1 /*
2 * Copyright (c) 2003, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #ifndef _WINDOWS
27 #include "alloca.h"
28 #endif
29 #include "asm/macroAssembler.hpp"
30 #include "asm/macroAssembler.inline.hpp"
31 #include "classfile/symbolTable.hpp"
32 #include "code/compiledIC.hpp"
33 #include "code/debugInfoRec.hpp"
34 #include "code/nativeInst.hpp"
35 #include "code/vtableStubs.hpp"
36 #include "compiler/oopMap.hpp"
37 #include "gc/shared/collectedHeap.hpp"
38 #include "gc/shared/gcLocker.hpp"
39 #include "gc/shared/barrierSet.hpp"
40 #include "gc/shared/barrierSetAssembler.hpp"
41 #include "interpreter/interpreter.hpp"
42 #include "logging/log.hpp"
43 #include "memory/resourceArea.hpp"
44 #include "memory/universe.hpp"
45 #include "oops/klass.inline.hpp"
46 #include "oops/method.inline.hpp"
47 #include "prims/methodHandles.hpp"
48 #include "runtime/continuation.hpp"
49 #include "runtime/continuationEntry.inline.hpp"
50 #include "runtime/globals.hpp"
51 #include "runtime/jniHandles.hpp"
52 #include "runtime/safepointMechanism.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "runtime/signature.hpp"
55 #include "runtime/stubRoutines.hpp"
56 #include "runtime/timerTrace.hpp"
57 #include "runtime/vframeArray.hpp"
58 #include "runtime/vm_version.hpp"
59 #include "utilities/align.hpp"
60 #include "utilities/checkedCast.hpp"
61 #include "utilities/formatBuffer.hpp"
62 #include "vmreg_x86.inline.hpp"
63 #ifdef COMPILER1
64 #include "c1/c1_Runtime1.hpp"
65 #endif
66 #ifdef COMPILER2
67 #include "opto/runtime.hpp"
68 #endif
69 #if INCLUDE_JVMCI
70 #include "jvmci/jvmciJavaClasses.hpp"
71 #endif
72
73 #define __ masm->
74
75 #ifdef PRODUCT
76 #define BLOCK_COMMENT(str) /* nothing */
77 #else
78 #define BLOCK_COMMENT(str) __ block_comment(str)
79 #endif // PRODUCT
80
81 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
82
83 class RegisterSaver {
84 // Capture info about frame layout. Layout offsets are in jint
85 // units because compiler frame slots are jints.
86 #define XSAVE_AREA_BEGIN 160
87 #define XSAVE_AREA_YMM_BEGIN 576
88 #define XSAVE_AREA_EGPRS 960
89 #define XSAVE_AREA_OPMASK_BEGIN 1088
90 #define XSAVE_AREA_ZMM_BEGIN 1152
91 #define XSAVE_AREA_UPPERBANK 1664
92 #define DEF_XMM_OFFS(regnum) xmm ## regnum ## _off = xmm_off + (regnum)*16/BytesPerInt, xmm ## regnum ## H_off
93 #define DEF_YMM_OFFS(regnum) ymm ## regnum ## _off = ymm_off + (regnum)*16/BytesPerInt, ymm ## regnum ## H_off
94 #define DEF_ZMM_OFFS(regnum) zmm ## regnum ## _off = zmm_off + (regnum)*32/BytesPerInt, zmm ## regnum ## H_off
95 #define DEF_OPMASK_OFFS(regnum) opmask ## regnum ## _off = opmask_off + (regnum)*8/BytesPerInt, opmask ## regnum ## H_off
96 #define DEF_ZMM_UPPER_OFFS(regnum) zmm ## regnum ## _off = zmm_upper_off + (regnum-16)*64/BytesPerInt, zmm ## regnum ## H_off
97 enum layout {
98 fpu_state_off = frame::arg_reg_save_area_bytes/BytesPerInt, // fxsave save area
99 xmm_off = fpu_state_off + XSAVE_AREA_BEGIN/BytesPerInt, // offset in fxsave save area
100 DEF_XMM_OFFS(0),
101 DEF_XMM_OFFS(1),
102 // 2..15 are implied in range usage
103 ymm_off = xmm_off + (XSAVE_AREA_YMM_BEGIN - XSAVE_AREA_BEGIN)/BytesPerInt,
104 DEF_YMM_OFFS(0),
105 DEF_YMM_OFFS(1),
106 // 2..15 are implied in range usage
107 r31_off = xmm_off + (XSAVE_AREA_EGPRS - XSAVE_AREA_BEGIN)/BytesPerInt,
108 r31H_off,
109 r30_off, r30H_off,
110 r29_off, r29H_off,
111 r28_off, r28H_off,
112 r27_off, r27H_off,
113 r26_off, r26H_off,
114 r25_off, r25H_off,
115 r24_off, r24H_off,
116 r23_off, r23H_off,
117 r22_off, r22H_off,
118 r21_off, r21H_off,
119 r20_off, r20H_off,
120 r19_off, r19H_off,
121 r18_off, r18H_off,
122 r17_off, r17H_off,
123 r16_off, r16H_off,
124 opmask_off = xmm_off + (XSAVE_AREA_OPMASK_BEGIN - XSAVE_AREA_BEGIN)/BytesPerInt,
125 DEF_OPMASK_OFFS(0),
126 DEF_OPMASK_OFFS(1),
127 // 2..7 are implied in range usage
128 zmm_off = xmm_off + (XSAVE_AREA_ZMM_BEGIN - XSAVE_AREA_BEGIN)/BytesPerInt,
129 DEF_ZMM_OFFS(0),
130 DEF_ZMM_OFFS(1),
131 zmm_upper_off = xmm_off + (XSAVE_AREA_UPPERBANK - XSAVE_AREA_BEGIN)/BytesPerInt,
132 DEF_ZMM_UPPER_OFFS(16),
133 DEF_ZMM_UPPER_OFFS(17),
134 // 18..31 are implied in range usage
135 fpu_state_end = fpu_state_off + ((FPUStateSizeInWords-1)*wordSize / BytesPerInt),
136 fpu_stateH_end,
137 r15_off, r15H_off,
138 r14_off, r14H_off,
139 r13_off, r13H_off,
140 r12_off, r12H_off,
141 r11_off, r11H_off,
142 r10_off, r10H_off,
143 r9_off, r9H_off,
144 r8_off, r8H_off,
145 rdi_off, rdiH_off,
146 rsi_off, rsiH_off,
147 ignore_off, ignoreH_off, // extra copy of rbp
148 rsp_off, rspH_off,
149 rbx_off, rbxH_off,
150 rdx_off, rdxH_off,
151 rcx_off, rcxH_off,
152 rax_off, raxH_off,
153 // 16-byte stack alignment fill word: see MacroAssembler::push/pop_IU_state
154 align_off, alignH_off,
155 flags_off, flagsH_off,
156 // The frame sender code expects that rbp will be in the "natural" place and
157 // will override any oopMap setting for it. We must therefore force the layout
158 // so that it agrees with the frame sender code.
159 rbp_off, rbpH_off, // copy of rbp we will restore
160 return_off, returnH_off, // slot for return address
161 reg_save_size // size in compiler stack slots
162 };
163
164 public:
165 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_wide_vectors);
166 static void restore_live_registers(MacroAssembler* masm, bool restore_wide_vectors = false);
167
168 // Offsets into the register save area
169 // Used by deoptimization when it is managing result register
170 // values on its own
171
172 static int rax_offset_in_bytes(void) { return BytesPerInt * rax_off; }
173 static int rdx_offset_in_bytes(void) { return BytesPerInt * rdx_off; }
174 static int rbx_offset_in_bytes(void) { return BytesPerInt * rbx_off; }
175 static int xmm0_offset_in_bytes(void) { return BytesPerInt * xmm0_off; }
176 static int return_offset_in_bytes(void) { return BytesPerInt * return_off; }
177
178 // During deoptimization only the result registers need to be restored,
179 // all the other values have already been extracted.
180 static void restore_result_registers(MacroAssembler* masm);
181 };
182
183 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words, bool save_wide_vectors) {
184 int off = 0;
185 int num_xmm_regs = XMMRegister::available_xmm_registers();
186 #if COMPILER2_OR_JVMCI
187 if (save_wide_vectors && UseAVX == 0) {
188 save_wide_vectors = false; // vectors larger than 16 byte long are supported only with AVX
189 }
190 assert(!save_wide_vectors || MaxVectorSize <= 64, "Only up to 64 byte long vectors are supported");
191 #else
192 save_wide_vectors = false; // vectors are generated only by C2 and JVMCI
193 #endif
194
195 // Always make the frame size 16-byte aligned, both vector and non vector stacks are always allocated
196 int frame_size_in_bytes = align_up(reg_save_size*BytesPerInt, num_xmm_regs);
197 // OopMap frame size is in compiler stack slots (jint's) not bytes or words
198 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
199 // CodeBlob frame size is in words.
200 int frame_size_in_words = frame_size_in_bytes / wordSize;
201 *total_frame_words = frame_size_in_words;
202
203 // Save registers, fpu state, and flags.
204 // We assume caller has already pushed the return address onto the
205 // stack, so rsp is 8-byte aligned here.
206 // We push rpb twice in this sequence because we want the real rbp
207 // to be under the return like a normal enter.
208
209 __ enter(); // rsp becomes 16-byte aligned here
210 __ pushf();
211 // Make sure rsp stays 16-byte aligned
212 __ subq(rsp, 8);
213 // Push CPU state in multiple of 16 bytes
214 __ save_legacy_gprs();
215 __ push_FPU_state();
216
217
218 // push cpu state handles this on EVEX enabled targets
219 if (save_wide_vectors) {
220 // Save upper half of YMM registers(0..15)
221 int base_addr = XSAVE_AREA_YMM_BEGIN;
222 for (int n = 0; n < 16; n++) {
223 __ vextractf128_high(Address(rsp, base_addr+n*16), as_XMMRegister(n));
224 }
225 if (VM_Version::supports_evex()) {
226 // Save upper half of ZMM registers(0..15)
227 base_addr = XSAVE_AREA_ZMM_BEGIN;
228 for (int n = 0; n < 16; n++) {
229 __ vextractf64x4_high(Address(rsp, base_addr+n*32), as_XMMRegister(n));
230 }
231 // Save full ZMM registers(16..num_xmm_regs)
232 base_addr = XSAVE_AREA_UPPERBANK;
233 off = 0;
234 int vector_len = Assembler::AVX_512bit;
235 for (int n = 16; n < num_xmm_regs; n++) {
236 __ evmovdqul(Address(rsp, base_addr+(off++*64)), as_XMMRegister(n), vector_len);
237 }
238 #if COMPILER2_OR_JVMCI
239 base_addr = XSAVE_AREA_OPMASK_BEGIN;
240 off = 0;
241 for(int n = 0; n < KRegister::number_of_registers; n++) {
242 __ kmov(Address(rsp, base_addr+(off++*8)), as_KRegister(n));
243 }
244 #endif
245 }
246 } else {
247 if (VM_Version::supports_evex()) {
248 // Save upper bank of XMM registers(16..31) for scalar or 16-byte vector usage
249 int base_addr = XSAVE_AREA_UPPERBANK;
250 off = 0;
251 int vector_len = VM_Version::supports_avx512vl() ? Assembler::AVX_128bit : Assembler::AVX_512bit;
252 for (int n = 16; n < num_xmm_regs; n++) {
253 __ evmovdqul(Address(rsp, base_addr+(off++*64)), as_XMMRegister(n), vector_len);
254 }
255 #if COMPILER2_OR_JVMCI
256 base_addr = XSAVE_AREA_OPMASK_BEGIN;
257 off = 0;
258 for(int n = 0; n < KRegister::number_of_registers; n++) {
259 __ kmov(Address(rsp, base_addr+(off++*8)), as_KRegister(n));
260 }
261 #endif
262 }
263 }
264
265 #if COMPILER2_OR_JVMCI
266 if (UseAPX) {
267 int base_addr = XSAVE_AREA_EGPRS;
268 off = 0;
269 for(int n = 16; n < Register::number_of_registers; n++) {
270 __ movq(Address(rsp, base_addr+(off++*8)), as_Register(n));
271 }
272 }
273 #endif
274
275 __ vzeroupper();
276 if (frame::arg_reg_save_area_bytes != 0) {
277 // Allocate argument register save area
278 __ subptr(rsp, frame::arg_reg_save_area_bytes);
279 }
280
281 // Set an oopmap for the call site. This oopmap will map all
282 // oop-registers and debug-info registers as callee-saved. This
283 // will allow deoptimization at this safepoint to find all possible
284 // debug-info recordings, as well as let GC find all oops.
285
286 OopMapSet *oop_maps = new OopMapSet();
287 OopMap* map = new OopMap(frame_size_in_slots, 0);
288
289 #define STACK_OFFSET(x) VMRegImpl::stack2reg((x))
290
291 map->set_callee_saved(STACK_OFFSET( rax_off ), rax->as_VMReg());
292 map->set_callee_saved(STACK_OFFSET( rcx_off ), rcx->as_VMReg());
293 map->set_callee_saved(STACK_OFFSET( rdx_off ), rdx->as_VMReg());
294 map->set_callee_saved(STACK_OFFSET( rbx_off ), rbx->as_VMReg());
295 // rbp location is known implicitly by the frame sender code, needs no oopmap
296 // and the location where rbp was saved by is ignored
297 map->set_callee_saved(STACK_OFFSET( rsi_off ), rsi->as_VMReg());
298 map->set_callee_saved(STACK_OFFSET( rdi_off ), rdi->as_VMReg());
299 map->set_callee_saved(STACK_OFFSET( r8_off ), r8->as_VMReg());
300 map->set_callee_saved(STACK_OFFSET( r9_off ), r9->as_VMReg());
301 map->set_callee_saved(STACK_OFFSET( r10_off ), r10->as_VMReg());
302 map->set_callee_saved(STACK_OFFSET( r11_off ), r11->as_VMReg());
303 map->set_callee_saved(STACK_OFFSET( r12_off ), r12->as_VMReg());
304 map->set_callee_saved(STACK_OFFSET( r13_off ), r13->as_VMReg());
305 map->set_callee_saved(STACK_OFFSET( r14_off ), r14->as_VMReg());
306 map->set_callee_saved(STACK_OFFSET( r15_off ), r15->as_VMReg());
307
308 if (UseAPX) {
309 map->set_callee_saved(STACK_OFFSET( r16_off ), r16->as_VMReg());
310 map->set_callee_saved(STACK_OFFSET( r17_off ), r17->as_VMReg());
311 map->set_callee_saved(STACK_OFFSET( r18_off ), r18->as_VMReg());
312 map->set_callee_saved(STACK_OFFSET( r19_off ), r19->as_VMReg());
313 map->set_callee_saved(STACK_OFFSET( r20_off ), r20->as_VMReg());
314 map->set_callee_saved(STACK_OFFSET( r21_off ), r21->as_VMReg());
315 map->set_callee_saved(STACK_OFFSET( r22_off ), r22->as_VMReg());
316 map->set_callee_saved(STACK_OFFSET( r23_off ), r23->as_VMReg());
317 map->set_callee_saved(STACK_OFFSET( r24_off ), r24->as_VMReg());
318 map->set_callee_saved(STACK_OFFSET( r25_off ), r25->as_VMReg());
319 map->set_callee_saved(STACK_OFFSET( r26_off ), r26->as_VMReg());
320 map->set_callee_saved(STACK_OFFSET( r27_off ), r27->as_VMReg());
321 map->set_callee_saved(STACK_OFFSET( r28_off ), r28->as_VMReg());
322 map->set_callee_saved(STACK_OFFSET( r29_off ), r29->as_VMReg());
323 map->set_callee_saved(STACK_OFFSET( r30_off ), r30->as_VMReg());
324 map->set_callee_saved(STACK_OFFSET( r31_off ), r31->as_VMReg());
325 }
326 // For both AVX and EVEX we will use the legacy FXSAVE area for xmm0..xmm15,
327 // on EVEX enabled targets, we get it included in the xsave area
328 off = xmm0_off;
329 int delta = xmm1_off - off;
330 for (int n = 0; n < 16; n++) {
331 XMMRegister xmm_name = as_XMMRegister(n);
332 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg());
333 off += delta;
334 }
335 if (UseAVX > 2) {
336 // Obtain xmm16..xmm31 from the XSAVE area on EVEX enabled targets
337 off = zmm16_off;
338 delta = zmm17_off - off;
339 for (int n = 16; n < num_xmm_regs; n++) {
340 XMMRegister zmm_name = as_XMMRegister(n);
341 map->set_callee_saved(STACK_OFFSET(off), zmm_name->as_VMReg());
342 off += delta;
343 }
344 }
345
346 #if COMPILER2_OR_JVMCI
347 if (save_wide_vectors) {
348 // Save upper half of YMM registers(0..15)
349 off = ymm0_off;
350 delta = ymm1_off - ymm0_off;
351 for (int n = 0; n < 16; n++) {
352 XMMRegister ymm_name = as_XMMRegister(n);
353 map->set_callee_saved(STACK_OFFSET(off), ymm_name->as_VMReg()->next(4));
354 off += delta;
355 }
356 if (VM_Version::supports_evex()) {
357 // Save upper half of ZMM registers(0..15)
358 off = zmm0_off;
359 delta = zmm1_off - zmm0_off;
360 for (int n = 0; n < 16; n++) {
361 XMMRegister zmm_name = as_XMMRegister(n);
362 map->set_callee_saved(STACK_OFFSET(off), zmm_name->as_VMReg()->next(8));
363 off += delta;
364 }
365 }
366 }
367 #endif // COMPILER2_OR_JVMCI
368
369 // %%% These should all be a waste but we'll keep things as they were for now
370 if (true) {
371 map->set_callee_saved(STACK_OFFSET( raxH_off ), rax->as_VMReg()->next());
372 map->set_callee_saved(STACK_OFFSET( rcxH_off ), rcx->as_VMReg()->next());
373 map->set_callee_saved(STACK_OFFSET( rdxH_off ), rdx->as_VMReg()->next());
374 map->set_callee_saved(STACK_OFFSET( rbxH_off ), rbx->as_VMReg()->next());
375 // rbp location is known implicitly by the frame sender code, needs no oopmap
376 map->set_callee_saved(STACK_OFFSET( rsiH_off ), rsi->as_VMReg()->next());
377 map->set_callee_saved(STACK_OFFSET( rdiH_off ), rdi->as_VMReg()->next());
378 map->set_callee_saved(STACK_OFFSET( r8H_off ), r8->as_VMReg()->next());
379 map->set_callee_saved(STACK_OFFSET( r9H_off ), r9->as_VMReg()->next());
380 map->set_callee_saved(STACK_OFFSET( r10H_off ), r10->as_VMReg()->next());
381 map->set_callee_saved(STACK_OFFSET( r11H_off ), r11->as_VMReg()->next());
382 map->set_callee_saved(STACK_OFFSET( r12H_off ), r12->as_VMReg()->next());
383 map->set_callee_saved(STACK_OFFSET( r13H_off ), r13->as_VMReg()->next());
384 map->set_callee_saved(STACK_OFFSET( r14H_off ), r14->as_VMReg()->next());
385 map->set_callee_saved(STACK_OFFSET( r15H_off ), r15->as_VMReg()->next());
386 if (UseAPX) {
387 map->set_callee_saved(STACK_OFFSET( r16H_off ), r16->as_VMReg()->next());
388 map->set_callee_saved(STACK_OFFSET( r17H_off ), r17->as_VMReg()->next());
389 map->set_callee_saved(STACK_OFFSET( r18H_off ), r18->as_VMReg()->next());
390 map->set_callee_saved(STACK_OFFSET( r19H_off ), r19->as_VMReg()->next());
391 map->set_callee_saved(STACK_OFFSET( r20H_off ), r20->as_VMReg()->next());
392 map->set_callee_saved(STACK_OFFSET( r21H_off ), r21->as_VMReg()->next());
393 map->set_callee_saved(STACK_OFFSET( r22H_off ), r22->as_VMReg()->next());
394 map->set_callee_saved(STACK_OFFSET( r23H_off ), r23->as_VMReg()->next());
395 map->set_callee_saved(STACK_OFFSET( r24H_off ), r24->as_VMReg()->next());
396 map->set_callee_saved(STACK_OFFSET( r25H_off ), r25->as_VMReg()->next());
397 map->set_callee_saved(STACK_OFFSET( r26H_off ), r26->as_VMReg()->next());
398 map->set_callee_saved(STACK_OFFSET( r27H_off ), r27->as_VMReg()->next());
399 map->set_callee_saved(STACK_OFFSET( r28H_off ), r28->as_VMReg()->next());
400 map->set_callee_saved(STACK_OFFSET( r29H_off ), r29->as_VMReg()->next());
401 map->set_callee_saved(STACK_OFFSET( r30H_off ), r30->as_VMReg()->next());
402 map->set_callee_saved(STACK_OFFSET( r31H_off ), r31->as_VMReg()->next());
403 }
404 // For both AVX and EVEX we will use the legacy FXSAVE area for xmm0..xmm15,
405 // on EVEX enabled targets, we get it included in the xsave area
406 off = xmm0H_off;
407 delta = xmm1H_off - off;
408 for (int n = 0; n < 16; n++) {
409 XMMRegister xmm_name = as_XMMRegister(n);
410 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg()->next());
