1 /*
2 * Copyright (c) 2016, 2019, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2016, 2019, 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/macroAssembler.inline.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/icBuffer.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "gc/shared/gcLocker.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "interpreter/interp_masm.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/compiledICHolder.hpp"
36 #include "oops/klass.inline.hpp"
37 #include "registerSaver_s390.hpp"
38 #include "runtime/safepointMechanism.hpp"
39 #include "runtime/sharedRuntime.hpp"
40 #include "runtime/vframeArray.hpp"
41 #include "utilities/align.hpp"
42 #include "vmreg_s390.inline.hpp"
43 #ifdef COMPILER1
44 #include "c1/c1_Runtime1.hpp"
45 #endif
46 #ifdef COMPILER2
47 #include "opto/ad.hpp"
48 #include "opto/runtime.hpp"
49 #endif
50
51 #ifdef PRODUCT
52 #define __ masm->
53 #else
54 #define __ (Verbose ? (masm->block_comment(FILE_AND_LINE),masm):masm)->
55 #endif
56
57 #define BLOCK_COMMENT(str) __ block_comment(str)
58 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
59
60 #define RegisterSaver_LiveIntReg(regname) \
61 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() }
62
63 #define RegisterSaver_LiveFloatReg(regname) \
64 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() }
65
66 // Registers which are not saved/restored, but still they have got a frame slot.
67 // Used to get same frame size for RegisterSaver_LiveRegs and RegisterSaver_LiveRegsWithoutR2
68 #define RegisterSaver_ExcludedIntReg(regname) \
69 { RegisterSaver::excluded_reg, regname->encoding(), regname->as_VMReg() }
70
71 // Registers which are not saved/restored, but still they have got a frame slot.
72 // Used to get same frame size for RegisterSaver_LiveRegs and RegisterSaver_LiveRegsWithoutR2.
73 #define RegisterSaver_ExcludedFloatReg(regname) \
74 { RegisterSaver::excluded_reg, regname->encoding(), regname->as_VMReg() }
75
76 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = {
77 // Live registers which get spilled to the stack. Register positions
78 // in this array correspond directly to the stack layout.
79 //
80 // live float registers:
81 //
82 RegisterSaver_LiveFloatReg(Z_F0 ),
83 // RegisterSaver_ExcludedFloatReg(Z_F1 ), // scratch (Z_fscratch_1)
84 RegisterSaver_LiveFloatReg(Z_F2 ),
85 RegisterSaver_LiveFloatReg(Z_F3 ),
86 RegisterSaver_LiveFloatReg(Z_F4 ),
87 RegisterSaver_LiveFloatReg(Z_F5 ),
88 RegisterSaver_LiveFloatReg(Z_F6 ),
89 RegisterSaver_LiveFloatReg(Z_F7 ),
90 RegisterSaver_LiveFloatReg(Z_F8 ),
91 RegisterSaver_LiveFloatReg(Z_F9 ),
92 RegisterSaver_LiveFloatReg(Z_F10),
93 RegisterSaver_LiveFloatReg(Z_F11),
94 RegisterSaver_LiveFloatReg(Z_F12),
95 RegisterSaver_LiveFloatReg(Z_F13),
96 RegisterSaver_LiveFloatReg(Z_F14),
97 RegisterSaver_LiveFloatReg(Z_F15),
98 //
99 // RegisterSaver_ExcludedIntReg(Z_R0), // scratch
100 // RegisterSaver_ExcludedIntReg(Z_R1), // scratch
101 RegisterSaver_LiveIntReg(Z_R2 ),
102 RegisterSaver_LiveIntReg(Z_R3 ),
103 RegisterSaver_LiveIntReg(Z_R4 ),
104 RegisterSaver_LiveIntReg(Z_R5 ),
105 RegisterSaver_LiveIntReg(Z_R6 ),
106 RegisterSaver_LiveIntReg(Z_R7 ),
107 RegisterSaver_LiveIntReg(Z_R8 ),
108 RegisterSaver_LiveIntReg(Z_R9 ),
109 RegisterSaver_LiveIntReg(Z_R10),
110 RegisterSaver_LiveIntReg(Z_R11),
111 RegisterSaver_LiveIntReg(Z_R12),
112 RegisterSaver_LiveIntReg(Z_R13),
113 // RegisterSaver_ExcludedIntReg(Z_R14), // return pc (Saved in caller frame.)
114 // RegisterSaver_ExcludedIntReg(Z_R15) // stack pointer
115 };
116
117 static const RegisterSaver::LiveRegType RegisterSaver_LiveIntRegs[] = {
118 // Live registers which get spilled to the stack. Register positions
119 // in this array correspond directly to the stack layout.
120 //
121 // live float registers: All excluded, but still they get a stack slot to get same frame size.
122 //
123 RegisterSaver_ExcludedFloatReg(Z_F0 ),
124 // RegisterSaver_ExcludedFloatReg(Z_F1 ), // scratch (Z_fscratch_1)
125 RegisterSaver_ExcludedFloatReg(Z_F2 ),
126 RegisterSaver_ExcludedFloatReg(Z_F3 ),
127 RegisterSaver_ExcludedFloatReg(Z_F4 ),
128 RegisterSaver_ExcludedFloatReg(Z_F5 ),
129 RegisterSaver_ExcludedFloatReg(Z_F6 ),
130 RegisterSaver_ExcludedFloatReg(Z_F7 ),
131 RegisterSaver_ExcludedFloatReg(Z_F8 ),
132 RegisterSaver_ExcludedFloatReg(Z_F9 ),
133 RegisterSaver_ExcludedFloatReg(Z_F10),
134 RegisterSaver_ExcludedFloatReg(Z_F11),
135 RegisterSaver_ExcludedFloatReg(Z_F12),
136 RegisterSaver_ExcludedFloatReg(Z_F13),
137 RegisterSaver_ExcludedFloatReg(Z_F14),
138 RegisterSaver_ExcludedFloatReg(Z_F15),
139 //
140 // RegisterSaver_ExcludedIntReg(Z_R0), // scratch
141 // RegisterSaver_ExcludedIntReg(Z_R1), // scratch
142 RegisterSaver_LiveIntReg(Z_R2 ),
143 RegisterSaver_LiveIntReg(Z_R3 ),
144 RegisterSaver_LiveIntReg(Z_R4 ),
145 RegisterSaver_LiveIntReg(Z_R5 ),
146 RegisterSaver_LiveIntReg(Z_R6 ),
147 RegisterSaver_LiveIntReg(Z_R7 ),
148 RegisterSaver_LiveIntReg(Z_R8 ),
149 RegisterSaver_LiveIntReg(Z_R9 ),
150 RegisterSaver_LiveIntReg(Z_R10),
151 RegisterSaver_LiveIntReg(Z_R11),
152 RegisterSaver_LiveIntReg(Z_R12),
153 RegisterSaver_LiveIntReg(Z_R13),
154 // RegisterSaver_ExcludedIntReg(Z_R14), // return pc (Saved in caller frame.)
155 // RegisterSaver_ExcludedIntReg(Z_R15) // stack pointer
156 };
157
158 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegsWithoutR2[] = {
159 // Live registers which get spilled to the stack. Register positions
160 // in this array correspond directly to the stack layout.
161 //
162 // live float registers:
163 //
164 RegisterSaver_LiveFloatReg(Z_F0 ),
165 // RegisterSaver_ExcludedFloatReg(Z_F1 ), // scratch (Z_fscratch_1)
166 RegisterSaver_LiveFloatReg(Z_F2 ),
167 RegisterSaver_LiveFloatReg(Z_F3 ),
168 RegisterSaver_LiveFloatReg(Z_F4 ),
169 RegisterSaver_LiveFloatReg(Z_F5 ),
170 RegisterSaver_LiveFloatReg(Z_F6 ),
171 RegisterSaver_LiveFloatReg(Z_F7 ),
172 RegisterSaver_LiveFloatReg(Z_F8 ),
173 RegisterSaver_LiveFloatReg(Z_F9 ),
174 RegisterSaver_LiveFloatReg(Z_F10),
175 RegisterSaver_LiveFloatReg(Z_F11),
176 RegisterSaver_LiveFloatReg(Z_F12),
177 RegisterSaver_LiveFloatReg(Z_F13),
178 RegisterSaver_LiveFloatReg(Z_F14),
179 RegisterSaver_LiveFloatReg(Z_F15),
180 //
181 // RegisterSaver_ExcludedIntReg(Z_R0), // scratch
182 // RegisterSaver_ExcludedIntReg(Z_R1), // scratch
183 RegisterSaver_ExcludedIntReg(Z_R2), // Omit saving R2.
184 RegisterSaver_LiveIntReg(Z_R3 ),
185 RegisterSaver_LiveIntReg(Z_R4 ),
186 RegisterSaver_LiveIntReg(Z_R5 ),
187 RegisterSaver_LiveIntReg(Z_R6 ),
188 RegisterSaver_LiveIntReg(Z_R7 ),
189 RegisterSaver_LiveIntReg(Z_R8 ),
190 RegisterSaver_LiveIntReg(Z_R9 ),
191 RegisterSaver_LiveIntReg(Z_R10),
192 RegisterSaver_LiveIntReg(Z_R11),
193 RegisterSaver_LiveIntReg(Z_R12),
194 RegisterSaver_LiveIntReg(Z_R13),
195 // RegisterSaver_ExcludedIntReg(Z_R14), // return pc (Saved in caller frame.)
196 // RegisterSaver_ExcludedIntReg(Z_R15) // stack pointer
197 };
198
199 // Live argument registers which get spilled to the stack.
200 static const RegisterSaver::LiveRegType RegisterSaver_LiveArgRegs[] = {
201 RegisterSaver_LiveFloatReg(Z_FARG1),
202 RegisterSaver_LiveFloatReg(Z_FARG2),
203 RegisterSaver_LiveFloatReg(Z_FARG3),
204 RegisterSaver_LiveFloatReg(Z_FARG4),
205 RegisterSaver_LiveIntReg(Z_ARG1),
206 RegisterSaver_LiveIntReg(Z_ARG2),
207 RegisterSaver_LiveIntReg(Z_ARG3),
208 RegisterSaver_LiveIntReg(Z_ARG4),
209 RegisterSaver_LiveIntReg(Z_ARG5)
210 };
211
212 static const RegisterSaver::LiveRegType RegisterSaver_LiveVolatileRegs[] = {
213 // Live registers which get spilled to the stack. Register positions
214 // in this array correspond directly to the stack layout.
215 //
216 // live float registers:
217 //
218 RegisterSaver_LiveFloatReg(Z_F0 ),
219 // RegisterSaver_ExcludedFloatReg(Z_F1 ), // scratch (Z_fscratch_1)
220 RegisterSaver_LiveFloatReg(Z_F2 ),
221 RegisterSaver_LiveFloatReg(Z_F3 ),
222 RegisterSaver_LiveFloatReg(Z_F4 ),
223 RegisterSaver_LiveFloatReg(Z_F5 ),
224 RegisterSaver_LiveFloatReg(Z_F6 ),
225 RegisterSaver_LiveFloatReg(Z_F7 ),
226 // RegisterSaver_LiveFloatReg(Z_F8 ), // non-volatile
227 // RegisterSaver_LiveFloatReg(Z_F9 ), // non-volatile
228 // RegisterSaver_LiveFloatReg(Z_F10), // non-volatile
229 // RegisterSaver_LiveFloatReg(Z_F11), // non-volatile
230 // RegisterSaver_LiveFloatReg(Z_F12), // non-volatile
231 // RegisterSaver_LiveFloatReg(Z_F13), // non-volatile
232 // RegisterSaver_LiveFloatReg(Z_F14), // non-volatile
233 // RegisterSaver_LiveFloatReg(Z_F15), // non-volatile
234 //
235 // RegisterSaver_ExcludedIntReg(Z_R0), // scratch
236 // RegisterSaver_ExcludedIntReg(Z_R1), // scratch
237 RegisterSaver_LiveIntReg(Z_R2 ),
238 RegisterSaver_LiveIntReg(Z_R3 ),
239 RegisterSaver_LiveIntReg(Z_R4 ),
240 RegisterSaver_LiveIntReg(Z_R5 ),
241 // RegisterSaver_LiveIntReg(Z_R6 ), // non-volatile
242 // RegisterSaver_LiveIntReg(Z_R7 ), // non-volatile
243 // RegisterSaver_LiveIntReg(Z_R8 ), // non-volatile
244 // RegisterSaver_LiveIntReg(Z_R9 ), // non-volatile
245 // RegisterSaver_LiveIntReg(Z_R10), // non-volatile
246 // RegisterSaver_LiveIntReg(Z_R11), // non-volatile
247 // RegisterSaver_LiveIntReg(Z_R12), // non-volatile
248 // RegisterSaver_LiveIntReg(Z_R13), // non-volatile
249 // RegisterSaver_ExcludedIntReg(Z_R14), // return pc (Saved in caller frame.)
250 // RegisterSaver_ExcludedIntReg(Z_R15) // stack pointer
251 };
252
253 int RegisterSaver::live_reg_save_size(RegisterSet reg_set) {
254 int reg_space = -1;
255 switch (reg_set) {
256 case all_registers: reg_space = sizeof(RegisterSaver_LiveRegs); break;
257 case all_registers_except_r2: reg_space = sizeof(RegisterSaver_LiveRegsWithoutR2); break;
258 case all_integer_registers: reg_space = sizeof(RegisterSaver_LiveIntRegs); break;
259 case all_volatile_registers: reg_space = sizeof(RegisterSaver_LiveVolatileRegs); break;
260 case arg_registers: reg_space = sizeof(RegisterSaver_LiveArgRegs); break;
261 default: ShouldNotReachHere();
262 }
263 return (reg_space / sizeof(RegisterSaver::LiveRegType)) * reg_size;
264 }
265
266
267 int RegisterSaver::live_reg_frame_size(RegisterSet reg_set) {
268 return live_reg_save_size(reg_set) + frame::z_abi_160_size;
269 }
270
271
272 // return_pc: Specify the register that should be stored as the return pc in the current frame.
273 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, RegisterSet reg_set, Register return_pc) {
274 // Record volatile registers as callee-save values in an OopMap so
275 // their save locations will be propagated to the caller frame's
276 // RegisterMap during StackFrameStream construction (needed for
277 // deoptimization; see compiledVFrame::create_stack_value).
278
279 // Calculate frame size.
280 const int frame_size_in_bytes = live_reg_frame_size(reg_set);
281 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
282 const int register_save_offset = frame_size_in_bytes - live_reg_save_size(reg_set);
283
284 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
285 OopMap* map = new OopMap(frame_size_in_slots, 0);
286
287 int regstosave_num = 0;
288 const RegisterSaver::LiveRegType* live_regs = NULL;
289
290 switch (reg_set) {
291 case all_registers:
292 regstosave_num = sizeof(RegisterSaver_LiveRegs)/sizeof(RegisterSaver::LiveRegType);
293 live_regs = RegisterSaver_LiveRegs;
294 break;
295 case all_registers_except_r2:
296 regstosave_num = sizeof(RegisterSaver_LiveRegsWithoutR2)/sizeof(RegisterSaver::LiveRegType);;
297 live_regs = RegisterSaver_LiveRegsWithoutR2;
298 break;
299 case all_integer_registers:
300 regstosave_num = sizeof(RegisterSaver_LiveIntRegs)/sizeof(RegisterSaver::LiveRegType);
301 live_regs = RegisterSaver_LiveIntRegs;
302 break;
303 case all_volatile_registers:
304 regstosave_num = sizeof(RegisterSaver_LiveVolatileRegs)/sizeof(RegisterSaver::LiveRegType);
305 live_regs = RegisterSaver_LiveVolatileRegs;
306 break;
307 case arg_registers:
308 regstosave_num = sizeof(RegisterSaver_LiveArgRegs)/sizeof(RegisterSaver::LiveRegType);;
309 live_regs = RegisterSaver_LiveArgRegs;
310 break;
311 default: ShouldNotReachHere();
312 }
313
314 // Save return pc in old frame.
315 __ save_return_pc(return_pc);
316
317 // Push a new frame (includes stack linkage).
318 // Use return_pc as scratch for push_frame. Z_R0_scratch (the default) and Z_R1_scratch are
319 // illegally used to pass parameters by RangeCheckStub::emit_code().
320 __ push_frame(frame_size_in_bytes, return_pc);
321 // We have to restore return_pc right away.
322 // Nobody else will. Furthermore, return_pc isn't necessarily the default (Z_R14).
323 // Nobody else knows which register we saved.
324 __ z_lg(return_pc, _z_abi16(return_pc) + frame_size_in_bytes, Z_SP);
325
326 // Register save area in new frame starts above z_abi_160 area.
327 int offset = register_save_offset;
328
329 Register first = noreg;
330 Register last = noreg;
331 int first_offset = -1;
332 bool float_spilled = false;
333
334 for (int i = 0; i < regstosave_num; i++, offset += reg_size) {
335 int reg_num = live_regs[i].reg_num;
336 int reg_type = live_regs[i].reg_type;
337
338 switch (reg_type) {
339 case RegisterSaver::int_reg: {
340 Register reg = as_Register(reg_num);
341 if (last != reg->predecessor()) {
342 if (first != noreg) {
343 __ z_stmg(first, last, first_offset, Z_SP);
344 }
345 first = reg;
346 first_offset = offset;
347 DEBUG_ONLY(float_spilled = false);
348 }
349 last = reg;
350 assert(last != Z_R0, "r0 would require special treatment");
351 assert(!float_spilled, "for simplicity, do not mix up ints and floats in RegisterSaver_LiveRegs[]");
352 break;
353 }
354
355 case RegisterSaver::excluded_reg: // Not saved/restored, but with dedicated slot.
356 continue; // Continue with next loop iteration.
357
358 case RegisterSaver::float_reg: {
359 FloatRegister freg = as_FloatRegister(reg_num);
360 __ z_std(freg, offset, Z_SP);
361 DEBUG_ONLY(float_spilled = true);
362 break;
363 }
364
365 default:
366 ShouldNotReachHere();
367 break;
368 }
369
370 // Second set_callee_saved is really a waste but we'll keep things as they were for now
371 map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2), live_regs[i].vmreg);
372 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size) >> 2), live_regs[i].vmreg->next());
373 }
374 assert(first != noreg, "Should spill at least one int reg.");
375 __ z_stmg(first, last, first_offset, Z_SP);
376
377 // And we're done.
378 return map;
379 }
380
381
382 // Generate the OopMap (again, regs where saved before).
383 OopMap* RegisterSaver::generate_oop_map(MacroAssembler* masm, RegisterSet reg_set) {
384 // Calculate frame size.
