Tue, 07 Jan 2014 17:24:59 +0100
8031319: PPC64: Some fixes in ppc and aix coding.
Reviewed-by: kvn
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 */
26 #include "precompiled.hpp"
27 #include "asm/assembler.hpp"
28 #include "asm/assembler.inline.hpp"
29 #include "asm/macroAssembler.inline.hpp"
30 #include "compiler/disassembler.hpp"
31 #include "gc_interface/collectedHeap.inline.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/cardTableModRefBS.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "prims/methodHandles.hpp"
36 #include "runtime/biasedLocking.hpp"
37 #include "runtime/interfaceSupport.hpp"
38 #include "runtime/objectMonitor.hpp"
39 #include "runtime/os.hpp"
40 #include "runtime/sharedRuntime.hpp"
41 #include "runtime/stubRoutines.hpp"
42 #include "utilities/macros.hpp"
43 #if INCLUDE_ALL_GCS
44 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
45 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
46 #include "gc_implementation/g1/heapRegion.hpp"
47 #endif // INCLUDE_ALL_GCS
49 #ifdef PRODUCT
50 #define BLOCK_COMMENT(str) // nothing
51 #else
52 #define BLOCK_COMMENT(str) block_comment(str)
53 #endif
55 #ifdef ASSERT
56 // On RISC, there's no benefit to verifying instruction boundaries.
57 bool AbstractAssembler::pd_check_instruction_mark() { return false; }
58 #endif
60 void MacroAssembler::ld_largeoffset_unchecked(Register d, int si31, Register a, int emit_filler_nop) {
61 assert(Assembler::is_simm(si31, 31) && si31 >= 0, "si31 out of range");
62 if (Assembler::is_simm(si31, 16)) {
63 ld(d, si31, a);
64 if (emit_filler_nop) nop();
65 } else {
66 const int hi = MacroAssembler::largeoffset_si16_si16_hi(si31);
67 const int lo = MacroAssembler::largeoffset_si16_si16_lo(si31);
68 addis(d, a, hi);
69 ld(d, lo, d);
70 }
71 }
73 void MacroAssembler::ld_largeoffset(Register d, int si31, Register a, int emit_filler_nop) {
74 assert_different_registers(d, a);
75 ld_largeoffset_unchecked(d, si31, a, emit_filler_nop);
76 }
78 void MacroAssembler::load_sized_value(Register dst, RegisterOrConstant offs, Register base,
79 size_t size_in_bytes, bool is_signed) {
80 switch (size_in_bytes) {
81 case 8: ld(dst, offs, base); break;
82 case 4: is_signed ? lwa(dst, offs, base) : lwz(dst, offs, base); break;
83 case 2: is_signed ? lha(dst, offs, base) : lhz(dst, offs, base); break;
84 case 1: lbz(dst, offs, base); if (is_signed) extsb(dst, dst); break; // lba doesn't exist :(
85 default: ShouldNotReachHere();
86 }
87 }
89 void MacroAssembler::store_sized_value(Register dst, RegisterOrConstant offs, Register base,
90 size_t size_in_bytes) {
91 switch (size_in_bytes) {
92 case 8: std(dst, offs, base); break;
93 case 4: stw(dst, offs, base); break;
94 case 2: sth(dst, offs, base); break;
95 case 1: stb(dst, offs, base); break;
96 default: ShouldNotReachHere();
97 }
98 }
100 void MacroAssembler::align(int modulus, int max, int rem) {
101 int padding = (rem + modulus - (offset() % modulus)) % modulus;
102 if (padding > max) return;
103 for (int c = (padding >> 2); c > 0; --c) { nop(); }
104 }
106 // Issue instructions that calculate given TOC from global TOC.
107 void MacroAssembler::calculate_address_from_global_toc(Register dst, address addr, bool hi16, bool lo16,
108 bool add_relocation, bool emit_dummy_addr) {
109 int offset = -1;
110 if (emit_dummy_addr) {
111 offset = -128; // dummy address
112 } else if (addr != (address)(intptr_t)-1) {
113 offset = MacroAssembler::offset_to_global_toc(addr);
114 }
116 if (hi16) {
117 addis(dst, R29, MacroAssembler::largeoffset_si16_si16_hi(offset));
118 }
119 if (lo16) {
120 if (add_relocation) {
121 // Relocate at the addi to avoid confusion with a load from the method's TOC.
122 relocate(internal_word_Relocation::spec(addr));
123 }
124 addi(dst, dst, MacroAssembler::largeoffset_si16_si16_lo(offset));
125 }
126 }
128 int MacroAssembler::patch_calculate_address_from_global_toc_at(address a, address bound, address addr) {
129 const int offset = MacroAssembler::offset_to_global_toc(addr);
131 const address inst2_addr = a;
132 const int inst2 = *(int *)inst2_addr;
134 // The relocation points to the second instruction, the addi,
135 // and the addi reads and writes the same register dst.
136 const int dst = inv_rt_field(inst2);
137 assert(is_addi(inst2) && inv_ra_field(inst2) == dst, "must be addi reading and writing dst");
139 // Now, find the preceding addis which writes to dst.
140 int inst1 = 0;
141 address inst1_addr = inst2_addr - BytesPerInstWord;
142 while (inst1_addr >= bound) {
143 inst1 = *(int *) inst1_addr;
144 if (is_addis(inst1) && inv_rt_field(inst1) == dst) {
145 // Stop, found the addis which writes dst.
146 break;
147 }
148 inst1_addr -= BytesPerInstWord;
149 }
151 assert(is_addis(inst1) && inv_ra_field(inst1) == 29 /* R29 */, "source must be global TOC");
152 set_imm((int *)inst1_addr, MacroAssembler::largeoffset_si16_si16_hi(offset));
153 set_imm((int *)inst2_addr, MacroAssembler::largeoffset_si16_si16_lo(offset));
154 return (int)((intptr_t)addr - (intptr_t)inst1_addr);
155 }
157 address MacroAssembler::get_address_of_calculate_address_from_global_toc_at(address a, address bound) {
158 const address inst2_addr = a;
159 const int inst2 = *(int *)inst2_addr;
161 // The relocation points to the second instruction, the addi,
162 // and the addi reads and writes the same register dst.
163 const int dst = inv_rt_field(inst2);
164 assert(is_addi(inst2) && inv_ra_field(inst2) == dst, "must be addi reading and writing dst");
166 // Now, find the preceding addis which writes to dst.
167 int inst1 = 0;
168 address inst1_addr = inst2_addr - BytesPerInstWord;
169 while (inst1_addr >= bound) {
170 inst1 = *(int *) inst1_addr;
171 if (is_addis(inst1) && inv_rt_field(inst1) == dst) {
172 // stop, found the addis which writes dst
173 break;
174 }
175 inst1_addr -= BytesPerInstWord;
176 }
178 assert(is_addis(inst1) && inv_ra_field(inst1) == 29 /* R29 */, "source must be global TOC");
180 int offset = (get_imm(inst1_addr, 0) << 16) + get_imm(inst2_addr, 0);
181 // -1 is a special case
182 if (offset == -1) {
183 return (address)(intptr_t)-1;
184 } else {
185 return global_toc() + offset;
186 }
187 }
189 #ifdef _LP64
190 // Patch compressed oops or klass constants.
191 // Assembler sequence is
192 // 1) compressed oops:
193 // lis rx = const.hi
194 // ori rx = rx | const.lo
195 // 2) compressed klass:
196 // lis rx = const.hi
197 // clrldi rx = rx & 0xFFFFffff // clearMS32b, optional
198 // ori rx = rx | const.lo
199 // Clrldi will be passed by.
200 int MacroAssembler::patch_set_narrow_oop(address a, address bound, narrowOop data) {
201 assert(UseCompressedOops, "Should only patch compressed oops");
203 const address inst2_addr = a;
204 const int inst2 = *(int *)inst2_addr;
206 // The relocation points to the second instruction, the ori,
207 // and the ori reads and writes the same register dst.
208 const int dst = inv_rta_field(inst2);
209 assert(is_ori(inst2) && inv_rs_field(inst2) == dst, "must be ori reading and writing dst");
210 // Now, find the preceding addis which writes to dst.
211 int inst1 = 0;
212 address inst1_addr = inst2_addr - BytesPerInstWord;
213 bool inst1_found = false;
214 while (inst1_addr >= bound) {
215 inst1 = *(int *)inst1_addr;
216 if (is_lis(inst1) && inv_rs_field(inst1) == dst) { inst1_found = true; break; }
217 inst1_addr -= BytesPerInstWord;
218 }
219 assert(inst1_found, "inst is not lis");
221 int xc = (data >> 16) & 0xffff;
222 int xd = (data >> 0) & 0xffff;
224 set_imm((int *)inst1_addr, (short)(xc)); // see enc_load_con_narrow_hi/_lo
225 set_imm((int *)inst2_addr, (xd)); // unsigned int
226 return (int)((intptr_t)inst2_addr - (intptr_t)inst1_addr);
227 }
229 // Get compressed oop or klass constant.
230 narrowOop MacroAssembler::get_narrow_oop(address a, address bound) {
231 assert(UseCompressedOops, "Should only patch compressed oops");
233 const address inst2_addr = a;
234 const int inst2 = *(int *)inst2_addr;
236 // The relocation points to the second instruction, the ori,
237 // and the ori reads and writes the same register dst.
238 const int dst = inv_rta_field(inst2);
239 assert(is_ori(inst2) && inv_rs_field(inst2) == dst, "must be ori reading and writing dst");
240 // Now, find the preceding lis which writes to dst.
241 int inst1 = 0;
242 address inst1_addr = inst2_addr - BytesPerInstWord;
243 bool inst1_found = false;
245 while (inst1_addr >= bound) {
246 inst1 = *(int *) inst1_addr;
247 if (is_lis(inst1) && inv_rs_field(inst1) == dst) { inst1_found = true; break;}
248 inst1_addr -= BytesPerInstWord;
249 }
250 assert(inst1_found, "inst is not lis");
252 uint xl = ((unsigned int) (get_imm(inst2_addr, 0) & 0xffff));
253 uint xh = (((get_imm(inst1_addr, 0)) & 0xffff) << 16);
255 return (int) (xl | xh);
256 }
257 #endif // _LP64
259 void MacroAssembler::load_const_from_method_toc(Register dst, AddressLiteral& a, Register toc) {
260 int toc_offset = 0;
261 // Use RelocationHolder::none for the constant pool entry, otherwise
262 // we will end up with a failing NativeCall::verify(x) where x is
263 // the address of the constant pool entry.
264 // FIXME: We should insert relocation information for oops at the constant
265 // pool entries instead of inserting it at the loads; patching of a constant
266 // pool entry should be less expensive.
267 address oop_address = address_constant((address)a.value(), RelocationHolder::none);
268 // Relocate at the pc of the load.
269 relocate(a.rspec());
270 toc_offset = (int)(oop_address - code()->consts()->start());
271 ld_largeoffset_unchecked(dst, toc_offset, toc, true);
272 }
274 bool MacroAssembler::is_load_const_from_method_toc_at(address a) {
275 const address inst1_addr = a;
276 const int inst1 = *(int *)inst1_addr;
278 // The relocation points to the ld or the addis.
279 return (is_ld(inst1)) ||
280 (is_addis(inst1) && inv_ra_field(inst1) != 0);
281 }
283 int MacroAssembler::get_offset_of_load_const_from_method_toc_at(address a) {
284 assert(is_load_const_from_method_toc_at(a), "must be load_const_from_method_toc");
286 const address inst1_addr = a;
287 const int inst1 = *(int *)inst1_addr;
289 if (is_ld(inst1)) {
290 return inv_d1_field(inst1);
291 } else if (is_addis(inst1)) {
292 const int dst = inv_rt_field(inst1);
294 // Now, find the succeeding ld which reads and writes to dst.
295 address inst2_addr = inst1_addr + BytesPerInstWord;
296 int inst2 = 0;
297 while (true) {
298 inst2 = *(int *) inst2_addr;
299 if (is_ld(inst2) && inv_ra_field(inst2) == dst && inv_rt_field(inst2) == dst) {
300 // Stop, found the ld which reads and writes dst.
301 break;
302 }
303 inst2_addr += BytesPerInstWord;
304 }
305 return (inv_d1_field(inst1) << 16) + inv_d1_field(inst2);
306 }
307 ShouldNotReachHere();
308 return 0;
309 }
311 // Get the constant from a `load_const' sequence.
312 long MacroAssembler::get_const(address a) {
313 assert(is_load_const_at(a), "not a load of a constant");
314 const int *p = (const int*) a;
315 unsigned long x = (((unsigned long) (get_imm(a,0) & 0xffff)) << 48);
316 if (is_ori(*(p+1))) {
317 x |= (((unsigned long) (get_imm(a,1) & 0xffff)) << 32);
318 x |= (((unsigned long) (get_imm(a,3) & 0xffff)) << 16);
319 x |= (((unsigned long) (get_imm(a,4) & 0xffff)));
320 } else if (is_lis(*(p+1))) {
321 x |= (((unsigned long) (get_imm(a,2) & 0xffff)) << 32);
322 x |= (((unsigned long) (get_imm(a,1) & 0xffff)) << 16);
323 x |= (((unsigned long) (get_imm(a,3) & 0xffff)));
324 } else {
325 ShouldNotReachHere();
326 return (long) 0;
327 }
328 return (long) x;
329 }
331 // Patch the 64 bit constant of a `load_const' sequence. This is a low
332 // level procedure. It neither flushes the instruction cache nor is it
333 // mt safe.
334 void MacroAssembler::patch_const(address a, long x) {
335 assert(is_load_const_at(a), "not a load of a constant");
336 int *p = (int*) a;
337 if (is_ori(*(p+1))) {
338 set_imm(0 + p, (x >> 48) & 0xffff);
339 set_imm(1 + p, (x >> 32) & 0xffff);
340 set_imm(3 + p, (x >> 16) & 0xffff);
341 set_imm(4 + p, x & 0xffff);
342 } else if (is_lis(*(p+1))) {
343 set_imm(0 + p, (x >> 48) & 0xffff);
344 set_imm(2 + p, (x >> 32) & 0xffff);
345 set_imm(1 + p, (x >> 16) & 0xffff);
346 set_imm(3 + p, x & 0xffff);
347 } else {
348 ShouldNotReachHere();
349 }
350 }
352 AddressLiteral MacroAssembler::allocate_metadata_address(Metadata* obj) {
353 assert(oop_recorder() != NULL, "this assembler needs a Recorder");
354 int index = oop_recorder()->allocate_metadata_index(obj);
355 RelocationHolder rspec = metadata_Relocation::spec(index);
356 return AddressLiteral((address)obj, rspec);
357 }
359 AddressLiteral MacroAssembler::constant_metadata_address(Metadata* obj) {
360 assert(oop_recorder() != NULL, "this assembler needs a Recorder");
361 int index = oop_recorder()->find_index(obj);
362 RelocationHolder rspec = metadata_Relocation::spec(index);
363 return AddressLiteral((address)obj, rspec);
364 }
366 AddressLiteral MacroAssembler::allocate_oop_address(jobject obj) {
367 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
368 int oop_index = oop_recorder()->allocate_oop_index(obj);
369 return AddressLiteral(address(obj), oop_Relocation::spec(oop_index));
370 }
372 AddressLiteral MacroAssembler::constant_oop_address(jobject obj) {
373 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
374 int oop_index = oop_recorder()->find_index(obj);
375 return AddressLiteral(address(obj), oop_Relocation::spec(oop_index));
376 }
378 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
379 Register tmp, int offset) {
380 intptr_t value = *delayed_value_addr;
381 if (value != 0) {
382 return RegisterOrConstant(value + offset);
383 }
385 // Load indirectly to solve generation ordering problem.
386 // static address, no relocation
387 int simm16_offset = load_const_optimized(tmp, delayed_value_addr, noreg, true);
388 ld(tmp, simm16_offset, tmp); // must be aligned ((xa & 3) == 0)
390 if (offset != 0) {
391 addi(tmp, tmp, offset);
392 }
394 return RegisterOrConstant(tmp);
395 }
397 #ifndef PRODUCT
398 void MacroAssembler::pd_print_patched_instruction(address branch) {
399 Unimplemented(); // TODO: PPC port
400 }
401 #endif // ndef PRODUCT
403 // Conditional far branch for destinations encodable in 24+2 bits.
404 void MacroAssembler::bc_far(int boint, int biint, Label& dest, int optimize) {
406 // If requested by flag optimize, relocate the bc_far as a
407 // runtime_call and prepare for optimizing it when the code gets
408 // relocated.
