mono/mini/tramp-amd64.c

   1 /**
   2  * \file
   3  * JIT trampoline code for amd64
   4  *
   5  * Authors:
   6  *   Dietmar Maurer (dietmar@ximian.com)
   7  *   Zoltan Varga (vargaz@gmail.com)
   8  *   Johan Lorensson (lateralusx.github@gmail.com)
   9  *
  10  * (C) 2001 Ximian, Inc.
  11  * Copyright 2003-2011 Novell, Inc (http://www.novell.com)
  12  * Copyright 2011 Xamarin, Inc (http://www.xamarin.com)
  13  * Licensed under the MIT license. See LICENSE file in the project root for full license information.
  14  */
  15
  16 #include <config.h>
  17 #include <glib.h>
  18
  19 #include <mono/metadata/abi-details.h>
  20 #include <mono/metadata/appdomain.h>
  21 #include <mono/metadata/marshal.h>
  22 #include <mono/metadata/tabledefs.h>
  23 #include <mono/metadata/profiler-private.h>
  24 #include <mono/metadata/gc-internals.h>
  25 #include <mono/arch/amd64/amd64-codegen.h>
  26
  27 #include <mono/utils/memcheck.h>
  28
  29 #include "mini.h"
  30 #include "mini-amd64.h"
  31 #include "debugger-agent.h"
  32
  33 #ifdef ENABLE_INTERPRETER
  34 #include "interp/interp.h"
  35 #endif
  36
  37 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
  38
  39 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
  40
  41 #ifndef DISABLE_JIT
  42 /*
  43  * mono_arch_get_unbox_trampoline:
  44  * @m: method pointer
  45  * @addr: pointer to native code for @m
  46  *
  47  * when value type methods are called through the vtable we need to unbox the
  48  * this argument. This method returns a pointer to a trampoline which does
  49  * unboxing before calling the method
  50  */
  51 gpointer
  52 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
  53 {
  54         guint8 *code, *start;
  55         GSList *unwind_ops;
  56         int this_reg, size = 20;
  57
  58         MonoDomain *domain = mono_domain_get ();
  59
  60         this_reg = mono_arch_get_this_arg_reg (NULL);
  61
  62         start = code = (guint8 *)mono_domain_code_reserve (domain, size + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0));
  63
  64         unwind_ops = mono_arch_get_cie_program ();
  65
  66         amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
  67         /* FIXME: Optimize this */
  68         amd64_mov_reg_imm (code, AMD64_RAX, addr);
  69         amd64_jump_reg (code, AMD64_RAX);
  70         g_assert ((code - start) < size);
  71         g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)));
  72
  73         mono_arch_flush_icache (start, code - start);
  74         mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_UNBOX_TRAMPOLINE, m);
  75
  76         mono_tramp_info_register (mono_tramp_info_create (NULL, start, code - start, NULL, unwind_ops), domain);
  77
  78         return start;
  79 }
  80
  81 /*
  82  * mono_arch_get_static_rgctx_trampoline:
  83  *
  84  *   Create a trampoline which sets RGCTX_REG to ARG, then jumps to ADDR.
  85  */
  86 gpointer
  87 mono_arch_get_static_rgctx_trampoline (gpointer arg, gpointer addr)
  88 {
  89         guint8 *code, *start;
  90         GSList *unwind_ops;
  91         int buf_len;
  92
  93         MonoDomain *domain = mono_domain_get ();
  94
  95 #ifdef MONO_ARCH_NOMAP32BIT
  96         buf_len = 32;
  97 #else
  98         /* AOTed code could still have a non-32 bit address */
  99         if ((((guint64)addr) >> 32) == 0)
 100                 buf_len = 16;
 101         else
 102                 buf_len = 30;
 103 #endif
 104
 105         start = code = (guint8 *)mono_domain_code_reserve (domain, buf_len + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0));
 106
 107         unwind_ops = mono_arch_get_cie_program ();
 108
 109         amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, arg);
 110         amd64_jump_code (code, addr);
 111         g_assert ((code - start) < buf_len);
 112         g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)));
 113
 114         mono_arch_flush_icache (start, code - start);
 115         mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
 116
 117         mono_tramp_info_register (mono_tramp_info_create (NULL, start, code - start, NULL, unwind_ops), domain);
 118
 119         return start;
 120 }
 121 #endif /* !DISABLE_JIT */
 122
 123 #ifdef _WIN64
 124 // Workaround lack of Valgrind support for 64-bit Windows
 125 #define VALGRIND_DISCARD_TRANSLATIONS(...)
 126 #endif
 127
 128 /*
 129  * mono_arch_patch_callsite:
 130  *
 131  *   Patch the callsite whose address is given by ORIG_CODE so it calls ADDR. ORIG_CODE
 132  * points to the pc right after the call.
 133  */
 134 void
 135 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
 136 {
 137         guint8 *code;
 138         guint8 buf [16];
 139         gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
 140
 141         code = buf + 14;
 142
 143         /* mov 64-bit imm into r11 (followed by call reg?)  or direct call*/
 144         if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
 145                 if (code [-5] != 0xe8) {
 146                         if (can_write) {
 147                                 InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
 148                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
 149                         }
 150                 } else {
 151                         gboolean disp_32bit = ((((gint64)addr - (gint64)orig_code)) < (1 << 30)) && ((((gint64)addr - (gint64)orig_code)) > -(1 << 30));
 152
 153                         if ((((guint64)(addr)) >> 32) != 0 && !disp_32bit) {
 154                                 /*
 155                                  * This might happen with LLVM or when calling AOTed code. Create a thunk.