385 const int frame_size_in_bytes = live_reg_frame_size(reg_set);
386 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
387 const int register_save_offset = frame_size_in_bytes - live_reg_save_size(reg_set);
388
389 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
390 OopMap* map = new OopMap(frame_size_in_slots, 0);
391
392 int regstosave_num = 0;
393 const RegisterSaver::LiveRegType* live_regs = NULL;
394
395 switch (reg_set) {
396 case all_registers:
397 regstosave_num = sizeof(RegisterSaver_LiveRegs)/sizeof(RegisterSaver::LiveRegType);
398 live_regs = RegisterSaver_LiveRegs;
399 break;
400 case all_registers_except_r2:
401 regstosave_num = sizeof(RegisterSaver_LiveRegsWithoutR2)/sizeof(RegisterSaver::LiveRegType);;
402 live_regs = RegisterSaver_LiveRegsWithoutR2;
403 break;
404 case all_integer_registers:
405 regstosave_num = sizeof(RegisterSaver_LiveIntRegs)/sizeof(RegisterSaver::LiveRegType);
406 live_regs = RegisterSaver_LiveIntRegs;
407 break;
408 case all_volatile_registers:
409 regstosave_num = sizeof(RegisterSaver_LiveVolatileRegs)/sizeof(RegisterSaver::LiveRegType);
410 live_regs = RegisterSaver_LiveVolatileRegs;
411 break;
412 case arg_registers:
413 regstosave_num = sizeof(RegisterSaver_LiveArgRegs)/sizeof(RegisterSaver::LiveRegType);;
414 live_regs = RegisterSaver_LiveArgRegs;
415 break;
416 default: ShouldNotReachHere();
417 }
418
419 // Register save area in new frame starts above z_abi_160 area.
420 int offset = register_save_offset;
421 for (int i = 0; i < regstosave_num; i++) {
422 if (live_regs[i].reg_type < RegisterSaver::excluded_reg) {
423 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), live_regs[i].vmreg);
424 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size)>>2), live_regs[i].vmreg->next());
425 }
426 offset += reg_size;
427 }
428 return map;
429 }
430
431
432 // Pop the current frame and restore all the registers that we saved.
433 void RegisterSaver::restore_live_registers(MacroAssembler* masm, RegisterSet reg_set) {
434 int offset;
435 const int register_save_offset = live_reg_frame_size(reg_set) - live_reg_save_size(reg_set);
436
437 Register first = noreg;
438 Register last = noreg;
439 int first_offset = -1;
440 bool float_spilled = false;
441
442 int regstosave_num = 0;
443 const RegisterSaver::LiveRegType* live_regs = NULL;
444
445 switch (reg_set) {
446 case all_registers:
447 regstosave_num = sizeof(RegisterSaver_LiveRegs)/sizeof(RegisterSaver::LiveRegType);;
448 live_regs = RegisterSaver_LiveRegs;
449 break;
450 case all_registers_except_r2:
451 regstosave_num = sizeof(RegisterSaver_LiveRegsWithoutR2)/sizeof(RegisterSaver::LiveRegType);;
452 live_regs = RegisterSaver_LiveRegsWithoutR2;
453 break;
454 case all_integer_registers:
455 regstosave_num = sizeof(RegisterSaver_LiveIntRegs)/sizeof(RegisterSaver::LiveRegType);
456 live_regs = RegisterSaver_LiveIntRegs;
457 break;
458 case all_volatile_registers:
459 regstosave_num = sizeof(RegisterSaver_LiveVolatileRegs)/sizeof(RegisterSaver::LiveRegType);;
460 live_regs = RegisterSaver_LiveVolatileRegs;
461 break;
462 case arg_registers:
463 regstosave_num = sizeof(RegisterSaver_LiveArgRegs)/sizeof(RegisterSaver::LiveRegType);;
464 live_regs = RegisterSaver_LiveArgRegs;
465 break;
466 default: ShouldNotReachHere();
467 }
468
469 // Restore all registers (ints and floats).
470
471 // Register save area in new frame starts above z_abi_160 area.
472 offset = register_save_offset;
473
474 for (int i = 0; i < regstosave_num; i++, offset += reg_size) {
475 int reg_num = live_regs[i].reg_num;
476 int reg_type = live_regs[i].reg_type;
477
478 switch (reg_type) {
479 case RegisterSaver::excluded_reg:
480 continue; // Continue with next loop iteration.
481
482 case RegisterSaver::int_reg: {
483 Register reg = as_Register(reg_num);
484 if (last != reg->predecessor()) {
485 if (first != noreg) {
486 __ z_lmg(first, last, first_offset, Z_SP);
487 }
488 first = reg;
489 first_offset = offset;
490 DEBUG_ONLY(float_spilled = false);
491 }
492 last = reg;
493 assert(last != Z_R0, "r0 would require special treatment");
494 assert(!float_spilled, "for simplicity, do not mix up ints and floats in RegisterSaver_LiveRegs[]");
495 break;
496 }
497
498 case RegisterSaver::float_reg: {
499 FloatRegister freg = as_FloatRegister(reg_num);
500 __ z_ld(freg, offset, Z_SP);
501 DEBUG_ONLY(float_spilled = true);
502 break;
503 }
504
505 default:
506 ShouldNotReachHere();
507 }
508 }
509 assert(first != noreg, "Should spill at least one int reg.");
510 __ z_lmg(first, last, first_offset, Z_SP);
511
512 // Pop the frame.
513 __ pop_frame();
514
515 // Restore the flags.
516 __ restore_return_pc();
517 }
518
519
520 // Pop the current frame and restore the registers that might be holding a result.
521 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
522 int i;
523 int offset;
524 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
525 sizeof(RegisterSaver::LiveRegType);
526 const int register_save_offset = live_reg_frame_size(all_registers) - live_reg_save_size(all_registers);
527
528 // Restore all result registers (ints and floats).
529 offset = register_save_offset;
530 for (int i = 0; i < regstosave_num; i++, offset += reg_size) {
531 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
532 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
533 switch (reg_type) {
534 case RegisterSaver::excluded_reg:
535 continue; // Continue with next loop iteration.
536 case RegisterSaver::int_reg: {
537 if (as_Register(reg_num) == Z_RET) { // int result_reg
538 __ z_lg(as_Register(reg_num), offset, Z_SP);
539 }
540 break;
541 }
542 case RegisterSaver::float_reg: {
543 if (as_FloatRegister(reg_num) == Z_FRET) { // float result_reg
544 __ z_ld(as_FloatRegister(reg_num), offset, Z_SP);
545 }
546 break;
547 }
548 default:
549 ShouldNotReachHere();
550 }
551 }
552 }
553
554 size_t SharedRuntime::trampoline_size() {
555 return MacroAssembler::load_const_size() + 2;
556 }
557
558 void SharedRuntime::generate_trampoline(MacroAssembler *masm, address destination) {
559 // Think about using pc-relative branch.
560 __ load_const(Z_R1_scratch, destination);
561 __ z_br(Z_R1_scratch);
562 }
563
564 // ---------------------------------------------------------------------------
565 void SharedRuntime::save_native_result(MacroAssembler * masm,
566 BasicType ret_type,
567 int frame_slots) {
568 Address memaddr(Z_SP, frame_slots * VMRegImpl::stack_slot_size);
569
570 switch (ret_type) {
571 case T_BOOLEAN: // Save shorter types as int. Do we need sign extension at restore??
572 case T_BYTE:
573 case T_CHAR:
574 case T_SHORT:
575 case T_INT:
576 __ reg2mem_opt(Z_RET, memaddr, false);
577 break;
578 case T_OBJECT: // Save pointer types as long.
579 case T_ARRAY:
580 case T_ADDRESS:
581 case T_VOID:
582 case T_LONG:
583 __ reg2mem_opt(Z_RET, memaddr);
584 break;
585 case T_FLOAT:
586 __ freg2mem_opt(Z_FRET, memaddr, false);
587 break;
588 case T_DOUBLE:
589 __ freg2mem_opt(Z_FRET, memaddr);
590 break;
591 default:
592 ShouldNotReachHere();
593 break;
594 }
595 }
596
597 void SharedRuntime::restore_native_result(MacroAssembler *masm,
598 BasicType ret_type,
599 int frame_slots) {
600 Address memaddr(Z_SP, frame_slots * VMRegImpl::stack_slot_size);
601
602 switch (ret_type) {
603 case T_BOOLEAN: // Restore shorter types as int. Do we need sign extension at restore??
604 case T_BYTE:
605 case T_CHAR:
606 case T_SHORT:
607 case T_INT:
608 __ mem2reg_opt(Z_RET, memaddr, false);
609 break;
610 case T_OBJECT: // Restore pointer types as long.
611 case T_ARRAY:
612 case T_ADDRESS:
613 case T_VOID:
614 case T_LONG:
615 __ mem2reg_opt(Z_RET, memaddr);
616 break;
617 case T_FLOAT:
618 __ mem2freg_opt(Z_FRET, memaddr, false);
619 break;
620 case T_DOUBLE:
621 __ mem2freg_opt(Z_FRET, memaddr);
622 break;
623 default:
624 ShouldNotReachHere();
625 break;
626 }
627 }
628
629 // ---------------------------------------------------------------------------
630 // Read the array of BasicTypes from a signature, and compute where the
631 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
632 // quantities. Values less than VMRegImpl::stack0 are registers, those above
633 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
634 // as framesizes are fixed.
635 // VMRegImpl::stack0 refers to the first slot 0(sp).
636 // VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Registers
637 // up to RegisterImpl::number_of_registers are the 64-bit integer registers.
638
639 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
640 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
641 // units regardless of build.
642
643 // The Java calling convention is a "shifted" version of the C ABI.
644 // By skipping the first C ABI register we can call non-static jni methods
645 // with small numbers of arguments without having to shuffle the arguments
646 // at all. Since we control the java ABI we ought to at least get some
647 // advantage out of it.
648 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
649 VMRegPair *regs,
650 int total_args_passed,
651 int is_outgoing) {
652 // c2c calling conventions for compiled-compiled calls.
653
654 // An int/float occupies 1 slot here.
655 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats.
656 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles.
657
658 const VMReg z_iarg_reg[5] = {
659 Z_R2->as_VMReg(),
660 Z_R3->as_VMReg(),
661 Z_R4->as_VMReg(),
662 Z_R5->as_VMReg(),
663 Z_R6->as_VMReg()
664 };
665 const VMReg z_farg_reg[4] = {
666 Z_F0->as_VMReg(),
667 Z_F2->as_VMReg(),
668 Z_F4->as_VMReg(),
669 Z_F6->as_VMReg()
670 };
671 const int z_num_iarg_registers = sizeof(z_iarg_reg) / sizeof(z_iarg_reg[0]);
672 const int z_num_farg_registers = sizeof(z_farg_reg) / sizeof(z_farg_reg[0]);
673
674 assert(RegisterImpl::number_of_arg_registers == z_num_iarg_registers, "iarg reg count mismatch");
675 assert(FloatRegisterImpl::number_of_arg_registers == z_num_farg_registers, "farg reg count mismatch");
676
677 int i;
678 int stk = 0;
679 int ireg = 0;
680 int freg = 0;
681
682 for (int i = 0; i < total_args_passed; ++i) {
683 switch (sig_bt[i]) {
684 case T_BOOLEAN:
685 case T_CHAR:
686 case T_BYTE:
687 case T_SHORT:
688 case T_INT:
689 if (ireg < z_num_iarg_registers) {
690 // Put int/ptr in register.
691 regs[i].set1(z_iarg_reg[ireg]);
692 ++ireg;
693 } else {
694 // Put int/ptr on stack.
695 regs[i].set1(VMRegImpl::stack2reg(stk));
696 stk += inc_stk_for_intfloat;
697 }
698 break;
699 case T_LONG:
700 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
701 if (ireg < z_num_iarg_registers) {
702 // Put long in register.
703 regs[i].set2(z_iarg_reg[ireg]);
704 ++ireg;
705 } else {
706 // Put long on stack and align to 2 slots.
707 if (stk & 0x1) { ++stk; }
708 regs[i].set2(VMRegImpl::stack2reg(stk));
709 stk += inc_stk_for_longdouble;
710 }
711 break;
712 case T_OBJECT:
713 case T_ARRAY:
714 case T_ADDRESS:
715 if (ireg < z_num_iarg_registers) {
716 // Put ptr in register.
717 regs[i].set2(z_iarg_reg[ireg]);
718 ++ireg;
719 } else {
720 // Put ptr on stack and align to 2 slots, because
721 // "64-bit pointers record oop-ishness on 2 aligned adjacent
722 // registers." (see OopFlow::build_oop_map).
723 if (stk & 0x1) { ++stk; }
724 regs[i].set2(VMRegImpl::stack2reg(stk));
725 stk += inc_stk_for_longdouble;
726 }
727 break;
728 case T_FLOAT:
729 if (freg < z_num_farg_registers) {
730 // Put float in register.
731 regs[i].set1(z_farg_reg[freg]);
732 ++freg;
733 } else {
734 // Put float on stack.
735 regs[i].set1(VMRegImpl::stack2reg(stk));
736 stk += inc_stk_for_intfloat;
737 }
738 break;
739 case T_DOUBLE:
740 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
741 if (freg < z_num_farg_registers) {
742 // Put double in register.
743 regs[i].set2(z_farg_reg[freg]);
744 ++freg;
745 } else {
746 // Put double on stack and align to 2 slots.
747 if (stk & 0x1) { ++stk; }
748 regs[i].set2(VMRegImpl::stack2reg(stk));
749 stk += inc_stk_for_longdouble;
750 }
751 break;
752 case T_VOID:
753 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
754 // Do not count halves.
755 regs[i].set_bad();
756 break;
757 default:
758 ShouldNotReachHere();
759 }
760 }
761 return align_up(stk, 2);
762 }
763
764 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
765 VMRegPair *regs,
766 VMRegPair *regs2,
767 int total_args_passed) {
768 assert(regs2 == NULL, "second VMRegPair array not used on this platform");
769
770 // Calling conventions for C runtime calls and calls to JNI native methods.
771 const VMReg z_iarg_reg[5] = {
772 Z_R2->as_VMReg(),
773 Z_R3->as_VMReg(),
774 Z_R4->as_VMReg(),
775 Z_R5->as_VMReg(),
776 Z_R6->as_VMReg()
777 };
778 const VMReg z_farg_reg[4] = {
779 Z_F0->as_VMReg(),
780 Z_F2->as_VMReg(),
781 Z_F4->as_VMReg(),
782 Z_F6->as_VMReg()
783 };
784 const int z_num_iarg_registers = sizeof(z_iarg_reg) / sizeof(z_iarg_reg[0]);
785 const int z_num_farg_registers = sizeof(z_farg_reg) / sizeof(z_farg_reg[0]);
786
787 // Check calling conventions consistency.
788 assert(RegisterImpl::number_of_arg_registers == z_num_iarg_registers, "iarg reg count mismatch");
789 assert(FloatRegisterImpl::number_of_arg_registers == z_num_farg_registers, "farg reg count mismatch");
790
791 // Avoid passing C arguments in the wrong stack slots.
792
793 // 'Stk' counts stack slots. Due to alignment, 32 bit values occupy
794 // 2 such slots, like 64 bit values do.
795 const int inc_stk_for_intfloat = 2; // 2 slots for ints and floats.
796 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles.
797
798 int i;
799 // Leave room for C-compatible ABI
800 int stk = (frame::z_abi_160_size - frame::z_jit_out_preserve_size) / VMRegImpl::stack_slot_size;
801 int freg = 0;
802 int ireg = 0;
803
804 // We put the first 5 arguments into registers and the rest on the
805 // stack. Float arguments are already in their argument registers
806 // due to c2c calling conventions (see calling_convention).
807 for (int i = 0; i < total_args_passed; ++i) {
808 switch (sig_bt[i]) {
809 case T_BOOLEAN:
810 case T_CHAR:
811 case T_BYTE:
812 case T_SHORT:
813 case T_INT:
814 // Fall through, handle as long.
815 case T_LONG:
816 case T_OBJECT:
817 case T_ARRAY:
818 case T_ADDRESS:
819 case T_METADATA:
820 // Oops are already boxed if required (JNI).
821 if (ireg < z_num_iarg_registers) {
822 regs[i].set2(z_iarg_reg[ireg]);
823 ++ireg;
824 } else {
825 regs[i].set2(VMRegImpl::stack2reg(stk));
826 stk += inc_stk_for_longdouble;
827 }
828 break;
829 case T_FLOAT:
830 if (freg < z_num_farg_registers) {
831 regs[i].set1(z_farg_reg[freg]);
832 ++freg;
833 } else {
834 regs[i].set1(VMRegImpl::stack2reg(stk+1));
835 stk += inc_stk_for_intfloat;
836 }
837 break;
838 case T_DOUBLE:
839 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
840 if (freg < z_num_farg_registers) {
841 regs[i].set2(z_farg_reg[freg]);
842 ++freg;
843 } else {
844 // Put double on stack.
845 regs[i].set2(VMRegImpl::stack2reg(stk));
846 stk += inc_stk_for_longdouble;
847 }
848 break;
849 case T_VOID:
850 // Do not count halves.
851 regs[i].set_bad();
852 break;
853 default:
854 ShouldNotReachHere();
855 }
856 }
857 return align_up(stk, 2);
858 }
859
860 ////////////////////////////////////////////////////////////////////////
861 //
862 // Argument shufflers
863 //
864 ////////////////////////////////////////////////////////////////////////
865
866 //----------------------------------------------------------------------
867 // The java_calling_convention describes stack locations as ideal slots on
868 // a frame with no abi restrictions. Since we must observe abi restrictions
869 // (like the placement of the register window) the slots must be biased by
870 // the following value.
871 //----------------------------------------------------------------------
872 static int reg2slot(VMReg r) {
873 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots();
874 }
875
876 static int reg2offset(VMReg r) {
877 return reg2slot(r) * VMRegImpl::stack_slot_size;
878 }
879
880 static void verify_oop_args(MacroAssembler *masm,
881 int total_args_passed,
882 const BasicType *sig_bt,
883 const VMRegPair *regs) {
884 if (!VerifyOops) { return; }
885
886 for (int i = 0; i < total_args_passed; i++) {
887 if (is_reference_type(sig_bt[i])) {
888 VMReg r = regs[i].first();
889 assert(r->is_valid(), "bad oop arg");
890
891 if (r->is_stack()) {
892 __ z_lg(Z_R0_scratch,
893 Address(Z_SP, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
894 __ verify_oop(Z_R0_scratch);
895 } else {
896 __ verify_oop(r->as_Register());
897 }
898 }
899 }
900 }
901
902 static void gen_special_dispatch(MacroAssembler *masm,
903 int total_args_passed,
904 vmIntrinsics::ID special_dispatch,
905 const BasicType *sig_bt,
906 const VMRegPair *regs) {
907 verify_oop_args(masm, total_args_passed, sig_bt, regs);
908
909 // Now write the args into the outgoing interpreter space.
