mono/mini/tramp-amd64.c

   1 /*
   2  * tramp-amd64.c: JIT trampoline code for amd64
   3  *
   4  * Authors:
   5  *   Dietmar Maurer (dietmar@ximian.com)
   6  *   Zoltan Varga (vargaz@gmail.com)
   7  *
   8  * (C) 2001 Ximian, Inc.
   9  */
  10
  11 #include <config.h>
  12 #include <glib.h>
  13
  14 #include <mono/metadata/appdomain.h>
  15 #include <mono/metadata/marshal.h>
  16 #include <mono/metadata/tabledefs.h>
  17 #include <mono/metadata/mono-debug-debugger.h>
  18 #include <mono/metadata/monitor.h>
  19 #include <mono/arch/amd64/amd64-codegen.h>
  20
  21 #ifdef HAVE_VALGRIND_MEMCHECK_H
  22 #include <valgrind/memcheck.h>
  23 #endif
  24
  25 #include "mini.h"
  26 #include "mini-amd64.h"
  27
  28 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
  29
  30 static guint8* nullified_class_init_trampoline;
  31
  32 /*
  33  * mono_arch_get_unbox_trampoline:
  34  * @gsctx: the generic sharing context
  35  * @m: method pointer
  36  * @addr: pointer to native code for @m
  37  *
  38  * when value type methods are called through the vtable we need to unbox the
  39  * this argument. This method returns a pointer to a trampoline which does
  40  * unboxing before calling the method
  41  */
  42 gpointer
  43 mono_arch_get_unbox_trampoline (MonoGenericSharingContext *gsctx, MonoMethod *m, gpointer addr)
  44 {
  45         guint8 *code, *start;
  46         int this_reg;
  47
  48         MonoDomain *domain = mono_domain_get ();
  49
  50         this_reg = mono_arch_get_this_arg_reg (mono_method_signature (m), gsctx, NULL);
  51
  52         start = code = mono_domain_code_reserve (domain, 20);
  53
  54         amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
  55         /* FIXME: Optimize this */
  56         amd64_mov_reg_imm (code, AMD64_RAX, addr);
  57         amd64_jump_reg (code, AMD64_RAX);
  58         g_assert ((code - start) < 20);
  59
  60         mono_arch_flush_icache (start, code - start);
  61
  62         return start;
  63 }
  64
  65 /*
  66  * mono_arch_patch_callsite:
  67  *
  68  *   Patch the callsite whose address is given by ORIG_CODE so it calls ADDR. ORIG_CODE
  69  * points to the pc right after the call.
  70  */
  71 void
  72 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
  73 {
  74         guint8 *code;
  75         guint8 buf [16];
  76         gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
  77
  78         code = buf + 14;
  79
  80         if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
  81                 if (code [-5] != 0xe8) {
  82                         if (can_write) {
  83                                 InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
  84 #ifdef HAVE_VALGRIND_MEMCHECK_H
  85                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
  86 #endif
  87                         }
  88                 } else {
  89                         if ((((guint64)(addr)) >> 32) != 0) {
  90                                 /* Print some diagnostics */
  91                                 MonoJitInfo *ji = mono_jit_info_table_find (mono_domain_get (), (char*)orig_code);
  92                                 if (ji)
  93                                         fprintf (stderr, "At %s, offset 0x%zx\n", mono_method_full_name (ji->method, TRUE), (guint8*)orig_code - (guint8*)ji->code_start);
  94                                 fprintf (stderr, "Addr: %p\n", addr);
  95                                 ji = mono_jit_info_table_find (mono_domain_get (), (char*)addr);
  96                                 if (ji)
  97                                         fprintf (stderr, "Callee: %s\n", mono_method_full_name (ji->method, TRUE));
  98                                 g_assert_not_reached ();
  99                         }
 100                         g_assert ((((guint64)(orig_code)) >> 32) == 0);
 101                         if (can_write) {
 102                                 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
 103 #ifdef HAVE_VALGRIND_MEMCHECK_H
 104                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
 105 #endif
 106                         }
 107                 }
 108         }
 109         else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
 110                 /* call *<OFFSET>(%rip) */
 111                 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
 112                 if (can_write) {
 113                         InterlockedExchangePointer (got_entry, addr);
 114 #ifdef HAVE_VALGRIND_MEMCHECK_H
 115                         VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
 116 #endif
 117                 }
 118         }
 119 }
 120
 121 void
 122 mono_arch_patch_plt_entry (guint8 *code, guint8 *addr)
 123 {
 124         gint32 disp;
 125         gpointer *plt_jump_table_entry;
 126
 127         /* A PLT entry: jmp *<DISP>(%rip) */
 128         g_assert (code [0] == 0xff);
 129         g_assert (code [1] == 0x25);
 130
 131         disp = *(gint32*)(code + 2);
 132
 133         plt_jump_table_entry = (gpointer*)(code + 6 + disp);
 134
 135         InterlockedExchangePointer (plt_jump_table_entry, addr);
 136 }
 137
 138 void
