1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2012, 2025 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 "asm/macroAssembler.inline.hpp"
27 #include "code/debugInfoRec.hpp"
28 #include "code/compiledIC.hpp"
29 #include "code/vtableStubs.hpp"
30 #include "frame_ppc.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "gc/shared/gcLocker.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "interpreter/interp_masm.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "oops/klass.inline.hpp"
37 #include "prims/methodHandles.hpp"
38 #include "runtime/continuation.hpp"
39 #include "runtime/continuationEntry.inline.hpp"
40 #include "runtime/jniHandles.hpp"
41 #include "runtime/os.inline.hpp"
42 #include "runtime/safepointMechanism.hpp"
43 #include "runtime/sharedRuntime.hpp"
44 #include "runtime/signature.hpp"
45 #include "runtime/stubRoutines.hpp"
46 #include "runtime/timerTrace.hpp"
47 #include "runtime/vframeArray.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/macros.hpp"
50 #include "vmreg_ppc.inline.hpp"
51 #ifdef COMPILER1
52 #include "c1/c1_Runtime1.hpp"
53 #endif
54 #ifdef COMPILER2
55 #include "opto/ad.hpp"
56 #include "opto/runtime.hpp"
57 #endif
58
59 #include <alloca.h>
60
61 #define __ masm->
62
63 #ifdef PRODUCT
64 #define BLOCK_COMMENT(str) // nothing
65 #else
66 #define BLOCK_COMMENT(str) __ block_comment(str)
67 #endif
68
69 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
70
71
72 class RegisterSaver {
73 // Used for saving volatile registers.
74 public:
75
76 // Support different return pc locations.
77 enum ReturnPCLocation {
78 return_pc_is_lr,
79 return_pc_is_pre_saved,
80 return_pc_is_thread_saved_exception_pc
81 };
82
83 static OopMap* push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
84 int* out_frame_size_in_bytes,
85 bool generate_oop_map,
86 int return_pc_adjustment,
87 ReturnPCLocation return_pc_location,
88 bool save_vectors = false);
89 static void restore_live_registers_and_pop_frame(MacroAssembler* masm,
90 int frame_size_in_bytes,
91 bool restore_ctr,
92 bool save_vectors = false);
93
94 static void push_frame_and_save_argument_registers(MacroAssembler* masm,
95 Register r_temp,
96 int frame_size,
97 int total_args,
98 const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
99 static void restore_argument_registers_and_pop_frame(MacroAssembler*masm,
100 int frame_size,
101 int total_args,
102 const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
103
104 // During deoptimization only the result registers need to be restored
105 // all the other values have already been extracted.
106 static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes);
107
108 // Constants and data structures:
109
110 typedef enum {
111 int_reg,
112 float_reg,
113 special_reg,
114 vec_reg
115 } RegisterType;
116
117 typedef enum {
118 reg_size = 8,
119 half_reg_size = reg_size / 2,
120 vec_reg_size = 16
121 } RegisterConstants;
122
123 typedef struct {
124 RegisterType reg_type;
125 int reg_num;
126 VMReg vmreg;
127 } LiveRegType;
128 };
129
130
131 #define RegisterSaver_LiveIntReg(regname) \
132 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() }
133
134 #define RegisterSaver_LiveFloatReg(regname) \
135 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() }
136
137 #define RegisterSaver_LiveSpecialReg(regname) \
138 { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() }
139
140 #define RegisterSaver_LiveVecReg(regname) \
141 { RegisterSaver::vec_reg, regname->encoding(), regname->as_VMReg() }
142
143 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = {
144 // Live registers which get spilled to the stack. Register
145 // positions in this array correspond directly to the stack layout.
146
147 //
148 // live special registers:
149 //
150 RegisterSaver_LiveSpecialReg(SR_CTR),
151 //
152 // live float registers:
153 //
154 RegisterSaver_LiveFloatReg( F0 ),
155 RegisterSaver_LiveFloatReg( F1 ),
156 RegisterSaver_LiveFloatReg( F2 ),
157 RegisterSaver_LiveFloatReg( F3 ),
158 RegisterSaver_LiveFloatReg( F4 ),
159 RegisterSaver_LiveFloatReg( F5 ),
160 RegisterSaver_LiveFloatReg( F6 ),
161 RegisterSaver_LiveFloatReg( F7 ),
162 RegisterSaver_LiveFloatReg( F8 ),
163 RegisterSaver_LiveFloatReg( F9 ),
164 RegisterSaver_LiveFloatReg( F10 ),
165 RegisterSaver_LiveFloatReg( F11 ),
166 RegisterSaver_LiveFloatReg( F12 ),
167 RegisterSaver_LiveFloatReg( F13 ),
168 RegisterSaver_LiveFloatReg( F14 ),
169 RegisterSaver_LiveFloatReg( F15 ),
170 RegisterSaver_LiveFloatReg( F16 ),
171 RegisterSaver_LiveFloatReg( F17 ),
172 RegisterSaver_LiveFloatReg( F18 ),
173 RegisterSaver_LiveFloatReg( F19 ),
174 RegisterSaver_LiveFloatReg( F20 ),
175 RegisterSaver_LiveFloatReg( F21 ),
176 RegisterSaver_LiveFloatReg( F22 ),
177 RegisterSaver_LiveFloatReg( F23 ),
178 RegisterSaver_LiveFloatReg( F24 ),
179 RegisterSaver_LiveFloatReg( F25 ),
180 RegisterSaver_LiveFloatReg( F26 ),
181 RegisterSaver_LiveFloatReg( F27 ),
182 RegisterSaver_LiveFloatReg( F28 ),
183 RegisterSaver_LiveFloatReg( F29 ),
184 RegisterSaver_LiveFloatReg( F30 ),
185 RegisterSaver_LiveFloatReg( F31 ),
186 //
187 // live integer registers:
188 //
189 RegisterSaver_LiveIntReg( R0 ),
190 //RegisterSaver_LiveIntReg( R1 ), // stack pointer
191 RegisterSaver_LiveIntReg( R2 ),
192 RegisterSaver_LiveIntReg( R3 ),
193 RegisterSaver_LiveIntReg( R4 ),
194 RegisterSaver_LiveIntReg( R5 ),
195 RegisterSaver_LiveIntReg( R6 ),
196 RegisterSaver_LiveIntReg( R7 ),
197 RegisterSaver_LiveIntReg( R8 ),
198 RegisterSaver_LiveIntReg( R9 ),
199 RegisterSaver_LiveIntReg( R10 ),
200 RegisterSaver_LiveIntReg( R11 ),
201 RegisterSaver_LiveIntReg( R12 ),
202 //RegisterSaver_LiveIntReg( R13 ), // system thread id
203 RegisterSaver_LiveIntReg( R14 ),
204 RegisterSaver_LiveIntReg( R15 ),
205 RegisterSaver_LiveIntReg( R16 ),
206 RegisterSaver_LiveIntReg( R17 ),
207 RegisterSaver_LiveIntReg( R18 ),
208 RegisterSaver_LiveIntReg( R19 ),
209 RegisterSaver_LiveIntReg( R20 ),
210 RegisterSaver_LiveIntReg( R21 ),
211 RegisterSaver_LiveIntReg( R22 ),
212 RegisterSaver_LiveIntReg( R23 ),
213 RegisterSaver_LiveIntReg( R24 ),
214 RegisterSaver_LiveIntReg( R25 ),
215 RegisterSaver_LiveIntReg( R26 ),
216 RegisterSaver_LiveIntReg( R27 ),
217 RegisterSaver_LiveIntReg( R28 ),
218 RegisterSaver_LiveIntReg( R29 ),
219 RegisterSaver_LiveIntReg( R30 ),
220 RegisterSaver_LiveIntReg( R31 ) // must be the last register (see save/restore functions below)
221 };
222
223 static const RegisterSaver::LiveRegType RegisterSaver_LiveVecRegs[] = {
224 //
225 // live vector registers (optional, only these ones are used by C2):
226 //
227 RegisterSaver_LiveVecReg( VR0 ),
228 RegisterSaver_LiveVecReg( VR1 ),
229 RegisterSaver_LiveVecReg( VR2 ),
230 RegisterSaver_LiveVecReg( VR3 ),
231 RegisterSaver_LiveVecReg( VR4 ),
232 RegisterSaver_LiveVecReg( VR5 ),
233 RegisterSaver_LiveVecReg( VR6 ),
234 RegisterSaver_LiveVecReg( VR7 ),
235 RegisterSaver_LiveVecReg( VR8 ),
236 RegisterSaver_LiveVecReg( VR9 ),
237 RegisterSaver_LiveVecReg( VR10 ),
238 RegisterSaver_LiveVecReg( VR11 ),
239 RegisterSaver_LiveVecReg( VR12 ),
240 RegisterSaver_LiveVecReg( VR13 ),
241 RegisterSaver_LiveVecReg( VR14 ),
242 RegisterSaver_LiveVecReg( VR15 ),
243 RegisterSaver_LiveVecReg( VR16 ),
244 RegisterSaver_LiveVecReg( VR17 ),
245 RegisterSaver_LiveVecReg( VR18 ),
246 RegisterSaver_LiveVecReg( VR19 ),
247 RegisterSaver_LiveVecReg( VR20 ),
248 RegisterSaver_LiveVecReg( VR21 ),
249 RegisterSaver_LiveVecReg( VR22 ),
250 RegisterSaver_LiveVecReg( VR23 ),
251 RegisterSaver_LiveVecReg( VR24 ),
252 RegisterSaver_LiveVecReg( VR25 ),
253 RegisterSaver_LiveVecReg( VR26 ),
254 RegisterSaver_LiveVecReg( VR27 ),
255 RegisterSaver_LiveVecReg( VR28 ),
256 RegisterSaver_LiveVecReg( VR29 ),
257 RegisterSaver_LiveVecReg( VR30 ),
258 RegisterSaver_LiveVecReg( VR31 )
259 };
260
261
262 OopMap* RegisterSaver::push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
263 int* out_frame_size_in_bytes,
264 bool generate_oop_map,
265 int return_pc_adjustment,
266 ReturnPCLocation return_pc_location,
267 bool save_vectors) {
268 // Push an abi_reg_args-frame and store all registers which may be live.
269 // If requested, create an OopMap: Record volatile registers as
270 // callee-save values in an OopMap so their save locations will be
271 // propagated to the RegisterMap of the caller frame during
272 // StackFrameStream construction (needed for deoptimization; see
273 // compiledVFrame::create_stack_value).
274 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment.
275 // Updated return pc is returned in R31 (if not return_pc_is_pre_saved).
276
277 // calculate frame size
278 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
279 sizeof(RegisterSaver::LiveRegType);
280 const int vecregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVecRegs) /
281 sizeof(RegisterSaver::LiveRegType))
282 : 0;
283 const int register_save_size = regstosave_num * reg_size + vecregstosave_num * vec_reg_size;
284 const int frame_size_in_bytes = align_up(register_save_size, frame::alignment_in_bytes)
285 + frame::native_abi_reg_args_size;
286
287 *out_frame_size_in_bytes = frame_size_in_bytes;
288 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
289 const int register_save_offset = frame_size_in_bytes - register_save_size;
290
291 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
292 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : nullptr;
293
294 BLOCK_COMMENT("push_frame_reg_args_and_save_live_registers {");
295
296 // push a new frame
297 __ push_frame(frame_size_in_bytes, noreg);
298
299 // Save some registers in the last (non-vector) slots of the new frame so we
300 // can use them as scratch regs or to determine the return pc.
301 __ std(R31, frame_size_in_bytes - reg_size - vecregstosave_num * vec_reg_size, R1_SP);
302 __ std(R30, frame_size_in_bytes - 2*reg_size - vecregstosave_num * vec_reg_size, R1_SP);
303
304 // save the flags
305 // Do the save_LR by hand and adjust the return pc if requested.
306 switch (return_pc_location) {
307 case return_pc_is_lr: __ mflr(R31); break;
308 case return_pc_is_pre_saved: assert(return_pc_adjustment == 0, "unsupported"); break;
309 case return_pc_is_thread_saved_exception_pc: __ ld(R31, thread_(saved_exception_pc)); break;
310 default: ShouldNotReachHere();
311 }
312 if (return_pc_location != return_pc_is_pre_saved) {
313 if (return_pc_adjustment != 0) {
314 __ addi(R31, R31, return_pc_adjustment);
315 }
316 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
317 }
318
319 // save all registers (ints and floats)
320 int offset = register_save_offset;
321
322 for (int i = 0; i < regstosave_num; i++) {
323 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
324 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
325
326 switch (reg_type) {
327 case RegisterSaver::int_reg: {
328 if (reg_num < 30) { // We spilled R30-31 right at the beginning.
329 __ std(as_Register(reg_num), offset, R1_SP);
330 }
331 break;
332 }
333 case RegisterSaver::float_reg: {
334 __ stfd(as_FloatRegister(reg_num), offset, R1_SP);
335 break;
336 }
337 case RegisterSaver::special_reg: {
338 if (reg_num == SR_CTR.encoding()) {
339 __ mfctr(R30);
340 __ std(R30, offset, R1_SP);
341 } else {
342 Unimplemented();
343 }
344 break;
345 }
346 default:
347 ShouldNotReachHere();
348 }
349
350 if (generate_oop_map) {
351 map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2),
352 RegisterSaver_LiveRegs[i].vmreg);
353 }
354 offset += reg_size;
355 }
356
357 // Note that generate_oop_map in the following loop is only used for the
358 // polling_page_vectors_safepoint_handler_blob.
359 // The order in which the vector contents are stored depends on Endianess and
360 // the utilized instructions (PowerArchitecturePPC64).
361 assert(is_aligned(offset, StackAlignmentInBytes), "should be");
362 if (PowerArchitecturePPC64 >= 10) {
363 assert(is_even(vecregstosave_num), "expectation");
364 for (int i = 0; i < vecregstosave_num; i += 2) {
365 int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
366 assert(RegisterSaver_LiveVecRegs[i + 1].reg_num == reg_num + 1, "or use other instructions!");
367
368 __ stxvp(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
369 // Note: The contents were read in the same order (see loadV16_Power9 node in ppc.ad).
370 if (generate_oop_map) {
371 map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2),
372 RegisterSaver_LiveVecRegs[i LITTLE_ENDIAN_ONLY(+1) ].vmreg);
373 map->set_callee_saved(VMRegImpl::stack2reg((offset + vec_reg_size) >> 2),
374 RegisterSaver_LiveVecRegs[i BIG_ENDIAN_ONLY(+1) ].vmreg);
375 }
376 offset += (2 * vec_reg_size);
377 }
378 } else {
379 for (int i = 0; i < vecregstosave_num; i++) {
380 int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
381
382 if (PowerArchitecturePPC64 >= 9) {
383 __ stxv(as_VectorRegister(reg_num)->to_vsr(), offset, R1_SP);
384 } else {
385 __ li(R31, offset);
386 __ stxvd2x(as_VectorRegister(reg_num)->to_vsr(), R31, R1_SP);
387 }
388 // Note: The contents were read in the same order (see loadV16_Power8 / loadV16_Power9 node in ppc.ad).
389 if (generate_oop_map) {
390 VMReg vsr = RegisterSaver_LiveVecRegs[i].vmreg;
391 map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2), vsr);
392 }
393 offset += vec_reg_size;
394 }
395 }
396
397 assert(offset == frame_size_in_bytes, "consistency check");
398
399 BLOCK_COMMENT("} push_frame_reg_args_and_save_live_registers");
400
401 // And we're done.
402 return map;
403 }
404
405
406 // Pop the current frame and restore all the registers that we
407 // saved.
408 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm,
409 int frame_size_in_bytes,
410 bool restore_ctr,
411 bool save_vectors) {
412 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
413 sizeof(RegisterSaver::LiveRegType);
414 const int vecregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVecRegs) /
415 sizeof(RegisterSaver::LiveRegType))
416 : 0;
417 const int register_save_size = regstosave_num * reg_size + vecregstosave_num * vec_reg_size;
418
419 const int register_save_offset = frame_size_in_bytes - register_save_size;
420
421 BLOCK_COMMENT("restore_live_registers_and_pop_frame {");
422
423 // restore all registers (ints and floats)
424 int offset = register_save_offset;
425
426 for (int i = 0; i < regstosave_num; i++) {
427 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
428 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
429
430 switch (reg_type) {
431 case RegisterSaver::int_reg: {
432 if (reg_num != 31) // R31 restored at the end, it's the tmp reg!
433 __ ld(as_Register(reg_num), offset, R1_SP);
434 break;
435 }
436 case RegisterSaver::float_reg: {
437 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
438 break;
439 }
440 case RegisterSaver::special_reg: {
441 if (reg_num == SR_CTR.encoding()) {
442 if (restore_ctr) { // Nothing to do here if ctr already contains the next address.
