#include <mono/metadata/marshal.h>
#include <mono/metadata/tabledefs.h>
#include <mono/metadata/mono-debug-debugger.h>
+#include <mono/metadata/monitor.h>
#include <mono/arch/amd64/amd64-codegen.h>
-#ifdef HAVE_VALGRIND_MEMCHECK_H
-#include <valgrind/memcheck.h>
-#endif
+#include <mono/utils/memcheck.h>
#include "mini.h"
#include "mini-amd64.h"
/*
* mono_arch_get_unbox_trampoline:
+ * @gsctx: the generic sharing context
* @m: method pointer
* @addr: pointer to native code for @m
*
* unboxing before calling the method
*/
gpointer
-mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
+mono_arch_get_unbox_trampoline (MonoGenericSharingContext *gsctx, MonoMethod *m, gpointer addr)
{
guint8 *code, *start;
int this_reg;
MonoDomain *domain = mono_domain_get ();
- this_reg = mono_arch_get_this_arg_reg (mono_method_signature (m));
+ this_reg = mono_arch_get_this_arg_reg (mono_method_signature (m), gsctx, NULL);
- mono_domain_lock (domain);
- start = code = mono_code_manager_reserve (domain->code_mp, 20);
- mono_domain_unlock (domain);
+ start = code = mono_domain_code_reserve (domain, 20);
amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
/* FIXME: Optimize this */
return start;
}
+/*
+ * mono_arch_get_static_rgctx_trampoline:
+ *
+ * Create a trampoline which sets RGCTX_REG to MRGCTX, then jumps to ADDR.
+ */
+gpointer
+mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr)
+{
+ guint8 *code, *start;
+ int buf_len;
+
+ MonoDomain *domain = mono_domain_get ();
+
+#ifdef MONO_ARCH_NOMAP32BIT
+ buf_len = 32;
+#else
+ /* AOTed code could still have a non-32 bit address */
+ if ((((guint64)addr) >> 32) == 0)
+ buf_len = 16;
+ else
+ buf_len = 30;
+#endif
+
+ start = code = mono_domain_code_reserve (domain, buf_len);
+
+ amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, mrgctx);
+ amd64_jump_code (code, addr);
+ g_assert ((code - start) < buf_len);
+
+ mono_arch_flush_icache (start, code - start);
+
+ return start;
+}
+
+gpointer
+mono_arch_get_llvm_imt_trampoline (MonoDomain *domain, MonoMethod *m, int vt_offset)
+{
+ guint8 *code, *start;
+ int buf_len;
+ int this_reg;
+
+ buf_len = 32;
+
+ start = code = mono_domain_code_reserve (domain, buf_len);
+
+ this_reg = mono_arch_get_this_arg_reg (mono_method_signature (m), NULL, NULL);
+
+ /* Set imt arg */
+ amd64_mov_reg_imm (code, MONO_ARCH_IMT_REG, m);
+ /* Load vtable address */
+ amd64_mov_reg_membase (code, AMD64_RAX, this_reg, 0, 8);
+ amd64_jump_membase (code, AMD64_RAX, vt_offset);
+ amd64_ret (code);
+
+ g_assert ((code - start) < buf_len);
+
+ mono_arch_flush_icache (start, code - start);
+
+ return start;
+}
+
/*
* mono_arch_patch_callsite:
*
* points to the pc right after the call.
*/
void
-mono_arch_patch_callsite (guint8 *orig_code, guint8 *addr)
+mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
{
guint8 *code;
guint8 buf [16];
- gboolean can_write = mono_breakpoint_clean_code (orig_code - 14, buf, sizeof (buf));
+ gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
code = buf + 14;
if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
if (code [-5] != 0xe8) {
- if (can_write)
+ if (can_write) {
InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
+ VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
+ }
} else {
if ((((guint64)(addr)) >> 32) != 0) {
+#ifdef MONO_ARCH_NOMAP32BIT
/* Print some diagnostics */
MonoJitInfo *ji = mono_jit_info_table_find (mono_domain_get (), (char*)orig_code);
if (ji)
if (ji)
fprintf (stderr, "Callee: %s\n", mono_method_full_name (ji->method, TRUE));
g_assert_not_reached ();
+#else
+ /*
+ * This might happen when calling AOTed code. Create a thunk.
