1 /*
2 * Copyright (c) 2016, 2018, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2016, 2018, SAP SE. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "asm/codeBuffer.hpp"
28 #include "asm/macroAssembler.inline.hpp"
29 #include "compiler/disassembler.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/barrierSetAssembler.hpp"
32 #include "gc/shared/collectedHeap.inline.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "gc/shared/cardTableBarrierSet.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "memory/universe.hpp"
37 #include "oops/accessDecorators.hpp"
38 #include "oops/compressedOops.inline.hpp"
39 #include "oops/klass.inline.hpp"
40 #include "opto/compile.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/matcher.hpp"
43 #include "prims/methodHandles.hpp"
44 #include "registerSaver_s390.hpp"
45 #include "runtime/biasedLocking.hpp"
46 #include "runtime/icache.hpp"
47 #include "runtime/interfaceSupport.inline.hpp"
48 #include "runtime/objectMonitor.hpp"
49 #include "runtime/os.hpp"
50 #include "runtime/safepoint.hpp"
51 #include "runtime/safepointMechanism.hpp"
52 #include "runtime/sharedRuntime.hpp"
53 #include "runtime/stubRoutines.hpp"
54 #include "utilities/events.hpp"
55 #include "utilities/macros.hpp"
56
57 #include <ucontext.h>
58
59 #define BLOCK_COMMENT(str) block_comment(str)
60 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
61
62 // Move 32-bit register if destination and source are different.
63 void MacroAssembler::lr_if_needed(Register rd, Register rs) {
64 if (rs != rd) { z_lr(rd, rs); }
65 }
66
67 // Move register if destination and source are different.
68 void MacroAssembler::lgr_if_needed(Register rd, Register rs) {
69 if (rs != rd) { z_lgr(rd, rs); }
70 }
71
72 // Zero-extend 32-bit register into 64-bit register if destination and source are different.
73 void MacroAssembler::llgfr_if_needed(Register rd, Register rs) {
74 if (rs != rd) { z_llgfr(rd, rs); }
75 }
76
77 // Move float register if destination and source are different.
78 void MacroAssembler::ldr_if_needed(FloatRegister rd, FloatRegister rs) {
79 if (rs != rd) { z_ldr(rd, rs); }
80 }
81
82 // Move integer register if destination and source are different.
83 // It is assumed that shorter-than-int types are already
84 // appropriately sign-extended.
85 void MacroAssembler::move_reg_if_needed(Register dst, BasicType dst_type, Register src,
86 BasicType src_type) {
87 assert((dst_type != T_FLOAT) && (dst_type != T_DOUBLE), "use move_freg for float types");
88 assert((src_type != T_FLOAT) && (src_type != T_DOUBLE), "use move_freg for float types");
89
90 if (dst_type == src_type) {
91 lgr_if_needed(dst, src); // Just move all 64 bits.
92 return;
93 }
94
95 switch (dst_type) {
96 // Do not support these types for now.
97 // case T_BOOLEAN:
98 case T_BYTE: // signed byte
99 switch (src_type) {
100 case T_INT:
101 z_lgbr(dst, src);
102 break;
103 default:
104 ShouldNotReachHere();
105 }
106 return;
107
108 case T_CHAR:
109 case T_SHORT:
110 switch (src_type) {
111 case T_INT:
112 if (dst_type == T_CHAR) {
113 z_llghr(dst, src);
114 } else {
115 z_lghr(dst, src);
116 }
117 break;
118 default:
119 ShouldNotReachHere();
120 }
121 return;
122
123 case T_INT:
124 switch (src_type) {
125 case T_BOOLEAN:
126 case T_BYTE:
127 case T_CHAR:
128 case T_SHORT:
129 case T_INT:
130 case T_LONG:
131 case T_OBJECT:
132 case T_ARRAY:
133 case T_VOID:
134 case T_ADDRESS:
135 lr_if_needed(dst, src);
136 // llgfr_if_needed(dst, src); // zero-extend (in case we need to find a bug).
137 return;
138
139 default:
140 assert(false, "non-integer src type");
141 return;
142 }
143 case T_LONG:
144 switch (src_type) {
145 case T_BOOLEAN:
146 case T_BYTE:
147 case T_CHAR:
148 case T_SHORT:
149 case T_INT:
150 z_lgfr(dst, src); // sign extension
151 return;
152
153 case T_LONG:
154 case T_OBJECT:
155 case T_ARRAY:
156 case T_VOID:
157 case T_ADDRESS:
158 lgr_if_needed(dst, src);
159 return;
160
161 default:
162 assert(false, "non-integer src type");
163 return;
164 }
165 return;
166 case T_OBJECT:
167 case T_ARRAY:
168 case T_VOID:
169 case T_ADDRESS:
170 switch (src_type) {
171 // These types don't make sense to be converted to pointers:
172 // case T_BOOLEAN:
173 // case T_BYTE:
174 // case T_CHAR:
175 // case T_SHORT:
176
177 case T_INT:
178 z_llgfr(dst, src); // zero extension
179 return;
180
181 case T_LONG:
182 case T_OBJECT:
183 case T_ARRAY:
184 case T_VOID:
185 case T_ADDRESS:
186 lgr_if_needed(dst, src);
187 return;
188
189 default:
190 assert(false, "non-integer src type");
191 return;
192 }
193 return;
194 default:
195 assert(false, "non-integer dst type");
196 return;
197 }
198 }
199
200 // Move float register if destination and source are different.
201 void MacroAssembler::move_freg_if_needed(FloatRegister dst, BasicType dst_type,
202 FloatRegister src, BasicType src_type) {
203 assert((dst_type == T_FLOAT) || (dst_type == T_DOUBLE), "use move_reg for int types");
204 assert((src_type == T_FLOAT) || (src_type == T_DOUBLE), "use move_reg for int types");
205 if (dst_type == src_type) {
206 ldr_if_needed(dst, src); // Just move all 64 bits.
207 } else {
208 switch (dst_type) {
209 case T_FLOAT:
210 assert(src_type == T_DOUBLE, "invalid float type combination");
211 z_ledbr(dst, src);
212 return;
213 case T_DOUBLE:
214 assert(src_type == T_FLOAT, "invalid float type combination");
215 z_ldebr(dst, src);
216 return;
217 default:
218 assert(false, "non-float dst type");
219 return;
220 }
221 }
222 }
223
224 // Optimized emitter for reg to mem operations.
225 // Uses modern instructions if running on modern hardware, classic instructions
226 // otherwise. Prefers (usually shorter) classic instructions if applicable.
227 // Data register (reg) cannot be used as work register.
228 //
229 // Don't rely on register locking, instead pass a scratch register (Z_R0 by default).
230 // CAUTION! Passing registers >= Z_R2 may produce bad results on old CPUs!
231 void MacroAssembler::freg2mem_opt(FloatRegister reg,
232 int64_t disp,
233 Register index,
234 Register base,
235 void (MacroAssembler::*modern) (FloatRegister, int64_t, Register, Register),
236 void (MacroAssembler::*classic)(FloatRegister, int64_t, Register, Register),
237 Register scratch) {
238 index = (index == noreg) ? Z_R0 : index;
239 if (Displacement::is_shortDisp(disp)) {
240 (this->*classic)(reg, disp, index, base);
241 } else {
242 if (Displacement::is_validDisp(disp)) {
243 (this->*modern)(reg, disp, index, base);
244 } else {
245 if (scratch != Z_R0 && scratch != Z_R1) {
246 (this->*modern)(reg, disp, index, base); // Will fail with disp out of range.
247 } else {
248 if (scratch != Z_R0) { // scratch == Z_R1
249 if ((scratch == index) || (index == base)) {
250 (this->*modern)(reg, disp, index, base); // Will fail with disp out of range.
251 } else {
252 add2reg(scratch, disp, base);
253 (this->*classic)(reg, 0, index, scratch);
254 if (base == scratch) {
255 add2reg(base, -disp); // Restore base.
256 }
257 }
258 } else { // scratch == Z_R0
259 z_lgr(scratch, base);
260 add2reg(base, disp);
261 (this->*classic)(reg, 0, index, base);
262 z_lgr(base, scratch); // Restore base.
263 }
264 }
265 }
266 }
267 }
268
269 void MacroAssembler::freg2mem_opt(FloatRegister reg, const Address &a, bool is_double) {
270 if (is_double) {
271 freg2mem_opt(reg, a.disp20(), a.indexOrR0(), a.baseOrR0(), MODERN_FFUN(z_stdy), CLASSIC_FFUN(z_std));
272 } else {
273 freg2mem_opt(reg, a.disp20(), a.indexOrR0(), a.baseOrR0(), MODERN_FFUN(z_stey), CLASSIC_FFUN(z_ste));
274 }
275 }
276
277 // Optimized emitter for mem to reg operations.
278 // Uses modern instructions if running on modern hardware, classic instructions
279 // otherwise. Prefers (usually shorter) classic instructions if applicable.
280 // data register (reg) cannot be used as work register.
281 //
282 // Don't rely on register locking, instead pass a scratch register (Z_R0 by default).
283 // CAUTION! Passing registers >= Z_R2 may produce bad results on old CPUs!
284 void MacroAssembler::mem2freg_opt(FloatRegister reg,
285 int64_t disp,
286 Register index,
287 Register base,
288 void (MacroAssembler::*modern) (FloatRegister, int64_t, Register, Register),
289 void (MacroAssembler::*classic)(FloatRegister, int64_t, Register, Register),
290 Register scratch) {
291 index = (index == noreg) ? Z_R0 : index;
292 if (Displacement::is_shortDisp(disp)) {
293 (this->*classic)(reg, disp, index, base);
294 } else {
295 if (Displacement::is_validDisp(disp)) {
296 (this->*modern)(reg, disp, index, base);
297 } else {
298 if (scratch != Z_R0 && scratch != Z_R1) {
299 (this->*modern)(reg, disp, index, base); // Will fail with disp out of range.
300 } else {
301 if (scratch != Z_R0) { // scratch == Z_R1
302 if ((scratch == index) || (index == base)) {
303 (this->*modern)(reg, disp, index, base); // Will fail with disp out of range.
304 } else {
305 add2reg(scratch, disp, base);
306 (this->*classic)(reg, 0, index, scratch);
307 if (base == scratch) {
308 add2reg(base, -disp); // Restore base.
309 }
310 }
311 } else { // scratch == Z_R0
312 z_lgr(scratch, base);
313 add2reg(base, disp);
314 (this->*classic)(reg, 0, index, base);
315 z_lgr(base, scratch); // Restore base.
316 }
317 }
318 }
319 }
320 }
321
322 void MacroAssembler::mem2freg_opt(FloatRegister reg, const Address &a, bool is_double) {
323 if (is_double) {
324 mem2freg_opt(reg, a.disp20(), a.indexOrR0(), a.baseOrR0(), MODERN_FFUN(z_ldy), CLASSIC_FFUN(z_ld));
325 } else {
326 mem2freg_opt(reg, a.disp20(), a.indexOrR0(), a.baseOrR0(), MODERN_FFUN(z_ley), CLASSIC_FFUN(z_le));
327 }
328 }
329
330 // Optimized emitter for reg to mem operations.
331 // Uses modern instructions if running on modern hardware, classic instructions
332 // otherwise. Prefers (usually shorter) classic instructions if applicable.
333 // Data register (reg) cannot be used as work register.
334 //
335 // Don't rely on register locking, instead pass a scratch register
336 // (Z_R0 by default)
337 // CAUTION! passing registers >= Z_R2 may produce bad results on old CPUs!
338 void MacroAssembler::reg2mem_opt(Register reg,
339 int64_t disp,
340 Register index,
341 Register base,
342 void (MacroAssembler::*modern) (Register, int64_t, Register, Register),
343 void (MacroAssembler::*classic)(Register, int64_t, Register, Register),
344 Register scratch) {
345 index = (index == noreg) ? Z_R0 : index;
346 if (Displacement::is_shortDisp(disp)) {
347 (this->*classic)(reg, disp, index, base);
348 } else {
349 if (Displacement::is_validDisp(disp)) {
350 (this->*modern)(reg, disp, index, base);
351 } else {
352 if (scratch != Z_R0 && scratch != Z_R1) {
353 (this->*modern)(reg, disp, index, base); // Will fail with disp out of range.
354 } else {
355 if (scratch != Z_R0) { // scratch == Z_R1
356 if ((scratch == index) || (index == base)) {
357 (this->*modern)(reg, disp, index, base); // Will fail with disp out of range.
358 } else {
359 add2reg(scratch, disp, base);
360 (this->*classic)(reg, 0, index, scratch);
361 if (base == scratch) {
362 add2reg(base, -disp); // Restore base.
363 }
364 }
365 } else { // scratch == Z_R0
366 if ((scratch == reg) || (scratch == base) || (reg == base)) {
367 (this->*modern)(reg, disp, index, base); // Will fail with disp out of range.
368 } else {
369 z_lgr(scratch, base);
370 add2reg(base, disp);
371 (this->*classic)(reg, 0, index, base);
372 z_lgr(base, scratch); // Restore base.
373 }
374 }
375 }
376 }
377 }
378 }
379
380 int MacroAssembler::reg2mem_opt(Register reg, const Address &a, bool is_double) {
381 int store_offset = offset();
382 if (is_double) {
383 reg2mem_opt(reg, a.disp20(), a.indexOrR0(), a.baseOrR0(), MODERN_IFUN(z_stg), CLASSIC_IFUN(z_stg));
384 } else {
385 reg2mem_opt(reg, a.disp20(), a.indexOrR0(), a.baseOrR0(), MODERN_IFUN(z_sty), CLASSIC_IFUN(z_st));
386 }
387 return store_offset;
388 }
389
390 // Optimized emitter for mem to reg operations.
391 // Uses modern instructions if running on modern hardware, classic instructions
392 // otherwise. Prefers (usually shorter) classic instructions if applicable.
393 // Data register (reg) will be used as work register where possible.
394 void MacroAssembler::mem2reg_opt(Register reg,
395 int64_t disp,
396 Register index,
397 Register base,
398 void (MacroAssembler::*modern) (Register, int64_t, Register, Register),
399 void (MacroAssembler::*classic)(Register, int64_t, Register, Register)) {
400 index = (index == noreg) ? Z_R0 : index;
401 if (Displacement::is_shortDisp(disp)) {
402 (this->*classic)(reg, disp, index, base);
403 } else {
404 if (Displacement::is_validDisp(disp)) {
405 (this->*modern)(reg, disp, index, base);
406 } else {
407 if ((reg == index) && (reg == base)) {
408 z_sllg(reg, reg, 1);
409 add2reg(reg, disp);
410 (this->*classic)(reg, 0, noreg, reg);
411 } else if ((reg == index) && (reg != Z_R0)) {
412 add2reg(reg, disp);
413 (this->*classic)(reg, 0, reg, base);
414 } else if (reg == base) {
415 add2reg(reg, disp);
416 (this->*classic)(reg, 0, index, reg);
417 } else if (reg != Z_R0) {
418 add2reg(reg, disp, base);
419 (this->*classic)(reg, 0, index, reg);
420 } else { // reg == Z_R0 && reg != base here
421 add2reg(base, disp);
422 (this->*classic)(reg, 0, index, base);
423 add2reg(base, -disp);
424 }
425 }
426 }
427 }
428
429 void MacroAssembler::mem2reg_opt(Register reg, const Address &a, bool is_double) {
430 if (is_double) {
431 z_lg(reg, a);
432 } else {
433 mem2reg_opt(reg, a.disp20(), a.indexOrR0(), a.baseOrR0(), MODERN_IFUN(z_ly), CLASSIC_IFUN(z_l));
434 }
435 }
436
437 void MacroAssembler::mem2reg_signed_opt(Register reg, const Address &a) {
438 mem2reg_opt(reg, a.disp20(), a.indexOrR0(), a.baseOrR0(), MODERN_IFUN(z_lgf), CLASSIC_IFUN(z_lgf));
439 }
440
441 void MacroAssembler::and_imm(Register r, long mask,
442 Register tmp /* = Z_R0 */,
443 bool wide /* = false */) {
444 assert(wide || Immediate::is_simm32(mask), "mask value too large");
445
446 if (!wide) {
447 z_nilf(r, mask);
448 return;
449 }
450
451 assert(r != tmp, " need a different temporary register !");
452 load_const_optimized(tmp, mask);
453 z_ngr(r, tmp);
454 }
455
456 // Calculate the 1's complement.
457 // Note: The condition code is neither preserved nor correctly set by this code!!!
458 // Note: (wide == false) does not protect the high order half of the target register
459 // from alteration. It only serves as optimization hint for 32-bit results.
460 void MacroAssembler::not_(Register r1, Register r2, bool wide) {
461
462 if ((r2 == noreg) || (r2 == r1)) { // Calc 1's complement in place.
463 z_xilf(r1, -1);
464 if (wide) {
465 z_xihf(r1, -1);
466 }
467 } else { // Distinct src and dst registers.
468 if (VM_Version::has_DistinctOpnds()) {
469 load_const_optimized(r1, -1);
470 z_xgrk(r1, r2, r1);
471 } else {
472 if (wide) {
473 z_lgr(r1, r2);
474 z_xilf(r1, -1);
475 z_xihf(r1, -1);
476 } else {
477 z_lr(r1, r2);
478 z_xilf(r1, -1);
479 }
480 }
481 }
482 }
483
484 unsigned long MacroAssembler::create_mask(int lBitPos, int rBitPos) {
485 assert(lBitPos >= 0, "zero is leftmost bit position");
486 assert(rBitPos <= 63, "63 is rightmost bit position");
487 assert(lBitPos <= rBitPos, "inverted selection interval");
488 return (lBitPos == 0 ? (unsigned long)(-1L) : ((1UL<<(63-lBitPos+1))-1)) & (~((1UL<<(63-rBitPos))-1));
489 }
490
491 // Helper function for the "Rotate_then_<logicalOP>" emitters.
492 // Rotate src, then mask register contents such that only bits in range survive.
493 // For oneBits == false, all bits not in range are set to 0. Useful for deleting all bits outside range.
494 // For oneBits == true, all bits not in range are set to 1. Useful for preserving all bits outside range.
495 // The caller must ensure that the selected range only contains bits with defined value.
496 void MacroAssembler::rotate_then_mask(Register dst, Register src, int lBitPos, int rBitPos,
497 int nRotate, bool src32bit, bool dst32bit, bool oneBits) {
498 assert(!(dst32bit && lBitPos < 32), "selection interval out of range for int destination");
499 bool sll4rll = (nRotate >= 0) && (nRotate <= (63-rBitPos)); // Substitute SLL(G) for RLL(G).
500 bool srl4rll = (nRotate < 0) && (-nRotate <= lBitPos); // Substitute SRL(G) for RLL(G).
501 // Pre-determine which parts of dst will be zero after shift/rotate.
502 bool llZero = sll4rll && (nRotate >= 16);
503 bool lhZero = (sll4rll && (nRotate >= 32)) || (srl4rll && (nRotate <= -48));
504 bool lfZero = llZero && lhZero;
505 bool hlZero = (sll4rll && (nRotate >= 48)) || (srl4rll && (nRotate <= -32));
506 bool hhZero = (srl4rll && (nRotate <= -16));
507 bool hfZero = hlZero && hhZero;
508
509 // rotate then mask src operand.
510 // if oneBits == true, all bits outside selected range are 1s.
511 // if oneBits == false, all bits outside selected range are 0s.
512 if (src32bit) { // There might be garbage in the upper 32 bits which will get masked away.
513 if (dst32bit) {
514 z_rll(dst, src, nRotate); // Copy and rotate, upper half of reg remains undisturbed.
515 } else {
516 if (sll4rll) { z_sllg(dst, src, nRotate); }
517 else if (srl4rll) { z_srlg(dst, src, -nRotate); }
518 else { z_rllg(dst, src, nRotate); }
519 }
520 } else {
521 if (sll4rll) { z_sllg(dst, src, nRotate); }
522 else if (srl4rll) { z_srlg(dst, src, -nRotate); }
523 else { z_rllg(dst, src, nRotate); }
524 }
525
526 unsigned long range_mask = create_mask(lBitPos, rBitPos);
527 unsigned int range_mask_h = (unsigned int)(range_mask >> 32);
528 unsigned int range_mask_l = (unsigned int)range_mask;
529 unsigned short range_mask_hh = (unsigned short)(range_mask >> 48);
530 unsigned short range_mask_hl = (unsigned short)(range_mask >> 32);
531 unsigned short range_mask_lh = (unsigned short)(range_mask >> 16);
532 unsigned short range_mask_ll = (unsigned short)range_mask;
533 // Works for z9 and newer H/W.
534 if (oneBits) {
535 if ((~range_mask_l) != 0) { z_oilf(dst, ~range_mask_l); } // All bits outside range become 1s.
536 if (((~range_mask_h) != 0) && !dst32bit) { z_oihf(dst, ~range_mask_h); }
537 } else {
538 // All bits outside range become 0s
539 if (((~range_mask_l) != 0) && !lfZero) {
540 z_nilf(dst, range_mask_l);
541 }
542 if (((~range_mask_h) != 0) && !dst32bit && !hfZero) {
543 z_nihf(dst, range_mask_h);
544 }
545 }
546 }
547
548 // Rotate src, then insert selected range from rotated src into dst.
549 // Clear dst before, if requested.
550 void MacroAssembler::rotate_then_insert(Register dst, Register src, int lBitPos, int rBitPos,
551 int nRotate, bool clear_dst) {
552 // This version does not depend on src being zero-extended int2long.
553 nRotate &= 0x003f; // For risbg, pretend it's an unsigned value.
554 z_risbg(dst, src, lBitPos, rBitPos, nRotate, clear_dst); // Rotate, then insert selected, clear the rest.
555 }
556
557 // Rotate src, then and selected range from rotated src into dst.
558 // Set condition code only if so requested. Otherwise it is unpredictable.
559 // See performance note in macroAssembler_s390.hpp for important information.
560 void MacroAssembler::rotate_then_and(Register dst, Register src, int lBitPos, int rBitPos,
561 int nRotate, bool test_only) {
562 guarantee(!test_only, "Emitter not fit for test_only instruction variant.");
563 // This version does not depend on src being zero-extended int2long.
564 nRotate &= 0x003f; // For risbg, pretend it's an unsigned value.
565 z_rxsbg(dst, src, lBitPos, rBitPos, nRotate, test_only); // Rotate, then xor selected.
566 }
567
568 // Rotate src, then or selected range from rotated src into dst.
569 // Set condition code only if so requested. Otherwise it is unpredictable.
570 // See performance note in macroAssembler_s390.hpp for important information.
571 void MacroAssembler::rotate_then_or(Register dst, Register src, int lBitPos, int rBitPos,
572 int nRotate, bool test_only) {
573 guarantee(!test_only, "Emitter not fit for test_only instruction variant.");
574 // This version does not depend on src being zero-extended int2long.
575 nRotate &= 0x003f; // For risbg, pretend it's an unsigned value.
576 z_rosbg(dst, src, lBitPos, rBitPos, nRotate, test_only); // Rotate, then xor selected.
577 }
578
579 // Rotate src, then xor selected range from rotated src into dst.
580 // Set condition code only if so requested. Otherwise it is unpredictable.
581 // See performance note in macroAssembler_s390.hpp for important information.
582 void MacroAssembler::rotate_then_xor(Register dst, Register src, int lBitPos, int rBitPos,
583 int nRotate, bool test_only) {
584 guarantee(!test_only, "Emitter not fit for test_only instruction variant.");
585 // This version does not depend on src being zero-extended int2long.
586 nRotate &= 0x003f; // For risbg, pretend it's an unsigned value.
587 z_rxsbg(dst, src, lBitPos, rBitPos, nRotate, test_only); // Rotate, then xor selected.
588 }
589
590 void MacroAssembler::add64(Register r1, RegisterOrConstant inc) {
591 if (inc.is_register()) {
592 z_agr(r1, inc.as_register());
593 } else { // constant
594 intptr_t imm = inc.as_constant();
595 add2reg(r1, imm);
596 }
597 }
598 // Helper function to multiply the 64bit contents of a register by a 16bit constant.
599 // The optimization tries to avoid the mghi instruction, since it uses the FPU for
600 // calculation and is thus rather slow.
601 //
602 // There is no handling for special cases, e.g. cval==0 or cval==1.
603 //
604 // Returns len of generated code block.
605 unsigned int MacroAssembler::mul_reg64_const16(Register rval, Register work, int cval) {
606 int block_start = offset();
607
608 bool sign_flip = cval < 0;
609 cval = sign_flip ? -cval : cval;
610
611 BLOCK_COMMENT("Reg64*Con16 {");
612
613 int bit1 = cval & -cval;
614 if (bit1 == cval) {
615 z_sllg(rval, rval, exact_log2(bit1));
616 if (sign_flip) { z_lcgr(rval, rval); }
617 } else {
618 int bit2 = (cval-bit1) & -(cval-bit1);
619 if ((bit1+bit2) == cval) {
620 z_sllg(work, rval, exact_log2(bit1));
621 z_sllg(rval, rval, exact_log2(bit2));
622 z_agr(rval, work);
623 if (sign_flip) { z_lcgr(rval, rval); }
624 } else {
625 if (sign_flip) { z_mghi(rval, -cval); }
626 else { z_mghi(rval, cval); }
627 }
628 }
629 BLOCK_COMMENT("} Reg64*Con16");
630
631 int block_end = offset();
632 return block_end - block_start;
633 }
634
635 // Generic operation r1 := r2 + imm.
636 //
637 // Should produce the best code for each supported CPU version.
638 // r2 == noreg yields r1 := r1 + imm
639 // imm == 0 emits either no instruction or r1 := r2 !
640 // NOTES: 1) Don't use this function where fixed sized
641 // instruction sequences are required!!!
642 // 2) Don't use this function if condition code
643 // setting is required!
644 // 3) Despite being declared as int64_t, the parameter imm
645 // must be a simm_32 value (= signed 32-bit integer).
646 void MacroAssembler::add2reg(Register r1, int64_t imm, Register r2) {
647 assert(Immediate::is_simm32(imm), "probably an implicit conversion went wrong");
648
649 if (r2 == noreg) { r2 = r1; }
650
651 // Handle special case imm == 0.
652 if (imm == 0) {
653 lgr_if_needed(r1, r2);
654 // Nothing else to do.
655 return;
656 }
657
658 if (!PreferLAoverADD || (r2 == Z_R0)) {
659 bool distinctOpnds = VM_Version::has_DistinctOpnds();
660
661 // Can we encode imm in 16 bits signed?
662 if (Immediate::is_simm16(imm)) {
663 if (r1 == r2) {
664 z_aghi(r1, imm);
665 return;
666 }
667 if (distinctOpnds) {
668 z_aghik(r1, r2, imm);
669 return;
670 }
671 z_lgr(r1, r2);
672 z_aghi(r1, imm);
673 return;
674 }
675 } else {
676 // Can we encode imm in 12 bits unsigned?
677 if (Displacement::is_shortDisp(imm)) {
678 z_la(r1, imm, r2);
679 return;
680 }
681 // Can we encode imm in 20 bits signed?
682 if (Displacement::is_validDisp(imm)) {
683 // Always use LAY instruction, so we don't need the tmp register.
684 z_lay(r1, imm, r2);
685 return;
686 }
687
688 }
689
690 // Can handle it (all possible values) with long immediates.
691 lgr_if_needed(r1, r2);
692 z_agfi(r1, imm);
693 }
694
695 // Generic operation r := b + x + d
696 //
697 // Addition of several operands with address generation semantics - sort of:
698 // - no restriction on the registers. Any register will do for any operand.
699 // - x == noreg: operand will be disregarded.
700 // - b == noreg: will use (contents of) result reg as operand (r := r + d).
701 // - x == Z_R0: just disregard
702 // - b == Z_R0: use as operand. This is not address generation semantics!!!
703 //
704 // The same restrictions as on add2reg() are valid!!!
705 void MacroAssembler::add2reg_with_index(Register r, int64_t d, Register x, Register b) {
706 assert(Immediate::is_simm32(d), "probably an implicit conversion went wrong");
707
708 if (x == noreg) { x = Z_R0; }
709 if (b == noreg) { b = r; }
710
711 // Handle special case x == R0.
712 if (x == Z_R0) {
713 // Can simply add the immediate value to the base register.
714 add2reg(r, d, b);
715 return;
716 }
717
718 if (!PreferLAoverADD || (b == Z_R0)) {
719 bool distinctOpnds = VM_Version::has_DistinctOpnds();
720 // Handle special case d == 0.
721 if (d == 0) {
722 if (b == x) { z_sllg(r, b, 1); return; }
723 if (r == x) { z_agr(r, b); return; }
724 if (r == b) { z_agr(r, x); return; }
725 if (distinctOpnds) { z_agrk(r, x, b); return; }
726 z_lgr(r, b);
727 z_agr(r, x);
728 } else {
729 if (x == b) { z_sllg(r, x, 1); }
730 else if (r == x) { z_agr(r, b); }
731 else if (r == b) { z_agr(r, x); }
732 else if (distinctOpnds) { z_agrk(r, x, b); }
733 else {
734 z_lgr(r, b);
735 z_agr(r, x);
736 }
737 add2reg(r, d);
738 }
739 } else {
740 // Can we encode imm in 12 bits unsigned?
741 if (Displacement::is_shortDisp(d)) {
742 z_la(r, d, x, b);
743 return;
744 }
745 // Can we encode imm in 20 bits signed?
746 if (Displacement::is_validDisp(d)) {
747 z_lay(r, d, x, b);
748 return;
749 }
750 z_la(r, 0, x, b);
751 add2reg(r, d);
752 }
753 }
754
755 // Generic emitter (32bit) for direct memory increment.
756 // For optimal code, do not specify Z_R0 as temp register.
757 void MacroAssembler::add2mem_32(const Address &a, int64_t imm, Register tmp) {
758 if (VM_Version::has_MemWithImmALUOps() && Immediate::is_simm8(imm)) {
759 z_asi(a, imm);
760 } else {
761 z_lgf(tmp, a);
762 add2reg(tmp, imm);
763 z_st(tmp, a);
764 }
765 }
766
767 void MacroAssembler::add2mem_64(const Address &a, int64_t imm, Register tmp) {
768 if (VM_Version::has_MemWithImmALUOps() && Immediate::is_simm8(imm)) {
769 z_agsi(a, imm);
770 } else {
771 z_lg(tmp, a);
772 add2reg(tmp, imm);
773 z_stg(tmp, a);
774 }
775 }
776
777 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed) {
778 switch (size_in_bytes) {
779 case 8: z_lg(dst, src); break;
780 case 4: is_signed ? z_lgf(dst, src) : z_llgf(dst, src); break;
781 case 2: is_signed ? z_lgh(dst, src) : z_llgh(dst, src); break;
782 case 1: is_signed ? z_lgb(dst, src) : z_llgc(dst, src); break;
783 default: ShouldNotReachHere();
784 }
785 }
786
787 void MacroAssembler::store_sized_value(Register src, Address dst, size_t size_in_bytes) {
788 switch (size_in_bytes) {
789 case 8: z_stg(src, dst); break;
790 case 4: z_st(src, dst); break;
791 case 2: z_sth(src, dst); break;
792 case 1: z_stc(src, dst); break;
793 default: ShouldNotReachHere();
794 }
795 }
796
797 // Split a si20 offset (20bit, signed) into an ui12 offset (12bit, unsigned) and
798 // a high-order summand in register tmp.
799 //
800 // return value: < 0: No split required, si20 actually has property uimm12.
801 // >= 0: Split performed. Use return value as uimm12 displacement and
802 // tmp as index register.
803 int MacroAssembler::split_largeoffset(int64_t si20_offset, Register tmp, bool fixed_codelen, bool accumulate) {
804 assert(Immediate::is_simm20(si20_offset), "sanity");
805 int lg_off = (int)si20_offset & 0x0fff; // Punch out low-order 12 bits, always positive.
806 int ll_off = (int)si20_offset & ~0x0fff; // Force low-order 12 bits to zero.
807 assert((Displacement::is_shortDisp(si20_offset) && (ll_off == 0)) ||
808 !Displacement::is_shortDisp(si20_offset), "unexpected offset values");
809 assert((lg_off+ll_off) == si20_offset, "offset splitup error");
810
811 Register work = accumulate? Z_R0 : tmp;
812
813 if (fixed_codelen) { // Len of code = 10 = 4 + 6.
814 z_lghi(work, ll_off>>12); // Implicit sign extension.
815 z_slag(work, work, 12);
816 } else { // Len of code = 0..10.
817 if (ll_off == 0) { return -1; }
818 // ll_off has 8 significant bits (at most) plus sign.
819 if ((ll_off & 0x0000f000) == 0) { // Non-zero bits only in upper halfbyte.
820 z_llilh(work, ll_off >> 16);
821 if (ll_off < 0) { // Sign-extension required.
822 z_lgfr(work, work);
823 }
824 } else {
825 if ((ll_off & 0x000f0000) == 0) { // Non-zero bits only in lower halfbyte.
826 z_llill(work, ll_off);
827 } else { // Non-zero bits in both halfbytes.
828 z_lghi(work, ll_off>>12); // Implicit sign extension.
829 z_slag(work, work, 12);
830 }
831 }
832 }
833 if (accumulate) { z_algr(tmp, work); } // len of code += 4
834 return lg_off;
835 }
836
837 void MacroAssembler::load_float_largeoffset(FloatRegister t, int64_t si20, Register a, Register tmp) {
838 if (Displacement::is_validDisp(si20)) {
839 z_ley(t, si20, a);
840 } else {
841 // Fixed_codelen = true is a simple way to ensure that the size of load_float_largeoffset
842 // does not depend on si20 (scratch buffer emit size == code buffer emit size for constant
843 // pool loads).
844 bool accumulate = true;
845 bool fixed_codelen = true;
846 Register work;
847
848 if (fixed_codelen) {
849 z_lgr(tmp, a); // Lgr_if_needed not applicable due to fixed_codelen.
850 } else {
851 accumulate = (a == tmp);
852 }
853 work = tmp;
854
855 int disp12 = split_largeoffset(si20, work, fixed_codelen, accumulate);
856 if (disp12 < 0) {
857 z_le(t, si20, work);
858 } else {
859 if (accumulate) {
860 z_le(t, disp12, work);
861 } else {
862 z_le(t, disp12, work, a);
863 }
864 }
865 }
866 }
867
868 void MacroAssembler::load_double_largeoffset(FloatRegister t, int64_t si20, Register a, Register tmp) {
869 if (Displacement::is_validDisp(si20)) {
870 z_ldy(t, si20, a);
871 } else {
872 // Fixed_codelen = true is a simple way to ensure that the size of load_double_largeoffset
873 // does not depend on si20 (scratch buffer emit size == code buffer emit size for constant
874 // pool loads).
875 bool accumulate = true;
876 bool fixed_codelen = true;
877 Register work;
878
879 if (fixed_codelen) {
880 z_lgr(tmp, a); // Lgr_if_needed not applicable due to fixed_codelen.
881 } else {
882 accumulate = (a == tmp);
883 }
884 work = tmp;
885
886 int disp12 = split_largeoffset(si20, work, fixed_codelen, accumulate);
887 if (disp12 < 0) {
888 z_ld(t, si20, work);
889 } else {
890 if (accumulate) {
891 z_ld(t, disp12, work);
892 } else {
893 z_ld(t, disp12, work, a);
894 }
895 }
896 }
897 }
898
899 // PCrelative TOC access.
900 // Returns distance (in bytes) from current position to start of consts section.
901 // Returns 0 (zero) if no consts section exists or if it has size zero.
902 long MacroAssembler::toc_distance() {
903 CodeSection* cs = code()->consts();
904 return (long)((cs != NULL) ? cs->start()-pc() : 0);
905 }
906
907 // Implementation on x86/sparc assumes that constant and instruction section are
908 // adjacent, but this doesn't hold. Two special situations may occur, that we must
909 // be able to handle:
910 // 1. const section may be located apart from the inst section.
911 // 2. const section may be empty
912 // In both cases, we use the const section's start address to compute the "TOC",
913 // this seems to occur only temporarily; in the final step we always seem to end up
914 // with the pc-relatice variant.
915 //
916 // PC-relative offset could be +/-2**32 -> use long for disp
917 // Furthermore: makes no sense to have special code for
918 // adjacent const and inst sections.
919 void MacroAssembler::load_toc(Register Rtoc) {
920 // Simply use distance from start of const section (should be patched in the end).
921 long disp = toc_distance();
922
923 RelocationHolder rspec = internal_word_Relocation::spec(pc() + disp);
924 relocate(rspec);
925 z_larl(Rtoc, RelAddr::pcrel_off32(disp)); // Offset is in halfwords.
926 }
927
928 // PCrelative TOC access.
929 // Load from anywhere pcrelative (with relocation of load instr)
930 void MacroAssembler::load_long_pcrelative(Register Rdst, address dataLocation) {
931 address pc = this->pc();
932 ptrdiff_t total_distance = dataLocation - pc;
933 RelocationHolder rspec = internal_word_Relocation::spec(dataLocation);
934
935 assert((total_distance & 0x01L) == 0, "halfword alignment is mandatory");
936 assert(total_distance != 0, "sanity");
937
938 // Some extra safety net.
939 if (!RelAddr::is_in_range_of_RelAddr32(total_distance)) {
940 guarantee(RelAddr::is_in_range_of_RelAddr32(total_distance), "load_long_pcrelative can't handle distance " INTPTR_FORMAT, total_distance);
941 }
942
943 (this)->relocate(rspec, relocInfo::pcrel_addr_format);
944 z_lgrl(Rdst, RelAddr::pcrel_off32(total_distance));
945 }
946
947
948 // PCrelative TOC access.
949 // Load from anywhere pcrelative (with relocation of load instr)
950 // loaded addr has to be relocated when added to constant pool.
951 void MacroAssembler::load_addr_pcrelative(Register Rdst, address addrLocation) {
952 address pc = this->pc();
953 ptrdiff_t total_distance = addrLocation - pc;
954 RelocationHolder rspec = internal_word_Relocation::spec(addrLocation);
955
956 assert((total_distance & 0x01L) == 0, "halfword alignment is mandatory");
957
958 // Some extra safety net.
959 if (!RelAddr::is_in_range_of_RelAddr32(total_distance)) {
960 guarantee(RelAddr::is_in_range_of_RelAddr32(total_distance), "load_long_pcrelative can't handle distance " INTPTR_FORMAT, total_distance);
961 }
962
963 (this)->relocate(rspec, relocInfo::pcrel_addr_format);
964 z_lgrl(Rdst, RelAddr::pcrel_off32(total_distance));
965 }
966
967 // Generic operation: load a value from memory and test.
968 // CondCode indicates the sign (<0, ==0, >0) of the loaded value.
969 void MacroAssembler::load_and_test_byte(Register dst, const Address &a) {
970 z_lb(dst, a);
971 z_ltr(dst, dst);
972 }
973
974 void MacroAssembler::load_and_test_short(Register dst, const Address &a) {
975 int64_t disp = a.disp20();
976 if (Displacement::is_shortDisp(disp)) {
977 z_lh(dst, a);
978 } else if (Displacement::is_longDisp(disp)) {
979 z_lhy(dst, a);
980 } else {
981 guarantee(false, "displacement out of range");
982 }
983 z_ltr(dst, dst);
984 }
985
986 void MacroAssembler::load_and_test_int(Register dst, const Address &a) {
987 z_lt(dst, a);
988 }
989
990 void MacroAssembler::load_and_test_int2long(Register dst, const Address &a) {
991 z_ltgf(dst, a);
992 }
993
994 void MacroAssembler::load_and_test_long(Register dst, const Address &a) {
995 z_ltg(dst, a);
996 }
997
998 // Test a bit in memory.
