Mercurial > hg > truffle
annotate src/cpu/x86/vm/macroAssembler_x86.cpp @ 13310:733cccc125ed
added subqWide to AMD64Assembler
renamed HotSpotHostBackend.stackShadowPages to pagesToBang
author | Doug Simon <doug.simon@oracle.com> |
---|---|
date | Fri, 13 Dec 2013 13:12:58 +0100 |
parents | 096c224171c4 |
children | 02f27ecb4f3a |
rev | line source |
---|---|
7199 | 1 /* |
11080
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved. |
7199 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA | |
20 * or visit www.oracle.com if you need additional information or have any | |
21 * questions. | |
22 * | |
23 */ | |
24 | |
25 #include "precompiled.hpp" | |
26 #include "asm/assembler.hpp" | |
27 #include "asm/assembler.inline.hpp" | |
28 #include "compiler/disassembler.hpp" | |
29 #include "gc_interface/collectedHeap.inline.hpp" | |
30 #include "interpreter/interpreter.hpp" | |
31 #include "memory/cardTableModRefBS.hpp" | |
32 #include "memory/resourceArea.hpp" | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
33 #include "memory/universe.hpp" |
7199 | 34 #include "prims/methodHandles.hpp" |
35 #include "runtime/biasedLocking.hpp" | |
36 #include "runtime/interfaceSupport.hpp" | |
37 #include "runtime/objectMonitor.hpp" | |
38 #include "runtime/os.hpp" | |
39 #include "runtime/sharedRuntime.hpp" | |
40 #include "runtime/stubRoutines.hpp" | |
8001
db9981fd3124
8005915: Unify SERIALGC and INCLUDE_ALTERNATE_GCS
jprovino
parents:
7477
diff
changeset
|
41 #include "utilities/macros.hpp" |
db9981fd3124
8005915: Unify SERIALGC and INCLUDE_ALTERNATE_GCS
jprovino
parents:
7477
diff
changeset
|
42 #if INCLUDE_ALL_GCS |
7199 | 43 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" |
44 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" | |
45 #include "gc_implementation/g1/heapRegion.hpp" | |
8001
db9981fd3124
8005915: Unify SERIALGC and INCLUDE_ALTERNATE_GCS
jprovino
parents:
7477
diff
changeset
|
46 #endif // INCLUDE_ALL_GCS |
7199 | 47 |
48 #ifdef PRODUCT | |
49 #define BLOCK_COMMENT(str) /* nothing */ | |
50 #define STOP(error) stop(error) | |
51 #else | |
52 #define BLOCK_COMMENT(str) block_comment(str) | |
53 #define STOP(error) block_comment(error); stop(error) | |
54 #endif | |
55 | |
56 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":") | |
57 | |
58 | |
7204
f0c2369fda5a
8003250: SPARC: move MacroAssembler into separate file
twisti
parents:
7199
diff
changeset
|
59 #ifdef ASSERT |
f0c2369fda5a
8003250: SPARC: move MacroAssembler into separate file
twisti
parents:
7199
diff
changeset
|
60 bool AbstractAssembler::pd_check_instruction_mark() { return true; } |
f0c2369fda5a
8003250: SPARC: move MacroAssembler into separate file
twisti
parents:
7199
diff
changeset
|
61 #endif |
f0c2369fda5a
8003250: SPARC: move MacroAssembler into separate file
twisti
parents:
7199
diff
changeset
|
62 |
7199 | 63 static Assembler::Condition reverse[] = { |
64 Assembler::noOverflow /* overflow = 0x0 */ , | |
65 Assembler::overflow /* noOverflow = 0x1 */ , | |
66 Assembler::aboveEqual /* carrySet = 0x2, below = 0x2 */ , | |
67 Assembler::below /* aboveEqual = 0x3, carryClear = 0x3 */ , | |
68 Assembler::notZero /* zero = 0x4, equal = 0x4 */ , | |
69 Assembler::zero /* notZero = 0x5, notEqual = 0x5 */ , | |
70 Assembler::above /* belowEqual = 0x6 */ , | |
71 Assembler::belowEqual /* above = 0x7 */ , | |
72 Assembler::positive /* negative = 0x8 */ , | |
73 Assembler::negative /* positive = 0x9 */ , | |
74 Assembler::noParity /* parity = 0xa */ , | |
75 Assembler::parity /* noParity = 0xb */ , | |
76 Assembler::greaterEqual /* less = 0xc */ , | |
77 Assembler::less /* greaterEqual = 0xd */ , | |
78 Assembler::greater /* lessEqual = 0xe */ , | |
79 Assembler::lessEqual /* greater = 0xf, */ | |
80 | |
81 }; | |
82 | |
83 | |
84 // Implementation of MacroAssembler | |
85 | |
86 // First all the versions that have distinct versions depending on 32/64 bit | |
87 // Unless the difference is trivial (1 line or so). | |
88 | |
89 #ifndef _LP64 | |
90 | |
91 // 32bit versions | |
92 | |
93 Address MacroAssembler::as_Address(AddressLiteral adr) { | |
94 return Address(adr.target(), adr.rspec()); | |
95 } | |
96 | |
97 Address MacroAssembler::as_Address(ArrayAddress adr) { | |
98 return Address::make_array(adr); | |
99 } | |
100 | |
101 int MacroAssembler::biased_locking_enter(Register lock_reg, | |
102 Register obj_reg, | |
103 Register swap_reg, | |
104 Register tmp_reg, | |
105 bool swap_reg_contains_mark, | |
106 Label& done, | |
107 Label* slow_case, | |
108 BiasedLockingCounters* counters) { | |
109 assert(UseBiasedLocking, "why call this otherwise?"); | |
110 assert(swap_reg == rax, "swap_reg must be rax, for cmpxchg"); | |
111 assert_different_registers(lock_reg, obj_reg, swap_reg); | |
112 | |
113 if (PrintBiasedLockingStatistics && counters == NULL) | |
114 counters = BiasedLocking::counters(); | |
115 | |
116 bool need_tmp_reg = false; | |
117 if (tmp_reg == noreg) { | |
118 need_tmp_reg = true; | |
119 tmp_reg = lock_reg; | |
120 } else { | |
121 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg); | |
122 } | |
123 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); | |
124 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes()); | |
125 Address klass_addr (obj_reg, oopDesc::klass_offset_in_bytes()); | |
126 Address saved_mark_addr(lock_reg, 0); | |
127 | |
128 // Biased locking | |
129 // See whether the lock is currently biased toward our thread and | |
130 // whether the epoch is still valid | |
131 // Note that the runtime guarantees sufficient alignment of JavaThread | |
132 // pointers to allow age to be placed into low bits | |
133 // First check to see whether biasing is even enabled for this object | |
134 Label cas_label; | |
135 int null_check_offset = -1; | |
136 if (!swap_reg_contains_mark) { | |
137 null_check_offset = offset(); | |
138 movl(swap_reg, mark_addr); | |
139 } | |
140 if (need_tmp_reg) { | |
141 push(tmp_reg); | |
142 } | |
143 movl(tmp_reg, swap_reg); | |
144 andl(tmp_reg, markOopDesc::biased_lock_mask_in_place); | |
145 cmpl(tmp_reg, markOopDesc::biased_lock_pattern); | |
146 if (need_tmp_reg) { | |
147 pop(tmp_reg); | |
148 } | |
149 jcc(Assembler::notEqual, cas_label); | |
150 // The bias pattern is present in the object's header. Need to check | |
151 // whether the bias owner and the epoch are both still current. | |
152 // Note that because there is no current thread register on x86 we | |
153 // need to store off the mark word we read out of the object to | |
154 // avoid reloading it and needing to recheck invariants below. This | |
155 // store is unfortunate but it makes the overall code shorter and | |
156 // simpler. | |
157 movl(saved_mark_addr, swap_reg); | |
158 if (need_tmp_reg) { | |
159 push(tmp_reg); | |
160 } | |
161 get_thread(tmp_reg); | |
162 xorl(swap_reg, tmp_reg); | |
163 if (swap_reg_contains_mark) { | |
164 null_check_offset = offset(); | |
165 } | |
166 movl(tmp_reg, klass_addr); | |
167 xorl(swap_reg, Address(tmp_reg, Klass::prototype_header_offset())); | |
168 andl(swap_reg, ~((int) markOopDesc::age_mask_in_place)); | |
169 if (need_tmp_reg) { | |
170 pop(tmp_reg); | |
171 } | |
172 if (counters != NULL) { | |
173 cond_inc32(Assembler::zero, | |
174 ExternalAddress((address)counters->biased_lock_entry_count_addr())); | |
175 } | |
176 jcc(Assembler::equal, done); | |
177 | |
178 Label try_revoke_bias; | |
179 Label try_rebias; | |
180 | |
181 // At this point we know that the header has the bias pattern and | |
182 // that we are not the bias owner in the current epoch. We need to | |
183 // figure out more details about the state of the header in order to | |
184 // know what operations can be legally performed on the object's | |
185 // header. | |
186 | |
187 // If the low three bits in the xor result aren't clear, that means | |
188 // the prototype header is no longer biased and we have to revoke | |
189 // the bias on this object. | |
190 testl(swap_reg, markOopDesc::biased_lock_mask_in_place); | |
191 jcc(Assembler::notZero, try_revoke_bias); | |
192 | |
193 // Biasing is still enabled for this data type. See whether the | |
194 // epoch of the current bias is still valid, meaning that the epoch | |
195 // bits of the mark word are equal to the epoch bits of the | |
196 // prototype header. (Note that the prototype header's epoch bits | |
197 // only change at a safepoint.) If not, attempt to rebias the object | |
198 // toward the current thread. Note that we must be absolutely sure | |
199 // that the current epoch is invalid in order to do this because | |
200 // otherwise the manipulations it performs on the mark word are | |
201 // illegal. | |
202 testl(swap_reg, markOopDesc::epoch_mask_in_place); | |
203 jcc(Assembler::notZero, try_rebias); | |
204 | |
205 // The epoch of the current bias is still valid but we know nothing | |
206 // about the owner; it might be set or it might be clear. Try to | |
207 // acquire the bias of the object using an atomic operation. If this | |
208 // fails we will go in to the runtime to revoke the object's bias. | |
209 // Note that we first construct the presumed unbiased header so we | |
210 // don't accidentally blow away another thread's valid bias. | |
211 movl(swap_reg, saved_mark_addr); | |
212 andl(swap_reg, | |
213 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place); | |
214 if (need_tmp_reg) { | |
215 push(tmp_reg); | |
216 } | |
217 get_thread(tmp_reg); | |
218 orl(tmp_reg, swap_reg); | |
219 if (os::is_MP()) { | |
220 lock(); | |
221 } | |
222 cmpxchgptr(tmp_reg, Address(obj_reg, 0)); | |
223 if (need_tmp_reg) { | |
224 pop(tmp_reg); | |
225 } | |
226 // If the biasing toward our thread failed, this means that | |
227 // another thread succeeded in biasing it toward itself and we | |
228 // need to revoke that bias. The revocation will occur in the | |
229 // interpreter runtime in the slow case. | |
230 if (counters != NULL) { | |
231 cond_inc32(Assembler::zero, | |
232 ExternalAddress((address)counters->anonymously_biased_lock_entry_count_addr())); | |
233 } | |
234 if (slow_case != NULL) { | |
235 jcc(Assembler::notZero, *slow_case); | |
236 } | |
237 jmp(done); | |
238 | |
239 bind(try_rebias); | |
240 // At this point we know the epoch has expired, meaning that the | |
241 // current "bias owner", if any, is actually invalid. Under these | |
242 // circumstances _only_, we are allowed to use the current header's | |
243 // value as the comparison value when doing the cas to acquire the | |
244 // bias in the current epoch. In other words, we allow transfer of | |
245 // the bias from one thread to another directly in this situation. | |
246 // | |
247 // FIXME: due to a lack of registers we currently blow away the age | |
248 // bits in this situation. Should attempt to preserve them. | |
249 if (need_tmp_reg) { | |
250 push(tmp_reg); | |
251 } | |
252 get_thread(tmp_reg); | |
253 movl(swap_reg, klass_addr); | |
254 orl(tmp_reg, Address(swap_reg, Klass::prototype_header_offset())); | |
255 movl(swap_reg, saved_mark_addr); | |
256 if (os::is_MP()) { | |
257 lock(); | |
258 } | |
259 cmpxchgptr(tmp_reg, Address(obj_reg, 0)); | |
260 if (need_tmp_reg) { | |
261 pop(tmp_reg); | |
262 } | |
263 // If the biasing toward our thread failed, then another thread | |
264 // succeeded in biasing it toward itself and we need to revoke that | |
265 // bias. The revocation will occur in the runtime in the slow case. | |
266 if (counters != NULL) { | |
267 cond_inc32(Assembler::zero, | |
268 ExternalAddress((address)counters->rebiased_lock_entry_count_addr())); | |
269 } | |
270 if (slow_case != NULL) { | |
271 jcc(Assembler::notZero, *slow_case); | |
272 } | |
273 jmp(done); | |
274 | |
275 bind(try_revoke_bias); | |
276 // The prototype mark in the klass doesn't have the bias bit set any | |
277 // more, indicating that objects of this data type are not supposed | |
278 // to be biased any more. We are going to try to reset the mark of | |
279 // this object to the prototype value and fall through to the | |
280 // CAS-based locking scheme. Note that if our CAS fails, it means | |
281 // that another thread raced us for the privilege of revoking the | |
282 // bias of this particular object, so it's okay to continue in the | |
283 // normal locking code. | |
284 // | |
285 // FIXME: due to a lack of registers we currently blow away the age | |
286 // bits in this situation. Should attempt to preserve them. | |
287 movl(swap_reg, saved_mark_addr); | |
288 if (need_tmp_reg) { | |
289 push(tmp_reg); | |
290 } | |
291 movl(tmp_reg, klass_addr); | |
292 movl(tmp_reg, Address(tmp_reg, Klass::prototype_header_offset())); | |
293 if (os::is_MP()) { | |
294 lock(); | |
295 } | |
296 cmpxchgptr(tmp_reg, Address(obj_reg, 0)); | |
297 if (need_tmp_reg) { | |
298 pop(tmp_reg); | |
299 } | |
300 // Fall through to the normal CAS-based lock, because no matter what | |
301 // the result of the above CAS, some thread must have succeeded in | |
302 // removing the bias bit from the object's header. | |
303 if (counters != NULL) { | |
304 cond_inc32(Assembler::zero, | |
305 ExternalAddress((address)counters->revoked_lock_entry_count_addr())); | |
306 } | |
307 | |
308 bind(cas_label); | |
309 | |
310 return null_check_offset; | |
311 } | |
312 void MacroAssembler::call_VM_leaf_base(address entry_point, | |
313 int number_of_arguments) { | |
314 call(RuntimeAddress(entry_point)); | |
315 increment(rsp, number_of_arguments * wordSize); | |
316 } | |
317 | |
318 void MacroAssembler::cmpklass(Address src1, Metadata* obj) { | |
319 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); | |
320 } | |
321 | |
322 void MacroAssembler::cmpklass(Register src1, Metadata* obj) { | |
323 cmp_literal32(src1, (int32_t)obj, metadata_Relocation::spec_for_immediate()); | |
324 } | |
325 | |
326 void MacroAssembler::cmpoop(Address src1, jobject obj) { | |
327 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); | |
328 } | |
329 | |
330 void MacroAssembler::cmpoop(Register src1, jobject obj) { | |
331 cmp_literal32(src1, (int32_t)obj, oop_Relocation::spec_for_immediate()); | |
332 } | |
333 | |
334 void MacroAssembler::extend_sign(Register hi, Register lo) { | |
335 // According to Intel Doc. AP-526, "Integer Divide", p.18. | |
336 if (VM_Version::is_P6() && hi == rdx && lo == rax) { | |
337 cdql(); | |
338 } else { | |
339 movl(hi, lo); | |
340 sarl(hi, 31); | |
341 } | |
342 } | |
343 | |
344 void MacroAssembler::jC2(Register tmp, Label& L) { | |
345 // set parity bit if FPU flag C2 is set (via rax) | |
346 save_rax(tmp); | |
347 fwait(); fnstsw_ax(); | |
348 sahf(); | |
349 restore_rax(tmp); | |
350 // branch | |
351 jcc(Assembler::parity, L); | |
352 } | |
353 | |
354 void MacroAssembler::jnC2(Register tmp, Label& L) { | |
355 // set parity bit if FPU flag C2 is set (via rax) | |
356 save_rax(tmp); | |
357 fwait(); fnstsw_ax(); | |
358 sahf(); | |
359 restore_rax(tmp); | |
360 // branch | |
361 jcc(Assembler::noParity, L); | |
362 } | |
363 | |
364 // 32bit can do a case table jump in one instruction but we no longer allow the base | |
365 // to be installed in the Address class | |
366 void MacroAssembler::jump(ArrayAddress entry) { | |
367 jmp(as_Address(entry)); | |
368 } | |
369 | |
370 // Note: y_lo will be destroyed | |
371 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { | |
372 // Long compare for Java (semantics as described in JVM spec.) | |
373 Label high, low, done; | |
374 | |
375 cmpl(x_hi, y_hi); | |
376 jcc(Assembler::less, low); | |
377 jcc(Assembler::greater, high); | |
378 // x_hi is the return register | |
379 xorl(x_hi, x_hi); | |
380 cmpl(x_lo, y_lo); | |
381 jcc(Assembler::below, low); | |
382 jcc(Assembler::equal, done); | |
383 | |
384 bind(high); | |
385 xorl(x_hi, x_hi); | |
386 increment(x_hi); | |
387 jmp(done); | |
388 | |
389 bind(low); | |
390 xorl(x_hi, x_hi); | |
391 decrementl(x_hi); | |
392 | |
393 bind(done); | |
394 } | |
395 | |
396 void MacroAssembler::lea(Register dst, AddressLiteral src) { | |
397 mov_literal32(dst, (int32_t)src.target(), src.rspec()); | |
398 } | |
399 | |
400 void MacroAssembler::lea(Address dst, AddressLiteral adr) { | |
401 // leal(dst, as_Address(adr)); | |
402 // see note in movl as to why we must use a move | |
403 mov_literal32(dst, (int32_t) adr.target(), adr.rspec()); | |
404 } | |
405 | |
406 void MacroAssembler::leave() { | |
407 mov(rsp, rbp); | |
408 pop(rbp); | |
409 } | |
410 | |
411 void MacroAssembler::lmul(int x_rsp_offset, int y_rsp_offset) { | |
412 // Multiplication of two Java long values stored on the stack | |
413 // as illustrated below. Result is in rdx:rax. | |
414 // | |
415 // rsp ---> [ ?? ] \ \ | |
416 // .... | y_rsp_offset | | |
417 // [ y_lo ] / (in bytes) | x_rsp_offset | |
418 // [ y_hi ] | (in bytes) | |
419 // .... | | |
420 // [ x_lo ] / | |
421 // [ x_hi ] | |
422 // .... | |
423 // | |
424 // Basic idea: lo(result) = lo(x_lo * y_lo) | |
425 // hi(result) = hi(x_lo * y_lo) + lo(x_hi * y_lo) + lo(x_lo * y_hi) | |
426 Address x_hi(rsp, x_rsp_offset + wordSize); Address x_lo(rsp, x_rsp_offset); | |
427 Address y_hi(rsp, y_rsp_offset + wordSize); Address y_lo(rsp, y_rsp_offset); | |
428 Label quick; | |
429 // load x_hi, y_hi and check if quick | |
430 // multiplication is possible | |
431 movl(rbx, x_hi); | |
432 movl(rcx, y_hi); | |
433 movl(rax, rbx); | |
434 orl(rbx, rcx); // rbx, = 0 <=> x_hi = 0 and y_hi = 0 | |
435 jcc(Assembler::zero, quick); // if rbx, = 0 do quick multiply | |
436 // do full multiplication | |
437 // 1st step | |
438 mull(y_lo); // x_hi * y_lo | |
439 movl(rbx, rax); // save lo(x_hi * y_lo) in rbx, | |
440 // 2nd step | |
441 movl(rax, x_lo); | |
442 mull(rcx); // x_lo * y_hi | |
443 addl(rbx, rax); // add lo(x_lo * y_hi) to rbx, | |
444 // 3rd step | |
445 bind(quick); // note: rbx, = 0 if quick multiply! | |
446 movl(rax, x_lo); | |
447 mull(y_lo); // x_lo * y_lo | |
448 addl(rdx, rbx); // correct hi(x_lo * y_lo) | |
449 } | |
450 | |
451 void MacroAssembler::lneg(Register hi, Register lo) { | |
452 negl(lo); | |
453 adcl(hi, 0); | |
454 negl(hi); | |
455 } | |
456 | |
457 void MacroAssembler::lshl(Register hi, Register lo) { | |
458 // Java shift left long support (semantics as described in JVM spec., p.305) | |
459 // (basic idea for shift counts s >= n: x << s == (x << n) << (s - n)) | |
460 // shift value is in rcx ! | |
461 assert(hi != rcx, "must not use rcx"); | |
462 assert(lo != rcx, "must not use rcx"); | |
463 const Register s = rcx; // shift count | |
464 const int n = BitsPerWord; | |
465 Label L; | |
466 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) | |
467 cmpl(s, n); // if (s < n) | |
468 jcc(Assembler::less, L); // else (s >= n) | |
469 movl(hi, lo); // x := x << n | |
470 xorl(lo, lo); | |
471 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! | |
472 bind(L); // s (mod n) < n | |
473 shldl(hi, lo); // x := x << s | |
474 shll(lo); | |
475 } | |
476 | |
477 | |
478 void MacroAssembler::lshr(Register hi, Register lo, bool sign_extension) { | |
479 // Java shift right long support (semantics as described in JVM spec., p.306 & p.310) | |
480 // (basic idea for shift counts s >= n: x >> s == (x >> n) >> (s - n)) | |
481 assert(hi != rcx, "must not use rcx"); | |
482 assert(lo != rcx, "must not use rcx"); | |
483 const Register s = rcx; // shift count | |
484 const int n = BitsPerWord; | |
485 Label L; | |
486 andl(s, 0x3f); // s := s & 0x3f (s < 0x40) | |
487 cmpl(s, n); // if (s < n) | |
488 jcc(Assembler::less, L); // else (s >= n) | |
489 movl(lo, hi); // x := x >> n | |
490 if (sign_extension) sarl(hi, 31); | |
491 else xorl(hi, hi); | |
492 // Note: subl(s, n) is not needed since the Intel shift instructions work rcx mod n! | |
493 bind(L); // s (mod n) < n | |
494 shrdl(lo, hi); // x := x >> s | |
495 if (sign_extension) sarl(hi); | |
496 else shrl(hi); | |
497 } | |
498 | |
499 void MacroAssembler::movoop(Register dst, jobject obj) { | |
500 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); | |
501 } | |
502 | |
503 void MacroAssembler::movoop(Address dst, jobject obj) { | |
504 mov_literal32(dst, (int32_t)obj, oop_Relocation::spec_for_immediate()); | |
505 } | |
506 | |
507 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { | |
508 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); | |
509 } | |
510 | |
511 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { | |
512 mov_literal32(dst, (int32_t)obj, metadata_Relocation::spec_for_immediate()); | |
513 } | |
514 | |
515 void MacroAssembler::movptr(Register dst, AddressLiteral src) { | |
516 if (src.is_lval()) { | |
517 mov_literal32(dst, (intptr_t)src.target(), src.rspec()); | |
518 } else { | |
519 movl(dst, as_Address(src)); | |
520 } | |
521 } | |
522 | |
523 void MacroAssembler::movptr(ArrayAddress dst, Register src) { | |
524 movl(as_Address(dst), src); | |
525 } | |
526 | |
527 void MacroAssembler::movptr(Register dst, ArrayAddress src) { | |
528 movl(dst, as_Address(src)); | |
529 } | |
530 | |
531 // src should NEVER be a real pointer. Use AddressLiteral for true pointers | |
532 void MacroAssembler::movptr(Address dst, intptr_t src) { | |
533 movl(dst, src); | |
534 } | |
535 | |
536 | |
537 void MacroAssembler::pop_callee_saved_registers() { | |
538 pop(rcx); | |
539 pop(rdx); | |
540 pop(rdi); | |
541 pop(rsi); | |
542 } | |
543 | |
544 void MacroAssembler::pop_fTOS() { | |
545 fld_d(Address(rsp, 0)); | |
546 addl(rsp, 2 * wordSize); | |
547 } | |
548 | |
549 void MacroAssembler::push_callee_saved_registers() { | |
550 push(rsi); | |
551 push(rdi); | |
552 push(rdx); | |
553 push(rcx); | |
554 } | |
555 | |
556 void MacroAssembler::push_fTOS() { | |
557 subl(rsp, 2 * wordSize); | |
558 fstp_d(Address(rsp, 0)); | |
559 } | |
560 | |
561 | |
562 void MacroAssembler::pushoop(jobject obj) { | |
563 push_literal32((int32_t)obj, oop_Relocation::spec_for_immediate()); | |
564 } | |
565 | |
566 void MacroAssembler::pushklass(Metadata* obj) { | |
567 push_literal32((int32_t)obj, metadata_Relocation::spec_for_immediate()); | |
568 } | |
569 | |
570 void MacroAssembler::pushptr(AddressLiteral src) { | |
571 if (src.is_lval()) { | |
572 push_literal32((int32_t)src.target(), src.rspec()); | |
573 } else { | |
574 pushl(as_Address(src)); | |
575 } | |
576 } | |
577 | |
578 void MacroAssembler::set_word_if_not_zero(Register dst) { | |
579 xorl(dst, dst); | |
580 set_byte_if_not_zero(dst); | |
581 } | |
582 | |
583 static void pass_arg0(MacroAssembler* masm, Register arg) { | |
584 masm->push(arg); | |
585 } | |
586 | |
587 static void pass_arg1(MacroAssembler* masm, Register arg) { | |
588 masm->push(arg); | |
589 } | |
590 | |
591 static void pass_arg2(MacroAssembler* masm, Register arg) { | |
592 masm->push(arg); | |
593 } | |
594 | |
595 static void pass_arg3(MacroAssembler* masm, Register arg) { | |
596 masm->push(arg); | |
597 } | |
598 | |
599 #ifndef PRODUCT | |
600 extern "C" void findpc(intptr_t x); | |
601 #endif | |
602 | |
603 void MacroAssembler::debug32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip, char* msg) { | |
604 // In order to get locks to work, we need to fake a in_VM state | |
605 JavaThread* thread = JavaThread::current(); | |
606 JavaThreadState saved_state = thread->thread_state(); | |
607 thread->set_thread_state(_thread_in_vm); | |
608 if (ShowMessageBoxOnError) { | |
609 JavaThread* thread = JavaThread::current(); | |
610 JavaThreadState saved_state = thread->thread_state(); | |
611 thread->set_thread_state(_thread_in_vm); | |
612 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { | |
613 ttyLocker ttyl; | |
614 BytecodeCounter::print(); | |
615 } | |
616 // To see where a verify_oop failed, get $ebx+40/X for this frame. | |
617 // This is the value of eip which points to where verify_oop will return. | |
618 if (os::message_box(msg, "Execution stopped, print registers?")) { | |
619 print_state32(rdi, rsi, rbp, rsp, rbx, rdx, rcx, rax, eip); | |
620 BREAKPOINT; | |
621 } | |
622 } else { | |
623 ttyLocker ttyl; | |
624 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", msg); | |
625 } | |
626 // Don't assert holding the ttyLock | |
627 assert(false, err_msg("DEBUG MESSAGE: %s", msg)); | |
628 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); | |
629 } | |
630 | |
631 void MacroAssembler::print_state32(int rdi, int rsi, int rbp, int rsp, int rbx, int rdx, int rcx, int rax, int eip) { | |
632 ttyLocker ttyl; | |
633 FlagSetting fs(Debugging, true); | |
634 tty->print_cr("eip = 0x%08x", eip); | |
635 #ifndef PRODUCT | |
636 if ((WizardMode || Verbose) && PrintMiscellaneous) { | |
637 tty->cr(); | |
638 findpc(eip); | |
639 tty->cr(); | |
640 } | |
641 #endif | |
642 #define PRINT_REG(rax) \ | |
643 { tty->print("%s = ", #rax); os::print_location(tty, rax); } | |
644 PRINT_REG(rax); | |
645 PRINT_REG(rbx); | |
646 PRINT_REG(rcx); | |
647 PRINT_REG(rdx); | |
648 PRINT_REG(rdi); | |
649 PRINT_REG(rsi); | |
650 PRINT_REG(rbp); | |
651 PRINT_REG(rsp); | |
652 #undef PRINT_REG | |
653 // Print some words near top of staack. | |
654 int* dump_sp = (int*) rsp; | |
655 for (int col1 = 0; col1 < 8; col1++) { | |
656 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); | |
657 os::print_location(tty, *dump_sp++); | |
658 } | |
659 for (int row = 0; row < 16; row++) { | |
660 tty->print("(rsp+0x%03x) 0x%08x: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (intptr_t)dump_sp); | |
661 for (int col = 0; col < 8; col++) { | |
662 tty->print(" 0x%08x", *dump_sp++); | |
663 } | |
664 tty->cr(); | |
665 } | |
666 // Print some instructions around pc: | |
667 Disassembler::decode((address)eip-64, (address)eip); | |
668 tty->print_cr("--------"); | |
669 Disassembler::decode((address)eip, (address)eip+32); | |
670 } | |
671 | |
672 void MacroAssembler::stop(const char* msg) { | |
673 ExternalAddress message((address)msg); | |
674 // push address of message | |
675 pushptr(message.addr()); | |
676 { Label L; call(L, relocInfo::none); bind(L); } // push eip | |
677 pusha(); // push registers | |
678 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug32))); | |
679 hlt(); | |
680 } | |
681 | |
682 void MacroAssembler::warn(const char* msg) { | |
683 push_CPU_state(); | |
684 | |
685 ExternalAddress message((address) msg); | |
686 // push address of message | |
687 pushptr(message.addr()); | |
688 | |
689 call(RuntimeAddress(CAST_FROM_FN_PTR(address, warning))); | |
690 addl(rsp, wordSize); // discard argument | |
691 pop_CPU_state(); | |
692 } | |
693 | |
694 void MacroAssembler::print_state() { | |
695 { Label L; call(L, relocInfo::none); bind(L); } // push eip | |
696 pusha(); // push registers | |
697 | |
698 push_CPU_state(); | |
699 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::print_state32))); | |
700 pop_CPU_state(); | |
701 | |
702 popa(); | |
703 addl(rsp, wordSize); | |
704 } | |
705 | |
706 #else // _LP64 | |
707 | |
708 // 64 bit versions | |
709 | |
710 Address MacroAssembler::as_Address(AddressLiteral adr) { | |
711 // amd64 always does this as a pc-rel | |
712 // we can be absolute or disp based on the instruction type | |
713 // jmp/call are displacements others are absolute | |
714 assert(!adr.is_lval(), "must be rval"); | |
715 assert(reachable(adr), "must be"); | |
716 return Address((int32_t)(intptr_t)(adr.target() - pc()), adr.target(), adr.reloc()); | |
717 | |
718 } | |
719 | |
720 Address MacroAssembler::as_Address(ArrayAddress adr) { | |
721 AddressLiteral base = adr.base(); | |
722 lea(rscratch1, base); | |
723 Address index = adr.index(); | |
724 assert(index._disp == 0, "must not have disp"); // maybe it can? | |
725 Address array(rscratch1, index._index, index._scale, index._disp); | |
726 return array; | |
727 } | |
728 | |
729 int MacroAssembler::biased_locking_enter(Register lock_reg, | |
730 Register obj_reg, | |
731 Register swap_reg, | |
732 Register tmp_reg, | |
733 bool swap_reg_contains_mark, | |
734 Label& done, | |
735 Label* slow_case, | |
736 BiasedLockingCounters* counters) { | |
737 assert(UseBiasedLocking, "why call this otherwise?"); | |
738 assert(swap_reg == rax, "swap_reg must be rax for cmpxchgq"); | |
739 assert(tmp_reg != noreg, "tmp_reg must be supplied"); | |
740 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg); | |
741 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits, "biased locking makes assumptions about bit layout"); | |
742 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes()); | |
743 Address saved_mark_addr(lock_reg, 0); | |
744 | |
745 if (PrintBiasedLockingStatistics && counters == NULL) | |
746 counters = BiasedLocking::counters(); | |
747 | |
748 // Biased locking | |
749 // See whether the lock is currently biased toward our thread and | |
750 // whether the epoch is still valid | |
751 // Note that the runtime guarantees sufficient alignment of JavaThread | |
752 // pointers to allow age to be placed into low bits | |
753 // First check to see whether biasing is even enabled for this object | |
754 Label cas_label; | |
755 int null_check_offset = -1; | |
756 if (!swap_reg_contains_mark) { | |
757 null_check_offset = offset(); | |
758 movq(swap_reg, mark_addr); | |
759 } | |
760 movq(tmp_reg, swap_reg); | |
761 andq(tmp_reg, markOopDesc::biased_lock_mask_in_place); | |
762 cmpq(tmp_reg, markOopDesc::biased_lock_pattern); | |
763 jcc(Assembler::notEqual, cas_label); | |
764 // The bias pattern is present in the object's header. Need to check | |
765 // whether the bias owner and the epoch are both still current. | |
766 load_prototype_header(tmp_reg, obj_reg); | |
767 orq(tmp_reg, r15_thread); | |
768 xorq(tmp_reg, swap_reg); | |
769 andq(tmp_reg, ~((int) markOopDesc::age_mask_in_place)); | |
770 if (counters != NULL) { | |
771 cond_inc32(Assembler::zero, | |
772 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr())); | |
773 } | |
774 jcc(Assembler::equal, done); | |
775 | |
776 Label try_revoke_bias; | |
777 Label try_rebias; | |
778 | |
779 // At this point we know that the header has the bias pattern and | |
780 // that we are not the bias owner in the current epoch. We need to | |
781 // figure out more details about the state of the header in order to | |
782 // know what operations can be legally performed on the object's | |
783 // header. | |
784 | |
785 // If the low three bits in the xor result aren't clear, that means | |
786 // the prototype header is no longer biased and we have to revoke | |
787 // the bias on this object. | |
788 testq(tmp_reg, markOopDesc::biased_lock_mask_in_place); | |
789 jcc(Assembler::notZero, try_revoke_bias); | |
790 | |
791 // Biasing is still enabled for this data type. See whether the | |
792 // epoch of the current bias is still valid, meaning that the epoch | |
793 // bits of the mark word are equal to the epoch bits of the | |
794 // prototype header. (Note that the prototype header's epoch bits | |
795 // only change at a safepoint.) If not, attempt to rebias the object | |
796 // toward the current thread. Note that we must be absolutely sure | |
797 // that the current epoch is invalid in order to do this because | |
798 // otherwise the manipulations it performs on the mark word are | |
799 // illegal. | |
800 testq(tmp_reg, markOopDesc::epoch_mask_in_place); | |
801 jcc(Assembler::notZero, try_rebias); | |
802 | |
803 // The epoch of the current bias is still valid but we know nothing | |
804 // about the owner; it might be set or it might be clear. Try to | |
805 // acquire the bias of the object using an atomic operation. If this | |
806 // fails we will go in to the runtime to revoke the object's bias. | |
807 // Note that we first construct the presumed unbiased header so we | |
808 // don't accidentally blow away another thread's valid bias. | |
809 andq(swap_reg, | |
810 markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place); | |
811 movq(tmp_reg, swap_reg); | |
812 orq(tmp_reg, r15_thread); | |
813 if (os::is_MP()) { | |
814 lock(); | |
815 } | |
816 cmpxchgq(tmp_reg, Address(obj_reg, 0)); | |
817 // If the biasing toward our thread failed, this means that | |
818 // another thread succeeded in biasing it toward itself and we | |
819 // need to revoke that bias. The revocation will occur in the | |
820 // interpreter runtime in the slow case. | |
821 if (counters != NULL) { | |
822 cond_inc32(Assembler::zero, | |
823 ExternalAddress((address) counters->anonymously_biased_lock_entry_count_addr())); | |
824 } | |
825 if (slow_case != NULL) { | |
826 jcc(Assembler::notZero, *slow_case); | |
827 } | |
828 jmp(done); | |
829 | |
830 bind(try_rebias); | |
831 // At this point we know the epoch has expired, meaning that the | |
832 // current "bias owner", if any, is actually invalid. Under these | |
833 // circumstances _only_, we are allowed to use the current header's | |
834 // value as the comparison value when doing the cas to acquire the | |
835 // bias in the current epoch. In other words, we allow transfer of | |
836 // the bias from one thread to another directly in this situation. | |
837 // | |
838 // FIXME: due to a lack of registers we currently blow away the age | |
839 // bits in this situation. Should attempt to preserve them. | |
840 load_prototype_header(tmp_reg, obj_reg); | |
841 orq(tmp_reg, r15_thread); | |
842 if (os::is_MP()) { | |
843 lock(); | |
844 } | |
845 cmpxchgq(tmp_reg, Address(obj_reg, 0)); | |
846 // If the biasing toward our thread failed, then another thread | |
847 // succeeded in biasing it toward itself and we need to revoke that | |
848 // bias. The revocation will occur in the runtime in the slow case. | |
849 if (counters != NULL) { | |
850 cond_inc32(Assembler::zero, | |
851 ExternalAddress((address) counters->rebiased_lock_entry_count_addr())); | |
852 } | |
853 if (slow_case != NULL) { | |
854 jcc(Assembler::notZero, *slow_case); | |
855 } | |
856 jmp(done); | |
857 | |
858 bind(try_revoke_bias); | |
859 // The prototype mark in the klass doesn't have the bias bit set any | |
860 // more, indicating that objects of this data type are not supposed | |
861 // to be biased any more. We are going to try to reset the mark of | |
862 // this object to the prototype value and fall through to the | |
863 // CAS-based locking scheme. Note that if our CAS fails, it means | |
864 // that another thread raced us for the privilege of revoking the | |
865 // bias of this particular object, so it's okay to continue in the | |
866 // normal locking code. | |
867 // | |
868 // FIXME: due to a lack of registers we currently blow away the age | |
869 // bits in this situation. Should attempt to preserve them. | |
870 load_prototype_header(tmp_reg, obj_reg); | |
871 if (os::is_MP()) { | |
872 lock(); | |
873 } | |
874 cmpxchgq(tmp_reg, Address(obj_reg, 0)); | |
875 // Fall through to the normal CAS-based lock, because no matter what | |
876 // the result of the above CAS, some thread must have succeeded in | |
877 // removing the bias bit from the object's header. | |
878 if (counters != NULL) { | |
879 cond_inc32(Assembler::zero, | |
880 ExternalAddress((address) counters->revoked_lock_entry_count_addr())); | |
881 } | |
882 | |
883 bind(cas_label); | |
884 | |
885 return null_check_offset; | |
886 } | |
887 | |
888 void MacroAssembler::call_VM_leaf_base(address entry_point, int num_args) { | |
889 Label L, E; | |
890 | |
891 #ifdef _WIN64 | |
892 // Windows always allocates space for it's register args | |