+ */
+ guint8 *thunk_start, *thunk_code;
+
+ thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
+ amd64_jump_membase (thunk_code, AMD64_RIP, 0);
+ *(guint64*)thunk_code = (guint64)addr;
+ addr = thunk_start;
+ g_assert ((((guint64)(addr)) >> 32) == 0);
+ mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
+#endif
}
g_assert ((((guint64)(orig_code)) >> 32) == 0);
- if (can_write)
+ if (can_write) {
InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
+ VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
+ }
}
}
else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
/* call *<OFFSET>(%rip) */
gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
- if (can_write)
+ if (can_write) {
InterlockedExchangePointer (got_entry, addr);
+ VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
+ }
}
}
void
-mono_arch_patch_plt_entry (guint8 *code, guint8 *addr)
+mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
{
gint32 disp;
gpointer *plt_jump_table_entry;
}
void
-mono_arch_nullify_class_init_trampoline (guint8 *orig_code, gssize *regs)
+mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
{
- guint8 *code;
guint8 buf [16];
- gboolean can_write = mono_breakpoint_clean_code (orig_code - 7, buf, sizeof (buf));
+ MonoJitInfo *ji = NULL;
+ gboolean can_write;
+
+ if (mono_use_llvm) {
+ /* code - 7 might be before the start of the method */
+ /* FIXME: Avoid this expensive call somehow */
+ ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
+ }
- code = buf + 4;
+ can_write = mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 7, buf, sizeof (buf));
if (!can_write)
return;
* really careful about the ordering of the cases. Longer sequences
* come first.
*/
- if ((code [-4] == 0x41) && (code [-3] == 0xff) && (code [-2] == 0x15)) {
+ if ((buf [0] == 0x41) && (buf [1] == 0xff) && (buf [2] == 0x15)) {
gpointer *vtable_slot;
/* call *<OFFSET>(%rip) */
- vtable_slot = mono_arch_get_vcall_slot_addr (code + 3, (gpointer*)regs);
+ vtable_slot = mono_get_vcall_slot_addr (code, regs);
g_assert (vtable_slot);
*vtable_slot = nullified_class_init_trampoline;
- } else if (code [-2] == 0xe8) {
+ } else if (buf [2] == 0xe8) {
/* call <TARGET> */
- guint8 *buf = code - 2;
-
+ //guint8 *buf = code - 2;
+
+ /*
+ * It would be better to replace the call with nops, but that doesn't seem
+ * to work on SMP machines even when the whole call is inside a cache line.
+ * Patching the call address seems to work.
+ */
+ /*
buf [0] = 0x66;
buf [1] = 0x66;
buf [2] = 0x90;
buf [3] = 0x66;
buf [4] = 0x90;
- } else if ((code [0] == 0x41) && (code [1] == 0xff)) {
- /* call <REG> */
- /* happens on machines without MAP_32BIT like freebsd */
- /* amd64_set_reg_template is 10 bytes long */
- guint8* buf = code - 10;
-
- /* FIXME: Make this thread safe */
- /* Padding code suggested by the AMD64 Opt Manual */
- buf [0] = 0x66;
- buf [1] = 0x66;
- buf [2] = 0x66;
- buf [3] = 0x90;
- buf [4] = 0x66;
- buf [5] = 0x66;
- buf [6] = 0x66;
- buf [7] = 0x90;
- buf [8] = 0x66;
- buf [9] = 0x66;
- buf [10] = 0x90;
- buf [11] = 0x66;
- buf [12] = 0x90;
- } else if (code [0] == 0x90 || code [0] == 0xeb || code [0] == 0x66) {
+ */
+
