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  * Copyright 2003-2011 Novell, Inc (http://www.novell.com)
  10  * Copyright 2011 Xamarin, Inc (http://www.xamarin.com)
  11  */
  12
  13 #include <config.h>
  14 #include <glib.h>
  15
  16 #include <mono/metadata/appdomain.h>
  17 #include <mono/metadata/marshal.h>
  18 #include <mono/metadata/tabledefs.h>
  19 #include <mono/metadata/mono-debug-debugger.h>
  20 #include <mono/metadata/monitor.h>
  21 #include <mono/metadata/monitor.h>
  22 #include <mono/metadata/gc-internal.h>
  23 #include <mono/arch/amd64/amd64-codegen.h>
  24
  25 #include <mono/utils/memcheck.h>
  26
  27 #include "mini.h"
  28 #include "mini-amd64.h"
  29
  30 #if defined(__native_client_codegen__) && defined(__native_client__)
  31 #include <malloc.h>
  32 #include <sys/nacl_syscalls.h>
  33 #endif
  34
  35 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
  36
  37 static guint8* nullified_class_init_trampoline;
  38
  39 /*
  40  * mono_arch_get_unbox_trampoline:
  41  * @m: method pointer
  42  * @addr: pointer to native code for @m
  43  *
  44  * when value type methods are called through the vtable we need to unbox the
  45  * this argument. This method returns a pointer to a trampoline which does
  46  * unboxing before calling the method
  47  */
  48 gpointer
  49 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
  50 {
  51         guint8 *code, *start;
  52         int this_reg;
  53
  54         MonoDomain *domain = mono_domain_get ();
  55
  56         this_reg = mono_arch_get_this_arg_reg (NULL);
  57
  58         start = code = mono_domain_code_reserve (domain, 20);
  59
  60         amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
  61         /* FIXME: Optimize this */
  62         amd64_mov_reg_imm (code, AMD64_RAX, addr);
  63         amd64_jump_reg (code, AMD64_RAX);
  64         g_assert ((code - start) < 20);
  65
  66         nacl_domain_code_validate (domain, &start, 20, &code);
  67
  68         mono_arch_flush_icache (start, code - start);
  69
  70         return start;
  71 }
  72
  73 /*
  74  * mono_arch_get_static_rgctx_trampoline:
  75  *
  76  *   Create a trampoline which sets RGCTX_REG to MRGCTX, then jumps to ADDR.
  77  */
  78 gpointer
  79 mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr)
  80 {
  81         guint8 *code, *start;
  82         int buf_len;
  83
  84         MonoDomain *domain = mono_domain_get ();
  85
  86 #ifdef MONO_ARCH_NOMAP32BIT
  87         buf_len = 32;
  88 #else
  89         /* AOTed code could still have a non-32 bit address */
  90         if ((((guint64)addr) >> 32) == 0)
  91                 buf_len = 16;
  92         else
  93                 buf_len = 30;
  94 #endif
  95
  96         start = code = mono_domain_code_reserve (domain, buf_len);
  97
  98         amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, mrgctx);
  99         amd64_jump_code (code, addr);
 100         g_assert ((code - start) < buf_len);
 101
 102         nacl_domain_code_validate (domain, &start, buf_len, &code);
 103         mono_arch_flush_icache (start, code - start);
 104
 105         return start;
 106 }
 107
 108 gpointer
 109 mono_arch_get_llvm_imt_trampoline (MonoDomain *domain, MonoMethod *m, int vt_offset)
 110 {
 111         guint8 *code, *start;
 112         int buf_len;
 113         int this_reg;
 114
 115         buf_len = 32;
 116
 117         start = code = mono_domain_code_reserve (domain, buf_len);
 118
 119         this_reg = mono_arch_get_this_arg_reg (NULL);
 120
 121         /* Set imt arg */
 122         amd64_mov_reg_imm (code, MONO_ARCH_IMT_REG, m);
 123         /* Load vtable address */
 124         amd64_mov_reg_membase (code, AMD64_RAX, this_reg, 0, 8);
 125         amd64_jump_membase (code, AMD64_RAX, vt_offset);
 126         amd64_ret (code);
 127
 128         g_assert ((code - start) < buf_len);
 129
 130         nacl_domain_code_validate (domain, &start, buf_len, &code);
 131
 132         mono_arch_flush_icache (start, code - start);
 133
 134         return start;
 135 }
 136
 137 /*
 138  * mono_arch_patch_callsite:
 139  *
 140  *   Patch the callsite whose address is given by ORIG_CODE so it calls ADDR. ORIG_CODE
 141  * points to the pc right after the call.
 142  */
 143 void
 144 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
 145 {
 146 #if defined(__default_codegen__)
 147         guint8 *code;
 148         guint8 buf [16];
 149         gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
 150
 151         code = buf + 14;
 152
 153         /* mov 64-bit imm into r11 (followed by call reg?)  or direct call*/
 154         if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
 155                 if (code [-5] != 0xe8) {
 156                         if (can_write) {
 157                                 InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
 158                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
 159                         }
 160                 } else {
 161                         if ((((guint64)(addr)) >> 32) != 0) {
 162 #ifdef MONO_ARCH_NOMAP32BIT
 163                                 /* Print some diagnostics */
 164                                 MonoJitInfo *ji = mono_jit_info_table_find (mono_domain_get (), (char*)orig_code);
 165                                 if (ji)
 166                                         fprintf (stderr, "At %s, offset 0x%zx\n", mono_method_full_name (ji->method, TRUE), (guint8*)orig_code - (guint8*)ji->code_start);
 167                                 fprintf (stderr, "Addr: %p\n", addr);
 168                                 ji = mono_jit_info_table_find (mono_domain_get (), (char*)addr);
 169                                 if (ji)
 170                                         fprintf (stderr, "Callee: %s\n", mono_method_full_name (ji->method, TRUE));
 171                                 g_assert_not_reached ();
 172 #else
 173                                 /*
 174                                  * This might happen when calling AOTed code. Create a thunk.
