Mercurial > hg > truffle

comparison src/cpu/ppc/vm/stubGenerator_ppc.cpp @ 14408:ec28f9c041ff

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
8019972: PPC64 (part 9): platform files for interpreter only VM. Summary: With this change the HotSpot core build works on Linux/PPC64. The VM succesfully executes simple test programs. Reviewed-by: kvn
author goetz
date Fri, 02 Aug 2013 16:46:45 +0200
parents
children 67fa91961822
comparison
equal deleted inserted replaced
14407:94c202aa2646 14408:ec28f9c041ff
1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * Copyright 2012, 2013 SAP AG. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "asm/assembler.hpp"
28 #include "asm/macroAssembler.inline.hpp"
29 #include "interpreter/interpreter.hpp"
30 #include "nativeInst_ppc.hpp"
31 #include "oops/instanceOop.hpp"
32 #include "oops/method.hpp"
33 #include "oops/objArrayKlass.hpp"
34 #include "oops/oop.inline.hpp"
35 #include "prims/methodHandles.hpp"
36 #include "runtime/frame.inline.hpp"
37 #include "runtime/handles.inline.hpp"
38 #include "runtime/sharedRuntime.hpp"
39 #include "runtime/stubCodeGenerator.hpp"
40 #include "runtime/stubRoutines.hpp"
41 #include "utilities/top.hpp"
42 #ifdef TARGET_OS_FAMILY_aix
43 # include "thread_aix.inline.hpp"
44 #endif
45 #ifdef TARGET_OS_FAMILY_linux
46 # include "thread_linux.inline.hpp"
47 #endif
48 #ifdef COMPILER2
49 #include "opto/runtime.hpp"
50 #endif
51
52 #define __ _masm->
53
54 #ifdef PRODUCT
55 #define BLOCK_COMMENT(str) // nothing
56 #else
57 #define BLOCK_COMMENT(str) __ block_comment(str)
58 #endif
59
60 class StubGenerator: public StubCodeGenerator {
61 private:
62
63 // Call stubs are used to call Java from C
64 //
65 // Arguments:
66 //
67 // R3 - call wrapper address : address
68 // R4 - result : intptr_t*
69 // R5 - result type : BasicType
70 // R6 - method : Method
71 // R7 - frame mgr entry point : address
72 // R8 - parameter block : intptr_t*
73 // R9 - parameter count in words : int
74 // R10 - thread : Thread*
75 //
76 address generate_call_stub(address& return_address) {
77 // Setup a new c frame, copy java arguments, call frame manager or
78 // native_entry, and process result.
79
80 StubCodeMark mark(this, "StubRoutines", "call_stub");
81
82 address start = __ emit_fd();
83
84 // some sanity checks
85 assert((sizeof(frame::abi_48) % 16) == 0, "unaligned");
86 assert((sizeof(frame::abi_112) % 16) == 0, "unaligned");
87 assert((sizeof(frame::spill_nonvolatiles) % 16) == 0, "unaligned");
88 assert((sizeof(frame::parent_ijava_frame_abi) % 16) == 0, "unaligned");
89 assert((sizeof(frame::entry_frame_locals) % 16) == 0, "unaligned");
90
91 Register r_arg_call_wrapper_addr = R3;
92 Register r_arg_result_addr = R4;
93 Register r_arg_result_type = R5;
94 Register r_arg_method = R6;
95 Register r_arg_entry = R7;
96 Register r_arg_thread = R10;
97
98 Register r_temp = R24;
99 Register r_top_of_arguments_addr = R25;
100 Register r_entryframe_fp = R26;
101
102 {
103 // Stack on entry to call_stub:
104 //
105 // F1 [C_FRAME]
106 // ...
107
108 Register r_arg_argument_addr = R8;
109 Register r_arg_argument_count = R9;
110 Register r_frame_alignment_in_bytes = R27;
111 Register r_argument_addr = R28;
112 Register r_argumentcopy_addr = R29;
113 Register r_argument_size_in_bytes = R30;
114 Register r_frame_size = R23;
115
116 Label arguments_copied;
117
118 // Save LR/CR to caller's C_FRAME.
119 __ save_LR_CR(R0);
120
121 // Zero extend arg_argument_count.
122 __ clrldi(r_arg_argument_count, r_arg_argument_count, 32);
123
124 // Save non-volatiles GPRs to ENTRY_FRAME (not yet pushed, but it's safe).
125 __ save_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
126
127 // Keep copy of our frame pointer (caller's SP).
128 __ mr(r_entryframe_fp, R1_SP);
129
130 BLOCK_COMMENT("Push ENTRY_FRAME including arguments");
131 // Push ENTRY_FRAME including arguments:
132 //
133 // F0 [TOP_IJAVA_FRAME_ABI]
134 // alignment (optional)
135 // [outgoing Java arguments]
136 // [ENTRY_FRAME_LOCALS]
137 // F1 [C_FRAME]
138 // ...
139
140 // calculate frame size
141
142 // unaligned size of arguments
143 __ sldi(r_argument_size_in_bytes,
144 r_arg_argument_count, Interpreter::logStackElementSize);
145 // arguments alignment (max 1 slot)
146 // FIXME: use round_to() here
147 __ andi_(r_frame_alignment_in_bytes, r_arg_argument_count, 1);
148 __ sldi(r_frame_alignment_in_bytes,
149 r_frame_alignment_in_bytes, Interpreter::logStackElementSize);
150
151 // size = unaligned size of arguments + top abi's size
152 __ addi(r_frame_size, r_argument_size_in_bytes,
153 frame::top_ijava_frame_abi_size);
154 // size += arguments alignment
155 __ add(r_frame_size,
156 r_frame_size, r_frame_alignment_in_bytes);
157 // size += size of call_stub locals
158 __ addi(r_frame_size,
159 r_frame_size, frame::entry_frame_locals_size);
160
161 // push ENTRY_FRAME
162 __ push_frame(r_frame_size, r_temp);
163
164 // initialize call_stub locals (step 1)
165 __ std(r_arg_call_wrapper_addr,
166 _entry_frame_locals_neg(call_wrapper_address), r_entryframe_fp);
167 __ std(r_arg_result_addr,
168 _entry_frame_locals_neg(result_address), r_entryframe_fp);
169 __ std(r_arg_result_type,
170 _entry_frame_locals_neg(result_type), r_entryframe_fp);
171 // we will save arguments_tos_address later
172
173
174 BLOCK_COMMENT("Copy Java arguments");
175 // copy Java arguments
176
177 // Calculate top_of_arguments_addr which will be R17_tos (not prepushed) later.
178 // FIXME: why not simply use SP+frame::top_ijava_frame_size?
179 __ addi(r_top_of_arguments_addr,
180 R1_SP, frame::top_ijava_frame_abi_size);
181 __ add(r_top_of_arguments_addr,
182 r_top_of_arguments_addr, r_frame_alignment_in_bytes);
183
184 // any arguments to copy?
185 __ cmpdi(CCR0, r_arg_argument_count, 0);
186 __ beq(CCR0, arguments_copied);
187
188 // prepare loop and copy arguments in reverse order
189 {
190 // init CTR with arg_argument_count
191 __ mtctr(r_arg_argument_count);
192
193 // let r_argumentcopy_addr point to last outgoing Java arguments P
194 __ mr(r_argumentcopy_addr, r_top_of_arguments_addr);
195
196 // let r_argument_addr point to last incoming java argument
197 __ add(r_argument_addr,
198 r_arg_argument_addr, r_argument_size_in_bytes);
199 __ addi(r_argument_addr, r_argument_addr, -BytesPerWord);
200
201 // now loop while CTR > 0 and copy arguments
202 {
203 Label next_argument;
204 __ bind(next_argument);
205
206 __ ld(r_temp, 0, r_argument_addr);
207 // argument_addr--;
208 __ addi(r_argument_addr, r_argument_addr, -BytesPerWord);
209 __ std(r_temp, 0, r_argumentcopy_addr);
210 // argumentcopy_addr++;
211 __ addi(r_argumentcopy_addr, r_argumentcopy_addr, BytesPerWord);
212
213 __ bdnz(next_argument);
214 }
215 }
216
217 // Arguments copied, continue.
218 __ bind(arguments_copied);
219 }
220
221 {
222 BLOCK_COMMENT("Call frame manager or native entry.");
223 // Call frame manager or native entry.
224 Register r_new_arg_entry = R14_state;
225 assert_different_registers(r_new_arg_entry, r_top_of_arguments_addr,
226 r_arg_method, r_arg_thread);
227
228 __ mr(r_new_arg_entry, r_arg_entry);
229
230 // Register state on entry to frame manager / native entry:
231 //
232 // R17_tos - intptr_t* sender tos (prepushed) Lesp = (SP) + copied_arguments_offset - 8
233 // R19_method - Method
234 // R16_thread - JavaThread*
235
236 // R17_tos must point to last argument - element_size.
237 __ addi(R17_tos, r_top_of_arguments_addr, -Interpreter::stackElementSize);
238
239 // initialize call_stub locals (step 2)
240 // now save R17_tos as arguments_tos_address
241 __ std(R17_tos, _entry_frame_locals_neg(arguments_tos_address), r_entryframe_fp);
242
243 // load argument registers for call
244 __ mr(R19_method, r_arg_method);
245 __ mr(R16_thread, r_arg_thread);
246 assert(R17_tos != r_arg_method, "trashed r_arg_method");
247 assert(R17_tos != r_arg_thread && R19_method != r_arg_thread, "trashed r_arg_thread");
248
249 // Set R15_prev_state to 0 for simplifying checks in callee.
250 __ li(R15_prev_state, 0);
251
252 // Stack on entry to frame manager / native entry:
253 //
254 // F0 [TOP_IJAVA_FRAME_ABI]
255 // alignment (optional)
256 // [outgoing Java arguments]
257 // [ENTRY_FRAME_LOCALS]
258 // F1 [C_FRAME]
259 // ...
260 //
261
262 // global toc register
263 __ load_const(R29, MacroAssembler::global_toc(), R11_scratch1);
264
265 // Load narrow oop base.
266 __ reinit_heapbase(R30, R11_scratch1);
267
268 // Remember the senderSP so we interpreter can pop c2i arguments off of the stack
269 // when called via a c2i.
270
271 // Pass initial_caller_sp to framemanager.
272 __ mr(R21_tmp1, R1_SP);
273
274 // Do a light-weight C-call here, r_new_arg_entry holds the address
275 // of the interpreter entry point (frame manager or native entry)
276 // and save runtime-value of LR in return_address.
