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                         gboolean disp_32bit = ((((gint64)addr - (gint64)orig_code)) < (1 << 30)) && ((((gint64)addr - (gint64)orig_code)) > -(1 << 30));
 162
 163                         if ((((guint64)(addr)) >> 32) != 0 && !disp_32bit) {
 164 #ifdef MONO_ARCH_NOMAP32BIT
 165                                 /* Print some diagnostics */
 166                                 MonoJitInfo *ji = mono_jit_info_table_find (mono_domain_get (), (char*)orig_code);
 167                                 if (ji)
 168                                         fprintf (stderr, "At %s, offset 0x%zx\n", mono_method_full_name (ji->method, TRUE), (guint8*)orig_code - (guint8*)ji->code_start);
 169                                 fprintf (stderr, "Addr: %p\n", addr);
 170                                 ji = mono_jit_info_table_find (mono_domain_get (), (char*)addr);
 171                                 if (ji)
 172                                         fprintf (stderr, "Callee: %s\n", mono_method_full_name (ji->method, TRUE));
 173                                 g_assert_not_reached ();
 174 #else
 175                                 /*
 176                                  * This might happen when calling AOTed code. Create a thunk.
 177                                  */
 178                                 guint8 *thunk_start, *thunk_code;
 179
 180                                 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
 181                                 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
 182                                 *(guint64*)thunk_code = (guint64)addr;
 183                                 addr = thunk_start;
 184                                 g_assert ((((guint64)(addr)) >> 32) == 0);
 185                                 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
 186 #endif
 187                         }
 188                         if (can_write) {
 189                                 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
 190                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
 191                         }
 192                 }
 193         }
 194         else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
 195                 /* call *<OFFSET>(%rip) */
 196                 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
 197                 if (can_write) {
 198                         InterlockedExchangePointer (got_entry, addr);
 199                         VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
 200                 }
 201         }
 202 #elif defined(__native_client__)
 203         /* These are essentially the same 2 cases as above, modified for NaCl*/
 204
 205         /* Target must be bundle-aligned */
 206         g_assert (((guint32)addr & kNaClAlignmentMask) == 0);
 207         /* Return target must be bundle-aligned */
 208         g_assert (((guint32)orig_code & kNaClAlignmentMask) == 0);
 209
 210         if (orig_code[-5] == 0xe8) {
 211                 /* Direct call */
 212                 int ret;
 213                 gint32 offset = (gint32)addr - (gint32)orig_code;
 214                 guint8 buf[sizeof(gint32)];
 215                 *((gint32*)(buf)) = offset;
 216                 ret = nacl_dyncode_modify (orig_code - sizeof(gint32), buf, sizeof(gint32));
 217                 g_assert (ret == 0);
 218         }
 219
 220         else if (is_nacl_call_reg_sequence (orig_code - 10) && orig_code[-16] == 0x41 && orig_code[-15] == 0xbb) {
 221                 int ret;
 222                 guint8 buf[sizeof(gint32)];
 223                 *((gint32 *)(buf)) = addr;
 224                 /* orig_code[-14] is the start of the immediate. */
 225                 ret = nacl_dyncode_modify (orig_code - 14, buf, sizeof(gint32));
 226                 g_assert (ret == 0);
 227         }
 228         else {
 229                 g_assert_not_reached ();
 230         }
 231
 232         return;
 233 #endif
 234 }
 235
 236 guint8*
 237 mono_arch_create_llvm_native_thunk (MonoDomain *domain, guint8 *addr)
 238 {
 239         /*
 240          * The caller is LLVM code and the call displacement might exceed 32 bits. We can't determine the caller address, so
 241          * we add a thunk every time.
 242          * Since the caller is also allocated using the domain code manager, hopefully the displacement will fit into 32 bits.
 243          * FIXME: Avoid this if possible if !MONO_ARCH_NOMAP32BIT and ADDR is 32 bits.
