1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2012, 2024 SAP SE. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/compiledIC.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "frame_ppc.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/shared/gcLocker.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "interpreter/interp_masm.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/klass.inline.hpp"
38 #include "prims/methodHandles.hpp"
39 #include "runtime/continuation.hpp"
40 #include "runtime/continuationEntry.inline.hpp"
41 #include "runtime/jniHandles.hpp"
42 #include "runtime/os.inline.hpp"
43 #include "runtime/safepointMechanism.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "runtime/signature.hpp"
46 #include "runtime/stubRoutines.hpp"
47 #include "runtime/timerTrace.hpp"
48 #include "runtime/vframeArray.hpp"
49 #include "utilities/align.hpp"
50 #include "utilities/macros.hpp"
51 #include "vmreg_ppc.inline.hpp"
52 #ifdef COMPILER1
53 #include "c1/c1_Runtime1.hpp"
54 #endif
55 #ifdef COMPILER2
56 #include "opto/ad.hpp"
57 #include "opto/runtime.hpp"
58 #endif
59
60 #include <alloca.h>
61
62 #define __ masm->
63
64 #ifdef PRODUCT
65 #define BLOCK_COMMENT(str) // nothing
66 #else
67 #define BLOCK_COMMENT(str) __ block_comment(str)
68 #endif
69
70 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
71
72
73 class RegisterSaver {
74 // Used for saving volatile registers.
75 public:
76
77 // Support different return pc locations.
78 enum ReturnPCLocation {
79 return_pc_is_lr,
80 return_pc_is_pre_saved,
81 return_pc_is_thread_saved_exception_pc
82 };
83
84 static OopMap* push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
85 int* out_frame_size_in_bytes,
86 bool generate_oop_map,
87 int return_pc_adjustment,
88 ReturnPCLocation return_pc_location,
89 bool save_vectors = false);
90 static void restore_live_registers_and_pop_frame(MacroAssembler* masm,
91 int frame_size_in_bytes,
92 bool restore_ctr,
93 bool save_vectors = false);
94
95 static void push_frame_and_save_argument_registers(MacroAssembler* masm,
96 Register r_temp,
97 int frame_size,
98 int total_args,
99 const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
100 static void restore_argument_registers_and_pop_frame(MacroAssembler*masm,
101 int frame_size,
102 int total_args,
103 const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
104
105 // During deoptimization only the result registers need to be restored
106 // all the other values have already been extracted.
107 static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes);
108
109 // Constants and data structures:
110
111 typedef enum {
112 int_reg,
113 float_reg,
114 special_reg,
115 vs_reg
116 } RegisterType;
117
118 typedef enum {
119 reg_size = 8,
120 half_reg_size = reg_size / 2,
121 vs_reg_size = 16
122 } RegisterConstants;
123
124 typedef struct {
125 RegisterType reg_type;
126 int reg_num;
127 VMReg vmreg;
128 } LiveRegType;
129 };
130
131
132 #define RegisterSaver_LiveIntReg(regname) \
133 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() }
134
135 #define RegisterSaver_LiveFloatReg(regname) \
136 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() }
137
138 #define RegisterSaver_LiveSpecialReg(regname) \
139 { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() }
140
141 #define RegisterSaver_LiveVSReg(regname) \
142 { RegisterSaver::vs_reg, regname->encoding(), regname->as_VMReg() }
143
144 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = {
145 // Live registers which get spilled to the stack. Register
146 // positions in this array correspond directly to the stack layout.
147
148 //
149 // live special registers:
150 //
151 RegisterSaver_LiveSpecialReg(SR_CTR),
152 //
153 // live float registers:
154 //
155 RegisterSaver_LiveFloatReg( F0 ),
156 RegisterSaver_LiveFloatReg( F1 ),
157 RegisterSaver_LiveFloatReg( F2 ),
158 RegisterSaver_LiveFloatReg( F3 ),
159 RegisterSaver_LiveFloatReg( F4 ),
160 RegisterSaver_LiveFloatReg( F5 ),
161 RegisterSaver_LiveFloatReg( F6 ),
162 RegisterSaver_LiveFloatReg( F7 ),
163 RegisterSaver_LiveFloatReg( F8 ),
164 RegisterSaver_LiveFloatReg( F9 ),
165 RegisterSaver_LiveFloatReg( F10 ),
166 RegisterSaver_LiveFloatReg( F11 ),
167 RegisterSaver_LiveFloatReg( F12 ),
168 RegisterSaver_LiveFloatReg( F13 ),
169 RegisterSaver_LiveFloatReg( F14 ),
170 RegisterSaver_LiveFloatReg( F15 ),
171 RegisterSaver_LiveFloatReg( F16 ),
172 RegisterSaver_LiveFloatReg( F17 ),
173 RegisterSaver_LiveFloatReg( F18 ),
174 RegisterSaver_LiveFloatReg( F19 ),
175 RegisterSaver_LiveFloatReg( F20 ),
176 RegisterSaver_LiveFloatReg( F21 ),
177 RegisterSaver_LiveFloatReg( F22 ),
178 RegisterSaver_LiveFloatReg( F23 ),
179 RegisterSaver_LiveFloatReg( F24 ),
180 RegisterSaver_LiveFloatReg( F25 ),
181 RegisterSaver_LiveFloatReg( F26 ),
182 RegisterSaver_LiveFloatReg( F27 ),
183 RegisterSaver_LiveFloatReg( F28 ),
184 RegisterSaver_LiveFloatReg( F29 ),
185 RegisterSaver_LiveFloatReg( F30 ),
186 RegisterSaver_LiveFloatReg( F31 ),
187 //
188 // live integer registers:
189 //
190 RegisterSaver_LiveIntReg( R0 ),
191 //RegisterSaver_LiveIntReg( R1 ), // stack pointer
192 RegisterSaver_LiveIntReg( R2 ),
193 RegisterSaver_LiveIntReg( R3 ),
194 RegisterSaver_LiveIntReg( R4 ),
195 RegisterSaver_LiveIntReg( R5 ),
196 RegisterSaver_LiveIntReg( R6 ),
197 RegisterSaver_LiveIntReg( R7 ),
198 RegisterSaver_LiveIntReg( R8 ),
199 RegisterSaver_LiveIntReg( R9 ),
200 RegisterSaver_LiveIntReg( R10 ),
201 RegisterSaver_LiveIntReg( R11 ),
202 RegisterSaver_LiveIntReg( R12 ),
203 //RegisterSaver_LiveIntReg( R13 ), // system thread id
204 RegisterSaver_LiveIntReg( R14 ),
205 RegisterSaver_LiveIntReg( R15 ),
206 RegisterSaver_LiveIntReg( R16 ),
207 RegisterSaver_LiveIntReg( R17 ),
208 RegisterSaver_LiveIntReg( R18 ),
209 RegisterSaver_LiveIntReg( R19 ),
210 RegisterSaver_LiveIntReg( R20 ),
211 RegisterSaver_LiveIntReg( R21 ),
212 RegisterSaver_LiveIntReg( R22 ),
213 RegisterSaver_LiveIntReg( R23 ),
214 RegisterSaver_LiveIntReg( R24 ),
215 RegisterSaver_LiveIntReg( R25 ),
216 RegisterSaver_LiveIntReg( R26 ),
217 RegisterSaver_LiveIntReg( R27 ),
218 RegisterSaver_LiveIntReg( R28 ),
219 RegisterSaver_LiveIntReg( R29 ),
220 RegisterSaver_LiveIntReg( R30 ),
221 RegisterSaver_LiveIntReg( R31 ) // must be the last register (see save/restore functions below)
222 };
223
224 static const RegisterSaver::LiveRegType RegisterSaver_LiveVSRegs[] = {
225 //
226 // live vector scalar registers (optional, only these ones are used by C2):
227 //
228 RegisterSaver_LiveVSReg( VSR32 ),
229 RegisterSaver_LiveVSReg( VSR33 ),
230 RegisterSaver_LiveVSReg( VSR34 ),
231 RegisterSaver_LiveVSReg( VSR35 ),
232 RegisterSaver_LiveVSReg( VSR36 ),
233 RegisterSaver_LiveVSReg( VSR37 ),
234 RegisterSaver_LiveVSReg( VSR38 ),
235 RegisterSaver_LiveVSReg( VSR39 ),
236 RegisterSaver_LiveVSReg( VSR40 ),
237 RegisterSaver_LiveVSReg( VSR41 ),
238 RegisterSaver_LiveVSReg( VSR42 ),
239 RegisterSaver_LiveVSReg( VSR43 ),
240 RegisterSaver_LiveVSReg( VSR44 ),
241 RegisterSaver_LiveVSReg( VSR45 ),
242 RegisterSaver_LiveVSReg( VSR46 ),
243 RegisterSaver_LiveVSReg( VSR47 ),
244 RegisterSaver_LiveVSReg( VSR48 ),
245 RegisterSaver_LiveVSReg( VSR49 ),
246 RegisterSaver_LiveVSReg( VSR50 ),
247 RegisterSaver_LiveVSReg( VSR51 )
248 };
249
250
251 OopMap* RegisterSaver::push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
252 int* out_frame_size_in_bytes,
253 bool generate_oop_map,
254 int return_pc_adjustment,
255 ReturnPCLocation return_pc_location,
256 bool save_vectors) {
257 // Push an abi_reg_args-frame and store all registers which may be live.
258 // If requested, create an OopMap: Record volatile registers as
259 // callee-save values in an OopMap so their save locations will be
260 // propagated to the RegisterMap of the caller frame during
261 // StackFrameStream construction (needed for deoptimization; see
262 // compiledVFrame::create_stack_value).
263 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment.
264 // Updated return pc is returned in R31 (if not return_pc_is_pre_saved).
265
266 // calculate frame size
267 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
268 sizeof(RegisterSaver::LiveRegType);
269 const int vsregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVSRegs) /
270 sizeof(RegisterSaver::LiveRegType))
271 : 0;
272 const int register_save_size = regstosave_num * reg_size + vsregstosave_num * vs_reg_size;
273 const int frame_size_in_bytes = align_up(register_save_size, frame::alignment_in_bytes)
274 + frame::native_abi_reg_args_size;
275
276 *out_frame_size_in_bytes = frame_size_in_bytes;
277 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
278 const int register_save_offset = frame_size_in_bytes - register_save_size;
279
280 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
281 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : nullptr;
282
283 BLOCK_COMMENT("push_frame_reg_args_and_save_live_registers {");
284
285 // push a new frame
286 __ push_frame(frame_size_in_bytes, noreg);
287
288 // Save some registers in the last (non-vector) slots of the new frame so we
289 // can use them as scratch regs or to determine the return pc.
290 __ std(R31, frame_size_in_bytes - reg_size - vsregstosave_num * vs_reg_size, R1_SP);
291 __ std(R30, frame_size_in_bytes - 2*reg_size - vsregstosave_num * vs_reg_size, R1_SP);
292
293 // save the flags
294 // Do the save_LR by hand and adjust the return pc if requested.
295 switch (return_pc_location) {
296 case return_pc_is_lr: __ mflr(R31); break;
297 case return_pc_is_pre_saved: assert(return_pc_adjustment == 0, "unsupported"); break;
298 case return_pc_is_thread_saved_exception_pc: __ ld(R31, thread_(saved_exception_pc)); break;
299 default: ShouldNotReachHere();
300 }
301 if (return_pc_location != return_pc_is_pre_saved) {
302 if (return_pc_adjustment != 0) {
303 __ addi(R31, R31, return_pc_adjustment);
304 }
305 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
306 }
307
308 // save all registers (ints and floats)
309 int offset = register_save_offset;
310
311 for (int i = 0; i < regstosave_num; i++) {
312 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
313 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
314
315 switch (reg_type) {
316 case RegisterSaver::int_reg: {
317 if (reg_num < 30) { // We spilled R30-31 right at the beginning.
318 __ std(as_Register(reg_num), offset, R1_SP);
319 }
320 break;
321 }
322 case RegisterSaver::float_reg: {
323 __ stfd(as_FloatRegister(reg_num), offset, R1_SP);
324 break;
325 }
326 case RegisterSaver::special_reg: {
327 if (reg_num == SR_CTR.encoding()) {
328 __ mfctr(R30);
329 __ std(R30, offset, R1_SP);
330 } else {
331 Unimplemented();
332 }
333 break;
334 }
335 default:
336 ShouldNotReachHere();
337 }
338
339 if (generate_oop_map) {
340 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
341 RegisterSaver_LiveRegs[i].vmreg);
342 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size)>>2),
343 RegisterSaver_LiveRegs[i].vmreg->next());
344 }
345 offset += reg_size;
346 }
347
348 for (int i = 0; i < vsregstosave_num; i++) {
349 int reg_num = RegisterSaver_LiveVSRegs[i].reg_num;
350 int reg_type = RegisterSaver_LiveVSRegs[i].reg_type;
351
352 __ li(R30, offset);
353 __ stxvd2x(as_VectorSRegister(reg_num), R30, R1_SP);
354
355 if (generate_oop_map) {
356 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
357 RegisterSaver_LiveVSRegs[i].vmreg);
358 }
359 offset += vs_reg_size;
360 }
361
362 assert(offset == frame_size_in_bytes, "consistency check");
363
364 BLOCK_COMMENT("} push_frame_reg_args_and_save_live_registers");
365
366 // And we're done.
367 return map;
368 }
369
370
371 // Pop the current frame and restore all the registers that we
372 // saved.
373 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm,
374 int frame_size_in_bytes,
375 bool restore_ctr,
376 bool save_vectors) {
377 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
378 sizeof(RegisterSaver::LiveRegType);
379 const int vsregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVSRegs) /
380 sizeof(RegisterSaver::LiveRegType))
381 : 0;
382 const int register_save_size = regstosave_num * reg_size + vsregstosave_num * vs_reg_size;
383
384 const int register_save_offset = frame_size_in_bytes - register_save_size;
385
386 BLOCK_COMMENT("restore_live_registers_and_pop_frame {");
387
388 // restore all registers (ints and floats)
389 int offset = register_save_offset;
390
391 for (int i = 0; i < regstosave_num; i++) {
392 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
393 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
394
395 switch (reg_type) {
396 case RegisterSaver::int_reg: {
397 if (reg_num != 31) // R31 restored at the end, it's the tmp reg!
398 __ ld(as_Register(reg_num), offset, R1_SP);
399 break;
400 }
401 case RegisterSaver::float_reg: {
402 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
403 break;
404 }
405 case RegisterSaver::special_reg: {
406 if (reg_num == SR_CTR.encoding()) {
407 if (restore_ctr) { // Nothing to do here if ctr already contains the next address.
408 __ ld(R31, offset, R1_SP);
409 __ mtctr(R31);
410 }
411 } else {
412 Unimplemented();
413 }
414 break;
415 }
416 default:
417 ShouldNotReachHere();
418 }
419 offset += reg_size;
420 }
421
422 for (int i = 0; i < vsregstosave_num; i++) {
423 int reg_num = RegisterSaver_LiveVSRegs[i].reg_num;
424 int reg_type = RegisterSaver_LiveVSRegs[i].reg_type;
425
426 __ li(R31, offset);
427 __ lxvd2x(as_VectorSRegister(reg_num), R31, R1_SP);
428
429 offset += vs_reg_size;
430 }
431
432 assert(offset == frame_size_in_bytes, "consistency check");
433
434 // restore link and the flags
435 __ ld(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
436 __ mtlr(R31);
437
438 // restore scratch register's value
439 __ ld(R31, frame_size_in_bytes - reg_size - vsregstosave_num * vs_reg_size, R1_SP);
440
441 // pop the frame
442 __ addi(R1_SP, R1_SP, frame_size_in_bytes);
443
444 BLOCK_COMMENT("} restore_live_registers_and_pop_frame");
445 }
446
447 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp,
448 int frame_size,int total_args, const VMRegPair *regs,
449 const VMRegPair *regs2) {
450 __ push_frame(frame_size, r_temp);
451 int st_off = frame_size - wordSize;
452 for (int i = 0; i < total_args; i++) {
453 VMReg r_1 = regs[i].first();
454 VMReg r_2 = regs[i].second();
455 if (!r_1->is_valid()) {
456 assert(!r_2->is_valid(), "");
457 continue;
458 }
459 if (r_1->is_Register()) {
460 Register r = r_1->as_Register();
461 __ std(r, st_off, R1_SP);
462 st_off -= wordSize;
463 } else if (r_1->is_FloatRegister()) {
464 FloatRegister f = r_1->as_FloatRegister();
465 __ stfd(f, st_off, R1_SP);
466 st_off -= wordSize;
467 }
468 }
469 if (regs2 != nullptr) {
470 for (int i = 0; i < total_args; i++) {
471 VMReg r_1 = regs2[i].first();
472 VMReg r_2 = regs2[i].second();
473 if (!r_1->is_valid()) {
474 assert(!r_2->is_valid(), "");
475 continue;
476 }
477 if (r_1->is_Register()) {
478 Register r = r_1->as_Register();
479 __ std(r, st_off, R1_SP);
480 st_off -= wordSize;
481 } else if (r_1->is_FloatRegister()) {
482 FloatRegister f = r_1->as_FloatRegister();
483 __ stfd(f, st_off, R1_SP);
484 st_off -= wordSize;
485 }
486 }
487 }
488 }
489
490 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size,
491 int total_args, const VMRegPair *regs,
492 const VMRegPair *regs2) {
493 int st_off = frame_size - wordSize;
494 for (int i = 0; i < total_args; i++) {
495 VMReg r_1 = regs[i].first();
496 VMReg r_2 = regs[i].second();
497 if (r_1->is_Register()) {
498 Register r = r_1->as_Register();
499 __ ld(r, st_off, R1_SP);
500 st_off -= wordSize;
501 } else if (r_1->is_FloatRegister()) {
502 FloatRegister f = r_1->as_FloatRegister();
503 __ lfd(f, st_off, R1_SP);
504 st_off -= wordSize;
505 }
506 }
507 if (regs2 != nullptr)
508 for (int i = 0; i < total_args; i++) {
509 VMReg r_1 = regs2[i].first();
510 VMReg r_2 = regs2[i].second();
511 if (r_1->is_Register()) {
512 Register r = r_1->as_Register();
513 __ ld(r, st_off, R1_SP);
514 st_off -= wordSize;
515 } else if (r_1->is_FloatRegister()) {
516 FloatRegister f = r_1->as_FloatRegister();
517 __ lfd(f, st_off, R1_SP);
518 st_off -= wordSize;
519 }
520 }
521 __ pop_frame();
522 }
523
524 // Restore the registers that might be holding a result.