910 bool has_receiver = false;
911 Register receiver_reg = noreg;
912 int member_arg_pos = -1;
913 Register member_reg = noreg;
914 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(special_dispatch);
915
916 if (ref_kind != 0) {
917 member_arg_pos = total_args_passed - 1; // trailing MemberName argument
918 member_reg = Z_R9; // Known to be free at this point.
919 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
920 } else {
921 guarantee(special_dispatch == vmIntrinsics::_invokeBasic, "special_dispatch=%d", special_dispatch);
922 has_receiver = true;
923 }
924
925 if (member_reg != noreg) {
926 // Load the member_arg into register, if necessary.
927 assert(member_arg_pos >= 0 && member_arg_pos < total_args_passed, "oob");
928 assert(sig_bt[member_arg_pos] == T_OBJECT, "dispatch argument must be an object");
929
930 VMReg r = regs[member_arg_pos].first();
931 assert(r->is_valid(), "bad member arg");
932
933 if (r->is_stack()) {
934 __ z_lg(member_reg, Address(Z_SP, reg2offset(r)));
935 } else {
936 // No data motion is needed.
937 member_reg = r->as_Register();
938 }
939 }
940
941 if (has_receiver) {
942 // Make sure the receiver is loaded into a register.
943 assert(total_args_passed > 0, "oob");
944 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
945
946 VMReg r = regs[0].first();
947 assert(r->is_valid(), "bad receiver arg");
948
949 if (r->is_stack()) {
950 // Porting note: This assumes that compiled calling conventions always
951 // pass the receiver oop in a register. If this is not true on some
952 // platform, pick a temp and load the receiver from stack.
953 assert(false, "receiver always in a register");
954 receiver_reg = Z_R13; // Known to be free at this point.
955 __ z_lg(receiver_reg, Address(Z_SP, reg2offset(r)));
956 } else {
957 // No data motion is needed.
958 receiver_reg = r->as_Register();
959 }
960 }
961
962 // Figure out which address we are really jumping to:
963 MethodHandles::generate_method_handle_dispatch(masm, special_dispatch,
964 receiver_reg, member_reg,
965 /*for_compiler_entry:*/ true);
966 }
967
968 ////////////////////////////////////////////////////////////////////////
969 //
970 // Argument shufflers
971 //
972 ////////////////////////////////////////////////////////////////////////
973
974 // Is the size of a vector size (in bytes) bigger than a size saved by default?
975 // 8 bytes registers are saved by default on z/Architecture.
976 bool SharedRuntime::is_wide_vector(int size) {
977 // Note, MaxVectorSize == 8 on this platform.
978 assert(size <= 8, "%d bytes vectors are not supported", size);
979 return size > 8;
980 }
981
982 //----------------------------------------------------------------------
983 // An oop arg. Must pass a handle not the oop itself
984 //----------------------------------------------------------------------
985 static void object_move(MacroAssembler *masm,
986 OopMap *map,
987 int oop_handle_offset,
988 int framesize_in_slots,
989 VMRegPair src,
990 VMRegPair dst,
991 bool is_receiver,
992 int *receiver_offset) {
993 int frame_offset = framesize_in_slots*VMRegImpl::stack_slot_size;
994
995 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)), "only one receiving object per call, please.");
996
997 // Must pass a handle. First figure out the location we use as a handle.
998
999 if (src.first()->is_stack()) {
1000 // Oop is already on the stack, put handle on stack or in register
1001 // If handle will be on the stack, use temp reg to calculate it.
1002 Register rHandle = dst.first()->is_stack() ? Z_R1 : dst.first()->as_Register();
1003 Label skip;
1004 int slot_in_older_frame = reg2slot(src.first());
1005
1006 guarantee(!is_receiver, "expecting receiver in register");
1007 map->set_oop(VMRegImpl::stack2reg(slot_in_older_frame + framesize_in_slots));
1008
1009 __ add2reg(rHandle, reg2offset(src.first())+frame_offset, Z_SP);
1010 __ load_and_test_long(Z_R0, Address(rHandle));
1011 __ z_brne(skip);
1012 // Use a NULL handle if oop is NULL.
1013 __ clear_reg(rHandle, true, false);
1014 __ bind(skip);
1015
1016 // Copy handle to the right place (register or stack).
1017 if (dst.first()->is_stack()) {
1018 __ z_stg(rHandle, reg2offset(dst.first()), Z_SP);
1019 } // else
1020 // nothing to do. rHandle uses the correct register
1021 } else {
1022 // Oop is passed in an input register. We must flush it to the stack.
1023 const Register rOop = src.first()->as_Register();
1024 const Register rHandle = dst.first()->is_stack() ? Z_R1 : dst.first()->as_Register();
1025 int oop_slot = (rOop->encoding()-Z_ARG1->encoding()) * VMRegImpl::slots_per_word + oop_handle_offset;
1026 int oop_slot_offset = oop_slot*VMRegImpl::stack_slot_size;
1027 NearLabel skip;
1028
1029 if (is_receiver) {
1030 *receiver_offset = oop_slot_offset;
1031 }
1032 map->set_oop(VMRegImpl::stack2reg(oop_slot));
1033
1034 // Flush Oop to stack, calculate handle.
1035 __ z_stg(rOop, oop_slot_offset, Z_SP);
1036 __ add2reg(rHandle, oop_slot_offset, Z_SP);
1037
1038 // If Oop == NULL, use a NULL handle.
1039 __ compare64_and_branch(rOop, (RegisterOrConstant)0L, Assembler::bcondNotEqual, skip);
1040 __ clear_reg(rHandle, true, false);
1041 __ bind(skip);
1042
1043 // Copy handle to the right place (register or stack).
1044 if (dst.first()->is_stack()) {
1045 __ z_stg(rHandle, reg2offset(dst.first()), Z_SP);
1046 } // else
1047 // nothing to do here, since rHandle = dst.first()->as_Register in this case.
1048 }
1049 }
1050
1051 //----------------------------------------------------------------------
1052 // A float arg. May have to do float reg to int reg conversion
1053 //----------------------------------------------------------------------
1054 static void float_move(MacroAssembler *masm,
1055 VMRegPair src,
1056 VMRegPair dst,
1057 int framesize_in_slots,
1058 int workspace_slot_offset) {
1059 int frame_offset = framesize_in_slots * VMRegImpl::stack_slot_size;
1060 int workspace_offset = workspace_slot_offset * VMRegImpl::stack_slot_size;
1061
1062 // We do not accept an argument in a VMRegPair to be spread over two slots,
1063 // no matter what physical location (reg or stack) the slots may have.
1064 // We just check for the unaccepted slot to be invalid.
1065 assert(!src.second()->is_valid(), "float in arg spread over two slots");
1066 assert(!dst.second()->is_valid(), "float out arg spread over two slots");
1067
1068 if (src.first()->is_stack()) {
1069 if (dst.first()->is_stack()) {
1070 // stack -> stack. The easiest of the bunch.
1071 __ z_mvc(Address(Z_SP, reg2offset(dst.first())),
1072 Address(Z_SP, reg2offset(src.first()) + frame_offset), sizeof(float));
1073 } else {
1074 // stack to reg
1075 Address memaddr(Z_SP, reg2offset(src.first()) + frame_offset);
1076 if (dst.first()->is_Register()) {
1077 __ mem2reg_opt(dst.first()->as_Register(), memaddr, false);
1078 } else {
1079 __ mem2freg_opt(dst.first()->as_FloatRegister(), memaddr, false);
1080 }
1081 }
1082 } else if (src.first()->is_Register()) {
1083 if (dst.first()->is_stack()) {
1084 // gpr -> stack
1085 __ reg2mem_opt(src.first()->as_Register(),
1086 Address(Z_SP, reg2offset(dst.first()), false ));
1087 } else {
1088 if (dst.first()->is_Register()) {
1089 // gpr -> gpr
1090 __ move_reg_if_needed(dst.first()->as_Register(), T_INT,
1091 src.first()->as_Register(), T_INT);
1092 } else {
1093 if (VM_Version::has_FPSupportEnhancements()) {
1094 // gpr -> fpr. Exploit z10 capability of direct transfer.
1095 __ z_ldgr(dst.first()->as_FloatRegister(), src.first()->as_Register());
1096 } else {
1097 // gpr -> fpr. Use work space on stack to transfer data.
1098 Address stackaddr(Z_SP, workspace_offset);
1099
1100 __ reg2mem_opt(src.first()->as_Register(), stackaddr, false);
1101 __ mem2freg_opt(dst.first()->as_FloatRegister(), stackaddr, false);
1102 }
1103 }
1104 }
1105 } else {
1106 if (dst.first()->is_stack()) {
1107 // fpr -> stack
1108 __ freg2mem_opt(src.first()->as_FloatRegister(),
1109 Address(Z_SP, reg2offset(dst.first())), false);
1110 } else {
1111 if (dst.first()->is_Register()) {
1112 if (VM_Version::has_FPSupportEnhancements()) {
1113 // fpr -> gpr.
1114 __ z_lgdr(dst.first()->as_Register(), src.first()->as_FloatRegister());
1115 } else {
1116 // fpr -> gpr. Use work space on stack to transfer data.
1117 Address stackaddr(Z_SP, workspace_offset);
1118
1119 __ freg2mem_opt(src.first()->as_FloatRegister(), stackaddr, false);
1120 __ mem2reg_opt(dst.first()->as_Register(), stackaddr, false);
1121 }
1122 } else {
1123 // fpr -> fpr
1124 __ move_freg_if_needed(dst.first()->as_FloatRegister(), T_FLOAT,
1125 src.first()->as_FloatRegister(), T_FLOAT);
1126 }
1127 }
1128 }
1129 }
1130
1131 //----------------------------------------------------------------------
1132 // A double arg. May have to do double reg to long reg conversion
1133 //----------------------------------------------------------------------
1134 static void double_move(MacroAssembler *masm,
1135 VMRegPair src,
1136 VMRegPair dst,
1137 int framesize_in_slots,
1138 int workspace_slot_offset) {
1139 int frame_offset = framesize_in_slots*VMRegImpl::stack_slot_size;
1140 int workspace_offset = workspace_slot_offset*VMRegImpl::stack_slot_size;
1141
1142 // Since src is always a java calling convention we know that the
1143 // src pair is always either all registers or all stack (and aligned?)
1144
1145 if (src.first()->is_stack()) {
1146 if (dst.first()->is_stack()) {
1147 // stack -> stack. The easiest of the bunch.
1148 __ z_mvc(Address(Z_SP, reg2offset(dst.first())),
1149 Address(Z_SP, reg2offset(src.first()) + frame_offset), sizeof(double));
1150 } else {
1151 // stack to reg
1152 Address stackaddr(Z_SP, reg2offset(src.first()) + frame_offset);
1153
1154 if (dst.first()->is_Register()) {
1155 __ mem2reg_opt(dst.first()->as_Register(), stackaddr);
1156 } else {
1157 __ mem2freg_opt(dst.first()->as_FloatRegister(), stackaddr);
1158 }
1159 }
1160 } else if (src.first()->is_Register()) {
1161 if (dst.first()->is_stack()) {
1162 // gpr -> stack
1163 __ reg2mem_opt(src.first()->as_Register(),
1164 Address(Z_SP, reg2offset(dst.first())));
1165 } else {
1166 if (dst.first()->is_Register()) {
1167 // gpr -> gpr
1168 __ move_reg_if_needed(dst.first()->as_Register(), T_LONG,
1169 src.first()->as_Register(), T_LONG);
1170 } else {
1171 if (VM_Version::has_FPSupportEnhancements()) {
1172 // gpr -> fpr. Exploit z10 capability of direct transfer.
1173 __ z_ldgr(dst.first()->as_FloatRegister(), src.first()->as_Register());
1174 } else {
1175 // gpr -> fpr. Use work space on stack to transfer data.
1176 Address stackaddr(Z_SP, workspace_offset);
1177 __ reg2mem_opt(src.first()->as_Register(), stackaddr);
1178 __ mem2freg_opt(dst.first()->as_FloatRegister(), stackaddr);
1179 }
1180 }
1181 }
1182 } else {
1183 if (dst.first()->is_stack()) {
1184 // fpr -> stack
1185 __ freg2mem_opt(src.first()->as_FloatRegister(),
1186 Address(Z_SP, reg2offset(dst.first())));
1187 } else {
1188 if (dst.first()->is_Register()) {
1189 if (VM_Version::has_FPSupportEnhancements()) {
1190 // fpr -> gpr. Exploit z10 capability of direct transfer.
1191 __ z_lgdr(dst.first()->as_Register(), src.first()->as_FloatRegister());
1192 } else {
1193 // fpr -> gpr. Use work space on stack to transfer data.
1194 Address stackaddr(Z_SP, workspace_offset);
1195
1196 __ freg2mem_opt(src.first()->as_FloatRegister(), stackaddr);
1197 __ mem2reg_opt(dst.first()->as_Register(), stackaddr);
1198 }
1199 } else {
1200 // fpr -> fpr
1201 // In theory these overlap but the ordering is such that this is likely a nop.
1202 __ move_freg_if_needed(dst.first()->as_FloatRegister(), T_DOUBLE,
1203 src.first()->as_FloatRegister(), T_DOUBLE);
1204 }
1205 }
1206 }
1207 }
1208
1209 //----------------------------------------------------------------------
1210 // A long arg.
1211 //----------------------------------------------------------------------
1212 static void long_move(MacroAssembler *masm,
1213 VMRegPair src,
1214 VMRegPair dst,
1215 int framesize_in_slots) {
1216 int frame_offset = framesize_in_slots*VMRegImpl::stack_slot_size;
1217
1218 if (src.first()->is_stack()) {
1219 if (dst.first()->is_stack()) {
1220 // stack -> stack. The easiest of the bunch.
1221 __ z_mvc(Address(Z_SP, reg2offset(dst.first())),
1222 Address(Z_SP, reg2offset(src.first()) + frame_offset), sizeof(long));
1223 } else {
1224 // stack to reg
1225 assert(dst.first()->is_Register(), "long dst value must be in GPR");
1226 __ mem2reg_opt(dst.first()->as_Register(),
1227 Address(Z_SP, reg2offset(src.first()) + frame_offset));
1228 }
1229 } else {
1230 // reg to reg
1231 assert(src.first()->is_Register(), "long src value must be in GPR");
1232 if (dst.first()->is_stack()) {
1233 // reg -> stack
1234 __ reg2mem_opt(src.first()->as_Register(),
1235 Address(Z_SP, reg2offset(dst.first())));
1236 } else {
1237 // reg -> reg
1238 assert(dst.first()->is_Register(), "long dst value must be in GPR");
1239 __ move_reg_if_needed(dst.first()->as_Register(),
1240 T_LONG, src.first()->as_Register(), T_LONG);
1241 }
1242 }
1243 }
1244
1245
1246 //----------------------------------------------------------------------
1247 // A int-like arg.
1248 //----------------------------------------------------------------------
1249 // On z/Architecture we will store integer like items to the stack as 64 bit
1250 // items, according to the z/Architecture ABI, even though Java would only store
1251 // 32 bits for a parameter.
1252 // We do sign extension for all base types. That is ok since the only
1253 // unsigned base type is T_CHAR, and T_CHAR uses only 16 bits of an int.
1254 // Sign extension 32->64 bit will thus not affect the value.
1255 //----------------------------------------------------------------------
1256 static void move32_64(MacroAssembler *masm,
1257 VMRegPair src,
1258 VMRegPair dst,
1259 int framesize_in_slots) {
1260 int frame_offset = framesize_in_slots * VMRegImpl::stack_slot_size;
1261
1262 if (src.first()->is_stack()) {
1263 Address memaddr(Z_SP, reg2offset(src.first()) + frame_offset);
1264 if (dst.first()->is_stack()) {
1265 // stack -> stack. MVC not posible due to sign extension.
1266 Address firstaddr(Z_SP, reg2offset(dst.first()));
1267 __ mem2reg_signed_opt(Z_R0_scratch, memaddr);
1268 __ reg2mem_opt(Z_R0_scratch, firstaddr);
1269 } else {
1270 // stack -> reg, sign extended
1271 __ mem2reg_signed_opt(dst.first()->as_Register(), memaddr);
1272 }
1273 } else {
1274 if (dst.first()->is_stack()) {
1275 // reg -> stack, sign extended
1276 Address firstaddr(Z_SP, reg2offset(dst.first()));
1277 __ z_lgfr(src.first()->as_Register(), src.first()->as_Register());
1278 __ reg2mem_opt(src.first()->as_Register(), firstaddr);
1279 } else {
1280 // reg -> reg, sign extended
1281 __ z_lgfr(dst.first()->as_Register(), src.first()->as_Register());
1282 }
1283 }
1284 }
1285
1286 static void save_or_restore_arguments(MacroAssembler *masm,
1287 const int stack_slots,
1288 const int total_in_args,
1289 const int arg_save_area,
1290 OopMap *map,
1291 VMRegPair *in_regs,
1292 BasicType *in_sig_bt) {
1293
1294 // If map is non-NULL then the code should store the values,
1295 // otherwise it should load them.
1296 int slot = arg_save_area;
1297 // Handle double words first.
1298 for (int i = 0; i < total_in_args; i++) {
1299 if (in_regs[i].first()->is_FloatRegister() && in_sig_bt[i] == T_DOUBLE) {
1300 int offset = slot * VMRegImpl::stack_slot_size;
1301 slot += VMRegImpl::slots_per_word;
1302 assert(slot <= stack_slots, "overflow (after DOUBLE stack slot)");
1303 const FloatRegister freg = in_regs[i].first()->as_FloatRegister();
1304 Address stackaddr(Z_SP, offset);
1305 if (map != NULL) {
1306 __ freg2mem_opt(freg, stackaddr);
1307 } else {
1308 __ mem2freg_opt(freg, stackaddr);
1309 }
1310 } else if (in_regs[i].first()->is_Register() &&
1311 (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_ARRAY)) {
1312 int offset = slot * VMRegImpl::stack_slot_size;
1313 const Register reg = in_regs[i].first()->as_Register();
1314 if (map != NULL) {
1315 __ z_stg(reg, offset, Z_SP);
1316 if (in_sig_bt[i] == T_ARRAY) {
1317 map->set_oop(VMRegImpl::stack2reg(slot));
1318 }
1319 } else {
1320 __ z_lg(reg, offset, Z_SP);
1321 }
1322 slot += VMRegImpl::slots_per_word;
1323 assert(slot <= stack_slots, "overflow (after LONG/ARRAY stack slot)");
1324 }
1325 }
1326
1327 // Save or restore single word registers.
1328 for (int i = 0; i < total_in_args; i++) {
1329 if (in_regs[i].first()->is_Register()) {
1330 int offset = slot * VMRegImpl::stack_slot_size;
1331 // Value lives in an input register. Save it on stack.