 139 mono_arch_nullify_class_init_trampoline (guint8 *code, gssize *regs)
 140 {
 141         guint8 buf [16];
 142         gboolean can_write = mono_breakpoint_clean_code (NULL, code, 7, buf, sizeof (buf));
 143
 144         if (!can_write)
 145                 return;
 146
 147         code -= 3;
 148
 149         /*
 150          * A given byte sequence can match more than case here, so we have to be
 151          * really careful about the ordering of the cases. Longer sequences
 152          * come first.
 153          */
 154         if ((code [-4] == 0x41) && (code [-3] == 0xff) && (code [-2] == 0x15)) {
 155                 gpointer *vtable_slot;
 156
 157                 /* call *<OFFSET>(%rip) */
 158                 vtable_slot = mono_arch_get_vcall_slot_addr (code + 3, (gpointer*)regs);
 159                 g_assert (vtable_slot);
 160
 161                 *vtable_slot = nullified_class_init_trampoline;
 162         } else if (code [-2] == 0xe8) {
 163                 /* call <TARGET> */
 164                 //guint8 *buf = code - 2;
 165
 166                 /*
 167                  * It would be better to replace the call with nops, but that doesn't seem
 168                  * to work on SMP machines even when the whole call is inside a cache line.
 169                  * Patching the call address seems to work.
 170                  */
 171                 /*
 172                 buf [0] = 0x66;
 173                 buf [1] = 0x66;
 174                 buf [2] = 0x90;
 175                 buf [3] = 0x66;
 176                 buf [4] = 0x90;
 177                 */
 178
 179                 mono_arch_patch_callsite (code - 2, code - 2 + 5, nullified_class_init_trampoline);
 180         } else if ((code [0] == 0x41) && (code [1] == 0xff)) {
 181                 /* call <REG> */
 182                 /* happens on machines without MAP_32BIT like freebsd */
 183                 /* amd64_set_reg_template is 10 bytes long */
 184                 guint8* buf = code - 10;
 185
 186                 /* FIXME: Make this thread safe */
 187                 /* Padding code suggested by the AMD64 Opt Manual */
 188                 buf [0] = 0x66;
 189                 buf [1] = 0x66;
 190                 buf [2] = 0x66;
 191                 buf [3] = 0x90;
 192                 buf [4] = 0x66;
 193                 buf [5] = 0x66;
 194                 buf [6] = 0x66;
 195                 buf [7] = 0x90;
 196                 buf [8] = 0x66;
 197                 buf [9] = 0x66;
 198                 buf [10] = 0x90;
 199                 buf [11] = 0x66;
 200                 buf [12] = 0x90;
 201         } else if (code [0] == 0x90 || code [0] == 0xeb || code [0] == 0x66) {
 202                 /* Already changed by another thread */
 203                 ;
 204         } else {
 205                 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", code [0], code [1], code [2], code [3],
 206                         code [4], code [5], code [6]);
 207                 g_assert_not_reached ();
 208         }
 209 }
 210
 211 void
 212 mono_arch_nullify_plt_entry (guint8 *code)
 213 {
 214         if (mono_aot_only && !nullified_class_init_trampoline)
 215                 nullified_class_init_trampoline = mono_aot_get_named_code ("nullified_class_init_trampoline");
 216
 217         mono_arch_patch_plt_entry (code, nullified_class_init_trampoline);
 218 }
 219
 220 guchar*
 221 mono_arch_create_trampoline_code (MonoTrampolineType tramp_type)
 222 {
 223         MonoJumpInfo *ji;
 224         guint32 code_size;
 225         guchar *code;
 226         GSList *unwind_ops, *l;
 227
 228         code = mono_arch_create_trampoline_code_full (tramp_type, &code_size, &ji, &unwind_ops, FALSE);
 229
 230         mono_save_trampoline_xdebug_info ("<generic_trampoline>", code, code_size, unwind_ops);
 231
 232         for (l = unwind_ops; l; l = l->next)
 233                 g_free (l->data);
 234         g_slist_free (unwind_ops);
 235
 236         return code;
 237 }
 238
 239 guchar*
 240 mono_arch_create_trampoline_code_full (MonoTrampolineType tramp_type, guint32 *code_size, MonoJumpInfo **ji, GSList **out_unwind_ops, gboolean aot)
 241 {
 242         guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code;
 243         int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset, saved_regs_offset;
 244         int saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
 245         gboolean has_caller;
 246         GSList *unwind_ops = NULL;
 247
 248         if (tramp_type == MONO_TRAMPOLINE_JUMP)
 249                 has_caller = FALSE;
 250         else
 251                 has_caller = TRUE;
 252
 253         code = buf = mono_global_codeman_reserve (538);
 254
 255         *ji = NULL;
 256
 257         framesize = 538 + sizeof (MonoLMF);
 258         framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
 259
 260         orig_rsp_to_rbp_offset = 0;
 261         r11_save_code = code;
 262         /* Reserve 5 bytes for the mov_membase_reg to save R11 */
 263         code += 5;
 264         after_r11_save_code = code;
 265
 266         // CFA = sp + 16 (the trampoline address is on the stack)
 267         cfa_offset = 16;