525 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) {
526 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
527 sizeof(RegisterSaver::LiveRegType);
528 const int register_save_size = regstosave_num * reg_size; // VS registers not relevant here.
529 const int register_save_offset = frame_size_in_bytes - register_save_size;
530
531 // restore all result registers (ints and floats)
532 int offset = register_save_offset;
533 for (int i = 0; i < regstosave_num; i++) {
534 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
535 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
536 switch (reg_type) {
537 case RegisterSaver::int_reg: {
538 if (as_Register(reg_num)==R3_RET) // int result_reg
539 __ ld(as_Register(reg_num), offset, R1_SP);
540 break;
541 }
542 case RegisterSaver::float_reg: {
543 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg
544 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
545 break;
546 }
547 case RegisterSaver::special_reg: {
548 // Special registers don't hold a result.
549 break;
550 }
551 default:
552 ShouldNotReachHere();
553 }
554 offset += reg_size;
555 }
556
557 assert(offset == frame_size_in_bytes, "consistency check");
558 }
559
560 // Is vector's size (in bytes) bigger than a size saved by default?
561 bool SharedRuntime::is_wide_vector(int size) {
562 // Note, MaxVectorSize == 8/16 on PPC64.
563 assert(size <= (SuperwordUseVSX ? 16 : 8), "%d bytes vectors are not supported", size);
564 return size > 8;
565 }
566
567 static int reg2slot(VMReg r) {
568 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots();
569 }
570
571 static int reg2offset(VMReg r) {
572 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
573 }
574
575 // ---------------------------------------------------------------------------
576 // Read the array of BasicTypes from a signature, and compute where the
577 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
578 // quantities. Values less than VMRegImpl::stack0 are registers, those above
579 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
580 // as framesizes are fixed.
581 // VMRegImpl::stack0 refers to the first slot 0(sp).
582 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register
583 // up to Register::number_of_registers) are the 64-bit
584 // integer registers.
585
586 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
587 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
588 // units regardless of build. Of course for i486 there is no 64 bit build
589
590 // The Java calling convention is a "shifted" version of the C ABI.
591 // By skipping the first C ABI register we can call non-static jni methods
592 // with small numbers of arguments without having to shuffle the arguments
593 // at all. Since we control the java ABI we ought to at least get some
594 // advantage out of it.
595
596 const VMReg java_iarg_reg[8] = {
597 R3->as_VMReg(),
598 R4->as_VMReg(),
599 R5->as_VMReg(),
600 R6->as_VMReg(),
601 R7->as_VMReg(),
602 R8->as_VMReg(),
603 R9->as_VMReg(),
604 R10->as_VMReg()
605 };
606
607 const VMReg java_farg_reg[13] = {
608 F1->as_VMReg(),
609 F2->as_VMReg(),
610 F3->as_VMReg(),
611 F4->as_VMReg(),
612 F5->as_VMReg(),
613 F6->as_VMReg(),
614 F7->as_VMReg(),
615 F8->as_VMReg(),
616 F9->as_VMReg(),
617 F10->as_VMReg(),
618 F11->as_VMReg(),
619 F12->as_VMReg(),
620 F13->as_VMReg()
621 };
622
623 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]);
624 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]);
625
626 STATIC_ASSERT(num_java_iarg_registers == Argument::n_int_register_parameters_j);
627 STATIC_ASSERT(num_java_farg_registers == Argument::n_float_register_parameters_j);
628
629 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
630 VMRegPair *regs,
631 int total_args_passed) {
632 // C2c calling conventions for compiled-compiled calls.
633 // Put 8 ints/longs into registers _AND_ 13 float/doubles into
634 // registers _AND_ put the rest on the stack.
635
636 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats
637 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles
638
639 int i;
640 VMReg reg;
641 int stk = 0;
642 int ireg = 0;
643 int freg = 0;
644
645 // We put the first 8 arguments into registers and the rest on the
646 // stack, float arguments are already in their argument registers
647 // due to c2c calling conventions (see calling_convention).
648 for (int i = 0; i < total_args_passed; ++i) {
649 switch(sig_bt[i]) {
650 case T_BOOLEAN:
651 case T_CHAR:
652 case T_BYTE:
653 case T_SHORT:
654 case T_INT:
655 if (ireg < num_java_iarg_registers) {
656 // Put int/ptr in register
657 reg = java_iarg_reg[ireg];
658 ++ireg;
659 } else {
660 // Put int/ptr on stack.
661 reg = VMRegImpl::stack2reg(stk);
662 stk += inc_stk_for_intfloat;
663 }
664 regs[i].set1(reg);
665 break;
666 case T_LONG:
667 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
668 if (ireg < num_java_iarg_registers) {
669 // Put long in register.
670 reg = java_iarg_reg[ireg];
671 ++ireg;
672 } else {
673 // Put long on stack. They must be aligned to 2 slots.
674 if (stk & 0x1) ++stk;
675 reg = VMRegImpl::stack2reg(stk);
676 stk += inc_stk_for_longdouble;
677 }
678 regs[i].set2(reg);
679 break;
680 case T_OBJECT:
681 case T_ARRAY:
682 case T_ADDRESS:
683 if (ireg < num_java_iarg_registers) {
684 // Put ptr in register.
685 reg = java_iarg_reg[ireg];
686 ++ireg;
687 } else {
688 // Put ptr on stack. Objects must be aligned to 2 slots too,
689 // because "64-bit pointers record oop-ishness on 2 aligned
690 // adjacent registers." (see OopFlow::build_oop_map).
691 if (stk & 0x1) ++stk;
692 reg = VMRegImpl::stack2reg(stk);
693 stk += inc_stk_for_longdouble;
694 }
695 regs[i].set2(reg);
696 break;
697 case T_FLOAT:
698 if (freg < num_java_farg_registers) {
699 // Put float in register.
700 reg = java_farg_reg[freg];
701 ++freg;
702 } else {
703 // Put float on stack.
704 reg = VMRegImpl::stack2reg(stk);
705 stk += inc_stk_for_intfloat;
706 }
707 regs[i].set1(reg);
708 break;
709 case T_DOUBLE:
710 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
711 if (freg < num_java_farg_registers) {
712 // Put double in register.
713 reg = java_farg_reg[freg];
714 ++freg;
715 } else {
716 // Put double on stack. They must be aligned to 2 slots.
717 if (stk & 0x1) ++stk;
718 reg = VMRegImpl::stack2reg(stk);
719 stk += inc_stk_for_longdouble;
720 }
721 regs[i].set2(reg);
722 break;
723 case T_VOID:
724 // Do not count halves.
725 regs[i].set_bad();
726 break;
727 default:
728 ShouldNotReachHere();
729 }
730 }
731 return stk;
732 }
733
734 #if defined(COMPILER1) || defined(COMPILER2)
735 // Calling convention for calling C code.
736 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
737 VMRegPair *regs,
738 int total_args_passed) {
739 // Calling conventions for C runtime calls and calls to JNI native methods.
740 //
741 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8
742 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist
743 // the first 13 flt/dbl's in the first 13 fp regs but additionally
744 // copy flt/dbl to the stack if they are beyond the 8th argument.
745
746 const VMReg iarg_reg[8] = {
747 R3->as_VMReg(),
748 R4->as_VMReg(),
749 R5->as_VMReg(),
750 R6->as_VMReg(),
751 R7->as_VMReg(),
752 R8->as_VMReg(),
753 R9->as_VMReg(),
754 R10->as_VMReg()
755 };
756
757 const VMReg farg_reg[13] = {
758 F1->as_VMReg(),
759 F2->as_VMReg(),
760 F3->as_VMReg(),
761 F4->as_VMReg(),
762 F5->as_VMReg(),
763 F6->as_VMReg(),
764 F7->as_VMReg(),
765 F8->as_VMReg(),
766 F9->as_VMReg(),
767 F10->as_VMReg(),
768 F11->as_VMReg(),
769 F12->as_VMReg(),
770 F13->as_VMReg()
771 };
772
773 // Check calling conventions consistency.
774 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c &&
775 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c,
776 "consistency");
777
778 const int additional_frame_header_slots = ((frame::native_abi_minframe_size - frame::jit_out_preserve_size)
779 / VMRegImpl::stack_slot_size);
780 const int float_offset_in_slots = Argument::float_on_stack_offset_in_bytes_c / VMRegImpl::stack_slot_size;
781
782 VMReg reg;
783 int arg = 0;
784 int freg = 0;
785 bool stack_used = false;
786
787 for (int i = 0; i < total_args_passed; ++i, ++arg) {
788 // Each argument corresponds to a slot in the Parameter Save Area (if not omitted)
789 int stk = (arg * 2) + additional_frame_header_slots;
790
791 switch(sig_bt[i]) {
792 //
793 // If arguments 0-7 are integers, they are passed in integer registers.
794 // Argument i is placed in iarg_reg[i].
795 //
796 case T_BOOLEAN:
797 case T_CHAR:
798 case T_BYTE:
799 case T_SHORT:
800 case T_INT:
801 // We must cast ints to longs and use full 64 bit stack slots
802 // here. Thus fall through, handle as long.
803 case T_LONG:
804 case T_OBJECT:
805 case T_ARRAY:
806 case T_ADDRESS:
807 case T_METADATA:
808 // Oops are already boxed if required (JNI).
809 if (arg < Argument::n_int_register_parameters_c) {
810 reg = iarg_reg[arg];
811 } else {
812 reg = VMRegImpl::stack2reg(stk);
813 stack_used = true;
814 }
815 regs[i].set2(reg);
816 break;
817
818 //
819 // Floats are treated differently from int regs: The first 13 float arguments
820 // are passed in registers (not the float args among the first 13 args).
821 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed
822 // in farg_reg[j] if argument i is the j-th float argument of this call.
823 //
824 case T_FLOAT:
825 if (freg < Argument::n_float_register_parameters_c) {
826 // Put float in register ...
827 reg = farg_reg[freg];
828 ++freg;
829 } else {
830 // Put float on stack.
831 reg = VMRegImpl::stack2reg(stk + float_offset_in_slots);
832 stack_used = true;
833 }
834 regs[i].set1(reg);
835 break;
836 case T_DOUBLE:
837 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
838 if (freg < Argument::n_float_register_parameters_c) {
839 // Put double in register ...
840 reg = farg_reg[freg];
841 ++freg;
842 } else {
843 // Put double on stack.
844 reg = VMRegImpl::stack2reg(stk);
845 stack_used = true;
846 }
847 regs[i].set2(reg);
848 break;
849
850 case T_VOID:
851 // Do not count halves.
852 regs[i].set_bad();
853 --arg;
854 break;
855 default:
856 ShouldNotReachHere();
857 }
858 }
859
860 // Return size of the stack frame excluding the jit_out_preserve part in single-word slots.
861 #if defined(ABI_ELFv2)
862 assert(additional_frame_header_slots == 0, "ABIv2 shouldn't use extra slots");
863 // ABIv2 allows omitting the Parameter Save Area if the callee's prototype
864 // indicates that all parameters can be passed in registers.
865 return stack_used ? (arg * 2) : 0;
866 #else
867 // The Parameter Save Area needs to be at least 8 double-word slots for ABIv1.
868 // We have to add extra slots because ABIv1 uses a larger header.
869 return MAX2(arg, 8) * 2 + additional_frame_header_slots;
870 #endif
871 }
872 #endif // COMPILER2
873
874 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
875 uint num_bits,
876 uint total_args_passed) {
877 Unimplemented();
878 return 0;
879 }
880
881 static address gen_c2i_adapter(MacroAssembler *masm,
882 int total_args_passed,
883 int comp_args_on_stack,
884 const BasicType *sig_bt,
885 const VMRegPair *regs,
886 Label& call_interpreter,
887 const Register& ientry) {
888
889 address c2i_entrypoint;
890
891 const Register sender_SP = R21_sender_SP; // == R21_tmp1
892 const Register code = R22_tmp2;
893 //const Register ientry = R23_tmp3;
894 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 };
895 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
896 int value_regs_index = 0;
897
898 const Register return_pc = R27_tmp7;
899 const Register tmp = R28_tmp8;
900
901 assert_different_registers(sender_SP, code, ientry, return_pc, tmp);
902
903 // Adapter needs TOP_IJAVA_FRAME_ABI.
904 const int adapter_size = frame::top_ijava_frame_abi_size +
905 align_up(total_args_passed * wordSize, frame::alignment_in_bytes);
906
907 // regular (verified) c2i entry point
908 c2i_entrypoint = __ pc();
909
910 // Does compiled code exists? If yes, patch the caller's callsite.
911 __ ld(code, method_(code));
912 __ cmpdi(CCR0, code, 0);
913 __ ld(ientry, method_(interpreter_entry)); // preloaded
914 __ beq(CCR0, call_interpreter);
915
916
917 // Patch caller's callsite, method_(code) was not null which means that
918 // compiled code exists.
919 __ mflr(return_pc);
920 __ std(return_pc, _abi0(lr), R1_SP);
921 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs);
922
923 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc);
924
925 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs);
926 __ ld(return_pc, _abi0(lr), R1_SP);
927 __ ld(ientry, method_(interpreter_entry)); // preloaded
928 __ mtlr(return_pc);
929
930
931 // Call the interpreter.
932 __ BIND(call_interpreter);
933 __ mtctr(ientry);
934
935 // Get a copy of the current SP for loading caller's arguments.
936 __ mr(sender_SP, R1_SP);
937
938 // Add space for the adapter.
939 __ resize_frame(-adapter_size, R12_scratch2);
940
941 int st_off = adapter_size - wordSize;
942
943 // Write the args into the outgoing interpreter space.
944 for (int i = 0; i < total_args_passed; i++) {
945 VMReg r_1 = regs[i].first();
946 VMReg r_2 = regs[i].second();
947 if (!r_1->is_valid()) {
948 assert(!r_2->is_valid(), "");
949 continue;
950 }
951 if (r_1->is_stack()) {
952 Register tmp_reg = value_regs[value_regs_index];
953 value_regs_index = (value_regs_index + 1) % num_value_regs;
954 // The calling convention produces OptoRegs that ignore the out
955 // preserve area (JIT's ABI). We must account for it here.
956 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
957 if (!r_2->is_valid()) {
958 __ lwz(tmp_reg, ld_off, sender_SP);
959 } else {
960 __ ld(tmp_reg, ld_off, sender_SP);
961 }
962 // Pretend stack targets were loaded into tmp_reg.
963 r_1 = tmp_reg->as_VMReg();
964 }
965
966 if (r_1->is_Register()) {
967 Register r = r_1->as_Register();
968 if (!r_2->is_valid()) {
969 __ stw(r, st_off, R1_SP);
970 st_off-=wordSize;
971 } else {
972 // Longs are given 2 64-bit slots in the interpreter, but the
973 // data is passed in only 1 slot.
974 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
975 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
976 st_off-=wordSize;
977 }
978 __ std(r, st_off, R1_SP);
979 st_off-=wordSize;
980 }
981 } else {
982 assert(r_1->is_FloatRegister(), "");
983 FloatRegister f = r_1->as_FloatRegister();
984 if (!r_2->is_valid()) {
985 __ stfs(f, st_off, R1_SP);
986 st_off-=wordSize;
987 } else {
988 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the
989 // data is passed in only 1 slot.