 175                                  */
 176                                 guint8 *thunk_start, *thunk_code;
 177
 178                                 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
 179                                 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
 180                                 *(guint64*)thunk_code = (guint64)addr;
 181                                 addr = thunk_start;
 182                                 g_assert ((((guint64)(addr)) >> 32) == 0);
 183                                 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
 184 #endif
 185                         }
 186                         g_assert ((((guint64)(orig_code)) >> 32) == 0);
 187                         if (can_write) {
 188                                 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
 189                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
 190                         }
 191                 }
 192         }
 193         else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
 194                 /* call *<OFFSET>(%rip) */
 195                 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
 196                 if (can_write) {
 197                         InterlockedExchangePointer (got_entry, addr);
 198                         VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
 199                 }
 200         }
 201 #elif defined(__native_client__)
 202         /* These are essentially the same 2 cases as above, modified for NaCl*/
 203
 204         /* Target must be bundle-aligned */
 205         g_assert (((guint32)addr & kNaClAlignmentMask) == 0);
 206         /* Return target must be bundle-aligned */
 207         g_assert (((guint32)orig_code & kNaClAlignmentMask) == 0);
 208
 209         if (orig_code[-5] == 0xe8) {
 210                 /* Direct call */
 211                 int ret;
 212                 gint32 offset = (gint32)addr - (gint32)orig_code;
 213                 guint8 buf[sizeof(gint32)];
 214                 *((gint32*)(buf)) = offset;
 215                 ret = nacl_dyncode_modify (orig_code - sizeof(gint32), buf, sizeof(gint32));
 216                 g_assert (ret == 0);
 217         }
 218
 219         else if (is_nacl_call_reg_sequence (orig_code - 10) && orig_code[-16] == 0x41 && orig_code[-15] == 0xbb) {
 220                 int ret;
 221                 guint8 buf[sizeof(gint32)];
 222                 *((gint32 *)(buf)) = addr;
 223                 /* orig_code[-14] is the start of the immediate. */
 224                 ret = nacl_dyncode_modify (orig_code - 14, buf, sizeof(gint32));
 225                 g_assert (ret == 0);
 226         }
 227         else {
 228                 g_assert_not_reached ();
 229         }
 230
 231         return;
 232 #endif
 233 }
 234
 235 void
 236 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
 237 {
 238         gint32 disp;
 239         gpointer *plt_jump_table_entry;
 240
 241 #if defined(__default_codegen__)
 242         /* A PLT entry: jmp *<DISP>(%rip) */
 243         g_assert (code [0] == 0xff);
 244         g_assert (code [1] == 0x25);
 245
 246         disp = *(gint32*)(code + 2);
 247
 248         plt_jump_table_entry = (gpointer*)(code + 6 + disp);
 249 #elif defined(__native_client_codegen__)
 250         /* A PLT entry:            */
 251         /* mov <DISP>(%rip), %r11d */
 252         /* nacljmp *%r11           */
 253
 254         /* Verify the 'mov' */
 255         g_assert (code [0] == 0x45);
 256         g_assert (code [1] == 0x8b);
 257         g_assert (code [2] == 0x1d);
 258
 259         disp = *(gint32*)(code + 3);
 260
 261         /* 7 = 3 (mov opcode) + 4 (disp) */
 262         /* This needs to resolve to the target of the RIP-relative offset */
 263         plt_jump_table_entry = (gpointer*)(code + 7 + disp);
 264
 265 #endif /* __native_client_codegen__ */
 266
 267
 268         InterlockedExchangePointer (plt_jump_table_entry, addr);
 269 }
 270
 271 static gpointer
 272 get_vcall_slot (guint8 *code, mgreg_t *regs, int *displacement)
 273 {
 274         guint8 buf [10];
 275         gint32 disp;
 276         MonoJitInfo *ji = NULL;
 277
 278 #ifdef ENABLE_LLVM
 279         /* code - 9 might be before the start of the method */
 280         /* FIXME: Avoid this expensive call somehow */
 281         ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
 282 #endif
 283
 284         mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 9, buf, sizeof (buf));
 285         code = buf + 9;
 286
 287         *displacement = 0;
 288
 289         code -= 7;
 290
 291         if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
 292                 /* call OFFSET(%rip) */
 293                 g_assert_not_reached ();
 294                 *displacement = *(guint32*)(code + 3);
 295                 return (gpointer*)(code + disp + 7);
 296         } else {
 297                 g_assert_not_reached ();
 298                 return NULL;
 299         }
 300 }
 301
 302 static gpointer*
 303 get_vcall_slot_addr (guint8* code, mgreg_t *regs)
 304 {
 305         gpointer vt;
 306         int displacement;
 307         vt = get_vcall_slot (code, regs, &displacement);
 308         if (!vt)
 309                 return NULL;
 310         return (gpointer*)((char*)vt + displacement);
 311 }
 312
 313 void
 314 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
 315 {
 316         guint8 buf [16];
 317         MonoJitInfo *ji = NULL;
 318         gboolean can_write;
 319
 320         if (mono_use_llvm) {
 321                 /* code - 7 might be before the start of the method */
 322                 /* FIXME: Avoid this expensive call somehow */
 323                 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
 324         }
 325
 326         can_write = mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 7, buf, sizeof (buf));
 327
 328         if (!can_write)
 329                 return;
 330
 331         /*
 332          * A given byte sequence can match more than case here, so we have to be
 333          * really careful about the ordering of the cases. Longer sequences
 334          * come first.