 244          */
 245         guint8 *thunk_start, *thunk_code;
 246
 247         thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
 248         amd64_jump_membase (thunk_code, AMD64_RIP, 0);
 249         *(guint64*)thunk_code = (guint64)addr;
 250         addr = thunk_start;
 251         mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
 252         return addr;
 253 }
 254
 255 void
 256 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
 257 {
 258         gint32 disp;
 259         gpointer *plt_jump_table_entry;
 260
 261 #if defined(__default_codegen__)
 262         /* A PLT entry: jmp *<DISP>(%rip) */
 263         g_assert (code [0] == 0xff);
 264         g_assert (code [1] == 0x25);
 265
 266         disp = *(gint32*)(code + 2);
 267
 268         plt_jump_table_entry = (gpointer*)(code + 6 + disp);
 269 #elif defined(__native_client_codegen__)
 270         /* A PLT entry:            */
 271         /* mov <DISP>(%rip), %r11d */
 272         /* nacljmp *%r11           */
 273
 274         /* Verify the 'mov' */
 275         g_assert (code [0] == 0x45);
 276         g_assert (code [1] == 0x8b);
 277         g_assert (code [2] == 0x1d);
 278
 279         disp = *(gint32*)(code + 3);
 280
 281         /* 7 = 3 (mov opcode) + 4 (disp) */
 282         /* This needs to resolve to the target of the RIP-relative offset */
 283         plt_jump_table_entry = (gpointer*)(code + 7 + disp);
 284
 285 #endif /* __native_client_codegen__ */
 286
 287
 288         InterlockedExchangePointer (plt_jump_table_entry, addr);
 289 }
 290
 291 static gpointer
 292 get_vcall_slot (guint8 *code, mgreg_t *regs, int *displacement)
 293 {
 294         guint8 buf [10];
 295         gint32 disp;
 296         MonoJitInfo *ji = NULL;
 297
 298 #ifdef ENABLE_LLVM
 299         /* code - 9 might be before the start of the method */
 300         /* FIXME: Avoid this expensive call somehow */
 301         ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
 302 #endif
 303
 304         mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 9, buf, sizeof (buf));
 305         code = buf + 9;
 306
 307         *displacement = 0;
 308
 309         code -= 7;
 310
 311         if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
 312                 /* call OFFSET(%rip) */
 313                 g_assert_not_reached ();
 314                 *displacement = *(guint32*)(code + 3);
 315                 return (gpointer*)(code + disp + 7);
 316         } else {
 317                 g_assert_not_reached ();
 318                 return NULL;
 319         }
 320 }
 321
 322 static gpointer*
 323 get_vcall_slot_addr (guint8* code, mgreg_t *regs)
 324 {
 325         gpointer vt;
 326         int displacement;
 327         vt = get_vcall_slot (code, regs, &displacement);
 328         if (!vt)
 329                 return NULL;
 330         return (gpointer*)((char*)vt + displacement);
 331 }
 332
 333 void
 334 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
 335 {
 336         guint8 buf [16];
 337         MonoJitInfo *ji = NULL;
 338         gboolean can_write;
 339
 340         if (mono_use_llvm) {
 341                 /* code - 7 might be before the start of the method */
 342                 /* FIXME: Avoid this expensive call somehow */
 343                 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
 344         }
 345
 346         can_write = mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 7, buf, sizeof (buf));
 347
 348         if (!can_write)
 349                 return;
 350
 351         /*
 352          * A given byte sequence can match more than case here, so we have to be
 353          * really careful about the ordering of the cases. Longer sequences
 354          * come first.
 355          */
 356         if ((buf [0] == 0x41) && (buf [1] == 0xff) && (buf [2] == 0x15)) {
 357                 gpointer *vtable_slot;
 358
 359                 /* call *<OFFSET>(%rip) */
 360                 vtable_slot = get_vcall_slot_addr (code, regs);
 361                 g_assert (vtable_slot);
 362
 363                 *vtable_slot = nullified_class_init_trampoline;
 364         } else if (buf [2] == 0xe8) {
 365                 /* call <TARGET> */
 366                 //guint8 *buf = code - 2;
 367
 368                 /*
 369                  * It would be better to replace the call with nops, but that doesn't seem
 370                  * to work on SMP machines even when the whole call is inside a cache line.
 371                  * Patching the call address seems to work.
 372                  */
 373                 /*
 374                 buf [0] = 0x66;
 375                 buf [1] = 0x66;
 376                 buf [2] = 0x90;
 377                 buf [3] = 0x66;
 378                 buf [4] = 0x90;
 379                 */
 380
 381                 mono_arch_patch_callsite (code - 5, code, nullified_class_init_trampoline);
 382         } else if ((buf [5] == 0xff) && x86_modrm_mod (buf [6]) == 3 && x86_modrm_reg (buf [6]) == 2) {
 383                 /* call *<reg> */
 384                 /* Generated by the LLVM JIT or on platforms without MAP_32BIT set */
 385                 mono_arch_patch_callsite (code - 13, code, nullified_class_init_trampoline);
 386         } else if (buf [4] == 0x90 || buf [5] == 0xeb || buf [6] == 0x66) {
 387                 /* Already changed by another thread */
 388                 ;
 389         } else {
 390                 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", buf [0], buf [1], buf [2], buf [3],
 391                         buf [4], buf [5], buf [6]);
 392                 g_assert_not_reached ();
 393         }
 394 }
 395
 396 void
 397 mono_arch_nullify_plt_entry (guint8 *code, mgreg_t *regs)
 398 {
 399         if (mono_aot_only && !nullified_class_init_trampoline)
 400                 nullified_class_init_trampoline = mono_aot_get_trampoline ("nullified_class_init_trampoline");
 401
 402         mono_arch_patch_plt_entry (code, NULL, regs, nullified_class_init_trampoline);
 403 }
 404
 405 static void
 406 stack_unaligned (MonoTrampolineType tramp_type)
 407 {
 408         printf ("%d\n", tramp_type);
 409         g_assert_not_reached ();
 410 }
 411
 412 guchar*
 413 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
 414 {
 415         guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code;
 416         int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset, saved_regs_offset;
 417         int saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
 418         gboolean has_caller;
 419         GSList *unwind_ops = NULL;
 420         MonoJumpInfo *ji = NULL;