411 off += delta;
412 }
413 if (UseAVX > 2) {
414 // Obtain xmm16..xmm31 from the XSAVE area on EVEX enabled targets
415 off = zmm16H_off;
416 delta = zmm17H_off - off;
417 for (int n = 16; n < num_xmm_regs; n++) {
418 XMMRegister zmm_name = as_XMMRegister(n);
419 map->set_callee_saved(STACK_OFFSET(off), zmm_name->as_VMReg()->next());
420 off += delta;
421 }
422 }
423 }
424
425 return map;
426 }
427
428 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_wide_vectors) {
429 int num_xmm_regs = XMMRegister::available_xmm_registers();
430 if (frame::arg_reg_save_area_bytes != 0) {
431 // Pop arg register save area
432 __ addptr(rsp, frame::arg_reg_save_area_bytes);
433 }
434
435 #if COMPILER2_OR_JVMCI
436 if (restore_wide_vectors) {
437 assert(UseAVX > 0, "Vectors larger than 16 byte long are supported only with AVX");
438 assert(MaxVectorSize <= 64, "Only up to 64 byte long vectors are supported");
439 }
440 #else
441 assert(!restore_wide_vectors, "vectors are generated only by C2");
442 #endif
443
444 __ vzeroupper();
445
446 // On EVEX enabled targets everything is handled in pop fpu state
447 if (restore_wide_vectors) {
448 // Restore upper half of YMM registers (0..15)
449 int base_addr = XSAVE_AREA_YMM_BEGIN;
450 for (int n = 0; n < 16; n++) {
451 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, base_addr+n*16));
452 }
453 if (VM_Version::supports_evex()) {
454 // Restore upper half of ZMM registers (0..15)
455 base_addr = XSAVE_AREA_ZMM_BEGIN;
456 for (int n = 0; n < 16; n++) {
457 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, base_addr+n*32));
458 }
459 // Restore full ZMM registers(16..num_xmm_regs)
460 base_addr = XSAVE_AREA_UPPERBANK;
461 int vector_len = Assembler::AVX_512bit;
462 int off = 0;
463 for (int n = 16; n < num_xmm_regs; n++) {
464 __ evmovdqul(as_XMMRegister(n), Address(rsp, base_addr+(off++*64)), vector_len);
465 }
466 #if COMPILER2_OR_JVMCI
467 base_addr = XSAVE_AREA_OPMASK_BEGIN;
468 off = 0;
469 for (int n = 0; n < KRegister::number_of_registers; n++) {
470 __ kmov(as_KRegister(n), Address(rsp, base_addr+(off++*8)));
471 }
472 #endif
473 }
474 } else {
475 if (VM_Version::supports_evex()) {
476 // Restore upper bank of XMM registers(16..31) for scalar or 16-byte vector usage
477 int base_addr = XSAVE_AREA_UPPERBANK;
478 int off = 0;
479 int vector_len = VM_Version::supports_avx512vl() ? Assembler::AVX_128bit : Assembler::AVX_512bit;
480 for (int n = 16; n < num_xmm_regs; n++) {
481 __ evmovdqul(as_XMMRegister(n), Address(rsp, base_addr+(off++*64)), vector_len);
482 }
483 #if COMPILER2_OR_JVMCI
484 base_addr = XSAVE_AREA_OPMASK_BEGIN;
485 off = 0;
486 for (int n = 0; n < KRegister::number_of_registers; n++) {
487 __ kmov(as_KRegister(n), Address(rsp, base_addr+(off++*8)));
488 }
489 #endif
490 }
491 }
492
493 #if COMPILER2_OR_JVMCI
494 if (UseAPX) {
495 int base_addr = XSAVE_AREA_EGPRS;
496 int off = 0;
497 for (int n = 16; n < Register::number_of_registers; n++) {
498 __ movq(as_Register(n), Address(rsp, base_addr+(off++*8)));
499 }
500 }
501 #endif
502
503 // Recover CPU state
504 __ pop_FPU_state();
505 __ restore_legacy_gprs();
506 __ addq(rsp, 8);
507 __ popf();
508 // Get the rbp described implicitly by the calling convention (no oopMap)
509 __ pop(rbp);
510 }
511
512 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
513
514 // Just restore result register. Only used by deoptimization. By
515 // now any callee save register that needs to be restored to a c2
516 // caller of the deoptee has been extracted into the vframeArray
517 // and will be stuffed into the c2i adapter we create for later
518 // restoration so only result registers need to be restored here.
519
520 // Restore fp result register
521 __ movdbl(xmm0, Address(rsp, xmm0_offset_in_bytes()));
522 // Restore integer result register
523 __ movptr(rax, Address(rsp, rax_offset_in_bytes()));
524 __ movptr(rdx, Address(rsp, rdx_offset_in_bytes()));
525
526 // Pop all of the register save are off the stack except the return address
527 __ addptr(rsp, return_offset_in_bytes());
528 }
529
530 // Is vector's size (in bytes) bigger than a size saved by default?
531 // 16 bytes XMM registers are saved by default using fxsave/fxrstor instructions.
532 bool SharedRuntime::is_wide_vector(int size) {
533 return size > 16;
534 }
535
536 // ---------------------------------------------------------------------------
537 // Read the array of BasicTypes from a signature, and compute where the
538 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
539 // quantities. Values less than VMRegImpl::stack0 are registers, those above
540 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
541 // as framesizes are fixed.
542 // VMRegImpl::stack0 refers to the first slot 0(sp).
543 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
544 // Register up to Register::number_of_registers are the 64-bit
545 // integer registers.
546
547 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
548 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
549 // units regardless of build. Of course for i486 there is no 64 bit build
550
551 // The Java calling convention is a "shifted" version of the C ABI.
552 // By skipping the first C ABI register we can call non-static jni methods
553 // with small numbers of arguments without having to shuffle the arguments
554 // at all. Since we control the java ABI we ought to at least get some
555 // advantage out of it.
556
557 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
558 VMRegPair *regs,
559 int total_args_passed) {
560
561 // Create the mapping between argument positions and
562 // registers.
563 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
564 j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5
565 };
566 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
567 j_farg0, j_farg1, j_farg2, j_farg3,
568 j_farg4, j_farg5, j_farg6, j_farg7
569 };
570
571
572 uint int_args = 0;
573 uint fp_args = 0;
574 uint stk_args = 0;
575
576 for (int i = 0; i < total_args_passed; i++) {
577 switch (sig_bt[i]) {
578 case T_BOOLEAN:
579 case T_CHAR:
580 case T_BYTE:
581 case T_SHORT:
582 case T_INT:
583 if (int_args < Argument::n_int_register_parameters_j) {
584 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
585 } else {
586 stk_args = align_up(stk_args, 2);
587 regs[i].set1(VMRegImpl::stack2reg(stk_args));
588 stk_args += 1;
589 }
590 break;
591 case T_VOID:
592 // halves of T_LONG or T_DOUBLE
593 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
594 regs[i].set_bad();
595 break;
596 case T_LONG:
597 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
598 // fall through
599 case T_OBJECT:
600 case T_ARRAY:
601 case T_ADDRESS:
602 if (int_args < Argument::n_int_register_parameters_j) {
603 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
604 } else {
605 stk_args = align_up(stk_args, 2);
606 regs[i].set2(VMRegImpl::stack2reg(stk_args));
607 stk_args += 2;
608 }
609 break;
610 case T_FLOAT:
611 if (fp_args < Argument::n_float_register_parameters_j) {
612 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
613 } else {
614 stk_args = align_up(stk_args, 2);
615 regs[i].set1(VMRegImpl::stack2reg(stk_args));
616 stk_args += 1;
617 }
618 break;
619 case T_DOUBLE:
620 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
621 if (fp_args < Argument::n_float_register_parameters_j) {
622 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
623 } else {
624 stk_args = align_up(stk_args, 2);
625 regs[i].set2(VMRegImpl::stack2reg(stk_args));
626 stk_args += 2;
627 }
628 break;
629 default:
630 ShouldNotReachHere();
631 break;
632 }
633 }
634
635 return stk_args;
636 }
637
638 // Same as java_calling_convention() but for multiple return
639 // values. There's no way to store them on the stack so if we don't
640 // have enough registers, multiple values can't be returned.
641 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j+1;
642 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
643 int SharedRuntime::java_return_convention(const BasicType *sig_bt,
644 VMRegPair *regs,
645 int total_args_passed) {
646 // Create the mapping between argument positions and
647 // registers.
648 static const Register INT_ArgReg[java_return_convention_max_int] = {
649 rax, j_rarg5, j_rarg4, j_rarg3, j_rarg2, j_rarg1, j_rarg0
650 };
651 static const XMMRegister FP_ArgReg[java_return_convention_max_float] = {
652 j_farg0, j_farg1, j_farg2, j_farg3,
653 j_farg4, j_farg5, j_farg6, j_farg7
654 };
655
656
657 uint int_args = 0;
658 uint fp_args = 0;
659
660 for (int i = 0; i < total_args_passed; i++) {
661 switch (sig_bt[i]) {
662 case T_BOOLEAN:
663 case T_CHAR:
664 case T_BYTE:
665 case T_SHORT:
666 case T_INT:
667 if (int_args < Argument::n_int_register_parameters_j+1) {
668 regs[i].set1(INT_ArgReg[int_args]->as_VMReg());
669 int_args++;
670 } else {
671 return -1;
672 }
673 break;
674 case T_VOID:
675 // halves of T_LONG or T_DOUBLE
676 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
677 regs[i].set_bad();
678 break;
679 case T_LONG:
680 assert(sig_bt[i + 1] == T_VOID, "expecting half");
681 // fall through
682 case T_OBJECT:
683 case T_ARRAY:
684 case T_ADDRESS:
685 case T_METADATA:
686 if (int_args < Argument::n_int_register_parameters_j+1) {
687 regs[i].set2(INT_ArgReg[int_args]->as_VMReg());
688 int_args++;
689 } else {
690 return -1;
691 }
692 break;
693 case T_FLOAT:
694 if (fp_args < Argument::n_float_register_parameters_j) {
695 regs[i].set1(FP_ArgReg[fp_args]->as_VMReg());
696 fp_args++;
697 } else {
698 return -1;
699 }
700 break;
701 case T_DOUBLE:
702 assert(sig_bt[i + 1] == T_VOID, "expecting half");
703 if (fp_args < Argument::n_float_register_parameters_j) {
704 regs[i].set2(FP_ArgReg[fp_args]->as_VMReg());
705 fp_args++;
706 } else {
707 return -1;
708 }
709 break;
710 default:
711 ShouldNotReachHere();
712 break;
713 }
714 }
715
716 return int_args + fp_args;
717 }
718
719 // Patch the callers callsite with entry to compiled code if it exists.
720 static void patch_callers_callsite(MacroAssembler *masm) {
721 Label L;
722 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), NULL_WORD);
723 __ jcc(Assembler::equal, L);
724
725 // Save the current stack pointer
726 __ mov(r13, rsp);
727 // Schedule the branch target address early.
728 // Call into the VM to patch the caller, then jump to compiled callee
729 // rax isn't live so capture return address while we easily can
730 __ movptr(rax, Address(rsp, 0));
731
732 // align stack so push_CPU_state doesn't fault
733 __ andptr(rsp, -(StackAlignmentInBytes));
734 __ push_CPU_state();
735 __ vzeroupper();
736 // VM needs caller's callsite
737 // VM needs target method
738 // This needs to be a long call since we will relocate this adapter to
739 // the codeBuffer and it may not reach
740
741 // Allocate argument register save area
742 if (frame::arg_reg_save_area_bytes != 0) {
743 __ subptr(rsp, frame::arg_reg_save_area_bytes);
744 }
745 __ mov(c_rarg0, rbx);
746 __ mov(c_rarg1, rax);
747 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
748
749 // De-allocate argument register save area
750 if (frame::arg_reg_save_area_bytes != 0) {
751 __ addptr(rsp, frame::arg_reg_save_area_bytes);
752 }
753
754 __ vzeroupper();
755 __ pop_CPU_state();
756 // restore sp
757 __ mov(rsp, r13);
758 __ bind(L);
759 }
760
761 // For each inline type argument, sig includes the list of fields of
762 // the inline type. This utility function computes the number of
763 // arguments for the call if inline types are passed by reference (the
764 // calling convention the interpreter expects).
765 static int compute_total_args_passed_int(const GrowableArray<SigEntry>* sig_extended) {
766 int total_args_passed = 0;
767 if (InlineTypePassFieldsAsArgs) {
768 for (int i = 0; i < sig_extended->length(); i++) {
769 BasicType bt = sig_extended->at(i)._bt;
770 if (bt == T_METADATA) {
771 // In sig_extended, an inline type argument starts with:
772 // T_METADATA, followed by the types of the fields of the
773 // inline type and T_VOID to mark the end of the value
774 // type. Inline types are flattened so, for instance, in the
775 // case of an inline type with an int field and an inline type
776 // field that itself has 2 fields, an int and a long:
777 // T_METADATA T_INT T_METADATA T_INT T_LONG T_VOID (second
778 // slot for the T_LONG) T_VOID (inner inline type) T_VOID
779 // (outer inline type)
780 total_args_passed++;
781 int vt = 1;
782 do {
783 i++;
784 BasicType bt = sig_extended->at(i)._bt;
785 BasicType prev_bt = sig_extended->at(i-1)._bt;
786 if (bt == T_METADATA) {
787 vt++;
788 } else if (bt == T_VOID &&
789 prev_bt != T_LONG &&
790 prev_bt != T_DOUBLE) {
791 vt--;
792 }
793 } while (vt != 0);
794 } else {
795 total_args_passed++;
796 }
797 }
798 } else {
799 total_args_passed = sig_extended->length();
800 }
801 return total_args_passed;
802 }
803
804
805 static void gen_c2i_adapter_helper(MacroAssembler* masm,
806 BasicType bt,
807 BasicType prev_bt,
808 size_t size_in_bytes,
809 const VMRegPair& reg_pair,
810 const Address& to,
811 int extraspace,
812 bool is_oop) {
813 if (bt == T_VOID) {
814 assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half");
815 return;
816 }
817
818 // Say 4 args:
819 // i st_off
820 // 0 32 T_LONG
821 // 1 24 T_VOID
822 // 2 16 T_OBJECT
823 // 3 8 T_BOOL
824 // - 0 return address
825 //
826 // However to make thing extra confusing. Because we can fit a long/double in
827 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
828 // leaves one slot empty and only stores to a single slot. In this case the
829 // slot that is occupied is the T_VOID slot. See I said it was confusing.
830
831 bool wide = (size_in_bytes == wordSize);
832 VMReg r_1 = reg_pair.first();
833 VMReg r_2 = reg_pair.second();
834 assert(r_2->is_valid() == wide, "invalid size");
835 if (!r_1->is_valid()) {
836 assert(!r_2->is_valid(), "must be invalid");
837 return;
838 }
839
840 if (!r_1->is_XMMRegister()) {
841 Register val = rax;
842 if (r_1->is_stack()) {
843 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
844 __ load_sized_value(val, Address(rsp, ld_off), size_in_bytes, /* is_signed */ false);
845 } else {
846 val = r_1->as_Register();
847 }
848 assert_different_registers(to.base(), val, rscratch1);
849 if (is_oop) {
850 __ push(r13);
851 __ push(rbx);
852 __ store_heap_oop(to, val, rscratch1, r13, rbx, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
853 __ pop(rbx);
854 __ pop(r13);
855 } else {
856 __ store_sized_value(to, val, size_in_bytes);
857 }
858 } else {
859 if (wide) {
860 __ movdbl(to, r_1->as_XMMRegister());
861 } else {
862 __ movflt(to, r_1->as_XMMRegister());
863 }
864 }
865 }
866
867 static void gen_c2i_adapter(MacroAssembler *masm,
868 const GrowableArray<SigEntry>* sig_extended,
869 const VMRegPair *regs,
870 bool requires_clinit_barrier,
871 address& c2i_no_clinit_check_entry,
872 Label& skip_fixup,
873 address start,
874 OopMapSet* oop_maps,
875 int& frame_complete,
876 int& frame_size_in_words,
877 bool alloc_inline_receiver) {
878 if (requires_clinit_barrier && VM_Version::supports_fast_class_init_checks()) {
879 Label L_skip_barrier;
880 Register method = rbx;
881
882 { // Bypass the barrier for non-static methods
883 Register flags = rscratch1;
884 __ movl(flags, Address(method, Method::access_flags_offset()));
885 __ testl(flags, JVM_ACC_STATIC);
886 __ jcc(Assembler::zero, L_skip_barrier); // non-static
887 }
888
889 Register klass = rscratch1;
890 __ load_method_holder(klass, method);
891 __ clinit_barrier(klass, r15_thread, &L_skip_barrier /*L_fast_path*/);
892
893 __ jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); // slow path
894
895 __ bind(L_skip_barrier);
896 c2i_no_clinit_check_entry = __ pc();
897 }
898
899 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
900 bs->c2i_entry_barrier(masm);
901
902 // Before we get into the guts of the C2I adapter, see if we should be here
903 // at all. We've come from compiled code and are attempting to jump to the
904 // interpreter, which means the caller made a static call to get here
905 // (vcalls always get a compiled target if there is one). Check for a
906 // compiled target. If there is one, we need to patch the caller's call.
907 patch_callers_callsite(masm);
908
909 __ bind(skip_fixup);
910
911 if (InlineTypePassFieldsAsArgs) {
912 // Is there an inline type argument?
913 bool has_inline_argument = false;
914 for (int i = 0; i < sig_extended->length() && !has_inline_argument; i++) {
915 has_inline_argument = (sig_extended->at(i)._bt == T_METADATA);
916 }
917 if (has_inline_argument) {
918 // There is at least an inline type argument: we're coming from
919 // compiled code so we have no buffers to back the inline types.