 335          */
 336         if ((buf [0] == 0x41) && (buf [1] == 0xff) && (buf [2] == 0x15)) {
 337                 gpointer *vtable_slot;
 338
 339                 /* call *<OFFSET>(%rip) */
 340                 vtable_slot = get_vcall_slot_addr (code, regs);
 341                 g_assert (vtable_slot);
 342
 343                 *vtable_slot = nullified_class_init_trampoline;
 344         } else if (buf [2] == 0xe8) {
 345                 /* call <TARGET> */
 346                 //guint8 *buf = code - 2;
 347
 348                 /*
 349                  * It would be better to replace the call with nops, but that doesn't seem
 350                  * to work on SMP machines even when the whole call is inside a cache line.
 351                  * Patching the call address seems to work.
 352                  */
 353                 /*
 354                 buf [0] = 0x66;
 355                 buf [1] = 0x66;
 356                 buf [2] = 0x90;
 357                 buf [3] = 0x66;
 358                 buf [4] = 0x90;
 359                 */
 360
 361                 mono_arch_patch_callsite (code - 5, code, nullified_class_init_trampoline);
 362         } else if ((buf [5] == 0xff) && x86_modrm_mod (buf [6]) == 3 && x86_modrm_reg (buf [6]) == 2) {
 363                 /* call *<reg> */
 364                 /* Generated by the LLVM JIT or on platforms without MAP_32BIT set */
 365                 mono_arch_patch_callsite (code - 13, code, nullified_class_init_trampoline);
 366         } else if (buf [4] == 0x90 || buf [5] == 0xeb || buf [6] == 0x66) {
 367                 /* Already changed by another thread */
 368                 ;
 369         } else {
 370                 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", buf [0], buf [1], buf [2], buf [3],
 371                         buf [4], buf [5], buf [6]);
 372                 g_assert_not_reached ();
 373         }
 374 }
 375
 376 void
 377 mono_arch_nullify_plt_entry (guint8 *code, mgreg_t *regs)
 378 {
 379         if (mono_aot_only && !nullified_class_init_trampoline)
 380                 nullified_class_init_trampoline = mono_aot_get_trampoline ("nullified_class_init_trampoline");
 381
 382         mono_arch_patch_plt_entry (code, NULL, regs, nullified_class_init_trampoline);
 383 }
 384
 385 static void
 386 stack_unaligned (MonoTrampolineType tramp_type)
 387 {
 388         printf ("%d\n", tramp_type);
 389         g_assert_not_reached ();
 390 }
 391
 392 guchar*
 393 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
 394 {
 395         guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code;
 396         int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset, saved_regs_offset;
 397         int saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
 398         gboolean has_caller;
 399         GSList *unwind_ops = NULL;
 400         MonoJumpInfo *ji = NULL;
 401         const guint kMaxCodeSize = NACL_SIZE (600, 600*2);
 402
 403 #if defined(__native_client_codegen__)
 404         const guint kNaClTrampOffset = 17;
 405 #endif
 406
 407         if (tramp_type == MONO_TRAMPOLINE_JUMP || tramp_type == MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD)
 408                 has_caller = FALSE;
 409         else
 410                 has_caller = TRUE;
 411
 412         code = buf = mono_global_codeman_reserve (kMaxCodeSize);
 413
 414         framesize = kMaxCodeSize + sizeof (MonoLMF);
 415         framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
 416
 417         orig_rsp_to_rbp_offset = 0;
 418         r11_save_code = code;
 419         /* Reserve 5 bytes for the mov_membase_reg to save R11 */
 420         code += 5;
 421         after_r11_save_code = code;
 422
 423         // CFA = sp + 16 (the trampoline address is on the stack)
 424         cfa_offset = 16;
 425         mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
 426         // IP saved at CFA - 8
 427         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
 428
 429         /* Pop the return address off the stack */
 430         amd64_pop_reg (code, AMD64_R11);
 431         orig_rsp_to_rbp_offset += sizeof(mgreg_t);
 432
 433         cfa_offset -= sizeof(mgreg_t);
 434         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 435
 436         /*
 437          * Allocate a new stack frame
 438          */
 439         amd64_push_reg (code, AMD64_RBP);
 440         cfa_offset += sizeof(mgreg_t);
 441         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 442         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
 443
 444         orig_rsp_to_rbp_offset -= sizeof(mgreg_t);
 445         amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
 446         mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
 447         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
 448
 449         offset = 0;
 450         rbp_offset = - offset;
 451
 452         offset += sizeof(mgreg_t);
 453         rax_offset = - offset;
 454
 455         offset += sizeof(mgreg_t);
 456         tramp_offset = - offset;
 457
 458         offset += sizeof(gpointer);
 459         arg_offset = - offset;
 460
 461         /* Compute the trampoline address from the return address */
 462         if (aot) {
 463 #if defined(__default_codegen__)
 464                 /* 7 = length of call *<offset>(rip) */
 465                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
 466 #elif defined(__native_client_codegen__)
 467                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, kNaClTrampOffset);
 468 #endif
 469         } else {
 470                 /* 5 = length of amd64_call_membase () */
 471                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
 472         }
 473         amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, sizeof(gpointer));
 474
 475         offset += sizeof(mgreg_t);
 476         res_offset = - offset;
 477
 478         /* Save all registers */
 479
 480         offset += AMD64_NREG * sizeof(mgreg_t);
 481         saved_regs_offset = - offset;
 482         for (i = 0; i < AMD64_NREG; ++i) {
 483                 if (i == AMD64_RBP) {
 484                         /* RAX is already saved */
 485                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, sizeof(mgreg_t));
 486                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_RAX, sizeof(mgreg_t));
 487                 } else if (i != AMD64_R11) {
 488                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
 489                 } else {
 490                         /* We have to save R11 right at the start of
 491                            the trampoline code because it's used as a
 492                            scratch register */
 493                         amd64_mov_membase_reg (r11_save_code, AMD64_RSP, saved_regs_offset + orig_rsp_to_rbp_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
 494                         g_assert (r11_save_code == after_r11_save_code);
 495                 }
 496         }
 497         offset += 8 * sizeof(mgreg_t);
 498         saved_fpregs_offset = - offset;
 499         for (i = 0; i < 8; ++i)
 500                 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)), i);
 501
 502         /* Check that the stack is aligned */
 503 #if defined(__default_codegen__)
 504         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof (mgreg_t));
 505         amd64_alu_reg_imm (code, X86_AND, AMD64_R11, 15);
 506         amd64_alu_reg_imm (code, X86_CMP, AMD64_R11, 0);
 507         br [0] = code;
 508         amd64_branch_disp (code, X86_CC_Z, 0, FALSE);
 509         if (aot) {
 510                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 511                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
 512         } else {
 513                 amd64_mov_reg_imm (code, AMD64_RDI, tramp_type);
 514                 amd64_mov_reg_imm (code, AMD64_R11, stack_unaligned);
 515                 amd64_call_reg (code, AMD64_R11);
 516         }
 517         mono_amd64_patch (br [0], code);
 518         //amd64_breakpoint (code);
 519 #endif
 520
