2 * tramp-x86.c: JIT trampoline code for x86
5 * Dietmar Maurer (dietmar@ximian.com)
7 * (C) 2001 Ximian, Inc.
13 #include <mono/metadata/abi-details.h>
14 #include <mono/metadata/appdomain.h>
15 #include <mono/metadata/metadata-internals.h>
16 #include <mono/metadata/marshal.h>
17 #include <mono/metadata/tabledefs.h>
18 #include <mono/metadata/mono-debug.h>
19 #include <mono/metadata/mono-debug-debugger.h>
20 #include <mono/metadata/monitor.h>
21 #include <mono/metadata/profiler-private.h>
22 #include <mono/metadata/gc-internal.h>
23 #include <mono/arch/x86/x86-codegen.h>
25 #include <mono/utils/memcheck.h>
30 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
33 * mono_arch_get_unbox_trampoline:
35 * @addr: pointer to native code for @m
37 * when value type methods are called through the vtable we need to unbox the
38 * this argument. This method returns a pointer to a trampoline which does
39 * unboxing before calling the method
42 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
45 int this_pos = 4, size = NACL_SIZE(16, 32);
46 MonoDomain *domain = mono_domain_get ();
48 start = code = mono_domain_code_reserve (domain, size);
50 x86_alu_membase_imm (code, X86_ADD, X86_ESP, this_pos, sizeof (MonoObject));
51 x86_jump_code (code, addr);
52 g_assert ((code - start) < size);
54 nacl_domain_code_validate (domain, &start, size, &code);
55 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_UNBOX_TRAMPOLINE, m);
61 mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr)
66 MonoDomain *domain = mono_domain_get ();
68 buf_len = NACL_SIZE (10, 32);
70 start = code = mono_domain_code_reserve (domain, buf_len);
72 x86_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, mrgctx);
73 x86_jump_code (code, addr);
74 g_assert ((code - start) <= buf_len);
76 nacl_domain_code_validate (domain, &start, buf_len, &code);
77 mono_arch_flush_icache (start, code - start);
78 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
84 mono_arch_get_llvm_imt_trampoline (MonoDomain *domain, MonoMethod *m, int vt_offset)
92 start = code = mono_domain_code_reserve (domain, buf_len);
94 this_offset = mono_x86_get_this_arg_offset (NULL, mono_method_signature (m));
97 x86_mov_reg_imm (code, MONO_ARCH_IMT_REG, m);
99 x86_mov_reg_membase (code, X86_EAX, X86_ESP, this_offset + 4, 4);
100 /* Load vtable address */
101 x86_mov_reg_membase (code, X86_EAX, X86_EAX, 0, 4);
102 x86_jump_membase (code, X86_EAX, vt_offset);
104 g_assert ((code - start) < buf_len);
106 nacl_domain_code_validate (domain, &start, buf_len, &code);
108 mono_arch_flush_icache (start, code - start);
109 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_IMT_TRAMPOLINE, NULL);
115 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
117 #if defined(__default_codegen__)
120 gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 8, buf, sizeof (buf));
124 /* go to the start of the call instruction
126 * address_byte = (m << 6) | (o << 3) | reg
127 * call opcode: 0xff address_byte displacement
133 if (code [1] == 0xe8) {
135 InterlockedExchange ((gint32*)(orig_code + 2), (guint)addr - ((guint)orig_code + 1) - 5);
137 /* Tell valgrind to recompile the patched code */
138 VALGRIND_DISCARD_TRANSLATIONS (orig_code + 2, 4);
140 } else if (code [1] == 0xe9) {
141 /* A PLT entry: jmp <DISP> */
143 InterlockedExchange ((gint32*)(orig_code + 2), (guint)addr - ((guint)orig_code + 1) - 5);
145 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", code [0], code [1], code [2], code [3],
146 code [4], code [5], code [6]);
147 g_assert_not_reached ();
149 #elif defined(__native_client__)
150 /* Target must be bundle-aligned */
151 g_assert (((guint32)addr & kNaClAlignmentMask) == 0);
153 /* 0xe8 = call <DISP>, 0xe9 = jump <DISP> */
154 if ((orig_code [-5] == 0xe8) || orig_code [-6] == 0xe9) {
156 gint32 offset = (gint32)addr - (gint32)orig_code;
157 guint8 buf[sizeof(gint32)];
158 *((gint32*)(buf)) = offset;
159 ret = nacl_dyncode_modify (orig_code - sizeof(gint32), buf, sizeof(gint32));
162 printf ("Invalid trampoline sequence %p: %02x %02x %02x %02x %02x\n", orig_code, orig_code [-5], orig_code [-4], orig_code [-3], orig_code [-2], orig_code[-1]);
163 g_assert_not_reached ();
169 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
173 /* Patch the jump table entry used by the plt entry */
175 #if defined(__native_client_codegen__) || defined(__native_client__)