 156                                  */
 157                                 guint8 *thunk_start, *thunk_code;
 158
 159                                 thunk_start = thunk_code = (guint8 *)mono_domain_code_reserve (mono_domain_get (), 32);
 160                                 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
 161                                 *(guint64*)thunk_code = (guint64)addr;
 162                                 addr = thunk_start;
 163                                 g_assert ((((guint64)(addr)) >> 32) == 0);
 164                                 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
 165                                 mono_profiler_code_buffer_new (thunk_start, thunk_code - thunk_start, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
 166                         }
 167                         if (can_write) {
 168                                 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
 169                                 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
 170                         }
 171                 }
 172         }
 173         else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
 174                 /* call *<OFFSET>(%rip) */
 175                 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
 176                 if (can_write) {
 177                         InterlockedExchangePointer (got_entry, addr);
 178                         VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
 179                 }
 180         }
 181 }
 182
 183 #ifndef DISABLE_JIT
 184 guint8*
 185 mono_arch_create_llvm_native_thunk (MonoDomain *domain, guint8 *addr)
 186 {
 187         /*
 188          * The caller is LLVM code and the call displacement might exceed 32 bits. We can't determine the caller address, so
 189          * we add a thunk every time.
 190          * Since the caller is also allocated using the domain code manager, hopefully the displacement will fit into 32 bits.
 191          * FIXME: Avoid this if possible if !MONO_ARCH_NOMAP32BIT and ADDR is 32 bits.
 192          */
 193         guint8 *thunk_start, *thunk_code;
 194
 195         thunk_start = thunk_code = (guint8 *)mono_domain_code_reserve (mono_domain_get (), 32);
 196         amd64_jump_membase (thunk_code, AMD64_RIP, 0);
 197         *(guint64*)thunk_code = (guint64)addr;
 198         addr = thunk_start;
 199         mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
 200         mono_profiler_code_buffer_new (thunk_start, thunk_code - thunk_start, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
 201         return addr;
 202 }
 203 #endif /* !DISABLE_JIT */
 204
 205 void
 206 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
 207 {
 208         gint32 disp;
 209         gpointer *plt_jump_table_entry;
 210
 211         /* A PLT entry: jmp *<DISP>(%rip) */
 212         g_assert (code [0] == 0xff);
 213         g_assert (code [1] == 0x25);
 214
 215         disp = *(gint32*)(code + 2);
 216
 217         plt_jump_table_entry = (gpointer*)(code + 6 + disp);
 218
 219         InterlockedExchangePointer (plt_jump_table_entry, addr);
 220 }
 221
 222 #ifndef DISABLE_JIT
 223 static void
 224 stack_unaligned (MonoTrampolineType tramp_type)
 225 {
 226         printf ("%d\n", tramp_type);
 227         g_assert_not_reached ();
 228 }
 229
 230 guchar*
 231 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
 232 {
 233         char *tramp_name;
 234         guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code, *br_ex_check;
 235         int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, ex_offset, tramp_offset, ctx_offset, saved_regs_offset;
 236         int r11_save_offset, saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
 237         gboolean has_caller;
 238         GSList *unwind_ops = NULL;
 239         MonoJumpInfo *ji = NULL;
 240         const guint kMaxCodeSize = 630;
 241
 242         if (tramp_type == MONO_TRAMPOLINE_JUMP || tramp_type == MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD)
 243                 has_caller = FALSE;
 244         else
 245                 has_caller = TRUE;
 246
 247         code = buf = (guint8 *)mono_global_codeman_reserve (kMaxCodeSize + MONO_MAX_TRAMPOLINE_UNWINDINFO_SIZE);
 248
 249         /* Compute stack frame size and offsets */
 250         offset = 0;
 251         rbp_offset = -offset;
 252
 253         offset += sizeof(mgreg_t);
 254         rax_offset = -offset;
 255
 256         offset += sizeof(mgreg_t);
 257         ex_offset = -offset;
 258
 259         offset += sizeof(mgreg_t);
 260         r11_save_offset = -offset;
 261
 262         offset += sizeof(mgreg_t);
 263         tramp_offset = -offset;
 264
 265         offset += sizeof(gpointer);
 266         arg_offset = -offset;
 267
 268         offset += sizeof(mgreg_t);
 269         res_offset = -offset;
 270
 271         offset += sizeof (MonoContext);
 272         ctx_offset = -offset;
 273         saved_regs_offset = ctx_offset + MONO_STRUCT_OFFSET (MonoContext, gregs);
 274         saved_fpregs_offset = ctx_offset + MONO_STRUCT_OFFSET (MonoContext, fregs);
 275
 276         offset += sizeof (MonoLMFTramp);
 277         lmf_offset = -offset;
 278
 279 #ifdef TARGET_WIN32
 280         /* Reserve space where the callee can save the argument registers */
 281         offset += 4 * sizeof (mgreg_t);
 282 #endif
 283
 284         framesize = ALIGN_TO (offset, MONO_ARCH_FRAME_ALIGNMENT);
 285
 286         // CFA = sp + 16 (the trampoline address is on the stack)
 287         cfa_offset = 16;
 288         mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
 289         // IP saved at CFA - 8
 290         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
 291
 292         orig_rsp_to_rbp_offset = 0;
 293         r11_save_code = code;
 294         /* Reserve space for the mov_membase_reg to save R11 */
 295         code += 5;
 296         after_r11_save_code = code;
 297
 298         /* Pop the return address off the stack */
 299         amd64_pop_reg (code, AMD64_R11);
 300         orig_rsp_to_rbp_offset += sizeof(mgreg_t);
 301
 302         cfa_offset -= sizeof(mgreg_t);
 303         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 304
 305         /*
 306          * Allocate a new stack frame
 307          */
 308         amd64_push_reg (code, AMD64_RBP);
 309         cfa_offset += sizeof(mgreg_t);
 310         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 311         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
 312
 313         orig_rsp_to_rbp_offset -= sizeof(mgreg_t);
 314         amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
 315         mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
 316         mono_add_unwind_op_fp_alloc (unwind_ops, code, buf, AMD64_RBP, 0);
 317         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
 318
 319         /* Compute the trampoline address from the return address */
 320         if (aot) {
 321                 /* 7 = length of call *<offset>(rip) */
 322                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
 323         } else {
 324                 /* 5 = length of amd64_call_membase () */
 325                 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
 326         }
 327         amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, sizeof(gpointer));
 328
 329         /* Save all registers */
 330         for (i = 0; i < AMD64_NREG; ++i) {
 331                 if (i == AMD64_RBP) {
 332                         /* RAX is already saved */
 333                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, sizeof(mgreg_t));
 334                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_RAX, sizeof(mgreg_t));
 335                 } else if (i == AMD64_RIP) {
 336                         if (has_caller)
 337                                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, sizeof(gpointer));
 338                         else
 339                                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 340                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_R11, sizeof(mgreg_t));
 341                 } else if (i == AMD64_RSP) {
 342                         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof(mgreg_t));
