-/*
- * mini-x86.c: x86 backend for the Mono code generator
+/**
+ * \file
+ * x86 backend for the Mono code generator
*
* Authors:
* Paolo Molaro (lupus@ximian.com)
case MONO_TYPE_U:
case MONO_TYPE_PTR:
case MONO_TYPE_FNPTR:
- case MONO_TYPE_CLASS:
case MONO_TYPE_OBJECT:
- case MONO_TYPE_SZARRAY:
- case MONO_TYPE_ARRAY:
- case MONO_TYPE_STRING:
cinfo->ret.storage = ArgInIReg;
cinfo->ret.reg = X86_EAX;
break;
case MONO_TYPE_U:
case MONO_TYPE_PTR:
case MONO_TYPE_FNPTR:
- case MONO_TYPE_CLASS:
case MONO_TYPE_OBJECT:
- case MONO_TYPE_STRING:
- case MONO_TYPE_SZARRAY:
- case MONO_TYPE_ARRAY:
add_general (&gr, param_regs, &stack_size, ainfo);
break;
case MONO_TYPE_GENERICINST:
* have locals larger than 8 bytes we need to make sure that
* they have the appropriate offset.
*/
- if (MONO_ARCH_FRAME_ALIGNMENT > 8 && locals_stack_align > 8)
- offset += MONO_ARCH_FRAME_ALIGNMENT - sizeof (gpointer) * 2;
+ if (MONO_ARCH_FRAME_ALIGNMENT > 8 && locals_stack_align > 8) {
+ int extra_size = MONO_ARCH_FRAME_ALIGNMENT - sizeof (gpointer) * 2;
+ offset += extra_size;
+ locals_stack_size += extra_size;
+ }
for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
if (offsets [i] != -1) {
MonoInst *inst = cfg->varinfo [i];
if (cfg->method->save_lmf) {
cfg->create_lmf_var = TRUE;
cfg->lmf_ir = TRUE;
-#ifndef HOST_WIN32
- cfg->lmf_ir_mono_lmf = TRUE;
-#endif
}
cfg->arch_eh_jit_info = 1;
}
}
+#define NEW_INS(cfg,ins,dest,op) do { \
+ MONO_INST_NEW ((cfg), (dest), (op)); \
+ (dest)->cil_code = (ins)->cil_code; \
+ mono_bblock_insert_before_ins (bb, ins, (dest)); \
+ } while (0)
+
/*
* mono_arch_lowering_pass:
*
break;
mono_decompose_op_imm (cfg, bb, ins);
break;
+#ifdef MONO_ARCH_SIMD_INTRINSICS
+ case OP_EXPAND_I1: {
+ MonoInst *temp;
+ int temp_reg1 = mono_alloc_ireg (cfg);
+ int temp_reg2 = mono_alloc_ireg (cfg);
+ int original_reg = ins->sreg1;
+
+ NEW_INS (cfg, ins, temp, OP_ICONV_TO_U1);
+ temp->sreg1 = original_reg;
+ temp->dreg = temp_reg1;
+
+ NEW_INS (cfg, ins, temp, OP_SHL_IMM);
+ temp->sreg1 = temp_reg1;
+ temp->dreg = temp_reg2;
+ temp->inst_imm = 8;
+
+ NEW_INS (cfg, ins, temp, OP_IOR);
+ temp->sreg1 = temp->dreg = temp_reg2;
+ temp->sreg2 = temp_reg1;
+
+ ins->opcode = OP_EXPAND_I2;
+ ins->sreg1 = temp_reg2;
+ }
+ break;
+#endif
default:
break;
}
return code;
}
-#ifdef __APPLE__
+#ifdef TARGET_MACH
static int tls_gs_offset;
#endif
gboolean
-mono_x86_have_tls_get (void)
+mono_arch_have_fast_tls (void)
{
#ifdef TARGET_MACH
- static gboolean have_tls_get = FALSE;
+ static gboolean have_fast_tls = FALSE;
static gboolean inited = FALSE;
+ guint32 *ins;
+ if (mini_get_debug_options ()->use_fallback_tls)
+ return FALSE;
if (inited)
- return have_tls_get;
-
-#ifdef MONO_HAVE_FAST_TLS
- guint32 *ins;
+ return have_fast_tls;
ins = (guint32*)pthread_getspecific;
/*
* mov 0x4(%esp),%eax
* mov %gs:[offset](,%eax,4),%eax
*/
- have_tls_get = ins [0] == 0x0424448b && ins [1] == 0x85048b65;
+ have_fast_tls = ins [0] == 0x0424448b && ins [1] == 0x85048b65;
tls_gs_offset = ins [2];
-#endif
-
inited = TRUE;
- return have_tls_get;
+ return have_fast_tls;
#elif defined(TARGET_ANDROID)
return FALSE;
#else
+ if (mini_get_debug_options ()->use_fallback_tls)
+ return FALSE;
return TRUE;
#endif
}
static guint8*
-mono_x86_emit_tls_set (guint8* code, int sreg, int tls_offset)
-{
-#if defined(__APPLE__)
- x86_prefix (code, X86_GS_PREFIX);
- x86_mov_mem_reg (code, tls_gs_offset + (tls_offset * 4), sreg, 4);
-#elif defined(TARGET_WIN32)
- g_assert_not_reached ();
-#else
- x86_prefix (code, X86_GS_PREFIX);
- x86_mov_mem_reg (code, tls_offset, sreg, 4);
-#endif
- return code;
-}
-
-/*
- * mono_x86_emit_tls_get:
- * @code: buffer to store code to
- * @dreg: hard register where to place the result
- * @tls_offset: offset info
- *
- * mono_x86_emit_tls_get emits in @code the native code that puts in
- * the dreg register the item in the thread local storage identified
- * by tls_offset.