176 /* for both compiler and runtime */
178 /* mov <DISP>(%ebx), %ecx */
179 /* and 0xffffffe0, %ecx */
181 g_assert (code [0] == 0x8b);
182 g_assert (code [1] == 0x8b);
184 offset = *(guint32*)(code + 2);
185 #elif defined(__default_codegen__)
186 /* A PLT entry: jmp *<DISP>(%ebx) */
187 g_assert (code [0] == 0xff);
188 g_assert (code [1] == 0xa3);
190 offset = *(guint32*)(code + 2);
191 #endif /* __native_client_codegen__ */
193 got = (gpointer*)(gsize) regs [MONO_ARCH_GOT_REG];
194 *(guint8**)((guint8*)got + offset) = addr;
198 get_vcall_slot (guint8 *code, mgreg_t *regs, int *displacement)
200 const int kBufSize = NACL_SIZE (8, 16);
205 mono_breakpoint_clean_code (NULL, code, kBufSize, buf, kBufSize);
210 if ((code [0] == 0xff) && ((code [1] & 0x18) == 0x10) && ((code [1] >> 6) == 2)) {
211 reg = code [1] & 0x07;
212 disp = *((gint32*)(code + 2));
213 #if defined(__native_client_codegen__) || defined(__native_client__)
214 } else if ((code[1] == 0x83) && (code[2] == 0xe1) && (code[4] == 0xff) &&
215 (code[5] == 0xd1) && (code[-5] == 0x8b)) {
216 disp = *((gint32*)(code - 3));
217 reg = code[-4] & 0x07;
218 } else if ((code[-2] == 0x8b) && (code[1] == 0x83) && (code[4] == 0xff)) {
219 reg = code[-1] & 0x07;
220 disp = (signed char)code[0];
223 g_assert_not_reached ();
227 *displacement = disp;
228 return (gpointer)regs [reg];
232 get_vcall_slot_addr (guint8* code, mgreg_t *regs)
236 vt = get_vcall_slot (code, regs, &displacement);
239 return (gpointer*)((char*)vt + displacement);
243 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
246 gboolean can_write = mono_breakpoint_clean_code (NULL, code, 6, buf, sizeof (buf));
247 gpointer tramp = mini_get_nullified_class_init_trampoline ();
253 if (code [0] == 0xe8) {
254 #if defined(__default_codegen__)
255 if (!mono_running_on_valgrind ()) {
258 * Thread safe code patching using the algorithm from the paper
259 * 'Practicing JUDO: Java Under Dynamic Optimizations'
262 * First atomically change the the first 2 bytes of the call to a
266 InterlockedExchange ((gint32*)code, ops);
268 /* Then change the other bytes to a nop */
273 /* Then atomically change the first 4 bytes to a nop as well */
275 InterlockedExchange ((gint32*)code, ops);
276 /* FIXME: the calltree skin trips on the self modifying code above */
278 /* Tell valgrind to recompile the patched code */
279 //VALGRIND_DISCARD_TRANSLATIONS (code, 8);
281 #elif defined(__native_client_codegen__)
282 mono_arch_patch_callsite (code, code + 5, tramp);
284 } else if (code [0] == 0x90 || code [0] == 0xeb) {
285 /* Already changed by another thread */
287 } else if ((code [-1] == 0xff) && (x86_modrm_reg (code [0]) == 0x2)) {
288 /* call *<OFFSET>(<REG>) -> Call made from AOT code */
289 gpointer *vtable_slot;
291 vtable_slot = get_vcall_slot_addr (code + 5, regs);
292 g_assert (vtable_slot);
294 *vtable_slot = tramp;
296 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", code [0], code [1], code [2], code [3],
297 code [4], code [5], code [6]);
298 g_assert_not_reached ();
303 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
306 guint8 *buf, *code, *tramp;
307 GSList *unwind_ops = NULL;
308 MonoJumpInfo *ji = NULL;
309 int i, offset, frame_size, regarray_offset, lmf_offset, caller_ip_offset, arg_offset;
311 unwind_ops = mono_arch_get_cie_program ();
313 code = buf = mono_global_codeman_reserve (256);
315 /* Note that there is a single argument to the trampoline
316 * and it is stored at: esp + pushed_args * sizeof (gpointer)
317 * the ret address is at: esp + (pushed_args + 1) * sizeof (gpointer)
320 // FIXME: Unwind info
322 /* Compute frame offsets relative to the frame pointer %ebp */
323 arg_offset = sizeof (mgreg_t);
324 caller_ip_offset = 2 * sizeof (mgreg_t);
326 offset += sizeof (MonoLMF);
327 lmf_offset = -offset;
328 offset += X86_NREG * sizeof (mgreg_t);
329 regarray_offset = -offset;
331 offset += 4 * sizeof (mgreg_t);
332 frame_size = ALIGN_TO (offset, MONO_ARCH_FRAME_ALIGNMENT);
335 x86_push_reg (code, X86_EBP);
336 x86_mov_reg_reg (code, X86_EBP, X86_ESP, sizeof (mgreg_t));
337 /* There are three words on the stack, adding + 4 aligns the stack to 16, which is needed on osx */
338 x86_alu_reg_imm (code, X86_SUB, X86_ESP, frame_size + sizeof (mgreg_t));
340 /* Save all registers */
341 for (i = X86_EAX; i <= X86_EDI; ++i) {
345 /* Save original ebp */
346 /* EAX is already saved */
