mono/mini/tramp-amd64.c

   1 /*
   2  * tramp-amd64.c: JIT trampoline code for amd64
   3  *
   4  * Authors:
   5  *   Dietmar Maurer (dietmar@ximian.com)
   6  *   Zoltan Varga (vargaz@gmail.com)
   7  *
   8  * (C) 2001 Ximian, Inc.
   9  */
  10
  11 #include <config.h>
  12 #include <glib.h>
  13
  14 #include <mono/metadata/appdomain.h>
  15 #include <mono/metadata/marshal.h>
  16 #include <mono/metadata/tabledefs.h>
  17 #include <mono/metadata/mono-debug-debugger.h>
  18 #include <mono/metadata/monitor.h>
  19 #include <mono/arch/amd64/amd64-codegen.h>
  20
  21 #include <mono/utils/memcheck.h>
  22
  23 #include "mini.h"
  24 #include "mini-amd64.h"
  25
  26 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
  27
  28 static guint8* nullified_class_init_trampoline;
  29
  30 /*
  31  * mono_arch_get_unbox_trampoline:
  32  * @gsctx: the generic sharing context
  33  * @m: method pointer
  34  * @addr: pointer to native code for @m
  35  *
  36  * when value type methods are called through the vtable we need to unbox the
  37  * this argument. This method returns a pointer to a trampoline which does
  38  * unboxing before calling the method
  39  */
  40 gpointer
  41 mono_arch_get_unbox_trampoline (MonoGenericSharingContext *gsctx, MonoMethod *m, gpointer addr)
  42 {
  43         guint8 *code, *start;
  44         int this_reg;
  45
  46         MonoDomain *domain = mono_domain_get ();
  47
  48         this_reg = mono_arch_get_this_arg_reg (mono_method_signature (m), gsctx, NULL);
  49
  50         start = code = mono_domain_code_reserve (domain, 20);
  51
  52         amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
  53         /* FIXME: Optimize this */
  54         amd64_mov_reg_imm (code, AMD64_RAX, addr);
  55         amd64_jump_reg (code, AMD64_RAX);
  56         g_assert ((code - start) < 20);
  57
  58         mono_arch_flush_icache (start, code - start);
  59
  60         return start;
  61 }
  62
  63 /*
  64  * mono_arch_get_static_rgctx_trampoline:
  65  *
  66  *   Create a trampoline which sets RGCTX_REG to MRGCTX, then jumps to ADDR.
  67  */
  68 gpointer
  69 mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr)
  70 {
  71         guint8 *code, *start;
  72         int buf_len;
  73
  74         MonoDomain *domain = mono_domain_get ();
  75
  76 #ifdef MONO_ARCH_NOMAP32BIT
  77         buf_len = 32;
  78 #else
  79         /* AOTed code could still have a non-32 bit address */
  80         if ((((guint64)addr) >> 32) == 0)
  81                 buf_len = 16;
  82         else
  83                 buf_len = 30;
  84 #endif
  85
  86         start = code = mono_domain_code_reserve (domain, buf_len);
  87
  88         amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, mrgctx);
  89         amd64_jump_code (code, addr);
  90         g_assert ((code - start) < buf_len);
  91
  92         mono_arch_flush_icache (start, code - start);
  93
  94         return start;
  95 }
  96
  97 gpointer
  98 mono_arch_get_llvm_imt_trampoline (MonoDomain *domain, MonoMethod *m, int vt_offset)
  99 {
 100         guint8 *code, *start;
 101         int buf_len;
 102         int this_reg;
 103
 104         buf_len = 32;
 105
 106         start = code = mono_domain_code_reserve (domain, buf_len);
 107
 108         this_reg = mono_arch_get_this_arg_reg (mono_method_signature (m), NULL, NULL);
 109
 110         /* Set imt arg */
 111         amd64_mov_reg_imm (code, MONO_ARCH_IMT_REG, m);
 112         /* Load vtable address */
 113         amd64_mov_reg_membase (code, AMD64_RAX, this_reg, 0, 8);
 114         amd64_jump_membase (code, AMD64_RAX, vt_offset);
 115         amd64_ret (code);
 116
 117         g_assert ((code - start) < buf_len);
 118
 119         mono_arch_flush_icache (start, code - start);
 120
 121         return start;
 122 }
 123
 124 /*
 125  * mono_arch_patch_callsite:
 126  *
 127  *   Patch the callsite whose address is given by ORIG_CODE so it calls ADDR. ORIG_CODE
 128  * points to the pc right after the call.