 421         const guint kMaxCodeSize = NACL_SIZE (600, 600*2);
 422
 423 #if defined(__native_client_codegen__)
 424         const guint kNaClTrampOffset = 17;
 425 #endif
 426
 427         if (tramp_type == MONO_TRAMPOLINE_JUMP || tramp_type == MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD)
 428                 has_caller = FALSE;
 429         else
 430                 has_caller = TRUE;
 431
 432         code = buf = mono_global_codeman_reserve (kMaxCodeSize);
 433
 434         framesize = kMaxCodeSize + sizeof (MonoLMF);
 435         framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
 436
 437         orig_rsp_to_rbp_offset = 0;
 438         r11_save_code = code;
 439         /* Reserve 5 bytes for the mov_membase_reg to save R11 */
 440         code += 5;
 441         after_r11_save_code = code;
 442
 443         // CFA = sp + 16 (the trampoline address is on the stack)
 444         cfa_offset = 16;
 445         mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
 446         // IP saved at CFA - 8
 447         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
 448
 449         /* Pop the return address off the stack */
 450         amd64_pop_reg (code, AMD64_R11);
 451         orig_rsp_to_rbp_offset += sizeof(mgreg_t);
 452
 453         cfa_offset -= sizeof(mgreg_t);
 454         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 455
 456         /*
 457          * Allocate a new stack frame
 458          */
 459         amd64_push_reg (code, AMD64_RBP);
 460         cfa_offset += sizeof(mgreg_t);
 461         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 462         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
 463
 464         orig_rsp_to_rbp_offset -= sizeof(mgreg_t);
 465         amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
 466         mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
 467         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
 468
 469         offset = 0;
 470         rbp_offset = - offset;
 471
 472         offset += sizeof(mgreg_t);
 473         rax_offset = - offset;
 474
 475         offset += sizeof(mgreg_t);
 476         tramp_offset = - offset;
 477
 478         offset += sizeof(gpointer);
 479         arg_offset = - offset;
 480
 481         /* Compute the trampoline address from the return address */
 482         if (aot) {
 483 #if defined(__default_codegen__)
 484                 /* 7 = length of call *<offset>(rip) */
 485                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
 486 #elif defined(__native_client_codegen__)
 487                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, kNaClTrampOffset);
 488 #endif
 489         } else {
 490                 /* 5 = length of amd64_call_membase () */
 491                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
 492         }
 493         amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, sizeof(gpointer));
 494
 495         offset += sizeof(mgreg_t);
 496         res_offset = - offset;
 497
 498         /* Save all registers */
 499
 500         offset += AMD64_NREG * sizeof(mgreg_t);
 501         saved_regs_offset = - offset;
 502         for (i = 0; i < AMD64_NREG; ++i) {
 503                 if (i == AMD64_RBP) {
 504                         /* RAX is already saved */
 505                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, sizeof(mgreg_t));
 506                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_RAX, sizeof(mgreg_t));
 507                 } else if (i != AMD64_R11) {
 508                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
 509                 } else {
 510                         /* We have to save R11 right at the start of
 511                            the trampoline code because it's used as a
 512                            scratch register */
 513                         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));
 514                         g_assert (r11_save_code == after_r11_save_code);
 515                 }
 516         }
 517         offset += 8 * sizeof(mgreg_t);
 518         saved_fpregs_offset = - offset;
 519         for (i = 0; i < 8; ++i)
 520                 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)), i);
 521
 522         /* Check that the stack is aligned */
 523 #if defined(__default_codegen__)
 524         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof (mgreg_t));
 525         amd64_alu_reg_imm (code, X86_AND, AMD64_R11, 15);
 526         amd64_alu_reg_imm (code, X86_CMP, AMD64_R11, 0);
 527         br [0] = code;
 528         amd64_branch_disp (code, X86_CC_Z, 0, FALSE);
 529         if (aot) {
 530                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 531                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
 532         } else {
 533                 amd64_mov_reg_imm (code, AMD64_RDI, tramp_type);
 534                 amd64_mov_reg_imm (code, AMD64_R11, stack_unaligned);
 535                 amd64_call_reg (code, AMD64_R11);
 536         }
 537         mono_amd64_patch (br [0], code);
 538         //amd64_breakpoint (code);
 539 #endif
 540
 541         if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
 542                 tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER &&
 543                 tramp_type != MONO_TRAMPOLINE_MONITOR_EXIT &&
 544                 tramp_type != MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) {
 545                 /* Obtain the trampoline argument which is encoded in the instruction stream */
 546                 if (aot) {
 547                         /* Load the GOT offset */
 548                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
 549 #if defined(__default_codegen__)
 550                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 7, 4);
 551 #elif defined(__native_client_codegen__)
 552                         /* The arg is hidden in a "push imm32" instruction, */
 553                         /* add one to skip the opcode.                      */
 554                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, kNaClTrampOffset+1, 4);
 555 #endif
 556                         /* Compute the address of the GOT slot */
 557                         amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, sizeof(gpointer));
 558                         /* Load the value */
 559                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer));
 560                 } else {
 561                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
 562 #if defined(__default_codegen__)