347 x86_mov_reg_membase (code, X86_EAX, X86_EBP, 0, sizeof (mgreg_t));
349 } else if (i == X86_ESP) {
350 /* Save original esp */
351 /* EAX is already saved */
352 x86_mov_reg_reg (code, X86_EAX, X86_EBP, sizeof (mgreg_t));
353 /* Saved ebp + trampoline arg + return addr */
354 x86_alu_reg_imm (code, X86_ADD, X86_EAX, 3 * sizeof (mgreg_t));
357 x86_mov_membase_reg (code, X86_EBP, regarray_offset + (i * sizeof (mgreg_t)), reg, sizeof (mgreg_t));
362 if (tramp_type == MONO_TRAMPOLINE_JUMP) {
363 x86_mov_membase_imm (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, eip), 0, sizeof (mgreg_t));
365 x86_mov_reg_membase (code, X86_EAX, X86_EBP, caller_ip_offset, sizeof (mgreg_t));
366 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, eip), X86_EAX, sizeof (mgreg_t));
369 if ((tramp_type == MONO_TRAMPOLINE_JIT) || (tramp_type == MONO_TRAMPOLINE_JUMP)) {
370 x86_mov_reg_membase (code, X86_EAX, X86_EBP, arg_offset, sizeof (mgreg_t));
371 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), X86_EAX, sizeof (mgreg_t));
373 x86_mov_membase_imm (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, sizeof (mgreg_t));
376 x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_ESP * sizeof (mgreg_t)), sizeof (mgreg_t));
377 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, esp), X86_EAX, sizeof (mgreg_t));
378 /* callee save registers */
379 x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_EBX * sizeof (mgreg_t)), sizeof (mgreg_t));
380 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebx), X86_EAX, sizeof (mgreg_t));
381 x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_EDI * sizeof (mgreg_t)), sizeof (mgreg_t));
382 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, edi), X86_EAX, sizeof (mgreg_t));
383 x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_ESI * sizeof (mgreg_t)), sizeof (mgreg_t));
384 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, esi), X86_EAX, sizeof (mgreg_t));
385 x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_EBP * sizeof (mgreg_t)), sizeof (mgreg_t));
386 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebp), X86_EAX, sizeof (mgreg_t));
389 /* get the address of lmf for the current thread */
391 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
392 x86_call_reg (code, X86_EAX);
394 x86_call_code (code, mono_get_lmf_addr);
396 /* lmf->lmf_addr = lmf_addr (%eax) */
397 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), X86_EAX, sizeof (mgreg_t));
398 /* lmf->previous_lmf = *(lmf_addr) */
399 x86_mov_reg_membase (code, X86_ECX, X86_EAX, 0, sizeof (mgreg_t));
400 /* Signal to mono_arch_find_jit_info () that this is a trampoline frame */
401 x86_alu_reg_imm (code, X86_ADD, X86_ECX, 1);
402 x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), X86_ECX, sizeof (mgreg_t));
403 /* *lmf_addr = lmf */
404 x86_lea_membase (code, X86_ECX, X86_EBP, lmf_offset);
405 x86_mov_membase_reg (code, X86_EAX, 0, X86_ECX, sizeof (mgreg_t));
407 /* Call trampoline function */
408 /* Arg 1 - registers */
409 x86_lea_membase (code, X86_EAX, X86_EBP, regarray_offset);
410 x86_mov_membase_reg (code, X86_ESP, (0 * sizeof (mgreg_t)), X86_EAX, sizeof (mgreg_t));
411 /* Arg2 - calling code */
412 if (tramp_type == MONO_TRAMPOLINE_JUMP) {
413 x86_mov_membase_imm (code, X86_ESP, (1 * sizeof (mgreg_t)), 0, sizeof (mgreg_t));
415 x86_mov_reg_membase (code, X86_EAX, X86_EBP, caller_ip_offset, sizeof (mgreg_t));
416 x86_mov_membase_reg (code, X86_ESP, (1 * sizeof (mgreg_t)), X86_EAX, sizeof (mgreg_t));
418 /* Arg3 - trampoline argument */
419 x86_mov_reg_membase (code, X86_EAX, X86_EBP, arg_offset, sizeof (mgreg_t));
420 x86_mov_membase_reg (code, X86_ESP, (2 * sizeof (mgreg_t)), X86_EAX, sizeof (mgreg_t));
421 /* Arg4 - trampoline address */
423 x86_mov_membase_imm (code, X86_ESP, (3 * sizeof (mgreg_t)), 0, sizeof (mgreg_t));
426 /* check the stack is aligned after the ret ip is pushed */
428 x86_mov_reg_reg (code, X86_EDX, X86_ESP, 4);
429 x86_alu_reg_imm (code, X86_AND, X86_EDX, 15);
430 x86_alu_reg_imm (code, X86_CMP, X86_EDX, 0);
431 x86_branch_disp (code, X86_CC_Z, 3, FALSE);
432 x86_breakpoint (code);
437 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
438 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
439 x86_call_reg (code, X86_EAX);
441 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
442 x86_call_code (code, tramp);
446 * Overwrite the trampoline argument with the address we need to jump to,