 129  */
 130 void
 131 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
 132 {
 133         guint8 *code;
 134         guint8 buf [16];
 135         gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
 136
 137         code = buf + 14;
 138
 139         if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
 140                 if (code [-5] != 0xe8) {
 141                         if (can_write) {
 142                                 InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
 143                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
 144                         }
 145                 } else {
 146                         if ((((guint64)(addr)) >> 32) != 0) {
 147 #ifdef MONO_ARCH_NOMAP32BIT
 148                                 /* Print some diagnostics */
 149                                 MonoJitInfo *ji = mono_jit_info_table_find (mono_domain_get (), (char*)orig_code);
 150                                 if (ji)
 151                                         fprintf (stderr, "At %s, offset 0x%zx\n", mono_method_full_name (ji->method, TRUE), (guint8*)orig_code - (guint8*)ji->code_start);
 152                                 fprintf (stderr, "Addr: %p\n", addr);
 153                                 ji = mono_jit_info_table_find (mono_domain_get (), (char*)addr);
 154                                 if (ji)
 155                                         fprintf (stderr, "Callee: %s\n", mono_method_full_name (ji->method, TRUE));
 156                                 g_assert_not_reached ();
 157 #else
 158                                 /*
 159                                  * This might happen when calling AOTed code. Create a thunk.
 160                                  */
 161                                 guint8 *thunk_start, *thunk_code;
 162
 163                                 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
 164                                 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
 165                                 *(guint64*)thunk_code = (guint64)addr;
 166                                 addr = thunk_start;
 167                                 g_assert ((((guint64)(addr)) >> 32) == 0);
 168                                 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
 169 #endif
 170                         }
 171                         g_assert ((((guint64)(orig_code)) >> 32) == 0);
 172                         if (can_write) {
 173                                 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
 174                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
 175                         }
 176                 }
 177         }
 178         else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
 179                 /* call *<OFFSET>(%rip) */
 180                 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
 181                 if (can_write) {
 182                         InterlockedExchangePointer (got_entry, addr);
 183                         VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
 184                 }
 185         }
 186 }
 187
 188 void
 189 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
 190 {
 191         gint32 disp;
 192         gpointer *plt_jump_table_entry;
 193
 194         /* A PLT entry: jmp *<DISP>(%rip) */
 195         g_assert (code [0] == 0xff);
 196         g_assert (code [1] == 0x25);
 197
 198         disp = *(gint32*)(code + 2);
 199
 200         plt_jump_table_entry = (gpointer*)(code + 6 + disp);
 201
 202         InterlockedExchangePointer (plt_jump_table_entry, addr);
 203 }
 204
 205 void
 206 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
 207 {
 208         guint8 buf [16];
 209         MonoJitInfo *ji = NULL;
 210         gboolean can_write;
 211
 212         if (mono_use_llvm) {
 213                 /* code - 7 might be before the start of the method */
 214                 /* FIXME: Avoid this expensive call somehow */
 215                 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
 216         }
 217
 218         can_write = mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 7, buf, sizeof (buf));
 219
 220         if (!can_write)
 221                 return;
 222
 223         code -= 3;
 224
 225         /*
 226          * A given byte sequence can match more than case here, so we have to be
 227          * really careful about the ordering of the cases. Longer sequences
 228          * come first.
 229          */
 230         if ((buf [0] == 0x41) && (buf [1] == 0xff) && (buf [2] == 0x15)) {
 231                 gpointer *vtable_slot;
 232
 233                 /* call *<OFFSET>(%rip) */
 234                 vtable_slot = mono_get_vcall_slot_addr (code + 3, regs);
 235                 g_assert (vtable_slot);
 236
 237                 *vtable_slot = nullified_class_init_trampoline;
 238         } else if (buf [2] == 0xe8) {
 239                 /* call <TARGET> */
 240                 //guint8 *buf = code - 2;
 241
 242                 /*
 243                  * It would be better to replace the call with nops, but that doesn't seem
 244                  * to work on SMP machines even when the whole call is inside a cache line.