893 assert(num_args <= 4, "only register arguments supported"); | |
894 subq(rsp, frame::arg_reg_save_area_bytes); | |
895 #endif | |
896 | |
897 // Align stack if necessary | |
898 testl(rsp, 15); | |
899 jcc(Assembler::zero, L); | |
900 | |
901 subq(rsp, 8); | |
902 { | |
903 call(RuntimeAddress(entry_point)); | |
904 } | |
905 addq(rsp, 8); | |
906 jmp(E); | |
907 | |
908 bind(L); | |
909 { | |
910 call(RuntimeAddress(entry_point)); | |
911 } | |
912 | |
913 bind(E); | |
914 | |
915 #ifdef _WIN64 | |
916 // restore stack pointer | |
917 addq(rsp, frame::arg_reg_save_area_bytes); | |
918 #endif | |
919 | |
920 } | |
921 | |
922 void MacroAssembler::cmp64(Register src1, AddressLiteral src2) { | |
923 assert(!src2.is_lval(), "should use cmpptr"); | |
924 | |
925 if (reachable(src2)) { | |
926 cmpq(src1, as_Address(src2)); | |
927 } else { | |
928 lea(rscratch1, src2); | |
929 Assembler::cmpq(src1, Address(rscratch1, 0)); | |
930 } | |
931 } | |
932 | |
933 int MacroAssembler::corrected_idivq(Register reg) { | |
934 // Full implementation of Java ldiv and lrem; checks for special | |
935 // case as described in JVM spec., p.243 & p.271. The function | |
936 // returns the (pc) offset of the idivl instruction - may be needed | |
937 // for implicit exceptions. | |
938 // | |
939 // normal case special case | |
940 // | |
941 // input : rax: dividend min_long | |
942 // reg: divisor (may not be eax/edx) -1 | |
943 // | |
944 // output: rax: quotient (= rax idiv reg) min_long | |
945 // rdx: remainder (= rax irem reg) 0 | |
946 assert(reg != rax && reg != rdx, "reg cannot be rax or rdx register"); | |
947 static const int64_t min_long = 0x8000000000000000; | |
948 Label normal_case, special_case; | |
949 | |
950 // check for special case | |
951 cmp64(rax, ExternalAddress((address) &min_long)); | |
952 jcc(Assembler::notEqual, normal_case); | |
953 xorl(rdx, rdx); // prepare rdx for possible special case (where | |
954 // remainder = 0) | |
955 cmpq(reg, -1); | |
956 jcc(Assembler::equal, special_case); | |
957 | |
958 // handle normal case | |
959 bind(normal_case); | |
960 cdqq(); | |
961 int idivq_offset = offset(); | |
962 idivq(reg); | |
963 | |
964 // normal and special case exit | |
965 bind(special_case); | |
966 | |
967 return idivq_offset; | |
968 } | |
969 | |
970 void MacroAssembler::decrementq(Register reg, int value) { | |
971 if (value == min_jint) { subq(reg, value); return; } | |
972 if (value < 0) { incrementq(reg, -value); return; } | |
973 if (value == 0) { ; return; } | |
974 if (value == 1 && UseIncDec) { decq(reg) ; return; } | |
975 /* else */ { subq(reg, value) ; return; } | |
976 } | |
977 | |
978 void MacroAssembler::decrementq(Address dst, int value) { | |
979 if (value == min_jint) { subq(dst, value); return; } | |
980 if (value < 0) { incrementq(dst, -value); return; } | |
981 if (value == 0) { ; return; } | |
982 if (value == 1 && UseIncDec) { decq(dst) ; return; } | |
983 /* else */ { subq(dst, value) ; return; } | |
984 } | |
985 | |
986 void MacroAssembler::incrementq(Register reg, int value) { | |
987 if (value == min_jint) { addq(reg, value); return; } | |
988 if (value < 0) { decrementq(reg, -value); return; } | |
989 if (value == 0) { ; return; } | |
990 if (value == 1 && UseIncDec) { incq(reg) ; return; } | |
991 /* else */ { addq(reg, value) ; return; } | |
992 } | |
993 | |
994 void MacroAssembler::incrementq(Address dst, int value) { | |
995 if (value == min_jint) { addq(dst, value); return; } | |
996 if (value < 0) { decrementq(dst, -value); return; } | |
997 if (value == 0) { ; return; } | |
998 if (value == 1 && UseIncDec) { incq(dst) ; return; } | |
999 /* else */ { addq(dst, value) ; return; } | |
1000 } | |
1001 | |
1002 // 32bit can do a case table jump in one instruction but we no longer allow the base | |
1003 // to be installed in the Address class | |
1004 void MacroAssembler::jump(ArrayAddress entry) { | |
1005 lea(rscratch1, entry.base()); | |
1006 Address dispatch = entry.index(); | |
1007 assert(dispatch._base == noreg, "must be"); | |
1008 dispatch._base = rscratch1; | |
1009 jmp(dispatch); | |
1010 } | |
1011 | |
1012 void MacroAssembler::lcmp2int(Register x_hi, Register x_lo, Register y_hi, Register y_lo) { | |
1013 ShouldNotReachHere(); // 64bit doesn't use two regs | |
1014 cmpq(x_lo, y_lo); | |
1015 } | |
1016 | |
1017 void MacroAssembler::lea(Register dst, AddressLiteral src) { | |
1018 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); | |
1019 } | |
1020 | |
1021 void MacroAssembler::lea(Address dst, AddressLiteral adr) { | |
1022 mov_literal64(rscratch1, (intptr_t)adr.target(), adr.rspec()); | |
1023 movptr(dst, rscratch1); | |
1024 } | |
1025 | |
1026 void MacroAssembler::leave() { | |
1027 // %%% is this really better? Why not on 32bit too? | |
7430
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
1028 emit_int8((unsigned char)0xC9); // LEAVE |
7199 | 1029 } |
1030 | |
1031 void MacroAssembler::lneg(Register hi, Register lo) { | |
1032 ShouldNotReachHere(); // 64bit doesn't use two regs | |
1033 negq(lo); | |
1034 } | |
1035 | |
1036 void MacroAssembler::movoop(Register dst, jobject obj) { | |
1037 mov_literal64(dst, (intptr_t)obj, oop_Relocation::spec_for_immediate()); | |
1038 } | |
1039 | |
1040 void MacroAssembler::movoop(Address dst, jobject obj) { | |
1041 mov_literal64(rscratch1, (intptr_t)obj, oop_Relocation::spec_for_immediate()); | |
1042 movq(dst, rscratch1); | |
1043 } | |
1044 | |
1045 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) { | |
1046 mov_literal64(dst, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); | |
1047 } | |
1048 | |
1049 void MacroAssembler::mov_metadata(Address dst, Metadata* obj) { | |
1050 mov_literal64(rscratch1, (intptr_t)obj, metadata_Relocation::spec_for_immediate()); | |
1051 movq(dst, rscratch1); | |
1052 } | |
1053 | |
1054 void MacroAssembler::movptr(Register dst, AddressLiteral src) { | |
1055 if (src.is_lval()) { | |
1056 mov_literal64(dst, (intptr_t)src.target(), src.rspec()); | |
1057 } else { | |
1058 if (reachable(src)) { | |
1059 movq(dst, as_Address(src)); | |
1060 } else { | |
1061 lea(rscratch1, src); | |
1062 movq(dst, Address(rscratch1,0)); | |
1063 } | |
1064 } | |
1065 } | |
1066 | |
1067 void MacroAssembler::movptr(ArrayAddress dst, Register src) { | |
1068 movq(as_Address(dst), src); | |
1069 } | |
1070 | |
1071 void MacroAssembler::movptr(Register dst, ArrayAddress src) { | |
1072 movq(dst, as_Address(src)); | |
1073 } | |
1074 | |
1075 // src should NEVER be a real pointer. Use AddressLiteral for true pointers | |
1076 void MacroAssembler::movptr(Address dst, intptr_t src) { | |
1077 mov64(rscratch1, src); | |
1078 movq(dst, rscratch1); | |
1079 } | |
1080 | |
1081 // These are mostly for initializing NULL | |
1082 void MacroAssembler::movptr(Address dst, int32_t src) { | |
1083 movslq(dst, src); | |
1084 } | |
1085 | |
1086 void MacroAssembler::movptr(Register dst, int32_t src) { | |
1087 mov64(dst, (intptr_t)src); | |
1088 } | |
1089 | |
1090 void MacroAssembler::pushoop(jobject obj) { | |
1091 movoop(rscratch1, obj); | |
1092 push(rscratch1); | |
1093 } | |
1094 | |
1095 void MacroAssembler::pushklass(Metadata* obj) { | |
1096 mov_metadata(rscratch1, obj); | |
1097 push(rscratch1); | |
1098 } | |
1099 | |
1100 void MacroAssembler::pushptr(AddressLiteral src) { | |
1101 lea(rscratch1, src); | |
1102 if (src.is_lval()) { | |
1103 push(rscratch1); | |
1104 } else { | |
1105 pushq(Address(rscratch1, 0)); | |
1106 } | |
1107 } | |
1108 | |
1109 void MacroAssembler::reset_last_Java_frame(bool clear_fp, | |
1110 bool clear_pc) { | |
1111 // we must set sp to zero to clear frame | |
1112 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); | |
1113 // must clear fp, so that compiled frames are not confused; it is | |
1114 // possible that we need it only for debugging | |
1115 if (clear_fp) { | |
1116 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); | |
1117 } | |
1118 | |
1119 if (clear_pc) { | |
1120 movptr(Address(r15_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); | |
1121 } | |
1122 } | |
1123 | |
1124 void MacroAssembler::set_last_Java_frame(Register last_java_sp, | |
1125 Register last_java_fp, | |
1126 address last_java_pc) { | |
1127 // determine last_java_sp register | |
1128 if (!last_java_sp->is_valid()) { | |
1129 last_java_sp = rsp; | |
1130 } | |
1131 | |
1132 // last_java_fp is optional | |
1133 if (last_java_fp->is_valid()) { | |
1134 movptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), | |
1135 last_java_fp); | |
1136 } | |
1137 | |
1138 // last_java_pc is optional | |
1139 if (last_java_pc != NULL) { | |
1140 Address java_pc(r15_thread, | |
1141 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()); | |
1142 lea(rscratch1, InternalAddress(last_java_pc)); | |
1143 movptr(java_pc, rscratch1); | |
1144 } | |
1145 | |
1146 movptr(Address(r15_thread, JavaThread::last_Java_sp_offset()), last_java_sp); | |
1147 } | |
1148 | |
1149 static void pass_arg0(MacroAssembler* masm, Register arg) { | |
1150 if (c_rarg0 != arg ) { | |
1151 masm->mov(c_rarg0, arg); | |
1152 } | |
1153 } | |
1154 | |
1155 static void pass_arg1(MacroAssembler* masm, Register arg) { | |
1156 if (c_rarg1 != arg ) { | |
1157 masm->mov(c_rarg1, arg); | |
1158 } | |
1159 } | |
1160 | |
1161 static void pass_arg2(MacroAssembler* masm, Register arg) { | |
1162 if (c_rarg2 != arg ) { | |
1163 masm->mov(c_rarg2, arg); | |
1164 } | |
1165 } | |
1166 | |
1167 static void pass_arg3(MacroAssembler* masm, Register arg) { | |
1168 if (c_rarg3 != arg ) { | |
1169 masm->mov(c_rarg3, arg); | |
1170 } | |
1171 } | |
1172 | |
1173 void MacroAssembler::stop(const char* msg) { | |
1174 address rip = pc(); | |
1175 pusha(); // get regs on stack | |
1176 lea(c_rarg0, ExternalAddress((address) msg)); | |
1177 lea(c_rarg1, InternalAddress(rip)); | |
1178 movq(c_rarg2, rsp); // pass pointer to regs array | |
1179 andq(rsp, -16); // align stack as required by ABI | |
1180 call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64))); | |
1181 hlt(); | |
1182 } | |
1183 | |
1184 void MacroAssembler::warn(const char* msg) { | |
1185 push(rbp); | |
1186 movq(rbp, rsp); | |
1187 andq(rsp, -16); // align stack as required by push_CPU_state and call | |
1188 push_CPU_state(); // keeps alignment at 16 bytes | |
1189 lea(c_rarg0, ExternalAddress((address) msg)); | |
1190 call_VM_leaf(CAST_FROM_FN_PTR(address, warning), c_rarg0); | |
1191 pop_CPU_state(); | |
1192 mov(rsp, rbp); | |
1193 pop(rbp); | |
1194 } | |
1195 | |
1196 void MacroAssembler::print_state() { | |
1197 address rip = pc(); | |
1198 pusha(); // get regs on stack | |
1199 push(rbp); | |
1200 movq(rbp, rsp); | |
1201 andq(rsp, -16); // align stack as required by push_CPU_state and call | |
1202 push_CPU_state(); // keeps alignment at 16 bytes | |
1203 | |
1204 lea(c_rarg0, InternalAddress(rip)); | |
1205 lea(c_rarg1, Address(rbp, wordSize)); // pass pointer to regs array | |
1206 call_VM_leaf(CAST_FROM_FN_PTR(address, MacroAssembler::print_state64), c_rarg0, c_rarg1); | |
1207 | |
1208 pop_CPU_state(); | |
1209 mov(rsp, rbp); | |
1210 pop(rbp); | |
1211 popa(); | |
1212 } | |
1213 | |
1214 #ifndef PRODUCT | |
1215 extern "C" void findpc(intptr_t x); | |
1216 #endif | |
1217 | |
1218 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[]) { | |
1219 // In order to get locks to work, we need to fake a in_VM state | |
1220 if (ShowMessageBoxOnError) { | |
1221 JavaThread* thread = JavaThread::current(); | |
1222 JavaThreadState saved_state = thread->thread_state(); | |
1223 thread->set_thread_state(_thread_in_vm); | |
1224 #ifndef PRODUCT | |
1225 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) { | |
1226 ttyLocker ttyl; | |
1227 BytecodeCounter::print(); | |
1228 } | |
1229 #endif | |
1230 // To see where a verify_oop failed, get $ebx+40/X for this frame. | |
1231 // XXX correct this offset for amd64 | |
1232 // This is the value of eip which points to where verify_oop will return. | |
1233 if (os::message_box(msg, "Execution stopped, print registers?")) { | |
1234 print_state64(pc, regs); | |
1235 BREAKPOINT; | |
1236 assert(false, "start up GDB"); | |
1237 } | |
1238 ThreadStateTransition::transition(thread, _thread_in_vm, saved_state); | |
1239 } else { | |
1240 ttyLocker ttyl; | |
1241 ::tty->print_cr("=============== DEBUG MESSAGE: %s ================\n", | |
1242 msg); | |
1243 assert(false, err_msg("DEBUG MESSAGE: %s", msg)); | |
1244 } | |
1245 } | |
1246 | |
1247 void MacroAssembler::print_state64(int64_t pc, int64_t regs[]) { | |
1248 ttyLocker ttyl; | |
1249 FlagSetting fs(Debugging, true); | |
1250 tty->print_cr("rip = 0x%016lx", pc); | |
1251 #ifndef PRODUCT | |
1252 tty->cr(); | |
1253 findpc(pc); | |
1254 tty->cr(); | |
1255 #endif | |
1256 #define PRINT_REG(rax, value) \ | |
1257 { tty->print("%s = ", #rax); os::print_location(tty, value); } | |
1258 PRINT_REG(rax, regs[15]); | |
1259 PRINT_REG(rbx, regs[12]); | |
1260 PRINT_REG(rcx, regs[14]); | |
1261 PRINT_REG(rdx, regs[13]); | |
1262 PRINT_REG(rdi, regs[8]); | |
1263 PRINT_REG(rsi, regs[9]); | |
1264 PRINT_REG(rbp, regs[10]); | |
1265 PRINT_REG(rsp, regs[11]); | |
1266 PRINT_REG(r8 , regs[7]); | |
1267 PRINT_REG(r9 , regs[6]); | |
1268 PRINT_REG(r10, regs[5]); | |
1269 PRINT_REG(r11, regs[4]); | |
1270 PRINT_REG(r12, regs[3]); | |
1271 PRINT_REG(r13, regs[2]); | |
1272 PRINT_REG(r14, regs[1]); | |
1273 PRINT_REG(r15, regs[0]); | |
1274 #undef PRINT_REG | |
1275 // Print some words near top of staack. | |
1276 int64_t* rsp = (int64_t*) regs[11]; | |
1277 int64_t* dump_sp = rsp; | |
1278 for (int col1 = 0; col1 < 8; col1++) { | |
1279 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp); | |
1280 os::print_location(tty, *dump_sp++); | |
1281 } | |
1282 for (int row = 0; row < 25; row++) { | |
1283 tty->print("(rsp+0x%03x) 0x%016lx: ", (int)((intptr_t)dump_sp - (intptr_t)rsp), (int64_t)dump_sp); | |
1284 for (int col = 0; col < 4; col++) { | |
1285 tty->print(" 0x%016lx", *dump_sp++); | |
1286 } | |
1287 tty->cr(); | |
1288 } | |
1289 // Print some instructions around pc: | |
1290 Disassembler::decode((address)pc-64, (address)pc); | |
1291 tty->print_cr("--------"); | |
1292 Disassembler::decode((address)pc, (address)pc+32); | |
1293 } | |
1294 | |
1295 #endif // _LP64 | |
1296 | |
1297 // Now versions that are common to 32/64 bit | |
1298 | |
1299 void MacroAssembler::addptr(Register dst, int32_t imm32) { | |
1300 LP64_ONLY(addq(dst, imm32)) NOT_LP64(addl(dst, imm32)); | |
1301 } | |
1302 | |
1303 void MacroAssembler::addptr(Register dst, Register src) { | |
1304 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); | |
1305 } | |
1306 | |
1307 void MacroAssembler::addptr(Address dst, Register src) { | |
1308 LP64_ONLY(addq(dst, src)) NOT_LP64(addl(dst, src)); | |
1309 } | |
1310 | |
1311 void MacroAssembler::addsd(XMMRegister dst, AddressLiteral src) { | |
1312 if (reachable(src)) { | |
1313 Assembler::addsd(dst, as_Address(src)); | |
1314 } else { | |
1315 lea(rscratch1, src); | |
1316 Assembler::addsd(dst, Address(rscratch1, 0)); | |
1317 } | |
1318 } | |
1319 | |
1320 void MacroAssembler::addss(XMMRegister dst, AddressLiteral src) { | |
1321 if (reachable(src)) { | |
1322 addss(dst, as_Address(src)); | |
1323 } else { | |
1324 lea(rscratch1, src); | |
1325 addss(dst, Address(rscratch1, 0)); | |
1326 } | |
1327 } | |
1328 | |
1329 void MacroAssembler::align(int modulus) { | |
1330 if (offset() % modulus != 0) { | |
1331 nop(modulus - (offset() % modulus)); | |
1332 } | |
1333 } | |
1334 | |
1335 void MacroAssembler::andpd(XMMRegister dst, AddressLiteral src) { | |
1336 // Used in sign-masking with aligned address. | |
1337 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); | |
1338 if (reachable(src)) { | |
1339 Assembler::andpd(dst, as_Address(src)); | |
1340 } else { | |
1341 lea(rscratch1, src); | |
1342 Assembler::andpd(dst, Address(rscratch1, 0)); | |
1343 } | |
1344 } | |
1345 | |
1346 void MacroAssembler::andps(XMMRegister dst, AddressLiteral src) { | |
1347 // Used in sign-masking with aligned address. | |
1348 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); | |
1349 if (reachable(src)) { | |
1350 Assembler::andps(dst, as_Address(src)); | |
1351 } else { | |
1352 lea(rscratch1, src); | |
1353 Assembler::andps(dst, Address(rscratch1, 0)); | |
1354 } | |
1355 } | |
1356 | |
1357 void MacroAssembler::andptr(Register dst, int32_t imm32) { | |
1358 LP64_ONLY(andq(dst, imm32)) NOT_LP64(andl(dst, imm32)); | |
1359 } | |
1360 | |
1361 void MacroAssembler::atomic_incl(AddressLiteral counter_addr) { | |
1362 pushf(); | |
1363 if (os::is_MP()) | |
1364 lock(); | |
1365 incrementl(counter_addr); | |
1366 popf(); | |
1367 } | |
1368 | |
1369 // Writes to stack successive pages until offset reached to check for | |
1370 // stack overflow + shadow pages. This clobbers tmp. | |
1371 void MacroAssembler::bang_stack_size(Register size, Register tmp) { | |
1372 movptr(tmp, rsp); | |
1373 // Bang stack for total size given plus shadow page size. | |
1374 // Bang one page at a time because large size can bang beyond yellow and | |
1375 // red zones. | |
1376 Label loop; | |
1377 bind(loop); | |
1378 movl(Address(tmp, (-os::vm_page_size())), size ); | |
1379 subptr(tmp, os::vm_page_size()); | |
1380 subl(size, os::vm_page_size()); | |
1381 jcc(Assembler::greater, loop); | |
1382 | |
1383 // Bang down shadow pages too. | |
13047
be525e91f65b
8026775: nsk/jvmti/RedefineClasses/StressRedefine crashes due to EXCEPTION_ACCESS_VIOLATION
mikael
parents:
13000
diff
changeset
|
1384 // At this point, (tmp-0) is the last address touched, so don't |
be525e91f65b
8026775: nsk/jvmti/RedefineClasses/StressRedefine crashes due to EXCEPTION_ACCESS_VIOLATION
mikael
parents:
13000
diff
changeset
|
1385 // touch it again. (It was touched as (tmp-pagesize) but then tmp |
be525e91f65b
8026775: nsk/jvmti/RedefineClasses/StressRedefine crashes due to EXCEPTION_ACCESS_VIOLATION
mikael
parents:
13000
diff
changeset
|
1386 // was post-decremented.) Skip this address by starting at i=1, and |
be525e91f65b
8026775: nsk/jvmti/RedefineClasses/StressRedefine crashes due to EXCEPTION_ACCESS_VIOLATION
mikael
parents:
13000
diff
changeset
|
1387 // touch a few more pages below. N.B. It is important to touch all |
be525e91f65b
8026775: nsk/jvmti/RedefineClasses/StressRedefine crashes due to EXCEPTION_ACCESS_VIOLATION
mikael
parents:
13000
diff
changeset
|
1388 // the way down to and including i=StackShadowPages. |
be525e91f65b
8026775: nsk/jvmti/RedefineClasses/StressRedefine crashes due to EXCEPTION_ACCESS_VIOLATION
mikael
parents:
13000
diff
changeset
|
1389 for (int i = 1; i <= StackShadowPages; i++) { |
7199 | 1390 // this could be any sized move but this is can be a debugging crumb |
1391 // so the bigger the better. | |
1392 movptr(Address(tmp, (-i*os::vm_page_size())), size ); | |
1393 } | |
1394 } | |
1395 | |
1396 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) { | |
1397 assert(UseBiasedLocking, "why call this otherwise?"); | |
1398 | |
1399 // Check for biased locking unlock case, which is a no-op | |
1400 // Note: we do not have to check the thread ID for two reasons. | |
1401 // First, the interpreter checks for IllegalMonitorStateException at | |
1402 // a higher level. Second, if the bias was revoked while we held the | |
1403 // lock, the object could not be rebiased toward another thread, so | |
1404 // the bias bit would be clear. | |
1405 movptr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes())); | |
1406 andptr(temp_reg, markOopDesc::biased_lock_mask_in_place); | |
1407 cmpptr(temp_reg, markOopDesc::biased_lock_pattern); | |
1408 jcc(Assembler::equal, done); | |
1409 } | |
1410 | |
1411 void MacroAssembler::c2bool(Register x) { | |
1412 // implements x == 0 ? 0 : 1 | |
1413 // note: must only look at least-significant byte of x | |
1414 // since C-style booleans are stored in one byte | |
1415 // only! (was bug) | |
1416 andl(x, 0xFF); | |
1417 setb(Assembler::notZero, x); | |
1418 } | |
1419 | |
1420 // Wouldn't need if AddressLiteral version had new name | |
1421 void MacroAssembler::call(Label& L, relocInfo::relocType rtype) { | |
1422 Assembler::call(L, rtype); | |
1423 } | |
1424 | |
1425 void MacroAssembler::call(Register entry) { | |
1426 Assembler::call(entry); | |
1427 } | |
1428 | |
1429 void MacroAssembler::call(AddressLiteral entry) { | |
1430 if (reachable(entry)) { | |
1431 Assembler::call_literal(entry.target(), entry.rspec()); | |
1432 } else { | |
1433 lea(rscratch1, entry); | |
1434 Assembler::call(rscratch1); | |
1435 } | |
1436 } | |
1437 | |
1438 void MacroAssembler::ic_call(address entry) { | |
1439 RelocationHolder rh = virtual_call_Relocation::spec(pc()); | |
1440 movptr(rax, (intptr_t)Universe::non_oop_word()); | |
1441 call(AddressLiteral(entry, rh)); | |
1442 } | |
1443 | |
1444 // Implementation of call_VM versions | |
1445 | |
1446 void MacroAssembler::call_VM(Register oop_result, | |
1447 address entry_point, | |
1448 bool check_exceptions) { | |
1449 Label C, E; | |
1450 call(C, relocInfo::none); | |
1451 jmp(E); | |
1452 | |
1453 bind(C); | |
1454 call_VM_helper(oop_result, entry_point, 0, check_exceptions); | |
1455 ret(0); | |
1456 | |
1457 bind(E); | |
1458 } | |
1459 | |
1460 void MacroAssembler::call_VM(Register oop_result, | |
1461 address entry_point, | |
1462 Register arg_1, | |
1463 bool check_exceptions) { | |
1464 Label C, E; | |
1465 call(C, relocInfo::none); | |
1466 jmp(E); | |
1467 | |
1468 bind(C); | |
1469 pass_arg1(this, arg_1); | |
1470 call_VM_helper(oop_result, entry_point, 1, check_exceptions); | |
1471 ret(0); | |
1472 | |
1473 bind(E); | |
1474 } | |
1475 | |
1476 void MacroAssembler::call_VM(Register oop_result, | |
1477 address entry_point, | |
1478 Register arg_1, | |
1479 Register arg_2, | |
1480 bool check_exceptions) { | |
1481 Label C, E; | |
1482 call(C, relocInfo::none); | |
1483 jmp(E); | |
1484 | |
1485 bind(C); | |
1486 | |
1487 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1488 | |
1489 pass_arg2(this, arg_2); | |
1490 pass_arg1(this, arg_1); | |
1491 call_VM_helper(oop_result, entry_point, 2, check_exceptions); | |
1492 ret(0); | |
1493 | |
1494 bind(E); | |
1495 } | |
1496 | |
1497 void MacroAssembler::call_VM(Register oop_result, | |
1498 address entry_point, | |
1499 Register arg_1, | |
1500 Register arg_2, | |
1501 Register arg_3, | |
1502 bool check_exceptions) { | |
1503 Label C, E; | |
1504 call(C, relocInfo::none); | |
1505 jmp(E); | |
1506 | |
1507 bind(C); | |
1508 | |
1509 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); | |
1510 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); | |
1511 pass_arg3(this, arg_3); | |
1512 | |
1513 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1514 pass_arg2(this, arg_2); | |
1515 | |
1516 pass_arg1(this, arg_1); | |
1517 call_VM_helper(oop_result, entry_point, 3, check_exceptions); | |
1518 ret(0); | |
1519 | |
1520 bind(E); | |
1521 } | |
1522 | |
1523 void MacroAssembler::call_VM(Register oop_result, | |
1524 Register last_java_sp, | |
1525 address entry_point, | |
1526 int number_of_arguments, | |
1527 bool check_exceptions) { | |
1528 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); | |
1529 call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); | |
1530 } | |
1531 | |
1532 void MacroAssembler::call_VM(Register oop_result, | |
1533 Register last_java_sp, | |
1534 address entry_point, | |
1535 Register arg_1, | |
1536 bool check_exceptions) { | |
1537 pass_arg1(this, arg_1); | |
1538 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); | |
1539 } | |
1540 | |
1541 void MacroAssembler::call_VM(Register oop_result, | |
1542 Register last_java_sp, | |
1543 address entry_point, | |
1544 Register arg_1, | |
1545 Register arg_2, | |
1546 bool check_exceptions) { | |
1547 | |
1548 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1549 pass_arg2(this, arg_2); | |
1550 pass_arg1(this, arg_1); | |
1551 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); | |
1552 } | |
1553 | |
1554 void MacroAssembler::call_VM(Register oop_result, | |
1555 Register last_java_sp, | |
1556 address entry_point, | |
1557 Register arg_1, | |
1558 Register arg_2, | |
1559 Register arg_3, | |
1560 bool check_exceptions) { | |
1561 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); | |
1562 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); | |
1563 pass_arg3(this, arg_3); | |
1564 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1565 pass_arg2(this, arg_2); | |
1566 pass_arg1(this, arg_1); | |
1567 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); | |
1568 } | |
1569 | |
1570 void MacroAssembler::super_call_VM(Register oop_result, | |
1571 Register last_java_sp, | |
1572 address entry_point, | |
1573 int number_of_arguments, | |
1574 bool check_exceptions) { | |
1575 Register thread = LP64_ONLY(r15_thread) NOT_LP64(noreg); | |
1576 MacroAssembler::call_VM_base(oop_result, thread, last_java_sp, entry_point, number_of_arguments, check_exceptions); | |
1577 } | |
1578 | |
1579 void MacroAssembler::super_call_VM(Register oop_result, | |
1580 Register last_java_sp, | |
1581 address entry_point, | |
1582 Register arg_1, | |
1583 bool check_exceptions) { | |
1584 pass_arg1(this, arg_1); | |
1585 super_call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions); | |
1586 } | |
1587 | |
1588 void MacroAssembler::super_call_VM(Register oop_result, | |
1589 Register last_java_sp, | |
1590 address entry_point, | |
1591 Register arg_1, | |
1592 Register arg_2, | |
1593 bool check_exceptions) { | |
1594 | |
1595 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1596 pass_arg2(this, arg_2); | |
1597 pass_arg1(this, arg_1); | |
1598 super_call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions); | |
1599 } | |
1600 | |
1601 void MacroAssembler::super_call_VM(Register oop_result, | |
1602 Register last_java_sp, | |
1603 address entry_point, | |
1604 Register arg_1, | |
1605 Register arg_2, | |
1606 Register arg_3, | |
1607 bool check_exceptions) { | |
1608 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); | |
1609 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); | |
1610 pass_arg3(this, arg_3); | |
1611 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1612 pass_arg2(this, arg_2); | |
1613 pass_arg1(this, arg_1); | |
1614 super_call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions); | |
1615 } | |
1616 | |
1617 void MacroAssembler::call_VM_base(Register oop_result, | |
1618 Register java_thread, | |
1619 Register last_java_sp, | |
1620 address entry_point, | |
1621 int number_of_arguments, | |
1622 bool check_exceptions) { | |
1623 // determine java_thread register | |
1624 if (!java_thread->is_valid()) { | |
1625 #ifdef _LP64 | |
1626 java_thread = r15_thread; | |
1627 #else | |
1628 java_thread = rdi; | |
1629 get_thread(java_thread); | |
1630 #endif // LP64 | |
1631 } | |
1632 // determine last_java_sp register | |
1633 if (!last_java_sp->is_valid()) { | |
1634 last_java_sp = rsp; | |
1635 } | |
1636 // debugging support | |
1637 assert(number_of_arguments >= 0 , "cannot have negative number of arguments"); | |
1638 LP64_ONLY(assert(java_thread == r15_thread, "unexpected register")); | |
1639 #ifdef ASSERT | |
1640 // TraceBytecodes does not use r12 but saves it over the call, so don't verify | |
1641 // r12 is the heapbase. | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
1642 LP64_ONLY(if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");) |
7199 | 1643 #endif // ASSERT |
1644 | |
1645 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result"); | |
1646 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp"); | |
1647 | |
1648 // push java thread (becomes first argument of C function) | |
1649 | |
1650 NOT_LP64(push(java_thread); number_of_arguments++); | |
1651 LP64_ONLY(mov(c_rarg0, r15_thread)); | |
1652 | |
1653 // set last Java frame before call | |
1654 assert(last_java_sp != rbp, "can't use ebp/rbp"); | |
1655 | |
1656 // Only interpreter should have to set fp | |
1657 set_last_Java_frame(java_thread, last_java_sp, rbp, NULL); | |
1658 | |
1659 // do the call, remove parameters | |
1660 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments); | |
1661 | |
1662 // restore the thread (cannot use the pushed argument since arguments | |
1663 // may be overwritten by C code generated by an optimizing compiler); | |
1664 // however can use the register value directly if it is callee saved. | |
1665 if (LP64_ONLY(true ||) java_thread == rdi || java_thread == rsi) { | |
1666 // rdi & rsi (also r15) are callee saved -> nothing to do | |
1667 #ifdef ASSERT | |
1668 guarantee(java_thread != rax, "change this code"); | |
1669 push(rax); | |
1670 { Label L; | |
1671 get_thread(rax); | |
1672 cmpptr(java_thread, rax); | |
1673 jcc(Assembler::equal, L); | |
1674 STOP("MacroAssembler::call_VM_base: rdi not callee saved?"); | |
1675 bind(L); | |
1676 } | |
1677 pop(rax); | |
1678 #endif | |
1679 } else { | |
1680 get_thread(java_thread); | |
1681 } | |
1682 // reset last Java frame | |
1683 // Only interpreter should have to clear fp | |
1684 reset_last_Java_frame(java_thread, true, false); | |
1685 | |
1686 #ifndef CC_INTERP | |
1687 // C++ interp handles this in the interpreter | |
1688 check_and_handle_popframe(java_thread); | |
1689 check_and_handle_earlyret(java_thread); | |
1690 #endif /* CC_INTERP */ | |
1691 | |
1692 if (check_exceptions) { | |
1693 // check for pending exceptions (java_thread is set upon return) | |
1694 cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t) NULL_WORD); | |
1695 #ifndef _LP64 | |
1696 jump_cc(Assembler::notEqual, | |
1697 RuntimeAddress(StubRoutines::forward_exception_entry())); | |
1698 #else | |
1699 // This used to conditionally jump to forward_exception however it is | |
1700 // possible if we relocate that the branch will not reach. So we must jump | |
1701 // around so we can always reach | |
1702 | |
1703 Label ok; | |
1704 jcc(Assembler::equal, ok); | |
1705 jump(RuntimeAddress(StubRoutines::forward_exception_entry())); | |
1706 bind(ok); | |
1707 #endif // LP64 | |
1708 } | |
1709 | |
1710 // get oop result if there is one and reset the value in the thread | |
1711 if (oop_result->is_valid()) { | |
1712 get_vm_result(oop_result, java_thread); | |
1713 } | |
1714 } | |
1715 | |
1716 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) { | |
1717 | |
1718 // Calculate the value for last_Java_sp | |
1719 // somewhat subtle. call_VM does an intermediate call | |
1720 // which places a return address on the stack just under the | |
1721 // stack pointer as the user finsihed with it. This allows | |
1722 // use to retrieve last_Java_pc from last_Java_sp[-1]. | |
1723 // On 32bit we then have to push additional args on the stack to accomplish | |
1724 // the actual requested call. On 64bit call_VM only can use register args | |
1725 // so the only extra space is the return address that call_VM created. | |
1726 // This hopefully explains the calculations here. | |
1727 | |
1728 #ifdef _LP64 | |
1729 // We've pushed one address, correct last_Java_sp | |
1730 lea(rax, Address(rsp, wordSize)); | |
1731 #else | |
1732 lea(rax, Address(rsp, (1 + number_of_arguments) * wordSize)); | |
1733 #endif // LP64 | |
1734 | |
1735 call_VM_base(oop_result, noreg, rax, entry_point, number_of_arguments, check_exceptions); | |
1736 | |
1737 } | |
1738 | |
1739 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) { | |
1740 call_VM_leaf_base(entry_point, number_of_arguments); | |
1741 } | |
1742 | |
1743 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) { | |
1744 pass_arg0(this, arg_0); | |
1745 call_VM_leaf(entry_point, 1); | |
1746 } | |
1747 | |
1748 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { | |
1749 | |
1750 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); | |
1751 pass_arg1(this, arg_1); | |
1752 pass_arg0(this, arg_0); | |
1753 call_VM_leaf(entry_point, 2); | |
1754 } | |
1755 | |
1756 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { | |
1757 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); | |
1758 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1759 pass_arg2(this, arg_2); | |
1760 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); | |
1761 pass_arg1(this, arg_1); | |
1762 pass_arg0(this, arg_0); | |
1763 call_VM_leaf(entry_point, 3); | |
1764 } | |
1765 | |
1766 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) { | |
1767 pass_arg0(this, arg_0); | |
1768 MacroAssembler::call_VM_leaf_base(entry_point, 1); | |
1769 } | |
1770 | |
1771 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) { | |
1772 | |
1773 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); | |
1774 pass_arg1(this, arg_1); | |
1775 pass_arg0(this, arg_0); | |
1776 MacroAssembler::call_VM_leaf_base(entry_point, 2); | |
1777 } | |
1778 | |
1779 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) { | |
1780 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); | |
1781 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1782 pass_arg2(this, arg_2); | |
1783 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); | |
1784 pass_arg1(this, arg_1); | |
1785 pass_arg0(this, arg_0); | |
1786 MacroAssembler::call_VM_leaf_base(entry_point, 3); | |
1787 } | |
1788 | |
1789 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) { | |
1790 LP64_ONLY(assert(arg_0 != c_rarg3, "smashed arg")); | |
1791 LP64_ONLY(assert(arg_1 != c_rarg3, "smashed arg")); | |
1792 LP64_ONLY(assert(arg_2 != c_rarg3, "smashed arg")); | |
1793 pass_arg3(this, arg_3); | |
1794 LP64_ONLY(assert(arg_0 != c_rarg2, "smashed arg")); | |
1795 LP64_ONLY(assert(arg_1 != c_rarg2, "smashed arg")); | |
1796 pass_arg2(this, arg_2); | |
1797 LP64_ONLY(assert(arg_0 != c_rarg1, "smashed arg")); | |
1798 pass_arg1(this, arg_1); | |
1799 pass_arg0(this, arg_0); | |
1800 MacroAssembler::call_VM_leaf_base(entry_point, 4); | |
1801 } | |
1802 | |
1803 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) { | |
1804 movptr(oop_result, Address(java_thread, JavaThread::vm_result_offset())); | |
1805 movptr(Address(java_thread, JavaThread::vm_result_offset()), NULL_WORD); | |
1806 verify_oop(oop_result, "broken oop in call_VM_base"); | |
1807 } | |
1808 | |
1809 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) { | |
1810 movptr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset())); | |
1811 movptr(Address(java_thread, JavaThread::vm_result_2_offset()), NULL_WORD); | |
1812 } | |
1813 | |
1814 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { | |
1815 } | |
1816 | |
1817 void MacroAssembler::check_and_handle_popframe(Register java_thread) { | |
1818 } | |
1819 | |
1820 void MacroAssembler::cmp32(AddressLiteral src1, int32_t imm) { | |
1821 if (reachable(src1)) { | |
1822 cmpl(as_Address(src1), imm); | |
1823 } else { | |
1824 lea(rscratch1, src1); | |
1825 cmpl(Address(rscratch1, 0), imm); | |
1826 } | |
1827 } | |
1828 | |
1829 void MacroAssembler::cmp32(Register src1, AddressLiteral src2) { | |
1830 assert(!src2.is_lval(), "use cmpptr"); | |
1831 if (reachable(src2)) { | |
1832 cmpl(src1, as_Address(src2)); | |
1833 } else { | |
1834 lea(rscratch1, src2); | |
1835 cmpl(src1, Address(rscratch1, 0)); | |
1836 } | |
1837 } | |
1838 | |
1839 void MacroAssembler::cmp32(Register src1, int32_t imm) { | |
1840 Assembler::cmpl(src1, imm); | |
1841 } | |
1842 | |
