 8) ^ byteTable[(crc ^ b) & 0xFF];
5160 ldub(buf, 0, G1);
5161 update_byte_crc32(crc, G1, table);
5162
5163 inc(buf);
5164 dec(len);
5165 cmp_and_br_short(len, 0, Assembler::greater, Assembler::pt, L_crc32c_tail);
5166
5167 bind(L_crc32c_return);
5168 nop();
5169 }