920 // Allocate the buffers here with a runtime call.
921 OopMap* map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, /*save_vectors*/ false);
922
923 frame_complete = __ offset();
924
925 __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1);
926
927 __ mov(c_rarg0, r15_thread);
928 __ mov(c_rarg1, rbx);
929 __ mov64(c_rarg2, (int64_t)alloc_inline_receiver);
930 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_inline_types)));
931
932 oop_maps->add_gc_map((int)(__ pc() - start), map);
933 __ reset_last_Java_frame(false);
934
935 RegisterSaver::restore_live_registers(masm);
936
937 Label no_exception;
938 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
939 __ jcc(Assembler::equal, no_exception);
940
941 __ movptr(Address(r15_thread, JavaThread::vm_result_offset()), NULL_WORD);
942 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
943 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
944
945 __ bind(no_exception);
946
947 // We get an array of objects from the runtime call
948 __ get_vm_result(rscratch2, r15_thread); // Use rscratch2 (r11) as temporary because rscratch1 (r10) is trashed by movptr()
949 __ get_vm_result_2(rbx, r15_thread); // TODO: required to keep the callee Method live?
950 }
951 }
952
953 // Since all args are passed on the stack, total_args_passed *
954 // Interpreter::stackElementSize is the space we need.
955 int total_args_passed = compute_total_args_passed_int(sig_extended);
956 assert(total_args_passed >= 0, "total_args_passed is %d", total_args_passed);
957
958 int extraspace = (total_args_passed * Interpreter::stackElementSize);
959
960 // stack is aligned, keep it that way
961 // This is not currently needed or enforced by the interpreter, but
962 // we might as well conform to the ABI.
963 extraspace = align_up(extraspace, 2*wordSize);
964
965 // set senderSP value
966 __ lea(r13, Address(rsp, wordSize));
967
968 #ifdef ASSERT
969 __ check_stack_alignment(r13, "sender stack not aligned");
970 #endif
971 if (extraspace > 0) {
972 // Pop the return address
973 __ pop(rax);
974
975 __ subptr(rsp, extraspace);
976
977 // Push the return address
978 __ push(rax);
979
980 // Account for the return address location since we store it first rather
981 // than hold it in a register across all the shuffling
982 extraspace += wordSize;
983 }
984
985 #ifdef ASSERT
986 __ check_stack_alignment(rsp, "callee stack not aligned", wordSize, rax);
987 #endif
988
989 // Now write the args into the outgoing interpreter space
990
991 // next_arg_comp is the next argument from the compiler point of
992 // view (inline type fields are passed in registers/on the stack). In
993 // sig_extended, an inline type argument starts with: T_METADATA,
994 // followed by the types of the fields of the inline type and T_VOID
995 // to mark the end of the inline type. ignored counts the number of
996 // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
997 // used to get the buffer for that argument from the pool of buffers
998 // we allocated above and want to pass to the
999 // interpreter. next_arg_int is the next argument from the
1000 // interpreter point of view (inline types are passed by reference).
1001 for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
1002 next_arg_comp < sig_extended->length(); next_arg_comp++) {
1003 assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
1004 assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
1005 BasicType bt = sig_extended->at(next_arg_comp)._bt;
1006 int st_off = (total_args_passed - next_arg_int) * Interpreter::stackElementSize;
1007 if (!InlineTypePassFieldsAsArgs || bt != T_METADATA) {
1008 int next_off = st_off - Interpreter::stackElementSize;
1009 const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
1010 const VMRegPair reg_pair = regs[next_arg_comp-ignored];
1011 size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
1012 gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
1013 size_in_bytes, reg_pair, Address(rsp, offset), extraspace, false);
1014 next_arg_int++;
1015 #ifdef ASSERT
1016 if (bt == T_LONG || bt == T_DOUBLE) {
1017 // Overwrite the unused slot with known junk
1018 __ mov64(rax, CONST64(0xdeadffffdeadaaaa));
1019 __ movptr(Address(rsp, st_off), rax);
1020 }
1021 #endif /* ASSERT */
1022 } else {
1023 ignored++;
1024 // get the buffer from the just allocated pool of buffers
1025 int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_OBJECT);
1026 __ load_heap_oop(r14, Address(rscratch2, index));
1027 next_vt_arg++; next_arg_int++;
1028 int vt = 1;
1029 // write fields we get from compiled code in registers/stack
1030 // slots to the buffer: we know we are done with that inline type
1031 // argument when we hit the T_VOID that acts as an end of inline
1032 // type delimiter for this inline type. Inline types are flattened
1033 // so we might encounter embedded inline types. Each entry in
1034 // sig_extended contains a field offset in the buffer.
1035 Label L_null;
1036 do {
1037 next_arg_comp++;
1038 BasicType bt = sig_extended->at(next_arg_comp)._bt;
1039 BasicType prev_bt = sig_extended->at(next_arg_comp-1)._bt;
1040 if (bt == T_METADATA) {
1041 vt++;
1042 ignored++;
1043 } else if (bt == T_VOID &&
1044 prev_bt != T_LONG &&
1045 prev_bt != T_DOUBLE) {
1046 vt--;
1047 ignored++;
1048 } else {
1049 int off = sig_extended->at(next_arg_comp)._offset;
1050 if (off == -1) {
1051 // Nullable inline type argument, emit null check
1052 VMReg reg = regs[next_arg_comp-ignored].first();
1053 Label L_notNull;
1054 if (reg->is_stack()) {
1055 int ld_off = reg->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
1056 __ testb(Address(rsp, ld_off), 1);
1057 } else {
1058 __ testb(reg->as_Register(), 1);
1059 }
1060 __ jcc(Assembler::notZero, L_notNull);
1061 __ movptr(Address(rsp, st_off), 0);
1062 __ jmp(L_null);
1063 __ bind(L_notNull);
1064 continue;
1065 }
1066 assert(off > 0, "offset in object should be positive");
1067 size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize;
1068 bool is_oop = is_reference_type(bt);
1069 gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
1070 size_in_bytes, regs[next_arg_comp-ignored], Address(r14, off), extraspace, is_oop);
1071 }
1072 } while (vt != 0);
1073 // pass the buffer to the interpreter
1074 __ movptr(Address(rsp, st_off), r14);
1075 __ bind(L_null);
1076 }
1077 }
1078
1079 // Schedule the branch target address early.
1080 __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset())));
1081 __ jmp(rcx);
1082 }
1083
1084 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg,
1085 address code_start, address code_end,
1086 Label& L_ok) {
1087 Label L_fail;
1088 __ lea(temp_reg, AddressLiteral(code_start, relocInfo::none));
1089 __ cmpptr(pc_reg, temp_reg);
1090 __ jcc(Assembler::belowEqual, L_fail);
1091 __ lea(temp_reg, AddressLiteral(code_end, relocInfo::none));
1092 __ cmpptr(pc_reg, temp_reg);
1093 __ jcc(Assembler::below, L_ok);
1094 __ bind(L_fail);
1095 }
1096
1097 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
1098 int comp_args_on_stack,
1099 const GrowableArray<SigEntry>* sig,
1100 const VMRegPair *regs) {
1101
1102 // Note: r13 contains the senderSP on entry. We must preserve it since
1103 // we may do a i2c -> c2i transition if we lose a race where compiled
1104 // code goes non-entrant while we get args ready.
1105 // In addition we use r13 to locate all the interpreter args as
1106 // we must align the stack to 16 bytes on an i2c entry else we
1107 // lose alignment we expect in all compiled code and register
1108 // save code can segv when fxsave instructions find improperly
1109 // aligned stack pointer.
1110
1111 // Adapters can be frameless because they do not require the caller
1112 // to perform additional cleanup work, such as correcting the stack pointer.
1113 // An i2c adapter is frameless because the *caller* frame, which is interpreted,
1114 // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
1115 // even if a callee has modified the stack pointer.
1116 // A c2i adapter is frameless because the *callee* frame, which is interpreted,
1117 // routinely repairs its caller's stack pointer (from sender_sp, which is set
1118 // up via the senderSP register).
1119 // In other words, if *either* the caller or callee is interpreted, we can
1120 // get the stack pointer repaired after a call.
1121 // This is why c2i and i2c adapters cannot be indefinitely composed.
1122 // In particular, if a c2i adapter were to somehow call an i2c adapter,
1123 // both caller and callee would be compiled methods, and neither would
1124 // clean up the stack pointer changes performed by the two adapters.
1125 // If this happens, control eventually transfers back to the compiled
1126 // caller, but with an uncorrected stack, causing delayed havoc.
1127
1128 if (VerifyAdapterCalls &&
1129 (Interpreter::code() != nullptr || StubRoutines::final_stubs_code() != nullptr)) {
1130 // So, let's test for cascading c2i/i2c adapters right now.
1131 // assert(Interpreter::contains($return_addr) ||
1132 // StubRoutines::contains($return_addr),
1133 // "i2c adapter must return to an interpreter frame");
1134 __ block_comment("verify_i2c { ");
1135 // Pick up the return address
1136 __ movptr(rax, Address(rsp, 0));
1137 Label L_ok;
1138 if (Interpreter::code() != nullptr) {
1139 range_check(masm, rax, r11,
1140 Interpreter::code()->code_start(),
1141 Interpreter::code()->code_end(),
1142 L_ok);
1143 }
1144 if (StubRoutines::initial_stubs_code() != nullptr) {
1145 range_check(masm, rax, r11,
1146 StubRoutines::initial_stubs_code()->code_begin(),
1147 StubRoutines::initial_stubs_code()->code_end(),
1148 L_ok);
1149 }
1150 if (StubRoutines::final_stubs_code() != nullptr) {
1151 range_check(masm, rax, r11,
1152 StubRoutines::final_stubs_code()->code_begin(),
1153 StubRoutines::final_stubs_code()->code_end(),
1154 L_ok);
1155 }
1156 const char* msg = "i2c adapter must return to an interpreter frame";
1157 __ block_comment(msg);
1158 __ stop(msg);
1159 __ bind(L_ok);
1160 __ block_comment("} verify_i2ce ");
1161 }
1162
1163 // Must preserve original SP for loading incoming arguments because
1164 // we need to align the outgoing SP for compiled code.
1165 __ movptr(r11, rsp);
1166
1167 // Pick up the return address
1168 __ pop(rax);
1169
1170 // Convert 4-byte c2 stack slots to words.
1171 int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1172
1173 if (comp_args_on_stack) {
1174 __ subptr(rsp, comp_words_on_stack * wordSize);
1175 }
1176
1177 // Ensure compiled code always sees stack at proper alignment
1178 __ andptr(rsp, -16);
1179
1180 // push the return address and misalign the stack that youngest frame always sees
1181 // as far as the placement of the call instruction
1182 __ push(rax);
1183
1184 // Put saved SP in another register
1185 const Register saved_sp = rax;
1186 __ movptr(saved_sp, r11);
1187
1188 // Will jump to the compiled code just as if compiled code was doing it.
1189 // Pre-load the register-jump target early, to schedule it better.
1190 __ movptr(r11, Address(rbx, in_bytes(Method::from_compiled_inline_offset())));
1191
1192 #if INCLUDE_JVMCI
1193 if (EnableJVMCI) {
1194 // check if this call should be routed towards a specific entry point
1195 __ cmpptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
1196 Label no_alternative_target;
1197 __ jcc(Assembler::equal, no_alternative_target);
1198 __ movptr(r11, Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
1199 __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())), 0);
1200 __ bind(no_alternative_target);
1201 }
1202 #endif // INCLUDE_JVMCI
1203
1204 int total_args_passed = sig->length();
1205
1206 // Now generate the shuffle code. Pick up all register args and move the
1207 // rest through the floating point stack top.
1208 for (int i = 0; i < total_args_passed; i++) {
1209 BasicType bt = sig->at(i)._bt;
1210 if (bt == T_VOID) {
1211 // Longs and doubles are passed in native word order, but misaligned
1212 // in the 32-bit build.
1213 BasicType prev_bt = (i > 0) ? sig->at(i-1)._bt : T_ILLEGAL;
1214 assert(i > 0 && (prev_bt == T_LONG || prev_bt == T_DOUBLE), "missing half");
1215 continue;
1216 }
1217
1218 // Pick up 0, 1 or 2 words from SP+offset.
1219
1220 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1221 "scrambled load targets?");
1222 // Load in argument order going down.
1223 int ld_off = (total_args_passed - i)*Interpreter::stackElementSize;
1224 // Point to interpreter value (vs. tag)
1225 int next_off = ld_off - Interpreter::stackElementSize;
1226 //
1227 //
1228 //
1229 VMReg r_1 = regs[i].first();
1230 VMReg r_2 = regs[i].second();
1231 if (!r_1->is_valid()) {
1232 assert(!r_2->is_valid(), "");
1233 continue;
1234 }
1235 if (r_1->is_stack()) {
1236 // Convert stack slot to an SP offset (+ wordSize to account for return address )
1237 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize;
1238
1239 // We can use r13 as a temp here because compiled code doesn't need r13 as an input
1240 // and if we end up going thru a c2i because of a miss a reasonable value of r13
1241 // will be generated.
1242 if (!r_2->is_valid()) {
1243 // sign extend???
1244 __ movl(r13, Address(saved_sp, ld_off));
1245 __ movptr(Address(rsp, st_off), r13);
1246 } else {
1247 //
1248 // We are using two optoregs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
1249 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
1250 // So we must adjust where to pick up the data to match the interpreter.
1251 //
1252 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
1253 // are accessed as negative so LSW is at LOW address
1254
1255 // ld_off is MSW so get LSW
1256 const int offset = (bt==T_LONG||bt==T_DOUBLE)?
1257 next_off : ld_off;
1258 __ movq(r13, Address(saved_sp, offset));
1259 // st_off is LSW (i.e. reg.first())
1260 __ movq(Address(rsp, st_off), r13);
1261 }
1262 } else if (r_1->is_Register()) { // Register argument
1263 Register r = r_1->as_Register();
1264 assert(r != rax, "must be different");
1265 if (r_2->is_valid()) {
1266 //
1267 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
1268 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
1269 // So we must adjust where to pick up the data to match the interpreter.
1270
1271 const int offset = (bt==T_LONG||bt==T_DOUBLE)?
1272 next_off : ld_off;
1273
1274 // this can be a misaligned move
1275 __ movq(r, Address(saved_sp, offset));
1276 } else {
1277 // sign extend and use a full word?
1278 __ movl(r, Address(saved_sp, ld_off));
1279 }
1280 } else {
1281 if (!r_2->is_valid()) {
1282 __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off));
1283 } else {
1284 __ movdbl(r_1->as_XMMRegister(), Address(saved_sp, next_off));
1285 }
1286 }
1287 }
1288
1289 __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1290
1291 // 6243940 We might end up in handle_wrong_method if
1292 // the callee is deoptimized as we race thru here. If that
1293 // happens we don't want to take a safepoint because the
1294 // caller frame will look interpreted and arguments are now
1295 // "compiled" so it is much better to make this transition
1296 // invisible to the stack walking code. Unfortunately if
1297 // we try and find the callee by normal means a safepoint
1298 // is possible. So we stash the desired callee in the thread
1299 // and the vm will find there should this case occur.
1300
1301 __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx);
1302
1303 // put Method* where a c2i would expect should we end up there
1304 // only needed because of c2 resolve stubs return Method* as a result in
1305 // rax
1306 __ mov(rax, rbx);
1307 __ jmp(r11);
1308 }
1309
1310 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
1311 Register data = rax;
1312 __ ic_check(1 /* end_alignment */);
1313 __ movptr(rbx, Address(data, CompiledICData::speculated_method_offset()));
1314
1315 // Method might have been compiled since the call site was patched to
1316 // interpreted if that is the case treat it as a miss so we can get
1317 // the call site corrected.
1318 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), NULL_WORD);
1319 __ jcc(Assembler::equal, skip_fixup);
1320 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1321 }
1322
1323 // ---------------------------------------------------------------
1324 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
1325 int comp_args_on_stack,
1326 const GrowableArray<SigEntry>* sig,
1327 const VMRegPair* regs,
1328 const GrowableArray<SigEntry>* sig_cc,
1329 const VMRegPair* regs_cc,
1330 const GrowableArray<SigEntry>* sig_cc_ro,
1331 const VMRegPair* regs_cc_ro,
1332 AdapterFingerPrint* fingerprint,
1333 AdapterBlob*& new_adapter,
1334 bool allocate_code_blob) {
1335 address i2c_entry = __ pc();
1336 gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);
1337
1338 // -------------------------------------------------------------------------
1339 // Generate a C2I adapter. On entry we know rbx holds the Method* during calls
1340 // to the interpreter. The args start out packed in the compiled layout. They
1341 // need to be unpacked into the interpreter layout. This will almost always
1342 // require some stack space. We grow the current (compiled) stack, then repack
1343 // the args. We finally end in a jump to the generic interpreter entry point.
1344 // On exit from the interpreter, the interpreter will restore our SP (lest the
1345 // compiled code, which relies solely on SP and not RBP, get sick).
1346
1347 address c2i_unverified_entry = __ pc();
1348 address c2i_unverified_inline_entry = __ pc();
1349 Label skip_fixup;
1350
1351 gen_inline_cache_check(masm, skip_fixup);
1352
1353 OopMapSet* oop_maps = new OopMapSet();
1354 int frame_complete = CodeOffsets::frame_never_safe;
1355 int frame_size_in_words = 0;
1356
1357 // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
1358 address c2i_no_clinit_check_entry = nullptr;
1359 address c2i_inline_ro_entry = __ pc();
1360 if (regs_cc != regs_cc_ro) {
1361 // No class init barrier needed because method is guaranteed to be non-static
1362 gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, c2i_no_clinit_check_entry,
1363 skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1364 skip_fixup.reset();
1365 }
1366
1367 // Scalarized c2i adapter
1368 address c2i_entry = __ pc();
1369 address c2i_inline_entry = __ pc();
1370 gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, c2i_no_clinit_check_entry,
1371 skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
1372
1373 // Non-scalarized c2i adapter
1374 if (regs != regs_cc) {
1375 c2i_unverified_inline_entry = __ pc();
1376 Label inline_entry_skip_fixup;
1377 gen_inline_cache_check(masm, inline_entry_skip_fixup);
1378
1379 c2i_inline_entry = __ pc();
1380 gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, c2i_no_clinit_check_entry,
1381 inline_entry_skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1382 }
1383
1384 // The c2i adapters might safepoint and trigger a GC. The caller must make sure that
1385 // the GC knows about the location of oop argument locations passed to the c2i adapter.
1386 if (allocate_code_blob) {
1387 bool caller_must_gc_arguments = (regs != regs_cc);
1388 new_adapter = AdapterBlob::create(masm->code(), frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
1389 }
1390
1391 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_inline_entry, c2i_inline_ro_entry, c2i_unverified_entry, c2i_unverified_inline_entry, c2i_no_clinit_check_entry);
1392 }
1393
1394 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1395 VMRegPair *regs,
1396 int total_args_passed) {
1397
1398 // We return the amount of VMRegImpl stack slots we need to reserve for all
1399 // the arguments NOT counting out_preserve_stack_slots.