- *
- * Returns: a pointer to the end of the stored code
- */
-guint8*
mono_x86_emit_tls_get (guint8* code, int dreg, int tls_offset)
{
-#if defined(__APPLE__)
+#if defined(TARGET_MACH)
x86_prefix (code, X86_GS_PREFIX);
x86_mov_reg_mem (code, dreg, tls_gs_offset + (tls_offset * 4), 4);
#elif defined(TARGET_WIN32)
- /*
- * See the Under the Hood article in the May 1996 issue of Microsoft Systems
+ /*
+ * See the Under the Hood article in the May 1996 issue of Microsoft Systems
* Journal and/or a disassembly of the TlsGet () function.
*/
x86_prefix (code, X86_FS_PREFIX);
}
static guint8*
-emit_tls_get_reg (guint8* code, int dreg, int offset_reg)
+mono_x86_emit_tls_set (guint8* code, int sreg, int tls_offset)
{
- /* offset_reg contains a value translated by mono_arch_translate_tls_offset () */
-#if defined(__APPLE__) || defined(__linux__)
- if (dreg != offset_reg)
- x86_mov_reg_reg (code, dreg, offset_reg, sizeof (mgreg_t));
+#if defined(TARGET_MACH)
x86_prefix (code, X86_GS_PREFIX);
- x86_mov_reg_membase (code, dreg, dreg, 0, sizeof (mgreg_t));
-#else
- g_assert_not_reached ();
-#endif
- return code;
-}
-
-guint8*
-mono_x86_emit_tls_get_reg (guint8* code, int dreg, int offset_reg)
-{
- return emit_tls_get_reg (code, dreg, offset_reg);
-}
-
-static guint8*
-emit_tls_set_reg (guint8* code, int sreg, int offset_reg)
-{
- /* offset_reg contains a value translated by mono_arch_translate_tls_offset () */
-#ifdef HOST_WIN32
+ x86_mov_mem_reg (code, tls_gs_offset + (tls_offset * 4), sreg, 4);
+#elif defined(TARGET_WIN32)
g_assert_not_reached ();
-#elif defined(__APPLE__) || defined(__linux__)
- x86_prefix (code, X86_GS_PREFIX);
- x86_mov_membase_reg (code, offset_reg, 0, sreg, sizeof (mgreg_t));
#else
- g_assert_not_reached ();
+ x86_prefix (code, X86_GS_PREFIX);
+ x86_mov_mem_reg (code, tls_offset, sreg, 4);
#endif
return code;
}
-
+
/*
* emit_setup_lmf:
*
/* Load ss_tramp_var */
/* This is equal to &ss_trampoline */
x86_mov_reg_membase (code, X86_ECX, var->inst_basereg, var->inst_offset, sizeof (mgreg_t));
- x86_alu_membase_imm (code, X86_CMP, X86_ECX, 0, 0);
+ x86_mov_reg_membase (code, X86_ECX, X86_ECX, 0, sizeof (mgreg_t));
+ x86_alu_reg_imm (code, X86_CMP, X86_ECX, 0);
br[0] = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
- x86_call_membase (code, X86_ECX, 0);
+ x86_call_reg (code, X86_ECX);
x86_patch (br [0], code);
}
code = mono_x86_emit_tls_get (code, ins->dreg, ins->inst_offset);
break;
}
- case OP_TLS_GET_REG: {
- code = emit_tls_get_reg (code, ins->dreg, ins->sreg1);
- break;
- }
case OP_TLS_SET: {
code = mono_x86_emit_tls_set (code, ins->sreg1, ins->inst_offset);
break;
}
- case OP_TLS_SET_REG: {
- code = emit_tls_set_reg (code, ins->sreg1, ins->sreg2);
- break;
- }
case OP_MEMORY_BARRIER: {
if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ) {