277 assert(r_new_arg_entry != R17_tos && r_new_arg_entry != R19_method && r_new_arg_entry != R16_thread,
278 "trashed r_new_arg_entry");
279 return_address = __ call_stub(r_new_arg_entry);
280 }
281
282 {
283 BLOCK_COMMENT("Returned from frame manager or native entry.");
284 // Returned from frame manager or native entry.
285 // Now pop frame, process result, and return to caller.
286
287 // Stack on exit from frame manager / native entry:
288 //
289 // F0 [ABI]
290 // ...
291 // [ENTRY_FRAME_LOCALS]
292 // F1 [C_FRAME]
293 // ...
294 //
295 // Just pop the topmost frame ...
296 //
297
298 Label ret_is_object;
299 Label ret_is_long;
300 Label ret_is_float;
301 Label ret_is_double;
302
303 Register r_entryframe_fp = R30;
304 Register r_lr = R7_ARG5;
305 Register r_cr = R8_ARG6;
306
307 // Reload some volatile registers which we've spilled before the call
308 // to frame manager / native entry.
309 // Access all locals via frame pointer, because we know nothing about
310 // the topmost frame's size.
311 __ ld(r_entryframe_fp, _abi(callers_sp), R1_SP);
312 assert_different_registers(r_entryframe_fp, R3_RET, r_arg_result_addr, r_arg_result_type, r_cr, r_lr);
313 __ ld(r_arg_result_addr,
314 _entry_frame_locals_neg(result_address), r_entryframe_fp);
315 __ ld(r_arg_result_type,
316 _entry_frame_locals_neg(result_type), r_entryframe_fp);
317 __ ld(r_cr, _abi(cr), r_entryframe_fp);
318 __ ld(r_lr, _abi(lr), r_entryframe_fp);
319
320 // pop frame and restore non-volatiles, LR and CR
321 __ mr(R1_SP, r_entryframe_fp);
322 __ mtcr(r_cr);
323 __ mtlr(r_lr);
324
325 // Store result depending on type. Everything that is not
326 // T_OBJECT, T_LONG, T_FLOAT, or T_DOUBLE is treated as T_INT.
327 __ cmpwi(CCR0, r_arg_result_type, T_OBJECT);
328 __ cmpwi(CCR1, r_arg_result_type, T_LONG);
329 __ cmpwi(CCR5, r_arg_result_type, T_FLOAT);
330 __ cmpwi(CCR6, r_arg_result_type, T_DOUBLE);
331
332 // restore non-volatile registers
333 __ restore_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
334
335
336 // Stack on exit from call_stub:
337 //
338 // 0 [C_FRAME]
339 // ...
340 //
341 // no call_stub frames left.
342
343 // All non-volatiles have been restored at this point!!
344 assert(R3_RET == R3, "R3_RET should be R3");
345
346 __ beq(CCR0, ret_is_object);
347 __ beq(CCR1, ret_is_long);
348 __ beq(CCR5, ret_is_float);
349 __ beq(CCR6, ret_is_double);
350
351 // default:
352 __ stw(R3_RET, 0, r_arg_result_addr);
353 __ blr(); // return to caller
354
355 // case T_OBJECT:
356 __ bind(ret_is_object);
357 __ std(R3_RET, 0, r_arg_result_addr);
358 __ blr(); // return to caller
359
360 // case T_LONG:
361 __ bind(ret_is_long);
362 __ std(R3_RET, 0, r_arg_result_addr);
363 __ blr(); // return to caller
364
365 // case T_FLOAT:
366 __ bind(ret_is_float);
367 __ stfs(F1_RET, 0, r_arg_result_addr);
368 __ blr(); // return to caller
369
370 // case T_DOUBLE:
371 __ bind(ret_is_double);
372 __ stfd(F1_RET, 0, r_arg_result_addr);
373 __ blr(); // return to caller
374 }
375
376 return start;
377 }
378
379 // Return point for a Java call if there's an exception thrown in
380 // Java code. The exception is caught and transformed into a
381 // pending exception stored in JavaThread that can be tested from
382 // within the VM.
383 //
384 address generate_catch_exception() {
385 StubCodeMark mark(this, "StubRoutines", "catch_exception");
386
387 address start = __ pc();
388
389 // Registers alive
390 //
391 // R16_thread
392 // R3_ARG1 - address of pending exception
393 // R4_ARG2 - return address in call stub
394
395 const Register exception_file = R21_tmp1;
396 const Register exception_line = R22_tmp2;
397
398 __ load_const(exception_file, (void*)__FILE__);
399 __ load_const(exception_line, (void*)__LINE__);
400
401 __ std(R3_ARG1, thread_(pending_exception));
402 // store into `char *'
403 __ std(exception_file, thread_(exception_file));
404 // store into `int'
405 __ stw(exception_line, thread_(exception_line));
406
407 // complete return to VM
408 assert(StubRoutines::_call_stub_return_address != NULL, "must have been generated before");
409
410 __ mtlr(R4_ARG2);
411 // continue in call stub
412 __ blr();
413
414 return start;
415 }
416
417 // Continuation point for runtime calls returning with a pending
418 // exception. The pending exception check happened in the runtime
419 // or native call stub. The pending exception in Thread is
420 // converted into a Java-level exception.
421 //
422 address generate_forward_exception() {
423 StubCodeMark mark(this, "StubRoutines", "forward_exception");
424 address start = __ pc();
425
426 #if !defined(PRODUCT)
427 if (VerifyOops) {
428 // Get pending exception oop.
429 __ ld(R3_ARG1,
430 in_bytes(Thread::pending_exception_offset()),
431 R16_thread);
432 // Make sure that this code is only executed if there is a pending exception.
433 {
434 Label L;
435 __ cmpdi(CCR0, R3_ARG1, 0);
436 __ bne(CCR0, L);
437 __ stop("StubRoutines::forward exception: no pending exception (1)");
438 __ bind(L);
439 }
440 __ verify_oop(R3_ARG1, "StubRoutines::forward exception: not an oop");
441 }
442 #endif
443
444 // Save LR/CR and copy exception pc (LR) into R4_ARG2.
445 __ save_LR_CR(R4_ARG2);
446 __ push_frame_abi112(0, R0);
447 // Find exception handler.
448 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
449 SharedRuntime::exception_handler_for_return_address),
450 R16_thread,
451 R4_ARG2);
452 // Copy handler's address.
453 __ mtctr(R3_RET);
454 __ pop_frame();
455 __ restore_LR_CR(R0);
456
457 // Set up the arguments for the exception handler:
458 // - R3_ARG1: exception oop
459 // - R4_ARG2: exception pc.
460
461 // Load pending exception oop.
462 __ ld(R3_ARG1,
463 in_bytes(Thread::pending_exception_offset()),
464 R16_thread);
465
466 // The exception pc is the return address in the caller.
467 // Must load it into R4_ARG2.
468 __ mflr(R4_ARG2);
469
470 #ifdef ASSERT
471 // Make sure exception is set.
472 {
473 Label L;
474 __ cmpdi(CCR0, R3_ARG1, 0);
475 __ bne(CCR0, L);
476 __ stop("StubRoutines::forward exception: no pending exception (2)");
477 __ bind(L);
478 }
479 #endif
480
481 // Clear the pending exception.
482 __ li(R0, 0);
483 __ std(R0,
484 in_bytes(Thread::pending_exception_offset()),
485 R16_thread);
486 // Jump to exception handler.
487 __ bctr();
488
489 return start;
490 }
491
492 #undef __
493 #define __ masm->
494 // Continuation point for throwing of implicit exceptions that are
495 // not handled in the current activation. Fabricates an exception
496 // oop and initiates normal exception dispatching in this
497 // frame. Only callee-saved registers are preserved (through the
498 // normal register window / RegisterMap handling). If the compiler
499 // needs all registers to be preserved between the fault point and
500 // the exception handler then it must assume responsibility for that
501 // in AbstractCompiler::continuation_for_implicit_null_exception or
502 // continuation_for_implicit_division_by_zero_exception. All other
503 // implicit exceptions (e.g., NullPointerException or
504 // AbstractMethodError on entry) are either at call sites or
505 // otherwise assume that stack unwinding will be initiated, so
506 // caller saved registers were assumed volatile in the compiler.
507 //
508 // Note that we generate only this stub into a RuntimeStub, because
509 // it needs to be properly traversed and ignored during GC, so we
510 // change the meaning of the "__" macro within this method.
511 //
512 // Note: the routine set_pc_not_at_call_for_caller in
513 // SharedRuntime.cpp requires that this code be generated into a
514 // RuntimeStub.
515 address generate_throw_exception(const char* name, address runtime_entry, bool restore_saved_exception_pc,
516 Register arg1 = noreg, Register arg2 = noreg) {
517 CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0);
518 MacroAssembler* masm = new MacroAssembler(&code);
519
520 OopMapSet* oop_maps = new OopMapSet();
521 int frame_size_in_bytes = frame::abi_112_size;
522 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
523
524 StubCodeMark mark(this, "StubRoutines", "throw_exception");
525
526 address start = __ pc();
527
528 __ save_LR_CR(R11_scratch1);
529
530 // Push a frame.
531 __ push_frame_abi112(0, R11_scratch1);
532
533 address frame_complete_pc = __ pc();
534
535 if (restore_saved_exception_pc) {
536 __ unimplemented("StubGenerator::throw_exception with restore_saved_exception_pc", 74);
537 }
538
539 // Note that we always have a runtime stub frame on the top of
540 // stack by this point. Remember the offset of the instruction
541 // whose address will be moved to R11_scratch1.
542 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1);
543
544 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
545
546 __ mr(R3_ARG1, R16_thread);
547 if (arg1 != noreg) {
548 __ mr(R4_ARG2, arg1);
549 }
550 if (arg2 != noreg) {
551 __ mr(R5_ARG3, arg2);
552 }
553 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, runtime_entry),
554 relocInfo::none);
555
556 // Set an oopmap for the call site.
557 oop_maps->add_gc_map((int)(gc_map_pc - start), map);
558
559 __ reset_last_Java_frame();
560
561 #ifdef ASSERT
562 // Make sure that this code is only executed if there is a pending
563 // exception.
564 {
565 Label L;
566 __ ld(R0,
567 in_bytes(Thread::pending_exception_offset()),
568 R16_thread);
569 __ cmpdi(CCR0, R0, 0);
570 __ bne(CCR0, L);
571 __ stop("StubRoutines::throw_exception: no pending exception");
572 __ bind(L);
573 }
574 #endif
575
576 // Pop frame.
577 __ pop_frame();
578
579 __ restore_LR_CR(R11_scratch1);
580
581 __ load_const(R11_scratch1, StubRoutines::forward_exception_entry());
582 __ mtctr(R11_scratch1);
583 __ bctr();
584
585 // Create runtime stub with OopMap.