+ mono_arch_patch_callsite (code - 5, code, nullified_class_init_trampoline);
+ } else if ((buf [5] == 0xff) && x86_modrm_mod (buf [6]) == 3 && x86_modrm_reg (buf [6]) == 2) {
+ /* call *<reg> */
+ /* Generated by the LLVM JIT or on platforms without MAP_32BIT set */
+ mono_arch_patch_callsite (code - 13, code, nullified_class_init_trampoline);
+ } else if (buf [4] == 0x90 || buf [5] == 0xeb || buf [6] == 0x66) {
/* Already changed by another thread */
;
} else {
- printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", code [0], code [1], code [2], code [3],
- code [4], code [5], code [6]);
+ printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", buf [0], buf [1], buf [2], buf [3],
+ buf [4], buf [5], buf [6]);
g_assert_not_reached ();
}
}
void
-mono_arch_nullify_plt_entry (guint8 *code)
+mono_arch_nullify_plt_entry (guint8 *code, mgreg_t *regs)
{
- mono_arch_patch_plt_entry (code, nullified_class_init_trampoline);
+ if (mono_aot_only && !nullified_class_init_trampoline)
+ nullified_class_init_trampoline = mono_aot_get_trampoline ("nullified_class_init_trampoline");
+
+ mono_arch_patch_plt_entry (code, NULL, regs, nullified_class_init_trampoline);
}
guchar*
-mono_arch_create_trampoline_code (MonoTrampolineType tramp_type)
+mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
{
- guint8 *buf, *code, *tramp, *br [2];
- int i, lmf_offset, offset, res_offset, arg_offset, tramp_offset, saved_regs_offset, saved_fpregs_offset, rbp_offset, framesize;
+ guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code;
+ int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset;
+ int buf_len, saved_regs_offset;
+ int saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
gboolean has_caller;
+ GSList *unwind_ops = NULL;
+ MonoJumpInfo *ji = NULL;
if (tramp_type == MONO_TRAMPOLINE_JUMP)
has_caller = FALSE;
else
has_caller = TRUE;
- code = buf = mono_global_codeman_reserve (512);
+ buf_len = 548;
+ code = buf = mono_global_codeman_reserve (buf_len);
- framesize = 512 + sizeof (MonoLMF);
+ framesize = 538 + sizeof (MonoLMF);
framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
- if (tramp_type == MONO_TRAMPOLINE_GENERIC_CLASS_INIT) {
- static int byte_offset = -1;
- static guint8 bitmask;
+ orig_rsp_to_rbp_offset = 0;
+ r11_save_code = code;
+ /* Reserve 5 bytes for the mov_membase_reg to save R11 */
+ code += 5;
+ after_r11_save_code = code;
- guint8 *jump;
+ // CFA = sp + 16 (the trampoline address is on the stack)
+ cfa_offset = 16;
+ mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
+ // IP saved at CFA - 8
+ mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
- if (byte_offset < 0)
- mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
+ /* Pop the return address off the stack */
+ amd64_pop_reg (code, AMD64_R11);
+ orig_rsp_to_rbp_offset += 8;
- amd64_test_membase_imm_size (code, MONO_ARCH_VTABLE_REG, byte_offset, bitmask, 1);
- jump = code;
- amd64_branch8 (code, X86_CC_Z, -1, 1);
-
- amd64_ret (code);
-
- x86_patch (jump, code);
- }
-
- /*
- * The generic class init trampoline is called directly by
- * JITted code, there is no specific trampoline. The lazy
- * fetch trampolines behave like generic class init
- * trampolines.