1400
1401 // NOTE: These arrays will have to change when c1 is ported
1402 #ifdef _WIN64
1403 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1404 c_rarg0, c_rarg1, c_rarg2, c_rarg3
1405 };
1406 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1407 c_farg0, c_farg1, c_farg2, c_farg3
1408 };
1409 #else
1410 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1411 c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5
1412 };
1413 static const XMMRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1414 c_farg0, c_farg1, c_farg2, c_farg3,
1415 c_farg4, c_farg5, c_farg6, c_farg7
1416 };
1417 #endif // _WIN64
1418
1419
1420 uint int_args = 0;
1421 uint fp_args = 0;
1422 uint stk_args = 0; // inc by 2 each time
1423
1424 for (int i = 0; i < total_args_passed; i++) {
1425 switch (sig_bt[i]) {
1426 case T_BOOLEAN:
1427 case T_CHAR:
1428 case T_BYTE:
1429 case T_SHORT:
1430 case T_INT:
1431 if (int_args < Argument::n_int_register_parameters_c) {
1432 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
1433 #ifdef _WIN64
1434 fp_args++;
1435 // Allocate slots for callee to stuff register args the stack.
1436 stk_args += 2;
1437 #endif
1438 } else {
1439 regs[i].set1(VMRegImpl::stack2reg(stk_args));
1440 stk_args += 2;
1441 }
1442 break;
1443 case T_LONG:
1444 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1445 // fall through
1446 case T_OBJECT:
1447 case T_ARRAY:
1448 case T_ADDRESS:
1449 case T_METADATA:
1450 if (int_args < Argument::n_int_register_parameters_c) {
1451 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
1452 #ifdef _WIN64
1453 fp_args++;
1454 stk_args += 2;
1455 #endif
1456 } else {
1457 regs[i].set2(VMRegImpl::stack2reg(stk_args));
1458 stk_args += 2;
1459 }
1460 break;
1461 case T_FLOAT:
1462 if (fp_args < Argument::n_float_register_parameters_c) {
1463 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
1464 #ifdef _WIN64
1465 int_args++;
1466 // Allocate slots for callee to stuff register args the stack.
1467 stk_args += 2;
1468 #endif
1469 } else {
1470 regs[i].set1(VMRegImpl::stack2reg(stk_args));
1471 stk_args += 2;
1472 }
1473 break;
1474 case T_DOUBLE:
1475 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1476 if (fp_args < Argument::n_float_register_parameters_c) {
1477 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
1478 #ifdef _WIN64
1479 int_args++;
1480 // Allocate slots for callee to stuff register args the stack.
1481 stk_args += 2;
1482 #endif
1483 } else {
1484 regs[i].set2(VMRegImpl::stack2reg(stk_args));
1485 stk_args += 2;
1486 }
1487 break;
1488 case T_VOID: // Halves of longs and doubles
1489 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
1490 regs[i].set_bad();
1491 break;
1492 default:
1493 ShouldNotReachHere();
1494 break;
1495 }
1496 }
1497 #ifdef _WIN64
1498 // windows abi requires that we always allocate enough stack space
1499 // for 4 64bit registers to be stored down.
1500 if (stk_args < 8) {
1501 stk_args = 8;
1502 }
1503 #endif // _WIN64
1504
1505 return stk_args;
1506 }
1507
1508 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
1509 uint num_bits,
1510 uint total_args_passed) {
1511 assert(num_bits == 64 || num_bits == 128 || num_bits == 256 || num_bits == 512,
1512 "only certain vector sizes are supported for now");
1513
1514 static const XMMRegister VEC_ArgReg[32] = {
1515 xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7,
1516 xmm8, xmm9, xmm10, xmm11, xmm12, xmm13, xmm14, xmm15,
1517 xmm16, xmm17, xmm18, xmm19, xmm20, xmm21, xmm22, xmm23,
1518 xmm24, xmm25, xmm26, xmm27, xmm28, xmm29, xmm30, xmm31
1519 };
1520
1521 uint stk_args = 0;
1522 uint fp_args = 0;
1523
1524 for (uint i = 0; i < total_args_passed; i++) {
1525 VMReg vmreg = VEC_ArgReg[fp_args++]->as_VMReg();
1526 int next_val = num_bits == 64 ? 1 : (num_bits == 128 ? 3 : (num_bits == 256 ? 7 : 15));
1527 regs[i].set_pair(vmreg->next(next_val), vmreg);
1528 }
1529
1530 return stk_args;
1531 }
1532
1533 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1534 // We always ignore the frame_slots arg and just use the space just below frame pointer
1535 // which by this time is free to use
1536 switch (ret_type) {
1537 case T_FLOAT:
1538 __ movflt(Address(rbp, -wordSize), xmm0);
1539 break;
1540 case T_DOUBLE:
1541 __ movdbl(Address(rbp, -wordSize), xmm0);
1542 break;
1543 case T_VOID: break;
1544 default: {
1545 __ movptr(Address(rbp, -wordSize), rax);
1546 }
1547 }
1548 }
1549
1550 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1551 // We always ignore the frame_slots arg and just use the space just below frame pointer
1552 // which by this time is free to use
1553 switch (ret_type) {
1554 case T_FLOAT:
1555 __ movflt(xmm0, Address(rbp, -wordSize));
1556 break;
1557 case T_DOUBLE:
1558 __ movdbl(xmm0, Address(rbp, -wordSize));
1559 break;
1560 case T_VOID: break;
1561 default: {
1562 __ movptr(rax, Address(rbp, -wordSize));
1563 }
1564 }
1565 }
1566
1567 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1568 for ( int i = first_arg ; i < arg_count ; i++ ) {
1569 if (args[i].first()->is_Register()) {
1570 __ push(args[i].first()->as_Register());
1571 } else if (args[i].first()->is_XMMRegister()) {
1572 __ subptr(rsp, 2*wordSize);
1573 __ movdbl(Address(rsp, 0), args[i].first()->as_XMMRegister());
1574 }
1575 }
1576 }
1577
1578 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1579 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
1580 if (args[i].first()->is_Register()) {
1581 __ pop(args[i].first()->as_Register());
1582 } else if (args[i].first()->is_XMMRegister()) {
1583 __ movdbl(args[i].first()->as_XMMRegister(), Address(rsp, 0));
1584 __ addptr(rsp, 2*wordSize);
1585 }
1586 }
1587 }
1588
1589 static void verify_oop_args(MacroAssembler* masm,
1590 const methodHandle& method,
1591 const BasicType* sig_bt,
1592 const VMRegPair* regs) {
1593 Register temp_reg = rbx; // not part of any compiled calling seq
1594 if (VerifyOops) {
1595 for (int i = 0; i < method->size_of_parameters(); i++) {
1596 if (is_reference_type(sig_bt[i])) {
1597 VMReg r = regs[i].first();
1598 assert(r->is_valid(), "bad oop arg");
1599 if (r->is_stack()) {
1600 __ movptr(temp_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
1601 __ verify_oop(temp_reg);
1602 } else {
1603 __ verify_oop(r->as_Register());
1604 }
1605 }
1606 }
1607 }
1608 }
1609
1610 static void check_continuation_enter_argument(VMReg actual_vmreg,
1611 Register expected_reg,
1612 const char* name) {
1613 assert(!actual_vmreg->is_stack(), "%s cannot be on stack", name);
1614 assert(actual_vmreg->as_Register() == expected_reg,
1615 "%s is in unexpected register: %s instead of %s",
1616 name, actual_vmreg->as_Register()->name(), expected_reg->name());
1617 }
1618
1619
1620 //---------------------------- continuation_enter_setup ---------------------------
1621 //
1622 // Arguments:
1623 // None.
1624 //
1625 // Results:
1626 // rsp: pointer to blank ContinuationEntry
1627 //
1628 // Kills:
1629 // rax
1630 //
1631 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
1632 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1633 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
1634 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1635
1636 stack_slots += checked_cast<int>(ContinuationEntry::size()) / wordSize;
1637 __ subptr(rsp, checked_cast<int32_t>(ContinuationEntry::size()));
1638
1639 int frame_size = (checked_cast<int>(ContinuationEntry::size()) + wordSize) / VMRegImpl::stack_slot_size;
1640 OopMap* map = new OopMap(frame_size, 0);
1641
1642 __ movptr(rax, Address(r15_thread, JavaThread::cont_entry_offset()));
1643 __ movptr(Address(rsp, ContinuationEntry::parent_offset()), rax);
1644 __ movptr(Address(r15_thread, JavaThread::cont_entry_offset()), rsp);
1645
1646 return map;
1647 }
1648
1649 //---------------------------- fill_continuation_entry ---------------------------
1650 //
1651 // Arguments:
1652 // rsp: pointer to blank Continuation entry
1653 // reg_cont_obj: pointer to the continuation
1654 // reg_flags: flags
1655 //
1656 // Results:
1657 // rsp: pointer to filled out ContinuationEntry
1658 //
1659 // Kills:
1660 // rax
1661 //
1662 static void fill_continuation_entry(MacroAssembler* masm, Register reg_cont_obj, Register reg_flags) {
1663 assert_different_registers(rax, reg_cont_obj, reg_flags);
1664 #ifdef ASSERT
1665 __ movl(Address(rsp, ContinuationEntry::cookie_offset()), ContinuationEntry::cookie_value());
1666 #endif
1667 __ movptr(Address(rsp, ContinuationEntry::cont_offset()), reg_cont_obj);
1668 __ movl (Address(rsp, ContinuationEntry::flags_offset()), reg_flags);
1669 __ movptr(Address(rsp, ContinuationEntry::chunk_offset()), 0);
1670 __ movl(Address(rsp, ContinuationEntry::argsize_offset()), 0);
1671 __ movl(Address(rsp, ContinuationEntry::pin_count_offset()), 0);
1672
1673 __ movptr(rax, Address(r15_thread, JavaThread::cont_fastpath_offset()));
1674 __ movptr(Address(rsp, ContinuationEntry::parent_cont_fastpath_offset()), rax);
1675 __ movq(rax, Address(r15_thread, JavaThread::held_monitor_count_offset()));
1676 __ movq(Address(rsp, ContinuationEntry::parent_held_monitor_count_offset()), rax);
1677
1678 __ movptr(Address(r15_thread, JavaThread::cont_fastpath_offset()), 0);
1679 __ movq(Address(r15_thread, JavaThread::held_monitor_count_offset()), 0);
1680 }
1681
1682 //---------------------------- continuation_enter_cleanup ---------------------------
1683 //
1684 // Arguments:
1685 // rsp: pointer to the ContinuationEntry
1686 //
1687 // Results:
1688 // rsp: pointer to the spilled rbp in the entry frame
1689 //
1690 // Kills:
1691 // rbx
1692 //
1693 void static continuation_enter_cleanup(MacroAssembler* masm) {
1694 #ifdef ASSERT
1695 Label L_good_sp;
1696 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
1697 __ jcc(Assembler::equal, L_good_sp);
1698 __ stop("Incorrect rsp at continuation_enter_cleanup");
1699 __ bind(L_good_sp);
1700 #endif
1701 __ movptr(rbx, Address(rsp, ContinuationEntry::parent_cont_fastpath_offset()));
1702 __ movptr(Address(r15_thread, JavaThread::cont_fastpath_offset()), rbx);
1703
1704 if (CheckJNICalls) {
1705 // Check if this is a virtual thread continuation
1706 Label L_skip_vthread_code;
1707 __ cmpl(Address(rsp, ContinuationEntry::flags_offset()), 0);
1708 __ jcc(Assembler::equal, L_skip_vthread_code);
1709
1710 // If the held monitor count is > 0 and this vthread is terminating then
1711 // it failed to release a JNI monitor. So we issue the same log message
1712 // that JavaThread::exit does.
1713 __ cmpptr(Address(r15_thread, JavaThread::jni_monitor_count_offset()), 0);
1714 __ jcc(Assembler::equal, L_skip_vthread_code);
1715
1716 // rax may hold an exception oop, save it before the call
1717 __ push(rax);
1718 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
1719 __ pop(rax);
1720
1721 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
1722 // on termination. The held count is implicitly zeroed below when we restore from
1723 // the parent held count (which has to be zero).
1724 __ movq(Address(r15_thread, JavaThread::jni_monitor_count_offset()), 0);
1725
1726 __ bind(L_skip_vthread_code);
1727 }
1728 #ifdef ASSERT
1729 else {
1730 // Check if this is a virtual thread continuation
1731 Label L_skip_vthread_code;
1732 __ cmpl(Address(rsp, ContinuationEntry::flags_offset()), 0);
1733 __ jcc(Assembler::equal, L_skip_vthread_code);
1734
1735 // See comment just above. If not checking JNI calls the JNI count is only
1736 // needed for assertion checking.
1737 __ movq(Address(r15_thread, JavaThread::jni_monitor_count_offset()), 0);
1738
1739 __ bind(L_skip_vthread_code);
1740 }
1741 #endif
1742
1743 __ movq(rbx, Address(rsp, ContinuationEntry::parent_held_monitor_count_offset()));
1744 __ movq(Address(r15_thread, JavaThread::held_monitor_count_offset()), rbx);
1745
1746 __ movptr(rbx, Address(rsp, ContinuationEntry::parent_offset()));
1747 __ movptr(Address(r15_thread, JavaThread::cont_entry_offset()), rbx);
1748 __ addptr(rsp, checked_cast<int32_t>(ContinuationEntry::size()));
1749 }
1750
1751 static void gen_continuation_enter(MacroAssembler* masm,
1752 const VMRegPair* regs,
1753 int& exception_offset,
1754 OopMapSet* oop_maps,
1755 int& frame_complete,
1756 int& stack_slots,
1757 int& interpreted_entry_offset,
1758 int& compiled_entry_offset) {
1759
1760 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1761 int pos_cont_obj = 0;
1762 int pos_is_cont = 1;
1763 int pos_is_virtual = 2;
1764
1765 // The platform-specific calling convention may present the arguments in various registers.
1766 // To simplify the rest of the code, we expect the arguments to reside at these known
1767 // registers, and we additionally check the placement here in case calling convention ever
1768 // changes.
1769 Register reg_cont_obj = c_rarg1;
1770 Register reg_is_cont = c_rarg2;
1771 Register reg_is_virtual = c_rarg3;
1772
1773 check_continuation_enter_argument(regs[pos_cont_obj].first(), reg_cont_obj, "Continuation object");
1774 check_continuation_enter_argument(regs[pos_is_cont].first(), reg_is_cont, "isContinue");
1775 check_continuation_enter_argument(regs[pos_is_virtual].first(), reg_is_virtual, "isVirtualThread");
1776
1777 // Utility methods kill rax, make sure there are no collisions
1778 assert_different_registers(rax, reg_cont_obj, reg_is_cont, reg_is_virtual);
1779
1780 AddressLiteral resolve(SharedRuntime::get_resolve_static_call_stub(),
1781 relocInfo::static_call_type);
1782
1783 address start = __ pc();
1784
1785 Label L_thaw, L_exit;
1786
1787 // i2i entry used at interp_only_mode only
1788 interpreted_entry_offset = __ pc() - start;
1789 {
1790 #ifdef ASSERT
1791 Label is_interp_only;
1792 __ cmpb(Address(r15_thread, JavaThread::interp_only_mode_offset()), 0);
1793 __ jcc(Assembler::notEqual, is_interp_only);
1794 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1795 __ bind(is_interp_only);
1796 #endif
1797
1798 __ pop(rax); // return address
1799 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1800 __ movptr(c_rarg1, Address(rsp, Interpreter::stackElementSize*2));
1801 __ movl(c_rarg2, Address(rsp, Interpreter::stackElementSize*1));
1802 __ movl(c_rarg3, Address(rsp, Interpreter::stackElementSize*0));
1803 __ andptr(rsp, -16); // Ensure compiled code always sees stack at proper alignment
1804 __ push(rax); // return address
1805 __ push_cont_fastpath();
1806
1807 __ enter();
1808
1809 stack_slots = 2; // will be adjusted in setup
1810 OopMap* map = continuation_enter_setup(masm, stack_slots);
1811 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1812 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1813
1814 __ verify_oop(reg_cont_obj);
1815
1816 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1817
1818 // If continuation, call to thaw. Otherwise, resolve the call and exit.
1819 __ testptr(reg_is_cont, reg_is_cont);
1820 __ jcc(Assembler::notZero, L_thaw);
1821
1822 // --- Resolve path
1823
1824 // Make sure the call is patchable
1825 __ align(BytesPerWord, __ offset() + NativeCall::displacement_offset);
1826 // Emit stub for static call
1827 address stub = CompiledDirectCall::emit_to_interp_stub(masm, __ pc());
1828 if (stub == nullptr) {
1829 fatal("CodeCache is full at gen_continuation_enter");
1830 }
1831 __ call(resolve);
1832 oop_maps->add_gc_map(__ pc() - start, map);
1833 __ post_call_nop();
1834
1835 __ jmp(L_exit);
1836 }
1837
1838 // compiled entry
1839 __ align(CodeEntryAlignment);
1840 compiled_entry_offset = __ pc() - start;
1841 __ enter();
1842
1843 stack_slots = 2; // will be adjusted in setup
1844 OopMap* map = continuation_enter_setup(masm, stack_slots);
1845
1846 // Frame is now completed as far as size and linkage.
1847 frame_complete = __ pc() - start;
1848
1849 __ verify_oop(reg_cont_obj);
1850
1851 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1852
1853 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1854 __ testptr(reg_is_cont, reg_is_cont);
1855 __ jccb(Assembler::notZero, L_thaw);
1856
1857 // --- call Continuation.enter(Continuation c, boolean isContinue)
1858
1859 // Make sure the call is patchable
1860 __ align(BytesPerWord, __ offset() + NativeCall::displacement_offset);
1861
1862 // Emit stub for static call
1863 address stub = CompiledDirectCall::emit_to_interp_stub(masm, __ pc());
1864 if (stub == nullptr) {
1865 fatal("CodeCache is full at gen_continuation_enter");
1866 }
1867
1868 // The call needs to be resolved. There's a special case for this in
1869 // SharedRuntime::find_callee_info_helper() which calls
1870 // LinkResolver::resolve_continuation_enter() which resolves the call to
1871 // Continuation.enter(Continuation c, boolean isContinue).
1872 __ call(resolve);
1873
1874 oop_maps->add_gc_map(__ pc() - start, map);
1875 __ post_call_nop();
1876
1877 __ jmpb(L_exit);
1878
1879 // --- Thawing path
1880
1881 __ bind(L_thaw);
1882
1883 __ call(RuntimeAddress(StubRoutines::cont_thaw()));
1884
1885 ContinuationEntry::_return_pc_offset = __ pc() - start;
1886 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1887 __ post_call_nop();
1888
1889 // --- Normal exit (resolve/thawing)
1890
1891 __ bind(L_exit);
1892
1893 continuation_enter_cleanup(masm);
1894 __ pop(rbp);
1895 __ ret(0);
1896
1897 // --- Exception handling path
1898
1899 exception_offset = __ pc() - start;
1900
1901 continuation_enter_cleanup(masm);
1902 __ pop(rbp);
1903
1904 __ movptr(c_rarg0, r15_thread);
1905 __ movptr(c_rarg1, Address(rsp, 0)); // return address
1906
1907 // rax still holds the original exception oop, save it before the call
1908 __ push(rax);
1909
1910 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 2);
1911 __ movptr(rbx, rax);
1912
1913 // Continue at exception handler:
1914 // rax: exception oop
1915 // rbx: exception handler
1916 // rdx: exception pc
1917 __ pop(rax);
1918 __ verify_oop(rax);
1919 __ pop(rdx);
1920 __ jmp(rbx);
1921 }
1922
1923 static void gen_continuation_yield(MacroAssembler* masm,
1924 const VMRegPair* regs,
1925 OopMapSet* oop_maps,
1926 int& frame_complete,
1927 int& stack_slots,
1928 int& compiled_entry_offset) {
1929 enum layout {
1930 rbp_off,
1931 rbpH_off,
1932 return_off,
1933 return_off2,
1934 framesize // inclusive of return address
1935 };
1936 stack_slots = framesize / VMRegImpl::slots_per_word;
1937 assert(stack_slots == 2, "recheck layout");
1938
1939 address start = __ pc();
1940 compiled_entry_offset = __ pc() - start;
1941 __ enter();
1942 address the_pc = __ pc();
1943
1944 frame_complete = the_pc - start;
1945
1946 // This nop must be exactly at the PC we push into the frame info.