990 // One of these should get known junk...
991 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
992 st_off-=wordSize;
993 __ stfd(f, st_off, R1_SP);
994 st_off-=wordSize;
995 }
996 }
997 }
998
999 // Jump to the interpreter just as if interpreter was doing it.
1000
1001 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
1002
1003 // load TOS
1004 __ addi(R15_esp, R1_SP, st_off);
1005
1006 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1.
1007 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register");
1008 __ bctr();
1009
1010 return c2i_entrypoint;
1011 }
1012
1013 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
1014 int total_args_passed,
1015 int comp_args_on_stack,
1016 const BasicType *sig_bt,
1017 const VMRegPair *regs) {
1018
1019 // Load method's entry-point from method.
1020 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
1021 __ mtctr(R12_scratch2);
1022
1023 // We will only enter here from an interpreted frame and never from after
1024 // passing thru a c2i. Azul allowed this but we do not. If we lose the
1025 // race and use a c2i we will remain interpreted for the race loser(s).
1026 // This removes all sorts of headaches on the x86 side and also eliminates
1027 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
1028
1029 // Note: r13 contains the senderSP on entry. We must preserve it since
1030 // we may do a i2c -> c2i transition if we lose a race where compiled
1031 // code goes non-entrant while we get args ready.
1032 // In addition we use r13 to locate all the interpreter args as
1033 // we must align the stack to 16 bytes on an i2c entry else we
1034 // lose alignment we expect in all compiled code and register
1035 // save code can segv when fxsave instructions find improperly
1036 // aligned stack pointer.
1037
1038 const Register ld_ptr = R15_esp;
1039 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 };
1040 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
1041 int value_regs_index = 0;
1042
1043 int ld_offset = total_args_passed*wordSize;
1044
1045 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
1046 // in registers, we will occasionally have no stack args.
1047 int comp_words_on_stack = 0;
1048 if (comp_args_on_stack) {
1049 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
1050 // registers are below. By subtracting stack0, we either get a negative
1051 // number (all values in registers) or the maximum stack slot accessed.
1052
1053 // Convert 4-byte c2 stack slots to words.
1054 comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1055 // Round up to miminum stack alignment, in wordSize.
1056 comp_words_on_stack = align_up(comp_words_on_stack, 2);
1057 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1);
1058 }
1059
1060 // Now generate the shuffle code. Pick up all register args and move the
1061 // rest through register value=Z_R12.
1062 BLOCK_COMMENT("Shuffle arguments");
1063 for (int i = 0; i < total_args_passed; i++) {
1064 if (sig_bt[i] == T_VOID) {
1065 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
1066 continue;
1067 }
1068
1069 // Pick up 0, 1 or 2 words from ld_ptr.
1070 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1071 "scrambled load targets?");
1072 VMReg r_1 = regs[i].first();
1073 VMReg r_2 = regs[i].second();
1074 if (!r_1->is_valid()) {
1075 assert(!r_2->is_valid(), "");
1076 continue;
1077 }
1078 if (r_1->is_FloatRegister()) {
1079 if (!r_2->is_valid()) {
1080 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr);
1081 ld_offset-=wordSize;
1082 } else {
1083 // Skip the unused interpreter slot.
1084 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr);
1085 ld_offset-=2*wordSize;
1086 }
1087 } else {
1088 Register r;
1089 if (r_1->is_stack()) {
1090 // Must do a memory to memory move thru "value".
1091 r = value_regs[value_regs_index];
1092 value_regs_index = (value_regs_index + 1) % num_value_regs;
1093 } else {
1094 r = r_1->as_Register();
1095 }
1096 if (!r_2->is_valid()) {
1097 // Not sure we need to do this but it shouldn't hurt.
1098 if (is_reference_type(sig_bt[i]) || sig_bt[i] == T_ADDRESS) {
1099 __ ld(r, ld_offset, ld_ptr);
1100 ld_offset-=wordSize;
1101 } else {
1102 __ lwz(r, ld_offset, ld_ptr);
1103 ld_offset-=wordSize;
1104 }
1105 } else {
1106 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
1107 // data is passed in only 1 slot.
1108 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
1109 ld_offset-=wordSize;
1110 }
1111 __ ld(r, ld_offset, ld_ptr);
1112 ld_offset-=wordSize;
1113 }
1114
1115 if (r_1->is_stack()) {
1116 // Now store value where the compiler expects it
1117 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size;
1118
1119 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN ||
1120 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) {
1121 __ stw(r, st_off, R1_SP);
1122 } else {
1123 __ std(r, st_off, R1_SP);
1124 }
1125 }
1126 }
1127 }
1128
1129 __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1130
1131 BLOCK_COMMENT("Store method");
1132 // Store method into thread->callee_target.
1133 // We might end up in handle_wrong_method if the callee is
1134 // deoptimized as we race thru here. If that happens we don't want
1135 // to take a safepoint because the caller frame will look
1136 // interpreted and arguments are now "compiled" so it is much better
1137 // to make this transition invisible to the stack walking
1138 // code. Unfortunately if we try and find the callee by normal means
1139 // a safepoint is possible. So we stash the desired callee in the
1140 // thread and the vm will find there should this case occur.
1141 __ std(R19_method, thread_(callee_target));
1142
1143 // Jump to the compiled code just as if compiled code was doing it.
1144 __ bctr();
1145 }
1146
1147 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1148 int total_args_passed,
1149 int comp_args_on_stack,
1150 const BasicType *sig_bt,
1151 const VMRegPair *regs,
1152 AdapterFingerPrint* fingerprint) {
1153 address i2c_entry;
1154 address c2i_unverified_entry;
1155 address c2i_entry;
1156
1157
1158 // entry: i2c
1159
1160 __ align(CodeEntryAlignment);
1161 i2c_entry = __ pc();
1162 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
1163
1164
1165 // entry: c2i unverified
1166
1167 __ align(CodeEntryAlignment);
1168 BLOCK_COMMENT("c2i unverified entry");
1169 c2i_unverified_entry = __ pc();
1170
1171 // inline_cache contains a CompiledICData
1172 const Register ic = R19_inline_cache_reg;
1173 const Register ic_klass = R11_scratch1;
1174 const Register receiver_klass = R12_scratch2;
1175 const Register code = R21_tmp1;
1176 const Register ientry = R23_tmp3;
1177
1178 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry);
1179 assert(R11_scratch1 == R11, "need prologue scratch register");
1180
1181 Label call_interpreter;
1182
1183 __ ic_check(4 /* end_alignment */);
1184 __ ld(R19_method, CompiledICData::speculated_method_offset(), ic);
1185 // Argument is valid and klass is as expected, continue.
1186
1187 __ ld(code, method_(code));
1188 __ cmpdi(CCR0, code, 0);
1189 __ ld(ientry, method_(interpreter_entry)); // preloaded
1190 __ beq_predict_taken(CCR0, call_interpreter);
1191
1192 // Branch to ic_miss_stub.
1193 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type);
1194
1195 // entry: c2i
1196
1197 c2i_entry = __ pc();
1198
1199 // Class initialization barrier for static methods
1200 address c2i_no_clinit_check_entry = nullptr;
1201 if (VM_Version::supports_fast_class_init_checks()) {
1202 Label L_skip_barrier;
1203
1204 { // Bypass the barrier for non-static methods
1205 __ lwz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1206 __ andi_(R0, R0, JVM_ACC_STATIC);
1207 __ beq(CCR0, L_skip_barrier); // non-static
1208 }
1209
1210 Register klass = R11_scratch1;
1211 __ load_method_holder(klass, R19_method);
1212 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
1213
1214 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
1215 __ mtctr(klass);
1216 __ bctr();
1217
1218 __ bind(L_skip_barrier);
1219 c2i_no_clinit_check_entry = __ pc();
1220 }
1221
1222 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1223 bs->c2i_entry_barrier(masm, /* tmp register*/ ic_klass, /* tmp register*/ receiver_klass, /* tmp register*/ code);
1224
1225 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry);
1226
1227 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry,
1228 c2i_no_clinit_check_entry);
1229 }
1230
1231 // An oop arg. Must pass a handle not the oop itself.
1232 static void object_move(MacroAssembler* masm,
1233 int frame_size_in_slots,
1234 OopMap* oop_map, int oop_handle_offset,
1235 bool is_receiver, int* receiver_offset,
1236 VMRegPair src, VMRegPair dst,
1237 Register r_caller_sp, Register r_temp_1, Register r_temp_2) {
1238 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)),
1239 "receiver has already been moved");
1240
1241 // We must pass a handle. First figure out the location we use as a handle.
1242
1243 if (src.first()->is_stack()) {
1244 // stack to stack or reg
1245
1246 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1247 Label skip;
1248 const int oop_slot_in_callers_frame = reg2slot(src.first());
1249
1250 guarantee(!is_receiver, "expecting receiver in register");
1251 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots));
1252
1253 __ addi(r_handle, r_caller_sp, reg2offset(src.first()));
1254 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp);
1255 __ cmpdi(CCR0, r_temp_2, 0);
1256 __ bne(CCR0, skip);
1257 // Use a null handle if oop is null.
1258 __ li(r_handle, 0);
1259 __ bind(skip);
1260
1261 if (dst.first()->is_stack()) {
1262 // stack to stack
1263 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1264 } else {
1265 // stack to reg
1266 // Nothing to do, r_handle is already the dst register.
1267 }
1268 } else {
1269 // reg to stack or reg
1270 const Register r_oop = src.first()->as_Register();
1271 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1272 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word
1273 + oop_handle_offset; // in slots
1274 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size;
1275 Label skip;
1276
1277 if (is_receiver) {
1278 *receiver_offset = oop_offset;
1279 }
1280 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot));
1281
1282 __ std( r_oop, oop_offset, R1_SP);
1283 __ addi(r_handle, R1_SP, oop_offset);
1284
1285 __ cmpdi(CCR0, r_oop, 0);
1286 __ bne(CCR0, skip);
1287 // Use a null handle if oop is null.
1288 __ li(r_handle, 0);
1289 __ bind(skip);
1290
1291 if (dst.first()->is_stack()) {
1292 // reg to stack
1293 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1294 } else {
1295 // reg to reg
1296 // Nothing to do, r_handle is already the dst register.
1297 }
1298 }
1299 }
1300
1301 static void int_move(MacroAssembler*masm,
1302 VMRegPair src, VMRegPair dst,
1303 Register r_caller_sp, Register r_temp) {
1304 assert(src.first()->is_valid(), "incoming must be int");
1305 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1306
1307 if (src.first()->is_stack()) {
1308 if (dst.first()->is_stack()) {
1309 // stack to stack
1310 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp);
1311 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1312 } else {
1313 // stack to reg
1314 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1315 }
1316 } else if (dst.first()->is_stack()) {
1317 // reg to stack
1318 __ extsw(r_temp, src.first()->as_Register());
1319 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1320 } else {
1321 // reg to reg
1322 __ extsw(dst.first()->as_Register(), src.first()->as_Register());
1323 }
1324 }
1325
1326 static void long_move(MacroAssembler*masm,
1327 VMRegPair src, VMRegPair dst,
1328 Register r_caller_sp, Register r_temp) {
1329 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long");
1330 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1331
1332 if (src.first()->is_stack()) {
1333 if (dst.first()->is_stack()) {
1334 // stack to stack
1335 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1336 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1337 } else {
1338 // stack to reg
1339 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1340 }
1341 } else if (dst.first()->is_stack()) {
1342 // reg to stack
1343 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP);
1344 } else {
1345 // reg to reg
1346 if (dst.first()->as_Register() != src.first()->as_Register())
1347 __ mr(dst.first()->as_Register(), src.first()->as_Register());
1348 }
1349 }
1350
1351 static void float_move(MacroAssembler*masm,
1352 VMRegPair src, VMRegPair dst,
1353 Register r_caller_sp, Register r_temp) {
1354 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float");
1355 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float");
1356
1357 if (src.first()->is_stack()) {
1358 if (dst.first()->is_stack()) {
1359 // stack to stack
1360 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp);
1361 __ stw(r_temp, reg2offset(dst.first()), R1_SP);
1362 } else {
1363 // stack to reg
1364 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1365 }
1366 } else if (dst.first()->is_stack()) {
1367 // reg to stack
1368 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1369 } else {
1370 // reg to reg
1371 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1372 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1373 }
1374 }
1375
1376 static void double_move(MacroAssembler*masm,
1377 VMRegPair src, VMRegPair dst,
1378 Register r_caller_sp, Register r_temp) {
1379 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double");
1380 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double");
1381
1382 if (src.first()->is_stack()) {
1383 if (dst.first()->is_stack()) {
1384 // stack to stack
1385 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1386 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1387 } else {
1388 // stack to reg
1389 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1390 }
1391 } else if (dst.first()->is_stack()) {
1392 // reg to stack
1393 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1394 } else {
1395 // reg to reg
1396 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1397 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1398 }
1399 }
1400
1401 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1402 switch (ret_type) {
1403 case T_BOOLEAN:
1404 case T_CHAR:
1405 case T_BYTE:
1406 case T_SHORT:
1407 case T_INT:
1408 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1409 break;
1410 case T_ARRAY:
1411 case T_OBJECT:
1412 case T_LONG:
1413 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1414 break;
1415 case T_FLOAT:
1416 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1417 break;
1418 case T_DOUBLE:
1419 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1420 break;
1421 case T_VOID:
1422 break;
1423 default:
1424 ShouldNotReachHere();
1425 break;
1426 }
1427 }
1428
1429 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1430 switch (ret_type) {
1431 case T_BOOLEAN:
1432 case T_CHAR:
1433 case T_BYTE:
1434 case T_SHORT:
1435 case T_INT:
1436 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1437 break;
1438 case T_ARRAY:
1439 case T_OBJECT:
1440 case T_LONG:
1441 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1442 break;
1443 case T_FLOAT:
1444 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1445 break;
1446 case T_DOUBLE:
1447 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1448 break;
1449 case T_VOID:
1450 break;
1451 default:
1452 ShouldNotReachHere();
1453 break;
1454 }
1455 }
1456
1457 static void verify_oop_args(MacroAssembler* masm,
1458 const methodHandle& method,
1459 const BasicType* sig_bt,
1460 const VMRegPair* regs) {
1461 Register temp_reg = R19_method; // not part of any compiled calling seq
1462 if (VerifyOops) {
1463 for (int i = 0; i < method->size_of_parameters(); i++) {
1464 if (is_reference_type(sig_bt[i])) {
1465 VMReg r = regs[i].first();
1466 assert(r->is_valid(), "bad oop arg");
1467 if (r->is_stack()) {
1468 __ ld(temp_reg, reg2offset(r), R1_SP);
1469 __ verify_oop(temp_reg, FILE_AND_LINE);
1470 } else {
1471 __ verify_oop(r->as_Register(), FILE_AND_LINE);
1472 }
1473 }
1474 }
1475 }
1476 }
1477
1478 static void gen_special_dispatch(MacroAssembler* masm,
1479 const methodHandle& method,
1480 const BasicType* sig_bt,
1481 const VMRegPair* regs) {
1482 verify_oop_args(masm, method, sig_bt, regs);
1483 vmIntrinsics::ID iid = method->intrinsic_id();
1484
1485 // Now write the args into the outgoing interpreter space
1486 bool has_receiver = false;
1487 Register receiver_reg = noreg;
1488 int member_arg_pos = -1;
1489 Register member_reg = noreg;
1490 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1491 if (ref_kind != 0) {
1492 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1493 member_reg = R19_method; // known to be free at this point
1494 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1495 } else if (iid == vmIntrinsics::_invokeBasic) {
1496 has_receiver = true;
1497 } else if (iid == vmIntrinsics::_linkToNative) {
1498 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1499 member_reg = R19_method; // known to be free at this point
1500 } else {
1501 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1502 }
1503
1504 if (member_reg != noreg) {
1505 // Load the member_arg into register, if necessary.
1506 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1507 VMReg r = regs[member_arg_pos].first();
1508 if (r->is_stack()) {
1509 __ ld(member_reg, reg2offset(r), R1_SP);
1510 } else {
1511 // no data motion is needed
1512 member_reg = r->as_Register();
1513 }
1514 }
1515
1516 if (has_receiver) {
1517 // Make sure the receiver is loaded into a register.
1518 assert(method->size_of_parameters() > 0, "oob");
1519 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1520 VMReg r = regs[0].first();
1521 assert(r->is_valid(), "bad receiver arg");
1522 if (r->is_stack()) {
1523 // Porting note: This assumes that compiled calling conventions always
1524 // pass the receiver oop in a register. If this is not true on some
1525 // platform, pick a temp and load the receiver from stack.
1526 fatal("receiver always in a register");
1527 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point?