- */
- if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
- tramp_type != MONO_TRAMPOLINE_RGCTX_LAZY_FETCH) {
- /* Pop the return address off the stack */
- amd64_pop_reg (code, AMD64_R11);
- }
+ cfa_offset -= 8;
+ mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
/*
* Allocate a new stack frame
*/
amd64_push_reg (code, AMD64_RBP);
+ cfa_offset += 8;
+ mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
+ mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
+
+ orig_rsp_to_rbp_offset -= 8;
amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, 8);
+ mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
offset = 0;
rbp_offset = - offset;
+ offset += 8;
+ rax_offset = - offset;
+
offset += 8;
tramp_offset = - offset;
offset += 8;
arg_offset = - offset;
- if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
- tramp_type != MONO_TRAMPOLINE_RGCTX_LAZY_FETCH) {
- /* Compute the trampoline address from the return address */
+ /* Compute the trampoline address from the return address */
+ if (aot) {
+ /* 7 = length of call *<offset>(rip) */
+ amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
+ } else {
/* 5 = length of amd64_call_membase () */
amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
- amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, 8);
- } else {
- amd64_mov_membase_imm (code, AMD64_RBP, tramp_offset, 0, 8);
}
+ amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, 8);
offset += 8;
res_offset = - offset;
/* RAX is already saved */
amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, 8);
amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * 8), AMD64_RAX, 8);
- } else {
+ } else if (i != AMD64_R11) {
amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * 8), i, 8);
+ } else {
+ /* We have to save R11 right at the start of
+ the trampoline code because it's used as a
+ scratch register */
+ amd64_mov_membase_reg (r11_save_code, AMD64_RSP, saved_regs_offset + orig_rsp_to_rbp_offset + (i * 8), i, 8);
+ g_assert (r11_save_code == after_r11_save_code);
}
}
offset += 8 * 8;
amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * 8), i);
if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
- tramp_type != MONO_TRAMPOLINE_RGCTX_LAZY_FETCH) {
+ tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER &&
+ tramp_type != MONO_TRAMPOLINE_MONITOR_EXIT) {
/* Obtain the trampoline argument which is encoded in the instruction stream */
- amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, 8);
- amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
- amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
- amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
- br [0] = code;
- x86_branch8 (code, X86_CC_NE, 6, FALSE);
- /* 32 bit immediate */
- amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
- br [1] = code;
- x86_jump8 (code, 10);
- /* 64 bit immediate */
- mono_amd64_patch (br [0], code);
- amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
- mono_amd64_patch (br [1], code);
+ if (aot) {
+ /* Load the GOT offset */
+ amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, 8);
+ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 7, 4);
+ /* Compute the address of the GOT slot */
+ amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, 8);
+ /* Load the value */
+ amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
+ } else {
+ amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, 8);
+ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
+ amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
+ amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
+ br [0] = code;
+ x86_branch8 (code, X86_CC_NE, 6, FALSE);
+ /* 32 bit immediate */
+ amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
+ br [1] = code;
+ x86_jump8 (code, 10);
+ /* 64 bit immediate */
+ mono_amd64_patch (br [0], code);
+ amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
+ mono_amd64_patch (br [1], code);
+ }
amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, 8);
} else {
- amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, MONO_ARCH_VTABLE_REG, 8);
+ amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * 8), 8);
+ amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, 8);
}
/* Save LMF begin */
amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, 8);
/* Save method */
- if (tramp_type == MONO_TRAMPOLINE_GENERIC || tramp_type == MONO_TRAMPOLINE_JUMP) {
+ if (tramp_type == MONO_TRAMPOLINE_JIT || tramp_type == MONO_TRAMPOLINE_JUMP) {
amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, arg_offset, 8);
amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, 8);
} else {
amd64_mov_membase_imm (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, 8);
}
/* Save callee saved regs */
-#ifdef PLATFORM_WIN32
+#ifdef TARGET_WIN32
amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), AMD64_RDI, 8);
amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), AMD64_RSI, 8);
#endif
amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, 8);
amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, 8);
- amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
+ if (aot) {
+ code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
+ } else {
+ amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
+ }
amd64_call_reg (code, AMD64_R11);
/* Save lmf_addr */
/* Arg4 is the trampoline address */
amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, 8);
- tramp = (guint8*)mono_get_trampoline_func (tramp_type);
- amd64_mov_reg_imm (code, AMD64_RAX, tramp);
- amd64_call_reg (code, AMD64_RAX);
+ if (aot) {
+ char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
+ code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
+ } else {
+ tramp = (guint8*)mono_get_trampoline_func (tramp_type);
+ amd64_mov_reg_imm (code, AMD64_R11, tramp);
+ }
+ amd64_call_reg (code, AMD64_R11);
/* Check for thread interruption */
/* This is not perf critical code so no need to check the interrupt flag */
+ /*
+ * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
+ */
amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, 8);
- amd64_mov_reg_imm (code, AMD64_RAX, (guint8*)mono_thread_interruption_checkpoint);
- amd64_call_reg (code, AMD64_RAX);
+ if (aot) {
+ code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
+ } else {
+ amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint);
+ }
+ amd64_call_reg (code, AMD64_R11);
+
amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, 8);
/* Restore LMF */
amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 8);
amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, 8);
- /* Restore argument registers and r10 (needed to pass rgctx to
- static shared generic methods). */
+ /*
+ * Save rax to the stack, after the leave instruction, this will become part of
+ * the red zone.