 268         mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
 269         // IP saved at CFA - 8
 270         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
 271
 272         /* Pop the return address off the stack */
 273         amd64_pop_reg (code, AMD64_R11);
 274         orig_rsp_to_rbp_offset += 8;
 275
 276         cfa_offset -= 8;
 277         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 278
 279         /*
 280          * Allocate a new stack frame
 281          */
 282         amd64_push_reg (code, AMD64_RBP);
 283         cfa_offset += 8;
 284         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 285         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
 286
 287         orig_rsp_to_rbp_offset -= 8;
 288         amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, 8);
 289         mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
 290         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
 291
 292         offset = 0;
 293         rbp_offset = - offset;
 294
 295         offset += 8;
 296         rax_offset = - offset;
 297
 298         offset += 8;
 299         tramp_offset = - offset;
 300
 301         offset += 8;
 302         arg_offset = - offset;
 303
 304         /* Compute the trampoline address from the return address */
 305         if (aot) {
 306                 /* 7 = length of call *<offset>(rip) */
 307                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
 308         } else {
 309                 /* 5 = length of amd64_call_membase () */
 310                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
 311         }
 312         amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, 8);
 313
 314         offset += 8;
 315         res_offset = - offset;
 316
 317         /* Save all registers */
 318
 319         offset += AMD64_NREG * 8;
 320         saved_regs_offset = - offset;
 321         for (i = 0; i < AMD64_NREG; ++i) {
 322                 if (i == AMD64_RBP) {
 323                         /* RAX is already saved */
 324                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, 8);
 325                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * 8), AMD64_RAX, 8);
 326                 } else if (i != AMD64_R11) {
 327                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * 8), i, 8);
 328                 } else {
 329                         /* We have to save R11 right at the start of
 330                            the trampoline code because it's used as a
 331                            scratch register */
 332                         amd64_mov_membase_reg (r11_save_code, AMD64_RSP, saved_regs_offset + orig_rsp_to_rbp_offset + (i * 8), i, 8);
 333                         g_assert (r11_save_code == after_r11_save_code);
 334                 }
 335         }
 336         offset += 8 * 8;
 337         saved_fpregs_offset = - offset;
 338         for (i = 0; i < 8; ++i)
 339                 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * 8), i);
 340
 341         if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
 342                         tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER &&
 343                         tramp_type != MONO_TRAMPOLINE_MONITOR_EXIT) {
 344                 /* Obtain the trampoline argument which is encoded in the instruction stream */
 345                 if (aot) {
 346                         /* Load the GOT offset */
 347                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, 8);
 348                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 7, 4);
 349                         /* Compute the address of the GOT slot */
 350                         amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, 8);
 351                         /* Load the value */
 352                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
 353                 } else {
 354                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, 8);
 355                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
 356                         amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
 357                         amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
 358                         br [0] = code;
 359                         x86_branch8 (code, X86_CC_NE, 6, FALSE);
 360                         /* 32 bit immediate */
 361                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
 362                         br [1] = code;
 363                         x86_jump8 (code, 10);
 364                         /* 64 bit immediate */
 365                         mono_amd64_patch (br [0], code);
 366                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
 367                         mono_amd64_patch (br [1], code);
 368                 }
 369                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, 8);
 370         } else {
 371                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * 8), 8);
 372                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, 8);
 373         }
 374
 375         /* Save LMF begin */
 376
 377         offset += sizeof (MonoLMF);