1947 // We use this nop for fast CodeBlob lookup, associate the OopMap
1948 // with it right away.
1949 __ post_call_nop();
1950 OopMap* map = new OopMap(framesize, 1);
1951 oop_maps->add_gc_map(frame_complete, map);
1952
1953 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch1);
1954 __ movptr(c_rarg0, r15_thread);
1955 __ movptr(c_rarg1, rsp);
1956 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1957 __ reset_last_Java_frame(true);
1958
1959 Label L_pinned;
1960
1961 __ testptr(rax, rax);
1962 __ jcc(Assembler::notZero, L_pinned);
1963
1964 __ movptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
1965 continuation_enter_cleanup(masm);
1966 __ pop(rbp);
1967 __ ret(0);
1968
1969 __ bind(L_pinned);
1970
1971 // Pinned, return to caller
1972
1973 // handle pending exception thrown by freeze
1974 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
1975 Label ok;
1976 __ jcc(Assembler::equal, ok);
1977 __ leave();
1978 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1979 __ bind(ok);
1980
1981 __ leave();
1982 __ ret(0);
1983 }
1984
1985 static void gen_special_dispatch(MacroAssembler* masm,
1986 const methodHandle& method,
1987 const BasicType* sig_bt,
1988 const VMRegPair* regs) {
1989 verify_oop_args(masm, method, sig_bt, regs);
1990 vmIntrinsics::ID iid = method->intrinsic_id();
1991
1992 // Now write the args into the outgoing interpreter space
1993 bool has_receiver = false;
1994 Register receiver_reg = noreg;
1995 int member_arg_pos = -1;
1996 Register member_reg = noreg;
1997 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1998 if (ref_kind != 0) {
1999 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
2000 member_reg = rbx; // known to be free at this point
2001 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
2002 } else if (iid == vmIntrinsics::_invokeBasic) {
2003 has_receiver = true;
2004 } else if (iid == vmIntrinsics::_linkToNative) {
2005 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
2006 member_reg = rbx; // known to be free at this point
2007 } else {
2008 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
2009 }
2010
2011 if (member_reg != noreg) {
2012 // Load the member_arg into register, if necessary.
2013 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
2014 VMReg r = regs[member_arg_pos].first();
2015 if (r->is_stack()) {
2016 __ movptr(member_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
2017 } else {
2018 // no data motion is needed
2019 member_reg = r->as_Register();
2020 }
2021 }
2022
2023 if (has_receiver) {
2024 // Make sure the receiver is loaded into a register.
2025 assert(method->size_of_parameters() > 0, "oob");
2026 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
2027 VMReg r = regs[0].first();
2028 assert(r->is_valid(), "bad receiver arg");
2029 if (r->is_stack()) {
2030 // Porting note: This assumes that compiled calling conventions always
2031 // pass the receiver oop in a register. If this is not true on some
2032 // platform, pick a temp and load the receiver from stack.
2033 fatal("receiver always in a register");
2034 receiver_reg = j_rarg0; // known to be free at this point
2035 __ movptr(receiver_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
2036 } else {
2037 // no data motion is needed
2038 receiver_reg = r->as_Register();
2039 }
2040 }
2041
2042 // Figure out which address we are really jumping to:
2043 MethodHandles::generate_method_handle_dispatch(masm, iid,
2044 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
2045 }
2046
2047 // ---------------------------------------------------------------------------
2048 // Generate a native wrapper for a given method. The method takes arguments
2049 // in the Java compiled code convention, marshals them to the native
2050 // convention (handlizes oops, etc), transitions to native, makes the call,
2051 // returns to java state (possibly blocking), unhandlizes any result and
2052 // returns.
2053 //
2054 // Critical native functions are a shorthand for the use of
2055 // GetPrimtiveArrayCritical and disallow the use of any other JNI
2056 // functions. The wrapper is expected to unpack the arguments before
2057 // passing them to the callee. Critical native functions leave the state _in_Java,
2058 // since they cannot stop for GC.
2059 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
2060 // block and the check for pending exceptions it's impossible for them
2061 // to be thrown.
2062 //
2063 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
2064 const methodHandle& method,
2065 int compile_id,
2066 BasicType* in_sig_bt,
2067 VMRegPair* in_regs,
2068 BasicType ret_type) {
2069 if (method->is_continuation_native_intrinsic()) {
2070 int exception_offset = -1;
2071 OopMapSet* oop_maps = new OopMapSet();
2072 int frame_complete = -1;
2073 int stack_slots = -1;
2074 int interpreted_entry_offset = -1;
2075 int vep_offset = -1;
2076 if (method->is_continuation_enter_intrinsic()) {
2077 gen_continuation_enter(masm,
2078 in_regs,
2079 exception_offset,
2080 oop_maps,
2081 frame_complete,
2082 stack_slots,
2083 interpreted_entry_offset,
2084 vep_offset);
2085 } else if (method->is_continuation_yield_intrinsic()) {
2086 gen_continuation_yield(masm,
2087 in_regs,
2088 oop_maps,
2089 frame_complete,
2090 stack_slots,
2091 vep_offset);
2092 } else {
2093 guarantee(false, "Unknown Continuation native intrinsic");
2094 }
2095
2096 #ifdef ASSERT
2097 if (method->is_continuation_enter_intrinsic()) {
2098 assert(interpreted_entry_offset != -1, "Must be set");
2099 assert(exception_offset != -1, "Must be set");
2100 } else {
2101 assert(interpreted_entry_offset == -1, "Must be unset");
2102 assert(exception_offset == -1, "Must be unset");
2103 }
2104 assert(frame_complete != -1, "Must be set");
2105 assert(stack_slots != -1, "Must be set");
2106 assert(vep_offset != -1, "Must be set");
2107 #endif
2108
2109 __ flush();
2110 nmethod* nm = nmethod::new_native_nmethod(method,
2111 compile_id,
2112 masm->code(),
2113 vep_offset,
2114 frame_complete,
2115 stack_slots,
2116 in_ByteSize(-1),
2117 in_ByteSize(-1),
2118 oop_maps,
2119 exception_offset);
2120 if (nm == nullptr) return nm;
2121 if (method->is_continuation_enter_intrinsic()) {
2122 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
2123 } else if (method->is_continuation_yield_intrinsic()) {
2124 _cont_doYield_stub = nm;
2125 }
2126 return nm;
2127 }
2128
2129 if (method->is_method_handle_intrinsic()) {
2130 vmIntrinsics::ID iid = method->intrinsic_id();
2131 intptr_t start = (intptr_t)__ pc();
2132 int vep_offset = ((intptr_t)__ pc()) - start;
2133 gen_special_dispatch(masm,
2134 method,
2135 in_sig_bt,
2136 in_regs);
2137 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
2138 __ flush();
2139 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
2140 return nmethod::new_native_nmethod(method,
2141 compile_id,
2142 masm->code(),
2143 vep_offset,
2144 frame_complete,
2145 stack_slots / VMRegImpl::slots_per_word,
2146 in_ByteSize(-1),
2147 in_ByteSize(-1),
2148 nullptr);
2149 }
2150 address native_func = method->native_function();
2151 assert(native_func != nullptr, "must have function");
2152
2153 // An OopMap for lock (and class if static)
2154 OopMapSet *oop_maps = new OopMapSet();
2155 intptr_t start = (intptr_t)__ pc();
2156
2157 // We have received a description of where all the java arg are located
2158 // on entry to the wrapper. We need to convert these args to where
2159 // the jni function will expect them. To figure out where they go
2160 // we convert the java signature to a C signature by inserting
2161 // the hidden arguments as arg[0] and possibly arg[1] (static method)
2162
2163 const int total_in_args = method->size_of_parameters();
2164 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
2165
2166 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2167 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2168 BasicType* in_elem_bt = nullptr;
2169
2170 int argc = 0;
2171 out_sig_bt[argc++] = T_ADDRESS;
2172 if (method->is_static()) {
2173 out_sig_bt[argc++] = T_OBJECT;
2174 }
2175
2176 for (int i = 0; i < total_in_args ; i++ ) {
2177 out_sig_bt[argc++] = in_sig_bt[i];
2178 }
2179
2180 // Now figure out where the args must be stored and how much stack space
2181 // they require.
2182 int out_arg_slots;
2183 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
2184
2185 // Compute framesize for the wrapper. We need to handlize all oops in
2186 // incoming registers
2187
2188 // Calculate the total number of stack slots we will need.
2189
2190 // First count the abi requirement plus all of the outgoing args
2191 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2192
2193 // Now the space for the inbound oop handle area
2194 int total_save_slots = 6 * VMRegImpl::slots_per_word; // 6 arguments passed in registers
2195
2196 int oop_handle_offset = stack_slots;
2197 stack_slots += total_save_slots;
2198
2199 // Now any space we need for handlizing a klass if static method
2200
2201 int klass_slot_offset = 0;
2202 int klass_offset = -1;
2203 int lock_slot_offset = 0;
2204 bool is_static = false;
2205
2206 if (method->is_static()) {
2207 klass_slot_offset = stack_slots;
2208 stack_slots += VMRegImpl::slots_per_word;
2209 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
2210 is_static = true;
2211 }
2212
2213 // Plus a lock if needed
2214
2215 if (method->is_synchronized()) {
2216 lock_slot_offset = stack_slots;
2217 stack_slots += VMRegImpl::slots_per_word;
2218 }
2219
2220 // Now a place (+2) to save return values or temp during shuffling
2221 // + 4 for return address (which we own) and saved rbp
2222 stack_slots += 6;
2223
2224 // Ok The space we have allocated will look like:
2225 //
2226 //
2227 // FP-> | |
2228 // |---------------------|
2229 // | 2 slots for moves |
2230 // |---------------------|
2231 // | lock box (if sync) |
2232 // |---------------------| <- lock_slot_offset
2233 // | klass (if static) |
2234 // |---------------------| <- klass_slot_offset
2235 // | oopHandle area |
2236 // |---------------------| <- oop_handle_offset (6 java arg registers)
2237 // | outbound memory |
2238 // | based arguments |
2239 // | |
2240 // |---------------------|
2241 // | |
2242 // SP-> | out_preserved_slots |
2243 //
2244 //
2245
2246
2247 // Now compute actual number of stack words we need rounding to make
2248 // stack properly aligned.
2249 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
2250
2251 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2252
2253 // First thing make an ic check to see if we should even be here
2254
2255 // We are free to use all registers as temps without saving them and
2256 // restoring them except rbp. rbp is the only callee save register
2257 // as far as the interpreter and the compiler(s) are concerned.
2258
2259 const Register receiver = j_rarg0;
2260
2261 Label exception_pending;
2262
2263 assert_different_registers(receiver, rscratch1, rscratch2);
2264 __ verify_oop(receiver);
2265 __ ic_check(8 /* end_alignment */);
2266
2267 int vep_offset = ((intptr_t)__ pc()) - start;
2268
2269 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
2270 Label L_skip_barrier;
2271 Register klass = r10;
2272 __ mov_metadata(klass, method->method_holder()); // InstanceKlass*
2273 __ clinit_barrier(klass, r15_thread, &L_skip_barrier /*L_fast_path*/);
2274
2275 __ jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub())); // slow path
2276
2277 __ bind(L_skip_barrier);
2278 }
2279
2280 #ifdef COMPILER1
2281 // For Object.hashCode, System.identityHashCode try to pull hashCode from object header if available.
2282 if ((InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) || (method->intrinsic_id() == vmIntrinsics::_identityHashCode)) {
2283 inline_check_hashcode_from_object_header(masm, method, j_rarg0 /*obj_reg*/, rax /*result*/);
2284 }
2285 #endif // COMPILER1
2286
2287 // The instruction at the verified entry point must be 5 bytes or longer
2288 // because it can be patched on the fly by make_non_entrant. The stack bang
2289 // instruction fits that requirement.
2290
2291 // Generate stack overflow check
2292 __ bang_stack_with_offset((int)StackOverflow::stack_shadow_zone_size());
2293
2294 // Generate a new frame for the wrapper.
2295 __ enter();
2296 // -2 because return address is already present and so is saved rbp
2297 __ subptr(rsp, stack_size - 2*wordSize);
2298
2299 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2300 // native wrapper is not hot enough to micro optimize the nmethod entry barrier with an out-of-line stub
2301 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */);
2302
2303 // Frame is now completed as far as size and linkage.
2304 int frame_complete = ((intptr_t)__ pc()) - start;
2305
2306 #ifdef ASSERT
2307 __ check_stack_alignment(rsp, "improperly aligned stack");
2308 #endif /* ASSERT */
2309
2310
2311 // We use r14 as the oop handle for the receiver/klass
2312 // It is callee save so it survives the call to native
2313
2314 const Register oop_handle_reg = r14;
2315
2316 //
2317 // We immediately shuffle the arguments so that any vm call we have to
2318 // make from here on out (sync slow path, jvmti, etc.) we will have
2319 // captured the oops from our caller and have a valid oopMap for
2320 // them.
2321
2322 // -----------------
2323 // The Grand Shuffle
2324
2325 // The Java calling convention is either equal (linux) or denser (win64) than the
2326 // c calling convention. However the because of the jni_env argument the c calling
2327 // convention always has at least one more (and two for static) arguments than Java.
2328 // Therefore if we move the args from java -> c backwards then we will never have
2329 // a register->register conflict and we don't have to build a dependency graph
2330 // and figure out how to break any cycles.
2331 //
2332
2333 // Record esp-based slot for receiver on stack for non-static methods
2334 int receiver_offset = -1;
2335
2336 // This is a trick. We double the stack slots so we can claim
2337 // the oops in the caller's frame. Since we are sure to have
2338 // more args than the caller doubling is enough to make
2339 // sure we can capture all the incoming oop args from the
2340 // caller.
2341 //
2342 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2343
2344 // Mark location of rbp (someday)
2345 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rbp));
2346
2347 // Use eax, ebx as temporaries during any memory-memory moves we have to do
2348 // All inbound args are referenced based on rbp and all outbound args via rsp.
2349
2350
2351 #ifdef ASSERT
2352 bool reg_destroyed[Register::number_of_registers];
2353 bool freg_destroyed[XMMRegister::number_of_registers];
2354 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
2355 reg_destroyed[r] = false;
2356 }
2357 for ( int f = 0 ; f < XMMRegister::number_of_registers ; f++ ) {
2358 freg_destroyed[f] = false;
2359 }
2360
2361 #endif /* ASSERT */
2362
2363 // For JNI natives the incoming and outgoing registers are offset upwards.
2364 GrowableArray<int> arg_order(2 * total_in_args);
2365
2366 VMRegPair tmp_vmreg;
2367 tmp_vmreg.set2(rbx->as_VMReg());
2368
2369 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
2370 arg_order.push(i);
2371 arg_order.push(c_arg);
2372 }
2373
2374 int temploc = -1;
2375 for (int ai = 0; ai < arg_order.length(); ai += 2) {
2376 int i = arg_order.at(ai);
2377 int c_arg = arg_order.at(ai + 1);
2378 __ block_comment(err_msg("move %d -> %d", i, c_arg));
2379 #ifdef ASSERT
2380 if (in_regs[i].first()->is_Register()) {
2381 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
2382 } else if (in_regs[i].first()->is_XMMRegister()) {
2383 assert(!freg_destroyed[in_regs[i].first()->as_XMMRegister()->encoding()], "destroyed reg!");
2384 }
2385 if (out_regs[c_arg].first()->is_Register()) {
2386 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
2387 } else if (out_regs[c_arg].first()->is_XMMRegister()) {
2388 freg_destroyed[out_regs[c_arg].first()->as_XMMRegister()->encoding()] = true;
2389 }
2390 #endif /* ASSERT */
2391 switch (in_sig_bt[i]) {
2392 case T_ARRAY:
2393 case T_OBJECT:
2394 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
2395 ((i == 0) && (!is_static)),
2396 &receiver_offset);
2397 break;
2398 case T_VOID:
2399 break;
2400
2401 case T_FLOAT:
2402 __ float_move(in_regs[i], out_regs[c_arg]);
2403 break;
2404
2405 case T_DOUBLE:
2406 assert( i + 1 < total_in_args &&
2407 in_sig_bt[i + 1] == T_VOID &&
2408 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
2409 __ double_move(in_regs[i], out_regs[c_arg]);
2410 break;
2411
2412 case T_LONG :
2413 __ long_move(in_regs[i], out_regs[c_arg]);
2414 break;
2415
2416 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2417
2418 default:
2419 __ move32_64(in_regs[i], out_regs[c_arg]);
2420 }
2421 }
2422
2423 int c_arg;
2424
2425 // Pre-load a static method's oop into r14. Used both by locking code and
2426 // the normal JNI call code.
2427 // point c_arg at the first arg that is already loaded in case we
2428 // need to spill before we call out
2429 c_arg = total_c_args - total_in_args;
2430
2431 if (method->is_static()) {
2432
2433 // load oop into a register
2434 __ movoop(oop_handle_reg, JNIHandles::make_local(method->method_holder()->java_mirror()));
2435
2436 // Now handlize the static class mirror it's known not-null.
2437 __ movptr(Address(rsp, klass_offset), oop_handle_reg);
2438 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2439
2440 // Now get the handle
2441 __ lea(oop_handle_reg, Address(rsp, klass_offset));
2442 // store the klass handle as second argument
2443 __ movptr(c_rarg1, oop_handle_reg);
2444 // and protect the arg if we must spill
2445 c_arg--;
2446 }
2447
2448 // Change state to native (we save the return address in the thread, since it might not
2449 // be pushed on the stack when we do a stack traversal). It is enough that the pc()
2450 // points into the right code segment. It does not have to be the correct return pc.
2451 // We use the same pc/oopMap repeatedly when we call out
2452
2453 intptr_t the_pc = (intptr_t) __ pc();
2454 oop_maps->add_gc_map(the_pc - start, map);
2455
2456 __ set_last_Java_frame(rsp, noreg, (address)the_pc, rscratch1);
2457
2458
2459 // We have all of the arguments setup at this point. We must not touch any register
2460 // argument registers at this point (what if we save/restore them there are no oop?