 563                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
 564                         amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
 565                         amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
 566                         br [0] = code;
 567                         x86_branch8 (code, X86_CC_NE, 6, FALSE);
 568                         /* 32 bit immediate */
 569                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
 570                         br [1] = code;
 571                         x86_jump8 (code, 10);
 572                         /* 64 bit immediate */
 573                         mono_amd64_patch (br [0], code);
 574                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
 575                         mono_amd64_patch (br [1], code);
 576 #elif defined(__native_client_codegen__)
 577                         /* All args are 32-bit pointers in NaCl */
 578                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
 579 #endif
 580                 }
 581                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
 582         } else {
 583                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * sizeof(mgreg_t)), sizeof(mgreg_t));
 584                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
 585         }
 586
 587         /* Save LMF begin */
 588
 589         offset += sizeof (MonoLMF);
 590         lmf_offset = - offset;
 591
 592         /* Save ip */
 593         if (has_caller)
 594                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, sizeof(gpointer));
 595         else
 596                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 597         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, sizeof(mgreg_t));
 598         /* Save fp */
 599         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, framesize, sizeof(mgreg_t));
 600         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), AMD64_R11, sizeof(mgreg_t));
 601         /* Save sp */
 602         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof(mgreg_t));
 603         amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
 604         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, sizeof(mgreg_t));
 605         /* Save method */
 606         if (tramp_type == MONO_TRAMPOLINE_JIT || tramp_type == MONO_TRAMPOLINE_JUMP) {
 607                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, arg_offset, sizeof(gpointer));
 608                 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, sizeof(gpointer));
 609         } else {
 610                 amd64_mov_membase_imm (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, sizeof(gpointer));
 611         }
 612         /* Save callee saved regs */
 613 #ifdef TARGET_WIN32
 614         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), AMD64_RDI, sizeof(mgreg_t));
 615         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), AMD64_RSI, sizeof(mgreg_t));
 616 #endif
 617         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, sizeof(mgreg_t));
 618         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, sizeof(mgreg_t));
 619         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), AMD64_R13, sizeof(mgreg_t));
 620         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, sizeof(mgreg_t));
 621         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, sizeof(mgreg_t));
 622
 623         if (aot) {
 624                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
 625         } else {
 626                 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
 627         }
 628         amd64_call_reg (code, AMD64_R11);
 629
 630         /* Save lmf_addr */
 631         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, sizeof(gpointer));
 632         /* Save previous_lmf */
 633         /* Set the lowest bit to 1 to signal that this LMF has the ip field set */
 634         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
 635         amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 1, sizeof(gpointer));
 636         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
 637         /* Set new lmf */
 638         amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
 639         amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
 640
 641         /* Save LMF end */
 642
 643         /* Arg1 is the pointer to the saved registers */
 644         amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
 645
 646         /* Arg2 is the address of the calling code */
 647         if (has_caller)
 648                 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, sizeof(gpointer));
 649         else
 650                 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
 651
 652         /* Arg3 is the method/vtable ptr */
 653         amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, sizeof(gpointer));
 654
 655         /* Arg4 is the trampoline address */
 656         amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, sizeof(gpointer));
 657
 658         if (aot) {
 659                 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
 660                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
 661         } else {
 662                 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
 663                 amd64_mov_reg_imm (code, AMD64_R11, tramp);
 664         }
 665         amd64_call_reg (code, AMD64_R11);
 666
 667         /* Check for thread interruption */
 668         /* This is not perf critical code so no need to check the interrupt flag */
 669         /*
 670          * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
 671          */
 672         amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, sizeof(mgreg_t));
 673         if (aot) {
 674                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
 675         } else {
 676                 amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint);
 677         }
 678         amd64_call_reg (code, AMD64_R11);
 679
 680         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t));
 681
 682         /* Restore LMF */
 683
 684         amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
 685         amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 1, sizeof(gpointer));
 686         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sizeof(gpointer));
 687         amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
 688
 689         /*
 690          * Save rax to the stack, after the leave instruction, this will become part of
 691          * the red zone.