 343                         amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
 344                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_R11, sizeof(mgreg_t));
 345                 } else if (i != AMD64_R11) {
 346                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
 347                 } else {
 348                         /* We have to save R11 right at the start of
 349                            the trampoline code because it's used as a
 350                            scratch register */
 351                         /* This happens before the frame is set up, so it goes into the redzone */
 352                         amd64_mov_membase_reg (r11_save_code, AMD64_RSP, r11_save_offset + orig_rsp_to_rbp_offset, i, sizeof(mgreg_t));
 353                         g_assert (r11_save_code == after_r11_save_code);
 354
 355                         /* Copy from the save slot into the register array slot */
 356                         amd64_mov_reg_membase (code, i, AMD64_RSP, r11_save_offset + orig_rsp_to_rbp_offset, sizeof(mgreg_t));
 357                         amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
 358                 }
 359                 /* cfa = rbp + cfa_offset */
 360                 mono_add_unwind_op_offset (unwind_ops, code, buf, i, - cfa_offset + saved_regs_offset + (i * sizeof (mgreg_t)));
 361         }
 362         for (i = 0; i < 8; ++i)
 363                 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)), i);
 364
 365         /* Check that the stack is aligned */
 366         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof (mgreg_t));
 367         amd64_alu_reg_imm (code, X86_AND, AMD64_R11, 15);
 368         amd64_alu_reg_imm (code, X86_CMP, AMD64_R11, 0);
 369         br [0] = code;
 370         amd64_branch_disp (code, X86_CC_Z, 0, FALSE);
 371         if (aot) {
 372                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 373                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
 374         } else {
 375                 amd64_mov_reg_imm (code, MONO_AMD64_ARG_REG1, tramp_type);
 376                 amd64_mov_reg_imm (code, AMD64_R11, stack_unaligned);
 377                 amd64_call_reg (code, AMD64_R11);
 378         }
 379         mono_amd64_patch (br [0], code);
 380         //amd64_breakpoint (code);
 381
 382         if (tramp_type != MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) {
 383                 /* Obtain the trampoline argument which is encoded in the instruction stream */
 384                 if (aot) {
 385                         /* Load the GOT offset */
 386                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
 387                         /*
 388                          * r11 points to a call *<offset>(%rip) instruction, load the
 389                          * pc-relative offset from the instruction itself.
 390                          */
 391                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 3, 4);
 392                         /* 7 is the length of the call, 8 is the offset to the next got slot */
 393                         amd64_alu_reg_imm_size (code, X86_ADD, AMD64_RAX, 7 + sizeof (gpointer), sizeof(gpointer));
 394                         /* Compute the address of the GOT slot */
 395                         amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, sizeof(gpointer));
 396                         /* Load the value */
 397                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer));
 398                 } else {
 399                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
 400                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
 401                         amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
 402                         amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
 403                         br [0] = code;
 404                         x86_branch8 (code, X86_CC_NE, 6, FALSE);
 405                         /* 32 bit immediate */
 406                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
 407                         br [1] = code;
 408                         x86_jump8 (code, 10);
 409                         /* 64 bit immediate */
 410                         mono_amd64_patch (br [0], code);
 411                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
 412                         mono_amd64_patch (br [1], code);
 413                 }
 414                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
 415         } else {
 416                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * sizeof(mgreg_t)), sizeof(mgreg_t));
 417                 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
 418         }
 419
 420         /* Save LMF begin */
 421
 422         /* Save ip */
 423         if (has_caller)
 424                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, sizeof(gpointer));
 425         else
 426                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 427         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, sizeof(mgreg_t));
 428         /* Save sp */
 429         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof(mgreg_t));
 430         amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
 431         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, sizeof(mgreg_t));
 432         /* Save pointer to context */
 433         amd64_lea_membase (code, AMD64_R11, AMD64_RBP, ctx_offset);
 434         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMFTramp, ctx), AMD64_R11, sizeof(mgreg_t));
 435
 436         if (aot) {
 437                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
 438         } else {
 439                 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
 440         }
 441         amd64_call_reg (code, AMD64_R11);
 442
 443         /* Save lmf_addr */
 444         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMFTramp, lmf_addr), AMD64_RAX, sizeof(gpointer));
 445         /* Save previous_lmf */
 446         /* Set the lowest bit to signal that this LMF has the ip field set */
 447         /* Set the third lowest bit to signal that this is a MonoLMFTramp structure */
 448         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
 449         amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 0x5, sizeof(gpointer));
 450         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
 451         /* Set new lmf */
 452         amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
 453         amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
 454
 455         /* Save LMF end */
 456
 457         /* Arg1 is the pointer to the saved registers */
 458         amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
 459
 460         /* Arg2 is the address of the calling code */
 461         if (has_caller)
 462                 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, sizeof(gpointer));
 463         else
 464                 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
 465
 466         /* Arg3 is the method/vtable ptr */
 467         amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, sizeof(gpointer));
 468
 469         /* Arg4 is the trampoline address */
 470         amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, sizeof(gpointer));
 471
 472         if (aot) {
 473                 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
 474                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
 475         } else {
 476                 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
 477                 amd64_mov_reg_imm (code, AMD64_R11, tramp);
 478         }
 479         amd64_call_reg (code, AMD64_R11);
 480         amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, sizeof(mgreg_t));
 481
 482         /* Restore LMF */
 483         amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
 484         amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 0x5, sizeof(gpointer));
 485         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMFTramp, lmf_addr), sizeof(gpointer));
 486         amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
 487
 488         /*
 489          * Save rax to the stack, after the leave instruction, this will become part of
 490          * the red zone.
