mono/mini/tramp-amd64.c

   1 /*
   2  * tramp-x86.c: JIT trampoline code for x86
   3  *
   4  * Authors:
   5  *   Dietmar Maurer (dietmar@ximian.com)
   6  *
   7  * (C) 2001 Ximian, Inc.
   8  */
   9
  10 #include <config.h>
  11 #include <glib.h>
  12
  13 #include <mono/metadata/appdomain.h>
  14 #include <mono/metadata/marshal.h>
  15 #include <mono/metadata/tabledefs.h>
  16 #include <mono/metadata/mono-debug-debugger.h>
  17 #include <mono/arch/amd64/amd64-codegen.h>
  18
  19 #ifdef HAVE_VALGRIND_MEMCHECK_H
  20 #include <valgrind/memcheck.h>
  21 #endif
  22
  23 #include "mini.h"
  24 #include "mini-amd64.h"
  25
  26 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
  27
  28 static guint8* nullified_class_init_trampoline;
  29
  30 /*
  31  * mono_arch_get_unbox_trampoline:
  32  * @m: method pointer
  33  * @addr: pointer to native code for @m
  34  *
  35  * when value type methods are called through the vtable we need to unbox the
  36  * this argument. This method returns a pointer to a trampoline which does
  37  * unboxing before calling the method
  38  */
  39 gpointer
  40 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
  41 {
  42         guint8 *code, *start;
  43         int this_reg = AMD64_RDI;
  44         MonoDomain *domain = mono_domain_get ();
  45
  46         if (!mono_method_signature (m)->ret->byref && MONO_TYPE_ISSTRUCT (mono_method_signature (m)->ret))
  47                 this_reg = AMD64_RSI;
  48
  49         mono_domain_lock (domain);
  50         start = code = mono_code_manager_reserve (domain->code_mp, 20);
  51         mono_domain_unlock (domain);
  52
  53         amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
  54         /* FIXME: Optimize this */
  55         amd64_mov_reg_imm (code, AMD64_RAX, addr);
  56         amd64_jump_reg (code, AMD64_RAX);
  57         g_assert ((code - start) < 20);
  58
  59         mono_arch_flush_icache (start, code - start);
  60
  61         return start;
  62 }
  63
  64 void
  65 mono_arch_patch_callsite (guint8 *code, guint8 *addr)
  66 {
  67         if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
  68                 if (code [-5] != 0xe8)
  69                         InterlockedExchangePointer ((gpointer*)(code - 11), addr);
  70                 else {
  71                         g_assert ((((guint64)(addr)) >> 32) == 0);
  72                         g_assert ((((guint64)(code)) >> 32) == 0);
  73                         InterlockedExchange ((gint32*)(code - 4), ((gint64)addr - (gint64)code));
  74                 }
  75         }
  76         else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
  77                 /* call *<OFFSET>(%rip) */
  78                 gpointer *got_entry = (gpointer*)((guint8*)code + (*(guint32*)(code - 4)));
  79                 InterlockedExchangePointer (got_entry, addr);
  80         }
  81 }
  82
  83 void
  84 mono_arch_patch_plt_entry (guint8 *code, guint8 *addr)
  85 {
  86         gint32 disp;
  87         gpointer *plt_jump_table_entry;
  88
  89         /* A PLT entry: jmp *<DISP>(%rip) */
  90         g_assert (code [0] == 0xff);
  91         g_assert (code [1] == 0x25);
  92
  93         disp = *(gint32*)(code + 2);
  94
  95         plt_jump_table_entry = (gpointer*)(code + 6 + disp);
  96
  97         InterlockedExchangePointer (plt_jump_table_entry, addr);
  98 }
  99
 100 void
 101 mono_arch_nullify_class_init_trampoline (guint8 *code, gssize *regs)
 102 {
 103         code -= 3;
 104
 105         /*
 106          * A given byte sequence can match more than case here, so we have to be
 107          * really careful about the ordering of the cases. Longer sequences
 108          * come first.