1843 void MacroAssembler::cmp32(Register src1, Address src2) { | |
1844 Assembler::cmpl(src1, src2); | |
1845 } | |
1846 | |
1847 void MacroAssembler::cmpsd2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { | |
1848 ucomisd(opr1, opr2); | |
1849 | |
1850 Label L; | |
1851 if (unordered_is_less) { | |
1852 movl(dst, -1); | |
1853 jcc(Assembler::parity, L); | |
1854 jcc(Assembler::below , L); | |
1855 movl(dst, 0); | |
1856 jcc(Assembler::equal , L); | |
1857 increment(dst); | |
1858 } else { // unordered is greater | |
1859 movl(dst, 1); | |
1860 jcc(Assembler::parity, L); | |
1861 jcc(Assembler::above , L); | |
1862 movl(dst, 0); | |
1863 jcc(Assembler::equal , L); | |
1864 decrementl(dst); | |
1865 } | |
1866 bind(L); | |
1867 } | |
1868 | |
1869 void MacroAssembler::cmpss2int(XMMRegister opr1, XMMRegister opr2, Register dst, bool unordered_is_less) { | |
1870 ucomiss(opr1, opr2); | |
1871 | |
1872 Label L; | |
1873 if (unordered_is_less) { | |
1874 movl(dst, -1); | |
1875 jcc(Assembler::parity, L); | |
1876 jcc(Assembler::below , L); | |
1877 movl(dst, 0); | |
1878 jcc(Assembler::equal , L); | |
1879 increment(dst); | |
1880 } else { // unordered is greater | |
1881 movl(dst, 1); | |
1882 jcc(Assembler::parity, L); | |
1883 jcc(Assembler::above , L); | |
1884 movl(dst, 0); | |
1885 jcc(Assembler::equal , L); | |
1886 decrementl(dst); | |
1887 } | |
1888 bind(L); | |
1889 } | |
1890 | |
1891 | |
1892 void MacroAssembler::cmp8(AddressLiteral src1, int imm) { | |
1893 if (reachable(src1)) { | |
1894 cmpb(as_Address(src1), imm); | |
1895 } else { | |
1896 lea(rscratch1, src1); | |
1897 cmpb(Address(rscratch1, 0), imm); | |
1898 } | |
1899 } | |
1900 | |
1901 void MacroAssembler::cmpptr(Register src1, AddressLiteral src2) { | |
1902 #ifdef _LP64 | |
1903 if (src2.is_lval()) { | |
1904 movptr(rscratch1, src2); | |
1905 Assembler::cmpq(src1, rscratch1); | |
1906 } else if (reachable(src2)) { | |
1907 cmpq(src1, as_Address(src2)); | |
1908 } else { | |
1909 lea(rscratch1, src2); | |
1910 Assembler::cmpq(src1, Address(rscratch1, 0)); | |
1911 } | |
1912 #else | |
1913 if (src2.is_lval()) { | |
1914 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); | |
1915 } else { | |
1916 cmpl(src1, as_Address(src2)); | |
1917 } | |
1918 #endif // _LP64 | |
1919 } | |
1920 | |
1921 void MacroAssembler::cmpptr(Address src1, AddressLiteral src2) { | |
1922 assert(src2.is_lval(), "not a mem-mem compare"); | |
1923 #ifdef _LP64 | |
1924 // moves src2's literal address | |
1925 movptr(rscratch1, src2); | |
1926 Assembler::cmpq(src1, rscratch1); | |
1927 #else | |
1928 cmp_literal32(src1, (int32_t) src2.target(), src2.rspec()); | |
1929 #endif // _LP64 | |
1930 } | |
1931 | |
1932 void MacroAssembler::locked_cmpxchgptr(Register reg, AddressLiteral adr) { | |
1933 if (reachable(adr)) { | |
1934 if (os::is_MP()) | |
1935 lock(); | |
1936 cmpxchgptr(reg, as_Address(adr)); | |
1937 } else { | |
1938 lea(rscratch1, adr); | |
1939 if (os::is_MP()) | |
1940 lock(); | |
1941 cmpxchgptr(reg, Address(rscratch1, 0)); | |
1942 } | |
1943 } | |
1944 | |
1945 void MacroAssembler::cmpxchgptr(Register reg, Address adr) { | |
1946 LP64_ONLY(cmpxchgq(reg, adr)) NOT_LP64(cmpxchgl(reg, adr)); | |
1947 } | |
1948 | |
1949 void MacroAssembler::comisd(XMMRegister dst, AddressLiteral src) { | |
1950 if (reachable(src)) { | |
1951 Assembler::comisd(dst, as_Address(src)); | |
1952 } else { | |
1953 lea(rscratch1, src); | |
1954 Assembler::comisd(dst, Address(rscratch1, 0)); | |
1955 } | |
1956 } | |
1957 | |
1958 void MacroAssembler::comiss(XMMRegister dst, AddressLiteral src) { | |
1959 if (reachable(src)) { | |
1960 Assembler::comiss(dst, as_Address(src)); | |
1961 } else { | |
1962 lea(rscratch1, src); | |
1963 Assembler::comiss(dst, Address(rscratch1, 0)); | |
1964 } | |
1965 } | |
1966 | |
1967 | |
1968 void MacroAssembler::cond_inc32(Condition cond, AddressLiteral counter_addr) { | |
1969 Condition negated_cond = negate_condition(cond); | |
1970 Label L; | |
1971 jcc(negated_cond, L); | |
1972 atomic_incl(counter_addr); | |
1973 bind(L); | |
1974 } | |
1975 | |
1976 int MacroAssembler::corrected_idivl(Register reg) { | |
1977 // Full implementation of Java idiv and irem; checks for | |
1978 // special case as described in JVM spec., p.243 & p.271. | |
1979 // The function returns the (pc) offset of the idivl | |
1980 // instruction - may be needed for implicit exceptions. | |
1981 // | |
1982 // normal case special case | |
1983 // | |
1984 // input : rax,: dividend min_int | |
1985 // reg: divisor (may not be rax,/rdx) -1 | |
1986 // | |
1987 // output: rax,: quotient (= rax, idiv reg) min_int | |
1988 // rdx: remainder (= rax, irem reg) 0 | |
1989 assert(reg != rax && reg != rdx, "reg cannot be rax, or rdx register"); | |
1990 const int min_int = 0x80000000; | |
1991 Label normal_case, special_case; | |
1992 | |
1993 // check for special case | |
1994 cmpl(rax, min_int); | |
1995 jcc(Assembler::notEqual, normal_case); | |
1996 xorl(rdx, rdx); // prepare rdx for possible special case (where remainder = 0) | |
1997 cmpl(reg, -1); | |
1998 jcc(Assembler::equal, special_case); | |
1999 | |
2000 // handle normal case | |
2001 bind(normal_case); | |
2002 cdql(); | |
2003 int idivl_offset = offset(); | |
2004 idivl(reg); | |
2005 | |
2006 // normal and special case exit | |
2007 bind(special_case); | |
2008 | |
2009 return idivl_offset; | |
2010 } | |
2011 | |
2012 | |
2013 | |
2014 void MacroAssembler::decrementl(Register reg, int value) { | |
2015 if (value == min_jint) {subl(reg, value) ; return; } | |
2016 if (value < 0) { incrementl(reg, -value); return; } | |
2017 if (value == 0) { ; return; } | |
2018 if (value == 1 && UseIncDec) { decl(reg) ; return; } | |
2019 /* else */ { subl(reg, value) ; return; } | |
2020 } | |
2021 | |
2022 void MacroAssembler::decrementl(Address dst, int value) { | |
2023 if (value == min_jint) {subl(dst, value) ; return; } | |
2024 if (value < 0) { incrementl(dst, -value); return; } | |
2025 if (value == 0) { ; return; } | |
2026 if (value == 1 && UseIncDec) { decl(dst) ; return; } | |
2027 /* else */ { subl(dst, value) ; return; } | |
2028 } | |
2029 | |
2030 void MacroAssembler::division_with_shift (Register reg, int shift_value) { | |
2031 assert (shift_value > 0, "illegal shift value"); | |
2032 Label _is_positive; | |
2033 testl (reg, reg); | |
2034 jcc (Assembler::positive, _is_positive); | |
2035 int offset = (1 << shift_value) - 1 ; | |
2036 | |
2037 if (offset == 1) { | |
2038 incrementl(reg); | |
2039 } else { | |
2040 addl(reg, offset); | |
2041 } | |
2042 | |
2043 bind (_is_positive); | |
2044 sarl(reg, shift_value); | |
2045 } | |
2046 | |
2047 void MacroAssembler::divsd(XMMRegister dst, AddressLiteral src) { | |
2048 if (reachable(src)) { | |
2049 Assembler::divsd(dst, as_Address(src)); | |
2050 } else { | |
2051 lea(rscratch1, src); | |
2052 Assembler::divsd(dst, Address(rscratch1, 0)); | |
2053 } | |
2054 } | |
2055 | |
2056 void MacroAssembler::divss(XMMRegister dst, AddressLiteral src) { | |
2057 if (reachable(src)) { | |
2058 Assembler::divss(dst, as_Address(src)); | |
2059 } else { | |
2060 lea(rscratch1, src); | |
2061 Assembler::divss(dst, Address(rscratch1, 0)); | |
2062 } | |
2063 } | |
2064 | |
2065 // !defined(COMPILER2) is because of stupid core builds | |
8265
ded5288f5b96
fix windows build of the server and client vms
Andreas Woess <andreas.woess@jku.at>
parents:
8042
diff
changeset
|
2066 #if !defined(_LP64) || defined(COMPILER1) || !defined(COMPILER2) || defined(GRAAL) |
7199 | 2067 void MacroAssembler::empty_FPU_stack() { |
2068 if (VM_Version::supports_mmx()) { | |
2069 emms(); | |
2070 } else { | |
2071 for (int i = 8; i-- > 0; ) ffree(i); | |
2072 } | |
2073 } | |
2074 #endif // !LP64 || C1 || !C2 | |
2075 | |
2076 | |
2077 // Defines obj, preserves var_size_in_bytes | |
2078 void MacroAssembler::eden_allocate(Register obj, | |
2079 Register var_size_in_bytes, | |
2080 int con_size_in_bytes, | |
2081 Register t1, | |
2082 Label& slow_case) { | |
2083 assert(obj == rax, "obj must be in rax, for cmpxchg"); | |
2084 assert_different_registers(obj, var_size_in_bytes, t1); | |
2085 if (CMSIncrementalMode || !Universe::heap()->supports_inline_contig_alloc()) { | |
2086 jmp(slow_case); | |
2087 } else { | |
2088 Register end = t1; | |
2089 Label retry; | |
2090 bind(retry); | |
2091 ExternalAddress heap_top((address) Universe::heap()->top_addr()); | |
2092 movptr(obj, heap_top); | |
2093 if (var_size_in_bytes == noreg) { | |
2094 lea(end, Address(obj, con_size_in_bytes)); | |
2095 } else { | |
2096 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); | |
2097 } | |
2098 // if end < obj then we wrapped around => object too long => slow case | |
2099 cmpptr(end, obj); | |
2100 jcc(Assembler::below, slow_case); | |
2101 cmpptr(end, ExternalAddress((address) Universe::heap()->end_addr())); | |
2102 jcc(Assembler::above, slow_case); | |
2103 // Compare obj with the top addr, and if still equal, store the new top addr in | |
2104 // end at the address of the top addr pointer. Sets ZF if was equal, and clears | |
2105 // it otherwise. Use lock prefix for atomicity on MPs. | |
2106 locked_cmpxchgptr(end, heap_top); | |
2107 jcc(Assembler::notEqual, retry); | |
2108 } | |
2109 } | |
2110 | |
2111 void MacroAssembler::enter() { | |
2112 push(rbp); | |
2113 mov(rbp, rsp); | |
2114 } | |
2115 | |
2116 // A 5 byte nop that is safe for patching (see patch_verified_entry) | |
2117 void MacroAssembler::fat_nop() { | |
2118 if (UseAddressNop) { | |
2119 addr_nop_5(); | |
2120 } else { | |
7430
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2121 emit_int8(0x26); // es: |
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2122 emit_int8(0x2e); // cs: |
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2123 emit_int8(0x64); // fs: |
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2124 emit_int8(0x65); // gs: |
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2125 emit_int8((unsigned char)0x90); |
7199 | 2126 } |
2127 } | |
2128 | |
2129 void MacroAssembler::fcmp(Register tmp) { | |
2130 fcmp(tmp, 1, true, true); | |
2131 } | |
2132 | |
2133 void MacroAssembler::fcmp(Register tmp, int index, bool pop_left, bool pop_right) { | |
2134 assert(!pop_right || pop_left, "usage error"); | |
2135 if (VM_Version::supports_cmov()) { | |
2136 assert(tmp == noreg, "unneeded temp"); | |
2137 if (pop_left) { | |
2138 fucomip(index); | |
2139 } else { | |
2140 fucomi(index); | |
2141 } | |
2142 if (pop_right) { | |
2143 fpop(); | |
2144 } | |
2145 } else { | |
2146 assert(tmp != noreg, "need temp"); | |
2147 if (pop_left) { | |
2148 if (pop_right) { | |
2149 fcompp(); | |
2150 } else { | |
2151 fcomp(index); | |
2152 } | |
2153 } else { | |
2154 fcom(index); | |
2155 } | |
2156 // convert FPU condition into eflags condition via rax, | |
2157 save_rax(tmp); | |
2158 fwait(); fnstsw_ax(); | |
2159 sahf(); | |
2160 restore_rax(tmp); | |
2161 } | |
2162 // condition codes set as follows: | |
2163 // | |
2164 // CF (corresponds to C0) if x < y | |
2165 // PF (corresponds to C2) if unordered | |
2166 // ZF (corresponds to C3) if x = y | |
2167 } | |
2168 | |
2169 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less) { | |
2170 fcmp2int(dst, unordered_is_less, 1, true, true); | |
2171 } | |
2172 | |
2173 void MacroAssembler::fcmp2int(Register dst, bool unordered_is_less, int index, bool pop_left, bool pop_right) { | |
2174 fcmp(VM_Version::supports_cmov() ? noreg : dst, index, pop_left, pop_right); | |
2175 Label L; | |
2176 if (unordered_is_less) { | |
2177 movl(dst, -1); | |
2178 jcc(Assembler::parity, L); | |
2179 jcc(Assembler::below , L); | |
2180 movl(dst, 0); | |
2181 jcc(Assembler::equal , L); | |
2182 increment(dst); | |
2183 } else { // unordered is greater | |
2184 movl(dst, 1); | |
2185 jcc(Assembler::parity, L); | |
2186 jcc(Assembler::above , L); | |
2187 movl(dst, 0); | |
2188 jcc(Assembler::equal , L); | |
2189 decrementl(dst); | |
2190 } | |
2191 bind(L); | |
2192 } | |
2193 | |
2194 void MacroAssembler::fld_d(AddressLiteral src) { | |
2195 fld_d(as_Address(src)); | |
2196 } | |
2197 | |
2198 void MacroAssembler::fld_s(AddressLiteral src) { | |
2199 fld_s(as_Address(src)); | |
2200 } | |
2201 | |
2202 void MacroAssembler::fld_x(AddressLiteral src) { | |
2203 Assembler::fld_x(as_Address(src)); | |
2204 } | |
2205 | |
2206 void MacroAssembler::fldcw(AddressLiteral src) { | |
2207 Assembler::fldcw(as_Address(src)); | |
2208 } | |
2209 | |
2210 void MacroAssembler::pow_exp_core_encoding() { | |
2211 // kills rax, rcx, rdx | |
2212 subptr(rsp,sizeof(jdouble)); | |
2213 // computes 2^X. Stack: X ... | |
2214 // f2xm1 computes 2^X-1 but only operates on -1<=X<=1. Get int(X) and | |
2215 // keep it on the thread's stack to compute 2^int(X) later | |
2216 // then compute 2^(X-int(X)) as (2^(X-int(X)-1+1) | |
2217 // final result is obtained with: 2^X = 2^int(X) * 2^(X-int(X)) | |
2218 fld_s(0); // Stack: X X ... | |
2219 frndint(); // Stack: int(X) X ... | |
2220 fsuba(1); // Stack: int(X) X-int(X) ... | |
2221 fistp_s(Address(rsp,0)); // move int(X) as integer to thread's stack. Stack: X-int(X) ... | |
2222 f2xm1(); // Stack: 2^(X-int(X))-1 ... | |
2223 fld1(); // Stack: 1 2^(X-int(X))-1 ... | |
2224 faddp(1); // Stack: 2^(X-int(X)) | |
2225 // computes 2^(int(X)): add exponent bias (1023) to int(X), then | |
2226 // shift int(X)+1023 to exponent position. | |
2227 // Exponent is limited to 11 bits if int(X)+1023 does not fit in 11 | |
2228 // bits, set result to NaN. 0x000 and 0x7FF are reserved exponent | |
2229 // values so detect them and set result to NaN. | |
2230 movl(rax,Address(rsp,0)); | |
2231 movl(rcx, -2048); // 11 bit mask and valid NaN binary encoding | |
2232 addl(rax, 1023); | |
2233 movl(rdx,rax); | |
2234 shll(rax,20); | |
2235 // Check that 0 < int(X)+1023 < 2047. Otherwise set rax to NaN. | |
2236 addl(rdx,1); | |
2237 // Check that 1 < int(X)+1023+1 < 2048 | |
2238 // in 3 steps: | |
2239 // 1- (int(X)+1023+1)&-2048 == 0 => 0 <= int(X)+1023+1 < 2048 | |
2240 // 2- (int(X)+1023+1)&-2048 != 0 | |
2241 // 3- (int(X)+1023+1)&-2048 != 1 | |
2242 // Do 2- first because addl just updated the flags. | |
2243 cmov32(Assembler::equal,rax,rcx); | |
2244 cmpl(rdx,1); | |
2245 cmov32(Assembler::equal,rax,rcx); | |
2246 testl(rdx,rcx); | |
2247 cmov32(Assembler::notEqual,rax,rcx); | |
2248 movl(Address(rsp,4),rax); | |
2249 movl(Address(rsp,0),0); | |
2250 fmul_d(Address(rsp,0)); // Stack: 2^X ... | |
2251 addptr(rsp,sizeof(jdouble)); | |
2252 } | |
2253 | |
2254 void MacroAssembler::increase_precision() { | |
2255 subptr(rsp, BytesPerWord); | |
2256 fnstcw(Address(rsp, 0)); | |
2257 movl(rax, Address(rsp, 0)); | |
2258 orl(rax, 0x300); | |
2259 push(rax); | |
2260 fldcw(Address(rsp, 0)); | |
2261 pop(rax); | |
2262 } | |
2263 | |
2264 void MacroAssembler::restore_precision() { | |
2265 fldcw(Address(rsp, 0)); | |
2266 addptr(rsp, BytesPerWord); | |
2267 } | |
2268 | |
2269 void MacroAssembler::fast_pow() { | |
2270 // computes X^Y = 2^(Y * log2(X)) | |
2271 // if fast computation is not possible, result is NaN. Requires | |
2272 // fallback from user of this macro. | |
2273 // increase precision for intermediate steps of the computation | |
2274 increase_precision(); | |
2275 fyl2x(); // Stack: (Y*log2(X)) ... | |
2276 pow_exp_core_encoding(); // Stack: exp(X) ... | |
2277 restore_precision(); | |
2278 } | |
2279 | |
2280 void MacroAssembler::fast_exp() { | |
2281 // computes exp(X) = 2^(X * log2(e)) | |
2282 // if fast computation is not possible, result is NaN. Requires | |
2283 // fallback from user of this macro. | |
2284 // increase precision for intermediate steps of the computation | |
2285 increase_precision(); | |
2286 fldl2e(); // Stack: log2(e) X ... | |
2287 fmulp(1); // Stack: (X*log2(e)) ... | |
2288 pow_exp_core_encoding(); // Stack: exp(X) ... | |
2289 restore_precision(); | |
2290 } | |
2291 | |
2292 void MacroAssembler::pow_or_exp(bool is_exp, int num_fpu_regs_in_use) { | |
2293 // kills rax, rcx, rdx | |
2294 // pow and exp needs 2 extra registers on the fpu stack. | |
2295 Label slow_case, done; | |
2296 Register tmp = noreg; | |
2297 if (!VM_Version::supports_cmov()) { | |
2298 // fcmp needs a temporary so preserve rdx, | |
2299 tmp = rdx; | |
2300 } | |
2301 Register tmp2 = rax; | |
2302 Register tmp3 = rcx; | |
2303 | |
2304 if (is_exp) { | |
2305 // Stack: X | |
2306 fld_s(0); // duplicate argument for runtime call. Stack: X X | |
2307 fast_exp(); // Stack: exp(X) X | |
2308 fcmp(tmp, 0, false, false); // Stack: exp(X) X | |
2309 // exp(X) not equal to itself: exp(X) is NaN go to slow case. | |
2310 jcc(Assembler::parity, slow_case); | |
2311 // get rid of duplicate argument. Stack: exp(X) | |
2312 if (num_fpu_regs_in_use > 0) { | |
2313 fxch(); | |
2314 fpop(); | |
2315 } else { | |
2316 ffree(1); | |
2317 } | |
2318 jmp(done); | |
2319 } else { | |
2320 // Stack: X Y | |
2321 Label x_negative, y_odd; | |
2322 | |
2323 fldz(); // Stack: 0 X Y | |
2324 fcmp(tmp, 1, true, false); // Stack: X Y | |
2325 jcc(Assembler::above, x_negative); | |
2326 | |
2327 // X >= 0 | |
2328 | |
2329 fld_s(1); // duplicate arguments for runtime call. Stack: Y X Y | |
2330 fld_s(1); // Stack: X Y X Y | |
2331 fast_pow(); // Stack: X^Y X Y | |
2332 fcmp(tmp, 0, false, false); // Stack: X^Y X Y | |
2333 // X^Y not equal to itself: X^Y is NaN go to slow case. | |
2334 jcc(Assembler::parity, slow_case); | |
2335 // get rid of duplicate arguments. Stack: X^Y | |
2336 if (num_fpu_regs_in_use > 0) { | |
2337 fxch(); fpop(); | |
2338 fxch(); fpop(); | |
2339 } else { | |
2340 ffree(2); | |
2341 ffree(1); | |
2342 } | |
2343 jmp(done); | |
2344 | |
2345 // X <= 0 | |
2346 bind(x_negative); | |
2347 | |
2348 fld_s(1); // Stack: Y X Y | |
2349 frndint(); // Stack: int(Y) X Y | |
2350 fcmp(tmp, 2, false, false); // Stack: int(Y) X Y | |
2351 jcc(Assembler::notEqual, slow_case); | |
2352 | |
2353 subptr(rsp, 8); | |
2354 | |
2355 // For X^Y, when X < 0, Y has to be an integer and the final | |
2356 // result depends on whether it's odd or even. We just checked | |
2357 // that int(Y) == Y. We move int(Y) to gp registers as a 64 bit | |
2358 // integer to test its parity. If int(Y) is huge and doesn't fit | |
2359 // in the 64 bit integer range, the integer indefinite value will | |
2360 // end up in the gp registers. Huge numbers are all even, the | |
2361 // integer indefinite number is even so it's fine. | |
2362 | |
2363 #ifdef ASSERT | |
2364 // Let's check we don't end up with an integer indefinite number | |
2365 // when not expected. First test for huge numbers: check whether | |
2366 // int(Y)+1 == int(Y) which is true for very large numbers and | |
2367 // those are all even. A 64 bit integer is guaranteed to not | |
2368 // overflow for numbers where y+1 != y (when precision is set to | |
2369 // double precision). | |
2370 Label y_not_huge; | |
2371 | |
2372 fld1(); // Stack: 1 int(Y) X Y | |
2373 fadd(1); // Stack: 1+int(Y) int(Y) X Y | |
2374 | |
2375 #ifdef _LP64 | |
2376 // trip to memory to force the precision down from double extended | |
2377 // precision | |
2378 fstp_d(Address(rsp, 0)); | |
2379 fld_d(Address(rsp, 0)); | |
2380 #endif | |
2381 | |
2382 fcmp(tmp, 1, true, false); // Stack: int(Y) X Y | |
2383 #endif | |
2384 | |
2385 // move int(Y) as 64 bit integer to thread's stack | |
2386 fistp_d(Address(rsp,0)); // Stack: X Y | |
2387 | |
2388 #ifdef ASSERT | |
2389 jcc(Assembler::notEqual, y_not_huge); | |
2390 | |
2391 // Y is huge so we know it's even. It may not fit in a 64 bit | |
2392 // integer and we don't want the debug code below to see the | |
2393 // integer indefinite value so overwrite int(Y) on the thread's | |
2394 // stack with 0. | |
2395 movl(Address(rsp, 0), 0); | |
2396 movl(Address(rsp, 4), 0); | |
2397 | |
2398 bind(y_not_huge); | |
2399 #endif | |
2400 | |
2401 fld_s(1); // duplicate arguments for runtime call. Stack: Y X Y | |
2402 fld_s(1); // Stack: X Y X Y | |
2403 fabs(); // Stack: abs(X) Y X Y | |
2404 fast_pow(); // Stack: abs(X)^Y X Y | |
2405 fcmp(tmp, 0, false, false); // Stack: abs(X)^Y X Y | |
2406 // abs(X)^Y not equal to itself: abs(X)^Y is NaN go to slow case. | |
2407 | |
2408 pop(tmp2); | |
2409 NOT_LP64(pop(tmp3)); | |
2410 jcc(Assembler::parity, slow_case); | |
2411 | |
2412 #ifdef ASSERT | |
2413 // Check that int(Y) is not integer indefinite value (int | |
2414 // overflow). Shouldn't happen because for values that would | |
2415 // overflow, 1+int(Y)==Y which was tested earlier. | |
2416 #ifndef _LP64 | |
2417 { | |
2418 Label integer; | |
2419 testl(tmp2, tmp2); | |
2420 jcc(Assembler::notZero, integer); | |
2421 cmpl(tmp3, 0x80000000); | |
2422 jcc(Assembler::notZero, integer); | |
2423 STOP("integer indefinite value shouldn't be seen here"); | |
2424 bind(integer); | |
2425 } | |
2426 #else | |
2427 { | |
2428 Label integer; | |
2429 mov(tmp3, tmp2); // preserve tmp2 for parity check below | |
2430 shlq(tmp3, 1); | |
2431 jcc(Assembler::carryClear, integer); | |
2432 jcc(Assembler::notZero, integer); | |
2433 STOP("integer indefinite value shouldn't be seen here"); | |
2434 bind(integer); | |
2435 } | |
2436 #endif | |
2437 #endif | |
2438 | |
2439 // get rid of duplicate arguments. Stack: X^Y | |
2440 if (num_fpu_regs_in_use > 0) { | |
2441 fxch(); fpop(); | |
2442 fxch(); fpop(); | |
2443 } else { | |
2444 ffree(2); | |
2445 ffree(1); | |
2446 } | |
2447 | |
2448 testl(tmp2, 1); | |
2449 jcc(Assembler::zero, done); // X <= 0, Y even: X^Y = abs(X)^Y | |
2450 // X <= 0, Y even: X^Y = -abs(X)^Y | |
2451 | |
2452 fchs(); // Stack: -abs(X)^Y Y | |
2453 jmp(done); | |
2454 } | |
2455 | |
2456 // slow case: runtime call | |
2457 bind(slow_case); | |
2458 | |
2459 fpop(); // pop incorrect result or int(Y) | |
2460 | |
2461 fp_runtime_fallback(is_exp ? CAST_FROM_FN_PTR(address, SharedRuntime::dexp) : CAST_FROM_FN_PTR(address, SharedRuntime::dpow), | |
2462 is_exp ? 1 : 2, num_fpu_regs_in_use); | |
2463 | |
2464 // Come here with result in F-TOS | |
2465 bind(done); | |
2466 } | |
2467 | |
2468 void MacroAssembler::fpop() { | |
2469 ffree(); | |
2470 fincstp(); | |
2471 } | |
2472 | |
2473 void MacroAssembler::fremr(Register tmp) { | |
2474 save_rax(tmp); | |
2475 { Label L; | |
2476 bind(L); | |
2477 fprem(); | |
2478 fwait(); fnstsw_ax(); | |
2479 #ifdef _LP64 | |
2480 testl(rax, 0x400); | |
2481 jcc(Assembler::notEqual, L); | |
2482 #else | |
2483 sahf(); | |
2484 jcc(Assembler::parity, L); | |
2485 #endif // _LP64 | |
2486 } | |
2487 restore_rax(tmp); | |
2488 // Result is in ST0. | |
2489 // Note: fxch & fpop to get rid of ST1 | |
2490 // (otherwise FPU stack could overflow eventually) | |
2491 fxch(1); | |
2492 fpop(); | |
2493 } | |
2494 | |
2495 | |
2496 void MacroAssembler::incrementl(AddressLiteral dst) { | |
2497 if (reachable(dst)) { | |
2498 incrementl(as_Address(dst)); | |
2499 } else { | |
2500 lea(rscratch1, dst); | |
2501 incrementl(Address(rscratch1, 0)); | |
2502 } | |
2503 } | |
2504 | |
2505 void MacroAssembler::incrementl(ArrayAddress dst) { | |
2506 incrementl(as_Address(dst)); | |
2507 } | |
2508 | |
2509 void MacroAssembler::incrementl(Register reg, int value) { | |
2510 if (value == min_jint) {addl(reg, value) ; return; } | |
2511 if (value < 0) { decrementl(reg, -value); return; } | |
2512 if (value == 0) { ; return; } | |
2513 if (value == 1 && UseIncDec) { incl(reg) ; return; } | |
2514 /* else */ { addl(reg, value) ; return; } | |
2515 } | |
2516 | |
2517 void MacroAssembler::incrementl(Address dst, int value) { | |
2518 if (value == min_jint) {addl(dst, value) ; return; } | |
2519 if (value < 0) { decrementl(dst, -value); return; } | |
2520 if (value == 0) { ; return; } | |
2521 if (value == 1 && UseIncDec) { incl(dst) ; return; } | |
2522 /* else */ { addl(dst, value) ; return; } | |
2523 } | |
2524 | |
2525 void MacroAssembler::jump(AddressLiteral dst) { | |
2526 if (reachable(dst)) { | |
2527 jmp_literal(dst.target(), dst.rspec()); | |
2528 } else { | |
2529 lea(rscratch1, dst); | |
2530 jmp(rscratch1); | |
2531 } | |
2532 } | |
2533 | |
2534 void MacroAssembler::jump_cc(Condition cc, AddressLiteral dst) { | |
2535 if (reachable(dst)) { | |
2536 InstructionMark im(this); | |
2537 relocate(dst.reloc()); | |
2538 const int short_size = 2; | |
2539 const int long_size = 6; | |
2540 int offs = (intptr_t)dst.target() - ((intptr_t)pc()); | |
2541 if (dst.reloc() == relocInfo::none && is8bit(offs - short_size)) { | |
2542 // 0111 tttn #8-bit disp | |
7430
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2543 emit_int8(0x70 | cc); |
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2544 emit_int8((offs - short_size) & 0xFF); |
7199 | 2545 } else { |
2546 // 0000 1111 1000 tttn #32-bit disp | |
7430
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2547 emit_int8(0x0F); |
d02120b7a34f
8004250: replace AbstractAssembler a_byte/a_long with emit_int8/emit_int32
twisti
parents:
7427
diff
changeset
|
2548 emit_int8((unsigned char)(0x80 | cc)); |
7476
ffa87474d7a4
8004537: replace AbstractAssembler emit_long with emit_int32
twisti
parents:
7475
diff
changeset
|
2549 emit_int32(offs - long_size); |
7199 | 2550 } |
2551 } else { | |
2552 #ifdef ASSERT | |
2553 warning("reversing conditional branch"); | |
2554 #endif /* ASSERT */ | |
2555 Label skip; | |
2556 jccb(reverse[cc], skip); | |
2557 lea(rscratch1, dst); | |
2558 Assembler::jmp(rscratch1); | |
2559 bind(skip); | |
2560 } | |
2561 } | |
2562 | |
2563 void MacroAssembler::ldmxcsr(AddressLiteral src) { | |
2564 if (reachable(src)) { | |
2565 Assembler::ldmxcsr(as_Address(src)); | |
2566 } else { | |
2567 lea(rscratch1, src); | |
2568 Assembler::ldmxcsr(Address(rscratch1, 0)); | |
2569 } | |
2570 } | |
2571 | |
2572 int MacroAssembler::load_signed_byte(Register dst, Address src) { | |
2573 int off; | |
2574 if (LP64_ONLY(true ||) VM_Version::is_P6()) { | |
2575 off = offset(); | |
2576 movsbl(dst, src); // movsxb | |
2577 } else { | |
2578 off = load_unsigned_byte(dst, src); | |
2579 shll(dst, 24); | |
2580 sarl(dst, 24); | |
2581 } | |
2582 return off; | |
2583 } | |
2584 | |
2585 // Note: load_signed_short used to be called load_signed_word. | |
2586 // Although the 'w' in x86 opcodes refers to the term "word" in the assembler | |
2587 // manual, which means 16 bits, that usage is found nowhere in HotSpot code. | |
2588 // The term "word" in HotSpot means a 32- or 64-bit machine word. | |
2589 int MacroAssembler::load_signed_short(Register dst, Address src) { | |
2590 int off; | |
2591 if (LP64_ONLY(true ||) VM_Version::is_P6()) { | |
2592 // This is dubious to me since it seems safe to do a signed 16 => 64 bit | |
2593 // version but this is what 64bit has always done. This seems to imply | |
2594 // that users are only using 32bits worth. | |
2595 off = offset(); | |
2596 movswl(dst, src); // movsxw | |
2597 } else { | |
2598 off = load_unsigned_short(dst, src); | |
2599 shll(dst, 16); | |
2600 sarl(dst, 16); | |
2601 } | |
2602 return off; | |
2603 } | |
2604 | |
2605 int MacroAssembler::load_unsigned_byte(Register dst, Address src) { | |
2606 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, | |
2607 // and "3.9 Partial Register Penalties", p. 22). | |
2608 int off; | |
2609 if (LP64_ONLY(true || ) VM_Version::is_P6() || src.uses(dst)) { | |
2610 off = offset(); | |
2611 movzbl(dst, src); // movzxb | |
2612 } else { | |
2613 xorl(dst, dst); | |
2614 off = offset(); | |
2615 movb(dst, src); | |
2616 } | |
2617 return off; | |
2618 } | |
2619 | |
2620 // Note: load_unsigned_short used to be called load_unsigned_word. | |
2621 int MacroAssembler::load_unsigned_short(Register dst, Address src) { | |
2622 // According to Intel Doc. AP-526, "Zero-Extension of Short", p.16, | |
2623 // and "3.9 Partial Register Penalties", p. 22). | |
2624 int off; | |
2625 if (LP64_ONLY(true ||) VM_Version::is_P6() || src.uses(dst)) { | |
2626 off = offset(); | |
2627 movzwl(dst, src); // movzxw | |
2628 } else { | |
2629 xorl(dst, dst); | |
2630 off = offset(); | |
2631 movw(dst, src); | |
2632 } | |
2633 return off; | |
2634 } | |
2635 | |
2636 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) { | |
2637 switch (size_in_bytes) { | |
2638 #ifndef _LP64 | |
2639 case 8: | |
2640 assert(dst2 != noreg, "second dest register required"); | |
2641 movl(dst, src); | |
2642 movl(dst2, src.plus_disp(BytesPerInt)); | |
2643 break; | |
2644 #else | |
2645 case 8: movq(dst, src); break; | |
2646 #endif | |
2647 case 4: movl(dst, src); break; | |
2648 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break; | |
2649 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break; | |
2650 default: ShouldNotReachHere(); | |
2651 } | |
2652 } | |
2653 | |
2654 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) { | |
2655 switch (size_in_bytes) { | |
2656 #ifndef _LP64 | |
2657 case 8: | |
2658 assert(src2 != noreg, "second source register required"); | |
2659 movl(dst, src); | |
2660 movl(dst.plus_disp(BytesPerInt), src2); | |
2661 break; | |
2662 #else | |
2663 case 8: movq(dst, src); break; | |
2664 #endif | |
2665 case 4: movl(dst, src); break; | |
2666 case 2: movw(dst, src); break; | |
2667 case 1: movb(dst, src); break; | |
2668 default: ShouldNotReachHere(); | |
2669 } | |
2670 } | |
2671 | |
2672 void MacroAssembler::mov32(AddressLiteral dst, Register src) { | |
2673 if (reachable(dst)) { | |
2674 movl(as_Address(dst), src); | |
2675 } else { | |
2676 lea(rscratch1, dst); | |
2677 movl(Address(rscratch1, 0), src); | |
2678 } | |
2679 } | |
2680 | |
2681 void MacroAssembler::mov32(Register dst, AddressLiteral src) { | |
2682 if (reachable(src)) { | |
2683 movl(dst, as_Address(src)); | |
2684 } else { | |
2685 lea(rscratch1, src); | |
2686 movl(dst, Address(rscratch1, 0)); | |
2687 } | |
2688 } | |
2689 | |
2690 // C++ bool manipulation | |
2691 | |
2692 void MacroAssembler::movbool(Register dst, Address src) { | |
2693 if(sizeof(bool) == 1) | |
2694 movb(dst, src); | |
2695 else if(sizeof(bool) == 2) | |
2696 movw(dst, src); | |
2697 else if(sizeof(bool) == 4) | |
2698 movl(dst, src); | |
2699 else | |
2700 // unsupported | |
2701 ShouldNotReachHere(); | |
2702 } | |
2703 | |
2704 void MacroAssembler::movbool(Address dst, bool boolconst) { | |
2705 if(sizeof(bool) == 1) | |
2706 movb(dst, (int) boolconst); | |
2707 else if(sizeof(bool) == 2) | |
2708 movw(dst, (int) boolconst); | |
2709 else if(sizeof(bool) == 4) | |
2710 movl(dst, (int) boolconst); | |
2711 else | |
2712 // unsupported | |
2713 ShouldNotReachHere(); | |
2714 } | |
2715 | |
2716 void MacroAssembler::movbool(Address dst, Register src) { | |
2717 if(sizeof(bool) == 1) | |
2718 movb(dst, src); | |
2719 else if(sizeof(bool) == 2) | |
2720 movw(dst, src); | |
2721 else if(sizeof(bool) == 4) | |
2722 movl(dst, src); | |
2723 else | |
2724 // unsupported | |
2725 ShouldNotReachHere(); | |
2726 } | |
2727 | |
2728 void MacroAssembler::movbyte(ArrayAddress dst, int src) { | |
2729 movb(as_Address(dst), src); | |
2730 } | |
2731 | |
2732 void MacroAssembler::movdl(XMMRegister dst, AddressLiteral src) { | |
2733 if (reachable(src)) { | |
2734 movdl(dst, as_Address(src)); | |
2735 } else { | |
2736 lea(rscratch1, src); | |
2737 movdl(dst, Address(rscratch1, 0)); | |
2738 } | |
2739 } | |
2740 | |
2741 void MacroAssembler::movq(XMMRegister dst, AddressLiteral src) { | |
2742 if (reachable(src)) { | |
2743 movq(dst, as_Address(src)); | |
2744 } else { | |
2745 lea(rscratch1, src); | |
2746 movq(dst, Address(rscratch1, 0)); | |
2747 } | |
2748 } | |
2749 | |
2750 void MacroAssembler::movdbl(XMMRegister dst, AddressLiteral src) { | |
2751 if (reachable(src)) { | |
2752 if (UseXmmLoadAndClearUpper) { | |
2753 movsd (dst, as_Address(src)); | |
2754 } else { | |
2755 movlpd(dst, as_Address(src)); | |
2756 } | |
2757 } else { | |
2758 lea(rscratch1, src); | |
2759 if (UseXmmLoadAndClearUpper) { | |
2760 movsd (dst, Address(rscratch1, 0)); | |
2761 } else { | |
2762 movlpd(dst, Address(rscratch1, 0)); | |
2763 } | |
2764 } | |
2765 } | |
2766 | |
2767 void MacroAssembler::movflt(XMMRegister dst, AddressLiteral src) { | |
2768 if (reachable(src)) { | |
2769 movss(dst, as_Address(src)); | |
2770 } else { | |
2771 lea(rscratch1, src); | |
2772 movss(dst, Address(rscratch1, 0)); | |
2773 } | |
2774 } | |
2775 | |
2776 void MacroAssembler::movptr(Register dst, Register src) { | |
2777 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); | |
2778 } | |
2779 | |
2780 void MacroAssembler::movptr(Register dst, Address src) { | |
2781 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); | |
2782 } | |
2783 | |
2784 // src should NEVER be a real pointer. Use AddressLiteral for true pointers | |
2785 void MacroAssembler::movptr(Register dst, intptr_t src) { | |
2786 LP64_ONLY(mov64(dst, src)) NOT_LP64(movl(dst, src)); | |
2787 } | |
2788 | |
2789 void MacroAssembler::movptr(Address dst, Register src) { | |
2790 LP64_ONLY(movq(dst, src)) NOT_LP64(movl(dst, src)); | |
2791 } | |
2792 | |
2793 void MacroAssembler::movdqu(XMMRegister dst, AddressLiteral src) { | |
2794 if (reachable(src)) { | |
2795 Assembler::movdqu(dst, as_Address(src)); | |
2796 } else { | |
2797 lea(rscratch1, src); | |
2798 Assembler::movdqu(dst, Address(rscratch1, 0)); | |
2799 } | |
2800 } | |
2801 | |
11080
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2802 void MacroAssembler::movdqa(XMMRegister dst, AddressLiteral src) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2803 if (reachable(src)) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2804 Assembler::movdqa(dst, as_Address(src)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2805 } else { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2806 lea(rscratch1, src); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2807 Assembler::movdqa(dst, Address(rscratch1, 0)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2808 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2809 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
2810 |
7199 | 2811 void MacroAssembler::movsd(XMMRegister dst, AddressLiteral src) { |
2812 if (reachable(src)) { | |
2813 Assembler::movsd(dst, as_Address(src)); | |
2814 } else { | |
2815 lea(rscratch1, src); | |
2816 Assembler::movsd(dst, Address(rscratch1, 0)); | |
2817 } | |
2818 } | |
2819 | |
2820 void MacroAssembler::movss(XMMRegister dst, AddressLiteral src) { | |
2821 if (reachable(src)) { | |
2822 Assembler::movss(dst, as_Address(src)); | |
2823 } else { | |
2824 lea(rscratch1, src); | |
2825 Assembler::movss(dst, Address(rscratch1, 0)); | |
2826 } | |
2827 } | |
2828 | |
2829 void MacroAssembler::mulsd(XMMRegister dst, AddressLiteral src) { | |
2830 if (reachable(src)) { | |
2831 Assembler::mulsd(dst, as_Address(src)); | |
2832 } else { | |
2833 lea(rscratch1, src); | |
2834 Assembler::mulsd(dst, Address(rscratch1, 0)); | |
2835 } | |
2836 } | |
2837 | |
2838 void MacroAssembler::mulss(XMMRegister dst, AddressLiteral src) { | |
2839 if (reachable(src)) { | |
2840 Assembler::mulss(dst, as_Address(src)); | |
2841 } else { | |
2842 lea(rscratch1, src); | |
2843 Assembler::mulss(dst, Address(rscratch1, 0)); | |
2844 } | |
2845 } | |
2846 | |
2847 void MacroAssembler::null_check(Register reg, int offset) { | |
2848 if (needs_explicit_null_check(offset)) { | |
2849 // provoke OS NULL exception if reg = NULL by | |
2850 // accessing M[reg] w/o changing any (non-CC) registers | |
2851 // NOTE: cmpl is plenty here to provoke a segv | |