 245                  * Patching the call address seems to work.
 246                  */
 247                 /*
 248                 buf [0] = 0x66;
 249                 buf [1] = 0x66;
 250                 buf [2] = 0x90;
 251                 buf [3] = 0x66;
 252                 buf [4] = 0x90;
 253                 */
 254
 255                 mono_arch_patch_callsite (code - 2, code - 2 + 5, nullified_class_init_trampoline);
 256         } else if ((buf [5] == 0xff) && x86_modrm_mod (buf [6]) == 3 && x86_modrm_reg (buf [6]) == 2) {
 257                 /* call *<reg> */
 258                 /* Generated by the LLVM JIT or on platforms without MAP_32BIT set */
 259                 guint8* buf = code;
 260
 261                 /* FIXME: Not thread safe */
 262                 buf [1] = 0x90;
 263                 buf [2] = 0x90;
 264         } else if (buf [4] == 0x90 || buf [5] == 0xeb || buf [6] == 0x66) {
 265                 /* Already changed by another thread */
 266                 ;
 267         } else {
 268                 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", buf [0], buf [1], buf [2], buf [3],
 269                         buf [4], buf [5], buf [6]);
 270                 g_assert_not_reached ();
 271         }
 272 }
 273
 274 void
 275 mono_arch_nullify_plt_entry (guint8 *code, mgreg_t *regs)
 276 {
 277         if (mono_aot_only && !nullified_class_init_trampoline)
 278                 nullified_class_init_trampoline = mono_aot_get_named_code ("nullified_class_init_trampoline");
 279
 280         mono_arch_patch_plt_entry (code, NULL, regs, nullified_class_init_trampoline);
 281 }
 282
 283 guchar*
 284 mono_arch_create_trampoline_code (MonoTrampolineType tramp_type)
 285 {
 286         MonoJumpInfo *ji;
 287         guint32 code_size;
 288         guchar *code;
 289         GSList *unwind_ops, *l;
 290
 291         code = mono_arch_create_trampoline_code_full (tramp_type, &code_size, &ji, &unwind_ops, FALSE);
 292
 293         mono_save_trampoline_xdebug_info ("<generic_trampoline>", code, code_size, unwind_ops);
 294
 295         for (l = unwind_ops; l; l = l->next)
 296                 g_free (l->data);
 297         g_slist_free (unwind_ops);
 298
 299         return code;
 300 }
 301
 302 guchar*
 303 mono_arch_create_trampoline_code_full (MonoTrampolineType tramp_type, guint32 *code_size, MonoJumpInfo **ji, GSList **out_unwind_ops, gboolean aot)
 304 {
 305         guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code;
 306         int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset, saved_regs_offset;
 307         int saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
 308         gboolean has_caller;
 309         GSList *unwind_ops = NULL;
 310
 311         if (tramp_type == MONO_TRAMPOLINE_JUMP)
 312                 has_caller = FALSE;
 313         else
 314                 has_caller = TRUE;
 315
 316         code = buf = mono_global_codeman_reserve (538);
 317
 318         *ji = NULL;
 319
 320         framesize = 538 + sizeof (MonoLMF);
 321         framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
 322
 323         orig_rsp_to_rbp_offset = 0;
 324         r11_save_code = code;
 325         /* Reserve 5 bytes for the mov_membase_reg to save R11 */
 326         code += 5;
 327         after_r11_save_code = code;
 328
 329         // CFA = sp + 16 (the trampoline address is on the stack)
 330         cfa_offset = 16;
 331         mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
 332         // IP saved at CFA - 8
 333         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
 334
 335         /* Pop the return address off the stack */
 336         amd64_pop_reg (code, AMD64_R11);
 337         orig_rsp_to_rbp_offset += 8;
 338
 339         cfa_offset -= 8;
 340         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 341
 342         /*
 343          * Allocate a new stack frame
 344          */
 345         amd64_push_reg (code, AMD64_RBP);
 346         cfa_offset += 8;
 347         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 348         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
 349
 350         orig_rsp_to_rbp_offset -= 8;
 351         amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, 8);
 352         mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
 353         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
 354
 355         offset = 0;
 356         rbp_offset = - offset;
 357
 358         offset += 8;
 359         rax_offset = - offset;
 360
 361         offset += 8;
 362         tramp_offset = - offset;
 363
 364         offset += 8;
 365         arg_offset = - offset;
 366
 367         /* Compute the trampoline address from the return address */
 368         if (aot) {
 369                 /* 7 = length of call *<offset>(rip) */
 370                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
 371         } else {
 372                 /* 5 = length of amd64_call_membase () */
 373                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
 374         }
 375         amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, 8);
 376
 377         offset += 8;
 378         res_offset = - offset;
 379
 380         /* Save all registers */
 381
 382         offset += AMD64_NREG * 8;
 383         saved_regs_offset = - offset;
 384         for (i = 0; i < AMD64_NREG; ++i) {
 385                 if (i == AMD64_RBP) {
 386                         /* RAX is already saved */
 387                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, 8);
 388                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * 8), AMD64_RAX, 8);