 692          */
 693         amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, sizeof(mgreg_t));
 694
 695         /* Restore argument registers, r10 (imt method/rgxtx)
 696            and rax (needed for direct calls to C vararg functions). */
 697         for (i = 0; i < AMD64_NREG; ++i)
 698                 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
 699                         amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
 700
 701         for (i = 0; i < 8; ++i)
 702                 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)));
 703
 704         /* Restore stack */
 705         amd64_leave (code);
 706
 707         if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
 708                 /* Load result */
 709                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - sizeof(mgreg_t), sizeof(mgreg_t));
 710                 amd64_ret (code);
 711         } else {
 712                 /* call the compiled method using the saved rax */
 713                 amd64_jump_membase (code, AMD64_RSP, rax_offset - sizeof(mgreg_t));
 714         }
 715
 716         g_assert ((code - buf) <= kMaxCodeSize);
 717
 718         nacl_global_codeman_validate (&buf, kMaxCodeSize, &code);
 719
 720         mono_arch_flush_icache (buf, code - buf);
 721
 722         if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
 723                 /* Initialize the nullified class init trampoline used in the AOT case */
 724                 nullified_class_init_trampoline = mono_arch_get_nullified_class_init_trampoline (NULL);
 725         }
 726
 727         if (info)
 728                 *info = mono_tramp_info_create (mono_get_generic_trampoline_name (tramp_type), buf, code - buf, ji, unwind_ops);
 729
 730         return buf;
 731 }
 732
 733 gpointer
 734 mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
 735 {
 736         guint8 *code, *buf;
 737
 738         code = buf = mono_global_codeman_reserve (16);
 739         amd64_ret (code);
 740
 741         nacl_global_codeman_validate(&buf, 16, &code);
 742
 743         mono_arch_flush_icache (buf, code - buf);
 744
 745         if (info)
 746                 *info = mono_tramp_info_create (g_strdup_printf ("nullified_class_init_trampoline"), buf, code - buf, NULL, NULL);
 747
 748         if (mono_jit_map_is_enabled ())
 749                 mono_emit_jit_tramp (buf, code - buf, "nullified_class_init_trampoline");
 750
 751         return buf;
 752 }
 753
 754 gpointer
 755 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
 756 {
 757         guint8 *code, *buf, *tramp;
 758         int size;
 759         gboolean far_addr = FALSE;
 760
 761         tramp = mono_get_trampoline_code (tramp_type);
 762
 763 #if defined(__default_codegen__)
 764         if ((((guint64)arg1) >> 32) == 0)
 765                 size = 5 + 1 + 4;
 766         else
 767                 size = 5 + 1 + 8;
 768
 769         code = buf = mono_domain_code_reserve_align (domain, size, 1);
 770
 771         if (((gint64)tramp - (gint64)code) >> 31 != 0 && ((gint64)tramp - (gint64)code) >> 31 != -1) {
 772 #ifndef MONO_ARCH_NOMAP32BIT
 773                 g_assert_not_reached ();
 774 #endif
 775                 far_addr = TRUE;
 776                 size += 16;
 777                 code = buf = mono_domain_code_reserve_align (domain, size, 1);
 778         }
 779 #elif defined(__native_client_codegen__)
 780         size = 5 + 1 + 4;
 781         /* Aligning the call site below could */
 782         /* add up to kNaClAlignment-1 bytes   */
 783         size += (kNaClAlignment-1);
 784         buf = mono_domain_code_reserve_align (domain, size, kNaClAlignment);
 785         code = buf;
 786 #endif
 787
 788         if (far_addr) {
 789                 amd64_mov_reg_imm (code, AMD64_R11, tramp);
 790                 amd64_call_reg (code, AMD64_R11);
 791         } else {
 792                 amd64_call_code (code, tramp);
 793         }
 794         /* The trampoline code will obtain the argument from the instruction stream */
 795 #if defined(__default_codegen__)
 796         if ((((guint64)arg1) >> 32) == 0) {
 797                 *code = 0x4;
 798                 *(guint32*)(code + 1) = (gint64)arg1;
 799                 code += 5;
 800         } else {
 801                 *code = 0x8;
 802                 *(guint64*)(code + 1) = (gint64)arg1;
 803                 code += 9;
 804         }
 805 #elif defined(__native_client_codegen__)
 806         /* For NaCl, all tramp args are 32-bit because they're pointers */
 807         *code = 0x68; /* push imm32 */
 808         *(guint32*)(code + 1) = (gint32)arg1;
 809         code += 5;
 810 #endif
 811
 812         g_assert ((code - buf) <= size);
 813
 814         if (code_len)
 815                 *code_len = size;
 816
 817         nacl_domain_code_validate(domain, &buf, size, &code);
 818
 819         mono_arch_flush_icache (buf, size);
 820
 821         return buf;
 822 }
 823
 824 gpointer
 825 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
 826 {
 827         guint8 *tramp;
 828         guint8 *code, *buf;
 829         guint8 **rgctx_null_jumps;
 830         int tramp_size;
 831         int depth, index;
 832         int i;
 833         gboolean mrgctx;
 834         MonoJumpInfo *ji = NULL;
 835         GSList *unwind_ops = NULL;
 836
 837         mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
 838         index = MONO_RGCTX_SLOT_INDEX (slot);
 839         if (mrgctx)
 840                 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
 841         for (depth = 0; ; ++depth) {
 842                 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
 843
 844                 if (index < size - 1)
 845                         break;
 846                 index -= size - 1;
 847         }
 848
 849         tramp_size = NACL_SIZE (64 + 8 * depth, 128 + 8 * depth);
 850
 851         code = buf = mono_global_codeman_reserve (tramp_size);
 852
 853         unwind_ops = mono_arch_get_cie_program ();
 854
 855         rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
 856
 857         if (mrgctx) {
 858                 /* get mrgctx ptr */
 859                 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