 521         if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
 522                 tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER &&
 523                 tramp_type != MONO_TRAMPOLINE_MONITOR_EXIT &&
 524                 tramp_type != MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) {
 525                 /* Obtain the trampoline argument which is encoded in the instruction stream */
 526                 if (aot) {
 527                         /* Load the GOT offset */
 528                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
 529 #if defined(__default_codegen__)
 530                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 7, 4);
 531 #elif defined(__native_client_codegen__)
 532                         /* The arg is hidden in a "push imm32" instruction, */
 533                         /* add one to skip the opcode.                      */
 534                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, kNaClTrampOffset+1, 4);
 535 #endif
 536                         /* Compute the address of the GOT slot */
 537                         amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, sizeof(gpointer));
 538                         /* Load the value */
 539                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer));
 540                 } else {
 541                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
 542 #if defined(__default_codegen__)
 543                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
 544                         amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
 545                         amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
 546                         br [0] = code;
 547                         x86_branch8 (code, X86_CC_NE, 6, FALSE);
 548                         /* 32 bit immediate */
 549                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
 550                         br [1] = code;
 551                         x86_jump8 (code, 10);
 552                         /* 64 bit immediate */
 553                         mono_amd64_patch (br [0], code);
 554                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
 555                         mono_amd64_patch (br [1], code);
 556 #elif defined(__native_client_codegen__)
 557                         /* All args are 32-bit pointers in NaCl */
 558                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
 559 #endif
 560                 }
 561                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
 562         } else {
 563                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * sizeof(mgreg_t)), sizeof(mgreg_t));
 564                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
 565         }
 566
 567         /* Save LMF begin */
 568
 569         offset += sizeof (MonoLMF);
 570         lmf_offset = - offset;
 571
 572         /* Save ip */
 573         if (has_caller)
 574                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, sizeof(gpointer));
 575         else
 576                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 577         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, sizeof(mgreg_t));
 578         /* Save fp */
 579         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, framesize, sizeof(mgreg_t));
 580         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), AMD64_R11, sizeof(mgreg_t));
 581         /* Save sp */
 582         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof(mgreg_t));
 583         amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
 584         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, sizeof(mgreg_t));
 585         /* Save method */
 586         if (tramp_type == MONO_TRAMPOLINE_JIT || tramp_type == MONO_TRAMPOLINE_JUMP) {
 587                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, arg_offset, sizeof(gpointer));
 588                 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, sizeof(gpointer));
 589         } else {
 590                 amd64_mov_membase_imm (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, sizeof(gpointer));
 591         }
 592         /* Save callee saved regs */
 593 #ifdef TARGET_WIN32
 594         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), AMD64_RDI, sizeof(mgreg_t));
 595         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), AMD64_RSI, sizeof(mgreg_t));
 596 #endif
 597         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, sizeof(mgreg_t));
 598         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, sizeof(mgreg_t));
 599         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), AMD64_R13, sizeof(mgreg_t));
 600         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, sizeof(mgreg_t));
 601         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, sizeof(mgreg_t));
 602
 603         if (aot) {
 604                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
 605         } else {
 606                 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
 607         }
 608         amd64_call_reg (code, AMD64_R11);
 609
 610         /* Save lmf_addr */
 611         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, sizeof(gpointer));
 612         /* Save previous_lmf */
 613         /* Set the lowest bit to 1 to signal that this LMF has the ip field set */
 614         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
 615         amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 1, sizeof(gpointer));
 616         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
 617         /* Set new lmf */
 618         amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
 619         amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
 620
 621         /* Save LMF end */
 622
 623         /* Arg1 is the pointer to the saved registers */
 624         amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
 625
 626         /* Arg2 is the address of the calling code */
 627         if (has_caller)
 628                 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, sizeof(gpointer));
 629         else
 630                 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
 631
 632         /* Arg3 is the method/vtable ptr */
 633         amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, sizeof(gpointer));
 634
 635         /* Arg4 is the trampoline address */
 636         amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, sizeof(gpointer));
 637
 638         if (aot) {
 639                 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
 640                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
 641         } else {
 642                 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
 643                 amd64_mov_reg_imm (code, AMD64_R11, tramp);
 644         }
 645         amd64_call_reg (code, AMD64_R11);
 646
 647         /* Check for thread interruption */
 648         /* This is not perf critical code so no need to check the interrupt flag */
 649         /*
 650          * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
 651          */
 652         amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, sizeof(mgreg_t));
 653         if (aot) {
 654                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
 655         } else {
 656                 amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint);
 657         }
 658         amd64_call_reg (code, AMD64_R11);
 659
 660         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t));
 661
 662         /* Restore LMF */
 663
 664         amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
 665         amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 1, sizeof(gpointer));
 666         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sizeof(gpointer));
 667         amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
 668
 669         /*
 670          * Save rax to the stack, after the leave instruction, this will become part of
 671          * the red zone.