 389                 } else if (i != AMD64_R11) {
 390                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * 8), i, 8);
 391                 } else {
 392                         /* We have to save R11 right at the start of
 393                            the trampoline code because it's used as a
 394                            scratch register */
 395                         amd64_mov_membase_reg (r11_save_code, AMD64_RSP, saved_regs_offset + orig_rsp_to_rbp_offset + (i * 8), i, 8);
 396                         g_assert (r11_save_code == after_r11_save_code);
 397                 }
 398         }
 399         offset += 8 * 8;
 400         saved_fpregs_offset = - offset;
 401         for (i = 0; i < 8; ++i)
 402                 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * 8), i);
 403
 404         if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
 405                         tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER &&
 406                         tramp_type != MONO_TRAMPOLINE_MONITOR_EXIT) {
 407                 /* Obtain the trampoline argument which is encoded in the instruction stream */
 408                 if (aot) {
 409                         /* Load the GOT offset */
 410                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, 8);
 411                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 7, 4);
 412                         /* Compute the address of the GOT slot */
 413                         amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, 8);
 414                         /* Load the value */
 415                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
 416                 } else {
 417                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, 8);
 418                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
 419                         amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
 420                         amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
 421                         br [0] = code;
 422                         x86_branch8 (code, X86_CC_NE, 6, FALSE);
 423                         /* 32 bit immediate */
 424                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
 425                         br [1] = code;
 426                         x86_jump8 (code, 10);
 427                         /* 64 bit immediate */
 428                         mono_amd64_patch (br [0], code);
 429                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
 430                         mono_amd64_patch (br [1], code);
 431                 }
 432                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, 8);
 433         } else {
 434                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * 8), 8);
 435                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, 8);
 436         }
 437
 438         /* Save LMF begin */
 439
 440         offset += sizeof (MonoLMF);
 441         lmf_offset = - offset;
 442
 443         /* Save ip */
 444         if (has_caller)
 445                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, 8);
 446         else
 447                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 448         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, 8);
 449         /* Save fp */
 450         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, framesize, 8);
 451         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), AMD64_R11, 8);
 452         /* Save sp */
 453         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, 8);
 454         amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
 455         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, 8);
 456         /* Save method */
 457         if (tramp_type == MONO_TRAMPOLINE_JIT || tramp_type == MONO_TRAMPOLINE_JUMP) {
 458                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, arg_offset, 8);
 459                 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, 8);
 460         } else {
 461                 amd64_mov_membase_imm (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, 8);
 462         }
 463         /* Save callee saved regs */
 464 #ifdef TARGET_WIN32
 465         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), AMD64_RDI, 8);
 466         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), AMD64_RSI, 8);
 467 #endif
 468         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, 8);
 469         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, 8);
 470         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), AMD64_R13, 8);
 471         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, 8);
 472         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, 8);
 473
 474         if (aot) {
 475                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
 476                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 477         } else {
 478                 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
 479         }
 480         amd64_call_reg (code, AMD64_R11);
 481
 482         /* Save lmf_addr */
 483         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, 8);
 484         /* Save previous_lmf */
 485         /* Set the lowest bit to 1 to signal that this LMF has the ip field set */
 486         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, 8);
 487         amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 1, 8);
 488         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, 8);