1332 switch (in_sig_bt[i]) {
1333 case T_BOOLEAN:
1334 case T_CHAR:
1335 case T_BYTE:
1336 case T_SHORT:
1337 case T_INT: {
1338 const Register reg = in_regs[i].first()->as_Register();
1339 Address stackaddr(Z_SP, offset);
1340 if (map != NULL) {
1341 __ z_st(reg, stackaddr);
1342 } else {
1343 __ z_lgf(reg, stackaddr);
1344 }
1345 slot++;
1346 assert(slot <= stack_slots, "overflow (after INT or smaller stack slot)");
1347 break;
1348 }
1349 case T_ARRAY:
1350 case T_LONG:
1351 // handled above
1352 break;
1353 case T_OBJECT:
1354 default: ShouldNotReachHere();
1355 }
1356 } else if (in_regs[i].first()->is_FloatRegister()) {
1357 if (in_sig_bt[i] == T_FLOAT) {
1358 int offset = slot * VMRegImpl::stack_slot_size;
1359 slot++;
1360 assert(slot <= stack_slots, "overflow (after FLOAT stack slot)");
1361 const FloatRegister freg = in_regs[i].first()->as_FloatRegister();
1362 Address stackaddr(Z_SP, offset);
1363 if (map != NULL) {
1364 __ freg2mem_opt(freg, stackaddr, false);
1365 } else {
1366 __ mem2freg_opt(freg, stackaddr, false);
1367 }
1368 }
1369 } else if (in_regs[i].first()->is_stack() &&
1370 in_sig_bt[i] == T_ARRAY && map != NULL) {
1371 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1372 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots));
1373 }
1374 }
1375 }
1376
1377 // Check GCLocker::needs_gc and enter the runtime if it's true. This
1378 // keeps a new JNI critical region from starting until a GC has been
1379 // forced. Save down any oops in registers and describe them in an OopMap.
1380 static void check_needs_gc_for_critical_native(MacroAssembler *masm,
1381 const int stack_slots,
1382 const int total_in_args,
1383 const int arg_save_area,
1384 OopMapSet *oop_maps,
1385 VMRegPair *in_regs,
1386 BasicType *in_sig_bt) {
1387 __ block_comment("check GCLocker::needs_gc");
1388 Label cont;
1389
1390 // Check GCLocker::_needs_gc flag.
1391 __ load_const_optimized(Z_R1_scratch, (long) GCLocker::needs_gc_address());
1392 __ z_cli(0, Z_R1_scratch, 0);
1393 __ z_bre(cont);
1394
1395 // Save down any values that are live in registers and call into the
1396 // runtime to halt for a GC.
1397 OopMap *map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1398
1399 save_or_restore_arguments(masm, stack_slots, total_in_args,
1400 arg_save_area, map, in_regs, in_sig_bt);
1401 address the_pc = __ pc();
1402 __ set_last_Java_frame(Z_SP, noreg);
1403
1404 __ block_comment("block_for_jni_critical");
1405 __ z_lgr(Z_ARG1, Z_thread);
1406
1407 address entry_point = CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical);
1408 __ call_c(entry_point);
1409 oop_maps->add_gc_map(__ offset(), map);
1410
1411 __ reset_last_Java_frame();
1412
1413 // Reload all the register arguments.
1414 save_or_restore_arguments(masm, stack_slots, total_in_args,
1415 arg_save_area, NULL, in_regs, in_sig_bt);
1416
1417 __ bind(cont);
1418
1419 if (StressCriticalJNINatives) {
1420 // Stress register saving
1421 OopMap *map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1422 save_or_restore_arguments(masm, stack_slots, total_in_args,
1423 arg_save_area, map, in_regs, in_sig_bt);
1424
1425 // Destroy argument registers.
1426 for (int i = 0; i < total_in_args; i++) {
1427 if (in_regs[i].first()->is_Register()) {
1428 // Don't set CC.
1429 __ clear_reg(in_regs[i].first()->as_Register(), true, false);
1430 } else {
1431 if (in_regs[i].first()->is_FloatRegister()) {
1432 FloatRegister fr = in_regs[i].first()->as_FloatRegister();
1433 __ z_lcdbr(fr, fr);
1434 }
1435 }
1436 }
1437
1438 save_or_restore_arguments(masm, stack_slots, total_in_args,
1439 arg_save_area, NULL, in_regs, in_sig_bt);
1440 }
1441 }
1442
1443 static void move_ptr(MacroAssembler *masm,
1444 VMRegPair src,
1445 VMRegPair dst,
1446 int framesize_in_slots) {
1447 int frame_offset = framesize_in_slots * VMRegImpl::stack_slot_size;
1448
1449 if (src.first()->is_stack()) {
1450 if (dst.first()->is_stack()) {
1451 // stack to stack
1452 __ mem2reg_opt(Z_R0_scratch, Address(Z_SP, reg2offset(src.first()) + frame_offset));
1453 __ reg2mem_opt(Z_R0_scratch, Address(Z_SP, reg2offset(dst.first())));
1454 } else {
1455 // stack to reg
1456 __ mem2reg_opt(dst.first()->as_Register(),
1457 Address(Z_SP, reg2offset(src.first()) + frame_offset));
1458 }
1459 } else {
1460 if (dst.first()->is_stack()) {
1461 // reg to stack
1462 __ reg2mem_opt(src.first()->as_Register(), Address(Z_SP, reg2offset(dst.first())));
1463 } else {
1464 __ lgr_if_needed(dst.first()->as_Register(), src.first()->as_Register());
1465 }
1466 }
1467 }
1468
1469 // Unpack an array argument into a pointer to the body and the length
1470 // if the array is non-null, otherwise pass 0 for both.
1471 static void unpack_array_argument(MacroAssembler *masm,
1472 VMRegPair reg,
1473 BasicType in_elem_type,
1474 VMRegPair body_arg,
1475 VMRegPair length_arg,
1476 int framesize_in_slots) {
1477 Register tmp_reg = Z_tmp_2;
1478 Register tmp2_reg = Z_tmp_1;
1479
1480 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg,
1481 "possible collision");
1482 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg,
1483 "possible collision");
1484
1485 // Pass the length, ptr pair.
1486 NearLabel set_out_args;
1487 VMRegPair tmp, tmp2;
1488
1489 tmp.set_ptr(tmp_reg->as_VMReg());
1490 tmp2.set_ptr(tmp2_reg->as_VMReg());
1491 if (reg.first()->is_stack()) {
1492 // Load the arg up from the stack.
1493 move_ptr(masm, reg, tmp, framesize_in_slots);
1494 reg = tmp;
1495 }
1496
1497 const Register first = reg.first()->as_Register();
1498
1499 // Don't set CC, indicate unused result.
1500 (void) __ clear_reg(tmp2_reg, true, false);
1501 if (tmp_reg != first) {
1502 __ clear_reg(tmp_reg, true, false); // Don't set CC.
1503 }
1504 __ compare64_and_branch(first, (RegisterOrConstant)0L, Assembler::bcondEqual, set_out_args);
1505 __ z_lgf(tmp2_reg, Address(first, arrayOopDesc::length_offset_in_bytes()));
1506 __ add2reg(tmp_reg, arrayOopDesc::base_offset_in_bytes(in_elem_type), first);
1507
1508 __ bind(set_out_args);
1509 move_ptr(masm, tmp, body_arg, framesize_in_slots);
1510 move32_64(masm, tmp2, length_arg, framesize_in_slots);
1511 }
1512
1513 //----------------------------------------------------------------------
1514 // Wrap a JNI call.
1515 //----------------------------------------------------------------------
1516 #undef USE_RESIZE_FRAME
1517 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
1518 const methodHandle& method,
1519 int compile_id,
1520 BasicType *in_sig_bt,
1521 VMRegPair *in_regs,
1522 BasicType ret_type,
1523 address critical_entry) {
1524 #ifdef COMPILER2
1525 int total_in_args = method->size_of_parameters();
1526 if (method->is_method_handle_intrinsic()) {
1527 vmIntrinsics::ID iid = method->intrinsic_id();
1528 intptr_t start = (intptr_t) __ pc();
1529 int vep_offset = ((intptr_t) __ pc()) - start;
1530
1531 gen_special_dispatch(masm, total_in_args,
1532 method->intrinsic_id(), in_sig_bt, in_regs);
1533
1534 int frame_complete = ((intptr_t)__ pc()) - start; // Not complete, period.
1535
1536 __ flush();
1537
1538 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // No out slots at all, actually.
1539
1540 return nmethod::new_native_nmethod(method,
1541 compile_id,
1542 masm->code(),
1543 vep_offset,
1544 frame_complete,
1545 stack_slots / VMRegImpl::slots_per_word,
1546 in_ByteSize(-1),
1547 in_ByteSize(-1),
1548 (OopMapSet *) NULL);
1549 }
1550
1551
1552 ///////////////////////////////////////////////////////////////////////
1553 //
1554 // Precalculations before generating any code
1555 //
1556 ///////////////////////////////////////////////////////////////////////
1557
1558 bool is_critical_native = true;
1559 address native_func = critical_entry;
1560 if (native_func == NULL) {
1561 native_func = method->native_function();
1562 is_critical_native = false;
1563 }
1564 assert(native_func != NULL, "must have function");
1565
1566 //---------------------------------------------------------------------
1567 // We have received a description of where all the java args are located
1568 // on entry to the wrapper. We need to convert these args to where
1569 // the jni function will expect them. To figure out where they go
1570 // we convert the java signature to a C signature by inserting
1571 // the hidden arguments as arg[0] and possibly arg[1] (static method).
1572 //
1573 // The first hidden argument arg[0] is a pointer to the JNI environment.
1574 // It is generated for every call.
1575 // The second argument arg[1] to the JNI call, which is hidden for static
1576 // methods, is the boxed lock object. For static calls, the lock object
1577 // is the static method itself. The oop is constructed here. for instance
1578 // calls, the lock is performed on the object itself, the pointer of
1579 // which is passed as the first visible argument.
1580 //---------------------------------------------------------------------
1581
1582 // Additionally, on z/Architecture we must convert integers
1583 // to longs in the C signature. We do this in advance in order to have
1584 // no trouble with indexes into the bt-arrays.
1585 // So convert the signature and registers now, and adjust the total number
1586 // of in-arguments accordingly.
1587 bool method_is_static = method->is_static();
1588 int total_c_args = total_in_args;
1589
1590 if (!is_critical_native) {
1591 int n_hidden_args = method_is_static ? 2 : 1;
1592 total_c_args += n_hidden_args;
1593 } else {
1594 // No JNIEnv*, no this*, but unpacked arrays (base+length).
1595 for (int i = 0; i < total_in_args; i++) {
1596 if (in_sig_bt[i] == T_ARRAY) {
1597 total_c_args ++;
1598 }
1599 }
1600 }
1601
1602 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1603 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1604 BasicType* in_elem_bt = NULL;
1605
1606 // Create the signature for the C call:
1607 // 1) add the JNIEnv*
1608 // 2) add the class if the method is static
1609 // 3) copy the rest of the incoming signature (shifted by the number of
1610 // hidden arguments)
1611
1612 int argc = 0;
1613 if (!is_critical_native) {
1614 out_sig_bt[argc++] = T_ADDRESS;
1615 if (method->is_static()) {
1616 out_sig_bt[argc++] = T_OBJECT;
1617 }
1618
1619 for (int i = 0; i < total_in_args; i++) {
1620 out_sig_bt[argc++] = in_sig_bt[i];
1621 }
1622 } else {
1623 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args);
1624 SignatureStream ss(method->signature());
1625 int o = 0;
1626 for (int i = 0; i < total_in_args; i++, o++) {
1627 if (in_sig_bt[i] == T_ARRAY) {
1628 // Arrays are passed as tuples (int, elem*).
1629 Symbol* atype = ss.as_symbol();
1630 const char* at = atype->as_C_string();
1631 if (strlen(at) == 2) {
1632 assert(at[0] == '[', "must be");
1633 switch (at[1]) {
1634 case 'B': in_elem_bt[o] = T_BYTE; break;
1635 case 'C': in_elem_bt[o] = T_CHAR; break;
1636 case 'D': in_elem_bt[o] = T_DOUBLE; break;
1637 case 'F': in_elem_bt[o] = T_FLOAT; break;
1638 case 'I': in_elem_bt[o] = T_INT; break;
1639 case 'J': in_elem_bt[o] = T_LONG; break;
1640 case 'S': in_elem_bt[o] = T_SHORT; break;
1641 case 'Z': in_elem_bt[o] = T_BOOLEAN; break;
1642 default: ShouldNotReachHere();
1643 }
1644 }
1645 } else {
1646 in_elem_bt[o] = T_VOID;
1647 }
1648 if (in_sig_bt[i] != T_VOID) {
1649 assert(in_sig_bt[i] == ss.type(), "must match");
1650 ss.next();
1651 }
1652 }
1653 assert(total_in_args == o, "must match");
1654
1655 for (int i = 0; i < total_in_args; i++) {
1656 if (in_sig_bt[i] == T_ARRAY) {
1657 // Arrays are passed as tuples (int, elem*).
1658 out_sig_bt[argc++] = T_INT;
1659 out_sig_bt[argc++] = T_ADDRESS;
1660 } else {
1661 out_sig_bt[argc++] = in_sig_bt[i];
1662 }
1663 }
1664 }
1665
1666 ///////////////////////////////////////////////////////////////////////
1667 // Now figure out where the args must be stored and how much stack space
1668 // they require (neglecting out_preserve_stack_slots but providing space
1669 // for storing the first five register arguments).
1670 // It's weird, see int_stk_helper.
1671 ///////////////////////////////////////////////////////////////////////
1672
1673 //---------------------------------------------------------------------
1674 // Compute framesize for the wrapper.
1675 //
1676 // - We need to handlize all oops passed in registers.
1677 // - We must create space for them here that is disjoint from the save area.
1678 // - We always just allocate 5 words for storing down these object.
1679 // This allows us to simply record the base and use the Ireg number to
1680 // decide which slot to use.
1681 // - Note that the reg number used to index the stack slot is the inbound
1682 // number, not the outbound number.
1683 // - We must shuffle args to match the native convention,
1684 // and to include var-args space.
1685 //---------------------------------------------------------------------
1686
1687 //---------------------------------------------------------------------
1688 // Calculate the total number of stack slots we will need:
1689 // - 1) abi requirements
1690 // - 2) outgoing args
1691 // - 3) space for inbound oop handle area
1692 // - 4) space for handlizing a klass if static method
1693 // - 5) space for a lock if synchronized method
1694 // - 6) workspace (save rtn value, int<->float reg moves, ...)
1695 // - 7) filler slots for alignment
1696 //---------------------------------------------------------------------
1697 // Here is how the space we have allocated will look like.
1698 // Since we use resize_frame, we do not create a new stack frame,
1699 // but just extend the one we got with our own data area.
1700 //
1701 // If an offset or pointer name points to a separator line, it is
1702 // assumed that addressing with offset 0 selects storage starting
1703 // at the first byte above the separator line.
1704 //
1705 //
1706 // ... ...
1707 // | caller's frame |
1708 // FP-> |---------------------|
1709 // | filler slots, if any|
1710 // 7| #slots == mult of 2 |
1711 // |---------------------|
1712 // | work space |
1713 // 6| 2 slots = 8 bytes |
1714 // |---------------------|
1715 // 5| lock box (if sync) |
1716 // |---------------------| <- lock_slot_offset
1717 // 4| klass (if static) |
1718 // |---------------------| <- klass_slot_offset
1719 // 3| oopHandle area |
1720 // | (save area for |
1721 // | critical natives) |
1722 // | |
1723 // | |
1724 // |---------------------| <- oop_handle_offset
1725 // 2| outbound memory |
1726 // ... ...
1727 // | based arguments |
1728 // |---------------------|
1729 // | vararg |
1730 // ... ...
1731 // | area |
1732 // |---------------------| <- out_arg_slot_offset
1733 // 1| out_preserved_slots |
1734 // ... ...
1735 // | (z_abi spec) |
1736 // SP-> |---------------------| <- FP_slot_offset (back chain)
1737 // ... ...
1738 //
1739 //---------------------------------------------------------------------
1740
1741 // *_slot_offset indicates offset from SP in #stack slots
1742 // *_offset indicates offset from SP in #bytes
1743
1744 int stack_slots = c_calling_convention(out_sig_bt, out_regs, /*regs2=*/NULL, total_c_args) + // 1+2
1745 SharedRuntime::out_preserve_stack_slots(); // see c_calling_convention
1746
1747 // Now the space for the inbound oop handle area.
1748 int total_save_slots = RegisterImpl::number_of_arg_registers * VMRegImpl::slots_per_word;
1749 if (is_critical_native) {
1750 // Critical natives may have to call out so they need a save area
1751 // for register arguments.
1752 int double_slots = 0;
1753 int single_slots = 0;
1754 for (int i = 0; i < total_in_args; i++) {
1755 if (in_regs[i].first()->is_Register()) {
1756 const Register reg = in_regs[i].first()->as_Register();
1757 switch (in_sig_bt[i]) {
1758 case T_BOOLEAN:
1759 case T_BYTE:
1760 case T_SHORT:
1761 case T_CHAR:
1762 case T_INT:
1763 // Fall through.
1764 case T_ARRAY:
1765 case T_LONG: double_slots++; break;
1766 default: ShouldNotReachHere();
1767 }
1768 } else {
1769 if (in_regs[i].first()->is_FloatRegister()) {
1770 switch (in_sig_bt[i]) {
1771 case T_FLOAT: single_slots++; break;
1772 case T_DOUBLE: double_slots++; break;
1773 default: ShouldNotReachHere();
1774 }
1775 }
1776 }
1777 } // for
1778 total_save_slots = double_slots * 2 + align_up(single_slots, 2); // Round to even.
1779 }
1780
1781 int oop_handle_slot_offset = stack_slots;
1782 stack_slots += total_save_slots; // 3)
1783
1784 int klass_slot_offset = 0;
1785 int klass_offset = -1;
1786 if (method_is_static && !is_critical_native) { // 4)
1787 klass_slot_offset = stack_slots;
1788 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1789 stack_slots += VMRegImpl::slots_per_word;
1790 }
1791
1792 int lock_slot_offset = 0;
1793 int lock_offset = -1;
1794 if (method->is_synchronized()) { // 5)
1795 lock_slot_offset = stack_slots;
1796 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size;
1797 stack_slots += VMRegImpl::slots_per_word;
1798 }
1799
1800 int workspace_slot_offset= stack_slots; // 6)
1801 stack_slots += 2;
1802
1803 // Now compute actual number of stack words we need.
1804 // Round to align stack properly.