409 if (optimize == bc_far_optimize_on_relocate) {
410 relocate(relocInfo::runtime_call_type);
411 }
413 // variant 2:
414 //
415 // b!cxx SKIP
416 // bxx DEST
417 // SKIP:
418 //
420 const int opposite_boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(boint)),
421 opposite_bcond(inv_boint_bcond(boint)));
423 // We emit two branches.
424 // First, a conditional branch which jumps around the far branch.
425 const address not_taken_pc = pc() + 2 * BytesPerInstWord;
426 const address bc_pc = pc();
427 bc(opposite_boint, biint, not_taken_pc);
429 const int bc_instr = *(int*)bc_pc;
430 assert(not_taken_pc == (address)inv_bd_field(bc_instr, (intptr_t)bc_pc), "postcondition");
431 assert(opposite_boint == inv_bo_field(bc_instr), "postcondition");
432 assert(boint == add_bhint_to_boint(opposite_bhint(inv_boint_bhint(inv_bo_field(bc_instr))),
433 opposite_bcond(inv_boint_bcond(inv_bo_field(bc_instr)))),
434 "postcondition");
435 assert(biint == inv_bi_field(bc_instr), "postcondition");
437 // Second, an unconditional far branch which jumps to dest.
438 // Note: target(dest) remembers the current pc (see CodeSection::target)
439 // and returns the current pc if the label is not bound yet; when
440 // the label gets bound, the unconditional far branch will be patched.
441 const address target_pc = target(dest);
442 const address b_pc = pc();
443 b(target_pc);
445 assert(not_taken_pc == pc(), "postcondition");
446 assert(dest.is_bound() || target_pc == b_pc, "postcondition");
447 }
449 bool MacroAssembler::is_bc_far_at(address instruction_addr) {
450 return is_bc_far_variant1_at(instruction_addr) ||
451 is_bc_far_variant2_at(instruction_addr) ||
452 is_bc_far_variant3_at(instruction_addr);
453 }
455 address MacroAssembler::get_dest_of_bc_far_at(address instruction_addr) {
456 if (is_bc_far_variant1_at(instruction_addr)) {
457 const address instruction_1_addr = instruction_addr;
458 const int instruction_1 = *(int*)instruction_1_addr;
459 return (address)inv_bd_field(instruction_1, (intptr_t)instruction_1_addr);
460 } else if (is_bc_far_variant2_at(instruction_addr)) {
461 const address instruction_2_addr = instruction_addr + 4;
462 return bxx_destination(instruction_2_addr);
463 } else if (is_bc_far_variant3_at(instruction_addr)) {
464 return instruction_addr + 8;
465 }
466 // variant 4 ???
467 ShouldNotReachHere();
468 return NULL;
469 }
470 void MacroAssembler::set_dest_of_bc_far_at(address instruction_addr, address dest) {
472 if (is_bc_far_variant3_at(instruction_addr)) {
473 // variant 3, far cond branch to the next instruction, already patched to nops:
474 //
475 // nop
476 // endgroup
477 // SKIP/DEST:
478 //
479 return;
480 }
482 // first, extract boint and biint from the current branch
483 int boint = 0;
484 int biint = 0;
486 ResourceMark rm;
487 const int code_size = 2 * BytesPerInstWord;
488 CodeBuffer buf(instruction_addr, code_size);
489 MacroAssembler masm(&buf);
490 if (is_bc_far_variant2_at(instruction_addr) && dest == instruction_addr + 8) {
491 // Far branch to next instruction: Optimize it by patching nops (produce variant 3).
492 masm.nop();
493 masm.endgroup();
494 } else {
495 if (is_bc_far_variant1_at(instruction_addr)) {
496 // variant 1, the 1st instruction contains the destination address:
497 //
498 // bcxx DEST
499 // endgroup
500 //
501 const int instruction_1 = *(int*)(instruction_addr);
502 boint = inv_bo_field(instruction_1);
503 biint = inv_bi_field(instruction_1);
504 } else if (is_bc_far_variant2_at(instruction_addr)) {
505 // variant 2, the 2nd instruction contains the destination address:
506 //
507 // b!cxx SKIP
508 // bxx DEST
509 // SKIP:
510 //
511 const int instruction_1 = *(int*)(instruction_addr);
512 boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(inv_bo_field(instruction_1))),
513 opposite_bcond(inv_boint_bcond(inv_bo_field(instruction_1))));
514 biint = inv_bi_field(instruction_1);
515 } else {
516 // variant 4???
517 ShouldNotReachHere();
518 }
520 // second, set the new branch destination and optimize the code
521 if (dest != instruction_addr + 4 && // the bc_far is still unbound!
522 masm.is_within_range_of_bcxx(dest, instruction_addr)) {
523 // variant 1:
524 //
525 // bcxx DEST
526 // endgroup
527 //
528 masm.bc(boint, biint, dest);
529 masm.endgroup();
530 } else {
531 // variant 2:
532 //
533 // b!cxx SKIP
534 // bxx DEST
535 // SKIP:
536 //
537 const int opposite_boint = add_bhint_to_boint(opposite_bhint(inv_boint_bhint(boint)),
538 opposite_bcond(inv_boint_bcond(boint)));
539 const address not_taken_pc = masm.pc() + 2 * BytesPerInstWord;
540 masm.bc(opposite_boint, biint, not_taken_pc);
541 masm.b(dest);
542 }
543 }
544 ICache::ppc64_flush_icache_bytes(instruction_addr, code_size);
545 }
547 // Emit a NOT mt-safe patchable 64 bit absolute call/jump.
548 void MacroAssembler::bxx64_patchable(address dest, relocInfo::relocType rt, bool link) {
549 // get current pc
550 uint64_t start_pc = (uint64_t) pc();
552 const address pc_of_bl = (address) (start_pc + (6*BytesPerInstWord)); // bl is last
553 const address pc_of_b = (address) (start_pc + (0*BytesPerInstWord)); // b is first
555 // relocate here
556 if (rt != relocInfo::none) {
557 relocate(rt);
558 }
560 if ( ReoptimizeCallSequences &&
561 (( link && is_within_range_of_b(dest, pc_of_bl)) ||
562 (!link && is_within_range_of_b(dest, pc_of_b)))) {
563 // variant 2:
564 // Emit an optimized, pc-relative call/jump.
566 if (link) {
567 // some padding
568 nop();
569 nop();
570 nop();
571 nop();
572 nop();
573 nop();
575 // do the call
576 assert(pc() == pc_of_bl, "just checking");
577 bl(dest, relocInfo::none);
578 } else {
579 // do the jump
580 assert(pc() == pc_of_b, "just checking");
581 b(dest, relocInfo::none);
583 // some padding
584 nop();
585 nop();
586 nop();
587 nop();
588 nop();
589 nop();
590 }
592 // Assert that we can identify the emitted call/jump.
593 assert(is_bxx64_patchable_variant2_at((address)start_pc, link),
594 "can't identify emitted call");
595 } else {
596 // variant 1:
598 mr(R0, R11); // spill R11 -> R0.
600 // Load the destination address into CTR,
601 // calculate destination relative to global toc.
602 calculate_address_from_global_toc(R11, dest, true, true, false);
604 mtctr(R11);
605 mr(R11, R0); // spill R11 <- R0.
606 nop();
608 // do the call/jump
609 if (link) {
610 bctrl();
611 } else{
612 bctr();
613 }
614 // Assert that we can identify the emitted call/jump.
615 assert(is_bxx64_patchable_variant1b_at((address)start_pc, link),
616 "can't identify emitted call");
617 }
619 // Assert that we can identify the emitted call/jump.
620 assert(is_bxx64_patchable_at((address)start_pc, link),
621 "can't identify emitted call");
622 assert(get_dest_of_bxx64_patchable_at((address)start_pc, link) == dest,
623 "wrong encoding of dest address");
624 }
626 // Identify a bxx64_patchable instruction.
627 bool MacroAssembler::is_bxx64_patchable_at(address instruction_addr, bool link) {
628 return is_bxx64_patchable_variant1b_at(instruction_addr, link)
629 //|| is_bxx64_patchable_variant1_at(instruction_addr, link)
630 || is_bxx64_patchable_variant2_at(instruction_addr, link);
631 }
633 // Does the call64_patchable instruction use a pc-relative encoding of
634 // the call destination?
635 bool MacroAssembler::is_bxx64_patchable_pcrelative_at(address instruction_addr, bool link) {
636 // variant 2 is pc-relative
637 return is_bxx64_patchable_variant2_at(instruction_addr, link);
638 }
640 // Identify variant 1.
641 bool MacroAssembler::is_bxx64_patchable_variant1_at(address instruction_addr, bool link) {
642 unsigned int* instr = (unsigned int*) instruction_addr;
643 return (link ? is_bctrl(instr[6]) : is_bctr(instr[6])) // bctr[l]
644 && is_mtctr(instr[5]) // mtctr
645 && is_load_const_at(instruction_addr);
646 }
648 // Identify variant 1b: load destination relative to global toc.
649 bool MacroAssembler::is_bxx64_patchable_variant1b_at(address instruction_addr, bool link) {
650 unsigned int* instr = (unsigned int*) instruction_addr;
651 return (link ? is_bctrl(instr[6]) : is_bctr(instr[6])) // bctr[l]
652 && is_mtctr(instr[3]) // mtctr
653 && is_calculate_address_from_global_toc_at(instruction_addr + 2*BytesPerInstWord, instruction_addr);
654 }
656 // Identify variant 2.
657 bool MacroAssembler::is_bxx64_patchable_variant2_at(address instruction_addr, bool link) {
658 unsigned int* instr = (unsigned int*) instruction_addr;
659 if (link) {
660 return is_bl (instr[6]) // bl dest is last
661 && is_nop(instr[0]) // nop
662 && is_nop(instr[1]) // nop
663 && is_nop(instr[2]) // nop
664 && is_nop(instr[3]) // nop
665 && is_nop(instr[4]) // nop
666 && is_nop(instr[5]); // nop
667 } else {
668 return is_b (instr[0]) // b dest is first
669 && is_nop(instr[1]) // nop
670 && is_nop(instr[2]) // nop
671 && is_nop(instr[3]) // nop
672 && is_nop(instr[4]) // nop
673 && is_nop(instr[5]) // nop
674 && is_nop(instr[6]); // nop
675 }
676 }
678 // Set dest address of a bxx64_patchable instruction.
679 void MacroAssembler::set_dest_of_bxx64_patchable_at(address instruction_addr, address dest, bool link) {
680 ResourceMark rm;
681 int code_size = MacroAssembler::bxx64_patchable_size;
682 CodeBuffer buf(instruction_addr, code_size);
683 MacroAssembler masm(&buf);
684 masm.bxx64_patchable(dest, relocInfo::none, link);
685 ICache::ppc64_flush_icache_bytes(instruction_addr, code_size);
686 }
688 // Get dest address of a bxx64_patchable instruction.
689 address MacroAssembler::get_dest_of_bxx64_patchable_at(address instruction_addr, bool link) {
690 if (is_bxx64_patchable_variant1_at(instruction_addr, link)) {
691 return (address) (unsigned long) get_const(instruction_addr);
692 } else if (is_bxx64_patchable_variant2_at(instruction_addr, link)) {
693 unsigned int* instr = (unsigned int*) instruction_addr;
694 if (link) {
695 const int instr_idx = 6; // bl is last
696 int branchoffset = branch_destination(instr[instr_idx], 0);
697 return instruction_addr + branchoffset + instr_idx*BytesPerInstWord;
698 } else {
699 const int instr_idx = 0; // b is first
700 int branchoffset = branch_destination(instr[instr_idx], 0);
701 return instruction_addr + branchoffset + instr_idx*BytesPerInstWord;
702 }
703 // Load dest relative to global toc.
704 } else if (is_bxx64_patchable_variant1b_at(instruction_addr, link)) {
705 return get_address_of_calculate_address_from_global_toc_at(instruction_addr + 2*BytesPerInstWord,
706 instruction_addr);
707 } else {
708 ShouldNotReachHere();
709 return NULL;
710 }
711 }
713 // Uses ordering which corresponds to ABI:
714 // _savegpr0_14: std r14,-144(r1)
715 // _savegpr0_15: std r15,-136(r1)
716 // _savegpr0_16: std r16,-128(r1)
717 void MacroAssembler::save_nonvolatile_gprs(Register dst, int offset) {
718 std(R14, offset, dst); offset += 8;
719 std(R15, offset, dst); offset += 8;
720 std(R16, offset, dst); offset += 8;
721 std(R17, offset, dst); offset += 8;
722 std(R18, offset, dst); offset += 8;
723 std(R19, offset, dst); offset += 8;
724 std(R20, offset, dst); offset += 8;
725 std(R21, offset, dst); offset += 8;
726 std(R22, offset, dst); offset += 8;
727 std(R23, offset, dst); offset += 8;
728 std(R24, offset, dst); offset += 8;
729 std(R25, offset, dst); offset += 8;
730 std(R26, offset, dst); offset += 8;
731 std(R27, offset, dst); offset += 8;
732 std(R28, offset, dst); offset += 8;
733 std(R29, offset, dst); offset += 8;
734 std(R30, offset, dst); offset += 8;
735 std(R31, offset, dst); offset += 8;
737 stfd(F14, offset, dst); offset += 8;
738 stfd(F15, offset, dst); offset += 8;
739 stfd(F16, offset, dst); offset += 8;
740 stfd(F17, offset, dst); offset += 8;
741 stfd(F18, offset, dst); offset += 8;
742 stfd(F19, offset, dst); offset += 8;
743 stfd(F20, offset, dst); offset += 8;
744 stfd(F21, offset, dst); offset += 8;
745 stfd(F22, offset, dst); offset += 8;
746 stfd(F23, offset, dst); offset += 8;
747 stfd(F24, offset, dst); offset += 8;
748 stfd(F25, offset, dst); offset += 8;
749 stfd(F26, offset, dst); offset += 8;
750 stfd(F27, offset, dst); offset += 8;
751 stfd(F28, offset, dst); offset += 8;
752 stfd(F29, offset, dst); offset += 8;
753 stfd(F30, offset, dst); offset += 8;
754 stfd(F31, offset, dst);
755 }
757 // Uses ordering which corresponds to ABI:
758 // _restgpr0_14: ld r14,-144(r1)
759 // _restgpr0_15: ld r15,-136(r1)
760 // _restgpr0_16: ld r16,-128(r1)
761 void MacroAssembler::restore_nonvolatile_gprs(Register src, int offset) {
762 ld(R14, offset, src); offset += 8;
763 ld(R15, offset, src); offset += 8;
764 ld(R16, offset, src); offset += 8;
765 ld(R17, offset, src); offset += 8;
766 ld(R18, offset, src); offset += 8;
767 ld(R19, offset, src); offset += 8;
768 ld(R20, offset, src); offset += 8;
769 ld(R21, offset, src); offset += 8;
770 ld(R22, offset, src); offset += 8;
771 ld(R23, offset, src); offset += 8;
772 ld(R24, offset, src); offset += 8;
773 ld(R25, offset, src); offset += 8;
774 ld(R26, offset, src); offset += 8;
775 ld(R27, offset, src); offset += 8;
776 ld(R28, offset, src); offset += 8;
777 ld(R29, offset, src); offset += 8;
778 ld(R30, offset, src); offset += 8;
779 ld(R31, offset, src); offset += 8;
781 // FP registers
782 lfd(F14, offset, src); offset += 8;
783 lfd(F15, offset, src); offset += 8;
784 lfd(F16, offset, src); offset += 8;
785 lfd(F17, offset, src); offset += 8;
786 lfd(F18, offset, src); offset += 8;
787 lfd(F19, offset, src); offset += 8;
788 lfd(F20, offset, src); offset += 8;
789 lfd(F21, offset, src); offset += 8;
790 lfd(F22, offset, src); offset += 8;
791 lfd(F23, offset, src); offset += 8;
792 lfd(F24, offset, src); offset += 8;
793 lfd(F25, offset, src); offset += 8;
794 lfd(F26, offset, src); offset += 8;
795 lfd(F27, offset, src); offset += 8;
796 lfd(F28, offset, src); offset += 8;
797 lfd(F29, offset, src); offset += 8;
798 lfd(F30, offset, src); offset += 8;
799 lfd(F31, offset, src);
800 }
802 // For verify_oops.
803 void MacroAssembler::save_volatile_gprs(Register dst, int offset) {
804 std(R3, offset, dst); offset += 8;
805 std(R4, offset, dst); offset += 8;
806 std(R5, offset, dst); offset += 8;
807 std(R6, offset, dst); offset += 8;
808 std(R7, offset, dst); offset += 8;
809 std(R8, offset, dst); offset += 8;
810 std(R9, offset, dst); offset += 8;
811 std(R10, offset, dst); offset += 8;
812 std(R11, offset, dst); offset += 8;
813 std(R12, offset, dst);
814 }
816 // For verify_oops.