 489         /* Set new lmf */
 490         amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
 491         amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, 8);
 492
 493         /* Save LMF end */
 494
 495         /* Arg1 is the pointer to the saved registers */
 496         amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
 497
 498         /* Arg2 is the address of the calling code */
 499         if (has_caller)
 500                 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, 8);
 501         else
 502                 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
 503
 504         /* Arg3 is the method/vtable ptr */
 505         amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, 8);
 506
 507         /* Arg4 is the trampoline address */
 508         amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, 8);
 509
 510         if (aot) {
 511                 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
 512                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
 513                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RIP, 0, 8);
 514         } else {
 515                 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
 516                 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
 517         }
 518         amd64_call_reg (code, AMD64_RAX);
 519
 520         /* Check for thread interruption */
 521         /* This is not perf critical code so no need to check the interrupt flag */
 522         /*
 523          * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
 524          */
 525         amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, 8);
 526         if (aot) {
 527                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
 528                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RIP, 0, 8);
 529         } else {
 530                 amd64_mov_reg_imm (code, AMD64_RAX, (guint8*)mono_thread_force_interruption_checkpoint);
 531         }
 532         amd64_call_reg (code, AMD64_RAX);
 533         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, 8);
 534
 535         /* Restore LMF */
 536
 537         amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 8);
 538         amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 1, 8);
 539         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 8);
 540         amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, 8);
 541
 542         /*
 543          * Save rax to the stack, after the leave instruction, this will become part of
 544          * the red zone.
 545          */
 546         amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, 8);
 547
 548         /* Restore argument registers, r10 (needed to pass rgctx to
 549            static shared generic methods), r11 (imt register for
 550            interface calls), and rax (needed for direct calls to C vararg functions). */
 551         for (i = 0; i < AMD64_NREG; ++i)
 552                 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_R11 || i == AMD64_RAX)
 553                         amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * 8), 8);
 554
 555         for (i = 0; i < 8; ++i)
 556                 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * 8));
 557
 558         /* Restore stack */
 559         amd64_leave (code);
 560
 561         if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
 562                 /* Load result */
 563                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - 0x8, 8);
 564                 amd64_ret (code);
 565         } else {
 566                 /* call the compiled method using the saved rax */
 567                 amd64_jump_membase (code, AMD64_RSP, rax_offset - 0x8);
 568         }
 569
 570         g_assert ((code - buf) <= 538);
 571
 572         mono_arch_flush_icache (buf, code - buf);
 573
 574         *code_size = code - buf;
 575
 576         if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
 577                 guint32 code_len;
 578
 579                 /* Initialize the nullified class init trampoline used in the AOT case */
 580                 nullified_class_init_trampoline = mono_arch_get_nullified_class_init_trampoline (&code_len);
 581         }
 582
 583         *out_unwind_ops = unwind_ops;
 584
 585         return buf;
 586 }
 587
 588 gpointer
 589 mono_arch_get_nullified_class_init_trampoline (guint32 *code_len)
 590 {
 591         guint8 *code, *buf;
 592
 593         code = buf = mono_global_codeman_reserve (16);
 594         amd64_ret (code);
 595
 596         mono_arch_flush_icache (buf, code - buf);
 597
 598         *code_len = code - buf;
 599
 600         return buf;
 601 }
 602
 603 gpointer
 604 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
 605 {
 606         guint8 *code, *buf, *tramp;
 607         int size;
 608
 609         tramp = mono_get_trampoline_code (tramp_type);
 610
 611         if ((((guint64)arg1) >> 32) == 0)
 612                 size = 5 + 1 + 4;
 613         else
 614                 size = 5 + 1 + 8;
 615
 616         code = buf = mono_domain_code_reserve_align (domain, size, 1);
 617
 618         amd64_call_code (code, tramp);
 619         /* The trampoline code will obtain the argument from the instruction stream */
 620         if ((((guint64)arg1) >> 32) == 0) {
 621                 *code = 0x4;
 622                 *(guint32*)(code + 1) = (gint64)arg1;
 623                 code += 5;
 624         } else {
 625                 *code = 0x8;
 626                 *(guint64*)(code + 1) = (gint64)arg1;