 672          */
 673         amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, sizeof(mgreg_t));
 674
 675         /* Restore argument registers, r10 (imt method/rgxtx)
 676            and rax (needed for direct calls to C vararg functions). */
 677         for (i = 0; i < AMD64_NREG; ++i)
 678                 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
 679                         amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
 680
 681         for (i = 0; i < 8; ++i)
 682                 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)));
 683
 684         /* Restore stack */
 685         amd64_leave (code);
 686
 687         if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
 688                 /* Load result */
 689                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - sizeof(mgreg_t), sizeof(mgreg_t));
 690                 amd64_ret (code);
 691         } else {
 692                 /* call the compiled method using the saved rax */
 693                 amd64_jump_membase (code, AMD64_RSP, rax_offset - sizeof(mgreg_t));
 694         }
 695
 696         g_assert ((code - buf) <= kMaxCodeSize);
 697
 698         nacl_global_codeman_validate (&buf, kMaxCodeSize, &code);
 699
 700         mono_arch_flush_icache (buf, code - buf);
 701
 702         if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
 703                 /* Initialize the nullified class init trampoline used in the AOT case */
 704                 nullified_class_init_trampoline = mono_arch_get_nullified_class_init_trampoline (NULL);
 705         }
 706
 707         if (info)
 708                 *info = mono_tramp_info_create (mono_get_generic_trampoline_name (tramp_type), buf, code - buf, ji, unwind_ops);
 709
 710         return buf;
 711 }
 712
 713 gpointer
 714 mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
 715 {
 716         guint8 *code, *buf;
 717
 718         code = buf = mono_global_codeman_reserve (16);
 719         amd64_ret (code);
 720
 721         nacl_global_codeman_validate(&buf, 16, &code);
 722
 723         mono_arch_flush_icache (buf, code - buf);
 724
 725         if (info)
 726                 *info = mono_tramp_info_create (g_strdup_printf ("nullified_class_init_trampoline"), buf, code - buf, NULL, NULL);
 727
 728         if (mono_jit_map_is_enabled ())
 729                 mono_emit_jit_tramp (buf, code - buf, "nullified_class_init_trampoline");
 730
 731         return buf;
 732 }
 733
 734 gpointer
 735 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
 736 {
 737         guint8 *code, *buf, *tramp;
 738         int size;
 739         gboolean far_addr = FALSE;
 740
 741         tramp = mono_get_trampoline_code (tramp_type);
 742
 743 #if defined(__default_codegen__)
 744         if ((((guint64)arg1) >> 32) == 0)
 745                 size = 5 + 1 + 4;
 746         else
 747                 size = 5 + 1 + 8;
 748
 749         code = buf = mono_domain_code_reserve_align (domain, size, 1);
 750
 751         if (((gint64)tramp - (gint64)code) >> 31 != 0 && ((gint64)tramp - (gint64)code) >> 31 != -1) {
 752 #ifndef MONO_ARCH_NOMAP32BIT
 753                 g_assert_not_reached ();
 754 #endif
 755                 far_addr = TRUE;
 756                 size += 16;
 757                 code = buf = mono_domain_code_reserve_align (domain, size, 1);
 758         }
 759 #elif defined(__native_client_codegen__)
 760         size = 5 + 1 + 4;
 761         /* Aligning the call site below could */
 762         /* add up to kNaClAlignment-1 bytes   */
 763         size += (kNaClAlignment-1);
 764         buf = mono_domain_code_reserve_align (domain, size, kNaClAlignment);
 765         code = buf;
 766 #endif
 767
 768         if (far_addr) {
 769                 amd64_mov_reg_imm (code, AMD64_R11, tramp);
 770                 amd64_call_reg (code, AMD64_R11);
 771         } else {
 772                 amd64_call_code (code, tramp);
 773         }
 774         /* The trampoline code will obtain the argument from the instruction stream */
 775 #if defined(__default_codegen__)
 776         if ((((guint64)arg1) >> 32) == 0) {
 777                 *code = 0x4;
 778                 *(guint32*)(code + 1) = (gint64)arg1;
 779                 code += 5;
 780         } else {
 781                 *code = 0x8;
 782                 *(guint64*)(code + 1) = (gint64)arg1;
 783                 code += 9;
 784         }
 785 #elif defined(__native_client_codegen__)
 786         /* For NaCl, all tramp args are 32-bit because they're pointers */
 787         *code = 0x68; /* push imm32 */
 788         *(guint32*)(code + 1) = (gint32)arg1;
 789         code += 5;
 790 #endif
 791
 792         g_assert ((code - buf) <= size);
 793
 794         if (code_len)
 795                 *code_len = size;
 796
 797         nacl_domain_code_validate(domain, &buf, size, &code);
 798
 799         mono_arch_flush_icache (buf, size);
 800
 801         return buf;
 802 }
 803
 804 gpointer
 805 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
 806 {
 807         guint8 *tramp;
 808         guint8 *code, *buf;
 809         guint8 **rgctx_null_jumps;
 810         int tramp_size;
 811         int depth, index;
 812         int i;
 813         gboolean mrgctx;
 814         MonoJumpInfo *ji = NULL;
 815         GSList *unwind_ops = NULL;
 816
 817         mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
 818         index = MONO_RGCTX_SLOT_INDEX (slot);
 819         if (mrgctx)
 820                 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
 821         for (depth = 0; ; ++depth) {
 822                 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
 823
 824                 if (index < size - 1)
 825                         break;
 826                 index -= size - 1;
 827         }
 828
 829         tramp_size = NACL_SIZE (64 + 8 * depth, 128 + 8 * depth);
 830
 831         code = buf = mono_global_codeman_reserve (tramp_size);
 832
 833         unwind_ops = mono_arch_get_cie_program ();
 834
 835         rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
 836
 837         if (mrgctx) {
 838                 /* get mrgctx ptr */
 839                 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