 491          */
 492         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t));
 493         amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, sizeof(mgreg_t));
 494
 495         /* Check for thread interruption */
 496         /* This is not perf critical code so no need to check the interrupt flag */
 497         /*
 498          * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
 499          */
 500         if (aot) {
 501                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint_noraise");
 502         } else {
 503                 amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint_noraise);
 504         }
 505         amd64_call_reg (code, AMD64_R11);
 506
 507         amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 508         br_ex_check = code;
 509         amd64_branch8 (code, X86_CC_Z, -1, 1);
 510
 511         /*
 512          * Exception case:
 513          * We have an exception we want to throw in the caller's frame, so pop
 514          * the trampoline frame and throw from the caller.
 515          */
 516 #if TARGET_WIN32
 517         amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, 0);
 518         amd64_pop_reg (code, AMD64_RBP);
 519         mono_add_unwind_op_same_value (unwind_ops, code, buf, AMD64_RBP);
 520 #else
 521         amd64_leave (code);
 522 #endif
 523         /* We are in the parent frame, the exception is in rax */
 524         /*
 525          * EH is initialized after trampolines, so get the address of the variable
 526          * which contains throw_exception, and load it from there.
 527          */
 528         if (aot) {
 529                 /* Not really a jit icall */
 530                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "throw_exception_addr");
 531         } else {
 532                 amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_get_throw_exception_addr ());
 533         }
 534         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer));
 535         amd64_mov_reg_reg (code, AMD64_ARG_REG1, AMD64_RAX, sizeof(mgreg_t));
 536         /*
 537          * We still have the original return value on the top of the stack, so the
 538          * throw trampoline will use that as the throw site.
 539          */
 540         amd64_jump_reg (code, AMD64_R11);
 541
 542         /* Normal case */
 543         mono_amd64_patch (br_ex_check, code);
 544
 545         /* Restore argument registers, r10 (imt method/rgxtx)
 546            and rax (needed for direct calls to C vararg functions). */
 547         for (i = 0; i < AMD64_NREG; ++i)
 548                 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
 549                         amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
 550         for (i = 0; i < 8; ++i)
 551                 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)));
 552
 553         /* Restore stack */
 554 #if TARGET_WIN32
 555         amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, 0);
 556         amd64_pop_reg (code, AMD64_RBP);
 557         mono_add_unwind_op_same_value (unwind_ops, code, buf, AMD64_RBP);
 558 #else
 559         amd64_leave (code);
 560 #endif
 561         cfa_offset -= sizeof (mgreg_t);
 562         mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, cfa_offset);
 563
 564         if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
 565                 /* Load result */
 566                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - sizeof(mgreg_t), sizeof(mgreg_t));
 567                 amd64_ret (code);
 568         } else {
 569                 /* call the compiled method using the saved rax */
 570                 amd64_jump_membase (code, AMD64_RSP, rax_offset - sizeof(mgreg_t));
 571         }
 572
 573         g_assert ((code - buf) <= kMaxCodeSize);
 574         g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_MAX_TRAMPOLINE_UNWINDINFO_SIZE));
 575
 576         mono_arch_flush_icache (buf, code - buf);
 577         mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
 578
 579         tramp_name = mono_get_generic_trampoline_name (tramp_type);
 580         *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);
 581         g_free (tramp_name);
 582
 583         return buf;
 584 }
 585
 586 gpointer
 587 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
 588 {
 589         guint8 *code, *buf, *tramp;
 590         int size;
 591         gboolean far_addr = FALSE;
 592
 593         tramp = mono_get_trampoline_code (tramp_type);
 594
 595         if ((((guint64)arg1) >> 32) == 0)
 596                 size = 5 + 1 + 4;
 597         else
 598                 size = 5 + 1 + 8;
 599
 600         code = buf = (guint8 *)mono_domain_code_reserve_align (domain, size, 1);
 601
 602         if (((gint64)tramp - (gint64)code) >> 31 != 0 && ((gint64)tramp - (gint64)code) >> 31 != -1) {
 603 #ifndef MONO_ARCH_NOMAP32BIT
 604                 g_assert_not_reached ();
 605 #endif
 606                 far_addr = TRUE;
 607                 size += 16;
 608                 code = buf = (guint8 *)mono_domain_code_reserve_align (domain, size, 1);
 609         }
 610
 611         if (far_addr) {
 612                 amd64_mov_reg_imm (code, AMD64_R11, tramp);
 613                 amd64_call_reg (code, AMD64_R11);
 614         } else {
 615                 amd64_call_code (code, tramp);
 616         }
 617         /* The trampoline code will obtain the argument from the instruction stream */
 618         if ((((guint64)arg1) >> 32) == 0) {
 619                 *code = 0x4;
 620                 *(guint32*)(code + 1) = (gint64)arg1;
 621                 code += 5;
 622         } else {
 623                 *code = 0x8;
 624                 *(guint64*)(code + 1) = (gint64)arg1;
 625                 code += 9;
 626         }
 627
 628         g_assert ((code - buf) <= size);
 629
 630         if (code_len)
 631                 *code_len = size;
 632
 633         mono_arch_flush_icache (buf, size);
 634         mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_SPECIFIC_TRAMPOLINE, mono_get_generic_trampoline_simple_name (tramp_type));
 635
 636         return buf;
 637 }
 638
 639 gpointer
 640 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
 641 {
 642         guint8 *tramp;
 643         guint8 *code, *buf;
 644         guint8 **rgctx_null_jumps;
 645         int tramp_size;
 646         int depth, index;
 647         int i;
 648         gboolean mrgctx;