 627                 code += 9;
 628         }
 629
 630         g_assert ((code - buf) <= size);
 631
 632         if (code_len)
 633                 *code_len = size;
 634
 635         mono_arch_flush_icache (buf, size);
 636
 637         return buf;
 638 }
 639
 640 gpointer
 641 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot)
 642 {
 643         guint32 code_size;
 644         MonoJumpInfo *ji;
 645
 646         return mono_arch_create_rgctx_lazy_fetch_trampoline_full (slot, &code_size, &ji, FALSE);
 647 }
 648
 649 gpointer
 650 mono_arch_create_rgctx_lazy_fetch_trampoline_full (guint32 slot, guint32 *code_size, MonoJumpInfo **ji, gboolean aot)
 651 {
 652         guint8 *tramp;
 653         guint8 *code, *buf;
 654         guint8 **rgctx_null_jumps;
 655         int tramp_size;
 656         int depth, index;
 657         int i;
 658         gboolean mrgctx;
 659         GSList *unwind_ops = NULL;
 660         char *name;
 661
 662         *ji = NULL;
 663
 664         mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
 665         index = MONO_RGCTX_SLOT_INDEX (slot);
 666         if (mrgctx)
 667                 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
 668         for (depth = 0; ; ++depth) {
 669                 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
 670
 671                 if (index < size - 1)
 672                         break;
 673                 index -= size - 1;
 674         }
 675
 676         tramp_size = 64 + 8 * depth;
 677
 678         code = buf = mono_global_codeman_reserve (tramp_size);
 679
 680         unwind_ops = mono_arch_get_cie_program ();
 681
 682         rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
 683
 684         if (mrgctx) {
 685                 /* get mrgctx ptr */
 686                 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
 687         } else {
 688                 /* load rgctx ptr from vtable */
 689                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), 8);
 690                 /* is the rgctx ptr null? */
 691                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 692                 /* if yes, jump to actual trampoline */
 693                 rgctx_null_jumps [0] = code;
 694                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 695         }
 696
 697         for (i = 0; i < depth; ++i) {
 698                 /* load ptr to next array */
 699                 if (mrgctx && i == 0)
 700                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, 8);
 701                 else
 702                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, 8);
 703                 /* is the ptr null? */
 704                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 705                 /* if yes, jump to actual trampoline */
 706                 rgctx_null_jumps [i + 1] = code;
 707                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 708         }
 709
 710         /* fetch slot */
 711         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), 8);
 712         /* is the slot null? */
 713         amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 714         /* if yes, jump to actual trampoline */
 715         rgctx_null_jumps [depth + 1] = code;
 716         amd64_branch8 (code, X86_CC_Z, -1, 1);
 717         /* otherwise return */
 718         amd64_ret (code);
 719
 720         for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
 721                 x86_patch (rgctx_null_jumps [i], code);
 722
 723         g_free (rgctx_null_jumps);
 724
 725         /* move the rgctx pointer to the VTABLE register */
 726         amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, 8);
 727
 728         if (aot) {
 729                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
 730                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 731                 amd64_jump_reg (code, AMD64_R11);
 732         } else {
 733                 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
 734
 735                 /* jump to the actual trampoline */
 736                 amd64_jump_code (code, tramp);
 737         }
 738
 739         mono_arch_flush_icache (buf, code - buf);
 740
 741         g_assert (code - buf <= tramp_size);
 742
 743         *code_size = code - buf;
 744
 745         name = g_strdup_printf ("rgctx_fetch_trampoline_%s_%d", mrgctx ? "mrgctx" : "rgctx", index);
 746         mono_save_trampoline_xdebug_info (name, buf, code - buf, unwind_ops);
 747         g_free (name);
 748
 749         return buf;
 750 }
 751
 752 gpointer
 753 mono_arch_create_generic_class_init_trampoline (void)
 754 {
 755         guint32 code_size;
 756         MonoJumpInfo *ji;
 757
 758         return mono_arch_create_generic_class_init_trampoline_full (&code_size, &ji, FALSE);
 759 }
 760
 761 gpointer
 762 mono_arch_create_generic_class_init_trampoline_full (guint32 *code_size, MonoJumpInfo **ji, gboolean aot)
 763 {
 764         guint8 *tramp;
 765         guint8 *code, *buf;
 766         static int byte_offset = -1;
 767         static guint8 bitmask;
 768         guint8 *jump;
 769         int tramp_size;
 770
 771         *ji = NULL;
 772
 773         tramp_size = 64;
 774
 775         code = buf = mono_global_codeman_reserve (tramp_size);
 776
 777         if (byte_offset < 0)
 778                 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
 779
 780         amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
 781         jump = code;
 782         amd64_branch8 (code, X86_CC_Z, -1, 1);
 783
 784         amd64_ret (code);