1528 __ ld(receiver_reg, reg2offset(r), R1_SP);
1529 } else {
1530 // no data motion is needed
1531 receiver_reg = r->as_Register();
1532 }
1533 }
1534
1535 // Figure out which address we are really jumping to:
1536 MethodHandles::generate_method_handle_dispatch(masm, iid,
1537 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1538 }
1539
1540 //---------------------------- continuation_enter_setup ---------------------------
1541 //
1542 // Frame setup.
1543 //
1544 // Arguments:
1545 // None.
1546 //
1547 // Results:
1548 // R1_SP: pointer to blank ContinuationEntry in the pushed frame.
1549 //
1550 // Kills:
1551 // R0, R20
1552 //
1553 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& framesize_words) {
1554 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1555 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
1556 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1557
1558 const int frame_size_in_bytes = (int)ContinuationEntry::size();
1559 assert(is_aligned(frame_size_in_bytes, frame::alignment_in_bytes), "alignment error");
1560
1561 framesize_words = frame_size_in_bytes / wordSize;
1562
1563 DEBUG_ONLY(__ block_comment("setup {"));
1564 // Save return pc and push entry frame
1565 const Register return_pc = R20;
1566 __ mflr(return_pc);
1567 __ std(return_pc, _abi0(lr), R1_SP); // SP->lr = return_pc
1568 __ push_frame(frame_size_in_bytes , R0); // SP -= frame_size_in_bytes
1569
1570 OopMap* map = new OopMap((int)frame_size_in_bytes / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1571
1572 __ ld_ptr(R0, JavaThread::cont_entry_offset(), R16_thread);
1573 __ st_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1574 __ st_ptr(R0, ContinuationEntry::parent_offset(), R1_SP);
1575 DEBUG_ONLY(__ block_comment("} setup"));
1576
1577 return map;
1578 }
1579
1580 //---------------------------- fill_continuation_entry ---------------------------
1581 //
1582 // Initialize the new ContinuationEntry.
1583 //
1584 // Arguments:
1585 // R1_SP: pointer to blank Continuation entry
1586 // reg_cont_obj: pointer to the continuation
1587 // reg_flags: flags
1588 //
1589 // Results:
1590 // R1_SP: pointer to filled out ContinuationEntry
1591 //
1592 // Kills:
1593 // R8_ARG6, R9_ARG7, R10_ARG8
1594 //
1595 static void fill_continuation_entry(MacroAssembler* masm, Register reg_cont_obj, Register reg_flags) {
1596 assert_different_registers(reg_cont_obj, reg_flags);
1597 Register zero = R8_ARG6;
1598 Register tmp2 = R9_ARG7;
1599 Register tmp3 = R10_ARG8;
1600
1601 DEBUG_ONLY(__ block_comment("fill {"));
1602 #ifdef ASSERT
1603 __ load_const_optimized(tmp2, ContinuationEntry::cookie_value());
1604 __ stw(tmp2, in_bytes(ContinuationEntry::cookie_offset()), R1_SP);
1605 #endif //ASSERT
1606
1607 __ li(zero, 0);
1608 __ st_ptr(reg_cont_obj, ContinuationEntry::cont_offset(), R1_SP);
1609 __ stw(reg_flags, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1610 __ st_ptr(zero, ContinuationEntry::chunk_offset(), R1_SP);
1611 __ stw(zero, in_bytes(ContinuationEntry::argsize_offset()), R1_SP);
1612 __ stw(zero, in_bytes(ContinuationEntry::pin_count_offset()), R1_SP);
1613
1614 __ ld_ptr(tmp2, JavaThread::cont_fastpath_offset(), R16_thread);
1615 __ ld(tmp3, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1616 __ st_ptr(tmp2, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1617 __ std(tmp3, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1618
1619 __ st_ptr(zero, JavaThread::cont_fastpath_offset(), R16_thread);
1620 __ std(zero, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1621 DEBUG_ONLY(__ block_comment("} fill"));
1622 }
1623
1624 //---------------------------- continuation_enter_cleanup ---------------------------
1625 //
1626 // Copy corresponding attributes from the top ContinuationEntry to the JavaThread
1627 // before deleting it.
1628 //
1629 // Arguments:
1630 // R1_SP: pointer to the ContinuationEntry
1631 //
1632 // Results:
1633 // None.
1634 //
1635 // Kills:
1636 // R8_ARG6, R9_ARG7, R10_ARG8, R15_esp
1637 //
1638 static void continuation_enter_cleanup(MacroAssembler* masm) {
1639 Register tmp1 = R8_ARG6;
1640 Register tmp2 = R9_ARG7;
1641 Register tmp3 = R10_ARG8;
1642
1643 #ifdef ASSERT
1644 __ block_comment("clean {");
1645 __ ld_ptr(tmp1, JavaThread::cont_entry_offset(), R16_thread);
1646 __ cmpd(CCR0, R1_SP, tmp1);
1647 __ asm_assert_eq(FILE_AND_LINE ": incorrect R1_SP");
1648 #endif
1649
1650 __ ld_ptr(tmp1, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1651 __ st_ptr(tmp1, JavaThread::cont_fastpath_offset(), R16_thread);
1652
1653 if (CheckJNICalls) {
1654 // Check if this is a virtual thread continuation
1655 Label L_skip_vthread_code;
1656 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1657 __ cmpwi(CCR0, R0, 0);
1658 __ beq(CCR0, L_skip_vthread_code);
1659
1660 // If the held monitor count is > 0 and this vthread is terminating then
1661 // it failed to release a JNI monitor. So we issue the same log message
1662 // that JavaThread::exit does.
1663 __ ld(R0, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1664 __ cmpdi(CCR0, R0, 0);
1665 __ beq(CCR0, L_skip_vthread_code);
1666
1667 // Save return value potentially containing the exception oop
1668 Register ex_oop = R15_esp; // nonvolatile register
1669 __ mr(ex_oop, R3_RET);
1670 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
1671 // Restore potental return value
1672 __ mr(R3_RET, ex_oop);
1673
1674 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
1675 // on termination. The held count is implicitly zeroed below when we restore from
1676 // the parent held count (which has to be zero).
1677 __ li(tmp1, 0);
1678 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1679
1680 __ bind(L_skip_vthread_code);
1681 }
1682 #ifdef ASSERT
1683 else {
1684 // Check if this is a virtual thread continuation
1685 Label L_skip_vthread_code;
1686 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1687 __ cmpwi(CCR0, R0, 0);
1688 __ beq(CCR0, L_skip_vthread_code);
1689
1690 // See comment just above. If not checking JNI calls the JNI count is only
1691 // needed for assertion checking.
1692 __ li(tmp1, 0);
1693 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1694
1695 __ bind(L_skip_vthread_code);
1696 }
1697 #endif
1698
1699 __ ld(tmp2, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1700 __ ld_ptr(tmp3, ContinuationEntry::parent_offset(), R1_SP);
1701 __ std(tmp2, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1702 __ st_ptr(tmp3, JavaThread::cont_entry_offset(), R16_thread);
1703 DEBUG_ONLY(__ block_comment("} clean"));
1704 }
1705
1706 static void check_continuation_enter_argument(VMReg actual_vmreg,
1707 Register expected_reg,
1708 const char* name) {
1709 assert(!actual_vmreg->is_stack(), "%s cannot be on stack", name);
1710 assert(actual_vmreg->as_Register() == expected_reg,
1711 "%s is in unexpected register: %s instead of %s",
1712 name, actual_vmreg->as_Register()->name(), expected_reg->name());
1713 }
1714
1715 static void gen_continuation_enter(MacroAssembler* masm,
1716 const VMRegPair* regs,
1717 int& exception_offset,
1718 OopMapSet* oop_maps,
1719 int& frame_complete,
1720 int& framesize_words,
1721 int& interpreted_entry_offset,
1722 int& compiled_entry_offset) {
1723
1724 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1725 int pos_cont_obj = 0;
1726 int pos_is_cont = 1;
1727 int pos_is_virtual = 2;
1728
1729 // The platform-specific calling convention may present the arguments in various registers.
1730 // To simplify the rest of the code, we expect the arguments to reside at these known
1731 // registers, and we additionally check the placement here in case calling convention ever
1732 // changes.
1733 Register reg_cont_obj = R3_ARG1;
1734 Register reg_is_cont = R4_ARG2;
1735 Register reg_is_virtual = R5_ARG3;
1736
1737 check_continuation_enter_argument(regs[pos_cont_obj].first(), reg_cont_obj, "Continuation object");
1738 check_continuation_enter_argument(regs[pos_is_cont].first(), reg_is_cont, "isContinue");
1739 check_continuation_enter_argument(regs[pos_is_virtual].first(), reg_is_virtual, "isVirtualThread");
1740
1741 address resolve_static_call = SharedRuntime::get_resolve_static_call_stub();
1742
1743 address start = __ pc();
1744
1745 Label L_thaw, L_exit;
1746
1747 // i2i entry used at interp_only_mode only
1748 interpreted_entry_offset = __ pc() - start;
1749 {
1750 #ifdef ASSERT
1751 Label is_interp_only;
1752 __ lwz(R0, in_bytes(JavaThread::interp_only_mode_offset()), R16_thread);
1753 __ cmpwi(CCR0, R0, 0);
1754 __ bne(CCR0, is_interp_only);
1755 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1756 __ bind(is_interp_only);
1757 #endif
1758
1759 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1760 __ ld(reg_cont_obj, Interpreter::stackElementSize*3, R15_esp);
1761 __ lwz(reg_is_cont, Interpreter::stackElementSize*2, R15_esp);
1762 __ lwz(reg_is_virtual, Interpreter::stackElementSize*1, R15_esp);
1763
1764 __ push_cont_fastpath();
1765
1766 OopMap* map = continuation_enter_setup(masm, framesize_words);
1767
1768 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1769 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1770
1771 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1772
1773 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1774 __ cmpwi(CCR0, reg_is_cont, 0);
1775 __ bne(CCR0, L_thaw);
1776
1777 // --- call Continuation.enter(Continuation c, boolean isContinue)
1778
1779 // Emit compiled static call. The call will be always resolved to the c2i
1780 // entry of Continuation.enter(Continuation c, boolean isContinue).
1781 // There are special cases in SharedRuntime::resolve_static_call_C() and
1782 // SharedRuntime::resolve_sub_helper_internal() to achieve this
1783 // See also corresponding call below.
1784 address c2i_call_pc = __ pc();
1785 int start_offset = __ offset();
1786 // Put the entry point as a constant into the constant pool.
1787 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1788 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1789 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1790
1791 // Emit the trampoline stub which will be related to the branch-and-link below.
1792 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1793 guarantee(stub != nullptr, "no space for trampoline stub");
1794
1795 __ relocate(relocInfo::static_call_type);
1796 // Note: At this point we do not have the address of the trampoline
1797 // stub, and the entry point might be too far away for bl, so __ pc()
1798 // serves as dummy and the bl will be patched later.
1799 __ bl(__ pc());
1800 oop_maps->add_gc_map(__ pc() - start, map);
1801 __ post_call_nop();
1802
1803 __ b(L_exit);
1804
1805 // static stub for the call above
1806 stub = CompiledDirectCall::emit_to_interp_stub(masm, c2i_call_pc);
1807 guarantee(stub != nullptr, "no space for static stub");
1808 }
1809
1810 // compiled entry
1811 __ align(CodeEntryAlignment);
1812 compiled_entry_offset = __ pc() - start;
1813
1814 OopMap* map = continuation_enter_setup(masm, framesize_words);
1815
1816 // Frame is now completed as far as size and linkage.
1817 frame_complete =__ pc() - start;
1818
1819 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1820
1821 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1822 __ cmpwi(CCR0, reg_is_cont, 0);
1823 __ bne(CCR0, L_thaw);
1824
1825 // --- call Continuation.enter(Continuation c, boolean isContinue)
1826
1827 // Emit compiled static call
1828 // The call needs to be resolved. There's a special case for this in
1829 // SharedRuntime::find_callee_info_helper() which calls
1830 // LinkResolver::resolve_continuation_enter() which resolves the call to
1831 // Continuation.enter(Continuation c, boolean isContinue).
1832 address call_pc = __ pc();
1833 int start_offset = __ offset();
1834 // Put the entry point as a constant into the constant pool.
1835 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1836 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1837 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1838
1839 // Emit the trampoline stub which will be related to the branch-and-link below.
1840 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1841 guarantee(stub != nullptr, "no space for trampoline stub");
1842
1843 __ relocate(relocInfo::static_call_type);
1844 // Note: At this point we do not have the address of the trampoline
1845 // stub, and the entry point might be too far away for bl, so __ pc()
1846 // serves as dummy and the bl will be patched later.
1847 __ bl(__ pc());
1848 oop_maps->add_gc_map(__ pc() - start, map);
1849 __ post_call_nop();
1850
1851 __ b(L_exit);
1852
1853 // --- Thawing path
1854
1855 __ bind(L_thaw);
1856 __ add_const_optimized(R0, R29_TOC, MacroAssembler::offset_to_global_toc(StubRoutines::cont_thaw()));
1857 __ mtctr(R0);
1858 __ bctrl();
1859 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1860 ContinuationEntry::_return_pc_offset = __ pc() - start;
1861 __ post_call_nop();
1862
1863 // --- Normal exit (resolve/thawing)
1864
1865 __ bind(L_exit);
1866 continuation_enter_cleanup(masm);
1867
1868 // Pop frame and return
1869 DEBUG_ONLY(__ ld_ptr(R0, 0, R1_SP));
1870 __ addi(R1_SP, R1_SP, framesize_words*wordSize);
1871 DEBUG_ONLY(__ cmpd(CCR0, R0, R1_SP));
1872 __ asm_assert_eq(FILE_AND_LINE ": inconsistent frame size");
1873 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1874 __ mtlr(R0);
1875 __ blr();
1876
1877 // --- Exception handling path
1878
1879 exception_offset = __ pc() - start;
1880
1881 continuation_enter_cleanup(masm);
1882 Register ex_pc = R17_tos; // nonvolatile register
1883 Register ex_oop = R15_esp; // nonvolatile register
1884 __ ld(ex_pc, _abi0(callers_sp), R1_SP); // Load caller's return pc
1885 __ ld(ex_pc, _abi0(lr), ex_pc);
1886 __ mr(ex_oop, R3_RET); // save return value containing the exception oop
1887 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, ex_pc);
1888 __ mtlr(R3_RET); // the exception handler
1889 __ ld(R1_SP, _abi0(callers_sp), R1_SP); // remove enterSpecial frame
1890
1891 // Continue at exception handler
1892 // See OptoRuntime::generate_exception_blob for register arguments
1893 __ mr(R3_ARG1, ex_oop); // pass exception oop
1894 __ mr(R4_ARG2, ex_pc); // pass exception pc
1895 __ blr();
1896
1897 // static stub for the call above
1898 stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1899 guarantee(stub != nullptr, "no space for static stub");
1900 }
1901
1902 static void gen_continuation_yield(MacroAssembler* masm,
1903 const VMRegPair* regs,
1904 OopMapSet* oop_maps,
1905 int& frame_complete,
1906 int& framesize_words,
1907 int& compiled_entry_offset) {
1908 Register tmp = R10_ARG8;
1909
1910 const int framesize_bytes = (int)align_up((int)frame::native_abi_reg_args_size, frame::alignment_in_bytes);
1911 framesize_words = framesize_bytes / wordSize;
1912
1913 address start = __ pc();
1914 compiled_entry_offset = __ pc() - start;
1915
1916 // Save return pc and push entry frame
1917 __ mflr(tmp);
1918 __ std(tmp, _abi0(lr), R1_SP); // SP->lr = return_pc
1919 __ push_frame(framesize_bytes , R0); // SP -= frame_size_in_bytes
1920
1921 DEBUG_ONLY(__ block_comment("Frame Complete"));
1922 frame_complete = __ pc() - start;
1923 address last_java_pc = __ pc();
1924
1925 // This nop must be exactly at the PC we push into the frame info.
1926 // We use this nop for fast CodeBlob lookup, associate the OopMap
1927 // with it right away.