 840         } else {
 841                 /* load rgctx ptr from vtable */
 842                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), sizeof(gpointer));
 843                 /* is the rgctx ptr null? */
 844                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 845                 /* if yes, jump to actual trampoline */
 846                 rgctx_null_jumps [0] = code;
 847                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 848         }
 849
 850         for (i = 0; i < depth; ++i) {
 851                 /* load ptr to next array */
 852                 if (mrgctx && i == 0)
 853                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, sizeof(gpointer));
 854                 else
 855                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, sizeof(gpointer));
 856                 /* is the ptr null? */
 857                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 858                 /* if yes, jump to actual trampoline */
 859                 rgctx_null_jumps [i + 1] = code;
 860                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 861         }
 862
 863         /* fetch slot */
 864         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), sizeof(gpointer));
 865         /* is the slot null? */
 866         amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 867         /* if yes, jump to actual trampoline */
 868         rgctx_null_jumps [depth + 1] = code;
 869         amd64_branch8 (code, X86_CC_Z, -1, 1);
 870         /* otherwise return */
 871         amd64_ret (code);
 872
 873         for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
 874                 mono_amd64_patch (rgctx_null_jumps [i], code);
 875
 876         g_free (rgctx_null_jumps);
 877
 878         /* move the rgctx pointer to the VTABLE register */
 879         amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
 880
 881         if (aot) {
 882                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
 883                 amd64_jump_reg (code, AMD64_R11);
 884         } else {
 885                 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
 886
 887                 /* jump to the actual trampoline */
 888                 amd64_jump_code (code, tramp);
 889         }
 890
 891         nacl_global_codeman_validate (&buf, tramp_size, &code);
 892         mono_arch_flush_icache (buf, code - buf);
 893
 894         g_assert (code - buf <= tramp_size);
 895
 896         if (info)
 897                 *info = mono_tramp_info_create (mono_get_rgctx_fetch_trampoline_name (slot), buf, code - buf, ji, unwind_ops);
 898
 899         return buf;
 900 }
 901
 902 gpointer
 903 mono_arch_create_generic_class_init_trampoline (MonoTrampInfo **info, gboolean aot)
 904 {
 905         guint8 *tramp;
 906         guint8 *code, *buf;
 907         static int byte_offset = -1;
 908         static guint8 bitmask;
 909         guint8 *jump;
 910         int tramp_size;
 911         GSList *unwind_ops = NULL;
 912         MonoJumpInfo *ji = NULL;
 913
 914         tramp_size = 64;
 915
 916         code = buf = mono_global_codeman_reserve (tramp_size);
 917
 918         if (byte_offset < 0)
 919                 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
 920
 921         amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
 922         jump = code;
 923         amd64_branch8 (code, X86_CC_Z, -1, 1);
 924
 925         amd64_ret (code);
 926
 927         x86_patch (jump, code);
 928
 929         if (aot) {
 930                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
 931                 amd64_jump_reg (code, AMD64_R11);
 932         } else {
 933                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
 934
 935                 /* jump to the actual trampoline */
 936                 amd64_jump_code (code, tramp);
 937         }
 938
 939         nacl_global_codeman_validate (&buf, tramp_size, &code);
 940
 941         mono_arch_flush_icache (buf, code - buf);
 942
 943         g_assert (code - buf <= tramp_size);
 944
 945         if (info)
 946                 *info = mono_tramp_info_create (g_strdup_printf ("generic_class_init_trampoline"), buf, code - buf, ji, unwind_ops);
 947
 948         return buf;
 949 }
 950
 951 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
 952
 953 gpointer
 954 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean aot)
 955 {
 956         guint8 *tramp;
 957         guint8 *code, *buf;
 958         guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid, *jump_sync_thin_hash = NULL;
 959         int tramp_size;
 960         int owner_offset, nest_offset, dummy;
 961         MonoJumpInfo *ji = NULL;
 962         GSList *unwind_ops = NULL;
 963
 964         g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
 965
 966         mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
 967         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
 968         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
 969         owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
 970         nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
 971
 972         tramp_size = 96;
 973
 974         code = buf = mono_global_codeman_reserve (tramp_size);
 975
 976         unwind_ops = mono_arch_get_cie_program ();
 977
 978         if (mono_thread_get_tls_offset () != -1) {
 979                 /* MonoObject* obj is in RDI */
 980                 /* is obj null? */
 981                 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
 982                 /* if yes, jump to actual trampoline */
 983                 jump_obj_null = code;
 984                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 985
 986                 /* load obj->synchronization to RCX */
 987                 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
 988
 989                 if (mono_gc_is_moving ()) {
 990                         /*if bit zero is set it's a thin hash*/
 991                         /*FIXME use testb encoding*/
 992                         amd64_test_reg_imm (code, AMD64_RCX, 0x01);
 993                         jump_sync_thin_hash = code;