2852 cmpptr(rax, Address(reg, 0)); | |
2853 // Note: should probably use testl(rax, Address(reg, 0)); | |
2854 // may be shorter code (however, this version of | |
2855 // testl needs to be implemented first) | |
2856 } else { | |
2857 // nothing to do, (later) access of M[reg + offset] | |
2858 // will provoke OS NULL exception if reg = NULL | |
2859 } | |
2860 } | |
2861 | |
2862 void MacroAssembler::os_breakpoint() { | |
2863 // instead of directly emitting a breakpoint, call os:breakpoint for better debugability | |
2864 // (e.g., MSVC can't call ps() otherwise) | |
2865 call(RuntimeAddress(CAST_FROM_FN_PTR(address, os::breakpoint))); | |
2866 } | |
2867 | |
2868 void MacroAssembler::pop_CPU_state() { | |
2869 pop_FPU_state(); | |
2870 pop_IU_state(); | |
2871 } | |
2872 | |
2873 void MacroAssembler::pop_FPU_state() { | |
2874 NOT_LP64(frstor(Address(rsp, 0));) | |
2875 LP64_ONLY(fxrstor(Address(rsp, 0));) | |
2876 addptr(rsp, FPUStateSizeInWords * wordSize); | |
2877 } | |
2878 | |
2879 void MacroAssembler::pop_IU_state() { | |
2880 popa(); | |
2881 LP64_ONLY(addq(rsp, 8)); | |
2882 popf(); | |
2883 } | |
2884 | |
2885 // Save Integer and Float state | |
2886 // Warning: Stack must be 16 byte aligned (64bit) | |
2887 void MacroAssembler::push_CPU_state() { | |
2888 push_IU_state(); | |
2889 push_FPU_state(); | |
2890 } | |
2891 | |
2892 void MacroAssembler::push_FPU_state() { | |
2893 subptr(rsp, FPUStateSizeInWords * wordSize); | |
2894 #ifndef _LP64 | |
2895 fnsave(Address(rsp, 0)); | |
2896 fwait(); | |
2897 #else | |
2898 fxsave(Address(rsp, 0)); | |
2899 #endif // LP64 | |
2900 } | |
2901 | |
2902 void MacroAssembler::push_IU_state() { | |
2903 // Push flags first because pusha kills them | |
2904 pushf(); | |
2905 // Make sure rsp stays 16-byte aligned | |
2906 LP64_ONLY(subq(rsp, 8)); | |
2907 pusha(); | |
2908 } | |
2909 | |
2910 void MacroAssembler::reset_last_Java_frame(Register java_thread, bool clear_fp, bool clear_pc) { | |
2911 // determine java_thread register | |
2912 if (!java_thread->is_valid()) { | |
2913 java_thread = rdi; | |
2914 get_thread(java_thread); | |
2915 } | |
2916 // we must set sp to zero to clear frame | |
2917 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), NULL_WORD); | |
2918 if (clear_fp) { | |
2919 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), NULL_WORD); | |
2920 } | |
2921 | |
2922 if (clear_pc) | |
2923 movptr(Address(java_thread, JavaThread::last_Java_pc_offset()), NULL_WORD); | |
2924 | |
2925 } | |
2926 | |
2927 void MacroAssembler::restore_rax(Register tmp) { | |
2928 if (tmp == noreg) pop(rax); | |
2929 else if (tmp != rax) mov(rax, tmp); | |
2930 } | |
2931 | |
2932 void MacroAssembler::round_to(Register reg, int modulus) { | |
2933 addptr(reg, modulus - 1); | |
2934 andptr(reg, -modulus); | |
2935 } | |
2936 | |
2937 void MacroAssembler::save_rax(Register tmp) { | |
2938 if (tmp == noreg) push(rax); | |
2939 else if (tmp != rax) mov(tmp, rax); | |
2940 } | |
2941 | |
2942 // Write serialization page so VM thread can do a pseudo remote membar. | |
2943 // We use the current thread pointer to calculate a thread specific | |
2944 // offset to write to within the page. This minimizes bus traffic | |
2945 // due to cache line collision. | |
2946 void MacroAssembler::serialize_memory(Register thread, Register tmp) { | |
2947 movl(tmp, thread); | |
2948 shrl(tmp, os::get_serialize_page_shift_count()); | |
2949 andl(tmp, (os::vm_page_size() - sizeof(int))); | |
2950 | |
2951 Address index(noreg, tmp, Address::times_1); | |
2952 ExternalAddress page(os::get_memory_serialize_page()); | |
2953 | |
2954 // Size of store must match masking code above | |
2955 movl(as_Address(ArrayAddress(page, index)), tmp); | |
2956 } | |
2957 | |
2958 // Calls to C land | |
2959 // | |
2960 // When entering C land, the rbp, & rsp of the last Java frame have to be recorded | |
2961 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp | |
2962 // has to be reset to 0. This is required to allow proper stack traversal. | |
2963 void MacroAssembler::set_last_Java_frame(Register java_thread, | |
2964 Register last_java_sp, | |
2965 Register last_java_fp, | |
2966 address last_java_pc) { | |
2967 // determine java_thread register | |
2968 if (!java_thread->is_valid()) { | |
2969 java_thread = rdi; | |
2970 get_thread(java_thread); | |
2971 } | |
2972 // determine last_java_sp register | |
2973 if (!last_java_sp->is_valid()) { | |
2974 last_java_sp = rsp; | |
2975 } | |
2976 | |
2977 // last_java_fp is optional | |
2978 | |
2979 if (last_java_fp->is_valid()) { | |
2980 movptr(Address(java_thread, JavaThread::last_Java_fp_offset()), last_java_fp); | |
2981 } | |
2982 | |
2983 // last_java_pc is optional | |
2984 | |
2985 if (last_java_pc != NULL) { | |
2986 lea(Address(java_thread, | |
2987 JavaThread::frame_anchor_offset() + JavaFrameAnchor::last_Java_pc_offset()), | |
2988 InternalAddress(last_java_pc)); | |
2989 | |
2990 } | |
2991 movptr(Address(java_thread, JavaThread::last_Java_sp_offset()), last_java_sp); | |
2992 } | |
2993 | |
2994 void MacroAssembler::shlptr(Register dst, int imm8) { | |
2995 LP64_ONLY(shlq(dst, imm8)) NOT_LP64(shll(dst, imm8)); | |
2996 } | |
2997 | |
2998 void MacroAssembler::shrptr(Register dst, int imm8) { | |
2999 LP64_ONLY(shrq(dst, imm8)) NOT_LP64(shrl(dst, imm8)); | |
3000 } | |
3001 | |
3002 void MacroAssembler::sign_extend_byte(Register reg) { | |
3003 if (LP64_ONLY(true ||) (VM_Version::is_P6() && reg->has_byte_register())) { | |
3004 movsbl(reg, reg); // movsxb | |
3005 } else { | |
3006 shll(reg, 24); | |
3007 sarl(reg, 24); | |
3008 } | |
3009 } | |
3010 | |
3011 void MacroAssembler::sign_extend_short(Register reg) { | |
3012 if (LP64_ONLY(true ||) VM_Version::is_P6()) { | |
3013 movswl(reg, reg); // movsxw | |
3014 } else { | |
3015 shll(reg, 16); | |
3016 sarl(reg, 16); | |
3017 } | |
3018 } | |
3019 | |
3020 void MacroAssembler::testl(Register dst, AddressLiteral src) { | |
3021 assert(reachable(src), "Address should be reachable"); | |
3022 testl(dst, as_Address(src)); | |
3023 } | |
3024 | |
3025 void MacroAssembler::sqrtsd(XMMRegister dst, AddressLiteral src) { | |
3026 if (reachable(src)) { | |
3027 Assembler::sqrtsd(dst, as_Address(src)); | |
3028 } else { | |
3029 lea(rscratch1, src); | |
3030 Assembler::sqrtsd(dst, Address(rscratch1, 0)); | |
3031 } | |
3032 } | |
3033 | |
3034 void MacroAssembler::sqrtss(XMMRegister dst, AddressLiteral src) { | |
3035 if (reachable(src)) { | |
3036 Assembler::sqrtss(dst, as_Address(src)); | |
3037 } else { | |
3038 lea(rscratch1, src); | |
3039 Assembler::sqrtss(dst, Address(rscratch1, 0)); | |
3040 } | |
3041 } | |
3042 | |
3043 void MacroAssembler::subsd(XMMRegister dst, AddressLiteral src) { | |
3044 if (reachable(src)) { | |
3045 Assembler::subsd(dst, as_Address(src)); | |
3046 } else { | |
3047 lea(rscratch1, src); | |
3048 Assembler::subsd(dst, Address(rscratch1, 0)); | |
3049 } | |
3050 } | |
3051 | |
3052 void MacroAssembler::subss(XMMRegister dst, AddressLiteral src) { | |
3053 if (reachable(src)) { | |
3054 Assembler::subss(dst, as_Address(src)); | |
3055 } else { | |
3056 lea(rscratch1, src); | |
3057 Assembler::subss(dst, Address(rscratch1, 0)); | |
3058 } | |
3059 } | |
3060 | |
3061 void MacroAssembler::ucomisd(XMMRegister dst, AddressLiteral src) { | |
3062 if (reachable(src)) { | |
3063 Assembler::ucomisd(dst, as_Address(src)); | |
3064 } else { | |
3065 lea(rscratch1, src); | |
3066 Assembler::ucomisd(dst, Address(rscratch1, 0)); | |
3067 } | |
3068 } | |
3069 | |
3070 void MacroAssembler::ucomiss(XMMRegister dst, AddressLiteral src) { | |
3071 if (reachable(src)) { | |
3072 Assembler::ucomiss(dst, as_Address(src)); | |
3073 } else { | |
3074 lea(rscratch1, src); | |
3075 Assembler::ucomiss(dst, Address(rscratch1, 0)); | |
3076 } | |
3077 } | |
3078 | |
3079 void MacroAssembler::xorpd(XMMRegister dst, AddressLiteral src) { | |
3080 // Used in sign-bit flipping with aligned address. | |
3081 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); | |
3082 if (reachable(src)) { | |
3083 Assembler::xorpd(dst, as_Address(src)); | |
3084 } else { | |
3085 lea(rscratch1, src); | |
3086 Assembler::xorpd(dst, Address(rscratch1, 0)); | |
3087 } | |
3088 } | |
3089 | |
3090 void MacroAssembler::xorps(XMMRegister dst, AddressLiteral src) { | |
3091 // Used in sign-bit flipping with aligned address. | |
3092 assert((UseAVX > 0) || (((intptr_t)src.target() & 15) == 0), "SSE mode requires address alignment 16 bytes"); | |
3093 if (reachable(src)) { | |
3094 Assembler::xorps(dst, as_Address(src)); | |
3095 } else { | |
3096 lea(rscratch1, src); | |
3097 Assembler::xorps(dst, Address(rscratch1, 0)); | |
3098 } | |
3099 } | |
3100 | |
3101 void MacroAssembler::pshufb(XMMRegister dst, AddressLiteral src) { | |
3102 // Used in sign-bit flipping with aligned address. | |
7427 | 3103 bool aligned_adr = (((intptr_t)src.target() & 15) == 0); |
3104 assert((UseAVX > 0) || aligned_adr, "SSE mode requires address alignment 16 bytes"); | |
7199 | 3105 if (reachable(src)) { |
3106 Assembler::pshufb(dst, as_Address(src)); | |
3107 } else { | |
3108 lea(rscratch1, src); | |
3109 Assembler::pshufb(dst, Address(rscratch1, 0)); | |
3110 } | |
3111 } | |
3112 | |
3113 // AVX 3-operands instructions | |
3114 | |
3115 void MacroAssembler::vaddsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { | |
3116 if (reachable(src)) { | |
3117 vaddsd(dst, nds, as_Address(src)); | |
3118 } else { | |
3119 lea(rscratch1, src); | |
3120 vaddsd(dst, nds, Address(rscratch1, 0)); | |
3121 } | |
3122 } | |
3123 | |
3124 void MacroAssembler::vaddss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { | |
3125 if (reachable(src)) { | |
3126 vaddss(dst, nds, as_Address(src)); | |
3127 } else { | |
3128 lea(rscratch1, src); | |
3129 vaddss(dst, nds, Address(rscratch1, 0)); | |
3130 } | |
3131 } | |
3132 | |
3133 void MacroAssembler::vandpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) { | |
3134 if (reachable(src)) { | |
3135 vandpd(dst, nds, as_Address(src), vector256); | |
3136 } else { | |
3137 lea(rscratch1, src); | |
3138 vandpd(dst, nds, Address(rscratch1, 0), vector256); | |
3139 } | |
3140 } | |
3141 | |
3142 void MacroAssembler::vandps(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) { | |
3143 if (reachable(src)) { | |
3144 vandps(dst, nds, as_Address(src), vector256); | |
3145 } else { | |
3146 lea(rscratch1, src); | |
3147 vandps(dst, nds, Address(rscratch1, 0), vector256); | |
3148 } | |
3149 } | |
3150 | |
3151 void MacroAssembler::vdivsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { | |
3152 if (reachable(src)) { | |
3153 vdivsd(dst, nds, as_Address(src)); | |
3154 } else { | |
3155 lea(rscratch1, src); | |
3156 vdivsd(dst, nds, Address(rscratch1, 0)); | |
3157 } | |
3158 } | |
3159 | |
3160 void MacroAssembler::vdivss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { | |
3161 if (reachable(src)) { | |
3162 vdivss(dst, nds, as_Address(src)); | |
3163 } else { | |
3164 lea(rscratch1, src); | |
3165 vdivss(dst, nds, Address(rscratch1, 0)); | |
3166 } | |
3167 } | |
3168 | |
3169 void MacroAssembler::vmulsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { | |
3170 if (reachable(src)) { | |
3171 vmulsd(dst, nds, as_Address(src)); | |
3172 } else { | |
3173 lea(rscratch1, src); | |
3174 vmulsd(dst, nds, Address(rscratch1, 0)); | |
3175 } | |
3176 } | |
3177 | |
3178 void MacroAssembler::vmulss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { | |
3179 if (reachable(src)) { | |
3180 vmulss(dst, nds, as_Address(src)); | |
3181 } else { | |
3182 lea(rscratch1, src); | |
3183 vmulss(dst, nds, Address(rscratch1, 0)); | |
3184 } | |
3185 } | |
3186 | |
3187 void MacroAssembler::vsubsd(XMMRegister dst, XMMRegister nds, AddressLiteral src) { | |
3188 if (reachable(src)) { | |
3189 vsubsd(dst, nds, as_Address(src)); | |
3190 } else { | |
3191 lea(rscratch1, src); | |
3192 vsubsd(dst, nds, Address(rscratch1, 0)); | |
3193 } | |
3194 } | |
3195 | |
3196 void MacroAssembler::vsubss(XMMRegister dst, XMMRegister nds, AddressLiteral src) { | |
3197 if (reachable(src)) { | |
3198 vsubss(dst, nds, as_Address(src)); | |
3199 } else { | |
3200 lea(rscratch1, src); | |
3201 vsubss(dst, nds, Address(rscratch1, 0)); | |
3202 } | |
3203 } | |
3204 | |
3205 void MacroAssembler::vxorpd(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) { | |
3206 if (reachable(src)) { | |
3207 vxorpd(dst, nds, as_Address(src), vector256); | |
3208 } else { | |
3209 lea(rscratch1, src); | |
3210 vxorpd(dst, nds, Address(rscratch1, 0), vector256); | |
3211 } | |
3212 } | |
3213 | |
3214 void MacroAssembler::vxorps(XMMRegister dst, XMMRegister nds, AddressLiteral src, bool vector256) { | |
3215 if (reachable(src)) { | |
3216 vxorps(dst, nds, as_Address(src), vector256); | |
3217 } else { | |
3218 lea(rscratch1, src); | |
3219 vxorps(dst, nds, Address(rscratch1, 0), vector256); | |
3220 } | |
3221 } | |
3222 | |
3223 | |
3224 ////////////////////////////////////////////////////////////////////////////////// | |
8001
db9981fd3124
8005915: Unify SERIALGC and INCLUDE_ALTERNATE_GCS
jprovino
parents:
7477
diff
changeset
|
3225 #if INCLUDE_ALL_GCS |
7199 | 3226 |
3227 void MacroAssembler::g1_write_barrier_pre(Register obj, | |
3228 Register pre_val, | |
3229 Register thread, | |
3230 Register tmp, | |
3231 bool tosca_live, | |
3232 bool expand_call) { | |
3233 | |
3234 // If expand_call is true then we expand the call_VM_leaf macro | |
3235 // directly to skip generating the check by | |
3236 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp. | |
3237 | |
3238 #ifdef _LP64 | |
3239 assert(thread == r15_thread, "must be"); | |
3240 #endif // _LP64 | |
3241 | |
3242 Label done; | |
3243 Label runtime; | |
3244 | |
3245 assert(pre_val != noreg, "check this code"); | |
3246 | |
3247 if (obj != noreg) { | |
3248 assert_different_registers(obj, pre_val, tmp); | |
3249 assert(pre_val != rax, "check this code"); | |
3250 } | |
3251 | |
3252 Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() + | |
3253 PtrQueue::byte_offset_of_active())); | |
3254 Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() + | |
3255 PtrQueue::byte_offset_of_index())); | |
3256 Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() + | |
3257 PtrQueue::byte_offset_of_buf())); | |
3258 | |
3259 | |
3260 // Is marking active? | |
3261 if (in_bytes(PtrQueue::byte_width_of_active()) == 4) { | |
3262 cmpl(in_progress, 0); | |
3263 } else { | |
3264 assert(in_bytes(PtrQueue::byte_width_of_active()) == 1, "Assumption"); | |
3265 cmpb(in_progress, 0); | |
3266 } | |
3267 jcc(Assembler::equal, done); | |
3268 | |
3269 // Do we need to load the previous value? | |
3270 if (obj != noreg) { | |
3271 load_heap_oop(pre_val, Address(obj, 0)); | |
3272 } | |
3273 | |
3274 // Is the previous value null? | |
3275 cmpptr(pre_val, (int32_t) NULL_WORD); | |
3276 jcc(Assembler::equal, done); | |
3277 | |
3278 // Can we store original value in the thread's buffer? | |
3279 // Is index == 0? | |
3280 // (The index field is typed as size_t.) | |
3281 | |
3282 movptr(tmp, index); // tmp := *index_adr | |
3283 cmpptr(tmp, 0); // tmp == 0? | |
3284 jcc(Assembler::equal, runtime); // If yes, goto runtime | |
3285 | |
3286 subptr(tmp, wordSize); // tmp := tmp - wordSize | |
3287 movptr(index, tmp); // *index_adr := tmp | |
3288 addptr(tmp, buffer); // tmp := tmp + *buffer_adr | |
3289 | |
3290 // Record the previous value | |
3291 movptr(Address(tmp, 0), pre_val); | |
3292 jmp(done); | |
3293 | |
3294 bind(runtime); | |
3295 // save the live input values | |
3296 if(tosca_live) push(rax); | |
3297 | |
3298 if (obj != noreg && obj != rax) | |
3299 push(obj); | |
3300 | |
3301 if (pre_val != rax) | |
3302 push(pre_val); | |
3303 | |
3304 // Calling the runtime using the regular call_VM_leaf mechanism generates | |
3305 // code (generated by InterpreterMacroAssember::call_VM_leaf_base) | |
3306 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL. | |
3307 // | |
3308 // If we care generating the pre-barrier without a frame (e.g. in the | |
3309 // intrinsified Reference.get() routine) then ebp might be pointing to | |
3310 // the caller frame and so this check will most likely fail at runtime. | |
3311 // | |
3312 // Expanding the call directly bypasses the generation of the check. | |
3313 // So when we do not have have a full interpreter frame on the stack | |
3314 // expand_call should be passed true. | |
3315 | |
3316 NOT_LP64( push(thread); ) | |
3317 | |
3318 if (expand_call) { | |
3319 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); ) | |
3320 pass_arg1(this, thread); | |
3321 pass_arg0(this, pre_val); | |
3322 MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), 2); | |
3323 } else { | |
3324 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread); | |
3325 } | |
3326 | |
3327 NOT_LP64( pop(thread); ) | |
3328 | |
3329 // save the live input values | |
3330 if (pre_val != rax) | |
3331 pop(pre_val); | |
3332 | |
3333 if (obj != noreg && obj != rax) | |
3334 pop(obj); | |
3335 | |
3336 if(tosca_live) pop(rax); | |
3337 | |
3338 bind(done); | |
3339 } | |
3340 | |
3341 void MacroAssembler::g1_write_barrier_post(Register store_addr, | |
3342 Register new_val, | |
3343 Register thread, | |
3344 Register tmp, | |
3345 Register tmp2) { | |
3346 #ifdef _LP64 | |
3347 assert(thread == r15_thread, "must be"); | |
3348 #endif // _LP64 | |
3349 | |
3350 Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() + | |
3351 PtrQueue::byte_offset_of_index())); | |
3352 Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() + | |
3353 PtrQueue::byte_offset_of_buf())); | |
3354 | |
3355 BarrierSet* bs = Universe::heap()->barrier_set(); | |
3356 CardTableModRefBS* ct = (CardTableModRefBS*)bs; | |
3357 Label done; | |
3358 Label runtime; | |
3359 | |
3360 // Does store cross heap regions? | |
3361 | |
3362 movptr(tmp, store_addr); | |
3363 xorptr(tmp, new_val); | |
3364 shrptr(tmp, HeapRegion::LogOfHRGrainBytes); | |
3365 jcc(Assembler::equal, done); | |
3366 | |
3367 // crosses regions, storing NULL? | |
3368 | |
3369 cmpptr(new_val, (int32_t) NULL_WORD); | |
3370 jcc(Assembler::equal, done); | |
3371 | |
3372 // storing region crossing non-NULL, is card already dirty? | |
3373 | |
3374 ExternalAddress cardtable((address) ct->byte_map_base); | |
3375 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); | |
3376 #ifdef _LP64 | |
3377 const Register card_addr = tmp; | |
3378 | |
3379 movq(card_addr, store_addr); | |
3380 shrq(card_addr, CardTableModRefBS::card_shift); | |
3381 | |
3382 lea(tmp2, cardtable); | |
3383 | |
3384 // get the address of the card | |
3385 addq(card_addr, tmp2); | |
3386 #else | |
3387 const Register card_index = tmp; | |
3388 | |
3389 movl(card_index, store_addr); | |
3390 shrl(card_index, CardTableModRefBS::card_shift); | |
3391 | |
3392 Address index(noreg, card_index, Address::times_1); | |
3393 const Register card_addr = tmp; | |
3394 lea(card_addr, as_Address(ArrayAddress(cardtable, index))); | |
3395 #endif | |
12835
69944b868a32
8014555: G1: Memory ordering problem with Conc refinement and card marking
mgerdin
parents:
12226
diff
changeset
|
3396 cmpb(Address(card_addr, 0), (int)G1SATBCardTableModRefBS::g1_young_card_val()); |
7199 | 3397 jcc(Assembler::equal, done); |
3398 | |
12835
69944b868a32
8014555: G1: Memory ordering problem with Conc refinement and card marking
mgerdin
parents:
12226
diff
changeset
|
3399 membar(Assembler::Membar_mask_bits(Assembler::StoreLoad)); |
69944b868a32
8014555: G1: Memory ordering problem with Conc refinement and card marking
mgerdin
parents:
12226
diff
changeset
|
3400 cmpb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val()); |
69944b868a32
8014555: G1: Memory ordering problem with Conc refinement and card marking
mgerdin
parents:
12226
diff
changeset
|
3401 jcc(Assembler::equal, done); |
69944b868a32
8014555: G1: Memory ordering problem with Conc refinement and card marking
mgerdin
parents:
12226
diff
changeset
|
3402 |
69944b868a32
8014555: G1: Memory ordering problem with Conc refinement and card marking
mgerdin
parents:
12226
diff
changeset
|
3403 |
7199 | 3404 // storing a region crossing, non-NULL oop, card is clean. |
3405 // dirty card and log. | |
3406 | |
12835
69944b868a32
8014555: G1: Memory ordering problem with Conc refinement and card marking
mgerdin
parents:
12226
diff
changeset
|
3407 movb(Address(card_addr, 0), (int)CardTableModRefBS::dirty_card_val()); |
7199 | 3408 |
3409 cmpl(queue_index, 0); | |
3410 jcc(Assembler::equal, runtime); | |
3411 subl(queue_index, wordSize); | |
3412 movptr(tmp2, buffer); | |
3413 #ifdef _LP64 | |
3414 movslq(rscratch1, queue_index); | |
3415 addq(tmp2, rscratch1); | |
3416 movq(Address(tmp2, 0), card_addr); | |
3417 #else | |
3418 addl(tmp2, queue_index); | |
3419 movl(Address(tmp2, 0), card_index); | |
3420 #endif | |
3421 jmp(done); | |
3422 | |
3423 bind(runtime); | |
3424 // save the live input values | |
3425 push(store_addr); | |
3426 push(new_val); | |
3427 #ifdef _LP64 | |
3428 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, r15_thread); | |
3429 #else | |
3430 push(thread); | |
3431 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread); | |
3432 pop(thread); | |
3433 #endif | |
3434 pop(new_val); | |
3435 pop(store_addr); | |
3436 | |
3437 bind(done); | |
3438 } | |
3439 | |
8001
db9981fd3124
8005915: Unify SERIALGC and INCLUDE_ALTERNATE_GCS
jprovino
parents:
7477
diff
changeset
|
3440 #endif // INCLUDE_ALL_GCS |
7199 | 3441 ////////////////////////////////////////////////////////////////////////////////// |
3442 | |
3443 | |
3444 void MacroAssembler::store_check(Register obj) { | |
3445 // Does a store check for the oop in register obj. The content of | |
3446 // register obj is destroyed afterwards. | |
3447 store_check_part_1(obj); | |
3448 store_check_part_2(obj); | |
3449 } | |
3450 | |
3451 void MacroAssembler::store_check(Register obj, Address dst) { | |
3452 store_check(obj); | |
3453 } | |
3454 | |
3455 | |
3456 // split the store check operation so that other instructions can be scheduled inbetween | |
3457 void MacroAssembler::store_check_part_1(Register obj) { | |
3458 BarrierSet* bs = Universe::heap()->barrier_set(); | |
3459 assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind"); | |
3460 shrptr(obj, CardTableModRefBS::card_shift); | |
3461 } | |
3462 | |
3463 void MacroAssembler::store_check_part_2(Register obj) { | |
3464 BarrierSet* bs = Universe::heap()->barrier_set(); | |
3465 assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind"); | |
3466 CardTableModRefBS* ct = (CardTableModRefBS*)bs; | |
3467 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code"); | |
3468 | |
3469 // The calculation for byte_map_base is as follows: | |
3470 // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift); | |
3471 // So this essentially converts an address to a displacement and | |
3472 // it will never need to be relocated. On 64bit however the value may be too | |
3473 // large for a 32bit displacement | |
3474 | |
3475 intptr_t disp = (intptr_t) ct->byte_map_base; | |
3476 if (is_simm32(disp)) { | |
3477 Address cardtable(noreg, obj, Address::times_1, disp); | |
3478 movb(cardtable, 0); | |
3479 } else { | |
3480 // By doing it as an ExternalAddress disp could be converted to a rip-relative | |
3481 // displacement and done in a single instruction given favorable mapping and | |
3482 // a smarter version of as_Address. Worst case it is two instructions which | |
3483 // is no worse off then loading disp into a register and doing as a simple | |
3484 // Address() as above. | |
3485 // We can't do as ExternalAddress as the only style since if disp == 0 we'll | |
3486 // assert since NULL isn't acceptable in a reloci (see 6644928). In any case | |
3487 // in some cases we'll get a single instruction version. | |
3488 | |
3489 ExternalAddress cardtable((address)disp); | |
3490 Address index(noreg, obj, Address::times_1); | |
3491 movb(as_Address(ArrayAddress(cardtable, index)), 0); | |
3492 } | |
3493 } | |
3494 | |
3495 void MacroAssembler::subptr(Register dst, int32_t imm32) { | |
3496 LP64_ONLY(subq(dst, imm32)) NOT_LP64(subl(dst, imm32)); | |
3497 } | |
3498 | |
3499 // Force generation of a 4 byte immediate value even if it fits into 8bit | |
3500 void MacroAssembler::subptr_imm32(Register dst, int32_t imm32) { | |
3501 LP64_ONLY(subq_imm32(dst, imm32)) NOT_LP64(subl_imm32(dst, imm32)); | |
3502 } | |
3503 | |
3504 void MacroAssembler::subptr(Register dst, Register src) { | |
3505 LP64_ONLY(subq(dst, src)) NOT_LP64(subl(dst, src)); | |
3506 } | |
3507 | |
3508 // C++ bool manipulation | |
3509 void MacroAssembler::testbool(Register dst) { | |
3510 if(sizeof(bool) == 1) | |
3511 testb(dst, 0xff); | |
3512 else if(sizeof(bool) == 2) { | |
3513 // testw implementation needed for two byte bools | |
3514 ShouldNotReachHere(); | |
3515 } else if(sizeof(bool) == 4) | |
3516 testl(dst, dst); | |
3517 else | |
3518 // unsupported | |
3519 ShouldNotReachHere(); | |
3520 } | |
3521 | |
3522 void MacroAssembler::testptr(Register dst, Register src) { | |
3523 LP64_ONLY(testq(dst, src)) NOT_LP64(testl(dst, src)); | |
3524 } | |
3525 | |
3526 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes. | |
3527 void MacroAssembler::tlab_allocate(Register obj, | |
3528 Register var_size_in_bytes, | |
3529 int con_size_in_bytes, | |
3530 Register t1, | |
3531 Register t2, | |
3532 Label& slow_case) { | |
3533 assert_different_registers(obj, t1, t2); | |
3534 assert_different_registers(obj, var_size_in_bytes, t1); | |
3535 Register end = t2; | |
3536 Register thread = NOT_LP64(t1) LP64_ONLY(r15_thread); | |
3537 | |
3538 verify_tlab(); | |
3539 | |
3540 NOT_LP64(get_thread(thread)); | |
3541 | |
3542 movptr(obj, Address(thread, JavaThread::tlab_top_offset())); | |
3543 if (var_size_in_bytes == noreg) { | |
3544 lea(end, Address(obj, con_size_in_bytes)); | |
3545 } else { | |
3546 lea(end, Address(obj, var_size_in_bytes, Address::times_1)); | |
3547 } | |
3548 cmpptr(end, Address(thread, JavaThread::tlab_end_offset())); | |
3549 jcc(Assembler::above, slow_case); | |
3550 | |
3551 // update the tlab top pointer | |
3552 movptr(Address(thread, JavaThread::tlab_top_offset()), end); | |
3553 | |
3554 // recover var_size_in_bytes if necessary | |
3555 if (var_size_in_bytes == end) { | |
3556 subptr(var_size_in_bytes, obj); | |
3557 } | |
3558 verify_tlab(); | |
3559 } | |
3560 | |
3561 // Preserves rbx, and rdx. | |
3562 Register MacroAssembler::tlab_refill(Label& retry, | |
3563 Label& try_eden, | |
3564 Label& slow_case) { | |
3565 Register top = rax; | |
3566 Register t1 = rcx; | |
3567 Register t2 = rsi; | |
3568 Register thread_reg = NOT_LP64(rdi) LP64_ONLY(r15_thread); | |
3569 assert_different_registers(top, thread_reg, t1, t2, /* preserve: */ rbx, rdx); | |
3570 Label do_refill, discard_tlab; | |
3571 | |
3572 if (CMSIncrementalMode || !Universe::heap()->supports_inline_contig_alloc()) { | |
3573 // No allocation in the shared eden. | |
3574 jmp(slow_case); | |
3575 } | |
3576 | |
3577 NOT_LP64(get_thread(thread_reg)); | |
3578 | |
3579 movptr(top, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); | |
3580 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); | |
3581 | |
3582 // calculate amount of free space | |
3583 subptr(t1, top); | |
3584 shrptr(t1, LogHeapWordSize); | |
3585 | |
3586 // Retain tlab and allocate object in shared space if | |
3587 // the amount free in the tlab is too large to discard. | |
3588 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset()))); | |
3589 jcc(Assembler::lessEqual, discard_tlab); | |
3590 | |
3591 // Retain | |
3592 // %%% yuck as movptr... | |
3593 movptr(t2, (int32_t) ThreadLocalAllocBuffer::refill_waste_limit_increment()); | |
3594 addptr(Address(thread_reg, in_bytes(JavaThread::tlab_refill_waste_limit_offset())), t2); | |
3595 if (TLABStats) { | |
3596 // increment number of slow_allocations | |
3597 addl(Address(thread_reg, in_bytes(JavaThread::tlab_slow_allocations_offset())), 1); | |
3598 } | |
3599 jmp(try_eden); | |
3600 | |
3601 bind(discard_tlab); | |
3602 if (TLABStats) { | |
3603 // increment number of refills | |
3604 addl(Address(thread_reg, in_bytes(JavaThread::tlab_number_of_refills_offset())), 1); | |
3605 // accumulate wastage -- t1 is amount free in tlab | |
3606 addl(Address(thread_reg, in_bytes(JavaThread::tlab_fast_refill_waste_offset())), t1); | |
3607 } | |
3608 | |
3609 // if tlab is currently allocated (top or end != null) then | |
3610 // fill [top, end + alignment_reserve) with array object | |
3611 testptr(top, top); | |
3612 jcc(Assembler::zero, do_refill); | |
3613 | |
3614 // set up the mark word | |
3615 movptr(Address(top, oopDesc::mark_offset_in_bytes()), (intptr_t)markOopDesc::prototype()->copy_set_hash(0x2)); | |
3616 // set the length to the remaining space | |
3617 subptr(t1, typeArrayOopDesc::header_size(T_INT)); | |
3618 addptr(t1, (int32_t)ThreadLocalAllocBuffer::alignment_reserve()); | |
3619 shlptr(t1, log2_intptr(HeapWordSize/sizeof(jint))); | |
3620 movl(Address(top, arrayOopDesc::length_offset_in_bytes()), t1); | |
3621 // set klass to intArrayKlass | |
3622 // dubious reloc why not an oop reloc? | |
3623 movptr(t1, ExternalAddress((address)Universe::intArrayKlassObj_addr())); | |
3624 // store klass last. concurrent gcs assumes klass length is valid if | |
3625 // klass field is not null. | |
3626 store_klass(top, t1); | |
3627 | |
3628 movptr(t1, top); | |
3629 subptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); | |
3630 incr_allocated_bytes(thread_reg, t1, 0); | |
3631 | |
3632 // refill the tlab with an eden allocation | |
3633 bind(do_refill); | |
3634 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset()))); | |
3635 shlptr(t1, LogHeapWordSize); | |
3636 // allocate new tlab, address returned in top | |
3637 eden_allocate(top, t1, 0, t2, slow_case); | |
3638 | |
3639 // Check that t1 was preserved in eden_allocate. | |
3640 #ifdef ASSERT | |
3641 if (UseTLAB) { | |
3642 Label ok; | |
3643 Register tsize = rsi; | |
3644 assert_different_registers(tsize, thread_reg, t1); | |
3645 push(tsize); | |
3646 movptr(tsize, Address(thread_reg, in_bytes(JavaThread::tlab_size_offset()))); | |
3647 shlptr(tsize, LogHeapWordSize); | |
3648 cmpptr(t1, tsize); | |
3649 jcc(Assembler::equal, ok); | |
3650 STOP("assert(t1 != tlab size)"); | |
3651 should_not_reach_here(); | |
3652 | |
3653 bind(ok); | |
3654 pop(tsize); | |
3655 } | |
3656 #endif | |
3657 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_start_offset())), top); | |
3658 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_top_offset())), top); | |
3659 addptr(top, t1); | |
3660 subptr(top, (int32_t)ThreadLocalAllocBuffer::alignment_reserve_in_bytes()); | |
3661 movptr(Address(thread_reg, in_bytes(JavaThread::tlab_end_offset())), top); | |
3662 verify_tlab(); | |
3663 jmp(retry); | |
3664 | |
3665 return thread_reg; // for use by caller | |
3666 } | |
3667 | |
3668 void MacroAssembler::incr_allocated_bytes(Register thread, | |
3669 Register var_size_in_bytes, | |
3670 int con_size_in_bytes, | |
3671 Register t1) { | |
3672 if (!thread->is_valid()) { | |
3673 #ifdef _LP64 | |
3674 thread = r15_thread; | |
3675 #else | |
3676 assert(t1->is_valid(), "need temp reg"); | |
3677 thread = t1; | |
3678 get_thread(thread); | |
3679 #endif | |
3680 } | |
3681 | |
3682 #ifdef _LP64 | |
3683 if (var_size_in_bytes->is_valid()) { | |
3684 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); | |
3685 } else { | |
3686 addq(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); | |
3687 } | |
3688 #else | |
3689 if (var_size_in_bytes->is_valid()) { | |
3690 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), var_size_in_bytes); | |
3691 } else { | |
3692 addl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())), con_size_in_bytes); | |
3693 } | |
3694 adcl(Address(thread, in_bytes(JavaThread::allocated_bytes_offset())+4), 0); | |
3695 #endif | |
3696 } | |
3697 | |
3698 void MacroAssembler::fp_runtime_fallback(address runtime_entry, int nb_args, int num_fpu_regs_in_use) { | |
3699 pusha(); | |
3700 | |
3701 // if we are coming from c1, xmm registers may be live | |
3702 int off = 0; | |
3703 if (UseSSE == 1) { | |
3704 subptr(rsp, sizeof(jdouble)*8); | |
3705 movflt(Address(rsp,off++*sizeof(jdouble)),xmm0); | |
3706 movflt(Address(rsp,off++*sizeof(jdouble)),xmm1); | |
3707 movflt(Address(rsp,off++*sizeof(jdouble)),xmm2); | |
3708 movflt(Address(rsp,off++*sizeof(jdouble)),xmm3); | |
3709 movflt(Address(rsp,off++*sizeof(jdouble)),xmm4); | |
3710 movflt(Address(rsp,off++*sizeof(jdouble)),xmm5); | |
3711 movflt(Address(rsp,off++*sizeof(jdouble)),xmm6); | |
3712 movflt(Address(rsp,off++*sizeof(jdouble)),xmm7); | |
3713 } else if (UseSSE >= 2) { | |
3714 #ifdef COMPILER2 | |
3715 if (MaxVectorSize > 16) { | |
3716 assert(UseAVX > 0, "256bit vectors are supported only with AVX"); | |
3717 // Save upper half of YMM registes | |
3718 subptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8)); | |
3719 vextractf128h(Address(rsp, 0),xmm0); | |
3720 vextractf128h(Address(rsp, 16),xmm1); | |
3721 vextractf128h(Address(rsp, 32),xmm2); | |
3722 vextractf128h(Address(rsp, 48),xmm3); | |
3723 vextractf128h(Address(rsp, 64),xmm4); | |
3724 vextractf128h(Address(rsp, 80),xmm5); | |
3725 vextractf128h(Address(rsp, 96),xmm6); | |
3726 vextractf128h(Address(rsp,112),xmm7); | |
3727 #ifdef _LP64 | |
3728 vextractf128h(Address(rsp,128),xmm8); | |
3729 vextractf128h(Address(rsp,144),xmm9); | |
3730 vextractf128h(Address(rsp,160),xmm10); | |
3731 vextractf128h(Address(rsp,176),xmm11); | |
3732 vextractf128h(Address(rsp,192),xmm12); | |
3733 vextractf128h(Address(rsp,208),xmm13); | |
3734 vextractf128h(Address(rsp,224),xmm14); | |
3735 vextractf128h(Address(rsp,240),xmm15); | |
3736 #endif | |
3737 } | |
3738 #endif | |
3739 // Save whole 128bit (16 bytes) XMM regiters | |
3740 subptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8)); | |
3741 movdqu(Address(rsp,off++*16),xmm0); | |
3742 movdqu(Address(rsp,off++*16),xmm1); | |
3743 movdqu(Address(rsp,off++*16),xmm2); | |
3744 movdqu(Address(rsp,off++*16),xmm3); | |
3745 movdqu(Address(rsp,off++*16),xmm4); | |
3746 movdqu(Address(rsp,off++*16),xmm5); | |
3747 movdqu(Address(rsp,off++*16),xmm6); | |
3748 movdqu(Address(rsp,off++*16),xmm7); | |
3749 #ifdef _LP64 | |
3750 movdqu(Address(rsp,off++*16),xmm8); | |
3751 movdqu(Address(rsp,off++*16),xmm9); | |
3752 movdqu(Address(rsp,off++*16),xmm10); | |
3753 movdqu(Address(rsp,off++*16),xmm11); | |
3754 movdqu(Address(rsp,off++*16),xmm12); | |
3755 movdqu(Address(rsp,off++*16),xmm13); | |
3756 movdqu(Address(rsp,off++*16),xmm14); | |
3757 movdqu(Address(rsp,off++*16),xmm15); | |
3758 #endif | |
3759 } | |
3760 | |
3761 // Preserve registers across runtime call | |
3762 int incoming_argument_and_return_value_offset = -1; | |
3763 if (num_fpu_regs_in_use > 1) { | |
3764 // Must preserve all other FPU regs (could alternatively convert | |
3765 // SharedRuntime::dsin, dcos etc. into assembly routines known not to trash | |
3766 // FPU state, but can not trust C compiler) | |
3767 NEEDS_CLEANUP; | |
3768 // NOTE that in this case we also push the incoming argument(s) to | |
3769 // the stack and restore it later; we also use this stack slot to | |
3770 // hold the return value from dsin, dcos etc. | |