1928 __ post_call_nop();
1929 OopMap* map = new OopMap(framesize_bytes / VMRegImpl::stack_slot_size, 1);
1930 oop_maps->add_gc_map(last_java_pc - start, map);
1931
1932 __ calculate_address_from_global_toc(tmp, last_java_pc); // will be relocated
1933 __ set_last_Java_frame(R1_SP, tmp);
1934 __ call_VM_leaf(Continuation::freeze_entry(), R16_thread, R1_SP);
1935 __ reset_last_Java_frame();
1936
1937 Label L_pinned;
1938
1939 __ cmpwi(CCR0, R3_RET, 0);
1940 __ bne(CCR0, L_pinned);
1941
1942 // yield succeeded
1943
1944 // Pop frames of continuation including this stub's frame
1945 __ ld_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1946 // The frame pushed by gen_continuation_enter is on top now again
1947 continuation_enter_cleanup(masm);
1948
1949 // Pop frame and return
1950 Label L_return;
1951 __ bind(L_return);
1952 __ pop_frame();
1953 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1954 __ mtlr(R0);
1955 __ blr();
1956
1957 // yield failed - continuation is pinned
1958
1959 __ bind(L_pinned);
1960
1961 // handle pending exception thrown by freeze
1962 __ ld(tmp, in_bytes(JavaThread::pending_exception_offset()), R16_thread);
1963 __ cmpdi(CCR0, tmp, 0);
1964 __ beq(CCR0, L_return); // return if no exception is pending
1965 __ pop_frame();
1966 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1967 __ mtlr(R0);
1968 __ load_const_optimized(tmp, StubRoutines::forward_exception_entry(), R0);
1969 __ mtctr(tmp);
1970 __ bctr();
1971 }
1972
1973 // ---------------------------------------------------------------------------
1974 // Generate a native wrapper for a given method. The method takes arguments
1975 // in the Java compiled code convention, marshals them to the native
1976 // convention (handlizes oops, etc), transitions to native, makes the call,
1977 // returns to java state (possibly blocking), unhandlizes any result and
1978 // returns.
1979 //
1980 // Critical native functions are a shorthand for the use of
1981 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1982 // functions. The wrapper is expected to unpack the arguments before
1983 // passing them to the callee. Critical native functions leave the state _in_Java,
1984 // since they cannot stop for GC.
1985 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1986 // block and the check for pending exceptions it's impossible for them
1987 // to be thrown.
1988 //
1989 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
1990 const methodHandle& method,
1991 int compile_id,
1992 BasicType *in_sig_bt,
1993 VMRegPair *in_regs,
1994 BasicType ret_type) {
1995 if (method->is_continuation_native_intrinsic()) {
1996 int exception_offset = -1;
1997 OopMapSet* oop_maps = new OopMapSet();
1998 int frame_complete = -1;
1999 int stack_slots = -1;
2000 int interpreted_entry_offset = -1;
2001 int vep_offset = -1;
2002 if (method->is_continuation_enter_intrinsic()) {
2003 gen_continuation_enter(masm,
2004 in_regs,
2005 exception_offset,
2006 oop_maps,
2007 frame_complete,
2008 stack_slots,
2009 interpreted_entry_offset,
2010 vep_offset);
2011 } else if (method->is_continuation_yield_intrinsic()) {
2012 gen_continuation_yield(masm,
2013 in_regs,
2014 oop_maps,
2015 frame_complete,
2016 stack_slots,
2017 vep_offset);
2018 } else {
2019 guarantee(false, "Unknown Continuation native intrinsic");
2020 }
2021
2022 #ifdef ASSERT
2023 if (method->is_continuation_enter_intrinsic()) {
2024 assert(interpreted_entry_offset != -1, "Must be set");
2025 assert(exception_offset != -1, "Must be set");
2026 } else {
2027 assert(interpreted_entry_offset == -1, "Must be unset");
2028 assert(exception_offset == -1, "Must be unset");
2029 }
2030 assert(frame_complete != -1, "Must be set");
2031 assert(stack_slots != -1, "Must be set");
2032 assert(vep_offset != -1, "Must be set");
2033 #endif
2034
2035 __ flush();
2036 nmethod* nm = nmethod::new_native_nmethod(method,
2037 compile_id,
2038 masm->code(),
2039 vep_offset,
2040 frame_complete,
2041 stack_slots,
2042 in_ByteSize(-1),
2043 in_ByteSize(-1),
2044 oop_maps,
2045 exception_offset);
2046 if (nm == nullptr) return nm;
2047 if (method->is_continuation_enter_intrinsic()) {
2048 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
2049 } else if (method->is_continuation_yield_intrinsic()) {
2050 _cont_doYield_stub = nm;
2051 }
2052 return nm;
2053 }
2054
2055 if (method->is_method_handle_intrinsic()) {
2056 vmIntrinsics::ID iid = method->intrinsic_id();
2057 intptr_t start = (intptr_t)__ pc();
2058 int vep_offset = ((intptr_t)__ pc()) - start;
2059 gen_special_dispatch(masm,
2060 method,
2061 in_sig_bt,
2062 in_regs);
2063 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
2064 __ flush();
2065 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
2066 return nmethod::new_native_nmethod(method,
2067 compile_id,
2068 masm->code(),
2069 vep_offset,
2070 frame_complete,
2071 stack_slots / VMRegImpl::slots_per_word,
2072 in_ByteSize(-1),
2073 in_ByteSize(-1),
2074 (OopMapSet*)nullptr);
2075 }
2076
2077 address native_func = method->native_function();
2078 assert(native_func != nullptr, "must have function");
2079
2080 // First, create signature for outgoing C call
2081 // --------------------------------------------------------------------------
2082
2083 int total_in_args = method->size_of_parameters();
2084 // We have received a description of where all the java args are located
2085 // on entry to the wrapper. We need to convert these args to where
2086 // the jni function will expect them. To figure out where they go
2087 // we convert the java signature to a C signature by inserting
2088 // the hidden arguments as arg[0] and possibly arg[1] (static method)
2089
2090 // Calculate the total number of C arguments and create arrays for the
2091 // signature and the outgoing registers.
2092 // On ppc64, we have two arrays for the outgoing registers, because
2093 // some floating-point arguments must be passed in registers _and_
2094 // in stack locations.
2095 bool method_is_static = method->is_static();
2096 int total_c_args = total_in_args + (method_is_static ? 2 : 1);
2097
2098 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2099 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2100 BasicType* in_elem_bt = nullptr;
2101
2102 // Create the signature for the C call:
2103 // 1) add the JNIEnv*
2104 // 2) add the class if the method is static
2105 // 3) copy the rest of the incoming signature (shifted by the number of
2106 // hidden arguments).
2107
2108 int argc = 0;
2109 out_sig_bt[argc++] = T_ADDRESS;
2110 if (method->is_static()) {
2111 out_sig_bt[argc++] = T_OBJECT;
2112 }
2113
2114 for (int i = 0; i < total_in_args ; i++ ) {
2115 out_sig_bt[argc++] = in_sig_bt[i];
2116 }
2117
2118
2119 // Compute the wrapper's frame size.
2120 // --------------------------------------------------------------------------
2121
2122 // Now figure out where the args must be stored and how much stack space
2123 // they require.
2124 //
2125 // Compute framesize for the wrapper. We need to handlize all oops in
2126 // incoming registers.
2127 //
2128 // Calculate the total number of stack slots we will need:
2129 // 1) abi requirements
2130 // 2) outgoing arguments
2131 // 3) space for inbound oop handle area
2132 // 4) space for handlizing a klass if static method
2133 // 5) space for a lock if synchronized method
2134 // 6) workspace for saving return values, int <-> float reg moves, etc.
2135 // 7) alignment
2136 //
2137 // Layout of the native wrapper frame:
2138 // (stack grows upwards, memory grows downwards)
2139 //
2140 // NW [ABI_REG_ARGS] <-- 1) R1_SP
2141 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset
2142 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset
2143 // klass <-- 4) R1_SP + klass_offset
2144 // lock <-- 5) R1_SP + lock_offset
2145 // [workspace] <-- 6) R1_SP + workspace_offset
2146 // [alignment] (optional) <-- 7)
2147 // caller [JIT_TOP_ABI_48] <-- r_callers_sp
2148 //
2149 // - *_slot_offset Indicates offset from SP in number of stack slots.
2150 // - *_offset Indicates offset from SP in bytes.
2151
2152 int stack_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args) + // 1+2)
2153 SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention.
2154
2155 // Now the space for the inbound oop handle area.
2156 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word;
2157
2158 int oop_handle_slot_offset = stack_slots;
2159 stack_slots += total_save_slots; // 3)
2160
2161 int klass_slot_offset = 0;
2162 int klass_offset = -1;
2163 if (method_is_static) { // 4)
2164 klass_slot_offset = stack_slots;
2165 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
2166 stack_slots += VMRegImpl::slots_per_word;
2167 }
2168
2169 int lock_slot_offset = 0;
2170 int lock_offset = -1;
2171 if (method->is_synchronized()) { // 5)
2172 lock_slot_offset = stack_slots;
2173 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size;
2174 stack_slots += VMRegImpl::slots_per_word;
2175 }
2176
2177 int workspace_slot_offset = stack_slots; // 6)
2178 stack_slots += 2;
2179
2180 // Now compute actual number of stack words we need.
2181 // Rounding to make stack properly aligned.
2182 stack_slots = align_up(stack_slots, // 7)
2183 frame::alignment_in_bytes / VMRegImpl::stack_slot_size);
2184 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size;
2185
2186
2187 // Now we can start generating code.
2188 // --------------------------------------------------------------------------
2189
2190 intptr_t start_pc = (intptr_t)__ pc();
2191 intptr_t vep_start_pc;
2192 intptr_t frame_done_pc;
2193 intptr_t oopmap_pc;
2194
2195 Label handle_pending_exception;
2196
2197 Register r_callers_sp = R21;
2198 Register r_temp_1 = R22;
2199 Register r_temp_2 = R23;
2200 Register r_temp_3 = R24;
2201 Register r_temp_4 = R25;
2202 Register r_temp_5 = R26;
2203 Register r_temp_6 = R27;
2204 Register r_return_pc = R28;
2205
2206 Register r_carg1_jnienv = noreg;
2207 Register r_carg2_classorobject = noreg;
2208 r_carg1_jnienv = out_regs[0].first()->as_Register();
2209 r_carg2_classorobject = out_regs[1].first()->as_Register();
2210
2211
2212 // Generate the Unverified Entry Point (UEP).
2213 // --------------------------------------------------------------------------
2214 assert(start_pc == (intptr_t)__ pc(), "uep must be at start");
2215
2216 // Check ic: object class == cached class?
2217 if (!method_is_static) {
2218 __ ic_check(4 /* end_alignment */);
2219 }
2220
2221 // Generate the Verified Entry Point (VEP).
2222 // --------------------------------------------------------------------------
2223 vep_start_pc = (intptr_t)__ pc();
2224
2225 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
2226 Label L_skip_barrier;
2227 Register klass = r_temp_1;
2228 // Notify OOP recorder (don't need the relocation)
2229 AddressLiteral md = __ constant_metadata_address(method->method_holder());
2230 __ load_const_optimized(klass, md.value(), R0);
2231 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
2232
2233 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
2234 __ mtctr(klass);
2235 __ bctr();
2236
2237 __ bind(L_skip_barrier);
2238 }
2239
2240 __ save_LR(r_temp_1);
2241 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame.
2242 __ mr(r_callers_sp, R1_SP); // Remember frame pointer.
2243 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame.
2244
2245 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2246 bs->nmethod_entry_barrier(masm, r_temp_1);
2247
2248 frame_done_pc = (intptr_t)__ pc();
2249
2250 // Native nmethod wrappers never take possession of the oop arguments.
2251 // So the caller will gc the arguments.
2252 // The only thing we need an oopMap for is if the call is static.
2253 //
2254 // An OopMap for lock (and class if static), and one for the VM call itself.
2255 OopMapSet *oop_maps = new OopMapSet();
2256 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2257
2258 // Move arguments from register/stack to register/stack.
2259 // --------------------------------------------------------------------------
2260 //
2261 // We immediately shuffle the arguments so that for any vm call we have
2262 // to make from here on out (sync slow path, jvmti, etc.) we will have
2263 // captured the oops from our caller and have a valid oopMap for them.
2264 //
2265 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2266 // (derived from JavaThread* which is in R16_thread) and, if static,
2267 // the class mirror instead of a receiver. This pretty much guarantees that
2268 // register layout will not match. We ignore these extra arguments during
2269 // the shuffle. The shuffle is described by the two calling convention
2270 // vectors we have in our possession. We simply walk the java vector to
2271 // get the source locations and the c vector to get the destinations.
2272
2273 // Record sp-based slot for receiver on stack for non-static methods.
2274 int receiver_offset = -1;
2275
2276 // We move the arguments backward because the floating point registers
2277 // destination will always be to a register with a greater or equal
2278 // register number or the stack.
2279 // in is the index of the incoming Java arguments
2280 // out is the index of the outgoing C arguments
2281
2282 #ifdef ASSERT
2283 bool reg_destroyed[Register::number_of_registers];
2284 bool freg_destroyed[FloatRegister::number_of_registers];
2285 for (int r = 0 ; r < Register::number_of_registers ; r++) {
2286 reg_destroyed[r] = false;
2287 }
2288 for (int f = 0 ; f < FloatRegister::number_of_registers ; f++) {
2289 freg_destroyed[f] = false;
2290 }
2291 #endif // ASSERT
2292
2293 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) {
2294
2295 #ifdef ASSERT
2296 if (in_regs[in].first()->is_Register()) {
2297 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!");
2298 } else if (in_regs[in].first()->is_FloatRegister()) {
2299 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!");
2300 }
2301 if (out_regs[out].first()->is_Register()) {
2302 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true;
2303 } else if (out_regs[out].first()->is_FloatRegister()) {
2304 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true;
2305 }
2306 #endif // ASSERT
2307
2308 switch (in_sig_bt[in]) {
2309 case T_BOOLEAN:
2310 case T_CHAR:
2311 case T_BYTE:
2312 case T_SHORT:
2313 case T_INT:
2314 // Move int and do sign extension.
2315 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2316 break;
2317 case T_LONG:
2318 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2319 break;
2320 case T_ARRAY:
2321 case T_OBJECT:
2322 object_move(masm, stack_slots,
2323 oop_map, oop_handle_slot_offset,
2324 ((in == 0) && (!method_is_static)), &receiver_offset,
2325 in_regs[in], out_regs[out],
2326 r_callers_sp, r_temp_1, r_temp_2);
2327 break;
2328 case T_VOID:
2329 break;
2330 case T_FLOAT:
2331 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2332 break;
2333 case T_DOUBLE:
2334 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2335 break;
2336 case T_ADDRESS:
2337 fatal("found type (T_ADDRESS) in java args");
2338 break;
2339 default:
2340 ShouldNotReachHere();
2341 break;
2342 }
2343 }
2344
2345 // Pre-load a static method's oop into ARG2.
2346 // Used both by locking code and the normal JNI call code.
2347 if (method_is_static) {
2348 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()),
2349 r_carg2_classorobject);
2350
2351 // Now handlize the static class mirror in carg2. It's known not-null.
2352 __ std(r_carg2_classorobject, klass_offset, R1_SP);
2353 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2354 __ addi(r_carg2_classorobject, R1_SP, klass_offset);
2355 }
2356
2357 // Get JNIEnv* which is first argument to native.
2358 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset()));
2359
2360 // NOTE:
2361 //
2362 // We have all of the arguments setup at this point.
2363 // We MUST NOT touch any outgoing regs from this point on.
2364 // So if we must call out we must push a new frame.
2365
2366 // Get current pc for oopmap, and load it patchable relative to global toc.
2367 oopmap_pc = (intptr_t) __ pc();
2368 __ calculate_address_from_global_toc(r_return_pc, (address)oopmap_pc, true, true, true, true);
2369
2370 // We use the same pc/oopMap repeatedly when we call out.
2371 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map);
2372
2373 // r_return_pc now has the pc loaded that we will use when we finally call
2374 // to native.
2375
2376 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below.
2377 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register");
2378
2379 # if 0
2380 // DTrace method entry
2381 # endif
2382
2383 // Lock a synchronized method.
2384 // --------------------------------------------------------------------------
2385
2386 if (method->is_synchronized()) {
2387 Register r_oop = r_temp_4;
2388 const Register r_box = r_temp_5;
2389 Label done, locked;
2390
2391 // Load the oop for the object or class. r_carg2_classorobject contains
2392 // either the handlized oop from the incoming arguments or the handlized
2393 // class mirror (if the method is static).
2394 __ ld(r_oop, 0, r_carg2_classorobject);
2395
2396 // Get the lock box slot's address.
2397 __ addi(r_box, R1_SP, lock_offset);
2398
2399 // Try fastpath for locking.
2400 if (LockingMode == LM_LIGHTWEIGHT) {
2401 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2402 __ compiler_fast_lock_lightweight_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2403 } else {
2404 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2405 __ compiler_fast_lock_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2406 }
2407 __ beq(CCR0, locked);
2408
2409 // None of the above fast optimizations worked so we have to get into the
2410 // slow case of monitor enter. Inline a special case of call_VM that
2411 // disallows any pending_exception.
2412
2413 // Save argument registers and leave room for C-compatible ABI_REG_ARGS.
2414 int frame_size = frame::native_abi_reg_args_size + align_up(total_c_args * wordSize, frame::alignment_in_bytes);
2415 __ mr(R11_scratch1, R1_SP);
2416 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs);
2417
2418 // Do the call.
2419 __ set_last_Java_frame(R11_scratch1, r_return_pc);
2420 assert(r_return_pc->is_nonvolatile(), "expecting return pc to be in non-volatile register");
2421 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread);
2422 __ reset_last_Java_frame();
2423
2424 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs);
2425
2426 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2427 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C");
2428
2429 __ bind(locked);
2430 }
2431
2432 // Use that pc we placed in r_return_pc a while back as the current frame anchor.