 649         MonoJumpInfo *ji = NULL;
 650         GSList *unwind_ops;
 651
 652         mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
 653         index = MONO_RGCTX_SLOT_INDEX (slot);
 654         if (mrgctx)
 655                 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
 656         for (depth = 0; ; ++depth) {
 657                 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
 658
 659                 if (index < size - 1)
 660                         break;
 661                 index -= size - 1;
 662         }
 663
 664         tramp_size = 64 + 8 * depth;
 665
 666         code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0));
 667
 668         unwind_ops = mono_arch_get_cie_program ();
 669
 670         rgctx_null_jumps = (guint8 **)g_malloc (sizeof (guint8*) * (depth + 2));
 671
 672         if (mrgctx) {
 673                 /* get mrgctx ptr */
 674                 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
 675         } else {
 676                 /* load rgctx ptr from vtable */
 677                 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoVTable, runtime_generic_context), sizeof(gpointer));
 678                 /* is the rgctx ptr null? */
 679                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 680                 /* if yes, jump to actual trampoline */
 681                 rgctx_null_jumps [0] = code;
 682                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 683         }
 684
 685         for (i = 0; i < depth; ++i) {
 686                 /* load ptr to next array */
 687                 if (mrgctx && i == 0)
 688                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, sizeof(gpointer));
 689                 else
 690                         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, sizeof(gpointer));
 691                 /* is the ptr null? */
 692                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 693                 /* if yes, jump to actual trampoline */
 694                 rgctx_null_jumps [i + 1] = code;
 695                 amd64_branch8 (code, X86_CC_Z, -1, 1);
 696         }
 697
 698         /* fetch slot */
 699         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), sizeof(gpointer));
 700         /* is the slot null? */
 701         amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
 702         /* if yes, jump to actual trampoline */
 703         rgctx_null_jumps [depth + 1] = code;
 704         amd64_branch8 (code, X86_CC_Z, -1, 1);
 705         /* otherwise return */
 706         amd64_ret (code);
 707
 708         for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
 709                 mono_amd64_patch (rgctx_null_jumps [i], code);
 710
 711         g_free (rgctx_null_jumps);
 712
 713         /* move the rgctx pointer to the VTABLE register */
 714         amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
 715
 716         if (aot) {
 717                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
 718                 amd64_jump_reg (code, AMD64_R11);
 719         } else {
 720                 tramp = (guint8 *)mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
 721
 722                 /* jump to the actual trampoline */
 723                 amd64_jump_code (code, tramp);
 724         }
 725
 726         mono_arch_flush_icache (buf, code - buf);
 727         mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
 728
 729         g_assert (code - buf <= tramp_size);
 730         g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)));
 731
 732         char *name = mono_get_rgctx_fetch_trampoline_name (slot);
 733         *info = mono_tramp_info_create (name, buf, code - buf, ji, unwind_ops);
 734         g_free (name);
 735
 736         return buf;
 737 }
 738
 739 gpointer
 740 mono_arch_create_general_rgctx_lazy_fetch_trampoline (MonoTrampInfo **info, gboolean aot)
 741 {
 742         guint8 *code, *buf;
 743         int tramp_size;
 744         MonoJumpInfo *ji = NULL;
 745         GSList *unwind_ops;
 746
 747         g_assert (aot);
 748         tramp_size = 64;
 749
 750         code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0));
 751
 752         unwind_ops = mono_arch_get_cie_program ();
 753
 754         // FIXME: Currently, we always go to the slow path.
 755         /* This receives a <slot, trampoline> in the rgctx arg reg. */
 756         /* Load trampoline addr */
 757         amd64_mov_reg_membase (code, AMD64_R11, MONO_ARCH_RGCTX_REG, 8, 8);
 758         /* move the rgctx pointer to the VTABLE register */
 759         amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
 760         /* Jump to the trampoline */
 761         amd64_jump_reg (code, AMD64_R11);
 762
 763         mono_arch_flush_icache (buf, code - buf);
 764         mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
 765
 766         g_assert (code - buf <= tramp_size);
 767         g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)));
 768
 769         if (info)
 770                 *info = mono_tramp_info_create ("rgctx_fetch_trampoline_general", buf, code - buf, ji, unwind_ops);
 771
 772         return buf;
 773 }
 774
 775 void
 776 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
 777 {
 778         /* FIXME: This is not thread safe */
 779         guint8 *code = (guint8 *)ji->code_start;
 780
 781         amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
 782         amd64_mov_reg_imm (code, AMD64_R11, func);
 783
 784         x86_push_imm (code, (guint64)func_arg);
 785         amd64_call_reg (code, AMD64_R11);
 786 }
 787 #endif /* !DISABLE_JIT */
 788
 789 gpointer
 790 mono_amd64_handler_block_trampoline_helper (void)
 791 {
 792         MonoJitTlsData *jit_tls = (MonoJitTlsData *)mono_tls_get_jit_tls ();
 793         return jit_tls->handler_block_return_address;
 794 }
 795
 796 #ifndef DISABLE_JIT
 797 gpointer
 798 mono_arch_create_handler_block_trampoline (MonoTrampInfo **info, gboolean aot)
 799 {
 800         guint8 *code, *buf;
 801         int tramp_size = 64;
 802         MonoJumpInfo *ji = NULL;
 803         GSList *unwind_ops;
 804
 805         code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0));
 806
 807         unwind_ops = mono_arch_get_cie_program ();
 808
 809         /*
 810          * This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
 811          * We get here from the ret emitted by CEE_ENDFINALLY.
 812          * The stack is misaligned.