 378         lmf_offset = - offset;
 379
 380         /* Save ip */
 381         if (has_caller)
 382                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, 8);
 383         else
 384                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 385         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, 8);
 386         /* Save fp */
 387         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, framesize, 8);
 388         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), AMD64_R11, 8);
 389         /* Save sp */
 390         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, 8);
 391         amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
 392         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, 8);
 393         /* Save method */
 394         if (tramp_type == MONO_TRAMPOLINE_JIT || tramp_type == MONO_TRAMPOLINE_JUMP) {
 395                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, arg_offset, 8);
 396                 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, 8);
 397         } else {
 398                 amd64_mov_membase_imm (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, 8);
 399         }
 400         /* Save callee saved regs */
 401 #ifdef PLATFORM_WIN32
 402         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), AMD64_RDI, 8);
 403         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), AMD64_RSI, 8);
 404 #endif
 405         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, 8);
 406         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, 8);
 407         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), AMD64_R13, 8);
 408         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, 8);
 409         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, 8);
 410
 411         if (aot) {
 412                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
 413                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 414         } else {
 415                 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
 416         }
 417         amd64_call_reg (code, AMD64_R11);
 418
 419         /* Save lmf_addr */
 420         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, 8);
 421         /* Save previous_lmf */
 422         /* Set the lowest bit to 1 to signal that this LMF has the ip field set */
 423         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, 8);
 424         amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 1, 8);
 425         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, 8);
 426         /* Set new lmf */
 427         amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
 428         amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, 8);
 429
 430         /* Save LMF end */
 431
 432         /* Arg1 is the pointer to the saved registers */
 433         amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
 434
 435         /* Arg2 is the address of the calling code */
 436         if (has_caller)
 437                 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, 8);
 438         else
 439                 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
 440
 441         /* Arg3 is the method/vtable ptr */
 442         amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, 8);
 443
 444         /* Arg4 is the trampoline address */
 445         amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, 8);
 446
 447         if (aot) {
 448                 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
 449                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
 450                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RIP, 0, 8);
 451         } else {
 452                 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
 453                 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
 454         }
 455         amd64_call_reg (code, AMD64_RAX);
 456
 457         /* Check for thread interruption */
 458         /* This is not perf critical code so no need to check the interrupt flag */
 459         /*
 460          * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
 461          */
 462         amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, 8);
 463         if (aot) {
 464                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
 465                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RIP, 0, 8);
 466         } else {
 467                 amd64_mov_reg_imm (code, AMD64_RAX, (guint8*)mono_thread_force_interruption_checkpoint);
 468         }
 469         amd64_call_reg (code, AMD64_RAX);
 470         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, 8);
 471
 472         /* Restore LMF */
 473
 474         amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 8);
 475         amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 1, 8);
 476         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 8);
 477         amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, 8);
 478
 479         /*
 480          * Save rax to the stack, after the leave instruction, this will become part of
 481          * the red zone.