2433 __ set_last_Java_frame(R1_SP, r_return_pc);
2434
2435 // Publish thread state
2436 // --------------------------------------------------------------------------
2437
2438 // Transition from _thread_in_Java to _thread_in_native.
2439 __ li(R0, _thread_in_native);
2440 __ release();
2441 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2442 __ stw(R0, thread_(thread_state));
2443
2444
2445 // The JNI call
2446 // --------------------------------------------------------------------------
2447 __ call_c(native_func, relocInfo::runtime_call_type);
2448
2449
2450 // Now, we are back from the native code.
2451
2452
2453 // Unpack the native result.
2454 // --------------------------------------------------------------------------
2455
2456 // For int-types, we do any needed sign-extension required.
2457 // Care must be taken that the return values (R3_RET and F1_RET)
2458 // will survive any VM calls for blocking or unlocking.
2459 // An OOP result (handle) is done specially in the slow-path code.
2460
2461 switch (ret_type) {
2462 case T_VOID: break; // Nothing to do!
2463 case T_FLOAT: break; // Got it where we want it (unless slow-path).
2464 case T_DOUBLE: break; // Got it where we want it (unless slow-path).
2465 case T_LONG: break; // Got it where we want it (unless slow-path).
2466 case T_OBJECT: break; // Really a handle.
2467 // Cannot de-handlize until after reclaiming jvm_lock.
2468 case T_ARRAY: break;
2469
2470 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1)
2471 __ normalize_bool(R3_RET);
2472 break;
2473 }
2474 case T_BYTE: { // sign extension
2475 __ extsb(R3_RET, R3_RET);
2476 break;
2477 }
2478 case T_CHAR: { // unsigned result
2479 __ andi(R3_RET, R3_RET, 0xffff);
2480 break;
2481 }
2482 case T_SHORT: { // sign extension
2483 __ extsh(R3_RET, R3_RET);
2484 break;
2485 }
2486 case T_INT: // nothing to do
2487 break;
2488 default:
2489 ShouldNotReachHere();
2490 break;
2491 }
2492
2493 Label after_transition;
2494
2495 // Publish thread state
2496 // --------------------------------------------------------------------------
2497
2498 // Switch thread to "native transition" state before reading the
2499 // synchronization state. This additional state is necessary because reading
2500 // and testing the synchronization state is not atomic w.r.t. GC, as this
2501 // scenario demonstrates:
2502 // - Java thread A, in _thread_in_native state, loads _not_synchronized
2503 // and is preempted.
2504 // - VM thread changes sync state to synchronizing and suspends threads
2505 // for GC.
2506 // - Thread A is resumed to finish this native method, but doesn't block
2507 // here since it didn't see any synchronization in progress, and escapes.
2508
2509 // Transition from _thread_in_native to _thread_in_native_trans.
2510 __ li(R0, _thread_in_native_trans);
2511 __ release();
2512 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2513 __ stw(R0, thread_(thread_state));
2514
2515
2516 // Must we block?
2517 // --------------------------------------------------------------------------
2518
2519 // Block, if necessary, before resuming in _thread_in_Java state.
2520 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2521 {
2522 Label no_block, sync;
2523
2524 // Force this write out before the read below.
2525 if (!UseSystemMemoryBarrier) {
2526 __ fence();
2527 }
2528
2529 Register sync_state_addr = r_temp_4;
2530 Register sync_state = r_temp_5;
2531 Register suspend_flags = r_temp_6;
2532
2533 // No synchronization in progress nor yet synchronized
2534 // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
2535 __ safepoint_poll(sync, sync_state, true /* at_return */, false /* in_nmethod */);
2536
2537 // Not suspended.
2538 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
2539 __ lwz(suspend_flags, thread_(suspend_flags));
2540 __ cmpwi(CCR1, suspend_flags, 0);
2541 __ beq(CCR1, no_block);
2542
2543 // Block. Save any potential method result value before the operation and
2544 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2545 // lets us share the oopMap we used when we went native rather than create
2546 // a distinct one for this pc.
2547 __ bind(sync);
2548 __ isync();
2549
2550 address entry_point =
2551 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans);
2552 save_native_result(masm, ret_type, workspace_slot_offset);
2553 __ call_VM_leaf(entry_point, R16_thread);
2554 restore_native_result(masm, ret_type, workspace_slot_offset);
2555
2556 __ bind(no_block);
2557
2558 // Publish thread state.
2559 // --------------------------------------------------------------------------
2560
2561 // Thread state is thread_in_native_trans. Any safepoint blocking has
2562 // already happened so we can now change state to _thread_in_Java.
2563
2564 // Transition from _thread_in_native_trans to _thread_in_Java.
2565 __ li(R0, _thread_in_Java);
2566 __ lwsync(); // Acquire safepoint and suspend state, release thread state.
2567 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2568 __ stw(R0, thread_(thread_state));
2569 __ bind(after_transition);
2570 }
2571
2572 // Reguard any pages if necessary.
2573 // --------------------------------------------------------------------------
2574
2575 Label no_reguard;
2576 __ lwz(r_temp_1, thread_(stack_guard_state));
2577 __ cmpwi(CCR0, r_temp_1, StackOverflow::stack_guard_yellow_reserved_disabled);
2578 __ bne(CCR0, no_reguard);
2579
2580 save_native_result(masm, ret_type, workspace_slot_offset);
2581 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2582 restore_native_result(masm, ret_type, workspace_slot_offset);
2583
2584 __ bind(no_reguard);
2585
2586
2587 // Unlock
2588 // --------------------------------------------------------------------------
2589
2590 if (method->is_synchronized()) {
2591 const Register r_oop = r_temp_4;
2592 const Register r_box = r_temp_5;
2593 const Register r_exception = r_temp_6;
2594 Label done;
2595
2596 // Get oop and address of lock object box.
2597 if (method_is_static) {
2598 assert(klass_offset != -1, "");
2599 __ ld(r_oop, klass_offset, R1_SP);
2600 } else {
2601 assert(receiver_offset != -1, "");
2602 __ ld(r_oop, receiver_offset, R1_SP);
2603 }
2604 __ addi(r_box, R1_SP, lock_offset);
2605
2606 // Try fastpath for unlocking.
2607 if (LockingMode == LM_LIGHTWEIGHT) {
2608 __ compiler_fast_unlock_lightweight_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2609 } else {
2610 __ compiler_fast_unlock_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2611 }
2612 __ beq(CCR0, done);
2613
2614 // Save and restore any potential method result value around the unlocking operation.
2615 save_native_result(masm, ret_type, workspace_slot_offset);
2616
2617 // Must save pending exception around the slow-path VM call. Since it's a
2618 // leaf call, the pending exception (if any) can be kept in a register.
2619 __ ld(r_exception, thread_(pending_exception));
2620 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile");
2621 __ li(R0, 0);
2622 __ std(R0, thread_(pending_exception));
2623
2624 // Slow case of monitor enter.
2625 // Inline a special case of call_VM that disallows any pending_exception.
2626 // Arguments are (oop obj, BasicLock* lock, JavaThread* thread).
2627 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box, R16_thread);
2628
2629 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2630 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C");
2631
2632 restore_native_result(masm, ret_type, workspace_slot_offset);
2633
2634 // Check_forward_pending_exception jump to forward_exception if any pending
2635 // exception is set. The forward_exception routine expects to see the
2636 // exception in pending_exception and not in a register. Kind of clumsy,
2637 // since all folks who branch to forward_exception must have tested
2638 // pending_exception first and hence have it in a register already.
2639 __ std(r_exception, thread_(pending_exception));
2640
2641 __ bind(done);
2642 }
2643
2644 # if 0
2645 // DTrace method exit
2646 # endif
2647
2648 // Clear "last Java frame" SP and PC.
2649 // --------------------------------------------------------------------------
2650
2651 __ reset_last_Java_frame();
2652
2653 // Unbox oop result, e.g. JNIHandles::resolve value.
2654 // --------------------------------------------------------------------------
2655
2656 if (is_reference_type(ret_type)) {
2657 __ resolve_jobject(R3_RET, r_temp_1, r_temp_2, MacroAssembler::PRESERVATION_NONE);
2658 }
2659
2660 if (CheckJNICalls) {
2661 // clear_pending_jni_exception_check
2662 __ load_const_optimized(R0, 0L);
2663 __ st_ptr(R0, JavaThread::pending_jni_exception_check_fn_offset(), R16_thread);
2664 }
2665
2666 // Reset handle block.
2667 // --------------------------------------------------------------------------
2668 __ ld(r_temp_1, thread_(active_handles));
2669 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
2670 __ li(r_temp_2, 0);
2671 __ stw(r_temp_2, in_bytes(JNIHandleBlock::top_offset()), r_temp_1);
2672
2673
2674 // Check for pending exceptions.
2675 // --------------------------------------------------------------------------
2676 __ ld(r_temp_2, thread_(pending_exception));
2677 __ cmpdi(CCR0, r_temp_2, 0);
2678 __ bne(CCR0, handle_pending_exception);
2679
2680 // Return
2681 // --------------------------------------------------------------------------
2682
2683 __ pop_frame();
2684 __ restore_LR(R11);
2685 __ blr();
2686
2687
2688 // Handler for pending exceptions (out-of-line).
2689 // --------------------------------------------------------------------------
2690 // Since this is a native call, we know the proper exception handler
2691 // is the empty function. We just pop this frame and then jump to
2692 // forward_exception_entry.
2693 __ bind(handle_pending_exception);
2694
2695 __ pop_frame();
2696 __ restore_LR(R11);
2697 __ b64_patchable((address)StubRoutines::forward_exception_entry(),
2698 relocInfo::runtime_call_type);
2699
2700 // Done.
2701 // --------------------------------------------------------------------------
2702
2703 __ flush();
2704
2705 nmethod *nm = nmethod::new_native_nmethod(method,
2706 compile_id,
2707 masm->code(),
2708 vep_start_pc-start_pc,
2709 frame_done_pc-start_pc,
2710 stack_slots / VMRegImpl::slots_per_word,
2711 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2712 in_ByteSize(lock_offset),
2713 oop_maps);
2714
2715 return nm;
2716 }
2717
2718 // This function returns the adjust size (in number of words) to a c2i adapter
2719 // activation for use during deoptimization.
2720 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2721 return align_up((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::frame_alignment_in_words);
2722 }
2723
2724 uint SharedRuntime::in_preserve_stack_slots() {
2725 return frame::jit_in_preserve_size / VMRegImpl::stack_slot_size;
2726 }
2727
2728 uint SharedRuntime::out_preserve_stack_slots() {
2729 #if defined(COMPILER1) || defined(COMPILER2)
2730 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size;
2731 #else
2732 return 0;
2733 #endif
2734 }
2735
2736 #if defined(COMPILER1) || defined(COMPILER2)
2737 // Frame generation for deopt and uncommon trap blobs.
2738 static void push_skeleton_frame(MacroAssembler* masm, bool deopt,
2739 /* Read */
2740 Register unroll_block_reg,
2741 /* Update */
2742 Register frame_sizes_reg,
2743 Register number_of_frames_reg,
2744 Register pcs_reg,
2745 /* Invalidate */
2746 Register frame_size_reg,
2747 Register pc_reg) {
2748
2749 __ ld(pc_reg, 0, pcs_reg);
2750 __ ld(frame_size_reg, 0, frame_sizes_reg);
2751 __ std(pc_reg, _abi0(lr), R1_SP);
2752 __ push_frame(frame_size_reg, R0/*tmp*/);
2753 __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP);
2754 __ addi(number_of_frames_reg, number_of_frames_reg, -1);
2755 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize);
2756 __ addi(pcs_reg, pcs_reg, wordSize);
2757 }
2758
2759 // Loop through the UnrollBlock info and create new frames.
2760 static void push_skeleton_frames(MacroAssembler* masm, bool deopt,
2761 /* read */
2762 Register unroll_block_reg,
2763 /* invalidate */
2764 Register frame_sizes_reg,
2765 Register number_of_frames_reg,
2766 Register pcs_reg,
2767 Register frame_size_reg,
2768 Register pc_reg) {
2769 Label loop;
2770
2771 // _number_of_frames is of type int (deoptimization.hpp)
2772 __ lwa(number_of_frames_reg,
2773 in_bytes(Deoptimization::UnrollBlock::number_of_frames_offset()),
2774 unroll_block_reg);
2775 __ ld(pcs_reg,
2776 in_bytes(Deoptimization::UnrollBlock::frame_pcs_offset()),
2777 unroll_block_reg);
2778 __ ld(frame_sizes_reg,
2779 in_bytes(Deoptimization::UnrollBlock::frame_sizes_offset()),
2780 unroll_block_reg);
2781
2782 // stack: (caller_of_deoptee, ...).
2783
2784 // At this point we either have an interpreter frame or a compiled
2785 // frame on top of stack. If it is a compiled frame we push a new c2i
2786 // adapter here
2787
2788 // Memorize top-frame stack-pointer.
2789 __ mr(frame_size_reg/*old_sp*/, R1_SP);
2790
2791 // Resize interpreter top frame OR C2I adapter.
2792
2793 // At this moment, the top frame (which is the caller of the deoptee) is
2794 // an interpreter frame or a newly pushed C2I adapter or an entry frame.
2795 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the
2796 // outgoing arguments.
2797 //
2798 // In order to push the interpreter frame for the deoptee, we need to
2799 // resize the top frame such that we are able to place the deoptee's
2800 // locals in the frame.
2801 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI
2802 // into a valid PARENT_IJAVA_FRAME_ABI.
2803
2804 __ lwa(R11_scratch1,
2805 in_bytes(Deoptimization::UnrollBlock::caller_adjustment_offset()),
2806 unroll_block_reg);
2807 __ neg(R11_scratch1, R11_scratch1);
2808
2809 // R11_scratch1 contains size of locals for frame resizing.
2810 // R12_scratch2 contains top frame's lr.
2811
2812 // Resize frame by complete frame size prevents TOC from being
2813 // overwritten by locals. A more stack space saving way would be
2814 // to copy the TOC to its location in the new abi.
2815 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size);
2816
2817 // now, resize the frame
2818 __ resize_frame(R11_scratch1, pc_reg/*tmp*/);
2819
2820 // In the case where we have resized a c2i frame above, the optional
2821 // alignment below the locals has size 32 (why?).
2822 __ std(R12_scratch2, _abi0(lr), R1_SP);
2823
2824 // Initialize initial_caller_sp.
2825 __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP);
2826
2827 #ifdef ASSERT
2828 // Make sure that there is at least one entry in the array.
2829 __ cmpdi(CCR0, number_of_frames_reg, 0);
2830 __ asm_assert_ne("array_size must be > 0");
2831 #endif
2832
2833 // Now push the new interpreter frames.
2834 //
2835 __ bind(loop);
2836 // Allocate a new frame, fill in the pc.
2837 push_skeleton_frame(masm, deopt,
2838 unroll_block_reg,
2839 frame_sizes_reg,
2840 number_of_frames_reg,
2841 pcs_reg,
2842 frame_size_reg,
2843 pc_reg);
2844 __ cmpdi(CCR0, number_of_frames_reg, 0);
2845 __ bne(CCR0, loop);
2846
2847 // Get the return address pointing into the frame manager.
2848 __ ld(R0, 0, pcs_reg);
2849 // Store it in the top interpreter frame.
2850 __ std(R0, _abi0(lr), R1_SP);
2851 // Initialize frame_manager_lr of interpreter top frame.
2852 }
2853 #endif
2854
2855 void SharedRuntime::generate_deopt_blob() {
2856 // Allocate space for the code
2857 ResourceMark rm;
2858 // Setup code generation tools
2859 const char* name = SharedRuntime::stub_name(SharedStubId::deopt_id);
2860 CodeBuffer buffer(name, 2048, 1024);
2861 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2862 Label exec_mode_initialized;
2863 int frame_size_in_words;
2864 OopMap* map = nullptr;
2865 OopMapSet *oop_maps = new OopMapSet();
2866
2867 // size of ABI112 plus spill slots for R3_RET and F1_RET.
2868 const int frame_size_in_bytes = frame::native_abi_reg_args_spill_size;
2869 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
2870 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info.
2871
2872 const Register exec_mode_reg = R21_tmp1;
2873
2874 const address start = __ pc();
2875
2876 #if defined(COMPILER1) || defined(COMPILER2)
2877 // --------------------------------------------------------------------------
2878 // Prolog for non exception case!
2879
2880 // We have been called from the deopt handler of the deoptee.
2881 //
2882 // deoptee:
2883 // ...
2884 // call X
2885 // ...
2886 // deopt_handler: call_deopt_stub
2887 // cur. return pc --> ...
2888 //
2889 // So currently SR_LR points behind the call in the deopt handler.