 994                         amd64_branch8 (code, X86_CC_NE, -1, 1);
 995
 996                         /*clear bits used by the gc*/
 997                         amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
 998                 }
 999
1000                 /* is synchronization null? */
1001                 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1002                 /* if yes, jump to actual trampoline */
1003                 jump_sync_null = code;
1004                 amd64_branch8 (code, X86_CC_Z, -1, 1);
1005
1006                 /* load MonoInternalThread* into RDX */
1007                 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1008                 /* load TID into RDX */
1009                 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1010
1011                 /* is synchronization->owner null? */
1012                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, owner_offset, 0, 8);
1013                 /* if not, jump to next case */
1014                 jump_tid = code;
1015                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1016
1017                 /* if yes, try a compare-exchange with the TID */
1018                 /* zero RAX */
1019                 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
1020                 /* compare and exchange */
1021                 amd64_prefix (code, X86_LOCK_PREFIX);
1022                 amd64_cmpxchg_membase_reg_size (code, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1023                 /* if not successful, jump to actual trampoline */
1024                 jump_cmpxchg_failed = code;
1025                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1026                 /* if successful, return */
1027                 amd64_ret (code);
1028
1029                 /* next case: synchronization->owner is not null */
1030                 x86_patch (jump_tid, code);
1031                 /* is synchronization->owner == TID? */
1032                 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1033                 /* if not, jump to actual trampoline */
1034                 jump_other_owner = code;
1035                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1036                 /* if yes, increment nest */
1037                 amd64_inc_membase_size (code, AMD64_RCX, nest_offset, 4);
1038                 /* return */
1039                 amd64_ret (code);
1040
1041                 x86_patch (jump_obj_null, code);
1042                 if (jump_sync_thin_hash)
1043                         x86_patch (jump_sync_thin_hash, code);
1044                 x86_patch (jump_sync_null, code);
1045                 x86_patch (jump_cmpxchg_failed, code);
1046                 x86_patch (jump_other_owner, code);
1047         }
1048
1049         /* jump to the actual trampoline */
1050 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1051         amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1052 #endif
1053
1054         if (aot) {
1055                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
1056                 amd64_jump_reg (code, AMD64_R11);
1057         } else {
1058                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
1059
1060                 /* jump to the actual trampoline */
1061                 amd64_jump_code (code, tramp);
1062         }
1063
1064         nacl_global_codeman_validate (&buf, tramp_size, &code);
1065
1066         mono_arch_flush_icache (code, code - buf);
1067         g_assert (code - buf <= tramp_size);
1068
1069         if (info)
1070                 *info = mono_tramp_info_create (g_strdup_printf ("monitor_enter_trampoline"), buf, code - buf, ji, unwind_ops);
1071
1072         return buf;
1073 }
1074
1075 gpointer
1076 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1077 {
1078         guint8 *tramp;
1079         guint8 *code, *buf;
1080         guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned, *jump_sync_thin_hash = NULL;
1081         guint8 *jump_next;
1082         int tramp_size;
1083         int owner_offset, nest_offset, entry_count_offset;
1084         MonoJumpInfo *ji = NULL;
1085         GSList *unwind_ops = NULL;
1086
1087         g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
1088
1089         mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
1090         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
1091         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
1092         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
1093         owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
1094         nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
1095         entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
1096
1097         tramp_size = 112;
1098
1099         code = buf = mono_global_codeman_reserve (tramp_size);
1100
1101         unwind_ops = mono_arch_get_cie_program ();
1102
1103         if (mono_thread_get_tls_offset () != -1) {
1104                 /* MonoObject* obj is in RDI */
1105                 /* is obj null? */
1106                 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
1107                 /* if yes, jump to actual trampoline */
1108                 jump_obj_null = code;
1109                 amd64_branch8 (code, X86_CC_Z, -1, 1);
1110
1111                 /* load obj->synchronization to RCX */
1112                 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1113
1114                 if (mono_gc_is_moving ()) {
1115                         /*if bit zero is set it's a thin hash*/
1116                         /*FIXME use testb encoding*/
1117                         amd64_test_reg_imm (code, AMD64_RCX, 0x01);
1118                         jump_sync_thin_hash = code;
1119                         amd64_branch8 (code, X86_CC_NE, -1, 1);
1120
1121                         /*clear bits used by the gc*/
1122                         amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1123                 }
1124
1125                 /* is synchronization null? */
1126                 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1127                 /* if yes, jump to actual trampoline */
1128                 jump_sync_null = code;
1129                 amd64_branch8 (code, X86_CC_Z, -1, 1);