999 void MacroAssembler::testbit(const Address &a, unsigned int bit) {
1000 assert(a.index() == noreg, "no index reg allowed in testbit");
1001 if (bit <= 7) {
1002 z_tm(a.disp() + 3, a.base(), 1 << bit);
1003 } else if (bit <= 15) {
1004 z_tm(a.disp() + 2, a.base(), 1 << (bit - 8));
1005 } else if (bit <= 23) {
1006 z_tm(a.disp() + 1, a.base(), 1 << (bit - 16));
1007 } else if (bit <= 31) {
1008 z_tm(a.disp() + 0, a.base(), 1 << (bit - 24));
1009 } else {
1010 ShouldNotReachHere();
1011 }
1012 }
1013
1014 // Test a bit in a register. Result is reflected in CC.
1015 void MacroAssembler::testbit(Register r, unsigned int bitPos) {
1016 if (bitPos < 16) {
1017 z_tmll(r, 1U<<bitPos);
1018 } else if (bitPos < 32) {
1019 z_tmlh(r, 1U<<(bitPos-16));
1020 } else if (bitPos < 48) {
1021 z_tmhl(r, 1U<<(bitPos-32));
1022 } else if (bitPos < 64) {
1023 z_tmhh(r, 1U<<(bitPos-48));
1024 } else {
1025 ShouldNotReachHere();
1026 }
1027 }
1028
1029 void MacroAssembler::prefetch_read(Address a) {
1030 z_pfd(1, a.disp20(), a.indexOrR0(), a.base());
1031 }
1032 void MacroAssembler::prefetch_update(Address a) {
1033 z_pfd(2, a.disp20(), a.indexOrR0(), a.base());
1034 }
1035
1036 // Clear a register, i.e. load const zero into reg.
1037 // Return len (in bytes) of generated instruction(s).
1038 // whole_reg: Clear 64 bits if true, 32 bits otherwise.
1039 // set_cc: Use instruction that sets the condition code, if true.
1040 int MacroAssembler::clear_reg(Register r, bool whole_reg, bool set_cc) {
1041 unsigned int start_off = offset();
1042 if (whole_reg) {
1043 set_cc ? z_xgr(r, r) : z_laz(r, 0, Z_R0);
1044 } else { // Only 32bit register.
1045 set_cc ? z_xr(r, r) : z_lhi(r, 0);
1046 }
1047 return offset() - start_off;
1048 }
1049
1050 #ifdef ASSERT
1051 int MacroAssembler::preset_reg(Register r, unsigned long pattern, int pattern_len) {
1052 switch (pattern_len) {
1053 case 1:
1054 pattern = (pattern & 0x000000ff) | ((pattern & 0x000000ff)<<8);
1055 case 2:
1056 pattern = (pattern & 0x0000ffff) | ((pattern & 0x0000ffff)<<16);
1057 case 4:
1058 pattern = (pattern & 0xffffffffL) | ((pattern & 0xffffffffL)<<32);
1059 case 8:
1060 return load_const_optimized_rtn_len(r, pattern, true);
1061 break;
1062 default:
1063 guarantee(false, "preset_reg: bad len");
1064 }
1065 return 0;
1066 }
1067 #endif
1068
1069 // addr: Address descriptor of memory to clear index register will not be used !
1070 // size: Number of bytes to clear.
1071 // !!! DO NOT USE THEM FOR ATOMIC MEMORY CLEARING !!!
1072 // !!! Use store_const() instead !!!
1073 void MacroAssembler::clear_mem(const Address& addr, unsigned size) {
1074 guarantee(size <= 256, "MacroAssembler::clear_mem: size too large");
1075
1076 if (size == 1) {
1077 z_mvi(addr, 0);
1078 return;
1079 }
1080
1081 switch (size) {
1082 case 2: z_mvhhi(addr, 0);
1083 return;
1084 case 4: z_mvhi(addr, 0);
1085 return;
1086 case 8: z_mvghi(addr, 0);
1087 return;
1088 default: ; // Fallthru to xc.
1089 }
1090
1091 z_xc(addr, size, addr);
1092 }
1093
1094 void MacroAssembler::align(int modulus) {
1095 while (offset() % modulus != 0) z_nop();
1096 }
1097
1098 // Special version for non-relocateable code if required alignment
1099 // is larger than CodeEntryAlignment.
1100 void MacroAssembler::align_address(int modulus) {
1101 while ((uintptr_t)pc() % modulus != 0) z_nop();
1102 }
1103
1104 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
1105 Register temp_reg,
1106 int64_t extra_slot_offset) {
1107 // On Z, we can have index and disp in an Address. So don't call argument_offset,
1108 // which issues an unnecessary add instruction.
1109 int stackElementSize = Interpreter::stackElementSize;
1110 int64_t offset = extra_slot_offset * stackElementSize;
1111 const Register argbase = Z_esp;
1112 if (arg_slot.is_constant()) {
1113 offset += arg_slot.as_constant() * stackElementSize;
1114 return Address(argbase, offset);
1115 }
1116 // else
1117 assert(temp_reg != noreg, "must specify");
1118 assert(temp_reg != Z_ARG1, "base and index are conflicting");
1119 z_sllg(temp_reg, arg_slot.as_register(), exact_log2(stackElementSize)); // tempreg = arg_slot << 3
1120 return Address(argbase, temp_reg, offset);
1121 }
1122
1123
1124 //===================================================================
1125 //=== START C O N S T A N T S I N C O D E S T R E A M ===
1126 //===================================================================
1127 //=== P A T CH A B L E C O N S T A N T S ===
1128 //===================================================================
1129
1130
1131 //---------------------------------------------------
1132 // Load (patchable) constant into register
1133 //---------------------------------------------------
1134
1135
1136 // Load absolute address (and try to optimize).
1137 // Note: This method is usable only for position-fixed code,
1138 // referring to a position-fixed target location.
1139 // If not so, relocations and patching must be used.
1140 void MacroAssembler::load_absolute_address(Register d, address addr) {
1141 assert(addr != NULL, "should not happen");
1142 BLOCK_COMMENT("load_absolute_address:");
1143 if (addr == NULL) {
1144 z_larl(d, pc()); // Dummy emit for size calc.
1145 return;
1146 }
1147
1148 if (RelAddr::is_in_range_of_RelAddr32(addr, pc())) {
1149 z_larl(d, addr);
1150 return;
1151 }
1152
1153 load_const_optimized(d, (long)addr);
1154 }
1155
1156 // Load a 64bit constant.
1157 // Patchable code sequence, but not atomically patchable.
1158 // Make sure to keep code size constant -> no value-dependent optimizations.
1159 // Do not kill condition code.
1160 void MacroAssembler::load_const(Register t, long x) {
1161 Assembler::z_iihf(t, (int)(x >> 32));
1162 Assembler::z_iilf(t, (int)(x & 0xffffffff));
1163 }
1164
1165 // Load a 32bit constant into a 64bit register, sign-extend or zero-extend.
1166 // Patchable code sequence, but not atomically patchable.
1167 // Make sure to keep code size constant -> no value-dependent optimizations.
1168 // Do not kill condition code.
1169 void MacroAssembler::load_const_32to64(Register t, int64_t x, bool sign_extend) {
1170 if (sign_extend) { Assembler::z_lgfi(t, x); }
1171 else { Assembler::z_llilf(t, x); }
1172 }
1173
1174 // Load narrow oop constant, no decompression.
1175 void MacroAssembler::load_narrow_oop(Register t, narrowOop a) {
1176 assert(UseCompressedOops, "must be on to call this method");
1177 load_const_32to64(t, a, false /*sign_extend*/);
1178 }
1179
1180 // Load narrow klass constant, compression required.
1181 void MacroAssembler::load_narrow_klass(Register t, Klass* k) {
1182 assert(UseCompressedClassPointers, "must be on to call this method");
1183 narrowKlass encoded_k = Klass::encode_klass(k);
1184 load_const_32to64(t, encoded_k, false /*sign_extend*/);
1185 }
1186
1187 //------------------------------------------------------
1188 // Compare (patchable) constant with register.
1189 //------------------------------------------------------
1190
1191 // Compare narrow oop in reg with narrow oop constant, no decompression.
1192 void MacroAssembler::compare_immediate_narrow_oop(Register oop1, narrowOop oop2) {
1193 assert(UseCompressedOops, "must be on to call this method");
1194
1195 Assembler::z_clfi(oop1, oop2);
1196 }
1197
1198 // Compare narrow oop in reg with narrow oop constant, no decompression.
1199 void MacroAssembler::compare_immediate_narrow_klass(Register klass1, Klass* klass2) {
1200 assert(UseCompressedClassPointers, "must be on to call this method");
1201 narrowKlass encoded_k = Klass::encode_klass(klass2);
1202
1203 Assembler::z_clfi(klass1, encoded_k);
1204 }
1205
1206 //----------------------------------------------------------
1207 // Check which kind of load_constant we have here.
1208 //----------------------------------------------------------
1209
1210 // Detection of CPU version dependent load_const sequence.
1211 // The detection is valid only for code sequences generated by load_const,
1212 // not load_const_optimized.
1213 bool MacroAssembler::is_load_const(address a) {
1214 unsigned long inst1, inst2;
1215 unsigned int len1, len2;
1216
1217 len1 = get_instruction(a, &inst1);
1218 len2 = get_instruction(a + len1, &inst2);
1219
1220 return is_z_iihf(inst1) && is_z_iilf(inst2);
1221 }
1222
1223 // Detection of CPU version dependent load_const_32to64 sequence.
1224 // Mostly used for narrow oops and narrow Klass pointers.
1225 // The detection is valid only for code sequences generated by load_const_32to64.
1226 bool MacroAssembler::is_load_const_32to64(address pos) {
1227 unsigned long inst1, inst2;
1228 unsigned int len1;
1229
1230 len1 = get_instruction(pos, &inst1);
1231 return is_z_llilf(inst1);
1232 }
1233
1234 // Detection of compare_immediate_narrow sequence.
1235 // The detection is valid only for code sequences generated by compare_immediate_narrow_oop.
1236 bool MacroAssembler::is_compare_immediate32(address pos) {
1237 return is_equal(pos, CLFI_ZOPC, RIL_MASK);
1238 }
1239
1240 // Detection of compare_immediate_narrow sequence.
1241 // The detection is valid only for code sequences generated by compare_immediate_narrow_oop.
1242 bool MacroAssembler::is_compare_immediate_narrow_oop(address pos) {
1243 return is_compare_immediate32(pos);
1244 }
1245
1246 // Detection of compare_immediate_narrow sequence.
1247 // The detection is valid only for code sequences generated by compare_immediate_narrow_klass.
1248 bool MacroAssembler::is_compare_immediate_narrow_klass(address pos) {
1249 return is_compare_immediate32(pos);
1250 }
1251
1252 //-----------------------------------
1253 // patch the load_constant
1254 //-----------------------------------
1255
1256 // CPU-version dependend patching of load_const.
1257 void MacroAssembler::patch_const(address a, long x) {
1258 assert(is_load_const(a), "not a load of a constant");
1259 set_imm32((address)a, (int) ((x >> 32) & 0xffffffff));
1260 set_imm32((address)(a + 6), (int)(x & 0xffffffff));
1261 }
1262
1263 // Patching the value of CPU version dependent load_const_32to64 sequence.
1264 // The passed ptr MUST be in compressed format!
1265 int MacroAssembler::patch_load_const_32to64(address pos, int64_t np) {
1266 assert(is_load_const_32to64(pos), "not a load of a narrow ptr (oop or klass)");
1267
1268 set_imm32(pos, np);
1269 return 6;
1270 }
1271
1272 // Patching the value of CPU version dependent compare_immediate_narrow sequence.
1273 // The passed ptr MUST be in compressed format!
1274 int MacroAssembler::patch_compare_immediate_32(address pos, int64_t np) {
1275 assert(is_compare_immediate32(pos), "not a compressed ptr compare");
1276
1277 set_imm32(pos, np);
1278 return 6;
1279 }
1280
1281 // Patching the immediate value of CPU version dependent load_narrow_oop sequence.
1282 // The passed ptr must NOT be in compressed format!
1283 int MacroAssembler::patch_load_narrow_oop(address pos, oop o) {
1284 assert(UseCompressedOops, "Can only patch compressed oops");
1285
1286 narrowOop no = CompressedOops::encode(o);
1287 return patch_load_const_32to64(pos, no);
1288 }
1289
1290 // Patching the immediate value of CPU version dependent load_narrow_klass sequence.
1291 // The passed ptr must NOT be in compressed format!
1292 int MacroAssembler::patch_load_narrow_klass(address pos, Klass* k) {
1293 assert(UseCompressedClassPointers, "Can only patch compressed klass pointers");
1294
1295 narrowKlass nk = Klass::encode_klass(k);
1296 return patch_load_const_32to64(pos, nk);
1297 }
1298
1299 // Patching the immediate value of CPU version dependent compare_immediate_narrow_oop sequence.
1300 // The passed ptr must NOT be in compressed format!
1301 int MacroAssembler::patch_compare_immediate_narrow_oop(address pos, oop o) {
1302 assert(UseCompressedOops, "Can only patch compressed oops");
1303
1304 narrowOop no = CompressedOops::encode(o);
1305 return patch_compare_immediate_32(pos, no);
1306 }
1307
1308 // Patching the immediate value of CPU version dependent compare_immediate_narrow_klass sequence.
1309 // The passed ptr must NOT be in compressed format!
1310 int MacroAssembler::patch_compare_immediate_narrow_klass(address pos, Klass* k) {
1311 assert(UseCompressedClassPointers, "Can only patch compressed klass pointers");
1312
1313 narrowKlass nk = Klass::encode_klass(k);
1314 return patch_compare_immediate_32(pos, nk);
1315 }
1316
1317 //------------------------------------------------------------------------
1318 // Extract the constant from a load_constant instruction stream.
1319 //------------------------------------------------------------------------
1320
1321 // Get constant from a load_const sequence.
1322 long MacroAssembler::get_const(address a) {
1323 assert(is_load_const(a), "not a load of a constant");
1324 unsigned long x;
1325 x = (((unsigned long) (get_imm32(a,0) & 0xffffffff)) << 32);
1326 x |= (((unsigned long) (get_imm32(a,1) & 0xffffffff)));
1327 return (long) x;
1328 }
1329
1330 //--------------------------------------
1331 // Store a constant in memory.
1332 //--------------------------------------
1333
1334 // General emitter to move a constant to memory.
1335 // The store is atomic.
1336 // o Address must be given in RS format (no index register)
1337 // o Displacement should be 12bit unsigned for efficiency. 20bit signed also supported.
1338 // o Constant can be 1, 2, 4, or 8 bytes, signed or unsigned.
1339 // o Memory slot can be 1, 2, 4, or 8 bytes, signed or unsigned.
1340 // o Memory slot must be at least as wide as constant, will assert otherwise.
1341 // o Signed constants will sign-extend, unsigned constants will zero-extend to slot width.
1342 int MacroAssembler::store_const(const Address &dest, long imm,
1343 unsigned int lm, unsigned int lc,
1344 Register scratch) {
1345 int64_t disp = dest.disp();
1346 Register base = dest.base();
1347 assert(!dest.has_index(), "not supported");
1348 assert((lm==1)||(lm==2)||(lm==4)||(lm==8), "memory length not supported");
1349 assert((lc==1)||(lc==2)||(lc==4)||(lc==8), "constant length not supported");
1350 assert(lm>=lc, "memory slot too small");
1351 assert(lc==8 || Immediate::is_simm(imm, lc*8), "const out of range");
1352 assert(Displacement::is_validDisp(disp), "displacement out of range");
1353
1354 bool is_shortDisp = Displacement::is_shortDisp(disp);
1355 int store_offset = -1;
1356
1357 // For target len == 1 it's easy.
1358 if (lm == 1) {
1359 store_offset = offset();
1360 if (is_shortDisp) {
1361 z_mvi(disp, base, imm);
1362 return store_offset;
1363 } else {
1364 z_mviy(disp, base, imm);
1365 return store_offset;
1366 }
1367 }
1368
1369 // All the "good stuff" takes an unsigned displacement.
1370 if (is_shortDisp) {
1371 // NOTE: Cannot use clear_mem for imm==0, because it is not atomic.
1372
1373 store_offset = offset();
1374 switch (lm) {
1375 case 2: // Lc == 1 handled correctly here, even for unsigned. Instruction does no widening.
1376 z_mvhhi(disp, base, imm);
1377 return store_offset;
1378 case 4:
1379 if (Immediate::is_simm16(imm)) {
1380 z_mvhi(disp, base, imm);
1381 return store_offset;
1382 }
1383 break;
1384 case 8:
1385 if (Immediate::is_simm16(imm)) {
1386 z_mvghi(disp, base, imm);
1387 return store_offset;
1388 }
1389 break;
1390 default:
1391 ShouldNotReachHere();
1392 break;
1393 }
1394 }
1395
1396 // Can't optimize, so load value and store it.
1397 guarantee(scratch != noreg, " need a scratch register here !");
1398 if (imm != 0) {
1399 load_const_optimized(scratch, imm); // Preserves CC anyway.
1400 } else {
1401 // Leave CC alone!!
1402 (void) clear_reg(scratch, true, false); // Indicate unused result.
1403 }
1404
1405 store_offset = offset();
1406 if (is_shortDisp) {
1407 switch (lm) {
1408 case 2:
1409 z_sth(scratch, disp, Z_R0, base);
1410 return store_offset;
1411 case 4:
1412 z_st(scratch, disp, Z_R0, base);
1413 return store_offset;
1414 case 8:
1415 z_stg(scratch, disp, Z_R0, base);
1416 return store_offset;
1417 default:
1418 ShouldNotReachHere();
1419 break;
1420 }
1421 } else {
1422 switch (lm) {
1423 case 2:
1424 z_sthy(scratch, disp, Z_R0, base);
1425 return store_offset;
1426 case 4:
1427 z_sty(scratch, disp, Z_R0, base);
1428 return store_offset;
1429 case 8:
1430 z_stg(scratch, disp, Z_R0, base);
1431 return store_offset;
1432 default:
1433 ShouldNotReachHere();
1434 break;
1435 }
1436 }
1437 return -1; // should not reach here
1438 }
1439
1440 //===================================================================
1441 //=== N O T P A T CH A B L E C O N S T A N T S ===
1442 //===================================================================
1443
1444 // Load constant x into register t with a fast instrcution sequence
1445 // depending on the bits in x. Preserves CC under all circumstances.
1446 int MacroAssembler::load_const_optimized_rtn_len(Register t, long x, bool emit) {
1447 if (x == 0) {
1448 int len;
1449 if (emit) {
1450 len = clear_reg(t, true, false);
1451 } else {
1452 len = 4;
1453 }
1454 return len;
1455 }
1456
1457 if (Immediate::is_simm16(x)) {
1458 if (emit) { z_lghi(t, x); }
1459 return 4;
1460 }
1461
1462 // 64 bit value: | part1 | part2 | part3 | part4 |
1463 // At least one part is not zero!
1464 int part1 = ((x >> 32) & 0xffff0000) >> 16;
1465 int part2 = (x >> 32) & 0x0000ffff;
1466 int part3 = (x & 0xffff0000) >> 16;
1467 int part4 = (x & 0x0000ffff);
1468
1469 // Lower word only (unsigned).
1470 if ((part1 == 0) && (part2 == 0)) {
1471 if (part3 == 0) {
1472 if (emit) z_llill(t, part4);
1473 return 4;
1474 }
1475 if (part4 == 0) {
1476 if (emit) z_llilh(t, part3);
1477 return 4;
1478 }
1479 if (emit) z_llilf(t, (int)(x & 0xffffffff));
1480 return 6;
1481 }
1482
1483 // Upper word only.
1484 if ((part3 == 0) && (part4 == 0)) {
1485 if (part1 == 0) {
1486 if (emit) z_llihl(t, part2);
1487 return 4;
1488 }
1489 if (part2 == 0) {
1490 if (emit) z_llihh(t, part1);
1491 return 4;
1492 }
1493 if (emit) z_llihf(t, (int)(x >> 32));
1494 return 6;
1495 }
1496
1497 // Lower word only (signed).
1498 if ((part1 == 0x0000ffff) && (part2 == 0x0000ffff) && ((part3 & 0x00008000) != 0)) {
1499 if (emit) z_lgfi(t, (int)(x & 0xffffffff));
1500 return 6;
1501 }
1502
1503 int len = 0;
1504
1505 if ((part1 == 0) || (part2 == 0)) {
1506 if (part1 == 0) {
1507 if (emit) z_llihl(t, part2);
1508 len += 4;
1509 } else {
1510 if (emit) z_llihh(t, part1);
1511 len += 4;
1512 }
1513 } else {
1514 if (emit) z_llihf(t, (int)(x >> 32));
1515 len += 6;
1516 }
1517
1518 if ((part3 == 0) || (part4 == 0)) {
1519 if (part3 == 0) {
1520 if (emit) z_iill(t, part4);
1521 len += 4;
1522 } else {
1523 if (emit) z_iilh(t, part3);
1524 len += 4;
1525 }
1526 } else {
1527 if (emit) z_iilf(t, (int)(x & 0xffffffff));
1528 len += 6;
1529 }
1530 return len;
1531 }
1532
1533 //=====================================================================
1534 //=== H I G H E R L E V E L B R A N C H E M I T T E R S ===
1535 //=====================================================================
1536
1537 // Note: In the worst case, one of the scratch registers is destroyed!!!
1538 void MacroAssembler::compare32_and_branch(Register r1, RegisterOrConstant x2, branch_condition cond, Label& lbl) {
1539 // Right operand is constant.
1540 if (x2.is_constant()) {
1541 jlong value = x2.as_constant();
1542 compare_and_branch_optimized(r1, value, cond, lbl, /*len64=*/false, /*has_sign=*/true);
1543 return;
1544 }
1545
1546 // Right operand is in register.
1547 compare_and_branch_optimized(r1, x2.as_register(), cond, lbl, /*len64=*/false, /*has_sign=*/true);
1548 }
1549
1550 // Note: In the worst case, one of the scratch registers is destroyed!!!
1551 void MacroAssembler::compareU32_and_branch(Register r1, RegisterOrConstant x2, branch_condition cond, Label& lbl) {
1552 // Right operand is constant.
1553 if (x2.is_constant()) {
1554 jlong value = x2.as_constant();
1555 compare_and_branch_optimized(r1, value, cond, lbl, /*len64=*/false, /*has_sign=*/false);
1556 return;
1557 }
1558
1559 // Right operand is in register.
1560 compare_and_branch_optimized(r1, x2.as_register(), cond, lbl, /*len64=*/false, /*has_sign=*/false);
1561 }
1562
1563 // Note: In the worst case, one of the scratch registers is destroyed!!!
1564 void MacroAssembler::compare64_and_branch(Register r1, RegisterOrConstant x2, branch_condition cond, Label& lbl) {
1565 // Right operand is constant.
1566 if (x2.is_constant()) {
1567 jlong value = x2.as_constant();
1568 compare_and_branch_optimized(r1, value, cond, lbl, /*len64=*/true, /*has_sign=*/true);
1569 return;
1570 }
1571
1572 // Right operand is in register.
1573 compare_and_branch_optimized(r1, x2.as_register(), cond, lbl, /*len64=*/true, /*has_sign=*/true);
1574 }
1575
1576 void MacroAssembler::compareU64_and_branch(Register r1, RegisterOrConstant x2, branch_condition cond, Label& lbl) {
1577 // Right operand is constant.
1578 if (x2.is_constant()) {
1579 jlong value = x2.as_constant();
1580 compare_and_branch_optimized(r1, value, cond, lbl, /*len64=*/true, /*has_sign=*/false);
1581 return;
1582 }
1583
1584 // Right operand is in register.
1585 compare_and_branch_optimized(r1, x2.as_register(), cond, lbl, /*len64=*/true, /*has_sign=*/false);
1586 }
1587
1588 // Generate an optimal branch to the branch target.
1589 // Optimal means that a relative branch (brc or brcl) is used if the
1590 // branch distance is short enough. Loading the target address into a
1591 // register and branching via reg is used as fallback only.
1592 //
1593 // Used registers:
1594 // Z_R1 - work reg. Holds branch target address.
1595 // Used in fallback case only.
1596 //
1597 // This version of branch_optimized is good for cases where the target address is known
1598 // and constant, i.e. is never changed (no relocation, no patching).
1599 void MacroAssembler::branch_optimized(Assembler::branch_condition cond, address branch_addr) {
1600 address branch_origin = pc();
1601
1602 if (RelAddr::is_in_range_of_RelAddr16(branch_addr, branch_origin)) {
1603 z_brc(cond, branch_addr);
1604 } else if (RelAddr::is_in_range_of_RelAddr32(branch_addr, branch_origin)) {
1605 z_brcl(cond, branch_addr);
1606 } else {
1607 load_const_optimized(Z_R1, branch_addr); // CC must not get killed by load_const_optimized.
1608 z_bcr(cond, Z_R1);
1609 }
1610 }
1611
1612 // This version of branch_optimized is good for cases where the target address
1613 // is potentially not yet known at the time the code is emitted.
1614 //
1615 // One very common case is a branch to an unbound label which is handled here.
1616 // The caller might know (or hope) that the branch distance is short enough
1617 // to be encoded in a 16bit relative address. In this case he will pass a
1618 // NearLabel branch_target.
1619 // Care must be taken with unbound labels. Each call to target(label) creates
1620 // an entry in the patch queue for that label to patch all references of the label
1621 // once it gets bound. Those recorded patch locations must be patchable. Otherwise,
1622 // an assertion fires at patch time.
1623 void MacroAssembler::branch_optimized(Assembler::branch_condition cond, Label& branch_target) {
1624 if (branch_target.is_bound()) {
1625 address branch_addr = target(branch_target);
1626 branch_optimized(cond, branch_addr);
1627 } else if (branch_target.is_near()) {
1628 z_brc(cond, branch_target); // Caller assures that the target will be in range for z_brc.
1629 } else {
1630 z_brcl(cond, branch_target); // Let's hope target is in range. Otherwise, we will abort at patch time.
1631 }
1632 }
1633
1634 // Generate an optimal compare and branch to the branch target.
1635 // Optimal means that a relative branch (clgrj, brc or brcl) is used if the
1636 // branch distance is short enough. Loading the target address into a
1637 // register and branching via reg is used as fallback only.
1638 //
1639 // Input:
1640 // r1 - left compare operand
1641 // r2 - right compare operand
1642 void MacroAssembler::compare_and_branch_optimized(Register r1,
1643 Register r2,
1644 Assembler::branch_condition cond,
1645 address branch_addr,
1646 bool len64,
1647 bool has_sign) {
1648 unsigned int casenum = (len64?2:0)+(has_sign?0:1);
1649
1650 address branch_origin = pc();
1651 if (VM_Version::has_CompareBranch() && RelAddr::is_in_range_of_RelAddr16(branch_addr, branch_origin)) {
1652 switch (casenum) {
1653 case 0: z_crj( r1, r2, cond, branch_addr); break;
1654 case 1: z_clrj (r1, r2, cond, branch_addr); break;
1655 case 2: z_cgrj(r1, r2, cond, branch_addr); break;
1656 case 3: z_clgrj(r1, r2, cond, branch_addr); break;
1657 default: ShouldNotReachHere(); break;
1658 }
1659 } else {
1660 switch (casenum) {
1661 case 0: z_cr( r1, r2); break;
1662 case 1: z_clr(r1, r2); break;
1663 case 2: z_cgr(r1, r2); break;
1664 case 3: z_clgr(r1, r2); break;
1665 default: ShouldNotReachHere(); break;
1666 }
1667 branch_optimized(cond, branch_addr);
1668 }
1669 }
1670
1671 // Generate an optimal compare and branch to the branch target.
1672 // Optimal means that a relative branch (clgij, brc or brcl) is used if the
1673 // branch distance is short enough. Loading the target address into a
1674 // register and branching via reg is used as fallback only.
1675 //
1676 // Input:
1677 // r1 - left compare operand (in register)
1678 // x2 - right compare operand (immediate)
1679 void MacroAssembler::compare_and_branch_optimized(Register r1,
1680 jlong x2,
1681 Assembler::branch_condition cond,
1682 Label& branch_target,
1683 bool len64,
1684 bool has_sign) {
1685 address branch_origin = pc();
1686 bool x2_imm8 = (has_sign && Immediate::is_simm8(x2)) || (!has_sign && Immediate::is_uimm8(x2));
1687 bool is_RelAddr16 = branch_target.is_near() ||
1688 (branch_target.is_bound() &&
1689 RelAddr::is_in_range_of_RelAddr16(target(branch_target), branch_origin));
1690 unsigned int casenum = (len64?2:0)+(has_sign?0:1);
1691
1692 if (VM_Version::has_CompareBranch() && is_RelAddr16 && x2_imm8) {
1693 switch (casenum) {
1694 case 0: z_cij( r1, x2, cond, branch_target); break;
1695 case 1: z_clij(r1, x2, cond, branch_target); break;
1696 case 2: z_cgij(r1, x2, cond, branch_target); break;
1697 case 3: z_clgij(r1, x2, cond, branch_target); break;
1698 default: ShouldNotReachHere(); break;
1699 }
1700 return;
1701 }
1702
1703 if (x2 == 0) {
1704 switch (casenum) {
1705 case 0: z_ltr(r1, r1); break;
1706 case 1: z_ltr(r1, r1); break; // Caution: unsigned test only provides zero/notZero indication!
1707 case 2: z_ltgr(r1, r1); break;
1708 case 3: z_ltgr(r1, r1); break; // Caution: unsigned test only provides zero/notZero indication!
1709 default: ShouldNotReachHere(); break;
1710 }
1711 } else {
1712 if ((has_sign && Immediate::is_simm16(x2)) || (!has_sign && Immediate::is_uimm(x2, 15))) {
1713 switch (casenum) {
1714 case 0: z_chi(r1, x2); break;
1715 case 1: z_chi(r1, x2); break; // positive immediate < 2**15
1716 case 2: z_cghi(r1, x2); break;
1717 case 3: z_cghi(r1, x2); break; // positive immediate < 2**15
1718 default: break;
1719 }
1720 } else if ( (has_sign && Immediate::is_simm32(x2)) || (!has_sign && Immediate::is_uimm32(x2)) ) {
1721 switch (casenum) {
1722 case 0: z_cfi( r1, x2); break;
1723 case 1: z_clfi(r1, x2); break;
1724 case 2: z_cgfi(r1, x2); break;
1725 case 3: z_clgfi(r1, x2); break;
1726 default: ShouldNotReachHere(); break;
1727 }
1728 } else {
1729 // No instruction with immediate operand possible, so load into register.
1730 Register scratch = (r1 != Z_R0) ? Z_R0 : Z_R1;
1731 load_const_optimized(scratch, x2);
1732 switch (casenum) {
1733 case 0: z_cr( r1, scratch); break;
1734 case 1: z_clr(r1, scratch); break;
1735 case 2: z_cgr(r1, scratch); break;
1736 case 3: z_clgr(r1, scratch); break;
1737 default: ShouldNotReachHere(); break;
1738 }
1739 }
1740 }
1741 branch_optimized(cond, branch_target);
1742 }
1743
1744 // Generate an optimal compare and branch to the branch target.
1745 // Optimal means that a relative branch (clgrj, brc or brcl) is used if the
1746 // branch distance is short enough. Loading the target address into a
1747 // register and branching via reg is used as fallback only.
1748 //
1749 // Input:
1750 // r1 - left compare operand
1751 // r2 - right compare operand
1752 void MacroAssembler::compare_and_branch_optimized(Register r1,
1753 Register r2,
1754 Assembler::branch_condition cond,
1755 Label& branch_target,
1756 bool len64,
1757 bool has_sign) {
1758 unsigned int casenum = (len64 ? 2 : 0) + (has_sign ? 0 : 1);
1759
1760 if (branch_target.is_bound()) {
1761 address branch_addr = target(branch_target);
1762 compare_and_branch_optimized(r1, r2, cond, branch_addr, len64, has_sign);
1763 } else {
1764 if (VM_Version::has_CompareBranch() && branch_target.is_near()) {
1765 switch (casenum) {
1766 case 0: z_crj( r1, r2, cond, branch_target); break;
1767 case 1: z_clrj( r1, r2, cond, branch_target); break;
1768 case 2: z_cgrj( r1, r2, cond, branch_target); break;
1769 case 3: z_clgrj(r1, r2, cond, branch_target); break;
1770 default: ShouldNotReachHere(); break;
1771 }
1772 } else {
1773 switch (casenum) {
1774 case 0: z_cr( r1, r2); break;
1775 case 1: z_clr(r1, r2); break;
1776 case 2: z_cgr(r1, r2); break;
1777 case 3: z_clgr(r1, r2); break;
1778 default: ShouldNotReachHere(); break;
1779 }
1780 branch_optimized(cond, branch_target);
1781 }
1782 }
1783 }
1784
1785 //===========================================================================
1786 //=== END H I G H E R L E V E L B R A N C H E M I T T E R S ===
1787 //===========================================================================
1788
1789 AddressLiteral MacroAssembler::allocate_metadata_address(Metadata* obj) {
1790 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1791 int index = oop_recorder()->allocate_metadata_index(obj);
1792 RelocationHolder rspec = metadata_Relocation::spec(index);
1793 return AddressLiteral((address)obj, rspec);
1794 }
1795
1796 AddressLiteral MacroAssembler::constant_metadata_address(Metadata* obj) {
1797 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1798 int index = oop_recorder()->find_index(obj);
1799 RelocationHolder rspec = metadata_Relocation::spec(index);
1800 return AddressLiteral((address)obj, rspec);
1801 }
1802
1803 AddressLiteral MacroAssembler::allocate_oop_address(jobject obj) {
1804 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1805 int oop_index = oop_recorder()->allocate_oop_index(obj);
1806 return AddressLiteral(address(obj), oop_Relocation::spec(oop_index));
1807 }
1808
1809 AddressLiteral MacroAssembler::constant_oop_address(jobject obj) {
1810 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
1811 int oop_index = oop_recorder()->find_index(obj);
1812 return AddressLiteral(address(obj), oop_Relocation::spec(oop_index));
1813 }
1814
1815 // NOTE: destroys r
1816 void MacroAssembler::c2bool(Register r, Register t) {
1817 z_lcr(t, r); // t = -r
1818 z_or(r, t); // r = -r OR r
1819 z_srl(r, 31); // Yields 0 if r was 0, 1 otherwise.
1820 }
1821
1822 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
1823 Register tmp,
1824 int offset) {
1825 intptr_t value = *delayed_value_addr;
1826 if (value != 0) {
1827 return RegisterOrConstant(value + offset);
1828 }
1829
1830 BLOCK_COMMENT("delayed_value {");
1831 // Load indirectly to solve generation ordering problem.
1832 load_absolute_address(tmp, (address) delayed_value_addr); // tmp = a;
1833 z_lg(tmp, 0, tmp); // tmp = *tmp;
1834
1835 #ifdef ASSERT
1836 NearLabel L;
1837 compare64_and_branch(tmp, (intptr_t)0L, Assembler::bcondNotEqual, L);
1838 z_illtrap();
1839 bind(L);
1840 #endif
1841
1842 if (offset != 0) {
1843 z_agfi(tmp, offset); // tmp = tmp + offset;
1844 }
1845
1846 BLOCK_COMMENT("} delayed_value");
1847 return RegisterOrConstant(tmp);
1848 }
1849
1850 // Patch instruction `inst' at offset `inst_pos' to refer to `dest_pos'
1851 // and return the resulting instruction.
1852 // Dest_pos and inst_pos are 32 bit only. These parms can only designate
1853 // relative positions.
1854 // Use correct argument types. Do not pre-calculate distance.
1855 unsigned long MacroAssembler::patched_branch(address dest_pos, unsigned long inst, address inst_pos) {
1856 int c = 0;
1857 unsigned long patched_inst = 0;
1858 if (is_call_pcrelative_short(inst) ||
1859 is_branch_pcrelative_short(inst) ||
1860 is_branchoncount_pcrelative_short(inst) ||
1861 is_branchonindex32_pcrelative_short(inst)) {
1862 c = 1;
1863 int m = fmask(15, 0); // simm16(-1, 16, 32);
1864 int v = simm16(RelAddr::pcrel_off16(dest_pos, inst_pos), 16, 32);
1865 patched_inst = (inst & ~m) | v;
1866 } else if (is_compareandbranch_pcrelative_short(inst)) {
1867 c = 2;
1868 long m = fmask(31, 16); // simm16(-1, 16, 48);
1869 long v = simm16(RelAddr::pcrel_off16(dest_pos, inst_pos), 16, 48);
1870 patched_inst = (inst & ~m) | v;
1871 } else if (is_branchonindex64_pcrelative_short(inst)) {
1872 c = 3;
1873 long m = fmask(31, 16); // simm16(-1, 16, 48);
1874 long v = simm16(RelAddr::pcrel_off16(dest_pos, inst_pos), 16, 48);
1875 patched_inst = (inst & ~m) | v;
1876 } else if (is_call_pcrelative_long(inst) || is_branch_pcrelative_long(inst)) {
1877 c = 4;
1878 long m = fmask(31, 0); // simm32(-1, 16, 48);
1879 long v = simm32(RelAddr::pcrel_off32(dest_pos, inst_pos), 16, 48);
1880 patched_inst = (inst & ~m) | v;
1881 } else if (is_pcrelative_long(inst)) { // These are the non-branch pc-relative instructions.
1882 c = 5;
1883 long m = fmask(31, 0); // simm32(-1, 16, 48);
1884 long v = simm32(RelAddr::pcrel_off32(dest_pos, inst_pos), 16, 48);
1885 patched_inst = (inst & ~m) | v;
1886 } else {
1887 print_dbg_msg(tty, inst, "not a relative branch", 0);
1888 dump_code_range(tty, inst_pos, 32, "not a pcrelative branch");
1889 ShouldNotReachHere();
1890 }
1891
1892 long new_off = get_pcrel_offset(patched_inst);
1893 if (new_off != (dest_pos-inst_pos)) {
1894 tty->print_cr("case %d: dest_pos = %p, inst_pos = %p, disp = %ld(%12.12lx)", c, dest_pos, inst_pos, new_off, new_off);
1895 print_dbg_msg(tty, inst, "<- original instruction: branch patching error", 0);
1896 print_dbg_msg(tty, patched_inst, "<- patched instruction: branch patching error", 0);
1897 #ifdef LUCY_DBG
1898 VM_Version::z_SIGSEGV();
1899 #endif
1900 ShouldNotReachHere();
1901 }
1902 return patched_inst;
1903 }
1904
1905 // Only called when binding labels (share/vm/asm/assembler.cpp)
1906 // Pass arguments as intended. Do not pre-calculate distance.
1907 void MacroAssembler::pd_patch_instruction(address branch, address target) {
1908 unsigned long stub_inst;
1909 int inst_len = get_instruction(branch, &stub_inst);
1910
1911 set_instruction(branch, patched_branch(target, stub_inst, branch), inst_len);
1912 }
1913
1914
1915 // Extract relative address (aka offset).
1916 // inv_simm16 works for 4-byte instructions only.
1917 // compare and branch instructions are 6-byte and have a 16bit offset "in the middle".