+ */
+ amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, 8);
+
+ /* Restore argument registers, r10 (imt method/rgxtx)
+ and rax (needed for direct calls to C vararg functions). */
for (i = 0; i < AMD64_NREG; ++i)
- if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10)
+ if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * 8), 8);
for (i = 0; i < 8; ++i)
/* Restore stack */
amd64_leave (code);
- if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT ||
- tramp_type == MONO_TRAMPOLINE_GENERIC_CLASS_INIT ||
- tramp_type == MONO_TRAMPOLINE_RGCTX_LAZY_FETCH)
+ if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
+ /* Load result */
+ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - 0x8, 8);
amd64_ret (code);
- else
- /* call the compiled method */
- amd64_jump_reg (code, X86_EAX);
+ } else {
+ /* call the compiled method using the saved rax */
+ amd64_jump_membase (code, AMD64_RSP, rax_offset - 0x8);
+ }
- g_assert ((code - buf) <= 512);
+ g_assert ((code - buf) <= buf_len);
mono_arch_flush_icache (buf, code - buf);
if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
/* Initialize the nullified class init trampoline used in the AOT case */
- nullified_class_init_trampoline = code = mono_global_codeman_reserve (16);
- x86_ret (code);
+ nullified_class_init_trampoline = mono_arch_get_nullified_class_init_trampoline (NULL);
}
+ if (info)
+ *info = mono_tramp_info_create (mono_get_generic_trampoline_name (tramp_type), buf, code - buf, ji, unwind_ops);
+
+ return buf;
+}
+
+gpointer
+mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
+{
+ guint8 *code, *buf;
+
+ code = buf = mono_global_codeman_reserve (16);
+ amd64_ret (code);
+
+ mono_arch_flush_icache (buf, code - buf);
+
+ if (info)
+ *info = mono_tramp_info_create (g_strdup_printf ("nullified_class_init_trampoline"), buf, code - buf, NULL, NULL);
+
return buf;
}
else
size = 5 + 1 + 8;
- mono_domain_lock (domain);
- code = buf = mono_code_manager_reserve_align (domain->code_mp, size, 1);
- mono_domain_unlock (domain);
+ code = buf = mono_domain_code_reserve_align (domain, size, 1);
amd64_call_code (code, tramp);
/* The trampoline code will obtain the argument from the instruction stream */
}
gpointer
-mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 encoded_offset)
+mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
{
- guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_RGCTX_LAZY_FETCH);
- gboolean indirect = MONO_RGCTX_OFFSET_IS_INDIRECT (encoded_offset);
- int offset = indirect ? MONO_RGCTX_OFFSET_INDIRECT_OFFSET (encoded_offset) :
- MONO_RGCTX_OFFSET_DIRECT_OFFSET (encoded_offset);
- guint8 *code, *buf, *jump;
- guint8 *dummy;
- int slots_reg;
-
- g_assert (tramp);
- if (indirect)
- g_assert (MONO_RGCTX_OFFSET_INDIRECT_SLOT (encoded_offset) == 0);
-
- code = buf = mono_global_codeman_reserve (32);
-
- /* load slots ptr */
- if (indirect) {
- /* if indirect, load extra_other_infos ptr */
- amd64_mov_reg_membase (buf, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoRuntimeGenericContext, extra_other_infos), 8);
- slots_reg = AMD64_RAX;
+ guint8 *tramp;
+ guint8 *code, *buf;
+ guint8 **rgctx_null_jumps;
+ int tramp_size;
+ int depth, index;
+ int i;
+ gboolean mrgctx;
+ MonoJumpInfo *ji = NULL;
+ GSList *unwind_ops = NULL;
+
+ mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
+ index = MONO_RGCTX_SLOT_INDEX (slot);
+ if (mrgctx)
+ index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
+ for (depth = 0; ; ++depth) {
+ int size = mono_class_rgctx_get_array_size (depth, mrgctx);
+
+ if (index < size - 1)
+ break;
+ index -= size - 1;
+ }
+
+ tramp_size = 64 + 8 * depth;
+
+ code = buf = mono_global_codeman_reserve (tramp_size);
+
+ unwind_ops = mono_arch_get_cie_program ();
+
+ rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
+