3771 for (int i = 0; i < num_fpu_regs_in_use; i++) { | |
3772 subptr(rsp, sizeof(jdouble)); | |
3773 fstp_d(Address(rsp, 0)); | |
3774 } | |
3775 incoming_argument_and_return_value_offset = sizeof(jdouble)*(num_fpu_regs_in_use-1); | |
3776 for (int i = nb_args-1; i >= 0; i--) { | |
3777 fld_d(Address(rsp, incoming_argument_and_return_value_offset-i*sizeof(jdouble))); | |
3778 } | |
3779 } | |
3780 | |
3781 subptr(rsp, nb_args*sizeof(jdouble)); | |
3782 for (int i = 0; i < nb_args; i++) { | |
3783 fstp_d(Address(rsp, i*sizeof(jdouble))); | |
3784 } | |
3785 | |
3786 #ifdef _LP64 | |
3787 if (nb_args > 0) { | |
3788 movdbl(xmm0, Address(rsp, 0)); | |
3789 } | |
3790 if (nb_args > 1) { | |
3791 movdbl(xmm1, Address(rsp, sizeof(jdouble))); | |
3792 } | |
3793 assert(nb_args <= 2, "unsupported number of args"); | |
3794 #endif // _LP64 | |
3795 | |
3796 // NOTE: we must not use call_VM_leaf here because that requires a | |
3797 // complete interpreter frame in debug mode -- same bug as 4387334 | |
3798 // MacroAssembler::call_VM_leaf_base is perfectly safe and will | |
3799 // do proper 64bit abi | |
3800 | |
3801 NEEDS_CLEANUP; | |
3802 // Need to add stack banging before this runtime call if it needs to | |
3803 // be taken; however, there is no generic stack banging routine at | |
3804 // the MacroAssembler level | |
3805 | |
3806 MacroAssembler::call_VM_leaf_base(runtime_entry, 0); | |
3807 | |
3808 #ifdef _LP64 | |
3809 movsd(Address(rsp, 0), xmm0); | |
3810 fld_d(Address(rsp, 0)); | |
3811 #endif // _LP64 | |
3812 addptr(rsp, sizeof(jdouble) * nb_args); | |
3813 if (num_fpu_regs_in_use > 1) { | |
3814 // Must save return value to stack and then restore entire FPU | |
3815 // stack except incoming arguments | |
3816 fstp_d(Address(rsp, incoming_argument_and_return_value_offset)); | |
3817 for (int i = 0; i < num_fpu_regs_in_use - nb_args; i++) { | |
3818 fld_d(Address(rsp, 0)); | |
3819 addptr(rsp, sizeof(jdouble)); | |
3820 } | |
3821 fld_d(Address(rsp, (nb_args-1)*sizeof(jdouble))); | |
3822 addptr(rsp, sizeof(jdouble) * nb_args); | |
3823 } | |
3824 | |
3825 off = 0; | |
3826 if (UseSSE == 1) { | |
3827 movflt(xmm0, Address(rsp,off++*sizeof(jdouble))); | |
3828 movflt(xmm1, Address(rsp,off++*sizeof(jdouble))); | |
3829 movflt(xmm2, Address(rsp,off++*sizeof(jdouble))); | |
3830 movflt(xmm3, Address(rsp,off++*sizeof(jdouble))); | |
3831 movflt(xmm4, Address(rsp,off++*sizeof(jdouble))); | |
3832 movflt(xmm5, Address(rsp,off++*sizeof(jdouble))); | |
3833 movflt(xmm6, Address(rsp,off++*sizeof(jdouble))); | |
3834 movflt(xmm7, Address(rsp,off++*sizeof(jdouble))); | |
3835 addptr(rsp, sizeof(jdouble)*8); | |
3836 } else if (UseSSE >= 2) { | |
3837 // Restore whole 128bit (16 bytes) XMM regiters | |
3838 movdqu(xmm0, Address(rsp,off++*16)); | |
3839 movdqu(xmm1, Address(rsp,off++*16)); | |
3840 movdqu(xmm2, Address(rsp,off++*16)); | |
3841 movdqu(xmm3, Address(rsp,off++*16)); | |
3842 movdqu(xmm4, Address(rsp,off++*16)); | |
3843 movdqu(xmm5, Address(rsp,off++*16)); | |
3844 movdqu(xmm6, Address(rsp,off++*16)); | |
3845 movdqu(xmm7, Address(rsp,off++*16)); | |
3846 #ifdef _LP64 | |
3847 movdqu(xmm8, Address(rsp,off++*16)); | |
3848 movdqu(xmm9, Address(rsp,off++*16)); | |
3849 movdqu(xmm10, Address(rsp,off++*16)); | |
3850 movdqu(xmm11, Address(rsp,off++*16)); | |
3851 movdqu(xmm12, Address(rsp,off++*16)); | |
3852 movdqu(xmm13, Address(rsp,off++*16)); | |
3853 movdqu(xmm14, Address(rsp,off++*16)); | |
3854 movdqu(xmm15, Address(rsp,off++*16)); | |
3855 #endif | |
3856 addptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8)); | |
3857 #ifdef COMPILER2 | |
3858 if (MaxVectorSize > 16) { | |
3859 // Restore upper half of YMM registes. | |
3860 vinsertf128h(xmm0, Address(rsp, 0)); | |
3861 vinsertf128h(xmm1, Address(rsp, 16)); | |
3862 vinsertf128h(xmm2, Address(rsp, 32)); | |
3863 vinsertf128h(xmm3, Address(rsp, 48)); | |
3864 vinsertf128h(xmm4, Address(rsp, 64)); | |
3865 vinsertf128h(xmm5, Address(rsp, 80)); | |
3866 vinsertf128h(xmm6, Address(rsp, 96)); | |
3867 vinsertf128h(xmm7, Address(rsp,112)); | |
3868 #ifdef _LP64 | |
3869 vinsertf128h(xmm8, Address(rsp,128)); | |
3870 vinsertf128h(xmm9, Address(rsp,144)); | |
3871 vinsertf128h(xmm10, Address(rsp,160)); | |
3872 vinsertf128h(xmm11, Address(rsp,176)); | |
3873 vinsertf128h(xmm12, Address(rsp,192)); | |
3874 vinsertf128h(xmm13, Address(rsp,208)); | |
3875 vinsertf128h(xmm14, Address(rsp,224)); | |
3876 vinsertf128h(xmm15, Address(rsp,240)); | |
3877 #endif | |
3878 addptr(rsp, 16 * LP64_ONLY(16) NOT_LP64(8)); | |
3879 } | |
3880 #endif | |
3881 } | |
3882 popa(); | |
3883 } | |
3884 | |
3885 static const double pi_4 = 0.7853981633974483; | |
3886 | |
3887 void MacroAssembler::trigfunc(char trig, int num_fpu_regs_in_use) { | |
3888 // A hand-coded argument reduction for values in fabs(pi/4, pi/2) | |
3889 // was attempted in this code; unfortunately it appears that the | |
3890 // switch to 80-bit precision and back causes this to be | |
3891 // unprofitable compared with simply performing a runtime call if | |
3892 // the argument is out of the (-pi/4, pi/4) range. | |
3893 | |
3894 Register tmp = noreg; | |
3895 if (!VM_Version::supports_cmov()) { | |
3896 // fcmp needs a temporary so preserve rbx, | |
3897 tmp = rbx; | |
3898 push(tmp); | |
3899 } | |
3900 | |
3901 Label slow_case, done; | |
3902 | |
3903 ExternalAddress pi4_adr = (address)&pi_4; | |
3904 if (reachable(pi4_adr)) { | |
3905 // x ?<= pi/4 | |
3906 fld_d(pi4_adr); | |
3907 fld_s(1); // Stack: X PI/4 X | |
3908 fabs(); // Stack: |X| PI/4 X | |
3909 fcmp(tmp); | |
3910 jcc(Assembler::above, slow_case); | |
3911 | |
3912 // fastest case: -pi/4 <= x <= pi/4 | |
3913 switch(trig) { | |
3914 case 's': | |
3915 fsin(); | |
3916 break; | |
3917 case 'c': | |
3918 fcos(); | |
3919 break; | |
3920 case 't': | |
3921 ftan(); | |
3922 break; | |
3923 default: | |
3924 assert(false, "bad intrinsic"); | |
3925 break; | |
3926 } | |
3927 jmp(done); | |
3928 } | |
3929 | |
3930 // slow case: runtime call | |
3931 bind(slow_case); | |
3932 | |
3933 switch(trig) { | |
3934 case 's': | |
3935 { | |
3936 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1, num_fpu_regs_in_use); | |
3937 } | |
3938 break; | |
3939 case 'c': | |
3940 { | |
3941 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1, num_fpu_regs_in_use); | |
3942 } | |
3943 break; | |
3944 case 't': | |
3945 { | |
3946 fp_runtime_fallback(CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1, num_fpu_regs_in_use); | |
3947 } | |
3948 break; | |
3949 default: | |
3950 assert(false, "bad intrinsic"); | |
3951 break; | |
3952 } | |
3953 | |
3954 // Come here with result in F-TOS | |
3955 bind(done); | |
3956 | |
3957 if (tmp != noreg) { | |
3958 pop(tmp); | |
3959 } | |
3960 } | |
3961 | |
3962 | |
3963 // Look up the method for a megamorphic invokeinterface call. | |
3964 // The target method is determined by <intf_klass, itable_index>. | |
3965 // The receiver klass is in recv_klass. | |
3966 // On success, the result will be in method_result, and execution falls through. | |
3967 // On failure, execution transfers to the given label. | |
3968 void MacroAssembler::lookup_interface_method(Register recv_klass, | |
3969 Register intf_klass, | |
3970 RegisterOrConstant itable_index, | |
3971 Register method_result, | |
3972 Register scan_temp, | |
3973 Label& L_no_such_interface) { | |
3974 assert_different_registers(recv_klass, intf_klass, method_result, scan_temp); | |
3975 assert(itable_index.is_constant() || itable_index.as_register() == method_result, | |
3976 "caller must use same register for non-constant itable index as for method"); | |
3977 | |
3978 // Compute start of first itableOffsetEntry (which is at the end of the vtable) | |
3979 int vtable_base = InstanceKlass::vtable_start_offset() * wordSize; | |
3980 int itentry_off = itableMethodEntry::method_offset_in_bytes(); | |
3981 int scan_step = itableOffsetEntry::size() * wordSize; | |
3982 int vte_size = vtableEntry::size() * wordSize; | |
3983 Address::ScaleFactor times_vte_scale = Address::times_ptr; | |
3984 assert(vte_size == wordSize, "else adjust times_vte_scale"); | |
3985 | |
3986 movl(scan_temp, Address(recv_klass, InstanceKlass::vtable_length_offset() * wordSize)); | |
3987 | |
3988 // %%% Could store the aligned, prescaled offset in the klassoop. | |
3989 lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base)); | |
3990 if (HeapWordsPerLong > 1) { | |
3991 // Round up to align_object_offset boundary | |
3992 // see code for InstanceKlass::start_of_itable! | |
3993 round_to(scan_temp, BytesPerLong); | |
3994 } | |
3995 | |
3996 // Adjust recv_klass by scaled itable_index, so we can free itable_index. | |
3997 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below"); | |
3998 lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off)); | |
3999 | |
4000 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) { | |
4001 // if (scan->interface() == intf) { | |
4002 // result = (klass + scan->offset() + itable_index); | |
4003 // } | |
4004 // } | |
4005 Label search, found_method; | |
4006 | |
4007 for (int peel = 1; peel >= 0; peel--) { | |
4008 movptr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes())); | |
4009 cmpptr(intf_klass, method_result); | |
4010 | |
4011 if (peel) { | |
4012 jccb(Assembler::equal, found_method); | |
4013 } else { | |
4014 jccb(Assembler::notEqual, search); | |
4015 // (invert the test to fall through to found_method...) | |
4016 } | |
4017 | |
4018 if (!peel) break; | |
4019 | |
4020 bind(search); | |
4021 | |
4022 // Check that the previous entry is non-null. A null entry means that | |
4023 // the receiver class doesn't implement the interface, and wasn't the | |
4024 // same as when the caller was compiled. | |
4025 testptr(method_result, method_result); | |
4026 jcc(Assembler::zero, L_no_such_interface); | |
4027 addptr(scan_temp, scan_step); | |
4028 } | |
4029 | |
4030 bind(found_method); | |
4031 | |
4032 // Got a hit. | |
4033 movl(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes())); | |
4034 movptr(method_result, Address(recv_klass, scan_temp, Address::times_1)); | |
4035 } | |
4036 | |
4037 | |
4038 // virtual method calling | |
4039 void MacroAssembler::lookup_virtual_method(Register recv_klass, | |
4040 RegisterOrConstant vtable_index, | |
4041 Register method_result) { | |
4042 const int base = InstanceKlass::vtable_start_offset() * wordSize; | |
4043 assert(vtableEntry::size() * wordSize == wordSize, "else adjust the scaling in the code below"); | |
4044 Address vtable_entry_addr(recv_klass, | |
4045 vtable_index, Address::times_ptr, | |
4046 base + vtableEntry::method_offset_in_bytes()); | |
4047 movptr(method_result, vtable_entry_addr); | |
4048 } | |
4049 | |
4050 | |
4051 void MacroAssembler::check_klass_subtype(Register sub_klass, | |
4052 Register super_klass, | |
4053 Register temp_reg, | |
4054 Label& L_success) { | |
4055 Label L_failure; | |
4056 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL); | |
4057 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL); | |
4058 bind(L_failure); | |
4059 } | |
4060 | |
4061 | |
4062 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass, | |
4063 Register super_klass, | |
4064 Register temp_reg, | |
4065 Label* L_success, | |
4066 Label* L_failure, | |
4067 Label* L_slow_path, | |
4068 RegisterOrConstant super_check_offset) { | |
4069 assert_different_registers(sub_klass, super_klass, temp_reg); | |
4070 bool must_load_sco = (super_check_offset.constant_or_zero() == -1); | |
4071 if (super_check_offset.is_register()) { | |
4072 assert_different_registers(sub_klass, super_klass, | |
4073 super_check_offset.as_register()); | |
4074 } else if (must_load_sco) { | |
4075 assert(temp_reg != noreg, "supply either a temp or a register offset"); | |
4076 } | |
4077 | |
4078 Label L_fallthrough; | |
4079 int label_nulls = 0; | |
4080 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } | |
4081 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } | |
4082 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; } | |
4083 assert(label_nulls <= 1, "at most one NULL in the batch"); | |
4084 | |
4085 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); | |
4086 int sco_offset = in_bytes(Klass::super_check_offset_offset()); | |
4087 Address super_check_offset_addr(super_klass, sco_offset); | |
4088 | |
4089 // Hacked jcc, which "knows" that L_fallthrough, at least, is in | |
4090 // range of a jccb. If this routine grows larger, reconsider at | |
4091 // least some of these. | |
4092 #define local_jcc(assembler_cond, label) \ | |
4093 if (&(label) == &L_fallthrough) jccb(assembler_cond, label); \ | |
4094 else jcc( assembler_cond, label) /*omit semi*/ | |
4095 | |
4096 // Hacked jmp, which may only be used just before L_fallthrough. | |
4097 #define final_jmp(label) \ | |
4098 if (&(label) == &L_fallthrough) { /*do nothing*/ } \ | |
4099 else jmp(label) /*omit semi*/ | |
4100 | |
4101 // If the pointers are equal, we are done (e.g., String[] elements). | |
4102 // This self-check enables sharing of secondary supertype arrays among | |
4103 // non-primary types such as array-of-interface. Otherwise, each such | |
4104 // type would need its own customized SSA. | |
4105 // We move this check to the front of the fast path because many | |
4106 // type checks are in fact trivially successful in this manner, | |
4107 // so we get a nicely predicted branch right at the start of the check. | |
4108 cmpptr(sub_klass, super_klass); | |
4109 local_jcc(Assembler::equal, *L_success); | |
4110 | |
4111 // Check the supertype display: | |
4112 if (must_load_sco) { | |
4113 // Positive movl does right thing on LP64. | |
4114 movl(temp_reg, super_check_offset_addr); | |
4115 super_check_offset = RegisterOrConstant(temp_reg); | |
4116 } | |
4117 Address super_check_addr(sub_klass, super_check_offset, Address::times_1, 0); | |
4118 cmpptr(super_klass, super_check_addr); // load displayed supertype | |
4119 | |
4120 // This check has worked decisively for primary supers. | |
4121 // Secondary supers are sought in the super_cache ('super_cache_addr'). | |
4122 // (Secondary supers are interfaces and very deeply nested subtypes.) | |
4123 // This works in the same check above because of a tricky aliasing | |
4124 // between the super_cache and the primary super display elements. | |
4125 // (The 'super_check_addr' can address either, as the case requires.) | |
4126 // Note that the cache is updated below if it does not help us find | |
4127 // what we need immediately. | |
4128 // So if it was a primary super, we can just fail immediately. | |
4129 // Otherwise, it's the slow path for us (no success at this point). | |
4130 | |
4131 if (super_check_offset.is_register()) { | |
4132 local_jcc(Assembler::equal, *L_success); | |
4133 cmpl(super_check_offset.as_register(), sc_offset); | |
4134 if (L_failure == &L_fallthrough) { | |
4135 local_jcc(Assembler::equal, *L_slow_path); | |
4136 } else { | |
4137 local_jcc(Assembler::notEqual, *L_failure); | |
4138 final_jmp(*L_slow_path); | |
4139 } | |
4140 } else if (super_check_offset.as_constant() == sc_offset) { | |
4141 // Need a slow path; fast failure is impossible. | |
4142 if (L_slow_path == &L_fallthrough) { | |
4143 local_jcc(Assembler::equal, *L_success); | |
4144 } else { | |
4145 local_jcc(Assembler::notEqual, *L_slow_path); | |
4146 final_jmp(*L_success); | |
4147 } | |
4148 } else { | |
4149 // No slow path; it's a fast decision. | |
4150 if (L_failure == &L_fallthrough) { | |
4151 local_jcc(Assembler::equal, *L_success); | |
4152 } else { | |
4153 local_jcc(Assembler::notEqual, *L_failure); | |
4154 final_jmp(*L_success); | |
4155 } | |
4156 } | |
4157 | |
4158 bind(L_fallthrough); | |
4159 | |
4160 #undef local_jcc | |
4161 #undef final_jmp | |
4162 } | |
4163 | |
4164 | |
4165 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass, | |
4166 Register super_klass, | |
4167 Register temp_reg, | |
4168 Register temp2_reg, | |
4169 Label* L_success, | |
4170 Label* L_failure, | |
4171 bool set_cond_codes) { | |
4172 assert_different_registers(sub_klass, super_klass, temp_reg); | |
4173 if (temp2_reg != noreg) | |
4174 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg); | |
4175 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg) | |
4176 | |
4177 Label L_fallthrough; | |
4178 int label_nulls = 0; | |
4179 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; } | |
4180 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; } | |
4181 assert(label_nulls <= 1, "at most one NULL in the batch"); | |
4182 | |
4183 // a couple of useful fields in sub_klass: | |
4184 int ss_offset = in_bytes(Klass::secondary_supers_offset()); | |
4185 int sc_offset = in_bytes(Klass::secondary_super_cache_offset()); | |
4186 Address secondary_supers_addr(sub_klass, ss_offset); | |
4187 Address super_cache_addr( sub_klass, sc_offset); | |
4188 | |
4189 // Do a linear scan of the secondary super-klass chain. | |
4190 // This code is rarely used, so simplicity is a virtue here. | |
4191 // The repne_scan instruction uses fixed registers, which we must spill. | |
4192 // Don't worry too much about pre-existing connections with the input regs. | |
4193 | |
4194 assert(sub_klass != rax, "killed reg"); // killed by mov(rax, super) | |
4195 assert(sub_klass != rcx, "killed reg"); // killed by lea(rcx, &pst_counter) | |
4196 | |
4197 // Get super_klass value into rax (even if it was in rdi or rcx). | |
4198 bool pushed_rax = false, pushed_rcx = false, pushed_rdi = false; | |
4199 if (super_klass != rax || UseCompressedOops) { | |
4200 if (!IS_A_TEMP(rax)) { push(rax); pushed_rax = true; } | |
4201 mov(rax, super_klass); | |
4202 } | |
4203 if (!IS_A_TEMP(rcx)) { push(rcx); pushed_rcx = true; } | |
4204 if (!IS_A_TEMP(rdi)) { push(rdi); pushed_rdi = true; } | |
4205 | |
4206 #ifndef PRODUCT | |
4207 int* pst_counter = &SharedRuntime::_partial_subtype_ctr; | |
4208 ExternalAddress pst_counter_addr((address) pst_counter); | |
4209 NOT_LP64( incrementl(pst_counter_addr) ); | |
4210 LP64_ONLY( lea(rcx, pst_counter_addr) ); | |
4211 LP64_ONLY( incrementl(Address(rcx, 0)) ); | |
4212 #endif //PRODUCT | |
4213 | |
4214 // We will consult the secondary-super array. | |
4215 movptr(rdi, secondary_supers_addr); | |
4216 // Load the array length. (Positive movl does right thing on LP64.) | |
4217 movl(rcx, Address(rdi, Array<Klass*>::length_offset_in_bytes())); | |
4218 // Skip to start of data. | |
4219 addptr(rdi, Array<Klass*>::base_offset_in_bytes()); | |
4220 | |
4221 // Scan RCX words at [RDI] for an occurrence of RAX. | |
4222 // Set NZ/Z based on last compare. | |
4223 // Z flag value will not be set by 'repne' if RCX == 0 since 'repne' does | |
4224 // not change flags (only scas instruction which is repeated sets flags). | |
4225 // Set Z = 0 (not equal) before 'repne' to indicate that class was not found. | |
4226 | |
4227 testptr(rax,rax); // Set Z = 0 | |
4228 repne_scan(); | |
4229 | |
4230 // Unspill the temp. registers: | |
4231 if (pushed_rdi) pop(rdi); | |
4232 if (pushed_rcx) pop(rcx); | |
4233 if (pushed_rax) pop(rax); | |
4234 | |
4235 if (set_cond_codes) { | |
4236 // Special hack for the AD files: rdi is guaranteed non-zero. | |
4237 assert(!pushed_rdi, "rdi must be left non-NULL"); | |
4238 // Also, the condition codes are properly set Z/NZ on succeed/failure. | |
4239 } | |
4240 | |
4241 if (L_failure == &L_fallthrough) | |
4242 jccb(Assembler::notEqual, *L_failure); | |
4243 else jcc(Assembler::notEqual, *L_failure); | |
4244 | |
4245 // Success. Cache the super we found and proceed in triumph. | |
4246 movptr(super_cache_addr, super_klass); | |
4247 | |
4248 if (L_success != &L_fallthrough) { | |
4249 jmp(*L_success); | |
4250 } | |
4251 | |
4252 #undef IS_A_TEMP | |
4253 | |
4254 bind(L_fallthrough); | |
4255 } | |
4256 | |
4257 | |
4258 void MacroAssembler::cmov32(Condition cc, Register dst, Address src) { | |
4259 if (VM_Version::supports_cmov()) { | |
4260 cmovl(cc, dst, src); | |
4261 } else { | |
4262 Label L; | |
4263 jccb(negate_condition(cc), L); | |
4264 movl(dst, src); | |
4265 bind(L); | |
4266 } | |
4267 } | |
4268 | |
4269 void MacroAssembler::cmov32(Condition cc, Register dst, Register src) { | |
4270 if (VM_Version::supports_cmov()) { | |
4271 cmovl(cc, dst, src); | |
4272 } else { | |
4273 Label L; | |
4274 jccb(negate_condition(cc), L); | |
4275 movl(dst, src); | |
4276 bind(L); | |
4277 } | |
4278 } | |
4279 | |
4280 void MacroAssembler::verify_oop(Register reg, const char* s) { | |
4281 if (!VerifyOops) return; | |
4282 | |
4283 // Pass register number to verify_oop_subroutine | |
8767
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4284 const char* b = NULL; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4285 { |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4286 ResourceMark rm; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4287 stringStream ss; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4288 ss.print("verify_oop: %s: %s", reg->name(), s); |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4289 b = code_string(ss.as_string()); |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4290 } |
7199 | 4291 BLOCK_COMMENT("verify_oop {"); |
4292 #ifdef _LP64 | |
4293 push(rscratch1); // save r10, trashed by movptr() | |
4294 #endif | |
4295 push(rax); // save rax, | |
4296 push(reg); // pass register argument | |
4297 ExternalAddress buffer((address) b); | |
4298 // avoid using pushptr, as it modifies scratch registers | |
4299 // and our contract is not to modify anything | |
4300 movptr(rax, buffer.addr()); | |
4301 push(rax); | |
4302 // call indirectly to solve generation ordering problem | |
4303 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); | |
4304 call(rax); | |
4305 // Caller pops the arguments (oop, message) and restores rax, r10 | |
4306 BLOCK_COMMENT("} verify_oop"); | |
4307 } | |
4308 | |
4309 | |
4310 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr, | |
4311 Register tmp, | |
4312 int offset) { | |
4313 intptr_t value = *delayed_value_addr; | |
4314 if (value != 0) | |
4315 return RegisterOrConstant(value + offset); | |
4316 | |
4317 // load indirectly to solve generation ordering problem | |
4318 movptr(tmp, ExternalAddress((address) delayed_value_addr)); | |
4319 | |
4320 #ifdef ASSERT | |
4321 { Label L; | |
4322 testptr(tmp, tmp); | |
4323 if (WizardMode) { | |
8767
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4324 const char* buf = NULL; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4325 { |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4326 ResourceMark rm; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4327 stringStream ss; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4328 ss.print("DelayedValue="INTPTR_FORMAT, delayed_value_addr[1]); |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4329 buf = code_string(ss.as_string()); |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4330 } |
7199 | 4331 jcc(Assembler::notZero, L); |
4332 STOP(buf); | |
4333 } else { | |
4334 jccb(Assembler::notZero, L); | |
4335 hlt(); | |
4336 } | |
4337 bind(L); | |
4338 } | |
4339 #endif | |
4340 | |
4341 if (offset != 0) | |
4342 addptr(tmp, offset); | |
4343 | |
4344 return RegisterOrConstant(tmp); | |
4345 } | |
4346 | |
4347 | |
4348 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot, | |
4349 int extra_slot_offset) { | |
4350 // cf. TemplateTable::prepare_invoke(), if (load_receiver). | |
4351 int stackElementSize = Interpreter::stackElementSize; | |
4352 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0); | |
4353 #ifdef ASSERT | |
4354 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1); | |
4355 assert(offset1 - offset == stackElementSize, "correct arithmetic"); | |
4356 #endif | |
4357 Register scale_reg = noreg; | |
4358 Address::ScaleFactor scale_factor = Address::no_scale; | |
4359 if (arg_slot.is_constant()) { | |
4360 offset += arg_slot.as_constant() * stackElementSize; | |
4361 } else { | |
4362 scale_reg = arg_slot.as_register(); | |
4363 scale_factor = Address::times(stackElementSize); | |
4364 } | |
4365 offset += wordSize; // return PC is on stack | |
4366 return Address(rsp, scale_reg, scale_factor, offset); | |
4367 } | |
4368 | |
4369 | |
4370 void MacroAssembler::verify_oop_addr(Address addr, const char* s) { | |
4371 if (!VerifyOops) return; | |
4372 | |
4373 // Address adjust(addr.base(), addr.index(), addr.scale(), addr.disp() + BytesPerWord); | |
4374 // Pass register number to verify_oop_subroutine | |
8767
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4375 const char* b = NULL; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4376 { |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4377 ResourceMark rm; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4378 stringStream ss; |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4379 ss.print("verify_oop_addr: %s", s); |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4380 b = code_string(ss.as_string()); |
a5de0cc2f91c
8008555: Debugging code in compiled method sometimes leaks memory
roland
parents:
8042
diff
changeset
|
4381 } |
7199 | 4382 #ifdef _LP64 |
4383 push(rscratch1); // save r10, trashed by movptr() | |
4384 #endif | |
4385 push(rax); // save rax, | |
4386 // addr may contain rsp so we will have to adjust it based on the push | |
4387 // we just did (and on 64 bit we do two pushes) | |
4388 // NOTE: 64bit seemed to have had a bug in that it did movq(addr, rax); which | |
4389 // stores rax into addr which is backwards of what was intended. | |
4390 if (addr.uses(rsp)) { | |
4391 lea(rax, addr); | |
4392 pushptr(Address(rax, LP64_ONLY(2 *) BytesPerWord)); | |
4393 } else { | |
4394 pushptr(addr); | |
4395 } | |
4396 | |
4397 ExternalAddress buffer((address) b); | |
4398 // pass msg argument | |
4399 // avoid using pushptr, as it modifies scratch registers | |
4400 // and our contract is not to modify anything | |
4401 movptr(rax, buffer.addr()); | |
4402 push(rax); | |
4403 | |
4404 // call indirectly to solve generation ordering problem | |
4405 movptr(rax, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address())); | |
4406 call(rax); | |
4407 // Caller pops the arguments (addr, message) and restores rax, r10. | |
4408 } | |
4409 | |
4410 void MacroAssembler::verify_tlab() { | |
4411 #ifdef ASSERT | |
4412 if (UseTLAB && VerifyOops) { | |
4413 Label next, ok; | |
4414 Register t1 = rsi; | |
4415 Register thread_reg = NOT_LP64(rbx) LP64_ONLY(r15_thread); | |
4416 | |
4417 push(t1); | |
4418 NOT_LP64(push(thread_reg)); | |
4419 NOT_LP64(get_thread(thread_reg)); | |
4420 | |
4421 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); | |
4422 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_start_offset()))); | |
4423 jcc(Assembler::aboveEqual, next); | |
4424 STOP("assert(top >= start)"); | |
4425 should_not_reach_here(); | |
4426 | |
4427 bind(next); | |
4428 movptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_end_offset()))); | |
4429 cmpptr(t1, Address(thread_reg, in_bytes(JavaThread::tlab_top_offset()))); | |
4430 jcc(Assembler::aboveEqual, ok); | |
4431 STOP("assert(top <= end)"); | |
4432 should_not_reach_here(); | |
4433 | |
4434 bind(ok); | |
4435 NOT_LP64(pop(thread_reg)); | |
4436 pop(t1); | |
4437 } | |
4438 #endif | |
4439 } | |
4440 | |
4441 class ControlWord { | |
4442 public: | |
4443 int32_t _value; | |
4444 | |
4445 int rounding_control() const { return (_value >> 10) & 3 ; } | |
4446 int precision_control() const { return (_value >> 8) & 3 ; } | |
4447 bool precision() const { return ((_value >> 5) & 1) != 0; } | |
4448 bool underflow() const { return ((_value >> 4) & 1) != 0; } | |
4449 bool overflow() const { return ((_value >> 3) & 1) != 0; } | |
4450 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } | |
4451 bool denormalized() const { return ((_value >> 1) & 1) != 0; } | |
4452 bool invalid() const { return ((_value >> 0) & 1) != 0; } | |
4453 | |
4454 void print() const { | |
4455 // rounding control | |
4456 const char* rc; | |
4457 switch (rounding_control()) { | |
4458 case 0: rc = "round near"; break; | |
4459 case 1: rc = "round down"; break; | |
4460 case 2: rc = "round up "; break; | |
4461 case 3: rc = "chop "; break; | |
4462 }; | |
4463 // precision control | |
4464 const char* pc; | |
4465 switch (precision_control()) { | |
4466 case 0: pc = "24 bits "; break; | |
4467 case 1: pc = "reserved"; break; | |
4468 case 2: pc = "53 bits "; break; | |
4469 case 3: pc = "64 bits "; break; | |
4470 }; | |
4471 // flags | |
4472 char f[9]; | |
4473 f[0] = ' '; | |
4474 f[1] = ' '; | |
4475 f[2] = (precision ()) ? 'P' : 'p'; | |
4476 f[3] = (underflow ()) ? 'U' : 'u'; | |
4477 f[4] = (overflow ()) ? 'O' : 'o'; | |
4478 f[5] = (zero_divide ()) ? 'Z' : 'z'; | |
4479 f[6] = (denormalized()) ? 'D' : 'd'; | |
4480 f[7] = (invalid ()) ? 'I' : 'i'; | |
4481 f[8] = '\x0'; | |
4482 // output | |
4483 printf("%04x masks = %s, %s, %s", _value & 0xFFFF, f, rc, pc); | |
4484 } | |
4485 | |
4486 }; | |
4487 | |
4488 class StatusWord { | |
4489 public: | |
4490 int32_t _value; | |
4491 | |
4492 bool busy() const { return ((_value >> 15) & 1) != 0; } | |
4493 bool C3() const { return ((_value >> 14) & 1) != 0; } | |
4494 bool C2() const { return ((_value >> 10) & 1) != 0; } | |
4495 bool C1() const { return ((_value >> 9) & 1) != 0; } | |
4496 bool C0() const { return ((_value >> 8) & 1) != 0; } | |
4497 int top() const { return (_value >> 11) & 7 ; } | |
4498 bool error_status() const { return ((_value >> 7) & 1) != 0; } | |
4499 bool stack_fault() const { return ((_value >> 6) & 1) != 0; } | |
4500 bool precision() const { return ((_value >> 5) & 1) != 0; } | |
4501 bool underflow() const { return ((_value >> 4) & 1) != 0; } | |
4502 bool overflow() const { return ((_value >> 3) & 1) != 0; } | |
4503 bool zero_divide() const { return ((_value >> 2) & 1) != 0; } | |
4504 bool denormalized() const { return ((_value >> 1) & 1) != 0; } | |
4505 bool invalid() const { return ((_value >> 0) & 1) != 0; } | |
4506 | |
4507 void print() const { | |
4508 // condition codes | |
4509 char c[5]; | |
4510 c[0] = (C3()) ? '3' : '-'; | |
4511 c[1] = (C2()) ? '2' : '-'; | |
4512 c[2] = (C1()) ? '1' : '-'; | |
4513 c[3] = (C0()) ? '0' : '-'; | |
4514 c[4] = '\x0'; | |
4515 // flags | |
4516 char f[9]; | |
4517 f[0] = (error_status()) ? 'E' : '-'; | |
4518 f[1] = (stack_fault ()) ? 'S' : '-'; | |
4519 f[2] = (precision ()) ? 'P' : '-'; | |
4520 f[3] = (underflow ()) ? 'U' : '-'; | |
4521 f[4] = (overflow ()) ? 'O' : '-'; | |
4522 f[5] = (zero_divide ()) ? 'Z' : '-'; | |
4523 f[6] = (denormalized()) ? 'D' : '-'; | |
4524 f[7] = (invalid ()) ? 'I' : '-'; | |
4525 f[8] = '\x0'; | |
4526 // output | |
4527 printf("%04x flags = %s, cc = %s, top = %d", _value & 0xFFFF, f, c, top()); | |
4528 } | |
4529 | |
4530 }; | |
4531 | |
4532 class TagWord { | |
4533 public: | |
4534 int32_t _value; | |
4535 | |
4536 int tag_at(int i) const { return (_value >> (i*2)) & 3; } | |
4537 | |
4538 void print() const { | |
4539 printf("%04x", _value & 0xFFFF); | |
4540 } | |
4541 | |
4542 }; | |
4543 | |
4544 class FPU_Register { | |
4545 public: | |
4546 int32_t _m0; | |
4547 int32_t _m1; | |
4548 int16_t _ex; | |
4549 | |
4550 bool is_indefinite() const { | |
4551 return _ex == -1 && _m1 == (int32_t)0xC0000000 && _m0 == 0; | |
4552 } | |
4553 | |
4554 void print() const { | |
4555 char sign = (_ex < 0) ? '-' : '+'; | |
4556 const char* kind = (_ex == 0x7FFF || _ex == (int16_t)-1) ? "NaN" : " "; | |
4557 printf("%c%04hx.%08x%08x %s", sign, _ex, _m1, _m0, kind); | |
4558 }; | |
4559 | |
4560 }; | |
4561 | |
4562 class FPU_State { | |
4563 public: | |
4564 enum { | |
4565 register_size = 10, | |
4566 number_of_registers = 8, | |
4567 register_mask = 7 | |
4568 }; | |
4569 | |
4570 ControlWord _control_word; | |
4571 StatusWord _status_word; | |
4572 TagWord _tag_word; | |
4573 int32_t _error_offset; | |
4574 int32_t _error_selector; | |
4575 int32_t _data_offset; | |
4576 int32_t _data_selector; | |
4577 int8_t _register[register_size * number_of_registers]; | |
4578 | |
4579 int tag_for_st(int i) const { return _tag_word.tag_at((_status_word.top() + i) & register_mask); } | |
4580 FPU_Register* st(int i) const { return (FPU_Register*)&_register[register_size * i]; } | |
4581 | |
4582 const char* tag_as_string(int tag) const { | |
4583 switch (tag) { | |
4584 case 0: return "valid"; | |
4585 case 1: return "zero"; | |
4586 case 2: return "special"; | |
4587 case 3: return "empty"; | |
4588 } | |
4589 ShouldNotReachHere(); | |
4590 return NULL; | |
4591 } | |
4592 | |
4593 void print() const { | |
4594 // print computation registers | |
4595 { int t = _status_word.top(); | |
4596 for (int i = 0; i < number_of_registers; i++) { | |
4597 int j = (i - t) & register_mask; | |
4598 printf("%c r%d = ST%d = ", (j == 0 ? '*' : ' '), i, j); | |
4599 st(j)->print(); | |
4600 printf(" %s\n", tag_as_string(_tag_word.tag_at(i))); | |
4601 } | |
4602 } | |
4603 printf("\n"); | |
4604 // print control registers | |
4605 printf("ctrl = "); _control_word.print(); printf("\n"); | |
4606 printf("stat = "); _status_word .print(); printf("\n"); | |
4607 printf("tags = "); _tag_word .print(); printf("\n"); | |
4608 } | |
4609 | |
4610 }; | |
4611 | |
4612 class Flag_Register { | |
4613 public: | |
4614 int32_t _value; | |
4615 | |
4616 bool overflow() const { return ((_value >> 11) & 1) != 0; } | |
4617 bool direction() const { return ((_value >> 10) & 1) != 0; } | |
4618 bool sign() const { return ((_value >> 7) & 1) != 0; } | |
4619 bool zero() const { return ((_value >> 6) & 1) != 0; } | |
4620 bool auxiliary_carry() const { return ((_value >> 4) & 1) != 0; } | |
4621 bool parity() const { return ((_value >> 2) & 1) != 0; } | |
4622 bool carry() const { return ((_value >> 0) & 1) != 0; } | |
4623 | |
4624 void print() const { | |
4625 // flags | |
4626 char f[8]; | |
4627 f[0] = (overflow ()) ? 'O' : '-'; | |
4628 f[1] = (direction ()) ? 'D' : '-'; | |
4629 f[2] = (sign ()) ? 'S' : '-'; | |
4630 f[3] = (zero ()) ? 'Z' : '-'; | |
4631 f[4] = (auxiliary_carry()) ? 'A' : '-'; | |
4632 f[5] = (parity ()) ? 'P' : '-'; | |
4633 f[6] = (carry ()) ? 'C' : '-'; | |
4634 f[7] = '\x0'; | |
4635 // output | |
4636 printf("%08x flags = %s", _value, f); | |
4637 } | |
4638 | |
4639 }; | |
4640 | |
4641 class IU_Register { | |
4642 public: | |
4643 int32_t _value; | |
4644 | |
4645 void print() const { | |
4646 printf("%08x %11d", _value, _value); | |
4647 } | |
4648 | |
4649 }; | |
4650 | |
4651 class IU_State { | |
4652 public: | |
4653 Flag_Register _eflags; | |
4654 IU_Register _rdi; | |
4655 IU_Register _rsi; | |
4656 IU_Register _rbp; | |
4657 IU_Register _rsp; | |
4658 IU_Register _rbx; | |
4659 IU_Register _rdx; | |
4660 IU_Register _rcx; | |
4661 IU_Register _rax; | |
4662 | |
4663 void print() const { | |
4664 // computation registers | |
4665 printf("rax, = "); _rax.print(); printf("\n"); | |
4666 printf("rbx, = "); _rbx.print(); printf("\n"); | |
4667 printf("rcx = "); _rcx.print(); printf("\n"); | |
4668 printf("rdx = "); _rdx.print(); printf("\n"); | |
4669 printf("rdi = "); _rdi.print(); printf("\n"); | |
4670 printf("rsi = "); _rsi.print(); printf("\n"); | |
4671 printf("rbp, = "); _rbp.print(); printf("\n"); | |
4672 printf("rsp = "); _rsp.print(); printf("\n"); | |