586 RuntimeStub* stub =
587 RuntimeStub::new_runtime_stub(name, &code,
588 /*frame_complete=*/ (int)(frame_complete_pc - start),
589 frame_size_in_bytes/wordSize,
590 oop_maps,
591 false);
592 return stub->entry_point();
593 }
594 #undef __
595 #define __ _masm->
596
597 // Generate G1 pre-write barrier for array.
598 //
599 // Input:
600 // from - register containing src address (only needed for spilling)
601 // to - register containing starting address
602 // count - register containing element count
603 // tmp - scratch register
604 //
605 // Kills:
606 // nothing
607 //
608 void gen_write_ref_array_pre_barrier(Register from, Register to, Register count, bool dest_uninitialized, Register Rtmp1) {
609 BarrierSet* const bs = Universe::heap()->barrier_set();
610 switch (bs->kind()) {
611 case BarrierSet::G1SATBCT:
612 case BarrierSet::G1SATBCTLogging:
613 // With G1, don't generate the call if we statically know that the target in uninitialized
614 if (!dest_uninitialized) {
615 const int spill_slots = 4 * wordSize;
616 const int frame_size = frame::abi_112_size + spill_slots;
617
618 __ save_LR_CR(R0);
619 __ push_frame_abi112(spill_slots, R0);
620 __ std(from, frame_size - 1 * wordSize, R1_SP);
621 __ std(to, frame_size - 2 * wordSize, R1_SP);
622 __ std(count, frame_size - 3 * wordSize, R1_SP);
623
624 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), to, count);
625
626 __ ld(from, frame_size - 1 * wordSize, R1_SP);
627 __ ld(to, frame_size - 2 * wordSize, R1_SP);
628 __ ld(count, frame_size - 3 * wordSize, R1_SP);
629 __ pop_frame();
630 __ restore_LR_CR(R0);
631 }
632 break;
633 case BarrierSet::CardTableModRef:
634 case BarrierSet::CardTableExtension:
635 case BarrierSet::ModRef:
636 break;
637 default:
638 ShouldNotReachHere();
639 }
640 }
641
642 // Generate CMS/G1 post-write barrier for array.
643 //
644 // Input:
645 // addr - register containing starting address
646 // count - register containing element count
647 // tmp - scratch register
648 //
649 // The input registers and R0 are overwritten.
650 //
651 void gen_write_ref_array_post_barrier(Register addr, Register count, Register tmp) {
652 BarrierSet* const bs = Universe::heap()->barrier_set();
653
654 switch (bs->kind()) {
655 case BarrierSet::G1SATBCT:
656 case BarrierSet::G1SATBCTLogging:
657 {
658 __ save_LR_CR(R0);
659 // We need this frame only that the callee can spill LR/CR.
660 __ push_frame_abi112(0, R0);
661
662 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), addr, count);
663
664 __ pop_frame();
665 __ restore_LR_CR(R0);
666 }
667 break;
668 case BarrierSet::CardTableModRef:
669 case BarrierSet::CardTableExtension:
670 {
671 Label Lskip_loop, Lstore_loop;
672 if (UseConcMarkSweepGC) {
673 // TODO PPC port: contribute optimization / requires shared changes
674 __ release();
675 }
676
677 CardTableModRefBS* const ct = (CardTableModRefBS*)bs;
678 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
679 assert_different_registers(addr, count, tmp);
680
681 __ sldi(count, count, LogBytesPerHeapOop);
682 __ addi(count, count, -BytesPerHeapOop);
683 __ add(count, addr, count);
684 // Use two shifts to clear out those low order two bits! (Cannot opt. into 1.)
685 __ srdi(addr, addr, CardTableModRefBS::card_shift);
686 __ srdi(count, count, CardTableModRefBS::card_shift);
687 __ subf(count, addr, count);
688 assert_different_registers(R0, addr, count, tmp);
689 __ load_const(tmp, (address)ct->byte_map_base);
690 __ addic_(count, count, 1);
691 __ beq(CCR0, Lskip_loop);
692 __ li(R0, 0);
693 __ mtctr(count);
694 // Byte store loop
695 __ bind(Lstore_loop);
696 __ stbx(R0, tmp, addr);
697 __ addi(addr, addr, 1);
698 __ bdnz(Lstore_loop);
699 __ bind(Lskip_loop);
700 }
701 break;
702 case BarrierSet::ModRef:
703 break;
704 default:
705 ShouldNotReachHere();
706 }
707 }
708
709 // Support for void zero_words_aligned8(HeapWord* to, size_t count)
710 //
711 // Arguments:
712 // to:
713 // count:
714 //
715 // Destroys:
716 //
717 address generate_zero_words_aligned8() {
718 StubCodeMark mark(this, "StubRoutines", "zero_words_aligned8");
719
720 // Implemented as in ClearArray.
721 address start = __ emit_fd();
722
723 Register base_ptr_reg = R3_ARG1; // tohw (needs to be 8b aligned)
724 Register cnt_dwords_reg = R4_ARG2; // count (in dwords)
725 Register tmp1_reg = R5_ARG3;
726 Register tmp2_reg = R6_ARG4;
727 Register zero_reg = R7_ARG5;
728
729 // Procedure for large arrays (uses data cache block zero instruction).
730 Label dwloop, fast, fastloop, restloop, lastdword, done;
731 int cl_size=VM_Version::get_cache_line_size(), cl_dwords=cl_size>>3, cl_dwordaddr_bits=exact_log2(cl_dwords);
732 int min_dcbz=2; // Needs to be positive, apply dcbz only to at least min_dcbz cache lines.
733
734 // Clear up to 128byte boundary if long enough, dword_cnt=(16-(base>>3))%16.
735 __ dcbtst(base_ptr_reg); // Indicate write access to first cache line ...
736 __ andi(tmp2_reg, cnt_dwords_reg, 1); // to check if number of dwords is even.
737 __ srdi_(tmp1_reg, cnt_dwords_reg, 1); // number of double dwords
738 __ load_const_optimized(zero_reg, 0L); // Use as zero register.
739
740 __ cmpdi(CCR1, tmp2_reg, 0); // cnt_dwords even?
741 __ beq(CCR0, lastdword); // size <= 1
742 __ mtctr(tmp1_reg); // Speculatively preload counter for rest loop (>0).
743 __ cmpdi(CCR0, cnt_dwords_reg, (min_dcbz+1)*cl_dwords-1); // Big enough to ensure >=min_dcbz cache lines are included?
744 __ neg(tmp1_reg, base_ptr_reg); // bit 0..58: bogus, bit 57..60: (16-(base>>3))%16, bit 61..63: 000
745
746 __ blt(CCR0, restloop); // Too small. (<31=(2*cl_dwords)-1 is sufficient, but bigger performs better.)
747 __ rldicl_(tmp1_reg, tmp1_reg, 64-3, 64-cl_dwordaddr_bits); // Extract number of dwords to 128byte boundary=(16-(base>>3))%16.
748
749 __ beq(CCR0, fast); // already 128byte aligned
750 __ mtctr(tmp1_reg); // Set ctr to hit 128byte boundary (0<ctr<cnt).
751 __ subf(cnt_dwords_reg, tmp1_reg, cnt_dwords_reg); // rest (>0 since size>=256-8)
752
753 // Clear in first cache line dword-by-dword if not already 128byte aligned.
754 __ bind(dwloop);
755 __ std(zero_reg, 0, base_ptr_reg); // Clear 8byte aligned block.
756 __ addi(base_ptr_reg, base_ptr_reg, 8);
757 __ bdnz(dwloop);
758
759 // clear 128byte blocks
760 __ bind(fast);
761 __ srdi(tmp1_reg, cnt_dwords_reg, cl_dwordaddr_bits); // loop count for 128byte loop (>0 since size>=256-8)
762 __ andi(tmp2_reg, cnt_dwords_reg, 1); // to check if rest even
763
764 __ mtctr(tmp1_reg); // load counter
765 __ cmpdi(CCR1, tmp2_reg, 0); // rest even?
766 __ rldicl_(tmp1_reg, cnt_dwords_reg, 63, 65-cl_dwordaddr_bits); // rest in double dwords
767
768 __ bind(fastloop);
769 __ dcbz(base_ptr_reg); // Clear 128byte aligned block.
770 __ addi(base_ptr_reg, base_ptr_reg, cl_size);
771 __ bdnz(fastloop);
772
773 //__ dcbtst(base_ptr_reg); // Indicate write access to last cache line.
774 __ beq(CCR0, lastdword); // rest<=1
775 __ mtctr(tmp1_reg); // load counter
776
777 // Clear rest.
778 __ bind(restloop);
779 __ std(zero_reg, 0, base_ptr_reg); // Clear 8byte aligned block.
780 __ std(zero_reg, 8, base_ptr_reg); // Clear 8byte aligned block.
781 __ addi(base_ptr_reg, base_ptr_reg, 16);
782 __ bdnz(restloop);
783
784 __ bind(lastdword);
785 __ beq(CCR1, done);
786 __ std(zero_reg, 0, base_ptr_reg);
787 __ bind(done);
788 __ blr(); // return
789
790 return start;
791 }
792
793 // The following routine generates a subroutine to throw an asynchronous
794 // UnknownError when an unsafe access gets a fault that could not be
795 // reasonably prevented by the programmer. (Example: SIGBUS/OBJERR.)
796 //
797 address generate_handler_for_unsafe_access() {
798 StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access");
799 address start = __ emit_fd();
800 __ unimplemented("StubRoutines::handler_for_unsafe_access", 93);
801 return start;
802 }
803
804 #if !defined(PRODUCT)
805 // Wrapper which calls oopDesc::is_oop_or_null()
806 // Only called by MacroAssembler::verify_oop
807 static void verify_oop_helper(const char* message, oop o) {
808 if (!o->is_oop_or_null()) {
809 fatal(message);
810 }
811 ++ StubRoutines::_verify_oop_count;
812 }
813 #endif
814
815 // Return address of code to be called from code generated by
816 // MacroAssembler::verify_oop.
817 //
818 // Don't generate, rather use C++ code.
819 address generate_verify_oop() {
820 StubCodeMark mark(this, "StubRoutines", "verify_oop");
821
822 // this is actually a `FunctionDescriptor*'.
823 address start = 0;
824
825 #if !defined(PRODUCT)
826 start = CAST_FROM_FN_PTR(address, verify_oop_helper);
827 #endif
828
829 return start;
830 }
831
832 // Fairer handling of safepoints for native methods.