+ if (mrgctx) {
+ /* get mrgctx ptr */
+ amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
} else {
- slots_reg = AMD64_ARG_REG1;
+ /* load rgctx ptr from vtable */
+ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), 8);
+ /* is the rgctx ptr null? */
+ amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
+ /* if yes, jump to actual trampoline */
+ rgctx_null_jumps [0] = code;
+ amd64_branch8 (code, X86_CC_Z, -1, 1);
}
- /* fetch slot */
- amd64_mov_reg_membase (buf, AMD64_RAX, slots_reg, offset, 8);
- dummy = buf;
+ for (i = 0; i < depth; ++i) {
+ /* load ptr to next array */
+ if (mrgctx && i == 0)
+ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, 8);
+ else
+ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, 8);
+ /* is the ptr null? */
+ amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
+ /* if yes, jump to actual trampoline */
+ rgctx_null_jumps [i + 1] = code;
+ amd64_branch8 (code, X86_CC_Z, -1, 1);
+ }
- /* is slot null? */
- amd64_test_reg_reg (buf, AMD64_RAX, AMD64_RAX);
- jump = buf;
+ /* fetch slot */
+ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), 8);
+ /* is the slot null? */
+ amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
/* if yes, jump to actual trampoline */
- amd64_branch8 (buf, X86_CC_Z, -1, 1);
+ rgctx_null_jumps [depth + 1] = code;
+ amd64_branch8 (code, X86_CC_Z, -1, 1);
+ /* otherwise return */
+ amd64_ret (code);
- /* if no, just return */
- amd64_ret (buf);
+ for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
+ x86_patch (rgctx_null_jumps [i], code);
+
+ g_free (rgctx_null_jumps);
- x86_patch (jump, buf);
/* move the rgctx pointer to the VTABLE register */
- amd64_mov_reg_reg (buf, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, 8);
- /* store the offset in RAX */
- amd64_mov_reg_imm (buf, AMD64_RAX, encoded_offset);
+ amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, 8);
+
+ if (aot) {
+ code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
+ amd64_jump_reg (code, AMD64_R11);
+ } else {
+ tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
+
+ /* jump to the actual trampoline */
+ amd64_jump_code (code, tramp);
+ }
+
+ mono_arch_flush_icache (buf, code - buf);
+
+ g_assert (code - buf <= tramp_size);
+
+ if (info)
+ *info = mono_tramp_info_create (mono_get_rgctx_fetch_trampoline_name (slot), buf, code - buf, ji, unwind_ops);
+
+ return buf;
+}
+
+gpointer
+mono_arch_create_generic_class_init_trampoline (MonoTrampInfo **info, gboolean aot)
+{
+ guint8 *tramp;
+ guint8 *code, *buf;
+ static int byte_offset = -1;
+ static guint8 bitmask;
+ guint8 *jump;
+ int tramp_size;
+ GSList *unwind_ops = NULL;
+ MonoJumpInfo *ji = NULL;
+
+ tramp_size = 64;
+
+ code = buf = mono_global_codeman_reserve (tramp_size);
+
+ if (byte_offset < 0)
+ mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
+
+ amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
+ jump = code;
+ amd64_branch8 (code, X86_CC_Z, -1, 1);
+
+ amd64_ret (code);
+
+ x86_patch (jump, code);
+
+ if (aot) {
+ code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
+ amd64_jump_reg (code, AMD64_R11);
+ } else {
+ tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
+
+ /* jump to the actual trampoline */
+ amd64_jump_code (code, tramp);
+ }
+
+ mono_arch_flush_icache (buf, code - buf);
+
+ g_assert (code - buf <= tramp_size);
+
+ if (info)
+ *info = mono_tramp_info_create (g_strdup_printf ("generic_class_init_trampoline"), buf, code - buf, ji, unwind_ops);
+
+ return buf;
+}
+
+#ifdef MONO_ARCH_MONITOR_OBJECT_REG
+
+gpointer
+mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean aot)
+{
+ guint8 *tramp;
+ guint8 *code, *buf;
+ guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid;
+ int tramp_size;
+ int owner_offset, nest_offset, dummy;
+ MonoJumpInfo *ji = NULL;
+ GSList *unwind_ops = NULL;