1918 long MacroAssembler::get_pcrel_offset(unsigned long inst) {
1919
1920 if (MacroAssembler::is_pcrelative_short(inst)) {
1921 if (((inst&0xFFFFffff00000000UL) == 0) && ((inst&0x00000000FFFF0000UL) != 0)) {
1922 return RelAddr::inv_pcrel_off16(inv_simm16(inst));
1923 } else {
1924 return RelAddr::inv_pcrel_off16(inv_simm16_48(inst));
1925 }
1926 }
1927
1928 if (MacroAssembler::is_pcrelative_long(inst)) {
1929 return RelAddr::inv_pcrel_off32(inv_simm32(inst));
1930 }
1931
1932 print_dbg_msg(tty, inst, "not a pcrelative instruction", 6);
1933 #ifdef LUCY_DBG
1934 VM_Version::z_SIGSEGV();
1935 #else
1936 ShouldNotReachHere();
1937 #endif
1938 return -1;
1939 }
1940
1941 long MacroAssembler::get_pcrel_offset(address pc) {
1942 unsigned long inst;
1943 unsigned int len = get_instruction(pc, &inst);
1944
1945 #ifdef ASSERT
1946 long offset;
1947 if (MacroAssembler::is_pcrelative_short(inst) || MacroAssembler::is_pcrelative_long(inst)) {
1948 offset = get_pcrel_offset(inst);
1949 } else {
1950 offset = -1;
1951 }
1952
1953 if (offset == -1) {
1954 dump_code_range(tty, pc, 32, "not a pcrelative instruction");
1955 #ifdef LUCY_DBG
1956 VM_Version::z_SIGSEGV();
1957 #else
1958 ShouldNotReachHere();
1959 #endif
1960 }
1961 return offset;
1962 #else
1963 return get_pcrel_offset(inst);
1964 #endif // ASSERT
1965 }
1966
1967 // Get target address from pc-relative instructions.
1968 address MacroAssembler::get_target_addr_pcrel(address pc) {
1969 assert(is_pcrelative_long(pc), "not a pcrelative instruction");
1970 return pc + get_pcrel_offset(pc);
1971 }
1972
1973 // Patch pc relative load address.
1974 void MacroAssembler::patch_target_addr_pcrel(address pc, address con) {
1975 unsigned long inst;
1976 // Offset is +/- 2**32 -> use long.
1977 ptrdiff_t distance = con - pc;
1978
1979 get_instruction(pc, &inst);
1980
1981 if (is_pcrelative_short(inst)) {
1982 *(short *)(pc+2) = RelAddr::pcrel_off16(con, pc); // Instructions are at least 2-byte aligned, no test required.
1983
1984 // Some extra safety net.
1985 if (!RelAddr::is_in_range_of_RelAddr16(distance)) {
1986 print_dbg_msg(tty, inst, "distance out of range (16bit)", 4);
1987 dump_code_range(tty, pc, 32, "distance out of range (16bit)");
1988 guarantee(RelAddr::is_in_range_of_RelAddr16(distance), "too far away (more than +/- 2**16");
1989 }
1990 return;
1991 }
1992
1993 if (is_pcrelative_long(inst)) {
1994 *(int *)(pc+2) = RelAddr::pcrel_off32(con, pc);
1995
1996 // Some Extra safety net.
1997 if (!RelAddr::is_in_range_of_RelAddr32(distance)) {
1998 print_dbg_msg(tty, inst, "distance out of range (32bit)", 6);
1999 dump_code_range(tty, pc, 32, "distance out of range (32bit)");
2000 guarantee(RelAddr::is_in_range_of_RelAddr32(distance), "too far away (more than +/- 2**32");
2001 }
2002 return;
2003 }
2004
2005 guarantee(false, "not a pcrelative instruction to patch!");
2006 }
2007
2008 // "Current PC" here means the address just behind the basr instruction.
2009 address MacroAssembler::get_PC(Register result) {
2010 z_basr(result, Z_R0); // Don't branch, just save next instruction address in result.
2011 return pc();
2012 }
2013
2014 // Get current PC + offset.
2015 // Offset given in bytes, must be even!
2016 // "Current PC" here means the address of the larl instruction plus the given offset.
2017 address MacroAssembler::get_PC(Register result, int64_t offset) {
2018 address here = pc();
2019 z_larl(result, offset/2); // Save target instruction address in result.
2020 return here + offset;
2021 }
2022
2023 void MacroAssembler::instr_size(Register size, Register pc) {
2024 // Extract 2 most significant bits of current instruction.
2025 z_llgc(size, Address(pc));
2026 z_srl(size, 6);
2027 // Compute (x+3)&6 which translates 0->2, 1->4, 2->4, 3->6.
2028 z_ahi(size, 3);
2029 z_nill(size, 6);
2030 }
2031
2032 // Resize_frame with SP(new) = SP(old) - [offset].
2033 void MacroAssembler::resize_frame_sub(Register offset, Register fp, bool load_fp)
2034 {
2035 assert_different_registers(offset, fp, Z_SP);
2036 if (load_fp) { z_lg(fp, _z_abi(callers_sp), Z_SP); }
2037
2038 z_sgr(Z_SP, offset);
2039 z_stg(fp, _z_abi(callers_sp), Z_SP);
2040 }
2041
2042 // Resize_frame with SP(new) = [newSP] + offset.
2043 // This emitter is useful if we already have calculated a pointer
2044 // into the to-be-allocated stack space, e.g. with special alignment properties,
2045 // but need some additional space, e.g. for spilling.
2046 // newSP is the pre-calculated pointer. It must not be modified.
2047 // fp holds, or is filled with, the frame pointer.
2048 // offset is the additional increment which is added to addr to form the new SP.
2049 // Note: specify a negative value to reserve more space!
2050 // load_fp == true only indicates that fp is not pre-filled with the frame pointer.
2051 // It does not guarantee that fp contains the frame pointer at the end.
2052 void MacroAssembler::resize_frame_abs_with_offset(Register newSP, Register fp, int offset, bool load_fp) {
2053 assert_different_registers(newSP, fp, Z_SP);
2054
2055 if (load_fp) {
2056 z_lg(fp, _z_abi(callers_sp), Z_SP);
2057 }
2058
2059 add2reg(Z_SP, offset, newSP);
2060 z_stg(fp, _z_abi(callers_sp), Z_SP);
2061 }
2062
2063 // Resize_frame with SP(new) = [newSP].
2064 // load_fp == true only indicates that fp is not pre-filled with the frame pointer.
2065 // It does not guarantee that fp contains the frame pointer at the end.
2066 void MacroAssembler::resize_frame_absolute(Register newSP, Register fp, bool load_fp) {
2067 assert_different_registers(newSP, fp, Z_SP);
2068
2069 if (load_fp) {
2070 z_lg(fp, _z_abi(callers_sp), Z_SP); // need to use load/store.
2071 }
2072
2073 z_lgr(Z_SP, newSP);
2074 if (newSP != Z_R0) { // make sure we generate correct code, no matter what register newSP uses.
2075 z_stg(fp, _z_abi(callers_sp), newSP);
2076 } else {
2077 z_stg(fp, _z_abi(callers_sp), Z_SP);
2078 }
2079 }
2080
2081 // Resize_frame with SP(new) = SP(old) + offset.
2082 void MacroAssembler::resize_frame(RegisterOrConstant offset, Register fp, bool load_fp) {
2083 assert_different_registers(fp, Z_SP);
2084
2085 if (load_fp) {
2086 z_lg(fp, _z_abi(callers_sp), Z_SP);
2087 }
2088 add64(Z_SP, offset);
2089 z_stg(fp, _z_abi(callers_sp), Z_SP);
2090 }
2091
2092 void MacroAssembler::push_frame(Register bytes, Register old_sp, bool copy_sp, bool bytes_with_inverted_sign) {
2093 #ifdef ASSERT
2094 assert_different_registers(bytes, old_sp, Z_SP);
2095 if (!copy_sp) {
2096 z_cgr(old_sp, Z_SP);
2097 asm_assert_eq("[old_sp]!=[Z_SP]", 0x211);
2098 }
2099 #endif
2100 if (copy_sp) { z_lgr(old_sp, Z_SP); }
2101 if (bytes_with_inverted_sign) {
2102 z_agr(Z_SP, bytes);
2103 } else {
2104 z_sgr(Z_SP, bytes); // Z_sgfr sufficient, but probably not faster.
2105 }
2106 z_stg(old_sp, _z_abi(callers_sp), Z_SP);
2107 }
2108
2109 unsigned int MacroAssembler::push_frame(unsigned int bytes, Register scratch) {
2110 long offset = Assembler::align(bytes, frame::alignment_in_bytes);
2111 assert(offset > 0, "should push a frame with positive size, size = %ld.", offset);
2112 assert(Displacement::is_validDisp(-offset), "frame size out of range, size = %ld", offset);
2113
2114 // We must not write outside the current stack bounds (given by Z_SP).
2115 // Thus, we have to first update Z_SP and then store the previous SP as stack linkage.
2116 // We rely on Z_R0 by default to be available as scratch.
2117 z_lgr(scratch, Z_SP);
2118 add2reg(Z_SP, -offset);
2119 z_stg(scratch, _z_abi(callers_sp), Z_SP);
2120 #ifdef ASSERT
2121 // Just make sure nobody uses the value in the default scratch register.
2122 // When another register is used, the caller might rely on it containing the frame pointer.
2123 if (scratch == Z_R0) {
2124 z_iihf(scratch, 0xbaadbabe);
2125 z_iilf(scratch, 0xdeadbeef);
2126 }
2127 #endif
2128 return offset;
2129 }
2130
2131 // Push a frame of size `bytes' plus abi160 on top.
2132 unsigned int MacroAssembler::push_frame_abi160(unsigned int bytes) {
2133 BLOCK_COMMENT("push_frame_abi160 {");
2134 unsigned int res = push_frame(bytes + frame::z_abi_160_size);
2135 BLOCK_COMMENT("} push_frame_abi160");
2136 return res;
2137 }
2138
2139 // Pop current C frame.
2140 void MacroAssembler::pop_frame() {
2141 BLOCK_COMMENT("pop_frame:");
2142 Assembler::z_lg(Z_SP, _z_abi(callers_sp), Z_SP);
2143 }
2144
2145 // Pop current C frame and restore return PC register (Z_R14).
2146 void MacroAssembler::pop_frame_restore_retPC(int frame_size_in_bytes) {
2147 BLOCK_COMMENT("pop_frame_restore_retPC:");
2148 int retPC_offset = _z_abi16(return_pc) + frame_size_in_bytes;
2149 // If possible, pop frame by add instead of load (a penny saved is a penny got :-).
2150 if (Displacement::is_validDisp(retPC_offset)) {
2151 z_lg(Z_R14, retPC_offset, Z_SP);
2152 add2reg(Z_SP, frame_size_in_bytes);
2153 } else {
2154 add2reg(Z_SP, frame_size_in_bytes);
2155 restore_return_pc();
2156 }
2157 }
2158
2159 void MacroAssembler::call_VM_leaf_base(address entry_point, bool allow_relocation) {
2160 if (allow_relocation) {
2161 call_c(entry_point);
2162 } else {
2163 call_c_static(entry_point);
2164 }
2165 }
2166
2167 void MacroAssembler::call_VM_leaf_base(address entry_point) {
2168 bool allow_relocation = true;
2169 call_VM_leaf_base(entry_point, allow_relocation);
2170 }
2171
2172 void MacroAssembler::call_VM_base(Register oop_result,
2173 Register last_java_sp,
2174 address entry_point,
2175 bool allow_relocation,
2176 bool check_exceptions) { // Defaults to true.
2177 // Allow_relocation indicates, if true, that the generated code shall
2178 // be fit for code relocation or referenced data relocation. In other
2179 // words: all addresses must be considered variable. PC-relative addressing
2180 // is not possible then.
2181 // On the other hand, if (allow_relocation == false), addresses and offsets
2182 // may be considered stable, enabling us to take advantage of some PC-relative
2183 // addressing tweaks. These might improve performance and reduce code size.
2184
2185 // Determine last_java_sp register.
2186 if (!last_java_sp->is_valid()) {
2187 last_java_sp = Z_SP; // Load Z_SP as SP.
2188 }
2189
2190 set_top_ijava_frame_at_SP_as_last_Java_frame(last_java_sp, Z_R1, allow_relocation);
2191
2192 // ARG1 must hold thread address.
2193 z_lgr(Z_ARG1, Z_thread);
2194
2195 address return_pc = NULL;
2196 if (allow_relocation) {
2197 return_pc = call_c(entry_point);
2198 } else {
2199 return_pc = call_c_static(entry_point);
2200 }
2201
2202 reset_last_Java_frame(allow_relocation);
2203
2204 // C++ interp handles this in the interpreter.
2205 check_and_handle_popframe(Z_thread);
2206 check_and_handle_earlyret(Z_thread);
2207
2208 // Check for pending exceptions.
2209 if (check_exceptions) {
2210 // Check for pending exceptions (java_thread is set upon return).
2211 load_and_test_long(Z_R0_scratch, Address(Z_thread, Thread::pending_exception_offset()));
2212
2213 // This used to conditionally jump to forward_exception however it is
2214 // possible if we relocate that the branch will not reach. So we must jump
2215 // around so we can always reach.
2216
2217 Label ok;
2218 z_bre(ok); // Bcondequal is the same as bcondZero.
2219 call_stub(StubRoutines::forward_exception_entry());
2220 bind(ok);
2221 }
2222
2223 // Get oop result if there is one and reset the value in the thread.
2224 if (oop_result->is_valid()) {
2225 get_vm_result(oop_result);
2226 }
2227
2228 _last_calls_return_pc = return_pc; // Wipe out other (error handling) calls.
2229 }
2230
2231 void MacroAssembler::call_VM_base(Register oop_result,
2232 Register last_java_sp,
2233 address entry_point,
2234 bool check_exceptions) { // Defaults to true.
2235 bool allow_relocation = true;
2236 call_VM_base(oop_result, last_java_sp, entry_point, allow_relocation, check_exceptions);
2237 }
2238
2239 // VM calls without explicit last_java_sp.
2240
2241 void MacroAssembler::call_VM(Register oop_result, address entry_point, bool check_exceptions) {
2242 // Call takes possible detour via InterpreterMacroAssembler.
2243 call_VM_base(oop_result, noreg, entry_point, true, check_exceptions);
2244 }
2245
2246 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, bool check_exceptions) {
2247 // Z_ARG1 is reserved for the thread.
2248 lgr_if_needed(Z_ARG2, arg_1);
2249 call_VM(oop_result, entry_point, check_exceptions);
2250 }
2251
2252 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2, bool check_exceptions) {
2253 // Z_ARG1 is reserved for the thread.
2254 lgr_if_needed(Z_ARG2, arg_1);
2255 assert(arg_2 != Z_ARG2, "smashed argument");
2256 lgr_if_needed(Z_ARG3, arg_2);
2257 call_VM(oop_result, entry_point, check_exceptions);
2258 }
2259
2260 void MacroAssembler::call_VM(Register oop_result, address entry_point, Register arg_1, Register arg_2,
2261 Register arg_3, bool check_exceptions) {
2262 // Z_ARG1 is reserved for the thread.
2263 lgr_if_needed(Z_ARG2, arg_1);
2264 assert(arg_2 != Z_ARG2, "smashed argument");
2265 lgr_if_needed(Z_ARG3, arg_2);
2266 assert(arg_3 != Z_ARG2 && arg_3 != Z_ARG3, "smashed argument");
2267 lgr_if_needed(Z_ARG4, arg_3);
2268 call_VM(oop_result, entry_point, check_exceptions);
2269 }
2270
2271 // VM static calls without explicit last_java_sp.
2272
2273 void MacroAssembler::call_VM_static(Register oop_result, address entry_point, bool check_exceptions) {
2274 // Call takes possible detour via InterpreterMacroAssembler.
2275 call_VM_base(oop_result, noreg, entry_point, false, check_exceptions);
2276 }
2277
2278 void MacroAssembler::call_VM_static(Register oop_result, address entry_point, Register arg_1, Register arg_2,
2279 Register arg_3, bool check_exceptions) {
2280 // Z_ARG1 is reserved for the thread.
2281 lgr_if_needed(Z_ARG2, arg_1);
2282 assert(arg_2 != Z_ARG2, "smashed argument");
2283 lgr_if_needed(Z_ARG3, arg_2);
2284 assert(arg_3 != Z_ARG2 && arg_3 != Z_ARG3, "smashed argument");
2285 lgr_if_needed(Z_ARG4, arg_3);
2286 call_VM_static(oop_result, entry_point, check_exceptions);
2287 }
2288
2289 // VM calls with explicit last_java_sp.
2290
2291 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, bool check_exceptions) {
2292 // Call takes possible detour via InterpreterMacroAssembler.
2293 call_VM_base(oop_result, last_java_sp, entry_point, true, check_exceptions);
2294 }
2295
2296 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1, bool check_exceptions) {
2297 // Z_ARG1 is reserved for the thread.
2298 lgr_if_needed(Z_ARG2, arg_1);
2299 call_VM(oop_result, last_java_sp, entry_point, check_exceptions);
2300 }
2301
2302 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1,
2303 Register arg_2, bool check_exceptions) {
2304 // Z_ARG1 is reserved for the thread.
2305 lgr_if_needed(Z_ARG2, arg_1);
2306 assert(arg_2 != Z_ARG2, "smashed argument");
2307 lgr_if_needed(Z_ARG3, arg_2);
2308 call_VM(oop_result, last_java_sp, entry_point, check_exceptions);
2309 }
2310
2311 void MacroAssembler::call_VM(Register oop_result, Register last_java_sp, address entry_point, Register arg_1,
2312 Register arg_2, Register arg_3, bool check_exceptions) {
2313 // Z_ARG1 is reserved for the thread.
2314 lgr_if_needed(Z_ARG2, arg_1);
2315 assert(arg_2 != Z_ARG2, "smashed argument");
2316 lgr_if_needed(Z_ARG3, arg_2);
2317 assert(arg_3 != Z_ARG2 && arg_3 != Z_ARG3, "smashed argument");
2318 lgr_if_needed(Z_ARG4, arg_3);
2319 call_VM(oop_result, last_java_sp, entry_point, check_exceptions);
2320 }
2321
2322 // VM leaf calls.
2323
2324 void MacroAssembler::call_VM_leaf(address entry_point) {
2325 // Call takes possible detour via InterpreterMacroAssembler.
2326 call_VM_leaf_base(entry_point, true);
2327 }
2328
2329 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1) {
2330 if (arg_1 != noreg) lgr_if_needed(Z_ARG1, arg_1);
2331 call_VM_leaf(entry_point);
2332 }
2333
2334 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1, Register arg_2) {
2335 if (arg_1 != noreg) lgr_if_needed(Z_ARG1, arg_1);
2336 assert(arg_2 != Z_ARG1, "smashed argument");
2337 if (arg_2 != noreg) lgr_if_needed(Z_ARG2, arg_2);
2338 call_VM_leaf(entry_point);
2339 }
2340
2341 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_1, Register arg_2, Register arg_3) {
2342 if (arg_1 != noreg) lgr_if_needed(Z_ARG1, arg_1);
2343 assert(arg_2 != Z_ARG1, "smashed argument");
2344 if (arg_2 != noreg) lgr_if_needed(Z_ARG2, arg_2);
2345 assert(arg_3 != Z_ARG1 && arg_3 != Z_ARG2, "smashed argument");
2346 if (arg_3 != noreg) lgr_if_needed(Z_ARG3, arg_3);
2347 call_VM_leaf(entry_point);
2348 }
2349
2350 // Static VM leaf calls.
2351 // Really static VM leaf calls are never patched.
2352
2353 void MacroAssembler::call_VM_leaf_static(address entry_point) {
2354 // Call takes possible detour via InterpreterMacroAssembler.
2355 call_VM_leaf_base(entry_point, false);
2356 }
2357
2358 void MacroAssembler::call_VM_leaf_static(address entry_point, Register arg_1) {
2359 if (arg_1 != noreg) lgr_if_needed(Z_ARG1, arg_1);
2360 call_VM_leaf_static(entry_point);
2361 }
2362
2363 void MacroAssembler::call_VM_leaf_static(address entry_point, Register arg_1, Register arg_2) {
2364 if (arg_1 != noreg) lgr_if_needed(Z_ARG1, arg_1);
2365 assert(arg_2 != Z_ARG1, "smashed argument");
2366 if (arg_2 != noreg) lgr_if_needed(Z_ARG2, arg_2);
2367 call_VM_leaf_static(entry_point);
2368 }
2369
2370 void MacroAssembler::call_VM_leaf_static(address entry_point, Register arg_1, Register arg_2, Register arg_3) {
2371 if (arg_1 != noreg) lgr_if_needed(Z_ARG1, arg_1);
2372 assert(arg_2 != Z_ARG1, "smashed argument");
2373 if (arg_2 != noreg) lgr_if_needed(Z_ARG2, arg_2);
2374 assert(arg_3 != Z_ARG1 && arg_3 != Z_ARG2, "smashed argument");
2375 if (arg_3 != noreg) lgr_if_needed(Z_ARG3, arg_3);
2376 call_VM_leaf_static(entry_point);
2377 }
2378
2379 // Don't use detour via call_c(reg).
2380 address MacroAssembler::call_c(address function_entry) {
2381 load_const(Z_R1, function_entry);
2382 return call(Z_R1);
2383 }
2384
2385 // Variant for really static (non-relocatable) calls which are never patched.
2386 address MacroAssembler::call_c_static(address function_entry) {
2387 load_absolute_address(Z_R1, function_entry);
2388 #if 0 // def ASSERT
2389 // Verify that call site did not move.
2390 load_const_optimized(Z_R0, function_entry);
2391 z_cgr(Z_R1, Z_R0);
2392 z_brc(bcondEqual, 3);
2393 z_illtrap(0xba);
2394 #endif
2395 return call(Z_R1);
2396 }
2397
2398 address MacroAssembler::call_c_opt(address function_entry) {
2399 bool success = call_far_patchable(function_entry, -2 /* emit relocation + constant */);
2400 _last_calls_return_pc = success ? pc() : NULL;
2401 return _last_calls_return_pc;
2402 }
2403
2404 // Identify a call_far_patchable instruction: LARL + LG + BASR
2405 //
2406 // nop ; optionally, if required for alignment
2407 // lgrl rx,A(TOC entry) ; PC-relative access into constant pool
2408 // basr Z_R14,rx ; end of this instruction must be aligned to a word boundary
2409 //
2410 // Code pattern will eventually get patched into variant2 (see below for detection code).
2411 //
2412 bool MacroAssembler::is_call_far_patchable_variant0_at(address instruction_addr) {
2413 address iaddr = instruction_addr;
2414
2415 // Check for the actual load instruction.
2416 if (!is_load_const_from_toc(iaddr)) { return false; }
2417 iaddr += load_const_from_toc_size();
2418
2419 // Check for the call (BASR) instruction, finally.
2420 assert(iaddr-instruction_addr+call_byregister_size() == call_far_patchable_size(), "size mismatch");
2421 return is_call_byregister(iaddr);
2422 }
2423
2424 // Identify a call_far_patchable instruction: BRASL
2425 //
2426 // Code pattern to suits atomic patching:
2427 // nop ; Optionally, if required for alignment.
2428 // nop ... ; Multiple filler nops to compensate for size difference (variant0 is longer).
2429 // nop ; For code pattern detection: Prepend each BRASL with a nop.
2430 // brasl Z_R14,<reladdr> ; End of code must be 4-byte aligned !
2431 bool MacroAssembler::is_call_far_patchable_variant2_at(address instruction_addr) {
2432 const address call_addr = (address)((intptr_t)instruction_addr + call_far_patchable_size() - call_far_pcrelative_size());
2433
2434 // Check for correct number of leading nops.
2435 address iaddr;
2436 for (iaddr = instruction_addr; iaddr < call_addr; iaddr += nop_size()) {
2437 if (!is_z_nop(iaddr)) { return false; }
2438 }
2439 assert(iaddr == call_addr, "sanity");
2440
2441 // --> Check for call instruction.
2442 if (is_call_far_pcrelative(call_addr)) {
2443 assert(call_addr-instruction_addr+call_far_pcrelative_size() == call_far_patchable_size(), "size mismatch");
2444 return true;
2445 }
2446
2447 return false;
2448 }
2449
2450 // Emit a NOT mt-safely patchable 64 bit absolute call.
2451 // If toc_offset == -2, then the destination of the call (= target) is emitted
2452 // to the constant pool and a runtime_call relocation is added
2453 // to the code buffer.
2454 // If toc_offset != -2, target must already be in the constant pool at
2455 // _ctableStart+toc_offset (a caller can retrieve toc_offset
2456 // from the runtime_call relocation).
2457 // Special handling of emitting to scratch buffer when there is no constant pool.
2458 // Slightly changed code pattern. We emit an additional nop if we would
2459 // not end emitting at a word aligned address. This is to ensure
2460 // an atomically patchable displacement in brasl instructions.
2461 //
2462 // A call_far_patchable comes in different flavors:
2463 // - LARL(CP) / LG(CP) / BR (address in constant pool, access via CP register)
2464 // - LGRL(CP) / BR (address in constant pool, pc-relative accesss)
2465 // - BRASL (relative address of call target coded in instruction)
2466 // All flavors occupy the same amount of space. Length differences are compensated
2467 // by leading nops, such that the instruction sequence always ends at the same
2468 // byte offset. This is required to keep the return offset constant.
2469 // Furthermore, the return address (the end of the instruction sequence) is forced
2470 // to be on a 4-byte boundary. This is required for atomic patching, should we ever
2471 // need to patch the call target of the BRASL flavor.
2472 // RETURN value: false, if no constant pool entry could be allocated, true otherwise.
2473 bool MacroAssembler::call_far_patchable(address target, int64_t tocOffset) {
2474 // Get current pc and ensure word alignment for end of instr sequence.
2475 const address start_pc = pc();
2476 const intptr_t start_off = offset();
2477 assert(!call_far_patchable_requires_alignment_nop(start_pc), "call_far_patchable requires aligned address");
2478 const ptrdiff_t dist = (ptrdiff_t)(target - (start_pc + 2)); // Prepend each BRASL with a nop.
2479 const bool emit_target_to_pool = (tocOffset == -2) && !code_section()->scratch_emit();
2480 const bool emit_relative_call = !emit_target_to_pool &&
2481 RelAddr::is_in_range_of_RelAddr32(dist) &&
2482 ReoptimizeCallSequences &&
2483 !code_section()->scratch_emit();
2484
2485 if (emit_relative_call) {
2486 // Add padding to get the same size as below.
2487 const unsigned int padding = call_far_patchable_size() - call_far_pcrelative_size();
2488 unsigned int current_padding;
2489 for (current_padding = 0; current_padding < padding; current_padding += nop_size()) { z_nop(); }
2490 assert(current_padding == padding, "sanity");
2491
2492 // relative call: len = 2(nop) + 6 (brasl)
2493 // CodeBlob resize cannot occur in this case because
2494 // this call is emitted into pre-existing space.
2495 z_nop(); // Prepend each BRASL with a nop.
2496 z_brasl(Z_R14, target);
2497 } else {
2498 // absolute call: Get address from TOC.
2499 // len = (load TOC){6|0} + (load from TOC){6} + (basr){2} = {14|8}
2500 if (emit_target_to_pool) {
2501 // When emitting the call for the first time, we do not need to use
2502 // the pc-relative version. It will be patched anyway, when the code
2503 // buffer is copied.
2504 // Relocation is not needed when !ReoptimizeCallSequences.
2505 relocInfo::relocType rt = ReoptimizeCallSequences ? relocInfo::runtime_call_w_cp_type : relocInfo::none;
2506 AddressLiteral dest(target, rt);
2507 // Store_oop_in_toc() adds dest to the constant table. As side effect, this kills
2508 // inst_mark(). Reset if possible.
2509 bool reset_mark = (inst_mark() == pc());
2510 tocOffset = store_oop_in_toc(dest);
2511 if (reset_mark) { set_inst_mark(); }
2512 if (tocOffset == -1) {
2513 return false; // Couldn't create constant pool entry.
2514 }
2515 }
2516 assert(offset() == start_off, "emit no code before this point!");
2517
2518 address tocPos = pc() + tocOffset;
2519 if (emit_target_to_pool) {
2520 tocPos = code()->consts()->start() + tocOffset;
2521 }
2522 load_long_pcrelative(Z_R14, tocPos);
2523 z_basr(Z_R14, Z_R14);
2524 }
2525
2526 #ifdef ASSERT
2527 // Assert that we can identify the emitted call.
2528 assert(is_call_far_patchable_at(addr_at(start_off)), "can't identify emitted call");
2529 assert(offset() == start_off+call_far_patchable_size(), "wrong size");
2530
2531 if (emit_target_to_pool) {
2532 assert(get_dest_of_call_far_patchable_at(addr_at(start_off), code()->consts()->start()) == target,
2533 "wrong encoding of dest address");
2534 }
2535 #endif
2536 return true; // success
2537 }
2538
2539 // Identify a call_far_patchable instruction.
2540 // For more detailed information see header comment of call_far_patchable.
2541 bool MacroAssembler::is_call_far_patchable_at(address instruction_addr) {
2542 return is_call_far_patchable_variant2_at(instruction_addr) || // short version: BRASL
2543 is_call_far_patchable_variant0_at(instruction_addr); // long version LARL + LG + BASR
2544 }
2545
2546 // Does the call_far_patchable instruction use a pc-relative encoding
2547 // of the call destination?
2548 bool MacroAssembler::is_call_far_patchable_pcrelative_at(address instruction_addr) {
2549 // Variant 2 is pc-relative.
2550 return is_call_far_patchable_variant2_at(instruction_addr);
2551 }
2552
2553 bool MacroAssembler::is_call_far_pcrelative(address instruction_addr) {
2554 // Prepend each BRASL with a nop.
2555 return is_z_nop(instruction_addr) && is_z_brasl(instruction_addr + nop_size()); // Match at position after one nop required.
2556 }
2557
2558 // Set destination address of a call_far_patchable instruction.
2559 void MacroAssembler::set_dest_of_call_far_patchable_at(address instruction_addr, address dest, int64_t tocOffset) {
2560 ResourceMark rm;
2561
2562 // Now that CP entry is verified, patch call to a pc-relative call (if circumstances permit).
2563 int code_size = MacroAssembler::call_far_patchable_size();
2564 CodeBuffer buf(instruction_addr, code_size);
2565 MacroAssembler masm(&buf);
2566 masm.call_far_patchable(dest, tocOffset);
2567 ICache::invalidate_range(instruction_addr, code_size); // Empty on z.
2568 }
2569
2570 // Get dest address of a call_far_patchable instruction.
2571 address MacroAssembler::get_dest_of_call_far_patchable_at(address instruction_addr, address ctable) {
2572 // Dynamic TOC: absolute address in constant pool.
2573 // Check variant2 first, it is more frequent.
2574
2575 // Relative address encoded in call instruction.
2576 if (is_call_far_patchable_variant2_at(instruction_addr)) {
2577 return MacroAssembler::get_target_addr_pcrel(instruction_addr + nop_size()); // Prepend each BRASL with a nop.
2578
2579 // Absolute address in constant pool.
2580 } else if (is_call_far_patchable_variant0_at(instruction_addr)) {
2581 address iaddr = instruction_addr;
2582
2583 long tocOffset = get_load_const_from_toc_offset(iaddr);
2584 address tocLoc = iaddr + tocOffset;
2585 return *(address *)(tocLoc);
2586 } else {
2587 fprintf(stderr, "MacroAssembler::get_dest_of_call_far_patchable_at has a problem at %p:\n", instruction_addr);
2588 fprintf(stderr, "not a call_far_patchable: %16.16lx %16.16lx, len = %d\n",
2589 *(unsigned long*)instruction_addr,
2590 *(unsigned long*)(instruction_addr+8),
2591 call_far_patchable_size());
2592 Disassembler::decode(instruction_addr, instruction_addr+call_far_patchable_size());
2593 ShouldNotReachHere();
2594 return NULL;
2595 }
2596 }
2597
2598 void MacroAssembler::align_call_far_patchable(address pc) {
2599 if (call_far_patchable_requires_alignment_nop(pc)) { z_nop(); }
2600 }
2601
2602 void MacroAssembler::check_and_handle_earlyret(Register java_thread) {
2603 }
2604
2605 void MacroAssembler::check_and_handle_popframe(Register java_thread) {
2606 }
2607
2608 // Read from the polling page.
2609 // Use TM or TMY instruction, depending on read offset.
2610 // offset = 0: Use TM, safepoint polling.
2611 // offset < 0: Use TMY, profiling safepoint polling.
2612 void MacroAssembler::load_from_polling_page(Register polling_page_address, int64_t offset) {
2613 if (Immediate::is_uimm12(offset)) {
2614 z_tm(offset, polling_page_address, mask_safepoint);
2615 } else {
2616 z_tmy(offset, polling_page_address, mask_profiling);
2617 }
2618 }
2619
2620 // Check whether z_instruction is a read access to the polling page
2621 // which was emitted by load_from_polling_page(..).
2622 bool MacroAssembler::is_load_from_polling_page(address instr_loc) {
2623 unsigned long z_instruction;
2624 unsigned int ilen = get_instruction(instr_loc, &z_instruction);
2625
2626 if (ilen == 2) { return false; } // It's none of the allowed instructions.
2627
2628 if (ilen == 4) {
2629 if (!is_z_tm(z_instruction)) { return false; } // It's len=4, but not a z_tm. fail.
2630
2631 int ms = inv_mask(z_instruction,8,32); // mask
2632 int ra = inv_reg(z_instruction,16,32); // base register
2633 int ds = inv_uimm12(z_instruction); // displacement
2634
2635 if (!(ds == 0 && ra != 0 && ms == mask_safepoint)) {
2636 return false; // It's not a z_tm(0, ra, mask_safepoint). Fail.
2637 }
2638
2639 } else { /* if (ilen == 6) */
2640
2641 assert(!is_z_lg(z_instruction), "old form (LG) polling page access. Please fix and use TM(Y).");
2642
2643 if (!is_z_tmy(z_instruction)) { return false; } // It's len=6, but not a z_tmy. fail.
2644
2645 int ms = inv_mask(z_instruction,8,48); // mask
2646 int ra = inv_reg(z_instruction,16,48); // base register
2647 int ds = inv_simm20(z_instruction); // displacement
2648 }
2649
2650 return true;
2651 }
2652
2653 // Extract poll address from instruction and ucontext.
2654 address MacroAssembler::get_poll_address(address instr_loc, void* ucontext) {
2655 assert(ucontext != NULL, "must have ucontext");
2656 ucontext_t* uc = (ucontext_t*) ucontext;
2657 unsigned long z_instruction;
2658 unsigned int ilen = get_instruction(instr_loc, &z_instruction);
2659
2660 if (ilen == 4 && is_z_tm(z_instruction)) {
2661 int ra = inv_reg(z_instruction, 16, 32); // base register
2662 int ds = inv_uimm12(z_instruction); // displacement
2663 address addr = (address)uc->uc_mcontext.gregs[ra];
2664 return addr + ds;
2665 } else if (ilen == 6 && is_z_tmy(z_instruction)) {
2666 int ra = inv_reg(z_instruction, 16, 48); // base register
2667 int ds = inv_simm20(z_instruction); // displacement
2668 address addr = (address)uc->uc_mcontext.gregs[ra];
2669 return addr + ds;
2670 }
2671
2672 ShouldNotReachHere();
2673 return NULL;
2674 }
2675
2676 // Extract poll register from instruction.
2677 uint MacroAssembler::get_poll_register(address instr_loc) {
2678 unsigned long z_instruction;
2679 unsigned int ilen = get_instruction(instr_loc, &z_instruction);
2680
2681 if (ilen == 4 && is_z_tm(z_instruction)) {
2682 return (uint)inv_reg(z_instruction, 16, 32); // base register
2683 } else if (ilen == 6 && is_z_tmy(z_instruction)) {
2684 return (uint)inv_reg(z_instruction, 16, 48); // base register
2685 }
2686
2687 ShouldNotReachHere();
2688 return 0;
2689 }
2690
2691 bool MacroAssembler::is_memory_serialization(int instruction, JavaThread* thread, void* ucontext) {
2692 ShouldNotCallThis();
2693 return false;
2694 }
2695
2696 // Write serialization page so VM thread can do a pseudo remote membar
2697 // We use the current thread pointer to calculate a thread specific
2698 // offset to write to within the page. This minimizes bus traffic
2699 // due to cache line collision.
2700 void MacroAssembler::serialize_memory(Register thread, Register tmp1, Register tmp2) {
2701 assert_different_registers(tmp1, tmp2);
2702 z_sllg(tmp2, thread, os::get_serialize_page_shift_count());
2703 load_const_optimized(tmp1, (long) os::get_memory_serialize_page());
2704
2705 int mask = os::get_serialize_page_mask();
2706 if (Immediate::is_uimm16(mask)) {
2707 z_nill(tmp2, mask);
2708 z_llghr(tmp2, tmp2);
2709 } else {
2710 z_nilf(tmp2, mask);
2711 z_llgfr(tmp2, tmp2);
2712 }
2713
2714 z_release();
2715 z_st(Z_R0, 0, tmp2, tmp1);
2716 }
2717
2718 void MacroAssembler::safepoint_poll(Label& slow_path, Register temp_reg) {
2719 if (SafepointMechanism::uses_thread_local_poll()) {
2720 const Address poll_byte_addr(Z_thread, in_bytes(Thread::polling_page_offset()) + 7 /* Big Endian */);
2721 // Armed page has poll_bit set.
2722 z_tm(poll_byte_addr, SafepointMechanism::poll_bit());
2723 z_brnaz(slow_path);
2724 } else {
2725 load_const_optimized(temp_reg, SafepointSynchronize::address_of_state());
2726 z_cli(/*SafepointSynchronize::sz_state()*/4-1, temp_reg, SafepointSynchronize::_not_synchronized);
2727 z_brne(slow_path);
2728 }
2729 }
2730
2731 // Don't rely on register locking, always use Z_R1 as scratch register instead.
2732 void MacroAssembler::bang_stack_with_offset(int offset) {
2733 // Stack grows down, caller passes positive offset.
2734 assert(offset > 0, "must bang with positive offset");
2735 if (Displacement::is_validDisp(-offset)) {
2736 z_tmy(-offset, Z_SP, mask_stackbang);
2737 } else {
2738 add2reg(Z_R1, -offset, Z_SP); // Do not destroy Z_SP!!!
2739 z_tm(0, Z_R1, mask_stackbang); // Just banging.
2740 }
2741 }
2742
2743 void MacroAssembler::reserved_stack_check(Register return_pc) {
2744 // Test if reserved zone needs to be enabled.
2745 Label no_reserved_zone_enabling;
2746 assert(return_pc == Z_R14, "Return pc must be in R14 before z_br() to StackOverflow stub.");
2747 BLOCK_COMMENT("reserved_stack_check {");
2748
2749 z_clg(Z_SP, Address(Z_thread, JavaThread::reserved_stack_activation_offset()));
2750 z_brl(no_reserved_zone_enabling);
2751
2752 // Enable reserved zone again, throw stack overflow exception.
2753 save_return_pc();
2754 push_frame_abi160(0);
2755 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone), Z_thread);
2756 pop_frame();
2757 restore_return_pc();
2758
2759 load_const_optimized(Z_R1, StubRoutines::throw_delayed_StackOverflowError_entry());
2760 // Don't use call() or z_basr(), they will invalidate Z_R14 which contains the return pc.