449 x86_mov_membase_reg (code, X86_EBP, arg_offset, X86_EAX, 4);
451 /* Check for interruptions */
453 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
454 x86_call_reg (code, X86_EAX);
456 x86_call_code (code, (guint8*)mono_thread_force_interruption_checkpoint);
460 x86_mov_reg_membase (code, X86_EAX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sizeof (mgreg_t));
461 x86_mov_reg_membase (code, X86_ECX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof (mgreg_t));
462 x86_alu_reg_imm (code, X86_SUB, X86_ECX, 1);
463 x86_mov_membase_reg (code, X86_EAX, 0, X86_ECX, sizeof (mgreg_t));
465 /* Restore registers */
466 for (i = X86_EAX; i <= X86_EDI; ++i) {
467 if (i == X86_ESP || i == X86_EBP)
469 if (i == X86_EAX && !((tramp_type == MONO_TRAMPOLINE_RESTORE_STACK_PROT) || (tramp_type == MONO_TRAMPOLINE_AOT_PLT)))
471 x86_mov_reg_membase (code, i, X86_EBP, regarray_offset + (i * 4), 4);
477 if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
478 /* Load the value returned by the trampoline */
479 x86_mov_reg_membase (code, X86_EAX, X86_ESP, 0, 4);
480 /* The trampoline returns normally, pop the trampoline argument */
481 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
484 /* The trampoline argument is at the top of the stack, and it contains the address we need to branch to */
485 if (tramp_type == MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) {
486 x86_pop_reg (code, X86_EAX);
487 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 0x8);
488 x86_jump_reg (code, X86_EAX);
494 nacl_global_codeman_validate (&buf, 256, &code);
495 g_assert ((code - buf) <= 256);
496 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
499 tramp_name = mono_get_generic_trampoline_name (tramp_type);
500 *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);
508 mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
511 int tramp_size = NACL_SIZE (16, kNaClAlignment);
513 code = buf = mono_global_codeman_reserve (tramp_size);
516 nacl_global_codeman_validate (&buf, tramp_size, &code);
518 mono_arch_flush_icache (buf, code - buf);
519 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
522 *info = mono_tramp_info_create ("nullified_class_init_trampoline", buf, code - buf, NULL, NULL);
527 #define TRAMPOLINE_SIZE 10
530 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
532 guint8 *code, *buf, *tramp;
534 tramp = mono_get_trampoline_code (tramp_type);
536 code = buf = mono_domain_code_reserve_align (domain, TRAMPOLINE_SIZE, NACL_SIZE (4, kNaClAlignment));
538 x86_push_imm (buf, arg1);
539 x86_jump_code (buf, tramp);
540 g_assert ((buf - code) <= TRAMPOLINE_SIZE);
542 nacl_domain_code_validate (domain, &code, NACL_SIZE (4, kNaClAlignment), &buf);
544 mono_arch_flush_icache (code, buf - code);
545 mono_profiler_code_buffer_new (code, buf - code, MONO_PROFILER_CODE_BUFFER_SPECIFIC_TRAMPOLINE, mono_get_generic_trampoline_simple_name (tramp_type));
548 *code_len = buf - code;
554 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
558 guint8 **rgctx_null_jumps;
563 MonoJumpInfo *ji = NULL;
564 GSList *unwind_ops = NULL;
566 unwind_ops = mono_arch_get_cie_program ();
568 mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
569 index = MONO_RGCTX_SLOT_INDEX (slot);
571 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
572 for (depth = 0; ; ++depth) {
573 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
575 if (index < size - 1)
580 #if defined(__default_codegen__)
581 tramp_size = (aot ? 64 : 36) + 6 * depth;
582 #elif defined(__native_client_codegen__)
583 tramp_size = (aot ? 64 : 36) + 2 * kNaClAlignment +
584 6 * (depth + kNaClAlignment);
587 code = buf = mono_global_codeman_reserve (tramp_size);
589 rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
591 /* load vtable/mrgctx ptr */
592 x86_mov_reg_membase (code, X86_EAX, X86_ESP, 4, 4);
594 /* load rgctx ptr from vtable */
595 x86_mov_reg_membase (code, X86_EAX, X86_EAX, MONO_STRUCT_OFFSET (MonoVTable, runtime_generic_context), 4);
596 /* is the rgctx ptr null? */
597 x86_test_reg_reg (code, X86_EAX, X86_EAX);
598 /* if yes, jump to actual trampoline */
599 rgctx_null_jumps [0] = code;
600 x86_branch8 (code, X86_CC_Z, -1, 1);
603 for (i = 0; i < depth; ++i) {
604 /* load ptr to next array */
605 if (mrgctx && i == 0)
606 x86_mov_reg_membase (code, X86_EAX, X86_EAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, 4);