+
+ g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
+
+ mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
+ g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
+ g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
+ owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
+ nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
+
+ tramp_size = 96;
+
+ code = buf = mono_global_codeman_reserve (tramp_size);
+
+ if (mono_thread_get_tls_offset () != -1) {
+ /* MonoObject* obj is in RDI */
+ /* is obj null? */
+ amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
+ /* if yes, jump to actual trampoline */
+ jump_obj_null = code;
+ amd64_branch8 (code, X86_CC_Z, -1, 1);
+
+ /* load obj->synchronization to RCX */
+ amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
+ /* is synchronization null? */
+ amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
+ /* if yes, jump to actual trampoline */
+ jump_sync_null = code;
+ amd64_branch8 (code, X86_CC_Z, -1, 1);
+
+ /* load MonoInternalThread* into RDX */
+ code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
+ /* load TID into RDX */
+ amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
+
+ /* is synchronization->owner null? */
+ amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, owner_offset, 0, 8);
+ /* if not, jump to next case */
+ jump_tid = code;
+ amd64_branch8 (code, X86_CC_NZ, -1, 1);
+
+ /* if yes, try a compare-exchange with the TID */
+ /* zero RAX */
+ amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
+ /* compare and exchange */
+ amd64_prefix (code, X86_LOCK_PREFIX);
+ amd64_cmpxchg_membase_reg_size (code, AMD64_RCX, owner_offset, AMD64_RDX, 8);
+ /* if not successful, jump to actual trampoline */
+ jump_cmpxchg_failed = code;
+ amd64_branch8 (code, X86_CC_NZ, -1, 1);
+ /* if successful, return */
+ amd64_ret (code);
+
+ /* next case: synchronization->owner is not null */
+ x86_patch (jump_tid, code);
+ /* is synchronization->owner == TID? */
+ amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
+ /* if not, jump to actual trampoline */
+ jump_other_owner = code;
+ amd64_branch8 (code, X86_CC_NZ, -1, 1);
+ /* if yes, increment nest */
+ amd64_inc_membase_size (code, AMD64_RCX, nest_offset, 4);
+ /* return */
+ amd64_ret (code);
+
+ x86_patch (jump_obj_null, code);
+ x86_patch (jump_sync_null, code);
+ x86_patch (jump_cmpxchg_failed, code);
+ x86_patch (jump_other_owner, code);
+ }
+
/* jump to the actual trampoline */
- amd64_jump_code (buf, tramp);
+#if MONO_AMD64_ARG_REG1 != AMD64_RDI
+ amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
+#endif
+
+ if (aot) {
+ code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
+ amd64_jump_reg (code, AMD64_R11);
+ } else {
+ tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
+
+ /* jump to the actual trampoline */
+ amd64_jump_code (code, tramp);
+ }
- mono_arch_flush_icache (code, buf - code);
+ mono_arch_flush_icache (code, code - buf);
+ g_assert (code - buf <= tramp_size);
- g_assert (buf - code <= 32);
+ if (info)
+ *info = mono_tramp_info_create (g_strdup_printf ("monitor_enter_trampoline"), buf, code - buf, ji, unwind_ops);
- return code;
+ return buf;
}
-guint32
-mono_arch_get_rgctx_lazy_fetch_offset (gpointer *regs)
+gpointer
+mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
{
- return (guint32)(gulong)(regs [AMD64_RAX]);
+ guint8 *tramp;
+ guint8 *code, *buf;
+ guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned;
+ guint8 *jump_next;
+ int tramp_size;
+ int owner_offset, nest_offset, entry_count_offset;
+ MonoJumpInfo *ji = NULL;
+ GSList *unwind_ops = NULL;
+
+ g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
+
+ mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
+ g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
+ g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