1805 stack_slots = align_up(stack_slots, // 7)
1806 frame::alignment_in_bytes / VMRegImpl::stack_slot_size);
1807 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size;
1808
1809
1810 ///////////////////////////////////////////////////////////////////////
1811 // Now we can start generating code
1812 ///////////////////////////////////////////////////////////////////////
1813
1814 unsigned int wrapper_CodeStart = __ offset();
1815 unsigned int wrapper_UEPStart;
1816 unsigned int wrapper_VEPStart;
1817 unsigned int wrapper_FrameDone;
1818 unsigned int wrapper_CRegsSet;
1819 Label handle_pending_exception;
1820 Label ic_miss;
1821
1822 //---------------------------------------------------------------------
1823 // Unverified entry point (UEP)
1824 //---------------------------------------------------------------------
1825 wrapper_UEPStart = __ offset();
1826
1827 // check ic: object class <-> cached class
1828 if (!method_is_static) __ nmethod_UEP(ic_miss);
1829 // Fill with nops (alignment of verified entry point).
1830 __ align(CodeEntryAlignment);
1831
1832 //---------------------------------------------------------------------
1833 // Verified entry point (VEP)
1834 //---------------------------------------------------------------------
1835 wrapper_VEPStart = __ offset();
1836
1837 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1838 Label L_skip_barrier;
1839 Register klass = Z_R1_scratch;
1840 // Notify OOP recorder (don't need the relocation)
1841 AddressLiteral md = __ constant_metadata_address(method->method_holder());
1842 __ load_const_optimized(klass, md.value());
1843 __ clinit_barrier(klass, Z_thread, &L_skip_barrier /*L_fast_path*/);
1844
1845 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub());
1846 __ z_br(klass);
1847
1848 __ bind(L_skip_barrier);
1849 }
1850
1851 __ save_return_pc();
1852 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame.
1853 #ifndef USE_RESIZE_FRAME
1854 __ push_frame(frame_size_in_bytes); // Create a new frame for the wrapper.
1855 #else
1856 __ resize_frame(-frame_size_in_bytes, Z_R0_scratch); // No new frame for the wrapper.
1857 // Just resize the existing one.
1858 #endif
1859
1860 wrapper_FrameDone = __ offset();
1861
1862 __ verify_thread();
1863
1864 // Native nmethod wrappers never take possession of the oop arguments.
1865 // So the caller will gc the arguments.
1866 // The only thing we need an oopMap for is if the call is static.
1867 //
1868 // An OopMap for lock (and class if static), and one for the VM call itself
1869 OopMapSet *oop_maps = new OopMapSet();
1870 OopMap *map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1871
1872 if (is_critical_native) {
1873 check_needs_gc_for_critical_native(masm, stack_slots, total_in_args,
1874 oop_handle_slot_offset, oop_maps, in_regs, in_sig_bt);
1875 }
1876
1877
1878 //////////////////////////////////////////////////////////////////////
1879 //
1880 // The Grand Shuffle
1881 //
1882 //////////////////////////////////////////////////////////////////////
1883 //
1884 // We immediately shuffle the arguments so that for any vm call we have
1885 // to make from here on out (sync slow path, jvmti, etc.) we will have
1886 // captured the oops from our caller and have a valid oopMap for them.
1887 //
1888 //--------------------------------------------------------------------
1889 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
1890 // (derived from JavaThread* which is in Z_thread) and, if static,
1891 // the class mirror instead of a receiver. This pretty much guarantees that
1892 // register layout will not match. We ignore these extra arguments during
1893 // the shuffle. The shuffle is described by the two calling convention
1894 // vectors we have in our possession. We simply walk the java vector to
1895 // get the source locations and the c vector to get the destinations.
1896 //
1897 // This is a trick. We double the stack slots so we can claim
1898 // the oops in the caller's frame. Since we are sure to have
1899 // more args than the caller doubling is enough to make
1900 // sure we can capture all the incoming oop args from the caller.
1901 //--------------------------------------------------------------------
1902
1903 // Record sp-based slot for receiver on stack for non-static methods.
1904 int receiver_offset = -1;
1905
1906 //--------------------------------------------------------------------
1907 // We move the arguments backwards because the floating point registers
1908 // destination will always be to a register with a greater or equal
1909 // register number or the stack.
1910 // jix is the index of the incoming Java arguments.
1911 // cix is the index of the outgoing C arguments.
1912 //--------------------------------------------------------------------
1913
1914 #ifdef ASSERT
1915 bool reg_destroyed[RegisterImpl::number_of_registers];
1916 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
1917 for (int r = 0; r < RegisterImpl::number_of_registers; r++) {
1918 reg_destroyed[r] = false;
1919 }
1920 for (int f = 0; f < FloatRegisterImpl::number_of_registers; f++) {
1921 freg_destroyed[f] = false;
1922 }
1923 #endif // ASSERT
1924
1925 for (int jix = total_in_args - 1, cix = total_c_args - 1; jix >= 0; jix--, cix--) {
1926 #ifdef ASSERT
1927 if (in_regs[jix].first()->is_Register()) {
1928 assert(!reg_destroyed[in_regs[jix].first()->as_Register()->encoding()], "ack!");
1929 } else {
1930 if (in_regs[jix].first()->is_FloatRegister()) {
1931 assert(!freg_destroyed[in_regs[jix].first()->as_FloatRegister()->encoding()], "ack!");
1932 }
1933 }
1934 if (out_regs[cix].first()->is_Register()) {
1935 reg_destroyed[out_regs[cix].first()->as_Register()->encoding()] = true;
1936 } else {
1937 if (out_regs[cix].first()->is_FloatRegister()) {
1938 freg_destroyed[out_regs[cix].first()->as_FloatRegister()->encoding()] = true;
1939 }
1940 }
1941 #endif // ASSERT
1942
1943 switch (in_sig_bt[jix]) {
1944 // Due to casting, small integers should only occur in pairs with type T_LONG.
1945 case T_BOOLEAN:
1946 case T_CHAR:
1947 case T_BYTE:
1948 case T_SHORT:
1949 case T_INT:
1950 // Move int and do sign extension.
1951 move32_64(masm, in_regs[jix], out_regs[cix], stack_slots);
1952 break;
1953
1954 case T_LONG :
1955 long_move(masm, in_regs[jix], out_regs[cix], stack_slots);
1956 break;
1957
1958 case T_ARRAY:
1959 if (is_critical_native) {
1960 int body_arg = cix;
1961 cix -= 1; // Point to length arg.
1962 unpack_array_argument(masm, in_regs[jix], in_elem_bt[jix], out_regs[body_arg], out_regs[cix], stack_slots);
1963 break;
1964 }
1965 // else fallthrough
1966 case T_OBJECT:
1967 assert(!is_critical_native, "no oop arguments");
1968 object_move(masm, map, oop_handle_slot_offset, stack_slots, in_regs[jix], out_regs[cix],
1969 ((jix == 0) && (!method_is_static)),
1970 &receiver_offset);
1971 break;
1972 case T_VOID:
1973 break;
1974
1975 case T_FLOAT:
1976 float_move(masm, in_regs[jix], out_regs[cix], stack_slots, workspace_slot_offset);
1977 break;
1978
1979 case T_DOUBLE:
1980 assert(jix+1 < total_in_args && in_sig_bt[jix+1] == T_VOID && out_sig_bt[cix+1] == T_VOID, "bad arg list");
1981 double_move(masm, in_regs[jix], out_regs[cix], stack_slots, workspace_slot_offset);
1982 break;
1983
1984 case T_ADDRESS:
1985 assert(false, "found T_ADDRESS in java args");
1986 break;
1987
1988 default:
1989 ShouldNotReachHere();
1990 }
1991 }
1992
1993 //--------------------------------------------------------------------
1994 // Pre-load a static method's oop into ARG2.
1995 // Used both by locking code and the normal JNI call code.
1996 //--------------------------------------------------------------------
1997 if (method_is_static && !is_critical_native) {
1998 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()), Z_ARG2);
1999
2000 // Now handlize the static class mirror in ARG2. It's known not-null.
2001 __ z_stg(Z_ARG2, klass_offset, Z_SP);
2002 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2003 __ add2reg(Z_ARG2, klass_offset, Z_SP);
2004 }
2005
2006 // Get JNIEnv* which is first argument to native.
2007 if (!is_critical_native) {
2008 __ add2reg(Z_ARG1, in_bytes(JavaThread::jni_environment_offset()), Z_thread);
2009 }
2010
2011 //////////////////////////////////////////////////////////////////////
2012 // We have all of the arguments setup at this point.
2013 // We MUST NOT touch any outgoing regs from this point on.
2014 // So if we must call out we must push a new frame.
2015 //////////////////////////////////////////////////////////////////////
2016
2017
2018 // Calc the current pc into Z_R10 and into wrapper_CRegsSet.
2019 // Both values represent the same position.
2020 __ get_PC(Z_R10); // PC into register
2021 wrapper_CRegsSet = __ offset(); // and into into variable.
2022
2023 // Z_R10 now has the pc loaded that we will use when we finally call to native.
2024
2025 // We use the same pc/oopMap repeatedly when we call out.
2026 oop_maps->add_gc_map((int)(wrapper_CRegsSet-wrapper_CodeStart), map);
2027
2028 // Lock a synchronized method.
2029
2030 if (method->is_synchronized()) {
2031 assert(!is_critical_native, "unhandled");
2032
2033 // ATTENTION: args and Z_R10 must be preserved.
2034 Register r_oop = Z_R11;
2035 Register r_box = Z_R12;
2036 Register r_tmp1 = Z_R13;
2037 Register r_tmp2 = Z_R7;
2038 Label done;
2039
2040 // Load the oop for the object or class. R_carg2_classorobject contains
2041 // either the handlized oop from the incoming arguments or the handlized
2042 // class mirror (if the method is static).
2043 __ z_lg(r_oop, 0, Z_ARG2);
2044
2045 lock_offset = (lock_slot_offset * VMRegImpl::stack_slot_size);
2046 // Get the lock box slot's address.
2047 __ add2reg(r_box, lock_offset, Z_SP);
2048
2049 #ifdef ASSERT
2050 if (UseBiasedLocking)
2051 // Making the box point to itself will make it clear it went unused
2052 // but also be obviously invalid.
2053 __ z_stg(r_box, 0, r_box);
2054 #endif // ASSERT
2055
2056 // Try fastpath for locking.
2057 // Fast_lock kills r_temp_1, r_temp_2. (Don't use R1 as temp, won't work!)
2058 __ compiler_fast_lock_object(r_oop, r_box, r_tmp1, r_tmp2);
2059 __ z_bre(done);
2060
2061 //-------------------------------------------------------------------------
2062 // None of the above fast optimizations worked so we have to get into the
2063 // slow case of monitor enter. Inline a special case of call_VM that
2064 // disallows any pending_exception.
2065 //-------------------------------------------------------------------------
2066
2067 Register oldSP = Z_R11;
2068
2069 __ z_lgr(oldSP, Z_SP);
2070
2071 RegisterSaver::save_live_registers(masm, RegisterSaver::arg_registers);
2072
2073 // Prepare arguments for call.
2074 __ z_lg(Z_ARG1, 0, Z_ARG2); // Ynboxed class mirror or unboxed object.
2075 __ add2reg(Z_ARG2, lock_offset, oldSP);
2076 __ z_lgr(Z_ARG3, Z_thread);
2077
2078 __ set_last_Java_frame(oldSP, Z_R10 /* gc map pc */);
2079
2080 // Do the call.
2081 __ load_const_optimized(Z_R1_scratch, CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C));
2082 __ call(Z_R1_scratch);
2083
2084 __ reset_last_Java_frame();
2085
2086 RegisterSaver::restore_live_registers(masm, RegisterSaver::arg_registers);
2087 #ifdef ASSERT
2088 { Label L;
2089 __ load_and_test_long(Z_R0, Address(Z_thread, Thread::pending_exception_offset()));
2090 __ z_bre(L);
2091 __ stop("no pending exception allowed on exit from IR::monitorenter");
2092 __ bind(L);
2093 }
2094 #endif
2095 __ bind(done);
2096 } // lock for synchronized methods
2097
2098
2099 //////////////////////////////////////////////////////////////////////
2100 // Finally just about ready to make the JNI call.
2101 //////////////////////////////////////////////////////////////////////
2102
2103 // Use that pc we placed in Z_R10 a while back as the current frame anchor.
2104 __ set_last_Java_frame(Z_SP, Z_R10);
2105
2106 // Transition from _thread_in_Java to _thread_in_native.
2107 __ set_thread_state(_thread_in_native);
2108
2109
2110 //////////////////////////////////////////////////////////////////////
2111 // This is the JNI call.
2112 //////////////////////////////////////////////////////////////////////
2113
2114 __ call_c(native_func);
2115
2116
2117 //////////////////////////////////////////////////////////////////////
2118 // We have survived the call once we reach here.
2119 //////////////////////////////////////////////////////////////////////
2120
2121
2122 //--------------------------------------------------------------------
2123 // Unpack native results.
2124 //--------------------------------------------------------------------
2125 // For int-types, we do any needed sign-extension required.
2126 // Care must be taken that the return value (in Z_ARG1 = Z_RET = Z_R2
2127 // or in Z_FARG0 = Z_FRET = Z_F0) will survive any VM calls for
2128 // blocking or unlocking.
2129 // An OOP result (handle) is done specially in the slow-path code.
2130 //--------------------------------------------------------------------
2131 switch (ret_type) {
2132 case T_VOID: break; // Nothing to do!
2133 case T_FLOAT: break; // Got it where we want it (unless slow-path)
2134 case T_DOUBLE: break; // Got it where we want it (unless slow-path)
2135 case T_LONG: break; // Got it where we want it (unless slow-path)
2136 case T_OBJECT: break; // Really a handle.
2137 // Cannot de-handlize until after reclaiming jvm_lock.
2138 case T_ARRAY: break;
2139
2140 case T_BOOLEAN: // 0 -> false(0); !0 -> true(1)
2141 __ z_lngfr(Z_RET, Z_RET); // Force sign bit on except for zero.
2142 __ z_srlg(Z_RET, Z_RET, 63); // Shift sign bit into least significant pos.
2143 break;
2144 case T_BYTE: __ z_lgbr(Z_RET, Z_RET); break; // sign extension
2145 case T_CHAR: __ z_llghr(Z_RET, Z_RET); break; // unsigned result
2146 case T_SHORT: __ z_lghr(Z_RET, Z_RET); break; // sign extension
2147 case T_INT: __ z_lgfr(Z_RET, Z_RET); break; // sign-extend for beauty.
2148
2149 default:
2150 ShouldNotReachHere();
2151 break;
2152 }
2153
2154
2155 // Switch thread to "native transition" state before reading the synchronization state.
2156 // This additional state is necessary because reading and testing the synchronization
2157 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2158 // - Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2159 // - VM thread changes sync state to synchronizing and suspends threads for GC.
2160 // - Thread A is resumed to finish this native method, but doesn't block here since it
2161 // didn't see any synchronization in progress, and escapes.
2162
2163 // Transition from _thread_in_native to _thread_in_native_trans.
2164 __ set_thread_state(_thread_in_native_trans);
2165
2166 // Safepoint synchronization
2167 //--------------------------------------------------------------------
2168 // Must we block?
2169 //--------------------------------------------------------------------
2170 // Block, if necessary, before resuming in _thread_in_Java state.
2171 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2172 //--------------------------------------------------------------------
2173 Label after_transition;
2174 {
2175 Label no_block, sync;
2176
2177 save_native_result(masm, ret_type, workspace_slot_offset); // Make Z_R2 available as work reg.
2178
2179 // Force this write out before the read below.
2180 __ z_fence();
2181
2182 __ safepoint_poll(sync, Z_R1);
2183
2184 __ load_and_test_int(Z_R0, Address(Z_thread, JavaThread::suspend_flags_offset()));
2185 __ z_bre(no_block);
2186
2187 // Block. Save any potential method result value before the operation and
2188 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2189 // lets us share the oopMap we used when we went native rather than create
2190 // a distinct one for this pc.
2191 //
2192 __ bind(sync);
2193 __ z_acquire();
2194
2195 address entry_point = is_critical_native ? CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition)
2196 : CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans);
2197
2198 __ call_VM_leaf(entry_point, Z_thread);
2199
2200 if (is_critical_native) {
2201 restore_native_result(masm, ret_type, workspace_slot_offset);
2202 __ z_bru(after_transition); // No thread state transition here.
2203 }
2204 __ bind(no_block);
2205 restore_native_result(masm, ret_type, workspace_slot_offset);
2206 }
2207
2208 //--------------------------------------------------------------------
2209 // Thread state is thread_in_native_trans. Any safepoint blocking has
2210 // already happened so we can now change state to _thread_in_Java.
2211 //--------------------------------------------------------------------
2212 // Transition from _thread_in_native_trans to _thread_in_Java.
2213 __ set_thread_state(_thread_in_Java);
2214 __ bind(after_transition);
2215
2216
2217 //--------------------------------------------------------------------
2218 // Reguard any pages if necessary.
2219 // Protect native result from being destroyed.
2220 //--------------------------------------------------------------------
2221
2222 Label no_reguard;
2223
2224 __ z_cli(Address(Z_thread, JavaThread::stack_guard_state_offset() + in_ByteSize(sizeof(JavaThread::StackGuardState) - 1)),
2225 JavaThread::stack_guard_yellow_reserved_disabled);
2226
2227 __ z_bre(no_reguard);
2228
2229 save_native_result(masm, ret_type, workspace_slot_offset);
2230 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages), Z_method);
2231 restore_native_result(masm, ret_type, workspace_slot_offset);
2232
2233 __ bind(no_reguard);
2234
2235
2236 // Synchronized methods (slow path only)
2237 // No pending exceptions for now.
2238 //--------------------------------------------------------------------
2239 // Handle possibly pending exception (will unlock if necessary).
2240 // Native result is, if any is live, in Z_FRES or Z_RES.
2241 //--------------------------------------------------------------------
2242 // Unlock
2243 //--------------------------------------------------------------------
2244 if (method->is_synchronized()) {
2245 const Register r_oop = Z_R11;
2246 const Register r_box = Z_R12;
2247 const Register r_tmp1 = Z_R13;
2248 const Register r_tmp2 = Z_R7;
2249 Label done;
2250
2251 // Get unboxed oop of class mirror or object ...
2252 int offset = method_is_static ? klass_offset : receiver_offset;
2253
2254 assert(offset != -1, "");
2255 __ z_lg(r_oop, offset, Z_SP);
2256
2257 // ... and address of lock object box.
2258 __ add2reg(r_box, lock_offset, Z_SP);
2259
2260 // Try fastpath for unlocking.
2261 __ compiler_fast_unlock_object(r_oop, r_box, r_tmp1, r_tmp2); // Don't use R1 as temp.
2262 __ z_bre(done);
2263
2264 // Slow path for unlocking.
2265 // Save and restore any potential method result value around the unlocking operation.
2266 const Register R_exc = Z_R11;
2267
2268 save_native_result(masm, ret_type, workspace_slot_offset);
2269
2270 // Must save pending exception around the slow-path VM call. Since it's a
2271 // leaf call, the pending exception (if any) can be kept in a register.