833 //
834 // Generate code which reads from the polling page. This special handling is needed as the
835 // linux-ppc64 kernel before 2.6.6 doesn't set si_addr on some segfaults in 64bit mode
836 // (cf. http://www.kernel.org/pub/linux/kernel/v2.6/ChangeLog-2.6.6), especially when we try
837 // to read from the safepoint polling page.
838 address generate_load_from_poll() {
839 StubCodeMark mark(this, "StubRoutines", "generate_load_from_poll");
840 address start = __ emit_fd();
841 __ unimplemented("StubRoutines::verify_oop", 95); // TODO PPC port
842 return start;
843 }
844
845 // -XX:+OptimizeFill : convert fill/copy loops into intrinsic
846 //
847 // The code is implemented(ported from sparc) as we believe it benefits JVM98, however
848 // tracing(-XX:+TraceOptimizeFill) shows the intrinsic replacement doesn't happen at all!
849 //
850 // Source code in function is_range_check_if() shows OptimizeFill relaxed the condition
851 // for turning on loop predication optimization, and hence the behavior of "array range check"
852 // and "loop invariant check" could be influenced, which potentially boosted JVM98.
853 //
854 // We leave the code here and see if Oracle has updates in later releases(later than HS20).
855 //
856 // Generate stub for disjoint short fill. If "aligned" is true, the
857 // "to" address is assumed to be heapword aligned.
858 //
859 // Arguments for generated stub:
860 // to: R3_ARG1
861 // value: R4_ARG2
862 // count: R5_ARG3 treated as signed
863 //
864 address generate_fill(BasicType t, bool aligned, const char* name) {
865 StubCodeMark mark(this, "StubRoutines", name);
866 address start = __ emit_fd();
867
868 const Register to = R3_ARG1; // source array address
869 const Register value = R4_ARG2; // fill value
870 const Register count = R5_ARG3; // elements count
871 const Register temp = R6_ARG4; // temp register
872
873 //assert_clean_int(count, O3); // Make sure 'count' is clean int.
874
875 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
876 Label L_fill_2_bytes, L_fill_4_bytes, L_fill_elements, L_fill_32_bytes;
877
878 int shift = -1;
879 switch (t) {
880 case T_BYTE:
881 shift = 2;
882 // clone bytes (zero extend not needed because store instructions below ignore high order bytes)
883 __ rldimi(value, value, 8, 48); // 8 bit -> 16 bit
884 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element
885 __ blt(CCR0, L_fill_elements);
886 __ rldimi(value, value, 16, 32); // 16 bit -> 32 bit
887 break;
888 case T_SHORT:
889 shift = 1;
890 // clone bytes (zero extend not needed because store instructions below ignore high order bytes)
891 __ rldimi(value, value, 16, 32); // 16 bit -> 32 bit
892 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element
893 __ blt(CCR0, L_fill_elements);
894 break;
895 case T_INT:
896 shift = 0;
897 __ cmpdi(CCR0, count, 2<<shift); // Short arrays (< 8 bytes) fill by element
898 __ blt(CCR0, L_fill_4_bytes);
899 break;
900 default: ShouldNotReachHere();
901 }
902
903 if (!aligned && (t == T_BYTE || t == T_SHORT)) {
904 // align source address at 4 bytes address boundary
905 if (t == T_BYTE) {
906 // One byte misalignment happens only for byte arrays
907 __ andi_(temp, to, 1);
908 __ beq(CCR0, L_skip_align1);
909 __ stb(value, 0, to);
910 __ addi(to, to, 1);
911 __ addi(count, count, -1);
912 __ bind(L_skip_align1);
913 }
914 // Two bytes misalignment happens only for byte and short (char) arrays.
915 __ andi_(temp, to, 2);
916 __ beq(CCR0, L_skip_align2);
917 __ sth(value, 0, to);
918 __ addi(to, to, 2);
919 __ addi(count, count, -(1 << (shift - 1)));
920 __ bind(L_skip_align2);
921 }
922
923 if (!aligned) {
924 // Align to 8 bytes, we know we are 4 byte aligned to start.
925 __ andi_(temp, to, 7);
926 __ beq(CCR0, L_fill_32_bytes);
927 __ stw(value, 0, to);
928 __ addi(to, to, 4);
929 __ addi(count, count, -(1 << shift));
930 __ bind(L_fill_32_bytes);
931 }
932
933 __ li(temp, 8<<shift); // prepare for 32 byte loop
934 // clone bytes int->long as above
935 __ rldimi(value, value, 32, 0); // 32 bit -> 64 bit
936
937 Label L_check_fill_8_bytes;
938 // Fill 32-byte chunks
939 __ subf_(count, temp, count);
940 __ blt(CCR0, L_check_fill_8_bytes);
941
942 Label L_fill_32_bytes_loop;
943 __ align(32);
944 __ bind(L_fill_32_bytes_loop);
945
946 __ std(value, 0, to);
947 __ std(value, 8, to);
948 __ subf_(count, temp, count); // update count
949 __ std(value, 16, to);
950 __ std(value, 24, to);
951
952 __ addi(to, to, 32);
953 __ bge(CCR0, L_fill_32_bytes_loop);
954
955 __ bind(L_check_fill_8_bytes);
956 __ add_(count, temp, count);
957 __ beq(CCR0, L_exit);
958 __ addic_(count, count, -(2 << shift));
959 __ blt(CCR0, L_fill_4_bytes);
960
961 //
962 // Length is too short, just fill 8 bytes at a time.
963 //
964 Label L_fill_8_bytes_loop;
965 __ bind(L_fill_8_bytes_loop);
966 __ std(value, 0, to);
967 __ addic_(count, count, -(2 << shift));
968 __ addi(to, to, 8);
969 __ bge(CCR0, L_fill_8_bytes_loop);
970
971 // fill trailing 4 bytes
972 __ bind(L_fill_4_bytes);
973 __ andi_(temp, count, 1<<shift);
974 __ beq(CCR0, L_fill_2_bytes);
975
976 __ stw(value, 0, to);
977 if (t == T_BYTE || t == T_SHORT) {
978 __ addi(to, to, 4);
979 // fill trailing 2 bytes
980 __ bind(L_fill_2_bytes);
981 __ andi_(temp, count, 1<<(shift-1));
982 __ beq(CCR0, L_fill_byte);
983 __ sth(value, 0, to);
984 if (t == T_BYTE) {
985 __ addi(to, to, 2);
986 // fill trailing byte
987 __ bind(L_fill_byte);
988 __ andi_(count, count, 1);
989 __ beq(CCR0, L_exit);
990 __ stb(value, 0, to);
991 } else {
992 __ bind(L_fill_byte);
993 }
994 } else {
995 __ bind(L_fill_2_bytes);
996 }
997 __ bind(L_exit);
998 __ blr();
999
1000 // Handle copies less than 8 bytes. Int is handled elsewhere.
1001 if (t == T_BYTE) {
1002 __ bind(L_fill_elements);
1003 Label L_fill_2, L_fill_4;
1004 __ andi_(temp, count, 1);
1005 __ beq(CCR0, L_fill_2);
1006 __ stb(value, 0, to);
1007 __ addi(to, to, 1);
1008 __ bind(L_fill_2);
1009 __ andi_(temp, count, 2);
1010 __ beq(CCR0, L_fill_4);
1011 __ stb(value, 0, to);
1012 __ stb(value, 0, to);
1013 __ addi(to, to, 2);
1014 __ bind(L_fill_4);
1015 __ andi_(temp, count, 4);
1016 __ beq(CCR0, L_exit);
1017 __ stb(value, 0, to);
1018 __ stb(value, 1, to);
1019 __ stb(value, 2, to);
1020 __ stb(value, 3, to);
1021 __ blr();
1022 }
1023
1024 if (t == T_SHORT) {
1025 Label L_fill_2;
1026 __ bind(L_fill_elements);
1027 __ andi_(temp, count, 1);
1028 __ beq(CCR0, L_fill_2);
1029 __ sth(value, 0, to);
1030 __ addi(to, to, 2);
1031 __ bind(L_fill_2);
1032 __ andi_(temp, count, 2);
1033 __ beq(CCR0, L_exit);
1034 __ sth(value, 0, to);
1035 __ sth(value, 2, to);
1036 __ blr();
1037 }
1038 return start;
1039 }
1040
1041
1042 // Generate overlap test for array copy stubs
1043 //
1044 // Input:
1045 // R3_ARG1 - from
1046 // R4_ARG2 - to
1047 // R5_ARG3 - element count
1048 //
1049 void array_overlap_test(address no_overlap_target, int log2_elem_size) {
1050 Register tmp1 = R6_ARG4;
1051 Register tmp2 = R7_ARG5;
1052
1053 Label l_overlap;
1054 #ifdef ASSERT
1055 __ srdi_(tmp2, R5_ARG3, 31);
1056 __ asm_assert_eq("missing zero extend", 0xAFFE);
1057 #endif
1058
1059 __ subf(tmp1, R3_ARG1, R4_ARG2); // distance in bytes
1060 __ sldi(tmp2, R5_ARG3, log2_elem_size); // size in bytes
1061 __ cmpld(CCR0, R3_ARG1, R4_ARG2); // Use unsigned comparison!
1062 __ cmpld(CCR1, tmp1, tmp2);
1063 __ crand(/*CCR0 lt*/0, /*CCR1 lt*/4+0, /*CCR0 lt*/0);
1064 __ blt(CCR0, l_overlap); // Src before dst and distance smaller than size.
1065
1066 // need to copy forwards
1067 if (__ is_within_range_of_b(no_overlap_target, __ pc())) {
1068 __ b(no_overlap_target);
1069 } else {
1070 __ load_const(tmp1, no_overlap_target, tmp2);
1071 __ mtctr(tmp1);
1072 __ bctr();
1073 }
1074
1075 __ bind(l_overlap);
1076 // need to copy backwards
1077 }
1078
1079 // The guideline in the implementations of generate_disjoint_xxx_copy
1080 // (xxx=byte,short,int,long,oop) is to copy as many elements as possible with
1081 // single instructions, but to avoid alignment interrupts (see subsequent
1082 // comment). Furthermore, we try to minimize misaligned access, even
1083 // though they cause no alignment interrupt.
1084 //
1085 // In Big-Endian mode, the PowerPC architecture requires implementations to
1086 // handle automatically misaligned integer halfword and word accesses,
1087 // word-aligned integer doubleword accesses, and word-aligned floating-point
1088 // accesses. Other accesses may or may not generate an Alignment interrupt
1089 // depending on the implementation.
1090 // Alignment interrupt handling may require on the order of hundreds of cycles,
1091 // so every effort should be made to avoid misaligned memory values.