608 x86_mov_reg_membase (code, X86_EAX, X86_EAX, 0, 4);
609 /* is the ptr null? */
610 x86_test_reg_reg (code, X86_EAX, X86_EAX);
611 /* if yes, jump to actual trampoline */
612 rgctx_null_jumps [i + 1] = code;
613 x86_branch8 (code, X86_CC_Z, -1, 1);
617 x86_mov_reg_membase (code, X86_EAX, X86_EAX, sizeof (gpointer) * (index + 1), 4);
618 /* is the slot null? */
619 x86_test_reg_reg (code, X86_EAX, X86_EAX);
620 /* if yes, jump to actual trampoline */
621 rgctx_null_jumps [depth + 1] = code;
622 x86_branch8 (code, X86_CC_Z, -1, 1);
623 /* otherwise return */
626 for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
627 x86_patch (rgctx_null_jumps [i], code);
629 g_free (rgctx_null_jumps);
631 x86_mov_reg_membase (code, MONO_ARCH_VTABLE_REG, X86_ESP, 4, 4);
634 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
635 x86_jump_reg (code, X86_EAX);
637 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
639 /* jump to the actual trampoline */
640 x86_jump_code (code, tramp);
643 nacl_global_codeman_validate (&buf, tramp_size, &code);
644 mono_arch_flush_icache (buf, code - buf);
645 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
647 g_assert (code - buf <= tramp_size);
650 char *name = mono_get_rgctx_fetch_trampoline_name (slot);
651 *info = mono_tramp_info_create (name, buf, code - buf, ji, unwind_ops);
659 * mono_arch_create_general_rgctx_lazy_fetch_trampoline:
661 * This is a general variant of the rgctx fetch trampolines. It receives a pointer to gpointer[2] in the rgctx reg. The first entry contains the slot, the second
662 * the trampoline to call if the slot is not filled.
665 mono_arch_create_general_rgctx_lazy_fetch_trampoline (MonoTrampInfo **info, gboolean aot)
669 MonoJumpInfo *ji = NULL;
670 GSList *unwind_ops = NULL;
674 unwind_ops = mono_arch_get_cie_program ();
678 code = buf = mono_global_codeman_reserve (tramp_size);
680 // FIXME: Currently, we always go to the slow path.
682 /* Load trampoline addr */
683 x86_mov_reg_membase (code, X86_EAX, MONO_ARCH_RGCTX_REG, 4, 4);
684 /* Load mrgctx/vtable */
685 x86_mov_reg_membase (code, MONO_ARCH_VTABLE_REG, X86_ESP, 4, 4);
687 x86_jump_reg (code, X86_EAX);
689 nacl_global_codeman_validate (&buf, tramp_size, &code);
690 mono_arch_flush_icache (buf, code - buf);
691 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
693 g_assert (code - buf <= tramp_size);
696 *info = mono_tramp_info_create ("rgctx_fetch_trampoline_general", buf, code - buf, ji, unwind_ops);
702 mono_arch_create_generic_class_init_trampoline (MonoTrampInfo **info, gboolean aot)
706 static int byte_offset = -1;
707 static guint8 bitmask;
710 GSList *unwind_ops = NULL;
711 MonoJumpInfo *ji = NULL;
715 code = buf = mono_global_codeman_reserve (tramp_size);
717 unwind_ops = mono_arch_get_cie_program ();
720 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
722 x86_test_membase_imm (code, MONO_ARCH_VTABLE_REG, byte_offset, bitmask);
724 x86_branch8 (code, X86_CC_Z, -1, 1);
728 x86_patch (jump, code);
730 /* Push the vtable so the stack is the same as in a specific trampoline */
731 x86_push_reg (code, MONO_ARCH_VTABLE_REG);
734 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "generic_trampoline_generic_class_init");
735 x86_jump_reg (code, X86_EAX);
737 tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_GENERIC_CLASS_INIT);
739 /* jump to the actual trampoline */
740 x86_jump_code (code, tramp);
743 mono_arch_flush_icache (code, code - buf);
745 g_assert (code - buf <= tramp_size);
746 #ifdef __native_client_codegen__
747 g_assert (code - buf <= kNaClAlignment);
750 nacl_global_codeman_validate (&buf, tramp_size, &code);
751 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
754 *info = mono_tramp_info_create ("generic_class_init_trampoline", buf, code - buf, ji, unwind_ops);
759 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
761 * The code produced by this trampoline is equivalent to this:
764 * if (obj->synchronisation) {
765 * if (obj->synchronisation->owner == 0) {
766 * if (cmpxch (&obj->synchronisation->owner, TID, 0) == 0)
769 * if (obj->synchronisation->owner == TID) {
770 * ++obj->synchronisation->nest;
775 * return full_monitor_enter ();
779 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean is_v4, gboolean aot)
782 guint8 *jump_obj_null, *jump_sync_null, *jump_other_owner, *jump_cmpxchg_failed, *jump_tid, *jump_sync_thin_hash = NULL;