4673 printf("\n"); | |
4674 // control registers | |
4675 printf("flgs = "); _eflags.print(); printf("\n"); | |
4676 } | |
4677 }; | |
4678 | |
4679 | |
4680 class CPU_State { | |
4681 public: | |
4682 FPU_State _fpu_state; | |
4683 IU_State _iu_state; | |
4684 | |
4685 void print() const { | |
4686 printf("--------------------------------------------------\n"); | |
4687 _iu_state .print(); | |
4688 printf("\n"); | |
4689 _fpu_state.print(); | |
4690 printf("--------------------------------------------------\n"); | |
4691 } | |
4692 | |
4693 }; | |
4694 | |
4695 | |
4696 static void _print_CPU_state(CPU_State* state) { | |
4697 state->print(); | |
4698 }; | |
4699 | |
4700 | |
4701 void MacroAssembler::print_CPU_state() { | |
4702 push_CPU_state(); | |
4703 push(rsp); // pass CPU state | |
4704 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _print_CPU_state))); | |
4705 addptr(rsp, wordSize); // discard argument | |
4706 pop_CPU_state(); | |
4707 } | |
4708 | |
4709 | |
4710 static bool _verify_FPU(int stack_depth, char* s, CPU_State* state) { | |
4711 static int counter = 0; | |
4712 FPU_State* fs = &state->_fpu_state; | |
4713 counter++; | |
4714 // For leaf calls, only verify that the top few elements remain empty. | |
4715 // We only need 1 empty at the top for C2 code. | |
4716 if( stack_depth < 0 ) { | |
4717 if( fs->tag_for_st(7) != 3 ) { | |
4718 printf("FPR7 not empty\n"); | |
4719 state->print(); | |
4720 assert(false, "error"); | |
4721 return false; | |
4722 } | |
4723 return true; // All other stack states do not matter | |
4724 } | |
4725 | |
4726 assert((fs->_control_word._value & 0xffff) == StubRoutines::_fpu_cntrl_wrd_std, | |
4727 "bad FPU control word"); | |
4728 | |
4729 // compute stack depth | |
4730 int i = 0; | |
4731 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) < 3) i++; | |
4732 int d = i; | |
4733 while (i < FPU_State::number_of_registers && fs->tag_for_st(i) == 3) i++; | |
4734 // verify findings | |
4735 if (i != FPU_State::number_of_registers) { | |
4736 // stack not contiguous | |
4737 printf("%s: stack not contiguous at ST%d\n", s, i); | |
4738 state->print(); | |
4739 assert(false, "error"); | |
4740 return false; | |
4741 } | |
4742 // check if computed stack depth corresponds to expected stack depth | |
4743 if (stack_depth < 0) { | |
4744 // expected stack depth is -stack_depth or less | |
4745 if (d > -stack_depth) { | |
4746 // too many elements on the stack | |
4747 printf("%s: <= %d stack elements expected but found %d\n", s, -stack_depth, d); | |
4748 state->print(); | |
4749 assert(false, "error"); | |
4750 return false; | |
4751 } | |
4752 } else { | |
4753 // expected stack depth is stack_depth | |
4754 if (d != stack_depth) { | |
4755 // wrong stack depth | |
4756 printf("%s: %d stack elements expected but found %d\n", s, stack_depth, d); | |
4757 state->print(); | |
4758 assert(false, "error"); | |
4759 return false; | |
4760 } | |
4761 } | |
4762 // everything is cool | |
4763 return true; | |
4764 } | |
4765 | |
4766 | |
4767 void MacroAssembler::verify_FPU(int stack_depth, const char* s) { | |
4768 if (!VerifyFPU) return; | |
4769 push_CPU_state(); | |
4770 push(rsp); // pass CPU state | |
4771 ExternalAddress msg((address) s); | |
4772 // pass message string s | |
4773 pushptr(msg.addr()); | |
4774 push(stack_depth); // pass stack depth | |
4775 call(RuntimeAddress(CAST_FROM_FN_PTR(address, _verify_FPU))); | |
4776 addptr(rsp, 3 * wordSize); // discard arguments | |
4777 // check for error | |
4778 { Label L; | |
4779 testl(rax, rax); | |
4780 jcc(Assembler::notZero, L); | |
4781 int3(); // break if error condition | |
4782 bind(L); | |
4783 } | |
4784 pop_CPU_state(); | |
4785 } | |
4786 | |
8873
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4787 void MacroAssembler::restore_cpu_control_state_after_jni() { |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4788 // Either restore the MXCSR register after returning from the JNI Call |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4789 // or verify that it wasn't changed (with -Xcheck:jni flag). |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4790 if (VM_Version::supports_sse()) { |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4791 if (RestoreMXCSROnJNICalls) { |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4792 ldmxcsr(ExternalAddress(StubRoutines::addr_mxcsr_std())); |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4793 } else if (CheckJNICalls) { |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4794 call(RuntimeAddress(StubRoutines::x86::verify_mxcsr_entry())); |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4795 } |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4796 } |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4797 if (VM_Version::supports_avx()) { |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4798 // Clear upper bits of YMM registers to avoid SSE <-> AVX transition penalty. |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4799 vzeroupper(); |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4800 } |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4801 |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4802 #ifndef _LP64 |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4803 // Either restore the x87 floating pointer control word after returning |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4804 // from the JNI call or verify that it wasn't changed. |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4805 if (CheckJNICalls) { |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4806 call(RuntimeAddress(StubRoutines::x86::verify_fpu_cntrl_wrd_entry())); |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4807 } |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4808 #endif // _LP64 |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4809 } |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4810 |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
4811 |
7199 | 4812 void MacroAssembler::load_klass(Register dst, Register src) { |
4813 #ifdef _LP64 | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
4814 if (UseCompressedClassPointers) { |
7199 | 4815 movl(dst, Address(src, oopDesc::klass_offset_in_bytes())); |
4816 decode_klass_not_null(dst); | |
4817 } else | |
4818 #endif | |
4819 movptr(dst, Address(src, oopDesc::klass_offset_in_bytes())); | |
4820 } | |
4821 | |
4822 void MacroAssembler::load_prototype_header(Register dst, Register src) { | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
4823 load_klass(dst, src); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
4824 movptr(dst, Address(dst, Klass::prototype_header_offset())); |
7199 | 4825 } |
4826 | |
4827 void MacroAssembler::store_klass(Register dst, Register src) { | |
4828 #ifdef _LP64 | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
4829 if (UseCompressedClassPointers) { |
7199 | 4830 encode_klass_not_null(src); |
4831 movl(Address(dst, oopDesc::klass_offset_in_bytes()), src); | |
4832 } else | |
4833 #endif | |
4834 movptr(Address(dst, oopDesc::klass_offset_in_bytes()), src); | |
4835 } | |
4836 | |
4837 void MacroAssembler::load_heap_oop(Register dst, Address src) { | |
4838 #ifdef _LP64 | |
4839 // FIXME: Must change all places where we try to load the klass. | |
4840 if (UseCompressedOops) { | |
4841 movl(dst, src); | |
4842 decode_heap_oop(dst); | |
4843 } else | |
4844 #endif | |
4845 movptr(dst, src); | |
4846 } | |
4847 | |
4848 // Doesn't do verfication, generates fixed size code | |
4849 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src) { | |
4850 #ifdef _LP64 | |
4851 if (UseCompressedOops) { | |
4852 movl(dst, src); | |
4853 decode_heap_oop_not_null(dst); | |
4854 } else | |
4855 #endif | |
4856 movptr(dst, src); | |
4857 } | |
4858 | |
4859 void MacroAssembler::store_heap_oop(Address dst, Register src) { | |
4860 #ifdef _LP64 | |
4861 if (UseCompressedOops) { | |
4862 assert(!dst.uses(src), "not enough registers"); | |
4863 encode_heap_oop(src); | |
4864 movl(dst, src); | |
4865 } else | |
4866 #endif | |
4867 movptr(dst, src); | |
4868 } | |
4869 | |
4870 void MacroAssembler::cmp_heap_oop(Register src1, Address src2, Register tmp) { | |
4871 assert_different_registers(src1, tmp); | |
4872 #ifdef _LP64 | |
4873 if (UseCompressedOops) { | |
4874 bool did_push = false; | |
4875 if (tmp == noreg) { | |
4876 tmp = rax; | |
4877 push(tmp); | |
4878 did_push = true; | |
4879 assert(!src2.uses(rsp), "can't push"); | |
4880 } | |
4881 load_heap_oop(tmp, src2); | |
4882 cmpptr(src1, tmp); | |
4883 if (did_push) pop(tmp); | |
4884 } else | |
4885 #endif | |
4886 cmpptr(src1, src2); | |
4887 } | |
4888 | |
4889 // Used for storing NULLs. | |
4890 void MacroAssembler::store_heap_oop_null(Address dst) { | |
4891 #ifdef _LP64 | |
4892 if (UseCompressedOops) { | |
4893 movl(dst, (int32_t)NULL_WORD); | |
4894 } else { | |
4895 movslq(dst, (int32_t)NULL_WORD); | |
4896 } | |
4897 #else | |
4898 movl(dst, (int32_t)NULL_WORD); | |
4899 #endif | |
4900 } | |
4901 | |
4902 #ifdef _LP64 | |
4903 void MacroAssembler::store_klass_gap(Register dst, Register src) { | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
4904 if (UseCompressedClassPointers) { |
7199 | 4905 // Store to klass gap in destination |
4906 movl(Address(dst, oopDesc::klass_gap_offset_in_bytes()), src); | |
4907 } | |
4908 } | |
4909 | |
4910 #ifdef ASSERT | |
4911 void MacroAssembler::verify_heapbase(const char* msg) { | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
4912 assert (UseCompressedOops, "should be compressed"); |
7199 | 4913 assert (Universe::heap() != NULL, "java heap should be initialized"); |
4914 if (CheckCompressedOops) { | |
4915 Label ok; | |
4916 push(rscratch1); // cmpptr trashes rscratch1 | |
4917 cmpptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); | |
4918 jcc(Assembler::equal, ok); | |
4919 STOP(msg); | |
4920 bind(ok); | |
4921 pop(rscratch1); | |
4922 } | |
4923 } | |
4924 #endif | |
4925 | |
4926 // Algorithm must match oop.inline.hpp encode_heap_oop. | |
4927 void MacroAssembler::encode_heap_oop(Register r) { | |
4928 #ifdef ASSERT | |
4929 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?"); | |
4930 #endif | |
4931 verify_oop(r, "broken oop in encode_heap_oop"); | |
4932 if (Universe::narrow_oop_base() == NULL) { | |
4933 if (Universe::narrow_oop_shift() != 0) { | |
4934 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); | |
4935 shrq(r, LogMinObjAlignmentInBytes); | |
4936 } | |
4937 return; | |
4938 } | |
4939 testq(r, r); | |
4940 cmovq(Assembler::equal, r, r12_heapbase); | |
4941 subq(r, r12_heapbase); | |
4942 shrq(r, LogMinObjAlignmentInBytes); | |
4943 } | |
4944 | |
4945 void MacroAssembler::encode_heap_oop_not_null(Register r) { | |
4946 #ifdef ASSERT | |
4947 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?"); | |
4948 if (CheckCompressedOops) { | |
4949 Label ok; | |
4950 testq(r, r); | |
4951 jcc(Assembler::notEqual, ok); | |
4952 STOP("null oop passed to encode_heap_oop_not_null"); | |
4953 bind(ok); | |
4954 } | |
4955 #endif | |
4956 verify_oop(r, "broken oop in encode_heap_oop_not_null"); | |
4957 if (Universe::narrow_oop_base() != NULL) { | |
4958 subq(r, r12_heapbase); | |
4959 } | |
4960 if (Universe::narrow_oop_shift() != 0) { | |
4961 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); | |
4962 shrq(r, LogMinObjAlignmentInBytes); | |
4963 } | |
4964 } | |
4965 | |
4966 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) { | |
4967 #ifdef ASSERT | |
4968 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?"); | |
4969 if (CheckCompressedOops) { | |
4970 Label ok; | |
4971 testq(src, src); | |
4972 jcc(Assembler::notEqual, ok); | |
4973 STOP("null oop passed to encode_heap_oop_not_null2"); | |
4974 bind(ok); | |
4975 } | |
4976 #endif | |
4977 verify_oop(src, "broken oop in encode_heap_oop_not_null2"); | |
4978 if (dst != src) { | |
4979 movq(dst, src); | |
4980 } | |
4981 if (Universe::narrow_oop_base() != NULL) { | |
4982 subq(dst, r12_heapbase); | |
4983 } | |
4984 if (Universe::narrow_oop_shift() != 0) { | |
4985 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); | |
4986 shrq(dst, LogMinObjAlignmentInBytes); | |
4987 } | |
4988 } | |
4989 | |
4990 void MacroAssembler::decode_heap_oop(Register r) { | |
4991 #ifdef ASSERT | |
4992 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?"); | |
4993 #endif | |
4994 if (Universe::narrow_oop_base() == NULL) { | |
4995 if (Universe::narrow_oop_shift() != 0) { | |
4996 assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); | |
4997 shlq(r, LogMinObjAlignmentInBytes); | |
4998 } | |
4999 } else { | |
5000 Label done; | |
5001 shlq(r, LogMinObjAlignmentInBytes); | |
5002 jccb(Assembler::equal, done); | |
5003 addq(r, r12_heapbase); | |
5004 bind(done); | |
5005 } | |
5006 verify_oop(r, "broken oop in decode_heap_oop"); | |
5007 } | |
5008 | |
5009 void MacroAssembler::decode_heap_oop_not_null(Register r) { | |
5010 // Note: it will change flags | |
5011 assert (UseCompressedOops, "should only be used for compressed headers"); | |
5012 assert (Universe::heap() != NULL, "java heap should be initialized"); | |
5013 // Cannot assert, unverified entry point counts instructions (see .ad file) | |
5014 // vtableStubs also counts instructions in pd_code_size_limit. | |
5015 // Also do not verify_oop as this is called by verify_oop. | |
5016 if (Universe::narrow_oop_shift() != 0) { | |
5017 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); | |
5018 shlq(r, LogMinObjAlignmentInBytes); | |
5019 if (Universe::narrow_oop_base() != NULL) { | |
5020 addq(r, r12_heapbase); | |
5021 } | |
5022 } else { | |
5023 assert (Universe::narrow_oop_base() == NULL, "sanity"); | |
5024 } | |
5025 } | |
5026 | |
5027 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) { | |
5028 // Note: it will change flags | |
5029 assert (UseCompressedOops, "should only be used for compressed headers"); | |
5030 assert (Universe::heap() != NULL, "java heap should be initialized"); | |
5031 // Cannot assert, unverified entry point counts instructions (see .ad file) | |
5032 // vtableStubs also counts instructions in pd_code_size_limit. | |
5033 // Also do not verify_oop as this is called by verify_oop. | |
5034 if (Universe::narrow_oop_shift() != 0) { | |
5035 assert(LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong"); | |
5036 if (LogMinObjAlignmentInBytes == Address::times_8) { | |
5037 leaq(dst, Address(r12_heapbase, src, Address::times_8, 0)); | |
5038 } else { | |
5039 if (dst != src) { | |
5040 movq(dst, src); | |
5041 } | |
5042 shlq(dst, LogMinObjAlignmentInBytes); | |
5043 if (Universe::narrow_oop_base() != NULL) { | |
5044 addq(dst, r12_heapbase); | |
5045 } | |
5046 } | |
5047 } else { | |
5048 assert (Universe::narrow_oop_base() == NULL, "sanity"); | |
5049 if (dst != src) { | |
5050 movq(dst, src); | |
5051 } | |
5052 } | |
5053 } | |
5054 | |
5055 void MacroAssembler::encode_klass_not_null(Register r) { | |
13000
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5056 if (Universe::narrow_klass_base() != NULL) { |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5057 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5058 assert(r != r12_heapbase, "Encoding a klass in r12"); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5059 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5060 subq(r, r12_heapbase); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5061 } |
7199 | 5062 if (Universe::narrow_klass_shift() != 0) { |
5063 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); | |
5064 shrq(r, LogKlassAlignmentInBytes); | |
5065 } | |
13000
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5066 if (Universe::narrow_klass_base() != NULL) { |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5067 reinit_heapbase(); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5068 } |
7199 | 5069 } |
5070 | |
5071 void MacroAssembler::encode_klass_not_null(Register dst, Register src) { | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5072 if (dst == src) { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5073 encode_klass_not_null(src); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5074 } else { |
13000
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5075 if (Universe::narrow_klass_base() != NULL) { |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5076 mov64(dst, (int64_t)Universe::narrow_klass_base()); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5077 negq(dst); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5078 addq(dst, src); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5079 } else { |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5080 movptr(dst, src); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5081 } |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5082 if (Universe::narrow_klass_shift() != 0) { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5083 assert (LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5084 shrq(dst, LogKlassAlignmentInBytes); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5085 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5086 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5087 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5088 |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5089 // Function instr_size_for_decode_klass_not_null() counts the instructions |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5090 // generated by decode_klass_not_null(register r) and reinit_heapbase(), |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5091 // when (Universe::heap() != NULL). Hence, if the instructions they |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5092 // generate change, then this method needs to be updated. |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5093 int MacroAssembler::instr_size_for_decode_klass_not_null() { |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
5094 assert (UseCompressedClassPointers, "only for compressed klass ptrs"); |
13000
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5095 if (Universe::narrow_klass_base() != NULL) { |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5096 // mov64 + addq + shlq? + mov64 (for reinit_heapbase()). |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5097 return (Universe::narrow_klass_shift() == 0 ? 20 : 24); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5098 } else { |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5099 // longest load decode klass function, mov64, leaq |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5100 return 16; |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5101 } |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5102 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5103 |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5104 // !!! If the instructions that get generated here change then function |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5105 // instr_size_for_decode_klass_not_null() needs to get updated. |
7199 | 5106 void MacroAssembler::decode_klass_not_null(Register r) { |
5107 // Note: it will change flags | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
5108 assert (UseCompressedClassPointers, "should only be used for compressed headers"); |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5109 assert(r != r12_heapbase, "Decoding a klass in r12"); |
7199 | 5110 // Cannot assert, unverified entry point counts instructions (see .ad file) |
5111 // vtableStubs also counts instructions in pd_code_size_limit. | |
5112 // Also do not verify_oop as this is called by verify_oop. | |
5113 if (Universe::narrow_klass_shift() != 0) { | |
5114 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); | |
5115 shlq(r, LogKlassAlignmentInBytes); | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5116 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5117 // Use r12 as a scratch register in which to temporarily load the narrow_klass_base. |
13000
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5118 if (Universe::narrow_klass_base() != NULL) { |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5119 mov64(r12_heapbase, (int64_t)Universe::narrow_klass_base()); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5120 addq(r, r12_heapbase); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5121 reinit_heapbase(); |
209aa13ab8c0
8024927: Nashorn performance regression with CompressedOops
coleenp
parents:
12835
diff
changeset
|
5122 } |
7199 | 5123 } |
5124 | |
5125 void MacroAssembler::decode_klass_not_null(Register dst, Register src) { | |
5126 // Note: it will change flags | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
5127 assert (UseCompressedClassPointers, "should only be used for compressed headers"); |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5128 if (dst == src) { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5129 decode_klass_not_null(dst); |
7199 | 5130 } else { |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5131 // Cannot assert, unverified entry point counts instructions (see .ad file) |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5132 // vtableStubs also counts instructions in pd_code_size_limit. |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5133 // Also do not verify_oop as this is called by verify_oop. |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5134 mov64(dst, (int64_t)Universe::narrow_klass_base()); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5135 if (Universe::narrow_klass_shift() != 0) { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5136 assert(LogKlassAlignmentInBytes == Universe::narrow_klass_shift(), "decode alg wrong"); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5137 assert(LogKlassAlignmentInBytes == Address::times_8, "klass not aligned on 64bits?"); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5138 leaq(dst, Address(dst, src, Address::times_8, 0)); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5139 } else { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5140 addq(dst, src); |
7199 | 5141 } |
5142 } | |
5143 } | |
5144 | |
5145 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) { | |
5146 assert (UseCompressedOops, "should only be used for compressed headers"); | |
5147 assert (Universe::heap() != NULL, "java heap should be initialized"); | |
5148 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); | |
5149 int oop_index = oop_recorder()->find_index(obj); | |
5150 RelocationHolder rspec = oop_Relocation::spec(oop_index); | |
5151 mov_narrow_oop(dst, oop_index, rspec); | |
5152 } | |
5153 | |
5154 void MacroAssembler::set_narrow_oop(Address dst, jobject obj) { | |
5155 assert (UseCompressedOops, "should only be used for compressed headers"); | |
5156 assert (Universe::heap() != NULL, "java heap should be initialized"); | |
5157 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); | |
5158 int oop_index = oop_recorder()->find_index(obj); | |
5159 RelocationHolder rspec = oop_Relocation::spec(oop_index); | |
5160 mov_narrow_oop(dst, oop_index, rspec); | |
5161 } | |
5162 | |
5163 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) { | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
5164 assert (UseCompressedClassPointers, "should only be used for compressed headers"); |
7199 | 5165 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); |
5166 int klass_index = oop_recorder()->find_index(k); | |
5167 RelocationHolder rspec = metadata_Relocation::spec(klass_index); | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5168 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); |
7199 | 5169 } |
5170 | |
5171 void MacroAssembler::set_narrow_klass(Address dst, Klass* k) { | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
5172 assert (UseCompressedClassPointers, "should only be used for compressed headers"); |
7199 | 5173 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); |
5174 int klass_index = oop_recorder()->find_index(k); | |
5175 RelocationHolder rspec = metadata_Relocation::spec(klass_index); | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5176 mov_narrow_oop(dst, Klass::encode_klass(k), rspec); |
7199 | 5177 } |
5178 | |
5179 void MacroAssembler::cmp_narrow_oop(Register dst, jobject obj) { | |
5180 assert (UseCompressedOops, "should only be used for compressed headers"); | |
5181 assert (Universe::heap() != NULL, "java heap should be initialized"); | |
5182 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); | |
5183 int oop_index = oop_recorder()->find_index(obj); | |
5184 RelocationHolder rspec = oop_Relocation::spec(oop_index); | |
5185 Assembler::cmp_narrow_oop(dst, oop_index, rspec); | |
5186 } | |
5187 | |
5188 void MacroAssembler::cmp_narrow_oop(Address dst, jobject obj) { | |
5189 assert (UseCompressedOops, "should only be used for compressed headers"); | |
5190 assert (Universe::heap() != NULL, "java heap should be initialized"); | |
5191 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); | |
5192 int oop_index = oop_recorder()->find_index(obj); | |
5193 RelocationHolder rspec = oop_Relocation::spec(oop_index); | |
5194 Assembler::cmp_narrow_oop(dst, oop_index, rspec); | |
5195 } | |
5196 | |
5197 void MacroAssembler::cmp_narrow_klass(Register dst, Klass* k) { | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
5198 assert (UseCompressedClassPointers, "should only be used for compressed headers"); |
7199 | 5199 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); |
5200 int klass_index = oop_recorder()->find_index(k); | |
5201 RelocationHolder rspec = metadata_Relocation::spec(klass_index); | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5202 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); |
7199 | 5203 } |
5204 | |
5205 void MacroAssembler::cmp_narrow_klass(Address dst, Klass* k) { | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
5206 assert (UseCompressedClassPointers, "should only be used for compressed headers"); |
7199 | 5207 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder"); |
5208 int klass_index = oop_recorder()->find_index(k); | |
5209 RelocationHolder rspec = metadata_Relocation::spec(klass_index); | |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5210 Assembler::cmp_narrow_oop(dst, Klass::encode_klass(k), rspec); |
7199 | 5211 } |
5212 | |
5213 void MacroAssembler::reinit_heapbase() { | |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
12056
diff
changeset
|
5214 if (UseCompressedOops || UseCompressedClassPointers) { |
12056
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5215 if (Universe::heap() != NULL) { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5216 if (Universe::narrow_oop_base() == NULL) { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5217 MacroAssembler::xorptr(r12_heapbase, r12_heapbase); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5218 } else { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5219 mov64(r12_heapbase, (int64_t)Universe::narrow_ptrs_base()); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5220 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5221 } else { |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5222 movptr(r12_heapbase, ExternalAddress((address)Universe::narrow_ptrs_base_addr())); |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5223 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5224 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5225 } |
740e263c80c6
8003424: Enable Class Data Sharing for CompressedOops
hseigel
parents:
11080
diff
changeset
|
5226 |
7199 | 5227 #endif // _LP64 |
5228 | |
5229 | |
5230 // C2 compiled method's prolog code. | |
5231 void MacroAssembler::verified_entry(int framesize, bool stack_bang, bool fp_mode_24b) { | |
5232 | |
5233 // WARNING: Initial instruction MUST be 5 bytes or longer so that | |
5234 // NativeJump::patch_verified_entry will be able to patch out the entry | |
5235 // code safely. The push to verify stack depth is ok at 5 bytes, | |
5236 // the frame allocation can be either 3 or 6 bytes. So if we don't do | |
5237 // stack bang then we must use the 6 byte frame allocation even if | |
5238 // we have no frame. :-( | |
5239 | |
5240 assert((framesize & (StackAlignmentInBytes-1)) == 0, "frame size not aligned"); | |
5241 // Remove word for return addr | |
5242 framesize -= wordSize; | |
5243 | |
5244 // Calls to C2R adapters often do not accept exceptional returns. | |
5245 // We require that their callers must bang for them. But be careful, because | |
5246 // some VM calls (such as call site linkage) can use several kilobytes of | |
5247 // stack. But the stack safety zone should account for that. | |
5248 // See bugs 4446381, 4468289, 4497237. | |
5249 if (stack_bang) { | |
5250 generate_stack_overflow_check(framesize); | |
5251 | |
5252 // We always push rbp, so that on return to interpreter rbp, will be | |
5253 // restored correctly and we can correct the stack. | |
5254 push(rbp); | |
5255 // Remove word for ebp | |
5256 framesize -= wordSize; | |
5257 | |
5258 // Create frame | |
5259 if (framesize) { | |
5260 subptr(rsp, framesize); | |
5261 } | |
5262 } else { | |
5263 // Create frame (force generation of a 4 byte immediate value) | |
5264 subptr_imm32(rsp, framesize); | |
5265 | |
5266 // Save RBP register now. | |
5267 framesize -= wordSize; | |
5268 movptr(Address(rsp, framesize), rbp); | |
5269 } | |
5270 | |
5271 if (VerifyStackAtCalls) { // Majik cookie to verify stack depth | |
5272 framesize -= wordSize; | |
5273 movptr(Address(rsp, framesize), (int32_t)0xbadb100d); | |
5274 } | |
5275 | |
5276 #ifndef _LP64 | |
5277 // If method sets FPU control word do it now | |
5278 if (fp_mode_24b) { | |
5279 fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_24())); | |
5280 } | |
5281 if (UseSSE >= 2 && VerifyFPU) { | |
5282 verify_FPU(0, "FPU stack must be clean on entry"); | |
5283 } | |
5284 #endif | |
5285 | |
5286 #ifdef ASSERT | |
5287 if (VerifyStackAtCalls) { | |
5288 Label L; | |
5289 push(rax); | |
5290 mov(rax, rsp); | |
5291 andptr(rax, StackAlignmentInBytes-1); | |
5292 cmpptr(rax, StackAlignmentInBytes-wordSize); | |
5293 pop(rax); | |
5294 jcc(Assembler::equal, L); | |
5295 STOP("Stack is not properly aligned!"); | |
5296 bind(L); | |
5297 } | |
5298 #endif | |
5299 | |
5300 } | |
5301 | |
7474
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5302 void MacroAssembler::clear_mem(Register base, Register cnt, Register tmp) { |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5303 // cnt - number of qwords (8-byte words). |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5304 // base - start address, qword aligned. |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5305 assert(base==rdi, "base register must be edi for rep stos"); |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5306 assert(tmp==rax, "tmp register must be eax for rep stos"); |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5307 assert(cnt==rcx, "cnt register must be ecx for rep stos"); |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5308 |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5309 xorptr(tmp, tmp); |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5310 if (UseFastStosb) { |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5311 shlptr(cnt,3); // convert to number of bytes |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5312 rep_stosb(); |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5313 } else { |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5314 NOT_LP64(shlptr(cnt,1);) // convert to number of dwords for 32-bit VM |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5315 rep_stos(); |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5316 } |
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
7430
diff
changeset
|
5317 } |
7199 | 5318 |
5319 // IndexOf for constant substrings with size >= 8 chars | |
5320 // which don't need to be loaded through stack. | |
5321 void MacroAssembler::string_indexofC8(Register str1, Register str2, | |
5322 Register cnt1, Register cnt2, | |
5323 int int_cnt2, Register result, | |
5324 XMMRegister vec, Register tmp) { | |
5325 ShortBranchVerifier sbv(this); | |
5326 assert(UseSSE42Intrinsics, "SSE4.2 is required"); | |
5327 | |
5328 // This method uses pcmpestri inxtruction with bound registers | |
5329 // inputs: | |
5330 // xmm - substring | |
5331 // rax - substring length (elements count) | |
5332 // mem - scanned string | |
5333 // rdx - string length (elements count) | |
5334 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) | |
5335 // outputs: | |
5336 // rcx - matched index in string | |
5337 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); | |
5338 | |
5339 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, | |
5340 RET_FOUND, RET_NOT_FOUND, EXIT, FOUND_SUBSTR, | |
5341 MATCH_SUBSTR_HEAD, RELOAD_STR, FOUND_CANDIDATE; | |
5342 | |
5343 // Note, inline_string_indexOf() generates checks: | |
5344 // if (substr.count > string.count) return -1; | |
5345 // if (substr.count == 0) return 0; | |
5346 assert(int_cnt2 >= 8, "this code isused only for cnt2 >= 8 chars"); | |
5347 | |
5348 // Load substring. | |
5349 movdqu(vec, Address(str2, 0)); | |
5350 movl(cnt2, int_cnt2); | |
5351 movptr(result, str1); // string addr | |
5352 | |
5353 if (int_cnt2 > 8) { | |
5354 jmpb(SCAN_TO_SUBSTR); | |
5355 | |
5356 // Reload substr for rescan, this code | |
5357 // is executed only for large substrings (> 8 chars) | |
5358 bind(RELOAD_SUBSTR); | |
5359 movdqu(vec, Address(str2, 0)); | |
5360 negptr(cnt2); // Jumped here with negative cnt2, convert to positive | |
5361 | |
5362 bind(RELOAD_STR); | |
5363 // We came here after the beginning of the substring was | |
5364 // matched but the rest of it was not so we need to search | |
5365 // again. Start from the next element after the previous match. | |
5366 | |
5367 // cnt2 is number of substring reminding elements and | |
5368 // cnt1 is number of string reminding elements when cmp failed. | |
5369 // Restored cnt1 = cnt1 - cnt2 + int_cnt2 | |
5370 subl(cnt1, cnt2); | |
5371 addl(cnt1, int_cnt2); | |
5372 movl(cnt2, int_cnt2); // Now restore cnt2 | |
5373 | |
5374 decrementl(cnt1); // Shift to next element | |
5375 cmpl(cnt1, cnt2); | |
5376 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring | |
5377 | |
5378 addptr(result, 2); | |
5379 | |
5380 } // (int_cnt2 > 8) | |
5381 | |
5382 // Scan string for start of substr in 16-byte vectors | |
5383 bind(SCAN_TO_SUBSTR); | |
5384 pcmpestri(vec, Address(result, 0), 0x0d); | |
5385 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 | |
5386 subl(cnt1, 8); | |
5387 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string | |
5388 cmpl(cnt1, cnt2); | |
5389 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring | |
5390 addptr(result, 16); | |
5391 jmpb(SCAN_TO_SUBSTR); | |
5392 | |
5393 // Found a potential substr | |
5394 bind(FOUND_CANDIDATE); | |
5395 // Matched whole vector if first element matched (tmp(rcx) == 0). | |
5396 if (int_cnt2 == 8) { | |
5397 jccb(Assembler::overflow, RET_FOUND); // OF == 1 | |
5398 } else { // int_cnt2 > 8 | |
5399 jccb(Assembler::overflow, FOUND_SUBSTR); | |
5400 } | |
5401 // After pcmpestri tmp(rcx) contains matched element index | |
5402 // Compute start addr of substr | |
5403 lea(result, Address(result, tmp, Address::times_2)); | |