817 void MacroAssembler::restore_volatile_gprs(Register src, int offset) {
818 ld(R3, offset, src); offset += 8;
819 ld(R4, offset, src); offset += 8;
820 ld(R5, offset, src); offset += 8;
821 ld(R6, offset, src); offset += 8;
822 ld(R7, offset, src); offset += 8;
823 ld(R8, offset, src); offset += 8;
824 ld(R9, offset, src); offset += 8;
825 ld(R10, offset, src); offset += 8;
826 ld(R11, offset, src); offset += 8;
827 ld(R12, offset, src);
828 }
830 void MacroAssembler::save_LR_CR(Register tmp) {
831 mfcr(tmp);
832 std(tmp, _abi(cr), R1_SP);
833 mflr(tmp);
834 std(tmp, _abi(lr), R1_SP);
835 // Tmp must contain lr on exit! (see return_addr and prolog in ppc64.ad)
836 }
838 void MacroAssembler::restore_LR_CR(Register tmp) {
839 assert(tmp != R1_SP, "must be distinct");
840 ld(tmp, _abi(lr), R1_SP);
841 mtlr(tmp);
842 ld(tmp, _abi(cr), R1_SP);
843 mtcr(tmp);
844 }
846 address MacroAssembler::get_PC_trash_LR(Register result) {
847 Label L;
848 bl(L);
849 bind(L);
850 address lr_pc = pc();
851 mflr(result);
852 return lr_pc;
853 }
855 void MacroAssembler::resize_frame(Register offset, Register tmp) {
856 #ifdef ASSERT
857 assert_different_registers(offset, tmp, R1_SP);
858 andi_(tmp, offset, frame::alignment_in_bytes-1);
859 asm_assert_eq("resize_frame: unaligned", 0x204);
860 #endif
862 // tmp <- *(SP)
863 ld(tmp, _abi(callers_sp), R1_SP);
864 // addr <- SP + offset;
865 // *(addr) <- tmp;
866 // SP <- addr
867 stdux(tmp, R1_SP, offset);
868 }
870 void MacroAssembler::resize_frame(int offset, Register tmp) {
871 assert(is_simm(offset, 16), "too big an offset");
872 assert_different_registers(tmp, R1_SP);
873 assert((offset & (frame::alignment_in_bytes-1))==0, "resize_frame: unaligned");
874 // tmp <- *(SP)
875 ld(tmp, _abi(callers_sp), R1_SP);
876 // addr <- SP + offset;
877 // *(addr) <- tmp;
878 // SP <- addr
879 stdu(tmp, offset, R1_SP);
880 }
882 void MacroAssembler::resize_frame_absolute(Register addr, Register tmp1, Register tmp2) {
883 // (addr == tmp1) || (addr == tmp2) is allowed here!
884 assert(tmp1 != tmp2, "must be distinct");
886 // compute offset w.r.t. current stack pointer
887 // tmp_1 <- addr - SP (!)
888 subf(tmp1, R1_SP, addr);
890 // atomically update SP keeping back link.
891 resize_frame(tmp1/* offset */, tmp2/* tmp */);
892 }
894 void MacroAssembler::push_frame(Register bytes, Register tmp) {
895 #ifdef ASSERT
896 assert(bytes != R0, "r0 not allowed here");
897 andi_(R0, bytes, frame::alignment_in_bytes-1);
898 asm_assert_eq("push_frame(Reg, Reg): unaligned", 0x203);
899 #endif
900 neg(tmp, bytes);
901 stdux(R1_SP, R1_SP, tmp);
902 }
904 // Push a frame of size `bytes'.
905 void MacroAssembler::push_frame(unsigned int bytes, Register tmp) {
906 long offset = align_addr(bytes, frame::alignment_in_bytes);
907 if (is_simm(-offset, 16)) {
908 stdu(R1_SP, -offset, R1_SP);
909 } else {
910 load_const(tmp, -offset);
911 stdux(R1_SP, R1_SP, tmp);
912 }
913 }
915 // Push a frame of size `bytes' plus abi112 on top.
916 void MacroAssembler::push_frame_abi112(unsigned int bytes, Register tmp) {
917 push_frame(bytes + frame::abi_112_size, tmp);
918 }
920 // Setup up a new C frame with a spill area for non-volatile GPRs and
921 // additional space for local variables.
922 void MacroAssembler::push_frame_abi112_nonvolatiles(unsigned int bytes,
923 Register tmp) {
924 push_frame(bytes + frame::abi_112_size + frame::spill_nonvolatiles_size, tmp);
925 }
927 // Pop current C frame.
928 void MacroAssembler::pop_frame() {
929 ld(R1_SP, _abi(callers_sp), R1_SP);
930 }
932 // Generic version of a call to C function via a function descriptor
933 // with variable support for C calling conventions (TOC, ENV, etc.).
934 // Updates and returns _last_calls_return_pc.
935 address MacroAssembler::branch_to(Register function_descriptor, bool and_link, bool save_toc_before_call,
936 bool restore_toc_after_call, bool load_toc_of_callee, bool load_env_of_callee) {
937 // we emit standard ptrgl glue code here
938 assert((function_descriptor != R0), "function_descriptor cannot be R0");
940 // retrieve necessary entries from the function descriptor
941 ld(R0, in_bytes(FunctionDescriptor::entry_offset()), function_descriptor);
942 mtctr(R0);
944 if (load_toc_of_callee) {
945 ld(R2_TOC, in_bytes(FunctionDescriptor::toc_offset()), function_descriptor);
946 }
947 if (load_env_of_callee) {
948 ld(R11, in_bytes(FunctionDescriptor::env_offset()), function_descriptor);
949 } else if (load_toc_of_callee) {
950 li(R11, 0);
951 }
953 // do a call or a branch
954 if (and_link) {
955 bctrl();
956 } else {
957 bctr();
958 }
959 _last_calls_return_pc = pc();
961 return _last_calls_return_pc;
962 }
964 // Call a C function via a function descriptor and use full C calling
965 // conventions.
966 // We don't use the TOC in generated code, so there is no need to save
967 // and restore its value.
968 address MacroAssembler::call_c(Register fd) {
969 return branch_to(fd, /*and_link=*/true,
970 /*save toc=*/false,
971 /*restore toc=*/false,
972 /*load toc=*/true,
973 /*load env=*/true);
974 }
976 address MacroAssembler::call_c_and_return_to_caller(Register fd) {
977 return branch_to(fd, /*and_link=*/false,
978 /*save toc=*/false,
979 /*restore toc=*/false,
980 /*load toc=*/true,
981 /*load env=*/true);
982 }
984 address MacroAssembler::call_c(const FunctionDescriptor* fd, relocInfo::relocType rt) {
985 if (rt != relocInfo::none) {
986 // this call needs to be relocatable
987 if (!ReoptimizeCallSequences
988 || (rt != relocInfo::runtime_call_type && rt != relocInfo::none)
989 || fd == NULL // support code-size estimation
990 || !fd->is_friend_function()
991 || fd->entry() == NULL) {
992 // it's not a friend function as defined by class FunctionDescriptor,
993 // so do a full call-c here.
994 load_const(R11, (address)fd, R0);
996 bool has_env = (fd != NULL && fd->env() != NULL);
997 return branch_to(R11, /*and_link=*/true,
998 /*save toc=*/false,
999 /*restore toc=*/false,
1000 /*load toc=*/true,
1001 /*load env=*/has_env);
1002 } else {
1003 // It's a friend function. Load the entry point and don't care about
1004 // toc and env. Use an optimizable call instruction, but ensure the
1005 // same code-size as in the case of a non-friend function.
1006 nop();
1007 nop();
1008 nop();
1009 bl64_patchable(fd->entry(), rt);
1010 _last_calls_return_pc = pc();
1011 return _last_calls_return_pc;
1012 }
1013 } else {
1014 // This call does not need to be relocatable, do more aggressive
1015 // optimizations.
1016 if (!ReoptimizeCallSequences
1017 || !fd->is_friend_function()) {
1018 // It's not a friend function as defined by class FunctionDescriptor,
1019 // so do a full call-c here.
1020 load_const(R11, (address)fd, R0);
1021 return branch_to(R11, /*and_link=*/true,
1022 /*save toc=*/false,
1023 /*restore toc=*/false,
1024 /*load toc=*/true,
1025 /*load env=*/true);
1026 } else {
1027 // it's a friend function, load the entry point and don't care about
1028 // toc and env.
1029 address dest = fd->entry();
1030 if (is_within_range_of_b(dest, pc())) {
1031 bl(dest);
1032 } else {
1033 bl64_patchable(dest, rt);
1034 }
1035 _last_calls_return_pc = pc();
1036 return _last_calls_return_pc;
1037 }
1038 }
1039 }
1041 // Call a C function. All constants needed reside in TOC.
1042 //
1043 // Read the address to call from the TOC.
1044 // Read env from TOC, if fd specifies an env.
1045 // Read new TOC from TOC.
1046 address MacroAssembler::call_c_using_toc(const FunctionDescriptor* fd,
1047 relocInfo::relocType rt, Register toc) {
1048 if (!ReoptimizeCallSequences
1049 || (rt != relocInfo::runtime_call_type && rt != relocInfo::none)
1050 || !fd->is_friend_function()) {
1051 // It's not a friend function as defined by class FunctionDescriptor,
1052 // so do a full call-c here.
1053 assert(fd->entry() != NULL, "function must be linked");
1055 AddressLiteral fd_entry(fd->entry());
1056 load_const_from_method_toc(R11, fd_entry, toc);
1057 mtctr(R11);
1058 if (fd->env() == NULL) {
1059 li(R11, 0);
1060 nop();
1061 } else {
1062 AddressLiteral fd_env(fd->env());
1063 load_const_from_method_toc(R11, fd_env, toc);
1064 }
1065 AddressLiteral fd_toc(fd->toc());
1066 load_toc_from_toc(R2_TOC, fd_toc, toc);
1067 // R2_TOC is killed.
1068 bctrl();
1069 _last_calls_return_pc = pc();
1070 } else {
1071 // It's a friend function, load the entry point and don't care about
1072 // toc and env. Use an optimizable call instruction, but ensure the
1073 // same code-size as in the case of a non-friend function.
1074 nop();
1075 bl64_patchable(fd->entry(), rt);
1076 _last_calls_return_pc = pc();
1077 }
1078 return _last_calls_return_pc;
1079 }
1081 void MacroAssembler::call_VM_base(Register oop_result,
1082 Register last_java_sp,
1083 address entry_point,
1084 bool check_exceptions) {
1085 BLOCK_COMMENT("call_VM {");
1086 // Determine last_java_sp register.
1087 if (!last_java_sp->is_valid()) {
1088 last_java_sp = R1_SP;
1089 }
1090 set_top_ijava_frame_at_SP_as_last_Java_frame(last_java_sp, R11_scratch1);
1092 // ARG1 must hold thread address.
1093 mr(R3_ARG1, R16_thread);
1095 address return_pc = call_c((FunctionDescriptor*)entry_point, relocInfo::none);
1097 reset_last_Java_frame();
1099 // Check for pending exceptions.
1100 if (check_exceptions) {
1101 // We don't check for exceptions here.
1102 ShouldNotReachHere();
1103 }
1105 // Get oop result if there is one and reset the value in the thread.
1106 if (oop_result->is_valid()) {
1107 get_vm_result(oop_result);
1108 }
1110 _last_calls_return_pc = return_pc;
1111 BLOCK_COMMENT("} call_VM");
1112 }
1114 void MacroAssembler::call_VM_leaf_base(address entry_point) {
1115 BLOCK_COMMENT("call_VM_leaf {");
1116 call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, entry_point), relocInfo::none);
1117 BLOCK_COMMENT("} call_VM_leaf");
1118 }
1120 void MacroAssembler::call_VM(Register oop_result, address entry_point, bool check_exceptions) {
1121 call_VM_base(oop_result, noreg, entry_point, check_exceptions);
1122 }
1124 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1,
1125 bool check_exceptions) {
1126 // R3_ARG1 is reserved for the thread.
1127 mr_if_needed(R4_ARG2, arg_1);
1128 call_VM(oop_result, entry_point, check_exceptions);
1129 }
1131 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2,
1132 bool check_exceptions) {
1133 // R3_ARG1 is reserved for the thread
1134 mr_if_needed(R4_ARG2, arg_1);
1135 assert(arg_2 != R4_ARG2, "smashed argument");
1136 mr_if_needed(R5_ARG3, arg_2);
1137 call_VM(oop_result, entry_point, check_exceptions);
1138 }
1140 void MacroAssembler::call_VM_leaf(address entry_point) {
1141 call_VM_leaf_base(entry_point);
1142 }
1144 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1) {
1145 mr_if_needed(R3_ARG1, arg_1);
1146 call_VM_leaf(entry_point);
1147 }
1149 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1, Register arg_2) {
1150 mr_if_needed(R3_ARG1, arg_1);
1151 assert(arg_2 != R3_ARG1, "smashed argument");
1152 mr_if_needed(R4_ARG2, arg_2);
1153 call_VM_leaf(entry_point);
1154 }
1156 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1, Register arg_2, Register arg_3) {
1157 mr_if_needed(R3_ARG1, arg_1);
1158 assert(arg_2 != R3_ARG1, "smashed argument");
1159 mr_if_needed(R4_ARG2, arg_2);
1160 assert(arg_3 != R3_ARG1 && arg_3 != R4_ARG2, "smashed argument");
1161 mr_if_needed(R5_ARG3, arg_3);
1162 call_VM_leaf(entry_point);
1163 }
1165 // Check whether instruction is a read access to the polling page
1166 // which was emitted by load_from_polling_page(..).
1167 bool MacroAssembler::is_load_from_polling_page(int instruction, void* ucontext,
1168 address* polling_address_ptr) {
1169 if (!is_ld(instruction))
1170 return false; // It's not a ld. Fail.
1172 int rt = inv_rt_field(instruction);
1173 int ra = inv_ra_field(instruction);
1174 int ds = inv_ds_field(instruction);
1175 if (!(ds == 0 && ra != 0 && rt == 0)) {
1176 return false; // It's not a ld(r0, X, ra). Fail.
1177 }
1179 if (!ucontext) {
1180 // Set polling address.
1181 if (polling_address_ptr != NULL) {
1182 *polling_address_ptr = NULL;
1183 }
1184 return true; // No ucontext given. Can't check value of ra. Assume true.
1185 }
1187 #ifdef LINUX
1188 // Ucontext given. Check that register ra contains the address of
1189 // the safepoing polling page.
1190 ucontext_t* uc = (ucontext_t*) ucontext;
1191 // Set polling address.
1192 address addr = (address)uc->uc_mcontext.regs->gpr[ra] + (ssize_t)ds;
1193 if (polling_address_ptr != NULL) {
1194 *polling_address_ptr = addr;
1195 }
1196 return os::is_poll_address(addr);
1197 #else
1198 // Not on Linux, ucontext must be NULL.
1199 ShouldNotReachHere();
1200 return false;
1201 #endif
1202 }
1204 bool MacroAssembler::is_memory_serialization(int instruction, JavaThread* thread, void* ucontext) {
1205 #ifdef LINUX
1206 ucontext_t* uc = (ucontext_t*) ucontext;
1208 if (is_stwx(instruction) || is_stwux(instruction)) {
1209 int ra = inv_ra_field(instruction);
1210 int rb = inv_rb_field(instruction);
1212 // look up content of ra and rb in ucontext
1213 address ra_val=(address)uc->uc_mcontext.regs->gpr[ra];
1214 long rb_val=(long)uc->uc_mcontext.regs->gpr[rb];
1215 return os::is_memory_serialize_page(thread, ra_val+rb_val);
1216 } else if (is_stw(instruction) || is_stwu(instruction)) {
1217 int ra = inv_ra_field(instruction);
1218 int d1 = inv_d1_field(instruction);
1220 // look up content of ra in ucontext
1221 address ra_val=(address)uc->uc_mcontext.regs->gpr[ra];
1222 return os::is_memory_serialize_page(thread, ra_val+d1);
1223 } else {
1224 return false;
1225 }
1226 #else
1227 // workaround not needed on !LINUX :-)
1228 ShouldNotCallThis();
1229 return false;
1230 #endif
1231 }
1233 void MacroAssembler::bang_stack_with_offset(int offset) {
1234 // When increasing the stack, the old stack pointer will be written
1235 // to the new top of stack according to the PPC64 abi.
1236 // Therefore, stack banging is not necessary when increasing
1237 // the stack by <= os::vm_page_size() bytes.
1238 // When increasing the stack by a larger amount, this method is
1239 // called repeatedly to bang the intermediate pages.
1241 // Stack grows down, caller passes positive offset.
1242 assert(offset > 0, "must bang with positive offset");
1244 long stdoffset = -offset;
1246 if (is_simm(stdoffset, 16)) {
1247 // Signed 16 bit offset, a simple std is ok.