783 guint8 *jump_lock_taken_true = NULL;
785 int status_offset, nest_offset;
786 MonoJumpInfo *ji = NULL;
787 GSList *unwind_ops = NULL;
789 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == X86_EAX);
790 #ifdef MONO_ARCH_MONITOR_LOCK_TAKEN_REG
791 g_assert (MONO_ARCH_MONITOR_LOCK_TAKEN_REG == X86_EDX);
796 mono_monitor_threads_sync_members_offset (&status_offset, &nest_offset);
797 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (status_offset) == sizeof (guint32));
798 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
799 status_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (status_offset);
800 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
802 tramp_size = NACL_SIZE (128, 192);
804 code = buf = mono_global_codeman_reserve (tramp_size);
806 x86_push_reg (code, X86_EAX);
807 if (mono_thread_get_tls_offset () != -1) {
809 x86_test_membase_imm (code, X86_EDX, 0, 1);
810 /* if *lock_taken is 1, jump to actual trampoline */
811 jump_lock_taken_true = code;
812 x86_branch8 (code, X86_CC_NZ, -1, 1);
813 x86_push_reg (code, X86_EDX);
815 /* MonoObject* obj is in EAX */
817 x86_test_reg_reg (code, X86_EAX, X86_EAX);
818 /* if yes, jump to actual trampoline */
819 jump_obj_null = code;
820 x86_branch8 (code, X86_CC_Z, -1, 1);
822 /* load obj->synchronization to ECX */
823 x86_mov_reg_membase (code, X86_ECX, X86_EAX, MONO_STRUCT_OFFSET (MonoObject, synchronisation), 4);
825 if (mono_gc_is_moving ()) {
826 /*if bit zero is set it's a thin hash*/
827 /*FIXME use testb encoding*/
828 x86_test_reg_imm (code, X86_ECX, 0x01);
829 jump_sync_thin_hash = code;
830 x86_branch8 (code, X86_CC_NE, -1, 1);
832 /*clear bits used by the gc*/
833 x86_alu_reg_imm (code, X86_AND, X86_ECX, ~0x3);
836 /* is synchronization null? */
837 x86_test_reg_reg (code, X86_ECX, X86_ECX);
839 /* if yes, jump to actual trampoline */
840 jump_sync_null = code;
841 x86_branch8 (code, X86_CC_Z, -1, 1);
843 /* load MonoInternalThread* into EDX */
845 /* load_aotconst () puts the result into EAX */
846 x86_mov_reg_reg (code, X86_EDX, X86_EAX, sizeof (mgreg_t));
847 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_TLS_OFFSET, GINT_TO_POINTER (TLS_KEY_THREAD));
848 code = mono_x86_emit_tls_get_reg (code, X86_EAX, X86_EAX);
849 x86_xchg_reg_reg (code, X86_EAX, X86_EDX, sizeof (mgreg_t));
851 code = mono_x86_emit_tls_get (code, X86_EDX, mono_thread_get_tls_offset ());
853 /* load TID into EDX */
854 x86_mov_reg_membase (code, X86_EDX, X86_EDX, MONO_STRUCT_OFFSET (MonoInternalThread, small_id), 4);
856 /* is synchronization->owner free */
857 x86_mov_reg_membase (code, X86_EAX, X86_ECX, status_offset, 4);
858 x86_test_reg_imm (code, X86_EAX, OWNER_MASK);
859 /* if not, jump to next case */
861 x86_branch8 (code, X86_CC_NZ, -1, 1);
863 /* if yes, try a compare-exchange with the TID */
864 /* Form new status */
865 x86_alu_reg_reg (code, X86_OR, X86_EDX, X86_EAX);
866 /* compare and exchange */
867 x86_prefix (code, X86_LOCK_PREFIX);
868 x86_cmpxchg_membase_reg (code, X86_ECX, status_offset, X86_EDX);
869 /* if not successful, jump to actual trampoline */
870 jump_cmpxchg_failed = code;
871 x86_branch8 (code, X86_CC_NZ, -1, 1);
872 /* if successful, pop and return */
874 x86_pop_reg (code, X86_EDX);
875 x86_mov_membase_imm (code, X86_EDX, 0, 1, 1);
877 x86_pop_reg (code, X86_EAX);
880 /* next case: synchronization->owner is not null */
881 x86_patch (jump_tid, code);
882 /* is synchronization->owner == TID? */
883 x86_alu_reg_imm (code, X86_AND, X86_EAX, OWNER_MASK);
884 x86_alu_reg_reg (code, X86_CMP, X86_EAX, X86_EDX);
885 /* if not, jump to actual trampoline */
886 jump_other_owner = code;
887 x86_branch8 (code, X86_CC_NZ, -1, 1);
888 /* if yes, increment nest */
889 x86_inc_membase (code, X86_ECX, nest_offset);
891 x86_pop_reg (code, X86_EDX);
892 x86_mov_membase_imm (code, X86_EDX, 0, 1, 1);
894 x86_pop_reg (code, X86_EAX);
898 /* obj is pushed, jump to the actual trampoline */
899 x86_patch (jump_obj_null, code);
900 if (jump_sync_thin_hash)
901 x86_patch (jump_sync_thin_hash, code);
902 x86_patch (jump_sync_null, code);
903 x86_patch (jump_other_owner, code);
904 x86_patch (jump_cmpxchg_failed, code);
907 x86_pop_reg (code, X86_EDX);
908 x86_patch (jump_lock_taken_true, code);
913 /* We are calling the generic trampoline directly, the argument is pushed
914 * on the stack just like a specific trampoline.