 482          */
 483         amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, 8);
 484
 485         /* Restore argument registers, r10 (needed to pass rgctx to
 486            static shared generic methods), r11 (imt register for
 487            interface calls), and rax (needed for direct calls to C vararg functions). */
 488         for (i = 0; i < AMD64_NREG; ++i)
 489                 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_R11 || i == AMD64_RAX)
 490                         amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * 8), 8);
 491
 492         for (i = 0; i < 8; ++i)
 493                 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * 8));
 494
 495         /* Restore stack */
 496         amd64_leave (code);
 497
 498         if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
 499                 /* Load result */
 500                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - 0x8, 8);
 501                 amd64_ret (code);
 502         } else {
 503                 /* call the compiled method using the saved rax */
 504                 amd64_jump_membase (code, AMD64_RSP, rax_offset - 0x8);
 505         }
 506
 507         g_assert ((code - buf) <= 538);
 508
 509         mono_arch_flush_icache (buf, code - buf);
 510
 511         *code_size = code - buf;
 512
 513         if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
 514                 guint32 code_len;
 515
 516                 /* Initialize the nullified class init trampoline used in the AOT case */
 517                 nullified_class_init_trampoline = mono_arch_get_nullified_class_init_trampoline (&code_len);
 518         }
 519
 520         *out_unwind_ops = unwind_ops;
 521
 522         return buf;
 523 }
 524
 525 gpointer
 526 mono_arch_get_nullified_class_init_trampoline (guint32 *code_len)
 527 {
 528         guint8 *code, *buf;
 529
 530         code = buf = mono_global_codeman_reserve (16);
 531         amd64_ret (code);
 532
 533         mono_arch_flush_icache (buf, code - buf);
 534
 535         *code_len = code - buf;
 536
 537         return buf;
 538 }
 539
 540 gpointer
 541 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
 542 {
 543         guint8 *code, *buf, *tramp;
 544         int size;
 545
 546         tramp = mono_get_trampoline_code (tramp_type);
 547
 548         if ((((guint64)arg1) >> 32) == 0)
 549                 size = 5 + 1 + 4;
 550         else
 551                 size = 5 + 1 + 8;
 552
 553         code = buf = mono_domain_code_reserve_align (domain, size, 1);
 554
 555         amd64_call_code (code, tramp);
 556         /* The trampoline code will obtain the argument from the instruction stream */
 557         if ((((guint64)arg1) >> 32) == 0) {
 558                 *code = 0x4;
 559                 *(guint32*)(code + 1) = (gint64)arg1;
 560                 code += 5;
 561         } else {
 562                 *code = 0x8;
 563                 *(guint64*)(code + 1) = (gint64)arg1;
 564                 code += 9;
 565         }
 566
 567         g_assert ((code - buf) <= size);
 568
 569         if (code_len)
 570                 *code_len = size;
 571
 572         mono_arch_flush_icache (buf, size);
 573
 574         return buf;
 575 }
 576
 577 gpointer
 578 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot)
 579 {
 580         guint32 code_size;
 581         MonoJumpInfo *ji;
 582
 583         return mono_arch_create_rgctx_lazy_fetch_trampoline_full (slot, &code_size, &ji, FALSE);
 584 }
 585
 586 gpointer
 587 mono_arch_create_rgctx_lazy_fetch_trampoline_full (guint32 slot, guint32 *code_size, MonoJumpInfo **ji, gboolean aot)
 588 {
 589         guint8 *tramp;
 590         guint8 *code, *buf;
 591         guint8 **rgctx_null_jumps;
 592         int tramp_size;
 593         int depth, index;
 594         int i;
 595         gboolean mrgctx;
 596
 597         *ji = NULL;
 598
 599         mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
 600         index = MONO_RGCTX_SLOT_INDEX (slot);
 601         if (mrgctx)
 602                 index += sizeof (MonoMethodRuntimeGenericContext) / sizeof (gpointer);
 603         for (depth = 0; ; ++depth) {
 604                 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
 605
 606                 if (index < size - 1)
 607                         break;
 608                 index -= size - 1;
 609         }
 610
 611         tramp_size = 64 + 8 * depth;
 612
 613         code = buf = mono_global_codeman_reserve (tramp_size);
 614
 615         rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
 616
 617         if (mrgctx) {
 618                 /* get mrgctx ptr */
 619                 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
 620         } else {
 621                 /* load rgctx ptr from vtable */
 622                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), 8);
 623                 /* is the rgctx ptr null? */