 109          */
 110         if ((code [-4] == 0x41) && (code [-3] == 0xff) && (code [-2] == 0x15)) {
 111                 gpointer *vtable_slot;
 112
 113                 /* call *<OFFSET>(%rip) */
 114                 vtable_slot = mono_arch_get_vcall_slot_addr (code + 3, (gpointer*)regs);
 115                 g_assert (vtable_slot);
 116
 117                 *vtable_slot = nullified_class_init_trampoline;
 118         } else if (code [-2] == 0xe8) {
 119                 /* call <TARGET> */
 120                 guint8 *buf = code - 2;
 121
 122                 buf [0] = 0x66;
 123                 buf [1] = 0x66;
 124                 buf [2] = 0x90;
 125                 buf [3] = 0x66;
 126                 buf [4] = 0x90;
 127         } else if ((code [0] == 0x49) && (code [1] == 0xff)) {
 128                 /* call <REG> */
 129                 /* amd64_set_reg_template is 10 bytes long */
 130                 guint8* buf = code - 10;
 131
 132                 /* FIXME: Make this thread safe */
 133                 /* Padding code suggested by the AMD64 Opt Manual */
 134                 buf [0] = 0x66;
 135                 buf [1] = 0x66;
 136                 buf [2] = 0x66;
 137                 buf [3] = 0x90;
 138                 buf [4] = 0x66;
 139                 buf [5] = 0x66;
 140                 buf [6] = 0x66;
 141                 buf [7] = 0x90;
 142                 buf [8] = 0x66;
 143                 buf [9] = 0x66;
 144                 buf [10] = 0x90;
 145                 buf [11] = 0x66;
 146                 buf [12] = 0x90;
 147         } else if (code [0] == 0x90 || code [0] == 0xeb || code [0] == 0x66) {
 148                 /* Already changed by another thread */
 149                 ;
 150         } else {
 151                 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", code [0], code [1], code [2], code [3],
 152                                 code [4], code [5], code [6]);
 153                 g_assert_not_reached ();
 154         }
 155 }
 156
 157 void
 158 mono_arch_nullify_plt_entry (guint8 *code)
 159 {
 160         mono_arch_patch_plt_entry (code, nullified_class_init_trampoline);
 161 }
 162
 163 void
 164 mono_arch_patch_delegate_trampoline (guint8 *code, guint8 *tramp, gssize *regs, guint8 *addr)
 165 {
 166         guint8 rex = 0;
 167         guint32 reg;
 168         guint32 disp;
 169
 170         if ((code [-3] == 0xff) && (amd64_modrm_reg (code [-2]) == 0x2) && (amd64_modrm_mod (code [-2]) == 0x1)) {
 171                 /* call *[reg+disp8] */
 172                 if (IS_REX (code [-4]))
 173                         rex = code [-4];
 174                 reg = amd64_modrm_rm (code [-2]);
 175                 disp = *(guint8*)(code - 1);
 176                 //printf ("B: [%%r%d+0x%x]\n", reg, disp);
 177         }
 178         else {
 179                 int i;
 180
 181                 for (i = -16; i < 0; ++i)
 182                         printf ("%d ", code [i]);
 183                 printf ("\n");
 184                 g_assert_not_reached ();
 185         }
 186
 187         reg += amd64_rex_b (rex);
 188
 189         /* R11 is clobbered by the trampoline code */
 190         g_assert (reg != AMD64_R11);
 191
 192         *(gpointer*)(((guint64)(regs [reg])) + disp) = addr;
 193 }
 194
 195 guchar*
 196 mono_arch_create_trampoline_code (MonoTrampolineType tramp_type)
 197 {
 198         guint8 *buf, *code, *tramp;
 199         int i, lmf_offset, offset, method_offset, tramp_offset, saved_regs_offset, saved_fpregs_offset, framesize;
 200         gboolean has_caller;
 201
 202         if (tramp_type == MONO_TRAMPOLINE_JUMP)
 203                 has_caller = FALSE;
 204         else
 205                 has_caller = TRUE;
 206
 207         code = buf = mono_global_codeman_reserve (512);
 208
 209         framesize = 512 + sizeof (MonoLMF);
 210         framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
 211
 212         offset = 0;
 213
 214         /*
 215          * Allocate a new stack frame and transfer the two arguments received on
 216          * the stack to our frame.
 217          */
 218         amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
 219         amd64_pop_reg (code, AMD64_R11);
 220
 221         amd64_push_reg (code, AMD64_RBP);
 222         amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, 8);
 223         amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
 224
 225         /*
 226          * The method is at offset -8 from the new RBP, so no need to
 227          * copy it.
 228          */
 229         offset += 8;
 230         method_offset = - offset;
 231
 232         offset += 8;
 233         tramp_offset = - offset;
 234         amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, 8);
 235
 236         /* Save all registers */
 237
 238         offset += AMD64_NREG * 8;
 239         saved_regs_offset = - offset;
 240         for (i = 0; i < AMD64_NREG; ++i)
 241                 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * 8), i, 8);
 242         offset += 8 * 8;
 243         saved_fpregs_offset = - offset;
 244         for (i = 0; i < 8; ++i)
 245                 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * 8), i);
 246
 247         /* Save LMF begin */
 248
 249         offset += sizeof (MonoLMF);
 250         lmf_offset = - offset;
 251
 252         /* Save ip */
 253         if (has_caller)
 254                 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, 8);
 255         else
 256                 amd64_mov_reg_imm (code, AMD64_R11, 0);
 257         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, 8);
 258         /* Save fp */
 259         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, framesize, 8);
 260         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebp), AMD64_R11, 8);
 261         /* Save sp */
 262         amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, 8);
 263         amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
 264         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, 8);
 265         /* Save method */
 266         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, method_offset, 8);