1092 //
1093 //
1094 // Generate stub for disjoint byte copy. If "aligned" is true, the
1095 // "from" and "to" addresses are assumed to be heapword aligned.
1096 //
1097 // Arguments for generated stub:
1098 // from: R3_ARG1
1099 // to: R4_ARG2
1100 // count: R5_ARG3 treated as signed
1101 //
1102 address generate_disjoint_byte_copy(bool aligned, const char * name) {
1103 StubCodeMark mark(this, "StubRoutines", name);
1104 address start = __ emit_fd();
1105
1106 Register tmp1 = R6_ARG4;
1107 Register tmp2 = R7_ARG5;
1108 Register tmp3 = R8_ARG6;
1109 Register tmp4 = R9_ARG7;
1110
1111
1112 Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8, l_9;
1113 // Don't try anything fancy if arrays don't have many elements.
1114 __ li(tmp3, 0);
1115 __ cmpwi(CCR0, R5_ARG3, 17);
1116 __ ble(CCR0, l_6); // copy 4 at a time
1117
1118 if (!aligned) {
1119 __ xorr(tmp1, R3_ARG1, R4_ARG2);
1120 __ andi_(tmp1, tmp1, 3);
1121 __ bne(CCR0, l_6); // If arrays don't have the same alignment mod 4, do 4 element copy.
1122
1123 // Copy elements if necessary to align to 4 bytes.
1124 __ neg(tmp1, R3_ARG1); // Compute distance to alignment boundary.
1125 __ andi_(tmp1, tmp1, 3);
1126 __ beq(CCR0, l_2);
1127
1128 __ subf(R5_ARG3, tmp1, R5_ARG3);
1129 __ bind(l_9);
1130 __ lbz(tmp2, 0, R3_ARG1);
1131 __ addic_(tmp1, tmp1, -1);
1132 __ stb(tmp2, 0, R4_ARG2);
1133 __ addi(R3_ARG1, R3_ARG1, 1);
1134 __ addi(R4_ARG2, R4_ARG2, 1);
1135 __ bne(CCR0, l_9);
1136
1137 __ bind(l_2);
1138 }
1139
1140 // copy 8 elements at a time
1141 __ xorr(tmp2, R3_ARG1, R4_ARG2); // skip if src & dest have differing alignment mod 8
1142 __ andi_(tmp1, tmp2, 7);
1143 __ bne(CCR0, l_7); // not same alignment -> to or from is aligned -> copy 8
1144
1145 // copy a 2-element word if necessary to align to 8 bytes
1146 __ andi_(R0, R3_ARG1, 7);
1147 __ beq(CCR0, l_7);
1148
1149 __ lwzx(tmp2, R3_ARG1, tmp3);
1150 __ addi(R5_ARG3, R5_ARG3, -4);
1151 __ stwx(tmp2, R4_ARG2, tmp3);
1152 { // FasterArrayCopy
1153 __ addi(R3_ARG1, R3_ARG1, 4);
1154 __ addi(R4_ARG2, R4_ARG2, 4);
1155 }
1156 __ bind(l_7);
1157
1158 { // FasterArrayCopy
1159 __ cmpwi(CCR0, R5_ARG3, 31);
1160 __ ble(CCR0, l_6); // copy 2 at a time if less than 32 elements remain
1161
1162 __ srdi(tmp1, R5_ARG3, 5);
1163 __ andi_(R5_ARG3, R5_ARG3, 31);
1164 __ mtctr(tmp1);
1165
1166 __ bind(l_8);
1167 // Use unrolled version for mass copying (copy 32 elements a time)
1168 // Load feeding store gets zero latency on Power6, however not on Power5.
1169 // Therefore, the following sequence is made for the good of both.
1170 __ ld(tmp1, 0, R3_ARG1);
1171 __ ld(tmp2, 8, R3_ARG1);
1172 __ ld(tmp3, 16, R3_ARG1);
1173 __ ld(tmp4, 24, R3_ARG1);
1174 __ std(tmp1, 0, R4_ARG2);
1175 __ std(tmp2, 8, R4_ARG2);
1176 __ std(tmp3, 16, R4_ARG2);
1177 __ std(tmp4, 24, R4_ARG2);
1178 __ addi(R3_ARG1, R3_ARG1, 32);
1179 __ addi(R4_ARG2, R4_ARG2, 32);
1180 __ bdnz(l_8);
1181 }
1182
1183 __ bind(l_6);
1184
1185 // copy 4 elements at a time
1186 __ cmpwi(CCR0, R5_ARG3, 4);
1187 __ blt(CCR0, l_1);
1188 __ srdi(tmp1, R5_ARG3, 2);
1189 __ mtctr(tmp1); // is > 0
1190 __ andi_(R5_ARG3, R5_ARG3, 3);
1191
1192 { // FasterArrayCopy
1193 __ addi(R3_ARG1, R3_ARG1, -4);
1194 __ addi(R4_ARG2, R4_ARG2, -4);
1195 __ bind(l_3);
1196 __ lwzu(tmp2, 4, R3_ARG1);
1197 __ stwu(tmp2, 4, R4_ARG2);
1198 __ bdnz(l_3);
1199 __ addi(R3_ARG1, R3_ARG1, 4);
1200 __ addi(R4_ARG2, R4_ARG2, 4);
1201 }
1202
1203 // do single element copy
1204 __ bind(l_1);
1205 __ cmpwi(CCR0, R5_ARG3, 0);
1206 __ beq(CCR0, l_4);
1207
1208 { // FasterArrayCopy
1209 __ mtctr(R5_ARG3);
1210 __ addi(R3_ARG1, R3_ARG1, -1);
1211 __ addi(R4_ARG2, R4_ARG2, -1);
1212
1213 __ bind(l_5);
1214 __ lbzu(tmp2, 1, R3_ARG1);
1215 __ stbu(tmp2, 1, R4_ARG2);
1216 __ bdnz(l_5);
1217 }
1218
1219 __ bind(l_4);
1220 __ blr();
1221
1222 return start;
1223 }
1224
1225 // Generate stub for conjoint byte copy. If "aligned" is true, the
1226 // "from" and "to" addresses are assumed to be heapword aligned.
1227 //
1228 // Arguments for generated stub:
1229 // from: R3_ARG1
1230 // to: R4_ARG2
1231 // count: R5_ARG3 treated as signed
1232 //
1233 address generate_conjoint_byte_copy(bool aligned, const char * name) {
1234 StubCodeMark mark(this, "StubRoutines", name);
1235 address start = __ emit_fd();
1236
1237 Register tmp1 = R6_ARG4;
1238 Register tmp2 = R7_ARG5;
1239 Register tmp3 = R8_ARG6;
1240
1241 address nooverlap_target = aligned ?
1242 ((FunctionDescriptor*)StubRoutines::arrayof_jbyte_disjoint_arraycopy())->entry() :
1243 ((FunctionDescriptor*)StubRoutines::jbyte_disjoint_arraycopy())->entry();
1244
1245 array_overlap_test(nooverlap_target, 0);
1246 // Do reverse copy. We assume the case of actual overlap is rare enough
1247 // that we don't have to optimize it.
1248 Label l_1, l_2;
1249
1250 __ b(l_2);
1251 __ bind(l_1);
1252 __ stbx(tmp1, R4_ARG2, R5_ARG3);
1253 __ bind(l_2);
1254 __ addic_(R5_ARG3, R5_ARG3, -1);
1255 __ lbzx(tmp1, R3_ARG1, R5_ARG3);
1256 __ bge(CCR0, l_1);
1257
1258 __ blr();
1259
1260 return start;
1261 }
1262
1263 // Generate stub for disjoint short copy. If "aligned" is true, the
1264 // "from" and "to" addresses are assumed to be heapword aligned.
1265 //
1266 // Arguments for generated stub:
1267 // from: R3_ARG1
1268 // to: R4_ARG2
1269 // elm.count: R5_ARG3 treated as signed
1270 //
1271 // Strategy for aligned==true:
1272 //
1273 // If length <= 9:
1274 // 1. copy 2 elements at a time (l_6)
1275 // 2. copy last element if original element count was odd (l_1)
1276 //
1277 // If length > 9:
1278 // 1. copy 4 elements at a time until less than 4 elements are left (l_7)
1279 // 2. copy 2 elements at a time until less than 2 elements are left (l_6)
1280 // 3. copy last element if one was left in step 2. (l_1)
1281 //
1282 //
1283 // Strategy for aligned==false:
1284 //
1285 // If length <= 9: same as aligned==true case, but NOTE: load/stores
1286 // can be unaligned (see comment below)
1287 //
1288 // If length > 9:
1289 // 1. continue with step 6. if the alignment of from and to mod 4
1290 // is different.
1291 // 2. align from and to to 4 bytes by copying 1 element if necessary
1292 // 3. at l_2 from and to are 4 byte aligned; continue with
1293 // 5. if they cannot be aligned to 8 bytes because they have
1294 // got different alignment mod 8.
1295 // 4. at this point we know that both, from and to, have the same
1296 // alignment mod 8, now copy one element if necessary to get
1297 // 8 byte alignment of from and to.
1298 // 5. copy 4 elements at a time until less than 4 elements are
1299 // left; depending on step 3. all load/stores are aligned or
1300 // either all loads or all stores are unaligned.
1301 // 6. copy 2 elements at a time until less than 2 elements are
1302 // left (l_6); arriving here from step 1., there is a chance
1303 // that all accesses are unaligned.
1304 // 7. copy last element if one was left in step 6. (l_1)
1305 //
1306 // There are unaligned data accesses using integer load/store
1307 // instructions in this stub. POWER allows such accesses.
1308 //
1309 // According to the manuals (PowerISA_V2.06_PUBLIC, Book II,
1310 // Chapter 2: Effect of Operand Placement on Performance) unaligned
1311 // integer load/stores have good performance. Only unaligned
1312 // floating point load/stores can have poor performance.
1313 //
1314 // TODO:
1315 //
1316 // 1. check if aligning the backbranch target of loops is beneficial
1317 //
1318 address generate_disjoint_short_copy(bool aligned, const char * name) {
1319 StubCodeMark mark(this, "StubRoutines", name);
1320
1321 Register tmp1 = R6_ARG4;
1322 Register tmp2 = R7_ARG5;
1323 Register tmp3 = R8_ARG6;
1324 Register tmp4 = R9_ARG7;
1325
1326 address start = __ emit_fd();
1327
1328 Label l_1, l_2, l_3, l_4, l_5, l_6, l_7, l_8;
1329 // don't try anything fancy if arrays don't have many elements
1330 __ li(tmp3, 0);
1331 __ cmpwi(CCR0, R5_ARG3, 9);
1332 __ ble(CCR0, l_6); // copy 2 at a time
1333
1334 if (!aligned) {
1335 __ xorr(tmp1, R3_ARG1, R4_ARG2);
1336 __ andi_(tmp1, tmp1, 3);
1337 __ bne(CCR0, l_6); // if arrays don't have the same alignment mod 4, do 2 element copy
1338
1339 // At this point it is guaranteed that both, from and to have the same alignment mod 4.