5404 | |
5405 // Make sure string is still long enough | |
5406 subl(cnt1, tmp); | |
5407 cmpl(cnt1, cnt2); | |
5408 if (int_cnt2 == 8) { | |
5409 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); | |
5410 } else { // int_cnt2 > 8 | |
5411 jccb(Assembler::greaterEqual, MATCH_SUBSTR_HEAD); | |
5412 } | |
5413 // Left less then substring. | |
5414 | |
5415 bind(RET_NOT_FOUND); | |
5416 movl(result, -1); | |
5417 jmpb(EXIT); | |
5418 | |
5419 if (int_cnt2 > 8) { | |
5420 // This code is optimized for the case when whole substring | |
5421 // is matched if its head is matched. | |
5422 bind(MATCH_SUBSTR_HEAD); | |
5423 pcmpestri(vec, Address(result, 0), 0x0d); | |
5424 // Reload only string if does not match | |
5425 jccb(Assembler::noOverflow, RELOAD_STR); // OF == 0 | |
5426 | |
5427 Label CONT_SCAN_SUBSTR; | |
5428 // Compare the rest of substring (> 8 chars). | |
5429 bind(FOUND_SUBSTR); | |
5430 // First 8 chars are already matched. | |
5431 negptr(cnt2); | |
5432 addptr(cnt2, 8); | |
5433 | |
5434 bind(SCAN_SUBSTR); | |
5435 subl(cnt1, 8); | |
5436 cmpl(cnt2, -8); // Do not read beyond substring | |
5437 jccb(Assembler::lessEqual, CONT_SCAN_SUBSTR); | |
5438 // Back-up strings to avoid reading beyond substring: | |
5439 // cnt1 = cnt1 - cnt2 + 8 | |
5440 addl(cnt1, cnt2); // cnt2 is negative | |
5441 addl(cnt1, 8); | |
5442 movl(cnt2, 8); negptr(cnt2); | |
5443 bind(CONT_SCAN_SUBSTR); | |
5444 if (int_cnt2 < (int)G) { | |
5445 movdqu(vec, Address(str2, cnt2, Address::times_2, int_cnt2*2)); | |
5446 pcmpestri(vec, Address(result, cnt2, Address::times_2, int_cnt2*2), 0x0d); | |
5447 } else { | |
5448 // calculate index in register to avoid integer overflow (int_cnt2*2) | |
5449 movl(tmp, int_cnt2); | |
5450 addptr(tmp, cnt2); | |
5451 movdqu(vec, Address(str2, tmp, Address::times_2, 0)); | |
5452 pcmpestri(vec, Address(result, tmp, Address::times_2, 0), 0x0d); | |
5453 } | |
5454 // Need to reload strings pointers if not matched whole vector | |
5455 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 | |
5456 addptr(cnt2, 8); | |
5457 jcc(Assembler::negative, SCAN_SUBSTR); | |
5458 // Fall through if found full substring | |
5459 | |
5460 } // (int_cnt2 > 8) | |
5461 | |
5462 bind(RET_FOUND); | |
5463 // Found result if we matched full small substring. | |
5464 // Compute substr offset | |
5465 subptr(result, str1); | |
5466 shrl(result, 1); // index | |
5467 bind(EXIT); | |
5468 | |
5469 } // string_indexofC8 | |
5470 | |
5471 // Small strings are loaded through stack if they cross page boundary. | |
5472 void MacroAssembler::string_indexof(Register str1, Register str2, | |
5473 Register cnt1, Register cnt2, | |
5474 int int_cnt2, Register result, | |
5475 XMMRegister vec, Register tmp) { | |
5476 ShortBranchVerifier sbv(this); | |
5477 assert(UseSSE42Intrinsics, "SSE4.2 is required"); | |
5478 // | |
5479 // int_cnt2 is length of small (< 8 chars) constant substring | |
5480 // or (-1) for non constant substring in which case its length | |
5481 // is in cnt2 register. | |
5482 // | |
5483 // Note, inline_string_indexOf() generates checks: | |
5484 // if (substr.count > string.count) return -1; | |
5485 // if (substr.count == 0) return 0; | |
5486 // | |
5487 assert(int_cnt2 == -1 || (0 < int_cnt2 && int_cnt2 < 8), "should be != 0"); | |
5488 | |
5489 // This method uses pcmpestri inxtruction with bound registers | |
5490 // inputs: | |
5491 // xmm - substring | |
5492 // rax - substring length (elements count) | |
5493 // mem - scanned string | |
5494 // rdx - string length (elements count) | |
5495 // 0xd - mode: 1100 (substring search) + 01 (unsigned shorts) | |
5496 // outputs: | |
5497 // rcx - matched index in string | |
5498 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); | |
5499 | |
5500 Label RELOAD_SUBSTR, SCAN_TO_SUBSTR, SCAN_SUBSTR, ADJUST_STR, | |
5501 RET_FOUND, RET_NOT_FOUND, CLEANUP, FOUND_SUBSTR, | |
5502 FOUND_CANDIDATE; | |
5503 | |
5504 { //======================================================== | |
5505 // We don't know where these strings are located | |
5506 // and we can't read beyond them. Load them through stack. | |
5507 Label BIG_STRINGS, CHECK_STR, COPY_SUBSTR, COPY_STR; | |
5508 | |
5509 movptr(tmp, rsp); // save old SP | |
5510 | |
5511 if (int_cnt2 > 0) { // small (< 8 chars) constant substring | |
5512 if (int_cnt2 == 1) { // One char | |
5513 load_unsigned_short(result, Address(str2, 0)); | |
5514 movdl(vec, result); // move 32 bits | |
5515 } else if (int_cnt2 == 2) { // Two chars | |
5516 movdl(vec, Address(str2, 0)); // move 32 bits | |
5517 } else if (int_cnt2 == 4) { // Four chars | |
5518 movq(vec, Address(str2, 0)); // move 64 bits | |
5519 } else { // cnt2 = { 3, 5, 6, 7 } | |
5520 // Array header size is 12 bytes in 32-bit VM | |
5521 // + 6 bytes for 3 chars == 18 bytes, | |
5522 // enough space to load vec and shift. | |
5523 assert(HeapWordSize*TypeArrayKlass::header_size() >= 12,"sanity"); | |
5524 movdqu(vec, Address(str2, (int_cnt2*2)-16)); | |
5525 psrldq(vec, 16-(int_cnt2*2)); | |
5526 } | |
5527 } else { // not constant substring | |
5528 cmpl(cnt2, 8); | |
5529 jccb(Assembler::aboveEqual, BIG_STRINGS); // Both strings are big enough | |
5530 | |
5531 // We can read beyond string if srt+16 does not cross page boundary | |
5532 // since heaps are aligned and mapped by pages. | |
5533 assert(os::vm_page_size() < (int)G, "default page should be small"); | |
5534 movl(result, str2); // We need only low 32 bits | |
5535 andl(result, (os::vm_page_size()-1)); | |
5536 cmpl(result, (os::vm_page_size()-16)); | |
5537 jccb(Assembler::belowEqual, CHECK_STR); | |
5538 | |
5539 // Move small strings to stack to allow load 16 bytes into vec. | |
5540 subptr(rsp, 16); | |
5541 int stk_offset = wordSize-2; | |
5542 push(cnt2); | |
5543 | |
5544 bind(COPY_SUBSTR); | |
5545 load_unsigned_short(result, Address(str2, cnt2, Address::times_2, -2)); | |
5546 movw(Address(rsp, cnt2, Address::times_2, stk_offset), result); | |
5547 decrement(cnt2); | |
5548 jccb(Assembler::notZero, COPY_SUBSTR); | |
5549 | |
5550 pop(cnt2); | |
5551 movptr(str2, rsp); // New substring address | |
5552 } // non constant | |
5553 | |
5554 bind(CHECK_STR); | |
5555 cmpl(cnt1, 8); | |
5556 jccb(Assembler::aboveEqual, BIG_STRINGS); | |
5557 | |
5558 // Check cross page boundary. | |
5559 movl(result, str1); // We need only low 32 bits | |
5560 andl(result, (os::vm_page_size()-1)); | |
5561 cmpl(result, (os::vm_page_size()-16)); | |
5562 jccb(Assembler::belowEqual, BIG_STRINGS); | |
5563 | |
5564 subptr(rsp, 16); | |
5565 int stk_offset = -2; | |
5566 if (int_cnt2 < 0) { // not constant | |
5567 push(cnt2); | |
5568 stk_offset += wordSize; | |
5569 } | |
5570 movl(cnt2, cnt1); | |
5571 | |
5572 bind(COPY_STR); | |
5573 load_unsigned_short(result, Address(str1, cnt2, Address::times_2, -2)); | |
5574 movw(Address(rsp, cnt2, Address::times_2, stk_offset), result); | |
5575 decrement(cnt2); | |
5576 jccb(Assembler::notZero, COPY_STR); | |
5577 | |
5578 if (int_cnt2 < 0) { // not constant | |
5579 pop(cnt2); | |
5580 } | |
5581 movptr(str1, rsp); // New string address | |
5582 | |
5583 bind(BIG_STRINGS); | |
5584 // Load substring. | |
5585 if (int_cnt2 < 0) { // -1 | |
5586 movdqu(vec, Address(str2, 0)); | |
5587 push(cnt2); // substr count | |
5588 push(str2); // substr addr | |
5589 push(str1); // string addr | |
5590 } else { | |
5591 // Small (< 8 chars) constant substrings are loaded already. | |
5592 movl(cnt2, int_cnt2); | |
5593 } | |
5594 push(tmp); // original SP | |
5595 | |
5596 } // Finished loading | |
5597 | |
5598 //======================================================== | |
5599 // Start search | |
5600 // | |
5601 | |
5602 movptr(result, str1); // string addr | |
5603 | |
5604 if (int_cnt2 < 0) { // Only for non constant substring | |
5605 jmpb(SCAN_TO_SUBSTR); | |
5606 | |
5607 // SP saved at sp+0 | |
5608 // String saved at sp+1*wordSize | |
5609 // Substr saved at sp+2*wordSize | |
5610 // Substr count saved at sp+3*wordSize | |
5611 | |
5612 // Reload substr for rescan, this code | |
5613 // is executed only for large substrings (> 8 chars) | |
5614 bind(RELOAD_SUBSTR); | |
5615 movptr(str2, Address(rsp, 2*wordSize)); | |
5616 movl(cnt2, Address(rsp, 3*wordSize)); | |
5617 movdqu(vec, Address(str2, 0)); | |
5618 // We came here after the beginning of the substring was | |
5619 // matched but the rest of it was not so we need to search | |
5620 // again. Start from the next element after the previous match. | |
5621 subptr(str1, result); // Restore counter | |
5622 shrl(str1, 1); | |
5623 addl(cnt1, str1); | |
5624 decrementl(cnt1); // Shift to next element | |
5625 cmpl(cnt1, cnt2); | |
5626 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring | |
5627 | |
5628 addptr(result, 2); | |
5629 } // non constant | |
5630 | |
5631 // Scan string for start of substr in 16-byte vectors | |
5632 bind(SCAN_TO_SUBSTR); | |
5633 assert(cnt1 == rdx && cnt2 == rax && tmp == rcx, "pcmpestri"); | |
5634 pcmpestri(vec, Address(result, 0), 0x0d); | |
5635 jccb(Assembler::below, FOUND_CANDIDATE); // CF == 1 | |
5636 subl(cnt1, 8); | |
5637 jccb(Assembler::lessEqual, RET_NOT_FOUND); // Scanned full string | |
5638 cmpl(cnt1, cnt2); | |
5639 jccb(Assembler::negative, RET_NOT_FOUND); // Left less then substring | |
5640 addptr(result, 16); | |
5641 | |
5642 bind(ADJUST_STR); | |
5643 cmpl(cnt1, 8); // Do not read beyond string | |
5644 jccb(Assembler::greaterEqual, SCAN_TO_SUBSTR); | |
5645 // Back-up string to avoid reading beyond string. | |
5646 lea(result, Address(result, cnt1, Address::times_2, -16)); | |
5647 movl(cnt1, 8); | |
5648 jmpb(SCAN_TO_SUBSTR); | |
5649 | |
5650 // Found a potential substr | |
5651 bind(FOUND_CANDIDATE); | |
5652 // After pcmpestri tmp(rcx) contains matched element index | |
5653 | |
5654 // Make sure string is still long enough | |
5655 subl(cnt1, tmp); | |
5656 cmpl(cnt1, cnt2); | |
5657 jccb(Assembler::greaterEqual, FOUND_SUBSTR); | |
5658 // Left less then substring. | |
5659 | |
5660 bind(RET_NOT_FOUND); | |
5661 movl(result, -1); | |
5662 jmpb(CLEANUP); | |
5663 | |
5664 bind(FOUND_SUBSTR); | |
5665 // Compute start addr of substr | |
5666 lea(result, Address(result, tmp, Address::times_2)); | |
5667 | |
5668 if (int_cnt2 > 0) { // Constant substring | |
5669 // Repeat search for small substring (< 8 chars) | |
5670 // from new point without reloading substring. | |
5671 // Have to check that we don't read beyond string. | |
5672 cmpl(tmp, 8-int_cnt2); | |
5673 jccb(Assembler::greater, ADJUST_STR); | |
5674 // Fall through if matched whole substring. | |
5675 } else { // non constant | |
5676 assert(int_cnt2 == -1, "should be != 0"); | |
5677 | |
5678 addl(tmp, cnt2); | |
5679 // Found result if we matched whole substring. | |
5680 cmpl(tmp, 8); | |
5681 jccb(Assembler::lessEqual, RET_FOUND); | |
5682 | |
5683 // Repeat search for small substring (<= 8 chars) | |
5684 // from new point 'str1' without reloading substring. | |
5685 cmpl(cnt2, 8); | |
5686 // Have to check that we don't read beyond string. | |
5687 jccb(Assembler::lessEqual, ADJUST_STR); | |
5688 | |
5689 Label CHECK_NEXT, CONT_SCAN_SUBSTR, RET_FOUND_LONG; | |
5690 // Compare the rest of substring (> 8 chars). | |
5691 movptr(str1, result); | |
5692 | |
5693 cmpl(tmp, cnt2); | |
5694 // First 8 chars are already matched. | |
5695 jccb(Assembler::equal, CHECK_NEXT); | |
5696 | |
5697 bind(SCAN_SUBSTR); | |
5698 pcmpestri(vec, Address(str1, 0), 0x0d); | |
5699 // Need to reload strings pointers if not matched whole vector | |
5700 jcc(Assembler::noOverflow, RELOAD_SUBSTR); // OF == 0 | |
5701 | |
5702 bind(CHECK_NEXT); | |
5703 subl(cnt2, 8); | |
5704 jccb(Assembler::lessEqual, RET_FOUND_LONG); // Found full substring | |
5705 addptr(str1, 16); | |
5706 addptr(str2, 16); | |
5707 subl(cnt1, 8); | |
5708 cmpl(cnt2, 8); // Do not read beyond substring | |
5709 jccb(Assembler::greaterEqual, CONT_SCAN_SUBSTR); | |
5710 // Back-up strings to avoid reading beyond substring. | |
5711 lea(str2, Address(str2, cnt2, Address::times_2, -16)); | |
5712 lea(str1, Address(str1, cnt2, Address::times_2, -16)); | |
5713 subl(cnt1, cnt2); | |
5714 movl(cnt2, 8); | |
5715 addl(cnt1, 8); | |
5716 bind(CONT_SCAN_SUBSTR); | |
5717 movdqu(vec, Address(str2, 0)); | |
5718 jmpb(SCAN_SUBSTR); | |
5719 | |
5720 bind(RET_FOUND_LONG); | |
5721 movptr(str1, Address(rsp, wordSize)); | |
5722 } // non constant | |
5723 | |
5724 bind(RET_FOUND); | |
5725 // Compute substr offset | |
5726 subptr(result, str1); | |
5727 shrl(result, 1); // index | |
5728 | |
5729 bind(CLEANUP); | |
5730 pop(rsp); // restore SP | |
5731 | |
5732 } // string_indexof | |
5733 | |
5734 // Compare strings. | |
5735 void MacroAssembler::string_compare(Register str1, Register str2, | |
5736 Register cnt1, Register cnt2, Register result, | |
5737 XMMRegister vec1) { | |
5738 ShortBranchVerifier sbv(this); | |
5739 Label LENGTH_DIFF_LABEL, POP_LABEL, DONE_LABEL, WHILE_HEAD_LABEL; | |
5740 | |
5741 // Compute the minimum of the string lengths and the | |
5742 // difference of the string lengths (stack). | |
5743 // Do the conditional move stuff | |
5744 movl(result, cnt1); | |
5745 subl(cnt1, cnt2); | |
5746 push(cnt1); | |
5747 cmov32(Assembler::lessEqual, cnt2, result); | |
5748 | |
5749 // Is the minimum length zero? | |
5750 testl(cnt2, cnt2); | |
5751 jcc(Assembler::zero, LENGTH_DIFF_LABEL); | |
5752 | |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5753 // Compare first characters |
7199 | 5754 load_unsigned_short(result, Address(str1, 0)); |
5755 load_unsigned_short(cnt1, Address(str2, 0)); | |
5756 subl(result, cnt1); | |
5757 jcc(Assembler::notZero, POP_LABEL); | |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5758 cmpl(cnt2, 1); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5759 jcc(Assembler::equal, LENGTH_DIFF_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5760 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5761 // Check if the strings start at the same location. |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5762 cmpptr(str1, str2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5763 jcc(Assembler::equal, LENGTH_DIFF_LABEL); |
7199 | 5764 |
5765 Address::ScaleFactor scale = Address::times_2; | |
5766 int stride = 8; | |
5767 | |
8042
91a23b11d8dc
8007708: compiler/6855215 assert(VM_Version::supports_sse4_2())
kvn
parents:
8002
diff
changeset
|
5768 if (UseAVX >= 2 && UseSSE42Intrinsics) { |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5769 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_WIDE_TAIL, COMPARE_SMALL_STR; |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5770 Label COMPARE_WIDE_VECTORS_LOOP, COMPARE_16_CHARS, COMPARE_INDEX_CHAR; |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5771 Label COMPARE_TAIL_LONG; |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5772 int pcmpmask = 0x19; |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5773 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5774 // Setup to compare 16-chars (32-bytes) vectors, |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5775 // start from first character again because it has aligned address. |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5776 int stride2 = 16; |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5777 int adr_stride = stride << scale; |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5778 int adr_stride2 = stride2 << scale; |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5779 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5780 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5781 // rax and rdx are used by pcmpestri as elements counters |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5782 movl(result, cnt2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5783 andl(cnt2, ~(stride2-1)); // cnt2 holds the vector count |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5784 jcc(Assembler::zero, COMPARE_TAIL_LONG); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5785 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5786 // fast path : compare first 2 8-char vectors. |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5787 bind(COMPARE_16_CHARS); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5788 movdqu(vec1, Address(str1, 0)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5789 pcmpestri(vec1, Address(str2, 0), pcmpmask); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5790 jccb(Assembler::below, COMPARE_INDEX_CHAR); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5791 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5792 movdqu(vec1, Address(str1, adr_stride)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5793 pcmpestri(vec1, Address(str2, adr_stride), pcmpmask); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5794 jccb(Assembler::aboveEqual, COMPARE_WIDE_VECTORS); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5795 addl(cnt1, stride); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5796 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5797 // Compare the characters at index in cnt1 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5798 bind(COMPARE_INDEX_CHAR); //cnt1 has the offset of the mismatching character |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5799 load_unsigned_short(result, Address(str1, cnt1, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5800 load_unsigned_short(cnt2, Address(str2, cnt1, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5801 subl(result, cnt2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5802 jmp(POP_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5803 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5804 // Setup the registers to start vector comparison loop |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5805 bind(COMPARE_WIDE_VECTORS); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5806 lea(str1, Address(str1, result, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5807 lea(str2, Address(str2, result, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5808 subl(result, stride2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5809 subl(cnt2, stride2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5810 jccb(Assembler::zero, COMPARE_WIDE_TAIL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5811 negptr(result); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5812 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5813 // In a loop, compare 16-chars (32-bytes) at once using (vpxor+vptest) |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5814 bind(COMPARE_WIDE_VECTORS_LOOP); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5815 vmovdqu(vec1, Address(str1, result, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5816 vpxor(vec1, Address(str2, result, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5817 vptest(vec1, vec1); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5818 jccb(Assembler::notZero, VECTOR_NOT_EQUAL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5819 addptr(result, stride2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5820 subl(cnt2, stride2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5821 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS_LOOP); |
8873
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
5822 // clean upper bits of YMM registers |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
5823 vzeroupper(); |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5824 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5825 // compare wide vectors tail |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5826 bind(COMPARE_WIDE_TAIL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5827 testptr(result, result); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5828 jccb(Assembler::zero, LENGTH_DIFF_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5829 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5830 movl(result, stride2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5831 movl(cnt2, result); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5832 negptr(result); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5833 jmpb(COMPARE_WIDE_VECTORS_LOOP); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5834 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5835 // Identifies the mismatching (higher or lower)16-bytes in the 32-byte vectors. |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5836 bind(VECTOR_NOT_EQUAL); |
8873
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
5837 // clean upper bits of YMM registers |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
5838 vzeroupper(); |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5839 lea(str1, Address(str1, result, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5840 lea(str2, Address(str2, result, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5841 jmp(COMPARE_16_CHARS); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5842 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5843 // Compare tail chars, length between 1 to 15 chars |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5844 bind(COMPARE_TAIL_LONG); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5845 movl(cnt2, result); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5846 cmpl(cnt2, stride); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5847 jccb(Assembler::less, COMPARE_SMALL_STR); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5848 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5849 movdqu(vec1, Address(str1, 0)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5850 pcmpestri(vec1, Address(str2, 0), pcmpmask); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5851 jcc(Assembler::below, COMPARE_INDEX_CHAR); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5852 subptr(cnt2, stride); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5853 jccb(Assembler::zero, LENGTH_DIFF_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5854 lea(str1, Address(str1, result, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5855 lea(str2, Address(str2, result, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5856 negptr(cnt2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5857 jmpb(WHILE_HEAD_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5858 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5859 bind(COMPARE_SMALL_STR); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5860 } else if (UseSSE42Intrinsics) { |
7199 | 5861 Label COMPARE_WIDE_VECTORS, VECTOR_NOT_EQUAL, COMPARE_TAIL; |
5862 int pcmpmask = 0x19; | |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5863 // Setup to compare 8-char (16-byte) vectors, |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5864 // start from first character again because it has aligned address. |
7199 | 5865 movl(result, cnt2); |
5866 andl(cnt2, ~(stride - 1)); // cnt2 holds the vector count | |
5867 jccb(Assembler::zero, COMPARE_TAIL); | |
5868 | |
5869 lea(str1, Address(str1, result, scale)); | |
5870 lea(str2, Address(str2, result, scale)); | |
5871 negptr(result); | |
5872 | |
5873 // pcmpestri | |
5874 // inputs: | |
5875 // vec1- substring | |
5876 // rax - negative string length (elements count) | |
5877 // mem - scaned string | |
5878 // rdx - string length (elements count) | |
5879 // pcmpmask - cmp mode: 11000 (string compare with negated result) | |
5880 // + 00 (unsigned bytes) or + 01 (unsigned shorts) | |
5881 // outputs: | |
5882 // rcx - first mismatched element index | |
5883 assert(result == rax && cnt2 == rdx && cnt1 == rcx, "pcmpestri"); | |
5884 | |
5885 bind(COMPARE_WIDE_VECTORS); | |
5886 movdqu(vec1, Address(str1, result, scale)); | |
5887 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); | |
5888 // After pcmpestri cnt1(rcx) contains mismatched element index | |
5889 | |
5890 jccb(Assembler::below, VECTOR_NOT_EQUAL); // CF==1 | |
5891 addptr(result, stride); | |
5892 subptr(cnt2, stride); | |
5893 jccb(Assembler::notZero, COMPARE_WIDE_VECTORS); | |
5894 | |
5895 // compare wide vectors tail | |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5896 testptr(result, result); |
7199 | 5897 jccb(Assembler::zero, LENGTH_DIFF_LABEL); |
5898 | |
5899 movl(cnt2, stride); | |
5900 movl(result, stride); | |
5901 negptr(result); | |
5902 movdqu(vec1, Address(str1, result, scale)); | |
5903 pcmpestri(vec1, Address(str2, result, scale), pcmpmask); | |
5904 jccb(Assembler::aboveEqual, LENGTH_DIFF_LABEL); | |
5905 | |
5906 // Mismatched characters in the vectors | |
5907 bind(VECTOR_NOT_EQUAL); | |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5908 addptr(cnt1, result); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5909 load_unsigned_short(result, Address(str1, cnt1, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5910 load_unsigned_short(cnt2, Address(str2, cnt1, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5911 subl(result, cnt2); |
7199 | 5912 jmpb(POP_LABEL); |
5913 | |
5914 bind(COMPARE_TAIL); // limit is zero | |
5915 movl(cnt2, result); | |
5916 // Fallthru to tail compare | |
5917 } | |
5918 // Shift str2 and str1 to the end of the arrays, negate min | |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5919 lea(str1, Address(str1, cnt2, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5920 lea(str2, Address(str2, cnt2, scale)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5921 decrementl(cnt2); // first character was compared already |
7199 | 5922 negptr(cnt2); |
5923 | |
5924 // Compare the rest of the elements | |
5925 bind(WHILE_HEAD_LABEL); | |
5926 load_unsigned_short(result, Address(str1, cnt2, scale, 0)); | |
5927 load_unsigned_short(cnt1, Address(str2, cnt2, scale, 0)); | |
5928 subl(result, cnt1); | |
5929 jccb(Assembler::notZero, POP_LABEL); | |
5930 increment(cnt2); | |
5931 jccb(Assembler::notZero, WHILE_HEAD_LABEL); | |
5932 | |
5933 // Strings are equal up to min length. Return the length difference. | |
5934 bind(LENGTH_DIFF_LABEL); | |
5935 pop(result); | |
5936 jmpb(DONE_LABEL); | |
5937 | |
5938 // Discard the stored length difference | |
5939 bind(POP_LABEL); | |
5940 pop(cnt1); | |
5941 | |
5942 // That's it | |
5943 bind(DONE_LABEL); | |
5944 } | |
5945 | |
5946 // Compare char[] arrays aligned to 4 bytes or substrings. | |
5947 void MacroAssembler::char_arrays_equals(bool is_array_equ, Register ary1, Register ary2, | |
5948 Register limit, Register result, Register chr, | |
5949 XMMRegister vec1, XMMRegister vec2) { | |
5950 ShortBranchVerifier sbv(this); | |
5951 Label TRUE_LABEL, FALSE_LABEL, DONE, COMPARE_VECTORS, COMPARE_CHAR; | |
5952 | |
5953 int length_offset = arrayOopDesc::length_offset_in_bytes(); | |
5954 int base_offset = arrayOopDesc::base_offset_in_bytes(T_CHAR); | |
5955 | |
5956 // Check the input args | |
5957 cmpptr(ary1, ary2); | |
5958 jcc(Assembler::equal, TRUE_LABEL); | |
5959 | |
5960 if (is_array_equ) { | |
5961 // Need additional checks for arrays_equals. | |
5962 testptr(ary1, ary1); | |
5963 jcc(Assembler::zero, FALSE_LABEL); | |
5964 testptr(ary2, ary2); | |
5965 jcc(Assembler::zero, FALSE_LABEL); | |
5966 | |
5967 // Check the lengths | |
5968 movl(limit, Address(ary1, length_offset)); | |
5969 cmpl(limit, Address(ary2, length_offset)); | |
5970 jcc(Assembler::notEqual, FALSE_LABEL); | |
5971 } | |
5972 | |
5973 // count == 0 | |
5974 testl(limit, limit); | |
5975 jcc(Assembler::zero, TRUE_LABEL); | |
5976 | |
5977 if (is_array_equ) { | |
5978 // Load array address | |
5979 lea(ary1, Address(ary1, base_offset)); | |
5980 lea(ary2, Address(ary2, base_offset)); | |
5981 } | |
5982 | |
5983 shll(limit, 1); // byte count != 0 | |
5984 movl(result, limit); // copy | |
5985 | |
7477
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5986 if (UseAVX >= 2) { |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5987 // With AVX2, use 32-byte vector compare |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5988 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5989 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5990 // Compare 32-byte vectors |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5991 andl(result, 0x0000001e); // tail count (in bytes) |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5992 andl(limit, 0xffffffe0); // vector count (in bytes) |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5993 jccb(Assembler::zero, COMPARE_TAIL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5994 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5995 lea(ary1, Address(ary1, limit, Address::times_1)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5996 lea(ary2, Address(ary2, limit, Address::times_1)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5997 negptr(limit); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5998 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
5999 bind(COMPARE_WIDE_VECTORS); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6000 vmovdqu(vec1, Address(ary1, limit, Address::times_1)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6001 vmovdqu(vec2, Address(ary2, limit, Address::times_1)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6002 vpxor(vec1, vec2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6003 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6004 vptest(vec1, vec1); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6005 jccb(Assembler::notZero, FALSE_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6006 addptr(limit, 32); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6007 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6008 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6009 testl(result, result); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6010 jccb(Assembler::zero, TRUE_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6011 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6012 vmovdqu(vec1, Address(ary1, result, Address::times_1, -32)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6013 vmovdqu(vec2, Address(ary2, result, Address::times_1, -32)); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6014 vpxor(vec1, vec2); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6015 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6016 vptest(vec1, vec1); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6017 jccb(Assembler::notZero, FALSE_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6018 jmpb(TRUE_LABEL); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6019 |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6020 bind(COMPARE_TAIL); // limit is zero |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6021 movl(limit, result); |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6022 // Fallthru to tail compare |
038dd2875b94
8005419: Improve intrinsics code performance on x86 by using AVX2
kvn
parents:
7476
diff
changeset
|
6023 } else if (UseSSE42Intrinsics) { |
7199 | 6024 // With SSE4.2, use double quad vector compare |
6025 Label COMPARE_WIDE_VECTORS, COMPARE_TAIL; | |
6026 | |
6027 // Compare 16-byte vectors | |
6028 andl(result, 0x0000000e); // tail count (in bytes) | |
6029 andl(limit, 0xfffffff0); // vector count (in bytes) | |
6030 jccb(Assembler::zero, COMPARE_TAIL); | |
6031 | |
6032 lea(ary1, Address(ary1, limit, Address::times_1)); | |
6033 lea(ary2, Address(ary2, limit, Address::times_1)); | |
6034 negptr(limit); | |
6035 | |
6036 bind(COMPARE_WIDE_VECTORS); | |
6037 movdqu(vec1, Address(ary1, limit, Address::times_1)); | |
6038 movdqu(vec2, Address(ary2, limit, Address::times_1)); | |
6039 pxor(vec1, vec2); | |
6040 | |
6041 ptest(vec1, vec1); | |
6042 jccb(Assembler::notZero, FALSE_LABEL); | |
6043 addptr(limit, 16); | |
6044 jcc(Assembler::notZero, COMPARE_WIDE_VECTORS); | |
6045 | |
6046 testl(result, result); | |
6047 jccb(Assembler::zero, TRUE_LABEL); | |
6048 | |
6049 movdqu(vec1, Address(ary1, result, Address::times_1, -16)); | |
6050 movdqu(vec2, Address(ary2, result, Address::times_1, -16)); | |
6051 pxor(vec1, vec2); | |
6052 | |
6053 ptest(vec1, vec1); | |
6054 jccb(Assembler::notZero, FALSE_LABEL); | |
6055 jmpb(TRUE_LABEL); | |
6056 | |
6057 bind(COMPARE_TAIL); // limit is zero | |
6058 movl(limit, result); | |
6059 // Fallthru to tail compare | |
6060 } | |
6061 | |
6062 // Compare 4-byte vectors | |
6063 andl(limit, 0xfffffffc); // vector count (in bytes) | |
6064 jccb(Assembler::zero, COMPARE_CHAR); | |
6065 | |
6066 lea(ary1, Address(ary1, limit, Address::times_1)); | |
6067 lea(ary2, Address(ary2, limit, Address::times_1)); | |
6068 negptr(limit); | |
6069 | |
6070 bind(COMPARE_VECTORS); | |
6071 movl(chr, Address(ary1, limit, Address::times_1)); | |
6072 cmpl(chr, Address(ary2, limit, Address::times_1)); | |
6073 jccb(Assembler::notEqual, FALSE_LABEL); | |
6074 addptr(limit, 4); | |
6075 jcc(Assembler::notZero, COMPARE_VECTORS); | |
6076 | |
6077 // Compare trailing char (final 2 bytes), if any | |
6078 bind(COMPARE_CHAR); | |
6079 testl(result, 0x2); // tail char | |
6080 jccb(Assembler::zero, TRUE_LABEL); | |
6081 load_unsigned_short(chr, Address(ary1, 0)); | |
6082 load_unsigned_short(limit, Address(ary2, 0)); | |
6083 cmpl(chr, limit); | |
6084 jccb(Assembler::notEqual, FALSE_LABEL); | |
6085 | |
6086 bind(TRUE_LABEL); | |
6087 movl(result, 1); // return true | |
6088 jmpb(DONE); | |
6089 | |
6090 bind(FALSE_LABEL); | |
6091 xorl(result, result); // return false | |
6092 | |
6093 // That's it | |
6094 bind(DONE); | |
8873
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6095 if (UseAVX >= 2) { |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6096 // clean upper bits of YMM registers |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6097 vzeroupper(); |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6098 } |
7199 | 6099 } |
6100 | |
6101 void MacroAssembler::generate_fill(BasicType t, bool aligned, | |
6102 Register to, Register value, Register count, | |
6103 Register rtmp, XMMRegister xtmp) { | |
6104 ShortBranchVerifier sbv(this); | |
6105 assert_different_registers(to, value, count, rtmp); | |
6106 Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte; | |
6107 Label L_fill_2_bytes, L_fill_4_bytes; | |
6108 | |
6109 int shift = -1; | |
6110 switch (t) { | |
6111 case T_BYTE: | |
6112 shift = 2; | |
6113 break; | |
6114 case T_SHORT: | |
6115 shift = 1; | |
6116 break; | |
6117 case T_INT: | |
6118 shift = 0; | |
6119 break; | |
6120 default: ShouldNotReachHere(); | |
6121 } | |
6122 | |
6123 if (t == T_BYTE) { | |
6124 andl(value, 0xff); | |
6125 movl(rtmp, value); | |
6126 shll(rtmp, 8); | |
6127 orl(value, rtmp); | |
6128 } | |
6129 if (t == T_SHORT) { | |
6130 andl(value, 0xffff); | |
6131 } | |
6132 if (t == T_BYTE || t == T_SHORT) { | |
6133 movl(rtmp, value); | |
6134 shll(rtmp, 16); | |
6135 orl(value, rtmp); | |
6136 } | |
6137 | |
6138 cmpl(count, 2<<shift); // Short arrays (< 8 bytes) fill by element | |
6139 jcc(Assembler::below, L_fill_4_bytes); // use unsigned cmp | |
6140 if (!UseUnalignedLoadStores && !aligned && (t == T_BYTE || t == T_SHORT)) { | |
6141 // align source address at 4 bytes address boundary | |
6142 if (t == T_BYTE) { | |
6143 // One byte misalignment happens only for byte arrays | |
6144 testptr(to, 1); | |
6145 jccb(Assembler::zero, L_skip_align1); | |
6146 movb(Address(to, 0), value); | |
6147 increment(to); | |
6148 decrement(count); | |