917 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "generic_trampoline_monitor_enter_v4");
919 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "generic_trampoline_monitor_enter");
920 x86_jump_reg (code, X86_EAX);
923 x86_jump_code (code, mono_get_trampoline_code (MONO_TRAMPOLINE_MONITOR_ENTER_V4));
925 x86_jump_code (code, mono_get_trampoline_code (MONO_TRAMPOLINE_MONITOR_ENTER));
928 mono_arch_flush_icache (buf, code - buf);
929 g_assert (code - buf <= tramp_size);
931 nacl_global_codeman_validate (&buf, tramp_size, &code);
932 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_MONITOR, NULL);
936 *info = mono_tramp_info_create ("monitor_enter_v4_trampoline", buf, code - buf, ji, unwind_ops);
938 *info = mono_tramp_info_create ("monitor_enter_trampoline", buf, code - buf, ji, unwind_ops);
945 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
947 guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_MONITOR_EXIT);
949 guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned, *jump_sync_thin_hash = NULL;
950 guint8 *jump_next, *jump_cmpxchg_failed;
952 int status_offset, nest_offset;
953 MonoJumpInfo *ji = NULL;
954 GSList *unwind_ops = NULL;
956 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == X86_EAX);
958 mono_monitor_threads_sync_members_offset (&status_offset, &nest_offset);
959 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (status_offset) == sizeof (guint32));
960 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
961 status_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (status_offset);
962 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
964 tramp_size = NACL_SIZE (128, 192);
966 code = buf = mono_global_codeman_reserve (tramp_size);
968 x86_push_reg (code, X86_EAX);
969 if (mono_thread_get_tls_offset () != -1) {
970 /* MonoObject* obj is in EAX */
972 x86_test_reg_reg (code, X86_EAX, X86_EAX);
973 /* if yes, jump to actual trampoline */
974 jump_obj_null = code;
975 x86_branch8 (code, X86_CC_Z, -1, 1);
977 /* load obj->synchronization to ECX */
978 x86_mov_reg_membase (code, X86_ECX, X86_EAX, MONO_STRUCT_OFFSET (MonoObject, synchronisation), 4);
980 if (mono_gc_is_moving ()) {
981 /*if bit zero is set it's a thin hash*/
982 /*FIXME use testb encoding*/
983 x86_test_reg_imm (code, X86_ECX, 0x01);
984 jump_sync_thin_hash = code;
985 x86_branch8 (code, X86_CC_NE, -1, 1);
987 /*clear bits used by the gc*/
988 x86_alu_reg_imm (code, X86_AND, X86_ECX, ~0x3);
991 /* is synchronization null? */
992 x86_test_reg_reg (code, X86_ECX, X86_ECX);
993 /* if yes, jump to actual trampoline */
994 jump_sync_null = code;
995 x86_branch8 (code, X86_CC_Z, -1, 1);
997 /* next case: synchronization is not null */
998 /* load MonoInternalThread* into EDX */
1000 /* load_aotconst () puts the result into EAX */
1001 x86_mov_reg_reg (code, X86_EDX, X86_EAX, sizeof (mgreg_t));
1002 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_TLS_OFFSET, GINT_TO_POINTER (TLS_KEY_THREAD));
1003 code = mono_x86_emit_tls_get_reg (code, X86_EAX, X86_EAX);
1004 x86_xchg_reg_reg (code, X86_EAX, X86_EDX, sizeof (mgreg_t));
1006 code = mono_x86_emit_tls_get (code, X86_EDX, mono_thread_get_tls_offset ());
1008 /* load TID into EDX */
1009 x86_mov_reg_membase (code, X86_EDX, X86_EDX, MONO_STRUCT_OFFSET (MonoInternalThread, small_id), 4);
1010 /* is synchronization->owner == TID */
1011 x86_mov_reg_membase (code, X86_EAX, X86_ECX, status_offset, 4);
1012 x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EAX);
1013 x86_test_reg_imm (code, X86_EDX, OWNER_MASK);
1014 /* if no, jump to actual trampoline */
1015 jump_not_owned = code;
1016 x86_branch8 (code, X86_CC_NZ, -1, 1);
1018 /* next case: synchronization->owner == TID */
1019 /* is synchronization->nest == 1 */
1020 x86_alu_membase_imm (code, X86_CMP, X86_ECX, nest_offset, 1);
1021 /* if not, jump to next case */
1023 x86_branch8 (code, X86_CC_NZ, -1, 1);
1024 /* if yes, is synchronization->entry_count greater than zero? */
1025 x86_test_reg_imm (code, X86_EAX, ENTRY_COUNT_WAITERS);
1026 /* if yes, jump to actual trampoline */
1027 jump_have_waiters = code;
1028 x86_branch8 (code, X86_CC_NZ, -1 , 1);
1029 /* if not, try to set synchronization->owner to null and return */
1030 x86_mov_reg_reg (code, X86_EDX, X86_EAX, 4);
1031 x86_alu_reg_imm (code, X86_AND, X86_EDX, ENTRY_COUNT_MASK);
1032 /* compare and exchange */
1033 x86_prefix (code, X86_LOCK_PREFIX);
1034 /* EAX contains the previous status */