 860         } else {
 861                 /* load rgctx ptr from vtable */
 862                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), sizeof(gpointer));
 863                 /* is the rgctx ptr null? */
 864                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 865                 /* if yes, jump to actual trampoline */
 866                 rgctx_null_jumps [0] = code;
 867                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 868         }
 869
 870         for (i = 0; i < depth; ++i) {
 871                 /* load ptr to next array */
 872                 if (mrgctx && i == 0)
 873                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, sizeof(gpointer));
 874                 else
 875                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, sizeof(gpointer));
 876                 /* is the ptr null? */
 877                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 878                 /* if yes, jump to actual trampoline */
 879                 rgctx_null_jumps [i + 1] = code;
 880                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 881         }
 882
 883         /* fetch slot */
 884         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), sizeof(gpointer));
 885         /* is the slot null? */
 886         amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 887         /* if yes, jump to actual trampoline */
 888         rgctx_null_jumps [depth + 1] = code;
 889         amd64_branch8 (code, X86_CC_Z, -1, 1);
 890         /* otherwise return */
 891         amd64_ret (code);
 892
 893         for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
 894                 mono_amd64_patch (rgctx_null_jumps [i], code);
 895
 896         g_free (rgctx_null_jumps);
 897
 898         /* move the rgctx pointer to the VTABLE register */
 899         amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
 900
 901         if (aot) {
 902                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
 903                 amd64_jump_reg (code, AMD64_R11);
 904         } else {
 905                 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
 906
 907                 /* jump to the actual trampoline */
 908                 amd64_jump_code (code, tramp);
 909         }
 910
 911         nacl_global_codeman_validate (&buf, tramp_size, &code);
 912         mono_arch_flush_icache (buf, code - buf);
 913
 914         g_assert (code - buf <= tramp_size);
 915
 916         if (info)
 917                 *info = mono_tramp_info_create (mono_get_rgctx_fetch_trampoline_name (slot), buf, code - buf, ji, unwind_ops);
 918
 919         return buf;
 920 }
 921
 922 gpointer
 923 mono_arch_create_generic_class_init_trampoline (MonoTrampInfo **info, gboolean aot)
 924 {
 925         guint8 *tramp;
 926         guint8 *code, *buf;
 927         static int byte_offset = -1;
 928         static guint8 bitmask;
 929         guint8 *jump;
 930         int tramp_size;
 931         GSList *unwind_ops = NULL;
 932         MonoJumpInfo *ji = NULL;
 933
 934         tramp_size = 64;
 935
 936         code = buf = mono_global_codeman_reserve (tramp_size);
 937
 938         if (byte_offset < 0)
 939                 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
 940
 941         amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
 942         jump = code;
 943         amd64_branch8 (code, X86_CC_Z, -1, 1);
 944
 945         amd64_ret (code);
 946
 947         x86_patch (jump, code);
 948
 949         if (aot) {
 950                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
 951                 amd64_jump_reg (code, AMD64_R11);
 952         } else {
 953                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
 954
 955                 /* jump to the actual trampoline */
 956                 amd64_jump_code (code, tramp);
 957         }
 958
 959         nacl_global_codeman_validate (&buf, tramp_size, &code);
 960
 961         mono_arch_flush_icache (buf, code - buf);
 962
 963         g_assert (code - buf <= tramp_size);
 964
 965         if (info)
 966                 *info = mono_tramp_info_create (g_strdup_printf ("generic_class_init_trampoline"), buf, code - buf, ji, unwind_ops);
 967
 968         return buf;
 969 }
 970
 971 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
 972
 973 gpointer
 974 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean aot)
 975 {
 976         guint8 *tramp;
 977         guint8 *code, *buf;
 978         guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid, *jump_sync_thin_hash = NULL;
 979         int tramp_size;
 980         int owner_offset, nest_offset, dummy;
 981         MonoJumpInfo *ji = NULL;
 982         GSList *unwind_ops = NULL;
 983
 984         g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
 985
 986         mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
 987         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
 988         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
 989         owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
 990         nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
 991
 992         tramp_size = 96;
 993
 994         code = buf = mono_global_codeman_reserve (tramp_size);
 995
 996         unwind_ops = mono_arch_get_cie_program ();
 997
 998         if (mono_thread_get_tls_offset () != -1) {
 999                 /* MonoObject* obj is in RDI */
1000                 /* is obj null? */
1001                 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
1002                 /* if yes, jump to actual trampoline */
1003                 jump_obj_null = code;
1004                 amd64_branch8 (code, X86_CC_Z, -1, 1);
1005
1006                 /* load obj->synchronization to RCX */
1007                 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1008
1009                 if (mono_gc_is_moving ()) {
1010                         /*if bit zero is set it's a thin hash*/
1011                         /*FIXME use testb encoding*/
1012                         amd64_test_reg_imm (code, AMD64_RCX, 0x01);
1013                         jump_sync_thin_hash = code;
1014                         amd64_branch8 (code, X86_CC_NE, -1, 1);