6149 BIND(L_skip_align1); | |
6150 } | |
6151 // Two bytes misalignment happens only for byte and short (char) arrays | |
6152 testptr(to, 2); | |
6153 jccb(Assembler::zero, L_skip_align2); | |
6154 movw(Address(to, 0), value); | |
6155 addptr(to, 2); | |
6156 subl(count, 1<<(shift-1)); | |
6157 BIND(L_skip_align2); | |
6158 } | |
6159 if (UseSSE < 2) { | |
6160 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; | |
6161 // Fill 32-byte chunks | |
6162 subl(count, 8 << shift); | |
6163 jcc(Assembler::less, L_check_fill_8_bytes); | |
6164 align(16); | |
6165 | |
6166 BIND(L_fill_32_bytes_loop); | |
6167 | |
6168 for (int i = 0; i < 32; i += 4) { | |
6169 movl(Address(to, i), value); | |
6170 } | |
6171 | |
6172 addptr(to, 32); | |
6173 subl(count, 8 << shift); | |
6174 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); | |
6175 BIND(L_check_fill_8_bytes); | |
6176 addl(count, 8 << shift); | |
6177 jccb(Assembler::zero, L_exit); | |
6178 jmpb(L_fill_8_bytes); | |
6179 | |
6180 // | |
6181 // length is too short, just fill qwords | |
6182 // | |
6183 BIND(L_fill_8_bytes_loop); | |
6184 movl(Address(to, 0), value); | |
6185 movl(Address(to, 4), value); | |
6186 addptr(to, 8); | |
6187 BIND(L_fill_8_bytes); | |
6188 subl(count, 1 << (shift + 1)); | |
6189 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); | |
6190 // fall through to fill 4 bytes | |
6191 } else { | |
6192 Label L_fill_32_bytes; | |
6193 if (!UseUnalignedLoadStores) { | |
6194 // align to 8 bytes, we know we are 4 byte aligned to start | |
6195 testptr(to, 4); | |
6196 jccb(Assembler::zero, L_fill_32_bytes); | |
6197 movl(Address(to, 0), value); | |
6198 addptr(to, 4); | |
6199 subl(count, 1<<shift); | |
6200 } | |
6201 BIND(L_fill_32_bytes); | |
6202 { | |
6203 assert( UseSSE >= 2, "supported cpu only" ); | |
6204 Label L_fill_32_bytes_loop, L_check_fill_8_bytes, L_fill_8_bytes_loop, L_fill_8_bytes; | |
6205 movdl(xtmp, value); | |
7475
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6206 if (UseAVX >= 2 && UseUnalignedLoadStores) { |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6207 // Fill 64-byte chunks |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6208 Label L_fill_64_bytes_loop, L_check_fill_32_bytes; |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6209 vpbroadcastd(xtmp, xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6210 |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6211 subl(count, 16 << shift); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6212 jcc(Assembler::less, L_check_fill_32_bytes); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6213 align(16); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6214 |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6215 BIND(L_fill_64_bytes_loop); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6216 vmovdqu(Address(to, 0), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6217 vmovdqu(Address(to, 32), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6218 addptr(to, 64); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6219 subl(count, 16 << shift); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6220 jcc(Assembler::greaterEqual, L_fill_64_bytes_loop); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6221 |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6222 BIND(L_check_fill_32_bytes); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6223 addl(count, 8 << shift); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6224 jccb(Assembler::less, L_check_fill_8_bytes); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6225 vmovdqu(Address(to, 0), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6226 addptr(to, 32); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6227 subl(count, 8 << shift); |
8873
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6228 |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6229 BIND(L_check_fill_8_bytes); |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6230 // clean upper bits of YMM registers |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6231 vzeroupper(); |
7199 | 6232 } else { |
7475
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6233 // Fill 32-byte chunks |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6234 pshufd(xtmp, xtmp, 0); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6235 |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6236 subl(count, 8 << shift); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6237 jcc(Assembler::less, L_check_fill_8_bytes); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6238 align(16); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6239 |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6240 BIND(L_fill_32_bytes_loop); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6241 |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6242 if (UseUnalignedLoadStores) { |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6243 movdqu(Address(to, 0), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6244 movdqu(Address(to, 16), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6245 } else { |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6246 movq(Address(to, 0), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6247 movq(Address(to, 8), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6248 movq(Address(to, 16), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6249 movq(Address(to, 24), xtmp); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6250 } |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6251 |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6252 addptr(to, 32); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6253 subl(count, 8 << shift); |
e2e6bf86682c
8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents:
7474
diff
changeset
|
6254 jcc(Assembler::greaterEqual, L_fill_32_bytes_loop); |
8873
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6255 |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6256 BIND(L_check_fill_8_bytes); |
7199 | 6257 } |
6258 addl(count, 8 << shift); | |
6259 jccb(Assembler::zero, L_exit); | |
6260 jmpb(L_fill_8_bytes); | |
6261 | |
6262 // | |
6263 // length is too short, just fill qwords | |
6264 // | |
6265 BIND(L_fill_8_bytes_loop); | |
6266 movq(Address(to, 0), xtmp); | |
6267 addptr(to, 8); | |
6268 BIND(L_fill_8_bytes); | |
6269 subl(count, 1 << (shift + 1)); | |
6270 jcc(Assembler::greaterEqual, L_fill_8_bytes_loop); | |
6271 } | |
6272 } | |
6273 // fill trailing 4 bytes | |
6274 BIND(L_fill_4_bytes); | |
6275 testl(count, 1<<shift); | |
6276 jccb(Assembler::zero, L_fill_2_bytes); | |
6277 movl(Address(to, 0), value); | |
6278 if (t == T_BYTE || t == T_SHORT) { | |
6279 addptr(to, 4); | |
6280 BIND(L_fill_2_bytes); | |
6281 // fill trailing 2 bytes | |
6282 testl(count, 1<<(shift-1)); | |
6283 jccb(Assembler::zero, L_fill_byte); | |
6284 movw(Address(to, 0), value); | |
6285 if (t == T_BYTE) { | |
6286 addptr(to, 2); | |
6287 BIND(L_fill_byte); | |
6288 // fill trailing byte | |
6289 testl(count, 1); | |
6290 jccb(Assembler::zero, L_exit); | |
6291 movb(Address(to, 0), value); | |
6292 } else { | |
6293 BIND(L_fill_byte); | |
6294 } | |
6295 } else { | |
6296 BIND(L_fill_2_bytes); | |
6297 } | |
6298 BIND(L_exit); | |
6299 } | |
7637
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6300 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6301 // encode char[] to byte[] in ISO_8859_1 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6302 void MacroAssembler::encode_iso_array(Register src, Register dst, Register len, |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6303 XMMRegister tmp1Reg, XMMRegister tmp2Reg, |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6304 XMMRegister tmp3Reg, XMMRegister tmp4Reg, |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6305 Register tmp5, Register result) { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6306 // rsi: src |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6307 // rdi: dst |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6308 // rdx: len |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6309 // rcx: tmp5 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6310 // rax: result |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6311 ShortBranchVerifier sbv(this); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6312 assert_different_registers(src, dst, len, tmp5, result); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6313 Label L_done, L_copy_1_char, L_copy_1_char_exit; |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6314 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6315 // set result |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6316 xorl(result, result); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6317 // check for zero length |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6318 testl(len, len); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6319 jcc(Assembler::zero, L_done); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6320 movl(result, len); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6321 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6322 // Setup pointers |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6323 lea(src, Address(src, len, Address::times_2)); // char[] |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6324 lea(dst, Address(dst, len, Address::times_1)); // byte[] |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6325 negptr(len); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6326 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6327 if (UseSSE42Intrinsics || UseAVX >= 2) { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6328 Label L_chars_8_check, L_copy_8_chars, L_copy_8_chars_exit; |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6329 Label L_chars_16_check, L_copy_16_chars, L_copy_16_chars_exit; |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6330 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6331 if (UseAVX >= 2) { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6332 Label L_chars_32_check, L_copy_32_chars, L_copy_32_chars_exit; |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6333 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6334 movdl(tmp1Reg, tmp5); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6335 vpbroadcastd(tmp1Reg, tmp1Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6336 jmpb(L_chars_32_check); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6337 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6338 bind(L_copy_32_chars); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6339 vmovdqu(tmp3Reg, Address(src, len, Address::times_2, -64)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6340 vmovdqu(tmp4Reg, Address(src, len, Address::times_2, -32)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6341 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector256 */ true); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6342 vptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6343 jccb(Assembler::notZero, L_copy_32_chars_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6344 vpackuswb(tmp3Reg, tmp3Reg, tmp4Reg, /* vector256 */ true); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6345 vpermq(tmp4Reg, tmp3Reg, 0xD8, /* vector256 */ true); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6346 vmovdqu(Address(dst, len, Address::times_1, -32), tmp4Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6347 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6348 bind(L_chars_32_check); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6349 addptr(len, 32); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6350 jccb(Assembler::lessEqual, L_copy_32_chars); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6351 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6352 bind(L_copy_32_chars_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6353 subptr(len, 16); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6354 jccb(Assembler::greater, L_copy_16_chars_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6355 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6356 } else if (UseSSE42Intrinsics) { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6357 movl(tmp5, 0xff00ff00); // create mask to test for Unicode chars in vector |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6358 movdl(tmp1Reg, tmp5); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6359 pshufd(tmp1Reg, tmp1Reg, 0); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6360 jmpb(L_chars_16_check); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6361 } |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6362 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6363 bind(L_copy_16_chars); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6364 if (UseAVX >= 2) { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6365 vmovdqu(tmp2Reg, Address(src, len, Address::times_2, -32)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6366 vptest(tmp2Reg, tmp1Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6367 jccb(Assembler::notZero, L_copy_16_chars_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6368 vpackuswb(tmp2Reg, tmp2Reg, tmp1Reg, /* vector256 */ true); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6369 vpermq(tmp3Reg, tmp2Reg, 0xD8, /* vector256 */ true); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6370 } else { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6371 if (UseAVX > 0) { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6372 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6373 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6374 vpor(tmp2Reg, tmp3Reg, tmp4Reg, /* vector256 */ false); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6375 } else { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6376 movdqu(tmp3Reg, Address(src, len, Address::times_2, -32)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6377 por(tmp2Reg, tmp3Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6378 movdqu(tmp4Reg, Address(src, len, Address::times_2, -16)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6379 por(tmp2Reg, tmp4Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6380 } |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6381 ptest(tmp2Reg, tmp1Reg); // check for Unicode chars in vector |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6382 jccb(Assembler::notZero, L_copy_16_chars_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6383 packuswb(tmp3Reg, tmp4Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6384 } |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6385 movdqu(Address(dst, len, Address::times_1, -16), tmp3Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6386 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6387 bind(L_chars_16_check); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6388 addptr(len, 16); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6389 jccb(Assembler::lessEqual, L_copy_16_chars); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6390 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6391 bind(L_copy_16_chars_exit); |
8873
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6392 if (UseAVX >= 2) { |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6393 // clean upper bits of YMM registers |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6394 vzeroupper(); |
e961c11b85fe
8011102: Clear AVX registers after return from JNI call
kvn
parents:
8767
diff
changeset
|
6395 } |
7637
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6396 subptr(len, 8); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6397 jccb(Assembler::greater, L_copy_8_chars_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6398 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6399 bind(L_copy_8_chars); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6400 movdqu(tmp3Reg, Address(src, len, Address::times_2, -16)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6401 ptest(tmp3Reg, tmp1Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6402 jccb(Assembler::notZero, L_copy_8_chars_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6403 packuswb(tmp3Reg, tmp1Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6404 movq(Address(dst, len, Address::times_1, -8), tmp3Reg); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6405 addptr(len, 8); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6406 jccb(Assembler::lessEqual, L_copy_8_chars); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6407 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6408 bind(L_copy_8_chars_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6409 subptr(len, 8); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6410 jccb(Assembler::zero, L_done); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6411 } |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6412 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6413 bind(L_copy_1_char); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6414 load_unsigned_short(tmp5, Address(src, len, Address::times_2, 0)); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6415 testl(tmp5, 0xff00); // check if Unicode char |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6416 jccb(Assembler::notZero, L_copy_1_char_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6417 movb(Address(dst, len, Address::times_1, 0), tmp5); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6418 addptr(len, 1); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6419 jccb(Assembler::less, L_copy_1_char); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6420 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6421 bind(L_copy_1_char_exit); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6422 addptr(result, len); // len is negative count of not processed elements |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6423 bind(L_done); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6424 } |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7477
diff
changeset
|
6425 |
11080
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6426 /** |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6427 * Emits code to update CRC-32 with a byte value according to constants in table |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6428 * |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6429 * @param [in,out]crc Register containing the crc. |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6430 * @param [in]val Register containing the byte to fold into the CRC. |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6431 * @param [in]table Register containing the table of crc constants. |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6432 * |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6433 * uint32_t crc; |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6434 * val = crc_table[(val ^ crc) & 0xFF]; |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6435 * crc = val ^ (crc >> 8); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6436 * |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6437 */ |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6438 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6439 xorl(val, crc); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6440 andl(val, 0xFF); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6441 shrl(crc, 8); // unsigned shift |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6442 xorl(crc, Address(table, val, Address::times_4, 0)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6443 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6444 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6445 /** |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6446 * Fold 128-bit data chunk |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6447 */ |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6448 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, Register buf, int offset) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6449 vpclmulhdq(xtmp, xK, xcrc); // [123:64] |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6450 vpclmulldq(xcrc, xK, xcrc); // [63:0] |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6451 vpxor(xcrc, xcrc, Address(buf, offset), false /* vector256 */); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6452 pxor(xcrc, xtmp); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6453 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6454 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6455 void MacroAssembler::fold_128bit_crc32(XMMRegister xcrc, XMMRegister xK, XMMRegister xtmp, XMMRegister xbuf) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6456 vpclmulhdq(xtmp, xK, xcrc); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6457 vpclmulldq(xcrc, xK, xcrc); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6458 pxor(xcrc, xbuf); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6459 pxor(xcrc, xtmp); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6460 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6461 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6462 /** |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6463 * 8-bit folds to compute 32-bit CRC |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6464 * |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6465 * uint64_t xcrc; |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6466 * timesXtoThe32[xcrc & 0xFF] ^ (xcrc >> 8); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6467 */ |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6468 void MacroAssembler::fold_8bit_crc32(XMMRegister xcrc, Register table, XMMRegister xtmp, Register tmp) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6469 movdl(tmp, xcrc); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6470 andl(tmp, 0xFF); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6471 movdl(xtmp, Address(table, tmp, Address::times_4, 0)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6472 psrldq(xcrc, 1); // unsigned shift one byte |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6473 pxor(xcrc, xtmp); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6474 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6475 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6476 /** |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6477 * uint32_t crc; |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6478 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6479 */ |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6480 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6481 movl(tmp, crc); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6482 andl(tmp, 0xFF); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6483 shrl(crc, 8); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6484 xorl(crc, Address(table, tmp, Address::times_4, 0)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6485 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6486 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6487 /** |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6488 * @param crc register containing existing CRC (32-bit) |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6489 * @param buf register pointing to input byte buffer (byte*) |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6490 * @param len register containing number of bytes |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6491 * @param table register that will contain address of CRC table |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6492 * @param tmp scratch register |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6493 */ |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6494 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len, Register table, Register tmp) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6495 assert_different_registers(crc, buf, len, table, tmp, rax); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6496 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6497 Label L_tail, L_tail_restore, L_tail_loop, L_exit, L_align_loop, L_aligned; |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6498 Label L_fold_tail, L_fold_128b, L_fold_512b, L_fold_512b_loop, L_fold_tail_loop; |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6499 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6500 lea(table, ExternalAddress(StubRoutines::crc_table_addr())); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6501 notl(crc); // ~crc |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6502 cmpl(len, 16); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6503 jcc(Assembler::less, L_tail); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6504 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6505 // Align buffer to 16 bytes |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6506 movl(tmp, buf); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6507 andl(tmp, 0xF); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6508 jccb(Assembler::zero, L_aligned); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6509 subl(tmp, 16); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6510 addl(len, tmp); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6511 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6512 align(4); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6513 BIND(L_align_loop); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6514 movsbl(rax, Address(buf, 0)); // load byte with sign extension |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6515 update_byte_crc32(crc, rax, table); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6516 increment(buf); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6517 incrementl(tmp); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6518 jccb(Assembler::less, L_align_loop); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6519 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6520 BIND(L_aligned); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6521 movl(tmp, len); // save |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6522 shrl(len, 4); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6523 jcc(Assembler::zero, L_tail_restore); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6524 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6525 // Fold crc into first bytes of vector |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6526 movdqa(xmm1, Address(buf, 0)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6527 movdl(rax, xmm1); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6528 xorl(crc, rax); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6529 pinsrd(xmm1, crc, 0); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6530 addptr(buf, 16); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6531 subl(len, 4); // len > 0 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6532 jcc(Assembler::less, L_fold_tail); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6533 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6534 movdqa(xmm2, Address(buf, 0)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6535 movdqa(xmm3, Address(buf, 16)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6536 movdqa(xmm4, Address(buf, 32)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6537 addptr(buf, 48); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6538 subl(len, 3); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6539 jcc(Assembler::lessEqual, L_fold_512b); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6540 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6541 // Fold total 512 bits of polynomial on each iteration, |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6542 // 128 bits per each of 4 parallel streams. |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6543 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 32)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6544 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6545 align(32); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6546 BIND(L_fold_512b_loop); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6547 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6548 fold_128bit_crc32(xmm2, xmm0, xmm5, buf, 16); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6549 fold_128bit_crc32(xmm3, xmm0, xmm5, buf, 32); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6550 fold_128bit_crc32(xmm4, xmm0, xmm5, buf, 48); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6551 addptr(buf, 64); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6552 subl(len, 4); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6553 jcc(Assembler::greater, L_fold_512b_loop); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6554 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6555 // Fold 512 bits to 128 bits. |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6556 BIND(L_fold_512b); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6557 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6558 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm2); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6559 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm3); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6560 fold_128bit_crc32(xmm1, xmm0, xmm5, xmm4); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6561 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6562 // Fold the rest of 128 bits data chunks |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6563 BIND(L_fold_tail); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6564 addl(len, 3); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6565 jccb(Assembler::lessEqual, L_fold_128b); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6566 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr() + 16)); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6567 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6568 BIND(L_fold_tail_loop); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6569 fold_128bit_crc32(xmm1, xmm0, xmm5, buf, 0); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6570 addptr(buf, 16); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6571 decrementl(len); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6572 jccb(Assembler::greater, L_fold_tail_loop); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6573 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6574 // Fold 128 bits in xmm1 down into 32 bits in crc register. |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6575 BIND(L_fold_128b); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6576 movdqu(xmm0, ExternalAddress(StubRoutines::x86::crc_by128_masks_addr())); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6577 vpclmulqdq(xmm2, xmm0, xmm1, 0x1); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6578 vpand(xmm3, xmm0, xmm2, false /* vector256 */); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6579 vpclmulqdq(xmm0, xmm0, xmm3, 0x1); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6580 psrldq(xmm1, 8); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6581 psrldq(xmm2, 4); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6582 pxor(xmm0, xmm1); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6583 pxor(xmm0, xmm2); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6584 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6585 // 8 8-bit folds to compute 32-bit CRC. |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6586 for (int j = 0; j < 4; j++) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6587 fold_8bit_crc32(xmm0, table, xmm1, rax); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6588 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6589 movdl(crc, xmm0); // mov 32 bits to general register |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6590 for (int j = 0; j < 4; j++) { |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6591 fold_8bit_crc32(crc, table, rax); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6592 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6593 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6594 BIND(L_tail_restore); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6595 movl(len, tmp); // restore |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6596 BIND(L_tail); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6597 andl(len, 0xf); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6598 jccb(Assembler::zero, L_exit); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6599 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6600 // Fold the rest of bytes |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6601 align(4); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6602 BIND(L_tail_loop); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6603 movsbl(rax, Address(buf, 0)); // load byte with sign extension |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6604 update_byte_crc32(crc, rax, table); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6605 increment(buf); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6606 decrementl(len); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6607 jccb(Assembler::greater, L_tail_loop); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6608 |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6609 BIND(L_exit); |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6610 notl(crc); // ~c |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6611 } |
b800986664f4
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents:
8873
diff
changeset
|
6612 |
7199 | 6613 #undef BIND |
6614 #undef BLOCK_COMMENT | |
6615 | |
6616 | |
6617 Assembler::Condition MacroAssembler::negate_condition(Assembler::Condition cond) { | |
6618 switch (cond) { | |
6619 // Note some conditions are synonyms for others | |
6620 case Assembler::zero: return Assembler::notZero; | |
6621 case Assembler::notZero: return Assembler::zero; | |
6622 case Assembler::less: return Assembler::greaterEqual; | |
6623 case Assembler::lessEqual: return Assembler::greater; | |
6624 case Assembler::greater: return Assembler::lessEqual; | |
6625 case Assembler::greaterEqual: return Assembler::less; | |
6626 case Assembler::below: return Assembler::aboveEqual; | |
6627 case Assembler::belowEqual: return Assembler::above; | |
6628 case Assembler::above: return Assembler::belowEqual; | |
6629 case Assembler::aboveEqual: return Assembler::below; | |
6630 case Assembler::overflow: return Assembler::noOverflow; | |
6631 case Assembler::noOverflow: return Assembler::overflow; | |
6632 case Assembler::negative: return Assembler::positive; | |
6633 case Assembler::positive: return Assembler::negative; | |
6634 case Assembler::parity: return Assembler::noParity; | |
6635 case Assembler::noParity: return Assembler::parity; | |
6636 } | |
6637 ShouldNotReachHere(); return Assembler::overflow; | |
6638 } | |
6639 | |
6640 SkipIfEqual::SkipIfEqual( | |
6641 MacroAssembler* masm, const bool* flag_addr, bool value) { | |
6642 _masm = masm; | |
6643 _masm->cmp8(ExternalAddress((address)flag_addr), value); | |
6644 _masm->jcc(Assembler::equal, _label); | |
6645 } | |
6646 | |
6647 SkipIfEqual::~SkipIfEqual() { | |
6648 _masm->bind(_label); | |
6649 } |