 813          */
 814         /* Align the stack before the call to mono_amd64_handler_block_trampoline_helper() */
 815 #ifdef TARGET_WIN32
 816         /* Also make room for the "register parameter stack area" as specified by the Windows x64 ABI (4 64-bit registers) */
 817         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8 + 4 * 8);
 818 #else
 819         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
 820 #endif
 821         if (aot) {
 822                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_amd64_handler_block_trampoline_helper");
 823                 amd64_call_reg (code, AMD64_R11);
 824         } else {
 825                 amd64_mov_reg_imm (code, AMD64_RAX, mono_amd64_handler_block_trampoline_helper);
 826                 amd64_call_reg (code, AMD64_RAX);
 827         }
 828         /* Undo stack alignment */
 829 #ifdef TARGET_WIN32
 830         amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8 + 4 * 8);
 831 #else
 832         amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
 833 #endif
 834         /* Save the result to the stack */
 835         amd64_push_reg (code, AMD64_RAX);
 836 #ifdef TARGET_WIN32
 837         /* Make room for the "register parameter stack area" as specified by the Windows x64 ABI (4 64-bit registers) */
 838         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 4 * 8);
 839 #endif
 840         if (aot) {
 841                 char *name = g_strdup_printf ("trampoline_func_%d", MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
 842                 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, name);
 843                 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_R11, 8);
 844         } else {
 845                 amd64_mov_reg_imm (code, AMD64_RAX, mono_get_trampoline_func (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD));
 846         }
 847         /* The stack is aligned */
 848         amd64_call_reg (code, AMD64_RAX);
 849 #ifdef TARGET_WIN32
 850         amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 4 * 8);
 851 #endif
 852         /* Load return address */
 853         amd64_pop_reg (code, AMD64_RAX);
 854         /* The stack is misaligned, thats what the code we branch to expects */
 855         amd64_jump_reg (code, AMD64_RAX);
 856
 857         mono_arch_flush_icache (buf, code - buf);
 858         mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
 859         g_assert (code - buf <= tramp_size);
 860         g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)));
 861
 862         *info = mono_tramp_info_create ("handler_block_trampoline", buf, code - buf, ji, unwind_ops);
 863
 864         return buf;
 865 }
 866 #endif /* !DISABLE_JIT */
 867
 868 /*
 869  * mono_arch_get_call_target:
 870  *
 871  *   Return the address called by the code before CODE if exists.
 872  */
 873 guint8*
 874 mono_arch_get_call_target (guint8 *code)
 875 {
 876         if (code [-5] == 0xe8) {
 877                 gint32 disp = *(gint32*)(code - 4);
 878                 guint8 *target = code + disp;
 879
 880                 return target;
 881         } else {
 882                 return NULL;
 883         }
 884 }
 885
 886 /*
 887  * mono_arch_get_plt_info_offset:
 888  *
 889  *   Return the PLT info offset belonging to the plt entry PLT_ENTRY.
 890  */
 891 guint32
 892 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
 893 {
 894         return *(guint32*)(plt_entry + 6);
 895 }
 896
 897 #ifndef DISABLE_JIT
 898 /*
 899  * mono_arch_create_sdb_trampoline:
 900  *
 901  *   Return a trampoline which captures the current context, passes it to
 902  * debugger_agent_single_step_from_context ()/debugger_agent_breakpoint_from_context (),
 903  * then restores the (potentially changed) context.
 904  */
 905 guint8*
 906 mono_arch_create_sdb_trampoline (gboolean single_step, MonoTrampInfo **info, gboolean aot)
 907 {
 908         int tramp_size = 512;
 909         int i, framesize, ctx_offset, cfa_offset, gregs_offset;
 910         guint8 *code, *buf;
 911         GSList *unwind_ops = NULL;
 912         MonoJumpInfo *ji = NULL;
 913
 914         code = buf = (guint8 *)mono_global_codeman_reserve (tramp_size + MONO_MAX_TRAMPOLINE_UNWINDINFO_SIZE);
 915
 916         framesize = 0;
 917 #ifdef TARGET_WIN32
 918         /* Reserve space where the callee can save the argument registers */
 919         framesize += 4 * sizeof (mgreg_t);
 920 #endif
 921
 922         ctx_offset = framesize;
 923         framesize += sizeof (MonoContext);
 924
 925         framesize = ALIGN_TO (framesize, MONO_ARCH_FRAME_ALIGNMENT);
 926
 927         // CFA = sp + 8
 928         cfa_offset = 8;
 929         mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 8);
 930         // IP saved at CFA - 8
 931         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -cfa_offset);
 932
 933         amd64_push_reg (code, AMD64_RBP);
 934         cfa_offset += sizeof(mgreg_t);
 935         mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
 936         mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
 937
 938         amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
 939         mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
 940         mono_add_unwind_op_fp_alloc (unwind_ops, code, buf, AMD64_RBP, 0);
 941         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
 942
 943         gregs_offset = ctx_offset + MONO_STRUCT_OFFSET (MonoContext, gregs);
 944
 945         /* Initialize a MonoContext structure on the stack */
 946         for (i = 0; i < AMD64_NREG; ++i) {
 947                 if (i != AMD64_RIP && i != AMD64_RSP && i != AMD64_RBP)
 948                         amd64_mov_membase_reg (code, AMD64_RSP, gregs_offset + (i * sizeof (mgreg_t)), i, sizeof (mgreg_t));
 949         }
 950         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 0, sizeof (mgreg_t));
 951         amd64_mov_membase_reg (code, AMD64_RSP, gregs_offset + (AMD64_RBP * sizeof (mgreg_t)), AMD64_R11, sizeof (mgreg_t));
 952         amd64_lea_membase (code, AMD64_R11, AMD64_RBP, 2 * sizeof (mgreg_t));