443 __ ld(R31, offset, R1_SP);
444 __ mtctr(R31);
445 }
446 } else {
447 Unimplemented();
448 }
449 break;
450 }
451 default:
452 ShouldNotReachHere();
453 }
454 offset += reg_size;
455 }
456
457 assert(is_aligned(offset, StackAlignmentInBytes), "should be");
458 if (PowerArchitecturePPC64 >= 10) {
459 for (int i = 0; i < vecregstosave_num; i += 2) {
460 int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
461 assert(RegisterSaver_LiveVecRegs[i + 1].reg_num == reg_num + 1, "or use other instructions!");
462
463 __ lxvp(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
464
465 offset += (2 * vec_reg_size);
466 }
467 } else {
468 for (int i = 0; i < vecregstosave_num; i++) {
469 int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
470
471 if (PowerArchitecturePPC64 >= 9) {
472 __ lxv(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
473 } else {
474 __ li(R31, offset);
475 __ lxvd2x(as_VectorRegister(reg_num).to_vsr(), R31, R1_SP);
476 }
477
478 offset += vec_reg_size;
479 }
480 }
481
482 assert(offset == frame_size_in_bytes, "consistency check");
483
484 // restore link and the flags
485 __ ld(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
486 __ mtlr(R31);
487
488 // restore scratch register's value
489 __ ld(R31, frame_size_in_bytes - reg_size - vecregstosave_num * vec_reg_size, R1_SP);
490
491 // pop the frame
492 __ addi(R1_SP, R1_SP, frame_size_in_bytes);
493
494 BLOCK_COMMENT("} restore_live_registers_and_pop_frame");
495 }
496
497 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp,
498 int frame_size,int total_args, const VMRegPair *regs,
499 const VMRegPair *regs2) {
500 __ push_frame(frame_size, r_temp);
501 int st_off = frame_size - wordSize;
502 for (int i = 0; i < total_args; i++) {
503 VMReg r_1 = regs[i].first();
504 VMReg r_2 = regs[i].second();
505 if (!r_1->is_valid()) {
506 assert(!r_2->is_valid(), "");
507 continue;
508 }
509 if (r_1->is_Register()) {
510 Register r = r_1->as_Register();
511 __ std(r, st_off, R1_SP);
512 st_off -= wordSize;
513 } else if (r_1->is_FloatRegister()) {
514 FloatRegister f = r_1->as_FloatRegister();
515 __ stfd(f, st_off, R1_SP);
516 st_off -= wordSize;
517 }
518 }
519 if (regs2 != nullptr) {
520 for (int i = 0; i < total_args; i++) {
521 VMReg r_1 = regs2[i].first();
522 VMReg r_2 = regs2[i].second();
523 if (!r_1->is_valid()) {
524 assert(!r_2->is_valid(), "");
525 continue;
526 }
527 if (r_1->is_Register()) {
528 Register r = r_1->as_Register();
529 __ std(r, st_off, R1_SP);
530 st_off -= wordSize;
531 } else if (r_1->is_FloatRegister()) {
532 FloatRegister f = r_1->as_FloatRegister();
533 __ stfd(f, st_off, R1_SP);
534 st_off -= wordSize;
535 }
536 }
537 }
538 }
539
540 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size,
541 int total_args, const VMRegPair *regs,
542 const VMRegPair *regs2) {
543 int st_off = frame_size - wordSize;
544 for (int i = 0; i < total_args; i++) {
545 VMReg r_1 = regs[i].first();
546 VMReg r_2 = regs[i].second();
547 if (r_1->is_Register()) {
548 Register r = r_1->as_Register();
549 __ ld(r, st_off, R1_SP);
550 st_off -= wordSize;
551 } else if (r_1->is_FloatRegister()) {
552 FloatRegister f = r_1->as_FloatRegister();
553 __ lfd(f, st_off, R1_SP);
554 st_off -= wordSize;
555 }
556 }
557 if (regs2 != nullptr)
558 for (int i = 0; i < total_args; i++) {
559 VMReg r_1 = regs2[i].first();
560 VMReg r_2 = regs2[i].second();
561 if (r_1->is_Register()) {
562 Register r = r_1->as_Register();
563 __ ld(r, st_off, R1_SP);
564 st_off -= wordSize;
565 } else if (r_1->is_FloatRegister()) {
566 FloatRegister f = r_1->as_FloatRegister();
567 __ lfd(f, st_off, R1_SP);
568 st_off -= wordSize;
569 }
570 }
571 __ pop_frame();
572 }
573
574 // Restore the registers that might be holding a result.
575 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) {
576 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
577 sizeof(RegisterSaver::LiveRegType);
578 const int register_save_size = regstosave_num * reg_size; // VS registers not relevant here.
579 const int register_save_offset = frame_size_in_bytes - register_save_size;
580
581 // restore all result registers (ints and floats)
582 int offset = register_save_offset;
583 for (int i = 0; i < regstosave_num; i++) {
584 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
585 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
586 switch (reg_type) {
587 case RegisterSaver::int_reg: {
588 if (as_Register(reg_num)==R3_RET) // int result_reg
589 __ ld(as_Register(reg_num), offset, R1_SP);
590 break;
591 }
592 case RegisterSaver::float_reg: {
593 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg
594 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
595 break;
596 }
597 case RegisterSaver::special_reg: {
598 // Special registers don't hold a result.
599 break;
600 }
601 default:
602 ShouldNotReachHere();
603 }
604 offset += reg_size;
605 }
606
607 assert(offset == frame_size_in_bytes, "consistency check");
608 }
609
610 // Is vector's size (in bytes) bigger than a size saved by default?
611 bool SharedRuntime::is_wide_vector(int size) {
612 // Note, MaxVectorSize == 8/16 on PPC64.
613 assert(size <= (SuperwordUseVSX ? 16 : 8), "%d bytes vectors are not supported", size);
614 return size > 8;
615 }
616
617 static int reg2slot(VMReg r) {
618 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots();
619 }
620
621 static int reg2offset(VMReg r) {
622 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
623 }
624
625 // ---------------------------------------------------------------------------
626 // Read the array of BasicTypes from a signature, and compute where the
627 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
628 // quantities. Values less than VMRegImpl::stack0 are registers, those above
629 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
630 // as framesizes are fixed.
631 // VMRegImpl::stack0 refers to the first slot 0(sp).
632 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register
633 // up to Register::number_of_registers) are the 64-bit
634 // integer registers.
635
636 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
637 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
638 // units regardless of build. Of course for i486 there is no 64 bit build
639
640 // The Java calling convention is a "shifted" version of the C ABI.
641 // By skipping the first C ABI register we can call non-static jni methods
642 // with small numbers of arguments without having to shuffle the arguments
643 // at all. Since we control the java ABI we ought to at least get some
644 // advantage out of it.
645
646 const VMReg java_iarg_reg[8] = {
647 R3->as_VMReg(),
648 R4->as_VMReg(),
649 R5->as_VMReg(),
650 R6->as_VMReg(),
651 R7->as_VMReg(),
652 R8->as_VMReg(),
653 R9->as_VMReg(),
654 R10->as_VMReg()
655 };
656
657 const VMReg java_farg_reg[13] = {
658 F1->as_VMReg(),
659 F2->as_VMReg(),
660 F3->as_VMReg(),
661 F4->as_VMReg(),
662 F5->as_VMReg(),
663 F6->as_VMReg(),
664 F7->as_VMReg(),
665 F8->as_VMReg(),
666 F9->as_VMReg(),
667 F10->as_VMReg(),
668 F11->as_VMReg(),
669 F12->as_VMReg(),
670 F13->as_VMReg()
671 };
672
673 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]);
674 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]);
675
676 STATIC_ASSERT(num_java_iarg_registers == Argument::n_int_register_parameters_j);
677 STATIC_ASSERT(num_java_farg_registers == Argument::n_float_register_parameters_j);
678
679 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
680 VMRegPair *regs,
681 int total_args_passed) {
682 // C2c calling conventions for compiled-compiled calls.
683 // Put 8 ints/longs into registers _AND_ 13 float/doubles into
684 // registers _AND_ put the rest on the stack.
685
686 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats
687 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles
688
689 int i;
690 VMReg reg;
691 int stk = 0;
692 int ireg = 0;
693 int freg = 0;
694
695 // We put the first 8 arguments into registers and the rest on the
696 // stack, float arguments are already in their argument registers
697 // due to c2c calling conventions (see calling_convention).
698 for (int i = 0; i < total_args_passed; ++i) {
699 switch(sig_bt[i]) {
700 case T_BOOLEAN:
701 case T_CHAR:
702 case T_BYTE:
703 case T_SHORT:
704 case T_INT:
705 if (ireg < num_java_iarg_registers) {
706 // Put int/ptr in register
707 reg = java_iarg_reg[ireg];
708 ++ireg;
709 } else {
710 // Put int/ptr on stack.
711 reg = VMRegImpl::stack2reg(stk);
712 stk += inc_stk_for_intfloat;
713 }
714 regs[i].set1(reg);
715 break;
716 case T_LONG:
717 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
718 if (ireg < num_java_iarg_registers) {
719 // Put long in register.
720 reg = java_iarg_reg[ireg];
721 ++ireg;
722 } else {
723 // Put long on stack. They must be aligned to 2 slots.
724 if (stk & 0x1) ++stk;
725 reg = VMRegImpl::stack2reg(stk);
726 stk += inc_stk_for_longdouble;
727 }
728 regs[i].set2(reg);
729 break;
730 case T_OBJECT:
731 case T_ARRAY:
732 case T_ADDRESS:
733 if (ireg < num_java_iarg_registers) {
734 // Put ptr in register.
735 reg = java_iarg_reg[ireg];
736 ++ireg;
737 } else {
738 // Put ptr on stack. Objects must be aligned to 2 slots too,
739 // because "64-bit pointers record oop-ishness on 2 aligned
740 // adjacent registers." (see OopFlow::build_oop_map).
741 if (stk & 0x1) ++stk;
742 reg = VMRegImpl::stack2reg(stk);
743 stk += inc_stk_for_longdouble;
744 }
745 regs[i].set2(reg);
746 break;
747 case T_FLOAT:
748 if (freg < num_java_farg_registers) {
749 // Put float in register.
750 reg = java_farg_reg[freg];
751 ++freg;
752 } else {
753 // Put float on stack.
754 reg = VMRegImpl::stack2reg(stk);
755 stk += inc_stk_for_intfloat;
756 }
757 regs[i].set1(reg);
758 break;
759 case T_DOUBLE:
760 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
761 if (freg < num_java_farg_registers) {
762 // Put double in register.
763 reg = java_farg_reg[freg];
764 ++freg;
765 } else {
766 // Put double on stack. They must be aligned to 2 slots.
767 if (stk & 0x1) ++stk;
768 reg = VMRegImpl::stack2reg(stk);
769 stk += inc_stk_for_longdouble;
770 }
771 regs[i].set2(reg);
772 break;
773 case T_VOID:
774 // Do not count halves.
775 regs[i].set_bad();
776 break;
777 default:
778 ShouldNotReachHere();
779 }
780 }
781 return stk;
782 }
783
784 #if defined(COMPILER1) || defined(COMPILER2)
785 // Calling convention for calling C code.
786 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
787 VMRegPair *regs,
788 int total_args_passed) {
789 // Calling conventions for C runtime calls and calls to JNI native methods.
790 //
791 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8
792 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist
793 // the first 13 flt/dbl's in the first 13 fp regs but additionally
794 // copy flt/dbl to the stack if they are beyond the 8th argument.
795
796 const VMReg iarg_reg[8] = {
797 R3->as_VMReg(),
798 R4->as_VMReg(),
799 R5->as_VMReg(),
800 R6->as_VMReg(),
801 R7->as_VMReg(),
802 R8->as_VMReg(),
803 R9->as_VMReg(),
804 R10->as_VMReg()
805 };
806
807 const VMReg farg_reg[13] = {
808 F1->as_VMReg(),
809 F2->as_VMReg(),
810 F3->as_VMReg(),
811 F4->as_VMReg(),
812 F5->as_VMReg(),
813 F6->as_VMReg(),
814 F7->as_VMReg(),
815 F8->as_VMReg(),
816 F9->as_VMReg(),
817 F10->as_VMReg(),
818 F11->as_VMReg(),
819 F12->as_VMReg(),
820 F13->as_VMReg()
821 };
822
823 // Check calling conventions consistency.
824 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c &&
825 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c,
826 "consistency");
827
828 const int additional_frame_header_slots = ((frame::native_abi_minframe_size - frame::jit_out_preserve_size)
829 / VMRegImpl::stack_slot_size);
830 const int float_offset_in_slots = Argument::float_on_stack_offset_in_bytes_c / VMRegImpl::stack_slot_size;
831
832 VMReg reg;
833 int arg = 0;
834 int freg = 0;
835 bool stack_used = false;
836
837 for (int i = 0; i < total_args_passed; ++i, ++arg) {
838 // Each argument corresponds to a slot in the Parameter Save Area (if not omitted)
839 int stk = (arg * 2) + additional_frame_header_slots;
840
841 switch(sig_bt[i]) {
842 //
843 // If arguments 0-7 are integers, they are passed in integer registers.
844 // Argument i is placed in iarg_reg[i].
845 //
846 case T_BOOLEAN:
847 case T_CHAR:
848 case T_BYTE:
849 case T_SHORT:
850 case T_INT:
851 // We must cast ints to longs and use full 64 bit stack slots
852 // here. Thus fall through, handle as long.
853 case T_LONG:
854 case T_OBJECT:
855 case T_ARRAY:
856 case T_ADDRESS:
857 case T_METADATA:
858 // Oops are already boxed if required (JNI).
859 if (arg < Argument::n_int_register_parameters_c) {
860 reg = iarg_reg[arg];
861 } else {
862 reg = VMRegImpl::stack2reg(stk);
863 stack_used = true;
864 }
865 regs[i].set2(reg);
866 break;
867
868 //
869 // Floats are treated differently from int regs: The first 13 float arguments
870 // are passed in registers (not the float args among the first 13 args).
871 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed
872 // in farg_reg[j] if argument i is the j-th float argument of this call.
873 //
874 case T_FLOAT:
875 if (freg < Argument::n_float_register_parameters_c) {
876 // Put float in register ...
877 reg = farg_reg[freg];
878 ++freg;
879 } else {
880 // Put float on stack.
881 reg = VMRegImpl::stack2reg(stk + float_offset_in_slots);
882 stack_used = true;
883 }
884 regs[i].set1(reg);
885 break;
886 case T_DOUBLE:
887 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
888 if (freg < Argument::n_float_register_parameters_c) {
889 // Put double in register ...
890 reg = farg_reg[freg];
891 ++freg;
892 } else {
893 // Put double on stack.
894 reg = VMRegImpl::stack2reg(stk);
895 stack_used = true;
896 }
897 regs[i].set2(reg);
898 break;
899
900 case T_VOID:
901 // Do not count halves.
902 regs[i].set_bad();
903 --arg;
904 break;
905 default:
906 ShouldNotReachHere();
907 }
908 }
909
910 // Return size of the stack frame excluding the jit_out_preserve part in single-word slots.
911 #if defined(ABI_ELFv2)
912 assert(additional_frame_header_slots == 0, "ABIv2 shouldn't use extra slots");
913 // ABIv2 allows omitting the Parameter Save Area if the callee's prototype
914 // indicates that all parameters can be passed in registers.
915 return stack_used ? (arg * 2) : 0;
916 #else
917 // The Parameter Save Area needs to be at least 8 double-word slots for ABIv1.
918 // We have to add extra slots because ABIv1 uses a larger header.
919 return MAX2(arg, 8) * 2 + additional_frame_header_slots;
920 #endif
921 }
922 #endif // COMPILER2
923
924 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
925 uint num_bits,
926 uint total_args_passed) {
927 Unimplemented();
928 return 0;
929 }
930
931 static address gen_c2i_adapter(MacroAssembler *masm,
932 int total_args_passed,
933 int comp_args_on_stack,
934 const BasicType *sig_bt,
935 const VMRegPair *regs,
936 Label& call_interpreter,
937 const Register& ientry) {
938
939 address c2i_entrypoint;
940
941 const Register sender_SP = R21_sender_SP; // == R21_tmp1
942 const Register code = R22_tmp2;
943 //const Register ientry = R23_tmp3;
944 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 };
945 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
946 int value_regs_index = 0;
947
948 const Register return_pc = R27_tmp7;
949 const Register tmp = R28_tmp8;
950
951 assert_different_registers(sender_SP, code, ientry, return_pc, tmp);
952
953 // Adapter needs TOP_IJAVA_FRAME_ABI.
954 const int adapter_size = frame::top_ijava_frame_abi_size +
955 align_up(total_args_passed * wordSize, frame::alignment_in_bytes);
956
957 // regular (verified) c2i entry point
958 c2i_entrypoint = __ pc();
959
960 // Does compiled code exists? If yes, patch the caller's callsite.
961 __ ld(code, method_(code));
962 __ cmpdi(CR0, code, 0);
963 __ ld(ientry, method_(interpreter_entry)); // preloaded
964 __ beq(CR0, call_interpreter);
965
966
967 // Patch caller's callsite, method_(code) was not null which means that
968 // compiled code exists.
969 __ mflr(return_pc);
970 __ std(return_pc, _abi0(lr), R1_SP);
971 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs);
972
973 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc);
974
975 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs);
976 __ ld(return_pc, _abi0(lr), R1_SP);
977 __ ld(ientry, method_(interpreter_entry)); // preloaded
978 __ mtlr(return_pc);
979
980
981 // Call the interpreter.
982 __ BIND(call_interpreter);
983 __ mtctr(ientry);
984
985 // Get a copy of the current SP for loading caller's arguments.
986 __ mr(sender_SP, R1_SP);
987
988 // Add space for the adapter.
989 __ resize_frame(-adapter_size, R12_scratch2);
990
991 int st_off = adapter_size - wordSize;
992
993 // Write the args into the outgoing interpreter space.