1248 if (UseLoadInstructionsForStackBangingPPC64) {
1249 ld(R0, (int)(signed short)stdoffset, R1_SP);
1250 } else {
1251 std(R0,(int)(signed short)stdoffset, R1_SP);
1252 }
1253 } else if (is_simm(stdoffset, 31)) {
1254 const int hi = MacroAssembler::largeoffset_si16_si16_hi(stdoffset);
1255 const int lo = MacroAssembler::largeoffset_si16_si16_lo(stdoffset);
1257 Register tmp = R11;
1258 addis(tmp, R1_SP, hi);
1259 if (UseLoadInstructionsForStackBangingPPC64) {
1260 ld(R0, lo, tmp);
1261 } else {
1262 std(R0, lo, tmp);
1263 }
1264 } else {
1265 ShouldNotReachHere();
1266 }
1267 }
1269 // If instruction is a stack bang of the form
1270 // std R0, x(Ry), (see bang_stack_with_offset())
1271 // stdu R1_SP, x(R1_SP), (see push_frame(), resize_frame())
1272 // or stdux R1_SP, Rx, R1_SP (see push_frame(), resize_frame())
1273 // return the banged address. Otherwise, return 0.
1274 address MacroAssembler::get_stack_bang_address(int instruction, void *ucontext) {
1275 #ifdef LINUX
1276 ucontext_t* uc = (ucontext_t*) ucontext;
1277 int rs = inv_rs_field(instruction);
1278 int ra = inv_ra_field(instruction);
1279 if ( (is_ld(instruction) && rs == 0 && UseLoadInstructionsForStackBangingPPC64)
1280 || (is_std(instruction) && rs == 0 && !UseLoadInstructionsForStackBangingPPC64)
1281 || (is_stdu(instruction) && rs == 1)) {
1282 int ds = inv_ds_field(instruction);
1283 // return banged address
1284 return ds+(address)uc->uc_mcontext.regs->gpr[ra];
1285 } else if (is_stdux(instruction) && rs == 1) {
1286 int rb = inv_rb_field(instruction);
1287 address sp = (address)uc->uc_mcontext.regs->gpr[1];
1288 long rb_val = (long)uc->uc_mcontext.regs->gpr[rb];
1289 return ra != 1 || rb_val >= 0 ? NULL // not a stack bang
1290 : sp + rb_val; // banged address
1291 }
1292 return NULL; // not a stack bang
1293 #else
1294 // workaround not needed on !LINUX :-)
1295 ShouldNotCallThis();
1296 return NULL;
1297 #endif
1298 }
1300 // CmpxchgX sets condition register to cmpX(current, compare).
1301 void MacroAssembler::cmpxchgw(ConditionRegister flag, Register dest_current_value,
1302 Register compare_value, Register exchange_value,
1303 Register addr_base, int semantics, bool cmpxchgx_hint,
1304 Register int_flag_success, bool contention_hint) {
1305 Label retry;
1306 Label failed;
1307 Label done;
1309 // Save one branch if result is returned via register and
1310 // result register is different from the other ones.
1311 bool use_result_reg = (int_flag_success != noreg);
1312 bool preset_result_reg = (int_flag_success != dest_current_value && int_flag_success != compare_value &&
1313 int_flag_success != exchange_value && int_flag_success != addr_base);
1315 // release/fence semantics
1316 if (semantics & MemBarRel) {
1317 release();
1318 }
1320 if (use_result_reg && preset_result_reg) {
1321 li(int_flag_success, 0); // preset (assume cas failed)
1322 }
1324 // Add simple guard in order to reduce risk of starving under high contention (recommended by IBM).
1325 if (contention_hint) { // Don't try to reserve if cmp fails.
1326 lwz(dest_current_value, 0, addr_base);
1327 cmpw(flag, dest_current_value, compare_value);
1328 bne(flag, failed);
1329 }
1331 // atomic emulation loop
1332 bind(retry);
1334 lwarx(dest_current_value, addr_base, cmpxchgx_hint);
1335 cmpw(flag, dest_current_value, compare_value);
1336 if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1337 bne_predict_not_taken(flag, failed);
1338 } else {
1339 bne( flag, failed);
1340 }
1341 // branch to done => (flag == ne), (dest_current_value != compare_value)
1342 // fall through => (flag == eq), (dest_current_value == compare_value)
1344 stwcx_(exchange_value, addr_base);
1345 if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1346 bne_predict_not_taken(CCR0, retry); // StXcx_ sets CCR0.
1347 } else {
1348 bne( CCR0, retry); // StXcx_ sets CCR0.
1349 }
1350 // fall through => (flag == eq), (dest_current_value == compare_value), (swapped)
1352 // Result in register (must do this at the end because int_flag_success can be the
1353 // same register as one above).
1354 if (use_result_reg) {
1355 li(int_flag_success, 1);
1356 }
1358 if (semantics & MemBarFenceAfter) {
1359 fence();
1360 } else if (semantics & MemBarAcq) {
1361 isync();
1362 }
1364 if (use_result_reg && !preset_result_reg) {
1365 b(done);
1366 }
1368 bind(failed);
1369 if (use_result_reg && !preset_result_reg) {
1370 li(int_flag_success, 0);
1371 }
1373 bind(done);
1374 // (flag == ne) => (dest_current_value != compare_value), (!swapped)
1375 // (flag == eq) => (dest_current_value == compare_value), ( swapped)
1376 }
1378 // Preforms atomic compare exchange:
1379 // if (compare_value == *addr_base)
1380 // *addr_base = exchange_value
1381 // int_flag_success = 1;
1382 // else
1383 // int_flag_success = 0;
1384 //
1385 // ConditionRegister flag = cmp(compare_value, *addr_base)
1386 // Register dest_current_value = *addr_base
1387 // Register compare_value Used to compare with value in memory
1388 // Register exchange_value Written to memory if compare_value == *addr_base
1389 // Register addr_base The memory location to compareXChange
1390 // Register int_flag_success Set to 1 if exchange_value was written to *addr_base
1391 //
1392 // To avoid the costly compare exchange the value is tested beforehand.
1393 // Several special cases exist to avoid that unnecessary information is generated.
1394 //
1395 void MacroAssembler::cmpxchgd(ConditionRegister flag,
1396 Register dest_current_value, Register compare_value, Register exchange_value,
1397 Register addr_base, int semantics, bool cmpxchgx_hint,
1398 Register int_flag_success, Label* failed_ext, bool contention_hint) {
1399 Label retry;
1400 Label failed_int;
1401 Label& failed = (failed_ext != NULL) ? *failed_ext : failed_int;
1402 Label done;
1404 // Save one branch if result is returned via register and result register is different from the other ones.
1405 bool use_result_reg = (int_flag_success!=noreg);
1406 bool preset_result_reg = (int_flag_success!=dest_current_value && int_flag_success!=compare_value &&
1407 int_flag_success!=exchange_value && int_flag_success!=addr_base);
1408 assert(int_flag_success == noreg || failed_ext == NULL, "cannot have both");
1410 // release/fence semantics
1411 if (semantics & MemBarRel) {
1412 release();
1413 }
1415 if (use_result_reg && preset_result_reg) {
1416 li(int_flag_success, 0); // preset (assume cas failed)
1417 }
1419 // Add simple guard in order to reduce risk of starving under high contention (recommended by IBM).
1420 if (contention_hint) { // Don't try to reserve if cmp fails.
1421 ld(dest_current_value, 0, addr_base);
1422 cmpd(flag, dest_current_value, compare_value);
1423 bne(flag, failed);
1424 }
1426 // atomic emulation loop
1427 bind(retry);
1429 ldarx(dest_current_value, addr_base, cmpxchgx_hint);
1430 cmpd(flag, dest_current_value, compare_value);
1431 if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1432 bne_predict_not_taken(flag, failed);
1433 } else {
1434 bne( flag, failed);
1435 }
1437 stdcx_(exchange_value, addr_base);
1438 if (UseStaticBranchPredictionInCompareAndSwapPPC64) {
1439 bne_predict_not_taken(CCR0, retry); // stXcx_ sets CCR0
1440 } else {
1441 bne( CCR0, retry); // stXcx_ sets CCR0
1442 }
1444 // result in register (must do this at the end because int_flag_success can be the same register as one above)
1445 if (use_result_reg) {
1446 li(int_flag_success, 1);
1447 }
1449 // POWER6 doesn't need isync in CAS.
1450 // Always emit isync to be on the safe side.
1451 if (semantics & MemBarFenceAfter) {
1452 fence();
1453 } else if (semantics & MemBarAcq) {
1454 isync();
1455 }
1457 if (use_result_reg && !preset_result_reg) {
1458 b(done);
1459 }
1461 bind(failed_int);
1462 if (use_result_reg && !preset_result_reg) {
1463 li(int_flag_success, 0);
1464 }
1466 bind(done);
1467 // (flag == ne) => (dest_current_value != compare_value), (!swapped)
1468 // (flag == eq) => (dest_current_value == compare_value), ( swapped)
1469 }
1471 // Look up the method for a megamorphic invokeinterface call.
1472 // The target method is determined by <intf_klass, itable_index>.
1473 // The receiver klass is in recv_klass.
1474 // On success, the result will be in method_result, and execution falls through.
1475 // On failure, execution transfers to the given label.
1476 void MacroAssembler::lookup_interface_method(Register recv_klass,
1477 Register intf_klass,
1478 RegisterOrConstant itable_index,
1479 Register method_result,
1480 Register scan_temp,
1481 Register sethi_temp,
1482 Label& L_no_such_interface) {
1483 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp);
1484 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
1485 "caller must use same register for non-constant itable index as for method");
1487 // Compute start of first itableOffsetEntry (which is at the end of the vtable).
1488 int vtable_base = InstanceKlass::vtable_start_offset() * wordSize;
1489 int itentry_off = itableMethodEntry::method_offset_in_bytes();
1490 int logMEsize = exact_log2(itableMethodEntry::size() * wordSize);
1491 int scan_step = itableOffsetEntry::size() * wordSize;
1492 int log_vte_size= exact_log2(vtableEntry::size() * wordSize);
1494 lwz(scan_temp, InstanceKlass::vtable_length_offset() * wordSize, recv_klass);
1495 // %%% We should store the aligned, prescaled offset in the klassoop.
1496 // Then the next several instructions would fold away.
1498 sldi(scan_temp, scan_temp, log_vte_size);
1499 addi(scan_temp, scan_temp, vtable_base);
1500 add(scan_temp, recv_klass, scan_temp);
1502 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
1503 if (itable_index.is_register()) {
1504 Register itable_offset = itable_index.as_register();
1505 sldi(itable_offset, itable_offset, logMEsize);
1506 if (itentry_off) addi(itable_offset, itable_offset, itentry_off);
1507 add(recv_klass, itable_offset, recv_klass);
1508 } else {
1509 long itable_offset = (long)itable_index.as_constant();
1510 load_const_optimized(sethi_temp, (itable_offset<<logMEsize)+itentry_off); // static address, no relocation
1511 add(recv_klass, sethi_temp, recv_klass);
1512 }
1514 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
1515 // if (scan->interface() == intf) {
1516 // result = (klass + scan->offset() + itable_index);
1517 // }
1518 // }
1519 Label search, found_method;
1521 for (int peel = 1; peel >= 0; peel--) {
1522 // %%%% Could load both offset and interface in one ldx, if they were
1523 // in the opposite order. This would save a load.
1524 ld(method_result, itableOffsetEntry::interface_offset_in_bytes(), scan_temp);
1526 // Check that this entry is non-null. A null entry means that
1527 // the receiver class doesn't implement the interface, and wasn't the
1528 // same as when the caller was compiled.
1529 cmpd(CCR0, method_result, intf_klass);
1531 if (peel) {
1532 beq(CCR0, found_method);
1533 } else {
1534 bne(CCR0, search);
1535 // (invert the test to fall through to found_method...)
1536 }
1538 if (!peel) break;
1540 bind(search);
1542 cmpdi(CCR0, method_result, 0);
1543 beq(CCR0, L_no_such_interface);
1544 addi(scan_temp, scan_temp, scan_step);
1545 }
1547 bind(found_method);
1549 // Got a hit.
1550 int ito_offset = itableOffsetEntry::offset_offset_in_bytes();
1551 lwz(scan_temp, ito_offset, scan_temp);
1552 ldx(method_result, scan_temp, recv_klass);
1553 }
1555 // virtual method calling
1556 void MacroAssembler::lookup_virtual_method(Register recv_klass,
1557 RegisterOrConstant vtable_index,
1558 Register method_result) {
1560 assert_different_registers(recv_klass, method_result, vtable_index.register_or_noreg());
1562 const int base = InstanceKlass::vtable_start_offset() * wordSize;
1563 assert(vtableEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1565 if (vtable_index.is_register()) {
1566 sldi(vtable_index.as_register(), vtable_index.as_register(), LogBytesPerWord);
1567 add(recv_klass, vtable_index.as_register(), recv_klass);
1568 } else {
1569 addi(recv_klass, recv_klass, vtable_index.as_constant() << LogBytesPerWord);
1570 }
1571 ld(R19_method, base + vtableEntry::method_offset_in_bytes(), recv_klass);
1572 }
1574 /////////////////////////////////////////// subtype checking ////////////////////////////////////////////
1576 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
1577 Register super_klass,
1578 Register temp1_reg,
1579 Register temp2_reg,
1580 Label& L_success,
1581 Label& L_failure) {
1583 const Register check_cache_offset = temp1_reg;
1584 const Register cached_super = temp2_reg;
1586 assert_different_registers(sub_klass, super_klass, check_cache_offset, cached_super);
1588 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1589 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
1591 // If the pointers are equal, we are done (e.g., String[] elements).
1592 // This self-check enables sharing of secondary supertype arrays among
1593 // non-primary types such as array-of-interface. Otherwise, each such
1594 // type would need its own customized SSA.
1595 // We move this check to the front of the fast path because many
1596 // type checks are in fact trivially successful in this manner,
1597 // so we get a nicely predicted branch right at the start of the check.
1598 cmpd(CCR0, sub_klass, super_klass);
1599 beq(CCR0, L_success);
1601 // Check the supertype display:
1602 lwz(check_cache_offset, sco_offset, super_klass);
1603 // The loaded value is the offset from KlassOopDesc.
1605 ldx(cached_super, check_cache_offset, sub_klass);
1606 cmpd(CCR0, cached_super, super_klass);
1607 beq(CCR0, L_success);
1609 // This check has worked decisively for primary supers.
1610 // Secondary supers are sought in the super_cache ('super_cache_addr').
1611 // (Secondary supers are interfaces and very deeply nested subtypes.)
1612 // This works in the same check above because of a tricky aliasing
1613 // between the super_cache and the primary super display elements.
1614 // (The 'super_check_addr' can address either, as the case requires.)
1615 // Note that the cache is updated below if it does not help us find
1616 // what we need immediately.
1617 // So if it was a primary super, we can just fail immediately.
1618 // Otherwise, it's the slow path for us (no success at this point).
1620 cmpwi(CCR0, check_cache_offset, sc_offset);
1621 bne(CCR0, L_failure);
1622 // bind(slow_path); // fallthru
1623 }
1625 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
1626 Register super_klass,
1627 Register temp1_reg,
1628 Register temp2_reg,
1629 Label* L_success,
1630 Register result_reg) {
1631 const Register array_ptr = temp1_reg; // current value from cache array
1632 const Register temp = temp2_reg;
1634 assert_different_registers(sub_klass, super_klass, array_ptr, temp);
1636 int source_offset = in_bytes(Klass::secondary_supers_offset());
1637 int target_offset = in_bytes(Klass::secondary_super_cache_offset());
1639 int length_offset = Array<Klass*>::length_offset_in_bytes();
1640 int base_offset = Array<Klass*>::base_offset_in_bytes();
1642 Label hit, loop, failure, fallthru;
1644 ld(array_ptr, source_offset, sub_klass);
1646 //assert(4 == arrayOopDesc::length_length_in_bytes(), "precondition violated.");
1647 lwz(temp, length_offset, array_ptr);
1648 cmpwi(CCR0, temp, 0);
1649 beq(CCR0, result_reg!=noreg ? failure : fallthru); // length 0
1651 mtctr(temp); // load ctr
1653 bind(loop);
1654 // Oops in table are NO MORE compressed.