1015
1016                         /*clear bits used by the gc*/
1017                         amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1018                 }
1019
1020                 /* is synchronization null? */
1021                 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1022                 /* if yes, jump to actual trampoline */
1023                 jump_sync_null = code;
1024                 amd64_branch8 (code, X86_CC_Z, -1, 1);
1025
1026                 /* load MonoInternalThread* into RDX */
1027                 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1028                 /* load TID into RDX */
1029                 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1030
1031                 /* is synchronization->owner null? */
1032                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, owner_offset, 0, 8);
1033                 /* if not, jump to next case */
1034                 jump_tid = code;
1035                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1036
1037                 /* if yes, try a compare-exchange with the TID */
1038                 /* zero RAX */
1039                 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
1040                 /* compare and exchange */
1041                 amd64_prefix (code, X86_LOCK_PREFIX);
1042                 amd64_cmpxchg_membase_reg_size (code, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1043                 /* if not successful, jump to actual trampoline */
1044                 jump_cmpxchg_failed = code;
1045                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1046                 /* if successful, return */
1047                 amd64_ret (code);
1048
1049                 /* next case: synchronization->owner is not null */
1050                 x86_patch (jump_tid, code);
1051                 /* is synchronization->owner == TID? */
1052                 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1053                 /* if not, jump to actual trampoline */
1054                 jump_other_owner = code;
1055                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1056                 /* if yes, increment nest */
1057                 amd64_inc_membase_size (code, AMD64_RCX, nest_offset, 4);
1058                 /* return */
1059                 amd64_ret (code);
1060
1061                 x86_patch (jump_obj_null, code);
1062                 if (jump_sync_thin_hash)
1063                         x86_patch (jump_sync_thin_hash, code);
1064                 x86_patch (jump_sync_null, code);
1065                 x86_patch (jump_cmpxchg_failed, code);
1066                 x86_patch (jump_other_owner, code);
1067         }
1068
1069         /* jump to the actual trampoline */
1070 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1071         amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1072 #endif
1073
1074         if (aot) {
1075                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
1076                 amd64_jump_reg (code, AMD64_R11);
1077         } else {
1078                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
1079
1080                 /* jump to the actual trampoline */
1081                 amd64_jump_code (code, tramp);
1082         }
1083
1084         nacl_global_codeman_validate (&buf, tramp_size, &code);
1085
1086         mono_arch_flush_icache (code, code - buf);
1087         g_assert (code - buf <= tramp_size);
1088
1089         if (info)
1090                 *info = mono_tramp_info_create (g_strdup_printf ("monitor_enter_trampoline"), buf, code - buf, ji, unwind_ops);
1091
1092         return buf;
1093 }
1094
1095 gpointer
1096 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1097 {
1098         guint8 *tramp;
1099         guint8 *code, *buf;
1100         guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned, *jump_sync_thin_hash = NULL;
1101         guint8 *jump_next;
1102         int tramp_size;
1103         int owner_offset, nest_offset, entry_count_offset;
1104         MonoJumpInfo *ji = NULL;
1105         GSList *unwind_ops = NULL;
1106
1107         g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
1108
1109         mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
1110         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
1111         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
1112         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
1113         owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
1114         nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
1115         entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
1116
1117         tramp_size = 112;
1118
1119         code = buf = mono_global_codeman_reserve (tramp_size);
1120
1121         unwind_ops = mono_arch_get_cie_program ();
1122
1123         if (mono_thread_get_tls_offset () != -1) {
1124                 /* MonoObject* obj is in RDI */
1125                 /* is obj null? */
1126                 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
1127                 /* if yes, jump to actual trampoline */
1128                 jump_obj_null = code;
1129                 amd64_branch8 (code, X86_CC_Z, -1, 1);
1130
1131                 /* load obj->synchronization to RCX */
1132                 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1133
1134                 if (mono_gc_is_moving ()) {
1135                         /*if bit zero is set it's a thin hash*/
1136                         /*FIXME use testb encoding*/
1137                         amd64_test_reg_imm (code, AMD64_RCX, 0x01);
1138                         jump_sync_thin_hash = code;
1139                         amd64_branch8 (code, X86_CC_NE, -1, 1);
1140
1141                         /*clear bits used by the gc*/
1142                         amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1143                 }
1144
1145                 /* is synchronization null? */
1146                 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1147                 /* if yes, jump to actual trampoline */
1148                 jump_sync_null = code;
1149                 amd64_branch8 (code, X86_CC_Z, -1, 1);