994 for (int i = 0; i < total_args_passed; i++) {
995 VMReg r_1 = regs[i].first();
996 VMReg r_2 = regs[i].second();
997 if (!r_1->is_valid()) {
998 assert(!r_2->is_valid(), "");
999 continue;
1000 }
1001 if (r_1->is_stack()) {
1002 Register tmp_reg = value_regs[value_regs_index];
1003 value_regs_index = (value_regs_index + 1) % num_value_regs;
1004 // The calling convention produces OptoRegs that ignore the out
1005 // preserve area (JIT's ABI). We must account for it here.
1006 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
1007 if (!r_2->is_valid()) {
1008 __ lwz(tmp_reg, ld_off, sender_SP);
1009 } else {
1010 __ ld(tmp_reg, ld_off, sender_SP);
1011 }
1012 // Pretend stack targets were loaded into tmp_reg.
1013 r_1 = tmp_reg->as_VMReg();
1014 }
1015
1016 if (r_1->is_Register()) {
1017 Register r = r_1->as_Register();
1018 if (!r_2->is_valid()) {
1019 __ stw(r, st_off, R1_SP);
1020 st_off-=wordSize;
1021 } else {
1022 // Longs are given 2 64-bit slots in the interpreter, but the
1023 // data is passed in only 1 slot.
1024 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
1025 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
1026 st_off-=wordSize;
1027 }
1028 __ std(r, st_off, R1_SP);
1029 st_off-=wordSize;
1030 }
1031 } else {
1032 assert(r_1->is_FloatRegister(), "");
1033 FloatRegister f = r_1->as_FloatRegister();
1034 if (!r_2->is_valid()) {
1035 __ stfs(f, st_off, R1_SP);
1036 st_off-=wordSize;
1037 } else {
1038 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the
1039 // data is passed in only 1 slot.
1040 // One of these should get known junk...
1041 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
1042 st_off-=wordSize;
1043 __ stfd(f, st_off, R1_SP);
1044 st_off-=wordSize;
1045 }
1046 }
1047 }
1048
1049 // Jump to the interpreter just as if interpreter was doing it.
1050
1051 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
1052
1053 // load TOS
1054 __ addi(R15_esp, R1_SP, st_off);
1055
1056 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1.
1057 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register");
1058 __ bctr();
1059
1060 return c2i_entrypoint;
1061 }
1062
1063 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
1064 int total_args_passed,
1065 int comp_args_on_stack,
1066 const BasicType *sig_bt,
1067 const VMRegPair *regs) {
1068
1069 // Load method's entry-point from method.
1070 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
1071 __ mtctr(R12_scratch2);
1072
1073 // We will only enter here from an interpreted frame and never from after
1074 // passing thru a c2i. Azul allowed this but we do not. If we lose the
1075 // race and use a c2i we will remain interpreted for the race loser(s).
1076 // This removes all sorts of headaches on the x86 side and also eliminates
1077 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
1078
1079 // Note: r13 contains the senderSP on entry. We must preserve it since
1080 // we may do a i2c -> c2i transition if we lose a race where compiled
1081 // code goes non-entrant while we get args ready.
1082 // In addition we use r13 to locate all the interpreter args as
1083 // we must align the stack to 16 bytes on an i2c entry else we
1084 // lose alignment we expect in all compiled code and register
1085 // save code can segv when fxsave instructions find improperly
1086 // aligned stack pointer.
1087
1088 const Register ld_ptr = R15_esp;
1089 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 };
1090 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
1091 int value_regs_index = 0;
1092
1093 int ld_offset = total_args_passed*wordSize;
1094
1095 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
1096 // in registers, we will occasionally have no stack args.
1097 int comp_words_on_stack = 0;
1098 if (comp_args_on_stack) {
1099 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
1100 // registers are below. By subtracting stack0, we either get a negative
1101 // number (all values in registers) or the maximum stack slot accessed.
1102
1103 // Convert 4-byte c2 stack slots to words.
1104 comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1105 // Round up to miminum stack alignment, in wordSize.
1106 comp_words_on_stack = align_up(comp_words_on_stack, 2);
1107 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1);
1108 }
1109
1110 // Now generate the shuffle code. Pick up all register args and move the
1111 // rest through register value=Z_R12.
1112 BLOCK_COMMENT("Shuffle arguments");
1113 for (int i = 0; i < total_args_passed; i++) {
1114 if (sig_bt[i] == T_VOID) {
1115 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
1116 continue;
1117 }
1118
1119 // Pick up 0, 1 or 2 words from ld_ptr.
1120 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1121 "scrambled load targets?");
1122 VMReg r_1 = regs[i].first();
1123 VMReg r_2 = regs[i].second();
1124 if (!r_1->is_valid()) {
1125 assert(!r_2->is_valid(), "");
1126 continue;
1127 }
1128 if (r_1->is_FloatRegister()) {
1129 if (!r_2->is_valid()) {
1130 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr);
1131 ld_offset-=wordSize;
1132 } else {
1133 // Skip the unused interpreter slot.
1134 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr);
1135 ld_offset-=2*wordSize;
1136 }
1137 } else {
1138 Register r;
1139 if (r_1->is_stack()) {
1140 // Must do a memory to memory move thru "value".
1141 r = value_regs[value_regs_index];
1142 value_regs_index = (value_regs_index + 1) % num_value_regs;
1143 } else {
1144 r = r_1->as_Register();
1145 }
1146 if (!r_2->is_valid()) {
1147 // Not sure we need to do this but it shouldn't hurt.
1148 if (is_reference_type(sig_bt[i]) || sig_bt[i] == T_ADDRESS) {
1149 __ ld(r, ld_offset, ld_ptr);
1150 ld_offset-=wordSize;
1151 } else {
1152 __ lwz(r, ld_offset, ld_ptr);
1153 ld_offset-=wordSize;
1154 }
1155 } else {
1156 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
1157 // data is passed in only 1 slot.
1158 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
1159 ld_offset-=wordSize;
1160 }
1161 __ ld(r, ld_offset, ld_ptr);
1162 ld_offset-=wordSize;
1163 }
1164
1165 if (r_1->is_stack()) {
1166 // Now store value where the compiler expects it
1167 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size;
1168
1169 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN ||
1170 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) {
1171 __ stw(r, st_off, R1_SP);
1172 } else {
1173 __ std(r, st_off, R1_SP);
1174 }
1175 }
1176 }
1177 }
1178
1179 __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1180
1181 BLOCK_COMMENT("Store method");
1182 // Store method into thread->callee_target.
1183 // We might end up in handle_wrong_method if the callee is
1184 // deoptimized as we race thru here. If that happens we don't want
1185 // to take a safepoint because the caller frame will look
1186 // interpreted and arguments are now "compiled" so it is much better
1187 // to make this transition invisible to the stack walking
1188 // code. Unfortunately if we try and find the callee by normal means
1189 // a safepoint is possible. So we stash the desired callee in the
1190 // thread and the vm will find there should this case occur.
1191 __ std(R19_method, thread_(callee_target));
1192
1193 // Jump to the compiled code just as if compiled code was doing it.
1194 __ bctr();
1195 }
1196
1197 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1198 int total_args_passed,
1199 int comp_args_on_stack,
1200 const BasicType *sig_bt,
1201 const VMRegPair *regs,
1202 address entry_address[AdapterBlob::ENTRY_COUNT]) {
1203 // entry: i2c
1204
1205 __ align(CodeEntryAlignment);
1206 entry_address[AdapterBlob::I2C] = __ pc();
1207 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
1208
1209
1210 // entry: c2i unverified
1211
1212 __ align(CodeEntryAlignment);
1213 BLOCK_COMMENT("c2i unverified entry");
1214 entry_address[AdapterBlob::C2I_Unverified] = __ pc();
1215
1216 // inline_cache contains a CompiledICData
1217 const Register ic = R19_inline_cache_reg;
1218 const Register ic_klass = R11_scratch1;
1219 const Register receiver_klass = R12_scratch2;
1220 const Register code = R21_tmp1;
1221 const Register ientry = R23_tmp3;
1222
1223 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry);
1224 assert(R11_scratch1 == R11, "need prologue scratch register");
1225
1226 Label call_interpreter;
1227
1228 __ ic_check(4 /* end_alignment */);
1229 __ ld(R19_method, CompiledICData::speculated_method_offset(), ic);
1230 // Argument is valid and klass is as expected, continue.
1231
1232 __ ld(code, method_(code));
1233 __ cmpdi(CR0, code, 0);
1234 __ ld(ientry, method_(interpreter_entry)); // preloaded
1235 __ beq_predict_taken(CR0, call_interpreter);
1236
1237 // Branch to ic_miss_stub.
1238 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type);
1239
1240 // entry: c2i
1241
1242 entry_address[AdapterBlob::C2I] = __ pc();
1243
1244 // Class initialization barrier for static methods
1245 entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
1246 if (VM_Version::supports_fast_class_init_checks()) {
1247 Label L_skip_barrier;
1248
1249 { // Bypass the barrier for non-static methods
1250 __ lhz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1251 __ andi_(R0, R0, JVM_ACC_STATIC);
1252 __ beq(CR0, L_skip_barrier); // non-static
1253 }
1254
1255 Register klass = R11_scratch1;
1256 __ load_method_holder(klass, R19_method);
1257 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
1258
1259 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
1260 __ mtctr(klass);
1261 __ bctr();
1262
1263 __ bind(L_skip_barrier);
1264 entry_address[AdapterBlob::C2I_No_Clinit_Check] = __ pc();
1265 }
1266
1267 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1268 bs->c2i_entry_barrier(masm, /* tmp register*/ ic_klass, /* tmp register*/ receiver_klass, /* tmp register*/ code);
1269
1270 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry);
1271 return;
1272 }
1273
1274 // An oop arg. Must pass a handle not the oop itself.
1275 static void object_move(MacroAssembler* masm,
1276 int frame_size_in_slots,
1277 OopMap* oop_map, int oop_handle_offset,
1278 bool is_receiver, int* receiver_offset,
1279 VMRegPair src, VMRegPair dst,
1280 Register r_caller_sp, Register r_temp_1, Register r_temp_2) {
1281 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)),
1282 "receiver has already been moved");
1283
1284 // We must pass a handle. First figure out the location we use as a handle.
1285
1286 if (src.first()->is_stack()) {
1287 // stack to stack or reg
1288
1289 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1290 Label skip;
1291 const int oop_slot_in_callers_frame = reg2slot(src.first());
1292
1293 guarantee(!is_receiver, "expecting receiver in register");
1294 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots));
1295
1296 __ addi(r_handle, r_caller_sp, reg2offset(src.first()));
1297 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp);
1298 __ cmpdi(CR0, r_temp_2, 0);
1299 __ bne(CR0, skip);
1300 // Use a null handle if oop is null.
1301 __ li(r_handle, 0);
1302 __ bind(skip);
1303
1304 if (dst.first()->is_stack()) {
1305 // stack to stack
1306 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1307 } else {
1308 // stack to reg
1309 // Nothing to do, r_handle is already the dst register.
1310 }
1311 } else {
1312 // reg to stack or reg
1313 const Register r_oop = src.first()->as_Register();
1314 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1315 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word
1316 + oop_handle_offset; // in slots
1317 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size;
1318 Label skip;
1319
1320 if (is_receiver) {
1321 *receiver_offset = oop_offset;
1322 }
1323 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot));
1324
1325 __ std( r_oop, oop_offset, R1_SP);
1326 __ addi(r_handle, R1_SP, oop_offset);
1327
1328 __ cmpdi(CR0, r_oop, 0);
1329 __ bne(CR0, skip);
1330 // Use a null handle if oop is null.
1331 __ li(r_handle, 0);
1332 __ bind(skip);
1333
1334 if (dst.first()->is_stack()) {
1335 // reg to stack
1336 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1337 } else {
1338 // reg to reg
1339 // Nothing to do, r_handle is already the dst register.
1340 }
1341 }
1342 }
1343
1344 static void int_move(MacroAssembler*masm,
1345 VMRegPair src, VMRegPair dst,
1346 Register r_caller_sp, Register r_temp) {
1347 assert(src.first()->is_valid(), "incoming must be int");
1348 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1349
1350 if (src.first()->is_stack()) {
1351 if (dst.first()->is_stack()) {
1352 // stack to stack
1353 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp);
1354 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1355 } else {
1356 // stack to reg
1357 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1358 }
1359 } else if (dst.first()->is_stack()) {
1360 // reg to stack
1361 __ extsw(r_temp, src.first()->as_Register());
1362 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1363 } else {
1364 // reg to reg
1365 __ extsw(dst.first()->as_Register(), src.first()->as_Register());
1366 }
1367 }
1368
1369 static void long_move(MacroAssembler*masm,
1370 VMRegPair src, VMRegPair dst,
1371 Register r_caller_sp, Register r_temp) {
1372 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long");
1373 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1374
1375 if (src.first()->is_stack()) {
1376 if (dst.first()->is_stack()) {
1377 // stack to stack
1378 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1379 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1380 } else {
1381 // stack to reg
1382 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1383 }
1384 } else if (dst.first()->is_stack()) {
1385 // reg to stack
1386 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP);
1387 } else {
1388 // reg to reg
1389 if (dst.first()->as_Register() != src.first()->as_Register())
1390 __ mr(dst.first()->as_Register(), src.first()->as_Register());
1391 }
1392 }
1393
1394 static void float_move(MacroAssembler*masm,
1395 VMRegPair src, VMRegPair dst,
1396 Register r_caller_sp, Register r_temp) {
1397 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float");
1398 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float");
1399
1400 if (src.first()->is_stack()) {
1401 if (dst.first()->is_stack()) {
1402 // stack to stack
1403 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp);
1404 __ stw(r_temp, reg2offset(dst.first()), R1_SP);
1405 } else {
1406 // stack to reg
1407 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1408 }
1409 } else if (dst.first()->is_stack()) {
1410 // reg to stack
1411 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1412 } else {
1413 // reg to reg
1414 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1415 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1416 }
1417 }
1418
1419 static void double_move(MacroAssembler*masm,
1420 VMRegPair src, VMRegPair dst,
1421 Register r_caller_sp, Register r_temp) {
1422 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double");
1423 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double");
1424
1425 if (src.first()->is_stack()) {
1426 if (dst.first()->is_stack()) {
1427 // stack to stack
1428 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1429 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1430 } else {
1431 // stack to reg
1432 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1433 }
1434 } else if (dst.first()->is_stack()) {
1435 // reg to stack
1436 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1437 } else {
1438 // reg to reg
1439 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1440 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1441 }
1442 }
1443
1444 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1445 switch (ret_type) {
1446 case T_BOOLEAN:
1447 case T_CHAR:
1448 case T_BYTE:
1449 case T_SHORT:
1450 case T_INT:
1451 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1452 break;
1453 case T_ARRAY:
1454 case T_OBJECT:
1455 case T_LONG:
1456 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1457 break;
1458 case T_FLOAT:
1459 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1460 break;
1461 case T_DOUBLE:
1462 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1463 break;
1464 case T_VOID:
1465 break;
1466 default:
1467 ShouldNotReachHere();
1468 break;
1469 }
1470 }
1471
1472 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1473 switch (ret_type) {
1474 case T_BOOLEAN:
1475 case T_CHAR:
1476 case T_BYTE:
1477 case T_SHORT:
1478 case T_INT:
1479 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1480 break;
1481 case T_ARRAY:
1482 case T_OBJECT:
1483 case T_LONG:
1484 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1485 break;
1486 case T_FLOAT:
1487 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1488 break;
1489 case T_DOUBLE:
1490 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1491 break;
1492 case T_VOID:
1493 break;
1494 default:
1495 ShouldNotReachHere();
1496 break;
1497 }
1498 }
1499
1500 static void verify_oop_args(MacroAssembler* masm,
1501 const methodHandle& method,
1502 const BasicType* sig_bt,
1503 const VMRegPair* regs) {
1504 Register temp_reg = R19_method; // not part of any compiled calling seq
1505 if (VerifyOops) {
1506 for (int i = 0; i < method->size_of_parameters(); i++) {
1507 if (is_reference_type(sig_bt[i])) {
1508 VMReg r = regs[i].first();
1509 assert(r->is_valid(), "bad oop arg");
1510 if (r->is_stack()) {
1511 __ ld(temp_reg, reg2offset(r), R1_SP);
1512 __ verify_oop(temp_reg, FILE_AND_LINE);
1513 } else {
1514 __ verify_oop(r->as_Register(), FILE_AND_LINE);
1515 }
1516 }
1517 }
1518 }
1519 }
1520
1521 static void gen_special_dispatch(MacroAssembler* masm,
1522 const methodHandle& method,
1523 const BasicType* sig_bt,
1524 const VMRegPair* regs) {
1525 verify_oop_args(masm, method, sig_bt, regs);
1526 vmIntrinsics::ID iid = method->intrinsic_id();
1527
1528 // Now write the args into the outgoing interpreter space
1529 bool has_receiver = false;
1530 Register receiver_reg = noreg;
1531 int member_arg_pos = -1;
1532 Register member_reg = noreg;
1533 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1534 if (ref_kind != 0) {
1535 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1536 member_reg = R19_method; // known to be free at this point
1537 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1538 } else if (iid == vmIntrinsics::_invokeBasic) {
1539 has_receiver = true;
1540 } else if (iid == vmIntrinsics::_linkToNative) {
1541 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1542 member_reg = R19_method; // known to be free at this point
1543 } else {
1544 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1545 }
1546
1547 if (member_reg != noreg) {
1548 // Load the member_arg into register, if necessary.