2461
2462 if (DTraceMethodProbes) {
2463 // protect the args we've loaded
2464 save_args(masm, total_c_args, c_arg, out_regs);
2465 __ mov_metadata(c_rarg1, method());
2466 __ call_VM_leaf(
2467 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2468 r15_thread, c_rarg1);
2469 restore_args(masm, total_c_args, c_arg, out_regs);
2470 }
2471
2472 // RedefineClasses() tracing support for obsolete method entry
2473 if (log_is_enabled(Trace, redefine, class, obsolete)) {
2474 // protect the args we've loaded
2475 save_args(masm, total_c_args, c_arg, out_regs);
2476 __ mov_metadata(c_rarg1, method());
2477 __ call_VM_leaf(
2478 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
2479 r15_thread, c_rarg1);
2480 restore_args(masm, total_c_args, c_arg, out_regs);
2481 }
2482
2483 // Lock a synchronized method
2484
2485 // Register definitions used by locking and unlocking
2486
2487 const Register swap_reg = rax; // Must use rax for cmpxchg instruction
2488 const Register obj_reg = rbx; // Will contain the oop
2489 const Register lock_reg = r13; // Address of compiler lock object (BasicLock)
2490 const Register old_hdr = r13; // value of old header at unlock time
2491
2492 Label slow_path_lock;
2493 Label lock_done;
2494
2495 if (method->is_synchronized()) {
2496 Label count_mon;
2497
2498 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
2499
2500 // Get the handle (the 2nd argument)
2501 __ mov(oop_handle_reg, c_rarg1);
2502
2503 // Get address of the box
2504
2505 __ lea(lock_reg, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size));
2506
2507 // Load the oop from the handle
2508 __ movptr(obj_reg, Address(oop_handle_reg, 0));
2509
2510 if (LockingMode == LM_MONITOR) {
2511 __ jmp(slow_path_lock);
2512 } else if (LockingMode == LM_LEGACY) {
2513 // Load immediate 1 into swap_reg %rax
2514 __ movl(swap_reg, 1);
2515
2516 // Load (object->mark() | 1) into swap_reg %rax
2517 __ orptr(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2518 if (EnableValhalla) {
2519 // Mask inline_type bit such that we go to the slow path if object is an inline type
2520 __ andptr(swap_reg, ~((int) markWord::inline_type_bit_in_place));
2521 }
2522
2523 // Save (object->mark() | 1) into BasicLock's displaced header
2524 __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
2525
2526 // src -> dest iff dest == rax else rax <- dest
2527 __ lock();
2528 __ cmpxchgptr(lock_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2529 __ jcc(Assembler::equal, count_mon);
2530
2531 // Hmm should this move to the slow path code area???
2532
2533 // Test if the oopMark is an obvious stack pointer, i.e.,
2534 // 1) (mark & 3) == 0, and
2535 // 2) rsp <= mark < mark + os::pagesize()
2536 // These 3 tests can be done by evaluating the following
2537 // expression: ((mark - rsp) & (3 - os::vm_page_size())),
2538 // assuming both stack pointer and pagesize have their
2539 // least significant 2 bits clear.
2540 // NOTE: the oopMark is in swap_reg %rax as the result of cmpxchg
2541
2542 __ subptr(swap_reg, rsp);
2543 __ andptr(swap_reg, 3 - (int)os::vm_page_size());
2544
2545 // Save the test result, for recursive case, the result is zero
2546 __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
2547 __ jcc(Assembler::notEqual, slow_path_lock);
2548 } else {
2549 assert(LockingMode == LM_LIGHTWEIGHT, "must be");
2550 __ lightweight_lock(lock_reg, obj_reg, swap_reg, r15_thread, rscratch1, slow_path_lock);
2551 }
2552 __ bind(count_mon);
2553 __ inc_held_monitor_count();
2554
2555 // Slow path will re-enter here
2556 __ bind(lock_done);
2557 }
2558
2559 // Finally just about ready to make the JNI call
2560
2561 // get JNIEnv* which is first argument to native
2562 __ lea(c_rarg0, Address(r15_thread, in_bytes(JavaThread::jni_environment_offset())));
2563
2564 // Now set thread in native
2565 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_native);
2566
2567 __ call(RuntimeAddress(native_func));
2568
2569 // Verify or restore cpu control state after JNI call
2570 __ restore_cpu_control_state_after_jni(rscratch1);
2571
2572 // Unpack native results.
2573 switch (ret_type) {
2574 case T_BOOLEAN: __ c2bool(rax); break;
2575 case T_CHAR : __ movzwl(rax, rax); break;
2576 case T_BYTE : __ sign_extend_byte (rax); break;
2577 case T_SHORT : __ sign_extend_short(rax); break;
2578 case T_INT : /* nothing to do */ break;
2579 case T_DOUBLE :
2580 case T_FLOAT :
2581 // Result is in xmm0 we'll save as needed
2582 break;
2583 case T_ARRAY: // Really a handle
2584 case T_OBJECT: // Really a handle
2585 break; // can't de-handlize until after safepoint check
2586 case T_VOID: break;
2587 case T_LONG: break;
2588 default : ShouldNotReachHere();
2589 }
2590
2591 Label after_transition;
2592
2593 // Switch thread to "native transition" state before reading the synchronization state.
2594 // This additional state is necessary because reading and testing the synchronization
2595 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2596 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2597 // VM thread changes sync state to synchronizing and suspends threads for GC.
2598 // Thread A is resumed to finish this native method, but doesn't block here since it
2599 // didn't see any synchronization is progress, and escapes.
2600 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_native_trans);
2601
2602 // Force this write out before the read below
2603 if (!UseSystemMemoryBarrier) {
2604 __ membar(Assembler::Membar_mask_bits(
2605 Assembler::LoadLoad | Assembler::LoadStore |
2606 Assembler::StoreLoad | Assembler::StoreStore));
2607 }
2608
2609 // check for safepoint operation in progress and/or pending suspend requests
2610 {
2611 Label Continue;
2612 Label slow_path;
2613
2614 __ safepoint_poll(slow_path, r15_thread, true /* at_return */, false /* in_nmethod */);
2615
2616 __ cmpl(Address(r15_thread, JavaThread::suspend_flags_offset()), 0);
2617 __ jcc(Assembler::equal, Continue);
2618 __ bind(slow_path);
2619
2620 // Don't use call_VM as it will see a possible pending exception and forward it
2621 // and never return here preventing us from clearing _last_native_pc down below.
2622 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are
2623 // preserved and correspond to the bcp/locals pointers. So we do a runtime call
2624 // by hand.
2625 //
2626 __ vzeroupper();
2627 save_native_result(masm, ret_type, stack_slots);
2628 __ mov(c_rarg0, r15_thread);
2629 __ mov(r12, rsp); // remember sp
2630 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
2631 __ andptr(rsp, -16); // align stack as required by ABI
2632 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans)));
2633 __ mov(rsp, r12); // restore sp
2634 __ reinit_heapbase();
2635 // Restore any method result value
2636 restore_native_result(masm, ret_type, stack_slots);
2637 __ bind(Continue);
2638 }
2639
2640 // change thread state
2641 __ movl(Address(r15_thread, JavaThread::thread_state_offset()), _thread_in_Java);
2642 __ bind(after_transition);
2643
2644 Label reguard;
2645 Label reguard_done;
2646 __ cmpl(Address(r15_thread, JavaThread::stack_guard_state_offset()), StackOverflow::stack_guard_yellow_reserved_disabled);
2647 __ jcc(Assembler::equal, reguard);
2648 __ bind(reguard_done);
2649
2650 // native result if any is live
2651
2652 // Unlock
2653 Label slow_path_unlock;
2654 Label unlock_done;
2655 if (method->is_synchronized()) {
2656
2657 Label fast_done;
2658
2659 // Get locked oop from the handle we passed to jni
2660 __ movptr(obj_reg, Address(oop_handle_reg, 0));
2661
2662 if (LockingMode == LM_LEGACY) {
2663 Label not_recur;
2664 // Simple recursive lock?
2665 __ cmpptr(Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size), NULL_WORD);
2666 __ jcc(Assembler::notEqual, not_recur);
2667 __ dec_held_monitor_count();
2668 __ jmpb(fast_done);
2669 __ bind(not_recur);
2670 }
2671
2672 // Must save rax if it is live now because cmpxchg must use it
2673 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2674 save_native_result(masm, ret_type, stack_slots);
2675 }
2676
2677 if (LockingMode == LM_MONITOR) {
2678 __ jmp(slow_path_unlock);
2679 } else if (LockingMode == LM_LEGACY) {
2680 // get address of the stack lock
2681 __ lea(rax, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size));
2682 // get old displaced header
2683 __ movptr(old_hdr, Address(rax, 0));
2684
2685 // Atomic swap old header if oop still contains the stack lock
2686 __ lock();
2687 __ cmpxchgptr(old_hdr, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2688 __ jcc(Assembler::notEqual, slow_path_unlock);
2689 __ dec_held_monitor_count();
2690 } else {
2691 assert(LockingMode == LM_LIGHTWEIGHT, "must be");
2692 __ lightweight_unlock(obj_reg, swap_reg, r15_thread, lock_reg, slow_path_unlock);
2693 __ dec_held_monitor_count();
2694 }
2695
2696 // slow path re-enters here
2697 __ bind(unlock_done);
2698 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2699 restore_native_result(masm, ret_type, stack_slots);
2700 }
2701
2702 __ bind(fast_done);
2703 }
2704 if (DTraceMethodProbes) {
2705 save_native_result(masm, ret_type, stack_slots);
2706 __ mov_metadata(c_rarg1, method());
2707 __ call_VM_leaf(
2708 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2709 r15_thread, c_rarg1);
2710 restore_native_result(masm, ret_type, stack_slots);
2711 }
2712
2713 __ reset_last_Java_frame(false);
2714
2715 // Unbox oop result, e.g. JNIHandles::resolve value.
2716 if (is_reference_type(ret_type)) {
2717 __ resolve_jobject(rax /* value */,
2718 r15_thread /* thread */,
2719 rcx /* tmp */);
2720 }
2721
2722 if (CheckJNICalls) {
2723 // clear_pending_jni_exception_check
2724 __ movptr(Address(r15_thread, JavaThread::pending_jni_exception_check_fn_offset()), NULL_WORD);
2725 }
2726
2727 // reset handle block
2728 __ movptr(rcx, Address(r15_thread, JavaThread::active_handles_offset()));
2729 __ movl(Address(rcx, JNIHandleBlock::top_offset()), NULL_WORD);
2730
2731 // pop our frame
2732
2733 __ leave();
2734
2735 // Any exception pending?
2736 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), NULL_WORD);
2737 __ jcc(Assembler::notEqual, exception_pending);
2738
2739 // Return
2740
2741 __ ret(0);
2742
2743 // Unexpected paths are out of line and go here
2744
2745 // forward the exception
2746 __ bind(exception_pending);
2747
2748 // and forward the exception
2749 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2750
2751 // Slow path locking & unlocking
2752 if (method->is_synchronized()) {
2753
2754 // BEGIN Slow path lock
2755 __ bind(slow_path_lock);
2756
2757 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
2758 // args are (oop obj, BasicLock* lock, JavaThread* thread)
2759
2760 // protect the args we've loaded
2761 save_args(masm, total_c_args, c_arg, out_regs);
2762
2763 __ mov(c_rarg0, obj_reg);
2764 __ mov(c_rarg1, lock_reg);
2765 __ mov(c_rarg2, r15_thread);
2766
2767 // Not a leaf but we have last_Java_frame setup as we want
2768 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
2769 restore_args(masm, total_c_args, c_arg, out_regs);
2770
2771 #ifdef ASSERT
2772 { Label L;
2773 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), NULL_WORD);
2774 __ jcc(Assembler::equal, L);
2775 __ stop("no pending exception allowed on exit from monitorenter");
2776 __ bind(L);
2777 }
2778 #endif
2779 __ jmp(lock_done);
2780
2781 // END Slow path lock
2782
2783 // BEGIN Slow path unlock
2784 __ bind(slow_path_unlock);
2785
2786 // If we haven't already saved the native result we must save it now as xmm registers
2787 // are still exposed.
2788 __ vzeroupper();
2789 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2790 save_native_result(masm, ret_type, stack_slots);
2791 }
2792
2793 __ lea(c_rarg1, Address(rsp, lock_slot_offset * VMRegImpl::stack_slot_size));
2794
2795 __ mov(c_rarg0, obj_reg);
2796 __ mov(c_rarg2, r15_thread);
2797 __ mov(r12, rsp); // remember sp
2798 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
2799 __ andptr(rsp, -16); // align stack as required by ABI
2800
2801 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2802 // NOTE that obj_reg == rbx currently
2803 __ movptr(rbx, Address(r15_thread, in_bytes(Thread::pending_exception_offset())));
2804 __ movptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), NULL_WORD);
2805
2806 // args are (oop obj, BasicLock* lock, JavaThread* thread)
2807 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C)));
2808 __ mov(rsp, r12); // restore sp
2809 __ reinit_heapbase();
2810 #ifdef ASSERT
2811 {
2812 Label L;
2813 __ cmpptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), NULL_WORD);
2814 __ jcc(Assembler::equal, L);
2815 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2816 __ bind(L);
2817 }
2818 #endif /* ASSERT */
2819
2820 __ movptr(Address(r15_thread, in_bytes(Thread::pending_exception_offset())), rbx);
2821
2822 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2823 restore_native_result(masm, ret_type, stack_slots);
2824 }
2825 __ jmp(unlock_done);
2826
2827 // END Slow path unlock
2828
2829 } // synchronized
2830
2831 // SLOW PATH Reguard the stack if needed
2832
2833 __ bind(reguard);
2834 __ vzeroupper();
2835 save_native_result(masm, ret_type, stack_slots);
2836 __ mov(r12, rsp); // remember sp
2837 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
2838 __ andptr(rsp, -16); // align stack as required by ABI
2839 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)));
2840 __ mov(rsp, r12); // restore sp
2841 __ reinit_heapbase();
2842 restore_native_result(masm, ret_type, stack_slots);
2843 // and continue
2844 __ jmp(reguard_done);
2845
2846
2847
2848 __ flush();
2849
2850 nmethod *nm = nmethod::new_native_nmethod(method,
2851 compile_id,
2852 masm->code(),
2853 vep_offset,
2854 frame_complete,
2855 stack_slots / VMRegImpl::slots_per_word,
2856 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2857 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2858 oop_maps);
2859
2860 return nm;
2861 }
2862
2863 // this function returns the adjust size (in number of words) to a c2i adapter
2864 // activation for use during deoptimization
2865 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals ) {
2866 return (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2867 }
2868
2869
2870 uint SharedRuntime::out_preserve_stack_slots() {
2871 return 0;
2872 }
2873
2874
2875 // Number of stack slots between incoming argument block and the start of
2876 // a new frame. The PROLOG must add this many slots to the stack. The
2877 // EPILOG must remove this many slots. amd64 needs two slots for
2878 // return address.
2879 uint SharedRuntime::in_preserve_stack_slots() {
2880 return 4 + 2 * VerifyStackAtCalls;
2881 }
2882
2883 //------------------------------generate_deopt_blob----------------------------
2884 void SharedRuntime::generate_deopt_blob() {
2885 // Allocate space for the code
2886 ResourceMark rm;
2887 // Setup code generation tools
2888 int pad = 0;
2889 if (UseAVX > 2) {
2890 pad += 1024;
2891 }
2892 if (UseAPX) {
2893 pad += 1024;
2894 }
2895 #if INCLUDE_JVMCI
2896 if (EnableJVMCI) {
2897 pad += 512; // Increase the buffer size when compiling for JVMCI
2898 }
2899 #endif
2900 CodeBuffer buffer("deopt_blob", 2560+pad, 1024);
2901 MacroAssembler* masm = new MacroAssembler(&buffer);
2902 int frame_size_in_words;
2903 OopMap* map = nullptr;
2904 OopMapSet *oop_maps = new OopMapSet();
2905
2906 // -------------
2907 // This code enters when returning to a de-optimized nmethod. A return
2908 // address has been pushed on the stack, and return values are in
2909 // registers.
2910 // If we are doing a normal deopt then we were called from the patched
2911 // nmethod from the point we returned to the nmethod. So the return
2912 // address on the stack is wrong by NativeCall::instruction_size
2913 // We will adjust the value so it looks like we have the original return
2914 // address on the stack (like when we eagerly deoptimized).
2915 // In the case of an exception pending when deoptimizing, we enter
2916 // with a return address on the stack that points after the call we patched
2917 // into the exception handler. We have the following register state from,
2918 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp).
2919 // rax: exception oop
2920 // rbx: exception handler
2921 // rdx: throwing pc
2922 // So in this case we simply jam rdx into the useless return address and
2923 // the stack looks just like we want.
2924 //
2925 // At this point we need to de-opt. We save the argument return
2926 // registers. We call the first C routine, fetch_unroll_info(). This
2927 // routine captures the return values and returns a structure which
2928 // describes the current frame size and the sizes of all replacement frames.
2929 // The current frame is compiled code and may contain many inlined
2930 // functions, each with their own JVM state. We pop the current frame, then
2931 // push all the new frames. Then we call the C routine unpack_frames() to
2932 // populate these frames. Finally unpack_frames() returns us the new target
2933 // address. Notice that callee-save registers are BLOWN here; they have
2934 // already been captured in the vframeArray at the time the return PC was
2935 // patched.
2936 address start = __ pc();
2937 Label cont;
2938
2939 // Prolog for non exception case!
2940
2941 // Save everything in sight.
2942 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, /*save_wide_vectors*/ true);
2943
2944 // Normal deoptimization. Save exec mode for unpack_frames.
2945 __ movl(r14, Deoptimization::Unpack_deopt); // callee-saved
2946 __ jmp(cont);
2947
2948 int reexecute_offset = __ pc() - start;
2949 #if INCLUDE_JVMCI && !defined(COMPILER1)
2950 if (EnableJVMCI && UseJVMCICompiler) {
2951 // JVMCI does not use this kind of deoptimization
2952 __ should_not_reach_here();
2953 }
2954 #endif
2955
2956 // Reexecute case
2957 // return address is the pc describes what bci to do re-execute at
2958
2959 // No need to update map as each call to save_live_registers will produce identical oopmap
2960 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, /*save_wide_vectors*/ true);
2961
2962 __ movl(r14, Deoptimization::Unpack_reexecute); // callee-saved
2963 __ jmp(cont);
2964
2965 #if INCLUDE_JVMCI
2966 Label after_fetch_unroll_info_call;
2967 int implicit_exception_uncommon_trap_offset = 0;
2968 int uncommon_trap_offset = 0;
2969
2970 if (EnableJVMCI) {
2971 implicit_exception_uncommon_trap_offset = __ pc() - start;
2972
2973 __ pushptr(Address(r15_thread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2974 __ movptr(Address(r15_thread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())), NULL_WORD);
2975
2976 uncommon_trap_offset = __ pc() - start;
2977
2978 // Save everything in sight.
2979 RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, /*save_wide_vectors*/ true);
2980 // fetch_unroll_info needs to call last_java_frame()
2981 __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1);
2982
2983 __ movl(c_rarg1, Address(r15_thread, in_bytes(JavaThread::pending_deoptimization_offset())));
2984 __ movl(Address(r15_thread, in_bytes(JavaThread::pending_deoptimization_offset())), -1);
2985
2986 __ movl(r14, Deoptimization::Unpack_reexecute);
2987 __ mov(c_rarg0, r15_thread);
2988 __ movl(c_rarg2, r14); // exec mode
2989 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap)));
2990 oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2991
2992 __ reset_last_Java_frame(false);
2993
2994 __ jmp(after_fetch_unroll_info_call);
2995 } // EnableJVMCI
2996 #endif // INCLUDE_JVMCI
2997
2998 int exception_offset = __ pc() - start;
2999
3000 // Prolog for exception case
3001
3002 // all registers are dead at this entry point, except for rax, and
3003 // rdx which contain the exception oop and exception pc
3004 // respectively. Set them in TLS and fall thru to the
3005 // unpack_with_exception_in_tls entry point.
3006
3007 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), rdx);
3008 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()), rax);
3009
3010 int exception_in_tls_offset = __ pc() - start;
3011
3012 // new implementation because exception oop is now passed in JavaThread
3013
3014 // Prolog for exception case
3015 // All registers must be preserved because they might be used by LinearScan
3016 // Exceptiop oop and throwing PC are passed in JavaThread
3017 // tos: stack at point of call to method that threw the exception (i.e. only
3018 // args are on the stack, no return address)
3019
3020 // make room on stack for the return address
3021 // It will be patched later with the throwing pc. The correct value is not
3022 // available now because loading it from memory would destroy registers.