1655 ld(temp, base_offset, array_ptr);
1656 cmpd(CCR0, temp, super_klass);
1657 beq(CCR0, hit);
1658 addi(array_ptr, array_ptr, BytesPerWord);
1659 bdnz(loop);
1661 bind(failure);
1662 if (result_reg!=noreg) li(result_reg, 1); // load non-zero result (indicates a miss)
1663 b(fallthru);
1665 bind(hit);
1666 std(super_klass, target_offset, sub_klass); // save result to cache
1667 if (result_reg != noreg) li(result_reg, 0); // load zero result (indicates a hit)
1668 if (L_success != NULL) b(*L_success);
1670 bind(fallthru);
1671 }
1673 // Try fast path, then go to slow one if not successful
1674 void MacroAssembler::check_klass_subtype(Register sub_klass,
1675 Register super_klass,
1676 Register temp1_reg,
1677 Register temp2_reg,
1678 Label& L_success) {
1679 Label L_failure;
1680 check_klass_subtype_fast_path(sub_klass, super_klass, temp1_reg, temp2_reg, L_success, L_failure);
1681 check_klass_subtype_slow_path(sub_klass, super_klass, temp1_reg, temp2_reg, &L_success);
1682 bind(L_failure); // Fallthru if not successful.
1683 }
1685 void MacroAssembler::check_method_handle_type(Register mtype_reg, Register mh_reg,
1686 Register temp_reg,
1687 Label& wrong_method_type) {
1688 assert_different_registers(mtype_reg, mh_reg, temp_reg);
1689 // Compare method type against that of the receiver.
1690 load_heap_oop_not_null(temp_reg, delayed_value(java_lang_invoke_MethodHandle::type_offset_in_bytes, temp_reg), mh_reg);
1691 cmpd(CCR0, temp_reg, mtype_reg);
1692 bne(CCR0, wrong_method_type);
1693 }
1695 RegisterOrConstant MacroAssembler::argument_offset(RegisterOrConstant arg_slot,
1696 Register temp_reg,
1697 int extra_slot_offset) {
1698 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
1699 int stackElementSize = Interpreter::stackElementSize;
1700 int offset = extra_slot_offset * stackElementSize;
1701 if (arg_slot.is_constant()) {
1702 offset += arg_slot.as_constant() * stackElementSize;
1703 return offset;
1704 } else {
1705 assert(temp_reg != noreg, "must specify");
1706 sldi(temp_reg, arg_slot.as_register(), exact_log2(stackElementSize));
1707 if (offset != 0)
1708 addi(temp_reg, temp_reg, offset);
1709 return temp_reg;
1710 }
1711 }
1713 void MacroAssembler::biased_locking_enter(ConditionRegister cr_reg, Register obj_reg,
1714 Register mark_reg, Register temp_reg,
1715 Register temp2_reg, Label& done, Label* slow_case) {
1716 assert(UseBiasedLocking, "why call this otherwise?");
1718 #ifdef ASSERT
1719 assert_different_registers(obj_reg, mark_reg, temp_reg, temp2_reg);
1720 #endif
1722 Label cas_label;
1724 // Branch to done if fast path fails and no slow_case provided.
1725 Label *slow_case_int = (slow_case != NULL) ? slow_case : &done;
1727 // Biased locking
1728 // See whether the lock is currently biased toward our thread and
1729 // whether the epoch is still valid
1730 // Note that the runtime guarantees sufficient alignment of JavaThread
1731 // pointers to allow age to be placed into low bits
1732 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits,
1733 "biased locking makes assumptions about bit layout");
1735 if (PrintBiasedLockingStatistics) {
1736 load_const(temp_reg, (address) BiasedLocking::total_entry_count_addr(), temp2_reg);
1737 lwz(temp2_reg, 0, temp_reg);
1738 addi(temp2_reg, temp2_reg, 1);
1739 stw(temp2_reg, 0, temp_reg);
1740 }
1742 andi(temp_reg, mark_reg, markOopDesc::biased_lock_mask_in_place);
1743 cmpwi(cr_reg, temp_reg, markOopDesc::biased_lock_pattern);
1744 bne(cr_reg, cas_label);
1746 load_klass_with_trap_null_check(temp_reg, obj_reg);
1748 load_const_optimized(temp2_reg, ~((int) markOopDesc::age_mask_in_place));
1749 ld(temp_reg, in_bytes(Klass::prototype_header_offset()), temp_reg);
1750 orr(temp_reg, R16_thread, temp_reg);
1751 xorr(temp_reg, mark_reg, temp_reg);
1752 andr(temp_reg, temp_reg, temp2_reg);
1753 cmpdi(cr_reg, temp_reg, 0);
1754 if (PrintBiasedLockingStatistics) {
1755 Label l;
1756 bne(cr_reg, l);
1757 load_const(mark_reg, (address) BiasedLocking::biased_lock_entry_count_addr());
1758 lwz(temp2_reg, 0, mark_reg);
1759 addi(temp2_reg, temp2_reg, 1);
1760 stw(temp2_reg, 0, mark_reg);
1761 // restore mark_reg
1762 ld(mark_reg, oopDesc::mark_offset_in_bytes(), obj_reg);
1763 bind(l);
1764 }
1765 beq(cr_reg, done);
1767 Label try_revoke_bias;
1768 Label try_rebias;
1770 // At this point we know that the header has the bias pattern and
1771 // that we are not the bias owner in the current epoch. We need to
1772 // figure out more details about the state of the header in order to
1773 // know what operations can be legally performed on the object's
1774 // header.
1776 // If the low three bits in the xor result aren't clear, that means
1777 // the prototype header is no longer biased and we have to revoke
1778 // the bias on this object.
1779 andi(temp2_reg, temp_reg, markOopDesc::biased_lock_mask_in_place);
1780 cmpwi(cr_reg, temp2_reg, 0);
1781 bne(cr_reg, try_revoke_bias);
1783 // Biasing is still enabled for this data type. See whether the
1784 // epoch of the current bias is still valid, meaning that the epoch
1785 // bits of the mark word are equal to the epoch bits of the
1786 // prototype header. (Note that the prototype header's epoch bits
1787 // only change at a safepoint.) If not, attempt to rebias the object
1788 // toward the current thread. Note that we must be absolutely sure
1789 // that the current epoch is invalid in order to do this because
1790 // otherwise the manipulations it performs on the mark word are
1791 // illegal.
1793 int shift_amount = 64 - markOopDesc::epoch_shift;
1794 // rotate epoch bits to right (little) end and set other bits to 0
1795 // [ big part | epoch | little part ] -> [ 0..0 | epoch ]
1796 rldicl_(temp2_reg, temp_reg, shift_amount, 64 - markOopDesc::epoch_bits);
1797 // branch if epoch bits are != 0, i.e. they differ, because the epoch has been incremented
1798 bne(CCR0, try_rebias);
1800 // The epoch of the current bias is still valid but we know nothing
1801 // about the owner; it might be set or it might be clear. Try to
1802 // acquire the bias of the object using an atomic operation. If this
1803 // fails we will go in to the runtime to revoke the object's bias.
1804 // Note that we first construct the presumed unbiased header so we
1805 // don't accidentally blow away another thread's valid bias.
1806 andi(mark_reg, mark_reg, (markOopDesc::biased_lock_mask_in_place |
1807 markOopDesc::age_mask_in_place |
1808 markOopDesc::epoch_mask_in_place));
1809 orr(temp_reg, R16_thread, mark_reg);
1811 assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0");
1813 // CmpxchgX sets cr_reg to cmpX(temp2_reg, mark_reg).
1814 fence(); // TODO: replace by MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq ?
1815 cmpxchgd(/*flag=*/cr_reg, /*current_value=*/temp2_reg,
1816 /*compare_value=*/mark_reg, /*exchange_value=*/temp_reg,
1817 /*where=*/obj_reg,
1818 MacroAssembler::MemBarAcq,
1819 MacroAssembler::cmpxchgx_hint_acquire_lock(),
1820 noreg, slow_case_int); // bail out if failed
1822 // If the biasing toward our thread failed, this means that
1823 // another thread succeeded in biasing it toward itself and we
1824 // need to revoke that bias. The revocation will occur in the
1825 // interpreter runtime in the slow case.
1826 if (PrintBiasedLockingStatistics) {
1827 load_const(temp_reg, (address) BiasedLocking::anonymously_biased_lock_entry_count_addr(), temp2_reg);
1828 lwz(temp2_reg, 0, temp_reg);
1829 addi(temp2_reg, temp2_reg, 1);
1830 stw(temp2_reg, 0, temp_reg);
1831 }
1832 b(done);
1834 bind(try_rebias);
1835 // At this point we know the epoch has expired, meaning that the
1836 // current "bias owner", if any, is actually invalid. Under these
1837 // circumstances _only_, we are allowed to use the current header's
1838 // value as the comparison value when doing the cas to acquire the
1839 // bias in the current epoch. In other words, we allow transfer of
1840 // the bias from one thread to another directly in this situation.
1841 andi(temp_reg, mark_reg, markOopDesc::age_mask_in_place);
1842 orr(temp_reg, R16_thread, temp_reg);
1843 load_klass_with_trap_null_check(temp2_reg, obj_reg);
1844 ld(temp2_reg, in_bytes(Klass::prototype_header_offset()), temp2_reg);
1845 orr(temp_reg, temp_reg, temp2_reg);
1847 assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0");
1849 // CmpxchgX sets cr_reg to cmpX(temp2_reg, mark_reg).
1850 fence(); // TODO: replace by MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq ?
1851 cmpxchgd(/*flag=*/cr_reg, /*current_value=*/temp2_reg,
1852 /*compare_value=*/mark_reg, /*exchange_value=*/temp_reg,
1853 /*where=*/obj_reg,
1854 MacroAssembler::MemBarAcq,
1855 MacroAssembler::cmpxchgx_hint_acquire_lock(),
1856 noreg, slow_case_int); // bail out if failed
1858 // If the biasing toward our thread failed, this means that
1859 // another thread succeeded in biasing it toward itself and we
1860 // need to revoke that bias. The revocation will occur in the
1861 // interpreter runtime in the slow case.
1862 if (PrintBiasedLockingStatistics) {
1863 load_const(temp_reg, (address) BiasedLocking::rebiased_lock_entry_count_addr(), temp2_reg);
1864 lwz(temp2_reg, 0, temp_reg);
1865 addi(temp2_reg, temp2_reg, 1);
1866 stw(temp2_reg, 0, temp_reg);
1867 }
1868 b(done);
1870 bind(try_revoke_bias);
1871 // The prototype mark in the klass doesn't have the bias bit set any
1872 // more, indicating that objects of this data type are not supposed
1873 // to be biased any more. We are going to try to reset the mark of
1874 // this object to the prototype value and fall through to the
1875 // CAS-based locking scheme. Note that if our CAS fails, it means
1876 // that another thread raced us for the privilege of revoking the
1877 // bias of this particular object, so it's okay to continue in the
1878 // normal locking code.
1879 load_klass_with_trap_null_check(temp_reg, obj_reg);
1880 ld(temp_reg, in_bytes(Klass::prototype_header_offset()), temp_reg);
1881 andi(temp2_reg, mark_reg, markOopDesc::age_mask_in_place);
1882 orr(temp_reg, temp_reg, temp2_reg);
1884 assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0");
1886 // CmpxchgX sets cr_reg to cmpX(temp2_reg, mark_reg).
1887 fence(); // TODO: replace by MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq ?
1888 cmpxchgd(/*flag=*/cr_reg, /*current_value=*/temp2_reg,
1889 /*compare_value=*/mark_reg, /*exchange_value=*/temp_reg,
1890 /*where=*/obj_reg,
1891 MacroAssembler::MemBarAcq,
1892 MacroAssembler::cmpxchgx_hint_acquire_lock());
1894 // reload markOop in mark_reg before continuing with lightweight locking
1895 ld(mark_reg, oopDesc::mark_offset_in_bytes(), obj_reg);
1897 // Fall through to the normal CAS-based lock, because no matter what
1898 // the result of the above CAS, some thread must have succeeded in
1899 // removing the bias bit from the object's header.
1900 if (PrintBiasedLockingStatistics) {
1901 Label l;
1902 bne(cr_reg, l);
1903 load_const(temp_reg, (address) BiasedLocking::revoked_lock_entry_count_addr(), temp2_reg);
1904 lwz(temp2_reg, 0, temp_reg);
1905 addi(temp2_reg, temp2_reg, 1);
1906 stw(temp2_reg, 0, temp_reg);
1907 bind(l);
1908 }
1910 bind(cas_label);
1911 }
1913 void MacroAssembler::biased_locking_exit (ConditionRegister cr_reg, Register mark_addr, Register temp_reg, Label& done) {
1914 // Check for biased locking unlock case, which is a no-op
1915 // Note: we do not have to check the thread ID for two reasons.
1916 // First, the interpreter checks for IllegalMonitorStateException at
1917 // a higher level. Second, if the bias was revoked while we held the
1918 // lock, the object could not be rebiased toward another thread, so
1919 // the bias bit would be clear.
1921 ld(temp_reg, 0, mark_addr);
1922 andi(temp_reg, temp_reg, markOopDesc::biased_lock_mask_in_place);
1924 cmpwi(cr_reg, temp_reg, markOopDesc::biased_lock_pattern);
1925 beq(cr_reg, done);
1926 }
1928 // "The box" is the space on the stack where we copy the object mark.
1929 void MacroAssembler::compiler_fast_lock_object(ConditionRegister flag, Register oop, Register box,
1930 Register temp, Register displaced_header, Register current_header) {
1931 assert_different_registers(oop, box, temp, displaced_header, current_header);
1932 assert(flag != CCR0, "bad condition register");
1933 Label cont;
1934 Label object_has_monitor;
1935 Label cas_failed;
1937 // Load markOop from object into displaced_header.
1938 ld(displaced_header, oopDesc::mark_offset_in_bytes(), oop);
1941 // Always do locking in runtime.
1942 if (EmitSync & 0x01) {
1943 cmpdi(flag, oop, 0); // Oop can't be 0 here => always false.
1944 return;
1945 }
1947 if (UseBiasedLocking) {
1948 biased_locking_enter(flag, oop, displaced_header, temp, current_header, cont);
1949 }
1951 // Handle existing monitor.
1952 if ((EmitSync & 0x02) == 0) {
1953 // The object has an existing monitor iff (mark & monitor_value) != 0.
1954 andi_(temp, displaced_header, markOopDesc::monitor_value);
1955 bne(CCR0, object_has_monitor);
1956 }
1958 // Set displaced_header to be (markOop of object | UNLOCK_VALUE).
1959 ori(displaced_header, displaced_header, markOopDesc::unlocked_value);
1961 // Load Compare Value application register.
1963 // Initialize the box. (Must happen before we update the object mark!)
1964 std(displaced_header, BasicLock::displaced_header_offset_in_bytes(), box);
1966 // Must fence, otherwise, preceding store(s) may float below cmpxchg.
1967 // Compare object markOop with mark and if equal exchange scratch1 with object markOop.
1968 // CmpxchgX sets cr_reg to cmpX(current, displaced).
1969 membar(Assembler::StoreStore);
1970 cmpxchgd(/*flag=*/flag,
1971 /*current_value=*/current_header,
1972 /*compare_value=*/displaced_header,
1973 /*exchange_value=*/box,
1974 /*where=*/oop,
1975 MacroAssembler::MemBarAcq,
1976 MacroAssembler::cmpxchgx_hint_acquire_lock(),
1977 noreg,
1978 &cas_failed);
1979 assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0");
1981 // If the compare-and-exchange succeeded, then we found an unlocked
1982 // object and we have now locked it.
1983 b(cont);
1985 bind(cas_failed);
1986 // We did not see an unlocked object so try the fast recursive case.
1988 // Check if the owner is self by comparing the value in the markOop of object
1989 // (current_header) with the stack pointer.
1990 sub(current_header, current_header, R1_SP);
1991 load_const_optimized(temp, (address) (~(os::vm_page_size()-1) |
1992 markOopDesc::lock_mask_in_place));
1994 and_(R0/*==0?*/, current_header, temp);
1995 // If condition is true we are cont and hence we can store 0 as the
1996 // displaced header in the box, which indicates that it is a recursive lock.
1997 mcrf(flag,CCR0);
1998 std(R0/*==0, perhaps*/, BasicLock::displaced_header_offset_in_bytes(), box);
2000 // Handle existing monitor.
2001 if ((EmitSync & 0x02) == 0) {
2002 b(cont);
2004 bind(object_has_monitor);
2005 // The object's monitor m is unlocked iff m->owner == NULL,
2006 // otherwise m->owner may contain a thread or a stack address.
2007 //
2008 // Try to CAS m->owner from NULL to current thread.
2009 addi(temp, displaced_header, ObjectMonitor::owner_offset_in_bytes()-markOopDesc::monitor_value);
2010 li(displaced_header, 0);
2011 // CmpxchgX sets flag to cmpX(current, displaced).