1549 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1550 VMReg r = regs[member_arg_pos].first();
1551 if (r->is_stack()) {
1552 __ ld(member_reg, reg2offset(r), R1_SP);
1553 } else {
1554 // no data motion is needed
1555 member_reg = r->as_Register();
1556 }
1557 }
1558
1559 if (has_receiver) {
1560 // Make sure the receiver is loaded into a register.
1561 assert(method->size_of_parameters() > 0, "oob");
1562 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1563 VMReg r = regs[0].first();
1564 assert(r->is_valid(), "bad receiver arg");
1565 if (r->is_stack()) {
1566 // Porting note: This assumes that compiled calling conventions always
1567 // pass the receiver oop in a register. If this is not true on some
1568 // platform, pick a temp and load the receiver from stack.
1569 fatal("receiver always in a register");
1570 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point?
1571 __ ld(receiver_reg, reg2offset(r), R1_SP);
1572 } else {
1573 // no data motion is needed
1574 receiver_reg = r->as_Register();
1575 }
1576 }
1577
1578 // Figure out which address we are really jumping to:
1579 MethodHandles::generate_method_handle_dispatch(masm, iid,
1580 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1581 }
1582
1583 //---------------------------- continuation_enter_setup ---------------------------
1584 //
1585 // Frame setup.
1586 //
1587 // Arguments:
1588 // None.
1589 //
1590 // Results:
1591 // R1_SP: pointer to blank ContinuationEntry in the pushed frame.
1592 //
1593 // Kills:
1594 // R0, R20
1595 //
1596 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& framesize_words) {
1597 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1598 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
1599 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1600
1601 const int frame_size_in_bytes = (int)ContinuationEntry::size();
1602 assert(is_aligned(frame_size_in_bytes, frame::alignment_in_bytes), "alignment error");
1603
1604 framesize_words = frame_size_in_bytes / wordSize;
1605
1606 DEBUG_ONLY(__ block_comment("setup {"));
1607 // Save return pc and push entry frame
1608 const Register return_pc = R20;
1609 __ mflr(return_pc);
1610 __ std(return_pc, _abi0(lr), R1_SP); // SP->lr = return_pc
1611 __ push_frame(frame_size_in_bytes , R0); // SP -= frame_size_in_bytes
1612
1613 OopMap* map = new OopMap((int)frame_size_in_bytes / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1614
1615 __ ld_ptr(R0, JavaThread::cont_entry_offset(), R16_thread);
1616 __ st_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1617 __ st_ptr(R0, ContinuationEntry::parent_offset(), R1_SP);
1618 DEBUG_ONLY(__ block_comment("} setup"));
1619
1620 return map;
1621 }
1622
1623 //---------------------------- fill_continuation_entry ---------------------------
1624 //
1625 // Initialize the new ContinuationEntry.
1626 //
1627 // Arguments:
1628 // R1_SP: pointer to blank Continuation entry
1629 // reg_cont_obj: pointer to the continuation
1630 // reg_flags: flags
1631 //
1632 // Results:
1633 // R1_SP: pointer to filled out ContinuationEntry
1634 //
1635 // Kills:
1636 // R8_ARG6, R9_ARG7, R10_ARG8
1637 //
1638 static void fill_continuation_entry(MacroAssembler* masm, Register reg_cont_obj, Register reg_flags) {
1639 assert_different_registers(reg_cont_obj, reg_flags);
1640 Register zero = R8_ARG6;
1641 Register tmp2 = R9_ARG7;
1642 Register tmp3 = R10_ARG8;
1643
1644 DEBUG_ONLY(__ block_comment("fill {"));
1645 #ifdef ASSERT
1646 __ load_const_optimized(tmp2, ContinuationEntry::cookie_value());
1647 __ stw(tmp2, in_bytes(ContinuationEntry::cookie_offset()), R1_SP);
1648 #endif //ASSERT
1649
1650 __ li(zero, 0);
1651 __ st_ptr(reg_cont_obj, ContinuationEntry::cont_offset(), R1_SP);
1652 __ stw(reg_flags, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1653 __ st_ptr(zero, ContinuationEntry::chunk_offset(), R1_SP);
1654 __ stw(zero, in_bytes(ContinuationEntry::argsize_offset()), R1_SP);
1655 __ stw(zero, in_bytes(ContinuationEntry::pin_count_offset()), R1_SP);
1656
1657 __ ld_ptr(tmp2, JavaThread::cont_fastpath_offset(), R16_thread);
1658 __ ld(tmp3, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1659 __ st_ptr(tmp2, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1660 __ std(tmp3, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1661
1662 __ st_ptr(zero, JavaThread::cont_fastpath_offset(), R16_thread);
1663 __ std(zero, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1664 DEBUG_ONLY(__ block_comment("} fill"));
1665 }
1666
1667 //---------------------------- continuation_enter_cleanup ---------------------------
1668 //
1669 // Copy corresponding attributes from the top ContinuationEntry to the JavaThread
1670 // before deleting it.
1671 //
1672 // Arguments:
1673 // R1_SP: pointer to the ContinuationEntry
1674 //
1675 // Results:
1676 // None.
1677 //
1678 // Kills:
1679 // R8_ARG6, R9_ARG7, R10_ARG8, R15_esp
1680 //
1681 static void continuation_enter_cleanup(MacroAssembler* masm) {
1682 Register tmp1 = R8_ARG6;
1683 Register tmp2 = R9_ARG7;
1684 Register tmp3 = R10_ARG8;
1685
1686 #ifdef ASSERT
1687 __ block_comment("clean {");
1688 __ ld_ptr(tmp1, JavaThread::cont_entry_offset(), R16_thread);
1689 __ cmpd(CR0, R1_SP, tmp1);
1690 __ asm_assert_eq(FILE_AND_LINE ": incorrect R1_SP");
1691 #endif
1692
1693 __ ld_ptr(tmp1, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1694 __ st_ptr(tmp1, JavaThread::cont_fastpath_offset(), R16_thread);
1695
1696 if (CheckJNICalls) {
1697 // Check if this is a virtual thread continuation
1698 Label L_skip_vthread_code;
1699 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1700 __ cmpwi(CR0, R0, 0);
1701 __ beq(CR0, L_skip_vthread_code);
1702
1703 // If the held monitor count is > 0 and this vthread is terminating then
1704 // it failed to release a JNI monitor. So we issue the same log message
1705 // that JavaThread::exit does.
1706 __ ld(R0, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1707 __ cmpdi(CR0, R0, 0);
1708 __ beq(CR0, L_skip_vthread_code);
1709
1710 // Save return value potentially containing the exception oop
1711 Register ex_oop = R15_esp; // nonvolatile register
1712 __ mr(ex_oop, R3_RET);
1713 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
1714 // Restore potental return value
1715 __ mr(R3_RET, ex_oop);
1716
1717 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
1718 // on termination. The held count is implicitly zeroed below when we restore from
1719 // the parent held count (which has to be zero).
1720 __ li(tmp1, 0);
1721 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1722
1723 __ bind(L_skip_vthread_code);
1724 }
1725 #ifdef ASSERT
1726 else {
1727 // Check if this is a virtual thread continuation
1728 Label L_skip_vthread_code;
1729 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1730 __ cmpwi(CR0, R0, 0);
1731 __ beq(CR0, L_skip_vthread_code);
1732
1733 // See comment just above. If not checking JNI calls the JNI count is only
1734 // needed for assertion checking.
1735 __ li(tmp1, 0);
1736 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1737
1738 __ bind(L_skip_vthread_code);
1739 }
1740 #endif
1741
1742 __ ld(tmp2, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1743 __ ld_ptr(tmp3, ContinuationEntry::parent_offset(), R1_SP);
1744 __ std(tmp2, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1745 __ st_ptr(tmp3, JavaThread::cont_entry_offset(), R16_thread);
1746 DEBUG_ONLY(__ block_comment("} clean"));
1747 }
1748
1749 static void check_continuation_enter_argument(VMReg actual_vmreg,
1750 Register expected_reg,
1751 const char* name) {
1752 assert(!actual_vmreg->is_stack(), "%s cannot be on stack", name);
1753 assert(actual_vmreg->as_Register() == expected_reg,
1754 "%s is in unexpected register: %s instead of %s",
1755 name, actual_vmreg->as_Register()->name(), expected_reg->name());
1756 }
1757
1758 static void gen_continuation_enter(MacroAssembler* masm,
1759 const VMRegPair* regs,
1760 int& exception_offset,
1761 OopMapSet* oop_maps,
1762 int& frame_complete,
1763 int& framesize_words,
1764 int& interpreted_entry_offset,
1765 int& compiled_entry_offset) {
1766
1767 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1768 int pos_cont_obj = 0;
1769 int pos_is_cont = 1;
1770 int pos_is_virtual = 2;
1771
1772 // The platform-specific calling convention may present the arguments in various registers.
1773 // To simplify the rest of the code, we expect the arguments to reside at these known
1774 // registers, and we additionally check the placement here in case calling convention ever
1775 // changes.
1776 Register reg_cont_obj = R3_ARG1;
1777 Register reg_is_cont = R4_ARG2;
1778 Register reg_is_virtual = R5_ARG3;
1779
1780 check_continuation_enter_argument(regs[pos_cont_obj].first(), reg_cont_obj, "Continuation object");
1781 check_continuation_enter_argument(regs[pos_is_cont].first(), reg_is_cont, "isContinue");
1782 check_continuation_enter_argument(regs[pos_is_virtual].first(), reg_is_virtual, "isVirtualThread");
1783
1784 address resolve_static_call = SharedRuntime::get_resolve_static_call_stub();
1785
1786 address start = __ pc();
1787
1788 Label L_thaw, L_exit;
1789
1790 // i2i entry used at interp_only_mode only
1791 interpreted_entry_offset = __ pc() - start;
1792 {
1793 #ifdef ASSERT
1794 Label is_interp_only;
1795 __ lwz(R0, in_bytes(JavaThread::interp_only_mode_offset()), R16_thread);
1796 __ cmpwi(CR0, R0, 0);
1797 __ bne(CR0, is_interp_only);
1798 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1799 __ bind(is_interp_only);
1800 #endif
1801
1802 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1803 __ ld(reg_cont_obj, Interpreter::stackElementSize*3, R15_esp);
1804 __ lwz(reg_is_cont, Interpreter::stackElementSize*2, R15_esp);
1805 __ lwz(reg_is_virtual, Interpreter::stackElementSize*1, R15_esp);
1806
1807 __ push_cont_fastpath();
1808
1809 OopMap* map = continuation_enter_setup(masm, framesize_words);
1810
1811 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1812 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1813
1814 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1815
1816 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1817 __ cmpwi(CR0, reg_is_cont, 0);
1818 __ bne(CR0, L_thaw);
1819
1820 // --- call Continuation.enter(Continuation c, boolean isContinue)
1821
1822 // Emit compiled static call. The call will be always resolved to the c2i
1823 // entry of Continuation.enter(Continuation c, boolean isContinue).
1824 // There are special cases in SharedRuntime::resolve_static_call_C() and
1825 // SharedRuntime::resolve_sub_helper_internal() to achieve this
1826 // See also corresponding call below.
1827 address c2i_call_pc = __ pc();
1828 int start_offset = __ offset();
1829 // Put the entry point as a constant into the constant pool.
1830 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1831 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1832 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1833
1834 // Emit the trampoline stub which will be related to the branch-and-link below.
1835 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1836 guarantee(stub != nullptr, "no space for trampoline stub");
1837
1838 __ relocate(relocInfo::static_call_type);
1839 // Note: At this point we do not have the address of the trampoline
1840 // stub, and the entry point might be too far away for bl, so __ pc()
1841 // serves as dummy and the bl will be patched later.
1842 __ bl(__ pc());
1843 oop_maps->add_gc_map(__ pc() - start, map);
1844 __ post_call_nop();
1845
1846 __ b(L_exit);
1847
1848 // static stub for the call above
1849 stub = CompiledDirectCall::emit_to_interp_stub(masm, c2i_call_pc);
1850 guarantee(stub != nullptr, "no space for static stub");
1851 }
1852
1853 // compiled entry
1854 __ align(CodeEntryAlignment);
1855 compiled_entry_offset = __ pc() - start;
1856
1857 OopMap* map = continuation_enter_setup(masm, framesize_words);
1858
1859 // Frame is now completed as far as size and linkage.
1860 frame_complete =__ pc() - start;
1861
1862 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1863
1864 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1865 __ cmpwi(CR0, reg_is_cont, 0);
1866 __ bne(CR0, L_thaw);
1867
1868 // --- call Continuation.enter(Continuation c, boolean isContinue)
1869
1870 // Emit compiled static call
1871 // The call needs to be resolved. There's a special case for this in
1872 // SharedRuntime::find_callee_info_helper() which calls
1873 // LinkResolver::resolve_continuation_enter() which resolves the call to
1874 // Continuation.enter(Continuation c, boolean isContinue).
1875 address call_pc = __ pc();
1876 int start_offset = __ offset();
1877 // Put the entry point as a constant into the constant pool.
1878 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1879 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1880 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1881
1882 // Emit the trampoline stub which will be related to the branch-and-link below.
1883 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1884 guarantee(stub != nullptr, "no space for trampoline stub");
1885
1886 __ relocate(relocInfo::static_call_type);
1887 // Note: At this point we do not have the address of the trampoline
1888 // stub, and the entry point might be too far away for bl, so __ pc()
1889 // serves as dummy and the bl will be patched later.
1890 __ bl(__ pc());
1891 oop_maps->add_gc_map(__ pc() - start, map);
1892 __ post_call_nop();
1893
1894 __ b(L_exit);
1895
1896 // --- Thawing path
1897
1898 __ bind(L_thaw);
1899 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1900 __ add_const_optimized(R0, R29_TOC, MacroAssembler::offset_to_global_toc(StubRoutines::cont_thaw()));
1901 __ mtctr(R0);
1902 __ bctrl();
1903 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1904 ContinuationEntry::_return_pc_offset = __ pc() - start;
1905 __ post_call_nop();
1906
1907 // --- Normal exit (resolve/thawing)
1908
1909 __ bind(L_exit);
1910 ContinuationEntry::_cleanup_offset = __ pc() - start;
1911 continuation_enter_cleanup(masm);
1912
1913 // Pop frame and return
1914 DEBUG_ONLY(__ ld_ptr(R0, 0, R1_SP));
1915 __ addi(R1_SP, R1_SP, framesize_words*wordSize);
1916 DEBUG_ONLY(__ cmpd(CR0, R0, R1_SP));
1917 __ asm_assert_eq(FILE_AND_LINE ": inconsistent frame size");
1918 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1919 __ mtlr(R0);
1920 __ blr();
1921
1922 // --- Exception handling path
1923
1924 exception_offset = __ pc() - start;
1925
1926 continuation_enter_cleanup(masm);
1927 Register ex_pc = R17_tos; // nonvolatile register
1928 Register ex_oop = R15_esp; // nonvolatile register
1929 __ ld(ex_pc, _abi0(callers_sp), R1_SP); // Load caller's return pc
1930 __ ld(ex_pc, _abi0(lr), ex_pc);
1931 __ mr(ex_oop, R3_RET); // save return value containing the exception oop
1932 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, ex_pc);
1933 __ mtlr(R3_RET); // the exception handler
1934 __ ld(R1_SP, _abi0(callers_sp), R1_SP); // remove enterSpecial frame
1935
1936 // Continue at exception handler
1937 // See OptoRuntime::generate_exception_blob for register arguments
1938 __ mr(R3_ARG1, ex_oop); // pass exception oop
1939 __ mr(R4_ARG2, ex_pc); // pass exception pc
1940 __ blr();
1941
1942 // static stub for the call above
1943 stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1944 guarantee(stub != nullptr, "no space for static stub");
1945 }
1946
1947 static void gen_continuation_yield(MacroAssembler* masm,
1948 const VMRegPair* regs,
1949 OopMapSet* oop_maps,
1950 int& frame_complete,
1951 int& framesize_words,
1952 int& compiled_entry_offset) {
1953 Register tmp = R10_ARG8;
1954
1955 const int framesize_bytes = (int)align_up((int)frame::native_abi_reg_args_size, frame::alignment_in_bytes);
1956 framesize_words = framesize_bytes / wordSize;
1957
1958 address start = __ pc();
1959 compiled_entry_offset = __ pc() - start;
1960
1961 // Save return pc and push entry frame
1962 __ mflr(tmp);
1963 __ std(tmp, _abi0(lr), R1_SP); // SP->lr = return_pc
1964 __ push_frame(framesize_bytes , R0); // SP -= frame_size_in_bytes
1965
1966 DEBUG_ONLY(__ block_comment("Frame Complete"));
1967 frame_complete = __ pc() - start;
1968 address last_java_pc = __ pc();
1969
1970 // This nop must be exactly at the PC we push into the frame info.
1971 // We use this nop for fast CodeBlob lookup, associate the OopMap
1972 // with it right away.