2272 __ z_lg(R_exc, Address(Z_thread, Thread::pending_exception_offset()));
2273 assert(R_exc->is_nonvolatile(), "exception register must be non-volatile");
2274
2275 // Must clear pending-exception before re-entering the VM. Since this is
2276 // a leaf call, pending-exception-oop can be safely kept in a register.
2277 __ clear_mem(Address(Z_thread, Thread::pending_exception_offset()), sizeof(intptr_t));
2278
2279 // Inline a special case of call_VM that disallows any pending_exception.
2280
2281 // Get locked oop from the handle we passed to jni.
2282 __ z_lg(Z_ARG1, offset, Z_SP);
2283 __ add2reg(Z_ARG2, lock_offset, Z_SP);
2284 __ z_lgr(Z_ARG3, Z_thread);
2285
2286 __ load_const_optimized(Z_R1_scratch, CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
2287
2288 __ call(Z_R1_scratch);
2289
2290 #ifdef ASSERT
2291 {
2292 Label L;
2293 __ load_and_test_long(Z_R0, Address(Z_thread, Thread::pending_exception_offset()));
2294 __ z_bre(L);
2295 __ stop("no pending exception allowed on exit from IR::monitorexit");
2296 __ bind(L);
2297 }
2298 #endif
2299
2300 // Check_forward_pending_exception jump to forward_exception if any pending
2301 // exception is set. The forward_exception routine expects to see the
2302 // exception in pending_exception and not in a register. Kind of clumsy,
2303 // since all folks who branch to forward_exception must have tested
2304 // pending_exception first and hence have it in a register already.
2305 __ z_stg(R_exc, Address(Z_thread, Thread::pending_exception_offset()));
2306 restore_native_result(masm, ret_type, workspace_slot_offset);
2307 __ z_bru(done);
2308 __ z_illtrap(0x66);
2309
2310 __ bind(done);
2311 }
2312
2313
2314 //--------------------------------------------------------------------
2315 // Clear "last Java frame" SP and PC.
2316 //--------------------------------------------------------------------
2317 __ verify_thread(); // Z_thread must be correct.
2318
2319 __ reset_last_Java_frame();
2320
2321 // Unpack oop result, e.g. JNIHandles::resolve result.
2322 if (is_reference_type(ret_type)) {
2323 __ resolve_jobject(Z_RET, /* tmp1 */ Z_R13, /* tmp2 */ Z_R7);
2324 }
2325
2326 if (CheckJNICalls) {
2327 // clear_pending_jni_exception_check
2328 __ clear_mem(Address(Z_thread, JavaThread::pending_jni_exception_check_fn_offset()), sizeof(oop));
2329 }
2330
2331 // Reset handle block.
2332 if (!is_critical_native) {
2333 __ z_lg(Z_R1_scratch, Address(Z_thread, JavaThread::active_handles_offset()));
2334 __ clear_mem(Address(Z_R1_scratch, JNIHandleBlock::top_offset_in_bytes()), 4);
2335
2336 // Check for pending exceptions.
2337 __ load_and_test_long(Z_R0, Address(Z_thread, Thread::pending_exception_offset()));
2338 __ z_brne(handle_pending_exception);
2339 }
2340
2341
2342 //////////////////////////////////////////////////////////////////////
2343 // Return
2344 //////////////////////////////////////////////////////////////////////
2345
2346
2347 #ifndef USE_RESIZE_FRAME
2348 __ pop_frame(); // Pop wrapper frame.
2349 #else
2350 __ resize_frame(frame_size_in_bytes, Z_R0_scratch); // Revert stack extension.
2351 #endif
2352 __ restore_return_pc(); // This is the way back to the caller.
2353 __ z_br(Z_R14);
2354
2355
2356 //////////////////////////////////////////////////////////////////////
2357 // Out-of-line calls to the runtime.
2358 //////////////////////////////////////////////////////////////////////
2359
2360
2361 if (!is_critical_native) {
2362
2363 //---------------------------------------------------------------------
2364 // Handler for pending exceptions (out-of-line).
2365 //---------------------------------------------------------------------
2366 // Since this is a native call, we know the proper exception handler
2367 // is the empty function. We just pop this frame and then jump to
2368 // forward_exception_entry. Z_R14 will contain the native caller's
2369 // return PC.
2370 __ bind(handle_pending_exception);
2371 __ pop_frame();
2372 __ load_const_optimized(Z_R1_scratch, StubRoutines::forward_exception_entry());
2373 __ restore_return_pc();
2374 __ z_br(Z_R1_scratch);
2375
2376 //---------------------------------------------------------------------
2377 // Handler for a cache miss (out-of-line)
2378 //---------------------------------------------------------------------
2379 __ call_ic_miss_handler(ic_miss, 0x77, 0, Z_R1_scratch);
2380 }
2381 __ flush();
2382
2383
2384 //////////////////////////////////////////////////////////////////////
2385 // end of code generation
2386 //////////////////////////////////////////////////////////////////////
2387
2388
2389 nmethod *nm = nmethod::new_native_nmethod(method,
2390 compile_id,
2391 masm->code(),
2392 (int)(wrapper_VEPStart-wrapper_CodeStart),
2393 (int)(wrapper_FrameDone-wrapper_CodeStart),
2394 stack_slots / VMRegImpl::slots_per_word,
2395 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2396 in_ByteSize(lock_offset),
2397 oop_maps);
2398
2399 if (is_critical_native) {
2400 nm->set_lazy_critical_native(true);
2401 }
2402
2403 return nm;
2404 #else
2405 ShouldNotReachHere();
2406 return NULL;
2407 #endif // COMPILER2
2408 }
2409
2410 static address gen_c2i_adapter(MacroAssembler *masm,
2411 int total_args_passed,
2412 int comp_args_on_stack,
2413 const BasicType *sig_bt,
2414 const VMRegPair *regs,
2415 Label &skip_fixup) {
2416 // Before we get into the guts of the C2I adapter, see if we should be here
2417 // at all. We've come from compiled code and are attempting to jump to the
2418 // interpreter, which means the caller made a static call to get here
2419 // (vcalls always get a compiled target if there is one). Check for a
2420 // compiled target. If there is one, we need to patch the caller's call.
2421
2422 // These two defs MUST MATCH code in gen_i2c2i_adapter!
2423 const Register ientry = Z_R11;
2424 const Register code = Z_R11;
2425
2426 address c2i_entrypoint;
2427 Label patch_callsite;
2428
2429 // Regular (verified) c2i entry point.
2430 c2i_entrypoint = __ pc();
2431
2432 // Call patching needed?
2433 __ load_and_test_long(Z_R0_scratch, method_(code));
2434 __ z_lg(ientry, method_(interpreter_entry)); // Preload interpreter entry (also if patching).
2435 __ z_brne(patch_callsite); // Patch required if code != NULL (compiled target exists).
2436
2437 __ bind(skip_fixup); // Return point from patch_callsite.
2438
2439 // Since all args are passed on the stack, total_args_passed*wordSize is the
2440 // space we need. We need ABI scratch area but we use the caller's since
2441 // it has already been allocated.
2442
2443 const int abi_scratch = frame::z_top_ijava_frame_abi_size;
2444 int extraspace = align_up(total_args_passed, 2)*wordSize + abi_scratch;
2445 Register sender_SP = Z_R10;
2446 Register value = Z_R12;
2447
2448 // Remember the senderSP so we can pop the interpreter arguments off of the stack.
2449 // In addition, frame manager expects initial_caller_sp in Z_R10.
2450 __ z_lgr(sender_SP, Z_SP);
2451
2452 // This should always fit in 14 bit immediate.
2453 __ resize_frame(-extraspace, Z_R0_scratch);
2454
2455 // We use the caller's ABI scratch area (out_preserved_stack_slots) for the initial
2456 // args. This essentially moves the callers ABI scratch area from the top to the
2457 // bottom of the arg area.
2458
2459 int st_off = extraspace - wordSize;
2460
2461 // Now write the args into the outgoing interpreter space.
2462 for (int i = 0; i < total_args_passed; i++) {
2463 VMReg r_1 = regs[i].first();
2464 VMReg r_2 = regs[i].second();
2465 if (!r_1->is_valid()) {
2466 assert(!r_2->is_valid(), "");
2467 continue;
2468 }
2469 if (r_1->is_stack()) {
2470 // The calling convention produces OptoRegs that ignore the preserve area (abi scratch).
2471 // We must account for it here.
2472 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
2473
2474 if (!r_2->is_valid()) {
2475 __ z_mvc(Address(Z_SP, st_off), Address(sender_SP, ld_off), sizeof(void*));
2476 } else {
2477 // longs are given 2 64-bit slots in the interpreter,
2478 // but the data is passed in only 1 slot.
2479 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
2480 #ifdef ASSERT
2481 __ clear_mem(Address(Z_SP, st_off), sizeof(void *));
2482 #endif
2483 st_off -= wordSize;
2484 }
2485 __ z_mvc(Address(Z_SP, st_off), Address(sender_SP, ld_off), sizeof(void*));
2486 }
2487 } else {
2488 if (r_1->is_Register()) {
2489 if (!r_2->is_valid()) {
2490 __ z_st(r_1->as_Register(), st_off, Z_SP);
2491 } else {
2492 // longs are given 2 64-bit slots in the interpreter, but the
2493 // data is passed in only 1 slot.
2494 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
2495 #ifdef ASSERT
2496 __ clear_mem(Address(Z_SP, st_off), sizeof(void *));
2497 #endif
2498 st_off -= wordSize;
2499 }
2500 __ z_stg(r_1->as_Register(), st_off, Z_SP);
2501 }
2502 } else {
2503 assert(r_1->is_FloatRegister(), "");
2504 if (!r_2->is_valid()) {
2505 __ z_ste(r_1->as_FloatRegister(), st_off, Z_SP);
2506 } else {
2507 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the
2508 // data is passed in only 1 slot.
2509 // One of these should get known junk...
2510 #ifdef ASSERT
2511 __ z_lzdr(Z_F1);
2512 __ z_std(Z_F1, st_off, Z_SP);
2513 #endif
2514 st_off-=wordSize;
2515 __ z_std(r_1->as_FloatRegister(), st_off, Z_SP);
2516 }
2517 }
2518 }
2519 st_off -= wordSize;
2520 }
2521
2522
2523 // Jump to the interpreter just as if interpreter was doing it.
2524 __ add2reg(Z_esp, st_off, Z_SP);
2525
2526 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in Z_R10.
2527 __ z_br(ientry);
2528
2529
2530 // Prevent illegal entry to out-of-line code.
2531 __ z_illtrap(0x22);
2532
2533 // Generate out-of-line runtime call to patch caller,
2534 // then continue as interpreted.
2535
2536 // IF you lose the race you go interpreted.
2537 // We don't see any possible endless c2i -> i2c -> c2i ...
2538 // transitions no matter how rare.
2539 __ bind(patch_callsite);
2540
2541 RegisterSaver::save_live_registers(masm, RegisterSaver::arg_registers);
2542 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), Z_method, Z_R14);
2543 RegisterSaver::restore_live_registers(masm, RegisterSaver::arg_registers);
2544 __ z_bru(skip_fixup);
2545
2546 // end of out-of-line code
2547
2548 return c2i_entrypoint;
2549 }
2550
2551 // On entry, the following registers are set
2552 //
2553 // Z_thread r8 - JavaThread*
2554 // Z_method r9 - callee's method (method to be invoked)
2555 // Z_esp r7 - operand (or expression) stack pointer of caller. one slot above last arg.
2556 // Z_SP r15 - SP prepared by call stub such that caller's outgoing args are near top
2557 //
2558 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
2559 int total_args_passed,
2560 int comp_args_on_stack,
2561 const BasicType *sig_bt,
2562 const VMRegPair *regs) {
2563 const Register value = Z_R12;
2564 const Register ld_ptr= Z_esp;
2565
2566 int ld_offset = total_args_passed * wordSize;
2567
2568 // Cut-out for having no stack args.
2569 if (comp_args_on_stack) {
2570 // Sig words on the stack are greater than VMRegImpl::stack0. Those in
2571 // registers are below. By subtracting stack0, we either get a negative
2572 // number (all values in registers) or the maximum stack slot accessed.
2573 // Convert VMRegImpl (4 byte) stack slots to words.
2574 int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
2575 // Round up to miminum stack alignment, in wordSize
2576 comp_words_on_stack = align_up(comp_words_on_stack, 2);
2577
2578 __ resize_frame(-comp_words_on_stack*wordSize, Z_R0_scratch);
2579 }
2580
2581 // Now generate the shuffle code. Pick up all register args and move the
2582 // rest through register value=Z_R12.
2583 for (int i = 0; i < total_args_passed; i++) {
2584 if (sig_bt[i] == T_VOID) {
2585 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
2586 continue;
2587 }
2588
2589 // Pick up 0, 1 or 2 words from ld_ptr.
2590 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
2591 "scrambled load targets?");
2592 VMReg r_1 = regs[i].first();
2593 VMReg r_2 = regs[i].second();
2594 if (!r_1->is_valid()) {
2595 assert(!r_2->is_valid(), "");
2596 continue;
2597 }
2598 if (r_1->is_FloatRegister()) {
2599 if (!r_2->is_valid()) {
2600 __ z_le(r_1->as_FloatRegister(), ld_offset, ld_ptr);
2601 ld_offset-=wordSize;
2602 } else {
2603 // Skip the unused interpreter slot.
2604 __ z_ld(r_1->as_FloatRegister(), ld_offset - wordSize, ld_ptr);
2605 ld_offset -= 2 * wordSize;
2606 }
2607 } else {
2608 if (r_1->is_stack()) {
2609 // Must do a memory to memory move.
2610 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
2611
2612 if (!r_2->is_valid()) {
2613 __ z_mvc(Address(Z_SP, st_off), Address(ld_ptr, ld_offset), sizeof(void*));
2614 } else {
2615 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
2616 // data is passed in only 1 slot.
2617 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
2618 ld_offset -= wordSize;
2619 }
2620 __ z_mvc(Address(Z_SP, st_off), Address(ld_ptr, ld_offset), sizeof(void*));
2621 }
2622 } else {
2623 if (!r_2->is_valid()) {
2624 // Not sure we need to do this but it shouldn't hurt.
2625 if (is_reference_type(sig_bt[i]) || sig_bt[i] == T_ADDRESS) {
2626 __ z_lg(r_1->as_Register(), ld_offset, ld_ptr);
2627 } else {
2628 __ z_l(r_1->as_Register(), ld_offset, ld_ptr);
2629 }
2630 } else {
2631 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
2632 // data is passed in only 1 slot.
2633 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
2634 ld_offset -= wordSize;
2635 }
2636 __ z_lg(r_1->as_Register(), ld_offset, ld_ptr);
2637 }
2638 }
2639 ld_offset -= wordSize;
2640 }
2641 }
2642
2643 // Jump to the compiled code just as if compiled code was doing it.
2644 // load target address from method oop:
2645 __ z_lg(Z_R1_scratch, Address(Z_method, Method::from_compiled_offset()));
2646
2647 // Store method oop into thread->callee_target.
2648 // 6243940: We might end up in handle_wrong_method if
2649 // the callee is deoptimized as we race thru here. If that
2650 // happens we don't want to take a safepoint because the
2651 // caller frame will look interpreted and arguments are now
2652 // "compiled" so it is much better to make this transition
2653 // invisible to the stack walking code. Unfortunately, if
2654 // we try and find the callee by normal means a safepoint
2655 // is possible. So we stash the desired callee in the thread
2656 // and the vm will find it there should this case occur.
2657 __ z_stg(Z_method, thread_(callee_target));
2658
2659 __ z_br(Z_R1_scratch);
2660 }
2661
2662 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
2663 int total_args_passed,
2664 int comp_args_on_stack,
2665 const BasicType *sig_bt,
2666 const VMRegPair *regs,
2667 AdapterFingerPrint* fingerprint) {
2668 __ align(CodeEntryAlignment);
2669 address i2c_entry = __ pc();
2670 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
2671
2672 address c2i_unverified_entry;
2673
2674 Label skip_fixup;
2675 {
2676 Label ic_miss;
2677 const int klass_offset = oopDesc::klass_offset_in_bytes();
2678 const int holder_klass_offset = CompiledICHolder::holder_klass_offset();
2679 const int holder_metadata_offset = CompiledICHolder::holder_metadata_offset();
2680
2681 // Out-of-line call to ic_miss handler.
2682 __ call_ic_miss_handler(ic_miss, 0x11, 0, Z_R1_scratch);
2683
2684 // Unverified Entry Point UEP
2685 __ align(CodeEntryAlignment);
2686 c2i_unverified_entry = __ pc();
2687
2688 // Check the pointers.
2689 if (!ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(klass_offset)) {
2690 __ z_ltgr(Z_ARG1, Z_ARG1);
2691 __ z_bre(ic_miss);
2692 }
2693 __ verify_oop(Z_ARG1);
2694
2695 // Check ic: object class <-> cached class
2696 // Compress cached class for comparison. That's more efficient.
2697 if (UseCompressedClassPointers) {
2698 __ z_lg(Z_R11, holder_klass_offset, Z_method); // Z_R11 is overwritten a few instructions down anyway.
2699 __ compare_klass_ptr(Z_R11, klass_offset, Z_ARG1, false); // Cached class can't be zero.
2700 } else {
2701 __ z_clc(klass_offset, sizeof(void *)-1, Z_ARG1, holder_klass_offset, Z_method);
2702 }
2703 __ z_brne(ic_miss); // Cache miss: call runtime to handle this.
2704
2705 // This def MUST MATCH code in gen_c2i_adapter!
2706 const Register code = Z_R11;
2707
2708 __ z_lg(Z_method, holder_metadata_offset, Z_method);
2709 __ load_and_test_long(Z_R0, method_(code));
2710 __ z_brne(ic_miss); // Cache miss: call runtime to handle this.
2711
2712 // Fallthru to VEP. Duplicate LTG, but saved taken branch.
2713 }
2714
2715 address c2i_entry = __ pc();
2716
2717 // Class initialization barrier for static methods
2718 address c2i_no_clinit_check_entry = NULL;
2719 if (VM_Version::supports_fast_class_init_checks()) {
2720 Label L_skip_barrier;
2721
2722 { // Bypass the barrier for non-static methods
2723 __ testbit(Address(Z_method, Method::access_flags_offset()), JVM_ACC_STATIC_BIT);
2724 __ z_bfalse(L_skip_barrier); // non-static
2725 }
2726
2727 Register klass = Z_R11;
2728 __ load_method_holder(klass, Z_method);
2729 __ clinit_barrier(klass, Z_thread, &L_skip_barrier /*L_fast_path*/);
2730
2731 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub());
2732 __ z_br(klass);
2733
2734 __ bind(L_skip_barrier);
2735 c2i_no_clinit_check_entry = __ pc();
2736 }
2737
2738 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
2739
2740 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry);
2741 }
2742
2743 // This function returns the adjust size (in number of words) to a c2i adapter
2744 // activation for use during deoptimization.