2761 z_br(Z_R1);
2762
2763 should_not_reach_here();
2764
2765 bind(no_reserved_zone_enabling);
2766 BLOCK_COMMENT("} reserved_stack_check");
2767 }
2768
2769 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
2770 void MacroAssembler::tlab_allocate(Register obj,
2771 Register var_size_in_bytes,
2772 int con_size_in_bytes,
2773 Register t1,
2774 Label& slow_case) {
2775 assert_different_registers(obj, var_size_in_bytes, t1);
2776 Register end = t1;
2777 Register thread = Z_thread;
2778
2779 z_lg(obj, Address(thread, JavaThread::tlab_top_offset()));
2780 if (var_size_in_bytes == noreg) {
2781 z_lay(end, Address(obj, con_size_in_bytes));
2782 } else {
2783 z_lay(end, Address(obj, var_size_in_bytes));
2784 }
2785 z_cg(end, Address(thread, JavaThread::tlab_end_offset()));
2786 branch_optimized(bcondHigh, slow_case);
2787
2788 // Update the tlab top pointer.
2789 z_stg(end, Address(thread, JavaThread::tlab_top_offset()));
2790
2791 // Recover var_size_in_bytes if necessary.
2792 if (var_size_in_bytes == end) {
2793 z_sgr(var_size_in_bytes, obj);
2794 }
2795 }
2796
2797 // Emitter for interface method lookup.
2798 // input: recv_klass, intf_klass, itable_index
2799 // output: method_result
2800 // kills: itable_index, temp1_reg, Z_R0, Z_R1
2801 // TODO: Temp2_reg is unused. we may use this emitter also in the itable stubs.
2802 // If the register is still not needed then, remove it.
2803 void MacroAssembler::lookup_interface_method(Register recv_klass,
2804 Register intf_klass,
2805 RegisterOrConstant itable_index,
2806 Register method_result,
2807 Register temp1_reg,
2808 Label& no_such_interface,
2809 bool return_method) {
2810
2811 const Register vtable_len = temp1_reg; // Used to compute itable_entry_addr.
2812 const Register itable_entry_addr = Z_R1_scratch;
2813 const Register itable_interface = Z_R0_scratch;
2814
2815 BLOCK_COMMENT("lookup_interface_method {");
2816
2817 // Load start of itable entries into itable_entry_addr.
2818 z_llgf(vtable_len, Address(recv_klass, Klass::vtable_length_offset()));
2819 z_sllg(vtable_len, vtable_len, exact_log2(vtableEntry::size_in_bytes()));
2820
2821 // Loop over all itable entries until desired interfaceOop(Rinterface) found.
2822 const int vtable_base_offset = in_bytes(Klass::vtable_start_offset());
2823
2824 add2reg_with_index(itable_entry_addr,
2825 vtable_base_offset + itableOffsetEntry::interface_offset_in_bytes(),
2826 recv_klass, vtable_len);
2827
2828 const int itable_offset_search_inc = itableOffsetEntry::size() * wordSize;
2829 Label search;
2830
2831 bind(search);
2832
2833 // Handle IncompatibleClassChangeError.
2834 // If the entry is NULL then we've reached the end of the table
2835 // without finding the expected interface, so throw an exception.
2836 load_and_test_long(itable_interface, Address(itable_entry_addr));
2837 z_bre(no_such_interface);
2838
2839 add2reg(itable_entry_addr, itable_offset_search_inc);
2840 z_cgr(itable_interface, intf_klass);
2841 z_brne(search);
2842
2843 // Entry found and itable_entry_addr points to it, get offset of vtable for interface.
2844 if (return_method) {
2845 const int vtable_offset_offset = (itableOffsetEntry::offset_offset_in_bytes() -
2846 itableOffsetEntry::interface_offset_in_bytes()) -
2847 itable_offset_search_inc;
2848
2849 // Compute itableMethodEntry and get method and entry point
2850 // we use addressing with index and displacement, since the formula
2851 // for computing the entry's offset has a fixed and a dynamic part,
2852 // the latter depending on the matched interface entry and on the case,
2853 // that the itable index has been passed as a register, not a constant value.
2854 int method_offset = itableMethodEntry::method_offset_in_bytes();
2855 // Fixed part (displacement), common operand.
2856 Register itable_offset = method_result; // Dynamic part (index register).
2857
2858 if (itable_index.is_register()) {
2859 // Compute the method's offset in that register, for the formula, see the
2860 // else-clause below.
2861 z_sllg(itable_offset, itable_index.as_register(), exact_log2(itableMethodEntry::size() * wordSize));
2862 z_agf(itable_offset, vtable_offset_offset, itable_entry_addr);
2863 } else {
2864 // Displacement increases.
2865 method_offset += itableMethodEntry::size() * wordSize * itable_index.as_constant();
2866
2867 // Load index from itable.
2868 z_llgf(itable_offset, vtable_offset_offset, itable_entry_addr);
2869 }
2870
2871 // Finally load the method's oop.
2872 z_lg(method_result, method_offset, itable_offset, recv_klass);
2873 }
2874 BLOCK_COMMENT("} lookup_interface_method");
2875 }
2876
2877 // Lookup for virtual method invocation.
2878 void MacroAssembler::lookup_virtual_method(Register recv_klass,
2879 RegisterOrConstant vtable_index,
2880 Register method_result) {
2881 assert_different_registers(recv_klass, vtable_index.register_or_noreg());
2882 assert(vtableEntry::size() * wordSize == wordSize,
2883 "else adjust the scaling in the code below");
2884
2885 BLOCK_COMMENT("lookup_virtual_method {");
2886
2887 const int base = in_bytes(Klass::vtable_start_offset());
2888
2889 if (vtable_index.is_constant()) {
2890 // Load with base + disp.
2891 Address vtable_entry_addr(recv_klass,
2892 vtable_index.as_constant() * wordSize +
2893 base +
2894 vtableEntry::method_offset_in_bytes());
2895
2896 z_lg(method_result, vtable_entry_addr);
2897 } else {
2898 // Shift index properly and load with base + index + disp.
2899 Register vindex = vtable_index.as_register();
2900 Address vtable_entry_addr(recv_klass, vindex,
2901 base + vtableEntry::method_offset_in_bytes());
2902
2903 z_sllg(vindex, vindex, exact_log2(wordSize));
2904 z_lg(method_result, vtable_entry_addr);
2905 }
2906 BLOCK_COMMENT("} lookup_virtual_method");
2907 }
2908
2909 // Factor out code to call ic_miss_handler.
2910 // Generate code to call the inline cache miss handler.
2911 //
2912 // In most cases, this code will be generated out-of-line.
2913 // The method parameters are intended to provide some variability.
2914 // ICM - Label which has to be bound to the start of useful code (past any traps).
2915 // trapMarker - Marking byte for the generated illtrap instructions (if any).
2916 // Any value except 0x00 is supported.
2917 // = 0x00 - do not generate illtrap instructions.
2918 // use nops to fill ununsed space.
2919 // requiredSize - required size of the generated code. If the actually
2920 // generated code is smaller, use padding instructions to fill up.
2921 // = 0 - no size requirement, no padding.
2922 // scratch - scratch register to hold branch target address.
2923 //
2924 // The method returns the code offset of the bound label.
2925 unsigned int MacroAssembler::call_ic_miss_handler(Label& ICM, int trapMarker, int requiredSize, Register scratch) {
2926 intptr_t startOffset = offset();
2927
2928 // Prevent entry at content_begin().
2929 if (trapMarker != 0) {
2930 z_illtrap(trapMarker);
2931 }
2932
2933 // Load address of inline cache miss code into scratch register
2934 // and branch to cache miss handler.
2935 BLOCK_COMMENT("IC miss handler {");
2936 BIND(ICM);
2937 unsigned int labelOffset = offset();
2938 AddressLiteral icmiss(SharedRuntime::get_ic_miss_stub());
2939
2940 load_const_optimized(scratch, icmiss);
2941 z_br(scratch);
2942
2943 // Fill unused space.
2944 if (requiredSize > 0) {
2945 while ((offset() - startOffset) < requiredSize) {
2946 if (trapMarker == 0) {
2947 z_nop();
2948 } else {
2949 z_illtrap(trapMarker);
2950 }
2951 }
2952 }
2953 BLOCK_COMMENT("} IC miss handler");
2954 return labelOffset;
2955 }
2956
2957 void MacroAssembler::nmethod_UEP(Label& ic_miss) {
2958 Register ic_reg = as_Register(Matcher::inline_cache_reg_encode());
2959 int klass_offset = oopDesc::klass_offset_in_bytes();
2960 if (!ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(klass_offset)) {
2961 if (VM_Version::has_CompareBranch()) {
2962 z_cgij(Z_ARG1, 0, Assembler::bcondEqual, ic_miss);
2963 } else {
2964 z_ltgr(Z_ARG1, Z_ARG1);
2965 z_bre(ic_miss);
2966 }
2967 }
2968 // Compare cached class against klass from receiver.
2969 compare_klass_ptr(ic_reg, klass_offset, Z_ARG1, false);
2970 z_brne(ic_miss);
2971 }
2972
2973 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
2974 Register super_klass,
2975 Register temp1_reg,
2976 Label* L_success,
2977 Label* L_failure,
2978 Label* L_slow_path,
2979 RegisterOrConstant super_check_offset) {
2980
2981 const int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
2982 const int sco_offset = in_bytes(Klass::super_check_offset_offset());
2983
2984 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
2985 bool need_slow_path = (must_load_sco ||
2986 super_check_offset.constant_or_zero() == sc_offset);
2987
2988 // Input registers must not overlap.
2989 assert_different_registers(sub_klass, super_klass, temp1_reg);
2990 if (super_check_offset.is_register()) {
2991 assert_different_registers(sub_klass, super_klass,
2992 super_check_offset.as_register());
2993 } else if (must_load_sco) {
2994 assert(temp1_reg != noreg, "supply either a temp or a register offset");
2995 }
2996
2997 const Register Rsuper_check_offset = temp1_reg;
2998
2999 NearLabel L_fallthrough;
3000 int label_nulls = 0;
3001 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
3002 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
3003 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
3004 assert(label_nulls <= 1 ||
3005 (L_slow_path == &L_fallthrough && label_nulls <= 2 && !need_slow_path),
3006 "at most one NULL in the batch, usually");
3007
3008 BLOCK_COMMENT("check_klass_subtype_fast_path {");
3009 // If the pointers are equal, we are done (e.g., String[] elements).
3010 // This self-check enables sharing of secondary supertype arrays among
3011 // non-primary types such as array-of-interface. Otherwise, each such
3012 // type would need its own customized SSA.
3013 // We move this check to the front of the fast path because many
3014 // type checks are in fact trivially successful in this manner,
3015 // so we get a nicely predicted branch right at the start of the check.
3016 compare64_and_branch(sub_klass, super_klass, bcondEqual, *L_success);
3017
3018 // Check the supertype display, which is uint.
3019 if (must_load_sco) {
3020 z_llgf(Rsuper_check_offset, sco_offset, super_klass);
3021 super_check_offset = RegisterOrConstant(Rsuper_check_offset);
3022 }
3023 Address super_check_addr(sub_klass, super_check_offset, 0);
3024 z_cg(super_klass, super_check_addr); // compare w/ displayed supertype
3025
3026 // This check has worked decisively for primary supers.
3027 // Secondary supers are sought in the super_cache ('super_cache_addr').
3028 // (Secondary supers are interfaces and very deeply nested subtypes.)
3029 // This works in the same check above because of a tricky aliasing
3030 // between the super_cache and the primary super display elements.
3031 // (The 'super_check_addr' can address either, as the case requires.)
3032 // Note that the cache is updated below if it does not help us find
3033 // what we need immediately.
3034 // So if it was a primary super, we can just fail immediately.
3035 // Otherwise, it's the slow path for us (no success at this point).
3036
3037 // Hacked jmp, which may only be used just before L_fallthrough.
3038 #define final_jmp(label) \
3039 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
3040 else { branch_optimized(Assembler::bcondAlways, label); } /*omit semicolon*/
3041
3042 if (super_check_offset.is_register()) {
3043 branch_optimized(Assembler::bcondEqual, *L_success);
3044 z_cfi(super_check_offset.as_register(), sc_offset);
3045 if (L_failure == &L_fallthrough) {
3046 branch_optimized(Assembler::bcondEqual, *L_slow_path);
3047 } else {
3048 branch_optimized(Assembler::bcondNotEqual, *L_failure);
3049 final_jmp(*L_slow_path);
3050 }
3051 } else if (super_check_offset.as_constant() == sc_offset) {
3052 // Need a slow path; fast failure is impossible.
3053 if (L_slow_path == &L_fallthrough) {
3054 branch_optimized(Assembler::bcondEqual, *L_success);
3055 } else {
3056 branch_optimized(Assembler::bcondNotEqual, *L_slow_path);
3057 final_jmp(*L_success);
3058 }
3059 } else {
3060 // No slow path; it's a fast decision.
3061 if (L_failure == &L_fallthrough) {
3062 branch_optimized(Assembler::bcondEqual, *L_success);
3063 } else {
3064 branch_optimized(Assembler::bcondNotEqual, *L_failure);
3065 final_jmp(*L_success);
3066 }
3067 }
3068
3069 bind(L_fallthrough);
3070 #undef local_brc
3071 #undef final_jmp
3072 BLOCK_COMMENT("} check_klass_subtype_fast_path");
3073 // fallthru (to slow path)
3074 }
3075
3076 void MacroAssembler::check_klass_subtype_slow_path(Register Rsubklass,
3077 Register Rsuperklass,
3078 Register Rarray_ptr, // tmp
3079 Register Rlength, // tmp
3080 Label* L_success,
3081 Label* L_failure) {
3082 // Input registers must not overlap.
3083 // Also check for R1 which is explicitely used here.
3084 assert_different_registers(Z_R1, Rsubklass, Rsuperklass, Rarray_ptr, Rlength);
3085 NearLabel L_fallthrough, L_loop;
3086 int label_nulls = 0;
3087 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
3088 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
3089 assert(label_nulls <= 1, "at most one NULL in the batch");
3090
3091 const int ss_offset = in_bytes(Klass::secondary_supers_offset());
3092 const int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
3093
3094 const int length_offset = Array<Klass*>::length_offset_in_bytes();
3095 const int base_offset = Array<Klass*>::base_offset_in_bytes();
3096
3097 // Hacked jmp, which may only be used just before L_fallthrough.
3098 #define final_jmp(label) \
3099 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
3100 else branch_optimized(Assembler::bcondAlways, label) /*omit semicolon*/
3101
3102 NearLabel loop_iterate, loop_count, match;
3103
3104 BLOCK_COMMENT("check_klass_subtype_slow_path {");
3105 z_lg(Rarray_ptr, ss_offset, Rsubklass);
3106
3107 load_and_test_int(Rlength, Address(Rarray_ptr, length_offset));
3108 branch_optimized(Assembler::bcondZero, *L_failure);
3109
3110 // Oops in table are NO MORE compressed.
3111 z_cg(Rsuperklass, base_offset, Rarray_ptr); // Check array element for match.
3112 z_bre(match); // Shortcut for array length = 1.
3113
3114 // No match yet, so we must walk the array's elements.
3115 z_lngfr(Rlength, Rlength);
3116 z_sllg(Rlength, Rlength, LogBytesPerWord); // -#bytes of cache array
3117 z_llill(Z_R1, BytesPerWord); // Set increment/end index.
3118 add2reg(Rlength, 2 * BytesPerWord); // start index = -(n-2)*BytesPerWord
3119 z_slgr(Rarray_ptr, Rlength); // start addr: += (n-2)*BytesPerWord
3120 z_bru(loop_count);
3121
3122 BIND(loop_iterate);
3123 z_cg(Rsuperklass, base_offset, Rlength, Rarray_ptr); // Check array element for match.
3124 z_bre(match);
3125 BIND(loop_count);
3126 z_brxlg(Rlength, Z_R1, loop_iterate);
3127
3128 // Rsuperklass not found among secondary super classes -> failure.
3129 branch_optimized(Assembler::bcondAlways, *L_failure);
3130
3131 // Got a hit. Return success (zero result). Set cache.
3132 // Cache load doesn't happen here. For speed it is directly emitted by the compiler.
3133
3134 BIND(match);
3135
3136 z_stg(Rsuperklass, sc_offset, Rsubklass); // Save result to cache.
3137
3138 final_jmp(*L_success);
3139
3140 // Exit to the surrounding code.
3141 BIND(L_fallthrough);
3142 #undef local_brc
3143 #undef final_jmp
3144 BLOCK_COMMENT("} check_klass_subtype_slow_path");
3145 }
3146
3147 // Emitter for combining fast and slow path.
3148 void MacroAssembler::check_klass_subtype(Register sub_klass,
3149 Register super_klass,
3150 Register temp1_reg,
3151 Register temp2_reg,
3152 Label& L_success) {
3153 NearLabel failure;
3154 BLOCK_COMMENT(err_msg("check_klass_subtype(%s subclass of %s) {", sub_klass->name(), super_klass->name()));
3155 check_klass_subtype_fast_path(sub_klass, super_klass, temp1_reg,
3156 &L_success, &failure, NULL);
3157 check_klass_subtype_slow_path(sub_klass, super_klass,
3158 temp1_reg, temp2_reg, &L_success, NULL);
3159 BIND(failure);
3160 BLOCK_COMMENT("} check_klass_subtype");
3161 }
3162
3163 // Increment a counter at counter_address when the eq condition code is
3164 // set. Kills registers tmp1_reg and tmp2_reg and preserves the condition code.
3165 void MacroAssembler::increment_counter_eq(address counter_address, Register tmp1_reg, Register tmp2_reg) {
3166 Label l;
3167 z_brne(l);
3168 load_const(tmp1_reg, counter_address);
3169 add2mem_32(Address(tmp1_reg), 1, tmp2_reg);
3170 z_cr(tmp1_reg, tmp1_reg); // Set cc to eq.
3171 bind(l);
3172 }
3173
3174 // Semantics are dependent on the slow_case label:
3175 // If the slow_case label is not NULL, failure to biased-lock the object
3176 // transfers control to the location of the slow_case label. If the
3177 // object could be biased-locked, control is transferred to the done label.
3178 // The condition code is unpredictable.
3179 //
3180 // If the slow_case label is NULL, failure to biased-lock the object results
3181 // in a transfer of control to the done label with a condition code of not_equal.
3182 // If the biased-lock could be successfully obtained, control is transfered to
3183 // the done label with a condition code of equal.
3184 // It is mandatory to react on the condition code At the done label.
3185 //
3186 void MacroAssembler::biased_locking_enter(Register obj_reg,
3187 Register mark_reg,
3188 Register temp_reg,
3189 Register temp2_reg, // May be Z_RO!
3190 Label &done,
3191 Label *slow_case) {
3192 assert(UseBiasedLocking, "why call this otherwise?");
3193 assert_different_registers(obj_reg, mark_reg, temp_reg, temp2_reg);
3194
3195 Label cas_label; // Try, if implemented, CAS locking. Fall thru to slow path otherwise.
3196
3197 BLOCK_COMMENT("biased_locking_enter {");
3198
3199 // Biased locking
3200 // See whether the lock is currently biased toward our thread and
3201 // whether the epoch is still valid.
3202 // Note that the runtime guarantees sufficient alignment of JavaThread
3203 // pointers to allow age to be placed into low bits.
3204 assert(markOopDesc::age_shift == markOopDesc::lock_bits + markOopDesc::biased_lock_bits,
3205 "biased locking makes assumptions about bit layout");
3206 z_lr(temp_reg, mark_reg);
3207 z_nilf(temp_reg, markOopDesc::biased_lock_mask_in_place);
3208 z_chi(temp_reg, markOopDesc::biased_lock_pattern);
3209 z_brne(cas_label); // Try cas if object is not biased, i.e. cannot be biased locked.
3210
3211 load_prototype_header(temp_reg, obj_reg);
3212 load_const_optimized(temp2_reg, ~((int) markOopDesc::age_mask_in_place));
3213
3214 z_ogr(temp_reg, Z_thread);
3215 z_xgr(temp_reg, mark_reg);
3216 z_ngr(temp_reg, temp2_reg);
3217 if (PrintBiasedLockingStatistics) {
3218 increment_counter_eq((address) BiasedLocking::biased_lock_entry_count_addr(), mark_reg, temp2_reg);
3219 // Restore mark_reg.
3220 z_lg(mark_reg, oopDesc::mark_offset_in_bytes(), obj_reg);
3221 }
3222 branch_optimized(Assembler::bcondEqual, done); // Biased lock obtained, return success.
3223
3224 Label try_revoke_bias;
3225 Label try_rebias;
3226 Address mark_addr = Address(obj_reg, oopDesc::mark_offset_in_bytes());
3227
3228 //----------------------------------------------------------------------------
3229 // At this point we know that the header has the bias pattern and
3230 // that we are not the bias owner in the current epoch. We need to
3231 // figure out more details about the state of the header in order to
3232 // know what operations can be legally performed on the object's
3233 // header.
3234
3235 // If the low three bits in the xor result aren't clear, that means
3236 // the prototype header is no longer biased and we have to revoke
3237 // the bias on this object.
3238 z_tmll(temp_reg, markOopDesc::biased_lock_mask_in_place);
3239 z_brnaz(try_revoke_bias);
3240
3241 // Biasing is still enabled for this data type. See whether the
3242 // epoch of the current bias is still valid, meaning that the epoch
3243 // bits of the mark word are equal to the epoch bits of the
3244 // prototype header. (Note that the prototype header's epoch bits
3245 // only change at a safepoint.) If not, attempt to rebias the object
3246 // toward the current thread. Note that we must be absolutely sure
3247 // that the current epoch is invalid in order to do this because
3248 // otherwise the manipulations it performs on the mark word are
3249 // illegal.
3250 z_tmll(temp_reg, markOopDesc::epoch_mask_in_place);
3251 z_brnaz(try_rebias);
3252
3253 //----------------------------------------------------------------------------
3254 // The epoch of the current bias is still valid but we know nothing
3255 // about the owner; it might be set or it might be clear. Try to
3256 // acquire the bias of the object using an atomic operation. If this
3257 // fails we will go in to the runtime to revoke the object's bias.
3258 // Note that we first construct the presumed unbiased header so we
3259 // don't accidentally blow away another thread's valid bias.
3260 z_nilf(mark_reg, markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place |
3261 markOopDesc::epoch_mask_in_place);
3262 z_lgr(temp_reg, Z_thread);
3263 z_llgfr(mark_reg, mark_reg);
3264 z_ogr(temp_reg, mark_reg);
3265
3266 assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0");
3267
3268 z_csg(mark_reg, temp_reg, 0, obj_reg);
3269
3270 // If the biasing toward our thread failed, this means that
3271 // another thread succeeded in biasing it toward itself and we
3272 // need to revoke that bias. The revocation will occur in the
3273 // interpreter runtime in the slow case.
3274
3275 if (PrintBiasedLockingStatistics) {
3276 increment_counter_eq((address) BiasedLocking::anonymously_biased_lock_entry_count_addr(),
3277 temp_reg, temp2_reg);
3278 }
3279 if (slow_case != NULL) {
3280 branch_optimized(Assembler::bcondNotEqual, *slow_case); // Biased lock not obtained, need to go the long way.
3281 }
3282 branch_optimized(Assembler::bcondAlways, done); // Biased lock status given in condition code.
3283
3284 //----------------------------------------------------------------------------
3285 bind(try_rebias);
3286 // At this point we know the epoch has expired, meaning that the
3287 // current "bias owner", if any, is actually invalid. Under these
3288 // circumstances _only_, we are allowed to use the current header's
3289 // value as the comparison value when doing the cas to acquire the
3290 // bias in the current epoch. In other words, we allow transfer of
3291 // the bias from one thread to another directly in this situation.
3292
3293 z_nilf(mark_reg, markOopDesc::biased_lock_mask_in_place | markOopDesc::age_mask_in_place | markOopDesc::epoch_mask_in_place);
3294 load_prototype_header(temp_reg, obj_reg);
3295 z_llgfr(mark_reg, mark_reg);
3296
3297 z_ogr(temp_reg, Z_thread);
3298
3299 assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0");
3300
3301 z_csg(mark_reg, temp_reg, 0, obj_reg);
3302
3303 // If the biasing toward our thread failed, this means that
3304 // another thread succeeded in biasing it toward itself and we
3305 // need to revoke that bias. The revocation will occur in the
3306 // interpreter runtime in the slow case.
3307
3308 if (PrintBiasedLockingStatistics) {
3309 increment_counter_eq((address) BiasedLocking::rebiased_lock_entry_count_addr(), temp_reg, temp2_reg);
3310 }
3311 if (slow_case != NULL) {
3312 branch_optimized(Assembler::bcondNotEqual, *slow_case); // Biased lock not obtained, need to go the long way.
3313 }
3314 z_bru(done); // Biased lock status given in condition code.
3315
3316 //----------------------------------------------------------------------------
3317 bind(try_revoke_bias);
3318 // The prototype mark in the klass doesn't have the bias bit set any
3319 // more, indicating that objects of this data type are not supposed
3320 // to be biased any more. We are going to try to reset the mark of
3321 // this object to the prototype value and fall through to the
3322 // CAS-based locking scheme. Note that if our CAS fails, it means
3323 // that another thread raced us for the privilege of revoking the
3324 // bias of this particular object, so it's okay to continue in the
3325 // normal locking code.
3326 load_prototype_header(temp_reg, obj_reg);
3327
3328 assert(oopDesc::mark_offset_in_bytes() == 0, "offset of _mark is not 0");
3329
3330 z_csg(mark_reg, temp_reg, 0, obj_reg);
3331
3332 // Fall through to the normal CAS-based lock, because no matter what
3333 // the result of the above CAS, some thread must have succeeded in
3334 // removing the bias bit from the object's header.
3335 if (PrintBiasedLockingStatistics) {
3336 // z_cgr(mark_reg, temp2_reg);
3337 increment_counter_eq((address) BiasedLocking::revoked_lock_entry_count_addr(), temp_reg, temp2_reg);
3338 }
3339
3340 bind(cas_label);
3341 BLOCK_COMMENT("} biased_locking_enter");
3342 }
3343
3344 void MacroAssembler::biased_locking_exit(Register mark_addr, Register temp_reg, Label& done) {
3345 // Check for biased locking unlock case, which is a no-op
3346 // Note: we do not have to check the thread ID for two reasons.
3347 // First, the interpreter checks for IllegalMonitorStateException at
3348 // a higher level. Second, if the bias was revoked while we held the
3349 // lock, the object could not be rebiased toward another thread, so
3350 // the bias bit would be clear.
3351 BLOCK_COMMENT("biased_locking_exit {");
3352
3353 z_lg(temp_reg, 0, mark_addr);
3354 z_nilf(temp_reg, markOopDesc::biased_lock_mask_in_place);
3355
3356 z_chi(temp_reg, markOopDesc::biased_lock_pattern);
3357 z_bre(done);
3358 BLOCK_COMMENT("} biased_locking_exit");
3359 }
3360
3361 void MacroAssembler::compiler_fast_lock_object(Register oop, Register box, Register temp1, Register temp2, bool try_bias) {
3362 Register displacedHeader = temp1;
3363 Register currentHeader = temp1;
3364 Register temp = temp2;
3365 NearLabel done, object_has_monitor;
3366
3367 BLOCK_COMMENT("compiler_fast_lock_object {");
3368
3369 // Load markOop from oop into mark.
3370 z_lg(displacedHeader, 0, oop);
3371
3372 if (try_bias) {
3373 biased_locking_enter(oop, displacedHeader, temp, Z_R0, done);
3374 }
3375
3376 // Handle existing monitor.
3377 // The object has an existing monitor iff (mark & monitor_value) != 0.
3378 guarantee(Immediate::is_uimm16(markOopDesc::monitor_value), "must be half-word");
3379 z_lr(temp, displacedHeader);
3380 z_nill(temp, markOopDesc::monitor_value);
3381 z_brne(object_has_monitor);
3382
3383 // Set mark to markOop | markOopDesc::unlocked_value.
3384 z_oill(displacedHeader, markOopDesc::unlocked_value);
3385
3386 // Load Compare Value application register.
3387
3388 // Initialize the box (must happen before we update the object mark).
3389 z_stg(displacedHeader, BasicLock::displaced_header_offset_in_bytes(), box);
3390
3391 // Memory Fence (in cmpxchgd)
3392 // Compare object markOop with mark and if equal exchange scratch1 with object markOop.
3393
3394 // If the compare-and-swap succeeded, then we found an unlocked object and we
3395 // have now locked it.
3396 z_csg(displacedHeader, box, 0, oop);
3397 assert(currentHeader==displacedHeader, "must be same register"); // Identified two registers from z/Architecture.
3398 z_bre(done);
3399
3400 // We did not see an unlocked object so try the fast recursive case.
3401
3402 z_sgr(currentHeader, Z_SP);
3403 load_const_optimized(temp, (~(os::vm_page_size()-1) | markOopDesc::lock_mask_in_place));
3404
3405 z_ngr(currentHeader, temp);
3406 // z_brne(done);
3407 // z_release();
3408 z_stg(currentHeader/*==0 or not 0*/, BasicLock::displaced_header_offset_in_bytes(), box);
3409
3410 z_bru(done);
3411
3412 Register zero = temp;
3413 Register monitor_tagged = displacedHeader; // Tagged with markOopDesc::monitor_value.
3414 bind(object_has_monitor);
3415 // The object's monitor m is unlocked iff m->owner == NULL,
3416 // otherwise m->owner may contain a thread or a stack address.
3417 //
3418 // Try to CAS m->owner from NULL to current thread.
3419 z_lghi(zero, 0);
3420 // If m->owner is null, then csg succeeds and sets m->owner=THREAD and CR=EQ.
3421 z_csg(zero, Z_thread, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), monitor_tagged);
3422 // Store a non-null value into the box.
3423 z_stg(box, BasicLock::displaced_header_offset_in_bytes(), box);
3424 #ifdef ASSERT
3425 z_brne(done);
3426 // We've acquired the monitor, check some invariants.
3427 // Invariant 1: _recursions should be 0.
3428 asm_assert_mem8_is_zero(OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions), monitor_tagged,
3429 "monitor->_recursions should be 0", -1);
3430 z_ltgr(zero, zero); // Set CR=EQ.
3431 #endif
3432 bind(done);
3433
3434 BLOCK_COMMENT("} compiler_fast_lock_object");
3435 // If locking was successful, CR should indicate 'EQ'.
3436 // The compiler or the native wrapper generates a branch to the runtime call
3437 // _complete_monitor_locking_Java.
3438 }
3439
3440 void MacroAssembler::compiler_fast_unlock_object(Register oop, Register box, Register temp1, Register temp2, bool try_bias) {
3441 Register displacedHeader = temp1;
3442 Register currentHeader = temp2;
3443 Register temp = temp1;
3444 Register monitor = temp2;
3445
3446 Label done, object_has_monitor;
3447
3448 BLOCK_COMMENT("compiler_fast_unlock_object {");
3449
3450 if (try_bias) {
3451 biased_locking_exit(oop, currentHeader, done);
3452 }
3453
3454 // Find the lock address and load the displaced header from the stack.
3455 // if the displaced header is zero, we have a recursive unlock.
3456 load_and_test_long(displacedHeader, Address(box, BasicLock::displaced_header_offset_in_bytes()));
3457 z_bre(done);
3458
3459 // Handle existing monitor.
3460 // The object has an existing monitor iff (mark & monitor_value) != 0.
3461 z_lg(currentHeader, oopDesc::mark_offset_in_bytes(), oop);
3462 guarantee(Immediate::is_uimm16(markOopDesc::monitor_value), "must be half-word");
3463 z_nill(currentHeader, markOopDesc::monitor_value);
3464 z_brne(object_has_monitor);
3465
3466 // Check if it is still a light weight lock, this is true if we see
3467 // the stack address of the basicLock in the markOop of the object
3468 // copy box to currentHeader such that csg does not kill it.
3469 z_lgr(currentHeader, box);
3470 z_csg(currentHeader, displacedHeader, 0, oop);
3471 z_bru(done); // Csg sets CR as desired.
3472
3473 // Handle existing monitor.
3474 bind(object_has_monitor);
3475 z_lg(currentHeader, oopDesc::mark_offset_in_bytes(), oop); // CurrentHeader is tagged with monitor_value set.
3476 load_and_test_long(temp, Address(currentHeader, OM_OFFSET_NO_MONITOR_VALUE_TAG(recursions)));
3477 z_brne(done);
3478 load_and_test_long(temp, Address(currentHeader, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner)));
3479 z_brne(done);
3480 load_and_test_long(temp, Address(currentHeader, OM_OFFSET_NO_MONITOR_VALUE_TAG(EntryList)));
3481 z_brne(done);
3482 load_and_test_long(temp, Address(currentHeader, OM_OFFSET_NO_MONITOR_VALUE_TAG(cxq)));
3483 z_brne(done);
3484 z_release();
3485 z_stg(temp/*=0*/, OM_OFFSET_NO_MONITOR_VALUE_TAG(owner), currentHeader);
3486
3487 bind(done);
3488
3489 BLOCK_COMMENT("} compiler_fast_unlock_object");
3490 // flag == EQ indicates success
3491 // flag == NE indicates failure
3492 }
3493
3494 void MacroAssembler::resolve_jobject(Register value, Register tmp1, Register tmp2) {
3495 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
3496 bs->resolve_jobject(this, value, tmp1, tmp2);
3497 }
3498
3499 // Last_Java_sp must comply to the rules in frame_s390.hpp.
3500 void MacroAssembler::set_last_Java_frame(Register last_Java_sp, Register last_Java_pc, bool allow_relocation) {
3501 BLOCK_COMMENT("set_last_Java_frame {");
3502
3503 // Always set last_Java_pc and flags first because once last_Java_sp
3504 // is visible has_last_Java_frame is true and users will look at the
3505 // rest of the fields. (Note: flags should always be zero before we
3506 // get here so doesn't need to be set.)
3507
3508 // Verify that last_Java_pc was zeroed on return to Java.
3509 if (allow_relocation) {
3510 asm_assert_mem8_is_zero(in_bytes(JavaThread::last_Java_pc_offset()),
3511 Z_thread,
3512 "last_Java_pc not zeroed before leaving Java",
3513 0x200);
3514 } else {
3515 asm_assert_mem8_is_zero_static(in_bytes(JavaThread::last_Java_pc_offset()),
3516 Z_thread,
3517 "last_Java_pc not zeroed before leaving Java",
3518 0x200);
3519 }
3520
3521 // When returning from calling out from Java mode the frame anchor's
3522 // last_Java_pc will always be set to NULL. It is set here so that
3523 // if we are doing a call to native (not VM) that we capture the
3524 // known pc and don't have to rely on the native call having a
3525 // standard frame linkage where we can find the pc.
3526 if (last_Java_pc!=noreg) {
3527 z_stg(last_Java_pc, Address(Z_thread, JavaThread::last_Java_pc_offset()));
3528 }
3529
3530 // This membar release is not required on z/Architecture, since the sequence of stores
3531 // in maintained. Nevertheless, we leave it in to document the required ordering.
3532 // The implementation of z_release() should be empty.
3533 // z_release();
3534
3535 z_stg(last_Java_sp, Address(Z_thread, JavaThread::last_Java_sp_offset()));
3536 BLOCK_COMMENT("} set_last_Java_frame");
3537 }
3538
3539 void MacroAssembler::reset_last_Java_frame(bool allow_relocation) {
3540 BLOCK_COMMENT("reset_last_Java_frame {");
3541
3542 if (allow_relocation) {
3543 asm_assert_mem8_isnot_zero(in_bytes(JavaThread::last_Java_sp_offset()),
3544 Z_thread,
3545 "SP was not set, still zero",
3546 0x202);
3547 } else {
3548 asm_assert_mem8_isnot_zero_static(in_bytes(JavaThread::last_Java_sp_offset()),
3549 Z_thread,
3550 "SP was not set, still zero",
3551 0x202);
3552 }
3553
3554 // _last_Java_sp = 0
3555 // Clearing storage must be atomic here, so don't use clear_mem()!
3556 store_const(Address(Z_thread, JavaThread::last_Java_sp_offset()), 0);
3557
3558 // _last_Java_pc = 0
3559 store_const(Address(Z_thread, JavaThread::last_Java_pc_offset()), 0);
3560
3561 BLOCK_COMMENT("} reset_last_Java_frame");
3562 return;
3563 }
3564
3565 void MacroAssembler::set_top_ijava_frame_at_SP_as_last_Java_frame(Register sp, Register tmp1, bool allow_relocation) {
3566 assert_different_registers(sp, tmp1);
3567
3568 // We cannot trust that code generated by the C++ compiler saves R14
3569 // to z_abi_160.return_pc, because sometimes it spills R14 using stmg at
3570 // z_abi_160.gpr14 (e.g. InterpreterRuntime::_new()).
3571 // Therefore we load the PC into tmp1 and let set_last_Java_frame() save
3572 // it into the frame anchor.
3573 get_PC(tmp1);
3574 set_last_Java_frame(/*sp=*/sp, /*pc=*/tmp1, allow_relocation);
3575 }
3576
3577 void MacroAssembler::set_thread_state(JavaThreadState new_state) {
3578 z_release();
3579
3580 assert(Immediate::is_uimm16(_thread_max_state), "enum value out of range for instruction");
3581 assert(sizeof(JavaThreadState) == sizeof(int), "enum value must have base type int");
3582 store_const(Address(Z_thread, JavaThread::thread_state_offset()), new_state, Z_R0, false);
3583 }
3584
3585 void MacroAssembler::get_vm_result(Register oop_result) {
3586 verify_thread();
3587
3588 z_lg(oop_result, Address(Z_thread, JavaThread::vm_result_offset()));
3589 clear_mem(Address(Z_thread, JavaThread::vm_result_offset()), sizeof(void*));
3590
3591 verify_oop(oop_result);
3592 }
3593
3594 void MacroAssembler::get_vm_result_2(Register result) {
3595 verify_thread();
3596
3597 z_lg(result, Address(Z_thread, JavaThread::vm_result_2_offset()));
3598 clear_mem(Address(Z_thread, JavaThread::vm_result_2_offset()), sizeof(void*));
3599 }
3600
3601 // We require that C code which does not return a value in vm_result will
3602 // leave it undisturbed.
3603 void MacroAssembler::set_vm_result(Register oop_result) {
3604 z_stg(oop_result, Address(Z_thread, JavaThread::vm_result_offset()));
3605 }
3606
3607 // Explicit null checks (used for method handle code).
3608 void MacroAssembler::null_check(Register reg, Register tmp, int64_t offset) {
3609 if (!ImplicitNullChecks) {
3610 NearLabel ok;
3611
3612 compare64_and_branch(reg, (intptr_t) 0, Assembler::bcondNotEqual, ok);
3613
3614 // We just put the address into reg if it was 0 (tmp==Z_R0 is allowed so we can't use it for the address).
3615 address exception_entry = Interpreter::throw_NullPointerException_entry();
3616 load_absolute_address(reg, exception_entry);
3617 z_br(reg);
3618
3619 bind(ok);
3620 } else {
3621 if (needs_explicit_null_check((intptr_t)offset)) {
3622 // Provoke OS NULL exception if reg = NULL by
3623 // accessing M[reg] w/o changing any registers.
3624 z_lg(tmp, 0, reg);
3625 }
3626 // else
3627 // Nothing to do, (later) access of M[reg + offset]
3628 // will provoke OS NULL exception if reg = NULL.
3629 }
3630 }
3631
3632 //-------------------------------------
3633 // Compressed Klass Pointers
3634 //-------------------------------------
3635
3636 // Klass oop manipulations if compressed.