1973 __ post_call_nop();
1974 OopMap* map = new OopMap(framesize_bytes / VMRegImpl::stack_slot_size, 1);
1975 oop_maps->add_gc_map(last_java_pc - start, map);
1976
1977 __ calculate_address_from_global_toc(tmp, last_java_pc); // will be relocated
1978 __ set_last_Java_frame(R1_SP, tmp);
1979 __ call_VM_leaf(Continuation::freeze_entry(), R16_thread, R1_SP);
1980 __ reset_last_Java_frame();
1981
1982 Label L_pinned;
1983
1984 __ cmpwi(CR0, R3_RET, 0);
1985 __ bne(CR0, L_pinned);
1986
1987 // yield succeeded
1988
1989 // Pop frames of continuation including this stub's frame
1990 __ ld_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1991 // The frame pushed by gen_continuation_enter is on top now again
1992 continuation_enter_cleanup(masm);
1993
1994 // Pop frame and return
1995 Label L_return;
1996 __ bind(L_return);
1997 __ pop_frame();
1998 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1999 __ mtlr(R0);
2000 __ blr();
2001
2002 // yield failed - continuation is pinned
2003
2004 __ bind(L_pinned);
2005
2006 // handle pending exception thrown by freeze
2007 __ ld(tmp, in_bytes(JavaThread::pending_exception_offset()), R16_thread);
2008 __ cmpdi(CR0, tmp, 0);
2009 __ beq(CR0, L_return); // return if no exception is pending
2010 __ pop_frame();
2011 __ ld(R0, _abi0(lr), R1_SP); // Return pc
2012 __ mtlr(R0);
2013 __ load_const_optimized(tmp, StubRoutines::forward_exception_entry(), R0);
2014 __ mtctr(tmp);
2015 __ bctr();
2016 }
2017
2018 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
2019 ::continuation_enter_cleanup(masm);
2020 }
2021
2022 // ---------------------------------------------------------------------------
2023 // Generate a native wrapper for a given method. The method takes arguments
2024 // in the Java compiled code convention, marshals them to the native
2025 // convention (handlizes oops, etc), transitions to native, makes the call,
2026 // returns to java state (possibly blocking), unhandlizes any result and
2027 // returns.
2028 //
2029 // Critical native functions are a shorthand for the use of
2030 // GetPrimtiveArrayCritical and disallow the use of any other JNI
2031 // functions. The wrapper is expected to unpack the arguments before
2032 // passing them to the callee. Critical native functions leave the state _in_Java,
2033 // since they cannot stop for GC.
2034 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
2035 // block and the check for pending exceptions it's impossible for them
2036 // to be thrown.
2037 //
2038 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
2039 const methodHandle& method,
2040 int compile_id,
2041 BasicType *in_sig_bt,
2042 VMRegPair *in_regs,
2043 BasicType ret_type) {
2044 if (method->is_continuation_native_intrinsic()) {
2045 int exception_offset = -1;
2046 OopMapSet* oop_maps = new OopMapSet();
2047 int frame_complete = -1;
2048 int stack_slots = -1;
2049 int interpreted_entry_offset = -1;
2050 int vep_offset = -1;
2051 if (method->is_continuation_enter_intrinsic()) {
2052 gen_continuation_enter(masm,
2053 in_regs,
2054 exception_offset,
2055 oop_maps,
2056 frame_complete,
2057 stack_slots,
2058 interpreted_entry_offset,
2059 vep_offset);
2060 } else if (method->is_continuation_yield_intrinsic()) {
2061 gen_continuation_yield(masm,
2062 in_regs,
2063 oop_maps,
2064 frame_complete,
2065 stack_slots,
2066 vep_offset);
2067 } else {
2068 guarantee(false, "Unknown Continuation native intrinsic");
2069 }
2070
2071 #ifdef ASSERT
2072 if (method->is_continuation_enter_intrinsic()) {
2073 assert(interpreted_entry_offset != -1, "Must be set");
2074 assert(exception_offset != -1, "Must be set");
2075 } else {
2076 assert(interpreted_entry_offset == -1, "Must be unset");
2077 assert(exception_offset == -1, "Must be unset");
2078 }
2079 assert(frame_complete != -1, "Must be set");
2080 assert(stack_slots != -1, "Must be set");
2081 assert(vep_offset != -1, "Must be set");
2082 #endif
2083
2084 __ flush();
2085 nmethod* nm = nmethod::new_native_nmethod(method,
2086 compile_id,
2087 masm->code(),
2088 vep_offset,
2089 frame_complete,
2090 stack_slots,
2091 in_ByteSize(-1),
2092 in_ByteSize(-1),
2093 oop_maps,
2094 exception_offset);
2095 if (nm == nullptr) return nm;
2096 if (method->is_continuation_enter_intrinsic()) {
2097 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
2098 } else if (method->is_continuation_yield_intrinsic()) {
2099 _cont_doYield_stub = nm;
2100 }
2101 return nm;
2102 }
2103
2104 if (method->is_method_handle_intrinsic()) {
2105 vmIntrinsics::ID iid = method->intrinsic_id();
2106 intptr_t start = (intptr_t)__ pc();
2107 int vep_offset = ((intptr_t)__ pc()) - start;
2108 gen_special_dispatch(masm,
2109 method,
2110 in_sig_bt,
2111 in_regs);
2112 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
2113 __ flush();
2114 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
2115 return nmethod::new_native_nmethod(method,
2116 compile_id,
2117 masm->code(),
2118 vep_offset,
2119 frame_complete,
2120 stack_slots / VMRegImpl::slots_per_word,
2121 in_ByteSize(-1),
2122 in_ByteSize(-1),
2123 (OopMapSet*)nullptr);
2124 }
2125
2126 address native_func = method->native_function();
2127 assert(native_func != nullptr, "must have function");
2128
2129 // First, create signature for outgoing C call
2130 // --------------------------------------------------------------------------
2131
2132 int total_in_args = method->size_of_parameters();
2133 // We have received a description of where all the java args are located
2134 // on entry to the wrapper. We need to convert these args to where
2135 // the jni function will expect them. To figure out where they go
2136 // we convert the java signature to a C signature by inserting
2137 // the hidden arguments as arg[0] and possibly arg[1] (static method)
2138
2139 // Calculate the total number of C arguments and create arrays for the
2140 // signature and the outgoing registers.
2141 // On ppc64, we have two arrays for the outgoing registers, because
2142 // some floating-point arguments must be passed in registers _and_
2143 // in stack locations.
2144 bool method_is_static = method->is_static();
2145 int total_c_args = total_in_args + (method_is_static ? 2 : 1);
2146
2147 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2148 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2149
2150 // Create the signature for the C call:
2151 // 1) add the JNIEnv*
2152 // 2) add the class if the method is static
2153 // 3) copy the rest of the incoming signature (shifted by the number of
2154 // hidden arguments).
2155
2156 int argc = 0;
2157 out_sig_bt[argc++] = T_ADDRESS;
2158 if (method->is_static()) {
2159 out_sig_bt[argc++] = T_OBJECT;
2160 }
2161
2162 for (int i = 0; i < total_in_args ; i++ ) {
2163 out_sig_bt[argc++] = in_sig_bt[i];
2164 }
2165
2166
2167 // Compute the wrapper's frame size.
2168 // --------------------------------------------------------------------------
2169
2170 // Now figure out where the args must be stored and how much stack space
2171 // they require.
2172 //
2173 // Compute framesize for the wrapper. We need to handlize all oops in
2174 // incoming registers.
2175 //
2176 // Calculate the total number of stack slots we will need:
2177 // 1) abi requirements
2178 // 2) outgoing arguments
2179 // 3) space for inbound oop handle area
2180 // 4) space for handlizing a klass if static method
2181 // 5) space for a lock if synchronized method
2182 // 6) workspace for saving return values, int <-> float reg moves, etc.
2183 // 7) alignment
2184 //
2185 // Layout of the native wrapper frame:
2186 // (stack grows upwards, memory grows downwards)
2187 //
2188 // NW [ABI_REG_ARGS] <-- 1) R1_SP
2189 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset
2190 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset
2191 // klass <-- 4) R1_SP + klass_offset
2192 // lock <-- 5) R1_SP + lock_offset
2193 // [workspace] <-- 6) R1_SP + workspace_offset
2194 // [alignment] (optional) <-- 7)
2195 // caller [JIT_TOP_ABI_48] <-- r_callers_sp
2196 //
2197 // - *_slot_offset Indicates offset from SP in number of stack slots.
2198 // - *_offset Indicates offset from SP in bytes.
2199
2200 int stack_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args) + // 1+2)
2201 SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention.
2202
2203 // Now the space for the inbound oop handle area.
2204 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word;
2205
2206 int oop_handle_slot_offset = stack_slots;
2207 stack_slots += total_save_slots; // 3)
2208
2209 int klass_slot_offset = 0;
2210 int klass_offset = -1;
2211 if (method_is_static) { // 4)
2212 klass_slot_offset = stack_slots;
2213 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
2214 stack_slots += VMRegImpl::slots_per_word;
2215 }
2216
2217 int lock_slot_offset = 0;
2218 int lock_offset = -1;
2219 if (method->is_synchronized()) { // 5)
2220 lock_slot_offset = stack_slots;
2221 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size;
2222 stack_slots += VMRegImpl::slots_per_word;
2223 }
2224
2225 int workspace_slot_offset = stack_slots; // 6)
2226 stack_slots += 2;
2227
2228 // Now compute actual number of stack words we need.
2229 // Rounding to make stack properly aligned.
2230 stack_slots = align_up(stack_slots, // 7)
2231 frame::alignment_in_bytes / VMRegImpl::stack_slot_size);
2232 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size;
2233
2234
2235 // Now we can start generating code.
2236 // --------------------------------------------------------------------------
2237
2238 intptr_t start_pc = (intptr_t)__ pc();
2239 intptr_t vep_start_pc;
2240 intptr_t frame_done_pc;
2241
2242 Label handle_pending_exception;
2243 Label last_java_pc;
2244
2245 Register r_callers_sp = R21;
2246 Register r_temp_1 = R22;
2247 Register r_temp_2 = R23;
2248 Register r_temp_3 = R24;
2249 Register r_temp_4 = R25;
2250 Register r_temp_5 = R26;
2251 Register r_temp_6 = R27;
2252 Register r_last_java_pc = R28;
2253
2254 Register r_carg1_jnienv = noreg;
2255 Register r_carg2_classorobject = noreg;
2256 r_carg1_jnienv = out_regs[0].first()->as_Register();
2257 r_carg2_classorobject = out_regs[1].first()->as_Register();
2258
2259
2260 // Generate the Unverified Entry Point (UEP).
2261 // --------------------------------------------------------------------------
2262 assert(start_pc == (intptr_t)__ pc(), "uep must be at start");
2263
2264 // Check ic: object class == cached class?
2265 if (!method_is_static) {
2266 __ ic_check(4 /* end_alignment */);
2267 }
2268
2269 // Generate the Verified Entry Point (VEP).
2270 // --------------------------------------------------------------------------
2271 vep_start_pc = (intptr_t)__ pc();
2272
2273 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
2274 Label L_skip_barrier;
2275 Register klass = r_temp_1;
2276 // Notify OOP recorder (don't need the relocation)
2277 AddressLiteral md = __ constant_metadata_address(method->method_holder());
2278 __ load_const_optimized(klass, md.value(), R0);
2279 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
2280
2281 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
2282 __ mtctr(klass);
2283 __ bctr();
2284
2285 __ bind(L_skip_barrier);
2286 }
2287
2288 __ save_LR(r_temp_1);
2289 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame.
2290 __ mr(r_callers_sp, R1_SP); // Remember frame pointer.
2291 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame.
2292
2293 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2294 bs->nmethod_entry_barrier(masm, r_temp_1);
2295
2296 frame_done_pc = (intptr_t)__ pc();
2297
2298 // Native nmethod wrappers never take possession of the oop arguments.
2299 // So the caller will gc the arguments.
2300 // The only thing we need an oopMap for is if the call is static.
2301 //
2302 // An OopMap for lock (and class if static), and one for the VM call itself.
2303 OopMapSet *oop_maps = new OopMapSet();
2304 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2305
2306 // Move arguments from register/stack to register/stack.
2307 // --------------------------------------------------------------------------
2308 //
2309 // We immediately shuffle the arguments so that for any vm call we have
2310 // to make from here on out (sync slow path, jvmti, etc.) we will have
2311 // captured the oops from our caller and have a valid oopMap for them.
2312 //
2313 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2314 // (derived from JavaThread* which is in R16_thread) and, if static,
2315 // the class mirror instead of a receiver. This pretty much guarantees that
2316 // register layout will not match. We ignore these extra arguments during
2317 // the shuffle. The shuffle is described by the two calling convention
2318 // vectors we have in our possession. We simply walk the java vector to
2319 // get the source locations and the c vector to get the destinations.
2320
2321 // Record sp-based slot for receiver on stack for non-static methods.
2322 int receiver_offset = -1;
2323
2324 // We move the arguments backward because the floating point registers
2325 // destination will always be to a register with a greater or equal
2326 // register number or the stack.
2327 // in is the index of the incoming Java arguments
2328 // out is the index of the outgoing C arguments
2329
2330 #ifdef ASSERT
2331 bool reg_destroyed[Register::number_of_registers];
2332 bool freg_destroyed[FloatRegister::number_of_registers];
2333 for (int r = 0 ; r < Register::number_of_registers ; r++) {
2334 reg_destroyed[r] = false;
2335 }
2336 for (int f = 0 ; f < FloatRegister::number_of_registers ; f++) {
2337 freg_destroyed[f] = false;
2338 }
2339 #endif // ASSERT
2340
2341 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) {
2342
2343 #ifdef ASSERT
2344 if (in_regs[in].first()->is_Register()) {
2345 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!");
2346 } else if (in_regs[in].first()->is_FloatRegister()) {
2347 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!");
2348 }
2349 if (out_regs[out].first()->is_Register()) {
2350 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true;
2351 } else if (out_regs[out].first()->is_FloatRegister()) {
2352 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true;
2353 }
2354 #endif // ASSERT
2355
2356 switch (in_sig_bt[in]) {
2357 case T_BOOLEAN:
2358 case T_CHAR:
2359 case T_BYTE:
2360 case T_SHORT:
2361 case T_INT:
2362 // Move int and do sign extension.
2363 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2364 break;
2365 case T_LONG:
2366 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2367 break;
2368 case T_ARRAY:
2369 case T_OBJECT:
2370 object_move(masm, stack_slots,
2371 oop_map, oop_handle_slot_offset,
2372 ((in == 0) && (!method_is_static)), &receiver_offset,
2373 in_regs[in], out_regs[out],
2374 r_callers_sp, r_temp_1, r_temp_2);
2375 break;
2376 case T_VOID:
2377 break;
2378 case T_FLOAT:
2379 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2380 break;
2381 case T_DOUBLE:
2382 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2383 break;
2384 case T_ADDRESS:
2385 fatal("found type (T_ADDRESS) in java args");
2386 break;
2387 default:
2388 ShouldNotReachHere();
2389 break;
2390 }
2391 }
2392
2393 // Pre-load a static method's oop into ARG2.
2394 // Used both by locking code and the normal JNI call code.
2395 if (method_is_static) {
2396 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()),
2397 r_carg2_classorobject);
2398
2399 // Now handlize the static class mirror in carg2. It's known not-null.
2400 __ std(r_carg2_classorobject, klass_offset, R1_SP);
2401 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2402 __ addi(r_carg2_classorobject, R1_SP, klass_offset);
2403 }
2404
2405 // Get JNIEnv* which is first argument to native.
2406 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset()));
2407
2408 // NOTE:
2409 //
2410 // We have all of the arguments setup at this point.
2411 // We MUST NOT touch any outgoing regs from this point on.
2412 // So if we must call out we must push a new frame.
2413
2414 // The last java pc will also be used as resume pc if this is the wrapper for wait0.
2415 // For this purpose the precise location matters but not for oopmap lookup.
2416 __ calculate_address_from_global_toc(r_last_java_pc, last_java_pc, true, true, true, true);
2417
2418 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below.
2419 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register");
2420
2421 # if 0
2422 // DTrace method entry
2423 # endif
2424
2425 // Lock a synchronized method.
2426 // --------------------------------------------------------------------------
2427
2428 if (method->is_synchronized()) {
2429 Register r_oop = r_temp_4;
2430 const Register r_box = r_temp_5;
2431 Label done, locked;
2432
2433 // Load the oop for the object or class. r_carg2_classorobject contains
2434 // either the handlized oop from the incoming arguments or the handlized
2435 // class mirror (if the method is static).
2436 __ ld(r_oop, 0, r_carg2_classorobject);
2437
2438 // Get the lock box slot's address.
2439 __ addi(r_box, R1_SP, lock_offset);
2440
2441 // Try fastpath for locking.
2442 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2443 Register r_temp_3_or_noreg = UseObjectMonitorTable ? r_temp_3 : noreg;
2444 __ compiler_fast_lock_lightweight_object(CR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3_or_noreg);
2445 __ beq(CR0, locked);
2446
2447 // None of the above fast optimizations worked so we have to get into the
2448 // slow case of monitor enter. Inline a special case of call_VM that
2449 // disallows any pending_exception.
2450
2451 // Save argument registers and leave room for C-compatible ABI_REG_ARGS.
2452 int frame_size = frame::native_abi_reg_args_size + align_up(total_c_args * wordSize, frame::alignment_in_bytes);
2453 __ mr(R11_scratch1, R1_SP);
2454 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs);
2455
2456 // Do the call.
2457 __ set_last_Java_frame(R11_scratch1, r_last_java_pc);
2458 assert(r_last_java_pc->is_nonvolatile(), "r_last_java_pc needs to be preserved accross complete_monitor_locking_C call");
2459 // The following call will not be preempted.