1340
1341 // Copy 1 element if necessary to align to 4 bytes.
1342 __ andi_(tmp1, R3_ARG1, 3);
1343 __ beq(CCR0, l_2);
1344
1345 __ lhz(tmp2, 0, R3_ARG1);
1346 __ addi(R3_ARG1, R3_ARG1, 2);
1347 __ sth(tmp2, 0, R4_ARG2);
1348 __ addi(R4_ARG2, R4_ARG2, 2);
1349 __ addi(R5_ARG3, R5_ARG3, -1);
1350 __ bind(l_2);
1351
1352 // At this point the positions of both, from and to, are at least 4 byte aligned.
1353
1354 // Copy 4 elements at a time.
1355 // Align to 8 bytes, but only if both, from and to, have same alignment mod 8.
1356 __ xorr(tmp2, R3_ARG1, R4_ARG2);
1357 __ andi_(tmp1, tmp2, 7);
1358 __ bne(CCR0, l_7); // not same alignment mod 8 -> copy 4, either from or to will be unaligned
1359
1360 // Copy a 2-element word if necessary to align to 8 bytes.
1361 __ andi_(R0, R3_ARG1, 7);
1362 __ beq(CCR0, l_7);
1363
1364 __ lwzx(tmp2, R3_ARG1, tmp3);
1365 __ addi(R5_ARG3, R5_ARG3, -2);
1366 __ stwx(tmp2, R4_ARG2, tmp3);
1367 { // FasterArrayCopy
1368 __ addi(R3_ARG1, R3_ARG1, 4);
1369 __ addi(R4_ARG2, R4_ARG2, 4);
1370 }
1371 }
1372
1373 __ bind(l_7);
1374
1375 // Copy 4 elements at a time; either the loads or the stores can
1376 // be unaligned if aligned == false.
1377
1378 { // FasterArrayCopy
1379 __ cmpwi(CCR0, R5_ARG3, 15);
1380 __ ble(CCR0, l_6); // copy 2 at a time if less than 16 elements remain
1381
1382 __ srdi(tmp1, R5_ARG3, 4);
1383 __ andi_(R5_ARG3, R5_ARG3, 15);
1384 __ mtctr(tmp1);
1385
1386 __ bind(l_8);
1387 // Use unrolled version for mass copying (copy 16 elements a time).
1388 // Load feeding store gets zero latency on Power6, however not on Power5.
1389 // Therefore, the following sequence is made for the good of both.
1390 __ ld(tmp1, 0, R3_ARG1);
1391 __ ld(tmp2, 8, R3_ARG1);
1392 __ ld(tmp3, 16, R3_ARG1);
1393 __ ld(tmp4, 24, R3_ARG1);
1394 __ std(tmp1, 0, R4_ARG2);
1395 __ std(tmp2, 8, R4_ARG2);
1396 __ std(tmp3, 16, R4_ARG2);
1397 __ std(tmp4, 24, R4_ARG2);
1398 __ addi(R3_ARG1, R3_ARG1, 32);
1399 __ addi(R4_ARG2, R4_ARG2, 32);
1400 __ bdnz(l_8);
1401 }
1402 __ bind(l_6);
1403
1404 // copy 2 elements at a time
1405 { // FasterArrayCopy
1406 __ cmpwi(CCR0, R5_ARG3, 2);
1407 __ blt(CCR0, l_1);
1408 __ srdi(tmp1, R5_ARG3, 1);
1409 __ andi_(R5_ARG3, R5_ARG3, 1);
1410
1411 __ addi(R3_ARG1, R3_ARG1, -4);
1412 __ addi(R4_ARG2, R4_ARG2, -4);
1413 __ mtctr(tmp1);
1414
1415 __ bind(l_3);
1416 __ lwzu(tmp2, 4, R3_ARG1);
1417 __ stwu(tmp2, 4, R4_ARG2);
1418 __ bdnz(l_3);
1419
1420 __ addi(R3_ARG1, R3_ARG1, 4);
1421 __ addi(R4_ARG2, R4_ARG2, 4);
1422 }
1423
1424 // do single element copy
1425 __ bind(l_1);
1426 __ cmpwi(CCR0, R5_ARG3, 0);
1427 __ beq(CCR0, l_4);
1428
1429 { // FasterArrayCopy
1430 __ mtctr(R5_ARG3);
1431 __ addi(R3_ARG1, R3_ARG1, -2);
1432 __ addi(R4_ARG2, R4_ARG2, -2);
1433
1434 __ bind(l_5);
1435 __ lhzu(tmp2, 2, R3_ARG1);
1436 __ sthu(tmp2, 2, R4_ARG2);
1437 __ bdnz(l_5);
1438 }
1439 __ bind(l_4);
1440 __ blr();
1441
1442 return start;
1443 }
1444
1445 // Generate stub for conjoint short copy. If "aligned" is true, the
1446 // "from" and "to" addresses are assumed to be heapword aligned.
1447 //
1448 // Arguments for generated stub:
1449 // from: R3_ARG1
1450 // to: R4_ARG2
1451 // count: R5_ARG3 treated as signed
1452 //
1453 address generate_conjoint_short_copy(bool aligned, const char * name) {
1454 StubCodeMark mark(this, "StubRoutines", name);
1455 address start = __ emit_fd();
1456
1457 Register tmp1 = R6_ARG4;
1458 Register tmp2 = R7_ARG5;
1459 Register tmp3 = R8_ARG6;
1460
1461 address nooverlap_target = aligned ?
1462 ((FunctionDescriptor*)StubRoutines::arrayof_jshort_disjoint_arraycopy())->entry() :
1463 ((FunctionDescriptor*)StubRoutines::jshort_disjoint_arraycopy())->entry();
1464
1465 array_overlap_test(nooverlap_target, 1);
1466
1467 Label l_1, l_2;
1468 __ sldi(tmp1, R5_ARG3, 1);
1469 __ b(l_2);
1470 __ bind(l_1);
1471 __ sthx(tmp2, R4_ARG2, tmp1);
1472 __ bind(l_2);
1473 __ addic_(tmp1, tmp1, -2);
1474 __ lhzx(tmp2, R3_ARG1, tmp1);
1475 __ bge(CCR0, l_1);
1476
1477 __ blr();
1478
1479 return start;
1480 }
1481
1482 // Generate core code for disjoint int copy (and oop copy on 32-bit). If "aligned"
1483 // is true, the "from" and "to" addresses are assumed to be heapword aligned.
1484 //
1485 // Arguments:
1486 // from: R3_ARG1
1487 // to: R4_ARG2
1488 // count: R5_ARG3 treated as signed
1489 //
1490 void generate_disjoint_int_copy_core(bool aligned) {
1491 Register tmp1 = R6_ARG4;
1492 Register tmp2 = R7_ARG5;
1493 Register tmp3 = R8_ARG6;
1494 Register tmp4 = R0;
1495
1496 Label l_1, l_2, l_3, l_4, l_5, l_6;
1497 // for short arrays, just do single element copy
1498 __ li(tmp3, 0);
1499 __ cmpwi(CCR0, R5_ARG3, 5);
1500 __ ble(CCR0, l_2);
1501
1502 if (!aligned) {
1503 // check if arrays have same alignment mod 8.
1504 __ xorr(tmp1, R3_ARG1, R4_ARG2);
1505 __ andi_(R0, tmp1, 7);
1506 // Not the same alignment, but ld and std just need to be 4 byte aligned.
1507 __ bne(CCR0, l_4); // to OR from is 8 byte aligned -> copy 2 at a time
1508
1509 // copy 1 element to align to and from on an 8 byte boundary
1510 __ andi_(R0, R3_ARG1, 7);
1511 __ beq(CCR0, l_4);
1512
1513 __ lwzx(tmp2, R3_ARG1, tmp3);
1514 __ addi(R5_ARG3, R5_ARG3, -1);
1515 __ stwx(tmp2, R4_ARG2, tmp3);
1516 { // FasterArrayCopy
1517 __ addi(R3_ARG1, R3_ARG1, 4);
1518 __ addi(R4_ARG2, R4_ARG2, 4);
1519 }
1520 __ bind(l_4);
1521 }
1522
1523 { // FasterArrayCopy
1524 __ cmpwi(CCR0, R5_ARG3, 7);
1525 __ ble(CCR0, l_2); // copy 1 at a time if less than 8 elements remain
1526
1527 __ srdi(tmp1, R5_ARG3, 3);
1528 __ andi_(R5_ARG3, R5_ARG3, 7);
1529 __ mtctr(tmp1);
1530
1531 __ bind(l_6);
1532 // Use unrolled version for mass copying (copy 8 elements a time).
1533 // Load feeding store gets zero latency on power6, however not on power 5.
1534 // Therefore, the following sequence is made for the good of both.
1535 __ ld(tmp1, 0, R3_ARG1);
1536 __ ld(tmp2, 8, R3_ARG1);
1537 __ ld(tmp3, 16, R3_ARG1);
1538 __ ld(tmp4, 24, R3_ARG1);
1539 __ std(tmp1, 0, R4_ARG2);
1540 __ std(tmp2, 8, R4_ARG2);
1541 __ std(tmp3, 16, R4_ARG2);
1542 __ std(tmp4, 24, R4_ARG2);
1543 __ addi(R3_ARG1, R3_ARG1, 32);
1544 __ addi(R4_ARG2, R4_ARG2, 32);
1545 __ bdnz(l_6);
1546 }
1547
1548 // copy 1 element at a time
1549 __ bind(l_2);
1550 __ cmpwi(CCR0, R5_ARG3, 0);
1551 __ beq(CCR0, l_1);
1552
1553 { // FasterArrayCopy
1554 __ mtctr(R5_ARG3);
1555 __ addi(R3_ARG1, R3_ARG1, -4);
1556 __ addi(R4_ARG2, R4_ARG2, -4);
1557
1558 __ bind(l_3);
1559 __ lwzu(tmp2, 4, R3_ARG1);
1560 __ stwu(tmp2, 4, R4_ARG2);
1561 __ bdnz(l_3);
1562 }
1563
1564 __ bind(l_1);
1565 return;
1566 }
1567
1568 // Generate stub for disjoint int copy. If "aligned" is true, the
1569 // "from" and "to" addresses are assumed to be heapword aligned.