 785
 786         x86_patch (jump, code);
 787
 788         if (aot) {
 789                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
 790                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 791                 amd64_jump_reg (code, AMD64_R11);
 792         } else {
 793                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
 794
 795                 /* jump to the actual trampoline */
 796                 amd64_jump_code (code, tramp);
 797         }
 798
 799         mono_arch_flush_icache (buf, code - buf);
 800
 801         g_assert (code - buf <= tramp_size);
 802
 803         *code_size = code - buf;
 804
 805         return buf;
 806 }
 807
 808 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
 809
 810 gpointer
 811 mono_arch_create_monitor_enter_trampoline (void)
 812 {
 813         guint32 code_size;
 814         MonoJumpInfo *ji;
 815
 816         return mono_arch_create_monitor_enter_trampoline_full (&code_size, &ji, FALSE);
 817 }
 818
 819 gpointer
 820 mono_arch_create_monitor_enter_trampoline_full (guint32 *code_size, MonoJumpInfo **ji, gboolean aot)
 821 {
 822
 823         guint8 *tramp;
 824         guint8 *code, *buf;
 825         guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid;
 826         int tramp_size;
 827         int owner_offset, nest_offset, dummy;
 828
 829         *ji = NULL;
 830
 831         g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
 832
 833         mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
 834         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
 835         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
 836         owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
 837         nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
 838
 839         tramp_size = 96;
 840
 841         code = buf = mono_global_codeman_reserve (tramp_size);
 842
 843         if (mono_thread_get_tls_offset () != -1) {
 844                 /* MonoObject* obj is in RDI */
 845                 /* is obj null? */
 846                 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
 847                 /* if yes, jump to actual trampoline */
 848                 jump_obj_null = code;
 849                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 850
 851                 /* load obj->synchronization to RCX */
 852                 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
 853                 /* is synchronization null? */
 854                 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
 855                 /* if yes, jump to actual trampoline */
 856                 jump_sync_null = code;
 857                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 858
 859                 /* load MonoInternalThread* into RDX */
 860                 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
 861                 /* load TID into RDX */
 862                 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
 863
 864                 /* is synchronization->owner null? */
 865                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, owner_offset, 0, 8);
 866                 /* if not, jump to next case */
 867                 jump_tid = code;
 868                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 869
 870                 /* if yes, try a compare-exchange with the TID */
 871                 /* zero RAX */
 872                 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
 873                 /* compare and exchange */
 874                 amd64_prefix (code, X86_LOCK_PREFIX);
 875                 amd64_cmpxchg_membase_reg_size (code, AMD64_RCX, owner_offset, AMD64_RDX, 8);
 876                 /* if not successful, jump to actual trampoline */
 877                 jump_cmpxchg_failed = code;
 878                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 879                 /* if successful, return */
 880                 amd64_ret (code);
 881
 882                 /* next case: synchronization->owner is not null */
 883                 x86_patch (jump_tid, code);
 884                 /* is synchronization->owner == TID? */
 885                 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
 886                 /* if not, jump to actual trampoline */
 887                 jump_other_owner = code;
 888                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 889                 /* if yes, increment nest */
 890                 amd64_inc_membase_size (code, AMD64_RCX, nest_offset, 4);
 891                 /* return */
 892                 amd64_ret (code);
 893
 894                 x86_patch (jump_obj_null, code);
 895                 x86_patch (jump_sync_null, code);
 896                 x86_patch (jump_cmpxchg_failed, code);
 897                 x86_patch (jump_other_owner, code);
 898         }
 899
 900         /* jump to the actual trampoline */
 901 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
 902         amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
 903 #endif
 904
 905         if (aot) {
 906                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
 907                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 908                 amd64_jump_reg (code, AMD64_R11);
 909         } else {
 910                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
 911
 912                 /* jump to the actual trampoline */
 913                 amd64_jump_code (code, tramp);
 914         }
 915
 916         mono_arch_flush_icache (code, code - buf);