2460 // push_cont_fastpath forces freeze slow path in case we try to preempt where we will pin the
2461 // vthread to the carrier (see FreezeBase::recurse_freeze_native_frame()).
2462 __ push_cont_fastpath();
2463 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread);
2464 __ pop_cont_fastpath();
2465 __ reset_last_Java_frame();
2466
2467 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs);
2468
2469 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2470 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C");
2471
2472 __ bind(locked);
2473 }
2474
2475 __ set_last_Java_frame(R1_SP, r_last_java_pc);
2476
2477 // Publish thread state
2478 // --------------------------------------------------------------------------
2479
2480 // Transition from _thread_in_Java to _thread_in_native.
2481 __ li(R0, _thread_in_native);
2482 __ release();
2483 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2484 __ stw(R0, thread_(thread_state));
2485
2486
2487 // The JNI call
2488 // --------------------------------------------------------------------------
2489 __ call_c(native_func, relocInfo::runtime_call_type);
2490
2491
2492 // Now, we are back from the native code.
2493
2494
2495 // Unpack the native result.
2496 // --------------------------------------------------------------------------
2497
2498 // For int-types, we do any needed sign-extension required.
2499 // Care must be taken that the return values (R3_RET and F1_RET)
2500 // will survive any VM calls for blocking or unlocking.
2501 // An OOP result (handle) is done specially in the slow-path code.
2502
2503 switch (ret_type) {
2504 case T_VOID: break; // Nothing to do!
2505 case T_FLOAT: break; // Got it where we want it (unless slow-path).
2506 case T_DOUBLE: break; // Got it where we want it (unless slow-path).
2507 case T_LONG: break; // Got it where we want it (unless slow-path).
2508 case T_OBJECT: break; // Really a handle.
2509 // Cannot de-handlize until after reclaiming jvm_lock.
2510 case T_ARRAY: break;
2511
2512 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1)
2513 __ normalize_bool(R3_RET);
2514 break;
2515 }
2516 case T_BYTE: { // sign extension
2517 __ extsb(R3_RET, R3_RET);
2518 break;
2519 }
2520 case T_CHAR: { // unsigned result
2521 __ andi(R3_RET, R3_RET, 0xffff);
2522 break;
2523 }
2524 case T_SHORT: { // sign extension
2525 __ extsh(R3_RET, R3_RET);
2526 break;
2527 }
2528 case T_INT: // nothing to do
2529 break;
2530 default:
2531 ShouldNotReachHere();
2532 break;
2533 }
2534
2535 // Publish thread state
2536 // --------------------------------------------------------------------------
2537
2538 // Switch thread to "native transition" state before reading the
2539 // synchronization state. This additional state is necessary because reading
2540 // and testing the synchronization state is not atomic w.r.t. GC, as this
2541 // scenario demonstrates:
2542 // - Java thread A, in _thread_in_native state, loads _not_synchronized
2543 // and is preempted.
2544 // - VM thread changes sync state to synchronizing and suspends threads
2545 // for GC.
2546 // - Thread A is resumed to finish this native method, but doesn't block
2547 // here since it didn't see any synchronization in progress, and escapes.
2548
2549 // Transition from _thread_in_native to _thread_in_native_trans.
2550 __ li(R0, _thread_in_native_trans);
2551 __ release();
2552 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2553 __ stw(R0, thread_(thread_state));
2554
2555
2556 // Must we block?
2557 // --------------------------------------------------------------------------
2558
2559 // Block, if necessary, before resuming in _thread_in_Java state.
2560 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2561 {
2562 Label no_block, sync;
2563
2564 // Force this write out before the read below.
2565 if (!UseSystemMemoryBarrier) {
2566 __ fence();
2567 }
2568
2569 Register sync_state_addr = r_temp_4;
2570 Register sync_state = r_temp_5;
2571 Register suspend_flags = r_temp_6;
2572
2573 // No synchronization in progress nor yet synchronized
2574 // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
2575 __ safepoint_poll(sync, sync_state, true /* at_return */, false /* in_nmethod */);
2576
2577 // Not suspended.
2578 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
2579 __ lwz(suspend_flags, thread_(suspend_flags));
2580 __ cmpwi(CR1, suspend_flags, 0);
2581 __ beq(CR1, no_block);
2582
2583 // Block. Save any potential method result value before the operation and
2584 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2585 // lets us share the oopMap we used when we went native rather than create
2586 // a distinct one for this pc.
2587 __ bind(sync);
2588 __ isync();
2589
2590 address entry_point =
2591 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans);
2592 save_native_result(masm, ret_type, workspace_slot_offset);
2593 __ call_VM_leaf(entry_point, R16_thread);
2594 restore_native_result(masm, ret_type, workspace_slot_offset);
2595
2596 __ bind(no_block);
2597
2598 // Publish thread state.
2599 // --------------------------------------------------------------------------
2600
2601 // Thread state is thread_in_native_trans. Any safepoint blocking has
2602 // already happened so we can now change state to _thread_in_Java.
2603
2604 // Transition from _thread_in_native_trans to _thread_in_Java.
2605 __ li(R0, _thread_in_Java);
2606 __ lwsync(); // Acquire safepoint and suspend state, release thread state.
2607 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2608 __ stw(R0, thread_(thread_state));
2609
2610 // Check preemption for Object.wait()
2611 if (method->is_object_wait0()) {
2612 Label not_preempted;
2613 __ ld(R0, in_bytes(JavaThread::preempt_alternate_return_offset()), R16_thread);
2614 __ cmpdi(CR0, R0, 0);
2615 __ beq(CR0, not_preempted);
2616 __ mtlr(R0);
2617 __ li(R0, 0);
2618 __ std(R0, in_bytes(JavaThread::preempt_alternate_return_offset()), R16_thread);
2619 __ blr();
2620 __ bind(not_preempted);
2621 }
2622 __ bind(last_java_pc);
2623 // We use the same pc/oopMap repeatedly when we call out above.
2624 intptr_t oopmap_pc = (intptr_t) __ pc();
2625 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map);
2626 }
2627
2628 // Reguard any pages if necessary.
2629 // --------------------------------------------------------------------------
2630
2631 Label no_reguard;
2632 __ lwz(r_temp_1, thread_(stack_guard_state));
2633 __ cmpwi(CR0, r_temp_1, StackOverflow::stack_guard_yellow_reserved_disabled);
2634 __ bne(CR0, no_reguard);
2635
2636 save_native_result(masm, ret_type, workspace_slot_offset);
2637 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2638 restore_native_result(masm, ret_type, workspace_slot_offset);
2639
2640 __ bind(no_reguard);
2641
2642
2643 // Unlock
2644 // --------------------------------------------------------------------------
2645
2646 if (method->is_synchronized()) {
2647 const Register r_oop = r_temp_4;
2648 const Register r_box = r_temp_5;
2649 const Register r_exception = r_temp_6;
2650 Label done;
2651
2652 // Get oop and address of lock object box.
2653 if (method_is_static) {
2654 assert(klass_offset != -1, "");
2655 __ ld(r_oop, klass_offset, R1_SP);
2656 } else {
2657 assert(receiver_offset != -1, "");
2658 __ ld(r_oop, receiver_offset, R1_SP);
2659 }
2660 __ addi(r_box, R1_SP, lock_offset);
2661
2662 // Try fastpath for unlocking.
2663 __ compiler_fast_unlock_lightweight_object(CR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2664 __ beq(CR0, done);
2665
2666 // Save and restore any potential method result value around the unlocking operation.
2667 save_native_result(masm, ret_type, workspace_slot_offset);
2668
2669 // Must save pending exception around the slow-path VM call. Since it's a
2670 // leaf call, the pending exception (if any) can be kept in a register.
2671 __ ld(r_exception, thread_(pending_exception));
2672 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile");
2673 __ li(R0, 0);
2674 __ std(R0, thread_(pending_exception));
2675
2676 // Slow case of monitor enter.
2677 // Inline a special case of call_VM that disallows any pending_exception.
2678 // Arguments are (oop obj, BasicLock* lock, JavaThread* thread).
2679 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box, R16_thread);
2680
2681 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2682 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C");
2683
2684 restore_native_result(masm, ret_type, workspace_slot_offset);
2685
2686 // Check_forward_pending_exception jump to forward_exception if any pending
2687 // exception is set. The forward_exception routine expects to see the
2688 // exception in pending_exception and not in a register. Kind of clumsy,
2689 // since all folks who branch to forward_exception must have tested
2690 // pending_exception first and hence have it in a register already.
2691 __ std(r_exception, thread_(pending_exception));
2692
2693 __ bind(done);
2694 }
2695
2696 # if 0
2697 // DTrace method exit
2698 # endif
2699
2700 // Clear "last Java frame" SP and PC.
2701 // --------------------------------------------------------------------------
2702
2703 // Last java frame won't be set if we're resuming after preemption
2704 bool maybe_preempted = method->is_object_wait0();
2705 __ reset_last_Java_frame(!maybe_preempted /* check_last_java_sp */);
2706
2707 // Unbox oop result, e.g. JNIHandles::resolve value.
2708 // --------------------------------------------------------------------------
2709
2710 if (is_reference_type(ret_type)) {
2711 __ resolve_jobject(R3_RET, r_temp_1, r_temp_2, MacroAssembler::PRESERVATION_NONE);
2712 }
2713
2714 if (CheckJNICalls) {
2715 // clear_pending_jni_exception_check
2716 __ load_const_optimized(R0, 0L);
2717 __ st_ptr(R0, JavaThread::pending_jni_exception_check_fn_offset(), R16_thread);
2718 }
2719
2720 // Reset handle block.
2721 // --------------------------------------------------------------------------
2722 __ ld(r_temp_1, thread_(active_handles));
2723 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
2724 __ li(r_temp_2, 0);
2725 __ stw(r_temp_2, in_bytes(JNIHandleBlock::top_offset()), r_temp_1);
2726
2727 // Prepare for return
2728 // --------------------------------------------------------------------------
2729 __ pop_frame();
2730 __ restore_LR(R11);
2731
2732 #if INCLUDE_JFR
2733 // We need to do a poll test after unwind in case the sampler
2734 // managed to sample the native frame after returning to Java.
2735 Label L_stub;
2736 int safepoint_offset = __ offset();
2737 if (!UseSIGTRAP) {
2738 __ relocate(relocInfo::poll_return_type);
2739 }
2740 __ safepoint_poll(L_stub, r_temp_2, true /* at_return */, true /* in_nmethod: frame already popped */);
2741 #endif // INCLUDE_JFR
2742
2743 // Check for pending exceptions.
2744 // --------------------------------------------------------------------------
2745 __ ld(r_temp_2, thread_(pending_exception));
2746 __ cmpdi(CR0, r_temp_2, 0);
2747 __ bne(CR0, handle_pending_exception);
2748
2749 // Return.
2750 __ blr();
2751
2752 // Handler for return safepoint (out-of-line).
2753 #if INCLUDE_JFR
2754 if (!UseSIGTRAP) {
2755 __ bind(L_stub);
2756 __ jump_to_polling_page_return_handler_blob(safepoint_offset);
2757 }
2758 #endif // INCLUDE_JFR
2759
2760 // Handler for pending exceptions (out-of-line).
2761 // --------------------------------------------------------------------------
2762 // Since this is a native call, we know the proper exception handler
2763 // is the empty function. We just pop this frame and then jump to
2764 // forward_exception_entry.
2765 __ bind(handle_pending_exception);
2766 __ b64_patchable((address)StubRoutines::forward_exception_entry(),
2767 relocInfo::runtime_call_type);
2768
2769 // Done.
2770 // --------------------------------------------------------------------------
2771
2772 __ flush();
2773
2774 nmethod *nm = nmethod::new_native_nmethod(method,
2775 compile_id,
2776 masm->code(),
2777 vep_start_pc-start_pc,
2778 frame_done_pc-start_pc,
2779 stack_slots / VMRegImpl::slots_per_word,
2780 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2781 in_ByteSize(lock_offset),
2782 oop_maps);
2783
2784 return nm;
2785 }
2786
2787 // This function returns the adjust size (in number of words) to a c2i adapter
2788 // activation for use during deoptimization.
2789 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2790 return align_up((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::frame_alignment_in_words);
2791 }
2792
2793 uint SharedRuntime::in_preserve_stack_slots() {
2794 return frame::jit_in_preserve_size / VMRegImpl::stack_slot_size;
2795 }
2796
2797 uint SharedRuntime::out_preserve_stack_slots() {
2798 #if defined(COMPILER1) || defined(COMPILER2)
2799 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size;
2800 #else
2801 return 0;
2802 #endif
2803 }
2804
2805 VMReg SharedRuntime::thread_register() {
2806 // On PPC virtual threads don't save the JavaThread* in their context (e.g. C1 stub frames).
2807 ShouldNotCallThis();
2808 return nullptr;
2809 }
2810
2811 #if defined(COMPILER1) || defined(COMPILER2)
2812 // Frame generation for deopt and uncommon trap blobs.
2813 static void push_skeleton_frame(MacroAssembler* masm, bool deopt,
2814 /* Read */
2815 Register unroll_block_reg,
2816 /* Update */
2817 Register frame_sizes_reg,
2818 Register number_of_frames_reg,
2819 Register pcs_reg,
2820 /* Invalidate */
2821 Register frame_size_reg,
2822 Register pc_reg) {
2823
2824 __ ld(pc_reg, 0, pcs_reg);
2825 __ ld(frame_size_reg, 0, frame_sizes_reg);
2826 __ std(pc_reg, _abi0(lr), R1_SP);
2827 __ push_frame(frame_size_reg, R0/*tmp*/);
2828 __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP);
2829 __ addi(number_of_frames_reg, number_of_frames_reg, -1);
2830 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize);
2831 __ addi(pcs_reg, pcs_reg, wordSize);
2832 }
2833
2834 // Loop through the UnrollBlock info and create new frames.
2835 static void push_skeleton_frames(MacroAssembler* masm, bool deopt,
2836 /* read */
2837 Register unroll_block_reg,
2838 /* invalidate */
2839 Register frame_sizes_reg,
2840 Register number_of_frames_reg,
2841 Register pcs_reg,
2842 Register frame_size_reg,
2843 Register pc_reg) {
2844 Label loop;
2845
2846 // _number_of_frames is of type int (deoptimization.hpp)
2847 __ lwa(number_of_frames_reg,
2848 in_bytes(Deoptimization::UnrollBlock::number_of_frames_offset()),
2849 unroll_block_reg);
2850 __ ld(pcs_reg,
2851 in_bytes(Deoptimization::UnrollBlock::frame_pcs_offset()),
2852 unroll_block_reg);
2853 __ ld(frame_sizes_reg,
2854 in_bytes(Deoptimization::UnrollBlock::frame_sizes_offset()),
2855 unroll_block_reg);
2856
2857 // stack: (caller_of_deoptee, ...).
2858
2859 // At this point we either have an interpreter frame or a compiled
2860 // frame on top of stack. If it is a compiled frame we push a new c2i
2861 // adapter here
2862
2863 // Memorize top-frame stack-pointer.
2864 __ mr(frame_size_reg/*old_sp*/, R1_SP);
2865
2866 // Resize interpreter top frame OR C2I adapter.
2867
2868 // At this moment, the top frame (which is the caller of the deoptee) is
2869 // an interpreter frame or a newly pushed C2I adapter or an entry frame.
2870 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the
2871 // outgoing arguments.
2872 //
2873 // In order to push the interpreter frame for the deoptee, we need to
2874 // resize the top frame such that we are able to place the deoptee's
2875 // locals in the frame.
2876 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI
2877 // into a valid PARENT_IJAVA_FRAME_ABI.
2878
2879 __ lwa(R11_scratch1,
2880 in_bytes(Deoptimization::UnrollBlock::caller_adjustment_offset()),
2881 unroll_block_reg);
2882 __ neg(R11_scratch1, R11_scratch1);
2883
2884 // R11_scratch1 contains size of locals for frame resizing.
2885 // R12_scratch2 contains top frame's lr.
2886
2887 // Resize frame by complete frame size prevents TOC from being
2888 // overwritten by locals. A more stack space saving way would be
2889 // to copy the TOC to its location in the new abi.
2890 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size);
2891
2892 // now, resize the frame
2893 __ resize_frame(R11_scratch1, pc_reg/*tmp*/);
2894
2895 // In the case where we have resized a c2i frame above, the optional
2896 // alignment below the locals has size 32 (why?).
2897 __ std(R12_scratch2, _abi0(lr), R1_SP);
2898
2899 // Initialize initial_caller_sp.
2900 __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP);
2901
2902 #ifdef ASSERT
2903 // Make sure that there is at least one entry in the array.
2904 __ cmpdi(CR0, number_of_frames_reg, 0);
2905 __ asm_assert_ne("array_size must be > 0");
2906 #endif
2907
2908 // Now push the new interpreter frames.
2909 //
2910 __ bind(loop);
2911 // Allocate a new frame, fill in the pc.
2912 push_skeleton_frame(masm, deopt,
2913 unroll_block_reg,
2914 frame_sizes_reg,
2915 number_of_frames_reg,
2916 pcs_reg,
2917 frame_size_reg,
2918 pc_reg);
2919 __ cmpdi(CR0, number_of_frames_reg, 0);
2920 __ bne(CR0, loop);
2921
2922 // Get the return address pointing into the template interpreter.
2923 __ ld(R0, 0, pcs_reg);
2924 // Store it in the top interpreter frame.