1570 //
1571 // Arguments for generated stub:
1572 // from: R3_ARG1
1573 // to: R4_ARG2
1574 // count: R5_ARG3 treated as signed
1575 //
1576 address generate_disjoint_int_copy(bool aligned, const char * name) {
1577 StubCodeMark mark(this, "StubRoutines", name);
1578 address start = __ emit_fd();
1579 generate_disjoint_int_copy_core(aligned);
1580 __ blr();
1581 return start;
1582 }
1583
1584 // Generate core code for conjoint int copy (and oop copy on
1585 // 32-bit). If "aligned" is true, the "from" and "to" addresses
1586 // are assumed to be heapword aligned.
1587 //
1588 // Arguments:
1589 // from: R3_ARG1
1590 // to: R4_ARG2
1591 // count: R5_ARG3 treated as signed
1592 //
1593 void generate_conjoint_int_copy_core(bool aligned) {
1594 // Do reverse copy. We assume the case of actual overlap is rare enough
1595 // that we don't have to optimize it.
1596
1597 Label l_1, l_2, l_3, l_4, l_5, l_6;
1598
1599 Register tmp1 = R6_ARG4;
1600 Register tmp2 = R7_ARG5;
1601 Register tmp3 = R8_ARG6;
1602 Register tmp4 = R0;
1603
1604 { // FasterArrayCopy
1605 __ cmpwi(CCR0, R5_ARG3, 0);
1606 __ beq(CCR0, l_6);
1607
1608 __ sldi(R5_ARG3, R5_ARG3, 2);
1609 __ add(R3_ARG1, R3_ARG1, R5_ARG3);
1610 __ add(R4_ARG2, R4_ARG2, R5_ARG3);
1611 __ srdi(R5_ARG3, R5_ARG3, 2);
1612
1613 __ cmpwi(CCR0, R5_ARG3, 7);
1614 __ ble(CCR0, l_5); // copy 1 at a time if less than 8 elements remain
1615
1616 __ srdi(tmp1, R5_ARG3, 3);
1617 __ andi(R5_ARG3, R5_ARG3, 7);
1618 __ mtctr(tmp1);
1619
1620 __ bind(l_4);
1621 // Use unrolled version for mass copying (copy 4 elements a time).
1622 // Load feeding store gets zero latency on Power6, however not on Power5.
1623 // Therefore, the following sequence is made for the good of both.
1624 __ addi(R3_ARG1, R3_ARG1, -32);
1625 __ addi(R4_ARG2, R4_ARG2, -32);
1626 __ ld(tmp4, 24, R3_ARG1);
1627 __ ld(tmp3, 16, R3_ARG1);
1628 __ ld(tmp2, 8, R3_ARG1);
1629 __ ld(tmp1, 0, R3_ARG1);
1630 __ std(tmp4, 24, R4_ARG2);
1631 __ std(tmp3, 16, R4_ARG2);
1632 __ std(tmp2, 8, R4_ARG2);
1633 __ std(tmp1, 0, R4_ARG2);
1634 __ bdnz(l_4);
1635
1636 __ cmpwi(CCR0, R5_ARG3, 0);
1637 __ beq(CCR0, l_6);
1638
1639 __ bind(l_5);
1640 __ mtctr(R5_ARG3);
1641 __ bind(l_3);
1642 __ lwz(R0, -4, R3_ARG1);
1643 __ stw(R0, -4, R4_ARG2);
1644 __ addi(R3_ARG1, R3_ARG1, -4);
1645 __ addi(R4_ARG2, R4_ARG2, -4);
1646 __ bdnz(l_3);
1647
1648 __ bind(l_6);
1649 }
1650 }
1651
1652 // Generate stub for conjoint int copy. If "aligned" is true, the
1653 // "from" and "to" addresses are assumed to be heapword aligned.
1654 //
1655 // Arguments for generated stub:
1656 // from: R3_ARG1
1657 // to: R4_ARG2
1658 // count: R5_ARG3 treated as signed
1659 //
1660 address generate_conjoint_int_copy(bool aligned, const char * name) {
1661 StubCodeMark mark(this, "StubRoutines", name);
1662 address start = __ emit_fd();
1663
1664 address nooverlap_target = aligned ?
1665 ((FunctionDescriptor*)StubRoutines::arrayof_jint_disjoint_arraycopy())->entry() :
1666 ((FunctionDescriptor*)StubRoutines::jint_disjoint_arraycopy())->entry();
1667
1668 array_overlap_test(nooverlap_target, 2);
1669
1670 generate_conjoint_int_copy_core(aligned);
1671
1672 __ blr();
1673
1674 return start;
1675 }
1676
1677 // Generate core code for disjoint long copy (and oop copy on
1678 // 64-bit). If "aligned" is true, the "from" and "to" addresses
1679 // are assumed to be heapword aligned.
1680 //
1681 // Arguments:
1682 // from: R3_ARG1
1683 // to: R4_ARG2
1684 // count: R5_ARG3 treated as signed
1685 //
1686 void generate_disjoint_long_copy_core(bool aligned) {
1687 Register tmp1 = R6_ARG4;
1688 Register tmp2 = R7_ARG5;
1689 Register tmp3 = R8_ARG6;
1690 Register tmp4 = R0;
1691
1692 Label l_1, l_2, l_3, l_4;
1693
1694 { // FasterArrayCopy
1695 __ cmpwi(CCR0, R5_ARG3, 3);
1696 __ ble(CCR0, l_3); // copy 1 at a time if less than 4 elements remain
1697
1698 __ srdi(tmp1, R5_ARG3, 2);
1699 __ andi_(R5_ARG3, R5_ARG3, 3);
1700 __ mtctr(tmp1);
1701
1702 __ bind(l_4);
1703 // Use unrolled version for mass copying (copy 4 elements a time).
1704 // Load feeding store gets zero latency on Power6, however not on Power5.
1705 // Therefore, the following sequence is made for the good of both.
1706 __ ld(tmp1, 0, R3_ARG1);
1707 __ ld(tmp2, 8, R3_ARG1);
1708 __ ld(tmp3, 16, R3_ARG1);
1709 __ ld(tmp4, 24, R3_ARG1);
1710 __ std(tmp1, 0, R4_ARG2);
1711 __ std(tmp2, 8, R4_ARG2);
1712 __ std(tmp3, 16, R4_ARG2);
1713 __ std(tmp4, 24, R4_ARG2);
1714 __ addi(R3_ARG1, R3_ARG1, 32);
1715 __ addi(R4_ARG2, R4_ARG2, 32);
1716 __ bdnz(l_4);
1717 }
1718
1719 // copy 1 element at a time
1720 __ bind(l_3);
1721 __ cmpwi(CCR0, R5_ARG3, 0);
1722 __ beq(CCR0, l_1);
1723
1724 { // FasterArrayCopy
1725 __ mtctr(R5_ARG3);
1726 __ addi(R3_ARG1, R3_ARG1, -8);
1727 __ addi(R4_ARG2, R4_ARG2, -8);
1728
1729 __ bind(l_2);
1730 __ ldu(R0, 8, R3_ARG1);
1731 __ stdu(R0, 8, R4_ARG2);
1732 __ bdnz(l_2);
1733
1734 }
1735 __ bind(l_1);
1736 }
1737
1738 // Generate stub for disjoint long copy. If "aligned" is true, the
1739 // "from" and "to" addresses are assumed to be heapword aligned.
1740 //
1741 // Arguments for generated stub:
1742 // from: R3_ARG1
1743 // to: R4_ARG2
1744 // count: R5_ARG3 treated as signed
1745 //
1746 address generate_disjoint_long_copy(bool aligned, const char * name) {
1747 StubCodeMark mark(this, "StubRoutines", name);
1748 address start = __ emit_fd();
1749 generate_disjoint_long_copy_core(aligned);
1750 __ blr();
1751
1752 return start;
1753 }
1754
1755 // Generate core code for conjoint long copy (and oop copy on
1756 // 64-bit). If "aligned" is true, the "from" and "to" addresses
1757 // are assumed to be heapword aligned.
1758 //
1759 // Arguments:
1760 // from: R3_ARG1
1761 // to: R4_ARG2
1762 // count: R5_ARG3 treated as signed
1763 //
1764 void generate_conjoint_long_copy_core(bool aligned) {
1765 Register tmp1 = R6_ARG4;
1766 Register tmp2 = R7_ARG5;
1767 Register tmp3 = R8_ARG6;
1768 Register tmp4 = R0;
1769
1770 Label l_1, l_2, l_3, l_4, l_5;
1771
1772 __ cmpwi(CCR0, R5_ARG3, 0);
1773 __ beq(CCR0, l_1);
1774
1775 { // FasterArrayCopy
1776 __ sldi(R5_ARG3, R5_ARG3, 3);
1777 __ add(R3_ARG1, R3_ARG1, R5_ARG3);
1778 __ add(R4_ARG2, R4_ARG2, R5_ARG3);
1779 __ srdi(R5_ARG3, R5_ARG3, 3);
1780
1781 __ cmpwi(CCR0, R5_ARG3, 3);
1782 __ ble(CCR0, l_5); // copy 1 at a time if less than 4 elements remain
1783
1784 __ srdi(tmp1, R5_ARG3, 2);
1785 __ andi(R5_ARG3, R5_ARG3, 3);
1786 __ mtctr(tmp1);
1787
1788 __ bind(l_4);
1789 // Use unrolled version for mass copying (copy 4 elements a time).
1790 // Load feeding store gets zero latency on Power6, however not on Power5.
1791 // Therefore, the following sequence is made for the good of both.
1792 __ addi(R3_ARG1, R3_ARG1, -32);
1793 __ addi(R4_ARG2, R4_ARG2, -32);
1794 __ ld(tmp4, 24, R3_ARG1);
1795 __ ld(tmp3, 16, R3_ARG1);
1796 __ ld(tmp2, 8, R3_ARG1);
1797 __ ld(tmp1, 0, R3_ARG1);
1798 __ std(tmp4, 24, R4_ARG2);
1799 __ std(tmp3, 16, R4_ARG2);
1800 __ std(tmp2, 8, R4_ARG2);
1801 __ std(tmp1, 0, R4_ARG2);
1802 __ bdnz(l_4);
1803
1804 __ cmpwi(CCR0, R5_ARG3, 0);
1805 __ beq(CCR0, l_1);
1806
1807 __ bind(l_5);
1808 __ mtctr(R5_ARG3);
1809 __ bind(l_3);
1810 __ ld(R0, -8, R3_ARG1);
1811 __ std(R0, -8, R4_ARG2);
1812 __ addi(R3_ARG1, R3_ARG1, -8);
1813 __ addi(R4_ARG2, R4_ARG2, -8);
1814 __ bdnz(l_3);
1815
1816 }
1817 __ bind(l_1);
1818 }
1819
1820 // Generate stub for conjoint long copy. If "aligned" is true, the
1821 // "from" and "to" addresses are assumed to be heapword aligned.