2890 // We adjust it such that it points to the start of the deopt handler.
2891 // The return_pc has been stored in the frame of the deoptee and
2892 // will replace the address of the deopt_handler in the call
2893 // to Deoptimization::fetch_unroll_info below.
2894 // We can't grab a free register here, because all registers may
2895 // contain live values, so let the RegisterSaver do the adjustment
2896 // of the return pc.
2897 const int return_pc_adjustment_no_exception = -MacroAssembler::bl64_patchable_size;
2898
2899 // Push the "unpack frame"
2900 // Save everything in sight.
2901 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2902 &first_frame_size_in_bytes,
2903 /*generate_oop_map=*/ true,
2904 return_pc_adjustment_no_exception,
2905 RegisterSaver::return_pc_is_lr);
2906 assert(map != nullptr, "OopMap must have been created");
2907
2908 __ li(exec_mode_reg, Deoptimization::Unpack_deopt);
2909 // Save exec mode for unpack_frames.
2910 __ b(exec_mode_initialized);
2911
2912 // --------------------------------------------------------------------------
2913 // Prolog for exception case
2914
2915 // An exception is pending.
2916 // We have been called with a return (interpreter) or a jump (exception blob).
2917 //
2918 // - R3_ARG1: exception oop
2919 // - R4_ARG2: exception pc
2920
2921 int exception_offset = __ pc() - start;
2922
2923 BLOCK_COMMENT("Prolog for exception case");
2924
2925 // Store exception oop and pc in thread (location known to GC).
2926 // This is needed since the call to "fetch_unroll_info()" may safepoint.
2927 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2928 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2929 __ std(R4_ARG2, _abi0(lr), R1_SP);
2930
2931 // Vanilla deoptimization with an exception pending in exception_oop.
2932 int exception_in_tls_offset = __ pc() - start;
2933
2934 // Push the "unpack frame".
2935 // Save everything in sight.
2936 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2937 &first_frame_size_in_bytes,
2938 /*generate_oop_map=*/ false,
2939 /*return_pc_adjustment_exception=*/ 0,
2940 RegisterSaver::return_pc_is_pre_saved);
2941
2942 // Deopt during an exception. Save exec mode for unpack_frames.
2943 __ li(exec_mode_reg, Deoptimization::Unpack_exception);
2944
2945 // fall through
2946
2947 int reexecute_offset = 0;
2948 #ifdef COMPILER1
2949 __ b(exec_mode_initialized);
2950
2951 // Reexecute entry, similar to c2 uncommon trap
2952 reexecute_offset = __ pc() - start;
2953
2954 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2955 &first_frame_size_in_bytes,
2956 /*generate_oop_map=*/ false,
2957 /*return_pc_adjustment_reexecute=*/ 0,
2958 RegisterSaver::return_pc_is_pre_saved);
2959 __ li(exec_mode_reg, Deoptimization::Unpack_reexecute);
2960 #endif
2961
2962 // --------------------------------------------------------------------------
2963 __ BIND(exec_mode_initialized);
2964
2965 const Register unroll_block_reg = R22_tmp2;
2966
2967 // We need to set `last_Java_frame' because `fetch_unroll_info' will
2968 // call `last_Java_frame()'. The value of the pc in the frame is not
2969 // particularly important. It just needs to identify this blob.
2970 __ set_last_Java_frame(R1_SP, noreg);
2971
2972 // With EscapeAnalysis turned on, this call may safepoint!
2973 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread, exec_mode_reg);
2974 address calls_return_pc = __ last_calls_return_pc();
2975 // Set an oopmap for the call site that describes all our saved registers.
2976 oop_maps->add_gc_map(calls_return_pc - start, map);
2977
2978 __ reset_last_Java_frame();
2979 // Save the return value.
2980 __ mr(unroll_block_reg, R3_RET);
2981
2982 // Restore only the result registers that have been saved
2983 // by save_volatile_registers(...).
2984 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes);
2985
2986 // reload the exec mode from the UnrollBlock (it might have changed)
2987 __ lwz(exec_mode_reg, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
2988 // In excp_deopt_mode, restore and clear exception oop which we
2989 // stored in the thread during exception entry above. The exception
2990 // oop will be the return value of this stub.
2991 Label skip_restore_excp;
2992 __ cmpdi(CCR0, exec_mode_reg, Deoptimization::Unpack_exception);
2993 __ bne(CCR0, skip_restore_excp);
2994 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2995 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2996 __ li(R0, 0);
2997 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2998 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2999 __ BIND(skip_restore_excp);
3000
3001 __ pop_frame();
3002
3003 // stack: (deoptee, optional i2c, caller of deoptee, ...).
3004
3005 // pop the deoptee's frame
3006 __ pop_frame();
3007
3008 // stack: (caller_of_deoptee, ...).
3009
3010 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3011 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3012 // and the frame is effectively not resized.
3013 Register caller_sp = R23_tmp3;
3014 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3015 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3016
3017 // Loop through the `UnrollBlock' info and create interpreter frames.
3018 push_skeleton_frames(masm, true/*deopt*/,
3019 unroll_block_reg,
3020 R23_tmp3,
3021 R24_tmp4,
3022 R25_tmp5,
3023 R26_tmp6,
3024 R27_tmp7);
3025
3026 // stack: (skeletal interpreter frame, ..., optional skeletal
3027 // interpreter frame, optional c2i, caller of deoptee, ...).
3028
3029 // push an `unpack_frame' taking care of float / int return values.
3030 __ push_frame(frame_size_in_bytes, R0/*tmp*/);
3031
3032 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3033 // skeletal interpreter frame, optional c2i, caller of deoptee,
3034 // ...).
3035
3036 // Spill live volatile registers since we'll do a call.
3037 __ std( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3038 __ stfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3039
3040 // Let the unpacker layout information in the skeletal frames just
3041 // allocated.
3042 __ calculate_address_from_global_toc(R3_RET, calls_return_pc, true, true, true, true);
3043 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET);
3044 // This is a call to a LEAF method, so no oop map is required.
3045 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3046 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/);
3047 __ reset_last_Java_frame();
3048
3049 // Restore the volatiles saved above.
3050 __ ld( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3051 __ lfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3052
3053 // Pop the unpack frame.
3054 __ pop_frame();
3055 __ restore_LR(R0);
3056
3057 // stack: (top interpreter frame, ..., optional interpreter frame,
3058 // optional c2i, caller of deoptee, ...).
3059
3060 // Initialize R14_state.
3061 __ restore_interpreter_state(R11_scratch1);
3062 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3063
3064 // Return to the interpreter entry point.
3065 __ blr();
3066 __ flush();
3067 #else // COMPILER2
3068 __ unimplemented("deopt blob needed only with compiler");
3069 int exception_offset = __ pc() - start;
3070 #endif // COMPILER2
3071
3072 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset,
3073 reexecute_offset, first_frame_size_in_bytes / wordSize);
3074 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3075 }
3076
3077 #ifdef COMPILER2
3078 void OptoRuntime::generate_uncommon_trap_blob() {
3079 // Allocate space for the code.
3080 ResourceMark rm;
3081 // Setup code generation tools.
3082 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024);
3083 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
3084 address start = __ pc();
3085
3086 Register unroll_block_reg = R21_tmp1;
3087 Register klass_index_reg = R22_tmp2;
3088 Register unc_trap_reg = R23_tmp3;
3089 Register r_return_pc = R27_tmp7;
3090
3091 OopMapSet* oop_maps = new OopMapSet();
3092 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3093 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3094
3095 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3096
3097 // Push a dummy `unpack_frame' and call
3098 // `Deoptimization::uncommon_trap' to pack the compiled frame into a
3099 // vframe array and return the `UnrollBlock' information.
3100
3101 // Save LR to compiled frame.
3102 __ save_LR(R11_scratch1);
3103
3104 // Push an "uncommon_trap" frame.
3105 __ push_frame_reg_args(0, R11_scratch1);
3106
3107 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...).
3108
3109 // Set the `unpack_frame' as last_Java_frame.
3110 // `Deoptimization::uncommon_trap' expects it and considers its
3111 // sender frame as the deoptee frame.
3112 // Remember the offset of the instruction whose address will be
3113 // moved to R11_scratch1.
3114 address gc_map_pc = __ pc();
3115 __ calculate_address_from_global_toc(r_return_pc, gc_map_pc, true, true, true, true);
3116 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3117
3118 __ mr(klass_index_reg, R3);
3119 __ li(R5_ARG3, Deoptimization::Unpack_uncommon_trap);
3120 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap),
3121 R16_thread, klass_index_reg, R5_ARG3);
3122
3123 // Set an oopmap for the call site.
3124 oop_maps->add_gc_map(gc_map_pc - start, map);
3125
3126 __ reset_last_Java_frame();
3127
3128 // Pop the `unpack frame'.
3129 __ pop_frame();
3130
3131 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3132
3133 // Save the return value.
3134 __ mr(unroll_block_reg, R3_RET);
3135
3136 // Pop the uncommon_trap frame.
3137 __ pop_frame();
3138
3139 // stack: (caller_of_deoptee, ...).
3140
3141 #ifdef ASSERT
3142 __ lwz(R22_tmp2, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3143 __ cmpdi(CCR0, R22_tmp2, (unsigned)Deoptimization::Unpack_uncommon_trap);
3144 __ asm_assert_eq("OptoRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap");
3145 #endif
3146
3147 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3148 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3149 // and the frame is effectively not resized.
3150 Register caller_sp = R23_tmp3;
3151 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3152 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3153
3154 // Allocate new interpreter frame(s) and possibly a c2i adapter
3155 // frame.
3156 push_skeleton_frames(masm, false/*deopt*/,
3157 unroll_block_reg,
3158 R22_tmp2,
3159 R23_tmp3,
3160 R24_tmp4,
3161 R25_tmp5,
3162 R26_tmp6);
3163
3164 // stack: (skeletal interpreter frame, ..., optional skeletal
3165 // interpreter frame, optional c2i, caller of deoptee, ...).
3166
3167 // Push a dummy `unpack_frame' taking care of float return values.
3168 // Call `Deoptimization::unpack_frames' to layout information in the
3169 // interpreter frames just created.
3170
3171 // Push a simple "unpack frame" here.
3172 __ push_frame_reg_args(0, R11_scratch1);
3173
3174 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3175 // skeletal interpreter frame, optional c2i, caller of deoptee,
3176 // ...).
3177
3178 // Set the "unpack_frame" as last_Java_frame.
3179 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3180
3181 // Indicate it is the uncommon trap case.
3182 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap);
3183 // Let the unpacker layout information in the skeletal frames just
3184 // allocated.
3185 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3186 R16_thread, unc_trap_reg);
3187
3188 __ reset_last_Java_frame();
3189 // Pop the `unpack frame'.
3190 __ pop_frame();
3191 // Restore LR from top interpreter frame.
3192 __ restore_LR(R11_scratch1);
3193
3194 // stack: (top interpreter frame, ..., optional interpreter frame,
3195 // optional c2i, caller of deoptee, ...).
3196
3197 __ restore_interpreter_state(R11_scratch1);
3198 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3199
3200 // Return to the interpreter entry point.
3201 __ blr();
3202
3203 masm->flush();
3204
3205 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize);
3206 }
3207 #endif // COMPILER2
3208
3209 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap.
3210 SafepointBlob* SharedRuntime::generate_handler_blob(SharedStubId id, address call_ptr) {
3211 assert(StubRoutines::forward_exception_entry() != nullptr,
3212 "must be generated before");
3213 assert(is_polling_page_id(id), "expected a polling page stub id");
3214
3215 ResourceMark rm;
3216 OopMapSet *oop_maps = new OopMapSet();
3217 OopMap* map;
3218
3219 // Allocate space for the code. Setup code generation tools.
3220 const char* name = SharedRuntime::stub_name(id);
3221 CodeBuffer buffer(name, 2048, 1024);
3222 MacroAssembler* masm = new MacroAssembler(&buffer);
3223
3224 address start = __ pc();
3225 int frame_size_in_bytes = 0;
3226
3227 RegisterSaver::ReturnPCLocation return_pc_location;
3228 bool cause_return = (id == SharedStubId::polling_page_return_handler_id);
3229 if (cause_return) {
3230 // Nothing to do here. The frame has already been popped in MachEpilogNode.
3231 // Register LR already contains the return pc.
3232 return_pc_location = RegisterSaver::return_pc_is_pre_saved;
3233 } else {
3234 // Use thread()->saved_exception_pc() as return pc.
3235 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc;
3236 }
3237
3238 bool save_vectors = (id == SharedStubId::polling_page_vectors_safepoint_handler_id);
3239
3240 // Save registers, fpu state, and flags. Set R31 = return pc.
3241 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3242 &frame_size_in_bytes,
3243 /*generate_oop_map=*/ true,
3244 /*return_pc_adjustment=*/0,
3245 return_pc_location, save_vectors);
3246
3247 // The following is basically a call_VM. However, we need the precise
3248 // address of the call in order to generate an oopmap. Hence, we do all the
3249 // work ourselves.
3250 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg);
3251
3252 // The return address must always be correct so that the frame constructor
3253 // never sees an invalid pc.
3254
3255 // Do the call
3256 __ call_VM_leaf(call_ptr, R16_thread);
3257 address calls_return_pc = __ last_calls_return_pc();
3258
3259 // Set an oopmap for the call site. This oopmap will map all
3260 // oop-registers and debug-info registers as callee-saved. This
3261 // will allow deoptimization at this safepoint to find all possible
3262 // debug-info recordings, as well as let GC find all oops.
3263 oop_maps->add_gc_map(calls_return_pc - start, map);
3264
3265 Label noException;
3266
3267 // Clear the last Java frame.
3268 __ reset_last_Java_frame();
3269
3270 BLOCK_COMMENT(" Check pending exception.");
3271 const Register pending_exception = R0;
3272 __ ld(pending_exception, thread_(pending_exception));
3273 __ cmpdi(CCR0, pending_exception, 0);
3274 __ beq(CCR0, noException);
3275
3276 // Exception pending
3277 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3278 frame_size_in_bytes,
3279 /*restore_ctr=*/true, save_vectors);
3280
3281 BLOCK_COMMENT(" Jump to forward_exception_entry.");
3282 // Jump to forward_exception_entry, with the issuing PC in LR
3283 // so it looks like the original nmethod called forward_exception_entry.
3284 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3285
3286 // No exception case.
3287 __ BIND(noException);
3288
3289 if (!cause_return) {
3290 Label no_adjust;
3291 // If our stashed return pc was modified by the runtime we avoid touching it
3292 __ ld(R0, frame_size_in_bytes + _abi0(lr), R1_SP);
3293 __ cmpd(CCR0, R0, R31);
3294 __ bne(CCR0, no_adjust);
3295
3296 // Adjust return pc forward to step over the safepoint poll instruction
3297 __ addi(R31, R31, 4);
3298 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
3299
3300 __ bind(no_adjust);
3301 }
3302
3303 // Normal exit, restore registers and exit.
3304 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3305 frame_size_in_bytes,
3306 /*restore_ctr=*/true, save_vectors);
3307
3308 __ blr();
3309
3310 // Make sure all code is generated
3311 masm->flush();
3312
3313 // Fill-out other meta info
3314 // CodeBlob frame size is in words.
3315 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize);
3316 }
3317
3318 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss)
3319 //
3320 // Generate a stub that calls into the vm to find out the proper destination
3321 // of a java call. All the argument registers are live at this point
3322 // but since this is generic code we don't know what they are and the caller
3323 // must do any gc of the args.
3324 //
3325 RuntimeStub* SharedRuntime::generate_resolve_blob(SharedStubId id, address destination) {
3326 assert(is_resolve_id(id), "expected a resolve stub id");
3327
3328 // allocate space for the code
3329 ResourceMark rm;
3330
3331 const char* name = SharedRuntime::stub_name(id);
3332 CodeBuffer buffer(name, 1000, 512);
3333 MacroAssembler* masm = new MacroAssembler(&buffer);
3334
3335 int frame_size_in_bytes;
3336
3337 OopMapSet *oop_maps = new OopMapSet();
3338 OopMap* map = nullptr;
3339
3340 address start = __ pc();
3341
3342 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3343 &frame_size_in_bytes,
3344 /*generate_oop_map*/ true,
3345 /*return_pc_adjustment*/ 0,
3346 RegisterSaver::return_pc_is_lr);
3347
3348 // Use noreg as last_Java_pc, the return pc will be reconstructed
3349 // from the physical frame.
3350 __ set_last_Java_frame(/*sp*/R1_SP, noreg);
3351
3352 int frame_complete = __ offset();
3353
3354 // Pass R19_method as 2nd (optional) argument, used by
3355 // counter_overflow_stub.
3356 __ call_VM_leaf(destination, R16_thread, R19_method);
3357 address calls_return_pc = __ last_calls_return_pc();
3358 // Set an oopmap for the call site.