 624                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 625                 /* if yes, jump to actual trampoline */
 626                 rgctx_null_jumps [0] = code;
 627                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 628         }
 629
 630         for (i = 0; i < depth; ++i) {
 631                 /* load ptr to next array */
 632                 if (mrgctx && i == 0)
 633                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (MonoMethodRuntimeGenericContext), 8);
 634                 else
 635                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, 8);
 636                 /* is the ptr null? */
 637                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 638                 /* if yes, jump to actual trampoline */
 639                 rgctx_null_jumps [i + 1] = code;
 640                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 641         }
 642
 643         /* fetch slot */
 644         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), 8);
 645         /* is the slot null? */
 646         amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 647         /* if yes, jump to actual trampoline */
 648         rgctx_null_jumps [depth + 1] = code;
 649         amd64_branch8 (code, X86_CC_Z, -1, 1);
 650         /* otherwise return */
 651         amd64_ret (code);
 652
 653         for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
 654                 x86_patch (rgctx_null_jumps [i], code);
 655
 656         g_free (rgctx_null_jumps);
 657
 658         /* move the rgctx pointer to the VTABLE register */
 659         amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, 8);
 660
 661         if (aot) {
 662                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
 663                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 664                 amd64_jump_reg (code, AMD64_R11);
 665         } else {
 666                 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
 667
 668                 /* jump to the actual trampoline */
 669                 amd64_jump_code (code, tramp);
 670         }
 671
 672         mono_arch_flush_icache (buf, code - buf);
 673
 674         g_assert (code - buf <= tramp_size);
 675
 676         *code_size = code - buf;
 677
 678         return buf;
 679 }
 680
 681 gpointer
 682 mono_arch_create_generic_class_init_trampoline (void)
 683 {
 684         guint32 code_size;
 685         MonoJumpInfo *ji;
 686
 687         return mono_arch_create_generic_class_init_trampoline_full (&code_size, &ji, FALSE);
 688 }
 689
 690 gpointer
 691 mono_arch_create_generic_class_init_trampoline_full (guint32 *code_size, MonoJumpInfo **ji, gboolean aot)
 692 {
 693         guint8 *tramp;
 694         guint8 *code, *buf;
 695         static int byte_offset = -1;
 696         static guint8 bitmask;
 697         guint8 *jump;
 698         int tramp_size;
 699
 700         *ji = NULL;
 701
 702         tramp_size = 64;
 703
 704         code = buf = mono_global_codeman_reserve (tramp_size);
 705
 706         if (byte_offset < 0)
 707                 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
 708
 709         amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
 710         jump = code;
 711         amd64_branch8 (code, X86_CC_Z, -1, 1);
 712
 713         amd64_ret (code);
 714
 715         x86_patch (jump, code);
 716
 717         if (aot) {
 718                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
 719                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 720                 amd64_jump_reg (code, AMD64_R11);
 721         } else {
 722                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
 723
 724                 /* jump to the actual trampoline */
 725                 amd64_jump_code (code, tramp);
 726         }
 727
 728         mono_arch_flush_icache (buf, code - buf);
 729
 730         g_assert (code - buf <= tramp_size);
 731
 732         *code_size = code - buf;
 733
 734         return buf;
 735 }
 736
 737 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
 738
 739 gpointer
 740 mono_arch_create_monitor_enter_trampoline (void)
 741 {
 742         guint32 code_size;
 743         MonoJumpInfo *ji;
 744
 745         return mono_arch_create_monitor_enter_trampoline_full (&code_size, &ji, FALSE);
 746 }
 747
 748 gpointer
 749 mono_arch_create_monitor_enter_trampoline_full (guint32 *code_size, MonoJumpInfo **ji, gboolean aot)
 750 {
 751
 752         guint8 *tramp;
 753         guint8 *code, *buf;
 754         guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid;
 755         int tramp_size;
 756         int owner_offset, nest_offset, dummy;
 757
 758         *ji = NULL;
 759
 760         g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
 761
 762         mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