3637 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
3638 Register current = (src != noreg) ? src : dst; // Klass is in dst if no src provided. (dst == src) also possible.
3639 address base = Universe::narrow_klass_base();
3640 int shift = Universe::narrow_klass_shift();
3641 assert(UseCompressedClassPointers, "only for compressed klass ptrs");
3642
3643 BLOCK_COMMENT("cKlass encoder {");
3644
3645 #ifdef ASSERT
3646 Label ok;
3647 z_tmll(current, KlassAlignmentInBytes-1); // Check alignment.
3648 z_brc(Assembler::bcondAllZero, ok);
3649 // The plain disassembler does not recognize illtrap. It instead displays
3650 // a 32-bit value. Issueing two illtraps assures the disassembler finds
3651 // the proper beginning of the next instruction.
3652 z_illtrap(0xee);
3653 z_illtrap(0xee);
3654 bind(ok);
3655 #endif
3656
3657 if (base != NULL) {
3658 unsigned int base_h = ((unsigned long)base)>>32;
3659 unsigned int base_l = (unsigned int)((unsigned long)base);
3660 if ((base_h != 0) && (base_l == 0) && VM_Version::has_HighWordInstr()) {
3661 lgr_if_needed(dst, current);
3662 z_aih(dst, -((int)base_h)); // Base has no set bits in lower half.
3663 } else if ((base_h == 0) && (base_l != 0)) {
3664 lgr_if_needed(dst, current);
3665 z_agfi(dst, -(int)base_l);
3666 } else {
3667 load_const(Z_R0, base);
3668 lgr_if_needed(dst, current);
3669 z_sgr(dst, Z_R0);
3670 }
3671 current = dst;
3672 }
3673 if (shift != 0) {
3674 assert (LogKlassAlignmentInBytes == shift, "decode alg wrong");
3675 z_srlg(dst, current, shift);
3676 current = dst;
3677 }
3678 lgr_if_needed(dst, current); // Move may be required (if neither base nor shift != 0).
3679
3680 BLOCK_COMMENT("} cKlass encoder");
3681 }
3682
3683 // This function calculates the size of the code generated by
3684 // decode_klass_not_null(register dst, Register src)
3685 // when (Universe::heap() != NULL). Hence, if the instructions
3686 // it generates change, then this method needs to be updated.
3687 int MacroAssembler::instr_size_for_decode_klass_not_null() {
3688 address base = Universe::narrow_klass_base();
3689 int shift_size = Universe::narrow_klass_shift() == 0 ? 0 : 6; /* sllg */
3690 int addbase_size = 0;
3691 assert(UseCompressedClassPointers, "only for compressed klass ptrs");
3692
3693 if (base != NULL) {
3694 unsigned int base_h = ((unsigned long)base)>>32;
3695 unsigned int base_l = (unsigned int)((unsigned long)base);
3696 if ((base_h != 0) && (base_l == 0) && VM_Version::has_HighWordInstr()) {
3697 addbase_size += 6; /* aih */
3698 } else if ((base_h == 0) && (base_l != 0)) {
3699 addbase_size += 6; /* algfi */
3700 } else {
3701 addbase_size += load_const_size();
3702 addbase_size += 4; /* algr */
3703 }
3704 }
3705 #ifdef ASSERT
3706 addbase_size += 10;
3707 addbase_size += 2; // Extra sigill.
3708 #endif
3709 return addbase_size + shift_size;
3710 }
3711
3712 // !!! If the instructions that get generated here change
3713 // then function instr_size_for_decode_klass_not_null()
3714 // needs to get updated.
3715 // This variant of decode_klass_not_null() must generate predictable code!
3716 // The code must only depend on globally known parameters.
3717 void MacroAssembler::decode_klass_not_null(Register dst) {
3718 address base = Universe::narrow_klass_base();
3719 int shift = Universe::narrow_klass_shift();
3720 int beg_off = offset();
3721 assert(UseCompressedClassPointers, "only for compressed klass ptrs");
3722
3723 BLOCK_COMMENT("cKlass decoder (const size) {");
3724
3725 if (shift != 0) { // Shift required?
3726 z_sllg(dst, dst, shift);
3727 }
3728 if (base != NULL) {
3729 unsigned int base_h = ((unsigned long)base)>>32;
3730 unsigned int base_l = (unsigned int)((unsigned long)base);
3731 if ((base_h != 0) && (base_l == 0) && VM_Version::has_HighWordInstr()) {
3732 z_aih(dst, base_h); // Base has no set bits in lower half.
3733 } else if ((base_h == 0) && (base_l != 0)) {
3734 z_algfi(dst, base_l); // Base has no set bits in upper half.
3735 } else {
3736 load_const(Z_R0, base); // Base has set bits everywhere.
3737 z_algr(dst, Z_R0);
3738 }
3739 }
3740
3741 #ifdef ASSERT
3742 Label ok;
3743 z_tmll(dst, KlassAlignmentInBytes-1); // Check alignment.
3744 z_brc(Assembler::bcondAllZero, ok);
3745 // The plain disassembler does not recognize illtrap. It instead displays
3746 // a 32-bit value. Issueing two illtraps assures the disassembler finds
3747 // the proper beginning of the next instruction.
3748 z_illtrap(0xd1);
3749 z_illtrap(0xd1);
3750 bind(ok);
3751 #endif
3752 assert(offset() == beg_off + instr_size_for_decode_klass_not_null(), "Code gen mismatch.");
3753
3754 BLOCK_COMMENT("} cKlass decoder (const size)");
3755 }
3756
3757 // This variant of decode_klass_not_null() is for cases where
3758 // 1) the size of the generated instructions may vary
3759 // 2) the result is (potentially) stored in a register different from the source.
3760 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
3761 address base = Universe::narrow_klass_base();
3762 int shift = Universe::narrow_klass_shift();
3763 assert(UseCompressedClassPointers, "only for compressed klass ptrs");
3764
3765 BLOCK_COMMENT("cKlass decoder {");
3766
3767 if (src == noreg) src = dst;
3768
3769 if (shift != 0) { // Shift or at least move required?
3770 z_sllg(dst, src, shift);
3771 } else {
3772 lgr_if_needed(dst, src);
3773 }
3774
3775 if (base != NULL) {
3776 unsigned int base_h = ((unsigned long)base)>>32;
3777 unsigned int base_l = (unsigned int)((unsigned long)base);
3778 if ((base_h != 0) && (base_l == 0) && VM_Version::has_HighWordInstr()) {
3779 z_aih(dst, base_h); // Base has not set bits in lower half.
3780 } else if ((base_h == 0) && (base_l != 0)) {
3781 z_algfi(dst, base_l); // Base has no set bits in upper half.
3782 } else {
3783 load_const_optimized(Z_R0, base); // Base has set bits everywhere.
3784 z_algr(dst, Z_R0);
3785 }
3786 }
3787
3788 #ifdef ASSERT
3789 Label ok;
3790 z_tmll(dst, KlassAlignmentInBytes-1); // Check alignment.
3791 z_brc(Assembler::bcondAllZero, ok);
3792 // The plain disassembler does not recognize illtrap. It instead displays
3793 // a 32-bit value. Issueing two illtraps assures the disassembler finds
3794 // the proper beginning of the next instruction.
3795 z_illtrap(0xd2);
3796 z_illtrap(0xd2);
3797 bind(ok);
3798 #endif
3799 BLOCK_COMMENT("} cKlass decoder");
3800 }
3801
3802 void MacroAssembler::load_klass(Register klass, Address mem) {
3803 if (UseCompressedClassPointers) {
3804 z_llgf(klass, mem);
3805 // Attention: no null check here!
3806 decode_klass_not_null(klass);
3807 } else {
3808 z_lg(klass, mem);
3809 }
3810 }
3811
3812 void MacroAssembler::load_klass(Register klass, Register src_oop) {
3813 if (UseCompressedClassPointers) {
3814 z_llgf(klass, oopDesc::klass_offset_in_bytes(), src_oop);
3815 // Attention: no null check here!
3816 decode_klass_not_null(klass);
3817 } else {
3818 z_lg(klass, oopDesc::klass_offset_in_bytes(), src_oop);
3819 }
3820 }
3821
3822 void MacroAssembler::load_prototype_header(Register Rheader, Register Rsrc_oop) {
3823 assert_different_registers(Rheader, Rsrc_oop);
3824 load_klass(Rheader, Rsrc_oop);
3825 z_lg(Rheader, Address(Rheader, Klass::prototype_header_offset()));
3826 }
3827
3828 void MacroAssembler::store_klass(Register klass, Register dst_oop, Register ck) {
3829 if (UseCompressedClassPointers) {
3830 assert_different_registers(dst_oop, klass, Z_R0);
3831 if (ck == noreg) ck = klass;
3832 encode_klass_not_null(ck, klass);
3833 z_st(ck, Address(dst_oop, oopDesc::klass_offset_in_bytes()));
3834 } else {
3835 z_stg(klass, Address(dst_oop, oopDesc::klass_offset_in_bytes()));
3836 }
3837 }
3838
3839 void MacroAssembler::store_klass_gap(Register s, Register d) {
3840 if (UseCompressedClassPointers) {
3841 assert(s != d, "not enough registers");
3842 // Support s = noreg.
3843 if (s != noreg) {
3844 z_st(s, Address(d, oopDesc::klass_gap_offset_in_bytes()));
3845 } else {
3846 z_mvhi(Address(d, oopDesc::klass_gap_offset_in_bytes()), 0);
3847 }
3848 }
3849 }
3850
3851 // Compare klass ptr in memory against klass ptr in register.
3852 //
3853 // Rop1 - klass in register, always uncompressed.
3854 // disp - Offset of klass in memory, compressed/uncompressed, depending on runtime flag.
3855 // Rbase - Base address of cKlass in memory.
3856 // maybeNULL - True if Rop1 possibly is a NULL.
3857 void MacroAssembler::compare_klass_ptr(Register Rop1, int64_t disp, Register Rbase, bool maybeNULL) {
3858
3859 BLOCK_COMMENT("compare klass ptr {");
3860
3861 if (UseCompressedClassPointers) {
3862 const int shift = Universe::narrow_klass_shift();
3863 address base = Universe::narrow_klass_base();
3864
3865 assert((shift == 0) || (shift == LogKlassAlignmentInBytes), "cKlass encoder detected bad shift");
3866 assert_different_registers(Rop1, Z_R0);
3867 assert_different_registers(Rop1, Rbase, Z_R1);
3868
3869 // First encode register oop and then compare with cOop in memory.
3870 // This sequence saves an unnecessary cOop load and decode.
3871 if (base == NULL) {
3872 if (shift == 0) {
3873 z_cl(Rop1, disp, Rbase); // Unscaled
3874 } else {
3875 z_srlg(Z_R0, Rop1, shift); // ZeroBased
3876 z_cl(Z_R0, disp, Rbase);
3877 }
3878 } else { // HeapBased
3879 #ifdef ASSERT
3880 bool used_R0 = true;
3881 bool used_R1 = true;
3882 #endif
3883 Register current = Rop1;
3884 Label done;
3885
3886 if (maybeNULL) { // NULL ptr must be preserved!
3887 z_ltgr(Z_R0, current);
3888 z_bre(done);
3889 current = Z_R0;
3890 }
3891
3892 unsigned int base_h = ((unsigned long)base)>>32;
3893 unsigned int base_l = (unsigned int)((unsigned long)base);
3894 if ((base_h != 0) && (base_l == 0) && VM_Version::has_HighWordInstr()) {
3895 lgr_if_needed(Z_R0, current);
3896 z_aih(Z_R0, -((int)base_h)); // Base has no set bits in lower half.
3897 } else if ((base_h == 0) && (base_l != 0)) {
3898 lgr_if_needed(Z_R0, current);
3899 z_agfi(Z_R0, -(int)base_l);
3900 } else {
3901 int pow2_offset = get_oop_base_complement(Z_R1, ((uint64_t)(intptr_t)base));
3902 add2reg_with_index(Z_R0, pow2_offset, Z_R1, Rop1); // Subtract base by adding complement.
3903 }
3904
3905 if (shift != 0) {
3906 z_srlg(Z_R0, Z_R0, shift);
3907 }
3908 bind(done);
3909 z_cl(Z_R0, disp, Rbase);
3910 #ifdef ASSERT
3911 if (used_R0) preset_reg(Z_R0, 0xb05bUL, 2);
3912 if (used_R1) preset_reg(Z_R1, 0xb06bUL, 2);
3913 #endif
3914 }
3915 } else {
3916 z_clg(Rop1, disp, Z_R0, Rbase);
3917 }
3918 BLOCK_COMMENT("} compare klass ptr");
3919 }
3920
3921 //---------------------------
3922 // Compressed oops
3923 //---------------------------
3924
3925 void MacroAssembler::encode_heap_oop(Register oop) {
3926 oop_encoder(oop, oop, true /*maybe null*/);
3927 }
3928
3929 void MacroAssembler::encode_heap_oop_not_null(Register oop) {
3930 oop_encoder(oop, oop, false /*not null*/);
3931 }
3932
3933 // Called with something derived from the oop base. e.g. oop_base>>3.
3934 int MacroAssembler::get_oop_base_pow2_offset(uint64_t oop_base) {
3935 unsigned int oop_base_ll = ((unsigned int)(oop_base >> 0)) & 0xffff;
3936 unsigned int oop_base_lh = ((unsigned int)(oop_base >> 16)) & 0xffff;
3937 unsigned int oop_base_hl = ((unsigned int)(oop_base >> 32)) & 0xffff;
3938 unsigned int oop_base_hh = ((unsigned int)(oop_base >> 48)) & 0xffff;
3939 unsigned int n_notzero_parts = (oop_base_ll == 0 ? 0:1)
3940 + (oop_base_lh == 0 ? 0:1)
3941 + (oop_base_hl == 0 ? 0:1)
3942 + (oop_base_hh == 0 ? 0:1);
3943
3944 assert(oop_base != 0, "This is for HeapBased cOops only");
3945
3946 if (n_notzero_parts != 1) { // Check if oop_base is just a few pages shy of a power of 2.
3947 uint64_t pow2_offset = 0x10000 - oop_base_ll;
3948 if (pow2_offset < 0x8000) { // This might not be necessary.
3949 uint64_t oop_base2 = oop_base + pow2_offset;
3950
3951 oop_base_ll = ((unsigned int)(oop_base2 >> 0)) & 0xffff;
3952 oop_base_lh = ((unsigned int)(oop_base2 >> 16)) & 0xffff;
3953 oop_base_hl = ((unsigned int)(oop_base2 >> 32)) & 0xffff;
3954 oop_base_hh = ((unsigned int)(oop_base2 >> 48)) & 0xffff;
3955 n_notzero_parts = (oop_base_ll == 0 ? 0:1) +
3956 (oop_base_lh == 0 ? 0:1) +
3957 (oop_base_hl == 0 ? 0:1) +
3958 (oop_base_hh == 0 ? 0:1);
3959 if (n_notzero_parts == 1) {
3960 assert(-(int64_t)pow2_offset != (int64_t)-1, "We use -1 to signal uninitialized base register");
3961 return -pow2_offset;
3962 }
3963 }
3964 }
3965 return 0;
3966 }
3967
3968 // If base address is offset from a straight power of two by just a few pages,
3969 // return this offset to the caller for a possible later composite add.
3970 // TODO/FIX: will only work correctly for 4k pages.
3971 int MacroAssembler::get_oop_base(Register Rbase, uint64_t oop_base) {
3972 int pow2_offset = get_oop_base_pow2_offset(oop_base);
3973
3974 load_const_optimized(Rbase, oop_base - pow2_offset); // Best job possible.
3975
3976 return pow2_offset;
3977 }
3978
3979 int MacroAssembler::get_oop_base_complement(Register Rbase, uint64_t oop_base) {
3980 int offset = get_oop_base(Rbase, oop_base);
3981 z_lcgr(Rbase, Rbase);
3982 return -offset;
3983 }
3984
3985 // Compare compressed oop in memory against oop in register.
3986 // Rop1 - Oop in register.
3987 // disp - Offset of cOop in memory.
3988 // Rbase - Base address of cOop in memory.
3989 // maybeNULL - True if Rop1 possibly is a NULL.
3990 // maybeNULLtarget - Branch target for Rop1 == NULL, if flow control shall NOT continue with compare instruction.
3991 void MacroAssembler::compare_heap_oop(Register Rop1, Address mem, bool maybeNULL) {
3992 Register Rbase = mem.baseOrR0();
3993 Register Rindex = mem.indexOrR0();
3994 int64_t disp = mem.disp();
3995
3996 const int shift = Universe::narrow_oop_shift();
3997 address base = Universe::narrow_oop_base();
3998
3999 assert(UseCompressedOops, "must be on to call this method");
4000 assert(Universe::heap() != NULL, "java heap must be initialized to call this method");
4001 assert((shift == 0) || (shift == LogMinObjAlignmentInBytes), "cOop encoder detected bad shift");
4002 assert_different_registers(Rop1, Z_R0);
4003 assert_different_registers(Rop1, Rbase, Z_R1);
4004 assert_different_registers(Rop1, Rindex, Z_R1);
4005
4006 BLOCK_COMMENT("compare heap oop {");
4007
4008 // First encode register oop and then compare with cOop in memory.
4009 // This sequence saves an unnecessary cOop load and decode.
4010 if (base == NULL) {
4011 if (shift == 0) {
4012 z_cl(Rop1, disp, Rindex, Rbase); // Unscaled
4013 } else {
4014 z_srlg(Z_R0, Rop1, shift); // ZeroBased
4015 z_cl(Z_R0, disp, Rindex, Rbase);
4016 }
4017 } else { // HeapBased
4018 #ifdef ASSERT
4019 bool used_R0 = true;
4020 bool used_R1 = true;
4021 #endif
4022 Label done;
4023 int pow2_offset = get_oop_base_complement(Z_R1, ((uint64_t)(intptr_t)base));
4024
4025 if (maybeNULL) { // NULL ptr must be preserved!
4026 z_ltgr(Z_R0, Rop1);
4027 z_bre(done);
4028 }
4029
4030 add2reg_with_index(Z_R0, pow2_offset, Z_R1, Rop1);
4031 z_srlg(Z_R0, Z_R0, shift);
4032
4033 bind(done);
4034 z_cl(Z_R0, disp, Rindex, Rbase);
4035 #ifdef ASSERT
4036 if (used_R0) preset_reg(Z_R0, 0xb05bUL, 2);
4037 if (used_R1) preset_reg(Z_R1, 0xb06bUL, 2);
4038 #endif
4039 }
4040 BLOCK_COMMENT("} compare heap oop");
4041 }
4042
4043 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators,
4044 const Address& addr, Register val,
4045 Register tmp1, Register tmp2, Register tmp3) {
4046 assert((decorators & ~(AS_RAW | IN_HEAP | IN_NATIVE | IS_ARRAY | IS_NOT_NULL |
4047 ON_UNKNOWN_OOP_REF)) == 0, "unsupported decorator");
4048 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4049 decorators = AccessInternal::decorator_fixup(decorators);
4050 bool as_raw = (decorators & AS_RAW) != 0;
4051 if (as_raw) {
4052 bs->BarrierSetAssembler::store_at(this, decorators, type,
4053 addr, val,
4054 tmp1, tmp2, tmp3);
4055 } else {
4056 bs->store_at(this, decorators, type,
4057 addr, val,
4058 tmp1, tmp2, tmp3);
4059 }
4060 }
4061
4062 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators,
4063 const Address& addr, Register dst,
4064 Register tmp1, Register tmp2, Label *is_null) {
4065 assert((decorators & ~(AS_RAW | IN_HEAP | IN_NATIVE | IS_ARRAY | IS_NOT_NULL |
4066 ON_PHANTOM_OOP_REF | ON_WEAK_OOP_REF)) == 0, "unsupported decorator");
4067 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4068 decorators = AccessInternal::decorator_fixup(decorators);
4069 bool as_raw = (decorators & AS_RAW) != 0;
4070 if (as_raw) {
4071 bs->BarrierSetAssembler::load_at(this, decorators, type,
4072 addr, dst,
4073 tmp1, tmp2, is_null);
4074 } else {
4075 bs->load_at(this, decorators, type,
4076 addr, dst,
4077 tmp1, tmp2, is_null);
4078 }
4079 }
4080
4081 void MacroAssembler::load_heap_oop(Register dest, const Address &a,
4082 Register tmp1, Register tmp2,
4083 DecoratorSet decorators, Label *is_null) {
4084 access_load_at(T_OBJECT, IN_HEAP | decorators, a, dest, tmp1, tmp2, is_null);
4085 }
4086
4087 void MacroAssembler::store_heap_oop(Register Roop, const Address &a,
4088 Register tmp1, Register tmp2, Register tmp3,
4089 DecoratorSet decorators) {
4090 access_store_at(T_OBJECT, IN_HEAP | decorators, a, Roop, tmp1, tmp2, tmp3);
4091 }
4092
4093 //-------------------------------------------------
4094 // Encode compressed oop. Generally usable encoder.
4095 //-------------------------------------------------
4096 // Rsrc - contains regular oop on entry. It remains unchanged.
4097 // Rdst - contains compressed oop on exit.
4098 // Rdst and Rsrc may indicate same register, in which case Rsrc does not remain unchanged.
4099 //
4100 // Rdst must not indicate scratch register Z_R1 (Z_R1_scratch) for functionality.
4101 // Rdst should not indicate scratch register Z_R0 (Z_R0_scratch) for performance.
4102 //
4103 // only32bitValid is set, if later code only uses the lower 32 bits. In this
4104 // case we must not fix the upper 32 bits.
4105 void MacroAssembler::oop_encoder(Register Rdst, Register Rsrc, bool maybeNULL,
4106 Register Rbase, int pow2_offset, bool only32bitValid) {
4107
4108 const address oop_base = Universe::narrow_oop_base();
4109 const int oop_shift = Universe::narrow_oop_shift();
4110 const bool disjoint = Universe::narrow_oop_base_disjoint();
4111
4112 assert(UseCompressedOops, "must be on to call this method");
4113 assert(Universe::heap() != NULL, "java heap must be initialized to call this encoder");
4114 assert((oop_shift == 0) || (oop_shift == LogMinObjAlignmentInBytes), "cOop encoder detected bad shift");
4115
4116 if (disjoint || (oop_base == NULL)) {
4117 BLOCK_COMMENT("cOop encoder zeroBase {");
4118 if (oop_shift == 0) {
4119 if (oop_base != NULL && !only32bitValid) {
4120 z_llgfr(Rdst, Rsrc); // Clear upper bits in case the register will be decoded again.
4121 } else {
4122 lgr_if_needed(Rdst, Rsrc);
4123 }
4124 } else {
4125 z_srlg(Rdst, Rsrc, oop_shift);
4126 if (oop_base != NULL && !only32bitValid) {
4127 z_llgfr(Rdst, Rdst); // Clear upper bits in case the register will be decoded again.
4128 }
4129 }
4130 BLOCK_COMMENT("} cOop encoder zeroBase");
4131 return;
4132 }
4133
4134 bool used_R0 = false;
4135 bool used_R1 = false;
4136
4137 BLOCK_COMMENT("cOop encoder general {");
4138 assert_different_registers(Rdst, Z_R1);
4139 assert_different_registers(Rsrc, Rbase);
4140 if (maybeNULL) {
4141 Label done;
4142 // We reorder shifting and subtracting, so that we can compare
4143 // and shift in parallel:
4144 //
4145 // cycle 0: potential LoadN, base = <const>
4146 // cycle 1: base = !base dst = src >> 3, cmp cr = (src != 0)
4147 // cycle 2: if (cr) br, dst = dst + base + offset
4148
4149 // Get oop_base components.
4150 if (pow2_offset == -1) {
4151 if (Rdst == Rbase) {
4152 if (Rdst == Z_R1 || Rsrc == Z_R1) {
4153 Rbase = Z_R0;
4154 used_R0 = true;
4155 } else {
4156 Rdst = Z_R1;
4157 used_R1 = true;
4158 }
4159 }
4160 if (Rbase == Z_R1) {
4161 used_R1 = true;
4162 }
4163 pow2_offset = get_oop_base_complement(Rbase, ((uint64_t)(intptr_t)oop_base) >> oop_shift);
4164 }
4165 assert_different_registers(Rdst, Rbase);
4166
4167 // Check for NULL oop (must be left alone) and shift.
4168 if (oop_shift != 0) { // Shift out alignment bits
4169 if (((intptr_t)oop_base&0xc000000000000000L) == 0L) { // We are sure: no single address will have the leftmost bit set.
4170 z_srag(Rdst, Rsrc, oop_shift); // Arithmetic shift sets the condition code.
4171 } else {
4172 z_srlg(Rdst, Rsrc, oop_shift);
4173 z_ltgr(Rsrc, Rsrc); // This is the recommended way of testing for zero.
4174 // This probably is faster, as it does not write a register. No!
4175 // z_cghi(Rsrc, 0);
4176 }
4177 } else {
4178 z_ltgr(Rdst, Rsrc); // Move NULL to result register.
4179 }
4180 z_bre(done);
4181
4182 // Subtract oop_base components.
4183 if ((Rdst == Z_R0) || (Rbase == Z_R0)) {
4184 z_algr(Rdst, Rbase);
4185 if (pow2_offset != 0) { add2reg(Rdst, pow2_offset); }
4186 } else {
4187 add2reg_with_index(Rdst, pow2_offset, Rbase, Rdst);
4188 }
4189 if (!only32bitValid) {
4190 z_llgfr(Rdst, Rdst); // Clear upper bits in case the register will be decoded again.
4191 }
4192 bind(done);
4193
4194 } else { // not null
4195 // Get oop_base components.
4196 if (pow2_offset == -1) {
4197 pow2_offset = get_oop_base_complement(Rbase, (uint64_t)(intptr_t)oop_base);
4198 }
4199
4200 // Subtract oop_base components and shift.
4201 if (Rdst == Z_R0 || Rsrc == Z_R0 || Rbase == Z_R0) {
4202 // Don't use lay instruction.
4203 if (Rdst == Rsrc) {
4204 z_algr(Rdst, Rbase);
4205 } else {
4206 lgr_if_needed(Rdst, Rbase);
4207 z_algr(Rdst, Rsrc);
4208 }
4209 if (pow2_offset != 0) add2reg(Rdst, pow2_offset);
4210 } else {
4211 add2reg_with_index(Rdst, pow2_offset, Rbase, Rsrc);
4212 }
4213 if (oop_shift != 0) { // Shift out alignment bits.
4214 z_srlg(Rdst, Rdst, oop_shift);
4215 }
4216 if (!only32bitValid) {
4217 z_llgfr(Rdst, Rdst); // Clear upper bits in case the register will be decoded again.
4218 }
4219 }
4220 #ifdef ASSERT
4221 if (used_R0 && Rdst != Z_R0 && Rsrc != Z_R0) { preset_reg(Z_R0, 0xb01bUL, 2); }
4222 if (used_R1 && Rdst != Z_R1 && Rsrc != Z_R1) { preset_reg(Z_R1, 0xb02bUL, 2); }
4223 #endif
4224 BLOCK_COMMENT("} cOop encoder general");
4225 }
4226
4227 //-------------------------------------------------
4228 // decode compressed oop. Generally usable decoder.
4229 //-------------------------------------------------
4230 // Rsrc - contains compressed oop on entry.
4231 // Rdst - contains regular oop on exit.
4232 // Rdst and Rsrc may indicate same register.
4233 // Rdst must not be the same register as Rbase, if Rbase was preloaded (before call).
4234 // Rdst can be the same register as Rbase. Then, either Z_R0 or Z_R1 must be available as scratch.
4235 // Rbase - register to use for the base
4236 // pow2_offset - offset of base to nice value. If -1, base must be loaded.
4237 // For performance, it is good to
4238 // - avoid Z_R0 for any of the argument registers.
4239 // - keep Rdst and Rsrc distinct from Rbase. Rdst == Rsrc is ok for performance.
4240 // - avoid Z_R1 for Rdst if Rdst == Rbase.
4241 void MacroAssembler::oop_decoder(Register Rdst, Register Rsrc, bool maybeNULL, Register Rbase, int pow2_offset) {
4242
4243 const address oop_base = Universe::narrow_oop_base();
4244 const int oop_shift = Universe::narrow_oop_shift();
4245 const bool disjoint = Universe::narrow_oop_base_disjoint();
4246
4247 assert(UseCompressedOops, "must be on to call this method");
4248 assert(Universe::heap() != NULL, "java heap must be initialized to call this decoder");
4249 assert((oop_shift == 0) || (oop_shift == LogMinObjAlignmentInBytes),
4250 "cOop encoder detected bad shift");
4251
4252 // cOops are always loaded zero-extended from memory. No explicit zero-extension necessary.
4253
4254 if (oop_base != NULL) {
4255 unsigned int oop_base_hl = ((unsigned int)((uint64_t)(intptr_t)oop_base >> 32)) & 0xffff;
4256 unsigned int oop_base_hh = ((unsigned int)((uint64_t)(intptr_t)oop_base >> 48)) & 0xffff;
4257 unsigned int oop_base_hf = ((unsigned int)((uint64_t)(intptr_t)oop_base >> 32)) & 0xFFFFffff;
4258 if (disjoint && (oop_base_hl == 0 || oop_base_hh == 0)) {
4259 BLOCK_COMMENT("cOop decoder disjointBase {");
4260 // We do not need to load the base. Instead, we can install the upper bits
4261 // with an OR instead of an ADD.
4262 Label done;
4263
4264 // Rsrc contains a narrow oop. Thus we are sure the leftmost <oop_shift> bits will never be set.
4265 if (maybeNULL) { // NULL ptr must be preserved!
4266 z_slag(Rdst, Rsrc, oop_shift); // Arithmetic shift sets the condition code.
4267 z_bre(done);
4268 } else {
4269 z_sllg(Rdst, Rsrc, oop_shift); // Logical shift leaves condition code alone.
4270 }
4271 if ((oop_base_hl != 0) && (oop_base_hh != 0)) {
4272 z_oihf(Rdst, oop_base_hf);
4273 } else if (oop_base_hl != 0) {
4274 z_oihl(Rdst, oop_base_hl);
4275 } else {
4276 assert(oop_base_hh != 0, "not heapbased mode");
4277 z_oihh(Rdst, oop_base_hh);
4278 }
4279 bind(done);
4280 BLOCK_COMMENT("} cOop decoder disjointBase");
4281 } else {
4282 BLOCK_COMMENT("cOop decoder general {");
4283 // There are three decode steps:
4284 // scale oop offset (shift left)
4285 // get base (in reg) and pow2_offset (constant)
4286 // add base, pow2_offset, and oop offset
4287 // The following register overlap situations may exist:
4288 // Rdst == Rsrc, Rbase any other
4289 // not a problem. Scaling in-place leaves Rbase undisturbed.
4290 // Loading Rbase does not impact the scaled offset.
4291 // Rdst == Rbase, Rsrc any other
4292 // scaling would destroy a possibly preloaded Rbase. Loading Rbase
4293 // would destroy the scaled offset.
4294 // Remedy: use Rdst_tmp if Rbase has been preloaded.
4295 // use Rbase_tmp if base has to be loaded.
4296 // Rsrc == Rbase, Rdst any other
4297 // Only possible without preloaded Rbase.
4298 // Loading Rbase does not destroy compressed oop because it was scaled into Rdst before.
4299 // Rsrc == Rbase, Rdst == Rbase
4300 // Only possible without preloaded Rbase.
4301 // Loading Rbase would destroy compressed oop. Scaling in-place is ok.
4302 // Remedy: use Rbase_tmp.
4303 //
4304 Label done;
4305 Register Rdst_tmp = Rdst;
4306 Register Rbase_tmp = Rbase;
4307 bool used_R0 = false;
4308 bool used_R1 = false;
4309 bool base_preloaded = pow2_offset >= 0;
4310 guarantee(!(base_preloaded && (Rsrc == Rbase)), "Register clash, check caller");
4311 assert(oop_shift != 0, "room for optimization");
4312
4313 // Check if we need to use scratch registers.
4314 if (Rdst == Rbase) {
4315 assert(!(((Rdst == Z_R0) && (Rsrc == Z_R1)) || ((Rdst == Z_R1) && (Rsrc == Z_R0))), "need a scratch reg");
4316 if (Rdst != Rsrc) {
4317 if (base_preloaded) { Rdst_tmp = (Rdst == Z_R1) ? Z_R0 : Z_R1; }
4318 else { Rbase_tmp = (Rdst == Z_R1) ? Z_R0 : Z_R1; }
4319 } else {
4320 Rbase_tmp = (Rdst == Z_R1) ? Z_R0 : Z_R1;
4321 }
4322 }
4323 if (base_preloaded) lgr_if_needed(Rbase_tmp, Rbase);
4324
4325 // Scale oop and check for NULL.
4326 // Rsrc contains a narrow oop. Thus we are sure the leftmost <oop_shift> bits will never be set.
4327 if (maybeNULL) { // NULL ptr must be preserved!
4328 z_slag(Rdst_tmp, Rsrc, oop_shift); // Arithmetic shift sets the condition code.
4329 z_bre(done);
4330 } else {
4331 z_sllg(Rdst_tmp, Rsrc, oop_shift); // Logical shift leaves condition code alone.
4332 }
4333
4334 // Get oop_base components.
4335 if (!base_preloaded) {
4336 pow2_offset = get_oop_base(Rbase_tmp, (uint64_t)(intptr_t)oop_base);
4337 }
4338
4339 // Add up all components.
4340 if ((Rbase_tmp == Z_R0) || (Rdst_tmp == Z_R0)) {
4341 z_algr(Rdst_tmp, Rbase_tmp);
4342 if (pow2_offset != 0) { add2reg(Rdst_tmp, pow2_offset); }
4343 } else {
4344 add2reg_with_index(Rdst_tmp, pow2_offset, Rbase_tmp, Rdst_tmp);
4345 }
4346
4347 bind(done);
4348 lgr_if_needed(Rdst, Rdst_tmp);
4349 #ifdef ASSERT
4350 if (used_R0 && Rdst != Z_R0 && Rsrc != Z_R0) { preset_reg(Z_R0, 0xb03bUL, 2); }
4351 if (used_R1 && Rdst != Z_R1 && Rsrc != Z_R1) { preset_reg(Z_R1, 0xb04bUL, 2); }
4352 #endif
4353 BLOCK_COMMENT("} cOop decoder general");
4354 }
4355 } else {
4356 BLOCK_COMMENT("cOop decoder zeroBase {");
4357 if (oop_shift == 0) {
4358 lgr_if_needed(Rdst, Rsrc);
4359 } else {
4360 z_sllg(Rdst, Rsrc, oop_shift);
4361 }
4362 BLOCK_COMMENT("} cOop decoder zeroBase");
4363 }
4364 }
4365
4366 // ((OopHandle)result).resolve();
4367 void MacroAssembler::resolve_oop_handle(Register result) {
4368 // OopHandle::resolve is an indirection.
4369 z_lg(result, 0, result);
4370 }
4371
4372 void MacroAssembler::load_mirror(Register mirror, Register method) {
4373 mem2reg_opt(mirror, Address(method, Method::const_offset()));
4374 mem2reg_opt(mirror, Address(mirror, ConstMethod::constants_offset()));
4375 mem2reg_opt(mirror, Address(mirror, ConstantPool::pool_holder_offset_in_bytes()));
4376 mem2reg_opt(mirror, Address(mirror, Klass::java_mirror_offset()));
4377 resolve_oop_handle(mirror);
4378 }
4379
4380 //---------------------------------------------------------------
4381 //--- Operations on arrays.
4382 //---------------------------------------------------------------
4383
4384 // Compiler ensures base is doubleword aligned and cnt is #doublewords.
4385 // Emitter does not KILL cnt and base arguments, since they need to be copied to
4386 // work registers anyway.
4387 // Actually, only r0, r1, and r5 are killed.
4388 unsigned int MacroAssembler::Clear_Array(Register cnt_arg, Register base_pointer_arg, Register src_addr, Register src_len) {
4389 // Src_addr is evenReg.
4390 // Src_len is odd_Reg.
4391
4392 int block_start = offset();
4393 Register tmp_reg = src_len; // Holds target instr addr for EX.
4394 Register dst_len = Z_R1; // Holds dst len for MVCLE.
4395 Register dst_addr = Z_R0; // Holds dst addr for MVCLE.
4396
4397 Label doXC, doMVCLE, done;
4398
4399 BLOCK_COMMENT("Clear_Array {");
4400
4401 // Check for zero len and convert to long.
4402 z_ltgfr(src_len, cnt_arg); // Remember casted value for doSTG case.
4403 z_bre(done); // Nothing to do if len == 0.
4404
4405 // Prefetch data to be cleared.
4406 if (VM_Version::has_Prefetch()) {
4407 z_pfd(0x02, 0, Z_R0, base_pointer_arg);
4408 z_pfd(0x02, 256, Z_R0, base_pointer_arg);
4409 }
4410
4411 z_sllg(dst_len, src_len, 3); // #bytes to clear.
4412 z_cghi(src_len, 32); // Check for len <= 256 bytes (<=32 DW).
4413 z_brnh(doXC); // If so, use executed XC to clear.
4414
4415 // MVCLE: initialize long arrays (general case).
4416 bind(doMVCLE);
4417 z_lgr(dst_addr, base_pointer_arg);
4418 clear_reg(src_len, true, false); // Src len of MVCLE is zero.
4419
4420 MacroAssembler::move_long_ext(dst_addr, src_addr, 0);
4421 z_bru(done);
4422
4423 // XC: initialize short arrays.
4424 Label XC_template; // Instr template, never exec directly!
4425 bind(XC_template);
4426 z_xc(0,0,base_pointer_arg,0,base_pointer_arg);
4427
4428 bind(doXC);
4429 add2reg(dst_len, -1); // Get #bytes-1 for EXECUTE.
4430 if (VM_Version::has_ExecuteExtensions()) {
4431 z_exrl(dst_len, XC_template); // Execute XC with var. len.
4432 } else {
4433 z_larl(tmp_reg, XC_template);
4434 z_ex(dst_len,0,Z_R0,tmp_reg); // Execute XC with var. len.
4435 }
4436 // z_bru(done); // fallthru
4437
4438 bind(done);
4439
4440 BLOCK_COMMENT("} Clear_Array");
4441
4442 int block_end = offset();
4443 return block_end - block_start;
4444 }
4445
4446 // Compiler ensures base is doubleword aligned and cnt is count of doublewords.
4447 // Emitter does not KILL any arguments nor work registers.
4448 // Emitter generates up to 16 XC instructions, depending on the array length.
4449 unsigned int MacroAssembler::Clear_Array_Const(long cnt, Register base) {
4450 int block_start = offset();
4451 int off;
4452 int lineSize_Bytes = AllocatePrefetchStepSize;
4453 int lineSize_DW = AllocatePrefetchStepSize>>LogBytesPerWord;
4454 bool doPrefetch = VM_Version::has_Prefetch();
4455 int XC_maxlen = 256;
4456 int numXCInstr = cnt > 0 ? (cnt*BytesPerWord-1)/XC_maxlen+1 : 0;
4457
4458 BLOCK_COMMENT("Clear_Array_Const {");
4459 assert(cnt*BytesPerWord <= 4096, "ClearArrayConst can handle 4k only");
4460
4461 // Do less prefetching for very short arrays.
4462 if (numXCInstr > 0) {
4463 // Prefetch only some cache lines, then begin clearing.
4464 if (doPrefetch) {
4465 if (cnt*BytesPerWord <= lineSize_Bytes/4) { // If less than 1/4 of a cache line to clear,
4466 z_pfd(0x02, 0, Z_R0, base); // prefetch just the first cache line.