2745 //
2746 // Actually only compiled frames need to be adjusted, but it
2747 // doesn't harm to adjust entry and interpreter frames, too.
2748 //
2749 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2750 assert(callee_locals >= callee_parameters,
2751 "test and remove; got more parms than locals");
2752 // Handle the abi adjustment here instead of doing it in push_skeleton_frames.
2753 return (callee_locals - callee_parameters) * Interpreter::stackElementWords +
2754 frame::z_parent_ijava_frame_abi_size / BytesPerWord;
2755 }
2756
2757 uint SharedRuntime::out_preserve_stack_slots() {
2758 return frame::z_jit_out_preserve_size/VMRegImpl::stack_slot_size;
2759 }
2760
2761 //
2762 // Frame generation for deopt and uncommon trap blobs.
2763 //
2764 static void push_skeleton_frame(MacroAssembler* masm,
2765 /* Unchanged */
2766 Register frame_sizes_reg,
2767 Register pcs_reg,
2768 /* Invalidate */
2769 Register frame_size_reg,
2770 Register pc_reg) {
2771 BLOCK_COMMENT(" push_skeleton_frame {");
2772 __ z_lg(pc_reg, 0, pcs_reg);
2773 __ z_lg(frame_size_reg, 0, frame_sizes_reg);
2774 __ z_stg(pc_reg, _z_abi(return_pc), Z_SP);
2775 Register fp = pc_reg;
2776 __ push_frame(frame_size_reg, fp);
2777 #ifdef ASSERT
2778 // The magic is required for successful walking skeletal frames.
2779 __ load_const_optimized(frame_size_reg/*tmp*/, frame::z_istate_magic_number);
2780 __ z_stg(frame_size_reg, _z_ijava_state_neg(magic), fp);
2781 // Fill other slots that are supposedly not necessary with eye catchers.
2782 __ load_const_optimized(frame_size_reg/*use as tmp*/, 0xdeadbad1);
2783 __ z_stg(frame_size_reg, _z_ijava_state_neg(top_frame_sp), fp);
2784 // The sender_sp of the bottom frame is set before pushing it.
2785 // The sender_sp of non bottom frames is their caller's top_frame_sp, which
2786 // is unknown here. Luckily it is not needed before filling the frame in
2787 // layout_activation(), we assert this by setting an eye catcher (see
2788 // comments on sender_sp in frame_s390.hpp).
2789 __ z_stg(frame_size_reg, _z_ijava_state_neg(sender_sp), Z_SP);
2790 #endif // ASSERT
2791 BLOCK_COMMENT(" } push_skeleton_frame");
2792 }
2793
2794 // Loop through the UnrollBlock info and create new frames.
2795 static void push_skeleton_frames(MacroAssembler* masm, bool deopt,
2796 /* read */
2797 Register unroll_block_reg,
2798 /* invalidate */
2799 Register frame_sizes_reg,
2800 Register number_of_frames_reg,
2801 Register pcs_reg,
2802 Register tmp1,
2803 Register tmp2) {
2804 BLOCK_COMMENT("push_skeleton_frames {");
2805 // _number_of_frames is of type int (deoptimization.hpp).
2806 __ z_lgf(number_of_frames_reg,
2807 Address(unroll_block_reg, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()));
2808 __ z_lg(pcs_reg,
2809 Address(unroll_block_reg, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()));
2810 __ z_lg(frame_sizes_reg,
2811 Address(unroll_block_reg, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()));
2812
2813 // stack: (caller_of_deoptee, ...).
2814
2815 // If caller_of_deoptee is a compiled frame, then we extend it to make
2816 // room for the callee's locals and the frame::z_parent_ijava_frame_abi.
2817 // See also Deoptimization::last_frame_adjust() above.
2818 // Note: entry and interpreted frames are adjusted, too. But this doesn't harm.
2819
2820 __ z_lgf(Z_R1_scratch,
2821 Address(unroll_block_reg, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()));
2822 __ z_lgr(tmp1, Z_SP); // Save the sender sp before extending the frame.
2823 __ resize_frame_sub(Z_R1_scratch, tmp2/*tmp*/);
2824 // The oldest skeletal frame requires a valid sender_sp to make it walkable
2825 // (it is required to find the original pc of caller_of_deoptee if it is marked
2826 // for deoptimization - see nmethod::orig_pc_addr()).
2827 __ z_stg(tmp1, _z_ijava_state_neg(sender_sp), Z_SP);
2828
2829 // Now push the new interpreter frames.
2830 Label loop, loop_entry;
2831
2832 // Make sure that there is at least one entry in the array.
2833 DEBUG_ONLY(__ z_ltgr(number_of_frames_reg, number_of_frames_reg));
2834 __ asm_assert_ne("array_size must be > 0", 0x205);
2835
2836 __ z_bru(loop_entry);
2837
2838 __ bind(loop);
2839
2840 __ add2reg(frame_sizes_reg, wordSize);
2841 __ add2reg(pcs_reg, wordSize);
2842
2843 __ bind(loop_entry);
2844
2845 // Allocate a new frame, fill in the pc.
2846 push_skeleton_frame(masm, frame_sizes_reg, pcs_reg, tmp1, tmp2);
2847
2848 __ z_aghi(number_of_frames_reg, -1); // Emit AGHI, because it sets the condition code
2849 __ z_brne(loop);
2850
2851 // Set the top frame's return pc.
2852 __ add2reg(pcs_reg, wordSize);
2853 __ z_lg(Z_R0_scratch, 0, pcs_reg);
2854 __ z_stg(Z_R0_scratch, _z_abi(return_pc), Z_SP);
2855 BLOCK_COMMENT("} push_skeleton_frames");
2856 }
2857
2858 //------------------------------generate_deopt_blob----------------------------
2859 void SharedRuntime::generate_deopt_blob() {
2860 // Allocate space for the code.
2861 ResourceMark rm;
2862 // Setup code generation tools.
2863 CodeBuffer buffer("deopt_blob", 2048, 1024);
2864 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2865 Label exec_mode_initialized;
2866 OopMap* map = NULL;
2867 OopMapSet *oop_maps = new OopMapSet();
2868
2869 unsigned int start_off = __ offset();
2870 Label cont;
2871
2872 // --------------------------------------------------------------------------
2873 // Normal entry (non-exception case)
2874 //
2875 // We have been called from the deopt handler of the deoptee.
2876 // Z_R14 points behind the call in the deopt handler. We adjust
2877 // it such that it points to the start of the deopt handler.
2878 // The return_pc has been stored in the frame of the deoptee and
2879 // will replace the address of the deopt_handler in the call
2880 // to Deoptimization::fetch_unroll_info below.
2881 // The (int) cast is necessary, because -((unsigned int)14)
2882 // is an unsigned int.
2883 __ add2reg(Z_R14, -(int)HandlerImpl::size_deopt_handler());
2884
2885 const Register exec_mode_reg = Z_tmp_1;
2886
2887 // stack: (deoptee, caller of deoptee, ...)
2888
2889 // pushes an "unpack" frame
2890 // R14 contains the return address pointing into the deoptimized
2891 // nmethod that was valid just before the nmethod was deoptimized.
2892 // save R14 into the deoptee frame. the `fetch_unroll_info'
2893 // procedure called below will read it from there.
2894 map = RegisterSaver::save_live_registers(masm, RegisterSaver::all_registers);
2895
2896 // note the entry point.
2897 __ load_const_optimized(exec_mode_reg, Deoptimization::Unpack_deopt);
2898 __ z_bru(exec_mode_initialized);
2899
2900 #ifndef COMPILER1
2901 int reexecute_offset = 1; // odd offset will produce odd pc, which triggers an hardware trap
2902 #else
2903 // --------------------------------------------------------------------------
2904 // Reexecute entry
2905 // - Z_R14 = Deopt Handler in nmethod
2906
2907 int reexecute_offset = __ offset() - start_off;
2908
2909 // No need to update map as each call to save_live_registers will produce identical oopmap
2910 (void) RegisterSaver::save_live_registers(masm, RegisterSaver::all_registers);
2911
2912 __ load_const_optimized(exec_mode_reg, Deoptimization::Unpack_reexecute);
2913 __ z_bru(exec_mode_initialized);
2914 #endif
2915
2916
2917 // --------------------------------------------------------------------------
2918 // Exception entry. We reached here via a branch. Registers on entry:
2919 // - Z_EXC_OOP (Z_ARG1) = exception oop
2920 // - Z_EXC_PC (Z_ARG2) = the exception pc.
2921
2922 int exception_offset = __ offset() - start_off;
2923
2924 // all registers are dead at this entry point, except for Z_EXC_OOP, and
2925 // Z_EXC_PC which contain the exception oop and exception pc
2926 // respectively. Set them in TLS and fall thru to the
2927 // unpack_with_exception_in_tls entry point.
2928
2929 // Store exception oop and pc in thread (location known to GC).
2930 // Need this since the call to "fetch_unroll_info()" may safepoint.
2931 __ z_stg(Z_EXC_OOP, Address(Z_thread, JavaThread::exception_oop_offset()));
2932 __ z_stg(Z_EXC_PC, Address(Z_thread, JavaThread::exception_pc_offset()));
2933
2934 // fall through
2935
2936 int exception_in_tls_offset = __ offset() - start_off;
2937
2938 // new implementation because exception oop is now passed in JavaThread
2939
2940 // Prolog for exception case
2941 // All registers must be preserved because they might be used by LinearScan
2942 // Exceptiop oop and throwing PC are passed in JavaThread
2943
2944 // load throwing pc from JavaThread and us it as the return address of the current frame.
2945 __ z_lg(Z_R1_scratch, Address(Z_thread, JavaThread::exception_pc_offset()));
2946
2947 // Save everything in sight.
2948 (void) RegisterSaver::save_live_registers(masm, RegisterSaver::all_registers, Z_R1_scratch);
2949
2950 // Now it is safe to overwrite any register
2951
2952 // Clear the exception pc field in JavaThread
2953 __ clear_mem(Address(Z_thread, JavaThread::exception_pc_offset()), 8);
2954
2955 // Deopt during an exception. Save exec mode for unpack_frames.
2956 __ load_const_optimized(exec_mode_reg, Deoptimization::Unpack_exception);
2957
2958
2959 #ifdef ASSERT
2960 // verify that there is really an exception oop in JavaThread
2961 __ z_lg(Z_ARG1, Address(Z_thread, JavaThread::exception_oop_offset()));
2962 __ verify_oop(Z_ARG1);
2963
2964 // verify that there is no pending exception
2965 __ asm_assert_mem8_is_zero(in_bytes(Thread::pending_exception_offset()), Z_thread,
2966 "must not have pending exception here", __LINE__);
2967 #endif
2968
2969 // --------------------------------------------------------------------------
2970 // At this point, the live registers are saved and
2971 // the exec_mode_reg has been set up correctly.
2972 __ bind(exec_mode_initialized);
2973
2974 // stack: ("unpack" frame, deoptee, caller_of_deoptee, ...).
2975
2976 {
2977 const Register unroll_block_reg = Z_tmp_2;
2978
2979 // we need to set `last_Java_frame' because `fetch_unroll_info' will
2980 // call `last_Java_frame()'. however we can't block and no gc will
2981 // occur so we don't need an oopmap. the value of the pc in the
2982 // frame is not particularly important. it just needs to identify the blob.
2983
2984 // Don't set last_Java_pc anymore here (is implicitly NULL then).
2985 // the correct PC is retrieved in pd_last_frame() in that case.
2986 __ set_last_Java_frame(/*sp*/Z_SP, noreg);
2987 // With EscapeAnalysis turned on, this call may safepoint
2988 // despite it's marked as "leaf call"!
2989 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), Z_thread, exec_mode_reg);
2990 // Set an oopmap for the call site this describes all our saved volatile registers
2991 int offs = __ offset();
2992 oop_maps->add_gc_map(offs, map);
2993
2994 __ reset_last_Java_frame();
2995 // save the return value.
2996 __ z_lgr(unroll_block_reg, Z_RET);
2997 // restore the return registers that have been saved
2998 // (among other registers) by save_live_registers(...).
2999 RegisterSaver::restore_result_registers(masm);
3000
3001 // reload the exec mode from the UnrollBlock (it might have changed)
3002 __ z_llgf(exec_mode_reg, Address(unroll_block_reg, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()));
3003
3004 // In excp_deopt_mode, restore and clear exception oop which we
3005 // stored in the thread during exception entry above. The exception
3006 // oop will be the return value of this stub.
3007 NearLabel skip_restore_excp;
3008 __ compare64_and_branch(exec_mode_reg, Deoptimization::Unpack_exception, Assembler::bcondNotEqual, skip_restore_excp);
3009 __ z_lg(Z_RET, thread_(exception_oop));
3010 __ clear_mem(thread_(exception_oop), 8);
3011 __ bind(skip_restore_excp);
3012
3013 // remove the "unpack" frame
3014 __ pop_frame();
3015
3016 // stack: (deoptee, caller of deoptee, ...).
3017
3018 // pop the deoptee's frame
3019 __ pop_frame();
3020
3021 // stack: (caller_of_deoptee, ...).
3022
3023 // loop through the `UnrollBlock' info and create interpreter frames.
3024 push_skeleton_frames(masm, true/*deopt*/,
3025 unroll_block_reg,
3026 Z_tmp_3,
3027 Z_tmp_4,
3028 Z_ARG5,
3029 Z_ARG4,
3030 Z_ARG3);
3031
3032 // stack: (skeletal interpreter frame, ..., optional skeletal
3033 // interpreter frame, caller of deoptee, ...).
3034 }
3035
3036 // push an "unpack" frame taking care of float / int return values.
3037 __ push_frame(RegisterSaver::live_reg_frame_size(RegisterSaver::all_registers));
3038
3039 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3040 // skeletal interpreter frame, caller of deoptee, ...).
3041
3042 // spill live volatile registers since we'll do a call.
3043 __ z_stg(Z_RET, offset_of(frame::z_abi_160_spill, spill[0]), Z_SP);
3044 __ z_std(Z_FRET, offset_of(frame::z_abi_160_spill, spill[1]), Z_SP);
3045
3046 // let the unpacker layout information in the skeletal frames just allocated.
3047 __ get_PC(Z_RET);
3048 __ set_last_Java_frame(/*sp*/Z_SP, /*pc*/Z_RET);
3049 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3050 Z_thread/*thread*/, exec_mode_reg/*exec_mode*/);
3051
3052 __ reset_last_Java_frame();
3053
3054 // restore the volatiles saved above.
3055 __ z_lg(Z_RET, offset_of(frame::z_abi_160_spill, spill[0]), Z_SP);
3056 __ z_ld(Z_FRET, offset_of(frame::z_abi_160_spill, spill[1]), Z_SP);
3057
3058 // pop the "unpack" frame.
3059 __ pop_frame();
3060 __ restore_return_pc();
3061
3062 // stack: (top interpreter frame, ..., optional interpreter frame,
3063 // caller of deoptee, ...).
3064
3065 __ z_lg(Z_fp, _z_abi(callers_sp), Z_SP); // restore frame pointer
3066 __ restore_bcp();
3067 __ restore_locals();
3068 __ restore_esp();
3069
3070 // return to the interpreter entry point.
3071 __ z_br(Z_R14);
3072
3073 // Make sure all code is generated
3074 masm->flush();
3075
3076 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, RegisterSaver::live_reg_frame_size(RegisterSaver::all_registers)/wordSize);
3077 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3078 }
3079
3080
3081 #ifdef COMPILER2
3082 //------------------------------generate_uncommon_trap_blob--------------------
3083 void SharedRuntime::generate_uncommon_trap_blob() {
3084 // Allocate space for the code
3085 ResourceMark rm;
3086 // Setup code generation tools
3087 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024);
3088 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
3089
3090 Register unroll_block_reg = Z_tmp_1;
3091 Register klass_index_reg = Z_ARG2;
3092 Register unc_trap_reg = Z_ARG2;
3093
3094 // stack: (deoptee, caller_of_deoptee, ...).
3095
3096 // push a dummy "unpack" frame and call
3097 // `Deoptimization::uncommon_trap' to pack the compiled frame into a
3098 // vframe array and return the `UnrollBlock' information.
3099
3100 // save R14 to compiled frame.
3101 __ save_return_pc();
3102 // push the "unpack_frame".
3103 __ push_frame_abi160(0);
3104
3105 // stack: (unpack frame, deoptee, caller_of_deoptee, ...).
3106
3107 // set the "unpack" frame as last_Java_frame.
3108 // `Deoptimization::uncommon_trap' expects it and considers its
3109 // sender frame as the deoptee frame.
3110 __ get_PC(Z_R1_scratch);
3111 __ set_last_Java_frame(/*sp*/Z_SP, /*pc*/Z_R1_scratch);
3112
3113 __ z_lgr(klass_index_reg, Z_ARG1); // passed implicitly as ARG2
3114 __ z_lghi(Z_ARG3, Deoptimization::Unpack_uncommon_trap); // passed implicitly as ARG3
3115 BLOCK_COMMENT("call Deoptimization::uncommon_trap()");
3116 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), Z_thread);
3117
3118 __ reset_last_Java_frame();
3119
3120 // pop the "unpack" frame
3121 __ pop_frame();
3122
3123 // stack: (deoptee, caller_of_deoptee, ...).
3124
3125 // save the return value.
3126 __ z_lgr(unroll_block_reg, Z_RET);
3127
3128 // pop the deoptee frame.
3129 __ pop_frame();
3130
3131 // stack: (caller_of_deoptee, ...).
3132
3133 #ifdef ASSERT
3134 assert(Immediate::is_uimm8(Deoptimization::Unpack_LIMIT), "Code not fit for larger immediates");
3135 assert(Immediate::is_uimm8(Deoptimization::Unpack_uncommon_trap), "Code not fit for larger immediates");
3136 const int unpack_kind_byte_offset = Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()
3137 #ifndef VM_LITTLE_ENDIAN
3138 + 3
3139 #endif
3140 ;
3141 if (Displacement::is_shortDisp(unpack_kind_byte_offset)) {
3142 __ z_cli(unpack_kind_byte_offset, unroll_block_reg, Deoptimization::Unpack_uncommon_trap);
3143 } else {
3144 __ z_cliy(unpack_kind_byte_offset, unroll_block_reg, Deoptimization::Unpack_uncommon_trap);
3145 }
3146 __ asm_assert_eq("SharedRuntime::generate_deopt_blob: expected Unpack_uncommon_trap", 0);
3147 #endif
3148
3149 __ zap_from_to(Z_SP, Z_SP, Z_R0_scratch, Z_R1, 500, -1);
3150
3151 // allocate new interpreter frame(s) and possibly resize the caller's frame
3152 // (no more adapters !)