 763         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
 764         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
 765         owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
 766         nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
 767
 768         tramp_size = 96;
 769
 770         code = buf = mono_global_codeman_reserve (tramp_size);
 771
 772         if (mono_thread_get_tls_offset () != -1) {
 773                 /* MonoObject* obj is in RDI */
 774                 /* is obj null? */
 775                 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
 776                 /* if yes, jump to actual trampoline */
 777                 jump_obj_null = code;
 778                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 779
 780                 /* load obj->synchronization to RCX */
 781                 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
 782                 /* is synchronization null? */
 783                 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
 784                 /* if yes, jump to actual trampoline */
 785                 jump_sync_null = code;
 786                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 787
 788                 /* load MonoThread* into RDX */
 789                 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
 790                 /* load TID into RDX */
 791                 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoThread, tid), 8);
 792
 793                 /* is synchronization->owner null? */
 794                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, owner_offset, 0, 8);
 795                 /* if not, jump to next case */
 796                 jump_tid = code;
 797                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 798
 799                 /* if yes, try a compare-exchange with the TID */
 800                 /* zero RAX */
 801                 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
 802                 /* compare and exchange */
 803                 amd64_prefix (code, X86_LOCK_PREFIX);
 804                 amd64_cmpxchg_membase_reg_size (code, AMD64_RCX, owner_offset, AMD64_RDX, 8);
 805                 /* if not successful, jump to actual trampoline */
 806                 jump_cmpxchg_failed = code;
 807                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 808                 /* if successful, return */
 809                 amd64_ret (code);
 810
 811                 /* next case: synchronization->owner is not null */
 812                 x86_patch (jump_tid, code);
 813                 /* is synchronization->owner == TID? */
 814                 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
 815                 /* if not, jump to actual trampoline */
 816                 jump_other_owner = code;
 817                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 818                 /* if yes, increment nest */
 819                 amd64_inc_membase_size (code, AMD64_RCX, nest_offset, 4);
 820                 /* return */
 821                 amd64_ret (code);
 822
 823                 x86_patch (jump_obj_null, code);
 824                 x86_patch (jump_sync_null, code);
 825                 x86_patch (jump_cmpxchg_failed, code);
 826                 x86_patch (jump_other_owner, code);
 827         }
 828
 829         /* jump to the actual trampoline */
 830 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
 831         amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
 832 #endif
 833
 834         if (aot) {
 835                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
 836                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 837                 amd64_jump_reg (code, AMD64_R11);
 838         } else {
 839                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
 840
 841                 /* jump to the actual trampoline */
 842                 amd64_jump_code (code, tramp);
 843         }
 844
 845         mono_arch_flush_icache (code, code - buf);
 846         g_assert (code - buf <= tramp_size);
 847
 848         *code_size = code - buf;
 849
 850         return buf;
 851 }
 852
 853 gpointer
 854 mono_arch_create_monitor_exit_trampoline (void)
 855 {
 856         guint32 code_size;
 857         MonoJumpInfo *ji;
 858
 859         return mono_arch_create_monitor_exit_trampoline_full (&code_size, &ji, FALSE);
 860 }
 861
 862 gpointer
 863 mono_arch_create_monitor_exit_trampoline_full (guint32 *code_size, MonoJumpInfo **ji, gboolean aot)
 864 {
 865         guint8 *tramp;
 866         guint8 *code, *buf;
 867         guint8 *jump_obj_null, *jump_have_waiters;
 868         guint8 *jump_next;
 869         int tramp_size;
 870         int owner_offset, nest_offset, entry_count_offset;
 871
 872         *ji = NULL;
 873
 874         g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
 875