 267         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, 8);
 268         /* Save callee saved regs */
 269         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, 8);
 270         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, 8);
 271         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), AMD64_R13, 8);
 272         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, 8);
 273         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, 8);
 274
 275         amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
 276         amd64_call_reg (code, AMD64_R11);
 277
 278         /* Save lmf_addr */
 279         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, 8);
 280         /* Save previous_lmf */
 281         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, 8);
 282         amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, 8);
 283         /* Set new lmf */
 284         amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
 285         amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, 8);
 286
 287         /* Save LMF end */
 288
 289         /* Arg1 is the pointer to the saved registers */
 290         amd64_lea_membase (code, AMD64_RDI, AMD64_RBP, saved_regs_offset);
 291
 292         /* Arg2 is the address of the calling code */
 293         if (has_caller)
 294                 amd64_mov_reg_membase (code, AMD64_RSI, AMD64_RBP, 8, 8);
 295         else
 296                 amd64_mov_reg_imm (code, AMD64_RSI, 0);
 297
 298         /* Arg3 is the method/vtable ptr */
 299         amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RBP, method_offset, 8);
 300
 301         /* Arg4 is the trampoline address */
 302         amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, tramp_offset, 8);
 303
 304         if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT)
 305                 tramp = (guint8*)mono_class_init_trampoline;
 306         else if (tramp_type == MONO_TRAMPOLINE_AOT)
 307                 tramp = (guint8*)mono_aot_trampoline;
 308         else if (tramp_type == MONO_TRAMPOLINE_AOT_PLT)
 309                 tramp = (guint8*)mono_aot_plt_trampoline;
 310         else if (tramp_type == MONO_TRAMPOLINE_DELEGATE)
 311                 tramp = (guint8*)mono_delegate_trampoline;
 312         else
 313                 tramp = (guint8*)mono_magic_trampoline;
 314
 315         amd64_mov_reg_imm (code, AMD64_RAX, tramp);
 316         amd64_call_reg (code, AMD64_RAX);
 317
 318         /* Restore LMF */
 319
 320         amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 8);
 321         amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 8);
 322         amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, 8);
 323
 324         /* Restore argument registers */
 325         for (i = 0; i < AMD64_NREG; ++i)
 326                 if (AMD64_IS_ARGUMENT_REG (i))
 327                         amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * 8), 8);
 328
 329         for (i = 0; i < 8; ++i)
 330                 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * 8));
 331
 332         /* Restore stack */
 333         amd64_leave (code);
 334
 335         if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT)
 336                 amd64_ret (code);
 337         else
 338                 /* call the compiled method */
 339                 amd64_jump_reg (code, X86_EAX);
 340
 341         g_assert ((code - buf) <= 512);
 342
 343         mono_arch_flush_icache (buf, code - buf);
 344
 345         if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
 346                 /* Initialize the nullified class init trampoline used in the AOT case */
 347                 nullified_class_init_trampoline = code = mono_global_codeman_reserve (16);
 348                 x86_ret (code);
 349         }
 350
 351         return buf;
 352 }
 353
 354 #define TRAMPOLINE_SIZE 34
 355
 356 gpointer
 357 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
 358 {
 359         guint8 *code, *buf, *tramp, *real_code;
 360         int size, jump_offset;
 361
 362         tramp = mono_get_trampoline_code (tramp_type);
 363
 364         code = buf = g_alloca (TRAMPOLINE_SIZE);
 365
 366         /* push trampoline address */
 367         amd64_lea_membase (code, AMD64_R11, AMD64_RIP, -7);
 368         amd64_push_reg (code, AMD64_R11);
 369
 370         /* push argument */
 371         if (amd64_is_imm32 ((gint64)arg1))
 372                 amd64_push_imm (code, (gint64)arg1);
 373         else {
 374                 amd64_mov_reg_imm (code, AMD64_R11, arg1);
 375                 amd64_push_reg (code, AMD64_R11);
 376         }
 377
 378         jump_offset = code - buf;
 379         amd64_jump_disp (code, 0xffffffff);
 380
 381         g_assert ((code - buf) <= TRAMPOLINE_SIZE);
 382
 383         mono_domain_lock (domain);
 384         real_code = mono_code_manager_reserve (domain->code_mp, code - buf);
 385         size = code - buf;
 386         mono_domain_unlock (domain);
 387
 388         memcpy (real_code, buf, size);
 389
 390         /* Fix up jump */
 391         code = (guint8*)real_code + jump_offset;
 392         g_assert (amd64_is_imm32 (((gint64)tramp - (gint64)code)));
 393         amd64_jump_disp (code, tramp - code);
 394
 395         if (code_len)
 396                 *code_len = size;
 397
 398         mono_arch_flush_icache (real_code, size);
 399
 400         return real_code;
 401 }
 402
 403 void
 404 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
 405 {
 406         /* FIXME: This is not thread safe */
 407         guint8 *code = ji->code_start;
 408
 409         amd64_mov_reg_imm (code, AMD64_RDI, func_arg);
 410         amd64_mov_reg_imm (code, AMD64_R11, func);
 411
 412         x86_push_imm (code, (guint64)func_arg);
 413         amd64_call_reg (code, AMD64_R11);
 414 }
 415
 416 /*
 417  * This method is only called when running in the Mono Debugger.
 418  */
 419 guint8 *
 420 mono_debugger_create_notification_function (MonoCodeManager *codeman)
 421 {
 422         guint8 *buf, *code;
 423
 424         code = buf = mono_code_manager_reserve (codeman, 2);
 425         x86_breakpoint (buf);
 426         x86_ret (buf);
 427         return code;
 428 }