2012 cmpxchgd(/*flag=*/flag,
2013 /*current_value=*/current_header,
2014 /*compare_value=*/displaced_header,
2015 /*exchange_value=*/R16_thread,
2016 /*where=*/temp,
2017 MacroAssembler::MemBarRel | MacroAssembler::MemBarAcq,
2018 MacroAssembler::cmpxchgx_hint_acquire_lock());
2020 // Store a non-null value into the box.
2021 std(box, BasicLock::displaced_header_offset_in_bytes(), box);
2023 # ifdef ASSERT
2024 bne(flag, cont);
2025 // We have acquired the monitor, check some invariants.
2026 addi(/*monitor=*/temp, temp, -ObjectMonitor::owner_offset_in_bytes());
2027 // Invariant 1: _recursions should be 0.
2028 //assert(ObjectMonitor::recursions_size_in_bytes() == 8, "unexpected size");
2029 asm_assert_mem8_is_zero(ObjectMonitor::recursions_offset_in_bytes(), temp,
2030 "monitor->_recursions should be 0", -1);
2031 // Invariant 2: OwnerIsThread shouldn't be 0.
2032 //assert(ObjectMonitor::OwnerIsThread_size_in_bytes() == 4, "unexpected size");
2033 //asm_assert_mem4_isnot_zero(ObjectMonitor::OwnerIsThread_offset_in_bytes(), temp,
2034 // "monitor->OwnerIsThread shouldn't be 0", -1);
2035 # endif
2036 }
2038 bind(cont);
2039 // flag == EQ indicates success
2040 // flag == NE indicates failure
2041 }
2043 void MacroAssembler::compiler_fast_unlock_object(ConditionRegister flag, Register oop, Register box,
2044 Register temp, Register displaced_header, Register current_header) {
2045 assert_different_registers(oop, box, temp, displaced_header, current_header);
2046 assert(flag != CCR0, "bad condition register");
2047 Label cont;
2048 Label object_has_monitor;
2050 // Always do locking in runtime.
2051 if (EmitSync & 0x01) {
2052 cmpdi(flag, oop, 0); // Oop can't be 0 here => always false.
2053 return;
2054 }
2056 if (UseBiasedLocking) {
2057 biased_locking_exit(flag, oop, current_header, cont);
2058 }
2060 // Find the lock address and load the displaced header from the stack.
2061 ld(displaced_header, BasicLock::displaced_header_offset_in_bytes(), box);
2063 // If the displaced header is 0, we have a recursive unlock.
2064 cmpdi(flag, displaced_header, 0);
2065 beq(flag, cont);
2067 // Handle existing monitor.
2068 if ((EmitSync & 0x02) == 0) {
2069 // The object has an existing monitor iff (mark & monitor_value) != 0.
2070 ld(current_header, oopDesc::mark_offset_in_bytes(), oop);
2071 andi(temp, current_header, markOopDesc::monitor_value);
2072 cmpdi(flag, temp, 0);
2073 bne(flag, object_has_monitor);
2074 }
2077 // Check if it is still a light weight lock, this is is true if we see
2078 // the stack address of the basicLock in the markOop of the object.
2079 // Cmpxchg sets flag to cmpd(current_header, box).
2080 cmpxchgd(/*flag=*/flag,
2081 /*current_value=*/current_header,
2082 /*compare_value=*/box,
2083 /*exchange_value=*/displaced_header,
2084 /*where=*/oop,
2085 MacroAssembler::MemBarRel,
2086 MacroAssembler::cmpxchgx_hint_release_lock(),
2087 noreg,
2088 &cont);
2090 assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0");
2092 // Handle existing monitor.
2093 if ((EmitSync & 0x02) == 0) {
2094 b(cont);
2096 bind(object_has_monitor);
2097 addi(current_header, current_header, -markOopDesc::monitor_value); // monitor
2098 ld(temp, ObjectMonitor::owner_offset_in_bytes(), current_header);
2099 ld(displaced_header, ObjectMonitor::recursions_offset_in_bytes(), current_header);
2100 xorr(temp, R16_thread, temp); // Will be 0 if we are the owner.
2101 orr(temp, temp, displaced_header); // Will be 0 if there are 0 recursions.
2102 cmpdi(flag, temp, 0);
2103 bne(flag, cont);
2105 ld(temp, ObjectMonitor::EntryList_offset_in_bytes(), current_header);
2106 ld(displaced_header, ObjectMonitor::cxq_offset_in_bytes(), current_header);
2107 orr(temp, temp, displaced_header); // Will be 0 if both are 0.
2108 cmpdi(flag, temp, 0);
2109 bne(flag, cont);
2110 release();
2111 std(temp, ObjectMonitor::owner_offset_in_bytes(), current_header);
2112 }
2114 bind(cont);
2115 // flag == EQ indicates success
2116 // flag == NE indicates failure
2117 }
2119 // Write serialization page so VM thread can do a pseudo remote membar.
2120 // We use the current thread pointer to calculate a thread specific
2121 // offset to write to within the page. This minimizes bus traffic
2122 // due to cache line collision.
2123 void MacroAssembler::serialize_memory(Register thread, Register tmp1, Register tmp2) {
2124 srdi(tmp2, thread, os::get_serialize_page_shift_count());
2126 int mask = os::vm_page_size() - sizeof(int);
2127 if (Assembler::is_simm(mask, 16)) {
2128 andi(tmp2, tmp2, mask);
2129 } else {
2130 lis(tmp1, (int)((signed short) (mask >> 16)));
2131 ori(tmp1, tmp1, mask & 0x0000ffff);
2132 andr(tmp2, tmp2, tmp1);
2133 }
2135 load_const(tmp1, (long) os::get_memory_serialize_page());
2136 release();
2137 stwx(R0, tmp1, tmp2);
2138 }
2141 // GC barrier helper macros
2143 // Write the card table byte if needed.
2144 void MacroAssembler::card_write_barrier_post(Register Rstore_addr, Register Rnew_val, Register Rtmp) {
2145 CardTableModRefBS* bs = (CardTableModRefBS*) Universe::heap()->barrier_set();
2146 assert(bs->kind() == BarrierSet::CardTableModRef ||
2147 bs->kind() == BarrierSet::CardTableExtension, "wrong barrier");
2148 #ifdef ASSERT
2149 cmpdi(CCR0, Rnew_val, 0);
2150 asm_assert_ne("null oop not allowed", 0x321);
2151 #endif
2152 card_table_write(bs->byte_map_base, Rtmp, Rstore_addr);
2153 }
2155 // Write the card table byte.
2156 void MacroAssembler::card_table_write(jbyte* byte_map_base, Register Rtmp, Register Robj) {
2157 assert_different_registers(Robj, Rtmp, R0);
2158 load_const_optimized(Rtmp, (address)byte_map_base, R0);
2159 srdi(Robj, Robj, CardTableModRefBS::card_shift);
2160 li(R0, 0); // dirty
2161 if (UseConcMarkSweepGC) membar(Assembler::StoreStore);
2162 stbx(R0, Rtmp, Robj);
2163 }
2165 #ifndef SERIALGC
2167 // General G1 pre-barrier generator.
2168 // Goal: record the previous value if it is not null.
2169 void MacroAssembler::g1_write_barrier_pre(Register Robj, RegisterOrConstant offset, Register Rpre_val,
2170 Register Rtmp1, Register Rtmp2, bool needs_frame) {
2171 Label runtime, filtered;
2173 // Is marking active?
2174 if (in_bytes(PtrQueue::byte_width_of_active()) == 4) {
2175 lwz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()), R16_thread);
2176 } else {
2177 guarantee(in_bytes(PtrQueue::byte_width_of_active()) == 1, "Assumption");
2178 lbz(Rtmp1, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()), R16_thread);
2179 }
2180 cmpdi(CCR0, Rtmp1, 0);
2181 beq(CCR0, filtered);
2183 // Do we need to load the previous value?
2184 if (Robj != noreg) {
2185 // Load the previous value...
2186 if (UseCompressedOops) {
2187 lwz(Rpre_val, offset, Robj);
2188 } else {
2189 ld(Rpre_val, offset, Robj);
2190 }
2191 // Previous value has been loaded into Rpre_val.
2192 }
2193 assert(Rpre_val != noreg, "must have a real register");
2195 // Is the previous value null?
2196 cmpdi(CCR0, Rpre_val, 0);
2197 beq(CCR0, filtered);
2199 if (Robj != noreg && UseCompressedOops) {
2200 decode_heap_oop_not_null(Rpre_val);
2201 }
2203 // OK, it's not filtered, so we'll need to call enqueue. In the normal
2204 // case, pre_val will be a scratch G-reg, but there are some cases in
2205 // which it's an O-reg. In the first case, do a normal call. In the
2206 // latter, do a save here and call the frameless version.
2208 // Can we store original value in the thread's buffer?
2209 // Is index == 0?
2210 // (The index field is typed as size_t.)
2211 const Register Rbuffer = Rtmp1, Rindex = Rtmp2;
2213 ld(Rindex, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_index()), R16_thread);
2214 cmpdi(CCR0, Rindex, 0);
2215 beq(CCR0, runtime); // If index == 0, goto runtime.
2216 ld(Rbuffer, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_buf()), R16_thread);
2218 addi(Rindex, Rindex, -wordSize); // Decrement index.
2219 std(Rindex, in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_index()), R16_thread);
2221 // Record the previous value.
2222 stdx(Rpre_val, Rbuffer, Rindex);
2223 b(filtered);
2225 bind(runtime);
2227 // VM call need frame to access(write) O register.
2228 if (needs_frame) {
2229 save_LR_CR(Rtmp1);
2230 push_frame_abi112(0, Rtmp2);
2231 }
2233 if (Rpre_val->is_volatile() && Robj == noreg) mr(R31, Rpre_val); // Save pre_val across C call if it was preloaded.
2234 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), Rpre_val, R16_thread);
2235 if (Rpre_val->is_volatile() && Robj == noreg) mr(Rpre_val, R31); // restore
2237 if (needs_frame) {
2238 pop_frame();
2239 restore_LR_CR(Rtmp1);
2240 }
2242 bind(filtered);
2243 }
2245 // General G1 post-barrier generator
2246 // Store cross-region card.
2247 void MacroAssembler::g1_write_barrier_post(Register Rstore_addr, Register Rnew_val, Register Rtmp1, Register Rtmp2, Register Rtmp3, Label *filtered_ext) {
2248 Label runtime, filtered_int;
2249 Label& filtered = (filtered_ext != NULL) ? *filtered_ext : filtered_int;
2250 assert_different_registers(Rstore_addr, Rnew_val, Rtmp1, Rtmp2);
2252 G1SATBCardTableModRefBS* bs = (G1SATBCardTableModRefBS*) Universe::heap()->barrier_set();
2253 assert(bs->kind() == BarrierSet::G1SATBCT ||
2254 bs->kind() == BarrierSet::G1SATBCTLogging, "wrong barrier");
2256 // Does store cross heap regions?
2257 if (G1RSBarrierRegionFilter) {
2258 xorr(Rtmp1, Rstore_addr, Rnew_val);
2259 srdi_(Rtmp1, Rtmp1, HeapRegion::LogOfHRGrainBytes);
2260 beq(CCR0, filtered);
2261 }
2263 // Crosses regions, storing NULL?
2264 #ifdef ASSERT
2265 cmpdi(CCR0, Rnew_val, 0);
2266 asm_assert_ne("null oop not allowed (G1)", 0x322); // Checked by caller on PPC64, so following branch is obsolete:
2267 //beq(CCR0, filtered);
2268 #endif
2270 // Storing region crossing non-NULL, is card already dirty?
2271 assert(sizeof(*bs->byte_map_base) == sizeof(jbyte), "adjust this code");
2272 const Register Rcard_addr = Rtmp1;
2273 Register Rbase = Rtmp2;
2274 load_const_optimized(Rbase, (address)bs->byte_map_base, /*temp*/ Rtmp3);
2276 srdi(Rcard_addr, Rstore_addr, CardTableModRefBS::card_shift);
2278 // Get the address of the card.
2279 lbzx(/*card value*/ Rtmp3, Rbase, Rcard_addr);
2281 assert(CardTableModRefBS::dirty_card_val() == 0, "otherwise check this code");
2282 cmpwi(CCR0, Rtmp3 /* card value */, 0);
2283 beq(CCR0, filtered);
2285 // Storing a region crossing, non-NULL oop, card is clean.
2286 // Dirty card and log.
2287 li(Rtmp3, 0); // dirty
2288 //release(); // G1: oops are allowed to get visible after dirty marking.
2289 stbx(Rtmp3, Rbase, Rcard_addr);
2291 add(Rcard_addr, Rbase, Rcard_addr); // This is the address which needs to get enqueued.
2292 Rbase = noreg; // end of lifetime
2294 const Register Rqueue_index = Rtmp2,
2295 Rqueue_buf = Rtmp3;
2296 ld(Rqueue_index, in_bytes(JavaThread::dirty_card_queue_offset() + PtrQueue::byte_offset_of_index()), R16_thread);
2297 cmpdi(CCR0, Rqueue_index, 0);
2298 beq(CCR0, runtime); // index == 0 then jump to runtime
2299 ld(Rqueue_buf, in_bytes(JavaThread::dirty_card_queue_offset() + PtrQueue::byte_offset_of_buf()), R16_thread);
2301 addi(Rqueue_index, Rqueue_index, -wordSize); // decrement index
2302 std(Rqueue_index, in_bytes(JavaThread::dirty_card_queue_offset() + PtrQueue::byte_offset_of_index()), R16_thread);
2304 stdx(Rcard_addr, Rqueue_buf, Rqueue_index); // store card
2305 b(filtered);
2307 bind(runtime);
2309 // Save the live input values.
2310 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), Rcard_addr, R16_thread);
2312 bind(filtered_int);
2313 }
2314 #endif // SERIALGC
2316 // Values for last_Java_pc, and last_Java_sp must comply to the rules
2317 // in frame_ppc64.hpp.
2318 void MacroAssembler::set_last_Java_frame(Register last_Java_sp, Register last_Java_pc) {
2319 // Always set last_Java_pc and flags first because once last_Java_sp
2320 // is visible has_last_Java_frame is true and users will look at the
2321 // rest of the fields. (Note: flags should always be zero before we
2322 // get here so doesn't need to be set.)
2324 // Verify that last_Java_pc was zeroed on return to Java
2325 asm_assert_mem8_is_zero(in_bytes(JavaThread::last_Java_pc_offset()), R16_thread,
2326 "last_Java_pc not zeroed before leaving Java", 0x200);
2328 // When returning from calling out from Java mode the frame anchor's
2329 // last_Java_pc will always be set to NULL. It is set here so that
2330 // if we are doing a call to native (not VM) that we capture the
2331 // known pc and don't have to rely on the native call having a
2332 // standard frame linkage where we can find the pc.
2333 if (last_Java_pc != noreg)
2334 std(last_Java_pc, in_bytes(JavaThread::last_Java_pc_offset()), R16_thread);
2336 // Set last_Java_sp last.
2337 std(last_Java_sp, in_bytes(JavaThread::last_Java_sp_offset()), R16_thread);
2338 }
2340 void MacroAssembler::reset_last_Java_frame(void) {
2341 asm_assert_mem8_isnot_zero(in_bytes(JavaThread::last_Java_sp_offset()),
2342 R16_thread, "SP was not set, still zero", 0x202);
2344 BLOCK_COMMENT("reset_last_Java_frame {");
2345 li(R0, 0);
2347 // _last_Java_sp = 0
2348 std(R0, in_bytes(JavaThread::last_Java_sp_offset()), R16_thread);
2350 // _last_Java_pc = 0
2351 std(R0, in_bytes(JavaThread::last_Java_pc_offset()), R16_thread);
2352 BLOCK_COMMENT("} reset_last_Java_frame");
2353 }
2355 void MacroAssembler::set_top_ijava_frame_at_SP_as_last_Java_frame(Register sp, Register tmp1) {
2356 assert_different_registers(sp, tmp1);
2358 // sp points to a TOP_IJAVA_FRAME, retrieve frame's PC via
2359 // TOP_IJAVA_FRAME_ABI.
2360 // FIXME: assert that we really have a TOP_IJAVA_FRAME here!