2925 __ std(R0, _abi0(lr), R1_SP);
2926 // Initialize frame_manager_lr of interpreter top frame.
2927 }
2928 #endif
2929
2930 void SharedRuntime::generate_deopt_blob() {
2931 // Allocate space for the code
2932 ResourceMark rm;
2933 // Setup code generation tools
2934 const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
2935 CodeBuffer buffer(name, 2048, 1024);
2936 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2937 Label exec_mode_initialized;
2938 int frame_size_in_words;
2939 OopMap* map = nullptr;
2940 OopMapSet *oop_maps = new OopMapSet();
2941
2942 // size of ABI112 plus spill slots for R3_RET and F1_RET.
2943 const int frame_size_in_bytes = frame::native_abi_reg_args_spill_size;
2944 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
2945 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info.
2946
2947 const Register exec_mode_reg = R21_tmp1;
2948
2949 const address start = __ pc();
2950
2951 #if defined(COMPILER1) || defined(COMPILER2)
2952 // --------------------------------------------------------------------------
2953 // Prolog for non exception case!
2954
2955 // We have been called from the deopt handler of the deoptee.
2956 //
2957 // deoptee:
2958 // ...
2959 // call X
2960 // ...
2961 // deopt_handler: call_deopt_stub
2962 // cur. return pc --> ...
2963 //
2964 // So currently SR_LR points behind the call in the deopt handler.
2965 // We adjust it such that it points to the start of the deopt handler.
2966 // The return_pc has been stored in the frame of the deoptee and
2967 // will replace the address of the deopt_handler in the call
2968 // to Deoptimization::fetch_unroll_info below.
2969 // We can't grab a free register here, because all registers may
2970 // contain live values, so let the RegisterSaver do the adjustment
2971 // of the return pc.
2972 const int return_pc_adjustment_no_exception = -MacroAssembler::bl64_patchable_size;
2973
2974 // Push the "unpack frame"
2975 // Save everything in sight.
2976 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2977 &first_frame_size_in_bytes,
2978 /*generate_oop_map=*/ true,
2979 return_pc_adjustment_no_exception,
2980 RegisterSaver::return_pc_is_lr);
2981 assert(map != nullptr, "OopMap must have been created");
2982
2983 __ li(exec_mode_reg, Deoptimization::Unpack_deopt);
2984 // Save exec mode for unpack_frames.
2985 __ b(exec_mode_initialized);
2986
2987 // --------------------------------------------------------------------------
2988 // Prolog for exception case
2989
2990 // An exception is pending.
2991 // We have been called with a return (interpreter) or a jump (exception blob).
2992 //
2993 // - R3_ARG1: exception oop
2994 // - R4_ARG2: exception pc
2995
2996 int exception_offset = __ pc() - start;
2997
2998 BLOCK_COMMENT("Prolog for exception case");
2999
3000 // Store exception oop and pc in thread (location known to GC).
3001 // This is needed since the call to "fetch_unroll_info()" may safepoint.
3002 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3003 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3004 __ std(R4_ARG2, _abi0(lr), R1_SP);
3005
3006 // Vanilla deoptimization with an exception pending in exception_oop.
3007 int exception_in_tls_offset = __ pc() - start;
3008
3009 // Push the "unpack frame".
3010 // Save everything in sight.
3011 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3012 &first_frame_size_in_bytes,
3013 /*generate_oop_map=*/ false,
3014 /*return_pc_adjustment_exception=*/ 0,
3015 RegisterSaver::return_pc_is_pre_saved);
3016
3017 // Deopt during an exception. Save exec mode for unpack_frames.
3018 __ li(exec_mode_reg, Deoptimization::Unpack_exception);
3019
3020 // fall through
3021
3022 int reexecute_offset = 0;
3023 #ifdef COMPILER1
3024 __ b(exec_mode_initialized);
3025
3026 // Reexecute entry, similar to c2 uncommon trap
3027 reexecute_offset = __ pc() - start;
3028
3029 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3030 &first_frame_size_in_bytes,
3031 /*generate_oop_map=*/ false,
3032 /*return_pc_adjustment_reexecute=*/ 0,
3033 RegisterSaver::return_pc_is_pre_saved);
3034 __ li(exec_mode_reg, Deoptimization::Unpack_reexecute);
3035 #endif
3036
3037 // --------------------------------------------------------------------------
3038 __ BIND(exec_mode_initialized);
3039
3040 const Register unroll_block_reg = R22_tmp2;
3041
3042 // We need to set `last_Java_frame' because `fetch_unroll_info' will
3043 // call `last_Java_frame()'. The value of the pc in the frame is not
3044 // particularly important. It just needs to identify this blob.
3045 __ set_last_Java_frame(R1_SP, noreg);
3046
3047 // With EscapeAnalysis turned on, this call may safepoint!
3048 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread, exec_mode_reg);
3049 address calls_return_pc = __ last_calls_return_pc();
3050 // Set an oopmap for the call site that describes all our saved registers.
3051 oop_maps->add_gc_map(calls_return_pc - start, map);
3052
3053 __ reset_last_Java_frame();
3054 // Save the return value.
3055 __ mr(unroll_block_reg, R3_RET);
3056
3057 // Restore only the result registers that have been saved
3058 // by save_volatile_registers(...).
3059 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes);
3060
3061 // reload the exec mode from the UnrollBlock (it might have changed)
3062 __ lwz(exec_mode_reg, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3063 // In excp_deopt_mode, restore and clear exception oop which we
3064 // stored in the thread during exception entry above. The exception
3065 // oop will be the return value of this stub.
3066 Label skip_restore_excp;
3067 __ cmpdi(CR0, exec_mode_reg, Deoptimization::Unpack_exception);
3068 __ bne(CR0, skip_restore_excp);
3069 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3070 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3071 __ li(R0, 0);
3072 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3073 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3074 __ BIND(skip_restore_excp);
3075
3076 __ pop_frame();
3077
3078 // stack: (deoptee, optional i2c, caller of deoptee, ...).
3079
3080 // pop the deoptee's frame
3081 __ pop_frame();
3082
3083 // stack: (caller_of_deoptee, ...).
3084
3085 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3086 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3087 // and the frame is effectively not resized.
3088 Register caller_sp = R23_tmp3;
3089 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3090 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3091
3092 // Loop through the `UnrollBlock' info and create interpreter frames.
3093 push_skeleton_frames(masm, true/*deopt*/,
3094 unroll_block_reg,
3095 R23_tmp3,
3096 R24_tmp4,
3097 R25_tmp5,
3098 R26_tmp6,
3099 R27_tmp7);
3100
3101 // stack: (skeletal interpreter frame, ..., optional skeletal
3102 // interpreter frame, optional c2i, caller of deoptee, ...).
3103
3104 // push an `unpack_frame' taking care of float / int return values.
3105 __ push_frame(frame_size_in_bytes, R0/*tmp*/);
3106
3107 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3108 // skeletal interpreter frame, optional c2i, caller of deoptee,
3109 // ...).
3110
3111 // Spill live volatile registers since we'll do a call.
3112 __ std( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3113 __ stfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3114
3115 // Let the unpacker layout information in the skeletal frames just
3116 // allocated.
3117 __ calculate_address_from_global_toc(R3_RET, calls_return_pc, true, true, true, true);
3118 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET);
3119 // This is a call to a LEAF method, so no oop map is required.
3120 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3121 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/);
3122 __ reset_last_Java_frame();
3123
3124 // Restore the volatiles saved above.
3125 __ ld( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3126 __ lfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3127
3128 // Pop the unpack frame.
3129 __ pop_frame();
3130 __ restore_LR(R0);
3131
3132 // stack: (top interpreter frame, ..., optional interpreter frame,
3133 // optional c2i, caller of deoptee, ...).
3134
3135 // Initialize R14_state.
3136 __ restore_interpreter_state(R11_scratch1);
3137 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3138
3139 // Return to the interpreter entry point.
3140 __ blr();
3141 __ flush();
3142 #else // COMPILER2
3143 __ unimplemented("deopt blob needed only with compiler");
3144 int exception_offset = __ pc() - start;
3145 #endif // COMPILER2
3146
3147 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset,
3148 reexecute_offset, first_frame_size_in_bytes / wordSize);
3149 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3150 }
3151
3152 #ifdef COMPILER2
3153 UncommonTrapBlob* OptoRuntime::generate_uncommon_trap_blob() {
3154 // Allocate space for the code.
3155 ResourceMark rm;
3156 // Setup code generation tools.
3157 const char* name = OptoRuntime::stub_name(StubId::c2_uncommon_trap_id);
3158 CodeBuffer buffer(name, 2048, 1024);
3159 if (buffer.blob() == nullptr) {
3160 return nullptr;
3161 }
3162 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
3163 address start = __ pc();
3164
3165 Register unroll_block_reg = R21_tmp1;
3166 Register klass_index_reg = R22_tmp2;
3167 Register unc_trap_reg = R23_tmp3;
3168 Register r_return_pc = R27_tmp7;
3169
3170 OopMapSet* oop_maps = new OopMapSet();
3171 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3172 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3173
3174 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3175
3176 // Push a dummy `unpack_frame' and call
3177 // `Deoptimization::uncommon_trap' to pack the compiled frame into a
3178 // vframe array and return the `UnrollBlock' information.
3179
3180 // Save LR to compiled frame.
3181 __ save_LR(R11_scratch1);
3182
3183 // Push an "uncommon_trap" frame.
3184 __ push_frame_reg_args(0, R11_scratch1);
3185
3186 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...).
3187
3188 // Set the `unpack_frame' as last_Java_frame.
3189 // `Deoptimization::uncommon_trap' expects it and considers its
3190 // sender frame as the deoptee frame.
3191 // Remember the offset of the instruction whose address will be
3192 // moved to R11_scratch1.
3193 address gc_map_pc = __ pc();
3194 __ calculate_address_from_global_toc(r_return_pc, gc_map_pc, true, true, true, true);
3195 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3196
3197 __ mr(klass_index_reg, R3);
3198 __ li(R5_ARG3, Deoptimization::Unpack_uncommon_trap);
3199 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap),
3200 R16_thread, klass_index_reg, R5_ARG3);
3201
3202 // Set an oopmap for the call site.
3203 oop_maps->add_gc_map(gc_map_pc - start, map);
3204
3205 __ reset_last_Java_frame();
3206
3207 // Pop the `unpack frame'.
3208 __ pop_frame();
3209
3210 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3211
3212 // Save the return value.
3213 __ mr(unroll_block_reg, R3_RET);
3214
3215 // Pop the uncommon_trap frame.
3216 __ pop_frame();
3217
3218 // stack: (caller_of_deoptee, ...).
3219
3220 #ifdef ASSERT
3221 __ lwz(R22_tmp2, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3222 __ cmpdi(CR0, R22_tmp2, (unsigned)Deoptimization::Unpack_uncommon_trap);
3223 __ asm_assert_eq("OptoRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap");
3224 #endif
3225
3226 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3227 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3228 // and the frame is effectively not resized.
3229 Register caller_sp = R23_tmp3;
3230 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3231 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3232
3233 // Allocate new interpreter frame(s) and possibly a c2i adapter
3234 // frame.
3235 push_skeleton_frames(masm, false/*deopt*/,
3236 unroll_block_reg,
3237 R22_tmp2,
3238 R23_tmp3,
3239 R24_tmp4,
3240 R25_tmp5,
3241 R26_tmp6);
3242
3243 // stack: (skeletal interpreter frame, ..., optional skeletal
3244 // interpreter frame, optional c2i, caller of deoptee, ...).
3245
3246 // Push a dummy `unpack_frame' taking care of float return values.
3247 // Call `Deoptimization::unpack_frames' to layout information in the
3248 // interpreter frames just created.
3249
3250 // Push a simple "unpack frame" here.
3251 __ push_frame_reg_args(0, R11_scratch1);
3252
3253 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3254 // skeletal interpreter frame, optional c2i, caller of deoptee,
3255 // ...).
3256
3257 // Set the "unpack_frame" as last_Java_frame.
3258 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3259
3260 // Indicate it is the uncommon trap case.
3261 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap);
3262 // Let the unpacker layout information in the skeletal frames just
3263 // allocated.
3264 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3265 R16_thread, unc_trap_reg);
3266
3267 __ reset_last_Java_frame();
3268 // Pop the `unpack frame'.
3269 __ pop_frame();
3270 // Restore LR from top interpreter frame.
3271 __ restore_LR(R11_scratch1);
3272
3273 // stack: (top interpreter frame, ..., optional interpreter frame,
3274 // optional c2i, caller of deoptee, ...).
3275
3276 __ restore_interpreter_state(R11_scratch1);
3277 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3278
3279 // Return to the interpreter entry point.
3280 __ blr();
3281
3282 masm->flush();
3283
3284 return UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize);
3285 }
3286 #endif // COMPILER2
3287
3288 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap.
3289 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
3290 assert(StubRoutines::forward_exception_entry() != nullptr,
3291 "must be generated before");
3292 assert(is_polling_page_id(id), "expected a polling page stub id");
3293
3294 ResourceMark rm;
3295 OopMapSet *oop_maps = new OopMapSet();
3296 OopMap* map;
3297
3298 // Allocate space for the code. Setup code generation tools.
3299 const char* name = SharedRuntime::stub_name(id);
3300 CodeBuffer buffer(name, 2048, 1024);
3301 MacroAssembler* masm = new MacroAssembler(&buffer);
3302
3303 address start = __ pc();
3304 int frame_size_in_bytes = 0;
3305
3306 RegisterSaver::ReturnPCLocation return_pc_location;
3307 bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
3308 if (cause_return) {
3309 // Nothing to do here. The frame has already been popped in MachEpilogNode.
3310 // Register LR already contains the return pc.
3311 return_pc_location = RegisterSaver::return_pc_is_pre_saved;
3312 } else {
3313 // Use thread()->saved_exception_pc() as return pc.
3314 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc;
3315 }
3316
3317 bool save_vectors = (id == StubId::shared_polling_page_vectors_safepoint_handler_id);
3318
3319 // Save registers, fpu state, and flags. Set R31 = return pc.
3320 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3321 &frame_size_in_bytes,
3322 /*generate_oop_map=*/ true,
3323 /*return_pc_adjustment=*/0,
3324 return_pc_location, save_vectors);
3325
3326 // The following is basically a call_VM. However, we need the precise
3327 // address of the call in order to generate an oopmap. Hence, we do all the
3328 // work ourselves.
3329 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg);
3330
3331 // The return address must always be correct so that the frame constructor
3332 // never sees an invalid pc.
3333
3334 // Do the call
3335 __ call_VM_leaf(call_ptr, R16_thread);
3336 address calls_return_pc = __ last_calls_return_pc();
3337
3338 // Set an oopmap for the call site. This oopmap will map all
3339 // oop-registers and debug-info registers as callee-saved. This
3340 // will allow deoptimization at this safepoint to find all possible
3341 // debug-info recordings, as well as let GC find all oops.
3342 oop_maps->add_gc_map(calls_return_pc - start, map);
3343
3344 Label noException;
3345
3346 // Clear the last Java frame.
3347 __ reset_last_Java_frame();
3348
3349 BLOCK_COMMENT(" Check pending exception.");
3350 const Register pending_exception = R0;
3351 __ ld(pending_exception, thread_(pending_exception));
3352 __ cmpdi(CR0, pending_exception, 0);
3353 __ beq(CR0, noException);
3354
3355 // Exception pending
3356 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3357 frame_size_in_bytes,
3358 /*restore_ctr=*/true, save_vectors);
3359
3360 BLOCK_COMMENT(" Jump to forward_exception_entry.");
3361 // Jump to forward_exception_entry, with the issuing PC in LR
3362 // so it looks like the original nmethod called forward_exception_entry.
3363 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3364
3365 // No exception case.
3366 __ BIND(noException);
3367
3368 if (!cause_return) {
3369 Label no_adjust;
3370 // If our stashed return pc was modified by the runtime we avoid touching it
3371 __ ld(R0, frame_size_in_bytes + _abi0(lr), R1_SP);
3372 __ cmpd(CR0, R0, R31);
3373 __ bne(CR0, no_adjust);
3374
3375 // Adjust return pc forward to step over the safepoint poll instruction
3376 __ addi(R31, R31, 4);
3377 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
3378
3379 __ bind(no_adjust);
3380 }
3381
3382 // Normal exit, restore registers and exit.
3383 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3384 frame_size_in_bytes,
3385 /*restore_ctr=*/true, save_vectors);
3386
3387 __ blr();
3388
3389 // Make sure all code is generated
3390 masm->flush();
3391
3392 // Fill-out other meta info
3393 // CodeBlob frame size is in words.
3394 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize);
3395 }
3396
3397 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss)
3398 //
3399 // Generate a stub that calls into the vm to find out the proper destination
3400 // of a java call. All the argument registers are live at this point
3401 // but since this is generic code we don't know what they are and the caller
3402 // must do any gc of the args.
3403 //
3404 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
3405 assert(is_resolve_id(id), "expected a resolve stub id");
3406
3407 // allocate space for the code
3408 ResourceMark rm;
3409
3410 const char* name = SharedRuntime::stub_name(id);
3411 CodeBuffer buffer(name, 1000, 512);
3412 MacroAssembler* masm = new MacroAssembler(&buffer);
3413
3414 int frame_size_in_bytes;
3415
3416 OopMapSet *oop_maps = new OopMapSet();
3417 OopMap* map = nullptr;
3418
3419 address start = __ pc();
3420
3421 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3422 &frame_size_in_bytes,
3423 /*generate_oop_map*/ true,
3424 /*return_pc_adjustment*/ 0,
3425 RegisterSaver::return_pc_is_lr);
3426
3427 // Use noreg as last_Java_pc, the return pc will be reconstructed
3428 // from the physical frame.