3023 __ push(0);
3024
3025 // Save everything in sight.
3026 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, /*save_wide_vectors*/ true);
3027
3028 // Now it is safe to overwrite any register
3029
3030 // Deopt during an exception. Save exec mode for unpack_frames.
3031 __ movl(r14, Deoptimization::Unpack_exception); // callee-saved
3032
3033 // load throwing pc from JavaThread and patch it as the return address
3034 // of the current frame. Then clear the field in JavaThread
3035
3036 __ movptr(rdx, Address(r15_thread, JavaThread::exception_pc_offset()));
3037 __ movptr(Address(rbp, wordSize), rdx);
3038 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), NULL_WORD);
3039
3040 #ifdef ASSERT
3041 // verify that there is really an exception oop in JavaThread
3042 __ movptr(rax, Address(r15_thread, JavaThread::exception_oop_offset()));
3043 __ verify_oop(rax);
3044
3045 // verify that there is no pending exception
3046 Label no_pending_exception;
3047 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
3048 __ testptr(rax, rax);
3049 __ jcc(Assembler::zero, no_pending_exception);
3050 __ stop("must not have pending exception here");
3051 __ bind(no_pending_exception);
3052 #endif
3053
3054 __ bind(cont);
3055
3056 // Call C code. Need thread and this frame, but NOT official VM entry
3057 // crud. We cannot block on this call, no GC can happen.
3058 //
3059 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
3060
3061 // fetch_unroll_info needs to call last_java_frame().
3062
3063 __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1);
3064 #ifdef ASSERT
3065 { Label L;
3066 __ cmpptr(Address(r15_thread, JavaThread::last_Java_fp_offset()), NULL_WORD);
3067 __ jcc(Assembler::equal, L);
3068 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
3069 __ bind(L);
3070 }
3071 #endif // ASSERT
3072 __ mov(c_rarg0, r15_thread);
3073 __ movl(c_rarg1, r14); // exec_mode
3074 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
3075
3076 // Need to have an oopmap that tells fetch_unroll_info where to
3077 // find any register it might need.
3078 oop_maps->add_gc_map(__ pc() - start, map);
3079
3080 __ reset_last_Java_frame(false);
3081
3082 #if INCLUDE_JVMCI
3083 if (EnableJVMCI) {
3084 __ bind(after_fetch_unroll_info_call);
3085 }
3086 #endif
3087
3088 // Load UnrollBlock* into rdi
3089 __ mov(rdi, rax);
3090
3091 __ movl(r14, Address(rdi, Deoptimization::UnrollBlock::unpack_kind_offset()));
3092 Label noException;
3093 __ cmpl(r14, Deoptimization::Unpack_exception); // Was exception pending?
3094 __ jcc(Assembler::notEqual, noException);
3095 __ movptr(rax, Address(r15_thread, JavaThread::exception_oop_offset()));
3096 // QQQ this is useless it was null above
3097 __ movptr(rdx, Address(r15_thread, JavaThread::exception_pc_offset()));
3098 __ movptr(Address(r15_thread, JavaThread::exception_oop_offset()), NULL_WORD);
3099 __ movptr(Address(r15_thread, JavaThread::exception_pc_offset()), NULL_WORD);
3100
3101 __ verify_oop(rax);
3102
3103 // Overwrite the result registers with the exception results.
3104 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax);
3105 // I think this is useless
3106 __ movptr(Address(rsp, RegisterSaver::rdx_offset_in_bytes()), rdx);
3107
3108 __ bind(noException);
3109
3110 // Only register save data is on the stack.
3111 // Now restore the result registers. Everything else is either dead
3112 // or captured in the vframeArray.
3113 RegisterSaver::restore_result_registers(masm);
3114
3115 // All of the register save area has been popped of the stack. Only the
3116 // return address remains.
3117
3118 // Pop all the frames we must move/replace.
3119 //
3120 // Frame picture (youngest to oldest)
3121 // 1: self-frame (no frame link)
3122 // 2: deopting frame (no frame link)
3123 // 3: caller of deopting frame (could be compiled/interpreted).
3124 //
3125 // Note: by leaving the return address of self-frame on the stack
3126 // and using the size of frame 2 to adjust the stack
3127 // when we are done the return to frame 3 will still be on the stack.
3128
3129 // Pop deoptimized frame
3130 __ movl(rcx, Address(rdi, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
3131 __ addptr(rsp, rcx);
3132
3133 // rsp should be pointing at the return address to the caller (3)
3134
3135 // Pick up the initial fp we should save
3136 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved)
3137 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset()));
3138
3139 #ifdef ASSERT
3140 // Compilers generate code that bang the stack by as much as the
3141 // interpreter would need. So this stack banging should never
3142 // trigger a fault. Verify that it does not on non product builds.
3143 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
3144 __ bang_stack_size(rbx, rcx);
3145 #endif
3146
3147 // Load address of array of frame pcs into rcx
3148 __ movptr(rcx, Address(rdi, Deoptimization::UnrollBlock::frame_pcs_offset()));
3149
3150 // Trash the old pc
3151 __ addptr(rsp, wordSize);
3152
3153 // Load address of array of frame sizes into rsi
3154 __ movptr(rsi, Address(rdi, Deoptimization::UnrollBlock::frame_sizes_offset()));
3155
3156 // Load counter into rdx
3157 __ movl(rdx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset()));
3158
3159 // Now adjust the caller's stack to make up for the extra locals
3160 // but record the original sp so that we can save it in the skeletal interpreter
3161 // frame and the stack walking of interpreter_sender will get the unextended sp
3162 // value and not the "real" sp value.
3163
3164 const Register sender_sp = r8;
3165
3166 __ mov(sender_sp, rsp);
3167 __ movl(rbx, Address(rdi,
3168 Deoptimization::UnrollBlock::
3169 caller_adjustment_offset()));
3170 __ subptr(rsp, rbx);
3171
3172 // Push interpreter frames in a loop
3173 Label loop;
3174 __ bind(loop);
3175 __ movptr(rbx, Address(rsi, 0)); // Load frame size
3176 __ subptr(rbx, 2*wordSize); // We'll push pc and ebp by hand
3177 __ pushptr(Address(rcx, 0)); // Save return address
3178 __ enter(); // Save old & set new ebp
3179 __ subptr(rsp, rbx); // Prolog
3180 // This value is corrected by layout_activation_impl
3181 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD);
3182 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), sender_sp); // Make it walkable
3183 __ mov(sender_sp, rsp); // Pass sender_sp to next frame
3184 __ addptr(rsi, wordSize); // Bump array pointer (sizes)
3185 __ addptr(rcx, wordSize); // Bump array pointer (pcs)
3186 __ decrementl(rdx); // Decrement counter
3187 __ jcc(Assembler::notZero, loop);
3188 __ pushptr(Address(rcx, 0)); // Save final return address
3189
3190 // Re-push self-frame
3191 __ enter(); // Save old & set new ebp
3192
3193 // Allocate a full sized register save area.
3194 // Return address and rbp are in place, so we allocate two less words.
3195 __ subptr(rsp, (frame_size_in_words - 2) * wordSize);
3196
3197 // Restore frame locals after moving the frame
3198 __ movdbl(Address(rsp, RegisterSaver::xmm0_offset_in_bytes()), xmm0);
3199 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax);
3200
3201 // Call C code. Need thread but NOT official VM entry
3202 // crud. We cannot block on this call, no GC can happen. Call should
3203 // restore return values to their stack-slots with the new SP.
3204 //
3205 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
3206
3207 // Use rbp because the frames look interpreted now
3208 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP.
3209 // Don't need the precise return PC here, just precise enough to point into this code blob.
3210 address the_pc = __ pc();
3211 __ set_last_Java_frame(noreg, rbp, the_pc, rscratch1);
3212
3213 __ andptr(rsp, -(StackAlignmentInBytes)); // Fix stack alignment as required by ABI
3214 __ mov(c_rarg0, r15_thread);
3215 __ movl(c_rarg1, r14); // second arg: exec_mode
3216 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
3217 // Revert SP alignment after call since we're going to do some SP relative addressing below
3218 __ movptr(rsp, Address(r15_thread, JavaThread::last_Java_sp_offset()));
3219
3220 // Set an oopmap for the call site
3221 // Use the same PC we used for the last java frame
3222 oop_maps->add_gc_map(the_pc - start,
3223 new OopMap( frame_size_in_words, 0 ));
3224
3225 // Clear fp AND pc
3226 __ reset_last_Java_frame(true);
3227
3228 // Collect return values
3229 __ movdbl(xmm0, Address(rsp, RegisterSaver::xmm0_offset_in_bytes()));
3230 __ movptr(rax, Address(rsp, RegisterSaver::rax_offset_in_bytes()));
3231 // I think this is useless (throwing pc?)
3232 __ movptr(rdx, Address(rsp, RegisterSaver::rdx_offset_in_bytes()));
3233
3234 // Pop self-frame.
3235 __ leave(); // Epilog
3236
3237 // Jump to interpreter
3238 __ ret(0);
3239
3240 // Make sure all code is generated
3241 masm->flush();
3242
3243 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
3244 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3245 #if INCLUDE_JVMCI
3246 if (EnableJVMCI) {
3247 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
3248 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
3249 }
3250 #endif
3251 }
3252
3253 //------------------------------generate_handler_blob------
3254 //
3255 // Generate a special Compile2Runtime blob that saves all registers,
3256 // and setup oopmap.
3257 //
3258 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
3259 assert(StubRoutines::forward_exception_entry() != nullptr,
3260 "must be generated before");
3261
3262 ResourceMark rm;
3263 OopMapSet *oop_maps = new OopMapSet();
3264 OopMap* map;
3265
3266 // Allocate space for the code. Setup code generation tools.
3267 CodeBuffer buffer("handler_blob", 2348, 1024);
3268 MacroAssembler* masm = new MacroAssembler(&buffer);
3269
3270 address start = __ pc();
3271 address call_pc = nullptr;
3272 int frame_size_in_words;
3273 bool cause_return = (poll_type == POLL_AT_RETURN);
3274 bool save_wide_vectors = (poll_type == POLL_AT_VECTOR_LOOP);
3275
3276 // Make room for return address (or push it again)
3277 if (!cause_return) {
3278 __ push(rbx);
3279 }
3280
3281 // Save registers, fpu state, and flags
3282 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, save_wide_vectors);
3283
3284 // The following is basically a call_VM. However, we need the precise
3285 // address of the call in order to generate an oopmap. Hence, we do all the
3286 // work ourselves.
3287
3288 __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1); // JavaFrameAnchor::capture_last_Java_pc() will get the pc from the return address, which we store next:
3289
3290 // The return address must always be correct so that frame constructor never
3291 // sees an invalid pc.
3292
3293 if (!cause_return) {
3294 // Get the return pc saved by the signal handler and stash it in its appropriate place on the stack.
3295 // Additionally, rbx is a callee saved register and we can look at it later to determine
3296 // if someone changed the return address for us!
3297 __ movptr(rbx, Address(r15_thread, JavaThread::saved_exception_pc_offset()));
3298 __ movptr(Address(rbp, wordSize), rbx);
3299 }
3300
3301 // Do the call
3302 __ mov(c_rarg0, r15_thread);
3303 __ call(RuntimeAddress(call_ptr));
3304
3305 // Set an oopmap for the call site. This oopmap will map all
3306 // oop-registers and debug-info registers as callee-saved. This
3307 // will allow deoptimization at this safepoint to find all possible
3308 // debug-info recordings, as well as let GC find all oops.
3309
3310 oop_maps->add_gc_map( __ pc() - start, map);
3311
3312 Label noException;
3313
3314 __ reset_last_Java_frame(false);
3315
3316 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
3317 __ jcc(Assembler::equal, noException);
3318
3319 // Exception pending
3320
3321 RegisterSaver::restore_live_registers(masm, save_wide_vectors);
3322
3323 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3324
3325 // No exception case
3326 __ bind(noException);
3327
3328 Label no_adjust;
3329 #ifdef ASSERT
3330 Label bail;
3331 #endif
3332 if (!cause_return) {
3333 Label no_prefix, not_special;
3334
3335 // If our stashed return pc was modified by the runtime we avoid touching it
3336 __ cmpptr(rbx, Address(rbp, wordSize));
3337 __ jccb(Assembler::notEqual, no_adjust);
3338
3339 // Skip over the poll instruction.
3340 // See NativeInstruction::is_safepoint_poll()
3341 // Possible encodings:
3342 // 85 00 test %eax,(%rax)
3343 // 85 01 test %eax,(%rcx)
3344 // 85 02 test %eax,(%rdx)
3345 // 85 03 test %eax,(%rbx)
3346 // 85 06 test %eax,(%rsi)
3347 // 85 07 test %eax,(%rdi)
3348 //
3349 // 41 85 00 test %eax,(%r8)
3350 // 41 85 01 test %eax,(%r9)
3351 // 41 85 02 test %eax,(%r10)
3352 // 41 85 03 test %eax,(%r11)
3353 // 41 85 06 test %eax,(%r14)
3354 // 41 85 07 test %eax,(%r15)
3355 //
3356 // 85 04 24 test %eax,(%rsp)
3357 // 41 85 04 24 test %eax,(%r12)
3358 // 85 45 00 test %eax,0x0(%rbp)
3359 // 41 85 45 00 test %eax,0x0(%r13)
3360
3361 __ cmpb(Address(rbx, 0), NativeTstRegMem::instruction_rex_b_prefix);
3362 __ jcc(Assembler::notEqual, no_prefix);
3363 __ addptr(rbx, 1);
3364 __ bind(no_prefix);
3365 #ifdef ASSERT
3366 __ movptr(rax, rbx); // remember where 0x85 should be, for verification below
3367 #endif
3368 // r12/r13/rsp/rbp base encoding takes 3 bytes with the following register values:
3369 // r12/rsp 0x04
3370 // r13/rbp 0x05
3371 __ movzbq(rcx, Address(rbx, 1));
3372 __ andptr(rcx, 0x07); // looking for 0x04 .. 0x05
3373 __ subptr(rcx, 4); // looking for 0x00 .. 0x01
3374 __ cmpptr(rcx, 1);
3375 __ jcc(Assembler::above, not_special);
3376 __ addptr(rbx, 1);
3377 __ bind(not_special);
3378 #ifdef ASSERT
3379 // Verify the correct encoding of the poll we're about to skip.
3380 __ cmpb(Address(rax, 0), NativeTstRegMem::instruction_code_memXregl);
3381 __ jcc(Assembler::notEqual, bail);
3382 // Mask out the modrm bits
3383 __ testb(Address(rax, 1), NativeTstRegMem::modrm_mask);
3384 // rax encodes to 0, so if the bits are nonzero it's incorrect
3385 __ jcc(Assembler::notZero, bail);
3386 #endif
3387 // Adjust return pc forward to step over the safepoint poll instruction
3388 __ addptr(rbx, 2);
3389 __ movptr(Address(rbp, wordSize), rbx);
3390 }
3391
3392 __ bind(no_adjust);
3393 // Normal exit, restore registers and exit.
3394 RegisterSaver::restore_live_registers(masm, save_wide_vectors);
3395 __ ret(0);
3396
3397 #ifdef ASSERT
3398 __ bind(bail);
3399 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
3400 #endif
3401
3402 // Make sure all code is generated
3403 masm->flush();
3404
3405 // Fill-out other meta info
3406 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
3407 }
3408
3409 //
3410 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3411 //
3412 // Generate a stub that calls into vm to find out the proper destination
3413 // of a java call. All the argument registers are live at this point
3414 // but since this is generic code we don't know what they are and the caller
3415 // must do any gc of the args.
3416 //
3417 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3418 assert (StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
3419
3420 // allocate space for the code
3421 ResourceMark rm;
3422
3423 CodeBuffer buffer(name, 1552, 512);
3424 MacroAssembler* masm = new MacroAssembler(&buffer);
3425
3426 int frame_size_in_words;
3427
3428 OopMapSet *oop_maps = new OopMapSet();
3429 OopMap* map = nullptr;
3430
3431 int start = __ offset();
3432
3433 // No need to save vector registers since they are caller-saved anyway.
3434 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_in_words, /*save_wide_vectors*/ false);
3435
3436 int frame_complete = __ offset();
3437
3438 __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1);
3439
3440 __ mov(c_rarg0, r15_thread);
3441
3442 __ call(RuntimeAddress(destination));
3443
3444
3445 // Set an oopmap for the call site.
3446 // We need this not only for callee-saved registers, but also for volatile
3447 // registers that the compiler might be keeping live across a safepoint.
3448
3449 oop_maps->add_gc_map( __ offset() - start, map);
3450
3451 // rax contains the address we are going to jump to assuming no exception got installed
3452
3453 // clear last_Java_sp
3454 __ reset_last_Java_frame(false);
3455 // check for pending exceptions
3456 Label pending;
3457 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
3458 __ jcc(Assembler::notEqual, pending);
3459
3460 // get the returned Method*
3461 __ get_vm_result_2(rbx, r15_thread);
3462 __ movptr(Address(rsp, RegisterSaver::rbx_offset_in_bytes()), rbx);
3463
3464 __ movptr(Address(rsp, RegisterSaver::rax_offset_in_bytes()), rax);
3465
3466 RegisterSaver::restore_live_registers(masm);
3467
3468 // We are back to the original state on entry and ready to go.
3469
3470 __ jmp(rax);
3471
3472 // Pending exception after the safepoint
3473
3474 __ bind(pending);
3475
3476 RegisterSaver::restore_live_registers(masm);
3477
3478 // exception pending => remove activation and forward to exception handler
3479
3480 __ movptr(Address(r15_thread, JavaThread::vm_result_offset()), NULL_WORD);
3481
3482 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
3483 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3484
3485 // -------------
3486 // make sure all code is generated
3487 masm->flush();
3488
3489 // return the blob
3490 // frame_size_words or bytes??
3491 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
3492 }
3493
3494 // Continuation point for throwing of implicit exceptions that are
3495 // not handled in the current activation. Fabricates an exception
3496 // oop and initiates normal exception dispatching in this
3497 // frame. Since we need to preserve callee-saved values (currently
3498 // only for C2, but done for C1 as well) we need a callee-saved oop
3499 // map and therefore have to make these stubs into RuntimeStubs
3500 // rather than BufferBlobs. If the compiler needs all registers to
3501 // be preserved between the fault point and the exception handler
3502 // then it must assume responsibility for that in
3503 // AbstractCompiler::continuation_for_implicit_null_exception or
3504 // continuation_for_implicit_division_by_zero_exception. All other
3505 // implicit exceptions (e.g., NullPointerException or
3506 // AbstractMethodError on entry) are either at call sites or
3507 // otherwise assume that stack unwinding will be initiated, so
3508 // caller saved registers were assumed volatile in the compiler.
3509 RuntimeStub* SharedRuntime::generate_throw_exception(const char* name, address runtime_entry) {
3510 // Information about frame layout at time of blocking runtime call.
3511 // Note that we only have to preserve callee-saved registers since
3512 // the compilers are responsible for supplying a continuation point
3513 // if they expect all registers to be preserved.