2361 #ifdef CC_INTERP
2362 ld(tmp1/*pc*/, _top_ijava_frame_abi(frame_manager_lr), sp);
2363 #else
2364 Unimplemented();
2365 #endif
2367 set_last_Java_frame(/*sp=*/sp, /*pc=*/tmp1);
2368 }
2370 void MacroAssembler::get_vm_result(Register oop_result) {
2371 // Read:
2372 // R16_thread
2373 // R16_thread->in_bytes(JavaThread::vm_result_offset())
2374 //
2375 // Updated:
2376 // oop_result
2377 // R16_thread->in_bytes(JavaThread::vm_result_offset())
2379 ld(oop_result, in_bytes(JavaThread::vm_result_offset()), R16_thread);
2380 li(R0, 0);
2381 std(R0, in_bytes(JavaThread::vm_result_offset()), R16_thread);
2383 verify_oop(oop_result);
2384 }
2386 void MacroAssembler::get_vm_result_2(Register metadata_result) {
2387 // Read:
2388 // R16_thread
2389 // R16_thread->in_bytes(JavaThread::vm_result_2_offset())
2390 //
2391 // Updated:
2392 // metadata_result
2393 // R16_thread->in_bytes(JavaThread::vm_result_2_offset())
2395 ld(metadata_result, in_bytes(JavaThread::vm_result_2_offset()), R16_thread);
2396 li(R0, 0);
2397 std(R0, in_bytes(JavaThread::vm_result_2_offset()), R16_thread);
2398 }
2401 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
2402 Register current = (src != noreg) ? src : dst; // Klass is in dst if no src provided.
2403 if (Universe::narrow_klass_base() != 0) {
2404 load_const(R0, Universe::narrow_klass_base(), (dst != current) ? dst : noreg); // Use dst as temp if it is free.
2405 sub(dst, current, R0);
2406 current = dst;
2407 }
2408 if (Universe::narrow_klass_shift() != 0) {
2409 srdi(dst, current, Universe::narrow_klass_shift());
2410 current = dst;
2411 }
2412 mr_if_needed(dst, current); // Move may be required.
2413 }
2415 void MacroAssembler::store_klass(Register dst_oop, Register klass, Register ck) {
2416 if (UseCompressedClassPointers) {
2417 encode_klass_not_null(ck, klass);
2418 stw(ck, oopDesc::klass_offset_in_bytes(), dst_oop);
2419 } else {
2420 std(klass, oopDesc::klass_offset_in_bytes(), dst_oop);
2421 }
2422 }
2424 int MacroAssembler::instr_size_for_decode_klass_not_null() {
2425 if (!UseCompressedClassPointers) return 0;
2426 int num_instrs = 1; // shift or move
2427 if (Universe::narrow_klass_base() != 0) num_instrs = 7; // shift + load const + add
2428 return num_instrs * BytesPerInstWord;
2429 }
2431 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
2432 if (src == noreg) src = dst;
2433 Register shifted_src = src;
2434 if (Universe::narrow_klass_shift() != 0 ||
2435 Universe::narrow_klass_base() == 0 && src != dst) { // Move required.
2436 shifted_src = dst;
2437 sldi(shifted_src, src, Universe::narrow_klass_shift());
2438 }
2439 if (Universe::narrow_klass_base() != 0) {
2440 load_const(R0, Universe::narrow_klass_base());
2441 add(dst, shifted_src, R0);
2442 }
2443 }
2445 void MacroAssembler::load_klass(Register dst, Register src) {
2446 if (UseCompressedClassPointers) {
2447 lwz(dst, oopDesc::klass_offset_in_bytes(), src);
2448 // Attention: no null check here!
2449 decode_klass_not_null(dst, dst);
2450 } else {
2451 ld(dst, oopDesc::klass_offset_in_bytes(), src);
2452 }
2453 }
2455 void MacroAssembler::load_klass_with_trap_null_check(Register dst, Register src) {
2456 if (!os::zero_page_read_protected()) {
2457 if (TrapBasedNullChecks) {
2458 trap_null_check(src);
2459 }
2460 }
2461 load_klass(dst, src);
2462 }
2464 void MacroAssembler::reinit_heapbase(Register d, Register tmp) {
2465 if (Universe::heap() != NULL) {
2466 if (Universe::narrow_oop_base() == NULL) {
2467 Assembler::xorr(R30, R30, R30);
2468 } else {
2469 load_const(R30, Universe::narrow_ptrs_base(), tmp);
2470 }
2471 } else {
2472 load_const(R30, Universe::narrow_ptrs_base_addr(), tmp);
2473 ld(R30, 0, R30);
2474 }
2475 }
2477 // Clear Array
2478 // Kills both input registers. tmp == R0 is allowed.
2479 void MacroAssembler::clear_memory_doubleword(Register base_ptr, Register cnt_dwords, Register tmp) {
2480 // Procedure for large arrays (uses data cache block zero instruction).
2481 Label startloop, fast, fastloop, small_rest, restloop, done;
2482 const int cl_size = VM_Version::get_cache_line_size(),
2483 cl_dwords = cl_size>>3,
2484 cl_dw_addr_bits = exact_log2(cl_dwords),
2485 dcbz_min = 1; // Min count of dcbz executions, needs to be >0.
2487 //2:
2488 cmpdi(CCR1, cnt_dwords, ((dcbz_min+1)<<cl_dw_addr_bits)-1); // Big enough? (ensure >=dcbz_min lines included).
2489 blt(CCR1, small_rest); // Too small.
2490 rldicl_(tmp, base_ptr, 64-3, 64-cl_dw_addr_bits); // Extract dword offset within first cache line.
2491 beq(CCR0, fast); // Already 128byte aligned.
2493 subfic(tmp, tmp, cl_dwords);
2494 mtctr(tmp); // Set ctr to hit 128byte boundary (0<ctr<cl_dwords).
2495 subf(cnt_dwords, tmp, cnt_dwords); // rest.
2496 li(tmp, 0);
2497 //10:
2498 bind(startloop); // Clear at the beginning to reach 128byte boundary.
2499 std(tmp, 0, base_ptr); // Clear 8byte aligned block.
2500 addi(base_ptr, base_ptr, 8);
2501 bdnz(startloop);
2502 //13:
2503 bind(fast); // Clear 128byte blocks.
2504 srdi(tmp, cnt_dwords, cl_dw_addr_bits); // Loop count for 128byte loop (>0).
2505 andi(cnt_dwords, cnt_dwords, cl_dwords-1); // Rest in dwords.
2506 mtctr(tmp); // Load counter.
2507 //16:
2508 bind(fastloop);
2509 dcbz(base_ptr); // Clear 128byte aligned block.
2510 addi(base_ptr, base_ptr, cl_size);
2511 bdnz(fastloop);
2512 if (InsertEndGroupPPC64) { endgroup(); } else { nop(); }
2513 //20:
2514 bind(small_rest);
2515 cmpdi(CCR0, cnt_dwords, 0); // size 0?
2516 beq(CCR0, done); // rest == 0
2517 li(tmp, 0);
2518 mtctr(cnt_dwords); // Load counter.
2519 //24:
2520 bind(restloop); // Clear rest.
2521 std(tmp, 0, base_ptr); // Clear 8byte aligned block.
2522 addi(base_ptr, base_ptr, 8);
2523 bdnz(restloop);
2524 //27:
2525 bind(done);
2526 }
2528 /////////////////////////////////////////// String intrinsics ////////////////////////////////////////////
2530 // Search for a single jchar in an jchar[].
2531 //
2532 // Assumes that result differs from all other registers.
2533 //
2534 // Haystack, needle are the addresses of jchar-arrays.
2535 // NeedleChar is needle[0] if it is known at compile time.
2536 // Haycnt is the length of the haystack. We assume haycnt >=1.
2537 //
2538 // Preserves haystack, haycnt, kills all other registers.
2539 //
2540 // If needle == R0, we search for the constant needleChar.
2541 void MacroAssembler::string_indexof_1(Register result, Register haystack, Register haycnt,
2542 Register needle, jchar needleChar,
2543 Register tmp1, Register tmp2) {
2545 assert_different_registers(result, haystack, haycnt, needle, tmp1, tmp2);
2547 Label L_InnerLoop, L_FinalCheck, L_Found1, L_Found2, L_Found3, L_NotFound, L_End;
2548 Register needle0 = needle, // Contains needle[0].
2549 addr = tmp1,
2550 ch1 = tmp2,
2551 ch2 = R0;
2553 //2 (variable) or 3 (const):
2554 if (needle != R0) lhz(needle0, 0, needle); // Preload needle character, needle has len==1.
2555 dcbtct(haystack, 0x00); // Indicate R/O access to haystack.
2557 srwi_(tmp2, haycnt, 1); // Shift right by exact_log2(UNROLL_FACTOR).
2558 mr(addr, haystack);
2559 beq(CCR0, L_FinalCheck);
2560 mtctr(tmp2); // Move to count register.
2561 //8:
2562 bind(L_InnerLoop); // Main work horse (2x unrolled search loop).
2563 lhz(ch1, 0, addr); // Load characters from haystack.
2564 lhz(ch2, 2, addr);
2565 (needle != R0) ? cmpw(CCR0, ch1, needle0) : cmplwi(CCR0, ch1, needleChar);
2566 (needle != R0) ? cmpw(CCR1, ch2, needle0) : cmplwi(CCR1, ch2, needleChar);
2567 beq(CCR0, L_Found1); // Did we find the needle?
2568 beq(CCR1, L_Found2);
2569 addi(addr, addr, 4);
2570 bdnz(L_InnerLoop);
2571 //16:
2572 bind(L_FinalCheck);
2573 andi_(R0, haycnt, 1);
2574 beq(CCR0, L_NotFound);
2575 lhz(ch1, 0, addr); // One position left at which we have to compare.
2576 (needle != R0) ? cmpw(CCR1, ch1, needle0) : cmplwi(CCR1, ch1, needleChar);
2577 beq(CCR1, L_Found3);
2578 //21:
2579 bind(L_NotFound);
2580 li(result, -1); // Not found.
2581 b(L_End);
2583 bind(L_Found2);
2584 addi(addr, addr, 2);
2585 //24:
2586 bind(L_Found1);
2587 bind(L_Found3); // Return index ...
2588 subf(addr, haystack, addr); // relative to haystack,
2589 srdi(result, addr, 1); // in characters.
2590 bind(L_End);
2591 }
2594 // Implementation of IndexOf for jchar arrays.
2595 //
2596 // The length of haystack and needle are not constant, i.e. passed in a register.
2597 //
2598 // Preserves registers haystack, needle.
2599 // Kills registers haycnt, needlecnt.
2600 // Assumes that result differs from all other registers.
2601 // Haystack, needle are the addresses of jchar-arrays.
2602 // Haycnt, needlecnt are the lengths of them, respectively.
2603 //
2604 // Needlecntval must be zero or 15-bit unsigned immediate and > 1.
2605 void MacroAssembler::string_indexof(Register result, Register haystack, Register haycnt,
2606 Register needle, ciTypeArray* needle_values, Register needlecnt, int needlecntval,
2607 Register tmp1, Register tmp2, Register tmp3, Register tmp4) {
2609 // Ensure 0<needlecnt<=haycnt in ideal graph as prerequisite!
2610 Label L_TooShort, L_Found, L_NotFound, L_End;
2611 Register last_addr = haycnt, // Kill haycnt at the beginning.
2612 addr = tmp1,
2613 n_start = tmp2,
2614 ch1 = tmp3,
2615 ch2 = R0;
2617 // **************************************************************************************************
2618 // Prepare for main loop: optimized for needle count >=2, bail out otherwise.
2619 // **************************************************************************************************
2621 //1 (variable) or 3 (const):
2622 dcbtct(needle, 0x00); // Indicate R/O access to str1.
2623 dcbtct(haystack, 0x00); // Indicate R/O access to str2.
2625 // Compute last haystack addr to use if no match gets found.
2626 if (needlecntval == 0) { // variable needlecnt
2627 //3:
2628 subf(ch1, needlecnt, haycnt); // Last character index to compare is haycnt-needlecnt.
2629 addi(addr, haystack, -2); // Accesses use pre-increment.
2630 cmpwi(CCR6, needlecnt, 2);
2631 blt(CCR6, L_TooShort); // Variable needlecnt: handle short needle separately.
2632 slwi(ch1, ch1, 1); // Scale to number of bytes.
2633 lwz(n_start, 0, needle); // Load first 2 characters of needle.
2634 add(last_addr, haystack, ch1); // Point to last address to compare (haystack+2*(haycnt-needlecnt)).
2635 addi(needlecnt, needlecnt, -2); // Rest of needle.
2636 } else { // constant needlecnt
2637 guarantee(needlecntval != 1, "IndexOf with single-character needle must be handled separately");
2638 assert((needlecntval & 0x7fff) == needlecntval, "wrong immediate");
2639 //5:
2640 addi(ch1, haycnt, -needlecntval); // Last character index to compare is haycnt-needlecnt.
2641 lwz(n_start, 0, needle); // Load first 2 characters of needle.
2642 addi(addr, haystack, -2); // Accesses use pre-increment.
2643 slwi(ch1, ch1, 1); // Scale to number of bytes.
2644 add(last_addr, haystack, ch1); // Point to last address to compare (haystack+2*(haycnt-needlecnt)).
2645 li(needlecnt, needlecntval-2); // Rest of needle.
2646 }
2648 // Main Loop (now we have at least 3 characters).
2649 //11:
2650 Label L_OuterLoop, L_InnerLoop, L_FinalCheck, L_Comp1, L_Comp2, L_Comp3;
2651 bind(L_OuterLoop); // Search for 1st 2 characters.
2652 Register addr_diff = tmp4;
2653 subf(addr_diff, addr, last_addr); // Difference between already checked address and last address to check.
2654 addi(addr, addr, 2); // This is the new address we want to use for comparing.
2655 srdi_(ch2, addr_diff, 2);
2656 beq(CCR0, L_FinalCheck); // 2 characters left?
2657 mtctr(ch2); // addr_diff/4
2658 //16:
2659 bind(L_InnerLoop); // Main work horse (2x unrolled search loop)
2660 lwz(ch1, 0, addr); // Load 2 characters of haystack (ignore alignment).
2661 lwz(ch2, 2, addr);
2662 cmpw(CCR0, ch1, n_start); // Compare 2 characters (1 would be sufficient but try to reduce branches to CompLoop).
2663 cmpw(CCR1, ch2, n_start);
2664 beq(CCR0, L_Comp1); // Did we find the needle start?
2665 beq(CCR1, L_Comp2);
2666 addi(addr, addr, 4);
2667 bdnz(L_InnerLoop);
2668 //24:
2669 bind(L_FinalCheck);
2670 rldicl_(addr_diff, addr_diff, 64-1, 63); // Remaining characters not covered by InnerLoop: (addr_diff>>1)&1.
2671 beq(CCR0, L_NotFound);
2672 lwz(ch1, 0, addr); // One position left at which we have to compare.
2673 cmpw(CCR1, ch1, n_start);
2674 beq(CCR1, L_Comp3);
2675 //29:
2676 bind(L_NotFound);
2677 li(result, -1); // not found
2678 b(L_End);
2681 // **************************************************************************************************
2682 // Special Case: unfortunately, the variable needle case can be called with needlecnt<2
2683 // **************************************************************************************************
2684 //31:
2685 if ((needlecntval>>1) !=1 ) { // Const needlecnt is 2 or 3? Reduce code size.
2686 int nopcnt = 5;
2687 if (needlecntval !=0 ) ++nopcnt; // Balance alignment (other case: see below).
2688 if (needlecntval == 0) { // We have to handle these cases separately.
2689 Label L_OneCharLoop;
2690 bind(L_TooShort);
2691 mtctr(haycnt);
2692 lhz(n_start, 0, needle); // First character of needle
2693 bind(L_OneCharLoop);
2694 lhzu(ch1, 2, addr);
2695 cmpw(CCR1, ch1, n_start);
2696 beq(CCR1, L_Found); // Did we find the one character needle?
2697 bdnz(L_OneCharLoop);
2698 li(result, -1); // Not found.
2699 b(L_End);
2700 } // 8 instructions, so no impact on alignment.
2701 for (int x = 0; x < nopcnt; ++x) nop();
2702 }
2704 // **************************************************************************************************
2705 // Regular Case Part II: compare rest of needle (first 2 characters have been compared already)
2706 // **************************************************************************************************
2708 // Compare the rest
2709 //36 if needlecntval==0, else 37:
2710 bind(L_Comp2);
2711 addi(addr, addr, 2); // First comparison has failed, 2nd one hit.
2712 bind(L_Comp1); // Addr points to possible needle start.
2713 bind(L_Comp3); // Could have created a copy and use a different return address but saving code size here.
2714 if (needlecntval != 2) { // Const needlecnt==2?
2715 if (needlecntval != 3) {
2716 if (needlecntval == 0) beq(CCR6, L_Found); // Variable needlecnt==2?
2717 Register ind_reg = tmp4;
2718 li(ind_reg, 2*2); // First 2 characters are already compared, use index 2.
2719 mtctr(needlecnt); // Decremented by 2, still > 0.
2720 //40:
2721 Label L_CompLoop;
2722 bind(L_CompLoop);
2723 lhzx(ch2, needle, ind_reg);
2724 lhzx(ch1, addr, ind_reg);
2725 cmpw(CCR1, ch1, ch2);
2726 bne(CCR1, L_OuterLoop);
2727 addi(ind_reg, ind_reg, 2);
2728 bdnz(L_CompLoop);
2729 } else { // No loop required if there's only one needle character left.