3429 __ set_last_Java_frame(/*sp*/R1_SP, noreg);
3430
3431 int frame_complete = __ offset();
3432
3433 // Pass R19_method as 2nd (optional) argument, used by
3434 // counter_overflow_stub.
3435 __ call_VM_leaf(destination, R16_thread, R19_method);
3436 address calls_return_pc = __ last_calls_return_pc();
3437 // Set an oopmap for the call site.
3438 // We need this not only for callee-saved registers, but also for volatile
3439 // registers that the compiler might be keeping live across a safepoint.
3440 // Create the oopmap for the call's return pc.
3441 oop_maps->add_gc_map(calls_return_pc - start, map);
3442
3443 // R3_RET contains the address we are going to jump to assuming no exception got installed.
3444
3445 // clear last_Java_sp
3446 __ reset_last_Java_frame();
3447
3448 // Check for pending exceptions.
3449 BLOCK_COMMENT("Check for pending exceptions.");
3450 Label pending;
3451 __ ld(R11_scratch1, thread_(pending_exception));
3452 __ cmpdi(CR0, R11_scratch1, 0);
3453 __ bne(CR0, pending);
3454
3455 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame.
3456
3457 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false);
3458
3459 // Get the returned method.
3460 __ get_vm_result_metadata(R19_method);
3461
3462 __ bctr();
3463
3464
3465 // Pending exception after the safepoint.
3466 __ BIND(pending);
3467
3468 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true);
3469
3470 // exception pending => remove activation and forward to exception handler
3471
3472 __ li(R11_scratch1, 0);
3473 __ ld(R3_ARG1, thread_(pending_exception));
3474 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_oop_offset()), R16_thread);
3475 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3476
3477 // -------------
3478 // Make sure all code is generated.
3479 masm->flush();
3480
3481 // return the blob
3482 // frame_size_words or bytes??
3483 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize,
3484 oop_maps, true);
3485 }
3486
3487 // Continuation point for throwing of implicit exceptions that are
3488 // not handled in the current activation. Fabricates an exception
3489 // oop and initiates normal exception dispatching in this
3490 // frame. Only callee-saved registers are preserved (through the
3491 // normal register window / RegisterMap handling). If the compiler
3492 // needs all registers to be preserved between the fault point and
3493 // the exception handler then it must assume responsibility for that
3494 // in AbstractCompiler::continuation_for_implicit_null_exception or
3495 // continuation_for_implicit_division_by_zero_exception. All other
3496 // implicit exceptions (e.g., NullPointerException or
3497 // AbstractMethodError on entry) are either at call sites or
3498 // otherwise assume that stack unwinding will be initiated, so
3499 // caller saved registers were assumed volatile in the compiler.
3500 //
3501 // Note that we generate only this stub into a RuntimeStub, because
3502 // it needs to be properly traversed and ignored during GC, so we
3503 // change the meaning of the "__" macro within this method.
3504 //
3505 // Note: the routine set_pc_not_at_call_for_caller in
3506 // SharedRuntime.cpp requires that this code be generated into a
3507 // RuntimeStub.
3508 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
3509 assert(is_throw_id(id), "expected a throw stub id");
3510
3511 const char* name = SharedRuntime::stub_name(id);
3512
3513 ResourceMark rm;
3514 const char* timer_msg = "SharedRuntime generate_throw_exception";
3515 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3516
3517 CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0);
3518 MacroAssembler* masm = new MacroAssembler(&code);
3519
3520 OopMapSet* oop_maps = new OopMapSet();
3521 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3522 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3523
3524 address start = __ pc();
3525
3526 __ save_LR(R11_scratch1);
3527
3528 // Push a frame.
3529 __ push_frame_reg_args(0, R11_scratch1);
3530
3531 address frame_complete_pc = __ pc();
3532
3533 // Note that we always have a runtime stub frame on the top of
3534 // stack by this point. Remember the offset of the instruction
3535 // whose address will be moved to R11_scratch1.
3536 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1);
3537
3538 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
3539
3540 __ mr(R3_ARG1, R16_thread);
3541 __ call_c(runtime_entry);
3542
3543 // Set an oopmap for the call site.
3544 oop_maps->add_gc_map((int)(gc_map_pc - start), map);
3545
3546 __ reset_last_Java_frame();
3547
3548 #ifdef ASSERT
3549 // Make sure that this code is only executed if there is a pending
3550 // exception.
3551 {
3552 Label L;
3553 __ ld(R0,
3554 in_bytes(Thread::pending_exception_offset()),
3555 R16_thread);
3556 __ cmpdi(CR0, R0, 0);
3557 __ bne(CR0, L);
3558 __ stop("SharedRuntime::throw_exception: no pending exception");
3559 __ bind(L);
3560 }
3561 #endif
3562
3563 // Pop frame.
3564 __ pop_frame();
3565
3566 __ restore_LR(R11_scratch1);
3567
3568 __ load_const(R11_scratch1, StubRoutines::forward_exception_entry());
3569 __ mtctr(R11_scratch1);
3570 __ bctr();
3571
3572 // Create runtime stub with OopMap.
3573 RuntimeStub* stub =
3574 RuntimeStub::new_runtime_stub(name, &code,
3575 /*frame_complete=*/ (int)(frame_complete_pc - start),
3576 frame_size_in_bytes/wordSize,
3577 oop_maps,
3578 false);
3579 return stub;
3580 }
3581
3582 //------------------------------Montgomery multiplication------------------------
3583 //
3584
3585 // Subtract 0:b from carry:a. Return carry.
3586 static unsigned long
3587 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) {
3588 long i = 0;
3589 unsigned long tmp, tmp2;
3590 __asm__ __volatile__ (
3591 "subfc %[tmp], %[tmp], %[tmp] \n" // pre-set CA
3592 "mtctr %[len] \n"
3593 "0: \n"
3594 "ldx %[tmp], %[i], %[a] \n"
3595 "ldx %[tmp2], %[i], %[b] \n"
3596 "subfe %[tmp], %[tmp2], %[tmp] \n" // subtract extended
3597 "stdx %[tmp], %[i], %[a] \n"
3598 "addi %[i], %[i], 8 \n"
3599 "bdnz 0b \n"
3600 "addme %[tmp], %[carry] \n" // carry + CA - 1
3601 : [i]"+b"(i), [tmp]"=&r"(tmp), [tmp2]"=&r"(tmp2)
3602 : [a]"r"(a), [b]"r"(b), [carry]"r"(carry), [len]"r"(len)
3603 : "ctr", "xer", "memory"
3604 );
3605 return tmp;
3606 }
3607
3608 // Multiply (unsigned) Long A by Long B, accumulating the double-
3609 // length result into the accumulator formed of T0, T1, and T2.
3610 inline void MACC(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3611 unsigned long hi, lo;
3612 __asm__ __volatile__ (
3613 "mulld %[lo], %[A], %[B] \n"
3614 "mulhdu %[hi], %[A], %[B] \n"
3615 "addc %[T0], %[T0], %[lo] \n"
3616 "adde %[T1], %[T1], %[hi] \n"
3617 "addze %[T2], %[T2] \n"
3618 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3619 : [A]"r"(A), [B]"r"(B)
3620 : "xer"
3621 );
3622 }
3623
3624 // As above, but add twice the double-length result into the
3625 // accumulator.
3626 inline void MACC2(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3627 unsigned long hi, lo;
3628 __asm__ __volatile__ (
3629 "mulld %[lo], %[A], %[B] \n"
3630 "mulhdu %[hi], %[A], %[B] \n"
3631 "addc %[T0], %[T0], %[lo] \n"
3632 "adde %[T1], %[T1], %[hi] \n"
3633 "addze %[T2], %[T2] \n"
3634 "addc %[T0], %[T0], %[lo] \n"
3635 "adde %[T1], %[T1], %[hi] \n"
3636 "addze %[T2], %[T2] \n"
3637 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3638 : [A]"r"(A), [B]"r"(B)
3639 : "xer"
3640 );
3641 }
3642
3643 // Fast Montgomery multiplication. The derivation of the algorithm is
3644 // in "A Cryptographic Library for the Motorola DSP56000,
3645 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237".
3646 static void
3647 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[],
3648 unsigned long m[], unsigned long inv, int len) {
3649 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3650 int i;
3651
3652 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3653
3654 for (i = 0; i < len; i++) {
3655 int j;
3656 for (j = 0; j < i; j++) {
3657 MACC(a[j], b[i-j], t0, t1, t2);
3658 MACC(m[j], n[i-j], t0, t1, t2);
3659 }
3660 MACC(a[i], b[0], t0, t1, t2);
3661 m[i] = t0 * inv;
3662 MACC(m[i], n[0], t0, t1, t2);
3663
3664 assert(t0 == 0, "broken Montgomery multiply");
3665
3666 t0 = t1; t1 = t2; t2 = 0;
3667 }
3668
3669 for (i = len; i < 2*len; i++) {
3670 int j;
3671 for (j = i-len+1; j < len; j++) {
3672 MACC(a[j], b[i-j], t0, t1, t2);
3673 MACC(m[j], n[i-j], t0, t1, t2);
3674 }
3675 m[i-len] = t0;
3676 t0 = t1; t1 = t2; t2 = 0;
3677 }
3678
3679 while (t0) {
3680 t0 = sub(m, n, t0, len);
3681 }
3682 }
3683
3684 // Fast Montgomery squaring. This uses asymptotically 25% fewer
3685 // multiplies so it should be up to 25% faster than Montgomery
3686 // multiplication. However, its loop control is more complex and it
3687 // may actually run slower on some machines.
3688 static void
3689 montgomery_square(unsigned long a[], unsigned long n[],
3690 unsigned long m[], unsigned long inv, int len) {
3691 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3692 int i;
3693
3694 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3695
3696 for (i = 0; i < len; i++) {
3697 int j;
3698 int end = (i+1)/2;
3699 for (j = 0; j < end; j++) {
3700 MACC2(a[j], a[i-j], t0, t1, t2);
3701 MACC(m[j], n[i-j], t0, t1, t2);
3702 }
3703 if ((i & 1) == 0) {
3704 MACC(a[j], a[j], t0, t1, t2);
3705 }
3706 for (; j < i; j++) {
3707 MACC(m[j], n[i-j], t0, t1, t2);
3708 }
3709 m[i] = t0 * inv;
3710 MACC(m[i], n[0], t0, t1, t2);
3711
3712 assert(t0 == 0, "broken Montgomery square");
3713
3714 t0 = t1; t1 = t2; t2 = 0;
3715 }
3716
3717 for (i = len; i < 2*len; i++) {
3718 int start = i-len+1;
3719 int end = start + (len - start)/2;
3720 int j;
3721 for (j = start; j < end; j++) {
3722 MACC2(a[j], a[i-j], t0, t1, t2);
3723 MACC(m[j], n[i-j], t0, t1, t2);
3724 }
3725 if ((i & 1) == 0) {
3726 MACC(a[j], a[j], t0, t1, t2);
3727 }
3728 for (; j < len; j++) {
3729 MACC(m[j], n[i-j], t0, t1, t2);
3730 }
3731 m[i-len] = t0;
3732 t0 = t1; t1 = t2; t2 = 0;
3733 }
3734
3735 while (t0) {
3736 t0 = sub(m, n, t0, len);
3737 }
3738 }
3739
3740 // The threshold at which squaring is advantageous was determined
3741 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz.
3742 // Doesn't seem to be relevant for Power8 so we use the same value.
3743 #define MONTGOMERY_SQUARING_THRESHOLD 64
3744
3745 // Copy len longwords from s to d, word-swapping as we go. The
3746 // destination array is reversed.
3747 static void reverse_words(unsigned long *s, unsigned long *d, int len) {
3748 d += len;
3749 while(len-- > 0) {
3750 d--;
3751 unsigned long s_val = *s;
3752 // Swap words in a longword on little endian machines.
3753 #ifdef VM_LITTLE_ENDIAN
3754 s_val = (s_val << 32) | (s_val >> 32);
3755 #endif
3756 *d = s_val;
3757 s++;
3758 }
3759 }
3760
3761 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints,
3762 jint len, jlong inv,
3763 jint *m_ints) {
3764 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3765 assert(len % 2 == 0, "array length in montgomery_multiply must be even");
3766 int longwords = len/2;
3767
3768 // Make very sure we don't use so much space that the stack might
3769 // overflow. 512 jints corresponds to an 16384-bit integer and
3770 // will use here a total of 8k bytes of stack space.
3771 int divisor = sizeof(unsigned long) * 4;
3772 guarantee(longwords <= 8192 / divisor, "must be");
3773 int total_allocation = longwords * sizeof (unsigned long) * 4;
3774 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3775
3776 // Local scratch arrays
3777 unsigned long
3778 *a = scratch + 0 * longwords,
3779 *b = scratch + 1 * longwords,
3780 *n = scratch + 2 * longwords,
3781 *m = scratch + 3 * longwords;
3782
3783 reverse_words((unsigned long *)a_ints, a, longwords);
3784 reverse_words((unsigned long *)b_ints, b, longwords);
3785 reverse_words((unsigned long *)n_ints, n, longwords);
3786
3787 ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords);
3788
3789 reverse_words(m, (unsigned long *)m_ints, longwords);
3790 }
3791
3792 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints,
3793 jint len, jlong inv,
3794 jint *m_ints) {
3795 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3796 assert(len % 2 == 0, "array length in montgomery_square must be even");
3797 int longwords = len/2;
3798
3799 // Make very sure we don't use so much space that the stack might
3800 // overflow. 512 jints corresponds to an 16384-bit integer and
3801 // will use here a total of 6k bytes of stack space.
3802 int divisor = sizeof(unsigned long) * 3;
3803 guarantee(longwords <= (8192 / divisor), "must be");
3804 int total_allocation = longwords * sizeof (unsigned long) * 3;
3805 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3806
3807 // Local scratch arrays
3808 unsigned long
3809 *a = scratch + 0 * longwords,
3810 *n = scratch + 1 * longwords,
3811 *m = scratch + 2 * longwords;
3812
3813 reverse_words((unsigned long *)a_ints, a, longwords);
3814 reverse_words((unsigned long *)n_ints, n, longwords);
3815
3816 if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3817 ::montgomery_square(a, n, m, (unsigned long)inv, longwords);
3818 } else {
3819 ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords);
3820 }
3821
3822 reverse_words(m, (unsigned long *)m_ints, longwords);
3823 }
3824
3825 #if INCLUDE_JFR
3826
3827 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3828 // It returns a jobject handle to the event writer.
3829 // The handle is dereferenced and the return value is the event writer oop.
3830 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3831 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
3832 CodeBuffer code(name, 512, 64);
3833 MacroAssembler* masm = new MacroAssembler(&code);
3834
3835 Register tmp1 = R10_ARG8;
3836 Register tmp2 = R9_ARG7;
3837
3838 int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3839 address start = __ pc();
3840 __ mflr(tmp1);
3841 __ std(tmp1, _abi0(lr), R1_SP); // save return pc
3842 __ push_frame_reg_args(0, tmp1);
3843 int frame_complete = __ pc() - start;
3844 __ set_last_Java_frame(R1_SP, noreg);
3845 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), R16_thread);
3846 address calls_return_pc = __ last_calls_return_pc();
3847 __ reset_last_Java_frame();
3848 // The handle is dereferenced through a load barrier.
3849 __ resolve_global_jobject(R3_RET, tmp1, tmp2, MacroAssembler::PRESERVATION_NONE);
3850 __ pop_frame();
3851 __ ld(tmp1, _abi0(lr), R1_SP);
3852 __ mtlr(tmp1);
3853 __ blr();
3854
3855 OopMapSet* oop_maps = new OopMapSet();
3856 OopMap* map = new OopMap(framesize, 0);
3857 oop_maps->add_gc_map(calls_return_pc - start, map);
3858
3859 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3860 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3861 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3862 oop_maps, false);
3863 return stub;
3864 }
3865
3866 // For c2: call to return a leased buffer.
3867 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
3868 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
3869 CodeBuffer code(name, 512, 64);
3870 MacroAssembler* masm = new MacroAssembler(&code);
3871
3872 Register tmp1 = R10_ARG8;
3873 Register tmp2 = R9_ARG7;
3874
3875 int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3876 address start = __ pc();
3877 __ mflr(tmp1);
3878 __ std(tmp1, _abi0(lr), R1_SP); // save return pc
3879 __ push_frame_reg_args(0, tmp1);
3880 int frame_complete = __ pc() - start;
3881 __ set_last_Java_frame(R1_SP, noreg);
3882 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), R16_thread);
3883 address calls_return_pc = __ last_calls_return_pc();
3884 __ reset_last_Java_frame();
3885 __ pop_frame();
3886 __ ld(tmp1, _abi0(lr), R1_SP);
3887 __ mtlr(tmp1);
3888 __ blr();
3889
3890 OopMapSet* oop_maps = new OopMapSet();
3891 OopMap* map = new OopMap(framesize, 0);
3892 oop_maps->add_gc_map(calls_return_pc - start, map);
3893
3894 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3895 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3896 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3897 oop_maps, false);
3898 return stub;
3899 }
3900
3901 #endif // INCLUDE_JFR