3514 enum layout {
3515 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
3516 rbp_off2,
3517 return_off,
3518 return_off2,
3519 framesize // inclusive of return address
3520 };
3521
3522 int insts_size = 512;
3523 int locs_size = 64;
3524
3525 ResourceMark rm;
3526 const char* timer_msg = "SharedRuntime generate_throw_exception";
3527 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3528
3529 CodeBuffer code(name, insts_size, locs_size);
3530 OopMapSet* oop_maps = new OopMapSet();
3531 MacroAssembler* masm = new MacroAssembler(&code);
3532
3533 address start = __ pc();
3534
3535 // This is an inlined and slightly modified version of call_VM
3536 // which has the ability to fetch the return PC out of
3537 // thread-local storage and also sets up last_Java_sp slightly
3538 // differently than the real call_VM
3539
3540 __ enter(); // required for proper stackwalking of RuntimeStub frame
3541
3542 assert(is_even(framesize/2), "sp not 16-byte aligned");
3543
3544 // return address and rbp are already in place
3545 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
3546
3547 int frame_complete = __ pc() - start;
3548
3549 // Set up last_Java_sp and last_Java_fp
3550 address the_pc = __ pc();
3551 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch1);
3552 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack
3553
3554 // Call runtime
3555 __ movptr(c_rarg0, r15_thread);
3556 BLOCK_COMMENT("call runtime_entry");
3557 __ call(RuntimeAddress(runtime_entry));
3558
3559 // Generate oop map
3560 OopMap* map = new OopMap(framesize, 0);
3561
3562 oop_maps->add_gc_map(the_pc - start, map);
3563
3564 __ reset_last_Java_frame(true);
3565
3566 __ leave(); // required for proper stackwalking of RuntimeStub frame
3567
3568 // check for pending exceptions
3569 #ifdef ASSERT
3570 Label L;
3571 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
3572 __ jcc(Assembler::notEqual, L);
3573 __ should_not_reach_here();
3574 __ bind(L);
3575 #endif // ASSERT
3576 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3577
3578
3579 // codeBlob framesize is in words (not VMRegImpl::slot_size)
3580 RuntimeStub* stub =
3581 RuntimeStub::new_runtime_stub(name,
3582 &code,
3583 frame_complete,
3584 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3585 oop_maps, false);
3586 return stub;
3587 }
3588
3589 //------------------------------Montgomery multiplication------------------------
3590 //
3591
3592 #ifndef _WINDOWS
3593
3594 // Subtract 0:b from carry:a. Return carry.
3595 static julong
3596 sub(julong a[], julong b[], julong carry, long len) {
3597 long long i = 0, cnt = len;
3598 julong tmp;
3599 asm volatile("clc; "
3600 "0: ; "
3601 "mov (%[b], %[i], 8), %[tmp]; "
3602 "sbb %[tmp], (%[a], %[i], 8); "
3603 "inc %[i]; dec %[cnt]; "
3604 "jne 0b; "
3605 "mov %[carry], %[tmp]; sbb $0, %[tmp]; "
3606 : [i]"+r"(i), [cnt]"+r"(cnt), [tmp]"=&r"(tmp)
3607 : [a]"r"(a), [b]"r"(b), [carry]"r"(carry)
3608 : "memory");
3609 return tmp;
3610 }
3611
3612 // Multiply (unsigned) Long A by Long B, accumulating the double-
3613 // length result into the accumulator formed of T0, T1, and T2.
3614 #define MACC(A, B, T0, T1, T2) \
3615 do { \
3616 unsigned long hi, lo; \
3617 __asm__ ("mul %5; add %%rax, %2; adc %%rdx, %3; adc $0, %4" \
3618 : "=&d"(hi), "=a"(lo), "+r"(T0), "+r"(T1), "+g"(T2) \
3619 : "r"(A), "a"(B) : "cc"); \
3620 } while(0)
3621
3622 // As above, but add twice the double-length result into the
3623 // accumulator.
3624 #define MACC2(A, B, T0, T1, T2) \
3625 do { \
3626 unsigned long hi, lo; \
3627 __asm__ ("mul %5; add %%rax, %2; adc %%rdx, %3; adc $0, %4; " \
3628 "add %%rax, %2; adc %%rdx, %3; adc $0, %4" \
3629 : "=&d"(hi), "=a"(lo), "+r"(T0), "+r"(T1), "+g"(T2) \
3630 : "r"(A), "a"(B) : "cc"); \
3631 } while(0)
3632
3633 #else //_WINDOWS
3634
3635 static julong
3636 sub(julong a[], julong b[], julong carry, long len) {
3637 long i;
3638 julong tmp;
3639 unsigned char c = 1;
3640 for (i = 0; i < len; i++) {
3641 c = _addcarry_u64(c, a[i], ~b[i], &tmp);
3642 a[i] = tmp;
3643 }
3644 c = _addcarry_u64(c, carry, ~0, &tmp);
3645 return tmp;
3646 }
3647
3648 // Multiply (unsigned) Long A by Long B, accumulating the double-
3649 // length result into the accumulator formed of T0, T1, and T2.
3650 #define MACC(A, B, T0, T1, T2) \
3651 do { \
3652 julong hi, lo; \
3653 lo = _umul128(A, B, &hi); \
3654 unsigned char c = _addcarry_u64(0, lo, T0, &T0); \
3655 c = _addcarry_u64(c, hi, T1, &T1); \
3656 _addcarry_u64(c, T2, 0, &T2); \
3657 } while(0)
3658
3659 // As above, but add twice the double-length result into the
3660 // accumulator.
3661 #define MACC2(A, B, T0, T1, T2) \
3662 do { \
3663 julong hi, lo; \
3664 lo = _umul128(A, B, &hi); \
3665 unsigned char c = _addcarry_u64(0, lo, T0, &T0); \
3666 c = _addcarry_u64(c, hi, T1, &T1); \
3667 _addcarry_u64(c, T2, 0, &T2); \
3668 c = _addcarry_u64(0, lo, T0, &T0); \
3669 c = _addcarry_u64(c, hi, T1, &T1); \
3670 _addcarry_u64(c, T2, 0, &T2); \
3671 } while(0)
3672
3673 #endif //_WINDOWS
3674
3675 // Fast Montgomery multiplication. The derivation of the algorithm is
3676 // in A Cryptographic Library for the Motorola DSP56000,
3677 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237.
3678
3679 static void NOINLINE
3680 montgomery_multiply(julong a[], julong b[], julong n[],
3681 julong m[], julong inv, int len) {
3682 julong t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3683 int i;
3684
3685 assert(inv * n[0] == ULLONG_MAX, "broken inverse in Montgomery multiply");
3686
3687 for (i = 0; i < len; i++) {
3688 int j;
3689 for (j = 0; j < i; j++) {
3690 MACC(a[j], b[i-j], t0, t1, t2);
3691 MACC(m[j], n[i-j], t0, t1, t2);
3692 }
3693 MACC(a[i], b[0], t0, t1, t2);
3694 m[i] = t0 * inv;
3695 MACC(m[i], n[0], t0, t1, t2);
3696
3697 assert(t0 == 0, "broken Montgomery multiply");
3698
3699 t0 = t1; t1 = t2; t2 = 0;
3700 }
3701
3702 for (i = len; i < 2*len; i++) {
3703 int j;
3704 for (j = i-len+1; j < len; j++) {
3705 MACC(a[j], b[i-j], t0, t1, t2);
3706 MACC(m[j], n[i-j], t0, t1, t2);
3707 }
3708 m[i-len] = t0;
3709 t0 = t1; t1 = t2; t2 = 0;
3710 }
3711
3712 while (t0)
3713 t0 = sub(m, n, t0, len);
3714 }
3715
3716 // Fast Montgomery squaring. This uses asymptotically 25% fewer
3717 // multiplies so it should be up to 25% faster than Montgomery
3718 // multiplication. However, its loop control is more complex and it
3719 // may actually run slower on some machines.
3720
3721 static void NOINLINE
3722 montgomery_square(julong a[], julong n[],
3723 julong m[], julong inv, int len) {
3724 julong t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3725 int i;
3726
3727 assert(inv * n[0] == ULLONG_MAX, "broken inverse in Montgomery square");
3728
3729 for (i = 0; i < len; i++) {
3730 int j;
3731 int end = (i+1)/2;
3732 for (j = 0; j < end; j++) {
3733 MACC2(a[j], a[i-j], t0, t1, t2);
3734 MACC(m[j], n[i-j], t0, t1, t2);
3735 }
3736 if ((i & 1) == 0) {
3737 MACC(a[j], a[j], t0, t1, t2);
3738 }
3739 for (; j < i; j++) {
3740 MACC(m[j], n[i-j], t0, t1, t2);
3741 }
3742 m[i] = t0 * inv;
3743 MACC(m[i], n[0], t0, t1, t2);
3744
3745 assert(t0 == 0, "broken Montgomery square");
3746
3747 t0 = t1; t1 = t2; t2 = 0;
3748 }
3749
3750 for (i = len; i < 2*len; i++) {
3751 int start = i-len+1;
3752 int end = start + (len - start)/2;
3753 int j;
3754 for (j = start; j < end; j++) {
3755 MACC2(a[j], a[i-j], t0, t1, t2);
3756 MACC(m[j], n[i-j], t0, t1, t2);
3757 }
3758 if ((i & 1) == 0) {
3759 MACC(a[j], a[j], t0, t1, t2);
3760 }
3761 for (; j < len; j++) {
3762 MACC(m[j], n[i-j], t0, t1, t2);
3763 }
3764 m[i-len] = t0;
3765 t0 = t1; t1 = t2; t2 = 0;
3766 }
3767
3768 while (t0)
3769 t0 = sub(m, n, t0, len);
3770 }
3771
3772 // Swap words in a longword.
3773 static julong swap(julong x) {
3774 return (x << 32) | (x >> 32);
3775 }
3776
3777 // Copy len longwords from s to d, word-swapping as we go. The
3778 // destination array is reversed.
3779 static void reverse_words(julong *s, julong *d, int len) {
3780 d += len;
3781 while(len-- > 0) {
3782 d--;
3783 *d = swap(*s);
3784 s++;
3785 }
3786 }
3787
3788 // The threshold at which squaring is advantageous was determined
3789 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz.
3790 #define MONTGOMERY_SQUARING_THRESHOLD 64
3791
3792 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints,
3793 jint len, jlong inv,
3794 jint *m_ints) {
3795 assert(len % 2 == 0, "array length in montgomery_multiply must be even");
3796 int longwords = len/2;
3797
3798 // Make very sure we don't use so much space that the stack might
3799 // overflow. 512 jints corresponds to an 16384-bit integer and
3800 // will use here a total of 8k bytes of stack space.
3801 int divisor = sizeof(julong) * 4;
3802 guarantee(longwords <= 8192 / divisor, "must be");
3803 int total_allocation = longwords * sizeof (julong) * 4;
3804 julong *scratch = (julong *)alloca(total_allocation);
3805
3806 // Local scratch arrays
3807 julong
3808 *a = scratch + 0 * longwords,
3809 *b = scratch + 1 * longwords,
3810 *n = scratch + 2 * longwords,
3811 *m = scratch + 3 * longwords;
3812
3813 reverse_words((julong *)a_ints, a, longwords);
3814 reverse_words((julong *)b_ints, b, longwords);
3815 reverse_words((julong *)n_ints, n, longwords);
3816
3817 ::montgomery_multiply(a, b, n, m, (julong)inv, longwords);
3818
3819 reverse_words(m, (julong *)m_ints, longwords);
3820 }
3821
3822 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints,
3823 jint len, jlong inv,
3824 jint *m_ints) {
3825 assert(len % 2 == 0, "array length in montgomery_square must be even");
3826 int longwords = len/2;
3827
3828 // Make very sure we don't use so much space that the stack might
3829 // overflow. 512 jints corresponds to an 16384-bit integer and
3830 // will use here a total of 6k bytes of stack space.
3831 int divisor = sizeof(julong) * 3;
3832 guarantee(longwords <= (8192 / divisor), "must be");
3833 int total_allocation = longwords * sizeof (julong) * 3;
3834 julong *scratch = (julong *)alloca(total_allocation);
3835
3836 // Local scratch arrays
3837 julong
3838 *a = scratch + 0 * longwords,
3839 *n = scratch + 1 * longwords,
3840 *m = scratch + 2 * longwords;
3841
3842 reverse_words((julong *)a_ints, a, longwords);
3843 reverse_words((julong *)n_ints, n, longwords);
3844
3845 if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3846 ::montgomery_square(a, n, m, (julong)inv, longwords);
3847 } else {
3848 ::montgomery_multiply(a, a, n, m, (julong)inv, longwords);
3849 }
3850
3851 reverse_words(m, (julong *)m_ints, longwords);
3852 }
3853
3854 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
3855 BufferBlob* buf = BufferBlob::create("inline types pack/unpack", 16 * K);
3856 CodeBuffer buffer(buf);
3857 short buffer_locs[20];
3858 buffer.insts()->initialize_shared_locs((relocInfo*)buffer_locs,
3859 sizeof(buffer_locs)/sizeof(relocInfo));
3860
3861 MacroAssembler* masm = new MacroAssembler(&buffer);
3862
3863 const Array<SigEntry>* sig_vk = vk->extended_sig();
3864 const Array<VMRegPair>* regs = vk->return_regs();
3865
3866 int pack_fields_jobject_off = __ offset();
3867 // Resolve pre-allocated buffer from JNI handle.
3868 // We cannot do this in generate_call_stub() because it requires GC code to be initialized.
3869 __ movptr(rax, Address(r13, 0));
3870 __ resolve_jobject(rax /* value */,
3871 r15_thread /* thread */,
3872 r12 /* tmp */);
3873 __ movptr(Address(r13, 0), rax);
3874
3875 int pack_fields_off = __ offset();
3876
3877 int j = 1;
3878 for (int i = 0; i < sig_vk->length(); i++) {
3879 BasicType bt = sig_vk->at(i)._bt;
3880 if (bt == T_METADATA) {
3881 continue;
3882 }
3883 if (bt == T_VOID) {
3884 if (sig_vk->at(i-1)._bt == T_LONG ||
3885 sig_vk->at(i-1)._bt == T_DOUBLE) {
3886 j++;
3887 }
3888 continue;
3889 }
3890 int off = sig_vk->at(i)._offset;
3891 assert(off > 0, "offset in object should be positive");
3892 VMRegPair pair = regs->at(j);
3893 VMReg r_1 = pair.first();
3894 VMReg r_2 = pair.second();
3895 Address to(rax, off);
3896 if (bt == T_FLOAT) {
3897 __ movflt(to, r_1->as_XMMRegister());
3898 } else if (bt == T_DOUBLE) {
3899 __ movdbl(to, r_1->as_XMMRegister());
3900 } else {
3901 Register val = r_1->as_Register();
3902 assert_different_registers(to.base(), val, r14, r13, rbx, rscratch1);
3903 if (is_reference_type(bt)) {
3904 __ store_heap_oop(to, val, r14, r13, rbx, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
3905 } else {
3906 __ store_sized_value(to, r_1->as_Register(), type2aelembytes(bt));
3907 }
3908 }
3909 j++;
3910 }
3911 assert(j == regs->length(), "missed a field?");
3912
3913 __ ret(0);
3914
3915 int unpack_fields_off = __ offset();
3916
3917 Label skip;
3918 __ testptr(rax, rax);
3919 __ jcc(Assembler::zero, skip);
3920
3921 j = 1;
3922 for (int i = 0; i < sig_vk->length(); i++) {
3923 BasicType bt = sig_vk->at(i)._bt;
3924 if (bt == T_METADATA) {
3925 continue;
3926 }
3927 if (bt == T_VOID) {
3928 if (sig_vk->at(i-1)._bt == T_LONG ||
3929 sig_vk->at(i-1)._bt == T_DOUBLE) {
3930 j++;
3931 }
3932 continue;
3933 }
3934 int off = sig_vk->at(i)._offset;
3935 assert(off > 0, "offset in object should be positive");
3936 VMRegPair pair = regs->at(j);
3937 VMReg r_1 = pair.first();
3938 VMReg r_2 = pair.second();
3939 Address from(rax, off);
3940 if (bt == T_FLOAT) {
3941 __ movflt(r_1->as_XMMRegister(), from);
3942 } else if (bt == T_DOUBLE) {
3943 __ movdbl(r_1->as_XMMRegister(), from);
3944 } else if (bt == T_OBJECT || bt == T_ARRAY) {
3945 assert_different_registers(rax, r_1->as_Register());
3946 __ load_heap_oop(r_1->as_Register(), from);
3947 } else {
3948 assert(is_java_primitive(bt), "unexpected basic type");
3949 assert_different_registers(rax, r_1->as_Register());
3950 size_t size_in_bytes = type2aelembytes(bt);
3951 __ load_sized_value(r_1->as_Register(), from, size_in_bytes, bt != T_CHAR && bt != T_BOOLEAN);
3952 }
3953 j++;
3954 }
3955 assert(j == regs->length(), "missed a field?");
3956
3957 __ bind(skip);
3958 __ ret(0);
3959
3960 __ flush();
3961
3962 return BufferedInlineTypeBlob::create(&buffer, pack_fields_off, pack_fields_jobject_off, unpack_fields_off);
3963 }
3964
3965 #if INCLUDE_JFR
3966
3967 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3968 // It returns a jobject handle to the event writer.
3969 // The handle is dereferenced and the return value is the event writer oop.
3970 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3971 enum layout {
3972 rbp_off,
3973 rbpH_off,
3974 return_off,
3975 return_off2,
3976 framesize // inclusive of return address
3977 };
3978
3979 CodeBuffer code("jfr_write_checkpoint", 1024, 64);
3980 MacroAssembler* masm = new MacroAssembler(&code);
3981 address start = __ pc();
3982
3983 __ enter();
3984 address the_pc = __ pc();
3985
3986 int frame_complete = the_pc - start;
3987
3988 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch1);
3989 __ movptr(c_rarg0, r15_thread);
3990 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
3991 __ reset_last_Java_frame(true);
3992
3993 // rax is jobject handle result, unpack and process it through a barrier.
3994 __ resolve_global_jobject(rax, r15_thread, c_rarg0);
3995
3996 __ leave();
3997 __ ret(0);
3998
3999 OopMapSet* oop_maps = new OopMapSet();
4000 OopMap* map = new OopMap(framesize, 1);
4001 oop_maps->add_gc_map(frame_complete, map);
4002
4003 RuntimeStub* stub =
4004 RuntimeStub::new_runtime_stub(code.name(),
4005 &code,
4006 frame_complete,
4007 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
4008 oop_maps,
4009 false);
4010 return stub;
4011 }
4012
4013 // For c2: call to return a leased buffer.
4014 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
4015 enum layout {
4016 rbp_off,
4017 rbpH_off,
4018 return_off,
4019 return_off2,
4020 framesize // inclusive of return address
4021 };
4022
4023 CodeBuffer code("jfr_return_lease", 1024, 64);
4024 MacroAssembler* masm = new MacroAssembler(&code);
4025 address start = __ pc();
4026
4027 __ enter();
4028 address the_pc = __ pc();
4029
4030 int frame_complete = the_pc - start;
4031
4032 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch2);
4033 __ movptr(c_rarg0, r15_thread);
4034 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
4035 __ reset_last_Java_frame(true);
4036
4037 __ leave();
4038 __ ret(0);
4039
4040 OopMapSet* oop_maps = new OopMapSet();
4041 OopMap* map = new OopMap(framesize, 1);
4042 oop_maps->add_gc_map(frame_complete, map);
4043
4044 RuntimeStub* stub =
4045 RuntimeStub::new_runtime_stub(code.name(),
4046 &code,
4047 frame_complete,
4048 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
4049 oop_maps,
4050 false);
4051 return stub;
4052 }
4053
4054 #endif // INCLUDE_JFR