 953         amd64_mov_membase_reg (code, AMD64_RSP, gregs_offset + (AMD64_RSP * sizeof (mgreg_t)), AMD64_R11, sizeof (mgreg_t));
 954         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, sizeof (mgreg_t), sizeof (mgreg_t));
 955         amd64_mov_membase_reg (code, AMD64_RSP, gregs_offset + (AMD64_RIP * sizeof (mgreg_t)), AMD64_R11, sizeof (mgreg_t));
 956
 957         /* Call the single step/breakpoint function in sdb */
 958         amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RSP, ctx_offset);
 959
 960         if (aot) {
 961                 if (single_step)
 962                         code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "debugger_agent_single_step_from_context");
 963                 else
 964                         code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "debugger_agent_breakpoint_from_context");
 965         } else {
 966                 if (single_step)
 967                         amd64_mov_reg_imm (code, AMD64_R11, debugger_agent_single_step_from_context);
 968                 else
 969                         amd64_mov_reg_imm (code, AMD64_R11, debugger_agent_breakpoint_from_context);
 970         }
 971         amd64_call_reg (code, AMD64_R11);
 972
 973         /* Restore registers from ctx */
 974         for (i = 0; i < AMD64_NREG; ++i) {
 975                 if (i != AMD64_RIP && i != AMD64_RSP && i != AMD64_RBP)
 976                         amd64_mov_reg_membase (code, i, AMD64_RSP, gregs_offset + (i * sizeof (mgreg_t)), sizeof (mgreg_t));
 977         }
 978         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, gregs_offset + (AMD64_RBP * sizeof (mgreg_t)), sizeof (mgreg_t));
 979         amd64_mov_membase_reg (code, AMD64_RBP, 0, AMD64_R11, sizeof (mgreg_t));
 980         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, gregs_offset + (AMD64_RIP * sizeof (mgreg_t)), sizeof (mgreg_t));
 981         amd64_mov_membase_reg (code, AMD64_RBP, sizeof (mgreg_t), AMD64_R11, sizeof (mgreg_t));
 982
 983 #if TARGET_WIN32
 984         amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, 0);
 985         amd64_pop_reg (code, AMD64_RBP);
 986         mono_add_unwind_op_same_value (unwind_ops, code, buf, AMD64_RBP);
 987 #else
 988         amd64_leave (code);
 989 #endif
 990         cfa_offset -= sizeof (mgreg_t);
 991         mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, cfa_offset);
 992         amd64_ret (code);
 993
 994         mono_arch_flush_icache (code, code - buf);
 995         mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
 996         g_assert (code - buf <= tramp_size);
 997         g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_MAX_TRAMPOLINE_UNWINDINFO_SIZE));
 998
 999         const char *tramp_name = single_step ? "sdb_single_step_trampoline" : "sdb_breakpoint_trampoline";
1000         *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);
1001
1002         return buf;
1003 }
1004
1005 /*
1006  * mono_arch_get_enter_icall_trampoline:
1007  *
1008  *   A trampoline that handles the transition from interpreter into native
1009  *   world. It requiers to set up a descriptor (InterpMethodArguments), so the
1010  *   trampoline can translate the arguments into the native calling convention.
1011  *
1012  *   See also `build_args_from_sig ()` in interp.c.
1013  */
1014 gpointer
1015 mono_arch_get_enter_icall_trampoline (MonoTrampInfo **info)
1016 {
1017 #ifdef ENABLE_INTERPRETER
1018         const int gregs_num = INTERP_ICALL_TRAMP_IARGS;
1019         const int fregs_num = INTERP_ICALL_TRAMP_FARGS;
1020         guint8 *start = NULL, *code, *label_gexits [gregs_num], *label_fexits [fregs_num], *label_leave_tramp [3], *label_is_float_ret;
1021         MonoJumpInfo *ji = NULL;
1022         GSList *unwind_ops = NULL;
1023         static int farg_regs[] = {AMD64_XMM0, AMD64_XMM1, AMD64_XMM2};
1024         int buf_len, i, framesize = 0, off_rbp, off_methodargs, off_targetaddr;
1025
1026         buf_len = 512 + MONO_TRAMPOLINE_UNWINDINFO_SIZE(0);
1027         start = code = (guint8 *) mono_global_codeman_reserve (buf_len);
1028
1029         off_rbp = -framesize;
1030
1031         framesize += sizeof (mgreg_t);
1032         off_methodargs = -framesize;
1033
1034         framesize += sizeof (mgreg_t);
1035         off_targetaddr = -framesize;
1036
1037         framesize += (gregs_num - PARAM_REGS) * sizeof (mgreg_t);
1038
1039         amd64_push_reg (code, AMD64_RBP);
1040         amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof (mgreg_t));
1041         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ALIGN_TO (framesize, MONO_ARCH_FRAME_ALIGNMENT));
1042
1043         /* save InterpMethodArguments* onto stack */
1044         amd64_mov_membase_reg (code, AMD64_RBP, off_methodargs, AMD64_ARG_REG2, sizeof (mgreg_t));
1045
1046         /* save target address on stack */
1047         amd64_mov_membase_reg (code, AMD64_RBP, off_targetaddr, AMD64_ARG_REG1, sizeof (mgreg_t));
1048
1049         /* load pointer to InterpMethodArguments* into R11 */
1050         amd64_mov_reg_reg (code, AMD64_R11, AMD64_ARG_REG2, 8);
1051
1052         /* move flen into RAX */
1053         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, flen), sizeof (mgreg_t));
1054         /* load pointer to fargs into R11 */
1055         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, fargs), sizeof (mgreg_t));
1056
1057         for (i = 0; i < fregs_num; ++i) {
1058                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
1059                 label_fexits [i] = code;
1060                 x86_branch8 (code, X86_CC_Z, 0, FALSE);
1061
1062                 amd64_sse_movsd_reg_membase (code, farg_regs [i], AMD64_R11, i * sizeof (double));
1063                 amd64_dec_reg_size (code, AMD64_RAX, 1);
1064         }
1065
1066         for (i = 0; i < fregs_num; i++)
1067                 x86_patch (label_fexits [i], code);
1068
1069         /* load pointer to InterpMethodArguments* into R11 */
1070         amd64_mov_reg_reg (code, AMD64_R11, AMD64_ARG_REG2, sizeof (mgreg_t));