1822 //
1823 // Arguments for generated stub:
1824 // from: R3_ARG1
1825 // to: R4_ARG2
1826 // count: R5_ARG3 treated as signed
1827 //
1828 address generate_conjoint_long_copy(bool aligned, const char * name) {
1829 StubCodeMark mark(this, "StubRoutines", name);
1830 address start = __ emit_fd();
1831
1832 address nooverlap_target = aligned ?
1833 ((FunctionDescriptor*)StubRoutines::arrayof_jlong_disjoint_arraycopy())->entry() :
1834 ((FunctionDescriptor*)StubRoutines::jlong_disjoint_arraycopy())->entry();
1835
1836 array_overlap_test(nooverlap_target, 3);
1837 generate_conjoint_long_copy_core(aligned);
1838
1839 __ blr();
1840
1841 return start;
1842 }
1843
1844 // Generate stub for conjoint oop copy. If "aligned" is true, the
1845 // "from" and "to" addresses are assumed to be heapword aligned.
1846 //
1847 // Arguments for generated stub:
1848 // from: R3_ARG1
1849 // to: R4_ARG2
1850 // count: R5_ARG3 treated as signed
1851 // dest_uninitialized: G1 support
1852 //
1853 address generate_conjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) {
1854 StubCodeMark mark(this, "StubRoutines", name);
1855
1856 address start = __ emit_fd();
1857
1858 address nooverlap_target = aligned ?
1859 ((FunctionDescriptor*)StubRoutines::arrayof_oop_disjoint_arraycopy())->entry() :
1860 ((FunctionDescriptor*)StubRoutines::oop_disjoint_arraycopy())->entry();
1861
1862 gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7);
1863
1864 // Save arguments.
1865 __ mr(R9_ARG7, R4_ARG2);
1866 __ mr(R10_ARG8, R5_ARG3);
1867
1868 if (UseCompressedOops) {
1869 array_overlap_test(nooverlap_target, 2);
1870 generate_conjoint_int_copy_core(aligned);
1871 } else {
1872 array_overlap_test(nooverlap_target, 3);
1873 generate_conjoint_long_copy_core(aligned);
1874 }
1875
1876 gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1);
1877
1878 __ blr();
1879
1880 return start;
1881 }
1882
1883 // Generate stub for disjoint oop copy. If "aligned" is true, the
1884 // "from" and "to" addresses are assumed to be heapword aligned.
1885 //
1886 // Arguments for generated stub:
1887 // from: R3_ARG1
1888 // to: R4_ARG2
1889 // count: R5_ARG3 treated as signed
1890 // dest_uninitialized: G1 support
1891 //
1892 address generate_disjoint_oop_copy(bool aligned, const char * name, bool dest_uninitialized) {
1893 StubCodeMark mark(this, "StubRoutines", name);
1894 address start = __ emit_fd();
1895
1896 gen_write_ref_array_pre_barrier(R3_ARG1, R4_ARG2, R5_ARG3, dest_uninitialized, R9_ARG7);
1897
1898 // save some arguments, disjoint_long_copy_core destroys them.
1899 // needed for post barrier
1900 __ mr(R9_ARG7, R4_ARG2);
1901 __ mr(R10_ARG8, R5_ARG3);
1902
1903 if (UseCompressedOops) {
1904 generate_disjoint_int_copy_core(aligned);
1905 } else {
1906 generate_disjoint_long_copy_core(aligned);
1907 }
1908
1909 gen_write_ref_array_post_barrier(R9_ARG7, R10_ARG8, R11_scratch1);
1910
1911 __ blr();
1912
1913 return start;
1914 }
1915
1916 void generate_arraycopy_stubs() {
1917 // Note: the disjoint stubs must be generated first, some of
1918 // the conjoint stubs use them.
1919
1920 // non-aligned disjoint versions
1921 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, "jbyte_disjoint_arraycopy");
1922 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, "jshort_disjoint_arraycopy");
1923 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_copy(false, "jint_disjoint_arraycopy");
1924 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_copy(false, "jlong_disjoint_arraycopy");
1925 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy", false);
1926 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(false, "oop_disjoint_arraycopy_uninit", true);
1927
1928 // aligned disjoint versions
1929 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(true, "arrayof_jbyte_disjoint_arraycopy");
1930 StubRoutines::_arrayof_jshort_disjoint_arraycopy = generate_disjoint_short_copy(true, "arrayof_jshort_disjoint_arraycopy");
1931 StubRoutines::_arrayof_jint_disjoint_arraycopy = generate_disjoint_int_copy(true, "arrayof_jint_disjoint_arraycopy");
1932 StubRoutines::_arrayof_jlong_disjoint_arraycopy = generate_disjoint_long_copy(true, "arrayof_jlong_disjoint_arraycopy");
1933 StubRoutines::_arrayof_oop_disjoint_arraycopy = generate_disjoint_oop_copy(true, "arrayof_oop_disjoint_arraycopy", false);
1934 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(true, "oop_disjoint_arraycopy_uninit", true);
1935
1936 // non-aligned conjoint versions
1937 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, "jbyte_arraycopy");
1938 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, "jshort_arraycopy");
1939 StubRoutines::_jint_arraycopy = generate_conjoint_int_copy(false, "jint_arraycopy");
1940 StubRoutines::_jlong_arraycopy = generate_conjoint_long_copy(false, "jlong_arraycopy");
1941 StubRoutines::_oop_arraycopy = generate_conjoint_oop_copy(false, "oop_arraycopy", false);
1942 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_oop_copy(false, "oop_arraycopy_uninit", true);
1943
1944 // aligned conjoint versions
1945 StubRoutines::_arrayof_jbyte_arraycopy = generate_conjoint_byte_copy(true, "arrayof_jbyte_arraycopy");
1946 StubRoutines::_arrayof_jshort_arraycopy = generate_conjoint_short_copy(true, "arrayof_jshort_arraycopy");
1947 StubRoutines::_arrayof_jint_arraycopy = generate_conjoint_int_copy(true, "arrayof_jint_arraycopy");
1948 StubRoutines::_arrayof_jlong_arraycopy = generate_conjoint_long_copy(true, "arrayof_jlong_arraycopy");
1949 StubRoutines::_arrayof_oop_arraycopy = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", false);
1950 StubRoutines::_arrayof_oop_arraycopy_uninit = generate_conjoint_oop_copy(true, "arrayof_oop_arraycopy", true);
1951
1952 // fill routines
1953 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill");
1954 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill");
1955 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill");
1956 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill");
1957 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
1958 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill");
1959 }
1960
1961 // Safefetch stubs.
1962 void generate_safefetch(const char* name, int size, address* entry, address* fault_pc, address* continuation_pc) {
1963 // safefetch signatures:
1964 // int SafeFetch32(int* adr, int errValue);
1965 // intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
1966 //
1967 // arguments:
1968 // R3_ARG1 = adr
1969 // R4_ARG2 = errValue
1970 //
1971 // result:
1972 // R3_RET = *adr or errValue
1973
1974 StubCodeMark mark(this, "StubRoutines", name);
1975
1976 // Entry point, pc or function descriptor.
1977 *entry = __ emit_fd();
1978
1979 // Load *adr into R4_ARG2, may fault.
1980 *fault_pc = __ pc();
1981 switch (size) {
1982 case 4:
1983 // int32_t, signed extended
1984 __ lwa(R4_ARG2, 0, R3_ARG1);
1985 break;
1986 case 8:
1987 // int64_t
1988 __ ld(R4_ARG2, 0, R3_ARG1);
1989 break;
1990 default:
1991 ShouldNotReachHere();
1992 }
1993
1994 // return errValue or *adr
1995 *continuation_pc = __ pc();
1996 __ mr(R3_RET, R4_ARG2);
1997 __ blr();
1998 }
1999
2000 // Initialization
2001 void generate_initial() {
2002 // Generates all stubs and initializes the entry points
2003
2004 // Entry points that exist in all platforms.
2005 // Note: This is code that could be shared among different platforms - however the
2006 // benefit seems to be smaller than the disadvantage of having a
2007 // much more complicated generator structure. See also comment in
2008 // stubRoutines.hpp.
2009
2010 StubRoutines::_forward_exception_entry = generate_forward_exception();
2011 StubRoutines::_call_stub_entry = generate_call_stub(StubRoutines::_call_stub_return_address);
2012 StubRoutines::_catch_exception_entry = generate_catch_exception();
2013 }
2014
2015 void generate_all() {
2016 // Generates all stubs and initializes the entry points
2017
2018 // These entry points require SharedInfo::stack0 to be set up in
2019 // non-core builds
2020 StubRoutines::_throw_AbstractMethodError_entry = generate_throw_exception("AbstractMethodError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_AbstractMethodError), false);
2021 // Handle IncompatibleClassChangeError in itable stubs.
2022 StubRoutines::_throw_IncompatibleClassChangeError_entry= generate_throw_exception("IncompatibleClassChangeError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_IncompatibleClassChangeError), false);
2023 StubRoutines::_throw_NullPointerException_at_call_entry= generate_throw_exception("NullPointerException at call throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_NullPointerException_at_call), false);
2024 StubRoutines::_throw_StackOverflowError_entry = generate_throw_exception("StackOverflowError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_StackOverflowError), false);
2025
2026 StubRoutines::_handler_for_unsafe_access_entry = generate_handler_for_unsafe_access();
2027
2028 // support for verify_oop (must happen after universe_init)
2029 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
2030
2031 // arraycopy stubs used by compilers
2032 generate_arraycopy_stubs();
2033
2034 // PPC uses stubs for safefetch.
2035 generate_safefetch("SafeFetch32", sizeof(int), &StubRoutines::_safefetch32_entry,
2036 &StubRoutines::_safefetch32_fault_pc,
2037 &StubRoutines::_safefetch32_continuation_pc);
2038 generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
2039 &StubRoutines::_safefetchN_fault_pc,
2040 &StubRoutines::_safefetchN_continuation_pc);
2041 }
2042
2043 public:
2044 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
2045 // replace the standard masm with a special one:
2046 _masm = new MacroAssembler(code);
2047 if (all) {
2048 generate_all();
2049 } else {
2050 generate_initial();
2051 }
2052 }
2053 };
2054
2055 void StubGenerator_generate(CodeBuffer* code, bool all) {
2056 StubGenerator g(code, all);
2057 }