4467 } else {
4468 assert(XC_maxlen == lineSize_Bytes, "ClearArrayConst needs 256B cache lines");
4469 for (off = 0; (off < AllocatePrefetchLines) && (off <= numXCInstr); off ++) {
4470 z_pfd(0x02, off*lineSize_Bytes, Z_R0, base);
4471 }
4472 }
4473 }
4474
4475 for (off=0; off<(numXCInstr-1); off++) {
4476 z_xc(off*XC_maxlen, XC_maxlen-1, base, off*XC_maxlen, base);
4477
4478 // Prefetch some cache lines in advance.
4479 if (doPrefetch && (off <= numXCInstr-AllocatePrefetchLines)) {
4480 z_pfd(0x02, (off+AllocatePrefetchLines)*lineSize_Bytes, Z_R0, base);
4481 }
4482 }
4483 if (off*XC_maxlen < cnt*BytesPerWord) {
4484 z_xc(off*XC_maxlen, (cnt*BytesPerWord-off*XC_maxlen)-1, base, off*XC_maxlen, base);
4485 }
4486 }
4487 BLOCK_COMMENT("} Clear_Array_Const");
4488
4489 int block_end = offset();
4490 return block_end - block_start;
4491 }
4492
4493 // Compiler ensures base is doubleword aligned and cnt is #doublewords.
4494 // Emitter does not KILL cnt and base arguments, since they need to be copied to
4495 // work registers anyway.
4496 // Actually, only r0, r1, r4, and r5 (which are work registers) are killed.
4497 //
4498 // For very large arrays, exploit MVCLE H/W support.
4499 // MVCLE instruction automatically exploits H/W-optimized page mover.
4500 // - Bytes up to next page boundary are cleared with a series of XC to self.
4501 // - All full pages are cleared with the page mover H/W assist.
4502 // - Remaining bytes are again cleared by a series of XC to self.
4503 //
4504 unsigned int MacroAssembler::Clear_Array_Const_Big(long cnt, Register base_pointer_arg, Register src_addr, Register src_len) {
4505 // Src_addr is evenReg.
4506 // Src_len is odd_Reg.
4507
4508 int block_start = offset();
4509 Register dst_len = Z_R1; // Holds dst len for MVCLE.
4510 Register dst_addr = Z_R0; // Holds dst addr for MVCLE.
4511
4512 BLOCK_COMMENT("Clear_Array_Const_Big {");
4513
4514 // Get len to clear.
4515 load_const_optimized(dst_len, (long)cnt*8L); // in Bytes = #DW*8
4516
4517 // Prepare other args to MVCLE.
4518 z_lgr(dst_addr, base_pointer_arg);
4519 // Indicate unused result.
4520 (void) clear_reg(src_len, true, false); // Src len of MVCLE is zero.
4521
4522 // Clear.
4523 MacroAssembler::move_long_ext(dst_addr, src_addr, 0);
4524 BLOCK_COMMENT("} Clear_Array_Const_Big");
4525
4526 int block_end = offset();
4527 return block_end - block_start;
4528 }
4529
4530 // Allocator.
4531 unsigned int MacroAssembler::CopyRawMemory_AlignedDisjoint(Register src_reg, Register dst_reg,
4532 Register cnt_reg,
4533 Register tmp1_reg, Register tmp2_reg) {
4534 // Tmp1 is oddReg.
4535 // Tmp2 is evenReg.
4536
4537 int block_start = offset();
4538 Label doMVC, doMVCLE, done, MVC_template;
4539
4540 BLOCK_COMMENT("CopyRawMemory_AlignedDisjoint {");
4541
4542 // Check for zero len and convert to long.
4543 z_ltgfr(cnt_reg, cnt_reg); // Remember casted value for doSTG case.
4544 z_bre(done); // Nothing to do if len == 0.
4545
4546 z_sllg(Z_R1, cnt_reg, 3); // Dst len in bytes. calc early to have the result ready.
4547
4548 z_cghi(cnt_reg, 32); // Check for len <= 256 bytes (<=32 DW).
4549 z_brnh(doMVC); // If so, use executed MVC to clear.
4550
4551 bind(doMVCLE); // A lot of data (more than 256 bytes).
4552 // Prep dest reg pair.
4553 z_lgr(Z_R0, dst_reg); // dst addr
4554 // Dst len already in Z_R1.
4555 // Prep src reg pair.
4556 z_lgr(tmp2_reg, src_reg); // src addr
4557 z_lgr(tmp1_reg, Z_R1); // Src len same as dst len.
4558
4559 // Do the copy.
4560 move_long_ext(Z_R0, tmp2_reg, 0xb0); // Bypass cache.
4561 z_bru(done); // All done.
4562
4563 bind(MVC_template); // Just some data (not more than 256 bytes).
4564 z_mvc(0, 0, dst_reg, 0, src_reg);
4565
4566 bind(doMVC);
4567
4568 if (VM_Version::has_ExecuteExtensions()) {
4569 add2reg(Z_R1, -1);
4570 } else {
4571 add2reg(tmp1_reg, -1, Z_R1);
4572 z_larl(Z_R1, MVC_template);
4573 }
4574
4575 if (VM_Version::has_Prefetch()) {
4576 z_pfd(1, 0,Z_R0,src_reg);
4577 z_pfd(2, 0,Z_R0,dst_reg);
4578 // z_pfd(1,256,Z_R0,src_reg); // Assume very short copy.
4579 // z_pfd(2,256,Z_R0,dst_reg);
4580 }
4581
4582 if (VM_Version::has_ExecuteExtensions()) {
4583 z_exrl(Z_R1, MVC_template);
4584 } else {
4585 z_ex(tmp1_reg, 0, Z_R0, Z_R1);
4586 }
4587
4588 bind(done);
4589
4590 BLOCK_COMMENT("} CopyRawMemory_AlignedDisjoint");
4591
4592 int block_end = offset();
4593 return block_end - block_start;
4594 }
4595
4596 //------------------------------------------------------
4597 // Special String Intrinsics. Implementation
4598 //------------------------------------------------------
4599
4600 // Intrinsics for CompactStrings
4601
4602 // Compress char[] to byte[].
4603 // Restores: src, dst
4604 // Uses: cnt
4605 // Kills: tmp, Z_R0, Z_R1.
4606 // Early clobber: result.
4607 // Note:
4608 // cnt is signed int. Do not rely on high word!
4609 // counts # characters, not bytes.
4610 // The result is the number of characters copied before the first incompatible character was found.
4611 // If precise is true, the processing stops exactly at this point. Otherwise, the result may be off
4612 // by a few bytes. The result always indicates the number of copied characters.
4613 // When used as a character index, the returned value points to the first incompatible character.
4614 //
4615 // Note: Does not behave exactly like package private StringUTF16 compress java implementation in case of failure:
4616 // - Different number of characters may have been written to dead array (if precise is false).
4617 // - Returns a number <cnt instead of 0. (Result gets compared with cnt.)
4618 unsigned int MacroAssembler::string_compress(Register result, Register src, Register dst, Register cnt,
4619 Register tmp, bool precise) {
4620 assert_different_registers(Z_R0, Z_R1, result, src, dst, cnt, tmp);
4621
4622 if (precise) {
4623 BLOCK_COMMENT("encode_iso_array {");
4624 } else {
4625 BLOCK_COMMENT("string_compress {");
4626 }
4627 int block_start = offset();
4628
4629 Register Rsrc = src;
4630 Register Rdst = dst;
4631 Register Rix = tmp;
4632 Register Rcnt = cnt;
4633 Register Rmask = result; // holds incompatibility check mask until result value is stored.
4634 Label ScalarShortcut, AllDone;
4635
4636 z_iilf(Rmask, 0xFF00FF00);
4637 z_iihf(Rmask, 0xFF00FF00);
4638
4639 #if 0 // Sacrifice shortcuts for code compactness
4640 {
4641 //---< shortcuts for short strings (very frequent) >---
4642 // Strings with 4 and 8 characters were fond to occur very frequently.
4643 // Therefore, we handle them right away with minimal overhead.
4644 Label skipShortcut, skip4Shortcut, skip8Shortcut;
4645 Register Rout = Z_R0;
4646 z_chi(Rcnt, 4);
4647 z_brne(skip4Shortcut); // 4 characters are very frequent
4648 z_lg(Z_R0, 0, Rsrc); // Treat exactly 4 characters specially.
4649 if (VM_Version::has_DistinctOpnds()) {
4650 Rout = Z_R0;
4651 z_ngrk(Rix, Z_R0, Rmask);
4652 } else {
4653 Rout = Rix;
4654 z_lgr(Rix, Z_R0);
4655 z_ngr(Z_R0, Rmask);
4656 }
4657 z_brnz(skipShortcut);
4658 z_stcmh(Rout, 5, 0, Rdst);
4659 z_stcm(Rout, 5, 2, Rdst);
4660 z_lgfr(result, Rcnt);
4661 z_bru(AllDone);
4662 bind(skip4Shortcut);
4663
4664 z_chi(Rcnt, 8);
4665 z_brne(skip8Shortcut); // There's more to do...
4666 z_lmg(Z_R0, Z_R1, 0, Rsrc); // Treat exactly 8 characters specially.
4667 if (VM_Version::has_DistinctOpnds()) {
4668 Rout = Z_R0;
4669 z_ogrk(Rix, Z_R0, Z_R1);
4670 z_ngr(Rix, Rmask);
4671 } else {
4672 Rout = Rix;
4673 z_lgr(Rix, Z_R0);
4674 z_ogr(Z_R0, Z_R1);
4675 z_ngr(Z_R0, Rmask);
4676 }
4677 z_brnz(skipShortcut);
4678 z_stcmh(Rout, 5, 0, Rdst);
4679 z_stcm(Rout, 5, 2, Rdst);
4680 z_stcmh(Z_R1, 5, 4, Rdst);
4681 z_stcm(Z_R1, 5, 6, Rdst);
4682 z_lgfr(result, Rcnt);
4683 z_bru(AllDone);
4684
4685 bind(skip8Shortcut);
4686 clear_reg(Z_R0, true, false); // #characters already processed (none). Precond for scalar loop.
4687 z_brl(ScalarShortcut); // Just a few characters
4688
4689 bind(skipShortcut);
4690 }
4691 #endif
4692 clear_reg(Z_R0); // make sure register is properly initialized.
4693
4694 if (VM_Version::has_VectorFacility()) {
4695 const int min_vcnt = 32; // Minimum #characters required to use vector instructions.
4696 // Otherwise just do nothing in vector mode.
4697 // Must be multiple of 2*(vector register length in chars (8 HW = 128 bits)).
4698 const int log_min_vcnt = exact_log2(min_vcnt);
4699 Label VectorLoop, VectorDone, VectorBreak;
4700
4701 VectorRegister Vtmp1 = Z_V16;
4702 VectorRegister Vtmp2 = Z_V17;
4703 VectorRegister Vmask = Z_V18;
4704 VectorRegister Vzero = Z_V19;
4705 VectorRegister Vsrc_first = Z_V20;
4706 VectorRegister Vsrc_last = Z_V23;
4707
4708 assert((Vsrc_last->encoding() - Vsrc_first->encoding() + 1) == min_vcnt/8, "logic error");
4709 assert(VM_Version::has_DistinctOpnds(), "Assumption when has_VectorFacility()");
4710 z_srak(Rix, Rcnt, log_min_vcnt); // # vector loop iterations
4711 z_brz(VectorDone); // not enough data for vector loop
4712
4713 z_vzero(Vzero); // all zeroes
4714 z_vgmh(Vmask, 0, 7); // generate 0xff00 mask for all 2-byte elements
4715 z_sllg(Z_R0, Rix, log_min_vcnt); // remember #chars that will be processed by vector loop
4716
4717 bind(VectorLoop);
4718 z_vlm(Vsrc_first, Vsrc_last, 0, Rsrc);
4719 add2reg(Rsrc, min_vcnt*2);
4720
4721 //---< check for incompatible character >---
4722 z_vo(Vtmp1, Z_V20, Z_V21);
4723 z_vo(Vtmp2, Z_V22, Z_V23);
4724 z_vo(Vtmp1, Vtmp1, Vtmp2);
4725 z_vn(Vtmp1, Vtmp1, Vmask);
4726 z_vceqhs(Vtmp1, Vtmp1, Vzero); // high half of all chars must be zero for successful compress.
4727 z_bvnt(VectorBreak); // break vector loop if not all vector elements compare eq -> incompatible character found.
4728 // re-process data from current iteration in break handler.
4729
4730 //---< pack & store characters >---
4731 z_vpkh(Vtmp1, Z_V20, Z_V21); // pack (src1, src2) -> tmp1
4732 z_vpkh(Vtmp2, Z_V22, Z_V23); // pack (src3, src4) -> tmp2
4733 z_vstm(Vtmp1, Vtmp2, 0, Rdst); // store packed string
4734 add2reg(Rdst, min_vcnt);
4735
4736 z_brct(Rix, VectorLoop);
4737
4738 z_bru(VectorDone);
4739
4740 bind(VectorBreak);
4741 add2reg(Rsrc, -min_vcnt*2); // Fix Rsrc. Rsrc was already updated, but Rdst and Rix are not.
4742 z_sll(Rix, log_min_vcnt); // # chars processed so far in VectorLoop, excl. current iteration.
4743 z_sr(Z_R0, Rix); // correct # chars processed in total.
4744
4745 bind(VectorDone);
4746 }
4747
4748 {
4749 const int min_cnt = 8; // Minimum #characters required to use unrolled loop.
4750 // Otherwise just do nothing in unrolled loop.
4751 // Must be multiple of 8.
4752 const int log_min_cnt = exact_log2(min_cnt);
4753 Label UnrolledLoop, UnrolledDone, UnrolledBreak;
4754
4755 if (VM_Version::has_DistinctOpnds()) {
4756 z_srk(Rix, Rcnt, Z_R0); // remaining # chars to compress in unrolled loop
4757 } else {
4758 z_lr(Rix, Rcnt);
4759 z_sr(Rix, Z_R0);
4760 }
4761 z_sra(Rix, log_min_cnt); // unrolled loop count
4762 z_brz(UnrolledDone);
4763
4764 bind(UnrolledLoop);
4765 z_lmg(Z_R0, Z_R1, 0, Rsrc);
4766 if (precise) {
4767 z_ogr(Z_R1, Z_R0); // check all 8 chars for incompatibility
4768 z_ngr(Z_R1, Rmask);
4769 z_brnz(UnrolledBreak);
4770
4771 z_lg(Z_R1, 8, Rsrc); // reload destroyed register
4772 z_stcmh(Z_R0, 5, 0, Rdst);
4773 z_stcm(Z_R0, 5, 2, Rdst);
4774 } else {
4775 z_stcmh(Z_R0, 5, 0, Rdst);
4776 z_stcm(Z_R0, 5, 2, Rdst);
4777
4778 z_ogr(Z_R0, Z_R1);
4779 z_ngr(Z_R0, Rmask);
4780 z_brnz(UnrolledBreak);
4781 }
4782 z_stcmh(Z_R1, 5, 4, Rdst);
4783 z_stcm(Z_R1, 5, 6, Rdst);
4784
4785 add2reg(Rsrc, min_cnt*2);
4786 add2reg(Rdst, min_cnt);
4787 z_brct(Rix, UnrolledLoop);
4788
4789 z_lgfr(Z_R0, Rcnt); // # chars processed in total after unrolled loop.
4790 z_nilf(Z_R0, ~(min_cnt-1));
4791 z_tmll(Rcnt, min_cnt-1);
4792 z_brnaz(ScalarShortcut); // if all bits zero, there is nothing left to do for scalar loop.
4793 // Rix == 0 in all cases.
4794 z_sllg(Z_R1, Rcnt, 1); // # src bytes already processed. Only lower 32 bits are valid!
4795 // Z_R1 contents must be treated as unsigned operand! For huge strings,
4796 // (Rcnt >= 2**30), the value may spill into the sign bit by sllg.
4797 z_lgfr(result, Rcnt); // all characters processed.
4798 z_slgfr(Rdst, Rcnt); // restore ptr
4799 z_slgfr(Rsrc, Z_R1); // restore ptr, double the element count for Rsrc restore
4800 z_bru(AllDone);
4801
4802 bind(UnrolledBreak);
4803 z_lgfr(Z_R0, Rcnt); // # chars processed in total after unrolled loop
4804 z_nilf(Z_R0, ~(min_cnt-1));
4805 z_sll(Rix, log_min_cnt); // # chars not yet processed in UnrolledLoop (due to break), broken iteration not included.
4806 z_sr(Z_R0, Rix); // fix # chars processed OK so far.
4807 if (!precise) {
4808 z_lgfr(result, Z_R0);
4809 z_sllg(Z_R1, Z_R0, 1); // # src bytes already processed. Only lower 32 bits are valid!
4810 // Z_R1 contents must be treated as unsigned operand! For huge strings,
4811 // (Rcnt >= 2**30), the value may spill into the sign bit by sllg.
4812 z_aghi(result, min_cnt/2); // min_cnt/2 characters have already been written
4813 // but ptrs were not updated yet.
4814 z_slgfr(Rdst, Z_R0); // restore ptr
4815 z_slgfr(Rsrc, Z_R1); // restore ptr, double the element count for Rsrc restore
4816 z_bru(AllDone);
4817 }
4818 bind(UnrolledDone);
4819 }
4820
4821 {
4822 Label ScalarLoop, ScalarDone, ScalarBreak;
4823
4824 bind(ScalarShortcut);
4825 z_ltgfr(result, Rcnt);
4826 z_brz(AllDone);
4827
4828 #if 0 // Sacrifice shortcuts for code compactness
4829 {
4830 //---< Special treatment for very short strings (one or two characters) >---
4831 // For these strings, we are sure that the above code was skipped.
4832 // Thus, no registers were modified, register restore is not required.
4833 Label ScalarDoit, Scalar2Char;
4834 z_chi(Rcnt, 2);
4835 z_brh(ScalarDoit);
4836 z_llh(Z_R1, 0, Z_R0, Rsrc);
4837 z_bre(Scalar2Char);
4838 z_tmll(Z_R1, 0xff00);
4839 z_lghi(result, 0); // cnt == 1, first char invalid, no chars successfully processed
4840 z_brnaz(AllDone);
4841 z_stc(Z_R1, 0, Z_R0, Rdst);
4842 z_lghi(result, 1);
4843 z_bru(AllDone);
4844
4845 bind(Scalar2Char);
4846 z_llh(Z_R0, 2, Z_R0, Rsrc);
4847 z_tmll(Z_R1, 0xff00);
4848 z_lghi(result, 0); // cnt == 2, first char invalid, no chars successfully processed
4849 z_brnaz(AllDone);
4850 z_stc(Z_R1, 0, Z_R0, Rdst);
4851 z_tmll(Z_R0, 0xff00);
4852 z_lghi(result, 1); // cnt == 2, second char invalid, one char successfully processed
4853 z_brnaz(AllDone);
4854 z_stc(Z_R0, 1, Z_R0, Rdst);
4855 z_lghi(result, 2);
4856 z_bru(AllDone);
4857
4858 bind(ScalarDoit);
4859 }
4860 #endif
4861
4862 if (VM_Version::has_DistinctOpnds()) {
4863 z_srk(Rix, Rcnt, Z_R0); // remaining # chars to compress in unrolled loop
4864 } else {
4865 z_lr(Rix, Rcnt);
4866 z_sr(Rix, Z_R0);
4867 }
4868 z_lgfr(result, Rcnt); // # processed characters (if all runs ok).
4869 z_brz(ScalarDone); // uses CC from Rix calculation
4870
4871 bind(ScalarLoop);
4872 z_llh(Z_R1, 0, Z_R0, Rsrc);
4873 z_tmll(Z_R1, 0xff00);
4874 z_brnaz(ScalarBreak);
4875 z_stc(Z_R1, 0, Z_R0, Rdst);
4876 add2reg(Rsrc, 2);
4877 add2reg(Rdst, 1);
4878 z_brct(Rix, ScalarLoop);
4879
4880 z_bru(ScalarDone);
4881
4882 bind(ScalarBreak);
4883 z_sr(result, Rix);
4884
4885 bind(ScalarDone);
4886 z_sgfr(Rdst, result); // restore ptr
4887 z_sgfr(Rsrc, result); // restore ptr, double the element count for Rsrc restore
4888 z_sgfr(Rsrc, result);
4889 }
4890 bind(AllDone);
4891
4892 if (precise) {
4893 BLOCK_COMMENT("} encode_iso_array");
4894 } else {
4895 BLOCK_COMMENT("} string_compress");
4896 }
4897 return offset() - block_start;
4898 }
4899
4900 // Inflate byte[] to char[].
4901 unsigned int MacroAssembler::string_inflate_trot(Register src, Register dst, Register cnt, Register tmp) {
4902 int block_start = offset();
4903
4904 BLOCK_COMMENT("string_inflate {");
4905
4906 Register stop_char = Z_R0;
4907 Register table = Z_R1;
4908 Register src_addr = tmp;
4909
4910 assert_different_registers(Z_R0, Z_R1, tmp, src, dst, cnt);
4911 assert(dst->encoding()%2 == 0, "must be even reg");
4912 assert(cnt->encoding()%2 == 1, "must be odd reg");
4913 assert(cnt->encoding() - dst->encoding() == 1, "must be even/odd pair");
4914
4915 StubRoutines::zarch::generate_load_trot_table_addr(this, table); // kills Z_R0 (if ASSERT)
4916 clear_reg(stop_char); // Stop character. Not used here, but initialized to have a defined value.
4917 lgr_if_needed(src_addr, src);
4918 z_llgfr(cnt, cnt); // # src characters, must be a positive simm32.
4919
4920 translate_ot(dst, src_addr, /* mask = */ 0x0001);
4921
4922 BLOCK_COMMENT("} string_inflate");
4923
4924 return offset() - block_start;
4925 }
4926
4927 // Inflate byte[] to char[].
4928 // Restores: src, dst
4929 // Uses: cnt
4930 // Kills: tmp, Z_R0, Z_R1.
4931 // Note:
4932 // cnt is signed int. Do not rely on high word!
4933 // counts # characters, not bytes.
4934 unsigned int MacroAssembler::string_inflate(Register src, Register dst, Register cnt, Register tmp) {
4935 assert_different_registers(Z_R0, Z_R1, src, dst, cnt, tmp);
4936
4937 BLOCK_COMMENT("string_inflate {");
4938 int block_start = offset();
4939
4940 Register Rcnt = cnt; // # characters (src: bytes, dst: char (2-byte)), remaining after current loop.
4941 Register Rix = tmp; // loop index
4942 Register Rsrc = src; // addr(src array)
4943 Register Rdst = dst; // addr(dst array)
4944 Label ScalarShortcut, AllDone;
4945
4946 #if 0 // Sacrifice shortcuts for code compactness
4947 {
4948 //---< shortcuts for short strings (very frequent) >---
4949 Label skipShortcut, skip4Shortcut;
4950 z_ltr(Rcnt, Rcnt); // absolutely nothing to do for strings of len == 0.
4951 z_brz(AllDone);
4952 clear_reg(Z_R0); // make sure registers are properly initialized.
4953 clear_reg(Z_R1);
4954 z_chi(Rcnt, 4);
4955 z_brne(skip4Shortcut); // 4 characters are very frequent
4956 z_icm(Z_R0, 5, 0, Rsrc); // Treat exactly 4 characters specially.
4957 z_icm(Z_R1, 5, 2, Rsrc);
4958 z_stm(Z_R0, Z_R1, 0, Rdst);
4959 z_bru(AllDone);
4960 bind(skip4Shortcut);
4961
4962 z_chi(Rcnt, 8);
4963 z_brh(skipShortcut); // There's a lot to do...
4964 z_lgfr(Z_R0, Rcnt); // remaining #characters (<= 8). Precond for scalar loop.
4965 // This does not destroy the "register cleared" state of Z_R0.
4966 z_brl(ScalarShortcut); // Just a few characters
4967 z_icmh(Z_R0, 5, 0, Rsrc); // Treat exactly 8 characters specially.
4968 z_icmh(Z_R1, 5, 4, Rsrc);
4969 z_icm(Z_R0, 5, 2, Rsrc);
4970 z_icm(Z_R1, 5, 6, Rsrc);
4971 z_stmg(Z_R0, Z_R1, 0, Rdst);
4972 z_bru(AllDone);
4973 bind(skipShortcut);
4974 }
4975 #endif
4976 clear_reg(Z_R0); // make sure register is properly initialized.
4977
4978 if (VM_Version::has_VectorFacility()) {
4979 const int min_vcnt = 32; // Minimum #characters required to use vector instructions.
4980 // Otherwise just do nothing in vector mode.
4981 // Must be multiple of vector register length (16 bytes = 128 bits).
4982 const int log_min_vcnt = exact_log2(min_vcnt);
4983 Label VectorLoop, VectorDone;
4984
4985 assert(VM_Version::has_DistinctOpnds(), "Assumption when has_VectorFacility()");
4986 z_srak(Rix, Rcnt, log_min_vcnt); // calculate # vector loop iterations
4987 z_brz(VectorDone); // skip if none
4988
4989 z_sllg(Z_R0, Rix, log_min_vcnt); // remember #chars that will be processed by vector loop
4990
4991 bind(VectorLoop);
4992 z_vlm(Z_V20, Z_V21, 0, Rsrc); // get next 32 characters (single-byte)
4993 add2reg(Rsrc, min_vcnt);
4994
4995 z_vuplhb(Z_V22, Z_V20); // V2 <- (expand) V0(high)
4996 z_vupllb(Z_V23, Z_V20); // V3 <- (expand) V0(low)
4997 z_vuplhb(Z_V24, Z_V21); // V4 <- (expand) V1(high)
4998 z_vupllb(Z_V25, Z_V21); // V5 <- (expand) V1(low)
4999 z_vstm(Z_V22, Z_V25, 0, Rdst); // store next 32 bytes
5000 add2reg(Rdst, min_vcnt*2);
5001
5002 z_brct(Rix, VectorLoop);
5003
5004 bind(VectorDone);
5005 }
5006
5007 const int min_cnt = 8; // Minimum #characters required to use unrolled scalar loop.
5008 // Otherwise just do nothing in unrolled scalar mode.
5009 // Must be multiple of 8.
5010 {
5011 const int log_min_cnt = exact_log2(min_cnt);
5012 Label UnrolledLoop, UnrolledDone;
5013
5014
5015 if (VM_Version::has_DistinctOpnds()) {
5016 z_srk(Rix, Rcnt, Z_R0); // remaining # chars to process in unrolled loop
5017 } else {
5018 z_lr(Rix, Rcnt);
5019 z_sr(Rix, Z_R0);
5020 }
5021 z_sra(Rix, log_min_cnt); // unrolled loop count
5022 z_brz(UnrolledDone);
5023
5024 clear_reg(Z_R0);
5025 clear_reg(Z_R1);
5026
5027 bind(UnrolledLoop);
5028 z_icmh(Z_R0, 5, 0, Rsrc);
5029 z_icmh(Z_R1, 5, 4, Rsrc);
5030 z_icm(Z_R0, 5, 2, Rsrc);
5031 z_icm(Z_R1, 5, 6, Rsrc);
5032 add2reg(Rsrc, min_cnt);
5033
5034 z_stmg(Z_R0, Z_R1, 0, Rdst);
5035
5036 add2reg(Rdst, min_cnt*2);
5037 z_brct(Rix, UnrolledLoop);
5038
5039 bind(UnrolledDone);
5040 z_lgfr(Z_R0, Rcnt); // # chars left over after unrolled loop.
5041 z_nilf(Z_R0, min_cnt-1);
5042 z_brnz(ScalarShortcut); // if zero, there is nothing left to do for scalar loop.
5043 // Rix == 0 in all cases.
5044 z_sgfr(Z_R0, Rcnt); // negative # characters the ptrs have been advanced previously.
5045 z_agr(Rdst, Z_R0); // restore ptr, double the element count for Rdst restore.
5046 z_agr(Rdst, Z_R0);
5047 z_agr(Rsrc, Z_R0); // restore ptr.
5048 z_bru(AllDone);
5049 }
5050
5051 {
5052 bind(ScalarShortcut);
5053 // Z_R0 must contain remaining # characters as 64-bit signed int here.
5054 // register contents is preserved over scalar processing (for register fixup).
5055
5056 #if 0 // Sacrifice shortcuts for code compactness
5057 {
5058 Label ScalarDefault;
5059 z_chi(Rcnt, 2);
5060 z_brh(ScalarDefault);
5061 z_llc(Z_R0, 0, Z_R0, Rsrc); // 6 bytes
5062 z_sth(Z_R0, 0, Z_R0, Rdst); // 4 bytes
5063 z_brl(AllDone);
5064 z_llc(Z_R0, 1, Z_R0, Rsrc); // 6 bytes
5065 z_sth(Z_R0, 2, Z_R0, Rdst); // 4 bytes
5066 z_bru(AllDone);
5067 bind(ScalarDefault);
5068 }
5069 #endif
5070
5071 Label CodeTable;
5072 // Some comments on Rix calculation:
5073 // - Rcnt is small, therefore no bits shifted out of low word (sll(g) instructions).
5074 // - high word of both Rix and Rcnt may contain garbage
5075 // - the final lngfr takes care of that garbage, extending the sign to high word
5076 z_sllg(Rix, Z_R0, 2); // calculate 10*Rix = (4*Rix + Rix)*2
5077 z_ar(Rix, Z_R0);
5078 z_larl(Z_R1, CodeTable);
5079 z_sll(Rix, 1);
5080 z_lngfr(Rix, Rix); // ix range: [0..7], after inversion & mult: [-(7*12)..(0*12)].
5081 z_bc(Assembler::bcondAlways, 0, Rix, Z_R1);
5082
5083 z_llc(Z_R1, 6, Z_R0, Rsrc); // 6 bytes
5084 z_sth(Z_R1, 12, Z_R0, Rdst); // 4 bytes
5085
5086 z_llc(Z_R1, 5, Z_R0, Rsrc);
5087 z_sth(Z_R1, 10, Z_R0, Rdst);
5088
5089 z_llc(Z_R1, 4, Z_R0, Rsrc);
5090 z_sth(Z_R1, 8, Z_R0, Rdst);
5091
5092 z_llc(Z_R1, 3, Z_R0, Rsrc);
5093 z_sth(Z_R1, 6, Z_R0, Rdst);
5094
5095 z_llc(Z_R1, 2, Z_R0, Rsrc);
5096 z_sth(Z_R1, 4, Z_R0, Rdst);
5097
5098 z_llc(Z_R1, 1, Z_R0, Rsrc);
5099 z_sth(Z_R1, 2, Z_R0, Rdst);
5100
5101 z_llc(Z_R1, 0, Z_R0, Rsrc);
5102 z_sth(Z_R1, 0, Z_R0, Rdst);
5103 bind(CodeTable);
5104
5105 z_chi(Rcnt, 8); // no fixup for small strings. Rdst, Rsrc were not modified.
5106 z_brl(AllDone);
5107
5108 z_sgfr(Z_R0, Rcnt); // # characters the ptrs have been advanced previously.
5109 z_agr(Rdst, Z_R0); // restore ptr, double the element count for Rdst restore.
5110 z_agr(Rdst, Z_R0);
5111 z_agr(Rsrc, Z_R0); // restore ptr.
5112 }
5113 bind(AllDone);
5114
5115 BLOCK_COMMENT("} string_inflate");
5116 return offset() - block_start;
5117 }
5118
5119 // Inflate byte[] to char[], length known at compile time.
5120 // Restores: src, dst
5121 // Kills: tmp, Z_R0, Z_R1.
5122 // Note:
5123 // len is signed int. Counts # characters, not bytes.
5124 unsigned int MacroAssembler::string_inflate_const(Register src, Register dst, Register tmp, int len) {
5125 assert_different_registers(Z_R0, Z_R1, src, dst, tmp);
5126
5127 BLOCK_COMMENT("string_inflate_const {");
5128 int block_start = offset();
5129
5130 Register Rix = tmp; // loop index
5131 Register Rsrc = src; // addr(src array)
5132 Register Rdst = dst; // addr(dst array)
5133 Label ScalarShortcut, AllDone;
5134 int nprocessed = 0;
5135 int src_off = 0; // compensate for saved (optimized away) ptr advancement.
5136 int dst_off = 0; // compensate for saved (optimized away) ptr advancement.
5137 bool restore_inputs = false;
5138 bool workreg_clear = false;
5139
5140 if ((len >= 32) && VM_Version::has_VectorFacility()) {
5141 const int min_vcnt = 32; // Minimum #characters required to use vector instructions.
5142 // Otherwise just do nothing in vector mode.
5143 // Must be multiple of vector register length (16 bytes = 128 bits).
5144 const int log_min_vcnt = exact_log2(min_vcnt);
5145 const int iterations = (len - nprocessed) >> log_min_vcnt;
5146 nprocessed += iterations << log_min_vcnt;
5147 Label VectorLoop;
5148
5149 if (iterations == 1) {
5150 z_vlm(Z_V20, Z_V21, 0+src_off, Rsrc); // get next 32 characters (single-byte)
5151 z_vuplhb(Z_V22, Z_V20); // V2 <- (expand) V0(high)
5152 z_vupllb(Z_V23, Z_V20); // V3 <- (expand) V0(low)
5153 z_vuplhb(Z_V24, Z_V21); // V4 <- (expand) V1(high)
5154 z_vupllb(Z_V25, Z_V21); // V5 <- (expand) V1(low)
5155 z_vstm(Z_V22, Z_V25, 0+dst_off, Rdst); // store next 32 bytes
5156
5157 src_off += min_vcnt;
5158 dst_off += min_vcnt*2;
5159 } else {
5160 restore_inputs = true;
5161
5162 z_lgfi(Rix, len>>log_min_vcnt);
5163 bind(VectorLoop);
5164 z_vlm(Z_V20, Z_V21, 0, Rsrc); // get next 32 characters (single-byte)
5165 add2reg(Rsrc, min_vcnt);
5166
5167 z_vuplhb(Z_V22, Z_V20); // V2 <- (expand) V0(high)
5168 z_vupllb(Z_V23, Z_V20); // V3 <- (expand) V0(low)
5169 z_vuplhb(Z_V24, Z_V21); // V4 <- (expand) V1(high)
5170 z_vupllb(Z_V25, Z_V21); // V5 <- (expand) V1(low)
5171 z_vstm(Z_V22, Z_V25, 0, Rdst); // store next 32 bytes
5172 add2reg(Rdst, min_vcnt*2);
5173
5174 z_brct(Rix, VectorLoop);
5175 }
5176 }
5177
5178 if (((len-nprocessed) >= 16) && VM_Version::has_VectorFacility()) {
5179 const int min_vcnt = 16; // Minimum #characters required to use vector instructions.
5180 // Otherwise just do nothing in vector mode.
5181 // Must be multiple of vector register length (16 bytes = 128 bits).
5182 const int log_min_vcnt = exact_log2(min_vcnt);
5183 const int iterations = (len - nprocessed) >> log_min_vcnt;
5184 nprocessed += iterations << log_min_vcnt;
5185 assert(iterations == 1, "must be!");
5186
5187 z_vl(Z_V20, 0+src_off, Z_R0, Rsrc); // get next 16 characters (single-byte)
5188 z_vuplhb(Z_V22, Z_V20); // V2 <- (expand) V0(high)
5189 z_vupllb(Z_V23, Z_V20); // V3 <- (expand) V0(low)
5190 z_vstm(Z_V22, Z_V23, 0+dst_off, Rdst); // store next 32 bytes
5191
5192 src_off += min_vcnt;
5193 dst_off += min_vcnt*2;
5194 }
5195
5196 if ((len-nprocessed) > 8) {
5197 const int min_cnt = 8; // Minimum #characters required to use unrolled scalar loop.
5198 // Otherwise just do nothing in unrolled scalar mode.
5199 // Must be multiple of 8.
5200 const int log_min_cnt = exact_log2(min_cnt);
5201 const int iterations = (len - nprocessed) >> log_min_cnt;
5202 nprocessed += iterations << log_min_cnt;
5203
5204 //---< avoid loop overhead/ptr increment for small # iterations >---
5205 if (iterations <= 2) {
5206 clear_reg(Z_R0);
5207 clear_reg(Z_R1);
5208 workreg_clear = true;
5209
5210 z_icmh(Z_R0, 5, 0+src_off, Rsrc);
5211 z_icmh(Z_R1, 5, 4+src_off, Rsrc);
5212 z_icm(Z_R0, 5, 2+src_off, Rsrc);
5213 z_icm(Z_R1, 5, 6+src_off, Rsrc);
5214 z_stmg(Z_R0, Z_R1, 0+dst_off, Rdst);
5215
5216 src_off += min_cnt;
5217 dst_off += min_cnt*2;
5218 }
5219
5220 if (iterations == 2) {
5221 z_icmh(Z_R0, 5, 0+src_off, Rsrc);
5222 z_icmh(Z_R1, 5, 4+src_off, Rsrc);
5223 z_icm(Z_R0, 5, 2+src_off, Rsrc);
5224 z_icm(Z_R1, 5, 6+src_off, Rsrc);
5225 z_stmg(Z_R0, Z_R1, 0+dst_off, Rdst);
5226
5227 src_off += min_cnt;
5228 dst_off += min_cnt*2;
5229 }
5230
5231 if (iterations > 2) {
5232 Label UnrolledLoop;
5233 restore_inputs = true;
5234
5235 clear_reg(Z_R0);
5236 clear_reg(Z_R1);
5237 workreg_clear = true;
5238
5239 z_lgfi(Rix, iterations);
5240 bind(UnrolledLoop);
5241 z_icmh(Z_R0, 5, 0, Rsrc);
5242 z_icmh(Z_R1, 5, 4, Rsrc);
5243 z_icm(Z_R0, 5, 2, Rsrc);
5244 z_icm(Z_R1, 5, 6, Rsrc);
5245 add2reg(Rsrc, min_cnt);
5246
5247 z_stmg(Z_R0, Z_R1, 0, Rdst);
5248 add2reg(Rdst, min_cnt*2);
5249
5250 z_brct(Rix, UnrolledLoop);
5251 }
5252 }
5253
5254 if ((len-nprocessed) > 0) {
5255 switch (len-nprocessed) {
5256 case 8:
5257 if (!workreg_clear) {
5258 clear_reg(Z_R0);
5259 clear_reg(Z_R1);
5260 }
5261 z_icmh(Z_R0, 5, 0+src_off, Rsrc);
5262 z_icmh(Z_R1, 5, 4+src_off, Rsrc);
5263 z_icm(Z_R0, 5, 2+src_off, Rsrc);
5264 z_icm(Z_R1, 5, 6+src_off, Rsrc);
5265 z_stmg(Z_R0, Z_R1, 0+dst_off, Rdst);
5266 break;
5267 case 7:
5268 if (!workreg_clear) {
5269 clear_reg(Z_R0);
5270 clear_reg(Z_R1);
5271 }
5272 clear_reg(Rix);
5273 z_icm(Z_R0, 5, 0+src_off, Rsrc);
5274 z_icm(Z_R1, 5, 2+src_off, Rsrc);
5275 z_icm(Rix, 5, 4+src_off, Rsrc);
5276 z_stm(Z_R0, Z_R1, 0+dst_off, Rdst);
5277 z_llc(Z_R0, 6+src_off, Z_R0, Rsrc);
5278 z_st(Rix, 8+dst_off, Z_R0, Rdst);
5279 z_sth(Z_R0, 12+dst_off, Z_R0, Rdst);
5280 break;
5281 case 6:
5282 if (!workreg_clear) {
5283 clear_reg(Z_R0);
5284 clear_reg(Z_R1);
5285 }
5286 clear_reg(Rix);
5287 z_icm(Z_R0, 5, 0+src_off, Rsrc);
5288 z_icm(Z_R1, 5, 2+src_off, Rsrc);
5289 z_icm(Rix, 5, 4+src_off, Rsrc);
5290 z_stm(Z_R0, Z_R1, 0+dst_off, Rdst);
5291 z_st(Rix, 8+dst_off, Z_R0, Rdst);
5292 break;
5293 case 5:
5294 if (!workreg_clear) {
5295 clear_reg(Z_R0);
5296 clear_reg(Z_R1);
5297 }
5298 z_icm(Z_R0, 5, 0+src_off, Rsrc);
5299 z_icm(Z_R1, 5, 2+src_off, Rsrc);
5300 z_llc(Rix, 4+src_off, Z_R0, Rsrc);
5301 z_stm(Z_R0, Z_R1, 0+dst_off, Rdst);
5302 z_sth(Rix, 8+dst_off, Z_R0, Rdst);
5303 break;
5304 case 4:
5305 if (!workreg_clear) {
5306 clear_reg(Z_R0);
5307 clear_reg(Z_R1);
5308 }
5309 z_icm(Z_R0, 5, 0+src_off, Rsrc);
5310 z_icm(Z_R1, 5, 2+src_off, Rsrc);
5311 z_stm(Z_R0, Z_R1, 0+dst_off, Rdst);
5312 break;
5313 case 3:
5314 if (!workreg_clear) {
5315 clear_reg(Z_R0);
5316 }
5317 z_llc(Z_R1, 2+src_off, Z_R0, Rsrc);
5318 z_icm(Z_R0, 5, 0+src_off, Rsrc);
5319 z_sth(Z_R1, 4+dst_off, Z_R0, Rdst);
5320 z_st(Z_R0, 0+dst_off, Rdst);
5321 break;
5322 case 2:
5323 z_llc(Z_R0, 0+src_off, Z_R0, Rsrc);
5324 z_llc(Z_R1, 1+src_off, Z_R0, Rsrc);
5325 z_sth(Z_R0, 0+dst_off, Z_R0, Rdst);
5326 z_sth(Z_R1, 2+dst_off, Z_R0, Rdst);
5327 break;
5328 case 1:
5329 z_llc(Z_R0, 0+src_off, Z_R0, Rsrc);
5330 z_sth(Z_R0, 0+dst_off, Z_R0, Rdst);
5331 break;
5332 default:
5333 guarantee(false, "Impossible");
5334 break;
5335 }
5336 src_off += len-nprocessed;
5337 dst_off += (len-nprocessed)*2;
5338 nprocessed = len;
5339 }
5340
5341 //---< restore modified input registers >---
5342 if ((nprocessed > 0) && restore_inputs) {
5343 z_agfi(Rsrc, -(nprocessed-src_off));
5344 if (nprocessed < 1000000000) { // avoid int overflow
5345 z_agfi(Rdst, -(nprocessed*2-dst_off));
5346 } else {
5347 z_agfi(Rdst, -(nprocessed-dst_off));
5348 z_agfi(Rdst, -nprocessed);
5349 }
5350 }
5351
5352 BLOCK_COMMENT("} string_inflate_const");
5353 return offset() - block_start;
5354 }
5355
5356 // Kills src.