1071         /* move ilen into RAX */
1072         amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, ilen), sizeof (mgreg_t));
1073
1074         int stack_offset = 0;
1075         for (i = 0; i < gregs_num; i++) {
1076                 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
1077                 label_gexits [i] = code;
1078                 x86_branch32 (code, X86_CC_Z, 0, FALSE);
1079
1080                 /* load pointer to InterpMethodArguments* into R11 */
1081                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_methodargs, sizeof (mgreg_t));
1082                 /* load pointer to iargs into R11 */
1083                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, iargs), sizeof (mgreg_t));
1084
1085                 if (i < PARAM_REGS) {
1086                         amd64_mov_reg_membase (code, param_regs [i], AMD64_R11, i * sizeof (mgreg_t), sizeof (mgreg_t));
1087                 } else {
1088                         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, i * sizeof (mgreg_t), sizeof (mgreg_t));
1089                         amd64_mov_membase_reg (code, AMD64_RSP, stack_offset, AMD64_R11, sizeof (mgreg_t));
1090                         stack_offset += sizeof (mgreg_t);
1091                 }
1092                 amd64_dec_reg_size (code, AMD64_RAX, 1);
1093         }
1094
1095         for (i = 0; i < gregs_num; i++)
1096                 x86_patch (label_gexits [i], code);
1097
1098         /* load target addr */
1099         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_targetaddr, sizeof (mgreg_t));
1100
1101         /* call into native function */
1102         amd64_call_reg (code, AMD64_R11);
1103
1104         /* load InterpMethodArguments */
1105         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_methodargs, sizeof (mgreg_t));
1106
1107         /* load is_float_ret */
1108         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, is_float_ret), sizeof (mgreg_t));
1109
1110         /* check if a float return value is expected */
1111         amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
1112
1113         label_is_float_ret = code;
1114         x86_branch8 (code, X86_CC_NZ, 0, FALSE);
1115
1116         /* greg return */
1117         /* load InterpMethodArguments */
1118         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_methodargs, sizeof (mgreg_t));
1119         /* load retval */
1120         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, retval), sizeof (mgreg_t));
1121
1122         amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
1123         label_leave_tramp [0] = code;
1124         x86_branch8 (code, X86_CC_Z, 0, FALSE);
1125
1126         amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RAX, sizeof (mgreg_t));
1127
1128         label_leave_tramp [1] = code;
1129         x86_jump8 (code, 0);
1130
1131         /* freg return */
1132         x86_patch (label_is_float_ret, code);
1133         /* load InterpMethodArguments */
1134         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, off_methodargs, sizeof (mgreg_t));
1135         /* load retval */
1136         amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (InterpMethodArguments, retval), sizeof (mgreg_t));
1137
1138         amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
1139         label_leave_tramp [2] = code;
1140         x86_branch8 (code, X86_CC_Z, 0, FALSE);
1141
1142         amd64_sse_movsd_membase_reg (code, AMD64_R11, 0, AMD64_XMM0);
1143
1144         for (i = 0; i < 3; i++)
1145                 x86_patch (label_leave_tramp [i], code);
1146
1147         amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, ALIGN_TO (framesize, MONO_ARCH_FRAME_ALIGNMENT));
1148         amd64_pop_reg (code, AMD64_RBP);
1149         amd64_ret (code);
1150
1151         g_assert (code - start < buf_len);
1152
1153         mono_arch_flush_icache (start, code - start);
1154         mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_EXCEPTION_HANDLING, NULL);
1155
1156         if (info)
1157                 *info = mono_tramp_info_create ("enter_icall_trampoline", start, code - start, ji, unwind_ops);
1158
1159         return start;
1160 #else
1161         g_assert_not_reached ();
1162         return NULL;
1163 #endif /* ENABLE_INTERPRETER */
1164 }
1165 #endif /* !DISABLE_JIT */
1166
1167 #ifdef DISABLE_JIT
1168 gpointer
1169 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
1170 {
1171         g_assert_not_reached ();
1172         return NULL;
1173 }
1174
1175 gpointer
1176 mono_arch_get_static_rgctx_trampoline (gpointer arg, gpointer addr)
1177 {
1178         g_assert_not_reached ();
1179         return NULL;
1180 }
1181
1182 gpointer
1183 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
1184 {
1185         g_assert_not_reached ();
1186         return NULL;
1187 }
1188
1189 guchar*
1190 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
1191 {
1192         g_assert_not_reached ();
1193         return NULL;
1194 }
1195
1196 gpointer
1197 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
1198 {
1199         g_assert_not_reached ();
1200         return NULL;
1201 }
1202
1203 gpointer
1204 mono_arch_create_general_rgctx_lazy_fetch_trampoline (MonoTrampInfo **info, gboolean aot)
1205 {
1206         g_assert_not_reached ();
1207         return NULL;
1208 }
1209
1210 gpointer
1211 mono_arch_create_handler_block_trampoline (MonoTrampInfo **info, gboolean aot)
1212 {
1213         g_assert_not_reached ();
1214         return NULL;
1215 }
1216
1217 void
1218 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1219 {
1220         g_assert_not_reached ();
1221         return;
1222 }
1223
1224 guint8*
1225 mono_arch_create_sdb_trampoline (gboolean single_step, MonoTrampInfo **info, gboolean aot)
1226 {
1227         g_assert_not_reached ();
1228         return NULL;
1229 }
1230
1231 gpointer
1232 mono_arch_get_enter_icall_trampoline (MonoTrampInfo **info)
1233 {
1234         g_assert_not_reached ();
1235         return NULL;
1236 }
1237 #endif /* DISABLE_JIT */