+ g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
+ owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
+ nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
+ entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
+
+ tramp_size = 94;
+
+ code = buf = mono_global_codeman_reserve (tramp_size);
+
+ if (mono_thread_get_tls_offset () != -1) {
+ /* MonoObject* obj is in RDI */
+ /* is obj null? */
+ amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
+ /* if yes, jump to actual trampoline */
+ jump_obj_null = code;
+ amd64_branch8 (code, X86_CC_Z, -1, 1);
+
+ /* load obj->synchronization to RCX */
+ amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
+ /* is synchronization null? */
+ amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
+ /* if yes, jump to actual trampoline */
+ jump_sync_null = code;
+ amd64_branch8 (code, X86_CC_Z, -1, 1);
+
+ /* next case: synchronization is not null */
+ /* load MonoInternalThread* into RDX */
+ code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
+ /* load TID into RDX */
+ amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
+ /* is synchronization->owner == TID */
+ amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
+ /* if no, jump to actual trampoline */
+ jump_not_owned = code;
+ amd64_branch8 (code, X86_CC_NZ, -1, 1);
+
+ /* next case: synchronization->owner == TID */
+ /* is synchronization->nest == 1 */
+ amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, nest_offset, 1, 4);
+ /* if not, jump to next case */
+ jump_next = code;
+ amd64_branch8 (code, X86_CC_NZ, -1, 1);
+ /* if yes, is synchronization->entry_count zero? */
+ amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, entry_count_offset, 0, 4);
+ /* if not, jump to actual trampoline */
+ jump_have_waiters = code;
+ amd64_branch8 (code, X86_CC_NZ, -1 , 1);
+ /* if yes, set synchronization->owner to null and return */
+ amd64_mov_membase_imm (code, AMD64_RCX, owner_offset, 0, 8);
+ amd64_ret (code);
+
+ /* next case: synchronization->nest is not 1 */
+ x86_patch (jump_next, code);
+ /* decrease synchronization->nest and return */
+ amd64_dec_membase_size (code, AMD64_RCX, nest_offset, 4);
+ amd64_ret (code);
+
+ x86_patch (jump_obj_null, code);
+ x86_patch (jump_have_waiters, code);
+ x86_patch (jump_not_owned, code);
+ x86_patch (jump_sync_null, code);
+ }
+
+ /* jump to the actual trampoline */
+#if MONO_AMD64_ARG_REG1 != AMD64_RDI
+ amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
+#endif
+
+ if (aot) {
+ code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
+ amd64_jump_reg (code, AMD64_R11);
+ } else {
+ tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
+ amd64_jump_code (code, tramp);
+ }
+
+ mono_arch_flush_icache (code, code - buf);
+ g_assert (code - buf <= tramp_size);
+
+ if (info)
+ *info = mono_tramp_info_create (g_strdup_printf ("monitor_exit_trampoline"), buf, code - buf, ji, unwind_ops);
+
+ return buf;
}
+#endif
void
mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
}
/*
- * This method is only called when running in the Mono Debugger.
+ * mono_arch_get_call_target:
+ *
+ * Return the address called by the code before CODE if exists.
*/
-gpointer
-mono_debugger_create_notification_function (void)
+guint8*
+mono_arch_get_call_target (guint8 *code)
{
- guint8 *buf, *code;
+ if (code [-5] == 0xe8) {
+ guint32 disp = *(guint32*)(code - 4);
+ guint8 *target = code + disp;
+
+ return target;
+ } else {
+ return NULL;
+ }
+}
- code = buf = mono_global_codeman_reserve (2);
- x86_breakpoint (buf);
- x86_ret (buf);
- return code;
+/*
+ * mono_arch_get_plt_info_offset:
+ *
+ * Return the PLT info offset belonging to the plt entry PLT_ENTRY.
+ */
+guint32
+mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
+{
+ return *(guint32*)(plt_entry + 6);
}