3359 // We need this not only for callee-saved registers, but also for volatile
3360 // registers that the compiler might be keeping live across a safepoint.
3361 // Create the oopmap for the call's return pc.
3362 oop_maps->add_gc_map(calls_return_pc - start, map);
3363
3364 // R3_RET contains the address we are going to jump to assuming no exception got installed.
3365
3366 // clear last_Java_sp
3367 __ reset_last_Java_frame();
3368
3369 // Check for pending exceptions.
3370 BLOCK_COMMENT("Check for pending exceptions.");
3371 Label pending;
3372 __ ld(R11_scratch1, thread_(pending_exception));
3373 __ cmpdi(CCR0, R11_scratch1, 0);
3374 __ bne(CCR0, pending);
3375
3376 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame.
3377
3378 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false);
3379
3380 // Get the returned method.
3381 __ get_vm_result_2(R19_method);
3382
3383 __ bctr();
3384
3385
3386 // Pending exception after the safepoint.
3387 __ BIND(pending);
3388
3389 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true);
3390
3391 // exception pending => remove activation and forward to exception handler
3392
3393 __ li(R11_scratch1, 0);
3394 __ ld(R3_ARG1, thread_(pending_exception));
3395 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_offset()), R16_thread);
3396 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3397
3398 // -------------
3399 // Make sure all code is generated.
3400 masm->flush();
3401
3402 // return the blob
3403 // frame_size_words or bytes??
3404 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize,
3405 oop_maps, true);
3406 }
3407
3408 // Continuation point for throwing of implicit exceptions that are
3409 // not handled in the current activation. Fabricates an exception
3410 // oop and initiates normal exception dispatching in this
3411 // frame. Only callee-saved registers are preserved (through the
3412 // normal register window / RegisterMap handling). If the compiler
3413 // needs all registers to be preserved between the fault point and
3414 // the exception handler then it must assume responsibility for that
3415 // in AbstractCompiler::continuation_for_implicit_null_exception or
3416 // continuation_for_implicit_division_by_zero_exception. All other
3417 // implicit exceptions (e.g., NullPointerException or
3418 // AbstractMethodError on entry) are either at call sites or
3419 // otherwise assume that stack unwinding will be initiated, so
3420 // caller saved registers were assumed volatile in the compiler.
3421 //
3422 // Note that we generate only this stub into a RuntimeStub, because
3423 // it needs to be properly traversed and ignored during GC, so we
3424 // change the meaning of the "__" macro within this method.
3425 //
3426 // Note: the routine set_pc_not_at_call_for_caller in
3427 // SharedRuntime.cpp requires that this code be generated into a
3428 // RuntimeStub.
3429 RuntimeStub* SharedRuntime::generate_throw_exception(SharedStubId id, address runtime_entry) {
3430 assert(is_throw_id(id), "expected a throw stub id");
3431
3432 const char* name = SharedRuntime::stub_name(id);
3433
3434 ResourceMark rm;
3435 const char* timer_msg = "SharedRuntime generate_throw_exception";
3436 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3437
3438 CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0);
3439 MacroAssembler* masm = new MacroAssembler(&code);
3440
3441 OopMapSet* oop_maps = new OopMapSet();
3442 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3443 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3444
3445 address start = __ pc();
3446
3447 __ save_LR(R11_scratch1);
3448
3449 // Push a frame.
3450 __ push_frame_reg_args(0, R11_scratch1);
3451
3452 address frame_complete_pc = __ pc();
3453
3454 // Note that we always have a runtime stub frame on the top of
3455 // stack by this point. Remember the offset of the instruction
3456 // whose address will be moved to R11_scratch1.
3457 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1);
3458
3459 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
3460
3461 __ mr(R3_ARG1, R16_thread);
3462 __ call_c(runtime_entry);
3463
3464 // Set an oopmap for the call site.
3465 oop_maps->add_gc_map((int)(gc_map_pc - start), map);
3466
3467 __ reset_last_Java_frame();
3468
3469 #ifdef ASSERT
3470 // Make sure that this code is only executed if there is a pending
3471 // exception.
3472 {
3473 Label L;
3474 __ ld(R0,
3475 in_bytes(Thread::pending_exception_offset()),
3476 R16_thread);
3477 __ cmpdi(CCR0, R0, 0);
3478 __ bne(CCR0, L);
3479 __ stop("SharedRuntime::throw_exception: no pending exception");
3480 __ bind(L);
3481 }
3482 #endif
3483
3484 // Pop frame.
3485 __ pop_frame();
3486
3487 __ restore_LR(R11_scratch1);
3488
3489 __ load_const(R11_scratch1, StubRoutines::forward_exception_entry());
3490 __ mtctr(R11_scratch1);
3491 __ bctr();
3492
3493 // Create runtime stub with OopMap.
3494 RuntimeStub* stub =
3495 RuntimeStub::new_runtime_stub(name, &code,
3496 /*frame_complete=*/ (int)(frame_complete_pc - start),
3497 frame_size_in_bytes/wordSize,
3498 oop_maps,
3499 false);
3500 return stub;
3501 }
3502
3503 //------------------------------Montgomery multiplication------------------------
3504 //
3505
3506 // Subtract 0:b from carry:a. Return carry.
3507 static unsigned long
3508 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) {
3509 long i = 0;
3510 unsigned long tmp, tmp2;
3511 __asm__ __volatile__ (
3512 "subfc %[tmp], %[tmp], %[tmp] \n" // pre-set CA
3513 "mtctr %[len] \n"
3514 "0: \n"
3515 "ldx %[tmp], %[i], %[a] \n"
3516 "ldx %[tmp2], %[i], %[b] \n"
3517 "subfe %[tmp], %[tmp2], %[tmp] \n" // subtract extended
3518 "stdx %[tmp], %[i], %[a] \n"
3519 "addi %[i], %[i], 8 \n"
3520 "bdnz 0b \n"
3521 "addme %[tmp], %[carry] \n" // carry + CA - 1
3522 : [i]"+b"(i), [tmp]"=&r"(tmp), [tmp2]"=&r"(tmp2)
3523 : [a]"r"(a), [b]"r"(b), [carry]"r"(carry), [len]"r"(len)
3524 : "ctr", "xer", "memory"
3525 );
3526 return tmp;
3527 }
3528
3529 // Multiply (unsigned) Long A by Long B, accumulating the double-
3530 // length result into the accumulator formed of T0, T1, and T2.
3531 inline void MACC(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3532 unsigned long hi, lo;
3533 __asm__ __volatile__ (
3534 "mulld %[lo], %[A], %[B] \n"
3535 "mulhdu %[hi], %[A], %[B] \n"
3536 "addc %[T0], %[T0], %[lo] \n"
3537 "adde %[T1], %[T1], %[hi] \n"
3538 "addze %[T2], %[T2] \n"
3539 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3540 : [A]"r"(A), [B]"r"(B)
3541 : "xer"
3542 );
3543 }
3544
3545 // As above, but add twice the double-length result into the
3546 // accumulator.
3547 inline void MACC2(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3548 unsigned long hi, lo;
3549 __asm__ __volatile__ (
3550 "mulld %[lo], %[A], %[B] \n"
3551 "mulhdu %[hi], %[A], %[B] \n"
3552 "addc %[T0], %[T0], %[lo] \n"
3553 "adde %[T1], %[T1], %[hi] \n"
3554 "addze %[T2], %[T2] \n"
3555 "addc %[T0], %[T0], %[lo] \n"
3556 "adde %[T1], %[T1], %[hi] \n"
3557 "addze %[T2], %[T2] \n"
3558 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3559 : [A]"r"(A), [B]"r"(B)
3560 : "xer"
3561 );
3562 }
3563
3564 // Fast Montgomery multiplication. The derivation of the algorithm is
3565 // in "A Cryptographic Library for the Motorola DSP56000,
3566 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237".
3567 static void
3568 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[],
3569 unsigned long m[], unsigned long inv, int len) {
3570 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3571 int i;
3572
3573 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3574
3575 for (i = 0; i < len; i++) {
3576 int j;
3577 for (j = 0; j < i; j++) {
3578 MACC(a[j], b[i-j], t0, t1, t2);
3579 MACC(m[j], n[i-j], t0, t1, t2);
3580 }
3581 MACC(a[i], b[0], t0, t1, t2);
3582 m[i] = t0 * inv;
3583 MACC(m[i], n[0], t0, t1, t2);
3584
3585 assert(t0 == 0, "broken Montgomery multiply");
3586
3587 t0 = t1; t1 = t2; t2 = 0;
3588 }
3589
3590 for (i = len; i < 2*len; i++) {
3591 int j;
3592 for (j = i-len+1; j < len; j++) {
3593 MACC(a[j], b[i-j], t0, t1, t2);
3594 MACC(m[j], n[i-j], t0, t1, t2);
3595 }
3596 m[i-len] = t0;
3597 t0 = t1; t1 = t2; t2 = 0;
3598 }
3599
3600 while (t0) {
3601 t0 = sub(m, n, t0, len);
3602 }
3603 }
3604
3605 // Fast Montgomery squaring. This uses asymptotically 25% fewer
3606 // multiplies so it should be up to 25% faster than Montgomery
3607 // multiplication. However, its loop control is more complex and it
3608 // may actually run slower on some machines.
3609 static void
3610 montgomery_square(unsigned long a[], unsigned long n[],
3611 unsigned long m[], unsigned long inv, int len) {
3612 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3613 int i;
3614
3615 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3616
3617 for (i = 0; i < len; i++) {
3618 int j;
3619 int end = (i+1)/2;
3620 for (j = 0; j < end; j++) {
3621 MACC2(a[j], a[i-j], t0, t1, t2);
3622 MACC(m[j], n[i-j], t0, t1, t2);
3623 }
3624 if ((i & 1) == 0) {
3625 MACC(a[j], a[j], t0, t1, t2);
3626 }
3627 for (; j < i; j++) {
3628 MACC(m[j], n[i-j], t0, t1, t2);
3629 }
3630 m[i] = t0 * inv;
3631 MACC(m[i], n[0], t0, t1, t2);
3632
3633 assert(t0 == 0, "broken Montgomery square");
3634
3635 t0 = t1; t1 = t2; t2 = 0;
3636 }
3637
3638 for (i = len; i < 2*len; i++) {
3639 int start = i-len+1;
3640 int end = start + (len - start)/2;
3641 int j;
3642 for (j = start; j < end; j++) {
3643 MACC2(a[j], a[i-j], t0, t1, t2);
3644 MACC(m[j], n[i-j], t0, t1, t2);
3645 }
3646 if ((i & 1) == 0) {
3647 MACC(a[j], a[j], t0, t1, t2);
3648 }
3649 for (; j < len; j++) {
3650 MACC(m[j], n[i-j], t0, t1, t2);
3651 }
3652 m[i-len] = t0;
3653 t0 = t1; t1 = t2; t2 = 0;
3654 }
3655
3656 while (t0) {
3657 t0 = sub(m, n, t0, len);
3658 }
3659 }
3660
3661 // The threshold at which squaring is advantageous was determined
3662 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz.
3663 // Doesn't seem to be relevant for Power8 so we use the same value.
3664 #define MONTGOMERY_SQUARING_THRESHOLD 64
3665
3666 // Copy len longwords from s to d, word-swapping as we go. The
3667 // destination array is reversed.
3668 static void reverse_words(unsigned long *s, unsigned long *d, int len) {
3669 d += len;
3670 while(len-- > 0) {
3671 d--;
3672 unsigned long s_val = *s;
3673 // Swap words in a longword on little endian machines.
3674 #ifdef VM_LITTLE_ENDIAN
3675 s_val = (s_val << 32) | (s_val >> 32);
3676 #endif
3677 *d = s_val;
3678 s++;
3679 }
3680 }
3681
3682 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints,
3683 jint len, jlong inv,
3684 jint *m_ints) {
3685 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3686 assert(len % 2 == 0, "array length in montgomery_multiply must be even");
3687 int longwords = len/2;
3688
3689 // Make very sure we don't use so much space that the stack might
3690 // overflow. 512 jints corresponds to an 16384-bit integer and
3691 // will use here a total of 8k bytes of stack space.
3692 int divisor = sizeof(unsigned long) * 4;
3693 guarantee(longwords <= 8192 / divisor, "must be");
3694 int total_allocation = longwords * sizeof (unsigned long) * 4;
3695 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3696
3697 // Local scratch arrays
3698 unsigned long
3699 *a = scratch + 0 * longwords,
3700 *b = scratch + 1 * longwords,
3701 *n = scratch + 2 * longwords,
3702 *m = scratch + 3 * longwords;
3703
3704 reverse_words((unsigned long *)a_ints, a, longwords);
3705 reverse_words((unsigned long *)b_ints, b, longwords);
3706 reverse_words((unsigned long *)n_ints, n, longwords);
3707
3708 ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords);
3709
3710 reverse_words(m, (unsigned long *)m_ints, longwords);
3711 }
3712
3713 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints,
3714 jint len, jlong inv,
3715 jint *m_ints) {
3716 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3717 assert(len % 2 == 0, "array length in montgomery_square must be even");
3718 int longwords = len/2;
3719
3720 // Make very sure we don't use so much space that the stack might
3721 // overflow. 512 jints corresponds to an 16384-bit integer and
3722 // will use here a total of 6k bytes of stack space.
3723 int divisor = sizeof(unsigned long) * 3;
3724 guarantee(longwords <= (8192 / divisor), "must be");
3725 int total_allocation = longwords * sizeof (unsigned long) * 3;
3726 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3727
3728 // Local scratch arrays
3729 unsigned long
3730 *a = scratch + 0 * longwords,
3731 *n = scratch + 1 * longwords,
3732 *m = scratch + 2 * longwords;
3733
3734 reverse_words((unsigned long *)a_ints, a, longwords);
3735 reverse_words((unsigned long *)n_ints, n, longwords);
3736
3737 if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3738 ::montgomery_square(a, n, m, (unsigned long)inv, longwords);
3739 } else {
3740 ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords);
3741 }
3742
3743 reverse_words(m, (unsigned long *)m_ints, longwords);
3744 }
3745
3746 #if INCLUDE_JFR
3747
3748 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3749 // It returns a jobject handle to the event writer.
3750 // The handle is dereferenced and the return value is the event writer oop.
3751 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3752 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_write_checkpoint_id);
3753 CodeBuffer code(name, 512, 64);
3754 MacroAssembler* masm = new MacroAssembler(&code);
3755
3756 Register tmp1 = R10_ARG8;
3757 Register tmp2 = R9_ARG7;
3758
3759 int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3760 address start = __ pc();
3761 __ mflr(tmp1);
3762 __ std(tmp1, _abi0(lr), R1_SP); // save return pc
3763 __ push_frame_reg_args(0, tmp1);
3764 int frame_complete = __ pc() - start;
3765 __ set_last_Java_frame(R1_SP, noreg);
3766 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), R16_thread);
3767 address calls_return_pc = __ last_calls_return_pc();
3768 __ reset_last_Java_frame();
3769 // The handle is dereferenced through a load barrier.
3770 __ resolve_global_jobject(R3_RET, tmp1, tmp2, MacroAssembler::PRESERVATION_NONE);
3771 __ pop_frame();
3772 __ ld(tmp1, _abi0(lr), R1_SP);
3773 __ mtlr(tmp1);
3774 __ blr();
3775
3776 OopMapSet* oop_maps = new OopMapSet();
3777 OopMap* map = new OopMap(framesize, 0);
3778 oop_maps->add_gc_map(calls_return_pc - start, map);
3779
3780 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3781 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3782 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3783 oop_maps, false);
3784 return stub;
3785 }
3786
3787 // For c2: call to return a leased buffer.
3788 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
3789 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_return_lease_id);
3790 CodeBuffer code(name, 512, 64);
3791 MacroAssembler* masm = new MacroAssembler(&code);
3792
3793 Register tmp1 = R10_ARG8;
3794 Register tmp2 = R9_ARG7;
3795
3796 int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3797 address start = __ pc();
3798 __ mflr(tmp1);
3799 __ std(tmp1, _abi0(lr), R1_SP); // save return pc
3800 __ push_frame_reg_args(0, tmp1);
3801 int frame_complete = __ pc() - start;
3802 __ set_last_Java_frame(R1_SP, noreg);
3803 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), R16_thread);
3804 address calls_return_pc = __ last_calls_return_pc();
3805 __ reset_last_Java_frame();
3806 __ pop_frame();
3807 __ ld(tmp1, _abi0(lr), R1_SP);
3808 __ mtlr(tmp1);
3809 __ blr();
3810
3811 OopMapSet* oop_maps = new OopMapSet();
3812 OopMap* map = new OopMap(framesize, 0);
3813 oop_maps->add_gc_map(calls_return_pc - start, map);
3814
3815 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3816 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3817 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3818 oop_maps, false);
3819 return stub;
3820 }
3821
3822 #endif // INCLUDE_JFR