5357 unsigned int MacroAssembler::has_negatives(Register result, Register src, Register cnt,
5358 Register odd_reg, Register even_reg, Register tmp) {
5359 int block_start = offset();
5360 Label Lloop1, Lloop2, Lslow, Lnotfound, Ldone;
5361 const Register addr = src, mask = tmp;
5362
5363 BLOCK_COMMENT("has_negatives {");
5364
5365 z_llgfr(Z_R1, cnt); // Number of bytes to read. (Must be a positive simm32.)
5366 z_llilf(mask, 0x80808080);
5367 z_lhi(result, 1); // Assume true.
5368 // Last possible addr for fast loop.
5369 z_lay(odd_reg, -16, Z_R1, src);
5370 z_chi(cnt, 16);
5371 z_brl(Lslow);
5372
5373 // ind1: index, even_reg: index increment, odd_reg: index limit
5374 z_iihf(mask, 0x80808080);
5375 z_lghi(even_reg, 16);
5376
5377 bind(Lloop1); // 16 bytes per iteration.
5378 z_lg(Z_R0, Address(addr));
5379 z_lg(Z_R1, Address(addr, 8));
5380 z_ogr(Z_R0, Z_R1);
5381 z_ngr(Z_R0, mask);
5382 z_brne(Ldone); // If found return 1.
5383 z_brxlg(addr, even_reg, Lloop1);
5384
5385 bind(Lslow);
5386 z_aghi(odd_reg, 16-1); // Last possible addr for slow loop.
5387 z_lghi(even_reg, 1);
5388 z_cgr(addr, odd_reg);
5389 z_brh(Lnotfound);
5390
5391 bind(Lloop2); // 1 byte per iteration.
5392 z_cli(Address(addr), 0x80);
5393 z_brnl(Ldone); // If found return 1.
5394 z_brxlg(addr, even_reg, Lloop2);
5395
5396 bind(Lnotfound);
5397 z_lhi(result, 0);
5398
5399 bind(Ldone);
5400
5401 BLOCK_COMMENT("} has_negatives");
5402
5403 return offset() - block_start;
5404 }
5405
5406 // kill: cnt1, cnt2, odd_reg, even_reg; early clobber: result
5407 unsigned int MacroAssembler::string_compare(Register str1, Register str2,
5408 Register cnt1, Register cnt2,
5409 Register odd_reg, Register even_reg, Register result, int ae) {
5410 int block_start = offset();
5411
5412 assert_different_registers(str1, cnt1, cnt2, odd_reg, even_reg, result);
5413 assert_different_registers(str2, cnt1, cnt2, odd_reg, even_reg, result);
5414
5415 // If strings are equal up to min length, return the length difference.
5416 const Register diff = result, // Pre-set result with length difference.
5417 min = cnt1, // min number of bytes
5418 tmp = cnt2;
5419
5420 // Note: Making use of the fact that compareTo(a, b) == -compareTo(b, a)
5421 // we interchange str1 and str2 in the UL case and negate the result.
5422 // Like this, str1 is always latin1 encoded, except for the UU case.
5423 // In addition, we need 0 (or sign which is 0) extend when using 64 bit register.
5424 const bool used_as_LU = (ae == StrIntrinsicNode::LU || ae == StrIntrinsicNode::UL);
5425
5426 BLOCK_COMMENT("string_compare {");
5427
5428 if (used_as_LU) {
5429 z_srl(cnt2, 1);
5430 }
5431
5432 // See if the lengths are different, and calculate min in cnt1.
5433 // Save diff in case we need it for a tie-breaker.
5434
5435 // diff = cnt1 - cnt2
5436 if (VM_Version::has_DistinctOpnds()) {
5437 z_srk(diff, cnt1, cnt2);
5438 } else {
5439 z_lr(diff, cnt1);
5440 z_sr(diff, cnt2);
5441 }
5442 if (str1 != str2) {
5443 if (VM_Version::has_LoadStoreConditional()) {
5444 z_locr(min, cnt2, Assembler::bcondHigh);
5445 } else {
5446 Label Lskip;
5447 z_brl(Lskip); // min ok if cnt1 < cnt2
5448 z_lr(min, cnt2); // min = cnt2
5449 bind(Lskip);
5450 }
5451 }
5452
5453 if (ae == StrIntrinsicNode::UU) {
5454 z_sra(diff, 1);
5455 }
5456 if (str1 != str2) {
5457 Label Ldone;
5458 if (used_as_LU) {
5459 // Loop which searches the first difference character by character.
5460 Label Lloop;
5461 const Register ind1 = Z_R1,
5462 ind2 = min;
5463 int stride1 = 1, stride2 = 2; // See comment above.
5464
5465 // ind1: index, even_reg: index increment, odd_reg: index limit
5466 z_llilf(ind1, (unsigned int)(-stride1));
5467 z_lhi(even_reg, stride1);
5468 add2reg(odd_reg, -stride1, min);
5469 clear_reg(ind2); // kills min
5470
5471 bind(Lloop);
5472 z_brxh(ind1, even_reg, Ldone);
5473 z_llc(tmp, Address(str1, ind1));
5474 z_llh(Z_R0, Address(str2, ind2));
5475 z_ahi(ind2, stride2);
5476 z_sr(tmp, Z_R0);
5477 z_bre(Lloop);
5478
5479 z_lr(result, tmp);
5480
5481 } else {
5482 // Use clcle in fast loop (only for same encoding).
5483 z_lgr(Z_R0, str1);
5484 z_lgr(even_reg, str2);
5485 z_llgfr(Z_R1, min);
5486 z_llgfr(odd_reg, min);
5487
5488 if (ae == StrIntrinsicNode::LL) {
5489 compare_long_ext(Z_R0, even_reg, 0);
5490 } else {
5491 compare_long_uni(Z_R0, even_reg, 0);
5492 }
5493 z_bre(Ldone);
5494 z_lgr(Z_R1, Z_R0);
5495 if (ae == StrIntrinsicNode::LL) {
5496 z_llc(Z_R0, Address(even_reg));
5497 z_llc(result, Address(Z_R1));
5498 } else {
5499 z_llh(Z_R0, Address(even_reg));
5500 z_llh(result, Address(Z_R1));
5501 }
5502 z_sr(result, Z_R0);
5503 }
5504
5505 // Otherwise, return the difference between the first mismatched chars.
5506 bind(Ldone);
5507 }
5508
5509 if (ae == StrIntrinsicNode::UL) {
5510 z_lcr(result, result); // Negate result (see note above).
5511 }
5512
5513 BLOCK_COMMENT("} string_compare");
5514
5515 return offset() - block_start;
5516 }
5517
5518 unsigned int MacroAssembler::array_equals(bool is_array_equ, Register ary1, Register ary2, Register limit,
5519 Register odd_reg, Register even_reg, Register result, bool is_byte) {
5520 int block_start = offset();
5521
5522 BLOCK_COMMENT("array_equals {");
5523
5524 assert_different_registers(ary1, limit, odd_reg, even_reg);
5525 assert_different_registers(ary2, limit, odd_reg, even_reg);
5526
5527 Label Ldone, Ldone_true, Ldone_false, Lclcle, CLC_template;
5528 int base_offset = 0;
5529
5530 if (ary1 != ary2) {
5531 if (is_array_equ) {
5532 base_offset = arrayOopDesc::base_offset_in_bytes(is_byte ? T_BYTE : T_CHAR);
5533
5534 // Return true if the same array.
5535 compareU64_and_branch(ary1, ary2, Assembler::bcondEqual, Ldone_true);
5536
5537 // Return false if one of them is NULL.
5538 compareU64_and_branch(ary1, (intptr_t)0, Assembler::bcondEqual, Ldone_false);
5539 compareU64_and_branch(ary2, (intptr_t)0, Assembler::bcondEqual, Ldone_false);
5540
5541 // Load the lengths of arrays.
5542 z_llgf(odd_reg, Address(ary1, arrayOopDesc::length_offset_in_bytes()));
5543
5544 // Return false if the two arrays are not equal length.
5545 z_c(odd_reg, Address(ary2, arrayOopDesc::length_offset_in_bytes()));
5546 z_brne(Ldone_false);
5547
5548 // string len in bytes (right operand)
5549 if (!is_byte) {
5550 z_chi(odd_reg, 128);
5551 z_sll(odd_reg, 1); // preserves flags
5552 z_brh(Lclcle);
5553 } else {
5554 compareU32_and_branch(odd_reg, (intptr_t)256, Assembler::bcondHigh, Lclcle);
5555 }
5556 } else {
5557 z_llgfr(odd_reg, limit); // Need to zero-extend prior to using the value.
5558 compareU32_and_branch(limit, (intptr_t)256, Assembler::bcondHigh, Lclcle);
5559 }
5560
5561
5562 // Use clc instruction for up to 256 bytes.
5563 {
5564 Register str1_reg = ary1,
5565 str2_reg = ary2;
5566 if (is_array_equ) {
5567 str1_reg = Z_R1;
5568 str2_reg = even_reg;
5569 add2reg(str1_reg, base_offset, ary1); // string addr (left operand)
5570 add2reg(str2_reg, base_offset, ary2); // string addr (right operand)
5571 }
5572 z_ahi(odd_reg, -1); // Clc uses decremented limit. Also compare result to 0.
5573 z_brl(Ldone_true);
5574 // Note: We could jump to the template if equal.
5575
5576 assert(VM_Version::has_ExecuteExtensions(), "unsupported hardware");
5577 z_exrl(odd_reg, CLC_template);
5578 z_bre(Ldone_true);
5579 // fall through
5580
5581 bind(Ldone_false);
5582 clear_reg(result);
5583 z_bru(Ldone);
5584
5585 bind(CLC_template);
5586 z_clc(0, 0, str1_reg, 0, str2_reg);
5587 }
5588
5589 // Use clcle instruction.
5590 {
5591 bind(Lclcle);
5592 add2reg(even_reg, base_offset, ary2); // string addr (right operand)
5593 add2reg(Z_R0, base_offset, ary1); // string addr (left operand)
5594
5595 z_lgr(Z_R1, odd_reg); // string len in bytes (left operand)
5596 if (is_byte) {
5597 compare_long_ext(Z_R0, even_reg, 0);
5598 } else {
5599 compare_long_uni(Z_R0, even_reg, 0);
5600 }
5601 z_lghi(result, 0); // Preserve flags.
5602 z_brne(Ldone);
5603 }
5604 }
5605 // fall through
5606
5607 bind(Ldone_true);
5608 z_lghi(result, 1); // All characters are equal.
5609 bind(Ldone);
5610
5611 BLOCK_COMMENT("} array_equals");
5612
5613 return offset() - block_start;
5614 }
5615
5616 // kill: haycnt, needlecnt, odd_reg, even_reg; early clobber: result
5617 unsigned int MacroAssembler::string_indexof(Register result, Register haystack, Register haycnt,
5618 Register needle, Register needlecnt, int needlecntval,
5619 Register odd_reg, Register even_reg, int ae) {
5620 int block_start = offset();
5621
5622 // Ensure 0<needlecnt<=haycnt in ideal graph as prerequisite!
5623 assert(ae != StrIntrinsicNode::LU, "Invalid encoding");
5624 const int h_csize = (ae == StrIntrinsicNode::LL) ? 1 : 2;
5625 const int n_csize = (ae == StrIntrinsicNode::UU) ? 2 : 1;
5626 Label L_needle1, L_Found, L_NotFound;
5627
5628 BLOCK_COMMENT("string_indexof {");
5629
5630 if (needle == haystack) {
5631 z_lhi(result, 0);
5632 } else {
5633
5634 // Load first character of needle (R0 used by search_string instructions).
5635 if (n_csize == 2) { z_llgh(Z_R0, Address(needle)); } else { z_llgc(Z_R0, Address(needle)); }
5636
5637 // Compute last haystack addr to use if no match gets found.
5638 if (needlecnt != noreg) { // variable needlecnt
5639 z_ahi(needlecnt, -1); // Remaining characters after first one.
5640 z_sr(haycnt, needlecnt); // Compute index succeeding last element to compare.
5641 if (n_csize == 2) { z_sll(needlecnt, 1); } // In bytes.
5642 } else { // constant needlecnt
5643 assert((needlecntval & 0x7fff) == needlecntval, "must be positive simm16 immediate");
5644 // Compute index succeeding last element to compare.
5645 if (needlecntval != 1) { z_ahi(haycnt, 1 - needlecntval); }
5646 }
5647
5648 z_llgfr(haycnt, haycnt); // Clear high half.
5649 z_lgr(result, haystack); // Final result will be computed from needle start pointer.
5650 if (h_csize == 2) { z_sll(haycnt, 1); } // Scale to number of bytes.
5651 z_agr(haycnt, haystack); // Point to address succeeding last element (haystack+scale*(haycnt-needlecnt+1)).
5652
5653 if (h_csize != n_csize) {
5654 assert(ae == StrIntrinsicNode::UL, "Invalid encoding");
5655
5656 if (needlecnt != noreg || needlecntval != 1) {
5657 if (needlecnt != noreg) {
5658 compare32_and_branch(needlecnt, (intptr_t)0, Assembler::bcondEqual, L_needle1);
5659 }
5660
5661 // Main Loop: UL version (now we have at least 2 characters).
5662 Label L_OuterLoop, L_InnerLoop, L_Skip;
5663 bind(L_OuterLoop); // Search for 1st 2 characters.
5664 z_lgr(Z_R1, haycnt);
5665 MacroAssembler::search_string_uni(Z_R1, result);
5666 z_brc(Assembler::bcondNotFound, L_NotFound);
5667 z_lgr(result, Z_R1);
5668
5669 z_lghi(Z_R1, n_csize);
5670 z_lghi(even_reg, h_csize);
5671 bind(L_InnerLoop);
5672 z_llgc(odd_reg, Address(needle, Z_R1));
5673 z_ch(odd_reg, Address(result, even_reg));
5674 z_brne(L_Skip);
5675 if (needlecnt != noreg) { z_cr(Z_R1, needlecnt); } else { z_chi(Z_R1, needlecntval - 1); }
5676 z_brnl(L_Found);
5677 z_aghi(Z_R1, n_csize);
5678 z_aghi(even_reg, h_csize);
5679 z_bru(L_InnerLoop);
5680
5681 bind(L_Skip);
5682 z_aghi(result, h_csize); // This is the new address we want to use for comparing.
5683 z_bru(L_OuterLoop);
5684 }
5685
5686 } else {
5687 const intptr_t needle_bytes = (n_csize == 2) ? ((needlecntval - 1) << 1) : (needlecntval - 1);
5688 Label L_clcle;
5689
5690 if (needlecnt != noreg || (needlecntval != 1 && needle_bytes <= 256)) {
5691 if (needlecnt != noreg) {
5692 compare32_and_branch(needlecnt, 256, Assembler::bcondHigh, L_clcle);
5693 z_ahi(needlecnt, -1); // remaining bytes -1 (for CLC)
5694 z_brl(L_needle1);
5695 }
5696
5697 // Main Loop: clc version (now we have at least 2 characters).
5698 Label L_OuterLoop, CLC_template;
5699 bind(L_OuterLoop); // Search for 1st 2 characters.
5700 z_lgr(Z_R1, haycnt);
5701 if (h_csize == 1) {
5702 MacroAssembler::search_string(Z_R1, result);
5703 } else {
5704 MacroAssembler::search_string_uni(Z_R1, result);
5705 }
5706 z_brc(Assembler::bcondNotFound, L_NotFound);
5707 z_lgr(result, Z_R1);
5708
5709 if (needlecnt != noreg) {
5710 assert(VM_Version::has_ExecuteExtensions(), "unsupported hardware");
5711 z_exrl(needlecnt, CLC_template);
5712 } else {
5713 z_clc(h_csize, needle_bytes -1, Z_R1, n_csize, needle);
5714 }
5715 z_bre(L_Found);
5716 z_aghi(result, h_csize); // This is the new address we want to use for comparing.
5717 z_bru(L_OuterLoop);
5718
5719 if (needlecnt != noreg) {
5720 bind(CLC_template);
5721 z_clc(h_csize, 0, Z_R1, n_csize, needle);
5722 }
5723 }
5724
5725 if (needlecnt != noreg || needle_bytes > 256) {
5726 bind(L_clcle);
5727
5728 // Main Loop: clcle version (now we have at least 256 bytes).
5729 Label L_OuterLoop, CLC_template;
5730 bind(L_OuterLoop); // Search for 1st 2 characters.
5731 z_lgr(Z_R1, haycnt);
5732 if (h_csize == 1) {
5733 MacroAssembler::search_string(Z_R1, result);
5734 } else {
5735 MacroAssembler::search_string_uni(Z_R1, result);
5736 }
5737 z_brc(Assembler::bcondNotFound, L_NotFound);
5738
5739 add2reg(Z_R0, n_csize, needle);
5740 add2reg(even_reg, h_csize, Z_R1);
5741 z_lgr(result, Z_R1);
5742 if (needlecnt != noreg) {
5743 z_llgfr(Z_R1, needlecnt); // needle len in bytes (left operand)
5744 z_llgfr(odd_reg, needlecnt);
5745 } else {
5746 load_const_optimized(Z_R1, needle_bytes);
5747 if (Immediate::is_simm16(needle_bytes)) { z_lghi(odd_reg, needle_bytes); } else { z_lgr(odd_reg, Z_R1); }
5748 }
5749 if (h_csize == 1) {
5750 compare_long_ext(Z_R0, even_reg, 0);
5751 } else {
5752 compare_long_uni(Z_R0, even_reg, 0);
5753 }
5754 z_bre(L_Found);
5755
5756 if (n_csize == 2) { z_llgh(Z_R0, Address(needle)); } else { z_llgc(Z_R0, Address(needle)); } // Reload.
5757 z_aghi(result, h_csize); // This is the new address we want to use for comparing.
5758 z_bru(L_OuterLoop);
5759 }
5760 }
5761
5762 if (needlecnt != noreg || needlecntval == 1) {
5763 bind(L_needle1);
5764
5765 // Single needle character version.
5766 if (h_csize == 1) {
5767 MacroAssembler::search_string(haycnt, result);
5768 } else {
5769 MacroAssembler::search_string_uni(haycnt, result);
5770 }
5771 z_lgr(result, haycnt);
5772 z_brc(Assembler::bcondFound, L_Found);
5773 }
5774
5775 bind(L_NotFound);
5776 add2reg(result, -1, haystack); // Return -1.
5777
5778 bind(L_Found); // Return index (or -1 in fallthrough case).
5779 z_sgr(result, haystack);
5780 if (h_csize == 2) { z_srag(result, result, exact_log2(sizeof(jchar))); }
5781 }
5782 BLOCK_COMMENT("} string_indexof");
5783
5784 return offset() - block_start;
5785 }
5786
5787 // early clobber: result
5788 unsigned int MacroAssembler::string_indexof_char(Register result, Register haystack, Register haycnt,
5789 Register needle, jchar needleChar, Register odd_reg, Register even_reg, bool is_byte) {
5790 int block_start = offset();
5791
5792 BLOCK_COMMENT("string_indexof_char {");
5793
5794 if (needle == haystack) {
5795 z_lhi(result, 0);
5796 } else {
5797
5798 Label Ldone;
5799
5800 z_llgfr(odd_reg, haycnt); // Preset loop ctr/searchrange end.
5801 if (needle == noreg) {
5802 load_const_optimized(Z_R0, (unsigned long)needleChar);
5803 } else {
5804 if (is_byte) {
5805 z_llgcr(Z_R0, needle); // First (and only) needle char.
5806 } else {
5807 z_llghr(Z_R0, needle); // First (and only) needle char.
5808 }
5809 }
5810
5811 if (!is_byte) {
5812 z_agr(odd_reg, odd_reg); // Calc #bytes to be processed with SRSTU.
5813 }
5814
5815 z_lgr(even_reg, haystack); // haystack addr
5816 z_agr(odd_reg, haystack); // First char after range end.
5817 z_lghi(result, -1);
5818
5819 if (is_byte) {
5820 MacroAssembler::search_string(odd_reg, even_reg);
5821 } else {
5822 MacroAssembler::search_string_uni(odd_reg, even_reg);
5823 }
5824 z_brc(Assembler::bcondNotFound, Ldone);
5825 if (is_byte) {
5826 if (VM_Version::has_DistinctOpnds()) {
5827 z_sgrk(result, odd_reg, haystack);
5828 } else {
5829 z_sgr(odd_reg, haystack);
5830 z_lgr(result, odd_reg);
5831 }
5832 } else {
5833 z_slgr(odd_reg, haystack);
5834 z_srlg(result, odd_reg, exact_log2(sizeof(jchar)));
5835 }
5836
5837 bind(Ldone);
5838 }
5839 BLOCK_COMMENT("} string_indexof_char");
5840
5841 return offset() - block_start;
5842 }
5843
5844
5845 //-------------------------------------------------
5846 // Constants (scalar and oop) in constant pool
5847 //-------------------------------------------------
5848
5849 // Add a non-relocated constant to the CP.
5850 int MacroAssembler::store_const_in_toc(AddressLiteral& val) {
5851 long value = val.value();
5852 address tocPos = long_constant(value);
5853
5854 if (tocPos != NULL) {
5855 int tocOffset = (int)(tocPos - code()->consts()->start());
5856 return tocOffset;
5857 }
5858 // Address_constant returned NULL, so no constant entry has been created.
5859 // In that case, we return a "fatal" offset, just in case that subsequently
5860 // generated access code is executed.
5861 return -1;
5862 }
5863
5864 // Returns the TOC offset where the address is stored.
5865 // Add a relocated constant to the CP.
5866 int MacroAssembler::store_oop_in_toc(AddressLiteral& oop) {
5867 // Use RelocationHolder::none for the constant pool entry.
5868 // Otherwise we will end up with a failing NativeCall::verify(x),
5869 // where x is the address of the constant pool entry.
5870 address tocPos = address_constant((address)oop.value(), RelocationHolder::none);
5871
5872 if (tocPos != NULL) {
5873 int tocOffset = (int)(tocPos - code()->consts()->start());
5874 RelocationHolder rsp = oop.rspec();
5875 Relocation *rel = rsp.reloc();
5876
5877 // Store toc_offset in relocation, used by call_far_patchable.
5878 if ((relocInfo::relocType)rel->type() == relocInfo::runtime_call_w_cp_type) {
5879 ((runtime_call_w_cp_Relocation *)(rel))->set_constant_pool_offset(tocOffset);
5880 }
5881 // Relocate at the load's pc.
5882 relocate(rsp);
5883
5884 return tocOffset;
5885 }
5886 // Address_constant returned NULL, so no constant entry has been created
5887 // in that case, we return a "fatal" offset, just in case that subsequently
5888 // generated access code is executed.
5889 return -1;
5890 }
5891
5892 bool MacroAssembler::load_const_from_toc(Register dst, AddressLiteral& a, Register Rtoc) {
5893 int tocOffset = store_const_in_toc(a);
5894 if (tocOffset == -1) return false;
5895 address tocPos = tocOffset + code()->consts()->start();
5896 assert((address)code()->consts()->start() != NULL, "Please add CP address");
5897
5898 load_long_pcrelative(dst, tocPos);
5899 return true;
5900 }
5901
5902 bool MacroAssembler::load_oop_from_toc(Register dst, AddressLiteral& a, Register Rtoc) {
5903 int tocOffset = store_oop_in_toc(a);
5904 if (tocOffset == -1) return false;
5905 address tocPos = tocOffset + code()->consts()->start();
5906 assert((address)code()->consts()->start() != NULL, "Please add CP address");
5907
5908 load_addr_pcrelative(dst, tocPos);
5909 return true;
5910 }
5911
5912 // If the instruction sequence at the given pc is a load_const_from_toc
5913 // sequence, return the value currently stored at the referenced position
5914 // in the TOC.
5915 intptr_t MacroAssembler::get_const_from_toc(address pc) {
5916
5917 assert(is_load_const_from_toc(pc), "must be load_const_from_pool");
5918
5919 long offset = get_load_const_from_toc_offset(pc);
5920 address dataLoc = NULL;
5921 if (is_load_const_from_toc_pcrelative(pc)) {
5922 dataLoc = pc + offset;
5923 } else {
5924 CodeBlob* cb = CodeCache::find_blob_unsafe(pc); // Else we get assertion if nmethod is zombie.
5925 assert(cb && cb->is_nmethod(), "sanity");
5926 nmethod* nm = (nmethod*)cb;
5927 dataLoc = nm->ctable_begin() + offset;
5928 }
5929 return *(intptr_t *)dataLoc;
5930 }
5931
5932 // If the instruction sequence at the given pc is a load_const_from_toc
5933 // sequence, copy the passed-in new_data value into the referenced
5934 // position in the TOC.
5935 void MacroAssembler::set_const_in_toc(address pc, unsigned long new_data, CodeBlob *cb) {
5936 assert(is_load_const_from_toc(pc), "must be load_const_from_pool");
5937
5938 long offset = MacroAssembler::get_load_const_from_toc_offset(pc);
5939 address dataLoc = NULL;
5940 if (is_load_const_from_toc_pcrelative(pc)) {
5941 dataLoc = pc+offset;
5942 } else {
5943 nmethod* nm = CodeCache::find_nmethod(pc);
5944 assert((cb == NULL) || (nm == (nmethod*)cb), "instruction address should be in CodeBlob");
5945 dataLoc = nm->ctable_begin() + offset;
5946 }
5947 if (*(unsigned long *)dataLoc != new_data) { // Prevent cache invalidation: update only if necessary.
5948 *(unsigned long *)dataLoc = new_data;
5949 }
5950 }
5951
5952 // Dynamic TOC. Getter must only be called if "a" is a load_const_from_toc
5953 // site. Verify by calling is_load_const_from_toc() before!!
5954 // Offset is +/- 2**32 -> use long.
5955 long MacroAssembler::get_load_const_from_toc_offset(address a) {
5956 assert(is_load_const_from_toc_pcrelative(a), "expected pc relative load");
5957 // expected code sequence:
5958 // z_lgrl(t, simm32); len = 6
5959 unsigned long inst;
5960 unsigned int len = get_instruction(a, &inst);
5961 return get_pcrel_offset(inst);
5962 }
5963
5964 //**********************************************************************************
5965 // inspection of generated instruction sequences for a particular pattern
5966 //**********************************************************************************
5967
5968 bool MacroAssembler::is_load_const_from_toc_pcrelative(address a) {
5969 #ifdef ASSERT
5970 unsigned long inst;
5971 unsigned int len = get_instruction(a+2, &inst);
5972 if ((len == 6) && is_load_pcrelative_long(a) && is_call_pcrelative_long(inst)) {
5973 const int range = 128;
5974 Assembler::dump_code_range(tty, a, range, "instr(a) == z_lgrl && instr(a+2) == z_brasl");
5975 VM_Version::z_SIGSEGV();
5976 }
5977 #endif
5978 // expected code sequence:
5979 // z_lgrl(t, relAddr32); len = 6
5980 //TODO: verify accessed data is in CP, if possible.
5981 return is_load_pcrelative_long(a); // TODO: might be too general. Currently, only lgrl is used.
5982 }
5983
5984 bool MacroAssembler::is_load_const_from_toc_call(address a) {
5985 return is_load_const_from_toc(a) && is_call_byregister(a + load_const_from_toc_size());
5986 }
5987
5988 bool MacroAssembler::is_load_const_call(address a) {
5989 return is_load_const(a) && is_call_byregister(a + load_const_size());
5990 }
5991
5992 //-------------------------------------------------
5993 // Emitters for some really CICS instructions
5994 //-------------------------------------------------
5995
5996 void MacroAssembler::move_long_ext(Register dst, Register src, unsigned int pad) {
5997 assert(dst->encoding()%2==0, "must be an even/odd register pair");
5998 assert(src->encoding()%2==0, "must be an even/odd register pair");
5999 assert(pad<256, "must be a padding BYTE");
6000
6001 Label retry;
6002 bind(retry);
6003 Assembler::z_mvcle(dst, src, pad);
6004 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6005 }
6006
6007 void MacroAssembler::compare_long_ext(Register left, Register right, unsigned int pad) {
6008 assert(left->encoding() % 2 == 0, "must be an even/odd register pair");
6009 assert(right->encoding() % 2 == 0, "must be an even/odd register pair");
6010 assert(pad<256, "must be a padding BYTE");
6011
6012 Label retry;
6013 bind(retry);
6014 Assembler::z_clcle(left, right, pad, Z_R0);
6015 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6016 }
6017
6018 void MacroAssembler::compare_long_uni(Register left, Register right, unsigned int pad) {
6019 assert(left->encoding() % 2 == 0, "must be an even/odd register pair");
6020 assert(right->encoding() % 2 == 0, "must be an even/odd register pair");
6021 assert(pad<=0xfff, "must be a padding HALFWORD");
6022 assert(VM_Version::has_ETF2(), "instruction must be available");
6023
6024 Label retry;
6025 bind(retry);
6026 Assembler::z_clclu(left, right, pad, Z_R0);
6027 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6028 }
6029
6030 void MacroAssembler::search_string(Register end, Register start) {
6031 assert(end->encoding() != 0, "end address must not be in R0");
6032 assert(start->encoding() != 0, "start address must not be in R0");
6033
6034 Label retry;
6035 bind(retry);
6036 Assembler::z_srst(end, start);
6037 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6038 }
6039
6040 void MacroAssembler::search_string_uni(Register end, Register start) {
6041 assert(end->encoding() != 0, "end address must not be in R0");
6042 assert(start->encoding() != 0, "start address must not be in R0");
6043 assert(VM_Version::has_ETF3(), "instruction must be available");
6044
6045 Label retry;
6046 bind(retry);
6047 Assembler::z_srstu(end, start);
6048 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6049 }
6050
6051 void MacroAssembler::kmac(Register srcBuff) {
6052 assert(srcBuff->encoding() != 0, "src buffer address can't be in Z_R0");
6053 assert(srcBuff->encoding() % 2 == 0, "src buffer/len must be an even/odd register pair");
6054
6055 Label retry;
6056 bind(retry);
6057 Assembler::z_kmac(Z_R0, srcBuff);
6058 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6059 }
6060
6061 void MacroAssembler::kimd(Register srcBuff) {
6062 assert(srcBuff->encoding() != 0, "src buffer address can't be in Z_R0");
6063 assert(srcBuff->encoding() % 2 == 0, "src buffer/len must be an even/odd register pair");
6064
6065 Label retry;
6066 bind(retry);
6067 Assembler::z_kimd(Z_R0, srcBuff);
6068 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6069 }
6070
6071 void MacroAssembler::klmd(Register srcBuff) {
6072 assert(srcBuff->encoding() != 0, "src buffer address can't be in Z_R0");
6073 assert(srcBuff->encoding() % 2 == 0, "src buffer/len must be an even/odd register pair");
6074
6075 Label retry;
6076 bind(retry);
6077 Assembler::z_klmd(Z_R0, srcBuff);
6078 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6079 }
6080
6081 void MacroAssembler::km(Register dstBuff, Register srcBuff) {
6082 // DstBuff and srcBuff are allowed to be the same register (encryption in-place).
6083 // DstBuff and srcBuff storage must not overlap destructively, and neither must overlap the parameter block.
6084 assert(srcBuff->encoding() != 0, "src buffer address can't be in Z_R0");
6085 assert(dstBuff->encoding() % 2 == 0, "dst buffer addr must be an even register");
6086 assert(srcBuff->encoding() % 2 == 0, "src buffer addr/len must be an even/odd register pair");
6087
6088 Label retry;
6089 bind(retry);
6090 Assembler::z_km(dstBuff, srcBuff);
6091 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6092 }
6093
6094 void MacroAssembler::kmc(Register dstBuff, Register srcBuff) {
6095 // DstBuff and srcBuff are allowed to be the same register (encryption in-place).
6096 // DstBuff and srcBuff storage must not overlap destructively, and neither must overlap the parameter block.
6097 assert(srcBuff->encoding() != 0, "src buffer address can't be in Z_R0");
6098 assert(dstBuff->encoding() % 2 == 0, "dst buffer addr must be an even register");
6099 assert(srcBuff->encoding() % 2 == 0, "src buffer addr/len must be an even/odd register pair");
6100
6101 Label retry;
6102 bind(retry);
6103 Assembler::z_kmc(dstBuff, srcBuff);
6104 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6105 }
6106
6107 void MacroAssembler::cksm(Register crcBuff, Register srcBuff) {
6108 assert(srcBuff->encoding() % 2 == 0, "src buffer addr/len must be an even/odd register pair");
6109
6110 Label retry;
6111 bind(retry);
6112 Assembler::z_cksm(crcBuff, srcBuff);
6113 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6114 }
6115
6116 void MacroAssembler::translate_oo(Register r1, Register r2, uint m3) {
6117 assert(r1->encoding() % 2 == 0, "dst addr/src len must be an even/odd register pair");
6118 assert((m3 & 0b1110) == 0, "Unused mask bits must be zero");
6119
6120 Label retry;
6121 bind(retry);
6122 Assembler::z_troo(r1, r2, m3);
6123 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6124 }
6125
6126 void MacroAssembler::translate_ot(Register r1, Register r2, uint m3) {
6127 assert(r1->encoding() % 2 == 0, "dst addr/src len must be an even/odd register pair");
6128 assert((m3 & 0b1110) == 0, "Unused mask bits must be zero");
6129
6130 Label retry;
6131 bind(retry);
6132 Assembler::z_trot(r1, r2, m3);
6133 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6134 }
6135
6136 void MacroAssembler::translate_to(Register r1, Register r2, uint m3) {
6137 assert(r1->encoding() % 2 == 0, "dst addr/src len must be an even/odd register pair");
6138 assert((m3 & 0b1110) == 0, "Unused mask bits must be zero");
6139
6140 Label retry;
6141 bind(retry);
6142 Assembler::z_trto(r1, r2, m3);
6143 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6144 }
6145
6146 void MacroAssembler::translate_tt(Register r1, Register r2, uint m3) {
6147 assert(r1->encoding() % 2 == 0, "dst addr/src len must be an even/odd register pair");
6148 assert((m3 & 0b1110) == 0, "Unused mask bits must be zero");
6149
6150 Label retry;
6151 bind(retry);
6152 Assembler::z_trtt(r1, r2, m3);
6153 Assembler::z_brc(Assembler::bcondOverflow /* CC==3 (iterate) */, retry);
6154 }
6155
6156
6157 void MacroAssembler::generate_type_profiling(const Register Rdata,
6158 const Register Rreceiver_klass,
6159 const Register Rwanted_receiver_klass,
6160 const Register Rmatching_row,
6161 bool is_virtual_call) {
6162 const int row_size = in_bytes(ReceiverTypeData::receiver_offset(1)) -
6163 in_bytes(ReceiverTypeData::receiver_offset(0));
6164 const int num_rows = ReceiverTypeData::row_limit();
6165 NearLabel found_free_row;
6166 NearLabel do_increment;
6167 NearLabel found_no_slot;
6168
6169 BLOCK_COMMENT("type profiling {");
6170
6171 // search for:
6172 // a) The type given in Rwanted_receiver_klass.
6173 // b) The *first* empty row.
6174
6175 // First search for a) only, just running over b) with no regard.
6176 // This is possible because
6177 // wanted_receiver_class == receiver_class && wanted_receiver_class == 0
6178 // is never true (receiver_class can't be zero).
6179 for (int row_num = 0; row_num < num_rows; row_num++) {
6180 // Row_offset should be a well-behaved positive number. The generated code relies
6181 // on that wrt constant code size. Add2reg can handle all row_offset values, but
6182 // will have to vary generated code size.
6183 int row_offset = in_bytes(ReceiverTypeData::receiver_offset(row_num));
6184 assert(Displacement::is_shortDisp(row_offset), "Limitation of generated code");
6185
6186 // Is Rwanted_receiver_klass in this row?
6187 if (VM_Version::has_CompareBranch()) {
6188 z_lg(Rwanted_receiver_klass, row_offset, Z_R0, Rdata);
6189 // Rmatching_row = Rdata + row_offset;
6190 add2reg(Rmatching_row, row_offset, Rdata);
6191 // if (*row_recv == (intptr_t) receiver_klass) goto fill_existing_slot;
6192 compare64_and_branch(Rwanted_receiver_klass, Rreceiver_klass, Assembler::bcondEqual, do_increment);
6193 } else {
6194 add2reg(Rmatching_row, row_offset, Rdata);
6195 z_cg(Rreceiver_klass, row_offset, Z_R0, Rdata);
6196 z_bre(do_increment);
6197 }
6198 }
6199
6200 // Now that we did not find a match, let's search for b).