x86_prefix (code, X86_LOCK_PREFIX);
}
break;
- case OP_EXPAND_I1:
- /*FIXME this causes a partial register stall, maybe it would not be that bad to use shift + mask + or*/
- /*The +4 is to get a mov ?h, ?l over the same reg.*/
- x86_mov_reg_reg (code, ins->dreg + 4, ins->dreg, 1);
- x86_sse_alu_pd_reg_reg_imm (code, X86_SSE_PINSRW, ins->dreg, ins->sreg1, 0);
- x86_sse_alu_pd_reg_reg_imm (code, X86_SSE_PINSRW, ins->dreg, ins->sreg1, 1);
- x86_sse_shift_reg_imm (code, X86_SSE_PSHUFD, ins->dreg, ins->dreg, 0);
- break;
case OP_EXPAND_I2:
x86_sse_alu_pd_reg_reg_imm (code, X86_SSE_PINSRW, ins->dreg, ins->sreg1, 0);
x86_sse_alu_pd_reg_reg_imm (code, X86_SSE_PINSRW, ins->dreg, ins->sreg1, 1);
guint8 *patch;
/* check if we need to restore protection of the stack after a stack overflow */
- /* FIXME */
-#if 0
- if (supported) {
- if (cfg->compile_aot) {
- code = emit_load_aotconst (NULL, code, cfg, NULL, X86_ECX, MONO_PATCH_INFO_TLS_OFFSET, GINT_TO_POINTER (TLS_KEY_JIT_TLS));
-
- code = emit_tls_get_reg (code, X86_ECX, X86_ECX);
- } else {
- code = mono_x86_emit_tls_get (code, X86_ECX, mono_get_jit_tls_offset ());
- }
+ if (!cfg->compile_aot && mono_arch_have_fast_tls () && mono_tls_get_tls_offset (TLS_KEY_JIT_TLS) != -1) {
+ code = mono_x86_emit_tls_get (code, X86_ECX, mono_tls_get_tls_offset (TLS_KEY_JIT_TLS));
/* we load the value in a separate instruction: this mechanism may be
* used later as a safer way to do thread interruption
x86_mov_reg_membase (code, X86_ECX, X86_ECX, MONO_STRUCT_OFFSET (MonoJitTlsData, restore_stack_prot), 4);
x86_alu_reg_imm (code, X86_CMP, X86_ECX, 0);
patch = code;
- x86_branch8 (code, X86_CC_Z, 0, FALSE);
+ x86_branch8 (code, X86_CC_Z, 0, FALSE);
/* note that the call trampoline will preserve eax/edx */
x86_call_reg (code, X86_ECX);
x86_patch (patch, code);
- } else {
- /* FIXME: maybe save the jit tls in the prolog */
}
-#endif
/* restore caller saved regs */
if (cfg->used_int_regs & (1 << X86_EBX)) {
void
mono_arch_finish_init (void)
{
- if (!g_getenv ("MONO_NO_TLS")) {
+ char *mono_no_tls = g_getenv ("MONO_NO_TLS");
+ if (!mono_no_tls) {
#ifndef TARGET_WIN32
#if MONO_XEN_OPT
optimize_for_xen = access ("/proc/xen", F_OK) == 0;
#endif
#endif
- }
+ } else {
+ g_free (mono_no_tls);
+ }
}
void
}
/**
- * mono_breakpoint_clean_code:
+ * \return TRUE if no sw breakpoint was present.
*
- * Copy @size bytes from @code - @offset to the buffer @buf. If the debugger inserted software
+ * Copy \p size bytes from \p code - \p offset to the buffer \p buf. If the debugger inserted software
* breakpoints in the original code, they are removed in the copy.
- *
- * Returns TRUE if no sw breakpoint was present.
*/
gboolean
mono_breakpoint_clean_code (guint8 *method_start, guint8 *code, int offset, guint8 *buf, int size)