1150
1151                 /* next case: synchronization is not null */
1152                 /* load MonoInternalThread* into RDX */
1153                 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1154                 /* load TID into RDX */
1155                 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1156                 /* is synchronization->owner == TID */
1157                 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1158                 /* if no, jump to actual trampoline */
1159                 jump_not_owned = code;
1160                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1161
1162                 /* next case: synchronization->owner == TID */
1163                 /* is synchronization->nest == 1 */
1164                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, nest_offset, 1, 4);
1165                 /* if not, jump to next case */
1166                 jump_next = code;
1167                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1168                 /* if yes, is synchronization->entry_count zero? */
1169                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, entry_count_offset, 0, 4);
1170                 /* if not, jump to actual trampoline */
1171                 jump_have_waiters = code;
1172                 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
1173                 /* if yes, set synchronization->owner to null and return */
1174                 amd64_mov_membase_imm (code, AMD64_RCX, owner_offset, 0, 8);
1175                 amd64_ret (code);
1176
1177                 /* next case: synchronization->nest is not 1 */
1178                 x86_patch (jump_next, code);
1179                 /* decrease synchronization->nest and return */
1180                 amd64_dec_membase_size (code, AMD64_RCX, nest_offset, 4);
1181                 amd64_ret (code);
1182
1183                 x86_patch (jump_obj_null, code);
1184                 x86_patch (jump_have_waiters, code);
1185                 x86_patch (jump_not_owned, code);
1186                 x86_patch (jump_sync_null, code);
1187         }
1188
1189         /* jump to the actual trampoline */
1190 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1191         amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1192 #endif
1193
1194         if (aot) {
1195                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
1196                 amd64_jump_reg (code, AMD64_R11);
1197         } else {
1198                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
1199                 amd64_jump_code (code, tramp);
1200         }
1201
1202         nacl_global_codeman_validate (&buf, tramp_size, &code);
1203
1204         mono_arch_flush_icache (code, code - buf);
1205         g_assert (code - buf <= tramp_size);
1206
1207         if (info)
1208                 *info = mono_tramp_info_create (g_strdup_printf ("monitor_exit_trampoline"), buf, code - buf, ji, unwind_ops);
1209
1210         return buf;
1211 }
1212 #endif
1213
1214 void
1215 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1216 {
1217         /* FIXME: This is not thread safe */
1218         guint8 *code = ji->code_start;
1219
1220         amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
1221         amd64_mov_reg_imm (code, AMD64_R11, func);
1222
1223         x86_push_imm (code, (guint64)func_arg);
1224         amd64_call_reg (code, AMD64_R11);
1225 }
1226
1227
1228 static void
1229 handler_block_trampoline_helper (gpointer *ptr)
1230 {
1231         MonoJitTlsData *jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
1232         *ptr = jit_tls->handler_block_return_address;
1233 }
1234
1235 gpointer
1236 mono_arch_create_handler_block_trampoline (void)
1237 {
1238         guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
1239         guint8 *code, *buf;
1240         int tramp_size = 64;
1241         code = buf = mono_global_codeman_reserve (tramp_size);
1242
1243         /*
1244         This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
1245         */
1246
1247         if (mono_get_jit_tls_offset () != -1) {
1248                 code = mono_amd64_emit_tls_get (code, AMD64_RDI, mono_get_jit_tls_offset ());
1249                 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RDI, G_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 8);
1250                 /* Simulate a call */
1251                 amd64_push_reg (code, AMD64_RAX);
1252                 amd64_jump_code (code, tramp);
1253         } else {
1254                 /*Slow path uses a c helper*/
1255                 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
1256                 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
1257                 amd64_push_reg (code, AMD64_RAX);
1258                 amd64_jump_code (code, handler_block_trampoline_helper);
1259         }
1260
1261         mono_arch_flush_icache (buf, code - buf);
1262         g_assert (code - buf <= tramp_size);
1263
1264         if (mono_jit_map_is_enabled ())
1265                 mono_emit_jit_tramp (buf, code - buf, "handler_block_trampoline");
1266
1267         return buf;
1268 }
1269
1270 /*
1271  * mono_arch_get_call_target:
1272  *
1273  *   Return the address called by the code before CODE if exists.
1274  */
1275 guint8*
1276 mono_arch_get_call_target (guint8 *code)
1277 {
1278         if (code [-5] == 0xe8) {
1279                 guint32 disp = *(guint32*)(code - 4);
1280                 guint8 *target = code + disp;
1281
1282                 return target;
1283         } else {
1284                 return NULL;
1285         }
1286 }
1287
1288 /*
1289  * mono_arch_get_plt_info_offset:
1290  *
1291  *   Return the PLT info offset belonging to the plt entry PLT_ENTRY.
1292  */
1293 guint32
1294 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
1295 {
1296 #if defined(__native_client__) || defined(__native_client_codegen__)
1297         /* 18 = 3 (mov opcode) + 4 (disp) + 10 (nacljmp) + 1 (push opcode) */
1298         /* See aot-compiler.c arch_emit_plt_entry for details.             */
1299         return *(guint32*)(plt_entry + 18);
1300 #else
1301         return *(guint32*)(plt_entry + 6);
1302 #endif
1303 }