1035 x86_cmpxchg_membase_reg (code, X86_ECX, status_offset, X86_EDX);
1036 /* if not successful, jump to actual trampoline */
1037 jump_cmpxchg_failed = code;
1038 x86_branch8 (code, X86_CC_NZ, -1, 1);
1040 x86_pop_reg (code, X86_EAX);
1043 /* next case: synchronization->nest is not 1 */
1044 x86_patch (jump_next, code);
1045 /* decrease synchronization->nest and return */
1046 x86_dec_membase (code, X86_ECX, nest_offset);
1047 x86_pop_reg (code, X86_EAX);
1050 /* push obj and jump to the actual trampoline */
1051 x86_patch (jump_obj_null, code);
1052 if (jump_sync_thin_hash)
1053 x86_patch (jump_sync_thin_hash, code);
1054 x86_patch (jump_have_waiters, code);
1055 x86_patch (jump_cmpxchg_failed, code);
1056 x86_patch (jump_not_owned, code);
1057 x86_patch (jump_sync_null, code);
1060 /* obj is pushed, jump to the actual trampoline */
1062 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "generic_trampoline_monitor_exit");
1063 x86_jump_reg (code, X86_EAX);
1065 x86_jump_code (code, tramp);
1068 nacl_global_codeman_validate (&buf, tramp_size, &code);
1070 mono_arch_flush_icache (buf, code - buf);
1071 g_assert (code - buf <= tramp_size);
1072 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_MONITOR, NULL);
1075 *info = mono_tramp_info_create ("monitor_exit_trampoline", buf, code - buf, ji, unwind_ops);
1083 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean is_v4, gboolean aot)
1085 g_assert_not_reached ();
1090 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1092 g_assert_not_reached ();
1099 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1101 /* FIXME: This is not thread safe */
1102 guint8 *code = ji->code_start;
1104 x86_push_imm (code, func_arg);
1105 x86_call_code (code, (guint8*)func);
1109 handler_block_trampoline_helper (void)
1111 MonoJitTlsData *jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
1112 return jit_tls->handler_block_return_address;
1116 mono_arch_create_handler_block_trampoline (MonoTrampInfo **info, gboolean aot)
1118 guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
1120 int tramp_size = 64;
1121 MonoJumpInfo *ji = NULL;
1122 GSList *unwind_ops = NULL;
1126 code = buf = mono_global_codeman_reserve (tramp_size);
1129 This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
1133 * We are in a method frame after the call emitted by OP_CALL_HANDLER.
1136 if (mono_get_jit_tls_offset () != -1) {
1137 code = mono_x86_emit_tls_get (code, X86_EAX, mono_get_jit_tls_offset ());
1138 x86_mov_reg_membase (code, X86_EAX, X86_EAX, MONO_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 4);
1140 /*Slow path uses a c helper*/
1141 x86_call_code (code, handler_block_trampoline_helper);
1143 /* Simulate a call */
1144 /*Fix stack alignment*/
1145 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x4);
1146 /* This is the address the trampoline will return to */
1147 x86_push_reg (code, X86_EAX);
1148 /* Dummy trampoline argument, since we call the generic trampoline directly */
1149 x86_push_imm (code, 0);
1150 x86_jump_code (code, tramp);
1152 nacl_global_codeman_validate (&buf, tramp_size, &code);
1154 mono_arch_flush_icache (buf, code - buf);
1155 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
1156 g_assert (code - buf <= tramp_size);
1159 *info = mono_tramp_info_create ("handler_block_trampoline", buf, code - buf, ji, unwind_ops);
1165 mono_arch_get_call_target (guint8 *code)
1167 if (code [-5] == 0xe8) {
1168 gint32 disp = *(gint32*)(code - 4);
1169 guint8 *target = code + disp;
1178 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
1180 return *(guint32*)(plt_entry + NACL_SIZE (6, 12));
1184 * mono_arch_get_gsharedvt_arg_trampoline:
1186 * Return a trampoline which passes ARG to the gsharedvt in/out trampoline ADDR.
1189 mono_arch_get_gsharedvt_arg_trampoline (MonoDomain *domain, gpointer arg, gpointer addr)
1191 guint8 *code, *start;
1196 start = code = mono_domain_code_reserve (domain, buf_len);
1198 x86_mov_reg_imm (code, X86_EAX, arg);
1199 x86_jump_code (code, addr);
1200 g_assert ((code - start) <= buf_len);
1202 nacl_domain_code_validate (domain, &start, buf_len, &code);
1203 mono_arch_flush_icache (start, code - start);
1204 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
1209 #if defined(ENABLE_GSHAREDVT)
1211 #include "../../../mono-extensions/mono/mini/tramp-x86-gsharedvt.c"
1216 mono_arch_get_gsharedvt_trampoline (MonoTrampInfo **info, gboolean aot)
1223 #endif /* !MONOTOUCH */