1130
1131                 /* next case: synchronization is not null */
1132                 /* load MonoInternalThread* into RDX */
1133                 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1134                 /* load TID into RDX */
1135                 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1136                 /* is synchronization->owner == TID */
1137                 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1138                 /* if no, jump to actual trampoline */
1139                 jump_not_owned = code;
1140                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1141
1142                 /* next case: synchronization->owner == TID */
1143                 /* is synchronization->nest == 1 */
1144                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, nest_offset, 1, 4);
1145                 /* if not, jump to next case */
1146                 jump_next = code;
1147                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1148                 /* if yes, is synchronization->entry_count zero? */
1149                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, entry_count_offset, 0, 4);
1150                 /* if not, jump to actual trampoline */
1151                 jump_have_waiters = code;
1152                 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
1153                 /* if yes, set synchronization->owner to null and return */
1154                 amd64_mov_membase_imm (code, AMD64_RCX, owner_offset, 0, 8);
1155                 amd64_ret (code);
1156
1157                 /* next case: synchronization->nest is not 1 */
1158                 x86_patch (jump_next, code);
1159                 /* decrease synchronization->nest and return */
1160                 amd64_dec_membase_size (code, AMD64_RCX, nest_offset, 4);
1161                 amd64_ret (code);
1162
1163                 x86_patch (jump_obj_null, code);
1164                 x86_patch (jump_have_waiters, code);
1165                 x86_patch (jump_not_owned, code);
1166                 x86_patch (jump_sync_null, code);
1167         }
1168
1169         /* jump to the actual trampoline */
1170 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1171         amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1172 #endif
1173
1174         if (aot) {
1175                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
1176                 amd64_jump_reg (code, AMD64_R11);
1177         } else {
1178                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
1179                 amd64_jump_code (code, tramp);
1180         }
1181
1182         nacl_global_codeman_validate (&buf, tramp_size, &code);
1183
1184         mono_arch_flush_icache (code, code - buf);
1185         g_assert (code - buf <= tramp_size);
1186
1187         if (info)
1188                 *info = mono_tramp_info_create (g_strdup_printf ("monitor_exit_trampoline"), buf, code - buf, ji, unwind_ops);
1189
1190         return buf;
1191 }
1192 #endif
1193
1194 void
1195 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1196 {
1197         /* FIXME: This is not thread safe */
1198         guint8 *code = ji->code_start;
1199
1200         amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
1201         amd64_mov_reg_imm (code, AMD64_R11, func);
1202
1203         x86_push_imm (code, (guint64)func_arg);
1204         amd64_call_reg (code, AMD64_R11);
1205 }
1206
1207
1208 static void
1209 handler_block_trampoline_helper (gpointer *ptr)
1210 {
1211         MonoJitTlsData *jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
1212         *ptr = jit_tls->handler_block_return_address;
1213 }
1214
1215 gpointer
1216 mono_arch_create_handler_block_trampoline (void)
1217 {
1218         guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
1219         guint8 *code, *buf;
1220         int tramp_size = 64;
1221         code = buf = mono_global_codeman_reserve (tramp_size);
1222
1223         /*
1224         This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
1225         */
1226
1227         if (mono_get_jit_tls_offset () != -1) {
1228                 code = mono_amd64_emit_tls_get (code, AMD64_RDI, mono_get_jit_tls_offset ());
1229                 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RDI, G_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 8);
1230                 /* Simulate a call */
1231                 amd64_push_reg (code, AMD64_RAX);
1232                 amd64_jump_code (code, tramp);
1233         } else {
1234                 /*Slow path uses a c helper*/
1235                 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
1236                 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
1237                 amd64_push_reg (code, AMD64_RAX);
1238                 amd64_jump_code (code, handler_block_trampoline_helper);
1239         }
1240
1241         mono_arch_flush_icache (buf, code - buf);
1242         g_assert (code - buf <= tramp_size);
1243
1244         if (mono_jit_map_is_enabled ())
1245                 mono_emit_jit_tramp (buf, code - buf, "handler_block_trampoline");
1246
1247         return buf;
1248 }
1249
1250 /*
1251  * mono_arch_get_call_target:
1252  *
1253  *   Return the address called by the code before CODE if exists.
1254  */
1255 guint8*
1256 mono_arch_get_call_target (guint8 *code)
1257 {
1258         if (code [-5] == 0xe8) {
1259                 guint32 disp = *(guint32*)(code - 4);
1260                 guint8 *target = code + disp;
1261
1262                 return target;
1263         } else {
1264                 return NULL;
1265         }
1266 }
1267
1268 /*
1269  * mono_arch_get_plt_info_offset:
1270  *
1271  *   Return the PLT info offset belonging to the plt entry PLT_ENTRY.
1272  */
1273 guint32
1274 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
1275 {
1276 #if defined(__native_client__) || defined(__native_client_codegen__)
1277         /* 18 = 3 (mov opcode) + 4 (disp) + 10 (nacljmp) + 1 (push opcode) */
1278         /* See aot-compiler.c arch_emit_plt_entry for details.             */
1279         return *(guint32*)(plt_entry + 18);
1280 #else
1281         return *(guint32*)(plt_entry + 6);
1282 #endif
1283 }