2730 lhz(ch2, 2*2, needle);
2731 lhz(ch1, 2*2, addr);
2732 cmpw(CCR1, ch1, ch2);
2733 bne(CCR1, L_OuterLoop);
2734 }
2735 }
2736 // Return index ...
2737 //46:
2738 bind(L_Found);
2739 subf(addr, haystack, addr); // relative to haystack, ...
2740 srdi(result, addr, 1); // in characters.
2741 //48:
2742 bind(L_End);
2743 }
2745 // Implementation of Compare for jchar arrays.
2746 //
2747 // Kills the registers str1, str2, cnt1, cnt2.
2748 // Kills cr0, ctr.
2749 // Assumes that result differes from the input registers.
2750 void MacroAssembler::string_compare(Register str1_reg, Register str2_reg, Register cnt1_reg, Register cnt2_reg,
2751 Register result_reg, Register tmp_reg) {
2752 assert_different_registers(result_reg, str1_reg, str2_reg, cnt1_reg, cnt2_reg, tmp_reg);
2754 Label Ldone, Lslow_case, Lslow_loop, Lfast_loop;
2755 Register cnt_diff = R0,
2756 limit_reg = cnt1_reg,
2757 chr1_reg = result_reg,
2758 chr2_reg = cnt2_reg,
2759 addr_diff = str2_reg;
2761 // Offset 0 should be 32 byte aligned.
2762 //-4:
2763 dcbtct(str1_reg, 0x00); // Indicate R/O access to str1.
2764 dcbtct(str2_reg, 0x00); // Indicate R/O access to str2.
2765 //-2:
2766 // Compute min(cnt1, cnt2) and check if 0 (bail out if we don't need to compare characters).
2767 subf(result_reg, cnt2_reg, cnt1_reg); // difference between cnt1/2
2768 subf_(addr_diff, str1_reg, str2_reg); // alias?
2769 beq(CCR0, Ldone); // return cnt difference if both ones are identical
2770 srawi(limit_reg, result_reg, 31); // generate signmask (cnt1/2 must be non-negative so cnt_diff can't overflow)
2771 mr(cnt_diff, result_reg);
2772 andr(limit_reg, result_reg, limit_reg); // difference or zero (negative): cnt1<cnt2 ? cnt1-cnt2 : 0
2773 add_(limit_reg, cnt2_reg, limit_reg); // min(cnt1, cnt2)==0?
2774 beq(CCR0, Ldone); // return cnt difference if one has 0 length
2776 lhz(chr1_reg, 0, str1_reg); // optional: early out if first characters mismatch
2777 lhzx(chr2_reg, str1_reg, addr_diff); // optional: early out if first characters mismatch
2778 addi(tmp_reg, limit_reg, -1); // min(cnt1, cnt2)-1
2779 subf_(result_reg, chr2_reg, chr1_reg); // optional: early out if first characters mismatch
2780 bne(CCR0, Ldone); // optional: early out if first characters mismatch
2782 // Set loop counter by scaling down tmp_reg
2783 srawi_(chr2_reg, tmp_reg, exact_log2(4)); // (min(cnt1, cnt2)-1)/4
2784 ble(CCR0, Lslow_case); // need >4 characters for fast loop
2785 andi(limit_reg, tmp_reg, 4-1); // remaining characters
2787 // Adapt str1_reg str2_reg for the first loop iteration
2788 mtctr(chr2_reg); // (min(cnt1, cnt2)-1)/4
2789 addi(limit_reg, limit_reg, 4+1); // compare last 5-8 characters in slow_case if mismatch found in fast_loop
2790 //16:
2791 // Compare the rest of the characters
2792 bind(Lfast_loop);
2793 ld(chr1_reg, 0, str1_reg);
2794 ldx(chr2_reg, str1_reg, addr_diff);
2795 cmpd(CCR0, chr2_reg, chr1_reg);
2796 bne(CCR0, Lslow_case); // return chr1_reg
2797 addi(str1_reg, str1_reg, 4*2);
2798 bdnz(Lfast_loop);
2799 addi(limit_reg, limit_reg, -4); // no mismatch found in fast_loop, only 1-4 characters missing
2800 //23:
2801 bind(Lslow_case);
2802 mtctr(limit_reg);
2803 //24:
2804 bind(Lslow_loop);
2805 lhz(chr1_reg, 0, str1_reg);
2806 lhzx(chr2_reg, str1_reg, addr_diff);
2807 subf_(result_reg, chr2_reg, chr1_reg);
2808 bne(CCR0, Ldone); // return chr1_reg
2809 addi(str1_reg, str1_reg, 1*2);
2810 bdnz(Lslow_loop);
2811 //30:
2812 // If strings are equal up to min length, return the length difference.
2813 mr(result_reg, cnt_diff);
2814 nop(); // alignment
2815 //32:
2816 // Otherwise, return the difference between the first mismatched chars.
2817 bind(Ldone);
2818 }
2821 // Compare char[] arrays.
2822 //
2823 // str1_reg USE only
2824 // str2_reg USE only
2825 // cnt_reg USE_DEF, due to tmp reg shortage
2826 // result_reg DEF only, might compromise USE only registers
2827 void MacroAssembler::char_arrays_equals(Register str1_reg, Register str2_reg, Register cnt_reg, Register result_reg,
2828 Register tmp1_reg, Register tmp2_reg, Register tmp3_reg, Register tmp4_reg,
2829 Register tmp5_reg) {
2831 // Str1 may be the same register as str2 which can occur e.g. after scalar replacement.
2832 assert_different_registers(result_reg, str1_reg, cnt_reg, tmp1_reg, tmp2_reg, tmp3_reg, tmp4_reg, tmp5_reg);
2833 assert_different_registers(result_reg, str2_reg, cnt_reg, tmp1_reg, tmp2_reg, tmp3_reg, tmp4_reg, tmp5_reg);
2835 // Offset 0 should be 32 byte aligned.
2836 Label Linit_cbc, Lcbc, Lloop, Ldone_true, Ldone_false;
2837 Register index_reg = tmp5_reg;
2838 Register cbc_iter = tmp4_reg;
2840 //-1:
2841 dcbtct(str1_reg, 0x00); // Indicate R/O access to str1.
2842 dcbtct(str2_reg, 0x00); // Indicate R/O access to str2.
2843 //1:
2844 andi(cbc_iter, cnt_reg, 4-1); // Remaining iterations after 4 java characters per iteration loop.
2845 li(index_reg, 0); // init
2846 li(result_reg, 0); // assume false
2847 srwi_(tmp2_reg, cnt_reg, exact_log2(4)); // Div: 4 java characters per iteration (main loop).
2849 cmpwi(CCR1, cbc_iter, 0); // CCR1 = (cbc_iter==0)
2850 beq(CCR0, Linit_cbc); // too short
2851 mtctr(tmp2_reg);
2852 //8:
2853 bind(Lloop);
2854 ldx(tmp1_reg, str1_reg, index_reg);
2855 ldx(tmp2_reg, str2_reg, index_reg);
2856 cmpd(CCR0, tmp1_reg, tmp2_reg);
2857 bne(CCR0, Ldone_false); // Unequal char pair found -> done.
2858 addi(index_reg, index_reg, 4*sizeof(jchar));
2859 bdnz(Lloop);
2860 //14:
2861 bind(Linit_cbc);
2862 beq(CCR1, Ldone_true);
2863 mtctr(cbc_iter);
2864 //16:
2865 bind(Lcbc);
2866 lhzx(tmp1_reg, str1_reg, index_reg);
2867 lhzx(tmp2_reg, str2_reg, index_reg);
2868 cmpw(CCR0, tmp1_reg, tmp2_reg);
2869 bne(CCR0, Ldone_false); // Unequal char pair found -> done.
2870 addi(index_reg, index_reg, 1*sizeof(jchar));
2871 bdnz(Lcbc);
2872 nop();
2873 bind(Ldone_true);
2874 li(result_reg, 1);
2875 //24:
2876 bind(Ldone_false);
2877 }
2880 void MacroAssembler::char_arrays_equalsImm(Register str1_reg, Register str2_reg, int cntval, Register result_reg,
2881 Register tmp1_reg, Register tmp2_reg) {
2882 // Str1 may be the same register as str2 which can occur e.g. after scalar replacement.
2883 assert_different_registers(result_reg, str1_reg, tmp1_reg, tmp2_reg);
2884 assert_different_registers(result_reg, str2_reg, tmp1_reg, tmp2_reg);
2885 assert(sizeof(jchar) == 2, "must be");
2886 assert(cntval >= 0 && ((cntval & 0x7fff) == cntval), "wrong immediate");
2888 Label Ldone_false;
2890 if (cntval < 16) { // short case
2891 if (cntval != 0) li(result_reg, 0); // assume false
2893 const int num_bytes = cntval*sizeof(jchar);
2894 int index = 0;
2895 for (int next_index; (next_index = index + 8) <= num_bytes; index = next_index) {
2896 ld(tmp1_reg, index, str1_reg);
2897 ld(tmp2_reg, index, str2_reg);
2898 cmpd(CCR0, tmp1_reg, tmp2_reg);
2899 bne(CCR0, Ldone_false);
2900 }
2901 if (cntval & 2) {
2902 lwz(tmp1_reg, index, str1_reg);
2903 lwz(tmp2_reg, index, str2_reg);
2904 cmpw(CCR0, tmp1_reg, tmp2_reg);
2905 bne(CCR0, Ldone_false);
2906 index += 4;
2907 }
2908 if (cntval & 1) {
2909 lhz(tmp1_reg, index, str1_reg);
2910 lhz(tmp2_reg, index, str2_reg);
2911 cmpw(CCR0, tmp1_reg, tmp2_reg);
2912 bne(CCR0, Ldone_false);
2913 }
2914 // fallthrough: true
2915 } else {
2916 Label Lloop;
2917 Register index_reg = tmp1_reg;
2918 const int loopcnt = cntval/4;
2919 assert(loopcnt > 0, "must be");
2920 // Offset 0 should be 32 byte aligned.
2921 //2:
2922 dcbtct(str1_reg, 0x00); // Indicate R/O access to str1.
2923 dcbtct(str2_reg, 0x00); // Indicate R/O access to str2.
2924 li(tmp2_reg, loopcnt);
2925 li(index_reg, 0); // init
2926 li(result_reg, 0); // assume false
2927 mtctr(tmp2_reg);
2928 //8:
2929 bind(Lloop);
2930 ldx(R0, str1_reg, index_reg);
2931 ldx(tmp2_reg, str2_reg, index_reg);
2932 cmpd(CCR0, R0, tmp2_reg);
2933 bne(CCR0, Ldone_false); // Unequal char pair found -> done.
2934 addi(index_reg, index_reg, 4*sizeof(jchar));
2935 bdnz(Lloop);
2936 //14:
2937 if (cntval & 2) {
2938 lwzx(R0, str1_reg, index_reg);
2939 lwzx(tmp2_reg, str2_reg, index_reg);
2940 cmpw(CCR0, R0, tmp2_reg);
2941 bne(CCR0, Ldone_false);
2942 if (cntval & 1) addi(index_reg, index_reg, 2*sizeof(jchar));
2943 }
2944 if (cntval & 1) {
2945 lhzx(R0, str1_reg, index_reg);
2946 lhzx(tmp2_reg, str2_reg, index_reg);
2947 cmpw(CCR0, R0, tmp2_reg);
2948 bne(CCR0, Ldone_false);
2949 }
2950 // fallthru: true
2951 }
2952 li(result_reg, 1);
2953 bind(Ldone_false);
2954 }
2957 void MacroAssembler::asm_assert(bool check_equal, const char *msg, int id) {
2958 #ifdef ASSERT
2959 Label ok;
2960 if (check_equal) {
2961 beq(CCR0, ok);
2962 } else {
2963 bne(CCR0, ok);
2964 }
2965 stop(msg, id);
2966 bind(ok);
2967 #endif
2968 }
2970 void MacroAssembler::asm_assert_mems_zero(bool check_equal, int size, int mem_offset,
2971 Register mem_base, const char* msg, int id) {
2972 #ifdef ASSERT
2973 switch (size) {
2974 case 4:
2975 lwz(R0, mem_offset, mem_base);
2976 cmpwi(CCR0, R0, 0);
2977 break;
2978 case 8:
2979 ld(R0, mem_offset, mem_base);
2980 cmpdi(CCR0, R0, 0);
2981 break;
2982 default:
2983 ShouldNotReachHere();
2984 }
2985 asm_assert(check_equal, msg, id);
2986 #endif // ASSERT
2987 }
2989 void MacroAssembler::verify_thread() {
2990 if (VerifyThread) {
2991 unimplemented("'VerifyThread' currently not implemented on PPC");
2992 }
2993 }
2995 // READ: oop. KILL: R0. Volatile floats perhaps.
2996 void MacroAssembler::verify_oop(Register oop, const char* msg) {
2997 if (!VerifyOops) {
2998 return;
2999 }
3000 // Will be preserved.
3001 Register tmp = R11;
3002 assert(oop != tmp, "precondition");
3003 unsigned int nbytes_save = 10*8; // 10 volatile gprs
3004 address/* FunctionDescriptor** */fd = StubRoutines::verify_oop_subroutine_entry_address();
3005 // save tmp
3006 mr(R0, tmp);
3007 // kill tmp
3008 save_LR_CR(tmp);
3009 push_frame_abi112(nbytes_save, tmp);
3010 // restore tmp
3011 mr(tmp, R0);
3012 save_volatile_gprs(R1_SP, 112); // except R0
3013 // load FunctionDescriptor**
3014 load_const(tmp, fd);
3015 // load FunctionDescriptor*
3016 ld(tmp, 0, tmp);
3017 mr(R4_ARG2, oop);
3018 load_const(R3_ARG1, (address)msg);
3019 // call destination for its side effect
3020 call_c(tmp);
3021 restore_volatile_gprs(R1_SP, 112); // except R0
3022 pop_frame();
3023 // save tmp
3024 mr(R0, tmp);
3025 // kill tmp
3026 restore_LR_CR(tmp);
3027 // restore tmp
3028 mr(tmp, R0);
3029 }
3031 const char* stop_types[] = {
3032 "stop",
3033 "untested",
3034 "unimplemented",
3035 "shouldnotreachhere"
3036 };
3038 static void stop_on_request(int tp, const char* msg) {
3039 tty->print("PPC assembly code requires stop: (%s) %s\n", (void *)stop_types[tp%/*stop_end*/4], msg);
3040 guarantee(false, err_msg("PPC assembly code requires stop: %s", msg));
3041 }
3043 // Call a C-function that prints output.
3044 void MacroAssembler::stop(int type, const char* msg, int id) {
3045 #ifndef PRODUCT
3046 block_comment(err_msg("stop: %s %s {", stop_types[type%stop_end], msg));
3047 #else
3048 block_comment("stop {");
3049 #endif
3051 // setup arguments
3052 load_const_optimized(R3_ARG1, type);
3053 load_const_optimized(R4_ARG2, (void *)msg, /*tmp=*/R0);
3054 call_VM_leaf(CAST_FROM_FN_PTR(address, stop_on_request), R3_ARG1, R4_ARG2);
3055 illtrap();
3056 emit_int32(id);
3057 block_comment("} stop;");
3058 }
3060 #ifndef PRODUCT
3061 // Write pattern 0x0101010101010101 in memory region [low-before, high+after].
3062 // Val, addr are temp registers.
3063 // If low == addr, addr is killed.
3064 // High is preserved.
3065 void MacroAssembler::zap_from_to(Register low, int before, Register high, int after, Register val, Register addr) {
3066 if (!ZapMemory) return;
3068 assert_different_registers(low, val);
3070 BLOCK_COMMENT("zap memory region {");
3071 load_const_optimized(val, 0x0101010101010101);
3072 int size = before + after;
3073 if (low == high && size < 5 && size > 0) {
3074 int offset = -before*BytesPerWord;
3075 for (int i = 0; i < size; ++i) {
3076 std(val, offset, low);
3077 offset += (1*BytesPerWord);
3078 }
3079 } else {
3080 addi(addr, low, -before*BytesPerWord);
3081 assert_different_registers(high, val);
3082 if (after) addi(high, high, after * BytesPerWord);
3083 Label loop;
3084 bind(loop);
3085 std(val, 0, addr);
3086 addi(addr, addr, 8);
3087 cmpd(CCR6, addr, high);
3088 ble(CCR6, loop);
3089 if (after) addi(high, high, -after * BytesPerWord); // Correct back to old value.
3090 }
3091 BLOCK_COMMENT("} zap memory region");
3092 }
3094 #endif // !PRODUCT