3153 push_skeleton_frames(masm, false/*deopt*/,
3154 unroll_block_reg,
3155 Z_tmp_2,
3156 Z_tmp_3,
3157 Z_tmp_4,
3158 Z_ARG5,
3159 Z_ARG4);
3160
3161 // stack: (skeletal interpreter frame, ..., optional skeletal
3162 // interpreter frame, (resized) caller of deoptee, ...).
3163
3164 // push a dummy "unpack" frame taking care of float return values.
3165 // call `Deoptimization::unpack_frames' to layout information in the
3166 // interpreter frames just created
3167
3168 // push the "unpack" frame
3169 const unsigned int framesize_in_bytes = __ push_frame_abi160(0);
3170
3171 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3172 // skeletal interpreter frame, (resized) caller of deoptee, ...).
3173
3174 // set the "unpack" frame as last_Java_frame
3175 __ get_PC(Z_R1_scratch);
3176 __ set_last_Java_frame(/*sp*/Z_SP, /*pc*/Z_R1_scratch);
3177
3178 // indicate it is the uncommon trap case
3179 BLOCK_COMMENT("call Deoptimization::Unpack_uncommon_trap()");
3180 __ load_const_optimized(unc_trap_reg, Deoptimization::Unpack_uncommon_trap);
3181 // let the unpacker layout information in the skeletal frames just allocated.
3182 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), Z_thread);
3183
3184 __ reset_last_Java_frame();
3185 // pop the "unpack" frame
3186 __ pop_frame();
3187 // restore LR from top interpreter frame
3188 __ restore_return_pc();
3189
3190 // stack: (top interpreter frame, ..., optional interpreter frame,
3191 // (resized) caller of deoptee, ...).
3192
3193 __ z_lg(Z_fp, _z_abi(callers_sp), Z_SP); // restore frame pointer
3194 __ restore_bcp();
3195 __ restore_locals();
3196 __ restore_esp();
3197
3198 // return to the interpreter entry point
3199 __ z_br(Z_R14);
3200
3201 masm->flush();
3202 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, NULL, framesize_in_bytes/wordSize);
3203 }
3204 #endif // COMPILER2
3205
3206
3207 //------------------------------generate_handler_blob------
3208 //
3209 // Generate a special Compile2Runtime blob that saves all registers,
3210 // and setup oopmap.
3211 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
3212 assert(StubRoutines::forward_exception_entry() != NULL,
3213 "must be generated before");
3214
3215 ResourceMark rm;
3216 OopMapSet *oop_maps = new OopMapSet();
3217 OopMap* map;
3218
3219 // Allocate space for the code. Setup code generation tools.
3220 CodeBuffer buffer("handler_blob", 2048, 1024);
3221 MacroAssembler* masm = new MacroAssembler(&buffer);
3222
3223 unsigned int start_off = __ offset();
3224 address call_pc = NULL;
3225 int frame_size_in_bytes;
3226
3227 bool cause_return = (poll_type == POLL_AT_RETURN);
3228 // Make room for return address (or push it again)
3229 if (!cause_return) {
3230 __ z_lg(Z_R14, Address(Z_thread, JavaThread::saved_exception_pc_offset()));
3231 }
3232
3233 // Save registers, fpu state, and flags
3234 map = RegisterSaver::save_live_registers(masm, RegisterSaver::all_registers);
3235
3236 if (SafepointMechanism::uses_thread_local_poll() && !cause_return) {
3237 // Keep a copy of the return pc to detect if it gets modified.
3238 __ z_lgr(Z_R6, Z_R14);
3239 }
3240
3241 // The following is basically a call_VM. However, we need the precise
3242 // address of the call in order to generate an oopmap. Hence, we do all the
3243 // work outselves.
3244 __ set_last_Java_frame(Z_SP, noreg);
3245
3246 // call into the runtime to handle the safepoint poll
3247 __ call_VM_leaf(call_ptr, Z_thread);
3248
3249
3250 // Set an oopmap for the call site. This oopmap will map all
3251 // oop-registers and debug-info registers as callee-saved. This
3252 // will allow deoptimization at this safepoint to find all possible
3253 // debug-info recordings, as well as let GC find all oops.
3254
3255 oop_maps->add_gc_map((int)(__ offset()-start_off), map);
3256
3257 Label noException;
3258
3259 __ reset_last_Java_frame();
3260
3261 __ load_and_test_long(Z_R1, thread_(pending_exception));
3262 __ z_bre(noException);
3263
3264 // Pending exception case, used (sporadically) by
3265 // api/java_lang/Thread.State/index#ThreadState et al.
3266 RegisterSaver::restore_live_registers(masm, RegisterSaver::all_registers);
3267
3268 // Jump to forward_exception_entry, with the issuing PC in Z_R14
3269 // so it looks like the original nmethod called forward_exception_entry.
3270 __ load_const_optimized(Z_R1_scratch, StubRoutines::forward_exception_entry());
3271 __ z_br(Z_R1_scratch);
3272
3273 // No exception case
3274 __ bind(noException);
3275
3276 if (SafepointMechanism::uses_thread_local_poll() && !cause_return) {
3277 Label no_adjust;
3278 // If our stashed return pc was modified by the runtime we avoid touching it
3279 const int offset_of_return_pc = _z_abi16(return_pc) + RegisterSaver::live_reg_frame_size(RegisterSaver::all_registers);
3280 __ z_cg(Z_R6, offset_of_return_pc, Z_SP);
3281 __ z_brne(no_adjust);
3282
3283 // Adjust return pc forward to step over the safepoint poll instruction
3284 __ instr_size(Z_R1_scratch, Z_R6);
3285 __ z_agr(Z_R6, Z_R1_scratch);
3286 __ z_stg(Z_R6, offset_of_return_pc, Z_SP);
3287
3288 __ bind(no_adjust);
3289 }
3290
3291 // Normal exit, restore registers and exit.
3292 RegisterSaver::restore_live_registers(masm, RegisterSaver::all_registers);
3293
3294 __ z_br(Z_R14);
3295
3296 // Make sure all code is generated
3297 masm->flush();
3298
3299 // Fill-out other meta info
3300 return SafepointBlob::create(&buffer, oop_maps, RegisterSaver::live_reg_frame_size(RegisterSaver::all_registers)/wordSize);
3301 }
3302
3303
3304 //
3305 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3306 //
3307 // Generate a stub that calls into vm to find out the proper destination
3308 // of a Java call. All the argument registers are live at this point
3309 // but since this is generic code we don't know what they are and the caller
3310 // must do any gc of the args.
3311 //
3312 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3313 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3314
3315 // allocate space for the code
3316 ResourceMark rm;
3317
3318 CodeBuffer buffer(name, 1000, 512);
3319 MacroAssembler* masm = new MacroAssembler(&buffer);
3320
3321 OopMapSet *oop_maps = new OopMapSet();
3322 OopMap* map = NULL;
3323
3324 unsigned int start_off = __ offset();
3325
3326 map = RegisterSaver::save_live_registers(masm, RegisterSaver::all_registers);
3327
3328 // We must save a PC from within the stub as return PC
3329 // C code doesn't store the LR where we expect the PC,
3330 // so we would run into trouble upon stack walking.
3331 __ get_PC(Z_R1_scratch);
3332
3333 unsigned int frame_complete = __ offset();
3334
3335 __ set_last_Java_frame(/*sp*/Z_SP, Z_R1_scratch);
3336
3337 __ call_VM_leaf(destination, Z_thread, Z_method);
3338
3339
3340 // Set an oopmap for the call site.
3341 // We need this not only for callee-saved registers, but also for volatile
3342 // registers that the compiler might be keeping live across a safepoint.
3343
3344 oop_maps->add_gc_map((int)(frame_complete-start_off), map);
3345
3346 // clear last_Java_sp
3347 __ reset_last_Java_frame();
3348
3349 // check for pending exceptions
3350 Label pending;
3351 __ load_and_test_long(Z_R0, Address(Z_thread, Thread::pending_exception_offset()));
3352 __ z_brne(pending);
3353
3354 __ z_lgr(Z_R1_scratch, Z_R2); // r1 is neither saved nor restored, r2 contains the continuation.
3355 RegisterSaver::restore_live_registers(masm, RegisterSaver::all_registers);
3356
3357 // get the returned method
3358 __ get_vm_result_2(Z_method);
3359
3360 // We are back the the original state on entry and ready to go.
3361 __ z_br(Z_R1_scratch);
3362
3363 // Pending exception after the safepoint
3364
3365 __ bind(pending);
3366
3367 RegisterSaver::restore_live_registers(masm, RegisterSaver::all_registers);
3368
3369 // exception pending => remove activation and forward to exception handler
3370
3371 __ z_lgr(Z_R2, Z_R0); // pending_exception
3372 __ clear_mem(Address(Z_thread, JavaThread::vm_result_offset()), sizeof(jlong));
3373 __ load_const_optimized(Z_R1_scratch, StubRoutines::forward_exception_entry());
3374 __ z_br(Z_R1_scratch);
3375
3376 // -------------
3377 // make sure all code is generated
3378 masm->flush();
3379
3380 // return the blob
3381 // frame_size_words or bytes??
3382 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, RegisterSaver::live_reg_frame_size(RegisterSaver::all_registers)/wordSize,
3383 oop_maps, true);
3384
3385 }
3386
3387 //------------------------------Montgomery multiplication------------------------
3388 //
3389
3390 // Subtract 0:b from carry:a. Return carry.
3391 static unsigned long
3392 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) {
3393 unsigned long i, c = 8 * (unsigned long)(len - 1);
3394 __asm__ __volatile__ (
3395 "SLGR %[i], %[i] \n" // initialize to 0 and pre-set carry
3396 "LGHI 0, 8 \n" // index increment (for BRXLG)
3397 "LGR 1, %[c] \n" // index limit (for BRXLG)
3398 "0: \n"
3399 "LG %[c], 0(%[i],%[a]) \n"
3400 "SLBG %[c], 0(%[i],%[b]) \n" // subtract with borrow
3401 "STG %[c], 0(%[i],%[a]) \n"
3402 "BRXLG %[i], 0, 0b \n" // while ((i+=8)<limit);
3403 "SLBGR %[c], %[c] \n" // save carry - 1
3404 : [i]"=&a"(i), [c]"+r"(c)
3405 : [a]"a"(a), [b]"a"(b)
3406 : "cc", "memory", "r0", "r1"
3407 );
3408 return carry + c;
3409 }
3410
3411 // Multiply (unsigned) Long A by Long B, accumulating the double-
3412 // length result into the accumulator formed of T0, T1, and T2.
3413 inline void MACC(unsigned long A[], long A_ind,
3414 unsigned long B[], long B_ind,
3415 unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3416 long A_si = 8 * A_ind,
3417 B_si = 8 * B_ind;
3418 __asm__ __volatile__ (
3419 "LG 1, 0(%[A_si],%[A]) \n"
3420 "MLG 0, 0(%[B_si],%[B]) \n" // r0r1 = A * B
3421 "ALGR %[T0], 1 \n"
3422 "LGHI 1, 0 \n" // r1 = 0
3423 "ALCGR %[T1], 0 \n"
3424 "ALCGR %[T2], 1 \n"
3425 : [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3426 : [A]"r"(A), [A_si]"r"(A_si), [B]"r"(B), [B_si]"r"(B_si)
3427 : "cc", "r0", "r1"
3428 );
3429 }
3430
3431 // As above, but add twice the double-length result into the
3432 // accumulator.
3433 inline void MACC2(unsigned long A[], long A_ind,
3434 unsigned long B[], long B_ind,
3435 unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3436 const unsigned long zero = 0;
3437 long A_si = 8 * A_ind,
3438 B_si = 8 * B_ind;
3439 __asm__ __volatile__ (
3440 "LG 1, 0(%[A_si],%[A]) \n"
3441 "MLG 0, 0(%[B_si],%[B]) \n" // r0r1 = A * B
3442 "ALGR %[T0], 1 \n"
3443 "ALCGR %[T1], 0 \n"
3444 "ALCGR %[T2], %[zero] \n"
3445 "ALGR %[T0], 1 \n"
3446 "ALCGR %[T1], 0 \n"
3447 "ALCGR %[T2], %[zero] \n"
3448 : [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3449 : [A]"r"(A), [A_si]"r"(A_si), [B]"r"(B), [B_si]"r"(B_si), [zero]"r"(zero)
3450 : "cc", "r0", "r1"
3451 );
3452 }
3453
3454 // Fast Montgomery multiplication. The derivation of the algorithm is
3455 // in "A Cryptographic Library for the Motorola DSP56000,
3456 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237".
3457 static void
3458 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[],
3459 unsigned long m[], unsigned long inv, int len) {
3460 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3461 int i;
3462
3463 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3464
3465 for (i = 0; i < len; i++) {
3466 int j;
3467 for (j = 0; j < i; j++) {
3468 MACC(a, j, b, i-j, t0, t1, t2);
3469 MACC(m, j, n, i-j, t0, t1, t2);
3470 }
3471 MACC(a, i, b, 0, t0, t1, t2);
3472 m[i] = t0 * inv;
3473 MACC(m, i, n, 0, t0, t1, t2);
3474
3475 assert(t0 == 0, "broken Montgomery multiply");
3476
3477 t0 = t1; t1 = t2; t2 = 0;
3478 }
3479
3480 for (i = len; i < 2 * len; i++) {
3481 int j;
3482 for (j = i - len + 1; j < len; j++) {
3483 MACC(a, j, b, i-j, t0, t1, t2);
3484 MACC(m, j, n, i-j, t0, t1, t2);
3485 }
3486 m[i-len] = t0;
3487 t0 = t1; t1 = t2; t2 = 0;
3488 }
3489
3490 while (t0) {
3491 t0 = sub(m, n, t0, len);
3492 }
3493 }
3494
3495 // Fast Montgomery squaring. This uses asymptotically 25% fewer
3496 // multiplies so it should be up to 25% faster than Montgomery
3497 // multiplication. However, its loop control is more complex and it
3498 // may actually run slower on some machines.
3499 static void
3500 montgomery_square(unsigned long a[], unsigned long n[],
3501 unsigned long m[], unsigned long inv, int len) {
3502 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3503 int i;
3504
3505 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3506
3507 for (i = 0; i < len; i++) {
3508 int j;
3509 int end = (i+1)/2;
3510 for (j = 0; j < end; j++) {
3511 MACC2(a, j, a, i-j, t0, t1, t2);
3512 MACC(m, j, n, i-j, t0, t1, t2);
3513 }
3514 if ((i & 1) == 0) {
3515 MACC(a, j, a, j, t0, t1, t2);
3516 }
3517 for (; j < i; j++) {
3518 MACC(m, j, n, i-j, t0, t1, t2);
3519 }
3520 m[i] = t0 * inv;
3521 MACC(m, i, n, 0, t0, t1, t2);
3522
3523 assert(t0 == 0, "broken Montgomery square");
3524
3525 t0 = t1; t1 = t2; t2 = 0;
3526 }
3527
3528 for (i = len; i < 2*len; i++) {
3529 int start = i-len+1;
3530 int end = start + (len - start)/2;
3531 int j;
3532 for (j = start; j < end; j++) {
3533 MACC2(a, j, a, i-j, t0, t1, t2);
3534 MACC(m, j, n, i-j, t0, t1, t2);
3535 }
3536 if ((i & 1) == 0) {
3537 MACC(a, j, a, j, t0, t1, t2);
3538 }
3539 for (; j < len; j++) {
3540 MACC(m, j, n, i-j, t0, t1, t2);
3541 }
3542 m[i-len] = t0;
3543 t0 = t1; t1 = t2; t2 = 0;
3544 }
3545
3546 while (t0) {
3547 t0 = sub(m, n, t0, len);
3548 }
3549 }
3550
3551 // The threshold at which squaring is advantageous was determined
3552 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz.
3553 // Value seems to be ok for other platforms, too.
3554 #define MONTGOMERY_SQUARING_THRESHOLD 64
3555
3556 // Copy len longwords from s to d, word-swapping as we go. The
3557 // destination array is reversed.
3558 static void reverse_words(unsigned long *s, unsigned long *d, int len) {
3559 d += len;
3560 while(len-- > 0) {
3561 d--;
3562 unsigned long s_val = *s;
3563 // Swap words in a longword on little endian machines.
3564 #ifdef VM_LITTLE_ENDIAN
3565 Unimplemented();
3566 #endif
3567 *d = s_val;
3568 s++;
3569 }
3570 }
3571
3572 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints,
3573 jint len, jlong inv,
3574 jint *m_ints) {
3575 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3576 assert(len % 2 == 0, "array length in montgomery_multiply must be even");
3577 int longwords = len/2;
3578
3579 // Make very sure we don't use so much space that the stack might
3580 // overflow. 512 jints corresponds to an 16384-bit integer and
3581 // will use here a total of 8k bytes of stack space.
3582 int total_allocation = longwords * sizeof (unsigned long) * 4;
3583 guarantee(total_allocation <= 8192, "must be");
3584 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3585
3586 // Local scratch arrays
3587 unsigned long
3588 *a = scratch + 0 * longwords,
3589 *b = scratch + 1 * longwords,
3590 *n = scratch + 2 * longwords,
3591 *m = scratch + 3 * longwords;
3592
3593 reverse_words((unsigned long *)a_ints, a, longwords);
3594 reverse_words((unsigned long *)b_ints, b, longwords);
3595 reverse_words((unsigned long *)n_ints, n, longwords);
3596
3597 ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords);
3598
3599 reverse_words(m, (unsigned long *)m_ints, longwords);
3600 }
3601
3602 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints,
3603 jint len, jlong inv,
3604 jint *m_ints) {
3605 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3606 assert(len % 2 == 0, "array length in montgomery_square must be even");
3607 int longwords = len/2;
3608
3609 // Make very sure we don't use so much space that the stack might
3610 // overflow. 512 jints corresponds to an 16384-bit integer and
3611 // will use here a total of 6k bytes of stack space.
3612 int total_allocation = longwords * sizeof (unsigned long) * 3;
3613 guarantee(total_allocation <= 8192, "must be");
3614 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3615
3616 // Local scratch arrays
3617 unsigned long
3618 *a = scratch + 0 * longwords,
3619 *n = scratch + 1 * longwords,
3620 *m = scratch + 2 * longwords;
3621
3622 reverse_words((unsigned long *)a_ints, a, longwords);
3623 reverse_words((unsigned long *)n_ints, n, longwords);
3624
3625 if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3626 ::montgomery_square(a, n, m, (unsigned long)inv, longwords);
3627 } else {
3628 ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords);
3629 }
3630
3631 reverse_words(m, (unsigned long *)m_ints, longwords);
3632 }
3633
3634 extern "C"
3635 int SpinPause() {
3636 return 0;
3637 }