 876         mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
 877         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
 878         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
 879         g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
 880         owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
 881         nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
 882         entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
 883
 884         tramp_size = 94;
 885
 886         code = buf = mono_global_codeman_reserve (tramp_size);
 887
 888         if (mono_thread_get_tls_offset () != -1) {
 889                 /* MonoObject* obj is in RDI */
 890                 /* is obj null? */
 891                 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
 892                 /* if yes, jump to actual trampoline */
 893                 jump_obj_null = code;
 894                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 895
 896                 /* load obj->synchronization to RCX */
 897                 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
 898                 /* is synchronization null? */
 899                 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
 900                 /* if not, jump to next case */
 901                 jump_next = code;
 902                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 903                 /* if yes, just return */
 904                 amd64_ret (code);
 905
 906                 /* next case: synchronization is not null */
 907                 x86_patch (jump_next, code);
 908                 /* load MonoThread* into RDX */
 909                 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
 910                 /* load TID into RDX */
 911                 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoThread, tid), 8);
 912                 /* is synchronization->owner == TID */
 913                 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
 914                 /* if yes, jump to next case */
 915                 jump_next = code;
 916                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 917                 /* if not, just return */
 918                 amd64_ret (code);
 919
 920                 /* next case: synchronization->owner == TID */
 921                 x86_patch (jump_next, code);
 922                 /* is synchronization->nest == 1 */
 923                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, nest_offset, 1, 4);
 924                 /* if not, jump to next case */
 925                 jump_next = code;
 926                 amd64_branch8 (code, X86_CC_NZ, -1, 1);
 927                 /* if yes, is synchronization->entry_count zero? */
 928                 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, entry_count_offset, 0, 4);
 929                 /* if not, jump to actual trampoline */
 930                 jump_have_waiters = code;
 931                 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
 932                 /* if yes, set synchronization->owner to null and return */
 933                 amd64_mov_membase_imm (code, AMD64_RCX, owner_offset, 0, 8);
 934                 amd64_ret (code);
 935
 936                 /* next case: synchronization->nest is not 1 */
 937                 x86_patch (jump_next, code);
 938                 /* decrease synchronization->nest and return */
 939                 amd64_dec_membase_size (code, AMD64_RCX, nest_offset, 4);
 940                 amd64_ret (code);
 941
 942                 x86_patch (jump_obj_null, code);
 943                 x86_patch (jump_have_waiters, code);
 944         }
 945
 946         /* jump to the actual trampoline */
 947 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
 948         amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
 949 #endif
 950
 951         if (aot) {
 952                 *ji = mono_patch_info_list_prepend (*ji, code - buf, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
 953                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
 954                 amd64_jump_reg (code, AMD64_R11);
 955         } else {
 956                 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
 957                 amd64_jump_code (code, tramp);
 958         }
 959
 960         mono_arch_flush_icache (code, code - buf);
 961         g_assert (code - buf <= tramp_size);
 962
 963         *code_size = code - buf;
 964
 965         return buf;
 966 }
 967 #endif
 968
 969 void
 970 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
 971 {
 972         /* FIXME: This is not thread safe */
 973         guint8 *code = ji->code_start;
 974
 975         amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
 976         amd64_mov_reg_imm (code, AMD64_R11, func);
 977
 978         x86_push_imm (code, (guint64)func_arg);
 979         amd64_call_reg (code, AMD64_R11);
 980 }