 917         g_assert (code - buf <= tramp_size);
 918
 919         *code_size = code - buf;
 920
 921         return buf;
 922 }
 923
 924 gpointer
 925 mono_arch_create_monitor_exit_trampoline (void)
 926 {
 927         guint32 code_size;
 928         MonoJumpInfo *ji;
 929
 930         return mono_arch_create_monitor_exit_trampoline_full (&code_size, &ji, FALSE);
 931 }
 932
 933 gpointer
 934 mono_arch_create_monitor_exit_trampoline_full (guint32 *code_size, MonoJumpInfo **ji, gboolean aot)
 935 {
 936         guint8 *tramp;
 937         guint8 *code, *buf;
 938         guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned;
 939         guint8 *jump_next;
 940         int tramp_size;
 941         int owner_offset, nest_offset, entry_count_offset;
 942
 943         *ji = NULL;
 944
 945         g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
 946
 947         mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
 948         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
 949         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
 950         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
 951         owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
 952         nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
 953         entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
 954
 955         tramp_size = 94;
 956
 957         code = buf = mono_global_codeman_reserve (tramp_size);
 958
 959         if (mono_thread_get_tls_offset () != -1) {
 960                 /* MonoObject* obj is in RDI */
 961                 /* is obj null? */
 962                 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
 963                 /* if yes, jump to actual trampoline */
 964                 jump_obj_null = code;
 965                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 966
 967                 /* load obj->synchronization to RCX */
 968                 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
 969                 /* is synchronization null? */
 970                 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
 971                 /* if yes, jump to actual trampoline */
 972                 jump_sync_null = code;
 973                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 974
 975                 /* next case: synchronization is not null */
 976                 /* load MonoInternalThread* into RDX */
 977                 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
 978                 /* load TID into RDX */
 979                 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
 980                 /* is synchronization->owner == TID */
 981                 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
 982                 /* if no, jump to actual trampoline */
 983                 jump_not_owned = code;
 984                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 985
 986                 /* next case: synchronization->owner == TID */
 987                 /* is synchronization->nest == 1 */
 988                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, nest_offset, 1, 4);
 989                 /* if not, jump to next case */
 990                 jump_next = code;
 991                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 992                 /* if yes, is synchronization->entry_count zero? */
 993                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, entry_count_offset, 0, 4);
 994                 /* if not, jump to actual trampoline */
 995                 jump_have_waiters = code;
 996                 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
 997                 /* if yes, set synchronization->owner to null and return */
 998                 amd64_mov_membase_imm (code, AMD64_RCX, owner_offset, 0, 8);
 999                 amd64_ret (code);
1000
1001                 /* next case: synchronization->nest is not 1 */
1002                 x86_patch (jump_next, code);
1003                 /* decrease synchronization->nest and return */
1004                 amd64_dec_membase_size (code, AMD64_RCX, nest_offset, 4);
1005                 amd64_ret (code);
1006
1007                 x86_patch (jump_obj_null, code);
1008                 x86_patch (jump_have_waiters, code);
1009                 x86_patch (jump_not_owned, code);
1010                 x86_patch (jump_sync_null, code);
1011         }
1012
1013         /* jump to the actual trampoline */
1014 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1015         amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1016 #endif
1017
1018         if (aot) {
1019                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
1020                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
1021                 amd64_jump_reg (code, AMD64_R11);
1022         } else {
1023                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
1024                 amd64_jump_code (code, tramp);
1025         }
1026
1027         mono_arch_flush_icache (code, code - buf);
1028         g_assert (code - buf <= tramp_size);
1029
1030         *code_size = code - buf;
1031
1032         return buf;
1033 }
1034 #endif
1035
1036 void
1037 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1038 {
1039         /* FIXME: This is not thread safe */
1040         guint8 *code = ji->code_start;
1041
1042         amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
1043         amd64_mov_reg_imm (code, AMD64_R11, func);
1044
1045         x86_push_imm (code, (guint64)func_arg);
1046         amd64_call_reg (code, AMD64_R11);
1047 }