6201
6202 // We could save the first calculation of Rmatching_row if we woud search for a) in reverse order.
6203 // We would then end up here with Rmatching_row containing the value for row_num == 0.
6204 // We would not see much benefit, if any at all, because the CPU can schedule
6205 // two instructions together with a branch anyway.
6206 for (int row_num = 0; row_num < num_rows; row_num++) {
6207 int row_offset = in_bytes(ReceiverTypeData::receiver_offset(row_num));
6208
6209 // Has this row a zero receiver_klass, i.e. is it empty?
6210 if (VM_Version::has_CompareBranch()) {
6211 z_lg(Rwanted_receiver_klass, row_offset, Z_R0, Rdata);
6212 // Rmatching_row = Rdata + row_offset
6213 add2reg(Rmatching_row, row_offset, Rdata);
6214 // if (*row_recv == (intptr_t) 0) goto found_free_row
6215 compare64_and_branch(Rwanted_receiver_klass, (intptr_t)0, Assembler::bcondEqual, found_free_row);
6216 } else {
6217 add2reg(Rmatching_row, row_offset, Rdata);
6218 load_and_test_long(Rwanted_receiver_klass, Address(Rdata, row_offset));
6219 z_bre(found_free_row); // zero -> Found a free row.
6220 }
6221 }
6222
6223 // No match, no empty row found.
6224 // Increment total counter to indicate polymorphic case.
6225 if (is_virtual_call) {
6226 add2mem_64(Address(Rdata, CounterData::count_offset()), 1, Rmatching_row);
6227 }
6228 z_bru(found_no_slot);
6229
6230 // Here we found an empty row, but we have not found Rwanted_receiver_klass.
6231 // Rmatching_row holds the address to the first empty row.
6232 bind(found_free_row);
6233 // Store receiver_klass into empty slot.
6234 z_stg(Rreceiver_klass, 0, Z_R0, Rmatching_row);
6235
6236 // Increment the counter of Rmatching_row.
6237 bind(do_increment);
6238 ByteSize counter_offset = ReceiverTypeData::receiver_count_offset(0) - ReceiverTypeData::receiver_offset(0);
6239 add2mem_64(Address(Rmatching_row, counter_offset), 1, Rdata);
6240
6241 bind(found_no_slot);
6242
6243 BLOCK_COMMENT("} type profiling");
6244 }
6245
6246 //---------------------------------------
6247 // Helpers for Intrinsic Emitters
6248 //---------------------------------------
6249
6250 /**
6251 * uint32_t crc;
6252 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
6253 */
6254 void MacroAssembler::fold_byte_crc32(Register crc, Register val, Register table, Register tmp) {
6255 assert_different_registers(crc, table, tmp);
6256 assert_different_registers(val, table);
6257 if (crc == val) { // Must rotate first to use the unmodified value.
6258 rotate_then_insert(tmp, val, 56-2, 63-2, 2, true); // Insert byte 7 of val, shifted left by 2, into byte 6..7 of tmp, clear the rest.
6259 z_srl(crc, 8); // Unsigned shift, clear leftmost 8 bits.
6260 } else {
6261 z_srl(crc, 8); // Unsigned shift, clear leftmost 8 bits.
6262 rotate_then_insert(tmp, val, 56-2, 63-2, 2, true); // Insert byte 7 of val, shifted left by 2, into byte 6..7 of tmp, clear the rest.
6263 }
6264 z_x(crc, Address(table, tmp, 0));
6265 }
6266
6267 /**
6268 * uint32_t crc;
6269 * timesXtoThe32[crc & 0xFF] ^ (crc >> 8);
6270 */
6271 void MacroAssembler::fold_8bit_crc32(Register crc, Register table, Register tmp) {
6272 fold_byte_crc32(crc, crc, table, tmp);
6273 }
6274
6275 /**
6276 * Emits code to update CRC-32 with a byte value according to constants in table.
6277 *
6278 * @param [in,out]crc Register containing the crc.
6279 * @param [in]val Register containing the byte to fold into the CRC.
6280 * @param [in]table Register containing the table of crc constants.
6281 *
6282 * uint32_t crc;
6283 * val = crc_table[(val ^ crc) & 0xFF];
6284 * crc = val ^ (crc >> 8);
6285 */
6286 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
6287 z_xr(val, crc);
6288 fold_byte_crc32(crc, val, table, val);
6289 }
6290
6291
6292 /**
6293 * @param crc register containing existing CRC (32-bit)
6294 * @param buf register pointing to input byte buffer (byte*)
6295 * @param len register containing number of bytes
6296 * @param table register pointing to CRC table
6297 */
6298 void MacroAssembler::update_byteLoop_crc32(Register crc, Register buf, Register len, Register table, Register data) {
6299 assert_different_registers(crc, buf, len, table, data);
6300
6301 Label L_mainLoop, L_done;
6302 const int mainLoop_stepping = 1;
6303
6304 // Process all bytes in a single-byte loop.
6305 z_ltr(len, len);
6306 z_brnh(L_done);
6307
6308 bind(L_mainLoop);
6309 z_llgc(data, Address(buf, (intptr_t)0));// Current byte of input buffer (zero extended). Avoids garbage in upper half of register.
6310 add2reg(buf, mainLoop_stepping); // Advance buffer position.
6311 update_byte_crc32(crc, data, table);
6312 z_brct(len, L_mainLoop); // Iterate.
6313
6314 bind(L_done);
6315 }
6316
6317 /**
6318 * Emits code to update CRC-32 with a 4-byte value according to constants in table.
6319 * Implementation according to jdk/src/share/native/java/util/zip/zlib-1.2.8/crc32.c.
6320 *
6321 */
6322 void MacroAssembler::update_1word_crc32(Register crc, Register buf, Register table, int bufDisp, int bufInc,
6323 Register t0, Register t1, Register t2, Register t3) {
6324 // This is what we implement (the DOBIG4 part):
6325 //
6326 // #define DOBIG4 c ^= *++buf4; \
6327 // c = crc_table[4][c & 0xff] ^ crc_table[5][(c >> 8) & 0xff] ^ \
6328 // crc_table[6][(c >> 16) & 0xff] ^ crc_table[7][c >> 24]
6329 // #define DOBIG32 DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4
6330 // Pre-calculate (constant) column offsets, use columns 4..7 for big-endian.
6331 const int ix0 = 4*(4*CRC32_COLUMN_SIZE);
6332 const int ix1 = 5*(4*CRC32_COLUMN_SIZE);
6333 const int ix2 = 6*(4*CRC32_COLUMN_SIZE);
6334 const int ix3 = 7*(4*CRC32_COLUMN_SIZE);
6335
6336 // XOR crc with next four bytes of buffer.
6337 lgr_if_needed(t0, crc);
6338 z_x(t0, Address(buf, bufDisp));
6339 if (bufInc != 0) {
6340 add2reg(buf, bufInc);
6341 }
6342
6343 // Chop crc into 4 single-byte pieces, shifted left 2 bits, to form the table indices.
6344 rotate_then_insert(t3, t0, 56-2, 63-2, 2, true); // ((c >> 0) & 0xff) << 2
6345 rotate_then_insert(t2, t0, 56-2, 63-2, 2-8, true); // ((c >> 8) & 0xff) << 2
6346 rotate_then_insert(t1, t0, 56-2, 63-2, 2-16, true); // ((c >> 16) & 0xff) << 2
6347 rotate_then_insert(t0, t0, 56-2, 63-2, 2-24, true); // ((c >> 24) & 0xff) << 2
6348
6349 // XOR indexed table values to calculate updated crc.
6350 z_ly(t2, Address(table, t2, (intptr_t)ix1));
6351 z_ly(t0, Address(table, t0, (intptr_t)ix3));
6352 z_xy(t2, Address(table, t3, (intptr_t)ix0));
6353 z_xy(t0, Address(table, t1, (intptr_t)ix2));
6354 z_xr(t0, t2); // Now t0 contains the updated CRC value.
6355 lgr_if_needed(crc, t0);
6356 }
6357
6358 /**
6359 * @param crc register containing existing CRC (32-bit)
6360 * @param buf register pointing to input byte buffer (byte*)
6361 * @param len register containing number of bytes
6362 * @param table register pointing to CRC table
6363 *
6364 * uses Z_R10..Z_R13 as work register. Must be saved/restored by caller!
6365 */
6366 void MacroAssembler::kernel_crc32_2word(Register crc, Register buf, Register len, Register table,
6367 Register t0, Register t1, Register t2, Register t3,
6368 bool invertCRC) {
6369 assert_different_registers(crc, buf, len, table);
6370
6371 Label L_mainLoop, L_tail;
6372 Register data = t0;
6373 Register ctr = Z_R0;
6374 const int mainLoop_stepping = 8;
6375 const int tailLoop_stepping = 1;
6376 const int log_stepping = exact_log2(mainLoop_stepping);
6377
6378 // Don't test for len <= 0 here. This pathological case should not occur anyway.
6379 // Optimizing for it by adding a test and a branch seems to be a waste of CPU cycles.
6380 // The situation itself is detected and handled correctly by the conditional branches
6381 // following aghi(len, -stepping) and aghi(len, +stepping).
6382
6383 if (invertCRC) {
6384 not_(crc, noreg, false); // 1s complement of crc
6385 }
6386
6387 #if 0
6388 {
6389 // Pre-mainLoop alignment did not show any positive effect on performance.
6390 // We leave the code in for reference. Maybe the vector instructions in z13 depend on alignment.
6391
6392 z_cghi(len, mainLoop_stepping); // Alignment is useless for short data streams.
6393 z_brnh(L_tail);
6394
6395 // Align buf to word (4-byte) boundary.
6396 z_lcr(ctr, buf);
6397 rotate_then_insert(ctr, ctr, 62, 63, 0, true); // TODO: should set cc
6398 z_sgfr(len, ctr); // Remaining len after alignment.
6399
6400 update_byteLoop_crc32(crc, buf, ctr, table, data);
6401 }
6402 #endif
6403
6404 // Check for short (<mainLoop_stepping bytes) buffer.
6405 z_srag(ctr, len, log_stepping);
6406 z_brnh(L_tail);
6407
6408 z_lrvr(crc, crc); // Revert byte order because we are dealing with big-endian data.
6409 rotate_then_insert(len, len, 64-log_stepping, 63, 0, true); // #bytes for tailLoop
6410
6411 BIND(L_mainLoop);
6412 update_1word_crc32(crc, buf, table, 0, 0, crc, t1, t2, t3);
6413 update_1word_crc32(crc, buf, table, 4, mainLoop_stepping, crc, t1, t2, t3);
6414 z_brct(ctr, L_mainLoop); // Iterate.
6415
6416 z_lrvr(crc, crc); // Revert byte order back to original.
6417
6418 // Process last few (<8) bytes of buffer.
6419 BIND(L_tail);
6420 update_byteLoop_crc32(crc, buf, len, table, data);
6421
6422 if (invertCRC) {
6423 not_(crc, noreg, false); // 1s complement of crc
6424 }
6425 }
6426
6427 /**
6428 * @param crc register containing existing CRC (32-bit)
6429 * @param buf register pointing to input byte buffer (byte*)
6430 * @param len register containing number of bytes
6431 * @param table register pointing to CRC table
6432 *
6433 * uses Z_R10..Z_R13 as work register. Must be saved/restored by caller!
6434 */
6435 void MacroAssembler::kernel_crc32_1word(Register crc, Register buf, Register len, Register table,
6436 Register t0, Register t1, Register t2, Register t3,
6437 bool invertCRC) {
6438 assert_different_registers(crc, buf, len, table);
6439
6440 Label L_mainLoop, L_tail;
6441 Register data = t0;
6442 Register ctr = Z_R0;
6443 const int mainLoop_stepping = 4;
6444 const int log_stepping = exact_log2(mainLoop_stepping);
6445
6446 // Don't test for len <= 0 here. This pathological case should not occur anyway.
6447 // Optimizing for it by adding a test and a branch seems to be a waste of CPU cycles.
6448 // The situation itself is detected and handled correctly by the conditional branches
6449 // following aghi(len, -stepping) and aghi(len, +stepping).
6450
6451 if (invertCRC) {
6452 not_(crc, noreg, false); // 1s complement of crc
6453 }
6454
6455 // Check for short (<4 bytes) buffer.
6456 z_srag(ctr, len, log_stepping);
6457 z_brnh(L_tail);
6458
6459 z_lrvr(crc, crc); // Revert byte order because we are dealing with big-endian data.
6460 rotate_then_insert(len, len, 64-log_stepping, 63, 0, true); // #bytes for tailLoop
6461
6462 BIND(L_mainLoop);
6463 update_1word_crc32(crc, buf, table, 0, mainLoop_stepping, crc, t1, t2, t3);
6464 z_brct(ctr, L_mainLoop); // Iterate.
6465
6466 z_lrvr(crc, crc); // Revert byte order back to original.
6467
6468 // Process last few (<8) bytes of buffer.
6469 BIND(L_tail);
6470 update_byteLoop_crc32(crc, buf, len, table, data);
6471
6472 if (invertCRC) {
6473 not_(crc, noreg, false); // 1s complement of crc
6474 }
6475 }
6476
6477 /**
6478 * @param crc register containing existing CRC (32-bit)
6479 * @param buf register pointing to input byte buffer (byte*)
6480 * @param len register containing number of bytes
6481 * @param table register pointing to CRC table
6482 */
6483 void MacroAssembler::kernel_crc32_1byte(Register crc, Register buf, Register len, Register table,
6484 Register t0, Register t1, Register t2, Register t3,
6485 bool invertCRC) {
6486 assert_different_registers(crc, buf, len, table);
6487 Register data = t0;
6488
6489 if (invertCRC) {
6490 not_(crc, noreg, false); // 1s complement of crc
6491 }
6492
6493 update_byteLoop_crc32(crc, buf, len, table, data);
6494
6495 if (invertCRC) {
6496 not_(crc, noreg, false); // 1s complement of crc
6497 }
6498 }
6499
6500 void MacroAssembler::kernel_crc32_singleByte(Register crc, Register buf, Register len, Register table, Register tmp,
6501 bool invertCRC) {
6502 assert_different_registers(crc, buf, len, table, tmp);
6503
6504 if (invertCRC) {
6505 not_(crc, noreg, false); // 1s complement of crc
6506 }
6507
6508 z_llgc(tmp, Address(buf, (intptr_t)0)); // Current byte of input buffer (zero extended). Avoids garbage in upper half of register.
6509 update_byte_crc32(crc, tmp, table);
6510
6511 if (invertCRC) {
6512 not_(crc, noreg, false); // 1s complement of crc
6513 }
6514 }
6515
6516 void MacroAssembler::kernel_crc32_singleByteReg(Register crc, Register val, Register table,
6517 bool invertCRC) {
6518 assert_different_registers(crc, val, table);
6519
6520 if (invertCRC) {
6521 not_(crc, noreg, false); // 1s complement of crc
6522 }
6523
6524 update_byte_crc32(crc, val, table);
6525
6526 if (invertCRC) {
6527 not_(crc, noreg, false); // 1s complement of crc
6528 }
6529 }
6530
6531 //
6532 // Code for BigInteger::multiplyToLen() intrinsic.
6533 //
6534
6535 // dest_lo += src1 + src2
6536 // dest_hi += carry1 + carry2
6537 // Z_R7 is destroyed !
6538 void MacroAssembler::add2_with_carry(Register dest_hi, Register dest_lo,
6539 Register src1, Register src2) {
6540 clear_reg(Z_R7);
6541 z_algr(dest_lo, src1);
6542 z_alcgr(dest_hi, Z_R7);
6543 z_algr(dest_lo, src2);
6544 z_alcgr(dest_hi, Z_R7);
6545 }
6546
6547 // Multiply 64 bit by 64 bit first loop.
6548 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart,
6549 Register x_xstart,
6550 Register y, Register y_idx,
6551 Register z,
6552 Register carry,
6553 Register product,
6554 Register idx, Register kdx) {
6555 // jlong carry, x[], y[], z[];
6556 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx--, kdx--) {
6557 // huge_128 product = y[idx] * x[xstart] + carry;
6558 // z[kdx] = (jlong)product;
6559 // carry = (jlong)(product >>> 64);
6560 // }
6561 // z[xstart] = carry;
6562
6563 Label L_first_loop, L_first_loop_exit;
6564 Label L_one_x, L_one_y, L_multiply;
6565
6566 z_aghi(xstart, -1);
6567 z_brl(L_one_x); // Special case: length of x is 1.
6568
6569 // Load next two integers of x.
6570 z_sllg(Z_R1_scratch, xstart, LogBytesPerInt);
6571 mem2reg_opt(x_xstart, Address(x, Z_R1_scratch, 0));
6572
6573
6574 bind(L_first_loop);
6575
6576 z_aghi(idx, -1);
6577 z_brl(L_first_loop_exit);
6578 z_aghi(idx, -1);
6579 z_brl(L_one_y);
6580
6581 // Load next two integers of y.
6582 z_sllg(Z_R1_scratch, idx, LogBytesPerInt);
6583 mem2reg_opt(y_idx, Address(y, Z_R1_scratch, 0));
6584
6585
6586 bind(L_multiply);
6587
6588 Register multiplicand = product->successor();
6589 Register product_low = multiplicand;
6590
6591 lgr_if_needed(multiplicand, x_xstart);
6592 z_mlgr(product, y_idx); // multiplicand * y_idx -> product::multiplicand
6593 clear_reg(Z_R7);
6594 z_algr(product_low, carry); // Add carry to result.
6595 z_alcgr(product, Z_R7); // Add carry of the last addition.
6596 add2reg(kdx, -2);
6597
6598 // Store result.
6599 z_sllg(Z_R7, kdx, LogBytesPerInt);
6600 reg2mem_opt(product_low, Address(z, Z_R7, 0));
6601 lgr_if_needed(carry, product);
6602 z_bru(L_first_loop);
6603
6604
6605 bind(L_one_y); // Load one 32 bit portion of y as (0,value).
6606
6607 clear_reg(y_idx);
6608 mem2reg_opt(y_idx, Address(y, (intptr_t) 0), false);
6609 z_bru(L_multiply);
6610
6611
6612 bind(L_one_x); // Load one 32 bit portion of x as (0,value).
6613
6614 clear_reg(x_xstart);
6615 mem2reg_opt(x_xstart, Address(x, (intptr_t) 0), false);
6616 z_bru(L_first_loop);
6617
6618 bind(L_first_loop_exit);
6619 }
6620
6621 // Multiply 64 bit by 64 bit and add 128 bit.
6622 void MacroAssembler::multiply_add_128_x_128(Register x_xstart, Register y,
6623 Register z,
6624 Register yz_idx, Register idx,
6625 Register carry, Register product,
6626 int offset) {
6627 // huge_128 product = (y[idx] * x_xstart) + z[kdx] + carry;
6628 // z[kdx] = (jlong)product;
6629
6630 Register multiplicand = product->successor();
6631 Register product_low = multiplicand;
6632
6633 z_sllg(Z_R7, idx, LogBytesPerInt);
6634 mem2reg_opt(yz_idx, Address(y, Z_R7, offset));
6635
6636 lgr_if_needed(multiplicand, x_xstart);
6637 z_mlgr(product, yz_idx); // multiplicand * yz_idx -> product::multiplicand
6638 mem2reg_opt(yz_idx, Address(z, Z_R7, offset));
6639
6640 add2_with_carry(product, product_low, carry, yz_idx);
6641
6642 z_sllg(Z_R7, idx, LogBytesPerInt);
6643 reg2mem_opt(product_low, Address(z, Z_R7, offset));
6644
6645 }
6646
6647 // Multiply 128 bit by 128 bit. Unrolled inner loop.
6648 void MacroAssembler::multiply_128_x_128_loop(Register x_xstart,
6649 Register y, Register z,
6650 Register yz_idx, Register idx,
6651 Register jdx,
6652 Register carry, Register product,
6653 Register carry2) {
6654 // jlong carry, x[], y[], z[];
6655 // int kdx = ystart+1;
6656 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
6657 // huge_128 product = (y[idx+1] * x_xstart) + z[kdx+idx+1] + carry;
6658 // z[kdx+idx+1] = (jlong)product;
6659 // jlong carry2 = (jlong)(product >>> 64);
6660 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry2;
6661 // z[kdx+idx] = (jlong)product;
6662 // carry = (jlong)(product >>> 64);
6663 // }
6664 // idx += 2;
6665 // if (idx > 0) {
6666 // product = (y[idx] * x_xstart) + z[kdx+idx] + carry;
6667 // z[kdx+idx] = (jlong)product;
6668 // carry = (jlong)(product >>> 64);
6669 // }
6670
6671 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
6672
6673 // scale the index
6674 lgr_if_needed(jdx, idx);
6675 and_imm(jdx, 0xfffffffffffffffcL);
6676 rshift(jdx, 2);
6677
6678
6679 bind(L_third_loop);
6680
6681 z_aghi(jdx, -1);
6682 z_brl(L_third_loop_exit);
6683 add2reg(idx, -4);
6684
6685 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 8);
6686 lgr_if_needed(carry2, product);
6687
6688 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry2, product, 0);
6689 lgr_if_needed(carry, product);
6690 z_bru(L_third_loop);
6691
6692
6693 bind(L_third_loop_exit); // Handle any left-over operand parts.
6694
6695 and_imm(idx, 0x3);
6696 z_brz(L_post_third_loop_done);
6697
6698 Label L_check_1;
6699
6700 z_aghi(idx, -2);
6701 z_brl(L_check_1);
6702
6703 multiply_add_128_x_128(x_xstart, y, z, yz_idx, idx, carry, product, 0);
6704 lgr_if_needed(carry, product);
6705
6706
6707 bind(L_check_1);
6708
6709 add2reg(idx, 0x2);
6710 and_imm(idx, 0x1);
6711 z_aghi(idx, -1);
6712 z_brl(L_post_third_loop_done);
6713
6714 Register multiplicand = product->successor();
6715 Register product_low = multiplicand;
6716
6717 z_sllg(Z_R7, idx, LogBytesPerInt);
6718 clear_reg(yz_idx);
6719 mem2reg_opt(yz_idx, Address(y, Z_R7, 0), false);
6720 lgr_if_needed(multiplicand, x_xstart);
6721 z_mlgr(product, yz_idx); // multiplicand * yz_idx -> product::multiplicand
6722 clear_reg(yz_idx);
6723 mem2reg_opt(yz_idx, Address(z, Z_R7, 0), false);
6724
6725 add2_with_carry(product, product_low, yz_idx, carry);
6726
6727 z_sllg(Z_R7, idx, LogBytesPerInt);
6728 reg2mem_opt(product_low, Address(z, Z_R7, 0), false);
6729 rshift(product_low, 32);
6730
6731 lshift(product, 32);
6732 z_ogr(product_low, product);
6733 lgr_if_needed(carry, product_low);
6734
6735 bind(L_post_third_loop_done);
6736 }
6737
6738 void MacroAssembler::multiply_to_len(Register x, Register xlen,
6739 Register y, Register ylen,
6740 Register z,
6741 Register tmp1, Register tmp2,
6742 Register tmp3, Register tmp4,
6743 Register tmp5) {
6744 ShortBranchVerifier sbv(this);
6745
6746 assert_different_registers(x, xlen, y, ylen, z,
6747 tmp1, tmp2, tmp3, tmp4, tmp5, Z_R1_scratch, Z_R7);
6748 assert_different_registers(x, xlen, y, ylen, z,
6749 tmp1, tmp2, tmp3, tmp4, tmp5, Z_R8);
6750
6751 z_stmg(Z_R7, Z_R13, _z_abi(gpr7), Z_SP);
6752
6753 // In openJdk, we store the argument as 32-bit value to slot.
6754 Address zlen(Z_SP, _z_abi(remaining_cargs)); // Int in long on big endian.
6755
6756 const Register idx = tmp1;
6757 const Register kdx = tmp2;
6758 const Register xstart = tmp3;
6759
6760 const Register y_idx = tmp4;
6761 const Register carry = tmp5;
6762 const Register product = Z_R0_scratch;
6763 const Register x_xstart = Z_R8;
6764
6765 // First Loop.
6766 //
6767 // final static long LONG_MASK = 0xffffffffL;
6768 // int xstart = xlen - 1;
6769 // int ystart = ylen - 1;
6770 // long carry = 0;
6771 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
6772 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
6773 // z[kdx] = (int)product;
6774 // carry = product >>> 32;
6775 // }
6776 // z[xstart] = (int)carry;
6777 //
6778
6779 lgr_if_needed(idx, ylen); // idx = ylen
6780 z_llgf(kdx, zlen); // C2 does not respect int to long conversion for stub calls, thus load zero-extended.
6781 clear_reg(carry); // carry = 0
6782
6783 Label L_done;
6784
6785 lgr_if_needed(xstart, xlen);
6786 z_aghi(xstart, -1);
6787 z_brl(L_done);
6788
6789 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
6790
6791 NearLabel L_second_loop;
6792 compare64_and_branch(kdx, RegisterOrConstant((intptr_t) 0), bcondEqual, L_second_loop);
6793
6794 NearLabel L_carry;
6795 z_aghi(kdx, -1);
6796 z_brz(L_carry);
6797
6798 // Store lower 32 bits of carry.
6799 z_sllg(Z_R1_scratch, kdx, LogBytesPerInt);
6800 reg2mem_opt(carry, Address(z, Z_R1_scratch, 0), false);
6801 rshift(carry, 32);
6802 z_aghi(kdx, -1);
6803
6804
6805 bind(L_carry);
6806
6807 // Store upper 32 bits of carry.
6808 z_sllg(Z_R1_scratch, kdx, LogBytesPerInt);
6809 reg2mem_opt(carry, Address(z, Z_R1_scratch, 0), false);
6810
6811 // Second and third (nested) loops.
6812 //
6813 // for (int i = xstart-1; i >= 0; i--) { // Second loop
6814 // carry = 0;
6815 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
6816 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
6817 // (z[k] & LONG_MASK) + carry;
6818 // z[k] = (int)product;
6819 // carry = product >>> 32;
6820 // }
6821 // z[i] = (int)carry;
6822 // }
6823 //
6824 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = rdx
6825
6826 const Register jdx = tmp1;
6827
6828 bind(L_second_loop);
6829
6830 clear_reg(carry); // carry = 0;
6831 lgr_if_needed(jdx, ylen); // j = ystart+1
6832
6833 z_aghi(xstart, -1); // i = xstart-1;
6834 z_brl(L_done);
6835
6836 // Use free slots in the current stackframe instead of push/pop.
6837 Address zsave(Z_SP, _z_abi(carg_1));
6838 reg2mem_opt(z, zsave);
6839
6840
6841 Label L_last_x;
6842
6843 z_sllg(Z_R1_scratch, xstart, LogBytesPerInt);
6844 load_address(z, Address(z, Z_R1_scratch, 4)); // z = z + k - j
6845 z_aghi(xstart, -1); // i = xstart-1;
6846 z_brl(L_last_x);
6847
6848 z_sllg(Z_R1_scratch, xstart, LogBytesPerInt);
6849 mem2reg_opt(x_xstart, Address(x, Z_R1_scratch, 0));
6850
6851
6852 Label L_third_loop_prologue;
6853
6854 bind(L_third_loop_prologue);
6855
6856 Address xsave(Z_SP, _z_abi(carg_2));
6857 Address xlensave(Z_SP, _z_abi(carg_3));
6858 Address ylensave(Z_SP, _z_abi(carg_4));
6859
6860 reg2mem_opt(x, xsave);
6861 reg2mem_opt(xstart, xlensave);
6862 reg2mem_opt(ylen, ylensave);
6863
6864
6865 multiply_128_x_128_loop(x_xstart, y, z, y_idx, jdx, ylen, carry, product, x);
6866
6867 mem2reg_opt(z, zsave);
6868 mem2reg_opt(x, xsave);
6869 mem2reg_opt(xlen, xlensave); // This is the decrement of the loop counter!
6870 mem2reg_opt(ylen, ylensave);
6871
6872 add2reg(tmp3, 1, xlen);
6873 z_sllg(Z_R1_scratch, tmp3, LogBytesPerInt);
6874 reg2mem_opt(carry, Address(z, Z_R1_scratch, 0), false);
6875 z_aghi(tmp3, -1);
6876 z_brl(L_done);
6877
6878 rshift(carry, 32);
6879 z_sllg(Z_R1_scratch, tmp3, LogBytesPerInt);
6880 reg2mem_opt(carry, Address(z, Z_R1_scratch, 0), false);
6881 z_bru(L_second_loop);
6882
6883 // Next infrequent code is moved outside loops.
6884 bind(L_last_x);
6885
6886 clear_reg(x_xstart);
6887 mem2reg_opt(x_xstart, Address(x, (intptr_t) 0), false);
6888 z_bru(L_third_loop_prologue);
6889
6890 bind(L_done);
6891
6892 z_lmg(Z_R7, Z_R13, _z_abi(gpr7), Z_SP);
6893 }
6894
6895 #ifndef PRODUCT
6896 // Assert if CC indicates "not equal" (check_equal==true) or "equal" (check_equal==false).
6897 void MacroAssembler::asm_assert(bool check_equal, const char *msg, int id) {
6898 Label ok;
6899 if (check_equal) {
6900 z_bre(ok);
6901 } else {
6902 z_brne(ok);
6903 }
6904 stop(msg, id);
6905 bind(ok);
6906 }
6907
6908 // Assert if CC indicates "low".
6909 void MacroAssembler::asm_assert_low(const char *msg, int id) {
6910 Label ok;
6911 z_brnl(ok);
6912 stop(msg, id);
6913 bind(ok);
6914 }
6915
6916 // Assert if CC indicates "high".
6917 void MacroAssembler::asm_assert_high(const char *msg, int id) {
6918 Label ok;
6919 z_brnh(ok);
6920 stop(msg, id);
6921 bind(ok);
6922 }
6923
6924 // Assert if CC indicates "not equal" (check_equal==true) or "equal" (check_equal==false)
6925 // generate non-relocatable code.
6926 void MacroAssembler::asm_assert_static(bool check_equal, const char *msg, int id) {
6927 Label ok;
6928 if (check_equal) { z_bre(ok); }
6929 else { z_brne(ok); }
6930 stop_static(msg, id);
6931 bind(ok);
6932 }
6933
6934 void MacroAssembler::asm_assert_mems_zero(bool check_equal, bool allow_relocation, int size, int64_t mem_offset,
6935 Register mem_base, const char* msg, int id) {
6936 switch (size) {
6937 case 4:
6938 load_and_test_int(Z_R0, Address(mem_base, mem_offset));
6939 break;
6940 case 8:
6941 load_and_test_long(Z_R0, Address(mem_base, mem_offset));
6942 break;
6943 default:
6944 ShouldNotReachHere();
6945 }
6946 if (allow_relocation) { asm_assert(check_equal, msg, id); }
6947 else { asm_assert_static(check_equal, msg, id); }
6948 }
6949
6950 // Check the condition
6951 // expected_size == FP - SP
6952 // after transformation:
6953 // expected_size - FP + SP == 0
6954 // Destroys Register expected_size if no tmp register is passed.
6955 void MacroAssembler::asm_assert_frame_size(Register expected_size, Register tmp, const char* msg, int id) {
6956 if (tmp == noreg) {
6957 tmp = expected_size;
6958 } else {
6959 if (tmp != expected_size) {
6960 z_lgr(tmp, expected_size);
6961 }
6962 z_algr(tmp, Z_SP);
6963 z_slg(tmp, 0, Z_R0, Z_SP);
6964 asm_assert_eq(msg, id);
6965 }
6966 }
6967 #endif // !PRODUCT
6968
6969 void MacroAssembler::verify_thread() {
6970 if (VerifyThread) {
6971 unimplemented("", 117);
6972 }
6973 }
6974
6975 // Plausibility check for oops.
6976 void MacroAssembler::verify_oop(Register oop, const char* msg) {
6977 if (!VerifyOops) return;
6978
6979 BLOCK_COMMENT("verify_oop {");
6980 Register tmp = Z_R0;
6981 unsigned int nbytes_save = 5*BytesPerWord;
6982 address entry = StubRoutines::verify_oop_subroutine_entry_address();
6983
6984 save_return_pc();
6985 push_frame_abi160(nbytes_save);
6986 z_stmg(Z_R1, Z_R5, frame::z_abi_160_size, Z_SP);
6987
6988 z_lgr(Z_ARG2, oop);
6989 load_const(Z_ARG1, (address) msg);
6990 load_const(Z_R1, entry);
6991 z_lg(Z_R1, 0, Z_R1);
6992 call_c(Z_R1);
6993
6994 z_lmg(Z_R1, Z_R5, frame::z_abi_160_size, Z_SP);
6995 pop_frame();
6996 restore_return_pc();
6997
6998 BLOCK_COMMENT("} verify_oop ");
6999 }
7000
7001 const char* MacroAssembler::stop_types[] = {
7002 "stop",
7003 "untested",
7004 "unimplemented",
7005 "shouldnotreachhere"
7006 };
7007
7008 static void stop_on_request(const char* tp, const char* msg) {
7009 tty->print("Z assembly code requires stop: (%s) %s\n", tp, msg);
7010 guarantee(false, "Z assembly code requires stop: %s", msg);
7011 }
7012
7013 void MacroAssembler::stop(int type, const char* msg, int id) {
7014 BLOCK_COMMENT(err_msg("stop: %s {", msg));
7015
7016 // Setup arguments.
7017 load_const(Z_ARG1, (void*) stop_types[type%stop_end]);
7018 load_const(Z_ARG2, (void*) msg);
7019 get_PC(Z_R14); // Following code pushes a frame without entering a new function. Use current pc as return address.
7020 save_return_pc(); // Saves return pc Z_R14.
7021 push_frame_abi160(0);
7022 call_VM_leaf(CAST_FROM_FN_PTR(address, stop_on_request), Z_ARG1, Z_ARG2);
7023 // The plain disassembler does not recognize illtrap. It instead displays
7024 // a 32-bit value. Issueing two illtraps assures the disassembler finds
7025 // the proper beginning of the next instruction.
7026 z_illtrap(); // Illegal instruction.
7027 z_illtrap(); // Illegal instruction.
7028
7029 BLOCK_COMMENT(" } stop");
7030 }
7031
7032 // Special version of stop() for code size reduction.
7033 // Reuses the previously generated call sequence, if any.
7034 // Generates the call sequence on its own, if necessary.
7035 // Note: This code will work only in non-relocatable code!
7036 // The relative address of the data elements (arg1, arg2) must not change.
7037 // The reentry point must not move relative to it's users. This prerequisite
7038 // should be given for "hand-written" code, if all chain calls are in the same code blob.
7039 // Generated code must not undergo any transformation, e.g. ShortenBranches, to be safe.
7040 address MacroAssembler::stop_chain(address reentry, int type, const char* msg, int id, bool allow_relocation) {
7041 BLOCK_COMMENT(err_msg("stop_chain(%s,%s): %s {", reentry==NULL?"init":"cont", allow_relocation?"reloc ":"static", msg));
7042
7043 // Setup arguments.
7044 if (allow_relocation) {
7045 // Relocatable version (for comparison purposes). Remove after some time.
7046 load_const(Z_ARG1, (void*) stop_types[type%stop_end]);
7047 load_const(Z_ARG2, (void*) msg);
7048 } else {
7049 load_absolute_address(Z_ARG1, (address)stop_types[type%stop_end]);
7050 load_absolute_address(Z_ARG2, (address)msg);
7051 }
7052 if ((reentry != NULL) && RelAddr::is_in_range_of_RelAddr16(reentry, pc())) {
7053 BLOCK_COMMENT("branch to reentry point:");
7054 z_brc(bcondAlways, reentry);
7055 } else {
7056 BLOCK_COMMENT("reentry point:");
7057 reentry = pc(); // Re-entry point for subsequent stop calls.
7058 save_return_pc(); // Saves return pc Z_R14.
7059 push_frame_abi160(0);
7060 if (allow_relocation) {
7061 reentry = NULL; // Prevent reentry if code relocation is allowed.
7062 call_VM_leaf(CAST_FROM_FN_PTR(address, stop_on_request), Z_ARG1, Z_ARG2);
7063 } else {
7064 call_VM_leaf_static(CAST_FROM_FN_PTR(address, stop_on_request), Z_ARG1, Z_ARG2);
7065 }
7066 z_illtrap(); // Illegal instruction as emergency stop, should the above call return.
7067 }
7068 BLOCK_COMMENT(" } stop_chain");
7069
7070 return reentry;
7071 }
7072
7073 // Special version of stop() for code size reduction.
7074 // Assumes constant relative addresses for data and runtime call.
7075 void MacroAssembler::stop_static(int type, const char* msg, int id) {
7076 stop_chain(NULL, type, msg, id, false);
7077 }
7078
7079 void MacroAssembler::stop_subroutine() {
7080 unimplemented("stop_subroutine", 710);
7081 }
7082
7083 // Prints msg to stdout from within generated code..
7084 void MacroAssembler::warn(const char* msg) {
7085 RegisterSaver::save_live_registers(this, RegisterSaver::all_registers, Z_R14);
7086 load_absolute_address(Z_R1, (address) warning);
7087 load_absolute_address(Z_ARG1, (address) msg);
7088 (void) call(Z_R1);
7089 RegisterSaver::restore_live_registers(this, RegisterSaver::all_registers);
7090 }
7091
7092 #ifndef PRODUCT
7093
7094 // Write pattern 0x0101010101010101 in region [low-before, high+after].
7095 void MacroAssembler::zap_from_to(Register low, Register high, Register val, Register addr, int before, int after) {
7096 if (!ZapEmptyStackFields) return;
7097 BLOCK_COMMENT("zap memory region {");
7098 load_const_optimized(val, 0x0101010101010101);
7099 int size = before + after;
7100 if (low == high && size < 5 && size > 0) {
7101 int offset = -before*BytesPerWord;
7102 for (int i = 0; i < size; ++i) {
7103 z_stg(val, Address(low, offset));
7104 offset +=(1*BytesPerWord);
7105 }
7106 } else {
7107 add2reg(addr, -before*BytesPerWord, low);
7108 if (after) {
7109 #ifdef ASSERT
7110 jlong check = after * BytesPerWord;
7111 assert(Immediate::is_simm32(check) && Immediate::is_simm32(-check), "value not encodable !");
7112 #endif
7113 add2reg(high, after * BytesPerWord);
7114 }
7115 NearLabel loop;
7116 bind(loop);
7117 z_stg(val, Address(addr));
7118 add2reg(addr, 8);
7119 compare64_and_branch(addr, high, bcondNotHigh, loop);
7120 if (after) {
7121 add2reg(high, -after * BytesPerWord);
7122 }
7123 }
7124 BLOCK_COMMENT("} zap memory region");
7125 }
7126 #endif // !PRODUCT
7127
7128 SkipIfEqual::SkipIfEqual(MacroAssembler* masm, const bool* flag_addr, bool value, Register _rscratch) {
7129 _masm = masm;
7130 _masm->load_absolute_address(_rscratch, (address)flag_addr);
7131 _masm->load_and_test_int(_rscratch, Address(_rscratch));
7132 if (value) {
7133 _masm->z_brne(_label); // Skip if true, i.e. != 0.
7134 } else {
7135 _masm->z_bre(_label); // Skip if false, i.e. == 0.
7136 }
7137 }
7138
7139 SkipIfEqual::~SkipIfEqual() {
7140 _masm->bind(_label);
7141 }